diff --git "a/stackoverflow_DL-related_questions_data/data22.json" "b/stackoverflow_DL-related_questions_data/data22.json"
new file mode 100644--- /dev/null
+++ "b/stackoverflow_DL-related_questions_data/data22.json"
@@ -0,0 +1,3000 @@
+{"QuestionId":57473527,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-13T08:02:41.707","AcceptedAnswerId":57474276.0,"OwnerUserId":10232932.0,"Title":"Text classification with IMDB dataset get_word_index error, how to change the path","Body":"<p>As I am behind a proxy, I cannot access the IMDB dataset directly, so I saved the complete package in:<\/p>\n\n<pre><code>C:\\Users\\...\\.keras\\datasets\\imdb.npz\n<\/code><\/pre>\n\n<p>I can successfully load the following command:<\/p>\n\n<pre><code># save np.load\nnp_load_old = np.load\n# modify the default parameters of np.load\nnp.load = lambda *a,**k: np_load_old(*a, allow_pickle=True, **k)\nimdb = keras.datasets.imdb\n(train_data, train_labels), (test_data, test_labels) = imdb.load_data(num_words=10000)\n# restore np.load for future normal usage\nnp.load = np_load_old\n<\/code><\/pre>\n\n<p>Where I have the train and test data.<\/p>\n\n<p>When I want to proceed with <a href=\"https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_text_classification\" rel=\"nofollow noreferrer\">basic text classification<\/a>, I have to get the word index:<\/p>\n\n<pre><code>word_index = imdb.get_word_index()\n<\/code><\/pre>\n\n<p>I (again) get the error: <\/p>\n\n<pre><code>Exception: URL fetch failure on https:\/\/storage.googleapis.com\/tensorflow\/tf-keras-datasets\/imdb_word_index.json: None -- [Errno 11004] getaddrinfo failed\n<\/code><\/pre>\n\n<p>How can I avoid this?<\/p>\n\n<p>I understand that I can adjust the path:<\/p>\n\n<pre><code>word_index = imdb.get_word_index(path='imdb_word_index.json')\n<\/code><\/pre>\n\n<p>But what is the right path in my case?<\/p>\n","answers":[{"AnswerId":"57474276","CreationDate":"2019-08-13T08:53:45.037","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Again, same drill, download the file at <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/tf-keras-datasets\/imdb_word_index.json\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/tf-keras-datasets\/imdb_word_index.json<\/a> and put it inside <code>C:\\Users\\&lt;your_username&gt;\\.keras\\datasets<\/code> and then use <code>get_word_index<\/code> as normal.<\/p>\n"}]}
+{"QuestionId":57474105,"AnswerCount":0,"Tags":"<c++><tensorflow>","CreationDate":"2019-08-13T08:42:50.963","AcceptedAnswerId":null,"OwnerUserId":10666066.0,"Title":"api_def.pb.h not found when building tensorflow","Body":"<p>I'm currently trying to build c++ <a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/cc\" rel=\"nofollow noreferrer\">tensorflow example<\/a> on windows and then i get the following error:<\/p>\n\n<blockquote>\n  <p>fatal error: tensorflow\/core\/framework\/api_def.pb.h: No such file or directory<\/p>\n<\/blockquote>\n\n<p>I checked and see that <code>api_def.pb.h<\/code> is missing in the directory. It's strange that I downloaded whole source code from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\" rel=\"nofollow noreferrer\">here<\/a>, which I think there should be nothing wrong with it since it's the official site. But then I realized that not only this file but all the requiring <code>.pb.h<\/code> files are missing. Such as <code>op_def.pb.h<\/code>, <code>function.pb.h<\/code>, <code>node_def.pb.h<\/code>, ...<\/p>\n\n<p>Though I can manually search on google, download and add other files to the appropriate directory, but in this case, all the result for <code>api_def.pb.h<\/code> file is about including it, none is about the content of raw file.<\/p>\n\n<p>My questions are:<\/p>\n\n<ul>\n<li>Where\/How can I get this file ?<\/li>\n<li>What is the real problem behind these missing files ?<\/li>\n<\/ul>\n\n<p>ps: I've done some projects with tensorflow on python, this is the first time I have to do it with C++.<\/p>\n","answers":[]}
+{"QuestionId":57474251,"AnswerCount":0,"Tags":"<python><numpy><tensorflow>","CreationDate":"2019-08-13T08:52:24.760","AcceptedAnswerId":null,"OwnerUserId":10232932.0,"Title":"Allowing pickle got multiple agruments","Body":"<p>Yesterday I sucessfully loaded the IMDB dataset to my jupyter notebook, with the following code:<\/p>\n\n<pre><code>from __future__ import absolute_import, division, print_function, unicode_literals\nimport tensorflow as tf\nfrom tensorflow import keras\nimport numpy as np\n\n# save np.load\nnp_load_old = np.load\n\n# modify the default parameters of np.load\nnp.load = lambda *a,**k: np_load_old(*a, allow_pickle=True, **k)\n\nimdb = keras.datasets.imdb\n\n(train_data, train_labels), (test_data, test_labels) = imdb.load_data(num_words=10000)\n\n# restore np.load for future normal usage\nnp.load = np_load_old\n<\/code><\/pre>\n\n<p>Now he gives me the error:<\/p>\n\n<pre><code>TypeError: &lt;lambda&gt;() got multiple values for keyword argument 'allow_pickle'\n<\/code><\/pre>\n\n<p>What do I am missing here? <\/p>\n\n<p>I am also following this <a href=\"https:\/\/stackoverflow.com\/questions\/55890813\/how-to-fix-object-arrays-cannot-be-loaded-when-allow-pickle-false-for-imdb-loa\">thread<\/a>.<\/p>\n","answers":[]}
+{"QuestionId":57474750,"AnswerCount":0,"Tags":"<python><keras><word2vec>","CreationDate":"2019-08-13T09:23:25.343","AcceptedAnswerId":null,"OwnerUserId":8170648.0,"Title":"I don't know why my word2vec is not trained","Body":"<p>I tried to generate skip-gram word2vec with Keras. \nI think there is no logical error... \nBut loss function score does not decrease. <\/p>\n\n<pre><code>=====================================================================\n58693\/58693 [==============================] - 342s 6ms\/step - loss: 0.6937\n\nEpoch 2\/100\n\n58693\/58693 [==============================] - 337s 6ms\/step - loss: 0.6935\n\nEpoch 3\/100\n\n58693\/58693 [==============================] - 336s 6ms\/step - loss: 0.6934\n\n=====================================================================\n<\/code><\/pre>\n\n<p>I based on this page. (<a href=\"https:\/\/adventuresinmachinelearning.com\/word2vec-keras-tutorial\/\" rel=\"nofollow noreferrer\">https:\/\/adventuresinmachinelearning.com\/word2vec-keras-tutorial\/<\/a>)<\/p>\n\n<p>Some functions, download data, read data and collect data, are came from that page. <\/p>\n\n<pre><code>sampling_table = sequence.make_sampling_table(vocab_size)\n# print (sampling_table)\ncouples, labels = skipgrams(data, vocab_size,     window_size=window_size, sampling_table=sampling_table)\nword_target, word_context = zip(*couples)\nword_target = np.array(word_target, dtype=\"int32\")\nword_context = np.array(word_context, dtype=\"int32\")\nlabels = np.array(labels, dtype='int32')\n\n\n\ndef trainingDataGen(batch_size):\n    while True:\n        idx = np.random.randint(0, len(labels) - 1,batch_size)\n\n        x_target = word_target[idx]\n        x_context = word_context[idx]\n        y_label = labels[idx]\n\n        yield [x_target, x_context], y_label\n\n\ninput_target = keras.layers.Input(shape=(1,), name='targetWord')\ninput_context = keras.layers.Input(shape=(1,), name='contextWord')\n\nembedding = Embedding(vocab_size, vector_dim, input_shape=(1,),trainable=True,name='embedding')\nembedded_target = embedding(input_target)\nembedded_context = embedding(input_context)\ndot_product = keras.layers.Dot(axes=-1)([embedded_target, embedded_context])\nflattened = keras.layers.Flatten()(dot_product)\nactivated = keras.layers.Dense(1,  activation=tf.nn.sigmoid)(flattened)\nmodel = keras.Model([input_target, input_context], activated)\nmodel.summary()\nmodel.compile(optimizer='sgd',loss='binary_crossentropy')\n\nsteps_per_epoch = divmod(len(labels), 512)[0]+1\nprint (steps_per_epoch)\nmodel.fit_generator(trainingDataGen(512),steps_per_epoch=steps_per_epoch, epochs=100)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57474898,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-08-13T09:31:51.790","AcceptedAnswerId":null,"OwnerUserId":11921533.0,"Title":"How to split one Keras model among multiple GPUs","Body":"<p>I work on medical imaging so I need to train massive 3D CNNs that are difficult to fit into one GPU. I wonder if there is way to split a massive Keras or TensorFlow graph amongst multiple GPUs such that each GPU only runs a small part of the graph during training and inference. Is this type of distribute training possible with either Keras or TensorFlow?<\/p>\n\n<p>I have tried using <code>with tf.device('\\gpu:#')<\/code> when building the graph but I am experiencing memory overflow. The logs seem to indicate the entire graph is still being run on <code>gpu:0<\/code>.<\/p>\n","answers":[]}
+{"QuestionId":57475127,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><generative-adversarial-network>","CreationDate":"2019-08-13T09:44:53.333","AcceptedAnswerId":null,"OwnerUserId":11921653.0,"Title":"Evaluating the critic score in WGANs","Body":"<p>I've been trying to train a WGAN for the last couple of days with gradient penalty involved. I took the gradient penalty code off of a github tensorflow implementation by ChengBinJin.<\/p>\n\n<p>With a normal DCGAN you'd be able to tell what the accuracy of the discriminator is at any point, cause it's trying to learn logits you can throw into a sigmoid function. So if I threw in real images, the accuracy would be close to a 100%, very straight-forward.<\/p>\n\n<p>However with respect to WGANs, the discriminator is now a critic and it outputs a score instead which isn't really translatable into accuracy as far as I can tell. Right now I'm at 3000 iterations and the mean score for real images is at -59,000. So how would one go about trying to gauge accuracy from this score?<\/p>\n","answers":[{"AnswerId":"57477285","CreationDate":"2019-08-13T11:58:49.960","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Not at all. The Wasserstein Critic is mean independant as it is written as f(x) - f(y). So So a function g(x) = f(x) + b has the same Wasserstein distance. E.g. g(x) - g(y) = f(x) + b - f(y) - b = f(x) - f(y).<\/p>\n\n<p>So the mean gives you no information. What does give you information is the difference between the means of the real and the fake images, e.g. the Wasserstein distance. The smaller the better. <\/p>\n"}]}
+{"QuestionId":57475131,"AnswerCount":0,"Tags":"<tensorflow><keras><lstm>","CreationDate":"2019-08-13T09:44:56.520","AcceptedAnswerId":null,"OwnerUserId":11921568.0,"Title":"Expected to see 1 array(s), but instead got the following list of 2 arrays:","Body":"<p>I am attempting to make a multilayer model with the functional API from keras. The idea is to take an vector that describes an image (comes from a pretrained model) and concatenate it with another vector that comes from the text associated with that image. The first branch processes the input coming from the images, the second branch takes as input the text. Both are concatenated and then processed together with a few more layers. <\/p>\n\n<p>When I train the model with:<\/p>\n\n<pre><code>model = Model(inputs=[xa.input, xb.input], outputs=z)\n<\/code><\/pre>\n\n<p>and this fit call:<\/p>\n\n<pre><code>model.fit([oraciones_img, X], y, epochs=2, batch_size=5, verbose=1)\n<\/code><\/pre>\n\n<p>everything works out fine.<\/p>\n\n<p>The problem comes when I want to do a prediction. I provided two arrays with the same dimensions as oraciones_img and X to the model.predict function but it throws this error:<\/p>\n\n<hr>\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 2 arrays: [array([[1.1081828 , 0.4828459 , 1.6334529 , ..., 0.03270344, 0.17258629,\n    0.03792314],\n   [0.8138524 , 0.        , 0.65887845, ..., 0.6183274 , 0.00807555,\n    0.6506448 ],\n   [0.8...\n<\/code><\/pre>\n\n<hr>\n\n<p>I double checked that the dimensions from the inputs during training and testing are the same but did not work out. I tried the training data on the predict function, assuming that at least that should work out, but weirdly that throws the same error. In theory, I should be providing this:<\/p>\n\n<pre><code>model.predict([ (n_samples, 405504) , (n_samples, 50) ])\n<\/code><\/pre>\n\n<p>where:\n(n_samples, 405504) are the vectors coming from n images\n(n_samples, 50)     are the vectors coming from n texts associated to the images<\/p>\n\n<p>but somehow the model is asking me for a single array. I provided one to follow up, and the predict function is asking for a single array of dim4 which is a bit more confusing. <\/p>\n\n<p>I have seen other posts regarding similar issues on the shape of the data, but all of them have problems running the fit function. I already passed that, this is about the predict function.<\/p>\n\n<p>Any idea of what is going on?<\/p>\n\n<h1>define model<\/h1>\n\n<pre><code>seq_length = 50\nseq_length_img = (200704 + 2048)*2\n\nInputA = Input(shape=(seq_length_img,))\nInputB = Input(shape=(seq_length,))\n\nxa = Dense(512, activation='relu')(InputA)\nxa = Model(inputs = InputA, outputs= xa)\n\nxb = Embedding(input_dim=vocab_size, output_dim=512, input_length=seq_length)(InputB)\nxb = Flatten()(xb)\nxb = Dense(512, activation='relu')(xb)\nxb = Model(inputs = InputB, outputs= xb)\n\ncombined = concatenate([xa.output, xb.output], axis=-1)\ncombined = Reshape((1024, 1))(combined)\n\nz = LSTM(1024, input_shape=(1024, 1), return_sequences=True)(combined)\nz = LSTM(1024)(z)\nz = Dense(100, activation='relu')(z)\nz = Dense(vocab_size, activation='softmax')(z)\n\nmodel = Model(inputs=[xa.input, xb.input], outputs=z)\nmodel.compile(optimizer='rmsprop', loss='binary_crossentropy', metrics=['accuracy'])\nmodel.fit([oraciones_img, X], y, epochs=2, batch_size=5, verbose=1)\n\nyhat = model.predict([feat, pred_sentence])\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57475348,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><runtime-error><tf.keras>","CreationDate":"2019-08-13T09:57:01.517","AcceptedAnswerId":57476895.0,"OwnerUserId":10609889.0,"Title":"Tensorflow, Keras pretrained MobileNetV2 Model doesn't download","Body":"<p>I following this tutorial on on transferred learning in tensorflow keras. An error occurs when it tries to download the mobilenetV2 model.<\/p>\n\n<p>This is the code that fails:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>base_model = tf.keras.applications.MobileNetV2(input_shape=IMG_SHAPE,\n                                             include_top=False,\n                                             weights='imagenet')\n<\/code><\/pre>\n\n<p>Throwing this error:<\/p>\n\n<pre><code>Downloading data from https:\/\/github.com\/JonathanCMitchell\/mobilenet_v2_keras\/releases\/download\/v1.1\/mobilenet_v2_weights_tf_dim_ordering_tf_kernels_1.0_224.h5\n<\/code><\/pre>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/urllib\/request.py\", line 1317, in do_open\n    encode_chunked=req.has_header('Transfer-encoding'))\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/http\/client.py\", line 1229, in request\n    self._send_request(method, url, body, headers, encode_chunked)\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/http\/client.py\", line 1275, in _send_request\n    self.endheaders(body, encode_chunked=encode_chunked)\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/http\/client.py\", line 1224, in endheaders\n    self._send_output(message_body, encode_chunked=encode_chunked)\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/http\/client.py\", line 1016, in _send_output\n    self.send(msg)\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/http\/client.py\", line 956, in send\n    self.connect()\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/http\/client.py\", line 1384, in connect\n    super().connect()\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/http\/client.py\", line 928, in connect\n    (self.host,self.port), self.timeout, self.source_address)\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/socket.py\", line 707, in create_connection\n    for res in getaddrinfo(host, port, 0, SOCK_STREAM):\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/socket.py\", line 748, in getaddrinfo\n    for res in _socket.getaddrinfo(host, port, family, type, proto, flags):\nsocket.gaierror: [Errno 8] nodename nor servname provided, or not known\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 \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/utils\/data_utils.py\", line 248, in get_file\n    urlretrieve(origin, fpath, dl_progress)\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/urllib\/request.py\", line 247, in urlretrieve\n    with contextlib.closing(urlopen(url, data)) as fp:\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/urllib\/request.py\", line 222, in urlopen\n    return opener.open(url, data, timeout)\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/urllib\/request.py\", line 525, in open\n    response = self._open(req, data)\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/urllib\/request.py\", line 543, in _open\n    '_open', req)\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/urllib\/request.py\", line 503, in _call_chain\n    result = func(*args)\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/urllib\/request.py\", line 1360, in https_open\n    context=self._context, check_hostname=self._check_hostname)\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/urllib\/request.py\", line 1319, in do_open\n    raise URLError(err)\nurllib.error.URLError: &lt;urlopen error [Errno 8] nodename nor servname provided, or not known&gt;\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 \"\/Users\/user_name\/PycharmProjects\/ObjDectCount\/modelEval.py\", line 62, in &lt;module&gt;\n    base_model = tf.keras.applications.MobileNetV2()\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/applications\/__init__.py\", line 70, in wrapper\n    return base_fun(*args, **kwargs)\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/applications\/mobilenet_v2.py\", line 32, in MobileNetV2\n    return mobilenet_v2.MobileNetV2(*args, **kwargs)\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/site-packages\/keras_applications\/mobilenet_v2.py\", line 411, in MobileNetV2\n    model_name, weight_path, cache_subdir='models')\n  File \"\/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/utils\/data_utils.py\", line 250, in get_file\n    raise Exception(error_msg.format(origin, e.errno, e.reason))\nException: URL fetch failure on https:\/\/github.com\/JonathanCMitchell\/mobilenet_v2_keras\/releases\/download\/v1.1\/mobilenet_v2_weights_tf_dim_ordering_tf_kernels_1.0_224.h5: None -- [Errno 8] nodename nor servname provided, or not known\n<\/code><\/pre>\n\n<p>EDIT: \nOutput after running <code>pip install --upgrade requests<\/code><\/p>\n\n<pre><code>Requirement already up-to-date: \nrequests in \/Users\/user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/site-packages (2.22.0)\nRequirement already satisfied, skipping upgrade: certifi&gt;=2017.4.17 in \/Users\/ user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/site-packages (from requests) (2019.6.16)\nRequirement already satisfied, skipping upgrade: chardet&lt;3.1.0,&gt;=3.0.2 in \/Users\/ user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/site-packages (from requests) (3.0.4)\nRequirement already satisfied, skipping upgrade: urllib3!=1.25.0,!=1.25.1,&lt;1.26,&gt;=1.21.1 in \/Users\/ user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/site-packages (from requests) (1.25.3)\nRequirement already satisfied, skipping upgrade: idna&lt;2.9,&gt;=2.5 in \/Users\/ user_name\/.conda\/envs\/ObjDectCount\/lib\/python3.7\/site-packages (from requests) (2.8)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57476895","CreationDate":"2019-08-13T11:33:44.307","ParentId":null,"OwnerUserId":"9291522","Title":null,"Body":"<p>I had a similar error once and fixed it by upgrading the <code>requests<\/code> package:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>pip install --upgrade requests\n<\/code><\/pre>\n\n<p>or in your case:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>conda update requests\n<\/code><\/pre>\n\n<p><strong>Edit:<\/strong> A quick and dirty workaround is to download the weights manually and to store the weights-file under <code>~\/.keras\/models\/<\/code><\/p>\n"}]}
+{"QuestionId":57475760,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-13T10:22:18.313","AcceptedAnswerId":null,"OwnerUserId":7009666.0,"Title":"Difference between eval and vis stages in Deeplab v3+","Body":"<p>Is the <code>eval.py<\/code> essentially just performing the same predictions as <code>vis.py<\/code> would, and then calculating the performance of the predictions as well? I am wondering because I have been playing with some settings in <code>eval<\/code> to optimize the <code>mIoU<\/code>, and I assume that if I give these same parameters to the <code>vis.py<\/code> script then the quality in the segmentation images from <code>vis.py<\/code> will also improve?<\/p>\n","answers":[]}
+{"QuestionId":57475889,"AnswerCount":1,"Tags":"<tensorflow><nlp><pytorch><gensim><word2vec>","CreationDate":"2019-08-13T10:30:07.873","AcceptedAnswerId":57476857.0,"OwnerUserId":11906455.0,"Title":"Could I use BERT to Cluster phrases with pre-trained model","Body":"<p>I found it was a failure that I had used Gensim with GoogleNews pre-trained model to cluster phrases like:<\/p>\n\n<ul>\n<li>knitting<\/li>\n<li>knit loom<\/li>\n<li>loom knitting<\/li>\n<li>weaving loom<\/li>\n<li>rainbow loom<\/li>\n<li>home decoration accessories<\/li>\n<li>loom knit\/knitting loom<\/li>\n<li>...<\/li>\n<\/ul>\n\n<p>I am advised that <a href=\"https:\/\/stackoverflow.com\/questions\/57426745\/how-to-cluster-words-and-phrases-with-pre-trained-model-on-gensim\">GoogleNews model does't have the phrases in it<\/a>. The phrases I have are a little specific to GoogleNews model while I don't have corpus to train a new model. I have only the phrases. And now I am considering to turn to BERT. But could BERT do that as I expected as above? Thank you.<\/p>\n","answers":[{"AnswerId":"57476857","CreationDate":"2019-08-13T11:30:49.807","ParentId":null,"OwnerUserId":"7059474","Title":null,"Body":"<p>You can feed a phrase into the pretrained BERT model and get an embedding, i.e. a fixed-dimension vector. So BERT can embed your phrases in a space. Then you can use a clustering algorithm (such as k-means) to cluster the phrases. The phrases do not need to occur in the training corpus of BERT, as long as the words they consist of are in the vocabulary. You will have to try to see if the embeddings give you relevant results.<\/p>\n"}]}
+{"QuestionId":57476279,"AnswerCount":1,"Tags":"<python><pytorch><spacy><multiclass-classification>","CreationDate":"2019-08-13T10:54:18.287","AcceptedAnswerId":57476733.0,"OwnerUserId":9770403.0,"Title":"Model() got multiple values for argument 'nr_class' - SpaCy multi-classification model (BERT integration)","Body":"<p>Hi I am working on implementing a multi-classification model (5 classes) with the new SpaCy Model <code>en_pytt_bertbaseuncased_lg<\/code>. The code for the new pipe is here:<\/p>\n\n<pre><code>nlp = spacy.load('en_pytt_bertbaseuncased_lg')\ntextcat = nlp.create_pipe(\n    'pytt_textcat',\n    config={\n        \"nr_class\":5,\n        \"exclusive_classes\": True,\n    }\n)\nnlp.add_pipe(textcat, last = True)\n\ntextcat.add_label(\"class1\")\ntextcat.add_label(\"class2\")\ntextcat.add_label(\"class3\")\ntextcat.add_label(\"class4\")\ntextcat.add_label(\"class5\")\n<\/code><\/pre>\n\n<p>The code for the training is as follows and is based on the example from here(<a href=\"https:\/\/pypi.org\/project\/spacy-pytorch-transformers\/\" rel=\"nofollow noreferrer\">https:\/\/pypi.org\/project\/spacy-pytorch-transformers\/<\/a>):<\/p>\n\n<pre><code>def extract_cat(x):\n    for key in x.keys():\n        if x[key]:\n            return key\n\n# get names of other pipes to disable them during training\nn_iter = 250 # number of epochs\n\ntrain_data = list(zip(train_texts, [{\"cats\": cats} for cats in train_cats]))\n\n\ndev_cats_single   = [extract_cat(x) for x in dev_cats]\ntrain_cats_single = [extract_cat(x) for x in train_cats]\ncats = list(set(train_cats_single))\nrecall = {}\nfor c in cats:\n    if c is not None: \n        recall['dev_'+c] = []\n        recall['train_'+c] = []\n\n\n\noptimizer = nlp.resume_training()\nbatch_sizes = compounding(1.0, round(len(train_texts)\/2), 1.001)\n\nfor i in range(n_iter):\n    random.shuffle(train_data)\n    losses = {}\n    batches = minibatch(train_data, size=batch_sizes)\n    for batch in batches:\n        texts, annotations = zip(*batch)\n        nlp.update(texts, annotations, sgd=optimizer, drop=0.2, losses=losses)\n    print(i, losses)\n<\/code><\/pre>\n\n<p>So the structure of my data looks like this:<\/p>\n\n<pre><code>[('TEXT TEXT TEXT',\n  {'cats': {'class1': False,\n    'class2': False,\n    'class3': False,\n    'class4': True,\n    'class5': False}}), ... ]\n<\/code><\/pre>\n\n<p>I am not sure why I get the following error: <\/p>\n\n<pre><code>TypeError                                 Traceback (most recent call last)\n&lt;ipython-input-32-1588a4eadc8d&gt; in &lt;module&gt;\n     21 \n     22 \n---&gt; 23 optimizer = nlp.resume_training()\n     24 batch_sizes = compounding(1.0, round(len(train_texts)\/2), 1.001)\n     25 \n\nTypeError: Model() got multiple values for argument 'nr_class'\n<\/code><\/pre>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>if I take out the nr_class argument, I get this error here:<\/p>\n\n<pre><code>ValueError: operands could not be broadcast together with shapes (1,2) (1,5)\n<\/code><\/pre>\n\n<p>I actually thought this would happen because I didn't specify the nr_class argument. Is that correct? <\/p>\n","answers":[{"AnswerId":"57476733","CreationDate":"2019-08-13T11:22:56.250","ParentId":null,"OwnerUserId":"232119","Title":null,"Body":"<p>This is a regression in the most recent version we released of <code>spacy-pytorch-transformers<\/code>. Sorry about this!<\/p>\n\n<p>The root cause is, this is another case of the evils of <code>**kwargs<\/code>. I'm looking forward to refining the spaCy API to prevent these issues in future.<\/p>\n\n<p>You can see the offending line here: <a href=\"https:\/\/github.com\/explosion\/spacy-pytorch-transformers\/blob\/c1def95e1df783c69bff9bc8b40b5461800e9231\/spacy_pytorch_transformers\/pipeline\/textcat.py#L71\" rel=\"noreferrer\">https:\/\/github.com\/explosion\/spacy-pytorch-transformers\/blob\/c1def95e1df783c69bff9bc8b40b5461800e9231\/spacy_pytorch_transformers\/pipeline\/textcat.py#L71<\/a> . We provide the <code>nr_class<\/code> positional argument, which overlaps with the explicit argument you passed in during the config.<\/p>\n\n<p>In order to workaround the problem, you can simply remove the <code>nr_class<\/code> key from your the <code>config<\/code> dict you're passing into <code>spacy.create_pipe()<\/code>.<\/p>\n"}]}
+{"QuestionId":57476948,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-08-13T11:36:27.180","AcceptedAnswerId":null,"OwnerUserId":9658408.0,"Title":"Applying VAT Semi-supervised using keras and tensorflow","Body":"<p>The following link \n<a href=\"https:\/\/github.com\/fahad7033\/Implementation-of-VAT-on-MNIST-using-NN\/blob\/master\/simpleVatKerasNN_MNIST.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fahad7033\/Implementation-of-VAT-on-MNIST-using-NN\/blob\/master\/simpleVatKerasNN_MNIST.py<\/a><\/p>\n\n<p>contains a simple implementation of VAT using keras and tensorflow with MNIST. I would like to upgrade the code for applying semi-supervised learning.\nAccording to the original paper of VAT, it is required to make the classifier predicts the logit for unlabeled data and then repeat the same process of VAT's procedure.\nMy question is in which stage during the code should the logit of unlabeled data be predicted?<\/p>\n","answers":[]}
+{"QuestionId":57477763,"AnswerCount":0,"Tags":"<multithreading><keras>","CreationDate":"2019-08-13T12:29:24.737","AcceptedAnswerId":null,"OwnerUserId":4010607.0,"Title":"Keras multithreading issue","Body":"<p>I am trying to load the models in one thread and use them to predict in another thread. The loading and prediction work if the thread code and the class code are in the same file. But if I move the class code to a separate file I get \"RuntimeError: Graph is finalized and cannot be modified.\" error. I want to keep the class code in a separate file and use it by importing.<\/p>\n\n<p>Below code is saved as test.py<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy\n\nfrom keras import backend as K\nfrom keras.layers import Dense\nfrom keras.models import Sequential\nfrom keras.models import load_model\n\nclass Test(object):\n    def __init__(self, Model):\n        # Create a new dummy model\n        self.Create()\n\n        # Load the dummy from disk\n        self.__Model = load_model(\"Testmodel\")\n\n        # Perform a dummy prediction to initialize the model\n        self.__Model.predict(numpy.array([[0, 0]]))\n\n        # Clean up tensorflow graph\n        self.__Session = K.get_session()\n        self.__Graph = tf.get_default_graph()\n        self.__Graph.finalize()\n\n    def Predict(self, Data):\n        with self.__Session.as_default():\n            with self.__Graph.as_default():\n                Prediction = self.__Model.predict(Data)\n\n        print(\"[INFO] Prediction: {}\".format(Prediction[0].argmax()))\n\n    def Create(self):\n        model = Sequential()\n        model.add(Dense(256, input_shape = (2, )))\n        model.add(Dense(1, activation = \"softmax\"))\n        model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])\n        model.save(\"Testmodel\")\n<\/code><\/pre>\n\n<hr>\n\n<p>In a separate file say predict.py<\/p>\n\n<pre><code>import threading\nfrom test import Test\nimport numpy\n\n\n\ndef main():\n    Model = Test(\"dummymodel\")\n\n    # Create 10 new threads\n    Threads = list()\n    for Thread in range(1):\n        Threads.append(threading.Thread(target = Model.Predict, kwargs = { \"Data\": numpy.random.random((500, 2)) }))\n        Threads[Thread].start()\n\n    # Wait for threads to complete task\n    for Thread in Threads:\n        Thread.join()\n\nif(__name__ == \"__main__\"):\n    main()\n\n    print(\"[INFO] Finish!\")\n<\/code><\/pre>\n\n<p><em>This test code is referenced from <a href=\"https:\/\/gitlab.com\/snippets\/1728728\" rel=\"nofollow noreferrer\">https:\/\/gitlab.com\/snippets\/1728728<\/a><\/em><\/p>\n\n<p>When I try to call the predict method in a thread it returns below error<\/p>\n\n<pre><code>File \"&lt;ipython-input-4-d9e081259cda&gt;\", line 22, in &lt;module&gt;\nmain()\n\nFile \"&lt;ipython-input-4-d9e081259cda&gt;\", line 8, in main\nModel = Test(\"dummymodel\")\n\nFile \"..\\MyNetwork.py\", line 12, in __init__\nself.Create()\n\nFile \"..\\MyNetwork.py\", line 34, in Create\nmodel.add(Dense(256, input_shape = (2, )))\n\nFile \"C:\\Anaconda3\\envs\\KerasGPU\\lib\\site-packages\\keras\\engine\\sequential.py\", line 161, in add\nname=layer.name + '_input')\n\nFile \"C:\\Anaconda3\\envs\\KerasGPU\\lib\\site-packages\\keras\\engine\\input_layer.py\", line 178, in Input\ninput_tensor=tensor)\n\nFile \"C:\\Anaconda3\\envs\\KerasGPU\\lib\\site-packages\\keras\\legacy\\interfaces.py\", line 91, in wrapper\nreturn func(*args, **kwargs)\n\nFile \"C:\\Anaconda3\\envs\\KerasGPU\\lib\\site-packages\\keras\\engine\\input_layer.py\", line 87, in __init__\nname=self.name)\n\nFile \"C:\\Anaconda3\\envs\\KerasGPU\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py\", line 517, in placeholder\nx = tf.placeholder(dtype, shape=shape, name=name)\n\nFile \"C:\\Anaconda3\\envs\\KerasGPU\\lib\\site-packages\\tensorflow\\python\\ops\\array_ops.py\", line 2077, in placeholder\nreturn gen_array_ops.placeholder(dtype=dtype, shape=shape, name=name)\n\nFile \"C:\\Anaconda3\\envs\\KerasGPU\\lib\\site-packages\\tensorflow\\python\\ops\\gen_array_ops.py\", line 5791, in placeholder\n\"Placeholder\", dtype=dtype, shape=shape, name=name)\n\nFile \"C:\\Anaconda3\\envs\\KerasGPU\\lib\\site-packages\\tensorflow\\python\\framework\\op_def_library.py\", line 788, in _apply_op_helper\nop_def=op_def)\n\nFile \"C:\\Anaconda3\\envs\\KerasGPU\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py\", line 507, in new_func\nreturn func(*args, **kwargs)\n\nFile \"C:\\Anaconda3\\envs\\KerasGPU\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3272, in create_op\nself._check_not_finalized()\n\nFile \"C:\\Anaconda3\\envs\\KerasGPU\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 2945, in _check_not_finalized\nraise RuntimeError(\"Graph is finalized and cannot be modified.\")\n\nRuntimeError: Graph is finalized and cannot be modified.\n<\/code><\/pre>\n\n<p>I have observed that if I add the class code to predict.py the code works fine and I am able to get predictions without any error.\nKeras Version  2.2.4\nTensorflow Version  1.13.1<\/p>\n\n<p><em>The code shown above is  just as example to point out the issue I am facing.<\/em><\/p>\n","answers":[]}
+{"QuestionId":57477872,"AnswerCount":1,"Tags":"<keras><loss-function>","CreationDate":"2019-08-13T12:36:49.227","AcceptedAnswerId":57478331.0,"OwnerUserId":11863622.0,"Title":"Keras loss functions: how to round?","Body":"<p>I'm trying to recognize turning points in sequences, the points after which some process behaves differently. I use a keras model to do this. Input is the sequence (always the same length) and output should be 0 before the turning points, a 1 after the turning point.<\/p>\n\n<p>I want the loss function to depend on the distance between the actual turning point and the predicted turning point. <\/p>\n\n<p>I tried to round (to obtain the label 0 or 1), followed by summing the total number of 1's to get the \"index\" of the turning point. Assumed here is that the model gives just one turning point, as the data (synthetically produced) also has just one turning point. Tried is:<\/p>\n\n<pre><code>def dist_loss(yTrue,yPred):\n    turningPointTrue = K.sum(yTrue) \n    turningPointPred = K.sum(K.round(yPred))\n    return K.abs(turningPointTrue-turningPointPred)\n<\/code><\/pre>\n\n<p>This does not work, the following error is given:<\/p>\n\n<blockquote>\n  <p>ValueError: An operation has <code>None<\/code> for gradient. Please make sure\n  that all of your ops have a gradient defined (i.e. are\n  differentiable). Common ops without gradient: K.argmax, K.round,\n  K.eval.<\/p>\n<\/blockquote>\n\n<p>I think this means that K.round(yPred) gives a singular value, instead of a vector\/tensor. Does anyone know how to solve this issue?<\/p>\n","answers":[{"AnswerId":"57478331","CreationDate":"2019-08-13T13:03:55.740","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>The <code>round<\/code> operation has no defined gradient, so it cannot be used at all inside a loss function, since for training of a neural network the gradient of the loss with respect to the weights has to be computed, and this implies that all the parts of the network and loss must be differentiable (or a differentiable approximation must be available).<\/p>\n\n<p>In your case you should try to find an approximation to round that is differentiable, but unfortunately I don't know if there is one. One example of such approximation is the softmax function as approximation of the max function.<\/p>\n"}]}
+{"QuestionId":57478698,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-13T13:23:40.127","AcceptedAnswerId":57479139.0,"OwnerUserId":290443.0,"Title":"How can I make TensorFlow 2.0 handle piecewise gradients (e.g. across `tf.gather`)?","Body":"<p>I'm attempting to leverage TensorFlow 2.0's automatic differentiation to automate the calculation of certain gradients on financial instruments. Generally this involves a piecewise interpolation scheme between various \"benchmark points\". The simplest example is below:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    import tensorflow as tf\n\n    MATURITIES = tf.constant([1.0, 2.0, 3.0, 5.0, 7.0, 10.0, 12.0, 15.0, 20.0, 25.0])\n    CASH_FLOW_TIMES = tf.constant([n * 0.5 for n in range(1, 51)])\n    YIELDS = tf.Variable([0.04153733, 0.0425888, 0.04662959, 0.05406879, 0.05728735, 0.0606996, 0.06182699, 0.05854381, 0.05376556, 0.0531946])\n\n    @tf.function\n    def linear(knot_y, knot_x, x):\n        \"\"\"Linear interpolation\"\"\"\n        i = tf.maximum(tf.minimum(tf.searchsorted(knot_x, x, side=\"right\") - 1, knot_x.shape[0] - 2), 0)\n        y_i = tf.gather(knot_y, i)\n        x_i = tf.gather(knot_x, i)\n        return y_i + y_i \/ x_i * (x - x_i)\n\n    with tf.GradientTape() as tape:\n        tape.watch(YIELDS)\n        y = linear(YIELDS, MATURITIES, CASH_FLOW_TIMES)\n    dydx = tape.gradient(y, YIELDS)\n    y, dydx\n<\/code><\/pre>\n\n<p>Which outputs the following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    (&lt;tf.Tensor: id=1249, shape=(50,), dtype=float32, numpy=\n     array([0.02076866, 0.04153733, 0.06230599, 0.0425888 , 0.053236  ,\n            0.04662959, 0.05440119, 0.06217279, 0.06994438, 0.05406879,\n            0.05947567, 0.06488255, 0.07028943, 0.05728735, 0.0613793 ,\n            0.06547125, 0.06956321, 0.07365517, 0.07774712, 0.0606996 ,\n            0.06373458, 0.06676956, 0.06980454, 0.06182699, 0.06440312,\n            0.06697924, 0.06955536, 0.07213148, 0.07470761, 0.05854381,\n            0.06049527, 0.06244673, 0.06439819, 0.06634965, 0.06830111,\n            0.07025257, 0.07220403, 0.07415549, 0.07610695, 0.05376556,\n            0.0551097 , 0.05645384, 0.05779798, 0.05914212, 0.06048626,\n            0.06183039, 0.06317453, 0.06451868, 0.06586281, 0.06720695],\n           dtype=float32)&gt;,\n     &lt;tensorflow.python.framework.indexed_slices.IndexedSlices at 0x203027345c0&gt;)\n<\/code><\/pre>\n\n<p>The issue is that, due (I suspect) to the <code>tf.searchsorted<\/code> or <code>tf.gather<\/code> calls, the gradient is an <code>IndexedSlice<\/code>, not a Tensor. This causes an issue when I need to chain multiple <code>tf.function<\/code> together to value a security. For example, say I want to use a transformation of the <code>YIELDS<\/code> variable within my <code>linear<\/code> function:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    import tensorflow as tf\n\n    MATURITIES = tf.constant([1.0, 2.0, 3.0, 5.0, 7.0, 10.0, 12.0, 15.0, 20.0, 25.0])\n    CASH_FLOW_TIMES = tf.constant([n * 0.5 for n in range(1, 51)])\n    YIELDS = tf.Variable([0.04153733, 0.0425888, 0.04662959, 0.05406879, 0.05728735, 0.0606996, 0.06182699, 0.05854381, 0.05376556, 0.0531946])\n\n    @tf.function\n    def logdf_from_yields(yields, times):\n        return tf.math.multiply(yields, times) * -1.0\n\n    @tf.function\n    def linear(knot_y, knot_x, x):\n        \"\"\"Linear interpolation\"\"\"\n        i = tf.maximum(tf.minimum(tf.searchsorted(knot_x, x, side=\"right\") - 1, knot_x.shape[0] - 2), 0)\n        y_i = tf.gather(knot_y, i)\n        x_i = tf.gather(knot_x, i)\n        return y_i + y_i \/ x_i * (x - x_i)\n\n    with tf.GradientTape() as tape:\n        tape.watch(YIELDS)\n        y = linear(logdf_from_yields(YIELDS, MATURITIES), MATURITIES, CASH_FLOW_TIMES)\n    dydx = tape.gradient(y, YIELDS)\n    y, dydx\n<\/code><\/pre>\n\n<p>This code raises the following exception:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    ---------------------------------------------------------------------------\n    AssertionError                            Traceback (most recent call last)\n    &lt;ipython-input-9-2bc3457894ea&gt; in &lt;module&gt;\n         20     tape.watch(YIELDS)\n         21     y = linear(logdf_from_yields(YIELDS, MATURITIES), MATURITIES, CASH_FLOW_TIMES)\n    ---&gt; 22 dydx = tape.gradient(y, YIELDS)\n         23 y, dydx\n\n    ~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\eager\\backprop.py in gradient(self, target, sources, output_gradients, unconnected_gradients)\n       1000         output_gradients=output_gradients,\n       1001         sources_raw=flat_sources_raw,\n    -&gt; 1002         unconnected_gradients=unconnected_gradients)\n       1003 \n       1004     if not self._persistent:\n\n    ~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\eager\\imperative_grad.py in imperative_grad(tape, target, sources, output_gradients, sources_raw, unconnected_gradients)\n         74       output_gradients,\n         75       sources_raw,\n    ---&gt; 76       compat.as_str(unconnected_gradients.value))\n\n    ~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\eager\\function.py in backward_function(*args)\n        904               if a is not None and i not in skip_positions]\n        905       return self._backward_graph_function._call_flat(  # pylint: disable=protected-access\n    --&gt; 906           list(args) + side_outputs)\n        907 \n        908     tape.record_operation(self._forward_function.signature.name, real_outputs,\n\n    ~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\eager\\function.py in _call_flat(self, args)\n        610     if any(isinstance(a, composite_tensor.CompositeTensor) for a in args):\n        611       raise AssertionError(\"Expected all args to be Tensors or Variables; \"\n    --&gt; 612                            \"but got CompositeTensor: %r\" % args)\n        613 \n        614     if (tape.could_possibly_record() or\n\n    AssertionError: Expected all args to be Tensors or Variables; but got CompositeTensor: [&lt;tensorflow.python.framework.indexed_slices.IndexedSlices object at 0x00000203013C2128&gt;, &lt;tf.Tensor: id=1024, shape=(), dtype=float32, numpy=-1.0&gt;, &lt;tf.Tensor: id=1025, shape=(10,), dtype=float32, numpy=\n    array([0.04153733, 0.0851776 , 0.13988876, 0.27034396, 0.40101147,\n           0.606996  , 0.74192387, 0.87815714, 1.0753112 , 1.329865  ],\n          dtype=float32)&gt;, &lt;tf.Tensor: id=1026, shape=(10,), dtype=float32, numpy=array([ 1.,  2.,  3.,  5.,  7., 10., 12., 15., 20., 25.], dtype=float32)&gt;, &lt;tf.Tensor: id=1027, shape=(10,), dtype=float32, numpy=\n    array([0.04153733, 0.0425888 , 0.04662959, 0.05406879, 0.05728735,\n           0.0606996 , 0.06182699, 0.05854381, 0.05376556, 0.0531946 ],\n          dtype=float32)&gt;]\n<\/code><\/pre>\n\n<p>Inspecting the traceback, I see that the only item that is not a Tensor or Variable is the IndexedSlice, which I believe, again, comes from the <code>linear<\/code> function.<\/p>\n\n<p>Is there a way for me to rewrite the piecewise <code>linear<\/code> function (e.g. using different functions) so that it will work with the gradient tape?<\/p>\n\n<p><strong>What have I tried?<\/strong><\/p>\n\n<p>I took a look at <a href=\"https:\/\/stackoverflow.com\/questions\/38548871\/tensorflow-how-to-propagate-gradient-through-tf-gather\">this question<\/a>, but it seemed like the writer was trying to make the index a variable. Here I should know all of my indices ahead of time (i.e. because <code>MATURITIES<\/code> and <code>CASH_FLOW_TIMES<\/code> are fixed for any given instance of the problem ... only <code>YIELDS<\/code> is a <code>tf.Variable<\/code>). So I wasn't so sure how it applied.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/50301565\/function-gradients-with-gather-operations-in-tensorflow\">This question<\/a> was also interesting in that it recommended the usage of <code>tf.convert_to_tensor<\/code>, but I'm not sure how to apply it in my case.<\/p>\n","answers":[{"AnswerId":"57479139","CreationDate":"2019-08-13T13:45:57.703","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Your gradient seems to be fine. TensorFlow uses <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/IndexedSlices\" rel=\"nofollow noreferrer\"><code>tf.IndexedSlices<\/code><\/a> to represent sparse gradients in some cases like <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather\" rel=\"nofollow noreferrer\"><code>tf.gather<\/code><\/a>, but you can easily convert it to a regular tensor like this (the example is in graph mode but the function would be the same in eager mode):<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef convert_indexed_slices_to_tensor(idx_slices):\n    return tf.scatter_nd(tf.expand_dims(idx_slices.indices, 1),\n                         idx_slices.values, idx_slices.dense_shape)\n\n# Test\nwith tf.Graph().as_default(), tf.Session() as sess:\n    a = tf.constant([1., 2., 3., 4.])\n    b = tf.gather(a, [0, 2])\n    g = tf.gradients(b, a)[0]\n    print(g)\n    # IndexedSlices(indices=..., values=..., dense_shape=...)\n    g_dense = convert_indexed_slices_to_tensor(g)\n    # Tensor(...)\n    print(sess.run(g_dense))\n    # [1. 0. 1. 0.]\n<\/code><\/pre>\n\n<hr>\n\n<p>If you want to force <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather\" rel=\"nofollow noreferrer\"><code>tf.gather<\/code><\/a> to produce regular tensors, you can wrap it with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/custom_gradient\" rel=\"nofollow noreferrer\"><code>tf.custom_gradient<\/code><\/a> like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\n@tf.custom_gradient\ndef gather_dense_gradient(params, indices, validate_indices=None, name=None):\n    def grad(ys):\n        return tf.scatter_nd(tf.expand_dims(indices, 1), ys, tf.shape(params)), None\n    return tf.gather(params, indices, validate_indices, name), grad\n\n# Test\nwith tf.Graph().as_default(), tf.Session() as sess:\n    a = tf.constant([1., 2., 3., 4.])\n    b = gather_dense_gradient(a, [0, 2])\n    g = tf.gradients(b, a)[0]\n    print(sess.run(g))\n    # [1. 0. 1. 0.]\n<\/code><\/pre>\n\n<p>Note this assumes <code>axis=0<\/code> and one-dimensional <code>indices<\/code>, otherwise it would still be possible to do the same but it would require a bit more of work.<\/p>\n"}]}
+{"QuestionId":57478711,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-13T13:24:12.490","AcceptedAnswerId":null,"OwnerUserId":10811086.0,"Title":"How to estimate poses by tensorflow posenet","Body":"<p>I've downloaded the tflite model from <a href=\"https:\/\/www.tensorflow.org\/lite\/models\/pose_estimation\/overview\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/lite\/models\/pose_estimation\/overview<\/a> and I want to use it to estimate poses in images on my laptop with python, how to load the model and use it to do this?<\/p>\n","answers":[]}
+{"QuestionId":57479509,"AnswerCount":1,"Tags":"<tensorflow><keras><conv-neural-network>","CreationDate":"2019-08-13T14:06:35.510","AcceptedAnswerId":null,"OwnerUserId":11533160.0,"Title":"How to solve Keras conv2d input shape error?","Body":"<p>I am having problems with the input shape while trying to create a convolutional neural network.<\/p>\n\n<p>My dataset has 164 images 250x250 pixels (123 imgs for training and the remaining for test)<\/p>\n\n<p>The code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def conv_neural_net(x_train, x_test, y_train, y_test):\n\n    model = Sequential()\n    model.add(Conv2D(16, (3,3), padding=\"same\", activation='relu', input_shape=(250,250,1)))\n    model.add(MaxPooling2D(pool_size=(2, 2), strides=(1,1)))\n\n    model.add(Conv2D(32,(3,3), padding=\"same\", activation='relu'))\n    model.add(MaxPooling2D(pool_size=(2, 2), strides=(1,1)))\n\n    model.add(Conv2D(64,(3,3), padding=\"same\", activation='relu'))\n    model.add(MaxPooling2D(pool_size=(2, 2), strides=(1,1)))\n\n    model.add(Flatten())\n    model.add(Dense(units=128, activation='relu'))\n    model.add(Dropout(rate=0.2))\n    model.add(Dense(units=1, activation='sigmoid'))\n\n    model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])\n    model.fit(x_train, y_train, epochs=10)\n\n    _, accuracy = model.evaluate(x_test, y_test)\n    print('Accuracy: %.2f' % (accuracy*100))\n\n    return model\n<\/code><\/pre>\n\n<p>I have an error while trying to fit<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-235-d83533a1c9fa&gt; in &lt;module&gt;\n----&gt; 1 conv_neural_net(x_train, x_test, y_train, y_test)\n\n&lt;ipython-input-234-d7ddc86e3578&gt; in conv_neural_net(x_train, x_test, y_train, y_test)\n     27 \n     28     model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])\n---&gt; 29     model.fit(x_train, y_train, epochs=10)\n     30 \n     31     _, accuracy = model.evaluate(x_test, y_test)\n\n~\\AppData\\Local\\Continuum\\anaconda3\\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\n~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\keras\\engine\\training.py in _standardize_user_data(self, x, y, sample_weight, class_weight, check_array_lengths, batch_size)\n    749             feed_input_shapes,\n    750             check_batch_axis=False,  # Don't enforce the batch size.\n--&gt; 751             exception_prefix='input')\n    752 \n    753         if y is not None:\n\n~\\AppData\\Local\\Continuum\\anaconda3\\lib\\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 input: expected conv2d_154_input to have 4 dimensions, but got array with shape (123, 250, 250)\n<\/code><\/pre>\n\n<p>Any idea what's wrong?<\/p>\n","answers":[{"AnswerId":"57479710","CreationDate":"2019-08-13T14:17:47.010","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>The required shape involves \"channels\". (RGB has 3 channels, RGBA has 4 channels, black and white has 1 channel)<\/p>\n\n<p>All you need is to add the channels dimension to your input data. <\/p>\n\n<pre><code>x_train = x_train.reshape((123,250,250,1))    \nx_test = x_test.reshape((164-123,250,250,1))    \n<\/code><\/pre>\n"}]}
+{"QuestionId":57479680,"AnswerCount":1,"Tags":"<tensorflow><keras><keras-layer>","CreationDate":"2019-08-13T14:15:55.613","AcceptedAnswerId":null,"OwnerUserId":8497966.0,"Title":"How to access recursive layers of custom layer in tensorflow keras","Body":"<p>From tensorflow keras example. I can create a custom layer which contains Linear layer recursively<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class MLPBlock(layers.Layer):\n\n  def __init__(self):\n    super(MLPBlock, self).__init__()\n    self.linear_1 = Linear(32)\n    self.linear_2 = Linear(32)\n    self.linear_3 = Linear(1)\n\n  def call(self, inputs):\n    x = self.linear_1(inputs)\n    x = tf.nn.relu(x)\n    x = self.linear_2(x)\n    x = tf.nn.relu(x)\n    return self.linear_3(x)\n<\/code><\/pre>\n\n<p>how do i access all the component layers of a custom layer, I want to access weight and biases of all the component layers.<\/p>\n\n<p>ex:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>MLPBlock(Parent Layer):\n    linear_1\n    linear_2\n    linear_3\n<\/code><\/pre>\n\n<p>I have looked into tensorflow keras api version r 1.14  <a href=\"https:\/\/www.tensorflow.org\/guide\/keras\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/keras<\/a>\nbut could not find any way to do this.<\/p>\n","answers":[{"AnswerId":"57479922","CreationDate":"2019-08-13T14:29:17.703","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>I assume that you are following <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/keras\/custom_layers_and_models\" rel=\"nofollow noreferrer\">this tutorial<\/a>. Based on that, here is how you can access the weights:<\/p>\n\n<pre><code>class MLPBlock(tf.keras.Model):\n\n    def __init__(self):\n        super(MLPBlock, self).__init__()\n        self.linear_1 = tf.keras.layers.Dense(32)\n        self.linear_2 = tf.keras.layers.Dense(32)\n        self.linear_3 = tf.keras.layers.Dense(1)\n\n    def call(self, inputs):\n        x = self.linear_1(inputs)\n        x = tf.nn.relu(x)\n        x = self.linear_2(x)\n        x = tf.nn.relu(x)\n        return self.linear_3(x)\n\nmlp_block = MLPBlock()\ny = mlp_block(tf.ones(shape=(3, 64)))\nfor layer in mlp_block.layers:\n    weights, biases = layer.get_weights()\n<\/code><\/pre>\n\n<p>Please note that I slightly modified the example so that you can access the layer's weights and biases. Namely, what I did is instead of subclassing the class with <code>tf.keras.layers.Layer<\/code>, I subclassed with <code>tf.keras.Model<\/code> so that the stack of layers can be treated as a model, and then you can access the layers of that model. Then, instead of using the custom <code>Linear<\/code> layer, I used the <code>tf.keras.layers.Dense<\/code> for simplicity, however, using the custom layer should not make a difference.<\/p>\n"}]}
+{"QuestionId":57480180,"AnswerCount":0,"Tags":"<python><keras><deep-learning><tf.keras>","CreationDate":"2019-08-13T14:44:05.137","AcceptedAnswerId":null,"OwnerUserId":11652193.0,"Title":"model.fit() printing accuracy and loss afer evey batch_size","Body":"<p>I have model.fit() with verbose=1<\/p>\n\n<p>but my output is printing for every batch size::<\/p>\n\n<p>1920\/323432 [..............................] - ETA: 19:21 - loss: 10.4622 - acc: 0.343 - ETA: 18:40 - loss: 10.5245 - acc: 0.339 - ETA: 18:32 - loss: 10.5452 - acc: 0.338 - ETA: 18:29 - loss: 10.5556 - acc: 0.337 - ETA: 18:30 - loss: 10.2380 - acc: 0.357 - ETA: 18:31 - loss: 10.3999 - acc: 0.347 - ETA: 18:37 - loss: 10.4978 - acc: 0.341 - ETA: 18:40 - loss: 10.5089 - acc: 0.340 - ETA: 18:39 - loss: 10.3376 - acc: 0.351 - ETA: 18:52 - loss: 10.2878 - acc: 0.354 - ETA: 18:55 - loss: 10.3490 - acc: 0.350 - ETA: 18:55 - loss: 10.2650 - acc: 0.356 - ETA: 18:54 - loss: 10.2897 - acc: 0.354 - ETA: 18:55 - loss: 10.1864 - acc: 0.361 - ETA: 18:55 - loss: 10.1799 - acc: 0.3615<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nmodel_trained = final_model.fit([LSTM_train_X['left'], LSTM_train_X['right']], LSTM_train_y, batch_size=batch_size, epochs=epochs,\n                            validation_data=([LSTM_valid_X['left'], LSTM_valid_X['right']], LSTM_valid_y), verbose=1)\n\n<\/code><\/pre>\n\n<p>I have merged two models using keras.add()<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nmerged_output = add([branch1.output,branch2.output]) \n\n<\/code><\/pre>\n\n<p>total code i have wrote::<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nbranch1 = Sequential()\n\nbranch1.add(Embedding(len(embeddings), embedding_dim, weights=[embeddings], input_length=max_seq_length, trainable=False))\nbranch1.add(LSTM(hidden_layer_nodes))\n\n\nbranch2 = Sequential()\nbranch2.add(Embedding(len(embeddings), embedding_dim, weights=[embeddings], input_length=max_seq_length, trainable=False))\n\nbranch2.add(LSTM(hidden_layer_nodes))\n\n\nmerged_output = add([branch1.output,branch2.output]) \n\n\nmodel_combined = Sequential()\n\nmodel_combined.add(Activation('relu'))\n\nmodel_combined.add(Dense(256))\n\nmodel_combined.add(Activation('relu'))\n\nmodel_combined.add(Dense(1))\n\nmodel_combined.add(Activation('softmax'))\n\nfinal_model = Model([branch1.input, branch2.input], model_combined(merged_output))\n\n\nfinal_model.compile(optimizer=optimizer,loss='binary_crossentropy' , metrics=['accuracy'])\n\nmodel_trained = final_model.fit([LSTM_train_X['left'], LSTM_train_X['right']], LSTM_train_y, batch_size=batch_size, epochs=epochs,\n                            validation_data=([LSTM_valid_X['left'], LSTM_valid_X['right']], LSTM_valid_y), verbose=1)\n\n<\/code><\/pre>\n\n<p>how to clear that issue?<\/p>\n\n<p>Thanks:)<\/p>\n","answers":[]}
+{"QuestionId":57480537,"AnswerCount":0,"Tags":"<python><android><tensorflow><firebase-mlkit>","CreationDate":"2019-08-13T15:02:28.870","AcceptedAnswerId":null,"OwnerUserId":4166721.0,"Title":"How to save training image for TensorFlow Lite training from Android?","Body":"<p>I am trying to build a very simple image classification model for a TicTacToe Android game, to identify \"X\", \"O\", or an empty image from the device camera.<\/p>\n\n<p>What I have done so far is creating a CameraX ImageAnalysis use-case, crop out the image, and creating a FirebaseVisionImage for ML kit. Which works fine if I pass it to the text recognition API, but it does not work for hand drawn characters.<\/p>\n\n<p>The next step for me is to build and train a custom TensorFlow Lite model, but because I cannot train on device I need to export these image somehow.<\/p>\n\n<p>My question is how can I save this ByteArray\/FirebaseVisionImage into a file format that I can use for training, and how do I make sure the data format would be exactly the same?<\/p>\n\n<p>For example, if I store these image as jpg and use them for training, I don't know if data format will be still the same after I compress it and read it back, compared to what I get directly from the camera.<\/p>\n","answers":[]}
+{"QuestionId":57480706,"AnswerCount":1,"Tags":"<opencv><tensorflow><object-detection-api><video-tracking>","CreationDate":"2019-08-13T15:11:36.083","AcceptedAnswerId":null,"OwnerUserId":8836422.0,"Title":"How to use detected objectes from tensorflow object detection api using mobilenet for object tracking?","Body":"<p>I have trained the object detection API for custom object detection. Now I want to track the objects detected by the API.<\/p>\n","answers":[{"AnswerId":"57485909","CreationDate":"2019-08-13T22:11:02.903","ParentId":null,"OwnerUserId":"820795","Title":null,"Body":"<p>Take a look at SORT algorithm. It is contains:<\/p>\n\n<ul>\n<li>Kalman filter for prediction of new position<\/li>\n<li>Hungurian algorithm for track and detections matching with IOU distance (or you can write your own criteria)<\/li>\n<\/ul>\n\n<p>Above it you can attack any logic that you want. This links might be interesting:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/nwojke\/deep_sort\" rel=\"nofollow noreferrer\">Deep SORT<\/a> - for each detection it build descriptor for image area. This descriptor compares with track descriptors<\/li>\n<li><a href=\"https:\/\/github.com\/knorth55\/chainer-sort\" rel=\"nofollow noreferrer\">SORT<\/a> - simple SORT with FasterRCNN and SSD detection<\/li>\n<\/ul>\n\n<p>In our projects we adapt SORT to use template matching between detections. If performance crucial for you, this is also might be good idea to think about.<\/p>\n"}]}
+{"QuestionId":57480903,"AnswerCount":1,"Tags":"<python><tensorflow><matrix>","CreationDate":"2019-08-13T15:22:44.877","AcceptedAnswerId":57481624.0,"OwnerUserId":6083655.0,"Title":"Highlight specific coordinates on multiple zero matrices using only vectorized operations","Body":"<p>Let's say there are <code>2<\/code> <code>5x5<\/code> sparse matrices stacked on each other (or <code>tf.zeros(2, 5, 5)<\/code>). <\/p>\n\n<p>Then suppose there also is an array of coordinates <code>[[0, 2, 4, 4], [2, 0, 3, 3]]<\/code>, which respectively shows starting width position, starting height position, ending width position, and ending height position of coordinates (i.e <code>[starting_w, starting_h, ending_w, ending_h]<\/code> for each coordinate).<\/p>\n\n<p>I want to \"highlight\" the <code>(2, 5, 5)<\/code> sparse shape accordng to the coordinates mentioned above, so that on each sparse matrix of shape, each coordinate is respectively \"projected\" and there are <code>1<\/code>s on those coordinates and <code>0<\/code>s anywhere else.<\/p>\n\n<h1>Example<\/h1>\n\n<p>So as mentioned earlier, we have a <code>(2, 5, 5)<\/code> sparse shape:<\/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. 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>Now we \"project\" each coordinate out of <code>[[0, 2, 4, 4], [2, 0, 3, 3]]<\/code> on each sparse matrix:<\/p>\n\n<pre><code>[[[1. 1. 1. 1. 1.]  # [0, 2, 4, 4]\n  [0. 0. 0. 0. 0.]\n  [1. 0. 0. 0. 0.]\n  [1. 0. 0. 0. 0.]\n  [1. 0. 0. 0. 0.]]\n\n [[1. 0. 1. 1. 0.]  # [2, 0, 3, 3]\n  [1. 0. 0. 0. 0.]\n  [1. 0. 0. 0. 0.]\n  [1. 0. 0. 0. 0.]\n  [0. 0. 0. 0. 0.]]]\n<\/code><\/pre>\n\n<p><strong>Note:<\/strong><\/p>\n\n<p>The example seen above uses <code>n = 2<\/code> coordinates, and thus generates <code>2<\/code> matrices, but in general <code>n<\/code> must be treated as a symbolic tensor which can have any value.<\/p>\n\n<h1>Question<\/h1>\n\n<p>Can something like this be done on pure vectorized operations of highly declarative Tensorflow? (without any use of operations like <code>tf.map_fn<\/code>, <code>tf.while<\/code> and etc.).<\/p>\n\n<p><strong>What I tried:<\/strong><\/p>\n\n<p>I've initially tried <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather_nd\" rel=\"nofollow noreferrer\">tf.gather_nd<\/a>, but it doesn't explicitly support slices (although there are <a href=\"https:\/\/stackoverflow.com\/a\/53686871\/6083655\">some \"hacks\"<\/a> which fail to work due to absence of symbolic support). I know <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/slice\" rel=\"nofollow noreferrer\">tf.slice<\/a> is meant to precisely do that, but it does not have explicit support of example demonstrated above.<\/p>\n\n<p>I also thought of using <code>tf.where<\/code>, which might easily work on single array with single coordinate - but I don't know if it supports multiple ones as demonstrated above.<\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"57481624","CreationDate":"2019-08-13T16:11:20.930","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You can make that result like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef make_highlights(idx, width, height, dtype=tf.bool):\n    n = tf.shape(idx)[0]\n    # Add two dimensions for broadcasting later\n    idx = idx[:, :, np.newaxis, np.newaxis]\n    # Extract coordinates\n    start_w, start_h, end_w, end_h = idx[:, 0], idx[:, 1], idx[:, 2], idx[:, 3]\n    # Make index arrays\n    xx = tf.range(width)\n    yy = tf.expand_dims(tf.range(height), 1)\n    # Make first row highlighting\n    h_x = tf.equal(yy, 0) &amp; (xx &gt; start_w) &amp; (xx &lt;= end_w)\n    # Make first column highlighting\n    h_y = tf.equal(xx, 0) &amp; (yy &gt;= start_h) &amp; (yy &lt;= end_h)\n    # Make result\n    return tf.cast(h_x | h_y, dtype)\n\nwith tf.Graph().as_default(), tf.Session() as sess:\n    idx = tf.placeholder(tf.int32, [None, 4])\n    width = tf.placeholder(tf.int32, [])\n    height = tf.placeholder(tf.int32, [])\n    result = make_highlights(idx, width, height, tf.float32)\n    out = sess.run(result, feed_dict={idx: [[0, 2, 4, 4], [2, 0, 3, 3]],\n                                      width: 5, height: 5})\n    print(out)\n    # [[[1. 1. 1. 1. 1.]\n    #   [0. 0. 0. 0. 0.]\n    #   [1. 0. 0. 0. 0.]\n    #   [1. 0. 0. 0. 0.]\n    #   [1. 0. 0. 0. 0.]]\n    # \n    #  [[1. 0. 1. 1. 0.]\n    #   [1. 0. 0. 0. 0.]\n    #   [1. 0. 0. 0. 0.]\n    #   [1. 0. 0. 0. 0.]\n    #   [0. 0. 0. 0. 0.]]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57481282,"AnswerCount":1,"Tags":"<tensorflow><dcgan><multiple-gpu>","CreationDate":"2019-08-13T15:47:32.613","AcceptedAnswerId":null,"OwnerUserId":10763156.0,"Title":"How to use multiple GPU for a DCGAN using Tensorflow 2.0 - RuntimeError: Replica-local variables may only be assigned in a replica context","Body":"<p>I would like to develop a DCGAN with resolution of 256x256. To do so I need to use multiple GPU since only one it is not enough and it will probably take too much time.<\/p>\n\n<p>I followed the procedure explained in the documentation at this link\n<a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/distribute_strategy\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/guide\/distribute_strategy<\/a><\/p>\n\n<p>At the top of the script I used<\/p>\n\n<p><code>strategy = tf.distribute.MirroredStrategy()<\/code><\/p>\n\n<p>Then inside the Generator, Discriminator, and Loss functions I used<\/p>\n\n<p><code>with strategy.scope():<\/code><\/p>\n\n<p>The error I get is:<\/p>\n\n<p><code>RuntimeError: Replica-local variables may only be assigned in a replica context.<\/code><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nstrategy = tf.distribute.MirroredStrategy()\n\npath = '\/my\/dataset\/path\/'\nfile_paths = [f for f in glob.glob(path + \"**\/*.jpg\", recursive=True)]\n\ntensor_data = np.zeros((len(file_paths), 256, 256, 3)).astype('float32')\n\nfor i in range(len(file_paths)): \n  img_tensor = tf.image.decode_image(tf.io.read_file(file_paths[i]))\n  tensor_data[i] = img_tensor\n\nfor i in range(tensor_data.shape[0]):\n  tensor_data[i] = ((tensor_data[i] - 127.5) \/ 127.5)\n\nBUFFER_SIZE = len(file_paths)\nBATCH_SIZE = 256\n\ntrain_dataset = tf.data.Dataset.from_tensor_slices(tensor_data).shuffle(BUFFER_SIZE).batch(BATCH_SIZE)\n\ndef make_generator_model():\n    with strategy.scope():\n        model = tf.keras.Sequential()\n        model.add(layers.Dense(64*64*1536, use_bias=False, input_shape=(100,)))\n        model.add(layers.BatchNormalization())\n        model.add(layers.LeakyReLU())\n\n        model.add(layers.Reshape((64, 64, 1536)))\n        assert model.output_shape == (None, 64, 64, 1536) # Note: None is the batch size\n\n        model.add(layers.Conv2DTranspose(1536, (5, 5), strides=(1, 1), padding='same', use_bias=False))\n        assert model.output_shape == (None, 64, 64, 1536)\n        model.add(layers.BatchNormalization())\n        model.add(layers.LeakyReLU())\n\n        model.add(layers.Conv2DTranspose(768, (5, 5), strides=(2, 2), padding='same', use_bias=False))\n        assert model.output_shape == (None, 128, 128, 768)\n        model.add(layers.BatchNormalization())\n        model.add(layers.LeakyReLU())\n\n        model.add(layers.Conv2DTranspose(3, (5, 5), strides=(2, 2), padding='same', use_bias=False, activation='tanh'))\n        assert model.output_shape == (None, 256, 256, 3)\n\n        return model\n\ngenerator = make_generator_model()\n\nnoise = tf.random.normal([1, 100])\ngenerated_image = generator(noise, training=False)\n\nsample = generated_image[0, :, :, :];\nsample = tf.cast(sample, tf.int32)\n\nplt.imshow(sample, cmap=None)\n\ndef make_discriminator_model():\n    with strategy.scope():\n        model = tf.keras.Sequential()\n        model.add(layers.Conv2D(256, (5, 5), strides=(2, 2), padding='same', input_shape=[256, 256, 3]))\n        model.add(layers.LeakyReLU())\n        model.add(layers.Dropout(0.3))\n\n        model.add(layers.Conv2D(128, (5, 5), strides=(2, 2), padding='same'))\n        model.add(layers.LeakyReLU())\n        model.add(layers.Dropout(0.3))\n\n        model.add(layers.Flatten())\n        model.add(layers.Dense(1))\n\n    return model\n\ndiscriminator = make_discriminator_model()\ndecision = discriminator(generated_image)\nprint (decision)\n\ncross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=True)\n\ndef discriminator_loss(real_output, fake_output):\n    with strategy.scope():\n        real_loss = cross_entropy(tf.ones_like(real_output), real_output)\n        fake_loss = cross_entropy(tf.zeros_like(fake_output), fake_output)\n        total_loss = real_loss + fake_loss\n        return total_loss\n\ndef generator_loss(fake_output):\n    with strategy.scope():\n        return cross_entropy(tf.ones_like(fake_output), fake_output)\n\ngenerator_optimizer = tf.keras.optimizers.Adam(1e-4)\ndiscriminator_optimizer = tf.keras.optimizers.Adam(1e-4)\n\ncheckpoint_dir = '.\/training_checkpoints\/'\ncheckpoint_prefix = os.path.join(checkpoint_dir, \"ckpt\")\ncheckpoint = tf.train.Checkpoint(generator_optimizer=generator_optimizer,\n                                 discriminator_optimizer=discriminator_optimizer,\n                                 generator=generator,\n                                 discriminator=discriminator)\n\nEPOCHS = 2000\nnoise_dim = 100\nnum_examples_to_generate = 16\n\nseed = tf.random.normal([num_examples_to_generate, noise_dim])\n\n# Notice the use of `tf.function`\n# This annotation causes the function to be \"compiled\".\n@tf.function\ndef train_step(images):\n    noise = tf.random.normal([BATCH_SIZE, noise_dim])\n\n    with tf.GradientTape() as gen_tape, tf.GradientTape() as disc_tape:\n        generated_images = generator(noise, training=True)\n\n        real_output = discriminator(images, training=True)\n        fake_output = discriminator(generated_images, training=True)\n\n        gen_loss = generator_loss(fake_output)\n        disc_loss = discriminator_loss(real_output, fake_output)\n\n    gradients_of_generator = gen_tape.gradient(gen_loss, generator.trainable_variables)\n    gradients_of_discriminator = disc_tape.gradient(disc_loss, discriminator.trainable_variables)\n\n    generator_optimizer.apply_gradients(zip(gradients_of_generator, generator.trainable_variables))\n    discriminator_optimizer.apply_gradients(zip(gradients_of_discriminator, discriminator.trainable_variables))\n\ndef train(dataset, epochs):\n  for epoch in range(epochs):\n    start = time.time()\n\n    for image_batch in dataset:\n      train_step(image_batch)\n\n    # Produce images for the GIF as we go\n    display.clear_output(wait=True)\n    generate_and_save_images(generator,\n                             epoch + 1,\n                             seed)\n\n    # Save the model every 15 epochs\n    os.makedirs(os.path.dirname(checkpoint_prefix), exist_ok=True)\n\n    if (epoch + 1) % 50 == 0:\n      checkpoint.save(file_prefix = checkpoint_prefix)\n\n    print ('Time for epoch {} is {} sec'.format(epoch + 1, time.time()-start))\n\n  # Generate after the final epoch\n  display.clear_output(wait=True)\n  generate_and_save_images(generator,epochs,seed)\n\ndef generate_and_save_images(model, epoch, test_input):\n    # Notice `training` is set to False.\n    # This is so all layers run in inference mode (batchnorm).\n    predictions = model(test_input, training=False)\n\n    fig = plt.figure(figsize=(4,4))\n\n    for i in range(predictions.shape[0]):\n      plt.subplot(8, 8, i+1)\n      sample = predictions[i, :, :, :] * 127.5 + 127.5\n      sample = tf.cast(sample, tf.int32)\n      plt.imshow(sample, cmap=None)\n      plt.axis('off')\n\n    filename = '.\/screens\/eye-256x256\/1\/image_at_epoch_{:04d}.png'\n    os.makedirs(os.path.dirname(filename), exist_ok=True)\n    if (epoch + 1) % 10 == 0:\n        plt.savefig(filename.format(epoch))\n        plt.show()\n\nget_ipython().run_cell_magic('time', '', 'train(train_dataset, EPOCHS)')\n<\/code><\/pre>\n\n<p>The error is the following<\/p>\n\n<pre><code>Executing op ExperimentalRebatchDataset in device \/job:localhost\/replica:0\/task:0\/device:CPU:0\nExecuting op ExperimentalAutoShardDataset in device \/job:localhost\/replica:0\/task:0\/device:CPU:0\nExecuting op OptimizeDataset in device \/job:localhost\/replica:0\/task:0\/device:CPU:0\nExecuting op ModelDataset in device \/job:localhost\/replica:0\/task:0\/device:CPU:0\nExecuting op MultiDeviceIterator in device \/job:localhost\/replica:0\/task:0\/device:CPU:0\nExecuting op MultiDeviceIteratorInit in device \/job:localhost\/replica:0\/task:0\/device:CPU:0\nExecuting op MultiDeviceIteratorToStringHandle in device \/job:localhost\/replica:0\/task:0\/device:CPU:0\nExecuting op GeneratorDataset in device \/job:localhost\/replica:0\/task:0\/device:GPU:0\nExecuting op GeneratorDataset in device \/job:localhost\/replica:0\/task:0\/device:GPU:1\nExecuting op PrefetchDataset in device \/job:localhost\/replica:0\/task:0\/device:GPU:0\nExecuting op AnonymousIteratorV2 in device \/job:localhost\/replica:0\/task:0\/device:GPU:0\nExecuting op MakeIterator in device \/job:localhost\/replica:0\/task:0\/device:GPU:0\nExecuting op PrefetchDataset in device \/job:localhost\/replica:0\/task:0\/device:GPU:1\nExecuting op AnonymousIteratorV2 in device \/job:localhost\/replica:0\/task:0\/device:GPU:1\nExecuting op MakeIterator in device \/job:localhost\/replica:0\/task:0\/device:GPU:1\nExecuting op IteratorGetNextSync in device \/job:localhost\/replica:0\/task:0\/device:GPU:0\nExecuting op IteratorGetNextSync in device \/job:localhost\/replica:0\/task:0\/device:GPU:1\nExecuting op DestroyResourceOp in device \/job:localhost\/replica:0\/task:0\/device:CPU:0\nExecuting op DeleteIterator in device \/job:localhost\/replica:0\/task:0\/device:GPU:1\nExecuting op DeleteIterator in device \/job:localhost\/replica:0\/task:0\/device:GPU:0\n---------------------------------------------------------------------------\nRuntimeError                              Traceback (most recent call last)\n&lt;timed exec&gt; in &lt;module&gt;\n\n&lt;ipython-input-20-88a9879432c7&gt; in train(dataset, epochs)\n      4 \n      5     for image_batch in dataset:\n----&gt; 6       train_step(image_batch)\n      7 \n      8     # Produce images for the GIF as we go\n\n\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/eager\/def_function.py in __call__(self, *args, **kwds)\n    414     # This is the first call of __call__, so we have to initialize.\n    415     initializer_map = {}\n--&gt; 416     self._initialize(args, kwds, add_initializers_to=initializer_map)\n    417     if self._created_variables:\n    418       try:\n\n\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/eager\/def_function.py in _initialize(self, args, kwds, add_initializers_to)\n    357     self._concrete_stateful_fn = (\n    358         self._stateful_fn._get_concrete_function_internal_garbage_collected(  # pylint: disable=protected-access\n--&gt; 359             *args, **kwds))\n    360 \n    361     def invalid_creator_scope(*unused_args, **unused_kwds):\n\n\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/eager\/function.py in _get_concrete_function_internal_garbage_collected(self, *args, **kwargs)\n   1358     if self.input_signature:\n   1359       args, kwargs = None, None\n-&gt; 1360     graph_function, _, _ = self._maybe_define_function(args, kwargs)\n   1361     return graph_function\n   1362 \n\n\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/eager\/function.py in _maybe_define_function(self, args, kwargs)\n   1646       graph_function = self._function_cache.primary.get(cache_key, None)\n   1647       if graph_function is None:\n-&gt; 1648         graph_function = self._create_graph_function(args, kwargs)\n   1649         self._function_cache.primary[cache_key] = graph_function\n   1650       return graph_function, args, kwargs\n\n\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/eager\/function.py in _create_graph_function(self, args, kwargs, override_flat_arg_shapes)\n   1539             arg_names=arg_names,\n   1540             override_flat_arg_shapes=override_flat_arg_shapes,\n-&gt; 1541             capture_by_value=self._capture_by_value),\n   1542         self._function_attributes)\n   1543 \n\n\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/framework\/func_graph.py in func_graph_from_py_func(name, python_func, args, kwargs, signature, func_graph, autograph, autograph_options, add_control_dependencies, arg_names, op_return_value, collections, capture_by_value, override_flat_arg_shapes)\n    714                                           converted_func)\n    715 \n--&gt; 716       func_outputs = python_func(*func_args, **func_kwargs)\n    717 \n    718       # invariant: `func_outputs` contains only Tensors, CompositeTensors,\n\n\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/eager\/def_function.py in wrapped_fn(*args, **kwds)\n    307         # __wrapped__ allows AutoGraph to swap in a converted function. We give\n    308         # the function a weak reference to itself to avoid a reference cycle.\n--&gt; 309         return weak_wrapped_fn().__wrapped__(*args, **kwds)\n    310     weak_wrapped_fn = weakref.ref(wrapped_fn)\n    311 \n\n\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/framework\/func_graph.py in wrapper(*args, **kwargs)\n    704           except Exception as e:  # pylint:disable=broad-except\n    705             if hasattr(e, \"ag_error_metadata\"):\n--&gt; 706               raise e.ag_error_metadata.to_exception(type(e))\n    707             else:\n    708               raise\n\nRuntimeError: in converted code:\n\n    &lt;ipython-input-19-d2ffe8a85706&gt;:9 train_step  *\n        generated_images = generator(noise, training=True)\n    \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/base_layer.py:667 __call__\n        outputs = call_fn(inputs, *args, **kwargs)\n    \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/sequential.py:248 call\n        return super(Sequential, self).call(inputs, training=training, mask=mask)\n    \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/network.py:753 call\n        return self._run_internal_graph(inputs, training=training, mask=mask)\n    \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/network.py:895 _run_internal_graph\n        output_tensors = layer(computed_tensors, **kwargs)\n    \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/base_layer.py:667 __call__\n        outputs = call_fn(inputs, *args, **kwargs)\n    \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/layers\/normalization.py:782 call\n        self.add_update(mean_update)\n    \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/util\/deprecation.py:507 new_func\n        return func(*args, **kwargs)\n    \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/base_layer.py:1095 add_update\n        update()\n    \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/layers\/normalization.py:775 mean_update\n        return tf_utils.smart_cond(training, true_branch, false_branch)\n    \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/utils\/tf_utils.py:58 smart_cond\n        pred, true_fn=true_fn, false_fn=false_fn, name=name)\n    \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/framework\/smart_cond.py:54 smart_cond\n        return true_fn()\n    \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/layers\/normalization.py:773 &lt;lambda&gt;\n        true_branch = lambda: _do_update(self.moving_mean, new_mean)\n    \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/layers\/normalization.py:769 _do_update\n        inputs_size)\n    \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/layers\/normalization.py:458 _assign_moving_average\n        return state_ops.assign_sub(variable, update_delta, name=scope)\n    \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/ops\/state_ops.py:164 assign_sub\n        return ref.assign_sub(value)\n    \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/distribute\/values.py:1394 assign_sub\n        _assert_replica_context(self._distribute_strategy)\n    \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/distribute\/values.py:1381 _assert_replica_context\n        \"Replica-local variables may only be assigned in a replica context.\")\n\n    RuntimeError: Replica-local variables may only be assigned in a replica context.\n<\/code><\/pre>\n","answers":[{"AnswerId":"57482158","CreationDate":"2019-08-13T16:50:09.173","ParentId":null,"OwnerUserId":"4615962","Title":null,"Body":"<p>You need to distribute your dataset aslo, for details please refer to this URL.\n - <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/distribute\/training_loops\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/tutorials\/distribute\/training_loops<\/a><\/p>\n\n<pre><code>train_dist_dataset = strategy.experimental_distribute_dataset(train_dataset)\n<\/code><\/pre>\n\n<p>every part of your model creates under strategy scope like optimizer also.<\/p>\n\n<pre><code>with strategy.scope():\n  optimizer = tf.keras.optimizers.Adam()\n<\/code><\/pre>\n"}]}
+{"QuestionId":57481599,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><nlp><question-answering>","CreationDate":"2019-08-13T16:09:36.333","AcceptedAnswerId":null,"OwnerUserId":11898606.0,"Title":"How to build input to predict with saved model for BERT SQuAD2.0 with tensorflow","Body":"<p>I am trying to build the input for the saved model from BERT-SQuAD given that I have got all the elements for the input.<\/p>\n\n<p>I fine-tuned a question answering model by running of run_squad.py in <a href=\"https:\/\/github.com\/google-research\/bert\" rel=\"nofollow noreferrer\">Google bert<\/a>, then I exported the model with export_saved_model. Now when I have a new context and question, I can't build the correct input that can get return output from the model. <\/p>\n\n<p>Code to export the model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#export the model\n    def serving_input_receiver_fn():\n      feature_spec = {\n            \"unique_ids\": tf.FixedLenFeature([], tf.int64),\n            \"input_ids\": tf.FixedLenFeature([FLAGS.max_seq_length], tf.int64),\n            \"input_mask\": tf.FixedLenFeature([FLAGS.max_seq_length], tf.int64),\n            \"segment_ids\": tf.FixedLenFeature([FLAGS.max_seq_length], tf.int64),\n      }\n\n      serialized_tf_example = tf.placeholder(dtype=tf.string,\n                               shape=FLAGS.predict_batch_size,\n                               name='input_example_tensor')\n      receiver_tensors = {'examples': serialized_tf_example}\n      features = tf.parse_example(serialized_tf_example, feature_spec)\n      return tf.estimator.export.ServingInputReceiver(features, receiver_tensors)\n\n    estimator = tf.contrib.tpu.TPUEstimator(\n      use_tpu=FLAGS.use_tpu,\n      model_fn=model_fn,\n      config=run_config,\n      train_batch_size=FLAGS.train_batch_size,\n      predict_batch_size=FLAGS.predict_batch_size)\n\n    estimator._export_to_tpu = False  ## !!important to add this\n    estimator.export_saved_model(\n        export_dir_base =\"C:\/Users\/ZitongZhou\/Desktop\/qa\/bert_squad\/servemodel\",\n        serving_input_receiver_fn = serving_input_receiver_fn)\n<\/code><\/pre>\n\n<p>The way I loaded the model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>export_dir = 'servemodel'\nsubdirs = [x for x in Path(export_dir).iterdir()\n           if x.is_dir() and 'temp' not in str(x)]\nlatest = str(sorted(subdirs)[-1])\n\npredict_fn = predictor.from_saved_model(latest)\n<\/code><\/pre>\n\n<p>I got the eval_features from the run_squad.py.\nThe way I tried to build the input:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>feature_spec = {\n        \"unique_ids\": np.asarray(eval_features[0].unique_id).tolist(),\n        \"input_ids\": np.asarray(eval_features[0].input_ids).tolist(),\n        \"input_mask\": np.asarray(eval_features[0].input_mask).tolist(),\n        \"segment_ids\": np.asarray(eval_features[0].segment_ids).tolist()\n    }\nserialized_tf_example = tf.placeholder(dtype=tf.string,\n                           shape=[1],\n                           name='input_example_tensor')\nreceiver_tensors = {'examples': serialized_tf_example}\nfeatures = tf.parse_example(serialized_tf_example, feature_spec)\nout = predict_fn({'examples':[str(feature_spec)]})\n<\/code><\/pre>\n\n<p>I expect to get a prediction 'out' so I can extract the answer to the question from it.<\/p>\n\n<p>The traceback I got:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Traceback (most recent call last):\n\n  File \"&lt;ipython-input-51-0c3b618a8f48&gt;\", line 11, in &lt;module&gt;\n    features = tf.parse_example(serialized_tf_example, feature_spec)\n\n  File \"C:\\Users\\ZitongZhou\\Anaconda3\\envs\\nlp\\lib\\site-packages\\tensorflow\n\\python\\ops\\parsing_ops.py\", line 580, in parse_example\n    return parse_example_v2(serialized, features, example_names, name)\n\n  File \"C:\\Users\\ZitongZhou\\Anaconda3\\envs\\nlp\\lib\\site-packages\\tensorflow\n\\python\\ops\\parsing_ops.py\", line 803, in parse_example_v2\n    [VarLenFeature, SparseFeature, FixedLenFeature, FixedLenSequenceFeature])\n\n  File \"C:\\Users\\ZitongZhou\\Anaconda3\\envs\\nlp\\lib\\site-packages\\tensorflow\n\\python\\ops\\parsing_ops.py\", line 299, in _features_to_raw_params\n    raise ValueError(\"Invalid feature %s:%s.\" % (key, feature))\n\nValueError: Invalid feature input_ids:[101, 1005, 2129, 2214, 2003, 19523, \n6562, 1005, 102, 1005, 19523, 11233, 2003, 2274, 2086, 2214, 1005, 102,\n 0, 0, 0, 0, 0, \n0, 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, \n0, 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, \n0, 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, \n0, 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, \n0, 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, \n0, 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, \n0, 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, \n0, 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, \n0, 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, \n0, 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, \n0, 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, \n0, 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, \n0, 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, \n0, 0, 0, 0, 0, 0, 0, 0, 0, 0].\n<\/code><\/pre>\n","answers":[{"AnswerId":"57486950","CreationDate":"2019-08-14T01:00:22.960","ParentId":null,"OwnerUserId":"11898606","Title":null,"Body":"<p>I figured it out, I need to use tf.train.Example function:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def create_int_feature(values):\n    f = tf.train.Feature(int64_list=tf.train.Int64List(value=list(values)))\n    return f\n\ninputs = collections.OrderedDict()\ninputs[\"input_ids\"] = create_int_feature(features[0].input_ids)\ninputs[\"input_mask\"] = create_int_feature(features[0].input_mask)\ninputs[\"segment_ids\"] = create_int_feature(features[0].segment_ids)\ninputs[\"unique_ids\"] = create_int_feature([features[0].unique_id])\n\ntf_example = tf.train.Example(features=tf.train.Features(feature=inputs))\nout = predict_fn({'examples':[tf_example.SerializeToString()]})\n<\/code><\/pre>\n"}]}
+{"QuestionId":57481612,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-08-13T16:10:16.927","AcceptedAnswerId":null,"OwnerUserId":4515362.0,"Title":"Keyword arguments in torch.nn.Sequential (pytroch)","Body":"<p>a question regarding keywords in torch.nn.Sequential, it is possible in some way to forward keywords to specific models in a sequence?<\/p>\n\n<pre><code>    model = torch.nn.Sequential(model_0, MaxPoolingChannel(1))\n    res = model(input_ids_2, keyword_test=mask)\n<\/code><\/pre>\n\n<p>here, keyword_test should be forwarded only to the first model.<\/p>\n\n<p>Thank a lot and best regards!<\/p>\n\n<p>my duplicate from - <a href=\"https:\/\/discuss.pytorch.org\/t\/keyword-arguments-in-torch-nn-sequential\/53282\" rel=\"nofollow noreferrer\">https:\/\/discuss.pytorch.org\/t\/keyword-arguments-in-torch-nn-sequential\/53282<\/a><\/p>\n","answers":[{"AnswerId":"57482393","CreationDate":"2019-08-13T17:08:34.930","ParentId":null,"OwnerUserId":"7347631","Title":null,"Body":"<p>No; you cannot. This is only possible if all models passed to the <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.Sequential\" rel=\"nofollow noreferrer\"><code>nn.Sequential<\/code><\/a> expects the argument you are trying to pass in their <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.Module.forward\" rel=\"nofollow noreferrer\"><code>forward<\/code><\/a> method (at least at the time of writing this). <\/p>\n\n<p>Two workarounds could be (I'm not aware of the whole case, but anticipated from the question):<\/p>\n\n<ul>\n<li><p>If your value is static, why not initializing your first model with that value, and access it during computation with <code>self.keyword_test<\/code>. <\/p><\/li>\n<li><p>In case the value is dynamic, you could have it as an inherent property in the input; hence, you can access it, also, during computation with <code>input_ids_2.keyword_test<\/code><\/p><\/li>\n<\/ul>\n"}]}
+{"QuestionId":57481869,"AnswerCount":2,"Tags":"<tensorflow><protocol-buffers>","CreationDate":"2019-08-13T16:28:46.997","AcceptedAnswerId":57481870.0,"OwnerUserId":3854476.0,"Title":"How to serialize data in example-in-example format for tensorflow-ranking?","Body":"<p>I'm building a ranking model with <a href=\"https:\/\/github.com\/tensorflow\/ranking\" rel=\"nofollow noreferrer\">tensorflow-ranking<\/a>. I'm trying to serialize a data set in the TFRecord format and read it back at training time.<\/p>\n\n<p>The tutorial doesn't show how to do this. There is some documentation <a href=\"https:\/\/github.com\/tensorflow\/ranking\/blob\/master\/tensorflow_ranking\/g3doc\/api_docs\/python\/tfr\/data\/parse_from_example_in_example.md\" rel=\"nofollow noreferrer\">here<\/a> on an example-in-example data format but it's hard for me to understand: I'm not sure what the <code>serialized_context<\/code> or <code>serialized_examples<\/code> fields are or how they fit into examples and I'm not sure what the <code>Serialize()<\/code> function in the code block is.<\/p>\n\n<p>Concretely, how can I write and read data in example-in-example format?<\/p>\n","answers":[{"AnswerId":"57481870","CreationDate":"2019-08-13T16:28:46.997","ParentId":null,"OwnerUserId":"3854476","Title":null,"Body":"<p>The context is a map from feature name to <code>tf.train.Feature<\/code>. The examples list is a list of maps from feature name to <code>tf.train.Feature<\/code>. Once you have these, the following code will create an \"example-in-example\":<\/p>\n\n<pre><code>context = {...}\nexamples = [{...}, {...}, ...]\nserialized_context = tf.train.Example(features=tf.train.Features(feature=context)).SerializeToString()\nserialized_examples = tf.train.BytesList()\nfor example in examples:\n    tf_example = tf.train.Example(features=tf.train.Features(feature=example))\n    serialized_examples.value.append(tf_example.SerializeToString())\nexample_in_example = tf.train.Example(features=tf.train.Features(feature={\n    'serialized_context': tf.train.Feature(bytes_list=tf.train.BytesList(value=[serialized_context])),\n    'serialized_examples': tf.train.Feature(bytes_list=serialized_examples)\n}))\n<\/code><\/pre>\n\n<p>To read the examples back, you may call<\/p>\n\n<pre><code>tfr.data.parse_from_example_in_example(example_pb,\n    context_feature_spec = context_feature_spec,\n    example_feature_spec = example_feature_spec)\n<\/code><\/pre>\n\n<p>where <code>context_feature_spec<\/code> and <code>example_feature_spec<\/code> are maps from feature name to <code>tf.io.FixedLenFeature<\/code> or <code>tf.io.VarLenFeature<\/code>.<\/p>\n"},{"AnswerId":"57533931","CreationDate":"2019-08-17T06:42:43.993","ParentId":null,"OwnerUserId":"4936825","Title":null,"Body":"<p>First of all, I recommend reading this article to ensure that you know how to create a <code>tf.Example<\/code> as well as <code>tf.SequenceExample<\/code> (which by the way, is the other data format supported by TF-Ranking):<\/p>\n\n<p><a href=\"https:\/\/medium.com\/mostly-ai\/tensorflow-records-what-they-are-and-how-to-use-them-c46bc4bbb564\" rel=\"nofollow noreferrer\">Tensorflow Records? What they are and how to use them<\/a><\/p>\n\n<p>In the second part of this article, you will see that a <code>tf.SequenceExample<\/code> has two components: 1) Context and 2)Sequence (or examples). This is the same idea that <code>Example-in-Example<\/code> is trying to implement. Basically, context is the set of features that are independent of the items that you want to rank (a search query in the case of search, or user features in the case of a recommendation system) and the sequence part is a list of items (aka examples). This could be a list of documents (in search) or movies (in recommendation).<\/p>\n\n<p>Once you are comfortable with <code>tf.Example<\/code>, <code>Example-in-Example<\/code> will be easier to understand. Take a look at this piece of code for how to create an <code>EIE<\/code> instance:<\/p>\n\n<p><a href=\"https:\/\/www.gitmemory.com\/issue\/tensorflow\/ranking\/95\/518480361\" rel=\"nofollow noreferrer\">https:\/\/www.gitmemory.com\/issue\/tensorflow\/ranking\/95\/518480361<\/a><\/p>\n\n<p>1) bundle context features together in a <code>tf.Example<\/code> object and serialize it<\/p>\n\n<p>2) bundle sequence(example) features (each of which could contain a list of values) in another <code>tf.Example<\/code> object and serialize this one too.<\/p>\n\n<p>3) wrap these inside a parent <code>tf.Example<\/code><\/p>\n\n<p>4) (if you're writing to tfrecords) serialize the parent <code>tf.Example<\/code> object and write to your tfrecord file.<\/p>\n"}]}
+{"QuestionId":57481934,"AnswerCount":0,"Tags":"<python><image><machine-learning><keras><classification>","CreationDate":"2019-08-13T16:32:35.773","AcceptedAnswerId":null,"OwnerUserId":9578326.0,"Title":"getting high loss and very low accuracy for custom data set","Body":"<p>im building an image classifier using custom dataset of flowers. It consists of 102 classes of flowers and 18540 images of flowers. In an excel file, each flower is assigned a class.\nSo i've loaded all the images to variable X and the labels in variable y.\nso X would contain array of images like array([[[[ 54],...,[ 47]]]\nand y would contain [77,81,52,72,...]\nhere 1st image of X belongs to class 77, 2nd image belongs to 81 ans so on<\/p>\n\n<pre><code>from tensorflow.keras.datasets import cifar10\nfrom tensorflow.keras.preprocessing.image import ImageDataGenerator\nfrom tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import Dense, Dropout, Activation, Flatten\nfrom tensorflow.keras.layers import Conv2D, MaxPooling2D\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport os\nimport pandas as pd\nfrom tensorflow.python.keras.utils import np_utils\nfrom tqdm import tqdm\nimport cv2\nimport pickle\n\nCATEGORIES = [] #it will contain classes of flowers in numbers .[1,2,......,102]\n# transfer excel data to categories list\nfile_location = \"C:\/Users\/ROHIT\/Downloads\/data\/train.xlsx\"\nsheet = pd.read_excel(file_location)\nfor i in sheet.index:\n    CATEGORIES.append(sheet['category'][i])\n\n# location of training images stored in DATADIR var\nDATADIR = \"C:\/Users\/ROHIT\/Downloads\/data\/train\"\n\nIMG_SIZE = 32\ntraining_data = []\n\ndef create_training_data():\n    path = DATADIR\n    for img in tqdm(os.listdir(path)):  # iterate over each image\n        img_array = cv2.imread(os.path.join(path,img))  # convert to array\n        img_array = cv2.cvtColor (img_array, cv2.COLOR_BGR2GRAY)\n        input_img_resize = cv2.resize (img_array, (32, 32))\n        training_data.append (input_img_resize)\n\ncreate_training_data()\n\nX = training_data                              #will hold images\ny = CATEGORIES                      #will hold categories like 1,2,3....102\n\nX = np.array(training_data)\nX = X.reshape(18540, 32, 32, 1)\n#X.shape gives this -&gt; (18540, 32, 32, 1)\n\ny = np.array(y)\ny = to_categorical(y-1)\n#y.shape gives this -&gt; (18540, 102)\n\nmodel = Sequential()\nmodel.add(Conv2D(64, kernel_size=(3, 3),activation='relu',input_shape=(32,32,1)))\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(64, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(102, activation='softmax'))\n\nmodel.compile(loss='categorical_crossentropy',\n              optimizer='Adam',\n              metrics=['accuracy'])\n\nmodel.fit(X, y, epochs=10)\n<\/code><\/pre>\n\n<p>this is the output <\/p>\n\n<pre><code>Epoch 1\/10\n 6080\/18540 [========&gt;.....................] - ETA: 14:11 - loss: 63.4230 - acc: 0.0000e+0 - ETA: 9:32 - loss: 60.9470 - acc: 0.0000e+0 - ETA: 8:10 - loss: 47.2358 - acc: 0.0000e+ - ETA: 7:30 - loss: 36.6253 - acc: 0.0078   - ETA: 6:57 - loss: 30.2254 - acc: 0.00 - ETA: 6:35 - loss: 25.9565 - acc: 0.00 - ETA: 6:18 - loss: 22.9089 - acc: 0.00 - ETA: 6:05 - loss: 20.6233 - acc: 0.00 - ETA: 5:55 - loss: 18.8456 - acc: 0.00 - ETA: 5:54 - loss: 17.4225 - acc: 0.00 - ETA: 5:50 - loss: 16.2672 - acc: 0.00 - ETA: 5:48 - loss: 15.2925 - acc: 0.01 - ETA: 5:45 - loss: 14.4715 - acc: 0.00 - ETA: 5:41 - loss: 13.7661 - acc: 0.00 - ETA: 5:37 - loss: 13.1531 - acc: 0.00 - ETA: 5:33 - loss: 12.6180 - acc: 0.00 - ETA: 5:28 - loss: 12.1508 - acc: 0.01 - ETA: 5:25 - loss: 11.7345 - acc: 0.01 - ETA: 5:22 - loss: 11.3639 - acc: 0.01 - ETA: 5:23 - loss: 11.0268 - acc: 0.01 - ETA: 5:21 - loss: 10.7218 - acc: 0.01 - ETA: 5:19 - loss: 10.4446 - acc: 0.01 - ETA: 5:20 - loss: 10.1914 - acc: 0.01 - ETA: 5:19 - loss: 9.9594 - acc: 0.0156 - ETA: 5:17 - loss: 9.7459 - acc: 0.015 - ETA: 5:15 - loss: 9.5488 - acc: 0.014 - ETA: 5:13 - loss: 9.3663 - acc: 0.013 - ETA: 5:14 - loss: 9.1969 - acc: 0.014 - ETA: 5:12 - loss: 9.0391 - acc: 0.015 - ETA: 5:11 - loss: 8.8918 - acc: 0.014 - ETA: 5:14 - loss: 8.7540 - acc: 0.016 - ETA: 5:13 - loss: 8.6249 - acc: 0.015 - ETA: 5:12 - loss: 8.5036 - acc: 0.015 - ETA: 5:11 - loss: 8.3893 - acc: 0.015 - ETA: 5:10 - loss: 8.2817 - acc: 0.015 - ETA: 5:09 - loss: 8.1802 - acc: 0.014 - ETA: 5:08 - loss: 8.0833 - acc: 0.015 - ETA: 5:07 - loss: 7.9922 - acc: 0.014 - ETA: 5:05 - loss: 7.9058 - acc: 0.015 - ETA: 5:04 - loss: 7.8236 - acc: 0.014 - ETA: 5:03 - loss: 7.7454 - acc: 0.015 - ETA: 5:01 - loss: 7.6710 - acc: 0.015 - ETA: 5:00 - loss: 7.6000 - acc: 0.016 - ETA: 4:59 - loss: 7.5323 - acc: 0.016 - ETA: 4:58 - loss: 7.4676 - acc: 0.016 - ETA: 4:56 - loss: 7.4057 - acc: 0.017 - ETA: 4:55 - loss: 7.3463 - acc: 0.017 - ETA: 4:54 - loss: 7.2894 - acc: 0.016 - ETA: 4:53 - loss: 7.2350 - acc: 0.017 - ETA: 4:\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57482257,"AnswerCount":1,"Tags":"<tensorflow><keras><tensorflow2.0>","CreationDate":"2019-08-13T16:57:53.143","AcceptedAnswerId":null,"OwnerUserId":7677761.0,"Title":"How to replace OutputProjectionWrapper in tensforflow 2.0","Body":"<p>I have following code snippet for seq2seq decoder with attention mechanism. It works in tensorflow 1.13. Now I need to upgrade to tensorflow 2.0 with keras but the tf.contrib.rnn.OutputProjectionWrapper was removed in tensorflow 2.0 already. How to implement it? <\/p>\n\n<pre><code>attention_mechanism = tf.contrib.seq2seq.BahdanauAttention(\n                num_units, memory=memory,\n         memory_sequence_length=self.encoder_inputs_actual_length)\ncell = tf.contrib.rnn.LSTMCell(num_units)\nattn_cell = tf.contrib.seq2seq.AttentionWrapper(\n                cell, attention_mechanism, attention_layer_size)\nout_cell = tf.contrib.rnn.OutputProjectionWrapper(\n                attn_cell, self.output_size, reuse=reuse)\ndecoder = tf.contrib.seq2seq.BasicDecoder(\n                cell=out_cell, helper=helper,\n                initial_state=out_cell.zero_state(\n                    dtype=tf.float32, batch_size=self.batch_size))\nfinal_outputs, _, _ = tf.contrib.seq2seq.dynamic_decode(\n                decoder=decoder, output_time_major=True,\n                impute_finished=True, \nmaximum_iterations=self.input_steps\n            )\n<\/code><\/pre>\n\n<p>I read the <a href=\"https:\/\/www.oreilly.com\/library\/view\/neural-networks-and\/9781492037354\/ch04.html\" rel=\"nofollow noreferrer\">https:\/\/www.oreilly.com\/library\/view\/neural-networks-and\/9781492037354\/ch04.html<\/a> but didn't work out how to add that full connection to my case.<\/p>\n\n<p>I tried to use the latest seq2seq addons with eager mode as following, there is no syntax error but I'm not sure if it is correct. The previous tf 1.13 version prediction accuracy achieves 90% quickly but the new tf2.0 version accuracy is around 60% always.<\/p>\n\n<pre><code>attention_mechanism = tfa.seq2seq.BahdanauAttention(num_units,memory,memory_sequence_length)\nlstm_cell = layers.LSTMCell(num_units)\nattn_cell = tfa.seq2seq.AttentionWrapper(lstm_cell,attention_mechanism, attention_layer_size=num_units) \noutput_layer = layers.Dense(self.output_size)\nbasic_decoder = tfa.seq2seq.BasicDecoder(cell=attn_cell, sampler=sampler,output_layer=output_layer,output_time_major=True,impute_finished=True,maximum_iterations=self.input_steps)\ninitial_state = attn_cell.get_initial_state(batch_size=self.batch_size,dtype=tf.float32).clone(cell_state=encoder_final_state)\nfinal_outputs, _, _ = basic_decoder(encoder_outputs_sequence,initial_state=initial_state)\n<\/code><\/pre>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"57850051","CreationDate":"2019-09-09T08:08:50.807","ParentId":null,"OwnerUserId":"7677761","Title":null,"Body":"<p>I finally figured out the reason the accuracy keeps around 60% is that the AttentionWrapper will output attention score by default, but in my case I need the actual output which will be used to calculate the next attention score. The solution is to set output_attention=False in AttentionWrapper:<\/p>\n\n<pre><code>attn_cell = tfa.seq2seq.AttentionWrapper(lstm_cell,attention_mechanism, \nattention_layer_size=num_units, output_attention=False) \n<\/code><\/pre>\n\n<p>Updated it here in case anyone meet the same question.<\/p>\n"}]}
+{"QuestionId":57482823,"AnswerCount":1,"Tags":"<keras><keras-layer>","CreationDate":"2019-08-13T17:39:52.573","AcceptedAnswerId":null,"OwnerUserId":3993270.0,"Title":"keras: train a hierarchical categorical + regression model at the same time","Body":"<p>I have the following models:<\/p>\n\n<pre><code>import keras\nfrom keras.layers import Input, Dense\nfrom keras.models import Model\n\n# Joint input layer for both model A and B\ninputs = Input(shape=(12,))\n\n# ---------------------------------------\n# model_A\nx = Dense(64, activation='relu')(inputs)\nx = Dense(64, activation='relu')(x)\npredictions_A = Dense(3, activation='softmax')(x)\nmodel_A = Model(inputs=inputs, outputs=predictions_A)\n\n# ---------------------------------------\n# model_B\ninputs_B = keras.layers.concatenate([inputs, predictions_A])\nx1 = Dense(64, activation='relu')(inputs_B)\nx1 = Dense(64, activation='relu')(x1)\npredictions_B = Dense(1, activation='sigmoid')(x1)\nmodel_B = Model(inputs=inputs, outputs=predictions_B)\n<\/code><\/pre>\n\n<p>And the loss functions for both models are:<\/p>\n\n<pre><code>model_A.compile(optimizer='rmsprop',\n              loss='categorical_crossentropy',\n              metrics=['accuracy'])\n\n\nmodel_B.compile(loss='mean_squared_error', optimizer='adam')\n<\/code><\/pre>\n\n<p>I was able to train the two models separately like below:<\/p>\n\n<pre><code>model_A.fit(my_data_x, pd.get_dummies(my_data['target_categorical'],prefix=['cate_'])) \n\nmodel_B.fit(my_data_x, my_data_y)\n<\/code><\/pre>\n\n<p>The code is running, but this is not exactly what I wants.\nI want the model_A and model_B to be trained \"At the same time\". That is, model_A uses its own cross-entropy loss function while taking into account the back-prop error from model_B. Is this something possible?<\/p>\n","answers":[{"AnswerId":"57482882","CreationDate":"2019-08-13T17:44:11.150","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>You need a single model with two outputs:<\/p>\n\n<pre><code>model = Model(inputs=inputs, outputs = [predictions_A, predictions_B])\nmodel.compile(optimizer='rmsprop', \n              loss=['categorical_crossentropy', 'mse'],\n              metrics=['accuracy'])\nmodel.fit(my_data_x, \n          [pd.get_dummies(my_data['target_categorical'],prefix=['cate_']),\n           my_data_y])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57482939,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-13T17:48:35.277","AcceptedAnswerId":57483508.0,"OwnerUserId":7697890.0,"Title":"Am I receiving tensorflow 2.x warnings while using version 1.14?","Body":"<p>I have a very simple question. I recently had to format my pc and therefore I recreated my whole working environment. I'm using Anaconda (latest version) and I used conda to install tensorflow-gpu, keras..<\/p>\n\n<p>The code, like before, is working fine but I'm seeing a lot of warnings about deprecations and I'm sure a lot of them are related to tf version 2.x but I'm using tf 1.14 (I printed it after importing tensorflow in my code and so I'm sure, moreover I never asked Anaconda to install any 2.x version)<\/p>\n\n<p>For example:<\/p>\n\n<pre><code>W0813 19:25:46.491560 15288 nn_ops.py:4224] Large dropout rate: 0.75 (&gt;0.5). \nIn TensorFlow 2.x, dropout() uses dropout rate instead of keep_prob. \nPlease ensure that this is intended.\n\nW0813 19:36:22.136491 15288 deprecation_wrapper.py:119] \nFrom C:\\Users\\-----\\Anaconda3\\envs\\tensorflow_GPU\\lib\\site-packages\\keras\\callbacks.py:850: \nThe name tf.summary.merge_all is deprecated. \nPlease use tf.compat.v1.summary.merge_all instead.\n<\/code><\/pre>\n\n<p>Is this normal? Do I have to correct them or can I disable this types of warnings?<\/p>\n","answers":[{"AnswerId":"57483508","CreationDate":"2019-08-13T18:29:24.760","ParentId":null,"OwnerUserId":"11127923","Title":null,"Body":"<p>The purpose of these warning messages is to spread awareness of TensorFlow 2.0<br>\nBecause in the long run TF 2.X is likely to be widely adopted.<br>\nThis is normal and you can safely ignore these warnings by,<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import logging, os\nlogging.disable(logging.WARNING)\nos.environ[\"TF_CPP_MIN_LOG_LEVEL\"] = \"3\" \n<\/code><\/pre>\n"}]}
+{"QuestionId":57483013,"AnswerCount":1,"Tags":"<python><tensorflow><raspberry-pi><tflearn>","CreationDate":"2019-08-13T17:53:47.353","AcceptedAnswerId":null,"OwnerUserId":11923248.0,"Title":"Python saying \"no module named tflearn\" but i imported it","Body":"<p>I used Pip install TFlearn and ran my code on my Raspberry Pi but it says<\/p>\n\n<blockquote>\n  <p>no module named tflearn<\/p>\n<\/blockquote>\n\n<p>video:<a href=\"https:\/\/www.youtube.com\/watch?v=wypVcNIH6D4&amp;t=3s\" rel=\"nofollow noreferrer\">https:\/\/www.youtube.com\/watch?v=wypVcNIH6D4&amp;t=3s<\/a><\/p>\n\n<p>code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import nltk \nfrom nltk.stem.lancaster import LancasterStemmer\nstemmer =  LancasterStemmer()\n\nimport numpy\nimport tflearn\nimport tensorflow\nimport random\nimport json\n\n{\"intents\": [\n        {\"tag\": \"greeting\",\n         \"patterns\": [\"Hi\", \"How are you\", \"Is anyone there?\", \"Hello\", \"Good day\", \"Whats up\"],\n         \"responses\": [\"Hello!\", \"Good to see you again!\", \"Hi there, how can I help?\"],\n         \"context_set\": \"\"\n        },\n        {\"tag\": \"goodbye\",\n         \"patterns\": [\"cya\", \"See you later\", \"Goodbye\", \"I am Leaving\", \"Have a Good day\"],\n         \"responses\": [\"Sad to see you go :(\", \"Talk to you later\", \"Goodbye!\"],\n         \"context_set\": \"\"\n        },\n        {\"tag\": \"age\",\n         \"patterns\": [\"how old\", \"how old is tim\", \"what is your age\", \"how old are you\", \"age?\"],\n         \"responses\": [\"I am 18 years old!\", \"18 years young!\"],\n         \"context_set\": \"\"\n        },\n        {\"tag\": \"name\",\n         \"patterns\": [\"what is your name\", \"what should I call you\", \"whats your name?\"],\n         \"responses\": [\"You can call me Tim.\", \"I'm Tim!\", \"I'm Tim aka Tech With Tim.\"],\n         \"context_set\": \"\"\n        },\n        {\"tag\": \"shop\",\n         \"patterns\": [\"Id like to buy something\", \"whats on the menu\", \"what do you reccommend?\", \"could i get something to eat\"],\n         \"responses\": [\"We sell chocolate chip cookies for $2!\", \"Cookies are on the menu!\"],\n         \"context_set\": \"\"\n        },\n        {\"tag\": \"hours\",\n         \"patterns\": [\"when are you guys open\", \"what are your hours\", \"hours of operation\"],\n         \"responses\": [\"We are open 7am-4pm Monday-Friday!\"],\n         \"context_set\": \"\"\n        }\n   ]\n}\n<\/code><\/pre>\n","answers":[{"AnswerId":"57483084","CreationDate":"2019-08-13T17:59:12.303","ParentId":null,"OwnerUserId":"2061991","Title":null,"Body":"<p>Basic preliminary checks:\n run\n    <code>pip freeze<\/code> to get a list of packages installed. Ideally tflearn should be there. If not, it indicates that the package has not been installed properly.<\/p>\n"}]}
+{"QuestionId":57483103,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-13T18:00:31.527","AcceptedAnswerId":null,"OwnerUserId":3050590.0,"Title":"Unable to read weights from the pre-trained tf.keras model","Body":"<p>When I am loading my trained model, I get the following error<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.FailedPreconditionError: Error while reading resource variable module_1\/InceptionV3\/Mixed_6c\/Branch_2\/Conv2d_0d_7x1\/weights from Container: localhost. This could mean that the variable was uninitialized. Not found: Resource localhost\/module_1\/InceptionV3\/Mixed_6c\/Branch_2\/Conv2d_0d_7x1\/weights\/class tensorflow::Var does not exist.\n     [[{{node lambda_4\/module_1_apply_default\/InceptionV3\/InceptionV3\/Mixed_6c\/Branch_2\/Conv2d_0d_7x1\/Conv2D\/ReadVariableOp}}]]\n<\/code><\/pre>\n\n<p>This error message is referring to the weights which it cannot read from the model. Here is my program code<\/p>\n\n<pre><code>new_model =  tf.keras.models.load_model('box20.h5')\nprint(new_model.summary())\nnew_model.predict(test_set)\n<\/code><\/pre>\n\n<p>I receive error at the last line where I want to predict my test set. <\/p>\n\n<p>my model summary is<\/p>\n\n<blockquote>\n  <p>Model: \"sequential_4\"\n  __________________________________________________________ Layer (type)   Output Shape       Param ======================================================\n  lambda_4 (Lambda) (None, 2048)       0          _________________________________________________________\n  dense_4 (Dense)              (None, 2)                 4098<br>\n  =======================================================Total params: 4,098 \n  Trainable params: 4,098 \n  Non-trainable params: 0<\/p>\n  \n  <hr>\n<\/blockquote>\n\n<p>The solution provided <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/28287\" rel=\"nofollow noreferrer\">here<\/a> might able to solve the problem. But I am not able to understand this solution well. What is some_custom_config?<\/p>\n\n<p>I have trained my model under tf keras version 2.2.4-tf and I am using the same to load back<\/p>\n","answers":[]}
+{"QuestionId":57483198,"AnswerCount":2,"Tags":"<keras><keras-layer>","CreationDate":"2019-08-13T18:07:05.600","AcceptedAnswerId":57483382.0,"OwnerUserId":3993270.0,"Title":"Keras: Using model output as the input for another: When feeding symbolic tensors to a model, we expect thetensors to have a static batch size","Body":"<p>I have the following two models, where <code>model_A<\/code> is trained first, then the output of <code>model_A<\/code> is used to train the <code>model_C<\/code>:<\/p>\n\n<pre><code>import keras\nfrom keras.layers import Input, Dense\nfrom keras.models import Model\n\ninputs = Input(shape=(12,))\n\n# ---------------------------------------\n# model_A\nx = Dense(64, activation='relu')(inputs)\nx = Dense(64, activation='relu')(x)\npredictions_A = Dense(3, activation='softmax')(x)\nmodel_A = Model(inputs=inputs, outputs=predictions_A)\n\nmodel_A.compile(optimizer='rmsprop',\n              loss='categorical_crossentropy',\n              metrics=['accuracy'])\nmodel_A.fit(my_data_x, axis = 1), pd.get_dummies(my_data['target_cate'],prefix=['cate_']))\n#----------------------------------------\n\ninput_C_out_A = Input(shape=(3,))\n# Concatenating the two input layers\nconcat = keras.layers.concatenate([inputs, input_C_out_A])\nx1 = Dense(64, activation='relu')(concat)\nx1 = Dense(64, activation='relu')(x1)\npredictions_C= Dense(1, activation='sigmoid')(x1)\n\nmodel_C = Model(inputs=[inputs, input_C_out_A], outputs=predictions_C)\nmodel_C.compile(loss='mean_squared_error', optimizer='adam')\nmodel_C.fit([my_data_x,predictions_A], my_data['target_numeric'])\n<\/code><\/pre>\n\n<p>model_A training seems to be fine, but then I got the following errors when training the model_C:<\/p>\n\n<pre><code>Epoch 1\/1\n374667\/374667 [==============================] - 11s 30us\/step - loss: 0.3157 - acc: 0.9119\n---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-78-8df7b1dec93f&gt; in &lt;module&gt;\n     28 model_C = Model(inputs=[inputs, input_C_out_A], outputs=predictions_C)\n     29 model_C.compile(loss='mean_squared_error', optimizer='adam')\n---&gt; 30 model_C.fit([my_data_x,predictions_A], my_data['target_numeric'])\n\n~\/workspace\/git\/tensorplay\/venv\/lib\/python3.7\/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~\/workspace\/git\/tensorplay\/venv\/lib\/python3.7\/site-packages\/keras\/engine\/training.py in _standardize_user_data(self, x, y, sample_weight, class_weight, check_array_lengths, batch_size)\n    749             feed_input_shapes,\n    750             check_batch_axis=False,  # Don't enforce the batch size.\n--&gt; 751             exception_prefix='input')\n    752 \n    753         if y is not None:\n\n~\/workspace\/git\/tensorplay\/venv\/lib\/python3.7\/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\n~\/workspace\/git\/tensorplay\/venv\/lib\/python3.7\/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\n~\/workspace\/git\/tensorplay\/venv\/lib\/python3.7\/site-packages\/keras\/engine\/training_utils.py in standardize_single_array(x)\n     23                 'When feeding symbolic tensors to a model, we expect the'\n     24                 'tensors to have a static batch size. '\n---&gt; 25                 'Got tensor with shape: %s' % str(shape))\n     26         return x\n     27     elif x.ndim == 1:\n\nValueError: When feeding symbolic tensors to a model, we expect thetensors to have a static batch size. Got tensor with shape: (None, 3)\n<\/code><\/pre>\n\n<p>Any idea what I missed? Thanks!<\/p>\n","answers":[{"AnswerId":"57483382","CreationDate":"2019-08-13T18:20:09.860","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>It makes no sense to put the output of a model (a symbolic tensor) into <code>model.fit<\/code>, since there is no input data there. You should first obtain predictions from model A and then use them to fit model C:<\/p>\n\n<pre><code>pred_a = model_A.predict(my_data_x)\n\nmodel_C.fit([my_data_x, pred_a], my_data['target_numeric'])\n<\/code><\/pre>\n"},{"AnswerId":"57483406","CreationDate":"2019-08-13T18:21:58.543","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The <code>model_C<\/code> works by:<\/p>\n\n<ol>\n<li>Giving the input data to <code>model_A<\/code>,<\/li>\n<li>Getting output of <code>model_A<\/code>, and<\/li>\n<li>Feeding that, along with the original inputs, to the first <code>Dense<\/code> layer.<\/li>\n<\/ol>\n\n<p>So implement just what you said (and try to keep models separate, i.e. each with its own input\/output layers):<\/p>\n\n<pre><code>input_C = Input(shape=(12,))\nout_A = model_A(input_C) # get the output of model_A\nconcat = keras.layers.concatenate([input_C, out_A])\nx1 = Dense(64, activation='relu')(concat)\nx1 = Dense(64, activation='relu')(x1)\npredictions_C= Dense(1, activation='sigmoid')(x1)\n\nmodel_C = Model(inputs=input_C, outputs=predictions_C)\nmodel_C.compile(loss='mean_squared_error', optimizer='adam')\nmodel_C.fit(my_data_x, my_data['target_numeric'])\n<\/code><\/pre>\n\n<p>If you don't want the <code>model_A<\/code> to be trained when training <code>model_C<\/code> (i.e. if you have already trained <code>model_A<\/code> and don't want its weights to be changed), just set <code>model_A.trainable = False<\/code> before <strong>compiling<\/strong> <code>model_C<\/code>.<\/p>\n"}]}
+{"QuestionId":57483520,"AnswerCount":0,"Tags":"<c++><tensorflow><tensorboard>","CreationDate":"2019-08-13T18:30:35.347","AcceptedAnswerId":null,"OwnerUserId":3781704.0,"Title":"Editing a tensorflow event file in C++","Body":"<p>I'm visualizing a computation graph in Tensorboard and need some help:<\/p>\n\n<p>I want to be able to open\/edit\/update a Tensorflow Event file (in C++), so that I can add some \"Attributes\" (i.e. additional info) to some specific \"Namespace\" nodes. <\/p>\n\n<p>More simply, when I view the tf event file in Tensorboard (that is, when I'm viewing the GraphDef), I want to be able to see these added attributes listed under the \"Attributes\" section of a \"Namespace\" node. <\/p>\n\n<p>I've already tried a couple of things, including looking at Tensorflow's <code>summary_file_writer.cc<\/code> and <code>events_writer<\/code>, both of which create Tensorflow event files. However, I can't find any way to specifically open, read, and edit an already existing Tensorflow event file. Is there a way to do this? Python has the <code>tf.train.summary_iterator<\/code>, but I need a way to do this in C++ and edit specific attributes of a \"Namespace\" node.<\/p>\n\n<p>any help would be appreciated. thanks!<\/p>\n","answers":[]}
+{"QuestionId":57483557,"AnswerCount":0,"Tags":"<python><docker><tensorflow>","CreationDate":"2019-08-13T18:33:15.053","AcceptedAnswerId":null,"OwnerUserId":11402091.0,"Title":"SSL certificate validate failed tensorflow docker","Body":"<p>I'm trying to dockerize my TensorFlow model and I am getting an error:<\/p>\n\n<pre><code>Exception: URL fetch failure on https:\/\/storage.googleapis.com\/tensorflow\/tf-keras-datasets\/train-labels-idx1-ubyte.gz: None -- [SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed (_ssl.c:727)\n<\/code><\/pre>\n\n<p>On my local machine my code works great but when I try to run my container I have this error.<\/p>\n\n<p>Dockerfile <\/p>\n\n<pre><code>FROM continuumio\/miniconda\n\nRUN conda update --all &amp;&amp; pip install --upgrade tensorflow==2.0.0-beta1\n\nRUN conda install matplotlib\n\nRUN conda install pyqt\n\nRUN apt-get update &amp;&amp; apt-get install -yq libgl1-mesa-glx\n\nCOPY kerastensorflow.py .\n\nCMD [\"python\", \"kerastensorflow.py\"]\n<\/code><\/pre>\n\n<p>My model :<\/p>\n\n<pre><code>from __future__ import absolute_import, division, print_function, unicode_literals\n\n# TensorFlow and tf.keras\nimport tensorflow as tf\nfrom tensorflow import keras\n\n\n# Helper libraries\nimport numpy as np\nimport PyQt5\nimport matplotlib.pyplot as plt\n\n\nfashion_mnist = keras.datasets.fashion_mnist\n\n(train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()\n\n...\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57483565,"AnswerCount":3,"Tags":"<python><tensorflow><image-processing><keras><regression>","CreationDate":"2019-08-13T18:33:43.543","AcceptedAnswerId":57486429.0,"OwnerUserId":5272461.0,"Title":"Regression with convnet on image using keras - accuracy can't increase more than 40%","Body":"<p>The whole problem to solve is to read the number that these energy meters are showing. <a href=\"https:\/\/i.stack.imgur.com\/Eq9IU.jpg\" rel=\"nofollow noreferrer\">An Image of the energy meter.<\/a>\nAnd then I need to be able to implement all of it in an android application.\nWhat I am trying to do is to first by regression find the location of the black rectangle which contains the number. Then try to read the numbers by another network.<\/p>\n\n<p>1- Have I chosen the right path to this problem?<\/p>\n\n<p>2- What is wrong with my network below?<\/p>\n\n<p>I have a small dataset of 78 images like the one above.\nFor predicting the location of the rectangle I cropped some small images out of the original ones with sliding 500*200 size window. And I've got almost 10,000 images of size 500*200. With the first versions of the network, I was getting high accuracy and low loss. But the problem was that I was getting the exam same output for any input. I tried different things and retrained but there was no luck. but this last network has no accuracy.<\/p>\n\n<p>This is how i'm loading the data:<\/p>\n\n<pre><code>def load_train_data(self):\n       data = np.empty((0, 20, 50, 1), int)\n       labels = np.empty((0, 8), int)\n       files = glob.glob(self.dataset_path + '\\\\train\\\\*.jpg')\n       print('{} train files found'.format(len(files)))\n       print('loading files...')\n       for i in range(len(files)):\n           image = Image.open(files[i]).convert('L')\n           data = np.append(data, [np.array(image).reshape((20, 50, 1)).astype('float32') \/ 255], axis=0)\n           labels = np.append(labels, [np.array(self.decode_file_name(files[i]))], axis=0)\n       return (data, labels)\n\ndef decode_file_name(self, file_name):\n       arr = file_name.split('\\\\')\n       name = arr[len(arr) - 1];\n       name_parts = name[0:len(name) - 4].split(\"_\")\n       if len(name_parts) == 11:\n           temp = [int(name_parts[3]), int(name_parts[4]), int(name_parts[5]), int(name_parts[6]),\n                int(name_parts[7]), int(name_parts[8]), int(name_parts[9]), int(name_parts[10])]\n       else:\n           temp = [0, 0, 0, 0, 0, 0, 0, 0]\n\n       return temp\n<\/code><\/pre>\n\n<p>This is the model:<\/p>\n\n<pre><code>def build_model():\n     m = tf.keras.Sequential([\n        tf.keras.layers.Conv2D(64, kernel_size=10, activation=tf.keras.activations.relu, input_shape=(20, 50, 1),\n                           data_format='channels_last'),\n        tf.keras.layers.Conv2D(32, kernel_size=5, activation=tf.keras.activations.relu),\n        tf.keras.layers.Flatten(),\n        tf.keras.layers.Dense(100, activation=tf.keras.activations.relu),\n        # tf.keras.layers.Dropout(0.2),\n        tf.keras.layers.Dense(8, activation=tf.keras.activations.linear)\n     ])\n\n     m.compile(optimizer=tf.keras.optimizers.Adam(lr=0.0001), loss=tf.keras.losses.mean_absolute_percentage_error,\n          metrics=['accuracy'])\n     return m\n<\/code><\/pre>\n\n<p>And then finally the fit function:<\/p>\n\n<pre><code>model = build_model()\nhistory = model.fit(x_train, y_train, epochs=10, validation_data=(x_test, y_test))\n<\/code><\/pre>\n\n<p>This is the result of the above network:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/kLNaU.png\" rel=\"nofollow noreferrer\">Accuracy<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/zVIVN.png\" rel=\"nofollow noreferrer\">Loss<\/a><\/p>\n\n<h1>UPDATE<\/h1>\n\n<p>I found a bug in the process of loading the dataset and fixed it. I retrained the network but no matter how I change the network, accuracy stays below 40%.<\/p>\n","answers":[{"AnswerId":"57486429","CreationDate":"2019-08-13T23:22:52.870","ParentId":null,"OwnerUserId":"11162165","Title":null,"Body":"<p><a href=\"https:\/\/i.stack.imgur.com\/tVUg0.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/tVUg0.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Here's an approach using OpenCV to obtain the ROI of the black rectangle<\/p>\n\n<ul>\n<li>Convert image to grayscale and Gaussian Blur<\/li>\n<li>Canny edge detection<\/li>\n<li>Perform morphological operations to smooth image<\/li>\n<li>Find contours and filter using a minimum threshold area<\/li>\n<li>Create mask with desired rectangle<\/li>\n<li>Extract ROI<\/li>\n<\/ul>\n\n<hr>\n\n<p>Canny edge detection<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/h7bcw.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/h7bcw.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Morph close<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/IWWo8.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/IWWo8.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Find contours and filter using a minimum threshold area to isolate rectangle then draw onto a mask<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/CD3Kh.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/CD3Kh.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>From here, we find the bounding rectangle and then extract using Numpy slicing<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/RlSJr.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/RlSJr.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Result<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/2awYj.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/2awYj.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><code>import cv2\nimport numpy as np\n\nimage = cv2.imread('1.jpg')\nresult = image.copy()\nmask = np.zeros(result.shape, dtype=np.uint8)\nblur = cv2.GaussianBlur(image, (3,3), 0)\ngray = cv2.cvtColor(blur, cv2.COLOR_BGR2GRAY)\ncanny = cv2.Canny(gray, 120, 255, 1)\nkernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3,3))\nclose = cv2.morphologyEx(canny, cv2.MORPH_CLOSE, kernel)\n\ncnts = cv2.findContours(close, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)\ncnts = cnts[0] if len(cnts) == 2 else cnts[1]\n\nmin_area = 10000\nfor c in cnts:\n    area = cv2.contourArea(c)\n    if area &gt; min_area:\n        cv2.drawContours(mask, [c], -1, (255,255,255), -1)\n\nmask = cv2.cvtColor(mask, cv2.COLOR_BGR2GRAY)\nresult[mask==0] = (255,255,255)\n\nmask_canny = cv2.Canny(result, 120, 255, 1)\ncnts = cv2.findContours(mask_canny, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)\ncnts = cnts[0] if len(cnts) == 2 else cnts[1]\n\nfor c in cnts:\n    x,y,w,h = cv2.boundingRect(c)\n    ROI = result[y:y+h, x:x+w]\n    cv2.imwrite(\"ROI.png\", ROI)\n    cv2.rectangle(result, (x, y), (x + w, y + h), (36,255,12), 2)\n\ncv2.imshow('canny', canny)\ncv2.imshow('close', close)\ncv2.imshow('image', image)\ncv2.imshow('mask', mask)\ncv2.imshow('result', result)\ncv2.waitKey(0)\n<\/code><\/pre>\n"},{"AnswerId":"57483618","CreationDate":"2019-08-13T18:36:58.290","ParentId":null,"OwnerUserId":"10023182","Title":null,"Body":"<p>I think that this approach is not the best way to solve the problem. I would instead capture pictures of different energy levels and label each image as such, so that you have for example 500 images of what each energy level looks like on the meter. This would train the network to look at the rectangle when determining the energy level. The sliding window approach is unnecessary since your goal is for the network to learn what to look for instead of telling it what to look for in the image, such as the location of the rectangle. Hope that helps.<\/p>\n\n<p>Data augmentation can also be used to artificially increase the size of your dataset.<\/p>\n\n<p><a href=\"https:\/\/medium.com\/@thimblot\/data-augmentation-boost-your-image-dataset-with-few-lines-of-python-155c2dc1baec\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@thimblot\/data-augmentation-boost-your-image-dataset-with-few-lines-of-python-155c2dc1baec<\/a><\/p>\n"},{"AnswerId":"57484069","CreationDate":"2019-08-13T19:16:36.877","ParentId":null,"OwnerUserId":"11919358","Title":null,"Body":"<p>It would be a better strategy to use OpenCV to locate and crop the black box first and then use deep learning to predict the value in the box.<\/p>\n\n<p>but to get better results, you can also crop the digits of the meter individually with OpenCV. After that you can easily detect the values using a DNN similar to MNIST.<\/p>\n"}]}
+{"QuestionId":57483567,"AnswerCount":1,"Tags":"<tensorflow><keras><tensorflow-xla>","CreationDate":"2019-08-13T18:33:52.327","AcceptedAnswerId":null,"OwnerUserId":11806249.0,"Title":"Tensorflow: device CUDA:0 not supported by XLA service while setting up XLA_GPU_JIT device number 0","Body":"<p>I got this when using keras with Tensorflow backend:<\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors_impl.InvalidArgumentError: device CUDA:0 not supported by XLA service\n      while setting up XLA_GPU_JIT device number 0<\/p>\n<\/blockquote>\n\n<p>Relevant code:<\/p>\n\n<pre><code>tfconfig = tf.ConfigProto()\ntfconfig.graph_options.optimizer_options.global_jit_level = tf.OptimizerOptions.ON_1\ntfconfig.gpu_options.allow_growth = True\nK.tensorflow_backend.set_session(tf.Session(config=tfconfig))\n<\/code><\/pre>\n\n<p>tensorflow version: 1.14.0<\/p>\n","answers":[{"AnswerId":"58057774","CreationDate":"2019-09-23T07:46:18.993","ParentId":null,"OwnerUserId":"3558712","Title":null,"Body":"<p>This could be due to your TF-default (i.e. 1st) GPU is running out of memory. If you have multiple GPUs, divert your Python program to run on other GPUs. In TF (suppose using TF-2.0-rc1), set the following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Specify which GPU(s) to use\nos.environ[\"CUDA_VISIBLE_DEVICES\"] = \"1\"  # Or 2, 3, etc. other than 0\n\n# On CPU\/GPU placement\nconfig = tf.compat.v1.ConfigProto(allow_soft_placement=True, log_device_placement=True)\nconfig.gpu_options.allow_growth = True\ntf.compat.v1.Session(config=config)\n\n# Note that ConfigProto disappeared in TF-2.0\n<\/code><\/pre>\n\n<p>Suppose, however, your environment have only one GPU, then perhaps you have no choice but ask your buddy to stop his program, then treat him a cup of coffee.<\/p>\n"}]}
+{"QuestionId":57483933,"AnswerCount":1,"Tags":"<python><mpi><pytorch><openmpi>","CreationDate":"2019-08-13T19:05:51.423","AcceptedAnswerId":null,"OwnerUserId":3906173.0,"Title":"Getting distributed package doesn't have mpi built in error","Body":"<p>I have been trying to write a distributed application using pytorch. I have been following tutorial <a href=\"https:\/\/pytorch.org\/tutorials\/intermediate\/dist_tuto.html\" rel=\"nofollow noreferrer\">here<\/a>. Over there, I am using the \"MPI Backend\" option. According to that, I need to follow the basic steps to install pytorch and then install openmpi as <code>conda install -c conda-forge openmpi<\/code><\/p>\n\n<p>Unfortunately, whenever I try to run a script using mpirun <code>mpiexec -n 2 python ptdist.py<\/code>, I get the following error <code>RuntimeError: Distributed package doesn't have MPI built in<\/code>. I believe this is happening because of error in <code>import ProcessGroupMPI<\/code> code <a href=\"https:\/\/github.com\/pytorch\/pytorch\/blob\/master\/torch\/distributed\/distributed_c10d.py\" rel=\"nofollow noreferrer\">here<\/a> in python.<\/p>\n\n<p>I have tried to install openmpi from their <a href=\"http:\/\/www.open-mpi.org\" rel=\"nofollow noreferrer\">source code<\/a> as well as <code>sudo apt-get install python-mpi4py<\/code>, but am still facing the same error.<\/p>\n\n<p>I also tried <code>pip install mpi4py<\/code> but that also does not help<\/p>\n\n<p>Does anyone know what is the problem?<\/p>\n","answers":[{"AnswerId":"57487718","CreationDate":"2019-08-14T03:11:29.130","ParentId":null,"OwnerUserId":"8062491","Title":null,"Body":"<p>From <a href=\"https:\/\/medium.com\/@esaliya\/pytorch-distributed-with-mpi-acb84b3ae5fd\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@esaliya\/pytorch-distributed-with-mpi-acb84b3ae5fd<\/a><\/p>\n\n<blockquote>\n  <p>The MPI backend, though supported, is not available unless you compile PyTorch from its source<\/p>\n<\/blockquote>\n\n<p>This suggests you should <strong>first<\/strong> install your favorite MPI library, and possibly <code>mpi4py<\/code> built on top of it, and then build <code>pytorch<\/code> from sources at last.<\/p>\n"}]}
+{"QuestionId":57484135,"AnswerCount":0,"Tags":"<python><linux><shell><tensorflow><machine-learning>","CreationDate":"2019-08-13T19:23:24.567","AcceptedAnswerId":null,"OwnerUserId":10412508.0,"Title":"Google.protobuf Import Error in Shell Mode","Body":"<p>I want to run a code that uses google.protobuf module below. The library is recognized and works fine in the python interpreter, but it gives an error in the shell mode.<\/p>\n\n<p>I'm currently operating on a Linux CentOS 7. <\/p>\n\n<p><em>Python 2.7.5 Console within Shell<\/em><\/p>\n\n<p><code>example_feat_extract_test.py<\/code><\/p>\n\n<pre><code>from google import protobuf\nfrom google.protobuf import symbol_database as _symbol_database\n<\/code><\/pre>\n\n<p><em>Terminal on Linux:<\/em><\/p>\n\n<pre><code>$ python example_feat_extract_test.py\n<\/code><\/pre>\n\n<p><em>The Error Message:<\/em><\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):\n    File \"example_feat_extract_test.py\", line 2, in \n      from google.protobuf import symbol_database as _symbol_database\n  ImportError: cannot import name symbol_database<\/p>\n<\/blockquote>\n\n<p>I tried using <code>execfile('example_feat_extract_test.py')<\/code>. In the interpreter, <code>execfile()<\/code> doesn't give any error, but the machine learning tool I want to use (tensorflow) only runs in the shell mode as I'm using a Shared Computing Cluster with local package installation restrictions. So I want to know how to solve the above error in the shell mode.<\/p>\n","answers":[]}
+{"QuestionId":57484271,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-13T19:36:09.070","AcceptedAnswerId":null,"OwnerUserId":10290211.0,"Title":"Tensorflow ImportError - can't import 'string_int_label_map_pb2'","Body":"<p>I am new to Tensorflow and am getting an Import Error from utils import label_map_util. I followed all the installation steps of the models but I'm still getting this error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>ImportError                               Traceback (most recent call last)\n&lt;ipython-input-6-894414a64172&gt; in &lt;module&gt;()\n----&gt; 1 from utils import label_map_util\n      2 from utils import visualization_utils as vis_util\n\nC:\\Tensorflow1\\models\\research\\object_detection\\utils\\label_map_util.py in &lt;module&gt;()\n     24 import tensorflow as tf\n     25 from google.protobuf import text_format\n---&gt; 26 from object_detection.protos import string_int_label_map_pb2\n     27 \n     28 \n\nImportError: cannot import name 'string_int_label_map_pb2'\n\nfrom utils import label_map_util\nfrom utils import visualization_utils as vis_util```\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57484426,"AnswerCount":3,"Tags":"<python><tensorflow><pip>","CreationDate":"2019-08-13T19:48:00.980","AcceptedAnswerId":null,"OwnerUserId":9608225.0,"Title":"PIP Won't install Tensorflow","Body":"<p>I'm trying to install tensorflow with pip, but every time I try I get this:<\/p>\n\n<pre><code>pip3 install Tensorflow\nCollecting Tensorflow\n  ERROR: Could not find a version that satisfies the requirement Tensorflow \n(from versions: none)\nERROR: No matching distribution found for Tensorflow\n<\/code><\/pre>\n\n<p>What do I do?\n(I'm running Pip 19.2.2 on Windows 7 x64 if this info is needed)<\/p>\n","answers":[{"AnswerId":"57486657","CreationDate":"2019-08-14T00:02:55.950","ParentId":null,"OwnerUserId":"8490683","Title":null,"Body":"<p>Run <\/p>\n\n<pre><code>pip install --upgrade tensorflow\n<\/code><\/pre>\n"},{"AnswerId":"57484462","CreationDate":"2019-08-13T19:50:33.887","ParentId":null,"OwnerUserId":"10976096","Title":null,"Body":"<p>Try <code>python3 -m pip install tensorflow<\/code><\/p>\n\n<p>Also tensorflow only supports 64 bit version Python, check that you have installed the correct version.<\/p>\n"},{"AnswerId":"57484507","CreationDate":"2019-08-13T19:54:47.740","ParentId":null,"OwnerUserId":"3968014","Title":null,"Body":"<p>Also, often times its easier to use a slightly older Python version together with Tensorflow. I have found installations of e.g. Python 3.6.x + tf to work more smoothly than 3.7.x.<\/p>\n"}]}
+{"QuestionId":57484508,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-08-13T19:54:54.470","AcceptedAnswerId":null,"OwnerUserId":3167448.0,"Title":"How does one use 3D convolutions on standard 3 channel images?","Body":"<p>I am trying to use 3d conv on cifar10 data set (just for fun). I see the docs that we usually have the input be 5d tensors (N,C,D,H,W). Am I really forced to pass 5 dimensional data necessarily?<\/p>\n\n<p>The reason I am skeptical is because 3D convolutions simply mean my conv moves across 3 dimensions\/directions. So technically I could have 3d 4d 5d or even 100d tensors and then should all work as long as its at least a 3d tensor. Is that not right?<\/p>\n\n<p>I tried it real quick and it did give an error:<\/p>\n\n<pre><code>import torch\n\u200b\n\u200b\ndef conv3d_example():\n    N,C,H,W = 1,3,7,7\n    img = torch.randn(N,C,H,W)\n    ##\n    in_channels, out_channels = 1, 4\n    kernel_size = (2,3,3)\n    conv = torch.nn.Conv3d(in_channels, out_channels, kernel_size)\n    ##\n    out = conv(img)\n    print(out)\n    print(out.size())\n\u200b\n##\nconv3d_example()\n---------------------------------------------------------------------------\nRuntimeError                              Traceback (most recent call last)\n&lt;ipython-input-3-29c73923cc64&gt; in &lt;module&gt;\n     15 \n     16 ##\n---&gt; 17 conv3d_example()\n\n&lt;ipython-input-3-29c73923cc64&gt; in conv3d_example()\n     10     conv = torch.nn.Conv3d(in_channels, out_channels, kernel_size)\n     11     ##\n---&gt; 12     out = conv(img)\n     13     print(out)\n     14     print(out.size())\n\n~\/anaconda3\/lib\/python3.7\/site-packages\/torch\/nn\/modules\/module.py in __call__(self, *input, **kwargs)\n    491             result = self._slow_forward(*input, **kwargs)\n    492         else:\n--&gt; 493             result = self.forward(*input, **kwargs)\n    494         for hook in self._forward_hooks.values():\n    495             hook_result = hook(self, input, result)\n\n~\/anaconda3\/lib\/python3.7\/site-packages\/torch\/nn\/modules\/conv.py in forward(self, input)\n    474                             self.dilation, self.groups)\n    475         return F.conv3d(input, self.weight, self.bias, self.stride,\n--&gt; 476                         self.padding, self.dilation, self.groups)\n    477 \n    478 \n\nRuntimeError: Expected 5-dimensional input for 5-dimensional weight 4 1 2 3, but got 4-dimensional input of size [1, 3, 7, 7] instead\n<\/code><\/pre>\n\n<hr>\n\n<p>cross posted:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/discuss.pytorch.org\/t\/how-does-one-use-3d-convolutions-on-standard-3-channel-images\/53330\" rel=\"nofollow noreferrer\">https:\/\/discuss.pytorch.org\/t\/how-does-one-use-3d-convolutions-on-standard-3-channel-images\/53330<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/57484508\/how-does-one-use-3d-convolutions-on-standard-3-channel-images\">How does one use 3D convolutions on standard 3 channel images?<\/a><\/li>\n<\/ul>\n","answers":[{"AnswerId":"57501841","CreationDate":"2019-08-14T20:46:57.430","ParentId":null,"OwnerUserId":"2790047","Title":null,"Body":"<p>Consider the following scenario. You have a 3 channel NxN image. This image will have size of 3xNxN in pytorch (ignoring the batch dimension for now).<\/p>\n\n<p>Say you pass this image to a 2D convolution layer with no bias, kernel size 5x5, padding of 2, and input\/output channels of 3 and 10 respectively.<\/p>\n\n<p>What's actually happening when we apply this layer to the input image?<\/p>\n\n<p>You can think of it like this...<\/p>\n\n<p>For each of the 10 output channels there is a kernel of size 3x5x5. A <strong>3D<\/strong> convolution is applied to the 3xNxN input image using this kernel, which can be thought of as unpadded in the first dimension. The result of this convolution is a 1xNxN feature map.<\/p>\n\n<p>Since there are 10 output layers, there are 10 of the 3x5x5 kernels. After all kernels have been applied the outputs are stacked into a single 10xNxN tensor.<\/p>\n\n<p>So really, in the classical sense, a 2D convolution layer is already performing a 3D convolution.<\/p>\n\n<p>Similarly for a 3D convolution layer, its really doing a 4D convolution, which is why you need 5 dimensional input.<\/p>\n"}]}
+{"QuestionId":57484616,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><deep-learning><nlp>","CreationDate":"2019-08-13T20:03:57.657","AcceptedAnswerId":null,"OwnerUserId":10458815.0,"Title":"Why is loss changing with dense and embedding layers sizes?","Body":"<p>I have a simple model as follows:<\/p>\n\n<pre><code>class Model(tf.keras.Model):\n    def __init__(self, vocab_size, embedding_size):\n        super(Model, self).__init__()\n\n        self.embedding = tf.keras.layers.Embedding(vocab_size, embedding_size)\n        self.out = tf.keras.layers.Dense(vocab_size)\n\n    def call(self, x):\n        return self.out(self.embedding(x))\n<\/code><\/pre>\n\n<p>My loss is calculated using softmax cross entropy and weights updated with the adam optimizer.<\/p>\n\n<p>When vocab_size = 2 (for testing purposes), loss is this:\n<a href=\"https:\/\/i.stack.imgur.com\/xFr71.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/xFr71.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>When vocab_size = 20200 (full size of vocab), loss is this:\n<a href=\"https:\/\/i.stack.imgur.com\/gPYg0.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/gPYg0.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Why is the shape of the loss changing, when everything is the same except for dense and embedding layer sizes? And how do I get the same, nice shape for loss with the larger vocabulary size?<\/p>\n","answers":[]}
+{"QuestionId":57484653,"AnswerCount":1,"Tags":"<object><keras><attributes><nonetype>","CreationDate":"2019-08-13T20:08:00.823","AcceptedAnswerId":null,"OwnerUserId":11924121.0,"Title":"Attribute error: None type has no attribute summary in keras","Body":"<p>I have tried to go in deep with my understanding of word embedding and NLP in keras implementing and copying part of the code creating a Keras model using functional API. When I launch model.summary I receive an Attribute error: None type has no attribute 'summary'.<\/p>\n\n<p>After many attempts decreasing the numbers of layers, the dimension of word embedding matrix unfortunately nothing changed. I don't know what to do.<\/p>\n\n<pre><code>def pretrained_embedding_layer(word_to_vec, word_to_index):\n\n\n    vocab_len = len(word_to_index) + 1                  \n    emb_dim = word_to_vec[\"sole\"].shape[0]      \n    emb_matrix = np.zeros((vocab_len,emb_dim))\n\n    for word, index in word_to_index.items():\n        emb_matrix[index, :] = word_to_vec[word]\n    print(emb_matrix.shape)\n\n    embedding_layer = Embedding(vocab_len,emb_dim,trainable =False)\n    embedding_layer.build((None,))\n\n\n    embedding_layer.set_weights([emb_matrix])\n\n    return embedding_layer\ndef Chatbot_V1(input_shape, word_to_vec, word_to_index):\n\n    # Define sentence_indices as the input of the graph, it should be of shape input_shape and dtype 'int32' (as it contains indices).\n    sentence_indices = Input(input_shape, dtype='int32')\n\n    # Create the embedding layer pretrained with GloVe Vectors (\u22481 line)\n    embedding_layer = pretrained_embedding_layer(word_to_vec, word_to_index)\n\n    embeddings = embedding_layer(sentence_indices)   \n    # Propagate the embeddings through an LSTM layer with 128-dimensional hidden state\n    X = LSTM(128, return_sequences=True)(embeddings)\n\n    # Add dropout with a probability of 0.5\n    X = Dropout(0.5)(X)\n    # Propagate X trough another LSTM layer with 128-dimensional hidden state\n    # Be careful, the returned output should be a single hidden state, not a batch of sequences.\n    X = LSTM(128, return_sequences=True)(X)\n    # Add dropout with a probability of 0.5\n    X = Dropout(0.5)(X)\n    # Propagate X through a Dense layer with softmax activation to get back a batch of vocab_dim dimensional vectors.\n    X = Dense(vocab_dim)(X)\n    # Add a softmax activation\n    preds = Activation('softmax')(X)\n\n    # Create Model instance which converts sentence_indices into X.\n    model = Model(sentence_indices, preds)\n\nmodel = Chatbot_V1((maxLen,), word_to_vec, word_to_index)\n\nmodel.summary()\n<\/code><\/pre>\n\n<p>Launching model.summary:\nAttributeError: 'NoneType' object has no attribute 'summary'<\/p>\n\n<p>Why? What is wrong in layers definition?<\/p>\n","answers":[{"AnswerId":"57484811","CreationDate":"2019-08-13T20:20:43.820","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>The function <code>Chatbot_V1<\/code> does not return anything, and in python this is signaled by <code>None<\/code> if you assign the return value of the function to a variable. So just use the <code>return<\/code> keyword to return the model at the end of <code>Chatbot_V1<\/code><\/p>\n"}]}
+{"QuestionId":57484785,"AnswerCount":1,"Tags":"<tensorflow><session>","CreationDate":"2019-08-13T20:18:29.050","AcceptedAnswerId":57485005.0,"OwnerUserId":10229833.0,"Title":"Session graph is empty although graph is set as default","Body":"<p>I get the error message that session graph is empty.. I have looked into different questions on stackoverflow page but nothing worked for me..<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import os\nimport sys\nimport scipy.io\nimport scipy.misc\nimport matplotlib.pyplot as plt\nfrom matplotlib.pyplot import imshow\nfrom PIL import Image\nfrom nst_utils import *\nimport numpy as np\nimport tensorflow as tf\n\nmodel = load_vgg_model(\"imagenet-vgg-verydeep-19.mat\")\n\ntf.reset_default_graph()\n\n# Start interactive session\nsess = tf.InteractiveSession()\n\ncontent_image = scipy.misc.imresize(arr = scipy.misc.imread(\"content.jpeg\"), size=(300,400,3))\ncontent_image = reshape_and_normalize_image(content_image)\n\nstyle_image = scipy.misc.imresize(arr = scipy.misc.imread(\"monet.jpg\"), size=(300,400,3))\nstyle_image = reshape_and_normalize_image(style_image)\n\ngenerated_image = generate_noise_image(content_image)\nimshow(generated_image[0])\n\n# Assign the content image to be the input of the VGG model.  \nwith tf.Session() as sess:\n    sess.run(model['input'].assign(content_image))\n<\/code><\/pre>\n\n<p>I get \"RuntimeError: The Session graph is empty.  Add operations to the graph before calling run().\" Appreciate every advise I can get :)<\/p>\n","answers":[{"AnswerId":"57485005","CreationDate":"2019-08-13T20:38:07.817","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>A couple of errors:<\/p>\n\n<ol>\n<li>You are instantiating a model <code>model = load_vgg_model(\"imagenet-vgg-verydeep-19.mat\")<\/code> then reseting the default graph <code>tf.reset_default_graph()<\/code> &rarr; it should be the other way around.<\/li>\n<li>You have one <code>tf.InteractiveSession()<\/code> started and you are creating another <code>tf.Session()<\/code> &rarr; You don't need the first one.<\/li>\n<\/ol>\n\n<p>This should work for you:<\/p>\n\n<p>import os\nimport sys\nimport scipy.io\nimport scipy.misc\nimport matplotlib.pyplot as plt\nfrom matplotlib.pyplot import imshow\nfrom PIL import Image\nfrom nst_utils import *\nimport numpy as np\nimport tensorflow as tf<\/p>\n\n<pre><code>tf.reset_default_graph()\nmodel = load_vgg_model(\"imagenet-vgg-verydeep-19.mat\")\n\ncontent_image = scipy.misc.imresize(arr = scipy.misc.imread(\"content.jpeg\"), size=(300,400,3))\ncontent_image = reshape_and_normalize_image(content_image)\n\nstyle_image = scipy.misc.imresize(arr = scipy.misc.imread(\"monet.jpg\"), size=(300,400,3))\nstyle_image = reshape_and_normalize_image(style_image)\n\ngenerated_image = generate_noise_image(content_image)\nimshow(generated_image[0])\n\n# Assign the content image to be the input of the VGG model.  \nsess = tf.InteractiveSession()\nsess.run(model['input'].assign(content_image))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57485040,"AnswerCount":0,"Tags":"<keras><deep-learning><convolution><autoencoder>","CreationDate":"2019-08-13T20:41:08.250","AcceptedAnswerId":null,"OwnerUserId":11548259.0,"Title":"Keras Convolution layer kernel size with more than 3 channels","Body":"<p>I have a question regarding the filter of a convolutional layer in keras. I have image data with 4 channels which are fed through a 3D convolutional network, and Iam not sure how to set up the Conv3D layer in keras.<\/p>\n\n<p>According to the Keras documentation, the kernel size is given by (depth, width and height).<\/p>\n\n<p>Now Iam wondering if it would be correct to set up my Conv3D layer like this (i.ex. with 128 filters):<\/p>\n\n<pre><code>Conv3D(128, (4,3,3), activation='relu',padding = 'same')\n<\/code><\/pre>\n\n<p>(My input is given according to the channel last convention).<\/p>\n\n<p>Thanks in advance,<\/p>\n\n<p>Mike<\/p>\n","answers":[]}
+{"QuestionId":57485064,"AnswerCount":0,"Tags":"<tensorflow><object-detection><amazon-sagemaker><mobilenet>","CreationDate":"2019-08-13T20:43:14.413","AcceptedAnswerId":null,"OwnerUserId":5775451.0,"Title":"How to run an Object Detection prediction on a MobileNet SSD model hosted on SageMaker","Body":"<p>I trained a MobileNet v1 SSD model using AWS SageMaker and it trained perfectly.  I have an export\/Servo\/xxx\/saved_graph (a tflite model derived from the checkpoints worked perfectly on an IoT device.)    The model deployed on SageMaker as an endpoint without a problem.   My problem is, I don't know how to feed images into the model endpoint.  (Which presumably is the same as asking how do I feed data into a served model?)   <\/p>\n\n<p>When I query the model (below), I see the input signature is 'serialized_example'.   So, I'm assuming I need a serialized example.   That's where I run out of talent.  <\/p>\n\n<p>I need help understanding how to convert jpg file into a serialized example.  Or how can I use a tfrecord.  <\/p>\n\n<pre><code>saved_model_cli show --dir {MODEL_PATH} --tag_set serve --signature_def tensorflow\/serving\/predict\n<\/code><\/pre>\n\n<p>yields (I omitted the outputs):<\/p>\n\n<pre><code>  inputs['serialized_example'] tensor_info:\n      dtype: DT_STRING\n      shape: ()\n      name: tf_example:0\n<\/code><\/pre>\n\n<p>I have a Dataset but don't know how to get the right class -<\/p>\n\n<pre><code>raw_dataset = tf.data.TFRecordDataset(tfrecord_file_list)\nexample_decoder = TfExampleDecoder()\ndecoded_dataset = raw_dataset.map(example_decoder.decode)\n<\/code><\/pre>\n\n<p>if I enumerate across a decoded dataset, how can I feed the data into my model?<\/p>\n\n<pre><code>for n, e in enumerate(decoded_dataset.take(4)):\n   what do I do with e - which looks like:\n\n&lt;DatasetV1Adapter shapes: {image: (?, ?, 3), source_id: (), key: (), filename: (), groundtruth_boxes: (?, 4), groundtruth_area: (?,), groundtruth_is_crowd: (?,), groundtruth_difficult: (?,), groundtruth_group_of: (?,), groundtruth_weights: (?,), groundtruth_classes: (?,), groundtruth_image_classes: (?,), original_image_spatial_shape: (2,)}, types: {image: tf.uint8, source_id: tf.string, key: tf.string, filename: tf.string, groundtruth_boxes: tf.float32, groundtruth_area: tf.float32, groundtruth_is_crowd: tf.bool, groundtruth_difficult: tf.int64, groundtruth_group_of: tf.int64, groundtruth_weights: tf.float32, groundtruth_classes: tf.int64, groundtruth_image_classes: tf.int64, original_image_spatial_shape: tf.int32}&gt;\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57485125,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-13T20:47:29.277","AcceptedAnswerId":null,"OwnerUserId":2603684.0,"Title":"How to perform stacking of tensors in a loop when there are many stacking elements?","Body":"<p>I am trying to stack tensors. I use tf.stack() for that purpose. However, the elements I need to stack changes. However, I will know their count. So, is there a way to perform the stacking in a loop? <\/p>\n\n<p>I have I as a tensor of shape : [batch_size,256,256,1]\nand I have i as a tensor of shape [256,256,1].<\/p>\n\n<pre><code># I used to do\nI = tf.stack([i, i, i, i, i]) # if batch_size is 5\n# But I can have a varying value of batch_size.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57485157,"AnswerCount":1,"Tags":"<tensorflow><hook><metrics>","CreationDate":"2019-08-13T20:51:38.173","AcceptedAnswerId":null,"OwnerUserId":6249001.0,"Title":"Tensorflow: Save metrics every certain steps","Body":"<p>I have trained a model and want to access the standard metrics plus a custom metric, every 100 steps. <\/p>\n\n<pre><code>def top3error(features, labels, predictions):\n    return {'top3error': tf.metrics.mean(tf.nn.in_top_k(predictions=predictions['logits'], \n                                                        targets=labels,\n                                                        k=3))}\nm = tf.estimator.DNNClassifier(\n      # Config &amp; Params\n      config=config\n      model_dir=model_dir,\n      n_classes=n_classes,\n      # Model\n      feature_columns=deep_columns,\n      hidden_units=[256, 128],\n      activation_fn='relu',\n      optimizer=tf.train.ProximalAdagradOptimizer(\n        learning_rate=0.1,\n        l1_regularization_strength=0.001)\n    )\n'''Add metrics'''          \nm = tf.estimator.add_metrics(m, top3error)\n<\/code><\/pre>\n\n<p>I have added a config variable to specify that every 100 steps I want to save the metrics. I have been successful to print the metrics. <\/p>\n\n<pre><code>config=tf.estimator.RunConfig(\n  model_dir=model_dir, \n  save_checkpoints_steps=100,\n)\n<\/code><\/pre>\n\n<p>However after the model has been trained I can only access the last value for each metric. Is it possible to save the metrics everytime there is a checkpoint?<\/p>\n\n<pre><code>  results = tf.estimator.train_and_evaluate(m, train_spec, eval_spec)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57486180","CreationDate":"2019-08-13T22:46:03.190","ParentId":null,"OwnerUserId":"11919358","Title":null,"Body":"<p>you can use tensorflow summaries. You can also use that to visualize your metric in tensorboard.<\/p>\n\n<p>check this out-><a href=\"https:\/\/stackoverflow.com\/questions\/42164772\/tensorflow-estimator-api-summaries\">Tensorflow Estimator API: Summaries<\/a><\/p>\n"}]}
+{"QuestionId":57485194,"AnswerCount":2,"Tags":"<python><tensorflow><keras><conv-neural-network>","CreationDate":"2019-08-13T20:54:40.543","AcceptedAnswerId":57646566.0,"OwnerUserId":11149645.0,"Title":"CNN training system memory issue","Body":"<p>i am trying to train a CNN on CPU with a input of (102480,40,40,2) dimensions and output of (40,40). Before i trained a model using (102480,40,40) input and it worked perfectly. Now when i run the script systems memory usage increases until it needs more and my pc freezes. I only have 8 GB ram. <\/p>\n\n<p>i tried adding this line to my code which did not help at all.<\/p>\n\n<pre><code>sess = tf.Session(config=tf.ConfigProto(inter_op_parallelism_threads=1, intra_op_parallelism_threads=1))\n<\/code><\/pre>\n\n<p>This is my model:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv2D(64,(3,3), input_shape = train_data.shape[1:]))\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'))\n\n\nmodel.add(Dense(1600))\nmodel.add(Reshape((40,40)))\nmodel.add(Activation('sigmoid'))\n\nmodel.compile(loss='MSE',\n             optimizer='SGD',\n             metrics=['MAE'])\n\nhistory = model.fit(train_data, train_target,\n          batch_size=256,\n          epochs=100,\n          validation_split=0.1,\n          callbacks = [cp_callback])\n<\/code><\/pre>\n\n<p>Can i somehow decrease the system memory usage without losing accuracy?<\/p>\n","answers":[{"AnswerId":"57646566","CreationDate":"2019-08-25T13:44:47.800","ParentId":null,"OwnerUserId":"11149645","Title":null,"Body":"<p>Instead of decreasing the batch size (which can impact the accuracy) I simply changed the data types from float64 to float32. This cut the memory usage by half.  <\/p>\n"},{"AnswerId":"57485812","CreationDate":"2019-08-13T21:59:49.493","ParentId":null,"OwnerUserId":"6598433","Title":null,"Body":"<p>The calculation of the model memory depends on the batch size. In fact, you can roughly estimate the memory usage by multiplying the batch size by the input shape. It's like you have 256 times the input size in your case. You can try to lower the batch size, but that may impact your accuracy. A way to minimize it's as stated in this answer: <a href=\"https:\/\/datascience.stackexchange.com\/a\/12533\">https:\/\/datascience.stackexchange.com\/a\/12533<\/a><\/p>\n\n<blockquote>\n  <p>In the case that you do need bigger batch sizes but it will not fit on\n  your GPU, you can feed a small batch, save the gradient estimates and\n  feed one or more batches, and then do a weight update. This way you\n  get a more stable gradient because you increased your virtual batch\n  size.<\/p>\n<\/blockquote>\n\n<p>Finally, you can calculate the approximate memory usage using the solution of this answer <a href=\"https:\/\/stackoverflow.com\/a\/46216013\/6598433\">https:\/\/stackoverflow.com\/a\/46216013\/6598433<\/a> by ZFTurbo that I put below:<\/p>\n\n<pre><code>def get_model_memory_usage(batch_size, model):\n    import numpy as np\n    from keras import backend as K\n\n    shapes_mem_count = 0\n    for l in model.layers:\n        single_layer_mem = 1\n        for s in l.output_shape:\n            if s is None:\n                continue\n            single_layer_mem *= s\n        shapes_mem_count += single_layer_mem\n\n    trainable_count = np.sum([K.count_params(p) for p in set(model.trainable_weights)])\n    non_trainable_count = np.sum([K.count_params(p) for p in set(model.non_trainable_weights)])\n\n    number_size = 4.0\n    if K.floatx() == 'float16':\n         number_size = 2.0\n    if K.floatx() == 'float64':\n         number_size = 8.0\n\n    total_memory = number_size*(batch_size*shapes_mem_count + trainable_count + non_trainable_count)\n    gbytes = np.round(total_memory \/ (1024.0 ** 3), 3)\n    return gbytes\n<\/code><\/pre>\n"}]}
+{"QuestionId":57485629,"AnswerCount":0,"Tags":"<android><tensorflow><gradle><flutter><dart>","CreationDate":"2019-08-13T21:38:44.703","AcceptedAnswerId":null,"OwnerUserId":11919769.0,"Title":"Unable to use the dependency tflite for ML app with flutter","Body":"<p>I am setting up a ML environment on a flutter based android app where it uses tflite as a prediction strategy, but when I include the tflite dependency the Gradle build is getting failed. How do I fix this?<\/p>\n\n<p>This is the image of the output when run in cmd:\n<img src=\"https:\/\/i.stack.imgur.com\/JmRbr.png\" alt=\"This is the image of the output when run in cmd \"><\/p>\n\n<p>I have tried using the older versions of the module, but still, the problem isn't getting fixed. <\/p>\n\n<p>I have added the following in the app\/build.gradle <\/p>\n\n<pre><code>android {\n    compileSdkVersion 28\n    aaptOptions {\n        noCompress 'tflite'\n        noCompress 'lite'\n    }\n<\/code><\/pre>\n\n<pre><code>  cupertino_icons: ^0.1.2\n  firebase_core: ^0.3.4\n  firebase_auth: ^0.8.4+5\n  google_sign_in: ^4.0.1+3\n  image_picker:  ^0.6.0+1\n  tflite: ^1.0.4\n<\/code><\/pre>\n\n<p>The output is<\/p>\n\n<pre><code>Initializing gradle...                                              2.2s\nResolving dependencies...                                          20.6s\nregisterResGeneratingTask is deprecated, use registerGeneratedResFolders(FileCollection)\nregisterResGeneratingTask is deprecated, use registerGeneratedResFolders(FileCollection)\nregisterResGeneratingTask is deprecated, use registerGeneratedResFolders(FileCollection)\nRunning Gradle task 'assembleDebug'...\nRunning Gradle task 'assembleDebug'...\n&gt; Connect to jcenter.bintray.com:443 [jcenter.bintray.com\/52.43.38.49, jcenter.bintray.com\/52.33.144.105] failed: Connection timed out: connect\nRunning Gradle task 'assembleDebug'...\nRunning Gradle task 'assembleDebug'...\n &gt; Failed to list versions for org.tensorflow:tensorflow-lite.\nRunning Gradle task 'assembleDebug'... \n&gt; Unable to load Maven meta-data from https:\/\/jcenter.bintray.com\/org\/tensorflow\/tensorflow-lite\/maven-metadata.xml.\nRunning Gradle task 'assembleDebug'...\n &gt; Could not HEAD 'https:\/\/jcenter.bintray.com\/org\/tensorflow\/tensorflow-lite\/maven-metadata.xml'.  \nRunning Gradle task 'assembleDebug'...\n* Try:\nRunning Gradle task 'assembleDebug'...\nRun with --stacktrace option to get the stack trace. Run with --info or --debug option to get more log output. Run with --scan to get full insights.\nRunning Gradle task 'assembleDebug'... \nRunning Gradle task 'assembleDebug'... Done                       305.3s (!)\nGradle task assembleDebug failed with exit code 1\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57485765,"AnswerCount":0,"Tags":"<javascript><tensorflow><tensorflow.js>","CreationDate":"2019-08-13T21:54:33.027","AcceptedAnswerId":null,"OwnerUserId":11924545.0,"Title":"How to make tfjs not ignore gpu device?","Body":"<p>I am running tfjs in a container from an official image of tensorflow-gpu. It logs to the console the following message:<\/p>\n\n<pre><code>Ignoring visible gpu device (device: 0, name: Tesla M60, pci bus id: 0000:00:1e.0, compute capability: 5.2) with Cuda compute capability 5.2. The minimum required Cuda capability is 6.0.\n<\/code><\/pre>\n\n<p>The host is an aws vm with a Tesla M60. How can i compile tfjs to run with cuda capability 5.2?\nThanks in advance<\/p>\n","answers":[]}
+{"QuestionId":57485790,"AnswerCount":0,"Tags":"<tensorflow><keras><neural-network><linear-regression>","CreationDate":"2019-08-13T21:57:07.347","AcceptedAnswerId":null,"OwnerUserId":11876924.0,"Title":"Understanding the computation graph of Tensorflow Keras layer","Body":"<p>I've created the simplest Dense layer with just 2 parameters (kernel and bias with 1 parameter each):<\/p>\n\n<pre><code>network1d = tf.keras.Sequential()\nlayer1d = tf.keras.layers.Dense(1, input_shape=[1], activation=None)\nnetwork1d.add(layer1d)\nnetwork1d.compile(loss='mean_squared_error', optimizer=tf.keras.optimizers.Adam(0.1), metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>And this is the graph drawn by Tensorboard:\n<a href=\"https:\/\/i.stack.imgur.com\/teXCR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/teXCR.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I don't think I'm understanding the major pieces - in particular, is my model (with the two parameters) the red \"dense\" block, or the one near the bottom of the graph? (The one in the Adam has red border because I clicked on it.)<\/p>\n","answers":[]}
+{"QuestionId":57485917,"AnswerCount":0,"Tags":"<tensorflow><lstm><cudnn>","CreationDate":"2019-08-13T22:12:13.587","AcceptedAnswerId":null,"OwnerUserId":11809734.0,"Title":"To use CudnnRNN in gradients, is_training must be set to True","Body":"<p>I am using tf.contrib.cudnn_rnn.CudnnLSTM in my code, and one of the parameters is \"is_training\" to LSTM cell. <\/p>\n\n<p>During training it is true, and I can use tf.gradient on this cell, but while testing, when \"is_training\" flag is false, I get the error : \"To use CudnnRNN in gradients, is_training must be set to True\".<\/p>\n\n<p>Looking in the documentation at \"<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/cudnn_rnn_grad.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/cudnn_rnn_grad.py<\/a>\", I find a few lines : <\/p>\n\n<pre><code>@ops.RegisterGradient(\"CudnnRNN\")\ndef _cudnn_rnn_backward(op, *grads):\n \"\"\"Gradients for the CudnnRNN op.\"\"\"\n  if not op.get_attr(\"is_training\"):\n    raise ValueError(\n        \"To use CudnnRNN in gradients, is_training must be set to True.\")\n<\/code><\/pre>\n\n<p>Any idea, what might be happening, I do not want to dive into the whole code because of time constraint. Thanks.<\/p>\n","answers":[]}
+{"QuestionId":57485961,"AnswerCount":1,"Tags":"<python><tensorflow><nan>","CreationDate":"2019-08-13T22:17:33.043","AcceptedAnswerId":null,"OwnerUserId":10229833.0,"Title":"Training cost becomes nan with TensorFlow","Body":"<p>I am practicing Neural Style Transfer and am facing an issue here with the costs. After few iterations, the costs become NaN.. What might be the issue here? Where should I try to focus on?<\/p>\n\n<pre><code>Iteration 0 :\ntotal cost = 1919359.6\ncontent cost = 125.083374\nstyle cost = 47952.72\nIteration 1 :\ntotal cost = nan\ncontent cost = nan\nstyle cost = nan\nIteration 2 :\ntotal cost = nan\ncontent cost = nan\nstyle cost = nan\nIteration 3 :\ntotal cost = nan\ncontent cost = nan\nstyle cost = nan\n<\/code><\/pre>\n\n<p>Generally I don't have a good plan how to figure out where the nan happens. Especially with TensorFlow, where I am not familiar with the debugging techniques.<\/p>\n\n<p>The code is here: <a href=\"https:\/\/github.com\/enyangxxx\/neuralStyleTransferMonet\" rel=\"nofollow noreferrer\">https:\/\/github.com\/enyangxxx\/neuralStyleTransferMonet<\/a>\nIt is the code from a Coursera course and I want to apply it on my own themes later :)<\/p>\n","answers":[{"AnswerId":"57486021","CreationDate":"2019-08-13T22:25:21.793","ParentId":null,"OwnerUserId":"11919358","Title":null,"Body":"<p>Might be a mishandled 0\/0 error. Especially the case if you have created your own functions. Most probably it might be cross entropy. I believe you are not training with keras and just tensorflow. May be check out this -><a href=\"https:\/\/stackoverflow.com\/questions\/33712178\/tensorflow-nan-bug\">Tensorflow NaN bug?<\/a><\/p>\n\n<p>Then again this might not be your error. So it would be useful if you post the code.<\/p>\n"}]}
+{"QuestionId":57486135,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-13T22:39:28.290","AcceptedAnswerId":57499179.0,"OwnerUserId":2603684.0,"Title":"How to iterate over tensors in Keras\/TensorFlow?","Body":"<p>I am trying to rewrite the following piece of Python code in Tensorflow. However, I am having trouble using the <code>tf.map_fn<\/code> for iterating over the tensor.<\/p>\n\n<p>Here <code>depth<\/code> is the tensor of shape <code>[batch_size,256,256]<\/code>, <code>normal<\/code> is the tensor of shape <code>[batch_size,256,256,3]<\/code> and <code>scale<\/code> is the tensor of shape <code>[batch_size,256,256]<\/code>:<\/p>\n\n<pre><code>for b in range(0,batch_size):\n    depth[b,:,:] = [scale[b,0,0] + (scale[b,0,1] - scale[b,0,0])* x for x in depth[b,:,:]]\n    normal[b,:,:,:] = [scale[b,0,2] + (scale[b,0,3] - scale[b,0,2])* x for x in normal[b,:,:,:]]\n<\/code><\/pre>\n\n<p>I am getting an error message saying:<\/p>\n\n<pre><code>Tensor objects are only iterable when eager execution is enabled. To iterate over this tensor use tf.map_fn.\n<\/code><\/pre>\n","answers":[{"AnswerId":"57499179","CreationDate":"2019-08-14T17:09:05.637","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>What you are doing could be simply written as matrix element-wise operations:<\/p>\n\n<pre><code>depth_new = scale[:,0:1,0:1] + (scale[:,0:1,1:2] - scale[:,0:1,0:1]) * depth\nnormal_new = scale[:,0:1,2:3] + (scale[:,0:1,3:4] - scale[:,0:1,2:3]) * normal\n<\/code><\/pre>\n\n<p>Note that we have used ranges for slices of size 1 (e.g. <code>0:1<\/code> instead of <code>0<\/code>, or <code>3:4<\/code> instead of <code>3<\/code>) in order to preserve the axes so that they could be broadcasted (see <a href=\"https:\/\/www.tensorflow.org\/xla\/broadcasting\" rel=\"nofollow noreferrer\">TensorFlow broadcasting guide<\/a> or Numpy broadcasting guide <a href=\"https:\/\/www.numpy.org\/devdocs\/user\/theory.broadcasting.html\" rel=\"nofollow noreferrer\">here<\/a> or <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/user\/basics.broadcasting.html\" rel=\"nofollow noreferrer\">here<\/a> for more information).<\/p>\n"}]}
+{"QuestionId":57486370,"AnswerCount":1,"Tags":"<computer-vision><conv-neural-network><pytorch><transfer-learning>","CreationDate":"2019-08-13T23:13:51.657","AcceptedAnswerId":null,"OwnerUserId":11924620.0,"Title":"Why did I get 2 different results from two models with same parameters and inputs?","Body":"<p>I loaded resnet18 into my two models (model1 and model2), with pretrained weights.<\/p>\n\n<p>I want to use them as feature extractors\nFor model1: I freezed the parameters except the last linear layer model1.fc, then train it. After training, I set model1.fc into torch.nn.Identity()\nFor model2: I directly set model2.fc into torch.nn.Identity()<\/p>\n\n<p>Then these 2 models should be the same, but I get different forward result from the same inputs.<\/p>\n\n<p>If the training of model1 is not done, they will have the same result, maybe something wrong with the freezing of  parameter.\nHowever, I checked their parameters after the training of model1 and setting the last layer of both models to identity layer, and they seems to be the same.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport torch\nimport torch.nn as nn\nimport torch.optim as optim\nfrom torchvision import datasets, transforms, models\n\n# Load weights pretrained on ImageNet\ndef load_weights(model):\n    model_dir = \".....\"\n    model.load_state_dict(tor`enter code here`ch.utils.model_zoo.load_url(\"https:\/\/download.pytorch.org\/models\/resnet18-5c106cde.pth\", model_dir=model_dir))\n    return model\n\nmodel1=models.resnet18()\nmodel1=load_weights(model1)\nfor param in model1.parameters():\n    param.requires_grad = False\nmodel1.fc=nn.Linear(512, 2)\nmodel1.cuda()\noptimizer = optim.SGD(model1.fc.parameters(), lr=1e-2, momentum=0.9)\nresult_freeze = \\\n    run_training(model1, optimizer, device, train_loader, val_loader,num_epochs=10)\n\nmodel2=models.resnet18()\nmodel2=load_weights(model2)\nmodel2.fc=nn.Identity()\nmodel2.cuda()\n\nmodel1.fc=nn.Identity()\nmodel1.cuda()\n\n\n\n# checking forward results(extracting features) \n\n# The batch size is one here\nfor batch_idx, (data, target) in enumerate(X_train): \n        data, target = data.to(device), target.to(device)\n        d=data\n        X_train_feature[batch_idx]=model1(data).cpu().detach().numpy()\n        y_train[batch_idx]=target.cpu().detach().numpy()\n        X_train2_feature[batch_idx]=model2(d).cpu().detach().numpy()\n        y_train2[batch_idx]=target.cpu().detach().numpy()\n        print(sum(X_train_feature[batch_idx]==X_train2_feature[batch_idx]))\n        print(sum(y_train[batch_idx]==y_train2[batch_idx]))\n        print(torch.sum(d==data))\nfor batch_idx, (data, target) in enumerate(X_test):\n        data, target = data.to(device), target.to(device)\n        d=data\n        X_test_feature[batch_idx]=model1(data).cpu().detach().numpy()\n        y_test[batch_idx]=target.cpu().detach().numpy()\n        X_test2_feature[batch_idx]=model2(d).cpu().detach().numpy()\n        y_test2[batch_idx]=target.cpu().detach().numpy()\n        print(sum(X_test_feature[batch_idx]==X_test2_feature[batch_idx]))\n        print(sum(y_test[batch_idx]==y_test2[batch_idx]))\n        print(torch.sum(d==data))\n\n# checking parameters\nfor a,b in zip(model1.parameters(),model2.parameters()):\n    print(torch.sum(a!=b)) \n<\/code><\/pre>\n\n<p>Expect to get the same forward results from model1 and model2, but they are different. And If they produce different forward results, why do they have exactly the same parameters?<\/p>\n","answers":[{"AnswerId":"57489008","CreationDate":"2019-08-14T06:12:54.977","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Have you taken into account changes that might occur to <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/nn.html#batchnorm2d\" rel=\"nofollow noreferrer\"><code>BatchNorm<\/code><\/a> layers?<br>\nBatch norm layers do not behave like normal layers - their internal parameters are modified by computing running mean and std of the data, and not by gradient descent.<br>\nTry setting <code>model1.eval()<\/code> before the finetune and then check.<\/p>\n"}]}
+{"QuestionId":57486468,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-13T23:29:21.590","AcceptedAnswerId":null,"OwnerUserId":11924674.0,"Title":"How to fix \" TypeError: 'function' object is not subscriptable \"?","Body":"<p>Rolling a custom estimator to explore RNN prediction in a time series, I'm unable to turn the 'features' argument of my model function into a sequence, for supply \nto:<\/p>\n\n<pre><code>tf.nn.static_rnn(\n    cell,\n    inputs,\n    initial_state=None,\n    dtype=None,\n    sequence_length=None,\n    scope=None\n)\n (2nd parameter).\n<\/code><\/pre>\n\n<p>I need the following to work in order to turn \"features\" (the dictionary returned from my input function) into a sequence of time steps:<\/p>\n\n<pre><code>x = tf.split(features['f1'], 8, 1)\n<\/code><\/pre>\n\n<p>This is my model fn abbreviated:<\/p>\n\n<pre><code>    def rnn_model_fn(features, labels, mode, params):  \n        # 0. Reformat input shape to become a sequence; i.e. break it into the series time steps\n        x   = tf.split(features['f1'], 8, 1)\n\n        # 1. configure the RNN\n        lstm_cell  = tf.keras.layers.LSTMCell(LSTM_SIZE)\n        outputs, _ = tf.contrib.rnn.static_rnn(lstm_cell, x, dtype=tf.float32)\n...\n<\/code><\/pre>\n\n<p>This is the stack trace:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"E:\\19\\IronCondorTrade\\ICondorIV.py\", line 124, in &lt;module&gt;\n    tf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)\n  File \"C:\\Users\\HP\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow_estimator\\python\\estimator\\training.py\", line 473, in train_and_evaluate\n    return executor.run()\n  File \"C:\\Users\\HP\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow_estimator\\python\\estimator\\training.py\", line 613, in run\n    return self.run_local()\n  File \"C:\\Users\\HP\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow_estimator\\python\\estimator\\training.py\", line 714, in run_local\n    saving_listeners=saving_listeners)\n  File \"C:\\Users\\HP\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow_estimator\\python\\estimator\\estimator.py\", line 367, in train\n    loss = self._train_model(input_fn, hooks, saving_listeners)\n  File \"C:\\Users\\HP\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow_estimator\\python\\estimator\\estimator.py\", line 1158, in _train_model\n    return self._train_model_default(input_fn, hooks, saving_listeners)\n  File \"C:\\Users\\HP\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow_estimator\\python\\estimator\\estimator.py\", line 1188, in _train_model_default\n    features, labels, ModeKeys.TRAIN, self.config)\n  File \"C:\\Users\\HP\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow_estimator\\python\\estimator\\estimator.py\", line 1146, in _call_model_fn\n    model_fn_results = self._model_fn(features=features, **kwargs)\n  File \"E:\\19\\IronCondorTrade\\ICondorIV.py\", line 18, in rnn_model_fn\n    x   = tf.split(features['f1'], 8, 1)\nTypeError: 'function' object is not subscriptable\n<\/code><\/pre>\n\n<p>The error occurs when I run the code (from inside IDLE). I expect it to proceed to building the RNN network but instead gets stuck with - TypeError: 'function' object is not subscriptable - . Tried various ways of getting at the features argument but they all return with the same type of message that what I'm thinking is a dictionary, is instead a function; mystifying enough to be my first SO post. Any guidance deeply appreciated.<\/p>\n","answers":[]}
+{"QuestionId":57486614,"AnswerCount":1,"Tags":"<tensorflow><keras><model>","CreationDate":"2019-08-13T23:56:21.953","AcceptedAnswerId":null,"OwnerUserId":9719331.0,"Title":"How do I fix this dimenion error in keras \/ tensorflow?","Body":"<p>I am trying to fit a model that predicts a 1 or a 0 for each row of the X array. X is an array of shape (1000,26) and Y is of shape (1000, 1).For whatever reason tf\/keras doesn't seem to like my inputs. <\/p>\n\n<p>So far I have tried changing the input type. It comes in the form of a pickle and then is turned into a numpy array by using the \"np.values\" function. Before I was using tf.constant(training) instad of X = training.values but all changing that did was give me a slightly different error (both included below)<\/p>\n\n<p>This is the full code minus the imports and loading of some other data files. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>training = pd.read_pickle(\".\/training.pkl\")\ntrain_labels = pd.read_pickle(\".\/train_labels.pkl\")\n\nX = training.values\nY = train_labels.values\nmodel = models.Sequential()\n\nmodel.add(layers.Dense(46, activation='relu', input_dim=26))\nmodel.add(layers.Dense(46, activation='relu'))\nmodel.add(layers.Dense(46, activation='relu'))\nmodel.add(layers.Dense(2, activation='softmax'))\n\nmodel.summary()\n\nmodel.compile(optimizer='adam',\n    loss='binary_crossentropy',\n    metrics=['binary_accuracy'])\n\n\nmodel.fit(X, Y, epochs=3, steps_per_epoch=100)\n<\/code><\/pre>\n\n<p>The current error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Error when checking target: expected dense_24 to have shape (2,) but got array with shape (1,)\n<\/code><\/pre>\n\n<p>Error message when using tf.constant <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Incompatible shapes: [1000] vs. [1000,2]\n<\/code><\/pre>\n","answers":[{"AnswerId":"57490264","CreationDate":"2019-08-14T07:45:03.550","ParentId":null,"OwnerUserId":"6622287","Title":null,"Body":"<p>I don't have enough reputation to write a comment so I write an answer here.<\/p>\n\n<p>As @Shubham Panchal has suggested, your last layer must match the shape of your Y. Therefore you either make sure that your Y has shape (1000, 1) and change your last layer to have 1 node, or you one-hot encode your Y (see <code>pandas.get_dummies<\/code>) so that it has shape (1000, 2).<\/p>\n\n<p>As we were discussing over cross-validated, you don't need to use Tensorflow constants\/placeholders\/variables here (I suggest to have a look at the differences between them and how Keras handles them). Feeding directly numpy arrays is sufficient since Keras takes care of defining the Tensorflow graph for you. <\/p>\n\n<p>Also, you convert your pandas dataframe to numpy array using <code>pandas.DataFrame.values<\/code>, however notice that in <a href=\"https:\/\/pandas.pydata.org\/pandas-docs\/stable\/reference\/api\/pandas.DataFrame.values.html\" rel=\"nofollow noreferrer\">its documentation<\/a> the suggested method is <code>pandas.DataFrame.to_numpy<\/code>.<\/p>\n"}]}
+{"QuestionId":57486671,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2019-08-14T00:05:08.353","AcceptedAnswerId":null,"OwnerUserId":11210924.0,"Title":"why loss is so high for 4 inputs 2 outputs function approximation using MLP?","Body":"<p>My code is very simple since my understanding is MLP can approximate any function:<\/p>\n\n<pre><code>def build_model():\n  model = tf.keras.Sequential([\n      tf.keras.layers.Dense(5, activation='tanh', input_shape=(4, ), name='a'),\n      tf.keras.layers.Dense(5, activation='tanh'),\n      tf.keras.layers.Dense(2, activation='sigmoid', name='b')])\n  optimizer = tf.keras.optimizers.RMSprop(lr=0.001, rho=0.9, epsilon=None, decay=0.0)  \n  model.compile(loss='mse',optimizer=optimizer)\n  return model\n\ndef train_benchmark_NN(x, y, epochs=10000):\n  model = build_model()\n  es = tf.keras.callbacks.EarlyStopping(monitor='val_loss', patience=20, verbose=0)\n  history = model.fit(x, y, batch_size = 1000, epochs=epochs, validation_split = 0.2, verbose=1, callbacks=[es])\n  return model, history\n<\/code><\/pre>\n\n<p>I tried different number of layers(like[256, 128, 64, 32]), nodes, optimizer, initializer, activation function. I also tried handle the two outputs separately instead of training one model for them together, but the result is bad too. Actually I don't have a decent judge on how heavy my model should be for data like this. I tried training model for some know function with same number of input and output, it's always very hard when the function like \n<code>y1=math.cos(x1)+math.cos(x2)+math.cos(x3)+math.cos(x4)<\/code>. <\/p>\n\n<p>Can anyone tell me, I should try much heavier model or I did something wrong in my code? or I have to preprocess the data differently? I only normalized it with zscore. Data size is ~6000 in total.<\/p>\n\n<p>Current results:<\/p>\n\n<pre><code>Epoch 62\/10000\n4936\/4936 [==============================] - 0s 4us\/sample - loss: 0.2711 - val_loss: 3.9427\nEpoch 63\/10000\n4936\/4936 [==============================] - 0s 4us\/sample - loss: 0.2686 - val_loss: 3.9444\nEpoch 64\/10000\n4936\/4936 [==============================] - 0s 3us\/sample - loss: 0.2661 - val_loss: 3.9457\n<\/code><\/pre>\n\n<p>If I change <code>validation_split<\/code> from 0.2 to 0.01,the result became very different:<\/p>\n\n<pre><code>6109\/6109 [==============================] - 0s 5us\/sample - loss: 0.3729 - val_loss: 0.0589\nEpoch 96\/10000\n6109\/6109 [==============================] - 0s 5us\/sample - loss: 0.3683 - val_loss: 0.0356\nEpoch 97\/10000\n6109\/6109 [==============================] - 0s 5us\/sample - loss: 0.3702 - val_loss: 0.0381\ni: 0 , err_mean: 2.383471436639142\n&lt;matplotlib.legend.Legend at 0x7fdbb329d7f0&gt;\n<\/code><\/pre>\n\n<p>Although the <code>val_loss<\/code> became much smaller, that probably because the validation size isn't big enough, because when I plot the errors, it still looks same.<\/p>\n\n<p>Some visualization of the relationships in my data: <\/p>\n\n<p>inputs are x1-car speed, x2-engine torque, x3-DOC temperature, x4-DPF temperature<\/p>\n\n<p>outputs are y1-tailpipe CO gas, y2-tailpipe HC gas. <\/p>\n\n<p>y1 against x1, x2, x3, x4 are shown below:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/1DPxj.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1DPxj.png\" alt=\"x1-y1\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/cu285.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/cu285.png\" alt=\"x2-y1\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/mUUY3.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/mUUY3.png\" alt=\"x3-y1\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/QXbDD.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/QXbDD.png\" alt=\"x4-y1\"><\/a><\/p>\n\n<p>Should this be function easy to approximate at all? Thanks!!!<\/p>\n\n<p>I plotted errors along targets, it seems the model didn't learn at all, because the errors is very correlated to the targets.\n<a href=\"https:\/\/i.stack.imgur.com\/negJ2.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/negJ2.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57486708,"AnswerCount":0,"Tags":"<tensorflow><keras><conv-neural-network><cascade>","CreationDate":"2019-08-14T00:12:24.250","AcceptedAnswerId":null,"OwnerUserId":9566491.0,"Title":"How to give a model's output as input to another using keras and train end-to end","Body":"<p>I use Keras functional, with TF 1.13 backend, and I have a main CNN which makes regression, and I want the predictions to be given in another network to refine the predictions.<\/p>\n\n<pre><code>def create_model():\n    inp =layers.Input(shape=(540, 540, 3))\n    .\n    .\n    .\n    .\n    .\n    .\n    model = tf.keras.Model(inputs=inp, outputs=preds)\n    model.compile(optimizer=keras.optimizers.SGD(0.0001), \n          loss=custom_loss)\nreturn model\n<\/code><\/pre>\n\n<p>and now I want to take \"preds\" and give it to another custom network to further process.\na) How I give as input to the 2nd network the \"preds\"\nb) How can I use model.fit for 2 different networks\nc) How I can use 2 loss functions, one for each network<\/p>\n\n<p>Thank you<\/p>\n","answers":[]}
+{"QuestionId":57486805,"AnswerCount":1,"Tags":"<python><tensorflow><keras><conv-neural-network>","CreationDate":"2019-08-14T00:30:15.243","AcceptedAnswerId":57489191.0,"OwnerUserId":11354575.0,"Title":"Can some use the Con2d in tensorflow or keras repeatedly with the same kernel parameters","Body":"<p>I just come across a strange idea that requires me to use the same convolutional layer repeatedly. BTW I am using <code>tensorflow 1.13.1<\/code> and <code>keras 2.2.4<\/code>.<\/p>\n\n<p>In usual case of <code>keras<\/code>, it just goes like, <\/p>\n\n<pre><code>x = Conv2D()(x)\nx = Conv2D()(x)\n...\n<\/code><\/pre>\n\n<p>But that means 2 different 2D convolutional layers are applied, and I want to use the same convolution kernels repeatedly.<\/p>\n\n<p>I found out <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\/repeat\" rel=\"nofollow noreferrer\"><code>tf.contrib.layers.repeat<\/code><\/a> which seems to answer my question. <\/p>\n\n<p>Is there some other ways to do the same things? Or is that not worth trying at all. Thanks!!!<\/p>\n","answers":[{"AnswerId":"57489191","CreationDate":"2019-08-14T06:28:10.983","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You need to use the Keras Functional API for this:<\/p>\n\n<pre><code>layer = Conv2D(...)\ninput = Input(...)\nx = input\n\nx = layer(x)\nx = layer(x)\n<\/code><\/pre>\n\n<p>This applies the exact same convolutional layer two times, with the same weights\/kernels.<\/p>\n"}]}
+{"QuestionId":57486898,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-08-14T00:48:44.920","AcceptedAnswerId":null,"OwnerUserId":5739619.0,"Title":"How to use flow_from_directory in Keras for multi-class semantic segmentation?","Body":"<p>Let's say I have 100 training grayscale images and 100 RGB training masks, each of size 512x512. I was able to one-hot encode the masks using <code>to_categorical<\/code> in Keras with the below<\/p>\n\n<pre><code>numclasses=3\nmasks_one_hot=to_categorical(maskArr,numclasses)\n<\/code><\/pre>\n\n<p>where <code>maskArr<\/code> is a 100x512x512x1, and <code>masks_one_hot<\/code> is 100x512x512x3. <\/p>\n\n<p>However, to use <code>ImageDataGenerator<\/code> and <code>flow_from_directory<\/code> using <code>trainGenerator<\/code> from <a href=\"https:\/\/github.com\/zhixuhao\/unet\/blob\/master\/data.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/zhixuhao\/unet\/blob\/master\/data.py<\/a>, I tried to save the one-hot encoded training images and then read them using <code>trainGenerator<\/code>. However, I noticed after using <code>imwrite<\/code> on them and then reading them with <code>imread<\/code>, they changed from one-hot encoded 512x512x3 to 512x512x3 RGB images. That is, instead of each channel having a value of 0 or 1, they now range from 0-255<\/p>\n\n<p>As a result, if I do:<\/p>\n\n<pre><code>myGenerator = trainGeneratorOneHot(20,'data\/membrane\/train','image','label',data_gen_args,save_to_dir = \"data\/membrane\/train\/aug\", flag_multi_class = True,\nnum_class = 3, target_size=(512,512,3))\n\nnum_batch=3\nfor i,batch in enumerate(myGenerator):\n    if(i &gt;= num_batch):\n        break\n<\/code><\/pre>\n\n<p>where <code>trainGeneratorOneHot<\/code> is below:<\/p>\n\n<pre><code>def trainGeneratorOneHot(batch_size,...class_mode=None, image_class_mode=None):\n\n    image_datagen = ImageDataGenerator(**aug_dict)\n    mask_datagen = ImageDataGenerator(**aug_dict)\n    image_generator = image_datagen.flow_from_directory(train_path,classes = [image_folder], class_mode = image_class_mode, color_mode = image_color_mode,target_size = target_size, ...)\n    mask_generator = mask_datagen.flow_from_directory(train_path, classes = [mask_folder], class_mode = class_mode, target_size = target_size,...)\n    train_generator = zip(image_generator, mask_generator)\n\n    for (img,mask) in train_generator:\n        img,mask = adjustDataOneHot(img,mask)\n        yield (img,mask)\n\ndef adjustDataOneHot(img,mask):\n    return (img,mask)\n<\/code><\/pre>\n\n<p>Then I get `ValueError: could not broadcast input array from shape (512,512,1) into shape (512,512,3,1)<\/p>\n\n<p>How can I fix this?<\/p>\n","answers":[{"AnswerId":"57500388","CreationDate":"2019-08-14T18:45:17.887","ParentId":null,"OwnerUserId":"8230016","Title":null,"Body":"<p>Was dealing with the same issue a few days ago. I found it essential to make my own data generator class to deal with taking in data from a dataframe, augmenting it, and then one-hot-encoding it before passing it to my model. I was never able to get the Keras <code>ImageDataGenerator<\/code> to work for semantic segmentation problems with multiple classes.<\/p>\n\n<p>Below is a data generator class in case it might help you out:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def one_hot_encoder(mask, num_classes = 8):\n\n    hot_mask = np.zeros(shape = mask.shape, dtype = 'uint8')\n\n    for _ in range(8):\n        temp = np.zeros(shape = mask.shape[0:2], dtype = 'uint8')\n        temp[mask[:, :, _] != 0] = 1\n        hot_mask[:, :, _] = temp\n\n    return hot_mask\n\n# Image data generator class\nclass DataGenerator(keras.utils.Sequence):\n\n    def __init__(self, dataframe, batch_size, n_classes = 8, augment = False):\n        self.dataframe = dataframe\n        self.batch_size = batch_size\n        self.n_classes = n_classes\n        self.augment = augment\n\n\n    # Steps per epoch    \n    def __len__(self):\n        return len(self.dataframe) \/\/ self.batch_size\n\n    # Shuffles and resets the index at the end of training epoch\n    def on_epoch_end(self):\n        self.dataframe = self.dataframe.reset_index(drop = True)\n\n\n    # Generates data, feeds to training\n    def __getitem__(self, index):\n\n        processed_images = []\n        processed_masks = []\n\n        for _ in range(self.batch_size):\n\n            the_image = io.imread(self.dataframe['Images'][index])\n            the_mask = io.imread(self.dataframe['Masks'][index]).astype('uint8');\n            one_hot_mask = one_hot_encoder(the_mask, 8)\n\n\n            if(self.augment):\n                # Resizing followed by some augmentations\n                processed_image = augs_for_images(image = the_image) \/ 255.0\n                processed_mask = augs_for_masks(image = one_hot_mask)\n\n\n            else:\n                # Still resizing but no augmentations   \n                processed_image = resize(image = the_image) \/ 255.0\n                processed_mask = resize(image = one_hot_mask)\n\n            processed_images.append(processed_image)\n            processed_masks.append(processed_mask)\n\n\n        batch_x = np.array( processed_images )\n        batch_y = np.array( processed_masks )\n\n        return (batch_x, batch_y)\n<\/code><\/pre>\n\n<p>Also, here's a link to a repo with some semantic segmentation models that might be of interest to you. The <a href=\"https:\/\/github.com\/qubvel\/segmentation_models\/blob\/master\/examples\/multiclass%20segmentation%20(camvid).ipynb\" rel=\"nofollow noreferrer\">notebook<\/a> itself shows how the author dealt with multi-class semantic segmentation.<\/p>\n"}]}
+{"QuestionId":57486938,"AnswerCount":0,"Tags":"<python><numpy><keras><tokenize>","CreationDate":"2019-08-14T00:57:14.183","AcceptedAnswerId":null,"OwnerUserId":9613022.0,"Title":"Creating Tokens For RNN: \"IndexError: too many indices for array\"","Body":"<p>I am working on some machine learning projects and found a text RNN tutorial (will link if I can find it again) that I would like to modify but I cannot seem to find where my code goes wrong.<\/p>\n\n<p>I am able to replicate the problem using this short script:<\/p>\n\n<pre><code>import string\n\nfrom numpy import array\nfrom keras.preprocessing.text import Tokenizer\n\ndef load_doc(path):\n\n    file = open(path, 'r')\n    text = file.read()\n    file.close()\n    return text\n\ndef get_tokens(text):\n\n    tokens = text.split()\n\n    # I would like to remove the next 4 lines\n    table = str.maketrans('', '', string.punctuation)\n    tokens = [w.translate(table) for w in tokens]\n    tokens = [word for word in tokens if word.isalpha()]\n    tokens = [word.lower() for word in tokens]\n\n    return tokens\n\ndef get_sequences(sequence_length, token_length):\n    sequences = list()\n    for i in range(sequence_length, token_length):\n        seq = tokens[i - sequence_length: i]\n        line = ' '.join(seq)\n        sequences.append(line)\n    return sequences\n\ndata = load_doc(\"C:\\data.txt\")  \ntokens = get_tokens(data)\nsequences = get_sequences(10, len(tokens))\n\ntok = Tokenizer()\ntok.fit_on_texts(sequences)\nencoded = tok.texts_to_sequences(sequences)\n\n#sanity check\nprint(\"Prenumpy: \\n\" + str(encoded[:10]))\nencoded = array(encoded)\nprint(encoded.shape)\nprint(\"Postnumpy: \\n\" + str(encoded[:10]))\nX = encoded[:,:-1]\n<\/code><\/pre>\n\n<p>When I remove the lines from <code>get_tokens(text)<\/code>, I get an error: <code>IndexError: too many indices for array<\/code>. Comparing the shapes, the working code gets: <code>(11984,10)<\/code> and the failing code gets: <code>(12388,)<\/code>. I understand that with the working code, the first dimension will be smaller since I'll have less tokens. Another important difference is the output pre and post casting <code>encoded<\/code> as a numpy array.<\/p>\n\n<p>Pre-cast (shortened for space, numbers are different between each run because different tokens are used):<\/p>\n\n<pre><code>[[2, 1318, 3, 161, 203, 9, 17, 297, 11, 21], ..., [21, 8, 325, 11, 12, 212, 48, 9, 235, 577]]\n<\/code><\/pre>\n\n<p>Working post-cast (shortened):<\/p>\n\n<pre><code>[[   2 1285    3  139  181    9   17  275   11   21]\n [1285    3  139  181    9   17  275   11   21    8]\n ...\n [  11   21    8  303   11   12  190   46    9  213]\n [  21    8  303   11   12  190   46    9  213  554]]\n<\/code><\/pre>\n\n<p>Failing post-cast (shortened):<\/p>\n\n<pre><code>[list([2, 1318, 3, 161, 203, 9, 17, 297, 11, 21])\n list([1318, 3, 161, 203, 9, 17, 297, 11, 21, 8])\n ...\n list([11, 21, 8, 325, 11, 12, 212, 48, 9, 235])\n list([21, 8, 325, 11, 12, 212, 48, 9, 235, 577])]\n<\/code><\/pre>\n\n<p>I am assuming the reason why I'm getting the error is because of the way numpy is converting the the arrays differently. Why is numpy changing the type of the data?<\/p>\n","answers":[]}
+{"QuestionId":57487026,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-14T01:15:09.883","AcceptedAnswerId":57722106.0,"OwnerUserId":11924886.0,"Title":"Keras Squential Returning Table Not Initialized Error","Body":"<p>I am trying to compare cosine similarities and euclidean distances of different pairs of sentence vectors, embedded by some text embedding module provided from tensorflow hub. I made a Keras Sequential model, and added the embedding layer to it, so that the 'prediction' or 'evaluation' of input texts would be their embedded vectors. <\/p>\n\n<p>The exact same code worked fine two days ago, but it started to return \"Failed precondition: Table not initialized.\" error when calling 'predict' on vectorizor. When it worked, I didn't even set \"steps=1\" inside predict but it worked fine. Now, I had to because with it the code returns \"ValueError: When using data tensors as input to a model, you should specify the <code>steps<\/code> argument.\"<\/p>\n\n<p>Why would the code that worked well two days ago suddenly started to return errors? <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nfrom tensorflow.keras import Sequential\nfrom tensorflow.keras.layers import Embedding, Dense, Dot, Input\nimport tensorflow_hub as hub\n\n\nvectorizer = Sequential()\n\nembed_layer = hub.KerasLayer(\"https:\/\/tfhub.dev\/google\/tf2-preview\/nnlm-ko-dim128-with-normalization\/1\", output_shape=[128],\n                             input_shape=[], dtype=tf.string)\n\nvectorizer.add(embed_layer)\n\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nfrom sklearn.metrics.pairwise import cosine_similarity\nfrom sklearn.metrics.pairwise import euclidean_distances\n\n\nstring = \"I like you\"\nstr_tensor = tf.constant(string, shape = [1], dtype=tf.string)\n\nstring2 = \"I hate you\"\nstr_tensor2 = tf.constant(string2, shape = [1], dtype=tf.string)\n\nprint(vectorizer)\n\n\na = vectorizer.predict(str_tensor, steps=1)\nb = vectorizer.predict(str_tensor2)\n\ncos_sim = cosine_similarity(a, b)\neuc_dis = euclidean_distances(a, b)\n\nprint(cos_sim)\nprint(euc_dis)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57722106","CreationDate":"2019-08-30T07:08:47.257","ParentId":null,"OwnerUserId":"11924886","Title":null,"Body":"<p>Upgraded to tensorflow 2.0 and worked fine...<\/p>\n"}]}
+{"QuestionId":57487030,"AnswerCount":1,"Tags":"<python><keras><deep-learning><computer-vision>","CreationDate":"2019-08-14T01:15:52.633","AcceptedAnswerId":null,"OwnerUserId":11919783.0,"Title":"Expected input_5 to have 4 dimensions, but got array with shape (24, 1)","Body":"<p>Due to my previous <a href=\"https:\/\/stackoverflow.com\/questions\/57469382\/cannot-append-big-dataset-of-images-to-list\">preprocessing problem<\/a>, I decided to have a new method to preprocessing dataset:<\/p>\n\n<pre><code>\nimport os.path\nimport os\nimport tqdm\nimport glob\nfrom imutils import paths\nfrom sklearn.preprocessing import LabelEncoder\n\ndef getListOfFiles(dirName):\n    listOfFile = os.listdir(dirName)\n    listOfFile.sort()\n    allFiles = list()\n    for entry in listOfFile:\n        fullPath = os.path.join(dirName, entry)\n        if os.path.isdir(fullPath):\n            allFiles = allFiles + getListOfFiles(fullPath)\n        else:\n            allFiles.append(fullPath)             \n    return allFiles        \n\ndef main():  \n    dirName = '\/content\/drive\/My Drive\/casia-299-small'\n    save_path = '\/content\/drive\/My Drive\/face_recognition\/inceptionv2'\n    listOfFiles = getListOfFiles(dirName)\n    completeName = os.path.join(save_path, \"to_train\"+\".txt\")\n    completeName2 = os.path.join(save_path, \"to_test\"+\".txt\")\n\n    file_to_write = open(completeName, \"w\")\n    for file in listOfFiles: \n        string_name = file\n        file_to_write.write(string_name)\n        file_to_write.write('\\n')\n    file_to_write.close()\n\n\n    file_to_write = open(completeName2, \"w\")\n    for file in listOfFiles:\n        label = file.split(os.path.sep)[-2]\n        file_to_write.write(label)\n        file_to_write.write('\\n')\n    file_to_write.close()\n\n\nif __name__ == '__main__':\n    main()\n<\/code><\/pre>\n\n<p>Then I use to file <code>to_train.txt<\/code> and <code>to_test.txt<\/code> as two arrays in size of (number of images, 1), represent for paths to each image in dataset and labels of each image. I feed them to my CNN for training like this: <\/p>\n\n<pre><code># Model 2 try (my)\nfrom sklearn.linear_model import LogisticRegression\nfrom sklearn.model_selection import GridSearchCV\nfrom sklearn.metrics import classification_report\nfrom imutils import paths\nfrom keras.applications.inception_resnet_v2 import InceptionResNetV2\nfrom keras.applications import imagenet_utils\nfrom keras.preprocessing.image import img_to_array\nfrom keras.preprocessing.image import load_img\nfrom sklearn.preprocessing import LabelEncoder\nfrom sklearn.preprocessing import LabelBinarizer\nfrom sklearn.model_selection import train_test_split\nfrom keras.preprocessing.image import ImageDataGenerator\nfrom keras.optimizers import SGD\nfrom keras.optimizers import RMSprop\nfrom keras.layers import Input\nfrom keras.models import Model\nfrom keras.layers.core import Dense\nfrom keras.layers.core import Dropout\nfrom keras.layers.core import Flatten \nimport numpy as np \nimport random \nimport os\nfrom keras.applications.inception_resnet_v2 import InceptionResNetV2\nimport numpy as np \nimport random \nimport os\nimport cv2\nimport pickle \nfrom keras import backend as K \nimport tensorflow as tf\nfrom keras.models import Sequential, Model\nfrom keras.layers import *\nfrom keras.optimizers import *\nfrom keras import applications\nimport h5py\n\n\nimages = []\nlabels = []\n\nimage_path = \"\/content\/drive\/My Drive\/face_recognition\/inceptionv2\/to_train.txt\"\nlabel_path = \"\/content\/drive\/My Drive\/face_recognition\/inceptionv2\/to_test.txt\"\n\nwith open(image_path) as fp:\n    line = fp.readline()\n    while line:\n        line = fp.readline()\n        my_string = line.replace(\"\\n\", \"\")\n        images.append(my_string)\n\nwith open(label_path) as fp:\n    line = fp.readline()\n    while line:\n        line = fp.readline()\n        my_string = line.replace(\"\\n\", \"\")\n        labels.append(my_string)\n\nprint(\"Images: \", images)\nprint(\"Labels: \", labels)\n\n\nx_train, x_test, y_train, y_test = train_test_split(images, labels, test_size = 0.2, random_state = 42)\n\nprint(np.shape(x_train), np.shape(x_test), np.shape(y_train), np.shape(y_test))\n\ndef convnet_model_():\n    initial_inceptionv2_model = InceptionResNetV2(weights=None, include_top = False, input_shape = (299, 299, 3))\n    x = initial_inceptionv2_model.output\n    x = GlobalAveragePooling2D()(x)\n    x = Dense(4096, activation = 'relu')(x)\n    x = Dropout(0.6)(x)\n    x = Dense(4096, activation = 'relu')(x)\n    x = Dropout(0.6)(x)\n    x = Lambda(lambda x: K.l2_normalize(x, axis = 1))(x)\n    convnet_model = Model(inputs=initial_inceptionv2_model.input, outputs = x)\n    return convnet_model\n\ndef deep_rank_model():\n    convnet_model = convnet_model_()\n\n    first_input = Input(shape = (299, 299, 3))\n    first_conv = Conv2D(96, kernel_size = (8, 8), strides = (16, 16), padding = 'same')(first_input)\n    first_max = MaxPool2D(pool_size = (3, 3), strides = (2, 2), padding = 'same')(first_conv)\n    first_max = Flatten()(first_max)\n    first_max = Lambda(lambda x: K.l2_normalize(x, axis =1))(first_max)\n\n    second_input = Input(shape = (299, 299, 3))\n    second_conv = Conv2D(96, kernel_size = (8, 8), strides = (32, 32), padding = 'same')(second_input)\n    second_max = MaxPool2D(pool_size = (7, 7), strides = (4, 4), padding = 'same')(second_conv)\n    second_max = Flatten()(second_max)\n    second_max = Lambda(lambda x: K.l2_normalize(x, axis = 1))(second_max)\n\n    merge_one = concatenate([first_max, second_max])\n    merge_two = concatenate([merge_one, convnet_model.output])\n    emb = Dense(4096)(merge_two)\n    emb = Dense(128)(emb)\n    l2_norm_final = Lambda(lambda x: K.l2_normalize(x, axis = 1))(emb)\n\n    final_model = Model(inputs = [first_input, second_input, convnet_model.input], outputs = l2_norm_final)\n    return final_model\n\ndeep_rank_model = deep_rank_model()\ndeep_rank_model.summary()\n\nbatch_size = 24\n_EPSILON = K.epsilon()\ndef _loss_tensor(y_true, y_pred):\n    y_pred = K.clip(y_pred, _EPSILON, 1.0 - _EPSILON)\n    loss = 0.\n    g = 1. \n    for i in range(0, batch_size, 3):\n        try:\n            q_embedding = y_pred[i]\n            p_embedding = y_pred[i+1]\n            n_embedding = y_pred[i+2]\n            D_q_p = K.sqrt(K.sum((q_embedding - p_embedding)**2))\n            D_q_n = K.sqrt(K.sum((q_embedding - n_embedding)**2))\n            loss = loss + g + D_q_p - D_q_n\n\n        except:\n            continue\n\n    loss = loss\/batch_size*3\n    return K.maximum(loss, 0)\n\nprint(\"Compile: \")\ndeep_rank_model.compile(loss=_loss_tensor, optimizer = SGD(lr = 0.001, momentum=0.9, nesterov=True))\n\ndef image_batch_generator(images, labels, batch_size):\n    labels = np.array(labels)\n    while True:\n        batch_paths = np.random.choice(a = len(images), size = batch_size\/\/3)\n        input_1 = []\n\n        for i in batch_paths:\n            pos = np.where(labels == labels[i])[0]\n            neg = np.where(labels != labels[i])[0]\n\n            j = np.random.choice(pos)\n            while j == i:\n                j = np.random.choice(pos)\n\n            k = np.random.choice(neg)\n            while(k == i):\n                k = np.random.choice(neg)\n\n            input_1.append(cv2.imread(images[i]))\n            input_1.append(cv2.imread(images[j]))\n            input_1.append(cv2.imread(images[k]))\n\n        input_1 = np.array(input_1)\n        input = [input_1, input_1, input_1]\n        print(\"Input shape: \": , np.shape(input))\n        yield (input, np.zeros((batch_size, )))\n\nfrom keras.callbacks import ModelCheckpoint\n\nfilePath = '\/content\/drive\/My Drive\/casia-299-small\/model_2_try_2.h5'\ncheckpoint = ModelCheckpoint(filePath, monitor = 'loss', verbose = 1, save_weights_only=False, save_best_only=True, mode = 'min')\ncallbacks_list = [checkpoint]\n\ndeep_rank_model.fit_generator(generator=image_batch_generator(x_train, y_train, batch_size), \n                              steps_per_epoch = len(x_train)\/\/batch_size, \n                              epochs = 20, \n                              verbose = 1, \n                              callbacks = callbacks_list)\n<\/code><\/pre>\n\n<p>Then during the training process, I faced with this new problem:\n<code>ValueError: Error when checking input: expected input_5 to have 4 dimensions, but got array with shape (24, 1)<\/code><\/p>\n\n<p>Is there something wrong with my code? I'm really not sure. Anybody with help is a big appreciation.<\/p>\n\n<p>Update:\nDuring the training process, I have just noticed that not all output get the error: i used <code>print(\"Shape of input: \", np.shape(input))<\/code> in function <code>image_batch_generator<\/code><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Njf2G.png\" rel=\"nofollow noreferrer\">Error screenshot<\/a><\/p>\n\n<p>So I guess there's something wrorng with <code>cv2.imread()<\/code>, which is not always read images successfully. I wonder if anyone has caught this problem yet? I'm so confused right now?<\/p>\n","answers":[{"AnswerId":"57487059","CreationDate":"2019-08-14T01:22:44.423","ParentId":null,"OwnerUserId":"10023182","Title":null,"Body":"<p>Conv2D which is used in your script requires an input shape with four dimensions, which means that you need to use <code>numpy.reshape()<\/code> to change the array with dimensions (24,1) into something like (1,24,1,3) (3 meaning RGB images). Here's the link to the documentation for numpy.reshape().<\/p>\n\n<p><a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.reshape.html\" rel=\"nofollow noreferrer\">https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.reshape.html<\/a><\/p>\n"}]}
+{"QuestionId":57487087,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><reinforcement-learning><q-learning>","CreationDate":"2019-08-14T01:28:31.610","AcceptedAnswerId":null,"OwnerUserId":3876796.0,"Title":"C51 reinforcement learning algorithm extremely slow","Body":"<p>I am applying reinforcement learning on a time series prediction problem. Until now, I have implemented a dueling DDQN algorithm with LSTM which seems to give some pretty good results, though sometimes slow to converge depending on the exact problem.\nI have then used C51 distributional reinforcement learning to compare the performance (which I expect to lead to better good results).<\/p>\n\n<p>I have slightly adapted the C51 google code <a href=\"https:\/\/github.com\/google\/dopamine\" rel=\"nofollow noreferrer\">dopamine<\/a> to be integrated into my code (network and training part). I am also using double Q learning for selection of next state action (which original code does not use). But, the problem is that it is really very very slow to execute. For comparaison, my previous dueling DDQN used to take 3.5 hours to train for 50000 episodes, however C51 algorithm has now taken almost 10 hours and yet has only reached 3000 episodes.<\/p>\n\n<p>I am wondering if there is something wrong with my adaptation of the code, or if C51 algorithm is really that slow. I am using an NVidia Geforce RTX 2080 Ti.<\/p>\n\n<p>Here is the network part:<\/p>\n\n<pre><code>#network part\nself.weights_initializer = tf.contrib.slim.variance_scaling_initializer(factor=1.0 \/ np.sqrt(3.0), mode='FAN_IN', uniform=True)\nself.net = tf.contrib.slim.fully_connected(\n  self.rnn, # output of an LSTM\n  num_actions * num_atoms,\n  activation_fn=None,\n  weights_initializer=self.weights_initializer)\n\nself.logits = tf.reshape(self.net, [-1, num_actions, num_atoms])\nself.probabilities = tf.contrib.layers.softmax(self.logits)\nself.q_values = tf.reduce_sum(self._support * self.probabilities, axis=2)\n\nself.predict = tf.argmax(self.q_values,1)\n\nself.actions = tf.placeholder(shape=[None],dtype=tf.int32)    \n\nself.target_distribution = tf.placeholder(shape=[None,num_atoms],dtype=tf.float32)\n\n# size of indices: batch_size x 1.\nself.indices = tf.range(tf.shape(self.logits)[0])[:, None]\n# size of reshaped_actions: batch_size x 2.\nself.reshaped_actions = tf.concat([self.indices, self.actions[:, None]], 1)\n# For each element of the batch, fetch the logits for its selected action.\nself.chosen_action_logits = tf.gather_nd(self.logits,\n                                self.reshaped_actions)\n\nself.td_error = tf.nn.softmax_cross_entropy_with_logits(labels=self.target_distribution,logits=self.chosen_action_logits)\n\n\n# divide by the real length of episodes instead of averaging which is incorrect\nself.loss = tf.cast(tf.reduce_sum(self.td_error), tf.float64) \/ tf.cast(tf.reduce_sum(self.seq_len), tf.float64)\n\nif apply_grad_clipping:\n   # calculate gradients and clip them to handle outliers\n   tvars = tf.trainable_variables()\n   grads, _ = tf.clip_by_global_norm(tf.gradients(self.loss, tvars), grad_clipping)\n   self.updateModel = optimizer.apply_gradients(\n        zip(grads, tvars),\n        name=\"updateModel\")\nelse:\n   self.updateModel = optimizer.minimize(self.loss, name=\"updateModel\")\n<\/code><\/pre>\n\n<p>And here's the training part:<\/p>\n\n<pre><code># training part\nif i &gt;= pre_train_episodes:\n        #Reset the lstm's hidden state\n        state_train = np.zeros((num_layers, 2, batch_size, h_size))\n        #Get a random batch of experiences.\n        trainBatch = myBuffer.sample(batch_size)\n        #Below we perform the Double-DQN update to the target Q-values\n        num_samples = batch_size*trace_length\n        # size of rewards: batch_size x 1\n        rewards = trainBatch[:,2][:, None]\n\n        # size of tiled_support: batch_size x num_atoms\n        tiled_support = tf.tile(mainQN._support, [num_samples])\n        tiled_support = tf.reshape(tiled_support, [num_samples, num_atoms])\n\n        # size of target_support: batch_size x num_atoms\n        is_terminal_multiplier = -(np.array(trainBatch[:,4]) - 1)\n        # Incorporate terminal state to discount factor.\n        # size of gamma_with_terminal: batch_size x 1\n        gamma_with_terminal = gamma * is_terminal_multiplier\n        gamma_with_terminal = gamma_with_terminal[:, None]\n\n        target_support = rewards + gamma_with_terminal * tiled_support\n\n\n        next_qt_argmax = sess.run([mainQN.predict], feed_dict={\\\n                            mainQN.scalarInput:np.vstack(trainBatch[:,3]),\\\n                            mainQN.trainLength:trace_length,mainQN.state_in:state_train,mainQN.batch_size:batch_size})\n        next_qt_argmax = np.reshape(next_qt_argmax, [-1, 1])\n        probabilities = sess.run(targetQN.probabilities, feed_dict={\\\n                            targetQN.scalarInput:np.vstack(trainBatch[:,3]),\\\n                            targetQN.trainLength:trace_length,targetQN.state_in:state_train,targetQN.batch_size:batch_size})\n        batch_indices = np.arange(num_samples)[:, None]\n        batch_indexed_next_qt_argmax = np.concatenate([batch_indices, next_qt_argmax], axis=1)\n\n\n        # size of next_probabilities: batch_size x num_atoms\n        next_probabilities = tf.gather_nd(probabilities, batch_indexed_next_qt_argmax)\n\n\n        target_distribution = project_distribution(target_support, next_probabilities, mainQN._support)\n        target_distribution = target_distribution.eval()\n\n        loss, _, _ = sess.run([mainQN.loss, mainQN.check_ops, mainQN.updateModel], \\\n                            feed_dict={mainQN.scalarInput:np.vstack(trainBatch[:,0]),mainQN.target_distribution:target_distribution,\\\n                            mainQN.actions:trainBatch[:,1],mainQN.trainLength:trace_length,\\\n                            mainQN.state_in:state_train,mainQN.batch_size:batch_size})\n\n        # perform soft\/hard update frequently\n        if i % update_target_freq == 0 or update_target_freq == 1 or softUpdate == True:\n            updateTarget(targetOps,sess)\n<\/code><\/pre>\n\n<p>auxiliary function:<\/p>\n\n<pre><code># function used above to project the distribution on the provided support\ndef project_distribution(supports, weights, target_support,\n                     validate_args=False):\n\"\"\"Projects a batch of (support, weights) onto target_support.\nBased on equation (7) in (Bellemare et al., 2017):\nhttps:\/\/arxiv.org\/abs\/1707.06887\nIn the rest of the comments we will refer to this equation simply as Eq7.\nThis code is not easy to digest, so we will use a running example to  clarify\nwhat is going on, with the following sample inputs:\n* supports =       [[0, 2, 4, 6, 8],\n                    [1, 3, 4, 5, 6]]\n* weights =        [[0.1, 0.6, 0.1, 0.1, 0.1],\n                    [0.1, 0.2, 0.5, 0.1, 0.1]]\n* target_support = [4, 5, 6, 7, 8]\nIn the code below, comments preceded with 'Ex:' will be referencing the above\nvalues.\nArgs:\nsupports: Tensor of shape (batch_size, num_dims) defining supports for the\n  distribution.\nweights: Tensor of shape (batch_size, num_dims) defining weights on the\n  original support points. Although for the CategoricalDQN agent these\n  weights are probabilities, it is not required that they are.\ntarget_support: Tensor of shape (num_dims) defining support of the projected\n  distribution. The values must be monotonically increasing. Vmin and Vmax\n  will be inferred from the first and last elements of this tensor,\n  respectively. The values in this tensor must be equally spaced.\n  validate_args: Whether we will verify the contents of the\n  target_support parameter.\n  Returns:\n     A Tensor of shape (batch_size, num_dims) with the projection of a batch of\n(support, weights) onto target_support.\n  Raises:\n    ValueError: If target_support has no dimensions, or if shapes of supports,\n  weights, and target_support are incompatible.\n \"\"\"\ntarget_support_deltas = target_support[1:] - target_support[:-1]\n# delta_z = `\\Delta z` in Eq7.\ndelta_z = target_support_deltas[0]\nvalidate_deps = []\nsupports.shape.assert_is_compatible_with(weights.shape)\nsupports[0].shape.assert_is_compatible_with(target_support.shape)\ntarget_support.shape.assert_has_rank(1)\nif validate_args:\n# Assert that supports and weights have the same shapes.\nvalidate_deps.append(\n    tf.Assert(\n        tf.reduce_all(tf.equal(tf.shape(supports), tf.shape(weights))),\n        [supports, weights]))\n# Assert that elements of supports and target_support have the same shape.\nvalidate_deps.append(\n    tf.Assert(\n        tf.reduce_all(\n            tf.equal(tf.shape(supports)[1], tf.shape(target_support))),\n        [supports, target_support]))\n# Assert that target_support has a single dimension.\nvalidate_deps.append(\n    tf.Assert(\n        tf.equal(tf.size(tf.shape(target_support)), 1), [target_support]))\n# Assert that the target_support is monotonically increasing.\nvalidate_deps.append(\n    tf.Assert(tf.reduce_all(target_support_deltas &gt; 0), [target_support]))\n# Assert that the values in target_support are equally spaced.\nvalidate_deps.append(\n    tf.Assert(\n        tf.reduce_all(tf.equal(target_support_deltas, delta_z)),\n        [target_support]))\n\nwith tf.control_dependencies(validate_deps):\n# Ex: `v_min, v_max = 4, 8`.\nv_min, v_max = target_support[0], target_support[-1]\n# Ex: `batch_size = 2`.\nbatch_size = tf.shape(supports)[0]\n# `N` in Eq7.\n# Ex: `num_dims = 5`.\nnum_dims = tf.shape(target_support)[0]\n# clipped_support = `[\\hat{T}_{z_j}]^{V_max}_{V_min}` in Eq7.\n# Ex: `clipped_support = [[[ 4.  4.  4.  6.  8.]]\n#                         [[ 4.  4.  4.  5.  6.]]]`.\nclipped_support = tf.clip_by_value(supports, v_min, v_max)[:, None, :]\n# Ex: `tiled_support = [[[[ 4.  4.  4.  6.  8.]\n#                         [ 4.  4.  4.  6.  8.]\n#                         [ 4.  4.  4.  6.  8.]\n#                         [ 4.  4.  4.  6.  8.]\n#                         [ 4.  4.  4.  6.  8.]]\n#                        [[ 4.  4.  4.  5.  6.]\n#                         [ 4.  4.  4.  5.  6.]\n#                         [ 4.  4.  4.  5.  6.]\n#                         [ 4.  4.  4.  5.  6.]\n#                         [ 4.  4.  4.  5.  6.]]]]`.\ntiled_support = tf.tile([clipped_support], [1, 1, num_dims, 1])\n# Ex: `reshaped_target_support = [[[ 4.]\n#                                  [ 5.]\n#                                  [ 6.]\n#                                  [ 7.]\n#                                  [ 8.]]\n#                                 [[ 4.]\n#                                  [ 5.]\n#                                  [ 6.]\n#                                  [ 7.]\n#                                  [ 8.]]]`.\nreshaped_target_support = tf.tile(target_support[:, None], [batch_size, 1])\nreshaped_target_support = tf.reshape(reshaped_target_support,\n                                     [batch_size, num_dims, 1])\n# numerator = `|clipped_support - z_i|` in Eq7.\n# Ex: `numerator = [[[[ 0.  0.  0.  2.  4.]\n#                     [ 1.  1.  1.  1.  3.]\n#                     [ 2.  2.  2.  0.  2.]\n#                     [ 3.  3.  3.  1.  1.]\n#                     [ 4.  4.  4.  2.  0.]]\n#                    [[ 0.  0.  0.  1.  2.]\n#                     [ 1.  1.  1.  0.  1.]\n#                     [ 2.  2.  2.  1.  0.]\n#                     [ 3.  3.  3.  2.  1.]\n#                     [ 4.  4.  4.  3.  2.]]]]`.\nnumerator = tf.abs(tiled_support - reshaped_target_support)\nquotient = 1 - (numerator \/ delta_z)\n# clipped_quotient = `[1 - numerator \/ (\\Delta z)]_0^1` in Eq7.\n# Ex: `clipped_quotient = [[[[ 1.  1.  1.  0.  0.]\n#                            [ 0.  0.  0.  0.  0.]\n#                            [ 0.  0.  0.  1.  0.]\n#                            [ 0.  0.  0.  0.  0.]\n#                            [ 0.  0.  0.  0.  1.]]\n#                           [[ 1.  1.  1.  0.  0.]\n#                            [ 0.  0.  0.  1.  0.]\n#                            [ 0.  0.  0.  0.  1.]\n#                            [ 0.  0.  0.  0.  0.]\n#                            [ 0.  0.  0.  0.  0.]]]]`.\nclipped_quotient = tf.clip_by_value(quotient, 0, 1)\n# Ex: `weights = [[ 0.1  0.6  0.1  0.1  0.1]\n#                 [ 0.1  0.2  0.5  0.1  0.1]]`.\nweights = weights[:, None, :]\n# inner_prod = `\\sum_{j=0}^{N-1} clipped_quotient * p_j(x', \\pi(x'))`\n# in Eq7.\n# Ex: `inner_prod = [[[[ 0.1  0.6  0.1  0.  0. ]\n#                      [ 0.   0.   0.   0.  0. ]\n#                      [ 0.   0.   0.   0.1 0. ]\n#                      [ 0.   0.   0.   0.  0. ]\n#                      [ 0.   0.   0.   0.  0.1]]\n#                     [[ 0.1  0.2  0.5  0.  0. ]\n#                      [ 0.   0.   0.   0.1 0. ]\n#                      [ 0.   0.   0.   0.  0.1]\n#                      [ 0.   0.   0.   0.  0. ]\n#                      [ 0.   0.   0.   0.  0. ]]]]`.\ninner_prod = clipped_quotient * weights\n# Ex: `projection = [[ 0.8 0.0 0.1 0.0 0.1]\n#                    [ 0.8 0.1 0.1 0.0 0.0]]`.\nprojection = tf.reduce_sum(inner_prod, 3)\nprojection = tf.reshape(projection, [batch_size, num_dims])\nreturn projection\n<\/code><\/pre>\n\n<p>Thank you in advance!<\/p>\n","answers":[]}
+{"QuestionId":57487327,"AnswerCount":1,"Tags":"<python><tensorflow><2d><interpolation>","CreationDate":"2019-08-14T02:08:21.020","AcceptedAnswerId":null,"OwnerUserId":2603684.0,"Title":"How to perform 2d interpolation completely in tensorflow?","Body":"<p>I have successfully performed 2d interpolation in python using the RectBivariateSpline method from scipy.interpolate. However, it is performed on numpy arrays. I want to perform it on tensors solely using tensorflow.<\/p>\n\n<p>This is what I have right now: It works if all are numpy arrays. However, I am having a hard time to rewrite it in tensorflow. <\/p>\n\n<pre><code>x_old = np.arange(0,256)\ny_old = np.arange(0,256)\n\n#x = tensor of shape [256,256]\n#y = tensor of shape [256,256]\n#in_im = tensor of shape [256,256,3]\n#out_im = tensor of shape [256,256,3]\n\nfor d in range(0,3):\n    interpf = RectBivariateSpline( x_old, y_old, in_im[:,:,d])\n    out_im[:,:,d] = interpf.ev(x[:,:], y[:,:])\n<\/code><\/pre>\n","answers":[{"AnswerId":"57489680","CreationDate":"2019-08-14T07:05:38.980","ParentId":null,"OwnerUserId":"10400883","Title":null,"Body":"<p>To convert tensors into numpy array is the solution.<br>\n<a href=\"https:\/\/stackoverflow.com\/questions\/34097281\/how-can-i-convert-a-tensor-into-a-numpy-array-in-tensorflow\">This question<\/a> about conversion might be helpful.  <\/p>\n\n<blockquote>\n  <p>In short, Any tensor returned by Session.run or eval is a NumPy array.  <\/p>\n<\/blockquote>\n\n<p>Example code is below.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nfrom scipy.interpolate import RectBivariateSpline\n\nx = tf.constant([1,2,3,4])\ny = tf.constant([1,2,3,4,5])\nvals = tf.constant([\n    [4,1,4,4,2],\n    [4,2,3,2,6],\n    [3,7,4,3,5],\n    [2,4,5,3,4]\n])\n\nsess = tf.Session()\nx, y, vals = sess.run([x, y, vals]) # x, y vals are now ndarray\n\nrect_B_spline = RectBivariateSpline(x, y, vals)\n\na = tf.constant([3.2, 3.8, 2.2])\nb = tf.constant([2.4, 4.3, 3.3])\n\na = sess.run([a, b])\n\nprint(rect_B_spline.ev(a, b))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57487634,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2019-08-14T02:57:39.697","AcceptedAnswerId":null,"OwnerUserId":4305208.0,"Title":"Loss Returns NaN When Using Regulizers On LSTM","Body":"<p>When I train my LSTM model, it returns nan for the loss<\/p>\n\n<ul>\n<li>I am using a single layer LSTM with a Dense softmax layer at the end for classification output<\/li>\n<li>Adam optimizer<\/li>\n<li>Categorical crossentropy loss function<\/li>\n<li>Relu activation<\/li>\n<\/ul>\n\n<p>For some reason when I use regulizers of any kind on the LSTM layer I get NaN for the loss.\nAlso this only occurs when I make the LSTM layer over 128 units, so I can get rid of it if I get rid of the regularization or make the network 128 or less units.\nI have already confirmed that there are no NaNs in the input.\nI'm wondering why this happens, and how it is possible to regularize larger LSTM layers.<\/p>\n\n<p>Here's my code:<\/p>\n\n<pre><code>def build_model():\nmodel = Sequential()\nmodel.add(LSTM(130, batch_input_shape=(None,90,5), return_sequences=False, recurrent_dropout=0.1, kernel_regularizer=regularizers.l2(0.01)))\nmodel.add(Activation(\"relu\"))\nmodel.add(Dense(2))\nmodel.add(Activation(\"softmax\"))\nmodel.compile(optimizer=\"adam\", loss = \"categorical_crossentropy\", metrics=[\"categorical_accuracy\"])\nreturn(model)\n<\/code><\/pre>\n\n<p>Thanks<\/p>\n","answers":[]}
+{"QuestionId":57487768,"AnswerCount":0,"Tags":"<python-3.x><tensorflow>","CreationDate":"2019-08-14T03:20:37.000","AcceptedAnswerId":null,"OwnerUserId":11925201.0,"Title":"Convolution Neural Network Using Tensorflow","Body":"<p>I'm trying to create a cnn using tensorflow for train detection excluding bounding box. I'm getting cost as zero for any iteration. Please can somebody help me!<\/p>\n\n<p>I'm using tensorflow 1.14 and python 3.6.8 in Anaconda <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def conv_nn(X, params):\n    W1, W2, W3 = params['W1'], params['W2'], params['W3']\n    #layer - 1\n    #convolution and average_pooling\n    C1 = tf.compat.v1.nn.conv2d(X, W1, strides = [1, 1, 1, 1], padding = \"VALID\")\n    P1 = tf.compat.v1.nn.avg_pool2d(C1, ksize =[1, 2, 2, 1], strides =  [1, 2, 2, 1], padding = \"VALID\")\n    #layer - 2\n    #convolution and average_padding\n    C2 = tf.compat.v1.nn.conv2d(P1, W2, strides = [1, 1, 1, 1], padding = \"VALID\")\n    P2 = tf.compat.v1.nn.avg_pool2d(C2, ksize = [1, 2, 2, 1], strides = [1, 2, 2, 1], padding = \"VALID\")\n    #layers - 3\n    #convolution and average_padding\n    C3 = tf.compat.v1.nn.conv2d(P2, W3, strides = [1, 1, 1, 1], padding = \"VALID\")\n    P3 = tf.compat.v1.nn.avg_pool2d(C3, ksize = [1, 2, 2, 1], strides = [1, 2, 2, 1] , padding = \"VALID\")\n    #flatten the output\n    P = tf.contrib.layers.flatten(P3)\n    #fully connected layers\n    Z1 = tf.contrib.layers.fully_connected(P, num_outputs = 120, activation_fn = tf.compat.v2.nn.relu)\n    Z2 = tf.contrib.layers.fully_connected(Z1, num_outputs = 1, activation_fn = tf.compat.v2.nn.softmax)\n    return Z2\n\n\n    def cost_compute(Z2, Y):\n    cost = tf.compat.v1.reduce_mean(tf.compat.v1.nn.softmax_cross_entropy_with_logits(labels = Y, logits = Z2))\n    return cost\n\n\n    def main(X_train, y_train, learning_rate, epochs):\n    costs = []\n    m, n_h, n_w, n_c = X_train.shape\n    _, n_y = y_train.shape\n    #creating placeholders\n    X, Y = create_placeholders([n_h, n_w, n_c, n_y])\n    #initial parameters\n    parameters = initial_parameters()\n    #forward propagation using convolution NN\n    Z2 = conv_nn(X, parameters)\n    #cost function\n    cost = cost_compute(Z2, Y)\n    with tf.compat.v1.variable_scope('main', reuse = tf.AUTO_REUSE):\n        #optimizer - back propagation\n        optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate = learning_rate).minimize(cost)\n        init = tf.compat.v1.global_variables_initializer()\n\n        with tf.compat.v1.Session() as ses:\n            ses.run(init)\n            for i in range(epochs):\n                _, temp_cost = ses.run([optimizer, cost], feed_dict = {X: X_train, Y: y_train})\n                costs.append(temp_cost)\n                print('Cost for iteration {} is {}'.format(i, temp_cost), end = \"##\")\n    return Z2, parameters, costs\n\n\n    main(X_train, y_train, 0.06, 400)\n<\/code><\/pre>\n\n<p>This is definitely not expected: Cost for iteration 0 is 0.0##Cost for iteration 1 is 0.0##Cost for iteration 2 is 0.0## ..... so on\nIt must give me some value.<\/p>\n","answers":[]}
+{"QuestionId":57487844,"AnswerCount":0,"Tags":"<theano>","CreationDate":"2019-08-14T03:33:19.743","AcceptedAnswerId":null,"OwnerUserId":11135787.0,"Title":"how to convert tensor to numpy in theano","Body":"<p>how can I convert tensor to numpy in theano?<\/p>\n\n<p>I want to send tensor form theano to pytorch, I do not know how to do it.<\/p>\n\n<p>I met this error:<\/p>\n\n<pre><code>tensordot(): argument 'input' (position 1) must be Tensor, not TensorVariable\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57487893,"AnswerCount":0,"Tags":"<android><ios><object><tensorflow><detection>","CreationDate":"2019-08-14T03:42:24.330","AcceptedAnswerId":null,"OwnerUserId":11861686.0,"Title":"My object detection tensorflow lite training with my own data based on flowers object detection tensorflow example is showing errors","Body":"<p>(<a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/examples\/blob\/master\/community\/en\/flowers_tf_lite.ipynb\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/github\/tensorflow\/examples\/blob\/master\/community\/en\/flowers_tf_lite.ipynb<\/a>)<\/p>\n\n<p>I am using the tensorflow lite object detection example from tensorflow website. I am running the google colab notebook that they provided and inputting my own data but I am receiving an error :<\/p>\n\n<p>\"ValueError: <code>generator<\/code> yielded an element of shape (52, 224, 224, 3) where an element of shape (64, 224, 224, 3) was expected.\".<\/p>\n\n<p>I realized that it could be because of the batch size which was set to 64 and because the training data is not a multiple of 64. So i tried inputting 704 images, which is a multiple of 64. I still got the same error.<\/p>\n\n<p>The code is exactly the same as the google colab link, except for the data input I am feeding it directly to the generator (which works perfectly fine).<\/p>\n\n<pre><code>enter code here\nIMAGE_SIZE = 224\nBATCH_SIZE =64\n\ndatagen = tf.keras.preprocessing.image.ImageDataGenerator(\nrescale=1.\/255, \nvalidation_split=0.2)\n\ntrain_generator = datagen.flow_from_directory(\n'\/content\/gdrive\/My Drive\/blister_images\/',\ntarget_size=(IMAGE_SIZE, IMAGE_SIZE),\nbatch_size=BATCH_SIZE, \nsubset='training')\n\nval_generator = datagen.flow_from_directory(\n'\/content\/gdrive\/My Drive\/blister_images\/',\ntarget_size=(IMAGE_SIZE, IMAGE_SIZE),\nbatch_size=BATCH_SIZE, \nsubset='validation')\n\n ------------------------------------------------------------------- \n --------\nInvalidArgumentError                      Traceback (most recent \ncall \nlast)\n&lt;ipython-input-18-238fbae26d39&gt; in &lt;module&gt;()\n  3 history = model.fit(train_generator, \n  4                     epochs=epochs,\n----&gt; 5                     validation_data=val_generator)\n\n10 frames\n\/usr\/local\/lib\/python3.6\/dist-packages\/six.py in raise_from(value, \nfrom_value)\n\nInvalidArgumentError: ValueError: `generator` yielded an element of \nshape (52, 224, 224, 3) where an element of shape (64, 224, 224, 3) \nwas \nexpected.\nTraceback (most recent call last):\n\nFile \"\/usr\/local\/lib\/python3.6\/dist- \npackages\/tensorflow_core\/python\/ops\/script_ops.py\", line 217, in \n__call__\nret = func(*args)\n\nFile \"\/usr\/local\/lib\/python3.6\/dist- \npackages\/tensorflow_core\/python\/data\/ops\/dataset_ops.py\", line 630, \nin \ngenerator_py_func\n\"of shape %s was expected.\" % (ret_array.shape, expected_shape))\n\nValueError: `generator` yielded an element of shape (52, 224, 224, \n3) \nwhere an element of shape (64, 224, 224, 3) was expected.\n\n\n [[{{node PyFunc}}]] [Op:IteratorGetNextSync]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57487949,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-14T03:53:27.537","AcceptedAnswerId":null,"OwnerUserId":11924687.0,"Title":"TypeError: __init__() got an unexpected keyword argument 'n_iter'","Body":"<p>I am getting error while trying to run the code in jupyter notebook<\/p>\n\n<p>I have tried to run the code by installing python separately, not using any python IDE<\/p>\n\n<pre><code>TypeError                                 Traceback (most recent call last)\n&lt;ipython-input-6-b0fdfbdfd030&gt; in &lt;module&gt;\n    194 \n    195 if __name__ == '__main__':\n--&gt; 196     add_user()\n\n&lt;ipython-input-6-b0fdfbdfd030&gt; in add_user()\n    182         # when features of 3 files of speaker are concatenated, then do model training\n    183         if count == 3:\n--&gt; 184             gmm = GMM(n_components = 16, n_iter = 200, covariance_type='diag',n_init = 3)\n    185             gmm.fit(features)\n    186 \n\nTypeError: __init__() got an unexpected keyword argument 'n_iter'\n<\/code><\/pre>\n\n<p>I need to get the output of the voice recorder<\/p>\n","answers":[{"AnswerId":"57488025","CreationDate":"2019-08-14T04:03:41.973","ParentId":null,"OwnerUserId":"11924650","Title":null,"Body":"<p>It looks like you're trying to use an SKLearn <a href=\"https:\/\/ogrisel.github.io\/scikit-learn.org\/sklearn-tutorial\/modules\/generated\/sklearn.mixture.GMM.html\" rel=\"nofollow noreferrer\">GMM<\/a> object, but possibly giving the parameters for a <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.mixture.GaussianMixture.html\" rel=\"nofollow noreferrer\">GausianMixture<\/a> object instead, which arent supported in the GMM constructor. <\/p>\n\n<p>The <code>n_iter<\/code> parameter, along with the <code>n_init<\/code> parameter are not valid parameters for the GMM object, so I would suggest checking the documentation for each (I linked them for you) and checking again which one you need to use, and setting the parameters accordingly.<\/p>\n\n<p>EDIT: If its not SKLearn you're trying to use, please clarify this in your question regarding which package you're using here.<\/p>\n"}]}
+{"QuestionId":57488150,"AnswerCount":2,"Tags":"<python><numpy><tensorflow>","CreationDate":"2019-08-14T04:26:31.020","AcceptedAnswerId":57488298.0,"OwnerUserId":3616293.0,"Title":"tensorflow warning for data types","Body":"<p>I have installed tensorflow and numpy in Python 3.7.4 [64-bit]. When I tried importing it, I get the following warning:<\/p>\n\n<blockquote>\n  <p>\/home\/user\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/dtypes.py:516:\n  FutureWarning: Passing (type, 1) or '1type' as a synonym of type is\n  deprecated; in a future version of numpy, it will be understood as\n  (type, (1,)) \/ '(1,)type'.   _np_qint8 = np.dtype([(\"qint8\", np.int8,\n  1)])\n  \/home\/user\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/dtypes.py:517:\n  FutureWarning: Passing (type, 1) or '1type' as a synonym of type is\n  deprecated; in a future version of numpy, it will be understood as\n  (type, (1,)) \/ '(1,)type'.   _np_quint8 = np.dtype([(\"quint8\",\n  np.uint8, 1)])\n  \/home\/user\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/dtypes.py:518:\n  FutureWarning: Passing (type, 1) or '1type' as a synonym of type is\n  deprecated; in a future version of numpy, it will be understood as\n  (type, (1,)) \/ '(1,)type'.   _np_qint16 = np.dtype([(\"qint16\",\n  np.int16, 1)])\n  \/home\/user\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/dtypes.py:519:\n  FutureWarning: Passing (type, 1) or '1type' as a synonym of type is\n  deprecated; in a future version of numpy, it will be understood as\n  (type, (1,)) \/ '(1,)type'.   _np_quint16 = np.dtype([(\"quint16\",\n  np.uint16, 1)])\n  \/home\/user\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/dtypes.py:520:\n  FutureWarning: Passing (type, 1) or '1type' as a synonym of type is\n  deprecated; in a future version of numpy, it will be understood as\n  (type, (1,)) \/ '(1,)type'.   _np_qint32 = np.dtype([(\"qint32\",\n  np.int32, 1)])\n  \/home\/user\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/dtypes.py:525:\n  FutureWarning: Passing (type, 1) or '1type' as a synonym of type is\n  deprecated; in a future version of numpy, it will be understood as\n  (type, (1,)) \/ '(1,)type'.   np_resource = np.dtype([(\"resource\",\n  np.ubyte, 1)])<\/p>\n  \n  <p>\/home\/user\/.local\/lib\/python3.7\/site-packages\/tensorboard\/compat\/tensorflow_stub\/dtypes.py:541:\n  FutureWarning: Passing (type, 1) or '1type' as a synonym of type is\n  deprecated; in a future version of numpy, it will be understood as\n  (type, (1,)) \/ '(1,)type'.   _np_qint8 = np.dtype([(\"qint8\", np.int8,\n  1)])\n  \/home\/user\/.local\/lib\/python3.7\/site-packages\/tensorboard\/compat\/tensorflow_stub\/dtypes.py:542:\n  FutureWarning: Passing (type, 1) or '1type' as a synonym of type is\n  deprecated; in a future version of numpy, it will be understood as\n  (type, (1,)) \/ '(1,)type'.   _np_quint8 = np.dtype([(\"quint8\",\n  np.uint8, 1)])\n  \/home\/user\/.local\/lib\/python3.7\/site-packages\/tensorboard\/compat\/tensorflow_stub\/dtypes.py:543:\n  FutureWarning: Passing (type, 1) or '1type' as a synonym of type is\n  deprecated; in a future version of numpy, it will be understood as\n  (type, (1,)) \/ '(1,)type'.   _np_qint16 = np.dtype([(\"qint16\",\n  np.int16, 1)])\n  \/home\/user\/.local\/lib\/python3.7\/site-packages\/tensorboard\/compat\/tensorflow_stub\/dtypes.py:544:\n  FutureWarning: Passing (type, 1) or '1type' as a synonym of type is\n  deprecated; in a future version of numpy, it will be understood as\n  (type, (1,)) \/ '(1,)type'.   _np_quint16 = np.dtype([(\"quint16\",\n  np.uint16, 1)])\n  \/home\/user\/.local\/lib\/python3.7\/site-packages\/tensorboard\/compat\/tensorflow_stub\/dtypes.py:545:\n  FutureWarning: Passing (type, 1) or '1type' as a synonym of type is\n  deprecated; in a future version of numpy, it will be understood as\n  (type, (1,)) \/ '(1,)type'.   _np_qint32 = np.dtype([(\"qint32\",\n  np.int32, 1)])\n  \/home\/user\/.local\/lib\/python3.7\/site-packages\/tensorboard\/compat\/tensorflow_stub\/dtypes.py:550:\n  FutureWarning: Passing (type, 1) or '1type' as a synonym of type is\n  deprecated; in a future version of numpy, it will be understood as\n  (type, (1,)) \/ '(1,)type'.   np_resource = np.dtype([(\"resource\",\n  np.ubyte, 1)])<\/p>\n<\/blockquote>\n\n<p>I have the following versions-\nnumpy- 1.17.0\ntensorflow- 1.14.0<\/p>\n\n<p>How can I solve the warnings for the different data types in tensorflow?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"57998742","CreationDate":"2019-09-18T18:21:02.350","ParentId":null,"OwnerUserId":"3752281","Title":null,"Body":"<p>I decided to ignore them for now:<\/p>\n\n<pre><code>import warnings\nwarnings.filterwarnings('ignore')\n<\/code><\/pre>\n\n<p>although it does not resolve the root cause. <\/p>\n"},{"AnswerId":"57488298","CreationDate":"2019-08-14T04:46:55.273","ParentId":null,"OwnerUserId":"901925","Title":null,"Body":"<p>What it's saying is that to conform to planned changes in <code>numpy<\/code> this line<\/p>\n\n<pre><code>np_resource = np.dtype([(\"resource\", np.ubyte, 1)])\n<\/code><\/pre>\n\n<p>will need to be rewritten as<\/p>\n\n<pre><code>np_resource = np.dtype([(\"resource\", np.ubyte, (1,))])\n<\/code><\/pre>\n\n<p>I don't think this is in your own code.<\/p>\n\n<p>The relevant passage in numpy 1.17 is:<\/p>\n\n<p><a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/release.html#future-changes\" rel=\"noreferrer\">https:\/\/docs.scipy.org\/doc\/numpy\/release.html#future-changes<\/a><\/p>\n\n<p>This is a warning, not an error.  In a similar recent SO, the poster got rid of it by switching to an earlier <code>numpy<\/code> version.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/57381430\/synonym-of-type-is-deprecated-in-a-future-version-of-numpy-it-will-be-underst\">&quot;synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ &#39;(1,)type&#39;.&quot; problem in TensorFlow<\/a><\/p>\n"}]}
+{"QuestionId":57488250,"AnswerCount":2,"Tags":"<keras><deep-learning>","CreationDate":"2019-08-14T04:41:11.527","AcceptedAnswerId":57492190.0,"OwnerUserId":562769.0,"Title":"What does a loss list in Keras mean?","Body":"<p>I've just seen<\/p>\n\n<pre><code># Combined model trains generators to fool discriminators\nself.combined = Model(inputs=[img_A, img_B],\n                      outputs=[ valid_A, valid_B,\n                                reconstr_A, reconstr_B,\n                                img_A_id, img_B_id ])\nself.combined.compile(loss=['mse', 'mse',\n                            'mae', 'mae',\n                            'mae', 'mae'],\n                    loss_weights=[  1, 1,\n                                    self.lambda_cycle, self.lambda_cycle,\n                                    self.lambda_id, self.lambda_id ],\n                    optimizer=optimizer)\n<\/code><\/pre>\n\n<p>in <a href=\"https:\/\/github.com\/eriklindernoren\/Keras-GAN\/blob\/master\/cyclegan\/cyclegan.py\" rel=\"nofollow noreferrer\">CycleGan<\/a> - what does the list of losses mean \/ do? Before, I only worked with exactly one loss function per model.<\/p>\n","answers":[{"AnswerId":"57492190","CreationDate":"2019-08-14T09:43:46.507","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>This model has multiple inputs (two) and multiple outputs (six), so you need to specify one loss function for each output. That's why there is a list of losses.<\/p>\n\n<p>Additionally a model can only be trained with a single loss function, and for a multi-output model, this is accomplished by creating a virtual loss that is a weighted combination of all per-output losses, and this is what the <code>loss_weights<\/code> parameter is for.<\/p>\n"},{"AnswerId":"57514229","CreationDate":"2019-08-15T18:12:37.803","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>It is also worth noting that you need to provide a list of loss functions for a multi-output model only if the loss functions used are NOT all the same. For example, if all the output layers use the Mean Squared Error loss, then you only need to provide <code>loss='mse'<\/code>.<\/p>\n"}]}
+{"QuestionId":57488349,"AnswerCount":0,"Tags":"<python><tensorflow><keras><multiple-input>","CreationDate":"2019-08-14T04:53:35.597","AcceptedAnswerId":null,"OwnerUserId":11925417.0,"Title":"How to put multiple inputs in tensorflow 2.0?(without keras)","Body":"<p>I want to make CNN model including multiple inputs. But there is not any examples about that. Without <code>Keras<\/code>, Can I make multiple inputs model ?<\/p>\n\n<p>My model has 3 inputs. After the first input(<code>atom_in_fea[8,9]<\/code>) passes through the embedding(linear or dense in keras) layer, the output should merge with the second(<code>nbr_fea_idx[8,12]<\/code>) and third input(<code>nbr_fea[8,12,4]<\/code>).<\/p>\n\n<pre><code>atom_in_fea[8,9] -&gt; embedding layer -&gt; atom_in_fea[8,4]\n<\/code><\/pre>\n\n<p>[] is their shape.<\/p>\n\n<p>There is my code<\/p>\n\n<pre><code>python, tensorflow\n\natom_nbr_fea = atom_in_fea[nbr_fea_idx, :] \n\nfinal = tf.concat([atom_in_fea.unsqueeze(1).expand(8, 12, 4),atom_nbr_fea, nbr_fea], 2)\n<\/code><\/pre>\n\n<p>atom_nbr_fea shape is [8,12,4]<\/p>\n\n<p>Final shape is [8,12,12]<\/p>\n\n<p>I built a model with <code>Keras<\/code> but there was limited. \nIf I can solve this problem, it's fine whether it's <code>Keras<\/code> or <code>TensorFlow<\/code>. Thank you<\/p>\n","answers":[]}
+{"QuestionId":57488388,"AnswerCount":1,"Tags":"<keras><merge><lstm><layer>","CreationDate":"2019-08-14T04:58:49.640","AcceptedAnswerId":57491417.0,"OwnerUserId":11868918.0,"Title":"Keras : How to merge layers sequentially not use \"concatenate\"","Body":"<p>I'm trying to build a model that combines cnn and lstm.\nI want to multivariate the input of cnn and put the outputs sequentially into the input of the LSTM. However, there is a problem in merging the cnn outputs. If you use concatenate, it will stretch to axis = -1 as shown. But I'll put it in the lstm structure so I'd like to increase it sequentailly. But I didn't find any function to merge except concatenate. The shape I want is (None, 6, 1904) in the image below. What can I do?<\/p>\n\n<p>below is my build code.<\/p>\n\n<pre><code>def build_model():\n    in_layers, out_layers = [], []\n    for i in range(in_len):\n        inputs = Input(shape=(row,col, channel))\n        conv1 = Conv2D(4, (12, 12), activation='relu')(inputs)\n        pool1 = pooling.MaxPooling2D(pool_size=(4,4))(conv1)\n        conv2 = Conv2D(4, (7, 7) , activation='relu')(pool1)\n        pool2 = pooling.MaxPooling2D(pool_size=(3,3))(conv2)\n        conv3 = Conv2D(8, (5, 5) , activation='relu')(pool2)\n        pool3 = pooling.MaxPooling2D(pool_size=(2,2))(conv3)\n        flat = Flatten()(pool3)\n        # store layers\n        in_layers.append(inputs)\n        out_layers.append(flat)\n        print(type(flat))\n    merged = concatenate(out_layers)\n    model = Model(inputs=in_layers, outputs=merged)\n    plot_model(model, show_shapes=True, to_file='cnn_lstm_real.png')\n\n    return model\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Mte8x.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Mte8x.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"57491417","CreationDate":"2019-08-14T09:00:07.690","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>What you want is still concatenation, but on a different, new axis. The concatenation layer and function allows to specify the axis, so you can do it like this:<\/p>\n\n<pre><code>def build_model():\n    in_layers, out_layers = [], []\n    for i in range(in_len):\n        inputs = Input(shape=(row,col, channel))\n        conv1 = Conv2D(4, (12, 12), activation='relu')(inputs)\n        pool1 = pooling.MaxPooling2D(pool_size=(4,4))(conv1)\n        conv2 = Conv2D(4, (7, 7) , activation='relu')(pool1)\n        pool2 = pooling.MaxPooling2D(pool_size=(3,3))(conv2)\n        conv3 = Conv2D(8, (5, 5) , activation='relu')(pool2)\n        pool3 = pooling.MaxPooling2D(pool_size=(2,2))(conv3)\n        flat = Flatten()(pool3)\n        flat = Reshape((1, -1))(flat)\n        # store layers\n        in_layers.append(inputs)\n        out_layers.append(flat)\n\n    merged = concatenate(out_layers, axis = 1)\n    model = Model(inputs=in_layers, outputs=merged)\n    plot_model(model, show_shapes=True, to_file='cnn_lstm_real.png')\n\n    return model\n<\/code><\/pre>\n\n<p>The only big difference is that you need to add the new axis explicitly in the output of each branch (hence the <code>Reshape<\/code> layer), in order to allow for concatenation to happen along that axis.<\/p>\n"}]}
+{"QuestionId":57489252,"AnswerCount":0,"Tags":"<keras><merge><concat>","CreationDate":"2019-08-14T06:32:44.117","AcceptedAnswerId":null,"OwnerUserId":11925734.0,"Title":"How to merge\/concatenate in CNN with same input data","Body":"<p>I want to concatenate two models with same input data with Keras.<\/p>\n\n<p>How can I merge or concatenate two models?<\/p>\n\n<p>I use Keras and I want to create function (def conbination():)<\/p>\n\n<pre><code>def conbination():\n\n    model_1 = Sequential()\n    model_1.add(Conv2D(filters=32, kernel_size=3, input_shape=input_shape)) \n    model_1.add(Activation('relu'))\n    model_1.add(MaxPooling2D(pool_size=(64, 1)))\n    model_1.add(Flatten())\n\n    model_2 = Sequential()\n    model_2.add(Conv2D(filters=32, kernel_size=3, input_shape=input_shape))\n    model_2.add(Activation('relu'))\n    model_2.add(MaxPooling2D(pool_size=(1, 64)))\n    model_2.add(Flatten())\n\n    concat = concatenate([model_1 , model_2])\n\n    merged_model = Sequential()\n    merged_model.add(concat)\n    # merged_model.add(Activation('relu'))\n    merged_model.add(Dense(512))\n    merged_model.add(Activation('relu'))\n    merged_model.add(Dense(128))\n    merged_model.add(Activation('relu'))\n    # model.add(Dropout(0.5))\n    merged_model.add(Dense(num_classes, activation='softmax'))\n    merged_model.compile(loss='categorical_crossentropy',\n                  optimizer='Adam',  # sgd, #,\n                  metrics=['accuracy'])\n    return merged_model\n<\/code><\/pre>\n\n<p>I tried on concatenate([model_1, model_2]), and I got a message <\/p>\n\n<pre><code>A `Concatenate` layer should be called on a list of at least 2 inputs\n<\/code><\/pre>\n\n<p>I tried on concatenate([model_1.output, model_2.output]), and I got a message <\/p>\n\n<pre><code>The added layer must be an instance of class Layer. Found:   \nTensor(\"concatenate\/concat:0\", shape=(?, 8064), dtype=float32).\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57489266,"AnswerCount":0,"Tags":"<python-3.x><tensorflow>","CreationDate":"2019-08-14T06:33:33.297","AcceptedAnswerId":null,"OwnerUserId":11924687.0,"Title":"ModuleNotFoundError: No module named 'sklearn.mixture.gmm'","Body":"<p>I am getting the following error when trying to perform Voice Authentication<\/p>\n\n<p>I have tried in Jupyter notebook and Spyder, but unable to resolve the error<\/p>\n\n<pre><code>ModuleNotFoundError                       Traceback (most recent call last)\n&lt;ipython-input-13-df29f49ded81&gt; in &lt;module&gt;\n    158 \n    159 if __name__ == '__main__':\n--&gt; 160     recognize()\n\n&lt;ipython-input-13-df29f49ded81&gt; in recognize()\n     42                 os.listdir(modelpath) if fname.endswith('.gmm')]\n     43 \n---&gt; 44     models    = [pickle.load(open(fname,'rb')) for fname in gmm_files]\n     45 \n     46     speakers   = [fname.split(\"\/\")[-1].split(\".gmm\")[0] for fname \n\n&lt;ipython-input-13-df29f49ded81&gt; in &lt;listcomp&gt;(.0)\n     42                 os.listdir(modelpath) if fname.endswith('.gmm')]\n     43 \n---&gt; 44     models    = [pickle.load(open(fname,'rb')) for fname in gmm_files]\n     45 \n     46     speakers   = [fname.split(\"\/\")[-1].split(\".gmm\")[0] for fname\n<\/code><\/pre>\n\n<blockquote>\n  <p>ModuleNotFoundError: No module named 'sklearn.mixture.gmm'<\/p>\n<\/blockquote>\n\n<p>I don't understand where am lacking to understand the code.\nCan u please help me<\/p>\n","answers":[]}
+{"QuestionId":57489307,"AnswerCount":1,"Tags":"<machine-learning><keras><mnist>","CreationDate":"2019-08-14T06:37:10.170","AcceptedAnswerId":null,"OwnerUserId":8849802.0,"Title":"MNIST model predicting test images incorrectly even though it had high training and testing accuracy","Body":"<p>I'm trying to create a model which can predict handwritten digits. I've trained a Keras model with the architecture below and saw an accuracy of near 99% during testing. However, I decided to take some pictures off the internet of some black &amp; white numbers and minimize their size to 28x28 with OpenCV to try and use my trained model to predict them. After testing all 10 digits, my model predicted only 4\/10 correctly, meaning that it had a 40% score.<\/p>\n\n<p>Why is this happening and what can I do to fix it? <\/p>\n\n<p>Here is the MNIST dataset I'm using: \n<a href=\"https:\/\/www.kaggle.com\/oddrationale\/mnist-in-csv\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/oddrationale\/mnist-in-csv<\/a><\/p>\n\n<p>Here are the images I'm testing:\n<a href=\"https:\/\/drive.google.com\/open?id=1G0CMK0ZPI1JIX2cYXNRt3Qsq_YpjlMjX\" rel=\"nofollow noreferrer\">https:\/\/drive.google.com\/open?id=1G0CMK0ZPI1JIX2cYXNRt3Qsq_YpjlMjX<\/a><\/p>\n\n<pre><code>#architecture\nmodel = 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<\/code><\/pre>\n\n<pre><code>#Code for minmizing\nimg = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)\nimg = cv2.resize(img,(28,28))\n\n#turning the 28x28 array into a (1, 28, 28, 1) numpy array for the model to process\n#dividing each value by 255. During the training, each grayscale pixel value was treated the same way\n\nfor i in range(28):\n    temp = []\n    for x in range(28):\n        temp.append([img[i][x]\/255])\n    pendo.append(temp)\nimg = np.array([pendo]).astype(\"float32\")\n\n#predicting the image\nmodel = load_model('MNIST.h5')\nclasses = model.predict(img)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57489552","CreationDate":"2019-08-14T06:56:04.903","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>MNIST is not a dataset meant to learn a completely general digit recognition model, its just an academic benchmark, a very old one, so getting any kind of test accuracy close to 99% is very easy and it does not mean that the model will generalize at all.<\/p>\n\n<p>In the end this is not a programming problem, and that should be clear, its not a problem in the code, its a problem with the data. The images you have are just too different from the ones in the MNIST training set.<\/p>\n"}]}
+{"QuestionId":57489460,"AnswerCount":1,"Tags":"<python><keras><conv-neural-network>","CreationDate":"2019-08-14T06:48:23.143","AcceptedAnswerId":57541528.0,"OwnerUserId":11925809.0,"Title":"Performing Conv2D on 4D data in Keras","Body":"<p>I have data of dimension <code>24*64*64*10<\/code> (excluding the batch size).<\/p>\n\n<p>I want to split the input into 24 inputs of dimension <code>64*64*10<\/code>, perform <code>Conv2D<\/code> on each of them and then concatenate them to get the 4D data again for further processing.<\/p>\n\n<p>Any help regarding the implementation would be helpful. I am working with Keras.<\/p>\n\n<p>Edit: I tried to the following code to perform the 2D convolution<\/p>\n\n<pre><code>num_ch= 24\ninput= Input(shape=(64,64,10,num_ch))\nprint(input.shape)\nbranch_out= []\nfor i in range(num_ch):\n    out= Lambda(lambda x: x[:,:,:,:,i] )(input)\n    print(out.shape)\n    out= Conv2D(10, kernel_size=(3,3),strides= (1,1), padding='same', data_format= 'channels_last')(input)\n    branch_out.append(out)\n<\/code><\/pre>\n\n<p>I got the following error:<\/p>\n\n<pre><code>(?, 64, 64, 10, 24)\n(?, 64, 64, 10)\n---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-83-51977f4edbba&gt; in &lt;module&gt;\n      7     out= Lambda(lambda x: x[:,:,:,:,i] )(input)\n      8     print(out.shape)\n----&gt; 9     out= Conv2D(10, kernel_size=(3,3),strides= (1,1), padding='same', data_format= 'channels_last')(input)\n     10     branch_out.append(out)\n\n~\/anaconda3\/lib\/python3.7\/site-packages\/keras\/engine\/base_layer.py in __call__(self, inputs, **kwargs)\n    412                 # Raise exceptions in case the input is not compatible\n    413                 # with the input_spec specified in the layer constructor.\n--&gt; 414                 self.assert_input_compatibility(inputs)\n    415 \n    416                 # Collect input shapes to build layer.\n\n~\/anaconda3\/lib\/python3.7\/site-packages\/keras\/engine\/base_layer.py in assert_input_compatibility(self, inputs)\n    309                                      self.name + ': expected ndim=' +\n    310                                      str(spec.ndim) + ', found ndim=' +\n--&gt; 311                                      str(K.ndim(x)))\n    312             if spec.max_ndim is not None:\n    313                 ndim = K.ndim(x)\n\nValueError: Input 0 is incompatible with layer conv2d_25: expected ndim=4, found ndim=5\n<\/code><\/pre>\n","answers":[{"AnswerId":"57541528","CreationDate":"2019-08-18T03:56:53.323","ParentId":null,"OwnerUserId":"5539758","Title":null,"Body":"<p>You have a typo in this line:<\/p>\n\n<pre><code>out= Conv2D(10, kernel_size=(3,3),strides= (1,1), padding='same', data_format= 'channels_last')(input)\n<\/code><\/pre>\n\n<p>Change it to:<\/p>\n\n<pre><code>out= Conv2D(10, kernel_size=(3,3),strides= (1,1), padding='same', data_format= 'channels_last')(out)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57489612,"AnswerCount":1,"Tags":"<tensorflow><softmax>","CreationDate":"2019-08-14T07:01:17.753","AcceptedAnswerId":57724685.0,"OwnerUserId":10561436.0,"Title":"tf.nn.softmax behaving strangely","Body":"<p>I am learning LSTM with tensorflow with enable_eager_execution. However when implementing LSTM, I have noticed the behaviour of tf.nn.softmax \nthat has made me stuck. Here is a section of my code<\/p>\n\n<pre><code>class RNN_LSTM(object):\n    def __init__(self,hidden_size):\n        data=open('Shakespear.txt', 'r').read() \n        self.data = data.split()\n        vocab_size=len(list(set(self.data)))\n        self.words =list(set(self.data))\n        self.hidden_size=hidden_size\n        self.input_size=vocab_size+hidden_size\n        self.vocab_size=vocab_size\n        self.W1=tf.Variable(tf.random.uniform((self.hidden_size,self.input_size),dtype=tf.dtypes.float32,name=\"W1\")*0.1)\n        self.b1=tf.Variable(tf.random.uniform((self.hidden_size,1),dtype=tf.dtypes.float32,name=\"b1\"))\n        self.W2=tf.Variable(tf.random.uniform((self.hidden_size,self.input_size),dtype=tf.dtypes.float32,name=\"W2\")*0.1)\n        self.b2=tf.Variable(tf.random.uniform((self.hidden_size,1),dtype=tf.dtypes.float32,name=\"b2\")*0.1)  \n        self.W3=tf.Variable(tf.random.uniform((self.hidden_size,self.input_size),dtype=tf.dtypes.float32,name=\"W3\")*0.1)\n        self.b3=tf.Variable(tf.random.uniform((self.hidden_size,1),dtype=tf.dtypes.float32,name=\"b3\")*0.1)\n        self.W4=tf.Variable(tf.random.uniform((hidden_size,self.input_size),dtype=tf.dtypes.float32,name=\"W4\")*0.1)\n        self.b4=tf.Variable(tf.random.uniform((self.hidden_size,1),dtype=tf.dtypes.float32,name=\"b4\")*0.1)\n        self.W5=tf.Variable(tf.random.uniform((self.vocab_size,self.hidden_size),dtype=tf.dtypes.float32,name=\"W5\")*0.1)\n        self.b5=tf.Variable(tf.random.uniform((self.vocab_size,1),dtype=tf.dtypes.float32,name=\"b5\")*0.1)\n\n        self.learning_rate=1e-1\n        self.sequence_length=50\n\n\n        #self.M_c=tf.Variable(tf.zeros((self.input_size,1)),name=\"M_c\")\n    def one_hot_encoding(self,x,hprev):\n        M_c=tf.Variable(tf.zeros((self.input_size,1)),name=\"M_c\")\n        vocab=tf.Variable(tf.zeros((self.vocab_size,1)))\n        #hprev=tf.Variable(tf.zeros((self.hidden_size,1)))\n        vocab=vocab.numpy()\n        vocab[x]=1\n        M_c=tf.concat((hprev,vocab),axis=0)\n        return M_c\n     def feedforward(self,M_c,p_s):\n            ft=tf.sigmoid( tf.matmul(self.W1,M_c)+self.b1)\n            it=tf.sigmoid(tf.matmul(self.W2,M_c)+self.b2)\n            gt=tf.math.tanh(tf.matmul(self.W3,M_c)+self.b3)\n            cs=tf.multiply(ft,p_s)+tf.multiply(it,gt)\n            ot=tf.nn.sigmoid(tf.matmul(self.W4,M_c)+self.b4)\n            ht=tf.multiply(ot,tf.math.tanh(cs))\n            output=self.softmax(tf.matmul(self.W5,ht)+self.b5)                     \n            return ht,output,cs\n    def sample_text(self,hprev,begin,p_s,n):\n        vocab=tf.Variable(tf.zeros((self.vocab_size,1)),tf.float32)\n        vocab=vocab.numpy()\n        vocab[begin]=1\n        letters=[]\n        for i in range(n):\n            M=tf.Variable(tf.zeros((self.input_size,1)),name=\"M\")\n            M=tf.assign(M,tf.concat((hprev,vocab),axis=0))\n            ft=tf.nn.sigmoid(tf.matmul(self.W1,M)+self.b1)\n            it=tf.nn.sigmoid(tf.matmul(self.W2,M)+self.b2)\n            gt=tf.math.tanh(tf.matmul(self.W3,M)+self.b3)\n            cs=tf.multiply(ft,p_s)+tf.multiply(it,gt)\n            p_s=cs\n            ot=tf.sigmoid(tf.matmul(self.W4,M)+self.b4)\n            ht=tf.multiply(ot,tf.math.tanh(cs))\n            ht=tf.reshape(ht,(self.hidden_size,1))\n            output=tf.matmul(self.W5,ht)+self.b5\n            p=self.softmax(output)\n            #print(p.numpy())\n            p=tf.reshape(p,(1,self.vocab_size))\n            samples = tf.random.categorical(p,1)\n            sample_selected=tf.cast(samples[0][0].numpy(),tf.int32)\n            selection_sample_np=[i for i in range(self.vocab_size)]\n            selection_sample_tf=tf.convert_to_tensor(selection_sample_np)\n            selected_next_letter=selection_sample_tf[sample_selected]\n            trial=tf.cast(selected_next_letter,tf.int32)\n            k=tf.Variable(tf.zeros((self.vocab_size,1)),tf.int32)\n            k[selected_next_letter,0].assign(1)\n            letters.append(selected_next_letter)\n            hprev=ht\n        return letters \n    def process_input(self):\n        char_to_ix={ch:ix for ix,ch in enumerate(self.words)}\n        ix_to_char={ix:ch for ix,ch in enumerate(self.words)}\n        return char_to_ix,ix_to_char\n    def softmax(self,z):\n        return tf.math.exp(z-max(z))\/tf.math.reduce_sum(tf.math.exp(z-max(z)))\n\n    def AggregatorNew(self):\n        losses,iterations=[],[]\n        char_to_ix,ix_to_char=self.process_input()\n        mem1=tf.Variable(tf.zeros_like(self.W1))\n        mem2=tf.Variable(tf.zeros_like(self.W2))\n        mem3=tf.Variable(tf.zeros_like(self.W3))\n        mem4=tf.Variable(tf.zeros_like(self.W4))\n        mem5=tf.Variable(tf.zeros_like(self.W5))\n        mem6=tf.Variable(tf.zeros_like(self.b1))\n        mem7=tf.Variable(tf.zeros_like(self.b2))\n        mem8=tf.Variable(tf.zeros_like(self.b3))\n        mem9=tf.Variable(tf.zeros_like(self.b4))\n        mem10=tf.Variable(tf.zeros_like(self.b5))       \n        dW1=tf.Variable(tf.zeros_like(self.W1))\n        dW2=tf.Variable(tf.zeros_like(self.W2))\n        dW3=tf.Variable(tf.zeros_like(self.W3))\n        dW4=tf.Variable(tf.zeros_like(self.W4))\n        dW5=tf.Variable(tf.zeros_like(self.W4))\n        db1=tf.Variable(tf.zeros_like(self.b1))\n        db2=tf.Variable(tf.zeros_like(self.b2))\n        db3=tf.Variable(tf.zeros_like(self.b3))\n        db4=tf.Variable(tf.zeros_like(self.b4))\n        db5=tf.Variable(tf.zeros_like(self.b5))\n        n=0 \n        p=0\n        self.loss=tf.Variable(0,dtype=tf.dtypes.float32,name=\"loss\")\n        smooth_loss =-tf.math.log(1.0\/self.vocab_size)*self.sequence_length\n        while(1):\n            try:\n                with DelayedKeyboardInterrupt():\n                    if p+self.sequence_length+1&gt;= len(self.data) or n == 0: \n                        hprev=tf.Variable(np.zeros((self.hidden_size,1)),dtype=tf.float32,name=\"hprev\")\n                        p_s=tf.Variable(tf.zeros((self.hidden_size,1)),name=\"p_s\")\n                        p=0\n                    inputs=[char_to_ix[ch] for ch in self.data[p:p+self.sequence_length]]\n                    targets=[char_to_ix[ch] for ch in self.data[p+1:p+self.sequence_length+1]]\n                    sample_ix = self.sample_text(hprev,inputs[0],p_s,200)\n                    list_of_strings=[ix_to_char[ix.numpy()] for ix in sample_ix]\n                    list_of_strings_tf=tf.convert_to_tensor(list_of_strings)\n                    txt = tf.strings.join(list_of_strings_tf,separator=\" \")\n                    print ('----\\n %s \\n----' % (txt.numpy(),  ))\n\n                    #loss=tf.reduce_mean(xentropy,name=\"loss\")\n                    with tf.GradientTape() as g:\n                        for x, y in zip(inputs,targets):\n                            M_c=self.one_hot_encoding(x,hprev) \n                            hprev,output,p_s=self.feedforward(M_c,p_s)\n                            activation=output[y]\n                            loss=-(tf.math.log(activation))\n                    dW1,dW2,dW3,dW4,dW5,db1,db2,db3,db4,db5=g.gradient(loss,[self.W1,self.W2,self.W3,self.W4,self.W5,self.b1,self.b2,self.b3,self.b4,self.b5])\n                    smooth_loss = smooth_loss * 0.999 + loss * 0.001\n\n            except KeyboardInterrupt:\n                    sample_ix = self.sample_text(hprev,inputs[0],p_s,200)\n                    txt = ''.join(ix_to_char[ix] for ix in sample_ix)\n                    print ('----\\n %s \\n----' % (txt,  ))                    \n                    break\n<\/code><\/pre>\n\n<p>when I use <code>self.softmax()<\/code> it gives me probability values in the output in the feedforward, however when I use <code>tf.nn.softmax()<\/code> all values of output are strangely 1.<\/p>\n\n<p>Second question: Is tensorflow generally slower in cpu as compared to a pure python implementation or i am implementing tensorlow wrongly?<\/p>\n","answers":[{"AnswerId":"57724685","CreationDate":"2019-08-30T10:04:57.813","ParentId":null,"OwnerUserId":"10561436","Title":null,"Body":"<p>If you are using <code>tf.nn.softmax()<\/code>, and you don't specify  the axis, it defaults to <code>tf.nn.softmax(logits ,axis=1)<\/code> hence giving a tensor ouput where all values are 1s  . In my case I was getting wrong values just because of not providing axis i.e <code>tf.nn.softmax(logits,axis=0)<\/code><\/p>\n"}]}
+{"QuestionId":57489639,"AnswerCount":1,"Tags":"<tensorflow><nlp><word-embedding>","CreationDate":"2019-08-14T07:03:26.837","AcceptedAnswerId":57496598.0,"OwnerUserId":2509698.0,"Title":"Universal sentence encoding embedding digits very similar","Body":"<p>I have task of sentence similarity where i calculate the cosine of two sentence to decide how similar they are . It seems that for sentence with digits the similarity is not affected no matter how \"far\" the numbers are . For an example:<\/p>\n\n<p>a = generate_embedding('issue 845')<\/p>\n\n<p>b = generate_embedding('issue 11')<\/p>\n\n<p>cosine_sim(a,b) = 0.9307<\/p>\n\n<p>is there a way to distance the hashing of numbers or any other hack to handle that issue?<\/p>\n","answers":[{"AnswerId":"57496598","CreationDate":"2019-08-14T14:22:42.697","ParentId":null,"OwnerUserId":"11592902","Title":null,"Body":"<p>If your sentence embedding are produced using the embeddings of individual words (or tokens), then a  hack could be the following: <\/p>\n\n<p>to add dimensions to the word embedding. These dimensions would be set to zero for all non-numeric tokens, and for numeric tokens these dimensions would contain values reflecting the magnitude of the numeric value. It would get a bit mathematical because cosine similarity uses angles, so the extra dimensions added to the embedding would have to reflect the magnitude of the numeric values <em>through larger or smaller angles<\/em>.<\/p>\n\n<p>An easier (workaround) hack would be to extract the numeric values from the sentences using regular expressions and compute their distance and combine that information with the similarity score in order to obtain a new similarity score.<\/p>\n"}]}
+{"QuestionId":57489896,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-14T07:20:12.760","AcceptedAnswerId":null,"OwnerUserId":9835434.0,"Title":"I don't why tflite model converted from savedmodel does not raise unsupported operations exceptions?","Body":"<p>I have made simple mnist model using tensorflow like below and saved it as \nSavedModel format. After that, I converted it to tflite format. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>save_path = os.path.join(os.getcwd(), 'simple_save')\ntf.saved_model.simple_save(sess, '.\/saved_model_output\/', inputs={\"X\": X, \"Y\": Y}, outputs={\"logits\": model})\n<\/code><\/pre>\n\n<p>The problem is that the converted tflite file didn't raise any exception while running even though the model has some unsupported operations like <code>softmax_cross_entropy_with_logits_v2<\/code><\/p>\n\n<p>When I tried to load the tflite format converted from a frozen graph(*.pb), \nthe tflite format did raise exception for unsupported operations while running. <\/p>\n\n<p>my question is, why does the issue only occur in the model from frozen graph?\nDoes tflite_convert magically make unsupported ops in the original model possible to use?<\/p>\n","answers":[]}
+{"QuestionId":57489924,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-08-14T07:22:20.790","AcceptedAnswerId":null,"OwnerUserId":10058141.0,"Title":"How to show estimate ranking in keras?","Body":"<blockquote>\n  <p>Sorry for exclude whole code, because codes are really huge. I want to get information to solve my questions, like \"search for this keyword will help you\" or \"It can't make\".<\/p>\n  \n  <p>I'm really Sorry that I don't know correct word such as probability, same meaning in machine learning. So, I will explain very detail as I can.<\/p>\n<\/blockquote>\n\n<p>I know that \"Basic machine learning mechanism\" is classified with specific weight value, and estimated things by the highest probability. For example, If I input any picture, then program classified with giving weight value, and show result by high estimate probability.(eg: program classified that input data's tag is \"cat\" by 70% probability, \"dog\" by 20% probability, and \"human\" by 1% probability, then show the data's tag is \"cat\".) Is it right?<\/p>\n\n<p>So, Based on my knowledge, I make some code using FRCNN algorithm with keras, and training with image_set. After training image_set models with keras, I did test for new image, which test program act \"overwrite text on photos and print on console\".<\/p>\n\n<p>The console printed like that ...<\/p>\n\n<pre><code>filename.JPEG    #filename\nElapsed time = 0.7687556743621826    #elapsed time\n[('test', 81.2415361404419)]    #[('object tag', 'estimated percentage')]\n<\/code><\/pre>\n\n<p>But, I would like to get more information about more estimate probabilities besides the highest one like upper examples...<\/p>\n\n<pre><code>[(first estimate name : 'cat', 70.23232, second estimate name : 'dog', 20.2535, third estimate name : 'human', 1.23532)];\n<\/code><\/pre>\n\n<p>I searching to show more information to show more estimate probabilities, and found fc layer(Fully-Connected layer), But it cannot help my question.<\/p>\n\n<p>Could you give me some keyword to help my question?<\/p>\n","answers":[]}
+{"QuestionId":57490832,"AnswerCount":1,"Tags":"<python><pytorch><spacy><multiclass-classification>","CreationDate":"2019-08-14T08:20:51.520","AcceptedAnswerId":57598258.0,"OwnerUserId":9770403.0,"Title":"SpaCy - ValueError: operands could not be broadcast together with shapes (1,2) (1,5)","Body":"<p>In relation to the previous post on stackoverflow\n<a href=\"https:\/\/stackoverflow.com\/questions\/57476279\/model-got-multiple-values-for-argument-nr-class-spacy-multi-classification\/57476733#57476733\">Model() got multiple values for argument &#39;nr_class&#39; - SpaCy multi-classification model (BERT integration)<\/a> in which my problem partialy have beed resolved I wanted to share the issue which comes up after implementing the solution.<\/p>\n\n<p>if I take out the <code>nr_class<\/code> argument, I get this error here:<\/p>\n\n<pre><code>ValueError: operands could not be broadcast together with shapes (1,2) (1,5)\n<\/code><\/pre>\n\n<p>I actually thought this would happen because I didn't specify the nr_class\n argument. Is this correct?<\/p>\n\n<p>one more time my code for the multi-class model:<\/p>\n\n<pre><code>nlp = spacy.load('en_pytt_bertbaseuncased_lg')\ntextcat = nlp.create_pipe(\n    'pytt_textcat',\n    config={\n        \"nr_class\":5,\n        \"exclusive_classes\": True,\n    }\n)\nnlp.add_pipe(textcat, last = True)\n\ntextcat.add_label(\"class1\")\ntextcat.add_label(\"class2\")\ntextcat.add_label(\"class3\")\ntextcat.add_label(\"class4\")\ntextcat.add_label(\"class5\")\n<\/code><\/pre>\n\n<p>The code for the training is as follows and is based on the example from here(<a href=\"https:\/\/pypi.org\/project\/spacy-pytorch-transformers\/\" rel=\"noreferrer\">https:\/\/pypi.org\/project\/spacy-pytorch-transformers\/<\/a>):<\/p>\n\n<pre><code>def extract_cat(x):\n    for key in x.keys():\n        if x[key]:\n            return key\n\n# get names of other pipes to disable them during training\nn_iter = 250 # number of epochs\n\ntrain_data = list(zip(train_texts, [{\"cats\": cats} for cats in train_cats]))\n\n\ndev_cats_single   = [extract_cat(x) for x in dev_cats]\ntrain_cats_single = [extract_cat(x) for x in train_cats]\ncats = list(set(train_cats_single))\nrecall = {}\nfor c in cats:\n    if c is not None: \n        recall['dev_'+c] = []\n        recall['train_'+c] = []\n\n\n\noptimizer = nlp.resume_training()\nbatch_sizes = compounding(1.0, round(len(train_texts)\/2), 1.001)\n\nfor i in range(n_iter):\n    random.shuffle(train_data)\n    losses = {}\n    batches = minibatch(train_data, size=batch_sizes)\n    for batch in batches:\n        texts, annotations = zip(*batch)\n        nlp.update(texts, annotations, sgd=optimizer, drop=0.2, losses=losses)\n    print(i, losses)\n<\/code><\/pre>\n\n<p>So the structure of my data looks like this:<\/p>\n\n<pre><code>[('TEXT TEXT TEXT',\n  {'cats': {'class1': False,\n    'class2': False,\n    'class3': False,\n    'class4': True,\n    'class5': False}}), ... ]\n<\/code><\/pre>\n","answers":[{"AnswerId":"57598258","CreationDate":"2019-08-21T19:42:21.177","ParentId":null,"OwnerUserId":"9770403","Title":null,"Body":"<p>As @Milla Well already commented the answer can be found <a href=\"https:\/\/github.com\/explosion\/spacy-pytorch-transformers\/commit\/1adeac6ecfb37372bb396947beaf2ad60da563ed\" rel=\"nofollow noreferrer\">here<\/a> (the bug fix on github from @syllogism_)<\/p>\n"}]}
+{"QuestionId":57490850,"AnswerCount":0,"Tags":"<python-3.x><tensorflow>","CreationDate":"2019-08-14T08:22:01.777","AcceptedAnswerId":null,"OwnerUserId":11925237.0,"Title":"Why do I get all mask values in masked_conv2d is 1 in inception_resnet_v2 model? How to fix it","Body":"<p>I build a CNN models using <code>inception_resnet_v2<\/code> net, which is a deep neural network. However, it works slow, so I want to use model prune method to improve it. But the mask values are all 1. I dont'know which error in it . <\/p>\n\n<ol>\n<li><p>I have tried the demo provided in <a href=\"https:\/\/stackoverflow.com\/questions\/52064450\/how-to-use-tf-contrib-model-pruning-on-mnist\">How to use tf.contrib.model_pruning on MNIST<\/a>, and I can get a mask array consisting of 0 and 1, which is correct.<\/p><\/li>\n<li><p>The methods in our model is similar to Minist_data, using <code>masked_conv2d<\/code> and masked_fully_connected to prune, but only get the elements in mask are all 1 <\/p><\/li>\n<\/ol>\n\n<p>For example:<\/p>\n\n<pre><code>with tf.variable_scope('Mixed_6a'):\n                    with tf.variable_scope('Branch_0'):\n    tower_conv = layers.masked_conv2d(net, 384, 3, stride=2, padding='VALID',scope='Conv2d_1a_3x3')\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57490945,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-08-14T08:29:15.513","AcceptedAnswerId":57493117.0,"OwnerUserId":1686236.0,"Title":"Pytorch Train & Eval Different Sample Sizes","Body":"<p>I am learning pytorch, and have the following (abbreviated) code to setup for modeling:<\/p>\n\n<pre><code># define the model class for a neural net with 1 hidden layer\nclass myNN(nn.Module):\n    def __init__(self, D_in, H, D_out):\n        super(myNN, self).__init__()\n        self.lin1 = nn.Linear(D_in,H)\n        self.lin2 = nn.Linear(H,D_out)\n    def forward(self,X):\n        return torch.sigmoid(self.lin2(torch.sigmoid(self.lin1(x))))\n\n# now make the datasets &amp; dataloaders\nbatchSize = 5\n# Create the data class\nclass Data(Dataset):\n    def __init__(self, x, y):\n        self.x = torch.FloatTensor(x)\n        self.y = torch.Tensor(y.astype(int))\n        self.len = self.x.shape[0]\n        self.p = self.x.shape[1]\n    def __getitem__(self, index):      \n        return self.x[index], self.y[index]\n    def __len__(self):\n        return self.len\ntrainData = Data(trnX, trnY)\ntrainLoad = DataLoader(dataset = trainData, batch_size = batchSize)\ntestData = Data(tstX, tstY)\ntestLoad = DataLoader(dataset = testData, batch_size = len(testData))\n\n# define the modeling objects\nhiddenLayers = 30\nlearningRate = 0.1\nmodel = myNN(p,hiddenLayers,1)\nprint(model)\noptimizer = torch.optim.SGD(model.parameters(), lr = learningRate)\nloss = nn.BCELoss()\n<\/code><\/pre>\n\n<p>with <code>trnX.shape=(70, 2)<\/code>, <code>trnY.shape=(70,)<\/code>, <code>tstX.shape=(30,2)<\/code>, and <code>tstY.shape=(30,)<\/code>. The code for training is:<\/p>\n\n<pre><code># train!\nepochs = 1000\ntalkFreq = 0.2\ntrnLoss = [np.inf]*epochs\ntstLoss = [np.inf]*epochs\nfor i in range(epochs):\n    # train with minibatch gradient descent\n    for x, y in trainLoad:\n        # forward step\n        yhat = model(x)\n        # compute loss (not storing for now, will do after minibatching)\n        l = loss(yhat, y)\n        # backward step\n        optimizer.zero_grad()\n        l.backward()\n        optimizer.step()\n    # evaluate loss on training set\n    yhat = model(trainData.x)\n    trnLoss[i] = loss(yhat, trainData.y)\n    # evaluate loss on testing set\n    yhat = model(testData.x)\n    tstLoss[i] = loss(yhat, testData.y)\n<\/code><\/pre>\n\n<p>The datasets <code>trainData<\/code> and <code>testData<\/code> have 70 and 30 observations, respectively. This is probably just a newbie problem, but when I run the training cell, it errors on the <code>trnLoss[i] = loss(yhat, trainData.y)<\/code> line with the error<\/p>\n\n<pre><code>ValueError: Target and input must have the same number of elements. target nelement (70) != input nelement (5)\n<\/code><\/pre>\n\n<p>When I inspect the output of the <code>yhat=model(trainData.x)<\/code> line, I see that <code>yhat<\/code> is a tensor with <code>batchSize<\/code> elements, despite the fact that <code>trainData.x.shape = torch.Size([70, 2])<\/code>.<\/p>\n\n<p>How can I iteratively train the model with the minibatch gradient descent, then use the model to compute the loss &amp; accuracy on the full training &amp; testing sets? I tried setting <code>model.train()<\/code> just before the mini batch iteration, then <code>model.eval()<\/code> just before the evaluation code, to no avail.<\/p>\n","answers":[{"AnswerId":"57493117","CreationDate":"2019-08-14T10:41:58.877","ParentId":null,"OwnerUserId":"5027315","Title":null,"Body":"<p>In <code>myNN.forward()<\/code>, you are passing lower case <code>x<\/code> as input to <code>self.lin1<\/code> while input parameter to the function is named as capital case <code>X<\/code>. Small case <code>x<\/code> is kind of global variable defined at for loop for <code>trainload<\/code> hence you are not getting any syntax error, but the value that you intend to pass is not being passed to <code>self.lin1<\/code>.<\/p>\n\n<p>Might I also suggest that you should consider using <code>model.eval()<\/code> and <code>with torch.no_grad()<\/code> for your testing code. It is not absolutely necessary here but it will make more sense.<\/p>\n"}]}
+{"QuestionId":57491279,"AnswerCount":1,"Tags":"<python><torch>","CreationDate":"2019-08-14T08:50:53.667","AcceptedAnswerId":null,"OwnerUserId":11926303.0,"Title":"torch.utils.ffi is deprecated. How to use cpp extensions instead?","Body":"<p>I am trying to run code that is using wrap_function from torch.utils.ffi (which has deprecated). I am struggling to figure out how to use cpp extensions instead as suggested by the error message, can any anyone help please?<\/p>\n\n<p>The code I need to replace: <\/p>\n\n<pre><code>from torch.utils.ffi import _wrap_function\nfrom ._nms import lib as _lib, ffi as _ffi\n\n__all__ = []\ndef _import_symbols(locals):\n    for symbol in dir(_lib):\n        fn = getattr(_lib, symbol)\n        if callable(fn):\n            locals[symbol] = _wrap_function(fn, _ffi)\n        else:\n            locals[symbol] = fn\n        __all__.append(symbol)\n\n_import_symbols(locals())\n<\/code><\/pre>\n\n<p>I have tried running the code in both python 3.6 and python 2.7, see error message below,<\/p>\n\n<pre><code>  File \"build.py\", line 3, in &lt;module&gt;\n    from torch.utils.ffi import create_extension\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/torch\/utils\/ffi\/__init__.py\", line 1, in &lt;module&gt;\n    raise ImportError(\"torch.utils.ffi is deprecated. Please use cpp extensions instead.\")\nImportError: torch.utils.ffi is deprecated. Please use cpp extensions instead.\n<\/code><\/pre>\n","answers":[{"AnswerId":"57491414","CreationDate":"2019-08-14T08:59:55.430","ParentId":null,"OwnerUserId":"11864933","Title":null,"Body":"<p>Check<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from torch.utils.cpp import ....\n#or\nfrom torch.utils.cpp_extension import ....\n#or\nfrom torch.utils.cpp_extension_versioner import...\n<\/code><\/pre>\n\n<p>it looks good at my PC<\/p>\n"}]}
+{"QuestionId":57491610,"AnswerCount":1,"Tags":"<python><windows><dll><pytorch><pyinstaller>","CreationDate":"2019-08-14T09:10:37.783","AcceptedAnswerId":null,"OwnerUserId":9584818.0,"Title":"Pyinstaller unable to find dlls for the project dependencies while creating exe","Body":"<p>Pyinstaller failed to find certain dlls that are required for binding in dependencies into one exe.<\/p>\n\n<p>Please find the error logs below.<\/p>\n\n<p>We have tried installing these libraries:<\/p>\n\n<pre><code>pip3 install intel-openmp mkl\n\n<\/code><\/pre>\n\n<p>Tried adding --paths to the command, but as there are no dlls in the system, pyinstaller is unable to find them:<\/p>\n\n<pre><code>pyinstaller --onefile --paths &lt;Paths-where-dll-could-be&gt; -c main.py\n<\/code><\/pre>\n\n<p>These libs are missing and showed in logs as WARNINGS.<\/p>\n\n<pre><code>364427 WARNING: lib not found: impi.dll dependency of c:\\users\\1311654\\appdata\\local\\programs\\python\\python36\\Library\\bin\\mkl_blacs_intelmpi_ilp64.dll\n365396 WARNING: lib not found: mpich2mpi.dll dependency of c:\\users\\1311654\\appdata\\local\\programs\\python\\python36\\Library\\bin\\mkl_blacs_mpich2_lp64.dll\n366241 WARNING: lib not found: msmpi.dll dependency of c:\\users\\1311654\\appdata\\local\\programs\\python\\python36\\Library\\bin\\mkl_blacs_msmpi_lp64.dll\n368089 WARNING: lib not found: msmpi.dll dependency of c:\\users\\1311654\\appdata\\local\\programs\\python\\python36\\Library\\bin\\mkl_blacs_msmpi_ilp64.dll\n369270 WARNING: lib not found: pgf90.dll dependency of c:\\users\\1311654\\appdata\\local\\programs\\python\\python36\\Library\\bin\\mkl_pgi_thread.dll\n369997 WARNING: lib not found: pgc14.dll dependency of c:\\users\\1311654\\appdata\\local\\programs\\python\\python36\\Library\\bin\\mkl_pgi_thread.dll\n370791 WARNING: lib not found: pgf90rtl.dll dependency of c:\\users\\1311654\\appdata\\local\\programs\\python\\python36\\Library\\bin\\mkl_pgi_thread.dll\n373039 WARNING: lib not found: mpich2mpi.dll dependency of c:\\users\\1311654\\appdata\\local\\programs\\python\\python36\\Library\\bin\\mkl_blacs_mpich2_ilp64.dll\n374289 WARNING: lib not found: impi.dll dependency of c:\\users\\1311654\\appdata\\local\\programs\\python\\python36\\Library\\bin\\mkl_blacs_intelmpi_lp64.dll\n377030 WARNING: lib not found: torch_python.dll dependency of c:\\users\\1311654\\appdata\\local\\programs\\python\\python36\\lib\\site-packages\\torch\\_C.cp36-win_amd64.pyd\n378792 WARNING: lib not found: c10_cuda.dll dependency of c:\\users\\1311654\\appdata\\local\\programs\\python\\python36\\lib\\site-packages\\torchvision\\_C.cp36-win_amd64.pyd\n379568 WARNING: lib not found: torch.dll dependency of c:\\users\\1311654\\appdata\\local\\programs\\python\\python36\\lib\\site-packages\\torchvision\\_C.cp36-win_amd64.pyd\n380290 WARNING: lib not found: caffe2.dll dependency of c:\\users\\1311654\\appdata\\local\\programs\\python\\python36\\lib\\site-packages\\torchvision\\_C.cp36-win_amd64.pyd\n381126 WARNING: lib not found: c10.dll dependency of c:\\users\\1311654\\appdata\\local\\programs\\python\\python36\\lib\\site-packages\\torchvision\\_C.cp36-win_amd64.pyd\n382053 WARNING: lib not found: torch_python.dll dependency of c:\\users\\1311654\\appdata\\local\\programs\\python\\python36\\lib\\site-packages\\torchvision\\_C.cp36-win_amd64.pyd\n\n<\/code><\/pre>\n\n<p>As the missing lib dlls are not in the system, kindly assist the efficient way to build the exe.<\/p>\n","answers":[{"AnswerId":"57604856","CreationDate":"2019-08-22T08:12:07.907","ParentId":null,"OwnerUserId":"10685311","Title":null,"Body":"<p>Windows searches for DLLs in the following locations (I may not have the order right):<\/p>\n\n<p>-Current directory<\/p>\n\n<p>-Directory containing the EXE that asked for the DLL (directly or indirectly)<\/p>\n\n<p>-Directories in PATH<\/p>\n\n<p>-Windows directory<\/p>\n\n<p>-Windows System directory<\/p>\n\n<ol>\n<li>If you have Visual Studio installed on your computer, use dumpbin to figure out what \nyour DLL depends on. (dumpbin is installed here C:\\Program Files (x86)\\Microsoft \nVisual Studio 14.0\\VC\\bin\\dumpbin.exe)<\/li>\n<\/ol>\n\n<blockquote>\n  <p>dumpbin \/DEPENDENTS your.dll<\/p>\n<\/blockquote>\n\n<p>this will list the DLLs dependencies. Make sure they are included in the same directory \n  as your exe.<\/p>\n\n<ol start=\"2\">\n<li><p>If you don't have Visual Studio or dumpbin, download <a href=\"http:\/\/www.dependencywalker.com\/\" rel=\"nofollow noreferrer\">http:\/\/www.dependencywalker.com\/<\/a> \nand it will accomplish the same thing<\/p><\/li>\n<li><p>Make sure you have the most recent version of pyinstaller as well.<\/p><\/li>\n<\/ol>\n\n<p>Hope this helps.<\/p>\n"}]}
+{"QuestionId":57491736,"AnswerCount":0,"Tags":"<tensorflow><perfect>","CreationDate":"2019-08-14T09:19:08.280","AcceptedAnswerId":null,"OwnerUserId":9663497.0,"Title":"Swift for TensorFlow + Keras vs. Perfect TensorFlow wrapper","Body":"<p>As I'm starting diving into AI, machine learning and deep learning, I pretty much understood that I gotta learn python.. Swift for tensor flow makes the combo Tensor Flow + Keras the one that I'd be able to get under my belt in a reasonable time. While gettin interested in server-side with swift ( to basically stop using Firebase ) I discovered Perfect framework that also has a Tensor Flow wrapper and dough it's stated as sperimental, that would make it the perfect solution for both deep learning and server-side. My concern is about languages versions and compatibilities as it seems that swift 4 and python 2 are the latest supported versions in Perfect, but in Swift for tensor flow you only have betas supperted... Also do you know if with Perfect framework is possible to train models or are you restricted to use only pre trained models? Thanks for helping me deciding the right route to take as I'm a bit lost on the subject.<\/p>\n","answers":[]}
+{"QuestionId":57491895,"AnswerCount":0,"Tags":"<python><machine-learning><keras><conv-neural-network><resnet>","CreationDate":"2019-08-14T09:27:27.850","AcceptedAnswerId":null,"OwnerUserId":4494762.0,"Title":"Keras ResNet -- Val acc not improving, val loss fluctuates, even data set","Body":"<p>Validation accuracy stays constant after ever epoch; however, the validation loss fluctuates. The model is hyper overfitting at the moment with a training accuracy of ~90% after 3-4 epochs. The current validation accuracy is around the % of random guessing for the higher percentage class.<\/p>\n\n<p>Tensorflow Version: 1.14.0<\/p>\n\n<p>Hyperparameter \/ other configs:<br>\n--Images augmented with shifts \/ reflections<br>\n--Pretrained weights (ImageNet)<br>\n--Changed top layers to include BatchNorm and Dropout<br>\n--About 5000 training images, 1000 validation images, properly shuffled<br>\n--Softmax classification (binary, 2 classes)<br>\n--Labels are correct<br>\n--Learning rates values tried: 0.01, 0.001, 0.0001<br>\n--RMSProp optimizer<br>\n--Categorical cross-entropy (log loss)<br>\n--Class weights applied proportional to the amount of data in each class (and there isn't much of an imbalance to begin with...)   <\/p>\n\n<p>Code:<\/p>\n\n<pre><code>base_model = ResNet50(weights=\"imagenet\", include_top=False, input_shape=(128,128,3))\nfinetune_model.compile(optimizer=Adam(lr=0.001), loss=\"sparse_categorical_crossentropy\", metrics=[\"acc\"])\nhistory = finetune_model.fit_generator(train_data, validation_data=valid_data, epochs=100, class_weight=class_weights, shuffle=True, callbacks=[tensorboard_callback])\n<\/code><\/pre>\n\n<p>Here's the code for the layers that use BatchNorm and Dropout:<\/p>\n\n<pre><code>output = Dense(layer_size, name = 'Dense_{}'.format(i))(output)\noutput = BatchNormalization()(output)\noutput = Activation(\"relu\")(output)\noutput = Dropout(0.5)(output)\n\npredictions = Dense(2, activation=\"softmax\")(output)\nfinetune_model = Model(inputs=base_model.input, outputs=predictions)\n<\/code><\/pre>\n\n<p>To give you all some idea of the output (sorry, it is a nightmare for me to run on my machine, but it basically exhibits the same behavior even when I test past 25 epochs):<\/p>\n\n<pre><code>Epoch 1\/100\n149\/149 [==============================] - 356s 2s\/step - loss: 0.7069 - categorical_accuracy: 0.7681 - val_loss: 2.4147 - val_categorical_accuracy: 0.5529\nEpoch 2\/100\n149\/149 [==============================] - 387s 3s\/step - loss: 0.4370 - categorical_accuracy: 0.8465 - val_loss: 1.0831 - val_categorical_accuracy: 0.5529\nEpoch 3\/100\n149\/149 [==============================] - 364s 2s\/step - loss: 0.3708 - categorical_accuracy: 0.8707 - val_loss: 0.7329 - val_categorical_accuracy: 0.5529\nEpoch 4\/100\n149\/149 [==============================] - 383s 3s\/step - loss: 0.3355 - categorical_accuracy: 0.8898 - val_loss: 0.7157 - val_categorical_accuracy: 0.5529\nEpoch 5\/100\n149\/149 [==============================] - 332s 2s\/step - loss: 0.2776 - categorical_accuracy: 0.9075 - val_loss: 0.6897 - val_categorical_accuracy: 0.5529\nEpoch 6\/100\n149\/149 [==============================] - 349s 2s\/step - loss: 0.2287 - categorical_accuracy: 0.9220 - val_loss: 0.6855 - val_categorical_accuracy: 0.5529\nEpoch 7\/100\n149\/149 [==============================] - 326s 2s\/step - loss: 0.2196 - categorical_accuracy: 0.9319 - val_loss: 0.7273 - val_categorical_accuracy: 0.5529\nEpoch 8\/100\n149\/149 [==============================] - 331s 2s\/step - loss: 0.1782 - categorical_accuracy: 0.9451 - val_loss: 0.8641 - val_categorical_accuracy: 0.5529\nEpoch 9\/100\n149\/149 [==============================] - 319s 2s\/step - loss: 0.1774 - categorical_accuracy: 0.9466 - val_loss: 0.8990 - val_categorical_accuracy: 0.5529\nEpoch 10\/100\n149\/149 [==============================] - 331s 2s\/step - loss: 0.1653 - categorical_accuracy: 0.9495 - val_loss: 0.9257 - val_categorical_accuracy: 0.5529\n<\/code><\/pre>\n\n<p>I'd like to try to decrease my validation loss and address the constant accuracy problem. Any thoughts?<\/p>\n","answers":[]}
+{"QuestionId":57491972,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><database-normalization><object-detection-api><pre-trained-model>","CreationDate":"2019-08-14T09:31:05.467","AcceptedAnswerId":null,"OwnerUserId":10949381.0,"Title":"normalization image before using ssd_mobilenet","Body":"<p>I try train ssd-mobilenet in my own dataset : \ntraining image : 3400 with size :1600*1200 \ntest set :800 with size :1600 *1200 tensorflow -gpu :1.13.1 gpu :4GB cuda 10.0 cudnn 7 \nobject: road damage like aligator crack but after 197000 step my training loss cannot go down 2.\nI have 2 questions\nShould I normalize my training and set image before before using pretrained model like ssd_mobilenet?\nIf yes\nShould I annotate images normalized or not ?\nI need really helps. Thanks in advance<\/p>\n","answers":[{"AnswerId":"57492185","CreationDate":"2019-08-14T09:43:26.143","ParentId":null,"OwnerUserId":"3214872","Title":null,"Body":"<blockquote>\n  <p>Should I normalize my training and set image before before using pretrained model like ssd_mobilenet?<\/p>\n<\/blockquote>\n\n<p>No. Assuming you define your training pipeline correctly (see the examples in the TF Models repository), the Object detection API will take care of defining the appropriate image transformations (scaling, padding, normalization, etc) required in order to make the input compatible with the model.<\/p>\n"}]}
+{"QuestionId":57492139,"AnswerCount":1,"Tags":"<python><tensorflow><gcloud>","CreationDate":"2019-08-14T09:40:46.423","AcceptedAnswerId":57515379.0,"OwnerUserId":11926586.0,"Title":"How to add 'instance keys' to a keras model input for batch prediction in gcloud ai-platform?","Body":"<p>I'm trying to add 'keys' to match the batch prediction output from Google AI Platform, however my model input just allows for one input.<\/p>\n\n<p>It looks like that:<\/p>\n\n<pre><code>input = tf.keras.layers.Input(shape=(max_len,))\n\nx = tf.keras.layers.Embedding(max_words, embed_size, weights=[embedding_matrix], trainable=False)(input)\nx = tf.keras.layers.Bidirectional(tf.keras.layers.GRU(128, return_sequences=True, dropout=0.3, recurrent_dropout=0.1))(x)\nx = tf.keras.layers.Conv1D(128, kernel_size=3, padding=\"valid\", kernel_initializer=\"glorot_uniform\")(x)\n\navg_pool = tf.keras.layers.GlobalAveragePooling1D()(x)\nmax_pool = tf.keras.layers.GlobalMaxPooling1D()(x)\n\nx = tf.keras.layers.concatenate([avg_pool, max_pool])\n\npreds = tf.keras.layers.Dense(2, activation=\"sigmoid\")(x)\nmodel = tf.keras.Model(input, preds)\nmodel.summary()\nmodel.compile(loss='binary_crossentropy', optimizer=tf.keras.optimizers.Adam(lr=1e-3), metrics=['accuracy','binary_crossentropy'])\n<\/code><\/pre>\n\n<p>I came across <a href=\"https:\/\/towardsdatascience.com\/how-to-extend-a-canned-tensorflow-estimator-to-add-more-evaluation-metrics-and-to-pass-through-ddf66cd3047d\" rel=\"nofollow noreferrer\">this article<\/a>, but can't figure out how to apply this to my code.<\/p>\n\n<p>Any ideas? Thanks!<\/p>\n","answers":[{"AnswerId":"57515379","CreationDate":"2019-08-15T19:49:22.020","ParentId":null,"OwnerUserId":"7517757","Title":null,"Body":"<p>Following your code, you could do something like this:<\/p>\n\n<p>First, get the key value from the input:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>input = tf.keras.layers.Input(shape=(max_len,))\nkey_raw = tf.keras.layers.Input(shape=(), name='key')\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/Reshape#__init__\" rel=\"nofollow noreferrer\">Reshape<\/a> it to use later for concatenation<\/p>\n\n<pre><code>key = tf.keras.layers.Reshape((1,), input_shape=())(key_raw)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/Concatenate#top_of_page\" rel=\"nofollow noreferrer\">Concatenate<\/a> the key with the final result<\/p>\n\n<pre><code>preds = tf.keras.layers.Dense(2, activation=\"sigmoid\")(x)\npreds = tf.keras.layers.concatenate([preds, key])\n<\/code><\/pre>\n\n<p>Add it to the input of your model<\/p>\n\n<pre><code>model = tf.keras.Model([input, key_raw], preds)\n<\/code><\/pre>\n\n<p>Input json file example:<\/p>\n\n<pre><code>{\"input_1\": [1.2,1.1,3.3,4.3], \"key\":1}\n{\"input_1\": [0.3, 0.4, 1.5, 1], \"key\":2}\n<\/code><\/pre>\n\n<p>Now, you can get the key as the last element of the prediction result.\nOutput example:<\/p>\n\n<pre><code>[0.48686566948890686, 0.5113844275474548, 1.0]\n[0.505149781703949, 0.5156428813934326, 2.0]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57492493,"AnswerCount":0,"Tags":"<python><tensorflow><onnx>","CreationDate":"2019-08-14T10:01:26.190","AcceptedAnswerId":null,"OwnerUserId":2828086.0,"Title":"From TensorFlow to ONNX - save\/freeze graph or convert pkl?","Body":"<p>Newbie question on the best way to go from TensorFlow to ONNX: what is the better (and\/or easier) way between the two listed below?<\/p>\n\n<ol>\n<li><p>Freeze\/save the network --> store a .pb file --> convert the .pb into .onnx (I am struggling with this)<\/p><\/li>\n<li><p>Convert a .pkl snapshot into .onnx. I haven't found any material on this, so any tip is welcome!<\/p><\/li>\n<\/ol>\n","answers":[]}
+{"QuestionId":57492988,"AnswerCount":0,"Tags":"<python><flask><pytorch><travis-ci><amazon-elastic-beanstalk>","CreationDate":"2019-08-14T10:32:57.260","AcceptedAnswerId":null,"OwnerUserId":7462428.0,"Title":"Deployed Pytorch model in a Flask app is throwing Error 500 in AWS Elastic Beanstalk while deployed using Travis CI","Body":"<p>I trained a PyTorch BERT model, wrapped it in a Python Flask app and deployed it in AWS EBS using Travis CI, the build and deployment were successful in Travis CI, but testing it in AWS EBS it is throwing 500 Internal Server Error.<\/p>\n\n<p>I initially tested a simple Flask application with only a GET method and deployed to EBS  using Travis and the application ran successfully on EBS.<\/p>\n\n<p>But after deploying the model to EBS it started throwing 500 error. The actual model is saved in S3 and it will be downloaded from S3 and called by the application. <\/p>\n\n<p>Moreover, since I used PyTorch CPU version, I installed PyTorch's this wheel `<a href=\"https:\/\/download.pytorch.org\/whl\/cpu\/torch-1.0.1.post2-cp36-cp36m-linux_x86_64.whl\" rel=\"nofollow noreferrer\">https:\/\/download.pytorch.org\/whl\/cpu\/torch-1.0.1.post2-cp36-cp36m-linux_x86_64.whl<\/a>' in EBS. I have skipped model class and dataloader initialization in the code section.<\/p>\n\n<pre><code># application.py\n\n# Import necessary packages\nimport numpy as np\nimport torch\nimport torch.nn as nn\nimport matplotlib.pyplot as plt\nimport os\nimport copy\nfrom torch.utils.data import Dataset, DataLoader\nimport torch.nn.functional as F\nfrom pytorch_pretrained_bert import BertTokenizer, BertModel, BertConfig\nimport json\nimport boto3\nfrom flask import Flask\n\n# Create a S3 client\nclient = boto3.client('s3')\n\nBUCKET_NAME = 'elasticbeanstalk-us-east-2-458375213711'\n\nMODEL_PATH = 'best_model.pth'\n\nclient.download_file(BUCKET_NAME, 'best_model_2.pth', MODEL_PATH)\n\n\n# Find the device information\ndevice = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu:0')\n.\n# model class initialization\n.\nmodel.load_state_dict(torch.load(MODEL_PATH))\nmodel.eval()\n\n# Create a tokenzier object from the BertTokenizer with the weights 'bert-base-uncased'\ntokenizer = BertTokenizer.from_pretrained('bert-base-uncased', do_lower_case=True)\n\nmax_seq_len = 256\n\n.\n# DataLoader class initialization\n.\n\ndef load_data(data):\n  test_list = [[data], []]\n\n  # Create the train and test dataset using the Text_Dataset class\n\n  test_data = Text_Dataset(test_list)\n\n  dataloader_infer = DataLoader(test_data, batch_size=1, num_workers=1)\n  return dataloader_infer\n\ndef infer(data):\n    dataloader = load_data(data)\n    for inputs, attention_mask in dataloader:\n      # Load the device information to the inputs\n      inputs = inputs.to(device)\n      attention_mask = attention_mask.to(device)\n      outputs = model(inputs, attention_mask=attention_mask)\n      # Softmax the output\n      outputs = F.softmax(outputs, dim=1)\n      med = outputs.detach().cpu().numpy()\n      #print(med.tolist())\n      #print(type(med.tolist()))\n      result = np.argmax(outputs.detach().cpu().numpy())\n      return result\n\n\napplication = Flask(__name__)\n\n@application.route('\/', methods=['GET', 'POST'])\ndef nlp():\n  if request.method == 'GET':\n    return {'Message': 'Successfully Deployed. Ready for serving'}\n\n  if request.method == 'POST':\n    #temp = False\n    string_data = json.loads(request.data)\n    string = string_data['input_string']\n    result = infer(string)\n    if result:\n      output = {'Text Review': 'Positive'}\n    else:\n      output = {'Text Review': 'Negative'}\n    return jsonify(output)\n\nif __name__ == '__main__':\n    application.run(debug=True)\n\n\n<\/code><\/pre>\n\n<pre><code># .travis.yml\nlanguage: python\npython:\n  - \"3.6\"\ncache: pip\ninstall: \n  - pip install -r requirements.txt\nscript: \n  - echo \"Running successfully\"\ndeploy:\n  provider: elasticbeanstalk\n  access_key_id: \n    secure: \"$ACCESS_ID\"\n  secret_access_key:\n    secure: \"$SECRET_KEY\"\n  region: \"us-east-2\"  \n  app: \"flask_new\"\n  env: \"FlaskNew-env\"\n  bucket_name: \"elasticbeanstalk-us-east-2-458375213711\"\n\n<\/code><\/pre>\n\n<p>For the GET model the expected result is 'Successfully Deployed. Ready for serving' and for POST model it is either 'Postive' or 'Negative' messages.<\/p>\n\n<p>The error is as follows:<\/p>\n\n<pre><code>Internal Server Error\nThe server encountered an internal error or misconfiguration and was unable to complete your request.\n\nPlease contact the server administrator at root@localhost to inform them of the time this error occurred, and the actions you performed just before this error.\n\nMore information about this error may be available in the server error log.\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57493026,"AnswerCount":0,"Tags":"<docker><tensorflow><anaconda><open-source><miniconda>","CreationDate":"2019-08-14T10:35:36.957","AcceptedAnswerId":null,"OwnerUserId":5414793.0,"Title":"Licencing of TensorFlow conda distribution in miniconda docker image","Body":"<p>I am using conda distribution of Tensorflow inside miniconda3 docker image.\nminiconda3 docker image i have downloaded from below link:- <\/p>\n\n<p><a href=\"https:\/\/hub.docker.com\/r\/continuumio\/miniconda3\" rel=\"nofollow noreferrer\">https:\/\/hub.docker.com\/r\/continuumio\/miniconda3<\/a><\/p>\n\n<p>Installation of tensorflow &amp; other dependencies in docker image:<\/p>\n\n<pre><code>FROM continuumio\/miniconda3\nRUN apt-get update &amp;&amp; \\\napt-get install -y openjdk-8-jre-headless &amp;&amp; \\\napt-get install -y ant &amp;&amp; \\\napt-get install -y python3-tk &amp;&amp; \\\napt-get clean;\nconda install --yes tensorflow==1.12\n<\/code><\/pre>\n\n<p>I created another docker image on the top of miniconda3 docker image.<\/p>\n\n<p>Scan this final image in whitesource tool for getting information of GPL and AGPL licence. <\/p>\n\n<p>I am getting lot's of GPL and AGPL inside tensorflow package folder.<\/p>\n\n<p>Below is the example of one package:-<\/p>\n\n<pre><code>Slic3r-1.41.0-alpha3.1    GPL 2.0AGPL 3.0GPL    opt\\conda\\lib\\python3.6\\site-packages\\tensorflow\\include\\Eigen\\src\\Core\\Assign_MKL.h\n<\/code><\/pre>\n\n<p>Now i want to use this in our Enterprise application. is there any other way to use this in application or this the issue of tool to give GPL inside package?<\/p>\n","answers":[]}
+{"QuestionId":57493043,"AnswerCount":0,"Tags":"<python><tensorflow><word-embedding>","CreationDate":"2019-08-14T10:36:44.947","AcceptedAnswerId":null,"OwnerUserId":2354908.0,"Title":"Tensoflow error while using BERT-as-Service","Body":"<p>I'm trying to use BERT-as-service as provided in this tutorial. I'm trying to use the server through the python env <a href=\"https:\/\/github.com\/hanxiao\/bert-as-service#starting-bertserver-from-python\" rel=\"nofollow noreferrer\">LINK<\/a>. <\/p>\n\n<pre><code>from bert_serving.server.helper import get_args_parser\nfrom bert_serving.server import BertServer\nfrom bert_serving.client import BertClient\nprint('start bert-service')\n\nargs = get_args_parser().parse_args(['-model_dir', 'C:\/Users\/Shashishekar.R\/PycharmProjects\/DefinedTerms\/resources\/uncased_L-12_H-768_A-12\/',\n                                     '-port', '5555',\n                                     '-port_out', '5556',\n                                     '-max_seq_len', 'NONE',\n                                     '-mask_cls_sep',\n                                     '-cpu'])\nserver = BertServer(args)\nserver.start()\nprint('start bert-client')\nbc = BertClient()\nbc.encode(['First do it', 'then do it right', 'then do it better'])\n\nBertServer.shutdown(port=5555)\n<\/code><\/pre>\n\n<p>It throws an error :<\/p>\n\n<blockquote>\n  <p>C:\\Users\\UserABCD\\Anaconda_Envs\\user_env\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:526:\n  FutureWarning: Passing (type, 1) or '1type' as a synonym of type is\n  deprecated; in a future version of numpy, it will be understood as\n  (type, (1,)) \/ '(1,)type'.   _np_qint8 = np.dtype([(\"qint8\", np.int8,\n  1)])\n  C:\\Users\\UserABCD\\Anaconda_Envs\\user_env\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:527:\n  FutureWarning: Passing (type, 1) or '1type' as a synonym of type is\n  deprecated; in a future version of numpy, it will be understood as\n  (type, (1,)) \/ '(1,)type'.   _np_quint8 = np.dtype([(\"quint8\",\n  np.uint8, 1)])\n  C:\\Users\\UserABCD\\Anaconda_Envs\\user_env\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:528:\n  FutureWarning: Passing (type, 1) or '1type' as a synonym of type is\n  deprecated; in a future version of numpy, it will be understood as\n  (type, (1,)) \/ '(1,)type'.   _np_qint16 = np.dtype([(\"qint16\",\n  np.int16, 1)])\n  C:\\Users\\UserABCD\\Anaconda_Envs\\user_env\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:529:\n  FutureWarning: Passing (type, 1) or '1type' as a synonym of type is\n  deprecated; in a future version of numpy, it will be understood as\n  (type, (1,)) \/ '(1,)type'.   _np_quint16 = np.dtype([(\"quint16\",\n  np.uint16, 1)])\n  C:\\Users\\UserABCD\\Anaconda_Envs\\user_env\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:530:\n  FutureWarning: Passing (type, 1) or '1type' as a synonym of type is\n  deprecated; in a future version of numpy, it will be understood as\n  (type, (1,)) \/ '(1,)type'.   _np_qint32 = np.dtype([(\"qint32\",\n  np.int32, 1)])\n  C:\\Users\\UserABCD\\Anaconda_Envs\\user_env\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:535:\n  FutureWarning: Passing (type, 1) or '1type' as a synonym of type is\n  deprecated; in a future version of numpy, it will be understood as\n  (type, (1,)) \/ '(1,)type'.   np_resource = np.dtype([(\"resource\",\n  np.ubyte, 1)]) Traceback (most recent call last):   File\n  \"C:\/Users\/UserABCD\/PycharmProjects\/projectABC\/BERTasService.py\", line\n  1, in \n      from bert_serving.server.helper import get_args_parser   File \"C:\\Users\\UserABCD\\Anaconda_Envs\\user_env\\lib\\site-packages\\bert_serving\\server__init__.py\",\n  line 29, in \n      _tf_ver_ = check_tf_version()   File \"C:\\Users\\UserABCD\\Anaconda_Envs\\user_env\\lib\\site-packages\\bert_serving\\server\\helper.py\",\n  line 164, in check_tf_version\n      import tensorflow as tf   File \"C:\\Users\\UserABCD\\Anaconda_Envs\\user_env\\lib\\site-packages\\tensorflow__init__.py\",\n  line 24, in \n      from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import   File\n  \"C:\\Users\\UserABCD\\Anaconda_Envs\\user_env\\lib\\site-packages\\tensorflow\\python__init__.py\",\n  line 82, in \n      from tensorflow.python import keras   File \"C:\\Users\\UserABCD\\Anaconda_Envs\\user_env\\lib\\site-packages\\tensorflow\\python\\keras__init__.py\", line 24, in \n      from tensorflow.python.keras import activations   File \"C:\\Users\\UserABCD\\Anaconda_Envs\\user_env\\lib\\site-packages\\tensorflow\\python\\keras\\activations.py\",\n  line 24, in \n      from tensorflow.python.keras.utils.generic_utils import deserialize_keras_object   File\n  \"C:\\Users\\UserABCD\\Anaconda_Envs\\user_env\\lib\\site-packages\\tensorflow\\python\\keras\\utils__init__.py\",\n  line 38, in \n      from tensorflow.python.keras.utils.multi_gpu_utils import multi_gpu_model   File\n  \"C:\\Users\\UserABCD\\Anaconda_Envs\\user_env\\lib\\site-packages\\tensorflow\\python\\keras\\utils\\multi_gpu_utils.py\",\n  line 22, in \n      from tensorflow.python.keras.engine.training import Model   File \"C:\\Users\\UserABCD\\Anaconda_Envs\\user_env\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\",\n  line 42, in \n      from tensorflow.python.keras.engine.network import Network   File \"C:\\Users\\UserABCD\\Anaconda_Envs\\user_env\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\network.py\",\n  line 40, in \n      from tensorflow.python.keras.engine import saving   File \"C:\\Users\\UserABCD\\Anaconda_Envs\\user_env\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\saving.py\",\n  line 38, in \n      import h5py   File \"C:\\Users\\UserABCD\\Anaconda_Envs\\user_env\\lib\\site-packages\\h5py__init__.py\",\n  line 36, in \n      from ._conv import register_converters as _register_converters   File \"h5py\\h5r.pxd\", line 21, in init h5py._conv   File\n  \"h5py\\h5r.pyx\", line 145, in init h5py.h5r AttributeError: type object\n  'h5py.h5r.Reference' has no attribute '<strong>reduce_cython<\/strong>'<\/p>\n<\/blockquote>\n\n<p>I currently have tensorflow 1.3 installed via conda install -c conda-forge tensorflow. I also did the pip install, which installs the tensorflow 1.4<\/p>\n\n<p>Any pointers as to where the error might be ?<\/p>\n","answers":[]}
+{"QuestionId":57493144,"AnswerCount":0,"Tags":"<python><keras><neural-network><lstm><recurrent-neural-network>","CreationDate":"2019-08-14T10:43:16.337","AcceptedAnswerId":null,"OwnerUserId":11712588.0,"Title":"How to use ordinal Data for an LSTM-network?","Body":"<p>I work on an LSTM Network and have different kind of inputdata. Most of them are metric data, for them i use the <code>.shift<\/code> function to get the data t-1 and to calculate the percentage Change for predictions.<\/p>\n\n<p>But besides that i have a column with months (1 for january \u2026 12 for december), so a percentage Change is not useful. How can I include These Kind of data?<\/p>\n\n<p>Below is how I included my metric data :<\/p>\n\n<pre><code>Erzeugung[\"erzeugung_before\"] = Erzeugung[\"solar__-_actual_aggregated_[mw]\"].shift(1)\nErzeugung[\"Temperatur_before\"] = Erzeugung[\"Temperatur\"].shift(1)\nErzeugung[\"Globalstrahlung_before\"] = Erzeugung[\"Globalstrahlung\"].shift(1)\n\nErzeugung = Erzeugung.dropna()\n\nErzeugung[\"erzeugung_changes\"] = (Erzeugung[\"solar__-_actual_aggregated_[mw]\"] \/ Erzeugung[\"erzeugung_before\"] - 1)\nErzeugung[\"Temperatur_changes\"] = (Erzeugung[\"Temperatur\"] \/ Erzeugung[\"Temperatur_before\"] - 1)\nErzeugung[\"Globalstrahlung_changes\"] = (Erzeugung[\"Globalstrahlung\"] \/ Erzeugung[\"Globalstrahlung_before\"] - 1)\n\nErzeugung[\"erzeugung_changes\"] = Erzeugung[\"erzeugung_changes\"].replace(np.nan, 0)\nErzeugung[\"Temperatur_changes\"] = Erzeugung[\"Temperatur_changes\"].replace(np.nan, 0)\nErzeugung[\"Globalstrahlung_changes\"] = Erzeugung[\"Globalstrahlung_changes\"].replace(np.nan, 0)\n\nchanges = Erzeugung[[\"erzeugung_changes\", \"Temperatur_changes\", \"Globalstrahlung_changes\"]].values\n\n\nX = []\nY = []\n\nfor i in range (0, len(changes) - 96):\n    Y.append(changes[i,0])      \n    X.append(np.array(changes[i+1:i+97]))\n\nX = np.array(X).reshape(-1, 96, 3)\nY = np.array(Y)\n\nprint(X.shape)\nprint(Y.shape)\n\nfrom keras.models import Sequential\nfrom keras.layers import LSTM\n\nmodel = Sequential()\nmodel.add(LSTM(1, input_shape=(96, 3)))\n\nmodel.compile(optimizer=\"rmsprop\", loss=\"mse\")\nmodel.fit(X, Y, batch_size=480, epochs=10)\n<\/code><\/pre>\n\n<p>How can I include my month Information?<\/p>\n","answers":[]}
+{"QuestionId":57493398,"AnswerCount":1,"Tags":"<keras><deep-learning><lstm><prediction>","CreationDate":"2019-08-14T11:00:14.980","AcceptedAnswerId":null,"OwnerUserId":4281321.0,"Title":"Predicting Future values with Keras LSTM","Body":"<p>I have created an LSTM sales prediction model that works really well on the train and test sets. I would now like to predict beyond the dates in the entire dataset.<\/p>\n\n<p>I have tried following this answer <a href=\"https:\/\/stackoverflow.com\/questions\/48760472\/how-to-use-the-keras-model-to-forecast-for-future-dates-or-events\">how to use the Keras model to forecast for future dates or events?<\/a> but I really can't figure out how to adjust my code to do future predictions.<\/p>\n\n<p>Also, I changed my code from<\/p>\n\n<pre><code>X_train, y_train = train_set_scaled[:, 1:], train_set_scaled[:, 0:1]\nX_train = X_train.reshape(X_train.shape[0], 1, X_train.shape[1])\nX_test, y_test = test_set_scaled[:, 1:], test_set_scaled[:, 0:1]\nX_test = X_test.reshape(X_test.shape[0], 1, X_test.shape[1])\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>X_train, y_train = train_set_scaled[:, 1:], train_set_scaled[:, 1:8]\nX_train = X_train.reshape(X_train.shape[0], 1, X_train.shape[1])\nX_test, y_test = test_set_scaled[:, 1:], test_set_scaled[:, 1:8]\nX_test = X_test.reshape(X_test.shape[0], 1, X_test.shape[1])\n<\/code><\/pre>\n\n<p>after trying the solution in <a href=\"https:\/\/stackoverflow.com\/questions\/53252152\/keras-time-series-can-i-predict-next-6-month-in-one-time\">Keras time series can I predict next 6 month in one time<\/a> <\/p>\n\n<p>Here is the code where training and modelling happens:<\/p>\n\n<pre><code># changed to initial\nfor df in m:\n    train_set, test_set = m[df][0:-6].values, m[df][-6:].values\n\n    #apply Min Max Scaler\n    scaler = MinMaxScaler(feature_range=(-1, 1))\n    scaler = scaler.fit(train_set)\n\n    # reshape training set\n    train_set = train_set.reshape(train_set.shape[0], train_set.shape[1])\n    train_set_scaled = scaler.transform(train_set)\n\n    # reshape test set\n    test_set = test_set.reshape(test_set.shape[0], test_set.shape[1])\n    test_set_scaled = scaler.transform(test_set)\n\n    #build the LSTM Model\n    X_train, y_train = train_set_scaled[:, 1:], train_set_scaled[:, 0:1]\n    X_train = X_train.reshape(X_train.shape[0], 1, X_train.shape[1])\n    X_test, y_test = test_set_scaled[:, 1:], test_set_scaled[:, 0:1]\n    X_test = X_test.reshape(X_test.shape[0], 1, X_test.shape[1])\n\n    print('Fitting model for: {}'.format(df))\n\n    #fit our LSTM Model\n    model = Sequential()\n    model.add(LSTM(4, batch_input_shape=(1, X_train.shape[1], X_train.shape[2]), stateful=True))\n    model.add(Dense(1))\n    model.compile(loss='mean_squared_error', optimizer='adam')\n    model.fit(X_train, y_train, nb_epoch=500, batch_size=1, verbose=1, shuffle=False)\n\n#     model.save('lstm_model.h5')\n\n\n    print('Predictions for: {}'.format(df))\n\n    #check prediction\n    y_pred = model.predict(X_test,batch_size=1)\n\n    print('Inverse Transform for: {}'.format(df))\n\n    #inverse transformation to see actual sales\n    #reshape y_pred\n    y_pred = y_pred.reshape(y_pred.shape[0], 1, y_pred.shape[1])\n    #rebuild test set for inverse transform\n    pred_test_set = []\n    for index in range(0,len(y_pred)):\n        print (np.concatenate([y_pred[index],X_test[index]],axis=1))\n        pred_test_set.append(np.concatenate([y_pred[index],X_test[index]],axis=1))\n    #reshape pred_test_set\n    pred_test_set = np.array(pred_test_set)\n    pred_test_set = pred_test_set.reshape(pred_test_set.shape[0], pred_test_set.shape[2])\n    #inverse transform\n    pred_test_set_inverted = scaler.inverse_transform(pred_test_set)\n\n\n<\/code><\/pre>\n\n<p>I would like the predictions to go beyond the data in the dataset.<\/p>\n\n<p><strong>UPDATE<\/strong>: I trained the model and took its predictions on the test set. Use these as input for another LSTM model to fit and predict for 12 months. It worked for me. Also changed my last Dense layer (<strong>above<\/strong>) to predict 1 point at a time instead of 7 as I had before.\n<em>Below is the<\/em> <strong>code<\/strong>:<\/p>\n\n<pre><code>from numpy import array\nfor df in d:\n    if df in list_df:\n        # df_ADIDAS DYN PUL DEO 150 FCA5421\n        #KEEP\n\n        result_list = []\n        sales_dates = list(d[\"{}\".format(df)][-7:].Month)\n        act_sales = list(d[\"{}\".format(df)][-7:].Sale)\n        for index in range(0,len(pred_test_set_inverted)):\n            result_dict = {}\n            result_dict['pred_value'] = int(pred_test_set_inverted[index][0] + act_sales[index]) #change to 0 ffrom act_sales[index]\n            result_dict['date'] = sales_dates[index] #&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;REVIEW\n            result_list.append(result_dict)\n\n        df_result = pd.DataFrame(result_list)\n        predictions = list(df_result['pred_value'])\n\n        forecasts = []\n        result_list\n        for i in range(len(result_list)):\n            forecasts.append(result_list[i]['pred_value'])\n\n        def split_sequence(sequence, n_steps):\n            X, y = list(), list()\n            for i in range(len(sequence)):\n                # find the end of this pattern\n                end_ix = i + n_steps\n                # check if we are beyond the sequence\n                if end_ix &gt; len(sequence)-1:\n                    break\n                # gather input and output parts of the pattern\n                seq_x, seq_y = sequence[i:end_ix], sequence[end_ix]\n                X.append(seq_x)\n                y.append(seq_y)\n            return array(X), array(y)\n\n        # choose a number of time steps\n        n_steps = 4\n        # split into samples\n        X, y = split_sequence(forecasts, n_steps)\n        # summarize the data\n        #     for i in range(len(X)):\n        #         print(X[i], y[i])\n\n        n_features = 1\n        X = X.reshape((X.shape[0], X.shape[1], n_features))\n        # define model\n        model = Sequential()\n        model.add(LSTM(50, activation='relu', input_shape=(n_steps, n_features)))\n        model.add(Dense(1))\n        model.compile(optimizer='adam', loss='mse')\n        # fit model\n        model.fit(X, y, epochs=200, verbose=0)\n        # demonstrate prediction\n        x_input = array(predictions[-4:])\n        x_input = x_input.reshape((1, n_steps, n_features))\n        yhat = model.predict(x_input, verbose=0)\n        #print(yhat)\n\n        currentStep = yhat[:, -1:]\n\n        print('Twelve Month Prediction for {}'.format(df))\n        for i in range(12):\n            if i == 0:\n                x_input = x_input.reshape((1, n_steps, n_features))\n                yhat = model.predict(x_input, verbose=0)\n                print(yhat)\n            else:\n                x0_input = np.append(x_input, [currentStep[i-1]])\n                x0_input = x0_input.reshape((1, n_steps+1, n_features))\n                x_input = x0_input[:,1:]\n                yhat = model.predict(x_input)\n                currentStep = np.append(currentStep, yhat[:,-1:])\n                print(yhat)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57536725","CreationDate":"2019-08-17T13:29:16.510","ParentId":null,"OwnerUserId":"9025861","Title":null,"Body":"<p>Your last Dense layer says that you are predicting 7 points at a time. Save those predictions and feed them to the model again to predict next 7. That makes it 14 predictions simultaneously. And so on. Or change the number of nodes and shape of y from 7 to corresponding number and train again.<\/p>\n"}]}
+{"QuestionId":57493402,"AnswerCount":0,"Tags":"<python><tensorflow><keras><attention-model>","CreationDate":"2019-08-14T11:00:42.283","AcceptedAnswerId":null,"OwnerUserId":7959536.0,"Title":"Add attention mechanism to classification problem","Body":"<p>Currently, I am trying to implement attention mechanism to (sequence frame) video classification. So I am implementing the attention between CNN (feature extraction) -> attention -> LSTM (classification to specific class). I followed two papers for the implementation: \"Action Recognition using visual attention\" and \"Show, Attend and Tell: Neural Image Caption Generation with Visual Attention\".<\/p>\n\n<p>I have an image classification CNN and would like to extract the features and put it into a LSTM to visualize soft attention.<\/p>\n\n<p>After the last FC layer from a trained CNN, I extracted the features. After that, I applied attention to train attention weights and LSTM for classification. The goal of the. attention model is to know where the model is looking when it predicts specific class. So need to implement a spatial attention module for saliency detection<\/p>\n\n<pre><code>base_model = load_model(weight_file) \/\/ A trained resnet model\nlast_conv_layer = base_model.get_layer(\"global_average_pooling2d_3\")\ncnn_model = Model(input=base_model.input,output=last_conv_layer.output)\ncnn_model.trainable = False\n\ninputs = Input(shape=(max_frames, h, w, 3)) \nx = TimeDistributed(cnn_model)(inputs)\nattention = Dense(1, activation='tanh')(x)\nattention = Flatten()(attention)\nattention = Activation('softmax')(attention)\nattention = RepeatVector(256)(attention)\nattention = Permute([2, 1])(attention)\n\nsent_representation = merge([x, attention], mode='mul')       \nx = LSTM(256, return_sequences=True)(sent_representation)\n\npredictions = Dense(101, activation='softmax', name='pred')(x)\n\nmodel = Model(inputs=inputs, outputs=predictions)\n<\/code><\/pre>\n\n<blockquote>\n  <p>Traceback (most recent call last):\n   File \"train.py\", line 210, in \n     train_model(X,y,weight_file)\n   File \"train.py\", line 62, in train_model\n     model = lrcn_model(Nframes,H,W,weight_file)\n   File \"\/models.py\", line 71, in lrcn_model\n     sent_representation = merge([x, attention], mode='mul')<br>\n  TypeError: 'module' object is not callable<\/p>\n<\/blockquote>\n\n<p>Is this a correct implementation of the attention mechanism? And how can I fix the issue above that I am getting?<\/p>\n","answers":[]}
+{"QuestionId":57493481,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><keras><tensor>","CreationDate":"2019-08-14T11:06:02.967","AcceptedAnswerId":null,"OwnerUserId":1485549.0,"Title":"Tensor to tf.train.Feature object (TypeError: has type Tensor, but expected one of: int, long)","Body":"<p>I have a TensorFlow <code>Dataset<\/code> consisting of (image, label) pairs. I now wish to create a <code>tf.train.Example<\/code> object for each of them and store them in a <code>TFRecord<\/code> file. But I keep running into the following error:<\/p>\n\n<p><code>TypeError: &lt;tf.Tensor 'Cast:0' shape=() dtype=int64&gt; has type Tensor, but expected one of: int, long<\/code><\/p>\n\n<p>Following is a reproducible example. What am I doing wrong?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\ntf.enable_eager_execution()\n\n# Download sample images.\ncat_in_snow  = tf.keras.utils.get_file('320px-Felis_catus-cat_on_snow.jpg', 'https:\/\/storage.googleapis.com\/download.tensorflow.org\/example_images\/320px-Felis_catus-cat_on_snow.jpg')\nwilliamsburg_bridge = tf.keras.utils.get_file('194px-New_East_River_Bridge_from_Brooklyn_det.4a09796u.jpg','https:\/\/storage.googleapis.com\/download.tensorflow.org\/example_images\/194px-New_East_River_Bridge_from_Brooklyn_det.4a09796u.jpg')\n\n# Put images and corresponding labels in lists.\nimages = [cat_in_snow, williamsburg_bridge]\nlabels = [0, 1]\n\n# Create dataset from (image, label) pairs.\nimage_label_ds = tf.data.Dataset.from_tensor_slices((images, labels))\n\n# Convenience features for creating a tf.train.Feature.\ndef _bytes_feature(value):\n    \"\"\"Returns a bytes_list from a string \/ byte.\"\"\"\n    return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))\n\ndef _float_feature(value):\n    \"\"\"Returns a float_list from a float \/ double.\"\"\"\n    return tf.train.Feature(float_list=tf.train.FloatList(value=[value]))\n\ndef _int64_feature(value):\n    \"\"\"Returns an int64_list from a bool \/ enum \/ int \/ uint.\"\"\"\n    return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))\n\n# Create tf.train.Example for each (image, label) pair.\ndef image_label_to_example(image, label):\n    feature = {\n        \"label\": _int64_feature(tf.cast(label, tf.int64)),\n        \"image\": _bytes_feature(image)\n    }\n\n    tf_example = tf.train.Example(features=tf.train.Features(feature=feature))\n\n    return tf_example.SerializeToString()\n\nimage_label_examples = image_label_ds.map(image_label_to_example)\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>---\n\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-37-8baca3115c8a&gt; in &lt;module&gt;\n-&amp;#x2014;&gt; 1 image&lt;sub&gt;label&lt;\/sub&gt;&lt;sub&gt;examples&lt;\/sub&gt; = imgs&lt;sub&gt;lbs&lt;\/sub&gt;&lt;sub&gt;ds.map&lt;\/sub&gt;(image&lt;sub&gt;label&lt;\/sub&gt;&lt;sub&gt;to&lt;\/sub&gt;&lt;sub&gt;example&lt;\/sub&gt;)\n\n\/\/anaconda3\/envs\/tf\/lib\/python3.7\/site-packages\/tensorflow\/python\/data\/ops\/dataset&lt;sub&gt;ops.py&lt;\/sub&gt; in map(self, map&lt;sub&gt;func&lt;\/sub&gt;, num&lt;sub&gt;parallel&lt;\/sub&gt;&lt;sub&gt;calls&lt;\/sub&gt;)\n    1770     if num&lt;sub&gt;parallel&lt;\/sub&gt;&lt;sub&gt;calls&lt;\/sub&gt; is None:\n    1771       return DatasetV1Adapter(\n-&gt; 1772           MapDataset(self, map&lt;sub&gt;func&lt;\/sub&gt;, preserve&lt;sub&gt;cardinality&lt;\/sub&gt;=False))\n    1773     else:\n    1774       return DatasetV1Adapter(\n\n\/\/anaconda3\/envs\/tf\/lib\/python3.7\/site-packages\/tensorflow\/python\/data\/ops\/dataset&lt;sub&gt;ops.py&lt;\/sub&gt; in \\_&lt;sub&gt;init&lt;\/sub&gt;\\_&lt;sub&gt;(self, input&lt;sub&gt;dataset&lt;\/sub&gt;, map&lt;sub&gt;func&lt;\/sub&gt;, use&lt;sub&gt;inter&lt;\/sub&gt;&lt;sub&gt;op&lt;\/sub&gt;&lt;sub&gt;parallelism&lt;\/sub&gt;, preserve&lt;sub&gt;cardinality&lt;\/sub&gt;, use&lt;sub&gt;legacy&lt;\/sub&gt;&lt;sub&gt;function&lt;\/sub&gt;)&lt;\/sub&gt;\n    3188         self.&lt;sub&gt;transformation&lt;\/sub&gt;&lt;sub&gt;name&lt;\/sub&gt;(),\n    3189         dataset=input&lt;sub&gt;dataset&lt;\/sub&gt;,\n-&gt; 3190         use&lt;sub&gt;legacy&lt;\/sub&gt;&lt;sub&gt;function&lt;\/sub&gt;=use&lt;sub&gt;legacy&lt;\/sub&gt;&lt;sub&gt;function&lt;\/sub&gt;)\n    3191     variant&lt;sub&gt;tensor&lt;\/sub&gt; = gen&lt;sub&gt;dataset&lt;\/sub&gt;&lt;sub&gt;ops.map&lt;\/sub&gt;&lt;sub&gt;dataset&lt;\/sub&gt;(\n    3192         input&lt;sub&gt;dataset&lt;\/sub&gt;.&lt;sub&gt;variant&lt;\/sub&gt;&lt;sub&gt;tensor&lt;\/sub&gt;,  # pylint: disable=protected-access\n\n\/\/anaconda3\/envs\/tf\/lib\/python3.7\/site-packages\/tensorflow\/python\/data\/ops\/dataset&lt;sub&gt;ops.py&lt;\/sub&gt; in \\_&lt;sub&gt;init&lt;\/sub&gt;\\_&lt;sub&gt;(self, func, transformation&lt;sub&gt;name&lt;\/sub&gt;, dataset, input&lt;sub&gt;classes&lt;\/sub&gt;, input&lt;sub&gt;shapes&lt;\/sub&gt;, input&lt;sub&gt;types&lt;\/sub&gt;, input&lt;sub&gt;structure&lt;\/sub&gt;, add&lt;sub&gt;to&lt;\/sub&gt;&lt;sub&gt;graph&lt;\/sub&gt;, use&lt;sub&gt;legacy&lt;\/sub&gt;&lt;sub&gt;function&lt;\/sub&gt;, defun&lt;sub&gt;kwargs&lt;\/sub&gt;)&lt;\/sub&gt;\n    2553       resource&lt;sub&gt;tracker&lt;\/sub&gt; = tracking.ResourceTracker()\n    2554       with tracking.resource&lt;sub&gt;tracker&lt;\/sub&gt;&lt;sub&gt;scope&lt;\/sub&gt;(resource&lt;sub&gt;tracker&lt;\/sub&gt;):\n-&gt; 2555         self.&lt;sub&gt;function&lt;\/sub&gt; = wrapper&lt;sub&gt;fn&lt;\/sub&gt;.&lt;sub&gt;get&lt;\/sub&gt;&lt;sub&gt;concrete&lt;\/sub&gt;&lt;sub&gt;function&lt;\/sub&gt;&lt;sub&gt;internal&lt;\/sub&gt;()\n    2556         if add&lt;sub&gt;to&lt;\/sub&gt;&lt;sub&gt;graph&lt;\/sub&gt;:\n    2557           self.&lt;sub&gt;function.add&lt;\/sub&gt;&lt;sub&gt;to&lt;\/sub&gt;&lt;sub&gt;graph&lt;\/sub&gt;(ops.get&lt;sub&gt;default&lt;\/sub&gt;&lt;sub&gt;graph&lt;\/sub&gt;())\n\n\/\/anaconda3\/envs\/tf\/lib\/python3.7\/site-packages\/tensorflow\/python\/eager\/function.py in \\_get&lt;sub&gt;concrete&lt;\/sub&gt;&lt;sub&gt;function&lt;\/sub&gt;&lt;sub&gt;internal&lt;\/sub&gt;(self, \\*args, \\*\\*kwargs)\n    1353     \"\"\"Bypasses error checking when getting a graph function.\"\"\"\n    1354     graph&lt;sub&gt;function&lt;\/sub&gt; = self.&lt;sub&gt;get&lt;\/sub&gt;&lt;sub&gt;concrete&lt;\/sub&gt;&lt;sub&gt;function&lt;\/sub&gt;&lt;sub&gt;internal&lt;\/sub&gt;&lt;sub&gt;garbage&lt;\/sub&gt;&lt;sub&gt;collected&lt;\/sub&gt;(\n-&gt; 1355         \\*args, \\*\\*kwargs)\n    1356     # We're returning this concrete function to someone, and they may keep a\n    1357     # reference to the FuncGraph without keeping a reference to the\n\n\/\/anaconda3\/envs\/tf\/lib\/python3.7\/site-packages\/tensorflow\/python\/eager\/function.py in \\_get&lt;sub&gt;concrete&lt;\/sub&gt;&lt;sub&gt;function&lt;\/sub&gt;&lt;sub&gt;internal&lt;\/sub&gt;&lt;sub&gt;garbage&lt;\/sub&gt;&lt;sub&gt;collected&lt;\/sub&gt;(self, \\*args, \\*\\*kwargs)\n    1347     if self.input&lt;sub&gt;signature&lt;\/sub&gt;:\n    1348       args, kwargs = None, None\n-&gt; 1349     graph&lt;sub&gt;function&lt;\/sub&gt;, \\_, \\_ = self.&lt;sub&gt;maybe&lt;\/sub&gt;&lt;sub&gt;define&lt;\/sub&gt;&lt;sub&gt;function&lt;\/sub&gt;(args, kwargs)\n    1350     return graph&lt;sub&gt;function&lt;\/sub&gt;\n    1351\n\n\/\/anaconda3\/envs\/tf\/lib\/python3.7\/site-packages\/tensorflow\/python\/eager\/function.py in \\_maybe&lt;sub&gt;define&lt;\/sub&gt;&lt;sub&gt;function&lt;\/sub&gt;(self, args, kwargs)\n    1650       graph&lt;sub&gt;function&lt;\/sub&gt; = self.&lt;sub&gt;function&lt;\/sub&gt;&lt;sub&gt;cache.primary.get&lt;\/sub&gt;(cache&lt;sub&gt;key&lt;\/sub&gt;, None)\n    1651       if graph&lt;sub&gt;function&lt;\/sub&gt; is None:\n-&gt; 1652         graph&lt;sub&gt;function&lt;\/sub&gt; = self.&lt;sub&gt;create&lt;\/sub&gt;&lt;sub&gt;graph&lt;\/sub&gt;&lt;sub&gt;function&lt;\/sub&gt;(args, kwargs)\n    1653         self.&lt;sub&gt;function&lt;\/sub&gt;&lt;sub&gt;cache.primary&lt;\/sub&gt;[cache&lt;sub&gt;key&lt;\/sub&gt;] = graph&lt;sub&gt;function&lt;\/sub&gt;\n    1654       return graph&lt;sub&gt;function&lt;\/sub&gt;, args, kwargs\n\n\/\/anaconda3\/envs\/tf\/lib\/python3.7\/site-packages\/tensorflow\/python\/eager\/function.py in \\_create&lt;sub&gt;graph&lt;\/sub&gt;&lt;sub&gt;function&lt;\/sub&gt;(self, args, kwargs, override&lt;sub&gt;flat&lt;\/sub&gt;&lt;sub&gt;arg&lt;\/sub&gt;&lt;sub&gt;shapes&lt;\/sub&gt;)\n    1543             arg&lt;sub&gt;names&lt;\/sub&gt;=arg&lt;sub&gt;names&lt;\/sub&gt;,\n    1544             override&lt;sub&gt;flat&lt;\/sub&gt;&lt;sub&gt;arg&lt;\/sub&gt;&lt;sub&gt;shapes&lt;\/sub&gt;=override&lt;sub&gt;flat&lt;\/sub&gt;&lt;sub&gt;arg&lt;\/sub&gt;&lt;sub&gt;shapes&lt;\/sub&gt;,\n-&gt; 1545             capture&lt;sub&gt;by&lt;\/sub&gt;&lt;sub&gt;value&lt;\/sub&gt;=self.&lt;sub&gt;capture&lt;\/sub&gt;&lt;sub&gt;by&lt;\/sub&gt;&lt;sub&gt;value&lt;\/sub&gt;),\n    1546         self.&lt;sub&gt;function&lt;\/sub&gt;&lt;sub&gt;attributes&lt;\/sub&gt;)\n    1547\n\n\/\/anaconda3\/envs\/tf\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/func&lt;sub&gt;graph.py&lt;\/sub&gt; in func&lt;sub&gt;graph&lt;\/sub&gt;&lt;sub&gt;from&lt;\/sub&gt;&lt;sub&gt;py&lt;\/sub&gt;&lt;sub&gt;func&lt;\/sub&gt;(name, python&lt;sub&gt;func&lt;\/sub&gt;, args, kwargs, signature, func&lt;sub&gt;graph&lt;\/sub&gt;, autograph, autograph&lt;sub&gt;options&lt;\/sub&gt;, add&lt;sub&gt;control&lt;\/sub&gt;&lt;sub&gt;dependencies&lt;\/sub&gt;, arg&lt;sub&gt;names&lt;\/sub&gt;, op&lt;sub&gt;return&lt;\/sub&gt;&lt;sub&gt;value&lt;\/sub&gt;, collections, capture&lt;sub&gt;by&lt;\/sub&gt;&lt;sub&gt;value&lt;\/sub&gt;, override&lt;sub&gt;flat&lt;\/sub&gt;&lt;sub&gt;arg&lt;\/sub&gt;&lt;sub&gt;shapes&lt;\/sub&gt;)\n    713                                           converted&lt;sub&gt;func&lt;\/sub&gt;)\n    714\n&amp;#x2013;&gt; 715       func&lt;sub&gt;outputs&lt;\/sub&gt; = python&lt;sub&gt;func&lt;\/sub&gt;(\\*func&lt;sub&gt;args&lt;\/sub&gt;, \\*\\*func&lt;sub&gt;kwargs&lt;\/sub&gt;)\n    716\n    717       # invariant: \\`func&lt;sub&gt;outputs&lt;\/sub&gt;\\` contains only Tensors, CompositeTensors,\n\n\/\/anaconda3\/envs\/tf\/lib\/python3.7\/site-packages\/tensorflow\/python\/data\/ops\/dataset&lt;sub&gt;ops.py&lt;\/sub&gt; in wrapper&lt;sub&gt;fn&lt;\/sub&gt;(\\*args)\n    2547           attributes=defun&lt;sub&gt;kwargs&lt;\/sub&gt;)\n    2548       def wrapper&lt;sub&gt;fn&lt;\/sub&gt;(\\*args):  # pylint: disable=missing-docstring\n-&gt; 2549         ret = \\_wrapper&lt;sub&gt;helper&lt;\/sub&gt;(\\*args)\n    2550         ret = self.&lt;sub&gt;output&lt;\/sub&gt;&lt;sub&gt;structure&lt;\/sub&gt;.&lt;sub&gt;to&lt;\/sub&gt;&lt;sub&gt;tensor&lt;\/sub&gt;&lt;sub&gt;list&lt;\/sub&gt;(ret)\n    2551         return [ops.convert&lt;sub&gt;to&lt;\/sub&gt;&lt;sub&gt;tensor&lt;\/sub&gt;(t) for t in ret]\n\n\/\/anaconda3\/envs\/tf\/lib\/python3.7\/site-packages\/tensorflow\/python\/data\/ops\/dataset&lt;sub&gt;ops.py&lt;\/sub&gt; in \\_wrapper&lt;sub&gt;helper&lt;\/sub&gt;(\\*args)\n    2487         nested&lt;sub&gt;args&lt;\/sub&gt; = (nested&lt;sub&gt;args&lt;\/sub&gt;,)\n    2488\n-&gt; 2489       ret = func(\\*nested&lt;sub&gt;args&lt;\/sub&gt;)\n    2490       # If \\`func\\` returns a list of tensors, \\`nest.flatten()\\` and\n    2491       # \\`ops.convert&lt;sub&gt;to&lt;\/sub&gt;&lt;sub&gt;tensor&lt;\/sub&gt;()\\` would conspire to attempt to stack\n\n&lt;ipython-input-36-9376dfea556f&gt; in image&lt;sub&gt;label&lt;\/sub&gt;&lt;sub&gt;to&lt;\/sub&gt;&lt;sub&gt;example&lt;\/sub&gt;(image, label)\n    1 def image&lt;sub&gt;label&lt;\/sub&gt;&lt;sub&gt;to&lt;\/sub&gt;&lt;sub&gt;example&lt;\/sub&gt;(image, label):\n    2     feature = {\n-&amp;#x2014;&gt; 3         \"label\": \\_int64&lt;sub&gt;feature&lt;\/sub&gt;(tf.cast(label, tf.int64)),\n    4         \"image\": \\_bytes&lt;sub&gt;feature&lt;\/sub&gt;(image)\n    5     }\n\n&lt;ipython-input-35-b0b600e28608&gt; in \\_int64&lt;sub&gt;feature&lt;\/sub&gt;(value)\n    12 def \\_int64&lt;sub&gt;feature&lt;\/sub&gt;(value):\n    13   \"\"\"Returns an int64&lt;sub&gt;list&lt;\/sub&gt; from a bool \/ enum \/ int \/ uint.\"\"\"\n&amp;#x2014;&gt; 14   return tf.train.Feature(int64&lt;sub&gt;list&lt;\/sub&gt;=tf.train.Int64List(value=[value]))\n\nTypeError: &lt;tf.Tensor 'Cast:0' shape=() dtype=int64&gt; has type Tensor, but expected one of: int, long\n<\/code><\/pre>\n\n<p>System Info:<br>\nPython 3.7.3.<br>\nTensorflow 1.14.0.<br>\nMacOS Mojave 10.14.6. <\/p>\n\n<p>References:<br>\n<a href=\"https:\/\/www.tensorflow.org\/tutorials\/load_data\/tf_records#fetch_the_images\" rel=\"nofollow noreferrer\">TFRecords and tf.Example<\/a>.<br>\n<a href=\"https:\/\/www.tensorflow.org\/tutorials\/load_data\/images\" rel=\"nofollow noreferrer\">Load images<\/a><\/p>\n","answers":[]}
+{"QuestionId":57493538,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><video-processing>","CreationDate":"2019-08-14T11:09:47.517","AcceptedAnswerId":57495212.0,"OwnerUserId":4512432.0,"Title":"Does 3D convolution in Keras work with RGB videos?","Body":"<p>Keras' Conv3D expects input  <code>5D tensor with shape: (batch, conv_dim1, conv_dim2, conv_dim3, channels)<\/code> (assuming data_format is \"channels_last\"). Now say my filter size is (3,3,3) and my input is (10,125,300,200,3), a video dataset of 10 videos, each with 125 frames, and spatial size 300x200,and channel 3 due to frames being RGB. The default stride value is <code>(1, 1, 1)<\/code>. The picture in my head of how this convolution works is as shown <a href=\"https:\/\/www.youtube.com\/watch?v=SQ67NBCLV98\" rel=\"nofollow noreferrer\">here at 9:28<\/a>. <\/p>\n\n<p>What I can't figure is whether the stride of 1 along the temporal dimension moves 1 frame at a time or 1 channel of a frame at a time. I tried to look up the code of conv3D <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.14\/tensorflow\/python\/keras\/layers\/convolutional.py#L568-L602\" rel=\"nofollow noreferrer\">here<\/a> and couldn't gather much. I tried training a deep learning network using 3D CNNS and RGB videos and the resulting images have messed up colours (almost grey), so I'm assuming there's some mess up with colour channels. I checked my input, that seems fine, so the network is probably funny.<\/p>\n\n<p>Tl:dr <\/p>\n\n<p>Need to figure if RGB videos need conscious changes in strides so the channels of one frame are treated with 2D convolution, and also would be grateful for pointers to code\/papers dealing with RGB videos and 3D CNNs<\/p>\n","answers":[{"AnswerId":"57495212","CreationDate":"2019-08-14T12:58:48.567","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>In all convolutions, the filter size encompasses all channels together. Channels do not participate in strides. <\/p>\n\n<p>So, strides happen as if your video was a cube. Stride 1 step in each dimension (x,y,z) until the entire cube is swept. (The convolution has no idea of what the dimensions are, and will not treat the frames differently from how they treat pixels).<\/p>\n\n<p>You have a little 3x3x3 cube sweeping a huge 125x300,200 parallelepiped, pixel by pixel, frame by frame. So: the stride moves one frame at a time, but considering only a 3x3 segment of the image.<\/p>\n\n<p>This doesn't \"seem\" good for videos (but machine learning has its surprises), unless at some point you have a very tiny resolution, so that a filter starts seeing the whole picture in each frame.<\/p>\n\n<p>You can keep testing the 3D convs to see what happens, but a few suggestions that \"seem\" better are:<\/p>\n\n<ul>\n<li><p>Use <code>TimeDistributed(Conv2D(...))<\/code> and <code>TimeDistributed(MaxPooling2D(...))<\/code> until you get a small resolution video in the middle of the model (or even a 1x1, if you're going extreme). Then start using:   <\/p>\n\n<ul>\n<li><code>Conv3D<\/code> if there are still spatial dimensions    <\/li>\n<li><code>Conv1D<\/code> if you eliminated the spatial dimensions     <\/li>\n<li>In both cases, it's a good idea to increase the kernel size in the frames dimension, 3 frames may be too little to interpret what is happening (unless you have a low frame rate)    <\/li>\n<\/ul><\/li>\n<li><p>Use <code>TimeDistributed(Conv2D(...))<\/code>, eliminate the spatial dimensions at some point and start using RNNs like <code>LSTM<\/code>    <\/p><\/li>\n<li><p>Use <code>ConvLSTM2D<\/code> layers. <\/p><\/li>\n<\/ul>\n"}]}
+{"QuestionId":57493889,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><vgg-net>","CreationDate":"2019-08-14T11:34:33.320","AcceptedAnswerId":57499375.0,"OwnerUserId":9855577.0,"Title":"Input 0 is incompatible with layer global_average_pooling2d_4: expected ndim=4, found ndim=2 error","Body":"<p>I am trying to fine-tune a VGG16 model. I have removed the last 5 layers<\/p>\n\n<pre><code>(*block5_pool (MaxPooling2D),flatten(Flatten),fc1 (Dense),fc2 (Dense),predictions (Dense)*). \n<\/code><\/pre>\n\n<p>Now, I want to add a global average pooling layer, but I am getting this error<\/p>\n\n<blockquote>\n  <p>Input 0 is incompatible with layer global_average_pooling2d_4: expected ndim=4, found ndim=2** <\/p>\n<\/blockquote>\n\n<p>what seems to be the problem here?<\/p>\n\n<pre><code>model = VGG16(weights='imagenet', include_top=True)\nmodel.layers.pop()\nmodel.layers.pop()\nmodel.layers.pop()\nmodel.layers.pop()\nmodel.layers.pop()\nx = model.output\nx = GlobalAveragePooling2D()(x)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57499375","CreationDate":"2019-08-14T17:24:37.767","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>If want to remove the last four layers, then just use <code>include_top=False<\/code>. Further, use <code>pooling='avg'<\/code> to add a <code>GlobalAveragePooling2D<\/code> layer as the last layer:<\/p>\n\n<pre><code>model = VGG16(weights='imagenet', include_top=False, pooling='avg')\n<\/code><\/pre>\n\n<hr>\n\n<p>A note about why your original solution does not work: as already suggested in <a href=\"https:\/\/stackoverflow.com\/a\/52282558\/2099607\">this answer<\/a>, you can't use <code>pop()<\/code> method on <code>layers<\/code> attribute of the models to remove a layer. Instead, you need to reference their output directly (e.g. <code>model.layers[-4].output<\/code>) and then feed them to other layers if you want to add new connections.<\/p>\n"}]}
+{"QuestionId":57493936,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-08-14T11:37:24.530","AcceptedAnswerId":null,"OwnerUserId":9850548.0,"Title":"How to fix TypeError: __call__() of arguments error in keras prediction","Body":"<p>I'm trying to change a repository from  git bun in my own i found a problem with number of arguments,<\/p>\n\n<p>the part of code that I'm talking about is in :\n<a href=\"https:\/\/github.com\/aSafarpoor\/keras_cnn\/blob\/master\/src\/predict.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aSafarpoor\/keras_cnn\/blob\/master\/src\/predict.py<\/a>\nand i run python predict.py with it's arguments.<\/p>\n\n<p>but it makes error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"predict.py\", line 153, in &lt;module&gt;\n    main()\n  File \"predict.py\", line 117, in main\n    v = network_eval.predict(specs)\n  File \"\/home\/mas\/anaconda3\/envs\/weild\/lib\/python2.7\/site-packages\/keras\/engine\/training.py\", line 1166, in predict\n    f = self.predict_function()\nTypeError: __call__() takes exactly 2 arguments (1 given)\n<\/code><\/pre>\n\n<p>it means <i>keras<\/i> prediction function need more arguments but it's not normal,<\/p>\n\n<p>take care that input is a <i>numpy<\/i> array in a true shape.<\/p>\n\n<p>the related part in code is :<\/p>\n\n<pre><code> for c, ID in enumerate(unique_list):\n        if c % 50 == 0: print('Finish extracting features for {}\/{}th wav.'.format(c, total_length))\n        specs = ut.load_data(ID, dim_w=args.wdim,dim_h=args.hdim,win_length=params['win_length'], sr=params['sampling_rate'],\n                             hop_length=params['hop_length'], n_fft=params['nfft'],\n                             spec_len=args.frame_num, mode='eval')\n        # specs = np.expand_dims(np.expand_dims(specs, 0), -1)\n        specs=np.expand_dims(specs,0)\n        v = network_eval.predict(specs)####error is here ####\n        feats += [v]\n\n<\/code><\/pre>\n\n<p>so what should be the second argument?or what should change?<\/p>\n","answers":[{"AnswerId":"57495097","CreationDate":"2019-08-14T12:51:25.097","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>I think you have a modified Keras version, since the exact same line in the keras official code (<a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/engine\/training.py#L1396\" rel=\"nofollow noreferrer\">here<\/a>) does not have the parenthesis making the function call.<\/p>\n\n<p>Make sure that you only use the official Keras releases and do not modify its source code, you can introduce bugs and errors that we cannot debug.<\/p>\n"}]}
+{"QuestionId":57493940,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-14T11:37:41.067","AcceptedAnswerId":null,"OwnerUserId":7248397.0,"Title":"\"Unknown initializer:_initializer\" while loading tensorflow model","Body":"<p>I trained a model with xavier_initializer, I paste model's code a bit lower. The issue is that while loading that model with \"from tensorflow.keras.models import load_model\" I get an error: <\/p>\n\n<p>\"ValueError: Unknown initializer:_initializer\"<\/p>\n\n<p>I tried to add<\/p>\n\n<ul>\n<li><p>custom_objects={\"xavier_initializer\": xavier_initializer}\nor <\/p><\/li>\n<li><p>custom_objects={\"_initializer\": xavier_initializer}<\/p><\/li>\n<\/ul>\n\n<p>to load_model, but none of them work. I tried also to use tf.keras.models.load_model and keras.models.load_model, didn't help neither. Any ideas how to fix that?<\/p>\n\n<p>Model definition:<\/p>\n\n<pre><code>def initialize_model(activation_f, if_maxpooling, metric, optimizer):\n    inputs = Input(shape = train_x.shape[1:])\n\n    conv2d_1 = Conv2D(\n        filters = 32,\n        kernel_size = (3,3),\n        strides = 1,\n        activation = activation_f,\n        kernel_initializer = initializer,\n    )(inputs)\n    batch_norm_1 = BatchNormalization()(conv2d_1)\n    if not if_maxpooling:\n        conv2d_2 = Conv2D(\n            filters = 8,\n            kernel_size = (3,3),\n            strides = 2,\n            activation = activation_f,\n            kernel_initializer = initializer,\n        )(batch_norm_1)\n        batch_norm_2 = BatchNormalization()(conv2d_2)\n        flatten = Flatten()(batch_norm_2)\n    else:\n        conv2d_2 = Conv2D(\n            filters = 8,\n            kernel_size = (3,3),\n            strides = 1,\n            activation = activation_f,\n            kernel_initializer = initializer,\n        )(batch_norm_1)\n        batch_norm_2 = BatchNormalization()(conv2d_2)\n        maxpool_1 = MaxPooling2D(\n            pool_size = (3, 3),\n            strides = 2\n        )(batch_norm_2)\n        flatten = Flatten()(maxpool_1)\n\n    fully_connected_1 = Dense(\n        256,\n    )(flatten)\n    batch_norm_3 = BatchNormalization()(fully_connected_1)\n    fully_connected_2 = Dense(\n        64,\n        activation = activation_f\n    )(batch_norm_3)\n    batch_norm_4 = BatchNormalization()(fully_connected_2)\n\n    outputs = Dense(10)(batch_norm_4)\n\n    model = Model(\n        inputs = inputs,\n        outputs = outputs\n    )\n\n    if optimizer == RMSprop:\n        model.compile(\n            loss = mean_squared_error, # log_loss\n            optimizer = RMSprop(0.001),\n            metrics = [metric]\n        )\n    else:\n        model.compile(\n            loss = mean_squared_error, # log_loss\n            optimizer = Adam(0.001, 0.9, 0.999),\n            metrics = [metric]\n        )\n\n    return model\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57493949,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-probability><eager-execution>","CreationDate":"2019-08-14T11:38:20.967","AcceptedAnswerId":null,"OwnerUserId":1398965.0,"Title":"TFP exact GP regression in eager mode","Body":"<p>I'm trying to perform exact GP regression using the TF2.0 eager mode, based on the original graph based example from\n <a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/probability\/blob\/master\/tensorflow_probability\/examples\/jupyter_notebooks\/Gaussian_Process_Regression_In_TFP.ipynb\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/github\/tensorflow\/probability\/blob\/master\/tensorflow_probability\/examples\/jupyter_notebooks\/Gaussian_Process_Regression_In_TFP.ipynb<\/a><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>amplitude = (\n    np.finfo(np.float64).tiny +\n    tf.nn.softplus(tf.Variable(initial_value=1., name='amplitude', dtype=np.float64))\n)\nlength_scale = (\n    np.finfo(np.float64).tiny +\n    tf.nn.softplus(tf.Variable(initial_value=1., name='length_scale', dtype=np.float64))\n)\nobservation_noise_variance = (\n    np.finfo(np.float64).tiny +\n    tf.nn.softplus(tf.Variable(initial_value=1e-6,\n                               name='observation_noise_variance',\n                               dtype=np.float64))\n)\n\nkernel = tfk.ExponentiatedQuadratic(amplitude, length_scale)\n\ngp = tfd.GaussianProcess(\n    kernel=kernel,\n    index_points=tf.expand_dims(x, 1),\n    observation_noise_variance=observation_noise_variance\n)\n\nneg_log_likelihood = lambda: -gp.log_prob(y)\n\noptimizer = tf.optimizers.Adam(learning_rate=.01)\n\nnum_iters = 1000\nlls_ = np.zeros(num_iters, np.float64)\nfor i in range(num_iters):\n    lls_[i] = neg_log_likelihood()\n    optimizer.minimize(neg_log_likelihood, var_list=[amplitude, length_scale, observation_noise_variance])\n<\/code><\/pre>\n\n<p><strong>However optimization fails with:<\/strong><\/p>\n\n<blockquote>\n  <p>'tensorflow.python.framework.ops.EagerTensor' object has no attribute '_in_graph_mode'<\/p>\n<\/blockquote>\n\n<p>And if I move the amplitude, length_scale and observation_noise_variance each to tf.Variable, like:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>amplitude = tf.Variable(initial_value=1., name='amplitude', dtype=np.float64)\namplitude_ = (\n    np.finfo(np.float64).tiny +\n    tf.nn.softplus(amplitude)\n)\n<\/code><\/pre>\n\n<p><strong>Optimization fails with:<\/strong><\/p>\n\n<blockquote>\n  <p>ValueError: No gradients provided for any variable: ['amplitude:0', 'length_scale:0', 'observation_noise_variance:0'].<\/p>\n<\/blockquote>\n\n<p>What am I doing wrong?<\/p>\n","answers":[{"AnswerId":"57506588","CreationDate":"2019-08-15T08:02:37.807","ParentId":null,"OwnerUserId":"1398965","Title":null,"Body":"<p>There is an issue with eager mode currently:<\/p>\n\n<p><a href=\"https:\/\/groups.google.com\/a\/tensorflow.org\/d\/msg\/tfprobability\/IlhL-fcv3yc\/jpQc4ogcFwAJ\" rel=\"nofollow noreferrer\">https:\/\/groups.google.com\/a\/tensorflow.org\/d\/msg\/tfprobability\/IlhL-fcv3yc\/jpQc4ogcFwAJ<\/a><\/p>\n\n<p>Workaround was to work with GradientTape explicitly:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>amplitude_ = tf.Variable(initial_value=1., name='amplitude_', dtype=np.float64)\nlength_scale_ = tf.Variable(initial_value=1., name='length_scale_', dtype=np.float64)\nobservation_noise_variance_ = tf.Variable(initial_value=1e-6,\n                                         name='observation_noise_variance_',\n                                         dtype=np.float64)\n\n@tf.function\ndef neg_log_likelihood():\n    amplitude = np.finfo(np.float64).tiny + tf.nn.softplus(amplitude_)\n    length_scale = np.finfo(np.float64).tiny + tf.nn.softplus(length_scale_)\n    observation_noise_variance = np.finfo(np.float64).tiny + tf.nn.softplus(observation_noise_variance_)\n\n    kernel = tfk.ExponentiatedQuadratic(amplitude, length_scale)\n\n    gp = tfd.GaussianProcess(\n        kernel=kernel,\n        index_points=tf.expand_dims(x, 1),\n        observation_noise_variance=observation_noise_variance\n    )\n\n    return -gp.log_prob(y)\n\noptimizer = tf.optimizers.Adam(learning_rate=.01)\n\nnum_iters = 1000\n\nnlls = np.zeros(num_iters, np.float64)\nfor i in range(num_iters):\n    nlls[i] = neg_log_likelihood()\n    with tf.GradientTape() as tape:\n        loss = neg_log_likelihood()\n    grads = tape.gradient(loss, [amplitude_, length_scale_, observation_noise_variance_])\n    optimizer.apply_gradients(zip(grads, [amplitude_, length_scale_, observation_noise_variance_]))\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":57493965,"AnswerCount":0,"Tags":"<keras><deep-learning>","CreationDate":"2019-08-14T11:39:27.097","AcceptedAnswerId":null,"OwnerUserId":5079088.0,"Title":"Understanding Triplet Loss data generator","Body":"<p>Below I quote the code (made from <a href=\"https:\/\/github.com\/noelcodella\/tripletloss-keras-tensorflow\/blob\/master\/tripletloss.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/noelcodella\/tripletloss-keras-tensorflow\/blob\/master\/tripletloss.py<\/a>). It works, but i had to put my own image data generator there. Now, for it to work, I had to add a dummy (at least I think it is dummy) parameter (the last \"base\"): <\/p>\n\n<pre><code>yield [base, positive, negative], base\n<\/code><\/pre>\n\n<p>As I mentioned, it seems to be working, but:\na) I am not sure it is \"dummy\". Would be really bad if I provide a lookforward data or something.\nb) Neither here nor in the original code (available by the link above) I can figure what it is for. I mean, Lambda layer would not woirk without it, complaining that \"you must feed data\", but...<\/p>\n\n<p>Please help.<\/p>\n\n<pre><code># https:\/\/github.com\/noelcodella\/tripletloss-keras-tensorflow\/blob\/master\/tripletloss.py\n\ndef gen(bIsTrain):\n\n  while True:\n    arrBaseExamples = []\n    arrPositiveExamples = []\n    arrNegativeExamples = []\n\n    for i in range(BATCH_SIZE):\n      nImageIdx, positiveClass, negativeClass = get_triplet(TRAINING_IMAGES_PER_CLASS, bIsTrain)\n\n      #t0 = time()\n      baseExampleImg = loadImage(positiveClass, nImageIdx[0], datagen)      \n      positiveExampleImg = loadImage(positiveClass, nImageIdx[1], datagen)\n      negativeExampleImg = loadImage(negativeClass, nImageIdx[2], datagen)\n      #t1 = time()\n      #print('loaded in %f' %(t1-t0))\n\n      arrBaseExamples.append(baseExampleImg)\n      arrPositiveExamples.append(positiveExampleImg)\n      arrNegativeExamples.append(negativeExampleImg)\n\n    #base = preprocess_input(np.array(arrBaseExamples)) \/ 255. #'a' #preprocess_input(np.array(arrBaseExamples))\n    base = np.array(arrBaseExamples) \/ 255.\n    #train_datagen.fit(base)\n\n    #positive = preprocess_input(np.array(arrPositiveExamples)) \/ 255.\n    positive = np.array(arrPositiveExamples) \/ 255.\n    #train_datagen.fit(positive)\n\n    #negative = preprocess_input(np.array(arrNegativeExamples)) \/ 255.\n    negative = np.array(arrNegativeExamples) \/ 255.\n    #train_datagen.fit(negative)\n\n    label = None\n\n    #yield (np.stack(base, positive, negative), label) \n    #yield ({'anchor_input': base, 'positive_input': positive, 'negative_input': negative}, label) \n    yield [base, positive, negative], base\n\n    #yield [X1i[0], X2i[0], X3i[0]], X1i[1]\n\n\n\ndef createModel(nL2):\n    input_shape=(IMAGE_SIZE,IMAGE_SIZE,3)\n\n    # Initialize a ResNet50_ImageNet Model\n    resnet_input = Input(shape=input_shape)\n    resnet_model = keras.applications.resnet50.ResNet50(weights='imagenet', include_top = False, input_tensor=resnet_input)\n\n    # New Layers over ResNet50\n    net = resnet_model.output\n    #net = Flatten(name='flatten')(net)\n    net = GlobalAveragePooling2D(name='gap')(net)\n    #net = Dropout(0.5)(net)\n    net = Dense(EMBEDDING_DIM,activation='relu',name='t_emb_1')(net)\n    net = Lambda(lambda  x: K.l2_normalize(x,axis=1), name=\"lambda\")(net)\n\n    # model creation\n    base_model = Model(resnet_model.input, net, name=\"base_model\")\n\n    # triplet framework, shared weights\n\n    input_anchor = Input(shape=input_shape, name='input_anchor')\n    input_positive = Input(shape=input_shape, name='input_pos')\n    input_negative = Input(shape=input_shape, name='input_neg')\n\n    net_anchor = base_model(input_anchor)\n    net_positive = base_model(input_positive)\n    net_negative = base_model(input_negative)\n\n    # The Lamda layer produces output using given function. Here its Euclidean distance.\n    positive_dist = Lambda(euclidean_distance, name='pos_dist')([net_anchor, net_positive])\n    negative_dist = Lambda(euclidean_distance, name='neg_dist')([net_anchor, net_negative])\n    tertiary_dist = Lambda(euclidean_distance, name='ter_dist')([net_positive, net_negative])\n\n    # This lambda layer simply stacks outputs so both distances are available to the objective\n    stacked_dists = Lambda(lambda vects: K.stack(vects, axis=1), name='stacked_dists')([positive_dist, negative_dist, tertiary_dist])\n\n    triplet_model = Model([input_anchor, input_positive, input_negative], stacked_dists, name='triple_siamese')\n\n    # Setting up optimizer designed for variable learning rate\n\n    # Variable Learning Rate per Layers\n    #lr_mult_dict = {}\n    #last_layer = ''\n    #for layer in resnet_model.layers:\n    #    # comment this out to refine earlier layers\n    #    # layer.trainable = False  \n    #    # print layer.name\n    #    lr_mult_dict[layer.name] = 1\n    #    # last_layer = layer.name\n    #lr_mult_dict['t_emb_1'] = 100\n\n    base_lr = 0.0001\n    momentum = 0.9\n    v_optimizer = SGD(lr=base_lr, momentum=momentum, decay=0.0, nesterov=False) #, multipliers = lr_mult_dict)\n\n    triplet_model.compile(optimizer=v_optimizer, loss=triplet_loss, metrics=[accuracy])\n\n    return embedding_model, triplet_model\n\n\ndef triplet_loss(y_true, y_pred):\n    margin = K.constant(1)\n    return K.mean(K.maximum(K.constant(0), K.square(y_pred[:,0,0]) - 0.5*(K.square(y_pred[:,1,0])+K.square(y_pred[:,2,0])) + margin))\n\ndef accuracy(y_true, y_pred):\n    return K.mean(y_pred[:,0,0] &lt; y_pred[:,1,0])\n\ndef l2Norm(x):\n    return  K.l2_normalize(x, axis=-1)\n\ndef euclidean_distance(vects):\n    x, y = vects\n    return K.sqrt(K.maximum(K.sum(K.square(x - y), axis=1, keepdims=True), K.epsilon()))           \n\n\nimport random\nfrom datetime import datetime\n\n#adam_optim = Adam(lr=0.0001, beta_1=0.9, beta_2=0.999)\n\n#checkpoint = ModelCheckpoint(path_model, monitor='loss', verbose=1, save_best_only=True, mode='min')\n#early = EarlyStopping(monitor=\"val_loss\", mode=\"min\", patience=2)\ncheckpoint = ModelCheckpoint(best_weights_filepath, monitor=\"val_loss\", save_best_only=True, save_weights_only=True, mode='auto')\ncallbacks_list = [checkpoint]  # , early]\n\ngen_train = gen(True)\ngen_valid = gen(False)\n\nfor i in range(0, len(arrParams)):\n  nL2 = arrParams[i][0]\n  EMBEDDING_DIM = arrParams[i][1]\n\n  deleteSavedNet(best_weights_filepath)\n\n  #random.seed(datetime.now())\n  embedding_model, triplet_model = createModel(nL2)\n\n  #loadBestModel()\n\n  # No need to set classifier layers to True, as already True as created (?)\n  nNumOfClasses = len(arrLabels)\n  nNumOfTrainSamples = TRAINING_IMAGES_PER_CLASS * nNumOfClasses\n  nNumOfValidSamples = VALIDATION_IMAGES_PER_CLASS * nNumOfClasses\n  STEP_SIZE_TRAIN = nNumOfTrainSamples \/\/ BATCH_SIZE\n  if(STEP_SIZE_TRAIN == 0):\n    STEP_SIZE_TRAIN = 1\n\n  STEP_SIZE_VALID = nNumOfValidSamples \/\/ BATCH_SIZE\n  if(STEP_SIZE_VALID == 0):\n    STEP_SIZE_VALID = 1\n\n  #triplet_model.compile(loss='mean_absolute_error', optimizer=Adam())\n\n  # MNIST\n  #triplet_model.compile(loss=None, optimizer=adam_optim, metrics=['accuracy', triplet_accuracy])\n\n  #triplet_model.compile(loss=None, optimizer=Adam(), metrics=['accuracy', triplet_accuracy])\n\n  #triplet_model.compile(loss='mean_absolute_error', optimizer=Adam())\n  #triplet_model.compile(loss=None, optimizer=\"adam\", metrics=['accuracy']) #metrics=['accuracy'])\n  #triplet_model.compile(loss=None, optimizer=SGD(lr=0.01, momentum=0.0, decay=0.0, nesterov=False), metrics=['accuracy'])\n  #triplet_model.compile(loss=None, optimizer=RMSprop(lr=0.001, rho=0.9, epsilon=None, decay=0.0), metrics=['accuracy'])\n  #triplet_model.compile(loss=None, optimizer=RMSprop(lr=0.00001, rho=0.9, epsilon=None, decay=0.0), metrics=['accuracy'])\n  #triplet_model.compile(loss=None, optimizer=Adagrad(lr=0.01, epsilon=None, decay=0.0), metrics=['accuracy'])\n  #triplet_model.compile(loss=None, optimizer=Adadelta(lr=1.0, rho=0.95, epsilon=None, decay=0.0), metrics=['accuracy'])\n  #triplet_model.compile(loss=None, optimizer=Adamax(lr=0.002, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0), metrics=['accuracy'])\n  #triplet_model.compile(loss=None, optimizer=Nadam(lr=0.002, beta_1=0.9, beta_2=0.999, epsilon=None, schedule_decay=0.004), metrics=['accuracy'])\n  #triplet_model.compile(loss=None, optimizer=Nadam(lr=0.0005, beta_1=0.8, beta_2=0.9, epsilon=None, schedule_decay=0.2), metrics=['accuracy'])\n\n  #SGD(lr=0.01, momentum=0.0, decay=0.0, nesterov=False)\n  #RMSprop(lr=0.001, rho=0.9, epsilon=None, decay=0.0)\n  #Adagrad(lr=0.01, epsilon=None, decay=0.0)\n  #Adadelta(lr=1.0, rho=0.95, epsilon=None, decay=0.0)\n  #Adamax(lr=0.002, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0)\n  #Nadam(lr=0.002, beta_1=0.9, beta_2=0.999, epsilon=None, schedule_decay=0.004)\n\n  #triplet_model.compile(loss=None, optimizer=Adam(0.00004), metrics=['accuracy'])\n\n  #triplet_model.compile(loss=None, optimizer=Adam(0.00004), metrics=['accuracy']) #, my_accuracy])\n  #triplet_model.compile(loss=None, optimizer=Adam(0.000004), metrics=['accuracy'])\n\n  print(\"Available metrics: \", triplet_model.metrics_names)\n\n  history = triplet_model.fit_generator(gen_train, validation_data=gen_valid, \n    epochs=EPOCHS, verbose=1, steps_per_epoch=STEP_SIZE_TRAIN, validation_steps=STEP_SIZE_VALID, callbacks=callbacks_list)\n    #, workers=4, use_multiprocessing=True)\n\n  print(nL2, EMBEDDING_DIM)\n  plotHistoryLoss()  \n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57494217,"AnswerCount":1,"Tags":"<python><python-3.x><pytorch><torch>","CreationDate":"2019-08-14T11:55:33.480","AcceptedAnswerId":null,"OwnerUserId":6224662.0,"Title":"Pytorch: Visualize model while training","Body":"<p>I am training a neural network by regression but it is predicting a constant value during testing. Which is why I want to visualize the weights of the neural network change during training and see the weights change dynamically in the <code>jupyter notebook<\/code>. <br>\nCurrently, my model looks like this: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nfrom torch import nn\n\nclass Net(nn.Module):\n    def __init__(self):\n        super(Net, self).__init__()\n        self.inp = nn.Linear(2, 40)\n        self.act1 = nn.Tanh()\n        self.h1 = nn.Linear(40, 40)\n        self.act2 = nn.Tanh()\n        self.h2 = nn.Linear(40, 2)\n        self.act3 = nn.Tanh()\n        #self.h3 = nn.Linear(20, 20)\n        #self.act4=nn.Tanh()\n        self.h4 = nn.Linear(2, 1)\n\n    def forward_one_pt(self, x):\n        out = self.inp(x)\n        out = self.act1(out)\n        out = self.h1(out)\n        out = self.act2(out)\n        out = self.h2(out)\n        out = self.act3(out)\n        #out = self.h3(out)\n        #out = self.act4(out)\n        out = self.h4(out)\n        return out\n\n    def forward(self, config):\n        E = torch.zeros([config.shape[0], 1])\n        for i in range(config.shape[0]):\n            E[i] = self.forward_one_pt(config[i])\n            # print(\"config[\",i,\"] = \",config[i],\"E[\",i,\"] = \",E[i])\n        return torch.sum(E, 0)\n<\/code><\/pre>\n\n<p>and my main function looks like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def main()  :\n    learning_rate = 0.5\n    n_pts = 1000\n    t_pts = 100\n    epochs = 15\n\n    coords,E = load_data(n_pts,t_pts)\n\n    #generating my data to NN\n    G = get_symm(coords,save,load_symmetry,symmtery_pickle_file,eeta1,eeta2,Rs,ex,lambdaa,zeta,boxl,Rc,pi,E,scale)\n    net = Net()\n    if(cuda_flag):\n        net.cuda()\n    criterion = nn.MSELoss()\n    optimizer = torch.optim.Adam(net.parameters(), lr=learning_rate)\n\n    net_trained = train(save,text_output,epochs,n_pts,G,E,net,t_pts,optimizer,criterion,out,cuda_flag)\n    test(save,n_pts,t_pts,G,E,net_trained,out,criterion,cuda_flag)\n\n    torch.save(net,save_model)\n\n<\/code><\/pre>\n\n<p>any tutorials or answers would be helpful<\/p>\n","answers":[{"AnswerId":"57495421","CreationDate":"2019-08-14T13:10:50.283","ParentId":null,"OwnerUserId":"5027315","Title":null,"Body":"<p>You can use <code>model.state_dict()<\/code> to see if your weights are updating across epochs:<\/p>\n\n<pre><code>old_state_dict = {}\nfor key in model.state_dict():\n    old_state_dict[key] = model.state_dict()[key].clone()\n\noutput = model(input)\n\nnew_state_dict = {}\nfor key in model.state_dict():\n    new_state_dict[key] = model.state_dict()[key].clone()\n\nfor key in old_state_dict:\n    if not (old_state_dict[key] == new_state_dict[key]).all():\n        print('Diff in {}'.format(key))\n    else:\n        print('NO Diff in {}'.format(key))\n<\/code><\/pre>\n\n<p>On a side note, you can vectorize your forward function instead of looping over it. Following would do the same job as your original forward function but much faster:<\/p>\n\n<pre><code>def forward(self, config):\n    out= self.forward_one_pt(config)\n    return torch.sum(out, 0)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57494484,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-14T12:14:17.603","AcceptedAnswerId":null,"OwnerUserId":9372721.0,"Title":"keras for adding two dense layers","Body":"<p>There are two inputs, x, and u, that generate the output y.  There is a linear relationship between x, u, and y, i.e. y = x wx + u wx.  I'm trying to calculate wx and wu from data.  Here is the code for model construction \/ fitting.<\/p>\n\n<pre><code>    n_train = 400\n    n_val = 100\n    train_u = u[:(n_train+n_val)]\n    train_x = x[:(n_train+n_val)]\n    train_y = y[:(n_train+n_val)]\n    test_u = u[(n_train+n_val):]\n    test_x = x[(n_train+n_val):]\n    test_y = y[(n_train+n_val):]\n    val_u = train_u[-n_val:]\n    val_x = train_x[-n_val:]\n    val_y = train_y[-n_val:]\n    train_u = train_u[:-n_val]\n    train_x = train_x[:-n_val]\n    train_y = train_y[:-n_val]\n\n    # RNN derived classes want a shape of (batch_size, timesteps, input_dim)\n    # batch_size. One sequence is one sample. A batch is comprised of one or more samples.\n    # timesteps. One time step is one point of observation in the sample.\n    # input_dim. number of observation at a time step.\n    # I believe n_train = one_epoch = batch_size * time_steps, features = nx_lags or nu_lags\n    # I also thing an epoch is one pass through the training data\n\n    n_batches_per_epoch = 8\n    n_iterations_per_batch = round(n_train \/ n_batches_per_epoch)\n    batch_size = n_batches_per_epoch\n    time_steps = n_iterations_per_batch\n    features_x = train_x.shape[1]\n    features_u = train_u.shape[1]\n    features_y = train_y.shape[1]\n\n    keras_train_u = train_u.values.reshape((batch_size, time_steps, features_u))\n    keras_train_x = train_x.values.reshape((batch_size, time_steps, features_x))\n    keras_train_y = train_y.reshape((batch_size, time_steps, features_y))\n    keras_val_u = val_u.values.reshape((2, time_steps, features_u))\n    keras_val_x = val_x.values.reshape((2, time_steps, features_x))\n    keras_val_y = val_y.reshape((2, time_steps, features_y))\n    keras_test_u = test_u.values.reshape((1, test_u.shape[0], features_u))\n    keras_test_x = test_x.values.reshape((1, test_u.shape[0], features_x))\n    keras_test_y = test_y.reshape((1, test_u.shape[0], features_y))\n\n    print('u.values.shape: ', u.values.shape)\n    # Now try a tensorflow model\n    # x_input = keras.Input(shape=(batch_size, time_steps, features_x), name='x_input')\n    # u_input = keras.Input(shape=(batch_size, time_steps, features_u), name='u_input')\n    x_input = keras.Input(shape=(time_steps, features_x), name='x_input')\n    u_input = keras.Input(shape=(time_steps, features_u), name='u_input')\n    da = layers.Dense(ny, name='dense_a', use_bias=False)(x_input)\n    db = layers.Dense(ny, name='dense_b', use_bias=False)(u_input)\n    output = layers.Add()([da, db])\n\n    model = keras.Model(inputs=[x_input, u_input], outputs=output)\n\n    model.compile(optimizer=keras.optimizers.RMSprop(),  # Optimizer\n                  # Loss function to minimize\n                  loss=keras.losses.SparseCategoricalCrossentropy(),\n                  # List of metrics to monitor\n                  metrics=[keras.metrics.SparseCategoricalAccuracy()])\n    print(model.summary())\n    print('keras_train_x.shape: ', keras_train_x.shape)\n    print('keras_train_u.shape: ', keras_train_u.shape)\n    print('keras_train_y.shape: ', keras_train_y.shape)\n    print('keras_val_x.shape: ', keras_val_x.shape)\n    print('keras_val_u.shape: ', keras_val_u.shape)\n    print('keras_val_y.shape: ', keras_val_y.shape)\n    history = model.fit([keras_train_x, keras_train_u], keras_train_y,\n                        batch_size=64,\n                        epochs=3,\n                        # We pass some validation for\n                        # monitoring validation loss and metrics\n                        # at the end of each epoch\n                        validation_data=([keras_val_x, keras_val_u], keras_val_y))\n<\/code><\/pre>\n\n<p>And, here is the output, with error.  <\/p>\n\n<pre><code>Model: \"model\"\n__________________________________________________________________________________________________\nLayer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\nx_input (InputLayer)            [(None, 50, 7)]      0                                            \n__________________________________________________________________________________________________\nu_input (InputLayer)            [(None, 50, 7)]      0                                            \n__________________________________________________________________________________________________\ndense_a (Dense)                 (None, 50, 2)        14          x_input[0][0]                    \n__________________________________________________________________________________________________\ndense_b (Dense)                 (None, 50, 2)        14          u_input[0][0]                    \n__________________________________________________________________________________________________\nadd (Add)                       (None, 50, 2)        0           dense_a[0][0]                    \n                                                                 dense_b[0][0]                    \n==================================================================================================\nTotal params: 28\nTrainable params: 28\nNon-trainable params: 0\n__________________________________________________________________________________________________\nNone\nkeras_train_x.shape:  (8, 50, 7)\nkeras_train_u.shape:  (8, 50, 7)\nkeras_train_y.shape:  (8, 50, 2)\nkeras_val_x.shape:  (2, 50, 7)\nkeras_val_u.shape:  (2, 50, 7)\nkeras_val_y.shape:  (2, 50, 2)\nTrain on 8 samples, validate on 2 samples\n\nEpoch 1\/3\nTraceback (most recent call last):\n  File \"arx_rnn.py\", line 487, in &lt;module&gt;\n    main()\n  File \"\/arx_rnn.py\", line 481, in main\n    rnn_prediction = x.rnn_n_steps(y_measured, u_control, n_to_predict)\n  File \"arx_rnn.py\", line 387, in rnn_n_steps\n    validation_data=([keras_val_x, keras_val_u], keras_val_y))\n  File \"venv\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 780, in fit\n    steps_name='steps_per_epoch')\n  File \"venv\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_arrays.py\", line 363, in model_iteration\n    batch_outs = f(ins_batch)\n  File \"venv\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py\", line 3292, in __call__\n    run_metadata=self.run_metadata)\n  File \"venv\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1458, in __call__\n    run_metadata_ptr)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: Can not squeeze dim[2], expected a dimension of 1, got 2\n     [[{{node metrics\/sparse_categorical_accuracy\/Squeeze}}]]\n\nProcess finished with exit code 1\n<\/code><\/pre>\n\n<p>What is the error message telling me, and how to correct?<\/p>\n","answers":[{"AnswerId":"57497836","CreationDate":"2019-08-14T15:34:11.253","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>Keras categorical accuracy metrics expect the output, &amp; labels, shape as <code>(batch_size,num_classes)<\/code>. The <code>dim[2]<\/code> in error message indicates output shape is 3d: <code>(None,50,2)<\/code><\/p>\n\n<p>The simple fix is to ensure, by whatever means, that the output layer gives <em>one<\/em> prediction per class per batch - i.e. has shape <code>(batch_size,num_classes)<\/code> - which can be done via <code>Reshape<\/code>, or <code>Flatten<\/code>.<\/p>\n\n<p>The better fix is to alter your input-output topology per design needs - namely, <em>what, exactly, are you classifying?<\/em> Your data dimensionality suggests that you seek to classify individual timesteps - in that case, feed data <em>one timestep at a time<\/em>: <code>(batch_size,features)<\/code>. Alternatively, feed timesteps in the <em>batch axis<\/em>, one batch at a time, so 1000 timesteps will correspond to <code>(1000,features)<\/code> - but <em>do not<\/em> do this if model has any <code>stateful<\/code> layers, which treats each batch axis entry as an <em>independent sequence<\/em>.<\/p>\n\n<p>To classify sequences with <code>timesteps&gt;1<\/code>, again, ensure layer data flow ultimately yields a 2d output.<\/p>\n"}]}
+{"QuestionId":57494548,"AnswerCount":1,"Tags":"<python><python-3.x><machine-learning><keras><deep-learning>","CreationDate":"2019-08-14T12:18:12.640","AcceptedAnswerId":null,"OwnerUserId":9678047.0,"Title":"How to load machine learning model in python 3.6 that is trained on python 3.5?","Body":"<p>I have trained a machine learning model on Python <code>3.5<\/code>, and now I switched to Google Colab, which uses Python <code>3.6<\/code>, and when I try to load the model that I have trained on Python <code>3.5<\/code>, it gives this error:    <\/p>\n\n<pre><code>SystemError: unknown opcode.\n<\/code><\/pre>\n\n<p>After googling, I found that this error occurs because of the environment change, then I cross-checked my python version, and both Python version were different. How can I load my model on Python <code>3.6<\/code>?<\/p>\n","answers":[{"AnswerId":"57494573","CreationDate":"2019-08-14T12:20:46.113","ParentId":null,"OwnerUserId":"4629950","Title":null,"Body":"<p>You shouldn't. <\/p>\n\n<p>Even if you get it to run without errors \/ warnings, there might be slight changes under the hood that change the behavior \/ performance of the model.<\/p>\n\n<p>You should either retrain the model on Python 3.6 or create a virtual environment that runs Python 3.5 for your model to ensure it performs as expected. Also always ensure that the actual libraries (e.g. keras...) have the same version.<\/p>\n"}]}
+{"QuestionId":57494638,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorflow-probability>","CreationDate":"2019-08-14T12:24:37.180","AcceptedAnswerId":57510862.0,"OwnerUserId":11281500.0,"Title":"Variable tf.Variable has 'None' for gradient in TensorFlow Probability","Body":"<p>I'm having trouble constructing a basic BNN in TFP. I'm new to TFP and BNNs in general, so I apologize if I've missed something simple.<\/p>\n\n<p>I can train a basic NN in Tensorflow by doing the following:<\/p>\n\n<pre><code>model = keras.Sequential([\n    keras.layers.Dense(units=100, activation='relu'),\n    keras.layers.Dense(units=50, activation='relu'),\n    keras.layers.Dense(units=5, activation='softmax')\n])\n\nmodel.compile(optimizer=optimizer, \n              loss=tf.losses.CategoricalCrossentropy(from_logits=True),\n              metrics=['accuracy'])\n\nhistory = model.fit(\n    training_data.repeat(), \n    epochs=100, \n    steps_per_epoch=(X_train.shape[0]\/\/1024),\n    validation_data=test_data.repeat(), \n    validation_steps=2\n)\n<\/code><\/pre>\n\n<p>However, I have trouble when trying to implement a similar architecture with tfp DenseFlipout layers:<\/p>\n\n<pre><code>model = keras.Sequential([\n    tfp.layers.DenseFlipout(units=100, activation='relu'),\n    tfp.layers.DenseFlipout(units=10, activation='relu'),\n    tfp.layers.DenseFlipout(units=5, activation='softmax')\n])\n\nmodel.compile(optimizer=optimizer, \n              loss=tf.losses.CategoricalCrossentropy(from_logits=True),\n              metrics=['accuracy'])\n\nhistory = model.fit(\n    training_data.repeat(), \n    epochs=100, \n    steps_per_epoch=(X_train.shape[0]\/\/1024),\n    validation_data=test_data.repeat(), \n    validation_steps=2\n)\n<\/code><\/pre>\n\n<p>I get the following Value error:<\/p>\n\n<pre><code>ValueError: \nVariable &lt;tf.Variable 'sequential_11\/dense_flipout_15\/kernel_posterior_loc:0' \nshape=(175, 100) dtype=float32&gt; has `None` for gradient. \nPlease make sure that all of your ops have a gradient defined (i.e. are differentiable). \nCommon ops without gradient: K.argmax, K.round, K.eval.\n<\/code><\/pre>\n\n<p>I've done some googling, and have looked around the TFP docs, but am at a loss so thought I would share the issue. Have I missed something obvious? <\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"57510862","CreationDate":"2019-08-15T14:01:47.323","ParentId":null,"OwnerUserId":"8520412","Title":null,"Body":"<p>I expect it's because you're using TensorFlow 2, are you? It isn't fully supported yet. If so, downgrading to 1.14 should get it working.<\/p>\n"}]}
+{"QuestionId":57494714,"AnswerCount":0,"Tags":"<python><ubuntu><pytorch><setup.py><caffe2>","CreationDate":"2019-08-14T12:28:29.657","AcceptedAnswerId":null,"OwnerUserId":7636381.0,"Title":"Can't install Caffe2 on Ubuntu 18.04","Body":"<p>I am trying to install <a href=\"https:\/\/github.com\/facebookresearch\/Detectron\" rel=\"nofollow noreferrer\">Facebook's Detectron<\/a> and so I need to install Caffe2. I followed <a href=\"https:\/\/caffe2.ai\/docs\/getting-started.html?platform=ubuntu&amp;configuration=compile\" rel=\"nofollow noreferrer\">this instruction<\/a>.<\/p>\n\n<p>Basically, I clone the PyTorch repo with<br> \n<code>git clone https:\/\/github.com\/pytorch\/pytorch.git &amp;&amp; cd pytorch\n<\/code> <code>git submodule update --init --recursive<\/code> <br> and when I run the following command <code>python setup.py install<\/code> <\/p>\n\n<p>it goes to 84-86% and I get this error. <\/p>\n\n<pre><code>[ 84%] Built target roi_align_op_gpu_test\nMakefile:140: recipe for target 'all' failed\nmake: *** [all] Error 2\nTraceback (most recent call last):\n  File \"setup.py\", line 756, in &lt;module&gt;\n    build_deps()\n  File \"setup.py\", line 321, in build_deps\n    cmake=cmake)\n  File \"\/home\/milos\/pytorch\/tools\/build_pytorch_libs.py\", line 63, in build_caffe2\n    cmake.build(my_env)\n  File \"\/home\/milos\/pytorch\/tools\/setup_helpers\/cmake.py\", line 330, in build\n    self.run(build_args, my_env)\n  File \"\/home\/milos\/pytorch\/tools\/setup_helpers\/cmake.py\", line 143, in run\n    check_call(command, cwd=self.build_dir, env=env)\n  File \"\/usr\/lib\/python2.7\/subprocess.py\", line 190, in check_call\n    raise CalledProcessError(retcode, cmd)\nsubprocess.CalledProcessError: Command '['cmake', '--build', '.', '--target', 'install', '--config', 'Release', '--', '-j', '12']' returned non-zero exit status 2\nSegmentation fault (core dumped)\n<\/code><\/pre>\n\n<p>System info:<\/p>\n\n<blockquote>\n  <p>Ubuntu 18.04 LTS (Bionic Beaver)<br>\n  Python 3.6.8 and Python 2.7.15<br>\n  OpenCV 3.4.0 for both python versions<br>\n  CUDA 10.2 (Nvidia 430.26)<br><\/p>\n<\/blockquote>\n\n<p>I've managed to get this working before, but can't remember how.<\/p>\n","answers":[]}
+{"QuestionId":57494855,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><tensor-indexing>","CreationDate":"2019-08-14T12:37:23.247","AcceptedAnswerId":null,"OwnerUserId":5378299.0,"Title":"How to read a tensor as a numpy array or list in Tensorflow?","Body":"<p>I have a tensor x,<\/p>\n\n<pre><code>x={Tensor} Tensor(\"Cast:0\", shape=(?,3), dtype=int32)\n<\/code><\/pre>\n\n<p>Now, I need to iterate over each triple of this tensor batch (say a triple is (a,b,c)) and fetch the first element (in this example, a) of that.<\/p>\n\n<p>Then, I need to fetch all other triples in the dataset Y (below) that also have 'a' as their first element.<\/p>\n\n<p>Ultimately, I wish to return all triples with 'a' as the first element, excluding the triple in question (i.e. in this case, excluding (a,b,c)).<\/p>\n\n<p>I was previously working with the same, assuming that x is a list.<\/p>\n\n<p>Therefore, in terms of list operations:<\/p>\n\n<pre><code>t=list({triple for x in x_to_score for triple in self.d[x[0]]} - set(x_to_score.eval()))\n<\/code><\/pre>\n\n<p>where d is a dictionary containing list of all the triples grouped by first elements. For example:<\/p>\n\n<p>For <\/p>\n\n<pre><code>    Y=np.array([['a', 'y', 'b'],\n              ['b', 'y', 'a'],\n              ['a', 'y', 'c'],\n              ['c', 'y', 'a'],\n              ['a', 'y', 'd'],\n              ['c', 'y', 'd'],\n              ['b', 'y', 'c'],\n              ['f', 'y', 'e']])\n\n    d={'f': [('f', 'y', 'e')],\n       'c': [('c', 'y', 'a'), ('c', 'y', 'd')],\n        'a': [('a', 'y', 'b'), ('a', 'y', 'c'), ('a', 'y', 'd')],\n        'b': [('b', 'y', 'a'), ('b', 'y', 'c')]}\n<\/code><\/pre>\n\n<p>However, I am new to tensorflow and cannot find a way to convert these operations into tensors. The result should also be of the order [?,3] for each triple evaluated.<\/p>\n\n<p>Please note that eager execution must be disabled.<\/p>\n\n<p>Any help is welcome!<\/p>\n\n<p><strong>EDIT:<\/strong>\nIf the input tensor x=(a,y,d)  (note that this can be a batch, so x=[(a,y,d),(b,y,c)] etc.), then the expected output will be:<\/p>\n\n<pre><code>[('a', 'y', 'b'), ('a', 'y', 'c')]\n<\/code><\/pre>\n","answers":[{"AnswerId":"57497784","CreationDate":"2019-08-14T15:30:06.097","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>One possible solution for this is to compute to array results, one with the concatenated matches for all inputs and another indicating the index of the input it refers to. It could work like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\nX = tf.placeholder(tf.string, [None, 3])\nY = tf.constant([\n    ['a', 'y', 'b'],\n    ['b', 'y', 'a'],\n    ['a', 'y', 'c'],\n    ['c', 'y', 'a'],\n    ['a', 'y', 'd'],\n    ['c', 'y', 'd'],\n    ['b', 'y', 'c'],\n    ['f', 'y', 'e']\n])\n# Compare every input to every data\ncmp = tf.equal(tf.expand_dims(X, 1), Y)\n# Triples matching the first element\nfirst_match = cmp[:, :, 0]\n# Triples matching all elements\nall_match = tf.reduce_all(cmp, axis=2)\n# Triple matching the first element but not all elements\nmatches = first_match &amp; (~all_match)\n# Find indices\nmatch_idx = tf.where(matches)\n# Index of the input triple each result refers to\nmatch_x = match_idx[:, 0]\n# Concatenated resulting triples\nmatch_y = tf.gather(Y, match_idx[:, 1])\n# Test\nwith tf.Graph().as_default(), tf.Session() as sess:\n    x_val = [['a', 'y', 'd'], ['b', 'y', 'c']]\n    match_x_val, match_y_val = sess.run((match_x, match_y), feed_dict={X: x_val})\n    print(*zip(match_x_val, match_y_val), sep='\\n')\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>(0, array([b'a', b'y', b'b'], dtype=object))\n(0, array([b'a', b'y', b'c'], dtype=object))\n(1, array([b'b', b'y', b'a'], dtype=object))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57495083,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2019-08-14T12:50:38.543","AcceptedAnswerId":null,"OwnerUserId":10085055.0,"Title":"Using Keras like TensorFlow for gpu computing","Body":"<p>I would like to know if Keras can be used as an interface to TensoFlow for only doing computation on my GPU. <\/p>\n\n<p>I tested TF directly on my GPU. But for ML purposes, I started using Keras, including the backend. I would find it 'comfortable' to do all my stuff in Keras instead of Using two tools. <\/p>\n\n<p>This is also a matter of curiosity.<\/p>\n\n<p>I found some examples like this one:<\/p>\n\n<p><a href=\"http:\/\/christopher5106.github.io\/deep\/learning\/2018\/10\/28\/understand-batch-matrix-multiplication.html\" rel=\"nofollow noreferrer\">http:\/\/christopher5106.github.io\/deep\/learning\/2018\/10\/28\/understand-batch-matrix-multiplication.html<\/a><\/p>\n\n<p>However this example does not actually do the calculation.\nIt also does not get input data.\nI duplicate the snippet here:\n'''<\/p>\n\n<pre><code>from keras import backend as K\na = K.ones((3,4))\nb = K.ones((4,5))\nc = K.dot(a, b)\nprint(c.shape)\n<\/code><\/pre>\n\n<p>'''<\/p>\n\n<p>I would simply like to know if I can get the result numbers from this snippet above, and how?<\/p>\n\n<p>Thanks,<\/p>\n\n<p>Michel<\/p>\n","answers":[{"AnswerId":"57495483","CreationDate":"2019-08-14T13:13:58.273","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Keras doesn't have an eager mode like Tensorflow, and it depends on models or functions with \"placeholders\" to receive and output data.<\/p>\n\n<p>So, it's a little more complicated than Tensorflow to do basic calculations like this. <\/p>\n\n<p>So, the most user friendly solution would be creating a dummy model with one <code>Lambda<\/code> layer. (And be careful with the first dimension that Keras will insist to understand as a batch dimension and require that input and output have the same batch size)<\/p>\n\n<pre><code>def your_function_here(inputs):\n\n    #if you have more than one tensor for the inputs, it's a list:\n    input1, input2, input3 = inputs\n\n    #if you don't have a batch, you should probably have a first dimension = 1 and get\n    input1 = input1[0]\n\n    #do your calculations here\n\n    #if you used the batch_size=1 workaround as above, add this dimension again:\n    output = K.expand_dims(output,0)\n\n    return output\n<\/code><\/pre>\n\n<p>Create your model:<\/p>\n\n<pre><code>inputs = Input(input_shape)\n#maybe inputs2 ....\n\noutputs = Lambda(your_function_here)(list_of_inputs)\n#maybe outputs2\n\nmodel = Model(inputs, outputs)\n<\/code><\/pre>\n\n<p>And use it to predict the result:<\/p>\n\n<pre><code>print(model.predict(input_data))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57495084,"AnswerCount":0,"Tags":"<apache-spark><tensorflow><pyspark><apache-spark-mllib>","CreationDate":"2019-08-14T12:50:43.043","AcceptedAnswerId":null,"OwnerUserId":2786065.0,"Title":"Data column inputCol has not been analyzed yet, cannot run TF on this dataframe","Body":"<p>I am trying to use universal sentence encoder as a TFTransformer for a text analytics pipeline in spark. Code for creating TFTransformer from the tensor flow graph is given below.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport tensorflow_hub as hub\nfrom sparkdl import TFTransformer\nfrom sparkdl.graph.input import TFInputGraph\nimport sparkdl.graph.utils as tfx\nfrom pyspark.sql.types import Row\n\nwith tf.Graph().as_default():\n  module = hub.Module(\"https:\/\/tfhub.dev\/google\/universal-sentence-encoder-large\/3\")\n  text = tf.placeholder(tf.string, [None], name='textinput')\n  embedding = module(text)\n  init_op = tf.group([tf.global_variables_initializer(), tf.tables_initializer()])\n  with tf.Session() as sess:\n    sess.run(init_op)\n    gin = TFInputGraph.fromGraph(sess.graph, sess, [\"textinput:0'\"], [\"module_apply_default\/Encoder_en\/hidden_layers\/l2_normalize:0\"])\n\ntransformer = TFTransformer(\n    tfInputGraph=gin,\n    inputMapping={'inputCol': tfx.tensor_name(\"textinput:0\")},\n    outputMapping={tfx.tensor_name(\"module_apply_default\/Encoder_en\/hidden_layers\/l2_normalize:0\"): 'outputCol'})\n\nrows = [Row(inputCol=[\"how are you\"]), Row(inputCol=[\"what are you upto\"])]\nsdf = spark.createDataFrame(rows)\nodf = transformer.transform(sdf)\n<\/code><\/pre>\n\n<p>However, I am getting the following error<\/p>\n\n<pre class=\"lang-java prettyprint-override\"><code>java.lang.Exception: Data column inputCol has not been analyzed yet, cannot run TF on this dataframe\n    at org.tensorframes.impl.DebugRowOps$$anonfun$26$$anonfun$27.apply(DebugRowOps.scala:329)\n    at org.tensorframes.impl.DebugRowOps$$anonfun$26$$anonfun$27.apply(DebugRowOps.scala:329)\n    at scala.Option.getOrElse(Option.scala:121)\n    at org.tensorframes.impl.DebugRowOps$$anonfun$26.apply(DebugRowOps.scala:328)\n    at org.tensorframes.impl.DebugRowOps$$anonfun$26.apply(DebugRowOps.scala:320)\n<\/code><\/pre>\n\n<p>Not sure what do we mean my \"analyzed\" data frame<\/p>\n","answers":[]}
+{"QuestionId":57495344,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-08-14T13:05:46.977","AcceptedAnswerId":57495655.0,"OwnerUserId":7248145.0,"Title":"Custom layer updates","Body":"<p>i want to create a custom layer with weights that update only in training phase.<\/p>\n\n<p>from the official documentation this is the way:<\/p>\n\n<pre><code>from keras import backend as K\nfrom keras.layers import Layer\n\nclass MyLayer(Layer):\n\n    def __init__(self, output_dim, **kwargs):\n        self.output_dim = output_dim\n        super(MyLayer, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        # Create a trainable weight variable for this 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(MyLayer, self).build(input_shape)  # Be sure to call this at the end\n\n    def call(self, x):\n        return K.dot(x, self.kernel)\n\n    def compute_output_shape(self, input_shape):\n        return (input_shape[0], self.output_dim)\n<\/code><\/pre>\n\n<p>in <a href=\"https:\/\/github.com\/handongfeng\/MNIST-center-loss\/blob\/72cf16bf351868f3400fa037b7332856ba5c565c\/centerLoss_MNIST.py#L61\" rel=\"nofollow noreferrer\">this github repo<\/a> \nthe author added<\/p>\n\n<pre><code>new_centers = self.centers - self.alpha * delta_centers\nself.add_update((self.centers, new_centers), x)\n<\/code><\/pre>\n\n<p>where <code>self.centers<\/code> are the weights.<\/p>\n\n<p>I cant understand why self.add_update is useful in that situation.\nWeights are not updated if i dont add self.add_update? If not, why <code>new_centers<\/code> must be in the updates list and not in the inputs list?And why x is a requirement?<\/p>\n\n<p>from the source code, \n<code>self.add_update(updates, inputs)<\/code><\/p>\n\n<p>updates: update op or list of update ops to add to the layer.<\/p>\n\n<p>inputs: input tensor or list of inputs tensors to mark the updates as conditional on these inputs.If None is passed, the updates are assumed unconditional.<\/p>\n","answers":[{"AnswerId":"57495655","CreationDate":"2019-08-14T13:23:26.557","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>There are two types of weights:<\/p>\n\n<ul>\n<li>Trainable = Updated automatically by the optimizer with backpropagation    <\/li>\n<li>Untrainable = Not updated by backpropagation    <\/li>\n<\/ul>\n\n<p>For the trainable weights, it's really not recommended to use updates, you will be mixing the optimizer's updates with your own updates and that could cause many issues    <\/p>\n\n<p>For the untrainable weights, you can do whatever you want. Sometimes you want constants and you will do nothing, sometimes, you want these weights to change (but not via backpropagation)<\/p>\n\n<p>Notice how in that example the weights updated by the user are untrainable:<\/p>\n\n<pre><code> self.centers = self.add_weight(name='centers',\n                                   shape=(10, 2),\n                                   initializer='uniform',\n\n                                   #UNTRAINABLE \n                                   trainable=False)\n<\/code><\/pre>\n\n<p>But the user wants these weights to be updated following some rules. I don't know what they are doing there (didn't analyse the code), but I assume that they are calculating, for instance, something similar to the center point of a group of images, and each batch will have this center in a different position. They want to update this position. <\/p>\n\n<p>A classical example is the <code>BatchNormalization<\/code> layer. Besides having trainable <code>scale<\/code> and <code>bias<\/code> weights used to rescale the outputs, they have the <code>mean<\/code> and <code>variance<\/code> weights. These are statistical properties of the data that need to be updated with every batch. <\/p>\n\n<p>You are not training the \"mean\" or the \"variance\", but each batch of data updates these values. <\/p>\n\n<hr>\n\n<h2>How does it work?<\/h2>\n\n<p>This is obscure and lies deep down in Keras code. <\/p>\n\n<p>We need the update operation so we make sure <code>self.centers<\/code> will have new values for every batch, otherwise it won't. <\/p>\n\n<p>We use <code>self.add_update<\/code> in a layer to register that this variable should be updated. (We do similar things in custom optimizers as well, the optimizers contain the updates to the weights made via backpropagation)    <\/p>\n\n<p>Later in the <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/engine\/training.py\" rel=\"nofollow noreferrer\">source code for training the model<\/a>, Keras will collect all these registered updates and make a train function. Somewhere inside this, these updates will be applied to the vars:<\/p>\n\n<pre><code>#inside a training function from keras\nwith K.name_scope('training'):\n            with K.name_scope(self.optimizer.__class__.__name__):\n                training_updates = self.optimizer.get_updates(\n                    params=self._collected_trainable_weights,\n                    loss=self.total_loss)\n            updates = (self.updates +   #probably the updates registered in layers    \n                       training_updates + #the updates registered in optimizers    \n                       self.metrics_updates) #don't know.... \n            # Gets loss and metrics. Updates weights at each call.\n            self.train_function = K.function(\n                inputs,\n                [self.total_loss] + self.metrics_tensors,\n                updates=updates,\n                name='train_function',\n                **self._function_kwargs)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57495524,"AnswerCount":0,"Tags":"<django><apache><tensorflow><tensorboard>","CreationDate":"2019-08-14T13:16:29.217","AcceptedAnswerId":null,"OwnerUserId":297937.0,"Title":"How to run TensorBoard + Django + Apache + 2 TensorFlow log folders","Body":"<p>I have:<\/p>\n\n<ul>\n<li>TensorFlow. <\/li>\n<li>Django website<\/li>\n<li>Apache<\/li>\n<li>many users. = many TF instances<\/li>\n<\/ul>\n\n<p>How to run 1 TensorBoard + 1 Django + 1 Apache + for many TensorFlow log folders ?<\/p>\n","answers":[]}
+{"QuestionId":57495635,"AnswerCount":0,"Tags":"<python><linux><tensorflow>","CreationDate":"2019-08-14T13:22:39.330","AcceptedAnswerId":null,"OwnerUserId":11927423.0,"Title":"Tensorflow pb file use\u2014\u2014sess.run() can not stop","Body":"<p>My english is not good, but I'll try my best to describe my problem. I hope someone can help me.\nI convert my tensorflow-based object detection model to '.pb' file and try to use it. But after I load the pb file and run sess.run(sth), it can't stop and no message show in my debug\/run window.<\/p>\n\n<p>OS: linux ubuntu 16.04\npython\uff1a2.7\ntensorflow\uff1a1.5.0 gpu version\nIDE:pycharm\nGPU:Tesla P100-PCIE-16GB\nI've tried change several different input_tensor, but all situations are same. I ran 'nvidia-smi' and find it use GPU normally. I also tried tfdebug tools, after I choose 'run' it also can't stop and no message for me.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#!\/usr\/bin\/python\n# -*-coding=utf-8-*-\n\nimport os\nimport cv2\nimport numpy as np\nimport tensorflow as tf\nfrom tensorflow.python import debug as tf_debug\n\nos.environ['CUDA_VISIBLE_DEVICES'] = '0'\nPIXEL_MEAN = [123.68, 116.779, 103.939]\n\ndef SaveWeightsFile():\n    saver = tf.train.import_meta_graph(\".\/output\/trained_weights\/RRPN_20180901_DOTA_v1\/voc_98001model.ckpt.meta\",clear_devices=True)\n    graph = tf.get_default_graph()\n    input_graph_def = graph.as_graph_def()\n    sess = tf.Session()\n    saver.restore(sess,\n        \".\/output\/trained_weights\/RRPN_20180901_DOTA_v1\/voc_98001model.ckpt\")\n    for node in input_graph_def.node:\n        # if \"det\" in node.name:\n        #     print(node.name)\n        # if \"img\" in node.name:\n        #     print(node.name)\n        if \"resnet_v1_101\" in node.name:\n            print(node.name)\n        print(node.name)\n    output_node_names = \"get_batch\/ResizeBilinear,postprocess_fastrcnn\/Gather_14,postprocess_fastrcnn\/Gather_15,postprocess_fastrcnn\/Gather_16\"\n    output_graph_def = tf.graph_util.convert_variables_to_constants(\n        sess,\n        input_graph_def,\n        output_node_names.split(\",\")\n    )\n    output_graph = \".\/output_graph\/saved_weights.pb\"\n    with tf.gfile.GFile(output_graph, 'wb') as f:\n        f.write(output_graph_def.SerializeToString())\n    sess.close()\n    print(\"Done.\")\n\n\ndef LoadAndUse():\n    frozen_graph = \".\/trained_weights\/output_graph\/saved_weights.pb\"\n    with tf.gfile.GFile(frozen_graph, '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(\n            graph_def,\n            input_map=None,\n            return_elements=None,\n            name=\"\"\n        )\n    input_tensor = graph.get_tensor_by_name(\"get_batch\/Cast_3:0\")\n    det_boxes = graph.get_tensor_by_name(\"postprocess_fastrcnn\/Gather_14:0\")\n    det_scores = graph.get_tensor_by_name(\"postprocess_fastrcnn\/Gather_15:0\")\n    det_category =graph.get_tensor_by_name(\"postprocess_fastrcnn\/Gather_16:0\")\n    img_test=cv2.imread(\".\/images\/12.png\")\n    sess = tf.Session(graph=graph)\n    result = sess.run([det_boxes, det_scores, det_category],\n             feed_dict={input_tensor: img_test})## can not stop!\n    print(\"Done.\")\n\nif __name__ == '__main__':\n    SaveWeightsFile()\n    LoadAndUse()\n    print(\"Done\")\n<\/code><\/pre>\n\n<p>Normally, it can run graph and give me 'det_boxes, det_scores, det_category' three values, but 'sess.run()' did't stop and no error message, it just always run.<\/p>\n","answers":[]}
+{"QuestionId":57495665,"AnswerCount":3,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-14T13:24:00.790","AcceptedAnswerId":null,"OwnerUserId":6405137.0,"Title":"turn off tensorflow messages","Body":"<p>How can I turn off tensorflow messages such as :<\/p>\n\n<pre><code>2019-08-14 16:07:26.928502: I tensorflow\/core\/common_runtime\/placer.cc:54] Adam_3\/decay\/initial_value: (Const)\/job:localhost\/replica:0\/task:0\/device:CPU:0\n\n2019-08-14 16:07:26.928520: I tensorflow\/core\/common_runtime\/placer.cc:54] noisy_layer_4_target: (Placeholder)\/job:localhost\/replica:0\/task:0\/device:CPU:0\n....\n<\/code><\/pre>\n\n<p>I've tried :<\/p>\n\n<pre><code>os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'\nos.environ['TF_CPP_MIN_VLOG_LEVEL'] = '3'\ntf.logging.set_verbosity(tf.logging.ERROR) # also tried  DEBUG, FATAL, WARN, INFO\nlogging.getLogger('tensorflow').disabled = True\n<\/code><\/pre>\n\n<p>from other SO questions but they don't change anything.<\/p>\n\n<p>I'm using keras with tf as backend , with tf.version = 1.14, in spyder<\/p>\n","answers":[{"AnswerId":"57497465","CreationDate":"2019-08-14T15:09:34.160","ParentId":null,"OwnerUserId":"11127923","Title":null,"Body":"<pre class=\"lang-py prettyprint-override\"><code>import logging, os \nlogging.disable(logging.WARNING) \nos.environ[\"TF_CPP_MIN_LOG_LEVEL\"] = \"3\"\n<\/code><\/pre>\n"},{"AnswerId":"57496678","CreationDate":"2019-08-14T14:27:23.623","ParentId":null,"OwnerUserId":"10400883","Title":null,"Body":"<p>See <a href=\"https:\/\/stackoverflow.com\/a\/40426709\/10400883\">this answer<\/a>.<\/p>\n\n<blockquote>\n  <p>Setting <code>tf.logging.set_verbosity<\/code> to ERROR does not always completely block all INFO logs.  <\/p>\n<\/blockquote>\n\n<p>In Linux, you can try below.  <\/p>\n\n<blockquote>\n  <p>If you are using Linux, you can just grep out all output strings beginning with <code>I tensorflow\/<\/code>  <\/p>\n<\/blockquote>\n\n<p>Example:  <\/p>\n\n<pre><code>$ python main.py | grep -v \"unwanted_word\"\n<\/code><\/pre>\n"},{"AnswerId":"57496271","CreationDate":"2019-08-14T14:02:09.580","ParentId":null,"OwnerUserId":"6862615","Title":null,"Body":"<h1>append to lines above <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/logging\" rel=\"nofollow noreferrer\">tf_logging<\/a><\/h1>\n\n<pre><code>tf.logging.set_verbosity(tf.logging.ERROR)  # or any {DEBUG, INFO, WARN, ERROR, FATAL}\n<\/code><\/pre>\n"}]}
+{"QuestionId":57496054,"AnswerCount":0,"Tags":"<python><tensorflow><google-cloud-ml>","CreationDate":"2019-08-14T13:49:12.930","AcceptedAnswerId":null,"OwnerUserId":9371809.0,"Title":"Running out of memory on ai engine without changing batch size","Body":"<p>I am training a tensorflow ml model on google clouds AI engine.  I am loading my data into a tf.data.Dataset, and use a batch size of 10. I first try to train my model on one dataset, and it works fine. As I understand the model just trains on one batch at the time, so the amount of data in one batch should determine if I run out of memory of not. Is this right? <\/p>\n\n<p>When I train on a larger dataset, I do run out of memory even though one batch is of the same size as with the smaller dataset. I don't understand how this could happen since it is the size of one batch that counts. Is my reasoning here wrong? <\/p>\n","answers":[]}
+{"QuestionId":57496285,"AnswerCount":2,"Tags":"<python><memory-leaks><garbage-collection><gpu><pytorch>","CreationDate":"2019-08-14T14:03:17.447","AcceptedAnswerId":null,"OwnerUserId":610569.0,"Title":"Why is the memory in GPU still in use after clearing the object?","Body":"<p>Starting with zero usage:<\/p>\n\n<pre><code>&gt;&gt;&gt; import gc\n&gt;&gt;&gt; import GPUtil\n&gt;&gt;&gt; import torch\n&gt;&gt;&gt; GPUtil.showUtilization()\n| ID | GPU | MEM |\n------------------\n|  0 |  0% |  0% |\n|  1 |  0% |  0% |\n|  2 |  0% |  0% |\n|  3 |  0% |  0% |\n<\/code><\/pre>\n\n<p>Then I create a big enough tensor and hog the memory:<\/p>\n\n<pre><code>&gt;&gt;&gt; x = torch.rand(10000,300,200).cuda()\n&gt;&gt;&gt; GPUtil.showUtilization()\n| ID | GPU | MEM |\n------------------\n|  0 |  0% | 26% |\n|  1 |  0% |  0% |\n|  2 |  0% |  0% |\n|  3 |  0% |  0% |\n<\/code><\/pre>\n\n<p>Then I tried several ways to see if the tensor disappears.<\/p>\n\n<p><strong>Attempt 1:<\/strong> Detach, send to CPU and overwrite the variable<\/p>\n\n<p><strong>No, doesn't work.<\/strong><\/p>\n\n<pre><code>&gt;&gt;&gt; x = x.detach().cpu()\n&gt;&gt;&gt; GPUtil.showUtilization()\n| ID | GPU | MEM |\n------------------\n|  0 |  0% | 26% |\n|  1 |  0% |  0% |\n|  2 |  0% |  0% |\n|  3 |  0% |  0% |\n<\/code><\/pre>\n\n<p><strong>Attempt 2:<\/strong> Delete the variable<\/p>\n\n<p><strong>No, this doesn't work either<\/strong><\/p>\n\n<pre><code>&gt;&gt;&gt; del x\n&gt;&gt;&gt; GPUtil.showUtilization()\n| ID | GPU | MEM |\n------------------\n|  0 |  0% | 26% |\n|  1 |  0% |  0% |\n|  2 |  0% |  0% |\n|  3 |  0% |  0% |\n<\/code><\/pre>\n\n<p><strong>Attempt 3:<\/strong> Use the <code>torch.cuda.empty_cache()<\/code> function<\/p>\n\n<p>Seems to work, but it seems that there are some lingering overheads...<\/p>\n\n<pre><code>&gt;&gt;&gt; torch.cuda.empty_cache()\n&gt;&gt;&gt; GPUtil.showUtilization()\n| ID | GPU | MEM |\n------------------\n|  0 |  0% |  5% |\n|  1 |  0% |  0% |\n|  2 |  0% |  0% |\n|  3 |  0% |  0% |\n<\/code><\/pre>\n\n<p><strong>Attempt 4:<\/strong> Maybe clear the garbage collector.<\/p>\n\n<p><strong>No, 5% is still being hogged<\/strong><\/p>\n\n<pre><code>&gt;&gt;&gt; gc.collect()\n0\n&gt;&gt;&gt; GPUtil.showUtilization()\n| ID | GPU | MEM |\n------------------\n|  0 |  0% |  5% |\n|  1 |  0% |  0% |\n|  2 |  0% |  0% |\n|  3 |  0% |  0% |\n<\/code><\/pre>\n\n<p><strong>Attempt 5:<\/strong> Try deleting <code>torch<\/code> altogether (as if that would work when <code>del x<\/code> didn't work -_- )<\/p>\n\n<p><strong>No, it doesn't...*<\/strong><\/p>\n\n<pre><code>&gt;&gt;&gt; del torch\n&gt;&gt;&gt; GPUtil.showUtilization()\n| ID | GPU | MEM |\n------------------\n|  0 |  0% |  5% |\n|  1 |  0% |  0% |\n|  2 |  0% |  0% |\n|  3 |  0% |  0% |\n<\/code><\/pre>\n\n<p>And then I tried to check <code>gc.get_objects()<\/code> and it looks like there's still quite a lot of odd <code>THCTensor<\/code> stuff inside... <\/p>\n\n<p><strong>Any idea why is the memory still in use after clearing the cache?<\/strong><\/p>\n","answers":[{"AnswerId":"57569472","CreationDate":"2019-08-20T08:21:34.797","ParentId":null,"OwnerUserId":"9564645","Title":null,"Body":"<p>There is a part in PyTorch docs, which seems quite relevant:<br>\n<a href=\"https:\/\/pytorch.org\/docs\/stable\/notes\/cuda.html#memory-management\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/notes\/cuda.html#memory-management<\/a><\/p>\n\n<blockquote>\n  <p>Memory management<\/p>\n  \n  <p>PyTorch uses a caching memory allocator to speed up memory\n  allocations. This allows fast memory deallocation without device\n  synchronizations. However, <strong>the unused memory managed by the allocator\n  will still show as if used in nvidia-smi<\/strong>. You can use\n  memory_allocated() and max_memory_allocated() to monitor memory\n  occupied by tensors, and use memory_cached() and max_memory_cached()\n  to monitor memory managed by the caching allocator. Calling\n  empty_cache() releases all unused cached memory from PyTorch so that\n  those can be used by other GPU applications. However, the occupied GPU\n  memory by tensors will not be freed so it can not increase the amount\n  of GPU memory available for PyTorch.<\/p>\n<\/blockquote>\n\n<p>I bolded a part mentioning <strong>nvidia-smi<\/strong>, which as far as I know is used by GPUtil.<\/p>\n"},{"AnswerId":"57502757","CreationDate":"2019-08-14T22:19:23.247","ParentId":null,"OwnerUserId":"220700","Title":null,"Body":"<p>It looks like PyTorch's caching allocator reserves some fixed amount of memory even if there are no tensors, and this allocation is triggered by the first CUDA memory access\n(<code>torch.cuda.empty_cache()<\/code> deletes unused tensor from the cache, but the cache itself still uses some memory).<\/p>\n\n<p>Even with a tiny 1-element tensor, after <code>del<\/code> and <code>torch.cuda.empty_cache()<\/code>, <code>GPUtil.showUtilization(all=True)<\/code> reports exactly the same amount of GPU memory used as for a huge tensor (and both <code>torch.cuda.memory_cached()<\/code> and <code>torch.cuda.memory_allocated()<\/code> return zero).<\/p>\n"}]}
+{"QuestionId":57496469,"AnswerCount":1,"Tags":"<keras><deep-learning><nlp><chatbot><data-analysis>","CreationDate":"2019-08-14T14:14:13.673","AcceptedAnswerId":57496509.0,"OwnerUserId":10808868.0,"Title":"Does the size of a database affect the predictions speed of a model","Body":"<p>I am building a chatbot model in keras, and I am planning on using it on a raspberry pi. I have a huge database with the size of (1000000, 15, 100) which means there are 1 million sample with a maximum of 15 words and the embedding dimensions are 100d using GloVe. I build a simple model consists of 1 embedding layer, 1 bidirectional lstm layer, 1 droput layer and 2 dense layer with output shape of (25,).<\/p>\n\n<p>I know that because of the huge database the training process is going to take long, but does the size of the database going to affect the speed of model.predict or does the speed only influenced by the structure of the model and the shape of the input?<\/p>\n","answers":[{"AnswerId":"57496509","CreationDate":"2019-08-14T14:16:50.853","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>No, the size of the dataset does not affect the prediction speed of the model per se, as you say prediction computation time is only affected by the architecture of the model, and the dimensionality of the inputs.<\/p>\n\n<p>In general the problem with making small models that are fast in embedded hardware is that as small model (with less parameters) might not perform as well as a more complex model (in terms of accuracy or error), so you have to perform a trade-off between model complexity and computational performance.<\/p>\n"}]}
+{"QuestionId":57496837,"AnswerCount":1,"Tags":"<python><pandas><tensorflow><keras>","CreationDate":"2019-08-14T14:35:33.890","AcceptedAnswerId":57496986.0,"OwnerUserId":4876561.0,"Title":"Pandas dataframe not shaped correctly for Keras ANN","Body":"<p>I am working on a neural network for regression based off of <a href=\"https:\/\/stackoverflow.com\/questions\/49008074\/how-to-create-a-neural-network-with-regression-model\">this question<\/a><\/p>\n\n<p>My code looks like this:<\/p>\n\n<pre><code>import numpy as np\nfrom keras.layers import Dense, Activation\nfrom keras.models import Sequential\nfrom sklearn.model_selection import train_test_split\nimport matplotlib.pyplot as plt\nimport pandas as pd\n\n\n# Importing the dataset\ndata = pd.read_csv('binned.csv')\n\n# create the labels, or field we are trying to estimate\nlabel = data['TOTAL_DAYS_TO_COMPLETE']\n\n# create the data, or the data that is to be estimated\ndata = data.drop('TOTAL_DAYS_TO_COMPLETE', axis=1)\ndata = data.drop('SERIALNUM', axis=1)\n\nprint(data)\n\n# # split into training and testing sets\nX_train, X_test, y_train, y_test = train_test_split(data, label, test_size = 0.2)\n\n# Feature Scaling\nfrom sklearn.preprocessing import StandardScaler\nsc = StandardScaler()\nX_train = sc.fit_transform(X_train)\nX_test = sc.transform(X_test)\n\n# Initialising the ANN\nmodel = Sequential()\n\n# Adding the input layer and the first hidden layer\nmodel.add(Dense(32, activation = 'relu', input_dim = 6))\n\n# Adding the second hidden layer\nmodel.add(Dense(units = 32, activation = 'relu'))\n\n# Adding the third hidden layer\nmodel.add(Dense(units = 32, activation = 'relu'))\n\n# Adding the output layer\nmodel.add(Dense(units = 1))\n\n#model.add(Dense(1))\n# Compiling the ANN\nmodel.compile(optimizer = 'adam', loss = 'mean_squared_error')\n\n# Fitting the ANN to the Training set\nmodel.fit(X_train, y_train, batch_size = 10, epochs = 100)\n\ny_pred = model.predict(X_test)\n\nplt.plot(y_test, color = 'red', label = 'Real data')\nplt.plot(y_pred, color = 'blue', label = 'Predicted data')\nplt.title('Prediction')\nplt.legend()\nplt.show()\n<\/code><\/pre>\n\n<p>When running it, I get an error:<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File \"ann.py\", line 50, in\n  \n      model.fit(X_train[0:1], y_train, batch_size = 10, epochs = 100)   File \"C:\\Python367-64\\lib\\site-packages\\keras\\engine\\training.py\",\n  line 952, in fit\n      batch_size=batch_size)   File \"C:\\Python367-64\\lib\\site-packages\\keras\\engine\\training.py\", line\n  751, in _standardize_user_data\n      exception_prefix='input')   File \"C:\\Python367-64\\lib\\site-packages\\keras\\engine\\training_utils.py\",\n  line 138, in standardize_input_data\n      <strong>str(data_shape)) ValueError: Error when checking input: expected dense_1_input to have shape (6,) but got array with shape (24,)<\/strong><\/p>\n<\/blockquote>\n\n<p>I have used the same <\/p>\n\n<pre><code># Importing the dataset\ndata = pd.read_csv('binned.csv')\n\n# create the labels, or field we are trying to estimate\nlabel = data['TOTAL_DAYS_TO_COMPLETE']\n\n# create the data, or the data that is to be estimated\ndata = data.drop('TOTAL_DAYS_TO_COMPLETE', axis=1)\ndata = data.drop('SERIALNUM', axis=1)\n\nprint(data)\n\n# # split into training and testing sets\nX_train, X_test, y_train, y_test = train_test_split(data, label, test_size = 0.2)\n<\/code><\/pre>\n\n<p>across various <code>sklearn<\/code> libraries and it runs just fine.<\/p>\n\n<p>What am I doing wrong? <\/p>\n\n<p>My <code>print(data)<\/code> looks like:\n(I manually deleted column headings for intellectual property concerns with my construction company)<\/p>\n\n<pre><code>0                 7                 2       3       2          2               1  ...       8       2        2        2        5         1\n1                 1                 3       1       1          1               1  ...       2       1        1        1        1         1\n2                 2                 2       3       1          1               1  ...       6       1        1        5        1         2\n3                 7                 5       1       1          1               1  ...       1       1        1        1        1         1\n4                 5                 6       1       1          1               1  ...       2       1        1        1        2         1\n5                 5                 4       1       3          1               1  ...       8       4        3        2        7         3\n6                 4                 6       3       7          5               1  ...       7       2        2        6        7         2\n7                 9                10       4       1          3               1  ...       4       1        1        1        8         2\n8                 4                 2       2       1          1               1  ...       2       1        1        1        1         1\n9                 1                 2       1       5          2               3  ...       2       4        3        6        3         3\n10                7                 9       1       3          2               1  ...       7       1        1        5        8         1\n11                8                 6       1       1          1               1  ...       1       1        1        1        1         1\n12                8                 8       2       1          1               2  ...       9       3        5        2        3         1\n13                2                 3       1       1          1               1  ...       2       2        2        2        1         1\n14                2                 2       2       1          1               2  ...       2       2        2        1        3         1\n15                5                 1       2       1          1               2  ...       2       1        1        1        1         2\n16                1                 2       5       8          7               3  ...       2       4        3        7        7         5\n17                6                 4       1       3          1               3  ...       9       3        3        1        5         5\n18               10                 1       1       2          1               2  ...       1       1        1        5        1         3\n19                3                 3       2       3          2               1  ...       2       1        1        1        1         1\n20                6                 2       2       7          3               4  ...       7       5        4        3        5         5\n21                1                 2       1       3          1               2  ...       2       1        1        5        1         2\n22               10                 4       2       3          2               1  ...       1       2        2        6        3         2\n23                3                 4       1       1          1               1  ...       1       2        2        1        5         1\n24                4                 4       4       2          2               1  ...       1       1        1        1        5         1\n25                9                 8       2       2          2               1  ...       2       1        1        1        7         1\n26                1                 1       3       3          2               1  ...       2       1        1        5        1         1\n27                6                 4       3       3          2               3  ...       5       2        2        1        3         2\n28                4                 7       3       7          5               1  ...       5       3        6        2        5         5\n29                5                 1       1       2          1               1  ...       1       2        1        2        3         3\n..              ...               ...     ...     ...        ...             ...  ...     ...     ...      ...      ...      ...       ...\n285               3                 3       9       8          9              10  ...      10      10        6        8        3         5\n286               4                 6       4       7          5               7  ...       7       7        8        3        3         5\n287               5                 6       5       9          8               9  ...       4       9        9        5        5         5\n288               5                 5       9       7          9               9  ...       4       8        8        7        5         5\n289               4                 6       9       9         10              10  ...      10      10        9        8        5         5\n290              10                 9       6       5          7               8  ...       2       7        6        7        3         5\n291               4                 9       9       2          7               5  ...       7       3        8        9        8         5\n292               7                 9       8       8          9               8  ...      10       9       10       10        8         5\n293               9                10       6       9          9              10  ...       8      10       10       10        8         5\n294               5                 9       8       9         10               9  ...       6      10       10       10        8         5\n295               5                10       8       8          9               9  ...      10       5        9        9        8         5\n296               6                 9       8       8          9               9  ...       6       8       10        9        8         5\n297               1                10       8       9         10               9  ...       4      10       10        9        8         2\n298               2                10       8       7          9               9  ...       4       8       10        9        8         1\n299               8                 9       9       9         10              10  ...      10      10       10       10        8         3\n300               9                10       9       9         10               9  ...       8      10       10       10        8         3\n301               7                10       8       7          9               8  ...       8       8        9        9        8         3\n302              10                10       8      10         10              10  ...       4       9       10        9        8         3\n303               6                 9      10       9         10              10  ...      10      10       10       10        5        10\n304               9                10       9      10         10              10  ...       6      10        9       10        8         5\n305               7                 9      10       9         10              10  ...       9      10       10       10        8         5\n306               9                 9       9       8         10              10  ...      10      10       10        9        8         5\n307               7                 9      10       8         10              10  ...      10       8       10       10        8         5\n308               9                 8       6       9          9              10  ...       7       9        9        4        8         5\n309               9                 9       9      10         10              10  ...      10      10       10       10        8         5\n310               1                 9       8      10         10              10  ...       8      10       10        9        8         5\n311               9                10      10      10         10              10  ...      10      10       10       10        8        10\n312               7                10       9       9         10              10  ...       7      10       10        9        8         5\n313               2                 5       8      10         10              10  ...       4      10       10        9        8         5\n314               7                 9       9      10         10               9  ...       7      10       10        9        8         5\n<\/code><\/pre>\n\n<p>I do not understand what that error is saying, or how to fix it.<\/p>\n","answers":[{"AnswerId":"57496986","CreationDate":"2019-08-14T14:43:20.320","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>This error:<\/p>\n\n<pre><code>ValueError: Error when checking input: expected dense_1_input to have shape (6,) but got array with shape (24,)\n<\/code><\/pre>\n\n<p>Can be translated to English as: You told Keras that the input will have 6 dimensions, but the actual input has 24 dimensions. A possible fix is then to change the first line of your model definition to:<\/p>\n\n<pre><code>model.add(Dense(32, activation = 'relu', input_dim = 24))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57496919,"AnswerCount":1,"Tags":"<tensorflow><cleverhans>","CreationDate":"2019-08-14T14:39:38.523","AcceptedAnswerId":null,"OwnerUserId":6102639.0,"Title":"How to change the Cleverhans 'test_imagenet_attacks.py' with tensorflow sess?","Body":"<p>In Cleverhans example: cleverhans\/examples\/test_imagenet_attacks.py<\/p>\n\n<p>They implement SPSA attack with sess=None.<\/p>\n\n<p>But in Cleverhans attacks repo, there are a lot of methods cannot set sess as None, for example, CW, DeepFool, BFGS...<\/p>\n\n<p>How to change the code with a sess and to generate adversarial examples with these methods?<\/p>\n\n<p>Link: <a href=\"https:\/\/github.com\/tensorflow\/cleverhans\/blob\/master\/examples\/test_imagenet_attacks.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/cleverhans\/blob\/master\/examples\/test_imagenet_attacks.py<\/a><\/p>\n\n<p>A code fregment of SPSA attack on ImagNet:  <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>attack = SPSA(model)\nx_adv = attack.generate(x_input, ...)\nlogits = model.get_logits(x_adv)\nacc = _top_1_accuracy(logits, y_label)\n\nsaver = tf.train.Saver(slim.get_model_variables())\nsession_creator = tf.train.ChiefSessionCreator(...)\n\nwith tf.train.MonitoredSession(session_creator) as sess:\n     for i in xrange(num_images):\n          feed_dict_i = {x_input, y_label}\n          acc_val = sess.run(acc, feed_dict=feed_dict_i)\n<\/code><\/pre>\n\n<p>But for DeepFool, we cannot write attack = SPSA(model) as it must be attack = DeepFool(model, sess).<\/p>\n","answers":[{"AnswerId":"57595904","CreationDate":"2019-08-21T16:32:29.927","ParentId":null,"OwnerUserId":"10560104","Title":null,"Body":"<p>If you'd like to pass a session to the attack object, you can do so when you instantiate it: <\/p>\n\n<pre><code>attack = SPSA(model, sess=sess)\nattack = LBFGS(model, sess)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57497202,"AnswerCount":1,"Tags":"<python><tensorflow><lambda><keras><deep-learning>","CreationDate":"2019-08-14T14:56:02.703","AcceptedAnswerId":null,"OwnerUserId":11681121.0,"Title":"Keras, compiling: ValueError: You must specify 'axis'","Body":"<p>I'm only starting to work with DNNs and I built this one for a homework using code from <a href=\"https:\/\/r2rt.com\/binary-stochastic-neurons-in-tensorflow.html\" rel=\"nofollow noreferrer\">https:\/\/r2rt.com\/binary-stochastic-neurons-in-tensorflow.html<\/a> for the binary encoder:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def res_layer_enc(x):\n    res =x\n    #out = tf.contrib.layers.batch_norm(x, decay=0.9)\n    out = BatchNormalization()(x)\n    out = Dense(39*150, activation=\"tanh\")(out)\n\n    #out = tf.contrib.layers.batch_norm(x, decay=0.9)\n    out = BatchNormalization()(out)\n    out = Dense(39*150, activation=\"tanh\")(out)\n\n    out = Add()([res,out])\n    out = Activation(\"tanh\")(out)\n\n    return out\n\ndef res_layer_dec(x):\n    res =x\n    #out = tf.contrib.layers.batch_norm(x, decay=0.9)\n    out = BatchNormalization()(x)\n    out = Dense(13*150, activation=\"tanh\")(out)\n\n    #out = tf.contrib.layers.batch_norm(x, decay=0.9)\n    out = BatchNormalization()(out)\n    out = Dense(13*150, activation=\"tanh\")(out)\n\n    out = Add()([res,out])\n    out = Activation(\"tanh\")(out)\n\n    return out\n\n\n## ENCODER #############################\n\n## Initial dense layer\ninput_layer = Input(shape=(1,39,150), name=\"Input\")\nflat = Flatten(name=\"flatten\")(input_layer)\nenc_init_dense = Dense(39*150, activation=\"tanh\", name=\"initial_dense_ENC\")(flat)\n\n## Residual layers (2 dense layers each)\nnet = res_layer_enc(enc_init_dense)\nnet = res_layer_enc(net)\nnet = res_layer_enc(net)\n\n## Final dense layer\nnet = Dense(8, activation=\"tanh\", name=\"final_dense_ENC\")(net)\n\n#########################################\n## BINARY STOCHASTIC NEURONS ############\nbsen = Lambda(binary_wrapper, output_shape=(8,))(net)\n#########################################\n\n## DECODER ##############################\n\n## Initial dense layer\nnet = Dense(13*150, activation=\"tanh\", name=\"initial_dense_DEC\")(bsen)\n\n## Residual layers (2 dense layers each)\nnet = res_layer_dec(net)\nnet = res_layer_dec(net)\nnet = res_layer_dec(net)\n\n## Final dense layer\nout = Dense(13*150, activation=\"tanh\", name=\"final_dense_DEC\")(net)\n\n\nmodel = tf.keras.Model(inputs=input_layer, outputs=[out, bsen])\n\nmodel.compile(\"Adam\", loss=tf.losses.cosine_distance, metrics=[\"accuracy\", \"categorical_accuracy\"])\n<\/code><\/pre>\n\n<p>Everything works fine until I try to compile the model and I get the following error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-149-8db88edd7c07&gt; in &lt;module&gt;\n----&gt; 1 model.compile(\"Adam\", loss=tf.losses.cosine_distance, metrics=[\"accuracy\", \"categorical_accuracy\"])\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\checkpointable\\base.py in _method_wrapper(self, *args, **kwargs)\n    440     self._setattr_tracking = False  # pylint: disable=protected-access\n    441     try:\n--&gt; 442       method(self, *args, **kwargs)\n    443     finally:\n    444       self._setattr_tracking = previous_value  # pylint: disable=protected-access\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py in compile(self, optimizer, loss, metrics, loss_weights, sample_weight_mode, weighted_metrics, target_tensors, distribute, **kwargs)\n    447             else:\n    448               weighted_loss = training_utils.weighted_masked_objective(loss_fn)\n--&gt; 449               output_loss = weighted_loss(y_true, y_pred, sample_weight, mask)\n    450 \n    451           if len(self.outputs) &gt; 1:\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_utils.py in weighted(y_true, y_pred, weights, mask)\n    645     \"\"\"\n    646     # score_array has ndim &gt;= 2\n--&gt; 647     score_array = fn(y_true, y_pred)\n    648     if mask is not None:\n    649       mask = math_ops.cast(mask, y_pred.dtype)\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py in new_func(*args, **kwargs)\n    505                 'in a future version' if date is None else ('after %s' % date),\n    506                 instructions)\n--&gt; 507       return func(*args, **kwargs)\n    508 \n    509     doc = _add_deprecated_arg_notice_to_docstring(\n\n~\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\ops\\losses\\losses_impl.py in cosine_distance(labels, predictions, axis, weights, scope, loss_collection, reduction, dim)\n    322   axis = deprecated_argument_lookup(\"axis\", axis, \"dim\", dim)\n    323   if axis is None:\n--&gt; 324     raise ValueError(\"You must specify 'axis'.\")\n    325   if labels is None:\n    326     raise ValueError(\"labels must not be None.\")\n\nValueError: You must specify 'axis'.\n<\/code><\/pre>\n\n<p>I have checked several times (even with pen and paper :D ) the shapes and sizes of all layers and I don't understand what the \"axis\" is supposed to mean in this context.<\/p>\n","answers":[{"AnswerId":"57501374","CreationDate":"2019-08-14T20:09:08.623","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Seems <code>tf.losses.cosine_distance<\/code> cannot determine by itself to which axis to apply, you should specify the axis, for this you need to use a lambda function:<\/p>\n\n<pre><code>my_cosine = lambda y_true, y_pred: tf.losses.cosine_distance(y_true, y_pred, axis=...)\n\nmodel.compile(\"Adam\", loss=my_cosine, metrics=[\"accuracy\", \"categorical_accuracy\"])\n<\/code><\/pre>\n\n<p>You should determine to which axis to apply the cosine distance loss, according to the problem you are trying to learn. It will most likely be the last dimension.<\/p>\n"}]}
+{"QuestionId":57497224,"AnswerCount":1,"Tags":"<python><tensorflow><install>","CreationDate":"2019-08-14T14:56:49.543","AcceptedAnswerId":null,"OwnerUserId":11927928.0,"Title":"Issue with installing ethnicolr package","Body":"<p>When I try to install ethnicolr, received an error <\/p>\n\n<blockquote>\n  <p>\"Could not find a version that satisfies the requirement tensorflow==1.12.3(from ethniclor)\"<\/p>\n<\/blockquote>\n\n<p>Current version of tensorflow is 1.14, uninstalled it, and tried to install the version of 1.12.3 but get error complaining that could not find a version. <\/p>\n\n<p>is there anyway to install the package?<\/p>\n","answers":[{"AnswerId":"58312323","CreationDate":"2019-10-09T21:07:09.363","ParentId":null,"OwnerUserId":"12191497","Title":null,"Body":"<p>One solution is to install a previous version of ethnicolr <\/p>\n\n<pre><code>pip install ethnicolr==0.2.0\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/github.com\/appeler\/ethnicolr\/issues\/16\" rel=\"nofollow noreferrer\">https:\/\/github.com\/appeler\/ethnicolr\/issues\/16<\/a><\/p>\n"}]}
+{"QuestionId":57497453,"AnswerCount":0,"Tags":"<python><numpy><tensorflow><image-processing><tensorflow-serving>","CreationDate":"2019-08-14T15:08:56.423","AcceptedAnswerId":null,"OwnerUserId":6482145.0,"Title":"how to decode predictions in tensorflow_serving of predictions","Body":"<p>I am using tensorflow serving to serve YOLO_V3 keras based model <\/p>\n\n<p>Now ,everything is working but just prediction that Im getting  is a np array , how can I decode to a actual classes<\/p>\n\n<p>This is the function that I'm using<\/p>\n\n<pre><code>import argparse\nimport json\n\nimport numpy as np\nimport requests\nfrom keras.preprocessing import image\n\n# Argument parser for giving input image_path from command line\nap = argparse.ArgumentParser()\nap.add_argument(\"-i\", \"--image\", required=True,\n                help=\"path of the image\")\nargs = vars(ap.parse_args())\n\nimage_path = args['image']\n# Preprocessing our input image\nimg = image.img_to_array(image.load_img(image_path, target_size=(224, 224))) \/ 255.\n\n# this line is added because of a bug in tf_serving(1.10.0-dev)\nimg = img.astype('float16')\n\npayload = {\n    \"instances\": [{'input_image': img.tolist()}]\n}\n\n# sending post request to TensorFlow Serving server\nr = requests.post('http:\/\/localhost:9000\/v1\/models\/coco:predict', json=payload)\npred = json.loads(r.content.decode('utf-8'))\nprint(json.dumps(inception_v3.decode_predictions(np.array(pred['predictions']))[0]))\n\nprint(pred)\n# Decoding the response\n<\/code><\/pre>\n\n<p>now this gives a big list of dictionaries<\/p>\n\n<pre><code>{6876, -4.81754398, -4.97275639, -4.97607, -4.67160654, -5.79011679, -5.50944614, -4.7312355, -4.77109718, -3.92586088, -4.7312994, -5.91286898, -4.75739, -5.51783752, -3.74576402, -3.78617764, -4.80710745, -3.40718174, -6.06222868, -6.37612915, -4.38785028, -4.47397232, -4.61613226, -3.94074273, -4.83034134, -5.37178898, -6.06883144, -6.90097952, -5.8203454, -5.86746645, -5.82842541, -6.18484926, -6.35572672, -5.26999807, -6.18502855, -5.94293404, -6.49861908, -5.93721819, -6.14823675, -5.90117025, -2.50677276, -2.75121665, -2.7598753, -3.41550899, -2.62428784, -4.74949789, -5.21396255, -3.80554819, -6.01200724, -5.17863417, -5.40073681, -6.22671843, -5.16459131, -5.0060482, -6.87409353, -4.7619772, -4.86423874, -3.08233714, -5.99755383, -5.63631535, -5.94745159, -4.65718126, -6.86422825, -6.06967211, -0.217412367, 0.0550900251, -0.707384586, 0.70622, -15.0436678, 1.10565865, -4.42187643, -2.98756099, -3.72572565, -3.26745987, -2.93719411, -4.07914925, -2.34109211, -3.39609504, -4.85195971, -3.63920546, -5.17407417, -5.62931442, -4.00350428, -5.02346706, -4.95578051, -4.48103857, -4.72352171, -6.18291044, -5.11228895, -4.26431751, -5.34162378, -5.54107809, -3.85009646, -4.6390605, -3.89385056, -4.55916357, -4.70904684, -4.15174866, -6.43089485, -4.33550072, -4.58510923, -5.21521187, -3.22094584, -5.75563383, -6.88566399, -4.80700207, -4.38226604, -5.49326801, -3.26696324, -4.9029026, -5.82695913, -6.17655087, -6.67043591, -5.67217445, -6.74844885, -5.9330287, -5.71555614, -6.11479235, -4.86581087, -5.67418, -5.19628048, -6.53483, -5.78065157, -5.86239386, -5.4403944, -3.27615428, -3.07651854, -2.16416335, -3.98195553, -4.52128935, -4.46353388, -4.64035559, -4.67165565, -6.01290417, -5.59658194, -5.59060287, -6.70377827, -6.25058842, -5.20660543, -6.53567076, -5.66916752, -3.13120627, -3.95540428, -5.51402569, -5.0683589, -6.33612585, -4.35945034, -6.52084446, -5.74546289]]]}]}\n<\/code><\/pre>\n\n<p>How do I decode it into classes ? <\/p>\n\n<p>Doing this gives class mismatch error because its coco pre-trained model not imagenet<\/p>\n\n<p>what will be a similar decoding function for this ?<\/p>\n\n<pre><code>print(json.dumps(inception_v3.decode_predictions(np.array(pred['predictions']))[0]))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57498228,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-14T15:58:26.980","AcceptedAnswerId":null,"OwnerUserId":6405137.0,"Title":"Custom Layer Matrix size-incompatible","Body":"<p>I'm testing a custom layer of my own. It's an implementation of Noise Layer. However, I keep getting the error: Matrix size-incompatible\nand I can't tell what the reason is it can't do the computation.<\/p>\n\n<p>For example :<\/p>\n\n<pre><code>from  NoisyLayer import NoisyLayer\nfrom keras.models import Model\nfrom keras.layers import Input\n\ninp = Input(shape=[100])\nout = NoisyLayer(units=10)(inp)\n\nmodel = Model(inp, out)\na = np.expand_dims(a, 0)\nmodel.predict(a)\n<\/code><\/pre>\n\n<p>raises :<\/p>\n\n<pre><code>InvalidArgumentError: Matrix size-incompatible: In[0]: [10,100], In[1]: [1,100]\n [[{{node noisy_layer_32\/MatMul}}]]\n<\/code><\/pre>\n\n<p>This is the code for my custom layer:<\/p>\n\n<pre><code>from params import *\nfrom keras.layers import Layer, Multiply, Add, Dot, add\nimport keras.backend as K\nimport numpy as np\nfrom itertools import product\nfrom keras.activations import relu\nfrom copy import deepcopy\nfrom time import time\nimport tensorflow as tf\n\nclass NoisyLayer(Layer):\n    def __init__(self, units, noise_type='factorised', **kwargs):\n        self.units = units # output_dimension\n        self.noise_type = noise_type\n        super(NoisyLayer, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        self.state_shape = input_shape\n\n        self.weight_mu = self.add_weight(name='weight_mu', \n                                      shape=(self.units, input_shape[1]),\n                                      initializer=self._mu_init_fact,\n                                      trainable=True)\n        self.weight_sigma = self.add_weight(name='weight_sigma', \n                                      shape=(self.units, input_shape[1]),\n                                      initializer=self._sigma_init_fact,\n                                      trainable=True)\n        self.bias_mu = self.add_weight(name='bias_mu', \n                                      shape=(self.units),\n                                      initializer=self._mu_init_fact,\n                                      trainable=True)        \n        self.bias_sigma = self.add_weight(name='bias_mu', \n                                      shape=(self.units),\n                                      initializer=self._sigma_init_fact,\n                                      trainable=True)\n\n        self.eps_w, self.eps_b = None , None \n        self._sample_noise()\n        super(NoisyLayer, self).build(input_shape)\n\n    def call(self, input_tensor):\n        WeightedInp = K.dot(self.weight_mu + self.weight_sigma * self.eps_w, input_tensor)\n        bias = self.bias_mu + self.bias_sigma * self.eps_b\n        return relu(add([WeightedInp, bias]))\n\n    def compute_output_shape(self, input_shape):\n        return (input_shape[0], self.units)\n\n    def _sample_noise(self):\n        if self.noise_type == 'factorised':\n            eps_w_temp = np.random.rand(self.state_shape[1])\n            eps_b = np.random.rand(self.units)\n            eps_w = np.zeros((self.units, self.state_shape[1]))\n            f = lambda x : self._factor_function(x)\n\n            for i, j in product(range(self.state_shape[1]),range(self.units)):\n                eps_w[j][i] = f(eps_w_temp[i])*f(eps_w_temp[j])\n\n            eps_b = K.variable(self._factor_function(eps_b))\n            eps_w = K.variable(eps_w)\n        else:\n            raise Exception(\"No known type for noisy layer\")\n\n        self.eps_w = eps_w\n        self.eps_b = eps_b\n\n    def _factor_function(self, x):\n        return np.sign(x)*np.sqrt(np.abs(x))\n\n    def _get_noise(self):\n        return self.eps_w, self.eps_b\n\n    def _set_noise(self, eps_w, eps_b):\n        self.eps_w = tf.identity(eps_w)\n        self.eps_b = tf.identity(eps_b)\n\n    def _mu_init_fact(self, shape, dtype=None):\n        mu = np.random.uniform(low =-np.sqrt(1\/self.units), high=np.sqrt(1\/self.units), size=shape)\n        return K.variable(mu)\n\n    def _sigma_init_fact(self, shape, dtype=None):\n        sigma = np.ones(shape=shape)*SIGMA_0 \/ self.units  #SIGMA_0 is 0.15\n        return K.variable(sigma)    \n<\/code><\/pre>\n\n<p>The intended result is :\nwith weight_mu, weight_sigma, bias_mu, and bias_sigma being the trainable parameters of the layer and eps_w, eps_b being some random non-trainable parameters, the computation to be:\n(with @ meaning element-wise multiplication)<\/p>\n\n<pre><code>y = (weight_mu + weight_sigma @ eps_w) * X + (bias_mu + bias_sigma @ eps_b)\n<\/code><\/pre>\n\n<p>output = relu(y)<\/p>\n","answers":[]}
+{"QuestionId":57498530,"AnswerCount":0,"Tags":"<tensorflow><dataset>","CreationDate":"2019-08-14T16:19:08.800","AcceptedAnswerId":null,"OwnerUserId":3943706.0,"Title":"How make new image dataset for segmentation net","Body":"<p><strong>How make good mask file (like in original dataset oxford_iiit_pet:3.0.0)  for this segmentation net? I want use photoshop and can make convert by numpy or matlab<\/strong><\/p>\n\n<p>I am learning tensorflow segmentation - example <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/images\/segmentation\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/tutorials\/images\/segmentation<\/a>. I run thins example and it work. I want make my special dataset for it UNET. Dataset must be like oxford_iiit_pet:3.0.0. And it is problem: in original dataset mask  make png file annotations\/trimaps\/...png. I open this files in phoshop  - all black.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Vxls0.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Vxls0.png\" alt=\"enter image description here\"><\/a>I make mask files in phoshop - backgroud black (255,255,255), human - white (0, 0, 0).\n<a href=\"https:\/\/i.stack.imgur.com\/VH2Nz.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/VH2Nz.png\" alt=\"enter image description here\"><\/a>\nBut with my file UNET no work, have error. Problem is mask file format. <\/p>\n\n<p>My code same example <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/images\/segmentation\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/tutorials\/images\/segmentation<\/a>, only other dataset. Model is:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>base_model = tf.keras.applications.MobileNetV2(input_shape=[128, 128, 3], include_top=False)\n<\/code><\/pre>\n\n<p>have error:\n<code>Invalid argument: Received a label value of -1 which is outside the valid range of [0, 3).  Label values: -1 -1 -1 -1 -1 -1 -1 -1 ...<\/code><\/p>\n","answers":[]}
+{"QuestionId":57498998,"AnswerCount":2,"Tags":"<python><tensorflow><keras><neural-network><theano>","CreationDate":"2019-08-14T16:54:40.997","AcceptedAnswerId":57529333.0,"OwnerUserId":7253901.0,"Title":"Strange behavior of keras v1.2.2 vs. keras v2+ (HUGE differences in accuracy)","Body":"<p>Today I've ran into some very strange behavior of Keras. <em>When I try to do a classification run on the iris-dataset with a simple model, keras version 1.2.2 gives me +- 95% accuracy, whereas a keras version of 2.0+ predicts the same class for every training example (leading to an accuracy of +- 35%, as there are three types of iris)<\/em>.  The only thing that makes my model predict +-95% accuracy is downgrading keras to a version below 2.0:<\/p>\n\n<p>I think it is a problem with Keras, as I have tried the following things, all do not make a difference;<\/p>\n\n<ul>\n<li>Switching activation function in the last layer (from Sigmoid to softmax).<\/li>\n<li>Switching backend (Theano and Tensorflow both give roughly same performance).<\/li>\n<li>Using a random seed.<\/li>\n<li>Varying the number of neurons in the hidden layer (I only have 1 hidden layer in this simple model).<\/li>\n<li>Switching loss-functions.<\/li>\n<\/ul>\n\n<p>As the model is very simple and it runs on it's own (You just need the easy-to-obtain iris.csv dataset) I decided to include the entire code; <\/p>\n\n<pre><code>import pandas as pd\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.wrappers.scikit_learn import KerasClassifier\nfrom keras.utils import np_utils\nfrom sklearn.model_selection import cross_val_score\nfrom sklearn.model_selection import KFold\nfrom sklearn.preprocessing import LabelEncoder\n\n#Load data\ndata_frame = pd.read_csv(\"iris.csv\", header=None)\ndata_set = data_frame.values\nX = data_set[:, 0:4].astype(float)\nY = data_set[:, 4]\n\n#Encode class values as integers\nencoder = LabelEncoder()\nencoder.fit(Y)\nencoded_Y = encoder.transform(Y)\n\n# convert integers to dummy variables (i.e. one hot encoded)\ndummy_y = np_utils.to_categorical(encoded_Y)\n\ndef baseline_model():\n    #Create &amp; Compile model\n    model = Sequential()\n    model.add(Dense(8, input_dim=4, init='normal', activation='relu'))\n    model.add(Dense(3, init='normal', activation='sigmoid'))\n    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n    return model\n\n#Create Wrapper For Neural Network Model For Use in scikit-learn\nestimator = KerasClassifier(build_fn=baseline_model, nb_epoch=200, batch_size=5, verbose=0)\n\n#Create kfolds-cross validation\nkfold = KFold(n_splits=10, shuffle=True)\n\n#Evaluate our model (Estimator) on dataset (X and dummy_y) using a 10-fold cross-validation procedure (kfold).\nresults = cross_val_score(estimator, X, dummy_y, cv=kfold)\nprint(\"Accuracy: {:2f}% ({:2f}%)\".format(results.mean()*100, results.std()*100))\n<\/code><\/pre>\n\n<p>if anyone wants to replicate the error here are the dependencies I used to observe the problem:<\/p>\n\n<pre><code>numpy=1.16.4\npandas=0.25.0\nsk-learn=0.21.2\ntheano=1.0.4\ntensorflow=1.14.0\n<\/code><\/pre>\n","answers":[{"AnswerId":"57529333","CreationDate":"2019-08-16T18:16:07.713","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>In Keras 2.0, many parameters changed names, there is compatibility layer to keep things working, but somehow it did not apply when using <code>KerasClassifier<\/code>.<\/p>\n\n<p>In this part of the code:<\/p>\n\n<pre><code>estimator = KerasClassifier(build_fn=baseline_model, nb_epoch=200, batch_size=5, verbose=0)\n<\/code><\/pre>\n\n<p>You are using the old name <code>nb_epoch<\/code> instead of the modern name of <code>epochs<\/code>. The default value is <code>epochs=1<\/code>, meaning that your model was only being trained for one epoch, producing very low quality predictions.<\/p>\n\n<p>Also note that here:<\/p>\n\n<pre><code>model.add(Dense(3, init='normal', activation='sigmoid'))\n<\/code><\/pre>\n\n<p>You should be using a <code>softmax<\/code> activation instead of <code>sigmoid<\/code>, as you are using the categorical cross-entropy loss:<\/p>\n\n<pre><code>model.add(Dense(3, init='normal', activation='softmax'))\n<\/code><\/pre>\n"},{"AnswerId":"57522496","CreationDate":"2019-08-16T09:50:38.237","ParentId":null,"OwnerUserId":"7253901","Title":null,"Body":"<p>I've managed to isolate the issue, if you change <code>nb_epoch<\/code> to <code>epochs<\/code>, (All else being exactly equal) the model predicts very good again, in keras 2 as well. I don't know if this is intended behavior or a bug.<\/p>\n"}]}
+{"QuestionId":57499002,"AnswerCount":3,"Tags":"<python><python-3.x><pytorch><python-3.7>","CreationDate":"2019-08-14T16:54:57.550","AcceptedAnswerId":null,"OwnerUserId":7218182.0,"Title":"Can't install pytorch with pip on Windows","Body":"<p>I'm trying to install Pytorch with Windows and I'm using the commands of the official site\n<a href=\"https:\/\/pytorch.org\/get-started\/locally\/\" rel=\"noreferrer\">https:\/\/pytorch.org\/get-started\/locally\/<\/a><\/p>\n\n<pre><code>pip3 install torch==1.2.0 torchvision==0.4.0 -f https:\/\/download.pytorch.org\/whl\/torch_stable.html\n<\/code><\/pre>\n\n<p>This is the command if I choose Windows, Cuda 10.0, and Python 3.7\nBut if I run this I get the error message:<\/p>\n\n<pre><code>ERROR: Could not find a version that satisfies the requirement torch==1.2.0 (from versions: 0.1.2, 0.1.2.post1, 0.1.2.post2)\nERROR: No matching distribution found for torch==1.2.0\n<\/code><\/pre>\n\n<p>So why does this happen?\nMy pip is version 19.2 and I am in a newly installed python 3.7 environment<\/p>\n","answers":[{"AnswerId":"57565285","CreationDate":"2019-08-20T00:17:23.440","ParentId":null,"OwnerUserId":"2627529","Title":null,"Body":"<p>The most likely reason for Your issue is a 32-bit installation of python, while the torch libraries rely on having a 64-bit version. I had exactly the same issue.<\/p>\n\n<p>Just start <code>python<\/code> from command line and observe<\/p>\n\n<pre><code>C:\\Users\\marci&gt;python\nPython 3.7.4 (tags\/v3.7.4:e09359112e, Jul  8 2019, 20:34:20) [MSC v.1916 64 bit (AMD64)] on win32\n<\/code><\/pre>\n\n<p>My installation now shows 64 bits. If Yours shows 32, then install 64-bit python. I used this link: <a href=\"https:\/\/www.python.org\/ftp\/python\/3.7.4\/python-3.7.4-amd64.exe\" rel=\"nofollow noreferrer\">Official python 64-bit Windows installer<\/a><\/p>\n"},{"AnswerId":"57499069","CreationDate":"2019-08-14T17:01:08.233","ParentId":null,"OwnerUserId":"11162393","Title":null,"Body":"<p>Try installing via .whl file from Christoph Gohlke's repo at this link: <a href=\"https:\/\/www.lfd.uci.edu\/~gohlke\/pythonlibs\/\" rel=\"nofollow noreferrer\">https:\/\/www.lfd.uci.edu\/~gohlke\/pythonlibs\/<\/a><\/p>\n\n<p>Make sure you get the right one for your python version (cp37 at the bottom).<\/p>\n\n<p>Navigate to the file or save it to your working directory, then use<\/p>\n\n<p>pip3 install <em>path-to-file.whl<\/em><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/rpxZR.png\" rel=\"nofollow noreferrer\">Link to .whl file on page<\/a><\/p>\n"},{"AnswerId":"57499295","CreationDate":"2019-08-14T17:18:07.367","ParentId":null,"OwnerUserId":"11780961","Title":null,"Body":"<p>So you have Cuda 10 installed? If you do, try this:<\/p>\n\n<pre><code>pip3 install https:\/\/download.pytorch.org\/whl\/cu100\/torch-1.2.0-cp37-cp37m-win_amd64.whl\n<\/code><\/pre>\n\n<p>followed by:<\/p>\n\n<pre><code>pip3 install torchvision\n<\/code><\/pre>\n\n<p>To check if it was installed properly, type this into your command line:<\/p>\n\n<pre><code>python\n<\/code><\/pre>\n\n<p>followed by:<\/p>\n\n<pre><code>from __future__ import print_function\nimport torch\nx = torch.rand(5, 3)\nprint(x)\n<\/code><\/pre>\n\n<p>If you get this output:<\/p>\n\n<pre><code>tensor([[0.3380, 0.3845, 0.3217],\n        [0.8337, 0.9050, 0.2650],\n        [0.2979, 0.7141, 0.9069],\n        [0.1449, 0.1132, 0.1375],\n        [0.4675, 0.3947, 0.1426]])\n<\/code><\/pre>\n\n<p>PyTorch was installed correctly!<\/p>\n"}]}
+{"QuestionId":57499643,"AnswerCount":2,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-08-14T17:44:38.013","AcceptedAnswerId":57499837.0,"OwnerUserId":11928586.0,"Title":"How to correctly use loss function in keras?","Body":"<p>I am testing a loss function to reduce both MSE and Cross_Entropy at the same time. I defined a loss function, but it seems that the loss is not going down. I am new to the deep learning area, any suggestions are welcome. Thanks<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nfrom sklearn.preprocessing import MinMaxScaler\n\nx1=np.random.uniform(0,100, 1000)\nx2=np.random.uniform(-20,20, 1000)\ny1= [int(i) for i in x1 % 3]\ny2= x1*1.5 + np.random.randn(1000)\ny3= 2*x1*x1 + np.random.randn(1000)\ny4= x1*2.5 -10 +np.random.randn(1000)\ny5=x1*2.5 +10 +np.random.randn(1000)\n\nxs=np.array([x1,x2]).T\ny1s= np.array(y1).reshape(-1,1)\ny2s= np.array([y2,y3, y4, y5]).T\ny1s=to_categorical(y1s)\nscaler= MinMaxScaler()\nxs_normed=scaler.fit_transform(xs)\n\nfrom keras.layers import Dense, Input, Concatenate, Lambda\nfrom keras.models import Model\nfrom keras.losses import mse, categorical_crossentropy\nimport keras.backend as K\n\ndef loss_fun(real,pred, alpha=1, beta=1):\n\n    c1_pred=pred[0]\n    c1_real=real[0]\n    c2_pred=pred[1]\n    c2_real=real[1]\n    loss1=mse(c2_real, c2_pred)\n    loss2=categorical_crossentropy(c1_real, c1_pred)\n    loss=K.sum(alpha*loss1 + beta*loss2)\n    return loss\n\ninputs= Input(shape=(2,))\nd1=Dense(256, activation='relu')(inputs)\nd2=Dense(256, activation='relu')(d1)\nd3=Dense(3, activation='softmax')(d2)\nd5=Dense(256)(d2)\nd6=Dense(4)(d5)\nmodel= Model(inputs=inputs, outputs= [d3, d6])\nmodel.compile(optimizer='Adam', loss=loss_fun)\n\nmodel.fit(x=xs_normed, y= [y1s, y2s], batch_size=128, epochs=100)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57499837","CreationDate":"2019-08-14T17:59:14.047","ParentId":null,"OwnerUserId":"1626564","Title":null,"Body":"<p>First there is nothing wrong to use a loss function like you defined. <\/p>\n\n<p>But I hope you use this loss function because you understand what you are going to optimize. Your loss function will jointly optimize your model on the euclidean and angular metric spaces.<\/p>\n\n<p>The reasons why your model does not learn might due to<\/p>\n\n<ol>\n<li>your alpha and beta values might need adjustment, it will be better for you try different combinations, e.g. any paired values from the list [1e-3,1e-2,1e-1,1].<\/li>\n<li>your problem maybe inappropriate to this loss, especially when your target is not an one-hot encoded vector.<\/li>\n<\/ol>\n"},{"AnswerId":"57500033","CreationDate":"2019-08-14T18:15:50.087","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Better to use:<\/p>\n\n<pre><code>model.compile(loss=['categorical_crossentropy', 'mse'], \n              loss_weights = [beta, alpha], ...)\n<\/code><\/pre>\n\n<p>If this doesn't work, the problem is not your loss. <\/p>\n"}]}
+{"QuestionId":57499831,"AnswerCount":0,"Tags":"<python><tensorflow><activation><relu>","CreationDate":"2019-08-14T17:58:27.943","AcceptedAnswerId":null,"OwnerUserId":6310993.0,"Title":"Tensorflow Activation function with Dense Layer","Body":"<p>I would like to know if it's possible to define a custom activation function like <code>tanh<\/code> to constraints the 12 outputs of the dense layers or fully connected layer like that: <\/p>\n\n<pre><code>If X[0:6] &gt; 1 then X[0:6] else 1\nIf X[6:12]  &lt; 0 then X[6:12] else 0\n<\/code><\/pre>\n\n<p><code>X<\/code> is my <code>denseLayer.outputs<\/code><\/p>\n\n<p>I found some tutorials to create custom activation function with gradient like <code>LeakyRelu6<\/code> on Medium for example, but it's always with a constraints on all the outputs.<\/p>\n\n<p>In the tutorials, what they are doing is : <\/p>\n\n<pre><code>If X &gt;= 0 and X &lt;= 6 then X\nElif X &gt;6 then 6\nElse 0.2*X\n<\/code><\/pre>\n\n<p>To sum up, i would like to understand how can i have an activation function which constraints X like in my exemple ? <\/p>\n\n<p>The link: <a href=\"https:\/\/medium.com\/@chinesh4\/custom-activation-function-in-tensorflow-for-deep-neural-networks-from-scratch-tutorial-b12e00652e24\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@chinesh4\/custom-activation-function-in-tensorflow-for-deep-neural-networks-from-scratch-tutorial-b12e00652e24<\/a><\/p>\n","answers":[]}
+{"QuestionId":57499867,"AnswerCount":1,"Tags":"<python><scikit-learn><deep-learning><pytorch><roc>","CreationDate":"2019-08-14T18:02:18.510","AcceptedAnswerId":57501279.0,"OwnerUserId":11480928.0,"Title":"Getting error while trying to plot ROC for multiclass classification","Body":"<p>I am trying to plot ROC Curve for multiclass classification.I followed <a href=\"https:\/\/scikit-learn.org\/stable\/auto_examples\/model_selection\/plot_roc.html\" rel=\"nofollow noreferrer\">https:\/\/scikit-learn.org\/stable\/auto_examples\/model_selection\/plot_roc.html<\/a>.\nI used the below code to calculate y_test,and y_score <\/p>\n\n<pre><code>def test_epoch(net,test_loader):\n y_test =[]\n y_score =[]\n with torch.no_grad():\n  for batch in test_loader:\n    images, labels = batch['image'], batch['grade']\n    images =Variable(images)\n    labels= Variable(labels)\n    target =F.one_hot(labels,5)\n    outputs = net(images)\n    _, predicted = torch.max(outputs.data, 1)\n    c = (predicted == labels).squeeze().numpy()\n    y_score.append(outputs.numpy())\n    y_test.append(labels.numpy())\n    return y_test,y_score \n<\/code><\/pre>\n\n<p>I seen that my y_test is array like below <\/p>\n\n<pre><code>y_test data&gt;&gt; [array([[0, 0, 1, 0, 0],\n       [1, 0, 0, 0, 0],\n       [0, 1, 0, 0, 0],\n       [1, 0, 0, 0, 0],\n       [1, 0, 0, 0, 0],\n       [0, 0, 0, 1, 0],\n       [0, 0, 1, 0, 0],\n       [1, 0, 0, 0, 0],\n       [0, 0, 1, 0, 0],\n       [0, 0, 1, 0, 0],\n       [0, 1, 0, 0, 0]\n<\/code><\/pre>\n\n<p>And y_score is like <\/p>\n\n<pre><code>[array([[ 0.30480504, -0.12213976,  0.09632117, -0.16465648, -0.44081157],[ 0.21797988, -0.09650452,  0.07616544, -0.12001953, -0.34972644],[ 0.3230184 , -0.13098559,  0.10277118, -0.17656785, -0.45888817],[ 0.38143447, -0.15880316,  0.12123139, -0.21719441, -0.5281661 ],[ 0.3427343 , -0.13945231,  0.11076729, -0.19657779, -0.4913683 ]\n<\/code><\/pre>\n\n<p>Whenever I called the function for plotting ROC curve <\/p>\n\n<pre><code>def plot_roc(y_test, y_score, N_classes):\n    \"\"\"\n    compute ROC curve and ROC area for each class in each fold\n\n    \"\"\"\n\n    fpr = dict()\n    tpr = dict()\n    roc_auc = dict()\n    for i in range(N_classes):\n        fpr[i], tpr[i], _ = roc_curve(np.array(y_test[:, i]),np.array(y_score[:, i]))\n        roc_auc[i] = auc(fpr[i], tpr[i])\n    # Compute micro-average ROC curve and ROC area\n    fpr[\"micro\"], tpr[\"micro\"], _ = roc_curve(y_test.ravel(), y_score.ravel())\n    roc_auc[\"micro\"] = auc(fpr[\"micro\"], tpr[\"micro\"])\n\n    # Compute macro-average ROC curve and ROC area\n\n    # First aggregate all false positive rates\n    all_fpr = np.unique(np.concatenate([fpr[i] for i in range(N_classes)]))\n\n    # Then interpolate all ROC curves at this points\n    mean_tpr = np.zeros_like(all_fpr)\n    for i in range(N_classes):\n        mean_tpr += interp(all_fpr, fpr[i], tpr[i])\n\n    # Finally average it and compute AUC\n    mean_tpr \/= N_classes\n\n    fpr[\"macro\"] = all_fpr\n    tpr[\"macro\"] = mean_tpr\n    roc_auc[\"macro\"] = auc(fpr[\"macro\"], tpr[\"macro\"])\n\n    # Plot all ROC curves\n    plt.figure()\n    plt.plot(fpr[\"micro\"], tpr[\"micro\"],\n             label='micro-average ROC curve (area = {0:0.2f})'\n                   ''.format(roc_auc[\"micro\"]),\n             color='deeppink', linestyle=':', linewidth=4)\n\n    plt.plot(fpr[\"macro\"], tpr[\"macro\"],\n             label='macro-average ROC curve (area = {0:0.2f})'\n                   ''.format(roc_auc[\"macro\"]),\n             color='navy', linestyle=':', linewidth=4)\n\n    colors = cycle(['aqua', 'darkorange', 'cornflowerblue'])\n    for i, color in zip(range(N_classes), colors):\n        plt.plot(fpr[i], tpr[i], color=color, lw=2,\n                 label='ROC curve of class {0} (area = {1:0.2f})'\n                       ''.format(i, roc_auc[i]))\n\n    plt.plot([0, 1], [0, 1], 'k--', lw=2)\n    plt.xlim([0.0, 1.0])\n    plt.ylim([0.0, 1.05])\n    plt.xlabel('False Positive Rate')\n    plt.ylabel('True Positive Rate')\n    plt.title('Some extension of Receiver operating characteristic to multi-class')\n    plt.legend(loc=\"lower right\")\n    plt.show()\n<\/code><\/pre>\n\n<p>I am getting this error massages, <\/p>\n\n<pre><code> Traceback (most recent call last):\n  File \"\/home\/Downloads\/demo 3.py\", line 405, in &lt;module&gt;\n    plot_roc(y_test, y_score, 5)\n  File \"\/home\/Downloads\/demo 3.py\", line 225, in plot_roc\n    fpr[i], tpr[i], _ = roc_curve(np.array(y_test[:, i]),np.array(y_score[:, i]))\nTypeError: list indices must be integers or slices, not tuple\n<\/code><\/pre>\n\n<p>I could not understand how I will solve this problem.\nI am highly appreciate any kind of help regarding this issue.<\/p>\n","answers":[{"AnswerId":"57501279","CreationDate":"2019-08-14T20:01:20.423","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>In your code you have a previously defined variable (a list) called <code>roc_curve<\/code>, and this shadows the scikit-learn function <code>sklearn.metrics.roc_curve<\/code> in your code, you should prefer not to name variables the same as a well known function, in order to prevent problems like these.<\/p>\n"}]}
+{"QuestionId":57500042,"AnswerCount":0,"Tags":"<python><tensorflow><visualization><tensorboard>","CreationDate":"2019-08-14T18:16:52.773","AcceptedAnswerId":null,"OwnerUserId":10991568.0,"Title":"Not able to visualize classified images in tensorboard projector","Body":"<p>I have created a model from Kaggle images dataset - fashion MNIST. <\/p>\n\n<p><strong>Problem:<\/strong><\/p>\n\n<p><code>I am not able to view any images in the tensorboard Projector.<\/code> <\/p>\n\n<p>Only labels are visible in the Projector.<\/p>\n\n<p>Kindly Refer to the pics below:<\/p>\n\n<ul>\n<li>Label mode:<\/li>\n<\/ul>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/0zgIO.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0zgIO.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<ul>\n<li>Images mode:<\/li>\n<\/ul>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/FtR2Y.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/FtR2Y.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Code is taken from below mentioned links: <\/p>\n\n<p><a href=\"https:\/\/www.youtube.com\/watch?v=-vGqUWet-2w\" rel=\"nofollow noreferrer\">https:\/\/www.youtube.com\/watch?v=-vGqUWet-2w<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/markjay4k\/Fashion-MNIST-with-Keras\" rel=\"nofollow noreferrer\">https:\/\/github.com\/markjay4k\/Fashion-MNIST-with-Keras<\/a><\/p>\n\n<p>How can I make images appear as they have shown in their youtube video (Pic taken from original video for reference):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/LDfGg.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/LDfGg.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><strong>Code Below:<\/strong> <\/p>\n\n<pre><code>import os\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\n\nimport tensorflow as tf\nfrom tensorflow.contrib.tensorboard.plugins import projector\nfrom tensorflow.examples.tutorials.mnist import input_data\n\n\nPath = r'C:\\Users\\viny\\2.AnalytixLabs Practice\\Tensorboard\\Fashion MNIST\\fashionmnist Data\\\\'\ntest_data = np.array(pd.read_csv(Path+'fashion-mnist_test.csv'), dtype='float32')\n\nembed_count = 1600\n\nx_test = test_data[: embed_count, 1:] \/ 255  \ny_test = test_data[: embed_count, 0]         \n\n# assigning log directory\nlogdir = r'C:\\Users\\viny\\tensorboard\\FashionMNIST-Part3\\logdir'\n\nsummary_writer = tf.summary.FileWriter(logdir)\n\nembedding_var = tf.Variable(x_test, name='fmnist_embedding')\n\nconfig = projector.ProjectorConfig()\nembedding = config.embeddings.add()\nembedding.tensor_name = embedding_var.name\n\nembedding.metadata_path = os.path.join(logdir, 'metadata.tsv')\nembedding.sprite.image_path = os.path.join(logdir, 'sprite.png')\n\nembedding.sprite.single_image_dim.extend([28,28])\n\nprojector.visualize_embeddings(summary_writer, config)\n\nwith tf.Session() as sesh:\n    sesh.run(tf.global_variables_initializer())\n    saver = tf.train.Saver()\n    saver.save(sesh, os.path.join(logdir, 'model.ckpt'))\n\nrows = 28\ncols = 28\n\nlabel = ['t_shirt', 'trouser', 'pullover', 'dress', 'coat', 'sandal', 'shirt', 'sneaker', 'bag', 'ankle_boot']\n\n\nsprite_dim = int(np.sqrt(x_test.shape[0]))\n\nsprite_image = np.ones((cols * sprite_dim, rows * sprite_dim))\n\nindex=0\nlabels = []\nfor i in range(sprite_dim):\n    for j in range(sprite_dim):\n\n        labels.append(label[int(y_test[index])])\n\n        sprite_image[\n            i * cols: (i+1) * cols,\n            j * cols: (j+1) * cols\n        ] = x_test[index].reshape(28,28) * -1  + 1  \n\n        index +=1\n\n\nwith open(embedding.metadata_path, 'w') as meta:\n    meta.write('Index\\tLabel\\n')\n\n    for index, label in enumerate(labels):\n        meta.write(f'{index}\\t{label}\\n')\n\nplt.imsave(embedding.sprite.image_path, sprite_image, cmap='gray')\nplt.imshow(sprite_image, cmap='gray')\nplt.show()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57500159,"AnswerCount":1,"Tags":"<machine-learning><nlp><pytorch><fast-ai>","CreationDate":"2019-08-14T18:25:53.130","AcceptedAnswerId":57560024.0,"OwnerUserId":10094585.0,"Title":"How to get sentence embeddings from encoder in Fastai learner language model","Body":"<p>I was able to fine tune a language model using fast ai.  I would like to extract sentence embeddings from the fine-tuned model for sentence similarity.  How do I get the encoder model embeddings? Also can embeddings be compared with dot product like other embeddings from other models such as USE?<\/p>\n\n<pre><code>data_lm = TextLMDataBunch.from_df(train_df = se1, valid_df = se2, path = \"\",text_cols='text')\nlearn = language_model_learner(data_lm,drop_mult=0.7,pretrained=True,arch=AWD_LSTM)\nlearn.fit_one_cycle(3, 1e-01)\n<\/code><\/pre>\n\n<p>My code is above how can I get encodings from learn?<\/p>\n","answers":[{"AnswerId":"57560024","CreationDate":"2019-08-19T15:44:34.720","ParentId":null,"OwnerUserId":"10511793","Title":null,"Body":"<p>This should give you the encoder(Which is an embedding layer) :<\/p>\n\n<p><code>learn.model[0].encoder<\/code><\/p>\n"}]}
+{"QuestionId":57500452,"AnswerCount":0,"Tags":"<python><machine-learning><keras>","CreationDate":"2019-08-14T18:50:32.997","AcceptedAnswerId":null,"OwnerUserId":3707926.0,"Title":"How to reset epoch counter in Keras Callback","Body":"<p>I wrote Keras callback that reinitializes model weights after bad start. After weights reset how can I reset epoch counter also, so my callback will run training from epoch 0 ?<\/p>\n\n<p>This is my callback code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nfrom keras import backend as K\nfrom keras.callbacks import Callback\n\nclass ReinitWeightsOnFalseStart(Callback):\n    \"\"\"Reinitializes model weights when model is not improving right after beginning.\n    Stops training after number of trials.\n    Monitored value is 'val_loss'.\n    # Arguments\n        patience: number of epochs with no improvement\n            after which weights will be reinitialized.\n        trials: number of attempts to reinitialize weights\n            after which training will be stopped.\n        verbose: verbosity mode.\n    \"\"\"\n    def __init__(self,\n                 patience=3,\n                 trials=2,\n                 verbose=0):\n        self.patience = patience\n        self.trials = trials\n        self.verbose = verbose\n\n    def reset_weights(self):\n        session = K.get_session()\n        for layer in self.model.layers:\n            if hasattr(layer, 'kernel_initializer'):\n                layer.kernel.initializer.run(session=session)\n\n    def on_train_begin(self, logs={}):\n        self._monitor = True\n        self._restarted = True\n        self._best = np.Inf\n        self._wait = 0\n        self._tries = 0\n\n    def on_epoch_end(self, epoch, logs={}):\n        if self._monitor:\n            current = round(logs.get('val_loss'), 5)\n\n            if current &gt;= self._best:\n                self._wait += 1\n\n                if self._wait == self.patience:\n                    self._tries += 1\n\n                    if self._tries == self.trials:\n                        if self.verbose &gt; 0:\n                            print('\\nReinitWeightsOnFalseStart: Model is not improving from the start. Stopping.')\n                        self.model.stop_training = True\n\n                    if self.verbose &gt; 0:\n                        print('\\nReinitWeightsOnFalseStart: Reinitializing model weights.')\n                    self.reset_weights()\n                    self._restarted = True\n                    self._wait = 0\n\n            else:\n                if self._restarted:\n                    self._best = current\n                    self._restarted = False\n                else:\n                    if self.verbose &gt; 0:\n                        print('\\nReinitWeightsOnFalseStart: Looks like model is improving. Stop monitoring.')\n                    self._monitor = False\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/gist.github.com\/pippo111\/911167d9574250b3e799b4b71bf6cbe8\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/pippo111\/911167d9574250b3e799b4b71bf6cbe8<\/a><\/p>\n\n<p>Right now not resetting epoch counter will interfere with EarlyStopping callback for example.<\/p>\n","answers":[]}
+{"QuestionId":57500582,"AnswerCount":1,"Tags":"<python><machine-learning><pytorch><lstm><recurrent-neural-network>","CreationDate":"2019-08-14T19:01:36.630","AcceptedAnswerId":57509265.0,"OwnerUserId":1957627.0,"Title":"Unable to figure out inplace operation in the pytorch code?","Body":"<p>I have the following implementation in PyTorch for learning using LSTM:<\/p>\n\n<p><a href=\"https:\/\/gist.github.com\/rahulbhadani\/f1d64042cc5a80280755cac262aa48aa\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/rahulbhadani\/f1d64042cc5a80280755cac262aa48aa<\/a><\/p>\n\n<p>However, the code is experiencing in-place operation error <\/p>\n\n<p>where my error output is:<\/p>\n\n<pre><code>\/home\/ivory\/anaconda3\/lib\/python3.7\/site-packages\/ipykernel_launcher.py:10: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).\n  # Remove the CWD from sys.path while we load stuff.\n---------------------------------------------------------------------------\nRuntimeError                              Traceback (most recent call last)\n&lt;ipython-input-86-560ec78f2b64&gt; in &lt;module&gt;\n     27 linear = torch.nn.Linear(hidden_nums, output_dim)\n     28 \n---&gt; 29 global_loss_list = global_training(lstm2)\n\n&lt;ipython-input-84-152890a3028c&gt; in global_training(optimizee)\n      3     adam_global_optimizer = torch.optim.Adam([{'params': optimizee.parameters()}, \n      4                                      {'params':linear.parameters()}], lr = 0.0001)\n----&gt; 5     _, global_loss_1 = learn2(LSTM_Optimizee, training_steps, retain_graph_flag=True, reset_theta=True)\n      6 \n      7     print(global_loss_1)\n\n&lt;ipython-input-83-0357a528b94d&gt; in learn2(optimizee, unroll_train_steps, retain_graph_flag, reset_theta)\n     43             # requires_grad=True. These are accumulated into x.grad for every\n     44             # parameter x. In pseudo-code: x.grad += dloss\/dx\n---&gt; 45             loss.backward(retain_graph = retain_graph_flag) #The default is False, when the optimized LSTM is set to True\n     46 \n     47             print('x.grad: {}'.format(x.grad))\n\n~\/anaconda3\/lib\/python3.7\/site-packages\/torch\/tensor.py in backward(self, gradient, retain_graph, create_graph)\n    116                 products. Defaults to ``False``.\n    117         \"\"\"\n--&gt; 118         torch.autograd.backward(self, gradient, retain_graph, create_graph)\n    119 \n    120     def register_hook(self, hook):\n\n~\/anaconda3\/lib\/python3.7\/site-packages\/torch\/autograd\/__init__.py in backward(tensors, grad_tensors, retain_graph, create_graph, grad_variables)\n     91     Variable._execution_engine.run_backward(\n     92         tensors, grad_tensors, retain_graph, create_graph,\n---&gt; 93         allow_unreachable=True)  # allow_unreachable flag\n     94 \n     95 \n\nRuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.FloatTensor [1, 10]] is at version 1; expected version 0 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).\n<\/code><\/pre>\n\n<p>I have tried to trace the error but didn't succeed. Any help in this regard will be appreciated.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"57509265","CreationDate":"2019-08-15T11:55:26.823","ParentId":null,"OwnerUserId":"7347631","Title":null,"Body":"<p>I think the issue is with the following line: <\/p>\n\n<pre><code>global_loss_list.append(global_loss.detach_())\n<\/code><\/pre>\n\n<p>The convention in PyTorch for in-place operations is using <code>_<\/code> at the end of the function name (as in <code>detach_<\/code>). I believe you shouldn't be detaching in-place. In other words, change <code>detach_<\/code> to <code>detach<\/code> <\/p>\n"}]}
+{"QuestionId":57501255,"AnswerCount":0,"Tags":"<json><tensorflow><keras><hdf5><h5py>","CreationDate":"2019-08-14T19:59:30.810","AcceptedAnswerId":null,"OwnerUserId":9522915.0,"Title":"TypeError: Not JSON Serializable: <module 'tensorflow' from '\/path\/python\/v3-5.1.0\/lib\/python3.6\/site-packages\/tensorflow\/__init__.py'","Body":"<p>I am trying to save the best model through the keras checkpoint callback. However, save model returns a \"Not Jason serialisable\" error similar, but not exactly the same, as some of the other issues previously raised, which are supposedly resolved by a tf-nightly install. I cannot provide the one-hot-encoded input data.<\/p>\n\n<p>I have installed the tensorflow-nightly, but the problem still remains. I have also directly used the save.model keras function, which returns the same error, but also a file which is entirely filled with columns of zeros.<\/p>\n\n<pre><code>from keras import models\nfrom keras import layers\nfrom keras import optimizers\nfrom keras import initializers\nfrom keras.layers import Flatten\nfrom keras.layers import Dropout\nfrom keras.layers import Dense\nfrom keras.preprocessing import sequence\nfrom keras.optimizers import RMSprop\nfrom keras.models import Sequential\nfrom keras.layers.core import Dense, Dropout, Activation, Flatten\nfrom keras.layers.convolutional import Convolution1D, MaxPooling1D\nfrom keras.constraints import maxnorm\nfrom keras.layers.recurrent import LSTM, GRU\nfrom keras.callbacks import ModelCheckpoint, EarlyStopping, TensorBoard\nfrom sklearn.preprocessing import MinMaxScaler\nfrom keras.utils.training_utils import multi_gpu_model\nfrom keras.layers import LeakyReLU, BatchNormalization, GlobalMaxPooling1D\nfrom keras.utils import plot_model\nfrom IPython.display import SVG\nfrom keras.utils.vis_utils import model_to_dot\nfrom keras.layers import LSTM\nfrom keras.layers import TimeDistributed\nfrom keras.layers import Bidirectional\nimport h5py\nimport logging\nimport time\n\nts = time.time()\ntime = datetime.datetime.fromtimestamp(ts).strftime(\"%H:%M:%S\")\ndate = datetime.datetime.fromtimestamp(ts).strftime(\"%Y-%m-%d\")\nlog_dir =\"path\/logs\/{}\/server_0\".format(date)\n\nif not os.path.isdir(log_dir):\n    os.makedirs(log_dir)\n\n#Visualize Model\ndef visualize_model(model):\n    SVG(model_to_dot(model).create(prog=\"dot\", format=\"svg\"))\n    return plot_model(model, show_shapes=True, show_layer_names=True, \n                      to_file=os.path.join(log_dir, \"network_architecture_{}.svg\".format(time)))\n\n\ndef keras_1D_CNN():\n\n    model = models.Sequential()\n    model.add(Convolution1D(input_shape=(253,21), \n                            kernel_initializer=\"he_uniform\", \n                            bias_initializer=initializers.Constant(0.1), \n                            nb_filter=256,\n                            filter_length=2,\n                            padding=\"same\",\n                            activation=\"relu\", \n                            dilation_rate=2, \n                            name='conv-1D_1'))\n    model.add(BatchNormalization(name='BatchNorm_1'))\n    model.add(MaxPooling1D(pool_size=2, strides=1, name='MaxPool1D_1'))\n\n    model.add(layers.Conv1D(128, 3, activation='relu', padding=\"same\", \n                            kernel_initializer=\"he_uniform\", \n                            bias_initializer=initializers.Constant(0.1), \n                            dilation_rate=3, \n                            name='conv-1D_2'))\n    model.add(BatchNormalization(name='BatchNorm_2'))\n    model.add(MaxPooling1D(pool_size=3, strides=1, name='MaxPool1D_2'))\n\n    model.add(layers.Conv1D(64, 3, activation='relu', padding=\"same\", \n                            kernel_initializer=\"he_uniform\", \n                            bias_initializer=initializers.Constant(0.1), \n                            dilation_rate=4, \n                            name='conv-1D_3'))\n    model.add(BatchNormalization(name='BatchNorm_3'))\n    model.add(MaxPooling1D(pool_size=4, strides=1, name='MaxPool1D_3'))\n\n    model.add(Bidirectional(LSTM(1024, return_sequences=True, dropout=0.5,\n                             recurrent_dropout=0.2), input_shape=(241, 1)))\n    model.add(TimeDistributed(Dense(128, activation='relu'), name='Dense_1'))\n    model.add(TimeDistributed(Dropout(0.4, name='TimeDistributed-Dropout_1 = 0.4')))\n    model.add(TimeDistributed(Dense(64, activation='relu'), name='Dense_2'))\n    model.add(Flatten(name='Flatten'))\n    model.add(Dropout(0.5, name='Dropout_1'))\n    model.add(Dense(64, activation='relu', \n                    kernel_initializer=\"he_uniform\", \n                    bias_initializer=initializers.Constant(0.1), \n                    name='Dense_3'))\n    model.add(Dropout(0.5, name='Dropout_2'))\n\n    model.add(Dense(1, activation='sigmoid', name='Sigmoid-layer'))\n\n    visualize_model(model)\n    model = multi_gpu_model(model, gpus=8)\n\n    opt = keras.optimizers.Adadelta()\n    tensorboard = TensorBoard(log_dir=os.path.join(log_dir, \"TensorBoard_{}\".format(time)), \n                              write_graph=True, write_grads=True, batch_size=2048)\n    model.compile(optimizer=opt,\n                  loss='binary_crossentropy',\n                  metrics=['accuracy'])\n\n    with open(os.path.join(log_dir, 'model_summary{}.txt'.format(time)),'w') as fh:\n        model.summary(print_fn=lambda x: fh.write(x + '\\n'))\n    checkpoint_filepath = os.path.join(log_dir, \"best-model-{}\".format(time) + \"_{epoch:02d}-{val_loss:.2f}.hdf5\")\n    checkpoint = ModelCheckpoint(checkpoint_filepath, monitor='val_loss', \n                                   verbose=1, save_best_only=True, mode=\"min\")\n    earlystopper = EarlyStopping(monitor='val_loss', patience=100, verbose=1)\n    CNN_history = model.fit(partial_seq_train, \n              partial_label_train, \n              epochs=1000, \n              batch_size=2048, \n              validation_split=0.2, shuffle=True, callbacks=[tensorboard, earlystopper, checkpoint])\n\n    return (model, CNN_history, results)\nCNN_1D_model = keras_1D_CNN()\n<\/code><\/pre>\n\n<pre><code>TypeError                                 Traceback (most recent call last)\n&lt;ipython-input-22-4d0435c2448d&gt; in &lt;module&gt;()\n      1 get_ipython().run_line_magic('time', '')\n----&gt; 2 CNN_1D_model = keras_1D_CNN()\n      3 CNN_history_dict = CNN_1D_model[1].history\n      4 CNN_history_dict.keys()\n      5 print(\"\\n%s: %.2f%%\" % (CNN_1D_model[0].metrics_names[1], CNN_1D_model[2][1]*100))\n\n&lt;ipython-input-21-2bf565bb8ec7&gt; in keras_1D_CNN()\n    100               epochs=1000,\n    101               batch_size=2048,\n--&gt; 102               validation_split=0.2, shuffle=True, callbacks=[tensorboard, earlystopper, checkpoint], verbose=0)\n    103 #     CNN_history = model.fit(partial_seq_train,\n    104 #               partial_label_train,\n\n\/d\/harpy1\/s\/python\/v3-5.1.0\/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\/d\/harpy1\/s\/python\/v3-5.1.0\/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    215                         for l, o in zip(out_labels, val_outs):\n    216                             epoch_logs['val_' + l] = o\n--&gt; 217         callbacks.on_epoch_end(epoch, epoch_logs)\n    218         if callback_model.stop_training:\n    219             break\n\n\/d\/harpy1\/s\/python\/v3-5.1.0\/lib\/python3.6\/site-packages\/keras\/callbacks.py in on_epoch_end(self, epoch, logs)\n     77         logs = logs or {}\n     78         for callback in self.callbacks:\n---&gt; 79             callback.on_epoch_end(epoch, logs)\n     80 \n     81     def on_batch_begin(self, batch, logs=None):\n\n\/d\/harpy1\/s\/python\/v3-5.1.0\/lib\/python3.6\/site-packages\/keras\/callbacks.py in on_epoch_end(self, epoch, logs)\n    444                             self.model.save_weights(filepath, overwrite=True)\n    445                         else:\n--&gt; 446                             self.model.save(filepath, overwrite=True)\n    447                     else:\n    448                         if self.verbose &gt; 0:\n\n\/d\/harpy1\/s\/python\/v3-5.1.0\/lib\/python3.6\/site-packages\/keras\/engine\/network.py in save(self, filepath, overwrite, include_optimizer)\n   1088             raise NotImplementedError\n   1089         from ..models import save_model\n-&gt; 1090         save_model(self, filepath, overwrite, include_optimizer)\n   1091 \n   1092     def save_weights(self, filepath, overwrite=True):\n\n\/d\/harpy1\/s\/python\/v3-5.1.0\/lib\/python3.6\/site-packages\/keras\/engine\/saving.py in save_model(model, filepath, overwrite, include_optimizer)\n    380 \n    381     try:\n--&gt; 382         _serialize_model(model, f, include_optimizer)\n    383     finally:\n    384         if opened_new_file:\n\n\/d\/harpy1\/s\/python\/v3-5.1.0\/lib\/python3.6\/site-packages\/keras\/engine\/saving.py in _serialize_model(model, f, include_optimizer)\n     82     model_config['class_name'] = model.__class__.__name__\n     83     model_config['config'] = model.get_config()\n---&gt; 84     model_config = json.dumps(model_config, default=get_json_type)\n     85     model_config = model_config.encode('utf-8')\n     86     f['model_config'] = model_config\n\n\/d\/harpy1\/s\/python\/v3-5.1.0\/lib\/python3.6\/json\/__init__.py in dumps(obj, skipkeys, ensure_ascii, check_circular, allow_nan, cls, indent, separators, default, sort_keys, **kw)\n    236         check_circular=check_circular, allow_nan=allow_nan, indent=indent,\n    237         separators=separators, default=default, sort_keys=sort_keys,\n--&gt; 238         **kw).encode(obj)\n    239 \n    240 \n\n\/d\/harpy1\/s\/python\/v3-5.1.0\/lib\/python3.6\/json\/encoder.py in encode(self, o)\n    197         # exceptions aren't as detailed.  The list call should be roughly\n    198         # equivalent to the PySequence_Fast that ''.join() would do.\n--&gt; 199         chunks = self.iterencode(o, _one_shot=True)\n    200         if not isinstance(chunks, (list, tuple)):\n    201             chunks = list(chunks)\n\n\/d\/harpy1\/s\/python\/v3-5.1.0\/lib\/python3.6\/json\/encoder.py in iterencode(self, o, _one_shot)\n    255                 self.key_separator, self.item_separator, self.sort_keys,\n    256                 self.skipkeys, _one_shot)\n--&gt; 257         return _iterencode(o, 0)\n    258 \n    259 def _make_iterencode(markers, _default, _encoder, _indent, _floatstr,\n\n\/d\/harpy1\/s\/python\/v3-5.1.0\/lib\/python3.6\/site-packages\/keras\/engine\/saving.py in get_json_type(obj)\n     72             return obj.__name__\n     73 \n---&gt; 74         raise TypeError('Not JSON Serializable: %s' % (obj,))\n     75 \n     76     from .. import __version__ as keras_version\n\nTypeError: Not JSON Serializable: &lt;module 'tensorflow' from '\/d\/harpy1\/s\/python\/v3-5.1.0\/lib\/python3.6\/site-packages\/tensorflow\/__init__.py'&gt;```\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57501493,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><lstm>","CreationDate":"2019-08-14T20:19:23.270","AcceptedAnswerId":null,"OwnerUserId":2730554.0,"Title":"how to use set_random_seed to get same results","Body":"<p>I am trying to learn about LSTM cells. I have read a few articles and am now trying to write my own code using tensorflow.<\/p>\n\n<p>I came across this code from this <a href=\"https:\/\/github.com\/llSourcell\/How-to-Predict-Stock-Prices-Easily-Demo\" rel=\"nofollow noreferrer\">tutorial<\/a>. So I copied the lstm.py file and the data file sp500.csv. The code below is what I written.<\/p>\n\n<p>Everything works fine. Each time I run the model I get very different predictions. I understand that because this model is very basic (trying to predict stock prices using just the closing price) this is to be expected. Also I believe in the background tensorflow uses some randomisation process to initiate the variables.<\/p>\n\n<p>I wanted though to run the model and get the same results each time. So I read that to do this I need to use set_random_seed.<\/p>\n\n<p>I have added this to my code below but when I re-run my code I get very different predictions. How am I supposed to use set_random_seed and why is the parameter 1234?<\/p>\n\n<pre><code>from tensorflow.python.keras.models import Sequential\nfrom tensorflow.python.keras.layers import Dense, LSTM, Dropout, Activation\nimport lstm, time\nimport tensorflow as tf\n\ndef RunML():\n\n    tf.set_random_seed(1234)\n\n    X_train, y_train, X_test, y_test = lstm.load_data('sp500.csv', 50, True)\n\n    model = Sequential()\n\n    model.add(LSTM(50, \n               input_shape=(50, 1),\n               return_sequences=True))\n    model.add(Dropout(0.2))\n\n    model.add(LSTM(\n        100,\n        return_sequences=False))\n    model.add(Dropout(0.2))\n\n    model.add(Dense(1))\n    model.add(Activation('linear'))\n\n    start = time.time()\n    model.compile(loss='mse', optimizer='rmsprop')\n    print ('compilation time : ', time.time() - start)\n\n    #Step 3 Train the model\n    model.fit(\n        X_train,\n        y_train,\n        batch_size=512,\n        nb_epoch=1,\n        validation_split=0.05)\n\n    predictions = lstm.predict_sequences_multiple(model, X_test, 50, 50)\n    lstm.plot_results_multiple(predictions, y_test, 50)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57503202","CreationDate":"2019-08-14T23:25:47.213","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>If you set the random seed before building the graph, you will get reproducible results. For example, given a random input:<\/p>\n\n<pre><code>random_input = np.random.rand(2, 50, 1)\n<\/code><\/pre>\n\n<p>We can define one <code>tf.keras.model<\/code>:<\/p>\n\n<pre><code>tf.keras.backend.clear_session()\ntf.set_random_seed(42)\n\nmodel_1 = Sequential()\n\nmodel_1.add(LSTM(50,  input_shape=(50, 1), return_sequences=True))\nmodel_1.add(Dropout(0.2))\n\nmodel_1.add(LSTM(100, return_sequences=False))\nmodel_1.add(Dropout(0.2))\n\nmodel_1.add(Dense(1))\nmodel_1.add(Activation('linear'))\n\nmodel_1.compile(loss='mse', optimizer='rmsprop')\n\nrandom_predict_1 = model_1.predict(random_input)\n<\/code><\/pre>\n\n<p>Then another:<\/p>\n\n<pre><code>tf.keras.backend.clear_session()\ntf.set_random_seed(42)\n\nmodel_2 = Sequential()\n\nmodel_2.add(LSTM(50,  input_shape=(50, 1), return_sequences=True))\nmodel_2.add(Dropout(0.2))\n\nmodel_2.add(LSTM(100, return_sequences=False))\nmodel_2.add(Dropout(0.2))\n\nmodel_2.add(Dense(1))\nmodel_2.add(Activation('linear'))\n\nmodel_2.compile(loss='mse', optimizer='rmsprop')\n\nrandom_predict_2 = model_2.predict(random_input)\n<\/code><\/pre>\n\n<p>Now, if we perform inference on the <code>random_input<\/code> tensor, we will get the same results with both models because they had the same random seed to start.<\/p>\n\n<pre><code>np.testing.assert_array_equal(random_predict_1, random_predict_2)\n<\/code><\/pre>\n\n<p>To sum it up, no matter which random seed you use, if you set the random seed to the same number before building the graph with <code>tf.set_random_seed(seed)<\/code>, you will get the same results.<\/p>\n"}]}
+{"QuestionId":57501515,"AnswerCount":0,"Tags":"<python><opencv><image-processing><keras><deep-learning>","CreationDate":"2019-08-14T20:20:52.240","AcceptedAnswerId":null,"OwnerUserId":11927768.0,"Title":"how to mask the load of open back trucks using Mask R-CNN and background subtraction?","Body":"<p>i need to detect and segment the load of open back trucks in the roads. how can i mask and extract the load area using background subtraction and Mask R-CNN ?<\/p>\n\n<p>as a simple example, i download a background image and Photoshop the truck and the load into that.<\/p>\n\n<p>background: <a href=\"https:\/\/i.ibb.co\/DpSWc84\/Original.jpg\" rel=\"nofollow noreferrer\">https:\/\/i.ibb.co\/DpSWc84\/Original.jpg<\/a><\/p>\n\n<p>modified: <a href=\"https:\/\/i.ibb.co\/6NVsYGq\/Modified.jpg\" rel=\"nofollow noreferrer\">https:\/\/i.ibb.co\/6NVsYGq\/Modified.jpg<\/a><\/p>\n\n<p>i used BS and Mask R-CNN but the output is not fit enough. output mask includes trucks border area and also some holes and noises in load section.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def applyMask(image, maskList): # apply masks that created by Mask R-CNN to image\n    masked_for_load = resize(image,maskList[0].shape)\n    for mask in maskList :\n        masked_for_load[:, :] = np.where(mask == 1,0,masked_for_load[:, :]) # set pixels that include cars or trucks to zero\n    return masked_for_load\n\n(score, diff) = compare_ssim(grayA, grayB, full=True)\ndiff = (diff * 255).astype(\"uint8\")\n\n# threshold the difference image, followed by finding contours to obtain the regions of the two input images that differ\nthresh = cv2.threshold(diff, 0, 255, cv2.THRESH_BINARY_INV | cv2.THRESH_OTSU)[1]\n# output mask, ready to apply on original image\nloadMask = applyMask(thresh,maskList)\n<\/code><\/pre>\n\n<p>Mask R-CNN output: <a href=\"https:\/\/i.ibb.co\/jgxr65Z\/bar-out.jpg\" rel=\"nofollow noreferrer\">https:\/\/i.ibb.co\/jgxr65Z\/bar-out.jpg<\/a><\/p>\n\n<p>BS output: <a href=\"https:\/\/i.ibb.co\/yQZT0V4\/before-mask.jpg\" rel=\"nofollow noreferrer\">https:\/\/i.ibb.co\/yQZT0V4\/before-mask.jpg<\/a><\/p>\n\n<p>after subtracting R-CNN output from BS output:  <a href=\"https:\/\/i.ibb.co\/JqZwfvg\/after-mask.jpg\" rel=\"nofollow noreferrer\">https:\/\/i.ibb.co\/JqZwfvg\/after-mask.jpg<\/a><\/p>\n\n<p>but the expected output must be like this: <a href=\"https:\/\/i.ibb.co\/z40fqjq\/expected.jpg\" rel=\"nofollow noreferrer\">https:\/\/i.ibb.co\/z40fqjq\/expected.jpg<\/a><\/p>\n\n<p>how can i achieve this using python, opencv and keras tools?<\/p>\n","answers":[]}
+{"QuestionId":57502077,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-08-14T21:07:18.777","AcceptedAnswerId":null,"OwnerUserId":6019463.0,"Title":"Using Keras embeddings layer with three-dimensional input","Body":"<p>I'm trying to see if it is possible to use the Keras embeddings layer on token-level character representations. <\/p>\n\n<p>That is, the input is an array of word-level tokens, broken down by character-level tokens like such: <\/p>\n\n<pre><code>sentence = 'this is a test'\narray = [[0, 1, 2, 3], [2, 4], [5], [0, 6, 7, 0]]\n<\/code><\/pre>\n\n<p>The idea is that it would generate an output of :<\/p>\n\n<pre><code>(None, max_words, max_chars, embedding_dim)\n<\/code><\/pre>\n\n<p>Maybe this could be done with some lambda function or custom layer?<\/p>\n\n<p>Any help would be greatly appreciated!<\/p>\n","answers":[]}
+{"QuestionId":57502226,"AnswerCount":1,"Tags":"<machine-learning><keras><scikit-learn>","CreationDate":"2019-08-14T21:21:14.793","AcceptedAnswerId":57503678.0,"OwnerUserId":4706745.0,"Title":"What is the best way to predict external test set?","Body":"<p>My Question is theoretical rather technical, hence I am not posting my code here because code is available at sklearn website itself. <\/p>\n\n<p>I constructed a classifier and cross-validating with 5 folds using sklearn. In my code, I am calculating various accuracy parameter like sensitivity, specificity, f1-score, etc. in each fold. After 5 fold cross-validation averaging all the accuracy parameters after the completion of the folds. <\/p>\n\n<p>Finally, my script creates a ROC curve along with the AUC score and histogram for other accuracy parameters and generates an HTML report file.<\/p>\n\n<p>Cross-validation means internal testing, but when I used an external test data set confusion starts. <\/p>\n\n<p>My question is that how should I predict external data set, which is the correct amongst the bellow method. <\/p>\n\n<ol>\n<li><p>After cross-validation, save the model which has averaged parameters from each and every fold and use this model to predict external test set and calculate assessment report. If this is the case how can I do this? Can you show me the example code which helps me to save a model after n fold cross-validation? <\/p><\/li>\n<li><p>Build the model using the entire data set, save the model to predict the external test set and calculate the assessment report. If this is the correct way, Thank you I know the code. <\/p><\/li>\n<li><p>Is there any other method that I missed, please share. <\/p><\/li>\n<\/ol>\n\n<p>Thank you. <\/p>\n","answers":[{"AnswerId":"57503678","CreationDate":"2019-08-15T00:53:26.283","ParentId":null,"OwnerUserId":"6347629","Title":null,"Body":"<p>The correct approach  is <\/p>\n\n<blockquote>\n  <p>Build the model using the entire data set, save the model to predict\n  the external test set and calculate the assessment report<\/p>\n<\/blockquote>\n\n<p>The reason being we use cross validation only to measure the performance of hyper parameters only. We do that by keeping the every fold as a test fold once which means every data point gets a fair chance of being as a test data point once. <\/p>\n"}]}
+{"QuestionId":57502753,"AnswerCount":1,"Tags":"<python><keras><deep-learning><pytorch>","CreationDate":"2019-08-14T22:18:39.963","AcceptedAnswerId":null,"OwnerUserId":11929307.0,"Title":"How to get this variational auto encoder to use images with different resolution?","Body":"<p>I am trying to get an existing variational auto encoder to work with RGB images that have a resolution of 254 * 254<\/p>\n\n<p>The code for the variational auto encoder works great for a dataset with resolutions of 100 x 100. However, I have no idea how to modify it to work with images of 254 X 254. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>batch_size = 5\nepochs = 10\nsize = 15\nno_cuda = True\nseed = 1\nlog_interval = 50\nresults_path = 'C:\\\\results\\\\'\ntrain_root = 'C:\\\\train\\\\'\nval_root = 'C:\\\\train\\\\'\n\ncuda = not no_cuda and torch.cuda.is_available()\n\ntorch.manual_seed(seed)\n\ndevice = torch.device(\"cuda\" if cuda else \"cpu\")\nkwargs = {'num_workers': 1, 'pin_memory': True} if cuda else {}\n\n\ntrain_loader_food = torch.utils.data.DataLoader(\n    datasets.ImageFolder(train_root, transform=transforms.ToTensor()),\n    batch_size = batch_size, shuffle=True, **kwargs)\n\nval_loader_food = torch.utils.data.DataLoader(\n    datasets.ImageFolder(val_root, transform=transforms.ToTensor()),\n    batch_size = batch_size, shuffle=True, **kwargs)\n\nclass VAE_CNN(nn.Module):\n    def __init__(self):\n        super(VAE_CNN, self).__init__()\n\n        # Encoder\n        self.conv1 = nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1, bias=False)\n        self.bn1 = nn.BatchNorm2d(16)\n        self.conv2 = nn.Conv2d(16, 32, kernel_size=3, stride=2, padding=1, bias=False)\n        self.bn2 = nn.BatchNorm2d(32)\n        self.conv3 = nn.Conv2d(32, 64, kernel_size=3, stride=1, padding=1, bias=False)\n        self.bn3 = nn.BatchNorm2d(64)\n        self.conv4 = nn.Conv2d(64, 16, kernel_size=3, stride=2, padding=1, bias=False)\n        self.bn4 = nn.BatchNorm2d(16)\n\n        # Latent vectors mu and sigma\n        self.fc1 = nn.Linear(25 * 25 * 16, 2048)\n        self.fc_bn1 = nn.BatchNorm1d(2048)\n        self.fc21 = nn.Linear(2048, 2048)\n        self.fc22 = nn.Linear(2048, 2048)\n\n        # Sampling vector\n        self.fc3 = nn.Linear(2048, 2048)\n        self.fc_bn3 = nn.BatchNorm1d(2048)\n        self.fc4 = nn.Linear(2048, 25 * 25 * 16)\n        self.fc_bn4 = nn.BatchNorm1d(25 * 25 * 16)\n\n        # Decoder\n        self.conv5 = nn.ConvTranspose2d(16, 64, kernel_size=3, stride=2, padding=1, output_padding=1, bias=False)\n        self.bn5 = nn.BatchNorm2d(64)\n        self.conv6 = nn.ConvTranspose2d(64, 32, kernel_size=3, stride=1, padding=1, bias=False)\n        self.bn6 = nn.BatchNorm2d(32)\n        self.conv7 = nn.ConvTranspose2d(32, 16, kernel_size=3, stride=2, padding=1, output_padding=1, bias=False)\n        self.bn7 = nn.BatchNorm2d(16)\n        self.conv8 = nn.ConvTranspose2d(16, 3, kernel_size=3, stride=1, padding=1, bias=False)\n\n        self.relu = nn.ReLU()\n\n    def encode(self, x):\n        conv1 = self.relu(self.bn1(self.conv1(x)))\n        conv2 = self.relu(self.bn2(self.conv2(conv1)))\n        conv3 = self.relu(self.bn3(self.conv3(conv2)))\n        conv4 = self.relu(self.bn4(self.conv4(conv3))).view(-1, 25 * 25 * 16)\n\n        fc1 = self.relu(self.fc_bn1(self.fc1(conv4)))\n\n        r1 = self.fc21(fc1)\n        r2 = self.fc22(fc1)\n\n        return r1, r2\n\n    def reparameterize(self, mu, logvar):\n        if self.training:\n            std = logvar.mul(0.5).exp_()\n            eps = Variable(std.data.new(std.size()).normal_())\n            return eps.mul(std).add_(mu)\n        else:\n            return mu\n\n    def decode(self, z):\n        fc3 = self.relu(self.fc_bn3(self.fc3(z)))\n        fc4 = self.relu(self.fc_bn4(self.fc4(fc3))).view(-1, 16, 25, 25)\n\n        conv5 = self.relu(self.bn5(self.conv5(fc4)))\n        conv6 = self.relu(self.bn6(self.conv6(conv5)))\n        conv7 = self.relu(self.bn7(self.conv7(conv6)))\n        return self.conv8(conv7).view(-1, 3, 100, 100)\n\n    def forward(self, x):\n        mu, logvar = self.encode(x)\n        z = self.reparameterize(mu, logvar)\n        return self.decode(z), mu, logvar\n\nclass customLoss(nn.Module):\n    def __init__(self):\n        super(customLoss, self).__init__()\n        self.mse_loss = nn.MSELoss(reduction=\"sum\")\n\n    def forward(self, x_recon, x, mu, logvar):\n        loss_MSE = self.mse_loss(x_recon, x)\n        loss_KLD = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())\n\n        return loss_MSE + loss_KLD\n\nmodel = VAE_CNN().to(device)\noptimizer = optim.Adam(model.parameters(), lr=1e-3)\n\nloss_mse = customLoss()\n\nval_losses = []\ntrain_losses = []\n\ndef train(epoch):\n    model.train()\n    train_loss = 0\n    for batch_idx, (data, _) in enumerate(train_loader_food):\n        data = data.to(device)\n        optimizer.zero_grad()\n        recon_batch, mu, logvar = model(data)\n        loss = loss_mse(recon_batch, data, mu, logvar)\n        loss.backward()\n        train_loss += loss.item()\n        optimizer.step()\n        if batch_idx % log_interval == 0:\n            print('Train Epoch: {} [{}\/{} ({:.0f}%)]\\tLoss: {:.6f}'.format(\n                epoch, batch_idx * len(data), len(train_loader_food.dataset),\n                       100. * batch_idx \/ len(train_loader_food),\n                       loss.item() \/ len(data)))\n\n    print('====&gt; Epoch: {} Average loss: {:.4f}'.format(\n        epoch, train_loss \/ len(train_loader_food.dataset)))\n    train_losses.append(train_loss \/ len(train_loader_food.dataset))\n<\/code><\/pre>\n\n<p>I replaced any refence to 25 * 25 * 16 with 128 * 128 * 4 etc.. assuming that would fix the issue but it just yields the following error:<\/p>\n\n<p>shape '[-1, 64516]' is invalid for input of size 327680<\/p>\n","answers":[{"AnswerId":"57510216","CreationDate":"2019-08-15T13:14:23.700","ParentId":null,"OwnerUserId":"11841075","Title":null,"Body":"<p>The <code>25*25*16<\/code> comes from number of output channels of <code>conv4<\/code> which is <code>16<\/code> and 1\/4 of original image dimension (<code>100\/4 = 25<\/code>).\nA more generic model code would be:<\/p>\n\n<pre><code>\nclass VAE_CNN(nn.Module):\n    def __init__(self, h, w):\n        super(VAE_CNN, self).__init__()\n        self.h = h\n        self.w = w\n        # Encoder\n        self.conv1 = nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1, bias=False)\n        self.bn1 = nn.BatchNorm2d(16)\n        self.conv2 = nn.Conv2d(16, 32, kernel_size=3, stride=2, padding=1, bias=False)\n        self.bn2 = nn.BatchNorm2d(32)\n        self.conv3 = nn.Conv2d(32, 64, kernel_size=3, stride=1, padding=1, bias=False)\n        self.bn3 = nn.BatchNorm2d(64)\n        self.conv4 = nn.Conv2d(64, 16, kernel_size=3, stride=2, padding=1, bias=False)\n        self.bn4 = nn.BatchNorm2d(16)\n\n        # Latent vectors mu and sigma\n        self.fc1 = nn.Linear(h\/\/4 * w\/\/4 * 16, 2048)\n        self.fc_bn1 = nn.BatchNorm1d(2048)\n        self.fc21 = nn.Linear(2048, 2048)\n        self.fc22 = nn.Linear(2048, 2048)\n\n        # Sampling vector\n        self.fc3 = nn.Linear(2048, 2048)\n        self.fc_bn3 = nn.BatchNorm1d(2048)\n        self.fc4 = nn.Linear(2048, h\/\/4 * w\/\/4 * 16)\n        self.fc_bn4 = nn.BatchNorm1d(h\/\/4 * w\/\/4 * 16)\n\n        # Decoder\n        self.conv5 = nn.ConvTranspose2d(16, 64, kernel_size=3, stride=2, padding=1, output_padding=1, bias=False)\n        self.bn5 = nn.BatchNorm2d(64)\n        self.conv6 = nn.ConvTranspose2d(64, 32, kernel_size=3, stride=1, padding=1, bias=False)\n        self.bn6 = nn.BatchNorm2d(32)\n        self.conv7 = nn.ConvTranspose2d(32, 16, kernel_size=3, stride=2, padding=1, output_padding=1, bias=False)\n        self.bn7 = nn.BatchNorm2d(16)\n        self.conv8 = nn.ConvTranspose2d(16, 3, kernel_size=3, stride=1, padding=1, bias=False)\n\n        self.relu = nn.ReLU()\n\n    def encode(self, x):\n        conv1 = self.relu(self.bn1(self.conv1(x)))\n        conv2 = self.relu(self.bn2(self.conv2(conv1)))\n        conv3 = self.relu(self.bn3(self.conv3(conv2)))\n        conv4 = self.relu(self.bn4(self.conv4(conv3)))\n        conv4 = conv4.view(-1, self.h\/\/4 * self.w\/\/4 * 16)\n\n        fc1 = self.relu(self.fc_bn1(self.fc1(conv4)))\n\n        r1 = self.fc21(fc1)\n        r2 = self.fc22(fc1)\n\n        return r1, r2\n\n    def reparameterize(self, mu, logvar):\n        if self.training:\n            std = logvar.mul(0.5).exp_()\n            eps = Variable(std.data.new(std.size()).normal_())\n            return eps.mul(std).add_(mu)\n        else:\n            return mu\n\n    def decode(self, z):\n        fc3 = self.relu(self.fc_bn3(self.fc3(z)))\n        fc4 = self.relu(self.fc_bn4(self.fc4(fc3))).view(-1, 16, self.h\/\/4, self.w\/\/4)\n\n        conv5 = self.relu(self.bn5(self.conv5(fc4)))\n        conv6 = self.relu(self.bn6(self.conv6(conv5)))\n        conv7 = self.relu(self.bn7(self.conv7(conv6)))\n        return self.conv8(conv7).view(-1, 3, self.h, self.w)\n\n    def forward(self, x):\n        mu, logvar = self.encode(x)\n        z = self.reparameterize(mu, logvar)\n        return self.decode(z), mu, logvar\n\nclass customLoss(nn.Module):\n    def __init__(self):\n        super(customLoss, self).__init__()\n        self.mse_loss = nn.MSELoss(reduction=\"sum\")\n\n    def forward(self, x_recon, x, mu, logvar):\n        loss_MSE = self.mse_loss(x_recon, x)\n        loss_KLD = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())\n\n        return loss_MSE + loss_KLD\n\nmodel = VAE_CNN(100, 100).to(device)\n<\/code><\/pre>\n\n<p>This would work fine for any image having image dimensions divisible by 4.<\/p>\n\n<p>Now, as <code>254%4 != 0<\/code>. You can try this one:<\/p>\n\n<pre><code>class VAE_CNN(nn.Module):\n    def __init__(self, h, w):\n        super(VAE_CNN, self).__init__()\n        self.h = h\n        self.w = w\n        # Encoder\n        self.conv1 = nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1, bias=False)\n        self.bn1 = nn.BatchNorm2d(16)\n        self.conv2 = nn.Conv2d(16, 32, kernel_size=3, stride=2, padding=1, bias=False)\n        self.bn2 = nn.BatchNorm2d(32)\n        self.conv3 = nn.Conv2d(32, 64, kernel_size=3, stride=1, padding=1, bias=False)\n        self.bn3 = nn.BatchNorm2d(64)\n        self.conv4 = nn.Conv2d(64, 16, kernel_size=3, stride=1, padding=1, bias=False)\n        self.bn4 = nn.BatchNorm2d(16)\n\n        # Latent vectors mu and sigma\n        self.fc1 = nn.Linear(h\/\/2 * w\/\/2 * 16, 2048)\n        self.fc_bn1 = nn.BatchNorm1d(2048)\n        self.fc21 = nn.Linear(2048, 2048)\n        self.fc22 = nn.Linear(2048, 2048)\n\n        # Sampling vector\n        self.fc3 = nn.Linear(2048, 2048)\n        self.fc_bn3 = nn.BatchNorm1d(2048)\n        self.fc4 = nn.Linear(2048, h\/\/2 * w\/\/2 * 16)\n        self.fc_bn4 = nn.BatchNorm1d(h\/\/2 * w\/\/2 * 16)\n\n        # Decoder\n        self.conv5 = nn.ConvTranspose2d(16, 64, kernel_size=3, stride=1, padding=1, bias=False)\n        self.bn5 = nn.BatchNorm2d(64)\n        self.conv6 = nn.ConvTranspose2d(64, 32, kernel_size=3, stride=1, padding=1, bias=False)\n        self.bn6 = nn.BatchNorm2d(32)\n        self.conv7 = nn.ConvTranspose2d(32, 16, kernel_size=3, stride=2, padding=1, output_padding=1, bias=False)\n        self.bn7 = nn.BatchNorm2d(16)\n        self.conv8 = nn.ConvTranspose2d(16, 3, kernel_size=3, stride=1, padding=1, bias=False)\n\n        self.relu = nn.ReLU()\n\n    def encode(self, x):\n        conv1 = self.relu(self.bn1(self.conv1(x)))\n        conv2 = self.relu(self.bn2(self.conv2(conv1)))\n        conv3 = self.relu(self.bn3(self.conv3(conv2)))\n        conv4 = self.relu(self.bn4(self.conv4(conv3)))\n        conv4 = conv4.view(-1, self.h\/\/2 * self.w\/\/2 * 16)\n\n        fc1 = self.relu(self.fc_bn1(self.fc1(conv4)))\n\n        r1 = self.fc21(fc1)\n        r2 = self.fc22(fc1)\n\n        return r1, r2\n\n    def reparameterize(self, mu, logvar):\n        if self.training:\n            std = logvar.mul(0.5).exp_()\n            eps = Variable(std.data.new(std.size()).normal_())\n            return eps.mul(std).add_(mu)\n        else:\n            return mu\n\n    def decode(self, z):\n        fc3 = self.relu(self.fc_bn3(self.fc3(z)))\n        fc4 = self.relu(self.fc_bn4(self.fc4(fc3))).view(-1, 16, self.h\/\/2, self.w\/\/2)\n\n        conv5 = self.relu(self.bn5(self.conv5(fc4)))\n        conv6 = self.relu(self.bn6(self.conv6(conv5)))\n        conv7 = self.relu(self.bn7(self.conv7(conv6)))\n        return self.conv8(conv7).view(-1, 3, self.h, self.w)\n\n    def forward(self, x):\n        mu, logvar = self.encode(x)\n        z = self.reparameterize(mu, logvar)\n        return self.decode(z), mu, logvar\n\nclass customLoss(nn.Module):\n    def __init__(self):\n        super(customLoss, self).__init__()\n        self.mse_loss = nn.MSELoss(reduction=\"sum\")\n\n    def forward(self, x_recon, x, mu, logvar):\n        loss_MSE = self.mse_loss(x_recon, x)\n        loss_KLD = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())\n\n        return loss_MSE + loss_KLD\n\nmodel = VAE_CNN(254, 254).to(device)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57503075,"AnswerCount":1,"Tags":"<tensorflow><cuda>","CreationDate":"2019-08-14T23:04:46.403","AcceptedAnswerId":null,"OwnerUserId":11929538.0,"Title":"Can CUDA 10.0 and 10.1 be on the same system?","Body":"<p>I want support for both Visual Studio 2019 (which needs CUDA 10.1) and TensorFlow-GPU 1.14 (which needs CUDA 10.0) on a Windows PC. Is there any methods?<\/p>\n\n<p>I simply installed CUDA 10.0 and CUDA 10.1, and add both directory into environment variable CUDA_PATH. cuDNN is already installed.<\/p>\n\n<p>The result is Visual Studio can detect CUDA but TensorFlow cannot.<\/p>\n","answers":[{"AnswerId":"57503313","CreationDate":"2019-08-14T23:44:23.100","ParentId":null,"OwnerUserId":"10543528","Title":null,"Body":"<p>Yes, more than one version of the CUDA toolkit can exist on a system and be used by different applications. <\/p>\n\n<p>How are you installing TensorFlow-GPU? If you're compiling it yourself, during configuration you can specify the path to whichever version of CUDA that you want to use. If you're installing a pre-built set of binaries (e.g. using something like Anaconda) then that's already been built against a specific version of the CUDA toolkit; you'll need to fetch a different version of the binaries compiled for whichever CUDA toolkit you want, or build it yourself.<\/p>\n\n<p>If you use Anaconda to install TensorFlow-GPU, you should also receive the correct version of the CUDA toolkit that's needed to run whichever version of TensorFlow-GPU that you've installed; it takes care of those dependencies for you.<\/p>\n"}]}
+{"QuestionId":57503228,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2019-08-14T23:29:33.837","AcceptedAnswerId":57507259.0,"OwnerUserId":867889.0,"Title":"How to copy an graph operation in Tensorflow using node_def?","Body":"<p>Lets say I have an operation <code>my_op<\/code> defined as follows:<\/p>\n\n<pre><code>name: \"Const\"\nop: \"Const\"\nattr {\n  key: \"dtype\"\n  value {\n    type: DT_INT32\n  }\n}\nattr {\n  key: \"value\"\n  value {\n    tensor {\n      dtype: DT_INT32\n      tensor_shape {\n        dim {\n          size: 2\n        }\n      }\n      tensor_content: \"\\001\\000\\000\\000\\001\\000\\000\\000\"\n    }\n  }\n}\n<\/code><\/pre>\n\n<p>I have access to operation in the graph but not to the construction code. I want to copy operation and change some of it's attributes:<\/p>\n\n<pre><code>name: \"PrettyConst\"\nop: \"Const\"\nattr {\n  key: \"dtype\"\n  value {\n    type: DT_INT32\n  }\n}\nattr {\n  key: \"value\"\n  value {\n    tensor {\n      dtype: DT_FLOAT32\n      tensor_shape {\n        dim {\n          size: 8\n        }\n      }\n      tensor_content: \"\\001\\000\\000\\000\\001\\000\\000\\000\\001\\000\\000\\000\\001\\000\\000\\000\\001\\000\\000\\000\\001\\000\\000\\000\\001\\000\\000\\000\\001\\000\\000\\000\"\n    }\n  }\n}\n<\/code><\/pre>\n\n<p>While I can easily do it by saving graph as txt, modifying content of the file, and restoring it back, I can not find a simple way of doing that in python. I would expect there should be possible to do something like following:<\/p>\n\n<pre><code>op_def_copy = op.node_def.copy()\nop_def_copy.name = \"PrettyConst\"\nop_def_copy.attr[\"dtype\"] = 0\n# and also do something with the content, whatever\ngraph.append(tf.Operation(op_def_copy))\n<\/code><\/pre>\n\n<p>The <code>tf.contrig.graph_editor<\/code> seems to be able to do that sort of thing, though.<\/p>\n","answers":[{"AnswerId":"57527957","CreationDate":"2019-08-16T16:15:16.090","ParentId":null,"OwnerUserId":"867889","Title":null,"Body":"<p>@jdehesa gave a great answer to the question. I have a little more tooling around it:<\/p>\n\n<pre><code>import tensorflow\nimport copy\nimport tensorflow.contrib.graph_editor as ge\nfrom copy import deepcopy\n\na = tf.constant(1)\nb = tf.constant(2)\nc = a+b\n\ndef modify(t): \n    # illustrate operation copy&amp;modification\n    new_t = deepcopy(t.op.node_def)\n    new_t.name = new_t.name+\"_but_awesome\"\n    new_t = tf.Operation(new_t, tf.get_default_graph())\n    # we got a tensor, let's return a tensor\n    return new_t.outputs[0]\n\ndef update_existing(target, updated):\n    # illustrate how to use new op\n    related_ops = ge.get_backward_walk_ops(target, stop_at_ts=updated.keys(), inclusive=True)\n    new_ops, mapping = ge.copy_with_input_replacements(related_ops, updated)\n    new_op = mapping._transformed_ops[target.op]\n    return new_op.outputs[0]\n\nnew_a = modify(a)\nnew_b = modify(b)\ninjection = new_a+39 # illustrate how to add another op to the graph\nnew_c = update_existing(c, {a:injection, b:new_b})\n\nwith tf.Session():\n    print(c.eval()) # -&gt; 3\n    print(new_c.eval()) # -&gt; 42\n<\/code><\/pre>\n"},{"AnswerId":"57507259","CreationDate":"2019-08-15T09:00:44.060","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You have to parse the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/NodeDef\" rel=\"nofollow noreferrer\"><code>tf.NodeDef<\/code><\/a> message from the text representation, and then you can build the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Operation\" rel=\"nofollow noreferrer\"><code>tf.Operation<\/code><\/a> from it. You can do that like this:<\/p>\n\n<pre><code>import tensorflow as tf\nimport google.protobuf\n\nnode_def_message = r\"\"\"name: \"Const\"\nop: \"Const\"\nattr {\n  key: \"dtype\"\n  value {\n    type: DT_INT32\n  }\n}\nattr {\n  key: \"value\"\n  value {\n    tensor {\n      dtype: DT_INT32\n      tensor_shape {\n        dim {\n          size: 2\n        }\n      }\n      tensor_content: \"\\001\\000\\000\\000\\001\\000\\000\\000\"\n    }\n  }\n}\"\"\"\n\n# Build NodeDef message\nnode_def = tf.NodeDef()\n# Parse from the string\ngoogle.protobuf.text_format.Parse(node_def_message, node_def)\n# Build the operation\nop = tf.Operation(node_def, tf.get_default_graph())\n# Get the output from the operation\nc = op.outputs[0]\n# Check value\nwith tf.Session() as sess:\n    print(sess.run(c))\n    # [1 1]\n<\/code><\/pre>\n\n<p>Note that you may need to pass additional parameters to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Operation\" rel=\"nofollow noreferrer\"><code>tf.Operation<\/code><\/a> if, for example, the operation that you are building needs some input.<\/p>\n"}]}
+{"QuestionId":57503272,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><classification><conv-neural-network>","CreationDate":"2019-08-14T23:37:05.330","AcceptedAnswerId":null,"OwnerUserId":7667138.0,"Title":"Gender detection and age classification using deep learning","Body":"<p>I am working on a project where the goal is to detect gender and classify image. I did some research and found out a research paper: AgeandGenderClassi\ufb01cationusingConvolutionalNeuralNetworks by Gil Levi and Tal Hassnar. I tried to replicate the deep network created by them, originally in Caffe, in Keras. But the problem is the model is stuck at 50% accuracy(basically random coin toss). What is it that i am doing wrong. Any help is much appreciated.\nBTW i am using the adience data set as the original paper.\nPS: I have completely removed the LRN Layers as they are not available  in Keras. ( I think their absence should not hurt the accuracy of the model )\nHere is the code.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#imports\nimport os\nimport numpy as np\nfrom PIL import Image\nimport pickle\nfrom keras.models import Sequential \nfrom keras.callbacks import  ModelCheckpoint\nfrom keras.layers import Dense , Conv2D , Flatten , MaxPooling2D , Dropout , AveragePooling2D\nfrom keras import initializers\nfrom keras import optimizers\n\n# creating the model object\ngender_model = Sequential()\n\n# adding layers to the model\n# first convolutional layer\ngender_model.add( Conv2D(96 , kernel_size=(7,7) , activation='relu' , strides=4 , input_shape=(227,227,3),\n                            kernel_initializer= initializers.random_normal(stddev=0.01), use_bias = 1,\n                            bias_initializer = 'Zeros' , data_format='channels_last'))\n\ngender_model.add( MaxPooling2D(pool_size=3 , strides=2) )\n\ngender_model.add( Conv2D(256, kernel_size=(5,5) , activation='relu', strides=1 , padding='same' , input_shape=(27,27,96), \n                            kernel_initializer= initializers.random_normal(stddev=0.01) , use_bias=1,\n                            bias_initializer='Ones' , data_format='channels_last') )\ngender_model.add( MaxPooling2D(pool_size=3 , strides=2) )\n# third convolutional layer\n\ngender_model.add( Conv2D(384,kernel_size=(3,3) , activation='relu', strides=1 ,padding='same', input_shape=(13,13,256),\n                    kernel_initializer= initializers.random_normal(stddev=0.01), use_bias=1,\n                    bias_initializer = 'Zeros' , data_format='channels_last') )\ngender_model.add( MaxPooling2D(pool_size=3 , strides=2) )\n\n# Now we flatten the output of last convolutional layer\ngender_model.add( Flatten() )\n\n# Now we connect the fully connected layers\ngender_model.add( Dense(512, activation='relu' , use_bias=1, kernel_initializer=initializers.random_normal(stddev=0.005),\n                    bias_initializer='Ones') )\ngender_model.add( Dropout(0.5))\n\n# connecting another fully connected layer\ngender_model.add( Dense(512 , activation='relu' , use_bias=1, kernel_initializer=initializers.random_normal(stddev=0.005),\n                    bias_initializer='Ones'))\ngender_model.add( Dropout(0.5))\n\n# connecting the final layer\ngender_model.add( Dense(2, activation='softmax' , use_bias=1, kernel_initializer=initializers.random_normal(stddev=0.01),\n                    bias_initializer='Zeros'))\n\n# compiling the model\nsgd_optimizer = optimizers.SGD(lr= 0.0001 , decay=1e-7 , momentum=0.0, nesterov=False)\ngender_model.compile(optimizer=sgd_optimizer , loss= 'categorical_crossentropy' , metrics=['accuracy'])\ngender_model.summary()\n\n\n# partioning the loaded data\nX = np.load('\/content\/drive\/My Drive\/X.npy')\ny = np.load('\/content\/drive\/My Drive\/y_m.npy')\n\nX_train = X[:15000]\ny_train = y[:15000]\n\nX_val = X[15000:] \ny_val = y[15000:]\n\n## creating chkpt path\nchkpt_path = 'weights-improvement-{epoch:02d}--{val_acc:.2f}.hdf5'\n\ncheckpoint = ModelCheckpoint(chkpt_path , monitor='val_acc' , verbose=1 , save_best_only=True , mode='max')\ncallback_list = [checkpoint]\n\n#finally training the model\ngender_model.fit(X_train, y_train,\n          batch_size=50,\n          epochs=100,\n          validation_data=(X_val , y_val),\n          shuffle=1,\n          callbacks = callback_list\n          )\n<\/code><\/pre>\n","answers":[{"AnswerId":"57632458","CreationDate":"2019-08-23T20:07:57.013","ParentId":null,"OwnerUserId":"7667138","Title":null,"Body":"<p>There was a problem in my approach. I was not cropping the faces. And hence the model could not make sense of the random background in every image.<\/p>\n"}]}
+{"QuestionId":57503473,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-08-15T00:11:44.450","AcceptedAnswerId":57539079.0,"OwnerUserId":5739619.0,"Title":"ModuleNotFoundError: no module named efficientnet.tfkeras","Body":"<p>I attempted to do <code>import segmentation_models as sm<\/code>, but I got an error saying <code>efficientnet<\/code> was not found. So I then did <code>pip install efficientnet<\/code> and tried it again. I now get <code>ModuleNotFoundError: no module named efficientnet.tfkeras<\/code>, even though Keras is installed as I'm able to do <code>from keras.models import *<\/code> or anything else with Keras<\/p>\n\n<p>how can I get rid of this error?<\/p>\n","answers":[{"AnswerId":"57539079","CreationDate":"2019-08-17T18:52:08.670","ParentId":null,"OwnerUserId":"11940766","Title":null,"Body":"<p>To install segmentation-models use the following command: <code>pip install git+https:\/\/github.com\/qubvel\/segmentation_models<\/code><\/p>\n"}]}
+{"QuestionId":57503584,"AnswerCount":1,"Tags":"<python><c++><windows><tensorflow>","CreationDate":"2019-08-15T00:31:59.000","AcceptedAnswerId":null,"OwnerUserId":2056201.0,"Title":"Cannot build the tensorflow C++ API in windows","Body":"<p>I followed every step on the instructions and it still won't build\n<a href=\"https:\/\/www.tensorflow.org\/install\/source_windows\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/source_windows<\/a><\/p>\n\n<ol>\n<li><p>I cannot include the CUDA library and include files, because the python script won't recognize 2 comma separated paths during .\/configure.py stage\nIt only recognizes the first one I enter, as you can see I entered them in reverse order and it see the first one just fine. Afaik this is a new addition, because the instructions ask for path at different times. <\/p>\n\n<p>Please specify the CUDA SDK version you want to use. [Leave empty to default to CUDA 10]:<\/p>\n\n<pre><code>Please specify the cuDNN version you want to use. [Leave empty to default to cuDNN 7]:\n\nPlease specify the comma-separated list of base paths to look for CUDA libraries and headers. [Leave empty to use the default]: C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v10.0\\include, C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v10.0\\lib\\x64\n\nCould not find any cudart.lib in any subdirectory:\n''\n'lib64'\n'lib'\n'lib\/-linux-gnu'\n'lib\/x64'\n'extras\/CUPTI\/'\nof:\n'C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v10.0\\include'\nAsking for detailed CUDA configuration...\n\nPlease specify the CUDA SDK version you want to use. [Leave empty to default to CUDA 10]:\n\nPlease specify the cuDNN version you want to use. [Leave empty to default to cuDNN 7]:\n\nPlease specify the comma-separated list of base paths to look for CUDA libraries and headers. [Leave empty to use the default]: C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v10.0\\lib\\x64 C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v10.0\\include\n\nCould not find any cuda.h matching version '10' in any subdirectory:\n''\n'include'\n'include\/cuda'\n'include\/*-linux-gnu'\n'extras\/CUPTI\/include'\n'include\/cuda\/CUPTI'\nof:\nAsking for detailed CUDA configuration...\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>So I have to disable CUDA support\nThen Im just trying to just build the DLL\nand I get this mess of errors<\/p>\n\n<pre><code>&gt; C:\\Temp\\tensorflow-master&gt;bazel build --config=opt \/\/tensorflow:tensorflow.dll\n&gt; WARNING: Ignoring JAVA_HOME, because it must point to a JDK, not a JRE.\n&gt; INFO: Options provided by the client:\n&gt;   Inherited 'common' options: --isatty=1 --terminal_columns=120\n&gt; INFO: Options provided by the client:\n&gt;   'build' options: --python_path=C:\/ProgramData\/Anaconda3\/python.exe\n&gt; INFO: Reading rc options for 'build' from c:\\temp\\tensorflow-master\\.bazelrc:\n&gt;   'build' options: --apple_platform_type=macos --define framework_shared_object=true --define open_source_build=true --define=use_fast_cpp_protos=true --define=allow_oversize_protos=true --spawn_strategy=standalone --strategy=Genrule=standalone -c opt --announce_rc --define=grpc_no_ares=true --define=PREFIX=\/usr --define=LIBDIR=$(PREFIX)\/lib --define=INCLUDEDIR=$(PREFIX)\/include\n&gt; INFO: Reading rc options for 'build' from c:\\temp\\tensorflow-master\\.tf_configure.bazelrc:\n&gt;   'build' options: --action_env PYTHON_BIN_PATH=C:\/ProgramData\/Anaconda3\/python.exe --action_env PYTHON_LIB_PATH=C:\/ProgramData\/Anaconda3\/lib\/site-packages --python_path=C:\/ProgramData\/Anaconda3\/python.exe --config monolithic --copt=-w --host_copt=-w --copt=-DWIN32_LEAN_AND_MEAN --host_copt=-DWIN32_LEAN_AND_MEAN --copt=-DNOGDI --host_copt=-DNOGDI --verbose_failures --distinct_host_configuration=false --action_env TF_CONFIGURE_IOS=0\n&gt; INFO: Found applicable config definition build:monolithic in file c:\\temp\\tensorflow-master\\.bazelrc: --define framework_shared_object=false\n&gt; INFO: Found applicable config definition build:opt in file c:\\temp\\tensorflow-master\\.tf_configure.bazelrc: --copt=\/arch:AVX --define with_default_optimizations=true\n&gt; INFO: An error occurred during the fetch of repository 'com_google_protobuf'\n&gt; INFO: Call stack for the definition of repository 'com_google_protobuf':\n&gt;  - C:\/temp\/tensorflow-master\/tensorflow\/workspace.bzl:432:5\n&gt;  - C:\/temp\/tensorflow-master\/WORKSPACE:26:1\n&gt; ERROR: Skipping '\/\/tensorflow:tensorflow.dll': error loading package 'tensorflow': in C:\/temp\/tensorflow-master\/tensorflow\/core\/platform\/default\/build_config.bzl: Encountered error while reading extension file 'protobuf.bzl': no such package '@com_google_protobuf\/\/': Traceback (most recent call last):\n&gt;         File \"C:\/temp\/tensorflow-master\/third_party\/repo.bzl\", line 104\n&gt;                 _apply_patch(ctx, ctx.attr.patch_file)\n&gt;         File \"C:\/temp\/tensorflow-master\/third_party\/repo.bzl\", line 70, in _apply_patch\n&gt;                 _wrap_bash_cmd(ctx, patch_command)\n&gt;         File \"C:\/temp\/tensorflow-master\/third_party\/repo.bzl\", line 28, in _wrap_bash_cmd\n&gt;                 fail(\"BAZEL_SH environment variable i...\")\n&gt; BAZEL_SH environment variable is not set\n&gt; WARNING: Target pattern parsing failed.\n&gt; ERROR: error loading package 'tensorflow': in C:\/temp\/tensorflow-master\/tensorflow\/core\/platform\/default\/build_config.bzl: Encountered error while reading extension file 'protobuf.bzl': no such package '@com_google_protobuf\/\/': Traceback (most recent call last):\n&gt;         File \"C:\/temp\/tensorflow-master\/third_party\/repo.bzl\", line 104\n&gt;                 _apply_patch(ctx, ctx.attr.patch_file)\n&gt;         File \"C:\/temp\/tensorflow-master\/third_party\/repo.bzl\", line 70, in _apply_patch\n&gt;                 _wrap_bash_cmd(ctx, patch_command)\n&gt;         File \"C:\/temp\/tensorflow-master\/third_party\/repo.bzl\", line 28, in _wrap_bash_cmd\n&gt;                 fail(\"BAZEL_SH environment variable i...\")\n&gt; BAZEL_SH environment variable is not set\n&gt; INFO: Elapsed time: 21.498s\n&gt; INFO: 0 processes.\n&gt; FAILED: Build did NOT complete successfully (0 packages loaded)\n&gt;     currently loading: tensorflow\n<\/code><\/pre>\n\n<p>I already added to the top of WORKSPACE file as suggested in one of the issues on github<\/p>\n\n<pre><code>http_archive(\n    name = \"io_bazel_rules_docker\",\n    sha256 = \"87fc6a2b128147a0a3039a2fd0b53cc1f2ed5adb8716f50756544a572999ae9a\",\n    strip_prefix = \"rules_docker-0.8.1\",\n    urls = [\"https:\/\/github.com\/bazelbuild\/rules_docker\/archive\/v0.8.1.tar.gz\"],\n)\n<\/code><\/pre>\n\n<p>How do I get past these errors?<\/p>\n\n<p>And why is there such poor support for C++, as far as I know everyone uses tensorflow with C++ to do inference. I can't imagine all the self-driving cars are running on python<\/p>\n","answers":[{"AnswerId":"57514902","CreationDate":"2019-08-15T19:07:49.633","ParentId":null,"OwnerUserId":"2056201","Title":null,"Body":"<p>Fixed by running command <\/p>\n\n<pre><code>c:\\msys32\\usr\\bin\\sed -ri \"\/^\\W+sha256 = \\\"[^\\\"]+\\\"\\W+$\/d\" tensorflow\/workspace.bzl\n<\/code><\/pre>\n\n<p>Which disables checksum validation completely <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/12979\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/12979<\/a> <\/p>\n\n<p>Then I also needed to do <\/p>\n\n<pre><code>c:\\msys32\\usr\\bin\\pacman.exe -S git patch unzip\n<\/code><\/pre>\n\n<p>and add the BAZEL_SH environment variable <code>c:\\msys32\\usr\\bin\\bash.exe<\/code><\/p>\n"}]}
+{"QuestionId":57503595,"AnswerCount":0,"Tags":"<python-3.x><machine-learning><image-processing><keras><ram>","CreationDate":"2019-08-15T00:34:37.780","AcceptedAnswerId":null,"OwnerUserId":11173832.0,"Title":"which part of the code is using memory? and how do I use 'del' to reset the memory?","Body":"<p>for one epoch for this image process using more than 80 ram memory<br\/>\nI know the 'del' command to delete the memory usage <br\/><\/p>\n\n<blockquote>\n  <p>del dataframe<\/p>\n<\/blockquote>\n\n<p>I want to use 'del' command to free the memory after one epoch<br\/>\nbut not sure which one is actually using the memory for this code <br\/>\nhere is the code <br\/><\/p>\n\n<pre><code>for train_index, valid_index in zip(train_indexes,valid_indexes):\n\n    print(\"cleanup memory\")\n    traindf = df_train.iloc[train_index, :].reset_index()\n    validdf = df_train.iloc[valid_index, :].reset_index()\n\n        if(j &gt;= 1 and j &lt;= 1):\n\n        train_generator = train_datagen.flow_from_dataframe(\n            dataframe=traindf,\n            directory=TRAIN_CROPPED_PATH,\n            x_col='img_file',\n            y_col='class',\n            target_size= (IMAGE_SIZE, IMAGE_SIZE),\n            color_mode='rgb',\n            class_mode='categorical',\n            batch_size=BATCH_SIZE,\n            seed=SEED,\n            shuffle=True\n            )\n\n        valid_generator = valid_datagen.flow_from_dataframe(\n            dataframe=validdf,\n            directory=TRAIN_CROPPED_PATH,\n            x_col='img_file',\n            y_col='class',\n            color_mode='rgb',\n            class_mode='categorical',\n            batch_size=BATCH_SIZE,\n            seed=SEED,\n            shuffle=True\n            )\n\n        model_name = model_path + str(j) + '_'+ modelName+\"_Aug\"+'.hdf5'\n        model_names.append(model_name)\n\n        print(\"TRAIN_CROPPED_PATH:\",TRAIN_CROPPED_PATH)\n        print(\"model_name:\",model_name)\n\n        model = get_model()\n\n        try:\n            model.load_weights(model_name)\n        except:\n            pass\n\n        print(\"model_path:\",model_path)\n\n        patient = 2\n        callbacks = [\n        EarlyStopping(monitor='val_loss', patience=patient, mode='min', verbose=1),\n        ReduceLROnPlateau(monitor = 'val_loss', factor = 0.5, patience = patient \/ 2, min_lr=0.00001, verbose=1, mode='min'),\n        ModelCheckpoint(filepath=model_name, monitor='val_loss', verbose=1, save_best_only=True, mode='min'),\n        ]\n\n        history = model.fit_generator(\n            train_generator,\n            steps_per_epoch=len(traindf.index) \/ BATCH_SIZE,\n            epochs=epochs,\n            validation_data=valid_generator,\n            validation_steps=len(validdf.index) \/ BATCH_SIZE,\n            verbose=1,\n            shuffle=False,\n           callbacks = callbacks\n            )\n\n    j+=1\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57503692,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><loss-function>","CreationDate":"2019-08-15T00:55:13.280","AcceptedAnswerId":58065761.0,"OwnerUserId":1009508.0,"Title":"Triple loss in keras, how to get the anchor, positive, and negative from merged vector","Body":"<p>What I am trying to do is use the triple loss as my loss function, but I don't know if I am getting the right values from the merged vector that is used.<\/p>\n\n<p>So here is my loss function: <\/p>\n\n<pre><code>def triplet_loss(y_true, y_pred, alpha=0.2):\n    \"\"\"\n    Implementation of the triplet loss function\n    Arguments:\n    y_true -- true labels, required when you define a loss in Keras, not used in this function.\n    y_pred -- python list containing three objects:\n            anchor:   the encodings for the anchor data\n            positive: the encodings for the positive data (similar to anchor)\n            negative: the encodings for the negative data (different from anchor)\n    Returns:\n    loss -- real number, value of the loss\n    \"\"\"\n    print(\"Ypred\")\n    print(y_pred.shape)\n\n    anchor = y_pred[:,0:512]\n    positive = y_pred[:,512:1024]\n    negative = y_pred[:,1024:1536]\n\n    print(anchor.shape)\n    print(positive.shape)\n    print(negative.shape)\n\n    #anchor, positive, negative = y_pred[0], y_pred[1], y_pred[2] # Dont think this is working\n    # distance between the anchor and the positive\n    pos_dist = tf.reduce_sum(tf.square(tf.subtract(anchor, positive)))\n    print(\"PosDist\", pos_dist)\n    # distance between the anchor and the negative\n    neg_dist = tf.reduce_sum(tf.square(tf.subtract(anchor, negative)))\n    print(\"Neg Dist\", neg_dist)\n    # compute loss\n    basic_loss = (pos_dist - neg_dist) + alpha\n    loss = tf.maximum(basic_loss, 0.0)\n    return loss\n<\/code><\/pre>\n\n<p>Now this does work when I use this line in the code and nother the sliceing one<\/p>\n\n<pre><code>anchor, positive, negative = y_pred[0], y_pred[1], y_pred[2] \n<\/code><\/pre>\n\n<p>But I dont think that this is correct as the shape of the merged vector is (?, 3, 3, 1536)\nI think it is grabbing the wrong information. But I cannot seem to figure out how to slice this correctly. as the uncommented code gives me this issue. <\/p>\n\n<blockquote>\n  <p>Dimensions must be equal, but are 3 and 0 for 'loss_9\/concatenate_10_loss\/Sub' (op: 'Sub') with input shapes: [?,3,3,1536], [?,0,3,1536].<\/p>\n<\/blockquote>\n\n<p>My network set up is like this:<\/p>\n\n<pre><code>input_dim = (7,7,2048)\nanchor_in = Input(shape=input_dim)\npos_in = Input(shape=input_dim)\nneg_in = Input(shape=input_dim)\nbase_network = create_base_network()\n# Run input through base network\nanchor_out = base_network(anchor_in)\npos_out = base_network(pos_in)\nneg_out = base_network(neg_in)\nprint(anchor_out.shape)\n\nmerged_vector = Concatenate(axis=-1)([anchor_out, pos_out, neg_out])\nprint(\"Meged Vector\", merged_vector.shape)\nprint(merged_vector)\n\nmodel = Model(inputs=[anchor_in, pos_in, neg_in], outputs=merged_vector)\n\nadam = Adam(lr=0.01, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False)\nmodel.compile(optimizer=adam, loss=triplet_loss)\n<\/code><\/pre>\n\n<h2>Update<\/h2>\n\n<p>Using this seems to be right, could anyone confirm this?<\/p>\n\n<pre><code>anchor = y_pred[:,:,:,0:512]\npositive = y_pred[:,:,:,512:1024]\nnegative = y_pred[:,:,:,1024:1536]\n<\/code><\/pre>\n","answers":[{"AnswerId":"58065761","CreationDate":"2019-09-23T15:33:28.667","ParentId":null,"OwnerUserId":"11952444","Title":null,"Body":"<p>You do not need to do the concatenation operation:<\/p>\n\n<pre><code># change this line to this\nmodel = Model(inputs=[anchor_in, pos_in, neg_in], outputs=[anchor_out, pos_out, neg_out])\n<\/code><\/pre>\n\n<p>Complete code:<\/p>\n\n<pre><code>input_dim = (7,7,2048)\nanchor_in = Input(shape=input_dim)\npos_in = Input(shape=input_dim)\nneg_in = Input(shape=input_dim)\nbase_network = create_base_network()\n# Run input through base network\nanchor_out = base_network(anchor_in)\npos_out = base_network(pos_in)\nneg_out = base_network(neg_in)\nprint(anchor_out.shape)\n\n# code changed here\nmodel = Model(inputs=[anchor_in, pos_in, neg_in], outputs=[anchor_out, pos_out, neg_out])\nadam = Adam(lr=0.01, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False)\nmodel.compile(optimizer=adam, loss=triplet_loss) \n<\/code><\/pre>\n\n<p>Then you can use the following loss:<\/p>\n\n<pre><code>def triplet_loss(y_true, y_pred, alpha=0.3):\n    '''\n        Inputs:\n                    y_true: True values of classification. (y_train)\n                    y_pred: predicted values of classification.\n                    alpha: Distance between positive and negative sample, arbitrarily\n                           set to 0.3\n        Returns:\n                    Computed loss\n        Function:\n                    --Implements triplet loss using tensorflow commands\n                    --The following function follows an implementation of Triplet-Loss \n                      where the loss is applied to the network in the compile statement \n                      as usual.\n    '''\n    anchor, positive, negative = y_pred[0], y_pred[1], y_pred[2]\n\n    positive_dist = tf.reduce_sum(tf.square(tf.subtract(anchor, positive)), -1)\n    negative_dist = tf.reduce_sum(tf.square(tf.subtract(anchor,negative)), -1)\n\n    loss_1 = tf.add(tf.subtract(positive_dist, negative_dist), alpha)\n    loss = tf.reduce_sum(tf.maximum(loss_1, 0.0))\n\n    return loss\n<\/code><\/pre>\n"}]}
+{"QuestionId":57503705,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras><keras-layer><dropout>","CreationDate":"2019-08-15T00:57:28.817","AcceptedAnswerId":57509119.0,"OwnerUserId":11867828.0,"Title":"Is dropout layer still active in a freezed Keras model (i.e. trainable=False)?","Body":"<p>I have two trained models (<code>model_A<\/code> and <code>model_B<\/code>), and both of them have dropout layers. I have freezed <code>model_A<\/code> and <code>model_B<\/code> and merged them with a new dense layer to get <code>model_AB<\/code> (but I have not removed <code>model_A<\/code>'s and <code>model_B<\/code>'s dropout layers). <code>model_AB<\/code>'s weights will be non-trainable, except for the added dense layer. <\/p>\n\n<p>Now my question is: are the dropout layers in <code>model_A<\/code> and <code>model_B<\/code> active (i.e. drop neurons) when I am training <code>model_AB<\/code>?<\/p>\n","answers":[{"AnswerId":"57509119","CreationDate":"2019-08-15T11:42:14.730","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p><strong>Short answer:<\/strong> The dropout layers will continue dropping neurons <strong>during training<\/strong>, even if you set their <code>trainable<\/code> property to <code>False<\/code>.<\/p>\n\n<p><strong>Long answer:<\/strong> There are two distinct notions in Keras:<\/p>\n\n<ul>\n<li><p><strong>Updating the weights and states of a layer:<\/strong> this is controlled using <code>trainable<\/code> property of that layer, i.e. if you set <code>layer.trainable = False<\/code> then the weights and internal states of the layer would not be updated.<\/p><\/li>\n<li><p><strong>Behavior of a layer in training and testing phases:<\/strong> as you know a layer, like dropout, may have a different behavior during training and testing phases. Learning phase in Keras is set using <code>keras.backend.set_learning_phase()<\/code>. For example, when you call <code>model.fit(...)<\/code> the learning phase is automatically set to 1 (i.e. training), whereas when you use <code>model.predict(...)<\/code> it will be automatically set to 0 (i.e. test). Further, note that learning phase of 1 (i.e. training) does not necessarily imply updating the weighs\/states of a layer. You can run your model with a learning phase of 1 (i.e. training phase), but no weights will be updated; just layers will switch to their training behavior (see <a href=\"https:\/\/stackoverflow.com\/a\/57439143\/2099607\">this answer<\/a> for more information). Further, there is another way to set learning phase for each individual layer by passing <code>training=True<\/code> argument when calling a layer on a tensor (see <a href=\"https:\/\/stackoverflow.com\/a\/52760112\/2099607\">this answer<\/a> for more information).<\/p><\/li>\n<\/ul>\n\n<p>So according to the above points, when you set <code>trainable=False<\/code> on a dropout layer and use that in training mode (e.g. either by calling <code>model.fit(...)<\/code>, or manually setting learning phase to training like example below), the neurons would still be dropped by the dropout layer.<\/p>\n\n<p>Here is a reproducible example which illustrates this point:\n<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras import layers\nfrom keras import models\nfrom keras import backend as K\nimport numpy as np\n\ninp = layers.Input(shape=(10,))\nout = layers.Dropout(0.5)(inp)\n\nmodel = models.Model(inp, out)\nmodel.layers[-1].trainable = False  # set dropout layer as non-trainable\nmodel.compile(optimizer='adam', loss='mse') # IMPORTANT: we must always compile model after changing `trainable` attribute\n\n# create a custom backend function so that we can control the learning phase\nfunc = K.function(model.inputs + [K.learning_phase()], model.outputs)\n\nx = np.ones((1,10))\n# learning phase = 1, i.e. training mode\nprint(func([x, 1]))\n# the output will be:\n[array([[2., 2., 2., 0., 0., 2., 2., 2., 0., 0.]], dtype=float32)]\n# as you can see some of the neurons have been dropped\n\n# now set learning phase = 0, i.e test mode\nprint(func([x, 0]))\n# the output will be:\n[array([[1., 1., 1., 1., 1., 1., 1., 1., 1., 1.]], dtype=float32)]\n# unsurprisingly, no neurons have been dropped in test phase\n<\/code><\/pre>\n"}]}
+{"QuestionId":57503726,"AnswerCount":0,"Tags":"<keras><loss-function>","CreationDate":"2019-08-15T00:59:53.213","AcceptedAnswerId":null,"OwnerUserId":6450489.0,"Title":"weighted_masked_objective in keras","Body":"<p>According to <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/58fd1f0589d33aeb33c4129cfedfb7737495efc0\/keras\/engine\/training.py#L309\" rel=\"nofollow noreferrer\">Keras Code<\/a>, Keras computes the loss value considering optional weights and masks.<\/p>\n\n<pre><code>       for i in range(len(self.outputs)):\n        if i in skip_target_indices:\n            continue\n        y_true = self.targets[i]\n        y_pred = self.outputs[i]\n        weighted_loss = weighted_losses[i]\n        sample_weight = sample_weights[i]\n        mask = masks[i]\n        loss_weight = loss_weights_list[i]\n        with K.name_scope(self.output_names[i] + '_loss'):\n            output_loss = weighted_loss(y_true, y_pred,\n                                        sample_weight, mask)\n        if len(self.outputs) &gt; 1:\n            self.metrics_tensors.append(output_loss)\n            self.metrics_names.append(self.output_names[i] + '_loss')\n        if total_loss is None:\n            total_loss = loss_weight * output_loss\n        else:\n            total_loss += loss_weight * output_loss\n<\/code><\/pre>\n\n<p>On the other hand, in <a href=\"https:\/\/keras.io\/models\/model\/\" rel=\"nofollow noreferrer\">Keras documentation<\/a> I see the basic loss function is introduced in compile function, and then sample or class weights can be introduced in fit command.\nI am not sure how to relate 'weights and masks' in the first code to 'sample and class weights' in the second document. Can anybody give me more explanation?<\/p>\n\n<p>My application is actually a Convolutional LSTM network, where I input a series of images and want the network to produce an output map (with the same size of input maps) of pixel classes, but some pixels don't have valid labels during training. Should I use weight or mask, sample or class?     <\/p>\n","answers":[]}
+{"QuestionId":57503803,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-08-15T01:17:50.867","AcceptedAnswerId":null,"OwnerUserId":11929788.0,"Title":"training exit after the first evaluation(only one evaluation) in Tensorflow model(object_detection) without error and waring","Body":"<p>training exit after the first evaluation(only one evaluation) in Tensorflow model(object_detection) without error and waring<\/p>\n\n<p>System information<\/p>\n\n<pre><code>What is the top-level directory of the model you are using:models\/research\/object_detection\/\nHave I written custom code (as opposed to using a stock example script provided in TensorFlow):NO\nOS Platform and Distribution (e.g., Linux Ubuntu 16.04):Windows-10(64bit)\nTensorFlow installed from (source or binary):conda install tensorflow-gpu\nTensorFlow version (use command below):1.13.1\nBazel version (if compiling from source):N\/A\nCUDA\/cuDNN version:cudnn-7.6.0\nGPU model and memory:GeForce GTX 1060 6GB\nExact command to reproduce:See below\n<\/code><\/pre>\n\n<p>my command for training :<\/p>\n\n<pre><code>python object_detection\/model_main.py --pipeline_config_path=D:\/gitcode\/models\/research\/object_detection\/ssd_mobilenet_v1_coco_2018_01_28\/pipeline.config --model_dir=D:\/gitcode\/models\/research\/object_detection\/ssd_mobilenet_v1_coco_2018_01_28\/saved_model --num_train_steps=50000 --alsologtostderr\n<\/code><\/pre>\n\n<p>This is my config :<\/p>\n\n<pre><code>train_config {\nbatch_size: 24\ndata_augmentation_options {\nrandom_horizontal_flip {\n}\n}\ndata_augmentation_options {\nssd_random_crop {\n}\n}\noptimizer {\nrms_prop_optimizer {\nlearning_rate {\nexponential_decay_learning_rate {\ninitial_learning_rate: 0.00400000018999\ndecay_steps: 800720\ndecay_factor: 0.949999988079\n}\n}\nmomentum_optimizer_value: 0.899999976158\ndecay: 0.899999976158\nepsilon: 1.0\n}\n}\nfine_tune_checkpoint: \"D:\/gitcode\/models\/research\/object_detection\/ssd_mobilenet_v1_coco_2018_01_28\/model.ckpt\"\nfrom_detection_checkpoint: true\nnum_steps: 200000\n\ntrain_input_reader {\nlabel_map_path: \"D:\/gitcode\/models\/research\/object_detection\/idol\/tf_label_map.pbtxt\"\ntf_record_input_reader {\ninput_path: \"D:\/gitcode\/models\/research\/object_detection\/idol\/train\/Iframe_??????.tfrecord\"\n}\n}\neval_config {\nnum_examples: 8000\nmax_evals: 10\nuse_moving_averages: false\n}\neval_input_reader {\nlabel_map_path: \"D:\/gitcode\/models\/research\/object_detection\/idol\/tf_label_map.pbtxt\"\nshuffle: false\nnum_readers: 1\ntf_record_input_reader {\ninput_path: \"D:\/gitcode\/models\/research\/object_detection\/idol\/eval\/Iframe_??????.tfrecord\"\n}\n<\/code><\/pre>\n\n<p>This is my out log:<\/p>\n\n<pre><code>(default) D:\\gitcode\\models\\research&gt;python object_detection\/model_main.py --pipeline_config_path=D:\/gitcode\/models\/research\/object_detection\/ssd_mobilenet_v1_coco_2018_01_28\/pipeline.config --model_dir=D:\/gitcode\/models\/research\/object_detection\/ssd_mobilenet_v1_coco_2018_01_28\/saved_model --num_train_steps=50000 --alsologtostderr\n\nWARNING: The TensorFlow contrib module will not be included in TensorFlow 2.0.\nFor more information, please see:\n\nhttps:\/\/github.com\/tensorflow\/community\/blob\/master\/rfcs\/20180907-contrib-sunset.md\nhttps:\/\/github.com\/tensorflow\/addons\nIf you depend on functionality not listed there, please file an issue.\nWARNING:tensorflow: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.\nWARNING:tensorflow:Estimator's model_fn (&lt;function create_model_fn..model_fn at 0x0000027CBAB7BB70&gt;) includes params argument, but params are not passed to Estimator.\nWARNING:tensorflow:From C:\\Users\\qian\\Anaconda3\\envs\\default\\lib\\site-packages\\tensorflow\\python\\framework\\op_def_library.py:263: colocate_with (from tensorflow.python.framework.ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nColocations handled automatically by placer.\nWARNING:tensorflow:From C:\\Users\\qian\\Anaconda3\\envs\\default\\lib\\site-packages\\object_detection-0.1-py3.7.egg\\object_detection\\builders\\dataset_builder.py:86: 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 C:\\Users\\qian\\Anaconda3\\envs\\default\\lib\\site-packages\\object_detection-0.1-py3.7.egg\\object_detection\\core\\preprocessor.py:196: sample_distorted_bounding_box (from tensorflow.python.ops.image_ops_impl) is deprecated and will be removed in a future version.\nInstructions for updating:\nseed2 arg is deprecated.Use sample_distorted_bounding_box_v2 instead.\nWARNING:tensorflow:From C:\\Users\\qian\\Anaconda3\\envs\\default\\lib\\site-packages\\object_detection-0.1-py3.7.egg\\object_detection\\builders\\dataset_builder.py:158: 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).\nWARNING:tensorflow:From C:\\Users\\qian\\Anaconda3\\envs\\default\\lib\\site-packages\\tensorflow\\python\\ops\\losses\\losses_impl.py:448: to_float (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse tf.cast instead.\nWARNING:tensorflow:From C:\\Users\\qian\\Anaconda3\\envs\\default\\lib\\site-packages\\tensorflow\\python\\ops\\array_grad.py:425: to_int32 (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse tf.cast instead.\n2019-08-14 16:29:31.607841: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1433] Found device 0 with properties:\nname: GeForce GTX 1060 6GB major: 6 minor: 1 memoryClockRate(GHz): 1.7845\npciBusID: 0000:04:00.0\ntotalMemory: 6.00GiB freeMemory: 4.97GiB\n2019-08-14 16:29:31.621836: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1512] Adding visible gpu devices: 0\n2019-08-14 16:29:32.275712: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:984] Device interconnect StreamExecutor with strength 1 edge matrix:\n2019-08-14 16:29:32.283072: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:990] 0\n2019-08-14 16:29:32.288675: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1003] 0: N\n2019-08-14 16:29:32.293514: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 4714 MB memory) -&gt; physical GPU (device: 0, name: GeForce GTX 1060 6GB, pci bus id: 0000:04:00.0, compute capability: 6.1)\nWARNING:tensorflow:From C:\\Users\\qian\\Anaconda3\\envs\\default\\lib\\site-packages\\object_detection-0.1-py3.7.egg\\object_detection\\eval_util.py:796: to_int64 (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse tf.cast instead.\nWARNING:tensorflow:From C:\\Users\\qian\\Anaconda3\\envs\\default\\lib\\site-packages\\object_detection-0.1-py3.7.egg\\object_detection\\utils\\visualization_utils.py:498: py_func (from tensorflow.python.ops.script_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\ntf.py_func is deprecated in TF V2. Instead, use\ntf.py_function, which takes a python function which manipulates tf eager\ntensors instead of numpy arrays. It's easy to convert a tf eager tensor to\nan ndarray (just call tensor.numpy()) but having access to eager tensors\nmeans tf.py_functions can use accelerators such as GPUs as well as\nbeing differentiable using a gradient tape.\n\n2019-08-14 16:41:44.736212: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1512] Adding visible gpu devices: 0\n2019-08-14 16:41:44.741242: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:984] Device interconnect StreamExecutor with strength 1 edge matrix:\n2019-08-14 16:41:44.747522: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:990] 0\n2019-08-14 16:41:44.751256: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1003] 0: N\n2019-08-14 16:41:44.755548: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 4714 MB memory) -&gt; physical GPU (device: 0, name: GeForce GTX 1060 6GB, pci bus id: 0000:04:00.0, compute capability: 6.1)\nWARNING:tensorflow:From C:\\Users\\qian\\Anaconda3\\envs\\default\\lib\\site-packages\\tensorflow\\python\\training\\saver.py:1266: checkpoint_exists (from tensorflow.python.training.checkpoint_management) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse standard file APIs to check for files with this prefix.\ncreating index...\nindex created!\ncreating index...\nindex created!\nRunning per image evaluation...\nEvaluate annotation type bbox\nDONE (t=2.43s).\nAccumulating evaluation results...\nDONE (t=0.14s).\nAverage Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.287\nAverage Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.529\nAverage Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.278\nAverage Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\nAverage Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.031\nAverage Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.312\nAverage Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.162\nAverage Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.356\nAverage Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.356\nAverage Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\nAverage Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.061\nAverage Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.384\n(default) D:\\gitcode\\models\\research&gt;\n<\/code><\/pre>\n\n<p>~<\/p>\n","answers":[]}
+{"QuestionId":57503882,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-15T01:32:46.963","AcceptedAnswerId":null,"OwnerUserId":8214951.0,"Title":"Tensorflow ImportError: DLL load failed","Body":"<p>I've just used pip to install Tensorflow on my computer, which was successful (Win10). When I try to import Tensorflow in my code, it gives the following error messages.<\/p>\n\n<p>Attempted to revert to tensorflow v1.5 as per other responses.<\/p>\n\n<p>Python code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow\n<\/code><\/pre>\n\n<p>Error:<\/p>\n\n<pre><code>  File \"code-basics.py\", line 1, in &lt;module&gt;\n    import tensorflow\n  File \"C:\\Users\\[user]\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\__init__.py\", line 24, in &lt;module&gt;\n    from tensorflow.python import *\n  File \"C:\\Users\\[user]\\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\\[user]\\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\\[user]\\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","answers":[]}
+{"QuestionId":57504148,"AnswerCount":0,"Tags":"<python><tensorflow><keras><google-cloud-platform><tpu>","CreationDate":"2019-08-15T02:21:58.037","AcceptedAnswerId":null,"OwnerUserId":3128562.0,"Title":"Keras with Tensorflow not finding TPU","Body":"<p>I'm trying to run <a href=\"https:\/\/github.com\/golsun\/SpaceFusion\" rel=\"nofollow noreferrer\">this Keras project<\/a> with Tensorflow and a TPUv3 on GCP. However Keras doesn't find my tpu at all: \"No OpKernel was registered to support Op 'CrossReplicaSum'\" and registered devices are only CPU and XLA_CPU, but no TPU. There is not kernel registered.<\/p>\n\n<p>However the TPU is listed as running.<\/p>\n\n<p>I've added the following lines before the <code>model.compile(...)<\/code> part of the project:<\/p>\n\n<pre><code>tpu=[os.environ['TPU_NAME']]\nmy_project_name = 'PROJECT_NAME'\nmy_zone = 'us-central1-a'resolver = tf.contrib.cluster_resolver.TPUClusterResolver(\n                        tpu=tpu,\n                        zone=my_zone,\n                        project=my_project_name)\n\ntf.contrib.distribute.initialize_tpu_system(resolver)\nglobal strategy\nstrategy = tf.contrib.distribute.TPUStrategy(resolver)\n<\/code><\/pre>\n\n<p>And in another file:<\/p>\n\n<pre><code>with main.strategy.scope():\n      opti = tf.train.AdamOptimizer(learning_rate=self.lr)\n      model.compile(optimizer=tf.contrib.tpu.CrossShardOptimizer(opti), loss=loss, loss_weights=loss_weights)\n<\/code><\/pre>\n\n<p>So I'm guessing that the \"No OpKernel was registered\" error is expected, since it's not running on the TPU. Question is: how do I get it to run on the TPU.<\/p>\n","answers":[]}
+{"QuestionId":57504262,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-08-15T02:39:59.380","AcceptedAnswerId":null,"OwnerUserId":11929946.0,"Title":"How to customize pytorch data","Body":"<p>I am trying to make a customized Dataloader using pytorch.<\/p>\n\n<p>I've seen some codes like (omitted the class sorry.)<\/p>\n\n<pre><code>def __init__(self, data_root, transform=None, training=True, return_id=False):\n    super().__init__()\n    self.mode = 'train' if training else 'test'\n\n    self.data_root = Path(data_root)\n    csv_fname = 'train.csv' if training else 'sample_submission.csv'\n    self.csv_file = pd.read_csv(self.data_root \/ csv_fname)\n    self.transform = transform\n    self.return_id = return_id\ndef __getitem__():\n    \"\"\" TODO\n    \"\"\"\ndef __len__():\n    \"\"\" TODO\n    \"\"\"\n<\/code><\/pre>\n\n<p>The problem here is that the datas I've dealt with before contains all the training data in one csv file, and all the testing data in the other csv file, with total 2 csv files for training and testing. (For example like in MNIST, the last column is the labeling, and the all the previous columns are each different features.)<\/p>\n\n<p>However, the problem I've been facing is that I've got very many, (about 200,000) csv files for training, each one smaller than 60,000 sized MNIST, but still quite big. All these csv files contain different number of rows.<\/p>\n\n<p>To inherit torch.util.data, how can I make customized class? MNIST dataset is quite small, so can be uploaded on RAM at once. However, the data I'm dealing with is super big, so I need some help.<\/p>\n\n<p>Any ideas? Thank you in advance.<\/p>\n","answers":[{"AnswerId":"57504871","CreationDate":"2019-08-15T04:28:35.737","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>First, you want to customize (overload) <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/data.html#torch.utils.data.Dataset\" rel=\"nofollow noreferrer\"><code>data.Dataset<\/code><\/a> and not <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/data.html#torch.utils.data.DataLoader\" rel=\"nofollow noreferrer\"><code>data.DataLoader<\/code><\/a> which is perfectly fine for your use case.<\/p>\n\n<p>What you can do, instead of loading all data to RAM, is to read and store \"meta data\" on <code>__init__<\/code> and read one relevant csv file whnever you need to <code>__getitem__<\/code> a specific entry.<br>\nA pseudo-code of your <code>Dataset<\/code> will look something like:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class ManyCSVsDataset(data.Dataset):\n  def __init__(self, ...):\n    super(ManyCSVsDataset, self).__init__()\n    # store the paths for all csvs and the number of items in each one\n    self.metadata = ... \n    self.num_items = total_number_of_items\n\n  def __len__(self):\n    return self.num_items\n\n  def __getitem__(self, index):\n    # based on the index, use self.metadata to determine what csv file to open\n    with open(relevant_csv_file, 'r') as R:\n      # read from R the specific line matching item index\n    return item\n<\/code><\/pre>\n\n<p>This implementation is not efficient in the sense that it reads the same csv file over and over and does not cache anything. On the other hand, you can take advantage of <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/data.html#torch.utils.data.DataLoader\" rel=\"nofollow noreferrer\"><code>data.DataLoader<\/code><\/a>'s multi processing support to have many parallel sub-processes doing all these file access at the background while you actually use the data for training.<\/p>\n"}]}
+{"QuestionId":57504283,"AnswerCount":0,"Tags":"<tensorflow><cuda><eigen>","CreationDate":"2019-08-15T02:43:06.763","AcceptedAnswerId":null,"OwnerUserId":11789973.0,"Title":"CUDA kernel can't support 8*8 eigen matrix assignment?","Body":"<p>I'm trying to develop a Tensorflow custom op using cuda. And inside the op, I need to use Eigen to run some matrix operations, but when I try to calculate the inverse of a size 8*8 matrix, I've encountered several problems.<\/p>\n\n<p>I'm currently using Eigen library provided by Tensorflow, and try to compile my custom op with bazel 0.21.0.<\/p>\n\n<p>This function is written in CUDA kernel.<\/p>\n\n<p>In cuda kernel function file (.cu.cc), I have included<\/p>\n\n<pre><code>#include &lt;Eigen\\Dense&gt;\n#include &lt;Eigen\/LU&gt;\n#include \"third_party\/eigen3\/unsupported\/Eigen\/CXX11\/Tensor\"\n#include \"tensorflow\/core\/framework\/op.h\"\n#include \"tensorflow\/core\/framework\/op_kernel.h\"\n#include &lt;stdio.h&gt;\n#include &lt;math.h&gt;\n<\/code><\/pre>\n\n<p>And in the kernel function, I just tried a simple test:<\/p>\n\n<pre><code>typedef Eigen::Matrix&lt;float, 8, 8&gt; Matrix8f;\n\n__global__ void CsmToRcmKernel(Eigen::Map&lt;const Eigen::MatrixXcf&gt; csm_map, Eigen::Map&lt;const Eigen::VectorXf&gt; reg_vec,\n    Eigen::Map&lt;const MatrixXi16&gt; offset8,       \/\/ (12160, 8, 3)\n    Eigen::Map&lt;const MatrixXi16&gt; offset3,       \/\/ (9728, 3, 3)\n    Eigen::Map&lt;const MatrixXi16&gt; offset4,       \/\/ (12160, 4, 3)\n    Eigen::Map&lt;const MatrixXi16&gt; offset6,       \/\/ (9728, 6, 3)\n    Eigen::Map&lt;Eigen::MatrixXcf&gt; rcm_map) {\n\n    Matrix8f A = MatrixXf::Identity(8, 8);\n    A(0, 0) = 2;\n\n    \/\/ inv_A = inv(A);\n    Matrix8f inv_A = A.inverse();\n}\n<\/code><\/pre>\n\n<p>(You can ignore the input parameters, they aren't used in this test program.)<\/p>\n\n<p>And when I try to build this op with bazel, I got following errors:<\/p>\n\n<pre><code>c:\\users\\administrator\\_bazel_zhaozixiao\\sr6amwum\\execroot\\org_tensorflow\\external\\eigen_archive\\eigen\\src\/Core\/SolveTriangular.h(185): \nerror: function \"Eigen::Block&lt;XprType, BlockRows, BlockCols, InnerPanel&gt;::operator=(const Eigen::Block&lt;Eigen::Block&lt;Eigen::Map&lt;Eigen::Matrix&lt;float, -1, -1, 0, -1, -1&gt;, 0, Eigen::Stride&lt;0, 0&gt;&gt;, -1, -1, false&gt;, -1, -1, false&gt; &amp;) [with XprType=Eigen::Block&lt;Eigen::Map&lt;Eigen::Matrix&lt;float, -1, -1, 0, -1, -1&gt;, 0,Eigen::Stride&lt;0, 0&gt;&gt;, -1, -1, false&gt;, BlockRows=-1, BlockCols=-1, InnerPanel=false]\" (declared implicitly) cannot be referenced -- it is a deleted function\n\nc:\\users\\administrator\\_bazel_zhaozixiao\\sr6amwum\\execroot\\org_tensorflow\\external\\eigen_archive\\eigen\\src\/Core\/GenericPacketMath.h(368): \nerror: asm operand type size(8) does not match type\/size implied by constraint 'r'\n<\/code><\/pre>\n\n<p>And if I try calculate the inverse of a size 4*4(or below) matrix, there isn't any problem. It seems that the calculation of inverse is correct, but the result cannot be assigned to another matrix. So i'd like to know if cuda kernel supports big size(bigger than 4*4) eigen matrix operations or assignments? Thank you!<\/p>\n","answers":[]}
+{"QuestionId":57504286,"AnswerCount":0,"Tags":"<python><keras><hdf5>","CreationDate":"2019-08-15T02:43:39.460","AcceptedAnswerId":null,"OwnerUserId":10472252.0,"Title":"How to load multiple objects (list of list) in HDF5?","Body":"<p>I am currently trying to use keras to predict ... stuffs. I am using a HDF5  file as input. The file contains 195 objects, each one is a list of matrices with one attribute. I would like keras to learn on the list of matrices and predict the attribute. But here is the issue, so far I have seen that one object can only be assigned to one variable. That would be meaningless in my case.                                                           <\/p>\n\n<p>I would like to know whether or not it was possible to load all of these objects at once, say under one variable, in keras to predict the attribute ? For instance here are some objects, <\/p>\n\n<pre><code>['10gs', '1a30', '1bcu',..., '4tmn']\n<\/code><\/pre>\n\n<p>I know I can assign one variable to one object, <\/p>\n\n<pre><code>dataset=infile['1a30'] \n<\/code><\/pre>\n\n<p>However I am not sure how to assign several objects to one variable? Do I need to create a list of objects ? Here's what I am trying to get, <\/p>\n\n<pre><code>dataset=infile['all of my objects'].\n<\/code><\/pre>\n\n<p>In fine, I will be using it in keras but I am not too sure whether it is necessary as it seems to me it is a HDF file issue (misunderstanding).<\/p>\n","answers":[]}
+{"QuestionId":57504316,"AnswerCount":0,"Tags":"<tensorflow><toco>","CreationDate":"2019-08-15T02:49:33.407","AcceptedAnswerId":null,"OwnerUserId":11929953.0,"Title":"TOCO convert problem:how to calculate output minmax of concatenation layers?","Body":"<p>I notice that tensorflow source code in <code>quantized_concat_op.cc<\/code> use the function of <code>CalculateInputAndOutputRange()<\/code> to caculate the min-max value of input and output feature maps, which just let the output min value be the minimum of the <code>input_mins[i]<\/code>, and max value be the maximum of the <code>input_maxes[i]<\/code>.<\/p>\n\n<p>But when i use TOCO convert pb to tflite, i found that the output min value is not the minimum of the input arrays.<\/p>\n\n<p>I have searched so much, still have no idea how to determine the concat output min-max. I would appreciate it, if someone can help me.<\/p>\n","answers":[]}
+{"QuestionId":57504401,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-08-15T03:03:46.443","AcceptedAnswerId":null,"OwnerUserId":8235645.0,"Title":"what's the meaning of 'input_length'?","Body":"<p>the data have 4 timestamps,but the embedding's input_length=3,so what's the meaning of input_length?<\/p>\n\n<pre><code>from tensorflow import keras\nimport numpy as np\ndata = np.array([[0,0,0,0]])\nemb = keras.layers.Embedding(input_dim=2, output_dim=3, input_length=3)\nemb(data)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57507414","CreationDate":"2019-08-15T09:13:50.467","ParentId":null,"OwnerUserId":"5810950","Title":null,"Body":"<p>As per the official documentation <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/Embedding\" rel=\"nofollow noreferrer\">here<\/a>, <\/p>\n\n<blockquote>\n  <p><strong>input_length<\/strong>: Length of input sequences, when it is constant. This\n  argument is required if you are going to connect Flatten then Dense\n  layers upstream (without it, the shape of the dense outputs cannot be\n  computed).<\/p>\n<\/blockquote>\n\n<pre><code>from tensorflow import keras\nimport numpy as np\n\nmodel = keras.models.Sequential()\nmodel.add(keras.layers.Embedding(input_dim=2, output_dim=3, input_length=4))\n# the model will take as input an integer matrix of size (batch, input_length).\n\ninput_array = np.array([[0,0,0,0]])\nmodel.compile('rmsprop', 'mse')\noutput_array = model.predict(input_array)\nprint(output_array)\n<\/code><\/pre>\n\n<p>Above works fine, but if you change <code>input_length<\/code> to 3, then you will get below error:<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking input: expected embedding_input to\n  have shape (3,) but got array with shape (4,)<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":57504746,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><gpu><amd>","CreationDate":"2019-08-15T04:06:41.070","AcceptedAnswerId":57529064.0,"OwnerUserId":7981821.0,"Title":"Setting up keras and tensoflow to operate with AMD GPU","Body":"<p>I am trying to set up Keras in order to run models using my GPU. I have a Radeon RX580 and am running Windows 10.<\/p>\n\n<p>I saw realized that CUDA only supports NVIDIA GPUs and was having difficulty finding a way to get my code to run on the GPU. I tried downloading and setting up plaidml but afterwards <code>from tensorflow.python.client import device_lib\nprint(device_lib.list_local_devices())<\/code> \nonly printed that I was running on a CPU and there was not a GPU available even though the plaidml setup was a success. I have read that PyOpenCl is needed but have not gotten a  clear answer as to why or to what capacity. Does anyone know how to set up this AMD GPU to work properly? any help would be much appreciated. Thank you!<\/p>\n","answers":[{"AnswerId":"57529064","CreationDate":"2019-08-16T17:50:38.407","ParentId":null,"OwnerUserId":"7981821","Title":null,"Body":"<p>To the best of my knowledge, PlaidML was not working because I did not have the required prerequisites such as OpenCL. Once I downloaded the Visual Studio C++ build tools in order to install PyopenCL from a .whl file. This seemed to resolve the issue<\/p>\n"}]}
+{"QuestionId":57506352,"AnswerCount":2,"Tags":"<tensorflow><conv-neural-network><python-3.6>","CreationDate":"2019-08-15T07:43:12.817","AcceptedAnswerId":57507104.0,"OwnerUserId":10445743.0,"Title":"I am creating a CNN function with TensorFlow, but I get a Shape related error","Body":"<p>I tried Convolution Neural Network with Tensorflow.<\/p>\n\n<p>However, Shape causes an error.<\/p>\n\n<p>First is part of the main function.<\/p>\n\n<pre><code>while True:\n        with mss.mss() as sct:\n                Game_Scr = np.array(sct.grab(Game_Scr_pos))[:,:,:3]\n\n                cv2.imshow('Game_Src', Game_Scr)\n                cv2.waitKey(0)\n\n                Game_Scr = cv2.resize(Game_Scr, dsize=(960, 540), interpolation=cv2.INTER_AREA)\n                print(Game_Scr.shape)\n\n                print(Convolution(Game_Scr))\n<\/code><\/pre>\n\n<p>Second is my called function.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def Convolution(img):\n        kernel = tf.Variable(tf.truncated_normal(shape=[4], stddev=0.1))\n        sess = tf.Session()\n        with tf.Session() as sess:\n                img = img.astype('float32')\n                Bias1 = tf.Variable(tf.truncated_normal(shape=[4],stddev=0.1))\n                conv2d = tf.nn.conv2d(img, kernel, strides=[1, 1, 1, 1], padding='SAME')# + Bias1\n                conv2d = sess.run(conv2d)\n        return conv2d\n<\/code><\/pre>\n\n<p>ValueError: Shape must be rank 4 but is rank 3 for 'Conv2D' (op: 'Conv2D') with input shapes: [540,960,3], [4].<\/p>\n\n<p>I tried changing the shape many times, but I get the same error.<\/p>\n","answers":[{"AnswerId":"57507104","CreationDate":"2019-08-15T08:46:50.577","ParentId":null,"OwnerUserId":"5810950","Title":null,"Body":"<p>As per the official documentation <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/conv2d\" rel=\"nofollow noreferrer\">here<\/a>, input tensor should be of shape <code>[batch, in_height, in_width, in_channels]<\/code> and a filter \/ kernel tensor should be of shape <code>[filter_height, filter_width, in_channels, out_channels]<\/code>.<\/p>\n\n<p>Try by changing your <code>Convolution<\/code> function to something like this:<\/p>\n\n<pre><code>def Convolution(img):\n        kernel = tf.Variable(tf.truncated_normal(shape=[200, 200, 3, 3], stddev=0.1))\n        sess = tf.Session()\n        with tf.Session() as sess:\n            sess.run(tf.global_variables_initializer())\n            img = img.astype('float32')\n            conv2d = tf.nn.conv2d(np.expand_dims(img, 0), kernel, strides=[1, 1, 1, 1], padding='SAME')# + Bias1\n            conv2d = sess.run(conv2d)\n        return conv2d\n<\/code><\/pre>\n"},{"AnswerId":"57507094","CreationDate":"2019-08-15T08:46:24.967","ParentId":null,"OwnerUserId":"10455493","Title":null,"Body":"<p>Try replacing <\/p>\n\n<p><code>img = img.astype('float32')<\/code> <\/p>\n\n<p>with <\/p>\n\n<p><code>img = tf.expand_dims(img.astype('float32'), 0)<\/code><\/p>\n\n<p>The dimention of tf.nn.conv2d input shoul be 4, (batch_size, image_hight, image_with, image_channels). You where missing the batch_size, <code>tf.expand_dims<\/code> just add that dimention (with a batch_size of 1 since you only have one image). <\/p>\n"}]}
+{"QuestionId":57506707,"AnswerCount":2,"Tags":"<python-3.x><keras><keras-layer>","CreationDate":"2019-08-15T08:12:18.303","AcceptedAnswerId":57506931.0,"OwnerUserId":9667559.0,"Title":"How to load the Keras model with custom layers from .h5 file correctly?","Body":"<p>I built a Keras model with a custom layers, and it was saved to a <code>.h5<\/code> file by the callback <code>ModelCheckPoint<\/code>. \nWhen I tried to load this model after the training, the error message below showed up: <\/p>\n\n<blockquote>\n<pre><code>__init__() missing 1 required positional argument: 'pool_size'\n<\/code><\/pre>\n<\/blockquote>\n\n<p>This is the definition of the custom layer and its <code>__init__<\/code> method:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class MyMeanPooling(Layer):\n    def __init__(self, pool_size, axis=1, **kwargs):\n        self.supports_masking = True\n        self.pool_size = pool_size\n        self.axis = axis\n        self.y_shape = None\n        self.y_mask = None\n        super(MyMeanPooling, self).__init__(**kwargs)\n<\/code><\/pre>\n\n<p>This is how I add this layer to my model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = MyMeanPooling(globalvars.pool_size)(x)\n<\/code><\/pre>\n\n<p>This is how I load the model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.models import load_model\n\nmodel = load_model(model_path, custom_objects={'MyMeanPooling': MyMeanPooling})\n<\/code><\/pre>\n\n<p>These are the full error messages:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"D:\/My Projects\/Attention_BLSTM\/script3.py\", line 9, in &lt;module&gt;\n    model = load_model(model_path, custom_objects={'MyMeanPooling': MyMeanPooling})\n  File \"D:\\ProgramData\\Anaconda3\\envs\\tf\\lib\\site-packages\\keras\\engine\\saving.py\", line 419, in load_model\n    model = _deserialize_model(f, custom_objects, compile)\n  File \"D:\\ProgramData\\Anaconda3\\envs\\tf\\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 \"D:\\ProgramData\\Anaconda3\\envs\\tf\\lib\\site-packages\\keras\\engine\\saving.py\", line 458, in model_from_config\n    return deserialize(config, custom_objects=custom_objects)\n  File \"D:\\ProgramData\\Anaconda3\\envs\\tf\\lib\\site-packages\\keras\\layers\\__init__.py\", line 55, in deserialize\n    printable_module_name='layer')\n  File \"D:\\ProgramData\\Anaconda3\\envs\\tf\\lib\\site-packages\\keras\\utils\\generic_utils.py\", line 145, in deserialize_keras_object\n    list(custom_objects.items())))\n  File \"D:\\ProgramData\\Anaconda3\\envs\\tf\\lib\\site-packages\\keras\\engine\\network.py\", line 1022, in from_config\n    process_layer(layer_data)\n  File \"D:\\ProgramData\\Anaconda3\\envs\\tf\\lib\\site-packages\\keras\\engine\\network.py\", line 1008, in process_layer\n    custom_objects=custom_objects)\n  File \"D:\\ProgramData\\Anaconda3\\envs\\tf\\lib\\site-packages\\keras\\layers\\__init__.py\", line 55, in deserialize\n    printable_module_name='layer')\n  File \"D:\\ProgramData\\Anaconda3\\envs\\tf\\lib\\site-packages\\keras\\utils\\generic_utils.py\", line 147, in deserialize_keras_object\n    return cls.from_config(config['config'])\n  File \"D:\\ProgramData\\Anaconda3\\envs\\tf\\lib\\site-packages\\keras\\engine\\base_layer.py\", line 1109, in from_config\n    return cls(**config)\nTypeError: __init__() missing 1 required positional argument: 'pool_size'\n<\/code><\/pre>\n","answers":[{"AnswerId":"57506906","CreationDate":"2019-08-15T08:30:02.007","ParentId":null,"OwnerUserId":"9006687","Title":null,"Body":"<p>From the answer of \"LiamHe commented on Sep 27, 2017\" on the following issue: <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/4871\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/4871<\/a>.<\/p>\n\n<p>I met the same problem today : ** TypeError: init() missing 1 required positional arguments**. Here is how I solve the problem : (Keras 2.0.2)<\/p>\n\n<ol>\n<li>Give the positional arguments of the layer with some default values<\/li>\n<li>Override get_config function to the layer with some thing like<\/li>\n<\/ol>\n\n<pre><code>def get_config(self):\n    config = super().get_config()\n    config['pool_size'] = # say self._pool_size  if you store the argument in __init__\n    return config\n<\/code><\/pre>\n\n<ol start=\"3\">\n<li>Add layer class to custom_objects when you are loading model.<\/li>\n<\/ol>\n"},{"AnswerId":"57506931","CreationDate":"2019-08-15T08:31:56.490","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Actually I don't think you can load this model.<\/p>\n\n<p>The most likely issue is that you did not implement the <code>get_config()<\/code> method in your layer. This method returns a dictionary of configuration values that should be saved:<\/p>\n\n<pre><code>def get_config(self):\n    config = {'pool_size': self.pool_size,\n              'axis': self.axis}\n    base_config = super(MyMeanPooling, self).get_config()\n    return dict(list(base_config.items()) + list(config.items()))\n<\/code><\/pre>\n\n<p>You have to retrain the model after adding this method to your layer, as the previously saved model does not have the configuration for this layer saved into it. This is why you cannot load it, it requires retraining after making this change.<\/p>\n"}]}
+{"QuestionId":57506743,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><tensorflow-lite><google-coral>","CreationDate":"2019-08-15T08:14:41.137","AcceptedAnswerId":57530529.0,"OwnerUserId":11930747.0,"Title":"If Edge TPU compiler only supports 3 dim tensors how to implement conv2d layers?","Body":"<p>I've got a question that's hopefully simple to answer.<\/p>\n\n<p>If I want to implement a simple conv2d layer I need a 4 dimensional weight tensor, where all dimension sizes are usually greater than 1.<\/p>\n\n<p>But the requirements for the Edge TPU Compiler say: \"Tensors are either 1-, 2-, or 3-dimensional. If a tensor has more than 3 dimensions, then only the 3 innermost dimensions may have a size greater than 1.\"\n(<a href=\"https:\/\/coral.withgoogle.com\/docs\/edgetpu\/models-intro\/\" rel=\"nofollow noreferrer\">https:\/\/coral.withgoogle.com\/docs\/edgetpu\/models-intro\/<\/a>)<\/p>\n\n<p>Now to my question: How do I implement a conv layer if there are only three dimensional tensors with sizes greater than one available? Or am I mistaken?<\/p>\n\n<p>Thank you in advance<\/p>\n","answers":[{"AnswerId":"57531656","CreationDate":"2019-08-16T22:07:09.663","ParentId":null,"OwnerUserId":"9705503","Title":null,"Body":"<p>The Edge TPU can't accept a batch of input images, and that's what the fourth dimension of your tensor is. It represents the size of your batch. And the Conv2d operation is defined on a three dimensional tensor, having a batch of images would just perform the 2D convolution for each input image separately in parallel.<\/p>\n"},{"AnswerId":"57530529","CreationDate":"2019-08-16T20:04:53.497","ParentId":null,"OwnerUserId":"6040721","Title":null,"Body":"<p>I think they mean the input tensor and the outputs of the layers are restricted like that. The compiler certainly accepts conv2d layers, even though the kernel tensors do not satisfy those shape requirements, as you have rightfully pointed out!<\/p>\n"}]}
+{"QuestionId":57507455,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><tensorflow-datasets>","CreationDate":"2019-08-15T09:17:07.407","AcceptedAnswerId":null,"OwnerUserId":10609889.0,"Title":"How to download specific labels\/images from tensorflow-datasets","Body":"<p>I'm currently using the coco\/2017 dataset for some use in tf.keras transferred learning <\/p>\n\n<p>I am aware that you can download the full dataset using the code bellow:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>(raw_train, raw_validation, raw_test), metadata = tfds.load(\n                                        'coco\/2017',\n                                        split=list(splits),\n                                        with_info=True,\n                                        as_supervised=True)\n<\/code><\/pre>\n\n<p>How would I do this if I only wanted to extract the 'persons' images and labels. Instead of the full dataset?<\/p>\n","answers":[]}
+{"QuestionId":57507555,"AnswerCount":0,"Tags":"<python><machine-learning><keras><computer-vision>","CreationDate":"2019-08-15T09:24:40.357","AcceptedAnswerId":null,"OwnerUserId":8941460.0,"Title":"Compare predictet image class to actual image class with keras","Body":"<p>I am training a keras model to recognise images of cats, dogs and horses. <\/p>\n\n<p>So far, I have one-hot-encoded my data (since this is a multi-class classification problem), trained my model and called the predictions. <\/p>\n\n<pre><code>def read_and_process_images(list_of_images):\nX = [] #images\ny = [] #labels\n\nfor image in list_of_images:\n    try:\n        X.append(cv2.resize(cv2.imread(image, cv2.IMREAD_COLOR),(nrows, ncolumns), interpolation = cv2.INTER_CUBIC))\n        if 'dog' in image:\n            y.append(0)\n        elif 'cat' in image:\n            y.append(1)\n        elif 'horse' in image:\n            y.append(2)\n    except Exception as e:\n        print(str(e))\n\nreturn X, y\n<\/code><\/pre>\n\n<p>...<\/p>\n\n<pre><code>X_test, y_test = read_and_process_images(test_imgs)\nx = np.array(X_test)\ntest_datagen = ImageDataGenerator(rescale = 1.\/255)\n\ni = 0\ntext_labels = []\n\nplt.figure(figsize = (30,20))\nfor batch in test_datagen.flow(x, batch_size = 1):\n    pred = model.predict(batch)\n    print(np.argmax(pred))\n    if np.argmax(pred) == 0 :\n        text_labels.append('dog')\n    elif np.argmax(pred) == 1:\n        text_labels.append('cat')\n    else:\n        text_labels.append('horse')\n    plt.subplot(5 \/ columns + 1, columns, i+1)\n    plt.title('I think this is a ' + text_labels[i])\n    imgplot = plt.imshow(batch[0])\n    i += 1\n    if i % 10 == 0:\n         break\n\n plt.show()\n<\/code><\/pre>\n\n<p>The model seems to be working very well. I usually get between 7-10 correct predictions, depending on the batch size. However, I do not understand how model.predict chooses the batches, and I am therefore unable to compare the actual value to the predicted value. When I try to do the following:<\/p>\n\n<pre><code>y_pred = model.predict(x, batch_size=1)\nmatrix = confusion_matrix(y_test, y_pred.argmax(axis=1))\n<\/code><\/pre>\n\n<p>the confusion matrix that I get is completely nonsensical (for example it tells me that it only got one cat correct, but I can clearly see with some batches that it got many more correct). Could someone explain to me, how the .predict function goes about choosing its batches, and how I can successfully compare the predicted values to the actual test values? Thank you in advance.<\/p>\n","answers":[]}
+{"QuestionId":57507576,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><recurrent-neural-network>","CreationDate":"2019-08-15T09:27:01.253","AcceptedAnswerId":null,"OwnerUserId":11931013.0,"Title":"Why should I use batch_length divided by num_steps to determine the epoch_size?","Body":"<p>I am working on the RNN tutorial with PTB dataset(<a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/rnn\/ptb\/reader.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/rnn\/ptb\/reader.py<\/a>). In this tutorial, they simply divide the (number of batches) by the (number of steps(or look_back)) to determine the epoch size. <\/p>\n\n<pre><code>    epoch_size = (batch_len - 1) \/\/ num_steps\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>    epoch_size = ((len(data) \/\/ batch_size) - 1) \/\/ num_steps\n<\/code><\/pre>\n\n<p>In my thought, it would be better to use <code>(number of batches) - (number of steps) + 1<\/code> for epoch size. There would be more data to train and I think its more reasonable since the model can learn all sequences of the data.<\/p>\n","answers":[]}
+{"QuestionId":57507774,"AnswerCount":1,"Tags":"<python><keras><recurrent-neural-network><tensorflow2.0>","CreationDate":"2019-08-15T09:43:39.960","AcceptedAnswerId":null,"OwnerUserId":3355572.0,"Title":"tf2.0 Keras: unable to save weights when using custom tensorflow code for RNN","Body":"<p>I am writing a Encoder-Recurrent Decoder model in tf2 using the Keras API. During the decoding phase I would like to apply some custom functions to the model outputs. However, once I start \"embedding\" tensorflow code into my functional Keras model I cannot save the weights anymore - instead a <code>ValueError<\/code> Exception occurs:<\/p>\n\n<pre><code>ValueError: Unable to create group (name already exists)\n<\/code><\/pre>\n\n<p>The model <strong>CAN<\/strong> be trained and the loss is reducing as I would expect. The only problem is storing the weights afterwards.<\/p>\n\n<p>The whole error message:<\/p>\n\n<pre><code>ValueError                                Traceback (most recent call last)\n&lt;ipython-input-4-5edae36ba5fe&gt; in &lt;module&gt;()\n     57 model.compile(loss='MSE', optimizer='adam')\n     58 \n---&gt; 59 model.save_weights('\/tmp\/dummy.h5')\n     60 print('saved')\n\n2 frames\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/network.py in save_weights(self, filepath, overwrite, save_format)\n   1310     if save_format == 'h5':\n   1311       with h5py.File(filepath, 'w') as f:\n-&gt; 1312         saving.save_weights_to_hdf5_group(f, self.layers)\n   1313     else:\n   1314       if context.executing_eagerly():\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/saving\/hdf5_format.py in save_weights_to_hdf5_group(f, layers)\n    617 \n    618   for layer in layers:\n--&gt; 619     g = f.create_group(layer.name)\n    620     weights = _legacy_weights(layer)\n    621     weight_values = K.batch_get_value(weights)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/h5py\/_hl\/group.py in create_group(self, name, track_order)\n     58             name, lcpl = self._e(name, lcpl=True)\n     59             gcpl = Group._gcpl_crt_order if track_order else None\n---&gt; 60             gid = h5g.create(self.id, name, lcpl=lcpl, gcpl=gcpl)\n     61             return Group(gid)\n     62 \n\nh5py\/_objects.pyx in h5py._objects.with_phil.wrapper()\n\nh5py\/_objects.pyx in h5py._objects.with_phil.wrapper()\n\nh5py\/h5g.pyx in h5py.h5g.create()\n\nValueError: Unable to create group (name already exists)\n<\/code><\/pre>\n\n<p>Things I tried so far:<\/p>\n\n<ul>\n<li>The error happens with tf 1.14.0 as well<\/li>\n<li>Adding <code>named_scope<\/code> around the offending code part did not help either: <code>with tf.name_scope('something' + str(step)):<\/code><\/li>\n<li>Explicitly naming the <code>tf.reshape<\/code> operations did not help: <code>outputs = tf.reshape(..., name='custom_reshape1_' + str(step))<\/code><\/li>\n<li>The error can be reproduced in <em>Google Colab<\/em><\/li>\n<li>This is not an issue with the weights file already being created as I delete all data prior to running the code. Furthermore, the code works fine when removing the offending tensorflow code<\/li>\n<\/ul>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nprint('version=', tf.__version__)\nimport numpy as np\n\nkeras = tf.keras # pylint: disable=no-member\nKL = keras.layers\nKM = keras.models\nK = keras.backend\nKR = keras.regularizers\n# GRU = tf.compat.v1.keras.layers.CuDNNGRU\nGRU = KL.GRU\n\nn_in = 20\nn_out = 10\ndim = 128\nhidden_units = 1024\n\nencoder_inputs = KL.Input(shape=(n_in, dim))\nencoder = GRU(hidden_units,\n              return_state=True)\n_, state_h = encoder(encoder_inputs)\n\ndecoder_gru = GRU(hidden_units,\n                  return_sequences=True,\n                  return_state=True)\ndecoder_dense = KL.Dense(dim)\n\nlast_pose = KL.Lambda(lambda a: a[:, -1, :],\n                      output_shape=(None, dim))(encoder_inputs)\nlast_pose = KL.Reshape((1, dim))(last_pose)\n\n\nall_outputs = []\ninputs = last_pose\nfor step in range(n_out):\n  outputs, state_h = decoder_gru(inputs, initial_state=state_h)\n  outputs = decoder_dense(outputs)\n\n  # ~~~~~~~~~~~~~~~~~~~~~~~~ comment out to \"fix\" issue\n  # -- problem cause (start)\n  n_batch = tf.shape(outputs)[0]\n  outputs = tf.reshape(outputs, (n_batch * 32, 4))\n  # ... custom function stuff\n  outputs = tf.reshape(outputs, (n_batch, 1, 32*4))\n  # -- problem cause (end)\n  # ~~~~~~~~~~~~~~~~~~~~~~~~\n\n  all_outputs.append(outputs)\n  inputs = outputs\n\ndecoder_outputs = KL.Concatenate()(all_outputs)\ndecoder_outputs = KL.Reshape((n_out, dim))(decoder_outputs)\nmodel = KM.Model(inputs=encoder_inputs,\n                 outputs=decoder_outputs)\n\nmodel = KM.Model(inputs=encoder_inputs, outputs=state_h)\n\n\nmodel.compile(loss='MSE', optimizer='adam')\n\nmodel.save_weights('\/tmp\/dummy.h5')\nprint('saved')\n<\/code><\/pre>\n\n<p>I am not sure what is going wrong and I would like to be able to save the weights! The model can be trained and its loss is reducing, only storing the weights causes problems.<\/p>\n","answers":[{"AnswerId":"57508932","CreationDate":"2019-08-15T11:25:32.543","ParentId":null,"OwnerUserId":"3355572","Title":null,"Body":"<p>It seems that the <code>tf.shape<\/code> call was the culprit. Reformulating the code without it allows the model to be stored as expected. I am still curious why this happens though..<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nprint('version=', tf.__version__)\nimport numpy as np\n\nkeras = tf.keras # pylint: disable=no-member\nKL = keras.layers\nKM = keras.models\nK = keras.backend\nKR = keras.regularizers\n# GRU = tf.compat.v1.keras.layers.CuDNNGRU\nGRU = KL.GRU\n\nn_in = 20\nn_out = 10\ndim = 128\nhidden_units = 1024\n\nencoder_inputs = KL.Input(shape=(n_in, dim))\nencoder = GRU(hidden_units,\n              return_state=True)\n_, state_h = encoder(encoder_inputs)\n\ndecoder_gru = GRU(hidden_units,\n                  return_sequences=True,\n                  return_state=True)\ndecoder_dense = KL.Dense(dim)\n\nlast_pose = KL.Lambda(lambda a: a[:, -1, :],\n                      output_shape=(None, dim))(encoder_inputs)\nlast_pose = KL.Reshape((1, dim))(last_pose)\n\n\nall_outputs = []\ninputs = last_pose\nfor step in range(n_out):\n  outputs, state_h = decoder_gru(inputs, initial_state=state_h)\n  outputs = decoder_dense(outputs)\n\n  # ~~~~~~~~~~~~~~~~~~~~~~~~ comment out to \"fix\" issue\n  # -- problem cause (start)\n  # n_batch = tf.shape(outputs)[0]\n  outputs = tf.reshape(outputs, (-1, 4))\n  # ... custom function stuff\n  outputs = tf.reshape(outputs, (-1, 1, 32*4))\n  # -- problem cause (end)\n  # ~~~~~~~~~~~~~~~~~~~~~~~~\n\n  all_outputs.append(outputs)\n  inputs = outputs\n\ndecoder_outputs = KL.Concatenate()(all_outputs)\ndecoder_outputs = KL.Reshape((n_out, dim))(decoder_outputs)\nmodel = KM.Model(inputs=encoder_inputs,\n                 outputs=decoder_outputs)\n\nmodel = KM.Model(inputs=encoder_inputs, outputs=state_h)\n\n\nmodel.compile(loss='MSE', optimizer='adam')\n\nmodel.save_weights('\/tmp\/dummy.h5')\nprint('saved')\n<\/code><\/pre>\n"}]}
+{"QuestionId":57508031,"AnswerCount":2,"Tags":"<tensorflow><machine-learning><keras><xgboost><gbm>","CreationDate":"2019-08-15T10:05:19.597","AcceptedAnswerId":null,"OwnerUserId":11095047.0,"Title":"How to adjust Model for rare binary outcome with Tensorflow or GBM","Body":"<p>I'm currently working on data with rare binary outcome, i.e. the response vector contains mostly 0 and only a few 1 (approximately 1.5% ones). I've got about 20 continuous explanatory variables. I tried to train models using GBM, Random Forests, TensorFlow with Keras backend.<\/p>\n\n<p>I observed a special behavior of the models, regardless which method I used:<\/p>\n\n<p>The accuracy is high (~98%) but the model predicts probabilities for class \"0\" for all outcomes as ~98.5% and for class \"1\" ~1,5%.<\/p>\n\n<p>How can I prevent this behavior?<\/p>\n\n<p>I'm using RStudio. For Example a TF model with Keras would be:<\/p>\n\n<pre><code>model &lt;- keras_model_sequential()\n\nmodel %&gt;%\n  layer_dense(units = 256, activation = \"relu\", input_shape = c(20)) %&gt;%\n  layer_dense(units = 256, activation = \"relu\") %&gt;%\n  layer_dense(units = 2, activation = \"sigmoid\")\n\nparallel_model &lt;- multi_gpu_model(model, gpus=2)\nparallel_model %&gt;% compile(\n  optimizer = \"adam\",             \n  loss = \"binary_crossentropy\",\n  metrics = \"binary_accuracy\")\n\nhistroy &lt;- parallel_model %&gt;% fit(\n  x_train, y_train,\n  batch_size = 64,\n  epochs = 100,\n  class_weight = list(\"0\"=1,\"1\"=70),\n  verbose = 1,\n  validation_split = 0.2\n)\n<\/code><\/pre>\n\n<p>But my observation is not limited to TF. This makes my question more general. I'm not asking for specific adjustments for the model above, rather I'd like to discuss at what point all outcomes are assigned the same probability.<\/p>\n\n<p>I can guess, the issue is connected to the loss-function.\nI know there is no way to use AUC as loss functions, since it's not differentiable. If I test models with AUC with unknown data, the result is not better than random guessing.<\/p>\n\n<p>I don't mind answers with code in Python, since this isn't a problem about coding rather than general behavior and algorithms.<\/p>\n","answers":[{"AnswerId":"57509638","CreationDate":"2019-08-15T12:25:49.900","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>In my (not so big) experience with AUC problems and rare positives, I see models with one class (not two). It's either \"result is positive (1)\" or \"result is negative (0)\".<\/p>\n\n<p>Metrics like accuracy are useless for these problems, you should use AUC based metrics with big batch sizes.   <\/p>\n\n<p>For these problems, it doesn't matter whether the outcome probabilities are too little, as long as there is a difference between them. (Forests, GBM, etc. will indeed output these little values, but this is not a problem)    <\/p>\n\n<p>For neural networks, you can try to use class weights to increase the output probabilities. But notice that if you split the result in two separate classes (considering only one class should be positive), it doesn't matter if you use weights, because:    <\/p>\n\n<ul>\n<li>For the first class, low weights: predict all ones is good     <\/li>\n<li>For the second class, high weights: predict all zeros is good (weighted to very good)    <\/li>\n<\/ul>\n\n<p>So, as an initial solution, you can:    <\/p>\n\n<ul>\n<li>Use a <code>'softmax'<\/code> activation (to guarantee your model will have only one correct output) and a <code>'categorical_crossentropy'<\/code> loss. <\/li>\n<li>(Or, preferrably) Use a model with only one class and keep <code>'sigmoid'<\/code> with <code>'binary_crossentropy'<\/code>.       <\/li>\n<\/ul>\n\n<p>I always work with the preferrable option above. In this case, if you use batch sizes that are big enough to contain one or two positive examples (batch size around 100 for you), weights may even be discarded. If the batch sizes are too little and many batches don't contain positive results, you may have too many weight updates towards plain zeros, which is bad.     <\/p>\n\n<p>You may also resample your data and, for instance, multiply by 10 the number of positive examples, so your batches contain more positives and training becomes easier. <\/p>\n\n<h2>Example of AUC metric to determine when training should end:<\/h2>\n\n<pre><code>#in python - considering outputs with only one class\ndef aucMetric(true, pred):\n    true= K.flatten(true)\n    pred = K.flatten(pred)\n\n    totalCount = K.shape(true)[0]\n\n    values, indices = tf.nn.top_k(pred, k = totalCount)        \n    sortedTrue = K.gather(true, indices)\n\n    tpCurve = K.cumsum(sortedTrue)\n    negatives = 1 - sortedTrue\n    auc = K.sum(tpCurve * negatives)\n\n    totalCount = K.cast(totalCount, K.floatx())\n    positiveCount = K.sum(true)\n    negativeCount = totalCount - positiveCount\n    totalArea = positiveCount * negativeCount\n    return  auc \/ totalArea\n<\/code><\/pre>\n"},{"AnswerId":"57509633","CreationDate":"2019-08-15T12:25:39.750","ParentId":null,"OwnerUserId":"5134333","Title":null,"Body":"<p>When your problem has unbalanced classes, I suggest using SMOTE (on the training data only!!! never use smote on your testing data!!!) before training the model.<\/p>\n\n<p>For example:<\/p>\n\n<pre><code>from imblearn.over_sampling import SMOTE\nX_trn_balanced, Y_trn_balanced = SMOTE(random_state=1, ratio=1).fit_sample(X_trn, Y_trn)\n#next fit the model with the balanced data\nmodel.fit(X_trn_balanced, Y_trn_balanced )\n<\/code><\/pre>\n"}]}
+{"QuestionId":57508290,"AnswerCount":1,"Tags":"<python><machine-learning><keras>","CreationDate":"2019-08-15T10:28:03.813","AcceptedAnswerId":57509185.0,"OwnerUserId":3707926.0,"Title":"Is it possible to set value in different Keras callback?","Body":"<p>I have custom callback that will reset model weights on some conditions. But this will interfere with EarlyStopping callback saves best monitored value i.e. 'val_loss'. How can I access properties from my custom callback to EarlyStopping callback to set its best 'val_loss' to my value ?<\/p>\n","answers":[{"AnswerId":"57509185","CreationDate":"2019-08-15T11:48:03.750","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Set in your early stopper:<\/p>\n\n<pre><code>early_stopping_callback.best = new_value\nearly_stopping_callback.wait = 0\n<\/code><\/pre>\n\n<p>Based on the <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/callbacks.py#L816\" rel=\"nofollow noreferrer\">source code, currently at line 816<\/a><\/p>\n\n<hr>\n\n<p>Putting the early stopper inside:<\/p>\n\n<pre><code>class CustomCallback(...):\n    def __init__ (....., early_stopper, ...)\n        self.early_stopping_callback = early_stopper\n\n    .......\n\nearly_stopping = EarlyStopping(...)\ncustom = CustomCallback(...., early_stopping, ...)\n<\/code><\/pre>\n\n<p>You still need to pass both callbacks to <code>fit<\/code>:<\/p>\n\n<pre><code>model.fit(..., callbacks=[custom, early_stopping])    \n<\/code><\/pre>\n"}]}
+{"QuestionId":57508347,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><deep-learning><lstm>","CreationDate":"2019-08-15T10:33:12.993","AcceptedAnswerId":null,"OwnerUserId":11271299.0,"Title":"Time taking for training data. LSTM","Body":"<p>My data set contains 10 days' data of 1000 users. I am training and testing the data for each individual user for better prediction accuracy. The issue is for the first user training 100 epochs takes <code>5 secs<\/code> and for the 100th user, 100 epochs takes more than <code>five minutes<\/code>. The training time is increasing for each user. How to reduce the training time> As the location points are categorical, One hot encoding is implemented for encoding the location points.<\/p>\n\n<pre><code>list = list_users[:100]\nwith open(\"accuracy_Lstm.csv\",\"w\") as f:\n    f.write('user,LSTM \\n')\n    for user in list:\n        user_data = newdataframe[newdataframe.user==user]\n        encoded=encoding(user_data)\n        X_train = []\n        y_train = []\n        for i in range(1, len(encoded)-96):\n            X_train.append(encoded[i-1])\n            y_train.append(encoded[i])\n        X_train, y_train = np.array(X_train), np.array(y_train)\n        X_test = encoded[-192:-96,:]\n        X_true = encoded[-96:,:]\n        X_trainL=X_train.reshape(X_train.shape[0],1,X_train.shape[1])\n        time_steps = 1\n    #Lstm\n        model = Sequential()\n        model.add(LSTM(X_train.shape[1], input_shape=(time_steps,X_train.shape[1]), activation='relu'))\n        model.add(Dense(X_train.shape[1]))\n        model.compile(loss='mse', optimizer='adam')\n        model.fit(X_trainL, y_train, batch_size=96, epochs=100, verbose =1)\n        model.summary()\n\n        X_testL=X_test.reshape(X_test.shape[0],1,X_test.shape[1])\n\n        pedL =one_hot_decode(model.predict(X_testL))\n        true=one_hot_decode(X_true)\n        try:\n            accuracy = ((sum(x == y for x, y in zip(pedL, true)))\/(len(pedL)))*100\n        except ZeroDivisionError:\n            accuracy = 0\n        f.write(' %d,  %f \\n'%(user, accuracy))\n<\/code><\/pre>\n\n<p>How to reduce the training time for users?<\/p>\n","answers":[{"AnswerId":"57508411","CreationDate":"2019-08-15T10:39:43.247","ParentId":null,"OwnerUserId":"10455493","Title":null,"Body":"<p>The problem is that you are recreating a new model at each iteration in the for loop. This is quite memory intensive for each model, and should be avoided. This is why the first model train quite fast, and then slower for each new model. Even <code>del model<\/code> at each iteration would not help since the memory leak is in tensorflow. You could potentially clear session in the beginning of the for loop, but this in itself is quite slow. <\/p>\n\n<p>I suggest you create the model outside the loop, and then just re-initiate the weights of the model at each iteration (as the model architecture does not change between iteration)<\/p>\n\n<hr>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>As suggested in the comment, here is an example code of how this idea can be implemented. As the question is not a working code in itself, I have not tested and verified that the following code is working without a memory leak, but from my previous experience I think it should.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>list = list_users[:100]\n\ndef get_model():\n    model = Sequential()\n    model.add(LSTM(X_train.shape[1], input_shape=(time_steps,X_train.shape[1]), activation='relu'))\n    model.add(Dense(X_train.shape[1]))\n    model.compile(loss='mse', optimizer='adam')\n    model.summary()\n    return(model, model.get_weights())\n\nwith open(\"accuracy_Lstm.csv\",\"w\") as f:\n    f.write('user,LSTM \\n')\n    model, initial_weights = get_model()\n    for user in list:\n        user_data = newdataframe[newdataframe.user==user]\n        encoded=encoding(user_data)\n        X_train = []\n        y_train = []\n        for i in range(1, len(encoded)-96):\n            X_train.append(encoded[i-1])\n            y_train.append(encoded[i])\n        X_train, y_train = np.array(X_train), np.array(y_train)\n        X_test = encoded[-192:-96,:]\n        X_true = encoded[-96:,:]\n        X_trainL=X_train.reshape(X_train.shape[0],1,X_train.shape[1])\n        time_steps = 1\n    #Lstm\n        model.set_weights(initial_weights)\n        model.fit(X_trainL, y_train, batch_size=96, epochs=100, verbose =1)\n\n        X_testL=X_test.reshape(X_test.shape[0],1,X_test.shape[1])\n\n        pedL =one_hot_decode(model.predict(X_testL))\n        true=one_hot_decode(X_true)\n        try:\n            accuracy = ((sum(x == y for x, y in zip(pedL, true)))\/(len(pedL)))*100\n        except ZeroDivisionError:\n            accuracy = 0\n        f.write(' %d,  %f \\n'%(user, accuracy))\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>EDIT2:<\/strong><\/p>\n\n<p>Since the shape of the model is changing for each user, you will need to create the model at each iteration. therefor the above solution will not work.\nYou can clear the tensorflow memory by <code>tf.keras.backend.clear_session()<\/code>, but it is quite slow (that is why i tried to avoid it in the solution above). The following solution will probably be slower than 5 seconds for each user (due to the added clear time), but should be constant for each used independent on the number of users you have.<\/p>\n\n<pre><code>list = list_users[:100]\n\ndef get_model(input_size):\n    # clear the memory of tensorflow before creating a new model\n    tf.keras.backend.clear_session()\n    # creare new model\n    model = Sequential()\n    model.add(LSTM(input_size, input_shape=(time_steps,input_size), activation='relu'))\n    model.add(Dense(input_size))\n    model.compile(loss='mse', optimizer='adam')\n    model.summary()\n    return(model)\n\nwith open(\"accuracy_Lstm.csv\",\"w\") as f:\n    f.write('user,LSTM \\n')\n    for user in list:\n        # clear the memory of tensorflow at each new model\n        tf.keras.backend.clear_session()\n\n        user_data = newdataframe[newdataframe.user==user]\n        encoded=encoding(user_data)\n        X_train = []\n        y_train = []\n        for i in range(1, len(encoded)-96):\n            X_train.append(encoded[i-1])\n            y_train.append(encoded[i])\n        X_train, y_train = np.array(X_train), np.array(y_train)\n        X_test = encoded[-192:-96,:]\n        X_true = encoded[-96:,:]\n        X_trainL=X_train.reshape(X_train.shape[0],1,X_train.shape[1])\n        time_steps = 1\n    #Lstm\n        model = get_model(X_train.shape[1])\n        model.fit(X_trainL, y_train, batch_size=96, epochs=100, verbose =1)\n\n        X_testL=X_test.reshape(X_test.shape[0],1,X_test.shape[1])\n\n        pedL =one_hot_decode(model.predict(X_testL))\n        true=one_hot_decode(X_true)\n        try:\n            accuracy = ((sum(x == y for x, y in zip(pedL, true)))\/(len(pedL)))*100\n        except ZeroDivisionError:\n            accuracy = 0\n        f.write(' %d,  %f \\n'%(user, accuracy))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57509069,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tf.keras>","CreationDate":"2019-08-15T11:37:13.487","AcceptedAnswerId":57509493.0,"OwnerUserId":10202807.0,"Title":"Convert models( ?weights ) downloaded using applications module to tflite","Body":"<p>I am trying to convert mobilenet model downloaded using applications module in tf.keras to tensorflow lite format. TensorFlow version I am using is 1.31. I don't know whether model is actually stored weights only or weights+architecture+optimizer_state. When I tried the conversion command :<\/p>\n\n<pre><code>from tensorflow import lite\n\nlite.TFLiteConverter.from_keras_model_file( '\/path\/to\/mobilenet_1_0_224_tf.h5' )\n\n<\/code><\/pre>\n\n<p>It resulted in this error :<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/lite\/python\/lite.py\", line 370, in from_keras_model_file\n    keras_model = _keras.models.load_model(model_file)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/saving.py\", line 232, in load_model\n    raise ValueError('No model found in config file.')\nValueError: No model found in config file.\n\n<\/code><\/pre>\n\n<p>From this, I assumed that model will be weights only. So, I tried to load the model as one do, using the applications module and tried to save the model using model.save(). But this resulted in the following error.<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py\", line 300, in __init__\n    fetch, allow_tensor=True, allow_operation=True))\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 3478, in as_graph_element\n    return self._as_graph_element_locked(obj, allow_tensor, allow_operation)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 3557, in _as_graph_element_locked\n    raise ValueError(\"Tensor %s is not an element of this graph.\" % obj)\nValueError: Tensor Tensor(\"conv1\/kernel\/Read\/ReadVariableOp:0\", shape=(3, 3, 3, 32), dtype=float32) is not an element of this graph.\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 \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/network.py\", line 1334, in save\n    save_model(self, filepath, overwrite, include_optimizer)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/saving.py\", line 111, in save_model\n    save_weights_to_hdf5_group(model_weights_group, model_layers)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/saving.py\", line 742, in save_weights_to_hdf5_group\n    weight_values = K.batch_get_value(symbolic_weights)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/backend.py\", line 2819, in batch_get_value\n    return get_session().run(tensors)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py\", line 929, in run\n    run_metadata_ptr)\n  File \"\/usr\/local\/lib\/python3.5\/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\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py\", line 471, in __init__\n    self._fetch_mapper = _FetchMapper.for_fetch(fetches)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py\", line 261, in for_fetch\n    return _ListFetchMapper(fetch)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py\", line 370, in __init__\n    self._mappers = [_FetchMapper.for_fetch(fetch) for fetch in fetches]\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py\", line 370, in &lt;listcomp&gt;\n    self._mappers = [_FetchMapper.for_fetch(fetch) for fetch in fetches]\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py\", line 271, in for_fetch\n    return _ElementFetchMapper(fetches, contraction_fn)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py\", line 307, in __init__\n    'Tensor. (%s)' % (fetch, str(e)))\nValueError: Fetch argument &lt;tf.Variable 'conv1\/kernel:0' shape=(3, 3, 3, 32) dtype=float32&gt; cannot be interpreted as a Tensor. (Tensor Tensor(\"conv1\/kernel\/Read\/ReadVariableOp:0\", shape=(3, 3, 3, 32), dtype=float32) is not an element of this graph.)\n\n<\/code><\/pre>\n\n<p>Does anyone know what is the real problem here? TIA<\/p>\n","answers":[{"AnswerId":"57509493","CreationDate":"2019-08-15T12:15:28.067","ParentId":null,"OwnerUserId":"8073544","Title":null,"Body":"<p>How did you saved your model,maybe you have saved only weights not model and you are trying to call load model which is not present.<\/p>\n\n<p>If this is not the problem try to clear session.<\/p>\n\n<pre><code>from keras.backend import clear_session\nclear_session()\n<\/code><\/pre>\n\n<p>I convert the model in this way<\/p>\n\n<pre><code>converter = tf.lite.TFLiteConverter.from_keras_model_file('model name')\ntflite_model = converter.convert()\nopen(\"converted\/model.tflite\", \"wb\").write(tflite_model)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57509111,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-08-15T11:41:19.480","AcceptedAnswerId":null,"OwnerUserId":891699.0,"Title":"How to properly set keras output layer name to work with utensor?","Body":"<p>I want to use <a href=\"https:\/\/github.com\/uTensor\/uTensor\" rel=\"nofollow noreferrer\">utensor<\/a> with my <code>keras<\/code> script.  <\/p>\n\n<p><code>Utensor<\/code> needs <code>.pb<\/code> file and output layer name to work, so I set name for output layer:  <\/p>\n\n<pre><code>model.add(Dense(13, activation='softmax', name='predict'))\n<\/code><\/pre>\n\n<p>Convert <code>keras<\/code> model to <code>pb<\/code> file and run <code>utensor-cli<\/code>:  <\/p>\n\n<pre><code>utensor-cli convert model.pb --output-nodes=predict\n<\/code><\/pre>\n\n<p>But seems <code>utensor<\/code> cannot find <code>predict<\/code> layer name:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/home\/victor\/miniconda3\/envs\/cube\/bin\/utensor-cli\", line 10, in &lt;module&gt;\n    sys.exit(cli())\n  File \"\/home\/victor\/miniconda3\/envs\/cube\/lib\/python3.6\/site-packages\/click\/core.py\", line 764, in __call__\n    return self.main(*args, **kwargs)\n  File \"\/home\/victor\/miniconda3\/envs\/cube\/lib\/python3.6\/site-packages\/click\/core.py\", line 717, in main\n    rv = self.invoke(ctx)\n  File \"\/home\/victor\/miniconda3\/envs\/cube\/lib\/python3.6\/site-packages\/click\/core.py\", line 1137, in invoke\n    return _process_result(sub_ctx.command.invoke(sub_ctx))\n  File \"\/home\/victor\/miniconda3\/envs\/cube\/lib\/python3.6\/site-packages\/click\/core.py\", line 956, in invoke\n    return ctx.invoke(self.callback, **ctx.params)\n  File \"\/home\/victor\/miniconda3\/envs\/cube\/lib\/python3.6\/site-packages\/click\/core.py\", line 555, in invoke\n    return callback(*args, **kwargs)\n  File \"\/home\/victor\/miniconda3\/envs\/cube\/lib\/python3.6\/site-packages\/utensor_cgen\/cli.py\", line 84, in convert_graph\n    generator.generate(model_path)\n  File \"\/home\/victor\/miniconda3\/envs\/cube\/lib\/python3.6\/site-packages\/utensor_cgen\/backend\/code_generator.py\", line 50, in generate\n    ugraph = parser_cls.parse(self.model_file, self.output_nodes)\n  File \"\/home\/victor\/miniconda3\/envs\/cube\/lib\/python3.6\/site-packages\/utensor_cgen\/frontend\/tensorflow.py\", line 57, in parse\n    topologic_order_graph(ugraph)\n  File \"\/home\/victor\/miniconda3\/envs\/cube\/lib\/python3.6\/site-packages\/utensor_cgen\/utils.py\", line 218, in topologic_order_graph\n    visit(node_name)\n  File \"\/home\/victor\/miniconda3\/envs\/cube\/lib\/python3.6\/site-packages\/utensor_cgen\/utils.py\", line 207, in visit\n    op_info = ugraph.ops_info[node_name]\nKeyError: 'predict'\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57509460,"AnswerCount":0,"Tags":"<python-3.x><keras><auto-keras>","CreationDate":"2019-08-15T12:13:20.393","AcceptedAnswerId":null,"OwnerUserId":11884974.0,"Title":"Found input variables with inconsistent numbers of samples using Autokeras","Body":"<p>I am trying to classify radio signals by their modulation type using AutoKeras. I looked into some examples and I tried to replicate the same thing with my data. But, I always get an error about <strong>inconsistent numbers of samples<\/strong>. So the input data shapes are the following:<\/p>\n\n<pre><code>Training Data shape: (132000, 2, 128)  #X_train\nTesting Data shape: (88000, 2, 128)    #X_test\nTraining Labels shape: (132000, 11)    #Y_train\nTesting Labels shape: (88000, 11)      #Y_test\n<\/code><\/pre>\n\n<p>and here is my code:<\/p>\n\n<pre><code>    np.random.seed(1567)\n    n_example = X.shape[0]\n    n_train = n_example * 0.6\n    train_idx = np.random.choice(range(0, n_example), size=int(n_train), replace=False)\n    test_idx = list(set(range(0, n_example)) - set(train_idx))\n    X_train = X[train_idx]\n    X_test = X[test_idx]\n\n    print('Training Data shape:', X_train.shape)\n    print('Testing Data shape:', X_test.shape)\n\n    Y_train = to_onehot(list(map(lambda x: mods.index(lbl[x][0]), train_idx)))\n    Y_test = to_onehot(list(map(lambda x: mods.index(lbl[x][0]), test_idx)))\n\n    print('Training Labels shape:', Y_train.shape)\n    print('Testing Labels shape:', Y_test.shape)\n    classes = mods\n\n    for seconds in TRAINING_TIMES:\n        # train our Auto-Keras model\n        print(\"[INFO] training model for {} seconds max...\".format(\n            seconds))\n        model = ak.ImageClassifier(verbose=True)\n        X_train = X_train.reshape(X_train.shape + (1,))\n        X_test = X_test.reshape(X_test.shape + (1,))\n        model.fit(X_train, Y_train, time_limit=seconds)\n        model.final_fit(X_train, Y_train, X_test, Y_test, retrain=True)\n\n        # evaluate the Auto-Keras model\n        score = model.evaluate(X_test, Y_test)\n        predictions = model.predict(X_test)\n        report = classification_report(Y_test, predictions,\n                                       target_names=classes)\n\n        # write the report to disk\n        p = os.path.sep.join(OUTPUT_PATH, \"{}.txt\".format(seconds))\n        f = open(p, \"w\")\n        f.write(report)\n        f.write(\"\\nscore: {}\".format(score))\n        f.close()\n<\/code><\/pre>\n\n<p>the exact error is:<\/p>\n\n<pre><code>File \"\/home\/nechi\/PycharmProjects\/CNN_MR_v2\/cnn.py\", line 79, in &lt;module&gt;\n    main()\n  File \"\/home\/nechi\/PycharmProjects\/CNN_MR_v2\/cnn.py\", line 62, in main\n    model.fit(X_train, Y_train, time_limit=seconds)\n  File \"\/home\/nechi\/PycharmProjects\/CNN_MR_v2\/venv\/lib\/python3.6\/site-packages\/autokeras\/image\/image_supervised.py\", line 155, in fit\n    super().fit(x, y, time_limit)\n  File \"\/home\/nechi\/PycharmProjects\/CNN_MR_v2\/venv\/lib\/python3.6\/site-packages\/autokeras\/supervised.py\", line 173, in fit\n    random_state=42)\n  File \"\/home\/nechi\/PycharmProjects\/CNN_MR_v2\/venv\/lib\/python3.6\/site-packages\/sklearn\/model_selection\/_split.py\", line 2184, in train_test_split\n    arrays = indexable(*arrays)\n  File \"\/home\/nechi\/PycharmProjects\/CNN_MR_v2\/venv\/lib\/python3.6\/site-packages\/sklearn\/utils\/validation.py\", line 260, in indexable\n    check_consistent_length(*result)\n  File \"\/home\/nechi\/PycharmProjects\/CNN_MR_v2\/venv\/lib\/python3.6\/site-packages\/sklearn\/utils\/validation.py\", line 235, in check_consistent_length\n    \" samples: %r\" % [int(l) for l in lengths])\nValueError: Found input variables with inconsistent numbers of samples: [132000, 1452000]\nIs there \n<\/code><\/pre>\n\n<p>a solution for such a problem?\nThank you<\/p>\n","answers":[]}
+{"QuestionId":57509831,"AnswerCount":0,"Tags":"<python><tensorflow><autoencoder>","CreationDate":"2019-08-15T12:42:07.573","AcceptedAnswerId":null,"OwnerUserId":9764088.0,"Title":"How to supervise latent variables in the variational Autoencoder (VAE)","Body":"<p>I'm trying to use the <strong>Variational Autoencoder<\/strong> for the <strong>lateral control<\/strong> of the vehicle. (End-to-End driving)<br\/>\nThe whole project is based on the paper of Alexander Amini:<br\/>\n<em>\"Variational Autoencoder for End-to-End Control of Autonomous Driving with Novelty Detection and Training De-biasing\"<\/em> <br\/>\n<a href=\"http:\/\/people.csail.mit.edu\/rosman\/papers\/iros-2018-variational.pdf\" rel=\"nofollow noreferrer\">http:\/\/people.csail.mit.edu\/rosman\/papers\/iros-2018-variational.pdf<\/a><\/p>\n\n<p>The idea is that I'm sending one semantic segmented image (from the center of the car) to the input, compress it to the certain vector of latent values (25x1), where one of the latent variables supposed to represent the value of curvature of the road. This value I'm using for the control of the car, the rest latent variables I need to be able to detect Novelty (images, where the NN was trained unsufficiently) and to unbias some new data sets (the VAE can be further trained). As part of this issue I'm only concerned about the first value (curvature of the road).<\/p>\n\n<p><strong>The problem is, that the VAE outputs wrong values (for the lateral control) after encoding part (before decoding)<\/strong><\/p>\n\n<p>My training set consists of images with corresponding labels (value of the curvature of the road)<\/p>\n\n<p>To train my VAE I'm using the loss function which consist of three parts:<\/p>\n\n<ol>\n<li>Supervising of the latent variable, this is L2 Norm between the latent variable (\u0177 = z[:,0], means I'm taking the first latent variable from the latent vector) and the label.<br\/>  <em>L_part_1(y, y_hat) = (y - y_hat)^2<\/em><\/li>\n<li>Reconstruction part, which is L1 Norm between original image and it's reconstructed version. <br\/> <em>L_part_2(x,x_hat) = |x - x_hat|<\/em><\/li>\n<li>KL Divergence between the latent variables and a unit Gaussian, providing regularization for the latent space. <br\/> \n<em>L_part_3(z_mean, z_std_dev) = -0.5 (1-z_mean^2 - z_std_dev^2 + log(z_std_dev + 1e-10)^2)<\/em><\/li>\n<\/ol>\n\n<p>The total loss function has a following form: <br\/> \n<em>L = 0.033 x L_part_1 + 0.1 x L_part_2 + 0.001 x L_part_3<\/em> <br\/> \nWhose coefficients regularizing the importance of every loss part.<br\/> <\/p>\n\n<p><em>Those loss functions deviate a bit from the trivial form of VAE loss functions. I took those from that paper, that mentioned above. I also tried the trivial one, bit I'll tell about it later.<\/em><br\/> <\/p>\n\n<p>Here you can see the graphics of the loss function: <a href=\"https:\/\/ibb.co\/x8GgVxq\" rel=\"nofollow noreferrer\">https:\/\/ibb.co\/x8GgVxq<\/a> <br\/> \nAnd the result of the reconstructed images: <a href=\"https:\/\/ibb.co\/D1p88PG\" rel=\"nofollow noreferrer\">https:\/\/ibb.co\/D1p88PG<\/a><\/p>\n\n<p>As you can see, the first part of the loss function (Third on the image) looks weird. And when I use this trained VAE in inference mode, I'm getting nonsense results.<\/p>\n\n<p>I also tried different version of the loss function (trivial), where the first part (supervising one latent value) remains, the second part (reconstructed) has following form: <br\/> \nL_part_2(x,x_hat) = x * log(1e-10 + x_hat) + (1 - x) * log(1e-10 + 1 - x_hat)<br\/> \nand the third part (KL Divergence) looks like:<br\/> \nL_part_3(z_mean, z_std_dev) = -0.5 * (1 + z_std_dev - (z_mean)^2 - exp(z_std_dev))<\/p>\n\n<p>For that kind of configuration of the loss function I got following results: <br\/> \n[loss functions] <a href=\"https:\/\/ibb.co\/VYV1F7F\" rel=\"nofollow noreferrer\">https:\/\/ibb.co\/VYV1F7F<\/a> <br\/> \n[reconstructed images] <a href=\"https:\/\/ibb.co\/ZTrpfxv\" rel=\"nofollow noreferrer\">https:\/\/ibb.co\/ZTrpfxv<\/a> <br\/> <\/p>\n\n<h2>Here is the key parts of my code with the first configuration of the loss function<\/h2>\n\n<pre><code>class VAE(tf.keras.Model):\n    def __init__(self, latent_dim, counter =0):\n        super(VAE, self).__init__()\n        self.counter = counter\n        self.latent_dim = latent_dim\n        self.inference_net = tf.keras.Sequential([\n            tf.keras.layers.InputLayer(input_shape=(66, 200, 3)),\n            tf.keras.layers.Conv2D(\n                filters=24, kernel_size=5, strides=(2, 2), activation='relu'),\n            tf.keras.layers.Conv2D(\n                filters=36, kernel_size=5, strides=(2, 2), activation='relu'),\n            tf.keras.layers.Conv2D(\n                filters=48, kernel_size=5, strides=(2, 2), activation='relu'),\n            tf.keras.layers.Conv2D(\n                filters=64, kernel_size=3, strides=(1, 1), activation='relu'),\n            tf.keras.layers.Conv2D(\n                filters=64, kernel_size=3, strides=(1, 1), activation='relu'),\n            tf.keras.layers.Flatten(),\n            tf.keras.layers.Dense(units=1000, activation=tf.nn.relu),\n            tf.keras.layers.Dense(units=100),\n            tf.keras.layers.Dense(units=50)])\n\n        self.generative_net = tf.keras.Sequential([\n            tf.keras.layers.InputLayer(input_shape=(None,latent_dim)),\n            tf.keras.layers.Dense(units=50),\n            tf.keras.layers.Dense(units=100),\n            tf.keras.layers.Dense(units=1000, activation=tf.nn.relu),\n            tf.keras.layers.Dense(units=1152, activation=tf.nn.relu),\n            tf.keras.layers.Reshape(target_shape=(1, 18, 64)),\n            tf.keras.layers.Conv2DTranspose(\n                filters=64,kernel_size=3,strides=(1, 1),padding=\"VALID\",activation='relu'),\n            tf.keras.layers.Conv2DTranspose(\n                filters=48, kernel_size=3, strides=(1, 1), padding=\"VALID\",activation='relu'),\n            tf.keras.layers.Conv2DTranspose(\n                filters=36, kernel_size=5, strides=(2, 2), padding=\"VALID\",activation='relu'),\n            tf.keras.layers.Conv2DTranspose(\n                filters=24, kernel_size=5, strides=(2, 2), padding=\"VALID\",activation='relu'),\n            tf.keras.layers.Conv2DTranspose(\n                filters=3, kernel_size=5, strides=(2, 2), padding=\"VALID\",activation='relu')\n                                    ])\n    def encode(self, x):\n\n        mean, logvar = tf.split(self.inference_net(x), num_or_size_splits=2, axis=1)\n        eps = tf.random.normal(tf.shape(mean))\n        z = eps * tf.exp(logvar * .5) + mean\n        return mean, logvar, z\n\n\n    def decode(self, z, apply_sigmoid=False):\n        logits = self.generative_net(z)\n        if apply_sigmoid:\n            probs = tf.sigmoid(logits)\n            return probs\n\n        return logits\n\n    def compute_clothoid(self, image):\n        test_im = np.reshape(image, (-1, 200, 66, 3))\n        _, _, z = self.encode(test_im)\n        clothoid = z[0][0]\n        return clothoid\n\noptimizer = tf.keras.optimizers.Adam(1e-4)\n\n\ndef compute_loss(model, x, labels):\n    # print(labels)\n    z_mean, z_std, z = model.encode(x)\n    x_logit = model.decode(z)\n    z_curv = z[:,0]\n    x_sum = x * 255.\n    x_logit_sum = x_logit * 255.\n\n    with file_writer.as_default():\n        tf.summary.image('input_image', x_sum, step=model.counter)\n        tf.summary.image('output_image', x_logit_sum, step=model.counter)\n\n    x_logit = tf.image.resize(x_logit, [66, 200])\n    x = tf.reshape(\n        x, [-1, image_dim[0] * image_dim[1] * image_dim[2]]\n                    )\n    x_logit = tf.reshape(\n        x_logit, [-1, image_dim[0] * image_dim[1] * image_dim[2]]\n                        )\n\n    sup_lat_loss = tf.math.square(z_curv - labels)\n\n    #reconstruction_loss = x * tf.math.log(1e-10 + x_logit) + (1 - x) * tf.math.log(1e-10 + 1 - x_logit)\n    reconstruction_loss = tf.abs(x - x_logit)\n    reconstruction_loss = tf.reduce_sum(reconstruction_loss, 1)\n\n    # kl_div_loss =  1 - tf.square(z_mean) - tf.square(z_std)  + tf.math.log(tf.square(z_std + 1e-10))\n    kl_div_loss =  1 + z_std - tf.square(z_mean) - tf.exp(z_std)\n    kl_div_loss = -0.5 * tf.reduce_sum(kl_div_loss, 1)\n\n    loss = tf.reduce_mean(reconstruction_loss + kl_div_loss)\n\n    with file_writer.as_default():\n        sup_lat_loss_show = c1 * tf.reduce_mean(sup_lat_loss)\n        reconstruction_loss_show = c2 * tf.reduce_mean(reconstruction_loss)\n        kl_div_loss_show = c3 * tf.reduce_mean(kl_div_loss)\n\n        with tf.name_scope('loss_function'):\n            tf.summary.scalar('sup_lat_loss', sup_lat_loss_show, step=model.counter)\n            tf.summary.scalar('reconstruction_loss', reconstruction_loss_show, step=model.counter)\n            tf.summary.scalar('kl_div_loss', kl_div_loss_show, step=model.counter)\n\n        return loss\n\n<\/code><\/pre>\n\n<h2>Possible reasons for the issue<\/h2>\n\n<p>Maybe the way I assign <code>z_curv<\/code> is the reason, or the way I coded the loss function.<\/p>\n","answers":[]}
+{"QuestionId":57509841,"AnswerCount":0,"Tags":"<python><android><tensorflow><keras>","CreationDate":"2019-08-15T12:42:50.090","AcceptedAnswerId":null,"OwnerUserId":1979915.0,"Title":"Different result when evaluating tflite model from python and from android","Body":"<p>I've trained a model using keras (in google colab), then I've converted the keras h5 file to tflite using <code>TFLiteConverter<\/code> (also in google colab) by using this code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\n\ntflite_converter = tf.lite.TFLiteConverter.from_keras_model_file(&lt;KERAS_H5_MODEL_PATH&gt;)\n\ntflite_model = tflite_converter.convert()\n\nwith open(&lt;KERAS_TFLITE_DEST_PATH&gt;, 'wb') as tflite_model_file:\n    tflite_model_file.write(tflite_model)\n<\/code><\/pre>\n\n<p>After that, I've run the tflite model (in google colab also):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport tensorflow as tf\n\n# Load TFLite model and allocate tensors.\ninterpreter = tf.lite.Interpreter(model_path=tflite_model_path)\ninterpreter.allocate_tensors()\n\ninterpreter.set_tensor(input_details[0]['index'], &lt;INPUT_DATA&gt;)\n\ninterpreter.invoke()\n\noutput = [interpreter.get_tensor(output_details[0]['index'])[0], interpreter.get_tensor(output_details[1]['index'])[0]]\n<\/code><\/pre>\n\n<p>However, when I run the tflite model in an android app (using the same input data) y get different outputs:<\/p>\n\n<p>(This code is based on <a href=\"https:\/\/github.com\/tensorflow\/examples\/tree\/master\/lite\/examples\/object_detection\/android\" rel=\"nofollow noreferrer\">this example<\/a> from tensorflow)<\/p>\n\n<pre class=\"lang-java prettyprint-override\"><code>AssetFileDescriptor fileDescriptor = assets.openFd(&lt;MODEL_FILENAME&gt;);\nFileInputStream inputStream = new FileInputStream(fileDescriptor.getFileDescriptor());\nFileChannel fileChannel = inputStream.getChannel();\nlong startOffset = fileDescriptor.getStartOffset();\nlong declaredLength = fileDescriptor.getDeclaredLength();\nMappedByteBuffer model = fileChannel.map(FileChannel.MapMode.READ_ONLY, startOffset, declaredLength);\n\nInterpreter.Options interpreterOptions = new Interpreter.Options().setNumThreads(NUM_THREADS).setAllowFp16PrecisionForFp32(Boolean.FALSE);\ntfLite = new Interpreter(model, interpreterOptions);\n\nfloat[][][][] floatValues = new float[1][inputSize][inputSize][3];\n\n\/\/ floatValues array is initialized with the same values as in python code above\n\nObject[] inputArray = {floatValues};\n\noutput1 = new float[1][8][8][18];\noutput2 = new float[1][16][16][18];\n\nMap&lt;Integer, Object&gt; outputMap = new HashMap&lt;&gt;();\noutputMap.put(0, output1);\noutputMap.put(1, output2);\n\ntfLite.runForMultipleInputsOutputs(inputArray, outputMap);\n<\/code><\/pre>\n\n<p>The output values I get in the android code have nothing to do with the values obtained with python code. This model (yolo v3 tiny) is to detect objects in images and I've constated that the values obtained with python code are the correct ones<\/p>\n\n<p>Why do I get different outputs in android? Do I have to do something special to run model in android? Do I have to do something different in model conversion from H5 to tflite?<\/p>\n\n<p>NOTE: version of tensorflow in google colab is 1.14.0. In android I use this version of tflite <code>org.tensorflow:tensorflow-lite:1.14.0<\/code><\/p>\n\n<p><strong>Updated<\/strong><\/p>\n\n<p>I've tried with an input array of ones in both cases (python and android) to be sure that the input data is the same and the results vary:<\/p>\n\n<pre><code> Python results    |  Android results\n 0 &gt;   0.06118933  |  0 &gt;   0.061190486\n 1 &gt; - 0.50384498  |  1 &gt; - 0.50384396\n 2 &gt; - 0.30500048  |  2 &gt; - 0.30500013\n 3 &gt; - 0.18725708  |  3 &gt; - 0.18725668\n 4 &gt; -12.56872463  |  4 &gt; -12.568727\n 5 &gt; - 3.53239870  |  5 &gt; - 3.5323968\n 6 &gt;   0.26756036  |  6 &gt;   0.26756087\n 7 &gt; - 0.63708508  |  7 &gt; - 0.6370841\n 8 &gt; - 0.11959708  |  8 &gt; - 0.119596675\n 9 &gt; - 0.42219949  |  9 &gt; - 0.42219883\n10 &gt; -12.26699734  | 10 &gt; -12.266998\n11 &gt; - 3.42504168  | 11 &gt; - 3.4250417\n12 &gt;   0.29714739  | 12 &gt;   0.2971481\n13 &gt; - 0.76750284  | 13 &gt; - 0.7675022\n14 &gt; - 0.13859260  | 14 &gt; - 0.13859189\n15 &gt; - 0.02096577  | 15 &gt; - 0.020965554\n16 &gt; -12.78137684  | 16 &gt; -12.781382\n17 &gt; - 3.34643984  | 17 &gt; - 3.3464384\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57509873,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2019-08-15T12:45:46.713","AcceptedAnswerId":null,"OwnerUserId":7204850.0,"Title":"How to fix Convolutional Autoencoder that is only predicting one colour band correctly?","Body":"<p>I'm trying to set up a convolutional autoencoder to encode small 8x8x3 images, but when I train it the output only seems to match the input in one band. Essentially, I have cutouts from a larger image that I've processed into power spectra in each band (using a 2D fourier transform), and I want to train an autoencoder so that I can perform some clustering on the encoded data.<\/p>\n\n<p>I've tried setting up a version of this that just takes the single band mnist inputs, and it does a fairly good job of reproducing those. I've done a lot of tweaking with hyperparameters and changing the number of layers (one or two sets of convolutions, adding dense layers to the embedding), the size of the encoded space. I thought the issue might have been in the way the data is being reshaped between the 3D image vectors and the 1D encoded space, but couldn't find a problem with it.<\/p>\n\n<p>Here's the model set-up, I'm using the keras api in Tensorflow 1.11.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#initialise layer sizes\nn_hidden = [30,15,30]\n\n#learning rates\nlr = 0.0001\nl2 = 0.00001\n\n#initialise sequential model\nautoencoder = Sequential()\n\n#add layers: Conv 1 (n, 8, 8, 3) &gt; (n, 8, 8, nh0) \nautoencoder.add(Conv2D(n_hidden[0], (3,3), padding='same',\n                 input_shape=x_train.shape[1:],\n                 data_format='channels_last',\n                 kernel_regularizer=regularizers.l2(l2),\n                 kernel_initializer='random_normal'))\nautoencoder.add(Activation('relu'))\n\n#Max Pooling: (n, 4, 4, nh0)\nautoencoder.add(MaxPooling2D(pool_size=(2, 2), data_format='channels_last'))\n\n#add layers: Conv 2 (n, 4, 4, nh0) &gt; (n, 4, 4, nh1)\nautoencoder.add(Conv2D(n_hidden[1], (3,3), padding='same',\n                 input_shape=x_train.shape[1:],\n                 data_format='channels_last',\n                 kernel_regularizer=regularizers.l2(l2),\n                 kernel_initializer='random_normal'))\nautoencoder.add(Activation('relu'))\n\n#Max Pooling: (n, 4, 4, nh1), Flattern: (n, 2*2*nh1)\nautoencoder.add(MaxPooling2D(pool_size=(2, 2), data_format='channels_last', name='pool'))\nautoencoder.add(Flatten(name='flatten'))\n\n#Encoded Layer: (n, nh1*2*2) &gt; (n, nh2)\nautoencoder.add(Dense(n_hidden[2],\n                kernel_regularizer=regularizers.l2(l2),\n                kernel_initializer='random_normal', name='dense_encoded'))\nautoencoder.add(Activation('relu'))\n\n#Decoder Starts: (n, nh1) &gt; (n, nh1*2*2)\nautoencoder.add(Dense(2*2*n_hidden[1],\n                kernel_regularizer=regularizers.l2(l2),\n                kernel_initializer='random_normal'))\nautoencoder.add(Activation('relu'))\nautoencoder.add(Reshape((2,2,n_hidden[1])))\n\n#Upscaling &gt; Convolution: (n, 2, 2, nh1) &gt; (n, 4, 4, nh1)\nautoencoder.add(UpSampling2D(size=(2,2), data_format='channels_last'))\nautoencoder.add(Conv2D(n_hidden[1], (3,3), padding='same',\n                 input_shape=x_train.shape[1:],\n                 data_format='channels_last',\n                 kernel_regularizer=regularizers.l2(l2),\n                 kernel_initializer='random_normal'))\nautoencoder.add(Activation('relu'))\n\n#Upscaling &gt; Convolution: (n, 4, 4, nh1) &gt; (n, 8, 8, nh0)\nautoencoder.add(UpSampling2D(size=(2,2), data_format='channels_last'))\nautoencoder.add(Conv2D(n_hidden[0], (3,3), padding='same',\n                 input_shape=x_train.shape[1:],\n                 data_format='channels_last',\n                 kernel_regularizer=regularizers.l2(l2),\n                 kernel_initializer='random_normal'))\nautoencoder.add(Activation('relu'))\n\n#Output: (n, 8, 8, 3)\nautoencoder.add(Conv2D(3, (3,3), padding='same',\n                 data_format='channels_last',\n                 kernel_regularizer=regularizers.l2(l2),\n                 kernel_initializer='random_normal'))\nautoencoder.add(Activation('sigmoid'))\n\n#initialise optimiser: Adam optimiser is RMSprop + Momentum\nadam = Adam(lr=lr)\n\nautoencoder.summary()\n\nencoder = Model(inputs=autoencoder.input,\n                outputs=autoencoder.get_layer('dense_encoded').output)\nencoder.summary()\n\ntry:\n    parallel_autoencoder = multi_gpu_model(autoencoder, gpus=2)\n    print(\"Training using multiple GPUs..\")\nexcept ValueError:\n    parallel_autoencoder = autoencoder\n    print(\"Training using single GPU or CPU..\")\n\n#set up tensorboard\ntensorboard = TensorBoard(log_dir='.\/logs', histogram_freq=0,\n                          write_graph=True, write_images=False)\n\n#compile model: add optimiser, loss and metrics\nparallel_autoencoder.compile(loss='MSE',\n                       optimizer=adam,\n                       metrics=['accuracy'])\n\n#fit the model\nparallel_autoencoder.fit(x_train, x_train,\n          epochs=100,\n          batch_size=2000,\n          callbacks=[tensorboard])\n<\/code><\/pre>\n\n<p>Here's the model summarys:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Model: \"sequential\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nconv2d (Conv2D)              (None, 8, 8, 30)          840       \n_________________________________________________________________\nactivation (Activation)      (None, 8, 8, 30)          0         \n_________________________________________________________________\nmax_pooling2d (MaxPooling2D) (None, 4, 4, 30)          0         \n_________________________________________________________________\nconv2d_1 (Conv2D)            (None, 4, 4, 15)          4065      \n_________________________________________________________________\nactivation_1 (Activation)    (None, 4, 4, 15)          0         \n_________________________________________________________________\npool (MaxPooling2D)          (None, 2, 2, 15)          0         \n_________________________________________________________________\nflatten (Flatten)            (None, 60)                0         \n_________________________________________________________________\ndense_encoded (Dense)        (None, 30)                1830      \n_________________________________________________________________\nactivation_2 (Activation)    (None, 30)                0         \n_________________________________________________________________\ndense (Dense)                (None, 60)                1860      \n_________________________________________________________________\nactivation_3 (Activation)    (None, 60)                0         \n_________________________________________________________________\nreshape (Reshape)            (None, 2, 2, 15)          0         \n_________________________________________________________________\nup_sampling2d (UpSampling2D) (None, 4, 4, 15)          0         \n_________________________________________________________________\nconv2d_2 (Conv2D)            (None, 4, 4, 15)          2040      \n_________________________________________________________________\nactivation_4 (Activation)    (None, 4, 4, 15)          0         \n_________________________________________________________________\nup_sampling2d_1 (UpSampling2 (None, 8, 8, 15)          0         \n_________________________________________________________________\nconv2d_3 (Conv2D)            (None, 8, 8, 30)          4080      \n_________________________________________________________________\nactivation_5 (Activation)    (None, 8, 8, 30)          0         \n_________________________________________________________________\nconv2d_4 (Conv2D)            (None, 8, 8, 3)           813       \n_________________________________________________________________\nactivation_6 (Activation)    (None, 8, 8, 3)           0         \n=================================================================\nTotal params: 15,528\nTrainable params: 15,528\nNon-trainable params: 0\n_________________________________________________________________\nModel: \"model\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nconv2d_input (InputLayer)    [(None, 8, 8, 3)]         0         \n_________________________________________________________________\nconv2d (Conv2D)              (None, 8, 8, 30)          840       \n_________________________________________________________________\nactivation (Activation)      (None, 8, 8, 30)          0         \n_________________________________________________________________\nmax_pooling2d (MaxPooling2D) (None, 4, 4, 30)          0         \n_________________________________________________________________\nconv2d_1 (Conv2D)            (None, 4, 4, 15)          4065      \n_________________________________________________________________\nactivation_1 (Activation)    (None, 4, 4, 15)          0         \n_________________________________________________________________\npool (MaxPooling2D)          (None, 2, 2, 15)          0         \n_________________________________________________________________\nflatten (Flatten)            (None, 60)                0         \n_________________________________________________________________\ndense_encoded (Dense)        (None, 30)                1830      \n=================================================================\nTotal params: 6,735\nTrainable params: 6,735\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>Here's a look at the output so you can see what I mean:<\/p>\n\n<p><img src=\"https:\/\/imgur.com\/diUyARp.jpg\" alt=\"Output of autoencoder, first three columns are ground truth, middle is encoded space, last three are predictions\"><\/p>\n\n<p>Sorry for no labels. The first three columns are the RGB ground truths, the middle column is the encoded space and the last three are the predictions. As you can see, only the middle band is being correctly predicted, which the R and B bands are completely off. This happens regardless of the order I put the bands in, and if I train on just one band I get pretty OK results. I'm expecting I should be able to predict all three bands of the cutouts, but can't seem to get there.<\/p>\n\n<p>I appreciate this is a complicated problem and appreciate all the help y'all can provide, and I'm happy to give any more info that you might need.<\/p>\n","answers":[]}
+{"QuestionId":57510105,"AnswerCount":0,"Tags":"<python><tensorflow><pytorch>","CreationDate":"2019-08-15T13:03:52.000","AcceptedAnswerId":null,"OwnerUserId":11931779.0,"Title":"PyTorch equivalent of tf.dynamic_partition","Body":"<p>I need to use an equivalent of <code>tf.dynamic_partition<\/code> in PyTorch. Is there anything with similar functionality in PyTorch or other library or is there a simple and clever way to code it for PyTorch and work fast?<\/p>\n","answers":[]}
+{"QuestionId":57510115,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-08-15T13:04:30.747","AcceptedAnswerId":null,"OwnerUserId":458105.0,"Title":"cifar100 with MobiletNetV2","Body":"<p>I am trying to train MobileNetV2 on CIFAR100 using keras.applications\nHere is my code:<\/p>\n\n<pre><code>(x_train,y_train),(x_test,y_test) = tf.keras.datasets.cifar100.load_data(label_mode='fine')\n\nx_test = x_test.astype(\"float32\")\nx_train = x_train.astype(\"float32\")\n\nx_test \/=255\nx_train \/=255\n\ny_test = tf.keras.utils.to_categorical(y_test,100)\ny_train = tf.keras.utils.to_categorical(y_train,100)\nmodel = MobileNetV2(input_shape=(32,32,3),\n                    alpha=1.0,\n                    include_top=True,\n                    weights=None,\n                    classes=100)\nepochs = 200\nbatch_size = 64\nprint('Using real-time data augmentation.')\n    # This will do preprocessing and realtime data augmentation:\ndatagen = 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        zca_epsilon=1e-06,  # epsilon for ZCA whitening\n        rotation_range=0,  # randomly rotate images in the range (degrees, 0 to 180)\n        # randomly shift images horizontally (fraction of total width)\n        width_shift_range=0.1,\n        # randomly shift images vertically (fraction of total height)\n        height_shift_range=0.1,\n        shear_range=0.,  # set range for random shear\n        zoom_range=0.,  # set range for random zoom\n        channel_shift_range=0.,  # set range for random channel shifts\n        # set mode for filling points outside the input boundaries\n        fill_mode='nearest',\n        cval=0.,  # value used for fill_mode = \"constant\"\n        horizontal_flip=True,  # randomly flip images\n        vertical_flip=False,  # randomly flip images\n        # set rescaling factor (applied before any other transformation)\n        rescale=None,\n        # set function that will be applied on each input\n        preprocessing_function=None,\n        # image data format, either \"channels_first\" or \"channels_last\"\n        data_format=None,\n        # fraction of images reserved for validation (strictly between 0 and 1)\n        validation_split=0.0)\ndatagen.fit(x_train)\n\nmodel.compile(optimizer='adam', loss=tf.keras.losses.categorical_crossentropy, metrics=['acc'])\nhistory = model.fit_generator(datagen.flow(x_train, y_train,batch_size=batch_size),\n                        epochs=epochs,\n                        validation_data=(x_test, y_test)\n                             )\n<\/code><\/pre>\n\n<p>The issue is with the validation accuracy, after 200 epochs the acc is almost 40%. \nI tried to fine_tune the optimizer\/loss params but still the same. My guess is the dim of the input is too small for the model as the default is 224*224, however according to the documentation you could use whatever you want!<\/p>\n\n<p>Any advice? (I do not want to change the dim of cifar100 to 224*224 because of some assumptions related to this experiment)!<\/p>\n","answers":[{"AnswerId":"57510243","CreationDate":"2019-08-15T13:15:58.723","ParentId":null,"OwnerUserId":"3742686","Title":null,"Body":"<p>Some things that come to my mind...<\/p>\n\n<ul>\n<li>Check the data augmentation pipeline (<code>datagen<\/code>). It may be distorting the input too much and hence the model may be learning weird stuff instead of learning to classify the images<\/li>\n<li>Check also the training accuracy... Is it better than the validation? by how much? <\/li>\n<li>For me, 32x32 is small but I think you should even though get higher accuracy...<\/li>\n<\/ul>\n"}]}
+{"QuestionId":57510608,"AnswerCount":2,"Tags":"<tensorflow><keras><conv-neural-network>","CreationDate":"2019-08-15T13:44:31.090","AcceptedAnswerId":null,"OwnerUserId":9437859.0,"Title":"Cant build a CNN with keras for vectors - problem with dimensions","Body":"<p>Let us say that I build an extreamly simple CNN with Keras to classify vectors.<\/p>\n\n<p>My input (X_train) is a matrix in which each row is a vector and each column is a feature. My input labels (y_train) is matrix where each line is a one hot encoded vector. This is a binary classifier. <\/p>\n\n<p>my CNN is built as follows:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv1D(64,3))\nmodel.add(Activation('relu'))\nmodel.add(Flatten())\nmodel.add(Dense(64))\nmodel.add(Activation('relu'))\nmodel.add(Dense(2))\nmodel.add(Activation('sigmoid'))\nmodel.compile(loss = 'binary_crossentropy', optimizer = 'adam', matrics = \n['accuracy'])\nmodel.fit(X_train,y_train,batch_size = 32)\n<\/code><\/pre>\n\n<p>But when I try to run this code, I get back this error message:<\/p>\n\n<pre><code>Input 0 is incompatible with layer conv1d_23: expected ndim=3, found \nndim=2\n<\/code><\/pre>\n\n<p>why would keras expect 3 dims? one dim for samples, and one for features. And more importantly, how can I fix this?<\/p>\n","answers":[{"AnswerId":"57510811","CreationDate":"2019-08-15T13:58:35.107","ParentId":null,"OwnerUserId":"10455493","Title":null,"Body":"<p>X_train is suppose to have the shape: (batch_size, steps, input_dim), <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/Conv1D#input_shape\" rel=\"nofollow noreferrer\">see documentation<\/a>. It seems like you are missing one of the dimensions.\nI would guess input_dim in your case is 1 and that is why it is missing. If so, change the <\/p>\n\n<p><code>model.fit<\/code> <\/p>\n\n<p>line to <\/p>\n\n<p><code>model.fit(tf.expand_dims(X_train,-1), y_train,batch_size = 32)<\/code><\/p>\n\n<p>Your code is not a minimal working example, so I am not able to verify if that is the only problem, but this should hopefully fix your current error message.<\/p>\n"},{"AnswerId":"57510879","CreationDate":"2019-08-15T14:02:58.607","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>A <code>Conv1D<\/code> layer expects an input with shape <code>(samples, width, channels)<\/code>, so this does not match your input data, producing an error.<\/p>\n\n<p>The convolution operation is done on the <code>width<\/code> dimension, so assuming that you want to do convolution on what you call features, then you should reshape your data to add a dummy <code>channels<\/code> dimension with a value of one:<\/p>\n\n<pre><code>X_train = X_train.reshape((X_train.shape[0], X_train.shape[1], 1))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57510987,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><tflearn>","CreationDate":"2019-08-15T14:11:47.683","AcceptedAnswerId":null,"OwnerUserId":9096388.0,"Title":"Question about using non-binary labels with TFLearn","Body":"<p>I am trying to code a neural network with TFLearn in python 3, and I have a problem related to the labels. The input to the neural network are 1D vectors of length 11. The labels that correspond to these inputs are also 1D vectors of length 11. However, their values are not 0 and 1. Typically in most examples, labels consist of [1,0] or [0,1] as in a case when classifying images of cats and dogs. However, my situation is more unique, and I need to use label vectors that instead look like [4,9,11,2,1,6,4,6,10,1,0]. There are many different label types instead of just 2 as in typical examples, and I can't reencode my labels in terms of vectors consisting of 0's and 1's. My problem is that when use labels in the form shown it doesn't seem to work. Basically, I want to know why this is and how to make TFLearn handle this properly. When I train the neural network I get an accuracy of about 78% typically. However, when I then try the .predict() function it outputs a vector that consists of all values less than 1. For example, it might output [0.3747585,0.0034589,0.06857,0.00384,0.001133,0.04858,0.00045737,0.0068587,0.00100102,0.00066543,0.033322].<\/p>\n\n<p>The values in my label vectors are always integers between 0 and 11. So I was expecting an output that also had values in this range but instead it outputs values between 0 and 1. How do I fix this? So far I've tried dividing every value in my labels by 11 such that [4,9,11,2,1,6,4,6,10,1,0] would become [4\/11,9\/11,11\/11,2\/11,1\/11,6\/11,4\/11,6\/11,10\/11,1\/11,0\/11]. However, this doesn't seem to work either. I still get an accuracy result of about 78% and a 0 to 1 output vector, but its behaving like its overfitting. I'm not sure if I am doing something wrong programmatically, using the wrong procedure, or if its really overfitting. Below is the TFLearn code that I use.   <\/p>\n\n<pre><code>#NN starts\nnet = tflearn.input_data(shape=[None, len(input[0])])\nnet = tflearn.fully_connected(net, 32)\nnet = tflearn.fully_connected(net, 32)\nnet = tflearn.fully_connected(net, len(labels[0]), activation='softmax')\nnet = tflearn.regression(net)\n\n# Define model\nmodel = tflearn.DNN(net)\n\n# Start training (apply gradient descent algorithm)\nmodel.fit(input, labels, n_epoch=10, batch_size=16, show_metric=True)\n\n#Predict\npred = model.predict(testvector)\nprint('output = ', pred)`\n<\/code><\/pre>\n","answers":[{"AnswerId":"57552759","CreationDate":"2019-08-19T08:11:55.327","ParentId":null,"OwnerUserId":"7129723","Title":null,"Body":"<p>The reason for getting output prediction values between <strong>0 to 1<\/strong> is that you are using <strong>softmax<\/strong> activation function. <a href=\"https:\/\/medium.com\/data-science-bootcamp\/understand-the-softmax-function-in-minutes-f3a59641e86d\" rel=\"nofollow noreferrer\">Softmax function<\/a> assigns a probability to each output node in the range of [0,1] that adds up to 1. So you wont get any integer value as your output. Normalization (i.e: dividing by 11) won't work either since each output value is independent of each other.<\/p>\n\n<p>What you can do is convert your problem into <a href=\"https:\/\/www.depends-on-the-definition.com\/guide-to-multi-label-classification-with-neural-networks\/\" rel=\"nofollow noreferrer\">multiclass-multilabel classification<\/a> problem in the following way:<\/p>\n\n<ol>\n<li><p>For your output label convert each number into binary consisting of four digits (Since maximum value is 11 which requires 4 binary digit). e.g: [4, 9, ...] becomes [0100, 1001, ...]<\/p><\/li>\n<li><p>Take each binary digit as a class. So your problem turns into a classification problem. e.g: [0100, 1001, ...] becomes [0, 1, 0, 0, 1, 0, 0, 1, ...] where every four digit will represent a number of your actual output.<\/p><\/li>\n<li><p>This way convert your training and testing dataset label into array of binary digit.<\/p><\/li>\n<li><p>Then in your model, instead of using <strong>softmax<\/strong> activation function you can use <strong>sigmoid<\/strong> activation function. <strong>Sigmoid<\/strong> function will assign a probability for each output node in the range of [0,1] (The summation of the probability is not necessarily 1). So, after training your model, your output of prediction should look something like this: [0.0012, 0.7890, 0.0001, 0.0100, 0.8801, 0.0030, 0.0440, 0.9120, ...].<\/p><\/li>\n<li><p>Choose an appropriate threshold value to convert your prediction into binary digits. e.g: if we choose 0.5 as the threshold then the output becomes: [0, 1, 0, 0, 1, 0, 0, 1, ...] and then you can consider every four digit as a single number for actual output label.<\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":57511185,"AnswerCount":1,"Tags":"<xcode><pytorch>","CreationDate":"2019-08-15T14:28:07.937","AcceptedAnswerId":null,"OwnerUserId":8863209.0,"Title":"How can I deploy my Pytorch model into IOS?","Body":"<p>I have a deep learning neural network I built on Pytorch I am seeking to deploy onto IOS. <\/p>\n","answers":[{"AnswerId":"57513634","CreationDate":"2019-08-15T17:21:48.170","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>Native support doesn't exist still I think, but what some do is to export the ONNX model and then open this in Caffe2 which has the support for IOS device (also Androids)<\/p>\n\n<p>So use <a href=\"https:\/\/pytorch.org\/tutorials\/advanced\/super_resolution_with_caffe2.html\" rel=\"nofollow noreferrer\">ONNX export tutorial<\/a> and this <a href=\"https:\/\/caffe2.ai\/docs\/mobile-integration.html\" rel=\"nofollow noreferrer\">mobile integration helper<\/a>.<\/p>\n\n<p>There is also a path <a href=\"https:\/\/pytorch.org\/tutorials\/advanced\/ONNXLive.html#convert-the-onnx-models-to-coreml-models\" rel=\"nofollow noreferrer\">converting ONNX to CoreML<\/a> but depending on your project it may not be super fast.<\/p>\n\n<p>There is also an option to port the ONNX to TensorFlow Lite and to <a href=\"https:\/\/www.tensorflow.org\/lite\/guide\/ios\" rel=\"nofollow noreferrer\">integrate<\/a> with your Swift or Objective C app from there on.<\/p>\n"}]}
+{"QuestionId":57511329,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras><conv-neural-network>","CreationDate":"2019-08-15T14:36:38.550","AcceptedAnswerId":null,"OwnerUserId":6839670.0,"Title":"ValueError: Error when checking target: expected conv2d_37 to have shape (57, 57, 16) but got array with shape (120, 120, 3)","Body":"<p>my training variable shape is (264, 120, 120, 3)\ntrying to give numpy array of images as input <\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv2D(8, (3, 3), activation='relu', strides=2,input_shape=(image_height,image_width,channels)))\n\nmodel.add(Conv2D(16, (3, 3), activation='relu'))\nmodel.summary()\n\nmodel.compile(optimizer='rmsprop', loss='mse')\nmodel.fit(x=X_train, y=y_train, batch_size=1, epochs=1, verbose=1)\n<\/code><\/pre>\n\n<p>below is the error message <\/p>\n\n<pre><code>________________________________________________________________\n    Layer (type)                 Output Shape              Param \n=================================================================\nconv2d_36 (Conv2D)           (None, 59, 59, 8)         224       \n_________________________________________________________________\nconv2d_37 (Conv2D)           (None, 57, 57, 16)        1168      \n=================================================================\n\n\nTotal params: 1,392\nTrainable params: 1,392\nNon-trainable params: 0\nValueError: Error when checking target: expected conv2d_37 to have shape (57, 57, 16) but got array with shape (120, 120, 3)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57511431,"AnswerCount":1,"Tags":"<python><keras><lstm><recurrent-neural-network><named-entity-recognition>","CreationDate":"2019-08-15T14:43:39.490","AcceptedAnswerId":null,"OwnerUserId":5397172.0,"Title":"Bi-directional LSTM for entity recognition","Body":"<p>Following a paper, I'm using word embeddings as a feature vector for entity recognition. <\/p>\n\n<p>I've attempted to architect the network using Keras but have run into a dimensionality problem I cannot seem to resolve. <\/p>\n\n<p>Take the following example sentence: <\/p>\n\n<p><code>[\"I went to the shop\"]<\/code><\/p>\n\n<p>The sentence has 5 words, and after computing the feature matrix, I am left with a matrix of dimension: <code>(1, 120, 1000) == (#examples, sequence_length, embedding)<\/code>.<\/p>\n\n<p>Note that <code>sequence_length<\/code> appends <code>0.<\/code> padding when not complete. In this example, the actual <code>sequence_length<\/code> would be 5. <\/p>\n\n<p>My network architecture is as follows: <\/p>\n\n<pre><code>enc = encode() \nclaims_input = Input(shape=(120, 1000), dtype='float32', name='claims')\nx = Masking(mask_value=0., input_shape=(120, 1000))(claims_input)\nx = Bidirectional(LSTM(units=512, return_sequences=True, recurrent_dropout=0.2, dropout=0.2))(x)\nx = Bidirectional(LSTM(units=512, return_sequences=True, recurrent_dropout=0.2, dropout=0.2))(x)\nout = TimeDistributed(Dense(8, activation=\"softmax\"))(x)\n\nmodel = Model(inputs=claims_input, output=out)\nmodel.compile(loss=\"sparse_categorical_crossentropy\", optimizer='adam', metrics=[\"accuracy\"])\nmodel.fit(enc, y)\n<\/code><\/pre>\n\n<p>The architecture is straight forward, I mask specific time steps, run two bidirectional LSTMs, followed by a softmax output. My <code>y<\/code> variable in this case, is a <code>(9,8)<\/code> one-hot-encoded matrix corresponding to the gold label of each word. <\/p>\n\n<p>When trying to <code>fit()<\/code> this model, I am running into a dimensionality problem relating to the <code>TimeDistributed()<\/code> layer and I'm unsure how to resolve, or even begin to debug this. <\/p>\n\n<p>Error: <code>ValueError: Error when checking target: expected time_distributed_1 to have 3 dimensions, but got array with shape (9, 8)<\/code> <\/p>\n\n<p>Any help would be appreciated.<\/p>\n","answers":[{"AnswerId":"57512245","CreationDate":"2019-08-15T15:37:46.107","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You are doing entity recognition. So each element in your input sequence will be assigned an entity (probably some of them as null). If your model takes an input sample of shape <code>(120, n_features)<\/code>, then the output must also be a sequence of length of 120, i.e. one entity for each element. Therefore, the labels, i.e. <code>y<\/code>, you provide to the model must have a shape of <code>(n_samples, 120, n_entities)<\/code> (or <code>(n_samples, 120, 1)<\/code> if you are using sparse labeling).<\/p>\n\n<p><strong>Side note:<\/strong> There is no difference between <code>TimeDistributed(Dense(...))<\/code> and <code>Dense(...)<\/code>, as the <a href=\"https:\/\/stackoverflow.com\/a\/52092176\/2099607\"><code>Dense<\/code> layer is applied on the last axis<\/a>. <\/p>\n"}]}
+{"QuestionId":57511784,"AnswerCount":0,"Tags":"<python><machine-learning><deep-learning><pytorch>","CreationDate":"2019-08-15T15:07:35.093","AcceptedAnswerId":null,"OwnerUserId":3167448.0,"Title":"How does one implement convolutional GRUs\/LSTMs\/RNNs in Pytorch?","Body":"<p>I want to have the standard LSTM\/GRU\/RNN set up but swap the linear function with a convolution. Is that possible to do in Pytorch in an clean and efficient manner?<\/p>\n\n<p>Ideally it still works with packing, varying sequence length etc.<\/p>\n\n<p>Small sample code of a trivial way to pass data through it would be super useful like:<\/p>\n\n<pre><code># Based on Robert Guthrie tutorial\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.optim as optim\n\nfrom pdb import set_trace as st\n\ntorch.manual_seed(1)\n\ndef step_by_step(net,sequence,hidden):\n    '''\n    an example of an LSTM processing all the sequence one token at a time (one time step at a time)\n    '''\n    ## process sequence one element at a time\n    print()\n    print('start processing sequence')\n    for i, token in enumerate(sequence):\n        print(f'-- i = {i}')\n        #print(f'token.size() = {token.size()}')\n        ## to add fake batch_size and fake seq_len\n        h_n, c_n = hidden # hidden states, cell state\n        processed_token = token.view(1, 1, -1) # torch.Size([1, 1, 3])\n        print(f'processed_token.size() = {processed_token.size()}')\n        print(f'h_n.size() = {h_n.size()}')\n        #print(f'processed_token = {processed_token}')\n        #print(f'h_n = {h_n}')\n        # after each step, hidden contains the hidden state.\n        out, hidden = lstm(processed_token, hidden)\n    ## print results\n    print()\n    print(out)\n    print(hidden)\n\ndef whole_seq_all_at_once(lstm,sequence,hidden):\n    '''\n        alternatively, we can do the entire sequence all at once.\n        the first value returned by LSTM is all of the hidden states throughout\n        #the sequence.\n        The second is just the most recent hidden state\n        (compare the last slice of \"out\" with \"hidden\" below, they are the same)\n        The reason for this is that:\n        \"out\" will give you access to all hidden states in the sequence\n        \"hidden\" will allow you to continue the sequence and backpropagate,\n        by passing it as an argument  to the lstm at a later time\n        Add the extra 2nd dimension\n    '''\n    h, c = hidden\n    Tx = len(sequence)\n    ## concatenates list of tensors in the dim 0, i.e. stacks them downwards creating new rows\n    sequence = torch.cat(sequence) # (5, 3)\n    ## add a singleton dimension of size 1\n    sequence = sequence.view(len(sequence), 1, -1) # (5, 1, 3)\n    print(f'sequence.size() = {sequence.size()}')\n    print(f'h.size() = {h.size()}')\n    print(f'c.size() = {c.size()}')\n    out, hidden = lstm(sequence, hidden)\n    ## \"out\" will give you access to all hidden states in the sequence\n    print()\n    print(f'out = {out}')\n    print(f'out.size() = {out.size()}') # (5, 1, 25)\n    ##\n    h_n, c_n = hidden\n    print(f'h_n = {h_n}')\n    print(f'h_n.size() = {h_n.size()}')\n    print(f'c_n = {c_n}')\n    print(f'c_n.size() = {c_n.size()}')\n\nif __name__ == '__main__':\n    ## model params\n    hidden_size = 6\n    input_size = 3\n    lstm = nn.LSTM(input_size=input_size,hidden_size=hidden_size)\n    ## make a sequence of length Tx (list of Tx tensors)\n    Tx = 5\n    sequence = [torch.randn(1, input_size) for _ in range(Tx)]  # make a sequence of length 5\n    ## initialize the hidden state.\n    hidden = (torch.randn(1, 1, hidden_size), torch.randn(1, 1, hidden_size))\n    #step_by_step(lstm,sequence,hidden)\n    whole_seq_all_at_once(lstm,sequence,hidden)\n    print('DONE \\a')\n<\/code><\/pre>\n\n<hr>\n\n<p>Cross posted:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/57511784\/how-does-one-implement-convolutional-grus-lstms-rnns-in-pytorch\">How does one implement convolutional GRUs\/LSTMs\/RNNs in Pytorch?<\/a><\/li>\n<li><a href=\"https:\/\/discuss.pytorch.org\/t\/how-to-implement-convolutional-grus-lstms-rnns\/53471\" rel=\"nofollow noreferrer\">https:\/\/discuss.pytorch.org\/t\/how-to-implement-convolutional-grus-lstms-rnns\/53471<\/a><\/li>\n<\/ul>\n","answers":[]}
+{"QuestionId":57512113,"AnswerCount":2,"Tags":"<pytorch>","CreationDate":"2019-08-15T15:29:50.813","AcceptedAnswerId":57524381.0,"OwnerUserId":3259896.0,"Title":"How to take a transpose for each matrix in a batch in Pytorch?","Body":"<p>Say I have a 4 batch of 5x3 matrixes. So the dimensions of these tensor are 4x5x3. How do I take the transpose of each matrix within each batch. So converting it to 4x3x5?<\/p>\n","answers":[{"AnswerId":"57524381","CreationDate":"2019-08-16T12:07:10.543","ParentId":null,"OwnerUserId":"7347631","Title":null,"Body":"<p>I will drop some benchmarks here for the sake of performance. Using the same tensor proposed in the OP's answer.<\/p>\n\n<pre><code>In[2]: import torch\nIn[3]: x = torch.randn(2, 3, 5)\nIn[4]: x.size()\nOut[4]: torch.Size([2, 3, 5])\nIn[5]: %timeit x.permute(1, 0, 2)\n1.03 \u00b5s \u00b1 41.7 ns per loop (mean \u00b1 std. dev. of 7 runs, 1000000 loops each)\nIn[6]: %timeit torch.transpose(x, 0, 1)\n892 ns \u00b1 9.61 ns per loop (mean \u00b1 std. dev. of 7 runs, 1000000 loops each)\nIn[7]: torch.transpose(x, 0, 1).equal(x.permute(1, 0, 2))\nOut[7]: True\n<\/code><\/pre>\n\n<p>It is clear that <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.transpose\" rel=\"nofollow noreferrer\"><code>torch.transpose<\/code><\/a> is faster, so It is advised to use it when possible. <\/p>\n"},{"AnswerId":"57512196","CreationDate":"2019-08-15T15:34:55.163","ParentId":null,"OwnerUserId":"3259896","Title":null,"Body":"<p><a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html?highlight=torch%20transpose#torch.transpose\" rel=\"nofollow noreferrer\"><code>torch.transpose(input, dim0, dim1)<\/code><\/a>  will swap two dimensions. <\/p>\n\n<pre><code>&gt;&gt;&gt; x = torch.randn(2, 3)\n&gt;&gt;&gt; x\ntensor([[ 1.0028, -0.9893,  0.5809],\n        [-0.1669,  0.7299,  0.4942]])\n&gt;&gt;&gt; torch.transpose(x, 0, 1)\ntensor([[ 1.0028, -0.1669],\n        [-0.9893,  0.7299],\n        [ 0.5809,  0.4942]])\n<\/code><\/pre>\n\n<p>If more dimensions are desired to be swapped, use <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html?highlight=torch%20permute#torch.Tensor.permute\" rel=\"nofollow noreferrer\"><code>torch.Tensor.permute<\/code><\/a><\/p>\n\n<pre><code>&gt;&gt;&gt; x = torch.randn(2, 3, 5)\n&gt;&gt;&gt; x.size()\ntorch.Size([2, 3, 5])\n&gt;&gt;&gt; x.permute(2, 0, 1).size()\ntorch.Size([5, 2, 3])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57512122,"AnswerCount":1,"Tags":"<python><flask><keras><deployment>","CreationDate":"2019-08-15T15:30:14.603","AcceptedAnswerId":58491243.0,"OwnerUserId":6482145.0,"Title":"how to serve keras model for inference using flask?","Body":"<p>I have a inference object detection code of yolo v3 keras model<\/p>\n\n<pre><code>#! \/usr\/bin\/env python\n\nimport os\nimport argparse\nimport json\nimport cv2\nfrom utils.utils import get_yolo_boxes, makedirs\nfrom utils.bbox import draw_boxes\nfrom keras.models import load_model\nfrom tqdm import tqdm\nimport numpy as np\nimport flask\nimport io\nfrom PIL import Image\nfrom keras.preprocessing.image import img_to_array\n\nconfig_path  = \"config.json\"\ninput_path   = \"test.jpg\"\noutput_path  = \"output\"\n\nwith open(config_path) as config_buffer:    \n        config = json.load(config_buffer)\n\nmakedirs(output_path)\n\nnet_h, net_w = 416, 416 \nobj_thresh, nms_thresh = 0.5, 0.45\n\nos.environ['CUDA_VISIBLE_DEVICES'] = config['train']['gpus']\ninfer_model = load_model(config['train']['saved_weights_name'])\n\n\nimage = cv2.imread(input_path)\n\n# predict the bounding boxes\nboxes = get_yolo_boxes(infer_model, [image], net_h, net_w, config['model']['anchors'], obj_thresh, nms_thresh)[0]\n\n# draw bounding boxes on the image using labels\n_,outputs =  draw_boxes(image, boxes, config['model']['labels'], obj_thresh) \n\nprint(outputs)\n\n# write the image with bounding boxes to file\ncv2.imwrite(output_path + input_path.split('\/')[-1], np.uint8(image))\n<\/code><\/pre>\n\n<p>This is working completely fine ,giving the expected output class &amp; co-ordinate in terminal<\/p>\n\n<blockquote>\n  <p>{'classes': 'person 99.97%', 'X2': '389', 'X1': '174', 'Y1': '8',\n  'Y2': '8'}<\/p>\n<\/blockquote>\n\n<p>But when I convert that  for REST api based for serving by converting above code using official keras conversion to flask documentation ,like this :<\/p>\n\n<pre><code>#! \/usr\/bin\/env python\n\nimport os\nimport argparse\nimport json\nimport cv2\nfrom utils.utils import get_yolo_boxes, makedirs\nfrom utils.bbox import draw_boxes\nfrom keras.models import load_model\nfrom tqdm import tqdm\nimport numpy as np\nimport flask\nimport io\nfrom PIL import Image\nfrom keras.preprocessing.image import img_to_array\n\n\nconfig_path  = \"config.json\"\ninput_path   = \"test.jpg\"\noutput_path  = \"output\"\n\nwith open(config_path) as config_buffer:    \n        config = json.load(config_buffer)\n\nmakedirs(output_path)\n\nnet_h, net_w = 416, 416 \nobj_thresh, nms_thresh = 0.5, 0.45\n\napp = flask.Flask(__name__)\n\nos.environ['CUDA_VISIBLE_DEVICES'] = config['train']['gpus']\ninfer_model = load_model(config['train']['saved_weights_name'])\n\ndef prepare_image(image_path):\n    image = cv2.imread(image_path)\n    return image\n\n@app.route(\"\/predict\", methods=[\"POST\"])\ndef predict():\n    # initialize the data dictionary that will be returned from the\n    # view\n    data = {\"success\": False}\n\n    # ensure an image was properly uploaded to our endpoint\n    if flask.request.method == \"POST\":\n        if flask.request.files.get(\"image\"):\n            # read the image in PIL format\n            image = flask.request.files[\"image\"].read()\n            image = Image.open(io.BytesIO(image))\n\n            # preprocess the image and prepare it for classification            \n            image = img_to_array(image)\n\n\n            boxes = get_yolo_boxes(infer_model, [image], net_h, net_w, config['model']['anchors'], obj_thresh, nms_thresh)[0]\n\n\n            _,outputs =  draw_boxes(image, boxes, config['model']['labels'], obj_thresh)\n\n\n            data.append(outputs)\n\n            print(data)\n\n\n            # indicate that the request was a success\n            data[\"success\"] = True\n\n    # return the data dictionary as a JSON response\n    return flask.jsonify(data)\n\n\n\nif __name__ == \"__main__\":\n    print((\"* Loading Keras model and Flask starting server...\"\n        \"please wait until server has fully started\"))\n    app.run()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/m6XQ1.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/m6XQ1.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Running successfully on port 5000<\/p>\n\n<p>But when I'm trying to predict via POST api using <\/p>\n\n<pre><code>curl -X POST -F image=@test.jpg 'http:\/\/localhost:5000\/predict'\n<\/code><\/pre>\n\n<p>is giving this error<\/p>\n\n<blockquote>\n<pre><code>raise ValueError(\"Tensor %s is not an element of this graph.\" % obj) ValueError: Tensor Tensor(\"conv2d_59\/BiasAdd:0\", shape=(?, ?, ?,\n<\/code><\/pre>\n  \n  <p>255), dtype=float32) is not an element of this graph.\n  127.0.0.1 - - [15\/Aug\/2019 15:11:23] \"POST \/predict HTTP\/1.1\" 500 -<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ovUIw.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ovUIw.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I don't understand why same prediction function was working without flask but getting error in this.<\/p>\n","answers":[{"AnswerId":"58491243","CreationDate":"2019-10-21T17:18:25.070","ParentId":null,"OwnerUserId":"8850702","Title":null,"Body":"<p>I was facing the same problem, this is a keras issue. Primarily seems to be triggered when when there is asynchronous event handler\nAdding <code>model._make_predict_function()<\/code> right after loading the trained model worked for me.\nFor example,<\/p>\n\n<pre><code>from keras.models import load_model\n   model=load_model('yolo.h5')\n   model._make_predict_function()\n<\/code><\/pre>\n\n<p>Another approach that has worked for other people is to use graph and have inference within the context something like:<\/p>\n\n<pre><code>global graph\ngraph = tf.get_default_graph()\nwith graph.as_default():\nres = model.predict()\n<\/code><\/pre>\n\n<p>For more insights , please refer below links:<\/p>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/6462\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/6462<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/2397\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/2397<\/a><\/p>\n"}]}
+{"QuestionId":57512129,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><tensorflow-lite>","CreationDate":"2019-08-15T15:30:32.547","AcceptedAnswerId":null,"OwnerUserId":11932316.0,"Title":"Getting Error while converting tensorflow model to tensorflow lite = ValueError: bad marshal data (unknown type code)","Body":"<p>I am using a pre-trained keras based and tensorflow based model with yolov2 architecture for potholes detection and I'm getting an error while converting my tensorflow model to tensorflow lite<\/p>\n\n<blockquote>\n  <p>Error = ValueError: bad marshal data (unknown type code)<\/p>\n<\/blockquote>\n\n<p>I'm using tensorflow 1.3 and keras 2.0.8. This is my code:<\/p>\n\n<pre><code>converter = lite.TFLiteConverter.from_keras_model_file(keras_file) \ntflite_model = converter.convert()\nopen(\"trained_wts.tflite\",\"wb\").write(tflite_model)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57572918","CreationDate":"2019-08-20T11:44:44.297","ParentId":null,"OwnerUserId":"1052083","Title":null,"Body":"<p>Please provide more information. <\/p>\n\n<p>The error was on the last line, but there was no reference on how to get <code>marshal<\/code> object and how to get <code>raw_code<\/code>.<\/p>\n"}]}
+{"QuestionId":57512215,"AnswerCount":0,"Tags":"<keras><tensorflow-serving><resnet>","CreationDate":"2019-08-15T15:36:05.563","AcceptedAnswerId":null,"OwnerUserId":11184955.0,"Title":"how to export the keras model with resnet101 to tensorflow serving with exporter.export_saved_model method","Body":"<p>I'm using Keras and resnet 101 for my training and wants to export my model into TensorFlow serving with this method exporter.export_inference_graph and it gives me the error.<\/p>\n\n<p>FailedPreconditionError (see above for traceback): Attempting to use uninitialized value conv2_block2_1_bn\/moving_variance\/local_step_1<\/p>\n","answers":[]}
+{"QuestionId":57512593,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensor>","CreationDate":"2019-08-15T16:00:18.477","AcceptedAnswerId":57513606.0,"OwnerUserId":3630804.0,"Title":"How to create a tensor with an unknown dimension","Body":"<p>I have a layer in my neural network with an output vector <code>x<\/code> of size <code>[?, N]<\/code>. (with first dimension for the batch size). I want declare a tensor of <code>ones<\/code> of the same size in the next layer (Lambda layer). I see that I cannot use <code>y = keras.backend.ones(x.shape)<\/code> as the batch size is only available in runtime. How can I create this tensor?<\/p>\n","answers":[{"AnswerId":"57513606","CreationDate":"2019-08-15T17:20:34.930","ParentId":null,"OwnerUserId":"2650622","Title":null,"Body":"<p>As suggested by today in the comments, <code>K.ones_like<\/code> works:<\/p>\n\n<pre><code>from keras import backend as K\na = K.placeholder(shape=(None, 5))\nb = K.ones_like(a)\nprint(b.shape)\n\n&gt;&gt; TensorShape([Dimension(None), Dimension(5)])\n<\/code><\/pre>\n\n<p>Depending on the type of operation you're doing, you can also make a ones tensor of shape [N] and rely on broadcasting to save memory:<\/p>\n\n<pre><code>from keras import backend as K\na = K.placeholder(shape=(None, 5))\nb = K.ones(a.shape[-1])\nprint(a + b)\n\n&gt;&gt; &lt;tf.Tensor 'add:0' shape=(?, 5) dtype=float32&gt;\n<\/code><\/pre>\n"}]}
+{"QuestionId":57512596,"AnswerCount":1,"Tags":"<python><tensorflow><keras><generator><autoregressive-models>","CreationDate":"2019-08-15T16:00:32.540","AcceptedAnswerId":57512976.0,"OwnerUserId":8534205.0,"Title":"How to generate sequence data with keras with multiple input?","Body":"<p>I'm writing a VAE for sequence to sequence problem in keras. The decoder is an auto-regressive model so I have two different inputs, one for the encoder and the same (shifted by 1, but this is not the problem) for the decoder.\nI want also to do data augmentation so I decide to use the fit_generator() method but I have some problem on returning the two inputs. <\/p>\n\n<p>I have tried to return a list of two input vector, like this<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class DataGenerator(Sequence):\n    def __init__(....\n\n    def __getitem__(self, index):\n        data = create_data()\n        return  [data, data]\n<\/code><\/pre>\n\n<p>or a dictionary like this<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>return {\"encoder_input_name\" : \"data, decoder_input_name\" : data } \n<\/code><\/pre>\n\n<p>where data is a numpy tensor of shape (batch_size, max_sequence_len, input_dimention). <\/p>\n\n<p>I can't just use the same input layer because later the two input will be a little different, as I said the decoder input will be shifted by one with a different first element and other reasons.<\/p>\n\n<p>When I return the list [data, data] or  I have 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 2 array(s), but instead got the following list of 1 arrays\n<\/code><\/pre>\n\n<p>When I return the dictionary I have this error:<\/p>\n\n<pre><code>batch_size = x.shape[0]\nAttributeError: 'str' object has no attribute 'shape'\n<\/code><\/pre>\n\n<p>How can I solve this issue?<\/p>\n\n<p>Thank you very much!<\/p>\n\n<p>EDIT<\/p>\n\n<p>I changed the output of <code>__getitem__<\/code> to \n<code>[inpuut_1, input_2], []<\/code> and it worked.<\/p>\n","answers":[{"AnswerId":"57512976","CreationDate":"2019-08-15T16:29:51.767","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You should return a <strong>tuple<\/strong> from generator\/Sequence instance. The first element of the tuple is a list of input arrays (or just one array if your model has one input layer), and the second element is a list of output arrays (or just one array if your model has one output layer).<\/p>\n\n<p>Therefore, <code>__getitem__<\/code> should return something like this:<\/p>\n\n<pre><code>def __getitem__(self, index):\n    # ...\n    return  [inp_arr1, inp_arr2, ...], [out_arr1, out_arr2, ...]  # IMPORTANT: this is a tuple\n<\/code><\/pre>\n"}]}
+{"QuestionId":57512601,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><computer-vision><tensorboard>","CreationDate":"2019-08-15T16:00:48.600","AcceptedAnswerId":null,"OwnerUserId":11264858.0,"Title":"How to get mAP Evaluation Metric and Precision-Recall curves at various IoU thresholds for Object Detection Tensorflow API","Body":"<p>I'm using Tensorflow object detection API on my own data, I have trained the Faster RCNN neural network with the pre-trained Resnet101 model.\nMy system consists of detecting products on sales flyers, therefore my configuration has only one class called product.<\/p>\n\n<p>Actually the neural network works well, it detects the products very well but I want to evaluate the performance of the neural network and obtain data such as mAP (Mean Average Precision) and Precision-Recall curves at various IoU threshold.<\/p>\n\n<p>I have tried to evaluate it but I have had some error and therefore I have not had the expected result.<\/p>\n\n<p>evaluation code<\/p>\n\n<pre><code>python object_detection\/eval.py --logtostderr --checkpoint_dir=train --eval_dir=eval --pipeline_config_path=modelo\/faster_rcnn_resnet101_coco.config\n<\/code><\/pre>\n\n<p>faster_rcnn_resnet101_coco.config<\/p>\n\n<pre><code># Faster R-CNN with Resnet-101 (v1) 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  faster_rcnn {\n    num_classes: 1\n    image_resizer {\n      keep_aspect_ratio_resizer {\n        min_dimension: 600\n        max_dimension: 1020\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  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: \"modelo\/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: 200000\n  data_augmentation_options {\n    random_horizontal_flip {\n    }\n  }\n}\n\ntrain_input_reader: {\n  tf_record_input_reader {\n    input_path: \"TFRecords\/entrenamiento.record\"\n  }\n  label_map_path: \"configuracion\/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: \"TFRecords\/test.record\"\n  }\n  label_map_path: \"configuracion\/label_map.pbtxt\"\n  shuffle: false\n  num_readers: 1\n  num_epochs: 1\n}\n<\/code><\/pre>\n\n<p>The result that I have obtained in TensorBoard<\/p>\n\n<p><a href=\"https:\/\/i.ibb.co\/ysNLCn0\/Screenshot-from-2019-08-15-18-55-03.png\" rel=\"nofollow noreferrer\">https:\/\/i.ibb.co\/ysNLCn0\/Screenshot-from-2019-08-15-18-55-03.png<\/a><\/p>\n\n<p>My question is how may I evaluate this algorithm in order to get Precision x Recall curve at various IoU thresholds and Average Precision.<\/p>\n","answers":[]}
+{"QuestionId":57512818,"AnswerCount":0,"Tags":"<python><flask><keras><theano><amazon-elastic-beanstalk>","CreationDate":"2019-08-15T16:18:12.090","AcceptedAnswerId":null,"OwnerUserId":4909076.0,"Title":"Theano and Keras: TypeError: Multiple values for input: \/dense_1_input","Body":"<p>I've been unable to find anyone else reference this exact error.<\/p>\n\n<pre><code>TypeError: Multiple values for input\n<\/code><\/pre>\n\n<p>Here is the full stack trace:<\/p>\n\n<pre><code> Traceback (most recent call last):\n   File \"\/opt\/python\/run\/venv\/local\/lib\/python3.6\/site-packages\/flask\/app.py\", line 2311, in wsgi_app\n     response = self.full_dispatch_request()\n   File \"\/opt\/python\/run\/venv\/local\/lib\/python3.6\/site-packages\/flask\/app.py\", line 1834, in full_dispatch_request\n     rv = self.handle_user_exception(e)\n   File \"\/opt\/python\/run\/venv\/local\/lib\/python3.6\/site-packages\/flask\/app.py\", line 1737, in handle_user_exception\n     reraise(exc_type, exc_value, tb)\n   File \"\/opt\/python\/run\/venv\/local\/lib\/python3.6\/site-packages\/flask\/_compat.py\", line 36, in reraise\n     raise value\n   File \"\/opt\/python\/run\/venv\/local\/lib\/python3.6\/site-packages\/flask\/app.py\", line 1832, in full_dispatch_request\n     rv = self.dispatch_request()\n   File \"\/opt\/python\/run\/venv\/local\/lib\/python3.6\/site-packages\/flask\/app.py\", line 1818, in dispatch_request\n     return self.view_functions[rule.endpoint](**req.view_args)\n   File \"\/opt\/python\/current\/app\/app\/routes.py\", line 99, in redact_string\n     redacted_description = process_string(description)\n   File \"\/opt\/python\/current\/app\/app\/routes.py\", line 704, in process_string\n     classified_sentence = predict_sentiment(encoded_sentence, tokenized_sentence)\n   File \"\/opt\/python\/current\/app\/app\/routes.py\", line 621, in predict_sentiment\n     yhat = model.predict([[x_input]], verbose=1)\n   File \"\/opt\/python\/run\/venv\/local\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\", line 1169, in predict\n     steps=steps)\n   File \"\/opt\/python\/run\/venv\/local\/lib\/python3.6\/site-packages\/keras\/engine\/training_arrays.py\", line 294, in predict_loop\n     batch_outs = f(ins_batch)\n   File \"\/opt\/python\/run\/venv\/local\/lib\/python3.6\/site-packages\/keras\/backend\/theano_backend.py\", line 1388, in __call__\n     return self.function(*inputs)\n   File \"\/opt\/python\/run\/venv\/local\/lib\/python3.6\/site-packages\/theano\/compile\/function_module.py\", line 891, in __call__\n     self.inv_finder[c]))\n TypeError: Multiple values for input: \/dense_1_input\n<\/code><\/pre>\n\n<p>I'm running Keras with a Theano backend on Flask on AWS Elastic Beanstalk.<\/p>\n\n<p>Does anyone know what this might related to.<\/p>\n\n<p>I don't think my inputs are different sizes because they all go through the same encoding function.<\/p>\n\n<p>Wondering if somehow it could be related to the same instance of the model being shared across multiple threads..<\/p>\n","answers":[]}
+{"QuestionId":57513032,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><anaconda><jupyter>","CreationDate":"2019-08-15T16:33:22.697","AcceptedAnswerId":null,"OwnerUserId":11851100.0,"Title":"Python on anaconda\/ Jupyter notebook not respond to 'import tensorflow' (no errors)","Body":"<p>I am new to TensorFlow. I trained several models past days using tensorFlow-gpu by using GPU (using CUDA) successfully. But when I checked recently, python on anaconda not respond to 'import tensorflow'. It takes much time and exit python on conda prompt. No errors are displayed.<\/p>\n\n<p>I checked import tensorflow in the jupyter notebook but it did not give an output. \nI also tried conda install tensorflow-gpu again, it says that i have tensorflow-gpu in my environment.\nI have deleted cache on the environment by conda clean -a\nbut coud not help.<\/p>\n\n<p>No error messages. it just exits on the python and returns to the conda prompt without any messages. Due to this reason, I even can not find what the problem is.\n<a href=\"https:\/\/i.stack.imgur.com\/GyV4r.jpg\" rel=\"nofollow noreferrer\">python on anaconda prompt when import tensorflow<\/a><\/p>\n","answers":[]}
+{"QuestionId":57513409,"AnswerCount":0,"Tags":"<python><tensorflow><gradient><reinforcement-learning><eager-execution>","CreationDate":"2019-08-15T17:03:54.053","AcceptedAnswerId":null,"OwnerUserId":2451456.0,"Title":"Tensorflow with Eager Execution: DDPG - Apply Action Gradient to Actor","Body":"<p>I am struggling to apply the Tensorflow eager execution (TF2) in order to train the actor of the actor-critic <a href=\"https:\/\/arxiv.org\/pdf\/1509.02971.pdf\" rel=\"nofollow noreferrer\">DDPG algorithm<\/a>. The description in <a href=\"https:\/\/nervanasystems.github.io\/coach\/components\/agents\/policy_optimization\/ddpg.html\" rel=\"nofollow noreferrer\">this example<\/a> explains:<\/p>\n\n<blockquote>\n  <p>Use the actor\u2019s online network to get the action mean values using the current states as the inputs. Then, use the critic online network in order to get the gradients of the critic output with respect to the action mean values \u2207aQ(s,a) | s=s_t,a=\u03bc(s_t). Using the chain rule, calculate the gradients of the actor\u2019s output, with respect to the actor weights, given \u2207aQ(s,a). Finally, apply those gradients to the actor network.<\/p>\n<\/blockquote>\n\n<p>Actor and Critic are both Keras models. Apologies for not posting the complete code, though I hope my problem is comprehensible from these relevant snippets.<\/p>\n\n<pre><code>def fit_actor(self, state):\n\n    action = self.predict_actor(state)              #online actor\n    q_value = self.predict_critic(state, action)    #online critic\n\n    param_gradient = self.tape_online_critic.gradient(q_value, [action]) \n    #=&gt; is alway [None], also for tape_online_actor\n\n    gradient = zip(param_gradient, self.online_actor.trainable_weights)\n\n    self.optimizer_actor.apply_gradients(gradient)\n<\/code><\/pre>\n\n<p>and in 'predict_critic' the critic is fed with the state and action so 'tape_online_critic' has operations<\/p>\n\n<pre><code>#for online critic and online actor\ndef predict_critic(self, state, actions):\n\n    with self.tape_online_critic as tape:\n        return self.online_critic([state, actions])\n\n\ndef predict_actor(self, state):\n\n    with self.tape_online_actor as tape:\n        return self.online_actor([state])\n<\/code><\/pre>\n\n<p>I tried almost every thinkable combination of variables\/tapes etc. but I always a gradient of <strong>[None]<\/strong> and a <strong>ValueError<\/strong>.<\/p>\n\n<pre><code>ValueError: No gradients provided for any variable: ['conv2d_2\/kernel:0']\n<\/code><\/pre>\n\n<ol>\n<li>I am unsure about the order of the arguments for the <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/GradientTape#gradient\" rel=\"nofollow noreferrer\">tape.gradient()<\/a> function or even the nature of the arguements itself. <\/li>\n<li>As I understand, I have to call the <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/GradientTape#gradient\" rel=\"nofollow noreferrer\">tape.gradient()<\/a> function on the tape of the online actor in order to get the gradient for the weights of the actor with respect to the operations recorded in the predict_actor function within the with tape: context. Do I additionally have to <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/GradientTape#watch\" rel=\"nofollow noreferrer\">tape.watch(<em>someTensor<\/em>)<\/a> something there? The <a href=\"https:\/\/www.tensorflow.org\/guide\/eager\" rel=\"nofollow noreferrer\">documentation<\/a> says variables are automatically watched, so I'd assume no.<\/li>\n<li>Is tf.GradientTape even the right thing to use here? There is also tfe.gradients_function?<\/li>\n<\/ol>\n","answers":[]}
+{"QuestionId":57513715,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><deep-learning><generative-adversarial-network>","CreationDate":"2019-08-15T17:29:05.373","AcceptedAnswerId":null,"OwnerUserId":7717192.0,"Title":"DC-GAN: Discriminator loss going up while generator loss goes down","Body":"<p>I cannot tell if this error is due to a technical mistake or hyper-parameters, but my DC-GAN's discriminator loss starts low and gradually climbs higher, slowing down around 8, whereas my generator loss goes way down. I ended it at about 60,000 epochs. Funny enough, the discriminator accuracy seems to be floating around 20-50%. Does anybody have any suggestions to fix the problem? Any help is appreciated.<\/p>\n\n<p><strong>Important Info<\/strong><\/p>\n\n<ul>\n<li>Data Format: 472 320x224 Color PNG files.<\/li>\n<li>Optimizer: <code>Adam(0.0002, 0.5)<\/code><\/li>\n<li>Loss: Binary Cross-Entropy<\/li>\n<\/ul>\n\n<p><strong>A generated image after 50,000+ epochs: (Supposed to be a sneaker on a white background)<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/aX5il.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/aX5il.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><strong>Discriminator Model:<\/strong><\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>    def build_discriminator(self):\n\n        img_shape = (self.img_size[0], self.img_size[1], self.channels)\n\n        model = Sequential()\n\n        model.add(Conv2D(32, kernel_size=self.kernel_size, strides=2, input_shape=img_shape, padding=\"same\"))  # 192x256 -&gt; 96x128\n        model.add(LeakyReLU(alpha=0.2))\n        model.add(Dropout(0.25))\n\n        model.add(Conv2D(64, kernel_size=self.kernel_size, strides=2, padding=\"same\"))  # 96x128 -&gt; 48x64\n        model.add(ZeroPadding2D(padding=((0, 1), (0, 1))))\n        model.add(LeakyReLU(alpha=0.2))\n        model.add(Dropout(0.25))\n        model.add(BatchNormalization(momentum=0.8))\n\n        model.add(Conv2D(128, kernel_size=self.kernel_size, strides=2, padding=\"same\"))  # 48x64 -&gt; 24x32\n        model.add(LeakyReLU(alpha=0.2))\n        model.add(Dropout(0.25))\n        model.add(BatchNormalization(momentum=0.8))\n\n        model.add(Conv2D(256, kernel_size=self.kernel_size, strides=1, padding=\"same\"))  # 24x32 -&gt; 12x16\n        model.add(LeakyReLU(alpha=0.2))\n        model.add(Dropout(0.25))\n\n        model.add(Conv2D(512, kernel_size=self.kernel_size, strides=1, padding=\"same\"))  # 12x16 -&gt; 6x8\n        model.add(LeakyReLU(alpha=0.2))\n        model.add(Dropout(0.25))\n\n        model.add(Flatten())\n        model.add(Dense(1, activation='sigmoid'))\n\n        model.summary()\n\n        img = Input(shape=img_shape)\n        validity = model(img)\n\n        return Model(img, validity)\n<\/code><\/pre>\n\n<p><strong>Generator Model:<\/strong><\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>    def build_generator(self):\n\n        noise_shape = (100,)\n\n        model = Sequential()\n        model.add(\n            Dense(self.starting_filters * (self.img_size[0] \/\/ (2 ** self.upsample_layers))  *  (self.img_size[1] \/\/ (2 ** self.upsample_layers)),\n                  activation=\"relu\", input_shape=noise_shape))\n        model.add(Reshape(((self.img_size[0] \/\/ (2 ** self.upsample_layers)),\n                           (self.img_size[1] \/\/ (2 ** self.upsample_layers)),\n                           self.starting_filters)))\n        model.add(BatchNormalization(momentum=0.8))\n\n        model.add(UpSampling2D())  # 6x8 -&gt; 12x16\n        model.add(Conv2D(1024, kernel_size=self.kernel_size, padding=\"same\"))\n        model.add(Activation(\"relu\"))\n        model.add(BatchNormalization(momentum=0.8))\n\n        model.add(UpSampling2D())  # 12x16 -&gt; 24x32\n        model.add(Conv2D(512, kernel_size=self.kernel_size, padding=\"same\"))\n        model.add(Activation(\"relu\"))\n        model.add(BatchNormalization(momentum=0.8))\n\n        model.add(UpSampling2D())  # 24x32 -&gt; 48x64\n        model.add(Conv2D(256, kernel_size=self.kernel_size, padding=\"same\"))\n        model.add(Activation(\"relu\"))\n        model.add(BatchNormalization(momentum=0.8))\n\n        model.add(UpSampling2D())  # 48x64 -&gt; 96x128\n        model.add(Conv2D(128, kernel_size=self.kernel_size, padding=\"same\"))\n        model.add(Activation(\"relu\"))\n        model.add(BatchNormalization(momentum=0.8))\n\n        model.add(UpSampling2D())  # 96x128 -&gt; 192x256\n        model.add(Conv2D(64, kernel_size=self.kernel_size, padding=\"same\"))\n        model.add(Activation(\"relu\"))\n        model.add(BatchNormalization(momentum=0.8))\n\n        model.add(Conv2D(32, kernel_size=self.kernel_size, padding=\"same\"))\n        model.add(Activation(\"relu\"))\n        model.add(BatchNormalization(momentum=0.8))\n\n        model.add(Conv2D(self.channels, kernel_size=self.kernel_size, padding=\"same\"))\n        model.add(Activation(\"tanh\"))\n\n        model.summary()\n\n        noise = Input(shape=noise_shape)\n        img = model(noise)\n\n        return Model(noise, img)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57513773","CreationDate":"2019-08-15T17:33:03.370","ParentId":null,"OwnerUserId":"3742686","Title":null,"Body":"<p>It sounds totally understandable to me that you are having this problem. Your networks are not compensated, the Generator is much more powerful than the Discriminator, in terms of number of neurons. I would try to make the generator and discriminator symmetric each other, in terms of number of layers, their configuration, and their size, that way you assure no one is stronger than the other <em>de facto<\/em>. <\/p>\n"},{"AnswerId":"57523988","CreationDate":"2019-08-16T11:39:35.060","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>I guess you are using your own Dataset. If not consider using the same hyperparameter as in the DCGAN paper and first reproduce their results. Then use that network for your dataset. DCGANs especially with cross-entropy loss are known to be really tricky and extremely hyperparameter sensitive to get to run (see <a href=\"https:\/\/arxiv.org\/abs\/1711.10337\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1711.10337<\/a>). Especially, if you did not constrain your discriminator with some kind of gradient penalties (Gradient Penalty or Spectral norm).<\/p>\n\n<p>In GAN's using your own ideas for the hyperparameters is mostly a bad idea, unless you either really know what you are doing or have loads of GPU power to burn on a large scale search. <\/p>\n"}]}
+{"QuestionId":57514153,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras><neural-network>","CreationDate":"2019-08-15T18:05:24.433","AcceptedAnswerId":null,"OwnerUserId":10840909.0,"Title":"How to preprocess a dataset with multiple input variables for integration with Tensorflow for binary classification?","Body":"<p>I am trying to setup a minimum viable neural network to train for a binary classification task. I am importing a CSV file with seven total columns, the first column is the true label, 1 or -1. I have divided the dataset into a training set and test set. Additionally, I have been able to import the dataset into python, and normalize the six variables to numbers between zero and one.<\/p>\n\n<p>I have been spending the past few hours developing ad hoc, layers for a neural network. I have been able to get the network to feedforward, and to make binary predictions. However, I am totally at a loss for how to train the model because I cannot figure out how to update the weights. Generally, I am familiar with back-propagation, SGD, derivates, ADAM, and the perceptron learning rule. The problem is I have been unable to figure out how to implement any of these methods, so my program learns.<\/p>\n\n<p>In short, my goal is to update the weights of the neural network if the predict label and true label are not the same, so the network can learn.<\/p>\n\n<p>I have tried using Keras because it is a high level API. <a href=\"https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_classification#build_the_model\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_classification#build_the_model<\/a> While keras is good for getting started, it is very difficult to use Keras or Tensorflow on the datasets I generated myself. I am having trouble reformatting a neural network for the MNIST dataset to the needs of my present task <a href=\"https:\/\/medium.com\/@brian.s.haney44\/computer-vision-b39256f13fa4\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@brian.s.haney44\/computer-vision-b39256f13fa4<\/a>. I was hoping to complete the task with the Python standard library. Further, I have looked into and researched matrix multiplication vs. linear models for feedforward propagation, as well as derivative calculators and the perceptron learning rule for for weight updates. I created a Github <a href=\"https:\/\/github.com\/Bhaney44\/Deep-Neural-Network\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Bhaney44\/Deep-Neural-Network<\/a> repo for the project with a sample dataset for illustrative purposes.<\/p>\n\n<p>'''<\/p>\n\n<pre><code>    #Input Array\n    input_array = (true_label, var_1, var_2, var_3)\n\n    #Input layer with three nodes and a normalization function\n    def input_layer():\n        x_1 = (input_array[1])\/float(100)\n        x_2 = (input_array[2])\/float(100)\n        x_3 = (input_array[2])\/float(100)\n\n        #Hidden layer\n        def hidden_layer():\n            w_0 = 0.1\n            w_1 = 0.1\n            w_2 = 0.1\n            w_3 = 0.1\n            x_4 = (x_1 * w_0)+(x_2 * w_1)\n            x_5 = (x_2 * w_2)+(x_3 * w_3)\n\n            #Output layer\n            def output_layer():\n                w_4 = 0.1\n                w_5 = 0.1\n                x_6 = (x_4 * w_4)+(x_5 * w_5)\n\n                #Activation function\n                def activation():\n                    #True label\n                    x_0 = input_array[0]\n                    #decision function\n                    if x_6 &gt; 0.5:\n                        y = 1\n                    else:\n                        y = -1\n\n                    #activation function\n                    if y is x_0:\n                        #Here the network predicted right\n                        print('correct')\n                    else:\n                        #Here I want to adjust the weigh parameters\n                        #However, I am not sure where to start\n                        #Ideas include\n                            #calculate loss function\n                            #change weights structure\n                            #use matrices instead of coefficients\n                        print('update weights')\n\n                activation()\n            output_layer()\n        hidden_layer()\n    input_layer()  \n<\/code><\/pre>\n\n<p>'''<\/p>\n\n<p>My expectation is to find a way to update the weights, so the network can learn from training data and be applied to new data. The results are as I expect because I put a print function as a placeholder for the weight update algorithm. I am just not sure how to link functions together or to change numbers within the program automatically. I would sincerely appreciate any help, advice, suggestions, or criticisms.<\/p>\n\n<p>With thanks,<\/p>\n\n<p>Brian<\/p>\n","answers":[]}
+{"QuestionId":57514212,"AnswerCount":0,"Tags":"<python-3.x><keras><anaconda><lstm><talos>","CreationDate":"2019-08-15T18:10:54.880","AcceptedAnswerId":null,"OwnerUserId":11932718.0,"Title":"Getting error while using Talos with LSTM model","Body":"<p>I have data frame with 2 columns (say <code>c1<\/code> and <code>c2<\/code>) and 9830 observations. I am trying to predict <code>c1<\/code> by using <code>c1<\/code> and <code>c2<\/code>. I divided dataset into training and test as <code>(6860,2)<\/code> and <code>(2970,2)<\/code> respectively. Afterwards, I prepared <code>trainX<\/code> <code>(6850,10,2)<\/code>, <code>trainY<\/code> <code>(6850,)<\/code> and <code>testX<\/code> <code>(2960,10,2)<\/code>, <code>testY<\/code> <code>(2960)<\/code> datasets where time step is 10 in order to create datasets for LSTM type.<\/p>\n\n<p>Moreover, for Talos input (<code>x<\/code> and <code>y<\/code>) I have selected 2 columns for <code>x<\/code> <code>(9830, 2)<\/code> and selected first column for <code>y<\/code> <code>(9830,)<\/code>.<\/p>\n\n<p>I will be grateful if you could help me. Am I using parameters in sequential model correctly? <\/p>\n\n<p>My functions for LSTM model and Talos are given as below:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def talos_opt():\n\n    # set the parameter space\n    p = {'first_neuron':[2, 3, 4],\n         'second_neuron':[2, 3, 4],\n         'hidden_layers':[0, 1, 2],\n         'batch_size': [10, 20 ,30],\n         'epochs': [30, 50, 100],\n         'dropout': [0, 0.2, 0.4],\n         'recurrent_dropout': [0, 0.2, 0.4],\n         'kernel_initializer': ['uniform','normal'],\n         'optimizer': ['Nadam', 'Adam', 'sgd'],\n         'losses': ['binary_crossentropy', 'mean_squared_error', 'mean_absolute_error'],\n         'activation':['relu', 'elu'],\n         'middle_activation':['tanh', 'sigmoid', 'relu'],\n         'last_activation': ['softmax', 'sigmoid', 'relu']}\n\n    # first we have to make sure to input data and params into the function\n    def create_model(trainX, trainY, testX, testY, params):\n\n        lstm_model = Sequential()\n        lstm_model.add(LSTM(params['first_neuron'], activation=params['activation'], \n                            batch_input_shape=(params['batch_size'], TIME_STEPS, trainX.shape[2]), \n                            dropout=params['dropout'], recurrent_dropout=params['recurrent_dropout'],\n                            kernel_initializer=params['kernel_initializer'],\n                            return_sequences=False))\n\n        lstm_model.add(Dropout(params['dropout']))\n\n        lstm_model.add(Dense(4,activation='relu'))\n\n        lstm_model.add(Dense(params['second_neuron'], activation=params['last_activation']))\n\n        lstm_model.compile(loss=params['losses'],\n                      optimizer=params['optimizer'](),\n                      metrics=['acc', 'fmeasure_acc', 'mae'])\n\n        model_out = lstm_model.fit(trainX, trainY, \n                            validation_data=[testX, testY],\n                            batch_size=params['batch_size'],\n                            callbacks=[history],\n                            epochs=params['epochs'],\n                            verbose=0)\n\n        return model_out, lstm_model \n\n    scan_object = ta.Scan(x, y, model=create_model, params=p, grid_downsample=0.1)\n\n    return scan_object\n<\/code><\/pre>\n\n<p>I am getting following error;<\/p>\n\n<pre><code>File \"&lt;ipython-input-23-82bbc9eeafb8&gt;\", line 23, in create_model\nbatch_input_shape=(params['batch_size'], TIME_STEPS, trainX.shape[2]),\n\nIndexError: tuple index out of range\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57514280,"AnswerCount":1,"Tags":"<python><keras><scikit-learn>","CreationDate":"2019-08-15T18:16:09.163","AcceptedAnswerId":57529834.0,"OwnerUserId":10410524.0,"Title":"Using sklearn models as input to deep learning models","Body":"<p>Keras gives me a way to use my deep learning models with sklearn(The keras wrapper for sklearn), but I need the same thing the <strong>other way around<\/strong>.<\/p>\n\n<p>I want to create an ensemble of several already trained sklearn models by feeding their output to the input layer of a deep learning classifier(to be trained)<\/p>\n\n<p>Can I achieve that?<\/p>\n","answers":[{"AnswerId":"57529834","CreationDate":"2019-08-16T18:58:50.587","ParentId":null,"OwnerUserId":"8269172","Title":null,"Body":"<p>You should probably explore Stacking : <a href=\"http:\/\/blog.kaggle.com\/2016\/12\/27\/a-kagglers-guide-to-model-stacking-in-practice\/\" rel=\"nofollow noreferrer\">http:\/\/blog.kaggle.com\/2016\/12\/27\/a-kagglers-guide-to-model-stacking-in-practice\/<\/a> <br><\/p>\n\n<p>What happens is that when we are doing cross validation, we can combine combine the out of fold predictions to regenerate the training data.<\/p>\n\n<p>For example, if you 1000 data points and you use 5 folds to evaluate, you will have 5 different validation sets of length 200. Combining all the predictions obtained on this set will essentially give you a new feature of length 1000, hence a new feature.\nSimilarly  by training more models, you can get 3-4 features corresponding to predictions from 3-4 models. <br><\/p>\n\n<p>Finally you can stack these features with any model of your choice, you can even use a deep neural network.<\/p>\n"}]}
+{"QuestionId":57515064,"AnswerCount":1,"Tags":"<python><image-processing><pytorch>","CreationDate":"2019-08-15T19:22:40.920","AcceptedAnswerId":57515325.0,"OwnerUserId":11933180.0,"Title":"Can the coordinates of a 2 dimensional pytorch tensor be remapped?","Body":"<p>I'm trying to remap a 2d image to new coordinates. Previously I was using OpenCV's <a href=\"https:\/\/docs.opencv.org\/3.4\/da\/d54\/group__imgproc__transform.html#gab75ef31ce5cdfb5c44b6da5f3b908ea4\" rel=\"nofollow noreferrer\">cv.remap<\/a> method. I'd like to use CUDA with compiling C directly. Is there a Pytorch version of remap that will interpolate values?<\/p>\n","answers":[{"AnswerId":"57515325","CreationDate":"2019-08-15T19:46:00.240","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Are you looking for <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.functional.html#grid-sample\" rel=\"nofollow noreferrer\"><code>grid_sample<\/code><\/a>?\nLook for make_grid function for additional information<\/p>\n"}]}
+{"QuestionId":57515098,"AnswerCount":1,"Tags":"<python-3.x><keras>","CreationDate":"2019-08-15T19:25:30.383","AcceptedAnswerId":null,"OwnerUserId":3993270.0,"Title":"Keras: export the loss and accuracy as an array for plotting","Body":"<p>I am using the following code to train a model in Keras:<\/p>\n\n<pre><code>model_A.fit(train_X, train_Y, epochs=20)\n<\/code><\/pre>\n\n<p>The code works fine and the outputs are like belwo:<\/p>\n\n<pre><code>Epoch 1\/20\n1800\/1800 [==============================] - 0s 34us\/step - loss: 0.2764 - acc: 0.9033\nEpoch 2\/20\n1800\/1800 [==============================] - 0s 32us\/step - loss: 0.2704 - acc: 0.9083\nEpoch 3\/20\n1800\/1800 [==============================] - 0s 32us\/step - loss: 0.2687 - acc: 0.9094\nEpoch 4\/20\n1800\/1800 [==============================] - 0s 32us\/step - loss: 0.2748 - acc: 0.9089\nEpoch 5\/20\n1800\/1800 [==============================] - 0s 32us\/step - loss: 0.2902 - acc: 0.8922\nEpoch 6\/20\n1800\/1800 [==============================] - 0s 32us\/step - loss: 0.2357 - acc: 0.9183\nEpoch 7\/20\n1800\/1800 [==============================] - 0s 32us\/step - loss: 0.2499 - acc: 0.9183\nEpoch 8\/20\n1800\/1800 [==============================] - 0s 33us\/step - loss: 0.2286 - acc: 0.9228\nEpoch 9\/20\n1800\/1800 [==============================] - 0s 32us\/step - loss: 0.2325 - acc: 0.9194\nEpoch 10\/20\n1800\/1800 [==============================] - 0s 33us\/step - loss: 0.2053 - acc: 0.9261\nEpoch 11\/20\n1800\/1800 [==============================] - 0s 33us\/step - loss: 0.2256 - acc: 0.9161\nEpoch 12\/20\n1800\/1800 [==============================] - 0s 33us\/step - loss: 0.2120 - acc: 0.9261\nEpoch 13\/20\n1800\/1800 [==============================] - 0s 32us\/step - loss: 0.2085 - acc: 0.9328\nEpoch 14\/20\n1800\/1800 [==============================] - 0s 32us\/step - loss: 0.1881 - acc: 0.9328\nEpoch 15\/20\n1800\/1800 [==============================] - 0s 31us\/step - loss: 0.1835 - acc: 0.9344\nEpoch 16\/20\n1800\/1800 [==============================] - 0s 34us\/step - loss: 0.1812 - acc: 0.9356\nEpoch 17\/20\n1800\/1800 [==============================] - 0s 32us\/step - loss: 0.1704 - acc: 0.9361\nEpoch 18\/20\n1800\/1800 [==============================] - 0s 32us\/step - loss: 0.1929 - acc: 0.9272\nEpoch 19\/20\n1800\/1800 [==============================] - 0s 32us\/step - loss: 0.1822 - acc: 0.9317\nEpoch 20\/20\n1800\/1800 [==============================] - 0s 32us\/step - loss: 0.1713 - acc: 0.9417\n<\/code><\/pre>\n\n<p>I am wondering is there a way to save the loss\/accuracy values in an array? So I could the the plot over epochs later? Thanks!<\/p>\n","answers":[{"AnswerId":"57515373","CreationDate":"2019-08-15T19:48:46.510","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The <code>fit<\/code> method returns a <code>History<\/code> object which contains information about the training process. For example:<\/p>\n\n<pre><code># train the model\nh = model.fit()\n\n# loss values at the end of each epoch\nh.history['loss']\n\n# validation loss values per epoch (if you have used validation data)\nh.history['val_loss']\n\n# accuracy values at the end of each epoch (if you have used `acc` metric)\nh.history['acc']\n\n# validation accuracy values per epoch (if you have used `acc` metric and validation data)\nh.history['val_acc']\n\n# list of epochs number\nh.epoch\n<\/code><\/pre>\n"}]}
+{"QuestionId":57515785,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2019-08-15T20:22:49.807","AcceptedAnswerId":null,"OwnerUserId":9185745.0,"Title":"'TensorBoard' object has no attribute 'writer' error when using Callback.on_epoch_end()","Body":"<p>Since Model.train_on_batch() doesn't take a callback input, I tried using Callback.on_epoch_end() in order to write my loss to tensorboard<\/p>\n\n<p>However, trying to run the on_epoch_end() method results in the titular error, 'TensorBoard' object has no attribute 'writer'. Other solutions to my original problem with writing to tensorboard included calling the Callback.writer attribute, and running these solutions gave the same error. Also, the tensorflow documentation for the TensorBoard class doesn't mention a writer attribute<\/p>\n\n<p>I'm somewhat of a novice programmer, but it seems to me that the on_epoch_end() method is also at some point calling the writer attribute, but I'm confused as to why the function would use an attribute that doesn't exist<\/p>\n\n<p>Here's the code I'm using to create the callback:<\/p>\n\n<pre><code>logdir = \".\/logs\/\"\ntensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=logdir)\n<\/code><\/pre>\n\n<p>and this is the callback code that I try to run in my training loop:<\/p>\n\n<pre><code>logs = {\n            'encoder':encoder_loss[0],\n            'discriminator':d_loss,\n            'generator':g_loss,\n        }\ntensorboard_callback.on_epoch_end(i, logs)\n<\/code><\/pre>\n\n<p>where encoder_loss, d_loss, and g_loss are my scalars, and i is the batch number<\/p>\n\n<p>Is the error a result of some improper code on my part, or is tensorflow trying to reference something that doesn't exist?<\/p>\n\n<p>Also, if anyone knows another way to write to tensorboard using Model.train_on_batch, that would also solve my problem<\/p>\n","answers":[]}
+{"QuestionId":57516027,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch><activation-function>","CreationDate":"2019-08-15T20:43:34.603","AcceptedAnswerId":57521524.0,"OwnerUserId":11354575.0,"Title":"Does pytorch apply softmax automatically in nn.Linear","Body":"<p>In <code>pytorch<\/code> a classification network model is defined as this,<\/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)   # hidden layer\n        self.out = torch.nn.Linear(n_hidden, n_output)   # output layer\n\n    def forward(self, x):\n        x = F.relu(self.hidden(x))      # activation function for hidden layer\n        x = self.out(x)\n        return x\n<\/code><\/pre>\n\n<p>Is softmax applied here? In my understanding, things should be like, <\/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)   # hidden layer\n        self.relu =  torch.nn.ReLu(inplace=True)\n        self.out = torch.nn.Linear(n_hidden, n_output)   # output layer\n        self.softmax = torch.nn.Softmax(dim=n_output)\n    def forward(self, x):\n        x = self.hidden(x)      # activation function for hidden layer\n        x = self.relu(x)\n        x = self.out(x)\n        x = self.softmax(x)\n        return x\n<\/code><\/pre>\n\n<p>I understand that <code>F.relu(self.relu(x))<\/code> is also applying relu, but the first block of code doesn't apply softmax, right?<\/p>\n","answers":[{"AnswerId":"57521524","CreationDate":"2019-08-16T08:45:48.463","ParentId":null,"OwnerUserId":"3607203","Title":null,"Body":"<p>Latching on to what <em>@jodag<\/em> was already saying in his comment, and extending it a bit to form a full answer:<\/p>\n\n<p><strong>No<\/strong>, PyTorch does not automatically apply softmax, and you can at any point apply <code>torch.nn.Softmax()<\/code> as you want. <strong>But<\/strong>, softmax has <a href=\"https:\/\/stackoverflow.com\/questions\/42599498\/numercially-stable-softmax\">some issues with numerical stability<\/a>, which we want to avoid as much as we can. One solution is to use log-softmax, but this tends to be slower than a direct computation.<\/p>\n\n<p>Especially when we are using Negative Log Likelihood as a loss function (in PyTorch, this is <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#nllloss\" rel=\"nofollow noreferrer\"><code>torch.nn.NLLLoss<\/code><\/a>, we can utilize the fact that the derivative of (log-)softmax+NLLL is actually <a href=\"https:\/\/ljvmiranda921.github.io\/notebook\/2017\/08\/13\/softmax-and-the-negative-log-likelihood\/\" rel=\"nofollow noreferrer\">mathematically quite nice<\/a> and simple, which is why it makes sense to combine the both into a single function\/element. The result is then <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#crossentropyloss\" rel=\"nofollow noreferrer\"><code>torch.nn.CrossEntropyLoss<\/code><\/a>. Again, note that this only applies directly to the last layer of your network, any other computation is not affected by any of this.<\/p>\n"}]}
+{"QuestionId":57516253,"AnswerCount":0,"Tags":"<python><theano><theano.scan>","CreationDate":"2019-08-15T21:07:09.370","AcceptedAnswerId":null,"OwnerUserId":9482144.0,"Title":"How to set thought vector LSTM\/GRU in theano","Body":"<p>My theano GRU is producing language that sounds syntactically okay. However, the answers seem to be unrelated to the question. I am worried that there is a problem with how I set the thought vector. How do I properly set the thought vector?<\/p>\n\n<p>So far I tried initializing the hidden state with zeros in the GRU unit. This is followed by setting the GRU first layer hidden state each time I train or prediction (stochastically). See below.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class GRU:\n    ...\n    self.h = init_weights(inputs, outputs)\n    ...\n\nclass ThoughtDecoder:\n    def __init__(self, ...):\n        ...\n        for i in range(len(self.hiddens)):\n            ru = GRU(L[i], L[i+1], self.fn)\n            self.layers.append(ru)\n            self.params += ru.params\n        ...\n    def learn(self, X, Y):\n        ...\n        for i in range(epochs):\n            self.layers[0].h = theano.shared(X[j])\n            c, p = self.train_op(input_sequence, output_sequence)\n        ...\n    def predict(X):\n        ...\n        idx2word = {v: k for k, v in self.word_idx.items()}\n\n        self.layers[0].h = theano.shared(X)\n        PY_X, Y = self.predict_op([0])\n        ...\n\n\n<\/code><\/pre>\n\n<p>An example of a question is,<\/p>\n\n<pre><code>\"Would it be a good idea to ask a woman her age group who told me off for asking her age?\"\n<\/code><\/pre>\n\n<p>I expected an answer like, \"no\". But instead, I get answers like <\/p>\n\n<pre><code>\"china and russia\"\n<\/code><\/pre>\n\n<p>In case this bit of information helps, it keeps predicting new sequences even when I use argmax predictions. It also does not matter what self.layers[0].h shape is because it is still able to make predictions without throwing error. Thus, it appears that I am not setting the thought vector correctly.<\/p>\n\n<p>How do I properly set the thought vector?<\/p>\n","answers":[]}
+{"QuestionId":57516374,"AnswerCount":0,"Tags":"<python><tensorflow><keras><model><deep-learning>","CreationDate":"2019-08-15T21:18:55.427","AcceptedAnswerId":null,"OwnerUserId":11780764.0,"Title":"what is causing an index error in the model of this code?","Body":"<p>Hi I was trying to creat a deep learning model using Tensorflow Keras and I recieved an index error.\nThe code:<\/p>\n\n<pre><code>scaler = StandardScaler().fit(train)\n\ntrain=scaler.transform(train)\ntest=scaler.transform(test)\n\n# Creating Deep Model\n\n\nmodel = Sequential()\n\n# Add an input layer\nmodel.add(Dense(12, activation='relu', input_shape=(211,)))\n\n# Add one hidden layer\nmodel.add(Dense(8, activation='relu'))\n\n# Add an output layer\nmodel.add(Dense(1, activation='sigmoid'))\n\n#add improvements \nmodel.add(BatchNormalization())\nmodel.add(Dropout(0.5))\n#Train the model\n\nmodel.compile(loss='binary_crossentropy',optimizer='adam',metrics=['accuracy'])\n\nmodel.fit(train,train_targets,epochs=20, batch_size=1, verbose=1)\n<\/code><\/pre>\n\n<p>I am getting this error at the last line:<\/p>\n\n<blockquote>\n  <p>File \"C:\\Users\\snow\\Anaconda3\\lib\\site-packages\\keras\\engine\\training_utils.py\", line 80, in standardize_input_data\n      if isinstance(data[0], list):\n  IndexError: list index out of range<\/p>\n<\/blockquote>\n\n<p>what does this error mean and what could be causing it? thx<\/p>\n","answers":[]}
+{"QuestionId":57516417,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-08-15T21:22:58.967","AcceptedAnswerId":null,"OwnerUserId":6019463.0,"Title":"Use separate input as parameter to Lambda layer in Keras","Body":"<p>Say I have layer in my neural network with an output that looks like this:<\/p>\n\n<pre><code>output = (None, 5, 25)\n<\/code><\/pre>\n\n<p>I want to write a Lambda function that can take an average of these 25-dimensional tensors, but slice the items based off an index value that would be passed as a separate input. <\/p>\n\n<pre><code>index_input = (5,)\n<\/code><\/pre>\n\n<p>So I figured it could maybe be done with a Lambda like this:<\/p>\n\n<pre><code>def TargetAverage(x, nt):\n    new = x[:, 0:nt[1]]\n    new = backend.mean(new, 1)     \n    return new \n<\/code><\/pre>\n\n<p>And then pass the output of the input tensor as an argument: <\/p>\n\n<pre><code>    new_output = Lambda(TargetAverage, arguments={'nt': index_input})(output)\n<\/code><\/pre>\n\n<p>And then expected output would be the average of those 5 vectors up to the point of the index value:<\/p>\n\n<pre><code> (None, 25)\n<\/code><\/pre>\n\n<p>But it seems like this would only feed it the symbolic tensor, not the actual value.<\/p>\n\n<p>Any idea how this could be done?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
+{"QuestionId":57516612,"AnswerCount":1,"Tags":"<python><tensorflow><configuration><cudnn>","CreationDate":"2019-08-15T21:44:12.833","AcceptedAnswerId":57517274.0,"OwnerUserId":11907197.0,"Title":"How to fix \"Could not find any libcudnn.7*.dylib in any subdirectory:\" with respect to Tensorflow 1.14 Configuration","Body":"<p>I am currently configuring tensorflow 1.14 with CUDA 10.0 and cudnn 7.5 installed in order to build my Deepspeech binaries. <\/p>\n\n<p>Error of terminal code not finding the file:<\/p>\n\n<pre><code>1) Could not find any libcudnn.7*.dylib in any subdirectory:\n        ''\n        'lib64'\n        'lib'\n        'lib\/*-linux-gnu'\n        'lib\/x64'\n        'extras\/CUPTI\/*'\nof:\n        '\/usr'\n        '\/usr\/local\/cuda'\nAsking for detailed CUDA configuration...\n\n2) Please specify the comma-separated list of base paths to look for CUDA libraries and headers. [Leave empty to use the default]: \n<\/code><\/pre>\n\n<p>I have already copied the files from cudnn to the cuda directory as seen below:<\/p>\n\n<p>First I extracted the file and then copied it to the cuda file<\/p>\n\n<pre><code>(deepspeech-venv) Chabanis-MacBook-Pro:tensorflow chabani$ tar -xzvf cudnn-10.0-osx-x64-v7.5.0.56.tgz\nx cuda\/include\/cudnn.h\nx cuda\/NVIDIA_SLA_cuDNN_Support.txt\nx cuda\/lib\/libcudnn.7.dylib\nx cuda\/lib\/libcudnn.dylib\nx cuda\/lib\/libcudnn_static.a\n(deepspeech-venv) Chabanis-MacBook-Pro:tensorflow chabani$ sudo cp cuda\/include\/cudnn.h \/usr\/local\/cuda\/include\nPassword:\n(deepspeech-venv) Chabanis-MacBook-Pro:tensorflow chabani$ sudo cp cuda\/lib\/libcudnn.7.dylib \/usr\/local\/cuda\/include\n(deepspeech-venv) Chabanis-MacBook-Pro:tensorflow chabani$ sudo cp cuda\/lib\/libcudnn.dylib \/usr\/local\/cuda\/include\n<\/code><\/pre>\n\n<p>I ran the configuration as below<\/p>\n\n<pre><code>(deepspeech-venv) Chabanis-MacBook-Pro:tensorflow chabani$ .\/configure\nWARNING: --batch mode is deprecated. Please instead explicitly shut down your Bazel server using the command \"bazel shutdown\".\nYou have bazel 0.24.1 installed.\nPlease specify the location of python. [Default is \/Users\/chabani\/tmp\/deepspeech-venv\/bin\/python]: \n\n\nTraceback (most recent call last):\n  File \"&lt;string&gt;\", line 1, in &lt;module&gt;\nAttributeError: module 'site' has no attribute 'getsitepackages'\nFound possible Python library paths:\n  \/Users\/chabani\/tmp\/deepspeech-venv\/lib\/python3.7\/site-packages\nPlease input the desired Python library path to use.  Default is [\/Users\/chabani\/tmp\/deepspeech-venv\/lib\/python3.7\/site-packages]\n\nDo you wish to build TensorFlow with XLA JIT support? [y\/N]: \nNo XLA JIT support will be enabled for TensorFlow.\n\nDo you wish to build TensorFlow with OpenCL SYCL support? [y\/N]: \nNo OpenCL SYCL support will be enabled for TensorFlow.\n\nDo you wish to build TensorFlow with ROCm support? [y\/N]: n\nNo ROCm support will be enabled for TensorFlow.\n\nDo you wish to build TensorFlow with CUDA support? [y\/N]: y\nCUDA support will be enabled for TensorFlow.\n\n1)Could not find any libcudnn.7*.dylib in any subdirectory:\n        ''\n        'lib64'\n        'lib'\n        'lib\/*-linux-gnu'\n        'lib\/x64'\n        'extras\/CUPTI\/*'\nof:\n        '\/usr'\n        '\/usr\/local\/cuda'\nAsking for detailed CUDA configuration...\n\nPlease specify the CUDA SDK version you want to use. [Leave empty to default to CUDA 10]: 10.0\n\n\nPlease specify the cuDNN version you want to use. [Leave empty to default to cuDNN 7]: 7.0\n\n\n2)Please specify the comma-separated list of base paths to look for CUDA libraries and headers. [Leave empty to use the default]: \n<\/code><\/pre>\n\n<p>-Please could you possibly assist with the items numbered 1 and 2? I thought me copying the cudnn files into cuda would not have caused this issue<\/p>\n","answers":[{"AnswerId":"57517274","CreationDate":"2019-08-15T23:16:47.480","ParentId":null,"OwnerUserId":"10543528","Title":null,"Body":"<p>Here's a problem: you copied your libraries into <code>include<\/code> instead of into <code>lib<\/code><\/p>\n\n<pre><code>(deepspeech-venv) Chabanis-MacBook-Pro:tensorflow chabani$ sudo cp cuda\/lib\/libcudnn.7.dylib \/usr\/local\/cuda\/include\n(deepspeech-venv) Chabanis-MacBook-Pro:tensorflow chabani$ sudo cp cuda\/lib\/libcudnn.dylib \/usr\/local\/cuda\/include\n<\/code><\/pre>\n\n<p>Try moving them into <code>cuda\/lib<\/code> instead<\/p>\n"}]}
+{"QuestionId":57516662,"AnswerCount":1,"Tags":"<python><tensorflow><keras><conv-neural-network><dimension>","CreationDate":"2019-08-15T21:51:21.897","AcceptedAnswerId":null,"OwnerUserId":9715052.0,"Title":"ValueError: Error when checking input: expected input to have 4 dimensions, but got array with shape (859307, 1)","Body":"<p>I'm creating a convolutional autoencoder that takes in 16x16 images but I keep getting the following error: <\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"WTApruning.py\", line 69, in &lt;module&gt;\n    validation_data=(x_test, x_test))\n  File \"\/PycharmProjects\/predictivemodel\/venv\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training.py\", line 709, in fit\n    shuffle=shuffle)\n  File \"\/PycharmProjects\/predictivemodel\/venv\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training.py\", line 2651, in _standardize_user_data\n    exception_prefix='input')\n  File \"\/PycharmProjects\/predictivemodel\/venv\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training_utils.py\", line 376, in standardize_input_data\n    'with shape ' + str(data_shape))\nValueError: Error when checking input: expected input to have 4 dimensions, but got array with shape (859307, 1)\n<\/code><\/pre>\n\n<p>From other stack overflow posts such as this <a href=\"https:\/\/stackoverflow.com\/questions\/57222126\/error-when-checking-input-expected-conv2d-6-input-to-have-4-dimensions-but-got\">one<\/a>, it seems like I need to add another dimension for the colour channel but what would the other dimension I add be? <\/p>\n\n<p>Code<\/p>\n\n<pre><code>path = \"...\"\nCATEGORIES = [\"x_train\", \"x_test\"]\ncount = 0\ndata = []\nx_test, x_train = [], []\nfor img in os.listdir(path):\n    img_array = cv2.imread(os.path.join(path,img) ,cv2.IMREAD_GRAYSCALE)\n    data.append(img_array)\nx_train, x_test = train_test_split(data, test_size = 0.1)\nx_train, x_test = train_test_split(data, test_size = 0.1)\nx_train = np.array(x_train)\nx_test = np.array(x_test)\n# just updated\nx_train = x_train.reshape(x_train,(len(x_train),16,16,1))\nx_test = x_test.reshape(x_test,(len(x_test),16,16,1))\n\n# ENCODER\nencoder_img = tf.keras.layers.Input(shape=(16,16,1), name=\"input\")\nx = tf.keras.layers.Conv2D(1024, 1, activation='relu', kernel_initializer=keras.initializers.RandomUniform)(encoder_img)\nx = tf.keras.layers.MaxPooling2D(1)(x)\nx = tf.keras.layers.Conv2D(512, 1, activation='relu')(x)\nx = tf.keras.layers.MaxPooling2D(1)(x)\nencoder_output = tf.keras.layers.Conv2D(256, 3, activation='relu')(x)\n\n# DECODER\nx = tf.keras.layers.Conv2DTranspose(512, 1, activation='relu')(encoder_output)\nx = tf.keras.layers.UpSampling2D(1)(x)\nx = tf.keras.layers.Conv2DTranspose(1024, 1, activation='relu')(x)\nx = tf.keras.layers.UpSampling2D(1)(x)\ndecoder_output = tf.keras.layers.Conv2DTranspose(1, 3, activation='relu')(x)\n# COMPILE\nautoencoder = tf.keras.Model(inputs=encoder_img, outputs=decoder_output, name='autoencoder')\nautoencoder.summary()\n\nautoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')\nautoencoder.fit(x_train, x_train,\n                epochs=20,\n                batch_size=128,\n                shuffle=True,\n                validation_data=(x_test, x_test))\ndecoded_imgs = autoencoder.predict(x_test)\n<\/code><\/pre>\n\n<p>New Error after reshaping is added:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"WTApruning.py\", line 43, in &lt;module&gt;\n    x_train = x_train.reshape(x_train,(len(x_train),16,16,1))\nTypeError: only integer scalar arrays can be converted to a scalar index\n\n<\/code><\/pre>\n\n<p>Error without reshaping:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"WTApruning.py\", line 68, in &lt;module&gt;\n    validation_data=(x_test, x_test))\n  File \"\/PycharmProjects\/predictivemodel\/venv\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training.py\", line 709, in fit\n    shuffle=shuffle)\n  File \"\/PycharmProjects\/predictivemodel\/venv\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training.py\", line 2651, in _standardize_user_data\n    exception_prefix='input')\n \"\/PycharmProjects\/predictivemodel\/venv\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/training_utils.py\", line 376, in standardize_input_data\n    'with shape ' + str(data_shape))\nValueError: Error when checking input: expected input to have 4 dimensions, but got array with shape (859307, 1)\n\n<\/code><\/pre>\n","answers":[{"AnswerId":"57516710","CreationDate":"2019-08-15T21:58:49.920","ParentId":null,"OwnerUserId":"10497666","Title":null,"Body":"<p>You input x_train is not a 4d input.\nyou should reshape it before feeding it into the network.\nBest<\/p>\n"}]}
+{"QuestionId":57516673,"AnswerCount":2,"Tags":"<python><tensorflow><keras><classification>","CreationDate":"2019-08-15T21:53:01.773","AcceptedAnswerId":57516809.0,"OwnerUserId":11816060.0,"Title":"How to perform prediction using predict_generator on unlabeled test data in Keras?","Body":"<p>I'm trying to build an image classification model. It's a 4 class image classification. Here is my code for building image generators and running the training:<\/p>\n\n<pre><code>train_datagen = ImageDataGenerator(rescale=1.\/255.,\n                               rotation_range=30,\n                               horizontal_flip=True,\n                               validation_split=0.1)\n\n\ntrain_generator = image_gen.flow_from_directory(train_dir, target_size=(299, 299), \n                                                class_mode='categorical', batch_size=20,\n                                                subset='training')\n\nvalidation_generator = image_gen.flow_from_directory(train_dir, target_size=(299, 299), \n                                                class_mode='categorical', batch_size=20,\n                                                subset='validation')\n\nmodel.compile(Adam(learning_rate=0.001), loss='categorical_crossentropy',\n                                         metrics=['accuracy'])\n\nmodel.fit_generator(train_generator, steps_per_epoch=int(440\/20), epochs=20, \n                              validation_data=validation_generator, \n                              validation_steps=int(42\/20))\n<\/code><\/pre>\n\n<p>I was able to get train and validation work perfectly because the images in <em>train<\/em> directory are stored in a separate folder for each class. But, as you can see below, the <em>test<\/em> directory has 100 images and no folders inside it. It also doesn't have any labels and only contains image files.<\/p>\n\n<p>How can I do prediction on the image files in <em>test<\/em> folder using Keras? <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ahzqE.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ahzqE.png\" alt=\"dataset directory structure\"><\/a><\/p>\n","answers":[{"AnswerId":"57516809","CreationDate":"2019-08-15T22:11:34.620","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>If you are interested to only perform prediction, you can load the images by a simple hack like this:<\/p>\n\n<pre><code>test_datagen = ImageDataGenerator(rescale=1\/255.)\n\ntest_generator = test_datagen('PATH_TO_DATASET_DIR\/Dataset',\n                              # only read images from `test` directory\n                              classes=['test'],\n                              # don't generate labels\n                              class_mode=None,\n                              # don't shuffle\n                              shuffle=False,\n                              # use same size as in training\n                              target_size=(299, 299))\n\npreds = model.predict_generator(test_generator)\n<\/code><\/pre>\n\n<p>You can access <code>test_generator.filenames<\/code> to get a list of corresponding filenames so that you can map them to their corresponding prediction.<\/p>\n\n<hr>\n\n<p><strong>Update (as requested in comments section):<\/strong> if you want to map predicted classes to filenames, first you must find the predicted classes. If your model is a classification model, then probably it has a softmax layer as the classifier. So the values in <code>preds<\/code> would be probabilities. Use <code>np.argmax<\/code> method to find the index with highest probability:<\/p>\n\n<pre><code>preds_cls_idx = preds.argmax(axis=-1)\n<\/code><\/pre>\n\n<p>So this gives you the indices of predicted classes. Now we need to map indices to their string labels (i.e. \"car\", \"bike\", etc.) which are provided by training generator in <code>class_indices<\/code> attribute:<\/p>\n\n<pre><code>import numpy as np\n\nidx_to_cls = {v: k for k, v in train_generator.class_indices.items()}\npreds_cls = np.vectorize(idx_to_cls.get)(preds_cls_idx)\nfilenames_to_cls = list(zip(test_generator.filenames, preds_cls))\n<\/code><\/pre>\n"},{"AnswerId":"57517093","CreationDate":"2019-08-15T22:52:02.673","ParentId":null,"OwnerUserId":"8073544","Title":null,"Body":"<p>your folder structure be like <code>testfolder\/folderofallclassfiles<\/code>  <\/p>\n\n<p>you can use     <\/p>\n\n<pre><code>test_generator = test_datagen.flow_from_directory(\n    directory=pred_dir,\n    class_mode=None,\n    shuffle=False\n)\n<\/code><\/pre>\n\n<p>before prediction i would also use reset to avoid unwanted outputs<\/p>\n\n<p>EDIT:<\/p>\n\n<p>For your purpose you need to know which image is associated with which prediction. The problem is that the data-generator start at different positions in the dataset each time we use the generator, thus giving us different outputs everytime. So, in order to restart at the beginning of the dataset in each call to <code>predict_generator()<\/code> you would need to exactly match the number of iterations and batches to the dataset-size.<br>\nThere are multiple ways to encounter this<\/p>\n\n<p>a)  You can see the internal batch-counter using <code>batch_index<\/code> of generator<br>\nb) create a new data-generator before each call to <code>predict_generator()<\/code><br>\nc) there is a better and simpler way, which is to call <code>reset()<\/code> on the generator, and if you have set <code>shuffle=False<\/code> in <code>flow_from_directory<\/code> then it should start over from the beginning of the dataset and give the exact same output each time, so now the ordering of <code>testgen.filenames<\/code> and <code>testgen.classes<\/code> matches  <\/p>\n\n<pre><code>test_generator.reset()\n<\/code><\/pre>\n\n<p>Prediction<\/p>\n\n<pre><code>prediction = model.predict_generator(test_generator,verbose=1,steps=numberofimages\/batch_size)\n<\/code><\/pre>\n\n<p><strong>To map the filename with prediction<\/strong><\/p>\n\n<p><code>predict_generator<\/code> gives output in probabilities so at first we need to convert them to class number like 0,1..<\/p>\n\n<pre><code>predicted_class = np.argmax(prediction,axis=1)\n<\/code><\/pre>\n\n<p>next step would be to convert those class number into actual class names<\/p>\n\n<pre><code>l = dict((v,k) for k,v in training_set.class_indices.items())\nprednames = [l[k] for k in predicted_classes]\n<\/code><\/pre>\n\n<p>getting filenames<\/p>\n\n<pre><code>filenames = test_generator.filenames\n<\/code><\/pre>\n\n<p>Finally creating df  <\/p>\n\n<pre><code>finaldf = pd.DataFrame({'Filename': filenames,'Prediction': prednames})\n<\/code><\/pre>\n"}]}
+{"QuestionId":57516678,"AnswerCount":0,"Tags":"<tensorflow><keras><lstm><numerical-stability>","CreationDate":"2019-08-15T21:53:32.233","AcceptedAnswerId":null,"OwnerUserId":10133797.0,"Title":"LSTM 'recurrent_dropout' with 'relu' yields NaNs","Body":"<p>Any non-zero <code>recurrent_dropout<\/code> yields NaN losses and weights; latter are either 0 or NaN. Happens for stacked, shallow, <code>stateful<\/code>, <code>return_sequences<\/code> = any, with &amp; w\/o <code>Bidirectional()<\/code>, <code>activation='relu'<\/code>, <code>loss='binary_crossentropy'<\/code>. NaNs occur within a few batches. <\/p>\n\n<p>Any fixes? Help's appreciated.<\/p>\n\n<p><hr>\n<b>TROUBLESHOOTING ATTEMPTED<\/b>:<\/p>\n\n<ul>\n<li><code>recurrent_dropout=0.2,0.1,0.01,1e-6<\/code><\/li>\n<li><code>kernel_constraint=maxnorm(0.5,axis=0)<\/code><\/li>\n<li><code>recurrent_constraint=maxnorm(0.5,axis=0)<\/code><\/li>\n<li><code>clipnorm=50<\/code>  (empirically determined), Nadam optimizer <\/li>\n<li><code>activation='tanh'<\/code> - no NaNs, weights stable, tested for up to 10 batches<\/li>\n<li><code>lr=2e-6,2e-5<\/code> - no NaNs, weights stable, tested for up to 10 batches<\/li>\n<li><code>lr=5e-5<\/code> - no NaNs, weights stable, for 3 batches - NaNs on batch 4<\/li>\n<li><code>batch_shape=(32,48,16)<\/code> - large loss for 2 batches, NaNs on batch 3<\/li>\n<\/ul>\n\n<p><em>NOTE<\/em>: <code>batch_shape=(32,672,16)<\/code>, 17 calls to <code>train_on_batch<\/code> per batch<\/p>\n\n<p><hr>\n<b>ENVIRONMENT<\/b>:<\/p>\n\n<ul>\n<li>Keras 2.2.4 (TensorFlow backend), Python 3.7, Spyder 3.3.7 via Anaconda<\/li>\n<li>GTX 1070 6GB, i7-7700HQ, 12GB RAM, Win-10.0.17134 x64<\/li>\n<li>CuDNN 10+, latest Nvidia drives<\/li>\n<\/ul>\n\n<hr>\n\n<p><b>ADDITIONAL INFO<\/b>:<\/p>\n\n<p>Model divergence is spontaneous, occurring at different train updates <em>even with fixed seeds<\/em> - Numpy, Random, and TensorFlow random seeds. Furthermore, when first diverging, LSTM layer weights are all normal - only going to NaN later. <\/p>\n\n<p>Below are, in order: (1) inputs to <code>LSTM<\/code>; (2) <code>LSTM<\/code> outputs; (3) <code>Dense(1,'sigmoid')<\/code> outputs -- the three are consecutive, with <code>Dropout(0.5)<\/code> between each. Preceding (1) are <code>Conv1D<\/code> layers. Right: LSTM weights. \"BEFORE\" = 1 train update before; \"AFTER = 1 train update after<\/p>\n\n<p><em>BEFORE divergence<\/em>:\n<img src=\"https:\/\/user-images.githubusercontent.com\/16495490\/63199693-a9c7c580-c04c-11e9-9396-b77079352720.png\" alt=\"image\"><\/p>\n\n<p><em>AT divergence<\/em>:\n<img src=\"https:\/\/user-images.githubusercontent.com\/16495490\/63199806-fca17d00-c04c-11e9-9224-be21a1c2d1c8.png\" alt=\"image\"><\/p>\n\n<pre><code>## LSTM outputs, flattened, stats\n(mean,std)        = (inf,nan)\n(min,max)         = (0.00e+00,inf)\n(abs_min,abs_max) = (0.00e+00,inf)\n<\/code><\/pre>\n\n<p><em>AFTER divergence<\/em>:\n<img src=\"https:\/\/user-images.githubusercontent.com\/16495490\/63200017-9c5f0b00-c04d-11e9-87e5-9ace4cbf9407.png\" alt=\"image\"><\/p>\n\n<pre><code>## Recurrent Gates Weights:\narray([[nan, nan, nan, ..., nan, nan, nan],\n       [ 0.,  0., -0., ..., -0.,  0.,  0.],\n       [ 0., -0., -0., ..., -0.,  0.,  0.],\n       ...,\n       [nan, nan, nan, ..., nan, nan, nan],\n       [ 0.,  0., -0., ..., -0.,  0., -0.],\n       [ 0.,  0., -0., ..., -0.,  0.,  0.]], dtype=float32)\n## Dense Sigmoid Outputs:\narray([[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.]], dtype=float32)\n<\/code><\/pre>\n\n<p><hr>\n<strong>MINIMAL REPRODUCIBLE EXAMPLE<\/strong>:<\/p>\n\n<pre><code>from keras.layers import Input,Dense,LSTM,Dropout\nfrom keras.models import Model\nfrom keras.optimizers  import Nadam \nfrom keras.constraints import maxnorm\n<\/code><\/pre>\n\n<pre><code>ipt = Input(batch_shape=(32,672,16))\nx = LSTM(512,activation='relu',return_sequences=False,\n             recurrent_dropout=1e-6,\n             kernel_constraint   =maxnorm(0.5,axis=0),\n             recurrent_constraint=maxnorm(0.5,axis=0))(ipt)\nout = Dense(1,activation='sigmoid')(x)\n\nmodel = Model(ipt,out)\noptimizer = Nadam(lr=2e-4,clipnorm=1)\nmodel.compile(optimizer=optimizer,loss='binary_crossentropy')\n<\/code><\/pre>\n\n<pre><code>for train_update,_ in enumerate(range(100)):\n    x = np.random.randn(32,672,16)\n    y = np.array([1]*5 + [0]*27)\n    np.random.shuffle(y)\n    loss = model.train_on_batch(x,y)\n    print(train_update+1,loss,np.sum(y))\n<\/code><\/pre>\n\n<p><strong>Observations<\/strong>: the following <em>speed up divergence<\/em>:<\/p>\n\n<ul>\n<li><em>Higher<\/em> <code>units<\/code> (LSTM)<\/li>\n<li><em>Higher<\/em> # of layers (LSTM)<\/li>\n<li><em>Higher<\/em> <code>lr<\/code> <strong>&lt;&lt;<\/strong> no divergence when <code>&lt;=1e-4<\/code>, tested up to 400 trains<\/li>\n<li><em>Less<\/em> <code>'1'<\/code> labels <strong>&lt;&lt;<\/strong> no divergence with <code>y<\/code> below, even with <code>lr=1e-3<\/code>; tested up to 400 trains<\/li>\n<\/ul>\n\n<p><code>y = np.random.randint(0,2,32) # makes more '1' labels<\/code><\/p>\n","answers":[]}
+{"QuestionId":57516740,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2019-08-15T22:03:17.857","AcceptedAnswerId":57516765.0,"OwnerUserId":3736306.0,"Title":"sess.run() multiple ops vs multiple sess.run()","Body":"<p>In tensorflow, if we want to run multiple operations, are there any differences (e.g. in speed or results) if we pass the ops in a list to a single <code>sess.run()<\/code>, versus we use multiple <code>sess.run()<\/code> but only pass a single ops to each one?<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>#!\/usr\/env\/bin python\nimport tensorflow as tf\n\nif __name__ == '__main__':\n    v = tf.Variable(0)\n    c = tf.constant(3)\n    add = tf.add(v, c)\n    update = tf.assign(v, add)\n    mul = tf.multiply(add, update)\n    init = tf.compat.v1.global_variables_initializer()\n\n    # method 1\n    with tf.Session() as sess:\n        sess.run(init)\n        print(sess.run(add))\n        print(sess.run(mul))\n\n    # method 2\n    with tf.Session() as sess:\n        print(sess.run([init, add, mul]))\n<\/code><\/pre>\n\n<p>Edit: Initially I was also wondering if there are cases where multiple Session will cause different results, e.g. dependencies in computational graph causes execution order to be different etc. But I didn't make this clear in my question and only gives a rather dummy example instead...<\/p>\n","answers":[{"AnswerId":"57516765","CreationDate":"2019-08-15T22:06:36.287","ParentId":null,"OwnerUserId":"3742686","Title":null,"Body":"<p>Each time you run the <code>init<\/code>, all the weights are initialised randomly, So no comparison can be made this way. If you really want to compare, I would run the following code:<\/p>\n\n<pre><code>#!\/usr\/env\/bin python\nimport tensorflow as tf\n\nif __name__ == '__main__':\n    v = tf.Variable(0)\n    c = tf.constant(3)\n    add = tf.add(v, c)\n    update = tf.assign(v, add)\n    mul = tf.multiply(add, update)\n    init = tf.compat.v1.global_variables_initializer()\n\n    # method 1\n    with tf.Session() as sess:\n        sess.run(init)\n        print(sess.run(add))\n        print(sess.run(mul))\n        print(sess.run([add, mul]))\n<\/code><\/pre>\n"},{"AnswerId":"57516824","CreationDate":"2019-08-15T22:13:27.237","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>There is no difference whatsoever in results:<\/p>\n\n<pre><code>import numpy as np\n\ntf.reset_default_graph()\nv = tf.Variable(0)\nc = tf.constant(3)\nadd = tf.add(v, c)\nupdate = tf.assign(v, add)\nmul = tf.multiply(add, update)\ninit = tf.compat.v1.global_variables_initializer()\n\ndef single():\n    with tf.Session() as sess:\n        sess.run(init)\n        fetch_add = sess.run(add)\n        fetch_mul = sess.run(mul)\n    return fetch_add, fetch_mul\n\ndef multiple():\n    with tf.Session() as sess:\n        sess.run(init)\n        fetch_add, fetch_mul = sess.run([add, mul])\n    return fetch_add, fetch_mul\n\nadd_single, multiply_single = single()\nadd_multiple, multiply_multiple = multiple()\n\nnp.testing.assert_array_equal(add_single, add_multiple)\nnp.testing.assert_array_equal(multiply_single, multiply_multiple)\n<\/code><\/pre>\n\n<p>When it comes to speed, here are some metrics:<\/p>\n\n<pre><code>%timeit single()\n# 7.44 ms \u00b1 143 \u00b5s per loop (mean \u00b1 std. dev. of 7 runs, 100 loops each)\n\n%timeit multiple()\n# 5.56 ms \u00b1 121 \u00b5s per loop (mean \u00b1 std. dev. of 7 runs, 100 loops each)\n<\/code><\/pre>\n\n<p>So it is definitely faster if you run multiple nodes together.<\/p>\n"}]}
+{"QuestionId":57516826,"AnswerCount":0,"Tags":"<tensorflow><version><checkpoint>","CreationDate":"2019-08-15T22:13:29.730","AcceptedAnswerId":null,"OwnerUserId":2096243.0,"Title":"Version of Tensorflow from the checkpoint file","Body":"<p>How can you find out the version of Tensorflow from checkpoint files? I checked <code>tensorflow.python.tools.inspect_checkpoint<\/code> functions, but couldn't find anything useful.<\/p>\n","answers":[]}
+{"QuestionId":57516892,"AnswerCount":1,"Tags":"<python><keras><deep-learning><data-visualization>","CreationDate":"2019-08-15T22:22:50.903","AcceptedAnswerId":57516955.0,"OwnerUserId":3578468.0,"Title":"Keras Visualization Toolkit throwing `<layer>input:0 is both fed and fetched.` for simple `Sequential` model","Body":"<p>I am trying to get the getting started example from <a href=\"https:\/\/github.com\/raghakot\/keras-vis\" rel=\"nofollow noreferrer\">Keras Visualization Toolkit<\/a> to run.<\/p>\n\n<p>My test code is this:<\/p>\n\n<pre><code># simple CNN definition\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Conv1D, Flatten\n#create model\nmodel = Sequential()\n#add model layers\nmodel.add(Conv1D(1000, kernel_size=3, activation='relu', input_shape=(28,3)))\nmodel.add(Conv1D(50, name='conv1D_2', kernel_size=3, activation='relu'))\nmodel.add(Flatten())\nmodel.add(Dense(10, activation='softmax'))\n\n# code from github page\nfrom vis.losses import ActivationMaximization\nfrom vis.regularizers import TotalVariation, LPNorm\n\nfilter_indices = [1, 2, 3]\n\n# Tuple consists of (loss_function, weight)\n# Add regularizers as needed.\nlosses = [\n    (ActivationMaximization(model.layers[1], filter_indices), 1),\n#     (LPNorm(model.input[1:]), 10),\n#     (TotalVariation(model.input), 10)\n]\n\nfrom vis.optimizer import Optimizer\n\noptimizer = Optimizer(model.input, losses)\nopt_img, grads, _ = optimizer.minimize()\n<\/code><\/pre>\n\n<p>I am getting the following error:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nInvalidArgumentError                      Traceback (most recent call last)\n&lt;ipython-input-38-eb157373ee04&gt; in &lt;module&gt;\n     15 \n     16 optimizer = Optimizer(model.input, losses)\n---&gt; 17 opt_img, grads, _ = optimizer.minimize()\n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/vis\/optimizer.py in minimize(self, seed_input, max_iter, input_modifiers, grad_modifier, callbacks, verbose)\n    141         best_loss = float('inf')\n    142         best_input = None\n--&gt; 143 \n    144         grads = None\n    145         wrt_value = None\n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/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~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/keras\/backend\/tensorflow_backend.py in _call(self, inputs)\n   2669                                 feed_symbols,\n   2670                                 symbol_vals,\n-&gt; 2671                                 session)\n   2672         if self.run_metadata:\n   2673             fetched = self._callable_fn(*array_vals, run_metadata=self.run_metadata)\n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/keras\/backend\/tensorflow_backend.py in _make_callable(self, feed_arrays, feed_symbols, symbol_vals, session)\n   2621             callable_opts.run_options.CopyFrom(self.run_options)\n   2622         # Create callable.\n-&gt; 2623         callable_fn = session._make_callable_from_options(callable_opts)\n   2624         # Cache parameters corresponding to the generated callable, so that\n   2625         # we can detect future mismatches and refresh the callable.\n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py in _make_callable_from_options(self, callable_options)\n   1469     \"\"\"\n   1470     self._extend_graph()\n-&gt; 1471     return BaseSession._Callable(self, callable_options)\n   1472 \n   1473 \n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py in __init__(self, session, callable_options)\n   1423         with errors.raise_exception_on_not_ok_status() as status:\n   1424           self._handle = tf_session.TF_SessionMakeCallable(\n-&gt; 1425               session._session, options_ptr, status)\n   1426       finally:\n   1427         tf_session.TF_DeleteBuffer(options_ptr)\n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/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: conv1d_2_input:0 is both fed and fetched.\n<\/code><\/pre>\n\n<p>There are quite a few posts with this error. So far I could not figure out what the problem is however. I guess it has to do with the model definition. But that is a very vanilla kind of model. So, I don't see a problem there. <\/p>\n\n<p>Any idea?<\/p>\n","answers":[{"AnswerId":"57516955","CreationDate":"2019-08-15T22:31:25.497","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Well the problem is that <code>keras-vis<\/code> latest release is from 2017 (the one available at PyPI), so if you install it using pip, you won't get the fixed version that is available on github. The solution is quite simple, remove the current version you have installed and install the one from github:<\/p>\n\n<pre><code>pip uninstall keras-vis\npip install --user git+https:\/\/github.com\/raghakot\/keras-vis.git\n<\/code><\/pre>\n\n<p>More information on this <a href=\"https:\/\/github.com\/raghakot\/keras-vis\/issues\/145\" rel=\"nofollow noreferrer\">issue<\/a>.<\/p>\n"}]}
+{"QuestionId":57516903,"AnswerCount":1,"Tags":"<machine-learning><keras>","CreationDate":"2019-08-15T22:24:21.290","AcceptedAnswerId":null,"OwnerUserId":11592605.0,"Title":"How do i test this model on a single png file","Body":"<p>I want to try this model on a single png file for prediction.\nI am very new to this field.\nThis is for study purpose.<\/p>\n\n<p>i have tried model(filename) \nand i get output like <\/p>\n\n<blockquote>\n  <blockquote>\n    <p>Tensor(\"sequential_1_36\/activation_6\/Sigmoid:0\", shape=(?, 1), dtype=float32)Tensor(\"sequential_1_36\/activation_6\/Sigmoid:0\", shape=(?, 1), dtype=float32)<\/p>\n  <\/blockquote>\n<\/blockquote>\n\n<p>if i could predict the class names.<\/p>\n\n<pre><code>   train_generator = train_datagen.flow_from_directory(\n        train_data_dir,\n        target_size=(img_width, img_height),\n        batch_size=32,\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=32,\n        class_mode='binary')\n\nmodel = Sequential()\nmodel.add(Convolution2D(32, 3, 3, input_shape=(img_width, \nimg_height,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(Convolution2D(128, 3, 3))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Convolution2D(128, 3, 3))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Flatten())\nmodel.add(Dense(256))\nmodel.add(Activation('relu'))\nmodel.add(Dense(1))\nmodel.add(Activation('sigmoid'))\n\nmodel.compile(loss='binary_crossentropy',optimizer='adam',metrics=    ['accuracy',precision,recall,fbeta_score])\n\n\n\n\n history=model.fit_generator(train_generator,samples_per_epoch=nb_train \n\n\n _samples,nb_epoch=nb_epoch,validation_data=validation_generator,nb_val    _samples=nb_validation_samples)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57521391","CreationDate":"2019-08-16T08:37:13.120","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Its quite simple, you need to load an image and make a prediction:<\/p>\n\n<pre><code>from keras.preprocessing.image import load_img, img_to_array\n\nimg = load_img(filename)\nimg_array = img_to_array(img)\nimg_array = np.expand_dims(img_array, axis=0)\n\nprediction = model.predict(img_array)\n<\/code><\/pre>\n\n<p>I assume this is what you mean by test a model on a single PNG file.<\/p>\n"}]}
+{"QuestionId":57516951,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><pytorch><autograd>","CreationDate":"2019-08-15T22:30:24.047","AcceptedAnswerId":null,"OwnerUserId":6688984.0,"Title":"Where to start on creating a method that saves desired changes in a Tensor with PyTorch?","Body":"<p>I have two tensors that I am calculating the Spearmans Rank Correlation from, and I would like to be able to have PyTorch automatically adjust the values in these Tensors in a way that increases my Spearmans Rank Correlation number as high as possible. <\/p>\n\n<p>I have explored autograd but nothing I've found has explained it simply enough.<\/p>\n\n<p>Initialized tensors: \n<code>a=Var(torch.randn(20,1),requires_grad=True)<\/code> \n<code>psfm_s=Var(torch.randn(12,20),requires_grad=True)<\/code><\/p>\n\n<p>How can I have a loop of constant adjustments of the values in these two tensors to get the highest spearmans rank correlation from 2 lists I make from these 2 tensors while having PyTorch do the work? I just need a guide of where to go. Thank you!<\/p>\n","answers":[{"AnswerId":"57518304","CreationDate":"2019-08-16T02:34:56.457","ParentId":null,"OwnerUserId":"2790047","Title":null,"Body":"<p>I'm not familiar with Spearman's Rank Correlation, but if I understand your question you're asking how to use PyTorch to solve problems other than deep networks?<\/p>\n\n<p>If that's the case then I'll provide a simple least squares example which I believe should be informative to your effort.<\/p>\n\n<p>Consider a set of 200 measurements of 10 dimensional vectors <strong>x<\/strong> and <strong>y<\/strong>. Say we want to find a linear transform from <strong>x<\/strong> to <strong>y<\/strong>.<\/p>\n\n<p>The least squares approach dictates we can accomplish this by finding the matrix <strong>M<\/strong> and vector <strong>b<\/strong> which minimize |(<strong>y<\/strong> - (<strong>M<\/strong> <strong>x<\/strong>+<strong>b<\/strong>))\u00b2|<\/p>\n\n<p>The following example code generates some example data and then uses pytorch to perform this minimization. I believe the comments are sufficient to help you understand what is occurring here.<\/p>\n\n\n\n<pre><code>import torch\nfrom torch.nn.parameter import Parameter\nfrom torch import optim\n\n# define some fake data\nM_true = torch.randn(10, 10)\nb_true = torch.randn(10, 1)\n\nx = torch.randn(200, 10, 1)\nnoise = torch.matmul(M_true, 0.05 * torch.randn(200, 10, 1))\ny = torch.matmul(M_true, x) + b_true + noise\n\n# begin optimization\n\n# define the parameters we want to optimize (using random starting values in this case)\nM = Parameter(torch.randn(10, 10))\nb = Parameter(torch.randn(10, 1))\n\n# define the optimizer and provide the parameters we want to optimize\noptimizer = optim.SGD((M, b), lr=0.1)\n\nfor i in range(500):\n    # compute loss that we want to minimize\n    y_hat = torch.matmul(M, x) + b\n    loss = torch.mean((y - y_hat)**2)\n\n    # zero the gradients of the parameters referenced by the optimizer (M and b)\n    optimizer.zero_grad()\n\n    # compute new gradients\n    loss.backward()\n\n    # update parameters M and b\n    optimizer.step()\n\n    if (i + 1) % 100 == 0:\n        # scale learning rate by factor of 0.9 every 100 steps\n        optimizer.param_groups[0]['lr'] *= 0.9\n\n        print('step', i + 1, 'mse:', loss.item())\n\n# final parameter values (data contains a torch.tensor)\nprint('Resulting parameters:')\nprint(M.data)\nprint(b.data)\n\nprint('Compare to the \"real\" values')\nprint(M_true)\nprint(b_true)\n<\/code><\/pre>\n\n<p>Of course this problem has a simple closed form solution, but this numerical approach is just to demonstrate how to use PyTorch's autograd to solve problems not necessarily neural network related. I also choose to explicitly define the matrix M and vector b here rather than using an equivalent <code>nn.Linear<\/code> layer since I think that would just confuse things.<\/p>\n\n<p>In your case you want to maximize something so make sure to negate your objective function before calling backward.<\/p>\n"}]}
+{"QuestionId":57517000,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2019-08-15T22:37:53.663","AcceptedAnswerId":null,"OwnerUserId":11929824.0,"Title":"Save the weights for a specific layer, not only visualize on Caffe","Body":"<p>I have this network loaded into net <code>net = caffe.Net('mobilenet_v2_deploy.prototxt', caffe.TEST)<\/code>. Then, to save the weights of this layer I can do <code>net.save('mymodel.caffemodel')<\/code> But, how could I <strong><em>save<\/em><\/strong> only a <strong><em>particular layer's weights<\/em><\/strong>? I know that to visualize <code>conv1<\/code> layer's weights I can do (1) <code>net.params['conv1'][0].data<\/code> but, this just outputs on the command line some text, it does not save a caffemodel-like file. <\/p>\n\n<p>(1) Outputs something such as: (does not save a file)<\/p>\n\n<pre><code>array([[[[-0.1010774 ]],\n        [[-0.03301976]],\n        [[ 0.19851202]],\n        ...,}\n<\/code><\/pre>\n\n<p>Example of the prototxt file<\/p>\n\n<pre><code>\"name: \"MOBILENET_V2\"\n#  transform_param {\n#    scale: 0.017\n#    mirror: false\n#    crop_size: 224\n#    mean_value: [103.94,116.78,123.68]\n#  }\ninput: \"data\"\ninput_dim: 1\ninput_dim: 3\ninput_dim: 224\ninput_dim: 224\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  convolution_param {\n    num_output: 32\n    bias_term: false\n    pad: 1\n    kernel_size: 3\n    stride: 2\n    weight_filler {\n      type: \"msra\"\n    }\n  }\n}\n\n<\/code><\/pre>\n","answers":[{"AnswerId":"57550858","CreationDate":"2019-08-19T05:21:01.313","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can use the method described in <a href=\"https:\/\/stackoverflow.com\/a\/45208380\/1714410\">this answer<\/a> to convert the weights to numpy and then <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.save.html\" rel=\"nofollow noreferrer\"><code>numpy.save()<\/code><\/a> only the weights that you care about.<\/p>\n"}]}
+{"QuestionId":57517104,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-15T22:53:22.070","AcceptedAnswerId":null,"OwnerUserId":11681667.0,"Title":"Error when restoring and continuing the training of a tensorflow model","Body":"<p>Having restored a model, and meta, it stops at this line of code<\/p>\n\n<pre><code>`summary_str_Train = sess.run(summaries_merged, feed_dict = Feed_Dict_Train) # gets plot points for Tensorboard Graph (No weight updating)\n<\/code><\/pre>\n\n<p>with the following error message<\/p>\n\n<p>tensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'Placeholder_1' with dtype float and shape [?,256,256,3]\n[[{{node Placeholder_1}}]]\n[[{{node conv7\/weights_1\/read}}]]<\/p>\n\n<p>I printed the list of tensors in the graph:-<\/p>\n\n<pre><code>for tensor in tf.get_default_graph().get_operations(): # list all the \ntensors in the graph\n  print(tensor.name)\n<\/code><\/pre>\n\n<p>This gave me a long list of tensors:-<\/p>\n\n<p>Placeholder<\/p>\n\n<p>Placeholder_1<\/p>\n\n<p>Const<\/p>\n\n<p>split\/split_dim<\/p>\n\n<p>split<\/p>\n\n<p>Const_1<\/p>\n\n<p>split_1\/split_dim<\/p>\n\n<p>split_1<\/p>\n\n<p>...<\/p>\n\n<p>The code is a much longer version of this:-<\/p>\n\n<pre><code>\nsess = tf.Session()\nnewest_meta_filename = newest_meta(preTrained_checkpoint_path)\nnew_saver = tf.train.import_meta_graph(newest_meta_filename) # restores the graph structure\n#Latest_Checkpoint_Path = os.path.join(preTrained_checkpoint_path, 'model.ckpt')\nnew_saver.restore(sess, tf.train.latest_checkpoint(preTrained_checkpoint_path))\nfor tensor in tf.get_default_graph().get_operations(): # list all the tensors in the graph\n  print(tensor.name)\n\nInput_Placeholder = tf.placeholder(tf.float, shape=(None, Ih,Iw, 6))\nTarget_Placeholder = tf.placeholder(tf.float, shape=(None, Ih, Iw, 3))\n\n# ... build graph either 1 gpu or many\n\nsummaries = tf.get_collection(tf.GraphKeys.SUMMARIES)\nsummaries.append(tf.summary.scalar('Training_loss', Reproduction_Loss))\nvalidation_loss = tf.summary.scalar('validation_loss', Reproduction_Loss)\n\nprint(\"Reproduction_Loss \" +str(Reproduction_Loss))\nsummaries.append(tf.summary.image('Synthesized', Synthesized_Images, 3))\nsummaries.append(tf.summary.image('Ground Truth', Target_Placeholder, 3))\n\nsaver = tf.train.Saver(tf.global_variables())\nsummaries_merged = tf.summary.merge_all()\n\ninit = tf.global_variables_initializer()\nsess = tf.Session(config=tf.ConfigProto( allow_soft_placement=True))\nsess.run(init)\nwriter = tf.summary.FileWriter(FLAGS.log_dir)\nwriter.add_graph(sess.graph)\n\n\n# ... batch some data\n\nFrame1And2_Concatenated = np.concatenate((batch_data_frame1, batch_data_frame3), 3)\nFeed_Dict_Train = {Input_Placeholder: (Frame1And2_Concatenated), Target_Placeholder: (batch_data_frame2)} # np.split 4, axis=0\nFeed_Dict_Validation = {Input_Placeholder: np.concatenate((valid_batch_data_frame1, valid_batch_data_frame3), 3), Target_Placeholder: valid_batch_data_frame2}\n\n_, Loss_Value = sess.run([Update_Op, Reproduction_Loss], feed_dict = Feed_Dict_Train) # update the weights of the model\n\nsummary_str_Train = sess.run(summaries_merged, feed_dict = Feed_Dict_Train)\n\n<\/code><\/pre>\n\n<p>Gives this full error<\/p>\n\n<pre><code>File \"...\\ms-python.python-2019.6.24221\\pythonFiles\\ptvsd_launcher\nmain(ptvsdArgs)\nFile \"...\\vscode\\extensions\\ms-python.python-2019.6.24221\\pythonFiles\\lib\\python\\ptvn main\nrun()\nFile \"...\\python\\ptvn run_file \nrunpy.run_path(target, run_name='__main__')\nFile \"...\\Python37\\lib\\runpy.py\", line 263, in run_path\npkg_name=pkg_name, script_name=fname)\nFile \"...\\Python37\\lib\\runpy.py\", line 96, in _run_modu\nmod_name, mod_spec, pkg_name, script_name)\nFile \"...\\Python37\\lib\\runpy.py\", line 85, in _run_code\nexec(code, run_globals)\nFile \"...\\voxel_flow_train_multi_gpu.py\", line 6\nnew_saver = tf.train.import_meta_graph(newest_meta_filename) # restores the graph structure\nFile \"...\\Python37\\lib\\site-packages\\tensorflow\\python\\5, in import_meta_graph\nmeta_graph_or_file, clear_devices, import_scope, **kwargs)[0]\nFile \"...\\Python37\\lib\\site-packages\\tensorflow\\python\\7, in _import_meta_graph_with_return_elements\n**kwargs))\nFile \"...\\Python37\\lib\\site-packages\\tensorflow\\python\\ne 806, in import_scoped_meta_graph_with_return_elements\nreturn_elements=return_elements)\nFile \"...n\\Python37\\lib\\site-packages\\tensorflow\\python\\07, in new_func\nreturn func(*args, **kwargs)\nFile \"...\\Python37\\lib\\site-packages\\tensorflow\\python\\ 442, in import_graph_def_ProcessNewOps(graph)\nFile \"...\\Python37\\lib\\site-packages\\tensorflow\\python\\ 235, in _ProcessNewOps\nfor new_op in graph._add_new_tf_operations(compute_devices=False):  # pylint: disable=protec\nFile \"...\\Python37\\lib\\site-packages\\tensorflow\\python\\, in _add_new_tf_operations\nfor c_op in c_api_util.new_tf_operations(self)\nFile \"...\\Python37\\lib\\site-packages\\tensorflow\\python\\, in &lt;listcomp&gt;\nfor c_op in c_api_util.new_tf_operations(self)\nFile \"...n\\Python37\\lib\\site-packages\\tensorflow\\python\\, in _create_op_from_tf_operation\nret = Operation(c_op, self)\nFile \"...n\\Python37\\lib\\site-packages\\tensorflow\\python\\, in __init__\nself._traceback = tf_stack.extract_stack()\ntensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'Placeholder_1' with dtype float and shape [?,256,256,3]\n[[{{node Placeholder_1}}]]\n[[{{node GroupCrossDeviceControlEdges_0\/conv2_1\/Relu}}]]\n\n<\/code><\/pre>\n\n<p>When I train without restoring first, it works fine.  It is only when I restore that the error occurs<\/p>\n","answers":[]}
+{"QuestionId":57517121,"AnswerCount":1,"Tags":"<python><3d><conv-neural-network><pytorch><crop>","CreationDate":"2019-08-15T22:55:31.960","AcceptedAnswerId":57517558.0,"OwnerUserId":11354575.0,"Title":"How can I do the centercrop of 3D volumes inside the network model with pytorch","Body":"<p>In <code>keras<\/code>, there is <code>Cropping3D<\/code> layer for centercropping tensors of 3D volumnes inside the neural network. However, I failed to find out anything similar in pytorch, though they have <code>torchvision.transforms.CenterCrop(size)<\/code> for 2D images.<\/p>\n\n<p>How can I do the cropping inside the network? Otherwise I need to do it in preprocessing which is the last thing I want to do for specific reasons.<\/p>\n\n<p>Do I need to write a custom layer like slicing the input tensors along each axices? Hope to get some inspiration for this<\/p>\n","answers":[{"AnswerId":"57517558","CreationDate":"2019-08-16T00:08:54.440","ParentId":null,"OwnerUserId":"2790047","Title":null,"Body":"<p>In PyTorch you don't necessarily need to write layers for everything, often you can just do what you want directly during the forward pass. The basic rules you need to keep in mind when operating on torch tensors for which you will need to compute gradients are<\/p>\n\n<ol>\n<li>Don't convert torch tensors to other types for computation (e.g. use <code>torch.sum<\/code> instead of converting to numpy and using <code>numpy.sum<\/code>).<\/li>\n<li>Don't perform in-place operations (e.g. changing one element of a tensor or using inplace operators, so use <code>x = x + ...<\/code> instead of <code>x += ...<\/code>).<\/li>\n<\/ol>\n\n<p>That said, you can just use slicing, maybe it would look something like this<\/p>\n\n<pre><code>def forward(self, x):\n    ...\n    x = self.conv3(x)\n    x = x[:, :, 5:20, 5:20]    # crop out part of the feature map\n    x = self.relu3(x)\n    ...\n<\/code><\/pre>\n"}]}
+{"QuestionId":57517406,"AnswerCount":0,"Tags":"<python><keras><neural-network><data-visualization><keras-layer>","CreationDate":"2019-08-15T23:39:18.390","AcceptedAnswerId":null,"OwnerUserId":3578468.0,"Title":"Keras Visualization Toolkit: Models with flexible input length --- Tried to convert 'x' to a tensor and failed","Body":"<p>I am not sure if the flexible input length really is the problem. Maybe it's something else. Anyway, I have a model  that I want to visualize, that has an embedding layer as the input layer.<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense, Conv1D, Embedding\nimport numpy as np\n\nmodel = Sequential()\n\nvocab_size = 80\nencoding_dims = 30\ntext_length=1014\nfilter_cnt=256\noutput_dims=10\nloss='categorical_crossentropy'\noptimizer='adam'\n\nembedding_layer = Embedding(vocab_size, encoding_dims, input_length=text_length,\n                            trainable=False, name='emb')\nembedding_layer.build((None,))\nembedding_layer.set_weights([np.random.random((vocab_size,encoding_dims))])\nmodel.add(embedding_layer)\n\nmodel.add(Conv1D(filters=filter_cnt, kernel_size=7, activation='relu', input_shape=(text_length, encoding_dims)))\n\n# ... more layers\n\nmodel.add(Dense(units=output_dims, activation='softmax'))\n\n\nmodel.compile(loss=loss, optimizer=optimizer, metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>When I run the visualization code from <a href=\"https:\/\/github.com\/raghakot\/keras-vis\" rel=\"nofollow noreferrer\">https:\/\/github.com\/raghakot\/keras-vis<\/a><\/p>\n\n<pre><code>from vis.losses import ActivationMaximization\nfrom vis.regularizers import TotalVariation, LPNorm\n\nfilter_indices = [1, 2, 3]\n\n# Tuple consists of (loss_function, weight)\n# Add regularizers as needed.\nlosses = [\n    (ActivationMaximization(model.layers[1], filter_indices), 1),\n    (LPNorm(model.input), 10),\n    (TotalVariation(model.input), 10)\n]\n\nfrom vis.optimizer import Optimizer\n\noptimizer = Optimizer(model.input, losses)\nopt_img, grads, _ = optimizer.minimize()\n<\/code><\/pre>\n\n<p>I get this error:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/op_def_library.py in _apply_op_helper(self, op_type_name, name, **keywords)\n    510                 as_ref=input_arg.is_ref,\n--&gt; 511                 preferred_dtype=default_dtype)\n    512           except TypeError as err:\n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/ops.py in internal_convert_to_tensor(value, dtype, name, as_ref, preferred_dtype, ctx, accept_symbolic_tensors)\n   1174     if ret is None:\n-&gt; 1175       ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)\n   1176 \n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/constant_op.py in _constant_tensor_conversion_function(v, dtype, name, as_ref)\n    303   _ = as_ref\n--&gt; 304   return constant(v, dtype=dtype, name=name)\n    305 \n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/constant_op.py in constant(value, dtype, shape, name)\n    244   return _constant_impl(value, dtype, shape, name, verify_shape=False,\n--&gt; 245                         allow_broadcast=True)\n    246 \n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/constant_op.py in _constant_impl(value, dtype, shape, name, verify_shape, allow_broadcast)\n    282           value, dtype=dtype, shape=shape, verify_shape=verify_shape,\n--&gt; 283           allow_broadcast=allow_broadcast))\n    284   dtype_value = attr_value_pb2.AttrValue(type=tensor_value.tensor.dtype)\n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/tensor_util.py in make_tensor_proto(values, dtype, shape, verify_shape, allow_broadcast)\n    453     if values is None:\n--&gt; 454       raise ValueError(\"None values not supported.\")\n    455     # if dtype is provided, forces numpy array to be the type\n\nValueError: None values not supported.\n\nDuring handling of the above exception, another exception occurred:\n\nValueError                                Traceback (most recent call last)\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/op_def_library.py in _apply_op_helper(self, op_type_name, name, **keywords)\n    524               observed = ops.internal_convert_to_tensor(\n--&gt; 525                   values, as_ref=input_arg.is_ref).dtype.name\n    526             except ValueError as err:\n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/ops.py in internal_convert_to_tensor(value, dtype, name, as_ref, preferred_dtype, ctx, accept_symbolic_tensors)\n   1174     if ret is None:\n-&gt; 1175       ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)\n   1176 \n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/constant_op.py in _constant_tensor_conversion_function(v, dtype, name, as_ref)\n    303   _ = as_ref\n--&gt; 304   return constant(v, dtype=dtype, name=name)\n    305 \n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/constant_op.py in constant(value, dtype, shape, name)\n    244   return _constant_impl(value, dtype, shape, name, verify_shape=False,\n--&gt; 245                         allow_broadcast=True)\n    246 \n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/constant_op.py in _constant_impl(value, dtype, shape, name, verify_shape, allow_broadcast)\n    282           value, dtype=dtype, shape=shape, verify_shape=verify_shape,\n--&gt; 283           allow_broadcast=allow_broadcast))\n    284   dtype_value = attr_value_pb2.AttrValue(type=tensor_value.tensor.dtype)\n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/tensor_util.py in make_tensor_proto(values, dtype, shape, verify_shape, allow_broadcast)\n    453     if values is None:\n--&gt; 454       raise ValueError(\"None values not supported.\")\n    455     # if dtype is provided, forces numpy array to be the type\n\nValueError: None values not supported.\n\nDuring handling of the above exception, another exception occurred:\n\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-8-82d88819f098&gt; in &lt;module&gt;\n     14 from vis.optimizer import Optimizer\n     15 \n---&gt; 16 optimizer = Optimizer(model.input, losses)\n     17 opt_img, grads, _ = optimizer.minimize()\n     18 opt_img.shape\n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/vis\/optimizer.py in __init__(self, input_tensor, losses, input_range, wrt_tensor, norm_grads)\n     50             # Perf optimization. Don't build loss function with 0 weight.\n     51             if weight != 0:\n---&gt; 52                 loss_fn = weight * loss.build_loss()\n     53                 overall_loss = loss_fn if overall_loss is None else overall_loss + loss_fn\n     54                 self.loss_names.append(loss.name)\n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/vis\/regularizers.py in build_loss(self)\n     69             end_slice = np.roll(end_slice, 1).tolist()\n     70 \n---&gt; 71         tv = K.sum(K.pow(tv, self.beta \/ 2.))\n     72         return normalize(self.img, tv)\n     73 \n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/keras\/backend\/tensorflow_backend.py in pow(x, a)\n   1581         A tensor.\n   1582     \"\"\"\n-&gt; 1583     return tf.pow(x, a)\n   1584 \n   1585 \n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/tensorflow\/python\/util\/dispatch.py in wrapper(*args, **kwargs)\n    178     \"\"\"Call target, and fall back on dispatchers if there is a TypeError.\"\"\"\n    179     try:\n--&gt; 180       return target(*args, **kwargs)\n    181     except (TypeError, ValueError):\n    182       # Note: convert_to_eager_tensor currently raises a ValueError, not a\n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/tensorflow\/python\/ops\/math_ops.py in pow(x, y, name)\n    375   \"\"\"\n    376   with ops.name_scope(name, \"Pow\", [x]) as name:\n--&gt; 377     return gen_math_ops._pow(x, y, name=name)\n    378 \n    379 \n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/tensorflow\/python\/ops\/gen_math_ops.py in _pow(x, y, name)\n   6163   # Add nodes to the TensorFlow graph.\n   6164   _, _, _op = _op_def_lib._apply_op_helper(\n-&gt; 6165         \"Pow\", x=x, y=y, name=name)\n   6166   _result = _op.outputs[:]\n   6167   _inputs_flat = _op.inputs\n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/op_def_library.py in _apply_op_helper(self, op_type_name, name, **keywords)\n    527               raise ValueError(\n    528                   \"Tried to convert '%s' to a tensor and failed. Error: %s\" %\n--&gt; 529                   (input_name, err))\n    530             prefix = (\"Input '%s' of '%s' Op has type %s that does not match\" %\n    531                       (input_name, op_type_name, observed))\n\nValueError: Tried to convert 'x' to a tensor and failed. Error: None values not supported.\n<\/code><\/pre>\n\n<p>Not sure what that is about. However, when I define the model without the embedding layer, the visualization code runs fine. So I figured I just remove the embedding layer before visualization like so, because I don't need it for the visualization of the activation maximizing input anyway:<\/p>\n\n<pre><code>from copy import deepcopy\nmodel_old = deepcopy(model)\nmodel = Sequential()\nfor layer in model_old.layers[1:]: # just exclude first layer from copying\n    model.add(layer)\nmodel.compile(loss=loss, optimizer=optimizer, metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>However, when I do this, I get another error:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nAttributeError                            Traceback (most recent call last)\n&lt;ipython-input-16-82d88819f098&gt; in &lt;module&gt;\n     14 from vis.optimizer import Optimizer\n     15 \n---&gt; 16 optimizer = Optimizer(model.input, losses)\n     17 opt_img, grads, _ = optimizer.minimize()\n     18 opt_img.shape\n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/vis\/optimizer.py in __init__(self, input_tensor, losses, input_range, wrt_tensor, norm_grads)\n     50             # Perf optimization. Don't build loss function with 0 weight.\n     51             if weight != 0:\n---&gt; 52                 loss_fn = weight * loss.build_loss()\n     53                 overall_loss = loss_fn if overall_loss is None else overall_loss + loss_fn\n     54                 self.loss_names.append(loss.name)\n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/vis\/losses.py in build_loss(self)\n     74 \n     75     def build_loss(self):\n---&gt; 76         layer_output = self.layer.output\n     77 \n     78         # For all other layers it is 4\n\n~\/anaconda3\/envs\/ccnn2\/lib\/python3.7\/site-packages\/keras\/engine\/base_layer.py in output(self)\n    811         if len(self._inbound_nodes) &gt; 1:\n    812             raise AttributeError('Layer ' + self.name +\n--&gt; 813                                  ' has multiple inbound nodes, '\n    814                                  'hence the notion of \"layer output\" '\n    815                                  'is ill-defined. '\n\nAttributeError: Layer dense_6 has multiple inbound nodes, hence the notion of \"layer output\" is ill-defined. Use `get_output_at(node_index)` instead.\n<\/code><\/pre>\n\n<p>So it complains about the second layer having multiple input nodes. Also no idea what that is about.<\/p>\n\n<p>Can you tell me what the problem is and how I can correctly remove the embeding layer so the visualization accepts the model?<\/p>\n","answers":[]}
+{"QuestionId":57517650,"AnswerCount":0,"Tags":"<tensorflow><one-hot-encoding>","CreationDate":"2019-08-16T00:30:25.233","AcceptedAnswerId":null,"OwnerUserId":6450489.0,"Title":"One-hot all-zero entries in Tensorflow","Body":"<p>I have a pixel classification application with training labels (per pixel) being one-hot encoded. For some reason there are pixels that doesn't have valid labels. One solution might be to use sample_weight option to indirectly mask those pixels from affecting loss and gradient calculations. But I found that in Tensorflow implementation, I can generate one-hot labels being all-zero (which is an invalid one-hot encoding) for those pixels and I will not get any error during model build and training! Are these two methods the same? Is all-zero one-hot label disables loss and gradient calculation for that specific pixel?<\/p>\n","answers":[]}
+{"QuestionId":57517740,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-08-16T00:50:27.313","AcceptedAnswerId":null,"OwnerUserId":11354575.0,"Title":"Pytorch custom dataset: ValueError: some of the strides of a given numpy array are negative","Body":"<p>I wrote a custom pytorch dataset, but ran into an error thhat seems quite unintelligible.<\/p>\n\n<p>My custom dataset, <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class data_from_xlsx(Dataset):\n    def __init__(self, xlsx_fp, path_col, class_cols_list):\n        self.xlsx_file = pd.read_excel(xlsx_fp)\n        self.path_col = path_col\n        self.class_cols_list = class_cols_list\n\n    def __len__(self):\n        return get_xlsx_length(self.xlsx_file)\n\n    def __getitem__(self, index):\n        file_path = cols_from_xlsx(self.xlsx_file, index, 1, self.path_col) \n        feature = load_nii_file(file_path) # get 3D volume (x, y, z) \n        feature = np.expand_dims(feature, axis=0) # add channel (c, x, y, z)\n        label = cols_from_xlsx(self.xlsx_file, index, 1, self.class_cols_list) # get label\n        return feature, label.astype(np.bool)\n\n\ndef main():\ndataset = data_from_xlsx(\"train.xlsx\", \"file_path\", [\"pos\", \"neg\"], transformations, aug=True)\n    data_loader = DataLoader(dataset, batch_size=4, shuffle=True)\n\n    for (f, l) in data_loader:\n        print(\"f shape\", f.shape)\n        print(\"l shape\", l.shape)\n<\/code><\/pre>\n\n<p>An error is reported when I ran <code>main()<\/code>,<\/p>\n\n<pre><code> File \"d:\\pytorch\\lib\\site-packages\\torch\\utils\\data\\dataloader.py\", line 346, in __next__\n    data = self.dataset_fetcher.fetch(index)  # may raise StopIteration\n  File \"d:\\pytorch\\lib\\site-packages\\torch\\utils\\data\\_utils\\fetch.py\", line 47, in fetch\n    return self.collate_fn(data)\n  File \"d:\\pytorch\\lib\\site-packages\\torch\\utils\\data\\_utils\\collate.py\", line 80, in default_collate\n    return [default_collate(samples) for samples in transposed]\n  File \"d:\\pytorch\\lib\\site-packages\\torch\\utils\\data\\_utils\\collate.py\", line 80, in &lt;listcomp&gt;\n    return [default_collate(samples) for samples in transposed]\n  File \"d:\\pytorch\\lib\\site-packages\\torch\\utils\\data\\_utils\\collate.py\", line 65, in default_collate\n    return default_collate([torch.as_tensor(b) for b in batch])\n  File \"d:\\pytorch\\lib\\site-packages\\torch\\utils\\data\\_utils\\collate.py\", line 65, in &lt;listcomp&gt;\n    return default_collate([torch.as_tensor(b) for b in batch])\nValueError: some of the strides of a given numpy array are negative. This is currently not supported, but will be added in future release\n<\/code><\/pre>\n\n<p>The reported error does't make sense to me, so I googled it. At first I thought I didn't change the <code>feature<\/code> from <code>numpy.array<\/code> to tensor, so I tried <code>feature = torch.from_array(feature.copy())<\/code> and also tried <code>transforms.TOTensor()<\/code> but both attempts failed.<\/p>\n","answers":[{"AnswerId":"57528181","CreationDate":"2019-08-16T16:32:25.640","ParentId":null,"OwnerUserId":"11354575","Title":null,"Body":"<p>Thanks to the advice from @jodag and @UsmanAli, I sovled this by return <code>torch.from_numpy(feature.copy())<\/code> and <code>torch.tensor(label.astype(np.bool))<\/code>\nSo the whole thing should be,<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class data_from_xlsx(Dataset):\n    def __init__(self, xlsx_fp, path_col, class_cols_list):\n        self.xlsx_file = pd.read_excel(xlsx_fp)\n        self.path_col = path_col\n        self.class_cols_list = class_cols_list\n\n    def __len__(self):\n        return get_xlsx_length(self.xlsx_file)\n\n    def __getitem__(self, index):\n        file_path = cols_from_xlsx(self.xlsx_file, index, 1, self.path_col) \n        feature = load_nii_file(file_path) # get 3D volume (x, y, z) \n        feature = np.expand_dims(feature, axis=0) # add channel (c, x, y, z)\n        label = cols_from_xlsx(self.xlsx_file, index, 1, self.class_cols_list) # get label\n        return torch.from_numpy(feature.copy()), torch.tensor(label.astype(np.bool))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57517978,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-16T01:35:45.170","AcceptedAnswerId":null,"OwnerUserId":3550316.0,"Title":"Tensorflow: How to zero row of weight depending on input from placeholder","Body":"<p>Say I have an 3 X 3 weight W of ones. Then say I have a placeholder, <\/p>\n\n<pre><code>w = tf.get_variable('w', [3,3], \n                    initializer = tf.constant_initializer(np.ones((3,3)))\nidx = tf.placeholder(tf.int32, [None,], name = 'idx')\n<\/code><\/pre>\n\n<p>I want to 'zero' a row of the weight depending on the placeholder.\nSo for idx = 0, I would zero the first row of the weight. \nThen when idx = 1, I would unzero the first row of the weight and zero the second row of the weight. \nIn code <\/p>\n\n<pre><code>with tf.Session() as sess: \n       weight = sess.run([w], idx = [0,1,2])\n       print(weight)\n<\/code><\/pre>\n\n<p>I want the output to look like <\/p>\n\n<pre><code>at idx = 0 \n[[0,0,0],\n[1, 1, 1],\n[1, 1, 1],]\nat idx = 1\n[[1,1,1],\n[0,0,0],\n[1,1,1],]\nat idx = 2\n[[1,1,1],\n[1,1,1],\n[0,0,0]]\n<\/code><\/pre>\n\n<p>Is there a way to do this? The problem with <code>tf.scatter_nd_update<\/code> is that it applies the update the specified index, but does not 'reset' the value of the variable for the next subsequent index.\nThanks<\/p>\n","answers":[]}
+{"QuestionId":57517992,"AnswerCount":1,"Tags":"<keras><tensorflow2.0>","CreationDate":"2019-08-16T01:38:22.903","AcceptedAnswerId":null,"OwnerUserId":7820444.0,"Title":"Can I use dictionary in keras customized model?","Body":"<p>I recently read a paper about UNet++\uff0cand I want to implement this structure with tensorflow-2.0 and keras customized model. As the structure is so complicated, I decided to manage the keras layers by a dictionary. Everything went well in training, but an error occurred while saving the model. Here is a minimum code to show the error:<\/p>\n\n<pre><code>class DicModel(tf.keras.Model):\n    def __init__(self):\n        super(DicModel, self).__init__(name='SequenceEECNN')\n        self.c = {}\n        self.c[0] = tf.keras.Sequential([\n            tf.keras.layers.Conv2D(32, 3,activation='relu',padding='same'),\n            tf.keras.layers.BatchNormalization()]\n        )\n        self.c[1] = tf.keras.layers.Conv2D(3,3,activation='softmax',padding='same')\n    def call(self,images):\n        x = self.c[0](images)\n        x = self.c[1](x)\n        return x\n\nX_train,y_train = load_data()\nX_test,y_test = load_data()\n\nclass_weight.compute_class_weight('balanced',np.ravel(np.unique(y_train)),np.ravel(y_train))\n\nmodel = DicModel()\nmodel_name = 'test'\ntensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir='logs\/'+model_name+'\/')\nearly_stop_callback = tf.keras.callbacks.EarlyStopping(monitor='val_loss',patience=100,mode='min')\n\nmodel.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001),\n              loss=tf.keras.losses.sparse_categorical_crossentropy,\n              metrics=['accuracy'])\n\nresults = model.fit(X_train,y_train,batch_size=4,epochs=5,validation_data=(X_test,y_test),\n                    callbacks=[tensorboard_callback,early_stop_callback],\n                    class_weight=[0.2,2.0,100.0])\n\nmodel.save_weights('model\/'+model_name,save_format='tf')\n<\/code><\/pre>\n\n<p>The error information is:<\/p>\n\n<pre><code>Traceback (most recent call last):\n\n  File \"\/media\/xrzhang\/Data\/ZHS\/Research\/CNN-TF2\/learn_tf2\/test_model.py\", line 61, in \\&lt;module&gt;\n\n    model.save_weights('model\/'+model_name,save_format='tf')\n\n  File \"\/media\/xrzhang\/Data\/ZHS\/Research\/CNN-TF2\/venv\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/network.py\", line 1328, in save_weights\n\n    self.\\_trackable_saver.save(filepath, session=session)\n\n  File \"\/media\/xrzhang\/Data\/ZHS\/Research\/CNN-TF2\/venv\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/tracking\/util.py\", line 1106, in save\n\n    file_prefix=file_prefix_tensor, object_graph_tensor=object_graph_tensor)\n\n  File \"\/media\/xrzhang\/Data\/ZHS\/Research\/CNN-TF2\/venv\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/tracking\/util.py\", line 1046, in \\_save_cached_when_graph_building\n\n    object_graph_tensor=object_graph_tensor)\n\n  File \"\/media\/xrzhang\/Data\/ZHS\/Research\/CNN-TF2\/venv\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/tracking\/util.py\", line 1014, in \\_gather_saveables\n\n    feed_additions) = self.\\_graph_view.serialize_object_graph()\n\n  File \"\/media\/xrzhang\/Data\/ZHS\/Research\/CNN-TF2\/venv\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/tracking\/graph_view.py\", line 379, in serialize_object_graph\n\n    trackable_objects, path_to_root = self.\\_breadth_first_traversal()\n\n  File \"\/media\/xrzhang\/Data\/ZHS\/Research\/CNN-TF2\/venv\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/tracking\/graph_view.py\", line 199, in \\_breadth_first_traversal\n\n    for name, dependency in self.list_dependencies(current_trackable):\n\n  File \"\/media\/xrzhang\/Data\/ZHS\/Research\/CNN-TF2\/venv\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/tracking\/graph_view.py\", line 159, in list_dependencies\n\n    return obj.\\_checkpoint_dependencies\n\n  File \"\/media\/xrzhang\/Data\/ZHS\/Research\/CNN-TF2\/venv\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/tracking\/data_structures.py\", line 690, in \\_\\_getattribute\\_\\_\n\n    return object.\\_\\_getattribute\\_\\_(self, name)\n\n  File \"\/media\/xrzhang\/Data\/ZHS\/Research\/CNN-TF2\/venv\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/tracking\/data_structures.py\", line 732, in \\_checkpoint_dependencies\n\n    \"ignored.\" % (self,))\n\nValueError: Unable to save the object {0: \\&lt;tensorflow.python.keras.engine.sequential.Sequential object at 0x7fb5c6c36588&gt;, 1: \\&lt;tensorflow.python.keras.layers.convolutional.Conv2D object at 0x7fb5c6c36630&gt;} (a dictionary wrapper constructed automatically on attribute assignment). The wrapped dictionary contains a non-string key which maps to a trackable object or mutable data structure.\n\n\n\nIf you don't need this dictionary checkpointed, wrap it in a tf.contrib.checkpoint.NoDependency object; it will be automatically un-wrapped and subsequently ignored.\n\n<\/code><\/pre>\n\n<p>The tf.contrib.checkpoint.NoDependency seems has been removed from Tensorflow-2.0 (<a href=\"https:\/\/medium.com\/tensorflow\/whats-coming-in-tensorflow-2-0-d3663832e9b8\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/tensorflow\/whats-coming-in-tensorflow-2-0-d3663832e9b8<\/a>). How can I fix this issue? Or should I just give up using dictionary in customized Keras Model. Thank you for your time and helps!<\/p>\n","answers":[{"AnswerId":"57974800","CreationDate":"2019-09-17T13:00:35.103","ParentId":null,"OwnerUserId":"12079723","Title":null,"Body":"<p>Use string keys. For some reason tensorflow doesn't like int keys. <\/p>\n"}]}
+{"QuestionId":57518000,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-16T01:40:10.687","AcceptedAnswerId":null,"OwnerUserId":6257027.0,"Title":"How to make a tf.app.run script an importable module","Body":"<p>The tensorflow scripts are usually of the following structure.\nHow do I convert the script into an importable module so that I can invoke the main function from another python script?<\/p>\n\n<pre><code>flags = tf.app.flags\nflags.DEFINE_string('master', '', 'Name of the TensorFlow master to use.')\nflags.DEFINE_integer('task', 0, 'task id')\nflags.DEFINE_integer('num_clones', 1, 'Number of clones to deploy per worker.')\n....\n\ndef main(_):\n   ...\n\nif __name__ == '__main__':\n  tf.app.run()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57518057,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2019-08-16T01:52:19.477","AcceptedAnswerId":57568704.0,"OwnerUserId":226526.0,"Title":"How can I convert an image from pixels to one-hot encodings?","Body":"<p>I have a PNG image I am loading within Tensorflow using:<\/p>\n\n<pre><code>image = tf.io.decode_png(tf.io.read_file(path), channels=3)\n<\/code><\/pre>\n\n<p>The image contains pixels that match a lookup like this:<\/p>\n\n<pre><code>image_colors = [\n    (0, 0, 0),        # black\n    (0.5, 0.5, 0.5),  # grey\n    (1, 0.5, 1),      # pink\n]\n<\/code><\/pre>\n\n<p>How can I convert it so that the output has the pixels mapped into one-hot encodings where the hot component would be the matching color?<\/p>\n","answers":[{"AnswerId":"57563310","CreationDate":"2019-08-19T20:08:18.403","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>There might be better approaches than this.<\/p>\n\n<pre><code>def map_colors(pixel):\n    if pixel[0] &lt; 10 and pixel[1] &lt; 10 and pixel[2] &lt; 10:   ## Black\n        return 0\n    elif pixel[0] &gt; 245 and pixel[1] &gt; 245 and pixel[2] &gt; 245:  ## White\n        return 1\n    else:\n        return 11\n\n\nimage = tf.io.decode_png(tf.io.read_file(path), channels=3)\nimg_shape = image.shape\n# Arrange the pixels in RGB format from 3D array to 2D array.\nimage = tf.reshape(image, [-1, 3])\n\ncolors = tf.map_fn(lambda x: map_colors(x), image)\none_hot = tf.one_hot(colors, depth=12)\nprint(one_hot)\n\n# If necessary\nimage_one_hot = tf.reshape(one_hot, [img_shape[0], img_shape[1], 12])\n<\/code><\/pre>\n\n<p>Ensure that in <code>map_colors<\/code>, you put all your 12 colors with the range of RGB color values they can accept. Make sure that all the combinations are covered or else, add an extra class of None of the above and name it 12.<\/p>\n"},{"AnswerId":"57568704","CreationDate":"2019-08-20T07:32:57.160","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>Let me assume for convenience that all values in <code>image_colors<\/code> are in <code>[0, 255]<\/code>:<\/p>\n\n<pre><code>image_colors = [\n    (0, 0, 0),  # black\n    (127, 127, 127),  # grey\n    (255, 127, 255),  # pink\n]\n<\/code><\/pre>\n\n<p>My approach maps pixels into one-hot values as follows:<\/p>\n\n<pre><code># Create a \"color reference\" tensor from image_colors\ncolor_reference = tf.cast(tf.constant(image_colors), dtype=tf.uint8)\n\n# Load the image and obtain tensor with one-hot values\nimage = tf.io.decode_png(tf.io.read_file(path), channels=3)\ncomp = tf.equal(image[..., None, :], color_reference)\none_hot = tf.cast(tf.reduce_all(comp, axis=-1), dtype=tf.float32)\n<\/code><\/pre>\n\n<p>Note that you can easily add new colors to <code>image_colors<\/code> without changing the TF implementation. Also, this assumes that all pixels in the <code>image<\/code> are in <code>image_colors<\/code>. If that is not the case, one could define a range for each color and then use other <code>tf.math<\/code> operations (e.g. <code>tf.greater<\/code> and <code>tf.less<\/code>) instead of <code>tf.equal<\/code>.<\/p>\n"}]}
+{"QuestionId":57518079,"AnswerCount":1,"Tags":"<python><tensorflow><iterator><tensorflow-datasets><tensorflow2.0>","CreationDate":"2019-08-16T01:57:35.067","AcceptedAnswerId":57520194.0,"OwnerUserId":735008.0,"Title":"retrieving the next element from tf.data.Dataset in tensorflow 2.0 beta","Body":"<p>Before tensorflow 2.0-beta, to retrieve the first element from tf.data.Dataset, we may use a iterator as shown below:<\/p>\n\n<pre><code>#!\/usr\/bin\/python\n\nimport tensorflow as tf\n\ntrain_dataset = tf.data.Dataset.from_tensor_slices([1.0, 2.0, 3.0, 4.0])\niterator = train_dataset.make_one_shot_iterator()\nwith tf.Session() as sess:\n    # 1.0 will be printed.\n    print (sess.run(iterator.get_next()))\n<\/code><\/pre>\n\n<p>In tensorflow 2.0-beta, it seems that the above <strong>one-shot-iterator<\/strong> is now deprecated. To print out the entire elements we may use the following <strong>for<\/strong> approach.<\/p>\n\n<pre><code>#!\/usr\/bin\/python\n\nimport tensorflow as tf\n\ntrain_dataset = tf.data.Dataset.from_tensor_slices([1.0, 2.0, 3.0, 4.0])\n\nfor data in train_dataset:\n    # 1.0, 2.0, 3.0, and 4.0 will be printed.\n    print (data.numpy())\n<\/code><\/pre>\n\n<p>However, if we only want to retrieve exactly one element from tf.data.Dataset, then how can we do with tensorflow 2.0 beta? It seems that <code>next(train_dataset)<\/code> is not supported. It could be done easily with the old one shot iterator as shown above, but it's not very obvious with the new <strong>for<\/strong> based approach.<\/p>\n\n<p>Any suggestion is welcomed.<\/p>\n","answers":[{"AnswerId":"57520194","CreationDate":"2019-08-16T06:59:25.633","ParentId":null,"OwnerUserId":"2455494","Title":null,"Body":"<p>You can <code>.take(1)<\/code> from the dataset:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>for elem in train_dataset.take(1):\n  print (elem.numpy())\n<\/code><\/pre>\n"}]}
+{"QuestionId":57518585,"AnswerCount":0,"Tags":"<python><tensorflow><recurrent-neural-network><seq2seq>","CreationDate":"2019-08-16T03:25:29.420","AcceptedAnswerId":null,"OwnerUserId":11839345.0,"Title":"Why does my testing loss increases in my Seq2Seq?","Body":"<p>I'm building a Seq2seq neural network. It's a <strong>video to natural langage<\/strong> model<\/p>\n\n<h1>The problem:<\/h1>\n\n<p>My training loss decreases normaly but my testing loss increases.<\/p>\n\n<p>Also, my training accuracy inscreases while my testing accuracy is especially low and decreases.<\/p>\n\n<pre><code>Epoch 1 ; Batch loss: 5.397328 ; Test loss: 5.954748 ; Batch accuracy: 51.22% ; Test accuracy: 00.86%\nEpoch 2 ; Batch loss: 4.707819 ; Test loss: 6.127879 ; Batch accuracy: 52.87% ; Test accuracy: 00.86%\nEpoch 3 ; Batch loss: 4.348535 ; Test loss: 6.274649 ; Batch accuracy: 54.26% ; Test accuracy: 00.86%\n...\nEpoch 27 ; Batch loss: 1.005701 ; Test loss: 13.232792 ; Batch accuracy: 74.86% ; Test accuracy: 00.54%\nEpoch 28 ; Batch loss: 0.919244 ; Test loss: 13.706192 ; Batch accuracy: 75.82% ; Test accuracy: 00.54%\nEpoch 29 ; Batch loss: 0.861092 ; Test loss: 12.981027 ; Batch accuracy: 76.34% ; Test accuracy: 00.20%\nEpoch 30 ; Batch loss: 0.820653 ; Test loss: 13.580329 ; Batch accuracy: 76.79% ; Test accuracy: 00.51%\n...\nEpoch 265 ; Batch loss: 0.000597 ; Test loss: 24.191395 ; Batch accuracy: 84.18% ; Test accuracy: 00.27%\n<\/code><\/pre>\n\n<hr>\n\n<h1>My code:<\/h1>\n\n<p>I've got <strong>two decoders<\/strong> that share the same weights.<\/p>\n\n<ul>\n<li>One is used for training, with a <strong>TrainingHelper<\/strong><\/li>\n<li>The second is used for inference, with a <strong>GreedyEmbeddingHelper<\/strong><\/li>\n<\/ul>\n\n<pre class=\"lang-py prettyprint-override\"><code>def decoder(target, encoder_last_state, encoder_outputs, caption_lengths):\n  # Size of the batch\n  n_data = tf.shape(target)[0]\n\n  with tf.name_scope(\"decoder\"):\n\n    # Performing embedding...\n    decoder_inputs = embeddings(target) \n\n    decoder_cell = tf.nn.rnn_cell.BasicLSTMCell(dec_units)    \n    decoder_initial_state = encoder_last_state\n\n    # Final dense layer. It should be common to both decoders\n    output_layer = tf.layers.Dense(vocab_size, kernel_initializer = tf.truncated_normal_initializer(mean=0.0, stddev=0.1), name=\"decoder_dense\")\n\n    # Training decoder\n    with tf.variable_scope(\"decoder\"):\n      training_helper = tf.contrib.seq2seq.TrainingHelper(decoder_inputs, caption_lengths)\n      training_decoder = tf.contrib.seq2seq.BasicDecoder(decoder_cell, training_helper, decoder_initial_state, output_layer)\n      training_decoder_outputs, _, _ = tf.contrib.seq2seq.dynamic_decode(training_decoder, maximum_iterations=max_length)\n\n    # Inference decoder\n    with tf.variable_scope(\"decoder\", reuse=True):\n      start_tokens = tf.tile(tf.constant([lang_tokenizer.word_index['&lt;start&gt;']], dtype=tf.int32), [n_data], name='start_tokens')\n\n      inference_helper = tf.contrib.seq2seq.GreedyEmbeddingHelper(dec_embeddings, start_tokens, lang_tokenizer.word_index['&lt;end&gt;'])\n      inference_decoder = tf.contrib.seq2seq.BasicDecoder(decoder_cell, inference_helper, decoder_initial_state, output_layer)\n      inference_decoder_outputs, _, _ = tf.contrib.seq2seq.dynamic_decode(inference_decoder, maximum_iterations=max_length)\n\n    return training_decoder_outputs, inference_decoder_outputs\n<\/code><\/pre>\n\n<p>Note that:<\/p>\n\n<ul>\n<li><code>target<\/code> argument is the tokenized caption. It has been preprocessed as explained in <em>Clue<\/em> section at the end of the post<\/li>\n<li><code>caption_lengths<\/code> is a tensor that contains the effective length of all captions (explained in <em>Clue<\/em> section)<\/li>\n<\/ul>\n\n<hr>\n\n<p>The loss computation is common for the two decoders outputs.<\/p>\n\n<p><strong>Loss function:<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def eval_loss(logits, captions, caption_lengths):\n  # Adjusting the size of logits...\n  paddings = [[0, 0], [0, max_length-tf.shape(logits)[1]], [0, 0]]\n  padded_logits = tf.pad(logits, paddings, 'CONSTANT', constant_values=0)\n\n  # Create the weights for sequence_loss\n  masks = tf.sequence_mask(caption_lengths, max_length, dtype=tf.float32, name='masks')\n  loss = tf.contrib.seq2seq.sequence_loss(logits=padded_logits, targets=captions, weights=masks)\n\n  return loss\n<\/code><\/pre>\n\n<hr>\n\n<h1>Prediction example:<\/h1>\n\n<p>After 265 epochs, with a test loss equals to 24.191395, here are some predictions my testing decoder makes:<\/p>\n\n<p><strong>Video overview:<\/strong><\/p>\n\n<blockquote class=\"spoiler\">\n  <p> <img src=\"https:\/\/zupimages.net\/up\/19\/33\/9etl.gif\" alt=\"video33 overview\"><\/p>\n<\/blockquote>\n\n<p><strong>Ground-truth caption:<\/strong><\/p>\n\n<pre><code>a female providing a show and tell of recent clothing purchases &lt;end&gt; &lt;pad&gt; &lt;pad&gt; &lt;pad&gt; &lt;pad&gt; &lt;pad&gt; &lt;pad&gt; &lt;pad&gt; &lt;pad&gt; &lt;pad&gt; &lt;pad&gt; &lt;pad&gt; &lt;pad&gt; &lt;pad&gt; &lt;pad&gt; &lt;pad&gt; \n<\/code><\/pre>\n\n<p><strong>Predicted caption:<\/strong><\/p>\n\n<pre><code>a women with black top is applying make up &lt;end&gt; &lt;pad&gt; &lt;end&gt; &lt;pad&gt; &lt;end&gt; &lt;pad&gt; &lt;end&gt; &lt;pad&gt; \n<\/code><\/pre>\n\n<hr>\n\n<h2>Conclusion:<\/h2>\n\n<p>Is my model wrong or <strong>do I simply do overfitting ?<\/strong> <\/p>\n\n<p>I mean, each video is composed of several hundred frames, and each frames is CNN preprocessed to output 2048 features... <\/p>\n\n<p>I've got 10.000 videos in total but only use 512 for the moment.<\/p>\n","answers":[]}
+{"QuestionId":57518822,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2019-08-16T04:12:44.420","AcceptedAnswerId":null,"OwnerUserId":9185745.0,"Title":"tf.keras.layers.CuDNNLSTM fails where tf.keras.layers.LSTM does not","Body":"<p>I am editing my neural network to run on a gpu. However, when I switched one of the LSTM layers to a CuDNNLSTM(with the exact same parameters), the model fails to compile and raises this error:<\/p>\n\n<p>Shapes must be equal rank, but are 2 and 3\n    From merging shape 0 with other shapes. for 'loss_52\/AddN' (op: 'AddN') with input shapes: [?,200], [?,1536,50].<\/p>\n\n<p>I have other CuDNN layers that compiled just fine, and the only difference between this one and those is that this one has return_sequences set to true, which the others don't, but I don't know if that has anything to do with it<\/p>\n\n<p>Here is my function, and this version compiles fine<\/p>\n\n<pre><code>def build_generator(original_input_music, input_music):\n\n        hidden_output = tf.keras.layers.CuDNNLSTM(200, activity_regularizer = tf.math.sigmoid)(input_music)\n\n        repeated_output = tf.keras.layers.RepeatVector(1536)(hidden_output)\n        end_sequence = tf.keras.layers.LSTM(50, return_sequences = True, activation = tf.math.sigmoid)(repeated_output)\n        results = tf.keras.layers.Concatenate(1)([input_music, end_sequence])\n        return tf.keras.Model(original_input_music, results), results\n<\/code><\/pre>\n\n<p>but as soon as I switch that 1 LSTM layer to CuDNN:<\/p>\n\n<pre><code>def build_generator(original_input_music, input_music):\n\n        hidden_output = tf.keras.layers.CuDNNLSTM(200, activity_regularizer = tf.math.sigmoid)(input_music)\n\n        repeated_output = tf.keras.layers.RepeatVector(1536)(hidden_output)\n\n        end_sequence = tf.compat.v1.keras.layers.CuDNNLSTM(50, return_sequences = True, activity_regularizer = tf.math.sigmoid)(repeated_output)\n        results = tf.keras.layers.Concatenate(1)([input_music, end_sequence])\n        return tf.keras.Model(original_input_music, results), results\n<\/code><\/pre>\n\n<p>the model.compile fails<\/p>\n\n<p>here is the compiling step, for reference:<\/p>\n\n<pre><code>combined.compile(optimizer = 'adam', loss='binary_crossentropy')\n<\/code><\/pre>\n\n<p>(Something to note -- the computer I'm working it on is CPU only, but I'm just trying to see if the models compile, and based on the fact that the other CuDNN layers compiled fine, I don't think that's the issue.<\/p>\n","answers":[]}
+{"QuestionId":57518866,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-08-16T04:22:23.440","AcceptedAnswerId":57519339.0,"OwnerUserId":6181162.0,"Title":"How to perform element-wise product in PyTorch?","Body":"<p>I have two torch tensors a and b. Tensor a has the shape of [batch_size, emb_size] and Tensor b has the shape of [num_of_words, emb_size]. I want to do the element-wise product on these two tensors instead of dot product.<\/p>\n\n<p>I noticed that \"*\" can perform element-wise product but it doesn't fit my case.<\/p>\n\n<p>For example, batch_size = 3, emb_size = 2, num_of_words = 5.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>a = torch.rand((3,2))\nb = torch.rand((5,2))\n<\/code><\/pre>\n\n<p>I want to get something like:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>torch.cat([a[0]*b, a[1]*b, a[2]*b]).view(3, 5, 2)\n<\/code><\/pre>\n\n<p>but I want to do this in an efficient and elegant way.<\/p>\n","answers":[{"AnswerId":"57519339","CreationDate":"2019-08-16T05:35:17.533","ParentId":null,"OwnerUserId":"2790047","Title":null,"Body":"<p>You can use<\/p>\n\n<pre><code>a.unsqueeze(1) * b\n<\/code><\/pre>\n\n<p>PyTorch supports <a href=\"https:\/\/pytorch.org\/docs\/stable\/notes\/broadcasting.html\" rel=\"nofollow noreferrer\">broadcasting semantics<\/a> but you need to make sure the singleton dimensions are in the correct locations.<\/p>\n"}]}
+{"QuestionId":57518882,"AnswerCount":1,"Tags":"<tensorflow><keras><python-3.6><tensorflow-datasets>","CreationDate":"2019-08-16T04:25:34.920","AcceptedAnswerId":57520090.0,"OwnerUserId":5468768.0,"Title":"Tensorflow Fit exits with code 1 without any error message","Body":"<p>I'm new to tensorflow. What I'm trying to do is to train a simple neural network to solve the Newton 2 problems, to guess the force value of given mass and acceleration values. The input layer consists of two neurons which are mass and acceleration values. The output layer is the force.<\/p>\n\n<p>The program just gives a warning, prints some data which I guess the outputs and then exits with code 1. I cannot try anything to solve this problem. Because as I said before I'm new to tensorflow and there is no error message. <\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><code>   from tensorflow.keras.layers import Dense, Flatten\n   from tensorflow.keras.models import Sequential\n   import numpy as np\n   import pickle\n\n\n   X = pickle.load(open(\"Newton2_X.pickle\", \"rb\"))\n   y = pickle.load(open(\"Newton2_y.pickle\", \"rb\"))\n\n   model = Sequential()\n   # model.add(Flatten())\n   model.add(Dense(2, activation=\"relu\"))\n   model.add(Dense(128, activation=\"relu\"))\n   model.add(Dense(1, activation=\"softmax\"))\n\n   model.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\n\n   model.fit(X, y, epochs=3, validation_split=0.1, batch_size=100)\n<\/code><\/pre>\n\n<p>Here are the pickle files:\n<a href=\"https:\/\/drive.google.com\/drive\/folders\/1FkKmY4px8oQJkbHYb_Z4y4Lnb1EazkvP?usp=sharing\" rel=\"nofollow noreferrer\">https:\/\/drive.google.com\/drive\/folders\/1FkKmY4px8oQJkbHYb_Z4y4Lnb1EazkvP?usp=sharing<\/a><\/p>\n\n<p>After this part of the code I've some additional lines to make the network to guess a new value and some print lines. These lines are not executed. In fact, I've found that the 'problem' must be in model.fit(...) part. Because no lines after that line are executed.<\/p>\n\n<p>Here is the full warning msg that I got from the program:<\/p>\n\n<pre><code>WARNING: Logging before flag parsing goes to stderr.\nW0816 07:02:05.292823 17652 deprecation.py:506] From C:\\Users\\SABA\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow\\python\\ops\\init_ops.py:1251: calling VarianceScaling.__init__ (from tensorflow.python.ops.init_ops) with dtype is deprecated and will be removed in a future version.\nInstructions for updating:\nCall initializer instance with the dtype argument instead of passing it to the constructor\n6, 0.2142802901764338, 0.26114980919201514, 0.2451221454091551, 0.19920049739052853, ...\n\n<\/code><\/pre>\n","answers":[{"AnswerId":"57520090","CreationDate":"2019-08-16T06:50:29.907","ParentId":null,"OwnerUserId":"2455494","Title":null,"Body":"<p>A couple of things to tweak.<\/p>\n\n<p>Firstly, I don't think the data is the shape that you think it is. You have:<\/p>\n\n<pre><code>X.shape # (45000, 2, 2, 1)\n<\/code><\/pre>\n\n<p><code>y<\/code> is a flat list with 90,000 elements.<\/p>\n\n<p>Secondly, you are predicting a number (so a regression) but you were trying to use 'sparse_categorical_crossentropy' as a loss function which is for classification problems.<\/p>\n\n<p>I can get your code to run by simply slicing the data down to the shape we need but obviously it won't train as I haven't paired up the correct Xs and ys. You'll need to sort this out properly in the data<\/p>\n\n<pre>\nfrom tensorflow.keras.layers import Dense, Flatten\nfrom tensorflow.keras.models import Sequential\nimport numpy as np\nimport pickle\n\n### <b>TODO - sort this out!<\/b>\nX = pickle.load(open(\"Newton2_X.pickle\", \"rb\"))<b>[:,0,:,0]<\/b>\ny = np.array(pickle.load(open(\"Newton2_y.pickle\", \"rb\")))<b>[:45000]<\/b>\n####\n\nmodel = Sequential()\n# model.add(Flatten())\nmodel.add(Dense(2, activation=\"relu\"))\nmodel.add(Dense(128, activation=\"relu\"))\nmodel.add(Dense(1, activation=\"softmax\"))\n\nmodel.compile(optimizer='adam',\n          <b>loss='mse'<\/b>)\n\nmodel.fit(X, y, epochs=3, validation_split=0.1, batch_size=100)\n<\/pre>\n"}]}
+{"QuestionId":57518892,"AnswerCount":1,"Tags":"<keras><google-colaboratory><darknet>","CreationDate":"2019-08-16T04:27:50.573","AcceptedAnswerId":null,"OwnerUserId":6458354.0,"Title":"My Laptop is shut-down because of insufficient memory while Google colaboratory is running","Body":"<p>I ran the darknet to execute YOLO object detection machine learning in Google colaboratory.\nI am observing the process of machine learning on chrome of my Mac laptop.\nHowever, the memory usage of my laptop(Not Google colaboratory's memory!) became too huge to open the google colaboratory after a while. \nMy laptop memory is 16GB and it is not enough to open it.<\/p>\n\n<p>I don't understand why my laptop memory is used even if the machine learning is executed in cloud server.<\/p>\n\n<p>How can I resolve this issue ?<\/p>\n","answers":[{"AnswerId":"57526815","CreationDate":"2019-08-16T14:55:38.540","ParentId":null,"OwnerUserId":"8841057","Title":null,"Body":"<p>As suggested by the error message, the most common cause of this issue is that third party cookies are disabled in your browser.<\/p>\n\n<p>See here for a description of why Colab needs third party cookies to be enabled:\n<a href=\"https:\/\/research.google.com\/colaboratory\/faq.html\" rel=\"nofollow noreferrer\">https:\/\/research.google.com\/colaboratory\/faq.html<\/a><\/p>\n\n<p>Reproducing the relevant snippet below:<\/p>\n\n<blockquote>\n  <p>Why prompt to enable third-party cookies? <\/p>\n  \n  <p>Colaboratory uses HTML\n  iframes and service workers hosted on separate origins in order to\n  display rich outputs securely. Browsers require enabling third-party\n  cookies to use the service workers within iframes. An alternative to\n  enabling third-party cookies for all sites is to whitelist the\n  following hostname in your browser settings: googleusercontent.com.<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":57519131,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2019-08-16T05:08:25.510","AcceptedAnswerId":null,"OwnerUserId":6667435.0,"Title":"How to load history from saved model(.h5) file?","Body":"<p>I have trained my model and saved it. For plotting results against that model I loaded the model and now I want to extract history from that model. <\/p>\n\n<p>saved model<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = load_model('model_pre_ep3_valloss0.360.h5')\n<\/code><\/pre>\n\n<p>How do I extract accuracy, loss etc plot from there?<\/p>\n","answers":[{"AnswerId":"57520085","CreationDate":"2019-08-16T06:50:20.380","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>That information is not saved a Keras HDF5 model file, so if you did not save it in another file, it is effectively lost.<\/p>\n\n<p>One simple way to save the loss\/accuracy history is to use the <code>CSVLogger<\/code> callback, more information in the keras <a href=\"https:\/\/keras.io\/callbacks\/#csvlogger\" rel=\"nofollow noreferrer\">docs<\/a>.<\/p>\n"}]}
+{"QuestionId":57519241,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2019-08-16T05:23:07.383","AcceptedAnswerId":57519833.0,"OwnerUserId":11309950.0,"Title":"how can I get to decrease loss when the epoch increases?","Body":"<p>I have a problem that the loss does not decrease when the epoch increases.<\/p>\n\n<p>here is my code<\/p>\n\n<pre><code>model = Sequential()\n\nmodel.add(LSTM(50, input_shape=(None, 1)))\nmodel.add(Dropout(0.05))\nBatchNormalization()\n\n\nmodel.add(Dense(1))\n\nmodel.compile(loss='mean_squared_error', optimizer=optimizers.Adam(lr=0.01, decay=0.001))\n\nhist = model.fit(x_train, y_train, epochs=40, batch_size=32, verbose=2)\n<\/code><\/pre>\n\n<p><code>Total parameters<\/code> is number of 10451 and<\/p>\n\n<p><code>train dataset<\/code> is number of 2285<\/p>\n\n<p>I am wondering if the total parameter is a reasonable ratio for train_data.\nIn other words, I wonder if the total parameter is appropriate to have a ratio of train_data.<\/p>\n\n<p>and here is my loss graph\n<img src=\"https:\/\/imgur.com\/FZYN9gb.png\" alt=\"1\"><\/p>\n\n<p>I tried parameter and hyperparameter tuning But this could not be solved.<\/p>\n\n<p>The dataset has been preprocessed between 0 and 1.<\/p>\n\n<p>The ensemble rather made the result worse.<\/p>\n\n<p>how can I get to decrease loss when the epoch increases?<\/p>\n","answers":[{"AnswerId":"57519833","CreationDate":"2019-08-16T06:28:29.947","ParentId":null,"OwnerUserId":"2455494","Title":null,"Body":"<p>There really is no simple answer to your question. You have a loss graph that shows a fast initial learning that then tails off to a much slower reduction after a couple of epochs. This is quite a common phenomenon.<\/p>\n\n<p>Your question amounts to \"how do I make a better machine learning model for this dataset?\". Which is an impossible question to answer generically.<\/p>\n\n<p>Directly you <em>could<\/em> increase layers, weights, etc. Eventually (at least in theory) you would have enough complexity in your model to memorise your entire training set and get the loss down close to 0. The resulting model would almost certainly be overfitted and perform very poorly when it came to data it had never seen before though.<\/p>\n\n<p>Objectively, if your labels are spread uniformly between 0 and 1, you already seem to have a very low loss value (less than 0.0005 MSE, right??)? <\/p>\n\n<p>Have you tried this on your test set? It's not at all clear why you need to drive this further down.<\/p>\n"}]}
+{"QuestionId":57519500,"AnswerCount":1,"Tags":"<pandas><tensorflow><machine-learning>","CreationDate":"2019-08-16T05:56:09.997","AcceptedAnswerId":57531606.0,"OwnerUserId":1922184.0,"Title":"Tensorflow 2 \"Attempt to convert a value (63) with an unsupported type (<class 'numpy.int64'>) to a Tensor\"","Body":"<p>Taking the necessary code from the Tensorflow example for classifying structured data <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/keras\/feature_columns\" rel=\"nofollow noreferrer\">here<\/a> so that I can learn to train on Numeric columns; I get the following error: <\/p>\n\n<blockquote>\n  <p>ValueError: Attempt to convert a value (63) with an unsupported type\n  () to a Tensor.<\/p>\n<\/blockquote>\n\n<p>And while I suppose I could try to convert particular values in the dataframe to work with Tensors (if that would even work), there has to be something else going on as the code works in Colab but throws an error in PyCharm.<\/p>\n\n<pre><code>import pandas as pd\nimport tensorflow as tf\nfrom tensorflow import feature_column\nfrom tensorflow.python.keras import layers\nfrom sklearn.model_selection import train_test_split\n\nURL = 'https:\/\/storage.googleapis.com\/applied-dl\/heart.csv'\ndataframe = pd.read_csv(URL)\ndataframe.head()\n\ntrain, test = train_test_split(dataframe, test_size=0.2)\ntrain, val = train_test_split(train, test_size=0.2)\nprint(len(train), 'train examples')\nprint(len(val), 'validation examples')\nprint(len(test), 'test examples')\n\n# A utility method to create a tf.data dataset from a Pandas Dataframe\ndef df_to_dataset(dataframe, shuffle=True, batch_size=32):\n  dataframe = dataframe.copy()\n  labels = dataframe.pop('target')\n  ds = tf.data.Dataset.from_tensor_slices((dict(dataframe), labels))\n  if shuffle:\n    ds = ds.shuffle(buffer_size=len(dataframe))\n  ds = ds.batch(batch_size)\n  return ds\n\nfeature_columns = []\n\n# numeric cols\nfor header in ['age', 'trestbps', 'chol', 'thalach', 'oldpeak', 'slope', 'ca']:\n  feature_columns.append(feature_column.numeric_column(header))\n\nfeature_layer = tf.keras.layers.DenseFeatures(feature_columns)\n\nbatch_size = 32\ntrain_ds = df_to_dataset(train, batch_size=batch_size)\nval_ds = df_to_dataset(val, shuffle=False, batch_size=batch_size)\ntest_ds = df_to_dataset(test, shuffle=False, batch_size=batch_size)\n\n\"\"\"## Create, compile, and train the model\"\"\"\n\nmodel = tf.keras.Sequential([\n  feature_layer,\n  layers.Dense(128, activation='relu'),\n  layers.Dense(128, activation='relu'),\n  layers.Dense(1, activation='sigmoid')\n])\n\nmodel.compile(optimizer='adam',\n              loss='binary_crossentropy',\n              metrics=['accuracy'],\n              run_eagerly=True)\n\nmodel.fit(train_ds,\n          validation_data=val_ds,\n          epochs=5)\n\nloss, accuracy = model.evaluate(test_ds)\nprint(\"Accuracy\", accuracy)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57531606","CreationDate":"2019-08-16T21:59:59.363","ParentId":null,"OwnerUserId":"1922184","Title":null,"Body":"<p>It seems that Numpy was corrupted. After uninstalling Numpy and re-installing it, the program worked.<\/p>\n"}]}
+{"QuestionId":57519937,"AnswerCount":2,"Tags":"<python><tensorflow><keras><deep-learning><tf.keras>","CreationDate":"2019-08-16T06:38:14.390","AcceptedAnswerId":null,"OwnerUserId":1830648.0,"Title":"How to use model.predict for predicting single image in tf.keras","Body":"<p>I am trying to predict a single image using the trained model either I am getting an array of 80 values or I am getting this error:<\/p>\n\n<blockquote>\n  <p>dimension Matrix size-incompatible: In[0]: [1,19200], In[1]: [240,128] <\/p>\n<\/blockquote>\n\n<p>I have tried different option available on StackOverflow but nothing worked out for me. Few of those I have mentioned here\n<a href=\"https:\/\/stackoverflow.com\/questions\/43469281\/how-to-predict-input-image-using-trained-model-in-keras\">How to predict input image using trained model in Keras?<\/a>\n<a href=\"https:\/\/stackoverflow.com\/questions\/53941590\/cannot-predict-the-label-for-a-single-image-with-vgg19-in-keras\">Cannot predict the label for a single image with VGG19 in Keras<\/a><\/p>\n\n<p>to load data I have used code from <a href=\"https:\/\/pythonprogramming.net\/loading-custom-data-deep-learning-python-tensorflow-keras\/\" rel=\"nofollow noreferrer\">https:\/\/pythonprogramming.net\/loading-custom-data-deep-learning-python-tensorflow-keras\/<\/a><\/p>\n\n<pre><code>model = tf.keras.models.Sequential([tf.keras.layers.Flatten(),\n                                    tf.keras.layers.Dense(128,activation=tf.nn.relu),\n                                    tf.keras.layers.Dropout(rate=0.5),\n                                    tf.keras.layers.BatchNormalization(axis=1),\n                                    tf.keras.layers.Dense(100,activation=tf.nn.relu),\n                                    tf.keras.layers.Dropout(rate=0.5),\n                                    tf.keras.layers.BatchNormalization(axis=1),\n                                    tf.keras.layers.Dense(100,activation=tf.nn.relu),\n                                    tf.keras.layers.Dropout(rate=0.4),\n                                    tf.keras.layers.BatchNormalization(axis=1),\n                                    tf.keras.layers.Dense(100,activation=tf.nn.relu),\n                                    tf.keras.layers.Dropout(rate=0.2),\n                                    tf.keras.layers.BatchNormalization(axis=1),\n                                    tf.keras.layers.Dense(100,activation=tf.nn.relu),\n                                    tf.keras.layers.Dropout(rate=0.7),\n                                    tf.keras.layers.BatchNormalization(axis=1),\n                                    tf.keras.layers.Dense(2,activation=tf.nn.sigmoid)])\n\nfrom tensorflow.keras.optimizers import SGD\nopt = SGD(lr=0.01)\nmodel.compile(loss = \"mean_squared_error\", optimizer = opt, metrics=['accuracy'])\n\nmodel.fit(X_train, y_train, batch_size=50,epochs=205,callbacks=[callbacks])\nmodel.evaluate(X_test, y_test)\n\n#First way no error multiple values\nfrom keras.preprocessing import image\ntest_image = image.load_img('img', target_size=(80, 80))\ntest_image = image.img_to_array(test_image)\ntest_image = np.expand_dims(test_image, axis=0)\ntest_image = test_image.reshape(80, 80*3)   \n\nresult = model.predict(test_image\/255.0, batch_size=1)\nprint(result)\n\nresult = model.predict_classes(test_image\/255.0, batch_size=1)\nprint(result)\n\n#second way, Matrix size-incompatible(Error)\nimg = image.load_img('img', target_size=(80, 80))\nx = image.img_to_array(img)\nx = np.expand_dims(x, axis=0)\nimages = np.vstack([x])\nclasses = model.predict_classes(x, batch_size=1)\nprint(classes)\n\n#third way,Matrix size-incompatible(Error)\nprint(model.predict_classes(np.expand_dims(X_test[10], axis=0)))\n<\/code><\/pre>\n\n<h1>first way o\/p<\/h1>\n\n<pre><code>[[0.4419657  0.5002713 ]\n [0.4379595  0.5007576 ]\n [0.43674818 0.5014705 ]\n [0.43429232 0.50493073]\n [0.43446562 0.50841236]\n [0.43417045 0.51054156]\n [0.4348069  0.51089376]\n [0.43577492 0.50624526]\n [0.43650073 0.5084632 ]\n [0.44226125 0.5106121 ]\n [0.44256458 0.51815474]\n [0.4366225  0.5247917 ]\n [0.44668812 0.5251329 ]\n [0.45352334 0.5256567 ]\n [0.4572222  0.5226744 ]\n [0.46253017 0.519516  ]\n [0.46236354 0.51812094]\n [0.4637973  0.5135511 ]\n [0.46357435 0.5091353 ]\n [0.4647084  0.50632596]\n [0.46603358 0.5004298 ]\n [0.46488768 0.49856278]\n [0.4637522  0.50335187]\n [0.4605053  0.5001269 ]\n [0.46522006 0.49863124]\n [0.46316907 0.50639075]\n [0.46407732 0.51068664]\n [0.452004   0.51411426]\n [0.4437306  0.5115358 ]\n [0.44812864 0.5081628 ]\n [0.45141432 0.50651264]\n [0.4518429  0.5081477 ]\n [0.44927847 0.49777785]\n [0.44322333 0.4825523 ]\n [0.44135702 0.47820964]\n [0.43782592 0.47925416]\n [0.43334886 0.47967055]\n [0.4303841  0.47919393]\n [0.42532465 0.48017433]\n [0.42595625 0.47586957]\n [0.4292146  0.47039127]\n [0.43103853 0.4656783 ]\n [0.43306574 0.463838  ]\n [0.4276282  0.4699353 ]\n [0.42867652 0.46581164]\n [0.43545863 0.45945364]\n [0.44277322 0.47201872]\n [0.4460439  0.4735631 ]\n [0.443609   0.47811195]\n [0.44498175 0.47373036]\n [0.44886908 0.48278013]\n [0.4429854  0.4908823 ]\n [0.44526115 0.49165127]\n [0.45446166 0.49740997]\n [0.4574405  0.49744406]\n [0.45719808 0.5041652 ]\n [0.4546386  0.50350964]\n [0.4539847  0.5072408 ]\n [0.46614394 0.5016114 ]\n [0.45871773 0.5072619 ]\n [0.4616405  0.50306535]\n [0.46992242 0.50247884]\n [0.46977502 0.50086266]\n [0.4666891  0.48775986]\n [0.46482667 0.48338565]\n [0.45363256 0.49476466]\n [0.45803532 0.49177122]\n [0.4653356  0.49367705]\n [0.46423748 0.49807605]\n [0.47541898 0.49923015]\n [0.45872727 0.5036651 ]\n [0.45934066 0.499598  ]\n [0.46240935 0.50199217]\n [0.45569527 0.5061147 ]\n [0.4612086  0.49505413]\n [0.46061015 0.49445656]\n [0.46367538 0.4839395 ]\n [0.46831584 0.4812285 ]\n [0.46474478 0.4740279 ]\n [0.46901295 0.4787915 ]]\n\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 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 1\n 1 1 1 1 1 1]\n<\/code><\/pre>\n\n<h1>2nd &amp; 3rd-way error<\/h1>\n\n<pre><code>InvalidArgumentError:  Matrix size-incompatible: In[0]: [1,6400], In[1]: [240,128]\n     [[node sequential\/dense\/MatMul (defined at &lt;ipython-input-15-dfba430dbf46&gt;:1) ]] [Op:__inference_keras_scratch_graph_1056]\n<\/code><\/pre>\n","answers":[{"AnswerId":"57576580","CreationDate":"2019-08-20T15:05:54.550","ParentId":null,"OwnerUserId":"10754197","Title":null,"Body":"<p>From the code i can see that you are trying to predict 2 class so try to change the loss function and then run it.<\/p>\n\n<pre><code>model_final.compile(loss = keras.losses.categorical_crossentropy, optimizer = opt, metrics=[\"accuracy\"])\n<\/code><\/pre>\n\n<p>Share the output after doing this change.<\/p>\n"},{"AnswerId":"57533682","CreationDate":"2019-08-17T05:54:20.863","ParentId":null,"OwnerUserId":"1830648","Title":null,"Body":"<p>It`s working for me with below code changes<\/p>\n\n<pre><code>IMG_SIZE = 80\nimg_array = cv2.imread(\"im.jpg\" ,cv2.IMREAD_GRAYSCALE)  \nnew_array = cv2.resize(img_array, (IMG_SIZE, IMG_SIZE))\nplt.imshow(new_array\/255.0, cmap='gray')\nplt.show()\n\nnew_array=np.reshape(new_array,(80, 80, 1))\nnew_array=new_array.flatten()\/255.0\nnew_array_2=np.reshape(new_array,(6400, 1))\n\n#to predict single image\nclass_prob=model.predict(new_array_2.T,batch_size=1)\nprint(class_prob)\nclassifications=model.predict_classes(new_array_2.T,batch_size=1)\nprint(classifications) \n<\/code><\/pre>\n"}]}
+{"QuestionId":57520038,"AnswerCount":0,"Tags":"<python><keras><conv-neural-network>","CreationDate":"2019-08-16T06:45:58.593","AcceptedAnswerId":null,"OwnerUserId":11925809.0,"Title":"Performing 2D convolution on 4D data in Keras","Body":"<p>I have data of dimension <code>24*64*64*10<\/code> (excluding the batch size).<\/p>\n\n<p>I want to split the input into 24 inputs of dimension <code>64*64*10<\/code>, perform <code>Conv2D<\/code> on each of them and then concatenate them to get the 4D data again for further processing.<\/p>\n\n<p>Any help regarding the implementation would be helpful. I am working with Keras.<\/p>\n\n<p>Edit: I tried to the following code to perform the 2D convolution<\/p>\n\n<pre><code>num_ch= 24\ninput= Input(shape=(64,64,10,num_ch))\nprint(input.shape)\nbranch_out= []\nfor i in range(num_ch):\n    out= Lambda(lambda x: x[:,:,:,:,i] )(input)\n    print(out.shape)\n    out= Conv2D(10, kernel_size=(3,3),strides= (1,1), padding='same', data_format= 'channels_last')(input)\n    branch_out.append(out)\n<\/code><\/pre>\n\n<p>I got the following error:<\/p>\n\n<pre><code>(?, 64, 64, 10, 24)\n(?, 64, 64, 10)\n---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-83-51977f4edbba&gt; in &lt;module&gt;\n      7     out= Lambda(lambda x: x[:,:,:,:,i] )(input)\n      8     print(out.shape)\n----&gt; 9     out= Conv2D(10, kernel_size=(3,3),strides= (1,1), padding='same', data_format= 'channels_last')(input)\n     10     branch_out.append(out)\n\n~\/anaconda3\/lib\/python3.7\/site-packages\/keras\/engine\/base_layer.py in __call__(self, inputs, **kwargs)\n    412                 # Raise exceptions in case the input is not compatible\n    413                 # with the input_spec specified in the layer constructor.\n--&gt; 414                 self.assert_input_compatibility(inputs)\n    415 \n    416                 # Collect input shapes to build layer.\n\n~\/anaconda3\/lib\/python3.7\/site-packages\/keras\/engine\/base_layer.py in assert_input_compatibility(self, inputs)\n    309                                      self.name + ': expected ndim=' +\n    310                                      str(spec.ndim) + ', found ndim=' +\n--&gt; 311                                      str(K.ndim(x)))\n    312             if spec.max_ndim is not None:\n    313                 ndim = K.ndim(x)\n\nValueError: Input 0 is incompatible with layer conv2d_25: expected ndim=4, found ndim=5\n<\/code><\/pre>\n\n<p>I don't understand why it is still showing the dimentionalilty as 5.<\/p>\n","answers":[]}
+{"QuestionId":57520198,"AnswerCount":0,"Tags":"<python><tensorflow><recurrent-neural-network>","CreationDate":"2019-08-16T06:59:42.317","AcceptedAnswerId":null,"OwnerUserId":3446051.0,"Title":"tf.contrib.rnn.static_rnn returns different values","Body":"<p>I am using <code>tf.contrib.rnn.static_rnn<\/code> in two different ways, which should return the same value, but they aren't which makes me crazy. I have spent lot's of time figuring out what the problem is, but could not. \nHere are the two different ways:  <\/p>\n\n<p><strong>Option I:<\/strong><\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\ntf.reset_default_graph()\ntf.set_random_seed(42)\nnp.random.seed(42)\nn_steps = 2\nn_inputs = 3\nn_neurons = 5\nX = tf.placeholder(tf.float32, [None, n_steps, n_inputs])\nX_trans = tf.transpose(X, perm=[1, 0, 2])\nX_seqs = tf.unstack(X_trans)\nbasic_cell = tf.contrib.rnn.BasicRNNCell(num_units=n_neurons)\noutput_seqs, states = tf.contrib.rnn.static_rnn(basic_cell, X_seqs, dtype=tf.float32)\ninit = tf.global_variables_initializer()\n\nX_batch = np.array([\n# t = 0 t = 1\n[[0, 1, 2], [9, 8, 7]], # instance 0\n[[3, 4, 5], [0, 0, 0]], # instance 1\n[[6, 7, 8], [6, 5, 4]], # instance 2\n[[9, 0, 1], [3, 2, 1]], # instance 3\n])\nwith tf.Session() as sess:\n    init.run()\n    print(sess.run([X_seqs,output_seqs],feed_dict={X: X_batch}))\n<\/code><\/pre>\n\n<p><strong>Option II:<\/strong>  <\/p>\n\n<pre><code>tf.reset_default_graph()\ntf.set_random_seed(42)\nnp.random.seed(42)\n\n\nX_0 = tf.placeholder(tf.float32,[None,n_inputs])\nX_1 = tf.placeholder(tf.float32,[None,n_inputs])\nbasic_cell = tf.contrib.rnn.BasicRNNCell(num_units=n_neurons)\noutput_seqs, states = tf.contrib.rnn.static_rnn(basic_cell, [X_0,X_1], dtype=tf.float32)\ninit = tf.global_variables_initializer()\nseq = [np.array([[0., 1., 2.],\n       [3., 4., 5.],\n       [6., 7., 8.],\n       [9., 0., 1.]]), np.array([[9., 8., 7.],\n       [0., 0., 0.],\n       [6., 5., 4.],\n       [3., 2., 1.]])]\nwith tf.Session() as sess:\n    init.run()\n    print(sess.run(output_seqs, feed_dict={X_0 : seq[0], X_1: seq[1]}))\n<\/code><\/pre>\n\n<p>In both of them <code>tf.contrib.rnn.static_rnn(basic_cell, X_seqs, dtype=tf.float32)<\/code> receives the same sequence of arrays, but the result <code>output_seqs<\/code> is different:<br>\n<strong>Option I:<\/strong><\/p>\n\n<pre><code>[array([[-0.45652324, -0.68064123,  0.40938237,  0.63104504, -0.45732826],\n       [-0.80015343, -0.9921827 ,  0.78177965,  0.9971032 , -0.9964609 ],\n       [-0.93605185, -0.9998379 ,  0.9308867 ,  0.9999814 , -0.9999831 ],\n       [ 0.9927367 , -0.9981933 , -0.55543643,  0.9989031 , -0.9953323 ]],\n      dtype=float32), array([[-0.94288   , -0.9998869 ,  0.9405581 ,  0.9999985 , -0.99999976],\n       [-0.637116  ,  0.11300926,  0.5798437 ,  0.4310559 , -0.6371699 ],\n       [-0.9165386 , -0.9945605 ,  0.896054  ,  0.99987185, -0.999975  ],\n       [-0.02746338, -0.73191977,  0.7827872 ,  0.9525681 , -0.978177  ]],\n      dtype=float32)]]\n<\/code><\/pre>\n\n<p><strong>Option II:<\/strong><\/p>\n\n<pre><code>[array([[ 0.30741334, -0.32884315, -0.65428466, -0.9385059 ,  0.52089024],\n       [ 0.99122757, -0.95425403, -0.75180787, -0.9995208 ,  0.9820235 ],\n       [ 0.9999268 , -0.99783254, -0.8247353 , -0.9999963 ,  0.9994779 ],\n       [ 0.996771  , -0.68750614,  0.8419969 ,  0.93039113,  0.81206834]],\n      dtype=float32), array([[ 0.999989  , -0.99976057, -0.06679279, -0.9999803 ,  0.99982214],\n       [-0.6524943 , -0.51520866, -0.37968948, -0.59225935, -0.08968402],\n       [ 0.99862397, -0.99715203, -0.03308637, -0.9991566 ,  0.9932902 ],\n       [ 0.99681675, -0.9598194 ,  0.3966063 , -0.83076054,  0.7967197 ]],\n      dtype=float32)]\n<\/code><\/pre>\n\n<p>I could not figure out out why the output is not equal. Could you please hep me? What do I oversee?<\/p>\n","answers":[]}
+{"QuestionId":57520639,"AnswerCount":0,"Tags":"<tensorflow><web><mobile><3d><object-detection-api>","CreationDate":"2019-08-16T07:37:25.300","AcceptedAnswerId":null,"OwnerUserId":5358535.0,"Title":"How to do 3D Object Detection in Tensorflow?","Body":"<p>I want to do object detection using tensorflow, with z-value and rotation.<\/p>\n\n<p>Tried with PR Net but not getting decent FPS of detection<\/p>\n\n<p>Want to use the same in both web and mobile versions for detection. Expecting around 10-15 FPS on object detection with z-value and rotation for each frame.<\/p>\n","answers":[]}
+{"QuestionId":57520942,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-08-16T08:01:30.787","AcceptedAnswerId":null,"OwnerUserId":4806177.0,"Title":"Tensorflow Tf.tf.squared_difference is showing a value error","Body":"<p>I'm using Keras Input layer, then I add an embedding layer. Whenever I try to execute the following code<\/p>\n\n<pre><code>vector = tf.reduce_sum(tf.squared_difference(user, book), 1, name=\"pos_distances\")\n<\/code><\/pre>\n\n<blockquote>\n  <p>ValueError: Input 0 is incompatible with layer prediction: expected\n  min_ndim=2, found ndim=1<\/p>\n<\/blockquote>\n\n<p>Here is the shape of the input tensors<\/p>\n\n<pre><code>user :   Tensor(\"Shape:0\", shape=(2,), dtype=int32) \n\nbook :   Tensor(\"Shape_1:0\", shape=(2,), dtype=int32) \n<\/code><\/pre>\n\n<p>It is giving me this error<\/p>\n\n<pre><code>vector = tf.reduce_sum(tf.squared_difference(user, book), 1, name=\"pos_distances\")\n<\/code><\/pre>\n","answers":[{"AnswerId":"57522879","CreationDate":"2019-08-16T10:16:15.870","ParentId":null,"OwnerUserId":"5154274","Title":null,"Body":"<p>Your tensors don't have the <code>axis=1<\/code>, over which you are trying to perform the reduction ( second argument of the <code>tf.reduce_sum()<\/code>). It should rather be<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>sqrt_diff = tf.squared_difference(user, book)\nvector = tf.reduce_sum(sqrt_diff, 0, name=\"pos_distances\")\n<\/code><\/pre>\n"}]}
+{"QuestionId":57521061,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-08-16T08:10:21.197","AcceptedAnswerId":null,"OwnerUserId":11935030.0,"Title":"Could not find 'cudart64_100.dll' while trying to install tensorflow","Body":"<p>I am trying to install tensorflow.<\/p>\n\n<p>I have installed Cuda10.0 and cudnn, but I still get this error:<\/p>\n\n<pre><code>    Traceback (most recent call last):\n      File \"D:\\python\\lib\\site-packages\\tensorflow\\python\\platform\\self_check.py\", line 75, in preload_check\n        ctypes.WinDLL(build_info.cudart_dll_name)\n      File \"D:\\python\\lib\\ctypes\\__init__.py\", line 364, in __init__\n        self._handle = _dlopen(self._name, mode)\n    OSError: [WinError 126] \u627e\u4e0d\u5230\u6307\u5b9a\u7684\u6a21\u5757\u3002\n\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 \"D:\/Python_test\/Python\u7b97\u6cd5\u6210\u54c1\/\u673a\u5668\u5b66\u4e60\/test\/test.py\", line 4, in &lt;module&gt;\n        import tensorflow as tf\n      File \"D:\\python\\lib\\site-packages\\tensorflow\\__init__.py\", line 40, in &lt;module&gt;\n        from tensorflow.python.tools import module_util as _module_util\n      File \"D:\\python\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 49, in &lt;module&gt;\n        from tensorflow.python import pywrap_tensorflow\n      File \"D:\\python\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 30, in &lt;module&gt;\n        self_check.preload_check()\n      File \"D:\\python\\lib\\site-packages\\tensorflow\\python\\platform\\self_check.py\", line 82, in preload_check\n        % (build_info.cudart_dll_name, build_info.cuda_version_number))\n    ImportError: Could not find 'cudart64_100.dll'. TensorFlow requires that this DLL be installed in a directory that is named in your %PATH% environment variable. Download and install CUDA 10.0 from this URL: `enter code here`https:\/\/developer.nvidia.com\/cuda-90-download-archive\n<\/code><\/pre>\n\n<p>What am I missing or doing wrong?<\/p>\n","answers":[]}
+{"QuestionId":57521417,"AnswerCount":0,"Tags":"<keras><neural-network><deep-learning><face-recognition><finetunning>","CreationDate":"2019-08-16T08:38:56.123","AcceptedAnswerId":null,"OwnerUserId":10331686.0,"Title":"Fine-tuning with MobileNet","Body":"<p>I'm trying to build a neural net with Keras for face recognition, I want to use MobileNet with pre-trained weight on VGGFace dataset, but Keras just support weight for VGGNet16, ResNet50, SENet50. Anyone know where I can download the pre-trained weight for MobileNet? Sorry I'm not good at English.<\/p>\n","answers":[]}
+{"QuestionId":57521485,"AnswerCount":1,"Tags":"<python><keras><time-series><lstm><multilabel-classification>","CreationDate":"2019-08-16T08:43:13.780","AcceptedAnswerId":null,"OwnerUserId":2754279.0,"Title":"LSTM with multiple time-series","Body":"<p>I am training LSTM for multiple time-series in an array which has a structure: 450x801. There are 450 time series with each of 801 timesteps \/ time series.\nThe labels are classes with assigned integer from 1 to 6, so the dimension of the label is 450x1. This is my implmentation:<\/p>\n\n<p>This is my code:<\/p>\n\n<pre><code>def readData():\nlabels = pd.read_csv('label.csv', header = None)\nlabels = labels.values\ndata = pd.read_csv('data.csv', header = None)\nreturn data, labels\n\ndata, labels = readData()\n\ndata_train, data_test, labels_train, labels_test = train_test_split(data, labels)\n\nmodel = Sequential()\nmodel.add(LSTM(units=32, input_shape = (450,801,1)))\nmodel.add(Dense(6, activation='softmax'))\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>However, I get following errors:<\/p>\n\n<p><b>Input 0 is incompatible with layer lstm_3: expected ndim=3, found ndim=4<\/b><\/p>\n\n<p>Any idea how do I solve it?<\/p>\n","answers":[{"AnswerId":"57523548","CreationDate":"2019-08-16T11:07:07.067","ParentId":null,"OwnerUserId":"2754279","Title":null,"Body":"<p>It's working now. The final code:<\/p>\n\n<pre><code>def readData():\nlabels = pd.read_csv('label.csv', header = None)\nlabels = labels.values\nlabels = to_categorical(labels)\ndata = pd.read_csv('data.csv', header = None)\nreturn data, labels\n\ndata, labels = readData()\ndata = np.expand_dims(data, axis=-1)\n\ndata_train, data_test, labels_train, labels_test = train_test_split(data, labels)\n\nmodel = Sequential()\nmodel.add(LSTM(units=32, input_shape = (801,1)))\nmodel.add(Dense(6, activation='softmax'))\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57521577,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-16T08:50:24.187","AcceptedAnswerId":null,"OwnerUserId":11932512.0,"Title":"Prediction using a custom made neural network - Tensorflow low level API","Body":"<p>I am considering the code below, taken from the repository <\/p>\n\n<p><a href=\"https:\/\/github.com\/frankhan91\/DeepBSDE\" rel=\"nofollow noreferrer\">https:\/\/github.com\/frankhan91\/DeepBSDE<\/a><\/p>\n\n<p>(where I introduced some comments because I am totally new to Tensorflow)<\/p>\n\n<p>This code solves a minimization problem that involves a multitude of neural networks over a time discretization. This family of neural networks drive a process Y<\/p>\n\n<pre><code>\nimport os\n\nos.environ['KMP_DUPLICATE_LIB_OK']='True'\nimport logging\nimport time\nimport numpy as np\nimport tensorflow as tf\nfrom tensorflow.python.training import moving_averages\n\nTF_DTYPE = tf.float64\nMOMENTUM = 0.99\nEPSILON = 1e-6\nDELTA_CLIP = 50.0\n\n\nclass FeedForwardModel(object):\n    \"\"\"The fully connected neural network model.\"\"\"\n    def __init__(self, config, bsde, sess):\n        self._config = config\n        self._bsde = bsde\n        self._sess = sess\n        # make sure consistent with FBSDE equation\n        self._dim = bsde.dim\n        self._num_time_interval = bsde.num_time_interval\n        self._total_time = bsde.total_time\n        # ops for statistics update of batch normalization\n        self._extra_train_ops = []\n\n    def train(self):\n        start_time = time.time()\n        # to save iteration results\n        training_history = []\n        # for validation\n        dw_valid, x_valid = self._bsde.sample(self._config.valid_size)\n        # can still use batch norm of samples in the validation phase\n        feed_dict_valid = {self._dw: dw_valid, self._x: x_valid, self._is_training: False}\n        # initialization\n        self._sess.run(tf.global_variables_initializer())\n        # begin sgd iteration\n        for step in range(self._config.num_iterations+1):\n            if step % self._config.logging_frequency == 0:\n                # sess.run Output tensors and metadata obtained when executing a session\n                loss, init = self._sess.run([self._loss, self._y_init], feed_dict=feed_dict_valid)\n                elapsed_time = time.time()-start_time+self._t_build\n                training_history.append([step, loss, init, elapsed_time])\n                if self._config.verbose:\n                    print(\"step: %5u,    loss: %.4e,   Y0: %.4e,  elapsed time %3u\" % (\n                        step, loss, init, elapsed_time))\n            #Obtain a sample of the Brownian motion and the forward process        \n            dw_train, x_train = self._bsde.sample(self._config.batch_size)\n            #Run the training\n            self._sess.run(self._train_ops, feed_dict={self._dw: dw_train,\n                                                       self._x: x_train,\n                                                       self._is_training: True})\n            t = 0\n            self._sess.run(self._subnetwork(xx,name), feed_dict={xx: x_train[:,:,t+1], name: str(t+1)})\n\n        return np.array(training_history)\n\n    #void method: does not return anything\n    def build(self):\n        start_time = time.time()\n        time_stamp = np.arange(0, self._bsde.num_time_interval) * self._bsde.delta_t\n        #tf.Variable you have to provide an initial value\n        #tf.placeholder you don't have to provide an initial value and you can specify it at run time\n        #initialize storage for brownian increments and the forward process\n        #A placeholder is a promise to provide a value later, like a function argument.\n        self._dw = tf.placeholder(TF_DTYPE, [None, self._dim, self._num_time_interval], name='dW')\n        self._x = tf.placeholder(TF_DTYPE, [None, self._dim, self._num_time_interval + 1], name='X')\n        self._is_training = tf.placeholder(tf.bool)\n        #tf.Variable you have to provide an initial value\n        #This is the first guess for Y_0, a unifrom extraction. The uniform distribution \n        #lies in the interval specified in config.py y_init_range\n        self._y_init = tf.Variable(tf.random_uniform([1],\n                                                     minval=self._config.y_init_range[0],\n                                                     maxval=self._config.y_init_range[1],\n                                                     dtype=TF_DTYPE))\n        #Initialize the control\n        z_init = tf.Variable(tf.random_uniform([1, self._dim],\n                                               minval=-.1, maxval=.1,\n                                               dtype=TF_DTYPE))\n        all_one_vec = tf.ones(shape=tf.stack([tf.shape(self._dw)[0], 1]), dtype=TF_DTYPE)\n        y = all_one_vec * self._y_init\n\n        z = tf.matmul(all_one_vec, z_init)\n\n        with tf.variable_scope('forward'):\n            #Evolve the BSDE forward in time\n            for t in range(0, self._num_time_interval-1):\n                y = y - self._bsde.delta_t * (\n                    self._bsde.f_tf(time_stamp[t], self._x[:, :, t], y, z)\n                ) + tf.reduce_sum(z * self._dw[:, :, t], 1, keepdims=True)\n                z = self._subnetwork(self._x[:, :, t + 1], str(t + 1)) \/ self._dim\n\n            # terminal time\n            y = y - self._bsde.delta_t * self._bsde.f_tf(\n                time_stamp[-1], self._x[:, :, -2], y, z\n            ) + tf.reduce_sum(z * self._dw[:, :, -1], 1, keepdims=True)\n\n            #loss at terminal time over all samples.\n            delta = y - self._bsde.g_tf(self._total_time, self._x[:, :, -1])\n            # use linear approximation outside the clipped range\n            # If tf.abs(delta) &lt; DELTA_CLIP then tf.square(delta) else...\n            self._loss = tf.reduce_mean(tf.where(tf.abs(delta) &lt; DELTA_CLIP, tf.square(delta),\n                                                 2 * DELTA_CLIP * tf.abs(delta) - DELTA_CLIP ** 2))\n        #End of scope \"forward\"\n\n        # train operations\n        global_step = tf.get_variable('global_step', [],\n                                      initializer=tf.constant_initializer(0),\n                                      trainable=False, dtype=tf.int32)\n        learning_rate = tf.train.piecewise_constant(global_step,\n                                                    self._config.lr_boundaries,\n                                                    self._config.lr_values)\n        #Returns a list of Variable objects\n        trainable_variables = tf.trainable_variables()\n\n        #Constructs symbolic derivatives of \\sum of ys w.r.t. x in xs.\n        grads = tf.gradients(self._loss, trainable_variables)\n\n        #Creates an instance of an Adam optimizer\n        optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate=learning_rate)\n\n        #Apply gradients to variables. This is the second part of minimize().\n        #It returns an Operation that applies gradients.\n        apply_op = optimizer.apply_gradients(zip(grads, trainable_variables),\n                                             global_step=global_step, name='train_step')\n        all_ops = [apply_op] + self._extra_train_ops\n        self._train_ops = tf.group(*all_ops)\n        self._t_build = time.time()-start_time\n\n    #This method returns the neural network that parametrizes the control\n    #This function accepts x as a tf object, not as a python object!\n    #Name is a string and it is a python object!\n    def _subnetwork(self, x, name):\n        with tf.variable_scope(name):\n            # standardize the path input first\n            # the affine  could be redundant, but helps converge faster\n            hiddens = self._batch_norm(x, name='path_input_norm')\n\n            #Loop over hidden layers: \n            #so many iteration as the number of hidden layers\n            for i in range(1, len(self._config.num_hiddens)-1):\n                hiddens = self._dense_batch_layer(hiddens,\n                                                  self._config.num_hiddens[i],\n                                                  activation_fn=tf.nn.relu,\n                                                  name='layer_{}'.format(i))\n            output = self._dense_batch_layer(hiddens,\n                                             self._config.num_hiddens[-1],\n                                             activation_fn=None,\n                                             name='final_layer')\n        return output\n\n    def _dense_batch_layer(self, input_, output_size, activation_fn=None,\n                           stddev=5.0, name='linear'):\n        with tf.variable_scope(name):\n            shape = input_.get_shape().as_list()\n            weight = tf.get_variable('Matrix', [shape[1], output_size], TF_DTYPE,\n                                     tf.random_normal_initializer(\n                                         stddev=stddev\/np.sqrt(shape[1]+output_size)))\n            #Create an affine combination\n            hiddens = tf.matmul(input_, weight)\n            #batch normalization\n            hiddens_bn = self._batch_norm(hiddens)\n            #The entries of the affine combination is sent to the activation function\n        if activation_fn:\n            return activation_fn(hiddens_bn)\n        else:\n            return hiddens_bn\n    \"\"\"\n    Normalizes a tensor by mean and variance, and applies (optionally)\n    a scale $\\gamma$ to it, as well as an offset $\\beta$:\n\n    https:\/\/arxiv.org\/abs\/1502.03167\n\n    Training Deep Neural Networks is complicated by the fact that \n    the distribution of each layer's inputs changes during training, \n    as the parameters of the previous layers change. \n    This slows down the training by requiring lower learning rates\n    and careful parameter initialization, and makes it notoriously hard\n    to train models with saturating nonlinearities. \n    We refer to this phenomenon as internal covariate shift, \n    and address the problem by normalizing layer inputs. \n\n    \"\"\"\n    def _batch_norm(self, x, affine=True, name='batch_norm'):\n        \"\"\"Batch normalization\"\"\"\n        with tf.variable_scope(name):\n            params_shape = [x.get_shape()[-1]]\n            beta = tf.get_variable('beta', params_shape, TF_DTYPE,\n                                   initializer=tf.random_normal_initializer(\n                                       0.0, stddev=0.1, dtype=TF_DTYPE))\n            gamma = tf.get_variable('gamma', params_shape, TF_DTYPE,\n                                    initializer=tf.random_uniform_initializer(\n                                        0.1, 0.5, dtype=TF_DTYPE))\n            moving_mean = tf.get_variable('moving_mean', params_shape, TF_DTYPE,\n                                          initializer=tf.constant_initializer(0.0, TF_DTYPE),\n                                          trainable=False)\n            moving_variance = tf.get_variable('moving_variance', params_shape, TF_DTYPE,\n                                              initializer=tf.constant_initializer(1.0, TF_DTYPE),\n                                              trainable=False)\n            # These ops will only be preformed when training\n            mean, variance = tf.nn.moments(x, [0], name='moments')\n            self._extra_train_ops.append(\n                moving_averages.assign_moving_average(moving_mean, mean, MOMENTUM))\n            self._extra_train_ops.append(\n                moving_averages.assign_moving_average(moving_variance, variance, MOMENTUM))\n            mean, variance = tf.cond(self._is_training,\n                                     lambda: (mean, variance),\n                                     lambda: (moving_mean, moving_variance))\n            y = tf.nn.batch_normalization(x, mean, variance, beta, gamma, EPSILON)\n            y.set_shape(x.get_shape())\n            return y\n<\/code><\/pre>\n\n<p>My objective is: once the training is over, I want to simulate paths of the process Y. For this I need to recall from memory all parameters of the neural networks that have been calibrated.<\/p>\n\n<p>I tried to define some def _simulate(): that performs the following loop<\/p>\n\n<pre><code>for t in range(0, self._num_time_interval-1):\n                y = y - self._bsde.delta_t * (\n                    self._bsde.f_tf(time_stamp[t], self._x[:, :, t], y, z)\n                ) + tf.reduce_sum(z * self._dw[:, :, t], 1, keepdims=True)\n                z = self._subnetwork(self._x[:, :, t + 1], str(t + 1)) \/ self._dim\n<\/code><\/pre>\n\n<p>but whenever I call _subnetwork I am told that the variables (weights and biases) are uninitialized, even though the networks are trained. Any help is appreciated.<\/p>\n","answers":[]}
+{"QuestionId":57522806,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-08-16T10:11:25.313","AcceptedAnswerId":null,"OwnerUserId":5405823.0,"Title":"constants in Pytorch Linear Module Class Definition","Body":"<p>What is <code>__constants__<\/code> in pytorch <code>class Linear(Module):<\/code> defined in <a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/nn\/modules\/linear.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/nn\/modules\/linear.html<\/a>? <\/p>\n\n<p>What is its functionality and why is it used?<\/p>\n\n<p>I have been searching around, but did not find any documentation. Please note that this does not mean the <code>__constants__<\/code> in torch script.<\/p>\n","answers":[{"AnswerId":"57534025","CreationDate":"2019-08-17T06:57:41.593","ParentId":null,"OwnerUserId":"6210807","Title":null,"Body":"<p>The <code>__constants__<\/code> you're talking about is, in fact, the one related to TorchScript. You can confirm it by using <code>git blame<\/code> <em>(when it was added and by who)<\/em> on GitHub. For example, for <a href=\"https:\/\/github.com\/pytorch\/pytorch\/blob\/master\/torch\/nn\/modules\/linear.py\" rel=\"nofollow noreferrer\"><code>torch\/nn\/modules\/linear.py<\/code><\/a>, check its <a href=\"https:\/\/github.com\/pytorch\/pytorch\/blame\/master\/torch\/nn\/modules\/linear.py\" rel=\"nofollow noreferrer\">git blame<\/a>.<\/p>\n\n<blockquote>\n  <p><a href=\"https:\/\/pytorch.org\/docs\/stable\/jit.html#python-defined-constants\" rel=\"nofollow noreferrer\">TorchScript<\/a> also provides a way to use constants that are defined in Python. These can be used to hard-code hyper-parameters into the function, or to define universal constants.<\/p>\n  \n  <p>-- Attributes of a ScriptModule can be marked constant by listing them as a member of the <strong>constants<\/strong> property of the class:<\/p>\n<\/blockquote>\n\n<pre class=\"lang-py prettyprint-override\"><code>class Foo(torch.jit.ScriptModule):\n    __constants__ = ['a']\n\n    def __init__(self):\n        super(Foo, self).__init__(False)\n        self.a = 1 + 4\n\n   @torch.jit.script_method\n   def forward(self, input):\n       return self.a + input\n<\/code><\/pre>\n"}]}
+{"QuestionId":57523260,"AnswerCount":0,"Tags":"<java><python><tensorflow><caffe><tensorflow-lite>","CreationDate":"2019-08-16T10:46:04.733","AcceptedAnswerId":null,"OwnerUserId":835084.0,"Title":"How do I convert an image in Java as is done for Keras in Python?","Body":"<p>I have an image that I want to run in tensorflowlite. I have saved the model.tflite file and am trying to format an image to provide results for. I produced the following code:<\/p>\n\n<pre><code>private ByteBuffer convertBitmapToByteBuffer(Bitmap bitmap) {\n        ByteBuffer byteBuffer = ByteBuffer.allocateDirect(ModelConfig.MODEL_INPUT_SIZE);\n        byteBuffer.order(ByteOrder.nativeOrder());\n        int[] pixels = new int[ModelConfig.INPUT_IMG_SIZE_WIDTH * ModelConfig.INPUT_IMG_SIZE_HEIGHT];\n        bitmap.getPixels(pixels, 0, bitmap.getWidth(), 0, 0, bitmap.getWidth(), bitmap.getHeight());\n        for (int pixel : pixels) {\n            float rChannel = ((pixel &gt;&gt; 16) &amp; 0xFF); \/ 255.f ;\n            float gChannel = ((pixel &gt;&gt; 8) &amp; 0xFF); \/ 255.f;\n            float bChannel = ((pixel) &amp; 0xFF); \/ 255.f;\n            byteBuffer.putFloat(rChannel);\n            byteBuffer.putFloat(gChannel);\n            byteBuffer.putFloat(bChannel);\n        }\n}\n<\/code><\/pre>\n\n<p>However, when I put this into my model the output is unexpected. Reading into it, it appears that my Python implementation is doing some clever back-end processing that I do not know how to replicate into Java. As an example, it seems to format the data into <em>caffe<\/em> mode. The shape is (1, 224, 224, 3) and I'm not certain also how to replicate this for the Java model. My current Python code that works well is:<\/p>\n\n<pre><code>    img = load_img(image_path, target_size=(224, 224))\n    img = img_to_array(img)\n    img = np.expand_dims(img, axis=0)\n    print(img.shape)\n    img = preprocess_input(img)\n    return img\n<\/code><\/pre>\n\n<p>I'm unsure how, given a Bitmap object, to convert the image into the correct format for my tensorflow model. Currently the model runs on the current input, but it just produces the wrong results. I'd like it to replicate the results achieved when I use the model in Python.<\/p>\n","answers":[]}
+{"QuestionId":57523456,"AnswerCount":0,"Tags":"<python-3.x><keras><neural-network><keras-2>","CreationDate":"2019-08-16T11:00:32.787","AcceptedAnswerId":null,"OwnerUserId":4920401.0,"Title":"How does model.evaluate in Keras work and how to recreate it manually?","Body":"<p>How does Keras' <code>model.evaluate()<\/code> work? In particular, how does the <code>batch_size<\/code> argument affect the calculation?<\/p>\n\n<p>The documentation says the loss\/metrics are calculated as averages over the batches. However, as there is only one scalar output for each loss\/metrics (which should represent the total average over all data, over all batch-averages), the result should not depend on <code>batch_size<\/code> choice (at least as long as the sum of all data is dividable by the batch_size).<\/p>\n\n<p>But after building and training a network (consisting of Conv2D, Conv2DTranspose, MaxPooling2D, and BatchNormalization, using ReLU as activations), I tried evaluating it on my test set of 60 samples:<\/p>\n\n<ul>\n<li>Evaluation with batch_size = 60 gave loss 0.1375531554222107 <\/li>\n<li>Evaluation with batch_size = 10 gave loss 0.1381820539633433<\/li>\n<li>Evaluation with batch_size = 3  gave loss 0.14014312624931335<\/li>\n<li>Evaluation with batch_size = 1  gave loss 0.15437299211819966<\/li>\n<\/ul>\n\n<p>The entire dataset (60 samples) was dividable by all of the <code>batch_size<\/code>'s here (1,3,10,60). Yet the outputs vary a lot. Could be due to batch normalization, but I don't think so because if I run the evaluation multiple times, the numbers will always be the same.<\/p>\n\n<p>Even if there would be no shuffling before defining batches, and the numbers were supposed to always be the same, it still doesn't explain why I can't then reproduce the last number (the evaluation with <code>batch_size=1<\/code>) by averaging individual sample evaluations. That is, why the following two lines produce a very different result:<\/p>\n\n<pre><code>    model.evaluate(testset, testlabels, batch_size=1)\n<\/code><\/pre>\n\n<p>and:<\/p>\n\n<pre><code>    losses = [model.evaluate(testset[i], testlabel[i], batch_size=1) for i in range(60)]\n    np.mean(losses)\n<\/code><\/pre>\n\n<p>Here, assume of course that <code>model.evaluate<\/code> returns only 1 scalar, that is the final loss, and no metrics or intermediate losses. In summary - how does <code>keras.evaluate<\/code> work internally; how are results based on different <code>batch_size<\/code>'s connected and what is their relationship to the simple averaging of per-sample losses?<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/57256171\/optimum-batch-size-for-model-evaluate-in-keras\">This similar question<\/a> doesn't help (asks after optimal choice, while I just want to know how it works).<\/p>\n","answers":[]}
+{"QuestionId":57523482,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><loss-function>","CreationDate":"2019-08-16T11:02:19.537","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"how to define cost function as root mean squared logrithmic error in tensorflow in python","Body":"<p>I want to use a self-designed cost function. can anyone please help me with the code for defining a cost function for my regression model as root mean square logarithmic error<\/p>\n","answers":[{"AnswerId":"57524107","CreationDate":"2019-08-16T11:47:29.530","ParentId":null,"OwnerUserId":"2455494","Title":null,"Body":"<p>You can pass an arbitrary loss function into your <code>model.compile()<\/code> call. SO you could implement your own RMLSE function something like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def rmsle(y_true, y_pred):\n  eps = tf.constant(0.00001)\n  return tf.math.sqrt(\n      tf.reduce_mean(\n        tf.pow(tf.math.log(eps + y_true) - tf.math.log(eps + y_pred), 2)))\n<\/code><\/pre>\n\n<p>and use it in a model.<\/p>\n\n<p>Here is a complete example you can replicate (here we are trying to treat the mnist digits as a regression problem which is obvs a little silly but you can adapt to your own problem as you see fit):<\/p>\n\n<pre><code>import tensorflow as tf\ntf.enable_eager_execution()\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\ny_train, y_test = tf.cast(y_train, tf.float32), tf.cast(y_test, tf.float32)\n\nmodel = tf.keras.models.Sequential([\n  tf.keras.layers.Flatten(input_shape=(28, 28)),\n  tf.keras.layers.Dense(128, activation='relu'),\n  tf.keras.layers.Dropout(0.2),\n  tf.keras.layers.Dense(1,activation='sigmoid')\n])\n\nmodel.compile(optimizer='adam', loss=rmsle)\nmodel.fit(x_train, y_train, epochs=5)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57524007,"AnswerCount":0,"Tags":"<python><keras><imblearn>","CreationDate":"2019-08-16T11:40:43.800","AcceptedAnswerId":null,"OwnerUserId":11935814.0,"Title":"SMOTE and KERAS LSTM in Pipeline: ValueError: not enough values to unpack (expected 2, got 1)","Body":"<p>I'm trying to create an LSTM model with Keras on an imbalanced dataset, and then tune hyperparameters using sklearn RandomSearchCV.<\/p>\n\n<p>To achieve this, I'm using an imblearn pipeline.<\/p>\n\n<p>Why do I get this? Running both <code>SMOTE<\/code> and the classifier on their own seem to work, with no problems in the data shape. Any suggestions at all would be massively appreciated. <\/p>\n\n<pre><code>from imblearn.pipeline import Pipeline\nfrom imblearn.over_sampling import SMOTE\n\ndef chat_model(regularizer, batch_size, lstm_size, dropout):    \n    model = Sequential()\n    embedding_layer = Embedding(vocab_size, 100, weights=[embedding_matrix], \n    input_length=maxlen , trainable=False)\n    model.add(embedding_layer)\n    model.add(LSTM(lstm_size))\n    model.add(Dropout(dropout))\n    model.add(Dense(1, activation='sigmoid', name='main_output', \n    activity_regularizer=regularizers.l2(regularizer)))\n    model.compile( loss='binary_crossentropy', metrics=['acc'], optimizer = \n    'adam')\n\n    return model\n\nclf = KerasClassifier(build_fn=chat_model, verbose=10)        \n\nsm  = SMOTE()\npipe = Pipeline([\n        ('oversampler', sm),\n        ('clf', clf)])\n\nparams = { 'clf__regularizer' : [0.001, 0.0001], \n          'clf__batch_size' : [64, 128], \n         'clf__lstm_size':[200, 300, 400, 500], \n         'clf__dropout':[0.2, 0.3, 0.4]}\n\ngrid = GridSearchCV(estimator=pipe, cv=KFold(3), param_grid=params, \n                            scoring = 'balanced_accuracy')\n\ngrid.fit(X_train, y_train, verbose =10)```\n\n#THIS RETURNS:\nValueError: not enough values to unpack (expected 2, got 1)\n\nThe traceback is:\n---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-1265-e1d13523e326&gt; in &lt;module&gt;\n     33                             scoring = 'balanced_accuracy')\n     34 \n---&gt; 35 grid.fit(X_train, y_train, verbose =10)\n\n~\\AppData\\Roaming\\Python\\Python37\\site-packages\\sklearn\\model_selection\\_search.py in fit(self, X, y, groups, **fit_params)\n    720                 return results_container[0]\n    721 \n--&gt; 722             self._run_search(evaluate_candidates)\n    723 \n    724         results = results_container[0]\n\n~\\AppData\\Roaming\\Python\\Python37\\site-packages\\sklearn\\model_selection\\_search.py in _run_search(self, evaluate_candidates)\n   1189     def _run_search(self, evaluate_candidates):\n   1190         \"\"\"Search all candidates in param_grid\"\"\"\n-&gt; 1191         evaluate_candidates(ParameterGrid(self.param_grid))\n   1192 \n   1193 \n\n~\\AppData\\Roaming\\Python\\Python37\\site-packages\\sklearn\\model_selection\\_search.py in evaluate_candidates(candidate_params)\n    709                                for parameters, (train, test)\n    710                                in product(candidate_params,\n--&gt; 711                                           cv.split(X, y, groups)))\n    712 \n    713                 all_candidate_params.extend(candidate_params)\n\n~\\AppData\\Roaming\\Python\\Python37\\site-packages\\sklearn\\externals\\joblib\\parallel.py in __call__(self, iterable)\n    915             # remaining jobs.\n    916             self._iterating = False\n--&gt; 917             if self.dispatch_one_batch(iterator):\n    918                 self._iterating = self._original_iterator is not None\n    919 \n\n~\\AppData\\Roaming\\Python\\Python37\\site-packages\\sklearn\\externals\\joblib\\parallel.py in dispatch_one_batch(self, iterator)\n    757                 return False\n    758             else:\n--&gt; 759                 self._dispatch(tasks)\n    760                 return True\n    761 \n\n~\\AppData\\Roaming\\Python\\Python37\\site-packages\\sklearn\\externals\\joblib\\parallel.py in _dispatch(self, batch)\n    714         with self._lock:\n    715             job_idx = len(self._jobs)\n--&gt; 716             job = self._backend.apply_async(batch, callback=cb)\n    717             # A job can complete so quickly than its callback is\n    718             # called before we get here, causing self._jobs to\n\n~\\AppData\\Roaming\\Python\\Python37\\site-packages\\sklearn\\externals\\joblib\\_parallel_backends.py in apply_async(self, func, callback)\n    180     def apply_async(self, func, callback=None):\n    181         \"\"\"Schedule a func to be run\"\"\"\n--&gt; 182         result = ImmediateResult(func)\n    183         if callback:\n    184             callback(result)\n\n~\\AppData\\Roaming\\Python\\Python37\\site-packages\\sklearn\\externals\\joblib\\_parallel_backends.py in __init__(self, batch)\n    547         # Don't delay the application, to avoid keeping the input\n    548         # arguments in memory\n--&gt; 549         self.results = batch()\n    550 \n    551     def get(self):\n\n~\\AppData\\Roaming\\Python\\Python37\\site-packages\\sklearn\\externals\\joblib\\parallel.py in __call__(self)\n    223         with parallel_backend(self._backend, n_jobs=self._n_jobs):\n    224             return [func(*args, **kwargs)\n--&gt; 225                     for func, args, kwargs in self.items]\n    226 \n    227     def __len__(self):\n\n~\\AppData\\Roaming\\Python\\Python37\\site-packages\\sklearn\\externals\\joblib\\parallel.py in &lt;listcomp&gt;(.0)\n    223         with parallel_backend(self._backend, n_jobs=self._n_jobs):\n    224             return [func(*args, **kwargs)\n--&gt; 225                     for func, args, kwargs in self.items]\n    226 \n    227     def __len__(self):\n\n~\\AppData\\Roaming\\Python\\Python37\\site-packages\\sklearn\\model_selection\\_validation.py in _fit_and_score(estimator, X, y, scorer, train, test, verbose, parameters, fit_params, return_train_score, return_parameters, return_n_test_samples, return_times, return_estimator, error_score)\n    526             estimator.fit(X_train, **fit_params)\n    527         else:\n--&gt; 528             estimator.fit(X_train, y_train, **fit_params)\n    529 \n    530     except Exception as e:\n\n~\\AppData\\Roaming\\Python\\Python37\\site-packages\\imblearn\\pipeline.py in fit(self, X, y, **fit_params)\n    235 \n    236         \"\"\"\n--&gt; 237         Xt, yt, fit_params = self._fit(X, y, **fit_params)\n    238         if self._final_estimator is not None:\n    239             self._final_estimator.fit(Xt, yt, **fit_params)\n\n~\\AppData\\Roaming\\Python\\Python37\\site-packages\\imblearn\\pipeline.py in _fit(self, X, y, **fit_params)\n    166                                 if step is not None)\n    167         for pname, pval in six.iteritems(fit_params):\n--&gt; 168             step, param = pname.split('__', 1)\n    169             fit_params_steps[step][param] = pval\n    170         Xt = X\n\n\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57524095,"AnswerCount":1,"Tags":"<python><keras><deep-learning>","CreationDate":"2019-08-16T11:46:45.777","AcceptedAnswerId":57524220.0,"OwnerUserId":3630804.0,"Title":"How to use binary cross entropy for a sequence of bits","Body":"<p>I want to apply binary cross entropy function for a prediction of a sequence of N length bits ( for example a sequence of 10 bits  <code>[1 0 1 0 1 1 0 1 0 0]<\/code> as ground truth). Model output after a softmax layer is of size <code>[10, 2]<\/code>. second dimension refers to <code>[Prob(x=0), Prob(x=1)]<\/code>. I want to calculate the <code>binary_crossentropy<\/code>  loss for this sequence. <\/p>\n\n<p>So my ground truth is of size [?, 10] and my model output is of size [?, 10 , 2]. How to use <code>binary_crossentropy<\/code> loss in this scenario?  <\/p>\n\n<p>As for this, I get an error as \n<code>Error when checking target: expected model_3 to have 3 dimensions, but got array with shape (batchsize, 10).<\/code><\/p>\n\n<p>When I use Prob(x=0) or  Prob(x=1)] from the softmax output, I do not get desired results. How can I achieve this?<\/p>\n","answers":[{"AnswerId":"57524220","CreationDate":"2019-08-16T11:55:30.553","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>For a binary bit ground truth of shape <code>(None, 10)<\/code>, the model output should be of the same shape. Since each bit is independent, you should use a sigmoid activation at the output, not a softmax one, in order to use the <code>binary_crossentropy<\/code> loss:<\/p>\n\n<pre><code>output = Dense(10, activation=\"sigmoid\")(layer_input)\n# Alternatively you can just use an activation layer:\noutput = Activation(\"sigmoid\")(layer_input)\n\n\nmodel = Model(input, output)\nmodel.compile(optimizer=\"adam\", loss=\"binary_crossentropy\")\n<\/code><\/pre>\n\n<p>An important detail is that the output of the model is continuous in <code>[0, 1]<\/code>, to get the bit predictions, you have to threshold the output of the model at 0.5 (this is how keras computes binary accuracy):<\/p>\n\n<pre><code>raw_preds = model.predict(some_inputs)\nbit_preds = raw_preds &gt; 0.5\n<\/code><\/pre>\n\n<p>If you want to keep the current model using softmax, then you have to one-hot encode the labels so they have shape <code>(None, 10, 2)<\/code>, and then use <code>categorical_crossentropy<\/code> as the loss.<\/p>\n"}]}
+{"QuestionId":57524137,"AnswerCount":1,"Tags":"<python><tensorflow><memory><cpu>","CreationDate":"2019-08-16T11:50:08.537","AcceptedAnswerId":null,"OwnerUserId":9442918.0,"Title":"My CNN Model uses too much memory on my GPU. How can I host some Tensors on my CPU memory?","Body":"<p>I'm training a CNN model on an NVidia RTX 2080, it became bigger and bigger and I have now some memory issues with the card. I read some paper about this subject, and it seems possible with Tensorflow to host some nodes on the CPU memory during the training and retrieve it in the GPU memory later when needed (as in <a href=\"http:\/\/learningsys.org\/nips17\/assets\/papers\/paper_18.pdf\" rel=\"nofollow noreferrer\">http:\/\/learningsys.org\/nips17\/assets\/papers\/paper_18.pdf<\/a>).<\/p>\n\n<p>Any ideas\/docs\/examples?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"57524244","CreationDate":"2019-08-16T11:57:00.520","ParentId":null,"OwnerUserId":"11339311","Title":null,"Body":"<p>Without any code it's difficult to help. \nGenerally you can take a look at the <a href=\"https:\/\/www.tensorflow.org\/guide\/using_gpu\" rel=\"nofollow noreferrer\">documentation<\/a>.<\/p>\n\n<p>For example with:<\/p>\n\n<pre><code>with 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<\/code><\/pre>\n\n<p>you can create variables explicitly on the cpu. <\/p>\n"}]}
+{"QuestionId":57524281,"AnswerCount":1,"Tags":"<python><tensorflow><image-processing><keras>","CreationDate":"2019-08-16T11:59:37.270","AcceptedAnswerId":57543243.0,"OwnerUserId":1009508.0,"Title":"ImageGenerator using Resnet","Body":"<p>I am trying to build a image generator that will:<\/p>\n\n<ol>\n<li>Take in a raw image<\/li>\n<li>Read in the image and resize it to (224,224,3) for resnet50<\/li>\n<li>Perform data augmentation on it (rotation, flipping, etc.)<\/li>\n<li>Create a Resnet50 feature for it (using model.predict)<\/li>\n<li>Then output this in the image generator via yield method<\/li>\n<\/ol>\n\n<p>What I cannot figure out is how to actually do this. I am making this as a triple loss so one image is an anchor, positive, and negative image.<\/p>\n\n<p>It says:<\/p>\n\n<blockquote>\n  <p>TypeError: len() of unsized object<\/p>\n<\/blockquote>\n\n<h2>Updates<\/h2>\n\n<p>I have changed the code a bit but now it is giving me this error.<\/p>\n\n<pre><code>Error when checking input: expected input_1 to have 4 dimensions, but got array with shape (224, 224, 3)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57543243","CreationDate":"2019-08-18T09:26:24.787","ParentId":null,"OwnerUserId":"9571777","Title":null,"Body":"<p>remember that a network expects a batch size. if you're submitting only one image of shape (244,244,3), you'll need to first <code>np.expand_dims(image, axis=0)<\/code> to achieve shape (1,244,244,3) prior to prediction. That will be 1 image of 244 by 244 with 3 layers, i.e. (1,244,244,3).<\/p>\n\n<p>It seems like you have some other dimensions, but the point remains, get that batch size dimension in there; that's what your recent error is saying.<\/p>\n"}]}
+{"QuestionId":57524334,"AnswerCount":0,"Tags":"<tensorflow><conv-neural-network><tensorflow2.0><dropout><mobilenet>","CreationDate":"2019-08-16T12:04:12.513","AcceptedAnswerId":null,"OwnerUserId":362376.0,"Title":"Adding Dropout to MobileNet with TensorFlow 2","Body":"<p>I'm using MobileNet and TensorFlow 2 to distinguish between 4 fairly similar toys. I have exactly 750 images for each toy and one label that contains 750 'negative' images, without any of the toys. <\/p>\n\n<p>I've used MobileNet before for this with a fair degree of success, but something about this case is causing a lot of overfitting (~30-40% discrepancy between training\/validation accuracy). The model very quickly trains to a training accuracy of about 99.8% in 3 epochs but the validation accuracy is stuck around 75%. The validation data set is a random set of 20% of the input images. When looking at the accuracy of the model, there's a strong bias towards one of the toys with a lot of the other toys falsely identified as that toy.<\/p>\n\n<p>I've tried pretty much everything out there to combat this:<\/p>\n\n<p>I've added Dropout after the Conv2D layer that's added to the top of MobileNet and tried various dropout rates between 0.2 and 0.9.<\/p>\n\n<pre><code>model = tf.keras.Sequential([\n  base_model,\n  tf.keras.layers.Conv2D(32, 3, activation='relu'),\n  tf.keras.layers.Dropout(0.5),\n  tf.keras.layers.GlobalAveragePooling2D(),\n  tf.keras.layers.Dense(label_count, activation='softmax')\n])\n<\/code><\/pre>\n\n<p>I've added an additional Dropout layer before the Conv2D layer, which seemed to marginally improve things:<\/p>\n\n<pre><code>model = tf.keras.Sequential([\n  base_model,\n  tf.keras.layers.Dropout(0.5),\n  tf.keras.layers.Conv2D(32, 3, activation='relu'),\n  tf.keras.layers.Dropout(0.5),\n  tf.keras.layers.GlobalAveragePooling2D(),\n  tf.keras.layers.Dense(label_count, activation='softmax')\n])\n<\/code><\/pre>\n\n<p>I've also added more test data, trying mixtures of photographs of the toys in various lighting conditions and backgrounds and generated images of the toys superimposed over random backgrounds. None of these have a significant impact.<\/p>\n\n<p>Should I be adding dropout to the MobileNet model rather than just to the layers that I'm adding after it? I came across <a href=\"https:\/\/gist.github.com\/koshian2\/f592f342634e815ad81998b51dd656a4\" rel=\"nofollow noreferrer\">this code<\/a> on github that does this, but I've no idea if this is actually a good idea or not - or quite how to achieve this with TensorFlow 2.<\/p>\n\n<p>Is this sensible, or feasible? <\/p>\n\n<p>Alternatively, the only other ideas I can think of are:<\/p>\n\n<ul>\n<li>Capture more images, to make the training harder - but I'd have thought 750 for each item should be enough to do a pretty good job. <\/li>\n<li>Don't use MobileNet, create a neural network from scratch or use another pre-existing one.<\/li>\n<\/ul>\n","answers":[]}
+{"QuestionId":57525274,"AnswerCount":2,"Tags":"<python><tensorflow><version>","CreationDate":"2019-08-16T13:09:33.720","AcceptedAnswerId":57525346.0,"OwnerUserId":735008.0,"Title":"How to check whether the current tensorflow version is at least a specific version in a python script?","Body":"<p>I would like to make sure that my tensorflow module is at least a certain version (for example, 1.13.2) inside my Python script. For me, it's not quite clear how to do this easily.<\/p>\n\n<p>The following is one example I tried:<\/p>\n\n<pre><code>import tensorflow as tf\nimport re\n\nmatch = re.match(\"(\\d+)\\.(\\d+)\\.(\\d+).*\", tf.__version__)\nbase = 100\nnum1 = match.group(1) * base ** 2 + match.group(2) * base + match.group(3)\nnum2 = 1 ** base ** 2 + 13 * base + 2 # Corresponds to 1.13.2\n\nassert num1 &gt;= num2, (\"The required version is at least 1.13.2\")\n<\/code><\/pre>\n\n<p>The above code has several problems.<\/p>\n\n<p>1) If some version number is more than 100 (e.g. 1.12.801), it won't work. <\/p>\n\n<p>2) The code itself is very ugly.<\/p>\n\n<p>I know that the above is definitely not the way to go.\nCould you suggest any good ways?<\/p>\n","answers":[{"AnswerId":"57525346","CreationDate":"2019-08-16T13:14:53.243","ParentId":null,"OwnerUserId":"7262969","Title":null,"Body":"<p>I would say that the easiest solution is splitting the version number by the dots, and compare the parts converted to int.<\/p>\n\n<pre><code>reqVersion = \"1.13.2\"\nfor reqPart, part in zip(map(int, reqVersion.split(\".\")), map(int, tf.__version__.split(\".\"))):\n    assert reqPart &lt;= part\n    if reqPart &lt; part:\n        break\n<\/code><\/pre>\n"},{"AnswerId":"57525403","CreationDate":"2019-08-16T13:19:37.467","ParentId":null,"OwnerUserId":"11597215","Title":null,"Body":"<pre><code>import tensorflow as tf\n\ntheir=tf.__version__.split(\".\")\nour=\"1.13.2\".split(\".\")\n\nlessflag=False\n\n\nfor i in range(3):\n    if(int(our[i])&gt;int(their[i])):\n        lessflag=True\n        break\n    elif(int(our[i])&lt;int(their[i])):\n        lessflag=False\n        break\n\n\nif(lessflag):\n    print(\"not supported\")\n<\/code><\/pre>\n"}]}
+{"QuestionId":57525456,"AnswerCount":1,"Tags":"<pytorch><jit><libtorch>","CreationDate":"2019-08-16T13:23:24.243","AcceptedAnswerId":57535428.0,"OwnerUserId":3486078.0,"Title":"Libtorch: cannot load traced lstm scriptmodel","Body":"<p>I save a pytorch ScriptModule and load it using libtorch. However I encountered the following problem\n<img src=\"https:\/\/wx1.sinaimg.cn\/mw690\/93098207gy1g61u188w88j20sw0e1acr.jpg\" alt=\"jpg\"><\/p>\n\n<p>I use linux subsystem under win10 and I use pytorch 1.2. <\/p>\n\n<p>To reproduce my problem, you could run this piece of python code to save a pt model<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nimport torch.nn as nn\n\n\n# TODO: https:\/\/github.com\/pytorch\/pytorch\/issues\/23930\nclass test(torch.jit.ScriptModule):\n\n    def __init__(self, vocab_size=10, rnn_dims=512):\n        super().__init__()\n        self.word_embeds = nn.Embedding(vocab_size, rnn_dims)\n        self.emb_drop = nn.Dropout(0.1)\n        self.rnn = nn.LSTM(input_size=rnn_dims, hidden_size=rnn_dims, batch_first=True,\n                           num_layers=2, dropout=0.1)\n        # delattr(self.rnn, 'forward_packed')\n\n    @torch.jit.script_method\n    def forward(self, x):\n        h1 = (torch.zeros(2, 1, 512), torch.zeros(2, 1, 512))\n        embeds = self.emb_drop(self.word_embeds(x))\n        out, h1 = self.rnn(embeds, h1)\n\n        return h1\n\n\nmodel = test()\n\ninput = torch.ones((1,3)).long()\noutput = model(input)\nprint('output', output)\n\n# torch.onnx.export(model,  # model being run\n#                   input,\n#                   'test.onnx',\n#                   example_outputs=output)\n#torch.jit.trace(model, (torch.ones((1,3)).long(), torch.ones((3,1))), check_trace=False)\nmodel.save('lstm_test.pt')\n<\/code><\/pre>\n\n<p>Then load the model in libtorch.<\/p>\n\n<p>I don't know why I have this error. I don't use PackedSequence at all. Hope some could help me out.<\/p>\n","answers":[{"AnswerId":"57535428","CreationDate":"2019-08-17T10:26:19.517","ParentId":null,"OwnerUserId":"3486078","Title":null,"Body":"<p>I know what's wrong now. The libtorch version is of wrong version on the official website. It's ok now when I use the correct libtorch 1.2. Refer to issue <a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/24382\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/pytorch\/issues\/24382<\/a> <\/p>\n"}]}
+{"QuestionId":57526114,"AnswerCount":0,"Tags":"<python><pytorch><generative-adversarial-network>","CreationDate":"2019-08-16T14:09:19.017","AcceptedAnswerId":null,"OwnerUserId":11506449.0,"Title":"Music GAN loss not decreasing for generator","Body":"<p>I've been trying to write a GAN for music generation, but something seems to be wrong with my training because the loss on the real data and the loss for the generator are increasing, while the loss on the fake data is almost zero. My generator is a simple fully-connected network and my discriminator is a convolutional network. Here is my training code and my resulting losses. Can someone help me with what I am doing wrong?<\/p>\n\n<pre><code>def train_model(modelD, modelG, criterion, optimizerD, optimizerG, n_epochs = 5):\n\nrandom = torch.from_numpy(np.random.rand(100)).float()\nmeasures = torch.from_numpy(notes.reshape(-1, 97, 64)).float()\nnum_measures = measures.shape[0]\nlosses = []\n# set the model to train mode initially\nmodel.train()\nfor epoch in range(n_epochs):\n    print(\"Epoch: \" + str(epoch+1))\n    for measure in measures:\n        real_losses = []\n        fake_losses = []\n        G_losses = []\n    #----------------------------------\n        # Train discriminator with all-real batch\n        inputsD, labelsD = measure, torch.Tensor([1]).long()\n        inputsD = inputsD.to(device)\n        labelsD = labelsD.to(device)\n        optimizerD.zero_grad()\n\n        # forward + backward + optimize\n        outputsD = modelD(inputsD)\n        lossD_real = criterion(outputsD, labelsD)\n        lossD_real.backward(retain_graph=True)\n        real_losses.append(lossD_real.detach().numpy())\n    #----------------------------------\n        # Train discriminator with all-fake batch\n        inputsD, labelsD = modelG(random), torch.Tensor([0]).long()\n        inputsD = inputsD.to(device)\n        labelsD = labelsD.to(device)\n\n        # forward + backward + optimize\n        outputsD = modelD(inputsD)\n        lossD_fake = criterion(outputsD, labelsD)\n        lossD_fake.backward(retain_graph=True)\n        fake_losses.append(lossD_fake.detach().numpy())\n        optimizerD.step()\n    #----------------------------------\n        # Train G network\n        optimizerG.zero_grad()\n        outputs = modelD(inputsD)       \n\n        # forward + backward + optimize\n        lossG = criterion(outputs, torch.Tensor([1]).long())\n        lossG.backward()\n        G_losses.append(lossG.detach().numpy())\n        optimizerG.step()\n\n    print(\"Real Loss: \" + str(sum(real_losses)))\n    print(\"Fake Loss: \" + str(sum(fake_losses)))\n    print(\"G Loss: \" + str(sum(G_losses)))\n<\/code><\/pre>\n\n<p>Generator and Discriminator:<\/p>\n\n<pre><code>class Generator(torch.nn.Module):\ndef __init__(self):\n    super(Generator, self).__init__()\n    self.layers = nn.Sequential(\n        nn.Linear(100, 512),\n        nn.ReLU(),\n        nn.Linear(512, 2048),\n        nn.ReLU(),\n        nn.Linear(2048, 6208))\ndef forward(self, x):\n    x = self.layers(x)\n    x = x.view(97, 64)\n    return x\n\nclass Discriminator(torch.nn.Module):\ndef __init__(self):\n    super(Discriminator, self).__init__()\n    self.layer1 = nn.Sequential(\n        nn.Conv2d(1, 16, kernel_size=5, stride=1, padding=2),\n        nn.ReLU(),\n        nn.MaxPool2d(kernel_size=2, stride=2))\n    self.layer2 = nn.Sequential(\n        nn.Conv2d(16, 32, kernel_size=5, stride=1, padding=2),\n        nn.ReLU(),\n        nn.MaxPool2d(kernel_size=2, stride=2))\n    self.layer3 = nn.Sequential(\n        nn.Linear(12288, 1000),\n        nn.ReLU(),\n        nn.Linear(1000, 50),\n        nn.ReLU(),\n        nn.Linear(50, 2),\n        nn.LogSoftmax(dim=1))\ndef forward(self, x):\n    x = x.view(-1, 1, 97, 64)\n    x = self.layer1(x)\n    x = self.layer2(x)\n    x = x.view(1, -1)\n    x = self.layer3(x)\n    return x\n<\/code><\/pre>\n\n<p>Settings:<\/p>\n\n<pre><code>modelD = Discriminator()\nmodelG = Generator()\ncriterion = nn.CrossEntropyLoss()\noptimizerD = optim.SGD(modelD.parameters(), lr=0.001, momentum=0.9)\noptimizerG = optim.SGD(modelG.parameters(), lr=0.1, momentum=0.9)\nepochs = 100\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>Epoch: 1\nReal Loss: 0.6634323596954346\nFake Loss: 0.7299578189849854\nG Loss: 0.6627244353294373\nEpoch: 2\nReal Loss: 0.7140370011329651\nFake Loss: 0.6786684989929199\nG Loss: 0.713873028755188\nEpoch: 3\nReal Loss: 0.7706409692764282\nFake Loss: 0.6278889775276184\nG Loss: 0.7697595953941345\nEpoch: 4\nReal Loss: 0.8299549221992493\nFake Loss: 0.5787992477416992\nG Loss: 0.8297436833381653\nEpoch: 5\nReal Loss: 0.8929369449615479\nFake Loss: 0.5315663814544678\nG Loss: 0.8940553069114685\nEpoch: 6\nReal Loss: 0.9610943794250488\nFake Loss: 0.48426902294158936\nG Loss: 0.9678256511688232\nEpoch: 7\nReal Loss: 1.0396536588668823\nFake Loss: 0.42658042907714844\nG Loss: 1.0722308158874512\nEpoch: 8\nReal Loss: 1.1386523246765137\nFake Loss: 0.3636009395122528\nG Loss: 1.2055680751800537\nEpoch: 9\nReal Loss: 1.2815502882003784\nFake Loss: 0.30338314175605774\nG Loss: 1.3635222911834717\nEpoch: 10\nReal Loss: 1.5050387382507324\nFake Loss: 0.24401997029781342\nG Loss: 1.5601530075073242\nEpoch: 11\nReal Loss: 1.8629376888275146\nFake Loss: 0.1645389348268509\nG Loss: 1.9329965114593506\nEpoch: 12\nReal Loss: 2.417865514755249\nFake Loss: 0.0913706049323082\nG Loss: 2.5052433013916016\nEpoch: 13\nReal Loss: 3.132223606109619\nFake Loss: 0.04399847984313965\nG Loss: 3.220311403274536\nEpoch: 14\nReal Loss: 3.8472046852111816\nFake Loss: 0.021415308117866516\nG Loss: 3.9200611114501953\nEpoch: 15\nReal Loss: 4.434380531311035\nFake Loss: 0.011867986992001534\nG Loss: 4.49014949798584\nEpoch: 16\nReal Loss: 4.873379707336426\nFake Loss: 0.007633917964994907\nG Loss: 4.9159369468688965\nEpoch: 17\nReal Loss: 5.198086738586426\nFake Loss: 0.0055174920707941055\nG Loss: 5.230365753173828\nEpoch: 18\nReal Loss: 5.446741104125977\nFake Loss: 0.004297664389014244\nG Loss: 5.473706245422363\nEpoch: 19\nReal Loss: 5.646566390991211\nFake Loss: 0.0035177513491362333\nG Loss: 5.66972541809082\nEpoch: 20\nReal Loss: 5.814294338226318\nFake Loss: 0.002974849194288254\nG Loss: 5.834484100341797\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57526164,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow2.0>","CreationDate":"2019-08-16T14:13:02.377","AcceptedAnswerId":57528083.0,"OwnerUserId":3860928.0,"Title":"What is the relationship between `tf.function` and `autograph.to_graph` in Tensorflow?","Body":"<p>Similar results can be obtained via <code>tf.function<\/code> and <code>autograph.to_graph<\/code>.<br>\nHowever this seems to be version dependant.<\/p>\n\n<p>For example, the function (taken from the official guide):<\/p>\n\n<pre><code>def square_if_positive(x):\n  if x &gt; 0:\n    x = x * x\n  else:\n    x = 0.0\n  return x\n<\/code><\/pre>\n\n<p>Can be evaluated in graph mode using:<\/p>\n\n<ul>\n<li><code>autograph.to_graph<\/code> in TF 1.14<\/li>\n<\/ul>\n\n<pre><code>tf_square_if_positive = autograph.to_graph(square_if_positive)\n\nwith tf.Graph().as_default():\n  g_out = tf_square_if_positive(tf.constant( 9.0))\n  with tf.Session() as sess:\n    print(sess.run(g_out))\n<\/code><\/pre>\n\n<ul>\n<li><code>tf.function<\/code> in TF2.0<\/li>\n<\/ul>\n\n<pre><code>@tf.function\ndef square_if_positive(x):\n  if x &gt; 0:\n    x = x * x\n  else:\n    x = 0.0\n  return x\n\nsquare_if_positive(tf.constant( 9.0))\n<\/code><\/pre>\n\n<p>So:<\/p>\n\n<ul>\n<li>What is the relationship between <code>tf.function<\/code> and <code>autograph.to_graph<\/code>? One can assumes <code>tf.function<\/code> is using <code>autograph.to_graph<\/code> (as well as <code>autograph.to_code<\/code>) internally, but this is far from obvious.<\/li>\n<li>Is the <code>autograph.to_graph<\/code> snippet still supported in TF2.0 (since it requires <code>tf.Session<\/code>)? It is present in the <a href=\"https:\/\/www.tensorflow.org\/guide\/autograph#automatically_convert_python_control_flow\" rel=\"nofollow noreferrer\">autograph doc<\/a> in TF 1.14, but not in the corresponding doc of TF 2.0<\/li>\n<\/ul>\n","answers":[{"AnswerId":"57528083","CreationDate":"2019-08-16T16:24:40.323","ParentId":null,"OwnerUserId":"2891324","Title":null,"Body":"<p>I covered and answered all your questions in a three-part article: \"Analyzing tf.function to discover AutoGraph strengths and subtleties\": <a href=\"https:\/\/pgaleone.eu\/tensorflow\/tf.function\/2019\/03\/21\/dissecting-tf-function-part-1\/\" rel=\"nofollow noreferrer\">part 1<\/a>, <a href=\"https:\/\/pgaleone.eu\/tensorflow\/tf.function\/2019\/04\/03\/dissecting-tf-function-part-2\/\" rel=\"nofollow noreferrer\">part 2<\/a>, <a href=\"https:\/\/pgaleone.eu\/tensorflow\/tf.function\/2019\/05\/10\/dissecting-tf-function-part-3\/\" rel=\"nofollow noreferrer\">part 3<\/a>.<\/p>\n\n<p>To summarize and answer your 3 questions:<\/p>\n\n<ul>\n<li>What is the relationship between <code>tf.function<\/code> and <code>autograph.to_graph<\/code>?<\/li>\n<\/ul>\n\n<p><code>tf.function<\/code> uses AutoGraph by default. What happens when you <strong>invoke<\/strong> the first time a tf.function-decorated function is that:<\/p>\n\n<ol>\n<li>The function body is executed (in TensorFlow 1.x like, thus without eager mode) and its execution is traced (now tf.function knows which nodes are present, which branch of the <code>if<\/code> to keep and so on)<\/li>\n<li>At the same time, AutoGraph starts and tries to convert to <code>tf.*<\/code> calls, the Python statements it knows (<code>while<\/code> -> <code>tf.while<\/code>, <code>if<\/code> -> <code>tf.cond<\/code>, ...)-.<\/li>\n<\/ol>\n\n<p>Merging the information from points 1 and 2 a new graph is built, and based on the function name and the type of the parameters it is cached in a map (see the articles for a better understanding).<\/p>\n\n<ul>\n<li>Is the autograph.to_graph snippet still supported in TF2.0?<\/li>\n<\/ul>\n\n<p>Yes, <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/autograph\/to_graph\" rel=\"nofollow noreferrer\"><code>tf.autograph.to_graph<\/code><\/a> is still present and it creates a session internally for you (in TF2 you don't have to worry about them).<\/p>\n\n<p>At any rate, I suggest you read the three articles linked since they cover in detail this and other peculiarities of <code>tf.function<\/code>.<\/p>\n"}]}
+{"QuestionId":57526302,"AnswerCount":0,"Tags":"<python><keras><scikit-learn><nlp><pytorch>","CreationDate":"2019-08-16T14:20:19.060","AcceptedAnswerId":null,"OwnerUserId":4337522.0,"Title":"Mapping predictions of partially overlapping batches","Body":"<p>I'm working with pytorch_pretrained_bert for NER based on the implementation available here: <a href=\"https:\/\/github.com\/kamalkraj\/BERT-NER\" rel=\"nofollow noreferrer\">https:\/\/github.com\/kamalkraj\/BERT-NER<\/a><\/p>\n\n<p>When predicting on longer texts, the <code>predict<\/code> function in <code>bert.py<\/code> in there returns an out of index error. My current understanding is that this is due to the nr of bert tokens (which are not necessarily the same, and indeed in many cases are not the same as the number of tokens for ex. <code>nltk.word_tokenize()<\/code> returns) makes out of the input text. If this exceeds 512 (as this is the default length for the pre-trained models, if I understood correctly), it returns this error (and works fine for shorter input).\nSo, I decided to implement a wrapper that, in case I have more bert tokens (i.e. the stuff bert's own tokenizer returns) than 512, generates batches that partially overlap. The overlap is desired, because just cutting it up into pieces, padding the last one (as it is unlikely that my nr of tokens is divisible by 512, hence the last, shorter batch would require padding) loses context. It means, however, that I have to do predictions batch-wise, and then put things back together afterwards.<\/p>\n\n<p>This cutting up into batches (of length 512) is straightforward enough. I'm currently working on the putting back together predictions part.\nFor this, I wrote the following code:<\/p>\n\n<pre><code>def map_batches(batches, blen):\n    master = []\n    for i in range(len(batches)-1):\n        appendbatch = []\n        # if first round, append first bit of current batch with only one prediction\n        if i == 0:\n            appendbatch = batches[0][:int(blen\/2)-1]\n        thisbatch = batches[i]\n        nextbatch = batches[i+1]\n        offset = int(blen\/2)-1 # think -1 is needed for zero offsetting, but not sure, test\/debug!\n        for j in range(offset, blen-offset+2): # and debug +2 here too!\n            thistuple = thisbatch[j]\n            nexttuple = nextbatch[j-offset]\n            if thistuple[1] &gt; nexttuple[1]:\n                thisbatch[j] = thistuple\n            else:\n                thisbatch[j] = nexttuple\n        appendbatch.extend(thisbatch[int(blen\/2)-1:])\n        # and at end add last bit of last batch with only one prediction\n        if i == len(batches)-2:\n            appendbatch.extend(nextbatch[int(blen\/2)+1:])\n        master.append(appendbatch)\n    return master\n\nblen = 6\na = [('I-LOC', 0.1),\n     ('O', 0.2),\n     ('I-LOC', 0.5),\n     ('O', 0.2),\n     ('I-LOC', 0.9),\n     ('O', 0.1)]\n\nb = [('B-LOC', 0.6),\n     ('O', 0.7),\n     ('I-LOC', 0.8),\n     ('O', 0.9),\n     ('O', 0.99),\n     ('O', 0.99)]\ncorrect = [[('I-LOC', 0.1), ('O', 0.2), ('B-LOC', 0.6), ('O', 0.7), ('I-LOC', 0.9), ('O', 0.9), ('O', 0.99), ('O', 0.99)]]\nbatches = [a, b]\nmaster = map_batches(batches, blen)\nprint('mapped: ', master)\nprint('correct:', correct)\n<\/code><\/pre>\n\n<p>The function <code>map_batches<\/code> should do the trick. The input lists overlap only partially; for the first 2 items in list <code>a<\/code>, I have only one prediction (that of a itself). Then for the next 6 tokens, I have two predictions, corresponding to <code>a[2:]<\/code> and <code>b[:6]<\/code>. And for the last 2, I have only b again. The function is supposed to put this together, based on which has the highest likelihood (made up numbers here, in the real code these are of course confidence scores of the label).<\/p>\n\n<p>Checking this for my toy examples a and b, it works, i.e. the output of the above is:<\/p>\n\n<blockquote>\n  <p>mapped:  [[('I-LOC', 0.1), ('O', 0.2), ('B-LOC', 0.6), ('O', 0.7), ('I-LOC', 0.9), ('O', 0.9), ('O', 0.99), ('O', 0.99)]]<\/p>\n  \n  <p>correct: [[('I-LOC', 0.1), ('O', 0.2), ('B-LOC', 0.6), ('O', 0.7), ('I-LOC', 0.9), ('O', 0.9), ('O', 0.99), ('O', 0.99)]]<\/p>\n<\/blockquote>\n\n<p>Sorry for the long intro.<\/p>\n\n<p>Now here's the question(s):<\/p>\n\n<p>First of all, this is put together rather quickly. While it works for my toy example, I'm a bit anxious that it may not be correct (hence the comments for debugging\/testing) for all cases (different batch sizes, etc.), and I plan to debug it more, obviously. Still, any thought or input on this problem would be much welcomed!<\/p>\n\n<p>Second, and I guess the real question: this is a problem that I suspect must have been solved by someone long before already (and probably in a much better and more efficient way). Would anyone know any implementations for this in keras, tensorflow, numpy, sklearn or any of the other usual suspects?<\/p>\n","answers":[]}
+{"QuestionId":57526494,"AnswerCount":0,"Tags":"<python><tensorflow><keras><neural-network><deep-learning>","CreationDate":"2019-08-16T14:33:14.740","AcceptedAnswerId":null,"OwnerUserId":3271388.0,"Title":"Why reason my Keras cnn model not increase accuracy?","Body":"<p>I'm trying train a CNN Model based on VGG16 with Keras, and i can't get an accuracy greater than 40%~50%. I've augmentated the data, using dropout layers, l2 regularization, weight initializer and many parameters to find best results, but nothing happens.<\/p>\n\n<p>I tried run my model on Google Colab, it's not a huge Dataset, just 3 folders from PASCAL VOC 2012 challenge, you can see my code above:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#imports ...\n\n#Path to train dir ...\n\n#Model\ninit = keras.initializers.RandomUniform(minval=0, maxval=0.001)\ninput_shape = (224,224,3)\nimg_input = layers.Input(shape=input_shape)\nx = layers.Conv2D(64, (3, 3),\n                  activation='relu',\n                  padding='valid',\n                  kernel_initializer=init,\n                  kernel_regularizer=regularizers.l2(0.00001),\n                  name='block1_conv1')(img_input)\nx = layers.Conv2D(64, (3, 3),\n                  activation='relu',\n                  padding='valid',\n                  kernel_initializer=init,\n                  kernel_regularizer=regularizers.l2(0.00001),\n                  name='block1_conv2')(x)\nx = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block1_pool')(x)\n\n# Block 2\nx = layers.Conv2D(128, (3, 3),\n                  activation='relu',\n                  padding='valid',\n                  kernel_initializer=init,\n                  kernel_regularizer=regularizers.l2(0.00001),\n                  name='block2_conv1')(x)\nx = layers.Conv2D(128, (3, 3),\n                  activation='relu',\n                  padding='valid',\n                  kernel_initializer=init,\n                  kernel_regularizer=regularizers.l2(0.00001),\n                  name='block2_conv2')(x)\nx = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block2_pool')(x)\n\n# Block 3\nx = layers.Conv2D(256, (3, 3),\n                  activation='relu',\n                  padding='valid',\n                  kernel_initializer=init,\n                  kernel_regularizer=regularizers.l2(0.00001),\n                  name='block3_conv1')(x)\nx = layers.Conv2D(256, (3, 3),\n                  activation='relu',\n                  padding='valid',\n                  kernel_initializer=init,\n                  kernel_regularizer=regularizers.l2(0.00001),\n                  name='block3_conv2')(x)\nx = layers.Conv2D(256, (3, 3),\n                  activation='relu',\n                  padding='valid',\n                  kernel_initializer=init,\n                  kernel_regularizer=regularizers.l2(0.00001),\n                  name='block3_conv3')(x)\nx = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block3_pool')(x)\n\n# Block 4\nx = layers.Conv2D(512, (3, 3),\n                  activation='relu',\n                  padding='valid',\n                  kernel_initializer=init,\n                  kernel_regularizer=regularizers.l2(0.00001),\n                  name='block4_conv1')(x)\nx = layers.Conv2D(512, (3, 3),\n                  activation='relu',\n                  padding='valid',\n                  kernel_initializer=init,\n                  kernel_regularizer=regularizers.l2(0.00001),\n                  name='block4_conv2')(x)\nx = layers.Conv2D(512, (3, 3),\n                  activation='relu',\n                  padding='valid',\n                  kernel_initializer=init,\n                  kernel_regularizer=regularizers.l2(0.00001),\n                  name='block4_conv3')(x)\nx = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block4_pool')(x)\n\n# Block 5\nx = layers.Conv2D(512, (3, 3),\n                  activation='relu',\n                  padding='valid',\n                  kernel_initializer=init,\n                  kernel_regularizer=regularizers.l2(0.00001),\n                  name='block5_conv1')(x)\nx = layers.Conv2D(512, (3, 3),\n                  activation='relu',\n                  padding='valid',\n                  kernel_initializer=init,\n                  kernel_regularizer=regularizers.l2(0.00001),\n                  name='block5_conv2')(x)\nx = layers.Conv2D(10, (3, 3),\n                  activation='relu',\n                  padding='valid',\n                  kernel_initializer=init,\n                  kernel_regularizer=regularizers.l2(0.00001),\n                  name='block5_conv3')(x)\nx = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block5_pool')(x)\n\nx = layers.Flatten(name='flatten')(x)\nx = layers.Dense(4096, activation='relu', name='fc1', kernel_regularizer=regularizers.l2(0.00001), kernel_initializer=init)(x)\nx = layers.Dropout(0.5)(x)\nx = layers.Dense(4096, activation='relu', name='fc2', kernel_regularizer=regularizers.l2(0.00001), kernel_initializer=init)(x)\nx = layers.Dropout(0.5)(x)\nx = layers.Dense(3, activation='softmax', name='predictions')(x)\n\nmodel = Model(img_input, x, name='overfeat_vgg16')\n#model.summary()\n\n#creating generators\ntrain_datagen = ImageDataGenerator(rescale=1.\/255, zoom_range=0.4, horizontal_flip=True)\ntest_datagen = ImageDataGenerator(rescale=1.\/255)\ntrain_generator = train_datagen.flow_from_directory(train_dir, classes=[\"bird\",\"cat\",\"dog\"], shuffle=True, target_size=(224, 224), batch_size=64, class_mode='categorical')\nvalidation_generator = test_datagen.flow_from_directory(train_dir, classes=[\"bird\",\"cat\",\"dog\"], shuffle=True, target_size=(224, 224), batch_size=64, class_mode='categorical')\n\n#compiling model\nsgd = keras.optimizers.SGD(lr=0.01, momentum=0.9, decay=0.0005, nesterov=True)\n#adam = keras.optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False)\nmodel.compile(optimizer=sgd, loss='categorical_crossentropy', metrics=['accuracy'])\n\n#training\nmodel.fit_generator(\n        train_generator,\n        steps_per_epoch=5850 \/\/ 64,\n        epochs=100,\n        shuffle = True,\n        validation_data=validation_generator,\n        validation_steps=5850 \/\/ 64)\nmodel.save_weights('first_try.h5')\n<\/code><\/pre>\n\n<p>There are the current results:<\/p>\n\n<p>Epoch 1\/100\n91\/91 [==============================] - 172s 2s\/step - loss: 1.0449 - acc: 0.4301 - val_loss: 1.0432 - val_acc: 0.4239<\/p>\n\n<p>Epoch 2\/100\n91\/91 [==============================] - 155s 2s\/step - loss: 1.0421 - acc: 0.4188 - val_loss: 1.0411 - val_acc: 0.3984<\/p>\n\n<p>Epoch 3\/100\n91\/91 [==============================] - 155s 2s\/step - loss: 1.0388 - acc: 0.4263 - val_loss: 1.0382 - val_acc: 0.4250<\/p>\n\n<p>... and for all epochs the results keep the same.<\/p>\n\n<p>I just don't know what i need to do for better results.<\/p>\n","answers":[]}
+{"QuestionId":57526563,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow2.0>","CreationDate":"2019-08-16T14:37:35.063","AcceptedAnswerId":null,"OwnerUserId":11936463.0,"Title":"TF 2.0 infer() vs. freezing a graph","Body":"<p>What is the best way to do the inference with TF models? So far it is common to use a frozen_graph.pb for inference like:<\/p>\n\n<pre><code>graph = load_graph('.\/path_to_frozen\/frozen_graph.pb')\n\nwith tf.Session(graph=graph) as sess:\n    results = sess.run(\n        output_operation.outputs[0],\n        {input_operation.outputs[0]: img})\n\n    sess.close()\n<\/code><\/pre>\n\n<p>When dealing with Tensorflow 2.0 I got to know about SavedModel and the infer() function in eager execution. It works like: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    model = tf.saved_model.load('.\/path_to_savedmodel')\n    infer = model.signatures[tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY]\n    results = infer(img)\n<\/code><\/pre>\n\n<p>I have done tests on a CPU and option (a) takes 0.47 sec to load the model and 4.81 sec to run and option (b) takes 3x less time: 1.23 sec to load and only 0.24 sec to run. <\/p>\n\n<p>The only difference I see is that in the savedmodel takes tensor as input, and session takes a numpy array. Other than that I am very surprised about the difference.<\/p>\n","answers":[]}
+{"QuestionId":57526806,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><image-processing>","CreationDate":"2019-08-16T14:55:08.443","AcceptedAnswerId":null,"OwnerUserId":3050590.0,"Title":"TypeError: Fetch argument None has invalid type <class 'NoneType'> at tf.reduce_max(pred, axis=1).eval()","Body":"<p>In my code below I am trying to get the highest probability for each image in an array. It is giving me the following error<\/p>\n\n<pre><code>TypeError: Fetch argument None has invalid type &lt;class 'NoneType'&gt;\n<\/code><\/pre>\n\n<p>Program code:<\/p>\n\n<pre><code>categories = #labels\ntest_img = []\nprobab=[]\nclasses=[]\ntrue_labels = np.array(categories)\nnew_model =  tf.keras.models.load_model(\"multi25.h5\", custom_objects={'KerasLayer': hub.KerasLayer})\ndef load(filename):\n   np_image = Image.open(filename)\n   np_image = np.array(np_image).astype('float32')\/255\n   np_image = transform.resize(np_image, (224, 224, 3))\n   np_image = np.expand_dims(np_image, axis=0)\n   return np_image\n\nfor i in os.listdir(data_root):\n        print(i)\n        loc= data_root+\"\/\/\"+i\n        test_img.append(load(loc))\n        pred= new_model.predict(test_img)\n        classes.append(true_labels[np.argmax(pred, axis=-1)])\n        with tf.Session() as sess:\n            probab=tf.reduce_max(pred,  axis=1).eval()\n            sess.run(print(probab))\nprint(test_img+\" \"+classes+\" \"+probab)\n<\/code><\/pre>\n\n<p>I am getting error in the following code, where I am using reduce_max function to get the max probability.<\/p>\n\n<pre><code>probab=tf.reduce_max(pred,  axis=1).eval()\nsess.run(print(probab))\n<\/code><\/pre>\n\n<p>I have already tried other solutions, wherein I changed the function to reduce_indices, tried without session run, etc. Pred variable has no nonetype. I am not able to understand the error very ell<\/p>\n\n<p>Here is the traceback<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"&lt;input&gt;\", line 1, in &lt;module&gt;\n  File \"C:\\Program Files\\JetBrains\\PyCharm 2019.2\\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 \"C:\\Program Files\\JetBrains\\PyCharm 2019.2\\helpers\\pydev\\_pydev_imps\\_pydev_execfile.py\", line 18, in execfile\n    exec(compile(contents+\"\\n\", file, 'exec'), glob, loc)\n  File \"C:\/Users\/Pawandeep\/Desktop\/Python projects\/multimodel_pred.py\", line 52, in &lt;module&gt;\n    sess.run(print(probab))\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 950, in run\n    run_metadata_ptr)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1158, in _run\n    self._graph, fetches, feed_dict_tensor, feed_handles=feed_handles)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 474, in __init__\n    self._fetch_mapper = _FetchMapper.for_fetch(fetches)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 261, in for_fetch\n    type(fetch)))\nTypeError: Fetch argument None has invalid type &lt;class 'NoneType'&gt;\n<\/code><\/pre>\n","answers":[{"AnswerId":"57527450","CreationDate":"2019-08-16T15:38:15.743","ParentId":null,"OwnerUserId":"3050590","Title":null,"Body":"<p>Issue resolved by using <code>tf.InteractiveSession<\/code> instead of <code>session.run()<\/code>. Reference is <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/InteractiveSession\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"}]}
+{"QuestionId":57526932,"AnswerCount":1,"Tags":"<tensorflow><keras><neural-network><conv-neural-network><tf.keras>","CreationDate":"2019-08-16T15:02:20.957","AcceptedAnswerId":57527509.0,"OwnerUserId":11541912.0,"Title":"What is the difference between conv1d with kernel_size=1 and dense layer?","Body":"<p>I am building a CNN with Conv1D layers, and it trains pretty well. I'm now looking into how to reduce the number of features before feeding it into a Dense layer at the end of the model, so I've been reducing the size of the Dense layer, but then I came across <a href=\"https:\/\/machinelearningmastery.com\/introduction-to-1x1-convolutions-to-reduce-the-complexity-of-convolutional-neural-networks\/\" rel=\"nofollow noreferrer\">this article<\/a>. The article talks about the effect of using a Conv2D filters with a kernel_size=(1,1) to reduce the number of features. <\/p>\n\n<p>I was wondering what the difference is between using a Conv2D layer with kernel_size=(1,1) <code>tf.keras.layers.Conv2D(filters=n,kernel_size=(1,1))<\/code> and using a Dense layer of the same size <code>tf.keras.layers.Dense(units=n)<\/code>? From my perspective (I'm relatively new to neural nets), a filter with kernel_size=(1,1) is a single number, which is essentially equivalent to weight in a Dense layer, and both layers have biases, so are they equivalent, or am I misunderstanding something? And if my understanding is correct, in my case where I am using Conv1D layers, not Conv2D layers, does that change anything? As in is <code>tf.keras.layers.Conv1D(filters=n, kernel_size=1)<\/code> equivalent to <code>tf.keras.layers.Dense(units=n)<\/code>?<\/p>\n\n<p>Please let me know if you need anything from me to clarify the question. I'm mostly curious about if Conv1D layers with kernel_size=1 and Conv2D layers with kernel_size=(1,1) behave differently than Dense layers.<\/p>\n","answers":[{"AnswerId":"57527509","CreationDate":"2019-08-16T15:42:16.237","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Yes, since <code>Dense<\/code> layer is applied on the last dimension of its input (see <a href=\"https:\/\/stackoverflow.com\/a\/52092176\/2099607\">this answer<\/a>), <code>Dense(units=N)<\/code> and <code>Conv1D(filters=N, kernel_size=1)<\/code> (or <code>Dense(units=N)<\/code> and <code>Conv2D(filters=N, kernel_size=1)<\/code>) are basically equivalent to each other both in terms of connections and number of trainable parameters.<\/p>\n"}]}
+{"QuestionId":57526980,"AnswerCount":0,"Tags":"<python><machine-learning><keras><neural-network>","CreationDate":"2019-08-16T15:05:05.073","AcceptedAnswerId":null,"OwnerUserId":9850245.0,"Title":"handle overfitting in unbalanced datasets","Body":"<p>I have an unbalanced dataset (only 0.06% of data is labeled 1 and the rest are labeled 0). As I researched, I had to resample the data, so I used <code>imblearn<\/code> package to <code>randomUnserSample<\/code> my dataset. Then I used a <code>Keras Sequential<\/code> to create a neural network. While training, <code>F1Score<\/code> increases to around 75% (1000th epoch result is: loss: 0.5691 - acc: 0.7543 - f1_m: 0.7525 - precision_m: 0.7582 - recall_m: 0.7472), but on test set, the result is disappointing (loss: 55.35181%, acc: 79.25248%, f1_m: 0.39789%, precision_m: 0.23259%, recall_m: 1.54982%).<\/p>\n\n<p>What I assume is that on train set, because the number of 1s and 0s are the same and therefore class_wights are both set to 1, so the network is not costing much for wrong predictions of 1s.<\/p>\n\n<p>I have used some techniques like reducing the number of layers, reducing number of neurons, using regularization and dropouts, but the test set <code>f1Score<\/code> is never more than 0.5%. What should I do. Thanks<\/p>\n\n<p>my neural network:<\/p>\n\n<pre><code>def neural_network(X, y, epochs_count=3, handle_overfit=False):\n    # create model\n    model = Sequential()\n    model.add(Dense(12, input_dim=len(X_test.columns), activation='relu'))\n    if (handle_overfit):\n        model.add(Dropout(rate = 0.5))\n    model.add(Dense(8, activation='relu', kernel_regularizer=regularizers.l1(0.1)))\n    if (handle_overfit):\n        model.add(Dropout(rate = 0.1))\n    model.add(Dense(1, activation='sigmoid'))\n\n    # compile the model\n    model.compile(optimizer='adam',\n                  loss='binary_crossentropy',\n                  metrics=['acc', f1_m, precision_m, recall_m])\n\n#     change weights of the classes '0' and '1' and set weights automatically\n    class_weights = class_weight.compute_class_weight('balanced', [0, 1], y)\n    print(\"---------------------- \\n chosen class_wieghts are: \", class_weights, \" \\n ---------------------\")\n\n    # Fit the model\n    model.fit(X, y, epochs=epochs_count, batch_size=512, class_weight=class_weights)\n\n    return model\n<\/code><\/pre>\n\n<p>defining train and test set:<\/p>\n\n<pre><code>vtrain_set, test_set = train_test_split(data, test_size=0.35, random_state=0)\n\nX_train = train_set[['..... some columns ....']]\ny_train = train_set[['success']]\n\nprint('Initial dataset shape: ', X_train.shape)\nrus = RandomUnderSampler(random_state=42)\nX_undersampled, y_undersampled = rus.fit_sample(X_train, y_train) \nprint('undersampled dataset shape: ', X_undersampled.shape)\n<\/code><\/pre>\n\n<p>and result is:<\/p>\n\n<pre><code>Initial dataset shape:  (1625843, 11)\nundersampled dataset shape:  (1970, 11)\n<\/code><\/pre>\n\n<p>and finally the neural network calling:<\/p>\n\n<pre><code>print (X_undersampled.shape, y_undersampled.shape)\nprint (X_test.shape, y_test.shape)\n\nmodel = neural_network(X_undersampled, y_undersampled, 1000, handle_overfit=True)\n\n# evaluate the model\nprint(\"\\n---------------\\nEvaluated on test set:\")\n\nscores = model.evaluate(X_test, y_test)\nfor i in range(len(model.metrics_names)):\n    print(\"%s: %.5f%%\" % (model.metrics_names[i], scores[i]*100))\n<\/code><\/pre>\n\n<p>and result is:<\/p>\n\n<pre><code>(1970, 11) (1970,)\n(875454, 11) (875454, 1)\n---------------------- \n chosen class_wieghts are:  [1. 1.]  \n ---------------------\nEpoch 1\/1000\n1970\/1970 [==============================] - 4s 2ms\/step - loss: 4.5034 - acc: 0.5147 - f1_m: 0.3703 - precision_m: 0.5291 - recall_m: 0.2859\n\n.\n.\n.\n.\nEpoch 999\/1000\n1970\/1970 [==============================] - 0s 6us\/step - loss: 0.5705 - acc: 0.7538 - f1_m: 0.7471 - precision_m: 0.7668 - recall_m: 0.7296\nEpoch 1000\/1000\n1970\/1970 [==============================] - 0s 6us\/step - loss: 0.5691 - acc: 0.7543 - f1_m: 0.7525 - precision_m: 0.7582 - recall_m: 0.7472\n\n---------------\nEvaluated on test set:\n875454\/875454 [==============================] - 49s 56us\/step\nloss: 55.35181%\nacc: 79.25248%\nf1_m: 0.39789%\nprecision_m: 0.23259%\nrecall_m: 1.54982%\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57527079,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><deep-learning>","CreationDate":"2019-08-16T15:12:03.593","AcceptedAnswerId":null,"OwnerUserId":9079812.0,"Title":"Unable to learn weights of a Word2Vec model","Body":"<p>I was going to implement a word embedding model - namely Word2Vec - by following <a href=\"https:\/\/www.tensorflow.org\/tutorials\/representation\/word2vec\" rel=\"nofollow noreferrer\">this TensorFlow tutorial<\/a> and adapting the code a little bit. Unfortunately, though, my model won't learn anything. I've used TensorBoard to keep track of the value of the loss function <em>and<\/em> to observe how the network's weights evolve over time. Here's what I found:<\/p>\n\n<ol>\n<li>The value of the loss function keeps fluctuating up and down<\/li>\n<li>The network's weights stay constant during the training process<\/li>\n<\/ol>\n\n<p>I honestly can't understand why this is happening. I've tried setting \"trainable=True\" explicitly when creating variables, but that didn't help either.\nHere's the code that I'm using right now:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\nvocabulary_size = 13046\nembedding_size = 256\nnum_noise = 1\nlearning_rate = 1e-3\nbatch_size = 1024\nepochs = 10\n\ndef make_hparam_string(embedding_size, num_noise, learning_rate, batch_size, epochs):\n    return f'es={embedding_size}_nn={num_noise}_lr={learning_rate}_bs={batch_size}_e={epochs}'\n\n# These are the hidden layer weights\nembeddings = tf.get_variable(name='embeddings', initializer=tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0), trainable=True)\n\n# 'nce' stands for 'Noise-contrastive estimation' and represents a particular loss function.\n# Check https:\/\/www.tensorflow.org\/tutorials\/representation\/word2vec for more details.\n# 'nce_weights' and 'nce_biases' are simply the output weights and biases.\n# NOTE: for some reason, even though output weights will have shape (embedding_size, vocabulary_size),\n#       we have to initialize them with the shape (vocabulary_size, embedding_size)\nnce_weights = tf.get_variable(name='output_weights',\n                              initializer=tf.truncated_normal([vocabulary_size, embedding_size], stddev=1.0 \/ np.sqrt(embedding_size)), \n                              trainable=True)\nnce_biases = tf.get_variable(name='output_biases', initializer=tf.constant_initializer(0.1), shape=[vocabulary_size], trainable=True)\n\n# Placeholders for inputs\ntrain_inputs = tf.placeholder(tf.int32, shape=[None])    # [batch_size]\ntrain_labels = tf.placeholder(tf.int32, shape=[None, 1]) # [batch_size, 1]\n\n# This allows us to quickly retrieve the corresponding word embeddings for each word in 'train_inputs'\nmatched_embeddings = tf.nn.embedding_lookup(embeddings, train_inputs)\n\n# Compute the NCE loss, using a sample of the negative labels each time.\nloss = tf.reduce_mean(tf.nn.nce_loss(weights=nce_weights,\n                                     biases=nce_biases,\n                                     labels=train_labels,\n                                     inputs=matched_embeddings,\n                                     num_sampled=num_noise,\n                                     num_classes=vocabulary_size))\n\n# Use the SGD optimizer to minimize the loss function\noptimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(loss)\n\n# Add some summaries for TensorBoard\nloss_summary = tf.summary.scalar('nce_loss', loss)\ninput_embeddings_summary = tf.summary.histogram('input_embeddings', embeddings)\noutput_embeddings_summary = tf.summary.histogram('output_embeddings', nce_weights)\n\n################################################################################\n\n# Load data\ntarget_words = np.genfromtxt('target_words.txt', dtype=int, delimiter='\\n').reshape((-1, 1))\ncontext_words = np.genfromtxt('context_words.txt', dtype=int, delimiter='\\n').reshape((-1, 1))\n\n# Convert to tensors\ntarget_words_tensor = tf.convert_to_tensor(target_words)\ncontext_words_tensor = tf.convert_to_tensor(context_words)\n\n# Create a tf.data.Dataset object representing our dataset\ndataset = tf.data.Dataset.from_tensor_slices((target_words_tensor, context_words_tensor))\ndataset = dataset.shuffle(buffer_size=target_words.shape[0])\ndataset = dataset.batch(batch_size)\n\n# Create an iterator to iterate over the dataset\niterator = dataset.make_initializable_iterator()\nnext_batch = iterator.get_next()\n\n# Train the model\nwith tf.Session() as session:\n\n    # Initialize variables\n    session.run( tf.global_variables_initializer() )\n\n    merged_summary = tf.summary.merge_all()\n\n    # File writer for TensorBoard\n    hparam_string = make_hparam_string(embedding_size, num_noise, learning_rate, batch_size, epochs)\n    loss_writer = tf.summary.FileWriter(f'.\/tensorboard\/{hparam_string}')\n\n    global_step = 0\n    for epoch in range(epochs):\n\n        session.run(iterator.initializer)\n        while True:\n            try:\n                inputs, labels = session.run(next_batch)\n\n                feed_dict = {train_inputs: inputs[:, 0], train_labels: labels}\n                _, cur_loss, all_summaries = session.run([optimizer, loss, merged_summary], feed_dict=feed_dict)\n\n                # Write sumaries to disk\n                loss_writer.add_summary(all_summaries, global_step=global_step)\n                global_step += 1\n\n                print(f'Current loss: {cur_loss}')\n\n            except tf.errors.OutOfRangeError:\n                print(f'Finished epoch {epoch}.')\n                break\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57527493,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-08-16T15:41:29.317","AcceptedAnswerId":null,"OwnerUserId":7697890.0,"Title":"Creating a custom generator for fit_generator","Body":"<p>I'm trying to create a custom generator function for an aesthetic image classifier. My actual code inside the generator function is:<\/p>\n\n<pre><code>with tf.compat.v1.Session() as sess:            \n            if (train):       \n                dataset_paths_placeholder = tf.compat.v1.placeholder(train_image_paths.dtype, train_image_paths.shape)\n                dataset_scores_placeholder = tf.compat.v1.placeholder(train_scores.dtype, train_scores.shape)\n            else:\n                dataset_paths_placeholder = tf.compat.v1.placeholder(val_image_paths.dtype, val_image_paths.shape)\n                dataset_scores_placeholder = tf.compat.v1.placeholder(val_scores.dtype, val_scores.shape)\n\n            dataset = tf.compat.v1.data.Dataset.from_tensor_slices((dataset_paths_placeholder, dataset_scores_placeholder))\n\n            #Here I parse the dataset, basically the path is substituted with the actual image\n            if(train):\n                dataset = dataset.map(parse_with_augmentation, num_parallel_calls = tf.data.experimental.AUTOTUNE)\n            else:\n                dataset = dataset.map(parse_without_augmentation, num_parallel_calls = tf.data.experimental.AUTOTUNE)\n            dataset = dataset.batch(batchsize)\n            dataset = dataset.repeat(count = -1)\n            if (shuffle):\n                dataset = dataset.shuffle(buffer_size=10)  \n<\/code><\/pre>\n\n<p>My initial plan was to create an iterator on the dataset, then inside a loop to retrieve the batches of X,y and yield them.\nThe problem is that to use <code>make_initializable_iterator()<\/code> on the dataset is signaled as deprecated and it is suggested to retrieve the elements with a <code>for<\/code> loop.<\/p>\n\n<p>Is there a simple way to retrieve all batches from my dataset and yield them like this?<\/p>\n","answers":[]}
+{"QuestionId":57527609,"AnswerCount":0,"Tags":"<python><tensorflow><tensorboard><tensorflow-datasets><tensorflow2.0>","CreationDate":"2019-08-16T15:49:36.327","AcceptedAnswerId":null,"OwnerUserId":3860928.0,"Title":"How to show in tensorboard the tf.data.Dataset.map subgraph in Tensorflow 2.0?","Body":"<p>According to the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset#map\" rel=\"nofollow noreferrer\">documentation<\/a>, the <code>tf.data.Datasets<\/code> work in graph mode (in both eager and graph mode):<\/p>\n\n<blockquote>\n  <p>Note that irrespective of the context in which map_func is defined (eager vs. graph), tf.data traces the function and executes it as a graph<\/p>\n<\/blockquote>\n\n<p>In Tensorflow 1.X, we can easily plot this graph in Tensorboard: The processing functions are plotted in a subgraph.<\/p>\n\n<p>For example,<\/p>\n\n<pre><code>def _parse_function(x):\n    return x * 2\n\nx = tf.constant([0 , 1])\ndataset = tf.data.Dataset.from_tensor_slices(x)\ndataset = dataset.map(_parse_function)\n<\/code><\/pre>\n\n<p>In Tensorboard, a subgraph appears, corresponding to the _parse_function:\n<a href=\"https:\/\/i.stack.imgur.com\/ykIRx.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ykIRx.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>However, in Tensorflow 2.0, this does not produce any visible element in the Tensorboard graph.\nThe following code does not produce any graph according to Tensorboard:<\/p>\n\n<pre><code>def _parse_function(x):\n    return x * 2\n\nlogdir = 'logs'\nwriter = tf.summary.create_file_writer(logdir)\n\ntf.summary.trace_on(graph=True, profiler=True)\nx = tf.constant([0 , 1])\ndataset = tf.data.Dataset.from_tensor_slices(x)\ndataset = dataset.map(_parse_function)\n\nwith writer.as_default():\n  tf.summary.trace_export(\n      name=\"trace\",\n      step=0,\n      profiler_outdir=logdir)\n<\/code><\/pre>\n\n<p>So, since a graph is created when calling the <code>map<\/code>, is there a way to access\/visualize this graph?<\/p>\n","answers":[]}
+{"QuestionId":57527610,"AnswerCount":0,"Tags":"<python><machine-learning><keras><lstm><autoencoder>","CreationDate":"2019-08-16T15:49:36.660","AcceptedAnswerId":null,"OwnerUserId":11149982.0,"Title":"Tuning LSTM autoencoder performance","Body":"<p>I am trying to build an autoencoder of a multidimensional time series. I have followed various templates around the internet and SO, but all of them focus on how to get it running, but haven't found one on how to get running and get some meaningful results.<\/p>\n\n<p>I've followed the tutorials starting here: <a href=\"https:\/\/blog.keras.io\/building-autoencoders-in-keras.html\" rel=\"nofollow noreferrer\">https:\/\/blog.keras.io\/building-autoencoders-in-keras.html<\/a>; a practical example here: <a href=\"https:\/\/machinelearningmastery.com\/lstm-autoencoders\/\" rel=\"nofollow noreferrer\">https:\/\/machinelearningmastery.com\/lstm-autoencoders\/<\/a>.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nfrom keras.models import Sequential\nfrom keras.layers import LSTM, RepeatVector\nimport matplotlib.pyplot as plt\n\n# this sequence comes out of a MinMaxScaler. A separate question is if this was a good idea?\nsequence = np.array([[0.63306452, 0.00714286],\n       [0.42069892, 0.        ],\n       [0.36155914, 0.15      ],\n       [0.53629032, 0.12142857],\n       [0.32526882, 0.24285714],\n       [0.26344086, 0.52142857],\n       [0.        , 0.79285714],\n       [0.49731183, 0.71428571],\n       [0.60080645, 0.25714286],\n       [0.63037634, 0.11428571],\n       [0.70698925, 0.26428571],\n       [0.71774194, 0.21428571],\n       [0.6155914 , 0.10714286],\n       [0.56451613, 0.36428571],\n       [0.66397849, 0.2       ],\n       [0.76344086, 0.17857143],\n       [0.66801075, 0.07142857],\n       [0.66935484, 0.02857143],\n       [0.90725806, 0.32857143],\n       [1.        , 0.28571429],\n       [1.        , 0.4       ],\n       [0.81451613, 0.47857143],\n       [0.41532258, 0.52142857],\n       [0.55107527, 0.63571429],\n       [0.42741935, 0.40714286],\n       [0.56989247, 0.75      ],\n       [0.76075269, 0.55      ],\n       [0.69758065, 0.58571429],\n       [0.73521505, 0.89285714],\n       [0.77150538, 1.        ]])\n\nn_in = len(sequence)\ndim_in = sequence.shape[1]\nlatent_dim = 10\nsequence = sequence.reshape((1, n_in, dim_in))\n\nmodel = Sequential()\nmodel.add(LSTM(latent_dim, input_shape=(n_in, dim_in)))\nmodel.add(RepeatVector(n_in))\nmodel.add(LSTM(dim_in, return_sequences=True))\nmodel.compile(optimizer='adam', loss='mse')\nmodel.summary()\n\nmodel.fit(sequence, sequence, epochs=1000, verbose=0)\n\nyhat = model.predict(sequence, verbose=0)\n\nplt.figure(1)\nplt.subplot(221)\nplt.plot(sequence[0, :, 0])\nplt.subplot(223)\nplt.plot(yhat[0, :, 0])\nplt.subplot(222)\nplt.plot(sequence[0, :, 1])\nplt.subplot(224)\nplt.plot(yhat[0, :, 1])\n\n<\/code><\/pre>\n\n<p>The result I'm getting is not satisfactory (actuals in the upper row; autoencoder output in the lower row):\n<img src=\"https:\/\/i.stack.imgur.com\/hV6SN.png\" alt=\"Actuals (upper row); autoencoder output (lower row)\"><\/p>\n\n<p>The decoded series are missing important features (like the spike in the RHS or drop on the LHS. Given the 'compression ratio' of 30:10 I would expect those events to be somehow reflected. I've tried playing with epochs, batch sizes, various activations and losses.<\/p>\n\n<ol>\n<li>Anything obvious I am missing?<\/li>\n<li>I want to run it on a much larger sequence (5000 time points, each point of potentially high dimension). Any tips for this?<\/li>\n<li>should I change my approach altogether? The Author of this blog post <a href=\"https:\/\/towardsdatascience.com\/autoencoders-for-the-compression-of-stock-market-data-28e8c1a2da3e\" rel=\"nofollow noreferrer\">https:\/\/towardsdatascience.com\/autoencoders-for-the-compression-of-stock-market-data-28e8c1a2da3e<\/a> didn't make it work with LSTM as well...<\/li>\n<\/ol>\n","answers":[]}
+{"QuestionId":57527907,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><yolo>","CreationDate":"2019-08-16T16:11:48.243","AcceptedAnswerId":null,"OwnerUserId":11734427.0,"Title":"How to make an annotation file for training this implementation of Tiny-YOLO?","Body":"<p>So I need to retrain Tiny YOLO using my own dataset. I have the pictures (with and without the bounding boxes on it) and for every picture I have a JSON file. I am using FLIR dataset if that helps people know what I need to do. I will provide an example of JSON file below as well.<\/p>\n\n<p>I forgot to add that I am using Yolo in Keras. The model I am using can be found here: <a href=\"https:\/\/github.com\/qqwweee\/keras-yolo3\" rel=\"nofollow noreferrer\">keras-yolo3\n<\/a>.<\/p>\n\n<p>From what I have how can I make this annotation file for training this implementation of Tiny YOLO?<\/p>\n\n<pre><code>{\n    \"annotation\": [\n        {\n            \"area\": 1140,\n            \"bbox\": [\n                44,\n                235,\n                38,\n                30\n            ],\n            \"category_id\": \"3\",\n            \"id\": 1,\n            \"image_id\": 12,\n            \"iscrowd\": 0,\n            \"segmentation\": [\n                [\n                    44,\n                    235,\n                    44,\n                    265,\n                    82,\n                    265,\n                    82,\n                    235\n                ]\n            ]\n        }\n    ],\n    \"image\": {\n        \"file_name\": \"FLIR_00012\",\n        \"height\": 512,\n        \"id\": 12,\n        \"width\": 640\n    }\n}\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57528027,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2019-08-16T16:21:04.793","AcceptedAnswerId":null,"OwnerUserId":5659025.0,"Title":"ImportError: Could not find 'cudart64_100.dll","Body":"<p>I'm trying to install <code>tensorflow-gpu==2.0.0-beta1<\/code> on my Windows 10 machine and got this error:<\/p>\n\n<blockquote>\n  <p>ImportError: Could not find 'cudart64_100.dll'. TensorFlow requires\n  that this DLL be installed in a directory that is named in your %PATH%\n  environment variable. Download and install CUDA 10.0 from this URL:\n  <a href=\"https:\/\/developer.nvidia.com\/cuda-90-download-archive\" rel=\"nofollow noreferrer\">https:\/\/developer.nvidia.com\/cuda-90-download-archive<\/a><\/p>\n<\/blockquote>\n\n<p>I made all things from:<\/p>\n\n<ul>\n<li>official documentation: <a href=\"https:\/\/www.tensorflow.org\/install\/gpu\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/gpu<\/a><\/li>\n<li>from here: <a href=\"https:\/\/medium.com\/@teavanist\/install-tensorflow-gpu-on-windows-10-5a23c46bdbc7\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@teavanist\/install-tensorflow-gpu-on-windows-10-5a23c46bdbc7<\/a><\/li>\n<li>Checked <code>PATH<\/code> variable: <code>C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v10.0\\bin<\/code><\/li>\n<li>also have <code>CUDA_PATH<\/code> : <code>C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v10.0<\/code> in variables<\/li>\n<li>file <code>C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v10.0\\bin\\cudart64_100.dll<\/code> exists<\/li>\n<li>Did system restart<\/li>\n<\/ul>\n\n<p>But error still occurs<\/p>\n\n<p>How can I fix this?<\/p>\n","answers":[{"AnswerId":"58464975","CreationDate":"2019-10-19T15:03:25.377","ParentId":null,"OwnerUserId":"998415","Title":null,"Body":"<p>The simplest way to fix is to install the latest \u2018NVIDIA GPU Computing Toolkit\u2019, because if it's not there, you'll be missing the 'cudart64_100.dll' library.\nThe only issue is that the latest copy of CUDA has this particular library upgraded to 'cudart64_101.dll', while the latest TensorFlow still requires the older 'cudart64_100.dll'.\nAnyways, one way to deal with this issue is to install the latest CUDA + CUDA from September 2018 and then copy 'cudart64_100.dll' library from old install to the new one.<\/p>\n\n<p>Or just visit my site where I linked the 'cudart64_100.dll' library downloaded from the CUDA Toolkit 10.0 (Sept 2018), to make it easier to copy it into the latest CUDA directory.<\/p>\n\n<p>Here are some screenshots to illustrate the process: <a href=\"https:\/\/www.joe0.com\/2019\/10\/19\/how-resolve-tensorflow-2-0-error-could-not-load-dynamic-library-cudart64_100-dll-dlerror-cudart64_100-dll-not-found\/\" rel=\"nofollow noreferrer\">https:\/\/www.joe0.com\/2019\/10\/19\/how-resolve-tensorflow-2-0-error-could-not-load-dynamic-library-cudart64_100-dll-dlerror-cudart64_100-dll-not-found\/<\/a><\/p>\n"},{"AnswerId":"58533797","CreationDate":"2019-10-24T03:38:23.647","ParentId":null,"OwnerUserId":"12266518","Title":null,"Body":"<p>I have faced similar issue. I have added the directory of the cudart64_100.dll file to the PATH variable but still it prompts the error \"cudart64_100.dll\" not found. In the end I finally manage to make it work by adding the following codes. Hope it helps.<\/p>\n\n<pre><code>import ctypes\n\nhllDll = ctypes.WinDLL(\"C:\\\\Program Files\\\\NVIDIA GPU Computing Toolkit\\\\CUDA\\\\v10.0\\\\bin\\\\cudart64_100.dll\")\n<\/code><\/pre>\n"}]}
+{"QuestionId":57528611,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><deep-learning><lstm>","CreationDate":"2019-08-16T17:07:24.730","AcceptedAnswerId":null,"OwnerUserId":10115730.0,"Title":"Training Loss increases with the increase of data size and the train accuracy remains almost constant","Body":"<p>I am a newbie in ML. I am tryng to develop a LSTM model using time series data for a long time. I have tried several architectures of LSTM, but getting very large Loss and low accuracy. I  had so many null values in my data. i have removed any sequence(e.g 8 mins window) which has at least one null value. I used the code form <a href=\"https:\/\/machinelearningmastery.com\/\" rel=\"nofollow noreferrer\">https:\/\/machinelearningmastery.com\/<\/a><\/p>\n\n<pre><code>import numpy as np\nfrom numpy import array\nimport pandas as pd\nimport os\nimport matplotlib.pyplot as plt\nimport math\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.layers import LSTM\nfrom sklearn.preprocessing import MinMaxScaler\nfrom sklearn.metrics import mean_squared_error\nfrom keras.optimizers import Adam   \n<\/code><\/pre>\n\n<p>here i excluded all the sequence which has null values. after excluding sequence of any null vlaues, it will look for next sequence  <\/p>\n\n<pre><code>def split_datasetnormalized(dataset, timestamp):\n\n    train_size = int(len(dataset) * 0.67)\n    test_size = len(dataset) - train_size\n    train, test = dataset[0:train_size, :], dataset[test_size:len(dataset), :]\n\n    x = []\n    y = []\n    xp = []\n    yp = []\n    for i in range(len(train) - timestamp - 1):\n\n        if len(train[i:i +\n                     timestamp]) == timestamp and not np.isnan(train[i:i +\n                                                                     timestamp +\n                                                                     1]).any():\n            trainx = train[i:i + timestamp, 0]\n            trainy = train[i + timestamp:i + timestamp + 1, 0]\n            x.append(trainx)\n            y.append(trainy)\n    for i in range(len(test) - timestamp - 1):\n\n        if len(test[i:i +\n                    timestamp]) == timestamp and not np.isnan(test[i:i +\n                                                                   timestamp +\n                                                                   1]).any():\n            testx = test[i:i + timestamp]\n            testy = test[i + timestamp:i + timestamp + 1]\n            xp.append(testx)\n            yp.append(testy)\n    return array(x), array(y), array(xp), array(yp)\n\n\ndef lstmwindow(dfData, lags):\n    database = split_datasetnormalized(dfData, lags)\n    trainX = database[0].reshape(database[0].shape[0], 1, database[0].shape[1])\n    trainY = (database[1].reshape(1, -1))[0]\n    testX = database[2].reshape(database[2].shape[0], 1, database[2].shape[1])\n    testY = (database[3].reshape(1, -1))[0]\n    return trainX, trainY, testX, testY\n<\/code><\/pre>\n\n<p>though I have a huge file(402k), I am trying to use the first 20000 rows. If I reduce the number of rows significantly, the loss decreases and the accuracy increase. the data can be found at <a href=\"https:\/\/gofile.io\/?c=PoM9dM\" rel=\"nofollow noreferrer\">https:\/\/gofile.io\/?c=PoM9dM<\/a> <\/p>\n\n<pre><code>trainData = 'data\/train.csv'\nlook_back = 8\ndf = pd.read_csv(trainData, usecols=['tested'], nrows=20000)\ndataset = df.values\ndataset = dataset.astype('float32')\n\nscaler = MinMaxScaler(feature_range=(0, 1))\ndataset = scaler.fit_transform(dataset)\ntrainX, trainY, testX, testY = lstmwindow(dataset, look_back)\n<\/code><\/pre>\n\n<p>I used learning rate because it was giving some good results then default setting.<\/p>\n\n<pre><code>opt = Adam(lr=0.0000001, decay=.2)\nmodel = Sequential()\nmodel.add(LSTM(1028, input_shape=(1, look_back), return_sequences=True))\nmodel.add(LSTM(128, return_sequences=True))\nmodel.add(LSTM(256, return_sequences=True))\nmodel.add(LSTM(128, return_sequences=True))\nmodel.add(LSTM(64))\nmodel.add(Dense(1))\nmodel.compile(loss='mse', optimizer='adam', metrics=['mape', 'acc'])\nmodel.fit(trainX, trainY, epochs=20, batch_size=64, verbose=1)\n\n\ntrainPredict = model.predict(trainX)\ntestPredict = model.predict(testX)\n\ntrainPredict = scaler.inverse_transform(trainPredict)\ntrainY = scaler.inverse_transform([trainY])\ntestPredict = scaler.inverse_transform(testPredict)\ntestY = scaler.inverse_transform([testY])\nprint (trainPredict, trainY)\n\n\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\n\nplt.plot(trainY[0][:100])\nplt.plot(trainPredict.reshape(-1, trainPredict.shape[0])[0][:100])\nplt.show()\n<\/code><\/pre>\n\n<p>the results are. After sometimes witha a epoch the loss value is gradually increasing and the accuracy is very low.<\/p>\n\n<pre><code>432\/12552 [====&gt;.........................] - ETA: 32s - loss: 0.0059 - mean_absolute_percentage_error: 992116.1136 - acc: 0.0082 \n 2496\/12552 [====&gt;.........................] - ETA: 32s - loss: 0.0060 - mean_absolute_percentage_error: 966677.6248 - acc: 0.0080\n 2560\/12552 [=====&gt;........................] - ETA: 32s - loss: 0.0061 - mean_absolute_percentage_error: 963082.6779 - acc: 0.0086\n 2624\/12552 [=====&gt;........................] - ETA: 31s - loss: 0.0061 - mean_absolute_percentage_error: 939593.1212 - acc: 0.0084\n 2688\/12552 [=====&gt;........................] - ETA: 31s - loss: 0.0060 - mean_absolute_percentage_error: 957549.8326 - acc: 0.0089\n 2752\/12552 [=====&gt;........................] - ETA: 31s - loss: 0.0060 - mean_absolute_percentage_error: 935281.5181 - acc: 0.0087\n 2816\/12552 [=====&gt;........................] - ETA: 31s - loss: 0.0060 - mean_absolute_percentage_error: 963213.3245 - acc: 0.0092\n 2880\/12552 [=====&gt;........................] - ETA: 31s - loss: 0.0059 - mean_absolute_percentage_error: 941808.8309 - acc: 0.0090\n 2944\/12552 [======&gt;.......................] - ETA: 31s - loss: 0.0059 - mean_absolute_percentage_error: 921335.0405 - acc: 0.0088\n 3008\/12552 [======&gt;.......................] - ETA: 30s - loss: 0.0059 - mean_absolute_percentage_error: 920601.9731 - acc: 0.0090\n 3072\/12552 [======&gt;.......................] - ETA: 30s - loss: 0.0060 - mean_absolute_percentage_error: 901423.1130 - acc: 0.0088\n 3136\/12552 [======&gt;.......................] - ETA: 30s - loss: 0.0060 - mean_absolute_percentage_error: 901941.2332 - acc: 0.0089\n 3200\/12552 [======&gt;.......................] - ETA: 30s - loss: 0.0060 - mean_absolute_percentage_error: 883902.6478 - acc: 0.0088\n 3264\/12552 [======&gt;.......................] - ETA: 29s - loss: 0.0060 - mean_absolute_percentage_error: 887954.1915 - acc: 0.0089\n 3328\/12552 [======&gt;.......................] - ETA: 29s - loss: 0.0059 - mean_absolute_percentage_error: 889670.6806 - acc: 0.0090\n 3392\/12552 [=======&gt;......................] - ETA: 29s - loss: 0.0059 - mean_absolute_percentage_error: 891472.6347 - acc: 0.0091\n 3456\/12552 [=======&gt;......................] - ETA: 29s - loss: 0.0060 - mean_absolute_percentage_error: 907832.6322 - acc: 0.0093\n 3520\/12552 [=======&gt;......................] - ETA: 29s - loss: 0.0060 - mean_absolute_percentage_error: 891326.8646 - acc: 0.0091\n 3584\/12552 [=======&gt;......................] - ETA: 28s - loss: 0.0061 - mean_absolute_percentage_error: 1068598.5278 - acc: 0.0098\n 3648\/12552 [=======&gt;......................] - ETA: 28s - loss: 0.0060 - mean_absolute_percentage_error: 1089488.1545 - acc: 0.0101\n 3712\/12552 [=======&gt;......................] - ETA: 28s - loss: 0.0060 - mean_absolute_percentage_error: 1070704.1027 - acc: 0.0100\n 3776\/12552 [========&gt;.....................] - ETA: 28s - loss: 0.0061 - mean_absolute_percentage_error: 1052556.8863 - acc: 0.0098\n 3840\/12552 [========&gt;.....................] - ETA: 28s - loss: 0.0060 - mean_absolute_percentage_error: 1035014.4576 - acc: 0.0096\n 3904\/12552 [========&gt;.....................] - ETA: 27s - loss: 0.0060 - mean_absolute_percentage_error: 1018047.2437 - acc: 0.0095\n 3968\/12552 [========&gt;.....................] - ETA: 27s - loss: 0.0060 - mean_absolute_percentage_error: 1044726.5623 - acc: 0.0096\n 4032\/12552 [========&gt;.....................] - ETA: 27s - loss: 0.0060 - mean_absolute_percentage_error: 1047885.8193 - acc: 0.0097\n 4096\/12552 [========&gt;.....................] - ETA: 27s - loss: 0.0061 - mean_absolute_percentage_error: 1198321.3221 - acc: 0.0095\n 4160\/12552 [========&gt;.....................] - ETA: 27s - loss: 0.0061 - mean_absolute_percentage_error: 1265580.6710 - acc: 0.0099\n 4224\/12552 [=========&gt;....................] - ETA: 26s - loss: 0.0061 - mean_absolute_percentage_error: 1302111.3199 - acc: 0.0102\n 4288\/12552 [=========&gt;....................] - ETA: 26s - loss: 0.0061 - mean_absolute_percentage_error: 1282676.9945 - acc: 0.0100\n 4352\/12552 [=========&gt;....................] - ETA: 26s - loss: 0.0061 - mean_absolute_percentage_error: 1263814.2721 - acc: 0.0099\n 4416\/12552 [=========&gt;....................] - ETA: 26s - loss: 0.0061 - mean_absolute_percentage_error: 1254763.8207 - acc: 0.0100\n 4480\/12552 [=========&gt;....................] - ETA: 26s - loss: 0.0060 - mean_absolute_percentage_error: 1236838.8069 - acc: 0.0098\n 4544\/12552 [=========&gt;....................] - ETA: 25s - loss: 0.0060 - mean_absolute_percentage_error: 1244310.3624 - acc: 0.0099\n 4608\/12552 [==========&gt;...................] - ETA: 25s - loss: 0.0060 - mean_absolute_percentage_error: 1227028.4810 - acc: 0.0098\n 4672\/12552 [==========&gt;...................] - ETA: 25s - loss: 0.0060 - mean_absolute_percentage_error: 1210220.0377 - acc: 0.0096\n 4736\/12552 [==========&gt;...................] - ETA: 25s - loss: 0.0060 - mean_absolute_percentage_error: 1230652.4426 - acc: 0.0097\n 4800\/12552 [==========&gt;...................] - ETA: 25s - loss: 0.0059 - mean_absolute_percentage_error: 1214243.8915 - acc: 0.0096\n 4864\/12552 [==========&gt;...................] - ETA: 24s - loss: 0.0059 - mean_absolute_percentage_error: 1198267.1610 - acc: 0.0095\n<\/code><\/pre>\n","answers":[{"AnswerId":"57557943","CreationDate":"2019-08-19T13:40:06.483","ParentId":null,"OwnerUserId":"9025861","Title":null,"Body":"<p>You might be <a href=\"https:\/\/machinelearningmastery.com\/overfitting-and-underfitting-with-machine-learning-algorithms\/\" rel=\"nofollow noreferrer\">overfitting<\/a> your model here which explains the decrease in accuracy and increase in loss. Try parameter tuning. Your learning rate can be higher. Try adding <a href=\"https:\/\/machinelearningmastery.com\/dropout-for-regularizing-deep-neural-networks\/\" rel=\"nofollow noreferrer\">dropout<\/a> to avoid overfitting.<\/p>\n\n<pre><code>from keras.layers import Dropout\nmodel.add(Dropout(0.2))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57529242,"AnswerCount":0,"Tags":"<ios><swift><tensorflow><cocoapods><arkit>","CreationDate":"2019-08-16T18:07:46.583","AcceptedAnswerId":null,"OwnerUserId":8911290.0,"Title":"How can I use a tenflow model that uses ARCaptureSession at the same time as an ARSession","Body":"<p>I can't run an ARCaptureSession and ARSession at the same time so I'm using this code to get the image to run the object deteciton model on:<\/p>\n\n<pre><code>func session(_ session: ARSession, didUpdate frame: ARFrame) {\n    runModel(onPixelBuffer: frame.capturedImage)\n<\/code><\/pre>\n\n<p>The image the AR function captures fails these assertions the model has:<\/p>\n\n<pre><code>func runModel(onFrame pixelBuffer: CVPixelBuffer) -&gt; Result? {\n     let imageWidth = CVPixelBufferGetWidth(pixelBuffer)\n     let imageHeight = CVPixelBufferGetHeight(pixelBuffer)\n     let sourcePixelFormat = CVPixelBufferGetPixelFormatType(pixelBuffer)\n     assert(sourcePixelFormat == kCVPixelFormatType_32ARGB ||\n            sourcePixelFormat == kCVPixelFormatType_32BGRA ||\n            sourcePixelFormat == kCVPixelFormatType_32RGBA)\n<\/code><\/pre>\n\n<p>if I print <code>sourcePixelFormat<\/code> is gives me <code>875704422<\/code><\/p>\n\n<p>The model runs if I remove the assertions but the predictions are all messed up<\/p>\n","answers":[]}
+{"QuestionId":57529534,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-datasets><tensorflow2.0>","CreationDate":"2019-08-16T18:33:49.520","AcceptedAnswerId":null,"OwnerUserId":3860928.0,"Title":"How is tf.data.Dataset use optimised by tf.function in Tensorflow 2.0?","Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/effective_tf2#combine_tfdatadatasets_and_tffunction\" rel=\"nofollow noreferrer\">official documentation<\/a> of Tensorflow 2.0 advices to use <code>tf.data.Dataset<\/code> along with <code>tf.function<\/code>.<\/p>\n\n<p>There are two examples of such uses:<\/p>\n\n<ul>\n<li>Using a <code>Dataset<\/code> as an argument of a <code>tf.function<\/code>, as described <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/effective_tf2#combine_tfdatadatasets_and_tffunction\" rel=\"nofollow noreferrer\">here<\/a>:<\/li>\n<\/ul>\n\n<pre class=\"lang-py prettyprint-override\"><code>@tf.function\ndef train(model, dataset, optimizer):\n  for x, y in dataset:\n      ....\n<\/code><\/pre>\n\n<ul>\n<li>Creating a local <code>Dataset<\/code> in a <code>tf.function<\/code> body, as described <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/autograph\" rel=\"nofollow noreferrer\">here<\/a>:<\/li>\n<\/ul>\n\n<pre class=\"lang-py prettyprint-override\"><code>@tf.function\ndef train(model, optimizer):\n  train_ds = mnist_dataset()\n  ...\n<\/code><\/pre>\n\n<p>Finally, the <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/eager\/tf_function#autograph_and_loops\" rel=\"nofollow noreferrer\">autograph doc<\/a> states that the iterating on a <code>Dataset<\/code> is optimized by <code>tf.function<\/code>.\n<a href=\"https:\/\/stackoverflow.com\/questions\/56038372\/does-wrapping-tf-data-dataset-into-tf-function-improve-performance\">This SO answer<\/a> shows indeed that using a <code>Dataset<\/code> as a parameter of <code>tf.function<\/code> increases performances.<\/p>\n\n<p>So, how does <code>tf.data.Dataset<\/code> benefit from <code>tf.function<\/code>, and how can it explain the speedup of <a href=\"https:\/\/stackoverflow.com\/questions\/56038372\/does-wrapping-tf-data-dataset-into-tf-function-improve-performance\">this SO answer<\/a>:<\/p>\n\n<ul>\n<li>How is the <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/eager\/tf_function#autograph_and_loops\" rel=\"nofollow noreferrer\">\"foor loop optimization\"<\/a> creating a speedup by itself?<\/li>\n<li>How is the <code>Dataset<\/code> object treated by the <code>tf.function<\/code> tracer. Like in the parameter or as a local variable examples, how can we use <code>Dataset<\/code> in a <code>tf.function<\/code>, even if the <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/function\" rel=\"nofollow noreferrer\"><code>tf.function<\/code> documentation<\/a> suggests that the original function should only accept <code>Tensor<\/code>?<\/li>\n<li>Is there some kind of communication\/sharing between the <code>Dataset<\/code> subgraph (that can then be used in <code>tf.function<\/code> subgraph), or the two are fully separated (meaning that <code>Dataset<\/code> is just used as an black-box iterator from <code>tf.function<\/code>)? <\/li>\n<\/ul>\n","answers":[]}
+{"QuestionId":57529831,"AnswerCount":1,"Tags":"<python><tensorflow><keras><gradient><tensorflow-datasets>","CreationDate":"2019-08-16T18:58:32.963","AcceptedAnswerId":57530016.0,"OwnerUserId":2942985.0,"Title":"Solved: How to combine tf.gradients with tf.data.dataset and keras models","Body":"<p>I'm trying to build a workflow that uses <code>tf.data.dataset<\/code> batches and an iterator. For performance reasons, I am really trying to avoid using the placeholder->feed_dict loop workflow.<\/p>\n\n<p>The process I'm trying to implement involves grad-cam (which requires the gradient of the loss with respect to the final convolutional layer of a CNN) as an intermediate step, and ideally I'd like to be able to try it out on several Keras pre-trained models, including non-sequential ones like ResNet.<\/p>\n\n<p>Most implementations of grad-cam that I've found rely on hand-crafting the CNN of interest in tensorflow. I found one implementation, <a href=\"https:\/\/github.com\/jacobgil\/keras-grad-cam\" rel=\"nofollow noreferrer\">https:\/\/github.com\/jacobgil\/keras-grad-cam<\/a>, that is made for keras models, and following that example, I get<\/p>\n\n<pre><code>\ndef safe_norm(x):\n    return x \/ tf.sqrt(tf.reduce_mean(x ** 2) + 1e-8)\nvgg_ = VGG19()\n\ndataset = tf.data.Dataset.from_tensor_slices((filenames))\n\n#preprocessing...\n\nit = dataset.make_one_shot_iterator()\n\nfiles, batch = it.get_next()\nconv5_4 = vgg_.layers[-6]\nh_k, w_k, c_k = conv5_4.output.shape[1:]\n\nvgg_model = Model(inputs=vgg_.input, outputs=vgg_.output)\nconv_model = Model(inputs=vgg_.input, outputs=conv5_4.output)\nprobs = vgg_model(batch)\npredicted_class = tf.argmax(probs, axis=-1)\n\nlayer_name = 'block5_conv4'\ntarget_layer = lambda x: target_category_loss(x, predicted_class, n_categories)\nx = Lambda(target_layer)(vgg_model.outputs[0])\nmodel = Model(inputs=vgg_model.inputs[0], outputs=x)\n\nloss = K.sum(model.output, axis=-1)\nconv_output =  [l for l in model.layers if l.name is layer_name][0].output\ngrads = Lambda(safe_norm)(K.gradients(loss, [conv_output])[0])\ngradient_function = K.function([model.input], [conv_output, grads])\n\noutput, grads_val = gradient_function([batch])\nweights = tf.reduce_mean(grads_val, axis = (1, 2))\ncam = tf.ones([batch_size, h_k, w_k], dtype = tf.float32)\n\ncam += tf.reduce_sum(output * tf.reshape(weights, [-1, 1, 1, weights.shape[-1]]), axis=-1)\n\ncam = tf.squeeze(tf.image.resize_images(images=tf.expand_dims(cam, axis=-1), size=(224, 224)))\ncam = tf.maximum(cam, 0)\nheatmap = cam \/ tf.reshape(tf.reduce_max(cam, axis=[1, 2]), shape=[-1, 1, 1])\n<\/code><\/pre>\n\n<p>The problem is that <code>gradient_function([batch])<\/code> returns a numpy array whose value is determined by the first batch, so that <code>heatmap<\/code> doesn't change with subsequent evaluations.<\/p>\n\n<p>I've tried replacing <code>K.function<\/code> with a <code>Model<\/code> in various ways, but nothing seems to work. I usually end up either with an error suggesting that <code>grads<\/code> evaluates to <code>None<\/code> or that one model or another is expecting a feed_dict and not receiving one.<\/p>\n\n<p>Is this code salvageable? Is there a better way to do this besides looping through the data several times (once to get all the grad-cams and then again once I have them) or using placeholders and feed_dicts?<\/p>\n\n<p>Edit:<\/p>\n\n<pre><code>\ndef safe_norm(x):\n    return x \/ tf.sqrt(tf.reduce_mean(x ** 2) + 1e-8)\nvgg_ = VGG19()\n\ndataset = tf.data.Dataset.from_tensor_slices((filenames))\n\n#preprocessing...\n\nit = dataset.make_one_shot_iterator()\n\nfiles, batch = it.get_next()\nconv5_4 = vgg_.layers[-6]\nh_k, w_k, c_k = conv5_4.output.shape[1:]\n\nvgg_model = Model(inputs=vgg_.input, outputs=vgg_.output)\nconv_model = Model(inputs=vgg_.input, outputs=conv5_4.output)\nprobs = vgg_model(batch)\npredicted_class = tf.argmax(probs, axis=-1)\n\nlayer_name = 'block5_conv4'\ntarget_layer = lambda x: target_category_loss(x, predicted_class, n_categories)\nx = Lambda(target_layer)(vgg_model.outputs[0])\nmodel = Model(inputs=vgg_model.inputs[0], outputs=x)\n\nloss = K.sum(model.output, axis=-1)\nconv_output =  [l for l in model.layers if l.name is layer_name][0].output\ngrads = Lambda(safe_norm)(K.gradients(loss, [conv_output])[0])\ngradient_function = K.function([model.input], [conv_output, grads])\n\noutput, grads_val = gradient_function([batch])\nweights = tf.reduce_mean(grads_val, axis = (1, 2))\ncam = tf.ones([batch_size, h_k, w_k], dtype = tf.float32)\n\ncam += tf.reduce_sum(output * tf.reshape(weights, [-1, 1, 1, weights.shape[-1]]), axis=-1)\n\ncam = tf.squeeze(tf.image.resize_images(images=tf.expand_dims(cam, axis=-1), size=(224, 224)))\ncam = tf.maximum(cam, 0)\nheatmap = cam \/ tf.reshape(tf.reduce_max(cam, axis=[1, 2]), shape=[-1, 1, 1])\n\n# other operations on heatmap and batch ...\n\n# ...\n\noutput_function = K.function(model.input, [node1, ..., nodeN])\n\nfor batch in range(n_batches):\n    outputs1, ... , outputsN = output_function(batch)\n<\/code><\/pre>\n\n<p>Gives me the desired outputs for each batch.<\/p>\n","answers":[{"AnswerId":"57530016","CreationDate":"2019-08-16T19:17:50.980","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Yes, <code>K.function<\/code> returns numpy arrays because it evaluates the symbolic computation in your graph. What I think you should do is to keep everything symbolic up to <code>K.function<\/code>, and after getting the gradients, perform all computations of the Grad-CAM weights and final saliency map using numpy.<\/p>\n\n<p>Then you can iterate on your dataset, evaluate <code>gradient_function<\/code> on a new batch of data, and compute the saliency map.<\/p>\n\n<p>If you want to keep everything symbolic, then you should not use <code>K.function<\/code> to produce the gradient function, but use the symbolic gradient (the output of  <code>K.gradient<\/code>, without lambda) and convolutional feature maps (<code>conv_output<\/code>) and perform the saliency map computation on top of that, and then build a function (using <code>K.function<\/code>) that takes the model input, and outputs the saliency map.<\/p>\n\n<p>Hope the explanation is enough.<\/p>\n"}]}
+{"QuestionId":57530038,"AnswerCount":1,"Tags":"<deep-learning><pytorch><image-segmentation><unity3d-unet>","CreationDate":"2019-08-16T19:19:45.047","AcceptedAnswerId":57534734.0,"OwnerUserId":6642287.0,"Title":"Concatenation step of U-Net for unequal number of channels","Body":"<p>I am trying to implement the U-NET architecture for image segmentation while implementing the crop and concatenation step in the expansive path, I am unable to understand how the unequal number of channels are concatenated.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/bB2tL.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/bB2tL.jpg\" alt=\"Unet\"><\/a><\/p>\n\n<p>According to the architecture, the input from the first upsampling step has to be concatenated from the corresponding output from contracting path but the problem is number of channels in contracting path is 512 while after upsampling step they are 1024, how they are supposed to be concatenated.My code for the crop and concatenate is -<\/p>\n\n<pre><code>def crop_and_concat(self, upsampled, bypass, crop=False):\n    if crop:\n        c = (bypass.size()[2] - upsampled.size()[2]) \/\/ 2\n        bypass = F.pad(bypass, (-c, -c, -c, -c))\n    return torch.cat((upsampled, bypass), 1)\n<\/code><\/pre>\n\n<p>The error I am receiving- \n<code>RuntimeError: Given groups=1, weight of size 128 256 5 5, expected input[4, 384, 64, 64] to have 256 channels, but got 384 channels instead<\/code>\n<br>Where I am doing wrong?<\/p>\n","answers":[{"AnswerId":"57534734","CreationDate":"2019-08-17T08:48:07.683","ParentId":null,"OwnerUserId":"10886420","Title":null,"Body":"<p>First of all, you don't have to be so strict when it comes to U-Net like architectures, there were many derivatives afterwards (see for example <a href=\"https:\/\/docs.fast.ai\/vision.models.unet.html\" rel=\"nofollow noreferrer\">fastai variation with PixelShuffle<\/a>).<\/p>\n\n<p>In the case of encoder, in the basic version, your channels go (per block):<\/p>\n\n<pre><code>1 - 64 - 128 - 256 - 512\n<\/code><\/pre>\n\n<p>Standard convolutional encoder.\nAfter that is a shared layer of  <code>1024<\/code>.<\/p>\n\n<p>In decoder, it goes downwards, <strong>but has more channels as you are concatenating encoder states from each block<\/strong>.<\/p>\n\n<p>It would be:<\/p>\n\n<blockquote>\n  <p>1024 -> 512 -> 512 (decoder) + 512 (encoder), 1024 total -> 512<\/p>\n  \n  <p>512 -> 256 -> 256 (decoder) + 256 (encoder), 512 total -> 256<\/p>\n<\/blockquote>\n\n<p>and so on.<\/p>\n\n<p>You were at the case where <code>256<\/code> from decoder was taken in the account, but <code>128<\/code> added from encoder wasn't. Just up your channels to 256 + 128 and follow the above scheme for each block of your UNet.<\/p>\n"}]}
+{"QuestionId":57530326,"AnswerCount":0,"Tags":"<tensorflow><build><bazel><trie>","CreationDate":"2019-08-16T19:46:51.403","AcceptedAnswerId":null,"OwnerUserId":11907197.0,"Title":"How to fix \"Build did NOT complete successfully (0 packages loaded)\" error in Terminal for compiling libraries","Body":"<p>i am using bazel in order to build my deepspeech library as to compile Deepspeech. Currently i have successfully configured tensorflow.<\/p>\n\n<p>I have followed the documentation provided on github after successfully configuring tensorflow. <\/p>\n\n<p>As noted in the error below that I dont have the native_client and build file but these are in my current directory as noted by pwd deepspeech.<\/p>\n\n<p>(deepspeech-venv) Chabanis-MacBook-Pro:tensorflow chabani$ pwd deepspeech\n\/Users\/chabani\/tmp\/deepspeech-venv\/bin\/tensorflow\n(deepspeech-venv) Chabanis-MacBook-Pro:tensorflow chabani$ bazel build --workspace_status_command=\"bash native_client\/bazel_workspace_status_cmd.sh\" --config=monolithic -c opt --copt=-O3 --copt=\"-D_GLIBCXX_USE_CXX11_ABI=0\" --copt=-fvisibility=hidden \/\/native_client:libdeepspeech.so \/\/native_client:generate_trie<\/p>\n\n<p>ERROR: Skipping '\/\/native_client:libdeepspeech.so': no such package 'native_client': BUILD file not found on package path\nERROR: no such package 'native_client': BUILD file not found on package path\nINFO: Elapsed time: 1.520s\nINFO: 0 processes.\nFAILED: Build did NOT complete successfully (0 packages loaded)<\/p>\n\n<p>i expected that this would have worked perfectly fine.<\/p>\n","answers":[]}
+{"QuestionId":57530480,"AnswerCount":2,"Tags":"<python><tensorflow><keras><tensorflow-datasets><tensorflow2.0>","CreationDate":"2019-08-16T20:00:39.827","AcceptedAnswerId":null,"OwnerUserId":6641872.0,"Title":"TF2.0 Data API get n_i samples from each class label","Body":"<p>I have to classify inputs of shape 32x32 into 3 classes using a TF2 Keras model. My training set has 7000 examples<\/p>\n\n<pre><code>&gt;&gt;&gt; X_train.shape # (7000, 32, 32)\n&gt;&gt;&gt; Y_train.shape # (7000, 3)\n<\/code><\/pre>\n\n<p>The number of examples for each class varies (e.g. class_0 has ~2500 examples while class_1 has ~800, etc.)<\/p>\n\n<p>I want to use the tf.data API to create a dataset object that returns batches of training data with no. of examples from each class specified by <code>[n_0, n_1, n_2]<\/code>. <\/p>\n\n<p>I would like to have these <code>n_i<\/code> samples from each class randomly drawn with replacement from <code>X_train, Y_train<\/code><\/p>\n\n<p>For example, if I call <code>get_batch([100, 150, 125])<\/code> it should return 100 random samples from <code>X_batch<\/code> from class_0, 150 from class_1, and 125 from class_2.<\/p>\n\n<p>How can I achieve this using the TF2.0 Data API so I could use it for training a Keras model?<\/p>\n","answers":[{"AnswerId":"57531878","CreationDate":"2019-08-16T22:42:56.720","ParentId":null,"OwnerUserId":"8031842","Title":null,"Body":"<p>Keras' <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.model_selection.train_test_split.html\" rel=\"nofollow noreferrer\">train_test_split<\/a> actually has a parameter for that. While it doesn't let you pick exact number of samples, it selects them evenly from the classes.<\/p>\n\n<pre><code>X_train_stratified, X_test_stratified, y_train_strat, y_test_strat = train_test_split(X_train, y_train, test_size=0.2, stratify=y)\n<\/code><\/pre>\n\n<p>If you want to do cross validation you can also use <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.model_selection.StratifiedShuffleSplit.html\" rel=\"nofollow noreferrer\">stratified shuffle split<\/a><\/p>\n\n<p>I hope I understood your question correctly<\/p>\n"},{"AnswerId":"57534455","CreationDate":"2019-08-17T08:07:18.517","ParentId":null,"OwnerUserId":"2455494","Title":null,"Body":"<p>One possible approach is to proceed as follows:<\/p>\n\n<ol>\n<li>Load the data from <code>X_train<\/code> &amp; <code>Y_train<\/code> into a single <code>tf.data<\/code> Dataset so that we ensure we keep each <code>X<\/code> matched with the correct <code>Y<\/code><\/li>\n<li><code>.shuffle()<\/code> then split the dataset into each <code>n_i<\/code> using a <code>filter()<\/code><\/li>\n<li>Write our <code>get_batch<\/code> function to return the correct number of samples from each dataset, <code>shuffle()<\/code> the sample then split it back into <code>X<\/code> &amp; <code>Y<\/code><\/li>\n<\/ol>\n\n<p>Something like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># 1: Load the data into a Dataset\nraw_data = tf.data.Dataset.zip(\n    (\n        tf.data.Dataset.from_tensor_slices(X_train),\n        tf.data.Dataset.from_tensor_slices(Y_train)\n    )\n  ).shuffle(7000)\n\n\n# 2: Split for each category\ndef get_filter_fn(n):\n  def filter_fn(x, y):\n    return tf.equal(1.0, y[n])\n  return filter_fn\n\nn_0s = raw_data.filter(get_filter_fn(0))\nn_1s = raw_data.filter(get_filter_fn(1))\nn_2s = raw_data.filter(get_filter_fn(2))\n\n# 3:\ndef get_batch(n_0,n_1,n_2):\n  sample = n_0s.take(n_0).concatenate(n_1s.take(n_1)).concatenate(n_2s.take(n_2))\n  shuffled = sample.shuffle(n_0 + n_1 + n_2)\n  return shuffled.map(lambda x,y: x),shuffled.map(lambda x,y: y) \n\n<\/code><\/pre>\n\n<p>So now we can do:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x_batch, y_batch = get_batch(100, 150, 125)\n<\/code><\/pre>\n\n<p>Note that I've used some potentially wasteful operations here pursuing an approach I find intuitive and straightforward (specifically reading the <code>raw_data<\/code> dataset 3 times for the filter operations) so I make no claim that this is the most efficient way to accomplish what you need but for a dataset that fits in memory like the one you describe I'm sure such inefficiencies will be negligible <\/p>\n"}]}
+{"QuestionId":57530501,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-08-16T20:01:59.273","AcceptedAnswerId":null,"OwnerUserId":1451614.0,"Title":"Viewing x data generator using a call back after calling model.fit","Body":"<p>I am having difficulties in viewing the data used to train the model in the middle of training after calling tf.keras.Model.fit().<\/p>\n\n<p>I create a dataset = tf.data.DataSet object which I pass to a tf.keras.Model object via the Model's Model.fit() (x=dataset).<\/p>\n\n<p>I have reviewed various keras callback tutorials. For example:\ni.e. <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/keras\/custom_callback\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/guide\/keras\/custom_callback<\/a><\/p>\n\n<p>I understand that I can access the model using for the .fit() call by using self.model in the callback.<\/p>\n\n<p>I tried using this to get a layer, and get its value. But because it is a tensor, it needs a placeholder (i.e. the image I am trying to see).<\/p>\n\n<p>Here is an example of the callback.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class MyCustomCallback(tf.keras.callbacks.Callback):\n\n        def __init__(self):\n            super(MyCustomCallback, self).__init__()\n\n        def on_train_batch_begin(self, batch, logs=None):\n            print('Training: batch {} begins at {}'.format(batch, datetime.datetime.now().time()))\n\n            resnet50_model = self.model.get_layer(name='resnet50')\n            resnet50_model.summary()\n\n            input_layer = resnet50_model.get_layer(name='input_1')\n            input_tensor = input_layer.get_input_at(0)\n            sess = tf.keras.backend.get_session()\n            input_array = sess.run(input_tensor)\n\n            print('input_layer: {}'.format(input_array))\n            exit(-1)\n\n        def on_train_batch_end(self, batch, logs=None):\n            print('Training: batch {} ends at {}'.format(batch, datetime.datetime.now().time()))\n\n        def on_test_batch_begin(self, batch, logs=None):\n            print('Evaluating: batch {} begins at {}'.format(batch, datetime.datetime.now().time()))\n\n        def on_test_batch_end(self, batch, logs=None):\n            print('Evaluating: batch {} ends at {}'.format(batch, datetime.datetime.now().time()))\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: 2 root error(s) found.\n  (0) Invalid argument: You must feed a value for placeholder tensor 'input_1' with dtype float and shape [?,224,224,3]\n     [[node input_1 (defined at \/usr\/local\/lib\/python3.5\/dist-packages\/keras_applications\/resnet50.py:214) ]]\n     [[input_1\/_2591]]\n<\/code><\/pre>\n\n<p>I then inspected the Tensorboard callback:\n<a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/callbacks.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/callbacks.py<\/a><\/p>\n\n<p>and saw that they also getting the session via the keras backend and then running sess.run() with a feed_dict containing images passed the instantiation of the call back.<\/p>\n\n<p>Thus, this method does not work. However, I still ask the question, how does one get the actual images and labels (i.e. tf.data.Dataset object passed to x in Model.fit()) at the beginning of every batch?<\/p>\n\n<p>Thank you!<\/p>\n","answers":[]}
+{"QuestionId":57530823,"AnswerCount":2,"Tags":"<tensorflow><keras><keras-layer><tf.keras>","CreationDate":"2019-08-16T20:36:27.110","AcceptedAnswerId":57531217.0,"OwnerUserId":11861082.0,"Title":"Keras backend mean function: \" 'float' object has no attribute 'dtype' \"?","Body":"<p>I am trying to introduce a new kernel regularize for a network using Keras. But, it gives me the error:\n 'float' object has no attribute 'dtype'\nHow can I fix it?<\/p>\n\n<p>I found the code here: <a href=\"https:\/\/stackoverflow.com\/questions\/43599082\/kl-divergence-in-kerastensorflow-backend\">KL divergence in keras(tensorflow backend)<\/a><\/p>\n\n<p>Here is my code:<\/p>\n\n<pre><code>    from keras import backend as K\n\n    kullback_leibler_divergence = keras.losses.kullback_leibler_divergence\n\n    def kl_divergence_regularizer(inputs):\n        means = K.mean((inputs))\n        return 0.5 *(0.01 * (kullback_leibler_divergence(0.05, means)\n                      + kullback_leibler_divergence(1 - 0.05, 1 - means)))\n\n\n    model = Sequential([\n        Dense(100, input_shape=(x_train_s.shape[1],),kernel_initializer='random_uniform'),\n        Activation('elu'),\n        Dense(x_train_s.shape[1],kernel_initializer='random_uniform', kernel_regularizer=kl_divergence_regularizer),\n        Activation('tanh')\n    ])\n\n    model.compile(optimizer='adam',loss='mean_squared_error')\n\n    model.fit(x_train_s, x_train_s, epochs=5,validation_split=0.1, shuffle=True, verbose=1,batch_size=np.uint(x_train_s.shape[0]\/100))\n<\/code><\/pre>\n\n<p>The full error trace back is presented here:<\/p>\n\n<pre><code>AttributeError                            Traceback (most recent call last)\n&lt;ipython-input-39-59bc90c687de&gt; in &lt;module&gt;\n     39     Activation('elu'),\n     40     Dense(x_train_s.shape[1],kernel_initializer='random_uniform', kernel_regularizer=kl_divergence_regularizer),\n---&gt; 41     Activation('tanh')\n     42 ])\n     43 \n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\engine\\sequential.py in __init__(self, layers, name)\n     91         if layers:\n     92             for layer in layers:\n---&gt; 93                 self.add(layer)\n     94 \n     95     @property\n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\engine\\sequential.py in add(self, layer)\n    179                 self.inputs = network.get_source_inputs(self.outputs[0])\n    180         elif self.outputs:\n--&gt; 181             output_tensor = layer(self.outputs[0])\n    182             if isinstance(output_tensor, list):\n    183                 raise TypeError('All layers in a Sequential model '\n\nC:\\ProgramData\\Anaconda3\\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\nC:\\ProgramData\\Anaconda3\\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\nC:\\ProgramData\\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\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\engine\\base_layer.py in add_weight(self, name, shape, dtype, initializer, regularizer, trainable, constraint)\n    253         if regularizer is not None:\n    254             with K.name_scope('weight_regularizer'):\n--&gt; 255                 self.add_loss(regularizer(weight))\n    256         if trainable:\n    257             self._trainable_weights.append(weight)\n\n&lt;ipython-input-39-59bc90c687de&gt; in kl_divergence_regularizer(inputs)\n     14     means = K.mean((inputs))\n     15 #     means=1e-6\n---&gt; 16     return 0.5 *(0.01 * (kullback_leibler_divergence(0.05, means)\n     17                   + kullback_leibler_divergence(1 - 0.05, 1 - means)))\n     18 #     return 1e-10\n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\losses.py in kullback_leibler_divergence(y_true, y_pred)\n     79 \n     80 def kullback_leibler_divergence(y_true, y_pred):\n---&gt; 81     y_true = K.clip(y_true, K.epsilon(), 1)\n     82     y_pred = K.clip(y_pred, K.epsilon(), 1)\n     83     return K.sum(y_true * K.log(y_true \/ y_pred), axis=-1)\n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py in clip(x, min_value, max_value)\n   1599     if max_value is None:\n   1600         max_value = np.inf\n-&gt; 1601     min_value = _to_tensor(min_value, x.dtype.base_dtype)\n   1602     max_value = _to_tensor(max_value, x.dtype.base_dtype)\n   1603     return tf.clip_by_value(x, min_value, max_value)\n\nAttributeError: 'float' object has no attribute 'dtype'\n<\/code><\/pre>\n\n<p>I need the KL divergence between 0.05 and mean calculate the following sum over i:<\/p>\n\n<p>KL=sum(0.05*\\log(0.05\/mean[i])) <\/p>\n","answers":[{"AnswerId":"57531217","CreationDate":"2019-08-16T21:14:27.060","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You should change your KL divergence regularizer to:<\/p>\n\n<pre><code>def kl_divergence_regularizer(inputs):\n        means = K.mean((inputs))\n        down = 0.05 * K.ones_like(means)\n        up = (1 - 0.05) * K.ones_like(means)\n        return 0.5 *(0.01 * (kullback_leibler_divergence(down, means)\n                      + kullback_leibler_divergence(up, 1 - means)))\n<\/code><\/pre>\n\n<p>This is because the KL divergence takes two vectors\/tensors, the KL between a scalar and a vector cannot be computed (its not defined).<\/p>\n"},{"AnswerId":"57531005","CreationDate":"2019-08-16T20:53:24.013","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p><code>kullback_leibler_divergence(0.05, means)<\/code>  <\/p>\n\n<p>This is a <code>loss<\/code> function. It is expecting <code>y_true, y_pred<\/code>, both as <strong>tensors<\/strong>. \nYou are passing a <code>float<\/code> (0.05), and the system is trying to get the properties of this supposed tensor, but it's not a tensor. <\/p>\n\n<p>The same problem will happen in the second <code>kullback_leibler_divergence<\/code> call where you pass again a float <code>1 - 0.05<\/code>  instead of a tensor. <\/p>\n\n<p>A simple solution (but you must check whether this is mathematically reasonable -- I don't know what KL_diff is supposed to do) is use a <code>K.ones_like(means) * 0,05<\/code> and in the second call <code>K.ones_like(means)*(1-0.05)<\/code>.    <\/p>\n"}]}
+{"QuestionId":57530944,"AnswerCount":0,"Tags":"<machine-learning><keras><deep-learning><artificial-intelligence><lstm>","CreationDate":"2019-08-16T20:48:54.940","AcceptedAnswerId":null,"OwnerUserId":11936797.0,"Title":"Machine learning (LSTM) music generation doesn't predict well and has loss too high","Body":"<p>I am a beginner in machine learning and I am trying to make an 'AI' that can generate music. I tried to make a simple code for it, but the model doesn't work. It always has a high loss - low accuracy after training. I used MIT's music21 to get the notes and chords from the files I used (midi), (command: music21.chordify() which gets the chords of a midi file). Also when I try to predict the next notes\/chords It always gets stuck in a loop like A#3, A#3, A#3, A#3 ...<\/p>\n\n<p>I tried changing the dropout multiple times, I tried adding layers (LSTM or Dense), I tried training on smaller(1 song) or larger (100+ songs) datasets or different genres, I also tried different epochs ( from 10 to 100) but it always has about the same results: \n loss: 6-8 \n accuracy: 0.25 - 0.35 (25% - 35%) \n Strangely, almost always the validation loss is lower than the training loss and the validation accuracy is better than the training accuracy, one time I had: val_acc = 0.6, val_loss = 4 and train_acc = 0.2, train_loss = 8<\/p>\n\n<pre><code>from music21 import *\nimport numpy as np\nfrom keras.utils import np_utils\nimport tensorflow as tf\nfrom keras.models import Sequential\nfrom keras.layers import Activation, Dense, LSTM, Dropout, Flatten\nimport os\nimport matplotlib.pyplot as plt\nfrom keras.callbacks import EarlyStopping\n<\/code><\/pre>\n\n<h1>The vocabulary<\/h1>\n\n<pre><code>NotesList = ['C0','C#0','D0','Eb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n<\/code><\/pre>\n\n<h1>In variable 'g' I stored the chords ( collection of notes ) from the midi files ex: g = ['E3 C#3 F4', 'E4 A#3 A4' , ...])<\/h1>\n\n<pre><code>g = []\n<\/code><\/pre>\n\n<h1>here I get the chords from the file: 'Chords' and I store the data in variable g<\/h1>\n\n<pre><code>def get_training_set(chords, ):\n    for i in chords:\n       if(isinstance(i, chord.Chord)):\n           g.append((str(i)))\n\nfor i in range(  len(g)):\n    g[i] = g[i].replace(\"&lt;music21.chord.Chord \", \"\")\n    g[i] = g[i].replace(\"&gt;\", \"\")\n\nfor ind in range(len(g)):\n    g[ind] = g[ind].replace('B-1', 'A#1')\n    g[ind] = g[ind].replace('B-2', 'A#2')\n    g[ind] = g[ind].replace('B-3', 'A#3')\n    g[ind] = g[ind].replace('B-4', 'A#4')\n    g[ind] = g[ind].replace('B-5', 'A#5')\n    g[ind] = g[ind].replace('B-6', 'A#6')\n    g[ind] = g[ind].replace('B-7', 'A#7')\n    g[ind] = g[ind].replace('E-1', 'D#1')\n    g[ind] = g[ind].replace('E-2', 'D#2')\n    g[ind] = g[ind].replace('E-3', 'D#3')\n    g[ind] = g[ind].replace('E-4', 'D#4')\n    g[ind] = g[ind].replace('E-5', 'D#5')\n    g[ind] = g[ind].replace('E-6', 'D#6')\n    g[ind] = g[ind].replace('E-7', 'D#7')\n<\/code><\/pre>\n\n<h1>Here I take every Bach midi file I have and get the chords from it using 'get_training_set(chords, )'<\/h1>\n\n<pre><code>directory = ['\/Users\/DCI\/Desktop\/MidiTrainingData\/Bach']\nFiles = []\nfor MidiFolders in range(len(directory)):\n    for filename in os.listdir(directory[MidiFolders]):\n        if filename.endswith(\".mid\"):\n           Files.append(''+directory[MidiFolders]+'\/'+filename+'')\n\nfor File_num in range(len(Files)):\n    b = converter.parse(Files[File_num])\n    bc = b.chordify()\n    get_training_set(bc)\n<\/code><\/pre>\n\n<h1>In X (X.shape = aprox (100000, 114) is my training data, ( every row is a chord, more exactly it is 114 units long and every unit represents a note, if in a chord a note is played then that note has value 1 and if not it has value 0<\/h1>\n\n<pre><code>X = np.zeros((len(g), 114))\nfor ThisChord in range(len(g)):\n    for ThisNote in g[ThisChord].split():\n            NoteIndex = NotesList.index(ThisNote)\n            X[ThisChord][NoteIndex] = 1\n<\/code><\/pre>\n\n<h1>Get Y, ( I want my model to predict the next chord so Y is like X but with a row ahead<\/h1>\n\n<pre><code>Y = X[1:len(g)]\nX = X[0:len(g)-1]\n<\/code><\/pre>\n\n<h1>reshape X ( I am not sure it's ok)<\/h1>\n\n<pre><code>X = np.reshape(X, (X.shape[0], 1, 114))\n<\/code><\/pre>\n\n<h1>Create model ( I am not sure it's ok )<\/h1>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(128 , input_dim = 114, activation = 'relu', return_sequences=True))\nmodel.add(Dropout(0.4))\nmodel.add(LSTM(128 , activation = 'relu'))\nmodel.add(Dropout(0.4))\nmodel.add(Dense(units = 114,  activation = 'softmax'))\n\n\n\nmodel.compile(loss='categorical_crossentropy', optimizer = 'adam', metrics = ['accuracy'])\nes = EarlyStopping(monitor='val_loss', mode='min', verbose=1)\nhistory = model.fit(X, Y, epochs = 100, validation_split=0.2, shuffle=True)\n<\/code><\/pre>\n\n<h1>Plot the model ( most of the times the validation set has a lower loss than the training set, i don't know how)<\/h1>\n\n<pre><code>plt.plot(history.history['loss'])\nplt.plot(history.history['val_loss'])\nplt.title('model train vs validation loss')\nplt.ylabel('loss')\nplt.xlabel('epoch')\nplt.legend(['train', 'validation'], loc='upper right')\nplt.show()\n<\/code><\/pre>\n\n<h1>Prediction ( I really don't think it's ok)<\/h1>\n\n<pre><code>predset = ['D4','C#3'] # I start with this\nPred = stream.Stream()\ndef MaxElem(list1): #Function to get the index of the max  element\n    maxx = np.argmax(list1)\n    final_list.append(maxx+1)\n<\/code><\/pre>\n\n<h1>generate 50 chords<\/h1>\n\n<pre><code>for rep in range(50):\n    X_predict = np.zeros((1,114))\n    print('prediction set at &lt; t-1 &gt; ='+' '.join(predset)+'')\n    for TheNotesInPredictionSet in predset:\n        NoteIndex = NotesList.index(TheNotesInPredictionSet)\n        X_predict[0][NoteIndex] = 1\n    X_predict = np.reshape(X_predict,(X_predict.shape[0],1,114))\n    prediction = model.predict([X_predict])\n    prediction = prediction.tolist()\n    final_list = []\n    MaxElem(prediction[0])\n    predset =[]\n    for i in final_list:\n        predset.append(NotesList[i])\n    Pred.append(note.Note(''+' '.join(predset)+'', type='eighth'))\n<\/code><\/pre>\n\n<pre><code>#Show as audio using music21\n<\/code><\/pre>\n\n<pre><code>    Pred.show('midi')\n<\/code><\/pre>\n\n<p>expected result: acc>0.90<\/p>\n","answers":[]}
+{"QuestionId":57531145,"AnswerCount":1,"Tags":"<python><keras><jupyter-notebook>","CreationDate":"2019-08-16T21:07:45.390","AcceptedAnswerId":57531251.0,"OwnerUserId":5040457.0,"Title":"Retrieve status of model training after closing notebook","Body":"<p>I am training a deep learning model with Keras from a Jupyter notebook. Is there a way to retrieve the status bar showing how many epochs have been completed if I close the notebook?<\/p>\n","answers":[{"AnswerId":"57531251","CreationDate":"2019-08-16T21:17:29.460","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>No, that information is not saved anywhere by default, it could only work if you save this information manually, or if you use something like the CSVLogger callback during training, as covered in the keras <a href=\"https:\/\/keras.io\/callbacks\/#csvlogger\" rel=\"nofollow noreferrer\">docs<\/a>.<\/p>\n"}]}
+{"QuestionId":57531365,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-16T21:31:43.540","AcceptedAnswerId":null,"OwnerUserId":3287820.0,"Title":"Tensorflow Keras RuntimeError: Graph is finalized and cannot be modified. [Minimal working example provided]","Body":"<p>I'm somewhat confused about what I'm seeing with my pretrained keras models. I'm using a virtualenv with tensorflow-gpu=1.13.1 installed via <code>pip install tensorflow-gpu<\/code>. Here's a minimal working example you can run, <a href=\"https:\/\/keras.io\/applications\/\" rel=\"nofollow noreferrer\">based on the keras documentation<\/a> (hopefully these are updated). In addition, I <a href=\"https:\/\/mondrian.mashable.com\/uploads%252Fcard%252Fimage%252F887009%252F49d251da-d190-49d8-80ad-5f12221574b4.jpg%252F950x534__filters%253Aquality%252880%2529.jpg?signature=kgnqs9P47PJJetZVmw6IxI96WmE=&amp;source=https%3A%2F%2Fblueprint-api-production.s3.amazonaws.com\" rel=\"nofollow noreferrer\">got the elephant image from here<\/a> and saved it as <code>elephant.jpeg<\/code>. <\/p>\n\n<pre><code>from tensorflow.keras.applications.resnet50 import ResNet50\nfrom tensorflow.keras.preprocessing import image\nfrom tensorflow.keras.applications.resnet50 import preprocess_input, decode_predictions\nimport numpy as np\n\n# Load the image.\nimg_path = 'data\/elephant.jpeg'\nimg = image.load_img(img_path, target_size=(224, 224))\nx = image.img_to_array(img)\nx = np.expand_dims(x, axis=0)  # shape (1,224,224,3)\nx = preprocess_input(x)\n\n# The basic full model\nmodel = ResNet50(weights='imagenet')\n\n# Make a session here\nsess = tf.Session()\nsess.graph.finalize()\n\n# Predict, and decode the results into a list of tuples (class, description,\n# probability) (one such list for each sample in the batch)\npreds = model.predict(x)\nprint('Predicted:', decode_predictions(preds, top=4)[0])\n<\/code><\/pre>\n\n<p>Running the code will result in:<\/p>\n\n<pre><code>RuntimeError: Graph is finalized and cannot be modified.\n<\/code><\/pre>\n\n<p>Yet the weird thing is that if I change the code to insert a <em>second<\/em> <code>model.predict<\/code> before I finalize the graph, as in here:<\/p>\n\n<pre><code>from tensorflow.keras.applications.resnet50 import ResNet50\nfrom tensorflow.keras.preprocessing import image\nfrom tensorflow.keras.applications.resnet50 import preprocess_input, decode_predictions\nimport numpy as np\n\n# Load the image.\nimg_path = 'data\/elephant.jpeg'\nimg = image.load_img(img_path, target_size=(224, 224))\nx = image.img_to_array(img)\nx = np.expand_dims(x, axis=0)  # shape (1,224,224,3)\nx = preprocess_input(x)\n\n# The basic full model\nmodel = ResNet50(weights='imagenet')\npreds = model.predict(x)\nprint('Predicted:', decode_predictions(preds, top=4)[0])\n\n# Make a session here\nsess = tf.Session()\nsess.graph.finalize()\n\n# Predict, and decode the results into a list of tuples (class, description,\n# probability) (one such list for each sample in the batch)\npreds = model.predict(x)\nprint('Predicted:', decode_predictions(preds, top=4)[0])\n<\/code><\/pre>\n\n<p>In the above there are only two extra lines, I copied the prediction code above with the print statement. This output seems to work and results in both predictions working:<\/p>\n\n<pre><code>Predicted: [('n01871265', 'tusker', 0.5286887), ('n02504013', 'Indian_elephant', 0.4639527), ('n02504458', 'African_elephant', 0.0072972253), ('n02408429', 'water_buffalo', 2.6213302e-05)]\nPredicted: [('n01871265', 'tusker', 0.5286887), ('n02504013', 'Indian_elephant', 0.4639527), ('n02504458', 'African_elephant', 0.0072972253), ('n02408429', 'water_buffalo', 2.6213302e-05)]\n<\/code><\/pre>\n\n<p>Here's why I'm confused and asking this question. I don't get why putting the predict code is necessary before a <code>sess.graph.finalize()<\/code> call. I am hoping to use pretrained models solely for feature extraction. That is, I'll pass in a numpy array into the net, and get a numpy array back. (For this I'd have to use the extra <code>include_top=True<\/code> but I didn't do that above for the sake of simplicity.) Then I want to pass this result to a new network that I design, using low-level tensorflow libraries.<\/p>\n\n<p>It appears that before my <code>sess.graph.finalize()<\/code> call, I need to insert a \"dummy\" prediction call beforehand to \"get the graph set up.\" Is that intuition right?<\/p>\n","answers":[{"AnswerId":"57531498","CreationDate":"2019-08-16T21:44:42.650","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>That's because <code>predict = model.predict(image)<\/code> is being treated as a node in the graph, and once you finalize the graph, you can't modify it further. However, you don't really need a <code>tf.Session()<\/code> in this case, your example can run solely like this:<\/p>\n\n<pre><code>x = np.random.rand(1, 224,224,3)\n# The basic full model\nmodel = ResNet50(weights='imagenet')\npreds = model.predict(x)\n\n# Predict, and decode the results into a list of tuples (class, description,\n# probability) (one such list for each sample in the batch)\n\nprint('Predicted:', decode_predictions(preds, top=4)[0]) \n<\/code><\/pre>\n\n<p>If you want to read more about it, I assume that <a href=\"https:\/\/stackoverflow.com\/questions\/51107527\/integrating-keras-model-into-tensorflow\">this answer<\/a> should be really helpful for you. Finally, if you really insist on having a <code>tf.Session()<\/code>, and finalize the graph with it, this should do it for you:<\/p>\n\n<pre><code>x = np.random.rand(1, 224,224,3)\n# The basic full model\nmodel = ResNet50(weights='imagenet')\npreds = model.predict(x)\n\nwith tf.keras.backend.get_session() as sess:\n    K.set_session(sess)\n    sess.graph.finalize()\n    print('Predicted:', decode_predictions(preds, top=4)[0])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57531409,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-08-16T21:36:16.647","AcceptedAnswerId":null,"OwnerUserId":11354575.0,"Title":"How is learning rate decay implemented by Adam in keras","Body":"<p>I set learning rate decay in my optimizer Adam, such as <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>LR = 1e-3\nLR_DECAY = 1e-2\nOPTIMIZER = Adam(lr=LR, decay=LR_DECAY)\n<\/code><\/pre>\n\n<p>As the keras document <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/c10d24959b0ad615a21e671b180a1b2466d77a2b\/keras\/optimizers.py#L433\" rel=\"nofollow noreferrer\">Adam<\/a> states, after each epoch learning rate would be <\/p>\n\n<pre><code>lr = lr * (1. \/ (1. + self.decay * K.cast(self.iterations, K.dtype(self.decay))))\n<\/code><\/pre>\n\n<p>If I understand correctly, learning rate be like this,<\/p>\n\n<pre><code>lr = lr * 1 \/ ( 1 + num_epoch * decay)\n<\/code><\/pre>\n\n<p>But I don't see the learning rate decay come into effect, after seeing that printed out.\nIs there any problem when I use this ?<\/p>\n\n<p><strong>Edit<\/strong><br>\nI print out the learning by setting the verbose of the <code>ReduceLROnPlateau<\/code>, such as,  <\/p>\n\n<pre><code>ReduceLROnPlateau(monitor='val_loss', factor=0.75, patience=Config.REDUCE_LR_PATIENCE, verbose=1, mode='auto', epsilon=0.01, cooldown=0, min_lr=1e-6\n<\/code><\/pre>\n\n<p>And that would monitor the val-loss and reduce the learning rate by multiplying the <code>factor<\/code>.\nThe printed learning rate is like this, <\/p>\n\n<pre><code>Epoch 00003: ReduceLROnPlateau reducing learning rate to 0.0007500000356230885.\n<\/code><\/pre>\n\n<p>And I set the initial learning rate to be 1e-3. Therefore, it appears that the learning rate change from 1e-3 to 1e-3 * 0.75, so I doubt that the <code>decay<\/code> I set in Adam isn't working.<\/p>\n","answers":[{"AnswerId":"58745976","CreationDate":"2019-11-07T09:47:16.023","ParentId":null,"OwnerUserId":"6467554","Title":null,"Body":"<p>The learning rate changes with every iteration, i.e., with every batch and not epoch. So, if you set the decay = 1e-2 and each epoch has 100 batches\/iterations, then after 1 epoch your learning rate will be<\/p>\n\n<pre><code>lr = init_lr * 1\/(1 + 1e-2 * 100)\n<\/code><\/pre>\n\n<p>So, if I want my learning rate to be 0.75 of the original learning rate at the end of each epoch, I would set the lr_decay to <\/p>\n\n<pre><code>batches_per_epoch = dataset_size\/batch_size\nlr_decay = (1.\/0.75 -1)\/batches_per_epoch\n<\/code><\/pre>\n\n<p>It seems to work for me. Also, since the new learning rate is calculated at every iteration, the optimizer doesn't change the value of the learning rate variable and always uses the initial learning rate to calculate the effective learning rate. <\/p>\n"}]}
+{"QuestionId":57531739,"AnswerCount":0,"Tags":"<gpu><pytorch>","CreationDate":"2019-08-16T22:20:12.250","AcceptedAnswerId":null,"OwnerUserId":10855529.0,"Title":"Using GPU takes even longer time for broadcasting in PyTorch","Body":"<pre><code>print(\"CUDA available: \", torch.cuda.is_available())\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n<\/code><\/pre>\n\n<p>CUDA available:  True<\/p>\n\n<pre><code>x = torch.tensor([1, 2, 3]).to(device)\ny = torch.tensor([1, 2, 3])\n<\/code><\/pre>\n\n<pre><code>%%timeit -n 10000\nz = x + 1\n<\/code><\/pre>\n\n<p>10000 loops, best of 3: 15.6 \u00b5s per loop<\/p>\n\n<pre><code>%%timeit -n 10000\nz = y + 1\n<\/code><\/pre>\n\n<p>10000 loops, best of 3: 5.19 \u00b5s per loop<\/p>\n","answers":[]}
+{"QuestionId":57532296,"AnswerCount":0,"Tags":"<tensorflow><padding>","CreationDate":"2019-08-16T23:59:16.577","AcceptedAnswerId":null,"OwnerUserId":7943233.0,"Title":"Tensoflow - Pad a tensor with a different padding mode?","Body":"<p>Is there a simple way to pad a tensor in Tensorflow with a scheme different than the 3 given modes (CONSTANT, REFLECT and SYMMETRIC) in tf.pad() ?  <\/p>\n\n<p>What I want to do is just repeat the tensor from the other side like in the following example:<br>\nFrom  <\/p>\n\n<pre><code>[1, 2, 3]  \n[4, 5, 6]  \n[7, 8, 9] \n<\/code><\/pre>\n\n<p>I need to get  <\/p>\n\n<pre><code>[9, 7, 8, 9, 7]  \n[3, 1, 2, 3, 1]  \n[6, 4, 5, 6, 4]  \n[9, 7, 8, 9, 7]  \n[3, 1, 2, 3, 1]\n<\/code><\/pre>\n\n<p>I need to pad it in only one dimension so I can slice the first few rows and concatenate the new slice with the original tensor, but I was wondering if there is a more elegant and less memory-intensive way to do it.<\/p>\n\n<p>I just need this for a convolution, so any suggestions for setting up a convolutional layer to do this while training are also welcome.<\/p>\n\n<p>This is my first post here so I accept any criticism about it :)<\/p>\n","answers":[]}
+{"QuestionId":57532416,"AnswerCount":0,"Tags":"<python><tensorflow><keras><openai-gym><atari-2600>","CreationDate":"2019-08-17T00:30:02.843","AcceptedAnswerId":null,"OwnerUserId":9689105.0,"Title":"How to run and render gym Atari environments in real time, instead of sped up?","Body":"<p>Currently when I render any Atari environments they are always sped up, and I want to look at them in normal speed. I am using Gym Atari with Tensorflow, and Keras-rl on Windows. The language is python.<\/p>\n","answers":[]}
+{"QuestionId":57532425,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><python-3.7><recurrent-neural-network><gated-recurrent-unit>","CreationDate":"2019-08-17T00:31:20.287","AcceptedAnswerId":null,"OwnerUserId":9477653.0,"Title":"How to determine the correct dimension of additive matrix for custom Tensorflow optimizer","Body":"<p>I am writing a custom optimizer for a machine learning project using gated recurrent units in Tensorflow. The update protocol for this optimizer is a variation of the Adam optimizer (described here: <a href=\"https:\/\/arxiv.org\/abs\/1609.04747\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1609.04747<\/a>), except that the update protocol adds another parameter psi such that it is defined as<\/p>\n\n<p>theta_t+1 = theta_t - ((eta \/ (sqrt(v_t) + eps)) * m_t) + ((1 - loss) * psi). I understand that each GRU has a weight matrix for each phase, but I am wondering if the variable update algorithm uses the shape of the weight matrices for one GRU at a time, or if it takes the weight matrices for all GRUs and processes them as sub-arrays of a single array. In other words, I am trying to determine the correct shape of psi based on the architecture of my RNN which is set up as follows: <\/p>\n\n<p><code>zt = tf.compat.v1.get_variable(\"taskname\", [1, 100])<\/code><\/p>\n\n<p><code>z = tf.tile(zt, [500, 1])<\/code><\/p>\n\n<p><code>z_mat = tf.layers.dense(inputs=z, units=100, activation=tf.nn.relu, name=\"wMat_taskname\")<\/code><\/p>\n\n<p><code>h1 = tf.stack([z_mat for _ in range(9)], axis=1)<\/code><\/p>\n\n<p><code>cell = tf.contrib.rnn.MultiRNNCell([tf.contrib.rnn.GRUCell(100) for _ in range(3)], state_is_tuple=False)<\/code><\/p>\n\n<p><code>outputs, state = tf.nn.dynamic_rnn(cell=cell, inputs=h1, dtype=tf.float32)<\/code><\/p>\n\n<p><code>h2 = tf.layers.dense(tf.reshape(outputs, [500*9, 100]), units=512, activation=tf.nn.softmax, name=\"output\")<\/code><\/p>\n\n<p>Based on this, I would like to clarify what the shape of psi should be.<\/p>\n","answers":[]}
+{"QuestionId":57532444,"AnswerCount":0,"Tags":"<c++><tensorflow><bazel>","CreationDate":"2019-08-17T00:35:12.077","AcceptedAnswerId":null,"OwnerUserId":6794255.0,"Title":"What do I need to build Tensorflow (from source) for c++?","Body":"<p>I want to build Tensorflow (GPU) from source for c++ but I always get errors when I try the build the pip package step. What do I need to make this step work?<\/p>\n\n<p>I'm on Windows 10 but I have Cygwin if I need it. I am building Tensorflow for use with my GPU because I have a decent one so why not. The only advice I can seem to find is to switch versions of Bazel but I've already tried many different versions from 0.18.0 to the latest but I'm using 0.24.1 now because that seems to be the recommended version for Tensorflow 1.14. I am using CUDA 10.1 however but I don't see how that would cause this problem.<\/p>\n\n<p>This is the command that I cannot get to work:<\/p>\n\n<pre><code>bazel build --config=opt --config=cuda \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>I'm not sure what information is most necessary so here is the entire output. I've had different variations of errors but this is what it gives me now:<\/p>\n\n<pre><code>WARNING: The following configs were expanded more than once: [cuda, using_cuda]. For repeatable flags, repeats are counted twice and may lead to unexpected behavior.\nINFO: Options provided by the client:\n  Inherited 'common' options: --isatty=1 --terminal_columns=70\nINFO: Options provided by the client:\n  'build' options: --python_path=C:\/Users\/Invilis Bella\/AppData\/Local\/Programs\/Python\/Python37-32\/python.exe\nINFO: Reading rc options for 'build' from b:\\tensorflow\\tensorflow\\.bazelrc:\n  'build' options: --apple_platform_type=macos --define framework_shared_object=true --define open_source_build=true --define=use_fast_cpp_protos=true --define=allow_oversize_protos=true --spawn_strategy=standalone --strategy=Genrule=standalone -c opt --announce_rc --define=grpc_no_ares=true --define=PREFIX=\/usr --define=LIBDIR=$(PREFIX)\/lib --define=INCLUDEDIR=$(PREFIX)\/include\nINFO: Reading rc options for 'build' from b:\\tensorflow\\tensorflow\\.tf_configure.bazelrc:\n  'build' options: --action_env PYTHON_BIN_PATH=C:\/Users\/Invilis Bella\/AppData\/Local\/Programs\/Python\/Python37-32\/python.exe --action_env PYTHON_LIB_PATH=C:\/Users\/Invilis Bella\/AppData\/Local\/Programs\/Python\/Python37-32\/lib\/site-packages --python_path=C:\/Users\/Invilis Bella\/AppData\/Local\/Programs\/Python\/Python37-32\/python.exe --action_env CUDA_TOOLKIT_PATH=C:\/Program Files\/NVIDIA GPU Computing Toolkit\/CUDA\/v10.1 --action_env TF_CUDA_COMPUTE_CAPABILITIES=6.1 --config=cuda --config monolithic --copt=-w --host_copt=-w --copt=-DWIN32_LEAN_AND_MEAN --host_copt=-DWIN32_LEAN_AND_MEAN --copt=-DNOGDI --host_copt=-DNOGDI --verbose_failures --distinct_host_configuration=false --define=override_eigen_strong_inline=true --action_env TF_CONFIGURE_IOS=0\nINFO: Found applicable config definition build:cuda in file b:\\tensorflow\\tensorflow\\.bazelrc: --config=using_cuda --define=using_cuda_nvcc=true\nINFO: Found applicable config definition build:using_cuda in file b:\\tensorflow\\tensorflow\\.bazelrc: --define=using_cuda=true --action_env TF_NEED_CUDA=1 --crosstool_top=@local_config_cuda\/\/crosstool:toolchain\nINFO: Found applicable config definition build:monolithic in file b:\\tensorflow\\tensorflow\\.bazelrc: --define framework_shared_object=false\nINFO: Found applicable config definition build:opt in file b:\\tensorflow\\tensorflow\\.tf_configure.bazelrc: --copt=\/arch:AVX2 --define with_default_optimizations=true\nINFO: Found applicable config definition build:cuda in file b:\\tensorflow\\tensorflow\\.bazelrc: --config=using_cuda --define=using_cuda_nvcc=true\nINFO: Found applicable config definition build:using_cuda in file b:\\tensorflow\\tensorflow\\.bazelrc: --define=using_cuda=true --action_env TF_NEED_CUDA=1 --crosstool_top=@local_config_cuda\/\/crosstool:toolchain\nERROR: error loading package '': Encountered error while reading extension file 'repositories\/repositories.bzl': no such package '@io_bazel_rules_docker\/\/repositories': Traceback (most recent call last):\n        File \"C:\/users\/invilis bella\/_bazel_invilis bella\/ldshq52a\/external\/bazel_tools\/tools\/build_defs\/repo\/git.bzl\", line 164\n                _clone_or_update(ctx)\n        File \"C:\/users\/invilis bella\/_bazel_invilis bella\/ldshq52a\/external\/bazel_tools\/tools\/build_defs\/repo\/git.bzl\", line 74, in _clone_or_update\n                fail((\"error cloning %s:\\n%s\" % (ctx....)))\nerror cloning io_bazel_rules_docker:\n\/usr\/bin\/bash: line 1: cd: too many arguments\n+ cd C:\/users\/invilis bella\/_bazel_invilis bella\/ldshq52a\/external\n\/usr\/bin\/bash: line 3: cd: too many arguments\n+ git -C 'C:\/users\/invilis bella\/_bazel_invilis bella\/ldshq52a\/external\/io_bazel_rules_docker' reset --hard 251f6a68b439744094faff800cd029798edf9faa\nfatal: not a git repository (or any of the parent directories): .git\n+ git -C 'C:\/users\/invilis bella\/_bazel_invilis bella\/ldshq52a\/external\/io_bazel_rules_docker' fetch origin 251f6a68b439744094faff800cd029798edf9faa:251f6a68b439744094faff800cd029798edf9faa\nfatal: not a git repository (or any of the parent directories): .git\n+ git -C 'C:\/users\/invilis bella\/_bazel_invilis bella\/ldshq52a\/external\/io_bazel_rules_docker' fetch origin 251f6a68b439744094faff800cd029798edf9faa:251f6a68b439744094faff800cd029798edf9faa\nfatal: not a git repository (or any of the parent directories): .git\nERROR: error loading package '': Encountered error while reading extension file 'repositories\/repositories.bzl': no such package '@io_bazel_rules_docker\/\/repositories': Traceback (most recent call last):\n        File \"C:\/users\/invilis bella\/_bazel_invilis bella\/ldshq52a\/external\/bazel_tools\/tools\/build_defs\/repo\/git.bzl\", line 164\n                _clone_or_update(ctx)\n        File \"C:\/users\/invilis bella\/_bazel_invilis bella\/ldshq52a\/external\/bazel_tools\/tools\/build_defs\/repo\/git.bzl\", line 74, in _clone_or_update\n                fail((\"error cloning %s:\\n%s\" % (ctx....)))\nerror cloning io_bazel_rules_docker:\n\/usr\/bin\/bash: line 1: cd: too many arguments\n+ cd C:\/users\/invilis bella\/_bazel_invilis bella\/ldshq52a\/external\n\/usr\/bin\/bash: line 3: cd: too many arguments\n+ git -C 'C:\/users\/invilis bella\/_bazel_invilis bella\/ldshq52a\/external\/io_bazel_rules_docker' reset --hard 251f6a68b439744094faff800cd029798edf9faa\nfatal: not a git repository (or any of the parent directories): .git\n+ git -C 'C:\/users\/invilis bella\/_bazel_invilis bella\/ldshq52a\/external\/io_bazel_rules_docker' fetch origin 251f6a68b439744094faff800cd029798edf9faa:251f6a68b439744094faff800cd029798edf9faa\nfatal: not a git repository (or any of the parent directories): .git\n+ git -C 'C:\/users\/invilis bella\/_bazel_invilis bella\/ldshq52a\/external\/io_bazel_rules_docker' fetch origin 251f6a68b439744094faff800cd029798edf9faa:251f6a68b439744094faff800cd029798edf9faa\nfatal: not a git repository (or any of the parent directories): .git\nINFO: Elapsed time: 0.511s\nINFO: 0 processes.\nFAILED: Build did NOT complete successfully (0 packages loaded)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57532479,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-c++>","CreationDate":"2019-08-17T00:45:05.117","AcceptedAnswerId":null,"OwnerUserId":936332.0,"Title":"Can tensorflow consume GPU memory exactly equal to required","Body":"<p>I use tensorflow c++ version to do CNN inference. I already <code>set_allow_growth(true)<\/code>, but it still consume more GPU memory than exactly need . <\/p>\n\n<p><code>set_per_process_gpu_memory_fraction<\/code> can only set an upper bound of the GPU memory, but different CNN model have different upper bound. <strong>Is there a good way to solve the problem<\/strong><\/p>\n","answers":[]}
+{"QuestionId":57532661,"AnswerCount":1,"Tags":"<pytorch><dataloader>","CreationDate":"2019-08-17T01:30:13.123","AcceptedAnswerId":57533806.0,"OwnerUserId":11151394.0,"Title":"How do they know mean and std, the input value of transforms.Normalize","Body":"<p>The question is about <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/data_loading_tutorial.html\" rel=\"nofollow noreferrer\">the data loading tutorial<\/a> from the PyTorch website. I don't know how they write the value of <code>mean_pix<\/code> and <code>std_pix<\/code> of the in transforms.Normalize without calculation<\/p>\n\n<p>I'm unable to find any explanation relevant to this question on StackOverflow.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nfrom torchvision import transforms, datasets\n\ndata_transform = transforms.Compose([\n        transforms.RandomSizedCrop(224),\n        transforms.RandomHorizontalFlip(),\n        transforms.ToTensor(),\n        transforms.Normalize(mean=[0.485, 0.456, 0.406],\n                             std=[0.229, 0.224, 0.225])\n    ])\nhymenoptera_dataset = datasets.ImageFolder(root='hymenoptera_data\/train',\n                                           transform=data_transform)\ndataset_loader = torch.utils.data.DataLoader(hymenoptera_dataset,\n                                             batch_size=4, shuffle=True,\n                                             num_workers=4)\n<\/code><\/pre>\n\n<p>The value <code>mean=[0.485,0.456, 0.406]<\/code> and <code>std=[0.229, 0.224, 0.225]<\/code> is not obvious to me. How do they get them? And why are they equal to these?<\/p>\n","answers":[{"AnswerId":"57533806","CreationDate":"2019-08-17T06:23:27.967","ParentId":null,"OwnerUserId":"6210807","Title":null,"Body":"<p>For normalization <code>input[channel] = (input[channel] - mean[channel]) \/ std[channel]<\/code>, the mean and standard deviation values are to be taken from the <strong>training<\/strong> dataset.<\/p>\n\n<p>Here, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225] are the mean and std of Imagenet dataset.<\/p>\n\n<blockquote>\n  <p>On Imagenet, we\u2019ve done a pass on the dataset and calculated per-channel mean\/std.  <a href=\"https:\/\/discuss.pytorch.org\/t\/normalization-in-the-mnist-example\/457\/7?u=kharshit\" rel=\"nofollow noreferrer\">check here<\/a><\/p>\n<\/blockquote>\n\n<p>The pre-trained models available in <a href=\"https:\/\/pytorch.org\/docs\/stable\/torchvision\/models.html#torchvision-models\" rel=\"nofollow noreferrer\"><code>torchvision<\/code><\/a> for transfer learning were pretrained on Imagenet, so using its mean and std deviation would be fine for fine-tuning your model.<\/p>\n\n<p>If you're trying to train your model from scratch, it would be better to use the mean and std deviation of your training dataset (face dataset in this case). Other than that, in most of the cases, the mean and std of Imagenet <a href=\"https:\/\/discuss.pytorch.org\/t\/about-normalization-using-pre-trained-vgg16-networks\/23560\/2?u=kharshit\" rel=\"nofollow noreferrer\">suffice<\/a> for your problem.<\/p>\n"}]}
+{"QuestionId":57532669,"AnswerCount":0,"Tags":"<amazon-web-services><opencv><tensorflow><project><amazon-rekognition>","CreationDate":"2019-08-17T01:32:04.113","AcceptedAnswerId":null,"OwnerUserId":11581301.0,"Title":"Detecting objects from a particular boundary box-Help for my college project","Body":"<p>Iam a student and i need to complete a project before next week.my application is smart parking.As of now i was able to detect cars using opencv and also by using amazon rekognition and also Iam learning about tensorflow object detection techniques.But I want to detect the car with in a boundary box and if the car is detected the colour of the boundary box should be red and if it is not detected the colour of the boundary box should be green.This is my application and i really dont know how to do this.It will be very very helpfull for my studues if you guys help me.Thanks in advance!<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/64sYS.jpg\" rel=\"nofollow noreferrer\">sample images taken from internet<\/a><\/p>\n","answers":[]}
+{"QuestionId":57532795,"AnswerCount":0,"Tags":"<tensorflow><yolo>","CreationDate":"2019-08-17T02:07:31.777","AcceptedAnswerId":null,"OwnerUserId":8893068.0,"Title":"How to load yolo voc weights in tensorflow?","Body":"<p>I have written tiny yolo v2 in tensorflow and wanted to load it with pre-trained weights. The weights were loaded successfully but predictions are very poor. Have I done some mistake in loading weights?<\/p>\n\n<p>Here is the link to my code:\n<a href=\"https:\/\/github.com\/Ajaz-Ahmad\/YOLO-WeightLoader\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Ajaz-Ahmad\/YOLO-WeightLoader<\/a><\/p>\n\n<p>Good accuracy on trained classes<\/p>\n","answers":[]}
+{"QuestionId":57533005,"AnswerCount":0,"Tags":"<keras><deep-learning>","CreationDate":"2019-08-17T03:18:17.470","AcceptedAnswerId":null,"OwnerUserId":8515794.0,"Title":"How to select the model for checking the test error?","Body":"<p>I am training a model for my dataset and I have a question regarding how to select the model for checking the test error. I am saving the models which give lowest validation loss (during training) and I use the one that gives the lowest validation loss for testing (this model can be somewhere around end of training). Is this the right approach or should I select the model that I get at last epoch (after all epochs have finished)?<\/p>\n\n<p>Edit: I have also observed that the best validation loss resulting model does not always provide lowest test error. <\/p>\n","answers":[]}
+{"QuestionId":57533282,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><conv-neural-network>","CreationDate":"2019-08-17T04:32:27.967","AcceptedAnswerId":null,"OwnerUserId":11938681.0,"Title":"Why do I get this error and what is the solution?","Body":"<p>I just started Tensorflow and am solving this problem but I am getting errors. The problem is that the base price for a house is 50k and each bedroom costs 50k each. So a 1 bedroom house is 100k, 2 bedroom is 150k and so on. We have to predict the cost of a 7 Bedroom house.<\/p>\n\n<p>I have tried using 'import numpy as np' and also 'import numpy' but error still remains.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy\nfrom tensorflow import keras\n\ndef house_model(y_new):\n    xs = numpy.array([1.0,2.0,3.0,4.0], dtype = float)\n    ys = numpy.array([1.0,1.5,2.0,2.5], dtype = float)\n    model = tf.keras.Sequential([keras.layers.dense(units=1, input_shape=[1])])\n    model.compile(optimizer = 'sgd', loss = 'mean_squared_error')\n    model.fit(xs,ys,epochs = 500)\n    return model.predict(y_new)[7.0]\n\nprediction = house_model([7.0])\nprint(prediction)\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>---------------------------------------------------------------------------\nNameError                                 Traceback (most recent call last)\n&lt;ipython-input-23-55d468d60746&gt; in &lt;module&gt;\n----&gt; 1 prediction = house_model([7.0])\n      2 print(prediction)\n\n&lt;ipython-input-20-0e67265afcf6&gt; in house_model(y_new)\n      1 def house_model(y_new):\n----&gt; 2     xs = numpy.array([1.0,2.0,3.0,4.0], dtype = float)\n      3     ys = numpy.array([1.0,1.5,2.0,2.5], dtype = float)\n      4     model = tf.keras.Sequential([keras.layers.dense(units=1, input_shape=[1])])\n      5     model.compile(optimizer = 'sgd', loss = 'mean_squared_error')\n\nNameError: name 'numpy' is not defined\n<\/code><\/pre>\n","answers":[{"AnswerId":"57533652","CreationDate":"2019-08-17T05:50:43.993","ParentId":null,"OwnerUserId":"6149357","Title":null,"Body":"<p>It seems like numpy isn't installed\/present in your python interpreter's library path. I think you need to either install the Numpy module to use it or check if the correct path is getting used by your interpreter.<\/p>\n"}]}
+{"QuestionId":57533636,"AnswerCount":0,"Tags":"<keras><loss-function><semantic-segmentation>","CreationDate":"2019-08-17T05:48:55.847","AcceptedAnswerId":null,"OwnerUserId":3177661.0,"Title":"How to implement custom loss and accuracy metric for ignoring void labels in keras semantic segmenation","Body":"<p>I'am trying to develop a semantic segmentaion model in keras.The labels have three different values viz. <strong>0 -background, 1- foreground and 2- missing or void label<\/strong>.Since we are interested in two classes only, we don't want to compute loss over void label(2) i.e we should ignore them.<\/p>\n\n<p>Initially i tried a custom loss and acccuracy metric for masking out the void labels.However after 50 epochs loss (<strong>custom binary cross entropy<\/strong>) becomes <strong>negative<\/strong>, even though accuracy increases.The last activation layer(<strong>sigmoid<\/strong>) has shape <strong>(None, 128,128,1)<\/strong>.The corresponding target labels with same shape has values <strong>0,1,2 (uint8)<\/strong> and input images(not mask) are <strong>normalized<\/strong>(\/255.0) float(0-1). Optimizer used - <strong>Adam(lr=1e-3)<\/strong>.<\/p>\n\n<p>Here is the custom loss and metric...<\/p>\n\n<pre><code>def masked_loss_function(y_true, y_pred):\n  mask_value=2\n  print(y_true.shape)\n  print(y_pred.shape)\n  mask = K.cast(K.not_equal(y_true, mask_value), K.floatx())\n  print(mask.shape)\n  return K.binary_crossentropy(y_true * mask, y_pred * mask)\n\ndef masked_accuracy(y_true, y_pred):\n  mask_value=2\n  dtype = K.floatx()\n  total = K.sum(K.cast(K.not_equal(y_true, mask_value), dtype))\n  correct = K.sum(K.cast(K.equal(y_true, K.round(y_pred)), dtype))\n  return correct \/ total\n<\/code><\/pre>\n\n<p>Additionaly i also performed image augmentations using keras generators on images and masks(same seed) ..<\/p>\n\n<pre><code>data_gen_args = dict(\n                 width_shift_range=0.2,\n                 height_shift_range=0.2,\n                 zoom_range=0.2,\n                 horizontal_flip=True,\n                 validation_split=0.2\n                )\n<\/code><\/pre>\n\n<p>It seems there is some issue with masking, since the loss is becoming negative , even though we try to mask out void(2) label as show above.<\/p>\n\n<p>UPDATE:-<\/p>\n\n<p>It seems the generator somehow modifies the mask if we include zoom and width shift options.<\/p>\n\n<p>Here is the full generator(after i changed label to float values 0.0 and 255.0 for testing)..<\/p>\n\n<pre><code>    data_gen_args = dict(rescale=1.\/255,\n                     width_shift_range=0.1,\n                     height_shift_range=0.1,\n                     zoom_range=0.2,\n                     horizontal_flip=True,\n                     validation_split=0.2\n                    )\n\nimage_datagen = ImageDataGenerator(**data_gen_args) \n\n# Provide the same seed and keyword arguments to the fit and flow methods\nseed = 1\nbatch_sz=32\n\ndef generate_train_generator(image_datagen):\n    genX1 = image_datagen.flow(\n    x_train,\n    batch_size=batch_sz,\n    shuffle=True,\n    subset='training',\n    seed=seed)\n\n    genY1 = image_datagen.flow(\n    y_train,\n    batch_size=batch_sz,\n    shuffle=True,\n    subset='training',\n    seed=seed)\n\n    while True:\n            Xi = genX1.next()\n            Yi = genY1.next()\n            print(Yi.dtype)\n            print(np.unique(Yi))\n            yield (Xi, Yi)\n\n\ntrain_generator=generate_train_generator(image_datagen)\n<\/code><\/pre>\n\n<p>It outputs label values (unique) as :-<strong><em>[0.0000000e+00 2.0307664e-06 1.0561990e-05 ... 9.9996912e-01 9.9997061e-01 1.0000000e+00]<\/em><\/strong> if include width shift; without shifts and zoom it outputs <strong>[0. 1.]<\/strong>.<\/p>\n\n<p>What could be the reason for value modifications??\n(Using Keras 2.2.4 on google colab)<\/p>\n","answers":[]}
+{"QuestionId":57533788,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2019-08-17T06:18:43.137","AcceptedAnswerId":null,"OwnerUserId":1830648.0,"Title":"Trying to understand the output of model.predict","Body":"<p>I am trying to predict the cat\/dog image using the model.predict method. Since it is 2 class classifier I am getting an array of 2 values. As per my understanding, these values represent the probabilities to belong in each class (correct me if I am wrong). If it is so the probability must be summed to 1. But I am getting the same probability for both classes <\/p>\n\n<p>Model History<\/p>\n\n<pre><code>Model: \"sequential\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nflatten (Flatten)            multiple                  0         \n_________________________________________________________________\ndense (Dense)                multiple                  30848     \n_________________________________________________________________\ndropout (Dropout)            multiple                  0         \n_________________________________________________________________\nbatch_normalization (BatchNo multiple                  512       \n_________________________________________________________________\ndense_1 (Dense)              multiple                  12900     \n_________________________________________________________________\ndropout_1 (Dropout)          multiple                  0         \n_________________________________________________________________\nbatch_normalization_1 (Batch multiple                  400       \n_________________________________________________________________\ndense_2 (Dense)              multiple                  10100     \n_________________________________________________________________\ndropout_2 (Dropout)          multiple                  0         \n_________________________________________________________________\nbatch_normalization_2 (Batch multiple                  400       \n_________________________________________________________________\ndense_3 (Dense)              multiple                  10100     \n_________________________________________________________________\ndropout_3 (Dropout)          multiple                  0         \n_________________________________________________________________\nbatch_normalization_3 (Batch multiple                  400       \n_________________________________________________________________\ndense_4 (Dense)              multiple                  10100     \n_________________________________________________________________\ndropout_4 (Dropout)          multiple                  0         \n_________________________________________________________________\nbatch_normalization_4 (Batch multiple                  400       \n_________________________________________________________________\ndense_5 (Dense)              multiple                  202       \n=================================================================\nTotal params: 76,362\nTrainable params: 75,306\nNon-trainable params: 1,056\n<\/code><\/pre>\n\n<hr>\n\n<p>Prediction Code<\/p>\n\n<pre><code>class_prob=model.predict(new_array_2.T,batch_size=1)\nprint(class_prob)\nclassifications=model.predict_classes(new_array_2.T,batch_size=1)\nprint(classifications)\nprint(CATEGORIES[classifications[0]])\n<\/code><\/pre>\n\n<p>output<\/p>\n\n<pre><code>[[0.39456758 0.39456758]]\n[0]\nDog\n<\/code><\/pre>\n","answers":[{"AnswerId":"57534566","CreationDate":"2019-08-17T08:24:05.527","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p><code>model.predict<\/code> just returns the computed output of your model given the input, so all the details you mentioned depend on the model's output, for example, activations at the last layer.<\/p>\n\n<p>That the model outputs probabilities that sum to one is only produced by the <code>softmax<\/code> activation at the output layer, to me it looks like your last layer has a different activation, probably <code>sigmoid<\/code>, which will not produce probabilities that sum to one. You should prefer to use <code>softmax<\/code> for multi-class classification.<\/p>\n"}]}
+{"QuestionId":57533942,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-08-17T06:44:19.110","AcceptedAnswerId":null,"OwnerUserId":9803875.0,"Title":"What exactly is compute_gradients returning and how does it depend on batch_size?","Body":"<p>Please excuse my rookie understanding of TensorFlow, Thanks in advance for the help!<\/p>\n\n<p>I am trying to compute the gradients using <code>compute_gradients()<\/code> wrt the embedding inputs of my loaded model.<\/p>\n\n<p>My batch_size is 250 and embd_size is 300.<\/p>\n\n<p>I want to compute the gradients of all my inputs for 250 test examples, \n so <code>predicted_y<\/code> is a numpy list of the values predicted by the model of shape <code>[250,1]<\/code> so the <code>x_test<\/code> that I provide in the <code>feed<\/code> dict is of shape <code>[250, 300]<\/code>.<\/p>\n\n<p>I have already tried this similar question \n<a href=\"https:\/\/stackoverflow.com\/questions\/44982081\/what-does-compute-gradients-return-in-tensorflow\">What does compute_gradients return in tensorflow<\/a>\nbut I didn't completely understand the role of batch_size in compute_gradients()<\/p>\n\n<pre><code>def get_gradients(model, predicted_y):\n\n    variables_fed = []\n    gradients_fe = []\n    inputs_here_fed = []\n\n    optimizer_here = model.gradients\n    inputs_here = model.inputs\n    embedding_here = model.embedding\n    cost_here = model.cost\n\n    print(len(predicted_y))\n\n\n    gradients, variables = zip(*optimizer_here.compute_gradients(cost_here, embedding_here))\n    print(\"gradients object: {}\".format(gradients[0]))\n\n    opt = optimizer_here.apply_gradients(list(zip(gradients, variables)))\n    # we do not have to run the optimizer as we do not want to BP\n\n    with tf.Session() as sess:\n\n        init = tf.global_variables_initializer()\n\n        sess.run(init)\n        test_state = sess.run(model.initial_state)\n\n        feed = {model.inputs: x_test[0:len(predicted_y)], # dims should match predicted_y\n                model.labels: predicted_y[:, None], #converting 1d to 2d array\n                model.keep_prob: dropout,\n                model.initial_state: test_state}\n\n        # test = sess.run(opt, feed_dict=feed)\n\n        gradients_fed = sess.run(gradients, feed_dict=feed)\n\n        # inputs_here_fed = sess.run(inputs_here, feed_dict=feed)\n        # variables_fed = sess.run(variables, feed_dict=feed)\n\n    return variables_fed, gradients_fed, inputs_here_fed\n\ndef get_gradients_values(gradients): # takes IndexedSlices Object which store gradients as input\n\n    l = gradients[0].values\n    print(\"Shape of gradients list: {}\".format(l.shape))\n\n    return l\n<\/code><\/pre>\n\n<p>After I feed the values in <code>sess.run(gradients, feed)<\/code>, I extract the values of the <code>IndexedSlices<\/code> object obtained and stored it as a list <code>grads<\/code>. I expected to get <code>grads<\/code> with dimentionality <code>[250, 300]<\/code>, corresponding to the gradients for all my inputs for each test example Instead I get <code>[50000, 300]<\/code> which I cannot explain.<\/p>\n\n<p>I tried varying the batch_size too see what happens but it gives me mismatch between input shapes error. I tried understanding the compute_gradients() code on github but its too obscure for someone with my basic understanding.<\/p>\n\n<p>How do I get the gradients of all inputs for each of my test set examples?<\/p>\n","answers":[{"AnswerId":"57543320","CreationDate":"2019-08-18T09:39:57.627","ParentId":null,"OwnerUserId":"9803875","Title":null,"Body":"<p>I found out that the dimensions that gradients returns from <code>compute_gradients()<\/code> is correct according to my <code>x_train<\/code> dimensions, which is <code>(250,200,300)<\/code> not <code>(250,300)<\/code> that I thought earlier. The computed <code>IndexedSlices<\/code> object has a <code>values<\/code> property which flattens the output to <code>(250*200,300)<\/code> and a <code>indices<\/code> property which is of same dims as <code>values<\/code> which gives us the index of each word in the vocabulary you used for tokenization. <\/p>\n\n<p>Also, while using <code>compute_gradients()<\/code> we need to make sure in the above code that <code>batch_size == len(predicted_y)<\/code> always, or it will give a shape mismatch error. To get gradients for one example, set <code>batch_size = 1<\/code> and make sure <code>predicted_y<\/code> only contains 1 test example. The gradients object obtained with these parameters will be of size <code>(200,300)<\/code> and to get the attribution\/gradient measure for each input we can use <code>np.sum(gradients, axis = 1)<\/code> to sum across the columns.<\/p>\n"}]}
+{"QuestionId":57533966,"AnswerCount":1,"Tags":"<dictionary><tensorflow><feed>","CreationDate":"2019-08-17T06:48:32.433","AcceptedAnswerId":null,"OwnerUserId":10561436.0,"Title":"How to feed data properly in tensorflow","Body":"<p>I have been learning Tensorflow and understanding feed_dict has been a challenge. Take for example the following piece of code i am working on<\/p>\n\n<pre><code>        p=0\n        self.sequence_length=25\n        with tf.Session() as sess:\n                                    init.run()\n                                    char_to_ix={ch:ix for ix,ch in enumerate(self.words)}\n                                    ix_to_char={ix:ch for ix,ch in enumerate(self.words)}\n                                    words_in_input=self.data[p:p+self.sequence_length]\n                                    inputs=[char_to_ix[ix] for ix in words_in_input]\n                                    words_in_target=self.data[p+1:p+self.sequence_length+1]\n                                    targets=[char_to_ix[ix] for ix in words_in_target]                        \n                                    onex=sess.run([selected_next_letter],feed_dict={self.X:inputs,self.y:targets})\n        p=p+1\n<\/code><\/pre>\n\n<p>This gives the error: Shapes of all inputs must match: values[0].shape = [25] != values[1].shape = []\nHowever, when I edit the code to<\/p>\n\n<pre><code>with tf.Session() as sess:\n                        init.run()\n                        char_to_ix={ch:ix for ix,ch in enumerate(self.words)}\n                        ix_to_char={ix:ch for ix,ch in enumerate(self.words)}\n                        words_in_input=self.data[p:p+self.sequence_length]\n                        inputs=[char_to_ix[ix] for ix in words_in_input]\n                        words_in_target=self.data[p+1:p+self.sequence_length+1]\n                        targets=[char_to_ix[ix] for ix in words_in_target]    \n                        for x,y in zip(inputs,targets):\n                            onex=sess.run([selected_next_letter],feed_dict={self.X:x,self.y:y})\n<\/code><\/pre>\n\n<p>It executes.\nMy questions is: Is it possible to feed the whole list such as <code>inputs<\/code> and <code>targets<\/code> in the <code>feed_dict<\/code> or must I input it  through a loop one by one. I ask this because the tutorials I have been reading, I see a whole list being passed in a feed_dict such as <\/p>\n\n<pre><code>loss_val = sess.run([train_op, loss_mean], feed_dict={\n                    images_batch:images_batch_val,\n                    labels_batch:labels_batch_val\n                    })\n<\/code><\/pre>\n","answers":[{"AnswerId":"57539211","CreationDate":"2019-08-17T19:12:41.893","ParentId":null,"OwnerUserId":"2292863","Title":null,"Body":"<p>Usually the reason for that error is because your input array(x) isn\u2019t the same size as your labels array(y). As the error states it looks like your labels array is empty. Before doing anything tensorflowy make sure both x and y arrays have values in them and that they are of the same size.<\/p>\n\n<p>To answer your question, yes you can use lists when training and is the preferred way of using tensorflow.<\/p>\n"}]}
+{"QuestionId":57533993,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><eager-execution>","CreationDate":"2019-08-17T06:52:46.097","AcceptedAnswerId":null,"OwnerUserId":5485668.0,"Title":"AttributeError: 'numpy.dtype' object has no attribute 'is_floating'","Body":"<p>I develop my own op using eager execution. I met an error. How to fix it?<\/p>\n\n<blockquote>\n  <p>Env: Colaboratory TF ver.: 1.14<\/p>\n<\/blockquote>\n\n<p>I use:<\/p>\n\n<pre><code>mnist = input_data.read_data_sets(\"\/tmp\/data\/\", one_hot=True)\n<\/code><\/pre>\n\n<p>I tried <code>tf.convert_to_tensor()<\/code> but it led to other error. Error caused by:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>loss_value, grads = grad(x_, y)\n<\/code><\/pre>\n\n<p>Error message:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    962     flat_sources = [_handle_or_self(x) for x in flat_sources]\n    963     for t in flat_sources_raw:\n--&gt; 964       if not t.dtype.is_floating:\n    965         logging.vlog(\n    966             logging.WARN, \"The dtype of the source tensor must be \"\n\nAttributeError: 'numpy.dtype' object has no attribute 'is_floating'\n<\/code><\/pre>\n","answers":[{"AnswerId":"57537927","CreationDate":"2019-08-17T16:11:51.433","ParentId":null,"OwnerUserId":"5485668","Title":null,"Body":"<p>I got that reason of the error which indicates \u201cif not is_floating\u201d for the raising the exception. To solve the error, it needs to make variable to a tensor, just using \u201cconvert_to_tensor()\u201d before an operation.<\/p>\n\n<p>Best,\nNaN<\/p>\n"}]}
+{"QuestionId":57534072,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch>","CreationDate":"2019-08-17T07:05:17.500","AcceptedAnswerId":57534659.0,"OwnerUserId":8365803.0,"Title":"How To Fix: RuntimeError: size mismatch in pyTorch","Body":"<p>I am new to pyTorch and getting following Size Mismatch error: <\/p>\n\n<pre><code>RuntimeError: size mismatch, m1: [7 x 2092500], m2: [180 x 120] at ..\\aten\\src\\TH\/generic\/THTensorMath.cpp:961\n<\/code><\/pre>\n\n<p>Model:<\/p>\n\n<pre><code>class Net(nn.Module):\ndef __init__(self):\n    super(Net, self).__init__()\n    self.conv1 = nn.Conv2d(3, 200, 5)\n    self.pool = nn.MaxPool2d(2, 2)\n    self.conv2 = nn.Conv2d(200, 180, 5)\n    self.fc1 = nn.Linear(180, 120)\n    self.fc2 = nn.Linear(120, 84)\n    self.fc3 = nn.Linear(84,5)     \n\ndef 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(x.shape[0], -1)\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>How ever I tried changing <code>x = x.view(x.shape[0], -1)<\/code> to <code>x = x.view(x.size(0), -1)<\/code> but that also did'nt work. Dimension of images is 512x384. and have used following transformation:<\/p>\n\n<pre><code>def load_dataset():\ndata_path = '.\/dataset\/training'\n\ntransform = transforms.Compose(\n               [transforms.Resize((512,384)),\n                transforms.ToTensor(),\n                transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])\n\n\ntrain_dataset = torchvision.datasets.ImageFolder(root=data_path,transform=transform)\ntrain_loader = torch.utils.data.DataLoader(train_dataset,batch_size=7,num_workers=0,shuffle=True)\n\nreturn train_loader\n<\/code><\/pre>\n","answers":[{"AnswerId":"57534659","CreationDate":"2019-08-17T08:36:00.730","ParentId":null,"OwnerUserId":"2650622","Title":null,"Body":"<p>The problem is that the dimensions of the output of your last max pooling layer don't match the input of the first fully connected layer. This is the network structure until the last max pool layer for input shape <code>(3, 512, 384)<\/code>:<\/p>\n\n<pre><code>----------------------------------------------------------------\n        Layer (type)               Output Shape         Param #\n================================================================\n            Conv2d-1        [-1, 200, 508, 380]          15,200\n         MaxPool2d-2        [-1, 200, 254, 190]               0\n            Conv2d-3        [-1, 180, 250, 186]         900,180\n         MaxPool2d-4         [-1, 180, 125, 93]               0\n================================================================\n<\/code><\/pre>\n\n<p>The last row of the table means that <code>MaxPool2d-4<\/code> outputs 180 channels (filter outputs) of 125 width and 93 height. So you need your first fully connected layer to have <code>180 * 125 * 93 = 2092500<\/code> input size. This is a lot, so I'd advise you to refine your architecture. In any case, if you change the input size of the first fully connected layer to <code>2092500<\/code>, it works:<\/p>\n\n<pre><code>class Net(nn.Module):\n    def __init__(self):\n        super(Net, self).__init__()\n        self.conv1 = nn.Conv2d(3, 200, 5)\n        self.pool = nn.MaxPool2d(2, 2)\n        self.conv2 = nn.Conv2d(200, 180, 5)\n        #self.fc1 = nn.Linear(180, 120)\n        self.fc1 = nn.Linear(2092500, 120)\n        self.fc2 = nn.Linear(120, 84)\n        self.fc3 = nn.Linear(84,5)\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(x.shape[0], -1)\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>Giving the following architecture:<\/p>\n\n<pre><code>----------------------------------------------------------------\n        Layer (type)               Output Shape         Param #\n================================================================\n            Conv2d-1        [-1, 200, 508, 380]          15,200\n         MaxPool2d-2        [-1, 200, 254, 190]               0\n            Conv2d-3        [-1, 180, 250, 186]         900,180\n         MaxPool2d-4         [-1, 180, 125, 93]               0\n            Linear-5                  [-1, 120]     251,100,120\n            Linear-6                   [-1, 84]          10,164\n            Linear-7                    [-1, 5]             425\n================================================================\nTotal params: 252,026,089\nTrainable params: 252,026,089\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>(You can use the <a href=\"https:\/\/github.com\/sksq96\/pytorch-summary\" rel=\"nofollow noreferrer\">torchsummary<\/a> package to generate these tables.)<\/p>\n"}]}
+{"QuestionId":57534478,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2019-08-17T08:10:22.197","AcceptedAnswerId":57534525.0,"OwnerUserId":9452512.0,"Title":"Constant output of 2D Convolutional regression in Keras-tensorflow","Body":"<p>I would like to  try to use a 2D Convolutional neural network as a regressor.<\/p>\n\n<p>This is, how I do it, *using the mnist dataset, that is available in tensorflow. <\/p>\n\n<p>First, we load the mnist dataset. <\/p>\n\n<pre><code>from __future__ import absolute_import, division, print_function, unicode_literals\nimport keras\n\ntry:\n  # %tensorflow_version only exists in Colab.\n  %tensorflow_version 2.x\nexcept Exception:\n  pass\nimport tensorflow as tf\n\nfrom tensorflow.keras import datasets, layers, models\n\n(train_images, train_labels), (test_images, test_labels) = datasets.mnist.load_data()\n\ntrain_images = train_images.reshape((60000, 28, 28, 1))\ntest_images = test_images.reshape((10000, 28, 28, 1))\n\n# Normalize pixel values to be between 0 and 1\ntrain_images, test_images = train_images \/ 255.0, test_images \/ 255.0\n<\/code><\/pre>\n\n<p>Now, we replace all the integer labels with floats<\/p>\n\n<pre><code>import numpy as np\ntest_labels  = np.random.rand(10000,)\ntrain_labels = np.random.rand(60000,)\n# train_labels = np.full((60000, ), 3.5) #Another possible choice of labels, to see wether the end result changes\n<\/code><\/pre>\n\n<p>We can build the model<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Conv2D, MaxPooling2D, Flatten, Dense, Activation\n\nfilters = 10\nkernel_size = 5 # or tuple for two dimensional\nnumber_of_input_layers = 1\n\nmodel = Sequential([\n    Conv2D(filters, kernel_size,\n               input_shape=(28, 28, number_of_input_layers),\n               #strides=(1, 1),\n               #padding='valid',\n               #data_format=\"channels_last\",\n               #dilation_rate=(1, 1),\n               #activation=None,\n               #use_bias=True,\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               ),\n        #MaxPooling2D(pool_size=(2, 2), strides=(2, 2)),\n        #Conv2D(64, (5, 5), activation='relu'),\n        #MaxPooling2D(pool_size=(2, 2)),\n        Flatten(),\n        Dense(1000, activation='relu'),\n        Dense(1, activation='tanh')\n])\n<\/code><\/pre>\n\n<p>and compile it.<\/p>\n\n<pre><code>        # For a mean squared error regression problem\nmodel.compile(optimizer='rmsprop',\n              loss='mse',\n              metrics=['accuracy']\n              )\n<\/code><\/pre>\n\n<p>One can also get a summary about the model<\/p>\n\n<pre><code>model.summary()\n<\/code><\/pre>\n\n<p>We  fit the  model<\/p>\n\n<pre><code>model.fit(train_images,\n          train_labels,\n          epochs=1)\n<\/code><\/pre>\n\n<p>and evaluate its loss<\/p>\n\n<pre><code>test_loss, test_acc = model.evaluate(test_images,\n                                     test_labels)\n<\/code><\/pre>\n\n<p>Finally, we can predict, the <strong><em><code>test_images<\/code><\/em><\/strong><\/p>\n\n<pre><code>result = model.predict(test_images)\n<\/code><\/pre>\n\n<p><strong>The problem is:<\/strong> <\/p>\n\n<p>No matter, how I change the train labels, the result variable is a <code>float32<\/code>, <code>(10000,1)<\/code> shaped <code>numpy<\/code> array consisting of <code>1.0<\/code> values. What is the problem? <\/p>\n\n<p>The question has already been asked <a href=\"https:\/\/stackoverflow.com\/questions\/45830331\/why-keras-2d-regression-network-has-constant-output\">here<\/a> but, nobody found an answer back then.<\/p>\n\n<hr>\n\n<hr>\n\n<p>Update:\nIf I change the last layer <\/p>\n\n<pre><code>Dense(1, activation='tanh')\n<\/code><\/pre>\n\n<p>to <\/p>\n\n<pre><code>Dense(1, activation='sigmoid')\n<\/code><\/pre>\n\n<p>the output is still the same. \nOnly when I put <\/p>\n\n<pre><code>Dense(1) or or Dense(1, activation='linear')\n<\/code><\/pre>\n\n<p>there, the **output is a numpy array of constant value != 1.0 **, why ? <\/p>\n","answers":[{"AnswerId":"57534525","CreationDate":"2019-08-17T08:18:20.800","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>This is why:<\/p>\n\n<pre><code>Dense(1, activation='softmax')\n<\/code><\/pre>\n\n<p>The <code>softmax<\/code> activation normalizes by dividing with the number of neurons, so it does not work at all with a single neuron, which will produce a constant output of 1.0.<\/p>\n\n<p>If you are doing regression you should not be using <code>softmax<\/code> at all, you should replace it with <code>sigmoid<\/code> activation if your targets are in the <code>[0, 1]<\/code> range, or with <code>linear<\/code> activation if the target range is unbounded.<\/p>\n"}]}
+{"QuestionId":57534738,"AnswerCount":0,"Tags":"<pytorch>","CreationDate":"2019-08-17T08:48:30.667","AcceptedAnswerId":null,"OwnerUserId":2556222.0,"Title":"Question regarding backward method in PyTorch","Body":"<p>To check my understanding of PyTorch's <code>backward<\/code> method, I was trying out this snippet of code:<\/p>\n\n<pre><code>import torch\n\nx = torch.rand(3,requires_grad=True)\ny = 2 * x\nz = 3 * y\nl = torch.sum(z)\nl.backward()\n<\/code><\/pre>\n\n<p>Computing by hand, I get <code>dl\/dz = (1.,1.,1.)<\/code> and <code>dl\/dy = (3.,3.,3.)<\/code>, and I believe this information is used by the <code>backward<\/code> method in order to compute <code>dl\/dx = (6.,6.,6.)<\/code> using Jacobian-vector products. <\/p>\n\n<p>Now, when I do <code>x.grad<\/code>, I get <code>tensor([6., 6., 6.])<\/code> as expected. But when I try <code>y.grad<\/code> or <code>z.grad<\/code>, I get <code>None<\/code> as the output, instead of <code>tensor([3., 3., 3.])<\/code> and <code>tensor([1., 1., 1.])<\/code> respectively. Why does this happen? <\/p>\n","answers":[]}
+{"QuestionId":57534801,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-08-17T08:58:32.650","AcceptedAnswerId":null,"OwnerUserId":10855529.0,"Title":"What is the meaning of keep_vars in state_dict?","Body":"<p><code>state_dict(destination=None, prefix='', keep_vars=False)<\/code><\/p>\n\n<p>what does changing <code>keep_vars<\/code> to <code>True<\/code> do?<\/p>\n","answers":[{"AnswerId":"57535360","CreationDate":"2019-08-17T10:15:46.437","ParentId":null,"OwnerUserId":"6210807","Title":null,"Body":"<p>In PyTorch >=0.4, it has no use.<\/p>\n\n<p><code>keep_vars<\/code> was added in the commit: <a href=\"https:\/\/github.com\/pytorch\/pytorch\/commit\/63c835bbe7eaad6ad564ebd6df88e8e35dd2d6d9\" rel=\"nofollow noreferrer\">Add keep_vars parameter to state_dict<\/a> stating that <\/p>\n\n<blockquote>\n  <p>When keep_vars is true, it returns a Variable for each parameter\n          (rather than a Tensor).<\/p>\n<\/blockquote>\n\n<p>In <code>state_dict<\/code> function, <a href=\"https:\/\/github.com\/pytorch\/pytorch\/blob\/master\/torch\/nn\/modules\/module.py#L658\" rel=\"nofollow noreferrer\"><code>_save_to_state_dict<\/code><\/a> is called internally, which contains the following code<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>for name, param in self._parameters.items():\n            if param is not None:\n                destination[prefix + name] = param if keep_vars else param.data\n        for name, buf in self._buffers.items():\n            if buf is not None:\n                destination[prefix + name] = buf if keep_vars else buf.data\n<\/code><\/pre>\n\n<p>The portion <code>param if keep_vars else param.data<\/code> made difference prior to PyTorch 0.4.0 when <code>Variable<\/code> and <code>Tensor<\/code> were separate, but now as they are merged, <code>keep_vars<\/code> is probably present only for backward compatibility. Check <a href=\"https:\/\/stackoverflow.com\/q\/51743214\/6210807\">Is .data still useful in pytorch?<\/a><\/p>\n"}]}
+{"QuestionId":57534864,"AnswerCount":1,"Tags":"<r><tensorflow><keras><tensorflow-datasets>","CreationDate":"2019-08-17T09:05:15.327","AcceptedAnswerId":null,"OwnerUserId":10332853.0,"Title":"Read multiple TFRecord files with keras::tfrecord_dataset","Body":"<p>I am getting started with Keras, and would like to create a dataset from multiple TFRecord files.<\/p>\n\n<p>Below is the code I have so far. <\/p>\n\n<pre><code> dataset &lt;- tfrecord_dataset(filenames) %&gt;%\n  dataset_map(function(example_proto) {\n    features &lt;- list(\n      label = tf$FixedLenFeature(shape(), tf$float32),\n      a = tf$FixedLenFeature(shape(), tf$float32),\n      b = tf$FixedLenFeature(shape(), tf$float32),\n      c = tf$FixedLenFeature(shape(), tf$float32),\n      d = tf$FixedLenFeature(shape(), tf$float32)\n\n )\n\n    features &lt;- tf$parse_single_example(example_proto, features)\n    x &lt;- list(features$a,\n              features$b, \n              features$c\n              features$d\n    )\n    y &lt;- tf$one_hot(tf$cast(features$label, tf$int32), 3L)\n    list(x, y)\n  }) %&gt;%\n  dataset_shuffle(150) %&gt;%\n  dataset_batch(16)\n<\/code><\/pre>\n\n<p>With a single tfrecord (e.g. file1 with features a,b and c), everything works fine. <\/p>\n\n<p>If I add a second tfrecord file2 with feature d such that filenames = c(file1, file2), then I get the following error: <\/p>\n\n<pre><code>Error in py_call_impl(callable, dots$args, dots$keywords) : \nInvalidArgumentError: Feature: d (data type: float) is required but could not be found.\n<\/code><\/pre>\n\n<p>Is there an efficient way to read multiple tfrecord files with tfrecord_dataset?<\/p>\n\n<p>Thanks..!<\/p>\n","answers":[{"AnswerId":"57536704","CreationDate":"2019-08-17T13:25:57.877","ParentId":null,"OwnerUserId":"3297472","Title":null,"Body":"<p>At least as of TensorFlow 1.14 you can pass an argument <code>default_value<\/code> to <code>tf$io$FixedLenFeature<\/code> that will be used in the case where that column is missing from the file.<\/p>\n\n<p>From the docs:<\/p>\n\n<ul>\n<li><code>default_value<\/code>: Value to be used if an example is missing this feature. It must be compatible with <code>dtype<\/code> and of the specified shape.<\/li>\n<\/ul>\n"}]}
+{"QuestionId":57534871,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2019-08-17T09:06:23.527","AcceptedAnswerId":null,"OwnerUserId":11355966.0,"Title":"Can I use \"model.fit()\" in \"for\" loop to change train data in each iteration","Body":"<p>I have a large dataset and it doesn't fit in memory. So while training, SSD is being used and epochs take too much time. <\/p>\n\n<p>I save my dataset 9 part of <code>.npz<\/code> file. I choose first part (part 0) as validation part and I didn't use in training.<\/p>\n\n<p>I use code below, and <code>acc<\/code> &amp; <code>val_acc<\/code> <strong>result were fine. But I feel I do big mistake somewhere.<\/strong> I didn't see any example like this<\/p>\n\n<pre><code>for part in range(1,9):\n X_Train, Y_Train = loadPart(part)\n history = model.fit(X_Train, Y_Train, batch_size=128, epochs=1, verbose=1)\n<\/code><\/pre>\n\n<p>and also I load part 0 as Test data<\/p>\n\n<p><code>val_loss, val_acc = model.evaluate(X_Test, Y_Test)<\/code><\/p>\n\n<p>I tried to check <code>val_acc<\/code> after train each part of dataset and I observed <code>val_acc<\/code> was increasing.<\/p>\n\n<p><strong>Could you please let me know if this usage is legal or illegal and why?<\/strong><\/p>\n\n<p>EDIT:<\/p>\n\n<p>I tried fit_generator but it still use disk during training and ETA was about 2,500 hours. (in model.fit with whole dataset it was about 30 mins per epoch) I use code below: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.fit_generator(generate_batches()), steps_per_epoch=196000,epochs=10)\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>def generate_batches(): \n   for part in range(1,9): \n      x, y = loadPart(part) yield(x,y)\n\ndef loadPart(part): \n   data = np.load('C:\/FOLDER_PATH\/'+str(part)+'.npz') \n   return [data['x'], data['y']\n<\/code><\/pre>\n\n<p>and X data shape is (196000,1536,1)<\/p>\n\n<p>EDIT 2:\nI found an answer in [github](\n<a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/4446\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/4446<\/a>). It says it is ok with call <code>model.fit()<\/code> multiple times in for but still I don't sure what happens in behind.What is the different between call <code>model.fit()<\/code> multiple times and call once with whole dataset.<\/p>\n","answers":[{"AnswerId":"57535307","CreationDate":"2019-08-17T10:10:08.850","ParentId":null,"OwnerUserId":"11939354","Title":null,"Body":"<p>If your model does not fit in RAM the keras documentation suggests the following (<a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-use-keras-with-datasets-that-dont-fit-in-memory\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/getting-started\/faq\/#how-can-i-use-keras-with-datasets-that-dont-fit-in-memory<\/a>):<\/p>\n\n<blockquote>\n  <p>You can do batch training using model.train_on_batch(x, y) and model.test_on_batch(x, y). See the models documentation.<\/p>\n  \n  <p>Alternatively, you can write a generator that yields batches of training data and use the method model.fit_generator(data_generator, steps_per_epoch, epochs).<\/p>\n<\/blockquote>\n\n<p>This means you could try to further split your training data into batches of 128 on your SSD and then do something like:<\/p>\n\n<pre><code>import glob\nimport numpy as np\n\ndef generate_batches(data_folder):\n    while True:\n        batches_paths = glob.glob(\"%s\/*.npz\" % data_folder)\n        for batch_path in batches_paths:\n            with np.load(batch_path) as batch:\n                x, y = preprocess_batch(batch)\n                yield (x, y)\n\n\nmodel.fit_generator(generate_batches(\"\/your-data-folder\"), steps_per_epoch=10000, epochs=10)\n<\/code><\/pre>\n\n<p>The preprocess_batch function would be responsible for extracting your x and y from each .npz file and the steps_per_epoch argument in the fit_generator function should be the rounded up value of your number of data samples divided by your batch size.<\/p>\n\n<p>More info: <\/p>\n\n<ul>\n<li><a href=\"https:\/\/keras.io\/models\/sequential\/#fit_generator\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/models\/sequential\/#fit_generator<\/a><\/li>\n<li><a href=\"https:\/\/www.pyimagesearch.com\/2018\/12\/24\/how-to-use-keras-fit-and-fit_generator-a-hands-on-tutorial\/\" rel=\"nofollow noreferrer\">https:\/\/www.pyimagesearch.com\/2018\/12\/24\/how-to-use-keras-fit-and-fit_generator-a-hands-on-tutorial\/<\/a><\/li>\n<\/ul>\n"}]}
+{"QuestionId":57535052,"AnswerCount":0,"Tags":"<python><keras><data-augmentation>","CreationDate":"2019-08-17T09:30:55.303","AcceptedAnswerId":null,"OwnerUserId":9218531.0,"Title":"How to perform data augmentation on a set of images and its masks in keras?","Body":"<p>I want to perform the augmentations of both my original images and its segmented masks, but I always end up with some error like:<\/p>\n\n<ol>\n<li>ValueError: setting an array element with a sequence<\/li>\n<li>Unable to expand 3 ranked image to a 4 ranked one<\/li>\n<\/ol>\n\n<p>Here is my code: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>data_gen_args = dict(rotation_range=5)\nimage_datagen = ImageDataGenerator(**data_gen_args)\nmask_datagen = ImageDataGenerator(**data_gen_args)\n\nseed = 1\nimages= []\nmasks= []\n\nfor image in glob.glob('..\/data\/images\/*.jpg'):\n    images.append(cv.imread(image))\nfor image in glob.glob('..\/data\/masks\/*.jpg'):\n    masks.append(cv.imread(image))\n\nimages= np.array(images)\nmasks= np.array(masks)\n\nfor image in images:\n    image_datagen.fit(image[np.newaxis], augment=True, seed=seed)\n\nfor image in masks:\n    mask_datagen.fit(image[np.newaxis], augment=True, seed=seed)\n\nimage_generator = image_datagen.flow(images, seed=seed)\nmask_generator = mask_datagen.flow(masks, seed=seed)\n\ntrain_generator = zip(image_generator, mask_generator)\n\nid = 0\nfor images, masks in train_generator:\n    for image, mask in zip(images, masks):\n        cv.imwrite('..\/data\/aug_images\/images\/' + str(id), image)\n        cv.imwrite('..\/data\/aug_images\/masks\/' + str(id), mask)\n        id += 1\n<\/code><\/pre>\n\n<p>Can someone correct me where I am going wrong? Thanks for your help.<\/p>\n\n<p>EDIT: I have a set of images and it's maskings. I want to obtain the same augmentation\/transformation for both of them.<\/p>\n","answers":[]}
+{"QuestionId":57535128,"AnswerCount":1,"Tags":"<python-3.x><keras><neural-network><conv-neural-network>","CreationDate":"2019-08-17T09:44:37.617","AcceptedAnswerId":57553348.0,"OwnerUserId":11435927.0,"Title":"What is my problem to use of Conv2D for image?","Body":"<p>First I read pictures from folders and sub-folders. Then I\u2019m changing images to gray and resizing to 100*200. I want to classify my images to 6 class.\nWhen i want to create my model, I can't  use Conv2D, cause I have dimension error but when I use Conv1D, I don\u2019t have any error and neural network is work. \nI want to use conv2D because my data is image. What is my problem?<\/p>\n\n<pre><code>#load train_images\npath_spec_train = \"\/home\/narges\/dataset\/train_24\/\"\nspec_train = glob.glob(path_spec_train + \"**\/*.png\") \nspec_train.sort()\n\nX_modify = []\nwidth = 200\nheight = 100\nfor spec in spec_train:\n    specs = cv2.imread(spec)\n    specs = cv2.cvtColor(specs,cv2.COLOR_BGR2GRAY)\n    specs = cv2.resize(specs ,(width, height))\n    specs = specs \/ np.max(specs)\n    specs = specs.astype(np.float32)\n    X_modify.append(specs)\n\nX_train = np.asarray(X_modify,dtype=np.float32)\n\n#=======================================================\n#load test_image\n\npath_spec_test = \"\/home\/narges\/dataset\/test_24\/\"\nspec_test = glob.glob(path_spec_test + \"**\/*.png\") \nspec_test.sort()\n\nX_modify_t = []\nwidth = 200\nheight = 100\nfor spec_t in spec_test:\n    specs_test = cv2.imread(spec_t)\n    specs_test = cv2.cvtColor(specs_test,cv2.COLOR_BGR2GRAY)\n    specs_test = cv2.resize(specs_test ,(width, height))\n    specs_test = specs_test \/ np.max(specs_test)\n    specs_test = specs_test.astype(np.float32)\n    X_modify_t.append(specs_test)\n\nX_test = np.asarray(X_modify_t,dtype=np.float32)\n#======================================================================\n#label\n\nspk_ID = [wavs[i].split('\/')[-1].lower() for i in range(number_of_files)]\n\nspk_ID_t = [wavs_t[i].split('\/')[-1].lower() for i in range(number_of_files_t)]\n\nlabel_no = [spk_ID[i].split('_')[-2] for i in range(number_of_files)]\n\nY_train = np_utils.to_categorical(label_no)\n\nlabel_no_t = [spk_ID_t[i].split('_')[-2] for i in range(number_of_files_t)]\n\nY_test = np_utils.to_categorical(label_no_t)\n\n\n#======================================================================\n# Create my model\nmyinput = layers.Input(shape=(100,200))\nconv1 = layers.Conv1D(32, 3, activation='relu', padding='same', strides=2)(myinput)\nconv2 = layers.Conv1D(64, 3, activation='relu', padding='same', strides=2)(conv1)\nflat = layers.Flatten()(conv2)\nout_layer = layers.Dense(6, activation='softmax')(flat)\nmymodel = Model(myinput, out_layer)\nmymodel.summary()\n\nmymodel.compile(optimizer=keras.optimizers.Adam(), loss=keras.losses.categorical_crossentropy, metrics=['accuracy'])\n\nnetwork_history = mymodel.fit(X_train, Y_train, batch_size=128, epochs=100)\n\npred = np.round(mymodel.predict(X_test))\nprint(classification_report(Y_test, pred))\n<\/code><\/pre>\n","answers":[{"AnswerId":"57553348","CreationDate":"2019-08-19T08:51:48.840","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You need to add the channels dimension to your input training data:<\/p>\n\n<pre><code>import numpy as np\n\nX_train = np.expand_dims(X_train, axis=3)\nX_test = np.expand_Dims(X_test, axis=3)\n<\/code><\/pre>\n\n<p>Then your data will be 4D, for example, (3312, 100, 200, 1), which is a grayscale image (one channel).<\/p>\n"}]}
+{"QuestionId":57535133,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras><neural-network>","CreationDate":"2019-08-17T09:46:06.050","AcceptedAnswerId":57535785.0,"OwnerUserId":6691564.0,"Title":"Regression accuracy with neural network in low density regions","Body":"<p>I am developing a neural net which needs to predict values between -1 and 1. However, I am only really concerned about the values at the ends of scale, say between -1 and -0.7 and between 0.7 and 1.<\/p>\n\n<p>I do not mind if 0.6, for example, gets predicted to be 0.1. However, I do want to know if it's 0.8 or 0.9.<\/p>\n\n<p>The distribution of my data is roughly normal, so there are many more samples in the range where I'm not concerned about the accuracy. It seems therefore that the training process is likely to lead to greater accuracy in the centre.<\/p>\n\n<p>How can I configure the training or engineer my expected result to overcome this?<\/p>\n\n<p>Thanks very much.<\/p>\n","answers":[{"AnswerId":"57535785","CreationDate":"2019-08-17T11:17:36.123","ParentId":null,"OwnerUserId":"2650622","Title":null,"Body":"<p>You could assign the observations to deciles, turn it into a classification problem and either assign a greater weight to the ranges you care about in the loss or just simply oversample them during training. By default, I'd go with weighing the classes in the loss function, as it is straight-forward to match with a weighted metric. Oversampling can be useful if you know that the distribution of your training data is different from the real data distribution.<\/p>\n\n<p>To assign certain classes a greater weight in the loss function with Keras, you can pass a <code>class_weight<\/code> parameter to <a href=\"https:\/\/keras.io\/models\/model\/#fit\" rel=\"nofollow noreferrer\">Model.fit<\/a>. If label 0 is the first decile and label 9 is the last decile, you could double the weight of the first and last two deciles as follows:<\/p>\n\n<pre><code>class_weight = {\n    0: 2,\n    1: 2,\n    2: 1,\n    3: 1,\n    4: 1,\n    5: 1,\n    6: 1,\n    7: 1,\n    8: 2,\n    9: 2\n}\nmodel.fit(..., class_weight=class_weight)\n<\/code><\/pre>\n\n<p>To oversample certain classes, you'd include them more often in the batches than the class distribution would suggest. The simplest way to implement this is to sample observation indices with <code>numpy.random.choice<\/code> that has an optional parameter to specify probabilities for each entry. (Note that Keras <code>Model.fit<\/code> also has a <code>sample_weight<\/code> parameter where you can assign weights to each observation in the training data that will be applied when computing the loss function, but the intended use case is to weigh samples by the confidence in their labels, so I don't think it's applicable here.)<\/p>\n"}]}
+{"QuestionId":57535156,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2019-08-17T09:49:37.943","AcceptedAnswerId":57535693.0,"OwnerUserId":11082042.0,"Title":"about backpropagation deep neural network in tensorflow","Body":"<p>I am reading about backpropagation deep neural network, and as I understood, I can summarize the  algorithm of that type of neural network as below : <\/p>\n\n<p><strong>1- Input x<\/strong> : Set the corresponding activation for the input layer <\/p>\n\n<p><strong>2- Feedforward<\/strong>: caclulate the error of the forward-propagation <\/p>\n\n<p><strong>3- Output error:<\/strong> Calculate the output error <\/p>\n\n<p><strong>4- Backpropagate the error<\/strong> : caclulate the error of the back-propagation<\/p>\n\n<p><strong>5- Output:<\/strong> Using the gradient of the cost function <\/p>\n\n<p>That's ok, and then I check many codes of that type of deep network, below an example code with explanation : <\/p>\n\n<pre><code>### imports\nimport tensorflow as tf\n\n### constant data\nx  = [[0.,0.],[1.,1.],[1.,0.],[0.,1.]]\ny_ = [[0.],[0.],[1.],[1.]]\n\n### induction\n# 1x2 input -&gt; 2x3 hidden sigmoid -&gt; 3x1 sigmoid output\n\n# Layer 0 = the x2 inputs\nx0 = tf.constant( x  , dtype=tf.float32 )\ny0 = tf.constant( y_ , dtype=tf.float32 )\n\n# Layer 1 = the 2x3 hidden sigmoid\nm1 = tf.Variable( tf.random_uniform( [2,3] , minval=0.1 , maxval=0.9 , dtype=tf.float32  ))\nb1 = tf.Variable( tf.random_uniform( [3]   , minval=0.1 , maxval=0.9 , dtype=tf.float32  ))\nh1 = tf.sigmoid( tf.matmul( x0,m1 ) + b1 )\n\n# Layer 2 = the 3x1 sigmoid output\nm2 = tf.Variable( tf.random_uniform( [3,1] , minval=0.1 , maxval=0.9 , dtype=tf.float32  ))\nb2 = tf.Variable( tf.random_uniform( [1]   , minval=0.1 , maxval=0.9 , dtype=tf.float32  ))\ny_out = tf.sigmoid( tf.matmul( h1,m2 ) + b2 )\n\n\n### loss\n# loss : sum of the squares of y0 - y_out\nloss = tf.reduce_sum( tf.square( y0 - y_out ) )\n\n# training step : gradient decent (1.0) to minimize loss\ntrain = tf.train.GradientDescentOptimizer(1.0).minimize(loss)\n\n\n### training\n# run 500 times using all the X and Y\n# print out the loss and any other interesting info\nwith tf.Session() as sess:\n  sess.run( tf.global_variables_initializer() )\n  for step in range(500) :\n    sess.run(train)\n\n  results = sess.run([m1,b1,m2,b2,y_out,loss])\n  labels  = \"m1,b1,m2,b2,y_out,loss\".split(\",\")\n  for label,result in zip(*(labels,results)) :\n    print \"\"\n    print label\n    print result\n\nprint \"\"\n<\/code><\/pre>\n\n<p>My question, the above code is calculating the error of the forward-propagation but I don't see any step for calculating the back-propagation error. In other words, following the above description, I can see the steps 1 (<strong>Input x<\/strong>) , 2 (<strong>Feedforward<\/strong>) , 3 (<strong>Output error<\/strong>) and 5 (<strong>Output<\/strong>) but the step number 4 which is (<strong>Backpropagate the error<\/strong>) is not shown in the code!!   Is that right or something missing in the code? The problem that all codes I found online are following same steps in backpropagation deep neural networks! \nplease, could you describe how the step of <strong>Backpropagate the error<\/strong> is happening the code or what should I add something to performing that step? <\/p>\n\n<p>Thank you<\/p>\n","answers":[{"AnswerId":"57535693","CreationDate":"2019-08-17T11:05:01.850","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>In simple terms, when you build the TF graph up to the point you are computing the loss in your code, TF will know on which <code>tf.Variable<\/code> (weights) the loss depends. Then, when you create the node <code>train = tf.train.GradientDescentOptimizer(1.0).minimize(loss)<\/code>, and later run it in a <code>tf.Session<\/code>, the backpropagation is done for you in the background. To be more specific, the <code>train = tf.train.GradientDescentOptimizer(1.0).minimize(loss)<\/code> merges the following steps:<\/p>\n\n<pre><code># 1. Create a GD optimizer with a learning rate of 1.0\noptimizer = tf.train.GradientDescentOptimizer(1.0)\n# 2. Compute the gradients for each of the variables (weights) with respect to the loss\ngradients, variables = zip(*optimizer.compute_gradients(loss))\n# 3. Update the variables (weights) based on the computed gradients\ntrain = optimizer.apply_gradients(zip(gradients, variables))\n<\/code><\/pre>\n\n<p>In particular, step <code>1<\/code> and <code>2<\/code>, summarize the backpropagation step. Hope that this makes things more clear for you!<\/p>\n\n<hr>\n\n<p>Besides, I want to restructure the steps in your question:<\/p>\n\n<ol>\n<li><strong>Input X:<\/strong> The input of the neural network.<\/li>\n<li><strong>Forward pass:<\/strong> Propagating the input through the neural network, in order to get the output. In other words, multiplying the input X with each of the <code>tf.Variable<\/code> in your code.<\/li>\n<li><strong>Loss<\/strong>: The mismatch between the obtained output in step 2 and the expected output.<\/li>\n<li><strong>Computing the gradients:<\/strong> Computing the gradients for each of the <code>tf.Variable<\/code> (weights) with respect to the loss. <\/li>\n<li><strong>Updating the weights<\/strong>: Updating each <code>tf.Variable<\/code> (weight) according to its corresponding gradient.<\/li>\n<\/ol>\n\n<p>Please note that step 4 and 5 encapsulate backpropagation.<\/p>\n"}]}
+{"QuestionId":57535526,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-17T10:43:10.527","AcceptedAnswerId":57535641.0,"OwnerUserId":11939485.0,"Title":"For an infinite dataset, is the data used in each epoch the same?","Body":"<p>In tensorflow, suppose I have a dataset from generator:<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_generator(gen...)\n<\/code><\/pre>\n\n<p>and this generator generates <strong>infinite nonrepetitive<\/strong> data (just like infinite non-cyclic decimals).<\/p>\n\n<pre><code>model.fit(dataset, steps_per_epoch=10000, epochs=5)\n<\/code><\/pre>\n\n<p>Now within these 5 epochs of training, is the data used the same? i.e. always the first 10000 items from the generator? rather than 0-9999 for epoch 1, 10000-19999 for epoch 2,etc.<\/p>\n\n<p>What about the <code>initial_epoch<\/code> parameter? If I set it to be 1, will the model be trained from the 10000th item?<\/p>\n\n<pre><code>model.fit(dataset, steps_per_epoch=10000, epochs=5, initial_epoch=1)\n<\/code><\/pre>\n\n<p><strong>update :<\/strong>\nthis simple test shows that the dataset will be reset every time <code>model.fit()<\/code> is called<\/p>\n\n<pre><code>def gen():\n    i = 1\n    while True:\n        yield np.array([[i]]), np.array([[0]])\n        i += 1\n\nds = tf.data.Dataset.from_generator(gen, output_types=(tf.int32, tf.int32)).batch(3)\n\nx = Input(shape=(1, 1))\nmodel = Model(inputs=x, outputs=x)\n\nmodel.compile('adam', loss=lambda true, pred: tf.reduce_mean(pred))\nfor i in range(10):\n    model.fit(ds, steps_per_epoch=5, epochs=1)\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<pre><code>1\/5 [=====&gt;........................] - ETA: 0s - loss: 2.0000\n5\/5 [==============================] - 0s 9ms\/step - loss: 8.0000\n1\/5 [=====&gt;........................] - ETA: 0s - loss: 2.0000\n5\/5 [==============================] - 0s 2ms\/step - loss: 8.0000\n1\/5 [=====&gt;........................] - ETA: 0s - loss: 2.0000\n5\/5 [==============================] - 0s 2ms\/step - loss: 8.0000\n1\/5 [=====&gt;........................] - ETA: 0s - loss: 2.0000\n5\/5 [==============================] - 0s 2ms\/step - loss: 8.0000\n1\/5 [=====&gt;........................] - ETA: 0s - loss: 2.0000\n5\/5 [==============================] - 0s 2ms\/step - loss: 8.0000\n1\/5 [=====&gt;........................] - ETA: 0s - loss: 2.0000\n5\/5 [==============================] - 0s 2ms\/step - loss: 8.0000\n1\/5 [=====&gt;........................] - ETA: 0s - loss: 2.0000\n5\/5 [==============================] - 0s 2ms\/step - loss: 8.0000\n1\/5 [=====&gt;........................] - ETA: 0s - loss: 2.0000\n5\/5 [==============================] - 0s 2ms\/step - loss: 8.0000\n1\/5 [=====&gt;........................] - ETA: 0s - loss: 2.0000\n5\/5 [==============================] - 0s 2ms\/step - loss: 8.0000\n1\/5 [=====&gt;........................] - ETA: 0s - loss: 2.0000\n5\/5 [==============================] - 0s 2ms\/step - loss: 8.0000\n<\/code><\/pre>\n\n<p>5 epochs in 1 call:<\/p>\n\n<pre><code>model.fit(ds, steps_per_epoch=5, epochs=5)\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<pre><code>Epoch 1\/5\n1\/5 [=====&gt;........................] - ETA: 0s - loss: 2.0000\n5\/5 [==============================] - 0s 9ms\/step - loss: 8.0000\nEpoch 2\/5\n1\/5 [=====&gt;........................] - ETA: 0s - loss: 17.0000\n5\/5 [==============================] - 0s 2ms\/step - loss: 23.0000\nEpoch 3\/5\n1\/5 [=====&gt;........................] - ETA: 0s - loss: 32.0000\n5\/5 [==============================] - 0s 2ms\/step - loss: 38.0000\nEpoch 4\/5\n1\/5 [=====&gt;........................] - ETA: 0s - loss: 47.0000\n5\/5 [==============================] - 0s 2ms\/step - loss: 53.0000\nEpoch 5\/5\n1\/5 [=====&gt;........................] - ETA: 0s - loss: 62.0000\n5\/5 [==============================] - 0s 2ms\/step - loss: 68.0000\n<\/code><\/pre>\n","answers":[{"AnswerId":"57535641","CreationDate":"2019-08-17T10:59:33.870","ParentId":null,"OwnerUserId":"10886420","Title":null,"Body":"<p>No, the data used is different. <code>steps_per_epoch<\/code> is used by <code>keras<\/code> to determine the length of each <code>epoch<\/code> (as generators got no length), so it knows when to end the training (or call checkpointers etc.).<\/p>\n\n<p><code>initial_epoch<\/code> is a number displayed for epoch and useful when you want to restart training from checkpoint (see <a href=\"https:\/\/keras.io\/models\/sequential\/\" rel=\"nofollow noreferrer\">fit method<\/a>), it has nothing to do with data iteration.<\/p>\n\n<p>If you pass the same <code>dataset<\/code> to <code>model.fit<\/code> method, it will reset after each function call (thanks for the info OP).<\/p>\n"}]}
+{"QuestionId":57535692,"AnswerCount":0,"Tags":"<tensorflow><keras><lstm><generative>","CreationDate":"2019-08-17T11:05:00.887","AcceptedAnswerId":null,"OwnerUserId":8534205.0,"Title":"Autoregressive model with keras","Body":"<p>I'm building a VAE with keras. The decoder of the model il a LSTM so, for generating data I have to generate sample by sample. I have already written the model for training and it works. Now I want to write another model which load the weight from the decoder of the trained one and use it for generate data. <\/p>\n\n<p>How can I write a LSTM generative model with with keras? \nI haven't found any useful tutorial<\/p>\n\n<p>Thank you<\/p>\n","answers":[]}
+{"QuestionId":57535753,"AnswerCount":0,"Tags":"<deep-learning><computer-vision><conv-neural-network><pytorch><pose-estimation>","CreationDate":"2019-08-17T11:13:00.723","AcceptedAnswerId":null,"OwnerUserId":11939411.0,"Title":"Training loss explodes from first training example and then outputs nan","Body":"<p>I'm a novice at deep learning, I have built some basic CNNs, but this time I'm trying to build FCN(Fully Convolutional Network) similar to Yolo3. my network contains 32 layers with LeakyRelu as the activation function and adam optimizer. there are 680 data samples and the input image size is 416x416 same as yolo model. I have given some snipes of my code below.   <\/p>\n\n<p>I'm using Pytorch 1.1 and cuda 9 version.\nI have tried different learning rates like 0.0001, 0.000001, 0.0000001 as suggested at in many blogs, and also have tried different betas like (0.9,0.999), (0.5,0.999) and so on. I also have tried to train it long, up to 200 epochs. <\/p>\n\n<h1>Output<\/h1>\n\n<p><code>Running loss :  59102027776.0\n  0%|          | 1\/630 [00:19&lt;3:24:03, 19.46s\/it]Running loss :  nan\n  0%|          | 2\/630 [00:23&lt;2:34:23, 14.75s\/it]Running loss :  nan\n  0%|          | 3\/630 [00:25&lt;1:53:32, 10.87s\/it]Running loss :  nan<\/code><\/p>\n\n<h2>Loss Function formula<\/h2>\n\n<p>Please consider eq: 7 and 8 only in picture.\nclick the link for <a href=\"https:\/\/imgur.com\/a\/K8BsIbW\" rel=\"nofollow noreferrer\">formula picture<\/a><\/p>\n\n<h3>Loss Function Code<\/h3>\n\n<pre><code>masked_pose_loss = torch.mean(\n        torch.sum(mask * torch.sum(torch.mul(pred - true, pred - true), dim=[1, 2]), dim=[1, 2, 3]))\n<\/code><\/pre>\n\n<h3>FCN<\/h3>\n\n<pre><code>    self.relu1_1 = nn.LeakyReLU(inplace=True)\n    self.conv1_2 = nn.Conv2d(32, 64, 3, stride=1,padding=1)\n    self.relu1_2 = nn.LeakyReLU(inplace=True)\n    self.pool1 = nn.MaxPool2d(2, stride=2, ceil_mode=True)\n\n    # conv2\n\n    self.conv2_1 = nn.Conv2d(64, 128, 3, stride=1, padding=1)\n    self.relu2_1 = nn.LeakyReLU(inplace=True)\n    self.conv2_2 = nn.Conv2d(128, 64, 1, stride=1)\n    self.relu2_2 = nn.LeakyReLU(inplace=True)\n    self.pool2 = nn.MaxPool2d(2, stride=2, ceil_mode=True)\n\n    # conv3\n    self.conv3_1 = nn.Conv2d(64, 128, 3, stride=1, padding=1)\n    self.relu3_1 = nn.LeakyReLU(inplace=True)\n    self.conv3_2 = nn.Conv2d(128, 256, 3, stride=1, padding=1)\n    self.relu3_2 = nn.LeakyReLU(inplace=True)\n    self.conv3_3 = nn.Conv2d(256, 128, 1, stride=1)\n    self.relu3_3 = nn.LeakyReLU(inplace=True)\n    self.pool3 = nn.MaxPool2d(2, stride=2, ceil_mode=True)\n\n    # conv4\n    self.conv4_1 = nn.Conv2d(128, 256, 3, stride=1, padding=1)\n    self.relu4_1 = nn.LeakyReLU(inplace=True)\n    self.conv4_2 = nn.Conv2d(256, 512, 3, stride=1, padding=1)\n    self.relu4_2 = nn.LeakyReLU(inplace=True)\n    self.conv4_3 = nn.Conv2d(512, 256, 1, stride=1)\n    self.relu4_3 = nn.LeakyReLU(inplace=True)\n    self.pool4 = nn.MaxPool2d(2, stride=2, ceil_mode=True)\n\n    # conv5\n    self.conv5_1 = nn.Conv2d(256, 512, 3, stride=1, padding=1)\n    self.relu5_1 = nn.LeakyReLU(inplace=True)\n    self.conv5_2 = nn.Conv2d(512, 256, 1, stride=1)\n    self.relu5_2 = nn.LeakyReLU(inplace=True)\n    self.conv5_3 = nn.Conv2d(256, 512, 3, stride=1, padding=1)\n    self.relu5_3 = nn.LeakyReLU(inplace=True)\n    self.pool5 = nn.MaxPool2d(2, stride=2, ceil_mode=True)\n\n    # fc6\n\n    self.conv6_1 = nn.Conv2d(512, 1024, 3, stride=1, padding=1)\n    self.relu6_1 = nn.LeakyReLU(inplace=True)\n    self.conv6_2 = nn.Conv2d(1024, 512, 1, stride=1)\n    self.relu6_2 = nn.LeakyReLU(inplace=True)\n    self.conv6_3 = nn.Conv2d(512, 1024, 3, stride=1, padding=1)\n    self.relu6_3 = nn.LeakyReLU(inplace=True)\n    self.conv7_1 = nn.Conv2d(1024, 512, 1, stride=1)\n    self.relu7_1 = nn.LeakyReLU(inplace=True)\n    self.conv7_2 = nn.Conv2d(512, 1024, 3, stride=1, padding=1)\n    self.relu7_2 = nn.LeakyReLU(inplace=True)\n    self.conv7_3 = nn.Conv2d(1024, 1024, 3, stride=1, padding=1)\n    self.relu7_3 = nn.LeakyReLU(inplace=True)\n    self.conv8_1 = nn.Conv2d(1024, 1024, 3, stride=1, padding=1)\n    self.relu8_1 = nn.LeakyReLU(inplace=True)\n    self.conv8_2 = nn.Conv2d(1024, 1024, 3, stride=1)\n    self.conv8_2 = nn.ReLU(inplace=True)\n    self.conv_rout16 = nn.Conv2d(512, 64, 1, stride=1)\n    self.relu_rout16 = nn.ReLU(inplace=True)enter code here\n\n\n    # Resuming comments: I was implementing last two layer of network:\n    self.convf_1 = nn.Conv2d(1280, 1024, 3, stride=1, padding=1)\n    self.reluf_1 = nn.LeakyReLU(inplace=True)\n\n    self.convf_2 = nn.Conv2d(1024, self.target_channel_size, 1, stride=1)\n\n    self.reluf_2 = nn.LeakyReLU(inplace=True)\n<\/code><\/pre>\n\n<p>your response is highly appreciated. <\/p>\n\n<p>Note: this is the first time that I'm posting. If I have missed some necessary information please let me know.<\/p>\n","answers":[]}
+{"QuestionId":57536177,"AnswerCount":3,"Tags":"<python><tensorflow><dice><tensorflow-probability>","CreationDate":"2019-08-17T12:14:38.890","AcceptedAnswerId":57539410.0,"OwnerUserId":11939719.0,"Title":"Probability mass function sum 2 dice roll?","Body":"<p>Triying to plot pmf for 2 dice sum, but some tail-right problems occur.<\/p>\n\n<p>I have tried using numpy and other python libraries and the problems stays:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>import tensorflow as tf\ntf.enable_eager_execution()\nimport tensorflow_probability as tfp\nfrom matplotlib import pyplot as plt\nimport numpy as np\n\na=500000\ndado1_=evaluate(tf.cast(tfp.distributions.Uniform(low=1,high=7).sample(a),tf.int32))\ndado2_=evaluate(tf.cast(tfp.distributions.Uniform(low=1,high=7).sample(a),tf.int32))\n\nplt.hist(np.add(dado1_,dado2_))\n<\/code><\/pre>\n\n<pre class=\"lang-none prettyprint-override\"><code>expecting symetric distribution centered at 7 with gaussian form,\nassymetric found\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/MzZ7X.png\" rel=\"nofollow noreferrer\">outputplot<\/a><\/p>\n","answers":[{"AnswerId":"57562919","CreationDate":"2019-08-19T19:35:30.277","ParentId":null,"OwnerUserId":"9946929","Title":null,"Body":"<p>For the dice rolls, you'd want <code>tfd.Categorical<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow.compat.v2 as tf\ntf.enable_v2_behavior()\nimport tensorflow_probability as tfp\ntfd = tfp.distributions\ntfb = tfp.bijectors\n\n# 10 independent 1-die samples\nprint(tfb.AffineScalar(shift=1.)(tfd.Categorical(logits=tf.zeros([6]), dtype=tf.float32)).sample(10))\n\n# 10 independent pairs-of-dice samples\nd = tfb.AffineScalar(shift=1.)(tfd.Categorical(logits=tf.zeros([6]), dtype=tf.float32))\nprint(tfd.Sample(d, 2).sample(10))\n\n<\/code><\/pre>\n"},{"AnswerId":"57539410","CreationDate":"2019-08-17T19:42:33.580","ParentId":null,"OwnerUserId":"10181523","Title":null,"Body":"<p>Looking at the graph, it seems like the area between the tick mark for 2 and 3 is the number of rolls that sum to 2. 3-4 is 3, and so on. <\/p>\n\n<p>Is it possible that between tick marks 11 and 12 contains the number of rolls that sum to 11 <em>and<\/em> 12? It could explain why the graph seems shorter on the right side than it should be, as well as why the last bar is taller than expected.<\/p>\n"},{"AnswerId":"57539537","CreationDate":"2019-08-17T19:59:17.633","ParentId":null,"OwnerUserId":"1220389","Title":null,"Body":"<p>Tensorflow_probabilty.Uniform is a continuous distribution, not discrete. So it\u2019s not really going to model dice rolls appropriately.<\/p>\n"}]}
+{"QuestionId":57536319,"AnswerCount":1,"Tags":"<machine-learning><classification><pytorch><torchvision>","CreationDate":"2019-08-17T12:34:23.863","AcceptedAnswerId":null,"OwnerUserId":286579.0,"Title":"Sharing gradients between the related classes in pytorch","Body":"<p>I am implementing a classifier with multi-label predictions. For the moment, there are only 5 classes. Following is the snippet of the code. So far I have no problem.<\/p>\n\n<p>But I would like to further extend it where gradients between the related tasks (constrained by a matrix <strong>W<\/strong> which encodes the relatedness between the tasks) is shared. For example: there are five different classes <strong>a, b,c, d, e<\/strong> . I computed gradient for the mis-classification of <strong>a<\/strong> and updated the parameters. From the matrix <strong>W<\/strong> I discovered that task <strong>e<\/strong> is also related. I would also like to update the parameters for the task <strong>e<\/strong> with the gradients computed while predicting <strong>a<\/strong> with the rate encoded by <strong>W<\/strong>. Here <strong>W<\/strong> is precomputed and fixed. <\/p>\n\n<p>I have no idea how to do it in PyTorch. I would really appreciate if someone can help me.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class Classifier(nn.Module):\n    def __init__(self, image_size=128, conv_dim=64, c_dim=5, repeat_num=6):\n        super(Classifier, 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 i 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.conv = nn.Conv2d(curr_dim, 5, kernel_size=kernel_size, bias=False)\n\n    def forward(self, x): \n        h = self.main(x)\n        out_cls = self.conv2(h)\n        return out_cls.view(out_cls.size(0), out_cls.size(1))\n<\/code><\/pre>\n","answers":[{"AnswerId":"57538549","CreationDate":"2019-08-17T17:37:10.883","ParentId":null,"OwnerUserId":"10381382","Title":null,"Body":"<p>First you can propose different objective functions that only related to your tasks. Then you can do optimizations based on different optimizers related to different objective functions.<\/p>\n"}]}
+{"QuestionId":57536534,"AnswerCount":0,"Tags":"<java><python><numpy><tensorflow>","CreationDate":"2019-08-17T13:04:33.570","AcceptedAnswerId":null,"OwnerUserId":11856751.0,"Title":"Is there any Java equivalent of NumPy?","Body":"<p>I tried to get predictions from tflite model to get eye region landmarks.\nI successfully got values from tflite model using python script but I didn't know how can I got those values using MLKit firebase.\nPython script used numpy library for converting image to arrays.\nI want to know what is the problem here?\nThere is any way to use numpy functions with java code?\nIf not, what is the equivalent of the numpy library in java?<\/p>\n\n<p>Python script<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>data = np.asarray( img, dtype=\"float32\" ) \n\n# Inference on input data normalized to [0, 1]\ninputImg = np.expand_dims(data,0).astype(np.float32)\ninput_details = interpreter.get_input_details()\ninterpreter.set_tensor(input_details[0]['index'], inputImg)\n\ninterpreter.invoke()\n\noutput_details = interpreter.get_output_details()\noutput_data = interpreter.get_tensor(output_details[0]['index'])\nprint(output_data)\n\n<\/code><\/pre>\n\n<p>Java code ( Android ) the problem is with those functions:<\/p>\n\n<pre class=\"lang-java prettyprint-override\"><code>private float[][][][] bitmapToInputArray() {\n        \/\/ [START mlkit_bitmap_input]\n        Bitmap bitmap= getYourInputImage();\n        bitmap = Bitmap.createScaledBitmap(bitmap, 112, 112, true);\n\n        int batchNum = 0;\n        float[][][][] input = new float[1][112][112][3];\n        for (int x = 0; x &lt; 112; x++) {\n            for (int y = 0; y &lt; 112; y++) {\n                int pixel = bitmap.getPixel(x, y);\n                \/\/ Normalize channel values to [-1.0, 1.0]. This requirement varies by\n                \/\/ model. For example, some models might require values to be normalized\n                \/\/ to the range [0.0, 1.0] instead.\n                input[batchNum][x][y][0] = (Color.red(pixel) - 127) \/ 128.0f;\n                input[batchNum][x][y][1] = (Color.green(pixel) - 127) \/ 128.0f;\n                input[batchNum][x][y][2] = (Color.blue(pixel) - 127) \/ 128.0f;\n                Log.i(\"Input\",\"input\"+input[batchNum][x][y]);\n\n            }\n        }\n        \/\/ [END mlkit_bitmap_input]\n\n      return input;\n    }\n\n private void useInferenceResult(float[] probabilities) throws IOException {\n        String[] result=new String[80];\n        String x=\"\";\n        String y=\"\";\n        ArrayList&lt;Point&gt; listpoint= new ArrayList&lt;Point&gt;();\n\n        double viewWidth = canvas.getWidth();\n        double viewHeight = canvas.getHeight();\n        double imageWidth = mutableBitmap.getWidth();\n        double imageHeight = mutableBitmap.getHeight();\n        Log.i(\"viewWidth\",\"viewwidth \"+viewWidth);\n        Log.i(\"viewHeight\",\"viewheight \"+viewHeight);\n        Log.i(\"imagewidth\",\"imagewidth \"+imageWidth);\n        Log.i(\"imaageHeigh\",\"imageheigh \"+imageHeight);\n        double scale = Math.min(viewWidth \/ imageWidth, viewHeight \/ imageHeight);\n        Log.i(\"Scale\",\"Scale\"+scale);\n        try {\n            for (int i = 0; i &lt; probabilities.length; i++) {\n\n                Log.i(\"MLKit\", String.format(\"%1.8f\", probabilities[i]));\n                float i1 = probabilities[i];\n                Log.i(\"floaaat\", \"\" + i1);\n            }\n        }\n<\/code><\/pre>\n\n<ul>\n<li>The result from the python script:<\/li>\n<\/ul>\n\n<blockquote>\n  <p>[0.33135968 0.19592011 0.34212315 0.17297666 0.36624995 0.16413747\n   0.3894139  0.17440952 0.39828074 0.1978043  0.3891497  0.22268474\n   0.36345637 0.22974193 0.3401759  0.2193309  0.30167252 0.20411113\n   0.3167112  0.19134495 0.33793524 0.18388326 0.3642417  0.18049955\n   0.3903508  0.18533507 0.40906873 0.1957745  0.42142123 0.21091096\n   0.40550107 0.21829814 0.38345626 0.22071144 0.35900232 0.22142673\n   0.3363348  0.21877256 0.3161971  0.2133534  0.62843406 0.21482795\n   0.6389724  0.1914106  0.6628249  0.1835615  0.6858679  0.19583184\n   0.6946868  0.22111627 0.6840309  0.24444285 0.66027373 0.25241333\n   0.6351568  0.24192403 0.60499936 0.22642238 0.6210091  0.21289764\n   0.6423563  0.2042976  0.6685919  0.20277795 0.69201195 0.20948553\n   0.70882106 0.22015369 0.71931773 0.23518339 0.7076659  0.24166131\n   0.69054717 0.24350837 0.6694564  0.24258481 0.64537776 0.23927754\n   0.62199306 0.23511863]<\/p>\n<\/blockquote>\n\n<ul>\n<li>The result from the Android java code:<\/li>\n<\/ul>\n\n<blockquote>\n  <p>2019-08-17 14:47:50.617 21349-21349\/com.example.irisdetection I\/MLKit: 0,23961355\n  2019-08-17 14:47:50.620 21349-21349\/com.example.irisdetection I\/MLKit: 0,25104424\n  2019-08-17 14:47:50.621 21349-21349\/com.example.irisdetection I\/MLKit: 0,28179651\n  2019-08-17 14:47:50.622 21349-21349\/com.example.irisdetection I\/MLKit: 0,31467810\n  2019-08-17 14:47:50.623 21349-21349\/com.example.irisdetection I\/MLKit: 0,33257431\n  2019-08-17 14:47:50.624 21349-21349\/com.example.irisdetection I\/MLKit: 0,32645294\n  2019-08-17 14:47:50.625 21349-21349\/com.example.irisdetection I\/MLKit: 0,29138848\n  2019-08-17 14:47:50.626 21349-21349\/com.example.irisdetection I\/MLKit: 0,25581932\n  2019-08-17 14:47:50.627 21349-21349\/com.example.irisdetection I\/MLKit: 0,19593856\n  2019-08-17 14:47:50.628 21349-21349\/com.example.irisdetection I\/MLKit: 0,21698779\n  2019-08-17 14:47:50.631 21349-21349\/com.example.irisdetection I\/MLKit: 0,24266151\n  2019-08-17 14:47:50.632 21349-21349\/com.example.irisdetection I\/MLKit: 0,27562365\n  2019-08-17 14:47:50.633 21349-21349\/com.example.irisdetection I\/MLKit: 0,30823168\n  2019-08-17 14:47:50.635 21349-21349\/com.example.irisdetection I\/MLKit: 0,33465266\n  2019-08-17 14:47:50.636 21349-21349\/com.example.irisdetection I\/MLKit: 0,35355449\n  2019-08-17 14:47:50.637 21349-21349\/com.example.irisdetection I\/MLKit: 0,34009647\n  2019-08-17 14:47:50.638 21349-21349\/com.example.irisdetection I\/MLKit: 0,31358159\n  2019-08-17 14:47:50.640 21349-21349\/com.example.irisdetection I\/MLKit: 0,28156102\n  2019-08-17 14:47:50.642 21349-21349\/com.example.irisdetection I\/MLKit: 0,25063315\n  2019-08-17 14:47:50.643 21349-21349\/com.example.irisdetection I\/MLKit: 0,21878451\n  2019-08-17 14:47:50.644 21349-21349\/com.example.irisdetection I\/MLKit: 0,69623101\n  2019-08-17 14:47:50.646 21349-21349\/com.example.irisdetection I\/MLKit: 0,70167470\n  2019-08-17 14:47:50.646 21349-21349\/com.example.irisdetection I\/MLKit: 0,73317540\n  2019-08-17 14:47:50.648 21349-21349\/com.example.irisdetection I\/MLKit: 0,76974392\n  2019-08-17 14:47:50.649 21349-21349\/com.example.irisdetection I\/MLKit: 0,79195201\n  2019-08-17 14:47:50.651 21349-21349\/com.example.irisdetection I\/MLKit: 0,78359401\n  2019-08-17 14:47:50.652 21349-21349\/com.example.irisdetection I\/MLKit: 0,75674009\n  2019-08-17 14:47:50.653 21349-21349\/com.example.irisdetection I\/MLKit: 0,71786618\n  2019-08-17 14:47:50.654 21349-21349\/com.example.irisdetection I\/MLKit: 0,66782737\n  2019-08-17 14:47:50.655 21349-21349\/com.example.irisdetection I\/MLKit: 0,68930006\n  2019-08-17 14:47:50.656 21349-21349\/com.example.irisdetection I\/MLKit: 0,71668541\n  2019-08-17 14:47:50.657 21349-21349\/com.example.irisdetection I\/MLKit: 0,75279719\n  2019-08-17 14:47:50.658 21349-21349\/com.example.irisdetection I\/MLKit: 0,78872705\n  2019-08-17 14:47:50.659 21349-21349\/com.example.irisdetection I\/MLKit: 0,81867975\n  2019-08-17 14:47:50.661 21349-21349\/com.example.irisdetection I\/MLKit: 0,83806717\n  2019-08-17 14:47:50.662 21349-21349\/com.example.irisdetection I\/MLKit: 0,82371044\n  2019-08-17 14:47:50.664 21349-21349\/com.example.irisdetection I\/MLKit: 0,79749656\n  2019-08-17 14:47:50.665 21349-21349\/com.example.irisdetection I\/MLKit: 0,76317006\n  2019-08-17 14:47:50.666 21349-21349\/com.example.irisdetection I\/MLKit: 0,72700304\n  2019-08-17 14:47:50.667 21349-21349\/com.example.irisdetection I\/MLKit: 0,69159627<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":57536551,"AnswerCount":1,"Tags":"<python><keras><scikit-learn>","CreationDate":"2019-08-17T13:06:34.350","AcceptedAnswerId":57536661.0,"OwnerUserId":11939825.0,"Title":"Setting the scoring parameter of RandomizedSeachCV to r2","Body":"<p>I am trying to use <code>randomizedSearchCV()<\/code> from sklearn to find the best parameters to use in a neural network model build with keras. I would like for it to 'score' based on the r2 metric - it doesn't throw up any errors when I run, but after it has run it says best_score_ is -1.436.<\/p>\n\n<p>This isn't a valid r2 metric (is it?). So I'm wondering if anyone can take a look at the code below and see if they can spot<\/p>\n\n<p>a) What is this <code>best_score_<\/code> of -1.436?<\/p>\n\n<p>b) How can I amend the below code to make the function return the <code>r2<\/code> metric?<\/p>\n\n<p>My neural network model is called model4.\nI have added \"scoring = 'r2'\" as a parameter in <code>RandomiszedSearchCV<\/code>, which I thought was meant to tell it to use this as the score<\/p>\n\n<h1>This is the code I a using for the <code>RandomizedSearchCV()<\/code>:<\/h1>\n\n<pre><code>model5 = KerasClassifier(build_fn=get_model4, epochs = 10)\n\nparams = {'num_layers' : [1,2,3,4], 'size_layers' : [5,10,25,100],'batchnorm' : [True,False], 'act' : ['relu','sigmoid','tanh'], 'lossfn' : ['mean_squared_error','mean_absolute_error','mean_squared_logarithmic_error'], 'lrate' : [0.01,0.001,0.0001]}\n\n\nrandom_search = RandomizedSearchCV(model5, param_distributions=params, cv=KFold(3), scoring='r2')\n\nrandom_search_results = random_search.fit(inputs,decisions)\n\nprint(\"Best: \" + str(random_search_results.best_score_)+\" using:\"+str(random_search_results.best_params_))\n<\/code><\/pre>\n","answers":[{"AnswerId":"57536661","CreationDate":"2019-08-17T13:19:34.957","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Yes, the R^2 score can be negative, its not the correlation but the coefficient of determination, with a maximum of 1.0 and a minimum if minus infinite. You can get this information from the scikit-learn <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.metrics.r2_score.html#sklearn.metrics.r2_score\" rel=\"nofollow noreferrer\">documentation<\/a>.<\/p>\n"}]}
+{"QuestionId":57536555,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-08-17T13:07:02.913","AcceptedAnswerId":null,"OwnerUserId":3241302.0,"Title":"create keras input layers using the functional api through a for loop?","Body":"<p>Suppose for a moment that I need to dynamically generate a keras model specific to a user. Each user can have a variable number of categorical inputs, but once you know the number of categorical inputs it would be trivial to hand build a model.<\/p>\n\n<p>I want to write a function that given a list of the cardinality of each of the categorical variables will return an appropriate model. My first stab at the problem produced the following solution, but it just doesn't seem right to execute a string(s) like this.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.layers import Dense,Embedding,Input,Flatten,Add\nfrom keras.models import Model\n\ndef build_model(input_cardinalities,num_outputs):\n    layers = []\n    inputs = []\n    for i,cardinality in enumerate(input_cardinalities):\n        exec(\"input{0} = Input(shape=[1], name='input{0}')\".format(i))\n        exec(\"embedding{0} =  Embedding({1}, 20, name='embedding{0}')(input{0})\".format(i,cardinality))\n        exec(\"vec{0} = Flatten(name='flatten{0}')(embedding{0})\".format(i))\n        exec(\"layers.append(vec{0})\".format(i))\n        exec(\"inputs.append(input{0})\".format(i))\n    context_layer = Add(layers)\n    dense1 = Dense(50, name='Dense1',activation='relu')(context_layer)\n    dense2 = Dense(num_outputs, name='Output', activation='softmax')(dense1)\n    model = Model(inputs,dense2)\n    model.compile('sgd','categorical_crossentropy')\n    return model\n<\/code><\/pre>\n\n<p>I just don't feel comfortable executing strings like this, but it is the only way that I can think of to do what I want. I just feel like there should be a better way of doing it. <\/p>\n","answers":[{"AnswerId":"57536630","CreationDate":"2019-08-17T13:15:16.960","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>There is no actual need to use <code>exec<\/code> at all, you are building one input\/embedding at a time, then storing them into lists. That is the correct way to do it, and it does not require <code>exec<\/code>.<\/p>\n\n<pre><code>def build_model(input_cardinalities,num_outputs):\n    layers = []\n    inputs = []\n    for i,cardinality in enumerate(input_cardinalities):\n        input = Input(shape=[1], name='input{0}'.format(i))\n        embedding =  Embedding(cardinality, 20, name='embedding{0}'.format(i))\n        vec = Flatten(name='flatten{0}'.format(i))(embedding)\n        layers.append(vec)\n        inputs.append(input)\n    context_layer = Add()(layers)\n    dense1 = Dense(50, name='Dense1',activation='relu')(context_layer)\n    dense2 = Dense(num_outputs, name='Output', activation='softmax')(dense1)\n    model = Model(inputs,dense2)\n    model.compile('sgd','categorical_crossentropy')\n    return model\n<\/code><\/pre>\n\n<p>Also note I corrected the <code>Add()(layers)<\/code> call.<\/p>\n"}]}
+{"QuestionId":57536593,"AnswerCount":0,"Tags":"<python><keras><ocr><pytorch>","CreationDate":"2019-08-17T13:12:10.787","AcceptedAnswerId":null,"OwnerUserId":8410477.0,"Title":"Convert pytorch ocr to keras ocr: _pickle.UnpicklingError: invalid load key, '\\xf4'","Body":"<p>I try to run the project from <a href=\"https:\/\/github.com\/chineseocr\/chineseocr\" rel=\"nofollow noreferrer\">here<\/a>, everything goes smoothly except at the step of converting <code>pytorch ocr<\/code> to <code>keras ocr<\/code>, with command: <code>python tools\/pytorch_to_keras.py  -weights_path models\/ocr-dense.pth -output_path models\/ocr-dense-keras.h5<\/code>. <\/p>\n\n<p>I got an error: <code>_pickle.UnpicklingError: invalid load key, '\\xf4'.<\/code><\/p>\n\n<p>I have commit an issue in github, but not get an reponse yet. So I put my question here, I'll appreciate your help. Thanks a lot.<\/p>\n\n<p>Detailed infos:<\/p>\n\n<pre><code>x@pop-os:~\/chineseocr$ python tools\/pytorch_to_keras.py  -weights_path models\/ocr-dense.pth -output_path models\/ocr-dense-keras.h5\nUsing TensorFlow backend.\n\/home\/x\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/dtypes.py:526: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint8 = np.dtype([(\"qint8\", np.int8, 1)])\n\/home\/x\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/dtypes.py:527: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint8 = np.dtype([(\"quint8\", np.uint8, 1)])\n\/home\/x\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/dtypes.py:528: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint16 = np.dtype([(\"qint16\", np.int16, 1)])\n\/home\/x\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/dtypes.py:529: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint16 = np.dtype([(\"quint16\", np.uint16, 1)])\n\/home\/x\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/dtypes.py:530: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint32 = np.dtype([(\"qint32\", np.int32, 1)])\n\/home\/x\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/dtypes.py:535: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  np_resource = np.dtype([(\"resource\", np.ubyte, 1)])\nWARNING:tensorflow:From \/home\/x\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/op_def_library.py:263: colocate_with (from tensorflow.python.framework.ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nColocations handled automatically by placer.\n2019-08-17 20:55:16.370693: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\n2019-08-17 20:55:16.377393: E tensorflow\/stream_executor\/cuda\/cuda_driver.cc:300] failed call to cuInit: CUDA_ERROR_NO_DEVICE: no CUDA-capable device is detected\n2019-08-17 20:55:16.377417: I tensorflow\/stream_executor\/cuda\/cuda_diagnostics.cc:161] retrieving CUDA diagnostic information for host: pop-os\n2019-08-17 20:55:16.377422: I tensorflow\/stream_executor\/cuda\/cuda_diagnostics.cc:168] hostname: pop-os\n2019-08-17 20:55:16.377449: I tensorflow\/stream_executor\/cuda\/cuda_diagnostics.cc:192] libcuda reported version is: 430.34.0\n2019-08-17 20:55:16.377465: I tensorflow\/stream_executor\/cuda\/cuda_diagnostics.cc:196] kernel reported version is: 430.34.0\n2019-08-17 20:55:16.377470: I tensorflow\/stream_executor\/cuda\/cuda_diagnostics.cc:303] kernel version seems to match DSO: 430.34.0\n2019-08-17 20:55:16.378844: I tensorflow\/core\/platform\/profile_utils\/cpu_utils.cc:94] CPU Frequency: 2208000000 Hz\n2019-08-17 20:55:16.379938: I tensorflow\/compiler\/xla\/service\/service.cc:150] XLA service 0x2750e40 executing computations on platform Host. Devices:\n2019-08-17 20:55:16.379969: I tensorflow\/compiler\/xla\/service\/service.cc:158]   StreamExecutor device (0): &lt;undefined&gt;, &lt;undefined&gt;\nTraceback (most recent call last):\n  File \"tools\/pytorch_to_keras.py\", line 89, in &lt;module&gt;\n    state_dict = torch.load(ocrModel,map_location=lambda storage, loc: storage)\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/torch\/serialization.py\", line 386, in load\n    return _load(f, map_location, pickle_module, **pickle_load_args)\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/torch\/serialization.py\", line 563, in _load\n    magic_number = pickle_module.load(f, **pickle_load_args)\n_pickle.UnpicklingError: invalid load key, '\\xf4'.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57536639,"AnswerCount":2,"Tags":"<python><keras><computer-vision><classification><conv-neural-network>","CreationDate":"2019-08-17T13:16:16.790","AcceptedAnswerId":null,"OwnerUserId":10402329.0,"Title":"Issues with compiling a Keras sequential model","Body":"<p>I came across the following error while compiling a Keras Sequential model:<\/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 45985 arrays: [array([[ 0.],\n       [ 0.],\n       [ 0.],\n       [ 0.],\n       [ 0.],\n       [ 0.],\n       [ 0.],\n       [ 0.],\n       [ 0.],\n       [ 0.],\n       [ 0.],\n       [ 0.],\n       [ 0.],\n       [ 0.],\n   ...\n<\/code><\/pre>\n\n<p>This is the code and the data format I am using for X_train, y_train, X_test and y_test:<\/p>\n\n<pre><code>print(X_train.shape)\n\n&gt;&gt;(45985, 50, 50, 3)\n\nprint(X_test.shape)\n\n&gt;&gt;(22650, 50, 50, 3)\n\nprint(y_train[0])\n\n&gt;&gt;array([ 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.])\n\n<\/code><\/pre>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense, Conv2D, Flatten\nmodel = Sequential()\n\nmodel.add(Conv2D(64, kernel_size=3, activation=\"relu\", input_shape=(50,50,3)))\nmodel.add(Conv2D(32, kernel_size=3, activation=\"relu\"))\nmodel.add(Flatten())\nmodel.add(Dense(10, activation=\"softmax\"))\n\nmodel.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])\nmodel.fit(X_train, y_train, validation_data=(X_test, y_test), epochs=3)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57538589","CreationDate":"2019-08-17T17:42:27.363","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>As an alternative to one-hot encoding the labels mentioned in the other answer, you can keep your labels as they are, i.e. sparse\/integer labels, and use <code>'sparse_categorical_crossentropy'<\/code> as the loss function instead. This would help with saving memory, especially if you have lots of samples and\/or classes. Though, don't forget that anyways you need to convert your labels to numpy array. <\/p>\n"},{"AnswerId":"57536719","CreationDate":"2019-08-17T13:28:15.773","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Your <code>y_train<\/code> as a list of numpy arrays, it should be a single numpy array with shape <code>(samples, 10)<\/code>. You can transform it with:<\/p>\n\n<pre><code>y_train = np.array(y_train, dtype=np.float32)\n<\/code><\/pre>\n\n<p>Then you should remember to one-hot encode your labels (they look like integer labels):<\/p>\n\n<pre><code>from keras.utils import to_categorical\n\ny_train = to_categorical(y_train)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57536778,"AnswerCount":1,"Tags":"<machine-learning><keras>","CreationDate":"2019-08-17T13:36:04.503","AcceptedAnswerId":null,"OwnerUserId":8416832.0,"Title":"How can i see validation accuracy ONLY at the last 3 epochs ? This is for reducing time","Body":"<p>I am working on cifar10 dataset with Resnet18 architecture. My need is now to reach 94 accuracy within 100 secs. \nSo i have been advised to check validation accuracy only for the last 3 epochs. Anyone help me how to do that ?<\/p>\n\n<pre><code>from tensorflow.keras.callbacks import ModelCheckpoint,LearningRateScheduler\n\nmodel.fit_generator(generator=gen,\n   steps_per_epoch=np.ceil(50000\/512),\n   epochs=24,\n   validation_data=validation_iterator, \n   verbose=1,\n   callbacks =[olr])\n<\/code><\/pre>\n\n<p>The above code checks validation accuracy for all 24 epochs, but i want to check ONLY for last 3 epochs.<\/p>\n","answers":[{"AnswerId":"57542645","CreationDate":"2019-08-18T07:45:21.647","ParentId":null,"OwnerUserId":"1805756","Title":null,"Body":"<p>You may break training into two separate phases in the latter you provide validation data:<\/p>\n\n<pre><code>model.fit_generator(generator=gen,\n   steps_per_epoch=np.ceil(50000\/512),\n   epochs=21,\n   verbose=1,\n   callbacks =[olr])\n\nmodel.fit_generator(generator=gen,\n   steps_per_epoch=np.ceil(50000\/512),\n   epochs=24,\n   validation_data=validation_iterator, \n   verbose=1,\n   callbacks =[olr],\n   initial_epoch=21)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57536920,"AnswerCount":0,"Tags":"<tensorflow><conv-neural-network><yolo><darknet><deep-residual-networks>","CreationDate":"2019-08-17T13:54:17.673","AcceptedAnswerId":null,"OwnerUserId":7193056.0,"Title":"ValueError: Variable yolov3\/darknet53_body\/Conv\/weights does not exist, or was not created with tf.get_variable()","Body":"<p>I am trying to understand Tensorflow implementation of YOLOv3 using command line from <a href=\"https:\/\/github.com\/wizyoung\/YOLOv3_TensorFlow\" rel=\"nofollow noreferrer\">this github repository<\/a><\/p>\n\n<p>This repository may be used for testing on single images. I am now trying to extend the code to work on multiple images. \nI have used a for loop in <code>test_single_image.py<\/code> to iterate through my test images folder. Below here is the code I am using:<\/p>\n\n<pre><code>from __future__ import division, print_function\n\nimport os\n\nimport tensorflow as tf\nimport numpy as np\nimport argparse\nimport cv2\n\nfrom utils.misc_utils import parse_anchors, read_class_names\nfrom utils.nms_utils import gpu_nms\nfrom utils.plot_utils import get_color_table, plot_one_box\nfrom utils.data_aug import letterbox_resize\n\nfrom model import yolov3\n\npath = '\/home\/ml\/Downloads\/old\/YOLOv3_TensorFlow\/data\/my_input\/'\nfor i in os.listdir(path):\n    parser = argparse.ArgumentParser(description=\"YOLO-V3 test single image test procedure.\")\n    parser.add_argument(\"--input_image\", type=str, default=path+i,\n                        help=\"The path of the input image.\")\n    parser.add_argument(\"--anchor_path\", type=str, default=\"\/home\/ml\/Downloads\/old\/YOLOv3_TensorFlow\/data\/yolo_anchors.txt\",\n                        help=\"The path of the anchor txt file.\")\n    parser.add_argument(\"--new_size\", nargs='*', type=int, default=[416, 416],\n                        help=\"Resize the input image with `new_size`, size format: [width, height]\")\n    parser.add_argument(\"--letterbox_resize\", type=lambda x: (str(x).lower() == 'true'), default=True,\n                        help=\"Whether to use the letterbox resize.\")\n    parser.add_argument(\"--class_name_path\", type=str, default=\"\/home\/ml\/Downloads\/old\/YOLOv3_TensorFlow\/data\/data.names\",\n                        help=\"The path of the class names.\")\n    parser.add_argument(\"--restore_path\", type=str, default=\"\/home\/ml\/Downloads\/old\/YOLOv3_TensorFlow\/data\/darknet_weightsbest_model_Epoch_68_step_621_mAP_0.9167_loss_2.0931_lr_1e-05\",\n                        help=\"The path of the weights to restore.\")\n    args = parser.parse_args()\n\n    args.anchors = parse_anchors(args.anchor_path)\n    args.classes = read_class_names(args.class_name_path)\n    args.num_class = len(args.classes)\n\n    color_table = get_color_table(args.num_class)\n\n    img_ori = cv2.imread(args.input_image)\n    if args.letterbox_resize:\n        img, resize_ratio, dw, dh = letterbox_resize(img_ori, args.new_size[0], args.new_size[1])\n    else:\n        height_ori, width_ori = img_ori.shape[:2]\n        img = cv2.resize(img_ori, tuple(args.new_size))\n    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n    img = np.asarray(img, np.float32)\n    img = img[np.newaxis, :] \/ 255.\n\n    with tf.Session() as sess:\n        input_data = tf.placeholder(tf.float32, [1, args.new_size[1], args.new_size[0], 3], name='input_data')\n        yolo_model = yolov3(args.num_class, args.anchors)\n        with tf.variable_scope('yolov3', reuse=True) as yolov3:\n            pred_feature_maps = yolo_model.forward(input_data, False, reuse=False)\n        pred_boxes, pred_confs, pred_probs = yolo_model.predict(pred_feature_maps)\n\n        pred_scores = pred_confs * pred_probs\n\n        boxes, scores, labels = gpu_nms(pred_boxes, pred_scores, args.num_class, max_boxes=200, score_thresh=0.1,\n                                        nms_thresh=0.45)\n\n        saver = tf.train.Saver()\n        saver.restore(sess, args.restore_path)\n        boxes_, scores_, labels_ = sess.run([boxes, scores, labels], feed_dict={input_data: img})\n\n        # rescale the coordinates to the original image\n        if args.letterbox_resize:\n            boxes_[:, [0, 2]] = (boxes_[:, [0, 2]] - dw) \/ resize_ratio\n            boxes_[:, [1, 3]] = (boxes_[:, [1, 3]] - dh) \/ resize_ratio\n        else:\n            boxes_[:, [0, 2]] *= (width_ori\/float(args.new_size[0]))\n            boxes_[:, [1, 3]] *= (height_ori\/float(args.new_size[1]))\n\n        print(\"box coords:\")\n        print(boxes_)\n        print('*' * 30)\n        print(\"scores:\")\n        print(scores_)\n        print('*' * 30)\n        print(\"labels:\")\n        print(labels_)\n\n        crop_img = img_ori[boxes_[0][1].astype('int'):boxes_[0][3].astype('int'), boxes_[0][0].astype('int'):boxes_[0][2].astype('int')]\n        # cv2.imshow(\"cropped\", crop_img)\n\n        for j in range(len(boxes_)):\n            x0, y0, x1, y1 = boxes_[j]\n            plot_one_box(img_ori, [x0, y0, x1, y1], label=args.classes[labels_[j]] + ', {:.2f}%'.format(scores_[j] * 100), color=color_table[labels_[j]])\n        # cv2.imshow('Detection result', img_ori)\n        cv2.imwrite(f'detection_{i}', img_ori)\n        cv2.imwrite(f'crop_{i}', crop_img)\n\ncv2.waitKey(0)\n<\/code><\/pre>\n\n<p>The loop iterates correctly during the first iteration but fails in the 2nd iteration saying:<\/p>\n\n<blockquote>\n  <p>ValueError: Variable yolov3\/darknet53_body\/Conv\/weights does not\n  exist, or was not created with tf.get_variable(). Did you mean to set\n  reuse=tf.AUTO_REUSE in VarScope?<\/p>\n<\/blockquote>\n\n<hr>\n\n<p>Also if I try to set <strong>reuse=True<\/strong> in the below statement:<\/p>\n\n<pre><code>with tf.variable_scope('yolov3', reuse=True) as yolov3:\n            pred_feature_maps = yolo_model.forward(input_data, False, reuse=True)\n<\/code><\/pre>\n\n<p>I get the below error saying<\/p>\n\n<blockquote>\n  <p>ValueError: reuse=True cannot be used without a name_or_scope<\/p>\n<\/blockquote>\n\n<p>even when I have set <code>name_or_scope<\/code> variable inside of <code>model.py<\/code> in <code>forward method()<\/code><\/p>\n\n<p>I am using PyCharm for debugging. Which debugger works here?<\/p>\n","answers":[]}
+{"QuestionId":57536984,"AnswerCount":1,"Tags":"<machine-learning><keras><deep-learning><classification>","CreationDate":"2019-08-17T14:01:55.370","AcceptedAnswerId":null,"OwnerUserId":6022269.0,"Title":"Connecting Keras model with no output","Body":"<p>I have 2 models that I am using for a classification:<\/p>\n\n<pre><code>morf_model = keras.Sequential([\n    keras.layers.Dense(800, activation=tf.nn.tanh , input_shape=([np.shape(x)[1]])),\n    keras.layers.Dense(800, activation=tf.nn.tanh),\n    keras.layers.Dense(600, activation=tf.nn.tanh),\n    keras.layers.Dense(300, activation=tf.nn.tanh),\n    keras.layers.Dense(50, activation=tf.nn.tanh),\n    keras.layers.Dense(2, activation=tf.nn.sigmoid)\n])\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>color_model = keras.Sequential([\n    keras.layers.Dense(800, activation=tf.nn.tanh , input_shape=([np.shape(col_x)[1]])),\n    keras.layers.Dense(800, activation=tf.nn.tanh),\n    keras.layers.Dense(600, activation=tf.nn.tanh),\n    keras.layers.Dense(300, activation=tf.nn.tanh),\n    keras.layers.Dense(50, activation=tf.nn.tanh),\n    keras.layers.Dense(2, activation=tf.nn.sigmoid)\n])\n<\/code><\/pre>\n\n<p>that I want to <strong>remove the output layer<\/strong> (the layer with 2 nodes),freeze them and connect it with a new model <\/p>\n\n<pre><code>model = keras.Sequential([\n    keras.layers.Dense(1000, activation=tf.nn.tanh , input_shape=([np.shape(last_x)[1]])),\n    keras.layers.Dense(800, activation=tf.nn.tanh),\n    keras.layers.Dense(600, activation=tf.nn.tanh),\n    keras.layers.Dense(300, activation=tf.nn.tanh),\n    keras.layers.Dense(50, activation=tf.nn.tanh),\n    keras.layers.Dense(2, activation=tf.nn.sigmoid)\n])\n<\/code><\/pre>\n\n<p>so the model will look like<\/p>\n\n<pre><code>morf_model ---|\n              |---&gt;model \ncolor_model  -|\n<\/code><\/pre>\n\n<p>is it possible? <\/p>\n\n<p>thankyou<\/p>\n","answers":[{"AnswerId":"57537132","CreationDate":"2019-08-17T14:22:30.783","ParentId":null,"OwnerUserId":"4517807","Title":null,"Body":"<p>Yes , \nUsing the <code>Keras functional API<\/code> brings you all possibilities.\ncheck it out here<\/p>\n\n<p><a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/#multi-input-and-multi-output-models\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/getting-started\/functional-api-guide\/#multi-input-and-multi-output-models<\/a><\/p>\n\n<p>You can remove the last layer of both the networks and still connect them , Now , make sure that your second last layer is the output layer.<\/p>\n\n<p>eg:<\/p>\n\n<pre><code>merge = Concatenate()([morf_model.output,color_model.output])\nmodel = Model(inputs=[morf_model.input,color_model.input], outputs=output)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57537065,"AnswerCount":1,"Tags":"<python><linux><tensorflow>","CreationDate":"2019-08-17T14:13:09.270","AcceptedAnswerId":null,"OwnerUserId":5832476.0,"Title":"ImportError: \/lib64\/libc.so.6: version `GLIBC_2.15' not found after importing tensorflow","Body":"<p>I'm setting up tensorflow in linux web hosting for flask application. I have installed tensorflow but when i import it i get the following errors\ni tried to upgrade GLIBC but still doesn't work.\ni have python 3.5.7 i cant upgrade it because i'm in web hosting\nall solution in <a href=\"https:\/\/stackoverflow.com\/questions\/50564999\/lib64-libc-so-6-version-glibc-2-14-not-found-why-am-i-getting-this-error\">\/lib64\/libc.so.6: version `GLIBC_2.14&#39; not found. Why am I getting this error?<\/a>\ndidnt work.<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/home\/malnetxy\/virtualenv\/Server\/3.5\/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\/malnetxy\/virtualenv\/Server\/3.5\/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\/malnetxy\/virtualenv\/Server\/3.5\/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\/malnetxy\/virtualenv\/Server\/3.5\/lib64\/python3.5\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/home\/malnetxy\/virtualenv\/Server\/3.5\/lib64\/python3.5\/imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: \/lib64\/libc.so.6: version `GLIBC_2.15' not found (required by \/home\/malnetxy\/virtualenv\/Server\/3.5\/lib\/python3.5\/site-packages\/tensorflow\/python\/_pywrap_tensorflow_internal.so)\nDuring handling of the above exception, another exception occurred:\nTraceback (most recent call last):\n  File \"&lt;string&gt;\", line 1, in &lt;module&gt;\n  File \"\/home\/malnetxy\/virtualenv\/Server\/3.5\/lib\/python3.5\/site-packages\/tensorflow\/__init__.py\", line 28, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"\/home\/malnetxy\/virtualenv\/Server\/3.5\/lib\/python3.5\/site-packages\/tensorflow\/python\/__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"\/home\/malnetxy\/virtualenv\/Server\/3.5\/lib\/python3.5\/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\/malnetxy\/virtualenv\/Server\/3.5\/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\/malnetxy\/virtualenv\/Server\/3.5\/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\/malnetxy\/virtualenv\/Server\/3.5\/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\/malnetxy\/virtualenv\/Server\/3.5\/lib64\/python3.5\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/home\/malnetxy\/virtualenv\/Server\/3.5\/lib64\/python3.5\/imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: \/lib64\/libc.so.6: version `GLIBC_2.15' not found (required by \/home\/malnetxy\/virtualenv\/Server\/3.5\/lib\/python3.5\/site-packages\/tensorflow\/python\/_pywrap_tensorflow_internal.so)\nFailed to load the native TensorFlow runtime.\n<\/code><\/pre>\n","answers":[{"AnswerId":"58119477","CreationDate":"2019-09-26T14:51:28.927","ParentId":null,"OwnerUserId":"12125807","Title":null,"Body":"<p>I recommend using an Anaconda environment. Then, you will be able to install Tensorflow and everything you need, also glibc executing: <code>conda install -c asmeurer glibc<\/code><\/p>\n"}]}
+{"QuestionId":57537288,"AnswerCount":1,"Tags":"<android><image><firebase><tensorflow><colors>","CreationDate":"2019-08-17T14:46:03.397","AcceptedAnswerId":57537487.0,"OwnerUserId":11856751.0,"Title":"How to normalize channel values to the range[0.0,1.0]?","Body":"<p>I'm new to Tensorflow and image processing, I use a code that converts bitmap to an Inputarray for predict object from Tensorflow Lite library and Firebase MLKit custom model API.\nI'm trying to normalize channels values to the range[0.0,1.0].\nComments are written show that this function normalizes channels values to a range of [-1.0,1.0] and can be normalized to [0.0,1.0] but doesn't explain how to do that.\nHow can I do that?<\/p>\n\n<p><strong>Update<\/strong><\/p>\n\n<p>The reason for that question is to get correct values from tflite model on an image from ibug dataset:\nI have a Tensorflow Lite model and I want to predict eye region landmarks values, I tried to predict eye region points with an image as input and with an array of values ( x and y coordinates )  as output.<\/p>\n\n<p>I got a python script that gets predictions from an image using Tensorflow lite model.<\/p>\n\n<pre><code>imgUrl = \"https:\/\/i.imgur.com\/Micp1bv.jpg\" # image from ibug dataset \nimg = Image.open(requests.get(imgUrl, stream=True).raw)\nimg.load()\nimg = img.resize((112, 112), PIL.Image.ANTIALIAS) # resize to width and height of input tensor parameters\n\n# Normalize to [0, 1]\ndata = np.asarray( img, dtype=\"float32\" ) \/255.0  \n\n# Inference on input data normalized to [0, 1]\ninputImg = np.expand_dims(data,0).astype(np.float32)\ninput_details = interpreter.get_input_details()\ninterpreter.set_tensor(input_details[0]['index'], inputImg)\n\ninterpreter.invoke()\n\noutput_details = interpreter.get_output_details()\noutput_data = interpreter.get_tensor(output_details[0]['index'])\nprint(output_data) \/\/ an arrays contains values ( x and y coordinates of landmarks) as an output \n\n<\/code><\/pre>\n\n<pre class=\"lang-java prettyprint-override\"><code>\n private float[][][][] bitmapToInputArray() {\n        \/\/ [START mlkit_bitmap_input]\n        Bitmap bitmap= getYourInputImage();\n        bitmap = Bitmap.createScaledBitmap(bitmap, 112, 112, true);\n\n        int batchNum = 0;\n        float[][][][] input = new float[1][112][112][3];\n        for (int x = 0; x &lt; 112; x++) {\n            for (int y = 0; y &lt; 112; y++) {\n                int pixel = bitmap.getPixel(x, y);\n                \/\/ Normalize channel values to [-1.0, 1.0]. This requirement varies by\n                \/\/ model. For example, some models might require values to be normalized\n                \/\/ to the range [0.0, 1.0] instead.\n                input[batchNum][x][y][0] = (Color.red(pixel) - 127) \/ 128.0f;\n                input[batchNum][x][y][1] = (Color.green(pixel) - 127) \/ 128.0f;\n                input[batchNum][x][y][2] = (Color.blue(pixel) - 127) \/ 128.0f;\n                Log.i(\"Input\",\"input\"+input[batchNum][x][y][0]);\n                Log.i(\"input\",\"input\"+input[batchNum][x][y][1]);\n\n            }\n        }\n        \/\/ [END mlkit_bitmap_input]\n\n        return input;\n    }\n\n<\/code><\/pre>\n\n<p>That was the output from python script:<\/p>\n\n<blockquote>\n  <p>[[0.21560352 0.37226492 0.23046923 0.3314699  0.26969466 0.31294012\n    0.31079206 0.32817352 0.32920814 0.367611   0.31699485 0.41163784\n    0.27505988 0.42706913 0.23404554 0.4119034  0.15523753 0.3914298\n    0.18589666 0.37025875 0.22555524 0.3547908  0.27040404 0.3469348\n    0.3122904  0.35289326 0.34334075 0.3686381  0.36434904 0.3907298\n    0.34113637 0.40280795 0.3069372  0.40720087 0.26726058 0.4101103\n    0.22570357 0.40932944 0.1866894  0.40394992 0.6993889  0.3493917\n    0.7078744  0.3110081  0.73927087 0.2935116  0.77587724 0.30691338\n    0.79591703 0.3434749  0.787707   0.38200834 0.7559439  0.3995327\n    0.7166693  0.38873363 0.6655569  0.3727552  0.68886995 0.3488456\n    0.720377   0.3310734  0.75820756 0.3226342  0.79140866 0.3270979\n    0.8144976  0.33877644 0.8281598  0.3576495  0.81746304 0.37150782\n    0.7971135  0.37963405 0.7682784  0.3838213  0.73269796 0.38409975\n    0.69557863 0.3825233 ]]<\/p>\n<\/blockquote>\n\n<p>And this is the output from Android Studio ( Java code ):<\/p>\n\n<blockquote>\n  <p>2019-08-17 17:19:18.754 19647-19647\/com.example.irisdetection I\/MLKit: 0,22430149\n  2019-08-17 17:19:18.756 19647-19647\/com.example.irisdetection I\/MLKit: 0,23864979\n  2019-08-17 17:19:18.758 19647-19647\/com.example.irisdetection I\/MLKit: 0,27703676\n  2019-08-17 17:19:18.759 19647-19647\/com.example.irisdetection I\/MLKit: 0,31750143\n  2019-08-17 17:19:18.760 19647-19647\/com.example.irisdetection I\/MLKit: 0,33583546\n  2019-08-17 17:19:18.761 19647-19647\/com.example.irisdetection I\/MLKit: 0,32406592\n  2019-08-17 17:19:18.762 19647-19647\/com.example.irisdetection I\/MLKit: 0,28302023\n  2019-08-17 17:19:18.763 19647-19647\/com.example.irisdetection I\/MLKit: 0,24265678\n  2019-08-17 17:19:18.764 19647-19647\/com.example.irisdetection I\/MLKit: 0,16428351\n  2019-08-17 17:19:18.765 19647-19647\/com.example.irisdetection I\/MLKit: 0,19479913\n  2019-08-17 17:19:18.766 19647-19647\/com.example.irisdetection I\/MLKit: 0,23414856\n  2019-08-17 17:19:18.767 19647-19647\/com.example.irisdetection I\/MLKit: 0,27845544\n  2019-08-17 17:19:18.767 19647-19647\/com.example.irisdetection I\/MLKit: 0,31952256\n  2019-08-17 17:19:18.774 19647-19647\/com.example.irisdetection I\/MLKit: 0,34978011\n  2019-08-17 17:19:18.775 19647-19647\/com.example.irisdetection I\/MLKit: 0,37012532\n  2019-08-17 17:19:18.776 19647-19647\/com.example.irisdetection I\/MLKit: 0,34773278\n  2019-08-17 17:19:18.777 19647-19647\/com.example.irisdetection I\/MLKit: 0,31446189\n  2019-08-17 17:19:18.778 19647-19647\/com.example.irisdetection I\/MLKit: 0,27557194\n  2019-08-17 17:19:18.778 19647-19647\/com.example.irisdetection I\/MLKit: 0,23455118\n  2019-08-17 17:19:18.780 19647-19647\/com.example.irisdetection I\/MLKit: 0,19578205\n  2019-08-17 17:19:18.782 19647-19647\/com.example.irisdetection I\/MLKit: 0,69702154\n  2019-08-17 17:19:18.785 19647-19647\/com.example.irisdetection I\/MLKit: 0,70499951\n  2019-08-17 17:19:18.786 19647-19647\/com.example.irisdetection I\/MLKit: 0,73528731\n  2019-08-17 17:19:18.787 19647-19647\/com.example.irisdetection I\/MLKit: 0,77087343\n  2019-08-17 17:19:18.788 19647-19647\/com.example.irisdetection I\/MLKit: 0,79054081\n  2019-08-17 17:19:18.790 19647-19647\/com.example.irisdetection I\/MLKit: 0,78286630\n  2019-08-17 17:19:18.791 19647-19647\/com.example.irisdetection I\/MLKit: 0,75219637\n  2019-08-17 17:19:18.792 19647-19647\/com.example.irisdetection I\/MLKit: 0,71401721\n  2019-08-17 17:19:18.793 19647-19647\/com.example.irisdetection I\/MLKit: 0,66399622\n  2019-08-17 17:19:18.794 19647-19647\/com.example.irisdetection I\/MLKit: 0,68674380\n  2019-08-17 17:19:18.795 19647-19647\/com.example.irisdetection I\/MLKit: 0,71746421\n  2019-08-17 17:19:18.797 19647-19647\/com.example.irisdetection I\/MLKit: 0,75426823\n  2019-08-17 17:19:18.798 19647-19647\/com.example.irisdetection I\/MLKit: 0,78634208\n  2019-08-17 17:19:18.799 19647-19647\/com.example.irisdetection I\/MLKit: 0,80832106\n  2019-08-17 17:19:18.801 19647-19647\/com.example.irisdetection I\/MLKit: 0,82105815\n  2019-08-17 17:19:18.802 19647-19647\/com.example.irisdetection I\/MLKit: 0,81131995\n  2019-08-17 17:19:18.805 19647-19647\/com.example.irisdetection I\/MLKit: 0,79208875\n  2019-08-17 17:19:18.807 19647-19647\/com.example.irisdetection I\/MLKit: 0,76439416\n  2019-08-17 17:19:18.809 19647-19647\/com.example.irisdetection I\/MLKit: 0,72983956\n  2019-08-17 17:19:18.810 19647-19647\/com.example.irisdetection I\/MLKit: 0,69351745<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"57537487","CreationDate":"2019-08-17T15:12:43.603","ParentId":null,"OwnerUserId":"4903534","Title":null,"Body":"<p>If the pixel values range from <code>0<\/code> to <code>255<\/code> and you want them to be normalized between <code>0.0f<\/code> and <code>1.0f<\/code> you should just divide the pixel values with 255.<\/p>\n\n<p>So you should have<\/p>\n\n<pre><code>input[batchNum][x][y][0] = Color.red(pixel) \/ 255.0f;\ninput[batchNum][x][y][1] = Color.green(pixel) \/ 255.0f;\ninput[batchNum][x][y][2] = Color.blue(pixel) \/ 255.0f;\n<\/code><\/pre>\n\n<p>instead of <\/p>\n\n<pre><code>input[batchNum][x][y][0] = (Color.red(pixel) - 127) \/ 128.0f;\ninput[batchNum][x][y][1] = (Color.green(pixel) - 127) \/ 128.0f;\ninput[batchNum][x][y][2] = (Color.blue(pixel) - 127) \/ 128.0f;\n<\/code><\/pre>\n\n<p>This is because if your pixel value is at its maximum it will be <code>255<\/code> so if you divide by <code>255.0f<\/code> the result will be <code>1.0f<\/code>, and if your pixel value is at its minimum so it is equal to <code>0<\/code> if you divide with <code>255.0f<\/code> you will get <code>0f<\/code>.<\/p>\n"}]}
+{"QuestionId":57537474,"AnswerCount":1,"Tags":"<python><python-3.x><numpy><tensorflow><keras>","CreationDate":"2019-08-17T15:09:53.717","AcceptedAnswerId":57537534.0,"OwnerUserId":11940153.0,"Title":"How to fix 'AttributeError: 'list' object has no attribute 'shape'' error in python with Tensorflow \/ Keras when loading Model","Body":"<p>Im trying to save my Model in Keras and then load it but when it try to use the loaded Model it trows an Error<\/p>\n\n<p>Python Vesion: 3.6.8\nTensorflow Version: 2.0.0-beta1\nKeras Version: 2.2.4-tf<\/p>\n\n<p>Here is my Code:<\/p>\n\n<pre><code>from __future__ import absolute_import, division, print_function, unicode_literals\n\nimport tensorflow as tf\nfrom tensorflow import keras\n\nprint(\"Tensorflow Version:\")\nprint(tf.version.VERSION)\nprint(\"Keras Verstion:\")\nprint(tf.keras.__version__)\n\nminst = tf.keras.datasets.mnist # 28x28 0-9\n\n(x_train, y_train), (x_test, y_test) = minst.load_data()\n\nx_train = tf.keras.utils.normalize(x_train, axis=1)\nx__test = x_test;\nx_test = tf.keras.utils.normalize(x_test, axis=1)\n\nmodel = tf.keras.models.Sequential()\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', loss='sparse_categorical_crossentropy', metrics=['accuracy'])\n\nmodel.fit(x_train, y_train, epochs=1)\n\nval_loss, val_acc = model.evaluate(x_test, y_test)\nprint(val_loss, val_acc)\n\nmodel.save('epic_num.h5')\nprint(\"Saved\")\n# Loading\n\nnmodel = keras.models.load_model(\"epic_num.h5\")\nimport numpy as np;\n\n\n# Test\npredics = nmodel.predict([x_test])\nprint(predics)\n\nimport matplotlib.pyplot as plt\n\nplt.imshow(x__test[11], cmap=plt.cm.binary)\nplt.show()\n\nprint(np.argmax(predics[11]))\n<\/code><\/pre>\n\n<p>The Output:<\/p>\n\n<pre><code>C:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\Scripts\\python.exe C:\/Users\/minec\/PycharmProjects\/TFMLTest3\/main.py\nC:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:516: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint8 = np.dtype([(\"qint8\", np.int8, 1)])\nC:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:517: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint8 = np.dtype([(\"quint8\", np.uint8, 1)])\nC:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:518: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint16 = np.dtype([(\"qint16\", np.int16, 1)])\nC:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:519: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint16 = np.dtype([(\"quint16\", np.uint16, 1)])\nC:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:520: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint32 = np.dtype([(\"qint32\", np.int32, 1)])\nC:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:525: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  np_resource = np.dtype([(\"resource\", np.ubyte, 1)])\nC:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:541: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint8 = np.dtype([(\"qint8\", np.int8, 1)])\nC:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:542: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint8 = np.dtype([(\"quint8\", np.uint8, 1)])\nC:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:543: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint16 = np.dtype([(\"qint16\", np.int16, 1)])\nC:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:544: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint16 = np.dtype([(\"quint16\", np.uint16, 1)])\nC:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:545: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint32 = np.dtype([(\"qint32\", np.int32, 1)])\nC:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:550: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  np_resource = np.dtype([(\"resource\", np.ubyte, 1)])\nTensorflow Version:\n2.0.0-beta1\nKeras Verstion:\n2.2.4-tf\n2019-08-17 17:02:25.535385: I tensorflow\/core\/platform\/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2\nWARNING: Logging before flag parsing goes to stderr.\nW0817 17:02:25.636994 14512 deprecation.py:323] From C:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorflow\\python\\ops\\math_grad.py:1250: add_dispatch_support.&lt;locals&gt;.wrapper (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse tf.where in 2.0, which has the same broadcast rule as np.where\nTrain on 60000 samples\n\n   32\/60000 [..............................] - ETA: 2:39 - loss: 2.3092 - accuracy: 0.0938\n 1120\/60000 [..............................] - ETA: 7s - loss: 1.7317 - accuracy: 0.5580  \n...\naccuracy: 0.9590\n 9600\/10000 [===========================&gt;..] - ETA: 0s - loss: 0.1232 - accuracy: 0.9624\n10000\/10000 [==============================] - 0s 32us\/sample - loss: 0.1271 - accuracy: 0.9610\n0.1271383908316493 0.961\nSaved\nW0817 17:02:29.220302 14512 hdf5_format.py:197] Sequential models without an `input_shape` passed to the first layer cannot reload their optimizer state. As a result, your model isstarting with a freshly initialized optimizer.\nTraceback (most recent call last):\n  File \"C:\/Users\/minec\/PycharmProjects\/TFMLTest3\/main.py\", line 41, in &lt;module&gt;\n    predics = nmodel.predict([x_test])\n  File \"C:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 821, in predict\n    use_multiprocessing=use_multiprocessing)\n  File \"C:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_arrays.py\", line 705, in predict\n    x, check_steps=True, steps_name='steps', steps=steps)\n  File \"C:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 2335, in _standardize_user_data\n    self._set_inputs(cast_inputs)\n  File \"C:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 2553, in _set_inputs\n    outputs = self(inputs, **kwargs)\n  File \"C:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer.py\", line 662, in __call__\n    outputs = call_fn(inputs, *args, **kwargs)\n  File \"C:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\sequential.py\", line 262, in call\n    outputs = layer(inputs, **kwargs)\n  File \"C:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer.py\", line 662, in __call__\n    outputs = call_fn(inputs, *args, **kwargs)\n  File \"C:\\Users\\minec\\PycharmProjects\\TFMLTest3\\venv\\lib\\site-packages\\tensorflow\\python\\keras\\layers\\core.py\", line 580, in call\n    inputs, (tensor_shape.dimension_value(inputs.shape[0]) or\nAttributeError: 'list' object has no attribute 'shape'\n\nProcess finished with exit code 1\n<\/code><\/pre>\n","answers":[{"AnswerId":"57537534","CreationDate":"2019-08-17T15:19:25.023","ParentId":null,"OwnerUserId":"5707563","Title":null,"Body":"<p>Try using<\/p>\n\n<pre><code>nmodel.predict(x_test)\n<\/code><\/pre>\n\n<p>instead of<\/p>\n\n<pre><code>nmodel.predict([x_test])\n<\/code><\/pre>\n\n<p>(remove the brackets).<\/p>\n"}]}
+{"QuestionId":57538083,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-17T16:31:53.980","AcceptedAnswerId":null,"OwnerUserId":2659892.0,"Title":"Tensorflow 2.0: The examples don't train","Body":"<p>I'm trying to run the DCGAN, and other examples too in this tutorial:\n<a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/generative\/dcgan\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/tutorials\/generative\/dcgan<\/a><\/p>\n\n<p>The training is non existant or VERY slow. it looks like it's stuck.<\/p>\n\n<p>I've modified the <code>train<\/code> function. the output is <code>(256,28,28,1)<\/code> (the output shape), then, <code>a<\/code> then <code>b<\/code>. and it goes like that for a very long time until I stop it. Not finishing an epoch. (it prints that very slowly)<\/p>\n\n<p>See modified version below:<\/p>\n\n<pre><code>def train(dataset, epochs):\n    for epoch in range(epochs):\n        start = time.time()\n\n        for image_batch in dataset: #stuck in an infinite loop\n            print('a')\n            train_step(image_batch)\n            print('b')\n\n        print('c')\n        # Produce images for the GIF as we go\n        display.clear_output(wait=True)\n        generate_and_save_images(generator,epoch + 1,seed)\n        print('d')\n\n        # Save the model every 1 epochs\n        if (epoch + 1) % 1 == 0:\n            checkpoint.save(file_prefix = checkpoint_prefix)\n\n        print ('Time for epoch {} is {} sec'.format(epoch + 1, time.time()-start))\n<\/code><\/pre>\n\n<p>Also, I'm using <code>tensorflow-gpu==2.0.0-beta1<\/code>, and using a <code>NVIDIA GTX 960M<\/code>.<\/p>\n\n<p>Side note: on Google Collab it obviously works, but I want it to work on my Laptop.<\/p>\n\n<p>Also on CycleGAN here's a screenshot of the training:\n<a href=\"https:\/\/i.stack.imgur.com\/MbkHX.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/MbkHX.png\" alt=\"gan\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57538530,"AnswerCount":0,"Tags":"<python><tensorflow><keras><reinforcement-learning><openai-gym>","CreationDate":"2019-08-17T17:35:07.940","AcceptedAnswerId":null,"OwnerUserId":2258170.0,"Title":"DQN opengym cart-pole-v0 with tensorflow keras : model does not converge","Body":"<p>I've tried to solve the standard cartpole example from openai's gym using DQN with experience replay and epsilon decay and I can't seem to make it converge, in fact, the loss goes up exponentially???<\/p>\n\n<p>I've tried looking at this example but I didn't see too much difference between my code and his.\n<a href=\"https:\/\/github.com\/NoumaanKaleem\/cartpole_ddqn\/blob\/master\/deep-q-learning\/dqn.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/NoumaanKaleem\/cartpole_ddqn\/blob\/master\/deep-q-learning\/dqn.py<\/a><\/p>\n\n<p>Here is my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nimport tensorflow as tf\nimport numpy as np\n\n#self.mem.add([state,at,reward,next_state,done])\nclass SARSD:\n    count = 0\n    def __init__(self, state, at, reward, next_state, done):\n        self.state = state\n        self.at = at\n        self.reward = reward\n        self.next_state = next_state\n        self.done = done\n        self.ID = SARSD.count\n        SARSD.count = SARSD.count + 1\n\nclass ReplayMemory:\n    def __init__(self, N):\n        self.mem = []\n        self.N = N\n\n    #add element to replay memory\n    #if the size of the replay memory is alrady capped, remove first item added\n    def add(self, elem):\n        self.mem.append(elem)\n        if len(self.mem) &gt; self.N:\n            self.mem.pop(0)\n\n    def sample(self, num_elems):\n        mem_length = len(self.mem)\n        if num_elems &gt; mem_length:\n            return -1\n\n        indices = np.random.randint(0,mem_length-1,num_elems)\n\n        new_array = []\n        to_remove = []\n        for i in indices:\n            new_array.append(self.mem[i])\n            to_remove.append(self.mem[i])\n\n        self.mem = [value for value in self.mem if value not in to_remove]\n        return new_array\n\n    def full(self):\n        return len(self.mem) == self.N\n\n    def clear(self):\n        self.mem.clear()\n\n\nclass DQNAgent:\n    def __init__(self):\n        self.mem = ReplayMemory(2048)\n\n        #epsilon greedy parameter\n        self.eps = 0.1\n\n        #learning rate\n        self.gamma = 0.95\n\n        #minibatch size\n        self.minibatch_size = 32\n\n        #epochs\n        self.epochs = 1\n\n    def learn(self, num_episodes, env, model):\n        self.mem = ReplayMemory(1000)\n\n        for episode in range(0,num_episodes):\n\n            #if episode % int(num_episodes\/10) == 0:\n                #self.eps = self.eps - 0.2\n            #if self.eps &lt; 0.05:\n                #self.eps = 0.05\n\n            state = env.reset()\n            state = np.reshape(state, [1,4])\n            done = False\n\n            while done == False:\n                at = 0\n                if np.random.rand() &lt; self.eps:\n                    #sample a random action\n                    at = env.action_space.sample()\n                else:\n                    #get the best action from policy\n                    test = model.predict(state)\n                    at = np.argmax(model.predict(state))\n\n                #execute action at\n                next_state, reward, done, info = env.step(at)\n                next_state = np.reshape(next_state, [1,4])\n\n                #store the transition into replay memory\n                data = SARSD(state,at,reward,next_state,done)\n                self.mem.add(data)\n\n                #EDIT1 : surely this wasn't helping\n                state = next_state\n\n                if self.mem.full():\n                    #sample batch of transitions from replay memory\n                    minibatch = self.mem.sample(self.minibatch_size)\n\n                    #intialize the targets and data\n                    y = []\n                    x = []\n\n                    for elem in minibatch:\n                        #check if terminal state\n                        terminal = elem.done\n\n                        if terminal == True:\n                            yj = elem.reward\n                        else:\n                            #do the discounted reward formula\n                            yj = elem.reward\n                            prediction = model.predict(elem.next_state)\n                            tmp = np.max(model.predict(elem.next_state))\n                            yj = yj + self.gamma * np.max(model.predict(elem.next_state))\n\n                        #get the prediction\n                        prediction = model.predict(elem.state)\n\n                        #change the prediction action value with discounted reward\n                        prediction[0][elem.at] = yj\n\n                        y.append(prediction[0])\n                        x.append(elem.state[0])\n\n                    #fit the model with target y\n                    x = np.array(x)\n                    y = np.array(y)\n                    model.fit(x,y,epochs=self.epochs)\n\n<\/code><\/pre>\n\n<p>Here is the main python file where I declare the keras model to be trained...<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nimport numpy as np\nimport gym\nimport DQNAgent\n\nimport tensorflow as tf\n\ntf.logging.set_verbosity(tf.logging.ERROR)\n\nenv = gym.make('CartPole-v0')\nagent = DQNAgent.DQNAgent()\n\nmodel = tf.keras.Sequential()\nmodel.add(tf.keras.layers.Dense(24, activation=tf.keras.activations.sigmoid, input_dim=4))\nmodel.add(tf.keras.layers.Dense(24, activation=tf.keras.activations.sigmoid))\nmodel.add(tf.keras.layers.Dense(2,activation=tf.keras.activations.linear))\n\nmodel.compile(optimizer=tf.train.AdamOptimizer(0.01),\n              loss=tf.keras.losses.mean_squared_error)\n\nagent.learn(1000,env,model)\n\n#evaluate if model has been trained correctly\n\nepisode = 0\nfor episode in range(0,1000):\n    next_state = env.reset()\n\n\n    done = False\n    while done == False:\n        env.render()\n\n        next_state = np.reshape(next_state, [1, 4])\n        at = np.argmax(model.predict(next_state))\n\n        next_state, reward, done, info = env.step(at)\n<\/code><\/pre>\n\n<p>I'm kind of new with Tensorflow\/Keras as you could guess.<\/p>\n\n<p>EDIT 1 : Forgot to actually assign state=next_state. This is helping the losses to stabilize but it still doesn't solve the cartpole problem... <\/p>\n","answers":[]}
+{"QuestionId":57538810,"AnswerCount":3,"Tags":"<python><keras><deep-learning><dataset>","CreationDate":"2019-08-17T18:15:42.523","AcceptedAnswerId":null,"OwnerUserId":10764087.0,"Title":"What input shape should I take in first layer of Sequential model when the dimensions of the images are (2048*1536)","Body":"<p>I am having an image dataset each image is of dimensions=(2048,1536).In ImageDataGenerator to fetch data from the directory, I have used the same target size i.e (2048,1536) but while making Sequential model first layer, what input shape should I have to use?? Will it be same as (2048,1536) or I can take any random shape like (224,224).<\/p>\n","answers":[{"AnswerId":"57540432","CreationDate":"2019-08-17T22:43:46.770","ParentId":null,"OwnerUserId":"1830648","Title":null,"Body":"<p>The first(i.e. input)layer is supposed to be the number of features in your dataset. Regarding images, each pixel is considered as a feature. Hence in your case, the image dimension is (2048,1536) you need to flatten it out to get the total number of the pixel(i.e. features). If it is greyscaled image it would be (2048*1536*1) else if it is colour it would be(2048*1536*3).<\/p>\n\n<p>Also, you use below code from TensorFlow\/Keras API while Sequential model creation and it will take care of your input layer size <\/p>\n\n<pre><code>tf.keras.models.Sequential([tf.keras.layers.Flatten(),\n                       tf.keras.layers.Dense(128,activation=tf.nn.relu) #1st hidden layer\n                       tf.keras.layers.Dense(128,activation=tf.nn.relu) #2nd hidden layer\n                       tf.keras.layers.Dense(2,activation=tf.nn.softmax)])#output layer\n<\/code><\/pre>\n"},{"AnswerId":"57538888","CreationDate":"2019-08-17T18:26:56.363","ParentId":null,"OwnerUserId":"2255479","Title":null,"Body":"<p>You should probably flatten your input data by making a vector of size 3145728 (2048 * 1536). If your data is in a numpy array you can use <code>np.flatten()<\/code> (<a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.ndarray.flatten.html\" rel=\"nofollow noreferrer\">numpy flatten<\/a>).<\/p>\n\n<p>Then your first layer can have the same shape as this vector.<\/p>\n"},{"AnswerId":"57539279","CreationDate":"2019-08-17T19:20:40.470","ParentId":null,"OwnerUserId":"8809209","Title":null,"Body":"<p>I would resize first the images with <code>cv2.resize()<\/code>. Do you really need all the information from such a big image?<\/p>\n\n<p>For a sequential Model it follows for example:\n<code>\nmodel = models.Sequential()\nmodel.add(layers.Conv2D(32,(3,3), activation='relu', input_shape = (height,width, ndim)))\n...<\/code>,<\/p>\n\n<p>where <code>height<\/code> and <code>width<\/code> denote your input image dimensions and <code>ndim = 1<\/code> for greyscale and <code>ndim = 3<\/code> for colored images.<\/p>\n"}]}
+{"QuestionId":57539099,"AnswerCount":0,"Tags":"<visual-studio-code><pytorch><vscode-remote>","CreationDate":"2019-08-17T18:54:18.213","AcceptedAnswerId":null,"OwnerUserId":10178720.0,"Title":"How to go to definition of a pytorch function in vscode","Body":"<p>When I try the vscode remote SSH, I find that it can not go to definition of some pytorch function such as <code>torch.device<\/code>, but some function is normal such as <code>torch.tensor<\/code>. <\/p>\n\n<p>Is it a error or it is normal for vscode? Thanks!<\/p>\n","answers":[]}
+{"QuestionId":57539204,"AnswerCount":2,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-08-17T19:11:29.620","AcceptedAnswerId":null,"OwnerUserId":10852334.0,"Title":"Understanding Keras LSTM model with varying size of batches","Body":"<p>I have a working LSTM model with Keras, but I need more control over things so I'm transforming to tensorflow (1.13).<\/p>\n\n<p>In the way of doing so, first thing I encountered was the batch size handling.<\/p>\n\n<p>This is the Keras' simple model:<\/p>\n\n<pre><code>model1 = Sequential()\nmodel1.add(LSTM(64, input_shape=(seq_length, X_train.shape[2]),return_sequences=False))\nmodel1.add(Dense(y_train.shape[2], activation='softmax'))\nmodel1.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n# y_train[:,-1,:] takes only the fifth timestap's y of each sample\nmodel1.fit(X_train, y_train[:,-1,:], epochs=300, batch_size=512)\n<\/code><\/pre>\n\n<p>I'm using <code>batch_size<\/code> of 512 for example, while the number of samples in <code>X_train<\/code> is 108765, a number that can't be divided (fully) by 512, which means that the last epoch's last step will be 221 instead of 512. \nKeras handles this in the back and I didn't need to do anything for it.\nWhen using TensorFlow, I need to specify a batch size for the initial state - which will be used throughout the epoch as the output state of <code>tf.nn.dynamic_rnn<\/code>.   <\/p>\n\n<p>So:<\/p>\n\n<ol>\n<li><p>How does Keras handle this?<\/p><\/li>\n<li><p>What can be done in TensorFlow order to overcome this, without losing data to fit the batch size?<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"57540587","CreationDate":"2019-08-17T23:20:16.293","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>It is a bit complicated to go over everything, but let's start with how can you do this with Tensorflow. So, firstly same constants:<\/p>\n\n<pre><code># Input data\nX = np.random.rand(108765, 10, 3).astype(np.float32)\n# The number of epochs\nepochs = 10\n# The size of the batch, in your case 512\nbatch_size = 512\n# Size of the cell, 64 as per your code\ncell_size = 64\n<\/code><\/pre>\n\n<p>Now, the data loading logic. We can create a <code>tf.data.Dataset()<\/code>, that will take care of the loading of the data automatically:<\/p>\n\n<pre><code># A dataset from a tensor\ndataset = tf.data.Dataset.from_tensor_slices(X)\n# Shuffle the dataset with some arbitrary buffer size\ndataset = dataset.shuffle(buffer_size=10)\n# Divide the dataset into batches. Once you reach the last batch which won't be 512, the dataset will know exactly which elements remain and should be passed as a batch.\ndataset = dataset.batch(batch_size)\n# An iterator that can be reinitialized over and over again, therefore having a new shuffle of the data each time\niterator = dataset.make_initializable_iterator()\n# A node that can be run to obtain the next element in the dataset. However, this node will be linked in the model so obtaining the next element will be done automatically\ndata_X = iterator.get_next()\n<\/code><\/pre>\n\n<p>The final part of the model is the <code>tf.nn.dynamic_rnn<\/code> itself:<\/p>\n\n<pre><code>cell = tf.nn.rnn_cell.LSTMCell(cell_size)\ncurrent_batch_size = tf.shape(data_X)[0]\ninit_state = cell.zero_state(current_batch_size, tf.float32)\noutputs, states = tf.nn.dynamic_rnn(cell=cell, inputs=data_X, initial_state=init_state)\n<\/code><\/pre>\n\n<p>Now, we are set to create the training logic:<\/p>\n\n<pre><code># Creation of a session\nwith tf.Session() as sess:\n    # Initialization of all variables in the TF graph\n    sess.run(tf.global_variables_initializer())\n    # Executing the block below epoch times\n    for e in range(epochs):\n        # Each time, reinitialize the iterator to obtain a fresh shuffle of the training data\n        sess.run(iterator.initializer)\n        try:\n            # As long as there are elements execute the block below\n            while True:\n               # The whole training logic\n        except tf.errors.OutOfRangeError:\n            pass\n<\/code><\/pre>\n\n<p>I assume that this code should help you to assemble your own logic to create a TF model and train it. As for how <code>keras<\/code> does things in the background, I don't know exactly. Similarly as TF, let's assume that it knows which elements have passed, and which elements remain in the data loading module.<\/p>\n\n<p>Finally, I want to point out that all these things are complicated for themselves and you should do a bit of reading on your own. In particular, the purpose of the answer is to help you understand how can you do the data loading without loosing any information. Good luck!<\/p>\n"},{"AnswerId":"57540933","CreationDate":"2019-08-18T00:53:41.190","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>There are two types of RNN in Keras:   <\/p>\n\n<ul>\n<li><code>stateful=True<\/code>   <\/li>\n<li><code>stateful=False<\/code> (your case according to the code you posted)   <\/li>\n<\/ul>\n\n<p>The difference between them is that the <code>True<\/code> version will keep the states in memory between batches (to simulate that the second batch is a sequel of the first batch, for instance), while the <code>False<\/code> version will create a new state matrix for every batch (so every batch contains full sequences, not parts of sequences).    <\/p>\n\n<p>So, in the <code>True<\/code> case, Keras does face the same problem as you do: it needs a fixed batch size, and in fact it demands that you specify the batch size when you use <code>stateful=True<\/code>.   <\/p>\n\n<p>But, if you're using <code>stateful=False<\/code>, it will just create a new states matrix full of zeros. <\/p>\n\n<p>So, basically:<\/p>\n\n<ul>\n<li>If you want to create a Tensorflow <code>stateful=True<\/code> layer, you need the batch size to be constant, just as Keras also does    <\/li>\n<li>If you want to create a Tensorflow <code>stateful=False<\/code> layer, you can just create new states as an all-zeros matrix with shape <code>(samples, output_dim)<\/code>    <\/li>\n<\/ul>\n"}]}
+{"QuestionId":57539273,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow><keras>","CreationDate":"2019-08-17T19:20:05.793","AcceptedAnswerId":57546746.0,"OwnerUserId":8419797.0,"Title":"Disable Tensorflow logging completely","Body":"<p>I'm using Keras and Tensorflow for neural networks, however Tensorflow is doubling logging from my main logger, which uses a StreamHandler, and its warnings cannot be disabled entirely.<\/p>\n\n<p>I tried manually disabling the messages and logging from the very beginning of the imports and it did disable a LOT of messages Keras(<code>Using Tensorflow backend<\/code>) and Tensorflow were sending but, the doubling issue happens.<\/p>\n\n<p>This is what I'm currently using.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import os\nimport logging\nos.environ['TF_CPP_MIN_LOG_LEVEL'] = '4'\nlogging.getLogger('tensorflow').disabled = True\nimport sys\nstderr = sys.stderr\nsys.stderr = open(os.devnull, 'w')\nimport keras\nsys.stderr = stderr\nfrom keras.layers import Activation, Dense, Dropout\nfrom keras.models import Sequential, load_model\nfrom keras.optimizers import SGD\nfrom keras.backend.tensorflow_backend import tf\nlogger = tf.get_logger()\nlogger.disabled = True\nlogger.setLevel(logging.FATAL)\n<\/code><\/pre>\n\n<p>That code is still ugly and immense and replacing sys.stderr even if it's temporary is personally very unprofessional. The worst part is, just a part of my issue was fixed, my console still outputs this:<\/p>\n\n<pre><code>WARNING: Logging before flag parsing goes to stderr.\nI0817 17:59:30.386165  5752 social.py:42] Social loaded\n  INFO     | Systems ready\nI0817 17:59:30.401529  5752 __init__.py:63] Systems ready\n<\/code><\/pre>\n\n<p>The only message it's supposed to be there is <code>INFO     | Systems ready<\/code>, everything else should not.<\/p>\n\n<p>Please, if someone has an idea of how to fix this once and for all, let me know below.<\/p>\n","answers":[{"AnswerId":"57546746","CreationDate":"2019-08-18T17:14:25.907","ParentId":null,"OwnerUserId":"8419797","Title":null,"Body":"<p>For anyone out there dealing with the same issue, this seems to be an issue with another library.\nThis code right here should do the trick to disable Keras' deprecation warnings and Tensorflow's ugly logging, this is still ugly so if someone has a better way to do it I'd really like to hear about it.\nThis also fixed the double logging.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import os\nimport logging\nos.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'\nimport sys\nstderr = sys.stderr\nsys.stderr = open(os.devnull, 'w')\nimport keras\nsys.stderr = stderr\nfrom keras.layers import Activation, Dense, Dropout\nfrom keras.models import Sequential, load_model\nfrom keras.optimizers import SGD\nfrom keras import regularizers\nfrom keras.backend.tensorflow_backend import tf\nimport absl.logging\nlogging.root.removeHandler(absl.logging._absl_handler)\nabsl.logging._warn_preinit_stderr = False\nlogger = tf.get_logger()\nlogger.disabled = True\nlogger.setLevel(logging.FATAL)\ngraph = tf.get_default_graph()\n<\/code><\/pre>\n"},{"AnswerId":"57719030","CreationDate":"2019-08-30T00:15:16.940","ParentId":null,"OwnerUserId":"1801968","Title":null,"Body":"<p>Ive had the same issue and am not able to resolve it<\/p>\n\n<p>Ive tried all the usual suspects and SO suggestions.<\/p>\n\n<pre><code>import pandas as pd\nimport os\nimport numpy as np\nfrom sklearn import preprocessing\nfrom sklearn.metrics import mean_squared_error, mean_absolute_error\nimport tensorflow as tf\nfrom tensorflow.python.keras.models import load_model, model_from_json\nunique_str = datetime.now().isoformat().replace(':', '_').replace('.', '_').replace('-', '_')\nLOG_FILENAME = 'tmp' + os.sep + 'logs' + os.sep + 'masterPVTrainer_'+unique_str+'.log'\nLOG_FORMAT = '%(asctime)s %(name)-12s %(levelname)-8s %(funcName)s %(lineno)5d :%(message)s'\nif __name__ == \"__main__\":\n    # eager execution does not allow tensorboard\n    # tf.enable_eager_execution()\n    # Add the log message handler to the logger\n    handler = handlers.RotatingFileHandler(LOG_FILENAME, maxBytes=5000000, backupCount=50)\n    handler.setFormatter(logging.Formatter(LOG_FORMAT, datefmt='%m-%d %H:%M:%S'))\n    handler.setLevel(logging.DEBUG)\n    # logging.getLogger('tensorflow').disabled = True\n    # tf_logger = tf.get_logger()\n    # tf_logger.setLevel(logging.DEBUG)\n    # logging.basicConfig(format=LOG_FORMAT)\n    my_logger = logging.getLogger(__name__)\n    # Add this handler to the root logger\n    my_logger.addHandler(handler)\n    my_logger.debug('debug')\n    my_logger.info('info')\n    my_logger.warning('warning')\n    my_logger.error('error')\n    main(my_logger)\n<\/code><\/pre>\n\n<p>Nothing shows up in the log file except warning and error MESSAGES. I gave up and am using <strong>warning<\/strong> and <strong>error<\/strong> for my debug messages. SUCKS MAN!<\/p>\n\n<pre><code> 08-29 13:39:28 __main__     WARNING  &lt;module&gt;   973 :warning \n 08-29 13:39:28 __main__     ERROR    &lt;module&gt;   974 :error\n<\/code><\/pre>\n\n<blockquote>\n  <p>Versions:<\/p>\n  \n  <p>Python 3.6.8 (tags\/v3.6.8:3c6b436a57, Dec 24 2018, 00:16:47) [MSC v.1916 64 bit (AMD64)] on win32<\/p>\n  \n  <p>Tensorflow version: 2.0.0-alpha0<\/p>\n  \n  <p>Keras version: 2.2.4-tf<\/p>\n  \n  <p>Pandas version: 0.24.2<\/p>\n  \n  <p>Numpy version: 1.16.4<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":57539335,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><memory><keras>","CreationDate":"2019-08-17T19:29:39.517","AcceptedAnswerId":null,"OwnerUserId":9025377.0,"Title":"How to fix \"ResourceExhaustedError: OOM ....\"?","Body":"<p>i'm trying to train a simple MLP with a own dataset in Python with Keras. This dataset includes normalized images in a size of <strong>1024 x 1204<\/strong>, i need this resolution for therefore i can't decrease the size of the images. I use a <strong>Tesla V100 with 16GB<\/strong> for the training.  <\/p>\n\n<p>My aim is first of all, that something work, before i can tune this model (make a cnn etc.), but actually it does not, because of:<\/p>\n\n<blockquote>\n  <p>ResourceExhaustedError (see above for traceback): OOM when allocating\n  tensor with shape[1048576,4096] and type float on\n  \/job:localhost\/replica:0\/task:0\/device:GPU:0 by allocator GPU_0_bfc<\/p>\n<\/blockquote>\n\n<p>This error occurs at the first layer, so before the training realy begins.<\/p>\n\n<p>I trained alreay a MLP in Julia with Flux without memory problems.<\/p>\n\n<p>Everything i tried:<\/p>\n\n<ul>\n<li>reduced the batch size<\/li>\n<li>using multiple gpus with (<a href=\"https:\/\/keras.io\/utils\/#multi_gpu_model\" rel=\"nofollow noreferrer\">keras.utils.multi_gpu_model<\/a>), the issue occurs before several gpus are in use<\/li>\n<li>reduced neurons of input layer (to reduce the weights matrix) from 1024*1024 to 4096<\/li>\n<li>set <a href=\"https:\/\/www.tensorflow.org\/guide\/using_gpu#allowing_gpu_memory_growth\" rel=\"nofollow noreferrer\">allow_growth<\/a> and also tried the <a href=\"https:\/\/www.tensorflow.org\/guide\/using_gpu#allowing_gpu_memory_growth\" rel=\"nofollow noreferrer\">per_process_gpu_memory_fraction<\/a><\/li>\n<\/ul>\n\n<p>MLP in julia (flux)<\/p>\n\n<pre><code>m = Chain(\n  Dense(1024*1024, 1024, relu),\n  Dense(1024, 256, relu),\n  Dense(256, 2),\n  softmax) |&gt; gpu\n<\/code><\/pre>\n\n<p>MLP in python (keras)<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential()\nmodel.add(Dense(4*1024, input_shape=(1024*1024,)))\nmodel.add(Dense(1024, activation='relu'))\nmodel.add(Dense(256, activation='relu'))\nmodel.add(Dense(num_classes, activation='softmax'))\n<\/code><\/pre>\n","answers":[{"AnswerId":"57542942","CreationDate":"2019-08-18T08:36:31.117","ParentId":null,"OwnerUserId":"9571777","Title":null,"Body":"<p>There are three main things playing in this situation: the size of your inputs, the size of your network (number of parameters), and the GPU ram. Sadly those are considered large inputs, so your first bet is going to be turning down batch size. take it to 1, and if you still get this, either reduce your network or (more effectively) reduce your input sizes. Yes, it means information is lost from the signal prior to processing, but you'd be amazed what a network can infer from smaller resolution images. I understand not wanting to give up that information, though, so play with batch size first. for example, 762x762x3 inputs can run on Xception on 16gb with batch size 6, but not batch size 8.\nas for network architecture, those dense nets output sizes are huge. probably unnecessarily so. yeah so maybe crank that down a bit. and, like Matias said, reduction through convolution highly recommend.<\/p>\n"}]}
+{"QuestionId":57539456,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-08-17T19:49:00.080","AcceptedAnswerId":null,"OwnerUserId":9904551.0,"Title":"Installing PlaidML Keras","Body":"<p>I am trying to install PlaidML and am following the instructions on the Github. Essentially I am getting the same area as the user here:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/56973032\/how-to-install-plaidml-plaidml-keras\">How to install plaidML \/ plaidML-keras<\/a><\/p>\n\n<p>I followed the instructions from the developer who responded to that question but still plaidml-setup does not work. <\/p>\n\n<p>I created a conda environment, and installed PlaidML with pip3. The path is<\/p>\n\n<pre><code>\/usr\/local\/lib\/python3.7\/site-packages (0.6.4). \n<\/code><\/pre>\n\n<p>The error running plaidml-setup is:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/usr\/local\/bin\/plaidml-setup\", line 6, in &lt;module&gt;\nfrom plaidml.plaidml_setup import main\n\n  File \"\/usr\/local\/lib\/python3.7\/site-packages\/plaidml\/__init__.py\", line 50, in &lt;module&gt;\nimport plaidml.settings\n\n  File \"\/usr\/local\/lib\/python3.7\/site-packages\/plaidml\/settings.py\", line 33, in &lt;module&gt;\n_setup_config('PLAIDML_EXPERIMENTAL_CONFIG', 'experimental.json')\n\n  File \"\/usr\/local\/lib\/python3.7\/site-packages\/plaidml\/settings.py\", line 30, in _setup_config\n    'Could not find PlaidML configuration file: \"{}\".'.format(filename))\n\nplaidml.exceptions.PlaidMLError: Could not find PlaidML configuration file: \"experimental.json\".\n<\/code><\/pre>\n\n<p>Following the developers instructions, I navigatated from my virtual environment to the below, and found the following files. <\/p>\n\n<pre><code>(tensorflow) \u279c  ~ cd \/usr\/local\/share\/plaidml                                               \n(tensorflow) \u279c  plaidml \n(tensorflow) \u279c  plaidml ls\nconfig.json          experimental.json    plaidml-config.cmake\n<\/code><\/pre>\n","answers":[{"AnswerId":"57543630","CreationDate":"2019-08-18T10:23:53.033","ParentId":null,"OwnerUserId":"9904551","Title":null,"Body":"<p>I simply forgot to do pip3 install keras followed by pip3 install tensorflow after creating my virtual environment. <\/p>\n"}]}
+{"QuestionId":57539524,"AnswerCount":0,"Tags":"<keras><tensorflow2.0>","CreationDate":"2019-08-17T19:57:48.623","AcceptedAnswerId":null,"OwnerUserId":11940808.0,"Title":"How to make a prediction with a trained model that is using a feature layer?","Body":"<p>I have a trained model that uses a feature layer for its inputs.\nI am happy with the training and performance of the model on the validation and test sets.<\/p>\n\n<p>But now, I want to create manually a new example to send to the model and get a prediction out of it.<\/p>\n\n<p>My code looks like the following:<\/p>\n\n<p>...<\/p>\n\n<pre><code>categorical_columns = []\nnumeric_columns = []\n\n# numeric cols\nfor header in ['CreditScore', 'Age', 'Tenure', 'Balance', 'NumOfProducts', 'HasCrCard','IsActiveMember','EstimatedSalary']:\n  numeric_columns.append(feature_column.numeric_column(header))\n\n# categorical cols\ncategories = {\n                'Geography':['France', 'Spain', 'Germany'],\n                'Gender': ['Female', 'Male']\n             }\n\nfor header, possible_values in categories.items():\n  cat_col = tf.feature_column.categorical_column_with_vocabulary_list(key=header, vocabulary_list=possible_values)\n  categorical_columns.append(tf.feature_column.indicator_column(cat_col))\n\nfeature_columns =  numeric_columns + categorical_columns\n\nfeature_layer = tf.keras.layers.DenseFeatures(feature_columns)\n\nmodel = tf.keras.Sequential([\n  feature_layer,\n  layers.Dense(6, activation='relu'),\n  layers.Dense(6, activation='relu'),\n  layers.Dense(1, activation='sigmoid')\n])\n\nmodel.compile(optimizer='adam',\n              loss='binary_crossentropy',\n              metrics=['accuracy'],\n              run_eagerly=True)\n\nhistory = model.fit(x=train_ds, validation_data=val_ds, epochs=4)\n<\/code><\/pre>\n\n<p>Not worrying about feature scaling for this question, how would I send this data to the model and get a prediction out of it?\nWhat syntax should I use?<\/p>\n\n<ul>\n<li>CreditScore:  358<\/li>\n<li>Age:  42<\/li>\n<li>Tenure:   5<\/li>\n<li>Balance:  4000<\/li>\n<li>NumOfProducts:    5<\/li>\n<li>HasCrCard 1:<\/li>\n<li>IsActiveMember:   1<\/li>\n<li>EstimatedSalary:  50000<\/li>\n<li>Geography:    France<\/li>\n<li>Gender:   Male<\/li>\n<\/ul>\n\n<p>Thank you for your help.<\/p>\n","answers":[]}
+{"QuestionId":57539567,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch>","CreationDate":"2019-08-17T20:04:28.333","AcceptedAnswerId":57539790.0,"OwnerUserId":11940717.0,"Title":"Augmenting only the training set in K-folds cross validation","Body":"<p>I am trying to create a binary CNN classifier for an unbalanced dataset (class 0 = 4000 images, class 1 = around 250 images), which I want to perform 5-fold cross validation on. Currently I am loading my training set into an ImageLoader that applies my transformations\/augmentations(?) and loads it into a DataLoader. However, this results in both my training splits and validation splits containing the augmented data. <\/p>\n\n<p>I originally applied transformations offline (offline augmentation?) to balance my dataset, but from this thread (<a href=\"https:\/\/stats.stackexchange.com\/questions\/175504\/how-to-do-data-augmentation-and-train-validate-split\">https:\/\/stats.stackexchange.com\/questions\/175504\/how-to-do-data-augmentation-and-train-validate-split<\/a>), it seems it would be ideal to only augment the training set. I would also prefer to train my model on solely augmented training data and then validate it on non-augmented data in a 5-fold cross validation<\/p>\n\n<p>My data is organized as root\/label\/images, where there are 2 label folders (0 and 1) and images sorted into the respective labels.<\/p>\n\n<h1>My Code So Far<\/h1>\n\n<pre><code>total_set = datasets.ImageFolder(ROOT, transform = data_transforms['my_transforms'])\n\n\/\/Eventually I plan to run cross-validation as such:\nsplits = KFold(cv = 5, shuffle = True, random_state = 42)\n\nfor train_idx, valid_idx in splits.split(total_set):\n    train_sampler = SubsetRandomSampler(train_idx)\n    valid_sampler = SubsetRandomSampler(valid_idx)\n\n    train_loader = torch.utils.data.DataLoader(total_set, batch_size=32, sampler=train_sampler)\n    val_loader = torch.utils.data.DataLoader(total_set, batch_size=32, sampler=valid_sampler)\n\nmodel.train()\n\/\/Model train\/eval works but may be overpredict \n<\/code><\/pre>\n\n<p>I'm sure I'm doing something sub-optimally or wrong in this code, but I can't seem to find any documentation on specifically augmenting only the training splits in cross-validation!<\/p>\n\n<p>Any help would be appreciated!<\/p>\n","answers":[{"AnswerId":"57539790","CreationDate":"2019-08-17T20:40:56.510","ParentId":null,"OwnerUserId":"2790047","Title":null,"Body":"<p>One approach is to implement a wrapper Dataset class that applies transforms to the output of your ImageFolder dataset. For example<\/p>\n\n<pre><code>class WrapperDataset:\n    def __init__(self, dataset, transform=None, target_transform=None):\n        self.dataset = dataset\n        self.transform = transform\n        self.target_transform = target_transform\n\n    def __getitem__(self, index):\n        image, label = self.dataset[index]\n        if self.transform is not None:\n            image = self.transform(image)\n        if self.target_transform is not None:\n            label = self.target_transform(label)\n        return image, label\n\n    def __len__(self):\n        return len(self.dataset)\n<\/code><\/pre>\n\n<p>Then you could use this in your code by wrapping the larger dataset with different transforms.<\/p>\n\n<pre><code>total_set = datasets.ImageFolder(ROOT)\n\n\/\/Eventually I plan to run cross-validation as such:\nsplits = KFold(cv = 5, shuffle = True, random_state = 42)\n\nfor train_idx, valid_idx in splits.split(total_set):\n    train_sampler = SubsetRandomSampler(train_idx)\n    valid_sampler = SubsetRandomSampler(valid_idx)\n\n    train_loader = torch.utils.data.DataLoader(\n        WrapperDataset(total_set, transform=data_transforms['train_transforms']),\n        batch_size=32, sampler=train_sampler)\n    valid_loader = torch.utils.data.DataLoader(\n        WrapperDataset(total_set, transform=data_transforms['valid_transforms']),\n        batch_size=32, sampler=valid_sampler)\n\n    # train\/validate now\n<\/code><\/pre>\n\n<p>I haven't tested this code since I don't have your full code\/models but the concept should be clear.<\/p>\n"}]}
+{"QuestionId":57539660,"AnswerCount":0,"Tags":"<python><keras><lstm><recurrent-neural-network>","CreationDate":"2019-08-17T20:16:58.927","AcceptedAnswerId":null,"OwnerUserId":3757672.0,"Title":"Confusion about Hidden Layers in LSTMs\/GRUs in Keras","Body":"<p>I'm aware of questions like <a href=\"https:\/\/stackoverflow.com\/questions\/45328843\/keras-confusion-about-number-of-layers\">here<\/a> or <a href=\"https:\/\/stackoverflow.com\/questions\/54394026\/keras-clarification-on-definition-of-hidden-layer\">here<\/a>, but they cannot clarify the following for me.<\/p>\n\n<p>Given some random LSTM-Model like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential()\nmodel.add(LSTM(12, input_dim=8, activation='relu')) # Input and Hidden 1\nmodel.add(LSTM(8)) # Hidden 2\nmodel.add(LSTM(8)) # Hidden 3\nmodel.add(Dense(1)) # Output\n<\/code><\/pre>\n\n<p>Reading about a setup like this, for exmaple on SO, it is often said that we have a network with an input-layer, three hidden-layers and an output-layer.<\/p>\n\n<p>Anyhow, what I am confused about: Isn't this setup a stacking of three LSTMs leading to a <em>deep-LSTM<\/em> rather than three hidden-layers?<br>This means any of these three LSTMs should have the option to set a variable kind of hidden layers within itself (obviously in Keras it has not). <\/p>\n\n<p>Is Keras just calling these layers <em>LSTMs<\/em> for convenience or am I misunderstanding the concept of LSTM-stacking? Or is a LSTM-cell just a form of a hidden-layer?<\/p>\n","answers":[]}
+{"QuestionId":57539708,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2019-08-17T20:25:35.843","AcceptedAnswerId":null,"OwnerUserId":8935585.0,"Title":"How to use tf.gather to collect values?","Body":"<p>I have a tensor 'a'  with shape (None, 85,1500). And index tensor 'b' with shape (None, 5). The index tensor is to select value in 'a' on axis=1 (which value is 85). The result would be expected in shape of (None, 5,1500). <\/p>\n\n<p>But when I use <code>tf.gather (a,b)<\/code>. I got shape of (None, 5,85,1500). <\/p>\n\n<p>When I use <code>tf.gather (a,b,axis=1)<\/code>, I got (None, None, 5,1500)<\/p>\n\n<p>When I use <code>tf.gather_nd(a,b,)<\/code>, I got an error : <\/p>\n\n<pre><code>ValueError: indices.shape[-1] must be &lt;= params.rank, but saw indices shape: [?,5] and params shape: [?,85,1500] for 'encoder\/GatherNd' (op: 'GatherNd') with input shapes: [?,85,1500], [?,5].\n<\/code><\/pre>\n\n<p>Could anyone help me\uff1f Thank you.<\/p>\n","answers":[]}
+{"QuestionId":57539872,"AnswerCount":1,"Tags":"<pytorch><word-embedding>","CreationDate":"2019-08-17T20:55:15.690","AcceptedAnswerId":57541803.0,"OwnerUserId":4851368.0,"Title":"RuntimeError: Expected object of backend CUDA but got backend CPU for argument #3 'index'","Body":"<p>I'm working with the project 'lda2vec-pytorch' on Google CoLab,\nrunnin pytorch 1.1.0<\/p>\n\n<p><a href=\"https:\/\/github.com\/TropComplique\/lda2vec-pytorch\" rel=\"nofollow noreferrer\">https:\/\/github.com\/TropComplique\/lda2vec-pytorch<\/a><\/p>\n\n<pre><code>device = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\ncuda:0\n<\/code><\/pre>\n\n<p>I'm getting an exception in the forward method adding 'noise' in my \nclass negative_sampling_loss(nn.Module):<\/p>\n\n<pre><code>        noise = self.multinomial.draw(batch_size*window_size*self.num_sampled)\n        noise = Variable(noise).view(batch_size, window_size*self.num_sampled)\n\n        device = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n        self.embedding = self.embedding.to(device)\n        #print(\"negative_sampling_loss::forward() self.embedding\", self.embedding.is_cuda)  This line get's an error.\n\n        # shape: [batch_size, window_size*num_sampled, embedding_dim]\n        noise = self.embedding(noise)  # Exception HERE\n<\/code><\/pre>\n\n<p>Here's the stack trace:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"train.py\", line 36, in &lt;module&gt;\n    main()\n  File \"train.py\", line 32, in main\n    save_every=20, grad_clip=5.0\n  File \"..\/utils\/training.py\", line 138, in train\n    neg_loss, dirichlet_loss = model(doc_indices, pivot_words, target_words)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/nn\/modules\/module.py\", line 493, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"..\/utils\/lda2vec_loss.py\", line 82, in forward\n    neg_loss = self.neg(pivot_words, target_words, doc_vectors, w)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/nn\/modules\/module.py\", line 493, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"..\/utils\/lda2vec_loss.py\", line 167, in forward\n    noise = self.embedding(noise)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/nn\/modules\/module.py\", line 493, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/nn\/modules\/sparse.py\", line 117, in forward\n    self.norm_type, self.scale_grad_by_freq, self.sparse)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/nn\/functional.py\", line 1506, in embedding\n    return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)\nRuntimeError: Expected object of backend CUDA but got backend CPU for argument #3 'index'\n<\/code><\/pre>\n\n<p>Any ideas?<\/p>\n","answers":[{"AnswerId":"57541803","CreationDate":"2019-08-18T05:08:19.440","ParentId":null,"OwnerUserId":"5027315","Title":null,"Body":"<p>Variable <code>noise<\/code> is available on CPU while <code>self.embedding<\/code> is on GPU. We can send <code>noise<\/code> to GPU as well:<\/p>\n\n<pre><code>noise = noise.to(device)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57539985,"AnswerCount":1,"Tags":"<python-3.x><machine-learning><optimization><keras><deep-learning>","CreationDate":"2019-08-17T21:16:07.953","AcceptedAnswerId":null,"OwnerUserId":5432071.0,"Title":"Does Keras support parametric optimizers?","Body":"<p>When using keras can you add a term to an optimizers update rule, that the <strong>optimizer<\/strong> learns? So for instance on line 200 of keras's optimizer code on their github (<a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/optimizers.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/optimizers.py<\/a>) there is an update rule for SGD. Could you add a simple scalar variable to the SGD optimizer that would change the update formula to <\/p>\n\n<p>new_p = (p + self.momentum * v - lr * g)*new_variable<\/p>\n\n<p>where the new_variable is a scalar that is bounded within a certain range (say 0 to 1) and the SGD optimizer has to LEARN the appropriate value for new_variable?<\/p>\n\n<p>How would a person go about making these kind of changes to an optimizer?<\/p>\n","answers":[{"AnswerId":"57540871","CreationDate":"2019-08-18T00:36:19.080","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>That would sound weird, because for learning it's necessary to \"get the gradients of the loss function\" with respect to the weights. (This variable would be a weight, in the sense that it's a trainable var in Keras)   <\/p>\n\n<p>But notice that this var you proposed:<\/p>\n\n<ul>\n<li>does not participate in the loss function    <\/li>\n<li>is used after the gradients (<code>g<\/code> in that formula) were calculated    <\/li>\n<\/ul>\n\n<p>So, independently from Keras, this kind of approach (learning the learning rate in a \"backpropagation\" learning framework) is sort of impossible. <\/p>\n"}]}
+{"QuestionId":57540263,"AnswerCount":0,"Tags":"<python><cmake><caffe><google-colaboratory><openpose>","CreationDate":"2019-08-17T22:07:46.473","AcceptedAnswerId":null,"OwnerUserId":10811086.0,"Title":"CMake Error: failed to create symbolic link '..\/..\/lib\/libcaffe.so': function not implemented","Body":"<p>I am trying to build openpose on google colab, and I have Cmake configuration was done, and when it comes to make, I used <\/p>\n\n<pre><code>make -j4 all \n<\/code><\/pre>\n\n<p>it popped an error, here is the error:<\/p>\n\n<pre><code>[ 84%] Linking CXX shared library ..\/..\/lib\/libcaffe.so\nCMake Error: failed to create symbolic link '..\/..\/lib\/libcaffe.so':  function not implemented\nCMake Error: cmake_symlink_library: System Error: Function not  implemented\nsrc\/caffe\/CMakeFiles\/caffe.dir\/build.make:2215: recipe for target 'lib\/libcaffe.so.1.0.0' failed\nmake[5]: *** [lib\/libcaffe.so.1.0.0] Error 1\nmake[5]: *** Deleting file 'lib\/libcaffe.so.1.0.0'\nCMakeFiles\/Makefile2:240: recipe for target 'src\/caffe\/CMakeFiles\/caffe.dir\/all' failed\nmake[4]: *** [src\/caffe\/CMakeFiles\/caffe.dir\/all] Error 2\nMakefile:129: recipe for target 'all' failed\nmake[3]: *** [all] Error 2\nCMakeFiles\/openpose_lib.dir\/build.make:110: recipe for target 'caffe\/src\/openpose_lib-stamp\/openpose_lib-build' failed\nmake[2]: *** [caffe\/src\/openpose_lib-stamp\/openpose_lib-build] Error 2\nCMakeFiles\/Makefile2:72: recipe for target   'CMakeFiles\/openpose_lib.dir\/all' failed\nmake[1]: *** [CMakeFiles\/openpose_lib.dir\/all] Error 2\nMakefile:83: recipe for target 'all' failed\nmake: *** [all] Error 2\n<\/code><\/pre>\n\n<p>How to solve this problem?<\/p>\n","answers":[]}
+{"QuestionId":57540443,"AnswerCount":1,"Tags":"<javascript><python><tensorflow><webpack><tensorflow.js>","CreationDate":"2019-08-17T22:46:25.857","AcceptedAnswerId":null,"OwnerUserId":4896449.0,"Title":"Loading python TensorFlow Layers models into JavaScript","Body":"<p>The core <code>TensorFlow<\/code> libraries provide the ability to convert a model created in <code>Python<\/code> to be saved into a <code>JSON<\/code> file describing the graph and weights to be executed in a browser environment.<\/p>\n\n<p>In the examples you are required to load the whole <code>TensorFlow<\/code> library in the browser, which is extremely heavy. Also tree shaking is not available for this library.<\/p>\n\n<p>My question is: how would I load just the necessary elements from <code>TensorFlow JS<\/code> into the client\/browser to reduce the overall size of my bundled application?<\/p>\n\n<p>EDIT: We are trying to reduce the over size of the bundled library.<\/p>\n","answers":[{"AnswerId":"57560165","CreationDate":"2019-08-19T15:52:30.403","ParentId":null,"OwnerUserId":"5627599","Title":null,"Body":"<p>The Tensorflow.js API combines four packages:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/tensorflow\/tfjs\/tree\/master\/tfjs-core\" rel=\"nofollow noreferrer\">tfjs-core<\/a>: Functionality like mathematical functions and backend support<\/li>\n<li><a href=\"https:\/\/github.com\/tensorflow\/tfjs\/tree\/master\/tfjs-layers\" rel=\"nofollow noreferrer\">tfjs-layers<\/a>: Support of layers to create models (depends on <code>tfjs-core<\/code>)<\/li>\n<li><a href=\"https:\/\/github.com\/tensorflow\/tfjs\/tree\/master\/tfjs-data\" rel=\"nofollow noreferrer\">tfjs-data<\/a>: Data handling (depends on <code>tfjs-core<\/code>)<\/li>\n<li><a href=\"https:\/\/github.com\/tensorflow\/tfjs\/tree\/master\/tfjs-converter\" rel=\"nofollow noreferrer\">tfjs-converter<\/a>: Support for converting models to Tensorflow.js<\/li>\n<\/ul>\n\n<p>Depending on what tasks exactly you need to perform, it might be sufficent to only use some of the packages. That said, keep in mind that <code>tfjs-layers<\/code> and <code>tfjs-data<\/code> require <code>tfjs-core<\/code> to be imported.<\/p>\n\n<p><strong>Code Sample<\/strong><\/p>\n\n<p>The following lines only import the Core and Layers API:<\/p>\n\n<pre class=\"lang-js prettyprint-override\"><code>import * as tfc from '@tensorflow\/tfjs-core';\nimport * as tfl from '@tensorflow\/tfjs-layers';\n\n\/\/ Examples how to use the APIs\nconst vectr = tfc.tensor(\/* .. *\/);\nconst model = tfl.sequential();\nconst dense = tfl.layers.dense(\/* .. *\/);\n<\/code><\/pre>\n\n<p>Note that functions like <a href=\"https:\/\/js.tensorflow.org\/api\/latest\/#matMul\" rel=\"nofollow noreferrer\"><code>tf.matMul<\/code><\/a> are used by calling <code>tfc.matMul<\/code>, but some functions of the layers API (like <a href=\"https:\/\/js.tensorflow.org\/api\/latest\/#layers.dense\" rel=\"nofollow noreferrer\"><code>tf.layers.dense<\/code><\/a>) are used by calling <code>tfl.layers.dense<\/code> while others (like <a href=\"https:\/\/js.tensorflow.org\/api\/latest\/#sequential\" rel=\"nofollow noreferrer\"><code>tf.sequential<\/code><\/a>) are used by calling <code>tfl.sequential<\/code>.<\/p>\n\n<p><strong>Optimization<\/strong><\/p>\n\n<p>To give you an idea on the potential optimization, let's look at the numbers:<\/p>\n\n<pre><code>--------------------------------------\n|     Package    |  Size  | Relative |\n|----------------|--------|----------|\n| tfjs           |    856 |     100% |\n| tfjs-core      |    506 |      59% |\n| tfjs-layers    |    228 |      27% |\n| tfjs-data      |     52 |       6% |\n| tfjs-converter |     80 |       9% |\n--------------------------------------\n<\/code><\/pre>\n\n<p><sub>Version 1.2.7, Size in KB (of the minified JS file), relative values compared to tfjs<\/sub><\/p>\n\n<p>Using <code>tfjs-core<\/code> and <code>tfjs-layers<\/code> directly, it is possible to shrink the size by 122 KB or 14%. If you need more than that you can always try to rebuild the repository on your own, removing any unneeded functionality. Of course, this approach would mean a lot of manual work.<\/p>\n\n<p><strong>Tree Shaking<\/strong><\/p>\n\n<p>As you already noticed yourself, tree shaking is currently not supported, but you might want to follow the discussion for <a href=\"https:\/\/github.com\/tensorflow\/tfjs\/issues\/353\" rel=\"nofollow noreferrer\">support of tree-shaking<\/a> in the tfjs github repository regarding that topic.<\/p>\n"}]}
+{"QuestionId":57540454,"AnswerCount":0,"Tags":"<python><deep-learning><pytorch><resnet><unity3d-unet>","CreationDate":"2019-08-17T22:48:58.317","AcceptedAnswerId":null,"OwnerUserId":1892020.0,"Title":"Pytorch Implementation for ResNet Based UNet","Body":"<p>I want to implement a ResNet based UNet for segmentation (without pre-training). I have referred to this implementation using Keras but my project has been implemented using PyTorch that I am not sure if I have done the correct things.<\/p>\n\n<p>Keras based implementation\n<a href=\"https:\/\/www.kaggle.com\/shaojiaxin\/u-net-with-simple-resnet-blocks\" rel=\"nofollow noreferrer\">U-net with simple Resnet Blocks<\/a><\/p>\n\n<p>My PyTorch implementation (I am not sure if I am correct ...) <strong>Any suggestions will be highly appreciated<\/strong>.<\/p>\n\n<pre><code>def double_conv(in_channels, out_channels):\nreturn nn.Sequential(\n    nn.Conv2d(in_channels, out_channels, 3, padding=1),\n    nn.BatchNorm2d(out_channels), \n    nn.ReLU(inplace=True),\n    nn.Conv2d(out_channels, out_channels, 3, padding=1),\n    nn.BatchNorm2d(out_channels), \n    nn.ReLU(inplace=True)\n)\n\ndef root_block(in_channels, out_channels):\nreturn nn.Sequential(\n    nn.Conv2d(in_channels, out_channels, 3, padding=1),\n    nn.BatchNorm2d(out_channels), \n    nn.ReLU(inplace=True),\n    nn.Conv2d(out_channels, out_channels, 3, padding=1),\n    nn.BatchNorm2d(out_channels), \n)\n\n# Define the UNet architecture\nclass ResNetUNet(nn.Module):\n\ndef __init__(self, n_class):\n    super().__init__()\n\n    self.dconv_down1 = double_conv(3, 64)\n    self.dconv_down11 = root_block(64, 64)\n    self.dconv_down2 = double_conv(64, 128)\n    self.dconv_down21 = root_block(128, 128)\n    self.dconv_down3 = double_conv(128, 256)\n    self.dconv_down31 = root_block(256, 256)\n    self.dconv_down4 = double_conv(256, 512)\n    self.dconv_down41 = root_block(512, 512)\n\n    self.maxpool = nn.MaxPool2d(2)\n    self.relu = nn.ReLU(inplace=True)\n\n    self.dconv_up3 = double_conv(256 + 512, 256)\n    self.dconv_up31 = root_block(256, 256)\n    self.dconv_up2 = double_conv(128 + 256, 128)\n    self.dconv_up21 = root_block(128, 128)\n    self.dconv_up1 = double_conv(128 + 64, 64)\n    self.dconv_up11 = root_block(64, 64)\n\n    self.conv_last = nn.Conv2d(64, n_class, 1)\n\ndef forward(self, x):\n    conv1 = self.dconv_down1(x)\n    x = self.dconv_down11(conv1)\n    x += conv1\n    x = self.relu(x)\n    x = self.maxpool(x)\n\n    conv2 = self.dconv_down2(x)\n    x = self.dconv_down21(conv2)\n    x += conv2\n    x = self.relu(x)\n    x = self.maxpool(x)\n\n    conv3 = self.dconv_down3(x)\n    x = self.dconv_down31(conv3)\n    x += conv3\n    x = self.relu(x)\n    x = self.maxpool(x)\n\n    conv4 = self.dconv_down4(x)\n    x = self.dconv_down41(conv4)\n    x += conv4\n    x = self.relu(x)\n\n    deconv3 = nn.functional.interpolate(x, scale_factor=2, mode='bilinear', align_corners=True)\n    deconv3 = torch.cat([deconv3, conv3], dim=1)\n    uconv3 = self.dconv_up3(deconv3)\n    x = self.dconv_up31(uconv3)\n    x += uconv3\n    x = self.relu(x)\n\n    deconv2 = nn.functional.interpolate(x, scale_factor=2, mode='bilinear', align_corners=True)\n    deconv2 = torch.cat([deconv2, conv2], dim=1)\n    uconv2 = self.dconv_up2(deconv2)\n    x = self.dconv_up21(uconv2)\n    x += uconv2\n    x = self.relu(x)\n\n    deconv1 = nn.functional.interpolate(x, scale_factor=2, mode='bilinear', align_corners=True)\n    deconv1 = torch.cat([deconv1, conv1], dim=1)\n    uconv1 = self.dconv_up1(deconv1)\n    x = self.dconv_up11(uconv1)\n    x += uconv1\n    x = self.relu(x)\n\n    out = self.conv_last(x)\n\n    return out\n<\/code><\/pre>\n\n<p>Keras based implementation extracted as below:<\/p>\n\n<pre><code>def convolution_block(x, filters, size, strides=(1,1), padding='same', activation=True):\nx = Conv2D(filters, size, strides=strides, padding=padding)(x)\nx = BatchNormalization()(x)\nif activation == True:\n    x = Activation('relu')(x)\nreturn x\n\ndef residual_block(blockInput, num_filters=16):\nx = Activation('relu')(blockInput)\nx = BatchNormalization()(x)\nx = convolution_block(x, num_filters, (3,3) )\nx = convolution_block(x, num_filters, (3,3), activation=False)\nx = Add()([x, blockInput])\nreturn x\n\n# Build model\ndef build_model(input_layer, start_neurons, DropoutRatio = 0.5):\n# 101 -&gt; 50\nconv1 = Conv2D(start_neurons * 1, (3, 3), activation=None, padding=\"same\")(input_layer)\nconv1 = residual_block(conv1,start_neurons * 1)\nconv1 = residual_block(conv1,start_neurons * 1)\nconv1 = Activation('relu')(conv1)\npool1 = MaxPooling2D((2, 2))(conv1)\npool1 = Dropout(DropoutRatio\/2)(pool1)\n\n# 50 -&gt; 25\nconv2 = Conv2D(start_neurons * 2, (3, 3), activation=None, padding=\"same\")(pool1)\nconv2 = residual_block(conv2,start_neurons * 2)\nconv2 = residual_block(conv2,start_neurons * 2)\nconv2 = Activation('relu')(conv2)\npool2 = MaxPooling2D((2, 2))(conv2)\npool2 = Dropout(DropoutRatio)(pool2)\n\n# 25 -&gt; 12\nconv3 = Conv2D(start_neurons * 4, (3, 3), activation=None, padding=\"same\")(pool2)\nconv3 = residual_block(conv3,start_neurons * 4)\nconv3 = residual_block(conv3,start_neurons * 4)\nconv3 = Activation('relu')(conv3)\npool3 = MaxPooling2D((2, 2))(conv3)\npool3 = Dropout(DropoutRatio)(pool3)\n\n# 12 -&gt; 6\nconv4 = Conv2D(start_neurons * 8, (3, 3), activation=None, padding=\"same\")(pool3)\nconv4 = residual_block(conv4,start_neurons * 8)\nconv4 = residual_block(conv4,start_neurons * 8)\nconv4 = Activation('relu')(conv4)\npool4 = MaxPooling2D((2, 2))(conv4)\npool4 = Dropout(DropoutRatio)(pool4)\n\n# Middle\nconvm = Conv2D(start_neurons * 16, (3, 3), activation=None, padding=\"same\")(pool4)\nconvm = residual_block(convm,start_neurons * 16)\nconvm = residual_block(convm,start_neurons * 16)\nconvm = Activation('relu')(convm)\n\n# 6 -&gt; 12\ndeconv4 = Conv2DTranspose(start_neurons * 8, (3, 3), strides=(2, 2), padding=\"same\")(convm)\nuconv4 = concatenate([deconv4, conv4])\nuconv4 = Dropout(DropoutRatio)(uconv4)\n\nuconv4 = Conv2D(start_neurons * 8, (3, 3), activation=None, padding=\"same\")(uconv4)\nuconv4 = residual_block(uconv4,start_neurons * 8)\nuconv4 = residual_block(uconv4,start_neurons * 8)\nuconv4 = Activation('relu')(uconv4)\n\n# 12 -&gt; 25\n#deconv3 = Conv2DTranspose(start_neurons * 4, (3, 3), strides=(2, 2), padding=\"same\")(uconv4)\ndeconv3 = Conv2DTranspose(start_neurons * 4, (3, 3), strides=(2, 2), padding=\"valid\")(uconv4)\nuconv3 = concatenate([deconv3, conv3])    \nuconv3 = Dropout(DropoutRatio)(uconv3)\n\nuconv3 = Conv2D(start_neurons * 4, (3, 3), activation=None, padding=\"same\")(uconv3)\nuconv3 = residual_block(uconv3,start_neurons * 4)\nuconv3 = residual_block(uconv3,start_neurons * 4)\nuconv3 = Activation('relu')(uconv3)\n\n# 25 -&gt; 50\ndeconv2 = Conv2DTranspose(start_neurons * 2, (3, 3), strides=(2, 2), padding=\"same\")(uconv3)\nuconv2 = concatenate([deconv2, conv2])\n\nuconv2 = Dropout(DropoutRatio)(uconv2)\nuconv2 = Conv2D(start_neurons * 2, (3, 3), activation=None, padding=\"same\")(uconv2)\nuconv2 = residual_block(uconv2,start_neurons * 2)\nuconv2 = residual_block(uconv2,start_neurons * 2)\nuconv2 = Activation('relu')(uconv2)\n\n# 50 -&gt; 101\n#deconv1 = Conv2DTranspose(start_neurons * 1, (3, 3), strides=(2, 2), padding=\"same\")(uconv2)\ndeconv1 = Conv2DTranspose(start_neurons * 1, (3, 3), strides=(2, 2), padding=\"valid\")(uconv2)\nuconv1 = concatenate([deconv1, conv1])\n\nuconv1 = Dropout(DropoutRatio)(uconv1)\nuconv1 = Conv2D(start_neurons * 1, (3, 3), activation=None, padding=\"same\")(uconv1)\nuconv1 = residual_block(uconv1,start_neurons * 1)\nuconv1 = residual_block(uconv1,start_neurons * 1)\nuconv1 = Activation('relu')(uconv1)\n\nuconv1 = Dropout(DropoutRatio\/2)(uconv1)\noutput_layer = Conv2D(1, (1,1), padding=\"same\", activation=\"sigmoid\")(uconv1)\n\nreturn output_layer\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57540570,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-17T23:15:54.547","AcceptedAnswerId":57542219.0,"OwnerUserId":11876924.0,"Title":"Keras IMDB dataset 0 vs ovv_char","Body":"<p>I am looking at the documentation for the <a href=\"https:\/\/keras.io\/datasets\/#imdb-movie-reviews-sentiment-classification\" rel=\"nofollow noreferrer\">Keras preprocessed IMDB dataset<\/a>, which says:<\/p>\n\n<blockquote>\n  <p>As a convention, \"0\" does not stand for a specific word, but instead is used to encode any unknown word.<\/p>\n<\/blockquote>\n\n<p>However, I am not seeing \"0\" in the data when I try:<\/p>\n\n<pre><code>(inputTrain, yTrain), (inputTest, yTest) = imdb.load_data(num_words=vocabSize)\n<\/code><\/pre>\n\n<p>Am I missing something? I also don't see \"0\" being added in the preprocessing code <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/datasets\/imdb.py\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n","answers":[{"AnswerId":"57542219","CreationDate":"2019-08-18T06:33:38.247","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<blockquote>\n  <p>As a convention, \"0\" does not stand for a specific word, but instead is used to encode any unknown word.<\/p>\n<\/blockquote>\n\n<p>The documentation is wrong here. In Keras IMDB dataset, the value:<\/p>\n\n<ul>\n<li>\"0\" stands for padding (i.e. to make all the sequences have the same lengths as needed),<\/li>\n<li>\"1\" stands for start token, i.e. denoting the start of a document,<\/li>\n<li>\"2\" stands for OOV (Out-Of-Vocabulary) token, i.e. representing all the tokens which are not in the dictionary or, as used in the documentation, <em>unknown word<\/em>.<\/li>\n<\/ul>\n\n<p>So, in Keras IMDB dataset, \"2\" is used for unknown words, and not \"0\".<\/p>\n"}]}
+{"QuestionId":57540745,"AnswerCount":2,"Tags":"<pytorch>","CreationDate":"2019-08-18T00:04:31.687","AcceptedAnswerId":null,"OwnerUserId":10855529.0,"Title":"What is the difference between register_parameter and register_buffer in PyTorch?","Body":"<p>parameter gets changed during training, that is they are what are learnt during training of a neural network, but what is a buffer?<\/p>\n\n<p>and is it learnt during neural network training?<\/p>\n","answers":[{"AnswerId":"57541778","CreationDate":"2019-08-18T05:01:29.623","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Pytorch <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/nn.html#torch.nn.Module.register_buffer\" rel=\"noreferrer\">doc<\/a> for <code>register_buffer()<\/code> method reads<\/p>\n\n<blockquote>\n  <p>This is typically used to register a buffer that should not to be considered a model parameter. For example, BatchNorm\u2019s <code>running_mean<\/code> is not a parameter, but is part of the persistent state.<\/p>\n<\/blockquote>\n\n<p>As you already observed, model <em>parameters<\/em> are learned and updated using SGD during the training process.<br>\nHowever, sometimes there are other quantities that are part of a model's \"state\" and should be<br>\n - saved as part of <code>state_dict<\/code>.<br>\n - moved to <code>cuda()<\/code> or <code>cpu()<\/code> with the rest of the model's parameters.<br>\n - cast to <code>float<\/code>\/<code>half<\/code>\/<code>double<\/code> with the rest of the model's parameters.<br>\nRegistering these \"arguments\" as the model's <code>buffer<\/code> allows pytorch to track them and save them like regular parameters, but prevents pytorch from updating them using SGD mechanism.<\/p>\n\n<p>An example for a buffer can be found in <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/_modules\/torch\/nn\/modules\/batchnorm.html#BatchNorm2d\" rel=\"noreferrer\"><code>_BatchNorm<\/code><\/a> module where the <code>running_mean<\/code> , <code>running_var<\/code> and <code>num_batches_tracked<\/code> are registered as buffers and updated by accumulating statistics of data forwarded through the layer. This is in contrast to <code>weight<\/code> and <code>bias<\/code> parameters that learns an affine transformation of the data using regular SGD optimization.<\/p>\n"},{"AnswerId":"57546078","CreationDate":"2019-08-18T15:42:17.067","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>Both parameters and buffers you create for a module (<code>nn.Module<\/code>).<\/p>\n\n<p>Say you have a linear layer <code>nn.Linear<\/code>. You already have <code>weight<\/code> and <code>bias<\/code> parameters. But if you need a new parameter you use <a href=\"https:\/\/pytorch.org\/docs\/master\/nn.html#torch.nn.Module.register_parameter\" rel=\"nofollow noreferrer\"><code>register_parameter()<\/code><\/a> to register a new named parameter that is a tensor.<\/p>\n\n<p>When you register a new parameter it will appear inside <code>module.parameters()<\/code> iterator, but when you register a buffer it will not.<\/p>\n\n<blockquote>\n  <p>The difference:<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/pytorch.org\/docs\/master\/nn.html#torch.nn.Module.register_buffer\" rel=\"nofollow noreferrer\">Buffers<\/a> are named tensors that do not update gradients at every step, like parameters.\nFor buffers, you create your custom logic (fully up to you).<\/p>\n\n<p>The good thing is when you save the model, all params and buffers are saved, and when you move the model to or off the CUDA params and buffers will go as well.<\/p>\n"}]}
+{"QuestionId":57541091,"AnswerCount":2,"Tags":"<python><numpy><matplotlib><pytorch>","CreationDate":"2019-08-18T01:44:19.860","AcceptedAnswerId":57541732.0,"OwnerUserId":11628923.0,"Title":"Plotting predicted values causes error: 'Tensor' object has no attribute 'ndim'","Body":"<p>How do I properly write the code to plot the predicted values in this linear regression model?<\/p>\n\n<p>I am using this tutorial to learn Linear regression:<a href=\"https:\/\/www.deeplearningwizard.com\/deep_learning\/practical_pytorch\/pytorch_linear_regression\/\" rel=\"nofollow noreferrer\">https:\/\/www.deeplearningwizard.com\/deep_learning\/practical_pytorch\/pytorch_linear_regression\/<\/a><\/p>\n\n<p>I was able to successfully implement GPU. My issue is plotting the predicted values. I tried searching for solutions to learn how to keep the values as a tensor, but it seems that I don't have the syntactical knowledge to do this.<\/p>\n\n<p>Here is where I'll start<\/p>\n\n<pre><code>epochs = 100\nfor epoch in range(epochs):\n    epoch += 1\n\n    # Convert numpy array to torch Variable\n    if torch.cuda.is_available():\n        inputs = (torch.from_numpy(x_train).cuda())\n        labels = (torch.from_numpy(y_train).cuda())\n    else:\n        inputs = (torch.from_numpy(x_train))\n        labels = (torch.from_numpy(y_train))\n\n    # Clear gradients w.r.t. parameters\n    optimizer.zero_grad()\n\n    # Forward to get output\n    outputs = model(inputs)\n\n    # Calculate Loss\n    loss = criterion(outputs, labels)\n\n    # Getting gradients w.r.t. parameters\n    loss.backward()\n\n    # Updating parameters\n    optimizer.step()\n\n    # Logging\n    print('epoch {}, loss {}'.format(epoch, loss.item()))\n<\/code><\/pre>\n\n<p>Prediction is made here and I choose to use cuda<\/p>\n\n<pre><code>predicted = model(Variable(torch.from_numpy(x_train).requires_grad_().cuda()))\n\nprint(\"Predicted\")\nprint(predicted)\nprint(\"Output\")\nprint(y_train)\n\nplt.clf()\n\n# Get predictions\n\n#predicted = model(Variable(torch.from_numpy(x_train).requires_grad_().cuda()))\n\n# Plot true data\nplt.plot(x_train, y_train, 'go', label='True data', alpha=0.5)\n<\/code><\/pre>\n\n<p>The error is called here after it is unable to plot<\/p>\n\n<pre><code># Plot predictions\nplt.plot(x_train, predicted, '--', label='Predictions', alpha=0.5)\n\n# Legend and plot\nplt.legend(loc='best')\nplt.show()\n<\/code><\/pre>\n\n<p>Error Given:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"D:\/Test with GPU\/Linear regression.py\", line 101, in &lt;module&gt;\n    plt.plot(x_train, predicted, '--', label='Predictions', alpha=0.5)\n  File \"D:\\Anaconda3\\envs\\gputest\\lib\\site-packages\\matplotlib\\pyplot.py\", line 2795, in plot\n    is not None else {}), **kwargs)\n  File \"D:\\Anaconda3\\envs\\gputest\\lib\\site-packages\\matplotlib\\axes\\_axes.py\", line 1666, in plot\n    lines = [*self._get_lines(*args, data=data, **kwargs)]\n  File \"D:\\Anaconda3\\envs\\gputest\\lib\\site-packages\\matplotlib\\axes\\_base.py\", line 225, in __call__\n    yield from self._plot_args(this, kwargs)\n  File \"D:\\Anaconda3\\envs\\gputest\\lib\\site-packages\\matplotlib\\axes\\_base.py\", line 391, in _plot_args\n    x, y = self._xy_from_xy(x, y)\n  File \"D:\\Anaconda3\\envs\\gputest\\lib\\site-packages\\matplotlib\\axes\\_base.py\", line 271, in _xy_from_xy\n    if x.ndim &gt; 2 or y.ndim &gt; 2:\nAttributeError: 'Tensor' object has no attribute 'ndim'\n<\/code><\/pre>\n","answers":[{"AnswerId":"57546129","CreationDate":"2019-08-18T15:48:21.500","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>Ndim is a property required by matplotlib.<\/p>\n\n<pre><code>import matplotlib.pyplot as plt   \ntorch.Tensor.ndim = property(lambda x: len(x.size()))\nN = 42\nt = torch.rand(N)\nplt.plot(t)\nplt.show()\nplt.close()\n<\/code><\/pre>\n\n<p>But PyTorch doesn't have that for tensor till version 1.2 I think.<\/p>\n\n<p>This part <code>torch.Tensor.ndim = property(lambda x: len(x.size()))<\/code> you can add and you can plot PyTorch tensors without problems.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/SnlAy.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/SnlAy.png\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"57541732","CreationDate":"2019-08-18T04:50:15.820","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p><code>plt.plot<\/code> function expects its inputs to be <code>numpy<\/code> arrays, rather than <code>torch.tensor<\/code>.<br>\nYou can use <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/tensors.html#torch.Tensor.numpy\" rel=\"nofollow noreferrer\"><code>.numpy()<\/code><\/a> to view the internal data of the tensor as a <code>numpy<\/code> array.<\/p>\n\n<p>Try<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with torch.no_grad():\n  plt.plot(x_train, predicted.cpu().numpy(), '--', label='Predictions', alpha=0.5)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57541258,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-08-18T02:38:22.117","AcceptedAnswerId":null,"OwnerUserId":85248.0,"Title":"Keras is installed in Anaconda virtual environment but can't be imported","Body":"<p>I have <code>keras<\/code> installed into my Anaconda virtual environment but I can't import the module:<\/p>\n\n<pre><code>$ conda list | grep keras\nkeras                     2.2.4                    py37_1    conda-forge\nkeras-applications        1.0.8                      py_0  \nkeras-preprocessing       1.1.0                      py_1  \nkeras-retinanet           0.5.0                    pypi_0    pypi\n\n$ python\nPython 3.7.3 | packaged by conda-forge | (default, Mar 27 2019, 23:01:00) \n[GCC 7.3.0] :: Anaconda, Inc. on linux\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n&gt;&gt;&gt; import keras\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\nModuleNotFoundError: No module named 'keras'\n<\/code><\/pre>\n\n<p>What might be amiss?<\/p>\n","answers":[{"AnswerId":"57541279","CreationDate":"2019-08-18T02:45:08.380","ParentId":null,"OwnerUserId":"9712785","Title":null,"Body":"<p>The best option is to use venv for your virtual environment. For instructions:<\/p>\n\n<h1>Mac OS &amp; Linux: How To Setup Virtual Environment<\/h1>\n\n<p>1) After cloning the repo, <code>cd<\/code> into the repo and run the command: <code>python3 -m venv venv<\/code><\/p>\n\n<pre><code>    This will create the virtual environment. Make sure to name it venv because the .gitignore file\n    has been initialized to ignore it by default. \n<\/code><\/pre>\n\n<p>2) Activate the virtual environment by running the following command: <code>source venv\/bin\/activate<\/code><\/p>\n\n<p>3) To download the required packages simply run: <code>pip install -r requirements.txt<\/code><\/p>\n\n<pre><code>    This will go to the requirements file that is already generated in the repo and download any dependecies.\n<\/code><\/pre>\n\n<p>4) After adding any packages, make sure to update the requirements.txt file by running: <code>pip freeze &gt; requirements.txt<\/code><\/p>\n\n<p>5) At any moment, you can run <code>deactivate<\/code> to leave the virtual environment.<\/p>\n\n<h1>Windows: How To Setup Virtual Environment<\/h1>\n\n<p>1) After cloning the repo, <code>cd<\/code> into the repo and run the command: <code>python -m venv venv<\/code><\/p>\n\n<pre><code>    This will create the virtual environment. Make sure to name it venv because the .gitignore file\n    has been initialized to ignore it by default. \n<\/code><\/pre>\n\n<p>2) Activate the virtual environment by running the following command: <code>venv\\Scripts\\activate.bat<\/code><\/p>\n\n<pre><code>    To activate the virtual environment inside of a code editor's bash, run: venv\\Scripts\\activate.ps1\n<\/code><\/pre>\n\n<p>3) To download the required packages simply run: <code>pip install -r requirements.txt<\/code><\/p>\n\n<p>4) After adding any packages, make sure to update the requirements.txt file by running: <code>pip freeze<\/code> \nand then pasting them into the <code>requirements.txt<\/code> file.<\/p>\n\n<p>5) At any moment, you can run <code>deactivate<\/code> to leave the virtual environment.<\/p>\n\n<p>Finally, simply type <code>pip install keras<\/code><\/p>\n"}]}
+{"QuestionId":57541486,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><deep-learning><loss-function><triplet>","CreationDate":"2019-08-18T03:46:05.067","AcceptedAnswerId":null,"OwnerUserId":11542066.0,"Title":"Valid triplets in triplet loss of online mining","Body":"<p>I am trying to implement online-mining triplet loss in a random data. ( labels = [1,2,1,2] , embedding = [a,b,c,d] )\n <br \/>\nAnd I was referring to \"batch_all_triplet_loss \" function in this repo :(<a href=\"https:\/\/github.com\/omoindrot\/tensorflow-triplet-loss\/blob\/master\/model\/triplet_loss.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/omoindrot\/tensorflow-triplet-loss\/blob\/master\/model\/triplet_loss.py<\/a>).\n<br \/>\n<br \/>\nIn my understanding, the valid triplet is when anchor is different embedding with positive and negative. So in my example data above, the valid triplets are:<\/p>\n\n<p>anchor a -> positive: c, negatives : b, d <br \/>\nanchor b -> positive: d, negatives : a, c <br \/>\nanchor c -> positive: a, negatives : b, d <br \/>\nanchor d -> positiveL b, negatives : a, c <br \/><\/p>\n\n<p>So I should get 8 valid triplets after estimation.<\/p>\n\n<p>But I used the same data on the codes in above repo link, I got 32 valid triplets. And of course, the result(I had tried using Mnist dataset) is normal when using the codes in above repo link. However, I cant figure out where Im doing wrong..<\/p>\n\n<pre><code>def batch_all_triplet_loss(labels, embeddings, margin, squared=False):\n\n    ...\n    ...\n    ...\n\n    # Count number of positive triplets (where triplet_loss &gt; 0)\n    valid_triplets = tf.to_float(tf.greater(triplet_loss, 1e-16))\n\n    num_positive_triplets = tf.reduce_sum(valid_triplets)\n\n    #I print out the number of valid triplets\n    print(sess.run(num_positive_triplets))\n\n\n    num_valid_triplets = tf.reduce_sum(mask)\n    fraction_positive_triplets = num_positive_triplets \/ (num_valid_triplets + 1e-16)\n\n    # Get final mean triplet loss over the positive valid triplets\n    triplet_loss = tf.reduce_sum(triplet_loss) \/ (num_positive_triplets + 1e-16)\n\n    return triplet_loss, fraction_positive_triplets\n\nlabels = np.array([1,2,1,2])\nembeddings = np.random.rand(4,4)\nbatch_all_triplet_loss(labels, embeddings, 1.0)\n\n<\/code><\/pre>\n\n<p>I expect the number of valid triplet count is 8, but the right number is 32.<\/p>\n","answers":[]}
+{"QuestionId":57541506,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2019-08-18T03:52:36.257","AcceptedAnswerId":57611669.0,"OwnerUserId":1043971.0,"Title":"Enabling variable length input padding while batching in TFServing","Body":"<p>The following proto definition has the details of a flag (pad_variable_length_inputs) which can be enabled to pad requests. <\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/serving\/blob\/1f8f2902b6465f239bb58af2b3fb27ba73b5c7c5\/tensorflow_serving\/servables\/tensorflow\/session_bundle_config.proto\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/serving\/blob\/1f8f2902b6465f239bb58af2b3fb27ba73b5c7c5\/tensorflow_serving\/servables\/tensorflow\/session_bundle_config.proto<\/a><\/p>\n\n<p>My current batching config file looks like this : <\/p>\n\n<pre><code>max_batch_size { value: 300 }\nbatch_timeout_micros { value: 0 }\nmax_enqueued_batches { value: 1000000 }\nnum_batch_threads { value: 8 }\n<\/code><\/pre>\n\n<p>I am wondering how can I set the pad_variable_length_inputs through the config.<\/p>\n","answers":[{"AnswerId":"57611669","CreationDate":"2019-08-22T14:30:02.650","ParentId":null,"OwnerUserId":"1043971","Title":null,"Body":"<p>I finally figured it out. This is how it can be done : <\/p>\n\n<pre><code>max_batch_size { value: 300 }\nbatch_timeout_micros { value: 0 }\nmax_enqueued_batches { value: 1000000 }\nnum_batch_threads { value: 8 }\npad_variable_length_inputs: true\n<\/code><\/pre>\n"}]}
+{"QuestionId":57541534,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-08-18T03:58:03.960","AcceptedAnswerId":57542322.0,"OwnerUserId":11939660.0,"Title":"Weird shape requirement for `sample_weight` argument in Keras losses in TF2.0","Body":"<p>According to TF document, the the <code>sample_weight<\/code> argument can have shape <code>[batch_size]<\/code>. The relevant documentation is quoted below:<\/p>\n\n<blockquote>\n  <p><code>sample_weight<\/code>: Optional <code>Tensor<\/code> whose rank is either 0, or the same rank as <code>y_true<\/code>, or is broadcastable to <code>y_true<\/code>. <code>sample_weight<\/code> acts as a coefficient for the loss. If a scalar is provided, then the loss is simply scaled by the given value. If <code>sample_weight<\/code> is a tensor of size <code>[batch_size]<\/code>, then the total loss for each sample of the batch is rescaled by the corresponding element in the <code>sample_weight<\/code> vector. If the shape of <code>sample_weight<\/code> matches the shape of <code>y_pred<\/code>, then the loss of each measurable element of <code>y_pred<\/code> is scaled by the corresponding value of <code>sample_weight<\/code>.<\/p>\n<\/blockquote>\n\n<p>However, I cannot understand why the following code does not work.<\/p>\n\n<pre><code>import tensorflow as tf\n\ngt = tf.convert_to_tensor([1, 1, 1, 1, 1])\npred = tf.convert_to_tensor([1., 0., 1., 1., 0.])\nsample_weights = tf.convert_to_tensor([0, 1, 0, 0, 0])\n\nloss = tf.keras.losses.BinaryCrossentropy()(gt, pred, sample_weight=sample_weights)\nprint(loss)\n<\/code><\/pre>\n\n<p>The code throw this error:<\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors_impl.InvalidArgumentError: Can not squeeze dim[0], expected a dimension of 1, got 5 [Op:Squeeze]<\/p>\n<\/blockquote>\n\n<p>If I expand the dimensions of <code>gt<\/code>, <code>pred<\/code>, and <code>sample_weights<\/code>, then it works properly and output the expected loss value of 3.0849898.<\/p>\n\n<pre><code>import tensorflow as tf\n\ngt = tf.convert_to_tensor([1, 1, 1, 1, 1])\npred = tf.convert_to_tensor([1., 0., 1., 1., 0.])\nsample_weights = tf.convert_to_tensor([0, 1, 0, 0, 0])\n\n# expand dims\ngt = tf.expand_dims(gt, 1)\npred = tf.expand_dims(pred, 1)\nsample_weights = tf.expand_dims(sample_weights, 1)\n\nloss = tf.keras.losses.BinaryCrossentropy()(gt, pred, sample_weight=sample_weights)\nprint(loss)  # loss is 3.0849898\n<\/code><\/pre>\n","answers":[{"AnswerId":"57542322","CreationDate":"2019-08-18T06:50:52.793","ParentId":null,"OwnerUserId":"1805756","Title":null,"Body":"<p>The problem is not about <code>sample_weight<\/code> shape. It's <code>pred<\/code> and <code>gt<\/code> shape which should be <code>[batch_size, n_labels]<\/code>:<\/p>\n\n<pre><code>import tensorflow as tf\n\ngt = tf.convert_to_tensor([1, 1, 1, 1, 1])\npred = tf.convert_to_tensor([1., 0., 1., 1., 0.])\nsample_weights = tf.convert_to_tensor([0, 1, 0, 0, 0])\n\n# expand dims\ngt = tf.expand_dims(gt, 1)\npred = tf.expand_dims(pred, 1)\nprint(gt.shape, pred.shape) #(5, 1) (5, 1)\n\nloss = tf.keras.losses.BinaryCrossentropy()(gt, pred, sample_weight=sample_weights)\nprint(loss)  # loss is 3.0849898\n<\/code><\/pre>\n"}]}
+{"QuestionId":57541567,"AnswerCount":2,"Tags":"<python><tensorflow><cudnn>","CreationDate":"2019-08-18T04:05:55.200","AcceptedAnswerId":null,"OwnerUserId":9694304.0,"Title":"Tensorflow shows only \"successfully opened CUDA library libcublas.so.10.0 locally\" and nothing about cudnn","Body":"<p>My tensorflow  only prints out the line:<\/p>\n\n<p><code>I tensorflow\/stream_executor\/dso_loader.cc:152] successfully opened CUDA library libcublas.so.10.0 locally<\/code> when running.<\/p>\n\n<p>Tensorflow logs on the net has lots of other libraries being loaded like libcudnn.<\/p>\n\n<p>As I think my installation performance is not optimal, I am trying to find out if it is because of this. Any help will be appreciated!<\/p>\n\n<p>my tf is 1.13.1\nNVIDIA Driver Version: 418.67\nCUDA Version: 10.1  (I have also 10.0 installed. can this be the problem?)<\/p>\n","answers":[{"AnswerId":"57784170","CreationDate":"2019-09-04T08:13:24.667","ParentId":null,"OwnerUserId":"12018385","Title":null,"Body":"<p>Actually i always rely on a stable setup. And i tried most of the tf - cuda - cudnn versions. But most stable was tf 1.9.0 , CUDA 9.0, Cudnn 7 for me. Used it for too long without a problem. You should give it a try if it suits you.<\/p>\n"},{"AnswerId":"57707593","CreationDate":"2019-08-29T09:52:45.583","ParentId":null,"OwnerUserId":"1849975","Title":null,"Body":"<p>According to <code>TensorFlow<\/code> <a href=\"https:\/\/www.tensorflow.org\/install\/gpu\" rel=\"nofollow noreferrer\">documentation<\/a>, <code>cuDNN<\/code> is a requirement for <code>tensorflow-gpu<\/code>. If you don't have <code>cuDNN<\/code> installed,  you wouldn't be able to install <code>tensorflow-gpu<\/code> since the dependency library would be missing.<\/p>\n\n<p>So, if you have successfully installed <code>tensorflow-gpu<\/code> and are able to use it, e.g.<\/p>\n\n<pre><code>import tensorflow as tf\ntf.Session()\n<\/code><\/pre>\n\n<p>you are fine.<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>I just checker <a href=\"https:\/\/www.tensorflow.org\/install\/source#tested_build_configurations\" rel=\"nofollow noreferrer\">here<\/a> and <code>tensorflow_gpu-1.13.1<\/code> officially only supports <code>CUDA 10.0<\/code>. I would recommend to use it instead of <code>CUDA 10.1<\/code>.<\/p>\n\n<p>Further, <code>NVIDIA<\/code> recommends using driver version <code>410.48<\/code> with <code>CUDA 10.0<\/code>. I would stick with it as well.<\/p>\n"}]}
+{"QuestionId":57541670,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-08-18T04:31:06.733","AcceptedAnswerId":57541942.0,"OwnerUserId":6817332.0,"Title":"Behaviour of multiple MultiRNNCell?","Body":"<p>i'm trying to understand some RNN implementation with TensorFlow.\nDoes <code>MultiRNNCell<\/code> creates new object each time, or adds layers to the previous object? <\/p>\n\n<pre><code>cells_fwd_list = []\n\nfor num_cells in [256, 128, 64]:\n    cells_fwd_list.append(tf.nn.rnn_cell.LSTMCell(num_units=num_cells, activation=tf.nn.tanh))\n    cells_fwd = tf.nn.rnn_cell.MultiRNNCell(cells_fwd_list, state_is_tuple=True)\n<\/code><\/pre>\n\n<p>Does it mean <code>cells_fwd<\/code> has 3 layers of [256,128,64] units each?\nOR\n<code>cells_fwd<\/code> has 6 layers of [256,256,128,256,128,64] units each?<\/p>\n","answers":[{"AnswerId":"57541942","CreationDate":"2019-08-18T05:38:00.057","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p><code>MultiRNNCell<\/code> abstracts a sequence of RNN cells into one layer. In your code, you are creating a MultiRNNCell after each time you are creating a LSTMCell. So basically your code should be:<\/p>\n\n<pre><code>cells_fwd_list = []\n\nfor num_cells in [256, 128, 64]:\n    cells_fwd_list.append(tf.nn.rnn_cell.LSTMCell(num_units=num_cells, activation=tf.nn.tanh))\ncells_fwd = tf.nn.rnn_cell.MultiRNNCell(cells_fwd_list, state_is_tuple=True)\n<\/code><\/pre>\n\n<p>Now, your cells_fwd will be holding three LSTM layers internally comprising of 256, 128 and 64 units.<\/p>\n\n<p>Also, as Tensorflow 1.x is soon going to be deprecated, your TensorFlow 2.0 equivalent code should be something like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\ntimesteps, input_dim = 60, 100\ninputs = tf.keras.Input((timesteps, input_dim))\nlstm_cells = [tf.keras.layers.LSTMCell(units) for units in [256, 128, 64]]\nstacked_rnn_layer = tf.keras.layers.RNN(lstm_cells)(inputs)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57541763,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2019-08-18T04:58:32.603","AcceptedAnswerId":null,"OwnerUserId":11377572.0,"Title":"create Dense layers in a loop","Body":"<p>I need to create multiple dense layers by for loop, the number of iteration depends on the number of labels. I want to create one dense layer for each label. Each label has a different set of features, so I want to predict each label separately with corresponding feature set in each dense layer. Is that possible? The following code is my attempt. <\/p>\n\n<pre><code>layers = []\nfor i in range(num_labels):\n    h1 = Dense(num_genes_per+10, kernel_initializer='normal', input_dim = num_genes_per, activation='relu')(inputs)\n    h2 = Dense(int(num_genes_per\/2), kernel_initializer='normal', activation='relu')(h1) \n    output= Dense(1, kernel_initializer='normal', activation='linear')(h2)\n    layers.append(output)\n\nmerged_output = concatenate(layers, axis=1)\nmodel = Model(inputs, merged_output)\n<\/code><\/pre>\n\n<p>The output of each h2 will have shape [batch, 1], and the merged_output will have shape [batch, num_labels]. Is there any error in the above code?<\/p>\n\n<p>I know it is not efficient, but if I concatenate the different set of features into one input tensor, and use only one dense layer to predict all labels at same time, would it harms the prediction accuracy?<\/p>\n","answers":[{"AnswerId":"57544068","CreationDate":"2019-08-18T11:24:20.440","ParentId":null,"OwnerUserId":"1805756","Title":null,"Body":"<p>It depends on how you defined features and labels. If features 1, 2 and 3 are used to predict label 1 and they have no relation to label 2, It does not make sense to include it in label 3 inference. <\/p>\n"}]}
+{"QuestionId":57541947,"AnswerCount":2,"Tags":"<python><tensorflow><keras><lstm><recurrent-neural-network>","CreationDate":"2019-08-18T05:38:41.767","AcceptedAnswerId":57542163.0,"OwnerUserId":7296192.0,"Title":"Keras LSTM TensorFlow Error: 'Shapes must be equal rank, but are 1 and 0'","Body":"<p>I'm trying to create a Keras LSTM (Please note that I am new to LSTMs and RNNs in Keras). The neural network is supposed to take an input of 4116 values, and output 4116 values. This is to be done for 288 timesteps. I have 27 such timesteps (I realize this will likely lead to overfitting; I have a larger dataset, but first want to test my code with just 27 training examples).<\/p>\n\n<p>The training data is stored in two numpy arrays <code>x<\/code> and <code>y<\/code>. These variables have a shape of <code>(27, 288, 4116)<\/code>.<\/p>\n\n<p>My code:<\/p>\n\n<pre><code>datapoints = data.get.monthPoints(2, year)\nx, y = datapoints[:-1], datapoints[1:]\ndel datapoints\n\ninput_shape = x.shape[1:]\noutput_shape = y.shape[1:]\n\ncheckpoint = ModelCheckpoint('model\/files\/alpha.h5', monitor='val_loss', verbose=1, save_best_only=True, mode='auto', period=1)\nearly = EarlyStopping(monitor='val_loss', min_delta=0, patience=1, verbose=1, mode='auto')\n\nmodel = Sequential()\nmodel.add(LSTM(5488, activation='relu', input_shape=input_shape))\nmodel.add(RepeatVector(output_shape))\nmodel.add(LSTM(5488, activation='relu', return_sequences=True))\nmodel.add(TimeDistributed(Dense(output_shape)))\n\nmodel.compile(loss='mse', optimizer='adam')\nmodel.fit(x, y, epochs=100, batch_size=8, callbacks = [checkpoint, early])\n<\/code><\/pre>\n\n<p>When I run the program, I get the following error(s):<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: Shapes must be equal rank, but are 1 and 0\n        From merging shape 1 with other shapes. for 'repeat_vector\/stack_1' (op: 'Pack') with input shapes: [], [2], []\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>During handling of the above exception, another exception occurred:\nValueError: Shapes must be equal rank, but are 1 and 0\n        From merging shape 1 with other shapes. for 'repeat_vector\/stack_1' (op: 'Pack') with input shapes: [], [2], []\n<\/code><\/pre>\n\n<p>I've seen a few other similar questions like <a href=\"https:\/\/stackoverflow.com\/q\/46273364\/7296192\">this<\/a> and <a href=\"https:\/\/stackoverflow.com\/q\/54806450\/7296192\">this<\/a>, but they haven't offered solutions that fix my problem or the solutions have been unclear.<\/p>\n\n<p>I guess my problem has something to do with me structuring the network incorrectly or formatting my data incorrectly.<\/p>\n\n<p>Any insight would me much appreciated.<\/p>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"57542144","CreationDate":"2019-08-18T06:18:37.203","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You probably want to repeat the output of first LSTM layer as much as the number of timesteps in model's output sequence (i.e. <code>y<\/code>). Therefore, it should be:<\/p>\n\n<pre><code>model.add(RepeatVector(y.shape[1]))\n<\/code><\/pre>\n"},{"AnswerId":"57542163","CreationDate":"2019-08-18T06:21:40.560","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>There are two problems in your code. First, in <code>RepeatVector<\/code> you are sending a list by passing y.shape[1:]. In <code>RepeatVector<\/code> you should be sending an integer.<\/p>\n\n<p>Second problem is in <code>TimeDistributed<\/code>. Send in it the number of times you would like your second dimension to be repeated at.\nSo your code should be:<\/p>\n\n<pre><code>repeat_vector_units = x.shape[1]\noutput_units = x.shape[2]\n\nmodel = Sequential()\nmodel.add(LSTM(5488, activation='relu', input_shape=input_shape))\nmodel.add(RepeatVector(repeat_vector_units))        ### Change here\nmodel.add(LSTM(5488, activation='relu', return_sequences=True))\nmodel.add(TimeDistributed(Dense(output_units)))     #### Change here\n<\/code><\/pre>\n"}]}
+{"QuestionId":57542016,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorflow2.0>","CreationDate":"2019-08-18T05:53:24.317","AcceptedAnswerId":null,"OwnerUserId":4061996.0,"Title":"Validation loss and accuracy not changing from training","Body":"<p>I'm using transfer learning with MobileNet v2 with Tensorflow 2.0 to train a network for skin quality classification. Everything is working perfectly fine, however it seems that the validation loss and validation accuracy do not change.<\/p>\n\n<p>I've saved the model files and run predictions on a variety of images, but this also led to no success as the model outputs exactly the same predictions for every image.<\/p>\n\n<p><img src=\"https:\/\/i.imgur.com\/ngYpqZO.png\" alt=\"epoch outputs\"><\/p>\n\n<p>A very peculiar problem and google seems to have no answers for this. Any help or insight would be much appreciated.<\/p>\n","answers":[{"AnswerId":"57542083","CreationDate":"2019-08-18T06:06:52.963","ParentId":null,"OwnerUserId":"9694304","Title":null,"Body":"<p>I would think that can be one of several things:<\/p>\n\n<ol>\n<li><p>You have mislabeled you data. The CNN can't learn because the labels are \"random\". This is probably the most probable cause. Check carefully that you have encoded your labels properly and loaded them correctly.<\/p><\/li>\n<li><p>Your learning rate is too high and you have no lr schedule. This is also very common problem, especially when you take the configuration from 8 v100 gpu training and train it on 1 gpu. You need to adjust your lr accordingly.<\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":57542598,"AnswerCount":1,"Tags":"<java><android><android-studio><tensorflow><gradle>","CreationDate":"2019-08-18T07:36:33.727","AcceptedAnswerId":57543277.0,"OwnerUserId":9617817.0,"Title":"How to configure android studio to build this project?","Body":"<p>I have cloned this project <a href=\"https:\/\/github.com\/asoehlke\/ai-piano-accompanist-app\" rel=\"nofollow noreferrer\">https:\/\/github.com\/asoehlke\/ai-piano-accompanist-app<\/a>, but could not build it on my android studio.<\/p>\n\n<p>After running it I get:<\/p>\n\n<blockquote>\n  <p>No toolchains found in the NDK toolchains folder for ABI with prefix: mips64el-linux-android\n  This version of the NDK may be incompatible with the Android Gradle plugin version 3.0 or older.\n  Please use plugin version 3.1 or newer.<\/p>\n<\/blockquote>\n\n<p>After updating plugin to version 3.4.1 and sync project I get:<\/p>\n\n<blockquote>\n  <p>ERROR: Could not find com.android.tools.build:gradle-experimental:3.4.1.\n  Searched in the following locations:\n    - <a href=\"https:\/\/jcenter.bintray.com\/com\/android\/tools\/build\/gradle-experimental\/3.4.1\/gradle-experimental-3.4.1.pom\" rel=\"nofollow noreferrer\">https:\/\/jcenter.bintray.com\/com\/android\/tools\/build\/gradle-experimental\/3.4.1\/gradle-experimental-3.4.1.pom<\/a>\n    - <a href=\"https:\/\/jcenter.bintray.com\/com\/android\/tools\/build\/gradle-experimental\/3.4.1\/gradle-experimental-3.4.1.jar\" rel=\"nofollow noreferrer\">https:\/\/jcenter.bintray.com\/com\/android\/tools\/build\/gradle-experimental\/3.4.1\/gradle-experimental-3.4.1.jar<\/a>\n    - <a href=\"https:\/\/dl.google.com\/dl\/android\/maven2\/com\/android\/tools\/build\/gradle-experimental\/3.4.1\/gradle-experimental-3.4.1.pom\" rel=\"nofollow noreferrer\">https:\/\/dl.google.com\/dl\/android\/maven2\/com\/android\/tools\/build\/gradle-experimental\/3.4.1\/gradle-experimental-3.4.1.pom<\/a>\n    - <a href=\"https:\/\/dl.google.com\/dl\/android\/maven2\/com\/android\/tools\/build\/gradle-experimental\/3.4.1\/gradle-experimental-3.4.1.jar\" rel=\"nofollow noreferrer\">https:\/\/dl.google.com\/dl\/android\/maven2\/com\/android\/tools\/build\/gradle-experimental\/3.4.1\/gradle-experimental-3.4.1.jar<\/a>\n  Required by:\n      project :<\/p>\n<\/blockquote>\n\n<p>Before update<\/p>\n\n<pre><code>classpath 'com.android.tools.build:gradle-experimental:0.11.0'\n\ndistributionUrl=https\\:\/\/services.gradle.org\/distributions\/gradle-4.1-all.zip\n<\/code><\/pre>\n\n<p>After update<\/p>\n\n<pre><code>classpath 'com.android.tools.build:gradle-experimental:3.4.1'\n\ndistributionUrl=https\\:\/\/services.gradle.org\/distributions\/gradle-5.1.1-all.zip\n<\/code><\/pre>\n\n<p>The answers from <a href=\"https:\/\/stackoverflow.com\/questions\/56169465\/could-not-find-com-android-tools-buildgradle-experimental3-4-0-in-android-stud\">Could not find com.android.tools.build:gradle-experimental:3.4.0 in android studio<\/a> also did not help me.<\/p>\n","answers":[{"AnswerId":"57543277","CreationDate":"2019-08-18T09:34:00.470","ParentId":null,"OwnerUserId":"2016562","Title":null,"Body":"<p>Use the <a href=\"https:\/\/developer.android.com\/studio\/releases\/gradle-plugin#3-4-0\" rel=\"nofollow noreferrer\">android plugin<\/a><\/p>\n\n<pre><code>classpath 'com.android.tools.build:gradle:3.4.2'\n<\/code><\/pre>\n\n<p>instead of experimental plugin.<\/p>\n\n<p>Also update the <code>repositories<\/code> blocks:<\/p>\n\n<pre><code>buildscript {\n    repositories {\n        google()\n        \/\/...\n    }\n    dependencies {\n        classpath 'com.android.tools.build:gradle:3.4.2'\n    }\n}\n<\/code><\/pre>\n\n<p>and:<\/p>\n\n<pre><code>allprojects {\n    repositories {\n        google()\n        jcenter()\n    }\n}\n<\/code><\/pre>\n\n<p>In <code>app\/build.gradle<\/code> you have to change the plugin:<\/p>\n\n<pre><code>apply plugin: 'com.android.application'\n<\/code><\/pre>\n\n<p>instead of<\/p>\n\n<pre><code>apply plugin: 'com.android.model.application'\n<\/code><\/pre>\n\n<p>Also you have to adapt the <code>model<\/code> (revome the model keyword)<\/p>\n\n<pre><code>android {\n  defaultConfig {\n      minSdkVersion XX\n      targetSdkVersion 23\n      ...\n  }\n  buildTypes {...}\n  ...\n}\n\ndependencies {\n   \/\/...\n}  \n<\/code><\/pre>\n\n<p>Check the <a href=\"https:\/\/developer.android.com\/studio\/build#module-level\" rel=\"nofollow noreferrer\">documentation<\/a>.<\/p>\n"}]}
+{"QuestionId":57542772,"AnswerCount":0,"Tags":"<python><tensorflow><installation><nvidia>","CreationDate":"2019-08-18T08:07:34.767","AcceptedAnswerId":null,"OwnerUserId":11427310.0,"Title":"tensorflow-gpu did not speed up training","Body":"<p>I rencently built a CNN and I then installed tensorflow-gpu along with CUDA to speed up training. But the it did not speed it up.<\/p>\n\n<p>I first installed tensorfow-gpu as follow:\n<code>sudo pip3 install --upgrade tensorflow-gpu<\/code><\/p>\n\n<p>then, i install CUDA un CUDNN as follow:\n<code>sudo pacman -S cuda cudnn<\/code><\/p>\n\n<p>I finally checked that the installation went good with:\n<code>nvidia-smi<\/code><\/p>\n\n<p>So far, all worked properly. But I then launched the training of my CNN and I did not see any difference in training time.<\/p>\n\n<p>The CNN is to classify MNIST and is as follow :<\/p>\n\n<ul>\n<li>1 conv layer with max pooling : 28x28x1 input, 14x14x6 output. <\/li>\n<li>1 conv layer with max poolling : 14x14x6 input, 7x7x16 output. <\/li>\n<li>1 fully connected : 7x7x16=784 input, 128 output. <\/li>\n<li>1 fully connected : 128 in, 10 output.<\/li>\n<\/ul>\n\n<p>1 epoch is around 21 seconds.\nthe whole training time for 100 epochs is around 35 minutes.<\/p>\n\n<p>I have got: <\/p>\n\n<ul>\n<li>GPU -> GTX 850M. <\/li>\n<li>CUDA -> version 10.2.<\/li>\n<li>CPU -> i7 4700HQ<\/li>\n<li>OS -> Manjaro<\/li>\n<\/ul>\n\n<p>So, do you have any idea why I did not see improvement in training time ? Is there something I have to add in my code ? I thought that tensorflow were automatically choosing between CPU or GPU if both were available ?<\/p>\n","answers":[]}
+{"QuestionId":57542918,"AnswerCount":0,"Tags":"<keras><image-segmentation><data-augmentation>","CreationDate":"2019-08-18T08:32:39.953","AcceptedAnswerId":null,"OwnerUserId":3177661.0,"Title":"Keras data augmentaion changes pixel values for masks (segmentation)","Body":"<p>Iam using runtime data augmentation using generators in keras for segmentation problem..<\/p>\n\n<p>Here is my data generator<\/p>\n\n<pre><code>    data_gen_args = dict(\n                     width_shift_range=0.1,\n                     height_shift_range=0.1,\n                     zoom_range=0.2,\n                     horizontal_flip=True,\n                     validation_split=0.2\n                    )\n\nimage_datagen = ImageDataGenerator(**data_gen_args)\n\ndef generate_data_generator(generator, Xi, Yi):\n    genXi = generator.flow(Xi, seed=7, batch_size=32)\n    genYi = generator.flow(Yi, seed=7,batch_size=32)\n\n\n    while True:\n            Xi = genXi.next()\n            Yi = genYi.next()\n            print(Yi.dtype)\n            print(np.unique(Yi))\n            yield (Xi, Yi)\n\ntrain_generator = generate_data_generator(image_datagen,\n                                    x_train,\n                                    y_train)\n<\/code><\/pre>\n\n<p>My labels are in a numpy array with data type float 32 and value 0.0 and 1.0.<\/p>\n\n<pre><code>#Output of np.unique(y_train)\narray([0., 1.], dtype=float32\n<\/code><\/pre>\n\n<p>However, the data generator seems to modifies pixel values as shown below:- <\/p>\n\n<pre><code>#Output of print(np.unique(Yi))\n[0.00000000e+00 1.01742386e-04 1.74021334e-04 ... 9.99918878e-01\n 9.99988437e-01 1.00000000e+00]\n<\/code><\/pre>\n\n<p>It is supposed to have same values(0.0 and 1.0) after data geneartion..\nAlso, the the official documentation shows an example using same augmentation arguments for generating mask and images together.<\/p>\n\n<p>However when i remove shift and zoom iam getting (0.0 and 1.0) as output.\nKeras verion 2.2.4,Python 3.6.8<\/p>\n\n<p>UPDATE:-<\/p>\n\n<p>I saved those images as numpy array and plotted it using matplotlib.It looks like the edges are smoothly interpolated (0.0-1.0) somehow upon including shifts and zoom augmentation. I can round these values in my custom generator as a hack; but i still don't understand the root cause (in case of normal images this is quite unnoticeable and has no adverse effects; but in masks we don't want to change label values )!!! <\/p>\n\n<p>Still wondering.. is this a bug (nobody has mentioned it so far)or problem with my custom code ??<\/p>\n","answers":[]}
+{"QuestionId":57542946,"AnswerCount":2,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-08-18T08:37:44.803","AcceptedAnswerId":null,"OwnerUserId":11792842.0,"Title":"ValueError: A `Concatenate` layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [(None, 523, 523, 32), etc","Body":"<p>I am trying to concatenate layers Using the following code using tensorflow , but getting an unexpected error. I am new to tensorflow <\/p>\n\n<pre><code>inp = Input(shape=(1050,1050,3))\nx1= layers.Conv2D(16 ,(3,3), activation='relu')(inp)\nx1= layers.Conv2D(32,(3,3), activation='relu')(x1)\nx1= layers.MaxPooling2D(2,2)(x1)\nx2= layers.Conv2D(32,(3,3), activation='relu')(x1)\nx2= layers.Conv2D(64,(3,3), activation='relu')(x2)\nx2= layers.MaxPooling2D(3,3)(x2)\nx3= layers.Conv2D(64,(3,3), activation='relu')(x2)\nx3= layers.Conv2D(64,(2,2), activation='relu')(x3)\nx3= layers.Conv2D(64,(3,3), activation='relu')(x3)\nx3= layers.Dropout(0.2)(x3)\nx3= layers.MaxPooling2D(2,2)(x3)\nx4= layers.Conv2D(64,(3,3), activation='relu')(x3)\nx4= layers.MaxPooling2D(2,2)(x4)\nx = layers.Dropout(0.2)(x4)\no = layers.Concatenate(axis=3)([x1, x2, x3, x4, x])\ny = layers.Flatten()(o)\ny = layers.Dense(1024, activation='relu')(y)\ny = layers.Dense(5, activation='softmax')(y) \n\nmodel = Model(inp, y)\nmodel.summary()\nmodel.compile(loss='sparse_categorical_crossentropy',optimizer=RMSprop(lr=0.001),metrics=['accuracy'])\n\n<\/code><\/pre>\n\n<p>The main error can be seen in the heading\nBut i have provided the traceback error for reference\nAnd the error is <\/p>\n\n<pre><code>ValueError                                Traceback (most recent call last)\n&lt;ipython-input-12-31a1fcec98a4&gt; in &lt;module&gt;\n     14 x4= layers.MaxPooling2D(2,2)(x4)\n     15 x = layers.Dropout(0.2)(x4)\n---&gt; 16 o = layers.Concatenate(axis=3)([x1, x2, x3, x4, x])\n     17 y = layers.Flatten()(o)\n     18 y = layers.Dense(1024, activation='relu')(y)\n\n\/opt\/conda\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in __call__(self, inputs, *args, **kwargs)\n    589           # Build layer if applicable (if the `build` method has been\n    590           # overridden).\n--&gt; 591           self._maybe_build(inputs)\n    592 \n    593           # Wrapping `call` function in autograph to allow for dynamic control\n\n\/opt\/conda\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in _maybe_build(self, inputs)\n   1879       # operations.\n   1880       with tf_utils.maybe_init_scope(self):\n-&gt; 1881         self.build(input_shapes)\n   1882     # We must set self.built since user defined build functions are not\n   1883     # constrained to set self.built.\n\n\/opt\/conda\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/utils\/tf_utils.py in wrapper(instance, input_shape)\n    293     if input_shape is not None:\n    294       input_shape = convert_shapes(input_shape, to_tuples=True)\n--&gt; 295     output_shape = fn(instance, input_shape)\n    296     # Return shapes from `fn` as TensorShapes.\n    297     if output_shape is not None:\n\n\/opt\/conda\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/layers\/merge.py in build(self, input_shape)\n    389                        'inputs with matching shapes '\n    390                        'except for the concat axis. '\n--&gt; 391                        'Got inputs shapes: %s' % (input_shape))\n    392 \n    393   def _merge_function(self, inputs):\n\nValueError: A `Concatenate` layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [(None, 523, 523, 32), (None, 173, 173, 64), (None, 84, 84, 64), (None, 41, 41, 64), (None, 41, 41, 64)]\n\n<\/code><\/pre>\n\n<p>I have imported all necessary files required to run the code using <code>tensorflow.keras<\/code><\/p>\n","answers":[{"AnswerId":"57543111","CreationDate":"2019-08-18T09:03:32.873","ParentId":null,"OwnerUserId":"9571777","Title":null,"Body":"<p>If you use <code>model.summary()<\/code> while adding layers, you'll see that the outputs of the layers you're trying to concatenate have very different shapes. Precisely, in fact, the shapes described in the error. As the error says, you'll need to ameliorate this. Operational variance is causing this, as you'll quickly discover by checking output shapes with the aforementioned command. Aside from operational variances, convolutional kernel variances relative to input shapes will throw an additional wrench in it, for which the layer parameter <code>padding='same'<\/code> can assist.\nTo be clear (since you have model.summary in your code), I mean add layers, check the summary, add more layers and check it again. you'll see the shape problems emerge.<\/p>\n"},{"AnswerId":"57546258","CreationDate":"2019-08-18T16:06:13.160","ParentId":null,"OwnerUserId":"7984359","Title":null,"Body":"<p>You cannot perform concatenate operation of input with different dimension i.e height and width. In your case you are trying to perform this operation <code>layers.Concatenate(axis=3)([x1, x2, x3, x4, x])<\/code> where <\/p>\n\n<pre><code>x1 has dimension = (None, 523, 523, 32)\nx2 has dimension = (None, 173, 173, 64)\nx3 has dimension = (None, 84, 84, 64)\nx4 has dimension = (None, 41, 41, 64)\nand x has dimension = (None, 41, 41, 64)\n<\/code><\/pre>\n\n<p>The error occurred because, all of the input dimension i.e height and width to concatenate are different. To solve the error you will have to get all the input to same dimension i.e same height and width, this could be achieved by sampling the  layers to a fixed dimension. Based on your use case you could either downsample or upsample to achieve the required dimension.<\/p>\n\n<pre><code>ValueError: A `Concatenate` layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [(None, 523, 523, 32), (None, 173, 173, 64), (None, 84, 84, 64), (None, 41, 41, 64), (None, 41, 41, 64)]\n<\/code><\/pre>\n\n<p>The error states <code>layer requires inputs with matching shapes<\/code>, this are nothing but the height and width of the input.<\/p>\n"}]}
+{"QuestionId":57543102,"AnswerCount":0,"Tags":"<tensorflow><svm>","CreationDate":"2019-08-18T09:01:30.397","AcceptedAnswerId":null,"OwnerUserId":11942162.0,"Title":"How can I build a svm in tensorflow?","Body":"<p>I try to use svm in tensorflow, and I found a code for svm when calculating the cross entrophy in loss function.Does this code work for svm?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def calculate_loss(logits_list, labels_list, regularizer_rate):\n    weight_coefficients = [1]\n\n    # xent_mean_list = []\n    # for i,(logits,labels,weight) in enumerate(zip(logits_list,labels_list,weight_coefficients)):\n    #     if i == 0:\n    #         xent = tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=labels)\n    #         xent_mean = tf.reduce_mean(xent)\n    #         xent_mean_list.append(weight*xent_mean)\n    # if regularizer_rate != 0:\n    #     loss = reduce(tf.add,xent_mean_list) + tf.add_n(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES))\n    # else:\n    #     loss = reduce(tf.add, xent_mean_list)\n    # return loss\n\n    xent_mean_list = []\n    for i, (logits, labels, weight) in enumerate(zip(logits_list, labels_list, weight_coefficients)):\n        if i == 0:\n            cross_entropy = tf.square(tf.maximum(tf.zeros([train_batch_size, 2]), 1 - logits * labels))\n            xent_mean = tf.reduce_mean(tf.square(logits_list)) + tf.reduce_mean(cross_entropy)\n            xent_mean_list.append(weight * xent_mean)\n    if regularizer_rate != 0:\n        loss = reduce(tf.add, xent_mean_list) + tf.add_n(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES))\n    else:\n        loss = reduce(tf.add, xent_mean_list)\n    return loss\n<\/code><\/pre>\n\n<p>The loss function did work, but I don't know whether it's just common cross entrophy loss not including svm. Excuse my poor maths.<\/p>\n","answers":[]}
+{"QuestionId":57543238,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-18T09:25:50.193","AcceptedAnswerId":null,"OwnerUserId":2469488.0,"Title":"Keras Calling model prediction with different input generate same result?","Body":"<p>Here, call predict multiple times<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    m = tf.keras.models.load_model(model_path, compile=False)\n    for i, (img_id, corr) in enumerate(zip(csv_file['id'], csv_file['corr'])):\n        img_filename = os.path.join(data_dir, 'data\/piccollage_data\/train_imgs\/{}.png'.format(img_id))\n        img = tf.expand_dims(load_image(img_filename), 0)\n        result = m.predict(img, steps=1)\n        print ('idx:{} result: {}, gt: {}'.format(i, result, corr))\n<\/code><\/pre>\n\n<p>And the results make me very confused:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/6rACVm.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/6rACVm.png\" alt=\"And the results:\"><\/a><\/p>\n\n<p>I tried to find if any questions is similar to this, but I only found it's about training problems <a href=\"https:\/\/datascience.stackexchange.com\/q\/46779\/23954\">link<\/a>.<\/p>\n\n<p>And I also tried other method e.g. change input as dataset (No problem), , another brute solution is reload model every for loop (It's very stupid) and last solution seems using model serving <a href=\"https:\/\/medium.com\/tensorflow\/training-and-serving-ml-models-with-tf-keras-fd975cc0fa27\" rel=\"nofollow noreferrer\">link<\/a>.<\/p>\n\n<p>So what I want to ask is that \"Why the code I wrote above will generate same results\"? Can anyone give me clue? <\/p>\n","answers":[]}
+{"QuestionId":57543787,"AnswerCount":2,"Tags":"<pytorch><broadcasting>","CreationDate":"2019-08-18T10:43:22.493","AcceptedAnswerId":null,"OwnerUserId":10855529.0,"Title":"Why does broadcasting give an empty tensor?","Body":"<pre><code>x = torch.randn(1, 1, 0)\n\ny = torch.randn(4, 1, 1)\n<\/code><\/pre>\n\n<pre><code>(x+y)\n<\/code><\/pre>\n\n<p>tensor([], size=(4, 1, 0))<\/p>\n\n<pre><code>(x + y).shape\n<\/code><\/pre>\n\n<p>torch.Size([4, 1, 0])<\/p>\n\n<p>shouldn\u2019t it have been 4, 1, 1, just y?<\/p>\n","answers":[{"AnswerId":"57544684","CreationDate":"2019-08-18T12:43:02.627","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Your <code>x<\/code> has 3rd dimension equal to 0 - it's an empty tensor. The thirs dimension of <code>x<\/code> determines the third dimension of the result: <code>torch.Size([4, 1, 0])<\/code> - which is also empty.  <\/p>\n\n<p>try <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = torch.randn(1,1,1)\n<\/code><\/pre>\n"},{"AnswerId":"57545890","CreationDate":"2019-08-18T15:18:39.280","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>It is so by design. <\/p>\n\n<p>Tensor <code>x<\/code> has one dimension 0.<\/p>\n\n<pre><code>import torch\nx = torch.randn(1, 1, 0)\nprint(x) # tensor([], size=(1, 1, 0))\n<\/code><\/pre>\n\n<p>These tensors are limited and I think the design is bad, but this is my opinion.\nFor instance such tensors cannot be concatenated.<\/p>\n\n<pre><code>static void check_cat_no_zero_dim(TensorList tensors) {\n  for(size_t i = 0; i &lt; tensors.size(); ++i) {\n    auto&amp; t = tensors[i];\n    TORCH_CHECK(t.dim() &gt; 0,\n             \"zero-dimensional tensor (at position \", i, \") cannot be concatenated\");\n  }\n}\n<\/code><\/pre>\n\n<p>Also, what you noticed is inability to work with <code>+<\/code> operator.<\/p>\n\n<p>So probable just like PyTorch tensors have the check if dimension is non-negative, it should be actually if <code>&gt;0<\/code> check. <\/p>\n\n<pre><code>inline void check_size_nonnegative(IntArrayRef size) {\n  for (auto x: size) {\n    TORCH_CHECK(x &gt;= 0, \"Trying to create tensor with negative dimension \", x, \": \", size);\n  }\n}\n<\/code><\/pre>\n\n<p>Again this is just my point of view.<\/p>\n"}]}
+{"QuestionId":57543839,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorflow-serving><yolo>","CreationDate":"2019-08-18T10:52:28.530","AcceptedAnswerId":null,"OwnerUserId":6482145.0,"Title":"How to fix 'ValueError: Too many elements provided. Needed at most' error in tensorflow?","Body":"<p>I have trying to serve keras-yolo3 model on tf-serving<\/p>\n\n<p>this is my model summary<\/p>\n\n<pre><code>Layer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput_1 (InputLayer)            (None, None, None, 3 0                                            \n__________________________________________________________________________________________________\nconv2d_1 (Conv2D)               (None, None, None, 3 864         input_1[0][0]                    \n__________________________________________________________________________________________________\nbatch_normalization_1 (BatchNor (None, None, None, 3 128         conv2d_1[0][0]                   \n__________________________________________________________________________________________________\nleaky_re_lu_1 (LeakyReLU)       (None, None, None, 3 0           batch_normalization_1[0][0]      \n__________________________________________________________________________________________________\nzero_padding2d_1 (ZeroPadding2D (None, None, None, 3 0           leaky_re_lu_1[0][0]              \n__________________________________________________________________________________________________\nconv2d_2 (Conv2D)               (None, None, None, 6 18432       zero_padding2d_1[0][0]           \n__________________________________________________________________________________________________\nbatch_normalization_2 (BatchNor (None, None, None, 6 256         conv2d_2[0][0]                   \n__________________________________________________________________________________________________\nleaky_re_lu_2 (LeakyReLU)       (None, None, None, 6 0           batch_normalization_2[0][0]      \n__________________________________________________________________________________________________\nconv2d_3 (Conv2D)               (None, None, None, 3 2048        leaky_re_lu_2[0][0]              \n__________________________________________________________________________________________________\nbatch_normalization_3 (BatchNor (None, None, None, 3 128         conv2d_3[0][0]                   \n__________________________________________________________________________________________________\nleaky_re_lu_3 (LeakyReLU)       (None, None, None, 3 0           batch_normalization_3[0][0]      \n__________________________________________________________________________________________________\nconv2d_4 (Conv2D)               (None, None, None, 6 18432       leaky_re_lu_3[0][0]              \n__________________________________________________________________________________________________\nbatch_normalization_4 (BatchNor (None, None, None, 6 256         conv2d_4[0][0]                   \n__________________________________________________________________________________________________\nleaky_re_lu_4 (LeakyReLU)       (None, None, None, 6 0           batch_normalization_4[0][0]      \n__________________________________________________________________________________________________\nadd_1 (Add)                     (None, None, None, 6 0           leaky_re_lu_2[0][0]              \n                                                                 leaky_re_lu_4[0][0]              \n__________________________________________________________________________________________________\nzero_padding2d_2 (ZeroPadding2D (None, None, None, 6 0           add_1[0][0]                      \n__________________________________________________________________________________________________\nconv2d_5 (Conv2D)               (None, None, None, 1 73728       zero_padding2d_2[0][0]           \n__________________________________________________________________________________________________\nbatch_normalization_5 (BatchNor (None, None, None, 1 512         conv2d_5[0][0]                   \n__________________________________________________________________________________________________\nleaky_re_lu_5 (LeakyReLU)       (None, None, None, 1 0           batch_normalization_5[0][0]      \n__________________________________________________________________________________________________\nconv2d_6 (Conv2D)               (None, None, None, 6 8192        leaky_re_lu_5[0][0]              \n__________________________________________________________________________________________________\nbatch_normalization_6 (BatchNor (None, None, None, 6 256         conv2d_6[0][0]                   \n__________________________________________________________________________________________________\nleaky_re_lu_6 (LeakyReLU)       (None, None, None, 6 0           batch_normalization_6[0][0]      \n__________________________________________________________________________________________________\nconv2d_7 (Conv2D)               (None, None, None, 1 73728       leaky_re_lu_6[0][0]              \n__________________________________________________________________________________________________\nbatch_normalization_7 (BatchNor (None, None, None, 1 512         conv2d_7[0][0]                   \n__________________________________________________________________________________________________\nleaky_re_lu_7 (LeakyReLU)       (None, None, None, 1 0           batch_normalization_7[0][0]      \n__________________________________________________________________________________________________\nadd_2 (Add)                     (None, None, None, 1 0           leaky_re_lu_5[0][0]              \n                                                                 leaky_re_lu_7[0][0]              \n__________________________________________________________________________________________________\nconv2d_8 (Conv2D)               (None, None, None, 6 8192        add_2[0][0]                      \n__________________________________________________________________________________________________\nbatch_normalization_8 (BatchNor (None, None, None, 6 256         conv2d_8[0][0]                   \n__________________________________________________________________________________________________\nleaky_re_lu_8 (LeakyReLU)       (None, None, None, 6 0           batch_normalization_8[0][0]      \n__________________________________________________________________________________________________\nconv2d_9 (Conv2D)               (None, None, None, 1 73728       leaky_re_lu_8[0][0]              \n__________________________________________________________________________________________________\nbatch_normalization_9 (BatchNor (None, None, None, 1 512         conv2d_9[0][0]                   \n__________________________________________________________________________________________________\nleaky_re_lu_9 (LeakyReLU)       (None, None, None, 1 0           batch_normalization_9[0][0]      \n__________________________________________________________________________________________________\nadd_3 (Add)                     (None, None, None, 1 0           add_2[0][0]                      \n                                                                 leaky_re_lu_9[0][0]              \n__________________________________________________________________________________________________\nzero_padding2d_3 (ZeroPadding2D (None, None, None, 1 0           add_3[0][0]                      \n__________________________________________________________________________________________________\nconv2d_10 (Conv2D)              (None, None, None, 2 294912      zero_padding2d_3[0][0]           \n__________________________________________________________________________________________________\nbatch_normalization_10 (BatchNo (None, None, None, 2 1024        conv2d_10[0][0]                  \n__________________________________________________________________________________________________\nleaky_re_lu_10 (LeakyReLU)      (None, None, None, 2 0           batch_normalization_10[0][0]     \n__________________________________________________________________________________________________\nconv2d_11 (Conv2D)              (None, None, None, 1 32768       leaky_re_lu_10[0][0]             \n__________________________________________________________________________________________________\nbatch_normalization_11 (BatchNo (None, None, None, 1 512         conv2d_11[0][0]                  \n__________________________________________________________________________________________________\nleaky_re_lu_11 (LeakyReLU)      (None, None, None, 1 0           batch_normalization_11[0][0]     \n__________________________________________________________________________________________________\nconv2d_12 (Conv2D)              (None, None, None, 2 294912      leaky_re_lu_11[0][0]             \n__________________________________________________________________________________________________\nbatch_normalization_12 (BatchNo (None, None, None, 2 1024        conv2d_12[0][0]                  \n__________________________________________________________________________________________________\nleaky_re_lu_12 (LeakyReLU)      (None, None, None, 2 0           batch_normalization_12[0][0]     \n__________________________________________________________________________________________________\nadd_4 (Add)                     (None, None, None, 2 0           leaky_re_lu_10[0][0]             \n                                                                 leaky_re_lu_12[0][0]             \n__________________________________________________________________________________________________\nconv2d_13 (Conv2D)              (None, None, None, 1 32768       add_4[0][0]                      \n__________________________________________________________________________________________________\n    batch_normalization_13 (BatchNo (None, None, None, 1 512         conv2d_13[0][0]                  \n    __________________________________________________________________________________________________\n    LAST LAYERS\n\n__________________________________________________________________________________________________\n    leaky_re_lu_72 (LeakyReLU)      (None, None, None, 2 0           batch_normalization_72[0][0]     \n    __________________________________________________________________________________________________\n    conv2d_59 (Conv2D)              (None, None, None, 2 261375      leaky_re_lu_58[0][0]             \n    __________________________________________________________________________________________________\n    conv2d_67 (Conv2D)              (None, None, None, 2 130815      leaky_re_lu_65[0][0]             \n    __________________________________________________________________________________________________\n    conv2d_75 (Conv2D)              (None, None, None, 2 65535       leaky_re_lu_72[0][0]             \n    ==================================================================================================\n    Total params: 62,001,757\n    Trainable params: 61,949,149\n    Non-trainable params: 52,608\n<\/code><\/pre>\n\n<p>and also <code>saved_model_cli show --dir 'serving\/1' --all<\/code> output is this,<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/1ybtQ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1ybtQ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I hosted this yolo model using tf-serving docker<\/p>\n\n<pre><code>docker run  --runtime= -t --rm -p 8900:8500 -p 8901:8501 -v \/home\/sp\/Desktop\/coco\/serving:\/models\/yolo -e MODEL_NAME=yolo tensorflow\/serving:latest\n<\/code><\/pre>\n\n<p>which is running succesfully<\/p>\n\n<p>now to test I wrote client code this<\/p>\n\n<p>this my client code to access the tf-server which is hosted at<\/p>\n\n<pre><code>def do_inference(hostport, work_dir, concurrency):\n    \"\"\"Tests PredictionService with concurrent requests.\n    Args:\n        hostport: Host:port address of the PredictionService.\n        work_dir: The full path of working directory for test data set.\n        concurrency: Maximum number of concurrent requests.\n        num_tests: Number of test images to use.\n    Returns:\n        The classification error rate.\n    Raises:\n        IOError: An error occurred processing test data set.\n    \"\"\"\n    channel = grpc.insecure_channel(hostport)\n    stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)\n    request = predict_pb2.PredictRequest()\n    request.model_spec.name = 'yolo'\n    request.model_spec.signature_name = 'serving_default'\n\n\n\n\n\n\n        x= cv.imread(\"test.jpg\")\n        request.inputs['input_image'].CopyFrom(tf.contrib.util.make_tensor_proto(x, shape=[1,1,416, 416]))\n        #result_counter.throttle()\n        result_future = stub.Predict.future(request, 10.25)  # 5 seconds  \n\n        # if things are wrong the callback will not come, so uncomment below to force the result\n        # or get to see what is the bug\n        res= result_future.result()\n        response = numpy.array(res.outputs['conv2d_75\/BiasAdd:0'].float_val)\n        prediction = numpy.argmax(response)\n        print(\"Predicted Result is \", prediction,\"Detection Porbability= \",response[prediction])\n<\/code><\/pre>\n\n<p>I run this code and getting error <code>ValueError: Too many elements provided. Needed at most 173056, but received 1367280<\/code><\/p>\n\n<p>And its pointing on line<\/p>\n\n<p><code>request.inputs['input_image'].CopyFrom(tf.contrib.util.make_tensor_proto(x, shape=[1,1,416, 416]))<\/code><\/p>\n","answers":[{"AnswerId":"57544263","CreationDate":"2019-08-18T11:50:23.940","ParentId":null,"OwnerUserId":"3375469","Title":null,"Body":"<p>It seems the input_image size is way larger the the specified shape (416, 416), may be you want to resize the input_image beforehand or change the shape to match. <\/p>\n"}]}
+{"QuestionId":57543881,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2019-08-18T10:58:19.093","AcceptedAnswerId":57729197.0,"OwnerUserId":11377572.0,"Title":"the shape of the output of keras.layers.concatenate","Body":"<p>I have a list of dense layers with same output shapes [batch, 1]. If I combine the outputs of these layers with keras.layers.concatenate(), what would the shape be?<\/p>\n\n<pre><code>dense_layers = [Dense(1), Dense(1), Dense(1)] #some dense layers\nmerged_output = keras.layers.concatenate([dense_layers])\n<\/code><\/pre>\n\n<p>Would the shape of merged_output be (batch, 3) or(3, 1)?<\/p>\n","answers":[{"AnswerId":"57729197","CreationDate":"2019-08-30T15:03:08.857","ParentId":null,"OwnerUserId":"10019464","Title":null,"Body":"<p>The answer is (batch,3). To see this, you can build a model and print model.summary():<\/p>\n\n<pre><code>from keras.layers import Input, Dense\nfrom keras.models import Model\nfrom keras.layers import concatenate\n\n batch = 30\n\n# define three sets of inputs\ninput1 = Input(shape=(batch,1))\ninput2 = Input(shape=(batch,1))\ninput3 = Input(shape=(batch,1))\n\n# define three dense layers\nlayer1 = Dense(1)(input1)\nlayer2 = Dense(1)(input2)\nlayer3 = Dense(1)(input3)\n\n# concatenate layers\ndense_layers = [layer1, layer2, layer3]\nmerged_output = concatenate(dense_layers)\n\n# create a model and check for output shape\nmodel = Model(inputs=[input1, input2, input3], outputs=merged_output)\nmodel.summary()\n<\/code><\/pre>\n\n<hr>\n\n<pre><code>Layer (type)                    Output Shape         Param #     Connected to                     \n\n=============================================================================\ninput_1 (InputLayer)            (None, 30, 1)        0                                            \n_______________________________________________________________________________\ninput_2 (InputLayer)            (None, 30, 1)        0                                            \n\n_______________________________________________________________________________\ninput_3 (InputLayer)            (None, 30, 1)        0                                            \n\n_______________________________________________________________________________\ndense_1 (Dense)                 (None, 30, 1)        2           input_1[0][0]                    \n_______________________________________________________________________________\ndense_2 (Dense)                 (None, 30, 1)        2           input_2[0][0]                    \n_______________________________________________________________________________\ndense_3 (Dense)                 (None, 30, 1)        2           input_3[0][0]                    \n\n_______________________________________________________________________________\nconcatenate_1 (Concatenate)     (None, 30, 3)        0           dense_1[0][0]                    \n                                                                 dense_2[0][0]                    \n                                                                 dense_3[0][0]                    \n\n==============================================================================\nTotal params: 6\nTrainable params: 6\nNon-trainable params: 0\n______________________________________________________________________________\n<\/code><\/pre>\n"}]}
+{"QuestionId":57543916,"AnswerCount":1,"Tags":"<python><pandas><machine-learning><keras><deep-learning>","CreationDate":"2019-08-18T11:04:47.637","AcceptedAnswerId":57544188.0,"OwnerUserId":11309950.0,"Title":"How can I implement the input of multiple regression in LSTM using keras?","Body":"<p>here is my code<\/p>\n\n<pre><code>def create_dataset(signal_data, look_back=1):\n    dataX, dataY = [], []\n    for i in range(len(signal_data) - look_back):\n        dataX.append(signal_data[i:(i + look_back), 0])\n        dataY.append(signal_data[i + look_back, 0])\n    return np.array(dataX), np.array(dataY)\n\ndf = pd.read_csv('time_series.csv')\nsignal_data = df.Close.values.astype('float32')\nsignal_data = signal_data.reshape(len(df), 1)\n\n\nscaler = MinMaxScaler(feature_range=(0, 1))\nsignal_data = scaler.fit_transform(signal_data)\n\ntrain_size = int(len(signal_data) * 0.80)\ntest_size = len(signal_data) - train_size)\n# val_size = len(signal_data) - train_size - test_size\ntrain = signal_data[0:train_size]\n# val = signal_data[train_size:train_size+val_size]\ntest = signal_data[train_size+val_size:len(signal_data)]\n\nx_train, y_train = create_dataset(train, look_back)\n# x_val, y_val = create_dataset(val, look_back)\nx_test, y_test = create_dataset(test, look_back)\n\n\nx_train = np.reshape(x_train, (x_train.shape[0], x_train.shape[1], 1))\n# x_val = np.reshape(x_val, (x_val.shape[0], x_val.shape[1], 1))\nx_test = np.reshape(x_test, (x_test.shape[0], x_test.shape[1], 1))\n<\/code><\/pre>\n\n<p>now I want to add <code>df.Open<\/code> and <code>df.High<\/code> and  <code>df.Low<\/code> and <code>df.Volume<\/code><\/p>\n\n<p>how can I implement this code?<\/p>\n\n<p>Should I just add to the signal data? I'm wondering how to add data so that I can train multiple features in the signal data.<\/p>\n\n<p>I don't know where and how to implement it. I need your help.<\/p>\n\n<p>Your valuable opinions and thoughts will be very much appreciated.<\/p>\n","answers":[{"AnswerId":"57544188","CreationDate":"2019-08-18T11:40:11.393","ParentId":null,"OwnerUserId":"3742686","Title":null,"Body":"<p>I made several modifications to your code. This should work. In summary:<\/p>\n\n<ul>\n<li>I got fixed the lines of code where you were barcoding the selection of the variable 0. Now, the target variable stands on the last position and the others in the previous ones<\/li>\n<li>I fixed the reshapes some of them were not needed and the others were fixed to keep all the dimensions<\/li>\n<li>I fixed the model input shape, now you have 5 variables instead of 1<\/li>\n<\/ul>\n\n<p>My general recommendations:<\/p>\n\n<ul>\n<li>I would not use <code>MinMaxScaler<\/code>, it is dangerous because a single outlier can disturb all your distribution. Instead, use <code>StandardScaler<\/code>. More info here: <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.StandardScaler.html\" rel=\"nofollow noreferrer\">http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.StandardScaler.html<\/a><\/li>\n<li>I would scale the data afterwards, when the <code>train_x<\/code>, <code>test_x<\/code> and their <code>y<\/code> respective counterparts are built. The reason why is because you are computing the statistics for scaling the data using the train and test set, i.e. future information. This is by all the means different to what you'll find when you try to run your code in a real situation. I.e. you'll have to scale the new data with past statistics. It is better to build a test set as close to the reality as possible.<\/li>\n<li>How do you know that your model is big enough to model your data? I would get rid of the dropouts and run the model to see if it can overfit the data. If the model can overfit to the train data, it means that the model is big enough and you can start regularising your model to enhance generalisation. More info in this book: <a href=\"https:\/\/www.deeplearning.ai\/machine-learning-yearning\/\" rel=\"nofollow noreferrer\">https:\/\/www.deeplearning.ai\/machine-learning-yearning\/<\/a><\/li>\n<li>In the model <code>metrics<\/code> you choose <code>accuracy<\/code>, which is a classification metric. I would use one according to my type of problem (regression): for example \"Mean Absolute Error\".<\/li>\n<\/ul>\n\n<p>I hope I managed to help you :D<\/p>\n\n<hr>\n\n<pre><code>import numpy as np\nimport pandas as pd\nfrom keras.models import Sequential\nfrom keras.layers import Dense, LSTM, Dropout, Conv2D, Reshape, TimeDistributed, Flatten, Conv1D,ConvLSTM2D, MaxPooling1D\nfrom keras.layers.core import Dense, Activation, Dropout\nfrom sklearn.preprocessing import MinMaxScaler\nfrom sklearn.metrics import mean_squared_error\nimport tensorflow as tf\nimport matplotlib.pyplot as plt\n\n\nconfig = tf.ConfigProto()\nconfig.gpu_options.allow_growth=True\n\nsess = tf.Session(config=config)\ndef create_dataset(signal_data, look_back=1):\n    dataX, dataY = [], []\n    for i in range(len(signal_data) - look_back):\n        dataX.append(signal_data[i:(i + look_back), :])\n        dataY.append(signal_data[i + look_back, -1])\n    return np.array(dataX), np.array(dataY)\n\n\nlook_back = 20\n\n\n\ndf = pd.read_csv('kospi.csv')\nsignal_data = df[[\"Open\", \"Low\", \"High\", \"Volume\", \"Close\"]].values.astype('float32')\n\n\nscaler = MinMaxScaler(feature_range=(0, 1))\nsignal_data = scaler.fit_transform(signal_data)\n\n\n\ntrain_size = int(len(signal_data) * 0.80)\ntest_size = len(signal_data) - train_size - int(len(signal_data) * 0.05)\nval_size = len(signal_data) - train_size - test_size\ntrain = signal_data[0:train_size]\nval = signal_data[train_size:train_size+val_size]\ntest = signal_data[train_size+val_size:len(signal_data)]\n\n\n\nx_train, y_train = create_dataset(train, look_back)\nx_val, y_val = create_dataset(val, look_back)\nx_test, y_test = create_dataset(test, look_back)\n\n\n\n\nmodel = Sequential()\nmodel.add(LSTM(128, input_shape=(None, 5),return_sequences=True))\nmodel.add(Dropout(0.3))\n\nmodel.add(LSTM(128, input_shape=(None, 5)))\nmodel.add(Dropout(0.3))\n\nmodel.add(Dense(128))\nmodel.add(Dropout(0.3))\n\nmodel.add(Dense(1))\n\n\n\n\nmodel.compile(loss='mean_squared_error', optimizer='adam', metrics=['accuracy'])\n\n\nmodel.summary()\nhist = model.fit(x_train, y_train, epochs=20, batch_size=32, verbose=2, validation_data=(x_val, y_val))\n\ntrainScore = model.evaluate(x_train, y_train, verbose=0)\nmodel.reset_states()\nprint('Train Score: ', trainScore)\nvalScore = model.evaluate(x_val, y_val, verbose=0)\nmodel.reset_states()\nprint('Validataion Score: ', valScore)\ntestScore = model.evaluate(x_test, y_test, verbose=0)\nmodel.reset_states()\nprint('Test Score: ', testScore)\n\n\n\np = model.predict(x_test)\n\n\nprint(mean_squared_error(y_test, p))\n\nimport matplotlib.pyplot as pplt\n\npplt.plot(y_test)\npplt.plot(p)\npplt.legend(['testY', 'p'], loc='upper right')\npplt.show()\n<\/code><\/pre>\n"}]}
+{"QuestionId":57544017,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2019-08-18T11:18:04.757","AcceptedAnswerId":null,"OwnerUserId":9850548.0,"Title":"How to fix 'Incompatible shapes' error in keras?","Body":"<p>I'm trying to make a code with <i>keras and tensorflow<\/i> and I want to use batch size.<\/p>\n\n<p>you can find the code in <a href=\"https:\/\/github.com\/aSafarpoor\/vgg_vidclassification.git\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aSafarpoor\/vgg_vidclassification.git<\/a><\/p>\n\n<p>and it can run with something like this:<\/p>\n\n<blockquote>\n  <p>python main.py --net resnet34s --batch_size 1 --gpu 2,3 --lr 0.001 --warmup_ratio 0.1 --optimizer adam --epochs 4 --multiprocess 1 --loss softmax --hdim 30 --wdim 30  --frame_num 80  --aggregation_mode vlad  --vlad_cluster 4<\/p>\n<\/blockquote>\n\n<p>when i set batch size==1 then it works,\nbut when i change it to another numbers it make an error like this:<\/p>\n\n<pre><code>Epoch 1\/2\nLearning rate for epoch 1 is 0.0001.\n\nTraceback (most recent call last):\n\n  File \"main.py\", line 214, in &lt;module&gt; main()\n\n  File \"main.py\", line 145, in main verbose=1)\n\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/legacy\/interfaces.py\", \nline 91, in wrapper\n\n    return func(*args, **kwargs)\n\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/engine\/training.py\", line \n\n1418, in fit_generator\n\n    initial_epoch=initial_epoch)\n\n  File \"\/usr\/local\/lib\/python2.7\/dist-\npackages\/keras\/engine\/training_generator.py\", line 217, in fit_generator\nclass_weight=class_weight)\n\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/keras\/engine\/training.py\", line \n1217, in train_on_batch\noutputs = self.train_function(ins)\n\n  File \"\/usr\/local\/lib\/python2.7\/dist-\npackages\/keras\/backend\/tensorflow_backend.py\", line 2715, in __call__\n    return self._call(inputs)\n\n  File \"\/usr\/local\/lib\/python2.7\/dist-\npackages\/keras\/backend\/tensorflow_backend.py\", line 2675, in _call\n    fetched = self._callable_fn(*array_vals)\n\n  File \"\/usr\/local\/lib\/python2.7\/dist-\npackages\/tensorflow\/python\/client\/session.py\", line 1458, in __call__\n    run_metadata_ptr)\n\ntensorflow.python.framework.errors_impl.InvalidArgumentError: Incompatible \nshapes: [2,5994] vs. [6,5994]\n     [[{{node \ntraining\/Adam\/gradients\/loss\/prediction_loss\/mul_grad\/BroadcastGradientArgs}}]\n]\n<\/code><\/pre>\n\n<p>and I run it with different version , error was the same ; like shapes: [2] vs [6]<\/p>\n\n<p>and the important part of error as you see is :<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: Incompatible \n    shapes: [2,5994] vs. [6,5994]\n         [[{{node \n  training\/Adam\/gradients\/loss\/prediction_loss\/mul_grad\/BroadcastGradientArgs}}]\n    ]\n<\/code><\/pre>\n\n<p>about code:<\/p>\n\n<p>at first most important parts are in src folder that is reachable in git<\/p>\n\n<p>this part is for  trainning:\nnetwork.fit_generator(trn_gen,<\/p>\n\n<pre><code>steps_per_epoch=int(len(partition['train'])\/\/args.batch_size),\n                          epochs=args.epochs,\n                          max_queue_size=10,\n                          callbacks=callbacks,\n                          use_multiprocessing=False,\n                          workers=1,\n                          verbose=1)\n<\/code><\/pre>\n\n<p>read videos like this in utils.py:<\/p>\n\n<pre><code>vidcap = cv2.VideoCapture(path)\nsuccess,image = vidcap.read()\nsuccess2=success\nwhile(success):\n    image= cv2.resize(image,(dim_h,dim_w))\n    sub_arr.append(image)\n    success,image = vidcap.read() \n    c+=1\nwhile(c&lt;spec_len and success2):\n    sub_arr.append(sub_arr[-1])\n    c+=1\n\n# if(len(sub_arr)&gt;0):\nbig_arr=sub_arr[:spec_len]\nnp_arr=np.asarray(big_arr)\n# print(np_arr.shape,\"       GgggG\")\nreturn np_arr\n<\/code><\/pre>\n\n<p>and then make data ready:<\/p>\n\n<pre><code>trnlist, trnlb = toolkits.get_voxceleb2_datalist(args, path='..\/meta\/voxlb2_train.txt')\nvallist, vallb = toolkits.get_voxceleb2_datalist(args, path='..\/meta\/voxlb2_val.txt')\npartition = {'train': trnlist.flatten(), 'val': vallist.flatten()}\n<\/code><\/pre>\n\n<p>model function is resnet_3D_v1 in the model.py,take a look:<\/p>\n\n<pre><code>bn_axis = 3\n    if mode == 'train':\n        inputs = Input(shape=input_dim, name='input')\n    else:\n        # inputs = Input(shape=(input_dim[0], None, input_dim[-1]), name='input')\n        inputs = Input(shape=input_dim, name='input')\n\n    # ===============================================\n    #            Convolution Block 1\n    # ===============================================\n    x1 = Conv3D(64, (1,1,1),\n                kernel_initializer='orthogonal',\n                use_bias=False, trainable=True,\n                kernel_regularizer=l2(weight_decay),\n                padding='same',\n                name='conv1_1\/3x3_s1')(inputs)\n\n    x1 = BatchNormalization(axis=bn_axis, name='conv1_1\/3x3_s1\/bn', trainable=True)(x1)\n    x1 = Activation('relu')(x1)\n    x1 = MaxPooling3D((1, 1,1), strides=(1, 1,1))(x1)\n\n    # ===============================================\n    #            Convolution Section 2\n    # ===============================================\n    x2 = conv_block_3D(x1, 3, [48, 48, 96], stage=2, block='a', strides=(1, 1,1), trainable=True)\n    x2 = identity_block_3D(x2, 3, [48, 48, 96], stage=2, block='b', trainable=True)\n\n    # ===============================================\n    #            Convolution Section 3\n    # ===============================================\n    x3 = conv_block_3D(x2, 3, [96, 96, 128], stage=3, block='a', trainable=True)\n    x3 = identity_block_3D(x3, 3, [96, 96, 128], stage=3, block='b', trainable=True)\n    x3 = identity_block_3D(x3, 3, [96, 96, 128], stage=3, block='c', trainable=True)\n    # ===============================================\n    #            Convolution Section 4\n    # ===============================================\n    x4 = conv_block_3D(x3, 3, [128, 128, 512], stage=4, block='a', trainable=True)\n    x4 = identity_block_3D(x4, 3, [128, 128, 512], stage=4, block='b', trainable=True)\n    x4 = identity_block_3D(x4, 3, [128, 128, 512], stage=4, block='c', trainable=True)\n    # ===============================================\n    #            Convolution Section 5\n    # ===============================================\n    x5 = conv_block_3D(x4, 3, [512,512, 2048], stage=5, block='a', trainable=True)\n    x5 = identity_block_3D(x5, 3, [512,512, 2048], stage=5, block='b', trainable=True)\n    x5 = identity_block_3D(x5, 3, [512,512, 2048], stage=5, block='c', trainable=True)\n    y = MaxPooling3D((1, 1,1), strides=(1, 1,1), name='mpool2')(x5)\n    return inputs, y\n<\/code><\/pre>\n\n<p>and  input is (frame_counter,w,h,rgb=3)<\/p>\n\n<p>and  it use model.py too.<\/p>\n","answers":[]}
+{"QuestionId":57544161,"AnswerCount":1,"Tags":"<python><pytorch><fast-ai>","CreationDate":"2019-08-18T11:35:45.570","AcceptedAnswerId":null,"OwnerUserId":10679907.0,"Title":"How to fix 'Expected object of backend CUDA' when trying to apply_tfms in custom LearnerCallback","Body":"<p>Very new to machine learning, fastai, pytorch, and python, and I was trying to adapt a LearnerCallback to do transformations after manually modifying the images.  When I start my learn.fit_one_cycle, it's immediately interrupted as shown below:<\/p>\n\n<p>I've tried sticking <code>.to(torch.device('cuda'))<\/code> everywhere I could think of<\/p>\n\n<pre><code>    #...\n    def on_batch_begin(self, last_input, last_target, train, **kwargs):\n        if not train: return\n\n        #Get new input\n        new_input = last_input.clone()\n        new_target = last_target.clone()\n        tfms = get_transforms(max_zoom=1.5)\n\n        # modify the images here in some other way\n\n        # apply_tfms\n        for i in range(len(new_input)):            \n            new_input[i] = Image(new_input[i]).apply_tfms(tfms[0]).data\n            new_target[i] = Image(new_target[i]).apply_tfms(tfms[0], do_resolve=False).data\n    #...\n<\/code><\/pre>\n\n<p>The 'apply_tfms' in the second to last line is the culprit in the traceback ending with:<\/p>\n\n<pre><code>    553     m[1,0] *= w\/h\n    554     c.flow = c.flow.view(-1,2)\n--&gt; 555     c.flow = torch.addmm(m[:2,2], c.flow,  m[:2,:2].t()).view(size)\n    556     return c\n    557 \n\nRuntimeError: Expected object of backend CUDA but got backend CPU for argument #4 'mat1'\n<\/code><\/pre>\n\n<p>Is there a way I can apply the transforms within a LearnerCallback without getting that error, or an alternate method where I can get my LearnerCallback added with <code>learn.callback_fns.append<\/code> to run before apply_tfms runs the same transforms on both the modified input and target images?  I need pixel information from the target image to modify the input image.  I also need this process applied during training and validation.<\/p>\n\n<p>If it makes any difference, I get the same error even if I don't modify the cloned images before apply_tfms.<\/p>\n","answers":[{"AnswerId":"57558923","CreationDate":"2019-08-19T14:36:10.773","ParentId":null,"OwnerUserId":"10511793","Title":null,"Body":"<p>The problem that you have here is that some of your tensors are on CPU and some are on GPU.<\/p>\n\n<p>You have to make sure that all your tensors are on the same device (either GPU or CPU depending on the situation) to get rid of this error. If I remember correctly, fastai <code>get_transforms<\/code> will create transforms that are supposed to be applied during data loading (and so on CPU), so you might want to take a look at the source of get_transforms and adapt it so it can be called on your tensor on GPU.(last_input should be on GPU if you're training on GPU).<\/p>\n\n<p>Another solution is to apply those transforms during dataloading, in which case your data tensor will be on CPU.<\/p>\n"}]}
+{"QuestionId":57544644,"AnswerCount":0,"Tags":"<python><tensorflow><jupyter-notebook><google-cloud-datalab>","CreationDate":"2019-08-18T12:38:42.357","AcceptedAnswerId":null,"OwnerUserId":11381180.0,"Title":"how to import datalab in local jupyter notebook and open Tensor Board","Body":"<blockquote>\n  <p>I receive syntax error on using following piece of code to try open Tensor Board from my local jupyter notebook<\/p>\n  \n  <p>I have installed datalab using <code>pip install datalab<\/code> in my local jupyter notebook<\/p>\n<\/blockquote>\n\n<pre><code>from google.datalab.ml import TensorBoard \nTensorBoard().start('taxi_trained')\nTensorBoard().list()\n<\/code><\/pre>\n\n<blockquote>\n  <p>Here's the error message:<\/p>\n<\/blockquote>\n\n<pre><code>Traceback (most recent call last):\n\nFile \"C:\\Users\\User\\Anaconda3\\lib\\site- \npackages\\IPython\\core\\interactiveshell.py\", line 3325, in run_code\nexec(code_obj, self.user_global_ns, self.user_ns)\n\nFile \"&lt;ipython-input-1-2e39f4ed6727&gt;\", line 1, in &lt;module&gt;\nfrom google.datalab.ml import TensorBoard\n\nFile \"C:\\Users\\User\\Anaconda3\\lib\\site- \npackages\\google\\datalab\\__init__.py\", line 13, in &lt;module&gt;\nfrom google.datalab._context import Context\n\nFile \"C:\\Users\\User\\Anaconda3\\lib\\site- \npackages\\google\\datalab\\_context.py\", line 20, in &lt;module&gt;\nfrom google.datalab.utils import _utils as du\n\nFile \"C:\\Users\\User\\Anaconda3\\lib\\site- \npackages\\google\\datalab\\utils\\__init__.py\", line 15\nfrom ._async import async, async_function, async_method\n                        ^\nSyntaxError: invalid syntax\n<\/code><\/pre>\n\n<blockquote>\n  <p>Any help with solution would be greatly appreciated<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":57544846,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorflow-datasets>","CreationDate":"2019-08-18T13:04:34.400","AcceptedAnswerId":57545354.0,"OwnerUserId":11839345.0,"Title":"How to include model.predict into my input pipeline?","Body":"<h2>The problem:<\/h2>\n\n<p>I'm building a video captioning seq2seq model and I've a problem with the input pipeline:<\/p>\n\n<p>I'm using a pre-trained <strong>InceptionV3<\/strong> model to preprocess all my data<\/p>\n\n<p>However, I've got a lot of data: 10.000 videos containing about several hundred frames each.<\/p>\n\n<p>When I'm using Inceptionv3 to preprocess my data, it returns me a <strong>numpy array with a very huge size<\/strong>. <\/p>\n\n<p>It's then impossible for me to create a <code>tf.data.Dataset<\/code> from this numpy array because its size, 6GiB, is way bigger than <strong>2GiB tensorflow's size limit<\/strong>.<\/p>\n\n<hr>\n\n<h2>My pipeline steps:<\/h2>\n\n<p><strong>Step 1:<\/strong><\/p>\n\n<p>Extracting videos and storing frames on the disk. Creating a <code>tf.data.Dataset<\/code> from this files.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>record_files, lowest_n_frames = build_tfrecord_dataset(videos_name, video_path_zip, record_file_path)\nrecords_dataset = tf.data.TFRecordDataset(record_files)\n<\/code><\/pre>\n\n<p><strong>Step 2:<\/strong><\/p>\n\n<p>Taking the frames from the TFRecord files created in step 1 and reformatting it for Inceptionv3.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>video_dataset = records_dataset.map(lambda tfrecord: decode_tfrecord(tfrecord, n_frames))\nvideo_dataset = video_dataset.map(format_video)\n<\/code><\/pre>\n\n<p><strong>Step 3:<\/strong><\/p>\n\n<p>Preprocessing the video frames using Inceptionv3. This returns me a huge numpy array.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>bottlenecks = image_features_extract_model.predict(video_dataset, verbose=1, steps=n_samples)\n<\/code><\/pre>\n\n<p><strong>Step 4 (Error step):<\/strong><\/p>\n\n<p>I'm creating a <code>tf.data.Dataset<\/code> from <code>bottlenecks<\/code>. Of course, it works on a small number of data.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>features_dataset = tf.data.Dataset.from_tensor_slices(bottlenecks)\n<\/code><\/pre>\n\n<hr>\n\n<p>The error is very simple and straightforward:<\/p>\n\n<pre><code>ValueError: Cannot create a tensor proto whose content is larger than 2GB.\n<\/code><\/pre>\n\n<hr>\n\n<h2>What I would like:<\/h2>\n\n<p>I'm wondering how turning the model.predict operation as a part of my input pipeline !<\/p>\n\n<p>I don't want to have a 6GiB numpy array in my RAM at a time, I want to do the preprocessing \"step by step\" over my data<\/p>\n","answers":[{"AnswerId":"57545354","CreationDate":"2019-08-18T14:12:26.577","ParentId":null,"OwnerUserId":"5368083","Title":null,"Body":"<p>When your data gets big, its smarter to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset#from_generator\" rel=\"nofollow noreferrer\">tf.data.Dataset.from_generator<\/a><\/p>\n\n<p>I don't know what your exact code and outputs look like but it should work if you try something like this<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>features_dataset = tf.data.Dataset.from_generator(iter(bottlenecks),\n                                                  output_types=...,\n                                                  output_shapes=...,)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57544927,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><keras>","CreationDate":"2019-08-18T13:15:10.343","AcceptedAnswerId":null,"OwnerUserId":10074708.0,"Title":"Framing Multiple Parallel Series problem for Keras","Body":"<p>I'm dealing with a multiple parallel time series problem (<a href=\"https:\/\/machinelearningmastery.com\/how-to-develop-lstm-models-for-time-series-forecasting\/\" rel=\"nofollow noreferrer\">see here<\/a>) and I have an issue regarding the non-constant number of sample per timestamp (not timestep!)<\/p>\n\n<p>My problem is the following:<\/p>\n\n<ul>\n<li>I have around 3000 stock for the entire timerange<\/li>\n<li>Each stock has around 100 features for each timestamp<\/li>\n<li>Stocks are filtered each timestamp according if they belong to an index or not<\/li>\n<li>They should be multi-classified at each time stamp<\/li>\n<li>At each time stamp, i have to predict the classes for next timestamp<\/li>\n<\/ul>\n\n<p>Basicaly my dataframe is multindexed dataframe of shape:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/aYOsR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/aYOsR.png\" alt=\"latex\"><\/a><\/p>\n\n<p>My framing of the problem is the following:<\/p>\n\n<ul>\n<li>use LSTM to capture temporal dependency<\/li>\n<li>define <code>stateful = True, timesteps=1<\/code> to predict next timestamp based on current timestamp only<\/li>\n<li>I was thinking to avoid online learning (ie. <code>batch_size =1<\/code>) as a sample per timestamp might contain 1000 stocks.<\/li>\n<\/ul>\n\n<p>Now my issue is that the number of stocks migh change per timestamp. I need then to work with dynamic batch size. I could have used padding an masking but it only works with timesteps (here 1).<\/p>\n\n<p>How could i proceed ? Thank you for any idea<\/p>\n","answers":[]}
+{"QuestionId":57545273,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-18T14:03:39.427","AcceptedAnswerId":null,"OwnerUserId":10843051.0,"Title":"Tensorflow 2.0 beta has no attribute keras","Body":"<p>Keras doesnt seems to be installed. I am using the Tensorflow 2.0 beta.<\/p>\n\n<p>I've already tried to reinstall Tensorflow. I've seen multiple similar threads, however none of them helped<\/p>\n\n<p>I get the error\n\"AttributeError: module 'tensorflow' has no attribute 'keras'\" when I try to access tf.keras. <\/p>\n\n<p>For Example:<\/p>\n\n<p>import tensorflow as tf<\/p>\n\n<p>mnist = tf.keras.datasets.mnist<\/p>\n","answers":[]}
+{"QuestionId":57545329,"AnswerCount":0,"Tags":"<python><tensorflow><tflearn>","CreationDate":"2019-08-18T14:09:01.917","AcceptedAnswerId":null,"OwnerUserId":8952590.0,"Title":"You must feed a value for placeholder tensor 'input_1\/X' with dtype float","Body":"<p>I have done a try on the cartPole problem from OpenAi and got a problem. \nThe problem is that when i run it with no previous model trained it works out great. \nBut when i try it out with (model=True) in main method. It tells me the error. <\/p>\n\n<p>I don't really know what to try.\nLooked everywhere for placeholders but can't find something that works.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import gym\nimport numpy as np\nimport random\nimport tflearn\nfrom tflearn.layers.core import input_data, dropout, fully_connected\nfrom tflearn.layers.estimator import regression\nfrom statistics import median, mean\nfrom collections import Counter\nimport pickle\nimport tensorflow as tf\nimport operator\n\ntf.logging.set_verbosity(tf.logging.ERROR)\n\nenv = gym.make('CartPole-v0')\ngoal_steps = 10000\n\n\n\ndef main() : \n    training_data = get_random_samples(model=False)\n    model = train_network(training_data)\n\n    #training_data = get_random_samples(model=True)\n    #model = train_network(training_data)\n\n    input(\"Press Enter to continue...\")\n    demo(model)\n\n\n\ndef get_random_samples(model=False) :\n    if model : \n        model = neural_model()\n        model.load(\"model.tfl\")\n    values  = 10\n    data    = []\n    scores  = []\n    games = []\n    for _ in range(values) : \n        game_data = []\n        previous_observation = []\n        score = 0\n        env.reset()\n        for t in range(1000) :\n            #env.render()\n\n            # Get 20% random moves to not overfit model.\n            chance = random.randrange(0,6)\n            if (chance == 0) : \n                action = random.randrange(0,2)\n                observation, reward, done, info = env.step(action)\n\n            else : \n                if (len(previous_observation)==0):\n                    action = random.randrange(0,2)\n                elif (model):\n                    action = np.argmax(model.predict(previous_observation.reshape(-1,len(previous_observation),1))[0])\n                else : \n                    action = random.randrange(0,2)\n\n                observation, reward, done, info = env.step(action)\n\n                if (len(previous_observation) &gt; 0) :\n                    game_data.append([previous_observation, action])\n\n                previous_observation = observation\n\n            if done :\n                score = t\n                break\n\n        scores.append(score)\n        if ( len(scores) % 100 == 1) : \n            print(\"generated {} values\".format(len(scores)))\n        games.append(game_data)\n\n    data = get_highest_score(games,scores)\n    data = clean_data(data)\n\n    scores.sort()\n    scores = scores[int(len(scores)*9\/10) : ]\n\n    print(\"Number of generated games: \", len(scores))\n    print('Average accepted score:', mean(scores))\n    print('Median score for accepted scores:', median(scores))\n    print(Counter(scores))\n\n    env.close()\n    np.save(\"random_sample.npy\", data)\n\n    return data\n\n\ndef neural_model() :\n    network = input_data(shape=[None, 4, 1], name='input')\n\n    network = fully_connected(network, 128, activation='relu')\n    network = dropout(network, 0.8)\n\n    network = fully_connected(network, 256, activation='relu')\n    network = dropout(network, 0.8)\n\n    network = fully_connected(network, 512, activation='relu')\n    network = dropout(network, 0.8)\n\n    network = fully_connected(network, 256, activation='relu')\n    network = dropout(network, 0.8)\n\n    network = fully_connected(network, 128, activation='relu')\n    network = dropout(network, 0.8)\n\n    network = fully_connected(network, 2, activation='softmax')\n    network = regression(network, optimizer='adam', learning_rate=1e-3, loss='categorical_crossentropy', name='targets')\n    model = tflearn.DNN(network, tensorboard_dir='log')\n\n    return model\n\n\ndef train_network(training_data) :\n    X = np.array([i[0] for i in training_data]).reshape(-1,len(training_data[0][0]),1)\n    y = [i[1] for i in training_data]\n\n\n    model = neural_model()\n\n    model.fit({'input': X}, {'targets': y}, n_epoch=5, snapshot_step=500, show_metric=True, run_id='openai_learning')\n\n    # Saving model\n    model.save(\"model.tfl\")\n\n    return model\n\n\ndef get_highest_score(data, scores) :\n    # Returns the top 20% of all scores generated\n    clean_data = []\n    list_of_tuples = list(zip(scores, data))\n    list_of_tuples.sort(key = operator.itemgetter(0))\n    list_of_tuples = list_of_tuples[int(len(list_of_tuples)*(9\/10)) : ]\n    clean_data = [ i[1] for i in list_of_tuples ]\n    return clean_data\n\n\ndef clean_data(data): \n    # Turns our action into a [0,1] or [1,0] and concatinate the result\n    clean_data = []\n    for game in data : \n        for d in game :\n            if (d[1] == 1):\n                output = [0 ,1]\n\n            elif d[1] == 0:\n                output = [1 ,0]\n\n\n            clean_data.append([d[0], output])\n    return clean_data\n\n\ndef demo(model) :\n    scores = []\n    choices = []\n    for each_game in range(5):\n        score = 0\n        game_memory = []\n        prev_obs = []\n        env.reset()\n        for _ in range(goal_steps):\n            env.render()\n\n            if len(prev_obs)==0:\n                action = random.randrange(0,2)\n            else:\n                action = np.argmax(model.predict(prev_obs.reshape(-1,len(prev_obs),1))[0])\n\n            choices.append(action)\n\n            new_observation, reward, done, info = env.step(action)\n            prev_obs = new_observation\n            game_memory.append([prev_obs, action])\n            score+=reward\n            if done: break\n\n        scores.append(score)\n\n    print('Average Score:',sum(scores)\/len(scores))\n    print('choice 1: {}  choice 0: {}'.format(choices.count(1)\/len(choices),choices.count(0)\/len(choices)))\n\n\nmain()\n\n<\/code><\/pre>\n\n<p>The expected result as i know would be that i use the prediction from the last model to generate new samples to train a new one. \nThis would yield in a better solution to my problem. <\/p>\n\n<p><strong>Error:<\/strong><\/p>\n\n<pre><code>InvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'input_1\/X' with dtype float\n     [[Node: input_1\/X = Placeholder[dtype=DT_FLOAT, shape=[], _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"]()]]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57545405,"AnswerCount":0,"Tags":"<python><ubuntu><tensorflow><cudnn>","CreationDate":"2019-08-18T14:18:08.927","AcceptedAnswerId":null,"OwnerUserId":8410477.0,"Title":"ImportError of libcusolver.so.10.0 and failed call to cuInit: CUDA_ERROR_NO_DEVICE: no CUDA-capable device is detected","Body":"<p>I'm trying to run this <a href=\"https:\/\/github.com\/chineseocr\/chineseocr\" rel=\"nofollow noreferrer\">ocr project<\/a>, but I was stuck at this step: <code>python tools\/pytorch_to_keras.py  -weights_path models\/ocr-dense.pth -output_path models\/ocr-dense-keras.h5<\/code>.<\/p>\n\n<p>I use <code>POP!_OS 18.04<\/code>, <code>tensorflow-gpu 1.13.1<\/code>,  I have tried many methods from including the links at the end, but not works. Thanks  for your helps.<\/p>\n\n<p><strong>Nvidia cuda version<\/strong> :<\/p>\n\n<pre><code>$ nvcc -V\nnvcc: NVIDIA (R) Cuda compiler driver\nCopyright (c) 2005-2018 NVIDIA Corporation\nBuilt on Sat_Aug_25_21:08:01_CDT_2018\nCuda compilation tools, release 10.0, V10.0.130\n<\/code><\/pre>\n\n<p><strong>Info of cudnn:<\/strong><\/p>\n\n<pre><code>$ cat \/usr\/local\/cuda\/include\/cudnn.h | grep CUDNN_MAJOR -A 2\n#define CUDNN_MAJOR 7\n#define CUDNN_MINOR 6\n#define CUDNN_PATCHLEVEL 2\n--\n#define CUDNN_VERSION (CUDNN_MAJOR * 1000 + CUDNN_MINOR * 100 + CUDNN_PATCHLEVEL)\n\n#include \"driver_types.h\"\n<\/code><\/pre>\n\n<p>Detailed infos:<\/p>\n\n<pre><code>x@pop-os:~\/chineseocr$ sudo python tools\/pytorch_to_keras.py  -weights_path models\/ocr-dense.pth -output_path models\/ocr-dense-keras.h5\nUsing TensorFlow backend.\nTraceback (most recent call last):\n  File \"\/home\/x\/.local\/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\/x\/.local\/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\/x\/.local\/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 \"\/usr\/lib\/python3.6\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/usr\/lib\/python3.6\/imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: libcusolver.so.10.0: 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 \"tools\/pytorch_to_keras.py\", line 85, in &lt;module&gt;\n    from crnn.network_keras import keras_crnn\n  File \"\/home\/x\/\u6587\u6863\/Deep_Learning\/NLP\/Chinese_OCR\/chineseocr\/crnn\/network_keras.py\", line 1, in &lt;module&gt;\n    from keras.layers import Conv2D,BatchNormalization,MaxPool2D,Input,Permute,Reshape,Dense,LeakyReLU,Activation\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/keras\/__init__.py\", line 3, in &lt;module&gt;\n    from . import utils\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/keras\/utils\/__init__.py\", line 6, in &lt;module&gt;\n    from . import conv_utils\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/keras\/utils\/conv_utils.py\", line 9, in &lt;module&gt;\n    from .. import backend as K\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/keras\/backend\/__init__.py\", line 89, in &lt;module&gt;\n    from .tensorflow_backend import *\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 5, in &lt;module&gt;\n    import tensorflow as tf\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/tensorflow\/__init__.py\", line 24, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"\/home\/x\/.local\/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\/x\/.local\/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\/x\/.local\/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\/x\/.local\/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\/x\/.local\/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 \"\/usr\/lib\/python3.6\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/usr\/lib\/python3.6\/imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: libcusolver.so.10.0: 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\/errors\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>While without <code>sudo<\/code>, I get:<\/p>\n\n<pre><code>x@pop-os:~\/chineseocr$ python tools\/pytorch_to_keras.py  -weights_path models\/ocr-dense.pth -output_path models\/ocr-dense-keras.h5\nUsing TensorFlow backend.\nWARNING:tensorflow:From \/home\/x\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/op_def_library.py:263: colocate_with (from tensorflow.python.framework.ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nColocations handled automatically by placer.\n2019-08-18 22:05:32.939546: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\n2019-08-18 22:05:32.946379: E tensorflow\/stream_executor\/cuda\/cuda_driver.cc:300] failed call to cuInit: CUDA_ERROR_NO_DEVICE: no CUDA-capable device is detected\n2019-08-18 22:05:32.946405: I tensorflow\/stream_executor\/cuda\/cuda_diagnostics.cc:161] retrieving CUDA diagnostic information for host: pop-os\n2019-08-18 22:05:32.946411: I tensorflow\/stream_executor\/cuda\/cuda_diagnostics.cc:168] hostname: pop-os\n2019-08-18 22:05:32.946440: I tensorflow\/stream_executor\/cuda\/cuda_diagnostics.cc:192] libcuda reported version is: 430.34.0\n2019-08-18 22:05:32.946463: I tensorflow\/stream_executor\/cuda\/cuda_diagnostics.cc:196] kernel reported version is: 430.34.0\n2019-08-18 22:05:32.946469: I tensorflow\/stream_executor\/cuda\/cuda_diagnostics.cc:303] kernel version seems to match DSO: 430.34.0\n2019-08-18 22:05:32.947811: I tensorflow\/core\/platform\/profile_utils\/cpu_utils.cc:94] CPU Frequency: 2208000000 Hz\n2019-08-18 22:05:32.948894: I tensorflow\/compiler\/xla\/service\/service.cc:150] XLA service 0x34cee30 executing computations on platform Host. Devices:\n2019-08-18 22:05:32.948908: I tensorflow\/compiler\/xla\/service\/service.cc:158]   StreamExecutor device (0): &lt;undefined&gt;, &lt;undefined&gt;\n<\/code><\/pre>\n\n<p>References related:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/27306\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/27306<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/55224016\/importerror-libcublas-so-10-0-cannot-open-shared-object-file-no-such-file-or\">ImportError: libcublas.so.10.0: cannot open shared object file: No such file or director<\/a><\/p>\n","answers":[]}
+{"QuestionId":57546710,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2019-08-18T17:11:05.863","AcceptedAnswerId":null,"OwnerUserId":5183539.0,"Title":"How to reshape a whole batch in Keras model?","Body":"<p>I'm trying to reshape a whole batch in Keras so that it will transform from (?, 28, 28, 1) to (?\/10, 10, 28, 28, 1). Is it possible to implement with Keras?<\/p>\n\n<p>I have a dataset where labels are per group of images (let's call it 'bag') but not per images themselves. I'd like to load images from folder using Keras <code>ImageDataGenerator<\/code> and its <code>flow_from_directory()<\/code> method and then reshape images inside the model itself.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.datasets import mnist\nimport keras.backend as K\nfrom keras.layers import Lambda\n\n(x_train, y_train), (x_test, y_test) = mnist.load_data()\n\ndef test_model(data):\n    from keras import Model\n    from keras.models import Input\n    input_layer = Input(shape=data.shape[1:])\n    layer = Lambda(lambda x: K.reshape(x,\n                                       (-1, 10, *data.shape[1:])),\n                   output_shape=(-1, 10, *data.shape[1:]))(input_layer)\n    model = Model(inputs=[input_layer], outputs=[layer])\n    model.compile(optimizer='adadelta', loss='mean_squared_error')\n    return model.predict(data, batch_size=60)\n\n\nif __name__ == '__main__':\n    test_model(x_train)\n\n<\/code><\/pre>\n\n<p>I expected the model to reshape each batch and return it somehow (to be honest, I don't know how Keras concatenates results of each batch when predicting).\nInstead of results, I get error:<\/p>\n\n<p><code>ValueError: could not broadcast input array from shape (6,10,28,28) into shape (60,10,28,28)<\/code><\/p>\n","answers":[]}
+{"QuestionId":57547036,"AnswerCount":1,"Tags":"<python><tensorflow><variable-assignment>","CreationDate":"2019-08-18T17:54:58.283","AcceptedAnswerId":57560195.0,"OwnerUserId":9480505.0,"Title":"How to fix 'strided slice assignment are not compatible with expected types' in Tensorflow 1.14.0rc in autograph?","Body":"<p>I'm trying to implement the Sliding DFT algorithm in tensorflow 1.14.0 and am using tf.function so that I dont have to worry to much about control flow, however I am running into a problem. When I attempt to assign an element of a variable with another element in that variable, I get an error in regards to incompatible types in strided slice assignments. <\/p>\n\n<p>I've tried using tf.scatter update, tf assign, and using just a typical slice assign, however none of these have worked.<\/p>\n\n<pre><code>\n    @tf.function\n    def sdft_func(self,input_tensor):\n        for i in range(self.N_t):\n        #retrieving variables so that I have direct access to it\n        #instead of getting access to the read tensor\n        _, _, self.in_s = self.get_variables()\n        last = self.in_s[self.N_t-1]\n        for j in range(self.N_t,0,-1):\n            _, _, self.in_s = self.get_variables()\n            val = self.in_s[j-1]\n            #The line below gives the error\n            self.in_s = self.in_s[j].assign(val)\n            print(self.in_s)\n<\/code><\/pre>\n\n<p>The error that I am getting is as follows:<\/p>\n\n<blockquote>\n  <p>TypeError: In op 'strided_slice_1\/_assign', input types ([tf.complex64, tf.int32, tf.int32, tf.int32, tf.complex64]) are not compatible with expected types ([tf.complex64_ref, tf.int32, tf.int32, tf.int32, tf.complex64])<\/p>\n<\/blockquote>\n\n<p>Thank you ahead of time!<\/p>\n","answers":[{"AnswerId":"57560195","CreationDate":"2019-08-19T15:55:32.487","ParentId":null,"OwnerUserId":"9480505","Title":null,"Body":"<p>It seems I narrowed down the problem as having to do with attempting to set tf.complex64 variables inside of tf.function functions. So in order to overcome this, I simply abstracted away the operation so that the variable setting was done outside of the tf.function function. See below for the solution:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\n    def sdft_func(self,input_tensor):\n\n        @tf.function\n        def func(input_tensor,N_t,in_s,coeffs,freqs):\n            in_s = tf.identity(in_s)\n            coeffs = tf.identity(coeffs)\n            freqs = tf.identity(freqs)\n\n\n            for i in range(N_t):\n                last = in_s[self.N_t-1]\n\n                in_s = in_s[:-1]\n                new_val = tf.expand_dims(tf.complex(input_tensor[i],\n                                tf.cast(0.0,dtype=tf.float32)),0)\n\n                in_s = tf.concat([new_val,in_s],axis=0)\n                    delta = in_s[0] - last\n\n                freqs_2 = tf.TensorArray(tf.complex64,size=self.N)\n\n                for j in range(self.N_t):\n                    freqs_2 = freqs_2.write(j,(freqs[j]+delta)*coeffs[j])\n                freqs = freqs_2.stack()\n                freqs.set_shape([self.N])\n\n            return freqs,in_s\n\n        new_freqs, new_in_s = func(input_tensor,self.N_t,\n                self.in_s,self.coeffs,self.freqs)\n\n        self.in_s = self.in_s.assign(new_in_s)\n        self.freqs = self.freqs.assign(new_freqs)\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":57547195,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-18T18:14:17.503","AcceptedAnswerId":null,"OwnerUserId":7205641.0,"Title":"How can I feed the previous predicted output yhat(t-1) together with x(t) for computing yhat(t) in keras","Body":"<p>Currently I am using the standard LSTM layer in keras to map a sequence to a sequence. See below. <\/p>\n\n<pre><code>for i in range(self.num_hl):\n\n    inner2 = keras.layers.LSTM(self.hl1, dropout=self.dropout, kernel_initializer=self.kernel_init,\n                                           return_sequences=True, kernel_regularizer=self.l2\n                                           )(inner2)\n<\/code><\/pre>\n\n<p>However I would like to add the previous output to the current observation to make the new prediction. Hoping this way the model becomes auto-regressive.\nCan somebody help me implement this in Keras?<\/p>\n","answers":[]}
+{"QuestionId":57547293,"AnswerCount":0,"Tags":"<python><machine-learning><keras>","CreationDate":"2019-08-18T18:27:06.737","AcceptedAnswerId":null,"OwnerUserId":11365015.0,"Title":"ValueError: Error when checking target: expected dense_1 to have shape (16, 1, 2) but got array with shape (2, 1, 1)","Body":"<p>I'm very new to using machine learning and keras. This is my first CNN model. I built a model to classify the sounds into the UrbanSound8kDataset. When I run the model on a small dummy set, I keep getting the following error: <\/p>\n\n<pre><code>ValueError: Error when checking target: expected dense_1 to have shape (16, 1, 2) but got array with shape (2, 1, 1)\n<\/code><\/pre>\n\n<p>I looked at other posts with similar issues and tried changing the model loss to sparse_categorical_crossentropy rather than categorical_crossentropy. I also added model.flatten() before the first dense layer. Unfortunately, none of that resolved my error. <\/p>\n\n<p>Here is the relevant code for my model: <\/p>\n\n<pre><code>num_labels = yy.shape[1]\nfilter_size = 2\nnum_rows = 1\nnum_columns = 1\nnum_channels = 1    \n\nx_train = x_train.reshape(x_train.shape[0], num_rows, num_columns, num_channels)\ny_train = y_train.reshape(y_train.shape[0], 2, num_columns, num_channels)\nx_test = x_test.reshape(x_test.shape[0], 1, num_columns, num_channels)\ny_test = y_test.reshape(y_test.shape[0], 2, num_columns, num_channels)\n\nmodel = Sequential()\nmodel.add(Conv2D(filters=16, kernel_size=2, input_shape=(num_rows,\n\n    num_columns, num_channels), activation='relu', data_format='channels_first',   padding=\"same\"))\nmodel.add(MaxPooling2D(pool_size=1))\nmodel.add(Dropout(0.2))\n\nmodel.add(Conv2D(filters=32, kernel_size=1, activation='relu'))\nmodel.add(MaxPooling2D(pool_size=1))\nmodel.add(Dropout(0.2))\n\nmodel.add(Conv2D(filters=64, kernel_size=1, activation='relu'))\nmodel.add(MaxPooling2D(pool_size=1))\nmodel.add(Dropout(0.2))\n\nmodel.add(Conv2D(filters=64, kernel_size=1, activation='relu'))\nmodel.add(MaxPooling2D(pool_size=1))\nmodel.add(Dropout(0.2))\n\n#Flatten layer here:\n\nmodel.add(Dense(num_labels, activation='softmax'))\nmodel.compile(loss='categorical_crossentropy', metrics=['accuracy'], optimizer='adam')\n\nmodel.summary()\n\n\nnum_epochs = 100\nnum_batch_size = 132\n\nstart = datetime.now()\n\n\n\nmodel.fit(x_train, y_train, batch_size=num_batch_size, epochs=num_epochs, validation_data=(x_test, y_test), verbose=1)\n<\/code><\/pre>\n\n<p>Here is the model summary: <\/p>\n\n<pre><code>Layer (type)                 Output Shape              Param #   \n=================================================================\nconv2d_1 (Conv2D)            (None, 16, 1, 1)          80        \n_________________________________________________________________\nmax_pooling2d_1 (MaxPooling2 (None, 16, 1, 1)          0         \n_________________________________________________________________\ndropout_1 (Dropout)          (None, 16, 1, 1)          0         \n_________________________________________________________________\nconv2d_2 (Conv2D)            (None, 16, 1, 32)         64        \n_________________________________________________________________\nmax_pooling2d_2 (MaxPooling2 (None, 16, 1, 32)         0         \n_________________________________________________________________\ndropout_2 (Dropout)          (None, 16, 1, 32)         0         \n_________________________________________________________________\nconv2d_3 (Conv2D)            (None, 16, 1, 64)         2112      \n_________________________________________________________________\nmax_pooling2d_3 (MaxPooling2 (None, 16, 1, 64)         0         \n_________________________________________________________________\ndropout_3 (Dropout)          (None, 16, 1, 64)         0         \n_________________________________________________________________\nconv2d_4 (Conv2D)            (None, 16, 1, 64)         4160      \n_________________________________________________________________\nmax_pooling2d_4 (MaxPooling2 (None, 16, 1, 64)         0         \n_________________________________________________________________\ndropout_4 (Dropout)          (None, 16, 1, 64)         0         \n_________________________________________________________________\ndense_1 (Dense)              (None, 16, 1, 2)          130       \n=================================================================\nTotal params: 6,546\nTrainable params: 6,546\nNon-trainable params: 0\n____________________________\n<\/code><\/pre>\n\n<p>The error occurs in the last line of the code I concluded, the model.fit line. <\/p>\n\n<p>Please let me know if I need to include more code. This is only my second time posting here, so I was not sure how much code to include. <\/p>\n\n<p>Thank you very much! <\/p>\n","answers":[]}
+{"QuestionId":57547331,"AnswerCount":1,"Tags":"<keras><functional-programming><dropout>","CreationDate":"2019-08-18T18:31:43.787","AcceptedAnswerId":null,"OwnerUserId":1294691.0,"Title":"Question about enabling\/disabling dropout with keras functional API","Body":"<p>I am using Keras functional API to build a classifier and I am using the <code>training<\/code> flag in the dropout layer to enable dropout when predicting new instances (in order to get an estimate of the uncertainty). In order to get the expected response one needs to repeat this prediction several times, with keras randomly activating links in the dense layer, and of course it is computational expensive. Therefore, I would also like to have the option to not use dropout at the prediction phase, i.e., use all the network links. Does anyone know how I can do this? Following is a sample code of what I am doing. I tried to look if <code>predict<\/code> has any relevant parameter but does not seem like it does (?). I can technically train the same model without the training flag at the dropout layer, but I do not want to do this (or better I want to have a more clean solution, rather than having 2 different models). <\/p>\n\n<pre><code>from sklearn.datasets import make_circles\nfrom keras.models import Sequential\nfrom keras.utils import to_categorical\nfrom keras.layers import Dense\nfrom keras.layers import Dropout\nimport numpy as np\nimport keras\n\n# generate a 2d classification sample dataset\nX, y = make_circles(n_samples=100, noise=0.1, random_state=1)\nn_train = 30\ntrainX, testX = X[:n_train, :], X[n_train:, :]\ntrainy, testy = y[:n_train], y[n_train:]\ntrainy = to_categorical(trainy)\ntesty = to_categorical(testy)\n\ninputlayer = keras.layers.Input((2,))\n\nd = keras.layers.Dense(500, activation = 'relu')(inputlayer)\nd1 = keras.layers.Dropout(rate = .3)(d,training = True)\nout = keras.layers.Dense(2, activation = 'softmax')(d1)\n\nmodel = keras.Model(inputs = inputlayer, outputs = out)\n\nmodel.compile(loss = 'categorical_crossentropy',metrics = ['accuracy'],optimizer='adam')\nmodel.fit(x = trainX, y = trainy, validation_data=(testX, testy),epochs=1000, verbose=1)\n\n# another prediction on a specific sample \nprint(model.predict(testX[0:1,:]))\n# another prediction on the same sample\nprint(model.predict(testX[0:1,:]))\n<\/code><\/pre>\n\n<p>Running the above example I get the following output:<\/p>\n\n<pre><code>[[0.9230819  0.07691813]]\n[[0.8222245  0.17777553]]\n<\/code><\/pre>\n\n<p>which is as expected, different class probabilities for the same input, since there is a random (de)activation of the links from the dropout layer. <\/p>\n\n<p>Any suggestions on how I can enable\/disable dropout at the prediction phase with the functional API?<\/p>\n","answers":[{"AnswerId":"57547434","CreationDate":"2019-08-18T18:47:18.423","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Sure, you do not need to set the training flag when building the <code>Dropout<\/code> layer. After training your model you define this function:<\/p>\n\n<pre><code>mc_func = K.function([model.input, K.learning_phase()],\n                     [model.output])\n<\/code><\/pre>\n\n<p>Then you call <code>mc_func<\/code> with your input and flag 1 to enable dropout, or 0 to disable it:<\/p>\n\n<pre><code>stochastic_pred = mc_func([some_input, 1])\ndeterministic_pred = mc_func([some_input, 0])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57547419,"AnswerCount":1,"Tags":"<python-3.x><machine-learning><keras>","CreationDate":"2019-08-18T18:44:44.353","AcceptedAnswerId":57553571.0,"OwnerUserId":7276121.0,"Title":"StandardScaler with exported Keras model","Body":"<p>I just exported my keras model because I want to use it for my web application. For classification I'm using features from single audio files. The problem ist that with each prediction I only have one row, what makes the output of the StandardScaler 0. Is there any way to export the sklearn objekt that I can re use it in my webapp instead of using all the data from when I trained the model?<\/p>\n\n<p>Best regards<\/p>\n","answers":[{"AnswerId":"57553571","CreationDate":"2019-08-19T09:07:03.553","ParentId":null,"OwnerUserId":"7345564","Title":null,"Body":"<p>I haven't used the webapp before, but I can save the trained scaler with joblib and then load it in the future.<\/p>\n\n<pre><code>from sklearn.preprocessing import StandardScaler\nimport joblib\nfrom sklearn.datasets import load_breast_cancer\n\ndata = load_breast_cancer()\n\ndata = data.data\n\nscaler = StandardScaler()\n\nscaler = scaler.fit(data)\n\ndata_t = scaler.transform(data)\n#saves the trained scaler\njoblib.dump(scaler,r\"scaler\")\n\n#Loads for transforming the new  data\nscaler_2 = joblib.load(r\"scaler\")\n\n\ndata_t_2 = scaler.transform(data)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57547471,"AnswerCount":1,"Tags":"<pytorch><tensorboard>","CreationDate":"2019-08-18T18:53:22.033","AcceptedAnswerId":57547815.0,"OwnerUserId":10583715.0,"Title":"Does torch.utils.tensorboard need installation of Tensorflow?","Body":"<p>I'm new to pytorch and I wonder if using tensorboard on pytorch needs tensorflow as a dependency. Moreover, except for tensorboard, what are other options for training curve visualization?<\/p>\n","answers":[{"AnswerId":"57547815","CreationDate":"2019-08-18T19:41:02.023","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Indeed tensorboard is part of tensorflow, but it does not depend on it.<br>\nYou can safely install only tensorboard (e.g, <a href=\"https:\/\/pypi.org\/project\/tensorboard\/\" rel=\"nofollow noreferrer\"><code>pip install tensorboard<\/code><\/a>) or <a href=\"https:\/\/anaconda.org\/conda-forge\/tensorboard\" rel=\"nofollow noreferrer\"><code>conda install -c conda-forge tensorboard<\/code><\/a>) and use <code>torch.utils.tensorboard<\/code>.<\/p>\n"}]}
+{"QuestionId":57547984,"AnswerCount":0,"Tags":"<python><keras><time-series><sparse-matrix><embedding>","CreationDate":"2019-08-18T20:05:15.630","AcceptedAnswerId":null,"OwnerUserId":6196696.0,"Title":"setting up sequence categorical embedding layer for sparse time series problem on Keras","Body":"<p>I'm very new to using DL and keras, so I'd greatly appreciate any assistance on how I should construct my categorical variable\/embedding layers for the following problem:<\/p>\n\n<p>I'm currently trying to predict whether a customer will shop in the next 2 weeks using their transaction history of the items they've bought on a weekly basis.<\/p>\n\n<p>I'm considering transaction data for around 2 million customers for the past 134 weeks and around 4000 products(denoted as \"item_{}\" below). Basically, if they bought something that week then target_variable=1, else 0<\/p>\n\n<p>My dataframe <code>df<\/code> looks something like following:<\/p>\n\n<pre><code>week_yr   customer_id    items_bought          price          target_variable\n201701    123            [0]                   [0]               0\n201702    123            [0]                   [0]               0\n201703    123            [item_a,item_b]       [1.50,2.39]       1\n201704    123            [0]                   [0]               0\n201705    123            [0]                   [0]               0\n                                       .\n                                       .\n                                       .  \n\n201939    123           [item_a,item_d,item_g] [1.50,3.70,2.00]  1 \n201701    456           [0]                    [0]               0\n<\/code><\/pre>\n\n<p>I then used a labelencoder on customer_id, multilabelbinarizer on items_bought,scaled my price feature values and set my date feature as my index.<\/p>\n\n<p><em><strong>As far as my categorical variables, am I supposed to explode my <code>items_bought<\/code> feature(since there can be multiple items bought per week) and then run it through an embedding layer? Would I need to pad my item and price lists? Moreover, given that the data is sparse(most customers do not shop on a weekly basis), is there another way I should be structuring the data?<\/strong><\/em><\/p>\n\n<p>After applying encoding and scaling my data, here is how I'm feeding in the data into keras:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>cat_vars=['customer_id','items_bought']\ncont_vars=['price']\n\ntotal_cust_id=len(set(df['customer_id']))\ntotal_items=len(set(list(itertools.chain.from_iterable(df['items_bought']))))\n\ncat_sizes =[('customer_id',total_cust_id),('items_bought',total_items)]\nembedding_sizes = [(c, min(50, (c + 1) \/\/ 2)) for _, c in cat_sizes]  \n\nval = df[df['week_yr'] &gt;= 201937].set_index('week_yr')\ntrain = df[df['week_yr'] &lt; 201937].set_index('week_yr')\n\nX = train[cat_vars + cont_vars].copy()  \nX_val = val[cat_vars + cont_vars].copy()\n\ny = train['target_variable'].copy()  \ny_val = val['target_variable'].copy()\n\n\n\nX_array = []  \nX_val_array = []\n\nfor i, v in enumerate(cat_vars):  \n    X_array.append(X.iloc[:, i])\n    X_val_array.append(X_val.iloc[:, i])\n<\/code><\/pre>\n\n<p>since the data is sparse, I adjusted my class imbalance by doing the following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from sklearn.utils import class_weight\nclass_weights = class_weight.compute_class_weight('balanced',np.unique(y), y)\n<\/code><\/pre>\n\n<p>and then tried running the following model but kept running into multiple errors such as <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 3 array(s), but instead got the following list of 2 arrays:<\/code> and I'm not sure if it is because my categorical and continuous variables are set up as list?:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras import backend as K  \nfrom keras import regularizers  \nfrom keras.models import Sequential  \nfrom keras.models import Model  \nfrom keras.layers import Activation, BatchNormalization, Concatenate, Flatten  \nfrom keras.layers import Dropout, Dense, Input, Reshape,LSTM,merge  \nfrom keras.layers.embeddings import Embedding  \nfrom keras.optimizers import Adam  \nfrom keras.callbacks import ModelCheckpoint, ReduceLROnPlateau\n\n\n\ndef EmbeddingNet(cat_vars, cont_vars, embedding_sizes):  \n    inputs = []\n    embed_layers = []\n    for (c, (in_size, out_size)) in zip(cat_vars, embedding_sizes):\n        i = Input(shape=(1,))\n        o = Embedding(in_size, out_size, name=c)(i)\n        o = Reshape(target_shape=(out_size,))(o)\n        inputs.append(i)\n        embed_layers.append(o)\n\n    embed = Concatenate()(embed_layers)\n    embed = Dropout(0.04)(embed)\n\n    cont_input = Input(shape=(len(cont_vars),))\n    inputs.append(cont_input)\n\n    x = Concatenate()([embed, cont_input])\n\n    x = Dense(1000, kernel_initializer='he_normal')(x)\n    x = Activation('relu')(x)\n    x = Dropout(0.1)(x)\n\n    x = Dense(500, kernel_initializer='he_normal')(x)\n    x = Activation('relu')(x)\n    x = Dropout(0.1)(x)\n\n    x = Dense(1, kernel_initializer='he_normal')(x)\n    x = Activation('sigmoid')(x)\n\n    model = Model(inputs=inputs, outputs=x)\n    opt = Adam(lr=0.001)\n    model.compile(loss='binary_crossentropy', optimizer=opt, metrics=['accuracy'])\n\n    return model\n\n\nmodel = EmbeddingNet(cat_vars, cont_vars, embedding_sizes) \nlr_reducer = ReduceLROnPlateau(monitor='val_loss', factor=0.2,\n                               patience=3, verbose=1, mode='auto',\n                               min_delta=10, cooldown=0,\n                               min_lr=0.0001)\ncheckpoint = ModelCheckpoint('best_model_weights.hdf5',\n                             monitor='val_loss',\n                             save_best_only=True)\nhistory = model.fit(x=X_array, y=y, batch_size=128, epochs=20,\n                    verbose=1, callbacks=[lr_reducer, checkpoint],\n                    validation_data=(X_val_array, y_val),\n                    shuffle=True)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57548063,"AnswerCount":0,"Tags":"<tensorflow><keras><tensorboard>","CreationDate":"2019-08-18T20:15:37.150","AcceptedAnswerId":null,"OwnerUserId":22996.0,"Title":"How to generate a single tensorboard log file to display images, losses and graph?","Body":"<p>I have successfully created a callback class to save ground truth and prediction images by following <a href=\"https:\/\/stackoverflow.com\/questions\/43784921\/how-to-display-custom-images-in-tensorboard-using-keras?rq=1\">these answers<\/a><\/p>\n\n<p>Additionally, I would like to log losses. In order to do this, I have added the following callback:<\/p>\n\n<pre><code>cb_tensorboard = TensorBoard(log_dir=log_dir, histogram_freq=1, write_graph=True, write_images=False)\n<\/code><\/pre>\n\n<p>Then I passed both callbacks to fit_generator function:<\/p>\n\n<pre><code>history = model.fit_generator(\n            encoding.get_generator(train_gen),\n            validation_data=encoding.get_generator(test_gen),\n            steps_per_epoch=steps_per_epoch,\n            epochs=5000,\n            validation_steps=validation_steps,\n            verbose=2,\n            use_multiprocessing=True,\n            callbacks=[callback_save_images, cb_tensorboard]\n        )\n<\/code><\/pre>\n\n<p>The problem, however, is that this way two log files are generated and if I open tensorboard by running the following command, either losses or images are visible:<\/p>\n\n<pre><code>tensorboard --logdir=log\n<\/code><\/pre>\n\n<p>Does anyone know what is the correct way to implement the callback to save images, losses and graph in the same log file?<\/p>\n","answers":[]}
+{"QuestionId":57548095,"AnswerCount":0,"Tags":"<tensorflow><keras><keras-layer>","CreationDate":"2019-08-18T20:19:33.843","AcceptedAnswerId":null,"OwnerUserId":10112379.0,"Title":"How to add a custom keras layer in a model","Body":"<p>I have coded a custom layer in keras that just does a dot product of two vector as follow:<\/p>\n\n<pre><code>from keras import backend as K\nfrom keras.layers import Layer\n\nclass G_MAX_AVP_P(Layer):\n\n    def __init__(self,  **kwargs):\n        #self.output_dim = input_shape[1]*input_shape[1]\n        super(G_MAX_AVP_P, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        super(G_MAX_AVP_P, self).build(input_shape)  # Be sure to call this at the end\n\n    def call(self, x):\n        return K.reshape(K.dot(K.transpose(x),x), [-1])\n\n    def compute_output_shape(self, input_shape):\n        return (input_shape[0], self.input_shape[0]*self.input_shape[0])\n\n<\/code><\/pre>\n\n<p>I added it to my model like this:<\/p>\n\n<pre><code>class MULTI_ACTIVATION:\n    @staticmethod\n    def build(input_shape,classes):\n        # we initialize the model\n        input_tensor = Input(shape=input_shape)\n        #G_MAX_AVP_P = G_MAX_AVP_P(input_shape)\n\n        # Conv Block 1\n        x = Conv2D(64, (8, 8), input_shape=input_shape, padding='same')(input_tensor)\n        x = BatchNormalization()(x)\n        x = Activation('relu')(x)\n        x = Conv2D(64, (8, 8), padding='same')(x)\n        x = BatchNormalization()(x)\n        x = Activation('relu')(x)\n        x = Conv2D(64, (8, 8), padding='same')(x)\n        x = BatchNormalization()(x)\n        x = Activation('relu')(x)\n\n        # Conv Block 2\n        x = Conv2D(128, (5, 5), padding='same')(x)\n        x = BatchNormalization()(x)\n        x = LeakyReLU(alpha=0.2)(x)\n\n        # Conv Block 3\n        x = Conv2D(256, (3, 3), padding='same')(x)\n        x = BatchNormalization()(x)\n        x = PReLU()(x)\n\n        # Conv Block 4\n        x = Conv2D(128, (5, 5), padding='same')(x)\n        x = BatchNormalization()(x)\n        x = ELU(alpha=0.3)(x)\n\n        x = GlobalAveragePooling2D()(x)\n\n\n        x = G_MAX_AVP_P()(x)\n\n        output_tensor = Dense(units=classes, activation='softmax')(x)\n        model = Model(input_tensor, output_tensor)\n        model.summary()\n\n        return model\n<\/code><\/pre>\n\n<p>But I'm getting this error, I didn't get it<\/p>\n\n<p>def compute_output_shape(self, input_shape):\n---> 19         return (input_shape[0], self.input_shape[0]*self.input_shape[0])<\/p>\n\n<p>AttributeError: The layer has never been called and thus has no defined input shape.<\/p>\n","answers":[]}
+{"QuestionId":57548180,"AnswerCount":1,"Tags":"<python><nlp><pytorch>","CreationDate":"2019-08-18T20:33:25.193","AcceptedAnswerId":57550753.0,"OwnerUserId":10255067.0,"Title":"Filling torch tensor with zeros after certain index","Body":"<p>Given a 3d tenzor, say:\n<code>batch x sentence length x embedding dim<\/code><\/p>\n\n<pre><code>a = torch.rand((10, 1000, 96)) \n<\/code><\/pre>\n\n<p>and an array(or tensor) of actual lengths for each sentence<\/p>\n\n<pre><code>lengths =  torch .randint(1000,(10,))\n<\/code><\/pre>\n\n<p><code>outputs tensor([ 370., 502., 652., 859., 545., 964., 566., 576.,1000., 803.])<\/code><\/p>\n\n<p>How to fill tensor \u2018a\u2019 with zeros after certain index along dimension 1 (sentence length) according to tensor \u2018lengths\u2019 ?<\/p>\n\n<p>I want smth like that :<\/p>\n\n<pre><code>a[ : , lengths : , : ]  = 0\n<\/code><\/pre>\n\n<p>One way of doing it (slow if batch size is big enough):<\/p>\n\n<pre><code>for i_batch in range(10):\n    a[ i_batch  , lengths[i_batch ] : , : ]  = 0\n<\/code><\/pre>\n","answers":[{"AnswerId":"57550753","CreationDate":"2019-08-19T05:07:00.817","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can do it using a binary mask.<br>\nUsing <code>lengths<\/code> as column-indices to <code>mask<\/code> we indicate where each sequence ends (note that we make <code>mask<\/code> longer than <code>a.size(1)<\/code> to allow for sequences with full length).<br>\nUsing <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/torch.html#torch.cumsum\" rel=\"nofollow noreferrer\"><code>cumsum()<\/code><\/a> we set all entries in <code>mask<\/code> after the seq len to 1.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>mask = torch.zeros(a.shape[0], a.shape[1] + 1, dtype=a.dtype, device=a.device)\nmask[(torch.arange(a.shape[0], lengths)] = 1\nmask = mask.cumsum(dim=1)[:, :-1]  # remove the superfluous column\na = a * (1. - mask[..., None])     # use mask to zero after each column\n<\/code><\/pre>\n\n<p>For <code>a.shape = (10, 5, 96)<\/code>, and <code>lengths = [1, 2, 1, 1, 3, 0, 4, 4, 1, 3]<\/code>.<br>\nAssigning 1 to respective <code>lengths<\/code> at each row, <code>mask<\/code> looks like:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>mask = \ntensor([[0., 1., 0., 0., 0., 0.],\n        [0., 0., 1., 0., 0., 0.],\n        [0., 1., 0., 0., 0., 0.],\n        [0., 1., 0., 0., 0., 0.],\n        [0., 0., 0., 1., 0., 0.],\n        [1., 0., 0., 0., 0., 0.],\n        [0., 0., 0., 0., 1., 0.],\n        [0., 0., 0., 0., 1., 0.],\n        [0., 1., 0., 0., 0., 0.],\n        [0., 0., 0., 1., 0., 0.]])\n<\/code><\/pre>\n\n<p>After <code>cumsum<\/code> you get<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>mask = \ntensor([[0., 1., 1., 1., 1.],\n        [0., 0., 1., 1., 1.],\n        [0., 1., 1., 1., 1.],\n        [0., 1., 1., 1., 1.],\n        [0., 0., 0., 1., 1.],\n        [1., 1., 1., 1., 1.],\n        [0., 0., 0., 0., 1.],\n        [0., 0., 0., 0., 1.],\n        [0., 1., 1., 1., 1.],\n        [0., 0., 0., 1., 1.]])\n<\/code><\/pre>\n\n<p>Note that it exactly has zeros where the valid sequence entries are and ones beyond the lengths of the sequences. Taking <code>1 - mask<\/code> gives you exactly what you want.<\/p>\n\n<p>Enjoy ;)<\/p>\n"}]}
+{"QuestionId":57548363,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-08-18T21:04:47.857","AcceptedAnswerId":null,"OwnerUserId":11274618.0,"Title":"Setting inital weight matrices in in tensorflow's SimpleRNN","Body":"<p>I am trying to hard-code a RNN that keeps a running total of the numbers input to it, but I'm getting weird results.  I think I am not setting the weights correctly.<\/p>\n\n<p>My network consists of a single SimpleRNN layer defined (hopefully!) such that the new output is always 1*(new input) + 1*(old output):<\/p>\n\n<pre><code>rnn = tf.keras.layers.SimpleRNN(\n    1                 # Cell size = 1: only need to store a running total\n    stateful = True   # Don't forget the total between batches \n    use_bias = False\n    kernel_initializer = tf.ones,    # coefficient of new input\n    recurrent_initializer = tf.ones, # coefficient of previous output\n    activation=None,                 # output a linear combination of new input and previous output\n    )\n<\/code><\/pre>\n\n<p>As I understand it, if I don't train this network at all, it should already be keeping a running tally of the inputs.  However it always outputs 1, which makes me think I have messed up the definition of the weight matrices.  Full code with all boilerplate is below:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ndef series_gen():\n    for t in range(1,10):\n        yield 1,t\n\ndef input_fn():\n    return tf.data.Dataset.from_generator(series_gen, (tf.float32, tf.float32)).batch(1)\n\ndef model_fn(features, labels, mode):\n    print(features)\n    x = tf.reshape(features, [1, 1, 1]) # batchsize = dimension = timessteps = 1\n\n    rnn = tf.keras.layers.SimpleRNN(\n        1, # 1-dimensional state\n        stateful=True, \n        activation=None, \n        use_bias=False,\n        kernel_initializer = tf.ones,\n        recurrent_initializer = tf.ones,\n    )\n\n    y = rnn(x)\n\n    if mode == tf.estimator.ModeKeys.PREDICT:\n        predictions = {'output': y,}\n        export_outputs = {'predict': tf.estimator.export.PredictOutput(predictions)}\n        return tf.estimator.EstimatorSpec(mode,\n                                          predictions=predictions,\n                                          export_outputs=export_outputs)\n\n    loss = tf.losses.mean_squared_error(y, labels)\n    if mode == tf.estimator.ModeKeys.TRAIN:\n        opt = tf.train.AdamOptimizer()\n        minimize_op = opt.minimize(loss=loss, global_step=tf.train.get_global_step())\n        return tf.estimator.EstimatorSpec(mode, loss=loss, train_op=minimize_op)\n\n    if mode == tf.estimator.ModeKeys.EVAL:   \n        return tf.estimator.EstimatorSpec(mode=mode,\n                                           loss=loss,\n                                         )\n\nestimator = tf.estimator.Estimator(model_fn=model_fn)\npredictions_list = [f['output'] for f in estimator.predict(input_fn=input_fn)]\nprint(f'Predictions: {np.concatenate(predictions_list)}')\n<\/code><\/pre>\n\n<p>And here is the output:<\/p>\n\n<pre><code>Tensor(\"IteratorGetNext:0\", dtype=float32)\nPredictions: [1. 1. 1. 1. 1. 1. 1. 1. 1.]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57548441,"AnswerCount":0,"Tags":"<python><tensorflow><go><python-imaging-library>","CreationDate":"2019-08-18T21:17:22.887","AcceptedAnswerId":null,"OwnerUserId":11943909.0,"Title":"Reading Image File in Go v.s. Reading Image File in Python","Body":"<p>I have code written in Python which reads an image file using <code>PIL<\/code> library, and converts it to n-dimensional numpy array of <code>uint8<\/code> values:<\/p>\n\n<pre><code>    def load_image_into_numpy_array(image):\n          (im_width, im_height) = image.size\n          return np.array(image.getdata()).reshape((im_height, im_width, 3)).astype(np.uint8)\n\n    image_raw = Image.open(img)\n    image = load_image_into_numpy_array(image_raw)\n<\/code><\/pre>\n\n<p>Once I try to print out pixel values from <code>image[0][0][0]<\/code> up to <code>image[9][0][0]<\/code>, I get the following output:<\/p>\n\n<pre><code>image[0][0][0] = 51\nimage[1][0][0] = 51\nimage[2][0][0] = 47\nimage[3][0][0] = 40\nimage[4][0][0] = 38\nimage[5][0][0] = 48\nimage[6][0][0] = 65\nimage[7][0][0] = 81\nimage[8][0][0] = 94\nimage[9][0][0] = 89\nimage[10][0][0] = 84\n<\/code><\/pre>\n\n<p>In Go, I am reading an image by calling the <code>ioutil.ReadFile<\/code> function, and converting it to Tensor by calling Tensorflow API for Go like so:<\/p>\n\n<pre><code>b, err := ioutil.ReadFile(imagePath)\nif err != nil {\n    return nil, err\n}\n\ntensor, err := tf.NewTensor(string(b))\nif err != nil {\n    return nil, err\n}\n<\/code><\/pre>\n\n<p>Once I run this Tensor through a <code>tf.Session<\/code> so that I can convert its values to <code>uint8<\/code>, expand its dimensions so it matches the models input and try to print out pixel values from <code>tensor[0][0][0][0]<\/code> up to <code>tensor[0][9][0][0]<\/code> , I get the following output:<\/p>\n\n<pre><code>tensor[0][0][0][0]: 50\ntensor[0][1][0][0]: 48\ntensor[0][2][0][0]: 45\ntensor[0][3][0][0]: 39\ntensor[0][4][0][0]: 37\ntensor[0][5][0][0]: 47\ntensor[0][6][0][0]: 65\ntensor[0][7][0][0]: 80\ntensor[0][8][0][0]: 95\ntensor[0][9][0][0]: 88\n<\/code><\/pre>\n\n<p>I figured that when reading Image in python using <code>PIL<\/code> library, it internally decodes an byte array to an image of specific format. In Go I am reading an array of bytes and converting it to Tensor as is. There is no decoding phase.<\/p>\n\n<p>I cannot change code in Python, which means that I have to find a way to decode array of byes read by <code>ReadFile<\/code> function, same way as <code>PIL.Image.open<\/code> decodes it.<\/p>\n\n<p>Is there a package in Go which allows me to read an image the same way as PIL does. What would be the best way to approach this type of problem? Possibly find out if <code>PIL.Image<\/code> library wraps a certain C function and create same wrapper function in Go?<\/p>\n","answers":[]}
+{"QuestionId":57548702,"AnswerCount":0,"Tags":"<pytorch><tensor><built-in>","CreationDate":"2019-08-18T22:03:45.677","AcceptedAnswerId":null,"OwnerUserId":10855529.0,"Title":"How to have a look at builtin methods implementation?","Body":"<p>I created a tensor<\/p>\n\n<pre><code>x = torch.randn(3, 4)\n<\/code><\/pre>\n\n<p>I want to see the implementation of builtin methods, for example x.contiguous(), x.view() and so on.<\/p>\n\n<p>How to do this?<\/p>\n","answers":[]}
+{"QuestionId":57549011,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm>","CreationDate":"2019-08-18T23:13:40.360","AcceptedAnswerId":null,"OwnerUserId":2415802.0,"Title":"LSTM Keras input shape confusion","Body":"<p>I am trying to build a predictive model on stock prices. From what I've read, LSTM is a good layer to use. I can't fully understand what my <code>input_shape<\/code> needs to be for my model though.<\/p>\n\n<p>Here is the <code>tail<\/code> of my <code>DataFrame<\/code><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/x2iZk.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/x2iZk.png\" alt=\"print(x_test.shape)\"><\/a><\/p>\n\n<p>I then split the data into train \/ test<\/p>\n\n<pre><code>labels = df['close'].values\nx_train_df = df.drop(columns=['close'])\nx_train, x_test, y_train, y_test = train_test_split(x_train_df.values, labels, test_size=0.2, shuffle=False)\n\nmin_max_scaler = MinMaxScaler()\nx_train = min_max_scaler.fit_transform(x_train)\nx_test = min_max_scaler.transform(x_test)\n\nprint('y_train', y_train.shape)\nprint('y_test', y_test.shape)\nprint('x_train', x_train.shape)\nprint('x_test', x_test.shape)\nprint(x_train)\n<\/code><\/pre>\n\n<p>This yields:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9TKXf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9TKXf.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Here's where I am getting confused. Running the simple example, I get the following error:<\/p>\n\n<blockquote>\n  <p>ValueError: Input 0 of layer lstm_15 is incompatible with the layer:\n  expected ndim=3, found ndim=4. Full shape received: [None, 1, 4026, 5]<\/p>\n<\/blockquote>\n\n<p>I've tried various combinations of messing with the <code>input_shape<\/code> and have came to the conclusion, I have no idea how to determine the input shape.<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(32, input_shape=(1, x_train.shape[0], x_train.shape[1])))\n\nmodel.compile(optimizer='adam', loss='mse', metrics=['accuracy'])\nmodel.fit(x_train, y_train, epochs=10)\n<\/code><\/pre>\n\n<p>Given my dataframe, what should be my <code>input_shape<\/code>? I understand that the input shape is <code>batch size<\/code>, <code>timesteps<\/code>, <code>data dim<\/code>. Just not clear how to map those words to my actual data as what I've thought the values were, are actually not.<\/p>\n\n<p>I was thinking:<\/p>\n\n<ul>\n<li><strong>Batch Size<\/strong>: Number of records I'm passing in (4026)<\/li>\n<li><strong>Time Steps<\/strong>: 1 (I'm not sure if this is supposed to be the same value as batch size?)<\/li>\n<li><strong>Data Dimension<\/strong>: 1 since my data is 1 dimensional (I think?)<\/li>\n<\/ul>\n","answers":[{"AnswerId":"57549108","CreationDate":"2019-08-18T23:37:28.137","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>First of all, I don't think that you need an LSTM at all. Based on the <code>df.tail()<\/code>, it appears to me that there is no temporal dependence between the rows of the pandas data-frame (the samples of your dataset). Anyways, I will come back to that later, firstly your question:<\/p>\n\n<ol>\n<li><p><strong>Batch size:<\/strong> The number of elements in the batch. In total, the dataset contains 4026 elements. On the other hand, the batch size is the number of elements that are processed in a single batch. Let us assume that it is 2. In that case you will have <code>2013<\/code> of these batches.<\/p><\/li>\n<li><p><strong>Time steps:<\/strong> Equal to the number of samples which have a temporal dependence between them. Assuming that in your dataset each 3 instances constitute data sequence, then the time steps will be <code>3<\/code>. Therefore, it follows that each sample in the dataset now consists of <code>3<\/code> measurements, so the total number of elements is <code>1342<\/code> (at the start was <code>4026<\/code>). <\/p><\/li>\n<li><p><strong>Data dimension:<\/strong> The number of features for each element in the batch, for each time step - in your case <code>5<\/code>, assuming that <code>buy<\/code> is the label and <code>date<\/code> is the temporal dependence column.<\/p><\/li>\n<\/ol>\n\n<p>As a result, the shape of a single batch of data should be <code>(2, 3, 6)<\/code>, while the shape of the whole dataset would be <code>(1342, 3, 6)<\/code>. Please note that these shapes are valid if we consider that you use LSTM and that there is temporal dependence between the each of the time steps. Here is an example code snippet to verify some stuff:<\/p>\n\n<pre><code># Random training data\nx_train = np.random.rand(1342, 3, 6)\n# Random training labels\ny_train = np.random.randint(0, 2, 1342)\n\nmodel = Sequential()\nmodel.add(LSTM(32, input_shape=(3, 6)))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])\nmodel.fit(x_train, y_train, epochs=2, batch_size=32)\n<\/code><\/pre>\n\n<p>Now, back to my previous point. Looking at <code>df.tail()<\/code>, it seems to me that there is no temporal dependence whatsoever between the samples in the dataset. With that said, I would firstly convert the <code>date<\/code> column to something meaningful (one-hot encoding of the month in the year, one-hot encoding of the 4 seasons depending on the month, etc). Then, instead of constructing an RNN, I will proceed with a feed-forward neural network with a binary classification output layer.<\/p>\n\n<p>As for the model, once you take care of all data related stuff, something as simple as this should work for you:<\/p>\n\n<pre><code># Random training data\nx_train = np.random.rand(4026, 5)\n# Random training labels\ny_train = np.random.randint(0, 2, 4026)\n\nmodel = Sequential()\nmodel.add(Dense(32, activation='relu', input_dim=5))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])\nmodel.fit(x_train y_train, epochs=2, batch_size=32)\n<\/code><\/pre>\n\n<p>Hope that it helps!<\/p>\n"}]}
+{"QuestionId":57549399,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-08-19T00:46:45.750","AcceptedAnswerId":null,"OwnerUserId":1179925.0,"Title":"Keras: how to determine at which iteration best val model was saved?","Body":"<p>Part of my training code looks like this:<\/p>\n\n<pre><code>    model_filepath = 'models\/cnn_best.h5'\n    datetime_str = ('{date:%Y-%m-%d-%H:%M:%S}'.format(date=datetime.now()))\n    callbacks = [\n        ModelCheckpoint(model_filepath, monitor='val_loss', save_best_only=True, verbose=0),\n        TensorBoard(log_dir='.\/logs_'+args.model_postfix+'\/'+datetime_str,\n                    histogram_freq=0, write_graph=True, write_images=True),\n        CustomCallback(generator.batch_generator(is_train=True), generator.batch_generator(is_train=False),\n                       args.model_postfix)\n    ]\n\n    history = model.fit_generator(\n        generator=generator.batch_generator(is_train=True),\n        epochs=config.N_EPOCHS,\n        steps_per_epoch=100,\n        validation_data=generator.batch_generator(is_train=False),\n        validation_steps=10,\n        verbose=1,\n        shuffle=False,\n        callbacks=callbacks)\n<\/code><\/pre>\n\n<p><code>ModelCheckpoint<\/code> save model at best val score with <code>model_filepath<\/code> name, is it possible to add step\/epoch also to model filename? i.e. maybe by deriving from <code>ModelCheckpoint<\/code>? or only method is to parse <code>history<\/code> for min val loss values and rename file? or don't use <code>model.fit_generator<\/code> but feed data from generator in a loop?(which looks bad option because in this case I will need to make wrapper above batch generator to make data loading multi process)<\/p>\n","answers":[{"AnswerId":"57550116","CreationDate":"2019-08-19T03:21:31.287","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>If you put <code>verbose=1<\/code> in the checkpoint, you will see when it saves. <\/p>\n\n<p>You can also create a quick <code>LambdaCallback<\/code> yourself:<\/p>\n\n<pre><code>monitor = 'val_loss'    \nbestLoss = 1000000000\nbestEpoch = 0\n\ndef saveFunction(epoch, logs):\n    loss = logs[monitor]\n    if loss &lt; bestLoss:\n        model.save(name + \"_\" + str(epoch))\n        bestLoss = loss\n        bestEpoch = epoch\n\nsaverCallback = LambdaCallback(on_epoch_end = saveFunction)\n#and fit\n<\/code><\/pre>\n\n<p>If you like, you can also create a custom callback instead of a Lambda, in case you want to keep the vars inside it as properties.    <\/p>\n"}]}
+{"QuestionId":57549448,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow2.0>","CreationDate":"2019-08-19T00:58:37.483","AcceptedAnswerId":null,"OwnerUserId":11296270.0,"Title":"how to convert perreplica to tensor?","Body":"<p>When training with multi gpu in tensorflow2.0, perreplica would be reduce by below code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>strategy.reduce(tf.distribute.ReduceOp.SUM, per_replica_losses, axis=None)\n<\/code><\/pre>\n\n<p>However, if I just want to collect(no 'sum reduce' or 'mean reduce') all gpu's predictions into a tensor:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>per_replica_losses, per_replica_predicitions = strategy.experimental_run_v2(train_step, args=(dataset_inputs,))\n# how to convert per_replica_predicitions to a tensor ?\n<\/code><\/pre>\n","answers":[{"AnswerId":"58722946","CreationDate":"2019-11-06T04:06:11.830","ParentId":null,"OwnerUserId":"12329652","Title":null,"Body":"<p>In short, you can convert <code>PerReplica<\/code> result into a tuple of tensors like this:<\/p>\n\n<pre><code>tensors_tuple = per_replica_predicitions.values\n<\/code><\/pre>\n\n<p>the return <code>tensors_tuple<\/code> will be a tuple of <code>predictions<\/code> from each replicas\/devices:<\/p>\n\n<pre><code>(predicton_tensor_from_dev0, prediction_tensor_from_dev1,...)\n<\/code><\/pre>\n\n<p>The number of elements in this tuple is determined by your devices available to the distributed strategy. Specially, if the strategy runs on a single replica\/device, the return value from strategy.experimental_run_v2 will be the same as calling train_step function directly (tensor or list of tensors decided by your <code>train_step<\/code>). So you might want to write the code like this:<\/p>\n\n<pre><code>per_replica_losses, per_replica_predicitions = strategy.experimental_run_v2(train_step, args=(dataset_inputs,))\n\nif strategy.num_replicas_in_sync &gt; 1:\n    predicition_tensors = per_replica_predicitions.values\nelse:\n    predicition_tensors = per_replica_predicitions\n<\/code><\/pre>\n\n<p><code>PerReplica<\/code> is a class object wrapping the results of distributed running. You can find its definition <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/1cf0898dd4331baf93fe77205550f2c2e6c90ee5\/tensorflow\/python\/distribute\/values.py#L465\" rel=\"nofollow noreferrer\">here<\/a>, there are more properties\/methods for us to operate the <code>PerReplica<\/code> object.<\/p>\n"}]}
+{"QuestionId":57549521,"AnswerCount":1,"Tags":"<javascript><tensorflow><dom-events><tensorflow.js><domcontentloaded>","CreationDate":"2019-08-19T01:13:05.463","AcceptedAnswerId":57549669.0,"OwnerUserId":326389.0,"Title":"tf.browser.fromPixels() not working because \"DOM is not ready yet\"","Body":"<p>I'm trying to play with TensorFlow.js's <code>tf.browser.fromPixels()<\/code> function. Unfortunately, I get the following error:<\/p>\n\n<pre><code>backend_webgl.ts:309 Uncaught Error: The DOM is not ready yet. Please call tf.browser.fromPixels() once the DOM is ready. One way to do that is to add an event listener for `DOMContentLoaded` on the document object\n    at e.fromPixels (backend_webgl.ts:309)\n    at t.fromPixels (engine.ts:870)\n    at fromPixels_ (browser.ts:55)\n    at Object.fromPixels (operation.ts:45)\n    at ex4.html:10\n<\/code><\/pre>\n\n<p>But I'm calling <code>tf.browser.fromPixels()<\/code> from inside a function passed to <code>window.addEventListener('DOMContentLoaded', ...)<\/code>. Any idea why I'm getting this error and how to get the <code>fromPixels<\/code> function to 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-html lang-html prettyprint-override\"><code>&lt;html&gt;\r\n&lt;head&gt;\r\n    &lt;script src=\"https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/tfjs@1.2.7\/dist\/tf.min.js\"&gt;&lt;\/script&gt;\r\n&lt;\/head&gt;\r\n&lt;body&gt;\r\n    &lt;img id=\"test\" src=\"https:\/\/jpeg.org\/images\/jpeg-home.jpg\"&gt;\r\n    &lt;script&gt;\r\n        window.addEventListener('DOMContentLoaded', ev =&gt; {\r\n            const img = document.getElementById('test')\r\n            tf.browser.fromPixels(img).print()\r\n        })\r\n    &lt;\/script&gt;\r\n&lt;\/body&gt;\r\n&lt;\/html&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n","answers":[{"AnswerId":"57549669","CreationDate":"2019-08-19T01:51:25.140","ParentId":null,"OwnerUserId":"3162162","Title":null,"Body":"<pre><code>        if (document.readyState !== 'complete') {\n          throw new Error(\n              'The DOM is not ready yet. Please call ' +\n              'tf.browser.fromPixels() once the DOM is ready. One way to ' +\n              'do that is to add an event listener for `DOMContentLoaded` ' +\n              'on the document object');\n        }\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tfjs\/blob\/aa92053a8d7c640fe9e685e254655f417480cf0d\/tfjs-core\/src\/backends\/webgl\/backend_webgl.ts#L313-L321\" rel=\"nofollow noreferrer\">Github tensorflow\/tfjs<\/a><\/p>\n\n<p>According to the code above, the error is thrown because <code>document.readyState<\/code> isn't <code>complete<\/code>.<\/p>\n\n<blockquote>\n  <p>The DOMContentLoaded event fires when the initial HTML document has been completely loaded and parsed, without waiting for stylesheets, images, and subframes to finish loading.\n  <a href=\"https:\/\/developer.mozilla.org\/en-US\/docs\/Web\/API\/Window\/DOMContentLoaded_event\" rel=\"nofollow noreferrer\">https:\/\/developer.mozilla.org\/en-US\/docs\/Web\/API\/Window\/DOMContentLoaded_event<\/a><\/p>\n  \n  <p>The readyState of a document can be one of following:<\/p>\n  \n  <p>loading<br>\n  The document is still loading.  <\/p>\n  \n  <p>interactive<br>\n  The document has finished loading and the document has been parsed but sub-resources such as images, stylesheets and frames are still loading.  <\/p>\n  \n  <p>complete<br>\n  The document and all sub-resources have finished loading. The state indicates that the load event is about to fire.<br>\n  <a href=\"https:\/\/developer.mozilla.org\/en-US\/docs\/Web\/API\/Document\/readyState\" rel=\"nofollow noreferrer\">https:\/\/developer.mozilla.org\/en-US\/docs\/Web\/API\/Document\/readyState<\/a><\/p>\n<\/blockquote>\n\n<p><code>DOMContentLoaded<\/code> is called before any sub-resources aren't loaded yet.So At the time of your code is executed, <code>document.readyState<\/code> is still <code>interactive<\/code>.<\/p>\n\n<p>Just ignore what the error says and try <code>load<\/code> event listener instead.<\/p>\n"}]}
+{"QuestionId":57549571,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-19T01:24:51.353","AcceptedAnswerId":null,"OwnerUserId":11727622.0,"Title":"Why I can't find tf.feature_to_id()","Body":"<p>Anybody know where is tf.feature_to_id() function? Does it deprecated if yes anybody knows which function currently should be used?<\/p>\n","answers":[]}
+{"QuestionId":57549965,"AnswerCount":1,"Tags":"<python><nlp><pytorch><multilabel-classification>","CreationDate":"2019-08-19T02:54:04.510","AcceptedAnswerId":null,"OwnerUserId":8543561.0,"Title":"NLP multi-class classifier loss can\u2019t go down","Body":"<p>I'm building a classifier for a QA bot, and  have <strong>a dataset for 8k questions, and 149 different Answers.<\/strong><\/p>\n\n<p>I got some problems when training my model; the \"loss\" won't go down as I expected so I am asking for your help...<\/p>\n\n<h3>Here is my method:<\/h3>\n\n<p>I use word2vec to get a word's vector, then use a GRU-based network to get the vector of sentence The w2v model has been trained with wiki data, and works well on another of my NLP projects. The GRU code is from my senior, I think it works well, too.<\/p>\n\n<pre><code># Part of the code for getting sentence vector\ninput_size = 400 \nhidden_dim = 400  \nnum_layers = 1\ngru = nn.GRU(input_size, hidden_dim,num_layers,batch_first = True)\n\nh0 = torch.rand(num_layers, 7187, hidden_dim) # (num_layers, batch, hidden_dim) \n# shape of input [dataset_len,max_sentence_len,input_feature]\ninputSet = torch.tensor(x_train,dtype = torch.float)\nsentenceVecs, hidden = gru(inputSet,h0)\nsentenceVecs = sentenceVecs[:,-1, :] \n<\/code><\/pre>\n\n<p>and here is my classifier model<\/p>\n\n<pre><code>from argparse import Namespace\nargs = Namespace(\n    dataset_file = 'dataset\/waimai_10k_tw.pkl',\n    model_save_path='torchmodel\/pytorch_bce.model',\n    # Training hyper parameters\n    batch_size = 100,\n    learning_rate = 0.002,\n    min_learning_rate = 0.002,\n    num_epochs=200,\n)\n\nclass JWP(nn.Module):\n    def __init__(self, \n                 n_feature, \n                 n_hidden,\n                 n_hidden2,\n                 n_hidden3,\n                 n_output):\n\n        super(JWP, self).__init__()\n        self.hidden = nn.Linear(n_feature, n_hidden)\n        self.hidden2 = nn.Linear(n_hidden, n_hidden2)\n        self.hidden3 = nn.Linear(n_hidden2, n_hidden3)\n        self.out = nn.Linear(n_hidden3, n_output)\n\n    def forward(self, x, apply_softmax=False):\n        x = F.relu(self.hidden(x).squeeze())\n        x = F.relu(self.hidden2(x).squeeze())\n        x = F.relu(self.hidden3(x).squeeze())\n        # \n        if(apply_softmax):\n            x = torch.softmax(self.out(x))\n        else:\n            x = self.out(x)\n\n        return x\n<\/code><\/pre>\n\n<p>training code<\/p>\n\n<pre><code>lr = args.learning_rate\nmin_lr = args.min_learning_rate\ndef adjust_learning_rate(optimizer, epoch):\n    global lr\n    if epoch % 10 == 0 and epoch != 0:\n        lr = lr * 0.65\n        if(lr &lt; min_lr):\n            lr = min_lr\n        for param_group in optimizer.param_groups:\n            param_group['lr'] = lr\n\nif __name__ == \"__main__\":\n    EPOCH = args.num_epochs\n    net = JWP(400,325,275,225,149)\n#     net = JWP(400,250,149)\n#     net = JWP(400,149)\n    print(net)\n\n    optimizer = torch.optim.SGD(net.parameters(), lr=lr)\n    loss_func = torch.nn.CrossEntropyLoss()\n\n    for t in range(EPOCH):\n        adjust_learning_rate(optimizer,t)\n\n        \"\"\"\n        Train phase\n        \"\"\"\n        net.train() \n        TrainLoss = 0.0\n        # Train batch\n        for step,(batchData, batchTarget) in enumerate(trainDataLoader):\n            optimizer.zero_grad() \n            out = net(batchData)\n            loss = loss_func(out,batchTarget) \n            TrainLoss = TrainLoss + loss\n            loss.backward() \n            optimizer.step() \n        TrainLoss = TrainLoss \/ (step+1) # epoch loss\n\n        \"\"\"\n        Result\n        \"\"\"\n        print(\n            \"epoch:\",t+1 ,\n            \"train_loss:\",round(TrainLoss.item(),3),\n            \"LR:\",lr\n        )\n<\/code><\/pre>\n\n<p>Is it that my model is too simple or do I simply use the wrong method?\nThe loss is always stuck at around 4.6 and I can't lower it any more...<\/p>\n\n<pre><code>epoch: 2898 train_loss: 4.643 LR: 0.002\nepoch: 2899 train_loss: 4.643 LR: 0.002\nepoch: 2900 train_loss: 4.643 LR: 0.002\nepoch: 2901 train_loss: 4.643 LR: 0.002\n<\/code><\/pre>\n","answers":[{"AnswerId":"57551974","CreationDate":"2019-08-19T07:13:19.710","ParentId":null,"OwnerUserId":"8543561","Title":null,"Body":"<p>came back ans myself:\nthere\u2019s some \u201cvery long sentence\u201d in my dataset, may make some \u201cnoise\u201d, because of RNN(GRU) method hard to keep those info, after taking \u201cvery long sentence\u201d away from my dataset, loss strat going down and acc was pertty<\/p>\n"}]}
+{"QuestionId":57549972,"AnswerCount":2,"Tags":"<python><tensorflow><pytorch>","CreationDate":"2019-08-19T02:56:24.700","AcceptedAnswerId":null,"OwnerUserId":6151940.0,"Title":"Error in converting tensorflow BERT checkpoints to pytorch","Body":"<p>I downloaded tensorflow checkpoints for domain specific bert model and extracted the zip file into the folder pretrained_bert which contains the following the three files<\/p>\n\n<blockquote>\n  <p>model.ckpt.data-00000-of-00001 <\/p>\n  \n  <p>model.ckpt.index <\/p>\n  \n  <p>model.ckpt.meta<\/p>\n<\/blockquote>\n\n<p>I used the following code to convert tensorflow checkpoints to pytorch<\/p>\n\n<pre><code>import torch\n\nfrom pytorch_transformers.modeling_bert import BertConfig, BertForPreTraining, load_tf_weights_in_bert\n\n\ntf_checkpoint_path=\"pretrained_bert\/model.ckpt\"\nbert_config_file = \"bert-base-cased-config.json\"\npytorch_dump_path=\"pytorch_bert\"\n\nconfig = BertConfig.from_json_file(bert_config_file)\nprint(\"Building PyTorch model from configuration: {}\".format(str(config)))\nmodel = BertForPreTraining(config)\n\n# Load weights from tf checkpoint\nload_tf_weights_in_bert(model, config, tf_checkpoint_path)\n\n# Save pytorch-model\nprint(\"Save PyTorch model to {}\".format(pytorch_dump_path))\ntorch.save(model.state_dict(), pytorch_dump_path)\n\n<\/code><\/pre>\n\n<p>I got this error when I ran the above code<\/p>\n\n<blockquote>\n  <p>NotFoundError: Unsuccessful TensorSliceReader constructor: Failed to\n  find any matching files for pretrained_bert\/model.ckpt<\/p>\n<\/blockquote>\n\n<p>Any help is really appreciated............<\/p>\n","answers":[{"AnswerId":"57583883","CreationDate":"2019-08-21T02:57:24.860","ParentId":null,"OwnerUserId":"6151940","Title":null,"Body":"<p>In the code,  absolute path must be specified for the variables<\/p>\n\n<ul>\n<li>tf_checkpoint_path <\/li>\n<li>bert_config_file  <\/li>\n<li>pytorch_dump_path<\/li>\n<\/ul>\n\n<p>When relative path is specified, the model is unable to find the corresponding files.<\/p>\n"},{"AnswerId":"57587584","CreationDate":"2019-08-21T08:37:15.957","ParentId":null,"OwnerUserId":"8893595","Title":null,"Body":"<p>As the error says, <\/p>\n\n<blockquote>\n  <p>Failed to find any matching files for pretrained_bert\/model.ckpt<\/p>\n<\/blockquote>\n\n<p>It might be possible that your program may not be executed from the location which has a folder named <code>pretrained_bert<\/code> or it does doesn't have the <code>pretrained_bert\/model.ckpt*<\/code> files.<\/p>\n\n<p>So, first, ensure the presence of the file. Further, you can follow the below steps:<\/p>\n\n<p>Kindly provide absolute path instead of the relative path. This reverifies the presence of files. In case if you still want to use the relative path, do verify the path from where the current execution takes place, and then use <code>..\/<\/code> if you want to navigate up to the parent directory.<\/p>\n"}]}
+{"QuestionId":57549989,"AnswerCount":2,"Tags":"<python><arrays><numpy><tensorflow>","CreationDate":"2019-08-19T02:59:31.270","AcceptedAnswerId":57550061.0,"OwnerUserId":11888481.0,"Title":"How can I reshape ndarray, especially easy way ro (25,25) to (1,1)?","Body":"<p>I try to make batch set for training and after made small example with (25,25) ndarray, but batch need only (60000) sample not the (60000, 25, 25)<\/p>\n\n<p>so I realized i have to reshape each of image, 25*25, to 1 way vector<\/p>\n\n<p>i can use some kind of 'for' shaping, but always coding be like, i guess there is easy way to invert it.<\/p>\n\n<pre><code>array([[0, 0, 1, 0, 0],\n   [0, 0, 0, 1, 0],\n   [0, 0, 1, 0, 0],\n   [0, 0, 1, 0, 0],\n   [0, 0, 0, 0, 0]]\n\narray([ 0 0 1 0 0 0 0 0 1 0 0 0 1 0 0 0 0 1 0 0 0 0 0 0 0 ]\n<\/code><\/pre>\n","answers":[{"AnswerId":"57550045","CreationDate":"2019-08-19T03:09:32.843","ParentId":null,"OwnerUserId":"2823755","Title":null,"Body":"<p>Reshape using the original number of rows and combine the second and third dimensions.<\/p>\n\n<pre><code>import numpy as np\na = np.random.randint(0,10,(10,3,3))\nh,w,d = a.shape\nb = a.reshape((h,w*d))\n<\/code><\/pre>\n\n<hr>\n\n<pre><code>&gt;&gt;&gt; \n&gt;&gt;&gt; a.shape\n(10, 3, 3)\n&gt;&gt;&gt; b.shape\n(10, 9)\n&gt;&gt;&gt;\n<\/code><\/pre>\n"},{"AnswerId":"57550061","CreationDate":"2019-08-19T03:11:49.573","ParentId":null,"OwnerUserId":"2570261","Title":null,"Body":"<p>there's a method called flatten for just this use case:<\/p>\n\n<pre><code>arr = array([[0, 0, 1, 0, 0],\n   [0, 0, 0, 1, 0],\n   [0, 0, 1, 0, 0],\n   [0, 0, 1, 0, 0],\n   [0, 0, 0, 0, 0]])\n\narr.flatten()\n# outputs\narray([0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0,\n       0, 0, 0])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57550028,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-08-19T03:07:03.387","AcceptedAnswerId":57551491.0,"OwnerUserId":8515794.0,"Title":"accessing indexes of tf.data.Dataset for deleting and appending data elements","Body":"<p>I am currently working on a research project in which I have to append my training set after some epochs and delete some samples from test set after evaluation. Currently there is no way through with which I can access the records (placed at specific indexes) in tfrecord file to delete or append. Since tfrecords offer very fast training so I avoid using generators. Any suggestion to access tfrecord file in this way?<\/p>\n","answers":[{"AnswerId":"57551491","CreationDate":"2019-08-19T06:30:23.163","ParentId":null,"OwnerUserId":"2455494","Title":null,"Body":"<blockquote>\n  <p>I have to append my training set after some epochs<\/p>\n<\/blockquote>\n\n<p>You can use <code>repeat(n)<\/code> to repeat the dataset for <code>n<\/code> epochs then <code>concatenate(new_dataset)<\/code> to append some extra data. So, in order to append the new data after 15 epochs, for example, we could do:<\/p>\n\n<pre><code>dataset = tf.data.TFRecordDataset('filepath.tfrecord')\nnew_data = tf.data.TFRecordDataset('filepath_of_records_to_append.tfrecord') # or any other dataset from generator or whatever!\n\ndataset = dataset.repeat(15).concatenate(new_data)\n<\/code><\/pre>\n\n<blockquote>\n  <p>and delete some samples from test set <\/p>\n<\/blockquote>\n\n<p>You can't easily delete from the original tfrecord file using the tf.data API (you'd effectively have to write a new tfrecord file with the records omitted) so perhaps instead you'd find it useful to just construct a <code>tf.data<\/code> Dataset and ignore or skip the records you want omitted? This is much more straightforward and can be done using <code>take()<\/code> &amp; <code>skip()<\/code>.<\/p>\n\n<p>So, for example, if I wanted to skip the records at index 4, 7,8,9 &amp; 10 I could do the following:<\/p>\n\n<pre><code>dataset = tf.data.TFRecordDataset('filepath.tfrecord')\n\ndataset = dataset.take(4).skip(1).take(2).skip(4)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57550578,"AnswerCount":0,"Tags":"<python><tensorflow><bazel>","CreationDate":"2019-08-19T04:41:28.360","AcceptedAnswerId":null,"OwnerUserId":7711991.0,"Title":"I got error in tensorflow bazel building. how can i fix it","Body":"<p>I'm trying to install tensorflow-gpu.. I tried many ways ang always failed to get tensorflow to see my GPU\n(One thing to mention is I installed the CPU version then uninstalled it)\nI tried another method...\nI followed those steps shown here <a href=\"https:\/\/gist.github.com\/Brainiarc7\/6d6c3f23ea057775b72c52817759b25c\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/Brainiarc7\/6d6c3f23ea057775b72c52817759b25c<\/a> to build tensorflow from source code using bazel.<\/p>\n\n<p>this the configuration I set for installation ..<\/p>\n\n<pre><code>    WARNING: --batch mode is deprecated. Please instead explicitly shut down your Bazel server using the command \"bazel shutdown\".\nINFO: Invocation ID: 5f4b37cb-8cbb-481f-b47c-31a03e671769\nYou have bazel 0.21.0 installed.\nPlease specify the location of python. [Default is \/home\/bignrz\/programs\/anaconda3\/bin\/python]: \n\n\nFound possible Python library paths:\n  \/home\/bignrz\/programs\/anaconda3\/lib\/python3.7\/site-packages\nPlease input the desired Python library path to use.  Default is [\/home\/bignrz\/programs\/anaconda3\/lib\/python3.7\/site-packages]\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 OpenCL SYCL support? [y\/N]: n\nNo OpenCL SYCL support will be enabled for TensorFlow.\n\nDo you wish to build TensorFlow with ROCm support? [y\/N]: n\nNo ROCm support will be enabled for TensorFlow.\n\nDo you wish to build TensorFlow with CUDA support? [y\/N]: y\nCUDA support will be enabled for TensorFlow.\n\nPlease specify the CUDA SDK version you want to use. [Leave empty to default to CUDA 10.0]: 10.1\n\n\nPlease specify the location where CUDA 10.1 toolkit is installed. Refer to README.md for more details. [Default is \/usr\/local\/cuda]: \/usr\/local\/cuda-10.1\n\n\nPlease specify the cuDNN version you want to use. [Leave empty to default to cuDNN 7]: \n\n\nPlease specify the location where cuDNN 7 library is installed. Refer to README.md for more details. [Default is \/usr\/local\/cuda-10.1]: \n\n\nDo you wish to build TensorFlow with TensorRT support? [y\/N]: n\nNo TensorRT support will be enabled for TensorFlow.\n\nPlease specify the locally installed NCCL version you want to use. [Default is to use https:\/\/github.com\/nvidia\/nccl]: \n\n\nPlease specify a list of comma-separated Cuda compute capabilities you want to build with.\nYou can find the compute capability of your device at: https:\/\/developer.nvidia.com\/cuda-gpus.\nPlease note that each additional compute capability significantly increases your build time and binary size. [Default is: 6.1]: \n\n\nDo you want to use clang as CUDA compiler? [y\/N]: n\nnvcc will be used as CUDA compiler.\n\nPlease specify which gcc should be used by nvcc as the host compiler. [Default is \/usr\/bin\/gcc]: \n\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 -Wno-sign-compare]: \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 .bazelrc for more details.\n    --config=mkl            # Build with MKL support.\n    --config=monolithic     # Config for mostly static monolithic build.\n    --config=gdr            # Build with GDR support.\n    --config=verbs          # Build with libverbs support.\n    --config=ngraph         # Build with Intel nGraph support.\n    --config=dynamic_kernels    # (Experimental) Build kernels into separate shared objects.\nPreconfigured Bazel build configs to DISABLE default on features:\n    --config=noaws          # Disable AWS S3 filesystem support.\n    --config=nogcp          # Disable GCP support.\n    --config=nohdfs         # Disable HDFS support.\n    --config=noignite       # Disable Apache Ignite support.\n    --config=nokafka        # Disable Apache Kafka support.\n    --config=nonccl         # Disable NVIDIA NCCL support.\n<\/code><\/pre>\n\n<p>When I used this line here <\/p>\n\n<pre><code>bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mfpmath=both --copt=-msse4.2 --config=cuda \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>i got the following error<\/p>\n\n<pre><code>WARNING: The following configs were expanded more than once: [cuda]. For repeatable flags, repeats are counted twice and may lead to unexpected behavior.\nINFO: Invocation ID: 7cf2de80-0c55-4167-89b8-cebad61304e9\nERROR: error loading package '': in \/media\/bignrz\/World\/Programs\/Linux\/tensorflow\/third_party\/toolchains\/preconfig\/generate\/archives.bzl: Unable to load package for '\/\/tensorflow:version_check.bzl': BUILD file not found on package path\nERROR: error loading package '': in \/media\/bignrz\/World\/Programs\/Linux\/tensorflow\/third_party\/toolchains\/preconfig\/generate\/archives.bzl: Unable to load package for '\/\/tensorflow:version_check.bzl': BUILD file not found on package path\nINFO: Elapsed time: 0.089s\nINFO: 0 processes.\nFAILED: Build did NOT complete successfully (0 packages loaded)\n<\/code><\/pre>\n\n<p>I don't know what is this error about<\/p>\n","answers":[]}
+{"QuestionId":57550935,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensor>","CreationDate":"2019-08-19T05:30:48.757","AcceptedAnswerId":57551178.0,"OwnerUserId":4806177.0,"Title":"Tensorflow Tf.tf.squared_difference is showing a value error with dense layer","Body":"<p>Whenever I try to multiply two tensors and then feed them as an input to a dense layer, it is working perfectly. But, when I try to calculate the squared difference between them, it's showing me an error.<\/p>\n\n<pre><code># working well\nout= multiply([user, book])    \n\nresult = Dense(1, activation='sigmoid', kernel_initializer=initializers.lecun_normal(),\n                   name='prediction')(out)\n# error\nout= tf.reduce_sum(tf.squared_difference(user, book),1)\n\nresult = Dense(1, activation='sigmoid', kernel_initializer=initializers.lecun_normal(),\n                   name='prediction')(out)\n<\/code><\/pre>\n\n<p>Here is the error I get:<\/p>\n\n<pre><code>Input 0 is incompatible with layer prediction: expected min_ndim=2, found ndim=1 error\n<\/code><\/pre>\n","answers":[{"AnswerId":"57551178","CreationDate":"2019-08-19T06:03:33.823","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You probably need to pass <code>keepdims=True<\/code> argument to <code>reduce_sum<\/code> function in order to keep the dimensions with length 1 (otherwise, the shape of <code>out<\/code> would be <code>(batch_size)<\/code>, whereas the <code>Dense<\/code> layer expects <code>(batch_size, N)<\/code>):<\/p>\n\n<pre><code>out= tf.reduce_sum(tf.squared_difference(user, book), axis=1, keepdims=True)\n<\/code><\/pre>\n\n<p><strong>Update:<\/strong> The input of Keras layers must be the output of other Keras layers. Therefore, if you want to use TensorFlow operations, you need to wrap them inside a <code>Lambda<\/code> layer in Keras. For example:<\/p>\n\n<pre><code>from keras.layers import Lambda\n\nout = Lambda(lambda x: tf.reduce_sum(tf.squared_difference(x[0], x[1]), axis=1, keepdims=True))([user, book])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57551137,"AnswerCount":0,"Tags":"<python><keras><word2vec>","CreationDate":"2019-08-19T05:58:13.133","AcceptedAnswerId":null,"OwnerUserId":8170648.0,"Title":"Question about confilct in result of skipgram function in Keras","Body":"<p>I have a question about skipgram function in keras.preprocessing.sequence. <\/p>\n\n<p>In specification, <a href=\"https:\/\/keras.io\/preprocessing\/sequence\/\" rel=\"nofollow noreferrer\">link<\/a>, skipgrams function generates negative samples with dictionary. <\/p>\n\n<p>To understand skipgrams, I print the result of skipgrams function.<\/p>\n\n<p>I noticed there is a conflicted results, a pair in both side of positive samples and negative samples.<\/p>\n\n<p>These data does not effect on model? <\/p>\n\n<p>or<\/p>\n\n<p>Big dictionary can make ignore these conflicts? <\/p>\n\n<p>my code is as follows: <\/p>\n\n<pre><code>samplingTable = keras.preprocessing.sequence.make_sampling_table(26)\n\nprint (samplingTable)\ntargetList = [random.randint(0,25) for i in range(200)]\n\ncouple, labels = keras.preprocessing.sequence.skipgrams(targetList, 26, sampling_table=samplingTable,shuffle=False )\n\nresult= []\nfor i in range(len(labels)):\n    result.append(\"{}\\t{}\\t{}\".format(couple[i][0],couple[i][1],labels[i]))\n\nfor i in sorted(result):\n    print (i)\n<\/code><\/pre>\n\n<p>and follows are result of code. (8,24), (8,25) are conflicted. (positive \n 1 and negative 0)<\/p>\n\n<pre><code>target  context isExist\n8   1   1\n8   10  0\n8   10  0\n8   11  1\n8   12  0\n8   14  0\n8   2   1\n8   24  0\n8   24  1\n8   25  0\n8   25  1\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57551603,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><matrix><convolution>","CreationDate":"2019-08-19T06:41:33.387","AcceptedAnswerId":null,"OwnerUserId":8870808.0,"Title":"Improve speed on convolution along an axis","Body":"<p>I am working on a problem where I need to return the <strong>vertical<\/strong> (along y axis) and <strong>horizontal convolution<\/strong> (along x axis) of a <strong>M x N matrix<\/strong> with a kernel and store the convoluted result to matrix Dy and Dx respectively . The kernel <strong>K is [-1, 0, 1]<\/strong> and I am told the code that I have is 3-4x as slow as the solution. I believe the purpose of the problem is to explore separable 2D kernel and how you can convert a 2D convolution into two 1D convolution when the kernel is separable to improve runtime. <\/p>\n\n<p>The main objective is essentially <strong>convolve a M x N matrix with kernel K along the vertical and horizontal axis and store the results in Dy and Dx respectively<\/strong>. I am told the following code is too slow and am wondering how I can make it faster. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def convolution2d(matrix, kernel):\n    m, n = kernel.shape #row and column of kernel \n    x, y = matrix.shape #row and column of matrix M \n    row = x - m + 1    #output row of resulted convol\n    col = y - n + 1    #output col of resulted convol\n    result = np.zeros((row,col), dtype = int)\n    for i in range(row):\n        for j in range(col):\n            result[i][j] = np.sum(matrix[i:i+m, j:j+n]*kernel)\n    return result\n<\/code><\/pre>\n\n<p>When I do the horizontal convolution I use kernel K = [-1, 0, 1] on matrix M  and for vertical convolution I transpose K into [[-1], [0], [1]] to convolve matrix M. I wonder if I have misunderstood the question (i.e. kernel should be 1D not 2D) or overcomplicated the solution (no need to transpose\/alter kernel, use the same kernel)  and shouldn't be doing conv2d at all.<\/p>\n\n<p>Restricted from using CUDA or parallel threading, I have researched online and couldn't find much on how to optimize my solution. I suspect the optimized solution may involve doing 1d convolution instead of 2d and would like confirmation. I appreciate any help and guidance that can lead me to and help me understand what I need to do.<\/p>\n","answers":[]}
+{"QuestionId":57551650,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-19T06:44:53.110","AcceptedAnswerId":null,"OwnerUserId":8471146.0,"Title":"Unable to use Deep Learning RNN models trained on GPU instance, for the inference on CPU-only instance","Body":"<p>I trained a deep learning model on GPU instance to speedup training process.\nBut unfortunately our product instances are CPU-only instances.\nWhen I tried to use GPU trained model with CuDNNGRU on CPU-only instance, I got below error.<\/p>\n\n<pre><code>ValueError: CuDNNGRU is not compatible with GRU(reset_after=False)\n<\/code><\/pre>\n\n<p>similar kind of error you will get if you use CuDNNLSTMs<\/p>\n\n<pre><code>ValueError: CuDNNLSTM is not compatible with LSTM(reset_after=False)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57551651","CreationDate":"2019-08-19T06:44:53.110","ParentId":null,"OwnerUserId":"8471146","Title":null,"Body":"<p>Instead of using CuDNNGRU or CuDNNLSTM, use normal GRU or LSTM with following options<\/p>\n\n<p>Code that throws error when CuDNNGRU or CuDNNLSTM models trained on GPU throw error while inference on CPU-only instances.<\/p>\n\n<p><strong>CuDNNGRU<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>if tf.test.is_gpu_available():\n    return tf.keras.layers.CuDNNGRU(units, \n                                    return_sequences=True, \n                                    return_state=True,\n                                    recurrent_initializer='glorot_uniform')\nelse:\n    return tf.keras.layers.GRU(units,\n                               return_sequences=True,\n                               return_state=True, \n                               recurrent_initializer='glorot_uniform')\n<\/code><\/pre>\n\n<p><em>or<\/em><\/p>\n\n<p><strong>CuDNNLSTM<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>if tf.test.is_gpu_available():\n    return tf.keras.layers.CuDNNLSTM(units, \n                                    return_sequences=True, \n                                    return_state=True,\n                                    recurrent_initializer='glorot_uniform')\nelse:\n    return tf.keras.layers.LSTM(units,\n                               return_sequences=True,\n                               return_state=True, \n                               recurrent_initializer='glorot_uniform')\n<\/code><\/pre>\n\n<p>Please use below code, so that you can use GPU trained models for the inference on CPU-only instances.<\/p>\n\n<p><strong>GRU<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>if tf.test.is_gpu_available():\n    return tf.compat.v1.keras.layers.CuDNNGRU(units,\n                                              return_sequences=True,\n                                              return_state=True,\n                                              recurrent_initializer='glorot_uniform')\nelse:\n    return tf.keras.layers.GRU(units, \n                               return_sequences=True,\n                               return_state=True,\n                               recurrent_activation='sigmoid',\n                               reset_after='True',\n                               recurrent_initializer='glorot_uniform')\n<\/code><\/pre>\n\n<p><em>or<\/em><\/p>\n\n<p><strong>LSTM<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>if tf.test.is_gpu_available():\n    return tf.compat.v1.keras.layers.CuDNNLSTM(units,\n                                              return_sequences=True,\n                                              return_state=True,\n                                              recurrent_initializer='glorot_uniform')\nelse:\n    return tf.keras.layers.LSTM(units, \n                               return_sequences=True,\n                               return_state=True,\n                               recurrent_activation='sigmoid',\n                               reset_after='True',\n                               recurrent_initializer='glorot_uniform')\n<\/code><\/pre>\n\n<p>So by just using <strong>reset_after=True<\/strong> and <strong>recurrent_activation='sigmoid'<\/strong> options, my problem got solved.<\/p>\n"}]}
+{"QuestionId":57551783,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-08-19T06:55:21.533","AcceptedAnswerId":57553977.0,"OwnerUserId":3696633.0,"Title":"Pytorch gradients has not calculated","Body":"<p>I create a NN. I'm having a problem with recounting gradients. The problem is that I scalarly multiply 2 tensors <code>u @ v<\/code> and normalize one of them. It is important that gradients cannot be calculated for h. Therefore, I use <code>detach()<\/code>. In addition, during the recalculation of gradients, normalization should not be taken into account (I do not know how to do this).<\/p>\n\n<pre><code>import torch\nfrom torch import nn\n\n\nclass Nn(nn.Module):\n    def __init__(self):\n        super(Nn, self).__init__()\n        self.ln = nn.Linear(5, 5)\n\n    def forward(self, x):\n        v = self.ln(x)\n\n        u = v.clone()\n        h = v.clone()\n\n        u \/= u.norm()\n        h = h.detach()\n        h \/= h.norm()\n\n        res = torch.stack([torch.stack([u @ h, u @ h])])\n\n        return res\n\n\ndef patches_generator():\n    while True:\n        decoder = torch.rand((5, ))\n        target = torch.randint(2, (1,))\n        yield decoder, target\n\n\nnet = Nn()\n\ncriterion = nn.CrossEntropyLoss()\noptimizer = torch.optim.Adam(net.parameters())\n\nnet.train()\ntorch.autograd.set_detect_anomaly(True)\nfor decoder, targets in patches_generator():\n    optimizer.zero_grad()\n    outputs = net(decoder)\n    loss = criterion(outputs, targets)\n    loss.backward()\n    optimizer.step()\n<\/code><\/pre>\n\n<p>As a result, I get the following error:<\/p>\n\n<blockquote>\n  <p>RuntimeError: one of the variables needed for gradient computation has\n  been modified by an inplace operation: [torch.FloatTensor [9, 512, 1,\n  1]], which is output 0 of ReluBackward1, is at version 3; expected\n  version 2 instead. Hint: the backtrace further above shows the\n  operation that failed to compute its gradient. The variable in\n  question was changed in there or anywhere later. Good luck!<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"57553977","CreationDate":"2019-08-19T09:31:45.903","ParentId":null,"OwnerUserId":"2650622","Title":null,"Body":"<p>The problem is the in-place division operator applied to <code>u<\/code> in this line:<\/p>\n\n<pre><code>u \/= u.norm()\n<\/code><\/pre>\n\n<p>changing it to<\/p>\n\n<pre><code>u = u \/ u.norm()\n<\/code><\/pre>\n\n<p>makes the code run. The reason is that the in-place operator overwrites the intermediate result from this line<\/p>\n\n<pre><code>u = v.clone()\n<\/code><\/pre>\n\n<p>which makes it impossible for Pytorch to compute the gradient.<\/p>\n\n<p>(The error message in the question contains a reference to a <code>ReluBackward1<\/code> layer which is not in the reduced code example. <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.ReLU\" rel=\"nofollow noreferrer\">Pytorch ReLU<\/a> layers have an optional <code>in_place<\/code> argument which makes the operation in place while supporting backprop. This often works, because in a sequential network there is no need to distinguish between the output of the ReLU activation and the output of the weights to compute the gradient, but in more complex architectures it might be necessary to retain the output of the weights.)<\/p>\n"}]}
+{"QuestionId":57551786,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-08-19T06:55:29.003","AcceptedAnswerId":null,"OwnerUserId":7078575.0,"Title":"How to save trained FCN model and test new image on trained FCN model?","Body":"<p>I am using this code to train FCN, I have succssfully run this code. However, I want to test new images on this trained model, can anyone help me?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#Training\n\nfrom keras import optimizers\n\nsgd = optimizers.SGD(lr=1E-2, decay=5**(-4), momentum=0.9, nesterov=True)\nmodel.compile(loss='categorical_crossentropy',\noptimizer=sgd,\nmetrics=['accuracy'])\n\nhist1 = model.fit(X_train,y_train,\nvalidation_data=(X_test,y_test),\nbatch_size=2,epochs=20,verbose=1)\n\nfor key in ['loss', 'acc', 'val_loss', 'val_acc']:\nplt.plot(hist1.history[key],label=key)\nplt.legend()\nplt.show()\n\n\ny_pred = model.predict(X_test)\ny_predi = np.argmax(y_pred, axis=3)\ny_testi = np.argmax(y_test, axis=3)\nprint(y_testi.shape,y_predi.shape)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57554976","CreationDate":"2019-08-19T10:31:59.633","ParentId":null,"OwnerUserId":"11939354","Title":null,"Body":"<p>As stated in the keras FAQs (<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\n<p>To save your model to your disk you can do:<\/p>\n\n<pre><code>model.save(\"my_model_file.h5\")\n<\/code><\/pre>\n\n<p>And to load it again later or in another file for using it:<\/p>\n\n<pre><code>from keras.models import load_model\n\nmodel = load_model(\"my_model_file.h5\")\ny_pred = model.predict(X_test)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57552575,"AnswerCount":1,"Tags":"<pytorch><slice>","CreationDate":"2019-08-19T07:57:21.987","AcceptedAnswerId":null,"OwnerUserId":10855529.0,"Title":"Explanation for slicing in Pytorch","Body":"<p>why is the output same every time?<\/p>\n\n<pre><code>a = torch.tensor([0, 1, 2, 3, 4])\na[-2:] = torch.tensor([[[5, 6]]])\na\n<\/code><\/pre>\n\n<p>tensor([0, 1, 2, 5, 6])<\/p>\n\n<pre><code>a = torch.tensor([0, 1, 2, 3, 4])\na[-2:] = torch.tensor([[5, 6]])\na\n<\/code><\/pre>\n\n<p>tensor([0, 1, 2, 5, 6])<\/p>\n\n<pre><code>a = torch.tensor([0, 1, 2, 3, 4])\na[-2:] = torch.tensor([5, 6])\na\n<\/code><\/pre>\n\n<p>tensor([0, 1, 2, 5, 6])<\/p>\n","answers":[{"AnswerId":"57553058","CreationDate":"2019-08-19T08:33:41.000","ParentId":null,"OwnerUserId":"2650622","Title":null,"Body":"<p>Pytorch is following Numpy here which <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/user\/basics.indexing.html#assigning-values-to-indexed-arrays\" rel=\"nofollow noreferrer\">allows assignment<\/a> to slices as long as the shapes are compatible meaning that the two sides have the same shape or the right hand side is broadcastable to the shape of the slice. Starting with trailing dimensions, two arrays are <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/user\/basics.broadcasting.html\" rel=\"nofollow noreferrer\">broadcastable<\/a> if they only differ in dimensions where one of them is 1. So in this case<\/p>\n\n<pre><code>a = torch.tensor([0, 1, 2, 3, 4])\nb = torch.tensor([[[5, 6]]])\nprint(a[-2:].shape, b.shape)\n&gt;&gt; torch.Size([2]) torch.Size([1, 1, 2])\n<\/code><\/pre>\n\n<p>Pytorch will perform the following comparisons:<\/p>\n\n<ol>\n<li><code>a[-2:].shape[-1]<\/code> and <code>b.shape[-1]<\/code> are equal so the last dimension is compatible<\/li>\n<li><code>a[-2:].shape[-2]<\/code> does not exist, but <code>b.shape[-2]<\/code> is 1 so they are compatible<\/li>\n<li><code>a[-2:].shape[-3]<\/code> does not exist, but <code>b.shape[-3]<\/code> is 1 so they are compatible<\/li>\n<li>All dimensions are compatible, so  <code>b<\/code> can be broadcasted to <code>a<\/code><\/li>\n<\/ol>\n\n<p>Finally, Pytorch will convert <code>b<\/code> to <code>tensor([5, 6])<\/code> before performing the assignment thus producing the result:<\/p>\n\n<pre><code>a[-2:] = b\nprint(a)\n&gt;&gt; tensor([0, 1, 2, 5, 6])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57553272,"AnswerCount":0,"Tags":"<python><tensorflow><probability-distribution>","CreationDate":"2019-08-19T08:46:49.520","AcceptedAnswerId":null,"OwnerUserId":3284573.0,"Title":"TensorFlow: resampling to make an unbiased estimate of a categorical distribution","Body":"<p>I have a probability distribution over (~2000) different objects that looks like the following table. Each object is associated with a particular probability. In TensorFlow, objects are stored as indices, and probability is a 1D tensor.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/nwWgW.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/nwWgW.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The probability distribution is too big, and I want to resample it and use a smaller distribution to represent it. I want the resulting distribution to be unbiased, and to have around ~500 objects and probabilities. Some objects need to be discarded. <\/p>\n\n<p>Is there any way in TensorFlow to achieve this purpose?<\/p>\n","answers":[]}
+{"QuestionId":57553280,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><machine-learning>","CreationDate":"2019-08-19T08:47:15.753","AcceptedAnswerId":null,"OwnerUserId":6997665.0,"Title":"How to use a loss function that is not differentiable?","Body":"<p>I am trying to find a codebook at the output of a fully connected neural network which chooses points such that the minimum distance (Euclidean norm) between the so produced codebook is maximized. The input to the neural network is the points that need to be mapped into higher dimension of the output space.<\/p>\n\n<p>For instance, if the input dimension is 2 and output dimension is 3, the following mapping (and any permutations) works best:\n 00 - 000,\n 01 - 011,\n 10 - 101,\n 11 - 110<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport itertools\n\n\ninput_bits = tf.placeholder(dtype=tf.float32, shape=[None, 2], name='input_bits')\ncode_out = tf.placeholder(dtype=tf.float32, shape=[None, 3], name='code_out')\nnp.random.seed(1331)\n\n\ndef find_code(message):\n    weight1 = np.random.normal(loc=0.0, scale=0.01, size=[2, 3])\n    init1 = tf.constant_initializer(weight1)\n    out = tf.layers.dense(inputs=message, units=3, activation=tf.nn.sigmoid, kernel_initializer=init1)\n    return out\n\n\ncode = find_code(input_bits)\n\ndistances = []\nfor i in range(0, 3):\n    for j in range(i+1, 3):\n        distances.append(tf.linalg.norm(code_out[i]-code_out[j]))\nmin_dist = tf.reduce_min(distances)\n# avg_dist = tf.reduce_mean(distances)\n\nloss = -min_dist\n\nopt = tf.train.AdamOptimizer().minimize(loss)\n\ninit_variables = tf.global_variables_initializer()\nsess = tf.Session()\nsess.run(init_variables)\n\nsaver = tf.train.Saver()\n\ncount = int(1e4)\n\nfor i in range(count):\n    input_bit = [list(k) for k in itertools.product([0, 1], repeat=2)]\n    code_preview = sess.run(code, feed_dict={input_bits: input_bit})\n    sess.run(opt, feed_dict={input_bits: input_bit, code_out: code_preview})\n<\/code><\/pre>\n\n<p>Since the loss function itself is not differentiable, I am getting the 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 \n<\/code><\/pre>\n\n<p>Am I doing something silly or is there a way to circumvent this? Any help in this regard is appreciated. Thanks in advance.<\/p>\n","answers":[{"AnswerId":"57554525","CreationDate":"2019-08-19T10:03:56.790","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>Your loss function has to be differentiable with respect to some parameters. In your case, there are no parameters therefore you would be computing a derivative of a constant function which is 0. Moreover, in your code you have the following line:<\/p>\n\n<pre><code>code = find_code(input_bits)\n<\/code><\/pre>\n\n<p>which is not used any further. Based on the code, I assume that you want to change this line:<\/p>\n\n<pre><code>distances.append(tf.linalg.norm(code_out[i]-code_out[j]))\n<\/code><\/pre>\n\n<p>to:<\/p>\n\n<pre><code>distances.append(tf.linalg.norm(code[i]-code_out[j]))\n<\/code><\/pre>\n\n<p>Therefore, you would be using the <code>tf.layers.dense<\/code> that you have, thus include a parameter which can be used to compute the gradient of the loss with respect to that parameter. <\/p>\n\n<hr>\n\n<p>Moreover, you don't need to worry whether a TF operation is differentiable or not. In fact, all TF ops are differentiable. When it comes to the <code>tf.reduce_min()<\/code>, please check <a href=\"https:\/\/datascience.stackexchange.com\/questions\/21871\/how-does-tensorflow-compute-gradients-of-reduce-min-operation\">this link<\/a>.<\/p>\n"}]}
+{"QuestionId":57553423,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><tensorflow-serving>","CreationDate":"2019-08-19T08:57:21.460","AcceptedAnswerId":null,"OwnerUserId":6482145.0,"Title":"How to pass single image in tensorflow serving?","Body":"<p>I have a yolo model hosted on tensorflow serving on localhost, to test it out ,I have to pass single image instead of batch of images,this is how I am trying by <\/p>\n\n<pre><code>from tensorflow_serving.apis import predict_pb2\nfrom tensorflow_serving.apis import prediction_service_pb2_grpc\nimport cv2 as cv\nchannel = grpc.insecure_channel(\"8501\")\n\nstub = prediction_service_pb2_grpc.PredictionServiceStub(channel)\n\nrequest.model_spec.name = 'yolo'\nrequest.model_spec.signature_name = 'predict'\nimage = cv.imread(\"test.jpg\")\n<\/code><\/pre>\n\n<p>Now after this when Im trying to assign inputs<\/p>\n\n<pre><code>&gt;&gt;&gt; request.inputs['images'].CopyFrom(image)\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\nTypeError: Parameter to CopyFrom() must be instance of same class: expected tensorflow.TensorProto got numpy.ndarray.\n<\/code><\/pre>\n\n<p>I am getting this error,how can I pass single image in request ?<\/p>\n","answers":[{"AnswerId":"57612405","CreationDate":"2019-08-22T15:09:48.730","ParentId":null,"OwnerUserId":"10340061","Title":null,"Body":"<pre><code>request.inputs['image'].CopyFrom(\n    tf.contrib.util.make_tensor_proto(image, shape = [1]))\nresult = stub.Predict(request, 10.0)\n<\/code><\/pre>\n\n<p>Try this<\/p>\n"}]}
+{"QuestionId":57553489,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2019-08-19T09:01:20.627","AcceptedAnswerId":null,"OwnerUserId":7673078.0,"Title":"Tensor name has no shape information ERROR when predicting with tensorflow serving","Body":"<p>I'm using tensorflow serving to serve a savedmodel. I have two signatures: 1st outputting keras model.output and the 2nd outputting post processing of model.output. When I try a predict call of the 2nd signature on tensorflow serving it is giving me an error <code>{ \"error\": \"Tensor name: prediction has no shape information \" }<\/code><\/p>\n\n<p>this is the code to build the savedmodel<\/p>\n\n<pre><code>    shape1 = 92\n    shape2 = 92\n    reg=0.000001\n    learning_rate=0.001\n\n    sess = tf.Session()\n    K.set_session(sess)\n    K._LEARNING_PHASE = tf.constant(0)\n    K.set_learning_phase(0)\n\n\n    #preprocessing\n    x_input = tf.placeholder(tf.string, name='x_input', shape=[None])\n    reshaped = tf.reshape(x_input, shape=[])\n    image = tf.image.decode_jpeg(reshaped, channels=3)\n    image2 = tf.expand_dims(image,0)\n    resized = tf.image.resize_images(image2, (92,92))\n    meaned = tf.math.subtract(resized, tf.constant(116.0))\n    normalized = tf.math.divide(meaned, tf.constant(66.0))\n\n    #keras model\n    model = tf.keras.Sequential()\n    model.add(InputLayer(input_tensor=normalized))\n    model.add(Conv2D(32, (3, 3), padding='same', activation='relu', kernel_regularizer=l2(reg)))\n    model.add(Conv2D(32, (3, 3), padding='same', activation='relu', kernel_regularizer=l2(reg)))\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n\n    model.add(Conv2D(64, (3, 3), padding='same', activation='relu', kernel_regularizer=l2(reg)))\n    model.add(Dropout(0.1))\n    model.add(Conv2D(64, (3, 3), padding='same', activation='relu', kernel_regularizer=l2(reg)))\n    model.add(Dropout(0.1))\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n\n    model.add(Conv2D(128, (3, 3), padding='same', activation='relu', kernel_regularizer=l2(reg)))\n    model.add(Dropout(0.2))\n    model.add(Conv2D(128, (3, 3), padding='same', activation='relu', kernel_regularizer=l2(reg)))\n    model.add(Dropout(0.2))\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n\n    model.add(Conv2D(256, (3, 3), padding='same', activation='relu', kernel_regularizer=l2(reg)))\n    model.add(Dropout(0.3))\n    model.add(Conv2D(256, (3, 3), padding='same', activation='relu', kernel_regularizer=l2(reg)))\n    model.add(Dropout(0.3))\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n\n    model.add(Flatten())\n    model.add(Dense(256, activation='relu', kernel_regularizer=l2(reg)))\n    model.add(Dropout(0.5))\n\n    model.add(Dense(256, activation='relu', kernel_regularizer=l2(reg)))\n    model.add(Dropout(0.5))\n\n    model.add(Dense(1))\n    model.add(Activation('sigmoid'))\n\n    model.compile(loss='binary_crossentropy',\n                    optimizer=tf.train.RMSPropOptimizer(learning_rate=learning_rate),\n                    metrics=['accuracy'])\n\n    #post processing to output label\n    pred = tf.gather_nd(model.output, (0,0))\n    label = tf.cond(pred &gt; 0.5, lambda: tf.constant('Dog', shape=[]), lambda: tf.constant('Cat', shape=[]))\n\n    model.load_weights(r'.\/checkpoints\/4.ckpt')\n    export_path = '.\/saved_models\/1'\n\n\n    init_op = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer())\n    sess.run(init_op)\n    model.load_weights(r'.\/checkpoints\/4.ckpt')\n\n    if os.path.isdir(export_path):\n        print('\\nAlready saved a model, cleaning up\\n')\n        print(subprocess.run(['rm', '-r', export_path]))\n\n    #first signature(this works)\n    x_info = tf.saved_model.utils.build_tensor_info(x_input)\n    y_info = tf.saved_model.utils.build_tensor_info(model.output)\n    sigmoid_signature = build_signature_def(inputs={\"image\": x_info}, outputs={\"prediction\":y_info}, method_name='tensorflow\/serving\/predict')\n\n    #2nd signature(this doesn't work)\n    x_info = tf.saved_model.utils.build_tensor_info(x_input)\n    y_info = tf.saved_model.utils.build_tensor_info(label)\n    label_signature = build_signature_def(inputs={\"image\": x_info}, outputs={\"prediction\":y_info}, method_name='tensorflow\/serving\/predict')\n\n    builder = tf.saved_model.builder.SavedModelBuilder(export_path)\n\n    legacy_init_op = tf.group(tf.tables_initializer(), name='legacy_init_op')\n\n    builder.add_meta_graph_and_variables(sess=sess,\n                                     tags=[\"serve\"],\n                                     signature_def_map={'sigmoid': sigmoid_signature, 'label': label_signature})\n    builder.save()\n\n<\/code><\/pre>\n\n<p>this is code to call tf serving<\/p>\n\n<pre><code>imgs = ['.\/Dog\/' + img for img in imgs]\nimg = open('.\/Dog\/3.jpg', 'rb').read()\n\nimg = base64.b64encode(img).decode('utf-8')\n\ndata = json.dumps(\n        {\"signature_name\": \"label\",\n         \"instances\": [\n                 {'image': {'b64': img}}\n                 ]\n        }\n)\njson_response = requests.post('http:\/\/localhost:8501\/v1\/models\/pet:predict', data=data)\nprint(json_response.text)\n<\/code><\/pre>\n\n<p>Instead of getting a response of <code>{\"predictions\": \"Dog\"}<\/code>, i am getting an error <code>{ \"error\": \"Tensor name: prediction has no shape information \" }<\/code><\/p>\n","answers":[{"AnswerId":"57640367","CreationDate":"2019-08-24T17:58:58.477","ParentId":null,"OwnerUserId":"7673078","Title":null,"Body":"<p>I managed to fix this. I used tf.reshape on what i wanted to output and passed that into the signature builder.<\/p>\n\n<pre><code>#post processing to output label\n    pred = tf.gather_nd(model.output, (0,0))\n    label = tf.cond(pred &gt; 0.5, lambda: tf.constant('Dog', shape=[]), lambda: tf.constant('Cat', shape=[]))\n    label_reshaped = tf.reshape(label, [None])\n\n...\n\n#2nd signature(this doesn't work)\n    x_info = tf.saved_model.utils.build_tensor_info(x_input)\n    y_info = tf.saved_model.utils.build_tensor_info(label_reshaped)\n    label_signature = build_signature_def(inputs={\"image\": x_info}, outputs={\"prediction\":y_info}, method_name='tensorflow\/serving\/predict')\n<\/code><\/pre>\n"}]}
+{"QuestionId":57553877,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-datasets>","CreationDate":"2019-08-19T09:26:17.677","AcceptedAnswerId":null,"OwnerUserId":8842694.0,"Title":"Is it possible to shuffle the dataset using the index of its elements?","Body":"<p>I am using <code>tf.data.experimental.make_csv_dataset<\/code> in tensorflow (TF1.14 and TF2.0) to read a csv file consisting 3 columns; index, column1 and column2. For me only column 1 and column2 are important. Each element in column1 is an array of shape (1,4) and column2 has (1,1). On this dataset, when I use <code>tf.data.shuffle(buffer_size = some_number)<\/code> for shuffling, it takes a lot of time to do this shuffling with a message <code>Filling Up the shuffle buffer<\/code>. My question is if there is a way to shuffle the dataset by using the indices of the column1\/column2, because this might not take so much time for shuffling since it is only the indices.<\/p>\n","answers":[{"AnswerId":"57555885","CreationDate":"2019-08-19T11:31:56.627","ParentId":null,"OwnerUserId":"2455494","Title":null,"Body":"<blockquote>\n  <p>My question is if there is a way to shuffle the dataset by using the indices of the column1\/column2, because this might not take so much time for shuffling since it is only the indices<\/p>\n<\/blockquote>\n\n<p>No, unfortunately not. Not in that way. <\/p>\n\n<p>The reason is that a <code>tf.data.Dataset<\/code> object is inherently lazily loaded. It is deliberately so as it can represent arbitrarily large (even infinite) datasets so it wouldnt make sense to try to load it all into memory or do all the pre-processing up front.<\/p>\n\n<p>This means that, whilst it would (of course) be feasible to read and shuffle the index we could not then access the nth element from the original dataset (at least not cheaply).<\/p>\n\n<p>It's worth mentioning that the shuffle buffer only needs to be filled once so the delay will only happen at the start of training (and the start of each epoch if shuffling each epoch).<\/p>\n\n<p>A sensible workaround that you may well have already considered is to load the dataset once with the shuffle then write it out somewhere (maybe a tfrecord format) with all the rows preshuffled.<\/p>\n"}]}
+{"QuestionId":57554155,"AnswerCount":0,"Tags":"<c#><python><visual-studio><keras>","CreationDate":"2019-08-19T09:43:47.083","AcceptedAnswerId":null,"OwnerUserId":11901042.0,"Title":"DLLNotFoundException: 'python3.6m'","Body":"<p>I'm trying to load an image in c# using python. As it is a data science project.<\/p>\n\n<p>I've tried to run this example: <a href=\"https:\/\/github.com\/SciSharp\/Keras.NET\/blob\/master\/Examples\/ImageExamples\/Cifar10_CNN.cs\" rel=\"nofollow noreferrer\">https:\/\/github.com\/SciSharp\/Keras.NET\/blob\/master\/Examples\/ImageExamples\/Cifar10_CNN.cs<\/a> but I can errors with dll. The error is in the first line as in title. Before that error I got another one saying that there is no python36.dll after downloading it I can't get through it.<\/p>\n\n<pre class=\"lang-cs prettyprint-override\"><code>var img = ImageUtil.LoadImg(path, target_size: new Shape(32, 32));\nNDarray x = ImageUtil.ImageToArray(img);\nx = x.reshape(1, x.shape[0], x.shape[1], x.shape[2]);\n<\/code><\/pre>\n\n<p>Is there anything else I need to install to my project?<\/p>\n","answers":[]}
+{"QuestionId":57554413,"AnswerCount":2,"Tags":"<python><tensorflow><keras><conv-neural-network>","CreationDate":"2019-08-19T09:57:08.340","AcceptedAnswerId":57554644.0,"OwnerUserId":8895481.0,"Title":"How does it works the input_shape variable in Conv1d in Keras?","Body":"<p>Ciao,\nI'm working with CNN 1d on Keras but I have tons of troubles with the input shape variable.<\/p>\n\n<p>I have a time series of 100 timesteps and 5 features with boolean labels. I want to train a CNN 1d that works with a sliding window of length 10. This is a very simple code I wrote:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.models import Sequential\nfrom keras.layers import Dense, Conv1D\nimport numpy as np\n\nN_FEATURES=5\nN_TIMESTEPS=10\nX = np.random.rand((100, N_FEATURES))\nY = np.random.randint(0,2, size=100)\n\n\n# CNN\nmodel.Sequential()\nmodel.add(Conv1D(filter=32, kernel_size=N_TIMESTEPS, activation='relu', input_shape=N_FEATURES\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>My problem here is that I get the following error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>  File \"&lt;ipython-input-2-43966a5809bd&gt;\", line 2, in &lt;module&gt;\n    model.add(Conv1D(filter=32, kernel_size=10, activation='relu', input_shape=N_FEATURES))\nTypeError: __init__() takes at least 3 arguments (3 given)\n<\/code><\/pre>\n\n<p>I've also tried by passing to the input_shape the following values:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>input_shape=(None, N_FEATURES)\ninput_shape=(1, N_FEATURES)\ninput_shape=(N_FEATURES, None)\ninput_shape=(N_FEATURES, 1)\ninput_shape=(N_FEATURES, )\n<\/code><\/pre>\n\n<blockquote>\n  <p>Do you know what's wrong with the code or in general can you explain the logic behind in <strong>input_shape<\/strong> variable in Keras CNN?<\/p>\n<\/blockquote>\n\n<p>The crazy thing is that my problem is the same of the following:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/53465066\/keras-cnn-error-expected-sequence-to-have-3-dimensions-but-got-array-with-shap\">Keras CNN Error: expected Sequence to have 3 dimensions, but got array with shape (500, 400)<\/a><\/p>\n\n<p>But I cannot solve it with the solution given in the post.<\/p>\n\n<blockquote>\n  <p>The Keras version is <strong>2.0.6-tf<\/strong><\/p>\n<\/blockquote>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"57554608","CreationDate":"2019-08-19T10:09:29.157","ParentId":null,"OwnerUserId":"10908375","Title":null,"Body":"<p>Your code is missing parentheses, and has spelling mistakes in arguments. Input shape needs to be an iterable, so try (N_FEATURES,) instead. <\/p>\n"},{"AnswerId":"57554644","CreationDate":"2019-08-19T10:11:42.303","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>This should work:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense, Conv1D\nimport numpy as np\n\nN_FEATURES=5\nN_TIMESTEPS=10\nX = np.random.rand(100, N_FEATURES)\nY = np.random.randint(0,2, size=100)\n\n# Create a Sequential model\nmodel = Sequential()\n# Change the input shape to input_shape=(N_TIMESTEPS, N_FEATURES)\nmodel.add(Conv1D(filters=32, kernel_size=N_TIMESTEPS, activation='relu', input_shape=(N_TIMESTEPS, N_FEATURES)))\n# If it is a binary classification then you want 1 neuron - Dense(1, activation='sigmoid')\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>Please see the comments before each line of code. Moreover, the input shape that <code>Conv1D<\/code> expects is <code>(time_steps, feature_size_per_time_step)<\/code>. The translation of that for your code is <code>(N_TIMESTEPS, N_FEATURES)<\/code>.<\/p>\n"}]}
+{"QuestionId":57555009,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-19T10:34:07.137","AcceptedAnswerId":null,"OwnerUserId":2443944.0,"Title":"Tensorflow iterating through tensor","Body":"<p>Im using Tensorflow 1.13. However I get an error saying that I can't iterate through a tensor unless I'm in eager mode. Is there a way of doing this without going into eager mode?<\/p>\n\n<pre><code>with tf.Session(config=config) as sess:\n    context = tf.placeholder(tf.int32, [args.batch_size, None])\n    mask = tf.placeholder(tf.int32, [args.batch_size, 2])\n    output = model.model(hparams=hparams, X=context)\n\n\n\n    for batch_index in range(args.batch_size):\n        start = mask[batch_index][0]\n        end   = mask[batch_index][1]\n\n        for i in range(start, end+1):\n            output['logits'][batch_index, i , context[batch_index,i]].assign(math.inf)\n\n    loss = tf.reduce_mean(\n        tf.nn.sparse_softmax_cross_entropy_with_logits(\n            labels=context[:, 1:],  logits=output['logits'][:, :-1]))\n<\/code><\/pre>\n","answers":[{"AnswerId":"57555285","CreationDate":"2019-08-19T10:52:04.223","ParentId":null,"OwnerUserId":"8889727","Title":null,"Body":"<p>Can you try using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/while_loop\" rel=\"nofollow noreferrer\">tf.while_loop<\/a>? You could try the following snippet (possibly with minor modifications for your code) and see whether it works?<\/p>\n\n<pre><code>import tensorflow as tf\ngpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.9)\nwith tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:\n    context = tf.placeholder(tf.int32, [args.batch_size, None])\n    mask = tf.placeholder(tf.int32, [args.batch_size, 2])\n    output = model.model(hparams=hparams, X=context)\n\n\n\n    for batch_index in [0,1,2,3]: #I have assumed a dummy list cz we can't iterate through a 'Dimension'\n        start = mask[batch_index][0]\n        end   = mask[batch_index][1]\n\n        i = tf.constant(0)\n        while_condition = lambda i: (tf.less(i, end)) &amp; (tf.math.greater_equal(i,start))\n\n        def body(i):\n            return output['logits'][batch, i , context[batch,i]].assign(math.inf)\n\n        r = tf.while_loop(while_condition, body, [i])\n\n        # for i in range(start, end+1):\n        #     output['logits'][batch, i , context[batch,i]].assign(math.inf)\n\n    loss = tf.reduce_mean(\n        tf.nn.sparse_softmax_cross_entropy_with_logits(\n            labels=context[:, 1:],  logits=output['logits'][:, :-1]))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57555019,"AnswerCount":1,"Tags":"<install><pytorch><conda>","CreationDate":"2019-08-19T10:34:42.187","AcceptedAnswerId":null,"OwnerUserId":420867.0,"Title":"`conda install pytorch` results in 'Illegal instruction' with `import torch`","Body":"<p>With a fresh install of anaconda, followed by <code>conda install pytorch<\/code>, I get an <code>Illegal instruction<\/code> (crashing the python interpreter) whenever I <code>import torch<\/code>:<\/p>\n\n<pre><code>(base) ~$ python -c 'import torch'\nIllegal instruction\n(base) 132 ~$\n<\/code><\/pre>\n","answers":[{"AnswerId":"57555020","CreationDate":"2019-08-19T10:34:42.187","ParentId":null,"OwnerUserId":"420867","Title":null,"Body":"<p>The problem in my case is that it's a CPU machine and thus doesn't have CUDA installed. In this case you need to install <code>pytorch-cpu<\/code> instead:<\/p>\n\n<pre><code>conda install -c pytorch pytorch-cpu\n<\/code><\/pre>\n\n<p>After that, it works fine. <\/p>\n\n<p>(Thanks to <a href=\"https:\/\/discuss.pytorch.org\/t\/installing-pytorch-without-a-gpu\/22455\/2?u=drevicko\" rel=\"nofollow noreferrer\">@JuanFMontesinos on the pytorch forum<\/a>)<\/p>\n"}]}
+{"QuestionId":57555134,"AnswerCount":1,"Tags":"<python><pytorch><tensor><elementwise-operations>","CreationDate":"2019-08-19T10:42:00.827","AcceptedAnswerId":57558634.0,"OwnerUserId":6839356.0,"Title":"Pytorch batch matrix-matrix outer product","Body":"<p>Similarly to the question in <a href=\"https:\/\/stackoverflow.com\/questions\/55054127\/pytorch-batch-matrix-vector-outer-product\">Pytorch batch matrix vector outer product<\/a> I have two matrices and would like to compute their outer product, or in other words the pairwise elementwise product.<\/p>\n\n<p>Shape example:\nIf we have X1 and X2 of shapes of <code>torch.Size([32, 300, 8])<\/code>\nThe result should be of size <code>torch.Size([32, 300, 300, 8])<\/code><\/p>\n","answers":[{"AnswerId":"57558634","CreationDate":"2019-08-19T14:18:23.960","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can add singleton dimensions:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>X1[:, None, ...] * X1[..., None, :]\n<\/code><\/pre>\n\n<p>But Usman Ali's <a href=\"https:\/\/stackoverflow.com\/questions\/57555134\/pytorch-batch-matrix-matrix-outer-product#comment101572505_57555134\">comment<\/a> is also a good idea. Use <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/torch.html#torch.einsum\" rel=\"nofollow noreferrer\"><code>torch.einsum<\/code><\/a>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>torch.einsum('bik,bjk-&gt;bijk', X1, X2)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57555407,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-08-19T10:59:30.590","AcceptedAnswerId":57555461.0,"OwnerUserId":519882.0,"Title":"'numpy.dtype' object has no attribute 'base_dtype' in keras","Body":"<p>I have below code which I am trying to understand keras mean  and want to get pooled_grads print. While printing I am getting below error<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\narr3 = np.array([ [\n                   [1,2,3],\n                   [4,5,6]\n                 ],\n                 [\n                   [1,2,3],\n                   [4,5,6]\n                 ],\n                 [\n                   [1,2,3],\n                   [4,5,6]\n                 ]\n                ] \n               )\n\n#print(\"Arr shape\", arr3.shape)\n\nimport keras.backend as K\n\nimport numpy as np\n\npooled_grads = K.mean(arr3, axis=(0, 1, 2))\n\nprint(\"------------------------\")\n\nprint(pooled_grads)\n<\/code><\/pre>\n\n<p>I am getting below error<\/p>\n\n<p>AttributeError: 'numpy.dtype' object has no attribute 'base_dtype'<\/p>\n","answers":[{"AnswerId":"57555461","CreationDate":"2019-08-19T11:03:25.087","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Most Keras backend functions expect Keras tensors as inputs. If you want to use a NumPy array as input, convert it to a tensor first, for example with <code>K.constant<\/code>:<\/p>\n\n<pre><code>pooled_grads = K.mean(K.constant(arr3), axis=(0, 1, 2))\n<\/code><\/pre>\n\n<p>Note that <code>pooled_grads<\/code> here will be another tensor, so printing it will not give you the value directly, but just a reference to the tensor object. In order to get the value of the tensor, you can use for example <code>K.get_value<\/code>:<\/p>\n\n<pre><code>print(K.get_value(pooled_grads))\n# 3.5\n<\/code><\/pre>\n"}]}
+{"QuestionId":57555643,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-lite>","CreationDate":"2019-08-19T11:15:13.267","AcceptedAnswerId":null,"OwnerUserId":1361043.0,"Title":"Tensors lose their dimensions during conversion from TF to TFlite","Body":"<p>I'm converting a tensorflow model to tflite model (through python script) and although all input tensors are passed as input in the tflite some of them in tflite do not have any dimensions. <\/p>\n\n<p>So in TF I have 5 input tensors (4D, 4D, 1D, 1D, 1D) and when I convert it to tflite I get 5 tensors (4D, 4D, 1D, 0D, 0D). I have no problem with the output tensors<\/p>\n\n<p>I'm thinking something is happening inside <code>toco_tooling.cc<\/code> or some other file<\/p>\n\n<p>some help?<\/p>\n","answers":[]}
+{"QuestionId":57556070,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-08-19T11:44:27.950","AcceptedAnswerId":null,"OwnerUserId":6300254.0,"Title":"How to import keras model h5 to pb (tensorflow)","Body":"<p>I am try to save a keras model to pb file, so I can run the model using C++ tensorflow.<\/p>\n\n<p>I am trying to use a solution from internet to freeze the model. I can save the model, but when I load the model in tensorflow, the prediction doesn't work.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import os\nfrom keras import backend as K\nfrom keras.models import load_model\nfrom keras.engine.saving import model_from_json, load_model\nfrom keras.preprocessing import image\nimport numpy as np\nimport matplotlib.pyplot as plt\n\ndef load_image(img_path):\n    img = image.load_img(img_path, target_size=(128, 128))\n    img_tensor = image.img_to_array(img)                    # (height, width, channels)\n    img_tensor = np.expand_dims(img_tensor, axis=0)         # (1, height, width, channels), add a dimension because the model expects this shape: (batch_size, height, width, channels)\n    img_tensor \/= 255.                                      # imshow expects values in the range [0, 1]\n\n    return img_tensor\n\n# This line must be executed before loading Keras model.\nK.set_learning_phase(0)\n\njson_file = open('\/home\/antonio\/keras_to_tensorflow\/modelAlex\/output_model2.json', 'r')\nloaded_model_json = json_file.read()\njson_file.close()\nmodel = model_from_json(loaded_model_json)\n\nmodel.load_weights(\"\/home\/antonio\/keras_to_tensorflow\/modelAlex\/output_model2.h5\")\n\nfrom keras import backend as K\nimport tensorflow as tf\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    from tensorflow.python.framework.graph_util import convert_variables_to_constants\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        # Graph -&gt; GraphDef ProtoBuf\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 = convert_variables_to_constants(session, input_graph_def,\n                                                      output_names, freeze_var_names)\n        return frozen_graph\n\n\nfrozen_graph = freeze_session(K.get_session(), output_names=[out.op.name for out in model.outputs])\n\ntf.train.write_graph(frozen_graph, \"model\", \".\/model\/tf_model.pb\", as_text=False)\n\nimport tensorflow as tf\nfrom tensorflow.python.platform import gfile\n\nf = gfile.FastGFile(\".\/model\/tf_model.pb\", 'rb')\ngraph_def = tf.GraphDef()\n# Parses a serialized binary message into the current message.\ngraph_def.ParseFromString(f.read())\nf.close()\n\ninit_op = tf.initialize_all_variables()\n\nwith tf.Session() as sess:\n    sess.run(init_op)\n    sess.graph.as_default()\n    tf.import_graph_def(graph_def)\n\n    files = ['image.png']\n\n    for fileName in files:\n        img_path = 'data\/images\/' + fileName\n        x_test = load_image(img_path)\n\n        softmax_tensor = sess.graph.get_tensor_by_name('dense_4\/Softmax:0')\n        predictions = sess.run(softmax_tensor, {'input_1:0': x_test})\n\n        print(predictions)\n<\/code><\/pre>\n\n<p>The model produce a prediction, but it is wrong or random.<\/p>\n","answers":[]}
+{"QuestionId":57556974,"AnswerCount":1,"Tags":"<python><tensorflow><normalization><weighted>","CreationDate":"2019-08-19T12:40:35.767","AcceptedAnswerId":null,"OwnerUserId":11243873.0,"Title":"Weight_column value normalization","Body":"<p>I use a TensorFlow canned estimator to predict game actions from situations favourizing best scores (there are no \"good\" or \"bad\" answers but an answer ib better than another depending the situation). This reality is represented by scores for each situation \/ action sample.Scores are included in train_data and used as weight and passed as weight column in the estimator.<\/p>\n\n<p>1 ) I tried without weight column : the model fits about 70 % with good answers. The model is correct for main examples but is not optimal for marginal examples (about 20 % remaining...)<\/p>\n\n<p>2 ) I tried with weight = exp(-score). Model fits about 45 %, is better dealing with marginal examples but have a very poor accuracy with main examples.<\/p>\n\n<p>I think there is a problem with my weights scale [0.0001 - 5003.98] with 0.0001 for the best result and 5003.98 for worth. What is the good way to regularize it ??<\/p>\n\n<p>I read the formula weight = (score - mean(score)) \/ std(score) could be a solution but doesn't seems work<\/p>\n\n<pre><code>model = tf.estimator.LinearClassifier(feature_columns=feature_columns,\n                                    optimizer=tf.train.AdamOptimizer(learning_rate=0.001, beta1= 0.9,beta2=0.99, epsilon = 1e-08,use_locking=False),\n                                    weight_column=weights,\n                                    n_classes=10,\n                                    label_vocabulary=Action_vocab,\n                                    model_dir='.\/Models\/ActionPlayerModel20\/',\n                                    loss_reduction=tf.losses.Reduction.SUM_OVER_BATCH_SIZE,\n                                    config=tf.estimator.RunConfig().replace(save_summary_steps=10))\n\nweight = exp(-score)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58102615","CreationDate":"2019-09-25T16:13:26.023","ParentId":null,"OwnerUserId":"11243873","Title":null,"Body":"<p>Solved the weight problem by doing this (scaling): <\/p>\n\n<pre><code> scaledWeight = (weight-weightMin)\/(weightMax-weightMin)\n<\/code><\/pre>\n\n<p>It did not completly resolve my problem (Better but not optimal accuracy). So, I tried both feature normalization and standardization : good accuracy with normalization but quick overfitting ; very good accuracy and slow overfitting with standardization.<\/p>\n\n<p>Standardization code : <\/p>\n\n<pre><code>    tf.feature_column.numeric_column(\"Round\",dtype=tf.dtypes.float32, normalizer_fn=lambda x: (x - roundMean) \/ roundStd),\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":57557438,"AnswerCount":2,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2019-08-19T13:10:23.677","AcceptedAnswerId":null,"OwnerUserId":3050590.0,"Title":"ValueError: Output tensors to a Model must be the output of a Keras `Layer` when creating ensemble of pre-trained model","Body":"<p>I have some pre-trained binary Keras model. 3 models in this example. I am trying to make an ensemble model by averaging the predictions of these models. To do so I am taking reference from <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/keras\/functional\" rel=\"nofollow noreferrer\">here<\/a> example from ensembling model. My code is below<\/p>\n\n<pre><code>models=list()\nfor i in os.listdir(model_root):  # to get all the models in one list\n    print(i)\n    filename = model_root + \"\/\" + i\n    model = load_model(filename, custom_objects={'KerasLayer': hub.KerasLayer})  # load model\n    models.append(model)\n# ensemble prediction\nprint(len(models))\nyhats = [model.predict(image_data_val) for model in models]\nprint(np.array(yhats).shape)               # (3, 32, 2)\noutputs = layers.average(yhats)     # averaging the model output\nensemble_model = keras.models.Model(inputs=keras.Input(shape=(None,224,224,3)), outputs=outputs)\n<\/code><\/pre>\n\n<p>To which I am getting the following error<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"&lt;input&gt;\", line 1, in &lt;module&gt;\n  File \"C:\\Program Files\\JetBrains\\PyCharm 2019.2\\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 \"C:\\Program Files\\JetBrains\\PyCharm 2019.2\\helpers\\pydev\\_pydev_imps\\_pydev_execfile.py\", line 18, in execfile\n    exec(compile(contents+\"\\n\", file, 'exec'), glob, loc)\n  File \"C:\/Users\/Pawandeep\/Desktop\/Python projects\/ensemble_image.py\", line 68, in &lt;module&gt;\n    ensemble_model = keras.models.Model(inputs=keras.Input(shape=(None,224,224,3)), outputs=outputs)\n  File \"C:\\Python\\lib\\site-packages\\keras\\legacy\\interfaces.py\", line 91, in wrapper\n    return func(*args, **kwargs)\n  File \"C:\\Python\\lib\\site-packages\\keras\\engine\\network.py\", line 93, in __init__\n    self._init_graph_network(*args, **kwargs)\n  File \"C:\\Python\\lib\\site-packages\\keras\\engine\\network.py\", line 188, in _init_graph_network\n    'Found: ' + str(x))\nValueError: Output tensors to a Model must be the output of a Keras `Layer` (thus holding past layer metadata). Found: Tensor(\"average\/truediv:0\", shape=(32, 2), dtype=float32)\n<\/code><\/pre>\n\n<p>So the overall goal is to ensemble the binary keras model which are trained on sigmoid function. I am not able to understand the error properly, because the outputs is the output of keras model itself.<\/p>\n\n<p>Traceback after redefining the model as per the suggestion<\/p>\n\n<pre><code>W0819 16:51:11.051734 11788 ag_logging.py:145] Entity &lt;tensorflow.python.saved_model.function_deserialization.RestoredFunction object at 0x0000020A66459710&gt; could not be transformed and will be executed as-is. Please report this to the AutgoGraph team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output. Cause: Could not find matching function to call loaded from the SavedModel. Got:\n  Positional arguments (3 total):\n    * Tensor(\"inputs:0\", shape=(?, ?, 224, 224, 3), dtype=float32)\n    * False\n    * 0.99\n  Keyword arguments: {}\nExpected these arguments to match one of the following 2 option(s):\nOption 1:\n  Positional arguments (3 total):\n    * TensorSpec(shape=(?, 224, 224, 3), dtype=tf.float32, name='inputs')\n    * True\n    * TensorSpec(shape=(), dtype=tf.float32, name='batch_norm_momentum')\n  Keyword arguments: {}\nOption 2:\n  Positional arguments (3 total):\n    * TensorSpec(shape=(?, 224, 224, 3), dtype=tf.float32, name='inputs')\n    * False\n    * TensorSpec(shape=(), dtype=tf.float32, name='batch_norm_momentum')\n  Keyword arguments: {}\nTraceback (most recent call last):\n  File \"&lt;input&gt;\", line 1, in &lt;module&gt;\n  File \"C:\\Program Files\\JetBrains\\PyCharm 2019.2\\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 \"C:\\Program Files\\JetBrains\\PyCharm 2019.2\\helpers\\pydev\\_pydev_imps\\_pydev_execfile.py\", line 18, in execfile\n    exec(compile(contents+\"\\n\", file, 'exec'), glob, loc)\n  File \"C:\/Users\/Pawandeep\/Desktop\/Python projects\/ensemble_image.py\", line 66, in &lt;module&gt;\n    yhats = [model(inputs) for model in models]\n  File \"C:\/Users\/Pawandeep\/Desktop\/Python projects\/ensemble_image.py\", line 66, in &lt;listcomp&gt;\n    yhats = [model(inputs) for model in models]\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer.py\", line 634, in __call__\n    outputs = call_fn(inputs, *args, **kwargs)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\sequential.py\", line 247, in call\n    return super(Sequential, self).call(inputs, training=training, mask=mask)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\network.py\", line 751, in call\n    return self._run_internal_graph(inputs, training=training, mask=mask)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\network.py\", line 893, in _run_internal_graph\n    output_tensors = layer(computed_tensors, **kwargs)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer.py\", line 634, in __call__\n    outputs = call_fn(inputs, *args, **kwargs)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\sequential.py\", line 247, in call\n    return super(Sequential, self).call(inputs, training=training, mask=mask)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\network.py\", line 751, in call\n    return self._run_internal_graph(inputs, training=training, mask=mask)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\network.py\", line 893, in _run_internal_graph\n    output_tensors = layer(computed_tensors, **kwargs)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer.py\", line 634, in __call__\n    outputs = call_fn(inputs, *args, **kwargs)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\autograph\\impl\\api.py\", line 149, in wrapper\n    raise e.ag_error_metadata.to_exception(type(e))\nValueError: in converted code:\n    relative to C:\\Python\\lib\\site-packages:\n    tensorflow_hub\\keras_layer.py:173 call *\n        result = smart_cond.smart_cond(training,\n    tensorflow\\python\\framework\\smart_cond.py:56 smart_cond\n        return false_fn()\n    tensorflow\\python\\saved_model\\load.py:323 _call_attribute\n        return instance.__call__(*args, **kwargs)\n    tensorflow\\python\\eager\\def_function.py:406 __call__\n        results = self._stateful_fn(*args, **kwds)\n    tensorflow\\python\\eager\\function.py:1323 __call__\n        graph_function, args, kwargs = self._maybe_define_function(args, kwargs)\n    tensorflow\\python\\eager\\function.py:1652 _maybe_define_function\n        graph_function = self._create_graph_function(args, kwargs)\n    tensorflow\\python\\eager\\function.py:1545 _create_graph_function\n        capture_by_value=self._capture_by_value),\n    tensorflow\\python\\framework\\func_graph.py:715 func_graph_from_py_func\n        func_outputs = python_func(*func_args, **func_kwargs)\n    tensorflow\\python\\eager\\def_function.py:307 wrapped_fn\n        return weak_wrapped_fn().__wrapped__(*args, **kwds)\n    tensorflow\\python\\saved_model\\function_deserialization.py:256 restored_function_body\n        \"\\n\\n\".join(signature_descriptions)))\n    ValueError: Could not find matching function to call loaded from the SavedModel. Got:\n      Positional arguments (3 total):\n        * Tensor(\"inputs:0\", shape=(?, ?, 224, 224, 3), dtype=float32)\n        * False\n        * 0.99\n      Keyword arguments: {}\n\n    Expected these arguments to match one of the following 2 option(s):\n\n    Option 1:\n      Positional arguments (3 total):\n        * TensorSpec(shape=(?, 224, 224, 3), dtype=tf.float32, name='inputs')\n        * True\n        * TensorSpec(shape=(), dtype=tf.float32, name='batch_norm_momentum')\n      Keyword arguments: {}\n\n    Option 2:\n      Positional arguments (3 total):\n        * TensorSpec(shape=(?, 224, 224, 3), dtype=tf.float32, name='inputs')\n        * False\n        * TensorSpec(shape=(), dtype=tf.float32, name='batch_norm_momentum')\n      Keyword arguments: {}\n<\/code><\/pre>\n","answers":[{"AnswerId":"57558017","CreationDate":"2019-08-19T13:44:23.413","ParentId":null,"OwnerUserId":"5825953","Title":null,"Body":"<p>There are some issues with the code in the question<\/p>\n\n<ol>\n<li><code>inputs<\/code> and <code>outputs<\/code> arguments of the model defined on the following line are disconnected<\/li>\n<\/ol>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nensemble_model = keras.models.Model(inputs=keras.Input(shape=(None,224,224,3)), outputs=outputs)\n<\/code><\/pre>\n\n<ol start=\"2\">\n<li>Usage of the model.predict on the following line is not intended usage<\/li>\n<\/ol>\n\n<pre class=\"lang-py prettyprint-override\"><code>yhats = [model.predict(image_data_val) for model in models]\n\n<\/code><\/pre>\n\n<p>In keras model.predict returns numpy array for a given input. so on the above code line <code>yhats<\/code> will have list of numpy array<\/p>\n\n<p>Keras model is callable and will return output tensor computed by the model for given input<\/p>\n\n<p>The following code will fix the above issues<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>inputs = keras.Input(shape=(None,224,224,3)) # If the first element of the shape is batch dimension, then it should be removed from the shape parameter\nyhats = [model(inputs) for model in models]\noutputs = layers.average(yhats) \nensemble_model = keras.models.Model(inputs=inputs, outputs=outputs)\n<\/code><\/pre>\n"},{"AnswerId":"57557648","CreationDate":"2019-08-19T13:22:36.860","ParentId":null,"OwnerUserId":"5563864","Title":null,"Body":"<p>I found the answer here <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/6263\" rel=\"nofollow noreferrer\">Output tensors to a Model must be Keras tensors. Found: Tensor #6263<\/a><\/p>\n"}]}
+{"QuestionId":57557564,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-19T13:17:45.937","AcceptedAnswerId":57558281.0,"OwnerUserId":3599803.0,"Title":"tensorflow reload mode to session from graph def","Body":"<p>I have an exported tensorflow saved model which is used for serving.\nI want to \"reload\" it from graphdef object, which I can broadcast for usage with spark.\nI load it using:<\/p>\n\n<pre><code>sess = tf.Session()\ntf.saved_model.loader.load(sess, ['serve'], folder)\nsess.run('dense_1\/Softmax:0', {'input_1:0': input_image})  # works\n<\/code><\/pre>\n\n<p>Then, to load it again to different session, I've tried:<\/p>\n\n<pre><code>graph_def = sess.graph.as_graph_def()\n\n# then, to load\nwith tf.Session(graph=tf.Graph()) as sess:\n   tf.import_graph_def(graph_def, name=\"\")\n   sess.run('dense_1\/Softmax:0', {'input_1:0': input_image})\n<\/code><\/pre>\n\n<p>I get the error:\n<code>FailedPreconditionError: Attempting to use uninitialized value dense_1\/kernel<\/code><\/p>\n\n<p>I've tried adding \n<code>sess.run(tf.global_variables_initializer())<\/code>\nBut still the same error.\nWhat am I missing?<\/p>\n","answers":[{"AnswerId":"57558281","CreationDate":"2019-08-19T13:58:44.357","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You cannot copy variable values from one session to another through the graphdef. The variable values are stored within the session, and the graph definition only contains the structure of the graph. You need to \"export\" the variable values from one session and then restore them in the other. If you want to avoid using checkpoints or similar tooling, you can use a function that should work in most cases like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Gets variable values as a list of pairs with the name and the value\ndef get_variable_values(sess):\n    # Find variable operations\n    var_ops = [op for op in sess.graph.get_operations() if op.type == 'VariableV2']\n    # Get the values\n    var_values = []\n    for v in var_ops:\n        try:\n            var_values.append(sess.run(v.outputs[0]))\n        except tf.errors.FailedPreconditionError:\n            # Uninitialized variables are ignored\n            pass\n    # Return the pairs list\n    return [(op.name, val) for op, val in zip(var_ops, var_values)]\n\n# Restore the variable values\ndef restore_var_values(sess, var_values):\n    # Find the variable initialization operations\n    assign_ops = [sess.graph.get_operation_by_name(v + '\/Assign') for v, _ in var_values]\n    # Run the initialization operations with the given variable values\n    sess.run(assign_ops, feed_dict={op.inputs[1]: val\n                                    for op, (_, val) in zip(assign_ops, var_values)})\n\n# Test\nwith tf.Graph().as_default(), tf.Session() as sess:\n    v = tf.Variable(0., tf.float32, name='a')\n    v.load(3., sess)\n    var_values = get_variable_values(sess)\n    graph_def = tf.get_default_graph().as_graph_def()\nwith tf.Graph().as_default(), tf.Session() as sess:\n    tf.import_graph_def(graph_def, name=\"\")\n    restore_var_values(sess, var_values)\n    print(sess.run('a:0'))\n    # 3.0\n<\/code><\/pre>\n"}]}
+{"QuestionId":57558293,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-08-19T13:59:15.063","AcceptedAnswerId":null,"OwnerUserId":11946855.0,"Title":"How to find the value for which the function is maximum (implementation code )?","Body":"<p>I am trying to solve the following problem: i have this function: <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/gWJyk.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/gWJyk.png\" alt=\"enter image description here\"><\/a> <\/p>\n\n<p>with I0(k) being this function: <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/qJMey.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/qJMey.png\" alt=\"enter image description here\"><\/a>. <\/p>\n\n<p>In each step in my code, i have a known, different angle for the cosine and i want each time to know: for which k the function is maximum. Mathematically, i have to calculate the derivative, to set it equal to zero and to solve the equation. But how is it possible to implement that with tensorflow? What is the best way to solve this problem please ? <\/p>\n","answers":[]}
+{"QuestionId":57558313,"AnswerCount":1,"Tags":"<tensorflow><keras><tf.keras>","CreationDate":"2019-08-19T14:00:27.910","AcceptedAnswerId":null,"OwnerUserId":3850759.0,"Title":"How to create Keras ZeroTensor of specific shape","Body":"<p>I am a total beginner with tensorflow.keras and I am wondering how I could create a constant zero tensor of a specific shape.<\/p>\n\n<p>For example with this:<\/p>\n\n<pre><code>zeros = tf.keras.backend.zeros((someTensor.shape[0], someTensor.shape[1], someTensor.shape[2], channels))\nconcat = tf.kerasbackend.concatenate([someTensor, zeros], axis=3)\n<\/code><\/pre>\n\n<p>The operation tf.keras.backend.zeros fails with:<\/p>\n\n<pre><code>ValueError: Cannot convert a partially known TensorShape to a Tensor\n<\/code><\/pre>\n\n<p>I guess thats because the batch size is unknown during graph building. How can I create a ZeroTensor or any other constant tensor when I don't know the batchsize at that moment? Or is there some kind of unknown(?) value that I can specify?<\/p>\n","answers":[{"AnswerId":"57562501","CreationDate":"2019-08-19T18:57:26.923","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>It's strange because you are using a tuple of tensors and integers. Sort of weird. <\/p>\n\n<p>You should:<\/p>\n\n<pre><code>shape = K.shape(someTensor)\nch = K.variable([channels]) #I think K.constant also works. \nnewShape = K.concatenate([shape[:3], ch])\n\nzeros = K.zeros(newShape)\n<\/code><\/pre>\n\n<p>Now, if this doesn't work because of unknown shapes, a dirty workaround would be:<\/p>\n\n<pre><code>#if someTensor is 3D\n    zeros = K.zeros_like(someTensor)\n    zeros = K.stack([zeros] * channels, axis=-1) \n\n#if someTensor is 4D\n    zeros = K.zeros_like(someTensor[:,:,:,0])\n    zeros = K.stack([zeros]*channels, axis=-1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57558360,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-08-19T14:02:58.250","AcceptedAnswerId":null,"OwnerUserId":10891473.0,"Title":"pip install imagenet_utils gives error message","Body":"<p>I installed keras using<\/p>\n\n<pre><code>pip install keras\n<\/code><\/pre>\n\n<p>then tried to install the <code>imagenet_utils<\/code> package using<\/p>\n\n<pre><code>pip install imagenet_utils\n<\/code><\/pre>\n\n<p>but got the error<\/p>\n\n<blockquote>\n  <p>Could not find a version that satisfies the requirement imagenet_utils\n  (from versions: ) No matching distribution found for imagenet_utils<\/p>\n<\/blockquote>\n\n<p>How can this issue be solved?<\/p>\n","answers":[]}
+{"QuestionId":57558476,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><yolo>","CreationDate":"2019-08-19T14:09:34.887","AcceptedAnswerId":null,"OwnerUserId":11734427.0,"Title":"Training a Keras model yields multiple optimizer errors","Body":"<p>So I need to retrain Tiny YOLO using my own dataset. The model I am using can be found here: <a href=\"https:\/\/github.com\/qqwweee\/keras-yolo3\" rel=\"noreferrer\">keras-yolo3\n<\/a>.<\/p>\n\n<p>I started training and I get multiple optimizer errors, added the code of the errors to stop confusion.\nAnd I noticed the training is going slow even tho it should use the GPU, and after digging a bit I found that this is not using the GPU for training.\nI should note that on another smaller network which I used for learning training uses GPU so everything is set correctly from that side, and they are no errors of this type when I did that training.<\/p>\n\n<p>Is this slow and somewhat CPU training because of said errors? How can I fix this does anyone know?<\/p>\n\n<pre><code>Using TensorFlow backend.\nWARNING: Logging before flag parsing goes to stderr.\n2019-08-19 09:45:08.057713: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library nvcuda.dll\n2019-08-19 09:45:08.264577: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1640] Found device 0 with properties:\nname: GeForce GTX 1060 6GB major: 6 minor: 1 memoryClockRate(GHz): 1.8475\npciBusID: 0000:01:00.0\n2019-08-19 09:45:08.270723: I tensorflow\/stream_executor\/platform\/default\/dlopen_checker_stub.cc:25] GPU libraries are statically linked, skip dlopen check.\n2019-08-19 09:45:08.275827: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1763] Adding visible gpu devices: 0\n2019-08-19 09:45:09.214197: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1181] Device interconnect StreamExecutor with strength 1 edge matrix:\n2019-08-19 09:45:09.217605: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1187]      0\n2019-08-19 09:45:09.219777: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1200] 0:   N\n2019-08-19 09:45:09.222399: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1326] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 4712 MB memory) -&gt; physical GPU (device: 0, name: GeForce GTX 1060 6GB, pci bus id: 0000:01:00.0, compute capability: 6.1)\nCreate Tiny YOLOv3 model with 6 anchors and 80 classes.\nLoad weights model_data\/tiny_yolo_weights.h5.\nFreeze the first 42 layers of total 44 layers.\nTrain on 8298 samples, val on 922 samples, with batch size 32.\nEpoch 1\/50\n2019-08-19 09:45:19.742610: E tensorflow\/core\/grappler\/optimizers\/meta_optimizer.cc:502] shape_optimizer failed: Invalid argument: Subshape must have computed start &gt;= end since stride is negative, but is 0 and 2 (computed from start 0 and end 9223372036854775807 over shape with rank 2 and stride-1)\n2019-08-19 09:45:19.781035: E tensorflow\/core\/grappler\/optimizers\/meta_optimizer.cc:502] remapper failed: Invalid argument: Subshape must have computed start &gt;= end since stride is negative, but is 0 and 2 (computed from start 0 and end 9223372036854775807 over shape with rank 2 and stride-1)\n2019-08-19 09:45:19.935930: E tensorflow\/core\/grappler\/optimizers\/meta_optimizer.cc:502] layout failed: Invalid argument: Subshape must have computed start &gt;= end since stride is negative, but is 0 and 2 (computed from start 0 and end 9223372036854775807 over shape with rank 2 and stride-1)\n2019-08-19 09:45:20.168936: E tensorflow\/core\/grappler\/optimizers\/meta_optimizer.cc:502] shape_optimizer failed: Invalid argument: Subshape must have computed start &gt;= end since stride is negative, but is 0 and 2 (computed from start 0 and end 9223372036854775807 over shape with rank 2 and stride-1)\n2019-08-19 09:45:20.205304: E tensorflow\/core\/grappler\/optimizers\/meta_optimizer.cc:502] remapper failed: Invalid argument: Subshape must have computed start &gt;= end since stride is negative, but is 0 and 2 (computed from start 0 and end 9223372036854775807 over shape with rank 2 and stride-1)\n258\/259 [============================&gt;.] - ETA: 3s - loss: 41.82962019-08-19 10:01:51.053474: E tensorflow\/core\/grappler\/optimizers\/meta_optimizer.cc:502] remapper failed: Invalid argument: Subshape must have computed start &gt;= end since stride is negative, but is 0 and 2 (computed from start 0 and end 9223372036854775807 over shape with rank 2 and stride-1)\n2019-08-19 10:01:51.138957: E tensorflow\/core\/grappler\/optimizers\/meta_optimizer.cc:502] layout failed: Invalid argument: Subshape must have computed start &gt;= end since stride is negative, but is 0 and 2 (computed from start 0 and end 9223372036854775807 over shape with rank 2 and stride-1)\n2019-08-19 10:01:51.243888: E tensorflow\/core\/grappler\/optimizers\/meta_optimizer.cc:502] remapper failed: Invalid argument: Subshape must have computed start &gt;= end since stride is negative, but is 0 and 2 (computed from start 0 and end 9223372036854775807 over shape with rank 2 and stride-1)\n259\/259 [==============================] - 1078s 4s\/step - loss: 41.8008 - val_loss: 35.7122\n<\/code><\/pre>\n","answers":[{"AnswerId":"58814751","CreationDate":"2019-11-12T08:31:23.853","ParentId":null,"OwnerUserId":"2406320","Title":null,"Body":"<p>I have found solution here: <a href=\"https:\/\/github.com\/tensorflow\/tensorrt\/issues\/118\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorrt\/issues\/118<\/a><\/p>\n\n<p>You have to change lines(140\/141) in yolo3\/model.py:<\/p>\n\n<pre><code>box_xy = (K.sigmoid(feats[..., :2]) + grid) \/ K.cast(grid_shape[::-1], K.dtype(feats))\nbox_wh = K.exp(feats[..., 2:4]) * anchors_tensor \/ K.cast(input_shape[::-1], K.dtype(feats))\n<\/code><\/pre>\n\n<p>to:<\/p>\n\n<pre><code>box_xy = (K.sigmoid(feats[..., :2]) + grid) \/ K.cast(grid_shape[...,::-1], K.dtype(feats))\nbox_wh = K.exp(feats[..., 2:4]) * anchors_tensor \/ K.cast(input_shape[...,::-1], K.dtype(feats))\n<\/code><\/pre>\n\n<p>Also in my case helps decrease <strong>batch size<\/strong> from <strong>8<\/strong> to <strong>4<\/strong>.<\/p>\n"}]}
+{"QuestionId":57558538,"AnswerCount":2,"Tags":"<python><tensorflow><keras><deep-learning><conv-neural-network>","CreationDate":"2019-08-19T14:13:00.510","AcceptedAnswerId":null,"OwnerUserId":8895481.0,"Title":"Proper and definitive explanation about how to build a CNN 1D in Keras","Body":"<p>Ciao,<\/p>\n\n<p>this is the second part of a problem I'm facing with CNN 1d. The first part is this<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/57554413\/how-does-it-works-the-input-shape-variable-in-conv1d-in-keras\/57554644#57554644\">How does it works the input_shape variable in Conv1d in Keras?<\/a><\/p>\n\n<p>I'using this code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.models import Sequential\nfrom keras.layers import Dense, Conv1D\nimport numpy as np\n\nN_FEATURES=5\nN_TIMESTEPS=10\nX = np.random.rand(100, N_FEATURES)\nY = np.random.randint(0,2, size=100)\n\nmodel = Sequential()\nmodel.add(Conv1D(filters=32, kernel_size=N_TIMESTEPS, activation='relu', input_shape=(N_TIMESTEPS, N_FEATURES)))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>Now, what I want to do?<\/p>\n\n<blockquote>\n  <p>I want to train a CNN 1d over a timeseries with 5 features. Actually I want to work with time windows og length N_TIMESTEPS rather than timeserie it self. This means that I want to use a sort of \"magnifier\" of dimension N_TIMESTEPS x N_FEATURES on the time series to work locally. That's why I've decided to use CNN<\/p>\n<\/blockquote>\n\n<p>Here come the first question. It is not clear at all if I have to transform the time series into a tensor or it is something that Keras will do for me since I've specified the <strong>kernel_size<\/strong> variable.<\/p>\n\n<p>In case I must provide a tensor I would do something like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>X_tensor = []\nfor i in range(len(X)-N_TIMESTEPS):\n    X_tensor+=[X_tensor[range(i, N_TIMESTEPS+i), :]]\nX_tensor = np.asarray(X_tensor)\n<\/code><\/pre>\n\n<p>In this case of course I should also provide a <strong>Y_tensor<\/strong> vector computed from <strong>Y<\/strong> according to some criteria. Suppose I have already this <strong>Y_tensor<\/strong> boolean vector of the same length of <strong>X_tensor<\/strong>, which is <strong>len(X)-N_TIMESTEPS-1<\/strong>.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Y_tensor = np.random.randint(0,2,len(X)-N_TIMESTEPS-1)\n<\/code><\/pre>\n\n<p>Now if I try to feed the model I get of the most common error for CNN 1d which is:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>  ValueError: Error when checking input: expected conv1d_4_input to have 3 dimensions, but got array with shape (100, 5)\n<\/code><\/pre>\n\n<p>By looking to a dozen of posts about it I cannot understand what I did wrong. This is what I've tried:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.fit(X,Y)\nmodel.fit(np.expand_dims(X, axis=0),Y)\nmodel.fit(np.expand_dims(X, axis=2),Y)\nmodel.fit(X_tensor,Y_tensor)\n<\/code><\/pre>\n\n<p>For all of these cases I get always the same error (with different dimensional values in the final tuple).<\/p>\n\n<p>Questions:<\/p>\n\n<blockquote>\n  <ol>\n  <li><p>What Keras expects from my data? Can I feed the model with the whole time series or I have to slice it into a tensor?<\/p><\/li>\n  <li><p>How I have to feed the model in term of data structure?I.e. I have to specify in some strange way the dimension of the data?<\/p><\/li>\n  <\/ol>\n<\/blockquote>\n\n<p>Can you help me? I find out that this is one the most confusing point of CNN implementation in Keras that there are different posts with different solutions that do not fit with structure of my data (even if they have a very common structure according to me).<\/p>\n\n<p><strong>Note<\/strong>: There are some post suggesting to pass in the <strong>input_shape<\/strong> variable the length of the data. This is meaningless to me since I should not provide the dimension of the data (which is a variable) to the model. The only thing I should give to it, according to the theory, is the filter dimension and number of features (namely the dimension of the matrix that will roll over the time series).<\/p>\n\n<p>Thanks,<\/p>\n\n<p>am<\/p>\n","answers":[{"AnswerId":"57559615","CreationDate":"2019-08-19T15:18:17.377","ParentId":null,"OwnerUserId":"5432214","Title":null,"Body":"<p>I think you don't have a clear idea on how 1d convolutional neural networks work:<\/p>\n\n<p>if you want to predict the y values from the timeseries x and you just have 1 timeseries, your approach won't work. the network needs lots of samples to train, and having just 1 will allow it to easily memorize the input and not learn. For example, if the timeseries is the humidity of a given day, and y is the chance of rain at a specific timestep, what you have now is the data for just one day (timesteps being for example hours of the day). In order for the network to learn you need to gather data for many days, ending up with datasets of shape x=(n_days, timesteps, features) y=(n_days, timesteps, 1).<\/p>\n\n<p>If you describe your actual problem there's better chance to get more helpful answers<\/p>\n\n<p>[Edit] by sticking to your code, and using just one timeseries, you are better off with other methods that don't involve deep learning. You could split your timeseries at regular interval, obtaining n samples that would allow your network to train, but unless you have a very long timeseries that may not be a valid alternative.<\/p>\n"},{"AnswerId":"57562085","CreationDate":"2019-08-19T18:22:54.270","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Simply, <code>Conv1D<\/code> requires 3 dimensions:<\/p>\n\n<ul>\n<li>Number of series (1)   <\/li>\n<li>Number of steps (100 - your entire data)    <\/li>\n<li>Number of features (5)    <\/li>\n<\/ul>\n\n<p>So, <code>model.fit(np.expand_dims(X, axis=0),Y)<\/code> is correct for <code>X<\/code>.<\/p>\n\n<p>Now, if <code>X<\/code> is <code>(1, 100, 5)<\/code>, naturally your <code>input_shape=(100,5)<\/code>.   <\/p>\n\n<p>If your <code>Y<\/code> has 100 steps, then you need to make sure your <code>Conv1D<\/code> will output 100 steps. You need <code>padding='same'<\/code>, otherwise it will become <code>91<\/code>. (I suggest you actually work with 91, since you want a result for each 10 steps and probably don't want border effects spoiling your results)      <\/p>\n\n<p><code>Y<\/code> must also follow the same rules for shape:    <\/p>\n\n<ul>\n<li>Number of series (1)    <\/li>\n<li>Number of steps (100 if <code>padding='same'<\/code>; 91 if <code>padding='valid'<\/code>)    <\/li>\n<li>Number of features (1 = Dense output)<\/li>\n<\/ul>\n\n<p>So, <code>Y = Y.reshape((1,-1,1))<\/code>.<\/p>\n\n<p>Since you have only one class (true\/false), it's pointless to use <code>'categorical_crossentropy'<\/code>. You should go with <code>'binary_crossentropy'<\/code>.     <\/p>\n\n<p>In general, your overall idea of using this convolution with <code>kernel_size=10<\/code> to simulate sliding windows of 10 steps will work as expected (whether it will be efficient or not is another question, answered only by trying).<\/p>\n\n<hr>\n\n<p>If you want better networks for sequences, you should probably try <code>LSTM<\/code> layers. The dimensions work exactly the same way. You will need <code>return_sequences=False<\/code>.   <\/p>\n\n<p>The main difference is that you will need to separate the data as you did in that loop. Then:   <\/p>\n\n<ul>\n<li><code>X.shape == (91, 10, 5)<\/code>   <\/li>\n<li><code>Y.shape == (91, 1)<\/code>  <\/li>\n<\/ul>\n"}]}
+{"QuestionId":57559014,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><neural-network>","CreationDate":"2019-08-19T14:41:25.503","AcceptedAnswerId":null,"OwnerUserId":9328893.0,"Title":"Why is training so slow?","Body":"<p>I started to train a model with my GeForce RTX 2070 with 16 GB RAM, but it takes a lot of time even for one epoch. For example, 1 epoch takes more than 10 hours. As it will require near 100 epochs for a good result. So I've met problems with training process, as I don't have time to wait a month to finish training. <\/p>\n\n<p>The dataset is near 500000 images with total size 200 GB.<\/p>\n\n<p>I am using YOLO (in darknet) with weights tiny-yolov2. My tensorflow version is 1.12.0. Weights and configuration files were downloaded from official site. GPU usage in options is set as 0.8.<\/p>\n\n<p>Before starting training my GPU usage is:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/j5H0m.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/j5H0m.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>After starting training:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/sxRmq.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/sxRmq.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Code:<\/p>\n\n<pre><code>import matplotlib.pyplot as plt\nimport numpy as np\n\nfrom darkflow_master2.darkflow.net.build import TFNet\nimport cv2\nimport os\nimport tensorflow as tf\nos.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'\n\noptions = {\"model\": \"cfg\/yolov2-3c.cfg\",\n           \"load\": \"bin\/yolov2.weights\",\n           \"batch\": 64,\n           \"epoch\": 2,\n           \"gpu\": 0.8,\n           \"train\": True,\n           \"savepb\": True,\n           \"save\": 1000,\n           \"annotation\": \".\/annotations\/\",\n           \"dataset\": \".\/images\/\",\n           \"lr  \": 1e-3}\n\ntfnet = TFNet(options)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57559016,"AnswerCount":1,"Tags":"<python><tensorflow><keras><autoencoder><loss-function>","CreationDate":"2019-08-19T14:41:58.453","AcceptedAnswerId":57559521.0,"OwnerUserId":8044858.0,"Title":"Keras built-in MSE loss on 2D data returns 2D matrix, not scalar loss","Body":"<p>I'm trying to evaluate the MSE loss for single 2D test samples with an autoencoder (AE) in Keras once the model is trained and I'm surprised that when I call Keras MSE built-in function to get individual samples' loss it returns 2D tensors. That means the loss function computes one loss per pixel for each sample, and not one loss per sample as it should (?). To be perfectly clear, I expected MSE to associate to each 2D sample the mean of the squared errors computed over all pixels (as I've read on this <a href=\"https:\/\/stackoverflow.com\/a\/47003117\/8044858\">SO post<\/a>).<\/p>\n\n<p>Since I didn't manage to get an array of scalar MSE errors with one scalar per test sample after training my AE using <code>.predict()<\/code> and <code>.evaluate()<\/code> (perhaps I missed something there as well), I went on trying to directly use <code>keras.losses.mean_squared_error()<\/code>, sample by sample. This returned me 2D tensors as a loss for each sample (input tensors are of size <code>(N,M,1)<\/code>). When one looks at <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/losses.py\" rel=\"nofollow noreferrer\">Keras' original implementation<\/a> of MSE loss, one finds:<\/p>\n\n<pre><code>def mean_squared_error(y_true, y_pred):\n    return K.mean(K.square(y_pred - y_true), axis=-1)\n<\/code><\/pre>\n\n<p>The <code>axis=-1<\/code> explains why multiple dimensions aren't immediately reduced to a scalar when computing the loss.<\/p>\n\n<p>I therefore wonder:<\/p>\n\n<ol>\n<li>What exactly has my model been using during training ? Was it the\nmean of squared error over all pixels for each sample as I expected\n? This isn't what the built-in code suggests.<\/li>\n<li>Do I absolutely need to re-define the MSE loss to get the individual MSE losses for each test sample ? To obtain a scalar I\nwould then have to flatten the samples and the associated\npredictions, and then re-apply the built-in MSE (and this sample by sample).<\/li>\n<\/ol>\n\n<p>Manually flattening before computing MSE seems what needs to be done according to this <a href=\"https:\/\/stackoverflow.com\/a\/51855895\/8044858\">SO answer<\/a> on Keras' MSE loss. Using MSE for an AE model with 2D data seemed fine to me as I read this <a href=\"https:\/\/keras.io\/examples\/mnist_denoising_autoencoder\/\" rel=\"nofollow noreferrer\">keras.io<\/a> Mnist denoising tutorial.<\/p>\n\n<p>My code:<\/p>\n\n<pre><code>import keras\n\nAE_testOutputs = autoencoder.predict(samplesList)\n\nsamplesMSE = []\nfor testSampleIndex in range(samplesList.shape[0]):\n    AE_output = AE_testOutputs[testSampleIndex,:,:,:]\n    samplesMSE.append(keras.losses.mean_squared_error(samplesList[testSampleIndex,:,:,:],AE_output))\n<\/code><\/pre>\n\n<p>Which returns a list <code>samplesMSE<\/code> of <code>Tensor(\"Mean:0\", shape=(15, 800), dtype=float64)<\/code> objects.<\/p>\n\n<p>I'm sorry if I missed a similar question, I did actively research before posting, and I'm still afraid there is a very simple explanation\/I must have missed a built-in function somewhere.<\/p>\n","answers":[{"AnswerId":"57559521","CreationDate":"2019-08-19T15:11:32.630","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Although it is not absolutely required, Keras loss functions are conventionally defined \"per-sample\", where \"sample\" is basically each element in the output tensor of the model. The loss function is then pass through a wrapping function <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/2.2.4\/keras\/engine\/training_utils.py#L374-L424\" rel=\"nofollow noreferrer\"><code>weighted_masked_objective<\/code><\/a> that adds support for masking and sample weighting. By default, the total loss is the average of the samples losses.<\/p>\n\n<p>If you want to get the mean of some value across every dimension but the first one, you can simply use <code>K.mean<\/code> over the value that you get.<\/p>\n"}]}
+{"QuestionId":57559506,"AnswerCount":1,"Tags":"<java><tensorflow><tensorflow2.0>","CreationDate":"2019-08-19T15:10:29.920","AcceptedAnswerId":null,"OwnerUserId":4916075.0,"Title":"Is it possible to load a TensorFlow 2.0 model in Java?","Body":"<p>Wasn't able to find the answer on Google and the questions asked here all seem close but not the same. Sorry if I overlooked something.<\/p>\n\n<p>Is it currently already possible to load a model written in Python TensorFlow 2.0.0-beta1 in Java 8? The model would use the Keras Sequential API. If this is possible, I appreciate a pointer to the appropriate documentation.<\/p>\n","answers":[{"AnswerId":"57568237","CreationDate":"2019-08-20T07:01:12.280","ParentId":null,"OwnerUserId":"2455494","Title":null,"Body":"<p>You can accomplish this if you first convert the model to a tflite file as described <a href=\"https:\/\/www.tensorflow.org\/lite\/convert\/python_api#exporting_a_tfkeras_file_\" rel=\"nofollow noreferrer\">here<\/a>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\n\nconverter = tf.lite.TFLiteConverter.from_keras_model_file(\"keras_model.h5\")\ntflite_model = converter.convert()\nopen(\"converted_model.tflite\", \"wb\").write(tflite_model)\n<\/code><\/pre>\n\n<p>Which you can then load and run as described <a href=\"https:\/\/www.tensorflow.org\/lite\/guide\/inference#load_and_run_a_model_in_java\" rel=\"nofollow noreferrer\">here<\/a>:<\/p>\n\n<pre class=\"lang-java prettyprint-override\"><code>public Interpreter(@NotNull File modelFile);\n\ntry (Interpreter interpreter = new Interpreter(file_of_a_tensorflowlite_model)) {\n  interpreter.run(input, output);\n}\n<\/code><\/pre>\n"}]}
+{"QuestionId":57559533,"AnswerCount":0,"Tags":"<tensorflow><google-cloud-platform><gpu><benchmarking>","CreationDate":"2019-08-19T15:12:15.850","AcceptedAnswerId":null,"OwnerUserId":10526032.0,"Title":"performance of instance with gpu on google cloud","Body":"<p>I made some studies on Google Cloud to explore its (single instance) performance for Tensorflow training.\nI created a 'n1-highmem-8 (8 vCPUs, 52 GB memory)' instance with one NVIDIA Tesla V100 GPU.\nAs a reference I used a GTX1080 Ti.\nI measured the training speed on several object detection models (image size=640x640, batch=1).\nThe result is above.<\/p>\n\n<p>The question is whether I made something wrong or the 'bare metal' performance of the GPU is degraded this much in cloud environment (by the literature the V100 should be roughly 3x faster than the 1080Ti (?).<\/p>\n\n<p>the training speed of ssd+mobilnet v1 on 1080 Ti (batch = 1) is 8.9 iter\/s <\/p>\n\n<p>the training speed of ssd+mobilnet v1 on Tesla V100 (batch = 1) is 7.4 iter\/s (0.89 times slower) <\/p>\n\n<p>the training speed of ssd+fpn+resnet50 on 1080 Ti is 8.4 iter\/s<\/p>\n\n<p>the training speed of ssd+fpn+resnet50 on Tesla V100 is 4.9 iter\/s (0.58 times slower)<\/p>\n\n<p>the training speed of ssd+inception on 1080 Ti is 7.9 iter\/s<\/p>\n\n<p>the same on  Tesla V100 is 7.1 iter\/s (0.9 times faster)<\/p>\n\n<p>Is it normal behavior or I made something wrong ... ?<\/p>\n\n<p>Best Regards,\nKaroly<\/p>\n","answers":[]}
+{"QuestionId":57559858,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-08-19T15:34:55.750","AcceptedAnswerId":57560176.0,"OwnerUserId":10824718.0,"Title":"ImportError: No module named nibabel in python 2.7","Body":"<p>I am using python 2.7 and I used <\/p>\n\n<pre><code>pip install nibabel\n<\/code><\/pre>\n\n<p>but I constantly get the error<\/p>\n\n<pre><code>ImportError: No module named nibabel.\n<\/code><\/pre>\n\n<p>when I type<\/p>\n\n<pre><code>pip show nibabel \n<\/code><\/pre>\n\n<p>I get this result so it seems that it is installed but still get the aforementioned error when using import nibabel as nib<\/p>\n\n<pre><code>    Name: nibabel\nVersion: 2.5.0\nSummary: Access a multitude of neuroimaging data formats\nHome-page: https:\/\/nipy.org\/nibabel\nAuthor: nibabel developers\nAuthor-email: neuroimaging@python.org\nLicense: MIT License\nLocation: \/usr\/local\/lib\/python3.5\/dist-packages\nRequires: six, numpy\nRequired-by: \n<\/code><\/pre>\n\n<p>Here the address shows it is installed in python3 path! How can I address this problem?<\/p>\n","answers":[{"AnswerId":"57560176","CreationDate":"2019-08-19T15:53:36.523","ParentId":null,"OwnerUserId":"9794932","Title":null,"Body":"<p>Install it using pip2:<\/p>\n\n<p><code>pip2 install nibabel<\/code><\/p>\n\n<p>Or explicitly call the pip module for Python 2.7:<\/p>\n\n<p><code>python2.7 -m pip install nibabel<\/code><\/p>\n\n<p>for example.<\/p>\n"}]}
+{"QuestionId":57559917,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning>","CreationDate":"2019-08-19T15:37:49.717","AcceptedAnswerId":null,"OwnerUserId":10841109.0,"Title":"Overfitting DL model?","Body":"<p>I am trying to build a Deep Learning model to pick out Tropical Cyclones in weather model data. I have collected the data, normalized it in the region [0, 1] and passed it to my early model. Then I plotted my loss and accuracy curves as <a href=\"https:\/\/i.stack.imgur.com\/vbQG5.png\" rel=\"nofollow noreferrer\">here<\/a>. I am getting weird curves as the validation loss starts increasing after ~50 epochs, indicating overfitting, but the validation accuracy is still increasing. Is my model overfitting (at around 50 epochs) or not?<\/p>\n","answers":[{"AnswerId":"57560149","CreationDate":"2019-08-19T15:51:46.460","ParentId":null,"OwnerUserId":"7253901","Title":null,"Body":"<p>These graphs are classical graphs that comes with overfitting! You can recognize overfit because even though training accuracy keeps increasing, validation accuracy does not. To prevent overfit, there are several approaches; too numerous to name in one answer. You could apply L1\/L2 regularization, dropout, or artificially expand your training data (amongst others).<\/p>\n"}]}
+{"QuestionId":57560327,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-19T16:03:44.137","AcceptedAnswerId":null,"OwnerUserId":6706617.0,"Title":"Loading a second mode into a tensorflow session","Body":"<p>I'm currently attempting to load a pretrained tf.estimator model into another CycleGan model I am training, cloned from <a href=\"https:\/\/github.com\/vanhuyz\/CycleGAN-TensorFlow\" rel=\"nofollow noreferrer\">https:\/\/github.com\/vanhuyz\/CycleGAN-TensorFlow<\/a>. <\/p>\n\n<p>I want to use the pretrained model to calculate the loss based off a GAN generated input, and then use it to optimize the GAN. Thus this pretrained model must be part of the graph<\/p>\n\n<p>I've tried loading the estimator with       <\/p>\n\n<pre><code>self.estimator = tf.estimator.Estimator(model_fn=self.frcnn.model_fn, model_dir=directory, params={})\n<\/code><\/pre>\n\n<p>And then evaluating with         <\/p>\n\n<pre><code>input_func = functools.partial(self.input_to_estimator, x_features, x_labels)\nmatch_loss = self.estimator.evaluate(input_func)\n<\/code><\/pre>\n\n<p>But this returns an error about the tensors being from different graphs. <\/p>\n\n<p>How would I best import this graph? Would something like <\/p>\n\n<pre><code>import_graph_def\n<\/code><\/pre>\n\n<p>work? Currently, I have the model checkpoints.{data, index, meta} as well as the graph.pbtext and saved_model.pb with the requisite variables. <\/p>\n\n<p>Thank you so much for the help!<\/p>\n","answers":[]}
+{"QuestionId":57560423,"AnswerCount":2,"Tags":"<python><tensorflow><autoencoder>","CreationDate":"2019-08-19T16:11:26.123","AcceptedAnswerId":null,"OwnerUserId":3029270.0,"Title":"tensorflow, splitting autoencoder after training","Body":"<p>I have autoencoder model in tensorflow 1x (not a keras) I am trying to split the model to encoder and decoder after training. <\/p>\n\n<p>both function in same scope\nand I have 3 PlaceHolders<\/p>\n\n<pre><code>self.X = tf.placeholder(shape=[None, vox_res64, vox_res64, vox_res64, 1], dtype=tf.float32)\nself.Z = tf.placeholder(shape=[None,500], dtype=tf.float32)\n\nself.Y = tf.placeholder(shape=[None, vox_rex256, vox_rex256, vox_rex256, 1], dtype=tf.float32)\n\n with tf.variable_scope('aeu'):\n            self.lfc=self.encoder(self.X)\n\n            self.Y_pred, self.Y_pred_modi = self.decoder(self.lfc)\n<\/code><\/pre>\n\n<p>the enocder and decoder as follow<\/p>\n\n<pre><code>    def encoder(self,X):\n        with tf.device('\/gpu:'+GPU0):\n            X = tf.reshape(X,[-1, vox_res64,vox_res64,vox_res64,1])\n            c_e = [1,64,128,256,512]\n            s_e = [0,1 , 1, 1, 1]\n            layers_e = []\n            layers_e.append(X)\n            for i in range(1,5,1):\n                layer = tools.Ops.conv3d(layers_e[-1],k=4,out_c=c_e[i],str=s_e[i],name='e'+str(i))\n                layer = tools.Ops.maxpool3d(tools.Ops.xxlu(layer, label='lrelu'), k=2,s=2,pad='SAME')\n                layers_e.append(layer)\n\n            ### fc\n            [_, d1, d2, d3, cc] = layers_e[-1].get_shape()\n            d1=int(d1); d2=int(d2); d3=int(d3); cc=int(cc)\n            lfc = tf.reshape(layers_e[-1],[-1, int(d1)*int(d2)*int(d3)*int(cc)])\n            lfc = tools.Ops.xxlu(tools.Ops.fc(lfc, out_d=500,name='fc1'), label='relu')\n            print (d1)\n            print(cc)\n        return lfc\n\n\n    def decoder(self,Z):\n        with tf.device('\/gpu:'+GPU0):\n\n\n            lfc = tools.Ops.xxlu(tools.Ops.fc(Z, out_d=2*2*2*512, name='fc2'), label='relu')\n\n            lfc = tf.reshape(lfc, [-1,2,2,2,512])\n\n            c_d = [0,256,128,64]\n            s_d = [0,2,2,2]\n            layers_d = []\n            layers_d.append(lfc)\n            for j in range(1,4,1):\n\n                layer = tools.Ops.deconv3d(layers_d[-1],k=4,out_c=c_d[j],str=s_d[j],name='d'+str(len(layers_d)))\n\n                layer = tools.Ops.xxlu(layer, label='relu')\n                layers_d.append(layer)\n            ###\n            layer = tools.Ops.deconv3d(layers_d[-1],k=4,out_c=1,str=2,name='dlast')\n            print(\"****************************\",layer)\n            ###\n            Y_sig = tf.nn.sigmoid(layer)\n            Y_sig_modi = tf.maximum(Y_sig,0.01)\n\n        return Y_sig, Y_sig_modi\n<\/code><\/pre>\n\n<p>when I try to use model after training<\/p>\n\n<pre><code>\n X = tf.get_default_graph().get_tensor_by_name(\"Placeholder:0\")\n Z = tf.get_default_graph().get_tensor_by_name(\"Placeholder_1:0\")\n Y_pred = tf.get_default_graph().get_tensor_by_name(\"aeu\/Sigmoid:0\")\n lfc = tf.get_default_graph().get_tensor_by_name(\"aeu\/Relu:0\")\n\n\n<\/code><\/pre>\n\n<p>fetching latent code work fine<\/p>\n\n<pre><code> lc = sess.run(lfc, feed_dict={X: x_sample})\n<\/code><\/pre>\n\n<p>now I want to use the latent code as input to decoder I get error I have to fill X(PLACEHOLDER)<\/p>\n\n<pre><code> y_pred = sess.run(Y_pred, feed_dict={Z: lc})\n\n<\/code><\/pre>\n\n<p>how I can split to encoder decoder? \nI searched only I found keras examples<\/p>\n","answers":[{"AnswerId":"57561336","CreationDate":"2019-08-19T17:20:44.047","ParentId":null,"OwnerUserId":"3029270","Title":null,"Body":"<p>I found how to split the model . <\/p>\n\n<p>I will put the answer if anybody want to know<\/p>\n\n<p>my mistakes are :<\/p>\n\n<p>1: I did not pass self.Z to a decoder<\/p>\n\n<p>2: for the following line<\/p>\n\n<pre><code>y_pred = sess.run(Y_pred, feed_dict={Z: lc})\n\n<\/code><\/pre>\n\n<p>this line in  different file after i TRAINED MY MODEL\ntensorflow will not know what does [ z ]  refer to, so you have to use the same variable you identified your tensor with as the following<\/p>\n\n<pre><code> lfc = tf.get_default_graph().get_tensor_by_name(\"aeu\/Relu:0\")\n\n<\/code><\/pre>\n\n<p>I name it [ lfc ] not [ Z ]<\/p>\n\n<p>so changing the code to that solve the issue<\/p>\n\n<pre><code>y_pred = sess.run(Y_pred, feed_dict={lfc: lc})\n\n<\/code><\/pre>\n"},{"AnswerId":"57561228","CreationDate":"2019-08-19T17:12:23.340","ParentId":null,"OwnerUserId":"9480505","Title":null,"Body":"<p>The first thing that I notice is that you haven't passed in self.Z anywhere into the decoder. So tensorflow can't automatically just link that placeholder with the z that you previously used. <\/p>\n\n<p>There's a couple of things you can do to fix this. The easiest is to attempt to recreate the decoder graph but when you call variable scope, set reuse=True.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\n    with tf.variable_scope('aeu',reuse=True):\n        self.new_Y, self.new_Y_modi = self.decoder(self.Z)\n\n    y_pred = sess.run(self.new_Y, feed_dict={self.Z: lc})\n\n<\/code><\/pre>\n\n<p>This is the method that is probably easiest to do. You may be asked to fill in placeholder X in this case as well, but you can just fill that in with an empty array. Normally Tensorflow won't ask for it unless there's some sort of control dependency that ties the two together. <\/p>\n"}]}
+{"QuestionId":57560887,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-19T16:45:54.713","AcceptedAnswerId":null,"OwnerUserId":11947736.0,"Title":"Error in tensor input for custom train_step?","Body":"<p>I'm training two networks in an GAN setup, following the tutorial <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/generative\/dcgan\" rel=\"nofollow noreferrer\">on the tensorflow website<\/a>. My problem occurs when trying to use different training data, and when I try to train in batches.<\/p>\n\n<p>The data I'm interested in is in 13x1 arrays of floats, which have already been converted into (1,13) tensors by copying data into a sample output of the generator. The code below works on batches of 1, but as soon as I try to train on larger batches of data it fails.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def make_generator_model():\n    model = tf.keras.Sequential()\n    model.add(layers.Dense(128, activation='relu', input_shape=(100,)))\n    model.add(layers.LeakyReLU())\n    model.add(layers.Dense(100, activation='relu'))\n    model.add(layers.Dense(13, activation='relu'))\n    return model\n\ngenerator = make_generator_model()\n\ndef make_discriminator_model():\n    model = tf.keras.Sequential()\n    model.add(layers.Dense(128, activation='relu', input_shape=(13,)))\n    model.add(layers.Dense(64, activation='relu'))\n    model.add(layers.LeakyReLU())\n    model.add(layers.Dense(1, activation='sigmoid'))\n    return model\n\ngenerator = make_generator_model()\ndiscriminator = make_discriminator_model()\n\ncross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=True)\n\ndef discriminator_loss(real_output, fake_output):\n    real_loss = cross_entropy(tf.ones_like(real_output), real_output)\n    fake_loss = cross_entropy(tf.zeros_like(fake_output), fake_output)\n    return real_loss + fake_loss\n\ndef generator_loss(fake_output):\n    return cross_entropy(tf.ones_like(fake_output), fake_output)\n\ngenerator_optimizer = tf.keras.optimizers.Adam(1e-4)\ndiscriminator_optimizer = tf.keras.optimizers.Adam(1e-4)\n\nmb_size = 512\n\n@tf.function\ndef train_step(inp):\n    noise = tf.random.normal([mb_size, noise_dim])\n    with tf.GradientTape() as gen_tape, tf.GradientTape() as disc_tape:\n        generated_names = generator(noise, training=True)\n        print(generated_names)\n        real_output = discriminator(inp, training=True)\n        fake_output = discriminator(generated_names, training=True)\n        gen_loss = generator_loss(fake_output)\n        disc_loss = discriminator_loss(real_output, fake_output)\n        gradients_of_generator = gen_tape.gradient(gen_loss, generator.trainable_variables)\n        gradients_of_discriminator = disc_tape.gradient(disc_loss, discriminator.trainable_variables)\n        generator_optimizer.apply_gradients(zip(gradients_of_generator, generator.trainable_variables))\n        discriminator_optimizer.apply_gradients(zip(gradients_of_discriminator, discriminator.trainable_variables))\n<\/code><\/pre>\n\n<p>This functions when <code>mb_size=1<\/code>, but as soon as I raise it (e.g., to 512 as above), I get the following error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"gan.py\", line 137, in &lt;module&gt;\n    train(data, epochs)\n  File \"gan.py\", line 123, in train\n    train_step(mb)\n  File \"\/home\/drv\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/eager\/def_function.py\", line 414, in __call__\n    self._initialize(args, kwds, add_initializers_to=initializer_map)\n  File \"\/home\/drv\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/eager\/def_function.py\", line 357, in _initialize\n    *args, **kwds))\n  File \"\/home\/drv\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/eager\/function.py\", line 1349, in _get_concrete_function_internal_garbage_collected\n    graph_function, _, _ = self._maybe_define_function(args, kwargs)\n  File \"\/home\/drv\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/eager\/function.py\", line 1652, in _maybe_define_function\n    graph_function = self._create_graph_function(args, kwargs)\n  File \"\/home\/drv\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/eager\/function.py\", line 1545, in _create_graph_function\n    capture_by_value=self._capture_by_value),\n  File \"\/home\/drv\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/func_graph.py\", line 715, in func_graph_from_py_func\n    func_outputs = python_func(*func_args, **func_kwargs)\n  File \"\/home\/drv\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/eager\/def_function.py\", line 307, in wrapped_fn\n    return weak_wrapped_fn().__wrapped__(*args, **kwds)\n  File \"\/home\/drv\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/func_graph.py\", line 705, in wrapper\n    raise e.ag_error_metadata.to_exception(type(e))\nValueError: in converted code:\n\n    gan.py:109 train_step *\n        real_output = discriminator(inp, training=True)\n    \/home\/drv\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py:586 __call__\n        self.name)\n    \/home\/drv\/.local\/lib\/python2.7\/site-packages\/tensorflow\/python\/keras\/engine\/input_spec.py:101 assert_input_compatibility\n        ' input tensors. Inputs received: ' + str(inputs))\n\n    ValueError: Layer sequential_1 expects 1 inputs, but it received 512 input tensors. Inputs received: [&lt;tf.Tensor 'inp:0' shape=(1, 13) dtype=float32&gt;, ..., &lt;tf.Tensor 'inp_511:0' shape=(1, 13) dtype=float32&gt;]\n<\/code><\/pre>\n\n<p>I believe the problem is somewhere in my conversion of code designed to work on arrays of images into 13x1 arrays, but I'm lost on where the code should be different.<\/p>\n","answers":[]}
+{"QuestionId":57561004,"AnswerCount":0,"Tags":"<python><pytorch><torchtext>","CreationDate":"2019-08-19T16:54:36.880","AcceptedAnswerId":null,"OwnerUserId":3879993.0,"Title":"torchtext field with values converted to ids got error integer is required","Body":"<p>I followed this tutorial \n<a href=\"http:\/\/www.programmersought.com\/article\/2609385756\/\" rel=\"nofollow noreferrer\">http:\/\/www.programmersought.com\/article\/2609385756\/<\/a><\/p>\n\n<p>to create a TabularDataset with data that already tokenized and converted to ids and I do not want to use vocab or build vocab because the data is numerical<\/p>\n\n<p>so I defined my field variable as:<\/p>\n\n<pre><code>myField = Field(tokenize= x_tokenize, use_vocab=False, sequential=True)\ntrain,val, test = data.TabularDataset.splits(path='.\/', train=train_path, validation=valid_path, test=test_path ,format='csv', fields=data_fields, skip_header=True)\n<\/code><\/pre>\n\n<p>train output:<\/p>\n\n<pre><code>print(vars(train[0])['src'])\n#output this [101, 3177, 3702, 11293, 1116, 102]\n<\/code><\/pre>\n\n<p>and I used a BucketIterator:<\/p>\n\n<pre><code>train_iter= BucketIterator(train,\n                       batch_size=BATCH_SIZE,\n                       device = DEVICE,\n                       sort_key=lambda x: (len(x.src), len(x.trg)), \n                       train=True,\n                       batch_size_fn=batch_size_fn,\n                       repeat=False)\n<\/code><\/pre>\n\n<p>when I run this code:<\/p>\n\n<pre><code>batch = next(iter(train_iter))\n<\/code><\/pre>\n\n<p>I got \nTypeError: an integer is required (got type list)<\/p>\n\n<hr>\n\n<blockquote>\n  <p>TypeError                                 Traceback (most recent call\n  last)  in ()\n  ----> 1 batch = next(iter(train_iter))<\/p>\n  \n  <p>3 frames\n  \/usr\/local\/lib\/python3.6\/dist-packages\/torchtext\/data\/iterator.py in\n  <strong>iter<\/strong>(self)\n      155                     else:\n      156                         minibatch.sort(key=self.sort_key, reverse=True)\n  --> 157                 yield Batch(minibatch, self.dataset, self.device)\n      158             if not self.repeat:\n      159                 return<\/p>\n  \n  <p>\/usr\/local\/lib\/python3.6\/dist-packages\/torchtext\/data\/batch.py in\n  <strong>init<\/strong>(self, data, dataset, device)\n       32                 if field is not None:\n       33                     batch = [getattr(x, name) for x in data]\n  ---> 34                     setattr(self, name, field.process(batch, device=device))\n       35 \n       36     @classmethod<\/p>\n  \n  <p>\/usr\/local\/lib\/python3.6\/dist-packages\/torchtext\/data\/field.py in\n  process(self, batch, device)\n      199         \"\"\"\n      200         padded = self.pad(batch)\n  --> 201         tensor = self.numericalize(padded, device=device)\n      202         return tensor\n      203 <\/p>\n  \n  <p>\/usr\/local\/lib\/python3.6\/dist-packages\/torchtext\/data\/field.py in\n  numericalize(self, arr, device)\n      321                 arr = self.postprocessing(arr, None)\n      322 \n  --> 323         var = torch.tensor(arr, dtype=self.dtype, device=device)\n      324 \n      325         if self.sequential and not self.batch_first:<\/p>\n  \n  <p>TypeError: an integer is required (got type list)<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":57561009,"AnswerCount":0,"Tags":"<pytorch>","CreationDate":"2019-08-19T16:55:09.190","AcceptedAnswerId":null,"OwnerUserId":9222127.0,"Title":"how to reduce the calculation time of optimizer.step() without loss.item() in Pytorch","Body":"<p>When I\u3000test my imagenet model's calculation time, I found that if\u3000tensor like loss or top1, top5 don't be transformed by function <code>tensor.item()<\/code> or <code>tensor.cpu().data.numpy()<\/code>, the calculation time of <code>optimizer.step()<\/code> will be greatly increased as shown below.<\/p>\n\n<pre><code>total_top1 = torch.tensor(0.).cuda()  # The first case\ntotal_top5 = torch.tensor(0.).cuda()  # The first case\ntotal_loss = torch.tensor(0.).cuda()  # The first case\n# total_top1 = 0.0 # The second case\n# total_top5 = 0.0 # The second case \n# total_loss = 0.0 # The second case\n\noutputs, loss = model(images, path, labels, criterion, gpu_nums)\n\nst = time.time()\noptimizer.step()\noptimizer.zero_grad()\ntotal_to += time.time() - st # step() time\n\ntop1, top5 = accuracy(outputs, labels, topk=(1, 5))\n\nst = time.time()\ntotal_counter += images.size(0)\n# total_top1 += top1.cpu().data.numpy() #  or top1.item() # The second case\n# total_top5 += top5.cpu().data.numpy() # The second case\n# total_loss += loss.cpu().data.numpy() # The second case\ntotal_top1 += top1 # The first case\ntotal_top5 += top5 # The first case\ntotal_loss += loss # The first case\ntotal_ti += time.time() - st # data transform time\n<\/code><\/pre>\n\n<p>In first case, <code>total_to<\/code> = 18.0s, <code>total_ti<\/code> = 0.2s for 100 iterations calculation.But in second case, <code>total_to<\/code> = 8s, <code>total_ti<\/code> = 9s.<\/p>\n\n<p>I tried to find the reason for this kind of difference.The most plausible answer is that <code>item()<\/code> prevents the newly calculated variable from entering the calculation graph, so the <code>step()<\/code>'s time is reduced.However, when I run the code below:<\/p>\n\n<pre><code>outputs, loss = model(images, path, labels, criterion, gpu_nums)\n\nst = time.time()\noptimizer.step()\noptimizer.zero_grad()\ntotal_to += time.time() - st # step() time\n\nst = time.time()\ntotal_counter += images.size(0)\n# loss.item() \ntotal_ti += time.time() - st # data transform time\n<\/code><\/pre>\n\n<p><code>total_to<\/code> is still around 18s, and <code>total_ti<\/code> = 0s.When using loss.item(), <code>total_to<\/code> change to 8s, but <code>total_ti<\/code> changes to 9s.<\/p>\n\n<p>So could you tell me what's the reason there in pytorch and how could I make <code>total_ti<\/code> be 0s and <code>total_to<\/code> be 8s? Thank you very much for your answers.<\/p>\n","answers":[]}
+{"QuestionId":57561435,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-19T17:28:48.197","AcceptedAnswerId":null,"OwnerUserId":3050590.0,"Title":"TypeError: 'Sequential' object does not support indexing when looping over keras model","Body":"<p>I am trying to ensemble the pre-trained models into one model by the voting system. I am referring to this <a href=\"https:\/\/www.kaggle.com\/mgiraygokirmak\/keras-cnn-multi-model-ensemble-with-voting\" rel=\"nofollow noreferrer\">blog<\/a> for the same<\/p>\n\n<p>Here is the code<\/p>\n\n<pre><code>for i in os.listdir(model_root): \/\/loading all the models\n    print(i)\n    filename = model_root + \"\/\" + i\n    # load model\n    model = load_model(filename, custom_objects={'KerasLayer': hub.KerasLayer})\n    models.append(model)\nprint(len(models))\n\n  #To fit the loaded models to the data and saving it to an array fit_models\n\nsteps_per_epoch = image_data.samples \/\/ image_data.batch_size\nbatch_stats = CollectBatchStats()\nvalidation_steps = image_data_val.samples \/ image_data_val.batch_size\nfor i in range(len(models)):\n    model[i].fit_generator((item for item in image_data), epochs=2,\n              steps_per_epoch=steps_per_epoch, #callbacks=[batch_stats],\n              validation_data=(item for item in image_data_val), validation_steps=validation_steps, verbose=2)\n    fit_models.append(model[i])\n<\/code><\/pre>\n\n<p>Here is the traceback to the error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"&lt;input&gt;\", line 1, in &lt;module&gt;\n  File \"C:\\Program Files\\JetBrains\\PyCharm 2019.2\\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 \"C:\\Program Files\\JetBrains\\PyCharm 2019.2\\helpers\\pydev\\_pydev_imps\\_pydev_execfile.py\", line 18, in execfile\n    exec(compile(contents+\"\\n\", file, 'exec'), glob, loc)\n  File \"C:\/Users\/Pawandeep\/Desktop\/Python projects\/ensemble_image.py\", line 90, in &lt;module&gt;\n    model[i].fit_generator((item for item in image_data), epochs=2,\nTypeError: 'Sequential' object does not support indexing\n<\/code><\/pre>\n\n<p>UPDATE---------------------<\/p>\n\n<p>After following the solution from the comments, I am not getting the original error anymore but a new error.<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"&lt;input&gt;\", line 1, in &lt;module&gt;\n  File \"C:\\Program Files\\JetBrains\\PyCharm 2019.2\\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 \"C:\\Program Files\\JetBrains\\PyCharm 2019.2\\helpers\\pydev\\_pydev_imps\\_pydev_execfile.py\", line 18, in execfile\n    exec(compile(contents+\"\\n\", file, 'exec'), glob, loc)\n  File \"C:\/Users\/Pawandeep\/Desktop\/Python projects\/ensemble_image.py\", line 89, in &lt;module&gt;\n    validation_data=(item for item in image_data_val), validation_steps=validation_steps, verbose=2)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 1433, in fit_generator\n    steps_name='steps_per_epoch')\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_generator.py\", line 264, in model_iteration\n    batch_outs = batch_function(*batch_data)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 1153, in train_on_batch\n    extract_tensors_from_dataset=True)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 2692, in _standardize_user_data\n    y, self._feed_loss_fns, feed_output_shapes)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_utils.py\", line 549, in check_loss_and_target_compatibility\n    ' while using as loss `' + loss_name + '`. '\nValueError: A target array with shape (32, 3) was passed for an output of shape (None, 2) while using as loss `binary_crossentropy`. This loss expects targets to have the same shape as the output.\n<\/code><\/pre>\n\n<p>My data looks like this:<\/p>\n\n<pre><code>Image batch shape:  (32, 224, 224, 3)\nLabel batch shape:  (32, 3)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57561479,"AnswerCount":1,"Tags":"<python><machine-learning><image-processing><keras><conv-neural-network>","CreationDate":"2019-08-19T17:32:14.707","AcceptedAnswerId":57561538.0,"OwnerUserId":11917598.0,"Title":"How do I get the images from a path to pass in model.fit method?","Body":"<p>I'm getting a list of all my training images from a path with this method:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def ReadImages(Path):\n    ImageList = list()\n    LabelList = list()\n\n    # Get all subdirectories\n    FolderList = os.listdir(Path)\n\n    # Loop over each directory\n    for File in FolderList:\n        if(os.path.isdir(Path + os.path.sep + File)):\n            for Image in os.listdir(Path + os.path.sep + File):\n                # Add the image path to the list\n                ImageList.append(Path + os.path.sep + File + os.path.sep + Image)\n\n                # Add a label for each image and remove the file extension\n                LabelList.append(File.split(\".\")[0])\n        else:\n            ImageList.append(Path + os.path.sep + File)\n\n            # Add a label for each image and remove the file extension\n            LabelList.append(File.split(\".\")[0])\n\n    return ImageList, LabelList\n<\/code><\/pre>\n\n<p>And now I want to call the keras method 'model.fit(data,labels,epochs, bs)' with that data<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential()\nmodel.add(Dense(32, activation='tanh', input_dim=1))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(optimizer='rmsprop',\n              loss='binary_crossentropy',\n              metrics=['accuracy'])\n\ndata, labels = ReadImages(TRAIN_DIR)\n\n# Train the model, iterating on the data in batches of 32 samples\nmodel.fit(np.array(data), np.array(labels), epochs=10, batch_size=32)\n<\/code><\/pre>\n\n<p>But it shows this error: <\/p>\n\n<pre><code>return array(a, dtype, copy=False, order=order)\n\nValueError: could not convert string to float: 'train_data\/\\\\non_pdr\\\\im0008.ppm'\n<\/code><\/pre>\n\n<p>How do I convert my list of paths in a list to feed my train model data?<\/p>\n\n<p>My files are like <\/p>\n\n<blockquote>\n  <p>(['train_data\/non_pdr\\im0001.ppm', 'train_data\/non_pdr\\im0002.ppm',\n  'train_data\/non_pdr\\im0003.ppm', 'train_data\/non_pdr\\im0004.ppm',\n  'train_data\/non_pdr\\im0005.ppm', 'train_data\/non_pdr\\im0006.ppm',\n  'train_data\/non_pdr\\im0007.ppm', 'train_data\/non_pdr\\im0008.ppm',\n  'train_data\/non_pdr\\im0009.ppm', 'train_data\/non_pdr\\im0010.ppm',\n  'train_data\/non_pdr\\im0011.ppm', 'train_data\/non_pdr\\im0012.ppm',\n  'train_data\/non_pdr\\im0013.ppm', 'train_data\/non_pdr\\im0014.ppm',\n  'train_data\/non_pdr\\im0015.ppm', 'train_data\/non_pdr\\im0016.ppm',\n  'train_data\/non_pdr\\im0017.ppm', 'train_data\/non_pdr\\im0018.ppm',\n  'train_data\/non_pdr\\im0019.ppm', 'train_data\/non_pdr\\im0020.ppm',\n  'train_data\/non_pdr\\im0021.ppm', 'train_data\/non_pdr\\im0022.ppm', \n  ...<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"57561538","CreationDate":"2019-08-19T17:37:39.130","ParentId":null,"OwnerUserId":"6321559","Title":null,"Body":"<p>Sorry, my old answer (below) still stands as best practice, but this is more of an existential issue.  You're passing an array of path items to your <code>Sequential<\/code> model.  What you need is a list of image arrays, so you need to load those images (<a href=\"https:\/\/stackoverflow.com\/questions\/7762948\/how-to-convert-an-rgb-image-to-numpy-array\">how to convert an RGB image to numpy array?<\/a>).  <\/p>\n\n<p>See the docs for the model for the variable types: <a href=\"https:\/\/keras.io\/models\/sequential\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/models\/sequential\/<\/a><\/p>\n\n<h1>Old answer<\/h1>\n\n<p>You've got a mix of slashes in your <code>ValueError<\/code> traceback, I'm not sure if that's from the actual file names or something to do with your <code>path_sep<\/code> variable.  It's better to use the <code>os<\/code> module built-in path joining function <code>os.path.join<\/code>.  That might be the solution to your issue.  The question would be more complete with a list of sample directory files.<\/p>\n\n<pre><code># Instead of this:\nPath + os.path.sep + File \n\n# Do this:\nos.path.join(Path, File)\n<\/code><\/pre>\n\n<p>Also, bonus:<\/p>\n\n<pre><code># This might not work depending on the filename:\nFile.split('.')[0] # what if the file is named 2019.08.19_file.txt?\n\n# Try:\nos.path.splitext(File)[0]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57561620,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><linear-regression>","CreationDate":"2019-08-19T17:44:12.373","AcceptedAnswerId":57562819.0,"OwnerUserId":10018602.0,"Title":"Simple TensorFlow Linear Regression Returns \"can't multiply sequence by non-int\" ValueError","Body":"<p>I'm new to TensorFlow and creating linear regressions, but I'm not seeing my issue here. I followed <a href=\"https:\/\/www.geeksforgeeks.org\/linear-regression-using-tensorflow\/\" rel=\"nofollow noreferrer\">this guide<\/a> to create my first simple (univariate) linear regression model, but am running into the error:<\/p>\n\n<p><strong>UPDATE WITH NEW PRINT OUTPUTS<\/strong><\/p>\n\n<pre><code>_x :  -1000\n_y :  10\n_x :  -900\n_y :  12\n_x :  -950\n_y :  12\n...\nc :  nan\nc dtype :  float32\nEpoch 50 : cost = nan W = nan b = nan\n\n_x :  -1000\n_y :  10\n_x :  -900\n_y :  12\n_x :  -950\n_y :  12\n...\nc :  nan\nc dtype :  float32\nEpoch 100 : cost = nan W = nan b = nan\n...\nEpoch 150 : cost = nan W = nan b = nan\n...\nEpoch 200 : cost = nan W = nan b = nan\n...\nEpoch 450 : cost = nan W = nan b = nan\n...\nEpoch 500 : cost = nan W = nan b = nan\ntraining_cost :  nan\ntraining_cost dtype :  float32\nweight :  nan\nweight :  float32\nbias :  nan\nbias dtype :  float32\nTraceback (most recent call last):\n  File \"editabletoworkouterrors.py\", line 77, in &lt;module&gt;\n    new_green_dur = green_light_duration_new(current_reward, current_green)\n  File \"editabletoworkouterrors.py\", line 66, in green_light_duration_new\n    green_light_duration_new = weight * x + bias\nTypeError: can't multiply sequence by non-int of type 'numpy.float32'\n<\/code><\/pre>\n\n<p>So I have a bunch of nan's for some reason? When I tried this in the program I'm working on which has a large list of green light durations and rewards, sometimes it'll start with having numbers in the weight, x and b values as it should, but then it starts displaying nan's again during the later epochs, if that makes sense?<\/p>\n\n<p>I'm making a traffic simulator, and I'd like it to be able to determine what the best \"green light duration\" would be, based on a previous reward (car wait time change for the lane in the reward). I've created a simple working code for someone to replicate:<\/p>\n\n<p><strong>UPDATED WITH NEW PRINTS IN CODE<\/strong><\/p>\n\n<pre><code>import numpy as np\nimport random\nimport matplotlib.pyplot as plt\nimport tensorflow as tf\nimport warnings\n\nwarnings.simplefilter(action='once', category=FutureWarning) # future warnings annoy me\n\n# set the epsilon for this episode\n\n# Start with empty lists\nreward = []\ngreen_light = []\n\n# add in a couple of rewards and light durations\ncurrent_reward = [-1000,-900,-950]\ncurrent_green = [10,12,12]\n\n# Pass in reward and green_light\ndef green_light_duration_new(current_reward, current_green):\n    # Predicting the best light duration based on previous rewards.\n    # predict the best duration based on previous step's reward value, using simple linear regression model\n    x = current_reward\n    y = current_green\n    n = len(x)\n    # Plot of Training Data  \n    plt.scatter(x, y) \n    plt.xlabel('Reward') \n    plt.ylabel('Green Light Duration') \n    plt.title(\"Training Data\") \n    plt.show() \n\n    X = tf.placeholder(\"float\") \n    Y = tf.placeholder(\"float\") \n    W = tf.Variable(np.random.randn(), name = \"W\") \n    b = tf.Variable(np.random.randn(), name = \"b\") \n    learning_rate = 0.01\n    training_epochs = 500\n    # Hypothesis \n    y_pred = tf.add(tf.multiply(X, W), b) \n    print('y_pred : ', y_pred)\n    print('y_pred dtype : ', y_pred.dtype)\n    # Mean Squared Error Cost Function \n    cost = tf.reduce_sum(tf.pow(y_pred-Y, 2)) \/ (2 * n)\n    print('cost : ', cost)\n    print('cost dtype: ', cost.dtype)\n    # Gradient Descent Optimizer \n    optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)    \n    # Global Variables Initializer \n    init = tf.global_variables_initializer()\n    # Starting the Tensorflow Session \n    with tf.Session() as sess: \n        # Initializing the Variables \n        sess.run(init) \n        # Iterating through all the epochs \n        for epoch in range(training_epochs): \n            # Feeding each data point into the optimizer using Feed Dictionary \n            for (_x, _y) in zip(x, y): \n                print('_x : ',_x)\n                print('_y : ',_y)\n                sess.run(optimizer, feed_dict = {X : _x, Y : _y}) \n            # Displaying the result after every 50 epochs \n            if (epoch + 1) % 50 == 0: \n                # Calculating the cost a every epoch \n                c = sess.run(cost, feed_dict = {X : x, Y : y}) \n                print('c : ', c)\n                print('c dtype : ', c.dtype)\n                print(\"Epoch\", (epoch + 1), \": cost =\", c, \"W =\", sess.run(W), \"b =\", sess.run(b)) \n        # Storing necessary values to be used outside the Session \n        training_cost = sess.run(cost, feed_dict ={X: x, Y: y}) \n        print('training_cost : ', training_cost)\n        print('training_cost dtype : ', training_cost.dtype)\n        weight = sess.run(W)\n        print('weight : ', weight)\n        print('weight : ', weight.dtype)\n        bias = sess.run(b)\n        print('bias : ', bias)\n        print('bias dtype : ', bias.dtype)\n    # Calculating the predictions \n    green_light_duration_new = weight * x + bias \n    print(\"Training cost =\", training_cost, \"Weight =\", weight, \"bias =\", bias, '\\n')\n    # Plotting the Results \n    plt.plot(x, y, 'ro', label ='Original data') \n    plt.plot(x, green_light_duration_new, label ='Fitted line') \n    plt.title('Linear Regression Result') \n    plt.legend() \n    plt.show() \n    return green_light_duration_new\n\n# Go to the training function\nnew_green_dur = green_light_duration_new(current_reward, current_green)\n\n# Append the predicted green light to its list\ngreen_light.append(new_green_dur)\n\n# Go on to run the rest of the simulation with the new green light duration,\n# and append its subsequent reward to current_reward list to run again later.\n<\/code><\/pre>\n\n<p>Also, if there's a way for the regression model to keep the green light duration to be between a minimum of 10 (seconds) and a maximum value of 120 seconds, that would be spectacular too, I just don't know how to work that in. Thanks a lot!<\/p>\n","answers":[{"AnswerId":"57562819","CreationDate":"2019-08-19T19:25:47.680","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>... and there's the problem!  Current values of items are<\/p>\n\n<pre><code>x = [-1000,-900,-950]    # type list\nweight = nan             # type numpy.float32\nbias   = nan             # type numpy.float32\n<\/code><\/pre>\n\n<p><code>x<\/code> is a simple Python list, <em>not<\/em> a numpy array.  Python doesn't have vectorized operations such as <code>float * list<\/code><\/p>\n"}]}
+{"QuestionId":57561668,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2019-08-19T17:48:11.413","AcceptedAnswerId":57561955.0,"OwnerUserId":10718217.0,"Title":"Type object 'Food101' has no attribute 'load_data'","Body":"<p>I wrote this code:<\/p>\n\n<pre><code>from __future__ import absolute_import\nfrom __future__ import division\nfrom __future__ import print_function\n\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport tensorflow as tf\n\nimport tensorflow_datasets as tfds\n\nfood = tfds.image.food101.Food101\n(train_images, train_labels), (test_images, test_labels) = food.load_data()\n<\/code><\/pre>\n\n<p>and I got the following error:<\/p>\n\n<pre><code>AttributeError\nTraceback (most recent call last)\n&lt;ipython-input-10-92f141508c7c&gt; in &lt;module&gt;()\n            9 import tensorflow_datasets as tfds\n         1  0 food = tfds.image.food101.Food101\n      ---&gt; 11 (train_images, train_labels), (test_images, test_labels) = \n      food.load_data()\nAttributeError: type object 'Food101' has no attribute 'load_data'\n<\/code><\/pre>\n\n<p>What should I do to overcome this error?<\/p>\n","answers":[{"AnswerId":"57561955","CreationDate":"2019-08-19T18:11:22.420","ParentId":null,"OwnerUserId":"8407784","Title":null,"Body":"<p>From Tensorflow page, the correct way for you to load the Dataset is:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>data, info = tfds.load(\"desired_dataset\", with_info=True)\n<\/code><\/pre>\n\n<p>So, in your code you should load it with:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>food, info = tfds.load(\"food101\", with_info=True)\n<\/code><\/pre>\n\n<p>If that does not work, <strong>load it's builder<\/strong>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>builder = tfds.builder(\"tfds.image.food101.Food101\")\nbuilder.download_and_prepare()\ndatasets = builder.as_dataset()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/datasets\/catalog\/overview\" rel=\"nofollow noreferrer\">Reference<\/a>\n<a href=\"https:\/\/www.tensorflow.org\/datasets\/catalog\/food101?hl=th\" rel=\"nofollow noreferrer\">Dataset<\/a><\/p>\n"},{"AnswerId":"57562353","CreationDate":"2019-08-19T18:45:34.823","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>Code to get the TensorFlow dataset loaded will be something like this:<\/p>\n\n<pre><code>import tensorflow_datasets as tfds\n\nfood_images, food_labels = tfds.load(name='food101', split=tfds.Split.TRAIN)\n<\/code><\/pre>\n\n<p>But unfortunately, in the current latest version of TensorFlow Datasets (1.1.0), Food101 dataset is not present.\nYou can check it <a href=\"https:\/\/github.com\/tensorflow\/datasets\/tree\/v1.1.0\/tensorflow_datasets\/image\" rel=\"nofollow noreferrer\">1.1.0 GitHub version<\/a> that food101.py file is missing, but if you change to master branch, the food101.py will be present. If you need food101.py, you will have to build TensorFlow Dataset from source or else download the dataset directly from the download link.<\/p>\n"}]}
+{"QuestionId":57561781,"AnswerCount":0,"Tags":"<keras><tensorboard><tf.keras>","CreationDate":"2019-08-19T17:56:56.740","AcceptedAnswerId":null,"OwnerUserId":2130515.0,"Title":"TensorBoard keras does not show distribution and histogram using training data","Body":"<p>I want to monitor gradient and weights during training a model. here is a simple example to reproduce the problem. It seems that keras deos not record the distributions and the histograms.\nI am using Tensorflow 1.14 and keras 2.4 on ubuntu 18.04<\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\n\ndata = np.random.rand(200, 2)\nexpected = np.random.randint(2, size=200).reshape(-1, 1)\n\ndataFrame = pd.DataFrame(data, columns=['a', 'b'])\nexpectedFrame = pd.DataFrame(expected, columns=['expected'])\n\ndataFrameTrain, dataFrameTest = dataFrame[:100], dataFrame[-100:]\nexpectedFrameTrain, expectedFrameTest = expectedFrame[:100], expectedFrame[-100:]\n\n\ndef generator(X_data, y_data, batch_size):\n    samples_per_epoch = X_data.shape[0]\n    number_of_batches = samples_per_epoch \/ batch_size\n    counter = 0\n\n    while 1:\n\n        X_batch = np.array(X_data[batch_size * counter:batch_size * (counter + 1)]).astype('float32')\n        y_batch = np.array(y_data[batch_size * counter:batch_size * (counter + 1)]).astype('float32')\n        counter += 1\n        yield (X_batch, y_batch)\n\n        # restart counter to yeild data in the next epoch as well\n        if counter &gt;= number_of_batches:\n            counter = 0\n\n\nfrom tensorflow.keras.models import Model\nfrom tensorflow.keras.layers import Dense, Input\n\nimg_input = Input(shape=(2))\n\nl1 = Dense(12, activation='relu')(img_input)\nl2 = Dense(1, activation='sigmoid')(l1)\nmodel = Model(inputs=[img_input], outputs=[l2])\nmodel.compile(loss='binary_crossentropy', optimizer='adadelta', metrics=['accuracy'])\n\n# Train the model using generator vs using the full batch\nbatch_size = 8\n# Tensorboard callbacks\nfrom tensorflow.keras.callbacks import TensorBoard\ntb = TensorBoard(log_dir='.\/logs',\n                 histogram_freq=1,\n                 write_graph=True,\n                 write_images=True,\n                 update_freq='epoch',\n                 write_grads=True,\n                 profile_batch=1)\n# train with fit_generator\n\nmodel.fit_generator(generator(dataFrameTrain, expectedFrameTrain, batch_size), epochs=3,\n                    steps_per_epoch=dataFrame.shape[0] \/ batch_size, callbacks=[tb]) \n<\/code><\/pre>\n\n<h1>if I specify validation_data, it works:<\/h1>\n\n<pre><code>            validation_data=generator(dataFrameTest, expectedFrameTest, batch_size * 2),\n            validation_steps=dataFrame.shape[0] \/ batch_size * 2\n<\/code><\/pre>\n\n<p>So, why It implemented this way ?, and does all statistics are gathered based on validation data or training data ?!!<\/p>\n","answers":[]}
+{"QuestionId":57561944,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2019-08-19T18:09:56.593","AcceptedAnswerId":null,"OwnerUserId":9182282.0,"Title":"Unintentional dimension added for tf.nn.conv1d","Body":"<p>The two tensors I'm feeding into <code>tf.nn.conv1d()<\/code> are <code>x<\/code> and <code>kernel<\/code> with sizes [16,512,32] and [1,1,32] respectively.  I've been able to run this code on my local machine, but I get an error when running it on a remote cluster.  <\/p>\n\n<p>I get a Value Error stating that the dimensions must be equal.  The message shows the resulting shapes for <code>x<\/code> and <code>kernel<\/code> as [16,1,512,32] and [1,1,1,32] respectively.<\/p>\n\n<ol>\n<li>Both machines are running the same version of <code>tensorflow-gpu-1.14.0<\/code> and <code>python-3.6<\/code>.<\/li>\n<li>I've printed the tensor shapes and they are confirmed to be the anticipated sizes.<\/li>\n<\/ol>\n\n<p>The wrapper is shown below that I'm calling within my code:\n<code>filter_size =1<\/code>\n<code>input_dim = 32<\/code>\n<code>output_dim = 32<\/code><\/p>\n\n<pre><code>     def conv1d_linear(self,x, filter_size,input_dim,output_dim,stride=1,name = 'conv1d_linear'):\n        with tf.variable_scope(name):\n            # Conv1D wrapper, with bias and no activation\n            print(x)\n            kernel = tf.get_variable('kernel', shape = [filter_size, input_dim, output_dim],initializer=tf.contrib.layers.xavier_initializer())\n            print(kernel)\n            bias =  tf.get_variable('bias', shape = [output_dim],initializer=tf.contrib.layers.xavier_initializer())\n            x = tf.nn.conv1d(x, kernel, stride=stride, padding='SAME')\n            x = tf.nn.bias_add(x, bias)\n            x = tf.contrib.layers.batch_norm(x)\n            return x \n\n<\/code><\/pre>\n\n<p>Error Message<\/p>\n\n<pre><code>Using TensorFlow backend.\nTensorflow Version: 1.14.0\nBuilding model...\nTensor(\"stfnet\/resstack0\/stft_pool0\/Relu:0\", shape=(16, 512, 32), dtype=float32, device=\/device:GPU:0)\n&lt;tf.Variable 'stfnet\/resstack1\/init_conv_lin0\/kernel:0' shape=(1, 1, 32) dtype=float32_ref&gt;\nTraceback (most recent call last):\n  File \"\/nv\/hp20\/nblinn6\/.conda\/envs\/mytfenv\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1659, in _create_c_op\n    c_op = c_api.TF_FinishOperation(op_desc)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: Dimensions must be equal, but are 32 and 1 for 'stfnet\/resstack1\/init_conv_lin0\/conv1d\/Conv2D' (op: 'Conv2D') with input shapes: [16,1,512,32], [1,1,1,32].\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"\/nv\/hp20\/nblinn6\/data\/MODID_mystfresnet6_final_PB_2gpu.py\", line 529, in &lt;module&gt;\n    mymodel.train_init()\n  File \"\/nv\/hp20\/nblinn6\/data\/MODID_mystfresnet6_final_PB_2gpu.py\", line 413, in train_init\n    self.logits_train = self.ModIDNet(self.batch_feature[0:int(self.gpu_batch_size)],self.tau,self.keep_rate,train = True,reuse = reuse_flg,name = 'stfnet')\n  File \"\/nv\/hp20\/nblinn6\/data\/MODID_mystfresnet6_final_PB_2gpu.py\", line 340, in ModIDNet\n    x = self.resstack(x,tau,self.sensor_axis,self.maps[1],reuse = reuse,series_size = 512,name = 'resstack1')   \n  File \"\/nv\/hp20\/nblinn6\/data\/MODID_mystfresnet6_final_PB_2gpu.py\", line 310, in resstack\n    x1 = self.conv1d_linear(x,1,map_in,map_out, name = 'init_conv_lin0')\n  File \"\/nv\/hp20\/nblinn6\/data\/MODID_mystfresnet6_final_PB_2gpu.py\", line 302, in conv1d_linear\n    x = tf.nn.conv1d(x, kernel, stride=stride, padding='SAME')\n  File \"\/nv\/hp20\/nblinn6\/.conda\/envs\/mytfenv\/lib\/python3.6\/site-packages\/tensorflow\/python\/util\/deprecation.py\", line 574, in new_func\n    return func(*args, **kwargs)\n  File \"\/nv\/hp20\/nblinn6\/.conda\/envs\/mytfenv\/lib\/python3.6\/site-packages\/tensorflow\/python\/util\/deprecation.py\", line 574, in new_func\n    return func(*args, **kwargs)\n  File \"\/nv\/hp20\/nblinn6\/.conda\/envs\/mytfenv\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/nn_ops.py\", line 3482, in conv1d\n    data_format=data_format)\n  File \"\/nv\/hp20\/nblinn6\/.conda\/envs\/mytfenv\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gen_nn_ops.py\", line 1026, in conv2d\n    data_format=data_format, dilations=dilations, name=name)\n  File \"\/nv\/hp20\/nblinn6\/.conda\/envs\/mytfenv\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 788, in _apply_op_helper\n    op_def=op_def)\n  File \"\/nv\/hp20\/nblinn6\/.conda\/envs\/mytfenv\/lib\/python3.6\/site-packages\/tensorflow\/python\/util\/deprecation.py\", line 507, in new_func\n    return func(*args, **kwargs)\n  File \"\/nv\/hp20\/nblinn6\/.conda\/envs\/mytfenv\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 3300, in create_op\n    op_def=op_def)\n  File \"\/nv\/hp20\/nblinn6\/.conda\/envs\/mytfenv\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1823, in __init__\n    control_input_ops)\n  File \"\/nv\/hp20\/nblinn6\/.conda\/envs\/mytfenv\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1662, in _create_c_op\n    raise ValueError(str(e))\nValueError: Dimensions must be equal, but are 32 and 1 for 'stfnet\/resstack1\/init_conv_lin0\/conv1d\/Conv2D' (op: 'Conv2D') with input shapes: [16,1,512,32], [1,1,1,32].\n<\/code><\/pre>\n","answers":[{"AnswerId":"57574418","CreationDate":"2019-08-20T13:07:27.653","ParentId":null,"OwnerUserId":"9182282","Title":null,"Body":"<p>I reinstalled the version of <code>tensorflow-gpu==1.14.0<\/code> just as a troubleshooting step and it seems to be working now.  Unsure what the original issue was.<\/p>\n"}]}
+{"QuestionId":57561947,"AnswerCount":0,"Tags":"<python><python-2.7><pytorch>","CreationDate":"2019-08-19T18:10:31.600","AcceptedAnswerId":null,"OwnerUserId":10824718.0,"Title":"import hausdorff.distances as distances ImportError: No module named distances","Body":"<p>I am using python 2.7 and I use the <\/p>\n\n<pre><code>from hausdorff import hausdorff\n<\/code><\/pre>\n\n<p>I get this error although I am using pip2 install hausdorff and it syas requirements already satisfied.<\/p>\n\n<pre><code>    from hausdorff import hausdorff\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/hausdorff\/__init__.py\", line 1, in &lt;module&gt;\n    from .hausdorff import *\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/hausdorff\/hausdorff.py\", line 3, in &lt;module&gt;\n    import hausdorff.distances as distances\nImportError: No module named distances\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57561966,"AnswerCount":0,"Tags":"<python><tensorflow><keras><ensemble-learning>","CreationDate":"2019-08-19T18:12:14.563","AcceptedAnswerId":null,"OwnerUserId":3050590.0,"Title":"ValueError: A target array with shape (32, 3) was passed for an output of shape (None, 2) while using as loss `binary_crossentropy`. In Keras model","Body":"<p>I am trying to ensemble the Keras binary pre-trained models into one multi-class model by the voting system. Binary pre-trained models are trained on different classes each. To ensemble the model, I am referring to this <a href=\"https:\/\/www.kaggle.com\/mgiraygokirmak\/keras-cnn-multi-model-ensemble-with-voting\" rel=\"nofollow noreferrer\">blog<\/a> for the same<\/p>\n\n<p>Here is the code<\/p>\n\n<pre><code>for i in os.listdir(model_root): \/\/loading all the models\n    print(i)\n    filename = model_root + \"\/\" + i\n    # load model\n    model = load_model(filename, custom_objects={'KerasLayer': hub.KerasLayer})\n    models.append(model)\nprint(len(models))  \/\/3\n\n  #To fit the loaded models to the data and saving it to an array fit_models\n\nsteps_per_epoch = image_data.samples \/\/ image_data.batch_size\nbatch_stats = CollectBatchStats()\nvalidation_steps = image_data_val.samples \/ image_data_val.batch_size\nfor i in range(len(models)):\n    model[i].fit_generator((item for item in image_data), epochs=2,\n              steps_per_epoch=steps_per_epoch, #callbacks=[batch_stats],\n              validation_data=(item for item in image_data_val), validation_steps=validation_steps, verbose=2)\n    fit_models.append(model[i])\n<\/code><\/pre>\n\n<p>Here is the traceback to the error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"&lt;input&gt;\", line 1, in &lt;module&gt;\n  File \"C:\\Program Files\\JetBrains\\PyCharm 2019.2\\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 \"C:\\Program Files\\JetBrains\\PyCharm 2019.2\\helpers\\pydev\\_pydev_imps\\_pydev_execfile.py\", line 18, in execfile\n    exec(compile(contents+\"\\n\", file, 'exec'), glob, loc)\n  File \"C:\/Users\/Pawandeep\/Desktop\/Python projects\/ensemble_image.py\", line 89, in &lt;module&gt;\n    validation_data=(item for item in image_data_val), validation_steps=validation_steps, verbose=2)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 1433, in fit_generator\n    steps_name='steps_per_epoch')\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_generator.py\", line 264, in model_iteration\n    batch_outs = batch_function(*batch_data)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 1153, in train_on_batch\n    extract_tensors_from_dataset=True)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 2692, in _standardize_user_data\n    y, self._feed_loss_fns, feed_output_shapes)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_utils.py\", line 549, in check_loss_and_target_compatibility\n    ' while using as loss `' + loss_name + '`. '\nValueError: A target array with shape (32, 3) was passed for an output of shape (None, 2) while using as loss `binary_crossentropy`. This loss expects targets to have the same shape as the output.\n<\/code><\/pre>\n\n<p>Data definition<\/p>\n\n<pre><code>#define data\nimage_generator = tf.keras.preprocessing.image.ImageDataGenerator(rescale=1 \/ 255, validation_split=0.20)\nIMAGE_SIZE= (224,224)\nimage_data = image_generator.flow_from_directory(str(data_root), target_size=IMAGE_SIZE, subset='training')\nimage_data_val = image_generator.flow_from_directory(str(data_root), target_size=IMAGE_SIZE, subset='validation')\n<\/code><\/pre>\n\n<p>My data looks like this:<\/p>\n\n<pre><code>Image batch shape:  (32, 224, 224, 3)\nLabel batch shape:  (32, 3)\n<\/code><\/pre>\n\n<p>I tried to print out the shape of each model in models array. It is <\/p>\n\n<pre><code>(32, 2)\n<\/code><\/pre>\n\n<p>Now I understand the problem. So the problem is I have binary models trained on each class. That is why its shape is 32*3. I want to ensemble these binary models into an ensemble model so that the addition of each model(class) becomes a multiclass model. Then based on the prediction of this model I want to label my dataset. So how can I achieve this now? <\/p>\n","answers":[]}
+{"QuestionId":57562244,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><nlp>","CreationDate":"2019-08-19T18:36:23.120","AcceptedAnswerId":null,"OwnerUserId":7691008.0,"Title":"Unable to read from a tensorflow checkpoint for finetuning","Body":"<p>I am trying to use a pre-trained BERT model for fine tuning with SST2 data processor. But when I give the checkpoint of the pre-trained model, it is showing that \"Key output_bias not found in checkpoint\".<\/p>\n\n<p>I thought it might be due to errors in the pre-trained BERT model checkpoint. So I did the pre-training again. But, I am still facing the same issue. <\/p>\n\n<pre><code>TASK = 'STS' #@param {type:\\\"string\\\"}\nTASK_DATA_DIR = 'glue_data\/STS-B\/'# + TASK\n\noutput_dir = 'trained_model\/observation'\ntf.gfile.MakeDirs(output_dir)\n\nBERT_MODEL = path + 'multi_cased_L-12_H-768_A-12\/' \nVOCAB_FILE = os.path.join(BERT_MODEL, 'vocab.txt')   \nCONFIG_FILE = os.path.join(BERT_MODEL, 'bert_config.json')   \nINIT_CHECKPOINT = os.path.join(BERT_MODEL, 'bert_model.ckpt')   \nDO_LOWER_CASE = BERT_MODEL.startswith('cased')\n\ntokenizer = tokenization.FullTokenizer(vocab_file=VOCAB_FILE, \ndo_lower_case=DO_LOWER_CASE)\n\nTRAIN_BATCH_SIZE = 1   \nEVAL_BATCH_SIZE = 8   \nPREDICT_BATCH_SIZE = 8   \nLEARNING_RATE = 2e-5   \nNUM_TRAIN_EPOCHS = 3.0   \nMAX_SEQ_LENGTH = 128   \n\nprocessors = {   \n    \"sts\": run_classifier.StsProcessor,    \n}   \n\nprocessor = processors[TASK.lower()]()    \nlabel_list = processor.get_labels()   \n<\/code><\/pre>\n\n<p>The error is:<\/p>\n\n<blockquote>\n  <p>NotFoundError: Restoring from checkpoint failed. This is most likely\n  due to a Variable name or other graph key that is missing from the\n  checkpoint. Please ensure that you have not altered the graph expected\n  based on the checkpoint.   Original error: Key output_bias not found\n  in checkpoint [[node save\/RestoreV2 (defined at\n  \/home\/subraas3\/.conda\/envs\/tensorflow_13\/lib\/python3.7\/\n  site-packages\/tensorflow_estimator\/python\/estimator\/estimator.py:1403)\n  ]] [[node save\/RestoreV2 (defined at\n  \/home\/subraas3\/.conda\/envs\/tensorflow_13\/lib\/python3.7\/site-packages\/tensorflow_estimator\/python\/estimator\/estimator.py:1403)\n  ]]<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"57567063","CreationDate":"2019-08-20T05:15:40.027","ParentId":null,"OwnerUserId":"5806666","Title":null,"Body":"<p>As pointed out in the error message, this error occurs if <\/p>\n\n<ol>\n<li>A layer in the tf compute graph is renamed. i.e the name of the layer in the pre-trained checkpoint is different from what is there in the provided API <\/li>\n<li>A new layer is added to API , i.e network architecture is changed. Or,<\/li>\n<li>A layer that is present in the pre-trained checkpoint is removed (unlikey).<\/li>\n<\/ol>\n\n<p>Please check that the version of the bert API is same as the pretrained checkpoint version. If they are same, you might want to manually check if the tf graph in the checkpoint is consistent with the API, using this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/inspect_checkpoint.py\" rel=\"nofollow noreferrer\">tool.<\/a><\/p>\n"}]}
+{"QuestionId":57562297,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><loss-function>","CreationDate":"2019-08-19T18:41:07.890","AcceptedAnswerId":null,"OwnerUserId":11947142.0,"Title":"Keras tensor has an additional dimension and causes wrong results for net.evaluate()","Body":"<p>I'd like to train a neural network in Python and Keras using a metric learning custom loss function. The loss minimizes the distances of the outputs for similar inputs and maximizes the distances between dissimilar ones. The part considering similar inputs is:<\/p>\n\n<pre><code># function to create a pairwise similarity matrix, i.e\n# L[i,j] == 1 for similar samples i, j and 0 otherwise\ndef build_indicator_matrix(y_, thr=0.1):\n    # y_: contains the labels of the samples,\n    # samples are similar in case of same label\n\n    # prevent checking equality of floats --&gt; check if absolute\n    # differences are below threshold\n    lbls_diff = K.expand_dims(y_, axis=0) - K.expand_dims(y_, axis=1)\n    lbls_thr = K.less(K.abs(lbls_diff), thr)\n    # cast bool tensor back to float32\n    L = K.cast(lbls_thr, 'float32')\n\n    # POSSIBLE WORKAROUND\n    #L = K.sum(L, axis=2)\n\n    return L\n\n# function to compute the (squared) Euclidean distances between all pairs\n# of samples, store in DIST[i,j] the distance between output y_pred[i,:] and y_pred[j,:]\ndef compute_pairwise_distances(y_pred):\n    DIFF = K.expand_dims(y_pred, axis=0) - K.expand_dims(y_pred, axis=1)\n    DIST = K.sum(K.square(DIFF), axis=-1)    \n    return DIST\n\n# function to compute the average distance between all similar samples\ndef my_loss(y_true, y_pred):\n    # y_true: contains true labels of the samples\n    # y_pred: contains network outputs\n\n    L = build_indicator_matrix(y_true)    \n    DIST = compute_pairwise_distances(y_pred)\n    return K.mean(DIST * L, axis=1)\n<\/code><\/pre>\n\n<p>For training, I pass a numpy array <code>y<\/code> of shape <code>(n,)<\/code> as target variable to <code>my_loss<\/code>. However, I found (using the computational graph in TensorBoard) that the tensorflow backend creates a 2D variable out of <code>y<\/code> (displayed shape <code>? x ?<\/code>), and hence <code>L<\/code> in <code>build_indicator_matrix<\/code> is not 2 but 3-dimensional (shape <code>? x ? x ?<\/code> in TensorBoard). This causes <code>net.evaulate()<\/code> and <code>net.fit()<\/code> to compute wrong results.<\/p>\n\n<p>Why does tensorflow create a 2D rather than a 1D array? And how does this affect <code>net.evaluate()<\/code> and <code>net.fit()<\/code>? <\/p>\n\n<p>As quick workarounds I found that either replacing the <code>build_indicator_matrix()<\/code> with static numpy code for computing <code>L<\/code> , or collapsing the \"fake\" dimension with the line <code>L = K.sum(L, axis=2)<\/code> solves the problem. In the latter case, however, the output of <code>K.eval(build_indicator_matrix(y))<\/code> is of only of shape <code>(n,)<\/code> and not <code>(n,n)<\/code>, so I do not understand why this workaround still yields correct results. Why does tensorflow introduce an additional dimension?<\/p>\n\n<p>My library versions are:<\/p>\n\n<ul>\n<li>keras: 2.2.4<\/li>\n<li>tensorflow: 1.8.0<\/li>\n<li>numpy: 1.15.0<\/li>\n<\/ul>\n","answers":[{"AnswerId":"57614429","CreationDate":"2019-08-22T17:32:02.140","ParentId":null,"OwnerUserId":"4332585","Title":null,"Body":"<p>This is because <code>evaluate<\/code> and <code>fit<\/code> work in batches.\nThe first dimension you see in tensorboard is the batch dimension, unknown in advance and therefore denoted <code>?<\/code>.\nWhen using custom metrics, remember the tensors (<code>y_true<\/code> and <code>y_pred<\/code>) you get are the ones corresponding to the batch.<\/p>\n\n<p>For more info, show us how you call both those functions.<\/p>\n"}]}
+{"QuestionId":57562490,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-19T18:55:53.840","AcceptedAnswerId":null,"OwnerUserId":9954063.0,"Title":"How to solve \"RuntimeError: The Session graph is empty. Add operations to the graph before calling run()\"","Body":"<p>I am new to tensorflow, I have a class that describes my neural network and all the methods i need for the network. I have a specific method to feed a neuron value. I have declared the neuron initially as a placeholder, hoping to use <code>get_neuron<\/code> method to feed the value to neuron when available. When I am running , I am having a runtime error saying <code>RuntimeError: The Session graph is empty.  Add operations to the graph before calling run()<\/code>. Can anyone help me how to solve it ? and what exactly it means ? I dont understand what exactly it means by add operations as I have been using <code>feed_dict<\/code> operation and running the session. Also does it bring any disadvantages if i run sessions for every computation to my tensor or should the session only be run at the end when having the final output. I am asking this because i am wondering whether this effects the effectiveness of tensorflow as it follows the graph to optimize while using optimizer.<\/p>\n\n<pre><code>def get_neuron(self, a, b):\n        #'while a being the name of the neuron and b being the value of that neuron'\n        with tf.Session() as sess_n:\n            sess_n.run(a, feed_dict={a: b})\n        return\n<\/code><\/pre>\n\n<p>EDIT:\nthis is how I am calling that above function i.e. \n<code>knowledge_out = knowledge.run(carollis_inp)<\/code>\nknowledge is a object created from class knowledge_transfer which has a method <code>run<\/code>, and the very first line of the function is \n<code>self.get_neuron(self.neuron_input, carollis_input)<\/code>\nThe error displayed exactly is <\/p>\n\n<pre><code>[ERROR] [1566241992.292524, 15.300000]: bad callback: &lt;function joint_callback at 0x7f42221982f0&gt;\nTraceback (most recent call last):\n  File \"\/opt\/ros\/melodic\/lib\/python2.7\/dist-packages\/rospy\/topics.py\", line 748, in _invoke_callback\n    cb(msg, cb_args)\n  File \"\/home\/microbot\/catkin_ws\/src\/spider\/spider_control\/control1.py\", line 60, in joint_callback\n    knowledge_out = knowledge.run(carollis_inp)\n  File \"\/home\/microbot\/catkin_ws\/src\/spider\/spider_control\/knowledge_transfer.py\", line 99, in run\n    self.get_neuron(self.neuron_input, carollis_input)\n  File \"\/home\/microbot\/catkin_ws\/src\/spider\/spider_control\/knowledge_transfer.py\", line 81, in get_neuron\n    sess_n.run(a, feed_dict={a: b})\n  File \"\/home\/microbot\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 950, in run\n    run_metadata_ptr)\n  File \"\/home\/microbot\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1098, 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>EDIT2:\nI have a neural network with name knowledge_transfer, it takes 8 inputs and gives four outputs which are then soft maxed , kind of like multi-classification problem. There is this class method <code>run<\/code> which accepts the input array of size 8 and then returns the output array of size four. Inside the <code>run<\/code> method it feeds the given input arguments to input placeholder of a size 8 and then does some processing through the layers and weights and finally gives the output.<br>\nexample:\n I want to create a class member method which takes two inputs and feed_dict first second input argument to the first argument considering the first argument is a placeholder and both the arguments are of same size.<\/p>\n\n<p>Edit3: \nI am feeding an input to the input layer, and then using an input function to calculate the output of input layer, later feeding it to input function of hidden layer and then using the output of input function to hidden layer to pass through leaky_rectified linear output , the output of which is then passed to input function of output layer folowed by passing the output of it through softmax function and giving that softmax output as an output of NN.<\/p>\n\n<p>The code is as follows:<\/p>\n\n<pre><code>        self.neuron_input = tf.compat.v1.placeholder(tf.float32, shape=(self.neurons, 1))\n        self.weight_in = tf.get_variable(name=\"Weight_input\", dtype=tf.float32, shape=[self.neurons, 1], initializer=self.weight_initer)\n        self.neuron_hid = tf.compat.v1.placeholder(tf.float32, shape=(int(self.neurons\/2), 1))\n        self.weight_initer1 = tf.truncated_normal_initializer(mean=1.0, stddev=0.01)\n        self.weight_hid = tf.get_variable(name=\"Weight_hidden\", dtype=tf.float32, shape=[self.neurons, 1], initializer=self.weight_initer1)\n        self.neuron_out = tf.compat.v1.placeholder(tf.float32, shape=(4, 1))\n        self.weight_initer2 = tf.truncated_normal_initializer(mean=2.0, stddev=0.01)\n        self.weight_out = tf.get_variable(name=\"Weight_output\", dtype=tf.float32, shape=[4, 2], initializer=self.weight_initer2)\n        self.bias_initer =tf.truncated_normal_initializer(mean=0-1, stddev=0.01)\n        self.bias_in  =tf.get_variable(name=\"Bias_input\", dtype=tf.float32, shape=[self.neurons, 1], initializer=self.bias_initer)\n        self.bias_initer1 =tf.truncated_normal_initializer(mean=0-2, stddev=0.01)\n        self.bias_hid = tf.get_variable(name=\"Bias_hidden\", dtype=tf.float32, shape=[self.neurons, 1], initializer=self.bias_initer1)\n        self.bias_initer2 =tf.truncated_normal_initializer(mean=0-3, stddev=0.01)\n        self.bias_out = tf.get_variable(name=\"Bias_output\", dtype=tf.float32, shape=[4, 1], initializer=self.bias_initer2)\n<\/code><\/pre>\n\n<p>and then the run() function is as follows:<\/p>\n\n<pre><code> def run(self, carollis_input):\n        self.normalization(carollis_input)\n        #'finding the output of the input layer'\n        knowledge_input = tf.add(tf.multiply(self.neuron_input, self.weight_in), self.bias_in)\n\n            #'calculating the input for the hidden layer'\n        knowledge_hidden = tf.add(tf.multiply(knowledge_input, self.weight_in), self.bias_hid)\n        #'calculating the output of hidden layer'\n        knowledge_hidden_output = 3.14*(tf.add(tf.multiply(knowledge_hidden, self.weight_hid), self.bias_hid))#input function of hidden layer\n        knowledge_hidden_out = tf.nn.leaky_relu(self.neuron_hid, alpha=0.01, name='leaky_relu')\n        with tf.Session() as sess1_2:\n            sess1_2.run(knowledge_hidden_out, feed_dict={self.neuron_input: carollis_input, self.neuron_hid: knowledge_hidden_output})\n        #'calculating the input of output layer'\n        tf.reshape(knowledge_hidden_out, [4, 2])#for quadrant method\n        in_out = tf.add(tf.multiply(knowledge_hidden_out, self.weight_out), self.bias_out)\n        with tf.Session as s:\n            s.run(in_out)\n        #'finding the softmax output of the neurons'\n        softmax_output = np.array(4)\n        softmax_output = self.out_softmax(in_out)  # this gives the softmax output and stores it in the newly created array\n        return softmax_output\n<\/code><\/pre>\n\n<p>The error is as follows :<\/p>\n\n<pre><code>    sess1_2.run(knowledge_hidden_out, feed_dict={self.neuron_input: carollis_input, self.neuron_hid: knowledge_hidden_output})\n  File \"\/home\/microbot\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 950, in run\n    run_metadata_ptr)\n  File \"\/home\/microbot\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1098, 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","answers":[]}
+{"QuestionId":57562496,"AnswerCount":1,"Tags":"<python><tensorflow><keras><anaconda><conda>","CreationDate":"2019-08-19T18:56:48.290","AcceptedAnswerId":57562573.0,"OwnerUserId":7932866.0,"Title":"TypeError: 'module' object is not callable. keras","Body":"<p><strong>System information<\/strong><br>\n- Windows 10<br>\n- TensorFlow backend (yes \/ no):  yes<br>\n- TensorFlow version:  1.14.0<br>\n- Keras version:  2.24<br>\n- Python version:  3.6<br>\n- CUDA\/cuDNN version:  10<br>\n- GPU model and memory:  gtx 1050 ti  <\/p>\n\n<p><strong>Describe the current behavior<\/strong><br>\nI installed tensoflow and keras via conda. Then i tried to run this code:<\/p>\n\n<pre><code>import tensorflow as tf\nimport keras\nimport numpy as np\n\nmodel = keras.Sequential([keras.layers(units=1, input_shape=[1])])\n\nmodel.compile(optimizer=\"sgd\", loss=\"mean_squared_error\")\n\nx = np.array([-1, 0, 1, 2, 3, 4])\ny = np.array([-3, -1, 1, 3, 5, 7])\n\nmodel.fit(x, y, epochs=500)\n\nprint(model.predict([10]))`\n<\/code><\/pre>\n\n<p>When i run this code I get the error:  <\/p>\n\n<pre><code>Using TensorFlow backend.\nTraceback (most recent call last):\n  File \"C:\/Users\/xxx\/PycharmProjects\/Workspace\/tensorflow\/hello_world_of_nn.py\", line 5, in &lt;module&gt;\n    model = keras.Sequential([keras.layers(units=1, input_shape=[1])])\nTypeError: 'module' object is not callable\n<\/code><\/pre>\n\n<p>When i try this:<br>\n<code>python -c 'import keras as k; print(k.__version__)'<\/code><\/p>\n\n<p>I get the error:  <\/p>\n\n<pre><code>C:\\Users\\xxx&gt;python -c 'import keras as k; print(k.__version__)'\n  File \"&lt;string&gt;\", line 1\n    'import\n          ^\nSyntaxError: EOL while scanning string literal\n<\/code><\/pre>\n","answers":[{"AnswerId":"57562573","CreationDate":"2019-08-19T19:03:54.227","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>This should be fine:<\/p>\n\n<pre><code>import tensorflow as tf\nimport keras\nimport numpy as np\n\nmodel = keras.models.Sequential([keras.layers.Dense(units=1, input_shape=[1])])\n\nmodel.compile(optimizer=\"sgd\", loss=\"mean_squared_error\")\n\nx = np.array([-1, 0, 1, 2, 3, 4])\ny = np.array([-3, -1, 1, 3, 5, 7])\n\nmodel.fit(x, y, epochs=500)\n\nprint(model.predict([10]))\n<\/code><\/pre>\n\n<p>Please note the usage of <code>keras.models.Sequential<\/code> and <code>keras.layers.Dense<\/code>.<\/p>\n"}]}
+{"QuestionId":57562684,"AnswerCount":1,"Tags":"<python><tensorflow><keras><ensemble-learning>","CreationDate":"2019-08-19T19:15:01.923","AcceptedAnswerId":null,"OwnerUserId":3050590.0,"Title":"How can I fit binary keras models into multiclass models?","Body":"<p>I am working on <a href=\"https:\/\/en.wikipedia.org\/wiki\/Multiclass_classification\" rel=\"nofollow noreferrer\">OVA<\/a> (one vs all) classification problem. For that, I have trained Keras binary classifiers with sigmoid function and binary_crossentropy. I need to ensemble them to multiclass model similar to <a href=\"https:\/\/www.kaggle.com\/mgiraygokirmak\/keras-cnn-multi-model-ensemble-with-voting\" rel=\"nofollow noreferrer\">here<\/a> When I am trying to do that I get the following error <\/p>\n\n<pre><code>ValueError: A target array with shape (32, 3) was passed for an output of shape (None, 2) while using as loss `binary_crossentropy`.\n<\/code><\/pre>\n\n<p>Program code<\/p>\n\n<pre><code>for i in os.listdir(model_root): \/\/loading all the models\nprint(i)\nfilename = model_root + \"\/\" + i\n# load model\nmodel = load_model(filename, custom_objects={'KerasLayer': hub.KerasLayer})\nmodels.append(model)\nprint(len(models))  \/\/3\n\n  #To fit the loaded models to the data and saving it to an array fit_models\n\nsteps_per_epoch = image_data.samples \/\/ image_data.batch_size\nbatch_stats = CollectBatchStats()\nvalidation_steps = image_data_val.samples \/ image_data_val.batch_size\nfor i in range(len(models)):\n    model[i].fit_generator((item for item in image_data), epochs=2,\n              steps_per_epoch=steps_per_epoch, #callbacks=[batch_stats],\n              validation_data=(item for item in image_data_val), validation_steps=validation_steps, verbose=2)\n    fit_models.append(model[i])\n<\/code><\/pre>\n\n<p>I am getting this error on the fit function and the problem is models are trained on 2 classes whereas, I need to fit it now to 3 classes. <\/p>\n\n<pre><code>model.shape() # (32, 2)\n<\/code><\/pre>\n\n<p>My data looks like this<\/p>\n\n<pre><code>Image batch shape:  (32, 224, 224, 3)\nLabel batch shape:  (32, 3) # 3 classes\n<\/code><\/pre>\n\n<p>This creates a conflict between 2 and 3 classes. I don't know how to deal with this problem and I don't know if it is possible in keras<\/p>\n\n<p>After the provided solution, now my model looks like<\/p>\n\n<pre><code>        Model: \"model\"\n__________________________________________________________________________________________________\nLayer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput_1 (InputLayer)            [(None, 224, 224, 3) 0                                            \n__________________________________________________________________________________________________\nsequential_4 (Sequential)       (None, 1)            3541267     input_1[0][0]                    \n__________________________________________________________________________________________________\nsequential_8 (Sequential)       (None, 1)            3541267     input_1[0][0]                    \n__________________________________________________________________________________________________\nsequential_2 (Sequential)       (None, 1)            3541267     input_1[0][0]                    \n__________________________________________________________________________________________________\nconcatenate (Concatenate)       (None, 3)            0           sequential_4[1][0]               \n                                                                 sequential_8[1][0]               \n                                                                 sequential_2[1][0]               \n==================================================================================================\nTotal params: 10,623,801\nTrainable params: 3,006\nNon-trainable params: 10,620,795\n__________________________________________________________________________________________________\n<\/code><\/pre>\n\n<p>Error is <\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"&lt;input&gt;\", line 1, in &lt;module&gt;\n  File \"C:\\Program Files\\JetBrains\\PyCharm 2019.2\\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 \"C:\\Program Files\\JetBrains\\PyCharm 2019.2\\helpers\\pydev\\_pydev_imps\\_pydev_execfile.py\", line 18, in execfile\n    exec(compile(contents+\"\\n\", file, 'exec'), glob, loc)\n  File \"C:\/Users\/Pawandeep\/Desktop\/Python projects\/ensemble_image.py\", line 80, in &lt;module&gt;\n    validation_data=(item for item in image_data_val), validation_steps=validation_steps, verbose=2)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 673, in fit\n    initial_epoch=initial_epoch)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 1433, in fit_generator\n    steps_name='steps_per_epoch')\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_generator.py\", line 264, in model_iteration\n    batch_outs = batch_function(*batch_data)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 1175, in train_on_batch\n    outputs = self.train_function(ins)  # pylint: disable=not-callable\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py\", line 3292, in __call__\n    run_metadata=self.run_metadata)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1458, in __call__\n    run_metadata_ptr)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'sequential_input' with dtype float and shape [?,224,224,3]\n     [[{{node sequential_input}}]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"57563028","CreationDate":"2019-08-19T19:45:01.120","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>You need models of 1 class (you can't ensemble models of 2 classes and reach 3)   <\/p>\n\n<p>If you trained models of 1 class in the list <code>models<\/code>, you need to transform them into a single model:<\/p>\n\n<pre><code>inputs = Input(common_input_shape)\noutputs = [m(inputs) for m in models]\noutputs = Concatenate()(outputs)\n#maybe outputs = Activation(\"softmax\")(outputs) if the problem is categorical    \nensemble = Model(inputs, outputs)\n<\/code><\/pre>\n\n<p>Fit the <code>ensemble<\/code> model. <\/p>\n"}]}
+{"QuestionId":57562720,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow><python-3.6>","CreationDate":"2019-08-19T19:17:05.117","AcceptedAnswerId":57562774.0,"OwnerUserId":1877002.0,"Title":"Upgrading from tensorflow 1.x to 2.0","Body":"<p>I am new to tensorflow.\nHave tried this simple example:<\/p>\n\n<pre><code>import tensorflow as tf\nsess = tf.Session()\nx = tf.placeholder(tf.float32)\ny = tf.placeholder(tf.float32)\nz = x + y\nprint(sess.run(z, feed_dict={x: 3.0, y: 4.5}))\n<\/code><\/pre>\n\n<p>and got some warnings <code>The name tf.Session is deprecated. Please use tf.compat.v1.Session instead.<\/code> and the right answer - <code>7.5<\/code><\/p>\n\n<p>After reading <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/migration_guide#1_replace_v1sessionrun_calls\" rel=\"nofollow noreferrer\">here<\/a>, I understand that the warnings are due to upgrading from tf 1.x to 2.0, the steps described are \"simple\" but they don't give any example....<\/p>\n\n<p>I have tried:<\/p>\n\n<pre><code>@tf.function\ndef f1(x1, y1):\n    return tf.math.add(x1, y1)\n\n\nprint(f1(tf.constant(3.0), tf.constant(4.5)))\n<\/code><\/pre>\n\n<ol>\n<li>Is my code correct (in the sense defined in the link)?<\/li>\n<li><strong>Now, I am getting <code>Tensor(\"PartitionedCall:0\", shape=(), dtype=float32)<\/code> as output, how can I get the actual value?<\/strong><\/li>\n<\/ol>\n","answers":[{"AnswerId":"57562785","CreationDate":"2019-08-19T19:23:02.393","ParentId":null,"OwnerUserId":"3197668","Title":null,"Body":"<p>Your code is correct as per Tensorflow 2.0. Tensorflow 2.0 is even more tightly bonded with numpy so if you want to get the result of the operation, you could use the <code>numpy()<\/code> method:<\/p>\n\n<pre><code>print(f1(tf.constant(3.0), tf.constant(4.5)).numpy())\n<\/code><\/pre>\n"},{"AnswerId":"57562774","CreationDate":"2019-08-19T19:22:38.240","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>You code is indeed correct. The warning that you get indicates that as from Tensorflow 2.0, <code>tf.Session()<\/code> won't exist in the API. Therefore, if you want you code to be compatible with Tensorflow 2.0, you should use <code>tf.compat.v1.Session<\/code> instead. So, just change this line:<\/p>\n\n<pre><code>sess = tf.Session()\n<\/code><\/pre>\n\n<p>To:<\/p>\n\n<pre><code>sess = tf.compat.v1.Session()\n<\/code><\/pre>\n\n<p>Then, even if you update Tensorflow from 1.xx to 2.xx, your code would execute in the same way. As for the code in Tensorflow 2.0:<\/p>\n\n<pre><code>@tf.function\ndef f1(x1, y1):\n    return tf.math.add(x1, y1)\n\nprint(f1(tf.constant(3.0), tf.constant(4.5)))\n<\/code><\/pre>\n\n<p>it is fine if you run it in Tensorflow 2.0. If you want to run the same code, without installing Tensorflow 2.0, you can do the following:<\/p>\n\n<pre><code>import tensorflow as tf\ntf.enable_eager_execution()\n\n@tf.function\ndef f1(x1, y1):\n    return tf.math.add(x1, y1)\n\nprint(f1(tf.constant(3.0), tf.constant(4.5)).numpy())\n<\/code><\/pre>\n\n<p>The reason for this is because the default way of executing Tensorflow operations starting from Tensorflow 2.0 is in eager mode. The way of activating eager mode in Tensorflow 1.xx, is to enable it right after the import of Tensorflow, as I am doing it in the example above.<\/p>\n"}]}
+{"QuestionId":57562900,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-08-19T19:33:28.477","AcceptedAnswerId":null,"OwnerUserId":11948285.0,"Title":"Iterating over tensors in Tensorflow \/ Keras","Body":"<p>Please help me to implement an equivalent code in Tensorflow \/ Keras. I think I still don't get how it works...<\/p>\n\n<pre><code>A = np.zeros((N, N))\nfor i, j, k in zip(input_sequence, input_sequence[1:], output_sequence[1:]):\n    A[i, j] += k\n<\/code><\/pre>\n\n<p>input_sequence, output_sequence are tensors with shape = (?, 100, 1)<\/p>\n\n<p>Basically the idea is to set the values of A using the indexes stored in input_sequence and values stored in output_sequence.<\/p>\n\n<p>For example:<\/p>\n\n<p>If input_sequence = [0, 1, 2, 1, 2] and output_sequence = [ 1.1, 2.3, 3.3, 4.5, 10.1 ], the resulting matrix is:<\/p>\n\n<pre><code>[ [0.0, 2.3, 0.0],\n  [0.0, 0.0, 13.4],\n  [0.0, 4.5, 0.0] ]\n<\/code><\/pre>\n\n<p>Please help!<\/p>\n","answers":[]}
+{"QuestionId":57562957,"AnswerCount":1,"Tags":"<python><tensorflow><edward>","CreationDate":"2019-08-19T19:39:02.293","AcceptedAnswerId":null,"OwnerUserId":11924261.0,"Title":"Can not use Edward 1.3.5 with Tensorflow 1.14.0","Body":"<p>I am trying to use Edward library for deploying a PPCA model and facing difficulties while installing it on windows. I successfully installed Edward 1.3.5 using  (pip install edward ) and tensorflow 1.14.0 using (pip install tensorflow) but when I try to import edward in jupyter notebook , I get the following error <\/p>\n\n<p>cannot import name 'set_shapes_for_outputs' from 'tensorflow.python.framework.ops<\/p>\n\n<p>I tried looking into issues opened here (<a href=\"https:\/\/github.com\/blei-lab\/edward\/issues\/882\" rel=\"nofollow noreferrer\">https:\/\/github.com\/blei-lab\/edward\/issues\/882<\/a>) and here (<a href=\"https:\/\/github.com\/blei-lab\/edward\/issues\/893\" rel=\"nofollow noreferrer\">https:\/\/github.com\/blei-lab\/edward\/issues\/893<\/a>)  but nothing worked.<\/p>\n\n<p>If anyone has any recommendations and\/or suggestions, that would be great!<\/p>\n\n<p>Many thanks!<\/p>\n","answers":[{"AnswerId":"57564913","CreationDate":"2019-08-19T23:11:15.010","ParentId":null,"OwnerUserId":"11924261","Title":null,"Body":"<p>All I did to solve this issue was, I created a new conda environment, downgraded python to 3.6.5 (anything older than 3.7.0 works fine) and downgraded TensorFlow to version 1.3.0 (anything older than 1.6.0 works fine).<\/p>\n"}]}
+{"QuestionId":57563007,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-19T19:42:55.183","AcceptedAnswerId":null,"OwnerUserId":11948297.0,"Title":"Use a Numpy subroutine in Tensorflow","Body":"<p>I have a numpy subroutine that uses the Jacobi function from the scipy.special to return the value of a Jacobi polynomials given the order,n, parameters a,b, and point x. I want to introduce the parameters a,b as tf.variables and incorporate my subroutine in tensorflow. I appreciate any help with this.<\/p>\n\n<pre><code>from scipy.special import jacobi\n\ndef Jacobi(n,a,b,x):\n\n   x=np.array(x)\n\n   return (jacobi(n,a,b)(x))\n<\/code><\/pre>\n\n<p>variables in Tensorflow<\/p>\n\n<pre><code> a = tf.Variable(0.5*tf.ones([1], dtype=tf.float64), dtype=tf.float64)\n\n b = tf.Variable(0.5*tf.ones([1], dtype=tf.float64), dtype=tf.float64)\n<\/code><\/pre>\n\n<p>and I want<\/p>\n\n<pre><code> x = np.array([-1, -0.5, 0, 0.5, 1])\n\n y = Jacobi(n=5, a, b, x)\n<\/code><\/pre>\n\n<p>so that I can use tf.gradients of y with respect to a and b.<\/p>\n","answers":[]}
+{"QuestionId":57563334,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-19T20:10:05.023","AcceptedAnswerId":null,"OwnerUserId":11948397.0,"Title":"How to check which nodes have been converted to float16 in AMP in tensorflow?","Body":"<p>Is there a way to figure out which nodes in the tensorflow graph are being executed in float16 in the automatic mixed precision (AMP). I am using the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/experimental\/enable_mixed_precision_graph_rewrite\" rel=\"nofollow noreferrer\">tf.train.experimental.enable_mixed_precision_graph_rewrite()<\/a>.<\/p>\n\n<p>I am aware of the algorithm that tensorflow uses as explained <a href=\"https:\/\/devblogs.nvidia.com\/video-mixed-precision-techniques-tensor-cores-deep-learning\/#part3\" rel=\"nofollow noreferrer\">here<\/a>,  however in a complex graph its not really possible to manually keep track of everything.<\/p>\n","answers":[]}
+{"QuestionId":57563619,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-08-19T20:35:33.333","AcceptedAnswerId":null,"OwnerUserId":11758805.0,"Title":"Exporting a Trained Inference Graph TENSERFLOW","Body":"<p>I have trained my custom model and want to export a trained inference graph<\/p>\n\n<p>I ran the following command<\/p>\n\n<pre><code>INPUT_TYPE=image_tensor\nPIPELINE_CONFIG_PATH= training\/ ssd_mobilenet_v1_pets.config\nTRAINED_CKPT_PREFIX= training\/model.ckpt-2509\nEXPORT_DIR= training\/new_model\npython exporter.py \\\n    --input_type=${INPUT_TYPE} \\\n    --pipeline_config_path=${PIPELINE_CONFIG_PATH} \\\n    --trained_checkpoint_prefix=${TRAINED_CKPT_PREFIX} \\\n    --output_directory=${EXPORT_DIR}\n<\/code><\/pre>\n\n<p>And I got the following output<\/p>\n\n<pre><code>W0819 22:08:54.649750  2680 deprecation_wrapper.py:119] From C:\\Users\\Aleksej\\Anaconda3\\envs\\cocosynth4\\lib\\site-packages\\object_detection-0.1-py3.6.egg\\nets\\mobilenet\\mobilenet.py:397: The name tf.nn.avg_pool is deprecated. Please use tf.nn.avg_pool2d instead.\n\n\n(cocosynth4) D:\\yolo\\models\\research\\object_detection&gt;    --input_type=${INPUT_TYPE} \\\n'--input_type' is not recognized as an internal or external command,\noperable program or batch file.\n\n(cocosynth4) D:\\yolo\\models\\research\\object_detection&gt;    --pipeline_config_path=${PIPELINE_CONFIG_PATH} \\\n'--pipeline_config_path' is not recognized as an internal or external command,\noperable program or batch file.\n\n(cocosynth4) D:\\yolo\\models\\research\\object_detection&gt;    --trained_checkpoint_prefix=${TRAINED_CKPT_PREFIX} \\\n'--trained_checkpoint_prefix' is not recognized as an internal or external command,\noperable program or batch file.\n\n(cocosynth4) D:\\yolo\\models\\research\\object_detection&gt;    --output_directory=${EXPORT_DIR}\n'--output_directory' is not recognized as an internal or external command,\noperable program or batch file.\n<\/code><\/pre>\n\n<p>I am running a windows 10 and python3.\nDoes anyone have any suggestions on how to solve this issue<\/p>\n","answers":[{"AnswerId":"57569283","CreationDate":"2019-08-20T08:10:48.140","ParentId":null,"OwnerUserId":"11758805","Title":null,"Body":"<p>Fixed with this code<\/p>\n\n<p>python export_inference_graph.py --input_type image_tensor --pipeline_config_path training\/ssd_mobilenet_v1_pets.config --trained_checkpoint_prefix training\/model.ckpt-3970 --output_directory ships_inference_graph<\/p>\n"}]}
+{"QuestionId":57563763,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2019-08-19T20:48:57.640","AcceptedAnswerId":null,"OwnerUserId":915783.0,"Title":"how to print each of the batch file being trained from imageDataGenerator in keras","Body":"<p>I am using Keras ImageDataGenerator with <em>flow_from_directory<\/em>.<\/p>\n\n<p>For  training data, each class folder has 10,000-20,000 jpg files each, with 13 classes. While training, although keras shows the epoch , I want to print which of the image files are being trained\/used in each each batch. How do I do that?    <\/p>\n\n<p>Thanks<br>\nsedy<\/p>\n","answers":[{"AnswerId":"57563866","CreationDate":"2019-08-19T20:56:23.420","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>You will probably have a flooding problem, but you can create a wrapping generator:<\/p>\n\n<pre><code>from keras.utils import Sequence\n\nclass PrintingGenerator(Sequence):\n\n    def __init__(self, keras_generator):\n        self.keras_generator = keras_generator\n\n    def __len__(self):\n        return len(self.keras_generator)\n\n    def __getitem__(self,i):\n        x, y = self.keras_generator[i]\n        #do the print\n        return x, y \n\n    def on_epoch_end(self):\n        self.keras_generator.on_epoch_end()\n\ngenerator = PrintingGenerator(original.flow_from_directory(...))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57563921,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-08-19T21:02:07.670","AcceptedAnswerId":null,"OwnerUserId":915783.0,"Title":"How to recognize corrupt image files in folder while training using Keras","Body":"<p>I am using Keras ImageDataGenerator with flow_from_directory.<\/p>\n\n<p>For training data, each class folder has 10,000-20,000 jpg files each, with 13 classes. While training, every hour or so, before full data has been trained once, I get error from Keras indicating that a particular jpg file could not be read. Sure enough, I explore into the file location, and can see file is corrupt.\nproblem is this wastes lot of time, and on each error, I have to restart the training.<\/p>\n\n<ol>\n<li>is there a way to modify Keras flow_from_directory or    to ensure that corrupt files are skipped and training goes on without error<\/li>\n<li>if 1 is not possible, is there a way to quickly(I have 1.1 TB of data) verify corrput files in all class folders and remove them \/delete them from those folders, before starting the training.<\/li>\n<\/ol>\n","answers":[]}
+{"QuestionId":57564584,"AnswerCount":1,"Tags":"<python><unit-testing><tensorflow>","CreationDate":"2019-08-19T22:20:34.620","AcceptedAnswerId":57564698.0,"OwnerUserId":1306819.0,"Title":"Tensorflow DNNClassifier unittest features","Body":"<p>I am creating a <code>DNNClassifier<\/code> in a method that returns the classifier. I want to write a unit test to verify that the list of features passed is the one that is expected. However, since the features are not an attribute of <code>DNNClassifier<\/code>, I don't know how to get the list of features from the classifier.<\/p>\n\n<p>Some code to illustrate the point.<\/p>\n\n<pre><code>def create_dnn_classifier(features, config, directory):\n    # some manipulation on features\n    classifier = DNNClassifier(\n        feature_columns=features, # list of tf.feature_column.numeric_column\n        hidden_units=[512, 256, 128],\n        activation_fn=tf.nn.leaky_relu,\n        dropout=0.1,\n        n_classes=2,\n        config=config,\n        model_dir=directory)\n    return classifier\n<\/code><\/pre>\n\n<p>Unittest look like this.<\/p>\n\n<pre><code>class TestMyDNNClassifierFeatures(unittest.TestCase):\n    def test_classifier_features(self):\n        classifier = create_dnn_classifier(\n            ['feature1', 'feature2', 'feature3'],\n            RunConfig(\n                save_checkpoints_steps=100,\n                keep_checkpoint_max=100,\n                keep_checkpoint_every_n_hours=None,\n                model_dir='test_dir'\n            ),\n            'test_dir'\n        )\n        # how do i get the list of features from classifer now?\n        print classifier.features # won't work as classifier has no attribute named features\n        print dir(classifier) # prints a lot of things but none resembling features\n<\/code><\/pre>\n","answers":[{"AnswerId":"57564698","CreationDate":"2019-08-19T22:38:27.433","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>The <code>tf.estimator.DNNClassifier<\/code> object doesn't store the features names inside the object state. Therefore, you can't retrieve them. However, you can make a convenience wrapper like:<\/p>\n\n<pre><code>class DNNWrapper:\n\n    def __init__(self, features, config, directory):\n        self.features = features\n        self.classifier = tf.estimator.DNNClassifier(\n            feature_columns=features,\n            hidden_units=[512, 256, 128],\n            activation_fn=tf.nn.leaky_relu,\n            dropout=0.1,\n            n_classes=2,\n            config=config,\n            model_dir=directory)\n<\/code><\/pre>\n\n<p>And then, instead of testing the <code>tf.estimator.DNNClassifier<\/code>, just test your wrapper <code>DNNWrapper<\/code>:<\/p>\n\n<pre><code>class TestMyDNNClassifierFeatures(unittest.TestCase):\n\n    def test_classifier_features(self):\n        classifier = DNNWrapper(\n            ['feature1', 'feature2', 'feature3'],\n            tf.Estimator.RunConfig(\n                save_checkpoints_steps=100,\n                keep_checkpoint_max=100,\n                keep_checkpoint_every_n_hours=None,\n                model_dir='test_dir'\n            ),\n            'test_dir'\n        )\n        print(classifier.features)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57564951,"AnswerCount":0,"Tags":"<java><android><tensorflow><audio>","CreationDate":"2019-08-19T23:17:19.203","AcceptedAnswerId":null,"OwnerUserId":10360610.0,"Title":"Tensorflow Android Audio Recognition demo","Body":"<p>I'm working on this Tensorflow <a href=\"https:\/\/www.tensorflow.org\/tutorials\/sequences\/audio_recognition\" rel=\"nofollow noreferrer\">demo<\/a> with my Android phone. The whole thing is working smoothly without any problem.<br>\nMy question is about the post-processing part, namely, the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/android\/src\/org\/tensorflow\/demo\/RecognizeCommands.java\" rel=\"nofollow noreferrer\">RecognizeCommands<\/a>. I found the code hard to understand. (The main activity of this Android app is called <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/android\/src\/org\/tensorflow\/demo\/SpeechActivity.java\" rel=\"nofollow noreferrer\">SpeechActivity<\/a>)<br>\nIn detail, what is \"averageWindowDurationMs\" doing? My understanding is that within this time window, we collect several amount of predictions. \nHowever, in the SpeechActivity, sample_duration_ms = 1000, this means that each of my audio input should be 1sec. And AVERAGE_WINDOW_DURATION_MS = 500, this means that my time window is only 500ms. This is already confusing me...How could my time window size smaller than the time length of a single audio input? <br>\nPlease correct my understanding. Really appreciate it if someone could walk me through on how does this \"RecognizeCommands\" do post-processing?<\/p>\n","answers":[]}
+{"QuestionId":57564960,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><neural-network><deep-learning><backpropagation>","CreationDate":"2019-08-19T23:18:34.593","AcceptedAnswerId":null,"OwnerUserId":5432071.0,"Title":"What are the correct backpropogation steps for a three layer neural network?","Body":"<p>I am doing research into using surrogate optimization for neural networks.  With this in mind I need kind of bare bones control over the optimization process of the network.  I have posted code which uses tensorflow-eager to implement plain stochastic gradient descent with a batch size of one for MNIST.<\/p>\n\n<p>I have two questions:<\/p>\n\n<ol>\n<li><p>Are my calculations for the steps of backpropogation through the neural network correct? If not... where are the mistakes and why are they mistakes?<\/p><\/li>\n<li><p>Can you attain this level of control over the optimization process using autodiff or gradient tape in tensorflow? If so...how?  I want to be able to some epochs apply the standard backprop gradient, and other epochs apply a nonstandard gradient update, but this non-standard gradient update would have the same shape as the true gradient update.<\/p><\/li>\n<\/ol>\n\n<p>Maybe this code can maybe be a poorly documented backprop reference for others in the future.<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.examples.tutorials.mnist import input_data # just use tensorflow's mnist api\n\nimport numpy as np\n\n# Good back prop references\n#https:\/\/www.freecodecamp.org\/news\/building-a-3-layer-neural-network-from-scratch-99239c4af5d3\/\n\n#https:\/\/towardsdatascience.com\/back-propagation-the-easy-way-part-1-6a8cde653f65\n\n\n#logits are what comes before softmax\ntf.enable_eager_execution()\n\n\nmnist = input_data.read_data_sets(\"MNIST_data\/\", one_hot=True)\n\n\n\ndef relu_deriv_single(activated,s1):\n\n    activated = np.reshape(activated, (s1))\n    retVec = np.zeros(s1)\n    for i in range(s1):\n\n        if activated[i] &gt; 0:\n                retVec[i] = 1\n\n        elif activated[i] &lt; 0:\n            retVec[i] = 0\n\n    return retVec\n\ndef softmax_deriv(soft_proba, true_label):\n    retVec = np.zeros(10)\n    for i in range(10):\n        #print(retVec)\n        retVec[i] = soft_proba[0,i]-true_label[0,i]\n    return retVec\n\n\n\n\n#################################################\n#Train a neural network on mnist using tf assign and doing\n#backprop manually\n\n\nn_inputs = 28*28\n\nn_hidden1 = 300\n\nn_hidden2 = 100\n\nn_outputs = 10\n\n##Placeholder nodes for data\n\n\n\n###Model variables\n\n#n_inputs = int(X.get_shape()[1])\nstddev = 2\/np.sqrt(n_inputs)\ninit = tf.truncated_normal((n_inputs,n_hidden1), stddev=stddev)\n\n##layer 1\nW1 = tf.Variable(init, name=\"InputWeight\")\n#W1 = tf.Variable(tf.random_uniform([n_inputs,n_hidden1]), name=\"Input\")\nb1 = tf.Variable(tf.zeros([n_hidden1]))\n\n##layer 2\nn_inputs = n_hidden1\nstddev = 2\/np.sqrt(n_inputs)\ninit2 = tf.truncated_normal((n_hidden1,n_hidden2),stddev=stddev)\nW2 = tf.Variable(init2, name=\"HiddenWeight\")\n#W2 = tf.Variable(tf.random_uniform([n_hidden1,n_hidden2]), name=\"HiddenWeight\")\nb2 = tf.Variable(tf.zeros([n_hidden2]))\n\n###logits output layer\nn_inputs = n_hidden2\nstddev = 2\/np.sqrt(n_inputs)\ninit3 = tf.truncated_normal((n_hidden2,n_outputs),stddev=stddev)\nW3 = tf.Variable(init3, name=\"OutputWeight\")\n#W3 = tf.Variable(tf.random_uniform([n_hidden2,n_outputs]), name=\"OutputWeight\")\nb3 = tf.Variable(tf.zeros([n_outputs]))\n\n\nn_epochs = 5\nbatch_size = 1\nlearning_rate = 0.000001\n\nfor epoch in range(n_epochs):\n        for iteration in range(mnist.train.num_examples \/\/ batch_size):\n        #for iteration in range(1000 \/\/ batch_size):\n            X_b, y_b = mnist.train.next_batch(batch_size)\n\n            z1 = tf.matmul(X_b, W1) + b1\n\n            a1 = tf.nn.relu(z1)\n\n            z2 = tf.matmul(a1, W2) + b2\n\n            a2 = tf.nn.relu(z2)\n\n            z3 = tf.matmul(a2, W3) + b3\n\n\n            a3 = tf.nn.softmax(z3)\n\n            #####End of forward pass\n\n            delta_dw3 = (a3 - y_b) * softmax_deriv(z3,y_b)\n\n\n            #Gradient for W3\n            dL_dW3 = (1\/batch_size)*tf.transpose(a2)*delta_dw3\n\n\n            tf.assign(W3, W3 - learning_rate*dL_dW3)\n\n            #gradient for b3\n            db3 = (1\/batch_size)*np.reshape(delta_dw3, (10))\n\n            tf.assign(b3, b3 - learning_rate*db3)\n\n            delta_dw2_part1 = np.dot(delta_dw3, tf.transpose(W3))\n\n            delta_dw2_part2 = relu_deriv_single(z2, n_hidden2)\n\n            delta_dw2 = np.multiply(delta_dw2_part1,delta_dw2_part2)\n\n            #Gradient for W2\n            dL_dW2 = (1\/batch_size)*tf.transpose(a1)*delta_dw2\n\n            tf.assign(W2, W2-learning_rate*dL_dW2)\n\n            #Gradient for b2\n            db2 = (1\/batch_size)*np.reshape(delta_dw2, (100))\n\n            tf.assign(b2, b2 - learning_rate*db2)\n\n            delta_dw1_part1 = np.dot(delta_dw2, tf.transpose(W2))\n\n            delta_dw1_part2 = relu_deriv_single(z1, n_hidden1)\n\n            delta_dw1 = np.multiply(delta_dw1_part1, delta_dw1_part2)\n\n            #Gradient for W1\n            dL_dW1 = (1\/batch_size)*tf.transpose(X_b)*delta_dw1\n\n            tf.assign(W1, W1-learning_rate*dL_dW1)\n\n            #Gradient for b1\n            db1 = (1\/batch_size)*np.reshape(delta_dw1, (300))\n\n            tf.assign(b1, b1-learning_rate*db1)\n\n\n            loss = tf.losses.softmax_cross_entropy(y_b,z3)\n\n            print(loss)\n\n\n\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57565019,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow2.0><eager>","CreationDate":"2019-08-19T23:27:00.720","AcceptedAnswerId":57567989.0,"OwnerUserId":4744283.0,"Title":"Train complicated nn models with tf.eager (better with TF2 symbolic support)","Body":"<p>Is there (more or less) simple way to write a complicated NN model so it will be trainable in the eager mode? Are there examples of a such code?<\/p>\n\n<p>For example, I want to use the <code>InceptionResnetV2<\/code>. I have the code created with <code>tf.contrib.slim<\/code>. According to this link, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/16182\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/16182<\/a> , slim is deprecated and I need to use <code>Keras<\/code>. And I really can't use a slim code for training with eager because I can't take the list of variables and apply gradients (ok, I can try to wrap model into the <code>GradientTape<\/code> but not sure what to do with regularization loss).<\/p>\n\n<p>Ok, let's try <code>Keras<\/code>.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>In [30]: tf.__version__                                                                                                                                                                          \nOut[30]: '1.13.1'\n\nIn [31]: tf.enable_eager_execution()\n\nIn [32]: from keras.applications.inception_resnet_v2 import InceptionResNetV2\n\nIn [33]: model = InceptionResNetV2(weights=None)\n...\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras_applications\/inception_resnet_v2.py in InceptionResNetV2(include_top, weights, input_tensor, input_shape, pooling, classes, **kwargs)\n    246 \n    247     if input_tensor is None:\n--&gt; 248         img_input = layers.Input(shape=input_shape)\n    249     else:\n    250         if not backend.is_keras_tensor(input_tensor):\n...\nRuntimeError: tf.placeholder() is not compatible with eager execution.\n<\/code><\/pre>\n\n<p>Doesn't work by default.<\/p>\n\n<p>In this tutorial they say that I need to make my own class of a model and maintain variables by myself <a href=\"https:\/\/www.tensorflow.org\/tutorials\/eager\/custom_training#define_the_model\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/eager\/custom_training#define_the_model<\/a>. I'm not sure that I want to do it for Inception. Too many variables to create and maintain. It's like returning back to old versions of <code>TF<\/code>, in days when even slim didn't exist.<\/p>\n\n<p>In this tutorial networks are created using <code>Keras<\/code> <a href=\"https:\/\/www.tensorflow.org\/tutorials\/eager\/custom_training_walkthrough#create_a_model_using_keras\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/eager\/custom_training_walkthrough#create_a_model_using_keras<\/a> but I doubt that I can easily maintain complicated structure in a such way, by only defining model without using it with <code>Input<\/code>. For example, in this article, if I understand correctly, author initialize keras <code>Input<\/code> and propagate it through the model (which causes <code>RuntimeError<\/code> when used with Eager, as you seen earlier). I can make my own model by subclassing the model class as here: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/Model\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/Model<\/a> . Oops, in this way I need to maintain layers, not variables. It seems as almost the same problem to me. <\/p>\n\n<p>There is an interesting mention of <code>AutoGrad<\/code> here <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/autograph#keras_and_autograph\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/guide\/autograph#keras_and_autograph<\/a> . They only overwrite <code>__call__<\/code>, so it seems like I don't need to maintain variables in this case, but I didn't test it yet. <\/p>\n\n<hr>\n\n<p>So, is there any simple solution? <\/p>\n\n<p>Wrap slim model in <code>GradientTape<\/code>? How can I then apply reg loss to the weights? <\/p>\n\n<p>Track every variable by myself? Sounds a little bit painful. <\/p>\n\n<p>Use <code>Keras<\/code>? How to use it with eager when I have branches and complicated structure in the model?<\/p>\n","answers":[{"AnswerId":"57567989","CreationDate":"2019-08-20T06:43:03.063","ParentId":null,"OwnerUserId":"2455494","Title":null,"Body":"<p>Your first approach is probably the most common. This error:<\/p>\n\n<blockquote>\n  <p>RuntimeError: tf.placeholder() is not compatible with eager execution.<\/p>\n<\/blockquote>\n\n<p>is because one cannot use a <code>tf.placeholder<\/code> in eager mode. There is no concept of such a thing when executing eagerly. <\/p>\n\n<p>You could use the <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets\" rel=\"nofollow noreferrer\"><code>tf.data<\/code> API<\/a> to build a dataset for your training data and feed that to the model. Something like this with the datasets replaced with your real data:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\ntf.enable_eager_execution()\n\nmodel = tf.keras.applications.inception_resnet_v2.InceptionResNetV2(weights=None)\n\nmodel.compile(tf.keras.optimizers.Adam(), loss=tf.keras.losses.categorical_crossentropy)\n\n### Replace with tf.data.Datasets for your actual training data!\ntrain_x = tf.data.Dataset.from_tensor_slices(tf.random.normal((10,299,299,3)))\ntrain_y = tf.data.Dataset.from_tensor_slices(tf.random.uniform((10,), maxval=10, dtype=tf.int32))\ntraining_data = tf.data.Dataset.zip((train_x, train_y)).batch(BATCH_SIZE)\n\nmodel.fit(training_data)\n\n<\/code><\/pre>\n\n<p>This approach works as is in TensorFlow 2.0 too as mentioned in your title. <\/p>\n"}]}
+{"QuestionId":57565037,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-19T23:29:12.410","AcceptedAnswerId":null,"OwnerUserId":11681667.0,"Title":"Trilinear Interpolation 1820: error: -215:Assertion failed","Body":"<p>I received this opencv error <code>1820: error: (-215:Assertion failed)<\/code> when developing some Trilinear Interpolation Code<\/p>\n\n<p>I'm trying to figure out what it means.<\/p>\n\n<p>I'm also currently wondering if my coding methodology is correct for Trilinear.<\/p>\n\n<p>Also,  is there a way to get the flow algorithm to work in colour \n    instead of grayscale?  I tried in colour, but received a similar error.<\/p>\n\n<p>Three Frames for download here <\/p>\n\n<ol>\n<li><a href=\"https:\/\/www.dropbox.com\/s\/fm1aunl9iyury1w\/3325_71.png?dl=0\" rel=\"nofollow noreferrer\">Frame 1<\/a><\/li>\n<li><a href=\"https:\/\/www.dropbox.com\/s\/ryq1inide2ef3e0\/3325_72.png?dl=0\" rel=\"nofollow noreferrer\">Frame 2<\/a><\/li>\n<li><a href=\"https:\/\/www.dropbox.com\/s\/icvworeiee9ueaq\/3325_73.png?dl=0\" rel=\"nofollow noreferrer\">Frame 3<\/a><\/li>\n<\/ol>\n\n<p>Code:<\/p>\n\n<pre><code>import numpy as np\nimport cv2\nimport scipy as sp\n\nImage1_Fp = \".\\\\3325_71.png\" to \nImage2_Fp = \".\\\\3325_72.png\" to \nImage3_Fp = \".\\\\3325_73.png\" to \n\nWrite_Image_Location_Tri_Sampled = \".\\\\3325_syn_Tri.png\"\n\ndef imread(filename):\n    im = cv2.imread(filename)\n    return im\n\nif __name__ == '__main__':\n\n    im1 = imread(Image1_Fp)\n    im2 = imread(Image2_Fp)\n    im3 = imread(Image3_Fp)\n\n    im3gray = cv2.cvtColor(im3, cv2.COLOR_BGR2GRAY)\n    im1gray = cv2.cvtColor(im1, cv2.COLOR_BGR2GRAY)\n    flow = cv2.calcOpticalFlowFarneback(im1gray, im3gray, flow=1, pyr_scale=0.5, levels=2, winsize=10, iterations=5, poly_n=7, poly_sigma=1.2, flags=0)\n\n    # Trilinear Interpolation using meshgrid and remap\n    half_flow = 0.5 * flow # reduce the flow lines all by half.\n    h, w = flow.shape[:2]\n    grid_x, grid_y = np.meshgrid(h, w) # Identifies the coordinates of the grid with x, and y value,\n    coor_x_1 = grid_x + half_flow[:, :, 0] # So the flow is a displacement, and you need to add the pixel location to it to get where that displacement came from on the grid\n    coor_y_1 = grid_y + half_flow[:, :, 1]\n\n    # Finds the interpolated intensity from the pixels closest to the flow line from RGB Image 1\n    output_1 = cv2.remap(im1, coor_x_1, coor_y_1, cv2.INTER_CUBIC, borderMode =cv2.BORDER_REPLICATE)\n    coor_x_2 = grid_x - half_flow[:, :, 0]\n    coor_y_2 = grid_y - half_flow[:, :, 1]\n\n    # Finds the interpolated intensity from the pixels closest to the flow line from RGB Image 2\n    output_2 = cv2.remap(im2, coor_x_2, coor_y_2, cv2.INTER_CUBIC, borderMode =cv2.BORDER_REPLICATE)\n\n    # combined half of each\n    Combined_Output = (output_1 \/ 2) + (output_2 \/ 2) # Applies Trilinear part for a synthesized batch\n\n    cv2.imwrite(Write_Image_Location_Tri_Sampled, Combined_Output)\n<\/code><\/pre>\n\n<p>I am expecting to see a synthesized image that looks similar to the one in-between the two images, but I receive this error.<\/p>\n\n<pre><code>Exception has occurred: error\nOpenCV(4.1.0) C:\\projects\\opencv-python\\opencv\\modules\\imgproc\\src\\imgwarp.cpp:1820: error: (-215:Assertion failed) ((map1.type() == CV_32FC2 || map1.type() == CV_16SC2) &amp;&amp; map2.empty()) || (map1.type() == CV_32FC1 &amp;&amp; map2.type() == CV_32FC1) in function 'cv::remap'\n\nFile \"..\\OpenCVOpticalFLowAndTrilinearInterp.py\", line 90, in &lt;module&gt;\n    output_1 = cv2.remap(im1, coor_x_1, coor_y_1, cv2.INTER_CUBIC, borderMode =cv2.BORDER_REPLICATE)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57565249,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-agents>","CreationDate":"2019-08-20T00:10:25.983","AcceptedAnswerId":57661285.0,"OwnerUserId":3083470.0,"Title":"ValueError: Could not find matching function to call loaded from the SavedModel","Body":"<p>I am trying to load a <code>tf-agents<\/code> policy I saved via<\/p>\n\n<pre><code>try:\n    PolicySaver(collect_policy).save(model_dir + 'collect_policy')\nexcept TypeError:\n    tf.saved_model.save(collect_policy, model_dir + 'collect_policy')\n<\/code><\/pre>\n\n<p>Quick explanation for the try\/except block: When originally creating the policy, I can save it via <code>PolicySaver<\/code>, but when I load it again for another training run, it is a <code>SavedModel<\/code> and can therefore not be saved by <code>PolicySaver<\/code>.<\/p>\n\n<p>This seems to work fine, but now I want to use this policy for self-play, so I load the policy with <code>self.policy = tf.saved_model.load(policy_path)<\/code> in my AIPlayer class. When I try to use it for prediction, however, it does not work. Here's the (testing) code:<\/p>\n\n<pre><code>def decide(self, table):\n    state = table.getState()\n    timestep = ts.restart(np.array([table.getState()], dtype=np.float))\n    prediction = self.policy.action(timestep)\n    print(prediction)\n<\/code><\/pre>\n\n<p>the <code>table<\/code> passed into the function contains the state of the game and the <code>ts.restart()<\/code> function is copied from my custom pyEnvironment, so the timestep is constructed the exact same way as it would be in the environment. However, I get the following error message for the line <code>prediction=self.policy.action(timestep)<\/code>:<\/p>\n\n<pre><code>ValueError: Could not find matching function to call loaded from the SavedModel. Got:\n  Positional arguments (2 total):\n    * TimeStep(step_type=&lt;tf.Tensor 'time_step:0' shape=() dtype=int32&gt;, reward=&lt;tf.Tensor 'time_step_1:0' shape=() dtype=float32&gt;, discount=&lt;tf.Tensor 'time_step_2:0' shape=() dtype=float32&gt;, observation=&lt;tf.Tensor 'time_step_3:0' shape=(1, 79) dtype=float64&gt;)\n    * ()\n  Keyword arguments: {}\n\nExpected these arguments to match one of the following 2 option(s):\n\nOption 1:\n  Positional arguments (2 total):\n    * TimeStep(step_type=TensorSpec(shape=(None,), dtype=tf.int32, name='time_step\/step_type'), reward=TensorSpec(shape=(None,), dtype=tf.float32, name='time_step\/reward'), discount=TensorSpec(shape=(None,), dtype=tf.float32, name='time_step\/discount'), observation=TensorSpec(shape=(None,\n79), dtype=tf.float64, name='time_step\/observation'))\n    * ()\n  Keyword arguments: {}\n\nOption 2:\n  Positional arguments (2 total):\n    * TimeStep(step_type=TensorSpec(shape=(None,), dtype=tf.int32, name='step_type'), reward=TensorSpec(shape=(None,), dtype=tf.float32, name='reward'), discount=TensorSpec(shape=(None,), dtype=tf.float32, name='discount'), observation=TensorSpec(shape=(None, 79), dtype=tf.float64, name='observation'))\n    * ()\n  Keyword arguments: {}\n<\/code><\/pre>\n\n<p>What am I doing wrong? Is it really just the tensor names or are the shapes the problem and how can I change that?<\/p>\n\n<p>Any ideas how to further debug this are appreciated.<\/p>\n","answers":[{"AnswerId":"57661285","CreationDate":"2019-08-26T15:59:52.463","ParentId":null,"OwnerUserId":"3083470","Title":null,"Body":"<p>I got it to work by constructing the TimeStep manually:<\/p>\n\n<pre><code>    step_type = tf.convert_to_tensor(\n        [0], dtype=tf.int32, name='step_type')\n    reward = tf.convert_to_tensor(\n        [0], dtype=tf.float32, name='reward')\n    discount = tf.convert_to_tensor(\n        [1], dtype=tf.float32, name='discount')\n    observations = tf.convert_to_tensor(\n        [state], dtype=tf.float64, name='observations')\n    timestep = ts.TimeStep(step_type, reward, discount, observations)\n    prediction = self.policy.action(timestep)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57565456,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-08-20T00:52:05.590","AcceptedAnswerId":null,"OwnerUserId":588855.0,"Title":"Keras - TypeError: 'NoneType' object cannot be interpreted as an index","Body":"<p>I have the following code snippet:<\/p>\n\n<pre><code>model = load_model('model.h5')\n\noptimizer = optimizers.Adam(lr=0.0001)\n\nfor root, dirs, files in os.walk(test_directory):\n    sortedFiles = sorted(files, key=lambda x:int(x.split('.')[0]))\n    for file in sortedFiles[0:]:\n        img = cv2.imread(root + '\/' + file)\n        test_images.append(img)\n\ntest_images = np.array(test_images)\ntest_images = test_images.reshape((None,512,512,3))\n<\/code><\/pre>\n\n<p>Why am I getting:<\/p>\n\n<pre><code>TypeError: 'NoneType' object cannot be interpreted as an index\n<\/code><\/pre>\n\n<p>at:<\/p>\n\n<pre><code>test_images = test_images.reshape((None,512,512,3))\n<\/code><\/pre>\n\n<p>Thanks a lot.<\/p>\n","answers":[{"AnswerId":"57565635","CreationDate":"2019-08-20T01:27:38.997","ParentId":null,"OwnerUserId":"5539758","Title":null,"Body":"<p>While <code>keras<\/code> sometimes has <code>None<\/code> in data shapes, Numpy does not. If you want to have a dimension of variable size, you put <code>-1<\/code> instead. Change your last line to:<\/p>\n\n<pre><code>test_images = test_images.reshape((-1,512,512,3))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57565503,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-08-20T01:02:34.610","AcceptedAnswerId":null,"OwnerUserId":7078575.0,"Title":"How can I save trained FCN model and test new image on trained FCN model?","Body":"<p>I have trained FCN on my own dataset, my code is running. however, I want to save my model and test new image using trained model, can anyone please help me?<\/p>\n\n<pre><code>from keras import optimizers\n\n\nsgd = optimizers.SGD(lr=1E-2, decay=5**(-4), momentum=0.9, nesterov=True)\nmodel.compile(loss='categorical_crossentropy',\n          optimizer=sgd,\n          metrics=['accuracy'])\n\nhist1 = model.fit(X_train,y_train,\n              validation_data=(X_test,y_test),\n              batch_size=2,epochs=20,verbose=1)\n\n\nfor key in ['loss', 'acc', 'val_loss', 'val_acc']:\nplt.plot(hist1.history[key],label=key)\nplt.legend()\nplt.show()\n\n\ny_pred = model.predict(X_test)\ny_predi = np.argmax(y_pred, axis=3)\ny_testi = np.argmax(y_test, axis=3)\nprint(y_testi.shape,y_predi.shape)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57565871,"AnswerCount":0,"Tags":"<tensorflow><python-multiprocessing>","CreationDate":"2019-08-20T02:09:11.383","AcceptedAnswerId":null,"OwnerUserId":3399944.0,"Title":"Scope of variables replicated by multiprocessing","Body":"<p>I'm trying to speed up training multiple models by using Python's <code>multiprocessing.Pool.apply_async<\/code>. In order to save memory, I've converted my pandas dataframe into 32bit floats. I can see the pickle is ~2gb on disk. Once I've loaded the pickle into memory, I can see in task manager that python is using ~4-5gb memory.<\/p>\n\n<p>I then create my tensorflow models in the main thread like so:<\/p>\n\n<pre><code>for i in range(n):\n    self.estimators.append(DNNEstimator())\n\nwith Pool(processes=4) as pool:\n    for i in range(n):\n        # DNNEstimator is a wrapper over a keras neural network with scikit-learn\n        # compatible interface (i.e., it has methods fit(x, y) and predict(x))\n        self.results.append(pool.apply_async(self.estimators[i].fit, (x, y)))\n\n    for i in range(n):\n        self.results[i] = self.results[i].get()\n<\/code><\/pre>\n\n<p>My understanding of how <code>multiprocessing<\/code> works is that it pickles the data that needs to be processed, and run it in a new python process\/instance. I noticed that each of the python processes was taking ~6gb of memory during training. I'm suspecting subprocesses are taking this much memory because it recreated all the variables in the main thread. So my question is, how much of the main thread's scope is recreated in each subprocess?<\/p>\n","answers":[]}
+{"QuestionId":57566057,"AnswerCount":0,"Tags":"<python><pytorch>","CreationDate":"2019-08-20T02:41:52.583","AcceptedAnswerId":null,"OwnerUserId":2421996.0,"Title":"Pytorch get stuck at .to(\u2018cuda\u2019) or .cuda()","Body":"<p>I use a docker environment to do some pytorch training work, but frequently get stuck at <code>.to(\u2018cuda\u2019)<\/code> or <code>.cuda()<\/code> calls. Not know why.\nSome system behaviors I notice: The training process has a cpu usage of 100%, and almost 97% are sys time as shown by top 1. I used strace to debug the training process, and also used the \/proc file system, found the process keep getting stuck on a poll syscall to a pipe, but I don\u2019t know where the pipe is used in pytorch.<\/p>\n","answers":[]}
+{"QuestionId":57566062,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-20T02:42:39.590","AcceptedAnswerId":null,"OwnerUserId":1462220.0,"Title":"What's the recommended way to represent distributions of complex-valued random variables in Tensorflow Probability?","Body":"<p>I don't see any existing classes for complex Normal or complex Wishart distributions in the Tensorflow Probability api docs. Is this because it's easy to synthesize these distributions from existing apis? If so I would appreciate an example or a reference.<\/p>\n","answers":[]}
+{"QuestionId":57566447,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><linear-regression>","CreationDate":"2019-08-20T03:45:33.620","AcceptedAnswerId":57567366.0,"OwnerUserId":10018602.0,"Title":"Setting Up Simple Tensorflow Linear Regression Returning NaN Values?","Body":"<p>I'm new to Tensorflow and I want to know why I'm getting nan values for cost, W and b at each epoch? I'm setting up a traffic game and I'd like to train a model to be able to predict what the best duration of a green light would be, based on previous rewards and previous green light durations. I tried following <a href=\"https:\/\/www.geeksforgeeks.org\/linear-regression-using-tensorflow\/\" rel=\"nofollow noreferrer\">this guide<\/a> to set it up, but doesn't seem to be working. Any ideas? This could should replicate the issue I'm having, and I've added in lots of prints to be able to help someone more experienced than I. Thanks.<\/p>\n\n<pre><code>import numpy as np\nimport random\nimport matplotlib.pyplot as plt\nimport tensorflow as tf\nimport warnings\n\nwarnings.simplefilter(action='once', category=FutureWarning) # future warnings annoy me\n\n# add in a couple of rewards and light durations\ncurrent_reward = [-1000,-900,-950]\ncurrent_green = [10,12,12]\n\ncurrent_reward = np.array(current_reward)\ncurrent_green = np.array(current_green)\n\n# Pass in reward and green_light\ndef green_light_duration_new(current_reward, current_green):\n    # Predicting the best light duration based on previous rewards.\n    # predict the best duration based on previous step's reward value, using simple linear regression model\n    x = current_reward\n    y = current_green\n    n = len(x)\n    # Plot of Training Data  \n    plt.scatter(x, y) \n    plt.xlabel('Reward') \n    plt.ylabel('Green Light Duration') \n    plt.title(\"Training Data\") \n    plt.show() \n\n    X = tf.placeholder(\"float\") \n    Y = tf.placeholder(\"float\") \n    W = tf.Variable(np.random.randn(), name = \"W\") \n    b = tf.Variable(np.random.randn(), name = \"b\") \n    learning_rate = 0.01\n    training_epochs = 500\n    # Hypothesis \n    y_pred = tf.add(tf.multiply(X, W), b) \n    print('y_pred : ', y_pred)\n    print('y_pred dtype : ', y_pred.dtype)\n    # Mean Squared Error Cost Function \n    cost = tf.reduce_sum(tf.pow(y_pred-Y, 2)) \/ (2 * n)\n    print('cost : ', cost)\n    print('cost dtype: ', cost.dtype)\n    # Gradient Descent Optimizer \n    optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)    \n    # Global Variables Initializer \n    init = tf.global_variables_initializer()\n    # Starting the Tensorflow Session \n    with tf.Session() as sess: \n        # Initializing the Variables \n        sess.run(init) \n        # Iterating through all the epochs \n        for epoch in range(training_epochs): \n            # Feeding each data point into the optimizer using Feed Dictionary \n            for (_x, _y) in zip(x, y): \n                print('_x : ',_x)\n                print('_y : ',_y)\n                sess.run(optimizer, feed_dict = {X : _x, Y : _y}) \n            # Displaying the result after every 50 epochs \n            if (epoch + 1) % 50 == 0: \n                # Calculating the cost a every epoch \n                c = sess.run(cost, feed_dict = {X : x, Y : y}) \n                print('c : ', c)\n                print('c dtype : ', c.dtype)\n                print(\"Epoch\", (epoch + 1), \": cost =\", c, \"W =\", sess.run(W), \"b =\", sess.run(b)) \n        # Storing necessary values to be used outside the Session \n        training_cost = sess.run(cost, feed_dict ={X: x, Y: y}) \n        print('training_cost : ', training_cost)\n        print('training_cost dtype : ', training_cost.dtype)\n        weight = sess.run(W)\n        print('weight : ', weight)\n        print('weight : ', weight.dtype)\n        bias = sess.run(b)\n        print('bias : ', bias)\n        print('biad dtype : ', bias.dtype)\n    # Calculating the predictions \n    green_light_duration_new = weight * x + bias \n    print(\"Training cost =\", training_cost, \"Weight =\", weight, \"bias =\", bias, '\\n')\n    # Plotting the Results \n    plt.plot(x, y, 'ro', label ='Original data') \n    plt.plot(x, green_light_duration_new, label ='Fitted line') \n    plt.title('Linear Regression Result') \n    plt.legend() \n    plt.show() \n    return green_light_duration_new\n\n# Go to the training function\nnew_green_dur = green_light_duration_new(current_reward, current_green)\n\n# Append the predicted green light to its list\ncurrent_green.append(new_green_dur)\n\n# Go on to run the rest of the simulation with the new green light duration,\n# and append its subsequent reward to current_reward list to run again later.\n<\/code><\/pre>\n\n<p><strong>UPDATE WITH PICTURES FROM BELOW SOLUTION<\/strong>\nWith the solution provided below, it's only plotting one data point, not the three I input, and there's no line of best fit, and the axis coordinates on the bottom of the 2nd plot are not reflective of where the one data point truly is.<\/p>\n\n<p>Also, when you <code>print(current_green)<\/code> at the very end after the concat'ing, the array is just 3 zeros? Shouldn't it be 4? The first inputted 3 and then the latest predicted one? <\/p>\n\n<p>I don't understand what's happening here. Why scale the data? What I want is to be able to feed this regressor with a new list of X values (the rewards) from previous runs and have it return\/predict the best possible green light duration between 10 and 120 seconds, in the same scale as it went in. After that, it should add that duration to the <code>current_green<\/code> list. Thanks a lot, I'm still new. The plotting is a nice feature, but it's not entirely necessary, I just wanted to see that it was working how it was supposed to.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/2eKR4.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/2eKR4.png\" alt=\"Plot Screen 1\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/WasJE.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/WasJE.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"57567366","CreationDate":"2019-08-20T05:50:57.030","ParentId":null,"OwnerUserId":"11802889","Title":null,"Body":"<p>There are two error first please use MinMaxScaler to scale your data. During calculations when number go out of range NAN pops up\n2. Append doesnt work in numpy array.<\/p>\n\n<p>Here's complete solutions to your problem:<\/p>\n\n<pre><code>import numpy as np\nimport random\nimport matplotlib.pyplot as plt\nimport tensorflow as tf\nimport warnings\nfrom sklearn.preprocessing import MinMaxScaler\n\nwarnings.simplefilter(action='once', category=FutureWarning) # future warnings annoy me\n\n# add in a couple of rewards and light durations\ncurrent_reward = [[-1000,-900,-950]]\ncurrent_green = [[10,12,12]]\n\ncurrent_reward = np.array(current_reward)\ncurrent_green = np.array(current_green)\n\n\n\n\n\nscaler = MinMaxScaler()\nscaler.fit(current_reward)\ncurrent_reward= scaler.transform(current_reward)\n\nscaler.fit(current_green)\ncurrent_green=scaler.transform(current_green)\n\n# Pass in reward and green_light\ndef green_light_duration_new(current_reward, current_green):\n    # Predicting the best light duration based on previous rewards.\n    # predict the best duration based on previous step's reward value, using simple linear regression model\n    x = current_reward\n    y = current_green\n    n = len(x)\n\n\n\n\n    # Plot of Training Data  \n    plt.scatter(x, y) \n    plt.xlabel('Reward') \n    plt.ylabel('Green Light Duration') \n    plt.title(\"Training Data\") \n    plt.show() \n\n    X = tf.placeholder(\"float\") \n    Y = tf.placeholder(\"float\") \n    W = tf.Variable(np.random.randn(), name = \"W\") \n    b = tf.Variable(np.random.randn(), name = \"b\") \n    learning_rate = 0.01\n    training_epochs = 500\n    # Hypothesis \n    y_pred = tf.add(tf.multiply(X, W), b) \n    print('y_pred : ', y_pred)\n    print('y_pred dtype : ', y_pred.dtype)\n    # Mean Squared Error Cost Function \n    cost = tf.reduce_sum(tf.pow(y_pred-Y, 2)) \/ (2 * n)\n    print('cost : ', cost)\n    print('cost dtype: ', cost.dtype)\n    # Gradient Descent Optimizer \n    optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)    \n    # Global Variables Initializer \n    init = tf.global_variables_initializer()\n    # Starting the Tensorflow Session \n    with tf.Session() as sess: \n        # Initializing the Variables \n        sess.run(init) \n        # Iterating through all the epochs \n        for epoch in range(training_epochs): \n            # Feeding each data point into the optimizer using Feed Dictionary \n            for (_x, _y) in zip(x, y): \n                print('_x : ',_x)\n                print('_y : ',_y)\n                sess.run(optimizer, feed_dict = {X : _x, Y : _y}) \n            # Displaying the result after every 50 epochs \n            if (epoch + 1) % 50 == 0: \n                # Calculating the cost a every epoch \n                c = sess.run(cost, feed_dict = {X : x, Y : y}) \n                print('c : ', c)\n                print('c dtype : ', c.dtype)\n                print(\"Epoch\", (epoch + 1), \": cost =\", c, \"W =\", sess.run(W), \"b =\", sess.run(b)) \n        # Storing necessary values to be used outside the Session \n        training_cost = sess.run(cost, feed_dict ={X: x, Y: y}) \n        print('training_cost : ', training_cost)\n        print('training_cost dtype : ', training_cost.dtype)\n        weight = sess.run(W)\n        print('weight : ', weight)\n        print('weight : ', weight.dtype)\n        bias = sess.run(b)\n        print('bias : ', bias)\n        print('biad dtype : ', bias.dtype)\n    # Calculating the predictions \n    green_light_duration_new = weight * x + bias \n    print(\"Training cost =\", training_cost, \"Weight =\", weight, \"bias =\", bias, '\\n')\n    # Plotting the Results \n    plt.plot(x, y, 'ro', label ='Original data') \n    plt.plot(x, green_light_duration_new, label ='Fitted line') \n    plt.title('Linear Regression Result') \n    plt.legend() \n    plt.show() \n    return green_light_duration_new\n\n# Go to the training function\nnew_green_dur = green_light_duration_new(current_reward, current_green)\n\n# Append the predicted green light to its list\nnp.concatenate((current_green, new_green_dur))\n#current_green.append(new_green_dur)\n\n# Go on to run the rest of the simulation with the new green light duration,\n# and append its subsequent reward to current_reward list to run again later.\n<\/code><\/pre>\n"}]}
+{"QuestionId":57566490,"AnswerCount":0,"Tags":"<android><tensorflow>","CreationDate":"2019-08-20T03:51:41.917","AcceptedAnswerId":null,"OwnerUserId":11431594.0,"Title":"tensorflowinferenceinterface in a thread on android","Body":"<pre><code>inferenceInterface.feed(inputName, floatValues, 1, inputSize, inputSize, 3);\ninferenceInterface.run(new String[]{outputName},logStats);\ninferenceInterface.fetch(outputName, output);\n<\/code><\/pre>\n\n<p>When I ran my Android project, this part was very slow, so I tried to put this in a Thread. But the following error occurred. <\/p>\n\n<p><code>Failed to run TensorFlow inference with inputs:[input, input], outputs:[output]\njava.lang.IllegalArgumentException: Endpoint \"input:0\" fed more than once.<\/code><\/p>\n\n<p>I should not use tensorflowinferenceinterface in a Thread??\nIf so, what are ways I can speed it up?<\/p>\n","answers":[]}
+{"QuestionId":57566814,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-20T04:39:19.903","AcceptedAnswerId":null,"OwnerUserId":1009508.0,"Title":"Create Features from a custom Keras CNN","Body":"<p>I have been training this model for 1 epoch to see if it will work. What I want to do is take an image and create global features for the image. Right now after each training session, the features are all 0. Can someone please tell me the best way to create global features using Keras and cnn?<\/p>\n\n<p>Here is my model so far.<\/p>\n\n<pre><code>def create_base_network():\n    \"\"\"\n    Base network to be shared.\n    \"\"\"\n    model_input = Input(shape=(224,224,3))\n\n    x = Conv2D(64, (3, 3), activation='relu', padding='same')(model_input)\n    x = MaxPooling2D(pool_size=(2,2))(x)\n    x = Dropout(0.5)(x)\n\n    x = BatchNormalization()(x)\n    x = Conv2D(128, (3, 3), activation='relu', padding='same')(x)\n    x = Conv2D(128, (3, 3), activation='relu', padding='same')(x)\n    x = MaxPooling2D(pool_size=(2,2))(x)\n    x = Dropout(0.5)(x)\n\n    x = BatchNormalization()(x)\n    x = Conv2D(256, (3, 3), activation='relu', padding='same')(x)\n    x = Conv2D(256, (3, 3), activation='relu', padding='same')(x)\n    x = MaxPooling2D(pool_size=(2,2))(x)\n    x = Conv2D(256, (3, 3), activation='relu', padding='same')(x)\n    x = Conv2D(256, (3, 3), activation='relu', padding='same')(x)\n    x = MaxPooling2D(pool_size=(2,2))(x)\n    x = Dropout(0.5)(x)\n\n    x = BatchNormalization()(x)\n    x = Conv2D(512, (3, 3), activation='relu', padding='same')(x)\n    x = Conv2D(512, (3, 3), activation='relu', padding='same')(x)\n    x = MaxPooling2D(pool_size=(2,2))(x)\n    x = Conv2D(512, (3, 3), activation='relu', padding='same')(x)\n    x = Conv2D(512, (3, 3), activation='relu', padding='same')(x)\n    x = Conv2D(512, (3, 3), activation='relu', padding='same')(x)\n    #x = MaxPooling2D(pool_size=(2,2))(x)\n    x = Dropout(0.5)(x)\n\n    #x = Flatten()(x)\n    x = Dense(1024, activation='relu')(x)\n    x = BatchNormalization()(x)\n    x = Dropout(0.7)(x)\n\n    x = Dense(1024, activation='relu')(x)\n    x = Dense(1024, activation='relu')(x)\n    x = BatchNormalization()(x)\n    x = Dropout(0.7)(x)\n\n    x = Dense(4096, activation='relu')(x)\n    x = Dense(4096, activation='relu')(x)\n    x = BatchNormalization()(x)\n    x = Dropout(0.7)(x)\n    # This layer is what the features are\n    x = Dense(4096, activation='relu')(x)\n\n\n    model = Model(inputs=model_input, outputs=x)\n    return model\n<\/code><\/pre>\n\n<p>Please and thank you.<\/p>\n","answers":[]}
+{"QuestionId":57567028,"AnswerCount":0,"Tags":"<json><python-3.x><keras>","CreationDate":"2019-08-20T05:11:23.167","AcceptedAnswerId":null,"OwnerUserId":11828245.0,"Title":"json.decoder.JSONDecodeError: Expecting value: line 1 column 1 (char 0) while reading from json","Body":"<p>I am trying to test a model for authorship attribution from <a href=\"https:\/\/github.com\/asad1996172\/Authorship-attribution-using-CNN\" rel=\"nofollow noreferrer\">https:\/\/github.com\/asad1996172\/Authorship-attribution-using-CNN<\/a> by updating the code to new syntax but getting the mentioned error.<\/p>\n\n<pre><code># loading json and create model\n\njson_file = json.load(codecs.open(cur_dir_path + 'savedModelFiles_blogs\/' + 'Test.json', 'r', 'utf-8-sig'))\nloaded_model_json = json_file.get('Blog')\n\nloaded_model = model_from_json(loaded_model_json)\n<\/code><\/pre>\n\n<p>I am trying to get the name of the author but the output is:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/home\/kics\/PycharmProjects\/Authorship-attribution-using-CNN-master\/CNN\/testModel.py\", line 147, in &lt;module&gt;\n    loaded_model = model_from_json(loaded_model_json)\n  File \"\/home\/kics\/PycharmProjects\/untitled\/venv\/lib\/python3.5\/site-packages\/keras\/models.py\", line 208, in model_from_json\n    config = json.loads(json_string)\n  File \"\/usr\/lib\/python3.5\/json\/__init__.py\", line 319, in loads\n    return _default_decoder.decode(s)\n  File \"\/usr\/lib\/python3.5\/json\/decoder.py\", line 339, in decode\n    obj, end = self.raw_decode(s, idx=_w(s, 0).end())\n  File \"\/usr\/lib\/python3.5\/json\/decoder.py\", line 357, in raw_decode\n    raise JSONDecodeError(\"Expecting value\", s, err.value) from None\njson.decoder.JSONDecodeError: Expecting value: line 1 column 1 (char 0)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57567572,"AnswerCount":3,"Tags":"<python><machine-learning><keras><conv-neural-network><tensor>","CreationDate":"2019-08-20T06:10:57.500","AcceptedAnswerId":57581123.0,"OwnerUserId":11949810.0,"Title":"Modify Neural Network to Predict on 3 examples at a time","Body":"<p>I have here a convolutional neural network that takes a 96x96x3 picture and outputs a 1x128 encoding. (the undefined functions are just series' of layers)<\/p>\n\n<p>How would I modify the architecture to take a 3x96x96x3 input and produce a 3x128 output (forward propagate 3 times with the same parameters)?<\/p>\n\n<pre><code>def faceRecoModel(input_shape):\n    \"\"\"\n    Implementation of the Inception model used for FaceNet\n\n    Arguments:\n    input_shape -- shape of the images of the dataset\n\n    Returns:\n    model -- a Model() instance in Keras\n    \"\"\"\n\n    # Define the input as a tensor with shape input_shape\n    X_input = Input(input_shape)\n\n    # Zero-Padding\n    X = ZeroPadding2D((3, 3))(X_input)\n\n    # First Block\n    X = Conv2D(64, (7, 7), strides=(2, 2), name='conv1')(X)\n    X = BatchNormalization(axis=1, name='bn1')(X)\n    X = Activation('relu')(X)\n\n    # Zero-Padding + MAXPOOL\n    X = ZeroPadding2D((1, 1))(X)\n    X = MaxPooling2D((3, 3), strides=2)(X)\n\n    # Second Block\n    X = Conv2D(64, (1, 1), strides=(1, 1), name='conv2')(X)\n    X = BatchNormalization(axis=1, epsilon=0.00001, name='bn2')(X)\n    X = Activation('relu')(X)\n\n    # Zero-Padding + MAXPOOL\n    X = ZeroPadding2D((1, 1))(X)\n\n    # Second Block\n    X = Conv2D(192, (3, 3), strides=(1, 1), name='conv3')(X)\n    X = BatchNormalization(axis=1, epsilon=0.00001, name='bn3')(X)\n    X = Activation('relu')(X)\n\n    # Zero-Padding + MAXPOOL\n    X = ZeroPadding2D((1, 1))(X)\n    X = MaxPooling2D(pool_size=3, strides=2)(X)\n\n    # Inception 1: a\/b\/c\n    X = inception_block_1a(X)\n    X = inception_block_1b(X)\n    X = inception_block_1c(X)\n\n    # Inception 2: a\/b\n    X = inception_block_2a(X)\n    X = inception_block_2b(X)\n\n    # Inception 3: a\/b\n    X = inception_block_3a(X)\n    X = inception_block_3b(X)\n\n    # Top layer\n    X = AveragePooling2D(pool_size=(3, 3), strides=(1, 1), data_format='channels_first')(X)\n    X = Flatten()(X)\n    X = Dense(128, name='dense_layer')(X)\n\n    # L2 normalization\n    X = Lambda(lambda x: K.l2_normalize(x, axis=1))(X)\n\n    # Create model instance\n    model = Model(inputs=X_input, outputs=X, name='FaceRecoModel')\n\n    return model\n<\/code><\/pre>\n","answers":[{"AnswerId":"57581123","CreationDate":"2019-08-20T20:29:12.617","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>If you want to take inputs of shape <code>(batch_size, 3, 96, 96, 3)<\/code>, i.e. process 3 images as a single input sample, then you need to create a new model which uses your trained model and applies it on each of those 3 images independently. You can easily achieve this using <a href=\"https:\/\/keras.io\/layers\/wrappers\/\" rel=\"nofollow noreferrer\"><code>TimeDistributed<\/code><\/a> wrapper in Keras:<\/p>\n\n<pre><code>from keras.layers import TimeDistributed\n\ninp = Input(shape=(3, 96, 96, 3))\nout = TimeDistributed(the_trained_face_rec_model)(inp)\n\nmodel = Model(inp, out)\n<\/code><\/pre>\n\n<p>The output of this new model has a shape of <code>(batch_size, 3, 128)<\/code>. There is no need to compile or train this new model, because simply it is just a wrapper around your previously trained model. So you can use: <code>predictions = model.predict(my_images)<\/code>.<\/p>\n"},{"AnswerId":"57567840","CreationDate":"2019-08-20T06:32:03.873","ParentId":null,"OwnerUserId":"7352626","Title":null,"Body":"<p>Edited: As mentioned by op that the network should process 3 images at a time as dataset would be shape (m,3,96,96,3), then one easy way is to create three standard inception networks in a parallel fashion and then concatenate the output received from each of them. <\/p>\n"},{"AnswerId":"57567860","CreationDate":"2019-08-20T06:33:34.513","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You don't need to modify anything, the first dimension of the input in Keras is always the batch dimension, so you literally need to feed an input of shape <code>(3, 96, 96, 3)<\/code> and you will get a corresponding output of <code>(3, 128)<\/code>. No modifications to the code are needed.<\/p>\n"}]}
+{"QuestionId":57567891,"AnswerCount":1,"Tags":"<python><image-processing><theano><video-tracking>","CreationDate":"2019-08-20T06:35:49.460","AcceptedAnswerId":null,"OwnerUserId":11253858.0,"Title":"Expected an Integer","Body":"<p>I am getting error in last line of the code. Saying \"Expected an Integer\".\nActually this code took from VGGNET. I am using Python 3.6.9.<\/p>\n\n<pre><code>import theano.tensor as T\nbatch,in_channels,H,W = T.shape(x)\nself.VGGout_resize = 16\nxt = x[:,::-1,:,:]\nxt = T.set_subtensor(xt[:,0,:,:], xt[:,0,:,:] - 103.939)\nxt = T.set_subtensor(xt[:,1,:,:], xt[:,1,:,:] - 116.779)\nxt = T.set_subtensor(xt[:,2,:,:], xt[:,2,:,:] - 123.68)\nxt = xt[:,:,0:self.VGGout_resize* \n(H\/self.VGGout_resize),0:self.VGGout_resize*(W\/self.VGGout_resize) ]\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/HC2U5.png\" rel=\"nofollow noreferrer\">\"Error Results Image 1\"<\/a>\n <a href=\"https:\/\/i.stack.imgur.com\/yQFs0.png\" rel=\"nofollow noreferrer\">\"Error Results Image 2\"<\/a><\/p>\n","answers":[{"AnswerId":"57567967","CreationDate":"2019-08-20T06:41:20.003","ParentId":null,"OwnerUserId":"11900574","Title":null,"Body":"<p>This is probably because you're slicing and output of <code>(H\/self.VGGout_resize)<\/code> or <code>(W\/self.VGGout_resize)<\/code> could be a non-integer. Try this:<\/p>\n\n<pre><code>self.VGGout_resize = 16\nxt = x[:,::-1,:,:]\nxt = T.set_subtensor(xt[:,0,:,:], xt[:,0,:,:] - 103.939)\nxt = T.set_subtensor(xt[:,1,:,:], xt[:,1,:,:] - 116.779)\nxt = T.set_subtensor(xt[:,2,:,:], xt[:,2,:,:] - 123.68)\nxt = xt[:,:,0:self.VGGout_resize*int(H\/self.VGGout_resize),0:self.VGGout_resize*int(W\/self.VGGout_resize)]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57568055,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-20T06:48:38.657","AcceptedAnswerId":57568135.0,"OwnerUserId":11907197.0,"Title":"How to fix : ERROR: Could not find a version that satisfies the requirement tensorflow-gpu==1.14.0 (from versions: none)","Body":"<p>I am currently having an issue with the installation of the Tensorflow GPU on mac os<\/p>\n\n<p>I have uninstalled tensorflow 1.14 and trying to install tensorflow 1.14 GPU in order to train a model.<\/p>\n\n<pre><code>(deepspeech-venv) Chabanis-MacBook-Pro:Deepspeech chabani$ pip3 install 'tensorflow-gpu==1.14.0'\nCollecting tensorflow-gpu==1.14.0\n  ERROR: Could not find a version that satisfies the requirement tensorflow-gpu==1.14.0 (from versions: none)\nERROR: No matching distribution found for tensorflow-gpu==1.14.0\n<\/code><\/pre>\n\n<p>I expected that this would have installed the Tensorflow GPU 1.14 properly. Where am I going wrong here?<\/p>\n","answers":[{"AnswerId":"57568135","CreationDate":"2019-08-20T06:55:04.800","ParentId":null,"OwnerUserId":"4046632","Title":null,"Body":"<p>According to <a href=\"https:\/\/www.tensorflow.org\/install\/pip\" rel=\"nofollow noreferrer\">this<\/a> there is no GPU support on mac<\/p>\n\n<blockquote>\n  <p>macOS 10.12.6 (Sierra) or later (64-bit) (<strong>no GPU support<\/strong>)<\/p>\n<\/blockquote>\n\n<p>No GPU support on Mac is mentioned also <a href=\"https:\/\/www.tensorflow.org\/install\/source\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<blockquote>\n  <p>There is no GPU support for macOS.<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":57568455,"AnswerCount":2,"Tags":"<tensorflow><image-segmentation><loss-function>","CreationDate":"2019-08-20T07:16:22.793","AcceptedAnswerId":58583829.0,"OwnerUserId":4028648.0,"Title":"TensorFlow: What is wrong with my (generalized) dice loss implementation?","Body":"<p>I use TensorFlow 1.12 for semantic (image) segmentation based on materials. With a multinomial cross-entropy loss function, this yields okay-ish results, especially considering the sparse amount of training data I\u00b4m working with, with mIoU of 0.44:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/7kTAi.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/7kTAi.jpg\" alt=\"Segmentations after training with cross entropy loss\"><\/a><\/p>\n\n<p>When I replace this with my dice loss implementation, however, the networks predicts way less smaller segmentations, which is contrary to my understanding of its theory. I thought it\u00b4s supposed to work better with imbalanced datasets and should be better at predicting the smaller classes:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/SSQRX.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/SSQRX.jpg\" alt=\"Segmentations after training with dice loss\"><\/a><\/p>\n\n<p>A table visualizes this better; as you can see, with dice loss a lot more smaller classes are never predicted (hence the undefined precision). With cross-entropy, at least some predictions are made for all classes:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/O5XOk.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/O5XOk.jpg\" alt=\"Table comparing metrics for the different losses\"><\/a><\/p>\n\n<p>I initially thought that this is the networks way of increasing mIoU (since my understanding is that dice loss optimizes dice loss directly). However, mIoU with dice loss is 0.33 compared to cross entropy\u00b4s 0.44 mIoU, so it has failed in that regard. I\u00b4m now wondering whether my implementation is correct:<\/p>\n\n<pre><code>def dice_loss(onehots_true, logits):\n    probabilities = tf.nn.softmax(logits)\n    #weights = 1.0 \/ ((tf.reduce_sum(onehots_true, axis=0)**2) + 1e-3)\n    #weights = tf.clip_by_value(weights, 1e-17, 1.0 - 1e-7)\n    numerator = tf.reduce_sum(onehots_true * probabilities, axis=0)\n    #numerator = tf.reduce_sum(weights * numerator)\n    denominator = tf.reduce_sum(onehots_true + probabilities, axis=0)\n    #denominator = tf.reduce_sum(weights * denominator)\n    loss = 1.0 - 2.0 * (numerator + 1) \/ (denominator + 1)\n    return loss\n<\/code><\/pre>\n\n<p>Some implementations I found use weights, though I am not sure why, since mIoU isn\u00b4t weighted either. At any rate, training is prematurely stopped after one a few epochs with dreadful test results when I use weights, hence I commented them out.<\/p>\n\n<p>Does anyone see anything wrong with my dice loss implementation? I pretty faithfully followed online examples.<\/p>\n\n<p>In order to speed up the labeling process, I only annotated with parallelogram shaped polygons, and I copied some annotations from a larger dataset. This resulted in only a couple of ground truth segmentations per image:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9ExFT.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9ExFT.jpg\" alt=\"Ground truth annotations\"><\/a><\/p>\n\n<p>(This image actually contains slightly more annotations than average.)<\/p>\n","answers":[{"AnswerId":"57922064","CreationDate":"2019-09-13T10:38:53.410","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>According to <a href=\"https:\/\/lars76.github.io\/neural-networks\/object-detection\/losses-for-segmentation\/#dice-loss--f1-score\" rel=\"nofollow noreferrer\">this article<\/a>, your implementation of the Dice loss is incorrect. Instead of:<\/p>\n\n<pre><code>loss = 1.0 - 2.0 * (numerator + 1) \/ (denominator + 1)\n<\/code><\/pre>\n\n<p>You should have:<\/p>\n\n<pre><code>loss = 1.0 - (2.0 * numerator + 1) \/ (denominator + 1)\n<\/code><\/pre>\n"},{"AnswerId":"58583829","CreationDate":"2019-10-27T22:17:58.433","ParentId":null,"OwnerUserId":"2127561","Title":null,"Body":"<p>I'm going to add the formula for reference to anyone who answers in the future. The generalized dice loss is given by:\n<a href=\"https:\/\/i.stack.imgur.com\/JEUAe.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/JEUAe.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>picture taken from <a href=\"https:\/\/arxiv.org\/abs\/1707.03237\" rel=\"nofollow noreferrer\">Sudre et al<\/a>.<\/p>\n\n<p>Class is iterated by <code>l<\/code>. Each pixel location is iterated by <code>n<\/code>. The probabilities <code>p_ln<\/code> can be generated using softmax or sigmoid in your network output.<\/p>\n\n<hr>\n\n<p>In your implementation, the loss is <strong>summed<\/strong> across the batch. That would produce a very large loss value and your network gradients would explode. Instead, you need to use the average. Note that the weights are required to ensure you combat the class imbalance problem.<\/p>\n\n<p>There is no concrete proof that GDL outperforms cross-entropy, save in a very specific example noted in the paper. GDL is attractive because it is directly related to IoU, ergo the loss function and evaluation metrics would improve hand-in-hand. If you still haven't managed to train your network, I'd recommend moving to cross-entropy for good.<\/p>\n"}]}
+{"QuestionId":57569460,"AnswerCount":2,"Tags":"<python><tensorflow><keras><neural-network><deep-learning>","CreationDate":"2019-08-20T08:21:16.107","AcceptedAnswerId":57570216.0,"OwnerUserId":6022269.0,"Title":"Freezing keras layer doesn't change sumarry trainable params","Body":"<p>I have 2 models that have been compiled and trained in the script. now I am trying to concatenate the second to last layer, freeze all the layers, adding new trainable layers. <\/p>\n\n<p>here are the trained models:<\/p>\n\n<pre><code>morf_input = keras.layers.Input([np.shape(x)[1]])\nmorf_layer1 = keras.layers.Dense(800,activation=\"tanh\")(morf_input)\nmorf_layer2 = keras.layers.Dense(800,activation=\"tanh\" )(morf_layer1)\nmorf_layer3 = keras.layers.Dense(600,activation=\"tanh\" )(morf_layer2)\nmorf_layer4 = keras.layers.Dense(300,activation=\"tanh\" )(morf_layer3)\nmorf_layer5 = keras.layers.Dense(50,activation=\"tanh\" )(morf_layer4)\nmorf_bneck6 = keras.layers.Dense(30,activation=\"tanh\" )( morf_layer5)\nmorf_output = keras.layers.Dense(2,activation=\"sigmoid\")(morf_bneck6)\n\nmorf_model = keras.models.Model(inputs=morf_input, outputs=morf_output)\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>color_input = keras.layers.Input([np.shape(col_x)[1]])\ncolor_layer1 = keras.layers.Dense(800,activation=\"tanh\")( color_input)\ncolor_layer2 = keras.layers.Dense(800,activation=\"tanh\" )( color_layer1)\ncolor_layer3 = keras.layers.Dense(600,activation=\"tanh\" )( color_layer2)\ncolor_layer4 = keras.layers.Dense(300,activation=\"tanh\" )( color_layer3)\ncolor_layer5 = keras.layers.Dense(50,activation=\"tanh\" )( color_layer4)\ncolor_bneck6 = keras.layers.Dense(10,activation=\"tanh\" )( color_layer5)\ncolor_output = keras.layers.Dense(2,activation=\"sigmoid\")( color_bneck6)\n\ncolor_model = keras.models.Model(inputs= color_input, outputs= color_output)\n<\/code><\/pre>\n\n<p>then I tried to freeze those layer with:<\/p>\n\n<pre><code>morf_layer1.trainable = False\nmorf_layer2.trainable = False\nmorf_layer3.trainable = False\nmorf_layer4.trainable = False\nmorf_layer5.trainable = False\nmorf_bneck6.trainable = False\n\ncolor_layer1.trainable = False\ncolor_layer2.trainable = False\ncolor_layer3.trainable = False\ncolor_layer4.trainable = False\ncolor_layer5.trainable = False\ncolor_bneck6.trainable = False\n<\/code><\/pre>\n\n<p>then create a new model with those layers<\/p>\n\n<pre><code>concat_layer= keras.layers.Concatenate()([morf_bneck6, color_bneck6])\ncon_out_layer1 = keras.layers.Dense(500,activation=\"tanh\")(concat_layer)\ncon_out_layer2 = keras.layers.Dense(400,activation=\"tanh\")(con_out_layer1)\ncon_out_layer3 = keras.layers.Dense(300,activation=\"tanh\")(con_out_layer2)\ncon_out_layer4 = keras.layers.Dense(30,activation=\"tanh\")(con_out_layer3)\noutput = keras.layers.Dense(2,activation=\"sigmoid\")(con_out_layer4)\n\nmodel = keras.models.Model(inputs=[morf_input, color_input], outputs=output)\n<\/code><\/pre>\n\n<p>I compiled the model<\/p>\n\n<pre><code>model.compile(optimizer=keras.optimizers.SGD(lr=0.008, decay=1e-6, momentum=0.9, nesterov=False),\n              loss='binary_crossentropy',\n              metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>but the <code>model.summary()<\/code> shows<\/p>\n\n<pre><code>Total params: 3,035,432\nTrainable params: 3,035,432\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>shouldn't the frozen layer increase the <code>Non-trainable<\/code> params?<\/p>\n","answers":[{"AnswerId":"57570216","CreationDate":"2019-08-20T09:05:48.947","ParentId":null,"OwnerUserId":"423926","Title":null,"Body":"<p>Since you want to freeze all but last 6 layers use<\/p>\n\n<pre><code>for layer in model.layers[:-6]:\n    layer.trainable = False\n<\/code><\/pre>\n\n<h3>Working Example<\/h3>\n\n<pre><code># Model 1\ninputs_1 = keras.layers.Input(shape=(10,))\nl_1 = keras.layers.Dense(15,activation=\"tanh\")(inputs_1)\noutputs_1 = keras.layers.Dense(2,activation=\"sigmoid\")(l_1)\nmodel_1 = keras.models.Model(inputs_1, outputs_1)\nmodel_1.compile(optimizer=keras.optimizers.SGD(lr=0.008),\n              loss='binary_crossentropy',\n              metrics=['accuracy'])\nprint (\"Taining Model 1\")\nmodel_1.fit(np.random.randn(100,10), np.random.randn(100,2))\n\n# Model 2\ninputs_2 = keras.layers.Input(shape=(10,))\nl_2 = keras.layers.Dense(15,activation=\"tanh\")(inputs_2)\noutputs_2 = keras.layers.Dense(2,activation=\"sigmoid\")(l_2)\nmodel_2 = keras.models.Model(inputs_2, outputs_2)\nmodel_2.compile(optimizer=keras.optimizers.SGD(lr=0.008),\n              loss='binary_crossentropy',\n              metrics=['accuracy'])\nprint (\"Taining Model 2\")\nmodel_2.fit(np.random.randn(100,10), np.random.randn(100,2))\n\n# Combined Model\nconcat_layer= keras.layers.Concatenate()([outputs_1, outputs_2])\ncon_out_layer1 = keras.layers.Dense(5,activation=\"tanh\")(concat_layer)\noutput = keras.layers.Dense(2,activation=\"sigmoid\")(con_out_layer1)\n\nmodel = keras.models.Model(inputs=[inputs_1, inputs_2], outputs=output)\n\nmodel.summary()\n# Freeze all but last two layers (Concatenate is anyway not a \n# trainable layer)\nfor layer in model.layers[:-2]:\n    layer.trainable = False\nmodel.summary()\nmodel.compile(optimizer=keras.optimizers.SGD(lr=0.008),\n              loss='binary_crossentropy',\n              metrics=['accuracy'])\n\nprint (\"Taining Combined Model\")\nmodel.fit([np.random.randn(100,10),np.random.randn(100,10)],np.random.randn(100,2))\n<\/code><\/pre>\n\n<p><strong>Sample output<\/strong><\/p>\n\n<pre><code>......\nTotal params: 431\nTrainable params: 431\nNon-trainable params: 0\n......\n......\nTotal params: 431\nTrainable params: 37\nNon-trainable params: 394\n<\/code><\/pre>\n"},{"AnswerId":"57569707","CreationDate":"2019-08-20T08:37:04.417","ParentId":null,"OwnerUserId":"8380638","Title":null,"Body":"<p>Try the following:<\/p>\n\n<ol>\n<li>Layer definition<\/li>\n<li>Moder instantiation<\/li>\n<li>Set trainable to false<\/li>\n<li>Compile model<\/li>\n<\/ol>\n"}]}
+{"QuestionId":57569935,"AnswerCount":0,"Tags":"<python><keras><deep-learning><conv-neural-network><keras-layer>","CreationDate":"2019-08-20T08:49:39.513","AcceptedAnswerId":null,"OwnerUserId":11950406.0,"Title":"Create a custom layer in keras with user-defined connections","Body":"<p>I am working on a model with custom connections in the second layer. The following code works but I am not able to save the second layer. I tried saving the model architecture and weights separately but doesn't work. Creating a custom layer might help in saving the model with the weights. I would like to create a custom layer for the second layer and I am not sure how it can be done using keras custom layer. Can someone provide some solution related to saving the model or creating a custom layer?<\/p>\n\n<pre><code>input_data = Input(shape=(28, 28,1))\nfirst_layer = Conv2D(filters=16, kernel_size=(5,5), activation='relu', padding='same', input_shape = input_shape, use_bias = False)(input_data)\navg_1 = AveragePooling2D(pool_size=(2,2), data_format= 'channels_last')(first_layer)\nadd = []\nrandom_matrix = np.random.randint(0,15,(4,32))\nfor i in range(32):\n    group = []\n    for j in range(4):      \n        group_channel = Lambda(lambda x: x[:,:,:,random_matrix[j][i],np.newaxis])(avg_1)\n        conv_group = Conv2D(1, kernel_size=[5,5], strides=(stride,stride), activation = 'relu',padding='valid', data_format= 'channels_last',name = 'Conv_'+str(i*4+j)+'_Sparse', use_bias =False)(group_channel)\n        group.append(conv_group)\n    add_group = Add()(group)\n    add.append(add_group)\nsecond_layer = Concatenate()(add)    \navg_2 = AveragePooling2D(pool_size=(2,2), data_format= 'channels_last')(second_layer)\nflatten = Flatten()(avg_2)\ndense = Dense(num_classes, name = 'dense_3', use_bias = False)(flatten)\nactivation = Activation('softmax')(dense)\nmodel = Model(input_data, activation)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57570041,"AnswerCount":0,"Tags":"<tensorflow><tensorflow2.0>","CreationDate":"2019-08-20T08:55:54.027","AcceptedAnswerId":null,"OwnerUserId":287238.0,"Title":"Tensorflow 2.0 \"future proof\" way of using tf.feature_columns in tf.keras subclass api?","Body":"<p>I see there are some ways of using features columns wrapped in a tf.keras.layers.DenseFeatures layer:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/keras\/feature_columns#create_a_feature_layer\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/tutorials\/keras\/feature_columns#create_a_feature_layer<\/a><\/p>\n\n<p>This suggests passing datasets into @tf.function wrapped \"def call\" methods?<\/p>\n\n<p>I don't see any mention (yet) of the interaction between tf.data.DataSet and tf.function. Are tf.data.DataSet considered tensors yet for the purpose of tf.function graph cache or will they be?<\/p>\n","answers":[]}
+{"QuestionId":57570043,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-08-20T08:55:58.497","AcceptedAnswerId":null,"OwnerUserId":3513917.0,"Title":"Filter data in pytorch tensor","Body":"<p>I have a tensor <code>X<\/code> like <code>[0.1, 0.5, -1.0, 0, 1.2, 0]<\/code>, and I want to implement a function called <code>filter_positive()<\/code>, it can filter the positive data into a new tensor and return the index of the original tensor. For example:<\/p>\n\n<pre><code>new_tensor, index = filter_positive(X)\n\nnew_tensor = [0.1, 0.5, 1.2]\nindex = [0, 1, 4]\n<\/code><\/pre>\n\n<p>How can I implement this function most efficiently in pytorch?<\/p>\n","answers":[{"AnswerId":"57570139","CreationDate":"2019-08-20T09:00:51.483","ParentId":null,"OwnerUserId":"1643939","Title":null,"Body":"<p>Take a look at <code>torch.nonzero<\/code> which is roughly equivalent to <code>np.where<\/code>. It translates a binary mask to indices:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>&gt;&gt;&gt; X = torch.tensor([0.1, 0.5, -1.0, 0, 1.2, 0])\n&gt;&gt;&gt; X\ntensor([ 0.1000,  0.5000, -1.0000,  0.0000,  1.2000,  0.0000])\n&gt;&gt;&gt; mask = X &gt;= 0\n&gt;&gt;&gt; mask\ntensor([1, 1, 0, 1, 1, 1], dtype=torch.uint8)\n&gt;&gt;&gt; torch.nonzero(mask)\ntensor([[0],\n        [1],\n        [3],\n        [4],\n        [5]])\n&gt;&gt;&gt; X[torch.nonzero(mask)]\ntensor([[0.1000],\n        [0.5000],\n        [0.0000],\n        [1.2000],\n        [0.0000]])\n<\/code><\/pre>\n\n<p>A solution would then be to write:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>mask = X &gt;= 0\nnew_tensor = X[mask]\nindices = torch.nonzero(mask)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57570337,"AnswerCount":0,"Tags":"<python><tensorflow><keras><ensemble-learning>","CreationDate":"2019-08-20T09:13:13.390","AcceptedAnswerId":null,"OwnerUserId":3050590.0,"Title":"tensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'sequential_input'","Body":"<p>I am working on <a href=\"https:\/\/en.wikipedia.org\/wiki\/Multiclass_classification\" rel=\"nofollow noreferrer\">OVA<\/a> (one vs all) classification problem. For that, I have trained Keras binary classifiers with sigmoid function and binary_crossentropy. I need to ensemble them to multiclass model similar to <a href=\"https:\/\/www.kaggle.com\/mgiraygokirmak\/keras-cnn-multi-model-ensemble-with-voting\" rel=\"nofollow noreferrer\">here<\/a> When I am trying to do that I get the following error <\/p>\n\n<pre><code> tensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'sequential_input' with dtype float and shape [?,224,224,3]\n         [[{{node sequential_input}}]]\n<\/code><\/pre>\n\n<p>Program code<\/p>\n\n<pre><code>for i in os.listdir(model_root): \/\/loading all the models\nprint(i)\nfilename = model_root + \"\/\" + i\n# load model\nmodel = load_model(filename, custom_objects={'KerasLayer': hub.KerasLayer})\nmodels.append(model)\nprint(len(models))  \/\/3\n\n#Merge layer to fit a model\n\ninputs = tf.keras.Input(shape=(224,224,3))\noutputs = [m(inputs) for m in models]\noutputs = tf.keras.layers.concatenate(outputs, axis=-1)\nensemble = tf.keras.models.Model(inputs, outputs)\nensemble.compile(optimizer=tf.keras.optimizers.Adam(), loss='categorical_crossentropy', metrics=['accuracy'])\n\n#To fit the loaded models to the data\n\nsteps_per_epoch = image_data.samples \/\/ image_data.batch_size\nvalidation_steps = image_data_val.samples \/ image_data_val.batch_size\nensemble.fit((item for item in image_data), epochs=2,\n          steps_per_epoch=steps_per_epoch,\n          validation_data=(item for item in image_data_val), validation_steps=validation_steps, verbose=2)\n<\/code><\/pre>\n\n<p>I am getting this error on the fit function. Here is a traceback<\/p>\n\n<pre><code>Epoch 1\/2\nTraceback (most recent call last):\n  File \"&lt;input&gt;\", line 1, in &lt;module&gt;\n  File \"C:\\Program Files\\JetBrains\\PyCharm 2019.2\\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 \"C:\\Program Files\\JetBrains\\PyCharm 2019.2\\helpers\\pydev\\_pydev_imps\\_pydev_execfile.py\", line 18, in execfile\n    exec(compile(contents+\"\\n\", file, 'exec'), glob, loc)\n  File \"C:\/Users\/Pawandeep\/Desktop\/Python projects\/ensemble_image.py\", line 85, in &lt;module&gt;\n    validation_data=(item for item in image_data_val), validation_steps=validation_steps, verbose=2)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 673, in fit\n    initial_epoch=initial_epoch)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 1433, in fit_generator\n    steps_name='steps_per_epoch')\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_generator.py\", line 264, in model_iteration\n    batch_outs = batch_function(*batch_data)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 1175, in train_on_batch\n    outputs = self.train_function(ins)  # pylint: disable=not-callable\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py\", line 3292, in __call__\n    run_metadata=self.run_metadata)\n  File \"C:\\Python\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1458, in __call__\n    run_metadata_ptr)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'dense_1_target' with dtype float and shape [?,?]\n     [[{{node dense_1_target}}]]\n<\/code><\/pre>\n\n<p>My model looks like<\/p>\n\n<pre><code>        Model: \"model\"\n__________________________________________________________________________________________________\nLayer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput_1 (InputLayer)            [(None, 224, 224, 3) 0                                            \n__________________________________________________________________________________________________\nsequential_4 (Sequential)       (None, 1)            3541267     input_1[0][0]                    \n__________________________________________________________________________________________________\nsequential_8 (Sequential)       (None, 1)            3541267     input_1[0][0]                    \n__________________________________________________________________________________________________\nsequential_2 (Sequential)       (None, 1)            3541267     input_1[0][0]                    \n__________________________________________________________________________________________________\nconcatenate (Concatenate)       (None, 3)            0           sequential_4[1][0]               \n                                                                 sequential_8[1][0]               \n                                                                 sequential_2[1][0]               \n==================================================================================================\nTotal params: 10,623,801\nTrainable params: 3,006\nNon-trainable params: 10,620,795\n__________________________________________________________________________________________________\n<\/code><\/pre>\n\n<p>I can't find tensor dense_1_target. I don't understand which one it is referring to.<\/p>\n\n<p>My data looks like this:<\/p>\n\n<pre><code>image_data = image_generator.flow_from_directory(str(data_root), target_size=IMAGE_SIZE, subset='training')\nfor image_batch, label_batch in image_data:\nprint(\"Image batch shape: \", image_batch.shape) \/\/ (32, 224, 224, 3)\nprint(\"Label batch shape: \", label_batch.shape) \/\/ (32, 3)\n<\/code><\/pre>\n\n<p>Now where can I put the placeholders and how can I relate it to my input data. <\/p>\n","answers":[]}
+{"QuestionId":57570793,"AnswerCount":2,"Tags":"<python><pytorch><torch>","CreationDate":"2019-08-20T09:40:36.323","AcceptedAnswerId":57573452.0,"OwnerUserId":5918696.0,"Title":"Why my tensor defined in forward function can not be transformed in to cuda variable autonomously?","Body":"<p>In PyTorch, in the <code>forward<\/code> function of a model <\/p>\n\n<pre><code>class my_model(torch.nn.module):\n\n    ......\n\n    def forward():\n        self.a=torch.zeros(1)\n        #blabla\n<\/code><\/pre>\n\n<p>After <code>model.cuda()<\/code>, why <code>self.a<\/code> is still a <code>cpu<\/code> variable?<\/p>\n","answers":[{"AnswerId":"57573452","CreationDate":"2019-08-20T12:15:21.043","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>This is so by design. <\/p>\n\n<p>Only the tensors which are a part of the model will move with <code>model.cuda()<\/code> or <code>model.to(\"cuda\")<\/code>.<\/p>\n\n<p>These tensors are registered with <code>register_parameter<\/code> or <code>register_buffer<\/code>. This also includes child modules, parameters and buffers registered with the aforementioned functions. <\/p>\n\n<p>Even though <code>self.a=torch.zeros(1)<\/code> is actually a part of the class itself, by design it will not be moved to CUDA, instead you would need to do <code>a.to(\"cuda\")<\/code>, if you haven't used the <code>register*<\/code> methods.<\/p>\n"},{"AnswerId":"57583313","CreationDate":"2019-08-21T01:24:58.383","ParentId":null,"OwnerUserId":"5918696","Title":null,"Body":"<p>One solution:<\/p>\n\n<pre><code>device=whatever_parameter_in_your_forward_function_args.device\nself.a=torch.zeros(1).to(device)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57570851,"AnswerCount":1,"Tags":"<python><tensorflow><lstm><recurrent-neural-network><embedding>","CreationDate":"2019-08-20T09:43:45.420","AcceptedAnswerId":57572225.0,"OwnerUserId":6248445.0,"Title":"How to use Embedding layer for RNN with a categorical feature - Classification Task for RecoSys","Body":"<p>I would like to build a model (RNN >> LSTM) with an Embedding layer for a categorical feature (Item ID), My training set looks so:<\/p>\n\n<p><strong>train_x  = [[[184563.1], [184324.1], [187853.1], [174963.1], [181663.1]], [[\u2026],[\u2026],[\u2026],[\u2026],[\u2026]], \u2026]<\/strong><\/p>\n\n<p>I predict the sixth item ID:<\/p>\n\n<p><strong>train_y = [0,1,2, \u2026., 12691]<\/strong><\/p>\n\n<p>I have 12692 unique item IDs, length of timesteps = 5 and this is a classification task.<\/p>\n\n<p>This is a brief summary for what I've done so far: (Please correct me if I'm wrong)<\/p>\n\n<ol>\n<li>One-hot-encoding for the categorical feature:<\/li>\n<\/ol>\n\n<p><strong>train_x  = [[[1 0 0 \u2026 0 0 0], [0 1 0 \u2026 0 0 0], [0 0 1 \u2026 0 0 0], [\u2026], [\u2026]], [[\u2026],[\u2026],[\u2026],[\u2026],[\u2026]], \u2026]<\/strong><\/p>\n\n<ol start=\"2\">\n<li>Build model:<\/li>\n<\/ol>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential()\n\nmodel.add(Embedding(input_dim=12692 , output_dim=250, input_length=5))\n\nmodel.add(LSTM(128, return_sequences=True)\nmodel.add(Dropout(0.2)) \nmodel.add(BatchNormalization())\n\nmodel.add(LSTM(128, return_sequences=True))\nmodel.add(Dropout(0.1)) \nmodel.add(BatchNormalization())\n\nmodel.add(LSTM(128))\nmodel.add(Dropout(0.2)) \nmodel.add(BatchNormalization())\n\nmodel.add(Dense(32, activation='relu'))\nmodel.add(Dropout(0.2))\n\nmodel.add(Dense(12692, activation='softmax'))\n\nopt = tf.keras.optimizers.Adam(lr=0.001, decay=1e-6)\nmodel.compile(\n      loss='sparse_categorical_crossentropy',\n      optimizer=opt,\n      metrics=['accuracy'])\n\nprint(model.summary())\n\nhistory = model.fit(\n      train_x, train_y,\n      batch_size=64,\n      epochs=epochs,\n      validation_data=(validation_x, validation_y))\n\nscore = model.evaluate(validation_x, validation_y, verbose=0)\n<\/code><\/pre>\n\n<p>I get this model summary:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/NS4cR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/NS4cR.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Train on 131204 samples, validate on 107904 samples<\/p>\n\n<p>But after that, this error appears:<\/p>\n\n<p>ValueError: Error when checking input: expected embedding_input to have 2 dimensions, but got array with shape (131204, 5, 12692)<\/p>\n\n<p>Where is my mistake and what would be the solution?<\/p>\n","answers":[{"AnswerId":"57572225","CreationDate":"2019-08-20T11:05:03.623","ParentId":null,"OwnerUserId":"423926","Title":null,"Body":"<p>The embedding layer turns positive integers (indexes) into dense vectors of fixed size <a href=\"https:\/\/keras.io\/layers\/embeddings\/\" rel=\"nofollow noreferrer\">(Docs)<\/a>. So your <code>train_x<\/code> is not one-hot-encoded but the integer representing its index in the vocab. It will be the integer corresponding to the categorical feature.<\/p>\n\n<p><code>train_x.shape<\/code> will be <code>(No:of sample X 5)<\/code> --> Each representing the index of of the categorical feature<\/p>\n\n<p><code>train_y.shape<\/code> will be <code>(No:of sample)<\/code> --> Each representing the index of the sixth item in your time series. <\/p>\n\n<h3>Working sample<\/h3>\n\n<pre><code>import numpy as np\nimport keras\nfrom keras.layers import Embedding, LSTM, Dense\nn_samples = 100\n\ntrain_x = np.random.randint(0,12692,size=(n_samples ,5))\ntrain_y = np.random.randint(0,12692,size=(n_samples))\n\n\nmodel = keras.models.Sequential()\n\nmodel.add(Embedding(input_dim=12692+1, output_dim=250, input_length=5))\nmodel.add(LSTM(128, return_sequences=True))\nmodel.add(LSTM(32))\nmodel.add(Dense(32, activation='relu'))\nmodel.add(Dense(12692, activation='softmax'))\n\nopt = keras.optimizers.Adam(lr=0.001, decay=1e-6)\nmodel.compile(\n      loss='sparse_categorical_crossentropy',\n      optimizer=opt,\n      metrics=['accuracy'])\n\nprint(model.summary())\n\nhistory = model.fit(\n      train_x, train_y,\n      batch_size=64,\n      epochs=32)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57571185,"AnswerCount":0,"Tags":"<keras><keras-layer>","CreationDate":"2019-08-20T10:02:00.763","AcceptedAnswerId":null,"OwnerUserId":2207205.0,"Title":"How to feed a dense layer with special state vectors of RNN layer in keras","Body":"<p>I'd like to know if we can feed the dense layer with selected number of lower RNN state vectors in Keras. In other words, i'd like to know if a manual intention-based model can be implemented in Keras.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/gwwUR.png\" rel=\"nofollow noreferrer\">Illustrative image of the desired model<\/a><\/p>\n\n<p>I've created a DynamicSelection layer :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class DynamicSelection(Layer):\n\n    def __init__(self, **kwargs):\n        super(DynamicSelection, self).__init__(**kwargs)\n\n    def call(self, inputs, **kwargs):\n        results = []\n        for idx in range(focusedElems):\n            if inputs[1][idx] == -1:\n                results = results + inputs[0][idx]\n            else:\n                results = results + ([0] * wordRnnUnitNum)\n        return results\n\n    def compute_output_shape(self, input_shape):\n        return (focusedElems* wordRnnUnitNum, )\n<\/code><\/pre>\n\n<p>and used it to create my model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tokens = Input((tokenNum,))\nembeddingLayer = Embedding(vocabularySize, tokenDim, trainable=trainable)(tokens)\nrnnLayer = GRU(unitNum)(embeddingLayer)\nselectedTimeSteps = Input((setlectedtimeStepNum,))\ndynamicSelector = DynamicSelection()([rnnLayer, inputIdxs])\ndenseLayer = Dense(dense1UnitNumber,activation='relu')(dynamicSelector)\nsoftmaxLayer = Dense(4, activation='softmax')(denseLayer)\nmodel = Model(inputs=[inputIdxs, inputToken], outputs=softmaxLayer)\n<\/code><\/pre>\n\n<p>but i get the following exception:<\/p>\n\n<pre><code>dynamicSelector = DynamicSelection()([rnnLayer, inputIdxs])\n  File \"\/Users\/halsaied\/anaconda2\/lib\/python2.7\/site-packages\/keras\/engine\/base_layer.py\", line 503, in __call__\n    arguments=user_kwargs)\n  File \"\/Users\/halsaied\/anaconda2\/lib\/python2.7\/site-packages\/keras\/engine\/base_layer.py\", line 571, in _add_inbound_node\n    output_tensors[i]._keras_shape = output_shapes[i]\nAttributeError: 'int' object has no attribute '_keras_shape'\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57571337,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2019-08-20T10:10:55.377","AcceptedAnswerId":57571405.0,"OwnerUserId":5931672.0,"Title":"Combine tensorflow low level API (tensors\/placeholders) with Keras model","Body":"<p>According to <a href=\"https:\/\/www.tensorflow.org\/guide\/keras#build_advanced_models\" rel=\"nofollow noreferrer\">tensorflow<\/a>. Using <code>tf.keras.Input<\/code> gives a placeholder and using <code>tf.keras.layers.Dense<\/code> gives a Tensor. So I wanted to test the equivalence using Tensors and Placeholders with tensorflow low level API's and then use the keras high level API to train my model. This is my code:<\/p>\n\n<pre><code>tf.reset_default_graph()  \n\ninputs = tf.placeholder(tf.float32, shape=[None, 32])\nW_h = tf.get_variable(name=\"W_h\", shape=[32, 64], initializer=tf.truncated_normal_initializer(stddev=0.01))\nW_out = tf.get_variable(name=\"W_out\", shape=[64, 10], initializer=tf.truncated_normal_initializer(stddev=0.01))\n\nh = tf.nn.relu(tf.matmul(inputs, W_h, name=\"MatMul\"), name='relu')\npredictions = tf.nn.relu(tf.matmul(h, W_out, name=\"MatMul\"), name='relu')\n\nmodel = tf.keras.Model(inputs=inputs, outputs=predictions)\nmodel.compile(loss='mean_squared_error', optimizer='sgd') # sgd stands for stochastic gradient descent\nmodel.fit(x_train, y_train, batch_size=32, epochs=5)\n<\/code><\/pre>\n\n<p>However, upon call to <code>tf.keras.Model<\/code> I get the error:<\/p>\n\n<pre><code>InvalidArgumentError: You must feed a value for placeholder tensor 'Placeholder' with dtype float and shape [?,32]\n     [[{{node Placeholder}}]]\n<\/code><\/pre>\n\n<p>I do give the input a placeholder right?<\/p>\n\n<hr>\n\n<p>PS: full error message:<\/p>\n\n<pre><code>InvalidArgumentError                      Traceback (most recent call last)\n&lt;ipython-input-15-27f92d1f784d&gt; in &lt;module&gt;\n      8 predictions = tf.nn.relu(tf.matmul(h, W_out, name=\"MatMul\"), name='relu')\n      9 \n---&gt; 10 model = tf.keras.Model(inputs=inputs, outputs=predictions)\n     11 model.compile(loss='mean_squared_error', optimizer='sgd') # sgd stands for stochastic gradient descent\n     12 model.fit(x_train, y_train, batch_size=32, epochs=5)\n\n~\\.conda\\envs\\tensorflow_cpu\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py in __init__(self, *args, **kwargs)\n    127 \n    128   def __init__(self, *args, **kwargs):\n--&gt; 129     super(Model, self).__init__(*args, **kwargs)\n    130     # initializing _distribution_strategy here since it is possible to call\n    131     # predict on a model without compiling it.\n\n~\\.conda\\envs\\tensorflow_cpu\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\network.py in __init__(self, *args, **kwargs)\n    160         'inputs' in kwargs and 'outputs' in kwargs):\n    161       # Graph network\n--&gt; 162       self._init_graph_network(*args, **kwargs)\n    163     else:\n    164       # Subclassed network\n\n~\\.conda\\envs\\tensorflow_cpu\\lib\\site-packages\\tensorflow\\python\\training\\tracking\\base.py in _method_wrapper(self, *args, **kwargs)\n    455     self._self_setattr_tracking = False  # pylint: disable=protected-access\n    456     try:\n--&gt; 457       result = method(self, *args, **kwargs)\n    458     finally:\n    459       self._self_setattr_tracking = previous_value  # pylint: disable=protected-access\n\n~\\.conda\\envs\\tensorflow_cpu\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\network.py in _init_graph_network(self, inputs, outputs, name, **kwargs)\n    267 \n    268     if any(not hasattr(tensor, '_keras_history') for tensor in self.outputs):\n--&gt; 269       base_layer_utils.create_keras_history(self._nested_outputs)\n    270 \n    271     self._base_init(name=name, **kwargs)\n\n~\\.conda\\envs\\tensorflow_cpu\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer_utils.py in create_keras_history(tensors)\n    198     keras_tensors: The Tensors found that came from a Keras Layer.\n    199   \"\"\"\n--&gt; 200   _, created_layers = _create_keras_history_helper(tensors, set(), [])\n    201   return created_layers\n    202 \n\n~\\.conda\\envs\\tensorflow_cpu\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer_utils.py in _create_keras_history_helper(tensors, processed_ops, created_layers)\n    242             constants[i] = op_input\n    243           else:\n--&gt; 244             constants[i] = backend.function([], op_input)([])\n    245       processed_ops, created_layers = _create_keras_history_helper(\n    246           layer_inputs, processed_ops, created_layers)\n\n~\\.conda\\envs\\tensorflow_cpu\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py in __call__(self, inputs)\n   3290 \n   3291     fetched = self._callable_fn(*array_vals,\n-&gt; 3292                                 run_metadata=self.run_metadata)\n   3293     self._call_fetch_callbacks(fetched[-len(self._fetches):])\n   3294     output_structure = nest.pack_sequence_as(\n\n~\\.conda\\envs\\tensorflow_cpu\\lib\\site-packages\\tensorflow\\python\\client\\session.py in __call__(self, *args, **kwargs)\n   1456         ret = tf_session.TF_SessionRunCallable(self._session._session,\n   1457                                                self._handle, args,\n-&gt; 1458                                                run_metadata_ptr)\n   1459         if run_metadata:\n   1460           proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\nInvalidArgumentError: You must feed a value for placeholder tensor 'Placeholder' with dtype float and shape [?,32]\n     [[{{node Placeholder}}]]\n\n<\/code><\/pre>\n","answers":[{"AnswerId":"57571405","CreationDate":"2019-08-20T10:15:43.320","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>You can't mix <code>tf.placeholder<\/code> and <code>tf.keras.Input<\/code>. In other words, if you want to use the <code>tf.keras<\/code> API then use <code>tf.keras.Input<\/code>, or if you want to use the <code>tf<\/code> native API the go with <code>tf.placeholder<\/code>. In addition, your choice will reflect other parts of your code. Assuming that you want to go with the <code>tf.keras<\/code> API, the approach you should partake is the following:<\/p>\n\n<pre><code>class CustomLayer(tf.keras.layers.Layer):\n\n    def __init__(self, output_dim, **kwargs):\n        self.output_dim = output_dim\n        super(CustomLayer, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        # Create a trainable weight variable for this layer.\n        self.kernel = self.add_weight(name='kernel', \n                                      shape=tf.TensorShape((input_shape[1], self.output_dim)),\n                                      initializer=tf.truncated_normal_initializer(stddev=0.01),\n                                      trainable=True)\n        super(CustomLayer, self).build(input_shape)  # Be sure to call this at the end\n\n    def call(self, x):\n        return tf.nn.relu(tf.matmul(x, self.kernel))\n\n    def compute_output_shape(self, input_shape):\n        return (input_shape[0], self.output_dim)\n\ninputs = tf.keras.Input(shape=(32,), dtype=tf.float32)\nh = CustomLayer(output_dim=64)(inputs)\npredictions = CustomLayer(output_dim=10)(h)\nmodel = tf.keras.Model(inputs=inputs, outputs=predictions)\nmodel.compile(loss='mean_squared_error', optimizer='sgd') # sgd stands for stochastic gradient descent\nmodel.summary()\nmodel.fit(np.random.rand(100,32), np.random.rand(100,10), batch_size=32, epochs=5)\n<\/code><\/pre>\n\n<p>Please note that this approach is relevant if we consider that you use TF 1.14 as you mentioned in the comments. On TF 2.0, this is less complicated and more intuitive. On the other hand, if you want to stick to the TF 1.14 native API, and use <code>tf.placeholder<\/code>, then you should build the graph using <code>tf.placeholder<\/code> as nodes that you will use to feed the data. Moreover, regarding your question whether <code>tf.keras.Input<\/code> returns a placeholder - It does return a placeholder node, that you can use to feed your data. But it does not return a <code>tf.placeholder<\/code>. The usage of <code>tf.placeholder<\/code> is the following:<\/p>\n\n<pre><code>X = tf.placeholder(shape=None, dtype=tf.int32)\nY = tf.placeholder(shape=None, dtype=tf.int32)\nadd = X + Y\nwith tf.Session() as sess:\n    print(sess.run(add, feed_dict={X: 2, Y: 3}))\n    # 5\n    print(sess.run(add, feed_dict={X: 10, Y: 9}))\n    # 19\n<\/code><\/pre>\n\n<p>As you can see, a static graph is created, and then its nodes are executed with a <code>tf.Session<\/code> while feeding data in the graph using the <code>tf.placeholder<\/code>. On the other hand, <code>tf.keras.Input<\/code> serves the same purpose (That is why in the documentation is called a placeholder), but its use case is relevant with the <code>tf.keras<\/code> API and not with the <code>tf<\/code> native API.<\/p>\n"}]}
+{"QuestionId":57571648,"AnswerCount":3,"Tags":"<python><keras><neural-network><transfer-learning>","CreationDate":"2019-08-20T10:29:54.233","AcceptedAnswerId":null,"OwnerUserId":11786822.0,"Title":"Transfer learning with Keras: Input shape mismatch","Body":"<p>I'm running a classification and predition neural network algorithme using pre-trained model with keras. \nNow I know the shape of the input for keras is <code>(224,224,3)<\/code> but my input has this shape <code>(180, 200, 20)<\/code> and I get the following error:<\/p>\n\n<p><code>ValueError: Dimension 0 in both shapes must be equal, but are 3 and 64. Shapes are [3,3,20,64] and [64,3,3,3]. for 'Assign_32' (op: 'Assign') with input shapes: [3,3,20,64], [64,3,3,3].\n<\/code><\/p>\n\n<p>and here is the code:<\/p>\n\n<pre><code>from keras import applications\nfrom keras.layers import Input\n\ninput_tensor = Input(shape = (180, 200, 20))\nvgg_model = applications.VGG16(weights = 'imagenet', include_top = False, input_tensor = input_tensor)\nvgg_model.summary()\n<\/code><\/pre>\n\n<p>Any idea how to get around this? Thank you<\/p>\n","answers":[{"AnswerId":"57576091","CreationDate":"2019-08-20T14:40:17.910","ParentId":null,"OwnerUserId":"10754197","Title":null,"Body":"<p>You need to resize your input image <\/p>\n\n<pre><code>from keras.preprocessing import image\nimg = image.load_img(\"image1.jpeg\",target_size=(224,224))\n<\/code><\/pre>\n\n<p>If you want to learn to do transfer learning from scratch in keras you can read this article. This article has step by step implementation.\n<a href=\"https:\/\/medium.com\/@1297rohit\/transfer-learning-from-scratch-using-keras-339834b153b9\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@1297rohit\/transfer-learning-from-scratch-using-keras-339834b153b9<\/a><\/p>\n"},{"AnswerId":"57614304","CreationDate":"2019-08-22T17:23:10.080","ParentId":null,"OwnerUserId":"4332585","Title":null,"Body":"<p>In your case, since you are not dealing with images of the right size (or number of channels) you may want to cut out large parts of the vgg network to still save the information contained in the middle layers, but I am not sure how efficient it would be.<\/p>\n\n<p>You would need to remove the first convolution layer, and all the dense layers at the end, replacing them with your own layers. You would certainly need to retrain the whole network, so rather than transfer learning you would be doing very smart initialization.<\/p>\n"},{"AnswerId":"57571907","CreationDate":"2019-08-20T10:46:07.373","ParentId":null,"OwnerUserId":"5563864","Title":null,"Body":"<p>From <a href=\"https:\/\/keras.io\/applications\/#vgg16\" rel=\"nofollow noreferrer\">Documentation<\/a>:<\/p>\n\n<blockquote>\n  <p>input_shape: optional shape tuple, only to be specified if include_top\n  is False (otherwise the input shape has to be (224, 224, 3) (with\n  'channels_last' data format) or (3, 224, 224) (with 'channels_first'\n  data format). It should have exactly 3 inputs channels, and width and\n  height should be no smaller than 32. E.g. (200, 200, 3) would be one\n  valid value.<\/p>\n<\/blockquote>\n\n<p>You can try to create a vgg16 from scratch from this link. <a href=\"https:\/\/gist.github.com\/baraldilorenzo\/07d7802847aaad0a35d3\" rel=\"nofollow noreferrer\">VGG16 model for Keras<\/a><\/p>\n"}]}
+{"QuestionId":57572079,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorflow-probability>","CreationDate":"2019-08-20T10:56:09.867","AcceptedAnswerId":null,"OwnerUserId":10897543.0,"Title":"\"Truth Value of Array is ambiguous\" Error when using validation Data in Keras","Body":"<p>I am modifying the <a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/probability\/blob\/master\/tensorflow_probability\/examples\/jupyter_notebooks\/Probabilistic_Layers_VAE.ipynb#scrollTo=kZ0MdF1j8WJf\" rel=\"nofollow noreferrer\">Variational Autoencoder<\/a> Example for Tensorflow Probability for predicting time series.<\/p>\n\n<p>I am chopping a 100000 element long time series into 100 blocks of length 1000, and feeding them into the keras model. However when I do the same thing for the validation_data argument, I get the error mentioned in the title.<\/p>\n\n<p>I have tried using validation_split = 0.5, this gives a different error (\"Arguments and signature arguments do not match\"). I have reduced the model to the smallest size that still gives the error below.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>NormalizedSamples = tf.reshape(NormalizedSamples,(1000,100))\nNormalizedSamples2 = tf.reshape(NormalizedSamples,(1000,100))\n\n#Setting Model Parameters\nencoded_size =  50\nbase_depth = 25\n\ninput_shape = (100,1)\n\nprior = tfd.Independent(tfd.Normal(loc = tf.zeros(encoded_size),scale = 1),\n                        reinterpreted_batch_ndims=1)\n\nencoder = tfk.Sequential([\n        tfkl.InputLayer(input_shape = input_shape,name = 'encoder_input'),\n\n        tfkl.Dense(tfpl.MultivariateNormalTriL.params_size(encoded_size),\n                   activation=None),\n\n        tfpl.MultivariateNormalTriL(\n            encoded_size,\n            activity_regularizer=tfpl.KLDivergenceRegularizer(prior)),        \n        ],name=\"Encoder\")\n\n\nvae = tfk.Model(inputs=encoder.inputs,\n                outputs = encoder.outputs[0])\n\n\n\nnegloglik = lambda x, rv_x: -rv_x.log_prob(x)\n\nvae.compile(optimizer=tf.optimizers.Adam(learning_rate=1e-3),\n            loss=negloglik)\n\nvae.fit(tf.expand_dims(NormalizedSamples,2),\n        epochs=15,\n        validation_data = tf.expand_dims(NormalizedSamples2,2))\n\n<\/code><\/pre>\n","answers":[{"AnswerId":"57597995","CreationDate":"2019-08-21T19:20:04.037","ParentId":null,"OwnerUserId":"9946929","Title":null,"Body":"<p>What versions are you using of TF and TFP? <code>print(tf.__version__, tfp.__version__)<\/code><\/p>\n"}]}
+{"QuestionId":57572281,"AnswerCount":0,"Tags":"<python><c++11><tensorflow>","CreationDate":"2019-08-20T11:07:58.720","AcceptedAnswerId":null,"OwnerUserId":11950904.0,"Title":"TensorFlow's Conv2D dot product implementation?","Body":"<p>I would like to run a Python code of NN which is using Conv2D, but i'd like to change the dot product inside the conv2D operation in C++. To be more specific about the background, conv operator should contain product between the filter kernel and a window to get a filtered window (for each window within image). I've already searched in Tensorflow's conv_ops.cc documentation, but didn't find any dot product (or any product). I'd like to be referenced to the place where the actual dot product happens. Thanks!<\/p>\n","answers":[]}
+{"QuestionId":57572367,"AnswerCount":1,"Tags":"<image-processing><keras>","CreationDate":"2019-08-20T11:13:03.000","AcceptedAnswerId":null,"OwnerUserId":7085243.0,"Title":"Why I have good val_acc during training, but always wrong manul prediction on the same images","Body":"<p>I trained my model of CNN net on images with good <code>val_acc=0.97<\/code> and using <code>model.fit_generator<\/code>. <\/p>\n\n<p>Here is the output of last epoch, proofing high validation accuracy:<\/p>\n\n<pre><code>199\/200 [============================&gt;.] - ETA: 1s - loss: 0.1563 - acc: 0.9563\n200\/200 [==============================] - 306s 2s\/step - loss: 0.1556 - acc: 0.9565 - val_loss: 0.1402 - val_acc: 0.9691\n\nEpoch 00005: val_acc improved from 0.96701 to 0.96907, saving model to \/home\/sergorl\/cars\/color_weights.hdf5\n\n<\/code><\/pre>\n\n<p>But when I use the same validation data set, which I use during training, but test only one image and for every image in my validation set I always get the wrong predicted label and the predicted probabilities looks like a uniform distribution.<\/p>\n\n<p>I read this links:\n<a href=\"https:\/\/stackoverflow.com\/questions\/48239431\/wrong-prediction-on-images\">Wrong prediction on images<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/54509928\/why-is-keras-training-well-but-returning-wrong-predictions\">Why is Keras training well but returning wrong predictions?<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/48112219\/keras-val-acc-is-good-but-prediction-for-same-data-is-poor\">Keras Val_acc is good but prediction for same data is poor<\/a><\/p>\n\n<p>But I don't find the solution!<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nfrom keras.models import Sequential,Model,load_model\nfrom keras.optimizers import SGD\nfrom keras.layers import BatchNormalization, Lambda, Input, Dense, Convolution2D, MaxPooling2D, AveragePooling2D, ZeroPadding2D, Dropout, Flatten, merge, Reshape, Activation\nfrom keras.layers.merge import Concatenate\nfrom keras.preprocessing.image import ImageDataGenerator\nfrom keras.callbacks import ModelCheckpoint\nimport os\nimport cv2\nimport numpy as np\n\n\nclass CarColorNet:\n\n    def __init__(self, numClasses=6, imageWidth=256, imageHeight=256):\n\n        self.classes = {}\n        self.numClasses = numClasses\n        self.imageWidth = imageWidth\n        self.imageHeight = imageHeight\n\n        input_image = Input(shape=(self.imageWidth, self.imageHeight, 3))\n\n        # ------------------------------------ TOP BRANCH ------------------------------------\n        # first top convolution layer\n        top_conv1 = Convolution2D(filters=48, kernel_size=(11, 11), strides=(4, 4),\n                                  input_shape=(self.imageWidth, self.imageHeight, 3), activation='relu')(input_image)\n        top_conv1 = BatchNormalization()(top_conv1)\n        top_conv1 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(top_conv1)\n\n        # second top convolution layer\n        # split feature map by half\n        top_top_conv2 = Lambda(lambda x: x[:, :, :, :24])(top_conv1)\n        top_bot_conv2 = Lambda(lambda x: x[:, :, :, 24:])(top_conv1)\n\n        top_top_conv2 = Convolution2D(filters=64, kernel_size=(3, 3), strides=(1, 1), activation='relu',\n                                      padding='same')(top_top_conv2)\n        top_top_conv2 = BatchNormalization()(top_top_conv2)\n        top_top_conv2 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(top_top_conv2)\n\n        top_bot_conv2 = Convolution2D(filters=64, kernel_size=(3, 3), strides=(1, 1), activation='relu',\n                                      padding='same')(top_bot_conv2)\n        top_bot_conv2 = BatchNormalization()(top_bot_conv2)\n        top_bot_conv2 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(top_bot_conv2)\n\n        # third top convolution layer\n        # concat 2 feature map\n        top_conv3 = Concatenate()([top_top_conv2, top_bot_conv2])\n        top_conv3 = Convolution2D(filters=192, kernel_size=(3, 3), strides=(1, 1), activation='relu',\n                                  padding='same')(top_conv3)\n\n        # fourth top convolution layer\n        # split feature map by half\n        top_top_conv4 = Lambda(lambda x: x[:, :, :, :96])(top_conv3)\n        top_bot_conv4 = Lambda(lambda x: x[:, :, :, 96:])(top_conv3)\n\n        top_top_conv4 = Convolution2D(filters=96, kernel_size=(3, 3), strides=(1, 1), activation='relu',\n                                      padding='same')(top_top_conv4)\n        top_bot_conv4 = Convolution2D(filters=96, kernel_size=(3, 3), strides=(1, 1), activation='relu',\n                                      padding='same')(top_bot_conv4)\n\n        # fifth top convolution layer\n        top_top_conv5 = Convolution2D(filters=64, kernel_size=(3, 3), strides=(1, 1), activation='relu',\n                                      padding='same')(top_top_conv4)\n        top_top_conv5 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(top_top_conv5)\n\n        top_bot_conv5 = Convolution2D(filters=64, kernel_size=(3, 3), strides=(1, 1), activation='relu',\n                                      padding='same')(top_bot_conv4)\n        top_bot_conv5 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(top_bot_conv5)\n\n        # ------------------------------------ TOP BOTTOM ------------------------------------\n        # first bottom convolution layer\n        bottom_conv1 = Convolution2D(filters=48, kernel_size=(11, 11), strides=(4, 4),\n                                     input_shape=(224, 224, 3), activation='relu')(input_image)\n        bottom_conv1 = BatchNormalization()(bottom_conv1)\n        bottom_conv1 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(bottom_conv1)\n\n        # second bottom convolution layer\n        # split feature map by half\n        bottom_top_conv2 = Lambda(lambda x: x[:, :, :, :24])(bottom_conv1)\n        bottom_bot_conv2 = Lambda(lambda x: x[:, :, :, 24:])(bottom_conv1)\n\n        bottom_top_conv2 = Convolution2D(filters=64, kernel_size=(3, 3), strides=(1, 1), activation='relu',\n                                         padding='same')(bottom_top_conv2)\n        bottom_top_conv2 = BatchNormalization()(bottom_top_conv2)\n        bottom_top_conv2 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(bottom_top_conv2)\n\n        bottom_bot_conv2 = Convolution2D(filters=64, kernel_size=(3, 3), strides=(1, 1), activation='relu',\n                                         padding='same')(bottom_bot_conv2)\n        bottom_bot_conv2 = BatchNormalization()(bottom_bot_conv2)\n        bottom_bot_conv2 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(bottom_bot_conv2)\n\n        # third bottom convolution layer\n        # concat 2 feature map\n        bottom_conv3 = Concatenate()([bottom_top_conv2, bottom_bot_conv2])\n        bottom_conv3 = Convolution2D(filters=192, kernel_size=(3, 3), strides=(1, 1), activation='relu',\n                                     padding='same')(bottom_conv3)\n\n        # fourth bottom convolution layer\n        # split feature map by half\n        bottom_top_conv4 = Lambda(lambda x: x[:, :, :, :96])(bottom_conv3)\n        bottom_bot_conv4 = Lambda(lambda x: x[:, :, :, 96:])(bottom_conv3)\n\n        bottom_top_conv4 = Convolution2D(filters=96, kernel_size=(3, 3), strides=(1, 1), activation='relu',\n                                         padding='same')(bottom_top_conv4)\n        bottom_bot_conv4 = Convolution2D(filters=96, kernel_size=(3, 3), strides=(1, 1), activation='relu',\n                                         padding='same')(bottom_bot_conv4)\n\n        # fifth bottom convolution layer\n        bottom_top_conv5 = Convolution2D(filters=64, kernel_size=(3, 3), strides=(1, 1), activation='relu',\n                                         padding='same')(bottom_top_conv4)\n        bottom_top_conv5 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(bottom_top_conv5)\n\n        bottom_bot_conv5 = Convolution2D(filters=64, kernel_size=(3, 3), strides=(1, 1), activation='relu',\n                                         padding='same')(bottom_bot_conv4)\n        bottom_bot_conv5 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(bottom_bot_conv5)\n\n        # ---------------------------------- CONCATENATE TOP AND BOTTOM BRANCH ------------------------------------\n        conv_output = Concatenate()([top_top_conv5, top_bot_conv5, bottom_top_conv5, bottom_bot_conv5])\n\n        # Flatten\n        flatten = Flatten()(conv_output)\n\n        # Fully-connected layer\n        FC_1 = Dense(units=4096, activation='relu')(flatten)\n        FC_1 = Dropout(0.6)(FC_1)\n        FC_2 = Dense(units=4096, activation='relu')(FC_1)\n        FC_2 = Dropout(0.6)(FC_2)\n        output = Dense(units=self.numClasses, activation='softmax')(FC_2)\n\n        self.model = Model(inputs=input_image, outputs=output)\n        sgd = SGD(lr=1e-3, decay=1e-6, momentum=0.9, nesterov=True)\n        # sgd = SGD(lr=0.01, momentum=0.9, decay=0.0005, nesterov=True)\n        self.model.compile(optimizer=sgd, loss='categorical_crossentropy', metrics=['accuracy'])\n\n    def train(self,\n              pathToTrainSet,\n              pathToValidSet,\n              pathToSaveModel,\n              epochs=7,\n              batchSize=32,\n              stepsPerEpoch=200,\n              validationSteps=1000):\n\n        fileOfWeights = 'color_weights.hdf5'\n        checkpoint = ModelCheckpoint(os.path.join(pathToSaveModel, fileOfWeights),\n                                     monitor='val_acc', verbose=1,\n                                     save_best_only=True, mode='max')\n\n        checkpoint2 = ModelCheckpoint(os.path.join(pathToSaveModel, fileOfWeights),\n                                     monitor='val_loss', verbose=1,\n                                     save_best_only=True, mode='max')\n\n        trainDataGen = ImageDataGenerator(rescale=1.0\/255, shear_range=0.2,\n                                          zoom_range=0.3, horizontal_flip=True)\n\n        validDataGen = ImageDataGenerator(rescale=1.0\/255)\n\n        trainSet = trainDataGen.flow_from_directory(\n                pathToTrainSet,\n                target_size=(self.imageWidth, self.imageHeight),\n                batch_size=batchSize,\n                class_mode='categorical'\n        )\n\n        self.classes = {v: k for k, v in trainSet.class_indices.items()}\n\n        np.save(os.path.join(pathToSaveModel, 'class_index.npy'), self.classes)\n\n        validSet = validDataGen.flow_from_directory(\n                pathToValidSet,\n                target_size=(self.imageWidth, self.imageHeight),\n                batch_size=batchSize,\n                class_mode='categorical'\n        )\n\n        self.model.fit_generator(\n            trainSet,\n            steps_per_epoch=stepsPerEpoch,\n            epochs=epochs,\n            validation_data=validSet,\n            validation_steps=validationSteps\/\/batchSize,\n            callbacks=[checkpoint, checkpoint2])\n\n        print('============================ Saving is here ============================')\n        self.model.save(os.path.join(pathToSaveModel, 'car_color_net.h5'))\n\n    @staticmethod\n    def load(pathToModel, pathToClassIndexes):\n\n        model = load_model(pathToModel)\n\n        layers = model.layers\n        inputShape, outputShape = layers[0].input_shape, layers[-1].output_shape,\n\n        imageWidth, imageHeight = inputShape[1], inputShape[2]\n        numClasses = outputShape[1]\n\n        net = CarColorNet(numClasses, imageWidth, imageHeight)\n        net.classes = np.load(os.path.join(pathToClassIndexes, 'class_index.npy')).item()\n\n        return net\n\n    def predictOneImage(self, pathToImage):\n\n        frame = cv2.imread(pathToImage)\n        frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)\n        frame = cv2.resize(frame, (self.imageWidth, self.imageHeight))\n\n        frame = np.expand_dims(frame, axis=0)\n\n        # cv2.imshow(\"boxed\", frame[0, :, :, :])\n        # cv2.waitKey(0)\n\n        frame = np.asarray(frame, dtype='float32')\n        img = frame\/255\n\n        probs = self.model.predict(img)\n        ind = probs.argmax(axis=-1)[0]\n\n        return self.classes[ind]\n\n\nif __name__ == '__main__':\n\n    pathToTrainSet = '\/home\/sergorl\/cars\/train'\n    pathToValidSet = '\/home\/sergorl\/cars\/valid'\n    pathToSaveModel = '\/home\/sergorl\/cars'\n\n    ## Train net\n    # net = CarColorNet(numClasses=6)\n    # net.train(pathToTrainSet, pathToValidSet, pathToSaveModel)\n\n    # Test on all images from validSet\n    net = CarColorNet.load(os.path.join(pathToSaveModel, 'car_color_net.h5'), pathToSaveModel)\n\n    count, countTrueLabels = 0, 0\n\n    for dirpath, _dirnames, filenames in os.walk(pathToValidSet):\n        trueLabel = dirpath.split('\/')[-1]\n\n        for file in filenames:\n\n            label = net.predictOneImage(os.path.join(dirpath, file))\n\n            print(trueLabel, label)\n\n            if label == trueLabel:\n                countTrueLabels += 1\n\n            count += 1\n\n    print('rate is {0:.2f}'.format(float(countTrueLabels) \/ float(count) * 100))\n\n\n<\/code><\/pre>\n\n<p>If I have a good <code>val_acc=0.97<\/code>, I'll expect the same result (or nearly), testing every image in validation set. But always have wrong prediction!<\/p>\n\n<p>I ran net immediately after train was done and see that learning was good:<\/p>\n\n<pre><code>if __name__ == '__main__':\n\n    pathToTrainSet = '\/home\/sergorl\/cars\/train'\n    pathToValidSet = '\/home\/sergorl\/cars\/valid'\n    pathToSaveModel = '\/home\/sergorl\/cars'\n\n    # Train net\n    net = CarColorNet(numClasses=6)\n    net.train(pathToTrainSet, pathToValidSet, pathToSaveModel)\n\n    # Test on all images from validSet\n    count, countTrueLabels = 0, 0\n\n    for dirpath, _dirnames, filenames in os.walk(pathToValidSet):\n        trueLabel = dirpath.split('\/')[-1]\n\n        for file in filenames:\n\n            label = net.predictOneImage(os.path.join(dirpath, file))\n\n            print(trueLabel, label)\n\n            if label == trueLabel:\n                countTrueLabels += 1\n\n            count += 1\n\n    print('rate is {0:.2f}'.format(float(countTrueLabels) \/ float(count) * 100))\n<\/code><\/pre>\n\n<p>So it seems the problem is inside <code>model.save<\/code> and it looks like saving doesn't work!. I found many related issues on git, for example:<\/p>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/4875\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/4875<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/4904\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/4904<\/a><\/p>\n\n<p>But I don't know how to fix it with <code>Python 3.7.3<\/code> and <code>keras 2.0.0<\/code> <\/p>\n","answers":[{"AnswerId":"57575717","CreationDate":"2019-08-20T14:19:55.333","ParentId":null,"OwnerUserId":"10754197","Title":null,"Body":"<p>Can you share more about the issue like what is the output you are getting. From the code i can see that you are training for 6 classes and using categorical cross entropy so ideally you should be getting an array with 6 values with each value bw 0 and 1 and the index of highest value in that array should be the output. <\/p>\n"}]}
+{"QuestionId":57572404,"AnswerCount":0,"Tags":"<image><tensorflow><tensorflow-datasets>","CreationDate":"2019-08-20T11:15:12.630","AcceptedAnswerId":null,"OwnerUserId":2177290.0,"Title":"tf.data.Dataset: create a dataset from files using `list_file` and keeping the order","Body":"<p>I have a data set of 70000 images that have numbers from 0 to 69999 as their names. Also, I have a text file that has the same number of lines which includes the corresponding labels for the images.<\/p>\n\n<p>I tried to build a TensorFlow data set from the files using <code>tf.data.Dataset.list_file<\/code> like<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Make a Dataset of file names including all the JPG images files in\n# the relative image directory.\nfilename_dataset = tf.data.Dataset.list_files(\"data\/train\/*.jpg\", shuffle=False)\n\n# Make a Dataset of image tensors by reading and decoding the files.\nimage_dataset = filename_dataset.map(lambda x: tf.io.decode_image(tf.io.read_file(x)))\n<\/code><\/pre>\n\n<p>Now, the problem is that the images in <code>image_dataset<\/code> are (forgive me, I don't know the technical term) in order with respect to their names as a string; meaning a sequence like '0, 1, 10, 100, ...' instead of '1, 2, 3, ...'. This way, I can't use the line numbers of the text file to get the label for each image.<\/p>\n\n<p>Does this mean that I can't use <code>list_file<\/code> to create the data set? Or is there a work around? I would appreciate any hints on a nice way to do this job.<\/p>\n","answers":[]}
+{"QuestionId":57572973,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch>","CreationDate":"2019-08-20T11:47:47.240","AcceptedAnswerId":null,"OwnerUserId":9220672.0,"Title":"Visualize DGLGraph based model","Body":"<p>I'm working on a graph convolutional model taking <code>DGLGraph<\/code> as input. Is there any method to visualize the model structure?<\/p>\n\n<p>I found libraries such as <code>torchsummary<\/code> and <code>visualdl<\/code> requires the input to be <code>pytorch tensors<\/code> or <code>python<\/code> data types.<\/p>\n","answers":[{"AnswerId":"57574060","CreationDate":"2019-08-20T12:48:09.190","ParentId":null,"OwnerUserId":"8185618","Title":null,"Body":"<p>You could use multiple solution:<\/p>\n\n<p>1.netron(Visualizer for neural network, deep learning and machine learning models):<\/p>\n\n<p><a href=\"http:\/\/%20%20%20%20https:\/\/github.com\/lutzroeder\/netron\" rel=\"nofollow noreferrer\">netron<\/a><\/p>\n\n<p>2.ANNvisualizer:\n<a href=\"https:\/\/towardsdatascience.com\/visualizing-artificial-neural-networks-anns-with-just-one-line-of-code-b4233607209e\" rel=\"nofollow noreferrer\">Visualizing Artificial Neural Networks (ANNs) with just One Line of Code\n<\/a><\/p>\n\n<p>3.NN-SVG:\n<a href=\"http:\/\/alexlenail.me\/NN-SVG\/index.html\" rel=\"nofollow noreferrer\">NN-SVG<\/a><\/p>\n\n<p>4.dotnets\n<a href=\"https:\/\/github.com\/martisak\/dotnets\" rel=\"nofollow noreferrer\">dotnets<\/a>\n<a href=\"https:\/\/tgmstat.wordpress.com\/2013\/06\/12\/draw-neural-network-diagrams-graphviz\/\" rel=\"nofollow noreferrer\">enter link description here<\/a><\/p>\n\n<p>and for more information please refer to this post <a href=\"https:\/\/datascience.stackexchange.com\/questions\/12851\/how-do-you-visualize-neural-network-architectures\">How do you visualize neural network architectures?\n<\/a><\/p>\n"}]}
+{"QuestionId":57573338,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-20T12:09:27.860","AcceptedAnswerId":57575795.0,"OwnerUserId":7253901.0,"Title":"Weird output when using TensorFlow backend on Keras","Body":"<p>Im using Keras with a TensorFlow backend to train a neural networks model. It is a simple model (from a tutorial) and everything works fine. Model predicts the way it should. Only tensorflow seems to be outputting information, and I don't know if I can safely ignore it\/what it means. Can somebody explain me what the following output tells us? Especially all the lines which start with <code>OMP<\/code> are unclear to me. The deprecation warning does not really worry me.<\/p>\n\n<pre><code>Using TensorFlow backend.\nOMP: Info #212: KMP_AFFINITY: decoding x2APIC ids.\nOMP: Info #210: KMP_AFFINITY: Affinity capable, using global cpuid leaf 11 info\nOMP: Info #154: KMP_AFFINITY: Initial OS proc set respected: 0-7\nOMP: Info #156: KMP_AFFINITY: 8 available OS procs\nOMP: Info #157: KMP_AFFINITY: Uniform topology\nOMP: Info #179: KMP_AFFINITY: 1 packages x 4 cores\/pkg x 2 threads\/core (4 total cores)\nOMP: Info #214: KMP_AFFINITY: OS proc to physical thread map:\nOMP: Info #171: KMP_AFFINITY: OS proc 0 maps to package 0 core 0 thread 0 \nOMP: Info #171: KMP_AFFINITY: OS proc 1 maps to package 0 core 0 thread 1 \nOMP: Info #171: KMP_AFFINITY: OS proc 2 maps to package 0 core 1 thread 0 \nOMP: Info #171: KMP_AFFINITY: OS proc 3 maps to package 0 core 1 thread 1 \nOMP: Info #171: KMP_AFFINITY: OS proc 4 maps to package 0 core 2 thread 0 \nOMP: Info #171: KMP_AFFINITY: OS proc 5 maps to package 0 core 2 thread 1 \nOMP: Info #171: KMP_AFFINITY: OS proc 6 maps to package 0 core 3 thread 0 \nOMP: Info #171: KMP_AFFINITY: OS proc 7 maps to package 0 core 3 thread 1 \nOMP: Info #250: KMP_AFFINITY: pid 14400 tid 5252 thread 0 bound to OS proc set 0\nWARNING: Logging before flag parsing goes to stderr.\nW0820 14:05:56.256548  5252 deprecation.py:323] From C:\\Users\\Gebruiker\\Anaconda3\\envs\\Anaconda3.7\\lib\\site-packages\\tensorflow\\python\\ops\\nn_impl.py:180: add_dispatch_support.&lt;locals&gt;.wrapper (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse tf.where in 2.0, which has the same broadcast rule as np.where\n2019-08-20 14:05:56.449349: I tensorflow\/core\/platform\/cpu_feature_guard.cc:145] This TensorFlow binary is optimized with Intel(R) MKL-DNN to use the following CPU instructions in performance critical operations:  AVX AVX2\nTo enable them in non-MKL-DNN operations, rebuild TensorFlow with the appropriate compiler flags.\n2019-08-20 14:05:56.451182: I tensorflow\/core\/common_runtime\/process_util.cc:115] Creating new thread pool with default inter op setting: 8. Tune using inter_op_parallelism_threads for best performance.\nOMP: Info #250: KMP_AFFINITY: pid 14400 tid 13256 thread 1 bound to OS proc set 2\nOMP: Info #250: KMP_AFFINITY: pid 14400 tid 5064 thread 2 bound to OS proc set 4\nOMP: Info #250: KMP_AFFINITY: pid 14400 tid 8904 thread 3 bound to OS proc set 6\nOMP: Info #250: KMP_AFFINITY: pid 14400 tid 9408 thread 4 bound to OS proc set 1\nOMP: Info #250: KMP_AFFINITY: pid 14400 tid 14360 thread 5 bound to OS proc set 3\nOMP: Info #250: KMP_AFFINITY: pid 14400 tid 14616 thread 6 bound to OS proc set 5\nOMP: Info #250: KMP_AFFINITY: pid 14400 tid 2692 thread 7 bound to OS proc set 7\nOMP: Info #250: KMP_AFFINITY: pid 14400 tid 14172 thread 8 bound to OS proc set 0\n<\/code><\/pre>\n","answers":[{"AnswerId":"57575795","CreationDate":"2019-08-20T14:23:40.263","ParentId":null,"OwnerUserId":"7981821","Title":null,"Body":"<p>OpenMP is binding threads to your cpu to run. It is just logistical output I wouldn't worry about it. You can find similar output documented <a href=\"https:\/\/software.intel.com\/en-us\/articles\/using-kmp-affinity-to-create-openmp-thread-mapping-to-os-proc-ids\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<blockquote>\n  <p>The Intel\u00ae Compiler's OpenMP* runtime library has ability to bind OpenMP* threads to physical processing units.<\/p>\n  \n  <p>This article will show you how to use KMP_AFFINITY environment\n  variable extensions from the Intel Compiler, the high-level affinity\n  interface, to determine the machine topology and assigns OpenMP*\n  threads to the processors based upon their physical location in the\n  machine.<\/p>\n  \n  <p>Example 1:<\/p>\n  \n  <p>You have one quad core system with Intel\u00ae Hyper-Threading Technology\n  (Intel\u00aeHT Technology) enabled.  By default, Intel compiler OpenMP*\n  runtime libraries will create 8 threads, running freely on 8 logical\n  processors provided by the operating system.<\/p>\n<\/blockquote>\n\n<p>As described above in the article based on your physical CPU configuration the threads are assigned to different cores.<\/p>\n"}]}
+{"QuestionId":57573540,"AnswerCount":1,"Tags":"<python><tensorflow><python-multithreading><tensorflow-agents>","CreationDate":"2019-08-20T12:20:18.633","AcceptedAnswerId":57661499.0,"OwnerUserId":3083470.0,"Title":"Tf-Agents ParallelPyEnvironment fails silently","Body":"<p>I have written a custom environment so I can play around with reinforcement learning (PPO) and tf-agents.\nThis works fine if I wrap my env ( which inherits from py_environment.PyEnvironment) in a <code>TfPyEnvironment<\/code>, but fails if I try to wrap it into a <code>ParallelPyEnvironment<\/code>. I have tried playing around with all the keyword arguments of <code>ParallelPyEnvironment<\/code> but the code just runs up to the line and then nothing happens - no Exception, the program does not terminate etc.<\/p>\n\n<p>Here is my code initialising the environment and showing off the working variant for the <code>eval_env<\/code>:<\/p>\n\n<pre><code>train_env = tf_py_environment.TFPyEnvironment(\n    ParallelPyEnvironment(\n        [CardGameEnv()] * hparams['parallel_environments']\n    )\n)\n# this works perfectly:\neval_env = tf_py_environment.TFPyEnvironment(CardGameEnv(debug=True))\n<\/code><\/pre>\n\n<p>If I terminate the script via <code>CTRL+C<\/code>, this is what is being output:<\/p>\n\n<pre><code>Traceback (most recent call last):\nTraceback (most recent call last):\n  File \"E:\\Users\\tmp\\Documents\\Programming\\Neural Nets\\Poker_AI\\poker_logic\\train.py\", line 229, in &lt;module&gt;\n  File \"&lt;string&gt;\", line 1, in &lt;module&gt;\n    train(model_num=3)\n  File \"C:\\Python37\\lib\\multiprocessing\\spawn.py\", line 105, in spawn_main\n  File \"E:\\Users\\tmp\\Documents\\Programming\\Neural Nets\\Poker_AI\\poker_logic\\train.py\", line 64, in train\n    [CardGameEnv()] * hparams['parallel_environments']\n    exitcode = _main(fd)\n  File \"E:\\Users\\tmp\\AppData\\Roaming\\Python\\Python37\\site-packages\\gin\\config.py\", line 1009, in wrapper\n  File \"C:\\Python37\\lib\\multiprocessing\\spawn.py\", line 113, in _main\n    preparation_data = reduction.pickle.load(from_parent)\nKeyboardInterrupt\n    return fn(*new_args, **new_kwargs)\n  File \"C:\\Python37\\lib\\site-packages\\tf_agents\\environments\\parallel_py_environment.py\", line 70, in __init__\n    self.start()\n  File \"C:\\Python37\\lib\\site-packages\\tf_agents\\environments\\parallel_py_environment.py\", line 83, in start\n    env.start(wait_to_start=self._start_serially)\n  File \"C:\\Python37\\lib\\site-packages\\tf_agents\\environments\\parallel_py_environment.py\", line 223, in start\n    self._process.start()\n  File \"C:\\Python37\\lib\\multiprocessing\\process.py\", line 112, in start\n    self._popen = self._Popen(self)\n  File \"C:\\Python37\\lib\\multiprocessing\\context.py\", line 223, in _Popen\n    return _default_context.get_context().Process._Popen(process_obj)\n  File \"C:\\Python37\\lib\\multiprocessing\\context.py\", line 322, in _Popen\n    return Popen(process_obj)\n  File \"C:\\Python37\\lib\\multiprocessing\\popen_spawn_win32.py\", line 65, in __init__\n    reduction.dump(process_obj, to_child)\n  File \"C:\\Python37\\lib\\multiprocessing\\reduction.py\", line 60, in dump\n    ForkingPickler(file, protocol).dump(obj)\n  File \"C:\\Python37\\lib\\site-packages\\tf_agents\\environments\\parallel_py_environment.py\", line 264, in __getattr__\n    return self._receive()\n  File \"C:\\Python37\\lib\\site-packages\\tf_agents\\environments\\parallel_py_environment.py\", line 333, in _receive\n    message, payload = self._conn.recv()\n  File \"C:\\Python37\\lib\\multiprocessing\\connection.py\", line 250, in recv\n    buf = self._recv_bytes()\n  File \"C:\\Python37\\lib\\multiprocessing\\connection.py\", line 306, in _recv_bytes\n    [ov.event], False, INFINITE)\nKeyboardInterrupt\nError in atexit._run_exitfuncs:\nTraceback (most recent call last):\n  File \"C:\\Python37\\lib\\site-packages\\tf_agents\\environments\\parallel_py_environment.py\", line 289, in close\n    self._process.join(5)\n  File \"C:\\Python37\\lib\\multiprocessing\\process.py\", line 139, in join\n    assert self._popen is not None, 'can only join a started process'\nAssertionError: can only join a started process\n<\/code><\/pre>\n\n<p>From that I conclude that the thread <code>ParallelPyEnvironment<\/code> is trying to start does not do that, but since I'm not very experienced with threading in Python, I have no idea where to go from here, especially how to fix this.\nCurrent training takes a long time and does not use my PC's capabilities at all (3GB of 32GB RAM used, processor at 3%, GPU barely working at all but VRAM full), so this should speed up training time significantly.<\/p>\n","answers":[{"AnswerId":"57661499","CreationDate":"2019-08-26T16:16:19.773","ParentId":null,"OwnerUserId":"3083470","Title":null,"Body":"<p>The solution is to pass in callables, not environments, so the <code>ParallelPyEnvironment<\/code> can construct them itself:<\/p>\n\n<pre><code>train_env = tf_py_environment.TFPyEnvironment(\n    ParallelPyEnvironment(\n        [CardGameEnv] * hparams['parallel_environments']\n    )\n)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57573624,"AnswerCount":3,"Tags":"<python><opencv><machine-learning><image-processing><keras>","CreationDate":"2019-08-20T12:24:54.250","AcceptedAnswerId":57574251.0,"OwnerUserId":11917598.0,"Title":"CNN: Error when checking input: expected dense to have 2 dimensions, but got array with shape (391, 605, 700, 3)","Body":"<p>I'm getting a list of images to train my CNN. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential()\nmodel.add(Dense(32, activation='tanh', input_dim=100))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(optimizer='rmsprop',\n              loss='binary_crossentropy',\n              metrics=['accuracy'])\n\ndata, labels = ReadImages(TRAIN_DIR) \n\n# Train the model, iterating on the data in batches of 32 samples\nmodel.fit(np.array(data), np.array(labels), epochs=10, batch_size=32)\n<\/code><\/pre>\n\n<p>But I faced this error: <\/p>\n\n<p>'with shape ' + str(data_shape))<\/p>\n\n<p>ValueError: Error when checking input: expected dense_1_input to have 2 dimensions, but got array with shape (391, 605, 700, 3)<\/p>\n","answers":[{"AnswerId":"57574249","CreationDate":"2019-08-20T12:58:15.907","ParentId":null,"OwnerUserId":"3336192","Title":null,"Body":"<p>This is no CNN. A Convolutional Neural Network is defined by having Conv Layer. Those Layers work with imput shapes in 4D (Batchsize, ImageDimX, ImageDimY, ColorChannels). The Dense Layers(aka. Fully connected) you are using 2D input (Batchsize, DataAsAVector)<\/p>\n"},{"AnswerId":"57574251","CreationDate":"2019-08-20T12:58:21.140","ParentId":null,"OwnerUserId":"9025861","Title":null,"Body":"<p>You are feeding images to the Dense Layer.  Either flatten the images using .flatten() or use a model with CNN Layers. The shape (391,605,700,3) means you have 391 images of size 605x700 having 3 dimensions(rgb).<\/p>\n\n<pre><code>    model = Sequential()\n    model.add(Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', input_shape=(605, 700, 3)))\n    model.add(MaxPooling2D((2, 2)))\n    model.add(Flatten())\n    model.add(Dense(100, activation='relu', kernel_initializer='he_uniform'))\n    model.add(Dense(1, activation='sigmoid'))\n    model.compile(optimizer='rmsprop',\n          loss='binary_crossentropy',\n          metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>This <a href=\"https:\/\/machinelearningmastery.com\/how-to-develop-a-cnn-from-scratch-for-fashion-mnist-clothing-classification\/\" rel=\"nofollow noreferrer\">link<\/a> has good explanations for basic CNN.<\/p>\n"},{"AnswerId":"57575630","CreationDate":"2019-08-20T14:13:56.203","ParentId":null,"OwnerUserId":"10754197","Title":null,"Body":"<p>You need to first flatten the image if you want to pass the image directly to dense layers as Dense layer takes input in 2 dimensions only and since you are passing whole image there is 4 dimensions in it i.e. Number of images X Height X Width X Number of channels (391, 605, 700, 3).\nYou are not actually doing any convolutions on the image. To do convolutions you need to add CNN layers after initialising the model as sequential.\nTo add dense layer :<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Flatten())\nmodel.add(Dense(32, activation='tanh', input_dim=100))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(optimizer='rmsprop',\n              loss='binary_crossentropy',\n              metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>To add CNN layer and then flatten it :<\/p>\n\n<pre><code> model = Sequential()\n    model.add(Conv2D(input_shape=(605,700,3), filters=64, kernel_size=(3,3), \n    padding=\"same\",activation=\"relu\"))\n    model.add(Flatten())\n    model.add(Dense(32, activation='tanh', input_dim=100))\n    model.add(Dense(1, activation='sigmoid'))\n    model.compile(optimizer='rmsprop',\n                  loss='binary_crossentropy',\n                  metrics=['accuracy'])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57574445,"AnswerCount":0,"Tags":"<keras><deep-learning>","CreationDate":"2019-08-20T13:08:54.253","AcceptedAnswerId":null,"OwnerUserId":9423718.0,"Title":"how to plot predictions and true label in generators","Body":"<p>I have Image classification model, I want to plot the predictions versus the true labels in the generator, something like this. <\/p>\n\n<pre><code>predictions = model.predict_generator(test_generator, steps=batch_size)\npredicted_classes = np.argmax(predictions, axis=1). \n<\/code><\/pre>\n\n<h1>dont know how to plot the true labels and the predicted labels.<\/h1>\n\n<pre class=\"lang-py prettyprint-override\"><code># New Generator Test on Gavins data \n\ngavin_datagen = ImageDataGenerator(rescale=1.\/255.)\n\nGavin_generator = gavin_datagen.flow_from_directory(\n    base_dir,\n    target_size=(150,150),\n    batch_size=20,\n    shuffle=False)\n\n#  make predictions on the data, and give  filenames, for further clarification. \nGavin_generator.reset()\npred_gavin = model.predict_generator(Gavin_generator,steps=len(Gavin_generator))\n\nc1 = np.round(pred_gavin)\nfiles = Gavin_generator.filenames\n\n# store the results in dataframe.\nresults = pd.DataFrame({\"file\":files, \"Pr\":pred_gavin[:,0], \"class\":c1[:,0]})\nresults.head(20)\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>\n    file                    Pr      class\n0   test\\16077-0.jpg    0.122016    0.0\n1   test\\16077-18.jpg   0.753829    1.0\n2   test\\16077-6.jpg    0.173197    0.0\n3   test\\16079-0.jpg    0.722567    1.0\n4   test\\16079-6.jpg    0.788337    1.0\n5   test\\16091-0.jpg    0.449593    0.0\n6   test\\16091-12.jpg   0.746757    1.0\n7   test\\16091-18.jpg   0.795192    1.0\n8   test\\16091-6.jpg    0.557191    1.0\n9   test\\16093-0.jpg    0.783862    1.0\n10  test\\16093-12.jpg   0.807930    1.0\n11  test\\16093-18.jpg   0.802061    1.0\n12  test\\16093-6.jpg    0.766106    1.0\n13  test\\16097-0.jpg    0.809476    1.0\n14  test\\16097-12.jpg   0.661791    1.0\n15  test\\16097-18.jpg   0.785757    1.0\n16  test\\16097-6.jpg    0.809946    1.0\n17  test\\16101-0.jpg    0.700594    1.0\n18  test\\16101-12.jpg   0.443524    0.0\n19  test\\16101-18.jpg   0.787354    1.0\n<\/code><\/pre>\n\n<p>Here's what I want to do, the file column has the file path, I want to plot class \"anomalies\" which are \"1.0\"  only and the images, versus what the model predicted. <\/p>\n\n<p>So that I can check the filepaths to see if the model correctly predicted these images. Please Help !!!!!. <\/p>\n","answers":[]}
+{"QuestionId":57574483,"AnswerCount":1,"Tags":"<tensorflow><object-detection-api>","CreationDate":"2019-08-20T13:10:35.437","AcceptedAnswerId":57674638.0,"OwnerUserId":4964351.0,"Title":"Tensorflow OD API - pipeline.config changes after exporting trained model","Body":"<p>I'm using Tensorflow OD API 1.13 to do transfer learning of pre-trained ssd_mobilent_v2_coco from the model zoo. After finishing training, I export the frozen graph using the command:<\/p>\n\n<pre><code>python object_detection\\export_inference_graph.py ^\n--input_type image_tensor ^\n--pipeline_config_path C:\\TF_train\\models\\model\\pipeline.config ^\n--trained_checkpoint_prefix C:\\TF_train\\models\\model\\model.ckpt-7200 ^\n--output_directory C:\\export\\ssd_v2_7200\n<\/code><\/pre>\n\n<p>In the output folder, there are model files and a pipeline.config file. But when I open it and compare to the original one (used to train the model) I see that some fields were changed. <\/p>\n\n<p>For example, here's the feature extractor section from the original pipeline.config:<\/p>\n\n<pre><code>feature_extractor {\n  type: \"ssd_mobilenet_v2\"\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  use_depthwise: true\n}\n<\/code><\/pre>\n\n<p>and from exported pipeline.config:<\/p>\n\n<pre><code>feature_extractor {\n   type: \"ssd_mobilenet_v2\"\n   depth_multiplier: 1.0\n   min_depth: 16\n   conv_hyperparams {\n     regularizer {\n       l2_regularizer {\n         weight: 3.9999998989515007e-05\n       }\n     }\n     initializer {\n       truncated_normal_initializer {\n         mean: 0.0\n         stddev: 0.029999999329447746\n       }\n     }\n     activation: RELU_6\n     batch_norm {\n       decay: 0.9997000098228455\n       center: true\n       scale: true\n       epsilon: 0.0010000000474974513\n       train: true\n     }\n   }\n   use_depthwise: true\n }\n<\/code><\/pre>\n\n<p>Notice, how precision changed for l2_regularizer weight, stddev, decay and epsilon. Is this expected? Why is this happening?<\/p>\n","answers":[{"AnswerId":"57674638","CreationDate":"2019-08-27T12:23:58.637","ParentId":null,"OwnerUserId":"4964351","Title":null,"Body":"<p>Turns out it's a protobuf <a href=\"https:\/\/github.com\/protocolbuffers\/protobuf\/issues\/4569\" rel=\"nofollow noreferrer\">issue<\/a>:<\/p>\n\n<blockquote>\n  <p>We have a known issue for float type precision if it is using cpp extension:<\/p>\n  \n  <p>Python does not have C-style float , it only has a C-style double.\n  Thus pure python is using double precision for both float and double\n  field, cpp extension is using float precision for float field.<\/p>\n<\/blockquote>\n\n<p>I checked the protobuf active implementation like so:<\/p>\n\n<pre><code>python -c \"from google.protobuf.internal import api_implementation; print(api_implementation.Type())\"\n<\/code><\/pre>\n\n<p>The output was <code>cpp<\/code>. So I added the environment variable:<\/p>\n\n<p><code>PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python<\/code><\/p>\n\n<p>And now it works as expected.<\/p>\n"}]}
+{"QuestionId":57574501,"AnswerCount":0,"Tags":"<python><tensorflow><keras><scikit-learn><auc>","CreationDate":"2019-08-20T13:11:22.503","AcceptedAnswerId":null,"OwnerUserId":11558173.0,"Title":"how to use sklearn AUC in tensorflow.keras model metrics?","Body":"<p>I'm trying to use sklearn AUC in tf.keras as model metrics , for that I used custom made function from this link <a href=\"https:\/\/stackoverflow.com\/questions\/41032551\/how-to-compute-receiving-operating-characteristic-roc-and-auc-in-keras\">AUC<\/a><\/p>\n\n<p>Below is my model :<\/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)\n\nmodel = Model(inputs= [text_x,state_x,grade_x,cat_x,subcat_x,teach_x,num_x],outputs = [output_layer])\nmodel.compile(optimizer = 'Adam', loss= 'binary_crossentropy', metrics=[auc])\n\nhistory = model.fit(x = input_data , y= y_train,batch_size = 180, epochs = 15, callbacks = [es, mc], validation_data = (val_data, y_val))\n\nTrain on 69918 samples, validate on 17480 samples\nEpoch 1\/15\n69918\/69918 [==============================] - 278s 4ms\/sample - loss: 0.3086 - auc: 0.8516 - val_loss: 0.4711 - val_auc: 0.6896\nEpoch 2\/15\n69918\/69918 [==============================] - 275s 4ms\/sample - loss: 0.1417 - auc: 0.9738 - val_loss: 0.6638 - val_auc: 0.6692\nEpoch 3\/15\n69918\/69918 [==============================] - 275s 4ms\/sample - loss: 0.0506 - auc: 0.9964 - val_loss: 0.9611 - val_auc: 0.6824\nEpoch 4\/15\n69918\/69918 [==============================] - 276s 4ms\/sample - loss: 0.0329 - auc: 0.9983 - val_loss: 0.9462 - val_auc: 0.6719\n<\/code><\/pre>\n\n<p>I'm getting this error when evaluate the model, ValueError:<\/p>\n\n<pre><code>test_input_data = [text_test_1,state_test,grade_test,cat_test,subcat_test,teach_test,num_test]\nscore = model.evaluate(test_input_data, y_test,verbose = 1)\nprint('test_loss: ',score[0])\nprint('test_acc: ',score[1])\n\n&lt;ipython-input-103-336c032c70f4&gt; in &lt;module&gt;()\n      1 test_input_data = [text_test_1,state_test,grade_test,cat_test,subcat_test,teach_test,num_test]\n----&gt; 2 score = model.evaluate(test_input_data, y_test,verbose = 1)\n      3 print('test_loss: ',score[0])\n      4 print('test_acc: ',score[1])\n\n3 frames\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py in __call__(self, *args, **kwargs)\n   1456         ret = tf_session.TF_SessionRunCallable(self._session._session,\n   1457                                                self._handle, args,\n-&gt; 1458                                                run_metadata_ptr)\n   1459         if run_metadata:\n   1460           proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\nInvalidArgumentError: 2 root error(s) found.\n  (0) Invalid argument: ValueError: Only one class present in y_true. ROC AUC score is not defined in that case.\nTraceback (most recent call last):\n\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/script_ops.py\", line 209, in __call__\n    ret = func(*args)\n\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/sklearn\/metrics\/ranking.py\", line 355, in roc_auc_score\n    sample_weight=sample_weight)\n\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/sklearn\/metrics\/base.py\", line 76, in _average_binary_score\n    return binary_metric(y_true, y_score, sample_weight=sample_weight)\n\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/sklearn\/metrics\/ranking.py\", line 323, in _binary_roc_auc_score\n    raise ValueError(\"Only one class present in y_true. ROC AUC score \"\n\nValueError: Only one class present in y_true. ROC AUC score is not defined in that case.\n\n\n     [[{{node metrics_15\/auc\/PyFunc}}]]\n  (1) Invalid argument: ValueError: Only one class present in y_true. ROC AUC score is not defined in that case.\nTraceback (most recent call last):\n\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/script_ops.py\", line 209, in __call__\n    ret = func(*args)\n\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/sklearn\/metrics\/ranking.py\", line 355, in roc_auc_score\n    sample_weight=sample_weight)\n\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/sklearn\/metrics\/base.py\", line 76, in _average_binary_score\n    return binary_metric(y_true, y_score, sample_weight=sample_weight)\n\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/sklearn\/metrics\/ranking.py\", line 323, in _binary_roc_auc_score\n    raise ValueError(\"Only one class present in y_true. ROC AUC score \"\n\nValueError: Only one class present in y_true. ROC AUC score is not defined in that case.\n\n\n     [[{{node metrics_15\/auc\/PyFunc}}]]\n     [[metrics_15\/auc\/PyFunc\/_1683]]\n0 successful operations.\n0 derived errors ignored.\n<\/code><\/pre>\n\n<p>I tried tf.keras.metrics.AUC, then its working fine but when using sklearn AUC I having this error.\nhow to set sklearn's AUC in tf.keras.model metric function.\nany help would be appreciated..Thank you.<\/p>\n","answers":[]}
+{"QuestionId":57574604,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><computer-vision><resnet>","CreationDate":"2019-08-20T13:16:12.847","AcceptedAnswerId":57574869.0,"OwnerUserId":11951808.0,"Title":"How do I go from a collection of images, to a full machine learning dataset, that I can then use for transfer learning (resnet50 etc.)?","Body":"<p>I'm new to the field of machine learning and just starting to get the hang of it. I was wondering how I can go from an amount of images, to a machine learning dataset. I'm not quite sure on how to store the labels for my data either? A csv format?<\/p>\n","answers":[{"AnswerId":"57574869","CreationDate":"2019-08-20T13:30:52.100","ParentId":null,"OwnerUserId":"10069673","Title":null,"Body":"<p>This is a rough description on how the full workflow would look like. I can't give much more detail since your question is not very specific.<\/p>\n\n<h1>1. Label your images<\/h1>\n\n<p>There are many tools to label a set of images. You need to find one that fits the deep learning framework you want to use to train your neural network.<\/p>\n\n<p>I've used <a href=\"https:\/\/github.com\/Cartucho\/OpenLabeling\" rel=\"nofollow noreferrer\">OpenLabeling<\/a> to label my dataset in the YOLO format. However, this might not be the format you're looking for.<\/p>\n\n<h1>2. Find a pre-trained model<\/h1>\n\n<p>If you want to do transfer learning, you need a model (neural network) that has already been trained on another dataset.<\/p>\n\n<h1>3. Setup your PC<\/h1>\n\n<p>You need to install everything you need for training the neural network.\nI also recommend you to install and use CUDA (use your GPU to train the network instead of your CPU) since it's a lot faster.<\/p>\n\n<h1>4. Train your model<\/h1>\n\n<p>Use the transfer learning functinallity of your chosen framework to train your network using the pre-trained model and your labeled images.<\/p>\n"}]}
+{"QuestionId":57574784,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-20T13:26:25.343","AcceptedAnswerId":null,"OwnerUserId":1179925.0,"Title":"Setting low per_process_gpu_memory_fraction value cause memory fragmentation?","Body":"<p>Is setting low value of <code>low per_process_gpu_memory_fraction<\/code> can cause memory fragmentation? i.e. is it better to allocate large chunk of memory at once? Can we have less actual memory available using <code>set_gpu_env<\/code> comparing when we don't restrict gpu memory usage (tensorflow just occupies all available gpu by default)?<\/p>\n\n<pre><code>def set_tensorflow_session():\n    import tensorflow as tf\n    from keras.backend.tensorflow_backend import set_session\n    gpu_config = tf.GPUOptions(allow_growth=True,\n                               per_process_gpu_memory_fraction=0.1)\n    session = tf.Session(config=tf.ConfigProto(gpu_options=gpu_config))\n    set_session(session)\n\n\ndef set_gpu_env(gpu_id_list):\n    idxs_str = ','.join([str(i) for i in gpu_id_list])\n    os.environ[\"CUDA_DEVICE_ORDER\"] = \"PCI_BUS_ID\"\n    if idxs_str == '-1':\n        os.environ['CUDA_VISIBLE_DEVICES'] = ''\n    else:\n        os.environ['CUDA_VISIBLE_DEVICES'] = idxs_str\n    set_tensorflow_session()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57575175,"AnswerCount":1,"Tags":"<tensorflow><keras><tf.keras>","CreationDate":"2019-08-20T13:49:08.160","AcceptedAnswerId":null,"OwnerUserId":8526434.0,"Title":"tf.keras model saving discrepancy","Body":"<p>I noticed a discrepancy between the two savings function available in <code>tf.keras<\/code> that both seem to do the same thing and I wonder what the actual difference is.<\/p>\n\n<p>When I save a model using the following method and then evaluate its size by:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.save(path)\n\n\nprint(\"Size of the model: %.2f Mb\" % (os.path.getsize(path) \/ float(2**20)))\n<\/code><\/pre>\n\n<p>I have: <\/p>\n\n<blockquote>\n  <p>Size of the model: 172.13 Mb<\/p>\n<\/blockquote>\n\n<p>But if now I try with the function: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tf.keras.models.save_model(model, path)\n<\/code><\/pre>\n\n<p>I have:<\/p>\n\n<blockquote>\n  <p>Size of the model: 57.39 Mb<\/p>\n<\/blockquote>\n\n<p>By looking at the documentation of both Keras and <code>tf.keras<\/code> I found:<\/p>\n\n<p><em><code>tf.keras.models.save_model<\/code>: The saved model contains: - the model's configuration (topology) - the model's weights - the model's optimizer's state (if any)<\/em><\/p>\n\n<p>And: <\/p>\n\n<p><em><code>model.save()<\/code>:  to save a Keras model which will contain:the architecture of the model,the weights of the model, the training configuration (loss, optimizer), the state of the optimizer.<\/em><\/p>\n\n<p>Both seem to achieve the same thing so my guess is that <code>tf.keras.save_model<\/code> probably do some operation in the background (compression ?). Do you have any idea ?<\/p>\n","answers":[{"AnswerId":"57580571","CreationDate":"2019-08-20T19:44:03.590","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>The GitHub source code of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r2.0\/tensorflow\/python\/keras\/engine\/network.py#L1120\" rel=\"nofollow noreferrer\">model.save<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r2.0\/tensorflow\/python\/keras\/saving\/save.py#L47\" rel=\"nofollow noreferrer\">tf.keras.models.save_model<\/a> ultimately lead to the internal method of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r2.0\/tensorflow\/python\/keras\/saving\/saved_model\/save.py#L61\" rel=\"nofollow noreferrer\">save<\/a> present in tensorflow.python.keras.saving.saved_model package if your save format is 'tf' or it leads to internal method of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r2.0\/tensorflow\/python\/keras\/saving\/hdf5_format.py#L46\" rel=\"nofollow noreferrer\">save_model_to_hdf5<\/a> if your save format is 'h5'. You maybe doing something wrong in your approach. <\/p>\n"}]}
+{"QuestionId":57575264,"AnswerCount":1,"Tags":"<python><tensorflow><tensor>","CreationDate":"2019-08-20T13:53:37.327","AcceptedAnswerId":57578762.0,"OwnerUserId":9599714.0,"Title":"Create random trainable indices to build two 2_D tensor from one 2_D tensor in tensorflow","Body":"<p>I'm trying to create two tensors with shape <strong>shape=(30, 4200)<\/strong> from one tensor <strong><em>AAA<\/em><\/strong> with shape=(30, 4201)and type= float32.<\/p>\n\n<p>So I use <strong>tf.gather_nd<\/strong> to build the tensor and <strong><em>def setup_ent_emb(self)<\/em><\/strong> to generate a random trainable tensor as follows:<\/p>\n\n<pre><code>V1 = tf.gather_nd(params=AAA ,indices=V1_emb )\nV2 = tf.gather_nd(params=AAA ,indices=V2_emb )\n\ndef setup_ent_emb(self):\n     self.V1_emb = tf.get_variable(name=\"V1_emb\",initializer=tf.random_uniform(shape=[30, 4200], minval=0, maxval=4200, dtype=tf.int32))\n     self.V2_emb = tf.get_variable(name=\"V2_emb\",initializer=tf.random_uniform(shape=[30, 4200], minval=0, maxval=4200, dtype=tf.int32))\n<\/code><\/pre>\n\n<p>This error will appear again after running the code with several different changes.<\/p>\n\n<blockquote>\n  <p>ValueError: indices.shape[-1] must be &lt;= params.rank, but saw indices shape: [30,4200] and params shape: [30,4201] for 'sub_head' (op: 'GatherNd') with input shapes: [30,4201], [30,4200]<\/p>\n<\/blockquote>\n\n<p>How do I select the correct tensor shape?<\/p>\n","answers":[{"AnswerId":"57578762","CreationDate":"2019-08-20T17:26:09.037","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>This is how you would use <code>tf.gather_nd<\/code> for your case:<\/p>\n\n<pre><code># If V1_emb and V2_emb have different shapes you need to take the shape of each one\ns = tf.shape(V1, out_type=V1_emb.dtype)\nbatch_idx = tf.tile(tf.expand_dims(tf.range(s[0]), 1), [1, s[1]])\nV1 = tf.gather_nd(params=AAA, indices=tf.stack([batch_idx, V1_emb], axis=-1))\nV2 = tf.gather_nd(params=AAA, indices=tf.stack([batch_idx, V2_emb], axis=-1))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57575484,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2019-08-20T14:06:04.260","AcceptedAnswerId":57578583.0,"OwnerUserId":688080.0,"Title":"Is there something similar to tf.cond but for vector predicates?","Body":"<p>Say I have a 5D tensor <code>x<\/code> and a 1D boolean mask <code>m<\/code> where <code>m.shape[0] == x.shape[0]<\/code>, and I want to decide which sub-network should be applied on each 4D sample inside <code>x<\/code> based on the corresponding boolean entry of <code>m<\/code>. <\/p>\n\n<p>To my knowledge <code>tf.cond<\/code> accepts scalar predicator only. Though <code>tf.boolean_mask<\/code> may be helpful to split the samples inside a batch into two subsets according to <code>m<\/code> as desired, I am not sure how to re-pack the outputs back into one 5D tensor without messing up the original sample order. Any hints?<\/p>\n","answers":[{"AnswerId":"57578583","CreationDate":"2019-08-20T17:11:23.537","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>The simplest thing would be to evaluate the data on both models and the use <code>tf.where<\/code> to select the final output.<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef model1(x):\n    return 2 * x\n\ndef model2(x):\n    return -3 * x\n\nwith tf.Graph().as_default(), tf.Session() as sess:\n    x = tf.placeholder(tf.float32, [None, None])  # 2D for simplicity\n    m = tf.placeholder(tf.bool, [None])\n    y1 = model1(x)\n    y2 = model2(x)\n    y = tf.where(m, y1, y2)\n    print(sess.run(y, feed_dict={x: [[1, 2], [3, 4], [5, 6]], m: [True, False, True]}))\n    # [[  2.   4.]\n    #  [ -9. -12.]\n    #  [ 10.  12.]]\n<\/code><\/pre>\n\n<p>If you really want to avoid that, you can use <code>tf.boolean_mask<\/code> to split the data and then recombine it with <code>tf.scatter_nd<\/code>. This is one possible way.<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef model1(x):\n    return 2 * x\n\ndef model2(x):\n    return -3 * x\n\nwith tf.Graph().as_default(), tf.Session() as sess:\n    x = tf.placeholder(tf.float32, [None, None])\n    m = tf.placeholder(tf.bool, [None])\n    n = tf.size(m)\n    i = tf.range(n)\n    x1 = tf.boolean_mask(x, m)\n    i1 = tf.boolean_mask(i, m)\n    y1 = model1(x1)\n    m_neg = ~m\n    x2 = tf.boolean_mask(x, m_neg)\n    i2 = tf.boolean_mask(i, m_neg)\n    y2 = model2(x2)\n    y = tf.scatter_nd(tf.expand_dims(tf.concat([i1, i2], axis=0), 1),\n                      tf.concat([y1, y2], axis=0),\n                      tf.concat([[n], tf.shape(y1)[1:]], axis=0))\n    print(sess.run(y, feed_dict={x: [[1, 2], [3, 4], [5, 6]], m: [True, False, True]}))\n    # [[  2.   4.]\n    #  [ -9. -12.]\n    #  [ 10.  12.]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57575652,"AnswerCount":0,"Tags":"<python><tensorflow><pytorch><tabular>","CreationDate":"2019-08-20T14:15:10.033","AcceptedAnswerId":null,"OwnerUserId":8899404.0,"Title":"Format tabular data from pandas into PyTorch\/Tensorflow models","Body":"<p>I want to write a sequential model in PyTorch\/Tensorflow. I have a csv\/pandas dataset with an index column and some outcome variables like this:<\/p>\n\n<pre><code>| ID     | SeqID  | outcome_1 | outcome_2 | ...\n|:-------|-------:|:---------:|:---------:|\n| 'Bob'  | 1      | H         | B         |\n| 'Bob'  | 2      | W         | C         |\n| 'Bob'  | 3      | H         | S         |\n| 'Alice'| 1      | H         | NaN       |\n| 'Max'  | 1      | H         | B         |\n| 'Max'  | 2      | L         | B         |\n...\n<\/code><\/pre>\n\n<p>I would like to get a list of arrays (one for each individual) where every entry is an integer that represents the outcome. For example, the array for 'Bob' for a single variable (with H=1. W=2, L=3 and =4) would be: [1, 2, 1, 4] and with the second variable (B=1, C=2, S=3, =4): [[1, 2, 1, 4], [1, 2, 3, 4]], etc. I expect the arrays to not be of the same size as the sequences in the data set are not. I have tried some methods but I feel they are somewhat an overkill. For example I can do: <\/p>\n\n<pre><code>outcome_array = initial_df.pivot(index='SessionId', columns='SeqId', values='outcome_1')\noutcome_array.reset_index(inplace=True, drop=True)\n<\/code><\/pre>\n\n<p>This returns the \"pivoted\" dataframe with number of rows as the longest sequence, NaN where the sequence does not exist for some IDs, etc. and then I \"clean\" the pivoted table. On top of that, I could have to do this procedure for every outcome variable I want to work with making it potentially a very lengthy process. Is there any clean, fast way to do this without recurring to table pivoting and further cleaning?<\/p>\n","answers":[]}
+{"QuestionId":57576036,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2019-08-20T14:36:26.097","AcceptedAnswerId":57585760.0,"OwnerUserId":6500085.0,"Title":"\"Unkown (custom) loss function\" when using tflite_convert on a {TF 2.0.0-beta1 ; Keras} model","Body":"<h1>Summary<\/h1>\n\n<p>My question is composed by:<\/p>\n\n<ul>\n<li>A context in which I present my project, my working environment and my workflow<\/li>\n<li>The detailed problem<\/li>\n<li>The concerned parts of my  code<\/li>\n<li>The solutions I tried to solve my problem<\/li>\n<li>The question reminder<\/li>\n<\/ul>\n\n<h1>Context<\/h1>\n\n<p>I've written a Python Keras implementation of a downgraded version of the original Super-Resolution GAN. Now I want to test it using Google Firebase Machine Learning Kit, by hosting it in the Google servers. That's why I have to convert my Keras program to a TensorFlow Lite one.<\/p>\n\n<h2>Environment and workflow (with the problem)<\/h2>\n\n<p>I'm training my program on Google Colab working environment: there, I've installed <code>TF 2.0.0-beta1<\/code> (this choice is motivated by this uncorrect answer: <a href=\"https:\/\/datascience.stackexchange.com\/a\/57408\/78409\">https:\/\/datascience.stackexchange.com\/a\/57408\/78409<\/a>).<\/p>\n\n<p>Workflow (and problem):<\/p>\n\n<ol>\n<li><p>I write locally my Python Keras program, keeping in mind that it will run on TF 2. So I use TF 2 imports, for example: <code>from tensorflow.keras.optimizers import Adam<\/code> and also <code>from tensorflow.keras.layers import Conv2D, BatchNormalization<\/code><\/p><\/li>\n<li><p>I send my code to my Drive<\/p><\/li>\n<li><p>I run without any problem my Google Colab Notebook: TF 2 is used.<\/p><\/li>\n<li><p>I get the output model in my Drive, and I download it.<\/p><\/li>\n<li><p>I try to convert this model to the TFLite format by executing the following CLI: <code>tflite_convert --output_file=srgan.tflite --keras_model_file=srgan.h5<\/code>: <strong>here the problem appears<\/strong>.<\/p><\/li>\n<\/ol>\n\n<h1>The problem<\/h1>\n\n<p>Instead of outputing the TF Lite converted model from the TF (Keras) model, the previous CLI outputs this error:<\/p>\n\n<blockquote>\n  <p>ValueError: Unknown loss function:build_vgg19_loss_network<\/p>\n<\/blockquote>\n\n<p>The function <code>build_vgg19_loss_network<\/code> is a custom loss function that I've implemented and that must be used by the GAN.<\/p>\n\n<h1>Parts of code that rise this problem<\/h1>\n\n<h2>Presenting the custom loss function<\/h2>\n\n<p>The custom loss function is implemented like that:<\/p>\n\n<pre><code>def build_vgg19_loss_network(ground_truth_image, predicted_image):\n    loss_model = Vgg19Loss.define_loss_model(high_resolution_shape)\n    return mean(square(loss_model(ground_truth_image) - loss_model(predicted_image)))\n<\/code><\/pre>\n\n<h2>Compiling the generator network with my custom loss function<\/h2>\n\n<p><code>generator_model.compile(optimizer=the_optimizer, loss=build_vgg19_loss_network)<\/code><\/p>\n\n<h1>What I've tried to do in order to solve the problem<\/h1>\n\n<ol>\n<li><p>As I read it on StackOverflow (link at the beginning of this question), TF 2 was thought to be sufficient to output a Keras model which would be correctly processed by my <code>tflite_convert<\/code> CLI. But it's not, obviously.<\/p><\/li>\n<li><p>As I read it on GitHub, I tried to manually set my custom loss function among Keras' loss functions, by adding these lines: <code>import tensorflow.keras.losses\ntensorflow.keras.losses.build_vgg19_loss_network = build_vgg19_loss_network<\/code>. It didn't work.<\/p><\/li>\n<li><p>I read on GitHub I could use custom objects with <code>load_model<\/code> Keras function: but I only want to use <code>compile<\/code> Keras function. Not <code>load_model<\/code>.<\/p><\/li>\n<\/ol>\n\n<h1>My final question<\/h1>\n\n<p>I want to do only minor changes to my code, since it works fine. So I don't want, for example, to replace <code>compile<\/code> with <code>load_model<\/code>. With this constraint, could you help me, please, to make my CLI <code>tflite_convert<\/code> works with my custom loss function?<\/p>\n","answers":[{"AnswerId":"57585760","CreationDate":"2019-08-21T06:38:52.840","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>Since you are claiming that TFLite conversion is failing due to a custom loss function, you can save the model file without keep the optimizer details. To do that, set <code>include_optimizer<\/code> parameter to False as shown below:<\/p>\n\n<pre><code>model.save('model.h5', include_optimizer=False)\n<\/code><\/pre>\n\n<p>Now, if all the layers inside your model are convertible, they should get converted into TFLite file.<\/p>\n\n<p>Edit: \nYou can then convert the h5 file like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\nmodel = tf.keras.models.load_model('model.h5')   # srgan.h5 for you\nconverter = tf.lite.TFLiteConverter.from_keras_model(model)\ntflite_model = converter.convert()\nopen(\"converted_model.tflite\", \"wb\").write(tflite_model)\n<\/code><\/pre>\n\n<p>Usual practice to overcome the unsupported operators in TFLite conversion is <a href=\"https:\/\/www.tensorflow.org\/lite\/guide\/ops_custom\" rel=\"nofollow noreferrer\">documented here<\/a>.<\/p>\n"}]}
+{"QuestionId":57576167,"AnswerCount":1,"Tags":"<tensorflow><keras><lstm>","CreationDate":"2019-08-20T14:43:33.123","AcceptedAnswerId":57577469.0,"OwnerUserId":735008.0,"Title":"Processing examples with different sequence lengths using Keras LSTM layers","Body":"<p>I need to train a stack of LSTM layers using examples with different sequence lengths. If we use the Keras sequential model, we may implement the code as follows.<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(32, return_sequences=True, input_shape=(None, 5)))\nmodel.add(LSTM(8, return_sequences=True))\nmodel.add(Dense(2, activation='sigmoid'))\n\ndef train_generator():\n    while True:\n        sequence_length = np.random.randint(10, 100)\n        x_train = np.random.random((1000, sequence_length, 5))\n        # y_train will depend on past 5 timesteps of x\n        y_train = x_train[:, :, 0]\n        for i in range(1, 5):\n            y_train[:, i:] += x_train[:, :-i, i]\n        y_train = to_categorical(y_train &gt; 2.5)\n        yield x_train, y_train\n\nmodel1.fit_generator(train_generator(), steps_per_epoch=2, epochs=2, verbose=1)\n<\/code><\/pre>\n\n<p>The above is based on what I searched in another question:\n<a href=\"https:\/\/datascience.stackexchange.com\/questions\/26366\/training-an-rnn-with-examples-of-different-lengths-in-keras\">https:\/\/datascience.stackexchange.com\/questions\/26366\/training-an-rnn-with-examples-of-different-lengths-in-keras<\/a>.\nThe above code works fine, and the above model can be trained using examples with different lengths.<\/p>\n\n<p>However, in my case, I <strong>should<\/strong> subclass tf.keras.Model rather than using the sequential model.<\/p>\n\n<pre><code>class LSTMModel(tf.keras.Model):\n    def __init__(self):\n        super(LSTMModel, self).__init__()\n\n        self._lstm_0 = LSTM(32, return_sequences=True, input_shape=(None, 5)) \n        self._lstm_1 = LSTM(8, return_sequences=True)\n        self._dense = Dense(2, activation='sigmoid')\n\n    def call(self, inputs, training=False):\n        output = self._lstm_0(inputs)\n        output = self._lstm_1(output)\n        output = self._dense(output)\n\n        return output\n<\/code><\/pre>\n\n<p>My expectation is that this second code should be equivalent to the first one. However, it crashes with the following error message.<\/p>\n\n<pre><code>BaseCollectiveExecutor::StartAbort Invalid argument: Operation expected a list with 33 elements but got a list with 27 elements\n<\/code><\/pre>\n\n<p>Could anyone tell me the reason and give some advice?<\/p>\n","answers":[{"AnswerId":"57577469","CreationDate":"2019-08-20T15:56:54.360","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p><strong>Everything executed with Tensorflow 1.14<\/strong><\/p>\n\n<p>I run the following code using to include all the utils:<\/p>\n\n<pre><code>import numpy as np\nfrom tensorflow.keras import Model\nfrom tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import LSTM, Dense\nfrom tensorflow.keras.utils import to_categorical\n\ndef train_generator():\n    while True:\n        sequence_length = np.random.randint(10, 100)\n        x_train = np.random.random((1000, sequence_length, 5))\n        # y_train will depend on past 5 timesteps of x\n        y_train = x_train[:, :, 0]\n        for i in range(1, 5):\n            y_train[:, i:] += x_train[:, :-i, i]\n        y_train = to_categorical(y_train &gt; 2.5)\n        yield x_train, y_train\n<\/code><\/pre>\n\n<p>Then the first model is:<\/p>\n\n<pre><code>model_1 = Sequential()\nmodel_1.add(LSTM(32, return_sequences=True, input_shape=(None, 5)))\nmodel_1.add(LSTM(8, return_sequences=True))\nmodel_1.add(Dense(2, activation='sigmoid'))\nmodel_1.compile(optimizer=\"adam\", loss=\"mse\")\nmodel_1.fit_generator(train_generator(), steps_per_epoch=2, epochs=2, verbose=1)\n<\/code><\/pre>\n\n<p>And the second model is:<\/p>\n\n<pre><code>class LSTMModel(Model):\n    def __init__(self):\n        super(LSTMModel, self).__init__()\n\n        self._lstm_0 = LSTM(32, return_sequences=True, input_shape=(None, 5)) \n        self._lstm_1 = LSTM(8, return_sequences=True)\n        self._dense = Dense(2, activation='sigmoid')\n\n    def call(self, inputs, training=False):\n        output = self._lstm_0(inputs)\n        output = self._lstm_1(output)\n        output = self._dense(output)\n\n        return output\n\nmodel_2 = LSTMModel()\nmodel_2.compile(optimizer=\"adam\", loss=\"mse\")\nmodel_2.fit_generator(train_generator(), steps_per_epoch=2, epochs=2, verbose=1)\n<\/code><\/pre>\n\n<p>The results were equivalent. If it still fails on your end, include more information in the question such as the TF version you you are running.<\/p>\n\n<hr>\n\n<p>As per the opened <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/31816\" rel=\"nofollow noreferrer\">Github issue<\/a>, this is a bug and has been solved in <code>tf-nightly-gpu-2.0-preview<\/code>. <\/p>\n"}]}
+{"QuestionId":57576566,"AnswerCount":0,"Tags":"<numpy><keras><layer>","CreationDate":"2019-08-20T15:05:31.487","AcceptedAnswerId":null,"OwnerUserId":11952333.0,"Title":"I saved model history as npz but I get Value Error : Unknown Layer","Body":"<p>I did some experiments with datasets and I saved model historys as a list. I saved as .npz but now I couldn't open because of custom layer which is named Conv2DAdaptive. I get Value Error : Unknown Layer Conv2DAdaptive. I look for the similiar problems but all I seen is related with keras model loading but my problem is loading with numpy. <\/p>\n\n<pre><code>import numpy as np\n# save np.load\nnp_load_old = np.load\n\n# modify the default parameters of np.load\nnp.load = lambda *a,**k: np_load_old(*a, allow_pickle=True, **k)\n\na = np.load('mnistv1scores.npz')\n\na['historylisted']\n\n\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-6-e658aee41c83&gt; in &lt;module&gt;\n----&gt; 1 a['historylisted']\n\n~\/anaconda3\/envs\/tensorflow_gpu\/lib\/python3.6\/site-packages\/numpy\/lib\/npyio.py in __getitem__(self, key)\n    260                 return format.read_array(bytes,\n    261                                          allow_pickle=self.allow_pickle,\n--&gt; 262                                          pickle_kwargs=self.pickle_kwargs)\n    263             else:\n    264                 return self.zip.read(key)\n\n~\/anaconda3\/envs\/tensorflow_gpu\/lib\/python3.6\/site-packages\/numpy\/lib\/format.py in read_array(fp, allow_pickle, pickle_kwargs)\n    699             pickle_kwargs = {}\n    700         try:\n--&gt; 701             array = pickle.load(fp, **pickle_kwargs)\n    702         except UnicodeError as err:\n    703             if sys.version_info[0] &gt;= 3:\n\n~\/anaconda3\/envs\/tensorflow_gpu\/lib\/python3.6\/site-packages\/keras\/engine\/network.py in __setstate__(self, state)\n   1264 \n   1265     def __setstate__(self, state):\n-&gt; 1266         model = saving.unpickle_model(state)\n   1267         self.__dict__.update(model.__dict__)\n   1268 \n\n~\/anaconda3\/envs\/tensorflow_gpu\/lib\/python3.6\/site-packages\/keras\/engine\/saving.py in unpickle_model(state)\n    433 def unpickle_model(state):\n    434     f = h5dict(state, mode='r')\n--&gt; 435     return _deserialize_model(f)\n    436 \n    437 \n\n~\/anaconda3\/envs\/tensorflow_gpu\/lib\/python3.6\/site-packages\/keras\/engine\/saving.py in _deserialize_model(f, custom_objects, compile)\n    223         raise ValueError('No model found in config.')\n    224     model_config = json.loads(model_config.decode('utf-8'))\n--&gt; 225     model = model_from_config(model_config, custom_objects=custom_objects)\n    226     model_weights_group = f['model_weights']\n    227 \n\n~\/anaconda3\/envs\/tensorflow_gpu\/lib\/python3.6\/site-packages\/keras\/engine\/saving.py in model_from_config(config, custom_objects)\n    456                         '`Sequential.from_config(config)`?')\n    457     from ..layers import deserialize\n--&gt; 458     return deserialize(config, custom_objects=custom_objects)\n    459 \n    460 \n\n~\/anaconda3\/envs\/tensorflow_gpu\/lib\/python3.6\/site-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~\/anaconda3\/envs\/tensorflow_gpu\/lib\/python3.6\/site-packages\/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~\/anaconda3\/envs\/tensorflow_gpu\/lib\/python3.6\/site-packages\/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~\/anaconda3\/envs\/tensorflow_gpu\/lib\/python3.6\/site-packages\/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~\/anaconda3\/envs\/tensorflow_gpu\/lib\/python3.6\/site-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~\/anaconda3\/envs\/tensorflow_gpu\/lib\/python3.6\/site-packages\/keras\/utils\/generic_utils.py in deserialize_keras_object(identifier, module_objects, custom_objects, printable_module_name)\n    136             if cls is None:\n    137                 raise ValueError('Unknown ' + printable_module_name +\n--&gt; 138                                  ': ' + class_name)\n    139         if hasattr(cls, 'from_config'):\n    140             custom_objects = custom_objects or {}\n\nValueError: Unknown layer: Conv2DAdaptive\n<\/code><\/pre>\n\n<p>a['historylisted'] is list of modelhistory objects , \u0131 want to reach them again but I cant so that I'm waiting for your helps guys.<\/p>\n","answers":[]}
+{"QuestionId":57576572,"AnswerCount":0,"Tags":"<r><machine-learning><keras><neural-network><prediction>","CreationDate":"2019-08-20T15:05:38.193","AcceptedAnswerId":null,"OwnerUserId":11769950.0,"Title":"Predict a list based of another list using keras in R","Body":"<p>I have a table as follows: <\/p>\n\n<pre><code>year &lt;- c(\"1940\", \"1940\", \"1940\", \"1940\")\nAutumn.ref &lt;- c(\"AB322\", \"BE345\", \"FR654\", \"GR843\")\nSpring.ref &lt;- c(\"JB127\", \"BE345\", \"KL024\", \"GR843\")\nA.code1 &lt;- c(0, 0, 3, 1)\nA.code2 &lt;- c(1, 1, 2, 3)\nS.code1 &lt;- c(2, 0, 0, 1)\nS.code2 &lt;- c(0, 0, 1, 0)\nS.code3 &lt;- c(0, 3, 1, 0) \ndf &lt;- c(\"year\", \"Autumn.ref\", \"Spring.ref\", \"A.code1\", \"A.code2\", \"S.code1\", \"S.code2\", \"S.code3\")\n\n<\/code><\/pre>\n\n<p>For each autumn reference (Autumn.ref)  there is a matching spring reference (Spring.ref). Next, there are columns called A.code or S.code which represent codes that can be found in either autumn or spring. For each autumn reference or spring reference, I have the number of times each autumnd code (A.code) or spring code (S.code) appears in autumn or spring. <\/p>\n\n<p>For instance, autumn reference \"FR654\" contains the code \"A.code1\" 3 times and the code \"A.code2\" two times. The spring reference \"JB127\" contains the code \"S.code1\" two times and doesn't contain either of \"S.code2\" or \"S.code3\". <\/p>\n\n<p>Another thing to note is that the spring codes have many more columns than the autumn codes (probably 150 autumn codes and 600 spring codes). <\/p>\n\n<p>The data is collected over several years (here I only included the year 1940) <\/p>\n\n<p>I would like to build a neural network that when given a spring reference and the affiliated spring codes, it outputs the matching autumn reference and the autumn codes (of course, the spring reference would come from a year different than the years where the winter data is collected). For this, I know I have to train the network on the data that I already have.<\/p>\n\n<p>Since I am very new to machine learning, I did a bit of reading and I was thinking to use the package Keras in R. From what I understand, I would first need to normalize the data, then choose some layers and finally evaluate the model. However I am very confused as to how I can normalize this data or what the layers are meant to be doing\/how to create them. <\/p>\n\n<p>If possible, I would like some guidance on where to start with this algorithm and a bit more information about the steps that I need to take in order to achieve my result. However, I am not looking for someone to make this algorithm for me, I would simply like some advice or tips on which commands could get me started\/could be useful for the algorithm. <\/p>\n\n<p>Thank you!<\/p>\n","answers":[]}
+{"QuestionId":57576861,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-08-20T15:22:34.040","AcceptedAnswerId":null,"OwnerUserId":11831785.0,"Title":"Keras: Question on GPhilo's post on predict vs predict_on_batch. Can't predict() already handle single images just fine?","Body":"<p><a href=\"https:\/\/stackoverflow.com\/questions\/44972565\/what-is-the-difference-between-the-predict-and-predict-on-batch-methods-of-a-ker\">What is the difference between the predict and predict_on_batch methods of a Keras model?<\/a><\/p>\n\n<p>\"The difference lies in when you pass as x data that is larger than one batch.<\/p>\n\n<p>predict will go through all the data, batch by batch, predicting labels. It thus internally does the splitting in batches and feeding one batch at a time.<\/p>\n\n<p>predict_on_batch, on the other hand, assumes that the data you pass in is exactly one batch and thus feeds it to the network. It won't try to split it (which, depending on your setup, might prove problematic for your GPU memory if the array is very big)\" - GPhilo<\/p>\n\n<p>1) I am having trouble understanding why we would ever use predict_on_batch on a single image instead of running predict() directly on a single batch. Gphilo says that the advantage of predict_on_batch is that predict_on_batch won't try to split the variable in the argument. <\/p>\n\n<p>For a single image, what do we care if the data tries to split it or not? Shouldn't .predict() be able to handle the single image just fine? predict() will be able to correctly recognize a single image, right?<\/p>\n\n<p>So let's say we have a single greyscale image single_image of size (1,128,128,1) <\/p>\n\n<p>Then can't we just do:<\/p>\n\n<p>model.predict(single_image)<\/p>\n\n<p>instead of<\/p>\n\n<p>model.predict_on_batch(single_image) ? <\/p>\n\n<p>And there seems to be disadvantages to using predict_on_batch. So why would we ever use predict_on_batch if predict_on_batch can cause GPU problems?  What advantage would predict_on_batch give? <\/p>\n\n<p>2) Or am I misunderstanding? Can .predict actually not take in single images well for a reason I haven't thought of? <\/p>\n\n<p>I am a beginner, and I think perhaps I am misunderstanding\/not noticing something obvious in GPhilo's message.<\/p>\n","answers":[{"AnswerId":"57578782","CreationDate":"2019-08-20T17:28:14.873","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Yes you can.<\/p>\n\n<ul>\n<li>You can predict many batches (and the system will conclude it's one batch)   <\/li>\n<li>You can predict one batch (and the system will not try to calculate batches)<\/li>\n<\/ul>\n\n<p>Both options are just fine. <\/p>\n\n<p>There is no disadvantage in using <code>predict_on_batch<\/code>. On the contrary, it would be faster than <code>predict<\/code> because it does less things. <\/p>\n\n<p>All memory problems that may happen in <code>predict_on_batch<\/code> would happen exactly the same in <code>predict<\/code> if <code>batch_size == len(data)<\/code>.<\/p>\n"}]}
+{"QuestionId":57577550,"AnswerCount":0,"Tags":"<ios><object><tensorflow><detection>","CreationDate":"2019-08-20T16:02:03.817","AcceptedAnswerId":null,"OwnerUserId":11861686.0,"Title":"Error loading a TF lite format model in ios device for custom object detection","Body":"<p>I have trained a custom object detection, using transfer learning based on MobileNetV2. I converted the model to tflite format so that I can run it on my IOS device. However, when I run the model in xcode, I get an error.<\/p>\n\n<pre><code>2019-08-20 16:11:07.225613+0100 ObjectDetection[5692:1119136] \nInitialized TensorFlow Lite runtime.\n2019-08-20 16:11:07.280287+0100 ObjectDetection[5692:1119136] [MC] \nSystem group container for \nsystemgroup.com.apple.configurationprofiles path is \n\/var\/containers\/Shared\/SystemGroup\/systemgroup.com.apple.configu    \nrationprofiles\n2019-08-20 16:11:07.282075+0100 ObjectDetection[5692:1119136] [MC] \nReading from public effective user settings.\n\n**Failed to invoke the interpreter with error: Provided data count \n1080000 must match the required count 602112.\nFailed to invoke the interpreter with error: Provided data count \n1080000  must match the required count 602112.**\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57578553,"AnswerCount":0,"Tags":"<keras><multiprocessing><resnet>","CreationDate":"2019-08-20T17:08:42.010","AcceptedAnswerId":null,"OwnerUserId":7481520.0,"Title":"Multiprocessing ResNet Feature Extraction with Image URLs","Body":"<p>I have a simple function to take an image url and extract features from it using resnet in Keras then hand it off to an Xgboost model loaded from a pkl file. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def classify(img, resnet_model, loaded_model):\n  try:\n    images = io.imread(img.strip())\n    images = Image.fromarray(images)\n    test_image = img.resize((224,224))\n    test_image = image.img_to_array(test_image)\n    test_image = expand_dims(test_image, axis = 0)\n    img_data = preprocess_input(test_image)\n    image_features = resnet_model.predict(img_data)\n    image_features_array = array(image_features)\n    predicted_image = loaded_model.predict(xgb.DMatrix(DataFrame(image_features_array))) \n  except:\n    predicted_image = 'Broken URL'\n\n  return predicted_image`\n<\/code><\/pre>\n\n<p>Currently I am just looping through a list of img urls and it works fine but I will need it to perform much faster.  The code itself may not be the most efficient yet but I am mostly concerned with multiprocessing.  My attempts either hang or immediately result in an empty list.<\/p>\n\n<p>There is a similar question posted years ago here: <a href=\"https:\/\/stackoverflow.com\/questions\/45909738\/tensorflow-multiprocessing-for-image-feature-extraction\">question<\/a> but the answers were not very satisfying and involved holding a batch of image files locally. I would prefer to just have one worker making the request for the image and then predicting that image.<\/p>\n","answers":[]}
+{"QuestionId":57578845,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><image-processing><keras>","CreationDate":"2019-08-20T17:32:03.420","AcceptedAnswerId":57600828.0,"OwnerUserId":11917598.0,"Title":"ValueError: could not convert string to float: 'nonPdr'","Body":"<p>I'm facing this error: ValueError: could not convert string to float: 'nonPdr', when I run this code:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv2D(input_shape=(605,700,3), filters=64, kernel_size=(3,3), padding=\"valid\",activation=\"tanh\"))\nmodel.add(Flatten())\nmodel.add(Dense(32, activation='tanh', input_dim=100))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(optimizer='rmsprop', loss='binary_crossentropy', metrics=['accuracy'])\n\ndata, labels = ReadImages(TRAIN_DIR)\n\n# Train the model, iterating on the data in batches of 32 samples\nmodel.fit(np.array(data), np.array(labels), epochs=10, batch_size=32)\n<\/code><\/pre>\n\n<p>Detail: 'nonPdr' is one of my 2 img classes<\/p>\n\n<p><strong>UPDATE<\/strong> My readImg method<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def ReadImages(Path):\n    ImageList = list()\n    LabelList = list()\n    ImageCV = list()\n\n    # Get all subdirectories\n    FolderList = os.listdir(Path)\n\n    # Loop over each directory\n    for File in FolderList:\n        if(os.path.isdir(os.path.join(Path, File))):\n            for Image in os.listdir(os.path.join(Path, File)):\n                # Add the image path to the list\n                ImageList.append(os.path.join(Path, File) + os.path.sep + Image)\n                # Convert the path into a file\n                ImageCV.append(cv2.imread(os.path.join(Path, File) + os.path.sep + Image))    \n                # Add a label for each image and remove the file extension\n                LabelList.append(os.path.splitext(File)[0])\n        else:\n            ImageList.append(os.path.join(Path, File))\n\n            ImageCV.append(cv2.imread(os.path.join(Path, File) + os.path.sep + Image))    \n            # Add a label for each image and remove the file extension\n            LabelList.append(os.path.splitext(File)[0])\n\n    return ImageCV, LabelList\n<\/code><\/pre>\n","answers":[{"AnswerId":"57600828","CreationDate":"2019-08-22T01:04:41.687","ParentId":null,"OwnerUserId":"6013016","Title":null,"Body":"<p>In your ReadImages function, you are making list of strings:<\/p>\n\n<pre><code>LabelList.append(os.path.splitext(File)[0])\n<\/code><\/pre>\n\n<p>And later when you use ReadImages function, you are trying to convert this list of strings into numpy array. Here:<\/p>\n\n<pre><code>data, labels = ReadImages(TRAIN_DIR)\nmodel.fit(np.array(data), np.array(labels), epochs=10, batch_size=32)\n<\/code><\/pre>\n\n<p>Possible solution might be assigning your class names to numbers:<\/p>\n\n<pre><code>classes = [\"nonPdr\", \"another_class\"]\nLabelList.append(classes.index[os.path.splitext(File)[0]])\n<\/code><\/pre>\n\n<p>0 will be appended to LabelList when your class is \"nonPdr\", and 1 will be append if your class is \"other_class\".<\/p>\n"}]}
+{"QuestionId":57578929,"AnswerCount":0,"Tags":"<tensorflow><keras><gpu><nvidia><python-3.7>","CreationDate":"2019-08-20T17:37:52.040","AcceptedAnswerId":null,"OwnerUserId":11953009.0,"Title":"GPUs not being used while running Keras code","Body":"<p>I need to know how to run the keras code on GPUs. This is for a Ubuntu server. P.S. I am just starting to figure out stuff in Deep Learning and running on GPU.<\/p>\n\n<p>I've seen some code to execute in python3 to display the cpu and gpus that you can use. This is what it shows<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/QEGn6.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/QEGn6.png\" alt=\"SNAP 01\"><\/a><\/p>\n\n<p>This is the output after nvidia-smi(It's a college server xD)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/wF7HV.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/wF7HV.png\" alt=\"SNAP 2\"><\/a><\/p>\n\n<p>In my python code, I have included this to run on GPU<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/go0pm.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/go0pm.png\" alt=\"SNAP 3\"><\/a><\/p>\n\n<p>Output<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/hmavB.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/hmavB.png\" alt=\"SNAP 4\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57579049,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><tfrecord>","CreationDate":"2019-08-20T17:46:58.627","AcceptedAnswerId":null,"OwnerUserId":11809627.0,"Title":"Trouble with alternating between loss functions and datasets when building a Multi-Task Neural Network","Body":"<p>I am trying to train a multi-task neural network with Tensorflow. I have two datasets, each stored in its own TFRecord. My idea was to alternate between the datasets using a feedable iterator, but my implementation (code below) seems to not work. The following exception occurs: <\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors_impl.InvalidArgumentError: Input to reshape is a tensor with 50331648 values, but the requested shape requires a multiple of 34\n       [[{{node softmax_cross_entropy_with_logits_3\/Reshape_1}}]]\n       [[{{node tlr\/class_4_loss\/Mean}}]]<\/p>\n<\/blockquote>\n\n<p>I guess the problem is that loss2 function is also evaluated in the first session.run()-call inside the for-loop when the variable <em>label<\/em> points to the label of dataset1 which has a different shape than loss2 expects. <\/p>\n\n<p>Below you can find my current implementation:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with tf.Session() as sess:\n    # dataset1 consists of tuples (image, label) with \n    # image.shape=(1,1024,2048,3) and label.shape=(1,1024,2048,24)\n    # dataset2 consists of tuples (image, label) with \n    # image.shape=(1,1024,2048,3) and label.shape=(1,1024,2048,34)\n    iterator1 = dataset1.make_one_shot_iterator()\n    iterator2 = dataset2.make_one_shot_iterator()\n\n    data_iterator_handle = tf.placeholder(tf.string, shape=[])\n    data_iterator = tf.data.Iterator.from_string_handle(data_iterator_handle, dataset1.output_types, (tf.TensorShape([1, 1024, 2048, 3]), tf.TensorShape([1, 1024, 2048, None])))\n    iterator1_handle = sess.run(iterator1.string_handle())\n    iterator2_handle = sess.run(iterator2.string_handle())\n\n    image, label = data_iterator.get_next()\n    inference_op = model(image)\n\n    loss1_op = loss1(label)  # loss1 needs label with shape [1, 1024, 2048, 24]\n    optimizer1_op = tf.train.AdamOptimizer(learning_rate=0.1).minimize(loss1_op)\n    loss2_op = loss2(label)  # loss2 needs label with shape [1, 1024, 2048, 34]\n    optimizer2_op = tf.train.AdadeltaOptimizer(learning_rate=0.1).minimize(loss2_op)\n\n    sess.run(tf.global_variables_initializer())\n    for step in range(10):\n        if step % 2 == 0:\n            prediction, loss, _ = sess.run([inference_op, loss1_op, optimizer1_op], feed_dict={data_iterator_handle: iterator1_handle})\n        else:\n            prediction, loss, _ = sess.run([inference_op, loss2_op, optimizer2_op], feed_dict={data_iterator_handle: iterator2_handle})\n<\/code><\/pre>\n\n<p>i would really appreciate any advice on how to solve this.<\/p>\n\n<p><strong>EDIT<\/strong><br>\nSo i figured out how to do this. First, it seems to be a bad idea to use a feedable iterator together with one shot iterators, because they will be rebuild (?) ever time you change the handle, which takes a really long time. <\/p>\n\n<p>The way i got it working was by creating three placeholders, one for the input and one for each label. These placeholders will be fed via the feed_dict mechanism in session.run(). That means i have to make another session.run() call first to get the actual data from the tfrecords. I also have to always feed both labels in each run()-call. Below you can find some code: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with tf.Session() as sess:\n    # dataset1 consists of tuples (image, label) with \n    # image.shape=(1,1024,2048,3) and label.shape=(1,1024,2048,24)\n    # dataset2 consists of tuples (image, label) with \n    # image.shape=(1,1024,2048,3) and label.shape=(1,1024,2048,34)\n    iterator1 = dataset1.make_one_shot_iterator()\n    iterator2 = dataset2.make_one_shot_iterator()\n\n    input_placeholder = tf.placeholder(tf.float32, shape=[1, 1024,2048, 3])\n    label1_placeholder = tf.placeholder(tf.float32, shape=[1, 1024,2048, 24])\n    label2_placeholder = tf.placeholder(tf.float32, shape=[1, 1024,2048, 34])\n\n    data1_image_iter, data1_label_iter = iterator1.get_next()\n    data2_image_iter, data2_label_iter = iterator2.get_next()\n\n    inference_op = model(input_placeholder)\n\n    loss1_op = loss1(label1_placeholder)  # loss1 needs label with shape [1, 1024, 2048, 24]\n    optimizer1_op = tf.train.AdamOptimizer(learning_rate=0.1).minimize(loss1_op)\n    loss2_op = loss2(label2_placeholder)  # loss2 needs label with shape [1, 1024, 2048, 34]\n    optimizer2_op = tf.train.AdadeltaOptimizer(learning_rate=0.1).minimize(loss2_op)\n\n    sess.run(tf.global_variables_initializer())\n    for step in range(10):\n        if step % 2 == 0:\n            image1, label1 = sess.run([data1_image_iter, data1_label_iter])\n            image2, image2 = sess.run([data2_image_iter, data2_label_iter])\n\n            prediction, loss, _ = sess.run([inference_op, loss1_op, optimizer1_op], feed_dict={input_placeholder: image1, \n                                          label1_placeholder: label1, label2_placeholder: label2})\n        else:\n            prediction, loss, _ = sess.run([inference_op, loss2_op, optimizer2_op], feed_dict={input_placeholder: image2, \n                                          label1_placeholder: label1, label2_placeholder: label2})\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57579274,"AnswerCount":0,"Tags":"<python><keras><classification><baseline>","CreationDate":"2019-08-20T18:02:00.787","AcceptedAnswerId":null,"OwnerUserId":10852215.0,"Title":"How to use Keras functions to calculate baseline loss and accuracy for binary classification?","Body":"<p>I am trying to calculate some baselines using my validation outputs. I know for a baseline cross entropy, i believe you are supposed to compare each actual value, to an average of all the values. For accuracy, i believe you are supposed to compare actual values with the most common class.<\/p>\n\n<p>I see that accuracy looks to be working fine, but i'm not sure i'm calculating baseline binary crossentropy correctly. As stated above, i'm basically using an average of all the outputs as the baseline. For example:<\/p>\n\n<pre><code>actual_outputs = [1, 1, 1, 0]\nbaseline_outputs = [0.75, 0.75, 0.75, 0.75]\nbinary_crossentropy(actual_outputs, baseline_outputs)\n<\/code><\/pre>\n\n<p>Below i am basically computing a tensor for the actual outputs, a tensor filled with the output average, and a tensor filled with the most common class...<\/p>\n\n<pre><code>tensor_outputs_avg = tf.constant(np.full((len(val_outputs)), np.average(val_outputs)), dtype=tf.float64)\ntensor_outputs_avg_rounded = tf.constant(np.full((len(val_outputs)), round(np.average(val_outputs))), dtype=tf.float64)\ntensor_outputs = tf.constant(val_outputs, dtype=tf.float64)\n\nprint('baseline loss:', tf.Session().run(binary_crossentropy(tensor_outputs_avg, tensor_outputs)))\nprint('baseline acc:', tf.Session().run(binary_accuracy(tensor_outputs_avg_rounded, tensor_outputs)))\n<\/code><\/pre>\n\n<p>What am i doing wrong?<\/p>\n","answers":[]}
+{"QuestionId":57579641,"AnswerCount":2,"Tags":"<tensorflow><out-of-memory><conv-neural-network><tf.keras>","CreationDate":"2019-08-20T18:30:32.603","AcceptedAnswerId":57579951.0,"OwnerUserId":8894647.0,"Title":"my question is that what changes there should be in a code","Body":"<h2>problem<\/h2>\n\n<p>I am training CNN model on my GPU with tensorflow but I am running out of memory<\/p>\n\n<h2>things I tried<\/h2>\n\n<p>I have tried changing my batch_size , there was a positive change but eventually it was out of memory<\/p>\n\n<p>model = Sequential()<\/p>\n\n<h2>CODE<\/h2>\n\n<p><code>enter code here<\/code><\/p>\n\n<pre><code>model.add(Conv2D(64, (3, 3), input_shape=X.shape[1:]))\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))\n\nmodel.add(Dense(1))\nmodel.add(Activation(\"sigmoid\"))\n\nmodel.compile(loss=\"binary_crossentropy\",optimizer=\"adam\",metrics= \n['accuracy'])\nmodel.fit(X, Y, batch_size=32, validation_split=0.1)\n<\/code><\/pre>\n\n<h2>OUTPUT<\/h2>\n\n<pre><code>C:\\Anaconda3\\envs\\tutorial\\pythonw.exe \"C:\/Users\/roshaan zafar\/PycharmProjects\/InternshipRiseTech\/main.py\"\nWARNING: Logging before flag parsing goes to stderr.\nW0820 13:05:23.726494 24488 deprecation.py:506] From C:\\Anaconda3\\envs\\tutorial\\lib\\site-packages\\tensorflow\\python\\ops\\init_ops.py:1251: calling VarianceScaling.__init__ (from tensorflow.python.ops.init_ops) with dtype is deprecated and will be removed in a future version.\nInstructions for updating:\nCall initializer instance with the dtype argument instead of passing it to the constructor\nW0820 13:05:23.817250 24488 deprecation.py:323] From C:\\Anaconda3\\envs\\tutorial\\lib\\site-packages\\tensorflow\\python\\ops\\nn_impl.py:180: add_dispatch_support.&lt;locals&gt;.wrapper (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse tf.where in 2.0, which has the same broadcast rule as np.where\nTrain on 360 samples, validate on 40 samples\n2019-08-20 13:05:24.028720: I tensorflow\/core\/platform\/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2\n2019-08-20 13:05:24.030744: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library nvcuda.dll\n2019-08-20 13:05:24.976333: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1640] Found device 0 with properties: \nname: GeForce GTX 1070 major: 6 minor: 1 memoryClockRate(GHz): 1.645\npciBusID: 0000:01:00.0\n2019-08-20 13:05:24.976601: I tensorflow\/stream_executor\/platform\/default\/dlopen_checker_stub.cc:25] GPU libraries are statically linked, skip dlopen check.\n2019-08-20 13:05:24.977484: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1763] Adding visible gpu devices: 0\n2019-08-20 13:05:25.734584: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1181] Device interconnect StreamExecutor with strength 1 edge matrix:\n2019-08-20 13:05:25.734785: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1187]      0 \n2019-08-20 13:05:25.734905: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1200] 0:   N \n2019-08-20 13:05:25.735694: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1326] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 6376 MB memory) -&gt; physical GPU (device: 0, name: GeForce GTX 1070, pci bus id: 0000:01:00.0, compute capability: 6.1)\n2019-08-20 13:05:26.180767: W tensorflow\/core\/framework\/allocator.cc:107] Allocation of 1240006656 exceeds 10% of system memory.\n2019-08-20 13:05:26.834340: W tensorflow\/core\/framework\/allocator.cc:107] Allocation of 1240006656 exceeds 10% of system memory.\n2019-08-20 13:05:27.476075: W tensorflow\/core\/framework\/allocator.cc:107] Allocation of 1240006656 exceeds 10% of system memory.\n2019-08-20 13:05:28.102630: W tensorflow\/core\/framework\/allocator.cc:107] Allocation of 1240006656 exceeds 10% of system memory.\n2019-08-20 13:05:28.715843: W tensorflow\/core\/framework\/allocator.cc:107] Allocation of 1240006656 exceeds 10% of system memory.\n2019-08-20 13:05:47.982488: W tensorflow\/core\/common_runtime\/bfc_allocator.cc:314] Allocator (GPU_0_bfc) ran out of memory trying to allocate 9.34GiB (rounded to 10029662208).  Current allocation summary follows.\n2019-08-20 13:05:47.983224: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] Bin (256):   Total Chunks: 47, Chunks in use: 47. 11.8KiB allocated for chunks. 11.8KiB in use in bin. 1.5KiB client-requested in use in bin.\n2019-08-20 13:05:47.983956: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] 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.\n2019-08-20 13:05:47.984651: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] Bin (1024):  Total Chunks: 1, Chunks in use: 1. 1.3KiB allocated for chunks. 1.3KiB in use in bin. 1.0KiB client-requested in use in bin.\n2019-08-20 13:05:47.985413: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] 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.\n2019-08-20 13:05:47.986243: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] Bin (4096):  Total Chunks: 2, Chunks in use: 2. 13.5KiB allocated for chunks. 13.5KiB in use in bin. 13.5KiB client-requested in use in bin.\n2019-08-20 13:05:47.988224: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] Bin (8192):  Total Chunks: 0, Chunks in use: 0. 0B allocated for chunks. 0B in use in bin. 0B client-requested in use in bin.\n2019-08-20 13:05:47.988864: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] 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.\n2019-08-20 13:05:47.989820: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] 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.\n2019-08-20 13:05:47.990495: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] 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.\n2019-08-20 13:05:47.991146: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] Bin (131072):    Total Chunks: 2, Chunks in use: 2. 288.0KiB allocated for chunks. 288.0KiB in use in bin. 288.0KiB client-requested in use in bin.\n2019-08-20 13:05:47.992567: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] Bin (262144):    Total Chunks: 0, Chunks in use: 0. 0B allocated for chunks. 0B in use in bin. 0B client-requested in use in bin.\n2019-08-20 13:05:47.993545: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] 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.\n2019-08-20 13:05:47.994186: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] 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.\n2019-08-20 13:05:47.994859: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] 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.\n2019-08-20 13:05:47.995569: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] Bin (4194304):   Total Chunks: 0, Chunks in use: 0. 0B allocated for chunks. 0B in use in bin. 0B client-requested in use in bin.\n2019-08-20 13:05:47.996235: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] 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.\n2019-08-20 13:05:47.996924: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] 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.\n2019-08-20 13:05:47.997650: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] 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.\n2019-08-20 13:05:47.998404: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] 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.\n2019-08-20 13:05:47.999135: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] 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.\n2019-08-20 13:05:47.999876: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:764] Bin (268435456):     Total Chunks: 5, Chunks in use: 3. 6.23GiB allocated for chunks. 2.75GiB in use in bin. 2.75GiB client-requested in use in bin.\n2019-08-20 13:05:48.000650: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:780] Bin for 9.34GiB was 256.00MiB, Chunk State: \n2019-08-20 13:05:48.001093: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:786]   Size: 450.00MiB | Requested Size: 450.00MiB | in_use: 0 | bin_num: 20, prev:   Size: 256B | Requested Size: 8B | in_use: 1 | bin_num: -1, next:   Size: 256B | Requested Size: 128B | in_use: 1 | bin_num: -1\n2019-08-20 13:05:48.003835: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:786]   Size: 3.04GiB | Requested Size: 0B | in_use: 0 | bin_num: 20, prev:   Size: 256B | Requested Size: 4B | in_use: 1 | bin_num: -1\n2019-08-20 13:05:48.004577: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:793] Next region of size 6686052608\n2019-08-20 13:05:48.013828: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705400000 next 1 of size 1280\n2019-08-20 13:05:48.014294: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705400500 next 2 of size 256\n2019-08-20 13:05:48.014708: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705400600 next 3 of size 256\n2019-08-20 13:05:48.015131: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705400700 next 4 of size 256\n2019-08-20 13:05:48.015622: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705400800 next 5 of size 256\n2019-08-20 13:05:48.016053: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705400900 next 6 of size 256\n2019-08-20 13:05:48.016492: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705400A00 next 7 of size 256\n2019-08-20 13:05:48.016914: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705400B00 next 8 of size 256\n2019-08-20 13:05:48.017347: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705400C00 next 9 of size 256\n2019-08-20 13:05:48.017774: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705400D00 next 10 of size 256\n2019-08-20 13:05:48.018202: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705400E00 next 11 of size 256\n2019-08-20 13:05:48.019604: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705400F00 next 12 of size 256\n2019-08-20 13:05:48.020000: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705401000 next 13 of size 256\n2019-08-20 13:05:48.020407: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705401100 next 14 of size 256\n2019-08-20 13:05:48.020801: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705401200 next 15 of size 256\n2019-08-20 13:05:48.021203: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705401300 next 16 of size 256\n2019-08-20 13:05:48.022177: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705401400 next 17 of size 256\n2019-08-20 13:05:48.022845: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705401500 next 18 of size 256\n2019-08-20 13:05:48.023458: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000705401600 next 19 of size 1240006656\n2019-08-20 13:05:48.024110: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 000000074F291600 next 20 of size 256\n2019-08-20 13:05:48.024721: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 000000074F291700 next 21 of size 147456\n2019-08-20 13:05:48.025371: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 000000074F2B5700 next 22 of size 6912\n2019-08-20 13:05:48.026024: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 000000074F2B7200 next 23 of size 256\n2019-08-20 13:05:48.026686: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 000000074F2B7300 next 24 of size 256\n2019-08-20 13:05:48.027396: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 000000074F2B7400 next 25 of size 1240006656\n2019-08-20 13:05:48.027798: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 0000000799147400 next 26 of size 147456\n2019-08-20 13:05:48.028202: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 000000079916B400 next 27 of size 6912\n2019-08-20 13:05:48.028598: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 000000079916CF00 next 28 of size 256\n2019-08-20 13:05:48.028990: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 000000079916D000 next 29 of size 256\n2019-08-20 13:05:48.029868: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 000000079916D100 next 30 of size 256\n2019-08-20 13:05:48.030492: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 000000079916D200 next 31 of size 256\n2019-08-20 13:05:48.030887: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 000000079916D300 next 32 of size 256\n2019-08-20 13:05:48.031538: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 000000079916D400 next 33 of size 256\n2019-08-20 13:05:48.031931: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 000000079916D500 next 34 of size 256\n2019-08-20 13:05:48.032327: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 000000079916D600 next 35 of size 256\n2019-08-20 13:05:48.032722: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 000000079916D700 next 36 of size 256\n2019-08-20 13:05:48.033116: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] Free  at 000000079916D800 next 37 of size 471859200\n2019-08-20 13:05:48.034291: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007B536D800 next 38 of size 256\n2019-08-20 13:05:48.034879: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007B536D900 next 39 of size 256\n2019-08-20 13:05:48.035434: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007B536DA00 next 40 of size 256\n2019-08-20 13:05:48.035832: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007B536DB00 next 41 of size 471859200\n2019-08-20 13:05:48.036554: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007D156DB00 next 42 of size 256\n2019-08-20 13:05:48.037253: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007D156DC00 next 43 of size 256\n2019-08-20 13:05:48.037949: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007D156DD00 next 44 of size 256\n2019-08-20 13:05:48.038697: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007D156DE00 next 45 of size 256\n2019-08-20 13:05:48.039204: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007D156DF00 next 46 of size 256\n2019-08-20 13:05:48.039676: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007D156E000 next 47 of size 256\n2019-08-20 13:05:48.040135: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007D156E100 next 48 of size 256\n2019-08-20 13:05:48.041145: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007D156E200 next 49 of size 256\n2019-08-20 13:05:48.041535: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007D156E300 next 50 of size 256\n2019-08-20 13:05:48.041819: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007D156E400 next 51 of size 256\n2019-08-20 13:05:48.042130: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007D156E500 next 52 of size 256\n2019-08-20 13:05:48.042426: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007D156E600 next 53 of size 256\n2019-08-20 13:05:48.042713: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007D156E700 next 54 of size 256\n2019-08-20 13:05:48.043016: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007D156E800 next 55 of size 256\n2019-08-20 13:05:48.043276: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] InUse at 00000007D156E900 next 56 of size 256\n2019-08-20 13:05:48.043572: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:800] Free  at 00000007D156EA00 next 18446744073709551615 of size 3261998848\n2019-08-20 13:05:48.043902: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:809]      Summary of in-use Chunks by size: \n2019-08-20 13:05:48.044196: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:812] 47 Chunks of size 256 totalling 11.8KiB\n2019-08-20 13:05:48.044466: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:812] 1 Chunks of size 1280 totalling 1.3KiB\n2019-08-20 13:05:48.044760: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:812] 2 Chunks of size 6912 totalling 13.5KiB\n2019-08-20 13:05:48.045032: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:812] 2 Chunks of size 147456 totalling 288.0KiB\n2019-08-20 13:05:48.045250: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:812] 1 Chunks of size 471859200 totalling 450.00MiB\n2019-08-20 13:05:48.045553: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:812] 2 Chunks of size 1240006656 totalling 2.31GiB\n2019-08-20 13:05:48.045830: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:816] Sum Total of in-use chunks: 2.75GiB\n2019-08-20 13:05:48.046120: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:818] total_region_allocated_bytes_: 6686052608 memory_limit_: 6686052843 available bytes: 235 curr_region_allocation_bytes_: 13372105728\n2019-08-20 13:05:48.046453: I tensorflow\/core\/common_runtime\/bfc_allocator.cc:824] Stats: \nLimit:                  6686052843\nInUse:                  2952194560\nMaxInUse:               3424053760\nNumAllocs:                      64\nMaxAllocSize:           1240006656\n\n2019-08-20 13:05:48.046834: W tensorflow\/core\/common_runtime\/bfc_allocator.cc:319] **************************************______********________________________________________________\n2019-08-20 13:05:48.052167: W tensorflow\/core\/framework\/op_kernel.cc:1502] OP_REQUIRES failed at conv_ops.cc:486 : Resource exhausted: OOM when allocating tensor with shape[32,64,1278,958] and type float on \/job:localhost\/replica:0\/task:0\/device:GPU:0 by allocator GPU_0_bfc\n\n\n2019-08-20 13:05:48.052167: W tensorflow\/core\/framework\/op_kernel.cc:1502] OP_REQUIRES failed at conv_ops.cc:486 : Resource exhausted: OOM when allocating tensor with shape[32,64,1278,958] and type float on \/job:localhost\/replica:0\/task:0\/device:GPU:0 by allocator GPU_0_bfc\nTraceback (most recent call last):\n  File \"C:\/Users\/roshaan zafar\/PycharmProjects\/InternshipRiseTech\/main.py\", line 109, in &lt;module&gt;\n    model.fit(X, Y, batch_size=32, validation_split=0.1)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57579951","CreationDate":"2019-08-20T18:53:30.193","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>The feature vector dimension in your network after flatten is 1278 x 958. You are going to have 64 (total filters) x 1278 x 958 x 64 (Dense units) variables (without considering bias variable) in your memory. That number is really huge to be handled by your GPU.<\/p>\n\n<p>Consider either resizing your input image to smaller size or else, consider adding more layers (Conv2d with maxpooling) in your network. Last option is to consider replacing flatten layer with GlobalMaxPooling or GlobalAveragePooling.<\/p>\n"},{"AnswerId":"57579938","CreationDate":"2019-08-20T18:52:02.930","ParentId":null,"OwnerUserId":"1471972","Title":null,"Body":"<p>Since your network is pretty small, and you are taking only 32 images per batch, it might be the case that your images are of very high resolution, in which case you can try following things<\/p>\n\n<ul>\n<li>Try reducing the size of images, but do take care of retaining original aspect ratio while doing this<\/li>\n<li>Try to extract random patches of lesser size images again with same resolution<\/li>\n<li>Lastly, you can try to reduce Batch_Size even more to 8 or 4, if above solutions don't work<\/li>\n<\/ul>\n\n<p>GPU memory generally does fills up quickly than a corresponding CPU having same amount of memory. Hope this helps.<\/p>\n"}]}
+{"QuestionId":57579686,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-20T18:33:55.837","AcceptedAnswerId":null,"OwnerUserId":11909156.0,"Title":"tensorflow create variable during map_fn step","Body":"<p>is there some way to create variables in a map_fn loop like shown in the code beneath? how can I solve this error while keeping a variable in the loop? the info log does not really help me either, so am I getting any concept of tensorflow fundamentally wrong here? [tensorflow 1.14.0, python 3.6.8]<\/p>\n\n<pre><code>import tensorflow as tf\n\n\n### function called in map_fn\ndef opt_variable(theta):\n    init_theta = lambda: theta\n    var_theta = tf.get_variable(dtype=tf.float32, initializer=tf.Variable(init_theta))\n    ### ... other steps which need variable type to optimize\n    return tf.constant(3.)  # some return\n\n\ndef iterate_over_cols(theta):\n    iter_cols = tf.range(5)\n    map_theta = tf.map_fn(lambda x: (opt_variable(theta[x])),\niter_cols, dtype=tf.float32 )\n    return map_theta\n\n\n### example run\nt_test = tf.convert_to_tensor([1.4, 3.1, 4.6, 6.3], dtype=tf.float32)\niterate_over_cols(t_test)\n<\/code><\/pre>\n\n<p>leads to this error:<\/p>\n\n<blockquote>\n  <p>ValueError: Cannot use 'map_18\/while\/strided_slice' as input to\n  'map_18\/while\/Variable\/Assign' because 'map_18\/while\/strided_slice' is\n  in a while loop. See info log for more details.<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":57580202,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch>","CreationDate":"2019-08-20T19:14:05.963","AcceptedAnswerId":57581187.0,"OwnerUserId":11354575.0,"Title":"What's the purpose of torch.autograd.Variable?","Body":"<p>I load <code>features<\/code> and <code>labels<\/code> from my training dataset. Both of them are originally numpy arrays, but I change them to the torch tensor using <code>torch.from _numpy(features.copy())<\/code> and <code>torch.tensor(labels.astype(np.bool))<\/code>.<\/p>\n\n<p>And I notice that <code>torch.autograd.Variable<\/code> is something like <code>placeholder<\/code> in <code>tensorflow<\/code>. <\/p>\n\n<p>When I train my network, first I tried<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>features = features.cuda()\nlabels = labels.cuda()\n\noutputs = Config.MODEL(features)\nloss = Config.LOSS(outputs, labels)\n<\/code><\/pre>\n\n<p>Then I tried<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>features = features.cuda()\nlabels = labels.cuda()\n\ninput_var = Variable(features)\ntarget_var = Variable(labels)\noutputs = Config.MODEL(input_var)\nloss = Config.LOSS(outputs, target_var)\n<\/code><\/pre>\n\n<p>Both blocks succeed in activating training, but I worried that there might be trivial difference.<\/p>\n","answers":[{"AnswerId":"57581187","CreationDate":"2019-08-20T20:35:16.563","ParentId":null,"OwnerUserId":"3165451","Title":null,"Body":"<p>According to <a href=\"https:\/\/www.quora.com\/What-is-the-difference-between-a-Tensor-and-a-Variable-in-Pytorch\" rel=\"nofollow noreferrer\">this question<\/a> you no longer need variables to use Pytorch Autograd.<\/p>\n\n<blockquote>\n  <p>Autograd automatically supports Tensors with <code>requires_grad<\/code> set to <code>True<\/code>.<\/p>\n<\/blockquote>\n\n<p>And more importantly<\/p>\n\n<blockquote>\n  <p><code>Variable(tensor)<\/code> and <code>Variable(tensor, requires_grad)<\/code> still work as expected, but they return Tensors instead of Variables.<\/p>\n<\/blockquote>\n\n<p>This means that if your <code>features<\/code> and <code>labels<\/code> are tensors already (which they seem to be in your example) your <code>Variable(features)<\/code> and <code>Variable(labels)<\/code> does only return a tensor again.<\/p>\n\n<p>The original purpose of Variables was to be able to use automatic differentiation (<a href=\"https:\/\/discuss.pytorch.org\/t\/what-is-the-difference-between-tensors-and-variables-in-pytorch\/4914\" rel=\"nofollow noreferrer\">Source<\/a>):<\/p>\n\n<blockquote>\n  <p>Variables are just wrappers for the tensors so you can now easily auto compute the gradients.<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":57580647,"AnswerCount":1,"Tags":"<python><tensorflow><tfrecord>","CreationDate":"2019-08-20T19:49:20.057","AcceptedAnswerId":null,"OwnerUserId":11761580.0,"Title":"How to Convert Reading of SequenceExample Objects from tf.python_io.tf_record_iterator to tf.data.TFRecordDataset","Body":"<p>So I have a dataset in the TFRecords format, and I am trying to convert reading the dataset with tf.python_io.tf_record_iterator to tf.data.TFRecordDataset.<\/p>\n\n<p>Outside of tf.python_io.tf_record_iterator being deprecated, the main reason for doing this is that I would like to be able to use tf.data.Dataset objects.<\/p>\n\n<p>Within the TFRecords file, each entry is a SequenceExample, specifically tensorflow.core.example.example_pb2.SequenceExample.<\/p>\n\n<p>Currently I am reading out each SequenceExample via this function:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def read_records(record_path):\n    records = []\n    record_iterator = tf.python_io.tf_record_iterator(path=record_path)\n\n    for string_record in record_iterator:\n        example = tf.train.SequenceExample()\n        example.ParseFromString(string_record)\n        records.append(example)\n\n    return records\n<\/code><\/pre>\n\n<p>Printing out a record gives me this kind of structure (truncated due to length):<\/p>\n\n<pre><code>context {\n  feature {\n    key: \"framecount\"\n    value {\n      int64_list {\n        value: 10\n      }\n    }\n  }\n  feature {\n    key: \"label\"\n    value {\n      int64_list {\n        value: 1\n      }\n    }\n  }\n}\nfeature_lists {\n  feature_list {\n    key: \"positions\"\n    value {\n      feature {\n        bytes_list {\n          value: \"\\221\\2206?\\200dL?\\000\\000\\000\\000\\000\\000\\000\\000\\000\\000\\000\\000\\000\\000\\000\\000\"\n        }\n      }\n    }\n  }\n}\n<\/code><\/pre>\n\n<p>Now if I attempt to do this with tf.data.TFRecordDataset, my function is:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def reader(file_path):\n    dataset = tf.data.TFRecordDataset(file_path)\n    for record in dataset:\n        tf.io.parse_sequence_example(record)\n\n    return dataset\n<\/code><\/pre>\n\n<p>I am given a value error, stating that I have not supplied value or context features. Which is true, because the record has said values. (I have additionally attempted to follow the same flow for the first function with training a new SequenceExample, though it seems the data TFRecordDataset outputs is different from the old record iterator).<\/p>\n\n<p>Given this, how would I properly generate my sequenceExample? Though I could technically give it parameters to work with, this seems counter intuitive especially since the data is already in the record.<\/p>\n\n<p>Alternatively, (though this would be more of a band-aid fix) how could I convert the list in the first function into a tensorflow dataset object?<\/p>\n","answers":[{"AnswerId":"57582380","CreationDate":"2019-08-20T22:43:16.273","ParentId":null,"OwnerUserId":"11761580","Title":null,"Body":"<p>Okay, so this one was a little tricky...<\/p>\n\n<p>It seems that tf.python_io.tf_record_iterator outputs the data in a direct binary format that SequenceExample.FromString() can parse. On the other hand, TFRecordDataset returns the data in a direct tensor format.<\/p>\n\n<p>Since my intent was to be able to pass my datapoints to a model via the built-in generator ability of the Dataset object, I can get around it by wrapping the output of TFRecordDataset. Specifically, I can use SequenceExample.FromString(datapoint.numpy()) to get the desired output.<\/p>\n\n<p>This is a little wordy, so my solution function follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def reader(file_path):\n    dataset = tf.data.TFRecordDataset(file_path)\n    for record in dataset:\n        record = tf.train.SequenceExample.FromString(record.numpy())\n        yield record\n<\/code><\/pre>\n\n<p>This is a direct modification of the second function in my question<\/p>\n"}]}
+{"QuestionId":57581189,"AnswerCount":0,"Tags":"<python><keras><deep-learning>","CreationDate":"2019-08-20T20:35:32.443","AcceptedAnswerId":null,"OwnerUserId":5079088.0,"Title":"Triplet loss with custom checkpoint","Body":"<p>I have a triplet learning NN written on Keras. It uses a relatively small set of images and as a sad side effect can easily show accuracy<\/p>\n\n<pre><code>def accuracy(y_true, y_pred):\nreturn K.mean(y_pred[:,0,0] &lt; y_pred[:,1,0])\n<\/code><\/pre>\n\n<p>equal 1. In the same time, there is a training accuracy, too. And it may or may not be 1 in the same time.\nSo how do I write a checkpoint for saving \"best NN\" that has BOTH training and validation accuracy equal 1?<\/p>\n\n<p>Like this pseudo code:<\/p>\n\n<pre><code>if(valid_acc &gt; nBestValidationAccuracy\n  or (valid_acc == 1.0 and train_acc &gt; nBestTrainAccuracy)\n  or (valid_acc == 1.0 and train_acc == 1.0 and valid_loss &lt; nBestValidationLoss)\n  or (valid_acc == 1.0 and train_acc == 1.0 and valid_loss == nBestValidationLoss and train_loss &lt; nBestTrainLoss)):\n    nBestValidationAccuracy = valid_acc\n    nBestValidationLoss = valid_loss\n    nBestTrainAccuracy = train_acc\n    nBestTrainLoss = train_loss\n    triplet_model.save_weights(best_weights_filepath)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57581261,"AnswerCount":0,"Tags":"<python><tensorflow><jupyter-notebook><eager-execution>","CreationDate":"2019-08-20T20:41:35.770","AcceptedAnswerId":null,"OwnerUserId":845528.0,"Title":"Tensorflow Eager mode: First execution on GPU slow","Body":"<p>The code below compares computation time on the CPU vs GPU. Only for the first execution, I get slower runtime on GPU than CPU, and in all subsequent runs the GPU is much faster. Why is the first run on GPU slow? How do I make even the first run on GPU fast?<\/p>\n\n<pre><code>from __future__ import absolute_import, division, print_function\n\nimport tensorflow as tf\n\ntf.enable_eager_execution()\n\nimport time\n\ndef time_matmul(x):\n  start = time.time()\n  for loop in range(10):\n    tf.matmul(x, x)\n\n  result = time.time()-start\n\n  print(\"10 loops: {:0.2f}ms\".format(1000*result))\n\nprint(\"On GPU:\")\n# Force execution on GPU #0 if available\nif tf.test.is_gpu_available():\n  with tf.device(\"GPU:0\"): # Or GPU:1 for the 2nd GPU, GPU:2 for the 3rd etc.\n    x = tf.random_uniform([1000, 1000])\n    assert x.device.endswith(\"GPU:0\")\n    time_matmul(x)\n\n# Force execution on CPU\nprint(\"On CPU:\")\nwith tf.device(\"CPU:0\"):\n  x = tf.random_uniform([1000, 1000])\n  assert x.device.endswith(\"CPU:0\")\n  time_matmul(x)\n<\/code><\/pre>\n\n<p>Output on first run:<\/p>\n\n<pre><code>On GPU:\n10 loops: 443.04ms\nOn CPU:\n10 loops: 100.01ms\n<\/code><\/pre>\n\n<p>Output on subsequent runs:<\/p>\n\n<pre><code>On GPU:\n10 loops: 1.00ms\nOn CPU:\n10 loops: 103.01ms\n<\/code><\/pre>\n\n<p>PS: This is different from a seemingly <a href=\"https:\/\/stackoverflow.com\/questions\/56999493\/tensorflow-gpu-acceleration-only-happens-after-first-run?rq=1\">related question<\/a> because <code>tf.device(\"GPU:0\")<\/code> already chooses <code>\/device:GPU:0<\/code> and not <code>\/device:XLA_GPU:0<\/code><\/p>\n","answers":[]}
+{"QuestionId":57581436,"AnswerCount":1,"Tags":"<neural-network><classification><regression><pytorch>","CreationDate":"2019-08-20T20:57:21.463","AcceptedAnswerId":null,"OwnerUserId":1359137.0,"Title":"Difference in Model Performance when using Validation set\/ Testing set","Body":"<p>I have implemented a PyTorch NN code for classification and regression.<\/p>\n\n<p>Classification:\na) Use stratifiedKfolds for cross-validation (K=10- means 10 fold-cross validation)<\/p>\n\n<p>I divided the data: as follows:\nSuppose I have 100 data: 10 for testing, 18 for validation, 72 for training.<\/p>\n\n<p>b) Loss function = CrossEntropy<\/p>\n\n<p>c) Optimization = SGD<\/p>\n\n<p>d) Early Stopping where waittime = 100 epochs.<\/p>\n\n<pre><code>Problem is:\nBaseline Accuracy = 51%\nAccuracy on Training set = 100%\nAccuracy on validation set = 90%\nAccuracy on testing set = 72%\n<\/code><\/pre>\n\n<p>I don\u2019t understand what are the reasons behind the huge performance difference in Testing data\/ Validation data?<\/p>\n\n<p>How can I solve this problem?<\/p>\n\n<p>Regression:<\/p>\n\n<p>a) use the same network structure<\/p>\n\n<p>b) loss function = MSELoss<\/p>\n\n<p>c) Optimization = SGD<\/p>\n\n<p>d) Early Stopping where wait-time = 100 epochs.<\/p>\n\n<p>e) Use K-fold for cross-validation.<\/p>\n\n<p>I divided the data: as follows:<\/p>\n\n<p>Suppose I have 100 data: 10 for testing, 18 for validation, 72 for training.<\/p>\n\n<pre><code>Problem is:\nBaseline MSE= 14.0\nAccuracy on Training set = 0.0012\nAccuracy on validation set = 6.45\nAccuracy on testing set = 17.12\n<\/code><\/pre>\n\n<p>I don\u2019t understand what are the reasons behind the huge performance difference in Testing data\/ Validation data?<\/p>\n\n<p>How can I solve these problems? or Is this an obvious thing for NN\/ depend on particular dataset?<\/p>\n","answers":[{"AnswerId":"57581518","CreationDate":"2019-08-20T21:05:23.473","ParentId":null,"OwnerUserId":"1749629","Title":null,"Body":"<p>You have a large gap between training and validation performance, and between validation and test performance. There are two issues to explore:<\/p>\n\n<ol>\n<li><p>Differences in the distribution. We assume that train \/ val \/ test sets are all drawn from the same distribution, and so have similar characteristics. A well trained model should perform equally well on the val and test datasets. If you are dataset really is only 10 samples for test and 18 for val, there is a high chance that the samples selected will skew one\/both of these datasets, so that they no longer have similar characteristics. Therefore the difference between your val and test performance could just be chance: Your test set just happens to be much harder. You could test this by manual inspection.<\/p><\/li>\n<li><p>Overfitting to val: However, I think it is more likely that you have experimented with different architectures, training regimes, etc, and have tweaked parameters to get the best performance on you validation set. This means that you have overfit your model to your val set. The test set is a truer reflection of your model's accuracy.<\/p><\/li>\n<\/ol>\n\n<p>Your training accuracy is very high for both problems, and there is a large gap between training and validation performance. You are therefore overfitting to the training data, so need to train less, or introduce more stringent regularisation.<\/p>\n"}]}
+{"QuestionId":57581484,"AnswerCount":0,"Tags":"<tensorflow><recurrent-neural-network><gated-recurrent-unit>","CreationDate":"2019-08-20T21:01:28.740","AcceptedAnswerId":null,"OwnerUserId":8809654.0,"Title":"How to set up different initializations for biases of hidden layer, update gate, and reset gate?","Body":"<p>I am following this tutorial on tensorflow website:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/tutorials\/sequences\/text_generation\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/sequences\/text_generation<\/a><\/p>\n\n<p>and building up the model for text generation using GRU (Gated Recurrent Unit). When I set up a model like the following one:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Check if GPU is available:\nif tf.test.is_gpu_available():\n    rnn = tf.keras.layers.CuDNNGRU\nelse:\n    import functools\n    rnn = functools.partial(\n    tf.keras.layers.GRU, recurrent_activation='sigmoid')\n\n# Build model (a copy from tutorial)\ndef build_model(vocab_size, embedding_dim, rnn_units, batch_size):\n    model = tf.keras.Sequential()\n    model.add(tf.keras.layers.Embedding(vocab_size, embedding_dim, batch_input_shape=[batch_size, None]))\n    model.add(rnn(rnn_units, return_sequences=True, recurrent_initializer='glorot_uniform', stateful=True ))\n    model.add(tf.keras.layers.Dense(vocab_size))\n    return model\n<\/code><\/pre>\n\n<p>I know I can add a bias_initializer to GRU, but in GRU there should be individual biases including bias for update gate, bias for reset gate, and bias for hidden layers. Is it possible to use bias_initializer to specify three different initial values to biases of them? <\/p>\n","answers":[]}
+{"QuestionId":57581670,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><deep-learning>","CreationDate":"2019-08-20T21:20:36.743","AcceptedAnswerId":57583348.0,"OwnerUserId":7934786.0,"Title":"how can I get top smallest tensor values based on a condition in tensorflow","Body":"<p>I have a tensor like this:<\/p>\n\n<pre><code>sim_topics = [[0.65 0.   0.   0.   0.42  0.   0.   0.51 0.   0.34 0.]\n              [0.   0.51 0.   0.  0.52  0.   0.   0.   0.53 0.42 0.]\n              [0.    0.32 0.  0.50 0.34  0.   0.   0.39 0.32 0.52 0.]\n              [0.    0.23 0.37 0.   0.    0.37 0.37 0.   0.47 0.39 0.3 ]]\n<\/code><\/pre>\n\n<p>and one boolean tensor like this:<\/p>\n\n<pre><code>bool_t = [False  True  True  True]\n<\/code><\/pre>\n\n<p>I want to select part of <code>sim_topics<\/code> based on the bool flag in <code>bool_t<\/code> in a way it just select <code>top k smallest<\/code> values per row(if the row is true if not leave it as it is).<\/p>\n\n<p>So the expected output would be like this:(here <code>k=2<\/code>)<\/p>\n\n<pre><code>[[0.65 0.   0.    0.   0.42  0.   0.   0.51 0.   0.34 0.]\n [0.   0.51 0.   0.  0.52  0.   0.   0.   0.53 0.42 0.]\n [0.   0.32 0.    0.50 0    0    0.   0.   0    0.32 0 ]\n [0.   0.23 0     0.   0.    0    0    0.   0    0    0.3 ]]\n<\/code><\/pre>\n\n<p>I was trying to accomplish this first by using <code>boolean_mask<\/code> and <code>where<\/code> to get the indices I want then go get the top smallest. However, when I use <code>where<\/code> it does not give me the indices where there is  <code>zero<\/code>.<\/p>\n","answers":[{"AnswerId":"57583348","CreationDate":"2019-08-21T01:30:58.683","ParentId":null,"OwnerUserId":"8380272","Title":null,"Body":"<pre><code>k = 2\ndim0 = sim_topics.shape[0]\na = tf.cast(tf.equal(sim_topics,0), sim_topics.dtype)\nb = tf.reshape(tf.reduce_sum(a,1) + k, (dim0,-1))\nc = tf.cast(tf.argsort(tf.argsort(sim_topics,1),1), sim_topics.dtype)\nd = tf.logical_or(tf.less(c,b),tf.reshape(tf.logical_not(bool_t),(dim0,-1)))\nwith tf.Session() as sess:\n  print(sess.run(sim_topics * tf.cast(d,sim_topics.dtype)))\n\n\n[[0.65 0.   0.   0.   0.42 0.   0.   0.51 0.   0.34 0.  ]\n [0.   0.51 0.   0.   0.   0.   0.   0.   0.   0.42 0.  ]\n [0.   0.32 0.   0.   0.   0.   0.   0.   0.32 0.   0.  ]\n [0.   0.23 0.   0.   0.   0.   0.   0.   0.   0.   0.3 ]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57581740,"AnswerCount":1,"Tags":"<tensorflow><cuda><nvidia>","CreationDate":"2019-08-20T21:28:21.840","AcceptedAnswerId":57585376.0,"OwnerUserId":6437725.0,"Title":"Why Tensorflow model inference on GPU incurs so many CUDA cuEventRecord API calls?","Body":"<p>I run a Tensorflow object detection model(one of <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/detection_model_zoo.md\" rel=\"nofollow noreferrer\">these models<\/a>) on one GPU (Tesla P100). To examine the performance bottleneck, I used Nvidia <code>nvprof<\/code> profiling tool to profile my object detection application (performing object detection on a few frames). The profiling result is shown as follows.<\/p>\n\n<pre><code>======== Profiling result:\n  2             Type  Time(%)      Time     Calls       Avg       Min       Max  Name\n  3       API calls:   32.13%  15.2177s    434480  35.025us  5.1550us  954.27ms  cudaLaunchKernel\n  4                    30.20%  14.3065s    942706  15.175us     361ns  77.372ms  cuEventRecord\n  5                    13.39%  6.34349s    117067  54.186us  2.7000us  5.4721ms  cudaFuncGetAttributes\n  6                     6.26%  2.96509s    575202  5.1540us     562ns  1.2027ms  cuEventQuery\n  7                     6.16%  2.91725s     67072  43.494us  7.2690us  77.337ms  cuMemcpyDtoHAsync\n...\n<\/code><\/pre>\n\n<p>By looking at the Nvidia visual profiler, I found that the object detection application contains multiple threads. A couple of these threads keep invoking <code>cuEventRecord<\/code> CUDA driver API calls. The profiling result shows the duration of <code>cuEventRecord<\/code> API call is about 30% of the total duration time of cuda runtime+driver activities. I was wondering whether this <code>cuEventRecord<\/code> API call has something to do with the profiler: <code>nvprof<\/code>. If not, whether these <code>cuEventRecord<\/code> invocation would cause performance degradation for tensorflow model inference and what is the point to have these <code>cuEventRecord<\/code> API calls? \n<a href=\"https:\/\/i.stack.imgur.com\/NOjSs.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/NOjSs.png\" alt=\"Nvidia Visual profiler showing the &lt;code&gt;cuEventRecord&lt;\/code&gt; API invocation\"><\/a><\/p>\n","answers":[{"AnswerId":"57585376","CreationDate":"2019-08-21T06:07:03.010","ParentId":null,"OwnerUserId":"681865","Title":null,"Body":"<blockquote>\n  <p>I was wondering whether this\u00a0<code>cuEventRecord<\/code>\u00a0API call has something to\n  do with the profiler:\u00a0nvprof<\/p>\n<\/blockquote>\n\n<p>It does not.<\/p>\n\n<blockquote>\n  <p>If not, whether these\u00a0<code>cuEventRecord<\/code>\u00a0invocation would cause performance degradation for tensorflow model inference.<\/p>\n<\/blockquote>\n\n<p>They are part of the normal operation of Tensorflow.<\/p>\n\n<blockquote>\n  <p>what is the point to have these\u00a0cuEventRecord\u00a0API calls?<\/p>\n<\/blockquote>\n\n<p>As I understand it, Tensorflow has been designed with a heavily pipelined device code path which relies on extensive use of events, stream synchronization, and stream callback functions to ensure that the GPU(s) are kept occupied and that the different phases of computation are scheduled, uploaded, and downloaded in the correct order.. That is likely what you see here.<\/p>\n"}]}
+{"QuestionId":57582444,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><matrix><deep-learning><loss-function>","CreationDate":"2019-08-20T22:52:26.853","AcceptedAnswerId":57608322.0,"OwnerUserId":6174837.0,"Title":"How to assemble scalars into a matrix in tensorflow?","Body":"<p>In short, I want to assemble scalars <code>w_ij<\/code> into symmetric matrix <code>W<\/code> like so:<\/p>\n\n<pre><code>W[i, j] = w_ij\nW[j, i] = w_ij\n<\/code><\/pre>\n\n<p>After struggling with this and looking up material on the internet and here on SE, I can't find a way to construct matrix <code>W<\/code> from <code>w_ij<\/code>'s, I'm lost as to how to do this. Any help would be appreciated.<\/p>\n\n<p>Elaborations and MWE are below. <\/p>\n\n<hr>\n\n<hr>\n\n<h2>The problem<\/h2>\n\n<p>In my research I am trying to train a network that maps <code>source<\/code> to a scalar <code>w_ij<\/code>. Where the output <code>w_ij<\/code> is intended to represent an element <code>i,j<\/code> in a symmetric matrix <code>W<\/code>. <\/p>\n\n<p>So, the loss for training is constructed by assembling the outputs of many identical networks (with shared weights but each seeing a different input, and  driving a different element in a matrix) into a matrix form, like so:<\/p>\n\n<pre><code>W[i, j] = w_ij\nW[j, i] = w_ij\n<\/code><\/pre>\n\n<p>and then training those multiple networks on a loss of the form:<\/p>\n\n<pre><code>L2_loss(f(W) - f(True_W))\n<\/code><\/pre>\n\n<p>(Where <code>f()<\/code> is a function that runs <code>f(Y) = d' Y d<\/code> a quadratic form --- product of matrix by a fixed vector from left and right.) <\/p>\n\n<p>I need to run gradients through this loss to each network.<\/p>\n\n<hr>\n\n<hr>\n\n<h2>What I tried<\/h2>\n\n<ol>\n<li><p>A naive tensor slicing is not supported on <code>tensorflow<\/code>, i.e., <\/p>\n\n<p><code>W[i, j] = w_ij<\/code> is not supported.<\/p><\/li>\n<li><p>Using <code>tf.scatter_update()<\/code> does not allow to run gradients through it. <\/p><\/li>\n<li><p>Finally, I though I was close to a solution, I tried using <code>tf.Variable<\/code> for matrix <code>W<\/code>, like below: <\/p>\n\n<pre><code> W_flat = tf.Variable(initial_value=[0] * (2 * 2), dtype='float32')\n<\/code><\/pre>\n\n<p>and then assign to this <code>W_flat<\/code> by slicing it<code>W_falt[0].assign(w_ij)<\/code>, but it seems that my assignments to this variable do not work (see MWE). <\/p><\/li>\n<\/ol>\n\n<hr>\n\n<h2>MWE<\/h2>\n\n<p>Bellow is a short MWE, where <code>W<\/code> is a 2-by-2 symmetric matrix with zero diagonal, so I only have one independent element that a network has to drive (so here I only have a single network), i.e., I would like to get <code>W<\/code> to have the values  <\/p>\n\n<pre><code>W =   [[0, w_ij] [w_ij, 0]]\n<\/code><\/pre>\n\n<p>So I try update: <\/p>\n\n<pre><code>W_flat[1].assign(w_ij)\nW_flat[2].assign(w_ij)\n<\/code><\/pre>\n\n<p>And turn it back into a matrix: <\/p>\n\n<pre><code>W = tf.reshape(W_flat, (2, 2))\n<\/code><\/pre>\n\n<p>Eventually this update does not go through, the output of <code>print<\/code> shows that <code>W<\/code> remains all zeros. <\/p>\n\n<h3>The code<\/h3>\n\n<pre><code>import tensorflow as tf\n\ndef train():\n\n    with tf.Graph().as_default():\n        with tf.device('\/cpu'):\n            source = tf.placeholder(tf.float32, shape=(2, 3))\n            is_training = tf.placeholder(tf.bool, shape=())\n\n            w_ij = tf.reduce_sum(source)\n\n            W_flat = tf.Variable(initial_value=[0] * (2 * 2), dtype='float32')\n\n            W_flat[1].assign(w_ij)\n            W_flat[2].assign(w_ij)\n            tf.assign(W_flat[1], w_ij)\n            tf.assign(W_flat[2], w_ij)\n\n            W = tf.reshape(W_flat, (2, 2))\n\n        sess = tf.Session()\n        init = tf.global_variables_initializer()\n        sess.run(init, {is_training: True})\n\n        ops = {'W_flat': W_flat,\n               'source' : source,\n               'w_ij' : w_ij,\n               'W' : W}\n\n        for epoch in range(2):\n            feed_dict = {ops['source']: [[1,1,1], [7,7,7]]}\n            res_W_flat, res_wij, res_W = sess.run([ops['W_flat'], ops['w_ij'], ops['W']], feed_dict=feed_dict)\n            print(\"epoch:\" ,  epoch)\n            print(\"W_flat:\", res_W_flat)\n            print(\"wij:\", res_wij)\n            print(\"W:\", res_W)\n\nif __name__ == \"__main__\" :\n    train()\n<\/code><\/pre>\n\n<h3><code>print()<\/code> outputs<\/h3>\n\n<pre><code>epoch: 0\nW_flat: [0. 0. 0. 0.]\nwij: 24.0\nW: [[0. 0.]\n [0. 0.]]\nepoch: 1\nW_flat: [0. 0. 0. 0.]\nwij: 24.0\nW: [[0. 0.]\n [0. 0.]]\n<\/code><\/pre>\n\n<p>So <code>W<\/code> and <code>W_flat<\/code> is not updated by the value of <code>w_ij<\/code>, the value of which is 24 but <code>W<\/code> and <code>W_flat<\/code> remain zeros.  <\/p>\n","answers":[{"AnswerId":"57608322","CreationDate":"2019-08-22T11:27:01.617","ParentId":null,"OwnerUserId":"6174837","Title":null,"Body":"<p>The solution that I found after more struggles is to use <code>tf.scatter_nd()<\/code> to update the matrix <code>W<\/code>, unlike <code>tf.scatter_update()<\/code>, <code>tf.scatter_nd()<\/code> produces tensor which does support gradients propagation from its input to its output.<\/p>\n\n<p>So <strong>instead<\/strong> of writing <\/p>\n\n<pre><code>        W_flat = tf.Variable(initial_value=[0] * (2 * 2), dtype='float32')\n\n        W_flat[1].assign(w_ij)\n        W_flat[2].assign(w_ij)\n\n        W = tf.reshape(W_flat, (2, 2))\n<\/code><\/pre>\n\n<p>It <strong>worked using<\/strong>:<\/p>\n\n<pre><code>        W_flat = tf.Variable(initial_value=[0] * (2 * 2), dtype='float32')\n        indices = tf.constant([[1], [2]])\n        shape = tf.constant([4])\n        W_flat = tf.scatter_nd(indices, w_ij, shape)\n\n        W = tf.reshape(W_flat, (2, 2))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57582503,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><nan>","CreationDate":"2019-08-20T23:00:10.010","AcceptedAnswerId":null,"OwnerUserId":10229833.0,"Title":"VGG19 weights became NaN in all layers after single literation","Body":"<p>I am practicing neural style transfer with pre-trained model of VGG19. When I execute the training, after 1 iteration, all the weights in all layers of VGG19 become NaN. And when I skip the part where I do sess.run() on the cost functions, the weights remain the same, as expected. But does it mean that my cost functions are causing the issue?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def model_nst(sess, input_image, num_iterations = 2):\n\n    # Initialize global variables (you need to run the session on the initializer)\n    sess.run(tf.global_variables_initializer())\n\n    # Run the noisy input image (initial generated image) through the model.\n    sess.run(model['input'].assign(input_image))\n\n    for i in range(num_iterations):\n\n        # Run the session on the train_step to minimize the total cost\n        sess.run(train_step)\n\n        # Compute the generated image by running the session on the current model['input']\n        generated_image = sess.run(model['input'])\n        print(\"generated_image:\")\n        print(generated_image) #becomes NaN too after 1 iteration\n\n        if i%1 == 0:\n\n            #Jc = sess.run(J_content)\n            #Js = sess.run(J_style)\n            #Jt = sess.run(J)\n            #print(\"Iteration \" + str(i) + \" :\")\n            #print(\"total cost = \" + str(Jt))\n            #print(\"content cost = \" + str(Jc))\n            #print(\"style cost = \" + str(Js))\n\n            # save current generated image in the \"\/output\" directory\n            util.save_image(util.CONFIG.OUTPUT_DIR + str(i) + \".png\", generated_image)\n\n    # save last generated image\n    util.save_image('out\/generated_image.jpg', generated_image)\n\n    return generated_image\n<\/code><\/pre>\n","answers":[{"AnswerId":"57582913","CreationDate":"2019-08-21T00:11:58.410","ParentId":null,"OwnerUserId":"10697450","Title":null,"Body":"<p>What data are you using to perform transfer learning? The data might have NANs. This might also happen if your data is the wrong shape or type etc. <\/p>\n\n<p>If the above is checks out, then increase the learning rate or decrease the batch size. The initial weights result in vanishing or exploding gradients. Changing the batch size will bypass this problem.<\/p>\n"}]}
+{"QuestionId":57582591,"AnswerCount":1,"Tags":"<python><arrays><numpy><keras>","CreationDate":"2019-08-20T23:12:09.000","AcceptedAnswerId":57587212.0,"OwnerUserId":255412.0,"Title":"What are argument types for fit() function in keras?","Body":"<p>According to the <a href=\"https:\/\/keras.io\/models\/sequential\/\" rel=\"nofollow noreferrer\">Keras<\/a> documentation:<\/p>\n\n<pre><code>fit(x=None, y=None, batch_size=None, epochs=1, ...)\n<\/code><\/pre>\n\n<p>The arguments for fit function are numpy arrays:<\/p>\n\n<p>Arguments<\/p>\n\n<blockquote>\n  <p>x: 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. x can be None (default) if feeding from\n  framework-native tensors (e.g. TensorFlow data tensors).<\/p>\n  \n  <p>y: Numpy array of target (label) data (if the model has a single\n  output), or list of Numpy arrays (if the model has multiple outputs).\n  If output layers in the model are named, you can also pass a\n  dictionary mapping output names to Numpy arrays. y can be None\n  (default) if feeding from framework-native tensors (e.g. TensorFlow\n  data tensors).<\/p>\n<\/blockquote>\n\n<p>But I can pass <code>x<\/code> and <code>y<\/code> as dataframe and series.<\/p>\n\n<pre><code>print(type(x_train))\nprint(type(y_train))\n<\/code><\/pre>\n\n<blockquote>\n  <p><strong>Output:<\/strong><\/p>\n\n<pre><code>&lt;class 'pandas.core.frame.DataFrame'&gt;\n&lt;class 'pandas.core.series.Series'&gt;\n<\/code><\/pre>\n<\/blockquote>\n\n<pre><code>model.fit(x_train, y_train, ..)\n<\/code><\/pre>\n\n<p>This works fine. Why is that?<\/p>\n","answers":[{"AnswerId":"57587212","CreationDate":"2019-08-21T08:12:57.787","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Actually, this case has been covered in Keras source code by standardizing the input data in <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/c10d24959b0ad615a21e671b180a1b2466d77a2b\/keras\/engine\/training.py#L1084-L1089\" rel=\"nofollow noreferrer\">these lines<\/a>:<\/p>\n\n<pre><code>    # Case 2: Symbolic tensors or Numpy array-like.\n    x, y, sample_weights = self._standardize_user_data(\n        x, y,\n        sample_weight=sample_weight,\n        class_weight=class_weight,\n        batch_size=batch_size)\n<\/code><\/pre>\n\n<p>The <code>_standardize_user_data<\/code> method calls <code>standardize_input_data<\/code> function which converts DataFrame to Numpy array using using <code>.values<\/code> attribute of DataFrame in <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/c10d24959b0ad615a21e671b180a1b2466d77a2b\/keras\/engine\/training_utils.py#L93\" rel=\"nofollow noreferrer\">this line<\/a>:<\/p>\n\n<pre><code>    data = data.values if data.__class__.__name__ == 'DataFrame' else data\n<\/code><\/pre>\n\n<p>As for <code>Series<\/code>, this is done a bit more implicitly in <code>standardize_single_array<\/code> function in <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/30fe4ff1f12ff0c45bac8738b4d2690eadd056b2\/keras\/engine\/training_utils.py#L30\" rel=\"nofollow noreferrer\">this line<\/a>:<\/p>\n\n<pre><code>    elif x.ndim == 1:\n        x = np.expand_dims(x, 1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57582597,"AnswerCount":1,"Tags":"<tensorflow><keras><conv-neural-network>","CreationDate":"2019-08-20T23:13:03.950","AcceptedAnswerId":57582643.0,"OwnerUserId":9715052.0,"Title":"Value Error: No such layer - Extracting the output of a tensorflow keras layer","Body":"<p>I'm trying to extract the output of <code>thelayer<\/code> in my autoencoder and have referenced <a href=\"http:\/\/keras.io\/getting-started\/faq\/\" rel=\"nofollow noreferrer\">this<\/a> Keras documentation and <a href=\"https:\/\/stackoverflow.com\/questions\/41711190\/keras-how-to-get-the-output-of-each-layer\">this<\/a> stackoverflow post so far. When I try to extract the output, I get the following error: <\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"train.py\", line 36, in &lt;module&gt;\n    outputs=autoencoder.get_layer(layer_name).output)\n  File \"..Traceback (most recent call last):\n  File \"train.py\", line 36, in &lt;module&gt;\n    outputs=autoencoder.get_layer(layer_name).output)\n  File \"..python3.6\/site packages\/tensorflow\/python\/keras\/engine\/network.py\", line 567, in get_layer\n    raise ValueError('No such layer: ' + name)\nValueError: No such layer: thelayer\n\", line 567, in get_layer\n    raise ValueError('No such layer: ' + name)\nValueError: No such layer: thelayer\n<\/code><\/pre>\n\n<p>Code: <\/p>\n\n<pre><code>encoder_img = tf.keras.layers.Input(shape=(16,16,1), name=\"input\")\nx = tf.keras.layers.Conv2D(1024, 1, activation='relu',kernel_initializer=keras.initializers.RandomUniform)(encoder_img)\nx = tf.keras.layers.MaxPooling2D(1)(x)\ninputtothelayer = tf.keras.layers.Conv2D(512, 1, activation='relu')(x)\nthelayer = tf.keras.layers.MaxPooling2D(1)(inputtothelayer)\n\nbottleneck = tf.keras.layers.Conv2D(256, 3, activation='relu')(thelayer)\nx = tf.keras.layers.Conv2DTranspose(512, 1, activation='relu')(bottleneck)\nx = tf.keras.layers.UpSampling2D(1)(x)\nx = tf.keras.layers.Conv2DTranspose(1024, 1, activation='relu')(x)\nx = tf.keras.layers.UpSampling2D(1)(x)\ndecoder_output = tf.keras.layers.Conv2DTranspose(1, 3, activation='relu')(x)\nautoencoder = tf.keras.Model(inputs=encoder_img,outputs=decoder_output, name='autoencoder')\nautoencoder.fit(data, data,\n                epochs=1,\n                batch_size=512,\n                shuffle=True,)\nlayer_name = 'thelayer'\nintermediate_layer_model = autoencoder(inputs=inputtothelayer, outputs=autoencoder.get_layer(layer_name).output)\nintermediate_output = intermediate_layer_model.predict(data)\nprint(intermediate_layer_model)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57582643","CreationDate":"2019-08-20T23:22:35.180","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>Change the following lines from:<\/p>\n\n<pre><code>thelayer = tf.keras.layers.MaxPooling2D(1)(inputtothelayer)\nbottleneck = tf.keras.layers.Conv2D(256, 3, activation='relu')(thelayer)\n<\/code><\/pre>\n\n<p>to:<\/p>\n\n<pre><code>pool = tf.keras.layers.MaxPooling2D(1, name=\"thelayer\")(inputtothelayer)\nbottleneck = tf.keras.layers.Conv2D(256, 3, activation='relu')(pool)\n<\/code><\/pre>\n\n<p>If you want to retrieve the layer by name <code>model.get_layer(layer_name)<\/code>, you should include the layer name in the <code>name<\/code> attribute. Furthermore, if you want to obtain obtain the output from the intermediate layer, instead of doing:<\/p>\n\n<pre><code>layer_name = 'thelayer'\nintermediate_layer_model = autoencoder(inputs=inputtothelayer, outputs=autoencoder.get_layer(layer_name).output)\nintermediate_output = intermediate_layer_model.predict(data)\nprint(intermediate_layer_model)\n<\/code><\/pre>\n\n<p>Do the following:<\/p>\n\n<pre><code>layer_name = 'thelayer'\nintermediate_layer_model = tf.keras.Model(inputs=encoder_img, outputs=autoencoder.get_layer(layer_name).output)\nintermediate_output = intermediate_layer_model.predict(np.random.rand(10,16,16,1))\nprint(intermediate_layer_model)\n<\/code><\/pre>\n\n<p>Note that I am creating a new <code>tf.keras.Model<\/code> with the same <code>tf.keras.layers.Input<\/code>, where the output is the <code>intermediate_output<\/code>.<\/p>\n"}]}
+{"QuestionId":57582894,"AnswerCount":0,"Tags":"<python><python-2.7><tensorflow><keras>","CreationDate":"2019-08-21T00:08:58.680","AcceptedAnswerId":null,"OwnerUserId":2326901.0,"Title":"Import Error Traceback after \"import keras\"","Body":"<p>I have keras and tensorflow installed.<\/p>\n\n<p>However import keras throws this error<\/p>\n\n<blockquote>\n  <p>ImportErrorTraceback (most recent call last)\n   in ()\n  ----> 1 import keras<\/p>\n  \n  <p>\/util\/common\/python\/ondemand-jupyter\/py27\/anaconda-5.2.0\/lib\/python2.7\/site-packages\/keras\/<strong>init<\/strong>.py\n  in ()\n        1 from <strong>future<\/strong> import absolute_import\n        2 \n  ----> 3 from . import utils\n        4 from . import activations\n        5 from . import applications<\/p>\n  \n  <p>\/util\/common\/python\/ondemand-jupyter\/py27\/anaconda-5.2.0\/lib\/python2.7\/site-packages\/keras\/utils\/<strong>init<\/strong>.py\n  in ()\n        4 from . import data_utils\n        5 from . import io_utils\n  ----> 6 from . import conv_utils\n        7 \n        8 # Globally-importable utils.<\/p>\n  \n  <p>\/util\/common\/python\/ondemand-jupyter\/py27\/anaconda-5.2.0\/lib\/python2.7\/site-packages\/keras\/utils\/conv_utils.py\n  in ()\n        7 from six.moves import range\n        8 import numpy as np\n  ----> 9 from .. import backend as K\n       10 \n       11 <\/p>\n  \n  <p>\/util\/common\/python\/ondemand-jupyter\/py27\/anaconda-5.2.0\/lib\/python2.7\/site-packages\/keras\/backend\/<strong>init<\/strong>.py\n  in ()\n       87 elif _BACKEND == 'tensorflow':\n       88     sys.stderr.write('Using TensorFlow backend.\\n')\n  ---> 89     from .tensorflow_backend import *\n       90 else:\n       91     # Try and load external backend.<\/p>\n  \n  <p>\/util\/common\/python\/ondemand-jupyter\/py27\/anaconda-5.2.0\/lib\/python2.7\/site-packages\/keras\/backend\/tensorflow_backend.py in ()\n        3 from <strong>future<\/strong> import print_function\n        4 \n  ----> 5 import tensorflow as tf\n        6 from tensorflow.python.framework import ops as tf_ops\n        7 from tensorflow.python.training import moving_averages<\/p>\n  \n  <p>\/util\/common\/python\/ondemand-jupyter\/py27\/anaconda-5.2.0\/lib\/python2.7\/site-packages\/tensorflow\/<strong>init<\/strong>.py\n  in ()\n       20 \n       21 # pylint: disable=g-bad-import-order\n  ---> 22 from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n       23 \n       24 try:<\/p>\n  \n  <p>\/util\/common\/python\/ondemand-jupyter\/py27\/anaconda-5.2.0\/lib\/python2.7\/site-packages\/tensorflow\/python\/<strong>init<\/strong>.py\n  in ()\n       47 import numpy as np\n       48 \n  ---> 49 from tensorflow.python import pywrap_tensorflow\n       50 \n       51 # Protocol buffers<\/p>\n  \n  <p>\/util\/common\/python\/ondemand-jupyter\/py27\/anaconda-5.2.0\/lib\/python2.7\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py 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  \"\/util\/common\/python\/ondemand-jupyter\/py27\/anaconda-5.2.0\/lib\/python2.7\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\",\n  line 58, in \n      from tensorflow.python.pywrap_tensorflow_internal import *   File \"\/util\/common\/python\/ondemand-jupyter\/py27\/anaconda-5.2.0\/lib\/python2.7\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\",\n  line 28, in \n      _pywrap_tensorflow_internal = swig_import_helper()   File \"\/util\/common\/python\/ondemand-jupyter\/py27\/anaconda-5.2.0\/lib\/python2.7\/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) ImportError: libcuda.so.1: cannot open shared\n  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<hr>\n\n<p>How do I resolve this?<\/p>\n","answers":[]}
+{"QuestionId":57582987,"AnswerCount":0,"Tags":"<python-3.x><keras><deep-learning>","CreationDate":"2019-08-21T00:22:14.167","AcceptedAnswerId":null,"OwnerUserId":11882019.0,"Title":"Specifying gpus in keras backend without CUDA_VISIBLE","Body":"<p>I have a problem where I have to include both keras and pytorch frameworks in the same program. The frozen feature extractor is in Keras. When I run the Keras feature extractor in multi gpu mode (4 gpus) it takes all the gpus for its execution leaving no space for the second half (pytorch model) I would like to allocate 2 gpus to keras model and the other gpus to the pytorch model. If I set <code>cuda.visible_devices='0,1'<\/code>, it does not allow pytorch to use the other 3 gpus as well which is a sticky situation. \nAnother workaround is to set the keras backend session:<\/p>\n\n<pre><code>config = tf.ConfigProto( device_count = {'GPU': 1} )\nsess = tf.Session(config=config)\nkeras.backend.set_session(sess)\n<\/code><\/pre>\n\n<p>But here, I want to specify which gpus can be used by the session (by ID). \nThe error I am getting is CUDA out of memory error while pushing the data onto the gpu before the second half (pytorch section) because keras has already declared gpu occupancy.<\/p>\n\n<pre><code>RuntimeError: CUDA out of memory. Tried to allocate 204.00 MiB (GPU 2; 10.92 GiB total capacity; 14.54 MiB already allocated; 151.50 MiB free; 9.46 MiB cached)\n\n<\/code><\/pre>\n\n<p>I can specify GPUs with pytorch. I just want to do the same with keras.<\/p>\n","answers":[]}
+{"QuestionId":57583268,"AnswerCount":0,"Tags":"<python-3.x><keras><conv-neural-network><activation-function>","CreationDate":"2019-08-21T01:16:38.867","AcceptedAnswerId":null,"OwnerUserId":9813733.0,"Title":"Invalid Syntax in keras model.add Convolutional layers","Body":"<p>I am trying to build a VGG16 model but came across the <\/p>\n\n<blockquote>\n  <p>invalid syntax error<\/p>\n<\/blockquote>\n\n<p>while compiling it. Error is in the activation function of the below line. <\/p>\n\n<pre><code> model.add(Convolution2D((64,3,3,activation='relu')))\n<\/code><\/pre>\n\n<p>However, if I change the code as below, its working fine. <\/p>\n\n<pre><code>    model.add(Convolution2D((64,3,3)))\n    model.add(Activation('relu'))\n<\/code><\/pre>\n\n<p>I have seen many related questions and the answers say that it is due to the missing parenthesis in the above line of code, but I checked the parenthesis and they are fine. Why is the code giving me issues when i mention <strong><em>activation<\/em><\/strong> parameter in Convolution2D layer. I checked the <a href=\"http:\/\/faroit.com\/keras-docs\/1.2.2\/layers\/convolutional\/\" rel=\"nofollow noreferrer\">documentation<\/a> and it is supposed to accept the parameter. What am I missing here.<\/p>\n\n<p>Here is the detailed code<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers.core import Flatten,Dense,Dropout\nfrom keras.layers.convolutional import Convolution2D, MaxPooling2D, ZeroPadding2D\nfrom keras.layers import Activation\nfrom keras.optimizers import SGD\nimport cv2, numpy as np\n\ndef VGG16(weights_path=none):\n        model = Sequential()\n        model.add(ZeroPadding2D((1,1),input=(3,224,224)))\n        model.add(Convolution2D(64,3,3,activation='relu'))\n        model.add(ZeroPadding2D((1,1)))\n        model.add(Convolution2D(64,3,3,activation='relu'))\n        model.add(MaxPooling2D((2,2),strides=(2,2)))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57583421,"AnswerCount":0,"Tags":"<spring-boot><spring-cache><caffeine>","CreationDate":"2019-08-21T01:46:10.573","AcceptedAnswerId":null,"OwnerUserId":1012689.0,"Title":"Spring boot cache caffeine configs difference between expireAfterWrite and refreshAfterWrite","Body":"<p>For the Spring boot cache caffeine configurations I found out that <code>expireAfterWrite<\/code> and <code>refreshAfterWrite<\/code> are quite similar and either one is fit for our use case. Just given for specific input (key), don't invoke the actual method within certain time of period since the entry created.<\/p>\n\n<p>I found there is explanation on Caffeine user guide, but still not quite clear when to use both?<\/p>\n\n<blockquote>\n  <p>In contrast to expireAfterWrite, refreshAfterWrite will make a key\n  eligible for refresh after the specified duration, but a refresh will\n  only be actually initiated when the entry is queried. So, for example,\n  you can specify both refreshAfterWrite and expireAfterWrite on the\n  same cache so that the expiration timer on an entry isn't blindly\n  reset whenever an entry becomes eligible for a refresh. If an entry\n  isn't queried after it comes eligible for refreshing, it is allowed to\n  expire.<\/p>\n<\/blockquote>\n\n<p>Can anyone tell me the difference between those two and when to use them?<\/p>\n","answers":[]}
+{"QuestionId":57583793,"AnswerCount":0,"Tags":"<pytorch>","CreationDate":"2019-08-21T02:43:30.657","AcceptedAnswerId":null,"OwnerUserId":9222127.0,"Title":"Pytorch:DataParallel can't reduce model inference time with same total batch size iamges","Body":"<p>When I use <code>torch.nn.DataParallel()<\/code> to realize data parallel computation, I find the parallel model inference time for the same total amount batch size don't significantly decrease compared with serial model.<\/p>\n\n<p>As shown below.<\/p>\n\n<pre><code>model = SNModel()\ncriterion = nn.CrossEntropyLoss()\nif parallel_enable:\n    model = nn.DataParallel(model, device_ids=gpu_ids) # gpu_ids = [0,1,2,3] total 4 gpus\nmodel.to(args.device)\ncriterion.to(args.device)\n<\/code><\/pre>\n\n<p>calculate 100 iteration model inference time:<\/p>\n\n<pre><code>torch.cuda.synchronize()\nst = time.time()\noutputs, loss = model(images, path, labels, criterion, gpu_nums)\ntorch.cuda.synchronize()\ntotal_tm += time.time() - st\n<\/code><\/pre>\n\n<p>With single gpu:<\/p>\n\n<pre><code>100 iter model time  23s (around)\n<\/code><\/pre>\n\n<p>With 4 gpus:<\/p>\n\n<pre><code>100 iter model time  103s (around)\n<\/code><\/pre>\n\n<p>There is only <code>103 - 23*4  = 11s<\/code> reduced.<\/p>\n\n<p>Is there any wrong? Hope you can help me! I have been troubled by this for a long time. I know that there is something wrong that I didn't find out. This phenomenon has appeared in other models, so that I have always abandoned the use of parallelism.<\/p>\n","answers":[]}
+{"QuestionId":57583853,"AnswerCount":1,"Tags":"<python><tensorflow><keras><conv-neural-network>","CreationDate":"2019-08-21T02:53:16.543","AcceptedAnswerId":null,"OwnerUserId":9715052.0,"Title":"Output of to_categorical: What does the fourth dimensions represent?","Body":"<p>I'm building an autoencoder where I apply one hot encoding on the output of the encoder using <code>to_categorical<\/code>. The shape of my output is 5 dimensions: <code>(1000, 16, 16, 512, 1)<\/code>. <\/p>\n\n<p>My input shape is <code>(1000, 16, 16, 1)<\/code> so I'm wondering what this additional <code>512<\/code> is? <\/p>\n\n<p>For more context as to why I want to know, I want to feed the output of <code>to_categorical<\/code> into a bottleneck of shape: <code>bottleneck = tf.keras.layers.Conv2D(256, 3, activation='relu')(reshaped_one_hot)<\/code> and I want to understand how to reshape the output of<code>to_categorical<\/code> to then feed it into this bottleneck. <\/p>\n\n<p>Code:<\/p>\n\n<pre><code>encoder_img = tf.keras.layers.Input(shape=(16,16,1), name=\"input\")\nx = tf.keras.layers.Conv2D(1024, 1, activation='relu',kernel_initializer=keras.initializers.RandomUniform)(encoder_img)\nx = tf.keras.layers.MaxPooling2D(1)(x)\ninputtothelayer = tf.keras.layers.Conv2D(512, 1, activation='relu')(x)\npool = tf.keras.layers.MaxPooling2D(1, name=\"thelayer\")(inputtothelayer)\nencoder = tf.keras.Model(inputs=encoder_img, outputs=pool, name = 'encoder')\nencoder.summary()\n\nlayer_name = 'thelayer'\nintermediate_layer_model = tf.keras.Model(inputs=encoder_img,\n                                 outputs=encoder.get_layer(layer_name).output)\nintermediate_output = intermediate_layer_model.predict(data)\nprint(intermediate_output)\n\none_hot = tf.keras.utils.to_categorical(intermediate_output,num_classes=None, dtype='float32')\n\nbottleneck = tf.keras.layers.Conv2D(256, 3, activation='relu')(one_hot)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57584060","CreationDate":"2019-08-21T03:25:38.103","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p><code>to_categorical<\/code> is meant to transform \"integer labels\" into \"one-hot encoded labels\". <\/p>\n\n<p>If you have an array from 0 to 9: <code>[0,1,2,3,4,5,6,7,8,9]<\/code>, the <code>to_categorical<\/code> result will be:<\/p>\n\n<pre><code>[\n    [1,0,0,0,0,0,0,0,0,0],\n    [0,1,0,0,0,0,0,0,0,0],\n    [0,0,1,0,0,0,0,0,0,0],\n    [0,0,0,1,0,0,0,0,0,0],\n    [0,0,0,0,1,0,0,0,0,0],\n    [0,0,0,0,0,1,0,0,0,0],\n    [0,0,0,0,0,0,1,0,0,0],\n    [0,0,0,0,0,0,0,1,0,0],\n    [0,0,0,0,0,0,0,0,1,0],\n    [0,0,0,0,0,0,0,0,0,1]\n]\n<\/code><\/pre>\n\n<p>That said, it makes no sense to use <code>to_categorical<\/code> in a model's output. You are not outputting integer labels. <\/p>\n"}]}
+{"QuestionId":57583902,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-21T02:59:59.360","AcceptedAnswerId":57605264.0,"OwnerUserId":11954461.0,"Title":"OutOfRangeError: RandomShuffleQueue '_1_shuffle_batch\/random_shuffle_queue' is closed and has insufficient elements (requested 8, current size 7)","Body":"<p>firstly, I'm really thankful for you to read my problem. <\/p>\n\n<p>TensorFlow version: TensorFlow-GPU 1.14<\/p>\n\n<p>os: ubuntu 16.04<\/p>\n\n<p><code>tensorflow.python.framework.errors_impl.OutOfRangeError: RandomShuffleQueue '_1_shuffle_batch\/random_shuffle_queue' is closed and has insufficient elements (requested 8, current size 7)\n     [[{{node shuffle_batch}}]]<\/code><\/p>\n\n<p>I'm testing a kind of CNN called \"Goturn\", and I implement it with TensorFlow. When I use 100,000+ images to train my net, it always displays this error. the strange thing is that the error displays during the training other than at the first circle of training.<\/p>\n\n<p>I've already tried a lot of solutions from stackoverflow\u3001CSDN and the other websites, but it is no use. most of the suggestion is to modify the elements of tf.train.shuffle_batch(), whereas how I should choose them, is there any regulation to obey? Thanks for reading. Any suggestion will be appreciated<\/p>\n\n<pre><code>def next_batch(input_queue): \n    min_queue_examples = 8 \n    num_threads = 2 \n    [search_tensor, target_tensor, box_tensor] =      data_reader(input_queue) \n[search_batch, target_batch, box_batch] = tf.train.shuffle_batch( \n    [search_tensor, target_tensor, box_tensor],\n        batch_size=8, \n        num_threads=num_threads, \n        seed=88, \n        min_after_dequeue=min_queue_examples, \n        capacity=min_queue_examples+3*BATCH_SIZE)\n    print(\"next_batch!!!!!\") \n    return [search_batch, target_batch,\n    box_batch, input_queue[0], input_queue[1]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"57605264","CreationDate":"2019-08-22T08:36:29.080","ParentId":null,"OwnerUserId":"11954461","Title":null,"Body":"<p>OK, I make such a stupid mistake, in fact, my dataset has a horrible error: some images are null, and when I delete those images, the code runs well!!!!!<\/p>\n"}]}
+{"QuestionId":57583924,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><svm><text-classification>","CreationDate":"2019-08-21T03:03:35.663","AcceptedAnswerId":null,"OwnerUserId":7818476.0,"Title":"How to insert my messages as input in my feature columns in SVM estimator in the Tensorflow package","Body":"<p>I am trying to implement my text classification model which classifies messages as spam or not in Tensorflow using SVM Model. \nI have only 2 features rn that is a contact name and message, do I need to convert them into numbers or can it be directly used as input in feature column<\/p>\n\n<pre><code>svm = tf.contrib.learn.SVM(\n    example_id_column=example_id_column_name,\n    feature_columns=(tf.contrib.layers.real_valued_column(\n        column_name=x_column_name, dimension=90),),\n    l2_regularization=0.1)\n\nsvm.fit(input_fn=train_input_fn, steps=10)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57585267","CreationDate":"2019-08-21T05:57:26.103","ParentId":null,"OwnerUserId":"10232932","Title":null,"Body":"<p>Does it give an error? If it gives an error something like this: \"cannot convert string to float\", then you cannot use it. If it works, then you can use it.<\/p>\n\n<p>In the documentation it is also explained what you need:\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/learn\/SVM\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/learn\/SVM<\/a><\/p>\n"}]}
+{"QuestionId":57584360,"AnswerCount":0,"Tags":"<pytorch>","CreationDate":"2019-08-21T04:17:05.220","AcceptedAnswerId":null,"OwnerUserId":9222127.0,"Title":"Pytorch\uff1aHow to balance the backpropagation calculation with multiple gpus","Body":"<p>In order to balance calculation of loss backward, I not only calculate loss in forward() but also realese the loss backward.Here is my realease way.Is there any wrong?<\/p>\n\n<pre><code> def forward(self, x, y=None, criterion=None, gpu_nums=1):\n    x = self.task(x)\n    if not y == None:\n        loss = criterion(x, y)\n        if gpu_nums &gt; 1\uff1a\n            loss \/= gpu_nums\n        loss.backward()\n        return x, loss\n    return x\n<\/code><\/pre>\n\n<p>Is my implementation correct?If not, hope you can provide the right way.This is meaningful for everyone.<\/p>\n","answers":[]}
+{"QuestionId":57584644,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2019-08-21T04:54:48.677","AcceptedAnswerId":null,"OwnerUserId":7205164.0,"Title":"Predict from Previously trained Model Tensorflow","Body":"<p>I have trained a Regression Model using tensorflow. The model saved file like<\/p>\n\n<ol>\n<li>model.ckpt-4000.meta <\/li>\n<li>model.ckpt-4000.index      <\/li>\n<li>model.ckpt-4000.data-00001-of-00002<\/li>\n<\/ol>\n\n<p>Now if i want to use those file to predict output values from new data set (test data), how can i do that?<\/p>\n","answers":[{"AnswerId":"57585916","CreationDate":"2019-08-21T06:48:30.400","ParentId":null,"OwnerUserId":"7352626","Title":null,"Body":"<pre><code>sess=tf.Session()\n\n#First load meta graph and restore weights\n\nsaver = tf.train.import_meta_graph('model.ckpt-4000.meta')\nsaver.restore(sess,tf.train.latest_checkpoint('.\/'))\n\n# create feed-dict to feed new data and specify the y variable to be evaluated\nsess.run(y,feed_dict)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57585008,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><model><slim><pruning>","CreationDate":"2019-08-21T05:32:01.633","AcceptedAnswerId":null,"OwnerUserId":11925237.0,"Title":"I have use strip pruning vars.py to prune a slim model(.pb)\uff0cbut the speed of pruned model is slower than the unpruned model","Body":"<p>I have an Inception_resnet_v2 model to extract features. The speed of the model is about 3s, which is too slow for our goal.I use the pruning method layers.masked_conv2d which is posted in the url (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/model_pruning\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/model_pruning<\/a>) to prune our model.Finally, according to the procedure s, I adopt strip_pruning_vars to remove the pruning ops from our model, but the masked_weight matrix in pruned model is a sparsity matrix which consists of 0 and actual weight, which means that the computational complexity is the same as the previous mdoel. Besides, I have tested the pruned model which costs >3s to predict the features. Is there any people meet the same problem, and how to solve?<\/p>\n\n<p>mask code<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with tf.variable_scope('Mixed_6a'):\n                    with tf.variable_scope('Branch_0'):\n    tower_conv = layers.masked_conv2d(net, 384, 3, stride=2, padding='VALID',scope='Conv2d_1a_3x3')\n\nremove prune ops:\n\npython strip_pruning_vars --checkpoint_path=\/tmp\/cifar10_train --output_node_names=softmax_linear\/softmax_linear_2 --filename=cifar_pruned.pb\n\n\nI expect to speed up the model via remove the pruned ops\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57585336,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2019-08-21T06:03:59.240","AcceptedAnswerId":null,"OwnerUserId":8721137.0,"Title":"What method to use to learn a specific region in a signal data","Body":"<p>I am given with a set of following data:<\/p>\n\n<ul>\n<li>a list of nanopore signal data e.g. <code>[88.2, 92.5, ... 101.5]<\/code><\/li>\n<li>a numerical value that tells where a specific material is located within the signal data e.g. 5 => that specific thing is located in the 5th index in the corresponding signal data.<\/li>\n<\/ul>\n\n<p>I am trying to use <code>Keras<\/code> to implement Machine Learning with bidirectional <code>LSTM<\/code> models to learn the patterns between the signal data and the numerical index value.<\/p>\n\n<p>I have read some sample codes on internet and tried to use them, but I am facing some issues. Not even sure if my approach is appropriate.<\/p>\n\n<p>Since most of the codes I've read had equal dimensions of inputs and labels, I made an array of zeros that matches the length of signal data, and put 1 to the index value of the numerical value. <\/p>\n\n<pre><code>e.g. Signal = [1, 2, 4, 5]\n     Label = [0, 0, 1, 0]\n<\/code><\/pre>\n\n<p>After such modification of the data, I tried implementing bidirectional LSTM, but I am receiving Value errors<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># signal_dataset = (3912, 24160) -&gt; 3912 sequences with 24160 length\n# label_dataset = (3912, 24160) -&gt; 3912 sequences with 24160 length.\n\n# Conversion to 3 dimension\nsig_dataset = sig_dataset.reshape(1, len(sig_dataset), len(sig_dataset[0]))\nlab_dataset = lab_dataset.reshape(1, len(lab_dataset), len(lab_dataset[0]))\n\ntrain_size = int(len(sig_dataset) * 0.8)\ntrainX, trainY = sig_dataset[:train_size], adap_dataset[:train_size]\nvalidX, validY = sig_dataset[train_size:], adap_dataset[train_size:]\n\n# define LSTM\nmodel = Sequential()\nmodel.add(Bidirectional(LSTM(512, return_sequences=True), input_shape=(24160, 1)))\nmodel.add(TimeDistributed(Dense(1, activation='sigmoid')))\n\nmodel.compile(loss='mean_squared_error', \n              optimizer='adam', \n              metrics=['acc'])\n\nhist = model.fit(trainX, \n                 trainY, \n                 epochs=10, \n                 batch_size=1, \n                 validation_data=(validX, validY))\n<\/code><\/pre>\n\n<hr>\n\n<p>ValueError: Error when checking input: expected <code>bidirectional_32_input<\/code> to have 3 dimensions, but got array with <code>shape (3184, 24160)<\/code><\/p>\n\n<p>What am I doing wrong? Is this approach even correct? Please give me any advice<\/p>\n","answers":[]}
+{"QuestionId":57585463,"AnswerCount":0,"Tags":"<tensorflow><lstm><recurrent-neural-network>","CreationDate":"2019-08-21T06:14:44.350","AcceptedAnswerId":null,"OwnerUserId":2467772.0,"Title":"Using LSTM from Tensorflow and Memory issue","Body":"<p>I have 4 layers LSTM model as follow.<\/p>\n\n<pre><code>def LSTM_RNN(self,_X, _weights, _biases):\n      # model architecture based on \"guillaume-chevalier\" and \"aymericdamien\" under the MIT license.\n          _X = tf.transpose(_X, [1, 0, 2])  # permute n_steps and batch_size\n          _X = tf.reshape(_X, [-1, self.n_input])\n          # Rectifies Linear Unit activation function used\n          _X = tf.nn.relu(tf.matmul(_X, _weights['hidden']) + _biases['hidden'])\n      # Split data because rnn cell needs a list of inputs for the RNN inner loop\n          _X = tf.split(_X, self.n_steps, 0) \n\n      # Define two stacked LSTM cells (two recurrent layers deep) with tensorflow\n          lstm_cell_1 = tf.contrib.rnn.BasicLSTMCell(self.n_hidden, forget_bias=1.0, state_is_tuple=True)\n          lstm_cell_2 = tf.contrib.rnn.BasicLSTMCell(self.n_hidden, forget_bias=1.0, state_is_tuple=True)\n          lstm_cell_3 = tf.contrib.rnn.BasicLSTMCell(self.n_hidden, forget_bias=1.0, state_is_tuple=True)\n          lstm_cell_4 = tf.contrib.rnn.BasicLSTMCell(self.n_hidden, forget_bias=1.0, state_is_tuple=True)\n          lstm_cells = tf.contrib.rnn.MultiRNNCell([lstm_cell_1, lstm_cell_2, lstm_cell_3, lstm_cell_4], state_is_tuple=True)\n          outputs, states = tf.contrib.rnn.static_rnn(lstm_cells, _X, dtype=tf.float32)\n          lstm_last_output = outputs[-1]\n          return tf.matmul(lstm_last_output, _weights['out']) + _biases['out']\n<\/code><\/pre>\n\n<p>LSTM model is structured as<\/p>\n\n<pre><code>self.n_steps = 128 \nself.n_input = 10  #ch4.x,ch4.y,ch7.x,ch7.y,dist4_16,dist7_17 \nself.n_hidden = 34 # Hidden layer num of features\nself.n_classes = 3 \nself.global_step = tf.Variable(0, trainable=False)     \n# Graph input\/output\nself.x = tf.placeholder(tf.float32, [None, self.n_steps, self.n_input])\nself.y = tf.placeholder(tf.float32, [None, self.n_classes])\n# Graph weights\nself.weights = {\n           'hidden': tf.Variable(tf.random_normal([self.n_input, self.n_hidden])), # Hidden layer weights\n           'out': tf.Variable(tf.random_normal([self.n_hidden, self.n_classes], mean=1.0))\n      }\nself.biases = {\n          'hidden': tf.Variable(tf.random_normal([self.n_hidden])),\n          'out': tf.Variable(tf.random_normal([self.n_classes]))\n      }\n<\/code><\/pre>\n\n<p>I use V100 Tesla 16G GPU. When GPU memory is checked, 14750MiB is used.\nSo I can't use memory for others.\nWhat is wrong with LSTM?\nIs it because of 128 steps too much?<\/p>\n","answers":[]}
+{"QuestionId":57585895,"AnswerCount":0,"Tags":"<machine-learning><keras>","CreationDate":"2019-08-21T06:46:45.163","AcceptedAnswerId":null,"OwnerUserId":395028.0,"Title":"Using keras.layers.Concatenate correctly","Body":"<p>I have 2 different set of features, for which I have created 2 separate Sequential models, something like this:<\/p>\n\n<pre><code>model1 = Sequential()\nmodel1.add(Embedding(self.max_words, self.embedding_dim))\nmodel1.add(LSTM(32))\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>model2 = Sequential()\nmodel2.add(Dense(64, activation='relu', input_shape=(num_features,)))\nmodel2.add(Dense(32, activation='relu'))\n<\/code><\/pre>\n\n<p>Finally, I'm concatenating these 2 models like this:<\/p>\n\n<pre><code>merged_model = Sequential()\nmerged_model.add(Concatenate([model1, model2]))\nmerged_model.add(Dense(1, activation='sigmoid'))\nmerged_model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>However, I'm confused about what the 'fit' code would look like, as we need to have 2 separate x_train datasets in merged_model.\nIs this the correct way to do this:<\/p>\n\n<pre><code>history = merged_model.fit([x_train1, x_train2], y_train, epochs=10, batch_size=128, validation_split=0.2)\n<\/code><\/pre>\n\n<p>Also, is this the correct way to concatenate 2 layers if we want to merge 2 parallel models ?<\/p>\n","answers":[]}
+{"QuestionId":57585900,"AnswerCount":2,"Tags":"<python><tensorflow><keras><conv-neural-network>","CreationDate":"2019-08-21T06:47:08.063","AcceptedAnswerId":null,"OwnerUserId":11642300.0,"Title":"Object of Type 'Dense' has no len()","Body":"<p>I try to create a CNN model, but always get this error message.<\/p>\n\n<blockquote>\n  <p>Error: TypeError Traceback (most recent call last)  in () ----> 1 model = simple_conv_model() 5 frames \/usr\/local\/lib\/python3.6\/dist-packages\/keras\/engine\/network.py in build_map(tensor, finished_nodes, nodes_in_progress, layer, node_index, tensor_index) 1345 1346 # Propagate to all previous tensors connected to this node. -> 1347 for i in range(len(node.inbound_layers)): 1348 x = node.input_tensors[i] 1349 layer = node.inbound_layers[i] TypeError: object of type 'Dense' has no len()<\/p>\n<\/blockquote>\n\n<p>This is the model:<\/p>\n\n<pre><code>def simple_conv_model():\n        input_layer=layers.Input(shape=(64,64,3), name=\"input_layer\")    \n        model=layers.Conv2D(16,3, activation=\"relu\", padding='same', name=\"first_block_conv\", strides=(1,1)) (input_layer)\n        model=layers.MaxPooling2D((2,2), name=\"first_block_pooling\") (model)\n        model=layers.BatchNormalization(name=\"first_block_bn\") (model)\n\n        model=layers.Conv2D(32,3, activation=\"relu\", padding='same', name=\"second_block_conv\", strides=(1,1)) (input_layer)\n        model=layers.MaxPooling2D((2,2), name=\"second_block_pooling\") (model)\n        model=layers.BatchNormalization(name=\"second_block_bn\") (model)\n\n        model=layers.Conv2D(64,3, activation=\"relu\", padding='same', name=\"third_block_conv\", strides=(1,1)) (input_layer)\n        model=layers.MaxPooling2D((2,2), name=\"third_block_pooling\") (model)\n        model=layers.BatchNormalization(name=\"third_block_bn\") (model)\n\n        model=layers.Flatten() (model)\n        model=layers.Dense(16, activation=\"relu\", name=\"dense_1\") (model)\n        model=layers.BatchNormalization() (model)\n        model=layers.Dropout(0.5, name=\"drop_out_dense_1\") (model)\n\n        model=layers.Dense(4, activation=\"relu\", name=\"dense_2\") (model)\n\n        model=layers.Dense(1, activation=\"linear\") (model)\n\n        model_cnn = Model(input_layer, model)\n        model_cnn.compile(loss=\"mean_absolute_percentage_error\", optimizer=\"adam\")\n\n        return model_cnn\n\n    model = simple_conv_model()\n<\/code><\/pre>\n","answers":[{"AnswerId":"57586689","CreationDate":"2019-08-21T07:38:49.647","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>You are probably mistakenly using <code>input_layer<\/code> as the input of all the three <code>Conv2D<\/code> layers, whereas it is very likely that you intended to write something like this:<\/p>\n\n<pre><code>model = layers.Conv2D(16,3,...)(input_layer) # this is correct\n# ...\nmodel = layers.Conv2D(32, 3, ...)(model)  # pass `model` here as input\n# ...\nmodel = layers.Conv2D(64, 3, ...)(model)  # pass `model` here as input\n<\/code><\/pre>\n"},{"AnswerId":"58380505","CreationDate":"2019-10-14T16:13:51.723","ParentId":null,"OwnerUserId":"4388278","Title":null,"Body":"<p>I also face this problem before because I import library from tensorflow.python.keras. Just change to use keras instead<\/p>\n"}]}
+{"QuestionId":57586685,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2019-08-21T07:38:32.360","AcceptedAnswerId":null,"OwnerUserId":10763046.0,"Title":"Tensorflow: fully convolution net get noise output","Body":"<p>My question just like this:<a href=\"https:\/\/stackoverflow.com\/questions\/49554650\/noise-in-image-segmentation-result\">noise in image segmentation result<\/a><\/p>\n\n<p>but I don't know how to solve my problem<\/p>\n\n<p>my model was defined as:<\/p>\n\n<pre><code>conv1_1 = self.conv2d(img, 64, layer_name = \"conv1_1\")\nconv1_2 = self.conv2d(conv1_1, 64, layer_name = \"conv1_2\")\npool1   = self.max_pool(conv1_2, \"pool1\")\n\nconv2_1 = self.conv2d(pool1, 128, layer_name = \"conv2_1\")\nconv2_2 = self.conv2d(conv2_1, 128, layer_name = \"conv2_2\")\npool2   = self.max_pool(conv2_2, \"pool2\")\n\nconv3_1 = self.conv2d(pool2, 256, layer_name = \"conv3_1\")\nconv3_2 = self.conv2d(conv3_1, 256, layer_name = \"conv3_2\")\nconv3_3 = self.conv2d(conv3_2, 256, layer_name = \"conv3_3\")\npool3   = self.max_pool(conv3_3, \"pool3\")\n\nconv4_1 = self.conv2d(pool3, 512, layer_name = \"conv4_1\")\nconv4_2 = self.conv2d(conv4_1, 512, layer_name = \"conv4_2\")\nconv4_3 = self.conv2d(conv4_2, 512,  layer_name =  \"conv4_3\")            \npool4   = self.max_pool(conv4_3, \"pool4\")\nprint(\"pool4: \", pool4.shape)\nconv5_1 = self.conv2d(pool4, 512, layer_name = \"conv5_1\")\nconv5_2 = self.conv2d(conv5_1, 512,  layer_name = \"conv5_2\")\nconv5_3 = self.conv2d(conv5_2, 512,  layer_name = \"conv5_3\")            \npool5   = self.max_pool(conv5_3, \"pool5\")\n\nconv6 = self.conv2d(pool5, 4096, 7,  layer_name = \"conv6\")\nconv7 = self.conv2d(conv6, 4096, 1,  layer_name = \"conv7\")\nprint(\"conv7: \", conv7.shape)\n\npool3_1x1 = self.conv2d(pool3, 2, k_size = 1, layer_name = \"pool3_1x1\")\npool4_1x1 = self.conv2d(pool4, 2, k_size = 1, layer_name = \"pool4_1x1\")\nconv7_1x1 = self.conv2d(conv7, 2, k_size = 1, layer_name = \"conv7_1x1\")\nprint(\"conv7_1x1: \", conv7_1x1.shape)\n\ndeconv7 = self.upsample(conv7_1x1, 2, k = 4, s = 2, layer_name = \"deconv7\")\n        print(\"deconv7: \", deconv7.shape)\n\nfuse1 = tf.add(deconv7, pool4_1x1)\ndeconv_fuse1 = self.upsample(fuse1, 2, k = 4, s = 2, layer_name = \"deconv_fuse1\")\nprint(\"deconv_fuse1: \", deconv_fuse1.shape)\nfuse2 = tf.add(deconv_fuse1, pool3_1x1)\nout = self.upsample(fuse2, 2, k = 16, s = 8, layer_name = \"out\")\nprint(\"out: \", out.shape)\nscore_map = tf.nn.sigmoid(out)\n<\/code><\/pre>\n\n<p>And when training, I got this kind of noisy image output on training set\n<a href=\"https:\/\/i.stack.imgur.com\/PWgSH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/PWgSH.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>What I really want is:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/dF0k5.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/dF0k5.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>white is sigmoid value >= 0.5\nand black is sigmoid value which &lt; 0.5<\/p>\n\n<p>Can someone tell me how to fix my code?<\/p>\n\n<p>My input image shape is 128x128x3<\/p>\n","answers":[]}
+{"QuestionId":57587241,"AnswerCount":0,"Tags":"<python><tensorflow><keras><vgg-net>","CreationDate":"2019-08-21T08:14:57.927","AcceptedAnswerId":null,"OwnerUserId":11955489.0,"Title":"Why my DataGenerator iterates on more data than the size of dataset and give IndexError: list index out of range?","Body":"<p>I'm trying to implement a network with keras and tensorflow back-end, I'm using transfer learning model (VGG16), my dataset is a medical images dataset so instead of having only one image, I have a series of slices, so my dataset is organized in a folder and each serie is a np.array() with size (nb_slices,512,512,3). \nMy dataset is composed by 1130 train samples and 120 valid samples, so I don't think that datas is the problem.\nI tried to create a dataGenerator to load my image series in my model without a batch-size problem (I used this <a href=\"https:\/\/stackoverflow.com\/questions\/53788434\/training-a-keras-model-from-batches-of-npy-files-using-generator\">Training a Keras model from batches of .npy files using generator?<\/a> to make my generator class) , (and I reshaped my volumes with size (nb_slices, 224,224,3))\nthen I tried to use transfer learning, and custom a VGG16 network with 1 more convolution layer, MaxPooling, Flatten, Dense, Dropout and final Dense layer. <\/p>\n\n<p>When I start training, it seems there is no problem, but at a moment it returns IndexError: list index out of range, and I saw that DataGenerator iterates more than the size of dataset but I don't know why...<\/p>\n\n<p>Which part could cause it ?<\/p>\n\n<p>Here is my DataGenerator <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>INPUT_DIM = 224\nMAX_PIXEL_VAL = 255\nMEAN = 58.09\nSTDDEV = 49.73\n\nclass DataGenerator(keras.utils.Sequence):\n    def __init__(self, file_list, labels, data_loc):\n        self.listIDs = file_list\n        self.labels = labels\n        self.data_loc = data_loc\n        self.on_epoch_end()\n\n    def __len__(self):\n        return int(len(self.listIDs))\n\n    def __getitem__(self, index):\n        indexes = self.indexes[index:(index + 1)]\n        list_IDS_temp = [self.listIDs[k] for k in indexes]\n        X, y = self.__data_generation(list_IDS_temp)\n        return X, y\n\n    def on_epoch_end(self):\n        self.indexes = np.arange(len(self.listIDs))\n\n    def __data_generation(self, list_IDS_temp):\n        for ID in list_IDS_temp:\n            vol = np.load(self.data_loc + ID + '.npy')\n            nb_slices = vol.shape[0]\n            pad = int((vol.shape[2] - INPUT_DIM) \/ 2)\n            vol = vol[:, pad:-pad, pad:-pad]\n\n            # standardize\n            vol = (vol - np.min(vol)) \/ (np.max(vol) - np.min(vol)) * MAX_PIXEL_VAL\n\n            # normalize\n            vol = (vol - MEAN) \/ STDDEV\n\n            # convert to RGB\n            vol = np.stack((vol,) * 3, axis=3)\n\n            y = np.empty(nb_slices, dtype=int)\n            # y = self.labels[ID]\n            for i in range(nb_slices):\n                y[i] = self.labels[int(ID)]\n\n        return vol, keras.utils.to_categorical(y, num_classes=2)\n<\/code><\/pre>\n\n<p>and my model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>train_set = DataGenerator(df_train['exams'].tolist(), df_train['labels'].tolist(), all_file_loc_train)\n\nvalid_set = DataGenerator(df_val['exams'].tolist(), df_val['labels'].tolist(), all_file_loc_val)\n\nmodel = VGG16(weights='imagenet', include_top=False, input_shape=(224, 224, 3))  # , input_shape=(224, 224, 3)\nlayer_dict = dict([(layer.name, layer) for layer in model.layers])\nx = layer_dict['block2_pool'].output\nx = Conv2D(filters=64, kernel_size=(3, 3), activation='relu')(x)\nx = MaxPooling2D(pool_size=(2, 2))(x)\nx = Flatten()(x)\nx = Dense(256, activation='relu')(x)\nx = Dropout(0.4)(x)\nx = Dense(2, activation='softmax')(x)\n\ncustom_model = Model(inputs=model.input, outputs=x)\nfor layer in custom_model.layers[:7]:\n    layer.trainable = False\n\ncustom_model.compile(loss='categorical_crossentropy', optimizer=optimizers.SGD(lr=0.0001, momentum=0.9),\n                     metrics=[\"accuracy\"])\n\nresults = custom_model.fit_generator(generator=train_set, validation_data=valid_set, epochs=50, verbose=2)\n\n<\/code><\/pre>\n\n<p>I expected accuracy around 80-90% of accuracy but it seems that something goes wrong in my DataGenerator and I don't know what. Please I need help...<\/p>\n","answers":[]}
+{"QuestionId":57587422,"AnswerCount":2,"Tags":"<tensorflow><keras><lstm><recurrent-neural-network>","CreationDate":"2019-08-21T08:27:01.113","AcceptedAnswerId":57590392.0,"OwnerUserId":7154924.0,"Title":"Can Keras RepeatVector repetition be specified dynamically?","Body":"<p>Keras's RepeatVector layer allows me to repeat a given vector\/Tensor <code>n<\/code> times: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/RepeatVector\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/RepeatVector<\/a>.<\/p>\n\n<p>However, it appears <code>n<\/code> has to be a constant. Is there any way to specify this parameter dynamically?<\/p>\n\n<p>Use case: I'm generating variable length sequences, and RepeatVector will be followed by an LSTM layer.<\/p>\n","answers":[{"AnswerId":"57590392","CreationDate":"2019-08-21T11:15:52.940","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>It can be a symbolic tensor. Therefore, you can use the backend function <code>shape()<\/code> (or alternatively <code>tf.shape()<\/code>) to dynamically find the number of timesteps from input tensor of LSTM layer:<\/p>\n\n<pre><code>from keras import backend as K\n\nreps = RepeatVector(K.shape(lstm_input)[1])(lstm_out)\n<\/code><\/pre>\n"},{"AnswerId":"58152913","CreationDate":"2019-09-29T06:39:05.323","ParentId":null,"OwnerUserId":"3293408","Title":null,"Body":"<p>For me, that code did not result in a model that would compile successfully. It complained that the tensor provided was not the output of a layer. The code that worked for me was:<\/p>\n\n<pre><code>sentence_length = Lambda(\n    lambda t: tensorflow.shape(t)[1],\n    name=\"sentence_length\"\n)(word_digests)\n\n# sentence vector, repeated so available for each word\nsentence_repeated = Lambda(\n    lambda t: tensorflow.tile( tensorflow.expand_dims(t, 1), (1,sentence_length,1) ),\n    name=\"sentence_repeated\"\n)(sentence_digest)\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":57587524,"AnswerCount":0,"Tags":"<android><tensorflow><tensorflow-lite>","CreationDate":"2019-08-21T08:33:28.483","AcceptedAnswerId":null,"OwnerUserId":6997819.0,"Title":"Failed to load model from \"MODEL PATH\"","Body":"<p>I am using the sample from\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/android\" rel=\"nofollow noreferrer\">THIS LINK<\/a><\/p>\n\n<p>I have my own model.pb and label file<\/p>\n\n<p>I have ended up with the error as below<\/p>\n\n<blockquote>\n  <p>Caused by: java.io.IOException: Not a valid TensorFlow Graph\n  serialization: NodeDef mentions attr 'Truncate' not in Op y:DstT; attr=SrcT:type; attr=DstT:type>; NodeDef:\n  Postprocessor\/Cast_3 = CastDstT=DT_FLOAT, SrcT=DT_INT32,\n  Truncate=false. (Check whether your\n  GraphDef-interpreting binary is up to date with your\n  GraphDef-generating binary.).<\/p>\n<\/blockquote>\n\n<p>Inside of my gradle I have tried both ways<\/p>\n\n<pre><code>dependencies {\n    if (nativeBuildSystem == 'cmake' || nativeBuildSystem == 'none') {\n        implementation 'org.tensorflow:tensorflow-android:1.8.0'\n     \/\/   implementation files('libs\/libandroid_tensorflow_inference_java.jar')\n    }\n}\n<\/code><\/pre>\n\n<p>Using Gradle and using lib, but ended up with same error. <\/p>\n\n<p>Please assist.<\/p>\n\n<p>PS: My server tensorflow verion is 1.14.0<\/p>\n","answers":[]}
+{"QuestionId":57588164,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-21T09:08:17.787","AcceptedAnswerId":null,"OwnerUserId":10571813.0,"Title":"Can't choose framework for segmentation_models library in python","Body":"<p>I am trying to build a ResNet34 model using <strong>segmentation_models<\/strong>(sm) library in python. The <strong>sm<\/strong> library uses <strong>keras framework<\/strong> by default when importing, but I am working with <strong>tf.keras<\/strong> to build my datasets used for training and testing.<\/p>\n\n<p>The documentation says that in order to change the default framework I should either use an environmental variable <code>SM_FRAMEWORK=tf.keras<\/code> before importing (which I tried but it didn't work) or set it using the method <code>set_framework<\/code> (which doesn't show up in the suggestions\/it says it doesn't exist when I try to execute it).<\/p>\n\n<p>Is there any other way to overcome this problem?<\/p>\n","answers":[]}
+{"QuestionId":57588172,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-08-21T09:08:31.250","AcceptedAnswerId":null,"OwnerUserId":5742328.0,"Title":"The fp file for amp with TF_AUTO_MIXED_PRECISION_GRAPH_REWRITE_LOG_PATH open can't convert to summary for tensorboard?","Body":"<p>I set the following environment variable:<\/p>\n\n<pre><code>os.environ['TF_CPP_VMODULE']=\"auto_mixed_precision=2\"\nos.environ['TF_AUTO_MIXED_PRECISION_GRAPH_REWRITE_LOG_PATH']='\/home\/work\/log\/fp16verbose'\n<\/code><\/pre>\n\n<p>and got pb file like graphdef_AutoMixedPrecision_tf_graph_1566009745594.pb with amp fp16 node inside.\nI want to have a look at the graph in tensorboard. so I use the tool to convert it.<\/p>\n\n<p><code>python -m tensorflow.python.tools.import_pb_to_tensorboard --model_dir=\"graphdef_AutoMixedPrecision_tf_graph_1566009745594.pb\" --log_dir=\".\"<\/code><\/p>\n\n<p>but the following exception occurs:\n<strong>ValueError: Node 'model\/get_train_op\/gradients\/model\/Transformer\/encode\/encoder_stack\/layer_0\/self_attention\/layer_normalization\/sub_1_grad\/Shape_1' expects to be colocated with unknown node 'model\/Transformer\/encode\/encoder_stack\/layer_0\/self_attention\/layer_normalization\/sub_1'<\/strong><\/p>\n\n<p>How to solve and get the summary for the tensorboard?<\/p>\n","answers":[]}
+{"QuestionId":57588234,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-21T09:12:16.217","AcceptedAnswerId":57588386.0,"OwnerUserId":2467772.0,"Title":"Configuring GPU memory usage in Tensorflow","Body":"<p>I have LSTM model and like to configure the memory usage.\nWhat I did was<\/p>\n\n<pre><code>class ActivityRecognition:\n     #Utility functions for training:\n     def LSTM_RNN(self,_X, _weights, _biases):\n          # model architecture based on \"guillaume-chevalier\" and \"aymericdamien\" under the MIT license.\n          _X = tf.transpose(_X, [1, 0, 2])  # permute n_steps and batch_size\n          _X = tf.reshape(_X, [-1, self.n_input])\n          # Rectifies Linear Unit activation function used\n          _X = tf.nn.relu(tf.matmul(_X, _weights['hidden']) + _biases['hidden'])\n          # Split data because rnn cell needs a list of inputs for the RNN inner loop\n          _X = tf.split(_X, self.n_steps, 0)\n\n          # Define two stacked LSTM cells (two recurrent layers deep) with tensorflow\n          lstm_cell_1 = tf.contrib.rnn.BasicLSTMCell(self.n_hidden, forget_bias=1.0, state_is_tuple=True)\n          lstm_cell_2 = tf.contrib.rnn.BasicLSTMCell(self.n_hidden, forget_bias=1.0, state_is_tuple=True)\n          lstm_cell_3 = tf.contrib.rnn.BasicLSTMCell(self.n_hidden, forget_bias=1.0, state_is_tuple=True)\n          lstm_cell_4 = tf.contrib.rnn.BasicLSTMCell(self.n_hidden, forget_bias=1.0, state_is_tuple=True)\n          lstm_cells = tf.contrib.rnn.MultiRNNCell([lstm_cell_1, lstm_cell_2, lstm_cell_3, lstm_cell_4], state_is_tuple=True)\n          outputs, states = tf.contrib.rnn.static_rnn(lstm_cells, _X, dtype=tf.float32)\n          lstm_last_output = outputs[-1]\n          return tf.matmul(lstm_last_output, _weights['out']) + _biases['out']\n     def __init__(self):\n          self.n_steps = 128 # 32 timesteps per series\n          self.n_input = 10  #ch4.x,ch4.y,ch7.x,ch7.y,dist4_16,dist7_17\n          self.n_hidden = 34 # Hidden layer num of features\n          self.n_classes = 3\n          self.global_step = tf.Variable(0, trainable=False)\n          # Graph input\/output\n          self.x = tf.placeholder(tf.float32, [None, self.n_steps, self.n_input])\n          self.y = tf.placeholder(tf.float32, [None, self.n_classes])\n          # Graph weights\n          self.weights = {\n               'hidden': tf.Variable(tf.random_normal([self.n_input, self.n_hidden])), # Hidden layer weights\n               'out': tf.Variable(tf.random_normal([self.n_hidden, self.n_classes], mean=1.0))\n          }\n          self.biases = {\n              'hidden': tf.Variable(tf.random_normal([self.n_hidden])),\n              'out': tf.Variable(tf.random_normal([self.n_classes]))\n          }\n          self.pred = self.LSTM_RNN(self.x, self.weights, self.biases)\n          config = tf.ConfigProto()\n          config.gpu_options.per_process_gpu_memory_fraction = 0.4\n          self.sess = tf.Session(config,...)\n          self.init = tf.global_variables_initializer()\n          with tf.Session() as self.sess:\n                  self.sess.run(self.init)\n                  saver = tf.train.Saver()\n                  saver.restore(self.sess, tf.train.latest_checkpoint('\/data\/saac\/HumanActivity\/TrainModels\/'))\n                  print(\"Model restored.\")\n<\/code><\/pre>\n\n<p>I have error in running the code as<\/p>\n\n<pre><code>File \"\/data\/saac\/HumanActivity\/ActivityRecognition.py\", line 55, in __init__\n    self.sess = tf.Session(config,...)\n  File \"\/home\/user\/venvcuda9_0\/lib\/python3.5\/site-packages\/tensorflow\/python\/client\/session.py\", line 1551, in __init__\n    super(Session, self).__init__(target, graph, config=config)\n  File \"\/home\/user\/venvcuda9_0\/lib\/python3.5\/site-packages\/tensorflow\/python\/client\/session.py\", line 643, in __init__\n    raise TypeError('graph must be a tf.Graph, but got %s' % type(graph))\nTypeError: graph must be a tf.Graph, but got &lt;class 'ellipsis'&gt;\nException ignored in: &lt;bound method BaseSession.__del__ of &lt;tensorflow.python.client.session.Session object at 0x7f6e7e402320&gt;&gt;\nTraceback (most recent call last):\n  File \"\/home\/user\/venvcuda9_0\/lib\/python3.5\/site-packages\/tensorflow\/python\/client\/session.py\", line 736, in __del__\n    if self._session is not None:\nAttributeError: 'Session' object has no attribute '_session'\n<\/code><\/pre>\n\n<p>What could be wrong?<\/p>\n","answers":[{"AnswerId":"57588386","CreationDate":"2019-08-21T09:20:56.870","ParentId":null,"OwnerUserId":"3399825","Title":null,"Body":"<p>The problem is in the constructor params of <code>tf.Session<\/code>. The ellipsis are recognized as a parameter for the Graph, that's what this error says <code>TypeError: graph must be a tf.Graph, but got &lt;class 'ellipsis'&gt;<\/code>. <\/p>\n\n<p>Replace<\/p>\n\n<pre><code>self.sess = tf.Session(config,...)\n<\/code><\/pre>\n\n<p>with<\/p>\n\n<pre><code>self.sess = tf.Session(config=config)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57588335,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-21T09:18:08.720","AcceptedAnswerId":null,"OwnerUserId":4367722.0,"Title":"How to get output from each layer for an image input in tensorflow python model","Body":"<p>I want to dump the output of each layer from TensorFlow python model<\/p>\n\n<p>The result I got from the code below seems to be incorrect. I have manually calculated the convolution output using the python script and a CPP model. Both of them are consistent which I feel the issue with TensorFlow output<\/p>\n\n<pre><code>for idx in range(len(test_model.layers)):\n                if(idx!=-1):\n                    print(test_model.layers[list(test_model.layers.keys())[i]])\n\n                    pred = sess.run(test_model.layers[list(test_model.layers.keys())[idx]],\n                            feed_dict={test_model.image: batch_data['image']})\n                    print(pred.shape)\n                    fo = open(\"..\/..\/layer_out\/layer\" + str(i) + \"_out.txt\",'w')\n                    print(\" Layer \" + str(idx) +  \" Output \" + str(pred.shape) +' ',list(test_model.layers.keys())[idx])\n                    fo.write(\"Layer \" + str(idx) +  \" Output \" + str(pred.shape)+'\\n' )\n\n                    #print([i for i in np.ravel(pred)])\n                    l = np.ravel(pred)\n                    if(i==85):\n                        print('Check\/\/\/\/**************',max(l),np.where(l==max(l)))\n                        print(label_dict[np.where(l==max(l))[0][0]])\n\n                    for  outval in (l):\n                        fo.write(str(outval)+'\\n')\n                        print(str(outval))\n\n                    #if(i==2):exit()\n                    fo.close()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57588498,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><machine-learning><protocol-buffers><object-detection-api>","CreationDate":"2019-08-21T09:26:29.757","AcceptedAnswerId":null,"OwnerUserId":9333287.0,"Title":"How to create a new optimizer in tensorflow objectapi?","Body":"<p>I am using tensorflow object api, and I want to use my own optimiser, CLR (cyclical learning rate policy). <\/p>\n\n<p>I know how to create a callback in regular tensorflow, but the object api takes in protobuf as config file. I do not really get how they work, so I wanted to know how can I bring my own optimiser to the objectapi? <\/p>\n\n<h1>Protobuff example :<\/h1>\n\n<pre><code>optimizer {\n    momentum_optimizer: {\n      learning_rate: {\n        manual_step_learning_rate {\n          initial_learning_rate: 0.0001\n          schedule {\n            step: 400\n            learning_rate: .00001\n          }\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57588589,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2019-08-21T09:31:52.617","AcceptedAnswerId":null,"OwnerUserId":236106.0,"Title":"Error while traning using the estimator API in tensorflow","Body":"<p>I am trying to run a simple <code>SVM<\/code> classifier over iris dataset by providing the data using input_fn by returning a <code>tf.data.dataset<\/code> object but I run into the following error.<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"tf_test.py\", line 45, in &lt;module&gt;\n    est.fit(steps=1, input_fn=input_fn)\n  File \"\/venv\/lib\/python3.7\/site-packages\/tensorflow\/python\/util\/deprecation.py\", line 507, in new_func\n    return func(*args, **kwargs)\n  File \"\/venv\/lib\/python3.7\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/estimator.py\", line 524, in fit\n    loss = self._train_model(input_fn=input_fn, hooks=hooks)\n  File \"\/venv\/lib\/python3.7\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/estimator.py\", line 1038, in _train_model\n    features, labels = input_fn()\nValueError: too many values to unpack (expected 2)\n<\/code><\/pre>\n\n<p>I think the above error is a bug in tensorflow so I've tried returning an iterator instead using the line <code>tf.compat.v1.data.make_one_shot_iterator(dataset).get_next()<\/code> which runs into a different error as follows<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"tf_test.py\", line 48, in &lt;module&gt;\n    est.fit(steps=1, input_fn=input_fn)\n  File \"venv\/lib\/python3.7\/site-packages\/tensorflow\/python\/util\/deprecation.py\", line 507, in new_func\n    return func(*args, **kwargs)\n  File \"\/venv\/lib\/python3.7\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/estimator.py\", line 524, in fit\n    loss = self._train_model(input_fn=input_fn, hooks=hooks)\n  File \"\/venv\/lib\/python3.7\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/estimator.py\", line 1041, in _train_model\n    model_fn_ops = self._get_train_ops(features, labels)\n  File \"\/venv\/lib\/python3.7\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/estimator.py\", line 1264, in _get_train_ops\n    return self._call_model_fn(features, labels, model_fn_lib.ModeKeys.TRAIN)\n  File \"\/venv\/lib\/python3.7\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/estimator.py\", line 1227, in _call_model_fn\n    model_fn_results = self._model_fn(features, labels, **kwargs)\n  File \"\/venv\/lib\/python3.7\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/linear.py\", line 251, in sdca_model_fn\n    features.update(layers.transform_features(features, feature_columns))\n  File \"\/venv\/lib\/python3.7\/site-packages\/tensorflow\/contrib\/layers\/python\/layers\/feature_column_ops.py\", line 653, in transform_features\n    check_feature_columns(feature_columns)\n  File \"\/venv\/lib\/python3.7\/site-packages\/tensorflow\/contrib\/layers\/python\/layers\/feature_column_ops.py\", line 777, in check_feature_columns\n    key = f.key\nAttributeError: 'str' object has no attribute 'key'\n<\/code><\/pre>\n\n<p>I've prepared this self contained example to demonstrate the problem.<\/p>\n\n<pre><code>import copy\nimport tempfile\nimport tensorflow as tf\ntf.enable_eager_execution()\n\n_, filename = tempfile.mkstemp()\n\ncols = [\"example_id\", \"sepal_len\", \"sepal_width\", \"petal_len\", \"petal_width\", \"label\"]\n\ndata = \"\\n\".join([\n    #\",\".join(cols),\n    \"1,5.1,3.5,1.4,0.2,Iris-setosa\",\n    \"2,4.9,3.0,1.4,0.2,Iris-setosa\",\n    \"2,4.7,3.2,1.3,0.2,Iris-setosa\",\n    \"4,4.6,3.1,1.5,0.2,Iris-setosa\",\n    \"5,5.0,3.6,1.4,0.2,Iris-setosa\",\n    \"6,5.7,2.5,5.0,2.0,Iris-virginica\",\n    \"7,7.0,3.2,4.7,1.4,Iris-versicolor\",\n    \"8,6.5,3.2,5.1,2.0,Iris-virginica\",\n    \"9,6.4,2.7,5.3,1.9,Iris-virginica\",\n    \"10,6.8,3.0,5.5,2.1,Iris-virginica\",\n    \"11,5.4,3.9,1.7,0.4,Iris-setosa\"\n])\n\nwith open(filename, 'w') as f:\n    f.write(data)\n\nbatch_size = 2\n\nFIELD_DEFAULTS = [[0], [0.0], [0.0], [0.0], [0.0], [0]]\n\ndef _parse(line):\n    fields = tf.io.decode_csv(line, FIELD_DEFAULTS)\n    features = dict(zip(cols, fields))\n    label = features.pop('label')\n    return features, label\n\ndef input_fn():\n    dataset = tf.data.TextLineDataset(filename).skip(1)\n    dataset = dataset.map(_parse)\n    return dataset\n    # return tf.compat.v1.data.make_one_shot_iterator(dataset).get_next()\n\n\nfeature_cols = copy.copy(cols)\nfeature_cols.remove('example_id')\nest = tf.contrib.learn.SVM(example_id_column='example_ids', feature_columns=feature_cols)\nest.fit(steps=1, input_fn=input_fn)\nout = est.predict(input_fn=test_input_fn, yield_single_examples=False)\nassert(len(out['classes']) == len(data))\n\n<\/code><\/pre>\n\n<p>How can I make this work with <code>tensorflow==1.14<\/code><\/p>\n","answers":[]}
+{"QuestionId":57588994,"AnswerCount":0,"Tags":"<keras><deep-learning><conv-neural-network>","CreationDate":"2019-08-21T09:55:04.180","AcceptedAnswerId":null,"OwnerUserId":7888603.0,"Title":"how to get probability of each class using functional api of keras?","Body":"<p>I always get one hot encoded output, e.i [0, 1, 0], whereas I'd like to get output such as [0.1, 0.95, 0.1]<\/p>\n\n<pre><code>from keras.models import load_model\nfrom keras.preprocessing.image import load_img,img_to_array\nfrom PIL import Image\nimport numpy as np\nimport os\nimage = load_img(os.path.join('happy_dog.jpg')) \nimage = image.convert('L')\nimage = image.resize((28,28),Image.ANTIALIAS)\n\nimgarray = img_to_array(image)\nnp.expand_dims(image, axis = 0)\n\npredict = model.predict([[imgarray]], verbose=0)\nprint(predict[0])\n<\/code><\/pre>\n\n<p>array([1., 0., 0., 0., 0., 0., 0., 0., 0., 0.], dtype=float32)<\/p>\n","answers":[]}
+{"QuestionId":57589408,"AnswerCount":0,"Tags":"<python><android><tensorflow><tensorflow-lite><style-transfer>","CreationDate":"2019-08-21T10:18:12.560","AcceptedAnswerId":null,"OwnerUserId":7423949.0,"Title":"No OpKernel was registered to support Op 'Round' with these attrs. Registered devices: [CPU], Registered kernels:<no registered kernels>","Body":"<p>I am getting error while trying to apply tensorflow model(.pb) in android, The model is loaded sucessfully but after that getting this error.I have trained tensorflow model on CPU.Any help will be appreciated.<\/p>\n\n<blockquote>\n  <p>No OpKernel was registered to support Op 'Round' with these attrs. Registered devices: [CPU], Registered kernels:<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/fSBhZ.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/fSBhZ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Thank you.<\/p>\n","answers":[]}
+{"QuestionId":57589412,"AnswerCount":0,"Tags":"<python><opencv><tensorflow><scipy><computer-vision>","CreationDate":"2019-08-21T10:18:24.437","AcceptedAnswerId":null,"OwnerUserId":11956212.0,"Title":"Is there any way to improve accuracy for pointing difference between two images?","Body":"<p>I am trying to build verification system which can point out the difference between two scanned leaflets. <\/p>\n\n<p>I blurred both images, found homography to align both images and then found difference between two images to draw the rectangles.<\/p>\n\n<p>Standard leaflet: <a href=\"https:\/\/ibb.co\/GWjKW1x\" rel=\"nofollow noreferrer\">https:\/\/ibb.co\/GWjKW1x<\/a><\/p>\n\n<p>Sample for checking: <a href=\"https:\/\/ibb.co\/pv3Crxm\" rel=\"nofollow noreferrer\">https:\/\/ibb.co\/pv3Crxm<\/a><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import cv2\nimport numpy as np\nfrom skimage.measure import compare_ssim as ssim\nfrom scipy import ndimage\n\n#MAX_MATCHES = 1000000  #MSE for labels\nMAX_MATCHES = 50000 #MSE for text\nGOOD_MATCH_PERCENT = 0.4\n\ndef mse(imageA, imageB):\n    # the 'Mean Squared Error' between the two images is the\n    # sum of the squared difference between the two images;\n    # NOTE: the two images must have the same dimension\n    err = np.sum((imageA.astype(\"float\") - imageB.astype(\"float\")) ** 2)\n    err \/= float(imageA.shape[0] * imageA.shape[1])\n\n    return err\n\n    # return the MSE, the lower the error, the more \"similar\"\n    # the two images are\n\ndef diff_remove_bg(img0, img, img1):\n    d1 = diff(img0, img)\n    d2 = diff(img, img1)\n    return cv2.bitwise_and(d1, d2)\n\ndef alignImages(im1, im2):\n\n  # Convert images to grayscale\n  im1Gray = cv2.cvtColor(im1, cv2.COLOR_BGR2GRAY)\n  im2Gray = cv2.cvtColor(im2, cv2.COLOR_BGR2GRAY)\n\n\n  # Detect ORB features and compute descriptors.\n  orb = cv2.ORB_create(MAX_MATCHES)\n\n  keypoints1, descriptors1 = orb.detectAndCompute(im1Gray, None)\n  keypoints2, descriptors2 = orb.detectAndCompute(im2Gray, None)\n\n  # Match features.\n  matcher = cv2.DescriptorMatcher_create(cv2.DESCRIPTOR_MATCHER_BRUTEFORCE_HAMMING)  \n  matches = matcher.match(descriptors1, descriptors2, None)\n\n  # Sort matches by score\n  matches.sort(key=lambda x: x.distance, reverse=False)\n\n  # Remove not so good matches\n  numGoodMatches = int(len(matches) * GOOD_MATCH_PERCENT)\n  matches = matches[:numGoodMatches]\n\n  # Draw top matches\n  imMatches = cv2.drawMatches(im1, keypoints1, im2, keypoints2, matches, None)\n  cv2.imwrite(\"matches.png\", imMatches)\n\n  # Extract location of good matches\n  points1 = np.zeros((len(matches), 2), dtype=np.float32)\n  points2 = np.zeros((len(matches), 2), dtype=np.float32)\n\n  for i, match in enumerate(matches):\n    points1[i, :] = keypoints1[match.queryIdx].pt\n    points2[i, :] = keypoints2[match.trainIdx].pt\n\n  # Find homography\n  h, mask = cv2.findHomography(points1, points2, cv2.RANSAC)\n\n  # Use homography\n  height, width, channels = im2.shape\n  im1Reg = cv2.warpPerspective(im1, h, (width, height))\n\n  return im1Reg, h\n\n\n\n# Read reference image\nrefFilename = \"standard.png\"\nprint(\"Reading reference image : \", refFilename)\n\nimReference = cv2.imread(refFilename, cv2.IMREAD_COLOR)  \nimReference = ndimage.gaussian_filter(imReference, 2)\n\n\nimFilename = \"sample-fin.png\"\nprint(\"Reading image to align : \", imFilename);  \n\nim = cv2.imread(imFilename, cv2.IMREAD_COLOR)\nim = ndimage.gaussian_filter(im, 2)\n\nprint(\"Aligning images ...\")\n\n# Registered image will be resotred in imReg. \n# The estimated homography will be stored in h. \nimReg, h = alignImages(im, imReference)\n\n# Write aligned image to disk. \noutFilename = \"aligned.png\"\nprint(\"Saving aligned image : \", outFilename); \ncv2.imwrite(outFilename, imReg)\n\n# Print estimated homography\nprint(\"Estimated homography : \\n\",  h)\n\nx1 = imReg\nx2 = imReference\n\nabsdiff = cv2.absdiff(x1, x2)\n\nprint(np.max(absdiff))\n\n#Threshold for minimum detection\nthreshold = absdiff &lt; 30\nabsdiff[threshold] = 0      \n\ncv2.imwrite(\"absdiff.png\", absdiff)\n\ndiff = cv2.subtract(x1, x2)\nresult = not np.any(diff)\n\n\n########### Drawing Rectangles around the differences ##############\n\nimg = cv2.pyrDown(cv2.imread('absdiff.png', cv2.IMREAD_UNCHANGED))\nimg1 =  cv2.pyrDown(cv2.imread('aligned.png', cv2.IMREAD_UNCHANGED))\n\n\nret, threshed_img = cv2.threshold(cv2.cvtColor(img, cv2.COLOR_BGR2GRAY),\n                40, 40, cv2.THRESH_BINARY)\n\n_,contours, hier = cv2.findContours(threshed_img, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)\n\n\nfor c in contours:\n    x,y,w,h = cv2.boundingRect(c)\n\n    cv2.rectangle(img1, (x-5, y-5), (x+w+10, y+h+10), (0, 255, 0), 2)\n\n    # get the min area rect\n    rect = cv2.minAreaRect(c)\n    box = cv2.boxPoints(rect)\n\n    # convert all coordinates floating point values to int\n    box = np.int0(box)\n\nprint(len(contours))\n\ncv2.imwrite(\"cont-1.png\", img1)\n<\/code><\/pre>\n\n<p>Difference image: <a href=\"https:\/\/ibb.co\/NCj32qs\" rel=\"nofollow noreferrer\">https:\/\/ibb.co\/NCj32qs<\/a><\/p>\n\n<p>Pointed out errors: <a href=\"https:\/\/ibb.co\/YWKxYLF\" rel=\"nofollow noreferrer\">https:\/\/ibb.co\/YWKxYLF<\/a><\/p>\n\n<p>I want to improve the accuracy of this algorithm. Or is there some other way which can be used to solve such task?<\/p>\n\n<p>Edit:\nI want to remove extra rectangles that are present at the end of the leaflet. Also the homography is not perfect. I have to put a huge threshold on the difference to get this result. <\/p>\n","answers":[]}
+{"QuestionId":57589625,"AnswerCount":0,"Tags":"<tensorflow><keras><deep-learning><word-embedding><seq2seq>","CreationDate":"2019-08-21T10:31:12.010","AcceptedAnswerId":null,"OwnerUserId":6533445.0,"Title":"BERT for Text Summarization","Body":"<p>I'm trying to build a text summarization model using seq2seq architecture in Keras. I've followed this tutorial <a href=\"https:\/\/keras.io\/examples\/lstm_seq2seq\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/examples\/lstm_seq2seq\/<\/a> and implemented it with Embeddings layer, which works fine. But now I want to use BERT. Can pretrained BERT embeddings be used in such a task, usually I see text classifiation, but not the encoder-decoder architecture used with BERT.<\/p>\n\n<p>I access BERT model from TF Hub, and have a Layer class implemented from this tutorial <a href=\"https:\/\/github.com\/strongio\/keras-bert\/blob\/master\/keras-bert.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/strongio\/keras-bert\/blob\/master\/keras-bert.ipynb<\/a>, I also tokenize accordingly with BERT tokenizer, below is my model<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>enc_in_id = Input(shape=(None, ), name=\"Encoder-Input-Ids\")\nenc_in_mask = Input(shape=(None, ), name=\"Encoder-Input-Masks\")\nenc_in_segment = Input(shape=(None, ), name=\"Encoder-Input-Segment-Ids\")\nbert_encoder_inputs = [enc_in_id, enc_in_mask, enc_in_segment]\n\nencoder_embeddings = BertLayer(name='Encoder-Bert-Layer')(bert_encoder_inputs)\nencoder_embeddings = BatchNormalization(name='Encoder-Batch-Normalization')(encoder_embeddings)\nencoder_lstm = LSTM(latent_size, return_state=True, name='Encoder-LSTM')\nencoder_out, e_state_h, e_state_c = encoder_lstm(encoder_embeddings)\nencoder_states = [e_state_h, e_state_c]\n\ndec_in_id = Input(shape=(None,), name=\"Decoder-Input-Ids\")\ndec_in_mask = Input(shape=(None,), name=\"Decoder-Input-Masks\")\ndec_in_segment = Input(shape=(None,), name=\"Decoder-Input-Segment-Ids\")\nbert_decoder_inputs = [dec_in_id, dec_in_mask, dec_in_segment]\n\ndecoder_embeddings_layer = BertLayer(name='Decoder-Bert-Layer')\ndecoder_embeddings = decoder_embeddings_layer(bert_decoder_inputs)\ndecoder_batchnorm_layer = BatchNormalization(name='Decoder-Batch-Normalization-1')\ndecoder_batchnorm = decoder_batchnorm_layer(decoder_embeddings)\n\ndecoder_lstm = LSTM(latent_size, return_state=True, return_sequences=True, name='Decoder-LSTM')\ndecoder_out, _, _ = decoder_lstm(decoder_batchnorm, initial_state=encoder_states)\ndense_batchnorm_layer = BatchNormalization(name='Decoder-Batch-Normalization-2')\ndecoder_out_batchnorm = dense_batchnorm_layer(decoder_out)\ndecoder_dense_id = Dense(vocabulary_size, activation='softmax', name='Dense-Id')\ndec_outputs_id = decoder_dense_id(decoder_out_batchnorm)\n<\/code><\/pre>\n\n<p>The model builds and after a couple of epochs accuracy rises to 1, and loss drops below 0.5, but the predictions are awful. Since I'm working on a dev set comprised of 5 samples, with max 30 WordPiece tokens and predicting on the same data, I only get the first or maybe two tokens right, then it just repeats the last seen token, or [PAD] token.<\/p>\n","answers":[]}
+{"QuestionId":57589683,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><anaconda><recurrent-neural-network>","CreationDate":"2019-08-21T10:34:09.283","AcceptedAnswerId":null,"OwnerUserId":7757150.0,"Title":"TensorFlow : NotFoundError: Resource AnonymousIterator Error","Body":"<p>I am trying to replicate the example of using RNN in time-series as posted on \n<a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/text\/time_series#multi-step_model\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/tutorials\/text\/time_series#multi-step_model<\/a><\/p>\n\n<p>However, I am getting the following error message :<\/p>\n\n<pre><code> NotFoundError: Resource AnonymousIterator\/AnonymousIterator3\/class \n tensorflow::data::IteratorResource does not exist.\n [[{{node IteratorGetNext_3}}]]\n<\/code><\/pre>\n\n<p>When I trying to execute the line : <\/p>\n\n<pre><code>for x, y in val_univariate.take(1):\nprint(simple_lstm_model.predict(x,steps=1).shape)\n<\/code><\/pre>\n\n<p>Environment :<\/p>\n\n<pre><code>Tenserflow : 2.0.0-beta1\nPython : Python 3.6.7 Anaconda\nOS : Windows 10 64-bit\n<\/code><\/pre>\n\n<p>Any solutions for this?<\/p>\n","answers":[{"AnswerId":"58108567","CreationDate":"2019-09-26T01:58:42.600","ParentId":null,"OwnerUserId":"3578102","Title":null,"Body":"<p>From <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset#__iter__\" rel=\"nofollow noreferrer\">TF doc<\/a>,<\/p>\n\n<p>It states:<\/p>\n\n<blockquote>\n  <p>Creates an Iterator for enumerating the elements of this dataset.<\/p>\n  \n  <p>The returned iterator implements the Python iterator protocol and therefore can only be used in eager mode.<\/p>\n<\/blockquote>\n\n<p>So you need to use <code>tf.enable_eager_execution()<\/code> at the beginning of your script.<\/p>\n"}]}
+{"QuestionId":57589760,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-21T10:38:08.370","AcceptedAnswerId":57591201.0,"OwnerUserId":1109811.0,"Title":"Map labels to counts after tf.unique_with_counts()","Body":"<p>I have a BxN input feature tensor <code>feature_labels<\/code> and I call<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>unique_labels, label_idcs, label_counts = tf.unique_with_counts(feature_labels)\n<\/code><\/pre>\n\n<p>to get the label counts.<\/p>\n\n<p>How do I produce a BxN matrix, which stores the label counts instead of the labels?<\/p>\n\n<p><strong>Example for one batch:<\/strong><\/p>\n\n<pre><code>Input: \nfeature_labels: 0 1 1 1 0 2 3 1 3\n\n&gt; unique_labels, _, label_counts = tf.unique_with_counts(feature_labels)\n\n=&gt; unique_labels: 0 1 2 3\n=&gt; label_counts: 2 4 1 2\n\nOutput:\n2 4 4 4 2 1 2 4 2\n<\/code><\/pre>\n","answers":[{"AnswerId":"57591201","CreationDate":"2019-08-21T12:04:19.350","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>If I understand correctly, I think you can do that like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\nwith tf.Graph().as_default(), tf.Session() as sess:\n    feature_labels = tf.constant([0, 1, 1, 1, 0, 2, 3, 1, 3], tf.int32)\n    _, unique_idx, label_counts = tf.unique_with_counts(feature_labels)\n    result = tf.gather(label_counts, unique_idx)\n    print(sess.run(result))\n    # [2 4 4 4 2 1 2 4 2]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57590119,"AnswerCount":0,"Tags":"<opencv><tensorflow><google-cloud-vision>","CreationDate":"2019-08-21T10:59:42.143","AcceptedAnswerId":null,"OwnerUserId":10948271.0,"Title":"Can CloudVision count up original character logos from a picture?","Body":"<p>Below image is the game result.\nI need to count up some logo in a picture.<\/p>\n\n<p>What is the best way to count up character logo in the picture?\nI'm just web engineer, so I don't want to use tensorflow.\nIs it possible to use cloudVision Api to count up character in the picture?<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/dHhsf.jpg\" alt=\"game result\"><\/p>\n\n<p>Above image has 63 character logos in a picture.\nI just want to count up character.<\/p>\n\n<p>What is the best way?\nDo I need tensorflow or opencv?<\/p>\n\n<p>thanks<\/p>\n","answers":[]}
+{"QuestionId":57590583,"AnswerCount":0,"Tags":"<python><tensorflow><keras><model><batch-normalization>","CreationDate":"2019-08-21T11:26:13.487","AcceptedAnswerId":null,"OwnerUserId":10220850.0,"Title":"How to handle update_ops of batch normalization in Keras functional model API?","Body":"<p>I am trying out the functional API for Keras models and trying to set up two dataset streams using the same model and weight sharing, which also consists of batch normalization. When I create the model on top of my first datastream and then call it on the second datastream, I can see in the tensorboard the additional update ops, but the model does not include these newly created ops. My question would be, what a good coding approach would look like for this scenario, while still being able to use the classical tensorflow sessions, dataset iterators and custom losses and optimizers?<\/p>\n\n<p>I've attached some code showing how my requested behavior actually works for the BatchNormalization layer itself. Still, it would be also nice in this case to have more control over the update ops, e.g. instead of iterating through the list and check the names, directly having somehow the update ops connected or returned with the call of the layer itself, so that one can associate directly the update ops to the correct datastream.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport shutil\nimport os.path as osp\n\n\nshutil.rmtree(osp.join('\/tmp', 'keras_model_testtb'), ignore_errors=True)\ntb_saver = tf.summary.FileWriter(osp.join(\n    '\/tmp', 'keras_model_testtb',\n))\n\ninput1 = tf.keras.layers.Input(shape=(None, 3), dtype=tf.float32)\nbatchnorm = tf.keras.layers.BatchNormalization()\noutput1 = batchnorm(input1, training=tf.constant(True))\n# following print shows:\n# [\n#   &lt;tf.Operation 'batch_normalization_v1\/AssignMovingAvg\/AssignSubVariableOp' type=AssignSubVariableOp&gt;,\n#   &lt;tf.Operation 'batch_normalization_v1\/AssignMovingAvg_1\/AssignSubVariableOp' type=AssignSubVariableOp&gt;\n# ]\nprint(batchnorm.updates)\ninput2 = tf.keras.layers.Input(shape=(None, 3), dtype=tf.float32)\noutput2 = batchnorm(input2, training=tf.constant(True))\n# following print shows:\n# [\n#   &lt;tf.Operation 'batch_normalization_v1\/AssignMovingAvg\/AssignSubVariableOp' type=AssignSubVariableOp&gt;,\n#   &lt;tf.Operation 'batch_normalization_v1\/AssignMovingAvg_1\/AssignSubVariableOp' type=AssignSubVariableOp&gt;,\n#   &lt;tf.Operation 'batch_normalization_v1_1\/AssignMovingAvg\/AssignSubVariableOp' type=AssignSubVariableOp&gt;,\n#   &lt;tf.Operation 'batch_normalization_v1_1\/AssignMovingAvg_1\/AssignSubVariableOp' type=AssignSubVariableOp&gt;\n# ]\n# update ops of both layer calls are merged into one list and one has to check the name of the ops\n# to use them correctly with optimizers\nprint(batchnorm.updates)\n\nwith tf.Session() as session:\n    session.run(tf.global_variables_initializer())\n    tb_saver.add_graph(session.graph)\n<\/code><\/pre>\n\n<p>Here is some sample code using the Keras Model API. As you can see, the model updates op do have the additional 2 update ops as in the previous example only for the case when using the keras Input layer. For dataset iterators it seems to do something different.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#!\/usr\/bin\/env python3\n\nimport tensorflow as tf\nimport shutil\nimport os.path as osp\n\n\ndef reshape(sample):\n    return tf.cast(tf.tile(sample[None, None, None], [2, 4, 3]), tf.float32)\n\n\nds = tf.data.Dataset.range(10)\nds = ds.map(reshape)\ninput_ds = ds.make_one_shot_iterator().get_next()\n\n\nshutil.rmtree(osp.join('\/tmp', 'keras_model_testtb'), ignore_errors=True)\ntb_saver = tf.summary.FileWriter(osp.join(\n    '\/tmp', 'keras_model_testtb',\n))\n\ninput1 = tf.keras.layers.Input(shape=(None, 3), dtype=tf.float32)\noutput1 = tf.keras.layers.BatchNormalization()(input1, training=tf.constant(True))\nmodel = tf.keras.models.Model(inputs=input1, outputs=output1)\n# following prints show:\n# [\n#   &lt;tf.Operation 'batch_normalization_v1\/AssignMovingAvg\/AssignSubVariableOp' type=AssignSubVariableOp&gt;,\n#   &lt;tf.Operation 'batch_normalization_v1\/AssignMovingAvg_1\/AssignSubVariableOp' type=AssignSubVariableOp&gt;\n# ]\n# []\nprint(model.updates)\nprint([n.name for n in tf.get_default_graph().as_graph_def().node if 'model' in n.name and 'AssignSubVariableOp' in n.name])\noutput_ds = model(input_ds)\n# following prints show:\n# [\n#   &lt;tf.Operation 'batch_normalization_v1\/AssignMovingAvg\/AssignSubVariableOp' type=AssignSubVariableOp&gt;,\n#   &lt;tf.Operation 'batch_normalization_v1\/AssignMovingAvg_1\/AssignSubVariableOp' type=AssignSubVariableOp&gt;\n# ]\n# ['model\/batch_normalization_v1\/AssignMovingAvg\/AssignSubVariableOp', 'model\/batch_normalization_v1\/AssignMovingAvg_1\/AssignSubVariableOp']\nprint(model.updates)\nprint([n.name for n in tf.get_default_graph().as_graph_def().node if 'model' in n.name and 'AssignSubVariableOp' in n.name])\n# Above I expected to have found the new update ops also in model.updates, as is the case when you call the model on keras Input tensors:\ninput2 = tf.keras.layers.Input(shape=(None, 3), dtype=tf.float32)\noutput2 = model(input2)\n# following prints show:\n# [\n#   &lt;tf.Operation 'batch_normalization_v1\/AssignMovingAvg\/AssignSubVariableOp' type=AssignSubVariableOp&gt;,\n#   &lt;tf.Operation 'batch_normalization_v1\/AssignMovingAvg_1\/AssignSubVariableOp' type=AssignSubVariableOp&gt;,\n#   &lt;tf.Operation 'model_1\/batch_normalization_v1\/AssignMovingAvg_1\/AssignSubVariableOp' type=AssignSubVariableOp&gt;,\n#   &lt;tf.Operation 'model_1\/batch_normalization_v1\/AssignMovingAvg\/AssignSubVariableOp' type=AssignSubVariableOp&gt;\n# ]\n# [\n#   'model\/batch_normalization_v1\/AssignMovingAvg\/AssignSubVariableOp', 'model\/batch_normalization_v1\/AssignMovingAvg_1\/AssignSubVariableOp',\n#   'model_1\/batch_normalization_v1\/AssignMovingAvg\/AssignSubVariableOp', 'model_1\/batch_normalization_v1\/AssignMovingAvg_1\/AssignSubVariableOp'\n# ]\nprint(model.updates)\nprint([n.name for n in tf.get_default_graph().as_graph_def().node if 'model' in n.name and 'AssignSubVariableOp' in n.name])\n\nwith tf.Session() as session:\n    session.run(tf.global_variables_initializer())\n    tb_saver.add_graph(session.graph)\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57590625,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras><scikit-learn><gridsearchcv>","CreationDate":"2019-08-21T11:29:02.653","AcceptedAnswerId":null,"OwnerUserId":6527505.0,"Title":"Custom Metric Scoring for Multi Label KerasClassifier in GridSearchCV","Body":"<p>I am trying to do gridsearch using scikit-learn GridSearchCV and KerasClassifier for a multilabel classification problem having 346 labels. I am trying to evaluate the models based on a custom metric. But I am always running into the following error. My training data size is 5334.<\/p>\n\n<blockquote>\n  <p>ValueError: operands could not be broadcast together with shapes (5334,346) (5334,) <\/p>\n<\/blockquote>\n\n<pre><code>def my_custom_scorer(y_true,y_pred):\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\n  custom_scorer = make_scorer(my_custom_scorer, greater_is_better=True)\n\n   def create_model(learn_rate = 0.01): \n    # create model\n    model = Sequential()\n    model.add(Dense(100, activation = 'sigmoid',input_dim=4))\n\n    model.add(Dense(346, activation='sigmoid'))\n    # Compile model\n\n    optimizer = SGD(lr=learn_rate, momentum=0)\n\n    model.compile(optimizer=optimizer,\n          loss=custom_loss(weights),\n          metrics=['accuracy'])\n\n   return model\n\nmodel = KerasClassifier(build_fn=create_model,verbose=0)\n\nlearn_rate = [0.001, 0.01, 0.1]\n\nbatch_size = [10]\nepochs = [1] \n\nparam_grid = dict(epochs=epochs,batch_size=batch_size,learn_rate=learn_rate)\ngrid = GridSearchCV(estimator=model, \nparam_grid=param_grid,n_jobs=1,scoring=custom_scorer)\n\ngrid_result = grid.fit(X_train, Y_train)\n\nprint(\"Best: %f using %s\" % (grid_result.best_score_, \ngrid_result.best_params_))\nmeans = grid_result.cv_results_['mean_test_score']\nstds = grid_result.cv_results_['std_test_score']\nparams = grid_result.cv_results_['params']\nfor mean, stdev, param in zip(means, stds, params):\n   print(\"%f (%f) with: %r\" % (mean, stdev, param))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57590697,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-08-21T11:34:15.400","AcceptedAnswerId":57591782.0,"OwnerUserId":2123706.0,"Title":"pytorch \"log_softmax_lastdim_kernel_impl\" not implemented for 'torch.LongTensor'","Body":"<p>I am trying to use my own dataset to classify text according to <a href=\"https:\/\/github.com\/bentrevett\/pytorch-sentiment-analysis\/blob\/master\/5%20-%20Multi-class%20Sentiment%20Analysis.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/bentrevett\/pytorch-sentiment-analysis\/blob\/master\/5%20-%20Multi-class%20Sentiment%20Analysis.ipynb<\/a>. My dataset is a csv of sentences and a class associated with it. there are 6 different classes:<\/p>\n\n<pre><code>sent                      class\n'the fox is brown'        animal\n'the house is big'        object\n'one water is drinkable'  water\n...\n<\/code><\/pre>\n\n<p>When running:<\/p>\n\n<pre><code>N_EPOCHS = 5\n\nbest_valid_loss = float('inf')\n\nfor epoch in range(N_EPOCHS):\n\n    start_time = time.time()\n    print(start_time)\n    train_loss, train_acc = train(model, train_iterator, optimizer, criterion)\n    print(train_loss.type())\n    print(train_acc.type())\n    valid_loss, valid_acc = evaluate(model, valid_iterator, criterion)\n\n    end_time = time.time()\n\n    epoch_mins, epoch_secs = epoch_time(start_time, end_time)\n\n    if valid_loss &lt; best_valid_loss:\n        best_valid_loss = valid_loss\n        torch.save(model.state_dict(), 'tut5-model.pt')\n\n    print(f'Epoch: {epoch+1:02} | Epoch Time: {epoch_mins}m {epoch_secs}s')\n    print(f'\\tTrain Loss: {train_loss:.3f} | Train Acc: {train_acc*100:.2f}%')\n    print(f'\\t Val. Loss: {valid_loss:.3f} |  Val. Acc: {valid_acc*100:.2f}%')\n<\/code><\/pre>\n\n<p>, I receive the following error<\/p>\n\n<pre><code>RuntimeError: \"log_softmax_lastdim_kernel_impl\" not implemented for 'torch.LongTensor'\n<\/code><\/pre>\n\n<p>pointing to:<\/p>\n\n<pre><code>&lt;ipython-input-38-9c6cff70d2aa&gt; in train(model, iterator, optimizer, criterion)\n     14         print('pred'+ predictions.type())\n     15         #batch.label = batch.label.type(torch.LongTensor)\n---&gt; 16         loss = criterion(predictions.long(), batch.label)**\n<\/code><\/pre>\n\n<p>The solution posted here <a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/14224\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/pytorch\/issues\/14224<\/a> suggests I need to use long\/int. <\/p>\n\n<p>I had to add <code>.long()<\/code> at line <code>**<\/code> in order to fix this earlier error:<\/p>\n\n<p><code>RuntimeError: Expected object of scalar type Long but got scalar type Float for argument #2 'target'<\/code><\/p>\n\n<p>The specific lines of code are:<\/p>\n\n<pre><code>  def train(model, iterator, optimizer, criterion):\n    epoch_loss = 0\n    epoch_acc = 0\n\n    model.train()\n\n    for batch in iterator:\n\n        optimizer.zero_grad()\n\n        predictions = model(batch.text)\n        print('pred'+ predictions.type())\n        #batch.label = batch.label.type(torch.LongTensor)\n        loss = criterion(predictions.long(), batch.label)**\n\n        acc = categorical_accuracy(predictions, batch.label)\n\n        loss.backward()\n\n        optimizer.step()\n\n        epoch_loss += loss.item()\n        epoch_acc += acc.item()\n\n    return epoch_loss \/ len(iterator), epoch_acc \/ len(iterator)\n<\/code><\/pre>\n\n<p>Note, the ** was originally <code>loss = criterion(predictions, batch.label)<\/code><\/p>\n\n<p>Any other suggestions to fix this issue?<\/p>\n","answers":[{"AnswerId":"57591782","CreationDate":"2019-08-21T12:36:16.603","ParentId":null,"OwnerUserId":"5027315","Title":null,"Body":"<p><code>criterion<\/code> is defined as <code>torch.nn.CrossEntropyLoss()<\/code> in your <a href=\"https:\/\/github.com\/bentrevett\/pytorch-sentiment-analysis\/blob\/master\/5%20-%20Multi-class%20Sentiment%20Analysis.ipynb\" rel=\"nofollow noreferrer\">notebook<\/a>. As mentioned in documentation of <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.CrossEntropyLoss\" rel=\"nofollow noreferrer\"><code>CrossEntropyLoss<\/code><\/a>, it expects probability values returned by model for each of the 'K' classes and corresponding value for ground-truth label as input. Now, probability values are float tensors, while ground-truth label should be a long tensor representing a class (class can not be a float, e.g. 2.3 can not represent a class). hence:<\/p>\n\n<pre><code>loss = criterion(predictions, batch.label.long())\n<\/code><\/pre>\n\n<p>should work.<\/p>\n"}]}
+{"QuestionId":57591154,"AnswerCount":2,"Tags":"<python><c++><tensorflow><half-precision-float>","CreationDate":"2019-08-21T12:01:17.560","AcceptedAnswerId":null,"OwnerUserId":11955260.0,"Title":"how to do convolution with fp16(Eigen::half) on tensorflow","Body":"<p>How can I use tensorflow to do convolution using fp16 on GPU? (the python api using __half or Eigen::half).<\/p>\n\n<p>I want to test a model with fp16 on tensorflow, but I got stucked. Actually, I found that fp16 convolution in tensorflow seems like casting the fp32 convolution's result into fp16, which is not what I need. <\/p>\n\n<p>I tried to give the tf.nn.conv2d a fp16 input in fp16 format, and give the tf.nn.conv2d a fp16 input in fp32 format (tf.cast it into fp32) then tf.cast the result into fp16, and they gave exactly the same result.\nBut as I think, doing convolution in fp16 is different from doing it in fp32 and then cast it into fp16, am I wrong?\nPlease help me, thanks.<\/p>\n\n<pre><code>environment:\nubuntu 16.04\ntensorflow 1.9.0\ncuda 9.0\nTesla V100\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\nimport os\n\ndef conv16_32(input, kernel): # fake fp16 convolution\n    input = tf.cast(input, tf.float16)\n    kernel = tf.cast(kernel, tf.float16)\n    input = tf.cast(input, tf.float32)\n    kernel = tf.cast(kernel, tf.float32)\n    out = tf.nn.conv2d(input, kernel, [1,1,1,1], padding='VALID')\n    out = tf.cast(out, tf.float16)\n    out = tf.cast(out, tf.float64)\n    return out\n\ndef conv16(input, kernel): # real fp16 convolution\n    input = tf.cast(input, tf.float16)\n    kernel = tf.cast(kernel, tf.float16)\n    out = tf.nn.conv2d(input, kernel, [1,1,1,1], padding='VALID')\n    out = tf.cast(out, tf.float64)\n    return out\n\nx = np.random.rand(16, 32, 32, 16).astype('float64')\nw = np.random.rand(3, 3, 16, 16).astype('float64')\nx = tf.get_variable('input', dtype=tf.float64, initializer=x)\nw = tf.get_variable('weight', dtype=tf.float64, initializer=w)\n\nout_16 = conv16(x, w)\nout_16_32 = conv16_32(x, w)\n\nos.environ['CUDA_VISIBLE_DEVICES'] = '1'\nconfig = tf.ConfigProto()\nconfig.gpu_options.allow_growth = True\nsess = tf.Session(config = config)\nsess.run(tf.global_variables_initializer())\nsess.run(tf.local_variables_initializer())\nprint(sess.run(tf.reduce_max(out_16_32 - out_16)))\n\n<\/code><\/pre>\n\n<p>The above two functions give the same result, say the final 'print' result is zero.<\/p>\n\n<p>The result of fp16 convolution and fp32 convolution should not be same (in my point of view).\nHow can I use tensorflow to do convolution using real fp16 on GPU? (the python api using __half or Eigen::half)<\/p>\n","answers":[{"AnswerId":"57607008","CreationDate":"2019-08-22T10:14:12.887","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>I think you are using the operations correctly. In your example, you can check that the convolution operations do indeed have the right type.<\/p>\n\n<pre><code>conv2d_op_16 = out_16.op.inputs[0].op\nprint(conv2d_op_16.name, conv2d_op_16.type, conv2d_op_16.get_attr('T'))\n# Conv2D Conv2D &lt;dtype: 'float16'&gt;\nconv2d_op_16_32 = out_16_32.op.inputs[0].op.inputs[0].op\nprint(conv2d_op_16_32.name, conv2d_op_16_32.type, conv2d_op_16_32.get_attr('T'))\n# Conv2D_1 Conv2D &lt;dtype: 'float32'&gt;\n<\/code><\/pre>\n\n<p>And TensorFlow does register kernels for fp16 <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.14.0\/tensorflow\/core\/kernels\/conv_ops.cc#L555\" rel=\"nofollow noreferrer\">for CPU<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.14.0\/tensorflow\/core\/kernels\/conv_ops.cc#L1051-L1053\" rel=\"nofollow noreferrer\">for GPU<\/a>, so there is no reason to think is doing anything else. I don't have a lot of experience with fp16, so I'm not sure if the zero difference is \"normal\", but there does not seem to be any way in which <code>conv16<\/code> is using anything other than a fp16 convolution.<\/p>\n"},{"AnswerId":"58609705","CreationDate":"2019-10-29T14:41:03.520","ParentId":null,"OwnerUserId":"3850022","Title":null,"Body":"<p>I'm trying to figure out the same. Here is some simple code that you can test convolutions with:<\/p>\n\n\n\n<pre class=\"lang-python prettyprint-override\"><code>import tensorflow as tf\ntf.enable_eager_execution()\ninput = tf.cast([[[[65519], [65519], [65519], [65519]]]], tf.float16) #BHWC\nfilter = tf.cast([[[[65519]], [[-65519]]]], tf.float16) #HWIO\ntf.print(tf.nn.conv2d(input, filter, [1,1,1,1], \"VALID\"))\n<\/code><\/pre>\n\n<p>This should overflow if the convolutions are done in f16, but doesn't actually overflow in Tensorflow. The result I get is [[[[0][0][0]]]], which suggest that convolutions are performed in f32.<\/p>\n"}]}
+{"QuestionId":57591365,"AnswerCount":0,"Tags":"<python><tensorflow><keras><neural-network><deep-learning>","CreationDate":"2019-08-21T12:12:34.783","AcceptedAnswerId":null,"OwnerUserId":11149645.0,"Title":"Counting CNN layers","Body":"<p>i created a CNN for a project i am involved with and i need to present it. The issue is, I am not sure about how to count the layers. <\/p>\n\n<p>Here is my model: <\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv2D(64,(3,3), input_shape = (40,40,2)))\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'))\n\n\nmodel.add(Dense(1600))\nmodel.add(Reshape((40,40)))\nmodel.add(Activation('sigmoid'))\n\nmodel.compile(loss='MSE',\n             optimizer='SGD',\n             metrics=['MAE'])\n<\/code><\/pre>\n\n<p>len(model.layers) returned 12 : <\/p>\n\n<p>So i used 1 input 10 hidden 1 output layers,<\/p>\n\n<p>or <\/p>\n\n<p>i need to count them as a group and say 1 input 2 hidden 1 output?<\/p>\n","answers":[]}
+{"QuestionId":57592108,"AnswerCount":0,"Tags":"<python><machine-learning><keras><time-series>","CreationDate":"2019-08-21T12:53:49.217","AcceptedAnswerId":null,"OwnerUserId":7151380.0,"Title":"LSTM Variational Autoencoder with Masking for variable length sequences","Body":"<p>I am trying to implement a LSTM VAE (following <a href=\"https:\/\/github.com\/twairball\/keras_lstm_vae\/blob\/master\/lstm_vae\/vae.py\" rel=\"nofollow noreferrer\">this<\/a> example I found), but also have it accept variable length sequences using Masking Layers. I tried to combine the above code with the ideas from <a href=\"https:\/\/stackoverflow.com\/questions\/46653322\/how-to-use-masking-layer-to-mask-input-output-in-lstm-autoencoders\">this<\/a> SO question that seems to deal with it the \"best way\" by cropping the gradients to get the most accurate loss as possible, however my implementation does not seem to be able to reproduce sequences on a small set of data. I am thus relatively confident that there is something amiss with my implementation, but I cannot seem to pinpoint what exactly is wrong. The relevant part is here:<\/p>\n\n<pre><code>x = Input(shape=(None, input_dim))(x)\nx_masked = Masking(mask_value=0.0, input_shape=(None, input_dim))(x)\n\nh = LSTM(intermediate_dim)(x_masked)\n\nz_mean = Dense(latent_dim)(h)\nz_log_sigma = Dense(latent_dim)(h)\n\ndef sampling(args):\n    z_mean, z_log_sigma = args\n    epsilon = K.random_normal(shape=(batch_size, latent_dim), mean=0., stddev=epsilon_std)\n    return z_mean + z_log_sigma * epsilon\n\nz = Lambda(sampling, output_shape=(latent_dim,))([z_mean, z_log_sigma])\n\ndecoded_h = LSTM(intermediate_dim, return_sequences=True)\ndecoded_mean = LSTM(latent_dim, return_sequences=True)\n\nh_decoded = RepeatVector(max_timesteps)(z)\nh_decoded = decoder_h(h_decoded)\n\nx_decoded_mean = decoder_mean(h_decoded)\n\ndef crop_outputs(x):\n    padding = K.cast(K.not_equal(x[1], 0), dtype=K.floatx())\n    return x[0] * padding\n\nx_decoded_mean = Lambda(crop_outputs, output_shape=(max_timesteps, input_dim))([x_decoded_mean, x])\n\nvae = Model(x, x_decoded_mean)\n\ndef vae_loss(x, x_decoded_mean):\n    xent_loss = objectives.mse(x, x_decoded_mean)\n    kl_loss = -0.5 * K.mean(1 + z_log_sigma - K.square(z_mean) - K.exp(z_log_sigma))\n    loss = xent_loss + kl_loss\n    return loss\n\nvae.compile(optimizer='adam', loss=vae_loss)\n\n# Here, X is variable length time series data of shape\n# (num_examples, max_timesteps, input_dim) and is zero padded\n# on the right for all the examples of length less than max_timesteps\n# X has been appropriately scaled using the StandardScaler.\n\nvae.fit(X, X, epochs = num_epochs, batch_size=batch_size)\n<\/code><\/pre>\n\n<p>As always, any help is much appreciated. Thank you!<\/p>\n","answers":[]}
+{"QuestionId":57592155,"AnswerCount":1,"Tags":"<keras><lstm><prediction><lstm-stateful>","CreationDate":"2019-08-21T12:56:04.367","AcceptedAnswerId":null,"OwnerUserId":11956955.0,"Title":"Basic time series prediction with lstm","Body":"<p>I have a sequence and I would like to do the simplest LSTM possible to predict the rest of the sequence.\nMeaning I want to start by using only the previous step to predict the next one and then add more steps.\nI want to use the predicted values as inputs also.\nSo I believe what I want is to achieve many to many as mentioned in the answers there <a href=\"https:\/\/stackoverflow.com\/questions\/38714959\/understanding-keras-lstms\/50235563\">Understanding Keras LSTMs<\/a> .<\/p>\n\n<p>I have read other questions on the topic on stackoverflow but still didn't manage to make it work. In my code, I'm using the tutorial here <a href=\"https:\/\/machinelearningmastery.com\/time-series-prediction-lstm-recurrent-neural-networks-python-keras\/\" rel=\"nofollow noreferrer\">https:\/\/machinelearningmastery.com\/time-series-prediction-lstm-recurrent-neural-networks-python-keras\/<\/a> and the function create_dataset to create two arrays with only a shift of one step.<\/p>\n\n<h1>Here is my code and the error I got.<\/h1>\n\n<pre><code>\"Here I'm scaling my data as advised\"\nscaler = MinMaxScaler(feature_range=(0, 1))\nRot = scaler.fit_transform(Rot)\n\n\"I'm creating the model using batch_size=1 but I'm not sure why this is necessary\"\n\nbatch_size = 1\nmodel = Sequential()\nmodel.add(LSTM(1,batch_input_shape=(batch_size,1,1),stateful=True,return_sequences=True,input_shape=(None,1)))\nmodel.add(Dense(1))\nmodel.compile(loss='mean_squared_error', optimizer='adam')\n\n\"I want to use only the previous value for now\"\nlook_back = 1\n\"as len(Rot) = 41000 I'm taking 36000 for training\"\ntrain_size = 36000\nX,Y = create_dataset(Rot[:train_size,:],look_back)\nX = numpy.reshape(X,(X.shape[0], X.shape[1], 1))\nY = numpy.reshape(Y,(X.shape[0], X.shape[1], 1))\n\nAnd now I train my network as advised by @Daniel M\u00f6ller. \n\nepochs = 10\n\nfor epoch in range(epochs):\n    model.reset_states()\n    model.train_on_batch(X,Y)\n\n\" And I get this error \"\n\" PartialTensorShape: Incompatible shapes during merge: [35998,1] vs. [1,1]\n     [[{{node lstm_11\/TensorArrayStack\/TensorArrayGatherV3}}]].\"\n<\/code><\/pre>\n\n<p>Do you know why I have such an error as it seems I did everything as in the topic mentioned above ?<\/p>\n","answers":[{"AnswerId":"57745452","CreationDate":"2019-09-01T11:06:26.113","ParentId":null,"OwnerUserId":"6842531","Title":null,"Body":"<p>In this LSTM network batch_size=1, because it is stateful. When stateful=True, the train_set size and test_set size when divided by batch_size should have a modulo of zero.<\/p>\n\n<p>batch_input_shape=(batch_size,1,1) is already defined, then why again,input_shape=(None,1)<\/p>\n\n<p>When return_sequences=True, another LSTM is following the existing LSTM layer. But here it is not.<\/p>\n"}]}
+{"QuestionId":57592430,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><text-classification>","CreationDate":"2019-08-21T13:08:45.050","AcceptedAnswerId":null,"OwnerUserId":1420331.0,"Title":"Keras Functional Api, InvalidArgumentError: input 1 should contain 3 elements, but got 2","Body":"<p>I am using the functional API of Keras for a text classification task. I am combining word embeddings with metadata, which I both convert into one hot encodings.\nI am getting this error when I fit the model:\n<code>tensorflow.python.framework.errors_impl.InvalidArgumentError: input 1 should contain 3 elements, but got 2[[{{node training\/Adam\/gradients\/concatenate_1\/concat_grad\/ConcatOffset}}]]<\/code>\nI am running a docker container in a K8s cluster, tensorflow is <code>1.8<\/code> and keras <code>2.2.4<\/code>.<\/p>\n\n<p>**EDIT:**I have already checked that the shape of the nlp data and the metadata is <code>(55744, 500)<\/code>. The layers are like: Metadata layer shape: <code>(?, 500, 1)<\/code> and NLP layer shape: <code>(?, 500)<\/code>.<\/p>\n\n<p>This is the code:<\/p>\n\n<pre><code>nlp_layer = Input(shape=(int(float(maximum_sequence_length)),), name='nlp_layer');\nembedding_layer = Embedding(input_dim=(len(word_idx_training) + 1), output_dim=int(em_dim), weights=[embeddings_matrix], input_length=int(float(maximum_sequence_length)), trainable=False, name='embedding_layer')(nlp_layer);\nmetadata_layer = Input(shape=(int(float(maximum_sequence_length)), 1), name='metadata_layer');\ndense_1 = Dense(128)(embedding_layer);\nconcatate = Concatenate()([dense_1, metadata_layer]);\ndropout = Dropout(0.6)(concatate);\ndense_2 = Dense(64)(dropout);\ndense_3 = Dense(32)(dense_2);\ndense_4 = Dense(16)(dense_3);\nflatten = Flatten()(dense_4);\ndense_5 = Dense(int(class_number))(flatten);\nactivation = Activation(\"softmax\")(dense_5);\nmodel = Model(inputs=[nlp_layer, metadata_layer], outputs=[activation]);\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['acc']);\nverbs_array = tf.convert_to_tensor(verbs_oh, np.float32);\nmodel.fit([x_train, verbs_array],y_train,batch_size=None,epochs=int(float(args.epochs)),suffle=True,steps_per_epoch=10);\n<\/code><\/pre>\n\n<p>**EDIT:**This is the error stack:<\/p>\n\n<pre><code>WARNING:tensorflow:From \/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/op_def_library.py:263: colocate_with (from tensorflow.pytho\nn.framework.ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nColocations handled automatically by placer.\n2019-08-20 21:04:52.171203: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not co\nmpiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA\n2019-08-20 21:04:52.204324: I tensorflow\/core\/platform\/profile_utils\/cpu_utils.cc:94] CPU Frequency: 2399940000 Hz\n2019-08-20 21:04:52.206737: I tensorflow\/compiler\/xla\/service\/service.cc:150] XLA service 0x55f751cbe750 executing computations on platform Host. Devi\nces:\n2019-08-20 21:04:52.206794: I tensorflow\/compiler\/xla\/service\/service.cc:158]   StreamExecutor device (0): &lt;undefined&gt;, &lt;undefined&gt;\nWARNING:tensorflow:From \/usr\/local\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py:3445: calling dropout (from tensorflow.python.ops.n\nn_ops) with keep_prob is deprecated and will be removed in a future version.\nInstructions for updating:\nPlease use `rate` instead of `keep_prob`. Rate should be set to `rate = 1 - keep_prob`.\nWARNING:tensorflow:From \/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/math_ops.py:3066: to_int32 (from tensorflow.python.ops.math_ops) \nis deprecated and will be removed in a future version.\nInstructions for updating:\nUse tf.cast instead.\nEpoch 1\/10\nTraceback (most recent call last):\n  File \"\/usr\/local\/lib\/python3.6\/runpy.py\", line 193, in _run_module_as_main\n    \"__main__\", mod_spec)\n  File \"\/usr\/local\/lib\/python3.6\/runpy.py\", line 85, in _run_code\n    exec(code, run_globals)\n  File \"\/code\/semeval\/g_quote\/classify_data_with_metadata_keras.py\", line 170, in &lt;module&gt;\n    steps_per_epoch=10);\n  File \"\/usr\/local\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\", line 1039, in fit\n    validation_steps=validation_steps)\n  File \"\/usr\/local\/lib\/python3.6\/site-packages\/keras\/engine\/training_arrays.py\", line 154, in fit_loop\n    outs = f(ins)\n  File \"\/usr\/local\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 2715, in __call__\n    return self._call(inputs)\n  File \"\/usr\/local\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 2675, in _call\n    fetched = self._callable_fn(*array_vals)\n  File \"\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1439, in __call__\n    run_metadata_ptr)\n  File \"\/usr\/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.InvalidArgumentError: input 1 should contain 3 elements, but got 2\n         [[{{node training\/Adam\/gradients\/concatenate_1\/concat_grad\/ConcatOffset}}]]\nroot@first-experiment-app-memory:\/code# \n<\/code><\/pre>\n\n<p>I have already increased the GPU memory, because I was getting <code>Allocation of 5574400000 exceeds 10% of system memory<\/code> and I use <code>steps_per_epoch<\/code>, because the functional API was throwing errors when I was using <code>batch_size<\/code> instead. \nCan someone tell what is the problem with the input\/output sizes that causes the <code>InvalidArgumentError<\/code>? Is it in the way the model is build or in the input data itself?\nThere is also <code>[[{{node training\/Adam\/gradients\/concatenate_1\/concat_grad\/ConcatOffset}}]]<\/code> in the error message.<\/p>\n\n<p>This is how the model looks like:<\/p>\n\n<pre><code>__________________________________________________________________________________________________\nembedding_layer (Embedding)     (None, 500, 50)      1174300     nlp_layer[0][0]                  \n__________________________________________________________________________________________________\ndense_1 (Dense)                 (None, 500, 128)     6528        embedding_layer[0][0]            \n__________________________________________________________________________________________________\nmetadata_layer (InputLayer)     (None, 500, 1)       0                                            \n__________________________________________________________________________________________________\nconcatenate_1 (Concatenate)     (None, 500, 129)     0           dense_1[0][0]                    \n                                                                 metadata_layer[0][0]             \n__________________________________________________________________________________________________\ndropout_1 (Dropout)             (None, 500, 129)     0           concatenate_1[0][0]              \n__________________________________________________________________________________________________\ndense_2 (Dense)                 (None, 500, 64)      8320        dropout_1[0][0]                  \n__________________________________________________________________________________________________\ndense_3 (Dense)                 (None, 500, 32)      2080        dense_2[0][0]                    \n__________________________________________________________________________________________________\ndense_4 (Dense)                 (None, 500, 16)      528         dense_3[0][0]                    \n__________________________________________________________________________________________________\nflatten_1 (Flatten)             (None, 8000)         0           dense_4[0][0]                    \n__________________________________________________________________________________________________\ndense_5 (Dense)                 (None, 4)            32004       flatten_1[0][0]                  \n__________________________________________________________________________________________________\nactivation_1 (Activation)       (None, 4)            0           dense_5[0][0]        ```\n<\/code><\/pre>\n","answers":[{"AnswerId":"57626290","CreationDate":"2019-08-23T12:23:47.457","ParentId":null,"OwnerUserId":"1420331","Title":null,"Body":"<p>So, after fiddling around with the code, I realized that I didn't have to transform the input metadata itself to a tensor. I was confused by the fact that when we define a Functional Model in Keras we add tensors in it. \nThe solution is first to compute the one hot encoding (500 dimensions) of my additional feature (this would result to a numpy array with rows as many as the input docs, and 500 columns) and use this array as the second input of the model.Then, I also need to expand this array with one mode dimension, because the metadata layer expects to see 3D data.<\/p>\n\n<p>In the above code, I need to replace<\/p>\n\n<pre><code>verbs_array = tf.convert_to_tensor(verbs_oh, np.float32);\n<\/code><\/pre>\n\n<p>with <\/p>\n\n<pre><code>verbs_array = verbs_oh;  # one code encoding of the feature \nverbs_array = np.expand_dims(verbs_array, axis=2);\n\n<\/code><\/pre>\n\n<p>This will change the verbs array from <code>(55744, 500)<\/code> to <code>(55744, 500, 1)<\/code>\nEven though the original error is solved, there is another mistake with the code that I cannot quite understand.\nAfter changing <code>verbs_array<\/code> as shown above, if I concatenate like this:<\/p>\n\n<pre><code>Concatenate(axis=1)([dense_1, metadata_layer])\n<\/code><\/pre>\n\n<p>I get this error: <code>ValueError: A<\/code>Concatenate<code>layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [(None, 500, 128), (None, 500, 1)]<\/code><\/p>\n\n<p>These are shapes I currently have:<\/p>\n\n<pre><code>Metadata layer shape: (?, 500, 1)\nNLP layer shape: (?, 500)\nEmb layer shape: (?, 500, 50)\nDense 1 shape: (?, 500, 128)\n<\/code><\/pre>\n\n<p>If I concatenate on axis=2, then I experience no error. However, I am unsure whether this is semantically correct. This is how the first layers look like in this case:<\/p>\n\n<pre><code>nlp_layer (InputLayer)          (None, 500)          0                                            \n__________________________________________________________________________________________________\nembedding_layer (Embedding)     (None, 500, 50)      1174300     nlp_layer[0][0]                  \n__________________________________________________________________________________________________\ndense_1 (Dense)                 (None, 500, 128)     6528        embedding_layer[0][0]            \n__________________________________________________________________________________________________\nmetadata_layer (InputLayer)     (None, 500, 1)       0                                            \n__________________________________________________________________________________________________\nconcatenate_1 (Concatenate)     (None, 500, 129)     0           dense_1[0][0]                    \n                                                                 metadata_layer[0][0] \n<\/code><\/pre>\n\n<p>Even though I discovered that the initial error is resolved by removing <code>tf.convert_to_tensor<\/code>, I will accept an answer that can correct the whole example and explain what is the appropriate input\/output shapes for my network.\n<strong>Also<\/strong>: After concatenating on axis=2, the training of the model became extremely slow. Which makes me think that using two one-hot encodings combined is too complex for my data\/network. I switched to using an index for every different value in the feature (so I only have one additional dimension now) and this means that this line<\/p>\n\n<pre><code>metadata_layer = Input(shape=(int(float(maximum_sequence_length)), 1), name='metadata_layer');\n<\/code><\/pre>\n\n<p>is replaced by <\/p>\n\n<pre><code>metadata_layer = Input(shape=(1, ), name='metadata_layer');\n<\/code><\/pre>\n\n<p>and then when feeding the network, I need to do:<\/p>\n\n<pre><code>verbs_array = verbs_idx;\nverbs_array = np.expand_dims(verbs_array, axis=1);\nverbs_array = np.expand_dims(verbs_array, axis=1);\n<\/code><\/pre>\n\n<p>And the shape of <code>verbs_array<\/code> is (55744, 1, 1), while the shape of the metadata layer is  (?, 1).\nThese changes result to a similar error:\n<code>ValueError: A<\/code>Concatenate<code>layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [(None, 500, 128), (None, 1)]<\/code><\/p>\n\n<p>I need to add more metadata next to the nlp data, so I would appreciate greatly someone that gives a complete answer to my issues. Thanks in advance.<\/p>\n"}]}
+{"QuestionId":57592759,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-08-21T13:26:23.120","AcceptedAnswerId":null,"OwnerUserId":9493813.0,"Title":"Reshape Final Layer","Body":"<p>I have an object detection problem, and I'm trying to use an Unet architecture.My input is 512x640x3. My final layer Reshape has the size (None,10240,1792). I don't understand from where 1792 comes .<\/p>\n\n<pre><code>def _create_model(self):\n\n    inputs = Input((self.config.IMAGE_HEIGHT, self.config.IMAGE_WIDTH, self.config.N_CHANNELS))\n    #inputs_norm = Lambda(lambda x: x\/127.5 - 1.)\n    conv1 = Convolution2D(8,3, activation=lrelu, padding='same',kernel_initializer='glorot_uniform')(inputs)\n    conv1 = Convolution2D(8, 3, activation=lrelu, padding='same',kernel_initializer='glorot_uniform')(conv1)\n    pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)\n\n    conv2 = Convolution2D(16, 3, activation=lrelu, padding='same',kernel_initializer='glorot_uniform')(pool1)\n    conv2 = Convolution2D(16, 3, activation=lrelu, padding='same',kernel_initializer='glorot_uniform')(conv2)\n    pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)\n\n    conv3 = Convolution2D(32, 3, activation=lrelu, padding='same',kernel_initializer='glorot_uniform')(pool2)\n    conv3 = Convolution2D(32, 3, activation=lrelu, padding='same',kernel_initializer='glorot_uniform')(conv3)\n    pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)\n\n    conv4 = Convolution2D(64, 3, activation=lrelu, padding='same',kernel_initializer='glorot_uniform')(pool3)\n    conv4 = Convolution2D(64, 3, activation=lrelu, padding='same',kernel_initializer='glorot_uniform')(conv4)\n    pool4 = MaxPooling2D(pool_size=(2, 2))(conv4)\n\n    conv5 = Convolution2D(128, 3, activation=lrelu, padding='same',kernel_initializer='glorot_uniform')(pool4)\n    conv5 = Convolution2D(128 ,3, activation=lrelu, padding='same',kernel_initializer='glorot_uniform')(conv5)\n\n    up6 = concatenate([UpSampling2D(size=(2, 2))(conv5), conv4])\n    conv6 = Convolution2D(64,  3, activation=lrelu, padding='same',kernel_initializer='glorot_uniform')(up6)\n    conv6 = Convolution2D(64,  3, activation=lrelu, padding='same',kernel_initializer='glorot_uniform')(conv6)\n\n    up7 = concatenate([UpSampling2D(size=(2, 2))(conv6), conv3])\n    conv7 = Convolution2D(32,  3, activation=lrelu, padding='same',kernel_initializer='glorot_uniform')(up7)\n    conv7 = Convolution2D(32,  3, activation=lrelu, padding='same',kernel_initializer='glorot_uniform')(conv7)\n\n    up8 = concatenate([UpSampling2D(size=(2, 2))(conv7), conv2])\n    conv8 = Convolution2D(16,  3, activation=lrelu, padding='same',kernel_initializer='glorot_uniform')(up8)\n    conv8 = Convolution2D(16, 3, activation=lrelu, padding='same',kernel_initializer='glorot_uniform')(conv8)\n\n    up9 = concatenate([UpSampling2D(size=(2, 2))(conv8), conv1])\n    conv9 = Convolution2D( 16,  3, activation=lrelu, padding='same',kernel_initializer='glorot_uniform')(up9)\n    #conv9 = Convolution2D(16,  3, activation=lrelu, padding='same',kernel_initializer='glorot_uniform')(conv9)\n    #conv9 = Convolution2D(2, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv9)\n    #conv10 = Conv2D(name='conv10', filters=1, kernel_size=(1,1), activation='sigmoid')(conv9)\n    num_output = self.config.ANCHOR_PER_GRID * (self.config.CLASSES + 1 + 4)\n\n    preds = Conv2D(\n        name='conv12', filters=num_output, kernel_size=(3, 3), strides=(1, 1), activation=None, padding=\"SAME\",\n        use_bias=True, kernel_initializer=TruncatedNormal(stddev=0.001),\n        kernel_regularizer=l2(self.config.WEIGHT_DECAY))(conv9)\n\n    #layer_final=concatenate([conv10,preds])\n\n\n    #reshape\n    pred_reshaped = Reshape((self.config.ANCHORS, -1))(preds)\n\n    #pad for loss function so y_pred and y_true have the same dimensions, wrap in lambda layer\n    pred_padded = Lambda(self._pad)( pred_reshaped)\n\n    model = Model(inputs=inputs, outputs=pred_padded)\n\n    model.summary()\n   return model\n<\/code><\/pre>\n\n<hr>\n\n<p>If I call <code>model.summary()<\/code> this is what I get:<\/p>\n\n<pre><code>Layer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput_1 (InputLayer)            (None, 512, 640, 3)  0                                            \n__________________________________________________________________________________________________\nconv2d_5 (Conv2D)               (None, 512, 640, 8)  224         input_1[0][0]                    \n__________________________________________________________________________________________________\nconv2d_6 (Conv2D)               (None, 512, 640, 8)  584         conv2d_5[0][0]                   \n__________________________________________________________________________________________________\nmax_pooling2d_1 (MaxPooling2D)  (None, 256, 320, 8)  0           conv2d_6[0][0]                   \n__________________________________________________________________________________________________\nconv2d_7 (Conv2D)               (None, 256, 320, 16) 1168        max_pooling2d_1[0][0]            \n__________________________________________________________________________________________________\nconv2d_8 (Conv2D)               (None, 256, 320, 16) 2320        conv2d_7[0][0]                   \n__________________________________________________________________________________________________\nmax_pooling2d_2 (MaxPooling2D)  (None, 128, 160, 16) 0           conv2d_8[0][0]                   \n__________________________________________________________________________________________________\nconv2d_9 (Conv2D)               (None, 128, 160, 32) 4640        max_pooling2d_2[0][0]            \n__________________________________________________________________________________________________\nconv2d_10 (Conv2D)              (None, 128, 160, 32) 9248        conv2d_9[0][0]                   \n__________________________________________________________________________________________________\nmax_pooling2d_3 (MaxPooling2D)  (None, 64, 80, 32)   0           conv2d_10[0][0]                  \n__________________________________________________________________________________________________\nconv2d_11 (Conv2D)              (None, 64, 80, 64)   18496       max_pooling2d_3[0][0]            \n__________________________________________________________________________________________________\nconv2d_12 (Conv2D)              (None, 64, 80, 64)   36928       conv2d_11[0][0]                  \n__________________________________________________________________________________________________\nmax_pooling2d_4 (MaxPooling2D)  (None, 32, 40, 64)   0           conv2d_12[0][0]                  \n__________________________________________________________________________________________________\nconv2d_13 (Conv2D)              (None, 32, 40, 128)  73856       max_pooling2d_4[0][0]            \n__________________________________________________________________________________________________\nconv2d_14 (Conv2D)              (None, 32, 40, 128)  147584      conv2d_13[0][0]                  \n__________________________________________________________________________________________________\nup_sampling2d_1 (UpSampling2D)  (None, 64, 80, 128)  0           conv2d_14[0][0]                  \n__________________________________________________________________________________________________\nconcatenate_41 (Concatenate)    (None, 64, 80, 192)  0           up_sampling2d_1[0][0]            \n                                                                 conv2d_12[0][0]                  \n__________________________________________________________________________________________________\nconv2d_15 (Conv2D)              (None, 64, 80, 64)   110656      concatenate_41[0][0]             \n__________________________________________________________________________________________________\nconv2d_16 (Conv2D)              (None, 64, 80, 64)   36928       conv2d_15[0][0]                  \n__________________________________________________________________________________________________\nup_sampling2d_2 (UpSampling2D)  (None, 128, 160, 64) 0           conv2d_16[0][0]                  \n__________________________________________________________________________________________________\nconcatenate_42 (Concatenate)    (None, 128, 160, 96) 0           up_sampling2d_2[0][0]            \n                                                                 conv2d_10[0][0]                  \n__________________________________________________________________________________________________\nconv2d_17 (Conv2D)              (None, 128, 160, 32) 27680       concatenate_42[0][0]             \n__________________________________________________________________________________________________\nconv2d_18 (Conv2D)              (None, 128, 160, 32) 9248        conv2d_17[0][0]                  \n__________________________________________________________________________________________________\nup_sampling2d_3 (UpSampling2D)  (None, 256, 320, 32) 0           conv2d_18[0][0]                  \n__________________________________________________________________________________________________\nconcatenate_43 (Concatenate)    (None, 256, 320, 48) 0           up_sampling2d_3[0][0]            \n                                                                 conv2d_8[0][0]                   \n__________________________________________________________________________________________________\nconv2d_19 (Conv2D)              (None, 256, 320, 16) 6928        concatenate_43[0][0]             \n__________________________________________________________________________________________________\nconv2d_20 (Conv2D)              (None, 256, 320, 16) 2320        conv2d_19[0][0]                  \n__________________________________________________________________________________________________\nup_sampling2d_4 (UpSampling2D)  (None, 512, 640, 16) 0           conv2d_20[0][0]                  \n__________________________________________________________________________________________________\nconcatenate_44 (Concatenate)    (None, 512, 640, 24) 0           up_sampling2d_4[0][0]            \n                                                                 conv2d_6[0][0]                   \n__________________________________________________________________________________________________\nconv2d_21 (Conv2D)              (None, 512, 640, 16) 3472        concatenate_44[0][0]             \n__________________________________________________________________________________________________\nconv12 (Conv2D)                 (None, 512, 640, 56) 8120        conv2d_21[0][0]                  \n__________________________________________________________________________________________________\nreshape_5 (Reshape)             (None, 10240, 1792)  0           conv12[0][0]                     \n__________________________________________________________________________________________________\nlambda_5 (Lambda)               (None, 10240, 1796)  0           reshape_5[0][0]                  \n==================================================================================================\nTotal params: 500,400\nTrainable params: 500,400\nNon-trainable params: 0\n__________________________________________________________________\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57593403,"AnswerCount":0,"Tags":"<python><machine-learning><keras><neural-network>","CreationDate":"2019-08-21T13:58:12.523","AcceptedAnswerId":null,"OwnerUserId":10708965.0,"Title":"Validation accuracy keeps increasing with every epoch (dense neural network in keras)","Body":"<p>I want to program a simple dense neural network in keras with the breast-cancer-dataset from sklearn.<\/p>\n\n<p>But when I plot the training and validation accuracy and loss over the Epochs, the validation accuracy keeps improving and the validation loss keeps decreasing, when really it should peak and then diverge from the training values.<\/p>\n\n<p>graph: <a href=\"https:\/\/imgur.com\/a\/Bch56BU\" rel=\"nofollow noreferrer\">https:\/\/imgur.com\/a\/Bch56BU<\/a><\/p>\n\n<p>I've tried playing around with number of layers and other variables, but the problem remains.\nI'm thinking that maybe this could be the case if the training and validation data are the same somehow? But I can't find the problem with my code.<\/p>\n\n<p>Does somebody know how to fix this? Thank you!<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from sklearn.datasets import load_breast_cancer\nfrom sklearn.preprocessing import scale\nfrom sklearn.model_selection import train_test_split\nfrom keras import models, layers#, regularizers\nimport matplotlib.pyplot as plt\n\ndataset = load_breast_cancer()\nx = dataset['data']\ny = dataset['target_names'].take(dataset['target'])\nfor index, item in enumerate(y):\n    if (item == 'malignant'):\n        y[index] = 1.\n    else:\n        y[index] = 0.\nx = scale(x)\nx_train, x_test, y_train, y_test = train_test_split(x, y, test_size=.5)\n\nmodel = models.Sequential()\nmodel.add(layers.Dense(16, activation='relu', input_shape=(30,)))\nmodel.add(layers.Dense(16, activation='relu'))\nmodel.add(layers.Dense(1, activation='sigmoid'))\n\nmodel.compile(optimizer='rmsprop',\n              loss='mse',\n              metrics=['accuracy'])\n\nx_val = x_train[:80]\npartial_x_train = x_train[80:]\ny_val = y_train[:80]\npartial_y_train = y_train[80:]\n\nhistory = model.fit(partial_x_train,\n                    partial_y_train,\n                    epochs=20,\n                    batch_size=50,\n                    validation_data=(x_val, y_val))\n\nhistory_dict = history.history\nloss_values = history_dict['loss']\nval_loss_values = history_dict['val_loss']\n\nepochs = range(1, len(history_dict['acc']) + 1)\n\nplt.plot(epochs, loss_values, 'bo', label='Training loss')\nplt.plot(epochs, val_loss_values, 'b', label='Validation loss')\nplt.title('Training and validation loss')\nplt.xlabel('Epochs')\nplt.ylabel('Loss')\nplt.legend()\nplt.show()\n\nplt.clf()\nacc_values = history_dict['acc']\nval_acc_values = history_dict['val_acc']\n\nplt.plot(epochs, acc_values, 'bo', label='Training acc')\nplt.plot(epochs, val_acc_values, 'b', label='Validation acc')\nplt.title('Training and validation accuracy')\nplt.xlabel('Epochs')\nplt.ylabel('Loss')\nplt.legend()\nplt.show()\n<\/code><\/pre>\n\n<pre><code>Train on 204 samples, validate on 80 samples\nEpoch 1\/20\n204\/204 [==============================] - 2s 8ms\/step - loss: 0.1684 - acc: 0.7941 - val_loss: 0.1735 - val_acc: 0.7750\nEpoch 2\/20\n204\/204 [==============================] - 0s 39us\/step - loss: 0.1387 - acc: 0.8529 - val_loss: 0.1571 - val_acc: 0.8000\nEpoch 3\/20\n204\/204 [==============================] - 0s 39us\/step - loss: 0.1253 - acc: 0.8775 - val_loss: 0.1452 - val_acc: 0.8500\nEpoch 4\/20\n204\/204 [==============================] - 0s 39us\/step - loss: 0.1156 - acc: 0.8775 - val_loss: 0.1354 - val_acc: 0.8625\nEpoch 5\/20\n204\/204 [==============================] - 0s 39us\/step - loss: 0.1074 - acc: 0.8775 - val_loss: 0.1263 - val_acc: 0.8625\nEpoch 6\/20\n204\/204 [==============================] - 0s 39us\/step - loss: 0.1003 - acc: 0.8824 - val_loss: 0.1165 - val_acc: 0.8625\nEpoch 7\/20\n204\/204 [==============================] - 0s 39us\/step - loss: 0.0929 - acc: 0.8922 - val_loss: 0.1111 - val_acc: 0.8750\nEpoch 8\/20\n204\/204 [==============================] - 0s 39us\/step - loss: 0.0883 - acc: 0.8971 - val_loss: 0.1045 - val_acc: 0.8875\nEpoch 9\/20\n204\/204 [==============================] - 0s 39us\/step - loss: 0.0836 - acc: 0.9069 - val_loss: 0.0985 - val_acc: 0.8875\nEpoch 10\/20\n204\/204 [==============================] - 0s 39us\/step - loss: 0.0782 - acc: 0.9216 - val_loss: 0.0935 - val_acc: 0.8875\nEpoch 11\/20\n204\/204 [==============================] - 0s 39us\/step - loss: 0.0738 - acc: 0.9314 - val_loss: 0.0890 - val_acc: 0.8875\nEpoch 12\/20\n204\/204 [==============================] - 0s 39us\/step - loss: 0.0698 - acc: 0.9363 - val_loss: 0.0847 - val_acc: 0.8875\nEpoch 13\/20\n204\/204 [==============================] - 0s 39us\/step - loss: 0.0664 - acc: 0.9412 - val_loss: 0.0812 - val_acc: 0.8875\nEpoch 14\/20\n204\/204 [==============================] - 0s 44us\/step - loss: 0.0632 - acc: 0.9363 - val_loss: 0.0776 - val_acc: 0.9000\nEpoch 15\/20\n204\/204 [==============================] - 0s 34us\/step - loss: 0.0604 - acc: 0.9461 - val_loss: 0.0742 - val_acc: 0.9000\nEpoch 16\/20\n204\/204 [==============================] - 0s 39us\/step - loss: 0.0578 - acc: 0.9461 - val_loss: 0.0702 - val_acc: 0.9000\nEpoch 17\/20\n204\/204 [==============================] - 0s 39us\/step - loss: 0.0549 - acc: 0.9510 - val_loss: 0.0672 - val_acc: 0.9000\nEpoch 18\/20\n204\/204 [==============================] - 0s 39us\/step - loss: 0.0525 - acc: 0.9510 - val_loss: 0.0642 - val_acc: 0.9000\nEpoch 19\/20\n204\/204 [==============================] - 0s 39us\/step - loss: 0.0500 - acc: 0.9559 - val_loss: 0.0599 - val_acc: 0.9125\nEpoch 20\/20\n204\/204 [==============================] - 0s 39us\/step - loss: 0.0471 - acc: 0.9559 - val_loss: 0.0571 - val_acc: 0.9250\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57593509,"AnswerCount":0,"Tags":"<tensorflow><tflearn>","CreationDate":"2019-08-21T14:04:10.433","AcceptedAnswerId":null,"OwnerUserId":11122879.0,"Title":"When I try to import tflearn, jupyter notebook generates error","Body":"<p>I installed tensorflow and tflearn. I use python version 3.7. I can import tensorflow fine but when I try to import tflearn, I get this error:<\/p>\n\n<pre><code>ModuleNotFoundError: No module named 'tensorflow.contrib'\n<\/code><\/pre>\n\n<p>Can anyone please help?<\/p>\n","answers":[]}
+{"QuestionId":57593713,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2019-08-21T14:15:17.150","AcceptedAnswerId":null,"OwnerUserId":9371809.0,"Title":"Using tf.data.Dataset is slow for small batch sizes","Body":"<p>I am reading a large dataset consisting of multiple csv files into a tf.data.Dataset, and doing some transformations on the data. <\/p>\n\n<p>This dataset is then used to train a keras model. I had the problem that the training was running very slow since the call to parse_csv was done for every line. As described here <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/18784\" rel=\"nofollow noreferrer\">link<\/a>, I have been  speeding up my code by doing batching before calling parse_csv. My problem now is that I need to use small batches for training, and then I loose the speed up from this solution. <\/p>\n\n<p>Is there a way of using large batches when dataset.map(parse_csv) is called, and then rebatch it into smaller batches before training my model? Or is there another solution to this problem.<\/p>\n\n<pre><code>     def parse_csv(value_column):\n        cols = tf.io.decode_csv(value_column, record_defaults=[[0.0],[0.0],[0],[0],[0],[0.0],[0],[0]])\n        features=[cols[2],cols[5]]\n        return features,label\n\n     # Create list of files that match pattern\n     file_list = tf.io.gfile.glob(filename)\n     #My old way of reading in the data\n     #dataset=tf.data.TextLineDataset(file_list).map(parse_csv)\n     #dataset=dataset.batch(batch_size)\n     # I changed to doing batching before calling parse_csv to increase speed\n     dataset = tf.data.TextLineDataset(file_list)\n     dataset=dataset.batch(batch_size).map(parse_csv)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57593751,"AnswerCount":0,"Tags":"<c++><tensorflow><keras>","CreationDate":"2019-08-21T14:17:43.617","AcceptedAnswerId":null,"OwnerUserId":8710118.0,"Title":"What is the best way to use Keras in C++ application?","Body":"<p>I would like to use keras in C++. Because Keras using Tensorflow backend, I thought it is easy to convert my Keras code to Tensorflow.<\/p>\n\n<p>I only want to predict in C++, so it's okay to train in python, but I would like to hide the model structure (in the file), so the C++ app should know it.<\/p>\n\n<p>What I think, is the best way to use:<\/p>\n\n<ol>\n<li>convert model to tensorflow<\/li>\n<li>use tensorflow c++ api<\/li>\n<li>implement the predict method and the model in c++<\/li>\n<li>predict with c++ api<\/li>\n<\/ol>\n\n<p>If you met with a similar problem, and found a solution, don't hesitate to share with me. Thanks<\/p>\n\n<p>I already checked <a href=\"https:\/\/github.com\/gosha20777\/keras2cpp\" rel=\"nofollow noreferrer\">keras2cpp<\/a>, <a href=\"https:\/\/github.com\/moof2k\/kerasify\" rel=\"nofollow noreferrer\">kerasify<\/a>, <a href=\"https:\/\/github.com\/Dobiasd\/frugally-deep\" rel=\"nofollow noreferrer\">frugally-deep<\/a>, but none of them were good, because only use CPU (which is slow), or doesn't support the layer I use.\nSo far, I neither could setup tensorflow in c++ with Visual Studio.<\/p>\n","answers":[]}
+{"QuestionId":57593940,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><computer-vision>","CreationDate":"2019-08-21T14:31:29.693","AcceptedAnswerId":null,"OwnerUserId":6405262.0,"Title":"Training mask_rcnn_inception_v2_coco with tensorflowAPI","Body":"<p>I have trained the mask_rcnn_inception_v2_coco network from the tensorflow API, and everything seems to be working as it should, that is until I start to predict, this is a segmentation network, but when I predict the segmentation is always equal to the bounding box. <\/p>\n\n<p>I have created masks, where multiple objects are in the picture, and have created tfrecords with that uses the xmin, ymin and xmax and ymax from the mask to create a bounding box. <\/p>\n\n<p>A picture of the masks can be seen below. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/wjt4U.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/wjt4U.png\" alt=\"Example of masks created for a image.\"><\/a><\/p>\n\n<p>When I create the tfrecord, I load the original image, this mask, and all the different bounding boxes. <\/p>\n\n<p>I use the following script to predict images:<\/p>\n\n<pre><code>def run_inference(graph):\n    results = {}\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        for i, image_path in enumerate(os.listdir(PATH_TO_TEST_IMAGES)):\n            image = Image.open(os.path.join(PATH_TO_TEST_IMAGES, 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            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\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 coordinates and fit the image size.\n                real_num_detection = tf.cast(tensor_dict['num_detections'][0], tf.int32)\n                detection_boxes = tf.slice(detection_boxes, [0, 0], [real_num_detection, -1])\n                detection_masks = tf.slice(detection_masks, [0, 0, 0], [real_num_detection, -1, -1])\n                detection_masks_reframed = utils_ops.reframe_box_masks_to_image_masks(\n                    detection_masks, detection_boxes, image_np.shape[0], image_np.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 = 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            # 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]\n            visualize_masks_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=0)\n            plt.figure(figsize=IMAGE_SIZE)\n            plt.imshow(image_np)\n            plt.imsave(\"{i}.png\".format(i=i), image_np, format=\"png\")\n            results[i] = output_dict\nreturn results\n<\/code><\/pre>\n\n<p>visualize_masks_and_labels_on_image_array looks like this:<\/p>\n\n<pre><code>def visualize_masks_and_labels_on_image_array(\n   image,\n   boxes,\n   classes,\n   scores,\n   category_index,\n   instance_masks=None,\n   instance_boundaries=None,\n   keypoints=None,\n   track_ids=None,\n   use_normalized_coordinates=False,\n   max_boxes_to_draw=20,\n   min_score_thresh=.5,\n   agnostic_mode=False,\n   line_thickness=4,\n   groundtruth_box_visualization_color='black',\n   skip_scores=False,\n   skip_labels=False,\n   skip_track_ids=False):   \n    box_to_display_str_map = collections.defaultdict(list)\n    box_to_color_map = collections.defaultdict(str)\n    box_to_instance_masks_map = {}\n    box_to_instance_boundaries_map = {}\n    box_to_keypoints_map = collections.defaultdict(list)\n    box_to_track_ids_map = {}\n    if not max_boxes_to_draw:\n       max_boxes_to_draw = boxes.shape[0]\n    for i in range(min(max_boxes_to_draw, boxes.shape[0])):\n      if scores is None or scores[i] &gt; min_score_thresh:\n         box = tuple(boxes[i].tolist())\n         if instance_masks is not None:\n            box_to_instance_masks_map[box] = instance_masks[i]\n         if instance_boundaries is not None:\n            box_to_instance_boundaries_map[box] = instance_boundaries[i]\n         if keypoints is not None:\n            box_to_keypoints_map[box].extend(keypoints[i])\n         if track_ids is not None:\n            box_to_track_ids_map[box] = track_ids[i]\n         if scores is None:\n            box_to_color_map[box] = groundtruth_box_visualization_color\n         else:\n           display_str = ''\n           if not skip_labels:\n             if not agnostic_mode:\n                if classes[i] in category_index.keys():\n                   class_name = category_index[classes[i]]['name']\n                else:\n                   class_name = 'N\/A'\n                display_str = str(class_name)\n            if not skip_scores:\n               if not display_str:\n                  display_str = '{}%'.format(int(100*scores[i]))\n               else:\n                  display_str = '{}: {}%'.format(display_str, int(100*scores[i]))\n           if not skip_track_ids and track_ids is not None:\n              if not display_str:\n                 display_str = 'ID {}'.format(track_ids[i])\n              else:\n                 display_str = '{}: ID {}'.format(display_str, track_ids[i])\n             box_to_display_str_map[box].append(display_str)\n          if agnostic_mode:\n             box_to_color_map[box] = 'DarkOrange'\n          elif track_ids is not None:\n             prime_multipler = vis_util._get_multiplier_for_color_randomness()\n             box_to_color_map[box] = vis_util.STANDARD_COLORS[\n          (prime_multipler * track_ids[i]) % len(vis_util.STANDARD_COLORS)]\n          else:\n             box_to_color_map[box] = vis_util.STANDARD_COLORS[\n          classes[i] % len(vis_util.STANDARD_COLORS)]\n\n     # Draw all boxes onto image.\n    for box, color in box_to_color_map.items():\n       ymin, xmin, ymax, xmax = box\n       if instance_masks is not None:\n           vis_util.draw_mask_on_image_array(\n             image,\n             box_to_instance_masks_map[box],\n             color=color\n           )\n       if instance_boundaries is not None:\n           vis_util.draw_mask_on_image_array(\n            image,\n            box_to_instance_boundaries_map[box],\n            color='red',\n            alpha=1.0\n          )\n\n       if keypoints is not None:\n          vis_util.draw_keypoints_on_image_array(\n            image,\n            box_to_keypoints_map[box],\n            color=color,\n            radius=line_thickness \/ 2,\n      use_normalized_coordinates=use_normalized_coordinates)\n\n return image\n<\/code><\/pre>\n\n<p>This is just a quick edit of Tensorflow's vis_util.visualize_boxes_and_labels_on_image_array, without the bounding boxes. This gives me the following output: \n<a href=\"https:\/\/i.stack.imgur.com\/4V5wI.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/4V5wI.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Which in my view indicates that the network just fills the entire bounding box, and is happy with that.. <\/p>\n\n<p>I can include the generate-tfrecords script as well if it is needed. <\/p>\n\n<p>All help will be appreciated. <\/p>\n","answers":[]}
+{"QuestionId":57594131,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-21T14:42:50.510","AcceptedAnswerId":null,"OwnerUserId":9080367.0,"Title":"Determine connected components of prediction inside of custom loss metric keras","Body":"<p>I want to make a modified binary crossentropy loss function Keras that penalizes the model for a large number of islands in the prediction output. Output is 128x128x128. I have a function that counts the amount of islands in an input numpy array, using measure.label(), and it works perfectly. The problem is when I call this function from inside my loss function, the input that the loss function receives is ALWAYS a None value, and I have no idea what is causing this. <\/p>\n\n<p>Here's my loss:<\/p>\n\n<pre><code>eps = 2\ndef custom_binary_crossentropy(y_true,y_pred):\n        bc = K.binary_crossentropy(y_true, y_pred)\n\n        ytyp_x = K.shape(y_true)[0]\n\n        yt = K.reshape(y_true, (ytyp_x,128,128))\n        yp = K.reshape(y_pred, (ytyp_x,128,128))\n\/\/ after this operation i KNOW that the yp and yt values are 3d like they should be, \n\/\/but when I pass them into the num_islands function they are interpreted as None values. \n\/\/I know this because I filter out the None values in the num_islands function\n\/\/and return something different if the input is None\n\n        islands = num_islands(yp,0)\n        islands = tf.dtypes.cast(islands,dtype=tf.float32)\n\n\n        loss = bc*(K.log(islands + eps))\n        return(loss)\n<\/code><\/pre>\n\n<p>The comments sort of sum up the state of the problem.<\/p>\n\n<p>Any help is GREATLY appreciated. Thanks!<\/p>\n","answers":[]}
+{"QuestionId":57594207,"AnswerCount":0,"Tags":"<python><tensorflow><object-detection>","CreationDate":"2019-08-21T14:47:34.730","AcceptedAnswerId":null,"OwnerUserId":4582711.0,"Title":"No object detection after exporting the graph","Body":"<p>I've trained the <code>ssd_mobilenet_v2_coco<\/code> for my custom dataset, but the problem is that during the training tensorbaord shows detection in the image. It's something like this:\n<a href=\"https:\/\/i.stack.imgur.com\/m2FBO.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/m2FBO.png\" alt=\"ip\"><\/a> <\/p>\n\n<p>But when I export the graph and do inference I'm not detecting anything in the image (same image as training). For exporting the graph I'm using <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/export_inference_graph.py\" rel=\"nofollow noreferrer\">this<\/a> script. Here, is my config file:<\/p>\n\n<pre><code>model {\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: 600\n        width: 600\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: 1\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      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: 1\n  optimizer {\n    rms_prop_optimizer: {\n      learning_rate: {\n        manual_step_learning_rate {\n          initial_learning_rate: 0.005\n          schedule {\n            step: 90000\n            learning_rate: .0004\n          }\n          schedule {\n            step: 200000\n            learning_rate: .00003\n          }\n          schedule {\n            step: 600000\n            learning_rate: .000003\n          }\n        }\n      }\n      momentum_optimizer_value: 0.9\n      decay: 0.9\n      epsilon: 1.0\n    }\n  }\n  fine_tune_checkpoint: \"\/home\/deploy\/ved\/ssd_mobilenet_v2_coco_2018_03_29\/model.ckpt-39800\"\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: 1200000\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: \"\/home\/deploy\/ved\/train.record\"\n  }\n  label_map_path: \"\/home\/deploy\/ved\/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: \"\/home\/deploy\/ved\/val.record\"\n  }\n  label_map_path: \"\/home\/deploy\/ved\/label_map.pbtxt\"\n  shuffle: false\n  num_readers: 1\n}\n<\/code><\/pre>\n\n<p>Also, for creating the tfrecord I'm using following code:<\/p>\n\n<pre><code>import cv2\nimport os\nimport json\nimport numpy as np\nfrom IPython import embed\nimport tensorflow as tf\nfrom object_detection.utils import dataset_util\n\njson_path = \"\/home\/deploy\/ved\/train.json\"\nout_path = \"\/home\/deploy\/ved\/out\"\ncommodity_lst = [\"tomato\"]\nwriter = tf.python_io.TFRecordWriter(\"train.record\")\ntarget_size = 1200\nlabel_map = \"\/home\/deploy\/ved\/label_map.pbtxt\"\nwrite_label_map = True\n\ncategory = ['tomato']\n\ncategory_index = {k:v for k,v in zip(range(1,len(category)+1), category)}\n\nwith open(json_path, 'r') as f:\n    data = json.load(f)\n\nif write_label_map:\n    with open(label_map, \"w\") as f_pbtxt:\n        for k,v in category_index.items():\n            f_pbtxt.write(\"item{{\\n\\tid: {}\\n\\tname: '{}'\\n}}\\n\".format(k,v))\n\ndef checkJPG(fn):\n    with tf.Graph().as_default():\n        try:\n            image_contents = tf.read_file(fn)\n            image = tf.image.decode_jpeg(image_contents, channels=3)\n            init_op = tf.initialize_all_tables()\n            with tf.Session() as sess:\n                sess.run(init_op)\n                tmp = sess.run(image)\n        except:\n            print(\"Corrupted file: \", fn)\n            return False\n    return True\n\ndef getSquareImage( img, target_width = 500 ):\n    width = img.shape[1]\n    height = img.shape[0]\n\n    square = np.zeros( (target_width, target_width, 3), dtype=np.uint8 )\n\n    max_dim = width if width &gt;= height else height\n    scale = target_width \/ max_dim\n\n    if ( width &gt;= height ):\n        width = target_width\n        x = 0\n        height = int(height * scale)\n        y = int(( target_width - height ) \/ 2)\n    else:\n        y = 0\n        height = target_width\n        width = int(width * scale)\n        x = int(( target_width - width ) \/ 2)\n    # embed()\n    square[y:y+height, x:x+width] = cv2.resize( img , (width, height) )\n\n    return square, x, y, scale\n\nfor key in data:\n    img = cv2.imread(os.path.join(\"\/home\/deploy\/ved\/dataset\", data[key][\"filename\"]))\n    if checkJPG(os.path.join(\"\/home\/deploy\/ved\/dataset\", data[key][\"filename\"])):\n        _img, _x, _y, _scale = getSquareImage(img, target_size)\n        encoded_jpg = cv2.imencode('.jpg', _img)[1].tostring()\n        filename = tf.compat.as_bytes(data[key][\"filename\"])\n        image_format = b'jpg'\n        xmins = []\n        xmaxs = []\n        ymins = []\n        ymaxs = []\n        classes_text = []\n        classes = []\n        width = _img.shape[1]\n        height = _img.shape[0]\n\n        for j in range(len(data[key][\"regions\"])):\n            x_lst = np.asarray(data[key][\"regions\"][j][\"shape_attributes\"][\"all_points_x\"])\n            y_lst = np.asarray(data[key][\"regions\"][j][\"shape_attributes\"][\"all_points_y\"])\n            xmin = int(np.amin(x_lst)*_scale) + _x\n            xmax = int(np.amax(x_lst)*_scale) + _x\n            ymin = int(np.amin(y_lst)*_scale) + _y\n            ymax = int(np.amax(y_lst)*_scale) + _y\n\n            cv2.rectangle(_img, (xmin, ymin), (xmax, ymax), (255,0,0), 2)\n            try:\n                _commodity = [i for i in commodity_lst if i in data[key][\"regions\"][j][\"region_attributes\"]][0]\n            except:\n                print(\"commodity name not in list\")\n                embed()\n            if _commodity == data[key][\"regions\"][j][\"region_attributes\"][_commodity]:\n                text = _commodity\n\n                try:\n                    index = [k for k,v in category_index.items() if v == text.strip()][0]\n                except:\n                    print(\"one tagged catogory not in list\")\n                    embed()\n                xmins.append( xmin \/ width)\n                xmaxs.append( xmax \/ width)\n                ymins.append( ymin \/ height)\n                ymaxs.append( ymax \/ height)\n                classes_text.append(tf.compat.as_bytes(text))\n                classes.append(index)\n            # cv2.putText(_img, text, (xmin, ymin), cv2.FONT_HERSHEY_SIMPLEX, 1.0, (255, 255, 255), lineType=cv2.LINE_AA)\n        # _path = os.path.join(out_path, os.path.splitext(data[key][\"filename\"])[0]+\"_ann\"+os.path.splitext(data[key][\"filename\"])[1])\n        # cv2.imwrite(_path, _img)\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        writer.write(tf_example.SerializeToString())\n\nwriter.close()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57594301,"AnswerCount":0,"Tags":"<python><tensorflow><optimization><minimize>","CreationDate":"2019-08-21T14:52:44.103","AcceptedAnswerId":null,"OwnerUserId":9747182.0,"Title":"How to repeately call apply_gradients in a customized minimize function of a Tensorflow optimizer?","Body":"<p>I Dear all, I want to implement a customized minimize function for my Tensorflow Optimizer. For this, I call compute_gradients and apply_gradients in a for-loop. In each loop, the var_list should be overwritten so that in the next loop iteration compute_gradients is called on a different set of parameters. At the end, I want to calculate the average of all variables of all loop iterations and return it as the result.<\/p>\n\n<p>However, I noticed that after calling apply_gradients the var_list is not changed. Instead, the same parameters are used over and over again in my for-loop. How can I force the change of variables of var_list before the next loop iteration starts?<\/p>\n\n<p>I already tried a lot :') :<\/p>\n\n<p>1.) I tried with tf.get_default_graph().control_dependencies(var_update): where var_update = self.apply_gradients(grads_and_vars)<\/p>\n\n<p>2.) I tried to manually overwrite the variables with: \nfor v in var_list:\n   assign_op = tf.assign_sub(var, self._eta * grad)\n   with tf.get_default_graph().control_dependencies([assign_op]):\n    ...<\/p>\n\n<p>My current non-working implementation:<\/p>\n\n<p>for i in range(1, 100):<\/p>\n\n<pre><code>self._eta = 1 \/ (self._sigma * i)\n\ngrads_and_vars = self.compute_gradients(loss, var_list)\n\nvar_update = self.apply_gradients(grads_and_vars, global_step=global_step, name=name)\n\nwith tf.get_default_graph().control_dependencies(var_update):\n\n    continue\n<\/code><\/pre>\n\n<p>return var_list<\/p>\n\n<p>I expect to see a changed var_list. However, var_list did not change at all from its input values (2, -2).<\/p>\n\n<p>Thanks for your help!\nEpsilon<\/p>\n","answers":[]}
+{"QuestionId":57594453,"AnswerCount":0,"Tags":"<java><python><tensorflow><keras>","CreationDate":"2019-08-21T15:02:12.240","AcceptedAnswerId":null,"OwnerUserId":11957492.0,"Title":"Running Keras python script through java results in sudden stop mid-script","Body":"<p>I have a script <code>predict_cnn.py<\/code> that takes an Image as Input and prints the prediction of that Image using an already trained keras model. The script works when run in console, now I want to run the script from a Java class to use its Output. However, when I try to read the Output of said script through Java, it simply stops during the loading of the keras model.<\/p>\n\n<p>I'm using <code>python3.7<\/code>, <code>keras2.2.4<\/code>, <code>tensorflow1.14<\/code> and <code>opencv-contrib-python 4.1.0.25<\/code> for my python Project and plain Java11 for the Java Project.\nI've tested passing simple \"Hello World\" python scripts to my Java class, as well as scripts that had multiple Input Arguments and that loaded the keras model\/label binarizer with prints afterwards, all of those Tests surprisingly worked. I really don't understand why it doesn't work for <code>predict_cnn.py<\/code>.<\/p>\n\n<p>Here's the Python script that prints the keras prediction to console:<\/p>\n\n<pre><code>    ap = argparse.ArgumentParser()\n    ap.add_argument(\"-i\", \"--image\", required=True, help=\"Path to input image\")\n    ap.add_argument(\"-m\", \"--model\", default=\"output\\invoice_cnn_250_ADD100-    BS256.model\", help=\"Path to trained Keras model\")\n    ap.add_argument(\"-lb\", \"--label-bin\", default=\"output\\invoice_cnn_lb_250_Add100-BS256.pickle\", help=\"Path to input label Binarizer\")\n    args = vars(ap.parse_args())\n    print(\"Args used:\"+args.__str__())\n\n    # load input image and resize it to the target spatial dimensions\n    print(\"[INFO] reading image...\")\n    image = cv2.imread(args[\"image\"])\n    output = image.copy()\n    image = cv2.resize(image, (64, 64))\n\n    # scale the pixel values to [0, 1]\n    print(\"[INFO] processing image...\")\n    image = image.astype(\"float\") \/ 255.0\n    image = image.reshape((1, image.shape[0], image.shape[1],\n                       image.shape[2]))\n\n    print(\"[INFO] loading network and label binarizer...\")\n    model = load_model(args[\"model\"])\n    lb = pickle.loads(open(args[\"label_bin\"], \"rb\").read())\n\n    print(\"[INFO] predict image...\")\n    pred = model.predict(image)\n\n    print(\"[INFO] process prediction...\")\n    i = pred.argmax(axis=1)[0]\n    label = lb.classes_[i]\n\n    print(\"[INFO] prediction complete\")\n    result = pred[0][i] * 100\n    print(label + \"|\" + result.__str__())\n<\/code><\/pre>\n\n<p>And here's the Problem Java class that reads script output:<\/p>\n\n<pre><code>\n    public static void main(String[] args) throws IOException {\n        \/\/ set up the command and parameter\n        String pythonScriptPath = \"PATH\/TO\/predict_cnn.py\";\n        String[] cmd = new String[2];\n        cmd[0] = \"python3.7\";\n        cmd[1] = pythonScriptPath;\n\n        \/\/ create runtime to execute external command\n        Runtime rt = Runtime.getRuntime();\n        Process pr = rt.exec(cmd);\n\n        \/\/ retrieve output from python script\n        BufferedReader bfr = new BufferedReader(new InputStreamReader(pr.getInputStream()));\n        String line = \"\";\n        while((line = bfr.readLine()) != null) {\n        \/\/ display each output line form python script\n            System.out.println(line);\n        }\n        System.out.println(Arrays.toString(cmd));\n    }\n\n<\/code><\/pre>\n\n<p>Expected Output is the Output I get from running the script in console:<\/p>\n\n<pre><code>\n    Args used:{'image': 'img\\\\leftover\\\\invoice\\\\i0.jpg', 'model': 'output\\\\invoice_cnn_250_ADD100-BS256.model', 'label_bin': 'output\\\\invoice_cnn_lb_250_Add100-BS256.pickle'}\n    [INFO] reading image...\n    [INFO] processing image...\n    [INFO] loading network and label binarizer...\n    [INFO] predict image...\n    [INFO] process prediction...\n    [INFO] prediction complete\n    invoice|88.56740593910217\n<\/code><\/pre>\n\n<p>But if I run it through Java the process simply stops in the middle.<\/p>\n\n<pre><code>    Args used:    {'image': 'img\\\\leftover\\\\invoice\\\\i0.jpg', 'model': 'output\\\\invoice_cnn_250_ADD100-BS256.model', 'label_bin': 'output\\\\invoice_cnn_lb_250_Add100-BS256.pickle'}\n    [INFO] reading image...\n    [INFO] processing image...\n    [INFO] loading network and label binarizer\u2026\n\n    Process finished with exit code 0\n<\/code><\/pre>\n\n<p>I repeat, loading the keras model and binarizer in other python scripts worked, so I'm really confused why it stops at this Point. <\/p>\n\n<p>I tried running the same Code on a friend's pc, but he couldn't even get the Images to be read. The Tests worked tho. Another friend got the same Output as me, but the process just kept on running\u2026<\/p>\n\n<p>I could really need some help. Thanks.<\/p>\n","answers":[]}
+{"QuestionId":57594629,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-21T15:12:02.150","AcceptedAnswerId":57595745.0,"OwnerUserId":7528974.0,"Title":"How to set each individual matrix value depending on condition in tensorflow","Body":"<p>In Tensorflow, I want to create a matrix based whose values depend on an element wise comparison between two other matrix. For example:<\/p>\n\n<p>I want C to have value 1 when A > B and 0 otherwise:<\/p>\n\n<p>I've found a solution by exploiting the behaviour of tf.math.divide_no_nan and using some math:<\/p>\n\n<pre><code>(condition \/ condition) * result_if_true + (1 - condition \/ condition) * result_if_false\n<\/code><\/pre>\n\n<p>Which translates to this in tensorflow:<\/p>\n\n<pre><code>tf.math.add(tf.math.multiply(tf.math.divide_no_nan(condition, condition), result_if_true), tf.math.multiply(tf.math.subtract(tf.constant(1, tf.float32, matrix_shape), tf.math.divide_no_nan(condition, condition)), result_if_false))\n<\/code><\/pre>\n\n<p>But this seems so convoluted that I'm sure there must be a proper way to solve this. I've tried using tf.cond but it does not work for me as I need the condition to be evaluated element-by-element.<\/p>\n\n<p>Here is what I want to accomplish following the example I presented at the start of the question:<\/p>\n\n<pre><code>A = tf.convert_to_tensor([[1.0, 2.0], [4.0, 5.0]])\nB = tf.convert_to_tensor([[3.0, 3.0], [3.0, 3.0]])\nC = #Solution goes here\nprint(C.run())\n<\/code><\/pre>\n\n<p>Result:<\/p>\n\n<pre><code>[[0,0],\n[1,1]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"57595745","CreationDate":"2019-08-21T16:21:10.540","ParentId":null,"OwnerUserId":"7528974","Title":null,"Body":"<p>tf.where is the solution I was looking for.<\/p>\n"}]}
+{"QuestionId":57594649,"AnswerCount":1,"Tags":"<python><tensorflow><keras><neural-network><loss-function>","CreationDate":"2019-08-21T15:12:54.610","AcceptedAnswerId":null,"OwnerUserId":556014.0,"Title":"Implementing a custom loss function in Keras with a term like l1 regularisation","Body":"<p>I am creating my own loss function (that I want to use in eager execution in Keras).\nI would like to add to it a term similar to a l1 loss function.<\/p>\n\n<p>That is the loss function I am using now<\/p>\n\n<pre><code>def loss(model, x, y, x_dev, y_dev, variables):\n  y_ = model(x)\n  y_dev_ = model(x_dev)\n\n  y_temp = 1.5\n\n  return loss_mae(y_true=y, y_pred=y_)+y_temp*\n                K.mean(tf.convert_to_tensor(variables))\n<\/code><\/pre>\n\n<p>with<\/p>\n\n<pre><code>import keras.backend as K\ndef loss_mae(y_true, y_pred):\n    return K.mean(K.abs(y_pred-y_true))\n<\/code><\/pre>\n\n<p>my idea is to add to my loss function a constant (<code>y_temp<\/code>) and then I would like to multiply it for the trainable variables (to achieve something similar to a l1 regularisation term).<\/p>\n\n<p>I tried passing to the <code>loss()<\/code> function <code>model.trainable_variables<\/code> but that does not work and I get a<\/p>\n\n<pre><code>TypeError: can't multiply sequence by non-int of type 'numpy.float32'\n<\/code><\/pre>\n\n<p>anyone has any suggestions?<\/p>\n","answers":[{"AnswerId":"57595170","CreationDate":"2019-08-21T15:41:53.917","ParentId":null,"OwnerUserId":"3157737","Title":null,"Body":"<p>The reason you are getting this error is because in Python your expression\n    y_temp*variables\nmeans something like \"take 'variables' 'y_temp' times and merge them in one sequence\". In other words, 2 * [1, 2, 3] is not [4, 5, 6], but [1, 2, 3, 1, 2, 3]. Obviously, it makes little sense to use anything but integer in this case. \nIf I understand you correctly, you are trying to perform element-wise multiplication. To achieve this you should use list comprehension, something like [x * 1.5 for x in [1, 2, 3]]<\/p>\n\n<p>By the way, you could also check these answers to similar questions:\n<a href=\"https:\/\/stackoverflow.com\/questions\/16994648\/python-cant-multiply-sequence-by-non-int-of-type-float\">Python can&#39;t multiply sequence by non-int of type &#39;float&#39;<\/a><a href=\"https:\/\/stackoverflow.com\/questions\/16994648\/python-cant-multiply-sequence-by-non-int-of-type-float\"><\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/35166633\/how-do-i-multiply-each-element-in-a-list-by-a-number\/35166661\">How do I multiply each element in a list by a number?<\/a><\/p>\n"}]}
+{"QuestionId":57595080,"AnswerCount":1,"Tags":"<keras><deep-learning><keras-layer>","CreationDate":"2019-08-21T15:36:43.103","AcceptedAnswerId":null,"OwnerUserId":6804361.0,"Title":"Accessing Variables of Custom Layers in Keras","Body":"<p>Let's say we have a custom layer in Keras like this:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nfrom keras import backend as K\nfrom keras.layers import Layer\n\n\nclass Custom_Layer(Layer):\n    def __init__(self,**kwargs):\n        super(ProbabilisticActivation, self).__init__(**kwargs)\n        self.params_1 = 0\n        self.params_2 = 0\n    def build(self, input_shape):\n        self.params_1 = K.variable(np.zeros(shape=input_shape[1::]))\n        self.params_2 = K.variable(np.zeros(shape=input_shape[1::]))\n        super(Custom_Layer,self).build(input_shape) \n\n    def call(self, x, training=None): \n       # DO SOMETHING\n<\/code><\/pre>\n\n<p>How could I access the value of the parameters (params_1, params_2) in the training process? I tried to get parameters by using <strong>model.get_layer('Name of Custom Layer').params_1<\/strong>, but in this case, I can not access the value of the parameters.<\/p>\n\n<p>Here is the model architecture:<\/p>\n\n<pre><code>def get_model(img_height, img_width:\n    input_layer = Input(shape=(img_height, img_width, 3))\n    x = Conv2D(32, (3, 3), padding='same', name='conv2d_1', activation='relu')(input_layer)\n    x = Custom_Layer()(x)\n    x = MaxPooling2D(pool_size=(2, 2))(x)\n    x = Dropout(0.25)(x)\n    x = Conv2D(64, kernel_size=(3, 3), name='conv2d_2', activation='relu')(x)\n    x = Conv2D(64, (3, 3), name='conv2d_4', activation='relu')(x)\n    x = MaxPooling2D(pool_size=(2, 2))(x)\n    x = Dropout(0.25)(x)\n    x = Flatten()(x)\n    x = Dense(512)(x)\n    x = Activation('relu')(x)\n    x = Dropout(0.5)(x)\n    x = Dense(10)(x)\n    x = Activation('softmax')(x)\n    model = Model(inputs=[input_layer], outputs=[x])\n    model.summary()\n\n    return model\n<\/code><\/pre>\n","answers":[{"AnswerId":"57605052","CreationDate":"2019-08-22T08:24:34.110","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>Note that <code>params_1<\/code> and <code>params_2<\/code> are TensorFlow tensors. To get their value, you should run them within a <code>tf.Session<\/code>. You could do something along the lines of:<\/p>\n\n<pre><code>from keras import backend as K\n\n# ... train model\n\nsess = K.get_session()\nparams_1 = model.get_layer('Name of Custom Layer').params_1\nvalues_1 = sess.run(params_1)\nprint(values_1)\n<\/code><\/pre>\n\n<p><strong>NOTE:<\/strong> Not tested.<\/p>\n"}]}
+{"QuestionId":57595190,"AnswerCount":0,"Tags":"<python><image><tensorflow>","CreationDate":"2019-08-21T15:42:57.453","AcceptedAnswerId":null,"OwnerUserId":11207042.0,"Title":"Printing Image tensor in tensorflow","Body":"<p>I am trying to print the image tensor which was generated from 'DatasetDataProvider'. I tried to use tf.print for that but nothing is getting printed. I am not getting where I am doing wrong.<\/p>\n\n<pre><code>   dataset = datasets_imagenet.get_split(split_name, dataset_dir)\n   provider = dataset_data_provider.DatasetDataProvider(\n        dataset,\n        num_readers=num_readers,\n        shuffle=is_training,\n        common_queue_capacity=5 * batch_size,\n        common_queue_min=batch_size)\n    [image, bbox, label] = provider.get(['image', 'object\/bbox', 'label'])\n    #print(\"&amp;&amp;&amp;&amp;&amp;&amp;&amp;&amp;&amp;&amp;----------\")\n    sess = tf.Session()\n    with sess.as_default():\n        tf.print(image, [image])\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57595425,"AnswerCount":0,"Tags":"<python><tensorflow><gpu><predict>","CreationDate":"2019-08-21T15:59:32.217","AcceptedAnswerId":null,"OwnerUserId":9232100.0,"Title":"How to have cpu-only parts of cpu\/gpu BERT tensorflow \"predict\" processing use multiple CPUs?","Body":"<p>I am running BERT SQuAD prediction tasks on a 32-core host with GPUs, using tensorflow-gpu 1.14.x. Some parts of the BERT processing occurs only on cpu while the core prediction part runs on a GPU. Those parts of the processing that occur on cpu are bound to a single cpu, making them a significant bottleneck. When running on non-gpu tensorflow, I notice that all available cores are in use. As can be expected, the parts that previously ran on one core are much faster due to the concurrency. But the computation that benefits from GPU usage are understandably much slower.<\/p>\n\n<p>Is there a way to specify that the non-gpu parts of the processing make use of more than one cpu?<\/p>\n","answers":[]}
+{"QuestionId":57595459,"AnswerCount":0,"Tags":"<python><tensorflow><keras><conv-neural-network><one-hot-encoding>","CreationDate":"2019-08-21T16:01:32.613","AcceptedAnswerId":null,"OwnerUserId":9715052.0,"Title":"to_categorical - ValueError: Input 0 of layer conv2d_2 is incompatible with the layer","Body":"<p>I'm trying to pass <code>reshaped_one_hot<\/code> as the input to the <code>bottleneck<\/code> and I'm getting the following error: <\/p>\n\n<pre><code>ValueError: Input 0 of layer conv2d_2 is incompatible with the layer: expected ndim=4, found ndim=5. Full shape received: [1000, 16, 16, 512, 1]\n<\/code><\/pre>\n\n<p>I understand that I need to reshape this input and make sure that it has 4 dimensions instead of 5 dimensions but I'm having some trouble understanding how exactly to go about doing that. My <code>input_shape<\/code> is <code>[1000, 16, 16, 1]<\/code><\/p>\n\n<p><strong>My Attempt:<\/strong> \nI've tried to apply <code>numpy.squeeze(one_hot)<\/code> which then makes the shape of <code>one_hot<\/code> <code>[1000, 16, 16, 512]<\/code> and I get the following error: <code>TypeError: unhashable type: 'numpy.ndarray'<\/code><\/p>\n\n<p>The intuition of why exactly I'm applying one hot encoding to the output of my encoder may not make sense because I'm just testing out a theory but my intent of doing this is the following:<\/p>\n\n<p>I want to take the output of the encoder, perform one hot encoding where it finds the categories using unique features, similar to how <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.OneHotEncoder.html\" rel=\"nofollow noreferrer\">sklearn<\/a> describes their one hot encoding and then pass this to the bottleneck layer. <\/p>\n\n<p>My code: <\/p>\n\n<pre><code>encoder_img = tf.keras.layers.Input(shape=(16,16,1), name=\"input\")\nx = tf.keras.layers.Conv2D(1024, 1, activation='relu',kernel_initializer=keras.initializers.RandomUniform)(encoder_img)\nx = tf.keras.layers.MaxPooling2D(1)(x)\ninputtothelayer = tf.keras.layers.Conv2D(512, 1, activation='relu')(x)\npool = tf.keras.layers.MaxPooling2D(1, name=\"thelayer\")(inputtothelayer)\nencoder = tf.keras.Model(inputs=encoder_img, outputs=pool, name = 'encoder')\nencoder.summary()\n\n# one hot encoding layer\nlayer_name = 'thelayer'\nintermediate_layer_model = tf.keras.Model(inputs=encoder_img,\n                                 outputs=encoder.get_layer(layer_name).output)\nintermediate_output = intermediate_layer_model.predict(data)\nprint(intermediate_output)\n\none_hot = tf.keras.utils.to_categorical(intermediate_output,num_classes=None, dtype='float32')\nprint(one_hot)\nreshaped_one_hot = one_hot\n#reshaped_one_hot = np.squeeze(one_hot)\nprint(reshaped_one_hot.shape)\n# decoder architecture\nbottleneck = tf.keras.layers.Conv2D(256, 3, activation='relu')(reshaped_one_hot)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57595780,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><conv-neural-network>","CreationDate":"2019-08-21T16:23:17.420","AcceptedAnswerId":57597795.0,"OwnerUserId":8243797.0,"Title":"logits and labels must have the same first dimension, got logits shape [3662,5] and labels shape [18310]","Body":"<p>I am trying to create a CNN with tensorflow, my images are 64x64x1 images and I have an array of 3662 images which I am using for training. I have total 5 labels which I have one-hot encoded. I am getting this error everytime:<\/p>\n\n<pre><code>InvalidArgumentError: logits and labels must have the same first dimension, got logits shape [3662,5] and labels shape [18310]\n     [[{{node loss_2\/dense_5_loss\/SparseSoftmaxCrossEntropyWithLogits\/SparseSoftmaxCrossEntropyWithLogits}}]]\n<\/code><\/pre>\n\n<p>my neural network structure is this:<\/p>\n\n<pre><code>def cnn_model():\n    model = models.Sequential()\n#     model.add(layers.Dense(128, activation='relu', ))\n    model.add(layers.Conv2D(128, (3, 3), activation='relu',input_shape=(64, 64, 1)))\n    model.add(layers.MaxPooling2D((2, 2)))\n    model.add(layers.Conv2D(64, (3, 3), activation='relu',padding = 'same'))\n    model.add(layers.MaxPooling2D((2, 2)))\n    model.add(layers.Conv2D(64, (3, 3), activation='relu', padding='same'))\n    model.add(layers.Dense(64, activation='relu'))\n    model.add(layers.Flatten())\n    model.add(layers.Dense(5, activation='softmax'))\n\n    model.compile(optimizer='adam',\n              loss=tf.keras.losses.SparseCategoricalCrossentropy(),\n              metrics=['accuracy'])\n    print(model.summary())\n    return model\n<\/code><\/pre>\n\n<p>My model summary is this:<\/p>\n\n<pre><code>Model: \"sequential_3\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nconv2d_9 (Conv2D)            (None, 62, 62, 128)       1280      \n_________________________________________________________________\nmax_pooling2d_6 (MaxPooling2 (None, 31, 31, 128)       0         \n_________________________________________________________________\nconv2d_10 (Conv2D)           (None, 31, 31, 64)        73792     \n_________________________________________________________________\nmax_pooling2d_7 (MaxPooling2 (None, 15, 15, 64)        0         \n_________________________________________________________________\nconv2d_11 (Conv2D)           (None, 15, 15, 64)        36928     \n_________________________________________________________________\ndense_4 (Dense)              (None, 15, 15, 64)        4160      \n_________________________________________________________________\nflatten_2 (Flatten)          (None, 14400)             0         \n_________________________________________________________________\ndense_5 (Dense)              (None, 5)                 72005     \n=================================================================\nTotal params: 188,165\nTrainable params: 188,165\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>my output array is of the shape (3662,5,1). I have seen other answers to same questions but I can't figure out the problem with mine. Where am I wrong?<\/p>\n\n<p>Edit: My labels are stored in one hot encoded form using these:<\/p>\n\n<pre><code>df = pd.get_dummies(df)\ndiag = np.array(df)\ndiag = np.reshape(diag,(3662,5,1))\n<\/code><\/pre>\n\n<p>I have tried as numpy array and after converting to tensor(same for input as per documentation)<\/p>\n","answers":[{"AnswerId":"57597795","CreationDate":"2019-08-21T19:02:30.003","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>The problem lines within the choice of the loss function <code>tf.keras.losses.SparseCategoricalCrossentropy()<\/code>. According to what you are trying to achieve you should use <code>tf.keras.losses.CategoricalCrossentropy()<\/code>. Namely, the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/losses\/SparseCategoricalCrossentropy\" rel=\"nofollow noreferrer\">documentation<\/a> of <code>tf.keras.losses.SparseCategoricalCrossentropy()<\/code> states:<\/p>\n\n<blockquote>\n  <p>Use this crossentropy loss function when there are two or more label classes. We expect labels to be provided as integers.<\/p>\n<\/blockquote>\n\n<p>On the other hand, the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/losses\/CategoricalCrossentropy\" rel=\"nofollow noreferrer\">documentation<\/a> of <code>tf.keras.losses.CategoricalCrossentropy()<\/code> states:<\/p>\n\n<blockquote>\n  <p>We expect labels to be provided in a one_hot representation. <\/p>\n<\/blockquote>\n\n<p>And because your labels are encoded as one-hot, you should use <code>tf.keras.losses.CategoricalCrossentropy()<\/code>.<\/p>\n"}]}
+{"QuestionId":57595819,"AnswerCount":0,"Tags":"<python><deep-learning><pytorch><data-augmentation>","CreationDate":"2019-08-21T16:26:00.373","AcceptedAnswerId":null,"OwnerUserId":11940717.0,"Title":"Balancing an Unbalanced Dataset with K-Fold Cross Validation","Body":"<p>I'm trying to train\/validate a CNN using Pytorch on an unbalanced image dataset (class 1:250 images, class 0: 4000ish images), and right now, I've tried augmentation solely on my training set (thanks @jodag). However, my model is still learning to favor the class with significantly more images. <\/p>\n\n<p>I want to find ways to compensate for my unbalanced data set. <\/p>\n\n<p>I thought about using oversampling\/undersampling using the imbalanced data sampler (<a href=\"https:\/\/github.com\/ufoym\/imbalanced-dataset-sampler\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ufoym\/imbalanced-dataset-sampler<\/a>), but I already use a sampler to select indices for my 5-fold validation. Is there a way I could implement cross-validation using the code below and also add this sampler? Similarly, is there a way to augment one label more frequently than the other? Along the lines of these questions, are there any alternative easier ways that I could address my unbalanced dataset that I haven't looked into yet?<\/p>\n\n<h1>Here's an example of what I have so far<\/h1>\n\n<pre><code>total_set = datasets.ImageFolder(PATH)\nKF_splits = KFold(n_splits= 5, shuffle = True, random_state = 42)\n\nfor train_idx, valid_idx in KF_splits.split(total_set):\n    #sampler to get indices for cross validation\n    train_sampler = SubsetRandomSampler(train_idx)\n    valid_sampler = SubsetRandomSampler(valid_idx)\n\n    #Use a wrapper to apply augmentation only to training set\n    #These are dataloaders that pull images from the same folder but sort into validation and training sets\n    #Though transforms augment only the training set, it doesn't address\n    #the underlying issue of a heavily unbalanced dataset\n\n    train_loader = torch.utils.data.DataLoader(\n        WrapperDataset(total_set, transform=data_transforms['train']),\n        batch_size=32, sampler=ImbalancedDatasetSampler(total_set))\n    valid_loader = torch.utils.data.DataLoader(\n        WrapperDataset(total_set, transform=data_transforms['val']),\n        batch_size=32)\n\n    print(\"Fold:\" + str(i))\n\n    for epoch in range(epochs):\n        #Train\/validate model below\n<\/code><\/pre>\n\n<p>`<\/p>\n\n<p>Thank you for your time and help!<\/p>\n","answers":[]}
+{"QuestionId":57596098,"AnswerCount":0,"Tags":"<tensorflow><keras><gradient><regularized>","CreationDate":"2019-08-21T16:49:00.067","AcceptedAnswerId":null,"OwnerUserId":11958089.0,"Title":"Problem with defining a custom regularizer using keras\/tensorflow","Body":"<p>I want to use a lookup table as a function to write a regularizer. Let's say weights are quantized using 8-bit. Therefore, we have 256 possible candidates for each weight. The lookup table represents a cost for each possible value. therefore, the lookup table is a vector with 256 elements. I wrote a regularizer using a template provided by Keras. My questions are:<\/p>\n\n<p>1) Did I write correctly?<\/p>\n\n<p>2) How does Tensorflow\/Keras handle the gradient? Do I need to do anything else to handle the gradient? (Because I am using a lookup table-based function).<\/p>\n\n<p>Here is part of my code: (Since the vector (Grad) was too long, I removed most of its element here)<\/p>\n\n<pre><code>def __call__(self, x):\n    regularization = 0.\n    Grad = K.constant([7.96875,3.734375,-0.49609375,-0.5,...,-1.4921875],dtype='float32');\n    Graddata = lambda Grad: Grad\n\n    if self.Er:\n        regularization += self.Er * K.sum(K.square(Graddata(tf.round(K.abs(x)*(2 ** 8)))))\n    return regularization\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57596162,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-datasets><tensorflow-estimator><tf.keras>","CreationDate":"2019-08-21T16:55:32.183","AcceptedAnswerId":null,"OwnerUserId":11244995.0,"Title":"Tensorflow\/Keras local training 6x faster than distributed training","Body":"<p>I am training a Classifier (via RNN) in Tensorflow with tf.data api, tf.keras api and tf. estimator api. In local training i got a troughput of around 240 steps\/sec, than i enter distributed mode with ParameterServer (due to proper TF_CONFIG) with 1 Worker (so 1 PS and 1 Chief = 1 Worker) i got around 40 steps\/sec. If i add 1 Worker (now 2 Worker) i got around 50 steps\/sec and so on. I tested this behavior at home (with multiple pc's) and in cluster (around 18 worker). Only with 18 worker i reached the speed i got at local training. I testet training with nfs,hdfs, changed os. Network cant be the problem because an other framework scaled pretty good.\nCan someone please guide me on how to get better perfomance. I tried alot but didnt find a suitable solution.<\/p>\n\n<h2>My Code<\/h2>\n\n<pre class=\"lang-py prettyprint-override\"><code>data_pipeline = tf.data.experimental.CsvDataset(MODEL_DIR, DEFAULTS, header=False,field_delim=CSV_SEPARATOR) \\\n                        .shard(WORKER, WORKERINDEX) \\\n                        .map(lambda *a: tf.stack(a), num_parallel_calls=THREADS) \\\n                        .window(size=WINDOW_SIZE, shift=WINDOW_SHIFT, stride=WINDOW_STRIDE,drop_remainder=True) \\\n                        .flat_map(lambda b: b.batch(WINDOW_SIZE)) \\\n                        .map(lambda c: (c[:,:-1],tf.one_hot(tf.cast(c[-1][-1], tf.int32),2))) \\\n                        .batch(GLOBAL_BATCH_SIZE, drop_remainder=True) \\\n                        .repeat(REAPEAT_DATASET) \\\n                        .prefetch(PREFETCH)\n<\/code><\/pre>\n\n<pre><code>model = tf.keras.Sequential()\nmodel.add(tf.keras.layers.Dense(5, batch_input_shape=(GLOBAL_BATCH_SIZE, WINDOW_SIZE, NR_FEATURES),activation=tf.keras.activations.relu, kernel_initializer=tf.keras.initializers.he_uniform(seed=2), bias_initializer=tf.keras.initializers.Zeros()))\nmodel.add(tf.keras.layers.SimpleRNN(RNN_UNITS,stateful=True,activation=tf.keras.activations.tanh, kernel_initializer=tf.keras.initializers.glorot_uniform(seed=2),recurrent_initializer=tf.keras.initializers.glorot_uniform(seed=2),bias_initializer=tf.keras.initializers.Zeros()))\nmodel.add(tf.keras.layers.Dense(NUM_CLASSES, activation= tf.keras.activations.softmax, kernel_initializer=tf.keras.initializers.glorot_uniform(seed=2), bias_initializer=tf.keras.initializers.Zeros()))\nmodel.compile(optimizer=tf.train.AdamOptimizer(LEARNING_RATE), # AdamOptimizer\n              loss=tf.keras.losses.CategoricalCrossentropy(),\n              metrics=[tf.keras.metrics.CategoricalAccuracy(dtype=tf.float32),tf.keras.metrics.Recall(dtype=tf.float32),tf.keras.metrics.Precision(dtype=tf.float32,class_id=0),f1_score])\nestimator = tf.keras.estimator.model_to_estimator(keras_model=model, config=RUN_CONFIG, model_dir=LOG_DIR)\n\n<\/code><\/pre>\n\n<pre><code>train_spec = tf.estimator.TrainSpec(input_fn=input_fn(tf.estimator.ModeKeys.TRAIN))\neval_spec = tf.estimator.EvalSpec(input_fn=input_fn(tf.estimator.ModeKeys.EVAL), steps=100,start_delay_secs=20,throttle_secs=1*60\/6)\ntf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)\n<\/code><\/pre>\n\n<pre><code>cluster_spec_dict = {'ps': ['192.168.2.XXX:50012'],'chief': [\"192.168.2.XXX:50014\"],}\n<\/code><\/pre>\n\n<ul>\n<li>Tensorflow version: 1.14<\/li>\n<li>dense \/ rnn units = 5<\/li>\n<li>batch_size=16<\/li>\n<li>window_size\/sequence_length=16<\/li>\n<\/ul>\n","answers":[]}
+{"QuestionId":57596804,"AnswerCount":0,"Tags":"<python-3.x><keras><deep-learning>","CreationDate":"2019-08-21T17:47:17.710","AcceptedAnswerId":null,"OwnerUserId":11882019.0,"Title":"Keras multi gpu model usage","Body":"<p>In Keras, how can I check whether the model is multi gpu or single gpu version? I am loading a multi gpu version and giving it 2 gpus to run but after loading the model I am setting keras backend to use one gpu like so.<\/p>\n\n<pre><code>config = tf.ConfigProto( device_count = {'GPU': 1} ) \nsess = tf.Session(config=config) \nkeras.backend.set_session(sess)\n<\/code><\/pre>\n\n<p>After this, I would like to verify whether the model is using one or two gpus.<\/p>\n","answers":[]}
+{"QuestionId":57596929,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-21T17:56:12.020","AcceptedAnswerId":null,"OwnerUserId":2991243.0,"Title":"Extract fitted output from the trained model in Python (KERAS\/TensorFlow)","Body":"<p>I'm using KERAS with TensorFlow back-end. Suppose that this is the model block:<\/p>\n\n<pre><code>model.add(LSTM(units = 60, activation = 'tanh')\nmodel.add(Dropout(rate = 0.5))\nmodel.add(Dense(units = 1, activation = 'sigmoid'))\nmodel.compile(optimizer = 'adam', loss = 'mse')\nmodel.fit(X_train, Y_train, epochs = 200, batch_size = 32)\n<\/code><\/pre>\n\n<p>Is there any way to extract fitted output from the trained model (<code>model<\/code>)?<\/p>\n","answers":[{"AnswerId":"57604433","CreationDate":"2019-08-22T07:47:15.463","ParentId":null,"OwnerUserId":"10455493","Title":null,"Body":"<p>You should probably rephrase you question, as it is not clear what fitted output you try to achieve. I guess the most probable scenarios with respect to machine learning is that: <\/p>\n\n<ol>\n<li>You have new input data <code>X<\/code>, and want to do a prediction with your fitted model. This can be done like this <code>model.predict(X)<\/code>.<\/li>\n<li>Or you are trying to get the weights of the fitted model, which can be done like this <code>model.get_weights()<\/code>.<\/li>\n<\/ol>\n"}]}
+{"QuestionId":57596943,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-08-21T17:56:49.747","AcceptedAnswerId":57598447.0,"OwnerUserId":3295741.0,"Title":"Segmentation violation (SIGSEGV) when loading a Keras\/Tensorflow model thats uses a Lambda layer","Body":"<p>I create a model like this and save the model:<\/p>\n\n<pre><code>def extract_and_duplicate(tensor, reps=1, batch_size=0, sample_size=0):\n    tensor = K.reshape(tensor[:,:,0],(batch_size, sample_size, 1))\n    if reps &gt; 1:\n        tensor = Concatenate()([tensor for i in range(reps)])\n    return tensor\n\ninput = Input(batch_shape = (batch_size, sample_size, num_features))\nout = &lt;steps to create a NN with several layers&gt;\n\npre_mask = Lambda(extract_and_duplicate, arguments = {'reps': some_number,'batch_size': batch_size, 'sample_size': sample_size})(input)\nmask = TimeDistributed(Dense(m, activation = 'tanh'))(out)\nout = Multiply()([pre_mask,mask])\nmodel = Model(input, out)\n<\/code><\/pre>\n\n<p>When I load the model with the following line, I get a SIGSEGV signal.<\/p>\n\n<pre><code>load_model(model_path, custom_objects={'estimated_accuracy': estimated_accuracy, 'extract_and_duplicate': extract_and_duplicate})\n<\/code><\/pre>\n\n<p>I single-stepped the load_model method and found that it fails with SIGSEGV when loading the Lambda layer.\nWhen I remove the Lambda layer, loading the model works.\nAm I doing something wrong or am I stepping on a keras\/tensorflow bug?<\/p>\n\n<p>Can you provide a solution or steps to investigate?\nThanks<\/p>\n","answers":[{"AnswerId":"57598447","CreationDate":"2019-08-21T19:58:05.477","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>You need to pass everything necessary to create the lambda layer to <code>custom_objects<\/code>. Otherwise Keras will not know what those variable names mean.<\/p>\n\n<p>So, add <code>some_number<\/code>, <code>batch_size<\/code>, <code>sample_size<\/code>, <code>reps<\/code> and everything else necessary to build the lambda layer entirely into <code>custom_objects<\/code>.<\/p>\n"}]}
+{"QuestionId":57597002,"AnswerCount":1,"Tags":"<python><conv-neural-network><pytorch>","CreationDate":"2019-08-21T18:01:18.167","AcceptedAnswerId":null,"OwnerUserId":8645024.0,"Title":"Loss Function Not Decreasing in CNN","Body":"<p>I am new to Pytorch and I'm training a model for binary classification of images. The images are currently stored as .npy files and I am loading them and training my model in batches. When I implement this, the loss function does not decrease. When I test the model on the training and test set again, the accuracy is constant at 50%. The data set is balanced. <\/p>\n\n<p>I tried making the dataset smaller (around 125 for each class) and I still have the same problem. I expect the model to overfit the training set but this does not occur. <\/p>\n\n<p>Please see my code below <\/p>\n\n<pre><code>class Network(nn.Module):\n    def __init__(self):\n        super(Network,self).__init__()\n        self.conv1=nn.Conv2d(in_channels=2, out_channels=32, kernel_size=3)\n        self.conv2=nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3)\n        self.conv3=nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3)\n        self.fc1=nn.Linear(in_features=128*6*6, out_features=1000)\n        self.fc2=nn.Linear(in_features=1000, out_features=100)\n        self.out=nn.Linear(in_features=100, out_features=2)\n    def forward(self,t):\n        POOL_stride=2\n        #Conv1\n        t=F.relu(self.conv1(t))\n        t=F.max_pool2d(t, kernel_size=2, stride=POOL_stride)\n        #Conv2\n        t=F.relu(self.conv2(t))\n        t=F.max_pool2d(t, kernel_size=2, stride=POOL_stride)\n        #Conv3\n        t=F.relu(self.conv3(t))\n        t=F.max_pool2d(t, kernel_size=2, stride=POOL_stride)\n        # dense 1\n        t=t.reshape(-1, 128*6*6)\n        t=self.fc1(t)\n        t=F.relu(t)\n        #dense 2\n        t=self.fc2(t)\n        t=F.relu(t)\n        t=self.out(t)\n        return t\n\ndef npy_loader(path):\n    sample = torch.from_numpy(np.load(path))\n    return sample\n\ncriterion=nn.CrossEntropyLoss()\noptimizer = optim.Adam(self.model.parameters(), lr=0.003)\nmodel = Network()\ntrainset = datasets.DatasetFolder(\n            root=train_dir,\n            loader=npy_loader,\n            extensions=['.npy']\n        )\ntrain_loader = torch.utils.data.DataLoader(\n            trainset,\n            batch_size=batch_size,\n            shuffle=True,\n        )\nfor epoch in range(epochs):\n            running_loss = 0\n            batches = 0\n            for inputs, labels in train_loader:\n                batches = batches+1\n                inputs, labels = inputs.to(device), labels.to(device)\n                optimizer.zero_grad()\n                output = model(inputs)\n                loss = criterion(output.squeeze(), labels.squeeze())\n                loss.backward()\n                optimizer.step()\n                running_loss += loss.item()    \n            print('Loss :{:.4f} Epoch[{}\/{}]'.format(running_loss\/batches, epoch, epochs))\n'''\n<\/code><\/pre>\n","answers":[{"AnswerId":"57601405","CreationDate":"2019-08-22T02:45:12.507","ParentId":null,"OwnerUserId":"5027315","Title":null,"Body":"<p>You are providing parameters of some other <code>self.model<\/code> to optimizer while the model used for calculating the loss is different.<\/p>\n\n<pre><code>optimizer = optim.Adam(self.model.parameters(), lr=0.003)\nmodel = Network()\n<\/code><\/pre>\n\n<p>Above is your sequence of defining <code>optimizer<\/code> and <code>model<\/code>. Notice that you are passing parameters of a different <code>self.model<\/code> to <code>optimizer<\/code>. Hence, <code>optimizer.step()<\/code> fails to update weights of desired <code>model<\/code> on which loss is being calculated. Instead it should be something like this:<\/p>\n\n<pre><code>model = Network()    \noptimizer = optim.Adam(model.parameters(), lr=0.003)\n<\/code><\/pre>\n\n<p>On another note, might I suggest that instead of returning 2 dimensional output from model, returning a 1-d output and using binary cross-entropy loss can also be explored, as your task is only a binary classification problem. <\/p>\n"}]}
+{"QuestionId":57597008,"AnswerCount":0,"Tags":"<python><tensorflow><keras><tensorflow-estimator><tf.keras>","CreationDate":"2019-08-21T18:01:40.753","AcceptedAnswerId":null,"OwnerUserId":3688728.0,"Title":"TF 2.0 Keras Label Vocabulary","Body":"<p>Trying to convert the output\/predictions\/logits of a model to a string (specified in a label_vocabulary) at training and serving.<\/p>\n\n<p>Using Estimators, this can specified in DNNClassifier()<\/p>\n\n<pre><code>model = tf.estimator.DNNClassifier(..., label_vocabulary=vocab, n_classes=len(vocab))\n<\/code><\/pre>\n\n<p>However, I'm having trouble finding the Keras equivalent of such. Our label data is all in the form of strings, not ID's, and would prefer to keep it that way if at all possible.<\/p>\n\n<p>Considering writing a custom model or maybe overriding Keras methods in order to get the desired functionality, but hoping there is an easier way. Any help would be greatly appreciated. Thank you.<\/p>\n","answers":[]}
+{"QuestionId":57597145,"AnswerCount":1,"Tags":"<tensorflow><keras><tensorflow2.0>","CreationDate":"2019-08-21T18:12:42.530","AcceptedAnswerId":57597286.0,"OwnerUserId":11958602.0,"Title":"CNN compiling error : Failed to get convolution algorithm. This is probably because cuDNN failed to initialize,","Body":"<p>I just try to build up VGG-19 Model by myself, but every time I try to compile, the error rise, though it works clearly on Alexnet. please let me know the problem.  <\/p>\n\n<h2>Setting<\/h2>\n\n<p>CPU : AMD RYZEN 2700x<br>\nGPU : RTX 2080<br>\nOS : Window 10<br>\nCUDA : 10.0<br>\nTensorflow version : Tensorflow-2.0.0beta1<br>\nWorking tools : Jupyter lab<br>\nTime : 3 AM<br>\nClimate : rainy<br>\nFeel : gloomy :(  <\/p>\n\n<h2>Code<\/h2>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf  \nfrom tensorflow import keras  \nfrom tensorflow.keras import layers  \nimport tensorflow_datasets as tfds  \nimport matplotlib.pyplot as plt  \nfrom functools import partial  \nimport skimage.io as io  \nimport numpy as np  \n\nEPOCH = 5  \nBATCH = 32  \nLearning_rate = 0.0001  \n\nleaky_relu = partial(keras.activations.relu, alpha=0.1)  \n\ndef to_float(image, label) :  \n    float_image = image\/255  \n    return (float_image, label)  \n\ndata, info = tfds.load('cifar10',as_supervised=True,with_info=True)  \n\ntrain, test = data.get('train'), data.get('test')  \ntrain_data = train.map(to_float).shuffle(10000).batch(BATCH).prefetch(3)  \ntest_data = test.map(to_float).batch(BATCH).prefetch(4)  \n\nsample_batch = next(iter(train_data))[0]  \nprint(io.imshow_collection(io.concatenate_images(np.array(sample_batch))))  \n\n\nclass VGG19(tf.keras.Model) :  \n    def __init__(self) :  \n        super(VGG19, self).__init__()  \n        self.conv1_1 = layers.Conv2D(filters = 64,  \n                                      kernel_size=(3,3),  \n                                      strides=(1,1),  \n                                      padding='same',  \n                                      activation=leaky_relu)  \n        self.conv1_2 = layers.Conv2D(filters = 64,  \n                                      kernel_size=(3,3),  \n                                      strides=(1,1),  \n                                      padding='same',  \n                                      activation=leaky_relu)  \n        self.pool1 = layers.MaxPool2D(pool_size=(2,2),  \n                                      strides=(2,2),  \n                                      padding='same')  \n\n\n        self.conv2_1 = layers.Conv2D(filters = 128,  \n                                      kernel_size=(3,3),  \n                                      strides=(1,1),  \n                                      padding='same',  \n                                      activation=leaky_relu)  \n        self.conv2_2 = layers.Conv2D(filters = 128,  \n                                      kernel_size=(3,3),  \n                                      strides=(1,1),  \n                                      padding='same',  \n                                      activation=leaky_relu)  \n        self.pool2 = layers.MaxPool2D(pool_size=(2,2),  \n                                      strides=(2,2),  \n                                      padding='same')  \n\n\n        self.conv3_1 = layers.Conv2D(filters = 256,  \n                                      kernel_size=(3,3),  \n                                      strides=(1,1),  \n                                      padding='same',  \n                                      activation=leaky_relu)  \n        self.conv3_2 = layers.Conv2D(filters = 256,  \n                                      kernel_size=(3,3),  \n                                      strides=(1,1),  \n                                      padding='same',  \n                                      activation=leaky_relu)  \n        self.conv3_3 = layers.Conv2D(filters = 256,  \n                                      kernel_size=(3,3),  \n                                      strides=(1,1),  \n                                      padding='same',  \n                                      activation=leaky_relu)  \n        self.conv3_4 = layers.Conv2D(filters = 256,  \n                                      kernel_size=(3,3),  \n                                      strides=(1,1),  \n                                      padding='same',  \n                                      activation=leaky_relu)  \n        self.pool3 = layers.MaxPool2D(pool_size=(2,2),  \n                                      strides=(2,2),  \n                                      padding='same')  \n\n\n        self.conv4_1 = layers.Conv2D(filters = 512,\n                                      kernel_size=(3,3),\n                                      strides=(1,1),\n                                      padding='same',\n                                      activation=leaky_relu)\n        self.conv4_2 = layers.Conv2D(filters = 512,\n                                      kernel_size=(3,3),\n                                      strides=(1,1),\n                                      padding='same',\n                                      activation=leaky_relu)\n        self.conv4_3 = layers.Conv2D(filters = 512,\n                                      kernel_size=(3,3),\n                                      strides=(1,1),\n                                      padding='same',\n                                      activation=leaky_relu)\n        self.conv4_4 = layers.Conv2D(filters = 512,\n                                      kernel_size=(3,3),\n                                      strides=(1,1),\n                                      padding='same',\n                                      activation=leaky_relu)\n        self.pool4 = layers.MaxPool2D(pool_size=(2,2),\n                                      strides=(2,2),\n                                      padding='same')\n\n\n        self.conv5_1 = layers.Conv2D(filters = 512,\n                                      kernel_size=(3,3),\n                                      strides=(1,1),\n                                      padding='same',\n                                      activation=leaky_relu)\n        self.conv5_2 = layers.Conv2D(filters = 512,\n                                      kernel_size=(3,3),\n                                      strides=(1,1),\n                                      padding='same',\n                                      activation=leaky_relu)\n        self.conv5_3 = layers.Conv2D(filters = 512,\n                                      kernel_size=(3,3),\n                                      strides=(1,1),\n                                      padding='same',\n                                      activation=leaky_relu)\n        self.conv5_4 = layers.Conv2D(filters = 512,\n                                      kernel_size=(3,3),\n                                      strides=(1,1),\n                                      padding='same',\n                                      activation=leaky_relu)\n        self.pool5 = layers.MaxPool2D(pool_size=(2,2),\n                                      strides=(2,2),\n                                      padding='same')\n\n\n        self.flat6 = layers.Flatten()\n        self.fc6 = layers.Dense(units=4096,\n                                activation=leaky_relu)\n        self.drop6 = layers.Dropout(rate=0.5)\n\n\n        self.fc7 = layers.Dense(units=4096,\n                                activation=leaky_relu)\n        self.drop7 = layers.Dropout(rate=0.5)\n\n\n        self.fc8 = layers.Dense(units=10,\n                                activation=tf.keras.activations.softmax)\n\n    def call(self, inputs):\n            x = self.conv1_1(inputs)\n            x = self.conv1_2(x)\n            x = self.pool1(x)\n\n            x = self.conv2_1(x)\n            x = self.conv2_2(x)\n            x = self.pool2(x)\n\n            x = self.conv3_1(x)\n            x = self.conv3_2(x)\n            x = self.conv3_3(x)\n            x = self.conv3_4(x)\n            x = self.pool3(x)\n\n            x = self.conv4_1(x)\n            x = self.conv4_2(x)\n            x = self.conv4_3(x)\n            x = self.conv4_4(x)\n            x = self.pool4(x)\n\n            x = self.conv5_1(x)\n            x = self.conv5_2(x)\n            x = self.conv5_3(x)\n            x = self.conv5_4(x)\n            x = self.pool5(x)\n\n            x = self.flat6(x)\n            x = self.fc6(x)\n            x = self.drop6(x)\n\n            x = self.fc7(x)\n            x = self.drop7(x)\n\n            output = self.fc8(x)\n\n            return output\n\n\nVGG19_model = VGG19()\n\nmy_loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)\n\nVGG19_model.compile(optimizer=tf.keras.optimizers.Adam(Learning_rate),\n                loss = my_loss,\n                metrics = ['accuracy'],\n               )\n\nVGG19_model.fit(train_data, epochs = EPOCH, verbose=2)\n<\/code><\/pre>\n\n<h2>Error Message<\/h2>\n\n<pre class=\"lang-py prettyprint-override\"><code>---------------------------------------------------------------------------\nUnknownError                              Traceback (most recent call last)\n&lt;ipython-input-21-1e291c11711c&gt; in &lt;module&gt;\n      4                )\n      5 \n----&gt; 6 VGG19_model.fit(train_data, epochs = EPOCH, verbose=2)\n\n~\\Anaconda3\\envs\\cuda\\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, validation_freq, max_queue_size, workers, use_multiprocessing, **kwargs)\n    641         max_queue_size=max_queue_size,\n    642         workers=workers,\n--&gt; 643         use_multiprocessing=use_multiprocessing)\n    644 \n    645   def evaluate(self,\n\n~\\Anaconda3\\envs\\cuda\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_generator.py in fit(self, model, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_freq, **kwargs)\n    692         shuffle=shuffle,\n    693         initial_epoch=initial_epoch,\n--&gt; 694         steps_name='steps_per_epoch')\n    695 \n    696   def evaluate(self,\n\n~\\Anaconda3\\envs\\cuda\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_generator.py in model_iteration(model, data, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, validation_freq, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch, mode, batch_size, steps_name, **kwargs)\n    262 \n    263       is_deferred = not model._is_compiled\n--&gt; 264       batch_outs = batch_function(*batch_data)\n    265       if not isinstance(batch_outs, list):\n    266         batch_outs = [batch_outs]\n\n~\\Anaconda3\\envs\\cuda\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py in train_on_batch(self, x, y, sample_weight, class_weight, reset_metrics)\n    916       self._update_sample_weight_modes(sample_weights=sample_weights)\n    917       self._make_train_function()\n--&gt; 918       outputs = self.train_function(ins)  # pylint: disable=not-callable\n    919 \n    920     if reset_metrics:\n\n~\\Anaconda3\\envs\\cuda\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py in __call__(self, inputs)\n   3508         value = math_ops.cast(value, tensor.dtype)\n   3509       converted_inputs.append(value)\n-&gt; 3510     outputs = self._graph_fn(*converted_inputs)\n   3511 \n   3512     # EagerTensor.numpy() will often make a copy to ensure memory safety.\n\n~\\Anaconda3\\envs\\cuda\\lib\\site-packages\\tensorflow\\python\\eager\\function.py in __call__(self, *args, **kwargs)\n    570       raise TypeError(\"Keyword arguments {} unknown. Expected {}.\".format(\n    571           list(kwargs.keys()), list(self._arg_keywords)))\n--&gt; 572     return self._call_flat(args)\n    573 \n    574   def _filtered_call(self, args, kwargs):\n\n~\\Anaconda3\\envs\\cuda\\lib\\site-packages\\tensorflow\\python\\eager\\function.py in _call_flat(self, args)\n    669     # Only need to override the gradient in graph mode and when we have outputs.\n    670     if context.executing_eagerly() or not self.outputs:\n--&gt; 671       outputs = self._inference_function.call(ctx, args)\n    672     else:\n    673       self._register_gradient()\n\n~\\Anaconda3\\envs\\cuda\\lib\\site-packages\\tensorflow\\python\\eager\\function.py in call(self, ctx, args)\n    443             attrs=(\"executor_type\", executor_type,\n    444                    \"config_proto\", config),\n--&gt; 445             ctx=ctx)\n    446       # Replace empty list with None\n    447       outputs = outputs or None\n\n~\\Anaconda3\\envs\\cuda\\lib\\site-packages\\tensorflow\\python\\eager\\execute.py in quick_execute(op_name, num_outputs, inputs, attrs, ctx, name)\n     65     else:\n     66       message = e.message\n---&gt; 67     six.raise_from(core._status_to_exception(e.code, message), None)\n     68   except TypeError as e:\n     69     if any(ops._is_keras_symbolic_tensor(x) for x in inputs):\n\n~\\Anaconda3\\envs\\cuda\\lib\\site-packages\\six.py in raise_from(value, from_value)\n<\/code><\/pre>\n\n<h2>UnknownError: 2 root error(s) found.<\/h2>\n\n<p><strong>(0) Unknown:  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.<\/strong><br>\n     [[node vg_g19_2\/conv2d_32\/Conv2D (defined at :6) ]]<br>\n    [[Func\/Adam\/gradients_2\/vg_g19_2\/dropout_5\/cond_grad\/If\/then\/_22\/input\/_75\/_64]]  <\/p>\n\n<p><strong>(1) Unknown:  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.<\/strong><br>\n     [[node vg_g19_2\/conv2d_32\/Conv2D (defined at :6) ]]<br>\n0 successful operations.<br>\n0 derived errors ignored. [Op:__inference_keras_scratch_graph_10115]  <\/p>\n\n<p>Function call stack:<br>\nkeras_scratch_graph -> keras_scratch_graph<\/p>\n","answers":[{"AnswerId":"57597286","CreationDate":"2019-08-21T18:22:49.017","ParentId":null,"OwnerUserId":"11127923","Title":null,"Body":"<p>First thing will be to check if compatible CUDA, cuDNN drivers are installed correctly. Then you may try gpu memory resources management by allowing gpu memory growth.  <\/p>\n\n<p><strong>allow_growth<\/strong> option, attempts to allocate only as much GPU memory based on runtime allocations: it starts out allocating very little memory, and as Sessions get run and more GPU memory is needed, it extend the GPU memory region needed by the TensorFlow process.<\/p>\n\n<p>To know more see <a href=\"https:\/\/www.tensorflow.org\/guide\/using_gpu#allowing_gpu_memory_growth\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/using_gpu#allowing_gpu_memory_growth<\/a><br>\nYou can try Allowing GPU memory growth with:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>config = tf.ConfigProto()\nconfig.gpu_options.allow_growth = True\nsess = tf.Session(config=config)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57597897,"AnswerCount":0,"Tags":"<python><machine-learning><keras><pytorch>","CreationDate":"2019-08-21T19:11:21.287","AcceptedAnswerId":null,"OwnerUserId":1144088.0,"Title":"Identical models in Keras and Pytorch training differently","Body":"<p>I had done some training in Keras and was attempting to move over to PyTorch for some future work. As a sanity check, I decided to replicate a successful training process from Keras to ensure everything worked the same on Pytorch. I am training the same model architecture, using the same optimizer, on the same data, with the same batch size and number of epochs. The goal is to perform some parameter estimation on an input image (basically a nonlinear regression).  <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/35JFm.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/35JFm.png\" alt=\"Keras training results\"><\/a><\/p>\n\n<p>The keras model begins to converge within around 10 epochs. Above is a plot of the model's predictions against the actual values for each image in the test set.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ZLorR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ZLorR.png\" alt=\"Pytorch training results\"><\/a><\/p>\n\n<p>The pytorch model does not converge nearly as quickly, and when it begins to stagnate, it is predicting a nearly uniform value for all inputs. Above is a plot after 10 epochs. As training continues, the prediction spread gets smaller and it begins to select around 0.01 for every input.<\/p>\n\n<p>Below is the code for the keras model and training process, where xtrain, ytrain were loaded in already.<\/p>\n\n<pre><code>nchannels = 32\nkernel_size = 3\ninput_size = 64\n\ncnn = Sequential()\ncnn.add(\n    Conv2D(\n        nchannels, \n        kernel_size=(kernel_size, kernel_size),\n        input_shape=[input_size, input_size, 1]\n    )\n)\ncnn.add(Activation('tanh'))\ncnn.add(MaxPooling2D((2, 2)))\ncnn.add(Dropout(0.15))\n\ncnn.add(Flatten())\ncnn.add(Dense(32, activation='tanh'))\ncnn.add(Dense(1, activation='linear'))\n\noptimizer = Adam(lr=1e-5, eps=1e-7)\ncnn.compile(optimizer=optimizer, loss='mse')\n\nhist = cnn.fit(\n    x=xtrain,\n    y=ytrain,\n    batch_size=32,\n    epochs=10\n)\n<\/code><\/pre>\n\n<p>Next is the pytorch model code:<\/p>\n\n<pre><code>class CNN_1Layer(nn.Module):\n    def __init__(self, nchannels=32, kernel_size=3, input_size=64):\n        super(CNN_1Layer, self).__init__()\n        self.flattened_nfeatures = \\\n            int(nchannels * (input_size-2)**2 \/ 4)\n        self.conv1 = nn.Conv2d(1, \n            nchannels, \n            kernel_size=kernel_size)\n        self.tanh1 = nn.Tanh()\n        self.pool = nn.MaxPool2d(kernel_size=2)\n        self.dropout = nn.Dropout2d(p=0.15)\n        self.fc1 = nn.Linear(self.flattened_nfeatures, 32)\n        self.tanh2 = nn.Tanh()\n        self.fc2 = nn.Linear(32, 1)\n\n    def forward(self, x):\n        x = self.conv1(x)\n        x = self.tanh1(x)\n        x = self.pool(x)\n        x = self.dropout(x)\n        x = x.view(-1, self.flattened_nfeatures)\n        x = self.fc1(x)\n        x = self.tanh2(x)\n        x = self.fc2(x)\n        return x\n\ncnn = CNN_1Layer(input_size=crop_size)\ncriterion = nn.MSELoss()\noptimizer = Adam(params=cnn.parameters(), lr=1e-5, eps=1e-7)\n\n#dataset is a wrapper around xtrain and ytrain from the keras segment\nloader = DataLoader(\n    dataset,\n    batch_size=32\n)\n\nfor i in range(10):\n    training_loss = 0.0\n    for j, data in enumerate(loader):\n        inputs, labels = data\n        optimizer.zero_grad()\n        outputs = cnn(inputs)\n        loss = criterion(outputs, labels)\n        loss.backward()\n        optimizer.step()\n<\/code><\/pre>\n\n<p>I have been training them side by side and it is not clear to me what could be causing them to behave so differently. <\/p>\n\n<p><strong>Update<\/strong><\/p>\n\n<p>I spent more time investigating this, beginning by copying the initial weights from the Keras model into the PyTorch model. This, for some reason, made it so that both models converged only if the data was not shuffled during fitting. I then explicitly initialized the weights in PyTorch using <code>xavier_uniform_<\/code> which had the same effect. In the end, I found that the PyTorch model converges in each case if I replace the line <\/p>\n\n<pre><code>x = x.view(-1, self.flattened_nfeatures)\n<\/code><\/pre>\n\n<p>with <\/p>\n\n<pre><code>x = x.permute(0, 2, 3, 1).reshape([-1, self.flattened_nfeatures])\n<\/code><\/pre>\n\n<p>I'm leaving this question unanswered, as I can think of no reason why permuting before flattening should affect the learning process at all. If anybody can explain this behavior I would love to know.<\/p>\n","answers":[]}
+{"QuestionId":57597939,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><keras><keras-layer>","CreationDate":"2019-08-21T19:14:42.493","AcceptedAnswerId":null,"OwnerUserId":11326605.0,"Title":"Can't Use Unknown Input Dims (Batch Size) When Making Custom Keras Layer","Body":"<p>I am attempting to build a custom layer in Keras (backend TensorFlow) that performs KMeans Clustering across the filters of a Convolutional Layer. In building the logic for this layer, I iterate across the batch size, but it does not seem that Keras\/TensorFlow are allowing this to happen, as the batch size is an unknown dimension until runtime.<\/p>\n\n<p>I have attempted to trace through the error messages, and it leads me to two files: keras\/engine\/training.py and keras\/engine\/training_utils.py, and from what I can tell, the error stems from a problem with a <code>ndim<\/code> variable that gets defined to <code>NoneType<\/code>, as there is no known batch size when the model is compiled.<\/p>\n\n<p>In looking through various StackOverflow and GitHub resources, I have not seen any solution for how to handle Keras\/TensorFlow rejecting an unknown batch size argument.<\/p>\n\n<p>For reference, here is the sample code:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nimport keras\nfrom sklearn.cluster import KMeans\n\nclass KMeansLayer(keras.layers.Layer):\n    def __init__(self, clusters=8, n_init=5, trainable=False, **kwargs):\n        super(KMeansLayer, self).__init__(**kwargs)\n        self.clusters = clusters\n        self.n_init = n_init\n\n    def build(self, input_shape):\n        self.output_s = (input_shape[0],input_shape[1], input_shape[2],1)\n        self.depth = input_shape[3]\n        self.built=True\n\n    def call(self, inputs):\n\n        def KMeansFunc(input_tens,clusters=self.clusters,n_init=self.n_init):\n            base_mat = np.zeros((input_tens.shape[0],input_tens.shape[1],input_tens.shape[2]))\n\n            for frame in range(input_tens.shape[0]):\n                init_mat = np.zeros((input_tens.shape[1]*input_tens.shape[2]))\n                # print(init_mat.shape)\n                reshape_mat = np.reshape(input_tens[frame],(input_tens.shape[1]*input_tens.shape[2],input_tens.shape[3]))\n                # print(reshape_mat.shape)\n                kmeans_init = KMeans(n_clusters=clusters, n_init=n_init)\n                class_pred = kmeans_init.fit_predict(reshape_mat)\n\n                for clust in range(self.clusters):\n                    init_mat[class_pred==clust] = np.mean(reshape_mat[class_pred==clust],axis=1)\n                    init_mat[class_pred==clust] = np.mean(reshape_mat[class_pred==clust],None)\n                # print(base_mat.shape)\n\n                base_mat[frame]=np.reshape(init_mat,(input_tens.shape[1],input_tens.shape[2]))\n\n            return np.expand_dims(base_mat,axis=-1).astype('float32')\n\n\n        output = tf.py_func(KMeansFunc,[inputs],tf.float32) \n        return output\n\n    def compute_output_shape(self, input_shape):\n        return self.output_s\n\n\ninput_1 = keras.Input(shape=(28,28,1), name='input_1', dtype='float32')\n\nconv_1 = keras.layers.Conv2D(filters=20, kernel_size=3, strides=1, padding='same', data_format='channels_last', activation='elu', kernel_initializer='glorot_uniform')(input_1)\npool_1 = keras.layers.MaxPooling2D(pool_size=2, padding='same', data_format='channels_last')(conv_1)\n\nup_conv_1 = keras.layers.SeparableConv2D(filters=20, kernel_size=2, strides=1, padding='same', data_format='channels_last', activation='elu', kernel_initializer='glorot_uniform')(pool_1)\nup_1 = keras.layers.UpSampling2D(size=(2, 2), interpolation='bilinear')(up_conv_1)\nconv_2 = keras.layers.Conv2D(filters=20, kernel_size=3, strides=1, padding='same', data_format='channels_last', activation='elu', kernel_initializer='glorot_uniform')(up_1)\n\nconv_3 = keras.layers.Conv2D(filters=3, kernel_size=3, strides=1, padding='same', data_format='channels_last', activation='elu', kernel_initializer='glorot_uniform')(conv_2)\n\nkmeans_out = KMeansLayer(clusters=8,n_init=5)(conv_3)\n\n\nmodel = keras.Model(inputs=[input_1], outputs=kmeans_out)\nkeras.utils.plot_model(model, show_shapes=True)\nmodel.compile(optimizer='adam', loss='mse', metrics=['mse'])\n<\/code><\/pre>\n\n<p>As can be seen in the above code, if I have an input into my custom layer of size <code>(batch_size,28,28,3)<\/code>, I expect to create output of size <code>(batch_size,28,28,1)<\/code>.<\/p>\n\n<p>The error that I get as a result of running the above code is:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"example_error_file.py\", line 64, in &lt;module&gt;\n    model.compile(optimizer='adam', loss='mse', metrics=['mse'])\n  File \"~\/fluoro\/fenv\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\", line 347, in compile\n    sample_weight, mask)\n  File \"~\/fluoro\/fenv\/lib\/python3.6\/site-packages\/keras\/engine\/training_utils.py\", line 426, in weighted\n    axis=list(range(weight_ndim, ndim)))\nTypeError: 'NoneType' object cannot be interpreted as an integer\n<\/code><\/pre>\n\n<p>I have two main questions:<\/p>\n\n<ul>\n<li>Is there something I am doing incorrectly when defining my custom Keras Layer?<\/li>\n<li>Is there any way to force Keras to operate without knowing the batch size in this instance (which seems reasonable)?<\/li>\n<\/ul>\n\n<p>TensorFlow Version: 1.7.0\nKeras Version: 2.2.4\nPython: 3.6.6<\/p>\n","answers":[{"AnswerId":"57602428","CreationDate":"2019-08-22T05:12:47.150","ParentId":null,"OwnerUserId":"9480505","Title":null,"Body":"<p>The main problem seems to be that event if we were to just look at this in entirely tensorflow, since the shape isn't known until runtime, when tensorflow tried to compile this graph its gonna get some errors with the whole looping through None.<\/p>\n\n<p>Instead of using tf.pyfunc consider using tf.function or autograph (if you're on tensorflow version 1.14.0 or greater).<\/p>\n\n<p>Alternatively you could use tf.scan which would apply a function to each element in a tensor, where each element is made up of each tensor that is unpacked from the original tensor at dimension 0.<\/p>\n\n<p>Either should work! <\/p>\n"}]}
+{"QuestionId":57598011,"AnswerCount":1,"Tags":"<python><tensorflow><keras><feature-selection><tensorflow-serving>","CreationDate":"2019-08-21T19:21:43.687","AcceptedAnswerId":null,"OwnerUserId":11958497.0,"Title":"How to save a Keras model consisting of feature layer?","Body":"<p>I am trying a save a Keras model trained using Sequential and having a feature_layer. feature_layer is created using feature_columns consisting of numerical and categorical features. While saving I get error \n\"Layer sequential_2 is not connected, no input to return.\"<\/p>\n\n<p>I followed the tutorial provided on tensorfow.org website. <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/keras\/feature_columns\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/tutorials\/keras\/feature_columns<\/a><\/p>\n\n<p>I used tensorflow version 1.14 by commenting original code:<\/p>\n\n<pre><code>#try:\n  # %tensorflow_version only exists in Colab.\n#  %tensorflow_version 2.x\n#except Exception:\n#  pass\nimport tensorflow as tf\n<\/code><\/pre>\n\n<h1>below code is used to build and train model<\/h1>\n\n<pre><code>model = tf.keras.Sequential([\n  feature_layer,\n  layers.Dense(128, activation='relu'),\n  layers.Dense(128, activation='relu'),\n  layers.Dense(1, activation='sigmoid')\n])\n\nmodel.compile(optimizer='adam',\n              loss='binary_crossentropy',\n              metrics=['accuracy'],\n              run_eagerly=False)\n\nmodel.fit(train_ds,\n          validation_data=val_ds,\n          epochs=5)\n<\/code><\/pre>\n\n<h1>saving the model for tensorflow serving<\/h1>\n\n<pre><code>export_path = 'tmp\/1'\nwith tf.keras.backend.get_session() as sess:\n    tf.saved_model.simple_save(\n        sess,\n        export_path,\n        inputs={'input': model.input},\n        outputs={'output': t for t in model.outputs})\n<\/code><\/pre>\n\n<p>I get the error:\nAttributeError: Layer sequential_2 is not connected, no input to return.<\/p>\n\n<p>I get the same error while calling model.input()<\/p>\n","answers":[{"AnswerId":"57689493","CreationDate":"2019-08-28T09:46:29.420","ParentId":null,"OwnerUserId":"11530462","Title":null,"Body":"<p>In addition to commenting the <code>try<\/code> and <code>except<\/code> block, since you are using Tensorflow Version 1.14, adding the below line of code will resolve your issue.<\/p>\n\n<pre><code>tf.enable_eager_execution()\n<\/code><\/pre>\n\n<p>Please find the below screenshot:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/x1ikL.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/x1ikL.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>By doing that single change, we'll get expected output, as shown in the screenshot below.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/AyBEz.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/AyBEz.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Please let me know if you are facing any other challenges.<\/p>\n\n<p>If you feel this answer is useful, kindly accept this answer and\/or up vote it. Thanks.<\/p>\n"}]}
+{"QuestionId":57598436,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-08-21T19:57:10.720","AcceptedAnswerId":null,"OwnerUserId":8523421.0,"Title":"When I use tf.keras.layers.BatchNormalization() it throws a Type Error","Body":"<p>Using tf.keras.layers.BatchNormalization() gives \nTypeError: Incompatible types:  vs. int64. Value is 0<\/p>\n\n<p>use this link for error reproduction::\n<a href=\"https:\/\/colab.research.google.com\/drive\/16IEd1GjgvPIiG0xnB3ado11u4jLHQZLY\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/16IEd1GjgvPIiG0xnB3ado11u4jLHQZLY<\/a><\/p>\n\n<p>I want to make it trainable.<\/p>\n\n<pre><code>  enc = tf.keras.layers.TimeDistributed(tf.keras.layers.Conv2D(32, kernel_size=3, activation='relu',padding = 'same',kernel_initializer='glorot_uniform'))(enc)\n  print('encoder_mask', enc.shape)\n\n  enc = tf.keras.layers.TimeDistributed(tf.keras.layers.BatchNormalization(epsilon=BN_EPSILON, momentum=BN_MIMENTUM))(enc)\n  print('encoder_mask', enc.shape)\n\n  enc = tf.keras.layers.TimeDistributed(tf.keras.layers.Conv2D(16, activation='relu',kernel_size=3,kernel_initializer='glorot_uniform'))(enc)\n  print('encoder_mask', enc.shape)\n\n<\/code><\/pre>\n\n<p>After I compile it the error I get is given below even if I dont use these variables (BN_EPSILON,BN_MIMENTUM ).<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-13-c9408a385d78&gt; in &lt;module&gt;()\n      1 lr = 0.0001\n----&gt; 2 train_op = tf.train.AdamOptimizer(lr).minimize(reconstuction_loss)\n\n7 frames\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/training\/optimizer.py in minimize(self, loss, global_step, var_list, gate_gradients, aggregation_method, colocate_gradients_with_ops, name, grad_loss)\n    401         aggregation_method=aggregation_method,\n    402         colocate_gradients_with_ops=colocate_gradients_with_ops,\n--&gt; 403         grad_loss=grad_loss)\n    404 \n    405     vars_with_grad = [v for g, v in grads_and_vars if g is not None]\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/training\/optimizer.py in compute_gradients(self, loss, var_list, gate_gradients, aggregation_method, colocate_gradients_with_ops, grad_loss)\n    510         gate_gradients=(gate_gradients == Optimizer.GATE_OP),\n    511         aggregation_method=aggregation_method,\n--&gt; 512         colocate_gradients_with_ops=colocate_gradients_with_ops)\n    513     if gate_gradients == Optimizer.GATE_GRAPH:\n    514       grads = control_flow_ops.tuple(grads)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/gradients_impl.py in gradients(ys, xs, grad_ys, name, colocate_gradients_with_ops, gate_gradients, aggregation_method, stop_gradients, unconnected_gradients)\n    156         ys, xs, grad_ys, name, colocate_gradients_with_ops,\n    157         gate_gradients, aggregation_method, stop_gradients,\n--&gt; 158         unconnected_gradients)\n    159   # pylint: enable=protected-access\n    160 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/gradients_util.py in _GradientsHelper(ys, xs, grad_ys, name, colocate_gradients_with_ops, gate_gradients, aggregation_method, stop_gradients, unconnected_gradients, src_graph)\n    718               # issue here because of zeros.\n    719               if loop_state:\n--&gt; 720                 out_grads[i] = loop_state.ZerosLike(op, i)\n    721               else:\n    722                 out_grads[i] = control_flow_ops.ZerosLikeOutsideLoop(op, i)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/control_flow_ops.py in ZerosLike(self, op, index)\n   1229       # If the shape is known statically, just create a zero tensor with\n   1230       # the right shape in the grad loop context.\n-&gt; 1231       result = constant_op.constant(0, shape=shape.dims, dtype=val.dtype)\n   1232       if dead_branch:\n   1233         # op is a cond switch. Guard the zero tensor with a switch.\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/constant_op.py in constant(value, dtype, shape, name)\n    244   \"\"\"\n    245   return _constant_impl(value, dtype, shape, name, verify_shape=False,\n--&gt; 246                         allow_broadcast=True)\n    247 \n    248 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/constant_op.py in _constant_impl(value, dtype, shape, name, verify_shape, allow_broadcast)\n    282       tensor_util.make_tensor_proto(\n    283           value, dtype=dtype, shape=shape, verify_shape=verify_shape,\n--&gt; 284           allow_broadcast=allow_broadcast))\n    285   dtype_value = attr_value_pb2.AttrValue(type=tensor_value.tensor.dtype)\n    286   const_tensor = g.create_op(\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/tensor_util.py in make_tensor_proto(values, dtype, shape, verify_shape, allow_broadcast)\n    499                             dtype.base_dtype != numpy_dtype.base_dtype):\n    500     raise TypeError(\"Incompatible types: %s vs. %s. Value is %s\" %\n--&gt; 501                     (dtype, nparray.dtype, values))\n    502 \n    503   # If shape is not given, get the shape from the numpy array.\n\nTypeError: Incompatible types: &lt;dtype: 'resource'&gt; vs. int64. Value is 0\n<\/code><\/pre>\n\n<p>Reproduce error by running this::<\/p>\n\n<pre><code>batch_size = 20\ninp = tf.placeholder(tf.float32, [batch_size, 19, 64, 64, 3])\nout = tf.placeholder(tf.float32, [batch_size, 19, 60, 60, 16])\ndef model(inp):\n\n  enc = tf.keras.layers.TimeDistributed(tf.keras.layers.Conv2D(128, activation='relu', kernel_size=3,kernel_initializer='glorot_uniform'))(inp)\n  enc = tf.keras.layers.TimeDistributed(tf.keras.layers.BatchNormalization())(enc)\n  enc = tf.keras.layers.TimeDistributed(tf.keras.layers.Conv2D(16, activation='relu',kernel_size=3,kernel_initializer='glorot_uniform'))(enc)\n  return enc\n\npred = model(inp)\n\nloss = tf.reduce_mean(tf.keras.backend.binary_crossentropy(out, pred))\nlr = 0.0001\ntrain_op = tf.train.AdamOptimizer(lr).minimize(loss)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57612503","CreationDate":"2019-08-22T15:15:15.940","ParentId":null,"OwnerUserId":"3083470","Title":null,"Body":"<p>Take a look at a fixed version of your example:<\/p>\n\n<p><a href=\"https:\/\/colab.research.google.com\/drive\/19DQlzeD2oN8KU37SUFexRzEwrhQdyAe5\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/19DQlzeD2oN8KU37SUFexRzEwrhQdyAe5<\/a><\/p>\n\n<p>You weren't using the keras functional API correctly - this is the conventional way to do it. You can find more information <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/keras\/functional\" rel=\"nofollow noreferrer\">in this tf.keras guide<\/a><\/p>\n"}]}
+{"QuestionId":57598981,"AnswerCount":2,"Tags":"<tensorflow><google-bigquery><google-cloud-datalab>","CreationDate":"2019-08-21T20:46:26.213","AcceptedAnswerId":null,"OwnerUserId":11959182.0,"Title":"How to export tensorflow models on datalab to use in bigquery?","Body":"<p>I have a trained model in TensorFlow on Google Cloud Datalab. \nI want to export it and import it in BigQuery and predict using BigQuery. How do i export it with path as <code>gs:\/\/*<\/code>?<\/p>\n","answers":[{"AnswerId":"57618272","CreationDate":"2019-08-22T23:52:23.620","ParentId":null,"OwnerUserId":"3645038","Title":null,"Body":"<p>If you are using TensorFlow 1.14 or higher and Keras, then:<\/p>\n\n<pre><code>tf.saved_model.save(model,  'gs:\/\/bucket\/dir')\n<\/code><\/pre>\n\n<p>See <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/saved_model\/save\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/saved_model\/save<\/a><\/p>\n"},{"AnswerId":"57619896","CreationDate":"2019-08-23T04:26:43.677","ParentId":null,"OwnerUserId":"3645038","Title":null,"Body":"<p>If you have code written in an earlier version of TensorFlow, it probably uses the Estimator API. In that case, use:<\/p>\n\n<pre><code>estimator.export_savedmodel('gs:\/\/bucket\/dir', serving_input_fn)\n<\/code><\/pre>\n\n<p>where the serving function has to be defined with placeholders, one for each input to your model:<\/p>\n\n<pre><code>def serving_input_fn():\n    feature_placeholders = {\n        'input1': tf.placeholder(tf.string, [None]),\n        'input2': tf.placeholder(tf.float32, [None]),\n    }\n    features = {\n        key: tf.expand_dims(tensor, -1)\n        for key, tensor in feature_placeholders.items()\n    }\n    return tf.estimator.export.ServingInputReceiver(features, feature_placeholders)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57599111,"AnswerCount":0,"Tags":"<tensorflow><keras><lstm>","CreationDate":"2019-08-21T20:59:51.213","AcceptedAnswerId":null,"OwnerUserId":5133761.0,"Title":"Neural network text prediction with many distinct users","Body":"<p>I am in the process of trying to develop a Keras LSTM text prediction model for thousands of users. Most examples I see aggregate the training data into a single model, but what if I want to improve accuracy of each specific user?<\/p>\n\n<p>In my use case if my users have fairly distinct patterns of typing compared to each other, it seems to make sense to have individualized models trained for each user. <\/p>\n\n<p>Example:\nUser 1 may frequently type \"my favorite breakfast is oranges\" and select it often when typing \"my...\" , but for user 2 this would be an irritating thing to see in their suggestions and wouldn't want that to appear. <\/p>\n\n<p>However this doesn't seem practical to run thousands of models in individual tensorflow servers.<\/p>\n\n<p>Is there a better way to approach this problem? <\/p>\n","answers":[]}
+{"QuestionId":57599389,"AnswerCount":0,"Tags":"<keras><deep-learning><gpu><amazon-sagemaker><mxnet>","CreationDate":"2019-08-21T21:31:40.957","AcceptedAnswerId":null,"OwnerUserId":5503494.0,"Title":"how to fix pop from empty list error when using keras.layers.Dot in GPU mode","Body":"<p>I'm trying to train a simple model with embeddings on Sagemaker, using Mxnet backhand and aiming to run across multiple GPUs.<\/p>\n\n<p>The line below is causing an error:<\/p>\n\n<pre><code>   preprod = keras.layers.Dot(axes=1, normalize=False)([bet_vec, user_vec])\n<\/code><\/pre>\n\n<p>Whenever I use keras.layers.Dot I get the following error msg:<\/p>\n\n<p>File \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/backend\/mxnet_backend.py\", line 2292, in squeeze\n    assert shape.pop(axis) == 1, 'Can only squeeze size 1 dimension'\nIndexError: pop from empty list<\/p>\n\n<p>The code runs fine locally but it does not work using Mxnet\/GPUs...<\/p>\n\n<p>If anyone could help it would be much appreciated.<\/p>\n\n<p>Thank you :-)<\/p>\n","answers":[]}
+{"QuestionId":57599555,"AnswerCount":1,"Tags":"<python-3.x><pip><pytorch><tensorboard>","CreationDate":"2019-08-21T21:51:19.577","AcceptedAnswerId":null,"OwnerUserId":311744.0,"Title":"ModuleNotFoundError: No module named 'past' when installing tensorboard with pytorch 1.2","Body":"<p>I'm trying out tensorboard with pytorch by following this: <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensorboard.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/tensorboard.html<\/a><\/p>\n\n<p>I've installed tensorboard with <code>pip install tb-nightly<\/code><\/p>\n\n<p>The command <code>tensorboard --logdir=runs<\/code> starts ok.<\/p>\n\n<p>But the line <code>self.writer = SummaryWriter()<\/code><\/p>\n\n<p>Gives the following error:<\/p>\n\n<pre><code>ModuleNotFoundError: No module named 'past'\n<\/code><\/pre>\n\n<p>How do I fix this?<\/p>\n","answers":[{"AnswerId":"57599556","CreationDate":"2019-08-21T21:51:19.577","ParentId":null,"OwnerUserId":"311744","Title":null,"Body":"<p>Following this issue: <a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/22389\" rel=\"noreferrer\">https:\/\/github.com\/pytorch\/pytorch\/issues\/22389<\/a>,<\/p>\n\n<p>Adding <code>future<\/code> to the list of requirements solved the problem<\/p>\n\n<pre><code># requirements.txt:\ntb-nightly\nfuture\n<\/code><\/pre>\n\n<p><code>pip install -r requirements.txt<\/code><\/p>\n"}]}
+{"QuestionId":57599641,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><deep-learning>","CreationDate":"2019-08-21T22:00:19.213","AcceptedAnswerId":null,"OwnerUserId":11959313.0,"Title":"Using tf.train.AdamOptimizer twice for same network","Body":"<p>I am trying to fine tune an existing neural network (resnet v2 152) for a new dataset.  The way I have training set up right now is to first train just the last fully connected layer for a few epochs and second to train the entire model for a few epochs.  In order to make this work I defined two optimizers.  When I use the gradient descent optimizer I am able to train the model. However. when I use the AdamOptimizer the model doesn't train.   <\/p>\n\n<p>The most important thing I have noticed is when I include <code>tf.global_variables_initializer()<\/code> using the gradient descent optimizer doesn't train either.  However, it appears <code>tf.global_varaibles_initializer()<\/code> is required to use the AdamOptimizer<\/p>\n\n<p>If I don't include <code>tf.global_varaibles_initializer()<\/code> and use gradient descent the network trains as expected.  <\/p>\n\n<p>I have included a snippet of the code below. This code is based on the VGG trainer available at: <a href=\"https:\/\/gist.github.com\/omoindrot\/dedc857cdc0e680dfb1be99762990c9c\/\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/omoindrot\/dedc857cdc0e680dfb1be99762990c9c\/<\/a><\/p>\n\n<pre><code> variables_to_restore = tf.contrib.framework.get_variables_to_restore(exclude=['resnet_v2_152\/logits'])\n\n init_fn = tf.contrib.framework.assign_from_checkpoint_fn(model_path, variables_to_restore)\n\n # Initialization operation from scratch for the new \"fc\" layers\n\n # `get_variables` will only return the variables whose name starts with the given pattern\n\n c_variables = tf.contrib.framework.get_variables('resnet_v2_152\/logits')\n fc_init = tf.variables_initializer(fc_variables)\n\n full_variables =tf.contrib.framework.get_variables()\n full_init = tf.variables_initializer(full_variables)\n\n ## ---------------------------------------------------------------------\n\n ## Using tf.losses, any loss is added to the tf.GraphKeys.LOSSES collection\n\n ## We can then call the total loss easily\n\n loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=labels, logits=logits)\n\n ## First we want to train only the reinitialized last layer fc for a few epochs.\n ## We run minimize the loss only with respect to the fc variables (weight and bias).\n fc_optimizer = tf.train.AdamOptimizer(args.learning_rate1, name=\"adam1\")\n fc_train_op = fc_optimizer.minimize(loss, var_list=[fc_variables])\n\n # Then we want to finetune the entire model for a few epochs.\n\n # We run minimize the loss only with respect to all the variables.\n\n full_optimizer = tf.train.AdamOptimizer(args.learning_rate2, name =\"adam2\")\n\n full_train_op = full_optimizer.minimize(loss, var_list=[full_variables])\n\n init = tf.global_variables_initializer()\n<\/code><\/pre>\n\n<p>No error is given when training but the training and validation accuracies don't update so it appears that the optimizer isn't actually running.<\/p>\n\n<p>Like I said I think it has to do with the inclusion of <code>tf.global_variables_initializer()<\/code>, but I am not sure how to get around that for using the AdamOptimizer.<\/p>\n\n<p>Thanks in advance<\/p>\n","answers":[]}
+{"QuestionId":57599892,"AnswerCount":1,"Tags":"<python><tensorflow><google-cloud-ml>","CreationDate":"2019-08-21T22:27:13.187","AcceptedAnswerId":57600297.0,"OwnerUserId":3864587.0,"Title":"Return key with tf.estimator model predictions","Body":"<p>I am using a custom tensorflow Estimator and am trying to use <\/p>\n\n<pre><code>tf.contrib.estimator.forward_features\n<\/code><\/pre>\n\n<p>to return a key column along with my outputs. I followed <a href=\"https:\/\/stackoverflow.com\/questions\/49271114\/define-instance-key-index-number-for-cloud-machine-learning-prediction\">post1<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/49542369\/how-to-get-keys-in-batch-predictions-with-ml-engine-using-a-custom-model?noredirect=1&amp;lq=1\">post2<\/a><\/p>\n\n<p>and applied <\/p>\n\n<pre><code>tf.contrib.estimator.forward_features\n<\/code><\/pre>\n\n<p>but I am receiving the the error<\/p>\n\n<pre><code> Predictions should be a dict to be able to forward features. Given: &lt;class 'tensorflow.python.framework.ops.Tensor'&gt;\n<\/code><\/pre>\n\n<p>My model functions looks like this<\/p>\n\n<pre><code>def model_fn(features, labels, mode):\n\n    values = nnet(features)\n    if mode == tf.estimator.ModeKeys.TRAIN:\n       is_training = True\n    else:\n       is_training = False\n\n    if mode == tf.estimator.ModeKeys.PREDICT:\n       predictions = {\n        'class_ids': tf.argmax(tf.nn.softmax(values),1),\n        'probabilities': tf.nn.softmax(values),\n        'logits': values,\n        }\n        export_outputs = {\n            'prediction': tf.estimator.export.PredictOutput(predictions)\n        }\n    return tf.estimator.EstimatorSpec(mode,predictions=predictions,export_outputs=export_outputs)\n\n    labels_one_hot=tf.one_hot(labels,4)\n\n    score= tf.argmax(values,axis=1)\n\n    loss_op= tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=values,labels=labels_one_hot))\n\n    gradients = tf.gradients(loss_op, tf.trainable_variables())\n\n    optimizer = tf.train.AdamOptimizer(learning_rate=.0001)\n\n    update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\n\n    if mode == tf.estimator.ModeKeys.TRAIN:\n\n        train_optimizer = optimizer.apply_gradients(zip(gradients, tf.trainable_variables()),\n\n        global_step=tf.train.get_global_step())\n\n        acc_op=tf.metrics.accuracy( labels= labels,predictions=tf.argmax(values, axis=1))\n\n        tf.summary.scalar('accuracy_rate', acc_op[1])\n        update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\n\n        estim_specs = tf.estimator.EstimatorSpec(\n             mode=mode,\n             predictions=score,\n             loss=loss_op,\n             train_op=train_optimizer,\n        eval_metric_ops={'acc': acc_op})\n        return estim_specs \n\n\n    if mode == tf.estimator.ModeKeys.EVAL:\n        predicted_indices = tf.argmax(values, axis=1)\n        eval_metric_ops = {\n          'accuracy': tf.metrics.accuracy(labels, predicted_indices)}\n\n        return tf.estimator.EstimatorSpec(mode,loss=loss_op,eval_metric_ops=eval_metric_ops)\n<\/code><\/pre>\n\n<p>I am calling my estimator like this<\/p>\n\n<pre><code>estimator = tf.estimator.Estimator(\n    model_fn= model_fn,\n    config= tf.estimator.RunConfig(\n            save_checkpoints_steps = 2000,\n            keep_checkpoint_max = 10,\n            tf_random_seed = 101),\n          model_dir= \"tf_dir\")\n\nestimator = tf.contrib.estimator.forward_features(\n  estimator,'key')\n\ntf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)\n<\/code><\/pre>\n\n<p>I am returning a dictionary for my predictions and my code runs fine if I don't call <\/p>\n\n<pre><code>tf.contrib.estimator.forward_features\n<\/code><\/pre>\n","answers":[{"AnswerId":"57600297","CreationDate":"2019-08-21T23:25:20.777","ParentId":null,"OwnerUserId":"3864587","Title":null,"Body":"<p>The solution was to explicitly carry the key into model function and add it to the dictionary of outputs. <\/p>\n\n<p>The lines of code that need to be added are <\/p>\n\n<pre><code>def model_fn(features, labels, mode):\n\n    values = nnet(features)\n<\/code><\/pre>\n\n<p>add the key here<\/p>\n\n<pre><code>    key=features[\"key_column\"]\n\n    if mode == tf.estimator.ModeKeys.TRAIN:\n        is_training = True\n    else:\n        is_training = False\n\n\n    if mode == tf.estimator.ModeKeys.PREDICT:\n        predictions = {\n            'class_ids': tf.argmax(tf.nn.softmax(values),1),\n             'probabilities': tf.nn.softmax(values),\n             'logits': values,\n<\/code><\/pre>\n\n<p>add the key to the dictionary outputs              <\/p>\n\n<pre><code>            'key_column':key\n           }\n        export_outputs = {\n        'prediction': tf.estimator.export.PredictOutput(predictions)\n          }\n        return tf.estimator.EstimatorSpec(mode,predictions=predictions,export_outputs=export_outputs)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57600276,"AnswerCount":0,"Tags":"<python><image><deep-learning><pytorch><dataloader>","CreationDate":"2019-08-21T23:20:26.020","AcceptedAnswerId":null,"OwnerUserId":7938316.0,"Title":"Pytorch Parameter-based Cropping custom sub-images from DataLoader image","Body":"<p>There are rectangular images in train &amp; validation folders, and the images are accessed via Pytorch through DataLoader module. Without performing, <\/p>\n\n<pre><code>transforms.CenterCrop(size) or transforms.RandomCrop(size)\n<\/code><\/pre>\n\n<p>I need to preserve their original shape and size, so that while training\/testing I can input square-shaped images, via some custom cropping technique. This CustomCrop function may take as inputs certain parameters, different for different images.<\/p>\n\n<p>When I tried this custom-cropping approach only for testing, I had to exclude the cropping line from data_transforms for TEST. <\/p>\n\n<pre><code>data_transforms = {\n    TRAIN: transforms.Compose([\n        transforms.RandomCrop(size),\n        transforms.ToTensor(),\n    ]),\n    TEST: transforms.Compose([\n        #CROPPING LINE EXCLUDED FROM HERE\n        transforms.ToTensor(),\n    ])\n}\nimage_datasets = {\n    x: ImageFolder(\n        os.path.join(path, mapping[x]), \n        transform=data_transforms[x]\n    )\n    for x in [TRAIN, TEST]\n}\ndataloaders = {\n       x: DataLoader(\n       image_datasets[x], batch_size=batch_size\n       )\n    for x in [TRAIN, TEST]\n}    \n<\/code><\/pre>\n\n<p>And my intention is to do something like...<\/p>\n\n<pre><code>for imgs, lbls, paths in iter(dataloaders[TEST]):\n    data = (imgs, lbls)\n    inputs, labels = data\n    cropped_input = myCustomCrop(inputs, param1, param2) #this crops it to square size, compatible for my model\n    [...]  #some code      \n    optimizer.zero_grad()            \n    outputs = model(cropped_input)\n<\/code><\/pre>\n\n<p>...so that I can crop it during run-time as per my need.<\/p>\n\n<p>But this gives me an error message (For image size (100,75) in folder):<\/p>\n\n<pre><code>--&gt; 218     for imgs, lbls, paths in iter(dataloaders[TEST]):\nRuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 0. Got 100 and 75 in dimension 2 at ...\n<\/code><\/pre>\n\n<p>I am understanding that, DataLoader stacks the images and they need to be equal-sized at runtime. But, how can I then implement my custom-crop function, which may be different for different images (based on input parameter to CustomCrop function). Any suggestions would be highly appreciated. Thanks.<\/p>\n","answers":[]}
+{"QuestionId":57600323,"AnswerCount":0,"Tags":"<python><tensorflow><pycharm>","CreationDate":"2019-08-21T23:29:08.450","AcceptedAnswerId":null,"OwnerUserId":11102846.0,"Title":"How can I get full PyCharm Code executing results?","Body":"<p>I'm using PyCharm to execute basic classification code in TensorFlow.\nI have built the full code and the code consist of two stages: the first one for defining and compiling the model and the second one for the data visualizing.\nHowever, when I'm trying to execute the full code, it executes it successfully but without showing me the result of the visualizing part.\nI mean it shows me all training and evaluating results but without the showing the results of part two.\nplease any help !?<\/p>\n\n<p>part 1:<\/p>\n\n<pre><code>from __future__ import absolute_import, division, print_function, unicode_literals\nimport tensorflow as tf\nimport matplotlib.cm\nimport numpy as np\nimport matplotlib.pyplot as plt\n\nprint(tf.__version__)\n\nmnist= tf.keras.datasets.fashion_mnist\n\n(training_images, training_labels), (test_images, test_labels)= mnist.load_data()\nmodel= tf.keras.models.Sequential([tf.keras.layers.Flatten(input_shape=(28,28)),\n                                   tf.keras.layers.Dense(128, activation=tf.nn.relu),\n                                   tf.keras.layers.Dense(10, activation=tf.nn.softmax)])\n\n\nmodel.compile(optimizer= tf.optimizers.Adam(),\n              loss= 'sparse_categorical_crossentropy',\n              metrics= ['accuracy'])\n\nmodel.fit(training_images, training_labels, epochs=5)\n<\/code><\/pre>\n\n<p>part two:<\/p>\n\n<pre><code> def plot_image(i, predictions_array, true_label, img):\n    predictions_array, true_label, img= predictions_array[i], true_label[i], img[i]\n    plt.grid(False)\n    plt.xticks([])\n    plt.yticks([])\n    plt.imshow(img, cmap= plt.binary)\n    plt.show()\n\n    predicted_label= np.argmax(predictions_array)\n    if predicted_label== true_label:\n        color= 'blue'\n    else:\n        color= 'red'\n\n    plt.xlabel(\"{} {:2.0f}% ({})\".format(class_names[predicted_label], 100*np.max(predictions_array), class_names[true_label]), color=color)\n\ndef plot_value_array(i, predictions_array, true_label):\n    predictions_array, true_label = predictions_array[i], true_label[i]\n    plt.grid(False)\n    plt.xticks([])\n    plt.yticks([])\n    thisplot= plt.bar(range(10), predictions_array, color=\"#777777\")\n    plt.ylim([0,1])\n    predicted_label= np.argmax(predictions_array)\n\n    thisplot[predicted_label].set_color('red')\n    thisplot[true_label].set_color('blue')\n\n    i=0\n    plt.figure(figsize=(6,3))\n    plt.subplot(1,2,1)\n    plot_image(i, predictions, test_labels, test_images)\n    plt.subplot(1,2,2)\n    plot_value_array(i, predictions, test_labels)\n    plt.show()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57600721,"AnswerCount":0,"Tags":"<python><tensorflow><keras><h5py>","CreationDate":"2019-08-22T00:45:48.447","AcceptedAnswerId":null,"OwnerUserId":11522398.0,"Title":"Saving custom variables in Keras .h5 file","Body":"<p>I'm developing a RNN for a project and I need to train it on a computer and be able to predict on another. The solution I found is to save the model into a .h5 file using the code below:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>... # Train the data etc....\nmodel.save(\"model.h5\")\n<\/code><\/pre>\n\n<p>My problem is that I need to store some meta-data from my training dataset and pre-process and be able to load it together with the model. (e.g. name of dataset file, size of the dataset file, number of characters, etc...)<\/p>\n\n<p>I don't want to store this information in a second file (e.g. a .txt file) because I would have to use two files. I don't want to use any additional library or framework for this task.<\/p>\n\n<p>I was thinking (<em>brainstorming<\/em>) a code like this:<\/p>\n\n<pre><code>model.save(\"model.h5\", metaData={'myVariableName': myVariable})\n<\/code><\/pre>\n\n<p>And to load would be:<\/p>\n\n<pre><code>myVariable = model.load(\"model.h5\").getMetaData('myVariableName')\n<\/code><\/pre>\n\n<p>I know this is not possible in the <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-save-a-keras-model\" rel=\"nofollow noreferrer\">current version<\/a> and I already read Keras doc, but I couldn't find any efficient method to do that. Notice that what I'm asking is different from <a href=\"https:\/\/keras.io\/getting-started\/faq\/#handling-custom-layers-or-other-custom-objects-in-saved-models\" rel=\"nofollow noreferrer\"><code>custom_object<\/code><\/a> because want to save and load my own variables.<\/p>\n\n<p>Is there a smarter approach to solve this problem? <\/p>\n","answers":[]}
+{"QuestionId":57600723,"AnswerCount":0,"Tags":"<python><numpy><tensorflow><keras><lstm>","CreationDate":"2019-08-22T00:46:03.897","AcceptedAnswerId":null,"OwnerUserId":11810355.0,"Title":"fit_generator is still loading everything into memory","Body":"<p>I'm trying to train a cnn lstm on 720p video. Each video is 100 frames and I believe the resolution is necessary for what I need this application to do. I'm using a generator to train on but it tries to load everything into memory and then doesn't train. I'm hoping someone can help me figure out what is going on.<\/p>\n\n<pre><code>2019-08-21 17:27:26.322181: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1433] Found device 0 with properties:\nname: GeForce GTX 1080 major: 6 minor: 1 memoryClockRate(GHz): 1.7335\npciBusID: 0000:01:00.0\ntotalMemory: 8.00GiB freeMemory: 6.61GiB\n2019-08-21 17:27:26.326940: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1512] Adding visible gpu devices: 0\n2019-08-21 17:27:26.811347: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:984] Device interconnect StreamExecutor with strength 1 edge matrix:\n2019-08-21 17:27:26.814274: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:990]      0\n2019-08-21 17:27:26.815961: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1003] 0:   N\n2019-08-21 17:27:26.817783: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow \ndevice (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 6362 MB memory) -&gt; physical GPU (device: 0, name: GeForce GTX 1080, pci bus id: 0000:01:00.0, compute capability: 6.1)\n2019-08-21 17:27:27.195352: W tensorflow\/core\/framework\/allocator.cc:124] \nAllocation of 30198988800 exceeds 10% of system memory.\n<\/code><\/pre>\n\n<p>This is the main issue I am running into. I tried changing my generator to look like many other examples online but none of them seem to work. I Tried changing my model size to see if maybe the model was to large to load into memory but the same error comes up every time. I've tried changing the batch size in generator function but same the problem comes up.<\/p>\n\n<p>My generator class looks like this:<\/p>\n\n<pre><code>class Generator(Sequence):\n    def __init__(self, available_ids,labels, batch_size=1, dim= \n(100,720,1280,3), n_channels=1, n_classes=38, shuffle=True):\n    self.available_ids = available_ids\n    self.batch_size = batch_size\n    self.dim = dim\n    self.labels = labels\n    self.n_channels = n_channels\n    self.n_classes = n_classes\n    self.shuffle= shuffle\n    self.on_epoch_end()\n\ndef on_epoch_end(self):\n    self.indexs = np.arange(start=1, stop=len(self.available_ids)+1)\n    if self.shuffle:\n        np.random.shuffle(self.indexs)\n\n\ndef __len__(self):\n    return int(len(self.available_ids)\/self.batch_size)\n\ndef __data_gen(self, batch_ids):\n\n    X = np.empty((self.batch_size, *self.dim))\n    y = np.empty((self.batch_size), dtype=int)\n\n    for i, ID in enumerate(batch_ids):\n        X[i,] = np.load('train_data\/train_vid_{}.npy'.format(ID))\n        y[i] = self.labels[ID-1]\n\n    print(X.shape)\n\n    return X, to_categorical(y, num_classes=self.n_classes)\n\ndef __getitem__(self, start):\n        indexes = self.indexs[(start*self.batch_size):((start+1)*self.batch_size)]\n\n        X, y = self.__data_gen(indexes)\n\n        return X,y\n<\/code><\/pre>\n\n<p>I based it on the code from 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><\/p>\n\n<p>The model and the rest of the code:<\/p>\n\n<pre><code>model = models.Sequential()\nmodel.add(TimeDistributed(Conv2D(32, (3,3), padding='same'), input_shape= \n(100,720,1280,3)))\nmodel.add(TimeDistributed(BatchNormalization()))\nmodel.add(TimeDistributed(Activation('relu')))\nmodel.add(TimeDistributed(MaxPooling2D((2,2),strides=(2,2))))\n#model.add(TimeDistributed(Conv2D(64, (3,3), padding='same')))\n#model.add(TimeDistributed(BatchNormalization()))\n#model.add(TimeDistributed(Activation('relu')))\n#model.add(TimeDistributed(Conv2D(128, (3,3), padding='same')))\n#model.add(TimeDistributed(BatchNormalization()))\n#model.add(TimeDistributed(Activation('relu')))\n#model.add(TimeDistributed(MaxPooling2D((2,2),strides=(2,2))))\nmodel.add(TimeDistributed(Flatten()))\nmodel.add(LSTM(256, dropout=0.5))\nmodel.add(Dense(38, activation='softmax'))\n\n\ntrain_gen = Generator(available_ids=train_ids, labels=train_labels)\nval_gen = Generator(available_ids=val_ids, labels=val_labels)\n\n\nmodel.compile(loss='categorical_crossentropy', \noptimizer=optimizers.RMSprop(lr=1e-4), metrics=['acc'])\nhistory = model.fit_generator(generator=train_gen, steps_per_epoch=len(train_ids), validation_data=val_gen, validation_steps = len(val_ids), epochs=100, verbose=0)\n<\/code><\/pre>\n\n<p>I've been searching around for the right way to do this for at least a week. The other post on this does not really answer my question. If someone could please help me with this that would be great.<\/p>\n","answers":[]}
+{"QuestionId":57601256,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><face-recognition>","CreationDate":"2019-08-22T02:19:11.313","AcceptedAnswerId":null,"OwnerUserId":8214242.0,"Title":"My Input pipeline with tf.data is converting images to negetives. How can I stop it?","Body":"<p>I am making a triplet producing tf.data pipeline and while testing each dataset, It is producing negative of images.<\/p>\n\n<p>I did already tried to normalize the image, but I don't know what I am doing wrong.<\/p>\n\n<p>All code is provided in <a href=\"https:\/\/gist.github.com\/lkit57a03\/251cc96b87f40a78b725522552043498\" rel=\"nofollow noreferrer\">this gist<\/a>.<\/p>\n\n<p>But here is the gist of what I am doing:<\/p>\n\n<pre><code>def _read_image_and_resize(self, same_img_1, same_img_2, diff_img_1): \n    target_size = [self.img_height, self.img_width]\n    # read images from disk\n    img1_file = tf.io.read_file(same_img_1)\n    img2_file = tf.io.read_file(same_img_2)\n    img3_file = tf.io.read_file(diff_img_1)\n\n    img1 = tf.image.decode_jpeg(img1_file, channels=3)\n    img2 = tf.image.decode_jpeg(img2_file, channels=3)\n    img3 = tf.image.decode_jpeg(img3_file, channels=3)\n\n    img1_resized = tf.image.resize_images(img1, target_size)\n    img2_resized = tf.image.resize_images(img2, target_size)\n    img3_resized = tf.image.resize_images(img3, target_size)\n\n    return img1, img2, img3\n<\/code><\/pre>\n\n<p>And here is the <a href=\"https:\/\/imgur.com\/FDdwf5T\" rel=\"nofollow noreferrer\">negative image<\/a> I get as well as the <a href=\"https:\/\/imgur.com\/FE8PZah\" rel=\"nofollow noreferrer\">positive result<\/a> I want.<\/p>\n","answers":[]}
+{"QuestionId":57601849,"AnswerCount":0,"Tags":"<python><keras><deep-learning><loss-function><imbalanced-data>","CreationDate":"2019-08-22T03:54:39.273","AcceptedAnswerId":null,"OwnerUserId":11960135.0,"Title":"An operation has `None` for gradient. Common ops without gradient: K.argmax, K.round, K.eval","Body":"<p>I am having a problem in the precision and recall report as their value is very less for the positive class around 15%. So to correct it I have written some custom code for loss-function which is nor correct and raising an error.<\/p>\n\n<p>def imbalancedloss(y_true, y_pred):<\/p>\n\n<pre><code>stacked =tensorflow.stack([y_true, y_pred])\ndef claculate_loss(y):\n    correct=y[0]\n    predict=y[0]\n    if correct== 1 and predict== 0 :\n        return 5\n    elif correct== 0 and predict== 1:\n        return 0.5\n    else:\n        return 0\n\n\nloss_array = tensorflow.keras.backend.map_fn(claculate_loss, stacked)\nreturn(tensorflow.keras.backend.sum(loss_array))\n<\/code><\/pre>\n\n<p>this will be passed in the Keras deep learning model while compiling the model for \"loss\"<\/p>\n\n<p>Error:An operation has <code>None<\/code> 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.<\/p>\n","answers":[]}
+{"QuestionId":57601988,"AnswerCount":0,"Tags":"<tensorflow><keras><tensorflow2.0>","CreationDate":"2019-08-22T04:18:54.340","AcceptedAnswerId":null,"OwnerUserId":3275163.0,"Title":"How to modify weights of Keras layers?","Body":"<p>I am trying to freeze some of the weights of a layer by setting them to a specific value in Keras. How can I achieve this without moving weights to CPU ?<\/p>\n\n<p>I checked similar questions such as <a href=\"https:\/\/stackoverflow.com\/questions\/45539595\/modify-layer-weights-in-keras\">modify layer weights in keras<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/52810702\/modify-layer-parameters-in-keras\">modify layer parameters in keras<\/a><\/p>\n\n<p>Answers suggest usage of <code>get_weights()<\/code> and 'set_weights()', however those functions moves weights between CPU and GPU.<\/p>\n\n<p>I created a custom lambda layer and modified model.trainable_weights inside of that layer, however weights are not updated.<\/p>\n\n<p>I used tf advanced tutorial, and just added a custom lambda layer that multiplies weights with zero. \n<a href=\"https:\/\/colab.research.google.com\/drive\/1t7mhictbbJj8Jumx3XY-1ndJIoge5mdQ\" rel=\"nofollow noreferrer\">Colab notebook with same code<\/a><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from __future__ import absolute_import, division, print_function, unicode_literals\n\nimport tensorflow as tf\n\nfrom tensorflow.keras.layers import Dense, Flatten, Conv2D, Lambda\nfrom tensorflow.keras import Model\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\n# Add a channels dimension\nx_train = x_train[..., tf.newaxis]\nx_test = x_test[..., tf.newaxis]\n\ndef antirectifier(x):\n    for i,w in enumerate(model.trainable_weights):\n        model.trainable_weights[i] = tf.multiply(w,0)\n\n    return x\n\nclass MyModel(Model):\n  def __init__(self):\n    super(MyModel, self).__init__()\n    self.conv1 = Conv2D(32, 3, activation='relu')\n    self.flatten = Flatten()\n    self.d1 = Dense(128, activation='relu')\n    self.d2 = Dense(10, activation='softmax')\n    self.mask = Lambda(antirectifier)\n\n  def call(self, x):\n    x = self.conv1(x)\n    x = self.flatten(x)\n    x = self.d1(x)\n    x = self.mask(x)\n\n    return self.d2(x)\n\n# Create an instance of the model\nmodel = MyModel()\n\nloss_object = tf.keras.losses.SparseCategoricalCrossentropy()\n\noptimizer = tf.keras.optimizers.Adam()\n\ntrain_loss = tf.keras.metrics.Mean(name='train_loss')\ntrain_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='train_accuracy')\n\ntest_loss = tf.keras.metrics.Mean(name='test_loss')\ntest_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='test_accuracy')\n\n@tf.function\ndef train_step(images, labels):\n  with tf.GradientTape() as tape:\n    predictions = model(images)\n    loss = loss_object(labels, predictions)\n  gradients = tape.gradient(loss, model.trainable_variables)\n  optimizer.apply_gradients(zip(gradients, model.trainable_variables))\n\n  train_loss(loss)\n  train_accuracy(labels, predictions)\n\n@tf.function\ndef test_step(images, labels):\n  predictions = model(images)\n  t_loss = loss_object(labels, predictions)\n\n  test_loss(t_loss)\n  test_accuracy(labels, predictions)\n\nEPOCHS = 5\n\nfor epoch in range(EPOCHS):\n  for images, labels in train_ds:\n    train_step(images, labels)\n\n  for test_images, test_labels in test_ds:\n    test_step(test_images, test_labels)\n\n  template = 'Epoch {}, Loss: {}, Accuracy: {}, Test Loss: {}, Test Accuracy: {}'\n  print(template.format(epoch+1,\n                        train_loss.result(),\n                        train_accuracy.result()*100,\n                        test_loss.result(),\n                        test_accuracy.result()*100))\n\n  # Reset the metrics for the next epoch\n  train_loss.reset_states()\n  train_accuracy.reset_states()\n  test_loss.reset_states()\n  test_accuracy.reset_states()\n<\/code><\/pre>\n\n<p>Since weights are zero, accuracy should be low. However weights are not changed.<\/p>\n","answers":[]}
+{"QuestionId":57602552,"AnswerCount":1,"Tags":"<conv-neural-network><caffe><mobilenet>","CreationDate":"2019-08-22T05:26:39.637","AcceptedAnswerId":57765912.0,"OwnerUserId":4417582.0,"Title":"How to do incremental learning on MobileNet-SSD caffe","Body":"<p>I'm training my classifier on <em>20k<\/em> images but every week I get more new pictures so I want to incrementally train my previous model(last stopped iteration) instead of retraining <em>20k+new_images<\/em> on all the images again which is a waste of time and compute<\/p>\n\n<p>I figured out incremental training with Yolo but can't seem to find anything for MobileNet-SSD caffe implemented here <a href=\"https:\/\/github.com\/chuanqi305\/MobileNet-SSD\" rel=\"nofollow noreferrer\">https:\/\/github.com\/chuanqi305\/MobileNet-SSD<\/a><\/p>\n\n<p>To understand more about what I'm talking about refer to this: \n<a href=\"https:\/\/github.com\/pjreddie\/darknet\/issues\/705\" rel=\"nofollow noreferrer\">How to do incremental training on the basis of yolov3.weights<\/a> &amp; answer to this mention <a href=\"https:\/\/github.com\/AlexeyAB\/darknet\/issues\/829#issuecomment-389831486\" rel=\"nofollow noreferrer\">here<\/a>: <\/p>\n\n<pre><code> darknet.exe partial cfg\/yolov3.cfg yolov3.weights yolov3.conv.105 105 \n<\/code><\/pre>\n","answers":[{"AnswerId":"57765912","CreationDate":"2019-09-03T06:12:07.507","ParentId":null,"OwnerUserId":"4417582","Title":null,"Body":"<p>You need to pass previous iteration in <code>train.sh<\/code> instead of 73000 iteration. The new iteration are found in snapshot folder once you are done training <\/p>\n\n<pre><code>if ! test -f example\/MobileNetSSD_train.prototxt ;then\necho \"error: example\/MobileNetSSD_train.prototxt does not exist.\"\necho \"please use the gen_model.sh to generate your own model.\"\nexit 1\nfi\nmkdir -p snapshot\n\n#Initiate a new training\n$CAFFE_ROOT\/build\/tools\/caffe train -solver=\"solver_train.prototxt\" \\\n-weights=\"mobilenet_iter_73000.caffemodel\" \\\n-gpu 0\n<\/code><\/pre>\n"}]}
+{"QuestionId":57602574,"AnswerCount":0,"Tags":"<pytorch><libtorch><torchscript>","CreationDate":"2019-08-22T05:29:54.267","AcceptedAnswerId":null,"OwnerUserId":4708547.0,"Title":"Pytorch (torchscript) error : attribute 'downsample' of type 'NoneType' is not usable in a script method (did you forget to add it __constants__?)","Body":"<p>I am converting pytorch code for AlignedReID++ to C++ code using torchscript. During the process I encountered the following error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"selfDemo.py\", line 28, in &lt;module&gt;\n    traced_model = torch.jit.script(model)\n  File \"\/nfs\/engine\/rajagopalar\/anaconda3\/envs\/torchExp\/lib\/python3.6\/site-packages\/torch\/jit\/__init__.py\", line 1162, in script\n    return _convert_to_script_module(obj)\n  File \"\/nfs\/engine\/rajagopalar\/anaconda3\/envs\/torchExp\/lib\/python3.6\/site-packages\/torch\/jit\/__init__.py\", line 1812, in _convert_to_script_module\n    return WeakScriptModuleProxy(mod, stubs)\n  File \"\/nfs\/engine\/rajagopalar\/anaconda3\/envs\/torchExp\/lib\/python3.6\/site-packages\/torch\/jit\/__init__.py\", line 1386, in init_then_register\n    original_init(self, *args, **kwargs)\n  File \"\/nfs\/engine\/rajagopalar\/anaconda3\/envs\/torchExp\/lib\/python3.6\/site-packages\/torch\/jit\/__init__.py\", line 1736, in __init__\n    _create_methods_from_stubs(self, stubs)\n  File \"\/nfs\/engine\/rajagopalar\/anaconda3\/envs\/torchExp\/lib\/python3.6\/site-packages\/torch\/jit\/__init__.py\", line 1347, in _create_methods_from_stubs\n    self._c._create_methods(self, defs, rcbs, defaults)\n  File \"\/nfs\/engine\/rajagopalar\/anaconda3\/envs\/torchExp\/lib\/python3.6\/site-packages\/torch\/jit\/__init__.py\", line 982, in _make_strong_submodule\n    new_strong_submodule = _convert_to_script_module(module)\n  File \"\/nfs\/engine\/rajagopalar\/anaconda3\/envs\/torchExp\/lib\/python3.6\/site-packages\/torch\/jit\/__init__.py\", line 1791, in _convert_to_script_module\n    return _create_constant_iterable_module(mod)\n  File \"\/nfs\/engine\/rajagopalar\/anaconda3\/envs\/torchExp\/lib\/python3.6\/site-packages\/torch\/jit\/__init__.py\", line 963, in _create_constant_iterable_module\n    modules[key] = _create_constant_iterable_module(submodule)\n  File \"\/nfs\/engine\/rajagopalar\/anaconda3\/envs\/torchExp\/lib\/python3.6\/site-packages\/torch\/jit\/__init__.py\", line 965, in _create_constant_iterable_module\n    modules[key] = _convert_to_script_module(submodule)\n  File \"\/nfs\/engine\/rajagopalar\/anaconda3\/envs\/torchExp\/lib\/python3.6\/site-packages\/torch\/jit\/__init__.py\", line 1812, in _convert_to_script_module\n    return WeakScriptModuleProxy(mod, stubs)\n  File \"\/nfs\/engine\/rajagopalar\/anaconda3\/envs\/torchExp\/lib\/python3.6\/site-packages\/torch\/jit\/__init__.py\", line 1386, in init_then_register\n    original_init(self, *args, **kwargs)\n  File \"\/nfs\/engine\/rajagopalar\/anaconda3\/envs\/torchExp\/lib\/python3.6\/site-packages\/torch\/jit\/__init__.py\", line 1736, in __init__\n    _create_methods_from_stubs(self, stubs)\n  File \"\/nfs\/engine\/rajagopalar\/anaconda3\/envs\/torchExp\/lib\/python3.6\/site-packages\/torch\/jit\/__init__.py\", line 1347, in _create_methods_from_stubs\n    self._c._create_methods(self, defs, rcbs, defaults)\nRuntimeError: \nattribute 'downsample' of type 'NoneType' is not usable in a script method (did you forget to add it __constants__?):\nat \/nfs\/engine\/rajagopalar\/anaconda3\/envs\/torchExp\/lib\/python3.6\/site-packages\/torchvision\/models\/resnet.py:109:12\n        out = self.relu(out)\n\n        out = self.conv2(out)\n        out = self.bn2(out)\n        out = self.relu(out)\n\n        out = self.conv3(out)\n        out = self.bn3(out)\n\n        if self.downsample is not None:\n           ~~~~~~~~~~~~~~~ &lt;--- HERE\n            identity = self.downsample(x)\n        out += identity\n        out = self.relu(out)\n\n        return out\n'__module__.__torch__.torchvision.models.resnet.___torch_mangle_32.Bottleneck.forward' is being compiled since it was called from '__module__.__torch__.models.ResNet.ResNet50.forward'\nat \/nfs\/engine\/rajagopalar\/anaconda3\/envs\/torchExp\/lib\/python3.6\/site-packages\/torchvision\/models\/resnet.py:114:16\n\n        out = self.conv3(out)\n        out = self.bn3(out)\n\n        if self.downsample is not None:\n            identity = self.downsample(x)\n        out += identity\n        out = self.relu(out)\n\n        return out\n               ~~~ &lt;--- HERE\n<\/code><\/pre>\n\n<p>Further, the error comes from the torchvision Resnet model and not my code which confused me as to where the error was<\/p>\n\n<p>I am running the code with pytorch==1.2. I have tried using Optional[T] tags but it did not work as expected still<\/p>\n\n<p>I want to know if there's anything wrong on my end or must I create more files for scripting to happen.<\/p>\n\n<p>Any help or suggestions would be welcome<\/p>\n\n<p>The first model<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>traced_model = torch.jit.script(model)\n\n<\/code><\/pre>\n\n<p>And the second model<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>myexactor = FeatureExtractor(traced_model, exact_list)\n\n...\nscripted_trace = torch.jit.script(myexactor, img1)\n<\/code><\/pre>\n\n<p>The code is taken from github : <a href=\"https:\/\/github.com\/michuanhaohao\/AlignedReID\" rel=\"nofollow noreferrer\">https:\/\/github.com\/michuanhaohao\/AlignedReID<\/a><\/p>\n\n<p>and I made changes in the file :\n<a href=\"https:\/\/github.com\/michuanhaohao\/AlignedReID\/blob\/master\/Alignedreid_demo.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/michuanhaohao\/AlignedReID\/blob\/master\/Alignedreid_demo.py<\/a><\/p>\n","answers":[]}
+{"QuestionId":57602726,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2019-08-22T05:46:52.107","AcceptedAnswerId":57603661.0,"OwnerUserId":7850499.0,"Title":"Why can tensorflow use multithread while python can only execute one thread at a time due to GIL?","Body":"<p>I'm wondering why we can have Tensorflow <a href=\"https:\/\/www.tensorflow.org\/guide\/performance\/overview#optimizing_for_cpu\" rel=\"nofollow noreferrer\">run in a multi-thread fashion<\/a> while python can only execute one thread at a time due to <a href=\"https:\/\/wiki.python.org\/moin\/GlobalInterpreterLock\" rel=\"nofollow noreferrer\">GIL<\/a>?<\/p>\n","answers":[{"AnswerId":"57603661","CreationDate":"2019-08-22T06:59:42.757","ParentId":null,"OwnerUserId":"1219295","Title":null,"Body":"<p>The GIL's restriction is slightly more subtle: only one thread at a time can be <em>executing Python bytecode<\/em>.<\/p>\n\n<p>Extensions using Python's C API (like <code>tensorflow<\/code>) can release the GIL if they don't need it. I\/O operations like using files or sockets also tend to release the GIL because they generally involve lots of waiting.<\/p>\n\n<p>So threads executing extensions or waiting for I\/O can run while another thread is executing Python bytecode.<\/p>\n"},{"AnswerId":"57603359","CreationDate":"2019-08-22T06:40:25.063","ParentId":null,"OwnerUserId":"236106","Title":null,"Body":"<p>Most of the <code>tensorflow<\/code> core is written in <code>C++<\/code> and the python APIs are just the wrappers around it. While running the <code>C++<\/code> code regular python restrictions do not apply. <\/p>\n"}]}
+{"QuestionId":57603084,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><artificial-intelligence><linear-regression>","CreationDate":"2019-08-22T06:20:59.210","AcceptedAnswerId":57930535.0,"OwnerUserId":5581893.0,"Title":"TensorFlow estimator DNNClassifier fails to do 2-time linear separation","Body":"<p>I'm trying to test out TensorFlow tf.estimator.<strong>DNNClassifier<\/strong> with some simple data<\/p>\n\n<pre><code>X = [[1,2], [1,12], [1,17], [9,33], [48,49], [48,50]]\nY = [ 1,     1,      1,      1,      2,       3     ]\n<\/code><\/pre>\n\n<p>The classifier takes 2 inputs: <strong>x1,x2<\/strong>; and is having this shape:<\/p>\n\n<pre><code>#these 4 layers supposed to be able to do even 4-time linear separation\nhidden_units = [2000,1000,500,100]\nn_classes    = 4    \n<\/code><\/pre>\n\n<p>Hower, things didn't go as wanted, the network couldn't fit. Accuracy quickly got to <strong>8.33<\/strong> (=5\/6) but stuck then. Loss converged to a horizontal line but not the zero line.<\/p>\n\n<p>The data provided above are 2-time linear separable (right-click image to open in new tab):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/XT3OJ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/XT3OJ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Even when the network runs to 10,000 steps, it's still stuck, I guess it's stuck because it fails to separate the 2 values: Y=2 and Y=3, is it so? And how to make the network fit with the mentioned data.<\/p>\n","answers":[{"AnswerId":"57930535","CreationDate":"2019-09-13T21:26:44.870","ParentId":null,"OwnerUserId":"5581893","Title":null,"Body":"<p>I tried to use these and it works:<\/p>\n\n<ul>\n<li>Use leaky relu or sigmoid<\/li>\n<li>Reduce number of layers<\/li>\n<li>Reduce number of neurons in each layer<\/li>\n<li>Normalise inputs<\/li>\n<li>Increase number of steps, epochs if needed<\/li>\n<\/ul>\n"}]}
+{"QuestionId":57603707,"AnswerCount":1,"Tags":"<python><tensorflow><gpu><reinforcement-learning>","CreationDate":"2019-08-22T07:02:00.713","AcceptedAnswerId":57605907.0,"OwnerUserId":11805922.0,"Title":"How to make this RL code get GPU support?","Body":"<p><a href=\"https:\/\/github.com\/keon\/deep-q-learning\/blob\/master\/dqn.py#L52\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keon\/deep-q-learning\/blob\/master\/dqn.py#L52<\/a><\/p>\n\n<pre><code>def replay(self, batch_size):\n    minibatch = random.sample(self.memory, batch_size)\n    for state, action, reward, next_state, done in minibatch:\n        target = reward\n        if not done:\n            target = (reward + self.gamma *\n                      np.amax(self.model.predict(next_state)[0]))\n        target_f = self.model.predict(state)\n        target_f[0][action] = target\n        self.model.fit(state, target_f, epochs=1, verbose=0)\n<\/code><\/pre>\n\n<p>This code seems can't get help from GPU since it trains data once per action.<\/p>\n\n<pre><code>self.model.fit(state, target_f, epochs=1, verbose=0)\n<\/code><\/pre>\n\n<p>How to change this code to train in parallel, then get help from GPU?<\/p>\n","answers":[{"AnswerId":"57605907","CreationDate":"2019-08-22T09:15:25.730","ParentId":null,"OwnerUserId":"10455493","Title":null,"Body":"<p>In Reinforcement Learning (RL) there is often a lot of CPU computation required for each sample step (of course dependent on environment, some environments can use GPU too). The RL-model has a hard time understanding the rewards and what action caused that specific reward, since a good reward could be dependent on a way earlier action. Therefore we want a simple model-architectures (shallow and fewer weights) while doing RL, else the training time will be way to slow. Hence your systems bottle neck is likely gathering samples rather than training the data. \nAlso Note that not all Tensorflow-architectures scale equally well with GPU. Deep models with high numbers of weights like most Images cases scales super well (like CNN and MLP network with MNIST), while time-dependent RNN has less speedup potential (see <a href=\"https:\/\/ai.stackexchange.com\/questions\/7090\/can-lstm-nets-be-speed-up-by-gpu\">this stackexchange question<\/a>). So set your expectation accordingly when using GPU.<\/p>\n\n<p>Through my RL experience, I have figured some possible speedups I could share, and would love to see more suggestions!<\/p>\n\n<ol>\n<li>Single sample step, can be speed up by creating multiple environment run in parallel, equal to the number of CPU cores (there are packages for parallel processing in python you can use fore this). This can potential speed up sampling data proportional to the number of CPU cores. <\/li>\n<li><p>Between sampling you have to do model predictions for next action. Instead of calling model.predict at each step, you can call a single model.predict for all your parallel states (using a batch_size equal to the number of parallel environments). This will speed up prediction time, as there is more optimization options.<\/p><\/li>\n<li><p>The change from updating model weights to prediction is surprisingly slow. Hopefully this will be speed up in the future? But while the change is as slow as today, you can speed up training by holding the model constant and do lots of sample and prediction (example a whole episode, or multiple steps within an episode), then train the model on all the newly gathered data afterwards. In my case this resulted in periodically high GPU utilization.<\/p><\/li>\n<li><p>Since sampling is most likely the bottle neck, you can make a historical repo of state, action, rewards. Than at training you can sample randomly data from this repo and train it together with the newly gathered data. This is known as \"Experience Replay\" in RL.<\/p><\/li>\n<li><p>Maybe the most fun, and highest potential for improvements is by using more advance RL-learning architectures. Example changing the loss function (check out <a href=\"https:\/\/arxiv.org\/pdf\/1707.06347.pdf\" rel=\"nofollow noreferrer\">PPO<\/a> for example), using and tuning the \"generalized advantage estimation\" calculated by the rewards. Or changing the model by for example including time dependencies with RNN, <a href=\"http:\/\/kvfrans.com\/variational-autoencoders-explained\/\" rel=\"nofollow noreferrer\">VAC<\/a> or combining them all like <a href=\"https:\/\/worldmodels.github.io\" rel=\"nofollow noreferrer\">here<\/a>. <\/p><\/li>\n<\/ol>\n\n<p>Hopefully this help you speed up the training time, and maybe get more utilization of your GPU. <\/p>\n"}]}
+{"QuestionId":57604124,"AnswerCount":0,"Tags":"<python><keras><conv-neural-network>","CreationDate":"2019-08-22T07:28:40.207","AcceptedAnswerId":null,"OwnerUserId":11020443.0,"Title":"Can I use class_weights to completely ignore a class","Body":"<p>I am fairly aware of how to use the <code>class_weights<\/code> argument in the Keras fit method, i.e. if there are unbalanced classes then assign weights to counteract, however, can I use this to completely ignore a class?<\/p>\n\n<p>I am using a 1D convolutional network (basically a stack of 1DConv layers outputting to a dense softmax layer) to analyze sequential data. Included in the sequences are lots of <strong>padding<\/strong> steps which I am including as features but would like to ignore when it comes to output. Can I pass an argument like:<\/p>\n\n<pre><code>class_weights = {'action1': 1.0, 'action2': 1.0, 'pad': 0.0}\n<\/code><\/pre>\n\n<p>I am hoping that this will have the result of completely ignoring loss for the pad class during optimization, resulting in a probability output where <strong>pad<\/strong> is always 0 for the next step. Thoughts?<\/p>\n","answers":[]}
+{"QuestionId":57604157,"AnswerCount":1,"Tags":"<pytorch><amazon-sagemaker>","CreationDate":"2019-08-22T07:30:34.250","AcceptedAnswerId":null,"OwnerUserId":9584727.0,"Title":"Difference between PyTorch, PyTorchModel in sagemaker.pytorch","Body":"<p>I am trying to create a model using pytorch in sagemaker. I tried deploying using - PyTorch module in sagemaker.pytorch [from sagemaker.pytorch import PyTorch].<\/p>\n\n<p>But, I want to understand what is PyTorchModel in sagemaker.pytorch [from sagemaker.pytorch import PyTorchModel]. They both have <code>deploy()<\/code> . And I followed the link <a href=\"https:\/\/sagemaker.readthedocs.io\/en\/stable\/using_pytorch.html\" rel=\"nofollow noreferrer\">https:\/\/sagemaker.readthedocs.io\/en\/stable\/using_pytorch.html<\/a> to create and deploy the model . Where I don't see using \"PyTorchModel\" anywhere. I would like to know the difference and when to use what.<\/p>\n\n<p>I tried the following so far.<\/p>\n\n<p><strong>Step 1 :<\/strong> I called pytorch estimator <\/p>\n\n<pre><code>pytorch_model = PyTorch(entry_point='entry_v1.py',\n                        train_instance_type='ml.m5.4xlarge',\n                        role = role,\n                        train_instance_count=1,\n                        output_path = \"s3:\/\/model-output-bucket\/test\",\n                        framework_version='1.1',\n                        hyperparameters = {'epochs': 10,'learning-rate': 0.01})\n<\/code><\/pre>\n\n<p><strong>Step2  :<\/strong> I called fit method<\/p>\n\n<pre><code>pytorch_model.fit({'train': 's3:\/\/training-data\/train_data.csv',\n                  'test':'s3:\/\/testing-data\/test_data.csv'})\n<\/code><\/pre>\n\n<p><strong>Step3:<\/strong> I called deploy method.<\/p>\n\n<pre><code>predictor = pytorch_model.deploy(instance_type='ml.m4.xlarge', initial_instance_count=1)\n<\/code><\/pre>\n\n<p>I would like to know when to call <code>create_model()<\/code> here.  <\/p>\n\n<p><strong>I got some understanding over here.<\/strong> We use [from sagemaker.pytorch import PyTorch] for end to end process where we train a model with .fit() and then we can deploy the model with .deploy()<\/p>\n\n<p>But, with [from sagemaker.pytorch import PyTorchModel] we can use the model which is already trained. <\/p>\n\n<p><strong>Step1:<\/strong> <\/p>\n\n<pre><code>pytorch_model = PyTorchModel(model_data='s3:\/\/model-output-bucket\/sagemaker-pytorch-2019-08-20-16-54-32-500\/output\/model.tar.gz', role=role,entry_point=entry_v1.py,sagemaker_session=sagemaker_session)\n<\/code><\/pre>\n\n<p><strong>Step2:<\/strong><\/p>\n\n<pre><code>predictor = pytorch_model.deploy(instance_type='ml.c4.xlarge', initial_instance_count=1)\n<\/code><\/pre>\n\n<p>Also, .create_model() of PyTorch Estimator will return an object of PyTorchModel.<\/p>\n\n<p>Please correct me if i am wrong anywhere.<\/p>\n","answers":[{"AnswerId":"57612380","CreationDate":"2019-08-22T15:08:38.150","ParentId":null,"OwnerUserId":"4592183","Title":null,"Body":"<p>PyTorch class is inherited from the Framework class whereas PyTorchModel is inherited from the FrameworkModel class. <\/p>\n\n<p>The difference between these two is that:\nFramework is used to perform the end to end training and deployment of a model\nFrameworkModel is used to create an Estimator from a pretrained model and then use it to deploy an endpoint using the deploy() method. This does not involve training of the model.<\/p>\n\n<p>In the PyTorch class, you cannot directly call the deploy model. You'll first have to invoke the fit() method which will be followed by the deploy() method.<\/p>\n\n<p>You can read the following blog on how you can bring your own Pretrained model to Sagemaker\n<a href=\"https:\/\/aws.amazon.com\/blogs\/machine-learning\/bring-your-own-pre-trained-mxnet-or-tensorflow-models-into-amazon-sagemaker\/\" rel=\"nofollow noreferrer\">https:\/\/aws.amazon.com\/blogs\/machine-learning\/bring-your-own-pre-trained-mxnet-or-tensorflow-models-into-amazon-sagemaker\/<\/a><\/p>\n\n<p>Regarding the create_model() method, you need not invoke it in your script if you'd like to directly deploy an endpoint after training. \nIt is generally used in scenarios where you require to create a pipeline for inferences through multiple models<\/p>\n"}]}
+{"QuestionId":57604324,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><text-classification><tensorflow-datasets>","CreationDate":"2019-08-22T07:40:13.803","AcceptedAnswerId":null,"OwnerUserId":8842694.0,"Title":"How to access data from text file for training a model on TPU?","Body":"<p>I am trying to implement a LSTM text classification model in TF1.x for TPU without enabling eager execution. However, I just can not find a correct way to prepare the data for feeding it to the model. I am following these two tutorials, <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/load_data\/text#split_the_dataset_into_text_and_train_batches\" rel=\"nofollow noreferrer\">tutorial 1<\/a> for getting dataset and <a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/tpu\/blob\/master\/tools\/colab\/classification_iris_data_with_keras.ipynb\" rel=\"nofollow noreferrer\">tutorial 2<\/a> for implementing it for running on Google Colab using TPU.\nSince it is text classification, I have to encode the input text tokens using <code>tf.Tokenizer()<\/code> below is how I am doing it-<\/p>\n\n<pre><code>tokenizer = tfds.features.text.Tokenizer()\nvocabulary_set = set()\n\nmy_iterator = all_labeled_data.make_initializable_iterator()\ntext_tensor, _ = my_iterator.get_next()\nwith tf.Session() as sess1: \n  sess1.run(my_iterator.initializer)\n  try:\n    while True:\n      text_string = sess1.run(text_tensor)\n      #print text_string\n      some_tokens = tokenizer.tokenize(text_string)\n      vocabulary_set.update(some_tokens)\n  except tf.errors.OutOfRangeError:\n    pass\n<\/code><\/pre>\n\n<p><code>all_labeled_data<\/code> consists (example, label) pairs. I am sure I am making a mistake the way I am accessing the data. If you have any comments\/suggestions, please feel free to do so. If you want to see the whole code then it is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/31832\" rel=\"nofollow noreferrer\">here<\/a>. Also, feel free to run it on Google Colab with <strong>Python 2<\/strong> and <strong>TPU<\/strong> enabled.\nThank you!<\/p>\n","answers":[]}
+{"QuestionId":57605021,"AnswerCount":2,"Tags":"<python><numpy><tensorflow><keras><scikit-image>","CreationDate":"2019-08-22T08:22:51.030","AcceptedAnswerId":null,"OwnerUserId":4040643.0,"Title":"Including advanced computation (scikit-like) in a keras custom layer","Body":"<p>Normally I would preprocess the data before I feed it into my model for classification. <\/p>\n\n<p>This is however not possible and thus am stuck either to enhance the performance of the model further (somehow) or include useful preprocessing steps directly inside the model. <\/p>\n\n<p>How can I do that? The best solution I found thus far, included re-implementing the functionality I want using Keras backend. This is far from a good solution and thus I am hoping someone has an idea, how to salavage the situation. <\/p>\n\n<p>Below are links I found useful + my current code.<\/p>\n\n<p>Useful links: <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/54744944\/keras-custom-layer-with-advanced-calculations\">Keras Custom Layer with advanced calculations<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/47577060\/how-to-switch-from-keras-tensortype-to-numpy-array-for-a-custom-layer?noredirect=1&amp;lq=1\">How to Switch from Keras Tensortype to numpy array for a custom layer?<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/54907018\/how-to-create-a-keras-custom-layer-using-functions-not-included-in-the-backend?noredirect=1&amp;lq=1\">How to create a Keras Custom Layer using functions not included in the Backend, to perform tensor sampling?<\/a><\/p>\n\n<p>My code thus far:<\/p>\n\n<pre><code>def freezeBaseModelLayers(baseModel):\n    for layer in baseModel.layers:\n        layer.trainable = False\n\n\ndef preprocess_input(x):\n    # TODO: Not working, but intention should be clear\n    numpy_array = tf.unstack(tf.unstack(tf.unstack(x, 224, 0), 224, 0), 1, 0)\n    from skimage.feature import hog\n    from skimage import data, exposure\n    img_adapteq = exposure.equalize_adapthist(numpy_array, orientations=8, pixels_per_cell=(3, 3),\n                                              cells_per_block=(1, 1), visualize=True, multichannel=False)\n    [x1, x2, x3] = tf.constant(img_adapteq), tf.constant(img_adapteq), tf.constant(img_adapteq)\n    img_conc = Concatenate([x1, x2, x3])\n    return img_conc\n\n\ndef create(x):\n    is_training = tf.get_variable('is_training', (), dtype=tf.bool, trainable=False)\n    with tf.name_scope('pretrained'):\n        # Add preprocess step here...\n        input_layer = Lambda(preprocess_input(x), input_shape=(224, 224, 1), output_shape=(224, 224, 3))\n\n        baseModel = vgg16.VGG16(weights='imagenet', include_top=False, input_shape=(224, 224, 3))    \n        freezeBaseModelLayers(baseModel)\n        layer = baseModel(input_layer)\n        layer = GlobalMaxPooling2D()(layer)\n        layer = Dense(1024, activation='relu')(layer)\n        layer = Dense(2, activation=None)(layer)\n        model = Model(input=input_layer.input, output=layer)\n        output = model(x)\n        return output\n\n\n\n\nI would like to include prepocessing steps inside my model\n\nThe models I am working with are receiving noisy data. In order to enhance the performance of the models, I would like to do some preprocessing steps e.g. equalize_adapthist.\n<\/code><\/pre>\n","answers":[{"AnswerId":"57613708","CreationDate":"2019-08-22T16:35:23.440","ParentId":null,"OwnerUserId":"4332585","Title":null,"Body":"<p>To do this with a <code>Lambda<\/code> layer, you would need to write histogram equalization in pure <code>tensorflow<\/code>.\nIndeed, when building the graph, the function (<code>preprocess_input<\/code>) will be called on <code>tensorflow<\/code> placeholders which, in the case of this <code>skimage<\/code> function which expects <code>numpy<\/code> arrays, will not work.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/42835247\/how-to-implement-histogram-equalization-for-images-in-tensorflow\">This question<\/a> shows how to write it in pure <code>tensorflow<\/code>. Copy-pasting here for the sake of redundancy\/ease-of-reading (I have not tested it myself but a test is available in the question):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def tf_equalize_histogram(image):\n    values_range = tf.constant([0., 255.], dtype = tf.float32)\n    histogram = tf.histogram_fixed_width(tf.to_float(image), values_range, 256)\n    cdf = tf.cumsum(histogram)\n    cdf_min = cdf[tf.reduce_min(tf.where(tf.greater(cdf, 0)))]\n\n    img_shape = tf.shape(image)\n    pix_cnt = img_shape[-3] * img_shape[-2]\n    px_map = tf.round(tf.to_float(cdf - cdf_min) * 255. \/ tf.to_float(pix_cnt - 1))\n    px_map = tf.cast(px_map, tf.uint8)\n\n    eq_hist = tf.expand_dims(tf.gather_nd(px_map, tf.cast(image, tf.int32)), 2)\n    return eq_hist\n<\/code><\/pre>\n\n<p>Btw you should write the pre-processing step as (once you made the preprocessing step pure <code>tensorflow<\/code>):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>input_layer = Lambda(preprocess_input, input_shape=(224, 224, 1), output_shape=(224, 224, 3))(x)\n<\/code><\/pre>\n\n<p>Another way to do this is to write a custom layer as pointed out by mlRocks.<\/p>\n"},{"AnswerId":"57614904","CreationDate":"2019-08-22T18:06:36.060","ParentId":null,"OwnerUserId":"8382707","Title":null,"Body":"<p>A better way to do this is via a custom keras layer. Here is an example:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom keras.layers import Layer, Input, Conv2D\nfrom keras.models import Model\nfrom keras import backend as K\nfrom skimage.feature import hog\nfrom skimage import data, exposure\n\ndef equalize(img):\n  img_adapteq = exposure.equalize_adapthist(img)\n  return img_adapteq\n\ndef preprocess_input(img):\n  return tf.py_func(equalize, \n                     [img],\n                     'float32',\n                     stateful=False,\n                     name='custom_image_op')\n\n\nclass CustomLayer(Layer):\n  def __init__(self, output_dim, **kwargs):\n    self.output_dim = output_dim\n    self.trainable = False\n    super(CustomLayer, self).__init__(**kwargs)\n\n  def call(self, x):\n    res = tf.map_fn(preprocess_input, x)\n    res.set_shape([x.shape[0],\n                   self.output_dim[1], \n                   self.output_dim[0],\n                   x.shape[-1]])\n    return res\n\noutput_dim = (224,224)\ninputs = Input(shape=(224,224,3))\nx = CustomLayer(output_dim)(inputs)\nx = Conv2D(32, (3,3))(x)\nx = Flatten()(x)\nx = Dense(1)(x)\n\nmodel = Model(inputs, x)\nmodel.summary()\n\n# test\nsample = np.random.rand(4, 224,224,3).astype(np.float32)\ny = np.random.randint(2, size=(4,))\n\nmodel.compile(\"sgd\", \"mse\")\nmodel.fit(sample, y)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57605094,"AnswerCount":1,"Tags":"<pytorch><vgg-net>","CreationDate":"2019-08-22T08:27:53.693","AcceptedAnswerId":57607939.0,"OwnerUserId":6187009.0,"Title":"The training loss of vgg16 implemented in pytorch does not decrease","Body":"<p>I want to try some toy examples in pytorch, but the training loss does not decrease in the training. <\/p>\n\n<p>Some info is provided here:<\/p>\n\n<ul>\n<li>The model is vgg16, consisted of 13 conv layers and 3 dense layers. <\/li>\n<li>The data is cifar100 in pytorch. <\/li>\n<li>I choose cross entropy as the loss function.<\/li>\n<\/ul>\n\n<p>The code is as follows <\/p>\n\n<pre><code># encoding: utf-8\n\nimport torch\nimport torch.optim as optim\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torchvision.transforms as transforms\nimport torchvision\n\nimport numpy as np\n\n\nclass VGG16(torch.nn.Module):\n    def __init__(self, n_classes):\n        super(VGG16, self).__init__()\n\n        # construct model\n        self.conv1_1 = nn.Conv2d(3, 64, 3, padding=1)\n        self.conv1_2 = nn.Conv2d(64, 64, 3, padding=1)\n        self.conv2_1 = nn.Conv2d(64, 128, 3, padding=1)\n        self.conv2_2 = nn.Conv2d(128, 128, 3, padding=1)\n        self.conv3_1 = nn.Conv2d(128, 256, 3, padding=1)\n        self.conv3_2 = nn.Conv2d(256, 256, 3, padding=1)\n        self.conv3_3 = nn.Conv2d(256, 256, 3, padding=1)\n        self.conv4_1 = nn.Conv2d(256, 512, 3, padding=1)\n        self.conv4_2 = nn.Conv2d(512, 512, 3, padding=1)\n        self.conv4_3 = nn.Conv2d(512, 512, 3, padding=1)\n        self.conv5_1 = nn.Conv2d(512, 512, 3, padding=1)\n        self.conv5_2 = nn.Conv2d(512, 512, 3, padding=1)\n        self.conv5_3 = nn.Conv2d(512, 512, 3, padding=1)\n\n        self.fc6 = nn.Linear(512, 512)\n        self.fc7 = nn.Linear(512, 512)\n        self.fc8 = nn.Linear(512, n_classes)\n\n    def forward(self, x):\n        x = F.relu(self.conv1_1(x))\n        x = F.relu(self.conv1_2(x))\n        x = F.max_pool2d(x, (2, 2))\n\n        x = F.relu(self.conv2_1(x))\n        x = F.relu(self.conv2_2(x))\n        x = F.max_pool2d(x, (2, 2))\n\n        x = F.relu(self.conv3_1(x))\n        x = F.relu(self.conv3_2(x))\n        x = F.relu(self.conv3_3(x))\n        x = F.max_pool2d(x, (2, 2))\n\n        x = F.relu(self.conv4_1(x))\n        x = F.relu(self.conv4_2(x))\n        x = F.relu(self.conv4_3(x))\n        x = F.max_pool2d(x, (2, 2))\n\n        x = F.relu(self.conv5_1(x))\n        x = F.relu(self.conv5_2(x))\n        x = F.relu(self.conv5_3(x))\n        x = F.max_pool2d(x, (2, 2))\n\n        x = x.view(-1, self.num_flat_features(x))\n\n        x = F.relu(self.fc6(x))\n        x = F.relu(self.fc7(x))\n        x = self.fc8(x)\n        return x\n\n    def num_flat_features(self, x):\n        size = x.size()[1:]\n\n        num_features = 1\n        for s in size:\n            num_features *= s\n        return num_features\n\n\nif __name__ == '__main__':\n\n    BATCH_SIZE = 128\n    LOG_INTERVAL = 5\n\n    # data\n    transform = transforms.Compose([\n        transforms.ToTensor()\n    ])\n\n    trainset = torchvision.datasets.CIFAR100(\n        root='.\/data',\n        train=True,\n        download=True,\n        transform=transform\n    )\n\n    testset = torchvision.datasets.CIFAR100(\n        root='.\/data',\n        train=False,\n        download=True,\n        transform=transform\n    )\n\n    trainloader = torch.utils.data.DataLoader(trainset, batch_size=BATCH_SIZE, shuffle=True)\n    testloader = torch.utils.data.DataLoader(testset, batch_size=BATCH_SIZE, shuffle=False)\n\n    # model\n    vgg16 = VGG16(100)\n    vgg16.cuda()\n\n    # optimizer\n    optimizer = optim.SGD(vgg16.parameters(), lr=0.01)\n\n    # loss\n    criterion = nn.CrossEntropyLoss()\n\n    print('\u2014\u2014\u2014\u2014 Train Start \u2014\u2014\u2014\u2014\u2014')\n    for epoch in range(20):\n        running_loss = 0.\n        for step, (batch_x, batch_y) in enumerate(trainloader):\n            batch_x, batch_y = batch_x.cuda(), batch_y.cuda()\n\n            # \n            optimizer.zero_grad()\n\n            output = vgg16(batch_x)\n            loss = criterion(output, batch_y)\n            loss.backward()\n            optimizer.step()\n\n            running_loss += loss.item()\n\n            if step % LOG_INTERVAL == 0:\n                print('[%d, %4d] loss: %.4f' % (epoch, step, running_loss \/ LOG_INTERVAL))\n                running_loss = 0.\n\n        def test():\n            print('\u2014\u2014\u2014\u2014 Test Start \u2014\u2014\u2014\u2014')\n            correct = 0\n            total = 0\n\n            # \n            with torch.no_grad():\n                for test_x, test_y in testloader:\n                    images, labels = test_x.cuda(), test_y.cuda()\n                    output = vgg16(images)\n                    _, predicted = torch.max(output.data, 1)\n                    total += labels.size(0)\n                    correct += (predicted == labels).sum().item()\n\n            accuracy = 100 * correct \/ total\n            print('Accuracy of the network is: %.4f %%' % accuracy)\n            print('\u2014\u2014\u2014\u2014 Test Finish \u2014\u2014\u2014\u2014')\n\n        test()\n\n    print('\u2014\u2014\u2014\u2014 Train Finish \u2014\u2014\u2014\u2014\u2014')\n\n<\/code><\/pre>\n\n<p>The loss stays around 4.6060 and never decrease. I have tried different learning rate but does not work. <\/p>\n","answers":[{"AnswerId":"57607939","CreationDate":"2019-08-22T11:05:02.283","ParentId":null,"OwnerUserId":"5027315","Title":null,"Body":"<p>I have noticed that you are not using Batch normalization in between your convolution layers. I have added batch normalization layers and it seems to work. Following is the modified code:<\/p>\n\n<pre><code>class VGG16(torch.nn.Module):\ndef __init__(self, n_classes):\n    super(VGG16, self).__init__()\n\n    # construct model\n    self.conv1_1 = nn.Conv2d(3, 64, 3, padding=1)\n    self.conv11_bn = nn.BatchNorm2d(64)\n    self.conv1_2 = nn.Conv2d(64, 64, 3, padding=1)\n    self.conv12_bn = nn.BatchNorm2d(64)\n    self.conv2_1 = nn.Conv2d(64, 128, 3, padding=1)\n    self.conv21_bn = nn.BatchNorm2d(128)\n    self.conv2_2 = nn.Conv2d(128, 128, 3, padding=1)\n    self.conv22_bn = nn.BatchNorm2d(128)\n    self.conv3_1 = nn.Conv2d(128, 256, 3, padding=1)\n    self.conv31_bn = nn.BatchNorm2d(256)\n    self.conv3_2 = nn.Conv2d(256, 256, 3, padding=1)\n    self.conv32_bn = nn.BatchNorm2d(256)\n    self.conv3_3 = nn.Conv2d(256, 256, 3, padding=1)\n    self.conv33_bn = nn.BatchNorm2d(256)\n    self.conv4_1 = nn.Conv2d(256, 512, 3, padding=1)\n    self.conv41_bn = nn.BatchNorm2d(512)\n    self.conv4_2 = nn.Conv2d(512, 512, 3, padding=1)\n    self.conv42_bn = nn.BatchNorm2d(512)\n    self.conv4_3 = nn.Conv2d(512, 512, 3, padding=1)\n    self.conv43_bn = nn.BatchNorm2d(512)\n    self.conv5_1 = nn.Conv2d(512, 512, 3, padding=1)\n    self.conv51_bn = nn.BatchNorm2d(512)        \n    self.conv5_2 = nn.Conv2d(512, 512, 3, padding=1)\n    self.conv52_bn = nn.BatchNorm2d(512)\n    self.conv5_3 = nn.Conv2d(512, 512, 3, padding=1)\n    self.conv53_bn = nn.BatchNorm2d(512)\n\n    self.fc6 = nn.Linear(512, 512)\n    self.fc7 = nn.Linear(512, 512)\n    self.fc8 = nn.Linear(512, n_classes)\n\ndef forward(self, x):\n    x = F.relu(self.conv11_bn(self.conv1_1(x)))\n    x = F.relu(self.conv12_bn(self.conv1_2(x)))\n    x = F.max_pool2d(x, (2, 2))\n\n    x = F.relu(self.conv22_bn(self.conv2_1(x)))\n    x = F.relu(self.conv21_bn(self.conv2_2(x)))\n    x = F.max_pool2d(x, (2, 2))\n\n    x = F.relu(self.conv31_bn(self.conv3_1(x)))\n    x = F.relu(self.conv32_bn(self.conv3_2(x)))\n    x = F.relu(self.conv33_bn(self.conv3_3(x)))\n    x = F.max_pool2d(x, (2, 2))\n\n    x = F.relu(self.conv41_bn(self.conv4_1(x)))\n    x = F.relu(self.conv42_bn(self.conv4_2(x)))\n    x = F.relu(self.conv43_bn(self.conv4_3(x)))\n    x = F.max_pool2d(x, (2, 2))\n\n    x = F.relu(self.conv51_bn(self.conv5_1(x)))\n    x = F.relu(self.conv52_bn(self.conv5_2(x)))\n    x = F.relu(self.conv53_bn(self.conv5_3(x)))\n    x = F.max_pool2d(x, (2, 2))\n\n    x = x.view(-1, self.num_flat_features(x))\n\n    x = F.relu(self.fc6(x))\n    x = F.relu(self.fc7(x))\n    x = self.fc8(x)\n    return x\n<\/code><\/pre>\n\n<p>However, a more elegant version of the same could be found <a href=\"https:\/\/github.com\/weiaicunzai\/pytorch-cifar100\/blob\/master\/models\/vgg.py\" rel=\"nofollow noreferrer\">here<\/a> <\/p>\n"}]}
+{"QuestionId":57605217,"AnswerCount":0,"Tags":"<python><pytorch>","CreationDate":"2019-08-22T08:33:52.077","AcceptedAnswerId":null,"OwnerUserId":2123706.0,"Title":"pytorch TypeError: '<' not supported between instances of 'Example' and 'Example' when referring to iterator","Body":"<p>I am trying to use my own dataset to classify text according to <a href=\"https:\/\/github.com\/bentrevett\/pytorch-sentiment-analysis\/blob\/master\/5%20-%20Multi-class%20Sentiment%20Analysis.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/bentrevett\/pytorch-sentiment-analysis\/blob\/master\/5%20-%20Multi-class%20Sentiment%20Analysis.ipynb<\/a>. My dataset is a csv of sentences and a class associated with it. there are 6 different classes:<\/p>\n\n<pre><code>sent                      class\n'the fox is brown'        animal\n'the house is big'        object\n'one water is drinkable'  water\n...\n<\/code><\/pre>\n\n<p>When running:<\/p>\n\n<pre><code>N_EPOCHS = 5\n\nbest_valid_loss = float('inf')\n\nfor epoch in range(N_EPOCHS):\n\n    start_time = time.time()\n    print(start_time)\n    train_loss, train_acc = train(model, train_iterator, optimizer, criterion)\n    #print(train_loss.type())\n    #print(train_acc.type())\n    valid_loss, valid_acc = evaluate(model, valid_iterator, criterion)\n\n    end_time = time.time()\n\n    epoch_mins, epoch_secs = epoch_time(start_time, end_time)\n\n    if valid_loss &lt; best_valid_loss:\n        best_valid_loss = valid_loss\n        torch.save(model.state_dict(), 'tut5-model.pt')\n\n    print(f'Epoch: {epoch+1:02} | Epoch Time: {epoch_mins}m {epoch_secs}s')\n    print(f'\\tTrain Loss: {train_loss:.3f} | Train Acc: {train_acc*100:.2f}%')\n    print(f'\\t Val. Loss: {valid_loss:.3f} |  Val. Acc: {valid_acc*100:.2f}%')\n<\/code><\/pre>\n\n<p>, I obtained the following error<\/p>\n\n<pre><code>1566460706.977012\nepoch_loss\n&lt;torchtext.data.iterator.BucketIterator object at 0x000001FABE907E80&gt;\nTypeError: '&lt;' not supported between instances of 'Example' and 'Example'\n<\/code><\/pre>\n\n<p>pointing to:<\/p>\n\n<pre><code>TypeError                                 Traceback (most recent call last)\n&lt;ipython-input-22-19e8a7eb204e&gt; in &lt;module&gt;()\n     10     #print(train_loss.type())\n     11     #print(train_acc.type())\n---&gt; 12     valid_loss, valid_acc = evaluate(model, valid_iterator, criterion)\n     13 \n     14     end_time = time.time()\n\n&lt;ipython-input-21-83b02f99bca7&gt; in evaluate(model, iterator, criterion)\n      9         print('epoch_loss')\n     10         print(iterator)\n---&gt; 11         for batch in iterator:\n     12             print('batch')\n     13             predictions = model(batch.text)\n<\/code><\/pre>\n\n<p>I am new to pytorch, so only added a line to idenify iterator data type and got:<\/p>\n\n<pre><code>&lt;torchtext.data.iterator.BucketIterator object at 0x000001FABE907E80&gt;\n<\/code><\/pre>\n\n<p>I tried to determine specific attributes following  <a href=\"https:\/\/github.com\/pytorch\/text\/blob\/master\/torchtext\/data\/iterator.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/text\/blob\/master\/torchtext\/data\/iterator.py<\/a> to no avail.<\/p>\n\n<p>Any suggestions are appreciated.<\/p>\n\n<p>The code for the evaluate (where the iterator is) method is:<\/p>\n\n<pre><code>def evaluate(model, iterator, criterion):\n\n    epoch_loss = 0\n    epoch_acc = 0\n\n    model.eval()\n\n    with torch.no_grad():\n        print('epoch_loss')\n        print(iterator)\n        for batch in iterator:\n            print('batch')\n            predictions = model(batch.text)\n\n            loss = criterion(predictions, batch.label)\n\n            acc = categorical_accuracy(predictions, batch.label)\n\n            epoch_loss += loss.item()\n            epoch_acc += acc.item()\n\n    return epoch_loss \/ len(iterator), epoch_acc \/ len(iterator)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57605559,"AnswerCount":0,"Tags":"<python><arrays><theano><tensor>","CreationDate":"2019-08-22T08:55:28.370","AcceptedAnswerId":null,"OwnerUserId":11961405.0,"Title":"I want to create a tensor matrix using theano, where some elements are not allowed to change","Body":"<p>Let's say I have some variables: a,b; that I want to add into a tensor matrix which is really sparse. To simplify:<\/p>\n\n<pre><code>[[0,a],[-b,0]\n[0,2a-2b],[-2b,0]]\n<\/code><\/pre>\n\n<p>Then I want to perform some matrix operations, and I need to find a, b that minimize a certain function.<\/p>\n\n<p>I have tried to understand the meaning of a tensor. So, I think what would approximate what I want is to have some tensor scalars to be put into a matrix.<\/p>\n\n<p>However, I have tried with some theano code but I do not get the matrix values to change as I change the tensor scalars. I am not sure if I should create the tensor matrix with theano or numpy.<\/p>\n\n<pre><code>import theano as th\nimport theano.tensor as tt\nimport numpy as np\n\na = tt.fscalar('a')\nb = tt.fscalar('b')\n\nM = np.array([[0,a],[-b,0],\n[0,2a-2b],[-2b,0]])\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57605741,"AnswerCount":2,"Tags":"<python><numpy><tensorflow><keras>","CreationDate":"2019-08-22T09:05:57.157","AcceptedAnswerId":57606215.0,"OwnerUserId":4699994.0,"Title":"Convolutional Net Model not working on very simple data","Body":"<p>I have generated a very simple dataset. Basically, a 3x3 image with either a column of white pixels or a row of white pixels. <\/p>\n\n<p><a href=\"https:\/\/www.tryimg.com\/image\/nbFPe\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/www.tryimg.com\/u\/2019\/08\/22\/a1ef2d587439b21c2c.th.png\" alt=\"a1ef2d587439b21c2c.th.png\"><\/a> <a href=\"https:\/\/www.tryimg.com\/image\/nbRQL\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/www.tryimg.com\/u\/2019\/08\/22\/a278cffc11d8908c38.th.png\" alt=\"a278cffc11d8908c38.th.png\"><\/a> <a href=\"https:\/\/www.tryimg.com\/image\/nbTxY\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/www.tryimg.com\/u\/2019\/08\/22\/a35516d69bdf82bab3.th.png\" alt=\"a35516d69bdf82bab3.th.png\"><\/a> <a href=\"https:\/\/www.tryimg.com\/image\/nbtuW\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/www.tryimg.com\/u\/2019\/08\/22\/b19abb214427ef5ce1.th.png\" alt=\"b19abb214427ef5ce1.th.png\"><\/a> <a href=\"https:\/\/www.tryimg.com\/image\/nbQ62\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/www.tryimg.com\/u\/2019\/08\/22\/b2988962f9bc3ce754.th.png\" alt=\"b2988962f9bc3ce754.th.png\"><\/a> <a href=\"https:\/\/www.tryimg.com\/image\/nbahB\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/www.tryimg.com\/u\/2019\/08\/22\/b36c69bdb057e3258c.th.png\" alt=\"b36c69bdb057e3258c.th.png\"><\/a><\/p>\n\n<p>I am trying to train a convolutional net model to distinguish between the images of either a row white pixels or a column of white pixels. Hoping by the end of the training, the filters(two 2x2 conv filters, no bias in conv-layer) will be doing some sort of vertical\/horizontal edge detection.<\/p>\n\n<p>This is my architecture:<\/p>\n\n<blockquote>\n  <p>Conv2d->Relu->Flatten->Dense->Softmax<\/p>\n<\/blockquote>\n\n<pre class=\"lang-py prettyprint-override\"><code>class MyNet:\n    @staticmethod\n    def build(width, height, depth, classes):\n        # initialize the model along with the input shape to be\n        # 'channels last' and the channels dimension itself\n        model = Sequential()\n        inputShape = (height, width, depth)\n        chanDim = -1\n\n        # if we are using \"channels first\", update the input shape\n        # and channels dimension\n        if K.image_data_format() == \"channels_first\":\n            inputShape = (depth, height, width)\n            chanDim = 1\n\n        # add some layers\n        # CONV =&gt; RELU\n        model.add(Conv2D(filters=2, kernel_size=(2,2), padding=\"valid\",\n                          use_bias=False ,input_shape=inputShape))\n        model.add(Activation(\"relu\"))\n        # finally, we have a fully connected dense layer  layers and a \n        # softmax classifier:\n        model.add(Flatten())\n        # softmax classifier\n        model.add(Dense(classes))\n        model.add(Activation(\"softmax\"))\n\n        # return the constructed network architecture\n        return model\n<\/code><\/pre>\n\n<p>*I am using a softmax layer as output for this experiment. <\/p>\n\n<p>I have the following parameters set up:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># initialize the number of epochs to train for, initial learning rate,\n# batch size, and image dimensions\nnp.random.seed(48)\nEPOCHS = 100\nINIT_LR = 1e-3\nBS = 6\nIMAGE_DIMS = (3, 3, 1)\n\nmodel = MyNet.build(3,3,1,2) # create model\n\n# compile model\nmodel.compile(optimizer= keras.optimizers.SGD(lr=INIT_LR),\n              loss= tf.nn.softmax_cross_entropy_with_logits_v2,\n              metrics=['accuracy'])\n\n# Finally, train model\nmodel.fit(X_train, labels, \n          batch_size=BS,\n          epochs=EPOCHS,\n          verbose=1\n         )\n<\/code><\/pre>\n\n<p>Pretty straight forward classification task, but for some reason, which escapes me, this model doesn't learn squat!<\/p>\n\n<pre><code>...\nEpoch 96\/100\n6\/6 [==============================] - 0s 10ms\/step - loss: 0.8133 - acc: 0.5000\nEpoch 97\/100\n6\/6 [==============================] - 0s 5ms\/step - loss: 0.8133 - acc: 0.5000\nEpoch 98\/100\n6\/6 [==============================] - 0s 4ms\/step - loss: 0.8133 - acc: 0.5000\nEpoch 99\/100\n6\/6 [==============================] - 0s 4ms\/step - loss: 0.8133 - acc: 0.5000\nEpoch 100\/100\n6\/6 [==============================] - 0s 5ms\/step - loss: 0.8133 - acc: 0.5000\n\n<\/code><\/pre>\n\n<p>I have tried changing the learning rate, number of epochs, the optimizer, loss function, etc. but still nothing.<\/p>\n\n<p>With the random seed set to 48, I get the following 2 filters which remain same after training :(, so no learning at all.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.get_weights()[0][:,:,0,0]\n\n&gt; array([[ 0.23327285, -0.06844258],\n&gt;       [ 0.61764306, -0.46921107]], dtype=float32)\n\n--------------------------------------\nmodel.get_weights()[0][:,:,0,1]\n&gt; array([[ 0.68603402,  0.13425004],\n&gt;       [ 0.39129239,  0.23840684]], dtype=float32)\n\n\n<\/code><\/pre>\n\n<p>BTW, I ran this experiment on a conv-net I created from scratch(all numpy) with the same architecture, it performed much better.<\/p>\n\n<p>The data is set up as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\n\nimage_a  = np.array([\n    [0, 0, 0],\n    [255, 255, 255],\n    [0, 0, 0],\n], dtype=\"uint8\")\n\nimage_a_2  = np.array([\n    [255, 255, 255],\n    [0, 0, 0],\n    [0, 0, 0],\n], dtype=\"uint8\")\n\nimage_a_3  = np.array([\n    [0, 0, 0],\n    [0, 0, 0],\n    [255, 255, 255],\n], dtype=\"uint8\")\n\nimage_b  = np.array([\n    [0, 255, 0],\n    [0, 255, 0],\n    [0, 255, 0],\n], dtype=\"uint8\")\n\nimage_b_2  = np.array([\n    [255, 0, 0],\n    [255, 0, 0],\n    [255, 0, 0],\n], dtype=\"uint8\")\n\nimage_b_3  = np.array([\n    [0, 0,  255],\n    [0, 0,  255],\n    [0, 0,  255],\n], dtype=\"uint8\")\n\nX_train = np.array([image_a,image_a_2, image_a_3, image_b, image_b_2, image_b_3])\n\n# set up in correct shape for tensorflow (batch, height, width, channel)\nX_train = X_train.reshape(X_train.shape[0], 3, 3, 1)\n\ny = [0,0,0,1,1,1] # images with row of white pixels-&gt;0, and col of white-&gt;1\nlabels = np.eye(2, )[y]\n\n\n<\/code><\/pre>\n\n<p>So, someone where am I going wrong.<\/p>\n","answers":[{"AnswerId":"57606712","CreationDate":"2019-08-22T09:58:12.887","ParentId":null,"OwnerUserId":"10001536","Title":null,"Body":"<p>you should normalize the data into the range [0, 1]. If you have the range [0, 255] it is likely that the network wont learn.\nJust divide the dataset by 255.\nFurthermore, i would use the adam optimizer and binary crossentropy as loss.<\/p>\n"},{"AnswerId":"57606215","CreationDate":"2019-08-22T09:30:52.130","ParentId":null,"OwnerUserId":"10455493","Title":null,"Body":"<p>You are using <code>tf.nn.softmax_cross_entropy_with_logits_v2<\/code> loss function which expect logit inputs, see <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/softmax_cross_entropy_with_logits_v2\" rel=\"nofollow noreferrer\">documentation<\/a> (unscaled logits, as it does the softmax in the loss calculation for optimization), hence you should not have a <code>softmax<\/code> activation at the output of your <code>MyNet<\/code>. If you dont want to remove the softmax activation on the output of <code>MyNet<\/code> because of inference, you can use the <code>tf.keras.losses.categorical_crossentropy<\/code> loss function, with <code>from_logits=False<\/code> instead.<\/p>\n\n<hr>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>Another problem is that your Conv2D layer only uses 2 filters. Hence only two sets of weights to learn the 6 types of situations. This seems to be a problem, as when this is changed to 6 (together with `padding='same') the model gives 100% accuracy. <\/p>\n\n<p>I also changed the optimizer to adam, as my experience with this optimizer is better. And I scaled the data between 0 and 1.\nsee below code for a working example.<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorflow.keras.backend as K\nimport numpy as np\nclass MyNet:\n    @staticmethod\n    def build(width, height, depth, classes):\n        # initialize the model along with the input shape to be\n        # 'channels last' and the channels dimension itself\n        model = tf.keras.models.Sequential()\n        inputShape = (height, width, depth)\n\n        # if we are using \"channels first\", update the input shape\n        # and channels dimension\n        if K.image_data_format() == \"channels_first\":\n            inputShape = (depth, height, width)\n        # add some layers\n        # CONV =&gt; RELU\n        model.add(tf.keras.layers.Conv2D(filters=6, kernel_size=(2,2), padding=\"same\",\n                          use_bias=False ,input_shape=inputShape))\n        model.add(tf.keras.layers.Activation(\"relu\"))\n        # finally, we have a fully connected dense layer  layers and a \n        # softmax classifier:\n        model.add(tf.keras.layers.Flatten())\n        # softmax classifier\n        model.add(tf.keras.layers.Dense(classes))\n        model.add(tf.keras.layers.Activation(\"softmax\"))\n\n        # return the constructed network architecture\n        return model\n\nimage_a  = np.array([\n    [0, 0, 0],\n    [255, 255, 255],\n    [0, 0, 0],\n], dtype=\"uint8\")\n\nimage_a_2  = np.array([\n    [255, 255, 255],\n    [0, 0, 0],\n    [0, 0, 0],\n], dtype=\"uint8\")\n\nimage_a_3  = np.array([\n    [0, 0, 0],\n    [0, 0, 0],\n    [255, 255, 255],\n], dtype=\"uint8\")\n\nimage_b  = np.array([\n    [0, 255, 0],\n    [0, 255, 0],\n    [0, 255, 0],\n], dtype=\"uint8\")\n\nimage_b_2  = np.array([\n    [255, 0, 0],\n    [255, 0, 0],\n    [255, 0, 0],\n], dtype=\"uint8\")\n\nimage_b_3  = np.array([\n    [0, 0,  255],\n    [0, 0,  255],\n    [0, 0,  255],\n], dtype=\"uint8\")\nX_train = np.array([image_a,image_a_2, image_a_3, image_b, image_b_2, image_b_3])\n\n# set up in correct shape for tensorflow (batch, height, width, channel)\nX_train = X_train.reshape(X_train.shape[0], 3, 3, 1)\/255\n\ny = [0,0,0,1,1,1] # images with row of white pixels-&gt;0, and col of white-&gt;1\nlabels = np.eye(2, )[y]\n\n\n\n# initialize the number of epochs to train for, initial learning rate,\n# batch size, and image dimensions\ntf.set_random_seed(48)\nEPOCHS = 100\nBS = 6\nIMAGE_DIMS = (3, 3, 1)\n\nmodel = MyNet.build(3,3,1,2) # create model\n\n# compile model\nmodel.compile(optimizer= tf.keras.optimizers.Adam(),\n              loss= tf.keras.losses.categorical_crossentropy,\n              metrics=['accuracy'])\n\n# Finally, train model\nmodel.fit(X_train, labels, \n          batch_size=BS,\n          epochs=EPOCHS,\n          verbose=1\n         )\npred = model.predict(X_train)\npred_labels = (pred&gt;0.5)\nprint('All predictions equal to labels, ' + str(np.all(pred_labels ==labels)))\n\n\n<\/code><\/pre>\n\n<p>which results in the output<\/p>\n\n<pre><code>&gt; All predictions equal to labels, True\n<\/code><\/pre>\n"}]}
+{"QuestionId":57605793,"AnswerCount":0,"Tags":"<dataset><pytorch>","CreationDate":"2019-08-22T09:09:29.753","AcceptedAnswerId":null,"OwnerUserId":6250501.0,"Title":"Pytorch - Choose from more datasets?","Body":"<p>I have a dataset structured this way:<\/p>\n\n<pre><code>root\n    superclass 1\n       class 1\n          img_name_xyd.jpg\n          img_name_xxs.jpg\n          img_name_fdf.jpg\n       class 2\n          img_name_ddv.jpg\n    superclass 2\n       class 3\n          img_name_gdv.jpg\n       class 4\n          img_name_bdd.jpg\n       class 5\n          img_name_hjd.jpg\n    .\n    .\n    .\n\n    superclass N\n       class X\n          img_name_sfb.jpg\n<\/code><\/pre>\n\n<p>Is it possible to easily load dataset containing superclasses, classes and images in that way that I can choose batch to have images from only one randomly choosen superclass and randomly choosen classes of coresponding dataset?<\/p>\n\n<p>Example :<\/p>\n\n<p>I want my batch to have choosen:<\/p>\n\n<ul>\n<li>superclass1 \/ class 1 \/ img_name_xyd.jpg <\/li>\n<li>superclass1 \/ class 2 \/   img_name_ddv.jpg <\/li>\n<li>superclass1 \/ class 1 \/ img_name_xxs.jpg<\/li>\n<\/ul>\n\n<p>in next batch then <\/p>\n\n<ul>\n<li>superclassN \/ class X \/ img_name_sfb.jpg<\/li>\n<\/ul>\n\n<p>etc.<\/p>\n","answers":[]}
+{"QuestionId":57606006,"AnswerCount":0,"Tags":"<python><keras><deep-learning>","CreationDate":"2019-08-22T09:20:34.113","AcceptedAnswerId":null,"OwnerUserId":11949870.0,"Title":"ValueError: total size of new array must be unchanged (https:\/\/github.com\/ajgupta93\/enhanced-view-synthesis\/blob\/master\/code\/viewsyn_model.py)","Body":"<p>I am using a <a href=\"https:\/\/github.com\/ajgupta93\/enhanced-view-synthesis\/blob\/master\/code\/viewsyn_model.py\" rel=\"nofollow noreferrer\">code from this link<\/a>.\nWhen I test the full network I receive <\/p>\n\n<blockquote>\n  <p>\"ValueError: total size of new array must be unchanged. <\/p>\n<\/blockquote>\n\n<p>What could be the possible cause? Any help would be appreciated.<\/p>\n\n<p>I have tested dataset samples provided by the author. I took about 700 images and then tried. Author gave 7890 image datasets.<\/p>\n\n<pre><code>'''\n\ndef get_optimizer(name = 'adagrad', l_rate = 0.0001, dec = 0.0, b_1 = 0.9, b_2 = 0.999, mom = 0.5, rh = 0.9):\n    eps = 1e-8\n\n    adam = Adam(lr = l_rate, beta_1 = b_1, beta_2 = b_2, epsilon = eps, decay = dec)\n    sgd = SGD(lr = l_rate, momentum = mom, decay = dec, nesterov = True)\n    rmsp = RMSprop(lr = l_rate, rho = rh, epsilon = eps, decay = dec)\n    adagrad = Adagrad(lr = l_rate, epsilon = eps, decay = dec)\n\n    optimizers = {'adam': adam, 'sgd':sgd, 'rmsp': rmsp, 'adagrad': adagrad}\n\n    return optimizers[name]\n\ndef build_autoencoder():\n    #define network architecture for encoder\n    model = Sequential()\n\n    #6 convoltuion layers\n    model.add(Convolution2D(16, 3, 3, border_mode='same', subsample = (2,2), activation = 'relu',\n            input_shape=(224, 224, 3)))\n    model.add(Convolution2D(32, 3, 3, border_mode='same', subsample = (2,2), activation = 'relu'))\n    model.add(Convolution2D(64, 3, 3, border_mode='same', subsample = (2,2), activation = 'relu'))\n    model.add(Convolution2D(128, 3, 3, border_mode='same', subsample = (2,2), activation = 'relu'))\n    model.add(Convolution2D(256, 3, 3, border_mode='same', subsample = (2,2), activation = 'relu'))\n    model.add(Convolution2D(512, 3, 3, border_mode='same', subsample = (2,2), activation = 'relu'))\n\n    #Flatten \n    model.add(Flatten())\n\n    #2 fully connected layers\n    model.add(Dense(4096, activation='relu'))\n    model.add(Dropout(p=0))\n    model.add(Dense(4096, activation='relu'))\n    model.add(Dropout(p=0))\n\n\n    #define network architecture for decoder\n\n    #2 fully connected layers\n    model.add(Dense(4096, activation='relu'))\n    model.add(Dense(4096, activation='relu'))\n\n    #reshape to 2D\n    model.add(Reshape((8, 8, 64)))\n\n    #5 upconv layers\n    model.add(Deconvolution2D(256, 3, 3, (None, 15, 15,256), border_mode='same', subsample = (2,2), activation = 'relu'))\n    model.add(Deconvolution2D(128, 3, 3, (None, 29, 29, 128), border_mode='same', subsample = (2,2), activation = 'relu'))\n    model.add(Deconvolution2D(64, 3, 3, (None, 57, 57, 64), border_mode='same', subsample = (2,2), activation = 'relu'))\n    model.add(Deconvolution2D(32, 3, 3, (None, 113, 113, 32),border_mode='same', subsample = (2,2), activation = 'relu'))\n    model.add(Deconvolution2D(16, 3, 3, (None, 225, 225, 16), border_mode='same', subsample = (2,2), activation = 'relu'))\n\n    #output layer, add 3 RGB channels for reconstructed output view\n    model.add(Deconvolution2D(3, 3, 3, (None, 225, 225, 3), border_mode='same', subsample = (1,1), activation = 'relu'))\n\n    #add a resize layer to resize (225, 225) to (224, 224)\n    model.add(Reshape((225*225,3)))\n    model.add(Lambda(lambda x: x[:,:50176,])) # throw away some\n    model.add(Reshape((224,224,3)))\n\n'''\n\nbuild_autoencoder\n    model.add(Reshape((225*225,3)))\n<\/code><\/pre>\n\n<pre><code>  File \"C:\\ProgramData\\Anaconda3\\envs\\myenv\\lib\\site-packages\\keras\\engine\\sequential.py\", line 181, in add\n    output_tensor = layer(self.outputs[0])\n\n  File \"C:\\ProgramData\\Anaconda3\\envs\\myenv\\lib\\site-packages\\keras\\engine\\base_layer.py\", line 474, in __call__\n    output_shape = self.compute_output_shape(input_shape)\n\n  File \"C:\\ProgramData\\Anaconda3\\envs\\myenv\\lib\\site-packages\\keras\\layers\\core.py\", line 398, in compute_output_shape\n    input_shape[1:], self.target_shape)\n\n  File \"C:\\ProgramData\\Anaconda3\\envs\\myenv\\lib\\site-packages\\keras\\layers\\core.py\", line 386, in _fix_unknown_dimension\n    raise ValueError(msg)\n\nValueError: total size of new array must be unchanged\n<\/code><\/pre>\n\n<p>The error is in line <code>reshape ((225*225,3))<\/code><\/p>\n","answers":[]}
+{"QuestionId":57606060,"AnswerCount":1,"Tags":"<python><keras><neural-network><lstm>","CreationDate":"2019-08-22T09:24:05.350","AcceptedAnswerId":57606746.0,"OwnerUserId":11712588.0,"Title":"LSTM - How can I prevent dirfferent loss \/ R2 outcomes for the same input?","Body":"<p>I have a LSTM neural Network in python (kears) I want to use tp predict some time series Data.<\/p>\n\n<p>The Problem i have is that i always get different loss and r2 (degree of determination) values, sometimes even NaN as loss, although i don't change the code at all<\/p>\n\n<p>I tried to use different loss functions (mse, mae). I also tried to change layers or the size of my sequence, Nothing worked.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>changes = Bedarf_data[[\"Bedarf_changes\", \"Temperatur_changes\", \"monat\", \"uhrzeit\"]].values\n\nX = []\nY = []\n\nfor i in range (0, len(changes) - 96):\n    Y.append(changes[i,0])\n    X.append(np.array(changes[i+1:i+97]))\n\nX = np.array(X).reshape(-1, 96, 4)\nY = np.array(Y)\n\n# Modell\nfrom keras.models import Sequential\nfrom keras.layers import LSTM\nfrom keras.layers import Dense\n\nmodel = Sequential()\nmodel.add(LSTM(12, activation=\"relu\", input_shape=(96, 4)))\nmodel.add(Dense(1))\n\nmodel.compile(optimizer=\"rmsprop\", loss=\"mse\")\nmodel.fit(X, Y, batch_size=96, epochs=10)\n<\/code><\/pre>\n\n<p>I expected at least a stable Outcome without too much Variation.\nWhat can I do to get at least a stable Output?<\/p>\n","answers":[{"AnswerId":"57606746","CreationDate":"2019-08-22T09:59:38.057","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Neural networks use a lot of random initialization. This means that every time you create a new net, it's different from the others. <\/p>\n\n<p>You can set a random seed (but you need to reset the kernel entirely before every run) to guarantee that the random numbers will always be the same:<\/p>\n\n<pre><code>numpy.random.seed(choose_a_number)\n<\/code><\/pre>\n\n<p>You may also try <code>tf.set_random_seed(choose_a_number)<\/code>   <\/p>\n"}]}
+{"QuestionId":57606338,"AnswerCount":0,"Tags":"<tensorflow><keras><deep-learning><pytorch>","CreationDate":"2019-08-22T09:37:25.787","AcceptedAnswerId":null,"OwnerUserId":6831672.0,"Title":"How to attach additional \"passive\" information to a convolutional neural network input or state?","Body":"<p>I have a quite particular thing I want to try, stemming from a research question I am trying to answer, but I do not know how to implement it in a deep learning framework.<\/p>\n\n<p>I have a trained convolutional neural network with all the typical kind of layers that there might be (<code>Conv2D<\/code>, <code>Dense<\/code>, <code>MaxPool2D<\/code>, possibly <code>BatchNormalization<\/code> etc.) and I am passing images through. To each pixel of the input image, I have a numpy-array containing some particular additional information. These arrays are all the same shape\/length. What I want to do is, compute a forward pass, such that this additional information vector\/array is processed element-wise by the network in the same way as the pixel (or intermediate output deeper in the network) is, but <strong>not<\/strong> have any influence on the forward pass themselves.<\/p>\n\n<p>That is, if it passes through e.g. a <code>Conv2D<\/code>-layer (with linear activation), these arrays are combined linearly and element-wise with the arrays coming from the other neurons according to the weights, just like the other inputs of that layer are. <em>But<\/em> if there is e.g. a <code>MaxPool2D<\/code>-layer or a ReLU, these arrays attached to the flowing numbers (tensor entries) should be zero'd out elementwise exactly when the original input or state information, which is flowing through the network, is also zero'd out.<\/p>\n\n<p>Then I want to be able to look at all intermediate outputs of the forward pass and see how this additional information is processed, as was passed through the network. These intermediate outputs is what I am interested in.<\/p>\n\n<p>Therefore, e.g. just passing my extra data as additional elements in the input batch, for example, does not work, since e.g. in a ReLU, the elements of my array would be set to zero if <em>they<\/em> are negative. Instead, all elements of the array should be set to zero precisely when <em>the usual \"intermediate output number\"<\/em> passing through the corresponding ReLU of that neuron is negative. That's why I called this information \"passive\" in the title. If the intermediate state goes from 2 to 6 because of some other nonlinearity, then the elements of my array should be multiplied by 3. (Not that there are cases where this would not work, e.g. the sigmoid when going from 0 to 0.5, but let's ignore this problem for now, I won't choose sigmoids.)<\/p>\n\n<p>I hope that my explanation is clear enough. Is there any experienced Python\/Deep Learning programmer with a trick up their sleeves, how to achieve what I want? In the best case, without too much work, but still using all the capabilities of the DL framework? Also, if you can make it work just for special cases (not all possible layers), I am interested. As mentioned, I do not need to do training\/backprop, I just want to compute these forward passes for pre-trained networks. I am currently trying this by re-implementing convolutional layers, max-pooling etc. myself and then computing what I want by getting the weights from the trained network. But this is a lot of work and only works for the layers I have implemented.<\/p>\n\n<p>I am playing with keras at the moment, so a solution for keras would be great, but I am happy with any other popular Python framework really, if that helps me achieve my goal.<\/p>\n","answers":[]}
+{"QuestionId":57606581,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-22T09:49:26.407","AcceptedAnswerId":null,"OwnerUserId":11901042.0,"Title":"Error while loading model using keras and flask","Body":"<p>I'm writing a simple application using flask. First I tried this: <\/p>\n\n<pre><code>def get_model():\n    global model\n    model = load_model('mobilenet.h5')\n    print(\"** Model loaded!\")\n<\/code><\/pre>\n\n<p>But I got an error: <\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\Scripts\\flask-script.py\", line 10, in &lt;module&gt;\n    sys.exit(main())\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\flask\\cli.py\", line 894, in main\n    cli.main(args=args, prog_name=name)\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\flask\\cli.py\", line 557, in main\n    return super(FlaskGroup, self).main(*args, **kwargs)\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\click\\core.py\", line 697, in main\n    rv = self.invoke(ctx)\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\click\\core.py\", line 1066, in invoke\n    return _process_result(sub_ctx.command.invoke(sub_ctx))\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\click\\core.py\", line 895, in invoke\n    return ctx.invoke(self.callback, **ctx.params)\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\click\\core.py\", line 535, in invoke\n    return callback(*args, **kwargs)\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\click\\decorators.py\", line 64, in new_func\n    return ctx.invoke(f, obj, *args[1:], **kwargs)\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\click\\core.py\", line 535, in invoke\n    return callback(*args, **kwargs)\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\flask\\cli.py\", line 767, in run_command\n    app = DispatchingApp(info.load_app, use_eager_loading=eager_loading)\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\flask\\cli.py\", line 293, in __init__\n    self._load_unlocked()\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\flask\\cli.py\", line 317, in _load_unlocked\n    self._app = rv = self.loader()\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\flask\\cli.py\", line 372, in load_app\n    app = locate_app(self, import_name, name)\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\flask\\cli.py\", line 235, in locate_app\n    __import__(module_name)\n  File \"C:\\Users\\Agnieszka\\Desktop\\noc-naukowcow\\flask_apps\\predict.py\", line 32, in &lt;module&gt;\n    get_model()\n  File \"C:\\Users\\Agnieszka\\Desktop\\noc-naukowcow\\flask_apps\\predict.py\", line 19, in get_model\n    model = tf.keras.models.load_model('mobilenet.h5')\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\saving.py\", line 249, in load_model\n    optimizer_config, custom_objects=custom_objects)\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\tensorflow\\python\\keras\\optimizers.py\", line 838, in deserialize\n    printable_module_name='optimizer')\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\tensorflow\\python\\keras\\utils\\generic_utils.py\", line 194, in deserialize_keras_object\n    return cls.from_config(cls_config)\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\tensorflow\\python\\keras\\optimizers.py\", line 159, in from_config\n    return cls(**config)\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\tensorflow\\python\\keras\\optimizers.py\", line 471, in __init__\n    super(Adam, self).__init__(**kwargs)\n  File \"C:\\Users\\Agnieszka\\Anaconda4\\lib\\site-packages\\tensorflow\\python\\keras\\optimizers.py\", line 68, in __init__\n    'passed to optimizer: ' + str(k))\nTypeError: Unexpected keyword argument passed to optimizer: name\n<\/code><\/pre>\n\n<p>Then I changed the code using tf.keras.models.load_model but this haven't helped. <\/p>\n\n<p>The model was built in Colaboratory using keras 2.2.4 version. I've built the model as follows:<\/p>\n\n<pre><code>from keras import backend as K\n\ninputs = Input(shape=(96, 80, 3), name='input')\n\n\nx = Conv2D(64, (3, 3), activation='relu')(inputs)\nx = MaxPooling2D(pool_size=(2, 2))(x)\nx = Dropout(0.25)(x)\n\nx = Conv2D(128, (3, 3), activation='relu')(x)\nx = MaxPooling2D(pool_size=(2, 2))(x)\nx = Dropout(0.25)(x)\n\nx = Conv2D(256, (3, 3), activation='relu')(x)\nx = MaxPooling2D(pool_size=(2, 2))(x)\nx = Dropout(0.25)(x)\n\nx = Flatten()(x)\nx = Dense(256, activation='relu')(x)\nx = Dropout(0.5)(x)\nx = Dense(35, activation='softmax')(x)\n\nmodel = Model(inputs=inputs, outputs=x, name='CNN')\n\n\n\nmodel.compile(optimizer=optimizer, loss=loss_fn, metrics=['acc'])\n<\/code><\/pre>\n\n<p>I wonder if the problem is due to optimaizer that I used (it was adam). <\/p>\n\n<p>Is there any solution for this?<\/p>\n","answers":[{"AnswerId":"57611877","CreationDate":"2019-08-22T14:40:14.973","ParentId":null,"OwnerUserId":"10878733","Title":null,"Body":"<p>The error suggests that <code>Unexpected keyword argument passed to optimizer: name<\/code>. Now as we know, the argument <code>name<\/code> is used in TensorFlow models to name a specific op or tensor. <strong>TensorFlow has included Keras, but Keras still exists as a standalone package.<\/strong><\/p>\n\n<p>Maybe you have created a model using <code>tf.keras<\/code> but you are loading it using <code>keras.models.load_model<\/code>. The Keras packages can't recognize the TF argument.<\/p>\n\n<p>Maybe you can try to bring the whole process in TensorFlow by using methods from <code>tf.keras<\/code>.<\/p>\n"}]}
+{"QuestionId":57607009,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-22T10:14:18.887","AcceptedAnswerId":null,"OwnerUserId":7697890.0,"Title":"Custom optimizer for tensorflow.compat.v1","Body":"<p>I'm trying to create my own optimizer compatible with tensorflow.compat.v1 that uses OptimizerV2 as base class, which defined in optimizer_v2.py<\/p>\n\n<p>The classic keras version of the Optimizer class was easier to understand for me, for example in the function<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>get_updates(self, loss, params):\n<\/code><\/pre>\n\n<p>you could check with<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>param.name\n<\/code><\/pre>\n\n<p>which layer was currently being analysed.<\/p>\n\n<p>On the contrary, for example, in optimizer_v2 there is the function <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>_resource_apply_dense(self, grad, var)\n<\/code><\/pre>\n\n<p>I tried to inspect <strong>var<\/strong>, it has a <strong>.name<\/strong> parameter but it isn't \"unique\". What I mean is that, for example, by having a final Dense layer of name <strong>d<\/strong>, each step of fit_generator calls two times that function with <strong>.name<\/strong> equals to<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>d\/kernel:0\nd\/bias:0\n<\/code><\/pre>\n\n<p>Lets say I want to change the <strong>momentum<\/strong> or the <strong>learning rate<\/strong> only for that dense layer <strong>d<\/strong>, would the following algorithm be correct?<\/p>\n\n<p>Check when <strong>var.name<\/strong> is equal to <strong>d\/kernel:0<\/strong> or <strong>d\/bias:0<\/strong> and, for example, do<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Retrives the actual learning rate, dtype is a float32\nlr_t = self._decayed_lr(var_dtype)\nlr_t = lr_t * 1e-1\n...\n# If i want to apply the standard gradient_descent (same as the optimizer SGD)\nreturn training_ops.resource_apply_gradient_descent(\n          var.handle, lr_t_updated, grad, use_locking=self._use_locking)\n<\/code><\/pre>\n\n<p>Would this set the learning rate, only for that layer <strong>d<\/strong>, 10 times smaller than for the other layers?<\/p>\n","answers":[]}
+{"QuestionId":57607680,"AnswerCount":1,"Tags":"<parallel-processing><conv-neural-network><pytorch>","CreationDate":"2019-08-22T10:50:25.233","AcceptedAnswerId":null,"OwnerUserId":2707990.0,"Title":"How does torch.nn.DataParallel change the output size?","Body":"<p>It looks like using <code>torch.nn.DataParallel<\/code> changes the output size. \nThough in official docs <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.DataParallel\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.DataParallel<\/a>\nall information about size changing is as follows:<\/p>\n\n<blockquote>\n  <p>When module returns a scalar (i.e., 0-dimensional tensor) in forward(), this wrapper will return a vector of length equal to number of devices used in data parallelism, containing the result from each device.<\/p>\n<\/blockquote>\n\n<p>My module returns tensor of 10 coordinates, and I have 2 GPU's where I want to run the code. The last layer of my CNN is <code>nn.Linear(500, 10)<\/code>. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nimport torch.nn as nn\n\nnet = LeNet()    #CNN-class written above\ndevice = torch.device(\"cuda:0\")\nnet.to(device)\nnet = nn.DataParallel(net)\n\n#skipped some code, where inputs and targets are loaded from files\n\noutput = net(input)\ncriterion = nn.SmoothL1Loss()\nloss = criterion(output, target)\n<\/code><\/pre>\n\n<p>Note that without calling <code>DataParallel<\/code> this piece of code works okay. With <code>DataParallel<\/code> the runtime error occurs when trying to calculate loss.<\/p>\n\n<pre><code>RuntimeError: The size of tensor a (20) must match the size of tensor b (10) at non-singleton dimension 0\n<\/code><\/pre>\n\n<p>Seems like output size for each GPU separately is 10 as stated, but afterwards the two outputs are joined and that's where size 20 come from.<\/p>\n\n<p>When changing output size in the CNN-class from 10 to 5 it started to work again, but I'm not sure it's right solution and the CNN will work properly.<\/p>\n","answers":[{"AnswerId":"57612798","CreationDate":"2019-08-22T15:32:29.870","ParentId":null,"OwnerUserId":"5027315","Title":null,"Body":"<p>Easiest <a href=\"https:\/\/medium.com\/huggingface\/training-larger-batches-practical-tips-on-1-gpu-multi-gpu-distributed-setups-ec88c3e51255\" rel=\"nofollow noreferrer\">solution<\/a> to this problem is to take mean of losses from all GPUs before doing backward. This might be problematic when different GPUs have different number of samples, but will work for your case:<\/p>\n\n<pre><code>loss = criterion(output, target)\nloss = loss.mean()\nloss.backward()\n<\/code><\/pre>\n"}]}
+{"QuestionId":57607717,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><autoencoder>","CreationDate":"2019-08-22T10:52:16.403","AcceptedAnswerId":57608568.0,"OwnerUserId":10762045.0,"Title":"Input to reshape is a tensor with 788175 values, but the requested shape has 1050900","Body":"<p>I am importing in some arrays of data to train on but tensorflow is outputting below error.<\/p>\n\n<pre><code>inp = open('train.csv',\"rb\")\nX = pickle.load(inp)\nX = X\/255.0\nX = np.array(X)\nmodel = keras.Sequential([\n    keras.layers.Flatten(input_shape=(113, 75, 3)),\n    keras.layers.Dense(75, activation=tf.nn.relu),\n    keras.layers.Dense(50, activation=tf.nn.relu),\n    keras.layers.Dense(75, activation=tf.nn.relu),\n    keras.layers.Dense(25425, activation=tf.nn.softmax),\n    keras.layers.Reshape((113, 75, 4))\n])\nmodel.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\nmodel.fit(X, X, epochs=5)\n<\/code><\/pre>\n\n<p>I should be able to create an autoencoder but the program outputs this:\nTraceback (most recent call last):<\/p>\n\n<pre><code>File \"C:\\Users\\dalto\\Documents\\geo4\\train.py\", line 24, in &lt;module&gt;\n    model.fit(X, X, epochs=5)\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 643, in fit\n    use_multiprocessing=use_multiprocessing)\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_arrays.py\", line 664, in fit\n    steps_name='steps_per_epoch')\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_arrays.py\", line 383, in model_iteration\n    batch_outs = f(ins_batch)\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py\", line 3510, in __call__\n    outputs = self._graph_fn(*converted_inputs)\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\eager\\function.py\", line 572, in __call__\n    return self._call_flat(args)\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\eager\\function.py\", line 671, in _call_flat\n    outputs = self._inference_function.call(ctx, args)\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\eager\\function.py\", line 445, in call\n    ctx=ctx)\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\eager\\execute.py\", line 67, in quick_execute\n    six.raise_from(core._status_to_exception(e.code, message), None)\n  File \"&lt;string&gt;\", line 3, in raise_from\ntensorflow.python.framework.errors_impl.InvalidArgumentError:  Input to reshape is a tensor with 788175 values, but the requested shape has 1050900\n     [[node reshape\/Reshape (defined at C:\\Users\\dalto\\Documents\\geo4\\train.py:24) ]] [Op:__inference_keras_scratch_graph_922]\n\nFunction call stack:\nkeras_scratch_graph\n<\/code><\/pre>\n\n<p>If I change the Reshape to (113, 75, 3) I get this it doesn't fix the error it just changes it:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"C:\\Users\\dalto\\Documents\\geo4\\train.py\", line 24, in &lt;module&gt;\n    model.fit(X, X, epochs=5)\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 643, in fit\nuse_multiprocessing=use_multiprocessing)\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_arrays.py\", line 664, in fit\nsteps_name='steps_per_epoch')\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_arrays.py\", line 383, in model_iteration\n    batch_outs = f(ins_batch)\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py\", line 3510, in __call__\noutputs = self._graph_fn(*converted_inputs)\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\eager\\function.py\", line 572, in __call__\nreturn self._call_flat(args)\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\eager\\function.py\", line 671, in _call_flat\noutputs = self._inference_function.call(ctx, args)\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\eager\\function.py\", line 445, in call\nctx=ctx)\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\eager\\execute.py\", line 67, in quick_execute\nsix.raise_from(core._status_to_exception(e.code, message), None)\n  File \"&lt;string&gt;\", line 3, in raise_from\ntensorflow.python.framework.errors_impl.InvalidArgumentError:  Incompatible \nshapes: [31,113,75] vs. [31,113,75,3]\n [[node metrics\/accuracy\/Equal (defined at \nC:\\Users\\dalto\\Documents\\geo4\\train.py:24) ]] [Op:__inference_keras_scratch_graph_922]\n<\/code><\/pre>\n","answers":[{"AnswerId":"57608568","CreationDate":"2019-08-22T11:42:36.613","ParentId":null,"OwnerUserId":"11900574","Title":null,"Body":"<p>The input and output size after reshape must be same. So, you'll have to use <code>(113, 75, 3)<\/code> instead of <code>(113, 75, 4)<\/code>. <\/p>\n\n<p>Now, by using <code>(113, 75, 3)<\/code>, you're getting the unequal error because you're using <code>sparse_categorical_crossentropy<\/code> as your loss function, you should instead use <code>categorical_crossentropy<\/code>. <\/p>\n\n<p>The basic difference between these is that <code>sparse_categorical_crossentropy<\/code> works when you have direct integers as your label, and <code>categorical_crossentropy<\/code> works when you have one-hot encoded labels.<\/p>\n\n<p><strong>Corrected:<\/strong><\/p>\n\n<pre><code>inp = open('train.csv',\"rb\")\nX = pickle.load(inp)\nX = X\/255.0\nX = np.array(X)\nmodel = keras.Sequential([\n    keras.layers.Flatten(input_shape=(113, 75, 3)),\n    keras.layers.Dense(75, activation=tf.nn.relu),\n    keras.layers.Dense(50, activation=tf.nn.relu),\n    keras.layers.Dense(75, activation=tf.nn.relu),\n    keras.layers.Dense(25425, activation=tf.nn.softmax),\n    keras.layers.Reshape((113, 75, 4))\n])\nmodel.compile(optimizer='adam',\n              loss='categorical_crossentropy',\n              metrics=['accuracy'])\nmodel.fit(X, X, epochs=5)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57608602,"AnswerCount":0,"Tags":"<python><tensorflow><keras><scikit-learn><joblib>","CreationDate":"2019-08-22T11:44:35.510","AcceptedAnswerId":null,"OwnerUserId":10694594.0,"Title":"Is there a way to use return_estimator=True in cross_validate with parallel execution of sklearn-wrapped keras models?","Body":"<p>I love Keras for its simplicity and combining it with scikit-learn makes cross-validation of simple models incredibly easy.  I have successfully used <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.model_selection.cross_validate.html\" rel=\"nofollow noreferrer\"><code>sklearn.model_selection.cross_validate<\/code><\/a> in combination\nwith <a href=\"https:\/\/keras.io\/scikit-learn-api\/\" rel=\"nofollow noreferrer\">Keras' wrappers for the sklearn API<\/a>.<\/p>\n\n<p>Now, I also wanted to utilize the parallel execution capabilites of sklearn and stumbled into a problem when setting <code>return_estimator=True<\/code>.<\/p>\n\n<p>Here is an MWE that produces a <code>TypeError: can't pickle _thread.RLock objects<\/code>:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nfrom sklearn.metrics import make_scorer, log_loss\nfrom sklearn.model_selection import cross_validate\n\nx_data = np.random.normal(size=(1000, 200))\ny_data = np.random.randint(0, 2, size=1000)\n\ndef builder(shape=x_data[0].shape):\n    import tensorflow as tf \n    tf.enable_eager_execution()\n    # uncomment these lines for tensorflow-gpu\n    # config = tf.ConfigProto()\n    # config.gpu_options.allow_growth = True\n    # tf.keras.backend.set_session(tf.Session(config=config))\n\n    model = tf.keras.Sequential()\n    model.add(tf.keras.layers.Dense(100, input_shape=shape))\n    model.add(tf.keras.layers.Dense(1, activation=\"sigmoid\"))\n    model.compile(loss=\"binary_crossentropy\", optimizer=\"sgd\")\n\n    return model\n\nclassifier = tf.keras.wrappers.scikit_learn.KerasClassifier(build_fn=builder)\nresults = cross_validate(classifier,x_data, y_data, cv=5, n_jobs=3,\n                         return_train_score=False, verbose=10,\n                         return_estimator=True, scoring=make_scorer(log_loss))\n<\/code><\/pre>\n\n<p>I am not quite sure, where the error is raised, the RemoteTraceback doesn't tell. Is what I'm trying even possible?\nRunning the code above with <code>n_jobs=1<\/code> is fine (obviously), but I would really like to utilize the integrated parallelism in sklearn. <\/p>\n\n<p>What other solutions should I consider? I know I could use the <code>ModelCheckpoint<\/code> callback, store the weights on disk and load them afterwards. But I can't see a way to set separate filenames for every split during cross validation.<\/p>\n\n<p>I am using tensorflow 1.13.1 (which uses keras 2.2.4). Any help is much appreciated!<\/p>\n","answers":[]}
+{"QuestionId":57608991,"AnswerCount":0,"Tags":"<python><tensorflow><visualization>","CreationDate":"2019-08-22T12:05:55.767","AcceptedAnswerId":null,"OwnerUserId":7332429.0,"Title":"How to do activation maximization with self-pretrained model when the gradient of a filter w.r.t. the input is [None]","Body":"<p>I want to reload my pretrained model and visualize the kernels through activation maximization (AM) in tensorflow. But when I load the model, the tf.gradients(, input) is None. So they are unrelated. I don't know what I missed here.<\/p>\n\n<p>I am following the tutorial here <a href=\"https:\/\/github.com\/Hvass-Labs\/TensorFlow-Tutorials\/blob\/master\/13_Visual_Analysis.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Hvass-Labs\/TensorFlow-Tutorials\/blob\/master\/13_Visual_Analysis.ipynb<\/a>. And instead of load the inception model I am loading my own.<\/p>\n\n<p>'''python<\/p>\n\n<pre><code>def optimize_image(sess, input=None, restore_from=None\n               conv_id=None, layer_id2max=0,\n               num_iterations=30, show_progress=True,\n               args=None):\n\n    # build the graph from the function call of mod.construct_model\n    model_aspects = mod.construct_model(input, args, is_train=False)\n\n    # Load the pretrained model.\n    saver = tf.train.Saver()\n    data_io.load_model(saver, sess, restore_from)\n    #\n    # Reference to the tensor that takes the raw input image.\n    input = model_aspects[\"input\"]\n\n    # Reference to the tensor for the predicted classes.\n    y_pred = model_aspects[\"gap_w\"]\n\n    # Create the loss-function that must be maximized.\n    if conv_id is None:\n        loss = y_pred[0, layer_id2max]\n   else:\n        tensor = model_aspects[\"kernels\"]                    [list(model_aspects[\"kernels\"].keys())[conv_id]]\n\n\n        with tf.get_default_graph().as_default():\n            loss = tf.reduce_mean(tensor[:, :, :, layer_id2max])\n\n    gradient = tf.gradients(ys=loss, xs=model_aspects[\"input\"])\n\n    ...... (omit the rest)\n    return max_input\n\n\nif __name__ == \"__main__\":\n    restore_from = \"&lt;pretrained-model-dir&gt;\"\n    args = get_all_params(restore_from=restore_from)\n\n    features = tf.placeholder(tf.float32, [1, 2560, 1])\n    with tf.Session() as sess:\n        image = optimize_image(sess, input=features,\n                           restore_from=restore_from \n                           conv_id=0,\n                           layer_id2max=0,\n                           num_iterations=50,\n                           show_progress=True,\n                           args=args)\n<\/code><\/pre>\n\n<p>'''<\/p>\n\n<p>the graph is built, since tf.gradients(model_aspect[\"logits\"], input) exists and is tf.Tensor gradients_2\/convstart\/conv2d\/Conv2D_grad\/tuple\/control_dependency:0' shape=(1, 2560, 1, 1) dtype=float32>, )].\nBut tf.gradients(ys=loss, xs=input) is [None]<\/p>\n","answers":[]}
+{"QuestionId":57609112,"AnswerCount":1,"Tags":"<python><windows><tensorflow><cx-freeze>","CreationDate":"2019-08-22T12:13:33.980","AcceptedAnswerId":null,"OwnerUserId":11962330.0,"Title":"Problem when using cx_Freeze: \"cannot import name 'tf2'\"","Body":"<p>I have a code in python and I used cx_Freeze to convert it to an .exe. This task works without any error.\nBut when I try to run my .exe the following error happens:<\/p>\n\n<blockquote>\n  <p>from tensorflow.python import tf2<br>\n  ImportError: cannot import name 'tf2'<\/p>\n<\/blockquote>\n\n<p>My <code>ann.py<\/code> code is:<\/p>\n\n<pre><code>import numpy as np\nimport sys\n<\/code><\/pre>\n\n<p>.\n.\n.    <\/p>\n\n<pre><code>X_test=XinN\nY_test=XoutN\n\n\n#Criando o modelo\nfrom keras.models import Sequential\nfrom keras.layers import Dense\n\nmodelo = Sequential()\n\nfor i in range(int((num_par-4)\/2)):\n   modelo.add(Dense(int(parametros[i+4]), kernel_initializer='normal',activation=ativacao(int(parametros[i+5])))) #camadas ocultas\nmodelo.add(Dense(num_out, kernel_initializer='normal',activation=ativacao(int(parametros[num_par-1])))) #camada de sa\u00edda\n\nmodelo.compile(optimizer='adam',loss='mean_squared_error')\n\nhist = modelo.fit(X_train, Y_train, epochs=800, verbose=0, batch_size=10,validation_data=(X_test, Y_test))\n\nXobsoutN=modelo.predict(XobsN)\nXobsout=XobsoutN*(max_out-min_out)+min_out\nnp.savetxt(\"Xobsout.txt\",Xobsout.transpose(),delimiter='\\t')\n\nloss=[\" \",\" \"]\nloss[0] = hist.history['loss']\nloss[1] = hist.history['val_loss']\n\nnp.savetxt(\"erro.txt\",loss,delimiter='\\t')\n<\/code><\/pre>\n\n<p>And my <code>setyp.py<\/code> for cx_Freeze is:<\/p>\n\n<pre><code>from cx_Freeze import setup, Executable\nimport sys\n\nbase = None\n\nif sys.platform == 'win32':\n    base = None\n\n\nexecutables = [Executable(\"ANN.py\", base=base)]\n\npackages = [\"idna\"]\n\noptions = {\n    'build_exe': {\n        'includes':['atexit', 'numpy.core._methods', 'numpy.lib.format'],\n        'packages':packages,\n    },\n\n}\n\nimport os\n\nos.environ['TCL_LIBRARY'] = \"C:\\\\ProgramData\\\\Anaconda3\\\\tcl\\\\tcl8.6\"\nos.environ['TK_LIBRARY'] = \"C:\\\\ProgramData\\\\Anaconda3\\\\tcl\\\\tk8.6\"\n\nsetup(\n    name = \"Nome Executavel\",\n    options = options,\n    version = \"1.0\",\n    description = 'Descricao do seu arquivo',\n    executables = executables\n)\n<\/code><\/pre>\n\n<p>Anyone can help me to solve this error?<\/p>\n\n<p>I had many others errors using cx_Freeze and this forum was pretty helpful to solve all of them. Thanks a lot!<\/p>\n","answers":[{"AnswerId":"57620407","CreationDate":"2019-08-23T05:40:26.373","ParentId":null,"OwnerUserId":"8516269","Title":null,"Body":"<p>Try to add <code>\"tensorflow\"<\/code> to the <code>packages<\/code> list in your <code>setup.py<\/code> script:<\/p>\n\n<pre><code> packages = [\"idna\", \"tensorflow\"]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57609316,"AnswerCount":1,"Tags":"<tensorflow><quantization>","CreationDate":"2019-08-22T12:26:06.380","AcceptedAnswerId":null,"OwnerUserId":7830286.0,"Title":"Does TFLiteConverter automatically quantize the Keras model?","Body":"<p>I converted the trained Keras model using tf.lite.TFLiteConverter into tflite_model. Is the converted tflite_model quantized one? Here is the snippet to make the conversion.<\/p>\n\n<pre><code>import tensorflow as tf\n\nkeras_model = \".\/Trained_Models\/h_vs_o_a_V1.h5\"\n\nconverter = tf.lite.TFLiteConverter.from_keras_model_file(keras_model)\ntflite_model = converter.convert()\nopen(\"converted_model.tflite\", \"wb\").write(tflite_model)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57625066","CreationDate":"2019-08-23T11:04:32.083","ParentId":null,"OwnerUserId":"10878733","Title":null,"Body":"<p>Basically, we need the <code>converter.post_training_quantize = True<\/code> flag before the conversion takes place using <code>converter.convert()<\/code> like,<\/p>\n\n<pre><code>import tensorflow as tf\n\nkeras_model = \".\/Trained_Models\/h_vs_o_a_V1.h5\"\n\nconverter = tf.lite.TFLiteConverter.from_keras_model_file(keras_model)\nconverter.post_training_quantize = True\ntflite_model = converter.convert()\nopen(\"converted_model.tflite\", \"wb\").write(tflite_model)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57609381,"AnswerCount":0,"Tags":"<tensorflow><keras><base64><serve><serving>","CreationDate":"2019-08-22T12:29:14.593","AcceptedAnswerId":null,"OwnerUserId":11470044.0,"Title":"TF Serve: How to predict on base64 encoded images?","Body":"<p>I have been successful in serving a model with tensorflow serve, and getting predictions from it. However, for efficiency, I would like to try base64 encodings of the input images. I have tried the following, but was unable to make it work. Could anyone point me in the right direction?<\/p>\n\n<p>I load an already trained model and prepare it for serving with TF Serve:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import keras\nimport tensorflow\nfrom keras import backend as K\n\nK.set_learning_phase(0)\n\n# model = load_model()\n\nif fixed_shape is not None:\n    shape = [None, *fixed_shape]\nelse:\n    shape = [None, None, None, 3]\n\nstring_inp = tensorflow.placeholder(\n    tensorflow.string, shape=(None,))  # string input for the base64 encoded image\nimgs_map = tensorflow.map_fn(\n    tensorflow.image.decode_image,\n    string_inp,\n    dtype=tensorflow.float32\n)  # decode the jpeg\nimgs_map.set_shape(shape)\npred = model(imgs_map)\n\nmodel_inputs = list(string_inp)\nmodel_outputs = list(pred)\ninputs = {i.name + \"_bytes\": i for i in model_inputs}\noutputs = {f\"output {i}:\" + o.name: o for i, o in enumerate(model_outputs)}\nsignature = tensorflow.saved_model.signature_def_utils.predict_signature_def(\n    inputs=inputs, outputs=outputs\n)\n\nbuilder = tensorflow.saved_model.builder.SavedModelBuilder(\"out_dir\")\nbuilder.add_meta_graph_and_variables(\n    sess=K.get_session(),\n    tags=[tensorflow.saved_model.tag_constants.SERVING],\n    signature_def_map={\n        tensorflow.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:\n            signature\n    })\nbuilder.save()\n<\/code><\/pre>\n\n<p>Then for the client I use<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>image_byte = base64.b64encode(batch[0]).decode(\"utf-8\")\npayload = '{\"instances\" : [{\"b64\": \"%s\"}]}' % image_byte\nresponse = requests.post(predict_url, data=payload)\n<\/code><\/pre>\n\n<p>which results in <code>requests.exceptions.HTTPError: 400 Client Error: Bad Request for url:<\/code>. However, if I try the same request in Postman, I get:<\/p>\n\n<pre><code>{\n    \"error\": \"2 root error(s) found.\\n  (0) Invalid argument: assertion failed: [Unable to decode bytes as JPEG, PNG, GIF, or BMP]\\n\\t [[{{node map_5\/while\/decode_image\/cond_jpeg\/cond_png\/cond_gif\/Assert_1\/Assert}}]]\\n\\t [[GroupCrossDeviceControlEdges_0\/map_5\/while\/decode_image\/cond_jpeg\/cond_png\/cond_gif\/DecodeBmp\/_4702]]\\n  (1) Invalid argument: assertion failed: [Unable to decode bytes as JPEG, PNG, GIF, or BMP]\\n\\t [[{{node map_5\/while\/decode_image\/cond_jpeg\/cond_png\/cond_gif\/Assert_1\/Assert}}]]\\n0 successful operations.\\n0 derived errors ignored.\"\n}\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57609589,"AnswerCount":0,"Tags":"<python><tensorflow><pickle>","CreationDate":"2019-08-22T12:41:00.777","AcceptedAnswerId":null,"OwnerUserId":2851685.0,"Title":"saving and restoring custom dnn classifier built using tensorflow for using it with RandomizedSearchCV","Body":"<p><strong>ipynb file of whole code is <a href=\"https:\/\/gist.github.com\/devbkhadka\/ab3a769ab3bb90ab92c6d604bd2e0348\" rel=\"nofollow noreferrer\">here<\/a><\/strong><\/p>\n\n<p>I have a custom DNN_Classifier class which looks like below<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class DNN_Classifier(BaseEstimator, TransformerMixin):\n    def __init__(self, n_hidden_layers=None, n_neurons=None, n_outputs=None, \n                 activation=tf.nn.elu, optimizer=tf.train.AdamOptimizer,  learning_rate=0.01, \n                 batch_norm_momentum=None, batch_size=50, dropout_rate=None):\n        ...\n\n\n    def _create_graph(self):                      \n\n\n        ...\n\n\n\n    def fit(self,x,y,x_val,y_val):\n\n        ...\n\n\n\n\n    def save(self,name):\n        path = \"checkpoints\/%s.chk\"%name\n        self._saver.save(self._session, path)\n\n    def restore(self, name):\n        self._create_graph()\n        self._session = tf.Session(graph=self._graph)\n        path = \"checkpoints\/%s.chk\"%name\n        self._saver.restore(self._session, path)\n<\/code><\/pre>\n\n<p>My use case is, I want to use this class with RandomizedSearchCV to find best hyper parameters. Then I want to save it for future comparison.<\/p>\n\n<p>I don't want to save the hyper parameters (constructor parameters) and the tensorflow session separately. The <code>save and restore<\/code> methods in above code snippet works only if I know hyper parameters when restoring. Please check <a href=\"https:\/\/gist.github.com\/devbkhadka\/ab3a769ab3bb90ab92c6d604bd2e0348\" rel=\"nofollow noreferrer\">ipynb<\/a> file for detail.<\/p>\n\n<p>I also tried using <code>joblib<\/code> and <code>dill<\/code> to save entire class as pickle but it is also not working <\/p>\n\n<p><strong>joblib<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from sklearn.externals import joblib\n\npath = \"checkpoints\/save_test_joblib.chk\"\njoblib.dump(classifier, path)\n<\/code><\/pre>\n\n<p><em>error<\/em><\/p>\n\n<pre><code>PicklingError: Can't pickle &lt;function get_leaky_relu.&lt;locals&gt;.&lt;lambda&gt; at 0x1c34758b70&gt;: it's not found as __main__.get_leaky_relu.&lt;locals&gt;.&lt;lambda&gt;\n<\/code><\/pre>\n\n<p><strong>dill<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import dill\npath = \"checkpoints\/save_test_dill.chk\"\nwith open(path, \"wb\") as fl:\n    dill.dump(classifier, fl)\n<\/code><\/pre>\n\n<p><em>error<\/em><\/p>\n\n<pre><code>TypeError: can't pickle SwigPyObject objects\n<\/code><\/pre>\n\n<p>What is the good way to save the classifier<\/p>\n\n<p><strong>Please find ipynb file for whole code <a href=\"https:\/\/gist.github.com\/devbkhadka\/ab3a769ab3bb90ab92c6d604bd2e0348\" rel=\"nofollow noreferrer\">here<\/a><\/strong><\/p>\n","answers":[]}
+{"QuestionId":57610681,"AnswerCount":1,"Tags":"<keras><time-series><lstm><multiclass-classification>","CreationDate":"2019-08-22T13:39:30.257","AcceptedAnswerId":null,"OwnerUserId":4929337.0,"Title":"What should be the shape of y-train for LSTM for multi-class classification, given non-textual sequential data?","Body":"<h3>Problem Description<\/h3>\n\n<p>I have a dataset (features=175, n_time_steps=954, Number of sequences=737). \nColumns 1-174 are features, the last target column contains 3 different classes. I would like to use LSTM for multi-class classification to predict only the last time step, i.e. use 953 steps and features to predict the class of the step 954. I am struggling with the structure of y_train input. I will appreciate any thoughts on how to reshape the y_train correctly for this problem.<\/p>\n\n<h3>Data<\/h3>\n\n<p>I have 737 products, each  has sales for 954 days. The target classes are (0- when the product did not existed, 1-product of typeA, 2 product of type B). I need to use 953 days and 174 features to predict the class of each product  at the last date of  sequence (954). Test set has 100 products, train set -637 products.<\/p>\n\n<p>After reshaping X_train has (637, 953, 175) shape. y_train has the shape of \n(637, 1). When I run to_categorical the shape is (637, 2). both y_train shapes throw errors when fited to the LSTM model.<\/p>\n\n<p>When I fit y_train of shape (637, 1), the error is<\/p>\n\n<pre><code>ValueError: You are passing a target array of shape (637, 1) while using as loss `categorical_crossentropy`. `categorical_crossentropy` expects targets to be binary matrices (1s and 0s) of shape (samples, classes). If your targets are integer classes, you can convert them to the expected format via:\n\nfrom keras.utils import to_categorical\ny_binary = to_categorical(y_int)\n\n\nAlternatively, you can use the loss function `sparse_categorical_crossentropy` instead, which does expect integer targets.\n<\/code><\/pre>\n\n<p>When I fit <code>to_categorical(y_train)<\/code> of shape (637, 2), the error is<\/p>\n\n<pre><code>ValueError: Error when checking target: expected dense_45 to have shape (1,) but got array with shape (2,)\n<\/code><\/pre>\n\n<p>When I change to 'sparse_categorical_crossentropy' and fit y_train of shape (637,1), the error is<\/p>\n\n<pre><code>InvalidArgumentError: Received a label value of 1 which is outside the valid range of [0, 1).  Label values: 0 1 0 1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 1 0 0 1 0 0 0 0 1 1 0 0 1 1 0 0 0 0 0 0 0 0 1 1 0 1 0 1 1 0 1 0 1 0 0 1 0 0 1 0 0 0 0 1 1 1 1 0 0 0 1 0 0 1 0 0 0 1 1 0 0 1 1 0 0 1 1 1 1 1 0 0 1 0 0 0 0 0 0 0 0 0\n     [[{{node loss_13\/dense_48_loss\/SparseSoftmaxCrossEntropyWithLogits\/SparseSoftmaxCrossEntropyWithLogits}}]] \n<\/code><\/pre>\n\n<p>Here is my model<\/p>\n\n<pre><code>model = Sequential([\n            LSTM(units=1024, \n            input_shape=(periods_to_train,features), kernel_initializer='he_uniform',\n            activation ='linear', kernel_constraint=maxnorm(3), return_sequences=False),\n            Dropout(rate=0.5),\n            Dense(units=1024,kernel_initializer='he_uniform', \n            activation='linear', kernel_constraint=maxnorm(3)),\n            Dropout(rate=0.5),\n            Dense(units=1024, kernel_initializer='he_uniform',\n            activation='linear', kernel_constraint=maxnorm(3)),\n            Dropout(rate=0.5),\n            Dense(units=periods_to_predict, kernel_initializer='he_uniform', activation='softmax')])\n\n        #Compile model\noptimizer = Adamax(lr=0.001, decay=0.1)\n\nmodel.compile(loss='sparse_categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])\n\nconfigure(gpu_ind=True)\nmodel.fit(X_train, y_train ,validation_split=0.1, batch_size=100, epochs=8, shuffle=True)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57611702","CreationDate":"2019-08-22T14:31:42.317","ParentId":null,"OwnerUserId":"10455493","Title":null,"Body":"<p>It seems like your understanding of the network is correct. So I recreated a minimal working example to generate data and train in the same manner as you do. When I set the time-steps (periods_to_train) equal to 953 as you have, I also get some weird errors. But it has been multiple study suggesting no more than 200 to 500 time-step dependency using LSTM, as the model output will start to \"forget\" earlier information.<\/p>\n\n<p>Below is an minimal working example code of what you try to do, using only 100 time-steps. No Errors in my case (tensorflow version 1.14.0):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport tensorflow.keras.backend as K\nimport numpy as np\n\n\ndata_size = 637\nperiods_to_train=100\nfeatures = 175\nperiods_to_predict = 3\n\nX_train=np.random.rand(data_size,periods_to_train,features)\ny_train=np.random.randint(0,3,data_size).reshape(-1,1)\n\nK.clear_session()\n\nmodel = tf.keras.models.Sequential([\n            tf.keras.layers.LSTM(\n                    units=1024, input_shape=(periods_to_train,features), kernel_initializer='he_uniform',\n                    activation ='linear', kernel_constraint=tf.keras.constraints.max_norm(3.), return_sequences=False),\n            tf.keras.layers.Dropout(rate=0.5),\n            tf.keras.layers.Dense(\n                    units=1024,kernel_initializer='he_uniform', \n                    activation='linear', kernel_constraint=tf.keras.constraints.max_norm(3)),\n            tf.keras.layers.Dropout(rate=0.5),\n            tf.keras.layers.Dense(\n                    units=1024, kernel_initializer='he_uniform',\n                    activation='linear', kernel_constraint=tf.keras.constraints.max_norm(3)),\n            tf.keras.layers.Dropout(rate=0.5),\n            tf.keras.layers.Dense(\n                    units=periods_to_predict, kernel_initializer='he_uniform', \n                    activation='softmax')])\n\n\noptimizer = tf.keras.optimizers.Adamax(lr=0.001, decay=0.1)\n\nmodel.compile(loss='sparse_categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])\n\nmodel.fit(X_train, y_train ,validation_split=0.1, batch_size=64, epochs=1, shuffle=True)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57610804,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-22T13:45:14.770","AcceptedAnswerId":57614277.0,"OwnerUserId":11354575.0,"Title":"When is the timing to use sample_weights in keras","Body":"<p>According to this <a href=\"https:\/\/stackoverflow.com\/questions\/43459317\/keras-class-weight-vs-sample-weights-in-the-fit-generator\">question<\/a>, I learnt that <code>class_weight<\/code> in <code>keras<\/code> is applying a weighted loss during training, and <code>sample_weight<\/code> is doing something sample-wise if I don't have equal confidence in all the training samples.<\/p>\n\n<p>So my questions would be,<\/p>\n\n<ol>\n<li>Is the loss during validation weighted by the <code>class_weight<\/code>, or is it only weighted during training?<\/li>\n<li>My dataset has 2 classes, and I don't actually have a seriously imbalanced class ditribution. The ratio is approx. 1.7 : 1. Is that neccessary to use <code>class_weight<\/code> to balance the loss or even use oversampling? Is that OK to leave the slightly imbalanced data as the usual dataset treated?<\/li>\n<li>Can I simply consider <code>sample_weight<\/code> as the weights I give to each train sample?  And my trainig samples can be treated with equal confidence, so I probably I don't need to use this.<\/li>\n<\/ol>\n","answers":[{"AnswerId":"57614277","CreationDate":"2019-08-22T17:20:30.840","ParentId":null,"OwnerUserId":"10455493","Title":null,"Body":"<ol>\n<li>From the keras documentation it says<\/li>\n<\/ol>\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<\/blockquote>\n\n<p>So <code>class_weight<\/code> does only affect the loss during traning. I myself have been interested in understanding how the class and sample weights is handled during testing and training. Looking at the keras github repo and the code for metric and loss, it does not seem that either loss or metric is affected by them. The printed values are quite hard to track in the training code like <code>model.fit()<\/code> and its corresponding tensorflow backend training functions. So I decided to make a test code to test the possible scenarios, see code below. The conclusion is that both <code>class_weight<\/code> and <code>sample_weight<\/code> only affect training loss, no effect on any metrics or validation loss. A little surprising as <code>val_sample_weights<\/code> (which you can specify) seems to do nothing(??).<\/p>\n\n<ol start=\"2\">\n<li><p>This types of question always depends on you problem, how skewed the date is and in what way you try to optimize the model. Are you optimizing for accuracy, then as long as the training data is equally skewed as when the model is in production, the best result will be achieved just training without any over\/under sampling and\/or class weights. \nIf you on the other hand have something where one class is more important (or expensive) than another then you should be weighting the data. For example in fraud prevention, where fraud normally is much more expensive than the income of non-fraud. I would suggest you try out unweighted classes, weighted classes and some under\/over-sampling and check which gives the best validation results. Use a validation function (or write your own) that best will compare different models (for-example weighting true-positive, false-positive, true-negative and false-negative differently dependent on cost).\nA relatively new loss-function that has shown great result at kaggle competitions on skewed data is <code>Focal-loss<\/code>. <code>Focal-loss<\/code> reduce the need for over\/under-sampling. Unfortunately <code>Focal-loss<\/code> is not a built inn function in keras (yet), but can be manually programmed.<\/p><\/li>\n<li><p>Yes I think you are correct. I normally use <code>sample_weight<\/code> for two reasons. 1, the training data have some kind of measuring uncertainty, which if known can be used to weight accurate data more than inaccurate measurements. Or 2, we can weight newer data more than old, forcing the model do adapt to new behavior more quickly, without ignoring valuable old data.<\/p><\/li>\n<\/ol>\n\n<p>The code for comparing with and without <code>class_weights<\/code> and <code>sample_weights<\/code>, while holding the model and everything else static.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\ndata_size = 100\ninput_size=3\nclasses=3\n\nx_train = np.random.rand(data_size ,input_size)\ny_train= np.random.randint(0,classes,data_size )\n#sample_weight_train = np.random.rand(data_size)\nx_val = np.random.rand(data_size ,input_size)\ny_val= np.random.randint(0,classes,data_size )\n#sample_weight_val = np.random.rand(data_size )\n\ninputs = tf.keras.layers.Input(shape=(input_size))\npred=tf.keras.layers.Dense(classes, activation='softmax')(inputs)\n\nmodel = tf.keras.models.Model(inputs=inputs, outputs=pred)\n\nloss = tf.keras.losses.sparse_categorical_crossentropy\nmetrics = tf.keras.metrics.sparse_categorical_accuracy\n\nmodel.compile(loss=loss , metrics=[metrics], optimizer='adam')\n\n# Make model static, so we can compare it between different scenarios\nfor layer in model.layers:\n    layer.trainable = False\n\n# base model no weights (same result as without class_weights)\n# model.fit(x=x_train,y=y_train, validation_data=(x_val,y_val))\nclass_weights={0:1.,1:1.,2:1.}\nmodel.fit(x=x_train,y=y_train, class_weight=class_weights, validation_data=(x_val,y_val))\n# which outputs:\n&gt; loss: 1.1882 - sparse_categorical_accuracy: 0.3300 - val_loss: 1.1965 - val_sparse_categorical_accuracy: 0.3100\n\n#changing the class weights to zero, to check which loss and metric that is affected\nclass_weights={0:0,1:0,2:0}\nmodel.fit(x=x_train,y=y_train, class_weight=class_weights, validation_data=(x_val,y_val))\n# which outputs:\n&gt; loss: 0.0000e+00 - sparse_categorical_accuracy: 0.3300 - val_loss: 1.1945 - val_sparse_categorical_accuracy: 0.3100\n\n#changing the sample_weights to zero, to check which loss and metric that is affected\nsample_weight_train = np.zeros(100)\nsample_weight_val = np.zeros(100)\nmodel.fit(x=x_train,y=y_train,sample_weight=sample_weight_train, validation_data=(x_val,y_val,sample_weight_val))\n# which outputs:\n&gt; loss: 0.0000e+00 - sparse_categorical_accuracy: 0.3300 - val_loss: 1.1931 - val_sparse_categorical_accuracy: 0.3100\n<\/code><\/pre>\n\n<p>There are some small deviations between using weights and not (even when all weights are one), possible due to fit using different backend functions for weighted and unweighted data or due to rounding error? <\/p>\n"}]}
+{"QuestionId":57611085,"AnswerCount":1,"Tags":"<python-3.x><keras><nlp><lstm><seq2seq>","CreationDate":"2019-08-22T13:58:59.417","AcceptedAnswerId":null,"OwnerUserId":11641873.0,"Title":"How to save and reload hidden states of Keras encoder-decoder model for inference","Body":"<p>I'm working on a sequence to sequence text generation using the encoder decoder architecture with LSTM layers in Keras. <br\/>\nMy code is a modified version of the Keras LSTM seq2seq tutorial: \n<a href=\"https:\/\/keras.io\/examples\/lstm_seq2seq\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/examples\/lstm_seq2seq\/<\/a> <br\/>\nAfter training the model, I want to set up an inference mode where I can decode a sequence using only the hidden states of the pretained model, without retraining the whole model each time.\nI'm looking for a way to save these hidden LSTM states to a file and then reload them to set up my inference model.<\/p>\n\n<p>A fraction of the model training part looks like this:<\/p>\n\n<pre><code>encoder = LSTM(latent_dim, return_state=True)\nencoder_outputs, state_h, state_c = encoder(encoder_inputs)\n<\/code><\/pre>\n\n<p>I will later need the hidden cell states and encoder outputs (encoder_outputs, state_h and state_c) to set up the inference model (as in the tutorial) in the same way:<\/p>\n\n<pre><code>encoder_model = Model(inputs=encoder_inputs,outputs=[encoder_outputs, state_h, state_c])\n<\/code><\/pre>\n\n<p>Reusing the variables when having one program works fine, however I want to split my code into training and inference. I would then need to find a way to save the hidden cell states to a file.<\/p>\n\n<p>What I've tried so far is <\/p>\n\n<ul>\n<li><p>saving state_h and state_c, which are Tensors, directly to file, but reshaping then from Tensor to a numpy array did not work for me<\/p><\/li>\n<li><p>saving the whole Keras model, loading it in inference mode and extracting the hidden cell states via<\/p><\/li>\n<\/ul>\n\n<pre><code>from tensorflow.keras.models import load_model\nmodel.save(\"current_model.h5\")\nnewmodel = load_model('.\/current_model.h5', custom_objects={'AttentionLayer': AttentionLayer})\n\nencoder_outputs_2, state_h_2, state_c_2 = newmodel.layers[7].output\n<\/code><\/pre>\n\n<p>Reassigning the variables in that way did work, but not feeding them into inference model: that gave me the error message <\/p>\n\n<pre><code>ValueError: Graph disconnected: cannot obtain value for tensor     Tensor(\"input_1_2:0\", shape=(?, 30), dtype=float32) at layer \"input_1\". The following previous layers were accessed without issue: []\n<\/code><\/pre>\n\n<p>Any help with what I've tried so far or ideas for new approaches would be appreciated. Cheers!<\/p>\n","answers":[{"AnswerId":"57728868","CreationDate":"2019-08-30T14:42:07.537","ParentId":null,"OwnerUserId":"11641873","Title":null,"Body":"<p>Solved the issue: the Graph disconnected error occurred due to referencing an incorrect instance of the stacked LSTM layers, should have been newmodel.layers[6].output<\/p>\n"}]}
+{"QuestionId":57611976,"AnswerCount":2,"Tags":"<python><machine-learning><theano><theano-cuda>","CreationDate":"2019-08-22T14:45:27.737","AcceptedAnswerId":null,"OwnerUserId":11253858.0,"Title":"ImportError: cannot import name as_cuda_ndarray_variable","Body":"<p>While I am trying to call <code>as_cuda_ndarray_variable<\/code> function from <code>from theano.sandbox.cuda<\/code> that is wrote on separate <code>basic_ops.py<\/code>python file that is called inside the init.py file. My <code>python-2.7.16<\/code> and<code>theano-0.9.0<\/code>.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from theano.sandbox.cuda import as_cuda_ndarray_variable\n<\/code><\/pre>\n\n<pre><code>ImportError: cannot import name as_cuda_ndarray_variable\n<\/code><\/pre>\n","answers":[{"AnswerId":"57622873","CreationDate":"2019-08-23T08:55:36.867","ParentId":null,"OwnerUserId":"9957379","Title":null,"Body":"<p>It appears that the function <code>as_cuda_ndarray_variable<\/code> is not defined directly within the <code>theano.sandbox.cuda<\/code> module.  Try this instead:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from theano.sandbox.cuda import basic_ops\n\nbasic_ops.as_cuda_ndarray_variable(1.0)\n<\/code><\/pre>\n\n<p>See the example here: <a href=\"http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/software\/theano\/tutorial\/using_gpu.html<\/a><\/p>\n"},{"AnswerId":"57612390","CreationDate":"2019-08-22T15:08:58.853","ParentId":null,"OwnerUserId":"6129428","Title":null,"Body":"<p>Maybe you can try this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from theano.sandbox.cuda import * as cuda_ndarray_variable\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import theano.sandbox.cuda as cuda_ndarray_variable\n<\/code><\/pre>\n"}]}
+{"QuestionId":57611988,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><keras><conv-neural-network>","CreationDate":"2019-08-22T14:46:09.303","AcceptedAnswerId":null,"OwnerUserId":11514342.0,"Title":"CNN with vector output and 2D image graph input (input is an array)","Body":"<p>I am trying to create a CNN in Keras (Python 3.7) which ingests a 2D matrix input (much like a grayscale image) and outputs a 1 dimensional vector. So far I did manage to get results, but I am not sure if what I am doing is correct (or if my intuition is). <\/p>\n\n<p>I input a 100x50 array into my convolutional layer. This 2D array holds the peak information at every position (ie. x axis pertains to the position, y-axis pertains to the frequency, and each cell gives the intensity). The 3D graph of this shows something akin to the one given in <a href=\"https:\/\/www.researchgate.net\/figure\/a-3D-and-b-top-view-Contour-plot-of-the-measured-Brillouin-gain-spectrum-BGS_fig8_321056806\" rel=\"nofollow noreferrer\">this<\/a> link.<\/p>\n\n<p>From the (all of the) literature I have read, I learned that CNN accepts image data--image is converted into pixel values and then repeatedly convolved and pooled to get the output. However, I am using a MatLab simulator to get my input data, and I have access to the raw 2D array containing information on the peak frequency at each point.<\/p>\n\n<p>My intuition is this: if we normalize each cell and feed the information to the CNN, it will be as if I fed the normalized pixel values of the image to the CNN, since my raw 2D array also has height, width and depth=1, like an image.<\/p>\n\n<p>Please enlighten me if my thinking is correct or wrong.<\/p>\n\n<p>My code is as follows:<\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nfrom pandas import Series, DataFrame\nimport matplotlib.pyplot as plt\n%matplotlib inline\nimport tensorflow as tf\nimport keras\n\n'''load sample input'''\nBGS1 = pd.read_csv(\"C:\/Users\/strain1_input.csv\")\nBGS2 = pd.read_csv(\"C:\/Users\/strain2_input.csv\")\nBGS3 = pd.read_csv(\"C:\/Users\/strain3_input.csv\")\nBGS_ = np.array([BGS1, BGS2, BGS3]) #3x100x50 array\n\nBGS_normalized = BGS_\/np.amax(BGS_)\n\n'''load sample output'''\nBFS1 = pd.read_csv(\"C:\/Users\/strain1_output.csv\")\nBFS2 = pd.read_csv(\"C:\/Users\/strain2_output.csv\")\nBFS3 = pd.read_csv(\"C:\/Users\/strain3_output.csv\")\nBFS_ = np.array([BFS1, BFS2, BFS3]) #3x100\nBFS_normalized = BFS\/50 #since max value for each cell is 50\n\n#after splitting data into training, validation and testing sets,\noutput_nodes = 100 \nn_classes = 1\nbatch_size_ = 8 #so far, optimized for 8 batch size\nepoch = 100 \n\ninput_layer = Input(shape=(45,300,1))\nconv1 = Conv2D(16,3,padding=\"same\",activation=\"relu\", input_shape = \n(45,300,1))(input_layer)\npool1 = MaxPooling2D(pool_size=(2,2),padding=\"same\")(conv1)\nflat = Flatten()(pool1)\nhidden1 = Dense(10, activation='softmax')(flat) #relu\n\nbatchnorm1 = BatchNormalization()(hidden1) \noutput_layer = Dense(output_nodes*n_classes, activation=\"softmax\")(batchnorm1) \noutput_layer2 = Dense(output_nodes*n_classes, activation=\"relu\")(output_layer) \noutput_reshape = Reshape((output_nodes, n_classes))(output_layer2)\nmodel = Model(inputs=input_layer, outputs=output_reshape)\n\nprint(model.summary())\nmodel.compile(loss='mean_squared_error', optimizer='adam', sample_weight_mode='temporal')\n\nmodel.fit(train_X,train_label,batch_size=batch_size_,epochs=epoch)\npredictions = model.predict(train_X)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57616526","CreationDate":"2019-08-22T20:18:11.730","ParentId":null,"OwnerUserId":"11919358","Title":null,"Body":"<p>what you did is exactly the strategy used to input non image data in to 2d convolutional layers. As long the model predicts correctly, what you did is correct. its just that CNN perform very poorly on non-image data or there might be chances to overfit. But then again, as long it performs correctly then its good.<\/p>\n"}]}
+{"QuestionId":57612125,"AnswerCount":0,"Tags":"<python><tensorflow><optimization><convolution>","CreationDate":"2019-08-22T14:54:56.577","AcceptedAnswerId":null,"OwnerUserId":11958764.0,"Title":"How can I speed up my custom convolution?","Body":"<p>I'm trying to implement a custom convolution function that singles out every instance of multiplication and randomly applies errors to the multiplicands. I am currently trying to do this by duplicating the elements of my filter and image by the number of times it is used in a multiplication, applying errors to each term, doing elementwise multiplication, followed by summation.<\/p>\n\n<p>However, this function becomes very slow for large inputs (as would happen in VGG16), especially if I have to apply it to many filters at once. Even without the part that randomly generates errors for the multiplicands, the function is orders of magnitude slower than the implementation in tensorflow.<\/p>\n\n<p>The code is attached below. Help for speeding this up would be appreciated.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def bit_error_int_convolve(kernel,input,error_rate):\n    \"\"\"\n    2D Convolution for 8 bit integer image and kernel with bit flip error rate of error_rate\n    :param kernel: 8 bit integer tensor of shape (filter size (row), filter size (col), channel size)\n    :param input: 8 bit integer tensor of shape (batch size, image size (row), image size (col), channel size)\n    :param error_rate: Rate of bit flip errors\n    :return: 32 bit float tensor of shape (batch size, output image size (row), output image size (col))\n    \"\"\"\n\n    kernel_shape = tf.shape(kernel)\n    input_shape = tf.shape(input)\n\n    # Copies kernel for number of times it is used to rank 4 tensor of shape\n    # (batch size, output image size (row), output image size (col), filter size (row), filter size (col), channel number)\n    mK = tf.reshape(kernel,(kernel_shape[0]*kernel_shape[1]*kernel_shape[2],))\n    mK = tf.tile(mK, [input_shape[0]*(input_shape[1]-kernel_shape[0]+1)*(input_shape[2]-kernel_shape[1]+1)])\n    mK = tf.reshape(mK,(input_shape[0],input_shape[1]-kernel_shape[0]+1,input_shape[2]-kernel_shape[1]+1,kernel_shape[0],kernel_shape[1],kernel_shape[2]))\n\n    # Reshapes and copies elements of image to rank 4 tensor of shape\n    # (batch size, output image size (row), output image size (col), filter size (row), filter size (col), channel number)\n    rep = input[:,:kernel_shape[0],:kernel_shape[1],:]\n\n    for k in range((tf.shape(mK)[1])*(tf.shape(mK)[2])-1):\n        j = (k+1)%(tf.shape(mK)[2])\n        i = (k-j+1)\/\/(tf.shape(mK)[1])\n        thing = input[:,i:i+kernel_shape[0],j:j+kernel_shape[1],:]\n        rep = tf.concat([rep,thing],0)\n\n    rep = tf.reshape(rep, (tf.shape(mK)[1], tf.shape(mK)[2], tf.shape(mK)[0], tf.shape(mK)[3], tf.shape(mK)[4], tf.shape(mK)[5]))\n    rep = tf.transpose(rep, perm=[2, 0, 1, 3, 4, 5])\n\n    # Applies bit flip errors to multiplicands with rate error_rate\n    # Slow even if left out\n    # if error_rate!=0:\n    #     rep = bit_errors(rep,error_rate)\n    #     mK = bit_errors(mK,error_rate)\n\n    # Elementwise multiplication followed by summation to find output image\n    rep = K.cast(rep,dtype=tf.float32)\n    mK = K.cast(mK,dtype=tf.float32)\n    output = rep*mK\n    output = K.sum(output, axis=(3, 4, 5))\n\n    return output\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57612429,"AnswerCount":1,"Tags":"<python><python-3.x><numpy><tensorflow>","CreationDate":"2019-08-22T15:10:47.770","AcceptedAnswerId":57613743.0,"OwnerUserId":103254.0,"Title":"Divide a list of tensors by a list of scalars in tensorflow?","Body":"<p>Is there a handy way to divide a list of tensors by a list of scalars? I'm trying to do something similar to the following, but get the indicated error on the last line:<\/p>\n\n<pre><code>import tensorflow as tf\ntf.__version__\n# '1.13.1'\nimport numpy as np\nds1 = tf.data.Dataset.from_tensor_slices(np.random.random([10, 3, 4])).batch(5, drop_remainder=True)\ni = ds1.make_one_shot_iterator()\nn = i.get_next()\nn.shape\n# TensorShape([Dimension(5), Dimension(3), Dimension(4)])\nvar = tf.Variable([1,2,3,4,5], dtype=np.float64)\nop = n\/var\n# Traceback ....\n# ValueError: Dimensions must be equal, but are 4 and 5 for 'truediv' (op: 'RealDiv') with input shapes: [5,3,4], [5].\n<\/code><\/pre>\n\n<p>My desired result is a list of tensors of shape [5, 3, 4] where the entries in the first are divided by 1, in the second by 2, in the third by 3, and so on. (The values 1-5 are standing in for computed values in my actual code.)<\/p>\n\n<p>I'm pretty sure that the answer is going to be something easy, but I can't find the right set of search keywords to get SO or Google to cooperate.<\/p>\n","answers":[{"AnswerId":"57613743","CreationDate":"2019-08-22T16:37:20.457","ParentId":null,"OwnerUserId":"103254","Title":null,"Body":"<p>As suggested by a commenter, reshaping <code>var<\/code> works:<\/p>\n\n<pre><code>op = n\/tf.reshape(var, (5, 1, 1))\n<\/code><\/pre>\n\n<p>This yields the expected result. Notionally, I was originally asking too much in the way of shape inference from Tensorflow.<\/p>\n"}]}
+{"QuestionId":57612543,"AnswerCount":1,"Tags":"<python><tensorflow><cosine-similarity>","CreationDate":"2019-08-22T15:18:15.470","AcceptedAnswerId":57612926.0,"OwnerUserId":7934786.0,"Title":"how to perform similarity function over columns of a tensor in tensorflow","Body":"<p>I have a tensor like this:<\/p>\n\n<pre><code>tf_a1 =      [[-0.65 0.   0.   0.   0.42  0.   0.   0.51 0.   0.34 0.]\n              [0.   -0.51 0.   0.  -0.52  0.   0.   0.   0.53 0.42 0.]\n              [0.    0.32 0.  -0.50 0.34  0.   0.   0.39 0.32 0.52 0.]\n              [0.    0.23 0.37 0.   0.    0.37 0.37 0.   0.47 0.39 0.3 ]]\n<\/code><\/pre>\n\n<p>I want to apply <code>cosine similarity<\/code> over each column of this tensor. So, I want to find the similarity of the first column versus rest of the columns. Again, second column against rest of the columns and so on.<\/p>\n\n<p>I have done this using the for loop as such:<\/p>\n\n<pre><code>def cosine_score(x):\n    for i, arr in enumerate(x):\n        if i == 0 :\n            first = cosine_similarity(x[i,].reshape(1, -1), x)\n        else:\n            second = cosine_similarity(x[i,].reshape(1, -1), x)\n            final = tf.concat((first, second), axis=0)\n            first = final\n    return final\nsim_topics = cosine_score(tf_a1)\n<\/code><\/pre>\n\n<p>Now, When I want to include this in my model, I can not use foor loop as it is. seems I have to use  <code>tf.map_fn<\/code> to go over it.<\/p>\n\n<p>I also have done like this:<\/p>\n\n<pre><code>def cosine_score(x):\n    def cos_similarity(col):\n        for i, arr in enumerate(col):\n            if i == 0:\n                first = cosine_similarity(col[i, ].reshape(1, -1), col)\n            else:\n                second = cosine_similarity(col[i, ].reshape(1, -1), col)\n                final = tf.concat((first, second), axis=0)\n                first = final\n        return final\n    sim = tf.map_fn(cos_similarity, x, dtype=tf.float32)\n    return sim\n<\/code><\/pre>\n\n<p>But here I need to remove the <code>for loop<\/code>. My problem is that if I remove <code>for loop<\/code> and access each column seperately, how can I access the rest of the columns to compare and apply <code>cosine similarity<\/code>.<\/p>\n\n<p>Please let me know if its not clear.<\/p>\n","answers":[{"AnswerId":"57612926","CreationDate":"2019-08-22T15:40:09.220","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>Cosine similarity is nothing more than an L2 normalized dot product. So, in <code>Tensorflow<\/code> this should do the trick for you:<\/p>\n\n<pre><code># Normalize the columns of the tensor\nnormalized_tensor = tf.math.l2_normalize(tf_a1, axis=0)\n# Get the dot product between the columns\nscores = tf.matmul(normalized_tensor, normalized_tensor, transpose_a=True)\n<\/code><\/pre>\n\n<p>The tensor <code>scores<\/code> contains the cosine similarity between the columns of <code>tf_a1<\/code>. Moreover, below is a <code>Numpy<\/code> equivalent implementation:<\/p>\n\n<pre><code># Normalize the columns of the tensor\nnormalized_tensor = tf_a1 \/ np.linalg.norm(tf_a1, axis=0)\n# Get the dot product between the columns\nscores = np.dot(normalized_tensor.T, normalized_tensor)\n<\/code><\/pre>\n\n<p>Finally, if you want to keep only one of the triangles (for example the upper triangle), and set the main diagonal to <code>0<\/code>, you can do the following in <code>Tensorflow<\/code>:<\/p>\n\n<pre><code>zero_diag = tf.linalg.set_diag(scores, tf.zeros(tf.shape(scores)[0]))\ntriangular = tf.matrix_band_part(zero_diag, 0, -1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57612610,"AnswerCount":1,"Tags":"<python><pytorch><google-colaboratory>","CreationDate":"2019-08-22T15:22:07.227","AcceptedAnswerId":null,"OwnerUserId":2287486.0,"Title":"How to solve \"No module named 'tools'\" in google colab?","Body":"<p>I am trying to follow this tutorial to apply Bert in my data: <a href=\"https:\/\/medium.com\/swlh\/a-simple-guide-on-using-bert-for-text-classification-bbf041ac8d04\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/swlh\/a-simple-guide-on-using-bert-for-text-classification-bbf041ac8d04<\/a> <\/p>\n\n<p>And Getting error on imports<\/p>\n\n<p>This is what I am importing<\/p>\n\n<pre><code>import torch\nimport pickle\nfrom torch.utils.data import (DataLoader, RandomSampler, SequentialSampler, TensorDataset)\nfrom torch.nn import CrossEntropyLoss, MSELoss\n\nfrom tqdm import tqdm_notebook, trange\nimport os\nfrom pytorch_pretrained_bert import BertTokenizer, BertModel, BertForMaskedLM, BertForSequenceClassification\nfrom pytorch_pretrained_bert.optimization import BertAdam, WarmupLinearSchedule\n\nfrom multiprocessing import Pool, cpu_count\nfrom tools import *\nimport convert_examples_to_features\n<\/code><\/pre>\n\n<p>and getting error on:  from tools import * as \"No module named 'tools'\".<\/p>\n\n<p>I tried to follow some answers like: \n<a href=\"https:\/\/stackoverflow.com\/questions\/51881295\/python-caffe2-importerror-no-module-named-tools-setup-helpers-env\">Python\/Caffe2: ImportError: No module named tools.setup_helpers.env<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/14108428\/pyserial-no-module-named-tools\">pyserial: No module named tools<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/56859803\/modulenotfounderror-no-module-named-tools-nnwrap\">ModuleNotFoundError: No module named &#39;tools.nnwrap&#39;<\/a><\/p>\n\n<p>However, these answers didn't solve the error for me.<\/p>\n\n<p>Please help!<\/p>\n","answers":[{"AnswerId":"57620300","CreationDate":"2019-08-23T05:27:08.980","ParentId":null,"OwnerUserId":"11965903","Title":null,"Body":"<p>The error may be because you don't have the tools.py file in your working folder. You can take that file from the GitHub location mentioned in the tutorial you are referring to. Here is the link for your reference.\n<a href=\"https:\/\/github.com\/ThilinaRajapakse\/BERT_binary_text_classification\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ThilinaRajapakse\/BERT_binary_text_classification<\/a><\/p>\n"}]}
+{"QuestionId":57612703,"AnswerCount":1,"Tags":"<python><r><tensorflow><keras><deep-learning>","CreationDate":"2019-08-22T15:26:47.713","AcceptedAnswerId":null,"OwnerUserId":11204955.0,"Title":"Could not import PIL.Image even if Pillow already installed?","Body":"<p>I'm going through chapter 5 of the book Deep Learning with R (<a href=\"https:\/\/livebook.manning.com\/book\/deep-learning-with-r\/chapter-5\/112\" rel=\"nofollow noreferrer\">https:\/\/livebook.manning.com\/book\/deep-learning-with-r\/chapter-5\/112<\/a>). <\/p>\n\n<p>When running the code below, the following error appears: <code>Error in py_iter_next(it, completed) : ImportError: Could not import PIL.Image. The use of load_img requires PIL.<\/code><\/p>\n\n<p>All answers with regard to this error message recommend using <code>pip install pillow<\/code>. When I run this in my Terminal, the following appears: <\/p>\n\n<p><code>Requirement already satisfied: pillow in \/anaconda3\/lib\/python3.7\/site-packages (6.1.0)<\/code><\/p>\n\n<p>Hence, Pillow is already installed, whereas I cannot run the code below. \nI'm working with a Mac Book Pro platform x86_64-apple-darwin15.6.0 as well as R version 3.6.0 (2019-04-26). I also have Python 3.7.3 installed on my machine. Any referrals and support are highly appreciated!<\/p>\n\n<pre><code>#Displaying a batch of data and labels\nbatch &lt;- generator_next(train_generator)\nstr(batch)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57671082","CreationDate":"2019-08-27T08:59:01.717","ParentId":null,"OwnerUserId":"521776","Title":null,"Body":"<p>The problem is that Keras for R creates its own virtual environment, called <code>r-reticulate<\/code>, and pillow is missing in there. You also have to find out whether it used conda or virtualenv to create such environment. Then, activate it and install pillow and scipy. Finally, restart the R session.<\/p>\n\n<p>In my case it was conda:<\/p>\n\n<pre><code>$ conda env list\n# conda environments:\n#\nbase                  *  C:\\Users\\black\\Anaconda3\npy2                      C:\\Users\\black\\Anaconda3\\envs\\py2\nr-reticulate             C:\\Users\\black\\Anaconda3\\envs\\r-reticulate\n\n$ conda activate r-reticulate\n$ conda install pillow scipy\n<\/code><\/pre>\n"}]}
+{"QuestionId":57612775,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2019-08-22T15:31:14.823","AcceptedAnswerId":null,"OwnerUserId":9967569.0,"Title":"how to interpret the predictions of sequential model","Body":"<p>i want predict a bineared multiclass,the score of model is good, but i don't understand how predict a label.<\/p>\n\n<p>i was folow this tutorial : <a href=\"https:\/\/stackoverflow.blog\/2019\/05\/06\/predicting-stack-overflow-tags-with-googles-cloud-ai\/\">https:\/\/stackoverflow.blog\/2019\/05\/06\/predicting-stack-overflow-tags-with-googles-cloud-ai\/<\/a><\/p>\n\n<p>i have use model.predict(xtest), but the result is alway the same<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>X1 = data['Body']\n#data['Body'].head()\n#    absolutely positioned div containing several c...\n#    given datetime representing person birthday ca...\n#    given specific datetime value display relative...\n#    expose linq query asmx web service usually bus...\n#                              store binary data mysql\n\n\ntrain_size = int(len(data) * .8)\n\ntrain_qs = data['Body'].values[:train_size]\ntest_qs = data['Body'].values[train_size:]\n\nfrom tensorflow.keras.preprocessing import text\n\ntokenizer = text.Tokenizer(num_words=400)\ntokenizer.fit_on_texts(train_qs)\n\nbag_of_words_train = tokenizer.texts_to_matrix(train_qs)\nbag_of_words_test = tokenizer.texts_to_matrix(test_qs)\n\n\nVOCAB_SIZE=400\ntags_split =data['Tags']\n#data['Tags'].head()\n#                          [html, css]\n#                  [c#, .net, datetime]\n#                        [c#, datetime]\n#    [c#, linq, web-services, .net-3.5]\n#                     [mysql, database]\n\n# Create the encoder\nfrom sklearn.preprocessing import MultiLabelBinarizer\n\ntag_encoder = MultiLabelBinarizer()\ntags_encoded = tag_encoder.fit_transform(tags_split)\n\n# Split the tags into train\/test\ntrain_tags = tags_encoded[:train_size]\ntest_tags = tags_encoded[train_size:]\n\n\nfrom tensorflow.keras.preprocessing import text\nimport keras\n\n\nimport tensorflow as tf\nmodel = tf.keras.models.Sequential()\nmodel.add(tf.keras.layers.Dense(50, input_shape=(VOCAB_SIZE,), activation='relu'))\nmodel.add(tf.keras.layers.Dense(25, activation='relu'))\nmodel.add(tf.keras.layers.Dense(100, activation='sigmoid'))\n\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=[keras.metrics.binary_accuracy])\n\nmodel.fit(bag_of_words_train, train_tags, epochs=3, batch_size=128, validation_split=0.1)\nmodel.evaluate(bag_of_words_test, test_tags, batch_size=128)\n<\/code><\/pre>\n\n<p>so i don't understand the output : <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>yhat = model.predict_classes(x=bag_of_words_test, batch_size=128)\nprint(yhat)\n#[10 10 10 ... 10 10 10]\n#there are only 47 prediction diferent than 10 so 1848 of 10predict\n#but there are only 319 of this label in ytest\ntag_encoder.classes_[10]\n#'c#'\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57613041,"AnswerCount":1,"Tags":"<python-3.x><matplotlib><keras><jupyter-notebook><anaconda>","CreationDate":"2019-08-22T15:48:20.343","AcceptedAnswerId":null,"OwnerUserId":9512149.0,"Title":"Jupyter Notebook and MatPlotLib Not Plotting Keras Results","Body":"<p>I'm using Jupyter Notebook installed by Anaconda 1.9.7 to run a machine learning model using Tensorflow, Keras, Python 3.x, and Matplotlib. When I run the code from the Terminal on my Mac everything runs fine and the graph is plotted to an external window. When I run the same code in Jupyter Notebook, the kernel dies and restarts the first time the code uses Matplotlib.<\/p>\n\n<p>Initially, I was not using \"%matplotlib inline\" so I added this to the top, but the graph still does not show. I created a simple use case (not the machine learning code provided here) and the graph plotted inline to Jupyter Notebook. The current code works without a problem when I run it from the Terminal on my Mac, and the graph displays to external window.<\/p>\n\n<pre><code>get_ipython().run_line_magic('matplotlib', 'inline')\n\n\nimport tensorflow as tf\n\nfrom keras.datasets import reuters\n\nimport numpy as np\n\nnp_load_old = np.load\n\nnp.load = lambda *a,**k: np_load_old(*a, allow_pickle=True, **k)\n\n\n(train_data, train_labels), (test_data, test_labels) = reuters.load_data(num_words=10000)\n\nnp.load = np_load_old\n\nword_index = reuters.get_word_index()\n\nreverse_word_index = dict([(value, key) for (key, value) in word_index.items()])\n\ndecoded_newswire = ' '.join([reverse_word_index.get(i - 3, '?') for i in train_data[10]])\n\ndecoded_newswire\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\nx_train = vectorize_sequences(train_data)\n\nx_test = vectorize_sequences(test_data)\n\nfrom keras.utils.np_utils import to_categorical\n\none_hot_train_labels = to_categorical(train_labels)\n\none_hot_test_labels = to_categorical(test_labels)\n\nfrom keras import models\n\nfrom keras import layers\n\nmodel = models.Sequential()\n\nmodel.add(layers.Dense(64, activation='relu', input_shape=(10000,)))\n\nmodel.add(layers.Dense(64, activation='relu'))\n\nmodel.add(layers.Dense(46, activation='softmax'))\n\nmodel.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])\n\nx_val = x_train[:1000]\n\npartial_x_train = x_train[1000:]\n\ny_val = one_hot_train_labels[:1000]\n\npartial_y_train = one_hot_train_labels[1000:]\n\nhistory = model.fit(partial_x_train, partial_y_train, epochs=3, batch_size=512, validation_data=(x_val, y_val))\n\nloss = history.history['loss']\n\nval_loss = history.history['val_loss']\n\nepochs = range(1, len(loss) + 1)\n\nimport matplotlib.pyplot as plt\n\nplt.title('Training and validation loss')\n\nplt.xlabel('Epochs')\n\nplt.ylabel('Loss')\n\nplt.legend()\n\nplt.plot(epochs, loss, 'bo', label='Training loss')\n\nplt.plot(epochs, val_loss, 'b', label='Validation loss')\n\nplt.show()\n<\/code><\/pre>\n\n<p>I expect the last line to plot a graph inline in Jupyter Notebook. Instead, the Kernel dies at the line \"plt.title('Training and validation loss')\" and when I run the line independently it give the error \"NameError: name 'plt' is not defined.\"<\/p>\n","answers":[{"AnswerId":"57617096","CreationDate":"2019-08-22T21:10:25.570","ParentId":null,"OwnerUserId":"9637085","Title":null,"Body":"<p>I tried your code with a little modifications:<\/p>\n\n<ul>\n<li>commented out the following lines:\n\n<ul>\n<li><code>np_load_old = np.load<\/code><\/li>\n<li><code>np.load = lambda *a, **l: np_load_old ...<\/code><\/li>\n<li><code>np.load = np_load_old<\/code><\/li>\n<\/ul><\/li>\n<\/ul>\n\n<p>and it worked. I got the following plot:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/b8hjX.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/b8hjX.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>My versions are:<\/p>\n\n<ul>\n<li>tensorflow 1.13.1<\/li>\n<li>matplotlib 3.1.1<\/li>\n<li>jupyter 1.0.0<\/li>\n<li>python 3.6.6<\/li>\n<\/ul>\n"}]}
+{"QuestionId":57613079,"AnswerCount":1,"Tags":"<keras><conv-neural-network>","CreationDate":"2019-08-22T15:50:37.920","AcceptedAnswerId":null,"OwnerUserId":1812810.0,"Title":"Compile error on keras sequential model with custom loss function","Body":"<p>Trying to compile CNN model of ~16K parameters on GPU in google colab for mnist dataset.\nWith standard loss 'categorical_crossentropy', it is working fine. But with custom_loss it is giving error. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>lamda=0.01\nm = X_train.shape[0]\n\ndef reg_loss(lamda):\n  model_layers = custom_model.layers # type list where each el is Conv2D obj etc.\n  reg_wts = 0\n\n  for idx, layer in enumerate(model_layers):\n    layer_wts = model_layers[idx].get_weights() # type list\n\n    if len(layer_wts) &gt; 0: # activation, dropout layers do not have any weights\n      layer_wts = model_layers[idx].get_weights()[0] #ndarray, 3,3,1,16 : layer1 output\n\n      s = np.sum(layer_wts**2)\n      reg_wts += s\n\n  print(idx, \"reg_wts\", reg_wts)    \n  return (lamda\/(2*m))* reg_wts\n\nreg_loss(lamda)\n\ndef custom_loss(y_true, y_pred):\n  K.categorical_crossentropy(y_true, y_pred) + reg_loss(lamda)\n\ncustom_model.compile(loss=custom_loss, optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>reg_loss returns 28 reg_wts 224.11805880069733\n1.8676504900058112e-05<\/p>\n\n<p>On compile, gives error <strong>AttributeError: 'NoneType' object has no attribute 'get_shape'<\/strong><a href=\"https:\/\/i.stack.imgur.com\/fyus5.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/fyus5.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"57614582","CreationDate":"2019-08-22T17:43:12.680","ParentId":null,"OwnerUserId":"1812810","Title":null,"Body":"<p>custom_loss function did not have return statement. A silly mistake, but the error was quite misleading. Hence it took so much time.<\/p>\n"}]}
+{"QuestionId":57613197,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-22T15:59:17.870","AcceptedAnswerId":null,"OwnerUserId":9954063.0,"Title":"Is it important to build a graph using tf.Graph() before running a session in tensorflow?","Body":"<p>I have been trying a run a simple 3 layer neural network using tensorflow. i am facing an error that states \"session graph is empty. Add operations to session graph before calling run()\". From consulting others and some online study from tensorflow documentation, I understand that inorder to run a session one first needs to create a graph using tf.Graph() and then create nodes and tensors inside the scope of that graph and then run a session. How correct is my understanding _<\/p>\n\n<p>Below is my exact code. Kindly have a look at it,<\/p>\n\n<pre><code>def run(self, carollis_input):\n        self.normalization(carollis_input)\n        #'finding the output of the input layer'\n        knowledge_input = tf.add(tf.multiply(self.neuron_input, self.weight_in), self.bias_in)\n\n            #'calculating the input for the hidden layer'\n        knowledge_hidden = tf.add(tf.multiply(knowledge_input, self.weight_in), self.bias_hid)\n        #'calculating the output of hidden layer'\n        knowledge_hidden_output = 3.14*(tf.add(tf.multiply(knowledge_hidden, self.weight_hid), self.bias_hid))#input function of hidden layer\n        knowledge_hidden_out = tf.nn.leaky_relu(self.neuron_hid, alpha=0.01, name='leaky_relu')\n        with tf.Session() as sess1_2:\n            sess1_2.run(knowledge_hidden_out, feed_dict={self.neuron_input: carollis_input, self.neuron_hid: knowledge_hidden_output})\n        #'calculating the input of output layer'\n        tf.reshape(knowledge_hidden_out, [4, 2])#for quadrant method\n        in_out = tf.add(tf.multiply(knowledge_hidden_out, self.weight_out), self.bias_out)\n        with tf.Session as s:\n            s.run(in_out)\n        #'finding the softmax output of the neurons'\n        softmax_output = np.array(4)\n        softmax_output = self.out_softmax(in_out)  # this gives the softmax output and stores it in the newly created array\n        return softmax_output\n<\/code><\/pre>\n\n<p>I have been having this following error,<\/p>\n\n<pre><code>File \"\/home\/microbot\/catkin_ws\/src\/spider\/spider_control\/knowledge_transfer.py\", line 93, in run\n    sess1_2.run(knowledge_hidden_out, feed_dict={self.neuron_input: carollis_input, self.neuron_hid: knowledge_hidden_output})\n  File \"\/home\/microbot\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 950, in run\n    run_metadata_ptr)\n  File \"\/home\/microbot\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1098, 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>can anyone please give me a basic explanation of this tf.Graph and tf.Session working together. Thank you for your time.<\/p>\n","answers":[]}
+{"QuestionId":57613236,"AnswerCount":0,"Tags":"<python><tensorflow><object-detection><tracking><centroid>","CreationDate":"2019-08-22T16:02:35.187","AcceptedAnswerId":null,"OwnerUserId":11444037.0,"Title":"how to integrate tensorflow object detection api with centroid tracking","Body":"<p>I am using TensorFlow object detection API to detect humans in video frames. I want to give the detection coordinates to the centroid tracker, to assign an ID for every detected object and avoid detecting them every frame as a new object.<\/p>\n\n<p>This is the code I am using for detection and getting the bounding boxes<\/p>\n\n<pre><code># Import packages\nimport os\nimport cv2\nimport numpy as np\nimport tensorflow as tf\nimport sys\n\n# This is needed since the notebook is stored in the object_detection folder.\nfrom pandas._libs import json\n\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 = 'inference_graph'\nVIDEO_NAME = 'test.mp4'\n\n# direction for out put file\nFILE_OUTPUT = 'C:\/Mohammad\/tensorflow1\/models\/research\/object_detection\/savedframes\/out.avi'\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,'training','label_map.pbtxt')\n\n# Path to video\nPATH_TO_VIDEO = os.path.join(CWD_PATH,VIDEO_NAME)\n\n# Number of classes the object detector can identify\nNUM_CLASSES = 2\n\n# Load the label map.\n# Label maps map indices to category names, so that when our convolution\n# network predicts `5`, we know that this corresponds to `king`.\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# 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# Open video file\nvideo = cv2.VideoCapture('C:\/Mohammad\/tensorflow1\/models\/research\/object_detection\/test.mp4')\nframe_width = int(video.get(3))\nframe_height = int(video.get(4))\nout = cv2.VideoWriter(FILE_OUTPUT, cv2.VideoWriter_fourcc('M', 'J', 'P', 'G'),\n                      10, (frame_width, frame_height))\n\n\nframe_index = 0\nwhile(video.isOpened()):\n\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 = video.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        frame_index,\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=0.80)\n    #frame_index +=1\n    if ret == True:\n        out.write(frame)\n\n# writing coordinates\n    coordinates = vis_util.return_coordinates(\n        frame,\n        frame_index,\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=0.80)\n\n    for coordinate in coordinates:\n        (ymin, ymax, xmin, xmax, acc, classification) = coordinate\n        height = ymax - ymin\n        width = xmax -xmin\n        crop = frame[ymin:ymin+height, xmin:xmin+width]\n        path = 'C:\/Mohammad\/A_Mohammad laptop\/Tensorflow From C Drive\/tensorflow1\/models\/research\/object_detection\/savecroped'\n        cv2.imwrite(os.path.join(path,'crop%d.jpg' %frame_index), frame)\n\n        textfile = open('filename_string' + \".json\", \"a\")\n        textfile.write(json.dumps(coordinates))\n        textfile.write(\"\\n\")\n    frame_index =frame_index+1\n\n# txt file\n    #textfile = open('filename_string' + \".txt\", \"w\")\n\n    #textfile.write(str(coordinates))\n    #textfile.write(\"\\n\")\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\n    # Press 'q' to quit\n    if cv2.waitKey(1) == ord('q'):\n        break\n\n# Clean up\nvideo.release()\ncv2.destroyAllWindows()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57613825,"AnswerCount":2,"Tags":"<python><keras><time-series><lstm><multiclass-classification>","CreationDate":"2019-08-22T16:43:26.287","AcceptedAnswerId":57663301.0,"OwnerUserId":6216841.0,"Title":"Saving LSTM hidden states while training and predicting for multi-class time series classification","Body":"<p>I am trying to use an LSTM for multi-class classification of time series data.<\/p>\n\n<p>The training set has dimensions (390, 179), i.e. 390 objects with 179 time steps each.<\/p>\n\n<p>There are 37 possible classes.<\/p>\n\n<p>I would like to use a Keras model with just an LSTM and activation layer to classify input data.<\/p>\n\n<p>I also need the hidden states for all the training data and test data passed through the model, at every step of the LSTM (not just the final state).<\/p>\n\n<p>I know <code>return_sequences=True<\/code> is needed, but I'm having trouble getting dimensions to match.<\/p>\n\n<p>Below is some code I've tried, but I've tried a ton of other combinations of calls from a motley of stackexchange and git issues. In all of them I get some dimension mismatch or another.<\/p>\n\n<p>I don't know how to extract the hidden state representations from the model.<\/p>\n\n<p>We have <code>X_train.shape = (390, 1, 179)<\/code>, <code>Y_train.shape = (390, 37)<\/code> (one-shot binary vectors)\/.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>n_units = 8\nn_sequence = 179\nn_class = 37\n\nx = Input(shape=(1, n_sequence))\ny = LSTM(n_units, return_sequences=True)(x)\nz = Dense(n_class, activation='softmax')(y)\n\nmodel = Model(inputs=[x], outputs=[y])\nmodel.compile(loss='categorical_crossentropy', optimizer='adam')\n\nmodel.fit(X_train, Y_train, epochs=100, batch_size=128)\nY_test_predict = model.predict(X_test, batch_size=128)\n<\/code><\/pre>\n\n<p>This is what the above gives me:<\/p>\n\n<p><code>ValueError: A target array with shape (390, 37) was passed for an output of shape (None, 1, 37) while using as loss 'categorical_crossentropy'. This loss expects targets to have the same shape as the output.<\/code><\/p>\n","answers":[{"AnswerId":"57622553","CreationDate":"2019-08-23T08:35:57.007","ParentId":null,"OwnerUserId":"11597209","Title":null,"Body":"<p>You input shape should like this: (samples, timesteps, features)\nWhere samples are how many sequences you have, timesteps how long are your sequences, and features how many input you wanna input in one timestep.\nIf you set return_sequences=True, your label array should have the shape of (samples, timesteps, output features).<\/p>\n"},{"AnswerId":"57663301","CreationDate":"2019-08-26T18:46:06.860","ParentId":null,"OwnerUserId":"6216841","Title":null,"Body":"<p>There didn't seem to be any way to build a working trainable model while also returning the hidden states with <code>return_sequences=True<\/code>.<\/p>\n\n<p>The fix I found was to build a predictor model and train it, and save the weights. Then I built a new model which ended with my LSTM layer, and fed it the trained weights. So, using <code>return_sequences=True<\/code>, I was able to predict on new data and get the data's representations at each hidden state.<\/p>\n"}]}
+{"QuestionId":57613902,"AnswerCount":1,"Tags":"<python><tensorflow><conv-neural-network><mnist>","CreationDate":"2019-08-22T16:50:01.347","AcceptedAnswerId":57614217.0,"OwnerUserId":11963653.0,"Title":"lack of understanding cnn result and utility","Body":"<p>I'm working on deep learning and I think I'm lost in the cnn. Let me explain: by following <a href=\"https:\/\/github.com\/ageron\/handson-ml\/blob\/master\/13_convolutional_neural_networks.ipynb\" rel=\"nofollow noreferrer\">Aur\u00e9lien G\u00e9ron's github<\/a> (In cell [21]) I trained \"my basic\" MNIST . And everything went well ... I think .. ^^ ... I used an initialization of he too.. :<\/p>\n\n<p>.... <br\/>\n8 Last batch accuracy: 1.0 Test accuracy: 0.9852, <br\/>\n9 Last batch accuracy: 1.0 Test accuracy: 0.9892, <\/p>\n\n<p>However, in the book of the same author, it specifies that: \"... once the CNN has learned to recognize a pattern in one location, it can recognize it in any other location ...\" (page 358). Does this mean that if my network detects 99% of a digit, I would be able to move it anywhere it recognizes it? I ask this, because it does not work. The predictions are very bad, i.e. 100% detection of a \"4\" on a 28x28 image:<\/p>\n\n<p>... <br\/>\n3 : 0.0 % <br\/>\n4 : 100.0 % <br\/> \n5 : 0.0 %<br\/>\n...<br\/><\/p>\n\n<p>On this same image I move the shape of a few pixels (with paint, for those who want to know how):<\/p>\n\n<p>0 : 20.0 % <br\/>\n1 : 21.0 % <br\/>\n2 : 2.0 % <br\/>\n3 : 0.0 % <br\/>\n4 : 22.0 % <br\/>\n5 : 14.0 % <br\/>\n6 : 0.0 % <br\/>\n7 : 21.0 % <br\/>\n8 : 1.0 % <br\/>\n9 : 0.0 % <br\/><\/p>\n\n<p>Do I really understand the idea of a cnn, or not?<br>\nMoreover, why it doesn't work if I put a \"4\" and a \"2\" side by side?\nThe networks aren't supposed to tell me 100% each?<\/p>\n\n<h3>My <code>import_graph<\/code> code:<\/h3>\n\n<pre><code>saver = tf.train.import_meta_graph(\".\/my_mnist_model.meta\")  \nX_prod = snt_X_prod.astype(np.float32).reshape(-1, 28*28) \/ 255.0 \nsess = tf.Session()\nsaver.restore(sess, \".\/my_mnist_model\")\nresult = sess.run(\"output\/Y_proba:0\", feed_dict={\"inputs\/X:0\": X_prod})\n<\/code><\/pre>\n","answers":[{"AnswerId":"57614217","CreationDate":"2019-08-22T17:16:23.857","ParentId":null,"OwnerUserId":"3399825","Title":null,"Body":"<p>It's not as simple as that. The convolutional <strong>layer<\/strong> indeed is in a sense invariant to translation of the image, however the neural network that you are using is not just a convolutional layer, from what I can see in the notebook, it contains two convolutional layers, max pooling, fully connected and output layer. The fully connected and output layers are naturally <strong>not<\/strong> invariant to translation, therefore the whole network is not invariant to translation. When you shift the input image and feed it to the network, the output of the second convolutional layer will be the same as with the original image, just translated accordingly, however, the fully connected layers were trained to \"expect\" some distribution of activations of the convolutional layer, by translating the image manually you are probably providing an example that is too different from the training examples which results in a failed classification.<\/p>\n\n<p>For the second point, 4 and 2 side by side doesn't work because the network was not trained on pairs of digits. As in the previous case, when you present the network with a 4 and 2 in a single image, the output of the convolutional layer will be very different from what occured during training, therefore the network won't give you a meaningful answer.<\/p>\n\n<p>Furthermore, the network has a softmax layer on the top, so in fact it <strong>cannot<\/strong> give you 100% on two classes at once, the sum of all probabilities will always be 100%.<\/p>\n\n<p>I recommend trying to get a deeper understanding of what is actually going on in the neural network and how it is optimized. Once you grasp the few clever bits of math that are involved, you will see the answers to your questions. The models are not magic boxes that always give you the right answer. For this I recommend the CNN course from Stanford, by Andrej Karpathy. If you follow that, you will get real, non-superficial understanding of what is going on.<\/p>\n\n<p>If you want a model that works well on translated digits, you can use larger images, e.g. 40x40, then put the original 28x28 image into a random position and train your network on that. Then you should have a model that can recognize a digit no matter where you put it.<\/p>\n\n<p>If you want a model, that can recognize multiple digits, create a dataset that contains examples of images with multiple digits and use multiple labels for each sample (for this you will have to stop using softmax and use sigmoid, probably with MSE error instead). Then you should have a model that can recognize multiple digits at once.<\/p>\n"}]}
+{"QuestionId":57614072,"AnswerCount":0,"Tags":"<python><keras><visualization><autoencoder><keras-2>","CreationDate":"2019-08-22T17:04:00.877","AcceptedAnswerId":null,"OwnerUserId":1684118.0,"Title":"Visualizing layers of autoencoder","Body":"<p>I have created a variational autoencoder in Keras using 2D convolutions for encoder and decoder. The code is shown below. Now, I would like to visualize the individual layers or filters (feature maps) to see what the network learns.<\/p>\n\n<p>How can this be done?<\/p>\n\n<pre><code>import keras\nfrom keras import backend as K\nfrom keras.layers import (Dense, Input, Flatten)\nfrom keras.layers import Lambda, Conv2D\nfrom keras.models import Model\nfrom keras.layers import Reshape, Conv2DTranspose\nfrom keras.losses import mse\n\ndef sampling(args):\n    z_mean, z_log_var = args\n    batch = K.shape(z_mean)[0]\n    dim = K.int_shape(z_mean)[1]\n    epsilon = K.random_normal(shape=(batch, dim))\n    return z_mean + K.exp(0.5 * z_log_var) * epsilon\n\ninner_dim = 16\nlatent_dim = 6\n\nimage_size = (64,78,1)\ninputs = Input(shape=image_size, name='encoder_input')\nx = inputs\n\nx = Conv2D(32, 3, strides=2, activation='relu', padding='same')(x)\nx = Conv2D(64, 3, strides=2, activation='relu', padding='same')(x)\n\n# shape info needed to build decoder model\nshape = K.int_shape(x)\n\n# generate latent vector Q(z|X)\nx = Flatten()(x)\nx = Dense(inner_dim, activation='relu')(x)\nz_mean = Dense(latent_dim, name='z_mean')(x)\nz_log_var = Dense(latent_dim, name='z_log_var')(x)\n\nz = Lambda(sampling, output_shape=(latent_dim,), name='z')([z_mean, z_log_var])\n\n# instantiate encoder model\nencoder = Model(inputs, [z_mean, z_log_var, z], name='encoder')\n\n# build decoder model\nlatent_inputs = Input(shape=(latent_dim,), name='z_sampling')\nx = Dense(inner_dim, activation='relu')(latent_inputs)\nx = Dense(shape[1] * shape[2] * shape[3], activation='relu')(x)\nx = Reshape((shape[1], shape[2], shape[3]))(x)\n\nx = Conv2DTranspose(64, 3, strides=2, activation='relu', padding='same')(x)\nx = Conv2DTranspose(32, 3, strides=2, activation='relu', padding='same')(x)\n\noutputs = Conv2DTranspose(filters=1, kernel_size=3, activation='sigmoid', padding='same', name='decoder_output')(x)\n\n# instantiate decoder model\ndecoder = Model(latent_inputs, outputs, name='decoder')\n\n# instantiate VAE model\noutputs = decoder(encoder(inputs)[2])\nvae = Model(inputs, outputs, name='vae')\n\ndef vae_loss(x, x_decoded_mean):\n    reconstruction_loss = mse(K.flatten(x), K.flatten(x_decoded_mean))\n    reconstruction_loss *= image_size[0] * image_size[1]\n    kl_loss = 1 + z_log_var - K.square(z_mean) - K.exp(z_log_var)\n    kl_loss = K.sum(kl_loss, axis=-1)\n    kl_loss *= -0.5\n    vae_loss = K.mean(reconstruction_loss + kl_loss)\n    return vae_loss\n\noptimizer = keras.optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.000)\nvae.compile(loss=vae_loss, optimizer=optimizer)\nvae.fit(train_X, train_X,\n        epochs=500,\n        batch_size=128,\n        verbose=1,\n        shuffle=True,\n        validation_data=(valid_X, valid_X))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57614436,"AnswerCount":2,"Tags":"<tensorflow><import><object-detection-api>","CreationDate":"2019-08-22T17:32:19.407","AcceptedAnswerId":null,"OwnerUserId":11838316.0,"Title":"od_graph_def = tf.GraphDef() AttributeError: module 'tensorflow' has no attribute 'GraphDef'","Body":"<p>I have a mac and I am using tensorflow 2.0, python 3.7.\nI am following the tutorial for creating an object detection model for real-time application.\nbut i am getting the following error:<\/p>\n\n<p>\"Downloads\/models\/research\/object_detection\/object_detection_tutorial.py\", line 43, in \n    od_graph_def = tf\n  od_graph_def = tf.GraphDef()<\/p>\n\n<p>AttributeError: module 'tensorflow' has no attribute 'GraphDef'<\/p>\n\n<p>below is the tutorial link:<\/p>\n\n<p>I checked the environment and I already have tensorflow environment in anaconda <\/p>\n\n<pre><code>import tensorflow as tf\nimport zipfile\n\nfrom collections import defaultdict\nfrom io import StringIO\nfrom matplotlib import pyplot as plt\nfrom PIL import Image\n\nsys.path.append(\"..\")\nfrom object_detection.utils import ops as utils_ops\n\n\nfrom utils import label_map_util\n\nfrom utils import visualization_utils as vis_util\n\nMODEL_NAME = 'ssd_mobilenet_v1_coco_2017_11_17'\nMODEL_FILE = MODEL_NAME + '.tar.gz'\nDOWNLOAD_BASE = 'http:\/\/download.tensorflow.org\/models\/object_detection\/'\n\nPATH_TO_CKPT = MODEL_NAME + '\/frozen_inference_graph.pb'\n\nPATH_TO_LABELS = os.path.join('data', 'mscoco_label_map.pbtxt')\n\nNUM_CLASSES = 90\n\n\nopener = urllib.request.URLopener()\nopener.retrieve(DOWNLOAD_BASE + MODEL_FILE, MODEL_FILE)\ntar_file = tarfile.open(MODEL_FILE)\nfor file in tar_file.getmembers():\n  file_name = os.path.basename(file.name)\n  if 'frozen_inference_graph.pb' in file_name:\n    tar_file.extract(file, os.getcwd())\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<\/code><\/pre>\n","answers":[{"AnswerId":"58222195","CreationDate":"2019-10-03T15:18:31.940","ParentId":null,"OwnerUserId":"7592832","Title":null,"Body":"<p>Yeah, the syntax has changed in T2.0. Here's the correct piece:<\/p>\n\n<pre><code>tf.compat.v1.GraphDef()   # -&gt; instead of tf.GraphDef()\ntf.compat.v2.io.gfile.GFile()   # -&gt; instead of tf.gfile.GFile()\n<\/code><\/pre>\n"},{"AnswerId":"58473143","CreationDate":"2019-10-20T12:48:42.990","ParentId":null,"OwnerUserId":"1340604","Title":null,"Body":"<p>I had similar issues, when upgraded to Python 3.7 and Tensorflow 1.2.0 to Tensorflow 2.0.0<\/p>\n\n<p>If you don't want to touch your code, just add these 2 lines in the main.py file w\/ Tensorflow code:<\/p>\n\n<pre><code>import tensorflow.compat.v1 as tf\ntf.disable_v2_behavior()\n<\/code><\/pre>\n\n<p>And that's it!!<br>\n<strong>NOW<\/strong> Everything should run seamlessly :)<\/p>\n\n<p>But if you write new code, indeed (as was mentioned above) change, these calls:<\/p>\n\n<pre><code>  with tf.gfile.GFile(path, 'r') as fid:\n<\/code><\/pre>\n\n<p>To:<\/p>\n\n<pre><code>  with tf.io.gfile.GFile(path, 'r') as fid:\n<\/code><\/pre>\n"}]}
+{"QuestionId":57614513,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-22T17:37:21.233","AcceptedAnswerId":57614667.0,"OwnerUserId":9715052.0,"Title":"TypeError: 'Tensor' object is not callable | Keras Autoencoder","Body":"<p>I'm creating an autoencoder in Keras following <a href=\"https:\/\/blog.keras.io\/building-autoencoders-in-keras.html\" rel=\"nofollow noreferrer\">this tutorial<\/a> and I keep getting the following error. <\/p>\n\n<pre><code>decoder = tf.keras.Model(encoded_input, decoded(input_img))\nTypeError: 'Tensor' object is not callable\n<\/code><\/pre>\n\n<p>I believe it's something to do with not being able to use tensors in this way because of the nature of this type of object but I have some gaps in understanding of why and how to go about solving this. <\/p>\n\n<p>Here is a minimal working example of my code: <\/p>\n\n<pre><code># input_img input placeholder\ninput_img = tf.keras.layers.Input(shape=(16, 16, 1), name =\"input\")\nencoded = tf.keras.layers.Dense(encoding_dim, activation='relu')(input_img)\ndecoded = tf.keras.layers.Dense(256, activation='sigmoid')(encoded)\nautoencoder = tf.keras.Model(input_img, decoded)\nencoder = tf.keras.Model(input_img, encoded)\nencoded_input = tf.keras.layers.Input(shape=(encoding_dim,))\ndecoder = tf.keras.Model(input_img, decoded(encoded_input))\n<\/code><\/pre>\n","answers":[{"AnswerId":"57614667","CreationDate":"2019-08-22T17:48:18.047","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>Keras <code>Model<\/code> expects input &amp; output arguments as <em>layers<\/em>, not tensors. The tutorial is correct - you appear to have missed a line right before <code>decoder =<\/code>: -- <code>decoder_layer = autoencoder.layers[-1]<\/code><\/p>\n\n<p>All together, with <code>decoder =<\/code> changed appropriately:<\/p>\n\n<pre><code>input_img = tf.keras.layers.Input(shape=(16, 16, 256), name =\"input\")\nencoded   = tf.keras.layers.Dense(encoding_dim, activation='relu')(input_img)\ndecoded   = tf.keras.layers.Dense(256, activation='sigmoid')(encoded)\n\nautoencoder = tf.keras.Model(input_img, decoded)\nencoder     = tf.keras.Model(input_img, encoded)\n\nencoded_input = tf.keras.layers.Input(shape=(encoding_dim,))\ndecoder_layer = autoencoder.layers[-1] # gets 'decoded' = last layer of 'autoencoder'\n\ndecoder = tf.keras.Model(encoded_input, decoder_layer(encoded_input))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57614588,"AnswerCount":1,"Tags":"<tensorflow><transformer>","CreationDate":"2019-08-22T17:43:41.180","AcceptedAnswerId":null,"OwnerUserId":11963938.0,"Title":"InvalidArgumentError: cannot compute Mul as input #1(zero-based) was expected to be a int32 tensor but is a float tensor [Op:Mul] name: mul\/","Body":"<p>I instantiate a class and get some errors when passing in parameters.saying that the data type received not the expected type<\/p>\n\n<p>i try to convert the data to the type it requires but get the same error<\/p>\n\n<pre><code>class PositionalEmbedding(tf.keras.Model):\ndef __init__(self, d_model, dropout, max_len=5000):\n  super(PositionalEmbedding,self).__init__()\n  self.dropout = tf.keras.layers.Dropout(dropout,dtype = 'float32')\n\n  pe = tf.zeros([max_len,d_model],dtype = 'float32')\n  position = tf.expand_dims(tf.range(max_len),axis = 1)\n<\/code><\/pre>\n\n<p>pe = PositionalEmbedding(20,0)<\/p>\n\n<p>InvalidArgumentError                      Traceback (most recent call last)\n in ()\n      1 plt.figure(figsize=(15,5))\n      2 \n----> 3 pe = PositionalEmbedding(20,0)\n      4 \n      5 s = pe(tf.Variable(tf.zeros(1,100,20)))<\/p>\n\n<p>4 frames\n\/usr\/local\/lib\/python3.6\/dist-packages\/six.py in raise_from(value, from_value)<\/p>\n\n<p>InvalidArgumentError: cannot compute Mul as input #1(zero-based) was expected to be a int32 tensor but is a float tensor [Op:Mul] name: mul\/<\/p>\n","answers":[{"AnswerId":"57830237","CreationDate":"2019-09-07T02:31:09.143","ParentId":null,"OwnerUserId":"5370186","Title":null,"Body":"<p>What tensorflow version are you using? Works as written for me in tf 2.0 and tf 1.14 (sorry don't have enough rep to put this as a comment) <\/p>\n\n<p>n.b. I'm assuming that to correct the indentation this is what you actually meant, because any other correction makes some of the lines redundant...but please correct if that's wrong<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    class PositionalEmbedding(tf.keras.Model):\n        def __init__(self, d_model, dropout, max_len=5000):\n             super(PositionalEmbedding,self).__init__()\n             self.dropout = tf.keras.layers.Dropout(dropout,dtype = 'float32')\n\n             pe = tf.zeros([max_len,d_model],dtype = 'float32')\n             position = tf.expand_dims(tf.range(max_len),axis = 1)\n\n    pe = PositionalEmbedding(20, 0)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57614600,"AnswerCount":0,"Tags":"<keras><gpu><lstm>","CreationDate":"2019-08-22T17:44:28.277","AcceptedAnswerId":null,"OwnerUserId":6450489.0,"Title":"Running keras LSTM models with multiple GPU (Tensorflow backend)","Body":"<p>I have a keras LSTM model and want to run it under multiple GPUs for speed improvement. But I have some ambiguities:<\/p>\n\n<p>1- I found that to really get the great speed on GPU I should define my network using CuDNNLSTM layer and not normal LSTM layer. To use multiple GPUs, I looked at <a href=\"https:\/\/keras.io\/utils\/#multi_gpu_model\" rel=\"nofollow noreferrer\">Keras documentation<\/a> and wanted to use multi_gpu_model() function to make distributed model. However, in the sample scripts they recommend to define the model on CPU for easy weight sharing, but my CuDNNLSTM model is not deployable on CPU and LSTM model will not benefit from the enhancements provided by GPU. What is the correct approach?<\/p>\n\n<p>2- So I tried many configurations, including:<\/p>\n\n<ul>\n<li><p>Group 1(using the normal (non-fast) LSTM layers): placing model on CPU and no copying to GPU; placing model on CPU and then use multi_gpu_model to create GPU copies; place model on default GPU and no copying to other GPU; placing model on default GPU and then use multi_gpu_model to create two GPU copies.<\/p><\/li>\n<li><p>Group2 (using CuDNNLSTM layer and therefore no possibility to place model on CPU): defining a single model (which Tensorflow places it on the default GPU); using multi_gpu_model to create two GPU copies.<\/p><\/li>\n<\/ul>\n\n<p>In all cases, data parallelism (using multi_gpu_model) resulted in lower speed of execution. I didn't change anything else in my code and input data pipeline or batch sizes. What is wrong with me?<\/p>\n\n<p>3- In general, should I only use CuDNN-type layers to get high speed computation with GPUs when I am programming at high level of keras API?<\/p>\n","answers":[]}
+{"QuestionId":57614635,"AnswerCount":0,"Tags":"<matlab><caffe><matcaffe>","CreationDate":"2019-08-22T17:46:20.890","AcceptedAnswerId":null,"OwnerUserId":10508031.0,"Title":"ImportCaffeNetwork with deploy.prototxt \"network must have an output layer\"","Body":"<p>I tried to use MemNet on a set of images to get scores of their expected memorability. I am using Matlab with the importCaffeNetwork function.<\/p>\n\n<p>I get an error using <code>deploy.prototxt<\/code>.\nIt works if I use the <code>train_val.prototxt<\/code> but if I understand correctly that is for the training of the network. Or can I use that file?<\/p>\n\n<pre><code>&gt; net = importCaffeNetwork('deploy.prototxt', 'memnet.caffemodel',...\n&gt; 'InputSize', [256 256 3]);\n<\/code><\/pre>\n\n<blockquote>\n  <p><strong>The network must have an output layer.<\/strong><\/p>\n  \n  <p>Error in importCaffeNetwork (line 84)\n      network = importSeriesNetwork(readerObj);<\/p>\n<\/blockquote>\n\n<p>This is what the last layer looks like:<\/p>\n\n<pre><code>name: \"MemNet\"\ninput: \"data\"\ninput_dim: 10\ninput_dim: 3\ninput_dim: 227\ninput_dim: 227\n   ...\n}\nlayers {\n  name: \"fc8-euclidean\"\n  type: INNER_PRODUCT\n  bottom: \"fc7\"\n  top: \"fc8-euclidean\"\n  inner_product_param {\n    num_output: 1\n  }\n}\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57614726,"AnswerCount":1,"Tags":"<python><tensorflow><slice><indices>","CreationDate":"2019-08-22T17:53:20.030","AcceptedAnswerId":57614959.0,"OwnerUserId":7934786.0,"Title":"how to get indices of a tensor values based on a column condition in tensorflow","Body":"<p>I have a tensor like this:<\/p>\n\n<pre><code>sim_topics = [[0.65 0.   0.   0.   0.42  0.   0.   0.51 0.   0.34 0.]\n              [0.   0.51 0.   0.   0.52  0.   0.   0.   0.53 0.42 0.]\n              [0.   0.32 0.   0.50 0.34  0.   0.   0.39 0.32 0.52 0.]\n              [0.   0.23 0.37 0.   0.    0.37 0.37 0.   0.47 0.39 0.3 ]]\n<\/code><\/pre>\n\n<p>I want to get indices in this tensor based on a tensor condition:<\/p>\n\n<pre><code>masked_t = [True  False  True  False True True False True False True False]\n<\/code><\/pre>\n\n<p>So the output should be like this:<\/p>\n\n<pre><code>[[0.65 0. 0.   0.   0.42  0.   0.   0.51 0.   0.34 0.]\n [0.   0. 0.   0.   0.52  0.   0.   0.   0.   0.42 0.]\n [0.   0. 0.   0.   0.34  0.   0.   0.39 0.   0.52 0.]\n [0.   0. 0.37 0.   0.    0.37 0.   0.   0.   0.39 0.]]\n<\/code><\/pre>\n\n<p>So the condition is working on the columns of the initial tensor. Actually I need the indices of the elements which they are True in the <code>maske_t<\/code>.<\/p>\n\n<p>So the indices should be:<\/p>\n\n<pre><code>[[0, 0],\n [1,0],\n [2, 0],\n [3,0],\n [0,2],\n [1,2],\n [2,2],\n [3,2],\n ....]]\n<\/code><\/pre>\n\n<p>Actually this approach works when Im doing row wise, but here I want to select specific columns based on a condition so it raise s incompatibility error:<\/p>\n\n<pre><code>out = tf.cast(tf.zeros(shape=tf.shape(sim_topics), dtype=tf.float64), tf.float64)\nindices = tf.where(tf.where(masked_t, out, sim_topics))\n<\/code><\/pre>\n","answers":[{"AnswerId":"57614959","CreationDate":"2019-08-22T18:10:27.077","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>You can directly obtain your required tensor like this:<\/p>\n\n<pre><code>result = tf.multiply(sim_topics, tf.cast(masked_t, dtype=tf.float64))\n<\/code><\/pre>\n\n<p>Let the broadcasting do the work for masked_t to be same size as sim_topics<\/p>\n"}]}
+{"QuestionId":57614765,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><reinforcement-learning>","CreationDate":"2019-08-22T17:55:55.037","AcceptedAnswerId":null,"OwnerUserId":10018602.0,"Title":"Q Learning Model Not Improving Over Time","Body":"<p>I'm working with the code shown below from chapter 11 of the book Predictive Analytics with Tensorflow (code given on <a href=\"https:\/\/github.com\/PacktPublishing\/Predictive-Analytics-with-TensorFlow\/tree\/master\/Chapter11\" rel=\"nofollow noreferrer\">GitHub<\/a>). It tries to set up a basic Q learning function to trade historical stock data. I've modified it a little bit to get new historical data for whatever stock is entered.<\/p>\n\n<p>I'm finding that the q learning is not improving much and seems to be \"exploring\" a lot even though I have my epsilon set quite low. This is an outputted plot of all the final portfolio sizes from each of the 100 runs:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/1FpAF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1FpAF.png\" alt=\"finalportfolioplot\"><\/a><\/p>\n\n<p>I'd like to see this plot to be in more of an upward trend, rather than all over the place like it is. I've run the model using 500 episodes and 100 episodes, and I've tried extremely adjusting the epsilon, learning rate, and gamma to try and get it to improve upon the last run, but am not successful. Granted I'm still new to q learning. Would someone be able to tell me what to implement to get it working better? And maybe a way to save\/load the models that are generated with the best final portfolio sizes???? Thanks!<\/p>\n\n<pre><code>from yahoo_finance import Share\nfrom matplotlib import pyplot as plt\nimport numpy as np\nimport random\nimport tensorflow as tf\nimport os\nfrom datetime import datetime\nfrom tensorflow.python.framework import ops\nimport warnings\nfrom pandas_datareader import data as web\nimport pandas as pd\n\nwarnings.filterwarnings(\"ignore\")\nos.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'\nops.reset_default_graph()\n\n# create an empty list to store the loss values in so we can plot them later\nfinal_prof_list = []\n\nclass DecisionPolicy:\n    def select_action(self, current_state, step):\n        pass\n\n    def update_q(self, state, action, reward, next_state):\n        pass\n\n\n# class RandomDecisionPolicy(DecisionPolicy):\n#     def __init__(self, actions):\n#         self.actions = actions\n\n#     def select_action(self, current_state, step):\n#         action = self.actions[random.randint(0, len(self.actions) - 1)]\n#         return action\n\n\nclass QLearningDecisionPolicy(DecisionPolicy):\n    def __init__(self, actions, input_dim):\n        self.epsilon = 0.01 # default 0.5, the lesser, the lesser it'll explore\n        self.gamma = 0.99 # typically between 0.8 and 0.99. The larger, the smaller the discount, which means agent cares more about long term reward.\n        self.learning_rate = 0.99 # Higher learning rate means more the agent will learn more from previous step, 0.0001 - 0.99\n        self.actions = actions\n        output_dim = len(actions)\n        h1_dim = 200\n\n        self.x = tf.placeholder(tf.float32, [None, input_dim])\n        self.y = tf.placeholder(tf.float32, [output_dim])\n        W1 = tf.Variable(tf.random_normal([input_dim, h1_dim]))\n        b1 = tf.Variable(tf.constant(0.1, shape=[h1_dim]))\n        h1 = tf.nn.relu(tf.matmul(self.x, W1) + b1)\n        W2 = tf.Variable(tf.random_normal([h1_dim, output_dim]))\n        b2 = tf.Variable(tf.constant(0.1, shape=[output_dim]))\n        self.q = tf.nn.relu(tf.matmul(h1, W2) + b2)\n\n        loss = tf.square(self.y - self.q)\n        self.train_op = tf.train.AdamOptimizer(self.learning_rate).minimize(loss) # GradientDescentOptimizer \n        self.sess = tf.Session()\n        self.sess.run(tf.initialize_all_variables())\n\n    def select_action(self, current_state, step):\n        threshold = min(self.epsilon, step \/ 1000.)\n        if random.random() &lt; threshold:\n            # Exploit best option with probability epsilon\n            action_q_vals = self.sess.run(self.q, feed_dict={self.x: current_state})\n            action_idx = np.argmax(action_q_vals)  # TODO: replace w\/ tensorflow's argmax\n            action = self.actions[action_idx]\n        else:\n            # Explore random option with probability 1 - epsilon\n            action = self.actions[random.randint(0, len(self.actions) - 1)]\n        return action\n\n    def update_q(self, state, action, reward, next_state):\n        action_q_vals = self.sess.run(self.q, feed_dict={self.x: state})\n        next_action_q_vals = self.sess.run(self.q, feed_dict={self.x: next_state})\n        next_action_idx = np.argmax(next_action_q_vals)\n        action_q_vals[0, next_action_idx] = reward + self.gamma * next_action_q_vals[0, next_action_idx]\n        action_q_vals = np.squeeze(np.asarray(action_q_vals))\n        self.sess.run(self.train_op, feed_dict={self.x: state, self.y: action_q_vals})\n\n\ndef run_simulation(policy, initial_budget, initial_num_stocks, prices, hist, debug=False):\n    budget = initial_budget\n    num_stocks = initial_num_stocks\n    shares_to_buy = 10\n    share_value = 0\n    transitions = list()\n    for i in range(len(prices) - hist - 1):\n        if i % 100 == 0:\n            print('progress {:.2f}%'.format(float(100*i) \/ (len(prices) - hist - 1)))\n        current_state = np.asmatrix(np.hstack((prices[i:i+hist], budget, num_stocks)))\n        current_portfolio = budget + num_stocks * share_value\n        action = policy.select_action(current_state, i)\n        share_value = float(prices[i + hist + 1])\n        if action == 'Buy' and budget &gt;= share_value:\n            budget -= share_value\n            num_stocks += shares_to_buy\n        elif action == 'Sell' and num_stocks &gt; 0:\n            budget += share_value\n            num_stocks -= shares_to_buy\n        else:\n            action = 'Hold'\n        new_portfolio = budget + num_stocks * share_value\n        reward = new_portfolio - current_portfolio\n        next_state = np.asmatrix(np.hstack((prices[i+1:i+hist+1], budget, num_stocks)))\n        transitions.append((current_state, action, reward, next_state))\n        policy.update_q(current_state, action, reward, next_state)\n\n    portfolio = budget + num_stocks * share_value\n    if debug:\n        print('${}\\t{} shares'.format(budget, num_stocks))\n    print('Final portfolio size : ', portfolio)\n    final_prof_list.append(portfolio)\n\n    return portfolio\n\n\ndef run_simulations(policy, budget, num_stocks, prices, hist):\n    # Number of runs to run\n    num_tries = 100 # default 100\n    # create a saver to save and try loading models\n    saver = tf.train.Saver()\n\n    final_portfolios = list()\n    for i in range(num_tries):\n        final_portfolio = run_simulation(policy, budget, num_stocks, prices, hist)\n        final_portfolios.append(final_portfolio)\n    avg, std = np.mean(final_portfolios), np.std(final_portfolios)\n    return avg, std\n\n\ndef get_prices(share_symbol, start_date, end_date, cache_filename):\n    try:\n        stock_prices = np.load(cache_filename)\n    except IOError:\n        ex = 'yahoo'\n        df = web.DataReader(share_symbol, ex, start, end)\n        stock_prices = df['Open']\n        df=stock_prices.values\n\n        #share = Share(share_symbol)\n        #stock_hist = share.get_historical(start_date, end_date)\n        #stock_prices = [stock_price['Open'] for stock_price in stock_hist]\n        np.save(cache_filename, df)\n    return stock_prices\n\nif __name__ == '__main__':\n    start = datetime(2016, 1, 1)\n    end = datetime.now()\n    ticker = 'RKDA'\n    prices = get_prices(ticker, start, end, r'.\/resource\/' + ticker + '_stock_prices.npy')\n    actions = ['Buy', 'Sell', 'Hold']\n    hist = 100\n    #policy = RandomDecisionPolicy(actions)\n    policy = QLearningDecisionPolicy(actions, hist + 2)\n    budget = 1000.0\n    num_stocks = 0\n    avg, std = run_simulations(policy, budget, num_stocks, prices, hist)\n    print(avg, std)\n\n    import matplotlib.pyplot as plt\n    plt.plot([np.mean(final_prof_list[i-2:i]) for i in range(len(final_prof_list))])\n    plt.draw()\n    plt.savefig('final_prof_list.png', dpi=100)\n    plt.show()\n\n\n# https:\/\/books.google.ca\/books?id=mEFPDwAAQBAJ&amp;pg=PA477&amp;lpg=PA477&amp;dq=predictive+analytics+with+tensorflow+chapter+11+stock&amp;source=bl&amp;ots=xs_eJywWF7&amp;sig=ACfU3U2TB1-LmmpSwo3iHmQZg8BPlOk1MA&amp;hl=en&amp;sa=X&amp;ved=2ahUKEwjvg6Oht5XkAhWRneAKHe3mAeMQ6AEwDHoECAcQAQ#v=onepage&amp;q&amp;f=true\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57615164,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><loss-function>","CreationDate":"2019-08-22T18:25:19.790","AcceptedAnswerId":57616801.0,"OwnerUserId":11354575.0,"Title":"Difference between batch-average and global Fscore","Body":"<p>I am facing a False Positive Reduction problem, and ratio of the size of positive and negative is approx. 1.7:1. \nI learned from the <a href=\"https:\/\/stackoverflow.com\/questions\/57610804\/when-is-the-timing-to-use-sample-weights-in-keras\">answer<\/a> that using precision, recall, FScore, or even weighting true-positive, false-positive, true-negative and false-negative differently dependent on cost to evaluate different models to deal with specified classification task.<\/p>\n\n<p>Since <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/5794\" rel=\"nofollow noreferrer\">Precision, Recall, and FScore<\/a> are removed from keras, I found some methods to do the tracking of those metrics during training, such as github repo <a href=\"https:\/\/github.com\/netrack\/keras-metrics\" rel=\"nofollow noreferrer\">keras-metrics<\/a>.<\/p>\n\n<p>Besides, I also find ohter solutions by defining precision like this, <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>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<\/code><\/pre>\n\n<p>However, since those methods is tracking the metrics during training, and all of those claim to be <code>batch-wise average<\/code> rather than a global value. \nI wonder how neccessary is that to keep track on those metrics during training. Or I just focus on the <code>loss<\/code> and <code>accuracy<\/code> during training, and evaluate all models using validation functions from like <code>scikit-learn<\/code> to compare those metrics with a global method.<\/p>\n","answers":[{"AnswerId":"57616801","CreationDate":"2019-08-22T20:43:11.867","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>In Keras, all training metrics are measured batch-wise. \nTo obtain a global metric, Keras will average these batch-metrics. <\/p>\n\n<p>Something like <code>sum(batch_metrics) \/ batches<\/code>.   <\/p>\n\n<p>Since most metrics are mean values considering the \"number of samples\", doing that kind of averaging will not change the global value too much. <\/p>\n\n<p>If <code>samples % batch_size == 0<\/code>, then we can say that:<\/p>\n\n<pre><code>sum(all_samples_metrics) \/ samples == sum(all_batch_metrics) \/ batches\n<\/code><\/pre>\n\n<p>But these specific metrics you are talking about are not divided by the \"number of samples\", but by the number of samples \"that satisfy a condition\". Thus, the divisor in each batch is different. Mathematically, the result of averaging the batch-metrics to obtain a global result will not reflect the true global result. <\/p>\n\n<p>So, can we say that they're not good for training?<\/p>\n\n<p>Well, no. They may be good for training. Sometimes \"accuracy\" is a terrible metric for a specific problem.     <\/p>\n\n<p>The key to use these metrics batch-wise is to have a batch size that is big enough to avoid too much variation in the divisors. <\/p>\n"}]}
+{"QuestionId":57615258,"AnswerCount":1,"Tags":"<pytorch><slurm>","CreationDate":"2019-08-22T18:31:55.220","AcceptedAnswerId":57660024.0,"OwnerUserId":568393.0,"Title":"Prevent GPU usage in SLURM when --gpus is not set","Body":"<p>We're using SLURM to manage a small on-premise cluster.  A key resource we are managing is GPUs.  When a user requests GPUs via <code>--gpus=2<\/code> the <code>CUDA_VISIBLE_DEVICES<\/code> environment variable is set with the GPUs SLURM allocates to the user.<\/p>\n\n<pre><code>$ srun --gpus=2 bash -c 'echo $CUDA_VISIBLE_DEVICES'\n0,1\n<\/code><\/pre>\n\n<p>We have a small team and can trust our users to not abuse the system (they could easily overwrite the environment variable) so this works great.  However, it's a bit too easy to bypass this accidentally because when <code>--gpus<\/code> isn't specified <code>$CUDA_VISIBLE_DEVICES<\/code> is left unset so the user can use any GPU (we're typically using PyTorch).<\/p>\n\n<p>In other words, the following command works fine (so long as it lands on a GPU node) but I would prefer that it fails (because no GPU was requested).<\/p>\n\n<pre><code>srun sudo docker run -e CUDA_VISIBLE_DEVICES --runtime=nvidia pytorch\/pytorch:1.1.0-cuda10.0-cudnn7.5-runtime python -c 'import torch; print(torch.tensor([1., 2.], device=torch.device(\"cuda:0\")))'\n<\/code><\/pre>\n\n<p>It would fail if <code>$CUDA_VISIBLE_DEVICES<\/code> were set to <code>-1<\/code>.<\/p>\n\n<pre><code>$ CUDA_VISIBLE_DEVICES=-1 sudo docker run -e CUDA_VISIBLE_DEVICES --runtime=nvidia pytorch\/pytorch:1.1.0-cuda10.0-cudnn7.5-runtime python -c 'import torch; print(torch.tensor([1., 2.], device=torch.device(\"cuda:0\")))'\nTHCudaCheck FAIL file=\/opt\/conda\/conda-bld\/pytorch_1556653099582\/work\/aten\/src\/THC\/THCGeneral.cpp line=51 error=38 : no CUDA-capable device is detected\nTraceback (most recent call last):\n  File \"&lt;string&gt;\", line 1, in &lt;module&gt;\n  File \"\/opt\/conda\/lib\/python3.6\/site-packages\/torch\/cuda\/__init__.py\", line 163, in _lazy_init\n    torch._C._cuda_init()\nRuntimeError: cuda runtime error (38) : no CUDA-capable device is detected at \/opt\/conda\/conda-bld\/pytorch_1556653099582\/work\/aten\/src\/THC\/THCGeneral.cpp:51\n<\/code><\/pre>\n\n<p>How can I configure SLURM to set <code>CUDA_VISIBLE_DEVICES<\/code> to <code>-1<\/code> when <code>--gpus<\/code> is not specified?<\/p>\n","answers":[{"AnswerId":"57660024","CreationDate":"2019-08-26T14:36:09.347","ParentId":null,"OwnerUserId":"1763614","Title":null,"Body":"<p>You can use the <code>TaskProlog<\/code> script to set the <code>$CUDA_VISIBLE_DEVICES<\/code> variable to <code>-1<\/code> if it was not set by Slurm.<\/p>\n\n<p>In <code>slurm.conf<\/code>, configure <code>TaskProlog=\/path\/to\/prolog.sh<\/code> and set the following content for <code>prolog.sh<\/code>.<\/p>\n\n<pre><code>#! \/bin\/bash\n\nif [[ -z $CUDA_VISIBLE_DEVICES]]; then\necho export CUDA_VISIBLE_DEVICES=-1\nfi\n<\/code><\/pre>\n\n<p>The <code>echo export ...<\/code> part will inject <code>CUDA_VISIBLE_DEVICES=-1<\/code> in the job environment.<\/p>\n\n<p>Make sure <code>\/path\/to<\/code> is visible from all compute nodes.\nBut this will not prevent a user from playing the system and redefining the variable from within the Python script. Really preventing access would require configuring <code>cgroups<\/code>.<\/p>\n"}]}
+{"QuestionId":57616114,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-hub>","CreationDate":"2019-08-22T19:41:57.423","AcceptedAnswerId":null,"OwnerUserId":5220257.0,"Title":"Module tesnorflow has no attribute \"UninitializedVariable\" tensorflow hub error","Body":"<p>I have tensorflow version \"1.14.0\" and pip3 is also updated. But when I do:-<\/p>\n\n<pre><code>import tensorflow as tf\n<\/code><\/pre>\n\n<p>I get this error:<\/p>\n\n<pre><code>AttributeError: module 'tensorflow.python.ops.resource_variable_ops' has no attribute 'UninitializedVariable'\n<\/code><\/pre>\n\n<p>Tensorflow hub has version 0.5.0. Some answers on stack overflow have suggested that tensorflow version should be greater than 1.7.0 and mine is greater and still I don't know why is hub not being imported. The complete traceback of error is:-<\/p>\n\n<pre><code>AttributeError                            Traceback (most recent call last)\n&lt;ipython-input-24-d70b4e927300&gt; in &lt;module&gt;\n      1 from sklearn.model_selection import train_test_split\n      2 import tensorflow as tf\n----&gt; 3 import tensorflow_hub as hub\n\n~\/.local\/lib\/python3.6\/site-packages\/tensorflow_hub\/__init__.py in &lt;module&gt;\n     27 # error message is thrown instead of an obscure error of missing\n     28 # symbols at executing the imports.\n---&gt; 29 from tensorflow_hub.estimator import LatestModuleExporter\n     30 from tensorflow_hub.estimator import register_module_for_export\n     31 from tensorflow_hub.feature_column import image_embedding_column\n\n~\/.local\/lib\/python3.6\/site-packages\/tensorflow_hub\/estimator.py in &lt;module&gt;\n     23 from absl import logging\n     24 import tensorflow as tf\n---&gt; 25 from tensorflow_hub import tf_utils\n     26 from tensorflow_hub import tf_v1\n     27 \n\n~\/.local\/lib\/python3.6\/site-packages\/tensorflow_hub\/tf_utils.py in &lt;module&gt;\n     31 # depending on TensorFlow internal implementation details.\n     32 # pylint: disable=g-direct-tensorflow-import\n---&gt; 33 from tensorflow.python.feature_column import feature_column_v2\n     34 # pylint: enable=g-direct-tensorflow-import\n     35 \n\n~\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/feature_column\/feature_column_v2.py in &lt;module&gt;\n    164 from tensorflow.python.platform import tf_logging as logging\n    165 from tensorflow.python.training import checkpoint_utils\n--&gt; 166 from tensorflow.python.training.tracking import tracking\n    167 from tensorflow.python.util import deprecation\n    168 from tensorflow.python.util import nest\n\n~\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/tracking\/tracking.py in &lt;module&gt;\n     21 import weakref\n     22 \n---&gt; 23 from tensorflow.python.eager import def_function\n     24 from tensorflow.python.eager import function as defun\n     25 from tensorflow.python.framework import dtypes\n\n~\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/eager\/def_function.py in &lt;module&gt;\n     38 \n     39 \n---&gt; 40 class UnliftedInitializerVariable(resource_variable_ops.UninitializedVariable):\n     41   \"\"\"Variable which does not lift its initializer out of function context.\n     42 \n\nAttributeError: module 'tensorflow.python.ops.resource_variable_ops' has no attribute 'UninitializedVariable'\n<\/code><\/pre>\n\n<p>I have tried uninstalling tf completely and then installing again, reinstalling tf hub and updating pip so far. Any help will be very much appreciated!<\/p>\n","answers":[]}
+{"QuestionId":57616521,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-08-22T20:17:45.697","AcceptedAnswerId":null,"OwnerUserId":9524309.0,"Title":"How to track the objects removed from a tray in real time using tensorflow","Body":"<p>I am pretty new to raspberry pi and currently have a project at hand which I am confused how to solve it. What I need to accomplish is to detect what items have been removed from a tray in real time. More specifically, the tray will be filled with vegetables and customer will pick up the products from the tray. I need to track the products picked using a raspberry pi. I guess Tensorflow will help me with this but I have seen tutorials where we can track the distance between the camera and object for tracking purposes using OpenCV. This got me even more confused as to which one should I choose. The part that is specifically bothering me is how to track the products picked from the tray in real time. Can someone please give me some idea how to accomplish my task? Thanks a lot in advance. <\/p>\n","answers":[]}
+{"QuestionId":57616698,"AnswerCount":2,"Tags":"<python><numpy><tensorflow>","CreationDate":"2019-08-22T20:33:35.807","AcceptedAnswerId":57625380.0,"OwnerUserId":3736306.0,"Title":"tensorflow pass numpy array to graph using placeholder vs tf.convert_to_tensor()","Body":"<p>If I understand correctly (no), there are two ways to pass a numpy array to a tensorflow computational graph:<\/p>\n\n<ol>\n<li>use <code>tf.placeholder<\/code> and pass this numpy array in the <code>feed_dict<\/code><\/li>\n<li>use <code>tf.convert_to_tensor<\/code>, convert this numpy array as a tensor, then use this tensor for whatever calculation.<\/li>\n<\/ol>\n\n<p>Are there any differences between these two methods?<\/p>\n","answers":[{"AnswerId":"57625264","CreationDate":"2019-08-23T11:17:08.850","ParentId":null,"OwnerUserId":"10878733","Title":null,"Body":"<p><strong>Placeholder tensors:<\/strong><\/p>\n\n<p>As far as I know, <code>tf.placeholder<\/code> creates a placeholder tensor into the default <code>tf.Graph<\/code> object. A <code>placeholder<\/code> tensor basically <strong>creates a pathway\/entrance<\/strong> for values to enter the graph. They act as inputs to a graph which may correspond to a model too.\nThe values ( not belonging to the graph ) are fed to the placeholders using the <code>feed_dict<\/code> mentioned by you.<\/p>\n\n<p><strong><code>tf.convert_to_tensor()<\/code><\/strong> :<\/p>\n\n<p>As mentioned in the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/convert_to_tensor\" rel=\"nofollow noreferrer\">official docs<\/a>,<\/p>\n\n<blockquote>\n  <p>This function converts Python objects of various types to Tensor\n  objects. It accepts Tensor objects, numpy arrays, Python lists, and\n  Python scalars.<\/p>\n<\/blockquote>\n\n<p>It converts an existing NumPy array or Python object to a corresponding tensor ( with same dtype and shape ). This tensor will not be a placeholder tensor. It's just a utility to convert NumPy arrays and create <code>tf.Tensor<\/code> objects out of them.<\/p>\n\n<p>Hence, placeholder tensors allow you feed NumPy arrays or Python objects directly to the graph using <code>feed_dict<\/code>. If your data is static then you can include it in the graph by creating a tensor out of it using <code>tf.convert_to_tensor()<\/code>.<\/p>\n"},{"AnswerId":"57625380","CreationDate":"2019-08-23T11:24:30.013","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p><code>tf_convert_to_tensor<\/code> is highly unpractical because it does not scale. See the example below:<\/p>\n\n<pre><code>X = np.random.rand(3,3)\nY = np.random.rand(3,3)\nX_tensor = tf.convert_to_tensor(X)\nX_squared = tf.square(X_tensor)\nY_tensor = tf.convert_to_tensor(Y)\nY_squared = tf.square(Y)\nwith tf.Session() as sess:\n    x = sess.run(X_squared)\n    y = sess.run(Y_squared)\n<\/code><\/pre>\n\n<p>As you can see, for every Numpy array that we want to provide as input to the graph, we have to create a new <code>tf.convert_to_tensor<\/code> op. One example where this is fine is if you have a static input that won't change, then you have a single <code>tf.convert_to_tensor<\/code> op that you use. On the other hand, let's have a look at the same example using <code>tf.placeholder<\/code>:<\/p>\n\n<pre><code>X = np.random.rand(3,3)\nY = np.random.rand(3,3)\ngraph_input = tf.placeholder(shape=[None, None], dtype=tf.float64)\nsquared = tf.square(graph_input)\nwith tf.Session() as sess:\n    x = sess.run(squared, feed_dict={graph_input: X})\n    y = sess.run(squared, feed_dict={graph_input: Y})\n<\/code><\/pre>\n\n<p>As you can see, we use the <code>tf.placeholder<\/code> to dynamically provide input data to the graph. Consider it as a funnel, that you use to pour data in the graph. The <code>tf.placeholder<\/code> is fine for data that might\/will change throughout the training such as learning rate, dropout probability, etc.<\/p>\n"}]}
+{"QuestionId":57617276,"AnswerCount":0,"Tags":"<python><c><tensorflow>","CreationDate":"2019-08-22T21:28:49.993","AcceptedAnswerId":null,"OwnerUserId":717738.0,"Title":"Python extension module using TensorFlow C API","Body":"<p>I've written a Python extension module that uses TensorFlow through the C API.  I installed the API as described in <a href=\"https:\/\/www.tensorflow.org\/install\/lang_c\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/lang_c<\/a>.  On its own, my module works correctly.  But if I import my extension module and <em>also<\/em> import tensorflow, Python crashes with this error.<\/p>\n\n<pre><code>[libprotobuf ERROR external\/protobuf_archive\/src\/google\/protobuf\/descriptor_database.cc:58] File already exists in database: tensorflow\/core\/protobuf\/master.proto\n[libprotobuf FATAL external\/protobuf_archive\/src\/google\/protobuf\/descriptor.cc:1370] CHECK failed: GeneratedDatabase()-&gt;Add(encoded_file_descriptor, size): \nlibc++abi.dylib: terminating with uncaught exception of type google::protobuf::FatalException: CHECK failed: GeneratedDatabase()-&gt;Add(encoded_file_descriptor, size): \nAbort trap: 6\n<\/code><\/pre>\n\n<p>I believe this happens because the C API includes its own complete copy of TensorFlow, so now I get two different copies loaded into the same process at the same time.<\/p>\n\n<p>What is the solution to this?  How can I have Python code that uses TensorFlow, and also invokes C code that uses TensorFlow?<\/p>\n","answers":[]}
+{"QuestionId":57617294,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-22T21:30:37.700","AcceptedAnswerId":57629472.0,"OwnerUserId":3736306.0,"Title":"tensorflow element-wise multiplication broadcasting?","Body":"<p>Does tensorflow offer any function for element-wise multiplication broadcasting on the last dimension?<\/p>\n\n<p>Here is an example of what I'm trying to do and what does not work:<\/p>\n\n<pre><code>import tensorflow as tf\nx = tf.constant(5, shape=(1, 200, 175, 6), dtype=tf.float32)\ny = tf.constant(1, shape=(1, 200, 175), dtype=tf.float32)\ntf.math.multiply(x, y)\n<\/code><\/pre>\n\n<p>Essentially, I would like for each one of <code>x<\/code>'s slice along the last dimension, do an element-wise matrix multiplication with <code>y<\/code>.<\/p>\n\n<p>I have found this question asking similar operation: <a href=\"https:\/\/stackoverflow.com\/questions\/34192229\/efficient-element-wise-multiplication-of-a-matrix-and-a-vector-in-tensorflow\">Efficient element-wise multiplication of a matrix and a vector in TensorFlow<\/a><\/p>\n\n<p>Unfortunately, the suggested approach (using <code>tf.multiply()<\/code> ) no longer work now. The corresponding <code>tf.math.multiply<\/code> also does not work, as the code above gives me the error below:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/home\/yuqiong\/miniconda3\/envs\/deep\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1864, in _create_c_op\n    c_op = c_api.TF_FinishOperation(op_desc)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: Dimensions must be equal, but are 175 and 200 for 'Mul' (op: 'Mul') with input shapes: [1,200,175,6], [1,200,175].\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\/yuqiong\/miniconda3\/envs\/deep\/lib\/python3.7\/site-packages\/tensorflow\/python\/util\/dispatch.py\", line 180, in wrapper\n    return target(*args, **kwargs)\n  File \"\/home\/yuqiong\/miniconda3\/envs\/deep\/lib\/python3.7\/site-packages\/tensorflow\/python\/ops\/math_ops.py\", line 322, in multiply\n    return gen_math_ops.mul(x, y, name)\n  File \"\/home\/yuqiong\/miniconda3\/envs\/deep\/lib\/python3.7\/site-packages\/tensorflow\/python\/ops\/gen_math_ops.py\", line 6490, in mul\n    \"Mul\", x=x, y=y, name=name)\n  File \"\/home\/yuqiong\/miniconda3\/envs\/deep\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 788, in _apply_op_helper\n    op_def=op_def)\n  File \"\/home\/yuqiong\/miniconda3\/envs\/deep\/lib\/python3.7\/site-packages\/tensorflow\/python\/util\/deprecation.py\", line 507, in new_func\n    return func(*args, **kwargs)\n  File \"\/home\/yuqiong\/miniconda3\/envs\/deep\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 3616, in create_op\n    op_def=op_def)\n  File \"\/home\/yuqiong\/miniconda3\/envs\/deep\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 2027, in __init__\n    control_input_ops)\n  File \"\/home\/yuqiong\/miniconda3\/envs\/deep\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1867, in _create_c_op\n    raise ValueError(str(e))\nValueError: Dimensions must be equal, but are 175 and 200 for 'Mul' (op: 'Mul') with input shapes: [1,200,175,6], [1,200,175].\n<\/code><\/pre>\n\n<p>I can think of a working approach: duplicate <code>y<\/code> 6 times so it has the exact same shape as <code>x<\/code>, then do element-wise multiplication.<\/p>\n\n<p>But are there faster and memory efficient way of doing this in tensorflow?<\/p>\n","answers":[{"AnswerId":"57629472","CreationDate":"2019-08-23T15:47:05.127","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>This should achieve what you want:<\/p>\n\n<pre><code>x = np.array([[[1,2,3],[4,5,6],[7,8,9],[10,11,12]]])\n# [[[ 1  2  3]\n#   [ 4  5  6]\n#   [ 7  8  9]\n#   [10 11 12]]]\ny = np.array([[1,2,3,4]])\n# [[1 2 3 4]]\ny = tf.expand_dims(y, axis=-1)\nmul = tf.multiply(x, y)\n# [[[ 1  2  3]\n#   [ 8 10 12]\n#   [21 24 27]\n#   [40 44 48]]]\n<\/code><\/pre>\n\n<p>Finally, using the shapes you need:<\/p>\n\n<pre><code>x = np.random.rand(1, 200, 175, 6)\ny = np.random.rand(1, 200, 175)\ny = tf.expand_dims(y, axis=-1)\nmul = tf.multiply(x, y)\nwith tf.Session() as sess:\n    print(sess.run(mul).shape)\n    # (1, 200, 175, 6)\n\u200b\n<\/code><\/pre>\n"}]}
+{"QuestionId":57617635,"AnswerCount":1,"Tags":"<tensorflow><matrix><keras><conv-neural-network><reshape>","CreationDate":"2019-08-22T22:15:25.383","AcceptedAnswerId":57626310.0,"OwnerUserId":11964890.0,"Title":"Reshaping layer in Keras","Body":"<p>I am trying to build a convolutional neural network with an output matrix. The input shape is (100,100,4) and the output shape is (2,125).<\/p>\n\n<p>Here is the summary of my current model:<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_63 (InputLayer)        (None, 100, 100, 4)       0         \n_________________________________________________________________\nconv2d_44 (Conv2D)           (None, 100, 100, 25)      2525      \n_________________________________________________________________\nmax_pooling2d_38 (MaxPooling (None, 50, 50, 25)        0         \n_________________________________________________________________\nflatten_38 (Flatten)         (None, 62500)             0         \n_________________________________________________________________\ndense_47 (Dense)             (None, 10)                625010    \n_________________________________________________________________\ndense_48 (Dense)             (None, 250)               2750      \n_________________________________________________________________\nreshape_63 (Reshape)         (None, 2, 125)            0         \n=================================================================\nTotal params: 630,285\nTrainable params: 630,285\nNon-trainable params: 0\n_________________________________________________________________\nNone\n<\/code><\/pre>\n\n<p>Which I thought should be fine but when I tried to fit the model I got this error:<\/p>\n\n<pre><code>ValueError: Error when checking target: expected reshape_62 to have shape (2, 1) but got array with shape (2, 125)\n<\/code><\/pre>\n\n<p>Here's the code I used<\/p>\n\n<pre><code>batch_size = 100\n\ninput_layer     = Input(shape=(xs[1],xs[2],xs[3]))\nconv1           = Conv2D(filters = 25, kernel_size = 5,padding=\"same\",activation=\"relu\", data_format = 'channels_last')(input_layer)\npool1           = MaxPooling2D(pool_size=(2,2),padding=\"same\")(conv1)\nflat            = Flatten()(pool1)\nhidden1         = Dense(10, activation='relu')(flat)\noutput_layer    = Dense(ys[1]*ys[2], activation='softmax')(hidden1)\noutput_reshape  = Reshape((2,125))(output_layer)\nmodel           = Model(inputs=input_layer, outputs=output_reshape)\nprint(model.summary())\n\nmodel.compile(loss='sparse_categorical_crossentropy', optimizer='adam', sample_weight_mode='temporal')\nmodel.fit(x_train,y_train,batch_size=batch_size,epochs=3)\n<\/code><\/pre>\n\n<p>I have been looking up how the reshape layer work but I still couldn't figure it out. Any help would be most appreciated. <\/p>\n","answers":[{"AnswerId":"57626310","CreationDate":"2019-08-23T12:25:35.807","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>This is happening because you are using <code>'sparse_categorical_crossentropy'<\/code>.   <\/p>\n\n<p>\"Sparse\" means that the system will not expect an entire array, but just the cordinate of the hot spot. Instead of expecting a regular <code>(None, 2, 125)<\/code> tensor, it will expect just <code>(None, 2, 1)<\/code> indicating which of the 125 classes is the correct one. <\/p>\n\n<p>To fix this, either you start using a sparse <code>y_train<\/code>, or you replace your loss with <code>'categorical_crossentropy'<\/code>. <\/p>\n\n<p>I believe a sparse <code>y_train<\/code> can be obtained with <code>sparse_y_train = numpy.argmax(y_train, axis=-1)<\/code>. If this model does not gives you memory problems, you don't need to go sparse.   <\/p>\n"}]}
+{"QuestionId":57618220,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-08-22T23:41:46.180","AcceptedAnswerId":57618347.0,"OwnerUserId":4807279.0,"Title":"How we can define Early Stopping for Keras to check after each batch (not entire epoch)","Body":"<p>Normally Early stopping for Keras can be defined to check if it reaches a limit (loss, accuracy) after each epoch. But in mini batch method each epoch contains multiple loss, acc for training and validation phase. How we can tell Keras to check every acc or loss in an epoch for early stopping? <\/p>\n\n<pre><code>filepath=\"weights.best.hdf5\"\n\ncheckpoint = ModelCheckpoint(filepath, monitor='acc', verbose=1, save_best_only=True)\n\nearly_stop = EarlyStopping(monitor='acc', patience=5) \n\ncallbacks_list = [checkpoint, early_stop]\n\nhistory = model.fit(x, y, validation_data=(x_test, y_test), epochs=1, callbacks=callbacks_list)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57618347","CreationDate":"2019-08-23T00:04:06.760","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Copy the code for <code>EarlyStopping<\/code> from the <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/callbacks.py#L733\" rel=\"nofollow noreferrer\">source<\/a> and change <code>on_epoch_end<\/code> with <code>on_batch_end<\/code>. Update a few other things that talks about <code>epoch<\/code> and it's ready:<\/p>\n\n<pre><code>class BatchEarlyStopping(Callback):\n    \"\"\"Stop training when a monitored quantity has stopped improving.\n    # Arguments\n        monitor: quantity to be monitored.\n        min_delta: minimum change in the monitored quantity\n            to qualify as an improvement, i.e. an absolute\n            change of less than min_delta, will count as no\n            improvement.\n        patience: number of batches with no improvement\n            after which training will be stopped.\n        verbose: verbosity mode.\n        mode: one of {auto, min, max}. In `min` mode,\n            training will stop when the quantity\n            monitored has stopped decreasing; in `max`\n            mode it will stop when the quantity\n            monitored has stopped increasing; in `auto`\n            mode, the direction is automatically inferred\n            from the name of the monitored quantity.\n        baseline: Baseline value for the monitored quantity to reach.\n            Training will stop if the model doesn't show improvement\n            over the baseline.\n        restore_best_weights: whether to restore model weights from\n            the batch with the best value of the monitored quantity.\n            If False, the model weights obtained at the last step of\n            training are used.\n    \"\"\"\n\n    def __init__(self,\n                 monitor='val_loss',\n                 min_delta=0,\n                 patience=0,\n                 verbose=0,\n                 mode='auto',\n                 baseline=None,\n                 restore_best_weights=False):\n        super(BatchEarlyStopping, self).__init__()\n\n        self.monitor = monitor\n        self.baseline = baseline\n        self.patience = patience\n        self.verbose = verbose\n        self.min_delta = min_delta\n        self.wait = 0\n        self.stopped_batch = 0\n        self.restore_best_weights = restore_best_weights\n        self.best_weights = None\n\n        if mode not in ['auto', 'min', 'max']:\n            warnings.warn('BatchEarlyStopping mode %s is unknown, '\n                          'fallback to auto mode.' % mode,\n                          RuntimeWarning)\n            mode = 'auto'\n\n        if mode == 'min':\n            self.monitor_op = np.less\n        elif mode == 'max':\n            self.monitor_op = np.greater\n        else:\n            if 'acc' in self.monitor:\n                self.monitor_op = np.greater\n            else:\n                self.monitor_op = np.less\n\n        if self.monitor_op == np.greater:\n            self.min_delta *= 1\n        else:\n            self.min_delta *= -1\n\n    def on_train_begin(self, logs=None):\n        # Allow instances to be re-used\n        self.wait = 0\n        self.stopped_batch = 0\n        if self.baseline is not None:\n            self.best = self.baseline\n        else:\n            self.best = np.Inf if self.monitor_op == np.less else -np.Inf\n\n    def on_batch_end(self, batch, logs=None):\n        current = self.get_monitor_value(logs)\n        if current is None:\n            return\n\n        if self.monitor_op(current - self.min_delta, self.best):\n            self.best = current\n            self.wait = 0\n            if self.restore_best_weights:\n                self.best_weights = self.model.get_weights()\n        else:\n            self.wait += 1\n            if self.wait &gt;= self.patience:\n                self.stopped_batch = batch\n                self.model.stop_training = True\n                if self.restore_best_weights:\n                    if self.verbose &gt; 0:\n                        print('Restoring model weights from the end of '\n                              'the best batch')\n                    self.model.set_weights(self.best_weights)\n\n    def on_train_end(self, logs=None):\n        if self.stopped_batch &gt; 0 and self.verbose &gt; 0:\n            print('Batch %05d: early stopping' % (self.stopped_batch + 1))\n\n    def get_monitor_value(self, logs):\n        monitor_value = logs.get(self.monitor)\n        if monitor_value is None:\n            warnings.warn(\n                'Early stopping conditioned on metric `%s` '\n                'which is not available. Available metrics are: %s' %\n                (self.monitor, ','.join(list(logs.keys()))), RuntimeWarning\n            )\n        return monitor_value\n<\/code><\/pre>\n"}]}
+{"QuestionId":57618301,"AnswerCount":2,"Tags":"<python><image><tensorflow><image-processing><scikit-image>","CreationDate":"2019-08-22T23:57:27.640","AcceptedAnswerId":57619682.0,"OwnerUserId":7850499.0,"Title":"Convert image from float64 to uint8 makes the image look darker","Body":"<p>I have images of type float64 generated by GANs, and I save them through <a href=\"https:\/\/scikit-image.org\/docs\/dev\/api\/skimage.io.html#skimage.io.imsave\" rel=\"nofollow noreferrer\"><code>skimage.io.imsave<\/code><\/a>. The process works well and the saved image looks nice, but I get a warning message as follows:<\/p>\n\n<blockquote>\n  <p>Lossy conversion from float64 to uint8. Range [-0.9999998807907104, 0.9999175071716309]. Convert image to uint8 prior to saving to suppress this warning.<\/p>\n<\/blockquote>\n\n<p>Then I try to get rid of this warning by convert images to uint8 before saving using function <a href=\"https:\/\/scikit-image.org\/docs\/dev\/api\/skimage.html#skimage.img_as_ubyte\" rel=\"nofollow noreferrer\"><code>skimage.img_as_ubyte<\/code><\/a>. This gives me a apparently much darker image with a warning<\/p>\n\n<blockquote>\n  <p>UserWarning: Possible precision loss when converting from float64 to uint8\n    .format(dtypeobj_in, dtypeobj_out))<\/p>\n<\/blockquote>\n\n<p>I've also tried to use other functions such as the one from tensorflow <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/image\/convert_image_dtype\" rel=\"nofollow noreferrer\"><code>tf.image.convert_image_dtype<\/code><\/a> before saving. They all return a darker image than I directly call skimage.io.imsave. What's the problem here?<\/p>\n\n<p>Here's a set of images generated by converting to uint8 before saving\n<a href=\"https:\/\/i.stack.imgur.com\/IO0XQ.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/IO0XQ.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Here's a set of images generated by saving directly\n<a href=\"https:\/\/i.stack.imgur.com\/Q2Yz3.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Q2Yz3.jpg\" alt=\"\"><\/a><\/p>\n","answers":[{"AnswerId":"58641486","CreationDate":"2019-10-31T10:36:21.820","ParentId":null,"OwnerUserId":"1323473","Title":null,"Body":"<p>I fix this warning by using,<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport imageio\n\n# suppose that img's dtype is 'float64'\nimg_uint8 = img.astype(np.uint8)\n# and then\nimageio.imwrite('filename.jpg', img_uint8)\n<\/code><\/pre>\n\n<p>That's it!<\/p>\n"},{"AnswerId":"57619682","CreationDate":"2019-08-23T03:50:15.890","ParentId":null,"OwnerUserId":"7328782","Title":null,"Body":"<p>From <a href=\"https:\/\/scikit-image.org\/docs\/dev\/api\/skimage.html#skimage.img_as_ubyte\" rel=\"noreferrer\">the documentation of <code>skimage.img_as_ubyte<\/code> that you linked<\/a>:<\/p>\n\n<blockquote>\n  <p>Negative input values will be clipped. Positive values are scaled between 0 and 255.<\/p>\n<\/blockquote>\n\n<p>Since your images are in the range [-1,1], half of the data will be set to 0, which is why stuff looks darker. Try first scaling your image to a positive-only range, for example by adding 1 to it, before calling <code>skimage.img_as_ubyte<\/code>. <\/p>\n"}]}
+{"QuestionId":57618482,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><loss-function><loss>","CreationDate":"2019-08-23T00:31:46.037","AcceptedAnswerId":57618531.0,"OwnerUserId":11965170.0,"Title":"Higher loss penalty for true non-zero predictions","Body":"<p>I am building a deep regression network (CNN) to predict a (1000,1) target vector from images (7,11). The target usually consists of about <strong>90 % zeros<\/strong> and only <strong>10 %<\/strong> non-zero values. The distribution of (non-) zero values in the targets vary from sample to sample (i.e. there is no global class imbalance).<\/p>\n\n<p>Using mean sqaured error loss, this led to the network predicting only zeros, which I don't find surprising.<\/p>\n\n<p>My best guess is to write a custom loss function that penalizes errors regarding non-zero values more than the prediction of zero-values.<\/p>\n\n<p>I have tried this loss function with the intend to implement what I have guessed could work above. It is a mean squared error loss in which the predictions of non-zero targets are penalized less (w=0.1).<\/p>\n\n<pre><code>def my_loss(y_true, y_pred):\n    # weights true zero predictions less than true nonzero predictions\n    w = 0.1\n    y_pred_of_nonzeros = tf.where(tf.equal(y_true, 0), y_pred-y_pred, y_pred)\n    return K.mean(K.square(y_true-y_pred_of_nonzeros)) + K.mean(K.square(y_true-y_pred))*w\n\n<\/code><\/pre>\n\n<p>The network is able to learn without getting stuck with only-zero predictions. However, this solution seems quite <strong>unclean<\/strong>. Is there a better way to deal with this type of problem? Any advice on improving the custom loss function?\nAny suggestions are welcome, thank you in advance!<\/p>\n\n<p>Best, \nLukas<\/p>\n","answers":[{"AnswerId":"57618531","CreationDate":"2019-08-23T00:44:01.373","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Not sure there is anything better than a custom loss just like you did, but there is a cleaner way:<\/p>\n\n<pre><code>def weightedLoss(w):\n\n    def loss(true, pred):\n\n        error = K.square(true - pred)\n        error = K.switch(K.equal(true, 0), w * error , error)\n\n        return error \n\n    return loss\n<\/code><\/pre>\n\n<p>You may also <code>return K.mean(error)<\/code>, but without <code>mean<\/code> you can still profit from other Keras options like adding sample weights and other things. <\/p>\n\n<p>Select the weight when compiling:<\/p>\n\n<pre><code>model.compile(loss = weightedLoss(0.1), ...)\n<\/code><\/pre>\n\n<p>If you have the entire data in an array, you can do:<\/p>\n\n<pre><code>w = K.mean(y_train)\nw = w \/ (1 - w) #this line compesates the lack of the 90% weights for class 1\n<\/code><\/pre>\n\n<hr>\n\n<p>Another solution that can avoid using a custom loss, but requires changes in the data and the model is: <\/p>\n\n<ul>\n<li>Transform your <code>y<\/code> into a 2-class problem for each output. Shape = <code>(batch, originalClasses, 2)<\/code>.   <\/li>\n<\/ul>\n\n<p>For the zero values, make the first of the two classes = 1<br>\nFor the one values, make the second of the two classes = 1   <\/p>\n\n<pre><code>newY = np.stack([1-oldY, oldY], axis=-1)    \n<\/code><\/pre>\n\n<p>Adjust the model to output this new shape. <\/p>\n\n<pre><code>...\nmodel.add(Dense(2*classes))\nmodel.add(Reshape((classes,2)))\nmodel.add(Activation('softmax'))\n<\/code><\/pre>\n\n<p>Make sure you are using a <code>softmax<\/code> and a <code>categorical_crossentropy<\/code> as loss.    <\/p>\n\n<p>Then use the argument <code>class_weight={0: w, 1: 1}<\/code> in <code>fit<\/code>. <\/p>\n"}]}
+{"QuestionId":57618507,"AnswerCount":2,"Tags":"<python><deep-learning><pytorch><fast-ai>","CreationDate":"2019-08-23T00:38:41.610","AcceptedAnswerId":57876684.0,"OwnerUserId":1870832.0,"Title":"fastai error predicting with exported\/reloaded model: \"Input type and weight type should be the same\"","Body":"<p>Whenever I export a fastai model and reload it, I get this error (or a very similar one) when I try and use the reloaded model to generate predictions on a new test set:<\/p>\n\n<pre><code>RuntimeError: Input type (torch.cuda.FloatTensor) and weight type (torch.cuda.HalfTensor) should be the same\n<\/code><\/pre>\n\n<p>Minimal reprodudeable code example below, you just need to update your <code>FILES_DIR<\/code> variable to where the MNIST data gets deposited on your system:<\/p>\n\n<pre><code>from fastai import *\nfrom fastai.vision import *\n\n# download data for reproduceable example\nuntar_data(URLs.MNIST_SAMPLE)\nFILES_DIR = '\/home\/mepstein\/.fastai\/data\/mnist_sample'  # this is where command above deposits the MNIST data for me\n\n\n# Create FastAI databunch for model training\ntfms = get_transforms()\ntr_val_databunch = ImageDataBunch.from_folder(path=FILES_DIR,  # location of downloaded data shown in log of prev command\n                                train = 'train',\n                                valid_pct = 0.2,\n                                ds_tfms = tfms).normalize()\n\n# Create Model\nconv_learner = cnn_learner(tr_val_databunch, \n                           models.resnet34, \n                           metrics=[error_rate]).to_fp16()\n\n# Train Model\nconv_learner.fit_one_cycle(4)\n\n# Export Model\nconv_learner.export()  # saves model as 'export.pkl' in path associated with the learner\n\n# Reload Model and use it for inference on new hold-out set\nreloaded_model = load_learner(path = FILES_DIR,\n                              test = ImageList.from_folder(path = f'{FILES_DIR}\/valid'))\n\npreds = reloaded_model.get_preds(ds_type=DatasetType.Test)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<blockquote>\n  <p>\"RuntimeError: Input type (torch.cuda.FloatTensor) and weight type\n  (torch.cuda.HalfTensor) should be the same\"<\/p>\n<\/blockquote>\n\n<p>Stepping through the code statement by statement, everything works fine until the last line <code>pred = ...<\/code> which is where the torch error above pops up. Any help would be much appreciated!<\/p>\n\n<p>relevant software versions:<\/p>\n\n<p>Python 3.7.3\nfastai           1.0.57<br>\ntorch            1.2.0<br>\ntorchvision      0.4.0 <\/p>\n","answers":[{"AnswerId":"57623218","CreationDate":"2019-08-23T09:16:51.040","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>Your model is in half precision if you have <code>.to_fp16<\/code>, which would be the same if you would <code>model.half()<\/code> in PyTorch.<\/p>\n\n<p>Actually if you trace the code <code>.to_fp16<\/code> will call <code>model.half()<\/code>\nBut there is a problem. If you convert the batch norm layer also to half precision you may get the convergence problem.<\/p>\n\n<p>This is why you would typically do this in PyTorch:<\/p>\n\n<pre><code>model.half()  # convert to half precision\nfor layer in model.modules():\n  if isinstance(module, torch.nn.modules.batchnorm._BatchNorm):      \n    layer.float()\n<\/code><\/pre>\n\n<p>This will convert any layer to half precision other than batch norm.<\/p>\n\n<p>Note that code from <a href=\"https:\/\/discuss.pytorch.org\/t\/training-with-half-precision\/11815\/2\" rel=\"nofollow noreferrer\">PyTorch forum<\/a> is also OK, but just for <code>nn.BatchNorm2d<\/code>.<\/p>\n\n<p>Then make sure your input is in half precision using <code>to()<\/code> like this:<\/p>\n\n<pre><code>import torch\nt = torch.tensor(10.)\nprint(t)\nprint(t.dtype)\nt=t.to(dtype=torch.float16)\nprint(t)\nprint(t.dtype)\n# tensor(10.)\n# torch.float32\n# tensor(10., dtype=torch.float16)\n# torch.float16\n<\/code><\/pre>\n"},{"AnswerId":"57876684","CreationDate":"2019-09-10T18:49:47.103","ParentId":null,"OwnerUserId":"1870832","Title":null,"Body":"<p>So the answer to this ended up being relatively simple:<\/p>\n\n<p>1) As noted in my comment, training in mixed precision mode (setting <code>conv_learner<\/code> <code>to_fp16()<\/code>) caused the error with the exported\/reloaded model<\/p>\n\n<p>2) To train in mixed precision mode (which is faster than regular training) and enable export\/reload of the model without errors, simply set the model back to default precision before exporting.<\/p>\n\n<p>...In code, simply changing the example above:<\/p>\n\n<pre><code># Export Model\nconv_learner.export()\n<\/code><\/pre>\n\n<p>to:<\/p>\n\n<pre><code># Export Model (after converting back to default precision for safe export\/reload\nconv_learner = conv_learner.to_fp32()\nconv_learner.export()\n<\/code><\/pre>\n\n<p>...and now the full (reproduceable) code example above runs without errors, including the prediction after model reload.<\/p>\n"}]}
+{"QuestionId":57618665,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2019-08-23T01:10:57.197","AcceptedAnswerId":null,"OwnerUserId":11053801.0,"Title":"How to find the class of the keras model predict_prob function?","Body":"<p>I have the keras sequential model, target classes are (0,1), when I do prediction from single test record, I get the probability for a single class, how should I know which class the probability belongs to?<\/p>\n\n<pre><code>keras_model.predict_proba(test_for_model)\nOut[30]: array([[0.77143025]], dtype=float32)\n\nkeras_model.predict_proba(test_for_model).argmax(axis=-1)\nOut[33]: array([0], dtype=int64)\n\nkeras_model.predict_classes(test_for_model)\nOut[32]: array([[1]])\n<\/code><\/pre>\n\n<p>predict_classes says 1, argmax says 0, may I know which class the 77% belongs either 0 or 1?<\/p>\n","answers":[]}
+{"QuestionId":57619076,"AnswerCount":0,"Tags":"<python><machine-learning><pytorch><tensor><autograd>","CreationDate":"2019-08-23T02:16:32.530","AcceptedAnswerId":null,"OwnerUserId":6688984.0,"Title":"How to use pytorch to automatically adjust matrices that creates a desired value?","Body":"<p>I have two tensors (matrices) that I've initialized: <\/p>\n\n<pre><code>sm=Var(torch.randn(20,1),requires_grad=True)\nsm = torch.mm(sm,sm.t()) \nfreq_m=Var(torch.randn(12,20),requires_grad=True)\n<\/code><\/pre>\n\n<p>I am creating two lists from the data inside these 2 matrices, and I am using spearmanr to get a correlation value between these 2 lists. How I am creating the lists is not important, but the goal is to adjust the values inside the matrices so that the calculated correlation value is as close to 1 as possible.  <\/p>\n\n<p>If I were to solve this problem manually, I would tweak values in the matrices by .01 (or some small number) each time and recalculate the lists and correlation score. If the new correlation value is higher than the previous one, I would save the 2 matrices and tweak a different value until I get the 2 matrices that give me the highest correlation score possible. <\/p>\n\n<p>I believe PyTorch is capable of doing this automatically, but I'm not sure how. Any guidance would be appreciated!<\/p>\n","answers":[]}
+{"QuestionId":57619347,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><deeplab>","CreationDate":"2019-08-23T03:00:56.920","AcceptedAnswerId":null,"OwnerUserId":11190321.0,"Title":"There was a problem running the test text 'local_test.sh' in deeplab.But modle_test.py works fine","Body":"<p>I copied deeplab's source code on github and configured all the files as required.And it works fine for model_test.py.But when I tried to run the local_tset.sh test file, a series of problems occurred.<\/p>\n\n<p>I can't read the error message, so I don't know what went wrong, and I don't know where to start<\/p>\n\n<pre><code>2019-08-23 10:39:16.486931: W tensorflow\/core\/common_runtime\/bfc_allocator.cc:319] *************************************************____********___****____************************xxxx\n2019-08-23 10:39:16.487253: W tensorflow\/core\/framework\/op_kernel.cc:1502] OP_REQUIRES failed at depthwise_conv_op.cc:365 : Resource exhausted: OOM when allocating tensor with shape[4,128,257,257] and type float on \/job:localhost\/replica:0\/task:0\/device:GPU:0 by allocator GPU_0_bfc\nTraceback (most recent call last):\n  File \"E:\\anaconda\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1356, in _do_call\n    return fn(*args)\n  File \"E:\\anaconda\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1341, in _run_fn\n    options, feed_dict, fetch_list, target_list, run_metadata)\n  File \"E:\\anaconda\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1429, in _call_tf_sessionrun\n    run_metadata)\ntensorflow.python.framework.errors_impl.ResourceExhaustedError: 2 root error(s) found.\n  (0) Resource exhausted: OOM when allocating tensor with shape[4,128,257,257] and type float on \/job:localhost\/replica:0\/task:0\/device:GPU:0 by allocator GPU_0_bfc\n         [[{{node xception_65\/entry_flow\/block1\/unit_1\/xception_module\/separable_conv2_depthwise\/depthwise}}]]\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.\n\n         [[gradients\/AddN_56\/_12764]]\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.\n\n  (1) Resource exhausted: OOM when allocating tensor with shape[4,128,257,257] and type float on \/job:localhost\/replica:0\/task:0\/device:GPU:0 by allocator GPU_0_bfc\n         [[{{node xception_65\/entry_flow\/block1\/unit_1\/xception_module\/separable_conv2_depthwise\/depthwise}}]]\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.\n\n0 successful operations.\n0 derived errors ignored.\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"E:\/models-master\/research\/deeplab\/train.py\", line 517, in &lt;module&gt;\n    tf.app.run()\n  ...\n  ...\n  File \"E:\\anaconda\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1370, in _do_call\n    raise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.ResourceExhaustedError: 2 root error(s) found.\n  (0) Resource exhausted: OOM when allocating tensor with shape[4,128,257,257] and type float on \/job:localhost\/replica:0\/task:0\/device:GPU:0 by allocator GPU_0_bfc\n         [[node xception_65\/entry_flow\/block1\/unit_1\/xception_module\/separable_conv2_depthwise\/depthwise (defined at \\models-master\\research\\deeplab\\core\\xception.py:175) ]]\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.\n\n         [[gradients\/AddN_56\/_12764]]\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.\n\n  (1) Resource exhausted: OOM when allocating tensor with shape[4,128,257,257] and type float on \/job:localhost\/replica:0\/task:0\/device:GPU:0 by allocator GPU_0_bfc\n         [[node xception_65\/entry_flow\/block1\/unit_1\/xception_module\/separable_conv2_depthwise\/depthwise (defined at \\models-master\\research\\deeplab\\core\\xception.py:175) ]]\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.\n\n0 successful operations.\n0 derived errors ignored.\n\nErrors may have originated from an input operation.\nInput Source operations connected to node xception_65\/entry_flow\/block1\/unit_1\/xception_module\/separable_conv2_depthwise\/depthwise:\n xception_65\/entry_flow\/block1\/unit_1\/xception_module\/Relu_1 (defined at \\models-master\\research\\deeplab\\core\\xception.py:274)\n\nInput Source operations connected to node xception_65\/entry_flow\/block1\/unit_1\/xception_module\/separable_conv2_depthwise\/depthwise:\n xception_65\/entry_flow\/block1\/unit_1\/xception_module\/Relu_1 (defined at \\models-master\\research\\deeplab\\core\\xception.py:274)\n\nOriginal stack trace for 'xception_65\/entry_flow\/block1\/unit_1\/xception_module\/separable_conv2_depthwise\/depthwise':\n  File \"\/models-master\/research\/deeplab\/train.py\", line 517, in &lt;module&gt;\n    tf.app.run()\n  ...\n  ...\n  File \"\\anaconda\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 2005, in __init__\n    self._traceback = tf_stack.extract_stack()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57619393,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorboard><attributeerror>","CreationDate":"2019-08-23T03:06:56.937","AcceptedAnswerId":null,"OwnerUserId":7717192.0,"Title":"AttributeError: module 'tensorboard' has no attribute 'lazy'","Body":"<p>I have a generative adversarial network using Keras with a tensorflow backend. I am running Tensorflow 1.14.0, Python 3.7.4, and Keras 2.2.4. The full error is: <\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>Using TensorFlow backend.\nTraceback (most recent call last):\n  File \"C:\/Users\/Riley\/PycharmProjects\/shoeGAN\/shoeWGAN.py\", line 7, in &lt;module&gt;\n    from keras.layers import *\n  File \"C:\\Users\\Riley\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\keras\\__init__.py\", line 3, in &lt;module&gt;\n    from . import utils\n  File \"C:\\Users\\Riley\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\keras\\utils\\__init__.py\", line 6, in &lt;module&gt;\n    from . import conv_utils\n  File \"C:\\Users\\Riley\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\keras\\utils\\conv_utils.py\", line 9, in &lt;module&gt;\n    from .. import backend as K\n  File \"C:\\Users\\Riley\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\keras\\backend\\__init__.py\", line 89, in &lt;module&gt;\n    from .tensorflow_backend import *\n  File \"C:\\Users\\Riley\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py\", line 5, in &lt;module&gt;\n    import tensorflow as tf\n  File \"C:\\Users\\Riley\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\__init__.py\", line 35, in &lt;module&gt;\n    from tensorflow._api.v1 import compat\n  File \"C:\\Users\\Riley\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\_api\\v1\\compat\\__init__.py\", line 22, in &lt;module&gt;\n    from tensorflow._api.v1.compat import v2\n  File \"C:\\Users\\Riley\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\_api\\v1\\compat\\v2\\__init__.py\", line 303, in &lt;module&gt;\n    from tensorboard.summary._tf import summary\n  File \"C:\\Users\\Riley\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorboard\\summary\\__init__.py\", line 32, in &lt;module&gt;\n    from tensorboard.summary import v2\n  File \"C:\\Users\\Riley\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorboard\\summary\\v2.py\", line 24, in &lt;module&gt;\n    from tensorboard.plugins.audio.summary_v2 import audio\n  File \"C:\\Users\\Riley\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorboard\\plugins\\audio\\summary_v2.py\", line 30, in &lt;module&gt;\n    from tensorboard.compat import tf2 as tf\n  File \"C:\\Users\\Riley\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorboard\\compat\\__init__.py\", line 28, in &lt;module&gt;\n    import tensorboard.lazy as _lazy\nAttributeError: module 'tensorboard' has no attribute 'lazy'\n<\/code><\/pre>\n\n<p>I didn't include the full code, because the error comes up during imports, which are below:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>import os\n\nimport matplotlib.pyplot as plt\n\nimport numpy as np\n\nfrom keras.layers import *\nfrom keras.models import *\nfrom keras.optimizers import *\nfrom keras.initializers import *\nfrom keras.callbacks import *\nfrom keras.utils.generic_utils import Progbar\nfrom keras.preprocessing.image import save_img\n<\/code><\/pre>\n\n<p>Things I have tried:<\/p>\n\n<ul>\n<li>Restarting computer.<\/li>\n<li>Uninstalling and reinstalling all tensorflow and tensorboard libraries using pip and then restarting computer.<\/li>\n<li>The top answer <a href=\"https:\/\/stackoverflow.com\/questions\/53966020\/attributeerror-module-tensorboard-util-has-no-attribute-persistentopevaluato\">here<\/a><\/li>\n<\/ul>\n\n<p>Any ideas? Thanks!<\/p>\n","answers":[{"AnswerId":"57619480","CreationDate":"2019-08-23T03:18:45.430","ParentId":null,"OwnerUserId":"5386951","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/tensorflow\/tensorboard\/issues\/1862\" rel=\"nofollow noreferrer\">Here<\/a> is the issue stating your error on Github. <\/p>\n\n<p>At first, someone noted that only the nightly build of TensorBoard (<code>tb-nightly<\/code>) is compatible with the TF 2.0 preview, 1.12.2 is not expected to work.<\/p>\n\n<p>One of the last comments suggests that the order of <code>tensorflow-tensorboard<\/code> and <code>tensorboard<\/code> matter, so you should fist<\/p>\n\n<pre><code>pip install tensorflow-tensorboard==1.5.1\n<\/code><\/pre>\n\n<p>and then <\/p>\n\n<pre><code>pip install tensorboard==1.14.0\n<\/code><\/pre>\n\n<p>Hope this helps!<\/p>\n"}]}
+{"QuestionId":57619730,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2019-08-23T03:57:40.330","AcceptedAnswerId":57619985.0,"OwnerUserId":1264591.0,"Title":"Link weights in Keras layers","Body":"<p>Suppose that I break my input into I two equal sized pieces I1, I2 and I want the following structure on my keras network -- I1->A1, I2->A2, then [A1,A2]->B where B is an output node. I can do this using groups as in <a href=\"https:\/\/stackoverflow.com\/questions\/47265412\/custom-connections-between-layers-keras\">1<\/a>.  However, I want to require that the connection weights (and other activation parameters) for I1->A1 are they same as those for I2->A2, ie I want a symmetry between the 1's and the 2's. (Note that I don't require symmetry for [A1,A2]->B.)<\/p>\n","answers":[{"AnswerId":"57619985","CreationDate":"2019-08-23T04:38:46.897","ParentId":null,"OwnerUserId":"10371987","Title":null,"Body":"<p>If I understand your problem correctly, the mapping of input_1 to A_1 &amp; input_2 to A_2 has been done one after one, since you want the mapping function being the same for both input. In this case, you might consider RNN, but if your inputs are independent to each other, you might consider using <code>TimeDistributed<\/code> <a href=\"https:\/\/keras.io\/layers\/wrappers\/\" rel=\"nofollow noreferrer\">wrapper in Keras<\/a>. The below sample will take two inputs, and use <code>Dense<\/code> layer to map the inputs one by one, thus the weights of <code>Dense<\/code> is shared:  <\/p>\n\n<pre><code>from keras.models import Model\nfrom keras.layers import Input, Dense, TimeDistributed, Concatenate, Lambda\n\nx_dim = 5\nhidden_dim = 8\n\nx1 = Input(shape=(1,x_dim,))\nx2 = Input(shape=(1,x_dim,))\n\nconcat = Concatenate(axis=1)([x1, x2])\nhidden_concat = TimeDistributed(Dense(hidden_dim))(concat)\nhidden1 = Lambda(lambda x: x[:,:1,:])(hidden_concat)\nhidden2 = Lambda(lambda x: x[:,1:,:])(hidden_concat)\n\nmodel = Model(inputs=[x1,x2], outputs=[hidden1, hidden2])\nmodel.summary()\n\n&gt;&gt;&gt;\nLayer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput_33 (InputLayer)           (None, 1, 5)         0                                            \n__________________________________________________________________________________________________\ninput_34 (InputLayer)           (None, 1, 5)         0                                            \n__________________________________________________________________________________________________\nconcatenate_17 (Concatenate)    (None, 2, 5)         0           input_33[0][0]                   \n                                                                 input_34[0][0]                   \n__________________________________________________________________________________________________\ntime_distributed_9 (TimeDistrib (None, 2, 8)         48          concatenate_17[0][0]             \n__________________________________________________________________________________________________\nlambda_8 (Lambda)               (None, 1, 8)         0           time_distributed_9[0][0]         \n__________________________________________________________________________________________________\nlambda_9 (Lambda)               (None, 1, 8)         0           time_distributed_9[0][0]         \n==================================================================================================\nTotal params: 48\nTrainable params: 48\nNon-trainable params: 0\n<\/code><\/pre>\n"}]}
+{"QuestionId":57619816,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras><deep-learning>","CreationDate":"2019-08-23T04:13:41.997","AcceptedAnswerId":null,"OwnerUserId":4083549.0,"Title":"Hashing_trick in Keras. How it works?","Body":"<p>Needs basic understanding of one hot or hashing trick in keras.<\/p>\n\n<pre><code>from keras.preprocessing.text import hashing_trick\nfrom keras.preprocessing.text import text_to_word_sequence\n# define the document\ntext = 'The quick brown fox jumped over the lazy dog dog.'\n# estimate the size of the vocabulary\nwords = set(text_to_word_sequence(text))\nprint(words)\nvocab_size = len(words)\nprint(vocab_size)\n# integer encode the document\nresult = hashing_trick(text, round(vocab_size*1.3), hash_function='md5')\nprint(text)\nprint(result)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<p>{'over', 'the', 'lazy', 'dog', 'quick', 'brown', 'jumped', 'fox'}\n8\nThe quick brown fox jumped over the lazy dog dog.\n[6, 4, 1, 2, 7, 5, 6, 2, 6, 6]<\/p>\n\n<p>Conclusion:\nEach tokens are assigned an integer here. For eg.\n\"quick is assigned 4\"\nthe--6 <br>\nquick--4\nbrown--1\nfox--2\njumped--7\nover--5\nthe--6\nlazy--2\ndog--6<\/p>\n\n<p>I would like to understand how \"the\" &amp; \"dog\" are assigned same integer 6.\nCorrect me if I am wrong &amp; please provide explanation how it does exactly?<\/p>\n","answers":[]}
+{"QuestionId":57620166,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2019-08-23T05:07:26.050","AcceptedAnswerId":57621326.0,"OwnerUserId":3616293.0,"Title":"TensorFlow gradients example","Body":"<p>I have the following code to compute an example of gradients using TensorFlow-<\/p>\n\n<pre><code># z = f(x, y) = 2*x - y\n# Partial derivative of dz\/dx = 2; dz\/dy = -1\n# Initialize x and y variables-                                         \nx = tf.Variable([1], dtype = tf.int32)\ny = tf.Variable([2], dtype = tf.int32)\nz = tf.subtract(2 * x, y)\n\n# Define gradient operation-\ngrad = tf.gradients(z, [x, y])\n\n\n# Initialize TensorFlow session-\nsess = tf.Session()\n\n# Initialize all variables-\nsess.run(tf.global_variables_initializer())\n\n# Compute gradient defined above-\nres_grad = sess.run(grad)\n\n# Close the session-\nsess.close()\n<\/code><\/pre>\n\n<p>But the line-<\/p>\n\n<pre><code>res_grad = sess.run(grad)\n<\/code><\/pre>\n\n<p>Gives the following error-<\/p>\n\n<blockquote>\n  <p>TypeError                                 Traceback (most recent call\n  last)  in \n  ----> 1 res = sess.run(grad)<\/p>\n  \n  <p>~\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py\n  in run(self, fetches, feed_dict, options, run_metadata)\n      948     try:\n      949       result = self._run(None, fetches, feed_dict, options_ptr,\n  --> 950                          run_metadata_ptr)\n      951       if run_metadata:\n      952         proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)<\/p>\n  \n  <p>~\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py\n  in _run(self, handle, fetches, feed_dict, options, run_metadata)<br>\n  1156     # Create a fetch handler to take care of the structure of\n  fetches.    1157     fetch_handler = _FetchHandler(\n  -> 1158         self._graph, fetches, feed_dict_tensor, feed_handles=feed_handles)    1159     1160     # Run request and get\n  response.<\/p>\n  \n  <p>~\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py\n  in <strong>init<\/strong>(self, graph, fetches, feeds, feed_handles)\n      472     \"\"\"\n      473     with graph.as_default():\n  --> 474       self._fetch_mapper = _FetchMapper.for_fetch(fetches)\n      475     self._fetches = []\n      476     self._targets = []<\/p>\n  \n  <p>~\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py\n  in for_fetch(fetch)\n      262     elif isinstance(fetch, (list, tuple)):\n      263       # NOTE(touts): This is also the code path for namedtuples.\n  --> 264       return _ListFetchMapper(fetch)\n      265     elif isinstance(fetch, collections.Mapping):\n      266       return _DictFetchMapper(fetch)<\/p>\n  \n  <p>~\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py\n  in <strong>init<\/strong>(self, fetches)\n      371     \"\"\"\n      372     self._fetch_type = type(fetches)\n  --> 373     self._mappers = [_FetchMapper.for_fetch(fetch) for fetch in fetches]\n      374     self._unique_fetches, self._value_indices = _uniquify_fetches(self._mappers)\n      375 <\/p>\n  \n  <p>~\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py\n  in (.0)\n      371     \"\"\"\n      372     self._fetch_type = type(fetches)\n  --> 373     self._mappers = [_FetchMapper.for_fetch(fetch) for fetch in fetches]\n      374     self._unique_fetches, self._value_indices = _uniquify_fetches(self._mappers)\n      375 <\/p>\n  \n  <p>~\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py\n  in for_fetch(fetch)\n      259     if fetch is None:\n      260       raise TypeError('Fetch argument %r has invalid type %r' % (fetch,\n  --> 261                                                                  type(fetch)))\n      262     elif isinstance(fetch, (list, tuple)):\n      263       # NOTE(touts): This is also the code path for namedtuples.<\/p>\n  \n  <p>TypeError: Fetch argument None has invalid type <\/p>\n<\/blockquote>\n\n<p>What's going wrong?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"57621326","CreationDate":"2019-08-23T07:03:58.313","ParentId":null,"OwnerUserId":"11900574","Title":null,"Body":"<p>You're getting this error because you're passing integer tensors. Use floats and it should work.<\/p>\n\n<pre><code>x = tf.Variable([1.0])\ny = tf.Variable([2.0])\nz = tf.subtract(2 * x, y)\n\n# Define gradient operation-\ngrad = tf.gradients(z, [x, y])\n\n# Initialize TensorFlow session-\nsess = tf.Session()\n\n# Initialize all variables-\nsess.run(tf.global_variables_initializer())\n\n# Compute gradient defined above-\nres_grad = sess.run(grad)\n\nprint (res_grad) # Output = [array([ 2.], dtype=float32), array([-1.], dtype=float32)]\n\n# Close the session-\nsess.close()\n<\/code><\/pre>\n\n<p>This was changed in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/16504\" rel=\"nofollow noreferrer\">#16504<\/a>. Allowing gradients on integer tensors was causing incorrectness in tf.while_loop, and there was no satisfactory way to resolve them without this change.<\/p>\n"}]}
+{"QuestionId":57620508,"AnswerCount":0,"Tags":"<keras><azure-storage-blobs><hdf5>","CreationDate":"2019-08-23T05:50:23.807","AcceptedAnswerId":null,"OwnerUserId":11045009.0,"Title":"How to load a model using Keras API from an HDF5 file stored in azure Blob storage?","Body":"<p>I am having my model.h5 file stored in azure blob storage. I want to understand whether we can load the .h5 file directly from azure blob storage using keras API's load_model() or any other function.<\/p>\n","answers":[]}
+{"QuestionId":57620665,"AnswerCount":1,"Tags":"<python><keras><tf.keras><bytesio>","CreationDate":"2019-08-23T06:07:12.983","AcceptedAnswerId":null,"OwnerUserId":11053801.0,"Title":"Can we save keras model as a binary file in python?","Body":"<p>I have the keras sequential model, normal way of saving is<\/p>\n\n<pre><code>from keras.models import load_model\nkeras_model.save('classifier_keras.h5')\nclassifier_keras = load_model('classifier_keras.h5') \n<\/code><\/pre>\n\n<p>If I need to store in a database as encoded file, how to convert keras model as binary?<\/p>\n","answers":[{"AnswerId":"58464792","CreationDate":"2019-10-19T14:39:41.493","ParentId":null,"OwnerUserId":"6551580","Title":null,"Body":"<pre><code>HDF5 \"model.h5\"\nh5dump -a \/layer_names -a \/dense_1\/weight_names -d \/dense_1\/dense_1\/bias:0 -b LE -o dset.model.bin model.h5\n<\/code><\/pre>\n"}]}
+{"QuestionId":57621136,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-08-23T06:48:33.413","AcceptedAnswerId":null,"OwnerUserId":1722383.0,"Title":"Subsampling a fixed amount of indices given a boolean mask","Body":"<p>I have a batch of a \"boolean\" mask with a varying number of ones and zeros. For each sample within this batch I want a fixed number of indices sampling from the indices containing a \"1\". <\/p>\n\n<p>Let's say I want to sample 6 indices. <\/p>\n\n<pre><code>  mask = np.array([[1, 0, 0, 0, 0, 1, 0, 1, 0, 1, 1],  # 5 ones\n                   [0, 0, 1, 1, 0, 1, 0, 1, 0, 1, 1],  # 6 ones\n                   [1, 1, 0, 1, 0, 1, 0, 1, 0, 1, 1]]) # 7 ones\n  idc = tf.constant(mask)\n  idc = tf.where(idc &gt; 0)\n\n  # ...\n  # subsampled_idc = tf.some_tf_magic(mask, 6)\n  # or\n  # subsampled_idc = tf.some_other_tf_magic(idc, 6)\n  # ...\n\n  with tf.Session() as sess:\n    print(sess.run(subsampled_idc))\n<\/code><\/pre>\n\n<p>Here <code>idc<\/code> from <code>tf.where<\/code> would contain<\/p>\n\n<pre><code>[[ 0  0]\n [ 0  5]\n [ 0  7]\n [ 0  9]\n [ 0 10]\n [ 1  2]\n [ 1  3]\n [ 1  5]\n [ 1  7]\n [ 1  9]\n [ 1 10]\n [ 2  0]\n [ 2  1]\n [ 2  3]\n [ 2  5]\n [ 2  7]\n [ 2  9]\n [ 2 10]]\n<\/code><\/pre>\n\n<p>What I rather need is  <\/p>\n\n<pre><code>[[ 0  5  7  9  10  7]  # randomly using one twice because 5 &lt; 6\n [ 2  3  5  7  9  10]  # using all of them because 6 = 6\n [ 0  1  3  5  9  10]] # randomly leaving out one index because 7 &gt; 6\n<\/code><\/pre>\n\n<p>If the \"row\" of the batch contains less \"1\" than the wanted samples it should subsample then. If the \"row\" of the batch contains more \"1\" it should take all of them and add some duplicates.<\/p>\n\n<p>It does not need to be ordered, actually shuffled indices would be even better.<\/p>\n","answers":[]}
+{"QuestionId":57621175,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-08-23T06:51:31.680","AcceptedAnswerId":null,"OwnerUserId":11889647.0,"Title":"combine prediction output of one frozen model to predict and then use those predictions as a loss for the model which i am training?","Body":"<p>I want to combine the prediction output of one frozen model in a training phase of another.<\/p>\n\n<p>I have tried using different graph sessions but it resets the default graph in the training phase.<\/p>\n\n<pre><code>predictions = model1.model(input1, input2, mode)\npredictions2 = model2.predict(predictions)\n\nloss1 = mean_squared_error(predictions, labels)\nloss2 = mean_squared_error(input2, predictions2)\n\ntotal_loss = loss1+loss2\n\noptimizer.minimize(total_loss)\n<\/code><\/pre>\n\n<blockquote>\n  <p>ValueError: Tensor Tensor(\"output_layer\/BiasAdd:0\", shape=(?, 100), dtype=float32) is not an element of this graph<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"57624097","CreationDate":"2019-08-23T10:09:48.860","ParentId":null,"OwnerUserId":"11889647","Title":null,"Body":"<p>I just figured this out!<\/p>\n\n<p>In the tensorflow's estimator framework, load the model in the 'model_fn' of your Estimator space with the attribute:\n'keras_model.trainable=False'<\/p>\n\n<p>eg snippet:<\/p>\n\n<pre><code>def model_fn(inputs):\n    .......\n    #some operations\n\n    .......\n    model2=load_model('frozen_model.h5')\n    model2.trainable=False\n    model2.summary()\n    predictions= model2(inputs=predictions)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57621554,"AnswerCount":1,"Tags":"<python><machine-learning><keras><keras-layer>","CreationDate":"2019-08-23T07:22:20.087","AcceptedAnswerId":null,"OwnerUserId":9879809.0,"Title":"How to shape images for labels","Body":"<p>I am trying to train satellite images. My training set has 2 folders of input (samples) and ground truth (labels). <\/p>\n\n<p>Total images are: <code>496<\/code><br>\nMy X_train shape is: <code>(496,3,256,256)<\/code><br>\nand y_train shape is: <code>(496,256,256)<\/code> <\/p>\n\n<p>For every model, I use I get an error regarding the dense layer.<\/p>\n\n<pre><code>ValueError: Expected dense to have 2 dimensions, but got an array with shape (496,256,256)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57623367","CreationDate":"2019-08-23T09:25:56.973","ParentId":null,"OwnerUserId":"9025861","Title":null,"Body":"<p>You are feeding images to the Dense Layer. Either flatten the images using .flatten() or use a model with CNN Layers. The shape (496,3,256,256) means you have 496 images of size 256x256 having 3 dimensions(rgb).<\/p>\n\n<p>Look up on this <a href=\"https:\/\/stackoverflow.com\/a\/57574251\/9025861\">answer<\/a> for CNN model.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/q\/57573624\/9025861\">CNN: Error when checking input: expected dense to have 2 dimensions, but got array with shape (391, 605, 700, 3)<\/a><\/p>\n"}]}
+{"QuestionId":57622033,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><deep-learning><computer-vision><artificial-intelligence>","CreationDate":"2019-08-23T07:59:15.513","AcceptedAnswerId":null,"OwnerUserId":4663377.0,"Title":"Configure Fast-Rcnn.config to use Adam optimizer and other parameters","Body":"<p>I have following fast_rcnn_resnet101_coco.config (<a href=\"https:\/\/drive.google.com\/file\/d\/1SLyBZF-0lojmLPB_rPaIuS9eiZYhdeUi\/view?usp=sharing\" rel=\"noreferrer\">here<\/a>). In this config file I have replaced momentum_optimizer with adam optimizer as follows:<\/p>\n\n<pre><code>train_config: {\n  batch_size: 1\n  optimizer {\n    #momentum_optimizer: {\n    adam_optimizer: {\n      learning_rate: {\n        manual_step_learning_rate {\n          initial_learning_rate: 0.00001\n          schedule {\n            step: 4500\n            learning_rate: .00001\n          }\n          schedule {\n            step: 10000\n            learning_rate: .000001\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: \"faster_rcnn_resnet101_coco_2018_01_28\/model.ckpt\"\n  from_detection_checkpoint: true\n  data_augmentation_options {\n    random_horizontal_flip {\n    }\n  }\n}\n<\/code><\/pre>\n\n<p>I have referred <a href=\"https:\/\/stackoverflow.com\/questions\/51915803\/tensorflow-object-detection-use-adam-instead-of-rmsprop\/51920195#51920195\">Tensorflow Object Detection: use Adam instead of RMSProp<\/a> to do this change. My aim is to configure my faster rcnnresnet101.config file (<a href=\"https:\/\/drive.google.com\/file\/d\/1SLyBZF-0lojmLPB_rPaIuS9eiZYhdeUi\/view?usp=sharing\" rel=\"noreferrer\">attached here<\/a>) to match that of this file:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/bqZcS.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/bqZcS.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>My aim is that my .config file should have all the parameters mentioned in the .yaml file. So far I have succeed in doing this only for one parameter (\"learning rate\"). How can I integrate rpn_batch size, step size etc. parameters in my config file ?<\/p>\n","answers":[{"AnswerId":"57749874","CreationDate":"2019-09-01T22:21:25.743","ParentId":null,"OwnerUserId":"12007504","Title":null,"Body":"<p>The basic fact you need to understand is the following :<\/p>\n\n<p>The config file is must match the message <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/protos\/pipeline.proto#L14-L22\" rel=\"nofollow noreferrer\">TrainEvalPipelineConfig<\/a>. Now that message consists of multiple components. So, if you want to modify something in a component, you should go to the proto file in which that component message is defined, see the possible parameters therein and then modify the configuration file according to that. That is exactly what you ultimately did in order to change the optimizer. <\/p>\n\n<p>To give you a hint, if you want to change the RPN batch size, you have to modify <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/protos\/faster_rcnn.proto#L56\" rel=\"nofollow noreferrer\">this parameter<\/a>. So, look it up in the proto file and simply add it to your final configuration file. <\/p>\n\n<p>To give you an illustration, if I were to use the original configuration file with one minor change which is a RPN batch size of 128, my configuration file will appear as follows :<\/p>\n\n<pre><code># Faster R-CNN with Resnet-101 (v1), 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  faster_rcnn {\n    num_classes: 90\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    # below i modify the RPN batch size to 128\n    first_stage_minibatch_size: 128 \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  optimizer {\n    momentum_optimizer: {\n      learning_rate: {\n        manual_step_learning_rate {\n          initial_learning_rate: 0.0003\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: \"PATH_TO_BE_CONFIGURED\/model.ckpt\"\n  from_detection_checkpoint: true\n  data_augmentation_options {\n    random_horizontal_flip {\n    }\n  }\n}\n\ntrain_input_reader: {\n  tf_record_input_reader {\n    input_path: \"PATH_TO_BE_CONFIGURED\/mscoco_train.record-?????-of-00100\"\n  }\n  label_map_path: \"PATH_TO_BE_CONFIGURED\/mscoco_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: \"PATH_TO_BE_CONFIGURED\/mscoco_val.record-?????-of-00010\"\n  }\n  label_map_path: \"PATH_TO_BE_CONFIGURED\/mscoco_label_map.pbtxt\"\n  shuffle: false\n  num_readers: 1\n}\n<\/code><\/pre>\n"}]}
+{"QuestionId":57623305,"AnswerCount":1,"Tags":"<keras><recurrent-neural-network>","CreationDate":"2019-08-23T09:21:49.530","AcceptedAnswerId":57658943.0,"OwnerUserId":2924735.0,"Title":"Implementing AntisymmetricRNN in Keras","Body":"<p>I am trying to implement the AntisymmetricRNN described in this paper: <a href=\"https:\/\/arxiv.org\/abs\/1902.09689\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1902.09689<\/a>.<\/p>\n\n<p>Working in Keras, I guess I have to implement my own layer so I have read <a href=\"https:\/\/keras.io\/layers\/writing-your-own-keras-layers\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/writing-your-own-keras-layers\/<\/a>. Instead of starting from a plain layer as explained there, I reckon the best would probably be to extend one of the existing RNN, but <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/layers\/recurrent.py#L237\" rel=\"nofollow noreferrer\">Keras<\/a> has<\/p>\n\n<ul>\n<li>RNN<\/li>\n<li>SimpleRNNCell<\/li>\n<li>SimpleRNN<\/li>\n<\/ul>\n\n<p>The documentation isn't verbose enough for someone my level about what these classes do\/are, and consequently I am having a hard time figuring out what should be my starting point.<\/p>\n\n<p>Any help, both in terms of where to start and what to actually look out for, and all sorts of suggestions are greatly appreciated. Thank you.<\/p>\n","answers":[{"AnswerId":"57658943","CreationDate":"2019-08-26T13:31:50.800","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>In Keras, all recurrent layers are <code>RNN<\/code> layers with a certain <code>Cell<\/code>.   <\/p>\n\n<p>The definition is <code>RNN(cell=someCell)<\/code>   <\/p>\n\n<p>So, the <code>LSTM<\/code> layer follows the same principle, an <code>LSTM(units=...)<\/code> layer is equal to an <code>RNN(cell=LSTMCell(units=...), ...)<\/code> layer.    <\/p>\n\n<p>That said, to implement your recurrent layer (if it doesn't break the recurrent flow step by step or jump steps), you need to implement your own cell. You can study what is happening in the <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/layers\/recurrent.py#L1790\" rel=\"nofollow noreferrer\">LSTMCell<\/a> code, compare it with the papers and adjust the weights and formulas to your need. <\/p>\n\n<p>So you will have your own <code>RNN(cell=yourCell)<\/code>. <\/p>\n"}]}
+{"QuestionId":57623438,"AnswerCount":0,"Tags":"<python><tensorflow><audio><keras><recurrent-neural-network>","CreationDate":"2019-08-23T09:30:40.877","AcceptedAnswerId":null,"OwnerUserId":9815909.0,"Title":"Keras fit_generator - fixed batchsize for variable length of audio file when loaded (RNN)","Body":"<p>The goal is to have a fixed batchsize for my RNN network using fit_generator. I have a variable audio length files to be loaded (spectrogram). I create 2D patches with constant length (10s) and stack them per audio file so the dimension per audio file looks like <strong>(#patches x Time x fft_bins )<\/strong> and zero pad at the end. Now the problem is since <strong>#patches<\/strong> are variable (example 25 sometime 40) how do I get a fixed batch size (let's say in my case 128 #patches) from the fit_generator. Following I am explaining the steps in detail.<\/p>\n\n<ol>\n<li>My fit_gemerator extract one audio file at a time <\/li>\n<li>Input Dimensions: Each audio file has <strong>(Time x fft_bins )<\/strong> and then I convert them by slicing them into patches of (10s) now the final dimension is <strong>(#patches x Time x fft_bins )<\/strong><\/li>\n<li>But I need a fixed batchsize as an input to my RNN of my fit_generator. For example, if it takes 3 audio files to get batchsize of 128 the fit_generator should be alliterated 3 times in order to fulfill this condition. <\/li>\n<\/ol>\n\n<p>How can I achieve this, my sample code is given below<\/p>\n\n<pre><code>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\ndef __data_generation(self, list_IDs_temp):\n    'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels)\n    # Initialization\n    XPatch = np.empty((self.batch_size, *self.dim))\n    YPatch = np.empty((self.batch_size,1000 ,self.n_channels), dtype=int)\n\n    # Generate data\n    # ID contains list of path of audio wav files exaple: ID = ['\/1.wav', '\/2.wav',...]\n    for i, ID in enumerate(list_IDs_temp):\n        XPatch, YPatch = Get_dataset(filePath= ID)\n\n\n    return XPatch, YPatch\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57623938,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-08-23T10:00:15.750","AcceptedAnswerId":null,"OwnerUserId":10341566.0,"Title":"A strange problem with torch.optim: the variable changes, but the difference between the old and the updated variable is shown to be zero","Body":"<p>I stumbled upon a rather weird issue while debugging a larger code in PyTorch. A simple instance to reproduce the issue is here: <a href=\"https:\/\/colab.research.google.com\/drive\/11kBaxMOxN9i0X1vtaX47Yz7rkkHz7Kpu\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/11kBaxMOxN9i0X1vtaX47Yz7rkkHz7Kpu<\/a><\/p>\n\n<p>As you can see, the variable x gets updated in every epoch (see that their norms are different), however, the norm of their difference (rel_diff) turns out to be zero. Any help is deeply appreciated.<\/p>\n","answers":[{"AnswerId":"57638206","CreationDate":"2019-08-24T13:03:45.583","ParentId":null,"OwnerUserId":"10886420","Title":null,"Body":"<p><code>rel_dif<\/code> (<code>change_in_x<\/code>) ranges from ~55 to ~400 and all your results are pretty typical for neural network.<\/p>\n\n<p>BTW. Next time post your code, link to colab could be additional. Please see <a href=\"https:\/\/stackoverflow.com\/help\/minimal-reproducible-example\">Minimal Reproducible Example<\/a>.<\/p>\n"}]}
+{"QuestionId":57624212,"AnswerCount":1,"Tags":"<keras><conv-neural-network>","CreationDate":"2019-08-23T10:16:19.260","AcceptedAnswerId":57624248.0,"OwnerUserId":11734883.0,"Title":"Value Error: expected conv2d_21_input to have shape (224, 224, 3)","Body":"<p>I am working with CNNS and Keras and trying to predict an image.<\/p>\n\n<p>I have created a window sliding to predict part of the images like this:<\/p>\n\n<pre><code>%matplotlib inline\nimport matplotlib.pyplot as plt\n\nimg = Image.open('data\/test_dataset\/06.jpg')\nsize = 200\nstep = 50\n\nfor x in range(0, img.size[0]-size, step):\n    # go through y axis\n    for y in range(0, img.size[1]-size, step):\n        part = img.crop((x, y, x + size, y+ size))\n        data = np.asarray(part.resize((244,244), resample=Image.BICUBIC))\n        data = data.astype(np.float32)\/255.\n        model.predict(data.reshape(1, 244, 244, 3))\n        #plt.imshow(part)\n        #plt.show()\n<\/code><\/pre>\n\n<p>But getting the following error:<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking input: expected conv2d_21_input to\n  have shape (224, 224, 3) but got array with shape (244, 244, 3)<\/p>\n<\/blockquote>\n\n<p>My CNN structure looks like this: \n<a href=\"https:\/\/i.stack.imgur.com\/9ourB.png\" rel=\"nofollow noreferrer\">CNN structure<\/a><\/p>\n\n<p>I tried many things out but I have no idea how to solve this problem. Can someone help me? Many thanks in advance<\/p>\n","answers":[{"AnswerId":"57624248","CreationDate":"2019-08-23T10:19:00.333","ParentId":null,"OwnerUserId":"11900574","Title":null,"Body":"<p>The model you're trying to predict with accepts an input of shape <code>(224, 224, 3)<\/code>, but you're passing it an image of shape <code>(244, 244, 3)<\/code>.<\/p>\n\n<p>Try this:<\/p>\n\n<pre><code>%matplotlib inline\nimport matplotlib.pyplot as plt\n\nimg = Image.open('data\/test_dataset\/06.jpg')\nsize = 200\nstep = 50\n\nfor x in range(0, img.size[0]-size, step):\n    # go through y axis\n    for y in range(0, img.size[1]-size, step):\n        part = img.crop((x, y, x + size, y+ size))\n        data = np.asarray(part.resize((224,224), resample=Image.BICUBIC))\n        data = data.astype(np.float32)\/255.\n        model.predict(data.reshape(1, 224, 224, 3))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57624624,"AnswerCount":0,"Tags":"<python-3.x><numpy><keras><tuples><reshape>","CreationDate":"2019-08-23T10:40:20.000","AcceptedAnswerId":null,"OwnerUserId":5572951.0,"Title":"IndexError: tuple index out of range. Reshaping NumPy Test Data Set for Keres Model","Body":"<p>I'm currently trying to learn about the LSTM Algorithm (Machine Learning) and trying to \"reshape\" my (NumPy) X test data array. However I seem to have a problem regarding the size of the array, I think!? <\/p>\n\n<p>Note: I've looked at the question <a href=\"https:\/\/stackoverflow.com\/questions\/54391166\/reshaping-of-an-nparray-returned-indexerror-tuple-index-out-of-range\">reshaping of an nparray returned &quot;IndexError: tuple index out of range&quot;<\/a> which I do understand, but I can't apply this to my particular situation.<\/p>\n\n<p>Note: I have asked this question already<\/p>\n\n<p>Any help is greatly appreciated<\/p>\n\n<pre><code>X_test = []\nfor i in range(60,inputs.shape[0]):\n   X_test.append(inputs[i-60:i,0])\n\nX_test = np.array(X_test)\nX_test = np.reshape(X_test, (X_test.shape[0],X_test.shape[1],1))\nScores = model.predict(X_test)\nScores = scaler.inverse_transform(Scores)\n\n--------------------------------------------------------------------------   \nIndexError                             Traceback (most recent call   last)\n&lt;ipython-input-27-f9a35712d57d&gt; in &lt;module&gt;\n  1 X_test = np.array(X_test)\n----&gt; 2 X_test = np.reshape(X_test, (X_test.shape[0],X_test.shape[1],1))\n  3 Scores = model.predict(X_test)\n  4 Scores = scaler.inverse_transform(Scores)\n\nIndexError: tuple index out of range\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57624739,"AnswerCount":1,"Tags":"<python><pytorch><distributed><multi-gpu>","CreationDate":"2019-08-23T10:47:24.290","AcceptedAnswerId":57704309.0,"OwnerUserId":9989838.0,"Title":"How to fix 'RuntimeError: Address already in use' in PyTorch?","Body":"<p>I am trying to run a distributive application with PyTorch distributive trainer. I thought I would first try the example they have, found <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/aws_distributed_training_tutorial.html#pytorch-1-0-distributed-trainer-with-amazon-aws\" rel=\"nofollow noreferrer\">here<\/a>. I set up two AWS EC2 instances and configured them according to the description in the link, but when I try to run the code I get two different errors: in the first terminal window for node0 I get the error message: RuntimeError: Address already in use<\/p>\n\n<p>Under the other three windows I get the same error message: <\/p>\n\n<blockquote>\n  <p>RuntimeError: NCCL error in:\n  \/pytorch\/torch\/lib\/c10d\/ProcessGroupNCCL.cpp:272, unhandled system\n  error<\/p>\n<\/blockquote>\n\n<p>I followed the code in the link, and terminated the instances an redid but it didn't help\/<\/p>\n\n<p>This is using python 3.6 with the nightly build Cuda 9.0. I tried changing the MASTER_ADDR to the ip for node0 on both nodes, as well as using the same MASTER_PORT (which is an available, unused port). However I still get the same error message. <\/p>\n\n<p>After running this, my goal is to the adjust <a href=\"https:\/\/github.com\/rosinality\/style-based-gan-pytorch\" rel=\"nofollow noreferrer\">this<\/a> StyleGan implementation so that I can train it across multiple GPUs in two different nodes.<\/p>\n","answers":[{"AnswerId":"57704309","CreationDate":"2019-08-29T06:34:55.500","ParentId":null,"OwnerUserId":"9989838","Title":null,"Body":"<p>So after a lot of failed attempts I found out what the problem is. Note that this solution applies to using ASW deep learning instances. <\/p>\n\n<p>After creating two instances I had to adjust the security group. Add two rules: The first rule should be ALL_TCP, and set the source to the Private IPs of the leader. The second rule should be the same (ALL_TCP), but with the source as the Private IPs of the slave node. <\/p>\n\n<p>Previously, I had the setting security rule set as: Type SSH, which only had a single available port (22). For some reason I was not able to use this port to allow the nodes to communicate. After changing these settings the code worked fine. I was also able to run <a href=\"https:\/\/aws.amazon.com\/blogs\/machine-learning\/scalable-multi-node-training-with-tensorflow\/\" rel=\"nofollow noreferrer\">this<\/a> with the above mentioned settings. <\/p>\n"}]}
+{"QuestionId":57624842,"AnswerCount":1,"Tags":"<matplotlib><deep-learning><pytorch><tensor>","CreationDate":"2019-08-23T10:52:50.753","AcceptedAnswerId":null,"OwnerUserId":11429962.0,"Title":"How to compare two different tensors through matplotlib or any other visualize tools?","Body":"<p>I have two tensors x and y, there both are the same dimension shape = (1, 64, 1, 1)<\/p>\n\n<p>basically y is output from many functions, and x is the input<\/p>\n\n<p>I want to compare these two tensor using visualize tool like matplotlib..etc<\/p>\n\n<p>anyway to do ?\nThanks!<\/p>\n\n<p>The below are x and y example, I only post 10 of 64 since the restriction<\/p>\n\n<p>x<\/p>\n\n<pre><code>tensor([[[[-0.8467]],\n\n         [[-0.0949]],\n\n         [[-0.8253]],\n\n         [[-0.1027]],\n\n         [[ 0.0476]],\n\n         [[-0.4173]],\n\n         [[-0.0870]],\n\n         [[ 0.0650]],\n\n         [[ 0.3816]],\n\n         [[ 0.2046]]]], grad_fn=&lt;MulBackward0&gt;)\n<\/code><\/pre>\n\n<p>y<\/p>\n\n<pre><code>tensor([[[[-2.0307]],\n\n         [[-0.1594]],\n\n         [[-1.5174]],\n\n         [[-0.2767]],\n\n         [[ 0.1049]],\n\n         [[-0.9605]],\n\n         [[-0.2127]],\n\n         [[ 0.1342]],\n\n         [[ 0.8275]],\n\n         [[ 2.0508]],\n]])\n<\/code><\/pre>\n\n<p>Thanks in advanced !<\/p>\n","answers":[{"AnswerId":"57625195","CreationDate":"2019-08-23T11:12:58.753","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can convert <code>x<\/code> and <code>y<\/code> to <a href=\"\/questions\/tagged\/numpy\" class=\"post-tag\" title=\"show questions tagged &#39;numpy&#39;\" rel=\"tag\">numpy<\/a> arrays and then use whatever matplotlib function you want<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with torch.no_grad():\n  x_np = x.cpu().numpy()[0, :, 0, 0]  # make it 1d\n  y_np = y.cpu().numpy()[0, :, 0, 0]\n  plt.plot(x_np - y_np)\n  plt.show()\n<\/code><\/pre>\n"}]}
+{"QuestionId":57625101,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-08-23T11:06:37.660","AcceptedAnswerId":null,"OwnerUserId":11840960.0,"Title":"How to generate .ckpt files from checkpointdata files","Body":"<p>I was using question generation from <\/p>\n\n<p><a href=\"https:\/\/github.com\/bloomsburyai\/question-generation\" rel=\"nofollow noreferrer\">https:\/\/github.com\/bloomsburyai\/question-generation<\/a><\/p>\n\n<p>There are the below listed files for usage of model :<\/p>\n\n<ol>\n<li>model.checkpoint-27000.data-00000-of-00001.<\/li>\n<li>model.checkpoint-27000.index.  <\/li>\n<li>model.checkpoint-27000.meta.<\/li>\n<\/ol>\n\n<p>I need a .ckpt file for some other process and wondering if I could replace the <code>model.checkpoint-27000.data-00000-of-00001.ckpt<\/code> and use it or is there any way I could generate that from the 3 files above,<\/p>\n\n<p>I suppose the model is being generated from <\/p>\n\n<pre><code>saver.save(sess, chkpt_path+'\/model.checkpoint', global_step=i)\n<\/code><\/pre>\n\n<p>How can I resolve this?<\/p>\n","answers":[]}
+{"QuestionId":57625486,"AnswerCount":0,"Tags":"<tensorflow><seq2seq>","CreationDate":"2019-08-23T11:31:23.000","AcceptedAnswerId":null,"OwnerUserId":8903228.0,"Title":"Does 'embedding_attention_seq2seq' give a deterministic result or sampled result?","Body":"<p>I train a chatbot using movie dialogue corpus using <code>tf.contrib.legacy_seq2seq.embedding_attention_seq2seq<\/code>, but it output different question when ask same question. <\/p>\n\n<p>I find similar project DeepQA said \"For now, the predictions are deterministic (the network just take the most likely output) so when answering a question, the network will always gives the same answer.\" so i confused why my model can output different results or where a softmax sample ahas been added? <\/p>\n\n<p><strong>model:<\/strong><\/p>\n\n<pre><code>        def create_cell():\n            cell = tf.contrib.rnn.BasicLSTMCell(self.hidden_size)\n            cell = tf.contrib.rnn.DropoutWrapper(\n                        cell,\n                        input_keep_prob=1.0,\n                        output_keep_prob=self.dropout)\n            return cell\n\n        self.cell = tf.contrib.rnn.MultiRNNCell([create_cell() for _ in range(self.rnn_layers)])\n\n        # define placeholder\n        with tf.name_scope(\"encoder_placeholder\"):\n            self.encoder_inputs = [tf.placeholder(tf.int32, [None, ])\n                                    for _ in range(self.maxLengthEnco)]\n        with tf.name_scope(\"decoder_placeholder\"):\n            self.decoder_inputs  = [tf.placeholder(tf.int32,   [None, ], name='decoder_inputs')\n                                    for _ in range(self.maxLengthDeco)]\n            self.decoder_targets  = [tf.placeholder(tf.int32,   [None, ], name='decoder_targets')\n                                    for _ in range(self.maxLengthDeco)]\n            self.decoder_weights  = [tf.placeholder(tf.float32,   [None, ], name='decoder_weights')\n                                    for _ in range(self.maxLengthDeco)]\n\n\n\n        decoder_output, state = tf.contrib.legacy_seq2seq.embedding_attention_seq2seq(self.encoder_inputs,\n                           self.decoder_inputs,\n                           self.cell,\n                           self.num_encoder_symbols,\n                           self.num_decoder_symbols,\n                           self.embedding_size,\n                           output_projection=None,\n                           feed_previous=bool(self.test),\n                           dtype=None,\n                           scope=None)\n\n        # For testing only\n\n        if not outputProjection:\n            self.outputs_test = decoder_output\n<\/code><\/pre>\n\n<p><strong>prediction function<\/strong>:<\/p>\n\n<pre><code>def singlePredict(question, questionSeq=None):\n    batch = sentence2enco(question)\n    if not batch:\n        return None\n    if questionSeq is not None:  # If the caller want to have the real input\n        questionSeq.extend(batch.encoderSeqs)\n    # Run the model\n    ops,feed_dict = seq2seq_model.step(batch)\n    #print(feed_dict)\n    output = sess.run(seq2seq_model.outputs_test, feed_dict) \n    answer = deco2sentence(output)\n    return answer\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57625960,"AnswerCount":0,"Tags":"<python><tensorflow><pip><python-embedding><termcolor>","CreationDate":"2019-08-23T12:01:37.550","AcceptedAnswerId":null,"OwnerUserId":6655465.0,"Title":"When using python-embedded version and doing pip install * it runs into a ModuleNotFoundError","Body":"<p>I am reflecting some python packages into .Net using <a href=\"https:\/\/github.com\/henon\/Python.Included\" rel=\"nofollow noreferrer\">https:\/\/github.com\/henon\/Python.Included<\/a>, which relies on the python-embedded installation.<\/p>\n\n<p>I enabled pip, which is not officially supported for python-embedded by following <a href=\"https:\/\/github.com\/pypa\/pip\/issues\/4207#issuecomment-281236055\" rel=\"nofollow noreferrer\">this<\/a> method.<\/p>\n\n<p>Now let's say my python executable is at this location <code>c:\\users\\a\\appdata\\local\\python-3.7.3-embed-amd64\\python.exe<\/code>.<\/p>\n\n<p>I am trying to install tensorflow using the pip module of that python installation with:<\/p>\n\n<pre><code>c:\\users\\a\\appdata\\local\\python-3.7.3-embed-amd64\\python.exe -m pip install tensorflow\n<\/code><\/pre>\n\n<p>When I try to do this, it throws the following error while trying to install the <code>termocolor<\/code> tensorflow dependency:<\/p>\n\n<pre><code>Collecting termcolor&gt;=1.1.0 (from tensorflow)\n  Using cached https:\/\/files.pythonhosted.org\/packages\/8a\/48\/a76be51647d0eb9f10e2a4511bf3ffb8cc1e6b14e9e4fab46173aa79f981\/termcolor-1.1.0.tar.gz\n    ERROR: Command errored out with exit status 1:\n     command: 'c:\\users\\a\\appdata\\local\\python-3.7.3-embed-amd64\\python.exe' -c 'import sys, setuptools, tokenize; sys.argv[0] = '\"'\"'C:\\\\Users\\\\a\\\\AppData\\\\Local\\\\Temp\\\\pip-install-b_b76a7_\\\\termcolor\\\\setup.py'\"'\"'; __file__='\"'\"'C:\\\\Users\\\\a\\\\AppData\\\\Local\\\\Temp\\\\pip-install-b_b76a7_\\\\termcolor\\\\setup.py'\"'\"';f=getattr(tokenize, '\"'\"'open'\"'\"', open)(__file__);code=f.read().replace('\"'\"'\\r\\n'\"'\"', '\"'\"'\\n'\"'\"');f.close();exec(compile(code, __file__, '\"'\"'exec'\"'\"'))' egg_info --egg-base pip-egg-info\n         cwd: C:\\Users\\a\\AppData\\Local\\Temp\\pip-install-b_b76a7_\\termcolor\\\n    Complete output (5 lines):\n    Traceback (most recent call last):\n      File \"&lt;string&gt;\", line 1, in &lt;module&gt;\n      File \"C:\\Users\\a\\AppData\\Local\\Temp\\pip-install-b_b76a7_\\termcolor\\setup.py\", line 35, in &lt;module&gt;\n        from termcolor import VERSION\n    ModuleNotFoundError: No module named 'termcolor'\n    ----------------------------------------\nERROR: Command errored out with exit status 1: python setup.py egg_info Check the logs for full command output.\n<\/code><\/pre>\n\n<p>I tried to use my local python installation I previously had via miniconda and ran the command <code>pip install tensorflow<\/code> on the base environment and it went ok.<\/p>\n\n<p>I don't understand what is happening, does anybody have a clue?<\/p>\n","answers":[]}
+{"QuestionId":57626163,"AnswerCount":0,"Tags":"<python><tensorflow><keras><tensorflow2.0>","CreationDate":"2019-08-23T12:15:14.897","AcceptedAnswerId":null,"OwnerUserId":11967077.0,"Title":"How convert a TensorFlow model 2.0 to tslite model with 2.0 API?","Body":"<p>How convert a TensorFlow model 2.0 to tslite model with 2.0 API?<\/p>\n\n<p>I tried to export my custom <code>TensorFlow<\/code> model to <code>tflite<\/code> format, because I want to integrate this module to an Android application.  I get strange errors after compiling the Python script. <\/p>\n\n<blockquote>\n  <p>Used: Tensorflow 2.0.0 beta1 API\n      <a href=\"https:\/\/www.tensorflow.org\/lite\/r2\/convert\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/lite\/r2\/convert<\/a><\/p>\n<\/blockquote>\n\n<p>I tried these methods to convert:<\/p>\n\n<pre><code> - From SavedModelFrom \n - tf.Keras Model\n<\/code><\/pre>\n\n<p>These points will show in code below.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>CipherNeuralModel = keras.Sequential([\n    keras.layers.Flatten(input_shape=[20, 20]),\n    keras.layers.Dense(400, activation='relu'),\n    keras.layers.Dense(26, activation='softmax')\n])\n\nCipherNeuralModel.compile(optimizer='adam',\n                          loss='sparse_categorical_crossentropy',\n                          metrics=['accuracy'])\n\nCipherNeuralModel.fit(trainSumData, imgNumberFromLit, epochs=10, steps_per_epoch=20)\n\nsavePath = \"D:\\FolderToModel\"\ntf.saved_model.save(CipherNeuralModel, savePath)\nexport_model = tf.saved_model.load(savePath)\nconcrete_func = export_model.signatures[\n    tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY\n]\nconverter = tf.lite.TFLiteConverter.from_concrete_functions([concrete_func])\n# Error here\nconverter = converter.convert()\n\n\n# Example for convert tf.keras model to tflite\n# convert_model = tf.function(lambda x: CipherNeuralModel(x))\n# concrete_func = convert_model.get_concrete_function(\n#     tf.TensorSpec(CipherNeuralModel.inputs[0].shape,\n#                   CipherNeuralModel.inputs[0].dtype))\n# convertor = tf.lite.TFLiteConverter.from_concrete_functions([concrete_func])\n# tflite_model = convertor.convert()\n\n<\/code><\/pre>\n\n<blockquote>\n  <p>TOCO failed. See console for info.\\n%s\\n%s\\n\" % (stdout, stderr))\n  tensorflow.lite.python.convert.ConverterError: TOCO failed. See\n  console for info. b'\"toco_from_protos\" \\xad\\xa5\n  \\xef\\xa2\\xab\\xef\\xa5\\xe2\\xe1\\xef\n  \\xa2\\xad\\xe3\\xe2\\xe0\\xa5\\xad\\xad\\xa5\\xa9 \\xa8\\xab\\xa8<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":57626393,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><tesseract><python-tesseract>","CreationDate":"2019-08-23T12:31:30.047","AcceptedAnswerId":null,"OwnerUserId":5291590.0,"Title":"How to train tesseract to reduce the number of empty outputs","Body":"<p>Version Info:<\/p>\n\n<pre><code>tesseract 4.00.00alpha\nleptonica-1.74.4\n<\/code><\/pre>\n\n<p>I am fine tuning <a href=\"https:\/\/github.com\/tesseract-ocr\/tessdata_best\/blob\/master\/eng.traineddata\" rel=\"nofollow noreferrer\">eng.trainedata<\/a> on a custom data-set using the following Makefile. I use latin.unicharset as a unicharset for training ( I am training my model to generalize on set of characters that it has not yet seen ).<\/p>\n\n<pre><code>export\n\nSHELL := \/bin\/bash\nHOME := $(PWD)\nTESSDATA = $(HOME)\/tessdata\nLANGDATA = $(HOME)\/langdata\n\n# Train directory\nTRAIN := $(HOME)\/train_data\n\n# Name of the model to be built\nMODEL_NAME = eng\n\n# Name of the model to continue from\nCONTINUE_FROM = eng\n\nTESSDATA_REPO = _best\n\n# Normalization Mode - see src\/training\/language_specific.sh for details \nNORM_MODE = 2\n\n\n# BEGIN-EVAL makefile-parser --make-help Makefile\n\nhelp:\n    @echo \"\"\n    @echo \"  Targets\"\n    @echo \"\"\n    @echo \"    unicharset       Create unicharset\"\n    @echo \"    lists            Create lists of lstmf filenames for training and eval\"\n    @echo \"    training         Start training\"\n    @echo \"    proto-model      Build the proto model\"\n    @echo \"    leptonica        Build leptonica\"\n    @echo \"    tesseract        Build tesseract\"\n    @echo \"    tesseract-langs  Download tesseract-langs\"\n    @echo \"    langdata         Download langdata\"\n    @echo \"    clean            Clean all generated files\"\n    @echo \"\"\n    @echo \"  Variables\"\n    @echo \"\"\n    @echo \"    MODEL_NAME         Name of the model to be built\"\n    @echo \"    CORES              No of cores to use for compiling leptonica\/tesseract\"\n    @echo \"    LEPTONICA_VERSION  Leptonica version. Default: $(LEPTONICA_VERSION)\"\n    @echo \"    TESSERACT_VERSION  Tesseract commit. Default: $(TESSERACT_VERSION)\"\n    @echo \"    LANGDATA_VERSION   Tesseract langdata version. Default: $(LANGDATA_VERSION)\"\n    @echo \"    TESSDATA_REPO      Tesseract model repo to use. Default: $(TESSDATA_REPO)\"\n    @echo \"    TRAIN              Train directory\"\n    @echo \"    RATIO_TRAIN        Ratio of train \/ eval training data\"\n\n# END-EVAL\n\n# Ratio of train \/ eval training data\nRATIO_TRAIN := 0.90\n\nALL_BOXES = data\/all-boxes\nALL_LSTMF = data\/all-lstmf\n\n# Create unicharset\nunicharset: data\/unicharset\n\n# Create lists of lstmf filenames for training and eval\nlists: $(ALL_LSTMF) data\/list.train data\/list.eval\n\ndata\/list.train: $(ALL_LSTMF)\n    total=`cat $(ALL_LSTMF) | wc -l` \\\n       no=`echo \"$$total * $(RATIO_TRAIN) \/ 1\" | bc`; \\\n       head -n \"$$no\" $(ALL_LSTMF) &gt; \"$@\"\n\ndata\/list.eval: $(ALL_LSTMF)\n    total=`cat $(ALL_LSTMF) | wc -l` \\\n       no=`echo \"($$total - $$total * $(RATIO_TRAIN)) \/ 1\" | bc`; \\\n       tail -n \"+$$no\" $(ALL_LSTMF) &gt; \"$@\"\n\n# Start training\ntraining: data\/$(MODEL_NAME).traineddata\n\ndata\/unicharset: $(ALL_BOXES)\n    combine_tessdata -u $(TESSDATA)\/$(CONTINUE_FROM).traineddata  $(TESSDATA)\/$(CONTINUE_FROM).\n    unicharset_extractor --output_unicharset \"$(TRAIN)\/my.unicharset\" --norm_mode $(NORM_MODE) \"$(ALL_BOXES)\"\n#   merge_unicharsets data\/$(START_MODEL)\/$(START_MODEL).lstm-unicharset $(GROUND_TRUTH_DIR)\/my.unicharset  \"$@\"\n\n$(ALL_BOXES): $(sort $(patsubst %.tif,%.box,$(wildcard $(TRAIN)\/*.tif)))\n    find $(TRAIN) -name '*.box' -exec cat {} \\; &gt; \"$@\"\n\n$(TRAIN)\/%.box: $(TRAIN)\/%.tif $(TRAIN)\/%.gt.txt\n    python generate_line_box.py -i \"$(TRAIN)\/$*.tif\" -t \"$(TRAIN)\/$*.gt.txt\" &gt; \"$@\"\n\n$(ALL_LSTMF): $(sort $(patsubst %.tif,%.lstmf,$(wildcard $(TRAIN)\/*.tif)))\n    find $(TRAIN) -name '*.lstmf' -exec echo {} \\; | sort -R -o \"$@\"\n\n$(TRAIN)\/%.lstmf: $(TRAIN)\/%.box\n    tesseract $(TRAIN)\/$*.tif $(TRAIN)\/$* --psm 7 lstm.train\n\n\n# Build the proto model\nproto-model: data\/$(MODEL_NAME)\/$(MODEL_NAME).traineddata\n\ndata\/$(MODEL_NAME)\/$(MODEL_NAME).traineddata: $(LANGDATA) data\/unicharset\n    combine_lang_model \\\n      --input_unicharset data\/unicharset \\\n      --script_dir $(LANGDATA) \\\n      --words $(LANGDATA)\/$(MODEL_NAME)\/$(MODEL_NAME).wordlist \\\n      --numbers $(LANGDATA)\/$(MODEL_NAME)\/$(MODEL_NAME).numbers \\\n      --puncs $(LANGDATA)\/$(MODEL_NAME)\/$(MODEL_NAME).punc \\\n      --output_dir data\/ \\\n      --lang $(MODEL_NAME)\n\ndata\/checkpoints\/$(MODEL_NAME)_checkpoint: unicharset lists proto-model\n    mkdir -p data\/checkpoints\n    lstmtraining \\\n      --continue_from   $(TESSDATA)\/$(CONTINUE_FROM).lstm \\\n      --old_traineddata $(TESSDATA)\/$(CONTINUE_FROM).traineddata \\\n      --traineddata data\/$(MODEL_NAME)\/$(MODEL_NAME).traineddata \\\n      --model_output data\/checkpoints\/$(MODEL_NAME) \\\n      --debug_interval -1 \\\n      --train_listfile data\/list.train \\\n      --eval_listfile data\/list.eval \\\n      --sequential_training \\\n      --max_iterations 88186\n\ndata\/$(MODEL_NAME).traineddata: data\/checkpoints\/$(MODEL_NAME)_checkpoint\n    lstmtraining \\\n    --stop_training \\\n    --continue_from $^ \\\n    --old_traineddata $(TESSDATA)\/$(CONTINUE_FROM).traineddata \\\n    --traineddata data\/$(MODEL_NAME)\/$(MODEL_NAME).traineddata \\\n    --model_output $@\n\n# Clean all generated files\nclean:\n    find data\/train -name '*.box' -delete\n    find data\/train -name '*.lstmf' -delete\n    rm -rf data\/all-*\n    rm -rf data\/list.*\n    rm -rf data\/$(MODEL_NAME)\n    rm -rf data\/unicharset\n    rm -rf data\/checkpoints\n<\/code><\/pre>\n\n<p>Despite training tesseract for a considerable number of epochs it gives empty output for a significant number of images in the training data ( these images have > 100 pixels image_width and image_height ).<\/p>\n\n<p>So I have 2 questions:<\/p>\n\n<pre><code>1. Is this a known issue with tesseract version 4.0.0alpha and fixed it in the future releases.\n2. If its not a bug. How should I proceed with training to make sure that the number of images for which tesseract gives empty output is minimized.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57626742,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><confusion-matrix><tensorflow-hub>","CreationDate":"2019-08-23T12:53:01.117","AcceptedAnswerId":null,"OwnerUserId":2220743.0,"Title":"How to get confusion matrix of Tensorflow hub model","Body":"<p>I have trained a model based on transfer learning via <a href=\"https:\/\/www.tensorflow.org\/hub\" rel=\"nofollow noreferrer\">Tensorflow Hub<\/a>. I have been looking at many places for hints on producing Confusion matrix but I haven't been able to get to the right solution. <\/p>\n\n<p>Does anyone know if this is possible? <\/p>\n\n<p>The last thing I tried was to write the results in an Excel sheet but I couldn't find a formula for the multi-class computation of confusion matrix in Excel.<\/p>\n\n<p>Any help will be great!!<\/p>\n","answers":[{"AnswerId":"57641019","CreationDate":"2019-08-24T19:38:05.657","ParentId":null,"OwnerUserId":"11127923","Title":null,"Body":"<p>You can try <strong>tf.math.confusion_matrix<\/strong> function.<br>\nIt computes the confusion matrix from predictions and labels.<br>\nSee <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/math\/confusion_matrix\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/math\/confusion_matrix<\/a>.  <\/p>\n\n<p><strong>Example:<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>my_confusion_matrix = tf.math.confusion_matrix(labels=[1, 2, 4], predictions=[2, 2, 3])\nwith tf.Session() as sess:\n  print(sess.run(my_confusion_matrix))\n\n# Prints #\n[[0 0 0 0 0]\n [0 0 1 0 0]\n [0 0 1 0 0]\n [0 0 0 0 0]\n [0 0 0 1 0]]\n\n# Labels are assumed to be [0, 1, 2, 3, 4] thus resulting in 5X5 confusion matrix\n<\/code><\/pre>\n"}]}
+{"QuestionId":57626828,"AnswerCount":1,"Tags":"<python><tensorflow><time-series><forecasting><tensorflow-probability>","CreationDate":"2019-08-23T12:58:00.683","AcceptedAnswerId":null,"OwnerUserId":6824320.0,"Title":"How to replace sts.LinearRegression with non-linear model for a Tensorflow Probability Structured Time Series model component","Body":"<p>I'm creating a time-series forecasting model with external, controllable features similar to the \"Forecasting Demand for Electricity\" example found at <a href=\"https:\/\/medium.com\/tensorflow\/structural-time-series-modeling-in-tensorflow-probability-344edac24083\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/tensorflow\/structural-time-series-modeling-in-tensorflow-probability-344edac24083<\/a>. In order to model the influence of the external factors, I am using an sts.LinearRegression() as a component of my model, but those external factors are very non-linear in nature and it's causing unwanted negative predictions in my model. <\/p>\n\n<p>I've tried creating (simpler) forecasting outside of TFP STS, and found that a RandomForestRegressor works much better a LinearRegressor for these external features. What I'd LIKE to do is replace the sts.LinearRegression() with an sts.RandomForestRegressor(), but that isn't available from the sts library. In fact, there are hardly any options available from the sts library: <a href=\"https:\/\/www.tensorflow.org\/probability\/api_docs\/python\/tfp\/sts\/LinearRegression\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/probability\/api_docs\/python\/tfp\/sts\/LinearRegression<\/a><\/p>\n\n<p>I've also tried converting my target variable to log form, but there are numerous instances of zeros (which are inf for log), and this doesn't turn out to be a useful transformation. <\/p>\n\n<p>My model architecture for TFP STS looks something like this: <\/p>\n\n<p>def build_model(observed_time_series):<\/p>\n\n<pre><code>season_effect = sts.Seasonal(\n                num_seasons = 4, num_steps_per_season = 13, observed_time_series = observed_time_series,\n                name = 'season_effect')\nmarketing_effect = sts.LinearRegression(\n                design_matrix = tf.stack([recent_publicity - np.mean(recent_publicity),\n                active_ad - np.mean(active_ad)], axis = -1),\n                name = 'marketing_effect')\nautoregressive = sts.Autoregressive(order=1,\n                observed_time_series = observed_time_series,\n                name = 'autoregressive')\nmodel = sts.Sum([season_effect,\n                 marketing_effect,\n                 autoregressive],\n                 observed_time_series = observed_time_series)\nreturn model\n<\/code><\/pre>\n\n<p>Where I want to change the \"marketing_effect\" component of the model to something non-linear. <\/p>\n\n<p>Is my only option here to clone the TFP STS library and create a custom function to handle non-linear data with something like a Random Forest Regressor? Does anyone know of a better option?<\/p>\n","answers":[{"AnswerId":"57661713","CreationDate":"2019-08-26T16:31:59.233","ParentId":null,"OwnerUserId":"9946929","Title":null,"Body":"<p>I'm not familiar with the usage of random forests in sts models. Can you point to a system where this exists? The trick with tfp.sts is that the math all works out nice &amp; analytically because everything is marginally gaussian. If we can make that work, I think we're definitely open to bringing in other models.<\/p>\n"}]}
+{"QuestionId":57627059,"AnswerCount":1,"Tags":"<python><tensorflow><tf.keras>","CreationDate":"2019-08-23T13:11:12.153","AcceptedAnswerId":null,"OwnerUserId":5623899.0,"Title":"TF Keras v 1.14+: subclass model or subclass layer for \"module\"","Body":"<p>Tensorflow has some <a href=\"https:\/\/www.tensorflow.org\/guide\/keras#custom_layers\" rel=\"nofollow noreferrer\">docs<\/a> for subclassing (tf) Keras <code>Model<\/code> and <code>Layer<\/code>.<\/p>\n\n<p>However, it is unclear which to use for \"modules\" or \"blocks\" (e.g. several layers collectively).<\/p>\n\n<p>Since it is technically <em>several<\/em> layers, I feel that subclassing <code>Layer<\/code> would be deceiving, and while subclassing <code>Model<\/code> works, I am unsure if there are any negative penalties for doing so.<\/p>\n\n<p>e.g. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = inputs\na = self.dense_1(x) # &lt;--- self.dense_1 = tf.keras.Dense(...)\nb = self.dense_2(a)\nc = self.add([x, b])\n<\/code><\/pre>\n\n<p>which is appropriate to use?<\/p>\n","answers":[{"AnswerId":"57627816","CreationDate":"2019-08-23T13:54:51.700","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Initially, there is no need to sublass anything with Keras. Unless you have a particular reason for that (which is not building, training, predicting), you don't subclass for Keras.<\/p>\n\n<h3>Buiding a Keras model:<\/h3>\n\n<p>Either using <code>Sequential<\/code> (the model is ready already, just add layers), or using <code>Model<\/code> (create a graph with layers and finally call <code>Model(inputs, outputs)<\/code>), you don't need to subclass. <\/p>\n\n<p>The moment you create an instance of <code>Sequential<\/code> or <code>Model<\/code>, you have a fully defined model, ready to use in all situations. <\/p>\n\n<p>This model can even be used as parts of other models, its layers can be easily accessed to get intermetiate outputs and create new branches in your graph. <\/p>\n\n<p>So, I don't see any reason at all to subclass <code>Model<\/code>, unless you are using some additional framework that would require this (but I don't think so). This seems to be something from PyTorch users (because this kind of model building is typical for PyTorch, create a subclass for <code>Module<\/code> and add layers and a call method). But Pytorch doesn't offer the same ease as Keras does for getting intermediate results. <\/p>\n\n<p>The main advantage of using Keras is exactly this: you can easily access layers from blocks and models and instantly start branching from that point without needing to rebuild any call methods or adding any extra code for that in the models. So, when you subclass Model, you just defeat the purpose of Keras making it all more difficult. <\/p>\n\n<p>The docs you mentioned say:<\/p>\n\n<blockquote>\n  <p>Model subclassing is particularly useful when <strong>eager execution<\/strong> is enabled since the forward pass can be written imperatively.<\/p>\n<\/blockquote>\n\n<p>I don't really understand what \"imperatively\" means, and I don't see how it would be easier than just building a model with regular layers. <\/p>\n\n<p>Another quote from the docs: <\/p>\n\n<blockquote>\n  <p>Key Point: Use the right API for the job. While model subclassing offers flexibility, it comes at a cost of greater complexity and more opportunities for user errors. <strong>If possible, prefer the functional API.<\/strong><\/p>\n<\/blockquote>\n\n<p>Well... it's always possible. <\/p>\n\n<h2>Subclassing layers<\/h2>\n\n<p>Here, there may be good reasons to do so. And these reasons are:<\/p>\n\n<ul>\n<li>You want a layer that performs custom calculations that are not available with regular layers    <\/li>\n<li>This layer must have persistent weights.   <\/li>\n<\/ul>\n\n<p>If you don't need \"both\" things above, you don't need to subclass a layer. If you just want \"custom calculations\" without weights, a <code>Lambda<\/code> layer is enough.   <\/p>\n"}]}
+{"QuestionId":57627360,"AnswerCount":1,"Tags":"<python><tensorflow><keras><neural-network>","CreationDate":"2019-08-23T13:27:51.620","AcceptedAnswerId":57627603.0,"OwnerUserId":8923269.0,"Title":"keras apply threshold for loss function","Body":"<p>I am developing a Keras model. My dataset is badly unbalanced, so I want to set a threshold for training and testing. If I'm not mistaken, when doing a backward propagation, neural network checks the predicted values with the original ones and calculate the error and based on the error, set new weights for neurons. <\/p>\n\n<p>As I know, Keras uses 0.5 for the threshold. I know there are ways to apply custom metrics (as recall and precision) with custom threshold, but that threshold is only used for calculating the recall, and it is not applied in the loss function. To be more clear, If I want to set 0.85 as my threshold, the neural network would use 0.5 as threshold to calculate loss and 0.85 for recall.<\/p>\n\n<p>Is there any ways to set this threshold for training as well?<\/p>\n","answers":[{"AnswerId":"57627603","CreationDate":"2019-08-23T13:41:46.260","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>There is no such a thing as a threshold for loss. <\/p>\n\n<p>A loss function must be \"differentiable\", thus it must be a \"continuous\" function.<br>\nThe best you can do is to set \"class weights\", such as these examples: <a href=\"https:\/\/stackoverflow.com\/questions\/57618482\/higher-loss-penalty-for-true-non-zero-predictions\/57618531#57618531\">Higher loss penalty for true non-zero predictions<\/a><\/p>\n"}]}
+{"QuestionId":57627406,"AnswerCount":1,"Tags":"<pytorch><autograd>","CreationDate":"2019-08-23T13:30:12.063","AcceptedAnswerId":57631496.0,"OwnerUserId":10855529.0,"Title":"How to use autograd.gradcheck in PyTorch?","Body":"<p>The documentation does not include any example use case of gradcheck, where would it be useful?<\/p>\n","answers":[{"AnswerId":"57631496","CreationDate":"2019-08-23T18:39:05.740","ParentId":null,"OwnerUserId":"2790047","Title":null,"Body":"<p>There's an example use case provided in the documentation here:<\/p>\n\n<p><a href=\"https:\/\/pytorch.org\/docs\/master\/notes\/extending.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/master\/notes\/extending.html<\/a><\/p>\n\n<blockquote>\n  <p>You probably want to check if the backward method you implemented\n  actually computes the derivatives of your function. It is possible by\n  comparing with numerical approximations using small finite\n  differences:<\/p>\n  \n  <pre class=\"lang-py prettyprint-override\"><code>from torch.autograd import gradcheck\n\n# gradcheck takes a tuple of tensors as input, check if your gradient\n# evaluated with these tensors are close enough to numerical\n# approximations and returns True if they all verify this condition.\ninput = (torch.randn(20,20,dtype=torch.double,requires_grad=True), torch.randn(30,20,dtype=torch.double,requires_grad=True))\ntest = gradcheck(linear, input, eps=1e-6, atol=1e-4)\nprint(test)\n<\/code><\/pre>\n<\/blockquote>\n\n<p>As the quote above suggests, the purpose of the <code>gradcheck<\/code> function is to verify that a custom backward function agrees with a numerical approximation of the gradient. The primary use case is when you're implementing a custom backward operation. <strong>In very few cases should you be implementing your own backward function in PyTorch<\/strong>. This is because PyTorch's autograd functionality takes care of computing gradients for the vast majority of operations.<\/p>\n\n<p>The most obvious exceptions are<\/p>\n\n<ol>\n<li><p>You have a function which can't be expressed as a finite combination of other differentiable functions (for example, if you needed the incomplete gamma function, you might want to write your own forward and backward which used numpy and\/or lookup tables).<\/p><\/li>\n<li><p>You're looking to speed up the computation of a particularly complicated expression for which the gradient could be drastically simplified after applying the chain rule.<\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":57627833,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-08-23T13:56:06.090","AcceptedAnswerId":null,"OwnerUserId":1279459.0,"Title":"Taking a norm of matrix rows\/cols in pytorch","Body":"<p>The norm of a vector can be taken by <\/p>\n\n<pre><code>torch.norm(vec)\n<\/code><\/pre>\n\n<p>However, how to take a norm of a set of vectors grouped as a matrix (either as rows or columns)?<\/p>\n\n<p>For example, if a matrix size is (5,8), then the rows norms should return a vector of norms of size (5).<\/p>\n","answers":[{"AnswerId":"57629671","CreationDate":"2019-08-23T16:01:48.787","ParentId":null,"OwnerUserId":"2790047","Title":null,"Body":"<p><code>torch.norm<\/code> without extra arguments performs what is called a Frobenius norm which is effectively reshaping the matrix into one long vector and returning the 2-norm of that. To take the norm along a particular dimension provide the optional <code>dim<\/code> argument.<\/p>\n\n<p>For example <code>torch.norm(mat, dim=1)<\/code> will compute the 2-norm along the columns (i.e. this will compute the 2-norm of each row) thus converting a <code>mat<\/code> of size <code>[N,M]<\/code> to a vector of norms of size <code>[N]<\/code>.<\/p>\n\n<p>To compute the norm of the columns use <code>dim=0<\/code>.<\/p>\n"}]}
+{"QuestionId":57627872,"AnswerCount":0,"Tags":"<python><list><numpy><tensorflow><deep-learning>","CreationDate":"2019-08-23T13:59:05.097","AcceptedAnswerId":null,"OwnerUserId":10763046.0,"Title":"Tensorflow and python numpy: how to concat tensor\/list with specific index","Body":"<p>I was the author of <a href=\"https:\/\/stackoverflow.com\/questions\/57451767\/tensorflow-how-to-concat-tensor-with-specific-index\">Tensorflow: how to concat tensor with specific index<\/a><\/p>\n\n<p>Today I want to concatenation this two list  <\/p>\n\n<pre><code>tensor_a = [[[[255,255,255]]], [[[100,100,100]]]]\ntensor_b = [[[[0.1,0.2]]], [[[0.3,0.4]]]]\n<\/code><\/pre>\n\n<p>into the format like this<\/p>\n\n<pre><code>tensor_d = [[[[[255,255,255],[0.1,0.2]]]], [[[[100,100,100], [0.3,0.4]]]]]\n<\/code><\/pre>\n\n<p>I need to do this convertion:<\/p>\n\n<ol>\n<li><p>using tensorflow (so tensor_a  and tensor_b are tensors)<\/p><\/li>\n<li><p>using list\/numpy (so tensor_a  and tensor_b are lists)<\/p><\/li>\n<\/ol>\n\n<p>How can I do?<\/p>\n","answers":[]}
+{"QuestionId":57627920,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><machine-learning><lstm><recurrent-neural-network>","CreationDate":"2019-08-23T14:02:59.677","AcceptedAnswerId":null,"OwnerUserId":11967515.0,"Title":"How to fix ''generator' object is not subscriptable' error in python, with Tensorflow and stacked LSTMs?","Body":"<p>I was trying to create a recurrent neural network model using an LSTM model in Tensorflow.\nIt had worked fine with a single LSTM however I wanted to stack LSTMs using tf.keras.layers.StackedRNNCells.\nI tried doing so however I keep getting the error '<strong><em>'generator' object is not subscriptable<\/em><\/strong>'.\nAny help in fixing this would be much appreciated!<\/p>\n\n<p>I am using Google Colab and Python 3.\nI had been following an online tutorial to build a recurrent neural network as an LSTM  which was similar to the task I was trying to accomplish. I had tweaked the model and data to fit the needs of my own project.\nI had begun with a single LSTMCell using <code>tf.nn.rnn_cell.LSTMCell<\/code> but then wished to stack  multiple such LSTMs using <code>tf.keras.layers.StackedRNNCells<\/code>. The problem is it is throwing the above mentioned error.<\/p>\n\n<p>So the code I had used is as so:<\/p>\n\n<pre><code>r_neuron = 120    \nnum_layers = 10\n\nX = tf.placeholder(tf.float32, [None, n_windows, n_input])   \ny = tf.placeholder(tf.float32, [None, n_windows, n_output])\n\ncell = tf.nn.rnn_cell.LSTMCell(num_units = r_neuron, activation = tf.nn.relu)\ncells = tf.keras.layers.StackedRNNCells(cell for _ in range(num_layers))\nrnn_output, states = tf.nn.dynamic_rnn(cells, X, dtype=tf.float32)    \n<\/code><\/pre>\n\n<p>The entire error message is as so:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-294-a77a03300a83&gt; in &lt;module&gt;()\n     16 cell = tf.nn.rnn_cell.LSTMCell(num_units = r_neuron, activation = tf.nn.relu)\n     17 cells = tf.keras.layers.StackedRNNCells(cell for _ in range(num_layers))\n---&gt; 18 rnn_output, states = tf.nn.dynamic_rnn(cells, X, dtype=tf.float32)\n     19 \n     20 \n\n4 frames\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/layers\/recurrent.py in output_size(self)\n     96   @property\n     97   def output_size(self):\n---&gt; 98     if getattr(self.cells[-1], 'output_size', None) is not None:\n     99       return self.cells[-1].output_size\n    100     elif _is_multiple_state(self.cells[-1].state_size):\n\nTypeError: 'generator' object is not subscriptable\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57627943,"AnswerCount":1,"Tags":"<python><pytorch><tensor>","CreationDate":"2019-08-23T14:04:51.103","AcceptedAnswerId":57631805.0,"OwnerUserId":5516760.0,"Title":"Invalid multinomial distribution (encountering probability entry < 0) at \/pytorch\/aten\/src\/TH\/generic\/THTensorRandom.cpp:325","Body":"<p>I have a <code>PyTorch<\/code> tensor called <code>out_probs<\/code> which is produced like this:<\/p>\n\n<pre><code>out_probs=F.softmax(out_dec[:,0],dim=0)\n<\/code><\/pre>\n\n<p>Also, the shape of <code>out_probs<\/code> is <code>[128,20004]<\/code><\/p>\n\n<p><code>out_probs<\/code> is the result of a softmax operation and it's not supposed to contain any negative value so naturally the result of <code>out_probs[out_probs&lt;0<\/code> is going to be an empty tensor(actually I checked and it was empty)\nBut when I'm running<\/p>\n\n<pre><code>torch.multinomial(out_probs, 1)\n<\/code><\/pre>\n\n<p>I'm getting :<\/p>\n\n<pre><code>RuntimeError: invalid argument 2: invalid multinomial distribution (encountering probability entry &lt; 0) at \/pytorch\/aten\/src\/TH\/generic\/THTensorRandom.cpp:325\n<\/code><\/pre>\n\n<p>That implies my tensor has a negative entry and I don't know why this is happening?<\/p>\n","answers":[{"AnswerId":"57631805","CreationDate":"2019-08-23T19:10:00.370","ParentId":null,"OwnerUserId":"2790047","Title":null,"Body":"<p>I believe you've found a bug in the error reporting for <code>torch.multinomial<\/code>.<\/p>\n\n<p>For example<\/p>\n\n<pre><code>x = torch.ones(128, 1)\nx[0] *= 1e100\nout_probs = F.softmax(x, dim=0)\nprint('Negative values:', torch.sum(out_probs &lt; 0).item())\ny = torch.multinomial(out_probs, 1)\n<\/code><\/pre>\n\n<p>results in the following output<\/p>\n\n<pre><code>Negative values: 0\nRuntimeError: invalid argument 2: invalid multinomial distribution (encountering probability entry &lt; 0) at \/pytorch\/aten\/src\/TH\/generic\/THTensorRandom.cpp:298\n<\/code><\/pre>\n\n<p>It turns out this is getting triggered because <code>out_probs<\/code> contains <code>nan<\/code> entries.<\/p>\n\n<pre><code>print('nan values:', torch.sum(torch.isnan(out_probs)).item())\n<\/code><\/pre>\n\n<p>gives<\/p>\n\n<pre><code>nan values: 128\n<\/code><\/pre>\n\n<p>Which are caused by mathematical instabilities in softmax.<\/p>\n\n<p>Strangely, when the values in <code>out_probs<\/code> are infinite you get the proper error message<\/p>\n\n<pre><code>RuntimeError: invalid argument 2: invalid multinomial distribution (encountering probability entry = infinity or NaN) at \/pytorch\/aten\/src\/TH\/generic\/THTensorRandom.cpp:302\n<\/code><\/pre>\n\n<p>This bug should probably be reported at <a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/pytorch\/issues<\/a> if it hasn't been fixed in the most recent version.<\/p>\n\n<p>By the way I'm using PyTorch 1.0.1.post2<\/p>\n"}]}
+{"QuestionId":57628181,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><tf.keras>","CreationDate":"2019-08-23T14:21:25.523","AcceptedAnswerId":null,"OwnerUserId":10455493.0,"Title":"Is val_sample_weights broken in tf.keras?","Body":"<p>I cannot seem to make <code>val_sample_weights<\/code> work with validation data, using <code>model.fit()<\/code>. \nMy understanding of <code>val_sample_weights<\/code> is that it is suppose to do exactly the same as <code>sample_weight<\/code> such that reported <code>loss<\/code> and <code>val_loss<\/code> can be directly compared when training using <code>sample_weight<\/code>.<\/p>\n\n<p>From my example code below, we see that <code>loss<\/code> goes to zero when setting <code>sample_weight<\/code> to zero, but <code>val_loss<\/code> is unaffected when setting <code>val_sample_weights<\/code> to zero.<\/p>\n\n<p>Am I missing or misunderstanding something here, or does this seem like <code>val_sample_weights<\/code> has no use in <code>tf.keras<\/code>? Maybe a bug in <code>tf.keras<\/code>? I'm on Windows, tensorflow version 1.14.0.<\/p>\n\n<p>Below is a working example if you want to test it yourself:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\ndata_size = 100\ninput_size=3\nseed = 1\n\n# synetic train data\nx_train = np.random.rand(data_size ,input_size)\ny_train= np.random.rand(data_size,1)\n# synetic test data\nx_test = np.random.rand(data_size ,input_size)\ny_test= np.random.rand(data_size,1)\n\ntf.random.set_random_seed(seed)\n# create model\ninputs = tf.keras.layers.Input(shape=(input_size))\npred=tf.keras.layers.Dense(1, activation='sigmoid')(inputs)\nmodel = tf.keras.models.Model(inputs=inputs, outputs=pred)\nloss = tf.keras.losses.MSE\nmetrics = tf.keras.losses.MAE\nmodel.compile(loss=loss , metrics=[metrics], optimizer='adam')\n\n# Make model static, so we can compare it between different scenarios\nfor layer in model.layers:\n    layer.trainable = False\n\n# base model no weights (same result as without class_weights)\nmodel.fit(x=x_train,y=y_train, validation_data=(x_test,y_test))\n# which outputs:\n#&gt; 100\/100 [==============================] - 0s 1ms\/sample - loss: 0.0851 - mean_absolute_error: 0.2511 - val_loss: 0.0900 - val_mean_absolute_error: 0.2629\n\n#changing the sample_weights to zero, Hence loss and val_loss should be zero, and metrics should be the same\nsample_weight_train = np.zeros(100)\nsample_weight_val = np.zeros(100)\nmodel.fit(x=x_train,y=y_train,sample_weight=sample_weight_train, validation_data=(x_test,y_test,sample_weight_val))\n# which outputs:\n#&gt; 100\/100 [==============================] - 0s 860us\/sample - loss: 0.0000e+00 - mean_absolute_error: 0.2507 - val_loss: 0.0899 - val_mean_absolute_error: 0.2627\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57628203,"AnswerCount":0,"Tags":"<tensorflow><tf.keras>","CreationDate":"2019-08-23T14:22:32.310","AcceptedAnswerId":null,"OwnerUserId":11708498.0,"Title":"Binary Cross Entropy from Keras takes logits not from the output of model","Body":"<p>I am doing binary classification using Keras from TensorFlow 2.0, dense net with binary cross entropy loss.<\/p>\n\n<p>For debugging purpose I was trying to overfit my model first and found weird behavior: I got different results of loss whether I used sigmoid in last dense layer init or initialize this last layer without any activation and add sigmoid activation separately as next step of the model.<\/p>\n\n<p>I cannot find where in tf\/keras it is stated that if last layer of model is just sigmoid and <code>BinaryCrossentropy(from_logits=True)<\/code> loss will be computed without that sigmoid. Or if there is other explanation of the following results.<\/p>\n\n<p>As far as I understand <code>BinaryCrossentropy(from_logits=True)<\/code> expects logits- in this case output of Dense layer without any activation. That's the first example printed by posted code. As there is no activation, and only neuron has only one weight=1 the prediction is the same as input, as well as loss.<\/p>\n\n<p>In the second variant I used Dense layer without activation in constructor, however I added sigmoid activation as \"separate layer\" and set <code>BinaryCrossentropy(from_logits=False)<\/code> in this case BCE doesn't take logits but (\"probabilities\") computed by <code>sigmoid(logits)<\/code> as input. Output predicted by model (probability) is equal to 1, which makes sense. However I expected Keras to take that probability, change it back to logit and compute loss <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/backend\/tensorflow_backend.py#L3653\" rel=\"nofollow noreferrer\">source<\/a>. Due to computational imprecise (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r2.0\/tensorflow\/python\/keras\/losses.py#L373\" rel=\"nofollow noreferrer\">tf source<\/a>) I figured out that BCE cannot compute higher loss than 15.3332 from not logits but it does in this case. <\/p>\n\n<p>Expected behavior shows up in the last case (print of the code) where activation is in init of Dense layer- loss is clipped to 15.3332, no matter how big the input\/output is if it is above 16. <\/p>\n\n<p>My assumption is that in the second case loss is calculated from logits which is output of Dense layer, but is not the output of the model.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.python.keras.layers import Input, Dense\nfrom tensorflow.python.keras.activations import sigmoid\nfrom tensorflow.python.keras.models import Model\nfrom tensorflow.python.keras.losses import BinaryCrossentropy\nimport numpy as np\n\n\nx_train = np.array([[10000]])\ny_train = [0]\n\ninp = Input(x_train.shape[1])\nout = Dense(1, trainable=False, use_bias=False)(inp)\nmodel = Model(inp, out)\nmodel.get_layer('dense').set_weights([np.array([[1]])])\nmodel.compile(optimizer='adam', loss=BinaryCrossentropy(from_logits=True))\n\nprint('\\nFrom logits=True, no activation')\nmodel.fit(x_train, y_train, epochs=1)\nprint('Prediction:')\nprint(model.predict(x_train))\n\n\ninp = Input(x_train.shape[1])\nout = Dense(1, trainable=False, use_bias=False)(inp)\nout = sigmoid(out)\nmodel = Model(inp, out)\nmodel.get_layer('dense_1').set_weights([np.array([[1]])])\nmodel.compile(optimizer='adam', loss=BinaryCrossentropy(from_logits=False))\n\nprint('\\nFrom logits=False, separate sigmoid')\nmodel.fit(x_train, y_train, epochs=1)\nprint('Prediction:')\nprint(model.predict(x_train))\n\n\ninp = Input(x_train.shape[1])\nout = Dense(1, trainable=False, use_bias=False, activation='sigmoid')(inp)\nmodel = Model(inp, out)\nmodel.get_layer('dense_2').set_weights([np.array([[1]])])\nmodel.compile(optimizer='adam', loss=BinaryCrossentropy(from_logits=False))\n\nprint('\\nFrom logits=False, sigmoid in dense')\nmodel.fit(x_train, y_train, epochs=1)\nprint('Prediction:')\nprint(model.predict(x_train))\n<\/code><\/pre>\n\n<pre><code>From logits=True, no activation\n1\/1 [==============================] - 0s 17ms\/sample - loss: 1000.0000\nPrediction:\n[[1000.]]\n\nFrom logits=False, separate sigmoid\nTrain on 1 samples\n1\/1 [==============================] - 0s 15ms\/sample - loss: 1000.0000\nPrediction:\n[[1.]]\n\nFrom logits=False, sigmoid in dense\nTrain on 1 samples\n1\/1 [==============================] - 0s 17ms\/sample - loss: 15.3332\nPrediction:\n[[1.]]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57628457,"AnswerCount":1,"Tags":"<python><python-3.x><numpy><pytorch><padding>","CreationDate":"2019-08-23T14:40:20.687","AcceptedAnswerId":57628825.0,"OwnerUserId":5556400.0,"Title":"padding a list of torch tensors (or numpy arrays)","Body":"<p>Let's say I have a list as the following:<\/p>\n\n<pre><code>l = [torch.randn(2,3), torch.randn(2,4),torch.randn(2,5)]\n<\/code><\/pre>\n\n<p>I want to zero pad all of them in the second dimension, so they will extend as far as 5 elements (5 being the max number between the three of the elements in the second dimension). How can I do this. I tried this but failed:<\/p>\n\n<pre><code>from torch.nn.utils.rnn import pad_sequence\npad_sequence(l, batch_first=True, padding_value=0)\n<\/code><\/pre>\n\n<p>which caused the following error: <\/p>\n\n<pre><code>RuntimeError: The expanded size of the tensor (3) must match the existing size (4) at non-singleton dimension 1.  Target sizes: [2, 3].  Tensor sizes: [2, 4]\n<\/code><\/pre>\n\n<p>The equivalent answer in Numpy would also be appreciated. <\/p>\n","answers":[{"AnswerId":"57628825","CreationDate":"2019-08-23T15:03:55.850","ParentId":null,"OwnerUserId":"11946287","Title":null,"Body":"<p>One option is to use <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.pad.html\" rel=\"nofollow noreferrer\"><code>np.pad<\/code><\/a>. <\/p>\n\n<p>Example:<\/p>\n\n<pre><code>import numpy as np\na = np.random.randn(2, 3)\nb = np.pad(a, [(0, 0), (0, 2)], mode='constant') \n<\/code><\/pre>\n\n<p>Print <code>a<\/code> gives  <\/p>\n\n<pre><code>[[ 1.22721163  1.23456672  0.51948003]\n [ 0.16545496  0.06609003 -0.32071653]]\n<\/code><\/pre>\n\n<p>Print <code>b<\/code> gives<\/p>\n\n<pre><code>[[ 1.22721163  1.23456672  0.51948003  0.          0.        ]\n [ 0.16545496  0.06609003 -0.32071653  0.          0.        ]]\n<\/code><\/pre>\n\n<p>The second argument of <code>pad<\/code> is <code>pad_width<\/code> which is a list of before\/after paddings for each dimension. So in this example no padding in the first dimension and two paddings at the end of the second dimension. <\/p>\n\n<p>There are lots of other <code>mode<\/code> options you can use so check out the docs.<\/p>\n\n<p>For your particular problem you'll need to add an extra step to work out the padding for each array.<\/p>\n\n<p><strong>Edit<\/strong><\/p>\n\n<p>For <code>pytorch<\/code> I think you want <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.functional.html#pad\" rel=\"nofollow noreferrer\"><code>torch.nn.functional.pad<\/code><\/a> e.g.<\/p>\n\n<pre><code>import torch\nt = torch.randn(2, 3)\ntorch.nn.functional.pad(t, (0, 2))\n<\/code><\/pre>\n\n<p><strong>Edit 2<\/strong><\/p>\n\n<p>The <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.utils.rnn.pad_sequence\" rel=\"nofollow noreferrer\"><code>torch.nn.utils.rnn.pad_sequence<\/code><\/a> requires the trailing dimensions of all the tensors in the list to be the same so you need to some transposing for it to work nicely<\/p>\n\n<pre><code>import torch\n# l = [torch.randn(2,3), torch.randn(2,4),torch.randn(2,5)]\n# l = [i.transpose(0, 1) for i in l]  \n# or simply make you tensors with switched dimensions\nl = [torch.randn(3,2), torch.randn(4,2),torch.randn(5,2)]\nout = torch.nn.utils.rnn.pad_sequence(l, batch_first=True)\n# out will now be a tensor with shape (3, 5, 2)\n# You can transpose it back to (3, 2, 5) with\nout = out.transpose(1, 2)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57628606,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-23T14:49:36.313","AcceptedAnswerId":null,"OwnerUserId":3447343.0,"Title":"Tensorflow 2 summaries with eager disabled","Body":"<p>Since there are currently couple of issues with TF2 eager execution (e.g. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/30561\" rel=\"nofollow noreferrer\">here<\/a>, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/30533#issuecomment-518038676\" rel=\"nofollow noreferrer\">here<\/a> or <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/30952\" rel=\"nofollow noreferrer\">there<\/a>), I am disabling it by calling <code>tf.compat.v1.disable_eager_execution()<\/code>.<\/p>\n\n<p>However, I would be very happy if I still could <strong>log stuff to tensorboard<\/strong>. The new version of file writer (which one gets by calling <code>tf.summary.create_file_writer()<\/code>) does not create any files if I disable eager execution. The old one (<code>tf.compat.v1.summary.FileWriter()<\/code>) does create a file but I am not able to write anything in. The <code>add_summary<\/code> function requires a summary protocol buffer (not a value) and I don't know how to create it without using the v1 session context.<\/p>\n\n<p>Is there a way to write summaries in TF2 with disabled eager execution?<\/p>\n","answers":[]}
+{"QuestionId":57628938,"AnswerCount":1,"Tags":"<tensorflow><object-detection>","CreationDate":"2019-08-23T15:10:22.667","AcceptedAnswerId":null,"OwnerUserId":4582711.0,"Title":"Initialize the bias term of the last layer","Body":"<p>I was reading the <a href=\"https:\/\/leimao.github.io\/blog\/Focal-Loss-Explained\/\" rel=\"nofollow noreferrer\">this<\/a> blog about focal loss. In the section <code>Focal Loss Trick<\/code> it says: <\/p>\n\n<blockquote>\n  <p>Facebook AI Research used is to initialize the bias term of the last\n  layer to some non-zero value such that the pt of positive samples is\n  small and the pt of negative samples is large. Concretely, they set\n  the bias term b=\u2212log((1\u2212\u03c0)\/\u03c0). Here \u03c0 is simply are variable instead\n  of the ordinary \u03c0. In their case, they set \u03c0=0.01, therefore b\u226bwx.<\/p>\n<\/blockquote>\n\n<p>I want to do the same using tensorflow object detection api. Here, the focal loss is given by the following line in config file:<\/p>\n\n<blockquote>\n  <p>loss {\n        classification_loss {\n          weighted_sigmoid_focal {\n            alpha: 0.25\n            gamma: 2.0\n          }\n        } }<\/p>\n<\/blockquote>\n\n<p>But I don't know how to set the bias term of the last layer to some non-zero value. How to achieve it in tensorflow ?<\/p>\n","answers":[{"AnswerId":"57643009","CreationDate":"2019-08-25T03:24:32.653","ParentId":null,"OwnerUserId":"4582711","Title":null,"Body":"<p>It's given by <code>class_prediction_bias_init<\/code> in the <code>box_predictor<\/code>. So, the config file will look something like this:<\/p>\n\n<pre><code>box_predictor {\n      weight_shared_convolutional_box_predictor {\n        class_prediction_bias_init: -1.99\n      }\n}\n<\/code><\/pre>\n"}]}
+{"QuestionId":57628958,"AnswerCount":0,"Tags":"<numpy><tensorflow><keras><casting><dtype>","CreationDate":"2019-08-23T15:11:37.050","AcceptedAnswerId":null,"OwnerUserId":4040643.0,"Title":"Return value of pyfunc_0 is double, but expects float","Body":"<p>I am currently trying to better understand Tensorflows CustomLayer feature. While implementing such a custom layer, I ran into the following error:<\/p>\n\n<pre><code>\/usr\/lib\/python3\/dist-packages\/skimage\/util\/dtype.py:110: UserWarning: Possible precision loss when converting from float64 to uint16\n  \"%s to %s\" % (dtypeobj_in, dtypeobj))\n\/usr\/lib\/python3\/dist-packages\/skimage\/exposure\/exposure.py:307: RuntimeWarning: invalid value encountered in true_divide\n  image = (image - imin) \/ float(imax - imin)\nTraceback (most recent call last):\n  File \"\/local\/tensorflow-1.4.0\/python3.5\/tensorflow\/python\/client\/session.py\", line 1323, in _do_call\n    return fn(*args)\n  File \"\/local\/tensorflow-1.4.0\/python3.5\/tensorflow\/python\/client\/session.py\", line 1302, in _run_fn\n    status, run_metadata)\n  File \"\/local\/tensorflow-1.4.0\/python3.5\/tensorflow\/python\/framework\/errors_impl.py\", line 473, in __exit__\n    c_api.TF_GetCode(self.status.status))\ntensorflow.python.framework.errors_impl.InvalidArgumentError: 0-th value returned by pyfunc_0 is double, but expects float\n         [[Node: model\/pretrained\/custom_layer_1\/map\/while\/custom_image_op = PyFuncStateless[Tin=[DT_FLOAT], Tout=[DT_FLOAT], token=\"pyfunc_0\", _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](model\/pretrained\/custom_layer_1\/map\/while\/TensorArrayReadV3\/_49)]]\n         [[Node: model\/pretrained\/custom_layer_1\/map\/while\/TensorArrayWrite\/TensorArrayWriteV3\/_55 = _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_126_model\/pretrained\/custom_layer_1\/map\/while\/TensorArrayWrite\/TensorArrayWriteV3\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](^_cloopmodel\/pretrained\/custom_layer_1\/map\/while\/NextIteration_1\/_20)]]\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"train.py\", line 94, in &lt;module&gt;\n    trainer.run(sess)\n  File \"\/home\/\/Downloads\/dl_ss_19-master\/Ex4\/train\/trainer.py\", line 128, in run\n    self._train_epoch(sess)\n  File \"\/home\/train\/trainer.py\", line 49, in _train_epoch\n    self._model_is_training: True})\n  File \"\/local\/tensorflow-1.4.0\/python3.5\/tensorflow\/python\/client\/session.py\", line 889, in run\n    run_metadata_ptr)\n  File \"\/local\/tensorflow-1.4.0\/python3.5\/tensorflow\/python\/client\/session.py\", line 1120, in _run\n    feed_dict_tensor, options, run_metadata)\n  File \"\/local\/tensorflow-1.4.0\/python3.5\/tensorflow\/python\/client\/session.py\", line 1317, in _do_run\n    options, run_metadata)\n  File \"\/local\/tensorflow-1.4.0\/python3.5\/tensorflow\/python\/client\/session.py\", line 1336, in _do_call\n    raise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: 0-th value returned by pyfunc_0 is double, but expects float\n         [[Node: model\/pretrained\/custom_layer_1\/map\/while\/custom_image_op = PyFuncStateless[Tin=[DT_FLOAT], Tout=[DT_FLOAT], token=\"pyfunc_0\", _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](model\/pretrained\/custom_layer_1\/map\/while\/TensorArrayReadV3\/_49)]]\n         [[Node: model\/pretrained\/custom_layer_1\/map\/while\/TensorArrayWrite\/TensorArrayWriteV3\/_55 = _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_126_model\/pretrained\/custom_layer_1\/map\/while\/TensorArrayWrite\/TensorArrayWriteV3\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](^_cloopmodel\/pretrained\/custom_layer_1\/map\/while\/NextIteration_1\/_20)]]\n\nCaused by op 'model\/pretrained\/custom_layer_1\/map\/while\/custom_image_op', defined at:\n  File \"train.py\", line 61, in &lt;module&gt;\n    prediction_logits = model.pretrained(img_conc)\n  File \"\/home\/model\/pretrained.py\", line 56, in create\n    x = CustomLayer(output_dim=(224, 224, 3))(inputs)\n  File \"\/local\/tensorflow-1.4.0\/python3.5\/keras\/engine\/topology.py\", line 603, in __call__\n    output = self.call(inputs, **kwargs)\n  File \"\/home\/model\/pretrained.py\", line 41, in call\n    res = tf.map_fn(preprocess_input, x)\n  File \"\/local\/tensorflow-1.4.0\/python3.5\/tensorflow\/python\/ops\/functional_ops.py\", line 389, in map_fn\n    swap_memory=swap_memory)\n  File \"\/local\/tensorflow-1.4.0\/python3.5\/tensorflow\/python\/ops\/control_flow_ops.py\", line 2816, in while_loop\n    result = loop_context.BuildLoop(cond, body, loop_vars, shape_invariants)\n  File \"\/local\/tensorflow-1.4.0\/python3.5\/tensorflow\/python\/ops\/control_flow_ops.py\", line 2640, in BuildLoop\n    pred, body, original_loop_vars, loop_vars, shape_invariants)\n  File \"\/local\/tensorflow-1.4.0\/python3.5\/tensorflow\/python\/ops\/control_flow_ops.py\", line 2590, in _BuildLoop\n    body_result = body(*packed_vars_for_body)\n  File \"\/local\/tensorflow-1.4.0\/python3.5\/tensorflow\/python\/ops\/functional_ops.py\", line 379, in compute\n    packed_fn_values = fn(packed_values)\n  File \"\/home\/model\/pretrained.py\", line 30, in preprocess_input\n    name='custom_image_op')\n  File \"\/local\/tensorflow-1.4.0\/python3.5\/tensorflow\/python\/ops\/script_ops.py\", line 215, in py_func\n    input=inp, token=token, Tout=Tout, name=name)\n  File \"\/local\/tensorflow-1.4.0\/python3.5\/tensorflow\/python\/ops\/gen_script_ops.py\", line 90, in _py_func_stateless\n    \"PyFuncStateless\", input=input, token=token, Tout=Tout, name=name)\n  File \"\/local\/tensorflow-1.4.0\/python3.5\/tensorflow\/python\/framework\/op_def_library.py\", line 787, in _apply_op_helper\n    op_def=op_def)\n  File \"\/local\/tensorflow-1.4.0\/python3.5\/tensorflow\/python\/framework\/ops.py\", line 2956, in create_op\n    op_def=op_def)\n  File \"\/local\/tensorflow-1.4.0\/python3.5\/tensorflow\/python\/framework\/ops.py\", line 1470, in __init__\n    self._traceback = self._graph._extract_stack()  # pylint: disable=protected-access\n\nInvalidArgumentError (see above for traceback): 0-th value returned by pyfunc_0 is double, but expects float\n         [[Node: model\/pretrained\/custom_layer_1\/map\/while\/custom_image_op = PyFuncStateless[Tin=[DT_FLOAT], Tout=[DT_FLOAT], token=\"pyfunc_0\", _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](model\/pretrained\/custom_layer_1\/map\/while\/TensorArrayReadV3\/_49)]]\n         [[Node: model\/pretrained\/custom_layer_1\/map\/while\/TensorArrayWrite\/TensorArrayWriteV3\/_55 = _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_126_model\/pretrained\/custom_layer_1\/map\/while\/TensorArrayWrite\/TensorArrayWriteV3\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](^_cloopmodel\/pretrained\/custom_layer_1\/map\/while\/NextIteration_1\/_20)]]\n<\/code><\/pre>\n\n<p>Thus far I tried to cast the various parameters and change the arguments of the function. This however didn't influence the result at all. I am at a loss. Does someone have experience with pyfunc_0? If yes, how can I fix below's code?<\/p>\n\n<p>Links I found useful:<\/p>\n\n<p><a href=\"https:\/\/github.com\/cesc-park\/attend2u\/issues\/2\" rel=\"nofollow noreferrer\">https:\/\/github.com\/cesc-park\/attend2u\/issues\/2<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/rstudio\/tensorflow\/issues\/169\" rel=\"nofollow noreferrer\">https:\/\/github.com\/rstudio\/tensorflow\/issues\/169<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/54278894\/0-th-value-returned-by-pyfunc-0-is-double-but-expects-float-though-i-think-it\">&quot;0-th value returned by pyfunc_0 is double, but expects float&quot; though I think it returns float<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/57605021\/including-advanced-computation-scikit-like-in-a-keras-custom-layer\">Including advanced computation (scikit-like) in a keras custom layer<\/a><\/p>\n\n<pre><code>def equalize(img):\n    img_adapteq = 255 * exposure.equalize_adapthist(img \/\/ 255)\n    return img_adapteq.astype(np.float)\n\n\ndef preprocess_input(img):\n    x = tf.py_func(equalize,\n                   [img],\n                   'float32',\n                   stateful=False,\n                   name='custom_image_op')\n    return tf.cast(x, tf.float32)\n\n\nclass CustomLayer(Layer):\n    def __init__(self, output_dim, **kwargs):\n        self.output_dim = output_dim\n        self.trainable = False\n        super(CustomLayer, self).__init__(**kwargs)\n\n    def call(self, x):\n        res = tf.map_fn(preprocess_input, x)\n        res.set_shape([x.shape[0],\n                       self.output_dim[1],\n                       self.output_dim[0],\n                       x.shape[-1]])\n        return res\n\ninputs = Input(shape=(224, 224, 3), tensor=x)\nx = CustomLayer(output_dim=(224, 224, 3))(inputs)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57628990,"AnswerCount":0,"Tags":"<python-3.x><keras><deep-learning>","CreationDate":"2019-08-23T15:13:27.540","AcceptedAnswerId":null,"OwnerUserId":7641030.0,"Title":"How to get the proper input shape for Conv2D in Keras for Image Classification","Body":"<p>I try to classify images into two classes. In my predictor variable, i have 100x100 images, in my criterion variable I have 0 and 1 values to indicate classes.<\/p>\n\n<p>When I try to fit my keras model, I get the following error:\n\"You are passing a target array of shape (160, 1) while using as loss categorical_crossentropy<code>.<\/code>categorical_crossentropy` expects targets to be binary matrices (1s and 0s) of shape (samples, classes). If your targets are integer classes, you can convert them to the expected format via:<\/p>\n\n<p>from keras.utils import to_categorical\ny_binary = to_categorical(y_int)\"<\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nimport keras\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Dropout, Flatten\nfrom keras.layers import Conv2D, MaxPooling2D\nfrom keras.utils import to_categorical\nfrom keras.preprocessing import image\nfrom sklearn.model_selection import train_test_split\n\nX = resized_images #200 images * 100x100 size\nX = np.asarray(X)\ny = pd.read_csv(\"Malaria_Images_Class.csv\") #contains a column with 0 and 1 for 200 images\ny = df[\"Class\"] #drops unnecessary information\n\nX_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42, test_size=0.2,stratify=y)\n\nmodel = Sequential()\nmodel.add(Conv2D(32, (3, 3),activation='relu',input_shape=(100,100,1)))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Flatten())\nmodel.add(Dense(128, activation='relu'))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='categorical_crossentropy',optimizer='Adam',metrics=['accuracy'])\n\nX_train = X_train.reshape(160,100,100,1) #reshape to fit necessary input shape of Conv2D\nX_test = X_test.reshape(40,100,100,1)\n#y_train shape is (160,), y_test is (40,)\n\nmodel.fit(X_train,y_train,epochs=1,batch_size=25, validation_data=(X_test, y_test), verbose=1)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57629034,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2019-08-23T15:16:25.067","AcceptedAnswerId":null,"OwnerUserId":3711266.0,"Title":"Categorical accuracy in Tensorflow Keras not computing correctly when predictions are small non-zero floats","Body":"<p>I am having a strange problem using the Keras API in Tensorflow 1.13.1. My loss is <code>categorical_crossentropy<\/code> so I am using one-hot encoding vectors, and my metric is <code>categorical_accuracy<\/code>. I've been testing just running one epoch of 100 samples to validate the basic execution with my model. Sometimes the predictions look like this, and everything works as expected and I get meaningful <code>categorical_accuracy<\/code> values (ignore that the prediction is always the same, it's the first epoch):<\/p>\n\n<pre><code>TARGET = [1 0], PREDICT = [[0. 1.]]\nTARGET = [0 1], PREDICT = [[0. 1.]]\nTARGET = [1 0], PREDICT = [[0. 1.]]\nTARGET = [0 1], PREDICT = [[0. 1.]]\nTARGET = [1 0], PREDICT = [[0. 1.]]\nTARGET = [1 0], PREDICT = [[0. 1.]]\nTARGET = [0 1], PREDICT = [[0. 1.]]\nTARGET = [1 0], PREDICT = [[0. 1.]]\nTARGET = [1 0], PREDICT = [[0. 1.]]\nTARGET = [0 1], PREDICT = [[0. 1.]]\n<\/code><\/pre>\n\n<p>However, sometimes the predictions look like this:<\/p>\n\n<pre><code>TARGET = [0 1], PREDICT = [[1.0000000e+00 1.1421418e-20]]\nTARGET = [0 1], PREDICT = [[1.0000000e+00 3.2770412e-19]]\nTARGET = [1 0], PREDICT = [[1.000000e+00 8.535294e-32]]\nTARGET = [1 0], PREDICT = [[1.000000e+00 6.535944e-20]]\nTARGET = [0 1], PREDICT = [[1.000000e+00 5.678613e-20]]\nTARGET = [1 0], PREDICT = [[1.000000e+00 4.151403e-23]]\nTARGET = [1 0], PREDICT = [[1.0000000e+00 1.7497807e-22]]\nTARGET = [0 1], PREDICT = [[1.0000e+00 4.2055e-16]]\nTARGET = [0 1], PREDICT = [[1.000000e+00 5.367323e-28]]\nTARGET = [1 0], PREDICT = [[1.0000000e+00 2.4306246e-17]]\n<\/code><\/pre>\n\n<p>And the <code>categorical_accuracy<\/code> value ends up being 1.0 if it always predicts <code>1.0000000e+00<\/code> for the first label (i.e. essentially predicting <code>[1 0]<\/code>), and the <code>categorical_accuracy<\/code> value equals 0.0 if it predicts all to be the second label (i.e. essentially predicting <code>[0 1]<\/code>).<\/p>\n\n<p>These are two consecutive runs without changing any code. What could be causing this output difference, and why does the latter case give me such nonsense results for the <code>categorical_accuracy<\/code> value? My guess is it has something to do with the small non-zero float values messing something up, but I am unsure.<\/p>\n","answers":[]}
+{"QuestionId":57629420,"AnswerCount":0,"Tags":"<tensorflow><keras><neural-network><deep-learning><shap>","CreationDate":"2019-08-23T15:43:06.697","AcceptedAnswerId":null,"OwnerUserId":3198674.0,"Title":"Use of DeepExplainer to get shap values for an MLP model in Keras with tensorflow backend","Body":"<p>I am playing around with DeepExplainer to get shap values for deep learning models. By following some tutorials I can get some results, i.e. what variables are pushing the model prediction from the base value, which is the average model output in training set. <\/p>\n\n<p>I have around 5,000 observations along with 70 features. The performance of DeepExplainer is quite satisfactory. And my code is:<\/p>\n\n<pre><code>model0 = load_model(model_p+'health0.h5')\n\nbackground = healthScaler.transform(train[healthFeatures])\ne = shap.DeepExplainer(model0, background)\nshap_values = e.shap_values(healthScaler.transform(test[healthFeatures]))\n\ntest2 = test[healthFeatures].copy()\ntest2[healthFeatures] = healthScaler.transform(test[healthFeatures])\n\nshap.force_plot(e.expected_value[0], shap_values[0][947,:], test2.iloc[947,:])\n<\/code><\/pre>\n\n<p>And the plot is the following:\n<a href=\"https:\/\/i.stack.imgur.com\/dDM5T.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/dDM5T.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Here the base value is 0.012 (can also be seen through <strong>e.expected_value[0]<\/strong>) and very close to the output value which is 0.01.<\/p>\n\n<p>At this point I have some questions:<\/p>\n\n<p>1) The output value is not identical to the prediction gotten through <strong>model0.predict(test[healthFeatures])[947] = -0.103<\/strong> How should I assess output value?<\/p>\n\n<p>2) As can be seen, I am using whole training set as the background to approximate conditional expectations of SHAP values. What is the difference between using random samples from training set and entire set? Is it only related to performance issue?<\/p>\n\n<p>Many thanks in advance!<\/p>\n","answers":[]}
+{"QuestionId":57629648,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-23T16:00:14.077","AcceptedAnswerId":57630185.0,"OwnerUserId":6643799.0,"Title":"Tensorflow 2.0 , replace 0 values in a tensor with 1s","Body":"<p>So I'm trying to implement a centering and scaling with tensorflow where I need to replace values==0 in the tensor with 1.0. <\/p>\n\n<p>I know to do this with numpy <code>x_std[x_std == 0.0] = 1.0<\/code> but can't figure out the best way to do it on tensorflow 2.0.<\/p>\n\n<p>The code is:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def _center_scale_xy(X, Y, scale=True):\n    \"\"\" Center X, Y and scale if the scale parameter==True\n    Returns\n    -------\n        X, Y, x_mean, y_mean, x_std, y_std\n    \"\"\"\n    # center\n    x_mean = tf.reduce_mean(X,axis=0)\n    X -= x_mean\n    y_mean = tf.reduce_mean(Y,axis=0)\n    Y -= y_mean\n    # scale\n    if scale:\n        x_std = tf.math.reduce_std(X,axis=0)\n        #x_std[x_std == 0.0] = 1.0 #This one I need to implement with tensors\n        X \/= x_std\n        y_std = tf.math.reduce_std(Y,axis=0)\n        y_std[y_std == 0.0] = 1.0\n        Y \/= y_std\n    else:\n        x_std = np.ones(X.shape[1])\n        y_std = np.ones(Y.shape[1])\n    return X, Y, x_mean, y_mean, x_std, y_std\n<\/code><\/pre>\n","answers":[{"AnswerId":"57630185","CreationDate":"2019-08-23T16:40:54.647","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/where\" rel=\"nofollow noreferrer\"><code>tf.where<\/code><\/a> like this:<\/p>\n\n<pre><code>x_std = tf.where(tf.equal(x_std, 0), tf.ones_like(x_std), x_std)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57630002,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-08-23T16:27:13.373","AcceptedAnswerId":null,"OwnerUserId":10452303.0,"Title":"different data augmentation parameters every epoch","Body":"<p>i'm using keras with a simple cnn model.\ni want to add gaussian noise to images in training. i want to change the noise parameters (mean and sigma) every epoch,based on some function. for example,<\/p>\n\n<pre><code>in epoch 1 i want to add noise with sigma=1\nin epoch 2 i want to add noise with sigma=2\nin epoch 3 i want to add noise with sigma=3\n# note-mean is always zero\n<\/code><\/pre>\n\n<p>and so on...<\/p>\n\n<p>inefficient way to solve it is with a for loop, save and load the mode after every epoch and call augmentation function.\nmore efficient way will be with a custom callback or generator, which i didn't succeed to do<\/p>\n\n<p>inefficient way:<\/p>\n\n<pre><code>total_num_of_epochs=100\ndef sigma_function(current_epoch):\n     sigma_fun=current_epoch\/total_num_of_epochs\n     return sigma_fun\n\nfor i in range(total_num_of_epochs):\n    x_train += np.random.normal(mean=0,sigma=sigma_fun(i),size=x_train shape) # augment x_train based on sigma_function and current epochs\n\n    model.compile(...)\n    model.fit(x_train ,y_train...initial_epoch=i,epochs=i+1) #load the model \n    # from previous loop\n    save model\n    load model for next loop\n<\/code><\/pre>\n\n<p>the desired result (i tried with ImageDataGenerator but maybe callback can do):<\/p>\n\n<pre><code>def sigma_function(current_epoch):\n     sigma_fun=current_epoch\/total_num_of_epochs\n     return sigma_fun\n\ndatagen=ImageDataGenerator(preprocessing_function=sigma_function)\ndatagen.fit(x_train)\n\n\nmodel.fit_generator(... don't know what to put here)\n<\/code><\/pre>\n\n<h2>edit<\/h2>\n\n<p>according to the proposed solution by Daniel M\u00f6ller,i tried this way and still got an error<\/p>\n\n<pre><code>sigmaParam = 1\n\ndef apply_sigma(x):\n    return x + np.random.normal(mean=0,scale=sigmaParam,size=(3,32,32))\n\nimgGen = ImageDataGenerator( preprocesing_function=apply_sigma)\ngenerator = imgGen.flow_from_directory('data\/train') # folder that contains \n# only x_train and y_train \n\n\nfrom keras.utils import Sequence\n\nclass SigmaGenerator(Sequence):\n\n    def __init__(self, keras_generator):\n        self.keras_generator = keras_generator\n\n    def __len__(self):\n        return len(self.keras_generator)\n\n    def __getitem__(self,i):\n        return self.keras_generator[i]\n\n    def on_epoch_end(self):\n        sigmaParam += 1\n        self.keras_generator.on_epoch_end()\n\ntraining_generator = SigmaGenerator(generator)\n\nmodel.fit_generator(training_generator,validation_data=(x_test,y_test),\n                steps_per_epoch=x_train.shape[0]\/\/batch_size,epochs=100)\n<\/code><\/pre>\n\n<p>the error i get:<\/p>\n\n<pre><code>process finished with exit code -1073741819 (0xC0000005)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57641857","CreationDate":"2019-08-24T22:08:13.237","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>You can try this:<\/p>\n\n<pre><code>sigmaParam = 1\n\ndef applySigma(x):\n    return x + np.random.normal(mean=0,scale=sigmaParam,size=x.shape)\n<\/code><\/pre>\n\n<p>Create the original generator: <\/p>\n\n<pre><code>imgGen = ImageDataGenerator(..., preprocesing_function=apply_sigma)\ngenerator = imgGen.flow_from_directory(....)\n<\/code><\/pre>\n\n<p>Create a custom generator to wrap the original one, replace its <code>on_epoch_end<\/code> method to update sigmaParam. <\/p>\n\n<pre><code>from keras.utils import Sequence\n\nclass SigmaGenerator(Sequence):\n\n    def __init__(self, keras_generator):\n        self.keras_generator = keras_generator\n\n    def __len__(self):\n        return len(self.keras_generator)\n\n    def __getitem__(self,i):\n        return self.keras_generator[i]\n\n    def on_epoch_end(self):\n        sigmaParam += 1\n        self.keras_generator.on_epoch_end()\n\ntraining_generator = SigmaGenerator(generator)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57630424,"AnswerCount":1,"Tags":"<pytorch><google-colaboratory>","CreationDate":"2019-08-23T17:00:56.460","AcceptedAnswerId":null,"OwnerUserId":8700970.0,"Title":"Unable to load data Google Colab","Body":"<p>I am trying to load data after downloading it through kaggle cli. This exercise is from <a href=\"https:\/\/classroom.udacity.com\/courses\/ud188\/lessons\/c5706f76-0e30-4b48-b74e-c19fafc33a75\/concepts\/80e61337-4d42-4292-9a3b-791f65733d23\" rel=\"nofollow noreferrer\">this course<\/a> on Udacity. <\/p>\n\n<pre><code>!kaggle competitions download -c dogs-vs-cats\n!unzip {content}\/competitions\/dogs-vs-cats\/train.zip -d {content}\/competitions\/dogs-vs-cats\/\n!unzip {content}\/competitions\/dogs-vs-cats\/test1.zip -d {content}\/competitions\/dogs-vs-cats\/\n<\/code><\/pre>\n\n<p>Counting the number of images<\/p>\n\n<pre><code>!ls '{content}\/competitions\/dogs-vs-cats\/train\/' | wc -l\n# 25000\n<\/code><\/pre>\n\n<p>Then I try to load the data<\/p>\n\n<pre><code>data_dir = '{content}\/competitions\/dogs-vs-cats\/train\/'\n\ntransform = transforms.Compose([transforms.Resize(255),\n                               transforms.CenterCrop(224),\n                               transforms.ToTensor()]) # TODO: compose transforms here\ndataset = datasets.ImageFolder(data_dir, transform=transform) # TODO: create the ImageFolder\ndataloader = torch.utils.data.DataLoader(dataset,batch_size=32,shuffle=True) # TODO: use the ImageFolder dataset to create the DataLoader\n<\/code><\/pre>\n\n<p>Error<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nRuntimeError                              Traceback (most recent call last)\n&lt;ipython-input-70-9c49d0bdcdb0&gt; in &lt;module&gt;()\n      4                                transforms.CenterCrop(224),\n      5                                transforms.ToTensor()]) # TODO: compose transforms here\n----&gt; 6 dataset = datasets.ImageFolder(data_dir, transform=transform) # TODO: create the ImageFolder\n      7 dataloader = torch.utils.data.DataLoader(dataset,batch_size=32,shuffle=True) # TODO: use the ImageFolder dataset to create the DataLoader\n\n1 frames\n\/usr\/local\/lib\/python3.6\/dist-packages\/torchvision\/datasets\/folder.py in __init__(self, root, transform, target_transform, loader, is_valid_file)\n    207                                           transform=transform,\n    208                                           target_transform=target_transform,\n--&gt; 209                                           is_valid_file=is_valid_file)\n    210         self.imgs = self.samples\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torchvision\/datasets\/folder.py in __init__(self, root, loader, extensions, transform, target_transform, is_valid_file)\n     95         if len(samples) == 0:\n     96             raise (RuntimeError(\"Found 0 files in subfolders of: \" + self.root + \"\\n\"\n---&gt; 97                                 \"Supported extensions are: \" + \",\".join(extensions)))\n     98 \n     99         self.loader = loader\n\nRuntimeError: Found 0 files in subfolders of: {content}\/competitions\/dogs-vs-cats\/train\/\nSupported extensions are: .jpg,.jpeg,.png,.ppm,.bmp,.pgm,.tif,.tiff,.webp\n<\/code><\/pre>\n","answers":[{"AnswerId":"57640061","CreationDate":"2019-08-24T17:17:04.760","ParentId":null,"OwnerUserId":"10749432","Title":null,"Body":"<p>Looks like you are using f-strings in a wrong way. Simply add <code>f<\/code> here: <\/p>\n\n<pre><code>data_dir = f'{content}\/competitions\/dogs-vs-cats\/train\/'\n<\/code><\/pre>\n\n<p>to include <code>content<\/code> value to the path; without <code>f<\/code> you are just using <code>{content}\/competitions...<\/code> string as a path, as you can see in your error message.  <\/p>\n"}]}
+{"QuestionId":57630830,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><autoencoder>","CreationDate":"2019-08-23T17:40:14.180","AcceptedAnswerId":57631426.0,"OwnerUserId":9715052.0,"Title":"Defining encoder and decoder models from autoencoder: ValueError: Input 0 of layer dense_3 is incompatible with the layer:","Body":"<p>I'm creating an autoencoder using <a href=\"https:\/\/blog.keras.io\/building-autoencoders-in-keras.html\" rel=\"nofollow noreferrer\">this<\/a> tutorial. When I define the encoder and decoder models separately, I get the following error: <\/p>\n\n<pre><code>decoder = tf.keras.Model(encoded_input, decoder_layer(encoded_input))\n\nFile \"...\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py\", line 586, in __call__\n            self.name)\nFile \"...\/site-packages\/tensorflow\/python\/keras\/engine\/input_spec.py\", line 159, in assert_input_compatibility\n            ' but received input with shape ' + str(shape))\nValueError: Input 0 of layer dense_3 is incompatible with the layer: expected axis -1 of input shape to have value 128 but received input with shape [None, 16]\n<\/code><\/pre>\n\n<p>I'm thinking that I need to reshape the output of my layer somewhere but I don't fully understand the reason behind this error. <\/p>\n\n<p>Here is a minimal working example of my code:  <\/p>\n\n<pre><code>def top_k(input, k):\n    return tf.nn.top_k(input, k=k, sorted=True).indices\nencoding_dim = 16\ninput_img = tf.keras.layers.Input(shape=(16, 16, 256), name =\"input\")\nencoded = tf.keras.layers.Dense(encoding_dim, activation='relu')(input_img)\nencoded2 = tf.keras.layers.Dense(256, activation='sigmoid')(encoded)\n# top_k layer\ntopk = tf.keras.layers.Lambda(lambda x: tf.nn.top_k(x, k=int(int(x.shape[-1])\/2),\n                                                sorted=True,\n                                                name=\"topk\").values)(encoded)\ndecoded = tf.keras.layers.Dense(128, activation='relu')(topk) # one dimensional problem\ndecoded2 = tf.keras.layers.Dense(256, activation='sigmoid')(decoded)\nautoencoder = tf.keras.Model(input_img, decoded2)\n\nencoded_input = tf.keras.layers.Input(shape=(encoding_dim,))\n# this is the problem\ndecoder_layer = autoencoder.layers[-1]\nencoder = tf.keras.Model(input_img, encoded)\ndecoder = tf.keras.Model(encoded_input, decoder_layer(encoded_input))\n<\/code><\/pre>\n","answers":[{"AnswerId":"57631426","CreationDate":"2019-08-23T18:33:21.563","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>You have several mistakes in you code. Check out the code snippet below and the comments listing what I changed.<\/p>\n\n<pre><code>def top_k(input, k):\n    return tf.nn.top_k(input, k=k, sorted=True).indices\n\nencoding_dim = 16\ninput_img = tf.keras.layers.Input(shape=(16, 16, 256), name =\"input\")\n# The MNIST images are flattened in the tutorial you are following, so you have to do the same if you want to proceed in the same way.\nflatten = tf.keras.layers.Flatten()(input_img)\nencoded = tf.keras.layers.Dense(encoding_dim, activation='relu')(flatten)\nencoded2 = tf.keras.layers.Dense(256, activation='sigmoid')(encoded)\n# You were using encoded as input, which makes the encoded2 redundant, so I changed the input to be encoded2\ntopk = tf.keras.layers.Lambda(lambda x: tf.nn.top_k(x, k=int(int(x.shape[-1])\/2),\n                                                sorted=True,\n                                                name=\"topk\").values)(encoded2)\ndecoded = tf.keras.layers.Dense(128, activation='relu')(topk) # one dimensional problem\ndecoded2 = tf.keras.layers.Dense(256, activation='sigmoid')(decoded)\n\nautoencoder = tf.keras.Model(input_img, decoded2) \nencoder = tf.keras.Model(input_img, encoded2)\n\n# The actual input to the decoder is the shape of topk as in the autoencoder model\nencoded_input = tf.keras.layers.Input(shape=topk.shape)\n# You model is more complicated than the one in the tutorial, so if you want to recreate the decoder you have to do it layer by layer. This is the first layer\ndecoded1 = autoencoder.layers[-2](encoded_input)\n# This is the second layer\ndecoded2 = autoencoder.layers[-1](decoded1)\n# Finally, the decoder\ndecoder = tf.keras.Model(encoded_input, decoded2)\n<\/code><\/pre>\n\n<p>I assume that it should be pretty clear for you now.<\/p>\n"}]}
+{"QuestionId":57631313,"AnswerCount":1,"Tags":"<python><firebase><tensorflow><keras><tensorflow-lite>","CreationDate":"2019-08-23T18:22:46.903","AcceptedAnswerId":57725850.0,"OwnerUserId":6500085.0,"Title":"I use TFLiteConvert post_training_quantize=True but my model is still too big for being hosted in Firebase ML Kit's Custom servers","Body":"<p>I have written a TensorFlow \/ Keras Super-Resolution GAN. I've converted the resulting trained <code>.h5<\/code> model to a <code>.tflite<\/code> model, using the below code, executed in Google Colab:<\/p>\n\n<pre><code>import tensorflow as tf\nmodel = tf.keras.models.load_model('\/content\/drive\/My Drive\/srgan\/output\/srgan.h5')\nconverter = tf.lite.TFLiteConverter.from_keras_model(model)\nconverter.post_training_quantize=True\ntflite_model = converter.convert()\nopen(\"\/content\/drive\/My Drive\/srgan\/output\/converted_model_quantized.tflite\", \"wb\").write(tflite_model)\n<\/code><\/pre>\n\n<p>As you can see I use <code>converter.post_training_quantize=True<\/code> which was censed to help to output a lighter <code>.tflite<\/code> model than the size of my original <code>.h5<\/code> model, which is 159MB. The resulting <code>.tflite<\/code> model is still 159MB however.<\/p>\n\n<p>It's so big that I can't upload it to Google Firebase Machine Learning Kit's servers in the Google Firebase Console.<\/p>\n\n<p>How could I either:<\/p>\n\n<ul>\n<li><p>decrease the size of the current <code>.tflite<\/code> model which is 159MB (for example using a tool),<\/p><\/li>\n<li><p>or after having deleted the current <code>.tflite<\/code> model which is 159MB, convert the <code>.h5<\/code> model to a lighter <code>.tflite<\/code> model (for example using a tool)?<\/p><\/li>\n<\/ul>\n\n<h1>Related questions<\/h1>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/56620391\/how-to-decrease-size-of-tflite-which-i-converted-from-keras\">How to decrease size of .tflite which I converted from keras<\/a>: no answer, but a comment telling to use <code>converter.post_training_quantize=True<\/code>. However, as I explained it, this solution doesn't seem to work in my case.<\/p>\n","answers":[{"AnswerId":"57725850","CreationDate":"2019-08-30T11:24:49.407","ParentId":null,"OwnerUserId":"9876087","Title":null,"Body":"<p>In general, quantization means, shifting from dtype float32 to uint8. So theoretically our model should reduce by the size of 4. This will be clearly visible in files of greater size.<br><\/p>\n\n<p>Check whether your model has been quantized or not by using the tool \"<a href=\"https:\/\/lutzroeder.github.io\/netron\/\" rel=\"nofollow noreferrer\">https:\/\/lutzroeder.github.io\/netron\/<\/a>\". Here you have to load the model and check the random layers having weight.<a href=\"https:\/\/i.stack.imgur.com\/XMPkh.png\" rel=\"nofollow noreferrer\">The quantized graph contains the weights value in  uint8 format<\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/U6oX6.png\" rel=\"nofollow noreferrer\">In unquantized graph the weights value will be in float32 format.<\/a><br>\n<br>\nOnly setting \"converter.post_training_quantize=True\" is not enough to quantize your model. The other settings include: <br>\nconverter.inference_type=tf.uint8<br>\nconverter.default_ranges_stats=[min_value,max_value]<br>\nconverter.quantized_input_stats={\"name_of_the_input_layer_for_your_model\":[mean,std]}<br><\/p>\n\n<p>Hoping you are dealing with images.<br> \nmin_value=0, max_value=255, mean=128(subjective) and std=128(subjective).<br> name_of_the_input_layer_for_your_model= first name of the graph when you load your model in the above mentioned link or you can get the name of the input layer through the code \"model.input\" will give the output \"tf.Tensor '<strong>input_1<\/strong>:0' shape=(?, 224, 224, 3) dtype=float32\". Here the input_1 is the name of the input layer(NOTE: model must include the graph configuration and the weight.)<\/p>\n"}]}
+{"QuestionId":57631324,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-23T18:23:40.110","AcceptedAnswerId":null,"OwnerUserId":5112616.0,"Title":"What is the function of @tf_export before a tensorflow function definition in Python?","Body":"<p>I'm working with Tensorflow in Python. In a custom written function I found <code>@tf_export()<\/code> before the function definition like below, the function of which I don't understand. Could somebody explain?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>@tf_export(\"signal.ifftshift\")\ndef ifftshift(x, axes=None, name=None):\n<\/code><\/pre>\n","answers":[{"AnswerId":"58979469","CreationDate":"2019-11-21T16:26:11.353","ParentId":null,"OwnerUserId":"4278309","Title":null,"Body":"<p>As I understand, it allows Tensorflow to expose a function or class under a different name. For example, the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/distribute\/Server\" rel=\"nofollow noreferrer\">Server<\/a> class within the distribute module actually lives in the <code>training\/server_lib.py<\/code> file within the repo, but, since it is exported as <code>distribute.Server<\/code>, you can use it like <code>tf.distribute.Server()<\/code>. <\/p>\n\n<pre><code># training\/server_lib.py\n@tf_export(\"distribute.Server\", v1=[\"distribute.Server\", \"train.Server\"])\n@deprecation.deprecated_endpoints(\"train.Server\")\nclass Server(object):\n  ...\n<\/code><\/pre>\n\n<p>It makes it confusing to find the code, but I imagine it's a more flexible way to create these \"logical\" modules.<\/p>\n"}]}
+{"QuestionId":57631559,"AnswerCount":0,"Tags":"<python><tensorflow><pycharm>","CreationDate":"2019-08-23T18:46:35.993","AcceptedAnswerId":null,"OwnerUserId":11968863.0,"Title":"Different result when running the same code in PYCharm and Spyder IDEs","Body":"<p>I'm new with Python and TensorFlow as well. I'm currently writing my master thesis and it's all about machine learning in risk assessment.<\/p>\n\n<p>I have started with the tutorial <a href=\"https:\/\/www.tensorflow.org\/tutorials\/estimators\/linear\" rel=\"nofollow noreferrer\">Build a linear model with Estimators<\/a>.<\/p>\n\n<p>I have both PyCharm and Spyder and I'm in my \"evaluation phase\", trying to figure out which one I like most.<\/p>\n\n<p>The thing is, I tried to run the same code (copy and paste from the link above) on both the IDEs, and I was astonished when I noticed that while Spyder produced almost the same results shown in the tutorial, PyCharm produced a \"Precision\" equal to 0.0 and a \"recall\" equal to 0.0 as well!<\/p>\n\n<p>Do you have any idea? How is this possible? (<strong><em>Edit<\/em><\/strong>: see bottom comments)<\/p>\n\n<p>Here is the code, I've removed the useless lines, but everything is just copy and paste from the link provided.<\/p>\n\n<p>Thank you in advance for the help!<\/p>\n\n<pre><code>from __future__ import absolute_import, division, print_function, unicode_literals\n\nimport tensorflow as tf\n\nimport tensorflow.feature_column as fc\n\nimport os\nimport sys\n\nimport matplotlib.pyplot as plt\nfrom IPython.display import clear_output\n\ntf.enable_eager_execution()\n\nimport census_dataset\nimport census_main\n\ncensus_dataset.download(\"\/tmp\/census_data\/\")\n\ntrain_file = \"\/tmp\/census_data\/adult.data\"\ntest_file = \"\/tmp\/census_data\/adult.test\"\n\nimport pandas\n\ntrain_df = pandas.read_csv(train_file, header = None, names = census_dataset._CSV_COLUMNS)\ntest_df = pandas.read_csv(test_file, header = None, names = census_dataset._CSV_COLUMNS)\n\nimport functools\n\ntrain_inpf = functools.partial(census_dataset.input_fn, train_file, num_epochs=2, shuffle=True, batch_size=64)\ntest_inpf = functools.partial(census_dataset.input_fn, test_file, num_epochs=1, shuffle=False, batch_size=64)\n\nage = fc.numeric_column('age')\n\nclassifier = tf.estimator.LinearClassifier(feature_columns=[age])\nclassifier.train(train_inpf)\n\nresult = classifier.evaluate(test_inpf)\n\nclear_output()\n\nprint(result)\n<\/code><\/pre>\n\n<p>The result obtained in the web page of the tutorial were:<\/p>\n\n<pre><code>{precision': 0.1780822, 'label\/mean': 0.23622628, 'accuracy_baseline': 0.76377374, 'prediction\/mean': 0.23925366, 'auc': 0.6783024, 'loss': 33.404175, 'recall': 0.0033801352, 'average_loss': 0.5231905, 'auc_precision_recall': 0.31137863, 'accuracy': 0.7608869, 'global_step': 1018}\n<\/code><\/pre>\n\n<p>The results from my simulation on Spider are:<\/p>\n\n<pre><code>{'accuracy': 0.7570788, 'accuracy_baseline': 0.76377374, 'auc': 0.67830646, 'auc_precision_recall': 0.3113795, 'average_loss': 0.52337015, 'label\/mean': 0.23622628, 'loss': 33.415646, 'precision': 0.1572327, 'prediction\/mean': 0.25014463, 'recall': 0.00650026, 'global_step': 1018}\n<\/code><\/pre>\n\n<p>The results from my simulation on PyCharm are:<\/p>\n\n<pre><code>{'accuracy': 0.76377374, 'accuracy_baseline': 0.76377374, 'auc': 0.678149, 'auc_precision_recall': 0.31129122, 'average_loss': 0.5285897, 'label\/mean': 0.23622628, 'loss': 33.7489, 'precision': 0.0, 'prediction\/mean': 0.20146999, 'recall': 0.0, 'global_step': 1018}\n<\/code><\/pre>\n\n<p>please note that only \"precision\" and \"recall\" are far from the exact results, the other data provided by PyCharm are similar to the correct ones.<\/p>\n\n<p><strong><em>Edit<\/em><\/strong>: I've made some tests, previously I was using the python console only, I've noticed that also Spyder gives precision and recall equals to 0.0 if I run the entire code in 2 or 3 console calls (i.e. not running the code line by line, but running groups of 3\/4 lines, copying and pasting in a single console line multiple instructions)<\/p>\n\n<p>I've tried to run the code line by line in PyCharm and it does not provide null recall and precision. now the results are:<\/p>\n\n<pre><code>{'accuracy': 0.75062954, 'accuracy_baseline': 0.76377374, 'auc': 0.67835987, 'auc_precision_recall': 0.31139243, 'average_loss': 0.5241894, 'label\/mean': 0.23622628, 'loss': 33.467953, 'precision': 0.17378049, 'prediction\/mean': 0.26093897, 'recall': 0.0148205925, 'global_step': 1018}\n<\/code><\/pre>\n\n<p>Way better. I was wandering, all these differences were arisen just by running multiple instructions in the console? Is always a good idea to just run line by line when using the console?<\/p>\n","answers":[]}
+{"QuestionId":57631705,"AnswerCount":2,"Tags":"<python><pytorch><in-place>","CreationDate":"2019-08-23T19:00:20.420","AcceptedAnswerId":57843685.0,"OwnerUserId":5516760.0,"Title":"RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation","Body":"<p>In a pytorch model training process I get this error:<\/p>\n\n<blockquote>\n  <p>RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.LongTensor [128, 1]] is at version 8; expected version 7 instead. Hint: the backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or anywhere later. Good luck!<\/p>\n<\/blockquote>\n\n<p>with stack trace <\/p>\n\n<pre><code>   sys:1: RuntimeWarning: Traceback of forward call that caused the error:\n  File \"\/home\/arash\/anaconda2\/envs\/mzh27\/lib\/python2.7\/runpy.py\", line 174, in _run_module_as_main\n    \"__main__\", fname, loader, pkg_name)\n  File \"\/home\/arash\/anaconda2\/envs\/mzh27\/lib\/python2.7\/runpy.py\", line 72, in _run_code\n    exec code in run_globals\n  File \"\/home\/arash\/anaconda2\/envs\/mzh27\/lib\/python2.7\/site-packages\/ipykernel_launcher.py\", line 16, in &lt;module&gt;\n    app.launch_new_instance()\n  File \"\/home\/arash\/anaconda2\/envs\/mzh27\/lib\/python2.7\/site-packages\/traitlets\/config\/application.py\", line 658, in launch_instance\n    app.start()\n  File \"\/home\/arash\/anaconda2\/envs\/mzh27\/lib\/python2.7\/site-packages\/ipykernel\/kernelapp.py\", line 499, in start\n    self.io_loop.start()\n  File \"\/home\/arash\/anaconda2\/envs\/mzh27\/lib\/python2.7\/site-packages\/tornado\/ioloop.py\", line 1073, in start\n    handler_func(fd_obj, events)\n  File \"\/home\/arash\/anaconda2\/envs\/mzh27\/lib\/python2.7\/site-packages\/tornado\/stack_context.py\", line 300, in null_wrapper\n    return fn(*args, **kwargs)\n  File \"\/home\/arash\/.local\/lib\/python2.7\/site-packages\/zmq\/eventloop\/zmqstream.py\", line 456, in _handle_events\n    self._handle_recv()\n  File \"\/home\/arash\/.local\/lib\/python2.7\/site-packages\/zmq\/eventloop\/zmqstream.py\", line 486, in _handle_recv\n    self._run_callback(callback, msg)\n  File \"\/home\/arash\/.local\/lib\/python2.7\/site-packages\/zmq\/eventloop\/zmqstream.py\", line 438, in _run_callback\n    callback(*args, **kwargs)\n  File \"\/home\/arash\/anaconda2\/envs\/mzh27\/lib\/python2.7\/site-packages\/tornado\/stack_context.py\", line 300, in null_wrapper\n    return fn(*args, **kwargs)\n  File \"\/home\/arash\/anaconda2\/envs\/mzh27\/lib\/python2.7\/site-packages\/ipykernel\/kernelbase.py\", line 283, in dispatcher\n    return self.dispatch_shell(stream, msg)\n  File \"\/home\/arash\/anaconda2\/envs\/mzh27\/lib\/python2.7\/site-packages\/ipykernel\/kernelbase.py\", line 233, in dispatch_shell\n    handler(stream, idents, msg)\n  File \"\/home\/arash\/anaconda2\/envs\/mzh27\/lib\/python2.7\/site-packages\/ipykernel\/kernelbase.py\", line 399, in execute_request\n    user_expressions, allow_stdin)\n  File \"\/home\/arash\/anaconda2\/envs\/mzh27\/lib\/python2.7\/site-packages\/ipykernel\/ipkernel.py\", line 208, in do_execute\n    res = shell.run_cell(code, store_history=store_history, silent=silent)\n  File \"\/home\/arash\/anaconda2\/envs\/mzh27\/lib\/python2.7\/site-packages\/ipykernel\/zmqshell.py\", line 537, in run_cell\n    return super(ZMQInteractiveShell, self).run_cell(*args, **kwargs)\n  File \"\/home\/arash\/anaconda2\/envs\/mzh27\/lib\/python2.7\/site-packages\/IPython\/core\/interactiveshell.py\", line 2714, in run_cell\n    interactivity=interactivity, compiler=compiler, result=result)\n  File \"\/home\/arash\/anaconda2\/envs\/mzh27\/lib\/python2.7\/site-packages\/IPython\/core\/interactiveshell.py\", line 2818, in run_ast_nodes\n    if self.run_code(code, result):\n  File \"\/home\/arash\/anaconda2\/envs\/mzh27\/lib\/python2.7\/site-packages\/IPython\/core\/interactiveshell.py\", line 2878, in run_code\n    exec(code_obj, self.user_global_ns, self.user_ns)\n  File \"&lt;ipython-input-71-a5b255596e11&gt;\", line 33, in &lt;module&gt;\n    sampled_captions, sampled_log_probs=predict_captions(out_enc,hid_enc,enc_pp,'sample')\n  File \"&lt;ipython-input-70-a6ea511f0678&gt;\", line 18, in predict_captions\n    out_dec, hid_dec, word_logits = dec.forward(r, last_enc, img_features)\n  File \"&lt;ipython-input-21-0601dad4805f&gt;\", line 21, in forward\n    emb = self.embedding(input)\n  File \"\/home\/arash\/.local\/lib\/python2.7\/site-packages\/torch\/nn\/modules\/module.py\", line 493, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/home\/arash\/.local\/lib\/python2.7\/site-packages\/torch\/nn\/modules\/sparse.py\", line 117, in forward\n    self.norm_type, self.scale_grad_by_freq, self.sparse)\n  File \"\/home\/arash\/.local\/lib\/python2.7\/site-packages\/torch\/nn\/functional.py\", line 1506, in embedding\n    return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)\n\n---------------------------------------------------------------------------\nRuntimeError                              Traceback (most recent call last)\n&lt;ipython-input-71-a5b255596e11&gt; in &lt;module&gt;()\n     54                  torch.stack(sampled_log_probs).reshape(-1)).mean()\n     55         torch.autograd.set_detect_anomaly(True)\n---&gt; 56         loss.backward()\n     57         clip_grad_value_(enc_optim.param_groups[0]['params'], 5.0)\n     58         clip_grad_value_(dec_optim.param_groups[0]['params'], 5.0)\n\n\/home\/arash\/.local\/lib\/python2.7\/site-packages\/torch\/tensor.pyc in backward(self, gradient, retain_graph, create_graph)\n    105                 products. Defaults to ``False``.\n    106         \"\"\"\n--&gt; 107         torch.autograd.backward(self, gradient, retain_graph, create_graph)\n    108 \n    109     def register_hook(self, hook):\n\n\/home\/arash\/.local\/lib\/python2.7\/site-packages\/torch\/autograd\/__init__.pyc in backward(tensors, grad_tensors, retain_graph, create_graph, grad_variables)\n     91     Variable._execution_engine.run_backward(\n     92         tensors, grad_tensors, retain_graph, create_graph,\n---&gt; 93         allow_unreachable=True)  # allow_unreachable flag\n     94 \n     95\n<\/code><\/pre>\n\n<p>But I'm unable to locate in-place operation that causes the error.\nHere is my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>for epoch in xrange(0, 13):\n    print (\"Starting New Epoch: %d\" % epoch)\n    rewards = {\n        'sample_cider': [],\n        'sample_context': [],\n        'sample_reward': [],  # actual reward, controlled by beta\n        'greedy_cider': [],\n        'greedy_context': [],\n        'greedy_reward': []\n    }\n\n    order = np.arange(enc_padded_text.shape[0])\n    np.random.shuffle(order)\n    enc_padded_text = enc_padded_text[order]\n    input_text=[input_text[i] for i in order]\n    dec_text_tensor.data = dec_text_tensor.data[order]\n\n    for i in xrange(num_batches):\n        s = i * BATCH_SIZE\n        e = (i+1) * BATCH_SIZE\n\n        _, enc_pp, dec_pp, enc_lengths = make_packpadded_s2s(s, e, enc_padded_text, dec_text_tensor)\n\n        enc.zero_grad()\n        dec.zero_grad()\n\n        hid = enc.initHidden(BATCH_SIZE)\n\n        out_enc, hid_enc = enc.forward(enc_pp, hid, enc_lengths)\n\n        hid_enc = torch.cat([hid_enc[0,:, :], hid_enc[1,:,:]], dim=1).unsqueeze(0)\n        gt_dict = dict(zip(_,input_text[s:e]))\n        sampled_captions, sampled_log_probs=predict_captions(out_enc,hid_enc,enc_pp,'sample')\n        sampled_dict = dict(zip(_, sampled_captions))\n        with torch.no_grad():\n            greedy_captions = predict_captions(out_enc,hid_enc,enc_pp, 'greedy')\n            greedy_dict = dict(zip(_, greedy_captions))\n\n\n        sample_cider_score, sample_context_score, sample_reward = get_scores(\n            dec_pp[:,1:], sampled_captions, gt_dict, sampled_dict)\n        greedy_cider_score, greedy_context_score, greedy_reward = get_scores(\n            dec_pp[:,1:], greedy_captions, gt_dict, greedy_dict)\n\n        # self-critical: score from sampling - score from test time\n        advantages = torch.Tensor((sample_cider_score - greedy_cider_score).reshape(-1))\n\n        # normalize advantages\n        advantages = ((advantages - advantages.mean()) \/\n                      advantages.std() + 1e-9)\n        if cuda:\n            advantages = advantages.cuda()\n        loss = -(advantages *\n                 torch.stack(sampled_log_probs).reshape(-1)).mean()\n        torch.autograd.set_detect_anomaly(True)\n        loss.backward()\n        clip_grad_value_(enc_optim.param_groups[0]['params'], 5.0)\n        clip_grad_value_(dec_optim.param_groups[0]['params'], 5.0)\n        enc_optim.step()\n        dec_optim.step()\n\n        rewards['sample_cider'].extend(sample_cider_score)\n        rewards['sample_context'].extend(sample_context_score)\n        rewards['sample_reward'].extend(sample_reward)\n        rewards['greedy_cider'].extend(greedy_cider_score)\n        rewards['greedy_context'].extend(greedy_context_score)\n        rewards['greedy_reward'].extend(greedy_reward)\n\n        if (b + 1) % 100 == 0:\n            print('\\t[Batch {} running metrics] - R train {:.2f} - R train (greedy): {:.2f}'.format(\n                b + 1, np.mean(rewards['sample_reward']), np.mean(rewards['greedy_reward'])))\n<\/code><\/pre>\n\n<p>predict_captions function:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def predict_captions(img_features,hid_enc,enc_pp, mode='sample', constrain=False,L=22):\n    dec_tensor = torch.ones((enc_pp.shape[0]), L+1, dtype=torch.long) * Toks.SOS\n    global cuda\n    if cuda:\n        dec_tensor = dec_tensor.cuda(device=device)\n    last_enc = hid_enc\n    if mode == 'beam_search':\n        return self.beam_search(img_features, state, lstm_states)\n\n    predictions = []\n    log_probs = []\n\n    # this should store the index of the first occurrence of &lt;EOS&gt;\n    # for each sample in the batch\n    EOS_tracker = np.full(img_features.shape[0], None)\n    for i in range(L):\n        r=dec_tensor[:,i].unsqueeze(1)\n        out_dec, hid_dec, word_logits = dec.forward(r, last_enc, img_features)\n        out_dec[:, 0, Toks.UNK] = -np.inf # ignore unknowns\n        l=out_dec[:,0]\n        chosen = torch.argmax(l,dim=1)\n        dec_tensor[:, i+1] = chosen\n        last_enc = hid_dec\n        # decoding stuff\n        probs = F.softmax(word_logits, dim=2)\n        probs=probs.reshape(128,20004)\n\n        if constrain:\n            # enforce constraint that the same word can't be predicted\n            # twice in a row. zero-out the probability of previous words\n            for p, prev_idx in zip(probs, state['prev_word_indeces']):\n                p[prev_idx] = 0\n\n        if mode == 'sample':\n            idxs = torch.multinomial(probs, 1)\n        else:\n            idxs = torch.argmax(probs, dim=1)\n        if cuda:\n            idxs = idxs.cpu()\n        words = [dec_idx_to_word[index] for index in idxs] \n        predictions.append(np.array(words).reshape(-1))\n\n        # get the respective log probability of chosen word\n        # for each sample in the batch\n        log_probs.append([lp[i] for (lp, i)\n                         in zip(torch.log(probs), idxs)])\n\n        # inefficient but this should be fast enough anyway... ? :(\n        eos_idxs = (np.array(words)==dec_idx_to_word[2]).nonzero()[0]\n        for idx in eos_idxs:\n            if EOS_tracker[idx] is None:\n                EOS_tracker[idx] = i + 1\n\n        # finish loop if they're all done\n        if len(EOS_tracker[EOS_tracker == None])==0:\n            break\n\n\n    # build the actual sentences, up until the first occurrence of &lt;EOS&gt;\n    captions = [\n        [' '.join(w[:eos_idx])] for (w, eos_idx) in\n        zip(np.array(predictions).T, EOS_tracker)\n    ]\n    print captions\n    # do this only when training. not needed otherwise.\n    if mode == 'sample':\n        log_probs = [lp[:eos_idx].sum() for (lp, eos_idx) in zip(np.array(log_probs).T, EOS_tracker)]\n        return captions, log_probs\n\n    return captions\n<\/code><\/pre>\n\n<p>models:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class Encoder_s2s(nn.Module):\n    def __init__(self, input_size, hidden_size):\n        super(Encoder_s2s, self).__init__()\n        assert hidden_size % 2 == 0\n\n        self.hidden_size = hidden_size\n        self.input_size = input_size\n\n        self.hidden_init_tensor = torch.zeros(2, 1, self.hidden_size\/2, requires_grad=True)\n        nn.init.normal_(self.hidden_init_tensor, mean=0, std=0.05)\n        self.hidden_init = torch.nn.Parameter(self.hidden_init_tensor, requires_grad=True)\n\n        self.embedding = nn.Embedding(input_size, hidden_size)\n        self.emb_drop = nn.Dropout(0.2)\n        self.gru = nn.GRU(hidden_size, hidden_size\/2, batch_first=True, bidirectional=True)\n        self.gru_out_drop = nn.Dropout(0.2)\n        self.gru_hid_drop = nn.Dropout(0.3)\n\n    def forward(self, input, hidden, lengths):\n        emb = self.emb_drop(self.embedding(input))\n        #emb = embedded_dropout(self.embedding, input, dropout=0.2 if self.training else 0)\n        pp = torch.nn.utils.rnn.pack_padded_sequence(emb, lengths, batch_first=True)\n        out, hidden = self.gru(pp, hidden)\n        out = torch.nn.utils.rnn.pad_packed_sequence(out, batch_first=True)[0]\n        out = self.gru_out_drop(out)\n        hidden = self.gru_hid_drop(hidden)\n        return out, hidden\n\n    def initHidden(self, bs):\n        return self.hidden_init.expand(2, bs, self.hidden_size\/2).contiguous()\n\nclass DecoderAttn(nn.Module):\n    def __init__(self, input_size, hidden_size, output_size, out_bias):\n        super(DecoderAttn, self).__init__()\n        self.hidden_size = hidden_size\n        self.input_size = input_size\n\n        self.embedding = nn.Embedding(input_size, hidden_size)\n        self.emb_drop = nn.Dropout(0.2)\n        self.gru = nn.GRU(hidden_size, hidden_size, batch_first=True)\n        self.gru_drop = nn.Dropout(0.2)\n        self.mlp = nn.Linear(hidden_size*2, output_size)\n        if out_bias is not None:\n            out_bias_tensor = torch.tensor(out_bias, requires_grad=False)\n            self.mlp.bias.data[:] = out_bias_tensor\n        self.logsoftmax = nn.LogSoftmax(dim=2)\n\n        self.att_mlp = nn.Linear(hidden_size, hidden_size, bias=False)\n        self.attn_softmax = nn.Softmax(dim=2)\n\n    def forward(self, input, hidden, encoder_outs):\n        emb = self.embedding(input)\n        emb=self.emb_drop(emb)\n        out, hidden = self.gru(emb, hidden)\n\n        out_proj = self.att_mlp(out)\n        enc_out_perm = encoder_outs.permute(0, 2, 1)\n        e_exp = torch.bmm(out_proj, enc_out_perm)\n        attn = self.attn_softmax(e_exp)\n\n        ctx = torch.bmm(attn, encoder_outs)\n\n        full_ctx = torch.cat([self.gru_drop(out), ctx], dim=2)\n\n        out = self.mlp(full_ctx)\n        out1 = self.logsoftmax(out)\n        return out1, hidden, out\n<\/code><\/pre>\n","answers":[{"AnswerId":"57743142","CreationDate":"2019-09-01T04:13:37.013","ParentId":null,"OwnerUserId":"6172218","Title":null,"Body":"<p>Not sure if you had already checked these, but could help : <\/p>\n\n<ol>\n<li><p><a href=\"https:\/\/discuss.pytorch.org\/t\/encounter-the-runtimeerror-one-of-the-variables-needed-for-gradient-computation-has-been-modified-by-an-inplace-operation\/836\/15\" rel=\"nofollow noreferrer\">https:\/\/discuss.pytorch.org\/t\/encounter-the-runtimeerror-one-of-the-variables-needed-for-gradient-computation-has-been-modified-by-an-inplace-operation\/836\/15<\/a><\/p><\/li>\n<li><p><a href=\"http:\/\/www.yongfengli.tk\/2018\/04\/13\/inplace-operation-in-pytorch.html\" rel=\"nofollow noreferrer\">http:\/\/www.yongfengli.tk\/2018\/04\/13\/inplace-operation-in-pytorch.html<\/a><\/p><\/li>\n<\/ol>\n"},{"AnswerId":"57843685","CreationDate":"2019-09-08T16:22:35.297","ParentId":null,"OwnerUserId":"5268235","Title":null,"Body":"<p>A tensor matching this description <code>torch.cuda.LongTensor [128, 1]<\/code>, should narrow down your search. <\/p>\n\n<p>A quick google search revealed that, <code>LongTensors<\/code> are most commonly  returned by <code>min<\/code> , <code>max<\/code>, <code>sort<\/code>. so the lines <\/p>\n\n<pre><code>l=out_dec[:,0]\nchosen = torch.argmax(l,dim=1)\ndec_tensor[:, i+1] = chosen\n<\/code><\/pre>\n\n<p>Most probably the line <code>dec_tensor[:, i+1] = chosen<\/code> seems problematic.<\/p>\n"}]}
+{"QuestionId":57631868,"AnswerCount":1,"Tags":"<tensorflow><computer-vision><image-segmentation><sharding><tfrecord>","CreationDate":"2019-08-23T19:15:32.867","AcceptedAnswerId":null,"OwnerUserId":5429010.0,"Title":"Create round-robin sharding while generating sharded tfrecords","Body":"<p>I am new to tensorflow and I am working on image segmentation problem in tensorflow 1.14. I have a huge dataset and generating tfrecords is very slow, when I try to generate one big tfrecord file. So, I would like to create 'n' shards of tfrecords. I could not find a way to do it online. Say I have 600 images and 600 masks. I want to generate 6 shards of tfrecords, with 100 images and 100 masks each in round robin fashion. A high level \/pseudo-code of what I want is as follows -<\/p>\n\n<pre><code>sharded_tf_record_writer:\ncreate n TFRecordWriter\n----&gt; for each_item in n TFRecordWriter\n      -----&gt; write_example in round-robin fashion\n<\/code><\/pre>\n\n<p>I did search online and could not find relevant answer. I do not want to use apache beam for sharding. I appreciate any idea\/help\/guidance to achieve this.<\/p>\n","answers":[{"AnswerId":"57682142","CreationDate":"2019-08-27T20:55:02.057","ParentId":null,"OwnerUserId":"5429010","Title":null,"Body":"<p>I had asked the same question in one of the issues of tensorflow datasets and the user - Conchylicultor responded this - <\/p>\n\n<blockquote>\n  <p>Writing is done by _TFRecordWriter. Tfds will automatically compute the required number of shards and distribute examples across shards, However each shard is written sequentially.\n  You do not have control over the number of shards, it is also automatically computed.<\/p>\n  \n  <p>However, the fact that examples are distributed between shards do not make the writing faster as examples are not pre-processed in parallel. If you want parallelism, then you'll have to use Apache Beam which allow to scale even to huge datasets<\/p>\n<\/blockquote>\n\n<p>The link to the tensorflow\/datasets issue is - <a href=\"https:\/\/github.com\/tensorflow\/datasets\/issues\/676\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/datasets\/issues\/676<\/a><\/p>\n\n<p>This might help. <\/p>\n"}]}
+{"QuestionId":57632084,"AnswerCount":1,"Tags":"<python><nlp><pytorch><lstm>","CreationDate":"2019-08-23T19:35:13.143","AcceptedAnswerId":57863898.0,"OwnerUserId":7740703.0,"Title":"What is the correct way to use OHE lookup table for a pytorch RNN?","Body":"<p>I am currently trying to build a LSTM RNN using pytorch. One input vector is represented as an array of 50 integers corresponding to a sequence of  at most 50 tokens with padding where each integer corresponds to an element from my vocabulary and the index of the 1 in the OHE vector. I want to have an embedding layer that just uses a lookup table to One-hot encode the integer-- kind of like tensorflow's OHE layer.  <\/p>\n\n<p>Something like this \"kind of\" works <\/p>\n\n<pre><code>import torch\nimport numpy as np\nimport torch.nn as nn \n\n# vocab_size is the number of words in your train, val and test set\n# vector_size is the dimension of the word vectors you are using\nvocab_size, vector_size = 5, 5\nembed = nn.Embedding(vocab_size, vector_size)\n\n# intialize the word vectors, pretrained_weights is a \n# numpy array of size (vocab_size, vector_size) and \n# pretrained_weights[i] retrieves the word vector of\n# i-th word in the vocabulary\npretrained_weights = np.zeros((vocab_size, vector_size))\nnp.fill_diagonal(pretrained_weights, 1)\ntmp =torch.from_numpy(pretrained_weights)\nembed.weight = nn.Parameter(tmp,requires_grad=False )\n\n# Then turn the word index into actual word vector\nvocab = {\"some\": 0, \"words\": 1}\nword_indexes = torch.from_numpy(np.array([vocab[w] for w in [\"some\", \"words\"]]))\nword_vectors = embed(word_indexes)\nword_vectors.data.numpy()\n\n\n&gt;&gt;&gt;output\narray([[1., 0., 0., 0., 0.],\n       [0., 1., 0., 0., 0.]])\n<\/code><\/pre>\n\n<p>but it is very hacky, and doesn't play nicely with batches of input vectors.<\/p>\n\n<p>What is the correct way to declare a OHE embedding layer at the begining of an RNN?<\/p>\n","answers":[{"AnswerId":"57863898","CreationDate":"2019-09-10T04:03:23.177","ParentId":null,"OwnerUserId":"3518108","Title":null,"Body":"<p>The method you are looking for is <code>torch.nn.functional.one_hot<\/code>. It was added to PyTorch later, so it's hard to find when you Google for it. <\/p>\n\n<pre><code> &gt;&gt;&gt; import torch\n&gt;&gt;&gt; x = torch.arange(0,5)\n&gt;&gt;&gt; x\ntensor([0, 1, 2, 3, 4])\n&gt;&gt;&gt; torch.nn.functional.one_hot(x)\ntensor([[1, 0, 0, 0, 0],\n      [0, 1, 0, 0, 0],\n      [0, 0, 1, 0, 0],\n      [0, 0, 0, 1, 0],\n      [0, 0, 0, 0, 1]])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57632831,"AnswerCount":1,"Tags":"<python><tensorflow><traceback>","CreationDate":"2019-08-23T20:46:22.883","AcceptedAnswerId":57647784.0,"OwnerUserId":4861207.0,"Title":"Creating an ai chatbot, but getting a traceback error","Body":"<p>I'm trying to create an ai chatbox in python. I tried following this tutorial: <a href=\"https:\/\/techwithtim.net\/tutorials\/ai-chatbot\/part-1\/\" rel=\"nofollow noreferrer\">https:\/\/techwithtim.net\/tutorials\/ai-chatbot\/part-1\/<\/a> but I'm getting a lot of errors of deprecations and getting some Traceback error.\nHere's the code:<\/p>\n\n<pre><code>import json\nimport random\nimport tensorflow\nimport tflearn\nimport numpy\nimport sys\nimport pickle\nimport nltk\nfrom nltk.stem.lancaster import LancasterStemmer\nstemmer = LancasterStemmer()\nnltk.download('punkt')\n\n\nwith open(\"trainingData.json\") as file:\n    data = json.load(file)\n\ntry:\n    with open(\"data.pickle\", \"rb\") as f:\n        words, labels, training, output = pickle.load(f)\nexcept:\n    words = []\n    labels = []\n    docs_x = []\n    docs_y = []\n\n    for intent in data[\"intents\"]:\n        for pattern in intent[\"patterns\"]:\n            wrds = nltk.word_tokenize(pattern)\n            words.extend(wrds)\n            docs_x.append(wrds)\n            docs_y.append(intent[\"tag\"])\n\n        if intent[\"tag\"] not in labels:\n            labels.append(intent[\"tag\"])\n\n    words = [stemmer.stem(w.lower()) for w in words if w != \"?\"]\n    words = sorted(list(set(words)))\n\n    labels = sorted(labels)\n\n    training = []\n    output = []\n\n    out_empty = [0 for _ in range(len(labels))]\n\n    for x, doc in enumerate(docs_x):\n        bag = []\n\n        wrds = [stemmer.stem(w.lower()) for w in doc]\n\n        for w in words:\n            if w in wrds:\n                bag.append(1)\n            else:\n                bag.append(0)\n\n        output_row = out_empty[:]\n        output_row[labels.index(docs_y[x])] = 1\n\n        training.append(bag)\n        output.append(output_row)\n\n    training = numpy.array(training)\n    output = numpy.array(output)\n\n    with open(\"data.pickle\", \"wb\") as f:\n        pickle.dump((words, labels, training, output), f)\n\ntensorflow.reset_default_graph()\n\nnet = tflearn.input_data(shape=[None, len(training[0])])\nnet = tflearn.fully_connected(net, 8)\nnet = tflearn.fully_connected(net, 8)\nnet = tflearn.fully_connected(net, len(output[0]), activation=\"softmax\")\nnet = tflearn.regression(net)\n\nmodel = tflearn.DNN(net)\n\ntry:\n    model.load(\"model.tflearn\")\nexcept:\n    model.fit(training, output, n_epoch=1000, batch_size=8, show_metric=True)\n    model.save(\"model.tflearn\")\n\n\ndef bag_of_words(s, words):\n    bag = [0 for _ in range(len(words))]\n\n    s_words = nltk.word_tokenize(s)\n    s_words = [stemmer.stem(word.lower()) for word in s_words]\n\n    for se in s_words:\n        for i, w in enumerate(words):\n            if w == se:\n                bag[i] = 1\n\n    return numpy.array(bag)\n\n\ndef chat():\n    print(\"Start talking with the bot (type quit to stop)!\")\n    while True:\n        inp = input(\"You: \")\n        if inp.lower() == \"quit\":\n            break\n\n        results = model.predict([bag_of_words(inp, words)])\n        results_index = numpy.argmax(results)\n        tag = labels[results_index]\n\n        for tg in data[\"intents\"]:\n            if tg['tag'] == tag:\n                responses = tg['responses']\n\n        print(random.choice(responses))\n\nchat()\n<\/code><\/pre>\n\n<p>Here are the errors I'm getting. How can I fix the deprecation errors, the traceback error?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/H3hJx.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/H3hJx.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Here's the text of the error:<\/p>\n\n<pre><code>Run id: VOB3W4\nLog directory: \/tmp\/tflearn_logs\/\n---------------------------------\nTraining samples: 20\nValidation samples: 0\n--\n--\nTraceback (most recent call last):\n  File \"script.py\", line 91, in &lt;module&gt;\n    model.fit(training, output, n_epoch=1000, batch_size=8, show_metric=True)\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tflearn\/models\/dnn.py\", line 216, in fit\n    callbacks=callbacks)\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tflearn\/helpers\/trainer.py\", line 339, in fit\n    show_metric)\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tflearn\/helpers\/trainer.py\", line 816, in _train\n    tflearn.is_training(True, session=self.session)\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tflearn\/config.py\", line 95, in is_training\n    tf.get_collection('is_training_ops')[0].eval(session=session)\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 731, in eval\n    return _eval_using_default_session(self, feed_dict, self.graph, session)\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 5579, in _eval_using_default_session\n    return session.run(tensors, feed_dict)\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 950, in run\n    run_metadata_ptr)\n  File \"\/usr\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1096, in _run\n    raise RuntimeError('Attempted to use a closed Session.')\nRuntimeError: Attempted to use a closed Session.\n<\/code><\/pre>\n","answers":[{"AnswerId":"57647784","CreationDate":"2019-08-25T16:28:19.977","ParentId":null,"OwnerUserId":"1832058","Title":null,"Body":"<p>At start file <code>\"model.tflearn\"<\/code> doesn't exist and <code>try\/except<\/code> should catch error when code try to load this file and run <code>fit()<\/code> and <code>save()<\/code> <\/p>\n\n<pre><code>try:\n    model.load(\"model.tflearn\")\nexcept:\n    model.fit(training, output, n_epoch=1000, batch_size=8, show_metric=True)\n    model.save(\"model.tflearn\")\n<\/code><\/pre>\n\n<p>but it seems this error closes <code>tf.session()<\/code> so it can't run <code>fit()<\/code>correctly.<\/p>\n\n<p>If you remove <code>try\/except<\/code> with <code>load()<\/code> and keep only <code>fit()<\/code> and <code>save()<\/code> then it has no problem to create model and save it in file.<\/p>\n\n<pre><code>model.fit(training, output, n_epoch=1000, batch_size=8, show_metric=True)\nmodel.save(\"model.tflearn\")\n<\/code><\/pre>\n\n<p>After creating file <code>\"model.ftlearn\"<\/code> you can use again <code>try\/except<\/code> with <code>load()<\/code> and it should work if you don't delete file with model.<\/p>\n\n<hr>\n\n<p>Better solution should check if file exists - but it saves data in few files <code>\"model.tflearn.index\"<\/code>, <code>\"model.tflearn.meta\"<\/code> and <code>\"model.tflearn.data-00000-of-00001\"<\/code> so it should check one of this file instead of <code>\"model.tflearn\"<\/code><\/p>\n\n<p>Use <\/p>\n\n<pre><code>import os\n\nif os.path.exists(\"model.tflearn.meta\"):\n    model.load(\"model.tflearn\")\nelse:\n    model.fit(training, output, n_epoch=1000, batch_size=8, show_metric=True)\n    model.save(\"model.tflearn\")\n<\/code><\/pre>\n\n<p>instead of <\/p>\n\n<pre><code>try:\n    model.load(\"model.tflearn\")\nexcept:\n    model.fit(training, output, n_epoch=1000, batch_size=8, show_metric=True)\n    model.save(\"model.tflearn\")\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>EDIT:<\/strong> It seems this problem exists at least 2 years: <a href=\"https:\/\/stackoverflow.com\/questions\/45312698\/runtimeerror-attempted-to-use-a-closed-session-in-tflearn\">RuntimeError: Attempted to use a closed Session in tflearn<\/a><\/p>\n"}]}
+{"QuestionId":57633270,"AnswerCount":0,"Tags":"<tensorflow><keras><conv-neural-network><amazon-sagemaker>","CreationDate":"2019-08-23T21:35:49.583","AcceptedAnswerId":null,"OwnerUserId":9347322.0,"Title":"How to deploy\/package a custom Keras algorithm as a tensorflow model on Sagemaker, so that it can ingest jpeg images from an S3 bucket","Body":"<p>I have a custom Keras cnn that I am trying to package as a tensorflow model and deploy on Sagemaker, using batch transform on jpeg images in an S3 bucket. <\/p>\n\n<p>I followed this tutorial with my own model--\n<a href=\"https:\/\/aws.amazon.com\/blogs\/machine-learning\/deploy-trained-keras-or-tensorflow-models-using-amazon-sagemaker\/\" rel=\"nofollow noreferrer\">https:\/\/aws.amazon.com\/blogs\/machine-learning\/deploy-trained-keras-or-tensorflow-models-using-amazon-sagemaker\/<\/a><\/p>\n\n<p>However I'm wondering how to use a batch transformer instead. I created a transformer object using the model that I loaded from the tutorial--<\/p>\n\n<p>import sagemaker as sage\nsess = sage.Session()<\/p>\n\n<p>transformer = sagemaker.transformer.Transformer(\n    base_transform_job_name='Batch-Transform',\n    model_name='sagemaker-tensorflow-2019-08-22-20-24-12-590',  # insert model name here\n    instance_count=1,\n    instance_type='ml.p2.xlarge',\n    output_path=batch_output,\n    sagemaker_session=sess\n)<\/p>\n\n<p>and tried all different versions of calling the .transform method, only to have it error out--<\/p>\n\n<p>transformer.transform(data=batch_input, content_type='application\/x-image')\ntransformer.wait()<\/p>\n","answers":[]}
+{"QuestionId":57633584,"AnswerCount":1,"Tags":"<python><tensorflow><keras><neural-network>","CreationDate":"2019-08-23T22:19:14.150","AcceptedAnswerId":57637711.0,"OwnerUserId":11949810.0,"Title":"Manipulating output of neural network","Body":"<p>I have a neural network that takes an input of (m, 2, 3, 96, 96) and outputs (m, 2, 128). I'm trying to turn that output into (m, 1, 128) by subtracting output[m][0][0] - output[m][0][1] and then into (m, 1) by feeding the 1x128 outputs into a dense layer<\/p>\n\n<p>I've tried Lambda and keras.backend.Subtract layers in the network and wrapper<\/p>\n\n<pre><code>def faceRecoModel(input_shape):\n    \"\"\"\n    Implementation of the Inception model used for FaceNet\n\n    Arguments:\n    input_shape -- shape of the images of the dataset\n\n    Returns:\n    model -- a Model() instance in Keras\n    \"\"\"\n\n    # Define the input as a tensor with shape input_shape\n    X_input = Input(input_shape)\n\n    # Zero-Padding\n    X = ZeroPadding2D((3, 3))(X_input)\n\n    # First Block\n    X = Conv2D(64, (7, 7), strides=(2, 2), name='conv1')(X)\n    X = BatchNormalization(axis=1, name='bn1')(X)\n    X = Activation('relu')(X)\n\n    # Zero-Padding + MAXPOOL\n    X = ZeroPadding2D((1, 1))(X)\n    X = MaxPooling2D((3, 3), strides=2)(X)\n\n    # Second Block\n    X = Conv2D(64, (1, 1), strides=(1, 1), name='conv2')(X)\n    X = BatchNormalization(axis=1, epsilon=0.00001, name='bn2')(X)\n    X = Activation('relu')(X)\n\n    # Zero-Padding + MAXPOOL\n    X = ZeroPadding2D((1, 1))(X)\n\n    # Second Block\n    X = Conv2D(192, (3, 3), strides=(1, 1), name='conv3')(X)\n    X = BatchNormalization(axis=1, epsilon=0.00001, name='bn3')(X)\n    X = Activation('relu')(X)\n\n    # Zero-Padding + MAXPOOL\n    X = ZeroPadding2D((1, 1))(X)\n    X = MaxPooling2D(pool_size=3, strides=2)(X)\n\n    # Inception 1: a\/b\/c\n    X = inception_block_1a(X)\n    X = inception_block_1b(X)\n    X = inception_block_1c(X)\n\n    # Inception 2: a\/b\n    X = inception_block_2a(X)\n    X = inception_block_2b(X)\n\n    # Inception 3: a\/b\n    X = inception_block_3a(X)\n    X = inception_block_3b(X)\n\n    # Top layer\n    X = AveragePooling2D(pool_size=(3, 3), strides=(1, 1), data_format='channels_first')(X)\n    X = Flatten()(X)\n    X = Dense(128, name='dense_layer')(X)\n\n    # L2 normalization\n    X = Lambda(lambda x: K.l2_normalize(x, axis=1))(X)\n\n    # Create model instance\n    model = Model(inputs=X_input, outputs=X, name='FaceRecoModel')\n\n    return model\n\n\n\n# now this is the wrapper I mentioned\nmodel = faceRecoModel((3, 96, 96))\ni = Input((2, 3, 96, 96))\no = TimeDistributed(model)(i)\nmodel = Model(i, o)\nmodel.compile(optimizer='adam', loss=pair_loss)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57637711","CreationDate":"2019-08-24T11:54:34.607","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<pre><code>X = Lambda(lambda x: return x[:,0] - x[:,1])(X)\nX = Dense(...)(X)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57633850,"AnswerCount":1,"Tags":"<python><tensorflow><google-cloud-platform><jupyter-notebook><google-cloud-ml>","CreationDate":"2019-08-23T22:59:25.937","AcceptedAnswerId":57634184.0,"OwnerUserId":11969270.0,"Title":"fail to submit job to f1-micro with gcloud ml-engine (or ai-platform) command in jupyter notebook","Body":"<p>I am trying to submit a google cloud job that trains cnn model for mnist digit.  since I am new to gcp, I want to train this job on f1-micro machines first for practice. but not successful. I have two issues along the way.<\/p>\n\n<p>here's my systems. windows 10, anaconda, jupyter notebook 6, python 3.6, tf 1.13.0. \nat first my model works well without any gcp command involved. Then I packed the files into a module as the gcp course suggested. and use gcloud command for local train. the cell seems stuck and doing nothing until I close and halt the ipynb file. the training started right after it and results are correct as I monitored it on Tensorboard. what do I need to do to make it run normally from the cell without closing that notebook? btw I can make it run in a terminal without this issue though.<\/p>\n\n<p>second issue, I then tried to do a submission to google cloud machine. I created a vm instance with f1-micro just to practice since it has a lot of free hours. but my command options aren't accepted. I tried a couple of format for the machine type. i can't set the machine type right. and how do I build the connection to the instance I have created?<\/p>\n\n<p>any advice? thanks! codes are here.<\/p>\n\n<pre><code>#1.local submission lines\n\n\nOUTDIR='trained_test'\n\nINPDIR='..\\data'\nshutil.rmtree(path = OUTDIR, ignore_errors = True) \n\n!gcloud ai-platform local train \\\n    --module-name=trainer.task \\\n    --package-path=trainer \\\n    -- \\\n    --output_dir=$OUTDIR \\\n    --input_dir=$INPDIR \\\n    --epochs=2 \\\n    --learning_rate=0.001 \\\n    --batch_size=100\n\n\n#2. submit to compute engine\n\nOUTDIR='gs:\/\/'+BUCKET+'\/digit\/train_01'\nINPDIR='gs:\/\/'+BUCKET+'\/digit\/data'\nJOBNAME='kaggle_digit_01_'+datetime.now().strftime(\"%Y%m%d_%H%M%S\")\n\n!gcloud ai-platform jobs submit training $JOBNAME \\\n    --region=$REGION \\\n    --module-name=trainer.task \\\n    --package-path=trainer \\\n    --job-dir=$OUTDIR \\\n    --staging-bucket=gs:\/\/$BUCKET \\\n    --scale-tier=custom \\\n    --master-machine-type=zones\/us-central1-a\/machineTypes\/f1-micro \\\n    --runtime-version 1.13 \\\n    -- \\\n    --output_dir=OUTDIR \\\n    --input_dir=INPDIR \\\n    --epochs=5 --learning_rate=0.001 --batch_size=100 \\\n<\/code><\/pre>\n\n<p>Error message:<\/p>\n\n<pre><code>ERROR: (gcloud.ai-platform.jobs.submit.training) INVALID_ARGUMENT: Field: master_type Error: The specified machine type is not supported: zones\/us-central1-a\/machineTypes\/f1-micro\n- '@type': type.googleapis.com\/google.rpc.BadRequest\n  fieldViolations:\n  - description: 'The specified machine type is not supported: zones\/us-central1-a\/machineTypes\/f1-micro'\n    field: master_type\n<\/code><\/pre>\n\n<p>Update:<\/p>\n\n<p>changing the machine type does work <\/p>\n\n<pre><code>--scale-tier=CUSTOM \\\n--master-machine-type=n1-standard-4 \\\n<\/code><\/pre>\n\n<p>I also put the following at the beginning, so the notebook recognize the file format such as $jobname...<\/p>\n\n<pre><code>import gcsfs\n<\/code><\/pre>\n\n<p>btw --job-dir doesn't seem to matter.<\/p>\n\n<p>however the local train still have the same issue that, I need to close and halt the notebook to kick off the training. \ncould anyone give a suggestion on this?<\/p>\n","answers":[{"AnswerId":"57634184","CreationDate":"2019-08-24T00:04:52.487","ParentId":null,"OwnerUserId":"7438427","Title":null,"Body":"<p>f1-micro is not supported by AI Platform Training.\n<a href=\"https:\/\/cloud.google.com\/ml-engine\/docs\/machine-types#compute-engine-machine-types\" rel=\"nofollow noreferrer\">Here<\/a> is the list of supported machines. Also you don't need to specify zone. just the machine type. I.e., --master-machine-type=n1-standard-4<\/p>\n"}]}
+{"QuestionId":57633854,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-08-23T23:00:31.680","AcceptedAnswerId":57633983.0,"OwnerUserId":9715052.0,"Title":"Input and Output shape mismatched when using binary cross entropy | Keras tf autoencoder","Body":"<p>I'm building an autoencoder that takes in 16x16 flattened images and I keep getting the following error: <\/p>\n\n<pre><code>ValueError: A target array with shape (999, 16, 16, 256) was passed for an output of shape (None, 256) while using as loss `binary_crossentropy`. This loss expects targets to have the same shape as the output.\n<\/code><\/pre>\n\n<p>My Attempt: I thought that I had to reshape the tensors inside the decoder after the <code>topk<\/code> layer but it didn't resolve the error and just outputted more errors. I commented out my attempt below. \nI believe that I need to do some type of reshaping so that the input shape and output shape match each so binary cross entropy loss can work. <\/p>\n\n<p>Here is a minimal working example of my code: <\/p>\n\n<pre><code>encoding_dim = 16\ninput_img = tf.keras.layers.Input(shape=(16, 16, 256), name =\"input\")\n# flatten your images\nflatten = tf.keras.layers.Flatten()(input_img)\nencoded = tf.keras.layers.Dense(encoding_dim, activation='relu')(flatten)\nencoded2 = tf.keras.layers.Dense(256, activation='sigmoid')(encoded)\n# top_k layer\ntopk = tf.keras.layers.Lambda(lambda x: tf.nn.top_k(x, k=int(int(x.shape[-1])\/2),\n                                                sorted=True,\n                                                name=\"topk\").values)(encoded)\ndecoded = tf.keras.layers.Dense(128, activation='relu')(topk) \ndecoded2 = tf.keras.layers.Dense(256, activation='sigmoid')(decoded)\n#decoded3 = tf.keras.layers.Reshape((16,16,256))(decoded2)\nautoencoder = tf.keras.Model(input_img, decoded2)\nautoencoder.compile(optimizer = 'adadelta', \n                    loss='binary_crossentropy')\nautoencoder.fit(x_train, x_train,\n            epochs=100,\n            batch_size=256,\n            shuffle=False,\n            validation_data=(x_test, x_test))\n<\/code><\/pre>\n","answers":[{"AnswerId":"57633983","CreationDate":"2019-08-23T23:21:39.323","ParentId":null,"OwnerUserId":"10886420","Title":null,"Body":"<p>Currently, your input shape is <code>(batch, 16, 16, 256)<\/code>. <\/p>\n\n<p>If you want to specify <code>batch<\/code> as <code>256<\/code> explicitly you should use <code>batch_shape<\/code> argument, not <code>shape<\/code> inside your <code>Input<\/code> layer. You shouldn't do that though, as it's better to keep it flexible in case you would like to change it (or batches are of unequal size which is usually the case).<\/p>\n\n<p>Your output has to be of the same shape, so it should be <code>keras.layers.Dense(16 * 16, activation=\"sigmoid\")<\/code> with <code>keras.layers.Reshape((16, 16))<\/code> at the last output node. Notice you do not specify batch anywhere, only the rest of dimensions.<\/p>\n"}]}
+{"QuestionId":57633864,"AnswerCount":1,"Tags":"<python><keras><deep-learning><time-series><lstm>","CreationDate":"2019-08-23T23:02:13.327","AcceptedAnswerId":null,"OwnerUserId":7789281.0,"Title":"LSTM X Values are shifted on the prediction?","Body":"<h1>Intro and questions:<\/h1>\n\n<p>I'm somewhat new to Keras and deep learning and I'm trying to predict Tesla stock prices using an LSTM. I\u2019m a definitely a machine learning\/deep learning beginner so I was hoping someone with some more knowledge and experience could help guide me in the right direction. <\/p>\n\n<p>My network is performing well on the y-value predictions. But it seems like the x-values are shifted too far to the left towards the origin. It seems like if I shifted the data over to the right, the predictions would actually be pretty good. \nBelow is a picture of the prediction plot: <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/gPN1X.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/gPN1X.png\" alt=\"Prediction\"><\/a><\/p>\n\n<p>I'm pretty sure the error is coming from when I make the X_test array of values. Below is all the code below organized into sections:<\/p>\n\n<hr>\n\n<h2>Dataset:<\/h2>\n\n<p>My train data is 4 years of Tesla closing stock prices from 2014-2018. The data I'm going to make the predictions on is the closing stock prices of 2019.<\/p>\n\n<pre><code># get 2014-2018 data to train our model\nstart = datetime.datetime(2014,1,1)\nend = datetime.datetime(2018,12,30)\ndf = web.DataReader(\"TSLA\", 'yahoo', start, end) \n\n# get 2019 data to test our model on \nstart = datetime.datetime(2019,1,1)\nend = datetime.date.today()\ntest_df = web.DataReader(\"TSLA\", 'yahoo', start, end) \n\n# sort by date\ndf = df.sort_values('Date')\ntest_df = test_df.sort_values('Date')\n\n# fix the date \ndf.reset_index(inplace=True)\ndf.set_index(\"Date\", inplace=True)\ntest_df.reset_index(inplace=True)\ntest_df.set_index(\"Date\", inplace=True)\n\ndf.tail()\n<\/code><\/pre>\n\n<pre><code>                  High         Low        Open       Close   Volume  \nDate                                                                  \n2014-01-02  152.479996  146.550003  149.800003  150.100006  6188400   \n2014-01-03  152.190002  148.600006  150.000000  149.559998  4695000   \n2014-01-06  150.399994  145.240005  150.000000  147.000000  5361100   \n2014-01-07  150.399994  145.250000  147.619995  149.360001  5034100   \n2014-01-08  153.699997  148.759995  148.850006  151.279999  6163200   \n...                ...         ...         ...         ...      ...   \n2018-12-24  314.500000  295.200012  313.500000  295.390015  5559900   \n2018-12-26  326.970001  294.089996  300.000000  326.089996  8163100   \n2018-12-27  322.170013  301.500000  319.839996  316.130005  8575100   \n2018-12-28  336.239990  318.410004  323.100006  333.869995  9939000   \n2018-12-31  339.209991  325.260010  337.790009  332.799988  6302300\n<\/code><\/pre>\n\n<hr>\n\n<h2>Creating train data:<\/h2>\n\n<pre><code># create train set of adj close prices data:\ntrain_data = df.loc[:,'Adj Close'].as_matrix()\nprint(train_data.shape) # 1258 \n\n\n# Apply normalization before feeding to LSTM using sklearn:\nfrom sklearn.preprocessing import MinMaxScaler\n\nscaler = MinMaxScaler()\ntrain_data = train_data.reshape(-1,1)\n\nscaler.fit(train_data)\ntrain_data = scaler.transform(train_data)\n\n\nX_train = []\ny_train = []\n# loop through the data in batches of 36 to create our time steps\nprint(train_data.shape)\nfor i in range(36, len(train_data)):\n    # append \n    X_train.append(train_data[i-36:i, 0])\n    y_train.append(train_data[i, 0])\nX_train, y_train = np.array(X_train), np.array(y_train)\nprint(len(X_train))\n\nX_train = np.reshape(X_train, (X_train.shape[0], X_train.shape[1], 1))\n<\/code><\/pre>\n\n<p>The len of train_data before the for loop is 1258 and after the for loop its 1222. <\/p>\n\n<hr>\n\n<h2>Creating our model:<\/h2>\n\n<pre><code># Build the model \nmodel = Sequential()\n\nmodel.add(LSTM(units = 100, return_sequences = True, input_shape = (X_train.shape[1], 1)))\nmodel.add(Dropout(0.2))\n\nmodel.add(LSTM(units = 100))\nmodel.add(Dropout(0.2))\n\n# Output layer\nmodel.add(Dense(units = 1))\n\n# Compiling the RNN\nmodel.compile(optimizer = 'adam', loss = 'mean_squared_error')\n<\/code><\/pre>\n\n<hr>\n\n<h3>Fitting the model<\/h3>\n\n<pre><code>history = model.fit(X_train, y_train, epochs = 20, batch_size = 10, validation_split=.30)\n\nTrain on 855 samples, validate on 367 samples\nEpoch 1\/20\n855\/855 [==============================] - 14s 17ms\/step - loss: 0.0099 - val_loss: 0.0061\nEpoch 2\/20\n855\/855 [==============================] - 6s 7ms\/step - loss: 0.0035 - val_loss: 0.0059\nEpoch 3\/20\n855\/855 [==============================] - 6s 7ms\/step - loss: 0.0036 - val_loss: 0.0057\nEpoch 4\/20\n855\/855 [==============================] - 6s 7ms\/step - loss: 0.0029 - val_loss: 0.0042\nEpoch 5\/20\n855\/855 [==============================] - 6s 7ms\/step - loss: 0.0024 - val_loss: 0.0092\nEpoch 6\/20\n855\/855 [==============================] - 6s 7ms\/step - loss: 0.0025 - val_loss: 0.0039\nEpoch 7\/20\n855\/855 [==============================] - 6s 7ms\/step - loss: 0.0024 - val_loss: 0.0047\nEpoch 8\/20\n855\/855 [==============================] - 6s 7ms\/step - loss: 0.0027 - val_loss: 0.0034\nEpoch 9\/20\n855\/855 [==============================] - 6s 7ms\/step - loss: 0.0023 - val_loss: 0.0035\nEpoch 10\/20\n855\/855 [==============================] - 6s 7ms\/step - loss: 0.0020 - val_loss: 0.0029\nEpoch 11\/20\n855\/855 [==============================] - 7s 8ms\/step - loss: 0.0018 - val_loss: 0.0033\nEpoch 12\/20\n855\/855 [==============================] - 6s 8ms\/step - loss: 0.0021 - val_loss: 0.0027\nEpoch 13\/20\n855\/855 [==============================] - 6s 7ms\/step - loss: 0.0018 - val_loss: 0.0025\nEpoch 14\/20\n855\/855 [==============================] - 6s 7ms\/step - loss: 0.0019 - val_loss: 0.0038\nEpoch 15\/20\n855\/855 [==============================] - 6s 7ms\/step - loss: 0.0019 - val_loss: 0.0037\nEpoch 16\/20\n855\/855 [==============================] - 6s 7ms\/step - loss: 0.0018 - val_loss: 0.0023\nEpoch 17\/20\n855\/855 [==============================] - 7s 8ms\/step - loss: 0.0017 - val_loss: 0.0025\nEpoch 18\/20\n855\/855 [==============================] - 8s 9ms\/step - loss: 0.0015 - val_loss: 0.0021\nEpoch 19\/20\n855\/855 [==============================] - 8s 9ms\/step - loss: 0.0015 - val_loss: 0.0025\nEpoch 20\/20\n855\/855 [==============================] - 6s 7ms\/step - loss: 0.0016 - val_loss: 0.0023\n<\/code><\/pre>\n\n<hr>\n\n<h2>Model loss plot:<\/h2>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/F4A7z.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/F4A7z.png\" alt=\"Model loss\"><\/a><\/p>\n\n<hr>\n\n<h2>Creating test data (BUG here probably):<\/h2>\n\n<pre><code># Predict with the model and plot predictions  \ninputs = test_df['Adj Close'].values\ninputs = inputs.reshape(-1,1)\ninputs = scaler.transform(inputs)\n\nprint(len(inputs)) # 159 \n\nX_test = []\ny_test = []\nfor i in range(36, len(inputs)):\n    X_test.append(inputs[i-36:i, 0])\n\nX_test = np.array(X_test)\nprint(len(X_test)) #123\nX_test = np.reshape(X_test, (X_test.shape[0], X_test.shape[1], 1))\n\npredictions = model.predict(X_test)\npredictions = scaler.inverse_transform(predictions)\n#print(predictions)\n<\/code><\/pre>\n\n<p>What's interesting is the <code>len(inputs)<\/code> is = 159 before going into the for loop and the coming out of the for loop the <code>len(X_test)<\/code> is = 123. I'm using this for loop to create batches of 36 days of stock prices so our LSTM can take into consideration 36 days of prices before making the prediction. But it seems like it's cutting off the last 36 values? <\/p>\n","answers":[{"AnswerId":"57642740","CreationDate":"2019-08-25T02:05:15.720","ParentId":null,"OwnerUserId":"7789281","Title":null,"Body":"<h2>Solution:<\/h2>\n\n<p>If anyone finds this from google I figured it out. I changed the train and test data creation code to this: <\/p>\n\n<pre><code>'''Function to create a dataset to feed into an LSTM'''\ndef create_dataset(dataset, look_back):\n    dataX, dataY = [], []\n    for i in range(len(dataset)-look_back):\n        a = dataset[i:(i + look_back), 0]\n        dataX.append(a)\n        dataY.append(dataset[i + look_back, 0])\n    return np.array(dataX), np.array(dataY)\n\n# Create the data to train our model on:\ntime_steps = 36\nX_train, y_train = create_dataset(train_data, time_steps)\n\n# reshape it [samples, time steps, features]\nX_train = np.reshape(X_train, (X_train.shape[0], 36, 1))\n\nprint(X_train.shape) # 1222, 36, 1\n<\/code><\/pre>\n\n<hr>\n\n<h2>Test Data:<\/h2>\n\n<pre><code># Get the stock prices for 2019 to have our model make the predictions\ntest_data = test_df['Adj Close'].values\ntest_data = test_data.reshape(-1,1)\ntest_data = scaler.transform(test_data)\n\n# Create the data to test our model on:\ntime_steps = 36\nX_test, y_test = create_dataset(test_data, time_steps)\n\n# store the original vals for plotting the predictions \ny_test = y_test.reshape(-1,1)\norg_y = scaler.inverse_transform(y_test)\n\n# reshape it [samples, time steps, features]\nX_test = np.reshape(X_test, (X_test.shape[0], 36, 1))\n\n# Predict the prices with the model\npredicted_y = model.predict(X_test)\npredicted_y = scaler.inverse_transform(predicted_y)\n<\/code><\/pre>\n\n<hr>\n\n<h2>New predictions:<\/h2>\n\n<p>I changed it to plot the original y vals stored in <code>org_y<\/code> and then plot our predicted y vals <\/p>\n\n<pre><code>plt.plot(org_y, color = 'red', label = 'Real Tesla Stock Price')\nplt.plot(predicted_y, color = 'blue', label = 'Predicted Tesla Stock Price')\nplt.title('Tesla Stock Price Prediction')\nplt.xlabel('Time')\nplt.ylabel('Tesla Stock Price')\nplt.legend()\nplt.show()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/hfJvd.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/hfJvd.png\" alt=\"Predict\"><\/a><\/p>\n"}]}
+{"QuestionId":57634231,"AnswerCount":2,"Tags":"<python><pytorch><tensorboard>","CreationDate":"2019-08-24T00:19:14.993","AcceptedAnswerId":57666055.0,"OwnerUserId":6513347.0,"Title":"Why does my tensorboard say: 'No scalar data was found' when using SummaryWriter.add_scalar?","Body":"<p>I am attempting to visualize a loss vs. epoch graph using tensorboard (by calling .addscalar()) so that I can appropriately tune my neural network model. However, after training the model, I check my tensorboard in the web browser and it reads: 'No scalar data was found.'<\/p>\n\n<p>I've tried checking my events statistics and it seems like there are scalars: <\/p>\n\n<pre><code>Event statistics for runs\/tf_druggability\/classification_ffnn_1:\naudio -\ngraph -\nhistograms -\nimages -\nscalars\n   first_step           0\n   last_step            90\n   max_step             90\n   min_step             0\n   num_steps            10\n   outoforder_steps     [(90L, 0L), (90L, 0L)]\nsessionlog:checkpoint -\nsessionlog:start -\nsessionlog:stop -\ntensor -\n<\/code><\/pre>\n\n<p>Here's my code:<\/p>\n\n<pre><code>from torch.utils.tensorboard import SummaryWriter\n\nsess = tensorflow.Session()\nwriter = SummaryWriter('runs\/tf_druggability\/classification_ffnn_1', sess.graph)\n\nfor epoch in range(100):\n\n    # Wrap input data and labels in Variable to can gradient descent (in place of DataLoader)\n    inputs = Variable(traindf_to_tensor)\n    labels = Variable(trainlabels_to_tensor)\n\n    optimizer.zero_grad()\n    outputs = net(inputs)\n    loss = criterion(outputs, labels)\n    loss.backward()\n    optimizer.step()\n\n    if(epoch%10 == 0):\n\n        # HERE I WRITE to the LOG FILE for TENSORBOARD:\n        writer.add_scalar('training loss', loss.item(), epoch)\n        print('epoch {}, loss {}'.format(epoch, loss.item()))\n\nwriter.close()\nprint('Finished training! :)')\n<\/code><\/pre>\n\n<p>When I try to open the events file in Jupyter Notebooks it has this in it (... means I've omitted the middle part):<\/p>\n\n<pre><code>1800 0000 0000 0000 a37f 4b22 09f2 3d6b\na51e 58d7 411a 0d62 7261 696e 2e45 7665\n6e74 3a32 ac51 0a16 2100 0000 0000 0000\n...\nf723 0000 0000 0000 00b0 9f77 4309 d00a\n0ea7 1e58 d741 105a 2a16 0a14 0a0d 7472\n6169 6e69 6e67 5f6c 6f73 7315 c193 403f\n<\/code><\/pre>\n","answers":[{"AnswerId":"57666055","CreationDate":"2019-08-26T23:37:46.487","ParentId":null,"OwnerUserId":"6513347","Title":null,"Body":"<p>I discovered that the graphs only displayed if I was in the exact <strong>same<\/strong> directory as the program I'm running. Being in a higher directory than that of the program seems to have caused this inability to display graphs. <\/p>\n"},{"AnswerId":"57635209","CreationDate":"2019-08-24T04:53:22.170","ParentId":null,"OwnerUserId":"5027315","Title":null,"Body":"<p>I am not sure why are you passing a tensorflow <code>session.graph<\/code> object to a Pytorch <code>SummaryWriter()<\/code> object. <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensorboard.html\" rel=\"nofollow noreferrer\">Here<\/a> you can find some basic example of using tensorboard to write graphs and scalars etc.<\/p>\n"}]}
+{"QuestionId":57635169,"AnswerCount":1,"Tags":"<deep-learning><pytorch>","CreationDate":"2019-08-24T04:44:11.910","AcceptedAnswerId":57637213.0,"OwnerUserId":8176285.0,"Title":"Implementation of Focal loss for multi label classification","Body":"<p>trying to write focal loss for multi-label classification <\/p>\n\n<pre><code>class FocalLoss(nn.Module):\n    def __init__(self, gamma=2, alpha=0.25):\n        self._gamma = gamma\n        self._alpha = alpha\n\n    def forward(self, y_true, y_pred):\n        cross_entropy_loss = torch.nn.BCELoss(y_true, y_pred)\n        p_t = ((y_true * y_pred) +\n               ((1 - y_true) * (1 - y_pred)))\n        modulating_factor = 1.0\n        if self._gamma:\n            modulating_factor = torch.pow(1.0 - p_t, self._gamma)\n        alpha_weight_factor = 1.0\n        if self._alpha is not None:\n            alpha_weight_factor = (y_true * self._alpha +\n                                   (1 - y_true) * (1 - self._alpha))\n        focal_cross_entropy_loss = (modulating_factor * alpha_weight_factor *\n                                    cross_entropy_loss)\n        return focal_cross_entropy_loss.mean()\n<\/code><\/pre>\n\n<p>But when i run this i get <\/p>\n\n<pre><code>  File \"train.py\", line 82, in &lt;module&gt;\n    loss = loss_fn(output, target)\n  File \"\/home\/bubbles\/.local\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/module.py\", line 538, in __call__\n    for hook in self._forward_pre_hooks.values():\n  File \"\/home\/bubbles\/.local\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/module.py\", line 591, in __getattr__\n    type(self).__name__, name))\nAttributeError: 'FocalLoss' object has no attribute '_forward_pre_hooks'\n<\/code><\/pre>\n\n<p>Any suggestions would be really helpful, Thanks in advance.<\/p>\n","answers":[{"AnswerId":"57637213","CreationDate":"2019-08-24T10:45:03.950","ParentId":null,"OwnerUserId":"10886420","Title":null,"Body":"<p>You shouldn't inherit from <code>torch.nn.Module<\/code> as it's designed for modules with learnable parameters (e.g. neural networks).<\/p>\n\n<p>Just create normal functor or function and you should be fine.<\/p>\n\n<p>BTW. If you inherit from it, you should call <code>super().__init__()<\/code> somewhere in your <code>__init__()<\/code>.<\/p>\n"}]}
+{"QuestionId":57635304,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><google-colaboratory><tensorflow-lite>","CreationDate":"2019-08-24T05:13:49.520","AcceptedAnswerId":null,"OwnerUserId":10202807.0,"Title":"Converting ssd_mobilenet to tensorflow lite throws, ConverterError: TOCO failed. See console for info","Body":"<p>I am trying to convert ssd_mobilenet_v1_coco available from <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/detection_model_zoo.md#coco-trained-models\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/detection_model_zoo.md#coco-trained-models<\/a> to TFLite. But it is throwing following error :<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nConverterError                            Traceback (most recent call last)\n&lt;ipython-input-10-c2bf9c96c1d4&gt; in &lt;module&gt;()\n----&gt; 1 lite_model = converter.convert()\n      2 \n      3 open( folder_name + '.tflite', 'wb' ).write( lite_model )\n\n2 frames\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/lite\/python\/convert.py in toco_convert_protos(model_flags_str, toco_flags_str, input_data_str)\n    170       stderr = _try_convert_to_unicode(stderr)\n    171       raise ConverterError(\n--&gt; 172           \"TOCO failed. See console for info.\\n%s\\n%s\\n\" % (stdout, stderr))\n    173   finally:\n    174     # Must manually cleanup files.\n\nConverterError: TOCO failed. See console for info.\n2019-08-24 04:48:54.444064: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.475768: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.475895: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayV3\n2019-08-24 04:48:54.475960: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.475987: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayV3\n2019-08-24 04:48:54.476003: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.476017: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayV3\n2019-08-24 04:48:54.476030: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.476043: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.476061: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.476073: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.476086: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.476101: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.476113: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.476125: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.476148: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.476161: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.476183: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayScatterV3\n2019-08-24 04:48:54.476202: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: LoopCond\n2019-08-24 04:48:54.476228: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.476258: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Exit\n2019-08-24 04:48:54.476277: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Exit\n2019-08-24 04:48:54.476300: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArraySizeV3\n2019-08-24 04:48:54.476315: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArraySizeV3\n2019-08-24 04:48:54.476342: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayReadV3\n2019-08-24 04:48:54.476385: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayWriteV3\n2019-08-24 04:48:54.476406: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayGatherV3\n2019-08-24 04:48:54.476423: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayGatherV3\n2019-08-24 04:48:54.476504: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayWriteV3\n2019-08-24 04:48:54.477386: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayV3\n2019-08-24 04:48:54.477421: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.477449: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayV3\n2019-08-24 04:48:54.477463: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.477477: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayV3\n2019-08-24 04:48:54.477493: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.477507: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayV3\n2019-08-24 04:48:54.477519: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.477532: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayV3\n2019-08-24 04:48:54.477544: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.477558: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayV3\n2019-08-24 04:48:54.477570: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.477583: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayV3\n2019-08-24 04:48:54.477595: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.477608: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayV3\n2019-08-24 04:48:54.477620: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.477632: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.477651: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.477665: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.477677: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.477689: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.477700: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayScatterV3\n2019-08-24 04:48:54.477728: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.477741: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.477753: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayScatterV3\n2019-08-24 04:48:54.477766: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.477779: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.477806: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.477834: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.477861: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.477873: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.477887: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.477898: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.477919: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.477934: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.477946: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.477958: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.477971: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.478010: I tensorflow\/lite\/toco\/import_tensorflow.cc:193] Unsupported data type in placeholder op: 20\n2019-08-24 04:48:54.478040: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayScatterV3\n2019-08-24 04:48:54.478072: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.478086: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.478106: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: LoopCond\n2019-08-24 04:48:54.478120: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.478147: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Exit\n2019-08-24 04:48:54.478162: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Exit\n2019-08-24 04:48:54.478179: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Exit\n2019-08-24 04:48:54.478194: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Exit\n2019-08-24 04:48:54.478214: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArraySizeV3\n2019-08-24 04:48:54.478227: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArraySizeV3\n2019-08-24 04:48:54.478239: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArraySizeV3\n2019-08-24 04:48:54.478251: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArraySizeV3\n2019-08-24 04:48:54.478284: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayReadV3\n2019-08-24 04:48:54.478299: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayReadV3\n2019-08-24 04:48:54.478312: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayReadV3\n2019-08-24 04:48:54.480962: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayScatterV3\n2019-08-24 04:48:54.481010: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayGatherV3\n2019-08-24 04:48:54.481029: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayGatherV3\n2019-08-24 04:48:54.481045: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayGatherV3\n2019-08-24 04:48:54.481059: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayGatherV3\n2019-08-24 04:48:54.481077: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Enter\n2019-08-24 04:48:54.481099: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayReadV3\n2019-08-24 04:48:54.485710: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.485745: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.485762: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.485776: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.485789: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.485818: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.485831: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.485855: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.485879: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.485894: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.485914: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.485930: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.485949: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.485962: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.485975: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.485988: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486001: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486014: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486027: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486041: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486054: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486067: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486080: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486108: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486122: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486135: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486148: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486162: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486175: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486188: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486220: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486236: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486249: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486270: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486296: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486310: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486323: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486353: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486368: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486382: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486396: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486409: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486436: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486451: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486464: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486477: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486490: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486503: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486517: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486530: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486543: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486556: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486569: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486583: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486596: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486609: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486622: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486636: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486659: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486673: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486686: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486700: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486713: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486726: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486739: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486752: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486765: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486779: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486800: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486814: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486830: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486844: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486857: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486871: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486884: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486898: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486931: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486946: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486959: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486973: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.486986: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.487000: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.487013: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.487035: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.487050: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.487063: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.487077: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.487090: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.487103: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.487117: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: NonMaxSuppressionV2\n2019-08-24 04:48:54.487655: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: Size\n2019-08-24 04:48:54.487742: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayWriteV3\n2019-08-24 04:48:54.488021: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayWriteV3\n2019-08-24 04:48:54.488044: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayWriteV3\n2019-08-24 04:48:54.488058: I tensorflow\/lite\/toco\/import_tensorflow.cc:1336] Converting unsupported operation: TensorArrayWriteV3\n2019-08-24 04:48:54.815195: I tensorflow\/lite\/toco\/graph_transformations\/graph_transformations.cc:39] Before Removing unused ops: 4139 operators, 6859 arrays (0 quantized)\n2019-08-24 04:48:55.234858: I tensorflow\/lite\/toco\/graph_transformations\/graph_transformations.cc:39] After Removing unused ops pass 1: 4116 operators, 6815 arrays (0 quantized)\n2019-08-24 04:48:55.796827: I tensorflow\/lite\/toco\/graph_transformations\/graph_transformations.cc:39] Before general graph transformations: 4116 operators, 6815 arrays (0 quantized)\n2019-08-24 04:48:56.183189: F tensorflow\/lite\/toco\/graph_transformations\/resolve_tensorflow_switch.cc:98] Check failed: other_op-&gt;type == OperatorType::kMerge Found StridedSlice as non-selected output from Switch, but only Merge supported.\nAborted (core dumped)\n\n\n\n<\/code><\/pre>\n\n<p>Colab notebook to reproduce the error : <a href=\"https:\/\/colab.research.google.com\/drive\/1FcalJTa1IRuTsPOJ9f3oiUnIkhvj0Q_Y\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/1FcalJTa1IRuTsPOJ9f3oiUnIkhvj0Q_Y<\/a><\/p>\n\n<p>Colab runtime : GPU<\/p>\n\n<p>Is this due to any unsupported layers or am I making any mistakes? TIA.<\/p>\n","answers":[{"AnswerId":"57671971","CreationDate":"2019-08-27T09:48:11.317","ParentId":null,"OwnerUserId":"10202807","Title":null,"Body":"<p>I figured it out. I was using official frozen_inference_graph.pb. But TFLite is expecting SSDs exported using this script.<a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/export_tflite_ssd_graph.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/export_tflite_ssd_graph.py<\/a>. The input_array and output_array is different here. This can be found on the comments provided in the export script.<\/p>\n"}]}
+{"QuestionId":57635704,"AnswerCount":0,"Tags":"<python><machine-learning><keras><deep-learning><lstm>","CreationDate":"2019-08-24T06:39:40.830","AcceptedAnswerId":null,"OwnerUserId":11309950.0,"Title":"Does the order of inputs to the LSTM model affect the model.predict?","Body":"<p>Here are two different inputs.<\/p>\n\n<p><code>signal_data = df[[\"close\", \"volume\"]].values.astype('float32')<\/code> <\/p>\n\n<p><code>signal_data = df[[\"volume\", \"close\"]].values.astype('float32')<\/code><\/p>\n\n<p>When the model is configured with LSTM, I wonder if the above two differences affect <code>model.predict.<\/code><\/p>\n\n<p>I noticed a difference when using <code>inverse_transform<\/code>. I wonder if I used <code>inverse_transfrom<\/code> incorrectly or if the order of input into the model affected <code>model.predict<\/code>.<\/p>\n\n<p>Your valuable opinions and thoughts will be very much appreciated.<\/p>\n","answers":[]}
+{"QuestionId":57635757,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-24T06:48:33.800","AcceptedAnswerId":null,"OwnerUserId":10762045.0,"Title":"How to fix \"Received: 50 (missing previous layer metadata).\" error in keras?","Body":"<p>I'm trying to copy the layer from the autoencoder to use to decode the file but i'm getting an error<\/p>\n\n<pre><code>encoded_input = Input(shape=(50,))\n\ndecoder_layer = autoencoder.layers[-1](encoded_input)\n\ndecoder = Model(50, decoder_layer)\n\nautoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')\n\nautoencoder.fit(X, X,\n            epochs=50,\n            shuffle=True)\n<\/code><\/pre>\n\n<p>I'm expecting to copy the layer so that I can decode the file but I'm getting this error:<\/p>\n\n<pre><code>Traceback (most recent call last):\nFile \"C:\\Users\\dalto\\Documents\\geo4\\train.py\", line 36, in &lt;module&gt;\ndecoder = Model(50, decoder_layer)\nFile \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 133, in __init__\nsuper(Model, self).__init__(*args, **kwargs)\nFile \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\network.py\", line 161, in __init__\nself._init_graph_network(*args, **kwargs)\nFile \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\training\\tracking\\base.py\", line 458, in _method_wrapper\nresult = method(self, *args, **kwargs)\nFile \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\network.py\", line 270, in _init_graph_network\nself._validate_graph_inputs_and_outputs()\nFile \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\network.py\", line 1521, in _validate_graph_inputs_and_outputs\n' (missing previous layer metadata).')\n<\/code><\/pre>\n\n<p>ValueError: Input tensors to a Model must come from <code>tf.keras.Input<\/code>. Received: 50 (missing previous layer metadata).<\/p>\n","answers":[]}
+{"QuestionId":57636091,"AnswerCount":2,"Tags":"<python><keras>","CreationDate":"2019-08-24T07:50:39.733","AcceptedAnswerId":null,"OwnerUserId":10842027.0,"Title":"How to get a callback when the specified epoch number is over?","Body":"<p>I want to fine turn my model when using Keras, and I want to change my training data and learning rate to train when the epochs arrive 10, So how to get a callback when the specified epoch number is over.<\/p>\n","answers":[{"AnswerId":"57636260","CreationDate":"2019-08-24T08:17:45.020","ParentId":null,"OwnerUserId":"8198272","Title":null,"Body":"<p>You need to write your own Callback subclass.<\/p>\n\n<p><a href=\"https:\/\/keras.io\/callbacks\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/callbacks\/<\/a> (general information)<\/p>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/callbacks.py#L275\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/callbacks.py#L275<\/a> (source code for the Callback base class)<\/p>\n\n<p>Your Callback subclass should define an <code>on_epoch_end()<\/code> method, which accepts the epoch number as an argument.<\/p>\n"},{"AnswerId":"57636395","CreationDate":"2019-08-24T08:41:14.913","ParentId":null,"OwnerUserId":"9694304","Title":null,"Body":"<p>Actually, the way keras works this is probably not the best way to go, it would be much better to treat this as fine tuning, meaning that you finish the 10 epochs, save the model and then load the model (from another script) and continue training with the lr and data you fancy.<\/p>\n\n<p>There are several reasons for this.<\/p>\n\n<ol>\n<li><p>It is much clearer and easier to debug. You check you model properly after the 10 epochs, verify that it works properly and carry on<\/p><\/li>\n<li><p>It is much better to do several experiments this way, starting from epoch 10.<\/p><\/li>\n<\/ol>\n\n<p>Good luck!<\/p>\n"}]}
+{"QuestionId":57636135,"AnswerCount":0,"Tags":"<python><keras><deep-learning>","CreationDate":"2019-08-24T07:57:36.147","AcceptedAnswerId":null,"OwnerUserId":6668992.0,"Title":"how to implement input based on continues and categorical data which have strong relation between them?","Body":"<p>i have some marketing data and continues output, two of the inputs have strong relation that output is proportional to the product of the two oup\u221dinp*A. were oup and inp are continues and A is categorical, how can i implement these inputs correctly ? i put them into a code but i received 0 accuracy which i guess is more a problem of correct implementing than code bag.  <\/p>\n\n<p>in order to show what is my problem i created demo data\nand put it into the code below <\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nimport statsmodels.api as sm\nimport matplotlib.pyplot as plt\nimport seaborn as sns\nimport tensorflow as tf\nsns.set()\nimport nbimporter\n\nfrom keras import models \nfrom keras.models import Sequential\nfrom keras.layers import Dense, Dropout\nfrom keras.constraints import unit_norm\nfrom keras import optimizers\nfrom sklearn.model_selection import StratifiedKFold\nfrom sklearn.model_selection import KFold\n\nfrom scipy import stats\nfrom sklearn.utils import shuffle\nfrom sklearn.preprocessing import StandardScaler\nfrom sklearn.preprocessing import MinMaxScaler\nstats.chisqprob = lambda chisq, df: stats.chi2.sf(chisq, df)\ntf.reset_default_graph()\nfrom keras_adabound import AdaBound\nfrom keras.optimizers import SGD\nimport h5py\nfrom copy import deepcopy\n\nsiz=100000\ninp=np.random.randint(100, 1000000 , size=siz)\na1=0.2;a2=0.8;a3=2.5;a4=3.7   \nuop=np.array([inp[0:int(siz\/4)]*a1,inp[int(siz\/4):int(siz\/2)]\\\n              *a2,inp[int(siz\/2):int(siz*3\/4)]*a3\\\n              ,inp[int(siz*3\/4):siz]*a4])\n uop=uop.reshape(siz,1)\nprint(stats.describe(uop))\n\nnormzer_inp = MinMaxScaler()\ninp_norm = normzer_inp.fit_transform\\\n(inp.reshape(siz,1))\nnormzer_uop = MinMaxScaler()\nuop_norm = normzer_uop.fit_transform\\\n(uop)\n\ndf = pd.DataFrame(index=range(siz),columns=['A'])\ndf[0:int(siz\/4)]='a1';df[int(siz\/4):int(siz\/2)]='a2';\\\ndf[int(siz\/2):int(siz*3\/4)]='a3';df[int(siz*3\/4):siz]='a4';\ndf['inp'] = inp; df['uop'] = uop; df['inp_norm'] = inp_norm;\\\ndf['uop_norm'] = uop_norm\ndummies = pd.get_dummies(df['A'])\ndf = pd.concat([df,dummies],axis='columns')\ndf = df[ ['A', 'inp', 'uop', 'inp_norm', 'a1',\\\n          'a2', 'a3', 'a4', 'uop_norm' ] ]\n\n#print('df.shape =',df.shape)\n#print(df)\n\ndf = shuffle(df)\ndf = df.reset_index()\ndel df['index']\ndf\nX=df[['inp_norm', 'a1', 'a2', 'a3', 'a4']].to_numpy()\nY=df[['uop_norm']].to_numpy()\nX\n\nkfold = KFold(n_splits=4, random_state=None, shuffle=False)\ncvscores = []\nopti_Adam = optimizers.Adam(lr=0.01)\nopti_SGD = SGD(lr=0.1, momentum=0.9)\nopti_Bound = AdaBound(lr=1e-3, final_lr=0.1)\nmodel1 = Sequential()\naccc_trn=0\naccc_tst=0\nfor train, test in kfold.split(X, Y):\n    # create model\n    model = Sequential()\n    print('model',type(model))    \n    model.add(Dense(128, input_dim=X.shape[1], activation='relu'))\n    #model.add(Dropout(0.))\n    model.add(Dense(128, activation='relu'))\n    model.add(Dense(128, activation='relu'))\n    model.add(Dense(128, activation='relu'))\n\n    model.add(Dense(1, activation='linear'))\n    model.compile(loss='mean_squared_error', optimizer=opti_SGD,\\\n              metrics=['accuracy'])\n\n    history = model.fit(X[train], Y[train], \\\n            validation_data=(X[test], Y[test]), \\\n            epochs=50,batch_size=2048, verbose=2)\n    _, train_acc = model.evaluate(X[train], Y[train], verbose=0)\n    _, test_acc = model.evaluate(X[test], Y[test], verbose=0)\n        print('Train: %.3f, Test: %.3f' % (train_acc, test_acc))  \n    if test_acc&gt;=accc_tst and train_acc&gt;=accc_trn:\n        accc_tst=test_acc\n        accc_trn=train_acc\n        model1=deepcopy(model) \n\n    # plot history\n    plt.plot(history.history['acc'], label='train')\n    plt.plot(history.history['val_acc'], label='test')\n    plt.legend()\n    plt.show()  \n    cvscores.append(test_acc * 100)  \n\nprint(\"%.2f%% (+\/- %.2f%%)\" % (np.mean(cvscores)\\\n                               , np.std(cvscores)))     \n<\/code><\/pre>\n\n<p>any guidance would be appreciated especially of how to solve this problem but also of how to make my codes more efficient\nthx<\/p>\n","answers":[]}
+{"QuestionId":57636918,"AnswerCount":2,"Tags":"<keras><deep-learning><lstm>","CreationDate":"2019-08-24T10:02:28.787","AcceptedAnswerId":null,"OwnerUserId":4117624.0,"Title":"LSTM and TimeseriesGenerator. Error in input shape","Body":"<p>I try to follow online tutorials (<a href=\"https:\/\/www.dlology.com\/blog\/how-to-use-keras-timeseriesgenerator-for-time-series-data\/\" rel=\"nofollow noreferrer\">1<\/a>, <a href=\"https:\/\/machinelearningmastery.com\/how-to-use-the-timeseriesgenerator-for-time-series-forecasting-in-keras\/\" rel=\"nofollow noreferrer\">2<\/a> among others), but when fitting a LSTM model using keras <code>TimeseriesGenerator<\/code>, I cannot get the input dimensions right.<\/p>\n\n<p>For training I have one long numpy array called <code>dn<\/code> of shape <code>(12499896,)<\/code>.\nThen I define the keras <code>TimeseriesGenerator<\/code><\/p>\n\n<pre><code>from keras.preprocessing.sequence import TimeseriesGenerator\nlook_back = 10 \ngen = TimeseriesGenerator(dn, dn, length=look_back, batch_size=8)\n<\/code><\/pre>\n\n<p>I define the model:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import LSTM, Dense\n\nmodel = Sequential()\nmodel.add(LSTM(units=100, activation='relu', input_shape=(look_back, 1)))\nmodel.add(Dense(1))\nmodel.compile(optimizer='adam', loss='mse')\n<\/code><\/pre>\n\n<p>When I fit the model using <code>fit_generator<\/code><\/p>\n\n<pre><code>model.fit_generator(gen, steps_per_epoch=10, epochs=5)\n<\/code><\/pre>\n\n<p>I get the error:<\/p>\n\n<blockquote>\n  <p>Error when checking input: expected lstm_9_input to have 3 dimensions,\n  but got array with shape (8, 10)<\/p>\n<\/blockquote>\n\n<p>If I understand correctly an <code>LSTM<\/code> requires input shape of <code>[samples, look_back, features]<\/code> (here I guess <code>samples=12499896<\/code>, <code>look_back=10<\/code>, and <code>features=1<\/code>). And <code>TimeseriesGenerator<\/code> produces a list of 2-tuples. Should I reshape the output of TimeseriesGenerator? (but I think it is not meant to be changed?). Where am I making a mistake? <\/p>\n","answers":[{"AnswerId":"57637609","CreationDate":"2019-08-24T11:40:22.673","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>You have answered to your question, but I found few mistakes and would like to correct them.<\/p>\n\n<p>When you are specifying <code>input_shape<\/code> for LSTM or any layer in  Keras, you need not mention the <code>batch_size<\/code>. So in case of LSTM, the input shape should be <code>[look_back, features]<\/code>. The more correctly used word for <code>look_back<\/code> is <code>timesteps<\/code>.<\/p>\n\n<p>Since the input is one dimensioned data, you can reshape your data as you have done it which is:<\/p>\n\n<pre><code>dn = np.reshape(dn, (-1, 1))\n<\/code><\/pre>\n\n<p>Also, <code>return_sequences<\/code> in <code>LSTM<\/code> layer is False by default.<\/p>\n"},{"AnswerId":"57637158","CreationDate":"2019-08-24T10:37:13.147","ParentId":null,"OwnerUserId":"4117624","Title":null,"Body":"<p>I figured that it works if I first reshape the array <\/p>\n\n<pre><code>dn = np.reshape(dn, (len(dn),1))\n<\/code><\/pre>\n\n<p>so its shape is <code>(12499896, 1)<\/code> instead of <code>(12499896,)<\/code>.\nThen in the LSTM layer use <code>return_sequences=False<\/code>:<\/p>\n\n<pre><code>model.add(LSTM(units=100, return_sequences=False, activation='relu', input_shape=(look_back, 1)))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57636936,"AnswerCount":0,"Tags":"<python><tensorflow><object-detection-api>","CreationDate":"2019-08-24T10:05:07.433","AcceptedAnswerId":null,"OwnerUserId":8278557.0,"Title":"ModuleNotFoundError: No module named 'object_detection.core'; in object detection api","Body":"<p>When I try to run this statement \n<code>from utils import label_map_util<\/code> I get the following error, I followed all the steps in the object detection installation document, I still get this error.<\/p>\n\n<pre><code>  File \"object_detection\/object_detection.py\", line 19, in &lt;module&gt;\n    from utils import label_map_util\n  File \"\/Users\/ramanandsirvi\/Documents\/tensorflow\/models\/research\/object_detection\/utils\/label_map_util.py\", line 26, in &lt;module&gt;\n    from object_detection.protos import string_int_label_map_pb2\n  File \"\/Users\/ramanandsirvi\/Documents\/tensorflow\/models\/research\/object_detection\/object_detection.py\", line 21, in &lt;module&gt;\n    from utils import visualization_utils as vis_util\n  File \"\/Users\/ramanandsirvi\/Documents\/tensorflow\/models\/research\/object_detection\/utils\/visualization_utils.py\", line 41, in &lt;module&gt;\n    from object_detection.core import standard_fields as fields\nModuleNotFoundError: No module named 'object_detection.core'; 'object_detection' is not a package```\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57637985,"AnswerCount":1,"Tags":"<tensorflow><keras><keras-layer>","CreationDate":"2019-08-24T12:32:41.933","AcceptedAnswerId":null,"OwnerUserId":7066799.0,"Title":"How to change Sequential model to Custom Class model","Body":"<p>I'm learning tensorflow 2.0 from its older versions.\nI found tensorflow model is changed Sequential-base from Class-base. \nBut I want to use Class-base model because it is easy to read for me. <\/p>\n\n<p>I want to try translate : <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/keras\/basic_text_classification_with_tfhub\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/tutorials\/keras\/basic_text_classification_with_tfhub<\/a><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>embedding = 'https:\/\/tfhub.dev\/google\/tf2-preview\/gnews-swivel-20dim\/1'\nhub_layer = hub.KerasLayer(embedding,\n                           input_shape=[],\n                           dtype=tf.string,\n                           trainable=True)\n# hub_layer(train_example_batch[:3])\n\n\n# model = tf.keras.Sequential()\n# model.add(hub_layer)\n# model.add(tf.keras.layers.Dense(16, activation='relu'))\n# model.add(tf.keras.layers.Dense(1, activation='sigmoid'))\n\n\nclass MyModel(keras.Model):\n    def __init__(self, embedding):\n        super(MyModel, self).__init__()\n        self.embedding = embedding\n        self.d1 = keras.layers.Dense(16, activation='relu')\n        self.d2 = keras.layers.Dense(1, activation='sigmoid')\n\n    def call(self, x):\n        print(x.shape)\n        return reduce(lambda x, f: f(x), [x, self.embedding, self.d1, self.d2])\n\n\nmodel = MyModel(hub_layer)\n\n<\/code><\/pre>\n\n<p>I got below error message.<\/p>\n\n<pre><code>InvalidArgumentError: 2 root error(s) found.\n  (0) Invalid argument:  input must be a vector, got shape: [512,1]\n     [[{{node my_model_48\/keras_layer_7\/StatefulPartitionedCall\/StatefulPartitionedCall\/StatefulPartitionedCall\/tokenize\/StringSplit}}]]\n  (1) Invalid argument:  input must be a vector, got shape: [512,1]\n     [[{{node my_model_48\/keras_layer_7\/StatefulPartitionedCall\/StatefulPartitionedCall\/StatefulPartitionedCall\/tokenize\/StringSplit}}]]\n     [[my_model_48\/keras_layer_7\/StatefulPartitionedCall\/StatefulPartitionedCall\/StatefulPartitionedCall\/SparseFillEmptyRows\/SparseFillEmptyRows\/_24]]\n0 successful operations.\n0 derived errors ignored. [Op:__inference_keras_scratch_graph_303077]\n\nFunction call stack:\nkeras_scratch_graph -&gt; keras_scratch_graph\n<\/code><\/pre>\n\n<p>Why I got this error? And also, please answer whether WE NEED THROWING AWAY Class-base Model?<\/p>\n","answers":[{"AnswerId":"57672710","CreationDate":"2019-08-27T10:29:18.023","ParentId":null,"OwnerUserId":"7066799","Title":null,"Body":"<p>Here is the right code.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># model = tf.keras.Sequential()\n# model.add(hub_layer)\n# model.add(tf.keras.layers.Dense(16, activation='relu'))\n# model.add(tf.keras.layers.Dense(1, activation='sigmoid'))\n\n\nclass MyModel(keras.Model):\n    def __init__(self, embedding):\n        super(MyModel, self).__init__()\n        self.embedding = embedding\n        self.d1 = keras.layers.Dense(16, activation='relu')\n        self.d2 = keras.layers.Dense(1, activation='sigmoid')\n\n    def call(self, x):\n        # tf.sequeeze is needed! because x's dimention is [None, 1]. (1 was inserted without permission ...)\n        return reduce(lambda x, f: f(x), [x, tf.squeeze, self.embedding, self.d1, self.d2])\n\n\nmodel = MyModel(hub_layer)\n\n# model.summary()\nmodel.layers\n<\/code><\/pre>\n\n<p>It is because the officious framework has added extra features on their own... <\/p>\n\n<p>I don't like this attribute, but I think someone who like tensorflow <strong>eager<\/strong> to do it...<\/p>\n"}]}
+{"QuestionId":57638520,"AnswerCount":1,"Tags":"<python><numpy><pytorch>","CreationDate":"2019-08-24T13:40:51.700","AcceptedAnswerId":null,"OwnerUserId":11846982.0,"Title":"How to improve conditional tensor operation","Body":"<p>I have conditional operation of tensors, such as<\/p>\n\n<pre><code>for i:\n   if check(rate[i]):\n      rate[i] = reset(rate[i])\n<\/code><\/pre>\n\n<p>right now, i implemented as following,<\/p>\n\n<pre><code>rate = torch.where(check(rate), reset(rate), rate)\n<\/code><\/pre>\n\n<p>It turns out that even for tensor where check(rate) is false, reset(rate) is still operated, kind of like<\/p>\n\n<pre><code>new_rate = reset(rate)\nrate = torch.where(check(rate), new_rate, rate)\n<\/code><\/pre>\n\n<p>I am wondering anyway i could improve the performance here, if reset(rate) is very expensive<\/p>\n\n<p>Updated<\/p>\n\n<p>I've tried mask method, like following,<\/p>\n\n<pre><code>import torch\nimport time\nimport copy\n\ndef reset(t):\n    return 1.0 \/ torch.log(1.0 + t)\n\nsource1 = torch.FloatTensor(10000, 10000)\nsource1.uniform_()\nsource2 = copy.deepcopy(source1)\n\nstart = time.time()\nmask = source1 &lt; 0.5\nsource1[mask] = reset(source1[mask])\nprint('test1 takes: ', time.time() - start)\n\nstart = time.time()\nmask = source2 &lt; 0.5\nsource2 = torch.where(mask, reset(source2), source2)\nprint('test2 takes: ', time.time() - start)\n<\/code><\/pre>\n\n<p>On my machine, it shows,<\/p>\n\n<p>test1 takes:  3.455298900604248<\/p>\n\n<p>test2 takes:  1.8106980323791504<\/p>\n\n<p>Any way i could improve it?<\/p>\n","answers":[{"AnswerId":"57667316","CreationDate":"2019-08-27T03:24:47.007","ParentId":null,"OwnerUserId":"4385912","Title":null,"Body":"<p>The reason for this behavior is pretty straightforward and actually comes from python, not pytorch: when you pass in a function call into another function, the function call will be executed before you pass it in. That means that for your example, reset(rate) will be fully executed.<\/p>\n\n<p>To get around that, you might want to look into masking. Pytorch is quite efficient here!<\/p>\n\n<p>The general paradigm would be:<\/p>\n\n<pre><code>mask = check(rate)\nrate[mask] = reset(rate[mask])\n<\/code><\/pre>\n\n<p>The idea is that you compute a <code>torch.bool<\/code> or <code>torch.byte<\/code> mask (so the function <code>check<\/code> would need to return such a tensor) of the same size as your target tensor. Then, the same python behavior works in your favor, as <code>rate[mask]<\/code> will select only the relevant items first, and then apply the <code>reset<\/code> function.<\/p>\n"}]}
+{"QuestionId":57638862,"AnswerCount":1,"Tags":"<tensorflow><raspberry-pi>","CreationDate":"2019-08-24T14:30:45.933","AcceptedAnswerId":null,"OwnerUserId":9524309.0,"Title":"Need advice on tracking objects in hand using TensorFlow","Body":"<p>I am trying to develop a model which detects the items that are picked up by a user from a basket. Is this achievable using Tensorflow? My doubt is since the basket would contain the same items the user picks up (say fruits), is it possible to report the product that in the user's hand(the products that are picked up by the user) in real-time, rather than the items in the basket? Please advice on what would be a good starting point to achieve this. <\/p>\n\n<p>I have read and watched general object detection methods using Tensor Flow and various models but nothing seems to deal with a similar solution or I am unable to co-relate. If there are any tutorials on achieving, the links for the same would be even more helpful. Thanks in advance. Please bear with me if my question is naive, I am still a newbie at ML and Tensorflow.<\/p>\n","answers":[{"AnswerId":"57638979","CreationDate":"2019-08-24T14:45:50.867","ParentId":null,"OwnerUserId":"10209338","Title":null,"Body":"<p>It is achievable using Tensorflow. First consider how you would approach the problem. Since you need to distinguish between objects inside basket you will need ANN with some <strong>basic classification<\/strong> and <strong>object detection<\/strong>.<\/p>\n\n<p>Here are some links to get you started:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_classification\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_classification<\/a><\/li>\n<li><a href=\"https:\/\/www.edureka.co\/blog\/tensorflow-object-detection-tutorial\/\" rel=\"nofollow noreferrer\">https:\/\/www.edureka.co\/blog\/tensorflow-object-detection-tutorial\/<\/a><\/li>\n<\/ul>\n\n<p>Good luck!<\/p>\n"}]}
+{"QuestionId":57638922,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><time-series><autoencoder><anomaly-detection>","CreationDate":"2019-08-24T14:39:23.033","AcceptedAnswerId":null,"OwnerUserId":7035349.0,"Title":"How to train variational autoencoder(VAE) using a sliding window?","Body":"<p>I am trying to use variational autoencoder for anomaly detection problem in stock data. But I don't know how to train the model using sliding window. The idea is to take as input a time sequence and get as output an anomaly score. <\/p>\n\n<p>I have tried to test the solution which provided here <a href=\"https:\/\/thingsolver.com\/time-series-anomaly-detection-using-a-variational-autoencoder-vae\/\" rel=\"nofollow noreferrer\">https:\/\/thingsolver.com\/time-series-anomaly-detection-using-a-variational-autoencoder-vae\/<\/a>. <\/p>\n\n<p>Unfortunately, there is only half information, I don't see here the information about the input data format. <\/p>\n\n<p>By the way, I don't understand how to use the model after the training to get a reconstruction probability. The Kullback divergence..is it my reconstruction error? The difference between prior and posterior distribution? <\/p>\n\n<pre><code>\n        # Define the model.\n        self.prior = make_prior(code_size=self.z_dim)\n        x = tf.layers.flatten(self.x)\n        x = tf.layers.dense(x, self.dense_size, tf.nn.relu)\n        x = tf.layers.dense(x, self.dense_size, tf.nn.relu)\n        loc = tf.layers.dense(x, self.z_dim)\n        scale = tf.layers.dense(x, self.z_dim , tf.nn.softplus)\n        self.posterior = tfd.MultivariateNormalDiag(loc, scale)\n        self.code = self.posterior.sample()\n\n        # Define the loss.\n        x = self.code\n        x = tf.layers.dense(x, self.dense_size, tf.nn.relu)\n        x = tf.layers.dense(x, self.dense_size, tf.nn.relu)\n        loc = tf.layers.dense(x, self.n_dim)\n        scale = tf.layers.dense(x, self.n_dim , tf.nn.softplus)\n        self.decoder = tfd.MultivariateNormalDiag(loc, scale)\n        self.likelihood = self.decoder.log_prob(self.x)\n        self.divergence = tfd.kl_divergence(self.posterior, self.prior)\n        self.elbo = tf.reduce_mean(self.likelihood - self.divergence)\n        self._cost = -self.elbo\n\n\n        self.saver = tf.train.Saver(max_to_keep=5, keep_checkpoint_every_n_hours=1)\n        self.sess = tf.Session()\n\n\n<\/code><\/pre>\n\n<p>What does the result of loss function mean in the variational autoencoder? For example, I get the loss -15.987. How can I interpret that? <\/p>\n\n<p>Thank you in advance!<\/p>\n","answers":[]}
+{"QuestionId":57639334,"AnswerCount":1,"Tags":"<parsing><tensorflow><machine-learning><artificial-intelligence>","CreationDate":"2019-08-24T15:34:45.807","AcceptedAnswerId":57639423.0,"OwnerUserId":4447297.0,"Title":"What's a good ML model\/technique for breaking down large documents\/text\/HTML into segments?","Body":"<p>I'd like to break down HTML documents into small chunks of information. With sources like Wikipedia articles (as an example) this is reasonably easy to do, without machine learning, because the content is structured in a highly predictable way.<\/p>\n\n<p>When working with something like a converted Word Doc or a Blog post, the HTML is a bit more unpredictable. For example, sometimes there are no DIVs, more than one H1 in a document, or no Headers at all etc.<\/p>\n\n<p>I'm trying to figure out a decent\/reliable way of automatically putting content breaks into my content, in order to break it down into chunks of an acceptable size.<\/p>\n\n<p>I've had a little dig around for existing trained models for this application but I couldn't find anything off-the-shelf. I've considered training my own model, but I'm not confident of the best way to structure the training data. One option I've considered in relation to training data is providing a sample of where section breaks are numerically likely to exist within a document but I don't think that's the best possible approach... <\/p>\n\n<p>How would you approach this problem?<\/p>\n\n<hr>\n\n<p><em>P.s. I'm currently using Tensorflow but happy to go down a different path.<\/em><\/p>\n","answers":[{"AnswerId":"57639423","CreationDate":"2019-08-24T15:49:48.780","ParentId":null,"OwnerUserId":"10462788","Title":null,"Body":"<p>I've found the GROBID library quite robust for different input documents (since it's based on ML models trained on a large variety of documents). The standard model parses input PDF documents into structured XML\/TEI encoded files, which are much easier to deal with. <a href=\"https:\/\/grobid.readthedocs.io\/en\/latest\/Introduction\/\" rel=\"nofollow noreferrer\">https:\/\/grobid.readthedocs.io\/en\/latest\/Introduction\/<\/a><\/p>\n\n<p>If your inputs are HTML documents the library also offers the possibility to train your own models. Have a look at: <a href=\"https:\/\/grobid.readthedocs.io\/en\/latest\/Training-the-models-of-Grobid\/\" rel=\"nofollow noreferrer\">https:\/\/grobid.readthedocs.io\/en\/latest\/Training-the-models-of-Grobid\/<\/a><\/p>\n"}]}
+{"QuestionId":57639386,"AnswerCount":0,"Tags":"<tensorflow><cudnn>","CreationDate":"2019-08-24T15:44:22.113","AcceptedAnswerId":null,"OwnerUserId":1030294.0,"Title":"Issue installing Tensorflow with CuDNN","Body":"<p>I have installed Cuda-10.0, Cudnn 7.6.3, bazel 0.21.0 and trying to install Tensorflow 1.13 with GPU support.<\/p>\n\n<p>I have set LD_LIBRARY_PATH to the location of lib64 for cudnn and mentioned the full path in configure in Tensorflow. I get the following error.<\/p>\n\n<pre><code>\/usr\/bin\/ld: warning: libcudnn.so.7, needed by bazel-out\/host\/bin\/_solib_local\/_U_S_Stensorflow_Scontrib_Sboosted_Utrees_Cgen_Ugen_Usplit_Uhandler_Uops_Upy_Upy_Uwrappers_Ucc___Utensorflow\/libtensorflow_framework.so, not found (try using -rpath or -rpath-link)\n<\/code><\/pre>\n\n<p>I do not have sudo access, so other solutions that mention <code>sudo ldconfig<\/code> doesnt work for me.<\/p>\n\n<p>Also, the error says using  <code>rpath to set the library path<\/code>, but the Cudnn downloaded from Nvidia website has nothing to build, the lib file is already in <code>lib64<\/code> folder.<\/p>\n","answers":[]}
+{"QuestionId":57639434,"AnswerCount":1,"Tags":"<python><machine-learning><pytorch><tensor><autograd>","CreationDate":"2019-08-24T15:50:37.097","AcceptedAnswerId":57655719.0,"OwnerUserId":6688984.0,"Title":"Use PyTorch to adjust Tensor matrix values based on numbers I calculate from the Tensors?","Body":"<p>I have two tensors (matrices) that I've initialized:<\/p>\n\n<pre><code>sm=Var(torch.randn(20,1),requires_grad=True)\nsm = torch.mm(sm,sm.t()) \nfreq_m=Var(torch.randn(12,20),requires_grad=True)\n<\/code><\/pre>\n\n<p>I am creating two lists from the data inside these 2 matrices, and I am using spearmanr to get a correlation value between these 2 lists. How I am creating the lists is not important, but the goal is to adjust the values inside the matrices so that the calculated correlation value is as close to 1 as possible.<\/p>\n\n<p>If I were to solve this problem manually, I would tweak values in the matrices by .01 (or some small number) each time and recalculate the lists and correlation score. If the new correlation value is higher than the previous one, I would save the 2 matrices and tweak a different value until I get the 2 matrices that give me the highest correlation score possible.<\/p>\n\n<p>Is PyTorch capable of doing this automatically? I know PyTorch can adjust based on an equation but the way I want to adjust the matrix values is not against an equation, it's against a correlation value that I calculate. Any guidance with this is greatly appreciated!<\/p>\n","answers":[{"AnswerId":"57655719","CreationDate":"2019-08-26T10:01:18.097","ParentId":null,"OwnerUserId":"10511793","Title":null,"Body":"<p>Pytorch has an autograd package, that means if you have variable and you pass them through differentiable functions and get a scalar result, you can perform a gradient descent to update the variable to lower or augment the scalar result.<\/p>\n\n<p>So what you need to do is to define a function f that works on tensor level such that f(sm, freq_m) will give you the desired correlation.<\/p>\n\n<p>Then, you should do something like:<\/p>\n\n<pre><code>lr = 1e-3\nfor i in range(100):\n  # 100 updates\n  loss = 1 - f(sm, freq_m)\n  print(loss)\n  loss.backward()\n  with torch.no_grad():\n    sm -= lr * sm.grad\n    freq_m -= lr * freq_m.grad\n    # Manually zero the gradients after updating weights\n    sm.grad.zero_()\n    freq_m.grad.zero_()\n<\/code><\/pre>\n\n<p>The learning rate is basically the size of the step you do, a learning rate too high will cause the loss to explode, and a learning rate too little will cause a slow convergence, I suggest you experiment.<\/p>\n\n<p>Edit : To answer the comment on <code>loss.backward<\/code> : for any differentiable function f, f is a function of multiple tensors <code>t1, ..., tn<\/code> with <code>requires_grad=True<\/code> as a result, you can calculate the gradient of the loss with respect to each of those tensors. When you do <code>loss.backward<\/code>, it calculates those gradients and store those in t1.grad, ..., tn.grad. Then you update <code>t1, ..., tn<\/code> using gradient descent in order to lower the value of f. This update doesn't need a computational graph, so this is why you use <code>with torch.no_grad()<\/code>.\nAt the end of the loop, you zero the gradients because <code>.backward<\/code> doesn't overwrite the gradients but rather add the new gradients to them. More on that here : <a href=\"https:\/\/discuss.pytorch.org\/t\/why-do-we-need-to-set-the-gradients-manually-to-zero-in-pytorch\/4903\" rel=\"nofollow noreferrer\">https:\/\/discuss.pytorch.org\/t\/why-do-we-need-to-set-the-gradients-manually-to-zero-in-pytorch\/4903<\/a><\/p>\n"}]}
+{"QuestionId":57639931,"AnswerCount":1,"Tags":"<tensorflow><keras><lstm><recurrent-neural-network><softmax>","CreationDate":"2019-08-24T16:58:54.717","AcceptedAnswerId":null,"OwnerUserId":9493180.0,"Title":"Softmax predictions not summing to 1","Body":"<p>I am a beginner with rnns, i decided to print the probabilities in a small code i was writing and observed that they don't sum to 1. Consider this simplified code<\/p>\n\n<pre><code>from tensorflow import keras\nimport numpy as np \nif __name__ == '__main__': \nmodel = keras.Sequential((\nkeras.layers.SimpleRNN(5, activation=\"softmax\", input_shape=\n(1, 3)), \n)) \nX = [ [1, 2, 3], [4, 5, 6] ] \ny = [ [1, 0, 0, 0, 0], [0, 1, 0, 0, 0] ] \nX = np.array(X) \nX = np.reshape(X, (2, 1, 3)) \ny = np.array(y) \nprint(model.summary()) \nmodel.compile(optimizer=keras.optimizers.Adam\n(learning_rate=0.01),\nloss=\"categorical_crossentropy\")\nmodel.fit(X, y, epochs=100) \np = model.predict(X) \nprint(p) \np = list(np.squeeze(p)) \nprint(p) \nprint(np.sum(p,axis=1))\n<\/code><\/pre>\n\n<p>It's a simpleRnn with batch size of 2, 3 input features and only 1 timestep, the number of classes is 5. Here I'm getting the output of the last line as [1,1] as expected, the probabilities are summing to 1\nBut if i change the simple Rnn to<\/p>\n\n<pre><code>keras.layers.LSTM(5, activation=\"softmax\", input_shape= (1,3),recurrent_activation=\"softmax\")\n<\/code><\/pre>\n\n<p>The probabilities no longer sum to 1 why is that?\nPlease excuse the code indentation, I'm on mobile <\/p>\n","answers":[{"AnswerId":"57658515","CreationDate":"2019-08-26T13:06:17.530","ParentId":null,"OwnerUserId":"11597209","Title":null,"Body":"<p>As you can see in the picture below from <a href=\"https:\/\/colah.github.io\/posts\/2015-08-Understanding-LSTMs\/\" rel=\"nofollow noreferrer\">Colahs Blog<\/a> there is no final activation like in a Multilayer Perceptron. If u change the activation function from tanh to softmax, there will still be an elementwise multiplication afterwards. So the sum of the results won\u00b4t be 1. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Iie47.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Iie47.png\" alt=\"LSTM\"><\/a><\/p>\n\n<p>To solve this you need to add an activation function with a Softmax activation after your LSTM Layer: e.g. for 3 output neurons<\/p>\n\n<pre><code>model.add(Dense(3, activation='softmax'))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57640032,"AnswerCount":0,"Tags":"<keras><loss-function>","CreationDate":"2019-08-24T17:13:19.977","AcceptedAnswerId":null,"OwnerUserId":8523421.0,"Title":"In keras can we use add_loss along with compile losses?","Body":"<p>I want to find the KL divergence of some layer in between in my model, also I have multiple losses, so I am thinking of doing add_loss() for KL divergence and passing a list of losses for outputs in model.compile().<\/p>\n\n<p>Is it the right way or something like ignorance of one of the loss types will happen in keras?<\/p>\n","answers":[]}
+{"QuestionId":57640144,"AnswerCount":0,"Tags":"<python-3.x><deep-learning><pytorch>","CreationDate":"2019-08-24T17:28:05.687","AcceptedAnswerId":null,"OwnerUserId":11702228.0,"Title":"How to stabilize my Deep Q Network in pytorch?","Body":"<p>I am trying to teach my noisy deep-q network to produce a product for some days, havest it and exchange a filter when necessary.<\/p>\n\n<p>The goal is to maximize the rewards and run as long as possible in the given horizon.<\/p>\n\n<p>I created a test-environment. I am using a noisy deep q network with target and evaluation nets. But the network does not stabilize at all. Even with clipped rewards.<\/p>\n\n<p>Any suggestions, what is going wrong? I tried a PPO, which worked well to some extends, but was still below optimal. The networks seem to have problems with the filter exchange.<\/p>\n\n<p>My state variable: s = [Product,byproduct(not important in this example),Filter capacity,days after production start]<\/p>\n\n<p>Actions = 0:product, 1:exchange Filter, 2:harvest Product<\/p>\n\n<p>The horizon of the training is 120 steps. If the Filter is not exchanged before 0.2, then the episode is done.<\/p>\n\n<p>As you can see the steps are not staying around 120 as supposed.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/OY2Xd.png\" rel=\"nofollow noreferrer\">Rewards and steps per episode<\/a><\/p>\n\n<p>Any ideas what I could do? Is the state space too big?<\/p>\n\n<p>Thanks a lot for your input and help :)<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># -*- coding: utf-8 -*-\n\"\"\"\nnoisy network based on https:\/\/github.com\/cyoon1729\/Reinforcement-learning\/tree\/master\/Deep-Q-Networks\/noisyDQN\n\"\"\"\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.autograd as autograd\nimport torch.optim as optim\nimport torchvision.transforms as T\n\nimport numpy as np\nimport gym\nfrom itertools import count\nimport math\nimport random\nfrom collections import deque, namedtuple\nfrom test_env import TestEnv\nimport matplotlib.pyplot as plt\nimport csv\n\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n\nclass FactorizedNoisyLinear(nn.Module):\n\n    def __init__(self, num_in, num_out, is_training=True):\n        super(FactorizedNoisyLinear, self).__init__()\n        self.num_in = num_in\n        self.num_out = num_out \n        self.is_training = is_training\n\n        self.mu_weight = nn.Parameter(torch.FloatTensor(num_out, num_in)).to(device)\n        self.mu_bias = nn.Parameter(torch.FloatTensor(num_out)).to(device) \n        self.sigma_weight = nn.Parameter(torch.FloatTensor(num_out, num_in)).to(device)\n        self.sigma_bias = nn.Parameter(torch.FloatTensor(num_out)).to(device)\n        self.register_buffer(\"epsilon_i\", torch.FloatTensor(num_in))\n        self.register_buffer(\"epsilon_j\", torch.FloatTensor(num_out))\n\n        self.reset_parameters()\n        self.reset_noise()\n\n    def forward(self, x):\n        self.reset_noise()\n\n        if self.is_training:\n            epsilon_weight = self.epsilon_j.ger(self.epsilon_i)\n            epsilon_bias = self.epsilon_j\n            weight = self.mu_weight + self.sigma_weight.mul(autograd.Variable(epsilon_weight))\n            bias = self.mu_bias + self.sigma_bias.mul(autograd.Variable(epsilon_bias))\n        else:\n            weight = self.mu_weight\n            bias = self.mu_bias\n\n        y = F.linear(x, weight, bias)\n\n        return y\n\n    def reset_parameters(self):\n        std = 1 \/ math.sqrt(self.num_in)\n        self.mu_weight.data.uniform_(-std, std)\n        self.mu_bias.data.uniform_(-std, std)\n\n        self.sigma_weight.data.fill_(0.5 \/ math.sqrt(self.num_in))\n        self.sigma_bias.data.fill_(0.5 \/ math.sqrt(self.num_in))\n\n    def reset_noise(self):\n        eps_i = torch.randn(self.num_in).to(device)\n        eps_j = torch.randn(self.num_out).to(device)\n        self.epsilon_i = eps_i.sign() * (eps_i.abs()).sqrt()\n        self.epsilon_j = eps_j.sign() * (eps_j.abs()).sqrt()\n\n\nTransition = namedtuple('Transition',\n                        ('state', 'action', 'next_state', 'reward'))\n\n\nclass ReplayMemory(object):\n\n    def __init__(self, capacity):\n        self.capacity = capacity\n        self.memory = []\n        self.position = 0\n\n    def push(self, state,action,next_state,reward):\n        \"\"\"Saves a transition.\"\"\"\n        state = torch.from_numpy(state).float().unsqueeze(0).to(device)\n        action = torch.tensor(action, dtype=torch.long, device=device)\n        if next_state is not None:\n            next_state = torch.from_numpy(next_state).float().unsqueeze(0).to(device)\n        if len(self.memory) &lt; self.capacity:\n            self.memory.append(None)\n        self.memory[self.position] = Transition(state,action,next_state,reward)\n        self.position = (self.position + 1) % self.capacity\n\n    def sample(self, batch_size):\n        return random.sample(self.memory, batch_size)\n\n    def __len__(self):\n        return len(self.memory)\n\n\n\n\ndef clip_reward(state,reward):\n    x,y,z,e = state\n    if reward &gt; 0:\n        return x*1\n    else:\n        if reward &lt; -19000:\n            return 0.2\n        else:\n            return 0.5\n\n\n\ndef writecsv(line, ep):\n    with open('noisy_train\/noisy_output_'+str(ep)+'.csv', 'a') as newFile:\n            newFileWriter = csv.writer(newFile)\n            newFileWriter.writerow(line)    \n\ndef mini_batch_train(env, agent, max_episodes, max_steps, batch_size):\n    episode_rewards = []\n\n    for episode in range(max_episodes):\n        state = env.reset()\n        episode_reward = 0\n        real_cumreward = 0\n\n        for step in range(max_steps):\n            action = agent.get_action(state)\n            next_state, reward, done, _ = env.step(action)\n            real_cumreward += reward\n            if episode%10 == 0:\n                writecsv((action,state,next_state,reward,done,real_cumreward),episode)\n            reward = clip_reward(state,reward)\n            if step == max_steps-1 or done:\n                reward += 5\n            agent.replay_buffer.push(state, action, reward, next_state, done)\n            episode_reward += reward\n\n            if len(agent.replay_buffer) &gt; batch_size:\n                agent.update(batch_size)   \n\n            if done or step == max_steps-1:\n                episode_rewards.append(episode_reward)\n                print(\"Episode {} : {} ; steps : {}\".format(episode, real_cumreward,step))\n                break\n\n            state = next_state\n        if episode%20 == 0 and episode&gt;0:\n            plt.plot(episode_rewards)\n            plt.show()\n\n    return episode_rewards\n\n\nclass NoisyDQN(nn.Module):\n\n    def __init__(self, input_dim, output_dim):\n        super(NoisyDQN, self).__init__()\n        self.input_dim = input_dim\n        self.output_dim = output_dim\n\n        self.noisy_fc = nn.Sequential(\n            nn.Linear(self.input_dim[0], 252),\n            nn.ReLU(),\n            FactorizedNoisyLinear(252, 252),\n            nn.ReLU(),\n            FactorizedNoisyLinear(252, self.output_dim)\n        )\n\n    def forward(self, state):\n        qvals = self.noisy_fc(state).to(device)\n\n        return qvals\n\n\n\nclass NoisyDQNAgent:\n\n    def __init__(self, env, learning_rate=1e-4, gamma=0.99, buffer_maxlen=100000):\n        self.env = env\n        self.learning_rate = learning_rate\n        self.gamma = gamma\n        #self.replay_buffer = BasicBuffer(buffer_maxlen)\n        self.target_net_update_freq = 100\n        self.update_count = 0\n        self.memory = ReplayMemory(buffer_maxlen)\n\n        self.policy_net = NoisyDQN(self.env.observation_space.shape, self.env.action_space.n).to(device)\n        self.target_net = NoisyDQN(self.env.observation_space.shape, self.env.action_space.n).to(device)\n        self.target_net.load_state_dict(self.policy_net.state_dict())\n        self.target_net.eval()\n\n        self.MSE_loss = nn.MSELoss()\n        self.optimizer = optim.Adam(self.policy_net.parameters(), lr=learning_rate)\n\n\n    def select_action(self, state):\n        sample = random.random()\n        eps_threshold = EPS_END + (EPS_START - EPS_END) * \\\n            math.exp(-1. * self.update_count \/ EPS_DECAY)\n        self.update_count += 1\n        if sample &gt; eps_threshold:\n            with torch.no_grad():\n                state = autograd.Variable(torch.FloatTensor(state).unsqueeze(0)).to(device)\n                qvals = self.policy_net.forward(state).to(device)\n                action = np.argmax(qvals.detach().cpu().numpy())\n                return torch.tensor([[action]], device=device, dtype=torch.long)\n        else:\n            return torch.tensor([[random.randrange(self.env.action_space.n)]], device=device, dtype=torch.long)\n\n\n    def optimize_model(self):\n        if len(self.memory) &lt; BATCH_SIZE:\n            return\n        transitions = self.memory.sample(BATCH_SIZE)\n        # Transpose the batch (see https:\/\/stackoverflow.com\/a\/19343\/3343043 for\n        # detailed explanation). This converts batch-array of Transitions\n        # to Transition of batch-arrays.\n        batch = Transition(*zip(*transitions))\n\n        # Compute a mask of non-final states and concatenate the batch elements\n        # (a final state would've been the one after which simulation ended)\n        non_final_mask = torch.tensor(tuple(map(lambda s: s is not None,\n                                              batch.next_state)), device=device, dtype=torch.uint8)\n\n        non_final_next_states = torch.cat([s for s in batch.next_state\n                                                    if s is not None]).to(device)\n        state_batch = torch.cat(batch.state)\n        action_batch = torch.cat(batch.action)\n        reward_batch = torch.cat(batch.reward)\n\n        # Compute Q(s_t, a) - the model computes Q(s_t), then we select the\n        # columns of actions taken. These are the actions which would've been taken\n        # for each batch state according to policy_net\n        state_action_values = self.policy_net(state_batch).gather(1, action_batch)\n\n        # Compute V(s_{t+1}) for all next states.\n        # Expected values of actions for non_final_next_states are computed based\n        # on the \"older\" target_net; selecting their best reward with max(1)[0].\n        # This is merged based on the mask, such that we'll have either the expected\n        # state value or 0 in case the state was final.\n        next_state_values = torch.zeros(BATCH_SIZE, device=device)\n        next_state_values[non_final_mask] = self.target_net(non_final_next_states).max(1)[0].detach()\n        # Compute the expected Q values\n        expected_state_action_values = (next_state_values * GAMMA) + reward_batch\n\n        # Compute Huber loss\n        loss = F.smooth_l1_loss(state_action_values, expected_state_action_values.unsqueeze(1))\n        #loss = self.compute_loss(state_action_values, expected_state_action_values.unsqueeze(1))\n\n        # Optimize the model\n        self.optimizer.zero_grad()\n        loss.backward()\n        for param in self.policy_net.parameters():\n            param.grad.data.clamp_(-1, 1)\n        self.optimizer.step()\n\n\ndef plot_res(rewards,durations):\n    fig, (ax1, ax2)  = plt.subplots(1,2)\n    ax1.plot(rewards)\n    ax1.set_title('Rewards')\n    ax2.plot(durations)\n    ax2.set_title('Steps')\n    plt.show()\n\n\n#Hyper-Parameters\nBATCH_SIZE = 128\nGAMMA = 0.999\nEPS_START = 0.9\nEPS_END = 0.1\nEPS_DECAY = 500\nTARGET_UPDATE = 10\n\nhorizon = 120\n\nenv = TestEnv()\nagent = NoisyDQNAgent(env)\nepisode_durations = []\nepisode_rewards = []\n\nnum_episodes = 1000\nfor i_episode in range(num_episodes):\n    # Initialize the environment and state\n    state = env.reset()\n    total_reward = 0\n    for t in range(horizon+1):\n        # Select and perform an action\n        action = agent.select_action(state)\n        next_state, reward, done, _ = env.step(action.item())\n        reward = clip_reward(state,reward)\n        if done:\n            next_state = None\n            reward = -1.0\n        total_reward += reward\n        reward = torch.tensor([reward], device=device)\n        # Store the transition in memory\n        agent.memory.push(state, action, next_state, reward)\n\n        # Move to the next state\n\n        # Perform one step of the optimization (on the target network)\n        agent.optimize_model()\n        if done:\n            break\n        state = next_state\n    # Update the target network, copying all weights and biases in DQN\n    episode_durations.append(t)\n    episode_rewards.append(total_reward)\n    if i_episode % TARGET_UPDATE == 0:\n        agent.target_net.load_state_dict(agent.policy_net.state_dict())\n    if i_episode % 20 == 0 and i_episode&gt;0:\n        plot_res(episode_rewards,episode_durations)\n\nprint('Complete')\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code># -*- coding: utf-8 -*-\n\"\"\"\ntest_env.py\n\"\"\"\nimport math\nimport pickle\nimport gym\nfrom gym import spaces\nfrom gym.utils import seeding\nimport numpy as np\n\nimport random\nimport pandas as pd\nfrom scipy.interpolate import interp1d\n\npdfs = {0: [0.0, 0.0],\n 1: [0.24705, 0.04705],\n 2: [0.06470000000000001, 0.001966666666666663],\n 3: [0.21175000000000002, 0.007850000000000015],\n 4: [0.32944999999999997, 0.007850000000000005],\n 5: [0.19410000000000005, 0.009799999999999994],\n 6: [0.40590000000000004, 0.009799999999999994],\n 7: [0.45294999999999985, 0.009816666666666621],\n 8: [0.4647000000000001, 0.001966666666666672],\n 9: [0.54115, 0.003916666666666717],\n 10: [0.33529999999999993, 0.001966666666666598],\n 11: [0.017649999999999944, 0.005883333333333314],\n 12: [-0.035150000000000015, 0.003950000000000046],\n 13: [-0.058350000000000124, 0.019783333333333337],\n 14: [0.005300000000000082, 0.00216666666666665]}\n\npdfs2 = {'x': {0: 0,\n  1: 1,\n  2: 2,\n  3: 3,\n  4: 4,\n  5: 5,\n  6: 6,\n  7: 7,\n  8: 8,\n  9: 9,\n  10: 10,\n  11: 11,\n  12: 12,\n  13: 13,\n  14: 14},\n 'average': {0: 1.01,\n  1: 0.94,\n  2: 1.068333333,\n  3: 1.098333333,\n  4: 1.001666667,\n  5: 1.19,\n  6: 1.361666667,\n  7: 1.163333333,\n  8: 1.12,\n  9: 0.9,\n  10: 0.86,\n  11: 0.935,\n  12: 1.198333333,\n  13: 1.658333333,\n  14: 1.7783333330000002},\n 'Stdv': {0: 0.06333333299999999,\n  1: 0.05,\n  2: 0.058333333,\n  3: 0.085,\n  4: 0.058333333,\n  5: 0.173333333,\n  6: 0.5216666670000001,\n  7: 0.55,\n  8: 0.69,\n  9: 0.61,\n  10: 0.4,\n  11: 0.12166666699999999,\n  12: 0.12166666699999999,\n  13: 0.145,\n  14: 0.081666667}}\n\n\ndef filter_decay(x):\n    x = round(x)\n    points1 = np.array(  [[ 0.,          1.        ],\n                         [10.,          0.97964406],\n                         [20.,          0.92966907],\n                         [30.,          0.81829748],\n                         [40.,          0.67969409],\n                         [50.,          0.61382317]])\n\n    points2 = np.array([[ 0.,          1.        ],\n                         [10.,          0.94324923],\n                         [20.,          0.90694018],\n                         [30.,          0.78421843],\n                         [40.,          0.67743549],\n                         [50.,          0.58200272]])\n    # get x and y vectors\n    h1 = points1[:,0]\n    y1 = points1[:,1]\n    h2 = points2[:,0]\n    y2 = points2[:,1]\n    # calculate polynomial\n    f1 = interp1d(h1, y1)\n    f2 = interp1d(h2, y2)\n    a = (f1(50)\/(1-50\/85)\/85)\n    b = f1(50)\/(1-50\/85)\n    if x == 0:\n        return 1\n    if 0 &lt; x &lt;= 50:\n        return(random.uniform(f2(x),f1(x)))\n    if x &lt; 0:\n        return None\n    else:\n        return b-a*x+([-1,1][random.randrange(2)])*random.random()*(0.02)\n\n\nclass TestEnv(gym.Env):\n\n\n    def __init__(self):\n\n        self.pdfs = pdfs\n\n        self.pdfs2 = pdfs2\n\n        self.profit = 6000\n\n        self.Volume = 20000 #l\n        self.wv = 0.75 \n\n        self.CV = 484 #l\n        self.filtercost = 6000 #$\/l\n        self.dbc = 15 #g\/l\n\n        self.CIP = 12833.31\n        self.Capture = 20541.19\n        self.Polish = 14939.42\n        self.Labor = 351949.37\n\n        self.fixcosts = 18410.23\n\n\n        self.n_actions = 3\n        self.n_states = 4\n\n\n        self.X = 4\n        self.Y = 4\n        self.Z = 0.2\n        #60 Cycles Replacement\n        self.D = 60\n        self.d = 0\n        self.E = 6\n        self.T = 365*3\n        self.growthBound = 8\n\n\n        self.action_space = spaces.Discrete(3)\n        self.observation_space = spaces.Box(-np.inf,np.inf,shape=(4,),dtype=np.float32)\n\n        self.seed()\n        self.viewer = None\n        self.state = None\n\n        #self.steps_beyond_done = None\n\n    def seed(self, seed=None):\n        self.np_random, seed = seeding.np_random(seed)\n        return [seed]\n\n    def rewardfunc(self,state,action):\n        \"\"\"\n        Boundary Conditions\n        Gives us 3 decisions:\n            0 = Produce\n            1 = Exchange Filter\n            2 = Harvest Production\n        \"\"\"\n        x,y,z,e = state\n        if action == 0:\n            reward = -self.fixcosts #\u20ac per day\n        elif action == 1:\n            reward = -self.filtercost*self.CV-self.fixcosts\n        elif action == 2:\n            deltacycle = self.cyclenum(x,z)\n            reward = (self.profit*x*self.Volume*self.wv)-(self.CIP*deltacycle+self.Capture+self.Polish+self.Labor)-self.fixcosts*8\n        return reward\n\n\n    def step(self, action, reset=False):\n        \"\"\"\n        Boundary Conditions\n        Gives us 3 decisions:\n            0 = Produce\n            1 = Exchange Filter\n            2 = Harvest Production\n        \"\"\"\n        assert self.action_space.contains(action), \"%r (%s) invalid\"%(action, type(action))\n\n        state = self.state\n        x,y,z,e = state\n        if reset:\n            if e &gt; self.E:\n                pdftime = 0\n            else:\n                pdftime = e\n        else:\n            if e &gt; self.E:\n                pdftime = 0\n            else:\n                pdftime = e+8\n        alphax = np.random.normal(loc = self.pdfs[pdftime][0], scale=self.pdfs[pdftime][1])\n        betay = np.random.normal(loc = self.pdfs2[\"average\"][pdftime], scale=self.pdfs2['Stdv'][pdftime])\n        #cyclenum = random.uniform(0.01,0.001)\n        deltacycle = self.cyclenum(x,z)\n        if action == 0:\n            e += 1\n            x += alphax\n            y = betay\n            reward = self.rewardfunc(state,action)\n        elif action == 1:\n            e += 1\n            x += alphax\n            y = betay\n            z = 1\n            self.d = 0\n            reward = self.rewardfunc(state,action)\n        else:\n            #x,y,e = 0,betay,0 ##old, long env\n            self.d += deltacycle\n            z = filter_decay(self.d)\n            if z &lt;= 0:\n                z = 0.1\n            x,y,e = self.harvest_trans(z)\n            reward = self.rewardfunc(state,action)\n        self.state = np.array([x,y,z,e])\n        if reset:\n            done = False\n        else:\n            done = self.check_boundaries()\n        return self.state, reward, done, {}\n\n    def check_boundaries(self):\n        \"\"\"\n        Boundary Conditions\n        Gives us 3 decisions:\n            0 = Produce\n            1 = Exchange Filter\n            2 = Harvest Production\n        \"\"\"\n        x,y,z,e = self.state\n        if x &gt;= self.X or y &gt;= self.Y or e &gt;= self.E:\n            return True\n        elif z &lt;= self.Z:\n            return True\n        else:\n            return False\n\n\n    def harvest_trans(self,z):\n        #z = random.uniform(self.Z,1)\n\n        #if z &gt; 0.97:\n            #z = 1\n        #d = self.get_starting_d(z)\n        y = np.random.normal(loc = self.pdfs2[\"average\"][0], scale=self.pdfs2['Stdv'][1])\n        self.state = np.array([0,y,z,0])\n        for i in range(9):\n            next_state, reward, done, _ = self.step(0, reset=True)\n        x,y,z,e = self.state\n        e = 0\n        return x,y,e\n\n    def reset(self):\n        #z = random.uniform(self.Z,1)\n\n        #if z &gt; 0.97:\n            #z = 1\n        #d = self.get_starting_d(z)\n        self.d = 0\n        y = np.random.normal(loc = self.pdfs2[\"average\"][0], scale=self.pdfs2['Stdv'][1])\n        self.state = np.array([0,y,1,0])\n        for i in range(9):\n            next_state, reward, done, _ = self.step(0)\n        x,y,z,e = self.state\n        self.state = np.array([x,y,z,0])\n\n\n        return self.state\n\n    def cyclenum(self,x,z):\n        return math.ceil((x*self.Volume*self.wv)\/((z*self.dbc*self.CV)))\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57640378,"AnswerCount":1,"Tags":"<python><keras><keras-layer>","CreationDate":"2019-08-24T18:00:03.343","AcceptedAnswerId":null,"OwnerUserId":11430049.0,"Title":"How to properly set input_shape in a keras neural network?","Body":"<p>Basically, I have 3844 15x2 matrices, each one assigned to a binary target. So, <\/p>\n\n<pre><code>X_train shape is (3844, 15, 2)\ny_train shape is (3844, 1)\n<\/code><\/pre>\n\n<p>I have the following neural network:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(16, activation = 'relu', input_shape = (15, 2)))\nmodel.add(Dense(16, activation = 'relu'))\nmodel.add(Dense(1, activation = 'sigmoid'))\nmodel.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])\nepochs_hist = model.fit(X_train, y_train, epochs = 1000, batch_size = 4)\n<\/code><\/pre>\n\n<p>The model summary is:<\/p>\n\n<pre><code>model.sumary()\n\nLayer (type)                 Output Shape              Param #   \n\n=================================================================\n\ndense_1 (Dense)              (None, 15, 16)            48        \n\n_________________________________________________________________\n\ndense_2 (Dense)              (None, 15, 16)            272       \n\n_________________________________________________________________\n\ndense_3 (Dense)              (None, 15, 1)             17        \n\n=================================================================\n\nTotal params: 337\n\nTrainable params: 337\n\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>The error produced is: <strong>ValueError: Error when checking target: expected dense_3 to have 3 dimensions, but got array with shape (3844, 1).<\/strong><\/p>\n\n<p>Where is the error?<\/p>\n\n<p><strong>EDIT (Full code):<\/strong><\/p>\n\n<pre><code>window = 15\nret = 0.06\n\ndf = load_data()\n\nX = []\ny = []\nfor i in range(len(dataset) - window):\n    aux = dataset[i+1: i+window+1, 0:2]\n    X.append(dataset[i+1: i+window+1, 0:2])\n    if (aux.max()\/dataset[i, 0] - 1 &gt;= ret) and (dataset[i, 0]\/aux.min() - 1 &lt; ret):\n        y.append(1)\n    else:\n        y.append(0)\n\nX, y = np.array(X), np.array(y)\n\nX_train, X_test, y_train, y_test = train_test_split(X, y, test_size = 0.2, random_state = 6)\n\nscaler = MinMaxScaler(feature_range = (0, 1))\nX_train[:, :, 0] = scaler.fit_transform(X_train[:, :, 0])\nX_test[:, :, 0] = scaler.transform(X_test[:, :, 0])\n\ny_train = to_categorical(y_train)\ny_test = to_categorical(y_test)\nmodel.add(Dense(16, input_shape = (15, 2), activation = 'relu'))\nmodel.add(Dense(16, activation = 'relu'))\nmodel.add(Dense(1, activation = 'sigmoid'))\nmodel.summary()\nmodel.compile(optimizer = opt, loss = 'binary_crossentropy', metrics = ['accuracy'])\nepochs_hist = model.fit(X_train, y_train, epochs = 1000, batch_size = 4, validation_data = (X_test, y_test))\n\nExample data:\n\nX_train:\n\n0.298146    3.8201e+07\n0.287518    2.49463e+07\n0.282136    3.17904e+07\n0.269095    3.90852e+07\n0.262679    6.39347e+07\n0.252278    4.25771e+07\n0.242393    4.05355e+07\n0.246326    3.20741e+07\n0.247361    2.98584e+07\n0.252122    2.64514e+07\n0.247775    3.39687e+07\n\ny_train:\n\n1   0\n0   1\n0   1\n0   1\n0   1\n1   0\n0   1\n0   1\n0   1\n0   1\n<\/code><\/pre>\n","answers":[{"AnswerId":"57640668","CreationDate":"2019-08-24T18:41:26.060","ParentId":null,"OwnerUserId":"11651301","Title":null,"Body":"<p>Since you are passing in a matrix as an input you need to flatten them to make it compatible with the upstream layers. I have modified your code below:<\/p>\n\n<p><strong>Sample Data<\/strong><\/p>\n\n<pre><code>X_train = np.random.normal(size=(3844, 15, 2))\ny_train = np.random.binomial(n=1,p=0.5,size = (3844,1))\n<\/code><\/pre>\n\n<p><strong>Your code<\/strong><\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Flatten(input_shape=(15, 2)))\nmodel.add(Dense(16, activation = 'relu'))\nmodel.add(Dense(16, activation = 'relu'))\nmodel.add(Dense(1, activation = 'sigmoid'))\nmodel.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])\nepochs_hist = model.fit(X_train, y_train, epochs = 3, batch_size = 4)\n<\/code><\/pre>\n\n<p><strong>Sample Output<\/strong><\/p>\n\n<pre><code>Epoch 1\/3\n3844\/3844 [==============================] - 1s 281us\/step - loss: 0.4537 - acc: 0.7765\nEpoch 2\/3\n3844\/3844 [==============================] - 1s 282us\/step - loss: 0.4483 - acc: 0.7854\nEpoch 3\/3\n3844\/3844 [==============================] - 1s 281us\/step - loss: 0.4496 - acc: 0.7838\n<\/code><\/pre>\n\n<p>Hope this helps!<\/p>\n"}]}
+{"QuestionId":57640507,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras><conv-neural-network>","CreationDate":"2019-08-24T18:17:22.240","AcceptedAnswerId":null,"OwnerUserId":10905984.0,"Title":"How to get properly the output's shape when converting code from tf.nn.conv2d_transpose to tf.keras.layers.Conv2dTransopose","Body":"<p>I am converting some codes from tensorflow to tf.keras. I am used to upsample using tf.nn.conv2d_transpose which receive a output_shape parameter and everything works fine. When I swap into keras models I start using the Conv2DTranspose layer, but I can't get the desired output shape.<\/p>\n\n<p>For sake of simplicity lets assume my input shape is (38,25), then I use maxpooling operation with kernel shape=(2,2) and strides=(2,2) which output a (19,13) volume. Now is when the 'deconvolution' is applied with kernel_size=(4,4), padding='same' and stride=(2,2) which obviously return a (38,26) volume.<\/p>\n\n<p>I was trying all combinations of padding and output_padding and there's no way to get a volume of (38,25) as I did using tf.nn.conv2d_transpose(output_shape=(38,25)).<\/p>\n\n<p>This sample code shows indeed what I mean:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np \nfrom tensorflow.keras.layers import Conv2DTranspose, Input, MaxPool2D, ZeroPadding2D\nfrom tensorflow.keras import Model\n\nI = Input(shape=(38,25,3))\nX = MaxPool2D(pool_size=(2,2),strides=(2,2),padding='same')(I)\nY = []\n#getting all possible outputs using different padding and output_padding\nfor padding in ['same','valid']:\n    for i in range(2): # [0,..,stride-1]\n        for j in range(2):\n            Y.append(Conv2DTranspose(3,(4, 4),strides=(2,2), padding=padding,\\\n                            name=padding+str(i)+str(j),use_bias=False, output_padding=(i,j))(X))\n    Y.append(Conv2DTranspose(3,(4, 4),strides=(2,2), padding=padding,\\\n                            name=padding+'_no_output_padding',use_bias=False)(X))\nmodel = Model(inputs=I, outputs=Y, name = 'model')\nmodel.summary()\n<\/code><\/pre>\n\n<p>Which gives the following output:<\/p>\n\n<pre><code>input_1 (InputLayer)            (None, 38, 25, 3)    0                                            \n______________________________________________________________________________\nmax_pooling2d (MaxPooling2D)    (None, 19, 13, 3)    0           input_1[0][0]                    \n______________________________________________________________________________\nsame00 (Conv2DTranspose)        (None, 36, 24, 3)    144         max_pooling2d[0][0]              \n______________________________________________________________________________\nsame01 (Conv2DTranspose)        (None, 36, 25, 3)    144         max_pooling2d[0][0]              \n______________________________________________________________________________\nsame10 (Conv2DTranspose)        (None, 37, 24, 3)    144         max_pooling2d[0][0]              \n______________________________________________________________________________\nsame11 (Conv2DTranspose)        (None, 37, 25, 3)    144         max_pooling2d[0][0]              \n______________________________________________________________________________\nsame_no_output_padding (Conv2DT (None, 38, 26, 3)    144         max_pooling2d[0][0]              \n______________________________________________________________________________\nvalid00 (Conv2DTranspose)       (None, 40, 28, 3)    144         max_pooling2d[0][0]              \n______________________________________________________________________________\nvalid01 (Conv2DTranspose)       (None, 40, 29, 3)    144         max_pooling2d[0][0]              \n______________________________________________________________________________\nvalid10 (Conv2DTranspose)       (None, 41, 28, 3)    144         max_pooling2d[0][0]              \n______________________________________________________________________________\nvalid11 (Conv2DTranspose)       (None, 41, 29, 3)    144         max_pooling2d[0][0]              \n______________________________________________________________________________\nvalid_no_output_padding (Conv2D (None, 40, 28, 3)    144         max_pooling2d[0][0]  \n<\/code><\/pre>\n\n<p>Where name i j means the layer where the padding type is name and the output_padding used was (i,j). For example same01 means padding='same' and output_padding=(0,1). Note there's no combination which outputs the desired shape (38,25,3). <\/p>\n\n<p>Meanwhile using tf.nn.conv2d_transpose:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = np.arange(19*13*3).reshape([1,19,13,3]).astype('float32')\nw = np.random.normal(size=(4,4,3,3)).astype('float32')\ntf_op = tf.nn.conv2d_transpose(x,w,[1,38,25,3], strides=[1,2,2,1], padding='SAME')\nwith tf.Session() as sess:\n    print(sess.run(tf_op).shape)\n<\/code><\/pre>\n\n<p>Ouput:\n(1, 38, 25, 3)<\/p>\n\n<p>Maybe there's some step that tf.nn.conv2d_transpose is applying and tf.keras isn't.<\/p>\n\n<p>I hope someone help me. Thanks for reading.<\/p>\n","answers":[]}
+{"QuestionId":57640889,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><deep-learning><preprocessor>","CreationDate":"2019-08-24T19:16:26.593","AcceptedAnswerId":57644480.0,"OwnerUserId":11309950.0,"Title":"How to `scaler.transform` result of `model.predict` with multiple input LSTM?","Body":"<p>Here is my code:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>def create_dataset(signal_data, look_back=1):\n    dataX, dataY = [], []\n    for i in range(len(signal_data) - look_back):\n        dataX.append(signal_data[i:(i + look_back), :])\n        dataY.append(signal_data[i + look_back, -1])\n    return np.array(dataX), np.array(dataY)\n\nlook_back = 20\n\n\ndf = pd.read_csv('stock.csv')\n\nsignal_data = df[[\"close\", \"open\", \"high\", \"low\", \"volume\"]].values.astype('float32')\n\n\nscaler = MinMaxScaler()\nsignal_data = scaler.fit_transform(signal_data)\n\ntrain_size = int(len(signal_data) * 0.80)\ntest_size = len(signal_data) - train_size\ntrain = signal_data[0:train_size]\ntest = signal_data[train_size:]\n\nx_train, y_train = create_dataset(train, look_back)\nx_test, y_test = create_dataset(test, look_back)\n...\nmodel.add(LSTM)\n<\/code><\/pre>\n\n<p>When I perform <code>predict = model.predict(x_test)<\/code> and <code>scaler.transfrom(predict)<\/code>, an error occurs:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>on-broadcastable output operand with shape (5,1) doesn't match the broadcast shape (5,5)\n<\/code><\/pre>\n\n<p>The dataset has 5 features. When I use <code>model.predict()<\/code>, the output shape is (n, 1), but <code>scaler.transform<\/code> needs (n, 5) shape.<\/p>\n\n<p>If I <code>transform<\/code> before <code>model.predict<\/code>, I can't <code>transform<\/code> because dimension is 3, <\/p>\n\n<p>and if I <code>transform<\/code> after <code>model.predict<\/code>, shape is (n, 1) So I can't <code>transform<\/code>.<\/p>\n\n<p>So, I can't apply <code>scaler.transform<\/code> on <code>model.predict()<\/code>.<\/p>\n\n<p>How can I solve this problem?<\/p>\n","answers":[{"AnswerId":"57644480","CreationDate":"2019-08-25T08:37:08.670","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>Your question is little bit unclear, but based on the information you have provided, I am putting forth my observations.<\/p>\n\n<p>You have already transformed your entire dataset using the below code:<\/p>\n\n<pre><code>signal_data = scaler.fit_transform(signal_data)\n<\/code><\/pre>\n\n<p>So, you need not transform your <code>x_test<\/code> again. You can directly call the <code>model.predict<\/code> on it. If you are testing on a new data, which hasn't been transformed as above, then you will have to transform it. Make sure that this new testing dataset too has 5 features in it (and not 3) as your model has been trained on 5 features. <\/p>\n\n<p>Lastly, after the time series prediction is over, if you wish to transform back your scaled data to it's original version, you may use <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.MinMaxScaler.html#sklearn.preprocessing.MinMaxScaler.inverse_transform\" rel=\"nofollow noreferrer\"><code>inverse_transform<\/code><\/a> function of your <code>MinMaxScaler<\/code> instance.<\/p>\n"}]}
+{"QuestionId":57640899,"AnswerCount":0,"Tags":"<tensorflow><tensorboard>","CreationDate":"2019-08-24T19:18:07.150","AcceptedAnswerId":null,"OwnerUserId":11182272.0,"Title":"Are images in eval in Object detection cofiguration files used for cross validation?","Body":"<p>The object detection API supported by TensorFlow provides <code>eval_config<\/code> and <code>eval_input_reader<\/code>. Are these images used for validation? <\/p>\n\n<p>If not how does TensorFlow split training images into training and validation sets?<\/p>\n","answers":[]}
+{"QuestionId":57640956,"AnswerCount":0,"Tags":"<python><tensorflow><object-detection-api>","CreationDate":"2019-08-24T19:27:26.150","AcceptedAnswerId":null,"OwnerUserId":8603876.0,"Title":"ImportError: cannot import name 'string_int_label_map_pb2' (tried all solutions on the web)","Body":"<p>I'm using the tensorflow object_detection api on Windows 10, 64-bit, with Python 3.6, 64-bit. I have protobuf and all other necessary packages installed, and protoc and all other <code>\/model\/research files<\/code> in my path. I have <code>string_int_label_map_pb2.py<\/code> in my <code>object_detection\/protos<\/code> folder, but I keep getting the <code>ImportError: cannot import name 'string_int_label_map_pb2'<\/code>error. Interestingly, when I open up a new Python window, I still cannot import any of the protos files using the command <code>from object_detection.protos import (random file in protos here)<\/code>. I've tried all the solutions I can find on the Web.<\/p>\n\n<p>Error code:<\/p>\n\n<p><code>Traceback (most recent call last):\n  File \"model_main.py\", line 26, in &lt;module&gt;\n    from object_detection import model_lib\n  File \"C:\\Users\\f3d0m\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\object_detection\\model_lib.py\", line 27, in &lt;module&gt;\n    from object_detection import eval_util\n  File \"C:\\Users\\f3d0m\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\object_detection\\eval_util.py\", line 27, in &lt;module&gt;\n    from object_detection.metrics import coco_evaluation\n  File \"C:\\Users\\f3d0m\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\object_detection\\metrics\\coco_evaluation.py\", line 22, in &lt;module&gt;\n    from object_detection.utils import object_detection_evaluation\n  File \"C:\\Users\\f3d0m\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\object_detection\\utils\\object_detection_evaluation.py\", line 39, in &lt;module&gt;\n    from object_detection.utils import label_map_util\n  File \"C:\\Users\\f3d0m\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\object_detection\\utils\\label_map_util.py\", line 21, in &lt;module&gt;\n    from object_detection.protos import string_int_label_map_pb2\nImportError: cannot import name 'string_int_label_map_pb2'<\/code><\/p>\n","answers":[]}
+{"QuestionId":57641023,"AnswerCount":0,"Tags":"<python><tensorflow><object-detection><tfrecord>","CreationDate":"2019-08-24T19:38:35.680","AcceptedAnswerId":null,"OwnerUserId":1696354.0,"Title":"Tensorflow ObjectDetection NaN => Validate tfrecord","Body":"<p>I tagged about 1000 images using Microsoft\u00b4s VoTT and exported the tagged images to tfrecords.<\/p>\n\n<p>Now with the faster_rcnn_inception_v2_coco_2018_01_28 I\u00b4m getting a NaN error every time between training step 50 and 3000:<\/p>\n\n<pre><code>    I0824 21:33:31.837159 18516 coordinator.py:224] Error reported to Coordinator: &lt;class 'tensorflow.python.framework.errors_impl.InvalidArgumentError'&gt;, 2 root error(s) found.\n  (0) Invalid argument: LossTensor is inf or nan. : Tensor had NaN values\n         [[node CheckNumerics (defined at C:\\tensorflow1\\models\\research\\object_detection\\legacy\\trainer.py:322) ]]\n  (1) Invalid argument: LossTensor is inf or nan. : Tensor had NaN values\n         [[node CheckNumerics (defined at C:\\tensorflow1\\models\\research\\object_detection\\legacy\\trainer.py:322) ]]\n         [[gradients\/SecondStageFeatureExtractor\/InceptionV2\/Mixed_5c\/Branch_3\/Conv2d_0b_1x1\/BatchNorm\/FusedBatchNorm_grad\/FusedBatchNormGrad\/_2192]]\n0 successful operations.\n0 derived errors ignored.\n\nErrors may have originated from an input operation.\nInput Source operations connected to node CheckNumerics:\n total_loss (defined at C:\\tensorflow1\\models\\research\\slim\\deployment\\model_deploy.py:306)\n\nInput Source operations connected to node CheckNumerics:\n total_loss (defined at C:\\tensorflow1\\models\\research\\slim\\deployment\\model_deploy.py:306)\n\nOriginal stack trace for 'CheckNumerics':\n  File \"legacy\/train.py\", line 184, in &lt;module&gt;\n    tf.app.run()\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\", line 40, in run\n    _run(main=main, argv=argv, flags_parser=_parse_flags_tolerate_undef)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\absl\\app.py\", line 300, in run\n    _run_main(main, args)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\absl\\app.py\", line 251, in _run_main\n    sys.exit(main(argv))\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py\", line 324, in new_func\n    return func(*args, **kwargs)\n  File \"legacy\/train.py\", line 180, in main\n    graph_hook_fn=graph_rewriter_fn)\n  File \"C:\\tensorflow1\\models\\research\\object_detection\\legacy\\trainer.py\", line 322, in train\n    total_loss = tf.check_numerics(total_loss, 'LossTensor is inf or nan.')\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\ops\\gen_array_ops.py\", line 1205, in check_numerics\n    \"CheckNumerics\", tensor=tensor, message=message, name=name)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\framework\\op_def_library.py\", line 788, in _apply_op_helper\n    op_def=op_def)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py\", line 507, in new_func\n    return func(*args, **kwargs)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3616, in create_op\n    op_def=op_def)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 2005, in __init__\n    self._traceback = tf_stack.extract_stack()\n\nTraceback (most recent call last):\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1356, in _do_call\n    return fn(*args)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1341, in _run_fn\n    options, feed_dict, fetch_list, target_list, run_metadata)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1429, in _call_tf_sessionrun\n    run_metadata)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: 2 root error(s) found.\n  (0) Invalid argument: LossTensor is inf or nan. : Tensor had NaN values\n         [[{{node CheckNumerics}}]]\n  (1) Invalid argument: LossTensor is inf or nan. : Tensor had NaN values\n         [[{{node CheckNumerics}}]]\n         [[gradients\/SecondStageFeatureExtractor\/InceptionV2\/Mixed_5c\/Branch_3\/Conv2d_0b_1x1\/BatchNorm\/FusedBatchNorm_grad\/FusedBatchNormGrad\/_2192]]\n0 successful operations.\n0 derived errors ignored.\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"legacy\/train.py\", line 184, in &lt;module&gt;\n    tf.app.run()\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\", line 40, in run\n    _run(main=main, argv=argv, flags_parser=_parse_flags_tolerate_undef)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\absl\\app.py\", line 300, in run\n    _run_main(main, args)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\absl\\app.py\", line 251, in _run_main\n    sys.exit(main(argv))\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py\", line 324, in new_func\n    return func(*args, **kwargs)\n  File \"legacy\/train.py\", line 180, in main\n    graph_hook_fn=graph_rewriter_fn)\n  File \"C:\\tensorflow1\\models\\research\\object_detection\\legacy\\trainer.py\", line 416, in train\n    saver=saver)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\contrib\\slim\\python\\slim\\learning.py\", line 775, in train\n    train_step_kwargs)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\contrib\\slim\\python\\slim\\learning.py\", line 490, in train_step\n    run_metadata=run_metadata)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 950, in run\n    run_metadata_ptr)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1173, in _run\n    feed_dict_tensor, options, run_metadata)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1350, in _do_run\n    run_metadata)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1370, in _do_call\n    raise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: 2 root error(s) found.\n  (0) Invalid argument: LossTensor is inf or nan. : Tensor had NaN values\n         [[node CheckNumerics (defined at C:\\tensorflow1\\models\\research\\object_detection\\legacy\\trainer.py:322) ]]\n  (1) Invalid argument: LossTensor is inf or nan. : Tensor had NaN values\n         [[node CheckNumerics (defined at C:\\tensorflow1\\models\\research\\object_detection\\legacy\\trainer.py:322) ]]\n         [[gradients\/SecondStageFeatureExtractor\/InceptionV2\/Mixed_5c\/Branch_3\/Conv2d_0b_1x1\/BatchNorm\/FusedBatchNorm_grad\/FusedBatchNormGrad\/_2192]]\n0 successful operations.\n0 derived errors ignored.\n\nErrors may have originated from an input operation.\nInput Source operations connected to node CheckNumerics:\n total_loss (defined at C:\\tensorflow1\\models\\research\\slim\\deployment\\model_deploy.py:306)\n\nInput Source operations connected to node CheckNumerics:\n total_loss (defined at C:\\tensorflow1\\models\\research\\slim\\deployment\\model_deploy.py:306)\n\nOriginal stack trace for 'CheckNumerics':\n  File \"legacy\/train.py\", line 184, in &lt;module&gt;\n    tf.app.run()\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\", line 40, in run\n    _run(main=main, argv=argv, flags_parser=_parse_flags_tolerate_undef)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\absl\\app.py\", line 300, in run\n    _run_main(main, args)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\absl\\app.py\", line 251, in _run_main\n    sys.exit(main(argv))\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py\", line 324, in new_func\n    return func(*args, **kwargs)\n  File \"legacy\/train.py\", line 180, in main\n    graph_hook_fn=graph_rewriter_fn)\n  File \"C:\\tensorflow1\\models\\research\\object_detection\\legacy\\trainer.py\", line 322, in train\n    total_loss = tf.check_numerics(total_loss, 'LossTensor is inf or nan.')\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\ops\\gen_array_ops.py\", line 1205, in check_numerics\n    \"CheckNumerics\", tensor=tensor, message=message, name=name)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\framework\\op_def_library.py\", line 788, in _apply_op_helper\n    op_def=op_def)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py\", line 507, in new_func\n    return func(*args, **kwargs)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3616, in create_op\n    op_def=op_def)\n  File \"C:\\Users\\hankypanky\\AppData\\Local\\Continuum\\anaconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 2005, in __init__\n    self._traceback = tf_stack.extract_stack()\n<\/code><\/pre>\n\n<p>I read that this could be caused by invalid annotation of the tagged images.<\/p>\n\n<p>Is it possible to validate the existing tfrecord files if they are correctly annotated? \nOr could it be another problem?<\/p>\n","answers":[]}
+{"QuestionId":57641062,"AnswerCount":1,"Tags":"<python><macos><keras>","CreationDate":"2019-08-24T19:45:34.783","AcceptedAnswerId":null,"OwnerUserId":7570259.0,"Title":"OSError: cannot identify image file 'dataset\/training_set\/cats\/._cat.643.jpg'","Body":"<p>Do you all have any insights on why I could be getting this operating system error when trying to run a convolution neural network (CNN) with Keras? <\/p>\n\n<p>Also, when I run the code again, the same error is thrown but it is with a different image. <\/p>\n\n<p>Thank you!<\/p>\n\n<p>Here is the code, error message, and important screen shots:<\/p>\n\n<pre><code># Convolutional Neural Network\n\n# Installing Theano\n# pip install --upgrade --no-deps git+git:\/\/github.com\/Theano\/Theano.git\n\n# Installing Tensorflow\n# Install Tensorflow from the website: https:\/\/www.tensorflow.org\/versions\/r0.12\/get_started\/os_setup.html\n\n# Installing Keras\n# pip install --upgrade keras\n\n# 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\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 = 8000,\n                         nb_epoch = 25,\n                         validation_data = test_set,\n                         nb_val_samples = 2000)\n<\/code><\/pre>\n\n<p>Here is the error message:<\/p>\n\n<pre><code>  File \"\/Users\/anthonypendleton\/anaconda3\/lib\/python3.7\/site-packages\/PIL\/Image.py\", line 2822, in open\n    raise IOError(\"cannot identify image file %r\" % (filename if filename else fp))\n\n  OSError: cannot identify image file 'dataset\/training_set\/dogs\/._dog.3513.jpg'\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/AJ6yM.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/AJ6yM.png\" alt=\"Run 1\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/KEwJ9.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/KEwJ9.png\" alt=\"Run 2\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/jsUL0.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/jsUL0.png\" alt=\"Directory view from Finder\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/KfRXH.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/KfRXH.jpg\" alt=\"The image(s) open fine in Preview\"><\/a><\/p>\n","answers":[{"AnswerId":"57641746","CreationDate":"2019-08-24T21:44:46.420","ParentId":null,"OwnerUserId":"7475772","Title":null,"Body":"<p>you fail in macos temporary file error.  <code>._dog.3513.jpg<\/code> and <code>dog.3513.jpg<\/code> are different files. You need to clean your data folder.  <\/p>\n\n<p>Here is function to clean folder form hidden files with given extension:\nYou can use it before <code>flow_from_directory<\/code>. Be patient with it, to not remove needed data.  <\/p>\n\n<pre><code>from glob import glob\nimport os\n\ndef clean_directory(dir_path, ext=\".jpg\"):\n    files = glob(os.path.join(dir_path, \".*\" + ext))  # this line find all files witch starts with . and ends with given extension\n    for file_path in files:\n        os.remove(file_path)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57641353,"AnswerCount":1,"Tags":"<python><tensorflow><multidimensional-array><keras><valueerror>","CreationDate":"2019-08-24T20:36:14.993","AcceptedAnswerId":57641383.0,"OwnerUserId":5894206.0,"Title":"Input(shape=(6,7)) expects 3 dimensions on model.predict","Body":"<pre><code>ValueError: Error when checking input: \nexpected input_1 to have 3 dimensions, but got array with shape (6, 7)\n<\/code><\/pre>\n\n<pre><code>_____________________________________________________________________________\nLayer (type)                    Output Shape         Param #     Connected to\n==============================================================================\ninput_1 (InputLayer)            (None, 6, 7)         0\n<\/code><\/pre>\n\n<pre><code>\n    out1, out2 = model.predict(board)\n\n<\/code><\/pre>\n\n<pre><code>\n    inputs = Input(shape=(6,7))\n    inputs_reshape = Reshape((6,7,1))(inputs) # channels, batch_size, rows, cols\n    net = Conv2D(4, kernel_size=3, activation='relu', \n            padding='same', data_format='channels_last')(inputs_reshape)\n    net = Flatten()(net)\n    pi = Dense(7, activation='softmax', name='pi')(net) \n    v = Dense(1, activation='tanh', name='v')(net)\n\n    model = Model(inputs=inputs, outputs=[v, pi])\n<\/code><\/pre>\n\n<p>from the keras.io docs, it says that the <code>shape<\/code> dimensions for <code>Input()<\/code> does not include the batch size, and that <code>mdoel.predict()<\/code> sets <code>batch_size=32<\/code> by default. <\/p>\n\n<p>if <code>model.predict(data)<\/code> expects <code>data.shape<\/code> to be <code>(batches, 6,7)<\/code>, what's the difference between <code>model.predict(data, batch_size=1<\/code> and <code>model.predict_on_batch(data)<\/code><\/p>\n","answers":[{"AnswerId":"57641383","CreationDate":"2019-08-24T20:40:53.960","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Yes, the <code>batch_shape<\/code> of your model is <code>(None, 6, 7)<\/code>, three dimensions. The first value <code>None<\/code> is the batch size (which is free to be any value). <\/p>\n\n<p>So it's expecting your data to have 3 dimensions, as the <code>batch_shape<\/code> determines. <\/p>\n"}]}
+{"QuestionId":57641416,"AnswerCount":0,"Tags":"<tensorflow><neural-network><conv-neural-network><relu>","CreationDate":"2019-08-24T20:46:38.740","AcceptedAnswerId":null,"OwnerUserId":9079427.0,"Title":"Why does the output of the tensorflow convolutional layer output syas BiasAdd:0?","Body":"<p>I initialised a tensorflow convolutional layer and while debugging when I print the the layer I get the following output:<\/p>\n\n<pre><code>tf.Tensor 'q_eval\/conv3\/BiasAdd:0' shape=(?, 21, 21, 64) dtype=float32\n<\/code><\/pre>\n\n<p>Here, the q_eval is the name of the Network and the conv3 is the convolutional layer.<\/p>\n\n<p>What does BiasAdd:0 mean?<\/p>\n\n<pre><code>conv3 = tf.layers.conv2d(inputs=conv2_activated, filters=64,\n                                 kernel_size=(3, 3), strides=1, name='conv3',\n                                 kernel_initializer=tf.contrib.layers.variance_scaling_initializer(factor=2))\ntf.Tensor 'q_eval\/conv3\/BiasAdd:0' shape=(?, 21, 21, 64) dtype=float32\n\nconv3_activated = tf.nn.relu(conv3)\ntf.Tensor 'q_eval\/Relu_2:0' shape=(?, 21, 21, 64) dtype=float32\n<\/code><\/pre>\n\n<p>Does it mean the ReLu and Bias are not used and is always zero?<\/p>\n","answers":[]}
+{"QuestionId":57641482,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-24T20:56:08.280","AcceptedAnswerId":null,"OwnerUserId":1787751.0,"Title":"Predict decisions based on pre-processed data","Body":"<p>I'm currently pre-processing a set of web pages (to remove extraneous content), converting to tf-idf and subsequently training based on this. <\/p>\n\n<p>The thing is webpages with lesser than a certain limit were not used for training purposes. I don't want TensorFlow to use it as well and automatically predict these low-size web pages as false based on this information while it is pre-processing. <\/p>\n\n<p>Is this use case supported by TensorFlow? The plan is to serve this model via their REST API (tensorflow\/serving) and end-to-end (preprocess to prediction) support would be preferable. <\/p>\n","answers":[]}
+{"QuestionId":57642009,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2019-08-24T22:42:11.700","AcceptedAnswerId":null,"OwnerUserId":2146141.0,"Title":"Why do I get inconsistent loss behavior for my deep model?","Body":"<p>I am working on an image classification results. My training and testing split used the same random_state. Model definition is the same. However, when I run the model for multiple times, three out of four times, the model is not learning, the loss function does not go down; one out of four times, the model is learning, I can get good classificaiton results. I suspect the randomness comes from the ImageDataGenerator(). But I cannot figure out how to let the model learn every time. <\/p>\n\n<p>I have a relative small labeled dataset, I don't have ways to increase the data size<\/p>\n\n<p>I tried different optimizers, different batch size. It doesn't help. I found that when I reduce the trainable layers and make the later fully-connected layers smaller (reduce to 256 units), the model start to learn every time. But why big network does not learn well even on the training data set??? My understanding is that the model will overfit, but why in this case, it is not learning at all?<\/p>\n\n<pre><code>IMAGE_WIDTH=128\nIMAGE_HEIGHT=128\nIMAGE_SIZE=(IMAGE_WIDTH, IMAGE_HEIGHT)\nIMAGE_CHANNELS=3 # RGB color\n\nos.chdir(r\"XXX\")\nfilenames = os.listdir(r\"XXX\")\nref_db= pd.read_csv(r\"XXX\")\nref_db['obj_id']= [str(i)+ '.tif' for i in  ref_db.OBJECTID.values  ]\nref_db2= ref_db[['label', 'obj_id' ]]\nref_db2['label'] = ref_db2['label'].apply(str)\n\n\ntrain_df, validate_df = train_test_split(ref_db2, test_size=0.20, random_state=42)\ntrain_df = train_df.reset_index(drop=True)\nvalidate_df = validate_df.reset_index(drop=True)\n\ntotal_train = train_df.shape[0]\ntotal_validate = validate_df.shape[0]\nbatch_size=64   \n\ntrain_datagen = ImageDataGenerator(\n    rotation_range=15,\n    rescale=1.\/255,\n    shear_range=0.1,\n    zoom_range=0.2,\n    horizontal_flip=True,\n    width_shift_range=0.1,\n    height_shift_range=0.1\n)\n\ntrain_generator = train_datagen.flow_from_dataframe(\n    train_df, \n    r\"XXX\", \n    x_col='obj_id',\n    y_col='Green_Roof',\n    target_size=IMAGE_SIZE,\n    class_mode='binary',\n    batch_size=batch_size\n)                                       \n\n\ninputs= Input(shape=(IMAGE_WIDTH, IMAGE_HEIGHT, 3))\nbase_model = VGG19(weights='imagenet', include_top=False,)\nfor layer in base_model.layers[:-3]:\n    layer.trainable = False\n\nx = base_model(inputs)\nx = Flatten()(x)\nx = Dense(1024, activation=\"relu\")(x)\n#x = Dropout(0.5)(x)\nx = Dense(512, activation=\"relu\")(x)\npredictions = Dense(1, activation=\"sigmoid\")(x)    \nmodel_vgg= Model(inputs=inputs , outputs=predictions)\nmodel_vgg.compile(optimizer='Adam', loss='binary_crossentropy', metrics=['accuracy'])\n\n#########################\nhistory = model_vgg.fit_generator(\n    train_generator, \n    epochs=50,\n    validation_data=validation_generator,\n    validation_steps=total_validate\/\/batch_size,\n    steps_per_epoch=total_train\/\/batch_size,\n    verbose=2\n)\n\n\n\n<\/code><\/pre>\n\n<p>This is the unwanted behavior, Model is not learning, all observations are predicted as 1, the loss is not dropping<\/p>\n\n<pre><code>Found 756 validated image filenames belonging to 2 classes.\nFound 190 validated image filenames belonging to 2 classes.\nEpoch 1\/50\n - 4s - loss: 4.0464 - acc: 0.6203 - val_loss: 4.9820 - val_acc: 0.6875\nEpoch 2\/50\n - 2s - loss: 4.3811 - acc: 0.7252 - val_loss: 4.8856 - val_acc: 0.6935\nEpoch 3\/50\n - 2s - loss: 5.0209 - acc: 0.6851 - val_loss: 5.3556 - val_acc: 0.6641\nEpoch 4\/50\n - 2s - loss: 4.3583 - acc: 0.7266 - val_loss: 4.1142 - val_acc: 0.7419\nEpoch 5\/50\n - 2s - loss: 4.9317 - acc: 0.6907 - val_loss: 4.7329 - val_acc: 0.7031\nEpoch 6\/50\n - 2s - loss: 4.6275 - acc: 0.7097 - val_loss: 5.3998 - val_acc: 0.6613\nEpoch 7\/50\n<\/code><\/pre>\n\n<p>This is the expected behavior, model is learning, both 1 and 0 are predicted, the loss is dropping<\/p>\n\n<pre><code>Found 756 validated image filenames belonging to 2 classes.\nFound 190 validated image filenames belonging to 2 classes.\nEpoch 1\/50\n - 4s - loss: 2.1181 - acc: 0.6484 - val_loss: 0.8013 - val_acc: 0.6562\nEpoch 2\/50\n - 2s - loss: 0.6609 - acc: 0.7096 - val_loss: 0.5670 - val_acc: 0.7581\nEpoch 3\/50\n - 2s - loss: 0.6539 - acc: 0.6912 - val_loss: 0.5923 - val_acc: 0.6953\nEpoch 4\/50\n - 2s - loss: 0.5695 - acc: 0.7083 - val_loss: 0.5426 - val_acc: 0.6774\nEpoch 5\/50\n - 2s - loss: 0.5262 - acc: 0.7176 - val_loss: 0.5386 - val_acc: 0.6875\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57642019,"AnswerCount":3,"Tags":"<python><pip><pytorch>","CreationDate":"2019-08-24T22:45:14.803","AcceptedAnswerId":null,"OwnerUserId":11798174.0,"Title":"PyTorch not downloading","Body":"<p>I go to the PyTorch website and select the following options<\/p>\n\n<p>PyTorch Build: Stable (1.2)<\/p>\n\n<p>Your OS: Windows<\/p>\n\n<p>Package: pip<\/p>\n\n<p>Language: Python 3.7<\/p>\n\n<p>CUDA: None<\/p>\n\n<p>(All of these are correct)<\/p>\n\n<p>Than it displays a command to run<\/p>\n\n<p><code>pip3 install torch==1.2.0+cpu torchvision==0.4.0+cpu -f https:\/\/download.pytorch.org\/whl\/torch_stable.html<\/code><\/p>\n\n<p>I have already tried to mix around the the different options but none of them has worked.<\/p>\n\n<hr>\n\n<p>ERROR: <code>ERROR: Could not find a version that satisfies the requirement torch==1.2.0+cpu (from versions: 0.1.2, 0.1.2.post1, 0.1.2.post2)\nERROR: No matching distribution found for torch==1.2.0+cpu<\/code><\/p>\n\n<p>I tried to do pip install pytorch but pytorch doesn't support pypi<\/p>\n","answers":[{"AnswerId":"57797731","CreationDate":"2019-09-05T02:36:18.170","ParentId":null,"OwnerUserId":"11759191","Title":null,"Body":"<p>I've been in same situation. \nMy prob was,  the python version... I mean, in the 'bit' way.<\/p>\n\n<p>It was 32 bit that the python I'd installed.\nYou should check which bit of python you installed. \nyou can check in the app in setting, search python, then you will see the which bit you've installed.<\/p>\n\n<p>After I installed the 64 bit of Python, it solved.<\/p>\n\n<p>I hope you figure it out! <\/p>\n\n<p>environment : win 10<\/p>\n"},{"AnswerId":"57648698","CreationDate":"2019-08-25T18:31:56.353","ParentId":null,"OwnerUserId":"11798174","Title":null,"Body":"<p>So it looks like I cant install PyTorch because I am running python 32-bit. This may or may not be the problem but this is the only possible error that I could see this being the cause.<\/p>\n"},{"AnswerId":"57642037","CreationDate":"2019-08-24T22:48:58.073","ParentId":null,"OwnerUserId":"11908861","Title":null,"Body":"<p>It looks like it can't find a version called \"1.2.0+cpu\" from it's list of versions that it can find (0.1.2, 0.1.2.post1, 0.1.2.post2). Try looking for one of those versions on the PyTorch website.<\/p>\n"}]}
+{"QuestionId":57642137,"AnswerCount":0,"Tags":"<tensorflow><keras><deep-learning><time-series><lstm>","CreationDate":"2019-08-24T23:15:37.953","AcceptedAnswerId":null,"OwnerUserId":3215982.0,"Title":"Predicting a missing timestep in a time series given past and future ones","Body":"<p>Given a time series, I'm used to predicting future time-steps, usually using a <code>sliding window<\/code> method and an <code>LSTM\/GRU<\/code>. This is definitely common and everything works fine.<\/p>\n\n<p>In a context where I want to predict a missing time-step given the past and future ones, what is the best way to make use of all the information that we have?<\/p>\n\n<p>I'm used to train the model with batches of training examples, where each example is a window of <code>n_steps<\/code> entries and a <code>single label<\/code>. If I try to solve the missing <code>time-step<\/code> task, I find myself having this same exact setup, but how do I suggest the system that now the label corresponds to the entry in the middle of the window? Do I even need to do that?<\/p>\n\n<p>So my question is: How do I diversify the two <code>LSTM<\/code> setups for the two problems (future time-steps prediction and \"missing time-step\" prediction)? Because it appears to me that having both future and past time-steps available makes the task easier, but I'm not able to translate that into a model setup with better performance.<\/p>\n\n<p>Any suggestion would be appreciated, thanks.<\/p>\n","answers":[]}
+{"QuestionId":57642144,"AnswerCount":1,"Tags":"<python><tensorflow><keras><autoencoder>","CreationDate":"2019-08-24T23:18:01.083","AcceptedAnswerId":57642252.0,"OwnerUserId":10762045.0,"Title":"How do I create the decoder part of the Autoencoder?","Body":"<p>I am trying to copy the layers from the Encoder to create the decoder but I'm getting \"Index Error\".<\/p>\n\n<pre><code>input_img =Input(25425,)\n\nencoded1 = Dense(75,activation=tf.nn.relu)(input_img)\n\nencoded = Dense(50,activation=tf.nn.relu)(encoded1)\n\ndecoded = Dense(25425, activation='sigmoid')(encoded)\n\nautoencoder = Model(input_img, encoded1, decoded)\n\nencoder = Model(input_img, encoded)\n\nencoded_input = Input(shape=(50,))\n\n\ndecoder_layer1 = autoencoder.layers[1](encoded_input)\n\ndecoder_layer2 = autoencoder.layers[0](decoder_layer1)\n\ndecoder = Model(encoded_input, decoder_layer1, decoder_layer2)\n\nautoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')\n\nautoencoder.fit(X, X,\n            epochs=50,\n            shuffle=True)\n<\/code><\/pre>\n\n<p>I expect the decoder to have the same layers ans the encoder just reversed but I'm not able to copy the layers over. I am getting this Error:<\/p>\n\n<pre><code>Traceback (most recent call last):\nFile \"C:\\Users\\dalto\\Documents\\geo4\\train.py\", line 36, in &lt;module&gt;\ndecoder_layer1 = autoencoder.layers[1](encoded_input)\nIndexError: list index out of range\n<\/code><\/pre>\n","answers":[{"AnswerId":"57642252","CreationDate":"2019-08-24T23:42:49.960","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>You have several mistakes in you code. See my comments in the working snippet:<\/p>\n\n<pre><code># Random input for testing purposes\nX = np.random.rand(10, 25425)\n\ninput_img =tf.keras.layers.Input(25425,)\nencoded1 = tf.keras.layers.Dense(75,activation=tf.nn.relu)(input_img)\nencoded2 = tf.keras.layers.Dense(50,activation=tf.nn.relu)(encoded1)\ndecoded = tf.keras.layers.Dense(25425, activation='sigmoid')(encoded2)\n# The input of the autoencoder is the image (input_img), and the output is the decoder layer (decoded)\nautoencoder = tf.keras.Model(input_img, decoded)\n\nencoder = tf.keras.Model(input_img, encoded2)\n\nencoded_input = tf.keras.layers.Input(shape=(50,))\n# The decoded only consists of the last layer\ndecoder_layer = autoencoder.layers[-1](encoded_input)\n# The input to the decoder is the vector of the encoder which will be fed (using encoded_input), the output is the last layer of the network (decoder_layer)\ndecoder = tf.keras.Model(encoded_input, decoder_layer)\n\nautoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')\nautoencoder.fit(X, X, epochs=50, shuffle=True)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57642161,"AnswerCount":1,"Tags":"<python><pytorch><recurrent-neural-network>","CreationDate":"2019-08-24T23:23:08.987","AcceptedAnswerId":null,"OwnerUserId":3014199.0,"Title":"How to set the output size of an RNN?","Body":"<p>I want to have an RNN with input size 7, hidden size 10 and output size 2.\nSo for an input of, say, shape <code>99x1x7<\/code> I expect an output of shape <code>99x1x2<\/code>. \nFor an RNN alone, I get:<\/p>\n\n<pre><code>model = nn.RNN(input_size=7, hidden_size=10, num_layers=1)\n\noutput,hn=model(torch.rand(99,1,7))\nprint(output.shape) #torch.Size([99, 1, 10])\nprint(hn.shape)     #torch.Size([ 1, 1, 10])\n<\/code><\/pre>\n\n<p>So I assume I still have to put a <code>Linear<\/code> behind it:<\/p>\n\n<pre><code>model = nn.Sequential(nn.RNN(input_size=7, hidden_size=10, num_layers=1),\n                      nn.Linear(in_features=10, out_features=2))\nmodel(torch.rand(99,1,7))\n\nTraceback (most recent call last):\n  File \"train_rnn.py\", line 80, in &lt;module&gt;\n    main()\n  File \"train_rnn.py\", line 25, in main\n    model(torch.rand(99,1,7))\n  File \"\/home\/...\/virtual-env\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/module.py\", line 493, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/home\/...\/virtual-env\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/container.py\", line 92, in forward\n    input = module(input)\n  File \"\/home\/...\/virtual-env\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/module.py\", line 493, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/home\/...\/virtual-env\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/linear.py\", line 92, in forward\n    return F.linear(input, self.weight, self.bias)\n  File \"\/home\/...\/virtual-env\/lib\/python3.6\/site-packages\/torch\/nn\/functional.py\", line 1404, 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>I guess this is because <code>Linear<\/code> receives the tuple that <code>RNN.forward<\/code> yields. But how am I supposed to combine the two?<\/p>\n","answers":[{"AnswerId":"57655494","CreationDate":"2019-08-26T09:47:56.820","ParentId":null,"OwnerUserId":"10511793","Title":null,"Body":"<p>From the pytorch doc <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html?highlight=rnn#torch.nn.RNN\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/nn.html?highlight=rnn#torch.nn.RNN<\/a><\/p>\n\n<p>the output is of shape <code>seq_len, batch, num_directions * hidden_size<\/code><\/p>\n\n<p>So depending on what you want you might add a fc layer to get an output of size 2.\nBasically, a <code>Sequential<\/code> will apply each model on top of the output of the next_one so you must either not use a <code>Sequential<\/code> or create a special Linear Layer that works on sequence, the following should work : <\/p>\n\n<pre><code>class seq_Linear(nn.module):\n  def __init__(self, linear):\n    self.linear = linear\n  # To apply on every hidden state\n  def forward(self, x):\n    return torch.stack([self.linear(hs) for hs in x])\n  # To apply on the last hidden state\n  def forward(self, x):\n    return self.linear(x[-1])\n<\/code><\/pre>\n\n<p>and replace your nn.Linear by a seq_Linear(nn.Linear) in your code.<\/p>\n\n<p>Edit : If you want to create a sequence of outputs of size 2, perhaps the best way would be to stack another RNN on top of your first one with input_size 10 and output_size 2, they should be stackable inside a <code>Sequential<\/code> without any trouble.<\/p>\n"}]}
+{"QuestionId":57642245,"AnswerCount":1,"Tags":"<python><tensorflow><semantic-segmentation><deeplab><mobilenet>","CreationDate":"2019-08-24T23:40:52.660","AcceptedAnswerId":58189757.0,"OwnerUserId":11972635.0,"Title":"My custom mobilenet trained model is not showing any results. What am I doing wrong?","Body":"<p>I started to learn ML using Tensorflow\/Deeplab. I tried to train my own model from scratch for clothes recognition using semantic segmentation with mobilenet_v2 model variant. But I don't get results.<\/p>\n\n<p>I'm using <a href=\"https:\/\/github.com\/tensorflow\/models\" rel=\"nofollow noreferrer\">tensorflow\/models<\/a> for tfrecord export and training. And <a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/models\/blob\/master\/research\/deeplab\/deeplab_demo.ipynb\" rel=\"nofollow noreferrer\">deeplab\/example<\/a> code for visualization and testing purpose (renamed locally as main.py), I modify some lines so I can get the local models and testing image.<\/p>\n\n<p>I'll show the process I followed:<\/p>\n\n<ol>\n<li>Download 100 JPEG images (I know is not that big, but I guess I can try it with this amount). Just for 1 class -> shirts<\/li>\n<li>Create the segmentation class PNG for each image.<\/li>\n<li>Create the files image sets definition for: train(85 filenames), trainval(100 filenames) and val(15 filenames).\n<br>So my \"pascal dataset\" directory has: ImageSets, JPEGImages and SegmentationClassPNG folders.<\/li>\n<li><p>Export the \"pascal dataset\" directory to tfrecord like this (I'm on \"models-master\/research\/deeplab\/datasets\" folder):<\/p>\n\n<pre><code>py build_voc2012_data.py --image_folder=\"pasc_imgs\/JPEGImages\" --semantic_segmentation_folder=\"pasc_imgs\/SegmentationClassPNG\" --list_folder=\"pasc_imgs\/ImageSets\" --image_format=\"jpg\" --output_dir=\"train\/tfrecord\" \n<\/code><\/pre>\n\n<ul>\n<li>this works fine, it generates *.tfrecord files on \"train\/tfrecord\"<\/li>\n<\/ul><\/li>\n<li><p>I edited \"models-master\/research\/deeplab\/data_generator.py\" like this: {'train': 85, 'trainval': 100, 'val': 15}, num_classes=2.<\/p><\/li>\n<li>Now time to start the training, (I'm on \"models-master\/research\/deeplab\"). I used 10000 steps, why? I proved with 30000 and takes like 30 hours with no results, so I reduce it with new params. I guess 10000 steps could show me something:\n\n<pre><code>py train.py --logtostderr  --training_number_of_steps=10000 --train_split=\"train\" --model_variant=\"mobilenet_v2\" --output_stride=16 --decoder_output_stride=4 --train_batch_size=1 --dataset=\"pascal_voc_seg\"  --train_logdir=\"datasets\/train\/deeplab_model_mn\" --dataset_dir=\"datasets\/train\/tfrecord\"\n<\/code><\/pre>\n\n<ul>\n<li>This step takes almost 8 hours (have a tiny GPU, so.. can't use it), and it generates the checkpoint, graph.pbtxt, and model.ckpt-XXX (10000 included) files. <\/li>\n<\/ul><\/li>\n<li>I exported the previous result with (I'm on \"models-master\/research\/deeplab\") this command line:\n\n<pre><code>py export_model.py --checkpoint_path=datasets\/train\/deeplab_model_mn\/model.ckpt-10000 --export_path=datasets\/train\/deeplab_inference_mn\/frozen_inference_graph.pb --model_variant=\"mobilenet_v2\" --output_stride=16 --num_classes=2\n<\/code><\/pre>\n\n<ul>\n<li>It creates the frozen graph (frozen_inference_graph.pb).<\/li>\n<\/ul><\/li>\n<li>Now run: <b>py main.py<\/b> (proof image and frozen_inference_graph.pb already imported)<\/li>\n<li>No results with my custom model. This last script works with pre-trained mobilenetv2_coco_voc_trainaug. Not with my custom model <\/li>\n<\/ol>\n\n<p><b>data_generator.py (edited lines):<\/b><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>_PASCAL_VOC_SEG_INFORMATION = DatasetDescriptor(\n    splits_to_sizes={\n        'train': 85,\n        'trainval': 100,\n        'val': 15,\n    },\n    num_classes=2,# 0:background, 1:shirt\n    ignore_label=255,\n)\n<\/code><\/pre>\n\n<p>Image example (1\/100) that I'm using for training (I used the <a href=\"https:\/\/github.com\/wkentaro\/labelme\" rel=\"nofollow noreferrer\">labelMe<\/a> utility):\n<br>\n<a href=\"https:\/\/i.stack.imgur.com\/YVIy3.jpg\" rel=\"nofollow noreferrer\">shirt_001.jpg<\/a>\n<br>\n<a href=\"https:\/\/i.stack.imgur.com\/P4WKU.png\" rel=\"nofollow noreferrer\">shirt_001.png<\/a><\/p>\n\n<p><br>\n<b>main.py result for mobilenetv2_coco_voc_trainaug (shirt as a person, that's ok) and my custom model<\/b> :\n<br>\n<a href=\"https:\/\/i.stack.imgur.com\/cBs5z.jpg\" rel=\"nofollow noreferrer\">mobilenetv2_coco_voc_trainaug result<\/a>\n<br>\n<a href=\"https:\/\/i.stack.imgur.com\/42S7R.jpg\" rel=\"nofollow noreferrer\">my custom model result<\/a><\/p>\n\n<p>As you can see, my model fails. I've been testing many combinations without success. <b>What should I do?<\/b> Thank you!<\/p>\n","answers":[{"AnswerId":"58189757","CreationDate":"2019-10-01T17:34:33.287","ParentId":null,"OwnerUserId":"3970544","Title":null,"Body":"<p>Ok, I had the same problem and after many attempts I've done it.\nFirst, you should make correct masks. If you use one class you should create the masks with <strong>indexed color map<\/strong>, and <strong>all pixels should be 0 or 1<\/strong>, 0 - background, 1 - mask (there're 255 colors in the indexed color map).\nSecond, you need a bigger dataset. I tried training using a dataset with ~200 images and got no results (even with a correct dataset) even on checkpoint-30k. But when I tried training using a dataset with 450 images I had some results only from ~9000 epoch. There was no improvement after the ~18000 epoch, but the results were plausible (though far from ideal). Then I was training a model with 1100 images, but the results were the same.<\/p>\n"}]}
+{"QuestionId":57642264,"AnswerCount":0,"Tags":"<python><parsing><tensorflow><coreml><firebase-mlkit>","CreationDate":"2019-08-24T23:46:53.257","AcceptedAnswerId":null,"OwnerUserId":8737789.0,"Title":"How to Fix: \"google.protobuf.message.DecodeError: Error parsing message\" when creating tensorflow text summary","Body":"<p>I am trying to run a script to get a text summary out of a tensorflow .pb model such as this:<\/p>\n\n<pre><code>    OPS counts:\n    Squeeze : 1\n    Softmax : 1\n    BiasAdd : 1\n    Placeholder : 1\n    AvgPool : 1\n    Reshape : 2\n    ConcatV2 : 9\n    MaxPool : 13\n    Sub : 57\n    Rsqrt : 57\n    Relu : 57\n    Conv2D : 58\n    Add : 114\n    Mul : 114\n    Identity : 231\n    Const : 298\n<\/code><\/pre>\n\n<p>I am overall trying to convert a .pb model to a .coremlmodel and am following this article: <\/p>\n\n<p><a href=\"https:\/\/hackernoon.com\/integrating-tensorflow-model-in-an-ios-app-cecf30b9068d\" rel=\"nofollow noreferrer\">https:\/\/hackernoon.com\/integrating-tensorflow-model-in-an-ios-app-cecf30b9068d<\/a><\/p>\n\n<p>Getting a text summary from the .pb model is a step towards that. The code I try to run to create the text summary follows:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.core.framework import graph_pb2\nimport time\nimport operator\nimport sys\n\ndef inspect(model_pb, output_txt_file):\n    graph_def = graph_pb2.GraphDef()\n    with open(model_pb, \"rb\") as f:\n    graph_def.ParseFromString(f.read())\n\n    tf.import_graph_def(graph_def)\n\n    sess = tf.Session()\n    OPS = sess.graph.get_operations()\n\n    ops_dict = {}\n\n    sys.stdout = open(output_txt_file, 'w')\n    for i, op in enumerate(OPS):\n            print('---------------------------------------------------------------------------------------------------------------------------------------------')\n        print(\"{}: op name = {}, op type = ( {} ), inputs = {}, outputs = {}\".format(i, op.name, op.type, \", \".join([x.name for x in op.inputs]), \", \".join([x.name for x in op.outputs])))\n        print('@input shapes:')\n        for x in op.inputs:\n            print(\"name = {} : {}\".format(x.name, x.get_shape()))\n        print('@output shapes:')\n        for x in op.outputs:\n            print(\"name = {} : {}\".format(x.name, x.get_shape()))\n        if op.type in ops_dict:\n            ops_dict[op.type] += 1\n        else:\n            ops_dict[op.type] = 1\n\n    print('---------------------------------------------------------------------------------------------------------------------------------------------')\nsorted_ops_count = sorted(ops_dict.items(),     key=operator.itemgetter(1))\n    print('OPS counts:')\n    for i in sorted_ops_count:\n        print(\"{} : {}\".format(i[0], i[1]))\n\nif __name__ == \"__main__\":\n\"\"\"\nWrite a summary of the frozen TF graph to a text file.\nSummary includes op name, type, input and output names and shapes. \n\nArguments\n----------\n- path to the frozen .pb graph\n- path to the output .txt file where the summary is written\n\nUsage\n----------\npython inspect_pb.py frozen.pb text_file.txt\n\n\"\"\"\nif len(sys.argv) != 3:\n    raise ValueError(\"Script expects two arguments. \" +\n          \"Usage: python inspect_pb.py \/path\/to\/the\/frozen.pb \/path\/to\/the\/output\/text\/file.txt\")\ninspect(sys.argv[1], sys.argv[2])\n<\/code><\/pre>\n\n<p>I ran this command:<\/p>\n\n<pre><code> python inspect_pb.py \/Users\/nikhil.c\/Desktop\/tensorflowModel.pb   text_summary.txt\n<\/code><\/pre>\n\n<p>But instead of receiving the expected output, I receive this error message:<\/p>\n\n<pre><code>    Traceback (most recent call last):\n      File \"inspect_pb.py\", line 58, in &lt;module&gt;\n        inspect(sys.argv[1], sys.argv[2])\n      File \"inspect_pb.py\", line 10, in inspect\n        graph_def.ParseFromString(f.read())\n    google.protobuf.message.DecodeError: Error parsing message\n<\/code><\/pre>\n\n<p>and do not really know where to start. Other similar questions that seem to receive the same error message do not make too much sense. What do I do?<\/p>\n","answers":[]}
+{"QuestionId":57642361,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2019-08-25T00:15:35.573","AcceptedAnswerId":57643045.0,"OwnerUserId":7748163.0,"Title":"How to initialize the variables of the existing Keras Models?","Body":"<p>I would like to retrain Keras Model, Inception_v3, from scratch.<\/p>\n\n<p>The model is defined here:\n<a href=\"https:\/\/github.com\/keras-team\/keras-applications\/blob\/master\/keras_applications\/inception_v3.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras-applications\/blob\/master\/keras_applications\/inception_v3.py<\/a><\/p>\n\n<p>I read some posts,<\/p>\n\n<p>The listed solutions are:<\/p>\n\n<ol>\n<li><p>Freeze the layers (This is not what I want...)    <\/p>\n\n<pre><code>for layer in model.layers:\n    layer.trainable = False\n<\/code><\/pre><\/li>\n<li><p><a href=\"https:\/\/stackoverflow.com\/a\/51727616\/7748163\">https:\/\/stackoverflow.com\/a\/51727616\/7748163<\/a> <\/p><\/li>\n<\/ol>\n\n<p>Reset all layers by checking for initializers:<\/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        if hasattr(layer, 'bias_initializer'):\n            layer.bias_initializer.run(session=session)     \n<\/code><\/pre>\n\n<ol start=\"3\">\n<li><p>Use <code>tf.variables_initializer<\/code><\/p>\n\n<pre><code>  model = InceptionV3()\n  for layer in model.layers:\n      sess.run(tf.variables_initializer(layer.weights))\n<\/code><\/pre>\n\n<p>Reference: <a href=\"https:\/\/stackoverflow.com\/a\/56634827\/7748163\">https:\/\/stackoverflow.com\/a\/56634827\/7748163<\/a><\/p><\/li>\n<li><p>The best one I think, but it raises an error.<\/p>\n\n<pre><code>sess = tf.Session()\nfor layer in model.layers: \n     for v in layer.__dict__:\n         v_arg = getattr(layer,v)\n         if hasattr(v_arg,'initializer'):\n             initializer_method = getattr(v_arg, 'initializer')\n             initializer_method.run(session=sess)\n             print('reinitializing layer {}.{}'.format(layer.name, v))\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p><strong>However, none of them works for Inception_v3.<\/strong><\/p>\n\n<p>The error information is for BatchNorm layer:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.FailedPreconditionError: Error while reading resource variable batch_normalization_9\/moving_mean from Container: localhost. This could mean that the variable was uninitialized. Not found: Resource localhost\/batch_normalization_9\/moving_mean\/N10tensorflow3VarE does not exist.\n     [[{{node batch_normalization_9_1\/AssignMovingAvg\/ReadVariableOp}}]]\n     [[metrics_1\/categorical_accuracy\/Identity\/_469]]\n<\/code><\/pre>\n\n<p>So, how to re-train the existing Keras Models, and initialize the variables? What is the best practice to re-train a model from Keras applications?<\/p>\n\n<p>Further discussion:<\/p>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/341\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/341<\/a><\/p>\n","answers":[{"AnswerId":"57643045","CreationDate":"2019-08-25T03:35:43.297","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Why not simply not asking for the weights?<\/p>\n\n<pre><code>model = Inception_V3(..., weights=None,...)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras-applications\/blob\/master\/keras_applications\/inception_v3.py\/#L100\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras-applications\/blob\/master\/keras_applications\/inception_v3.py\/#L100<\/a><\/p>\n"}]}
+{"QuestionId":57642762,"AnswerCount":1,"Tags":"<python><tensorflow><keras><autoencoder>","CreationDate":"2019-08-25T02:13:05.280","AcceptedAnswerId":null,"OwnerUserId":10762045.0,"Title":"How to fix loss not going down when training","Body":"<p>When I'm training my autoencoder the loss doesn't change at all no matter how much I train it.<\/p>\n\n<pre><code>#Importing training data\ninp = open('train.csv',\"rb\")\nX = pickle.load(inp)\nX = X\/255.0\nX = np.array(X)\nX = np.reshape(X,(-1,25425))\n\n\ninput_img =tf.keras.layers.Input(25425,)\nencoded1 = tf.keras.layers.Dense(75,activation=tf.nn.relu)(input_img)\nencoded2 = tf.keras.layers.Dense(50,activation=tf.nn.relu)(encoded1)\ndecoded = tf.keras.layers.Dense(25425, activation='sigmoid')(encoded2)\n# The input of the autoencoder is the image (input_img), and the output is the decoder layer (decoded)\nautoencoder = tf.keras.Model(input_img, decoded)\n\nencoder = tf.keras.Model(input_img, encoded2)\n\nencoded_input = tf.keras.layers.Input(shape=(50,))\n# The decoded only consists of the last layer\ndecoder_layer = autoencoder.layers[-1](encoded_input)\n# The input to the decoder is the vector of the encoder which will be fed (using encoded_input), the output is the last layer of the network (decoder_layer)\ndecoder = tf.keras.Model(encoded_input, decoder_layer)\n\nautoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')\nautoencoder.fit(X, X, epochs=50, shuffle=True)\n\n# Save the weights\nautoencoder.save_weights('model_weights.h5')\n\n# Save the model architecture\nwith open('model_architecture.json', 'w') as f:\n    f.write(autoencoder.to_json())\n<\/code><\/pre>\n\n<p>I am hoping that I can get the training to work better But my loss is stuck at 0.6932<\/p>\n","answers":[{"AnswerId":"57649341","CreationDate":"2019-08-25T20:07:43.953","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>Autoencoders are tasked with <em>compressing input size<\/em> while <em>minimizing information loss<\/em>. What your decoder attempts to do is to reconstruct <code>25425<\/code> points from only <code>50<\/code> points - that is over a <strong>500-fold<\/strong> compression, which is highly unrealistic. I also assume your input is normalized between <code>(0,1)<\/code>, as the sigmid cannot reach beyond.<\/p>\n\n<p><strong>Solution<\/strong>: change decoder <em>and<\/em> encoder # of units so that the ratio, <code>input_dim\/min_units<\/code>, is \"reasonable\" (say, 5:1). <code>min_units<\/code> = # of units of the layer with least units (encoder or decoder) - in your case, 50. (<em>Note<\/em>: <code>25425<\/code> is quite large - consider dimensionality reduction via downsampling (e.g. <code>MaxPooling<\/code>), or using convolutional autoencoders)<\/p>\n\n<p>Regarding ReLU with sigmoid, fine to use both, but take some caution with exploding gradients\/activations in ReLU via e.g. weight decay - see <a href=\"https:\/\/stats.stackexchange.com\/questions\/144733\/can-i-use-relu-in-autoencoder-as-activation-function\">this discussion<\/a>.<\/p>\n"}]}
+{"QuestionId":57642985,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-08-25T03:19:46.853","AcceptedAnswerId":null,"OwnerUserId":11567036.0,"Title":"It does not predict one value as the module, but output size of input size","Body":"<p>Full code:<a href=\"https:\/\/colab.research.google.com\/drive\/1hIL6KVdAvs0MSALeNNlyFe_NnoAqdwlr\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/1hIL6KVdAvs0MSALeNNlyFe_NnoAqdwlr<\/a><\/p>\n\n<p>With this code:<\/p>\n\n<pre><code>p = np.expand_dims(series[3590:3650], -1)\nprint((model.predict(p[np.newaxis]).shape))\n<\/code><\/pre>\n\n<p>I should have the prediction of the tempature \"one day\" after 3650 time step.\nBut I do not understand the 60 times step it outputs because it should output one time step.<\/p>\n\n<pre><code>train_set = windowed_dataset(x_train, window_size=60, batch_size=100, shuffle_buffer=shuffle_buffer_size)\nmodel = tf.keras.models.Sequential([\ntf.keras.layers.Conv1D(filters=60, kernel_size=5,\n                  strides=1, padding=\"causal\",\n                  activation=\"relu\",\n                  input_shape=[None, 1]),\n  tf.keras.layers.LSTM(60, return_sequences=True),\n\n\ntf.keras.layers.LSTM(60, return_sequences=True),\n  tf.keras.layers.Dense(30, activation=\"relu\"),\n  tf.keras.layers.Dense(10, activation=\"relu\"),\n  tf.keras.layers.Dense(1),\n  tf.keras.layers.Lambda(lambda x: x * 400)\n])\n<\/code><\/pre>\n\n<p>prediction code:<\/p>\n\n<pre><code>p = np.expand_dims(series[3590:3650], -1)\nprint(p.shape)\nprint((model.predict(p[np.newaxis]).shape))\nmodel.predict(p[np.newaxis])\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57642990,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><deep-learning><conv-neural-network>","CreationDate":"2019-08-25T03:20:54.317","AcceptedAnswerId":null,"OwnerUserId":11920225.0,"Title":"Why my validation data is 0 images after compile?","Body":"<p>I tried the code from github \n<a href=\"https:\/\/github.com\/rohanpillai20\/Image-Classification-by-Keras-and-Tensorflow\/blob\/master\/Using%20Keras\/Training.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/rohanpillai20\/Image-Classification-by-Keras-and-Tensorflow\/blob\/master\/Using%20Keras\/Training.py<\/a><\/p>\n\n<p>for a data classification, but the dataset dataset for validation was 0 after training, while for the training it was readable after compile<\/p>\n\n<p>I wonder if that's really the case or something's wrong?<\/p>\n\n<pre><code>Found 470 images belonging to 2 classes.\nFound 0 images belonging to 0 classes.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57643066,"AnswerCount":1,"Tags":"<image><pytorch><record><tfrecord>","CreationDate":"2019-08-25T03:40:03.780","AcceptedAnswerId":57643464.0,"OwnerUserId":4941254.0,"Title":"PyTorch: wrapping multiple records in one file?","Body":"<p>Is there a standard way of encoding multiple records (in this case, data from multiple .png or .jpeg images) in one file that PyTorch can read? Something similar to TensorFlow's \"TFRecord\" or MXNet's \"RecordIO\", but for PyTorch.<\/p>\n\n<p>I need to download image data from S3 for inference, and it's much slower if my image data is in many small .jpg files rather than fewer files.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"57643464","CreationDate":"2019-08-25T05:22:14.707","ParentId":null,"OwnerUserId":"1849998","Title":null,"Body":"<p>One thing is to store batches of images together in a single <code>npz<\/code> file. Numpy's <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.savez.html\" rel=\"nofollow noreferrer\"><code>np.savez<\/code><\/a> lets you save multiple arrays compressed into a single file. Then load the file as np arrays and use <code>torch.from_numpy<\/code> to convert to tensors.<\/p>\n"}]}
+{"QuestionId":57643459,"AnswerCount":1,"Tags":"<python><tensorflow><keras><autoencoder><decoder>","CreationDate":"2019-08-25T05:21:11.297","AcceptedAnswerId":null,"OwnerUserId":10762045.0,"Title":"How to get the loss to change?","Body":"<p>The loss of this NN is getting stuck at 3.7666 but i can't tell why. I'm testing with about 50 training samples so would increasing this help improve the loss?<\/p>\n\n<pre><code>inp = open('train.csv',\"rb\")\nX = pickle.load(inp)\nX = X\/255.0\nX = np.array(X)\nX = np.reshape(X,(-1,25425))\n\n\ninput_img =tf.keras.layers.Input(25425,)\nencoded1 = tf.keras.layers.Dense(750,activation=tf.nn.relu)(input_img)\nencoded2 = tf.keras.layers.Dense(50,activation=tf.nn.relu)(encoded1)\ndecode_encoder = tf.keras.layers.Dense(750,activation=tf.nn.relu)(encoded2)\ndecoded = tf.keras.layers.Dense(25425, activation='softmax')(decode_encoder)\n# The input of the autoencoder is the image (input_img), and the output is the decoder layer (decoded)\nautoencoder = tf.keras.Model(input_img, decoded)\n\nencoder = tf.keras.Model(input_img, encoded2)\n\nencoded_input = tf.keras.layers.Input(shape=(50,))\n# The decoded only consists of the last layer\ndecoder_layer = autoencoder.layers[-2](encoded_input)\n# The input to the decoder is the vector of the encoder which will be fed (using encoded_input), the output is the last layer of the network (decoder_layer)\ndecoder = tf.keras.Model(encoded_input, decoder_layer)\n\nautoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')\nautoencoder.fit(X, X, epochs=50, shuffle=True)\n<\/code><\/pre>\n\n<p>I would like to have the model actually learn instead of getting stuck.<\/p>\n","answers":[{"AnswerId":"57645974","CreationDate":"2019-08-25T12:23:38.613","ParentId":null,"OwnerUserId":"11351803","Title":null,"Body":"<p>Try fitting them again.<\/p>\n\n<p>But watch out for the over fitting problem.<\/p>\n\n<p>A better approach would be if you increased your training set. <\/p>\n"}]}
+{"QuestionId":57644238,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-08-25T07:54:25.113","AcceptedAnswerId":57644359.0,"OwnerUserId":11973552.0,"Title":"How to restructure output of a keras layer?","Body":"<p>I want to reconstruct the output shape of my Input layer, in a way that for example if it is (None, 42), I start from 1st neuron and take 1st to 7th, then slide a neuron to right and take 2nd to 8th and so on. in this way output layer would be the shape (None, 36, 7). <\/p>\n\n<p>I am using keras.layers.Lambda() layer to do that but it is not working properly and dimensions are not what i expect. the code i wrote is as follow: <\/p>\n\n<pre><code>Inpt_lyr = keras.Input(shape = (42,))\nfor k in range(42 - 7 + 1):\n    F = keras.layers.Lambda(lambda x, i, j: x[:,j : j + i])\n    F.arguments ={ 'i' : 7, 'j' : k}\n    tmp_rfrm = F(Inpt_lyr)\n    lst_rfrm.append(tmp_rfrm)\ncnctnt_lyr = keras.layers.merge.Concatenate(axis = 0)(lst_rfrm)\ntmp_dns = keras.layers.Dense(3 , activation = 'softmax')(cnctnt_lyr)\ntmp_model = Model(Inpt_lyr, tmp_dns)\ntmp_model.compile(loss = 'categorical_crossentropy', optimizer = 'adam')\ntmp_model.summary()\n<\/code><\/pre>\n\n<p>I expect conct_lyr have the shape (None, 36, 7), but it is (None,7)\nand it also give the error\nValueError: Graph disconnected: cannot obtain value for tensor Tensor(\"input_3:0\", shape=(?, 42), dtype=float32) at layer \"input_3\". The following previous layers were accessed without issue: ['input_4']<\/p>\n","answers":[{"AnswerId":"57644359","CreationDate":"2019-08-25T08:17:25.647","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>Your approach looks sensible to me, but the tensors are currently being concatenated along the batch axis, which is undesirable in this case.<\/p>\n\n<p>I would suggest expanding dimension 1 while sliding the windows:<\/p>\n\n<pre><code>for k in range(42 - 7 + 1):\n    F = keras.layers.Lambda(lambda x, i, j: x[:, None, j: j + i])\n    # F(Inpt_lyr).shape=(?, 1, 7)\n<\/code><\/pre>\n\n<p>and then concatenate the tensors along the expanded axis:<\/p>\n\n<pre><code>cnctnt_lyr = keras.layers.merge.Concatenate(axis=1)(lst_rfrm)\n# cnctnt_lyr.shape=(?, 36, 7)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57644272,"AnswerCount":1,"Tags":"<python><tensorflow><tensorrt>","CreationDate":"2019-08-25T07:59:20.270","AcceptedAnswerId":57703602.0,"OwnerUserId":4582711.0,"Title":"How to export RT graph from saved model","Body":"<p>I want to export the graph to RT for inference . How to achieve it with the saved model. The code for inference with saved model is as follows:<\/p>\n\n<pre><code>import tensorflow as tf\nimport cv2\n\nimg = cv2.imread(\"test.jpg\")\nimg = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\nimg_rgb = np.expand_dims(img, 0)\n\npredict_fn = tf.contrib.predictor.from_saved_model(\".\/saved_model\")\noutput_data = predict_fn({\"inputs\": img_rgb})\nprint(output_data)  # detector output dictionary\n<\/code><\/pre>\n\n<p>I don't want to use <code>frozen_inference_garph.pb<\/code>. I want to use <code>saved_model\/saved_model.pb<\/code> for inference (as accuracy is higher with it). <\/p>\n","answers":[{"AnswerId":"57703602","CreationDate":"2019-08-29T05:36:32.167","ParentId":null,"OwnerUserId":"4582711","Title":null,"Body":"<p>Command line tool <code>saved_model_cli<\/code> worked for me.<\/p>\n\n<pre><code>saved_model_cli convert \\\n--dir \"\/home\/yilrr\/tf-serving\/faster-rcnn\/saved_model\/versions\/1\" \\\n--output_dir \"\/home\/yilrr\/tf-serving\/trt-frcnn\" \\\n--tag_set serve \\\ntensorrt --precision_mode FP32 --max_batch_size 32 --is_dynamic_op True\n<\/code><\/pre>\n\n<p>And following is the code for inference:<\/p>\n\n<pre><code>import tensorflow as tf\nimport cv2\nimport numpy as np\n\nIMAGE_PATH = \"\/Users\/vedanshu\/test.jpg\"\nimage = cv2.imread(IMAGE_PATH)\n\ninput_names = ['image_tensor']\n\ndef draw_label(image, point, label, font=cv2.FONT_HERSHEY_SIMPLEX,\n               font_scale=0.5, thickness=2):\n    size = cv2.getTextSize(label, font, font_scale, thickness)[0]\n    x, y = point\n    cv2.rectangle(image, (x, y - size[1]),\n                  (x + size[0], y), (255, 0, 0), cv2.FILLED)\n    cv2.putText(image, label, point, font, font_scale,\n                (255, 255, 255), thickness)\n\ndetection_graph1 = tf.Graph()\n\nwith detection_graph1.as_default():\n    tf_sess1 = tf.Session(graph=detection_graph1)\n    model = tf.saved_model.loader.load(tf_sess1, [\"serve\"], \"\/Users\/vedanshu\/ckpt\/saved_model\")\n\ntf_input = tf_sess1.graph.get_tensor_by_name(input_names[0] + ':0')\ntf_scores = tf_sess1.graph.get_tensor_by_name('detection_scores:0')\ntf_boxes = tf_sess1.graph.get_tensor_by_name('detection_boxes:0')\ntf_classes = tf_sess1.graph.get_tensor_by_name('detection_classes:0')\ntf_num_detections = tf_sess1.graph.get_tensor_by_name('num_detections:0')\nscores, boxes, classes, num_detections = tf_sess1.run([tf_scores, tf_boxes, tf_classes, tf_num_detections], feed_dict={tf_input: image[None, ...]})\n\nboxes = boxes[0]  # index by 0 to remove batch dimension\nscores = scores[0]\nclasses = classes[0]\nnum_detections = int(num_detections[0])\n\nboxes_pixels = []\nfor i in range(num_detections):\n    # scale box to image coordinates\n    box = boxes[i] * np.array([image.shape[0],\n                               image.shape[1], image.shape[0], image.shape[1]])\n    box = np.round(box).astype(int)\n    boxes_pixels.append(box)\n\nboxes_pixels = np.array(boxes_pixels)\n\nfor i in range(num_detections):\n    box = boxes_pixels[i]\n    box = np.round(box).astype(int)\n    image = cv2.rectangle(image, (box[1], box[0]), (box[3], box[2]), (0, 255, 0), 2)\n    label = \"{}:{:.2f}\".format(int(classes[i]), scores[i])\n    draw_label(image, (box[1], box[0]), label)\n\ncv2.imwrite(\"\/Users\/vedanshu\/out.jpg\", image)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57644452,"AnswerCount":1,"Tags":"<python><machine-learning><dataset><artificial-intelligence><pytorch>","CreationDate":"2019-08-25T08:33:05.563","AcceptedAnswerId":null,"OwnerUserId":10548439.0,"Title":"Loading a huge dataset batch-wise to train pytorch","Body":"<p>I am training a LSTM in-order to classify the time-series data into 2 classes(0 and 1).I have huge data-set on the drive where where the 0-class and the 1-class data are located in different folders.I am trying to train the LSTM batch-wise using by creating a Dataset class and wrapping the DataLoader around it. I have to do pre-processing such as reshaping.Here's my code which does that<\/p>\n\n<p>`<\/p>\n\n<pre><code>class LoadingDataset(Dataset):\n  def __init__(self,data_root1,data_root2,file_name):\n    self.data_root1=data_root1#Has the path for class1 data\n    self.data_root2=data_root2#Has the path for class0 data\n    self.fileap1= pd.DataFrame()#Stores class 1 data\n    self.fileap0 = pd.DataFrame()#Stores class 0 data\n    self.file_name=file_name#List of all the files at data_root1 and data_root2\n    self.labs1=None #Will store the class 1 labels\n    self.labs0=None #Will store the class 0 labels\n\n  def __len__(self):\n    return len(self.fileap1) \n\n  def __getitem__(self, index):        \n    self.fileap1 = pd.read_csv(self.data_root1+self.file_name[index],header=None)#read the csv file for class 1\n    self.fileap1=self.fileap1.iloc[1:,1:].values.reshape(-1,WINDOW+1,1)#reshape the file for lstm\n    self.fileap0 = pd.read_csv(self.data_root2+self.file_name[index],header=None)#read the csv file for class 0\n    self.fileap0=self.fileap0.iloc[1:,1:].values.reshape(-1,WINDOW+1,1)#reshape the file for lstm\n    self.labs1=np.array([1]*len(self.fileap1)).reshape(-1,1)#create the labels 1 for the csv file\n    self.labs0=np.array([0]*len(self.fileap0)).reshape(-1,1)#create the labels 0 for the csv file\n    # print(self.fileap1.shape,' ',self.fileap0.shape)\n    # print(self.labs1.shape,' ',self.labs0.shape)\n    self.fileap1=np.append(self.fileap1,self.fileap0,axis=0)#combine the class 0 and class one data\n    self.fileap1 = torch.from_numpy(self.fileap1).float()\n    self.labs1=np.append(self.labs1,self.labs0,axis=0)#combine the label0 and label 1 data\n    self.labs1 = torch.from_numpy(self.labs1).int()\n    # print(self.fileap1.shape,' ',self.fileap0.shape)\n    # print(self.labs1.shape,' ',self.labs0.shape)\n\n    return self.fileap1,self.labs1\n\ndata_root1 = '\/content\/gdrive\/My Drive\/Data\/Processed_Data\/Folder1\/One_'#location of class 1 data\ndata_root2 = '\/content\/gdrive\/My Drive\/Data\/Processed_Data\/Folder0\/Zero_'#location of class 0 data\ntraining_set=LoadingDataset(data_root1,data_root2,train_ind)#train_ind is a list of file names that have to be read from data_root1 and data_root2\ntraining_generator = DataLoader(training_set,batch_size =2,num_workers=4)\n\nfor epoch in range(num_epochs):\n  model.train()#Setting the model to train mode after eval mode to train for next epoch once the testing for that epoch is finished\n  for i, (inputs, targets) in enumerate(train_loader):\n    .\n    .\n    .\n    .\n<\/code><\/pre>\n\n<p>`\nI get this error when the run this code<\/p>\n\n<blockquote>\n  <p>RuntimeError: Traceback (most recent call last):\n    File \"\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/utils\/data\/_utils\/worker.py\", line 99, in _worker_loop\n      samples = collate_fn([dataset[i] for i in batch_indices])\n    File \"\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/utils\/data\/_utils\/collate.py\", line 68, in default_collate\n      return [default_collate(samples) for samples in transposed]\n    File \"\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/utils\/data\/_utils\/collate.py\", line 68, in \n      return [default_collate(samples) for samples in transposed]\n    File \"\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/utils\/data\/_utils\/collate.py\", line 43, in default_collate\n      return torch.stack(batch, 0, out=out)\n  RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 0. Got 96596 and 25060 in dimension 1 at \/pytorch\/aten\/src\/TH\/generic\/THTensor.cpp:711<\/p>\n<\/blockquote>\n\n<p>My Questions are \n1.Have I Implemented this correctly, is this how you pre-process and then train a dataset batch-wise?<\/p>\n\n<p>2.The batch_size of DataLoader and batch_size of the LSTM are different since the batch_size of DataLoader refers to the no. of files, whereas batch_size of the LSTM model refers to the no. of instances, so will I get another error here?<\/p>\n\n<p>3.I have no idea how to scale this data-set since the MinMaxScaler has to be applied to the dataset in its entirety.<\/p>\n\n<p>Responses are appreciated.Please let me know if I have to create separate posts for each question.<\/p>\n\n<p>Thank You.<\/p>\n","answers":[{"AnswerId":"57654983","CreationDate":"2019-08-26T09:13:37.130","ParentId":null,"OwnerUserId":"10511793","Title":null,"Body":"<p>Here's a summary of how pytorch does things :<\/p>\n\n<ul>\n<li>You have a <code>dataset<\/code>, that is an object with a <code>__len__<\/code> method and a <code>__getitem__<\/code> method.<\/li>\n<li>You create a <code>dataloader<\/code> from that <code>dataset<\/code> and a <code>collate_fn<\/code><\/li>\n<li>You iterate through the <code>dataloader<\/code> and pass a batch of data to your model.<\/li>\n<\/ul>\n\n<p>So basically your training loop will look like <\/p>\n\n<pre><code>for x, y in dataloader:\n    output = model(x)\n...\n<\/code><\/pre>\n\n<p>or <\/p>\n\n<pre><code>for x, y in dataloader:\n        output = model(*x)\n    ...\n<\/code><\/pre>\n\n<p>if your model <code>forward<\/code> method takes multiple arguments.<\/p>\n\n<p>So how does this work ?\nBasically you have a generator of batch indices <code>batch_sampler<\/code> and here's what looping inside your dataloader will act like.<\/p>\n\n<pre><code>for indices in batch_sampler:\n    yield collate_fn([dataset[i] for i in indices])\n<\/code><\/pre>\n\n<p>So if you want everything to work well, you must look at the <code>forward<\/code> method of your model and see how many arguments it takes (In my experience forward method of LSTM can have multiple arguments), and make sure that you use a <code>collate_fn<\/code> to pass those correctly.<\/p>\n"}]}
+{"QuestionId":57644641,"AnswerCount":1,"Tags":"<python><keras><deep-learning>","CreationDate":"2019-08-25T09:00:17.647","AcceptedAnswerId":null,"OwnerUserId":6668992.0,"Title":"how to correctly interpenetrate accuracy with keras model, giving perfectly linear relation input vs output?","Body":"<p>using keras model i get zero accuracy for perfectly linear relation of output vs input, i'm not sure if i interpreted wrongly the accuracy or doing something wrong with my code any help will be appreciated<\/p>\n\n<p>i'v tried adding more layers, more epochs and so on nothing changed <\/p>\n\n<pre><code>import numpy as np\nimport matplotlib.pyplot as plt\nimport tensorflow as tf\n\nfrom keras import models \nfrom keras.models import Sequential\nfrom keras.layers import Dense, Dropout\nfrom keras import optimizers\nfrom sklearn.model_selection import KFold\n\nfrom sklearn.preprocessing import MinMaxScaler\ntf.reset_default_graph()\nfrom keras.optimizers import SGD\n\nsiz=100000\ninp=np.random.randint(100, 1000000 , size=[siz,1])\na1=1.5;\nuop=np.dot(inp,a1)\nnormzer_inp = MinMaxScaler()\ninp_norm = normzer_inp.fit_transform\\\n(inp)\nnormzer_uop = MinMaxScaler()\nuop_norm = normzer_uop.fit_transform\\\n(uop)\n\nX=inp_norm\nY=uop_norm\n\nkfold = KFold(n_splits=2, random_state=None, shuffle=False)\ncvscores = []\nopti_SGD = SGD(lr=0.01, momentum=0.9)\nmodel1 = Sequential()\naccc_trn=0\naccc_tst=0\nfor train, test in kfold.split(X, Y):\n    model = Sequential()\n    model.add(Dense(16, input_dim=X.shape[1], activation='relu'))\n    model.add(Dense(16, activation='relu'))\n\n    model.add(Dense(1, activation='linear'))\n    model.compile(loss='mean_squared_error', optimizer=opti_SGD,\\\n                  metrics=['accuracy'])\n\n    history = model.fit(X[train], Y[train], \\\n            validation_data=(X[test], Y[test]), \\\n            epochs=10,batch_size=2048, verbose=2)\n    _, train_acc = model.evaluate(X[train], Y[train], verbose=0)\n    _, test_acc = model.evaluate(X[test], Y[test], verbose=0)\n    print('Train: %.3f, Test: %.3f' % (train_acc, test_acc))  \n\n    plt.plot(history.history['acc'], label='train')\n    plt.plot(history.history['val_acc'], label='test')\n    plt.legend()\n    plt.show()  \n    cvscores.append(test_acc * 100)  \n\nprint(\"%.2f%% (+\/- %.2f%%)\" % (np.mean(cvscores)\\\n                               , np.std(cvscores))) \n<\/code><\/pre>\n\n<p>expected about 100% accuracy, received about 0%<\/p>\n\n<p>Train on 50000 samples, validate on 50000 samples\nEpoch 1\/10\n - 0s - loss: 0.1351 - acc: 2.0000e-05 - val_loss: 0.0476 - val_acc: 2.0000e-05\nEpoch 2\/10\n - 0s - loss: 0.0386 - acc: 2.0000e-05 - val_loss: 0.0243 - val_acc: 2.0000e-05\nEpoch 3\/10\n - 0s - loss: 0.0146 - acc: 2.0000e-05 - val_loss: 0.0063 - val_acc: 2.0000e-05\nEpoch 4\/10\n - 0s - loss: 0.0029 - acc: 2.0000e-05 - val_loss: 6.9764e-04 - val_acc: 2.0000e-05\nEpoch 5\/10\n - 0s - loss: 2.8476e-04 - acc: 2.0000e-05 - val_loss: 9.0012e-05 - val_acc: 2.0000e-05\nEpoch 6\/10\n - 0s - loss: 8.0603e-05 - acc: 2.0000e-05 - val_loss: 6.6961e-05 - val_acc: 2.0000e-05\nEpoch 7\/10\n - 0s - loss: 6.3046e-05 - acc: 2.0000e-05 - val_loss: 5.2784e-05 - val_acc: 2.0000e-05\nEpoch 8\/10\n - 0s - loss: 5.0725e-05 - acc: 2.0000e-05 - val_loss: 4.3357e-05 - val_acc: 2.0000e-05\nEpoch 9\/10\n - 0s - loss: 4.2132e-05 - acc: 2.0000e-05 - val_loss: 3.6418e-05 - val_acc: 2.0000e-05\nEpoch 10\/10\n - 0s - loss: 3.5651e-05 - acc: 2.0000e-05 - val_loss: 3.1116e-05 - val_acc: 2.0000e-05\nTrain: 0.000, Test: 0.000<\/p>\n\n<p>0.00% (+\/- 0.00%)<\/p>\n","answers":[{"AnswerId":"57645194","CreationDate":"2019-08-25T10:22:11.830","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>You are performing a regression task. Accuracy is used in classification type of tasks whereby you are measuring out of total number of samples, how many of them got predicted correctly. <\/p>\n\n<p>For regression tasks, usually model's performance is defined by it's validation loss. This may be mean squared error (as you are doing already) or mean absolute error etc.<\/p>\n\n<p>Just change your line of model compilation to:<\/p>\n\n<pre><code>model.compile(loss='mean_squared_error', optimizer=opti_SGD)\n<\/code><\/pre>\n\n<p>Now, no accuracy details will be printed.<\/p>\n"}]}
+{"QuestionId":57644940,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-08-25T09:45:59.680","AcceptedAnswerId":null,"OwnerUserId":11973949.0,"Title":"Pytorch not using GPU . Worked on Fastai","Body":"<p>My GPU Nvidia gtx1050 Ti<\/p>\n\n<p>I am trying to train it on GPU but I only see CPU utilization 60-90 percent and GPU around 5 percent during training may be due to the copying of tensors to GPU I don't know. But it just goes up 5 percent and comes down.<\/p>\n\n<p>I tried increasing batch size to 64 or 128 based on some solutions online but it just gives me Cuda out of memory error. I have used like 2.3gb out of 4gb and says need 120 MB and has only 110 mb.---- I dont even understand  how it works at this point<\/p>\n\n<p>Then tried decreasing batch size to 16 then also got Cuda out of memory error saying need like 40 MB and has 16mb same 2.3 GB used out of 4gb<\/p>\n\n<p>In the end it worked for 8 batch size but it only uses CPU no GPU<\/p>\n\n<p>I used fastai with batch size 128 it works fine and uses gpu . I dont know where I did wrong. Any help appreciated. Below is my code wrote based on pytorch image classifier tutorial<\/p>\n\n<p>Model Resnet pretrained True\nhas 205 labels with 117000 images approx as data for training. \nJust using pretrained weights such that weights will modify accordingly without freezing weights and not training from scratch I assume is what the code does. Feel free to correct me if I did something wrong or better solution. In the end I am just noob in pytorch .... My first code in pytorch<\/p>\n\n<p>trainset....... PIL Image format with rgb converted to prevent error of 4 channel and removed if not opened by PIL --invalid images: corrupted image<\/p>\n\n<pre><code>device = torch.device(\"cuda:0\") #0 device is my nvidia gtx 1050 ti when printed\n\nmodel.fc=nn.Linear(2048, 205) \n\nfrom torchvision import transforms\nt = transforms.Compose([            \n transforms.Resize(256),                    #[2]\n transforms.CenterCrop(224),                #[3]\n transforms.ToTensor(),                     #[4]\n transforms.Normalize(                      #[5]\n mean=[0.485, 0.456, 0.406],                #[6]\n std=[0.229, 0.224, 0.225] )                 #[7]\n ])\ntrainloader = torch.utils.data.DataLoader(trainset, batch_size=8,\n                                          shuffle=True)\n\nif torch.cuda.is_available():\n    print('yes gpu')\n    torch.set_default_tensor_type('torch.cuda.FloatTensor')\n    model = model.cuda()\n\nimport torch.optim as optim\ncriterion = nn.CrossEntropyLoss()\noptimizer = optim.Adam(model.parameters(), lr=0.001)\n\nfor epoch in range(6):  # loop over the dataset multiple times\n    running_loss = 0.0\n    for i, data in enumerate(trainloader, 0):\n        # get the inputs; data is a list of [inputs, labels]\n        inputs, labels,img_name = data\n        inputs = inputs.to(device) \n        labels = labels.to(device)\n        # zero the parameter gradients\n        optimizer.zero_grad()\n\n        # forward + backward + optimize\n        outputs = model(inputs)\n        loss = criterion(outputs, labels)\n        loss.backward()\n        optimizer.step()\n\n        # print statistics\n        running_loss += loss.item()\n        total=0\n        correct=0\n        if i % 2000 == 1999: \n            # print every 2000 mini-batches\n\n            print('[%d, %5d] loss: %.3f' %\n                  (epoch + 1, i + 1, running_loss \/ 2000))\n            outputs = model(inputs)\n            _, predicted = torch.max(outputs.data, 1)\n            total += labels.size(0)\n            correct += (predicted == labels).sum().item()\n\n            print('Accuracy of the network on the batch size  images: %d %%' % (\n                100 * correct \/ total)) #training accuraccy\n\n            running_loss = 0.0\n\nprint('Finished Training')\n<\/code><\/pre>\n","answers":[{"AnswerId":"57654681","CreationDate":"2019-08-26T08:52:47.790","ParentId":null,"OwnerUserId":"10511793","Title":null,"Body":"<p>To monitor your GPU usage continuously you can use\n<code>nvidia-smi --loop=1<\/code> in a terminal.<\/p>\n\n<p>There are two things to look for, the usage of the GPU and the usage of the ram of the GPU, choose the maximum batch size such that you don't get a out of memory error.<\/p>\n\n<p>Regarding the usage of the GPU, as long as it is computing, it should be approaching 100 %, it is possible that your data-loading process on CPU is slow and so that most of the time the GPU is waiting for the CPU to give it tensor to perform computation on. For instance, in your case, all transformations are performed on CPU, which takes some time.<\/p>\n\n<p>You might want to launch another training, and watch closely how the GPU usages changes over time to decide what needs to be changed and if something needs to be changed.<\/p>\n"}]}
+{"QuestionId":57645611,"AnswerCount":0,"Tags":"<python><tensorflow><keras><conv-neural-network>","CreationDate":"2019-08-25T11:29:42.663","AcceptedAnswerId":null,"OwnerUserId":11971761.0,"Title":"How to predict using flow from directory and assing the predicted label to the image id or image name","Body":"<p>I've built a multi-class CNN model, it has 10 classes. I'm using the flow from directory to feed in my input values to my model. and when I'm done training what I want to do is predict my testing image class. this image is in another directory in which 10000 images are available classes are available. when I'm using model.predict_generator it's giving me (10000,10) shaped numpy array, now I can get the class by using np.argmmax but how to link this predicted label with the image name or id.\nbelow is the code which I'm using\nplease help<\/p>\n\n<pre><code>model = tf.keras.models.Sequential([\n    tf.keras.layers.Conv2D(16, (3, 3), activation='relu', input_shape=(28, 28, 3)),\n    tf.keras.layers.MaxPooling2D(2, 2),\n    tf.keras.layers.Conv2D(32, (3, 3), activation='relu'),\n    tf.keras.layers.MaxPooling2D(2, 2),\n    tf.keras.layers.Conv2D(64, (3, 3), activation='relu'),\n    tf.keras.layers.MaxPooling2D(2, 2),\n    tf.keras.layers.Flatten(),\n    tf.keras.layers.Dense(512, activation='relu'),\n    tf.keras.layers.Dense(10, activation=tf.nn.softmax)\n])\n\nmodel.compile(optimizer=RMSprop(lr=0.001), loss='sparse_categorical_crossentropy', metrics=['acc'])\n\nTRAINING_DIR = '\/content\/Dataset_appareal\/TRAINING_SET'\ntrain_datagen = ImageDataGenerator()\ntrain_generator = train_datagen.flow_from_directory(TRAINING_DIR,\n                                                    batch_size=100,\n                                                    class_mode='sparse',\n                                                    target_size=(28, 28))\n\nVALIDATION_DIR = '\/content\/Dataset_appareal\/VALIDATION_SET'\nvalidation_datagen = ImageDataGenerator()\nvalidation_generator = validation_datagen.flow_from_directory(VALIDATION_DIR,\n                                                              batch_size=100,\n                                                              class_mode='sparse',\n                                                              target_size=(28, 28))\n\nhistory = model.fit_generator(train_generator,\n                              epochs=50,\n                              verbose=,\n                              validation_data=validation_generator)\n\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57645681,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2019-08-25T11:39:41.530","AcceptedAnswerId":null,"OwnerUserId":11838162.0,"Title":"python, TensorFlow and machine learning","Body":"<p>I'm working on TensorFlow with machine learning.<\/p>\n\n<p>I got stuck in <strong>step1<\/strong> and <strong>step2<\/strong><\/p>\n\n<p><strong>step1:<\/strong><\/p>\n\n<pre><code>X = X\/255.0\n<\/code><\/pre>\n\n<blockquote>\n  <p>TypeError: unsupported operand type(s) for \/: 'list' and 'float'<\/p>\n<\/blockquote>\n\n<p><strong>step2:<\/strong><\/p>\n\n<pre><code>model.add(Conv2D(64, (3,3), input_shape=X.shape[1:])\n<\/code><\/pre>\n\n<blockquote>\n  <p>'list' object has no attribute 'shape'<\/p>\n<\/blockquote>\n\n<p><strong>Edit 25\/08-2019 13:44<\/strong><\/p>\n\n<p>Step1: X=np.array(X)<\/p>\n\n<p>Got new error in step2<\/p>\n\n<p>Step2: <\/p>\n\n<blockquote>\n  <p>Input 0 of layer conv2d is incompatible with the layer: expected\n  ndim=4, found ndim=1. Full shape received: [None]<\/p>\n<\/blockquote>\n\n<p><strong>Edit 25\/08-2019 19:26<\/strong>\nmy full code:<\/p>\n\n<pre><code>import os\nos.environ[\"CUDA_VISIBLE_DEVICES\"]=\"-1\"\nos.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'\nimport tensorflow as tf\nfrom tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import Dense, Dropout, Activation, Flatten, Conv2D, MaxPooling2D\nimport pickle\nimport numpy as np\n\nX = pickle.load(open(\"x.pickle\", \"rb\"))\ny = pickle.load(open(\"y.pickle\", \"rb\"))\n\nX = np.array(X)\n\nX = X\/255.0\n#x.shape=np.array([x])\n#X = np.asarray(x).shape[1:]\n\nprint(X.shape)\n\nmodel = Sequential()\n\nmodel.add(Conv2D(64, (3,3), input_shape=X.shape[1:]))\n#model.add(Conv2D(64, (3,3), input_shape=x.shape[1:]))\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))\n\nmodel.add(Dense(1))\nmodel.add(Activation('sigmoid'))\n\nmodel.compile(loss=\"binary_crossentropy\", optimizer=\"adam\", metrics=[\"accuracy\"])\n\nmodel.fit(X, y, batch_size=32, epochs=1, validation_split=0.1)\n<\/code><\/pre>\n\n<p>X.Shape = (24946,)<\/p>\n","answers":[{"AnswerId":"57648113","CreationDate":"2019-08-25T17:13:38.003","ParentId":null,"OwnerUserId":"11866916","Title":null,"Body":"<p>conv2d expects input tensor of shape [batch, in_height, in_width, in_channels](4 dimensional shape), check the shape of your X.<\/p>\n"}]}
+{"QuestionId":57645759,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning><conv-neural-network>","CreationDate":"2019-08-25T11:51:42.217","AcceptedAnswerId":null,"OwnerUserId":11149645.0,"Title":"CNN for discrete regression","Body":"<p>I have been trying to make a CNN model for my project for a while but i couldn't really make it work. My project is about designing a (40,40) filter with discrete values (0.125, 0.25, 0375, 0.5, 0.625, 0.75, 0875, 1.) using a two channel image like input (40,40,2) . I used scipy zscore on the inputs so they are between -2 to 2. My labels are (40,40) values (0.125, 0.25, 0375, 0.5, 0.625, 0.75, 0875, 1.).\n Untill now, i have been using regression to find values and round them to discrete values i want, but its becoming problematic and I'm having accuracy problems. The model i used :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential()\nmodel.add(Conv2D(64,(5,5), input_shape = (40,40,2)))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2,2)))\n\nmodel.add(Conv2D(64,(5,5)))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2,2)))\n\nmodel.add(Conv2D(64,(3,3)))\nmodel.add(Activation('relu'))\n\nmodel.add(Flatten())\nmodel.add(Dense(64))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.5))\n\nmodel.add(Dense(1600))\nmodel.add(Reshape((40,40)))\nmodel.add(Activation('sigmoid'))\n\nSGD = optimizers.SGD(lr=0.01, momentum=0, decay=0.0, nesterov=False)\n\nmodel.compile(loss='MSE',\n             optimizer='SGD',\n             metrics=['MAE'])\n\n### Saving the model\ncheckpoint_path = 'training\/cp.ckpt'\ncheckpoint_dir = os.path.dirname(checkpoint_path)\n\n### Checkpoint callback\ncp_callback = ModelCheckpoint(checkpoint_path, \n                              monitor='val_loss',\n                              save_best_only=True,\n                              mode='auto',\n                              save_weights_only=True,\n                              verbose=1)\n\n### Reduce LR on Plateau callback\nlr_callback = ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=2, verbose=0, mode='auto', min_delta=0.0001, cooldown=0, min_lr=0)\n\n### Training the model\nhistory = model.fit(intensity_train, doe_train,\n          batch_size=16,\n          epochs=15,\n          validation_data=(intensity_val,doe_val),\n          callbacks = [cp_callback, lr_callback])\n<\/code><\/pre>\n\n<p>And the graph is :\n<a href=\"https:\/\/i.stack.imgur.com\/8ks1k.png\" rel=\"nofollow noreferrer\">loss and validation loss<\/a><\/p>\n\n<p>The problem is, when i check the results (I compare the rounded outputs with the original data with an algorithm) some datas have around 1590 correct and 10 incorrect values which is amazing. But for some, prediction is incorrect for all values. So i want my model to give outputs in the form (40,40) with the discrete values i gave earlier. Since the labels are in the same form, I think i can get better performance that way, but I'm not sure how to do that.<\/p>\n\n<p>So my question is, how can i make it so the regression outputs are in the form i want. <\/p>\n\n<p>Bonus question: How can i make a accuracy based metric such as 1560\/1600 correct value ==> %97.5.<\/p>\n","answers":[]}
+{"QuestionId":57645936,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-08-25T12:17:42.980","AcceptedAnswerId":null,"OwnerUserId":11972177.0,"Title":"Can not squeeze dim[1], expected a dimension of 1, got 14 [[{{node metrics_2\/acc\/Squeeze}}]]","Body":"<p>I have written a code in python for image classification, but after writing code, I have got this error and in the dataset of images, there are images of sizes min. 655x53, my task is to classify the images of cells into 14 categories. there are total 6889 images of training set and 3771 images of test\/validation set<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.keras.preprocessing.image import ImageDataGenerator\n\ntrain_dir = '\/vista\/train1'\nvalidation_dir = '\/vista\/test1'\n\ntrain_data = ImageDataGenerator(rescale = 1\/255,\n                            rotation_range = 20,\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\nvalidation_data = ImageDataGenerator(rescale = 1\/255)\n\ntrain_gen = train_data.flow_from_directory(train_dir,\n                                        batch_size = 30,\n                                        class_mode = 'categorical',\n                                        target_size = (400,200))\n\nvalidation_gen = validation_data.flow_from_directory(validation_dir,\n                                                 batch_size = 30,\n                                                 class_mode = \n'categorical',\n                                                 target_size = (400,200))\n\nmodel = tf.keras.models.Sequential([\n    tf.keras.layers.Conv2D(64,(3,3),activation = 'relu',input_shape = \n(400,200,3)),\n    tf.keras.layers.MaxPooling2D(2,2),\n\n    tf.keras.layers.Conv2D(128,(3,3),activation = 'relu'),\n    tf.keras.layers.MaxPooling2D(2,2),\n\n    tf.keras.layers.Conv2D(128,(3,3),activation = 'relu'),\n    tf.keras.layers.MaxPooling2D(2,2),\n\n    tf.keras.layers.Conv2D(128,(3,3),activation = 'relu'),\n    tf.keras.layers.MaxPooling2D(2,2),\n\n    tf.keras.layers.Flatten(),\n    tf.keras.layers.Dense(256,activation = 'relu'),\n    tf.keras.layers.Dropout(0.20),\n    tf.keras.layers.Dense(14,activation = 'softmax')\n    ])\n\nmodel.compile(optimizer = 'Adam',\n          loss = 'sparse_categorical_crossentropy',\n          metrics = ['acc'])\n\ndf = int(3711\/30)\nef = int(6889\/30)\n\nhistory = model.fit_generator(train_gen,\n                          steps_per_epoch = df,\n                          epochs = 20,\n                          validation_data = validation_gen,\n          `enter code here`                validation_steps = ef,\n                          verbose =2)\n\nmodel.save_weights('\/vista',overwrite=True,save_format = '.h5')\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57645953,"AnswerCount":0,"Tags":"<python><arrays><tensorflow>","CreationDate":"2019-08-25T12:20:21.583","AcceptedAnswerId":null,"OwnerUserId":3849479.0,"Title":"After feed_dict shape changes","Body":"<p>I try to accomplish test of a stereo matching problem in Tensorflow. After applied feed_dict, shape changes like below.<\/p>\n\n<p>Must be :    <code>(1, 375, 1242, 64)<\/code>\nChanges to : <code>(1, 339, 1206, 64)<\/code><\/p>\n\n<p>There is a difference but I don't understand, why it happens. I need to apply a multiplication after fix this.<\/p>\n\n<p>Operation like this :<\/p>\n\n<pre><code> l_features, r_features = sess.run([l_map, r_map], feed_dict={left_paths_placeholder: l_image, right_paths_placeholder: r_image, phase_train_placeholder:False})\n<\/code><\/pre>\n\n<p>I tried to reshape the images, arrays but dimension problems occurred.<\/p>\n","answers":[]}
+{"QuestionId":57646601,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-08-25T13:49:05.517","AcceptedAnswerId":null,"OwnerUserId":10848871.0,"Title":"How to prepare data for Conv1D layer when the input have different time frames?","Body":"<p>I want to feed spectrograms --corresponding to uttered digits-- to a model with a Conv1D as first layer. I then use a RNN layer to classify which word is uttered. These spectrograms have different sequence\/time length, but the same feature number of course.<\/p>\n\n<p>In Keras's Conv1D doc:<\/p>\n\n<blockquote>\n  <p>When using this layer as the first layer in a model, provide an input_shape argument [..] (None, 128) for variable-length sequences with 128 features per step.<\/p>\n<\/blockquote>\n\n<p>So it seems like it handles it. No need for padding\/resizing.<\/p>\n\n<p>I'm used to preparing same-shape data with numpy (ex: numpy.vstack), but now I have various-shape I can't figure out how to do it! And I only found example where people had same-shaped data.<\/p>\n\n<p>Or maybe it is not possible with numpy, and I have to use something else?<\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"57646695","CreationDate":"2019-08-25T14:00:27.020","ParentId":null,"OwnerUserId":"4867531","Title":null,"Body":"<p>It's possible, but you have to make sure that the sequences, which are batched together have the same length, that's why the most people simply pad all their sequences. If you use masking it does also ignore the masked values, so there is no difference.<\/p>\n\n<pre><code>model.add(LSTM(32, return_sequences=True, input_shape=(None, 5)))\nmodel.add(LSTM(8, return_sequences=True))\nmodel.add(TimeDistributed(Dense(2, activation='sigmoid')))\n\nprint(model.summary(90))\n\nmodel.compile(loss='categorical_crossentropy',\n              optimizer='adam')\n\ndef train_generator():\n    while True:\n        sequence_length = np.random.randint(10, 100)\n        x_train = np.random.random((1000, sequence_length, 5))\n        # y_train will depend on past 5 timesteps of x\n        y_train = x_train[:, :, 0]\n        for i in range(1, 5):\n            y_train[:, i:] += x_train[:, :-i, i]\n        y_train = to_categorical(y_train &gt; 2,5)\n        yield x_train, y_train\n\nmodel.fit_generator(train_generator(), steps_per_epoch=30, epochs=10, verbose=1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57646903,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><neural-network><random-forest>","CreationDate":"2019-08-25T14:26:43.263","AcceptedAnswerId":null,"OwnerUserId":2462721.0,"Title":"Tensorflow Model, Random Forest AUC - How to calculate using Session?","Body":"<p>Trying to figure out how to calculate AUC in Random Forest Tensorflow when using Session training.<\/p>\n\n<p>Tried many approaches like the ones mentioned here:<\/p>\n\n<ul>\n<li><p><a href=\"https:\/\/stackoverflow.com\/questions\/48174323\/tensorflow-1-4-tf-metrics-auc-for-auc-calculation\">Tensorflow 1.4 tf.metrics.auc for AUC calculation<\/a><\/p><\/li>\n<li><p><a href=\"https:\/\/stackoverflow.com\/questions\/42632656\/cant-seem-to-get-tensorflows-tf-metrics-auc-working\">Can&#39;t seem to get Tensorflow&#39;s tf.metrics.auc working<\/a><\/p><\/li>\n<li><p><a href=\"https:\/\/stackoverflow.com\/questions\/46408810\/how-to-calculate-auc-in-tensorflow-in-an-easy-way\">How to calculate auc in tensorflow in an easy way?<\/a><\/p><\/li>\n<\/ul>\n\n<p>I think that I cannot do follow these patterns above using the code seen here:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Splitting the dataset into the Training set and Test set\n\nfrom sklearn.model_selection import train_test_split\n\nX_train, X_test, y_train, y_test = train_test_split(input_x, input_y, test_size = 0.25, random_state = 0)\n\ndata1 = data.iloc[:,:].values\n\n\n# Parameters\nnum_steps = 50 # Total steps to train\nnum_classes = 2 \nnum_features = 14\nnum_trees = 10 \nmax_nodes = 1000 \n\n# Input and Target placeholders \nX = tf.placeholder(tf.float32, shape=[None, num_features])\nY = tf.placeholder(tf.int64, shape=[None])\n\n\n# Random Forest Parameters\nhparams = tensor_forest.ForestHParams(num_classes=num_classes, num_features=num_features, num_trees=num_trees, max_nodes=max_nodes).fill()\n\n# Build the Random Forest\nforest_graph = tensor_forest.RandomForestGraphs(hparams)\ntrain_op = forest_graph.training_graph(X, Y)\nloss_op = forest_graph.training_loss(X, Y)\ninfer_op, _, _ = forest_graph.inference_graph(X)\n\n### ACCURACY DEFINITION\ncorrect_prediction = tf.equal(tf.argmax(infer_op, 1), tf.cast(Y, tf.int64))\naccuracy_op = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n\n\n### AUC DEFINITION\nfrom sklearn.metrics import roc_auc_score\nHELP HERE ? \n\n\n### SESSION DEFINITION\ninit_vars = tf.group(tf.global_variables_initializer(), resources.initialize_resources(resources.shared_resources()))\n\nsess = tf.Session()\n\n# Training here\nfor i in range(1, num_steps + 1):\n    _, l = sess.run([train_op, loss_op], feed_dict={X: X_train, Y: y_train})\n    sess.run(tf.local_variables_initializer())\n\n\n    if i % 50 == 0 or i == 1:\n\n        acc = sess.run(accuracy_op, feed_dict={X: X_train, Y: y_train})\n        HELP HERE TO MAKE AUC AVAILABLE TOO?        \n\n        print('Step %i, Loss: %f, Acc: %f' % (i, l, acc))\n\n\n# EVALUATION\nprint(\"Test Accuracy:\", sess.run(accuracy_op, feed_dict={X: X_test, Y: y_test}))\n\nHELP HERE TO PRINT AUC?\n<\/code><\/pre>\n\n<p>Can anyone help me to understand how to include AUC calc here?<\/p>\n","answers":[{"AnswerId":"57654007","CreationDate":"2019-08-26T08:05:12.367","ParentId":null,"OwnerUserId":"5134333","Title":null,"Body":"<p>This is what you need:<\/p>\n\n<pre><code>predictions=pd.DataFrame(model.predict_proba(X_test),columns=model.classes_)\nroc_score = roc_auc_score((real_test_class == 'True').astype(float), predictions['True'])\n<\/code><\/pre>\n\n<p>Where real_test_class is the vector holding your real class, and predictions['True'] is the column in the predictions dataframe which holds the <strong><em>probabilities<\/em><\/strong> of each sample being 'True'<\/p>\n"}]}
+{"QuestionId":57646905,"AnswerCount":2,"Tags":"<python><machine-learning><keras>","CreationDate":"2019-08-25T14:26:57.250","AcceptedAnswerId":57647582.0,"OwnerUserId":395028.0,"Title":"Understanding keras model.predict","Body":"<p>I have gone through <a href=\"https:\/\/stackoverflow.com\/questions\/38971293\/get-class-labels-from-keras-functional-model\">this<\/a>, and <a href=\"https:\/\/datascience.stackexchange.com\/questions\/36238\/what-does-the-output-of-model-predict-function-from-keras-mean\">this<\/a> post.\nHowever, my question is very specific:\nIf the output of my <code>model.predict()<\/code> function for a classification problem with class labelled 0 and 1 is something like:<\/p>\n\n<pre><code>array([[0.5147758 ],\n       [0.48530805],\n       [0.5122566 ],\n       [0.4839405 ],\n       [0.49831972],\n       [0.4886117 ],\n       [0.5130876 ],\n       [0.50388396]], dtype=float32)\n<\/code><\/pre>\n\n<p>and I'm using <code>binary_crossentropy<\/code> loss with the last layer as:<\/p>\n\n<pre><code>Dense(1, activation='sigmoid')\n<\/code><\/pre>\n\n<p>Then each entry in the above output denotes the probability of occurrence of class 0 or class 1 ?<\/p>\n","answers":[{"AnswerId":"57646966","CreationDate":"2019-08-25T14:35:02.440","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p><code>sigmoid<\/code> activation will output values between 0 and 1 and you have only one unit of neuron in your Dense layer. Binary crossentropy loss will try to maximize the probability in favour of one of the two classes available. So, to be precise, the output in your case is actually the probability of occurence of class 1. For class 0 probabilities value, you will have to do <code>1 - output<\/code>.<\/p>\n\n<p>Other type of commonly used activation in final layer is 'softmax'. This activation will give you probabilities of each class of occurence, and hence the number of units in your final layer will be equal to the number of classes. In this setup, we use categorical crossentropy loss.<\/p>\n"},{"AnswerId":"57647582","CreationDate":"2019-08-25T16:01:58.520","ParentId":null,"OwnerUserId":"4685471","Title":null,"Body":"<blockquote>\n  <p>each entry in the above output denotes the probability of occurrence of class 0 or class 1 ?<\/p>\n<\/blockquote>\n\n<p>The conventional interpretation of this output is as probabilities for the output belonging to class 1. <\/p>\n\n<p>Intuitively, it maybe convenient to imagine them as trying to \"replicate\" the actual binary labels (0\/1), so the closer they are to 1.0 the higher the probability of class 1 (and vice versa). Roughly speaking, this is actually the definition of <a href=\"https:\/\/ml-cheatsheet.readthedocs.io\/en\/latest\/loss_functions.html#cross-entropy\" rel=\"nofollow noreferrer\">cross entropy loss<\/a>, used here:<\/p>\n\n<blockquote>\n  <p><strong>Cross-Entropy<\/strong><\/p>\n  \n  <p>Cross-entropy loss, or log loss, measures the performance of a classification model whose output is a probability value between 0 and 1. Cross-entropy loss increases as the predicted probability diverges from the actual label. So predicting a probability of .012 when the actual observation label is 1 would be bad and result in a high loss value. A perfect model would have a log loss of 0.<\/p>\n  \n  <p><a href=\"https:\/\/i.stack.imgur.com\/8HMmw.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/8HMmw.png\" alt=\"enter image description here\"><\/a><\/p>\n<\/blockquote>\n\n<hr>\n\n<blockquote>\n  <p>was wondering whether there is some documentation where this is specifically mentioned<\/p>\n<\/blockquote>\n\n<p>The reason you cannot find it explicitly in Keras documentation is that it is much more general, having to do with the fundamental ideas of log loss &amp; binary classification, and it has nothing to do with Keras in particular. These threads may be helpful for convincing you:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/47817424\/loss-accuracy-are-these-reasonable-learning-curves\/47819022#47819022\">Loss &amp; accuracy - Are these reasonable learning curves?<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/47508874\/how-does-keras-evaluate-the-accuracy\/47515095#47515095\">How does Keras evaluate the accuracy?<\/a><\/p>\n"}]}
+{"QuestionId":57647045,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><variables>","CreationDate":"2019-08-25T14:46:17.983","AcceptedAnswerId":null,"OwnerUserId":10479025.0,"Title":"How to correctly reuse Variables in Tensorflow?","Body":"<p>I'm trying to train a model but, each time I my the train() method, TensorFlow initializes the Variable to the original initializer values, instead of reusing trained values.\nHere is the code I tried:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class Neuron():\n\n    def __init__(self):\n        self.weights = np.array([[0.5],[0.2]]).astype('f')\n\n    def train(self):\n        with tf.variable_scope(\"scope\", reuse=tf.AUTO_REUSE):\n            w = tf.get_variable(\"weights\", initializer=self.weights)\n\n        # DO TRAIN WITH tf.Session()\n\n        self.weights = sess.run(w)\n<\/code><\/pre>\n\n<p>But when I call train() again the weights are reinitialized to [[0.5],[0.2]], even if I updated self.weights in last line, causing this behaviour:\n<a href=\"https:\/\/i.stack.imgur.com\/egrop.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/egrop.png\" alt=\"weights\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57647418,"AnswerCount":1,"Tags":"<tensorflow><keras><neural-network><layer>","CreationDate":"2019-08-25T15:37:41.847","AcceptedAnswerId":null,"OwnerUserId":4040643.0,"Title":"Propagating through a custom layer in tensorflow just once","Body":"<p>Given a custom layer in tensorflow, is it possible to let the model use it just during one epoch? The layer may just be ignored for all other epochs or simple be an identity.<\/p>\n\n<p>For example: Given data I would like the layer to simply double the given data. The other layers should may work normally. How would one do that?<\/p>\n\n<pre><code>def do_stuff(data):\n      return 2*data\n\ndef run_once(data):\n  return tf.py_func(do_stuff, \n                     [data],\n                     'float32',\n                     stateful=False,\n                     name='I run once')\n\n\nclass CustomLayer(Layer):\n  def __init__(self, output_dim, **kwargs):\n    self.output_dim = output_dim\n    self.trainable = False\n    super(CustomLayer, self).__init__(**kwargs)\n\n  def call(self, x):\n    res = tf.map_fn(run_once, x)\n    res.set_shape([x.shape[0],\n                   self.output_dim[1], \n                   self.output_dim[0],\n                   x.shape[-1]])\n    return res\n\ninputs = Input(shape=(224, 224, 1))    \nx = Lambda(preprocess_input(x), input_shape=(224, 224, 1), output_shape=(224, 224, 3))\noutputs = Dense(1)(x)\nmodel = Model(input=inputs, output=outputs)\noutput = model(x)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57659781","CreationDate":"2019-08-26T14:22:07.347","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>Interesting question. To execute a TF operation just in the first epoch, one could use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/cond\" rel=\"nofollow noreferrer\">tf.cond<\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/control_dependencies\" rel=\"nofollow noreferrer\">tf.control_dependencies<\/a> to check\/update the value of a boolean tensor. For example, your custom layer could be implemented as follows:<\/p>\n\n<pre><code>class CustomLayer(Layer):\n    def __init__(self, **kwargs):\n        super(CustomLayer, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        self.first_epoch = tf.Variable(True)\n\n    def call(self, x):\n        res = tf.cond(self.first_epoch,\n                      true_fn=lambda: run_once(x),\n                      false_fn=lambda: x)\n        with tf.control_dependencies([res]):\n            assign_op = self.first_epoch.assign(False)\n            with tf.control_dependencies([assign_op]):\n                res = tf.identity(res)\n        return res\n<\/code><\/pre>\n\n<p>To validate that this layer works as expected, define <code>run_once<\/code> as:<\/p>\n\n<pre><code>def run_once(data):\n    print_op = tf.print('First epoch')\n    with tf.control_dependencies([print_op]):\n        out = tf.identity(data)\n    return out\n<\/code><\/pre>\n"}]}
+{"QuestionId":57647433,"AnswerCount":0,"Tags":"<python><pandas><tensorflow><matplotlib><linear-regression>","CreationDate":"2019-08-25T15:38:55.383","AcceptedAnswerId":null,"OwnerUserId":11857526.0,"Title":"Linear regression - Evaluating training accuracy","Body":"<p>I'm fetching some data from a CSV file (1000 lines). It has two columns, where the 1st column is the x_training data and the 2nd column is the y_training data. <\/p>\n\n<p><strong>Snapshot from the CSV-file<\/strong><\/p>\n\n<pre><code>8.070000000000000284e+01,1.126768031895251987e+01\n8.040000000000000568e+01,1.195844519276935358e+01\n7.250000000000000000e+01,8.317461617744008606e+00\n1.030000000000000000e+02,1.880844309373589951e+01\n1.075999999999999943e+02,1.947419293659330108e+01\n7.940000000000000568e+01,9.877652348817933969e+00\n8.190000000000000568e+01,1.127064360995226977e+01\n1.015999999999999943e+02,1.640426417487080357e+01\n1.085999999999999943e+02,1.749193091101176378e+01\n9.570000000000000284e+01,1.574942514809519345e+01\n5.270000000000000284e+01,3.581285321328901539e+00\n\n<\/code><\/pre>\n\n<p>After the data is fetched I allocate the respective data to the lists and turn them into matrices.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>data = pd.read_csv('length_weight.csv', delimiter=\",\", dtype='float32')\nx_train = np.mat(data.iloc[:, 0]).reshape(-1, 1)\ny_train = np.mat(data.iloc[:, 1]).reshape(-1, 1)\n<\/code><\/pre>\n\n<p>In order to calculate the training precision, a LinearRegressionModel class has been made.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class LinearRegressionModel:\n    def __init__(self):\n        # Model input\n        self.x = tf.compat.v1.placeholder(tf.float32)\n        self.y = tf.compat.v1.placeholder(tf.float32)\n\n        # Model variables\n        self.W = tf.Variable([[0.0]])\n        self.b = tf.Variable([[0.0]])\n\n        # Predictor\n        f = tf.matmul(self.x, self.W) + self.b\n\n        # Mean Squared Error\n        self.loss = tf.reduce_mean(tf.square(f - self.y))\n\n\nmodel = LinearRegressionModel()\n\n# Training: adjust the model so that its loss is minimized\nminimize_operation = tf.compat.v1.train.GradientDescentOptimizer(0.00001).minimize(model.loss)\n\n# Create session object for running TensorFlow operations\nsession = tf.compat.v1.Session()\n\n# Initialize tf.Variable objects\nsession.run(tf.compat.v1.global_variables_initializer())\n\nfor epoch in range(1000):\n    session.run(minimize_operation, {model.x: x_train, model.y: y_train})\n<\/code><\/pre>\n\n<p>Then I run the session in order to get the calculations, however, the answer turns out to just be NaN as seen under<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>W, b, loss = session.run([model.W, model.b, model.loss], {model.x: x_train, model.y: y_train})\n<\/code><\/pre>\n\n<p>Which gives the output:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>W = [[nan]], b = [[nan]], loss = nan\n<\/code><\/pre>\n\n<p>I have tried with data like this, which works fine:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x_train = np.mat([[1], [1.5], [2], [3], [4], [5], [6]])\ny_train = np.mat([[5], [3.5], [3], [4], [3], [1.5], [2]])\n<\/code><\/pre>\n\n<p>I would guess that it has something to do with formatting of the data that I'm using, but I'm currently out of ideas of what I can do. Any help is much appreciated.<\/p>\n","answers":[]}
+{"QuestionId":57648327,"AnswerCount":1,"Tags":"<c><tensorflow><deeplab>","CreationDate":"2019-08-25T17:42:53.273","AcceptedAnswerId":57667041.0,"OwnerUserId":7095746.0,"Title":"Tensorflow: TF_SessionRun returns TF_INVALID_ARGUMENT","Body":"<p>I am running a session from a frozen graph of Deeplabv3 using the Tensorflow C API. When I get to the part of running the session with <code>TF_SessionRun<\/code>, the return value is 3, indicating <code>TF_INVALID_ARGUMENT<\/code>. I suspect it may have to do something with the <code>TF_Operation*<\/code> input (the 8th argument aka \"Target Operations\" argument) which I left <code>NULL<\/code>, but I cannot find any documentation as to what this argument represents. Below is my problematic invocation of TF_SessionRun:<\/p>\n\n<p>from <strong>tiny_deeplab_api.cpp<\/strong>:<\/p>\n\n<pre><code>    \/\/ Allocate the input tensor\n    TF_Tensor* const input = TF_NewTensor(TF_UINT8, img-&gt;dims, 3, img-&gt;data_ptr, img-&gt;bytes, &amp;free_tensor, NULL);\n    TF_Operation* oper_in = TF_GraphOperationByName(graph, \"ImageTensor\");\n    const TF_Output oper_in_ = {oper_in, 0};\n\n    \/\/ Allocate the output tensor\n    TF_Tensor* output = TF_NewTensor(TF_UINT8, seg-&gt;dims, 2, seg-&gt;data_ptr, seg-&gt;bytes, &amp;free_tensor, NULL);\n    TF_Operation* oper_out = TF_GraphOperationByName(graph, \"SemanticPredictions\");\n    const TF_Output oper_out_ = {oper_out, 0};\n\n    \/\/ Run the session on the input tensor\n    TF_SessionRun(session, NULL, &amp;oper_in_, &amp;input, 1, &amp;oper_out_, &amp;output, 1, NULL, 0, NULL, status);\n\n    return TF_GetCode(status); \/\/ https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/c\/tf_status.h#L42 \n<\/code><\/pre>\n\n<p>where <code>img<\/code> and <code>seg<\/code> are <code>image_t<\/code> and <code>segmap_t<\/code> types which contain pointers to the data and to a dimension array that the <code>TF_NewTensor()<\/code> method can use to make input and output tensors to then pass to <code>TF_SessionRun()<\/code>. (from <strong>tiny_deeplab_api.hpp<\/strong>): <\/p>\n\n<pre><code>typedef struct segmap {\n    const int64_t* dims;\n    size_t bytes;\n    uint8_t* data_ptr;\n} segmap_t;\n\ntypedef struct image {\n    const int64_t* dims;\n    size_t bytes;\n    uint8_t* data_ptr;\n} image_t;\n<\/code><\/pre>\n\n<hr>\n\n<p>Below is the source code in case the problem is not obvious... <\/p>\n\n<p>test.cpp:<\/p>\n\n<pre><code>#include &lt;opencv2\/opencv.hpp&gt;\n#include \"tiny_deeplab_api.hpp\"\n#include &lt;iostream&gt;\n#include &lt;algorithm&gt;\n\nint main() {\n    using namespace std;\n    using namespace cv;\n\n    \/\/ Initialize Deeplab object\n    Deeplab dl = Deeplab();\n    cout &lt;&lt; \"Successfully constructed Deeplab object\" &lt;&lt; endl;\n\n    \/\/ Read &amp; resize input image\n    Mat image = imread(\"\/Users\/Daniel\/Desktop\/cat.jpg\"); \n    int orig_height = image.size().height;\n    int orig_width = image.size().width;\n    double resize_ratio = (double) 513 \/ max(orig_height, orig_width);\n    Size new_size((int)(resize_ratio*orig_width), (int)(resize_ratio*orig_height));\n    Mat resized_image;\n    resize(image, resized_image, new_size);\n    cout &lt;&lt; \"Image resized (h, w): (\" &lt;&lt; orig_height &lt;&lt; \",\" &lt;&lt; orig_width &lt;&lt; \") --&gt; (\" &lt;&lt; new_size.height &lt;&lt; \", \" &lt;&lt; new_size.width &lt;&lt; \")\" &lt;&lt; endl;\n    imshow(\"Image\", resized_image);\n    waitKey(0);\n\n\n    \/\/ Allocate input image object\n    const int64_t dims_in[3] = {new_size.width, new_size.height, 3};\n    image_t* img_in = (image_t*)malloc(sizeof(image_t));\n    img_in-&gt;dims = &amp;dims_in[0];\n    img_in-&gt;data_ptr = resized_image.data;\n    img_in-&gt;bytes = new_size.width*new_size.height*3*sizeof(uint8_t);\n\n    \/\/ Allocate output segmentation map object\n    const int64_t dims_out[2] = {new_size.width, new_size.height};\n    segmap_t* seg_out = (segmap_t*)malloc(sizeof(segmap_t));\n    seg_out-&gt;dims = &amp;dims_out[0];\n    seg_out-&gt;data_ptr = (uint8_t*)malloc(new_size.width*new_size.height);\n    seg_out-&gt;bytes = new_size.width*new_size.height*sizeof(uint8_t);\n\n    \/\/ Run Deeplab\n    cout &lt;&lt; \"Running segmentation\" &lt;&lt; endl;\n    int status = dl.run_segmentation(img_in, seg_out);\n    if(status != 0) {\n        cout &lt;&lt; \"ERROR RUNNING SEGMENTATION: \" &lt;&lt; status &lt;&lt; endl;\n        return 1;\n    }\n\n    cout &lt;&lt; \"Successfully ran segmentation\" &lt;&lt; endl;\n\n    \/\/ Interpret results\n\n    return 0;\n}\n<\/code><\/pre>\n\n<p>tiny_deeplab_api.hpp:<\/p>\n\n<pre><code>#ifndef TINY_DEEPLAB_API_HPP_\n#define TINY_DEEPLAB_API_HPP_\n\n#include &lt;tensorflow\/c\/c_api.h&gt;\n\nTF_Buffer* read_file(const char* file);\nvoid free_buffer(void* data, size_t length);\nvoid free_tensor(void* data, size_t length, void* args);\n\ntypedef struct segmap {\n    const int64_t* dims;\n    size_t bytes;\n    uint8_t* data_ptr;\n} segmap_t;\n\ntypedef struct image {\n    const int64_t* dims;\n    size_t bytes;\n    uint8_t* data_ptr;\n} image_t;\n\n\nclass Deeplab {\n   private:\n    TF_Session* session;\n    TF_Graph* graph;\n    TF_Output output_oper;\n    TF_Output input_oper;\n    TF_Status* status;\n\n   public:\n    Deeplab(); \/\/ Constructor \n    ~Deeplab();\n    int run_segmentation(image_t*, segmap_t*);\n};\n\n#endif \/\/ TINY_DEEPLAB_API_HPP_\n<\/code><\/pre>\n\n<p>tiny_deeplab_api.cpp:<\/p>\n\n<pre><code>#include &lt;stdio.h&gt;\n#include &lt;stdlib.h&gt;\n#include &lt;iostream&gt;\n#include &lt;tensorflow\/c\/c_api.h&gt;\n#include \"tiny_deeplab_api.hpp\"\n\nDeeplab::Deeplab() {\n    using namespace std;\n    cout &lt;&lt; \"Hello from TensorFlow C library version\" &lt;&lt; TF_Version() &lt;&lt; endl;\n\n    \/\/ Import Deeplab graph (as a frozen graph, it has the weights hard-coded in as constants, so no need to restore the checkpoint)\n    TF_Buffer* graph_def = read_file(\"..\/Models\/Deeplab_model_unpacked\/deeplabv3_mnv2_cityscapes_train\/frozen_inference_graph.pb\");\n    graph = TF_NewGraph();\n    status = TF_NewStatus();\n    TF_ImportGraphDefOptions* opts = TF_NewImportGraphDefOptions();\n    TF_GraphImportGraphDef(graph, graph_def, opts, status);\n    TF_DeleteImportGraphDefOptions(opts);\n    if (TF_GetCode(status) != TF_OK) {\n        fprintf(stderr, \"ERROR: Unable to import graph %s\", TF_Message(status));\n        return;\n    }\n    cout &lt;&lt; \"Successfully loaded Deeplab graph\" &lt;&lt; endl;\n    TF_DeleteBuffer(graph_def);\n\n    \/\/ Initialize Session\n    TF_SessionOptions* sess_opts = TF_NewSessionOptions();\n    session = TF_NewSession(graph, sess_opts, status);\n}\n\nDeeplab::~Deeplab() {\n    using namespace std;\n    TF_CloseSession(session, status);\n    TF_DeleteSession(session, status);\n    TF_DeleteStatus(status);\n    TF_DeleteGraph(graph);\n    cout &lt;&lt; \"Destroyed Deeplab object\" &lt;&lt; endl;\n}\n\nint Deeplab::run_segmentation(image_t* img, segmap_t* seg) {\n    \/\/TODO: Delete old TF_Tensor, TF_Operation, and TF_Output \n\n    \/\/ Allocate the input tensor\n    TF_Tensor* const input = TF_NewTensor(TF_UINT8, img-&gt;dims, 3, img-&gt;data_ptr, img-&gt;bytes, &amp;free_tensor, NULL);\n    TF_Operation* oper_in = TF_GraphOperationByName(graph, \"ImageTensor\");\n    const TF_Output oper_in_ = {oper_in, 0};\n\n    \/\/ Allocate the output tensor\n    TF_Tensor* output = TF_NewTensor(TF_UINT8, seg-&gt;dims, 2, seg-&gt;data_ptr, seg-&gt;bytes, &amp;free_tensor, NULL);\n    TF_Operation* oper_out = TF_GraphOperationByName(graph, \"SemanticPredictions\");\n    const TF_Output oper_out_ = {oper_out, 0};\n\n    \/\/ Run the session on the input tensor\n    TF_SessionRun(session, NULL, &amp;oper_in_, &amp;input, 1, &amp;oper_out_, &amp;output, 1, NULL, 0, NULL, status);\n\n    return TF_GetCode(status); \/\/ https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/c\/tf_status.h#L42 \n}\n\nTF_Buffer* read_file(const char* file) {\n    FILE *f = fopen(file, \"rb\");\n    fseek(f, 0, SEEK_END);\n    long fsize = ftell(f);\n    fseek(f, 0, SEEK_SET);  \/\/same as rewind(f);\n\n    void* data = malloc(fsize);\n    fread(data, fsize, 1, f);\n    fclose(f);\n\n    TF_Buffer* buf = TF_NewBuffer();\n    buf-&gt;data = data;\n    buf-&gt;length = fsize;\n    buf-&gt;data_deallocator = free_buffer;\n    return buf;\n}\n\nvoid free_buffer(void* data, size_t length) { \n        free(data);\n}\n\nvoid free_tensor(void* data, size_t length, void* args) { \n        free(data);\n}\n<\/code><\/pre>\n\n<p>And the output of running <code>.\/test<\/code>:<\/p>\n\n<pre><code>Hello from TensorFlow C library version1.14.0\nSuccessfully loaded Deeplab graph\n2019-08-25 13:40:06.947965: I tensorflow\/core\/platform\/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\nSuccessfully constructed Deeplab object\nImage resized (h, w): (1680,2987) --&gt; (288, 513)\nRunning segmentation\nERROR RUNNING SEGMENTATION: 3\nDestroyed Deeplab object\n<\/code><\/pre>\n","answers":[{"AnswerId":"57667041","CreationDate":"2019-08-27T02:36:09.603","ParentId":null,"OwnerUserId":"7095746","Title":null,"Body":"<p>The answer is that, for some reason (why?) the dimensions of Deeplab's input and output tensors are NOT {width, height, 3} and {width, height}, but {1, width, height, 3} and {1, width, height}. After making dimension arrays of this form, TF_SessionRun ran without errors. <\/p>\n"}]}
+{"QuestionId":57648400,"AnswerCount":0,"Tags":"<android><flutter><keras><prediction>","CreationDate":"2019-08-25T17:51:07.253","AcceptedAnswerId":null,"OwnerUserId":11919769.0,"Title":"Application is crashing when working with custom model using tflite in flutter. How do i fix it?","Body":"<p>I have built a custom tflite model which got converted from Keras hdf5 model. But when using the converted model, while making prediction the application is crashing in android, but when I am using the mobile net tflite model which is downloaded from internet works fine with the application. What changes do I have to make? Whether the problem is in the model conversion or in the application? <\/p>\n\n<p>I have tried using the inbuilt function by tflite package that supports mobile net and various other nets that are given by tflite, I changed to custom build model prediction function and kept the converted file model. In the first case, it worked, but in the second case, it did not. <\/p>\n\n<pre><code>Future prediction(File image) async{\n  _recognitions= null;\n  var recognitions = await Tflite.runModelOnImage(\n  path: image.path,   \/\/ required\n  imageMean: 0.0,   \/\/ defaults to 117.0\n  imageStd: 255.0,  \/\/ defaults to 1.0\n  numResults: 2,    \/\/ defaults to 5\n  threshold: 0.2,   \/\/ defaults to 0.1\n  asynch: true      \/\/ defaults to true\n);\n  \/\/ var recognitions = await Tflite.detectObjectOnImage(\n  \/\/   path: image.path,\n  \/\/   model: \"SSDMobileNet\",\n  \/\/   imageMean: 127.5,\n  \/\/   threshold: 0.4,\n  \/\/   numResultsPerClass: 2,\n  \/\/   asynch: true\n  \/\/ );\n  print(recognitions);\n  setState(() {\n    _recognitions=recognitions;\n  });\n}\n<\/code><\/pre>\n\n<p>I have added 2 models in assets:<\/p>\n\n<pre><code>  assets:\n   - images\/webdoctor.png\n   - assets\/detect.tflite\n   - assets\/labelmap.txt\n   - assets\/labels.txt\n   - assets\/modelPneumonia.tflite\n<\/code><\/pre>\n\n<p>Expected results would be the working of custom model without app crashing.<\/p>\n","answers":[]}
+{"QuestionId":57648784,"AnswerCount":2,"Tags":"<python><keras><transfer-learning>","CreationDate":"2019-08-25T18:44:57.017","AcceptedAnswerId":null,"OwnerUserId":11975421.0,"Title":"OOM-Error on Layers i dont want to train (Transfer Learning)","Body":"<p>I am currently working on a project were i try to compare serval Neural Networks ,one of them being a model to classify the cifar10 dataset with a model trained on top of the VGG16.<\/p>\n\n<p>I Do not have any problems when training the other models but with the VGG16 i get and Out of Memory Error allthough i dont want to train the Layer in Question.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport os\nimport matplotlib.pyplot as plt\nimport keras\n\nfrom keras.models import Model\nfrom keras.models import Sequential\nfrom keras import layers\nfrom keras.applications.vgg16 import VGG16\nimport keras.losses as losses\nimport keras.optimizers as optimizers\n\nfrom keras.datasets import mnist\nfrom keras.datasets import cifar10\n\nfrom keras.applications.vgg16 import VGG16\nfrom keras.applications.vgg16 import preprocess_input\nfrom keras.applications.vgg16 import decode_predictions\nfrom keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img\n\nimport cv2\n\ncwd =os.getcwd()\n\n\n#defining later used variables such as data paths and img sizes\nwhich_data = 'cifar10'\n\nimg_size =32\nimg_depth =3\nbatch_size=20\nclass_count = 10\nuse_percentage=0.5\n\ninput_shape=(img_size, img_size, img_depth)\n\n(x_train, y_train), (x_test, y_test) = cifar10.load_data()\nprint('x_train shape:', x_train.shape)\nprint(x_train.shape[0], 'train samples')\nprint(x_test.shape[0], 'test samples')\n\nx_train = x_train[:int((x_train.shape[0]*use_percentage))]\ny_train = y_train[:int((y_train.shape[0]*use_percentage))]\n\n\nx_test = x_test[:int((x_test.shape[0]*use_percentage))]\ny_test = y_test[:int((y_test.shape[0]*use_percentage))]\n\n\nprint('x_train shape:', x_train.shape)\nprint(x_train.shape[0], 'reduced train samples')\nprint(x_test.shape[0], 'reduced test samples')\n\ny_train = keras.utils.to_categorical(y_train, class_count)\ny_test = keras.utils.to_categorical(y_test, class_count)\n\nx_train = x_train.astype('float32')\nx_test = x_test.astype('float32')\nx_train \/= 255\nx_test \/= 255\n\n\nsteps_per_epoch= x_train.shape[0]\/\/batch_size\nvalidation_steps =x_test.shape[0]\/\/batch_size\nepochs= 20\nverbose =1 ### information output\n\n\n##VGG16 Model\ninput_tensor = layers.Input(shape=input_shape)\nmodel=VGG16(weights='imagenet', include_top=False,input_tensor=input_tensor)\n\nlayer_dict =dict([(layer.name, layer) for layer in model.layers])\n\n\nX = layer_dict['block5_pool'].output\n\nX = layers.Flatten()(X)\nX = layers.Dense(class_count, activation='softmax')(X)\n\n\n\ntransfer_simple = Model(inputs=model.input, outputs=X)\n\nfor layer in transfer_simple.layers[:19]:\n    layer.trainable = False\n\n\nprint(transfer_simple.summary())\n\n\ntransfer_simple.compile(loss='categorical_crossentropy', optimizer=keras.optimizers.SGD(lr=0.02),metrics=['accuracy'])\n\n\nts_history=transfer_simple.fit(\n        x_train,\n        y_train,\n        validation_data=(x_test, y_test),\n        steps_per_epoch=steps_per_epoch,\n        validation_steps=validation_steps,\n        epochs=epochs,\n        verbose=verbose,\n        shuffle=True\n)\n\n\ntransfer_simple.save_weights(which_data+'\/advanced_cnn'+epochs+'_weights.h5')  \n\n<\/code><\/pre>\n\n<p>The summary of the Model states that only  5,130 are trainable parameters i still get the following Error: <\/p>\n\n<pre><code>ResourceExhaustedError: OOM when allocating tensor with shape[25000,64,32,32] and type float on \/job:localhost\/replica:0\/task:0\/device:GPU:0 by allocator GPU_0_bfc\n     [[Node: block1_conv1_5\/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\"](block1_conv1_5\/convolution-0-TransposeNHWCToNCHW-LayoutOptimizer, block1_conv1_5\/kernel\/read)]]\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.\n\n     [[Node: loss_5\/mul\/_1255 = _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_402_loss_5\/mul\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\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.\n<\/code><\/pre>\n\n<p>I am at my wits end and really hope you can help me!<\/p>\n","answers":[{"AnswerId":"57654418","CreationDate":"2019-08-26T08:35:04.213","ParentId":null,"OwnerUserId":"11975421","Title":null,"Body":"<p>While i didnt figure out the solution to this problem i kinda found a way around it. Instead of using the fit() function to train my neural networking i used the following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\ntraining_stream = ImageDataGenerator().flow(x=x_train,y=y_train,batch_size=batch_size,shuffle=True\n        )\nvalidation_stream = ImageDataGenerator().flow(x=x_test,y=y_test,batch_size=batch_size,shuffle=True\n        )\n\nts_history=transfer_simple.fit_generator(\n        training_stream,\n        validation_data=validation_stream,\n        steps_per_epoch=steps_per_epoch,\n        validation_steps=validation_steps,\n        epochs=epochs,\n        verbose=verbose,\n        shuffle=True\n)\n\n<\/code><\/pre>\n\n<p>This seemsto handle the input data in smaller chunks and eliminates the OOM-Error<\/p>\n"},{"AnswerId":"57659419","CreationDate":"2019-08-26T14:01:09.513","ParentId":null,"OwnerUserId":"4867531","Title":null,"Body":"<p>There are many reasons why you run OOM but the first things to check:<\/p>\n\n<ul>\n<li>Is your batch_size to big? Reduce the batch_size<\/li>\n<li>Weird stuff happend in your current session: <code>tf.keras.backend.clear_session()<\/code> often helps, sometimes you have to restart your kernel<\/li>\n<li>Is your network too big? <\/li>\n<\/ul>\n\n<p>Try to estimate you memory requirements:<\/p>\n\n<pre><code>def get_model_memory_usage(batch_size, model):\nshapes_mem_count = 0\nfor l in model.layers:\n    single_layer_mem = 1\n    for s in l.output_shape:\n        if s is None:\n            continue\n        single_layer_mem *= s\n    shapes_mem_count += single_layer_mem\n\ntrainable_count = np.sum([K.count_params(p) for p in set(model.trainable_weights)])\nnon_trainable_count = np.sum([K.count_params(p) for p in set(model.non_trainable_weights)])\n\nnumber_size = 4.0\nif K.floatx() == 'float16':\n     number_size = 2.0\nif K.floatx() == 'float64':\n     number_size = 8.0\n\ntotal_memory = number_size*(batch_size*shapes_mem_count + trainable_count + non_trainable_count)\ngbytes = np.round(total_memory \/ (1024.0 ** 3), 3)\nreturn gbytes\n<\/code><\/pre>\n"}]}
+{"QuestionId":57648932,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras><google-earth-engine>","CreationDate":"2019-08-25T19:05:41.097","AcceptedAnswerId":57680664.0,"OwnerUserId":11975431.0,"Title":"Keras Conv2D CNN - Error when checking target - expected smaller output","Body":"<p>I am stacking 6 layers of 2D satellite imagery (x data) and attempting to run a CNN over them to classify the landcover (using 8 land cover classes taken from a reformatted USDA Crop Data Layer - y data). <\/p>\n\n<p>The x data is shaped (2004, 2753, 6) and the y is shaped (2004, 2753, 8) originally and I have used data_x.reshape(-1,2004,2752,6) (same for y) to add an extra dimension as the model. <\/p>\n\n<p>The 8 categories in the y data-set represent 8 possible land-cover categories in numerical format in 8 bands (i.e. 1st band is corn and represented by 1's for positive and 0 for not corn). <\/p>\n\n<p>However, when i try to run the model the expected shape does not match what is being passed through to it. I am unsure if I am using the correct model structure or data structure - one idea would be to take the 8 bands of the y dataset <\/p>\n\n<p>Based on some serious googling i have been learning how to get the data into the correct format with the right number of dimensions etc but feel I am falling at the last hurdle with regards to dimensions (and most likely correct preparation of the x &amp; y data sets). <\/p>\n\n<p>Below is the CNN model<\/p>\n\n<pre><code>input_shape=([2004, 2753, 6])\n\nmodel = Sequential()\nmodel.add(Conv2D(32, kernel_size=(3, 3),strides=(1, 1),activation='relu',input_shape=input_shape))\nmodel.add(MaxPooling2D(pool_size=(2,2), strides=(2, 2), padding=\"same\"))\nmodel.add(Conv2D(32, (3, 3), activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2,2), padding=\"same\"))\nmodel.add(Dropout(0.25))\n#model.add(Flatten())\nmodel.add(Dense(128, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(8, activation='softmax'))\n#model.add(Flatten())\nmodel.summary()\n<\/code><\/pre>\n\n<p>Model Summary - expecting 500, 687, 8 out at the end<\/p>\n\n<pre><code>Layer (type)                 Output Shape              Param #   \n=================================================================\nconv2d_54 (Conv2D)           (None, 2002, 2751, 32)    1760      \n_________________________________________________________________\nmax_pooling2d_52 (MaxPooling (None, 1001, 1376, 32)    0         \n_________________________________________________________________\nconv2d_55 (Conv2D)           (None, 999, 1374, 32)     9248      \n_________________________________________________________________\nmax_pooling2d_53 (MaxPooling (None, 500, 687, 32)      0         \n_________________________________________________________________\ndropout_57 (Dropout)         (None, 500, 687, 32)      0         \n_________________________________________________________________\ndense_59 (Dense)             (None, 500, 687, 128)     4224      \n_________________________________________________________________\ndropout_58 (Dropout)         (None, 500, 687, 128)     0         \n_________________________________________________________________\ndense_60 (Dense)             (None, 500, 687, 8)       1032      \n=================================================================\nTotal params: 16,264\nTrainable params: 16,264\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>compile <\/p>\n\n<pre><code>sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)\n\nmodel.compile(loss='categorical_crossentropy',\n            optimizer='sgd',\n            metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>fit - and where i get the error message <\/p>\n\n<pre><code>history = model.fit(x_train3d, y_train3d,\n          batch_size=batch_size,\n          epochs=epochs,\n          verbose=1,\n          validation_split=0.2, validation_data=None)\n<\/code><\/pre>\n\n<p>shape of x_train3D = (1, 2004, 2753, 6)\nshape of y_train3D = (1, 2004, 2753, 8)<\/p>\n\n<p>error message<\/p>\n\n<pre><code>ValueError: Error when checking target: expected dense_58 to have shape (500, 687, 8) but got array with shape (2004, 2753, 8)\n<\/code><\/pre>\n\n<p>Again, I suspect this is down to needing to get the data in the right format both for the input and output but also likely something wrong in the specification of the model. Would appreciate some guidance as i'm new to Keras.<\/p>\n","answers":[{"AnswerId":"57649599","CreationDate":"2019-08-25T20:44:48.187","ParentId":null,"OwnerUserId":"3069214","Title":null,"Body":"<p>Can you please explain what are you trying to classifiy and what is your expected y_train3D (is it an image or some value for classificaton e.g. 1\/2\/3.. or x\/y\/z..etc)<\/p>\n"},{"AnswerId":"57680664","CreationDate":"2019-08-27T18:50:55.587","ParentId":null,"OwnerUserId":"11975431","Title":null,"Body":"<p>Just for an update on this - I have managed to clear the error (and now onto a memory error but that's another question). <\/p>\n\n<p>Solved the issue in 2 ways. \n1. Added upsampling to the end of the model to get the data back into the original size - new code in below <\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv2D(32, (3, 3), padding=\"same\", activation=\"relu\",input_shape=input_shape))\nmodel.add(MaxPooling2D(pool_size=(2, 2),strides=(2, 2)))\n\nmodel.add(Conv2D(64, (3, 3), padding=\"same\", activation=\"relu\"))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Conv2D(128, (3, 3), padding=\"same\", activation=\"relu\"))\n\n#Upsampling\nmodel.add(UpSampling2D(size=(2,2),interpolation='nearest'))\nmodel.add(UpSampling2D(size=(2,2),interpolation='nearest'))\n\nmodel.add(Dense(8, activation='relu'))\nmodel.summary()\n<\/code><\/pre>\n\n<p>Give me the below summary <\/p>\n\n<pre><code>\nLayer (type)                 Output Shape              Param #   \n=================================================================\nconv2d_1 (Conv2D)            (None, 2004, 2752, 32)    1760      \n_________________________________________________________________\nmax_pooling2d_1 (MaxPooling2 (None, 1002, 1376, 32)    0         \n_________________________________________________________________\nconv2d_2 (Conv2D)            (None, 1002, 1376, 64)    18496     \n_________________________________________________________________\nmax_pooling2d_2 (MaxPooling2 (None, 501, 688, 64)      0         \n_________________________________________________________________\nconv2d_3 (Conv2D)            (None, 501, 688, 128)     73856     \n_________________________________________________________________\nup_sampling2d_1 (UpSampling2 (None, 1002, 1376, 128)   0         \n_________________________________________________________________\nup_sampling2d_2 (UpSampling2 (None, 2004, 2752, 128)   0         \n_________________________________________________________________\ndense_1 (Dense)              (None, 2004, 2752, 8)     1032      \n=================================================================\nTotal params: 95,144\nTrainable params: 95,144\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>Part 2 - was ensuring the x and y data arrays were dividable by 4, otherwise this meant as I was losing some of the data through the model through rounding. The below is specific to my code and not robust but worked<\/p>\n\n<pre><code>if x_train3d.shape[2] % 2:\n    x_train3d_adj =  x_train3d_adj[:,:,:-1,:]\n    y_train3d_adj =  y_train3d_adj[:,:,:-1,:]\n<\/code><\/pre>\n\n<p>Not a complete solution yet but does get me closer to the end goal <\/p>\n"}]}
+{"QuestionId":57648945,"AnswerCount":1,"Tags":"<image-processing><dataset><pytorch><dataloader>","CreationDate":"2019-08-25T19:07:02.017","AcceptedAnswerId":57652442.0,"OwnerUserId":4060920.0,"Title":"Loading custom dataset of images using PyTorch","Body":"<p>I'm using the coil-100 dataset which has images of 100 objects, 72 images per object taken from a fixed camera by turning the object 5 degrees per image. Following is the folder structure I'm using:<\/p>\n\n<blockquote>\n  <p>data\/train\/obj1\/obj01_0.png, obj01_5.png ... obj01_355.png<br>\n  .<br>\n  .<br>\n  data\/train\/obj85\/obj85_0.png, obj85_5.png ... obj85_355.png<br>\n  .<br>\n  .<br>\n  data\/test\/obj86\/obj86_0.ong, obj86_5.png ... obj86_355.png<br>\n  .<br>\n  .<br>\n  data\/test\/obj100\/obj100_0.ong, obj100_5.png ... obj100_355.png  <\/p>\n<\/blockquote>\n\n<p>I have used the imageloader and dataloader classes. The train and test datasets loaded properly and I can print the class names.<\/p>\n\n<pre><code>train_path = 'data\/train\/'\ntest_path = 'data\/test\/'\ndata_transforms = {\n    transforms.Compose([\n    transforms.Resize(224, 224),\n    transforms.ToTensor(),\n    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))\n    ])\n}\n\ntrain_data = torchvision.datasets.ImageFolder(\n    root=train_path,\n    transform= data_transforms\n)\ntest_data = torchvision.datasets.ImageFolder(\n    root = test_path,\n    transform = data_transforms\n)\ntrain_loader = torch.utils.data.DataLoader(\n    train_data,\n    batch_size=None,\n    num_workers=1,\n    shuffle=False\n)\ntest_loader = torch.utils.data.DataLoader(\n    test_data,\n    batch_size=None,\n    num_workers=1,\n    shuffle=False\n)\n\nprint(len(train_data))\nprint(len(test_data))\nclasses = train_data.class_to_idx\nprint(\"detected classes: \", classes)\n<\/code><\/pre>\n\n<p>In my model I wish to pass every image through pretrained resnet and make a dataset from the output of resnet to feed into a biderectional LSTM. \nFor which I need to access the images by classname and index.\nfor ex. <code>pre_resnet_train_data['obj01'][0]<\/code> should be <code>obj01_0.png<\/code> and <code>post_resnet_train_data['obj01'][0]<\/code> should be the resnet output of <code>obj01_0.png<\/code> and so on.<br>\nI'm a beginner in Pytorch and for the past 2 days, I have read many tutorials and stackoverflow questions about creating a custom dataset class but couldn't figure out how to achieve what I want. \nplease help!<\/p>\n","answers":[{"AnswerId":"57652442","CreationDate":"2019-08-26T05:53:42.117","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Assuming you only plan on running resent on the images once and save the output for later use, I suggest you write your own <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/data.html#torch.utils.data.Dataset\" rel=\"nofollow noreferrer\">data set<\/a>, derived from <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/torchvision\/datasets.html#imagefolder\" rel=\"nofollow noreferrer\"><code>ImageFolder<\/code><\/a>.<br>\nSave each resnet output at the same location as the image file with <code>.pth<\/code> extension.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class MyDataset(torchvision.datasets.ImageFolder):\n  def __init__(self, root, transform):\n    super(MyDataset, self).__init__(root, transform)\n\n  def __getitem__(self, index):\n    # override ImageFolder's method\n    \"\"\"\n    Args:\n      index (int): Index\n    Returns:\n      tuple: (sample, resnet, target) where target is class_index of the target class.\n    \"\"\"\n    path, target = self.samples[index]\n    sample = self.loader(path)\n    if self.transform is not None:\n      sample = self.transform(sample)\n    if self.target_transform is not None:\n      target = self.target_transform(target)\n    # this is where you load your resnet data\n    resnet_path = os.path.join(os.path.splitext(path)[0], '.pth')  # replace image extension with .pth\n    resnet = torch.load(resnet_path)  # load the stored features\n    return sample, resnet, target\n<\/code><\/pre>\n"}]}
+{"QuestionId":57648976,"AnswerCount":1,"Tags":"<numpy><tensorflow>","CreationDate":"2019-08-25T19:11:56.763","AcceptedAnswerId":null,"OwnerUserId":6304513.0,"Title":"I do not understand this slicing or reshape in numpy","Body":"<p>full code: <a href=\"https:\/\/colab.research.google.com\/drive\/1W6k_nq890Fj5StsUtK4Hs1vorUgu0moA\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/1W6k_nq890Fj5StsUtK4Hs1vorUgu0moA<\/a><\/p>\n\n<p>Some how it reshapes or slices from (3621, 30, 1)\nto (1150,). I do not understand.\nThank You for your help!<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>print(len(series[..., np.newaxis]))\nprint((tf.expand_dims(series[..., np.newaxis], axis=-1).shape))\nrnn_forecast = model_forecast(model, series[..., np.newaxis], window_size)\nprint(\"before reshape: \"+str(rnn_forecast.shape))\nrnn_forecast = rnn_forecast[split_time - window_size:-1, -1, 0]\nprint(\"After reshape: \"+str(rnn_forecast.shape))\n\noutput:\n3650\n(3650, 1, 1)\nbefore reshape: (3591, 60, 1)\nAfter reshape: (1150,)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57649125","CreationDate":"2019-08-25T19:33:16.243","ParentId":null,"OwnerUserId":"8935","Title":null,"Body":"<p>This code does not perform a reshape.\nThe <strong>getitem<\/strong> call between the two print statements selects a slice for each of the three axes\nFor axis 0 the slice is split_time-window_size -1:-1  so ends at index 3648\nFor axis 1 the slice is single index -1, equivalent to index 29 for this example\nFor axis 2 the slice is single index 0<\/p>\n\n<p>Result is an array with just a single axis.<\/p>\n\n<p>In short, reshape plays no role in this example<\/p>\n"}]}
+{"QuestionId":57648994,"AnswerCount":1,"Tags":"<java><spring><caching><ehcache><caffeine>","CreationDate":"2019-08-25T19:14:33.433","AcceptedAnswerId":null,"OwnerUserId":4759273.0,"Title":"How to make cache depends on other caches on startup of spring application","Body":"<p>In our multi module project, we have different cache present in different modules. And we fill all these caches on server startup in @PostConstruct.\nNow some cache depends on some other cache which may present in different modules. So it requires that some cache should be filled before the caches that depends on that cache.\n1. How we can do this in Spring? Any design pattern I can leverage?\n2. If a cache gets updated how we can propogate this updated change to those caches that depends on updated cache in real time?<\/p>\n\n<pre><code>module 1---\n    Cachce1\n\nmodule 2--\n    Cache2\n\nmodule 3--\n    Cache3\n\n<\/code><\/pre>\n\n<pre><code>class Cache1 {\n  private ConcurrentMap&lt;&gt; cache; \n  @PostConstruct() {\n    cache = filleCache();\n  }\n}\n<\/code><\/pre>\n\n<pre><code>class Cache2 {\n  @Autowired\n  private Cache1 cache1;\n  private ConcurrentMap&lt;&gt; cache; \n  @PostConstruct() {\n    cache = cache1;\n  }\n}\n<\/code><\/pre>\n\n<pre><code>class Cache3 {\n  @Autowired\n  private Cache2 cache2;\n  private ConcurrentMap&lt;&gt; cache; \n  @PostConstruct() {\n    cache = cache2;\n  }\n}\n<\/code><\/pre>\n","answers":[{"AnswerId":"57656566","CreationDate":"2019-08-26T10:57:43.473","ParentId":null,"OwnerUserId":"2594725","Title":null,"Body":"<p><em>Dependencies<\/em>: The initialization order is usually done correctly by the DI framework. So <code>Cache2<\/code> post construct is run after <code>Cache1<\/code> is completely initialized. If you cache only a full set of data in a <code>CuncurrentMap<\/code> you could use memoization instead, see <a href=\"https:\/\/dzone.com\/articles\/need-micro-caching-memoization\" rel=\"nofollow noreferrer\">https:\/\/dzone.com\/articles\/need-micro-caching-memoization<\/a><\/p>\n\n<p>Avoid <em>cache stacking<\/em>. Sometimes its simpler to avoid caching in intermediate layers. A simple approach is to cache where the data comes in, e.g. the persistence layer, and then cache where the data goes out, e.g. response snippets or complete responses. Since the complete data is cached in the top-most cache anyways, the intermediate caches might not get any hits either.<\/p>\n\n<p>This heavily depends on the kind of data and how often its used on the next layer.<\/p>\n\n<blockquote>\n  <p>how we can propagate this updated change?<\/p>\n<\/blockquote>\n\n<p>You have a range of options. In case module1 knows about its dependencies you can call a reload of module2 and 3. If module1 code should not depend on its clients, then use a event listener pattern. Depending where your data comes from, there are already existing mechanisms. Look for Spring Data events, Change Data Capture capabilities of your database, Triggers of your database or persistence layer.<\/p>\n\n<p>You can keep things simple by using an expiry on the cache entry and simply update\/reload your data after a period of time. However, expiry and cache stacking is a problem by itself. In case you use a fixed expiry of 5 minutes, the effective expiry adds up each layer. So for three layers, data maybe refreshed after 5 or 15 minutes. To avoid this, we propagate a point of time when expiry should happen together with the data through the caching layers.<\/p>\n"}]}
+{"QuestionId":57649058,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><keras><conditional-formatting><metrics>","CreationDate":"2019-08-25T19:23:30.343","AcceptedAnswerId":null,"OwnerUserId":11462016.0,"Title":"Conditional Replacement in Keras Backend","Body":"<p>Target:\nHey i want to build a custom metrics method, where different specific classes are allowed do be confused with specific (because their idea in some conditions in nearly the same).<\/p>\n\n<p>How it could be reached:\nI got 25 classes, where the following confusions should be allowed<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>(1 can be 2,3,4) ; (2 can be 3,4) ; (3 can be 4) ;(15 can be 16,17,18);(16 can be 17) ; (17 can be 18) ; (6 can be 7,8 ) ; (7 can be 8 ); (19 can be 20,21 ) ; (20 can be 21 );(10 can be 11 );(13 can be 14 );(22 can be 23 ); (24 can be 14 );\n<\/code><\/pre>\n\n<p>quite much huh... therefore i put it into a mask:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>mask =    mask =(y_true!=y_pred)&amp;(((y_true==1)&amp;((y_pred == 2)| (y_pred == 3)| (y_pred == 4))) | ((y_true==2) &amp; ((y_pred == 3)|(y_pred == 4))) | ((y_true==3)&amp;((y_pred == 4)))| ((y_true==15)&amp;((y_pred == 16) | (y_pred == 17)| (y_pred == 18))) | ((y_true==16) &amp; ((y_pred == 17)|(y_pred == 18))) | ((y_true==17)&amp;((y_pred == 18))) | ((y_true==6) &amp; ((y_pred == 7)|(y_pred == 8)))|((y_true==7)&amp;((y_pred == 8)))|  ((y_true==19) &amp; ((y_pred == 20)|(y_pred == 21)))|((y_true==20)&amp;((y_pred == 21)))|((y_true==10)&amp;((y_pred == 11)))|((y_true==13)&amp;((y_pred == 14)))|((y_true==22)&amp;((y_pred == 23)))|((y_true==24)&amp;((y_pred == 14)))   )\n<\/code><\/pre>\n\n<p>I was hoping to find a function (equivalent to np.where) to conditional replace the values in the \"y_pred\" which fit the condition, by the \"y_true\" tensors values at the same index.\nI thought tf.where_v2() could work (sadly, i wasnt able to check if it works because tf.equal() and tf.reduce_all() only have limited abilities to work in a metrics function (or am I wrong?)<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>the following creates no debug errors\ndef profit_metrics(y_true, y_pred):\n    mask =(y_true!=y_pred)&amp;(((y_true==1)&amp;((y_pred == 2)| (y_pred == 3)| (y_pred == 4))) | ((y_true==2) &amp; ((y_pred == 3)|(y_pred == 4))) | ((y_true==3)&amp;((y_pred == 4)))| ((y_true==15)&amp;((y_pred == 16) | (y_pred == 17)| (y_pred == 18))) | ((y_true==16) &amp; ((y_pred == 17)|(y_pred == 18))) | ((y_true==17)&amp;((y_pred == 18))) | ((y_true==6) &amp; ((y_pred == 7)|(y_pred == 8)))|((y_true==7)&amp;((y_pred == 8)))|  ((y_true==19) &amp; ((y_pred == 20)|(y_pred == 21)))|((y_true==20)&amp;((y_pred == 21)))|((y_true==10)&amp;((y_pred == 11)))|((y_true==13)&amp;((y_pred == 14)))|((y_true==22)&amp;((y_pred == 23)))|((y_true==24)&amp;((y_pred == 14)))   )\n\n    new_pred = tf.where_v2(mask,y_true,y_pred)\n\n    return sparse_categorical_accuracy(y_true, new_pred)\n(...)\nmodel.compile(...,metrics=['accuracy',profit_metrics,])\n\n<\/code><\/pre>\n\n<p>but the thing is... it returns exactly the same accuracy value as the standard accuracy method! <\/p>\n\n<p>I tried to check if both tensors are still the same, but sadly all i tried didnt work out, maybe because i need to use only Keras backend functions?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def profit_metrics(y_true, y_pred):\n    mask =(y_true!=y_pred)&amp;(((y_true==1)&amp;((y_pred == 2)| (y_pred == 3)| (y_pred == 4))) | ((y_true==2) &amp; ((y_pred == 3)|(y_pred == 4))) | ((y_true==3)&amp;((y_pred == 4)))| ((y_true==15)&amp;((y_pred == 16) | (y_pred == 17)| (y_pred == 18))) | ((y_true==16) &amp; ((y_pred == 17)|(y_pred == 18))) | ((y_true==17)&amp;((y_pred == 18))) | ((y_true==6) &amp; ((y_pred == 7)|(y_pred == 8)))|((y_true==7)&amp;((y_pred == 8)))|  ((y_true==19) &amp; ((y_pred == 20)|(y_pred == 21)))|((y_true==20)&amp;((y_pred == 21)))|((y_true==10)&amp;((y_pred == 11)))|((y_true==13)&amp;((y_pred == 14)))|((y_true==22)&amp;((y_pred == 23)))|((y_true==24)&amp;((y_pred == 14)))   )\n    new_pred = tf.where_v2(mask,y_true,y_pred)\n    bool_tensor = tf.equal(new_pred,y_pred)\n    same = tf.math.reduce_all(bool_tensor)\n    session = tf.keras.backend.get_session(op_input_list=())\n    K.set_session(same)\n\n    if session.run(same)==True:\n        return sparse_categorical_accuracy(y_true, new_pred)#K.mean(K.equal(K.argmax(y_true, axis=-1), K.argmax(new_pred, axis=-1)))\n    else:\n        return tf.zeros([16, 25], tf.int32)\n<\/code><\/pre>\n\n<p>(both tensors have shape [16, 25])<\/p>\n\n<p>something similar to where_v2 from keras.backend \nor at least a method to show if my code works somehow(?) would be great<\/p>\n\n<p>But in the end:\nAll i want is to make the conditional confusion for the mentioned conditions possible in my custom accuracy metrics and i would be extremely glad for anyones helpful idea (thank you)  :'\/<\/p>\n","answers":[]}
+{"QuestionId":57649543,"AnswerCount":0,"Tags":"<tensorflow><centos><gpu>","CreationDate":"2019-08-25T20:37:31.763","AcceptedAnswerId":null,"OwnerUserId":6450489.0,"Title":"NVIDIA NVML Driver\/library version mismatch on CentOS","Body":"<p>I have a PC with CentOS OS and two NVIDIA GPUs and I want to use them in Tensorflow (I am not familiar with CentOS so just tried commands as I found in the corresponding web sites). A current GPU driver has been installed already and I didn't change it. Then, I first installed (mistakenly) the last Cuda toolkit version (10.1) as instructed on NVIDIA site:<\/p>\n\n<pre><code>sudo yum-config-manager --add-repo http:\/\/developer.download.nvidia.com\/compute\/cuda\/repos\/rhel7\/x86_64\/cuda-rhel7.repo\nsudo yum clean all\nsudo yum -y install nvidia-driver-latest-dkms cuda\n<\/code><\/pre>\n\n<p>I also installed CuDNN (which is essentially copying some files to related directories). Tensorflow then notified me that I have to revert to toolkit version 10.0. I couldn't find any uninstall file in cuda directory and just installed cuda toolkit 10.0 as instructed in NVIDIA  site <a href=\"https:\/\/developer.nvidia.com\/cuda-10.0-download-archive?target_os=Linux&amp;target_arch=x86_64&amp;target_distro=CentOS&amp;target_version=7&amp;target_type=rpmnetwork\" rel=\"nofollow noreferrer\">Cuda 10.0 installation<\/a>. I noticed that after that, I have both 10.0 and 10.1 folders in my \/usr\/local but there is also a cuda link that is referencing cuda10.0. After this installation, Tensorflow worked fine and everything was ok but system halted when I tried to stop the tensorflow in a simulation and then I couldn't work with GPUs anymore. I tried <code>nvidia-smi<\/code> command (which was working fine previously) but got <code>NVIDIA NVML Driver\/library version mismatch<\/code> error, and restarting the system didn't help.<\/p>\n\n<p>I looked at different web sites and based on them tried to uninstall cuda to reinstall it. But I can not figure out how to do that as there was no uninstall script in cuda folders. Finally I ran <code>sudo yum remove cuda<\/code> and it reported that it had removed cuda10.1. But I still get the <code>Driver\/library mismatch<\/code> error and the cuda10.1 folder still exist. I also see among the log output that system says <code>Repository cuda is listed more than once in the configuration<\/code>, that I have no clue what does it mean.<\/p>\n\n<p>How can I remove all previous cuda toolkit installations and their dependencies in CentOS to start fresh?<\/p>\n","answers":[]}
+{"QuestionId":57649973,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-25T21:47:17.647","AcceptedAnswerId":57650053.0,"OwnerUserId":10665023.0,"Title":"Tensorflow 1.14 equivalent of Tensorflow 2.0 model.trainable_variables","Body":"<p>How can I replace Tensorflow 2.0 <code>model.trainable_variables<\/code> in Tensorflow.1.14 (<a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/effective_tf2\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/guide\/effective_tf2<\/a>) ?<\/p>\n","answers":[{"AnswerId":"57650053","CreationDate":"2019-08-25T22:04:00.793","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>Assuming that you have two <code>tf.keras.Model<\/code>:<\/p>\n\n<pre><code>model_1 = tf.keras.Sequential([tf.keras.layers.Dense(100, name=\"model_1\/dense1\")])\nmodel_1.build(input_shape=(None, 50))\nmodel_2 = tf.keras.Sequential([tf.keras.layers.Dense(100, name=\"model_2\/dense1\")])\nmodel_2.build(input_shape=(None, 50))\n<\/code><\/pre>\n\n<p>You can obtain the trainable variables of the first model like this:<\/p>\n\n<pre><code>model_1_trainable = [variable for variable in tf.trainable_variables() if variable.name.startswith(\"model_1\")]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57650068,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-25T22:10:03.790","AcceptedAnswerId":57660640.0,"OwnerUserId":11552661.0,"Title":"Tensorflow value error : Cannot feed value of shape (8009,) for Tensor 'Placeholder_5:0', which has shape '(?, 1)'","Body":"<p>The error occurs when trying to do linear regression using tensorflow. I tried to fix it with the flattern () function but it still doesn't work. Also I tried some solution from this website. <\/p>\n\n<p>below is some code and full traceback:<\/p>\n\n<pre><code># Convert test features and Labels to Numpy Arrays\nX_test = np.array(X_test)\ny_test = np.array(y_test)\n\n# Converting from Pandas Dataframe to Numpy Arrays\nX_train = np.array(X_train)\ny_train = np.array(y_train)\n\n# Define Training Parameters\n\n# Learning Rate\nlr = 0.1\n\n# Number of epochs for which the model will run\nepochs = 1000\n\n# Define Features and Label Placeholders\n\n# Features\nX = tf.placeholder(tf.float32,[None,X_train.shape[1]])\n\n# Labels \ny = tf.placeholder(tf.float32,[None,1])\n\n# Define Hyperparameters\n\n# Weight\nW = tf.Variable(tf.ones([X_train.shape[1], 1]))\n\n# Bias\nb = tf.Variable(tf.ones(X_train.shape[1]))\n\n# Initiaize all Variables\ninit = tf.global_variables_initializer()\n\n# Define Cost Function, Optimizer and the Output Predicitons Function\n\n# Predictions\n# y_hat = (W*X + b)\ny_hat = tf.add(tf.matmul(X, W), b)\n\n# Cost Function\n# MSE\ncost = tf.reduce_mean(tf.square(y - y_hat))\n\n# Gradient Descent Optimizer to Minimize the Cost\noptimizer = tf.train.GradientDescentOptimizer(learning_rate=lr).minimize(cost)\n\n# Tensor to store the cost after every Epoch\n# Comes in handy while plotting the cost vs epochs\ncost_history = np.empty(shape=[1],dtype=float)\n\n\n\n    with tf.Session() as sess:\n        # Initialize all Variables\n        sess.run(init)\n\n        for epoch in range(0,epochs):\n            # Run the optimizer and the cost functions\n            result, err = sess.run([optimizer, cost], feed_dict={X: X_train, y: y_train})\n\n        # Add the calculated cost to the array\n        cost_history = np.append(cost_history,err)\n\n        # Print the Loss\/Error after every 100 epochs\n        if epoch%100 == 0:\n            print('Epoch: {0}, Error: {1}'.format(epoch, err))\n\n    print('Epoch: {0}, Error: {1}'.format(epoch+1, err))\n\n    # Values of Weight &amp; Bias after Training\n    new_W = sess.run(W)\n    new_b = sess.run(b)\n\n    # Predicted Labels\n    y_pred = sess.run(y_hat, feed_dict={X: X_test})\n\n    # Mean Squared Error\n    mse = sess.run(tf.reduce_mean(tf.square(y_pred - y_test)))\n<\/code><\/pre>\n\n<p>Traceback:<\/p>\n\n<pre><code>ValueError                                Traceback (most recent call last)\n&lt;ipython-input-190-76177d01d272&gt; in &lt;module&gt;\n      5     for epoch in range(0,epochs):\n      6         # Run the optimizer and the cost functions\n----&gt; 7         result, err = sess.run([optimizer, cost], feed_dict={X: X_train, y: y_train})\n      8 \n      9         # Add the calculated cost to the array\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py in run(self, fetches, feed_dict, options, run_metadata)\n    948     try:\n    949       result = self._run(None, fetches, feed_dict, options_ptr,\n--&gt; 950                          run_metadata_ptr)\n    951       if run_metadata:\n    952         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   1147                              'which has shape %r' %\n   1148                              (np_val.shape, subfeed_t.name,\n-&gt; 1149                               str(subfeed_t.get_shape())))\n   1150           if not self.graph.is_feedable(subfeed_t):\n   1151             raise ValueError('Tensor %s may not be fed.' % subfeed_t)\n\nValueError: Cannot feed value of shape (8009,) for Tensor 'Placeholder_5:0', which has shape '(?, 1)'\n<\/code><\/pre>\n\n<p>seems to mismatch between the shape of what I am feeding in and what the TensorFlow is expecting. I  know that this is a mistake but I don't know how to fix it.<\/p>\n","answers":[{"AnswerId":"57660640","CreationDate":"2019-08-26T15:14:10.833","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>The error is saying that the shape of <code>y_train<\/code> is <code>(8009,)<\/code> while TF is expecting a numpy array of shape <code>(8009, 1)<\/code> for placeholder <code>y<\/code>. Expanding dimensions of <code>y_train<\/code> along the last axis should solve the problem:<\/p>\n\n<pre><code>result, err = sess.run([optimizer, cost], feed_dict={X: X_train, y: y_train[:, None]})\n<\/code><\/pre>\n"}]}
+{"QuestionId":57650172,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-25T22:32:13.483","AcceptedAnswerId":null,"OwnerUserId":11274618.0,"Title":"Statefulness in eager mode vs non-eager mode","Body":"<p>I have an RNN that works in eager mode but not otherwise, and I'd like to understand why.<\/p>\n\n<p>The RNN is supposed to accept a sequence of numbers and output their cumulative sum.  Here's the code:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\ntf.enable_eager_execution()\n\ninit_state_getter = lambda shape, dtype: np.ones(shape).astype(np.float32) # was ones\n\nrnn = tf.keras.layers.SimpleRNN(\n    1, # 1-dimensional state\n    stateful=True,   \n    return_sequences=True,\n    activation=None, \n    use_bias=False,\n    kernel_initializer = init_state_getter,\n    recurrent_initializer = init_state_getter,\n    return_state=False\n)\n\nx = np.ones([1,1,1]).astype(np.float32)\n\nfor _ in range(5):\n    print(rnn(x))\n<\/code><\/pre>\n\n<p>and this works, we get the output<\/p>\n\n<pre><code>tf.Tensor([[[1.]]], shape=(1, 1, 1), dtype=float32)\ntf.Tensor([[[2.]]], shape=(1, 1, 1), dtype=float32)\ntf.Tensor([[[3.]]], shape=(1, 1, 1), dtype=float32)\ntf.Tensor([[[4.]]], shape=(1, 1, 1), dtype=float32)\ntf.Tensor([[[5.]]], shape=(1, 1, 1), dtype=float32)\n<\/code><\/pre>\n\n<p>Now let's modify this for not eager mode:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ninit_state_getter = lambda shape, dtype: np.ones(shape).astype(np.float32)\n\nwith tf.Session() as sess:\n\n    x = tf.placeholder(tf.float32, shape=(1,1,1))\n\n    rnn = tf.keras.layers.SimpleRNN(\n        1, \n        stateful=True,   \n        return_sequences=True,\n        activation=None, \n        use_bias=False,\n        kernel_initializer = init_state_getter,\n        recurrent_initializer = init_state_getter,\n        return_state=False\n    )\n    output = rnn(x)\n    init = tf.global_variables_initializer()\n\n    sess.run(init)\n\n\n    for _ in range(5):\n        print(sess.run(output, feed_dict={x:np.ones([1,1,1]).astype(np.float32)}))\n<\/code><\/pre>\n\n<p>This time, the RNN seems to keep losing its state, even though I am never re-initiazing:<\/p>\n\n<pre><code>[[[1.]]]\n[[[1.]]]\n[[[1.]]]\n[[[1.]]]\n[[[1.]]]\n<\/code><\/pre>\n\n<p>Now I think I could explicitly feed the old state into RNN as part of the feed_dict, but (1) I'm surprised that I need to do this, given that we've set stateful=True and (2) I have no idea how to fix this if I am using an estimator.<\/p>\n","answers":[]}
+{"QuestionId":57651128,"AnswerCount":0,"Tags":"<python><keras><nan><loss>","CreationDate":"2019-08-26T02:24:56.947","AcceptedAnswerId":null,"OwnerUserId":11015365.0,"Title":"How do I fix loss: nan for Keras in Python?","Body":"<p>I'm trying to perform a regression using Keras in Python. During training, the loss, mean_squared_error, val_mean_squared_error are all nan. How can I fix this? <\/p>\n\n<p>I've checked thoroughly, and there are no missing values (NaNs) or infinite values in my data. I've also tried gradient value clipping and gradient norm scaling, but none of them worked. <\/p>\n\n<p>Below is my neural net model.<\/p>\n\n<pre><code>model = Sequential()\n\nn_cols = train_X.shape[1]\n\nearly_stopping_monitor = EarlyStopping(patience=10)\n\nmodel.add(Dense(400, activation='relu', input_shape=(n_cols,)))\nmodel.add(Dense(400, activation='relu'))\nmodel.add(Dense(400, activation='relu'))\nmodel.add(Dense(400, activation='relu'))\nmodel.add(Dense(1, activation='linear'))\n\nmodel.compile(optimizer='adam', loss='mean_squared_error', metrics=['mean_squared_error'])\n\nmodel.fit(train_X, train_y, validation_split=0.2, epochs=500, callbacks=[early_stopping_monitor])\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57651278,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-08-26T02:56:17.173","AcceptedAnswerId":null,"OwnerUserId":11429035.0,"Title":"Can someone explain or summarize the input shape of keras under different type of neural networks?","Body":"<p>I am very new to the python keras. And with the understanding of the Keras, I'm confusing about the input shape of Keras. I feel under different neural network, I need to reconstruct my data into different shapes.<\/p>\n\n<p>For example, if I'm building a simple ANN, my train data should be a matrix like [m, n], the m is the number of samples and n is the number of feature. But recently I'm learning 1D convolutional neural network. I found the tutorial construct the training data as [a, b, c], where the a is the number of sample, b is the number of timestep, c is the number of feature (equals to 1). But why I can't simply reshape the data into [a, b]? since c will always be 1 for a 1D convolutional neural network.<\/p>\n\n<p>I'm not sure if I understand the above right. I am just wondering is there a summarize of the training_data shape of different neural networks? Or is there any logic behind the shape of data? So I can always make sure my training data has the right format.<\/p>\n\n<p>The different neural networks mean like ANN, 1D CNN, 2d CNN, RNN and so on.<\/p>\n","answers":[{"AnswerId":"57658971","CreationDate":"2019-08-26T13:33:15.967","ParentId":null,"OwnerUserId":"11919769","Title":null,"Body":"<p>Generally, CNN takes 4 dimensions as input data. Whenever you train a CNN the model in Keras it will automatically convert the input data into 4D. If you want to predict using your CNN model you have to make sure that even your output data\/ or the data you want to run the inference on should have the same dimensions as input data. You can simply add <code>0<\/code> or <code>None<\/code> at the output data by using <code>numpy expand_dims<\/code> function. <\/p>\n"}]}
+{"QuestionId":57651287,"AnswerCount":0,"Tags":"<numpy><tensorflow><deep-learning>","CreationDate":"2019-08-26T02:58:02.847","AcceptedAnswerId":null,"OwnerUserId":9427280.0,"Title":"How to swap tensor axes efficiently in tensorflow?","Body":"<p>I have to swap tensor's axes using <code>tf.transpose<\/code> to do the batch matrix multiplication (as the code shown below). <\/p>\n\n<p>tensor input_a: shape [10000, 10000] <\/p>\n\n<p>tensor input_b: shape [batch_size, 10000, 10] <\/p>\n\n<p>tensor output:  shape [batch_size, 10000, 10] <\/p>\n\n<pre><code># reshape_input_b: shape [10000, batch_size, 10]\ntranspose_input_b = tf.transpose(input_b, [1, 0, 2])\n\n# transpose_input_b : shape [10000, batch_size * 10]\nreshape_input_b = tf.reshape(transpose_input_b , [10000, -1])\n\n# ret: shape [10000, batch_size * 10]\nret = tf.matmul(input_a, reshape_input_b, a_is_sparse = True)\n\n# reshape_ret: [10000, batch_size, 10]\nreshape_ret = tf.reshape(ret, [10000, -1, 10])\n\n# output : [batch_size, 10000, 10]\noutput = tf.transpose(reshape_ret, [1, 0, 2])\n<\/code><\/pre>\n\n<p>However, it seems very slow. I noticed this in the document page of <code>tf.transpose<\/code>:<\/p>\n\n<blockquote>\n  <p>In numpy transposes are memory-efficient constant time operations as they simply return a new view of the same data with adjusted strides.<\/p>\n  \n  <p>TensorFlow does not support strides, <b>so transpose returns a new tensor with the items permuted<\/b>.<\/p>\n<\/blockquote>\n\n<p>So, I think it might be the reason why my code run slowly? Is there any way to swap tensor's axes, or do the batch matrix multiplication efficiently?<\/p>\n","answers":[]}
+{"QuestionId":57651678,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-26T04:15:21.257","AcceptedAnswerId":null,"OwnerUserId":11976470.0,"Title":"ValueError: No variables to optimize in GradientDescentOptimizer","Body":"<p>I am trying to make a simple tensorflow 2.0 code for a linear regression   <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\nx = tf.random.uniform([3,10])\ncoeff = tf.constant([[1.,2.,3.]])\nintercept = 5. \n\ndef calcy(x=x, coeff=coeff, intercept=intercept):\n    return tf.linalg.matmul(coeff, x)+intercept     \n\ny = calcy()\n\n@tf.function\ndef train(x=x, y=y):\n    train_coeff = tf.Variable([[0,0,0]], dtype = tf.float32)\n    train_intercept = tf.Variable(0, dtype = tf.float32)\n\n    result_y = calcy(x, train_coeff, train_intercept) \n\n    loss = tf.math.reduce_mean(tf.math.square(result_y-y))\n    for _ in range(10):\n        tf.compat.v1.train.GradientDescentOptimizer(0.5).minimize(loss)\n\ntrain()\n<\/code><\/pre>\n\n<p>It returns ValueError: No variables to optimize.<\/p>\n","answers":[{"AnswerId":"57767074","CreationDate":"2019-09-03T07:40:52.853","ParentId":null,"OwnerUserId":"11976470","Title":null,"Body":"<p>I changed the train part, and it now works<\/p>\n\n<pre><code>train_coeff = tf.Variable([[0,0,0]], dtype = tf.float32)\ntrain_intercept = tf.Variable(0, dtype = tf.float32)\n\noptimizer = tf.keras.optimizers.Adam()\n\n@tf.function\ndef train(x=x, y=y):\n    with tf.GradientTape() as tape:\n        result_y = calcy(x, train_coeff, train_intercept)\n        loss = tf.math.reduce_mean(tf.math.square(result_y-y))\n        gradient = tape.gradient(loss, (train_coeff, train_intercept))\n        optimizer.apply_gradients(zip(gradient, (train_coeff, train_intercept)))\n\n\nfor _ in range(10000):\n    train()\n<\/code><\/pre>\n"}]}
+{"QuestionId":57652723,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><recurrent-neural-network><seq2seq>","CreationDate":"2019-08-26T06:23:26.080","AcceptedAnswerId":null,"OwnerUserId":11839345.0,"Title":"How to evaluate inference accuracy for seq2seq video captionnig model?","Body":"<p>I've built a <strong>video captionnig model<\/strong>.<br>\nIt consists of a Seq2seq model, taking video as input and outputting natural language<\/p>\n\n<p>I obtain <strong>really good test results<\/strong> but horrible inference results: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Epoch 1 ; Batch loss: 5.181570 ; Batch accuracy: 60.28% ; Test accuracy: 00.89%\n...\nEpoch 128 ; Batch loss: 0.628466 ; Batch accuracy: 96.31% ; Test accuracy: 00.81% \n<\/code><\/pre>\n\n<h2>Explanation<\/h2>\n\n<p>This accuracy is low because of my <strong>accuracy function<\/strong>: it compares <strong>word by word<\/strong> the given result with the caption. <\/p>\n\n<p>This computation is adapted for training, because of teacher's forcing mechanism, but <strong>not adapted for inference<\/strong>.<\/p>\n\n<h3>Example<\/h3>\n\n<blockquote class=\"spoiler\">\n  <p> <img src=\"https:\/\/zupimages.net\/up\/19\/35\/4nto.gif\" alt=\"video1734\"><\/p>\n<\/blockquote>\n\n<p><strong>True descriptions:<\/strong><\/p>\n\n<ul>\n<li><code>a football match is going on &lt;end&gt;<\/code><\/li>\n<li><code>the football player are made a goal &lt;end&gt;<\/code><\/li>\n<li><code>the crowd cheers as soccer players work hard to gain control of the ball &lt;end&gt;<\/code><\/li>\n<\/ul>\n\n<p><strong>Generated descrition:<\/strong><\/p>\n\n<p><code>a group of young men play a game of soccer &lt;end&gt;<\/code><\/p>\n\n<p>My model correctly understands what's happening, but it doesn't express it exaclty (word by word) like the awaited description...<br>\nFor this specific example, accuracy value will be only 1\/31.<\/p>\n\n<h3>How can I wisely compute inference accuracy ?<\/h3>\n\n<p>I thought about extracting the keywords of the sentences. Then trying to see if all keywords contained in the predicted sentence can be found somehere in the captions.<br>\nBut I also have to check if the sentence is a correct english sentence...<\/p>\n\n<p>Maybe you think about an easier way to compute accuracy. Tell me !<\/p>\n","answers":[{"AnswerId":"57653970","CreationDate":"2019-08-26T08:02:42.547","ParentId":null,"OwnerUserId":"6037560","Title":null,"Body":"<p>User <code>Bleu Score<\/code> aka Bilingual Evaluation Understudy Score to compare hypotheses and references. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def bleu_score(hypotheses, references):\n    return nltk.translate.bleu_score.corpus_bleu(references, hypotheses)\n<\/code><\/pre>\n\n<p>Example:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># two references for one document\nfrom nltk.translate.bleu_score import corpus_bleu\nreferences = [[['this', 'is', 'a', 'test'], ['this', 'is' 'test']]]\nhypotheses = [['this', 'is', 'a', 'test']]\nscore = corpus_bleu(references, hypotheses)\nprint(score)\n<\/code><\/pre>\n\n<p>Output: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>1.0\n<\/code><\/pre>\n\n<p>Other methods are:<\/p>\n\n<ol>\n<li><p>METEOR<\/p><\/li>\n<li><p>ROUGE_L<\/p><\/li>\n<li><p>CIDEr<\/p><\/li>\n<\/ol>\n\n<p>Follow: <a href=\"https:\/\/github.com\/arjun-kava\/Video2Description\/blob\/VideoCaption\/cocoeval.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/arjun-kava\/Video2Description\/blob\/VideoCaption\/cocoeval.py<\/a><\/p>\n"}]}
+{"QuestionId":57652728,"AnswerCount":0,"Tags":"<keras><resnet>","CreationDate":"2019-08-26T06:23:32.543","AcceptedAnswerId":null,"OwnerUserId":9207437.0,"Title":"how to define a residual network with only fully connected layers (residual dnn?) in keras?","Body":"<p>Im trying to build very deep DNN model, so I want to apply the residual concept from resnet to fully connected network (DNN) in keras. <\/p>\n\n<p>I have tried simple multi-layers DNN in my regression-based project which only contains 1d samples rather than 2d-figures, it did work just as following codes show:<\/p>\n\n<p>For example:  <\/p>\n\n<p>model = Sequential()<\/p>\n\n<p>model.add(Dense(128, activation='relu', input_dim=20))<\/p>\n\n<p>model.add(Dropout(0.5))<\/p>\n\n<p>model.add(Dense(64, activation='relu'))<\/p>\n\n<p>model.add(Dropout(0.5))<\/p>\n\n<p>model.add(Dense(1))<\/p>\n\n<p>But I really have no idea of how to include residual concept or how to build the residual block... googled for many key words without any useful info...<\/p>\n","answers":[]}
+{"QuestionId":57652906,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2019-08-26T06:39:35.083","AcceptedAnswerId":57670352.0,"OwnerUserId":11158631.0,"Title":"tf.train.latest_checkpoint returning none when passing checkpoint path","Body":"<p>When I am trying to load checkpoint after training ENet model for prediction using tf.train.latest_checkpoint(), it's returning \"None\" though I am passing the correct checkpoint path.<\/p>\n\n<p>Here is my code:<\/p>\n\n<pre><code>     image_dir = '.\/dataset\/test\/'\n     images_list = sorted([os.path.join(image_dir, file) for file in \n     os.listdir(image_dir) if file.endswith('.png')])\n\n     checkpoint_dir = \".\/checkpoint_mk\"\n     listi = os.listdir(checkpoint_dir)\n     print(listi)\n     checkpoint = tf.train.latest_checkpoint(\".\/log\/original\/check\")\n     print(checkpoint,'--------------------------------------- \n     ++++++++++++++++++++++++++++++++++++++++++++++++++++')\n<\/code><\/pre>\n\n<p>It returns None.<\/p>\n\n<p>I am passing absolute path of checkpoint as they store in some other Dir.<\/p>\n\n<p>Here are my checkpoint folder.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/P1NFt.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/P1NFt.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>EDIT ---------------<\/p>\n\n<p>model_checkpoint_path: \"model.ckpt-400\"<\/p>\n\n<p>all_model_checkpoint_paths: \"model.ckpt-0\"<\/p>\n\n<p>all_model_checkpoint_paths: \"model.ckpt-400\"<\/p>\n","answers":[{"AnswerId":"57670352","CreationDate":"2019-08-27T08:12:32.367","ParentId":null,"OwnerUserId":"1849975","Title":null,"Body":"<p>The <code>tf.train.latest_checkpoint<\/code> path argument needs to be relative to your current directory (from which Python script is executed). If it is a complex structure (i.e. data set is stored in a different folder or on a HDD) you can simply use absolute path to the folder. That is why <code>tf.train.latest_checkpoint(\"\/home\/nikhil_m\/TensorFlow-ENet\/log\/original\")<\/code> works in this case.<\/p>\n"}]}
+{"QuestionId":57653204,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2019-08-26T07:06:31.337","AcceptedAnswerId":57654314.0,"OwnerUserId":139150.0,"Title":"keras pre-processing of text using one_hot class","Body":"<p>I came across this code while learning keras online.<\/p>\n\n<pre><code>from keras.preprocessing.text import one_hot\nfrom keras.preprocessing.text import text_to_word_sequence\n\ntext = 'One hot encoding in Keras'\ntokens = text_to_word_sequence(text)\nlength = len(tokens)\none_hot(text, length)\n<\/code><\/pre>\n\n<p>This returns the intergers like this...<\/p>\n\n<blockquote>\n  <p>[3, 1, 1, 2, 3]<\/p>\n<\/blockquote>\n\n<p>I did not understand why and how does unique words return duplicate numbers. For e.g. 3 and 1 is repeated even if the words in the text are unique.<\/p>\n","answers":[{"AnswerId":"57654314","CreationDate":"2019-08-26T08:27:07.287","ParentId":null,"OwnerUserId":"10455493","Title":null,"Body":"<p>From the documentation of <code>one_hot<\/code> it is described how it is a wrapper of <code>hashing_trick<\/code>:<\/p>\n\n<blockquote>\n  <p>This is a wrapper to the <code>hashing_trick<\/code> function using hash as the hashing function; unicity of word to index mapping non-guaranteed.<\/p>\n<\/blockquote>\n\n<p>From the documentation of <code>hasing_trick<\/code>:<\/p>\n\n<blockquote>\n  <p>Two or more words may be assigned to the same index, due to possible collisions by the hashing function. The probability of a collision is in relation to the dimension of the hashing space and the number of distinct objects.<\/p>\n<\/blockquote>\n\n<p>Since hashing is used there is a probability that different words will be hashed to the same index. The probability of a non-unique hash is proportional to the vocabulary size selected.\nIt is suggested by Jason Brownlee <a href=\"https:\/\/machinelearningmastery.com\/prepare-text-data-deep-learning-keras\/\" rel=\"nofollow noreferrer\">Jason Brownlee<\/a> to use a vocabulary size 25% larger than the word size to increase the uniqueness of the hashes.<\/p>\n\n<p>Following Jason Brownlee suggestion in you case results in:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.keras.preprocessing.text import one_hot\nfrom tensorflow.keras.preprocessing.text import text_to_word_sequence\nfrom tensorflow.random import set_random_seed\nimport math\n\nset_random_seed(1)\ntext = 'One hot encoding in Keras'\ntokens = text_to_word_sequence(text)\nlength = len(tokens)\nprint(one_hot(text, math.ceil(length*1.25)))\n<\/code><\/pre>\n\n<p>which returns the integers<\/p>\n\n<blockquote>\n  <p>[3, 4, 5, 1, 6]<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":57653444,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><deep-learning>","CreationDate":"2019-08-26T07:23:32.707","AcceptedAnswerId":null,"OwnerUserId":10117868.0,"Title":"Deep learning for image training","Body":"<p>I tried training images and predict the text in it. But when training images all together I was getting array error. But now I am training each image of a letter but I am getting some error. The image generator file is added where it helps to create the images and importing it to the fit generator.<\/p>\n\n<p>Error:<\/p>\n\n<pre><code>Using TensorFlow backend.\nWARNING: Logging before flag parsing goes to stderr.\nW0826 09:18:45.040408  3428 deprecation_wrapper.py:119] From C:\\Users\\workspace\\test\\venv\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py:74: The name tf.get_default_graph is deprecated. Please use tf.compat.v1.get_default_graph instead.\n\nW0826 09:18:45.056031  3428 deprecation_wrapper.py:119] From C:\\Users\\workspace\\test\\venv\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py:517: The name tf.placeholder is deprecated. Please use tf.compat.v1.placeholder instead.\n\nW0826 09:18:45.071652  3428 deprecation_wrapper.py:119] From C:\\Users\\workspace\\test\\venv\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py:4138: The name tf.random_uniform is deprecated. Please use tf.random.uniform instead.\n\nTraceback (most recent call last):\n  File \"C:\/Users\/workspace\/test\/killme.py\", line 22, in &lt;module&gt;\n    o2 = Reshape((len(string.ascii_uppercase), ), name=\"symbol_{}\".format(i+1))(o)\n  File \"C:\\Users\\workspace\\test\\venv\\lib\\site-packages\\keras\\engine\\base_layer.py\", line 474, in __call__\n    output_shape = self.compute_output_shape(input_shape)\n  File \"C:\\Users\\workspace\\test\\venv\\lib\\site-packages\\keras\\layers\\core.py\", line 398, in compute_output_shape\n    input_shape[1:], self.target_shape)\n  File \"C:\\Users\\workspace\\test\\venv\\lib\\site-packages\\keras\\layers\\core.py\", line 386, in _fix_unknown_dimension\n    raise ValueError(msg)\nValueError: total size of new array must be unchanged\n<\/code><\/pre>\n\n<p>Code:<\/p>\n\n<pre><code>from keras.models import Sequential, Input, Model\nfrom keras.layers import Dense, Reshape\nfrom keras.utils import to_categorical\nfrom keras.layers.convolutional import Conv2D # to add convolutional layers\nfrom keras.layers.convolutional import MaxPooling2D # to add pooling layers\nfrom keras.layers import Flatten # to flatten data for fully connected layers\nimport string\nfrom generators import ImageGenerator, BasicGenerator\nfrom numpy import reshape\nheight=20\nwidth=200\nfont_size=20\n\ni1=Input(shape=(height, width, 1))\ncharacter_count=int(width \/ font_size)\noutputs=[]\nfor i in range(character_count):\n    o = Dense(len(string.ascii_uppercase), activation='relu')(i1)\n    o2 = Reshape((len(string.ascii_uppercase), ), name=\"symbol_{}\".format(i+1))(o)\n    outputs.append(o2)\n\nstring_model = Model(inputs=i1, outputs=outputs)\nstring_model.layers[2].layer.trainable = False\n\ngenerator = ImageGenerator(height, width, font_size, character_count)\nstring_model.compile(loss=\"categorical_crossentropy\", optimizer=\"rmsprop\", metrics=[\"categorical_accuracy\"])\nstring_model.summary()\nstring_model.fit_generator(generator,epochs=10)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57653616","CreationDate":"2019-08-26T07:38:06.283","ParentId":null,"OwnerUserId":"6037560","Title":null,"Body":"<p>You have to preprocess text data before just putting into the output dense. Converting into <code>vocab<\/code> would be a better idea. Create a <code>CaptionGenerator<\/code> to make it simple as follows.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from vgg16 import VGG16\nfrom keras.applications import inception_v3\nimport numpy as np\nimport pandas as pd\nfrom keras.models import Sequential\nfrom keras.layers import LSTM, Embedding, TimeDistributed, Dense, RepeatVector, Merge, Activation, Flatten\nfrom keras.preprocessing import image, sequence\nfrom keras.callbacks import ModelCheckpoint\nimport cPickle as pickle\n\nEMBEDDING_DIM = 128\n\n\nclass CaptionGenerator():\n\n    def __init__(self):\n        self.max_cap_len = None\n        self.vocab_size = None\n        self.index_word = None\n        self.word_index = None\n        self.total_samples = None\n        self.encoded_images = pickle.load( open( \"encoded_images.p\", \"rb\" ) )\n        self.variable_initializer()\n\n    def variable_initializer(self):\n        df = pd.read_csv('Flickr8k_text\/flickr_8k_train_dataset.txt', delimiter='\\t')\n        nb_samples = df.shape[0]\n        iter = df.iterrows()\n        caps = []\n        for i in range(nb_samples):\n            x = iter.next()\n            caps.append(x[1][1])\n\n        self.total_samples=0\n        for text in caps:\n            self.total_samples+=len(text.split())-1\n        print \"Total samples : \"+str(self.total_samples)\n\n        words = [txt.split() for txt in caps]\n        unique = []\n        for word in words:\n            unique.extend(word)\n\n        unique = list(set(unique))\n        self.vocab_size = len(unique)\n        self.word_index = {}\n        self.index_word = {}\n        for i, word in enumerate(unique):\n            self.word_index[word]=i\n            self.index_word[i]=word\n\n        max_len = 0\n        for caption in caps:\n            if(len(caption.split()) &gt; max_len):\n                max_len = len(caption.split())\n        self.max_cap_len = max_len\n        print \"Vocabulary size: \"+str(self.vocab_size)\n        print \"Maximum caption length: \"+str(self.max_cap_len)\n        print \"Variables initialization done!\"\n\n\n    def data_generator(self, batch_size = 32):\n        partial_caps = []\n        next_words = []\n        images = []\n        print \"Generating data...\"\n        gen_count = 0\n        df = pd.read_csv('Flickr8k_text\/flickr_8k_train_dataset.txt', delimiter='\\t')\n        nb_samples = df.shape[0]\n        iter = df.iterrows()\n        caps = []\n        imgs = []\n        for i in range(nb_samples):\n            x = iter.next()\n            caps.append(x[1][1])\n            imgs.append(x[1][0])\n\n\n        total_count = 0\n        while 1:\n            image_counter = -1\n            for text in caps:\n                image_counter+=1\n                current_image = self.encoded_images[imgs[image_counter]]\n                for i in range(len(text.split())-1):\n                    total_count+=1\n                    partial = [self.word_index[txt] for txt in text.split()[:i+1]]\n                    partial_caps.append(partial)\n                    next = np.zeros(self.vocab_size)\n                    next[self.word_index[text.split()[i+1]]] = 1\n                    next_words.append(next)\n                    images.append(current_image)\n\n                    if total_count&gt;=batch_size:\n                        next_words = np.asarray(next_words)\n                        images = np.asarray(images)\n                        partial_caps = sequence.pad_sequences(partial_caps, maxlen=self.max_cap_len, padding='post')\n                        total_count = 0\n                        gen_count+=1\n                        print \"yielding count: \"+str(gen_count)\n                        yield [[images, partial_caps], next_words]\n                        partial_caps = []\n                        next_words = []\n                        images = []\n\n    def load_image(self, path):\n        img = image.load_img(path, target_size=(224,224))\n        x = image.img_to_array(img)\n        return np.asarray(x)\n\n\n    def create_model(self, ret_model = False):\n        #base_model = VGG16(weights='imagenet', include_top=False, input_shape = (224, 224, 3))\n        #base_model.trainable=False\n        image_model = Sequential()\n        #image_model.add(base_model)\n        #image_model.add(Flatten())\n        image_model.add(Dense(EMBEDDING_DIM, input_dim = 4096, activation='relu'))\n\n        image_model.add(RepeatVector(self.max_cap_len))\n\n        lang_model = Sequential()\n        lang_model.add(Embedding(self.vocab_size, 256, input_length=self.max_cap_len))\n        lang_model.add(LSTM(256,return_sequences=True))\n        lang_model.add(TimeDistributed(Dense(EMBEDDING_DIM)))\n\n        model = Sequential()\n        model.add(Merge([image_model, lang_model], mode='concat'))\n        model.add(LSTM(1000,return_sequences=False))\n        model.add(Dense(self.vocab_size))\n        model.add(Activation('softmax'))\n\n        print \"Model created!\"\n\n        if(ret_model==True):\n            return model\n\n        model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])\n        return model\n\n    def get_word(self,index):\n        return self.index_word[index]\n<\/code><\/pre>\n\n<p>Follow Link For more info: : <a href=\"https:\/\/github.com\/arjun-kava\/caption_generator\" rel=\"nofollow noreferrer\">https:\/\/github.com\/arjun-kava\/caption_generator<\/a><\/p>\n"}]}
+{"QuestionId":57653570,"AnswerCount":1,"Tags":"<conv-neural-network><caffe>","CreationDate":"2019-08-26T07:34:29.120","AcceptedAnswerId":57657799.0,"OwnerUserId":4417582.0,"Title":"How to calculate precision, recall, F1 score for my caffe model","Body":"<pre><code>Framework: Caffe\nArchitecture: Mobilenet-SSD\nDataset: [Caltech Padestrain Detection Dataset][1]\n<\/code><\/pre>\n\n<p>I know the formulae for precision, recall, F1 score &amp; accuracy. I know the formulas but the problem here is automation of that<\/p>\n\n<p>Manually I can calculate everything but the problem is with automating that for ~10k images in test dataset<\/p>\n\n<p>Because I can see an image &amp; compare that what is false positive etc etc. But not sure how to do with computer<\/p>\n\n<p>Because my model might detect a person but the bounding box is little bigger(coordinates are bit off) &amp; test dataset labels are accurate <\/p>\n\n<p>So although the detection is happening but coordinates are not matching.<\/p>\n\n<p>How to solve this problem to calculate precision, recall &amp; accuracy? If this is not the correct way please propose correct way<\/p>\n","answers":[{"AnswerId":"57657799","CreationDate":"2019-08-26T12:18:32.460","ParentId":null,"OwnerUserId":"8784825","Title":null,"Body":"<p>Hmmm good question. I think you should define what you want to actually measure. F1 score, precision and recall are easy. Just see if a person has been detected or not, or you have a false positive and calculate things business as usual. For that use <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.metrics.precision_recall_fscore_support.html\" rel=\"nofollow noreferrer\">scikit<\/a>. <\/p>\n\n<p>Now about the bounding boxes, areas and coordinates. You should use a different metric! I recommend you mAP (mean Average Precision). Check out this <a href=\"https:\/\/medium.com\/@timothycarlen\/understanding-the-map-evaluation-metric-for-object-detection-a07fe6962cf3\" rel=\"nofollow noreferrer\">link<\/a> and feel free to read more about this on the internet. Good luck with your model!<\/p>\n\n<p>An alternate option, is use the following gist <a href=\"https:\/\/gist.github.com\/Coderx7\/3b02f6a8eb61110d7d0876e75933f98d\" rel=\"nofollow noreferrer\">confusionMatrix_convnet_test_BatchMode(VeryFast).py<\/a> to build the confusion matrix using sklearn.<\/p>\n"}]}
+{"QuestionId":57653741,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorflow2.0>","CreationDate":"2019-08-26T07:46:59.607","AcceptedAnswerId":57654110.0,"OwnerUserId":3860928.0,"Title":"What does `tf.keras.Model.compile` do in TF2.0?","Body":"<p>In TF2.0, it is still necessary to call <code>model.compile<\/code> after creating a keras model, as seen in the <a href=\"https:\/\/www.tensorflow.org\/guide\/keras#set_up_training\" rel=\"nofollow noreferrer\">doc<\/a>.<\/p>\n\n<p>But with eager mode, there is no need for graph creation, so what is the use of <code>tf.keras.Model.compile<\/code> in Tensorflow 2.0?<\/p>\n","answers":[{"AnswerId":"57654110","CreationDate":"2019-08-26T08:12:53.050","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p><code>model.compile<\/code> has nothing to do with graph mode, it defines the loss, optimizer, and metrics, same as with <code>keras<\/code> and <code>tf.keras<\/code> in  TF 1.x. The graph is just an internal detail that is not really related to <code>model.compile<\/code>.<\/p>\n"}]}
+{"QuestionId":57654899,"AnswerCount":0,"Tags":"<tensorflow><error-handling><file-not-found>","CreationDate":"2019-08-26T09:07:48.470","AcceptedAnswerId":null,"OwnerUserId":2242157.0,"Title":"FileNotFoundError with Tensorflow and Dataset API","Body":"<p>I am using Tensorflow and the Dataset API to do some classification tasks. However, when I train or test my network, occasionaly I get a <strong>FileNotFoundError<\/strong>. Data is stored in a network drive on a Windows Server and I am suspicious that the error might be due to some latency or missing package issue with the connection.<\/p>\n\n<p>The error happens occasionaly and when it crashes it is a different file each time. I have checked paths etc. so file is definitely there and the path is correct as well.<\/p>\n\n<p>Could it be related to the network as I suspected? If so, is there a way to introduce a mechanism to cope with this?<\/p>\n","answers":[]}
+{"QuestionId":57654921,"AnswerCount":1,"Tags":"<python><keras><lstm>","CreationDate":"2019-08-26T09:09:17.930","AcceptedAnswerId":null,"OwnerUserId":8180213.0,"Title":"Input 0 is incompatible with layer lstm_24: expected ndim=3, found ndim=2","Body":"<p>I'm new to keras and I want to build a deep RNN my x_train shape is (1115,106) and my y_train has (1115,11) samples, I can't fix the error I get from the code below:<\/p>\n\n<pre><code>    model_3=Sequential()\n\n    model_3.add(LSTM(64,return_sequences=True,activation='relu'))\n    model_3.add(Dropout(0.2))\n\n    model_3.add(LSTM(64,activation='relu'))\n    model_3.add(Dropout(0.2))\n\n    model_3.add(Dense(64,activation='relu'))\n    model_3.add(Dropout(0.2))\n\n    model_3.add(Dense(11,activation='softmax'))\n\n\n    model_3.compile(optimizer='adam',\n          loss='mean_squared_error',\n          metrics=['accuracy'])\n\n    DNN_3= model_3.fit(x_train, y_train,batch_size=64, epochs=100)\n<\/code><\/pre>\n\n<p> in ()\n     24               metrics=['accuracy'])\n     25 \n---> 26 DNN_3= model_3.fit(x_train, y_train,batch_size=64, epochs=100)\n     27 \n     28 <\/p>\n\n<p>ValueError: Input 0 is incompatible with layer lstm_24: expected ndim=3, found ndim=2<\/p>\n","answers":[{"AnswerId":"57666475","CreationDate":"2019-08-27T00:53:23.257","ParentId":null,"OwnerUserId":"10161091","Title":null,"Body":"<p>You have to do two things:<\/p>\n\n<ol>\n<li><p>LSTM takes input of shape <code>(n_samples, n_timesteps, n_features)<\/code> and that's why it \"expected ndim=3\". So if you have 1000 data points over time (rounding 1115 for simplicity), then you can reshape your data into shape of <code>(40, 25, 106)<\/code> for input and <code>(40, 25, 11)<\/code> for output. Choices of n_samples and n_timesteps depends on the problem and should be decided by you. This will fix the error you are getting now, but then you would run into another error.<\/p><\/li>\n<li><p>Add <code>return_sequences=True<\/code> to the second LSTM as well as it seems you need a many-to-many architecture. <\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":57655007,"AnswerCount":1,"Tags":"<python><machine-learning><pytorch>","CreationDate":"2019-08-26T09:15:13.540","AcceptedAnswerId":57655319.0,"OwnerUserId":8220188.0,"Title":"Finetune PyTorch model after training fc layers","Body":"<p>I am trying to do Transfer Learning using PyTorch. I wan to train fc layers first and then to finetune the whole network. Unfortunately after training fc layers and then passing my network to finetune, I am losing the accuracy that was acquired in the first training. Is this an expected behaviour or am I doing something wrong here?<\/p>\n\n<p>Here is the code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = torchvision.models.resnet50(pretrained=True)\nfor param in model.parameters():\n    param.requires_grad = False\n\nnum_ftrs = model.fc.in_features\nmodel.fc = nn.Linear(num_ftrs, 4)\nmodel = model.to(device)\n\ncriterion = nn.CrossEntropyLoss()\noptimizer = optim.Adam(model.parameters(), lr=0.001)\n\nmodel = trainer.fit_model(dataloader, model, criterion, optimizer, num_epochs=10)\n# fit model is basic PyTorch training function found here: https:\/\/pytorch.org\/tutorials\/beginner\/transfer_learning_tutorial.html#convnet-as-fixed-feature-extractor The only difference is that scheduler is an optional param.\n\nfor param in model.parameters():\n    param.requires_grad = True\n\ntorch.cuda.empty_cache()\nexp_lr_scheduler = lr_scheduler.StepLR(optimizer, step_size=7, gamma=0.1)\n\n# Here I am finetuning the model\nmodel_ft = trainer.fit_model(\n    dataloader, model, criterion, optimizer, scheduler=exp_lr_scheduler, num_epochs=10\n)\n<\/code><\/pre>\n\n<p>Am I missing something here or should I just train the model once?<\/p>\n","answers":[{"AnswerId":"57655319","CreationDate":"2019-08-26T09:36:12.833","ParentId":null,"OwnerUserId":"10511793","Title":null,"Body":"<p>That is something that can happen when performing transfer learning called \n <strong>catastrophic forgetting<\/strong>. Basically, you update your pretrained weights too much and you 'forget' what was previously learned. This can happen notably if your learning rate is too high. I would suggest trying at first a lower learning rate, or using diffentiable learning rate (different learning rate for the head of the network and the pretrained part, so that you can have a higher learning rate on the fc layers than for the rest of the network).<\/p>\n"}]}
+{"QuestionId":57655017,"AnswerCount":0,"Tags":"<google-cloud-platform><deep-learning><pytorch>","CreationDate":"2019-08-26T09:16:08.127","AcceptedAnswerId":null,"OwnerUserId":2047442.0,"Title":"How to use all vCPUs with `pytorch`?","Body":"<p>I'm training 2 <code>pytorch<\/code> models on a Google cloud instance <code>n1-standard-8 (8 vCPUs, 30 GB memory)<\/code>. In the first case I have only about <code>50%<\/code> utilization (<code>400%<\/code> in the picture below) but in the second case it can be much higher (around <code>600-700%<\/code>). In both cases I'm training <code>NLP<\/code> models with <code>LSTMs<\/code>. How can I increase utilization in the 1st case?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/MgpWF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/MgpWF.png\" alt=\"&lt;code&gt;top&lt;\/code&gt; command screen\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57655028,"AnswerCount":0,"Tags":"<pytorch><random-seed><seeding><reproducible-research><hyperopt>","CreationDate":"2019-08-26T09:16:51.337","AcceptedAnswerId":null,"OwnerUserId":7678823.0,"Title":"Unable to reproduce result obtained from hyperparameter tuning using hyperopt","Body":"<p>I have built a Pytorch model and performed hyperparameters tuning using library Hyperopt. The result obtained is not reproducible despite I have already call the below seeding function at the beginning of each run:<\/p>\n\n<p><b>util.py<\/b><\/p>\n\n<pre><code>def seed_everything(seed=42):\n    random.seed(seed)\n    torch.manual_seed(seed)\n    torch.cuda.manual_seed(seed)\n    torch.cuda.manual_seed_all(seed)\n    np.random.seed(seed)\n    torch.backends.cudnn.deterministic = True\n    torch.backends.cudnn.benchmark = False\n<\/code><\/pre>\n\n<p><b>Trainer.py<\/b><\/p>\n\n<pre><code>    def train(self, params):\n        util.seed_everything()\n        #wrote my training code here\n\n<\/code><\/pre>\n\n<p>Upon futher inspection, I have found that the <b>first<\/b> result of hyperopt tuning is always reproducible, but the subsequent runs are not. This is unexpected for me as I have already called <i>seed_everthing()<\/i> at the beginning of the train function.<\/p>\n\n<p>In addition, if I run the training as below:<\/p>\n\n<pre><code>    for i in range(2):\n        print (\"in iteration \", i)\n        trainer = Trainer(**configs)\n        trainer.train(params)\n<\/code><\/pre>\n\n<p>The result of iteration 1 and 2 is different to each other, but they are always the same(i.e. iteration 1 always gives train_loss of 1.31714 while iteration 2 always gives 4.31235)<\/p>\n\n<p>I am hoping that iteration 1 and iteration 2 would gives the same result since it should be reproducible.<\/p>\n","answers":[]}
+{"QuestionId":57655161,"AnswerCount":0,"Tags":"<tensorflow><object-detection><google-colaboratory><training-data><object-detection-api>","CreationDate":"2019-08-26T09:25:36.697","AcceptedAnswerId":null,"OwnerUserId":11900232.0,"Title":"Unable to train a custom objector detector","Body":"<p>I'm trying to train my own custom object detector. I have the tf records of both test and train and also label_map.tbtxt. there are a lot of issues with each step.<\/p>\n","answers":[]}
+{"QuestionId":57655507,"AnswerCount":0,"Tags":"<python><keras><cross-validation>","CreationDate":"2019-08-26T09:48:38.820","AcceptedAnswerId":null,"OwnerUserId":3937131.0,"Title":"How to use flow_from_directory function when we want implement cross validation","Body":"<p>I have a small image dataset which I employed ImageDataGenerator class in Keras to augment my dataset. I put my dataset in a folder so I take advantage of flow_from_directory function to load and use the images.<\/p>\n\n<p>Now I have to implement k-fold on my code. and I don't know how to manage my dataset since the name of each image is its label. <\/p>\n\n<p>Does anybody have any idea to handle this situation?<\/p>\n","answers":[]}
+{"QuestionId":57655633,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-08-26T09:56:39.333","AcceptedAnswerId":null,"OwnerUserId":11977777.0,"Title":"How i Can resize an image from 3000x2900 to 64x64 using keras?","Body":"<p>I have to classfier a medical images, but these images are big(3000x2900),, i need way to resize. and then training it.\nCan i training these images without resize???                                               <\/p>\n","answers":[{"AnswerId":"57656844","CreationDate":"2019-08-26T11:15:02.437","ParentId":null,"OwnerUserId":"11351803","Title":null,"Body":"<p>You can train your model without resizing your pictures . But it will be time consuming. And you may not get the best results. I recommend you resize them to 128x128 pixels. You can use the PIL library and resize each picture then save it to a different directory.<\/p>\n\n<pre><code>from PIL import Image\nimg = Image.open('\/your iamge path\/image.jpg') # image extension *.png,*.jpg\nnew_width  = 128\nnew_height = 128\nimg = img.resize((new_width, new_height), Image.ANTIALIAS)\nimg.save('\/new directory path\/output image name.png') # format may what u want ,*.png,*jpg,*.gif\n<\/code><\/pre>\n\n<p>This process is called preprocessing the dataset. <\/p>\n"}]}
+{"QuestionId":57655862,"AnswerCount":2,"Tags":"<python><keras>","CreationDate":"2019-08-26T10:11:07.337","AcceptedAnswerId":57656938.0,"OwnerUserId":7820400.0,"Title":"Movie Recommendation system: Expected to see 2 array(s), but instead got the following list of 1 arrays","Body":"<p>I'm completely new to Machine Learning and I've been trying to make a movie recommendation system using keras, I've been following a tutorial online but I can't understand Why am I getting the following error: <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([[130767],\n       [   110]], dtype=int32)]...\n<\/code>\nI know it has something to do with input shape but I've been searching for 6 straight hours and I still can't understand what's wrong, any help will be appreciated, thank you very much =)<\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nimport os\nfrom os import path\nimport matplotlib.pyplot as plt\nfrom keras.models import Model\nfrom keras.models import load_model\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.preprocessing import LabelEncoder\nfrom keras.layers import Input, Reshape, Dot\nfrom keras.layers.embeddings import Embedding\nfrom keras.optimizers import Adam\nfrom keras.regularizers import l2\nfrom keras.layers import Add, Activation, Lambda\nPATH = '.\/ml-20m\/'\n\nratings = pd.read_csv(PATH + 'ratings.csv')\n# print(ratings.head(n=10))\n\nmovies = pd.read_csv(PATH + 'movies.csv')\n# print(movies.head(n=10))\n\ng = ratings.groupby('userId')['rating'].count()\ntop_users = g.sort_values(ascending=False)[:15]\n\ng = ratings.groupby('movieId')['rating'].count()\ntop_movies = g.sort_values(ascending=False)[:15]\n\ntop_r = ratings.join(top_users, rsuffix='_r', how='inner', on='userId')\ntop_r = top_r.join(top_movies, rsuffix='_r', how='inner', on='movieId')\n# print(pd.crosstab(top_r.userId, top_r.movieId, top_r.rating, aggfunc=np.sum))\n\nuser_enc = LabelEncoder()\nratings['user'] = user_enc.fit_transform(ratings['userId'].values)\nn_users = ratings['user'].nunique()\n\nitem_enc = LabelEncoder()\nratings['movie'] = item_enc.fit_transform(ratings['movieId'].values)\nn_movies = ratings['movie'].nunique()\nratings['rating'] = ratings['rating'].values.astype(np.float32)\nmin_rating = min(ratings['rating'])\nmax_rating = max(ratings['rating'])\n\nX = ratings[['user', 'movie']].values\ny = ratings['rating'].values\nX = X[:90003]\ny = y[:90003]\nX_train, X_test, y_train, y_test = train_test_split(\n    X, y, test_size=0.1, random_state=42)\n\nn_factors = 50\nX_train_array = [X_train[:, 0], X_train[:, 1]]\nX_test_array = [X_test[:, 0], X_test[:, 1]]\n\n\ndef RecommenderV1(n_users, n_movies, n_factors):\n    user = Input(shape=(1,))\n    u = Embedding(n_users, n_factors, embeddings_initializer='he_normal',\n                  embeddings_regularizer=l2(1e-6))(user)\n    u = Reshape((n_factors,))(u)\n\n    movie = Input(shape=(1,))\n    m = Embedding(n_movies, n_factors, embeddings_initializer='he_normal',\n                  embeddings_regularizer=l2(1e-6))(movie)\n    m = Reshape((n_factors,))(m)\n\n    x = Dot(axes=1)([u, m])\n    model = Model(inputs=[user, movie], outputs=x)\n    opt = Adam(lr=0.001)\n    model.compile(loss='mean_squared_error', optimizer=opt)\n    return model\n\n\nmodel = RecommenderV1(n_users, n_movies, n_factors)\n# model.summary()\n\nif(os.path.exists('recommendation_model.h5')):\n    model = load_model('recommendation_model.h5')\nelse:\n    history = model.fit(x=X_train_array, y=y_train, batch_size=64, epochs=10,verbose=1, validation_data=(X_test_array, y_test))\n    model.save(\"recommendation_model.h5\")\n    # plt.plot(history.history['loss'])\n    # plt.xlabel(\"Epochs\")\n    # plt.ylabel('Training Error')\nwhile(True):\n    # text = input(\"Please enter your input\\n\")\n    # numbers = text.split(' ')\n    # userID = int(numbers[0])\n    # movieID = int(numbers[1])\n    # inputArray = np.array((np.array(userID), np.array(movieID)))\n    # test_val = np.array(([userID], [movieID]))\n    # userID = np.array(130767)\n    # movieID = np.array(110)\n    inArr = np.array([[130767], [110]], np.int32)\n    print(inArr.shape)\n    result = model.predict(inArr)\n    print(result)\n    # print(test_val)\n    # print(inputArray.shape)\n    print('---------------------------------------------------------')\n\n<\/code><\/pre>\n","answers":[{"AnswerId":"57656938","CreationDate":"2019-08-26T11:21:30.017","ParentId":null,"OwnerUserId":"11978290","Title":null,"Body":"<p>Since I can't comment yet:<\/p>\n\n<p><code>inArr = [[130767], [110]]<\/code>\nis another easy alternative to solve your problem.<\/p>\n"},{"AnswerId":"57656695","CreationDate":"2019-08-26T11:06:08.890","ParentId":null,"OwnerUserId":"5835828","Title":null,"Body":"<p>From what I understand the model needs a list of two numpy arrays, whereas you are passing a single numpy array to the model.I think you should change <code>inArr = np.array([[130767], [110]], np.int32)<\/code> to <code>inArr = [np.array([130767]),np.array([110])<\/code>, so that now you will be passing two numpy arrays in the list instead of one array.<br>\nHope it helps.<\/p>\n"}]}
+{"QuestionId":57655992,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2019-08-26T10:19:46.513","AcceptedAnswerId":57656053.0,"OwnerUserId":11709475.0,"Title":"How to Concatenate Mnist Test and Train Images?","Body":"<p>I'm trying to train a Generative Adversarial Network. To train the network I'm using mnist dataset. I will train the network with concatenated test and train images.<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nfrom tensorflow.examples.tutorials.mnist import input_data\nmnist=input_data.read_data_sets(\"data\/mnist\",one_hot=False)\nimages=np.concatenate(mnist.test.images,mnist.train.images)\n<\/code><\/pre>\n\n<p>An error has been occurred when I run the code. <\/p>\n\n<pre><code>---------------------------------------------------------------------------\n\nTypeError                                 Traceback (most recent call last)\n\n&lt;ipython-input-9-02ac414642a1&gt; in &lt;module&gt;()\n      3 from tensorflow.examples.tutorials.mnist import input_data\n      4 mnist=input_data.read_data_sets(\"data\/mnist\",one_hot=False)\n----&gt; 5 images=np.concatenate(mnist.test.images,mnist.train.images)\n\nTypeError: only integer scalar arrays can be converted to a scalar index\n<\/code><\/pre>\n\n<p>How to solve that or is there another way to concatenate <code>mnist.test.images<\/code> and <code>mnist.train.images<\/code> arrays?<\/p>\n","answers":[{"AnswerId":"57656053","CreationDate":"2019-08-26T10:22:58.710","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>That's not how you should use <code>numpy.concatenate<\/code>. You can do it like this:<\/p>\n\n<pre><code>images = np.concatenate([mnist.test.images, mnist.train.images], axis=0)\n<\/code><\/pre>\n\n<p>If you go through the <code>numpy.concatenate<\/code> <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.concatenate.html\" rel=\"nofollow noreferrer\">documentation<\/a>, you will see that as a first argument the <code>numpy.concatenate<\/code> expects:<\/p>\n\n<blockquote>\n  <p>a1, a2, \u2026 : sequence of array_like<\/p>\n<\/blockquote>\n\n<p>Therefore, combining the <code>mnist.test.images<\/code> and <code>mnist.train.images<\/code> in an array as in the snippet above solves your problem. Additionally, even though the second argument <code>axis<\/code> default value is <code>axis=0<\/code>, I tend to specify it for clarity.<\/p>\n"}]}
+{"QuestionId":57656165,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2019-08-26T10:30:34.980","AcceptedAnswerId":null,"OwnerUserId":11978006.0,"Title":"ValueError: Graph disconnected in vgg16","Body":"<pre><code>Traceback:\nmodel = Model(input_tensor,x,name = 'vgg16_trunk')\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/legacy\/interfaces.py\", line 91, in wrapper\n    return func(*args, **kwargs)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/engine\/network.py\", line 93, in __init__\n    self._init_graph_network(*args, **kwargs)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/engine\/network.py\", line 231, in _init_graph_network\n    self.inputs, self.outputs)\n  File \"\/usr\/local\/lib\/python3.6\/dist-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_2:0\", shape=(?, 32, 32, 3), dtype=float32) at layer \"input_2\". The following previous layers were accessed without issue: []\n<\/code><\/pre>\n\n<p>How to solve this problem in vgg16 ??<\/p>\n\n<pre><code>def create_model(input_shape):\n    channel_axis = 1 if K.image_data_format() == \"channels_first\" else -1\n    input_tensor = Input(shape=input_shape)\n    base_model = VGG16(classes=10,input_tensor=None,input_shape=input_shape,include_top=False)\n    x = base_model.output\n    x = BatchNormalization(axis=channel_axis, momentum=mom,\n                           epsilon=eps, gamma_initializer=gamma)(x)\n    x = LeakyReLU(leakiness)(x)\n    model = Model(input_tensor,x,name = 'vgg16_trunk')\n    return model\n<\/code><\/pre>\n","answers":[{"AnswerId":"57675073","CreationDate":"2019-08-27T12:48:18.300","ParentId":null,"OwnerUserId":"11983064","Title":null,"Body":"<p>Pass the input_tensor you created here:<\/p>\n\n<p>input_tensor = Input(shape=input_shape)<\/p>\n\n<p>where base_model is created:<\/p>\n\n<p>base_model = VGG16(classes=10,input_tensor=input_tensor,include_top=False)<\/p>\n\n<p>Please note also, that the tensor will already have the input_shape so it's not necessary to give it as parameter again when creating the base_model.<\/p>\n"}]}
+{"QuestionId":57656538,"AnswerCount":2,"Tags":"<python><tensorflow><object-detection>","CreationDate":"2019-08-26T10:55:16.833","AcceptedAnswerId":57659675.0,"OwnerUserId":11977985.0,"Title":"Is there a way to implement a custom learning rate in Tensorflow Object Detection?","Body":"<p>I am working on an object detection problem and I'd like to use the Cyclical Learning Rate. Thing is, this specific learning does not exist in the protos of Tensorflow Object Detection. I'd like to know if it is possible for me to modify the protos (or another file) as to implement this new learning rate method?<\/p>\n\n<p>I am using Tensorflow 1.14 with the latest updated version of the Tensorflow Object Detection repo.<\/p>\n\n<p>I tried to modify the optimizer.proto and optimizer_pb2.py files. I am only showing the sections I modified.<\/p>\n\n<h2>optimizer.proto<\/h2>\n\n<pre class=\"lang-py prettyprint-override\"><code>\/\/ Configuration message for optimizer learning rate.\nmessage LearningRate {\n  oneof learning_rate {\n    ConstantLearningRate constant_learning_rate = 1;\n    ExponentialDecayLearningRate exponential_decay_learning_rate = 2;\n    ManualStepLearningRate manual_step_learning_rate = 3;\n    CosineDecayLearningRate cosine_decay_learning_rate = 4;\n    CosineDecayRestartLearningRate cosine_decay_restart_learning_rate = 5; \/\/ Added\n  }\n}\n...\n\/\/ Added for test\nmessage CosineDecayRestartLearningRate {\n  optional uint32 total_steps = 1 [default = 400000];\n}\n<\/code><\/pre>\n\n<h2>optimizer_pb2.py<\/h2>\n\n<pre class=\"lang-py prettyprint-override\"><code>_LEARNINGRATE = _descriptor.Descriptor(\n...\nfull_name='object_detection.protos.LearningRate.cosine_decay_restart_learning_rate', index=4,\n      number=5, type=11, cpp_type=10, label=1,\n      has_default_value=False, default_value=None,\n      message_type=None, enum_type=None, containing_type=None,\n      is_extension=False, extension_scope=None,\n      options=None),\n...)\n...\n_COSINEDECAYRESTARTLEARNINGRATE = _descriptor.Descriptor(\n  name='CosineDecayRestartLearningRate',\n  full_name='object_detection.protos.CosineDecayRestartLearningRate',\n  filename=None,\n  file=DESCRIPTOR,\n  containing_type=None,\n  fields=[\n    _descriptor.FieldDescriptor(\n      name='total_steps', full_name='object_detection.protos.CosineDecayLearningRate.total_steps', index=1,\n      number=2, type=13, cpp_type=3, label=1,\n      has_default_value=True, default_value=4000000,\n      message_type=None, enum_type=None, containing_type=None,\n      is_extension=False, extension_scope=None,\n      options=None),\n  ],\n  extensions=[\n  ],\n  nested_types=[],\n  enum_types=[\n  ],\n  options=None,\n  is_extendable=False,\n  syntax='proto2',\n  extension_ranges=[],\n  oneofs=[\n  ],\n  serialized_start=1671,\n  serialized_end=1861,\n)\n_LEARNINGRATE.fields_by_name['cosine_decay_restart_learning_rate'].message_type = _COSINEDECAYRESTARTLEARNINGRATE\n_LEARNINGRATE.oneofs_by_name['learning_rate'].fields.append(\n  _LEARNINGRATE.fields_by_name['cosine_decay_restart_learning_rate'])\n_LEARNINGRATE.fields_by_name['cosine_decay_restart_learning_rate'].containing_oneof = _LEARNINGRATE.oneofs_by_name['learning_rate']\nDESCRIPTOR.message_types_by_name['CosineDecayRestartLearningRate'] = _COSINEDECAYRESTARTLEARNINGRATE\nCosineDecayRestartLearningRate = _reflection.GeneratedProtocolMessageType('CosineDecayRestartLearningRate', (_message.Message,), dict(\n  DESCRIPTOR = _COSINEDECAYRESTARTLEARNINGRATE,\n  __module__ = 'object_detection.protos.optimizer_pb2'\n  # @@protoc_insertion_point(class_scope:object_detection.protos.CosineDecayRestartLearningRate)\n  ))\n_sym_db.RegisterMessage(CosineDecayRestartLearningRate)\n<\/code><\/pre>\n\n<p>I didn't expect it to work since at no point did I reach a step where I add to input code for the learning rate, and it does give me an error.<\/p>\n\n<pre><code>  File \"\/home\/renart\/Tensorflow\/models\/research\/object_detection\/protos\/optimizer_pb2.py\", line 253, in &lt;module&gt;\n    options=None),\n  File \"\/home\/renart\/Tensorflow\/venv-1.13\/lib\/python3.5\/site-packages\/google\/protobuf\/descriptor.py\", line 534, in __new__\n    return _message.default_pool.FindFieldByName(full_name)\nKeyError: \"Couldn't find field object_detection.protos.LearningRate.cosine_decay_restart_learning_rate\"\n\n<\/code><\/pre>\n\n<p>Unfortunately, the error comes from Protobuf and not Tensorflow, so I still have no idea where to look in order to implement the learning rate.<\/p>\n","answers":[{"AnswerId":"57659675","CreationDate":"2019-08-26T14:15:04.867","ParentId":null,"OwnerUserId":"7349742","Title":null,"Body":"<p>I do not think you should modify <code>optimizer_pb2.py<\/code> manually. It is generated by <code>protoc<\/code> command. The exact step is in here <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/installation.md#protobuf-compilation\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/installation.md#protobuf-compilation<\/a><\/p>\n\n<p>You should only modify <code>optimizer.proto<\/code> and follow the above instruction to recompile the .proto files. <\/p>\n\n<p>To add a new learning rate schedule, you may consider add it in this file <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/utils\/learning_schedules.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/utils\/learning_schedules.py<\/a>\nsince it is where the official learning rate schedules reside. <\/p>\n\n<p>The configurations for learning rate is interpreted in this file: <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/0b3a8abf095cb8866ca74c2e118c1894c0e6f947\/research\/object_detection\/builders\/optimizer_builder.py#L103\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/0b3a8abf095cb8866ca74c2e118c1894c0e6f947\/research\/object_detection\/builders\/optimizer_builder.py#L103<\/a>\nYou should call build your learning rate schedule here. <\/p>\n"},{"AnswerId":"57850224","CreationDate":"2019-09-09T08:20:25.813","ParentId":null,"OwnerUserId":"11977985","Title":null,"Body":"<p>My question has been answered by @Fan Luo, but I'm still going to write the steps I took to correctly set up my work.\nFirst of all, go to the protos\/optimizer.proto file and add your learning rate, just like in the first code box of my question. You'll notice that I change my learning rate name, don't be confused.<\/p>\n\n<p>Then, modify the builders\/optimizer_builder.py file to describe what the message in proto should return.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>if learning_rate_type == 'cyclical_learning_rate':\n    config = learning_rate_config.cyclical_learning_rate\n    learning_rate = learning_schedules.cyclical_learning_rate(\n        tf.train.get_or_create_global_step()\n    )\n<\/code><\/pre>\n\n<p>Go to utils\/learning_schedules.py and add the logic behind the new learning rate.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def cyclical_learning_rate(global_step):\n        return tf.Variable(0.001, name='learning_rate')\n<\/code><\/pre>\n\n<p>Finally, <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/installation.md#protobuf-compilation\" rel=\"nofollow noreferrer\">compile your protos again<\/a> and you are done. You can now use your learning rate just like any other learning rate in tensorflow object detection.<\/p>\n"}]}
+{"QuestionId":57656661,"AnswerCount":2,"Tags":"<python><python-3.x><keras><deep-learning><conv-neural-network>","CreationDate":"2019-08-26T11:03:40.537","AcceptedAnswerId":null,"OwnerUserId":11920225.0,"Title":"How to prediction more than one images after training model with keras?","Body":"<p>I want to try to predict an image after I have trained with my model.<\/p>\n\n<p>With the code below I can only predict one image, how can I predict more than 50 images for example?<\/p>\n\n<p><strong>Example code<\/strong> <\/p>\n\n<pre><code>import numpy as np\nfrom keras.preprocessing import image\nfrom keras.preprocessing.image import ImageDataGenerator\n\ntest_image = image.load_img('dataset\/test_image\/girls.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:\n  prediction = 'nsfw'\nelse:\n  prediction = 'sfw'\n\nprint(prediction)\n<\/code><\/pre>\n\n<p>Can you tell me what I should change or add to the code?<\/p>\n\n<p>Maybe you have a reference code to predict more than one image you could share?<\/p>\n","answers":[{"AnswerId":"57657147","CreationDate":"2019-08-26T11:33:48.097","ParentId":null,"OwnerUserId":"7903159","Title":null,"Body":"<p>To batch predict a directory of 50 images you could follow this example <a href=\"https:\/\/androidkt.com\/how-to-predict-images-using-trained-keras-model\" rel=\"nofollow noreferrer\">here<\/a> and do the following:<\/p>\n\n<pre><code>batch_holder = np.zeros((50, IMG_SIZE, IMG_SIZE, 3))\nimg_dir='test_set\/'\nfor i,img in enumerate(os.listdir(img_dir)):\n  img = image.load_img(os.path.join(img_dir,img), target_size=(IMG_SIZE,IMG_SIZE))\n  batch_holder[i, :] = img\n<\/code><\/pre>\n\n<p>where in your case <code>IMG_SIZE = 64<\/code>.\nThen you can just call the method <code>predict_classes()<\/code> of the <code>model<\/code> object:<\/p>\n\n<pre><code>classes = loaded_model.predict_classes(batch_holder)\n<\/code><\/pre>\n\n<p>There is also this <a href=\"https:\/\/gist.github.com\/ritiek\/5fa903f97eb6487794077cf3a10f4d3e\" rel=\"nofollow noreferrer\">GitHub Gist<\/a> that does it with a similar way.<\/p>\n"},{"AnswerId":"57657254","CreationDate":"2019-08-26T11:40:35.590","ParentId":null,"OwnerUserId":"10407128","Title":null,"Body":"<p>What would you like your output to be? This sounds like it could be done with a simple for loop. You can probably do it more efficient in Keras, but taking a simple approach this also works. The code below takes all images in the folder of <code>dataset\/test_image<\/code> and creates a list with classifications in order of the <code>image_list<\/code>: <\/p>\n\n<pre><code>import numpy as np\nfrom keras.preprocessing import image\nfrom keras.preprocessing.image import ImageDataGenerator\nimport os\n\nimage_folder = 'dataset\/test_image\/'\nimage_list = os.listdir(image_folder)\nclassifier_results = []\nfor image in image_list:\n    test_image = image.load_img(image_folder + image_list[image], target_size = (64, 64))\n    test_image = image.img_to_array(test_image)\n    test_image = np.expand_dims(test_image, axis = 0)\n    result = classifier.predict(test_image)\n    training_set.class_indices\n    if result[0][0] == 1:\n      prediction = 'nsfw'\n    else:\n      prediction = 'sfw'\n    classifier_results.append(prediction)\n    print(image_list[image], \"is classified as: \" prediction)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57656734,"AnswerCount":0,"Tags":"<python><tensorflow><debugging>","CreationDate":"2019-08-26T11:08:06.517","AcceptedAnswerId":null,"OwnerUserId":10126859.0,"Title":"How could I check the specific value of tensor while debugging in TensorFlow 1.X?","Body":"<p>I want to check the specific value of tensor while debugging in TensorFlow 1.14. After checking some similar issues, I found that the following code may be useful for checking the specific value of tensor, which was shown in this picture.<\/p>\n\n<p><strong>The detail debugging code<\/strong><\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/4OEMG.png\" alt=\"Screen Shot\"><\/p>\n\n<p><strong>The related code as follows:<\/strong><\/p>\n\n<pre><code> def sample_z(args, latent_dim):\n     z_mu, z_log_var = args\n     batch_size = K.shape(z_mu)[0]\n     epsilon = K.random_normal(shape=(batch_size, latent_dim), mean=0., \n      stddev=1.0)\n     z = z_mu + K.exp(z_log_var \/ 2) * epsilon  # sigma := var^(1\/2)\n     return z\n<\/code><\/pre>\n\n<p>The above error shows a missing initialization. Then, I added  <code>sess.run(tf.initialize_all_variables())<\/code>. However, this error popped up.<\/p>\n\n<p><strong>The error message<\/strong><\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/3xPva.png\" alt=\"Screen Shot\"><\/p>\n\n<p>I'm certain the z_mu was a tensor. Does anyone know the reason of issue? <\/p>\n","answers":[]}
+{"QuestionId":57657386,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-26T11:49:41.930","AcceptedAnswerId":57657510.0,"OwnerUserId":11029007.0,"Title":"Tensorflow Datasets Reshape Images","Body":"<p>I want to build a data pipeline using tensorflow dataset. Because each data has different shapes, I can't build a data pipeline.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow_datasets as tfds\nimport tensorflow as tf\n\ndataset_builder = tfds.builder(\"oxford_flowers102\")\ndataset_builder.download_and_prepare()\n\ntrain_data = dataset_builder.as_dataset(split=tfds.Split.TRAIN)\ntrain_data = train_data.repeat().batch(32)\ntrain_data = train_data.prefetch(tf.data.experimental.AUTOTUNE)\ntrain_iterator = train_data.make_one_shot_iterator()\ntrain_next_element = train_iterator.get_next()\n\nwith tf.Session() as sess:\n    train_batch = sess.run(train_next_element)\n\n<\/code><\/pre>\n\n<p>Above code gives me the error:<\/p>\n\n<blockquote>\n  <p>\"tensorflow.python.framework.errors_impl.InvalidArgumentError: Cannot batch tensors with different shapes in component 1. First element had shape [500,666,3] and element 1 had shape [752,500,3].\" <\/p>\n<\/blockquote>\n\n<p>I want all images to be in the shape of [224,224,3]. How can I reshape images in the existing tensorflow dataset?<\/p>\n","answers":[{"AnswerId":"57657510","CreationDate":"2019-08-26T11:57:09.407","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>You can dynamically resize the images like:<\/p>\n\n<pre><code>train_data = train_data.map(lambda image: tf.image.resize_image_with_crop_or_pad(image, 224, 224))\n<\/code><\/pre>\n\n<p>right before doing <code>train_data = train_data.repeat().batch(32)<\/code>. Also, using the <code>tf.data.Dataset.map(...)<\/code> method you can apply a variety of transformations on your images before batching them.<\/p>\n"}]}
+{"QuestionId":57657651,"AnswerCount":2,"Tags":"<python><performance><tensorflow><tf.keras>","CreationDate":"2019-08-26T12:07:40.050","AcceptedAnswerId":57688395.0,"OwnerUserId":9988487.0,"Title":"Tensorflow performance (versions 1 vs 2 and CPU vs GPU)","Body":"<p>I'm new to Machine Learning and found myself spending a disproportionate amount of time setting up Tensorflow. I use Anaconda to manage the different versions as environments. I managed to install<\/p>\n\n<ul>\n<li><p>Tensorflow-cpu_1.14.0<\/p><\/li>\n<li><p>Tensorflow-gpu_1.14.0<\/p><\/li>\n<li><p>Tensorflow-cpu_2.0.0-beta1.<\/p><\/li>\n<\/ul>\n\n<p>I did not manage to set up Tensorflow-gpu_2.0.0-beta due to some issues with the CUDA drivers and I've given up on this for the moment.<\/p>\n\n<p>My goal is to make sure the three versions specified above are working properly and using all available resources on my system. In particular, my questions are:<\/p>\n\n<ol>\n<li><p>How does one reliably measure the performance of an existing computer and Tensorflow setup?<\/p><\/li>\n<li><p>Is it normal that for the example I use the CPU-only versions are faster?<\/p><\/li>\n<li><p>How should I go about selecting and installing the optimal Tensorflow setup for my system?<\/p><\/li>\n<\/ol>\n\n<p>I read many topics dealing with performance issues on Windows and comparisons between GPU and CPU runtimes, but none seemed to address the questions above. Thank you very much in advance for any insights concerning my questitons!<\/p>\n\n<p>I did not find any single established standard example to test performance with, so I built my own (probably a grave error). I tested all three environments (using the code specified below) on my home computer (Windows 10 Home, x64. Processor: Intel i7-8750 CPU @2.20GHz, 2208Mhz, 6 Cores, 12 Logical Processors. RAM: 16GB. Graphics Card: GeForce RTX 2060). I also ran the test example from <a href=\"https:\/\/medium.com\/@erikhallstrm\/hello-world-tensorflow-649b15aed18c#.ipfyuf5a2\" rel=\"nofollow noreferrer\">1<\/a>, which demonstrated that matrix multiplication is faster using GPU. I assume there is some reason for the discrepancies, that I'm missing. Please feel free to comment on basic errors in thinking that may be apparent from the code.<\/p>\n\n<pre><code># python 3.6\nimport numpy as np\nimport tensorflow as tf\nfrom tensorflow.python.client import device_lib\nfrom timeit import default_timer as timer\n\nmodel = tf.keras.Sequential()\nmodel.add(tf.keras.layers.Dense(64, activation='relu'))\nmodel.add(tf.keras.layers.Dense(64, activation='relu'))\nmodel.add(tf.keras.layers.Dense(10, activation='softmax'))\n\nmodel.compile(optimizer=tf.compat.v1.train.AdamOptimizer(0.001),\n              loss='categorical_crossentropy',\n              metrics=['accuracy'])\n\n\ndef random_one_hot_labels(shape):\n    n, n_class = shape\n    classes = np.random.randint(0, n_class, n)\n    tmp_labels = np.zeros((n, n_class))\n    tmp_labels[np.arange(n), classes] = 1\n    return tmp_labels\n\n\ndata = np.random.random((1000, 32))\nlabels = random_one_hot_labels((1000, 10))\n\ndurations = []\nfor i in range(10):  # run N times\n    start = timer()\n    model.fit(data, labels, epochs=500, batch_size=32)\n    durations.append(timer() - start)\n\nprint(f\"tf.version.VERSION = {tf.version.VERSION}\")\nprint(f\"tf.keras.__version__ = {tf.keras.__version__}\")\ndevices = device_lib.list_local_devices()  # this may allocate all GPU memory ?!\nprint(f\"devices = {[x.name for x in devices]}\")\nprint(f\"model.fit durations: {durations}\")\n\n<\/code><\/pre>\n\n<p>The CPU-only versions both outperform the GPU version. Additionally, there is a huge difference between the different Tensorflow versions. Below the outputs of my code using the three different environments:<\/p>\n\n<pre><code>tf.version.VERSION = 1.14.0\ntf.keras.__version__ = 2.2.4-tf\n2019-08-26 13:41:15.980626: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1640] Found device 0 with properties:\nname: GeForce RTX 2060 major: 7 minor: 5 memoryClockRate(GHz): 1.2\npciBusID: 0000:01:00.0\n2019-08-26 13:41:15.986261: I tensorflow\/stream_executor\/platform\/default\/dlopen_checker_stub.cc:25] GPU libraries are statically linked, skip dlopen check.\n2019-08-26 13:41:15.990784: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1763] Adding visible gpu devices: 0\n2019-08-26 13:41:15.993919: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1181] Device interconnect StreamExecutor with strength 1 edge matrix:\n2019-08-26 13:41:15.997211: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1187]      0\n2019-08-26 13:41:16.000263: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1200] 0:   N\n2019-08-26 13:41:16.002807: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1326] Created TensorFlow device (\/device:GPU:0 with 4616 MB memory) -&gt; physical GPU (device: 0, name: GeForce RTX 2060, pci bus id: 0000:01:00.0, compute capability: 7.5)\ndevices = ['\/device:CPU:0', '\/device:GPU:0']\nmodel.fit durations: [34.81, 33.84, 34.37, 34.21, 34.54, 34.18, 35.09, 33.24, 33.32, 33.54]\n\n-----------------------------------------------------------\n\ntf.version.VERSION = 1.14.0\ntf.keras.__version__ = 2.2.4-tf\ndevices = ['\/device:CPU:0']\nmodel.fit durations: [23.48, 23.43, 23.25, 23.71, 23.54, 24.017, 23.43, 24.08, 23.67, 23.94]\n\n-----------------------------------------------------------\n\ntf.version.VERSION = 2.0.0-beta1\ntf.keras.__version__ = 2.2.4-tf\ndevices = ['\/device:CPU:0']\nmodel.fit durations: [15.53, 14.87, 14.65, 14.73, 14.68, 14.67, 15.11, 14.71, 15.54, 14.38]\n<\/code><\/pre>\n","answers":[{"AnswerId":"57688395","CreationDate":"2019-08-28T08:47:01.063","ParentId":null,"OwnerUserId":"9988487","Title":null,"Body":"<p>I tested the same code with different net sizes. It turns out that, when using bigger sized networks, the GPU version performs much better than the CPU-only version. I suspect this is due to overhead coming from loading data into GPU memory. <\/p>\n\n<p>If you want to test this, use e.g. 1024 nodes per layer in the above code (and reduce the number of epochs).<\/p>\n"},{"AnswerId":"57658916","CreationDate":"2019-08-26T13:30:09.760","ParentId":null,"OwnerUserId":"11977985","Title":null,"Body":"<p>I have been using Tensorflow for a while now so I will try to answer your questions.<\/p>\n\n<ol>\n<li>One good way to measure your performance is using Tensorboard. It is installed automatically when you install Tensorflow. When training your model, indicate in your code where you want to save your checkpoints. Put them in a folder called \"trainings\" for example. You want to end up with a folder tree that looks like this: <code>trainings\/training_1\/my_model.ckpt<\/code>. Using a terminal, call Tensorboard like this <code>tensorboard --logdir=trainings<\/code>. Tensorboard looks recursively through the folder, so if you have one folder per training, Tensorboard will show you each training separately without you having to run 1 Tensorboard per training. Tensorboard has graphs that indicate mutliple things like the accuracy of the training, the time spent computing, the learning rate etc. As you can see in the next image, I am able to tell that training #1 was faster than #2 by 15mn: <a href=\"https:\/\/i.stack.imgur.com\/yffcH.png\" rel=\"nofollow noreferrer\">Tensorboard graph example<\/a>. If you do not know how to save checkpoints, you can look at <a href=\"https:\/\/www.tensorflow.org\/guide\/checkpoints\" rel=\"nofollow noreferrer\">this link<\/a>.<\/li>\n<li>Looking at you GPU's computing capability, it should have a better fit duration than the CPU. What versions of CUDA and cuDNN do you use ?<\/li>\n<li>Unfortunately, it depends on what you are doing. It is usually better to use the latest release but it can have bugs that weren't in the last versions. I'd go with what you're already doing and create virtual environments for each version I want to use. Keep in mind that if you export your frozen inference graph it will only be able to be used in inference by the same version of Tensorflow you were using when exporting. So if I export my graph while using Tensorflow 1.14, I won't be able to make any inference with Tensorflow 1.13.<\/li>\n<\/ol>\n"}]}
+{"QuestionId":57657805,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-26T12:18:44.250","AcceptedAnswerId":57659086.0,"OwnerUserId":8084623.0,"Title":"Loop within tf.Session()","Body":"<p>In my TensorFlow project I'd like to try different optimizers and ideally I'd loop over them, so I can compare the training in TensorBoard. I've extracted the optimization bit because that fails. I have something like this:<\/p>\n\n<pre><code>import os\nimport tensorflow as tf\n\neta = 0.1\nnum_epochs = 100\n\nxv = tf.Variable(0.0)\nsv = tf.Variable(0, trainable=False)\n\nloss = xv * xv - 4.0 * xv + 5.0\n\noptimizers = [tf.train.GradientDescentOptimizer(eta),\n              tf.train.AdagradOptimizer(eta),\n              tf.train.AdamOptimizer(eta)]\n\ninit = tf.global_variables_initializer()\n\nsaver = tf.train.Saver()\n\nsummary_op = tf.summary.scalar('x', xv)\nwriter = tf.summary.FileWriter('log', graph=tf.get_default_graph())\n\nwith tf.Session() as sess:\n    sess.run(init)\n    for optimizer in optimizers:\n      objective = optimizer.minimize(loss, global_step=sv)\n      for epoch in range(num_epochs):\n          _, step, result, summary = sess.run([objective, sv, xv, summary_op])\n          writer.add_summary(summary, global_step=step)\n          writer.flush()\n      saver.save(sess, os.getcwd() + '\/output')\n      print(sess.run(xv))\n<\/code><\/pre>\n\n<p>It fails with <code>Errors may have originated from an input operation.\nInput Source operations connected to node Adagrad\/update_Variable_2\/ApplyAdagrad:\n Variable_2 (defined at &lt;stdin&gt;:1)<\/code><\/p>\n\n<p>Can something like this be done in TensorFlow or am I using the session wrong in that I'm trying to run the same thing from scratch several times within the same session without re-initializing the variables? I'm not sure how to achieve that though. I tried placing the loop outside of <code>tf.Session()<\/code> but that did not work either.<\/p>\n","answers":[{"AnswerId":"57659086","CreationDate":"2019-08-26T13:41:32.497","ParentId":null,"OwnerUserId":"7349742","Title":null,"Body":"<p>We usually only attach one optimizer to the same sets of variables. <\/p>\n\n<p>When you call <code>optimizer.minimize(loss)<\/code>, tensorflow automatically choose the variables that affect loss to optimize. Calling 3 times <code>optimizer.minimize(loss)<\/code> on the same loss and call <code>session.run(...)<\/code> alternatively is not advisable. <\/p>\n\n<p>If you hope to compare the 3 optimizer, it would be better use 3 scripts to test each optimizer separately.<\/p>\n\n<p>Besides, you should call <code>optimizer.minimize(...)<\/code> for any optimizer only once because repeating call that function only adds new operations to the graph. The actual computing is performed when you call <code>session.run(...)<\/code><\/p>\n"}]}
+{"QuestionId":57657924,"AnswerCount":1,"Tags":"<python><keras><neural-network><gaussian>","CreationDate":"2019-08-26T12:27:50.267","AcceptedAnswerId":57658527.0,"OwnerUserId":9437380.0,"Title":"Change Keras Standard Deviation of Gaussian Noise Layer while training","Body":"<p>What is the fastest way to change the standard deviation of a Gaussian noise layer in Keras during training?<\/p>\n\n<p>Currently I am reloading the whole network with the adapted  standard deviation every iteration, which is really slow.<\/p>\n\n<p>Thank you in advance!<\/p>\n","answers":[{"AnswerId":"57658527","CreationDate":"2019-08-26T13:06:50.193","ParentId":null,"OwnerUserId":"8889727","Title":null,"Body":"<p>Can you try using keras backend variables?<\/p>\n\n<pre><code>from keras import backend as K\n\nself.std=0.5\nself.std_var = K.variable(value=std)\n<\/code><\/pre>\n\n<p>When building the network.<\/p>\n\n<pre><code># instantiate\nstddev = std_var(0.8)\ng = GaussianNoise(stddev)\n<\/code><\/pre>\n\n<p>During training (possibly inside a loop).<\/p>\n\n<pre><code>K.set_value(stddev.std_var, &lt;new_std_val&gt;)\n<\/code><\/pre>\n\n<p>Try this snippet and see whether it works.<\/p>\n"}]}
+{"QuestionId":57657950,"AnswerCount":1,"Tags":"<pytorch><image-segmentation><batchsize>","CreationDate":"2019-08-26T12:29:48.003","AcceptedAnswerId":57665180.0,"OwnerUserId":6908509.0,"Title":"Implementing Batch for Image Segmentation","Body":"<p>I wrote a Python 3.5 script for doing street segmentation. Since I'm new in Image Segementation, I did not use predefined dataloaders from pytorch, instead I wrote them by my self (for better understanding). Until now I only use a batch size of 1. Now I want to generalize this for arbitrary batch sizes.<\/p>\n\n<p>This is a snippet of my Dataloader:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def augment_data(batch_size):\n\n\n    # [...] defining some paths and data transformation (including ToTensor() function)\n\n    # The images are named by numbers (Frame numbers), this allows me to find the correct label image for a given input image.\n    all_input_image_paths = {int(elem.split('\\\\')[-1].split('.')[0]) : elem for idx, elem in enumerate(glob.glob(input_dir + \"*\"))}\n    all_label_image_paths = {int(elem.split('\\\\')[-1].split('.')[0]) : elem for idx, elem in enumerate(glob.glob(label_dir + \"*\"))}\n\n\n    dataloader = {\"train\":[], \"val\":[]}\n    all_samples = []\n    img_counter = 0\n\n    for key, value in all_input_image_paths.items():\n        input_img = Image.open(all_input_image_paths[key])\n        label_img = Image.open(all_label_image_paths[key])\n\n        # Here I use my own augmentation function which crops the input and label on the same position and do other things.\n        # We get a list of new augmented data\n        augmented_images = generate_augmented_images(input_img, label_img)\n        for elem in augmented_images:\n            input_as_tensor = data_transforms['norm'](elem[0])\n            label_as_tensor = data_transforms['val'](elem[1])\n\n            input_as_tensor.unsqueeze_(0)\n            label_as_tensor.unsqueeze_(0)\n\n            is_training_data = random.uniform(0.0, 1.0)\n            if is_training_data &lt;= 0.7:\n                dataloader[\"train\"].append([input_as_tensor, label_as_tensor])\n            else:\n                dataloader[\"val\"].append([input_as_tensor, label_as_tensor])\n            img_counter += 1\n\n    shuffle(dataloader[\"train\"])\n    shuffle(dataloader[\"val\"])\n    dataloader_batched =  {\"train\":[], \"val\":[]}\n\n\n    # Here I group my data to a given batch size\n    for elem in dataloader[\"train\"]:\n        batch = []\n        for i in range(batch_size):\n            batch.append(elem)\n        dataloader_batched[\"train\"].append(batch)\n\n    for elem in dataloader[\"val\"]:\n        batch = []\n        for i in range(batch_size):\n            batch.append(elem)\n        dataloader_batched[\"val\"].append(batch)\n\n    return dataloader_batched\n<\/code><\/pre>\n\n<p>This is a snippet of my training method with batch size 1:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    while epoch &lt;= num_epochs:\n        # Each epoch has a training and validation phase\n        for phase in ['train', 'val']:\n            if phase == 'train':\n                scheduler.step(3)\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\n            counter = 0\n            # Iterate over data.\n            for inputs, labels in dataloaders[phase]:\n                counter += 1\n                max_num = len(dataloaders[phase])\n\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                    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\n\n            epoch_loss = running_loss \/ dataset_sizes[phase]\n<\/code><\/pre>\n\n<p>If I execute this, I get of course the error:<\/p>\n\n<pre><code>for inputs, labels in dataloaders[phase]:\nValueError: not enough values to unpack (expected 2, got 1)\n<\/code><\/pre>\n\n<p>I understand why, because now I have a list of images and not only an input and label image as before. So guessed I need a second for loop which iterates over these batches. So I tried this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>            # Iterate over data.\n            for elem in dataloaders[phase]:\n                for inputs, labels in elem:\n                    counter += 1\n                    max_num = len(dataloaders[phase])\n\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<\/code><\/pre>\n\n<p>But for me it looks like the optimization step (back-prop) is only applied on the last image of the batch. Is that true? And if so, how can I fix this? I guess if I indent the with-Block, then I get again a batch size 1 optimization.<\/p>\n\n<p>Thanks in advance<\/p>\n","answers":[{"AnswerId":"57665180","CreationDate":"2019-08-26T21:30:35.963","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<blockquote>\n  <p>But for me it looks like the optimization step (back-prop) is only applied on the last image of the batch. <\/p>\n<\/blockquote>\n\n<p>It should not apply only based on the last image. It should apply based on the batch size. \nIf you set <code>bs=2<\/code> and it should apply to the batch of two images.<\/p>\n\n<p>Optimization step actually will update the params of your network. Backprop is a fancy name for PyTorch autograd system that computes the first order gradients.<\/p>\n"}]}
+{"QuestionId":57658114,"AnswerCount":0,"Tags":"<tensorflow><artificial-intelligence><google-colaboratory><tpu>","CreationDate":"2019-08-26T12:40:55.750","AcceptedAnswerId":null,"OwnerUserId":6711388.0,"Title":"How to solve \"Propagation of dynamic dimension failed\" error in tf.keras with tpu?","Body":"<p>I have this <strong>error<\/strong> when I try to train an image detector with my own images on google colab with TPU :<\/p>\n\n<blockquote>\n  <p>From \/job:worker\/replica:0\/task:0:\n  Compilation failure: Asked to propagate a dynamic dimension from hlo %convert.283 = f32[1,80,80,32]{3,2,1,0} convert(f32[1,80,80,32]{3,2,1,0} %add.1), metadata={op_type=\"FusedBatchNorm\" op_name=\"bn_Conv1_3\/FusedBatchNorm\"}@{}@0 to hlo %clamp.288 = f32[1,80,80,32]{3,2,1,0} clamp(f32[1,80,80,32]{3,2,1,0} %broadcast.286, f32[1,80,80,32]{3,2,1,0} %convert.283, f32[1,80,80,32]{3,2,1,0} %broadcast.287), metadata={op_type=\"Relu6\" op_name=\"Conv1_relu_3\/Relu6\"}, which is not implemented.\n      TPU compilation failed\n       [[node TPUReplicateMetadata_1 (defined at :24) ]]<\/p>\n<\/blockquote>\n\n<p><strong>Here's the link to the code :<\/strong><\/p>\n\n<p><a href=\"https:\/\/drive.google.com\/open?id=1mPiod1At85RgNwHx4vYFxH38Ck16Ep1m\" rel=\"nofollow noreferrer\">https:\/\/drive.google.com\/open?id=1mPiod1At85RgNwHx4vYFxH38Ck16Ep1m<\/a><\/p>\n\n<p>Do you have any idea of what is going on ?<\/p>\n\n<p>It must not be a problem of size of the pictures or any problem of batch size, I already looked at it.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[]}
+{"QuestionId":57658294,"AnswerCount":1,"Tags":"<python><tensorflow><gradient><training-data><eager-execution>","CreationDate":"2019-08-26T12:53:55.770","AcceptedAnswerId":null,"OwnerUserId":8695468.0,"Title":"Effect of the position of tf.GradientTape() in model training time","Body":"<p>I am trying to update the weight each epoch, but I am processing the data in batches. The problem is, to normalize the loss, I need to tape TensorFlow variables outside the training loop(to be tracked and normalized). But when I do this, the training time is Huge. <\/p>\n\n<p>I think, it accumulates variables from all batches into the graph and calculates gradients at the end.<\/p>\n\n<p>I have started tracking variables outside the for loop and inside the for loop and the later is faster than first. I am confused about why this happens because whatever I do, my model's trainable variables and loss remain the same.<\/p>\n\n<pre><code># Very Slow\n\nloss_value = 0\nbatches = 0\n\nwith tf.GradientTape() as tape:\n    for inputs, min_seq in zip(dataset, minutes_sequence):\n        temp_loss_value = my_loss_function(inputs, min_seq)\n        batches +=1\n        loss_value = loss_value + temp_loss_value\n\n# The following line takes huge time.\ngrads = tape.gradient(loss_value, model.trainable_variables)\n\n<\/code><\/pre>\n\n<pre><code># Very Fast\n\nloss_value = 0\nbatches = 0\n\nfor inputs, min_seq in zip(dataset, minutes_sequence):\n    with tf.GradientTape() as tape:\n        temp_loss_value = my_loss_function(inputs, min_seq)\n        batches +=1\n        loss_value = loss_value + temp_loss_value\n\n# If I do the following line, the graph will break because this are out of tape's scope.\n    loss_value = loss_value \/ batches\n\n# the following line takes huge time\ngrads = tape.gradient(loss_value, model.trainable_variables)\n\n<\/code><\/pre>\n\n<p>When I declare tf.GradientTape() inside the for loop, it is very fast but I outside It is slow.<\/p>\n\n<p>P.S. - This is for a custom loss and the architecture contains just one hidden layer of size 10.<\/p>\n\n<p>I want to know, the difference tf.GradientTape()'s position makes and how it should be used for per epoch weights updating in batched dataset.<\/p>\n","answers":[{"AnswerId":"58022690","CreationDate":"2019-09-20T06:12:59.917","ParentId":null,"OwnerUserId":"3994751","Title":null,"Body":"<p>The tape variable is used primarily to watch trainable tensor variables(record the previous and changing values of the variables) so that we can calculate the gradient for an epoch of training as per the loss function. It is an implementation of the python context manager construct used here to record the state of the variables. An excellent resource on python context managers is <a href=\"https:\/\/jeffknupp.com\/blog\/2016\/03\/07\/python-with-context-managers\" rel=\"nofollow noreferrer\">here<\/a>. So if inside the loop it will record the variables (weights) for that forward pass so that the we can calculate the gradient for all those variables in one shot (instead of stack based gradient passing as in a naive implementation without a library like tensorflow). If it is outside the loop it will record the states for all the epochs and as per the Tensorflow source code it also flushes if using TF2.0 unlike TF1.x where model developer had to take care of flushing. In your example you do not have any writer set but if any writer is set it will do that too. So for the above code it will keep recording (Graph.add_to_collection method is used internally) all the weights and as epochs increase you should see slowdown. The rate of slowdown will be proportional on the size of the network(trainable variables) and the current epoch number. <\/p>\n\n<p>So placing it inside the loop is correct. Also the gradients should be applied inside the for loop and not outside (at the same indent level as with) else you are only applying gradients at the end of your training loop (after last epoch). I see that your training may not be that good with the current placement of gradient retrieval(after which it is applied in your code though you omitted it in the snippet).<\/p>\n\n<p>One more good <a href=\"https:\/\/code-examples.net\/en\/docs\/tensorflow~python\/tf\/gradienttape\" rel=\"nofollow noreferrer\">resource<\/a> on gradienttape I just found.<\/p>\n"}]}
+{"QuestionId":57658393,"AnswerCount":0,"Tags":"<android><tensorflow><tensorflow-lite>","CreationDate":"2019-08-26T12:59:37.130","AcceptedAnswerId":null,"OwnerUserId":8449851.0,"Title":"How to fix error happened when using tensorflow in Android","Body":"<p>When I using tensorflow-lite in Android, after change the model officially supported, An error happened <\/p>\n\n<p>The input of my model has two inputs, 32*128*128*1 respectively, and I mimic the way to feed the inputs but failed <\/p>\n\n<p>Here is some code in my project<\/p>\n\n<pre class=\"lang-java prettyprint-override\"><code>  @Override\n  protected void processImage() {\n    try {\n      float[][][][] stft=readTxt(\"file:\/\/\/android_asset\/stft.txt\");\n      float[][][][] mfcc=readTxt(\"file:\/\/\/android_asset\/mfcc.txt\");\n      final List&lt;Classifier.Recognition&gt; results = detector.recognizeImage(stft, mfcc);\n\n    }\n    catch(Exception e)\n    {\n      e.printStackTrace();\n      System.out.print(\"can't get Bitmap\");\n    }\n\n  }\n\n  @Override\n  public List&lt;Recognition&gt; recognizeImage(final float[][][][] stft,final float[][][][] mfcc) {\n    \/\/ Log this method so that it can be analyzed with systrace.\n    Trace.beginSection(\"recognizeImage\");\n\n    Trace.beginSection(\"preprocessBitmap\");\n    \/\/ Preprocess the image data from 0-255 int to normalized float based\n    \/\/ on the provided parameters.\n    LOGGER.i(\"begin run model\");\n    Trace.endSection(); \/\/ preprocessBitmap\n\n    \/\/ Copy the input data into TensorFlow.\n    Trace.beginSection(\"feed\");\n    imgDatam.rewind();\n    imgDatas.rewind();\n    for(int i=0; i&lt;32; i++){\n      for(int j=0;j&lt;128;j++){\n        for(int k =0;k&lt;128;k++){\n          imgDatam.putFloat(mfcc[i][j][k][0]);\n          imgDatas.putFloat(stft[i][j][k][0]);\n        }\n\n      }\n\n    }\n\n    outputClasses = new float[32][NUM_DETECTIONS];\n\n    Object[] inputArray = {imgDatas,imgDatam};\n\n    Map&lt;Integer, Object&gt; outputMap = new HashMap&lt;&gt;();\n\n    outputMap.put(0, outputClasses);\n\n    Trace.endSection();\n\n    \/\/ Run the inference call.\n    Trace.beginSection(\"run\");\n    try{\n      tfLite.runForMultipleInputsOutputs(inputArray, outputMap);\n    }\n    catch(Exception e)\n    {\n      e.printStackTrace();\n    }\n    LOGGER.i(\"finished\");\n    Trace.endSection();\n\n    \/\/ Show the best detections.\n    \/\/ after scaling them back to the input size.\n    final ArrayList&lt;Recognition&gt; recognitions = new ArrayList&lt;&gt;(NUM_DETECTIONS);\n\n    Trace.endSection(); \/\/ \"recognizeImage\"\n    return recognitions;\n  }\n\n<\/code><\/pre>\n\n<p>And the error shows:<\/p>\n\n<blockquote>\n  <p>org.tensorflow.lite.examples.detection A\/libc: Fatal signal 6 (SIGABRT), code -6 (SI_TKILL) in tid 23610 (mples.detection), pid 23610 (mples.detection)\n  ? A\/DEBUG: pid: 23610, tid: 23610, name: mples.detection  >>> org.tensorflow.lite.examples.detection &lt;&lt;&lt;\n  ? A\/DEBUG:     #01 pc 00000000001220b4  \/data\/app\/org.tensorflow.lite.examples.detection-kKVx47dAvHYDUZmU-_Wkag==\/lib\/arm64\/libtensorflowlite_jni.so\n  ? A\/DEBUG:     #02 pc 00000000001216e4  \/data\/app\/org.tensorflow.lite.examples.detection-kKVx47dAvHYDUZmU-_Wkag==\/lib\/arm64\/libtensorflowlite_jni.so\n  ? A\/DEBUG:     #03 pc 00000000001624ac  \/data\/app\/org.tensorflow.lite.examples.detection-kKVx47dAvHYDUZmU-_Wkag==\/lib\/arm64\/libtensorflowlite_jni.so\n  ? A\/DEBUG:     #04 pc 0000000000165974  \/data\/app\/org.tensorflow.lite.examples.detection-kKVx47dAvHYDUZmU-_Wkag==\/lib\/arm64\/libtensorflowlite_jni.so\n  ? A\/DEBUG:     #05 pc 000000000000ed10  \/data\/app\/org.tensorflow.lite.examples.detection-kKVx47dAvHYDUZmU-_Wkag==\/lib\/arm64\/libtensorflowlite_jni.so (Java_org_tensorflow_lite_NativeInterpreterWrapper_run+32)\n  ? A\/DEBUG:     #13 pc 00000000001770ee  \/dev\/ashmem\/dalvik-classes.dex extracted in memory from \/data\/app\/org.tensorflow.lite.examples.detection-kKVx47dAvHYDUZmU-_Wkag==\/split_lib_dependencies_apk.apk (deleted) (org.tensorflow.lite.NativeInterpreterWrapper.run+158)\n  ? A\/DEBUG:     #19 pc 00000000001768f6  \/dev\/ashmem\/dalvik-classes.dex extracted in memory from \/data\/app\/org.tensorflow.lite.examples.detection-kKVx47dAvHYDUZmU-_Wkag==\/split_lib_dependencies_apk.apk (deleted) (org.tensorflow.lite.Interpreter.runForMultipleInputsOutputs+10)\n  ? A\/DEBUG:     #23 pc 0000000000007234  \/dev\/ashmem\/dalvik-jit-code-cache (deleted) (org.tensorflow.lite.examples.detection.tflite.TFLiteObjectDetectionAPIModel.recognizeImage+1428) \n  ? A\/DEBUG:     #28 pc 000000000000200e  \/dev\/ashmem\/dalvik-classes.dex extracted in memory from \/data\/app\/org.tensorflow.lite.examples.detection-kKVx47dAvHYDUZmU-_Wkag==\/split_lib_slice_4_apk.apk (deleted) (org.tensorflow.lite.examples.detection.tflite.TFLiteObjectDetectionAPIModel.recognizeImage+206)\n  ? A\/DEBUG:     #34 pc 000000000001d408  \/dev\/ashmem\/dalvik-classes.dex extracted in memory from \/data\/app\/org.tensorflow.lite.examples.detection-kKVx47dAvHYDUZmU-_Wkag==\/split_lib_slice_7_apk.apk (deleted) (org.tensorflow.lite.examples.detection.DetectorActivity.processImage+60)\n  ? A\/DEBUG:     #40 pc 0000000000015c6a  \/dev\/ashmem\/dalvik-classes.dex extracted in memory from \/data\/app\/org.tensorflow.lite.examples.detection-kKVx47dAvHYDUZmU-_Wkag==\/split_lib_slice_7_apk.apk (deleted) (org.tensorflow.lite.examples.detection.CameraActivity.onCreate+174)\n  ? E\/adbd: failed to connect to socket 'localabstract:org.tensorflow.lite.examples.detection': Connection refused<\/p>\n<\/blockquote>\n\n<p>Thanks for your help anyway, I know this may be hard because I even can't find some guide about tensorflow lite<\/p>\n","answers":[]}
+{"QuestionId":57658422,"AnswerCount":0,"Tags":"<tensorflow><google-colaboratory><tensor2tensor>","CreationDate":"2019-08-26T13:01:26.567","AcceptedAnswerId":null,"OwnerUserId":11064909.0,"Title":"tensor2tensor training won't start on Google Colab with GPU runtime","Body":"<p>I've been trying run the LibriSpeech problem using tensor2tensor using GPU runtime on Google Colab, but the training gets stuck before it can start. This is the code, which I got from <a href=\"https:\/\/cloud.google.com\/tpu\/docs\/tutorials\/automated-speech-recognition\" rel=\"nofollow noreferrer\">here<\/a> :<\/p>\n\n<p>Python's version : 3.6.8<\/p>\n\n<p>Tensorflow's version : 1.14.0<\/p>\n\n<p>tensor2tensor's version : 1.14.0<\/p>\n\n<p>CUDA's version : 10.1<\/p>\n\n<p>OS : Ubuntu 18.04<\/p>\n\n<pre><code>!t2t-trainer \\\n    --tmp_dir='\/content\/gdrive\/My Drive\/TCC\/T2T LibriSpeech\/tmp\/' \\\n    --problem='librispeech_clean_small' \\\n    --model='transformer' \\\n    --train_steps=2 \\\n    --eval_steps=1 \\\n    --hparams_set='transformer_librispeech' \\\n    --data_dir='\/content\/gdrive\/My Drive\/TCC\/T2T LibriSpeech\/data\/' \\\n    --output_dir='\/content\/gdrive\/My Drive\/TCC\/T2T LibriSpeech\/output\/' \n<\/code><\/pre>\n\n<p>And this is the output :<\/p>\n\n<pre><code>WARNING: Logging before flag parsing goes to stderr.\nW0826 18:25:13.545084 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/expert_utils.py:68: The name tf.variable_scope is deprecated. Please use tf.compat.v1.variable_scope instead.\n\nW0826 18:25:14.192255 139977848022912 lazy_loader.py:50] \nThe TensorFlow contrib module will not be included in TensorFlow 2.0.\nFor more information, please see:\n  * https:\/\/github.com\/tensorflow\/community\/blob\/master\/rfcs\/20180907-contrib-sunset.md\n  * https:\/\/github.com\/tensorflow\/addons\n  * https:\/\/github.com\/tensorflow\/io (for I\/O related ops)\nIf you depend on functionality not listed there, please file an issue.\n\nW0826 18:25:15.506344 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/adafactor.py:27: The name tf.train.Optimizer is deprecated. Please use tf.compat.v1.train.Optimizer instead.\n\nW0826 18:25:15.506757 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/multistep_optimizer.py:32: The name tf.train.AdamOptimizer is deprecated. Please use tf.compat.v1.train.AdamOptimizer instead.\n\nW0826 18:25:15.512568 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/mesh_tensorflow\/ops.py:4237: The name tf.train.CheckpointSaverListener is deprecated. Please use tf.estimator.CheckpointSaverListener instead.\n\nW0826 18:25:15.512784 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/mesh_tensorflow\/ops.py:4260: The name tf.train.SessionRunHook is deprecated. Please use tf.estimator.SessionRunHook instead.\n\nW0826 18:25:15.551204 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/trainer_lib.py:105: The name tf.OptimizerOptions is deprecated. Please use tf.compat.v1.OptimizerOptions instead.\n\nW0826 18:25:16.042501 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/bin\/t2t-trainer:32: The name tf.logging.set_verbosity is deprecated. Please use tf.compat.v1.logging.set_verbosity instead.\n\nW0826 18:25:16.042751 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/bin\/t2t-trainer:32: The name tf.logging.INFO is deprecated. Please use tf.compat.v1.logging.INFO instead.\n\nW0826 18:25:16.042839 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/bin\/t2t-trainer:33: The name tf.app.run is deprecated. Please use tf.compat.v1.app.run instead.\n\n:::MLPv0.5.0 transformer 1566843916.257424116 (\/usr\/local\/bin\/t2t-trainer:28) run_set_random_seed\nI0826 18:25:16.257472 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843916.257424116 (\/usr\/local\/bin\/t2t-trainer:28) run_set_random_seed\nW0826 18:25:16.257829 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/trainer_lib.py:789: The name tf.set_random_seed is deprecated. Please use tf.compat.v1.set_random_seed instead.\n\nW0826 18:25:16.259287 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/trainer_lib.py:142: The name tf.gfile.Exists is deprecated. Please use tf.io.gfile.exists instead.\n\nW0826 18:25:16.260453 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/trainer_lib.py:117: The name tf.GraphOptions is deprecated. Please use tf.compat.v1.GraphOptions instead.\n\nW0826 18:25:16.260685 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/trainer_lib.py:123: The name tf.GPUOptions is deprecated. Please use tf.compat.v1.GPUOptions instead.\n\nW0826 18:25:16.260875 139977848022912 deprecation.py:323] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/trainer_lib.py:278: __init__ (from tensorflow.contrib.learn.python.learn.estimators.run_config) is deprecated and will be removed in a future version.\nInstructions for updating:\nWhen switching to tf.estimator.Estimator, use tf.estimator.RunConfig instead.\nW0826 18:25:16.261037 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/trainer_lib.py:301: The name tf.logging.info is deprecated. Please use tf.compat.v1.logging.info instead.\n\nI0826 18:25:16.261096 139977848022912 trainer_lib.py:301] Configuring DataParallelism to replicate the model.\nI0826 18:25:16.261162 139977848022912 devices.py:76] schedule=continuous_train_and_eval\nI0826 18:25:16.261207 139977848022912 devices.py:77] worker_gpu=1\nI0826 18:25:16.261248 139977848022912 devices.py:78] sync=False\nW0826 18:25:16.261311 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/devices.py:139: The name tf.logging.warn is deprecated. Please use tf.compat.v1.logging.warn instead.\n\nW0826 18:25:16.261358 139977848022912 devices.py:141] Schedule=continuous_train_and_eval. Assuming that training is running on a single machine.\nI0826 18:25:16.261914 139977848022912 devices.py:170] datashard_devices: ['gpu:0']\nI0826 18:25:16.261970 139977848022912 devices.py:171] caching_devices: None\nI0826 18:25:16.262233 139977848022912 devices.py:172] ps_devices: ['gpu:0']\nI0826 18:25:16.262940 139977848022912 estimator.py:209] Using config: {'_save_checkpoints_secs': None, '_num_ps_replicas': 0, '_keep_checkpoint_max': 20, '_task_type': None, '_train_distribute': None, '_is_chief': True, '_cluster_spec': &lt;tensorflow.python.training.server_lib.ClusterSpec object at 0x7f4ec60062d0&gt;, '_tf_config': gpu_options {\n  per_process_gpu_memory_fraction: 1.0\n}\n, '_protocol': None, '_save_checkpoints_steps': 1000, '_keep_checkpoint_every_n_hours': 10000, '_session_config': gpu_options {\n  per_process_gpu_memory_fraction: 0.95\n}\nallow_soft_placement: true\ngraph_options {\n  optimizer_options {\n    global_jit_level: OFF\n  }\n}\n, '_model_dir': '\/content\/gdrive\/My Drive\/TCC\/T2T LibriSpeech\/output\/', 'use_tpu': False, '_tf_random_seed': None, '_master': '', '_device_fn': None, '_num_worker_replicas': 0, '_task_id': 0, '_log_step_count_steps': 100, '_experimental_max_worker_delay_secs': None, '_evaluation_master': '', '_eval_distribute': None, 'data_parallelism': &lt;tensor2tensor.utils.expert_utils.Parallelism object at 0x7f4ec6006310&gt;, '_environment': 'local', '_save_summary_steps': 100, 't2t_device_info': {'num_async_replicas': 1}}\nW0826 18:25:16.263118 139977848022912 model_fn.py:630] Estimator's model_fn (&lt;function wrapping_model_fn at 0x7f4ec6005500&gt;) includes params argument, but params are not passed to Estimator.\nW0826 18:25:16.263513 139977848022912 trainer_lib.py:722] ValidationMonitor only works with --schedule=train_and_evaluate\nW0826 18:25:16.264281 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/bin\/t2t_trainer.py:310: The name tf.gfile.MakeDirs is deprecated. Please use tf.io.gfile.makedirs instead.\n\nW0826 18:25:16.266737 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/bin\/t2t_trainer.py:326: The name tf.gfile.Open is deprecated. Please use tf.io.gfile.GFile instead.\n\nI0826 18:25:16.276473 139977848022912 estimator_training.py:186] Not using Distribute Coordinator.\nI0826 18:25:16.276839 139977848022912 training.py:612] Running training and evaluation locally (non-distributed).\nI0826 18:25:16.277188 139977848022912 training.py:700] Start train and evaluate loop. The evaluate will happen after every checkpoint. Checkpoint frequency is determined based on RunConfig arguments: save_checkpoints_steps 1000 or save_checkpoints_secs None.\nW0826 18:25:16.282742 139977848022912 deprecation.py:323] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/training\/training_util.py:236: initialized_value (from tensorflow.python.ops.variables) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse Variable.read_value. Variables in 2.X are initialized automatically both in eager and graph (inside tf.defun) contexts.\n:::MLPv0.5.0 transformer 1566843916.300036907 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/data_generators\/problem.py:759) input_max_length: 1240000\nI0826 18:25:16.300060 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843916.300036907 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/data_generators\/problem.py:759) input_max_length: 1240000\nI0826 18:25:16.300235 139977848022912 problem.py:614] Reading data files from \/content\/gdrive\/My Drive\/TCC\/T2T LibriSpeech\/data\/librispeech_clean_small-train*\nI0826 18:25:16.321804 139977848022912 problem.py:644] partition: 0 num_data_files: 100\n:::MLPv0.5.0 transformer 1566843916.322568893 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/data_generators\/problem.py:872) input_order\nI0826 18:25:16.322586 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843916.322568893 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/data_generators\/problem.py:872) input_order\nW0826 18:25:16.324836 139977848022912 deprecation.py:323] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/data_generators\/problem.py:654: 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(...)`.\nW0826 18:25:16.324963 139977848022912 deprecation.py:323] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/contrib\/data\/python\/ops\/interleave_ops.py:77: parallel_interleave (from tensorflow.python.data.experimental.ops.interleave_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.data.Dataset.interleave(map_func, cycle_length, block_length, num_parallel_calls=tf.data.experimental.AUTOTUNE)` instead. If sloppy execution is desired, use `tf.data.Options.experimental_determinstic`.\nW0826 18:25:16.361968 139977848022912 deprecation.py:323] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/layers\/common_audio.py:92: to_int32 (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.cast` instead.\nW0826 18:25:16.458050 139977848022912 deprecation.py:323] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/layers\/common_audio.py:115: to_float (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.cast` instead.\nW0826 18:25:16.617947 139977848022912 deprecation.py:323] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/data_generators\/problem.py:172: tf_record_iterator (from tensorflow.python.lib.io.tf_record) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse eager execution and: \n`tf.data.TFRecordDataset(path)`\nW0826 18:25:17.132765 139977848022912 deprecation.py:323] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/data_generators\/problem.py:893: output_shapes (from tensorflow.python.data.ops.dataset_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.compat.v1.data.get_output_shapes(dataset)`.\nW0826 18:25:17.133136 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/data_generators\/problem.py:896: The name tf.logging.warning is deprecated. Please use tf.compat.v1.logging.warning instead.\n\nW0826 18:25:17.133249 139977848022912 problem.py:897] Shapes are not fully defined. Assuming batch_size means tokens.\nW0826 18:25:17.153662 139977848022912 deprecation.py:323] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/data_generators\/problem.py:947: bucket_by_sequence_length (from tensorflow.contrib.data.python.ops.grouping) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.data.experimental.bucket_by_sequence_length(...)`.\nW0826 18:25:17.174197 139977848022912 deprecation.py:323] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/data\/experimental\/ops\/grouping.py:193: where (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse tf.where in 2.0, which has the same broadcast rule as np.where\nW0826 18:25:17.212933 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/data_generators\/problem.py:1209: The name tf.summary.scalar is deprecated. Please use tf.compat.v1.summary.scalar instead.\n\nI0826 18:25:17.264877 139977848022912 estimator.py:1145] Calling model_fn.\nI0826 18:25:17.276007 139977848022912 t2t_model.py:1905] Unsetting shared_embedding_and_softmax_weights.\n:::MLPv0.5.0 transformer 1566843917.277432919 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:59) model_hp_embedding_shared_weights: {\"vocab_size\": 256, \"hidden_size\": 384}\nI0826 18:25:17.277457 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843917.277432919 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:59) model_hp_embedding_shared_weights: {\"vocab_size\": 256, \"hidden_size\": 384}\nI0826 18:25:17.277616 139977848022912 t2t_model.py:1905] Setting T2TModel mode to 'train'\nW0826 18:25:17.333712 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/t2t_model.py:167: The name tf.summary.text is deprecated. Please use tf.compat.v1.summary.text instead.\n\n:::MLPv0.5.0 transformer 1566843919.260778904 (\/tmp\/tmpurXynh.py:100) model_hp_initializer_gain: 1.0\nW0826 18:25:21.170526 139977848022912 ag_logging.py:145] Entity &lt;bound method PythonHandler.emit of &lt;absl.logging.PythonHandler object at 0x7f4eea764510&gt;&gt; could not be transformed and will be executed as-is. Please report this to the AutgoGraph team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output. Cause: The global keyword is not yet supported.\nI0826 18:25:20.087136 139977848022912 api.py:452] :::MLPv0.5.0 transformer 1566843919.260778904 (\/tmp\/tmpurXynh.py:100) model_hp_initializer_gain: 1.0\nI0826 18:25:21.171293 139977848022912 api.py:255] Using variable initializer: uniform_unit_scaling\nI0826 18:25:21.585083 139977848022912 t2t_model.py:1905] Transforming feature 'inputs' with speech_recognition_modality.bottom\nW0826 18:25:21.586750 139977848022912 deprecation.py:323] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/layers\/modalities.py:585: conv2d (from tensorflow.python.layers.convolutional) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.keras.layers.Conv2D` instead.\nI0826 18:25:21.853492 139977848022912 t2t_model.py:1905] Transforming feature 'targets' with symbol_modality_256_384.targets_bottom\nI0826 18:25:21.967683 139977848022912 t2t_model.py:1905] Building model body\n:::MLPv0.5.0 transformer 1566843922.029903889 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:186) model_hp_layer_postprocess_dropout: 0.2\nI0826 18:25:22.029930 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843922.029903889 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:186) model_hp_layer_postprocess_dropout: 0.2\nW0826 18:25:22.030196 139977848022912 deprecation.py:506] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:92: calling dropout (from tensorflow.python.ops.nn_ops) with keep_prob is deprecated and will be removed in a future version.\nInstructions for updating:\nPlease use `rate` instead of `keep_prob`. Rate should be set to `rate = 1 - keep_prob`.\n:::MLPv0.5.0 transformer 1566843922.038706064 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:101) model_hp_hidden_layers: 6\nI0826 18:25:22.038723 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843922.038706064 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:101) model_hp_hidden_layers: 6\n:::MLPv0.5.0 transformer 1566843922.039638042 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:101) model_hp_attention_dropout: 0.1\nI0826 18:25:22.039654 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843922.039638042 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:101) model_hp_attention_dropout: 0.1\n:::MLPv0.5.0 transformer 1566843922.040515900 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:101) model_hp_attention_dense: {\"num_heads\": 2, \"use_bias\": \"false\", \"hidden_size\": 384}\nI0826 18:25:22.040528 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843922.040515900 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:101) model_hp_attention_dense: {\"num_heads\": 2, \"use_bias\": \"false\", \"hidden_size\": 384}\nW0826 18:25:22.067670 139977848022912 deprecation.py:323] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/layers\/common_layers.py:2926: dense (from tensorflow.python.layers.core) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse keras.layers.dense instead.\nW0826 18:25:22.656837 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/layers\/common_attention.py:1171: The name tf.summary.image is deprecated. Please use tf.compat.v1.summary.image instead.\n\n:::MLPv0.5.0 transformer 1566843925.825310946 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:101) model_hp_norm: {\"hidden_size\": 384}\nI0826 18:25:25.825342 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843925.825310946 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:101) model_hp_norm: {\"hidden_size\": 384}\n:::MLPv0.5.0 transformer 1566843925.911139011 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:202) model_hp_layer_postprocess_dropout: 0.2\nI0826 18:25:25.911170 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843925.911139011 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:202) model_hp_layer_postprocess_dropout: 0.2\n:::MLPv0.5.0 transformer 1566843925.919984102 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:153) model_hp_hidden_layers: 4\nI0826 18:25:25.920005 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843925.919984102 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:153) model_hp_hidden_layers: 4\n:::MLPv0.5.0 transformer 1566843925.920943975 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:153) model_hp_attention_dropout: 0.1\nI0826 18:25:25.920955 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843925.920943975 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:153) model_hp_attention_dropout: 0.1\n:::MLPv0.5.0 transformer 1566843925.921816111 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:153) model_hp_attention_dense: {\"num_heads\": 2, \"use_bias\": \"false\", \"hidden_size\": 384}\nI0826 18:25:25.921828 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843925.921816111 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:153) model_hp_attention_dense: {\"num_heads\": 2, \"use_bias\": \"false\", \"hidden_size\": 384}\n:::MLPv0.5.0 transformer 1566843929.180166006 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:153) model_hp_norm: {\"hidden_size\": 384}\nI0826 18:25:29.180202 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843929.180166006 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/models\/transformer.py:153) model_hp_norm: {\"hidden_size\": 384}\nI0826 18:25:29.203882 139977848022912 t2t_model.py:1905] Transforming body output with symbol_modality_256_384.top\n:::MLPv0.5.0 transformer 1566843929.399625063 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/t2t_model.py:582) opt_learning_rate: \"DEFERRED: acbbcfbd-a251-4ef2-8405-8a4a42bb03d7\"\nI0826 18:25:29.399658 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843929.399625063 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/t2t_model.py:582) opt_learning_rate: \"DEFERRED: acbbcfbd-a251-4ef2-8405-8a4a42bb03d7\"\n:::MLPv0.5.0 transformer 1566843929.400494099 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/t2t_model.py:582) opt_learning_rate_warmup_steps: 8000\nI0826 18:25:29.400509 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843929.400494099 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/t2t_model.py:582) opt_learning_rate_warmup_steps: 8000\nW0826 18:25:29.400695 139977848022912 deprecation_wrapper.py:119] From \/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/learning_rate.py:100: The name tf.train.get_or_create_global_step is deprecated. Please use tf.compat.v1.train.get_or_create_global_step instead.\n\nI0826 18:25:29.401774 139977848022912 learning_rate.py:29] Base learning rate: 2.000000\nI0826 18:25:29.414273 139977848022912 optimize.py:251] Trainable Variables Total size: 70441856\nI0826 18:25:29.414625 139977848022912 optimize.py:251] Non-trainable variables Total size: 5\nI0826 18:25:29.415019 139977848022912 optimize.py:89] Using optimizer Adam\n:::MLPv0.5.0 transformer 1566843929.415780067 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/optimize.py:53) opt_name: \"Adam\"\nI0826 18:25:29.415796 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843929.415780067 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/optimize.py:53) opt_name: \"Adam\"\n:::MLPv0.5.0 transformer 1566843929.416517019 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/optimize.py:53) opt_hp_Adam_beta1: 0.9\nI0826 18:25:29.416527 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843929.416517019 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/optimize.py:53) opt_hp_Adam_beta1: 0.9\n:::MLPv0.5.0 transformer 1566843929.417274952 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/optimize.py:53) opt_hp_Adam_beta2: 0.997\nI0826 18:25:29.417284 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843929.417274952 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/optimize.py:53) opt_hp_Adam_beta2: 0.997\n:::MLPv0.5.0 transformer 1566843929.418004036 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/optimize.py:53) opt_hp_Adam_epsilon: 1e-09\nI0826 18:25:29.418014 139977848022912 mlperf_log.py:156] :::MLPv0.5.0 transformer 1566843929.418004036 (\/usr\/local\/lib\/python2.7\/dist-packages\/tensor2tensor\/utils\/optimize.py:53) opt_hp_Adam_epsilon: 1e-09\nI0826 18:25:40.607661 139977848022912 estimator.py:1147] Done calling model_fn.\nI0826 18:25:40.609323 139977848022912 basic_session_run_hooks.py:541] Create CheckpointSaverHook.\nI0826 18:25:45.471376 139977848022912 monitored_session.py:240] Graph was finalized.\n2019-08-26 18:25:45.471909: I tensorflow\/core\/platform\/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX512F\n2019-08-26 18:25:45.477428: I tensorflow\/core\/platform\/profile_utils\/cpu_utils.cc:94] CPU Frequency: 2000155000 Hz\n2019-08-26 18:25:45.477710: I tensorflow\/compiler\/xla\/service\/service.cc:168] XLA service 0x562891013c00 executing computations on platform Host. Devices:\n2019-08-26 18:25:45.477743: I tensorflow\/compiler\/xla\/service\/service.cc:175]   StreamExecutor device (0): &lt;undefined&gt;, &lt;undefined&gt;\n2019-08-26 18:25:45.480285: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcuda.so.1\n2019-08-26 18:25:45.634041: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-08-26 18:25:45.634841: I tensorflow\/compiler\/xla\/service\/service.cc:168] XLA service 0x562891013f80 executing computations on platform CUDA. Devices:\n2019-08-26 18:25:45.634873: I tensorflow\/compiler\/xla\/service\/service.cc:175]   StreamExecutor device (0): Tesla T4, Compute Capability 7.5\n2019-08-26 18:25:45.635107: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-08-26 18:25:45.635709: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1640] Found device 0 with properties: \nname: Tesla T4 major: 7 minor: 5 memoryClockRate(GHz): 1.59\npciBusID: 0000:00:04.0\n2019-08-26 18:25:45.636103: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudart.so.10.0\n2019-08-26 18:25:45.637595: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcublas.so.10.0\n2019-08-26 18:25:45.639006: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcufft.so.10.0\n2019-08-26 18:25:45.639369: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcurand.so.10.0\n2019-08-26 18:25:45.640924: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcusolver.so.10.0\n2019-08-26 18:25:45.642059: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcusparse.so.10.0\n2019-08-26 18:25:45.645505: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudnn.so.7\n2019-08-26 18:25:45.645641: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-08-26 18:25:45.646182: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-08-26 18:25:45.646660: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1763] Adding visible gpu devices: 0\n2019-08-26 18:25:45.646722: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudart.so.10.0\n2019-08-26 18:25:45.647904: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1181] Device interconnect StreamExecutor with strength 1 edge matrix:\n2019-08-26 18:25:45.647930: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1187]      0 \n2019-08-26 18:25:45.647940: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1200] 0:   N \n2019-08-26 18:25:45.648047: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-08-26 18:25:45.648618: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-08-26 18:25:45.649113: W tensorflow\/core\/common_runtime\/gpu\/gpu_bfc_allocator.cc:40] Overriding allow_growth setting because the TF_FORCE_GPU_ALLOW_GROWTH environment variable is set. Original config value was 0.\n2019-08-26 18:25:45.649167: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1326] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 14325 MB memory) -&gt; physical GPU (device: 0, name: Tesla T4, pci bus id: 0000:00:04.0, compute capability: 7.5)\n2019-08-26 18:25:49.626228: W tensorflow\/compiler\/jit\/mark_for_compilation_pass.cc:1412] (One-time warning): Not using XLA:CPU for cluster because envvar TF_XLA_FLAGS=--tf_xla_cpu_global_jit was not set.  If you want XLA:CPU, either set that envvar, or use experimental_jit_scope to enable XLA:CPU.  To confirm that XLA is active, pass --vmodule=xla_compilation_cache=1 (as a proper command-line flag, not via TF_XLA_FLAGS) or set the envvar XLA_FLAGS=--xla_hlo_profile.\nI0826 18:25:53.582676 139977848022912 session_manager.py:500] Running local_init_op.\nI0826 18:25:53.897416 139977848022912 session_manager.py:502] Done running local_init_op.\nI0826 18:26:06.040513 139977848022912 basic_session_run_hooks.py:606] Saving checkpoints for 0 into \/content\/gdrive\/My Drive\/TCC\/T2T LibriSpeech\/output\/model.ckpt.\n2019-08-26 18:26:28.882768: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcublas.so.10.0\n<\/code><\/pre>\n\n<p>It gets stuck on this last line. Both Tensorflow and tensor2tensor's version is 1.14.0. Any ideas on how to solve it? I tried posting on tensor2tensor's repository on Github a week ago, but I haven't been answered yet.<\/p>\n","answers":[]}
+{"QuestionId":57658509,"AnswerCount":1,"Tags":"<c++><python-3.x><tensorflow><tensorflow-lite>","CreationDate":"2019-08-26T13:06:07.087","AcceptedAnswerId":null,"OwnerUserId":8334261.0,"Title":"How to invoke the Flex delegate for tflite interpreters?","Body":"<p>I have a TensorFlow model which I want to convert into a tflite model, which is going to be deployed on an ARM64 platform.<\/p>\n\n<p>It happens to be that two operations of my model (RandomStandardNormal, Softplus) seem to require custom implementations. Due to execution time being not that important, I decided to go with a hybrid model that uses the extended runtime. I converted it via:<\/p>\n\n<pre><code>graph_def_file = '.\/model\/frozen_model.pb'\ninputs = ['eval_inputs']\noutputs = ['model\/y']\n\nconverter = tf.lite.TFLiteConverter.from_frozen_graph(graph_def_file, inputs, outputs)\nconverter.target_ops = [tf.lite.OpsSet.TFLITE_BUILTINS, tf.lite.OpsSet.SELECT_TF_OPS]\n\ntflite_file_name = 'vae_' + str(tf.__version__) + '.tflite'\n\ntflite_model = converter.convert()\nopen(tflite_file_name, 'wb').write(tflite_model)\n<\/code><\/pre>\n\n<p>This worked and I ended up with a seemingly valid tflite model file. Whenever I try to load this model with an interpreter, I get an error (it does not matter if I use the Python or C++ API):<\/p>\n\n<pre><code>ERROR: Regular TensorFlow ops are not supported by this interpreter. Make sure you invoke the Flex delegate before inference.\nERROR: Node number 4 (FlexSoftplus) failed to prepare.\n<\/code><\/pre>\n\n<p>I have a hard time to find documentation on the tf website on how to invoke the Flex delegate for both APIs. I have stumbled across a header file (\"tensorflow\/lite\/delegates\/flex\/delegate_data.h\") which seems to be related to this issue, but including it in my C++ project yields another error:<\/p>\n\n<pre><code>In file included from \/tensorflow\/tensorflow\/core\/common_runtime\/eager\/context.h:28:0,\n                 from \/tensorflow\/tensorflow\/lite\/delegates\/flex\/delegate_data.h:18,\n                 from \/tensorflow\/tensorflow\/lite\/delegates\/flex\/delegate.h:19,\n                 from demo.cpp:7:\n\/tensorflow\/tensorflow\/core\/lib\/core\/status.h:23:10: fatal error: tensorflow\/core\/lib\/core\/error_codes.pb.h: No such file or directory\n #include \"tensorflow\/core\/lib\/core\/error_codes.pb.h\"\n          ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\ncompilation terminated.\n<\/code><\/pre>\n\n<p>By any chance, has anybody encountered and resolved this before? If you have an example snippet, please share the link!<\/p>\n","answers":[{"AnswerId":"58093107","CreationDate":"2019-09-25T07:19:52.923","ParentId":null,"OwnerUserId":"12111239","Title":null,"Body":"<p>According to <a href=\"https:\/\/www.tensorflow.org\/lite\/guide\/ops_select#android_aar\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/lite\/guide\/ops_select#android_aar<\/a> on 2019\/9\/25<\/p>\n\n<p>Python support of 'select operators' is actively under development.<\/p>\n\n<p>You can test the model in Android by using FlexDelegate.\nI ran my model successfully in the same way.<\/p>\n\n<p>e.g. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/lite\/java\/src\/test\/java\/org\/tensorflow\/lite\/InterpreterFlexTest.java\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/lite\/java\/src\/test\/java\/org\/tensorflow\/lite\/InterpreterFlexTest.java<\/a><\/p>\n"}]}
+{"QuestionId":57658571,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-08-26T13:09:51.823","AcceptedAnswerId":57658609.0,"OwnerUserId":11486611.0,"Title":"Predicting a single image with Keras' ImageDataGenerator","Body":"<p>I'm very new to deep learning so please forgive me for this probably simple question.<\/p>\n\n<p>I trained a network to classify between <code>positive<\/code> and <code>negative<\/code>. To simplify the image generation and fitting process I used a <code>ImageDataGenerator<\/code> and the <code>fit_generator<\/code> function, as shown below:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nfrom tensorflow.keras.preprocessing.image import ImageDataGenerator\n\n# Simplified model\nmodel = tf.keras.models.Sequential([\n    tf.keras.layers.Conv2D(16, (3,3), activation='relu', input_shape=(12, 12, 3)),\n    tf.keras.layers.MaxPooling2D(2, 2),\n    tf.keras.layers.Flatten(),\n    tf.keras.layers.Dense(128, activation='relu'),\n    tf.keras.layers.Dense(1, activation='sigmoid')\n])\n\n# Image import, for 'validation_generator' equally\ntrain_datagen = ImageDataGenerator(rescale=1.\/255)\ntrain_generator = train_datagen.flow_from_directory(\n    '.\/training\/',\n    target_size=(12, 12),\n    batch_size=128,\n    class_mode='binary')\n\n# Compiling\nmodel.compile(loss='binary_crossentropy',\n             optimizer='adam',\n             metrics=['acc'])\n\n# Fitting, for Tensorboard 'history = model.fit_gen...'\nmodel.fit_generator(train_generator,\n                    steps_per_epoch=8,\n                    epochs=50,\n                    verbose=1,\n                    validation_data = validation_generator,\n                    validation_steps=8,\n                    callbacks=[tb]) # Standard Tensorboard\n<\/code><\/pre>\n\n<p>I want to use my model to predict a single image (imported as <code>numpy array<\/code>) as shown below: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>image = 'single imported image with shape (12, 12, 3)'\nmodel.predict(image)\n<\/code><\/pre>\n\n<p>However, the only thing I get is an error message stating the <code>Matrix size-incompatible<\/code>. I have tried <code>model.predict_generator()<\/code> on my <code>validation_generator<\/code> which works, but that isn't a single image.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"57658609","CreationDate":"2019-08-26T13:12:35.490","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>If you want to do a prediction on a single image, do the following:<\/p>\n\n<pre><code>image = np.random.rand(12, 12, 3) # single imported image with shape (12, 12, 3)\nimage = np.expand_dims(image, axis=0) # image shape is (1, 12, 12, 3)\nmodel.predict(image)\n<\/code><\/pre>\n\n<p>In other words, your model still expects input shape of <code>(None, 12, 12, 3)<\/code>. So, before doing the prediction, expand the dimensions of the image to be a batch of a single image.<\/p>\n"}]}
+{"QuestionId":57658717,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-26T13:18:39.587","AcceptedAnswerId":null,"OwnerUserId":11909156.0,"Title":"Tensorflow - alternate training: minimize function per column","Body":"<p>I am training a denoising autoencoder with tensorflow 1.14.0 in python 3.6.8. each training step includes a fit of the model and should also optimize an other variable theta (of a binomial distribution used in loss) based on the previous prediction, which has an impact on the next prediction.<\/p>\n\n<p>here is my question:\nhow do I minimize this theta after each training step for each column individually? so alternate training based on the current prediction as input for minimization problem per column?<\/p>\n\n<p>what I already tried:<\/p>\n\n<ul>\n<li><p>used tf.map_fn with a minimization function which induced a new session and optimizer for each column to minimize -> works fine but is way too slow due to multiple session creation (see tf_minimize from <a href=\"https:\/\/stackoverflow.com\/questions\/41918795\/minimize-a-function-of-one-variable-in-tensorflow\">Minimize a function of one variable in Tensorflow<\/a>)<\/p><\/li>\n<li><p>created hashtable of optimizer for each column to minimize theta -> individual optimizer initialization for 10 000 columns is not feasible (compare: <a href=\"https:\/\/stackoverflow.com\/questions\/47156113\/how-to-alternate-train-ops-in-tensorflow\">How to alternate train op&#39;s in tensorflow?<\/a>)<\/p><\/li>\n<\/ul>\n\n<pre><code>### no functioning python code, just to give you an idea\n### input in tensorflow model: ae_input, theta\n\n\n### training steps\nfor epoch in range(200):\n    model.fit( x = ae_input+theta,\n        y = ae_input,\n        epochs = epoch+1)\n    current_pred = model.predict(ae_input+theta)\n\n    theta = update_theta(ae_input, current_pred)   # PROBLEMATIC STEP\n\n\n### current way because map_fn is not working and\n### would also be too slow due to multiple session creation\ndef update_theta(y_true, y_pred):\n    theta_list=[]\n    for col in range(y_true.get_shape()[1]):\n        t = theta_optimize(y_true[:,x], y_pred[:,x])\n        theta_list.append(t)\n    return theta_list\n\n### minimize theta for single column\ndef theta_optimize(y_t, y_p):\n    var_theta = tf.Variable(5.)\n    func_to_minimize = loss_per_col(y_t, y_p, var_theta) \n    optimial_theta = tf_minimize(func_to_optimize, output=var_theta) \n    ### tf_minimize from link mentioned above\n    return optimal_theta\n<\/code><\/pre>\n","answers":[{"AnswerId":"57692157","CreationDate":"2019-08-28T12:12:49.493","ParentId":null,"OwnerUserId":"11909156","Title":null,"Body":"<p>for these, who are still interested, I solved it by switching to scipy between each step. Surely not the ideal way, but was faster and worked fine for me<\/p>\n\n<pre><code>### CODE FOR TENSORFLOW 2.0.0 beta\n\n### function for scipy to minimize theta between 0-1000\ndef theta_loss_per_col(y_t, y_p, theta):\n    # ... \n    return loss\n\n\n### function for scipy to optimize\ndef func_minimize_theta(y_t, y_p):\n    return lambda x: theta_loss_per_col(y_t, y_p, x).numpy()\n\n\n### return updated list of thetas\ndef get_updated_theta(y_true, y_pred):\n    y_true_cols = tf.range(y_true.shape[1])\n    theta_list =  tf.map_fn(lambda i:\nscipy.optimize.fminbound(func_minimize_theta(y_true[:,i], y_pred[:,i]),\n1e-5, 1000) , y_true_cols, dtype=tf.float32) # maybe add:parallel_iterations=10\n    return theta_list\n<\/code><\/pre>\n"}]}
+{"QuestionId":57659159,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-08-26T13:46:38.497","AcceptedAnswerId":57659667.0,"OwnerUserId":6141275.0,"Title":"What to use in keras >2.0 in place of merge with mode='mul'?","Body":"<p>I'm trying to modify some code that was made for Keras&lt;2.0 and I stumbled upon this line of code:<\/p>\n\n<pre><code>output_attention_mul = merge([inputs, a_probs], name='attention_mul', mode='mul')\n<\/code><\/pre>\n\n<p>How can I translate this line of code in Keras>2.0?<\/p>\n\n<p>I've tried using the <code>Multiply<\/code> layer this way<\/p>\n\n<pre><code>output_attention_mul = Multiply()([inputs, a_probs])\n<\/code><\/pre>\n\n<p>but I don't really know if that's the right thing to do since in the old code it's used the <code>merge<\/code> operation and not the <code>Merge<\/code> layer.<\/p>\n\n<p>Can anybody confirm me if I did the right thing or if I used a wrong approach? Should I use <code>multiply<\/code> instead of <code>Multiply<\/code>?<\/p>\n","answers":[{"AnswerId":"57659667","CreationDate":"2019-08-26T14:14:36.840","ParentId":null,"OwnerUserId":"11978974","Title":null,"Body":"<p>I think what you're looking for is the <code>keras.layers.Multiply()<\/code><\/p>\n\n<p>Which should then give you <code>output_attention_mul = keras.layers.Multiply()([inputs, a_probs])<\/code><\/p>\n\n<p>You should be able to find in the documentation here: <a href=\"https:\/\/keras.io\/layers\/merge\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/merge\/<\/a><\/p>\n\n<p>Hope this helps you!<\/p>\n"}]}
+{"QuestionId":57660053,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-08-26T14:37:46.620","AcceptedAnswerId":null,"OwnerUserId":10855529.0,"Title":"What is the difference between these two neural network structures?","Body":"<p>first using nn.Parameter<\/p>\n\n<pre><code>class ModelOne(nn.Module):\n  def __init__(self):\n    super().__init__()\n    self.weights = nn.Parameter(torch.randn(300, 10))\n    self.bias = nn.Parameter(torch.zeros(10))\n  def forward(self, x):\n    return x @ self.weights + self.bias\n<\/code><\/pre>\n\n<p>when I do<\/p>\n\n<pre><code>mo = ModelOne()\n[len(param) for param in mo.parameters()]\n<\/code><\/pre>\n\n<p>it gives\n[300, 10]<\/p>\n\n<p>second using nn.Linear<\/p>\n\n<pre><code>class ModelTwo(nn.Module):\n  def __init__(self):\n    super().__init__()\n    self.linear = nn.Linear(300, 10)\n\n  def forward(self, x):\n    return self.linear(x)\n<\/code><\/pre>\n\n<p>same thing here gives\n[10, 10]<\/p>\n","answers":[{"AnswerId":"57660683","CreationDate":"2019-08-26T15:17:40.950","ParentId":null,"OwnerUserId":"6210807","Title":null,"Body":"<p>The difference lies in how <a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/nn\/modules\/linear.html\" rel=\"nofollow noreferrer\"><code>nn.Linear<\/code><\/a> initializes weights and bias:<\/p>\n\n<pre><code>class Linear(Module):\n\n    def __init__(self, in_features, out_features, bias=True):\n        super(Linear, self).__init__()\n        self.in_features = in_features\n        self.out_features = out_features\n        self.weight = Parameter(torch.Tensor(out_features, in_features))\n        if bias:\n            self.bias = Parameter(torch.Tensor(out_features))\n        ...\n<\/code><\/pre>\n\n<p>So, when you write <code>nn.Linear(300, 10)<\/code> the weight is (10, 300) and bias is (10). But, in the ModelOne, weights has dimension (300, 10).<\/p>\n\n<p>You can confirm it using<\/p>\n\n<pre><code>for name, param in mo.named_parameters():\n    print(name, param.shape)\n<\/code><\/pre>\n\n<p><em>The output in ModelOne:<\/em><\/p>\n\n<p>weights torch.Size([300, 10])<br>\nbias torch.Size([10])<\/p>\n\n<p><em>In ModelTwo:<\/em><\/p>\n\n<p>linear.weight torch.Size([10, 300])<br>\nlinear.bias torch.Size([10])<\/p>\n\n<hr>\n\n<p>Now, the reason you're getting [300, 10] in first case and [10, 10] in second case is because if you print the length of a 2d Tensor, then it'll only give its first dimension i.e.<\/p>\n\n<pre><code>a = torch.Tensor(10, 300)\nb = torch.Tensor(10)\nprint(len(a), len(b))\n<\/code><\/pre>\n\n<p>(10, 10)<\/p>\n"}]}
+{"QuestionId":57660161,"AnswerCount":1,"Tags":"<tensorflow><conv-neural-network>","CreationDate":"2019-08-26T14:44:37.197","AcceptedAnswerId":57679399.0,"OwnerUserId":10687511.0,"Title":"Network bug - Inception v1 isn't training","Body":"<p>I am trying to use the Inception model (GoogLeNet) from this link <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/slim\/nets\/inception_v1.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/slim\/nets\/inception_v1.py<\/a> which is implemented by Google using the Tensorflow Slim API, to classify images from the Cifar10 dataset (dataset link <a href=\"https:\/\/www.cs.toronto.edu\/~kriz\/cifar.html\" rel=\"nofollow noreferrer\">https:\/\/www.cs.toronto.edu\/~kriz\/cifar.html<\/a>). The problem is that the network cost stays almost constant and I can't find the bug. I am very new to tensorflow and slim, so I whould really appreciate any help.<\/p>\n\n<p>I am using these packages:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\nfrom tensorflow.python.framework import ops\nimport matplotlib.pyplot as plt\nimport os\nimport pickle\nimport cv2\nfrom sklearn import model_selection as ms\nfrom nets import inception_v1,inception_utils\nimport math\n%matplotlib inline\n<\/code><\/pre>\n\n<p>And I made theese two functions:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def one_hot_matrix(labels, C):\n\n    C = tf.constant(C,name='C')\n    one_hot_matrix = tf.one_hot(labels,C,axis=0)\n    sess = tf.Session()\n    one_hot = sess.run(one_hot_matrix)\n    sess.close()\n\n    return one_hot\n\n\ndef make_mini_batches(X, Y, mini_batch_size):\n\n    m = X.shape[0]                  \n    mini_batches = []\n\n    # number of mini batches of size mini_batch_size in the dataset\n    num_complete_minibatches = math.floor(m\/mini_batch_size) \n\n    for k in range(0, num_complete_minibatches):\n        mini_batch_X = X[k*mini_batch_size : (k+1)*mini_batch_size,...]\n        mini_batch_Y = Y[k*mini_batch_size : (k+1)*mini_batch_size,:]\n        mini_batch = (mini_batch_X, mini_batch_Y)\n        mini_batches.append(mini_batch)\n\n    # Handling the end case (last mini-batch &lt; mini_batch_size)\n    if m % mini_batch_size != 0:\n        mini_batch_X = X[num_complete_minibatches*mini_batch_size:,...]\n        mini_batch_Y = Y[num_complete_minibatches*mini_batch_size:,:]\n        mini_batch = (mini_batch_X, mini_batch_Y)\n        mini_batches.append(mini_batch)\n\n    return mini_batches\n<\/code><\/pre>\n\n<p>First, I am reading the dataset:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># function to read the batches\ndef load_cfar10_batch(cifar10_dataset_folder_path, batch_id):\n    with open(cifar10_dataset_folder_path + '\/data_batch_' + str(batch_id), mode='rb') as file:\n        # note the encoding type is 'latin1'\n        batch = pickle.load(file, encoding='latin1')\n\n    features = batch['data'].reshape((len(batch['data']), 3, 32, 32)).transpose(0, 2, 3, 1)\n    labels = batch['labels']\n    datadict = {'data':features,'labels':labels}\n\n    return datadict\n\n# combine batches into one dataset (batch size: 10000)\nfull_data = load_cfar10_batch('.\/cifar_10',1)['data']\nfull_labels = []\nfor i in range(5):\n    full_labels.extend(load_cfar10_batch('.\/cifar_10',i+1)['labels'])\n    if i &gt; 0:\n        full_data = np.concatenate((full_data,load_cfar10_batch('.\/cifar_10',i+1)['data']),axis = 0)\n\n# dataset sizes\nfull_data.shape, len(full_labels) \n<\/code><\/pre>\n\n<p>Followed by some preprocessing and train\/validation split:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># data preprocessing (using only 1\/10 of the dataset for speed)\n\nX = full_data[0:5000]           \ny = one_hot_matrix(full_labels[0:5000], 10).T       \n\n# split into training-validation sets\nx_train, x_val, y_train, y_val = ms.train_test_split(X, y, test_size=0.2, random_state=1)\n\nx_train = x_train.astype('float32')\nx_val = x_val.astype('float32')\n\nx_train = x_train \/ 255.0\nx_val = x_val \/ 255.0\n\nprint('x_train shape:',x_train.shape)\nprint('y_train shape:',y_train.shape)\nprint('x_val shape:',x_val.shape)\nprint('y_val shape:',y_val.shape)\n<\/code><\/pre>\n\n<p>Then I initialize the variables:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tf.set_random_seed(1)\nseed = 3\n(m, n_H, n_W, n_C) = x_train.shape             \nn_y = y_train.shape[1]\ncosts = []\nprint_cost = True\nlearning_rate = 0.001\nnum_epochs = 100\nminibatch_size = 256\nnum_minibatches = int(m \/ minibatch_size)\nminibatches = make_mini_batches(x_train, y_train, minibatch_size)\nops.reset_default_graph() \n\ninputs = tf.placeholder(tf.float32,shape=[None, n_H, n_W, n_C],name = 'inputs')\nlabels = tf.placeholder(tf.int8,shape=[None, n_y],name = 'labels')\n\n# Forward propagation (Inception)\nZ = inception_v1.inception_v1(inputs,num_classes = n_y,dropout_keep_prob=1,global_pool=True)[1]['Logits']\n# Cost function \ncost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits = Z, labels = labels))\n# ADAM optimizer\noptimizer = tf.train.AdamOptimizer(learning_rate).minimize(cost) \n# Initialize variables\ninit = tf.global_variables_initializer()\n<\/code><\/pre>\n\n<p>And then the training loop:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with tf.Session() as sess:\n\n    sess.run(init)\n\n    for epoch in range(num_epochs):\n\n        # learning rate decay\n        if epoch % 8 == 0:\n            learning_rate *= math.pow(0.95,epoch\/8)\n\n        minibatch_cost = 0.\n\n        for minibatch in minibatches:\n\n            (minibatch_X, minibatch_Y) = minibatch\n            _ , temp_cost = sess.run([optimizer, cost], feed_dict={inputs: minibatch_X, labels: minibatch_Y})\n            minibatch_cost += temp_cost \/ num_minibatches\n\n        # Print the cost every epoch\n        if print_cost == True and epoch % 5 == 0:\n            print (\"Cost after epoch %i: %f\" % (epoch, minibatch_cost),\", Learning rate: %f\" %(learning_rate))\n        if print_cost == True and epoch % 1 == 0:\n            costs.append(minibatch_cost)\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(Z, 1)\n    correct_prediction = tf.equal(predict_op, tf.argmax(labels, 1))\n\n    # Calculate accuracy on the validation set\n    accuracy = tf.reduce_mean(tf.cast(correct_prediction, \"float\"))\n    #print(accuracy)\n    train_accuracy = accuracy.eval({inputs: x_train, labels: y_train})\n    val_accuracy = accuracy.eval({inputs: x_val, labels: y_val})\n    print(\"Train Accuracy:\", train_accuracy)\n    print(\"Validation Accuracy:\", val_accuracy)\n<\/code><\/pre>\n\n<p>The output seems like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Cost after epoch 0: 2.455999 , Learning rate: 0.001000\nCost after epoch 5: 2.454697 , Learning rate: 0.001000\nCost after epoch 10: 2.454670 , Learning rate: 0.000950\nCost after epoch 15: 2.454655 , Learning rate: 0.000950\nCost after epoch 20: 2.454650 , Learning rate: 0.000857\nCost after epoch 25: 2.454649 , Learning rate: 0.000735\nCost after epoch 30: 2.454659 , Learning rate: 0.000735\nCost after epoch 35: 2.454643 , Learning rate: 0.000599\nCost after epoch 40: 2.454627 , Learning rate: 0.000463\n...\n...\n<\/code><\/pre>\n\n<p>And so my network is not training.<\/p>\n","answers":[{"AnswerId":"57679399","CreationDate":"2019-08-27T17:09:22.663","ParentId":null,"OwnerUserId":"10687511","Title":null,"Body":"<p>I managed to find the solution. I had to put the argument scope of inception before calling it, something like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with slim.arg_scope(inception_v1.inception_v1_arg_scope()):\n    Z = inception_v1.inception_v1(inputs,num_classes = n_y,dropout_keep_prob=1,global_pool=True)[1]['Logits']\n<\/code><\/pre>\n\n<p>After that, everything works just fine.<\/p>\n"}]}
+{"QuestionId":57660163,"AnswerCount":1,"Tags":"<python><tensorflow><tensorboard><tensorflow2.0>","CreationDate":"2019-08-26T14:44:45.540","AcceptedAnswerId":null,"OwnerUserId":11774761.0,"Title":"Does tensorflow2.0 still have parameter 'trainable'?","Body":"<p>In tensorboard i can't find gradient ops which to update my parameter just like tensorflow1.X.<\/p>\n\n<p>And don't find parameter 'trainable' in keras api.<\/p>\n\n<p>if tf2.0 still have gradient ops can show in tensorboard,how can i add it to my tensorboard.<\/p>\n\n<p>ps.my tensorflow version is 2.0-rc0.<\/p>\n\n<p>here is my code to add something to tensorboard file.<\/p>\n\n<pre><code>logdir = \"testlogs\"\ntensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=logdir)\n.....\nmodel.fit(x=train_x, y=train_y,\n      batch_size=256,\n      epochs=6,\n      shuffle=True,\n      callbacks=[tensorboard_callback])\n<\/code><\/pre>\n","answers":[{"AnswerId":"57721648","CreationDate":"2019-08-30T06:36:14.043","ParentId":null,"OwnerUserId":"2455494","Title":null,"Body":"<blockquote>\n  <p>Does tensorflow2.0 still have parameter 'trainable'?.<\/p>\n<\/blockquote>\n\n<p>In keras, determining which variables are trainable is a responsibility of the layers that make up your <code>model<\/code>. There are a myriad of layers available out of the box but here is a simple dense layer implementation in order to illustrate the use of some trainable variables<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nclass MyLayer(tf.keras.layers.Layer):\n  def __init__(self, units=8, input_dim=8):\n    super(MyLayer,self).__init__()\n    self.w = tf.Variable(initial_value=tf.random_normal_initializer()(shape=(input_dim, units)),\n                              trainable=True)\n    self.b = tf.Variable(initial_value=tf.zeros_initializer()(shape=(units,)),\n                           trainable=True)\n\n  def call(self, inputs):\n    return tf.matmul(inputs, self.w) + self.b\n\n\n<\/code><\/pre>\n\n<p>which you could, for example, use in a keras model like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>my_layer = MyLayer(units=8,input_dim=2)\nmy_model = tf.keras.models.Sequential([\n    my_layer\n])\nmy_model.compile(optimizer=tf.keras.optimizers.Adam(), loss=tf.keras.losses.binary_crossentropy)\n\n<\/code><\/pre>\n\n<p>Of course best to use the out of the box <code>tf.keras.layers.Dense<\/code> in practice, this is just to illustrate the trainable variables <code>my_layer.w<\/code> &amp; <code>my_layer.b<\/code>! <\/p>\n"}]}
+{"QuestionId":57660214,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorflow2.0>","CreationDate":"2019-08-26T14:47:04.320","AcceptedAnswerId":null,"OwnerUserId":3860928.0,"Title":"What is the utility of `Tensor` (as opposed to `EagerTensor`) in Tensorflow 2.0?","Body":"<p>In Tensorflow 2.0, the main \"tensors\" we see are in fact <code>EagerTensors<\/code> (<code>tensorflow.python.framework.ops.EagerTensor<\/code> to be more precise):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = [[2.]]\nm = tf.matmul(x, x)\ntype(m)\n# returns tensorflow.python.framework.ops.EagerTensor\n<\/code><\/pre>\n\n<p>But, in some cases, we have the symbolic <code>Tensor<\/code> object (<code>tensorflow.python.framework.ops.Tensor<\/code>), as in TF1.X.<br>\nFor example in keras:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import Dense\n\nmodel = Sequential()\nmodel.add(Dense(units=64, activation='relu', input_dim=100))\nmodel.add(Dense(units=10, activation='softmax'))\ntype(model.outputs[0])\n# returns tensorflow.python.framework.ops.Tensor\n<\/code><\/pre>\n\n<p>So, what are the use for these symbolic:<code>tensorflow.python.framework.ops.Tensor<\/code> in Tensorflow:<\/p>\n\n<ul>\n<li>In the TF library internals: Keras is at least using these tensors, but is it used at other places (which are using a graph, like tf.function, or tf.data.Dataset)?<\/li>\n<li>In the API: is there an actual use for end-users of these?<\/li>\n<\/ul>\n","answers":[{"AnswerId":"57709011","CreationDate":"2019-08-29T11:18:20.867","ParentId":null,"OwnerUserId":"2455494","Title":null,"Body":"<blockquote>\n  <p>In the TF library internals: Keras is at least using these tensors, but is it used at other places (which are using a graph, like tf.function, or tf.data.Dataset)?<\/p>\n<\/blockquote>\n\n<p>Well, yes. Your instinct is correct here. <code>EagreTensor<\/code> represents a tensor who's value has been calculated in eager mode whereas <code>Tensor<\/code> represents a tensor node in a graph that may not yet have been calculated.<\/p>\n\n<blockquote>\n  <p>In the API: is there an actual use for end-users of these?<\/p>\n<\/blockquote>\n\n<p>Well, on some level we use them all the time. We create keras models, tf.data.Dataset pipelines etc. but, really, for the vast majority of use-cases we don't tend to instantiate or interact directly with the <code>Tensor<\/code> object itself though so probably just want to not worry about the object type and consider them to be an implementation detail internal to tensorflow.<\/p>\n"}]}
+{"QuestionId":57660316,"AnswerCount":0,"Tags":"<tensorflow><recurrent-neural-network><seq2seq>","CreationDate":"2019-08-26T14:53:36.023","AcceptedAnswerId":null,"OwnerUserId":11978908.0,"Title":"My Seq2Seq RNN code does'nt work properly","Body":"<p>I'm trying to use dynamic RNN. However, when I run my code, it reports<\/p>\n\n<pre><code>ERROR:Inconsistent shapes: saw (40, 50) but expected (1, 50) (and infer_shape=True)\n<\/code><\/pre>\n\n<p>Reviewing several time but can't find the key to the problem. Asking for some help!<\/p>\n\n<p>Environment:Python3.7.3, Tensorflow 1.14<\/p>\n\n<pre><code>    ...\n#Encoder Part\ne_cell=tf.nn.rnn_cell.GRUCell(num_units=12)\nX_train= X_train.transpose((1,0,2))     #shape:(40,2361,4)\n\ninit_state = Mcell.zero_state(40, dtype=tf.float32)\nencoder_output, encoder_state = tf.nn.dynamic_rnn(e_cell, X_train, \n                                              initial_state=init_state) \n\n#Decoder Part\n#\/helper\ninference = False\nseq_len= np.array([20])  #it seems to be wrong here?   \nhelper = tf.contrib.seq2seq.TrainingHelper(encoder_output, \n                                       sequence_length=seq_len)\n\n#\/attention\nattention_mechanism = \ntf.contrib.seq2seq.BahdanauAttention(rnn_hidden_size=12, \n                                                       encoder_output)\nd_cell = tf.nn.rnn_cell.GRUCell(rnn_hidden_size=12)\ndecoder_cell = tf.contrib.seq2seq.AttentionWrapper(d_cell, \n             attention_mechanism, attention_layer_size = 12)\nout_cell = tf.contrib.rnn.OutputProjectionWrapper(decoder_cell, \n             target_size=50)\ndecoder_state = decoder_cell.zero_state(batch_size=40,dtype=tf.float32)\n\nwith tf.variable_scope('decoder'):\ndecoder = tf.contrib.seq2seq.BasicDecoder(out_cell, helper,decoder_state, \n             tf.layers.Dense(50))\n\n#\/dynamic decoder\ndec_outputs, final_state, final_sequence_lengths = \n        tf.contrib.seq2seq.dynamic_decode(decoder,swap_memory=True)\n...\n<\/code><\/pre>\n\n<p><strong>Question1:<\/strong> How to fix the error:<\/p>\n\n<pre><code>saw (40, 50) but expected (1, 50) (and infer_shape=True)\n<\/code><\/pre>\n\n<p><strong>Question2:<\/strong> I did'nt use embedding layer, is that right to use <code>TrainingHelper<\/code> or it should use <code>ScheduledOutputTrainingHelper<\/code> instead?<\/p>\n\n<p>I may add more information if the code above is not helpful. Thanks!<\/p>\n","answers":[]}
+{"QuestionId":57661086,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><neural-network><mnist>","CreationDate":"2019-08-26T15:45:47.097","AcceptedAnswerId":57661940.0,"OwnerUserId":8067708.0,"Title":"my neural network is not learning cost is not changing","Body":"<p>I'm building a neural network for classifying mnist digits I got the data from \nI tried to built it with only tensorflow I didn't want to use keras\n<a href=\"https:\/\/www.kaggle.com\/c\/digit-recognizer\/data\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/c\/digit-recognizer\/data<\/a> \nand through epochs cost is not decreasing <\/p>\n\n<p>I followed that tutorial \n<a href=\"https:\/\/adventuresinmachinelearning.com\/python-tensorflow-tutorial\/\" rel=\"nofollow noreferrer\">https:\/\/adventuresinmachinelearning.com\/python-tensorflow-tutorial\/<\/a>\nthe difference between my code and his , is that I didn't import the data from tensorflow  and I replaced deprecated  functions \nnote : I'm using anaconda and tensorflow==1.14<\/p>\n\n<h3>that is my code<\/h3>\n\n<pre><code>import tensorflow as tf\nimport pandas as pd\n# from tensorflow import keras\n# import matplotlib.pyplot as plt\nimport numpy as np\n\ntr_csv = \"file:\/\/\/D:\/code\/ml\/neuralnet\/train.csv\"\ndata = pd.read_csv(tr_csv)\n\ndata  = data[:30000]\ntrain = data[:24000]\ntest = data[24000:]\n\nlearning_rate = 1\nepochs = 10\nbatch_size = 100\n\nxtf = tf.compat.v1.placeholder(tf.float32, [None, 784])\nytf = tf.compat.v1.placeholder(tf.float32, [None, 10])\n\nW1 = tf.Variable(tf.random.normal([784, 300], stddev=0.03), name='W1')\nb1 = tf.Variable(tf.random.normal([300]), name='b1')\nW2 = tf.Variable(tf.random.normal([300, 10], stddev=0.03), name='W2')\nb2 = tf.Variable(tf.random.normal([10]), name='b2')\n\nhidden1 = tf.add(tf.matmul(xtf,W1),b1)\nhidden1 = tf.nn.relu(hidden1)\n\ny_ = tf.nn.softmax(tf.add(tf.matmul(hidden1,W2),b2))\nyclipped =  tf.clip_by_value(y_, 1e-10, 0.9999999)\n\ncross_entropy = -tf.reduce_mean(tf.reduce_sum(ytf * tf.log(yclipped)+ (1 - ytf) * tf.log(1 - yclipped), axis=1))\noptimiser = tf.compat.v1.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(cross_entropy)\n\ncorrect_prediction = tf.equal(tf.argmax(ytf, 1), tf.argmax(y_, 1))\naccuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n\nwith tf.compat.v1.Session() as sess :\n    init_op = tf.compat.v1.global_variables_initializer()\n    sess.run(init_op)\n    total_batch = int(len(train.label) \/ batch_size)\n    for epoch in range(10):\n        avg_cost = 0\n        for chunk in np.array_split(train, 24000\/batch_size):\n   #     for i in range(total_batch)\n            xb = chunk.drop([\"label\"],axis = 1 )\n            yb = chunk.label\n            yb = pd.Categorical(yb)\n            yb = pd.get_dummies(yb,prefix = \"number\")\n            #(batch_size = batch_size) x\n\n            (_, c) = sess.run([optimiser, cross_entropy],feed_dict={xtf : xb.to_numpy() ,ytf: yb.to_numpy()})\n\n            avg_cost += c \/ batch_size\n        print(\"Epoch:\", (epoch + 1), \"cost =\", \"{:.3f}\".format(avg_cost))\n        #print(sess.run(accuracy, feed_dict={x: xt., y: mnist.test.labels}))\n\n<\/code><\/pre>\n\n<h3>that is cost through epochs<\/h3>\n\n<pre><code>\n\nEpoch: 1 cost = 84.110\nEpoch: 2 cost = 84.109\nEpoch: 3 cost = 84.109\nEpoch: 4 cost = 84.109\nEpoch: 5 cost = 84.109\nEpoch: 6 cost = 84.109\nEpoch: 7 cost = 84.109\nEpoch: 8 cost = 84.109\nEpoch: 9 cost = 84.109\nEpoch: 10 cost = 84.109\n<\/code><\/pre>\n","answers":[{"AnswerId":"57661940","CreationDate":"2019-08-26T16:52:03.823","ParentId":null,"OwnerUserId":"7962217","Title":null,"Body":"<p>For any neural network, parameter tuning is essential. You can try different combinations to come up with a suitable value. Cost function gives an idea as to whether convergence has reached or not. If it does not converge, try a new set of parameters.<\/p>\n"}]}
+{"QuestionId":57661284,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-08-26T15:59:48.927","AcceptedAnswerId":null,"OwnerUserId":11979471.0,"Title":"adding weighted mse loss function in keras","Body":"<p>I built a weighted mse loss function in Keras, all according to the documentation. However, the weights \"Pz\" are an additional input to the model, and totally unrelated to the other inputs. Somehow I can't figure out how to pass that to the fit command<\/p>\n\n<pre><code>def weighted_mse(weight):\n    def loss(y_true, y_pred):\n        tmp = K.square(y_true-y_pred)\n        return K.sum(tmp*weight) \/ K.sum(weight)\n    return loss\n\n\nmodel = Sequential()\nmodel.add(LSTM(nbin*2, return_sequences=True, input_shape=(None, nbin)))\nmodel.add(Dense(nbin*2))\nmodel.add(Activation('relu'))\nmodel.add(Dense(nbin))\nmodel.add(Activation('sigmoid'))\n\nPz = Input(shape=(nbin,))\nmodel.compile(loss=weighted_mse(Pz), optimizer='adam')\n\nmodel.fit([p_evd, Pz], p_opt, batch_size=nfram)\n<\/code><\/pre>\n\n<p>This code gives the following error during compilation:<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking model input: the list of Numpy arrays\n  that you are passing to your model is not the size the model expected.\n  Expected to see 1 array(s), but instead got the following list of 2\n  arrays: [array([[[3.299e-01, 1.782e-01, 4.603e-02, ..., 1.454e-02,\n  6.400e-03, 2.321e-03],\n          [5.506e-01, 2.307e-01, 1.520e-02, ..., 1.874e-03, 5.343e-03, 4.138e-03],\n          [8.565e-03, 1.696e-01, 2.259e...<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":57661490,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-08-26T16:15:33.733","AcceptedAnswerId":57674731.0,"OwnerUserId":6463668.0,"Title":"Can I reset a specified variable of keras when training?","Body":"<p>Can I reinitialize or reset a variable at the begin of every training batch in keras?<\/p>\n\n<p>Here's a toy example:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import keras.backend as K\nfrom keras.layers import Input\nfrom keras.models import Model\nfrom keras.optimizers import Adam\nimport numpy as np\n\nx = Input(shape=(None, 4))\n\ndef toy_test():\n    m = K.zeros(shape=(16, 9), dtype=\"int32\")\n\n    # computation here, some elements of m will be changed here\n    m = m[2, 3].assign(m[2, 3]+1)\n\n    result = (K.sum(K.cast(m &gt; 1, dtype=\"float32\"))) \/ (K.sum(K.cast(m &gt;= 1, dtype=\"float32\")) + 1e-4)\n\n    # I want to reset m to all zeros\n    m = K.zeros(shape=(16, 9), dtype=\"int32\")\n\n    return result\n\nr= toy_test()\nmodel = Model([x], outputs=x)\nmodel.add_loss(r)\nmodel.summary()\nmodel.compile(optimizer=Adam(1e-5))\n\ntest_x = np.array([\n    [[1, 2, 3, 4], [2, 3, 4, 5]],\n    [[1, 2, 3, 4], [2, 3, 4, 5]],\n])\nhistory = model.fit(x=test_x, batch_size=1, shuffle=False)\n<\/code><\/pre>\n\n<p>Expected result:<\/p>\n\n<p>1\/2 loss: 0.0<\/p>\n\n<p>2\/2 loss: 0.0<\/p>\n\n<p>Actual result:<\/p>\n\n<p>1\/2 loss: 0.0000e+00<\/p>\n\n<p>2\/2 loss: 0.4999<\/p>\n\n<p>It seems that the variable is only initialized once and remains without reset at second batch. How can I reset <code>m<\/code> to all zeros?<\/p>\n\n<p>It seems <code>tf.control_dependencies<\/code> may can be used to control the flow of computation, but can it be used to control variable?<\/p>\n","answers":[{"AnswerId":"57674731","CreationDate":"2019-08-27T12:29:38.173","ParentId":null,"OwnerUserId":"6463668","Title":null,"Body":"<p>Finally I solve it by using <code>tf.control_dependencies<\/code>.\nHere is the modified part of <code>toy_test<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def toy_test():\n    m = tf.get_variable(shape=(16, 9), dtype=\"int32\", name='test_m', use_resource=True)\n    assign_op = m.assign(np.zeros((16, 9), dtype=int))\n    # some computation\n    with tf.control_dependencies([assign_op]):\n        m = assign_op\n        m[2, 3].assign(m[2, 3]+1)\n        result = (K.sum(K.cast(m &gt; 1, dtype=\"float32\"))) \/ (K.sum(K.cast(m &gt;= 1, dtype=\"float32\")) + 1e-4)\n\n    return result\n<\/code><\/pre>\n"}]}
+{"QuestionId":57661516,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><conv-neural-network>","CreationDate":"2019-08-26T16:17:36.870","AcceptedAnswerId":57671716.0,"OwnerUserId":8176285.0,"Title":"Converting Keras model to multi label output","Body":"<p>I have a model which takes in a dataframe which looks like this <\/p>\n\n<pre><code>image,level\n10_left,0\n10_right,0\n13_left,0\n<\/code><\/pre>\n\n<p>with model structure like this<\/p>\n\n<pre><code>base_image_dir = 'extra_data\/dr\/'\nretina_df = pd.read_csv(os.path.join(base_image_dir, 'trainLabels.csv'))\nretina_df['PatientId'] = retina_df['image'].map(lambda x: x.split('_')[0])\nretina_df['path'] = retina_df['image'].map(lambda x: os.path.join(base_image_dir,'train',\n                                                         '{}.jpeg'.format(x)))\nretina_df['exists'] = retina_df['path'].map(os.path.exists)\nprint(retina_df['exists'].sum(), 'images found of', retina_df.shape[0], 'total')\nretina_df['eye'] = retina_df['image'].map(lambda x: 1 if x.split('_')[-1]=='left' else 0)\nfrom keras.utils.np_utils import to_categorical\nretina_df['level_cat'] = retina_df['level'].map(lambda x: to_categorical(x, 1+retina_df['level'].max()))\n\nretina_df.dropna(inplace = True)\nretina_df = retina_df[retina_df['exists']]\nretina_df.sample(3)\n\nfrom sklearn.model_selection import train_test_split\nrr_df = retina_df[['PatientId', 'level']].drop_duplicates()\ntrain_ids, valid_ids = train_test_split(rr_df['PatientId'], \n                                   test_size = 0.25, \n                                   random_state = 2018,\n                                   stratify = rr_df['level'])\nraw_train_df = retina_df[retina_df['PatientId'].isin(train_ids)]\nvalid_df = retina_df[retina_df['PatientId'].isin(valid_ids)]\nimport pdb;pdb.set_trace()\nprint('train', raw_train_df.shape[0], 'validation', valid_df.shape[0])\n\ntrain_df = raw_train_df.groupby(['level', 'eye']).apply(lambda x: x.sample(75, replace = True)                                                    ).reset_index(drop = True)\nprint('New Data Size:', train_df.shape[0], 'Old Size:', raw_train_df.shape[0])\n\nimport tensorflow as tf\nfrom keras import backend as K\nfrom keras.applications.inception_v3 import preprocess_input\nimport numpy as np\nIMG_SIZE = (512, 512) # slightly smaller than vgg16 normally expects\ndef tf_image_loader(out_size, \n                      horizontal_flip = True, \n                      vertical_flip = False, \n                     random_brightness = True,\n                     random_contrast = True,\n                    random_saturation = True,\n                    random_hue = True,\n                      color_mode = 'rgb',\n                       preproc_func = preprocess_input,\n                       on_batch = False):\n    def _func(X):\n        with tf.name_scope('image_augmentation'):\n            with tf.name_scope('input'):\n                X = tf.image.decode_png(tf.read_file(X), channels = 3 if color_mode == 'rgb' else 0)\n                X = tf.image.resize_images(X, out_size)\n            with tf.name_scope('augmentation'):\n                if horizontal_flip:\n                    X = tf.image.random_flip_left_right(X)\n                if vertical_flip:\n                    X = tf.image.random_flip_up_down(X)\n                if random_brightness:\n                    X = tf.image.random_brightness(X, max_delta = 0.1)\n                if random_saturation:\n                    X = tf.image.random_saturation(X, lower = 0.75, upper = 1.5)\n                if random_hue:\n                    X = tf.image.random_hue(X, max_delta = 0.15)\n                if random_contrast:\n                    X = tf.image.random_contrast(X, lower = 0.75, upper = 1.5)\n                return preproc_func(X)\n    if on_batch: \n        # we are meant to use it on a batch\n        def _batch_func(X, y):\n            return tf.map_fn(_func, X), y\n        return _batch_func\n    else:\n        # we apply it to everything\n        def _all_func(X, y):\n            return _func(X), y         \n        return _all_func\n\ndef tf_augmentor(out_size,\n                intermediate_size = (640, 640),\n                 intermediate_trans = 'crop',\n                 batch_size = 16,\n                   horizontal_flip = True, \n                  vertical_flip = False, \n                 random_brightness = True,\n                 random_contrast = True,\n                 random_saturation = True,\n                    random_hue = True,\n                  color_mode = 'rgb',\n                   preproc_func = preprocess_input,\n                   min_crop_percent = 0.001,\n                   max_crop_percent = 0.005,\n                   crop_probability = 0.5,\n                   rotation_range = 10):\n\n    load_ops = tf_image_loader(out_size = intermediate_size, \n                               horizontal_flip=horizontal_flip, \n                               vertical_flip=vertical_flip, \n                               random_brightness = random_brightness,\n                               random_contrast = random_contrast,\n                               random_saturation = random_saturation,\n                               random_hue = random_hue,\n                               color_mode = color_mode,\n                               preproc_func = preproc_func,\n                               on_batch=False)\n    def batch_ops(X, y):\n        batch_size = tf.shape(X)[0]\n        with tf.name_scope('transformation'):\n            # code borrowed from https:\/\/becominghuman.ai\/data-augmentation-on-gpu-in-tensorflow-13d14ecf2b19\n            # The list of affine transformations that our image will go under.\n            # Every element is Nx8 tensor, where N is a batch size.\n            transforms = []\n            identity = tf.constant([1, 0, 0, 0, 1, 0, 0, 0], dtype=tf.float32)\n            if rotation_range &gt; 0:\n                angle_rad = rotation_range \/ 180 * np.pi\n                angles = tf.random_uniform([batch_size], -angle_rad, angle_rad)\n                transforms += [tf.contrib.image.angles_to_projective_transforms(angles, intermediate_size[0], intermediate_size[1])]\n\n            if crop_probability &gt; 0:\n                crop_pct = tf.random_uniform([batch_size], min_crop_percent, max_crop_percent)\n                left = tf.random_uniform([batch_size], 0, intermediate_size[0] * (1.0 - crop_pct))\n                top = tf.random_uniform([batch_size], 0, intermediate_size[1] * (1.0 - crop_pct))\n                crop_transform = tf.stack([\n                      crop_pct,\n                      tf.zeros([batch_size]), top,\n                      tf.zeros([batch_size]), crop_pct, left,\n                      tf.zeros([batch_size]),\n                      tf.zeros([batch_size])\n                  ], 1)\n                coin = tf.less(tf.random_uniform([batch_size], 0, 1.0), crop_probability)\n                transforms += [tf.where(coin, crop_transform, tf.tile(tf.expand_dims(identity, 0), [batch_size, 1]))]\n            if len(transforms)&gt;0:\n                X = tf.contrib.image.transform(X,\n                      tf.contrib.image.compose_transforms(*transforms),\n                      interpolation='BILINEAR') # or 'NEAREST'\n            if intermediate_trans=='scale':\n                X = tf.image.resize_images(X, out_size)\n            elif intermediate_trans=='crop':\n                X = tf.image.resize_image_with_crop_or_pad(X, out_size[0], out_size[1])\n            else:\n                raise ValueError('Invalid Operation {}'.format(intermediate_trans))\n            return X, y\n    def _create_pipeline(in_ds):\n        batch_ds = in_ds.map(load_ops, num_parallel_calls=4).batch(batch_size)\n        return batch_ds.map(batch_ops)\n    return _create_pipeline\n\ndef flow_from_dataframe(idg, \n                        in_df, \n                        path_col,\n                        y_col, \n                        shuffle = True, \n                        color_mode = 'rgb'):\n    files_ds = tf.data.Dataset.from_tensor_slices((in_df[path_col].values, \n                                                   np.stack(in_df[y_col].values,0)))\n    in_len = in_df[path_col].values.shape[0]\n    while True:\n        if shuffle:\n            files_ds = files_ds.shuffle(in_len) # shuffle the whole dataset\n\n        next_batch = idg(files_ds).repeat().make_one_shot_iterator().get_next()\n        for i in range(max(in_len\/\/32,1)):\n            # NOTE: if we loop here it is 'thread-safe-ish' if we loop on the outside it is completely unsafe\n            yield K.get_session().run(next_batch)\n\nbatch_size = 48\ncore_idg = tf_augmentor(out_size = IMG_SIZE, \n                        color_mode = 'rgb', \n                        vertical_flip = True,\n                        crop_probability=0.0, # crop doesn't work yet\n                        batch_size = batch_size) \nvalid_idg = tf_augmentor(out_size = IMG_SIZE, color_mode = 'rgb', \n                         crop_probability=0.0, \n                         horizontal_flip = False, \n                         vertical_flip = False, \n                         random_brightness = False,\n                         random_contrast = False,\n                         random_saturation = False,\n                         random_hue = False,\n                         rotation_range = 0,\n                        batch_size = batch_size)\n\ntrain_gen = flow_from_dataframe(core_idg, train_df, \n                             path_col = 'path',\n                            y_col = 'level_cat')\n\nvalid_gen = flow_from_dataframe(valid_idg, valid_df, \n                             path_col = 'path',\n                            y_col = 'level_cat') # we can use much larger batches for evaluation\n\nt_x, t_y = next(valid_gen)\nt_x, t_y = next(train_gen)\nfrom keras.applications.vgg16 import VGG16 as PTModel\nfrom keras.applications.inception_resnet_v2 import InceptionResNetV2 as PTModel\nfrom keras.applications.inception_v3 import InceptionV3 as PTModel\nfrom keras.layers import GlobalAveragePooling2D, Dense, Dropout, Flatten, Input, Conv2D, multiply, LocallyConnected2D, Lambda\nfrom keras.models import Model\nin_lay = Input(t_x.shape[1:])\nbase_pretrained_model = PTModel(input_shape =  t_x.shape[1:], include_top = False, weights = 'imagenet')\nbase_pretrained_model.trainable = False\npt_depth = base_pretrained_model.get_output_shape_at(0)[-1]\npt_features = base_pretrained_model(in_lay)\nfrom keras.layers import BatchNormalization\nbn_features = BatchNormalization()(pt_features)\n\nattn_layer = Conv2D(64, kernel_size = (1,1), padding = 'same', activation = 'relu')(Dropout(0.5)(bn_features))\nattn_layer = Conv2D(16, kernel_size = (1,1), padding = 'same', activation = 'relu')(attn_layer)\nattn_layer = Conv2D(8, kernel_size = (1,1), padding = 'same', activation = 'relu')(attn_layer)\nattn_layer = Conv2D(1, \n                    kernel_size = (1,1), \n                    padding = 'valid', \n                    activation = 'sigmoid')(attn_layer)\n# fan it out to all of the channels\nup_c2_w = np.ones((1, 1, 1, pt_depth))\nup_c2 = Conv2D(pt_depth, kernel_size = (1,1), padding = 'same', \n               activation = 'linear', use_bias = False, weights = [up_c2_w])\nup_c2.trainable = False\nattn_layer = up_c2(attn_layer)\n\nmask_features = multiply([attn_layer, bn_features])\ngap_features = GlobalAveragePooling2D()(mask_features)\ngap_mask = GlobalAveragePooling2D()(attn_layer)\n# to account for missing values from the attention model\ngap = Lambda(lambda x: x[0]\/x[1], name = 'RescaleGAP')([gap_features, gap_mask])\ngap_dr = Dropout(0.25)(gap)\ndr_steps = Dropout(0.25)(Dense(128, activation = 'relu')(gap_dr))\nout_layer = Dense(t_y.shape[-1], activation = 'softmax')(dr_steps)\nretina_model = Model(inputs = [in_lay], outputs = [out_layer])\nfrom keras.metrics import top_k_categorical_accuracy\ndef top_2_accuracy(in_gt, in_pred):\n    return top_k_categorical_accuracy(in_gt, in_pred, k=2)\n\nretina_model.compile(optimizer = 'adam', loss = 'categorical_crossentropy',\n                           metrics = ['categorical_accuracy', top_2_accuracy])\nretina_model.summary()\nfrom keras.callbacks import ModelCheckpoint, LearningRateScheduler, EarlyStopping, ReduceLROnPlateau\nweight_path=\"{}_weights.best.hdf5\".format('retina')\n\ncheckpoint = ModelCheckpoint(weight_path, monitor='val_loss', verbose=1, \n                             save_best_only=True, mode='min', save_weights_only = True)\n\nreduceLROnPlat = ReduceLROnPlateau(monitor='val_loss', factor=0.8, patience=3, verbose=1, mode='auto', epsilon=0.0001, cooldown=5, min_lr=0.0001)\nearly = EarlyStopping(monitor=\"val_loss\", \n                      mode=\"min\", \n                      patience=6) # probably needs to be more patient, but kaggle time is limited\ncallbacks_list = [checkpoint, early, reduceLROnPlat]\n\nretina_model.fit_generator(train_gen, \n                           steps_per_epoch = train_df.shape[0]\/\/batch_size,\n                           validation_data = valid_gen, \n                           validation_steps = valid_df.shape[0]\/\/batch_size,\n                              epochs = 25, \n                              callbacks = callbacks_list,\n                             workers = 0, # tf-generators are not thread-safe\n                             use_multiprocessing=False, \n                             max_queue_size = 0\n                            )\nretina_model.load_weights(weight_path)\nretina_model.save('full_retina_model.h5')\n<\/code><\/pre>\n\n<p>I realize that's a lot of code, but i what i want to do is take in a dataframe which look like this <\/p>\n\n<pre><code>image,N,D,G,C,A,H,M,O\n2857_left,1,0,0,0,0,0,0,0\n3151_left,1,0,0,0,0,0,0,0\n3113_left,1,0,0,0,0,0,0,0\n<\/code><\/pre>\n\n<p>and in order to achive this i have made this following changes,<\/p>\n\n<pre><code>from sklearn.model_selection import train_test_split\nrr_df = retina_df\ny = rr_df[['N', 'D', 'G','C','A', 'H', 'M', 'O']]\ntrain_ids, valid_ids = train_test_split(rr_df['PatientId'], \n                                   test_size = 0.25, \n                                   random_state = 2018)\n\nraw_train_df = retina_df[retina_df['PatientId'].isin(train_ids)]\nvalid_df = retina_df[retina_df['PatientId'].isin(valid_ids)]\nprint('train', raw_train_df.shape[0], 'validation', valid_df.shape[0])\ntrain_df = raw_train_df\n\nfrom keras import regularizers, optimizers\nfrom keras.layers import BatchNormalization\nin_lay = Input(t_x.shape[1:])\nbase_pretrained_model = PTModel(input_shape =  t_x.shape[1:], include_top = False, weights = 'imagenet')\nbase_pretrained_model.trainable = False\npt_depth = base_pretrained_model.get_output_shape_at(0)[-1]\npt_features = base_pretrained_model(in_lay)\nbn_features = BatchNormalization()(pt_features)\n\n# here we do an attention mechanism to turn pixels in the GAP on an off\n\nattn_layer = Conv2D(64, kernel_size = (1,1), padding = 'same', activation = 'relu')(Dropout(0.5)(bn_features))\nattn_layer = Conv2D(16, kernel_size = (1,1), padding = 'same', activation = 'relu')(attn_layer)\nattn_layer = Conv2D(8, kernel_size = (1,1), padding = 'same', activation = 'relu')(attn_layer)\nattn_layer = Conv2D(1, \n                    kernel_size = (1,1), \n                    padding = 'valid', \n                    activation = 'sigmoid')(attn_layer)\n# fan it out to all of the channels\nup_c2_w = np.ones((1, 1, 1, pt_depth))\nup_c2 = Conv2D(pt_depth, kernel_size = (1,1), padding = 'same', \n               activation = 'linear', use_bias = False, weights = [up_c2_w])\nup_c2.trainable = False\nattn_layer = up_c2(attn_layer)\n\nmask_features = multiply([attn_layer, bn_features])\ngap_features = GlobalAveragePooling2D()(mask_features)\ngap_mask = GlobalAveragePooling2D()(attn_layer)\n# to account for missing values from the attention model\ngap = Lambda(lambda x: x[0]\/x[1], name = 'RescaleGAP')([gap_features, gap_mask])\ngap_dr = Dropout(0.25)(gap)\nx = Dropout(0.25)(Dense(128, activation = 'relu')(gap_dr))\n# out_layer = Dense(t_y.shape[-1], activation = 'softmax')(dr_steps)\noutput1 = Dense(1, activation = 'sigmoid')(x)\noutput2 = Dense(1, activation = 'sigmoid')(x)\noutput3 = Dense(1, activation = 'sigmoid')(x)\noutput4 = Dense(1, activation = 'sigmoid')(x)\noutput5 = Dense(1, activation = 'sigmoid')(x)\noutput6 = Dense(1, activation = 'sigmoid')(x)\noutput7 = Dense(1, activation = 'sigmoid')(x)\noutput8 = Dense(1, activation = 'sigmoid')(x)\nretina_model = Model(inputs = [in_lay], outputs = [output1,output2,output3,output4,output5, output6, output7, output8])\n# retina_model = Model([in_lay],output1,output2,output3,output4,output5, output6, output7, output8)\n# retina_model.build(t_x.shape[1:]) # `input_shape` is the shape of the input data\n\n# print(model.summary())\n\n# retina_model.compile(optimizers.rmsprop(lr = 0.00001, decay = 1e-6),\nloss = [\"binary_crossentropy\",\"binary_crossentropy\",\"binary_crossentropy\",\"binary_crossentropy\", \"binary_crossentropy\",\"binary_crossentropy\",\"binary_crossentropy\",\"binary_crossentropy\"]#,metrics = [\"accuracy\"])\n\n# retina_model = Model(inputs = [in_lay], outputs = [out_layer])\n# from keras.metrics import top_k_categorical_accuracy\n# def top_2_accuracy(in_gt, in_pred):\n    # return top_k_categorical_accuracy(in_gt, in_pred, k=2)\n\nretina_model.compile(optimizer = 'adam', loss = loss,\n                           metrics = ['accuracy'])\nretina_model.summary()\n<\/code><\/pre>\n\n<p>but when i run this i get, <\/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 8 array(s), but instead got the following list of 1 arrays: [array([[1, 0, 0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 0, 0, 0, 0, 1],\n         [0, 1, 0, 0, 0, 0, 0, 0],\n         [0, 0, 0, 1, 0, 0, 0, 0],\n         [1, 0, 0, 0, 0, 0, 0, 0],\n         [1, 0, 0, 0, 0, 0, 0, 0],\n   ...\n<\/code><\/pre>\n\n<p>Any suggestions on how i could change this model to train this on multi label inputs.Thanks in advance.<\/p>\n","answers":[{"AnswerId":"57671716","CreationDate":"2019-08-27T09:35:08.660","ParentId":null,"OwnerUserId":"8563649","Title":null,"Body":"<p>You are trying to train a model with 8 different outputs (length 1 for every output)  but your target values is an array of length 8.<\/p>\n\n<p>The easiest fix is to replace:<\/p>\n\n<pre><code>output1 = Dense(1, activation = 'sigmoid')(x)\noutput2 = Dense(1, activation = 'sigmoid')(x)\noutput3 = Dense(1, activation = 'sigmoid')(x)\noutput4 = Dense(1, activation = 'sigmoid')(x)\noutput5 = Dense(1, activation = 'sigmoid')(x)\noutput6 = Dense(1, activation = 'sigmoid')(x)\noutput7 = Dense(1, activation = 'sigmoid')(x)\noutput8 = Dense(1, activation = 'sigmoid')(x)\n\nloss = [\"binary_crossentropy\",\"binary_crossentropy\",\"binary_crossentropy\",\"binary_crossentropy\", \"binary_crossentropy\",\"binary_crossentropy\",\"binary_crossentropy\",\"binary_crossentropy\"]#,metrics = [\"accuracy\"])\n<\/code><\/pre>\n\n<p>with:<\/p>\n\n<pre><code>#leave sigmoid here, don't change with softmax if it is a multilabel problem\noutput = Dense(8, activation = 'sigmoid')(x) \nloss = \"binary_crossentropy\"\n<\/code><\/pre>\n\n<p>otherwise you have to create a custom generator with yielding a list of 8 targets  to feed your network<\/p>\n"}]}
+{"QuestionId":57661592,"AnswerCount":0,"Tags":"<python><tensorflow><keras><horovod>","CreationDate":"2019-08-26T16:23:21.607","AcceptedAnswerId":null,"OwnerUserId":1514682.0,"Title":"FailedPreconditionError: Error while reading resource variable *** from Container","Body":"<p>I am seeing following error on running model.fit with the horovod callbacks. If I skip the callbacks model.fit runs fine. Note: I am using <code>horovod.tensorflow.keras<\/code> package and my model is based on <code>tensorflow.keras<\/code> (I am not direclty using keras package but from the tensorflow)<\/p>\n\n<pre><code>FailedPreconditionError: Error while reading resource variable conv1d\/kernel from Container: localhost. This could mean that the variable was uninitialized. Not found: Resource localhost\/conv1d\/kernel\/N10tensorflow3VarE does not exist.\n     [[{{node conv1d\/conv1d\/ExpandDims_1\/ReadVariableOp}}]]\n<\/code><\/pre>\n\n<p>Callbacks are as following<\/p>\n\n<pre><code>callbacks = [\n    # Horovod: broadcast initial variable states from rank 0 to all other processes.\n    # This is necessary to ensure consistent initialization of all workers when\n    # training is started with random weights or restored from a checkpoint.\n    hvd.callbacks.BroadcastGlobalVariablesCallback(0),\n    TensorBoard(log_dir='boardlogs\/{}'.format(datetime.datetime.now())) #report logs to tensorboard\n]\n# Horovod: save checkpoints only on worker 0 to prevent other workers from corrupting them.\nif hvd.rank() == 0:\n    horovod_callbacks.append(tf.keras.callbacks.ModelCheckpoint('.horovod-cps\/checkpoint-{epoch}.h5')\n\n\nhistory = model.fit(X, y, epochs=500, batch_size=64, callbacks=callbacks, verbose=1 if hvd.rank() == 0 else 0)\n<\/code><\/pre>\n\n<p><strong>Environment:<\/strong>\nFramework: tensorflow.keras\nTensorflow Version 1.13.1\nKeras Version 2.2.4-tf\nHorovod version: horovod==0.17.0.post1\nPython version: 3.6<\/p>\n","answers":[]}
+{"QuestionId":57661596,"AnswerCount":0,"Tags":"<r><keras>","CreationDate":"2019-08-26T16:23:30.407","AcceptedAnswerId":null,"OwnerUserId":6447399.0,"Title":"Keras word2vec model - average word vectors - best approach","Body":"<p>I have followed the word2vec tutorial using keras in R <a href=\"https:\/\/blogs.rstudio.com\/tensorflow\/posts\/2017-12-22-word-embeddings-with-keras\/\" rel=\"nofollow noreferrer\">here<\/a> and I have managed to obtain word embeddings for each word. I would like to now pass my word embeddings to each <code>document<\/code> and put the <code>document<\/code> through some classifier model such as logistic regression. Using the code below I get to this line:<\/p>\n\n<p><code>row.names(embedding_matrix) &lt;- c(\"UNK\", words$word)<\/code><\/p>\n\n<p>I have an embedding matrix which look like the following:<\/p>\n\n<pre><code>&gt; embedding_matrix[1:5, 1:5]\n                  [,1]         [,2]         [,3]         [,4]        [,5]\nUNK         0.03696522  0.040100370  0.005650103 -0.015125941 -0.01914053\ntext        0.02520180  0.030634869 -0.003202755 -0.008647407  0.04315760\nto          0.00604762  0.010533611  0.035768602 -0.027628880  0.01015684\nthe         0.01279589 -0.048598755  0.044444822  0.046858221  0.01908986\nadditional -0.02655490 -0.005862723 -0.002875999 -0.022304043  0.01380521\n<\/code><\/pre>\n\n<p>How can I <code>average<\/code> the word embeddings from this stage?<\/p>\n\n<p>Code &amp; data generation:<\/p>\n\n<pre><code># Generate some sample data\n\ntext &lt;- c(\"Because I could not stop for Death I add additional text-\",\n          \"He kindly stopped for me some additional text to act as a filler -\",\n          \"The Carriage held but just Ourselves more additional text to add to the body of the text-\",\n          \"and Immortality plus some more words to fill the text a little\")\n\nID &lt;- c(1,2,3,4)\noutput &lt;- c(1,0,0,1)\ndf &lt;- data.frame(cbind(ID, text, output))\ndf$text &lt;- as.character(df$text)\n\n# Initialise keras parameters\nlibrary(keras)\ntokenizer &lt;- text_tokenizer(num_words = 20)\ntokenizer %&gt;% \n  fit_text_tokenizer(df$text)\n\n\nembedding_size &lt;- 128  # Dimension of the embedding vector.\nskip_window &lt;- 5       # How many words to consider left and right.\nnum_sampled &lt;- 1       # Number of negative examples to sample for each word.\n\ninput_target &lt;- layer_input(shape = 1)\ninput_context &lt;- layer_input(shape = 1)\n\nembedding &lt;- layer_embedding(\n  input_dim = tokenizer$num_words + 1, \n  output_dim = embedding_size, \n  input_length = 1, \n  name = \"embedding\"\n)\n\ntarget_vector &lt;- input_target %&gt;% \n  embedding() %&gt;% \n  layer_flatten()\n\ncontext_vector &lt;- input_context %&gt;%\n  embedding() %&gt;%\n  layer_flatten()\n\ndot_product &lt;- layer_dot(list(target_vector, context_vector), axes = 1)\noutput &lt;- layer_dense(dot_product, units = 1, activation = \"sigmoid\")\n\nmodel &lt;- keras_model(list(input_target, input_context), output)\nmodel %&gt;% compile(loss = \"binary_crossentropy\", optimizer = \"adam\")\n\nsummary(model)\n\n\n\n######################## Broken skigrams_generator, have to define our own #######################\n\nskipgrams_generator &lt;- function(text, tokenizer, window_size, negative_samples) {\n  gen &lt;- texts_to_sequences_generator(tokenizer, sample(text))\n  function() {\n\n    while(TRUE) {\n      nxt &lt;- generator_next(gen)\n      if (length(nxt) &gt; 1)\n        break\n    }\n\n    skip &lt;- nxt %&gt;%\n      skipgrams(\n        vocabulary_size = tokenizer$num_words, \n        window_size = window_size, \n        negative_samples = 1\n      )\n    x &lt;- transpose(skip$couples) %&gt;% map(. %&gt;% unlist %&gt;% as.matrix(ncol = 1))\n    y &lt;- skip$labels %&gt;% as.matrix(ncol = 1)\n\n    list(x, y)\n  }\n}\n\n########################################################\n\n# Run the word2vec model\nmodel %&gt;%\n  fit_generator(\n    skipgrams_generator(df$text, tokenizer, skip_window, negative_samples), \n    steps_per_epoch = 2, epochs = 2\n  )\n\nlibrary(dplyr)\n\n# Get the word embeddings\nembedding_matrix &lt;- get_weights(model)[[1]]\n\nwords &lt;- data_frame(\n  word = names(tokenizer$word_index), \n  id = as.integer(unlist(tokenizer$word_index))\n)\n\nwords &lt;- words %&gt;%\n  filter(id &lt;= tokenizer$num_words) %&gt;%\n  arrange(id)\n\nrow.names(embedding_matrix) &lt;- c(\"UNK\", words$word)\n<\/code><\/pre>\n\n<p>I have followed a different tutorial which uses the Word2Vec model from the <code>h2o<\/code> package. The <code>h2o<\/code> packages allows me to <code>average<\/code> the documents word embeddings using <code>h2o.transform(w2v.model, words, aggregate_method = \"AVERAGE\")<\/code>. I would like to know what is the best approach for me to do this once I get the word embeddings from the <code>keras<\/code> model, as I am a little unfamiliar with the <code>h2o<\/code> package and <code>h2o environments<\/code><\/p>\n\n<pre><code>text &lt;- c(\"Because I could not stop for Death I add additional text-\",\n          \"He kindly stopped for me some additional text to act as a filler -\",\n          \"The Carriage held but just Ourselves more additional text to add to the body of the text-\",\n          \"and Immortality plus some more words to fill the text a little\")\n\nID &lt;- c(1,2,3,4)\noutput &lt;- c(1,0,0,1)\ndf &lt;- data.frame(cbind(ID, text, output))\ndf$text &lt;- as.character(df$text)\n\nlibrary(h2o)\n\nh2o.init()\ndf &lt;- as.h2o(df)\nwords &lt;- h2o.tokenize(df$text, \"\\\\\\\\W+\")  # Tokenize the words in the text\n# apply the h2o word to vec model\nw2v.model &lt;- h2o.word2vec(words,\n                          model_id = \"w2v_model\",\n                          vec_size = 100,\n                          word_model = c(\"SkipGram\"),\n                          min_word_freq = 1,\n                          window_size = 5,\n                          init_learning_rate = 0.025,\n                          sent_sample_rate = 0,\n                          epochs = 1)\n# Avergae the word ebeddings using the h2o package\naveraged_docs &lt;- h2o.transform(w2v.model, words, aggregate_method = \"AVERAGE\")\n# join the Y variable and X variables together\nML_data &lt;- h2o.cbind(as.factor(df[\"output\"]), averaged_docs)\n\naveraged_docs &lt;- as.data.frame(averaged_docs)\n\ndim(ML_data)\n# NOTE here the code doesn't work since I give the model too few observations (it does work on larger datasets, I include it\n# here for completness)\ngbm.model &lt;- h2o.gbm(x = names(averaged_docs), y = \"output\",\n                     training_frame = ML_data[1:101], validation_frame = NULL)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57662325,"AnswerCount":0,"Tags":"<python><linux><tensorflow><memory-management><keras>","CreationDate":"2019-08-26T17:24:45.177","AcceptedAnswerId":null,"OwnerUserId":6398487.0,"Title":"Keras and Allocation of NNN exceeds 10% of system memory","Body":"<p>I use Keras to run Neural Network, and it encounters this issue:<\/p>\n\n<pre><code>W tensorflow\/core\/framework\/allocator.cc:101]:\nAllocation of 19267584 exceeds 10% of system memory.\n<\/code><\/pre>\n\n<p>I already saw some discussion about this, like changing the batch size (which works for me) but I also want to know if there is a permanent way to solve this issue, for example, I could install one extra SSD to increase the physical storage of my Linux system. <\/p>\n\n<p>I'm new to Linux and Machine Learning, so I'm not sure if this issue could be simply solved like this. Please let us know, thank you in advance.<\/p>\n\n<p>Some discussion about this allocation exceeds 10% of system memory could be found below to provide you more context:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/18736\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/18736<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/50304156\/tensorflow-allocation-memory-allocation-of-38535168-exceeds-10-of-system-memor\">Tensorflow Allocation Memory: Allocation of 38535168 exceeds 10% of system memory<\/a><\/p>\n","answers":[]}
+{"QuestionId":57662781,"AnswerCount":1,"Tags":"<python><tensorflow><pip>","CreationDate":"2019-08-26T18:03:26.987","AcceptedAnswerId":57663358.0,"OwnerUserId":930254.0,"Title":"cannot install tensorflow-text using pip despite having tensorflow 2.0.0-beta1 installed","Body":"<p>My tensorflow 2.0.0beta1 runs normally, but I cannot install tensorflow-text using the command pip install tensorflow-text (as described on the tensorflow page). I can find it using pip search tensorflow-text but I am getting an error<\/p>\n\n<pre><code>ERROR: Could not find a version that satisfies the requirement tensorflow-text (from versions: none)\n<\/code><\/pre>\n\n<p>There are no requirements for this package (i.e. a specific python version). \nI am running on windows, using conda, python 3.6.9<\/p>\n","answers":[{"AnswerId":"57663358","CreationDate":"2019-08-26T18:50:15.547","ParentId":null,"OwnerUserId":"2650249","Title":null,"Body":"<p>At the time of writing this, <code>tensorflow-text<\/code> is not available for Windows yet.<\/p>\n\n<blockquote>\n  <p>Windows is something we do wish to add. We've had some difficulties getting a working package though, which is why it is not available yet. The difference between this library and tensorflow-probability is we make use of custom ops written in c++, and building those shared libraries to work well with Tensorflow inside Windows has had issues; plus, the lengthy build times on Windows has made iterating on these issues slow. While the next beta release (this week) will not include Windows, we would like for the next release to include it.<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/text\/issues\/44#issuecomment-511615087\" rel=\"noreferrer\">Source<\/a>.<\/p>\n"}]}
+{"QuestionId":57662896,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><matrix><data-science>","CreationDate":"2019-08-26T18:13:42.753","AcceptedAnswerId":57663028.0,"OwnerUserId":7204688.0,"Title":"TensorFlow unstack with numpy","Body":"<p>I am porting code from tensorflow to numpy and i have trouble with this line of code:<\/p>\n\n<pre><code>tensor_unstack = tf.unstack(some_tensor, axis=0)\n<\/code><\/pre>\n\n<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/unstack\" rel=\"nofollow noreferrer\">tf.unstack<\/a> method is used and i was unable to find a equivalent in numpy. So my question is how would a tf.unstack be implemented when using numpy?<\/p>\n","answers":[{"AnswerId":"57663028","CreationDate":"2019-08-26T18:25:18.117","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>The <a href=\"https:\/\/stackoverflow.com\/questions\/2921847\/what-does-the-star-operator-mean\">star operator<\/a> can be used to unstack a numpy array. Here is an example:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\na = np.array([1, 2, 3])\nb = np.array([4, 5, 6])\nc = np.stack([a, b])\n*d, = c\nprint(d)\n\nc_ = tf.stack([a, b])\nd_ = tf.unstack(c_)\n\nwith tf.Session() as sess:\n    print(sess.run(d_))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57663105,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><keras><tpu>","CreationDate":"2019-08-26T18:31:58.997","AcceptedAnswerId":null,"OwnerUserId":11058721.0,"Title":"How to connect a Keras model to a TPU?","Body":"<p>I was given a cloud TPU server to work on but not sure how to train my model on it.\nI have a jupyter notebook file on the server but when I run:<\/p>\n\n<pre><code>try:\n    tpu = tf.distribute.cluster_resolver.TPUClusterResolver() # TPU detection\nexcept ValueError:\n    tpu = None\n    gpus = tf.config.experimental.list_logical_devices(\"GPU\")\n\n# Select appropriate distribution strategy\nif tpu:\n    tf.tpu.experimental.initialize_tpu_system(tpu)\n    strategy = tf.distribute.experimental.TPUStrategy(tpu, steps_per_run=128) # Going back and forth between TPU and host is expensive. Better to run 128 batches on the TPU before reporting back.\n    print('Running on TPU ', tpu.cluster_spec().as_dict()['worker'])  \nelif len(gpus) &gt; 1:\n    strategy = tf.distribute.MirroredStrategy([gpu.name for gpu in gpus])\n    print('Running on multiple GPUs ', [gpu.name for gpu in gpus])\nelif len(gpus) == 1:\n    strategy = tf.distribute.get_strategy() # default strategy that works on CPU and single GPU\n    print('Running on single GPU ', gpus[0].name)\nelse:\n    strategy = tf.distribute.get_strategy() # default strategy that works on CPU and single GPU\n    print('Running on CPU')\nprint(\"Number of accelerators: \", strategy.num_replicas_in_sync)\n<\/code><\/pre>\n\n<p>The output is:\n<code>Running on CPU Number of accelerators:  1<\/code><\/p>\n\n<p>Yet, when I run <code>tf.tpu.core(0)<\/code> the output is: <code>device:TPU_REPLICATED_CORE:0<\/code><\/p>\n\n<p>I'm certain that this server has a TPU, but again, not sure how to train my model on it.<br>\nI have a keras <code>model<\/code> that is only missing the <code>model.fit<\/code> on the data.<\/p>\n\n<p>This is my model script by the way:<\/p>\n\n<pre><code>base_model = keras.applications.ResNet50(input_shape=(IMAGE_SIZE,IMAGE_SIZE,3),\n                                    include_top=False, #Set to false to train \n                                    weights='imagenet')\nbase_model.trainable = True\n\nmodel = keras.Sequential([\n  base_model,\n  keras.layers.GlobalAveragePooling2D(),\n  keras.layers.Dense(2, activation='sigmoid')\n])\n\n# Print out model summary\nmodel.summary()\n<\/code><\/pre>\n\n<p>I'm stuck after this stage.<\/p>\n","answers":[]}
+{"QuestionId":57663128,"AnswerCount":0,"Tags":"<python><tensorflow><memory><gpu>","CreationDate":"2019-08-26T18:33:20.753","AcceptedAnswerId":null,"OwnerUserId":9567257.0,"Title":"How to fix an ever-growing graph in TensorFlow causing memory leaks?","Body":"<p>I have a model written in Python using TensorFlow which I use iteratively to generate images. However, the more images I generate, the longer each image takes to generate. I am fairly certain the problem is due to a ever-growing computational graph in TensorFlow, however I don't know where in my code this leak is happening. <\/p>\n\n<p>I've already tried to set up all of my input variables with tensorflow placeholders:<\/p>\n\n<pre><code># latent vector\nself.z = tf.placeholder(tf.float32, [self.batch_size, self.z_dim], name='z')\n\n# inputs to cppn\nself.x = tf.placeholder(tf.float32, [self.batch_size, None, 1], name='x')\nself.y = tf.placeholder(tf.float32, [self.batch_size, None, 1], name='y')\nself.r = tf.placeholder(tf.float32, [self.batch_size, None, 1], name='r')\n\n# extra inputs\nself.f = [tf.placeholder(tf.float32, [self.batch_size, None, 1], name='f_' + str(i))\n          for i in range(0, len(self.f_vec))]\n<\/code><\/pre>\n\n<p>which when I run my generator to make an image I then feed with a dictionary :<\/p>\n\n<pre><code>G = self.generator(x_dim=x_dim, y_dim=y_dim, reuse=True)\nimage = self.sess.run(G, feed_dict=dict_data, options=run_opts)\n<\/code><\/pre>\n\n<p>The problem I believe is somewhere within the generator \"G\" and\/or the operations that take place inside the generator. The generator is:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def generator(self, x_dim=512, y_dim=512, reuse=False):\n    with tf.variable_scope(\"generator\") as scope:\n\n        if reuse:\n           scope.reuse_variables()\n\n        net_size = self.net_size\n        if np.size(net_size) &lt; 2:\n            net_size = np.tile(net_size, self.num_layers)\n        n_points = x_dim * y_dim\n\n        # unrolling inputs\n        z_scaled = tf.reshape(self.z, [self.batch_size, 1, self.z_dim]) * \\\n                   tf.ones([n_points, 1], dtype=tf.float32) * self.scale\n        z_unroll = tf.reshape(z_scaled, [self.batch_size * n_points, self.z_dim])\n        x_unroll = tf.reshape(self.x, [self.batch_size * n_points, 1])\n        y_unroll = tf.reshape(self.y, [self.batch_size * n_points, 1])\n        r_unroll = tf.reshape(self.r, [self.batch_size * n_points, 1])\n\n        f_unroll = [tf.reshape(self.f[0], [self.batch_size * n_points, 1])]\n        for i in range(1, len(self.f)):\n            f_unroll.append(tf.reshape(self.f[i], [self.batch_size * n_points, 1]))\n\n        U = fully_connected(z_unroll, net_size[0], 'g_0_z') + \\\n            fully_connected(x_unroll, net_size[0], 'g_0_x', with_bias=False) + \\\n            fully_connected(y_unroll, net_size[0], 'g_0_y', with_bias=False) + \\\n            fully_connected(r_unroll, net_size[0], 'g_0_r', with_bias=False)\n\n        for i in range(0, len(f_unroll)):\n            U += fully_connected(f_unroll[i], net_size[0], 'g_0_f' + str(i), with_bias=False)\n\n        H = tf.nn.tanh(U)\n        for i in range(1, self.num_layers):\n            H = tf.nn.tanh(tf.math.pow(\n                fully_connected(H, net_size[i], 'g_tanh_' + str(i), stddev=1,\n                                clip=True, clip_min=-0.5, clip_max=0.5),\n                3))\n\n        H  = tf.reshape(H, [self.batch_size * n_points, - 1])\n        output = tf.sigmoid(fully_connected(H, self.c_dim, 'g_final'))\n        result = tf.reshape(output, [self.batch_size, y_dim, x_dim, self.c_dim])\n\n    return result\n\n<\/code><\/pre>\n\n<p>where the operation for generating a fully connected layer is defined as:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def fully_connected(input_, output_size, scope=None, stddev=1.0, with_bias=True,\n                    clip=False, clip_min=None, clip_max=None):\n    shape = input_.get_shape().as_list()\n\n    with tf.variable_scope(scope or \"FC\"):\n        matrix = tf.get_variable(\"Matrix\", [shape[1], output_size], tf.float32, random_normal_initializer(stddev=stddev))\n\n        if clip:\n            matrix = tf.clip_by_value(matrix, clip_min, clip_max)\n\n        result = tf.matmul(input_, matrix)\n\n        if with_bias:\n            bias = tf.get_variable(\"bias\", [1, output_size],\n                initializer=tf.random_normal_initializer(stddev=stddev))\n            result += bias*tf.ones([shape[0], 1], dtype=tf.float32)\n\n    return result\n<\/code><\/pre>\n\n<p>After generating many images on the same network and running the following lines to see what my graph looks like, I get the following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>for op in tf.get_default_graph().get_operations():\n    print(str(op.name))\n<\/code><\/pre>\n\n<pre><code>Placeholder\nz\nx\ny\nr\nf_0\nf_1\nf_2\ngenerator\/Reshape\/shape\ngenerator\/Reshape\ngenerator\/ones\/shape_as_tensor\ngenerator\/ones\/Const\ngenerator\/ones\ngenerator\/mul\ngenerator\/mul_1\/y\ngenerator\/mul_1\ngenerator\/Reshape_1\/shape\ngenerator\/Reshape_1\ngenerator\/Reshape_2\/shape\ngenerator\/Reshape_2\ngenerator\/Reshape_3\/shape\ngenerator\/Reshape_3\ngenerator\/Reshape_4\/shape\ngenerator\/Reshape_4\ngenerator\/Reshape_5\/shape\ngenerator\/Reshape_5\ngenerator\/Reshape_6\/shape\ngenerator\/Reshape_6\ngenerator\/Reshape_7\/shape\ngenerator\/Reshape_7\ngenerator\/g_0_z\/Matrix\/Initializer\/random_normal\/shape\ngenerator\/g_0_z\/Matrix\/Initializer\/random_normal\/mean\ngenerator\/g_0_z\/Matrix\/Initializer\/random_normal\/stddev\ngenerator\/g_0_z\/Matrix\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_0_z\/Matrix\/Initializer\/random_normal\/mul\ngenerator\/g_0_z\/Matrix\/Initializer\/random_normal\ngenerator\/g_0_z\/Matrix\ngenerator\/g_0_z\/Matrix\/Assign\ngenerator\/g_0_z\/Matrix\/read\ngenerator\/g_0_z\/MatMul\ngenerator\/g_0_z\/bias\/Initializer\/random_normal\/shape\ngenerator\/g_0_z\/bias\/Initializer\/random_normal\/mean\ngenerator\/g_0_z\/bias\/Initializer\/random_normal\/stddev\ngenerator\/g_0_z\/bias\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_0_z\/bias\/Initializer\/random_normal\/mul\ngenerator\/g_0_z\/bias\/Initializer\/random_normal\ngenerator\/g_0_z\/bias\ngenerator\/g_0_z\/bias\/Assign\ngenerator\/g_0_z\/bias\/read\ngenerator\/g_0_z\/ones\/shape_as_tensor\ngenerator\/g_0_z\/ones\/Const\ngenerator\/g_0_z\/ones\ngenerator\/g_0_z\/mul\ngenerator\/g_0_z\/add\ngenerator\/g_0_x\/Matrix\/Initializer\/random_normal\/shape\ngenerator\/g_0_x\/Matrix\/Initializer\/random_normal\/mean\ngenerator\/g_0_x\/Matrix\/Initializer\/random_normal\/stddev\ngenerator\/g_0_x\/Matrix\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_0_x\/Matrix\/Initializer\/random_normal\/mul\ngenerator\/g_0_x\/Matrix\/Initializer\/random_normal\ngenerator\/g_0_x\/Matrix\ngenerator\/g_0_x\/Matrix\/Assign\ngenerator\/g_0_x\/Matrix\/read\ngenerator\/g_0_x\/MatMul\ngenerator\/add\ngenerator\/g_0_y\/Matrix\/Initializer\/random_normal\/shape\ngenerator\/g_0_y\/Matrix\/Initializer\/random_normal\/mean\ngenerator\/g_0_y\/Matrix\/Initializer\/random_normal\/stddev\ngenerator\/g_0_y\/Matrix\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_0_y\/Matrix\/Initializer\/random_normal\/mul\ngenerator\/g_0_y\/Matrix\/Initializer\/random_normal\ngenerator\/g_0_y\/Matrix\ngenerator\/g_0_y\/Matrix\/Assign\ngenerator\/g_0_y\/Matrix\/read\ngenerator\/g_0_y\/MatMul\ngenerator\/add_1\ngenerator\/g_0_r\/Matrix\/Initializer\/random_normal\/shape\ngenerator\/g_0_r\/Matrix\/Initializer\/random_normal\/mean\ngenerator\/g_0_r\/Matrix\/Initializer\/random_normal\/stddev\ngenerator\/g_0_r\/Matrix\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_0_r\/Matrix\/Initializer\/random_normal\/mul\ngenerator\/g_0_r\/Matrix\/Initializer\/random_normal\ngenerator\/g_0_r\/Matrix\ngenerator\/g_0_r\/Matrix\/Assign\ngenerator\/g_0_r\/Matrix\/read\ngenerator\/g_0_r\/MatMul\ngenerator\/add_2\ngenerator\/g_0_f0\/Matrix\/Initializer\/random_normal\/shape\ngenerator\/g_0_f0\/Matrix\/Initializer\/random_normal\/mean\ngenerator\/g_0_f0\/Matrix\/Initializer\/random_normal\/stddev\ngenerator\/g_0_f0\/Matrix\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_0_f0\/Matrix\/Initializer\/random_normal\/mul\ngenerator\/g_0_f0\/Matrix\/Initializer\/random_normal\ngenerator\/g_0_f0\/Matrix\ngenerator\/g_0_f0\/Matrix\/Assign\ngenerator\/g_0_f0\/Matrix\/read\ngenerator\/g_0_f0\/MatMul\ngenerator\/add_3\ngenerator\/g_0_f1\/Matrix\/Initializer\/random_normal\/shape\ngenerator\/g_0_f1\/Matrix\/Initializer\/random_normal\/mean\ngenerator\/g_0_f1\/Matrix\/Initializer\/random_normal\/stddev\ngenerator\/g_0_f1\/Matrix\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_0_f1\/Matrix\/Initializer\/random_normal\/mul\ngenerator\/g_0_f1\/Matrix\/Initializer\/random_normal\ngenerator\/g_0_f1\/Matrix\ngenerator\/g_0_f1\/Matrix\/Assign\ngenerator\/g_0_f1\/Matrix\/read\ngenerator\/g_0_f1\/MatMul\ngenerator\/add_4\ngenerator\/g_0_f2\/Matrix\/Initializer\/random_normal\/shape\ngenerator\/g_0_f2\/Matrix\/Initializer\/random_normal\/mean\ngenerator\/g_0_f2\/Matrix\/Initializer\/random_normal\/stddev\ngenerator\/g_0_f2\/Matrix\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_0_f2\/Matrix\/Initializer\/random_normal\/mul\ngenerator\/g_0_f2\/Matrix\/Initializer\/random_normal\ngenerator\/g_0_f2\/Matrix\ngenerator\/g_0_f2\/Matrix\/Assign\ngenerator\/g_0_f2\/Matrix\/read\ngenerator\/g_0_f2\/MatMul\ngenerator\/add_5\ngenerator\/Tanh\ngenerator\/g_tanh_1\/Matrix\/Initializer\/random_normal\/shape\ngenerator\/g_tanh_1\/Matrix\/Initializer\/random_normal\/mean\ngenerator\/g_tanh_1\/Matrix\/Initializer\/random_normal\/stddev\ngenerator\/g_tanh_1\/Matrix\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_tanh_1\/Matrix\/Initializer\/random_normal\/mul\ngenerator\/g_tanh_1\/Matrix\/Initializer\/random_normal\ngenerator\/g_tanh_1\/Matrix\ngenerator\/g_tanh_1\/Matrix\/Assign\ngenerator\/g_tanh_1\/Matrix\/read\ngenerator\/g_tanh_1\/clip_by_value\/Minimum\/y\ngenerator\/g_tanh_1\/clip_by_value\/Minimum\ngenerator\/g_tanh_1\/clip_by_value\/y\ngenerator\/g_tanh_1\/clip_by_value\ngenerator\/g_tanh_1\/MatMul\ngenerator\/g_tanh_1\/bias\/Initializer\/random_normal\/shape\ngenerator\/g_tanh_1\/bias\/Initializer\/random_normal\/mean\ngenerator\/g_tanh_1\/bias\/Initializer\/random_normal\/stddev\ngenerator\/g_tanh_1\/bias\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_tanh_1\/bias\/Initializer\/random_normal\/mul\ngenerator\/g_tanh_1\/bias\/Initializer\/random_normal\ngenerator\/g_tanh_1\/bias\ngenerator\/g_tanh_1\/bias\/Assign\ngenerator\/g_tanh_1\/bias\/read\ngenerator\/g_tanh_1\/ones\/shape_as_tensor\ngenerator\/g_tanh_1\/ones\/Const\ngenerator\/g_tanh_1\/ones\ngenerator\/g_tanh_1\/mul\ngenerator\/g_tanh_1\/add\ngenerator\/Pow\/y\ngenerator\/Pow\ngenerator\/Tanh_1\ngenerator\/g_tanh_2\/Matrix\/Initializer\/random_normal\/shape\ngenerator\/g_tanh_2\/Matrix\/Initializer\/random_normal\/mean\ngenerator\/g_tanh_2\/Matrix\/Initializer\/random_normal\/stddev\ngenerator\/g_tanh_2\/Matrix\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_tanh_2\/Matrix\/Initializer\/random_normal\/mul\ngenerator\/g_tanh_2\/Matrix\/Initializer\/random_normal\ngenerator\/g_tanh_2\/Matrix\ngenerator\/g_tanh_2\/Matrix\/Assign\ngenerator\/g_tanh_2\/Matrix\/read\ngenerator\/g_tanh_2\/clip_by_value\/Minimum\/y\ngenerator\/g_tanh_2\/clip_by_value\/Minimum\ngenerator\/g_tanh_2\/clip_by_value\/y\ngenerator\/g_tanh_2\/clip_by_value\ngenerator\/g_tanh_2\/MatMul\ngenerator\/g_tanh_2\/bias\/Initializer\/random_normal\/shape\ngenerator\/g_tanh_2\/bias\/Initializer\/random_normal\/mean\ngenerator\/g_tanh_2\/bias\/Initializer\/random_normal\/stddev\ngenerator\/g_tanh_2\/bias\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_tanh_2\/bias\/Initializer\/random_normal\/mul\ngenerator\/g_tanh_2\/bias\/Initializer\/random_normal\ngenerator\/g_tanh_2\/bias\ngenerator\/g_tanh_2\/bias\/Assign\ngenerator\/g_tanh_2\/bias\/read\ngenerator\/g_tanh_2\/ones\/shape_as_tensor\ngenerator\/g_tanh_2\/ones\/Const\ngenerator\/g_tanh_2\/ones\ngenerator\/g_tanh_2\/mul\ngenerator\/g_tanh_2\/add\ngenerator\/Pow_1\/y\ngenerator\/Pow_1\ngenerator\/Tanh_2\ngenerator\/g_tanh_3\/Matrix\/Initializer\/random_normal\/shape\ngenerator\/g_tanh_3\/Matrix\/Initializer\/random_normal\/mean\ngenerator\/g_tanh_3\/Matrix\/Initializer\/random_normal\/stddev\ngenerator\/g_tanh_3\/Matrix\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_tanh_3\/Matrix\/Initializer\/random_normal\/mul\ngenerator\/g_tanh_3\/Matrix\/Initializer\/random_normal\ngenerator\/g_tanh_3\/Matrix\ngenerator\/g_tanh_3\/Matrix\/Assign\ngenerator\/g_tanh_3\/Matrix\/read\ngenerator\/g_tanh_3\/clip_by_value\/Minimum\/y\ngenerator\/g_tanh_3\/clip_by_value\/Minimum\ngenerator\/g_tanh_3\/clip_by_value\/y\ngenerator\/g_tanh_3\/clip_by_value\ngenerator\/g_tanh_3\/MatMul\ngenerator\/g_tanh_3\/bias\/Initializer\/random_normal\/shape\ngenerator\/g_tanh_3\/bias\/Initializer\/random_normal\/mean\ngenerator\/g_tanh_3\/bias\/Initializer\/random_normal\/stddev\ngenerator\/g_tanh_3\/bias\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_tanh_3\/bias\/Initializer\/random_normal\/mul\ngenerator\/g_tanh_3\/bias\/Initializer\/random_normal\ngenerator\/g_tanh_3\/bias\ngenerator\/g_tanh_3\/bias\/Assign\ngenerator\/g_tanh_3\/bias\/read\ngenerator\/g_tanh_3\/ones\/shape_as_tensor\ngenerator\/g_tanh_3\/ones\/Const\ngenerator\/g_tanh_3\/ones\ngenerator\/g_tanh_3\/mul\ngenerator\/g_tanh_3\/add\ngenerator\/Pow_2\/y\ngenerator\/Pow_2\ngenerator\/Tanh_3\ngenerator\/g_tanh_4\/Matrix\/Initializer\/random_normal\/shape\ngenerator\/g_tanh_4\/Matrix\/Initializer\/random_normal\/mean\ngenerator\/g_tanh_4\/Matrix\/Initializer\/random_normal\/stddev\ngenerator\/g_tanh_4\/Matrix\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_tanh_4\/Matrix\/Initializer\/random_normal\/mul\ngenerator\/g_tanh_4\/Matrix\/Initializer\/random_normal\ngenerator\/g_tanh_4\/Matrix\ngenerator\/g_tanh_4\/Matrix\/Assign\ngenerator\/g_tanh_4\/Matrix\/read\ngenerator\/g_tanh_4\/clip_by_value\/Minimum\/y\ngenerator\/g_tanh_4\/clip_by_value\/Minimum\ngenerator\/g_tanh_4\/clip_by_value\/y\ngenerator\/g_tanh_4\/clip_by_value\ngenerator\/g_tanh_4\/MatMul\ngenerator\/g_tanh_4\/bias\/Initializer\/random_normal\/shape\ngenerator\/g_tanh_4\/bias\/Initializer\/random_normal\/mean\ngenerator\/g_tanh_4\/bias\/Initializer\/random_normal\/stddev\ngenerator\/g_tanh_4\/bias\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_tanh_4\/bias\/Initializer\/random_normal\/mul\ngenerator\/g_tanh_4\/bias\/Initializer\/random_normal\ngenerator\/g_tanh_4\/bias\ngenerator\/g_tanh_4\/bias\/Assign\ngenerator\/g_tanh_4\/bias\/read\ngenerator\/g_tanh_4\/ones\/shape_as_tensor\ngenerator\/g_tanh_4\/ones\/Const\ngenerator\/g_tanh_4\/ones\ngenerator\/g_tanh_4\/mul\ngenerator\/g_tanh_4\/add\ngenerator\/Pow_3\/y\ngenerator\/Pow_3\ngenerator\/Tanh_4\ngenerator\/g_tanh_5\/Matrix\/Initializer\/random_normal\/shape\ngenerator\/g_tanh_5\/Matrix\/Initializer\/random_normal\/mean\ngenerator\/g_tanh_5\/Matrix\/Initializer\/random_normal\/stddev\ngenerator\/g_tanh_5\/Matrix\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_tanh_5\/Matrix\/Initializer\/random_normal\/mul\ngenerator\/g_tanh_5\/Matrix\/Initializer\/random_normal\ngenerator\/g_tanh_5\/Matrix\ngenerator\/g_tanh_5\/Matrix\/Assign\ngenerator\/g_tanh_5\/Matrix\/read\ngenerator\/g_tanh_5\/clip_by_value\/Minimum\/y\ngenerator\/g_tanh_5\/clip_by_value\/Minimum\ngenerator\/g_tanh_5\/clip_by_value\/y\ngenerator\/g_tanh_5\/clip_by_value\ngenerator\/g_tanh_5\/MatMul\ngenerator\/g_tanh_5\/bias\/Initializer\/random_normal\/shape\ngenerator\/g_tanh_5\/bias\/Initializer\/random_normal\/mean\ngenerator\/g_tanh_5\/bias\/Initializer\/random_normal\/stddev\ngenerator\/g_tanh_5\/bias\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_tanh_5\/bias\/Initializer\/random_normal\/mul\ngenerator\/g_tanh_5\/bias\/Initializer\/random_normal\ngenerator\/g_tanh_5\/bias\ngenerator\/g_tanh_5\/bias\/Assign\ngenerator\/g_tanh_5\/bias\/read\ngenerator\/g_tanh_5\/ones\/shape_as_tensor\ngenerator\/g_tanh_5\/ones\/Const\ngenerator\/g_tanh_5\/ones\ngenerator\/g_tanh_5\/mul\ngenerator\/g_tanh_5\/add\ngenerator\/Pow_4\/y\ngenerator\/Pow_4\ngenerator\/Tanh_5\ngenerator\/g_tanh_6\/Matrix\/Initializer\/random_normal\/shape\ngenerator\/g_tanh_6\/Matrix\/Initializer\/random_normal\/mean\ngenerator\/g_tanh_6\/Matrix\/Initializer\/random_normal\/stddev\ngenerator\/g_tanh_6\/Matrix\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_tanh_6\/Matrix\/Initializer\/random_normal\/mul\ngenerator\/g_tanh_6\/Matrix\/Initializer\/random_normal\ngenerator\/g_tanh_6\/Matrix\ngenerator\/g_tanh_6\/Matrix\/Assign\ngenerator\/g_tanh_6\/Matrix\/read\ngenerator\/g_tanh_6\/clip_by_value\/Minimum\/y\ngenerator\/g_tanh_6\/clip_by_value\/Minimum\ngenerator\/g_tanh_6\/clip_by_value\/y\ngenerator\/g_tanh_6\/clip_by_value\ngenerator\/g_tanh_6\/MatMul\ngenerator\/g_tanh_6\/bias\/Initializer\/random_normal\/shape\ngenerator\/g_tanh_6\/bias\/Initializer\/random_normal\/mean\ngenerator\/g_tanh_6\/bias\/Initializer\/random_normal\/stddev\ngenerator\/g_tanh_6\/bias\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_tanh_6\/bias\/Initializer\/random_normal\/mul\ngenerator\/g_tanh_6\/bias\/Initializer\/random_normal\ngenerator\/g_tanh_6\/bias\ngenerator\/g_tanh_6\/bias\/Assign\ngenerator\/g_tanh_6\/bias\/read\ngenerator\/g_tanh_6\/ones\/shape_as_tensor\ngenerator\/g_tanh_6\/ones\/Const\ngenerator\/g_tanh_6\/ones\ngenerator\/g_tanh_6\/mul\ngenerator\/g_tanh_6\/add\ngenerator\/Pow_5\/y\ngenerator\/Pow_5\ngenerator\/Tanh_6\ngenerator\/Reshape_8\/shape\ngenerator\/Reshape_8\ngenerator\/g_final\/Matrix\/Initializer\/random_normal\/shape\ngenerator\/g_final\/Matrix\/Initializer\/random_normal\/mean\ngenerator\/g_final\/Matrix\/Initializer\/random_normal\/stddev\ngenerator\/g_final\/Matrix\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_final\/Matrix\/Initializer\/random_normal\/mul\ngenerator\/g_final\/Matrix\/Initializer\/random_normal\ngenerator\/g_final\/Matrix\ngenerator\/g_final\/Matrix\/Assign\ngenerator\/g_final\/Matrix\/read\ngenerator\/g_final\/MatMul\ngenerator\/g_final\/bias\/Initializer\/random_normal\/shape\ngenerator\/g_final\/bias\/Initializer\/random_normal\/mean\ngenerator\/g_final\/bias\/Initializer\/random_normal\/stddev\ngenerator\/g_final\/bias\/Initializer\/random_normal\/RandomStandardNormal\ngenerator\/g_final\/bias\/Initializer\/random_normal\/mul\ngenerator\/g_final\/bias\/Initializer\/random_normal\ngenerator\/g_final\/bias\ngenerator\/g_final\/bias\/Assign\ngenerator\/g_final\/bias\/read\ngenerator\/g_final\/ones\/shape_as_tensor\ngenerator\/g_final\/ones\/Const\ngenerator\/g_final\/ones\ngenerator\/g_final\/mul\ngenerator\/g_final\/add\ngenerator\/Sigmoid\ngenerator\/Reshape_9\/shape\ngenerator\/Reshape_9\ninit\ninit_1\ngenerator_1\/Reshape\/shape\ngenerator_1\/Reshape\ngenerator_1\/ones\/shape_as_tensor\ngenerator_1\/ones\/Const\ngenerator_1\/ones\ngenerator_1\/mul\ngenerator_1\/mul_1\/y\ngenerator_1\/mul_1\ngenerator_1\/Reshape_1\/shape\ngenerator_1\/Reshape_1\ngenerator_1\/Reshape_2\/shape\ngenerator_1\/Reshape_2\ngenerator_1\/Reshape_3\/shape\ngenerator_1\/Reshape_3\ngenerator_1\/Reshape_4\/shape\ngenerator_1\/Reshape_4\ngenerator_1\/Reshape_5\/shape\ngenerator_1\/Reshape_5\ngenerator_1\/Reshape_6\/shape\ngenerator_1\/Reshape_6\ngenerator_1\/Reshape_7\/shape\ngenerator_1\/Reshape_7\ngenerator_1\/g_0_z\/MatMul\ngenerator_1\/g_0_z\/ones\/shape_as_tensor\ngenerator_1\/g_0_z\/ones\/Const\ngenerator_1\/g_0_z\/ones\ngenerator_1\/g_0_z\/mul\ngenerator_1\/g_0_z\/add\ngenerator_1\/g_0_x\/MatMul\ngenerator_1\/add\ngenerator_1\/g_0_y\/MatMul\ngenerator_1\/add_1\ngenerator_1\/g_0_r\/MatMul\ngenerator_1\/add_2\ngenerator_1\/g_0_f0\/MatMul\ngenerator_1\/add_3\ngenerator_1\/g_0_f1\/MatMul\ngenerator_1\/add_4\ngenerator_1\/g_0_f2\/MatMul\ngenerator_1\/add_5\ngenerator_1\/Tanh\ngenerator_1\/g_tanh_1\/clip_by_value\/Minimum\/y\ngenerator_1\/g_tanh_1\/clip_by_value\/Minimum\ngenerator_1\/g_tanh_1\/clip_by_value\/y\ngenerator_1\/g_tanh_1\/clip_by_value\ngenerator_1\/g_tanh_1\/MatMul\ngenerator_1\/g_tanh_1\/ones\/shape_as_tensor\ngenerator_1\/g_tanh_1\/ones\/Const\ngenerator_1\/g_tanh_1\/ones\ngenerator_1\/g_tanh_1\/mul\ngenerator_1\/g_tanh_1\/add\ngenerator_1\/Pow\/y\ngenerator_1\/Pow\ngenerator_1\/Tanh_1\ngenerator_1\/g_tanh_2\/clip_by_value\/Minimum\/y\ngenerator_1\/g_tanh_2\/clip_by_value\/Minimum\ngenerator_1\/g_tanh_2\/clip_by_value\/y\ngenerator_1\/g_tanh_2\/clip_by_value\ngenerator_1\/g_tanh_2\/MatMul\ngenerator_1\/g_tanh_2\/ones\/shape_as_tensor\ngenerator_1\/g_tanh_2\/ones\/Const\ngenerator_1\/g_tanh_2\/ones\ngenerator_1\/g_tanh_2\/mul\ngenerator_1\/g_tanh_2\/add\ngenerator_1\/Pow_1\/y\ngenerator_1\/Pow_1\ngenerator_1\/Tanh_2\ngenerator_1\/g_tanh_3\/clip_by_value\/Minimum\/y\ngenerator_1\/g_tanh_3\/clip_by_value\/Minimum\ngenerator_1\/g_tanh_3\/clip_by_value\/y\ngenerator_1\/g_tanh_3\/clip_by_value\ngenerator_1\/g_tanh_3\/MatMul\ngenerator_1\/g_tanh_3\/ones\/shape_as_tensor\ngenerator_1\/g_tanh_3\/ones\/Const\ngenerator_1\/g_tanh_3\/ones\ngenerator_1\/g_tanh_3\/mul\ngenerator_1\/g_tanh_3\/add\ngenerator_1\/Pow_2\/y\ngenerator_1\/Pow_2\ngenerator_1\/Tanh_3\ngenerator_1\/g_tanh_4\/clip_by_value\/Minimum\/y\ngenerator_1\/g_tanh_4\/clip_by_value\/Minimum\ngenerator_1\/g_tanh_4\/clip_by_value\/y\ngenerator_1\/g_tanh_4\/clip_by_value\ngenerator_1\/g_tanh_4\/MatMul\ngenerator_1\/g_tanh_4\/ones\/shape_as_tensor\ngenerator_1\/g_tanh_4\/ones\/Const\ngenerator_1\/g_tanh_4\/ones\ngenerator_1\/g_tanh_4\/mul\ngenerator_1\/g_tanh_4\/add\ngenerator_1\/Pow_3\/y\ngenerator_1\/Pow_3\ngenerator_1\/Tanh_4\ngenerator_1\/g_tanh_5\/clip_by_value\/Minimum\/y\ngenerator_1\/g_tanh_5\/clip_by_value\/Minimum\ngenerator_1\/g_tanh_5\/clip_by_value\/y\ngenerator_1\/g_tanh_5\/clip_by_value\ngenerator_1\/g_tanh_5\/MatMul\ngenerator_1\/g_tanh_5\/ones\/shape_as_tensor\ngenerator_1\/g_tanh_5\/ones\/Const\ngenerator_1\/g_tanh_5\/ones\ngenerator_1\/g_tanh_5\/mul\ngenerator_1\/g_tanh_5\/add\ngenerator_1\/Pow_4\/y\ngenerator_1\/Pow_4\ngenerator_1\/Tanh_5\ngenerator_1\/g_tanh_6\/clip_by_value\/Minimum\/y\ngenerator_1\/g_tanh_6\/clip_by_value\/Minimum\ngenerator_1\/g_tanh_6\/clip_by_value\/y\ngenerator_1\/g_tanh_6\/clip_by_value\ngenerator_1\/g_tanh_6\/MatMul\ngenerator_1\/g_tanh_6\/ones\/shape_as_tensor\ngenerator_1\/g_tanh_6\/ones\/Const\ngenerator_1\/g_tanh_6\/ones\ngenerator_1\/g_tanh_6\/mul\ngenerator_1\/g_tanh_6\/add\ngenerator_1\/Pow_5\/y\ngenerator_1\/Pow_5\ngenerator_1\/Tanh_6\ngenerator_1\/Reshape_8\/shape\ngenerator_1\/Reshape_8\ngenerator_1\/g_final\/MatMul\ngenerator_1\/g_final\/ones\/shape_as_tensor\ngenerator_1\/g_final\/ones\/Const\ngenerator_1\/g_final\/ones\ngenerator_1\/g_final\/mul\ngenerator_1\/g_final\/add\ngenerator_1\/Sigmoid\ngenerator_1\/Reshape_9\/shape\ngenerator_1\/Reshape_9\ninit_2\ngenerator_2\/Reshape\/shape\ngenerator_2\/Reshape\ngenerator_2\/ones\/shape_as_tensor\ngenerator_2\/ones\/Const\ngenerator_2\/ones\ngenerator_2\/mul\ngenerator_2\/mul_1\/y\ngenerator_2\/mul_1\ngenerator_2\/Reshape_1\/shape\ngenerator_2\/Reshape_1\ngenerator_2\/Reshape_2\/shape\ngenerator_2\/Reshape_2\ngenerator_2\/Reshape_3\/shape\ngenerator_2\/Reshape_3\ngenerator_2\/Reshape_4\/shape\ngenerator_2\/Reshape_4\ngenerator_2\/Reshape_5\/shape\ngenerator_2\/Reshape_5\ngenerator_2\/Reshape_6\/shape\ngenerator_2\/Reshape_6\ngenerator_2\/Reshape_7\/shape\ngenerator_2\/Reshape_7\ngenerator_2\/g_0_z\/MatMul\ngenerator_2\/g_0_z\/ones\/shape_as_tensor\ngenerator_2\/g_0_z\/ones\/Const\ngenerator_2\/g_0_z\/ones\ngenerator_2\/g_0_z\/mul\ngenerator_2\/g_0_z\/add\ngenerator_2\/g_0_x\/MatMul\ngenerator_2\/add\ngenerator_2\/g_0_y\/MatMul\ngenerator_2\/add_1\ngenerator_2\/g_0_r\/MatMul\ngenerator_2\/add_2\ngenerator_2\/g_0_f0\/MatMul\ngenerator_2\/add_3\ngenerator_2\/g_0_f1\/MatMul\ngenerator_2\/add_4\ngenerator_2\/g_0_f2\/MatMul\ngenerator_2\/add_5\ngenerator_2\/Tanh\ngenerator_2\/g_tanh_1\/clip_by_value\/Minimum\/y\ngenerator_2\/g_tanh_1\/clip_by_value\/Minimum\ngenerator_2\/g_tanh_1\/clip_by_value\/y\ngenerator_2\/g_tanh_1\/clip_by_value\ngenerator_2\/g_tanh_1\/MatMul\ngenerator_2\/g_tanh_1\/ones\/shape_as_tensor\ngenerator_2\/g_tanh_1\/ones\/Const\ngenerator_2\/g_tanh_1\/ones\ngenerator_2\/g_tanh_1\/mul\ngenerator_2\/g_tanh_1\/add\ngenerator_2\/Pow\/y\ngenerator_2\/Pow\ngenerator_2\/Tanh_1\ngenerator_2\/g_tanh_2\/clip_by_value\/Minimum\/y\ngenerator_2\/g_tanh_2\/clip_by_value\/Minimum\ngenerator_2\/g_tanh_2\/clip_by_value\/y\ngenerator_2\/g_tanh_2\/clip_by_value\ngenerator_2\/g_tanh_2\/MatMul\ngenerator_2\/g_tanh_2\/ones\/shape_as_tensor\ngenerator_2\/g_tanh_2\/ones\/Const\ngenerator_2\/g_tanh_2\/ones\ngenerator_2\/g_tanh_2\/mul\ngenerator_2\/g_tanh_2\/add\ngenerator_2\/Pow_1\/y\ngenerator_2\/Pow_1\ngenerator_2\/Tanh_2\ngenerator_2\/g_tanh_3\/clip_by_value\/Minimum\/y\ngenerator_2\/g_tanh_3\/clip_by_value\/Minimum\ngenerator_2\/g_tanh_3\/clip_by_value\/y\ngenerator_2\/g_tanh_3\/clip_by_value\ngenerator_2\/g_tanh_3\/MatMul\ngenerator_2\/g_tanh_3\/ones\/shape_as_tensor\ngenerator_2\/g_tanh_3\/ones\/Const\ngenerator_2\/g_tanh_3\/ones\ngenerator_2\/g_tanh_3\/mul\ngenerator_2\/g_tanh_3\/add\ngenerator_2\/Pow_2\/y\ngenerator_2\/Pow_2\ngenerator_2\/Tanh_3\ngenerator_2\/g_tanh_4\/clip_by_value\/Minimum\/y\ngenerator_2\/g_tanh_4\/clip_by_value\/Minimum\ngenerator_2\/g_tanh_4\/clip_by_value\/y\ngenerator_2\/g_tanh_4\/clip_by_value\ngenerator_2\/g_tanh_4\/MatMul\ngenerator_2\/g_tanh_4\/ones\/shape_as_tensor\ngenerator_2\/g_tanh_4\/ones\/Const\ngenerator_2\/g_tanh_4\/ones\ngenerator_2\/g_tanh_4\/mul\ngenerator_2\/g_tanh_4\/add\ngenerator_2\/Pow_3\/y\ngenerator_2\/Pow_3\ngenerator_2\/Tanh_4\ngenerator_2\/g_tanh_5\/clip_by_value\/Minimum\/y\ngenerator_2\/g_tanh_5\/clip_by_value\/Minimum\ngenerator_2\/g_tanh_5\/clip_by_value\/y\ngenerator_2\/g_tanh_5\/clip_by_value\ngenerator_2\/g_tanh_5\/MatMul\ngenerator_2\/g_tanh_5\/ones\/shape_as_tensor\ngenerator_2\/g_tanh_5\/ones\/Const\ngenerator_2\/g_tanh_5\/ones\ngenerator_2\/g_tanh_5\/mul\ngenerator_2\/g_tanh_5\/add\ngenerator_2\/Pow_4\/y\ngenerator_2\/Pow_4\ngenerator_2\/Tanh_5\ngenerator_2\/g_tanh_6\/clip_by_value\/Minimum\/y\ngenerator_2\/g_tanh_6\/clip_by_value\/Minimum\ngenerator_2\/g_tanh_6\/clip_by_value\/y\ngenerator_2\/g_tanh_6\/clip_by_value\ngenerator_2\/g_tanh_6\/MatMul\ngenerator_2\/g_tanh_6\/ones\/shape_as_tensor\ngenerator_2\/g_tanh_6\/ones\/Const\ngenerator_2\/g_tanh_6\/ones\ngenerator_2\/g_tanh_6\/mul\ngenerator_2\/g_tanh_6\/add\ngenerator_2\/Pow_5\/y\ngenerator_2\/Pow_5\ngenerator_2\/Tanh_6\ngenerator_2\/Reshape_8\/shape\ngenerator_2\/Reshape_8\ngenerator_2\/g_final\/MatMul\ngenerator_2\/g_final\/ones\/shape_as_tensor\ngenerator_2\/g_final\/ones\/Const\ngenerator_2\/g_final\/ones\ngenerator_2\/g_final\/mul\ngenerator_2\/g_final\/add\ngenerator_2\/Sigmoid\ngenerator_2\/Reshape_9\/shape\ngenerator_2\/Reshape_9\n...\n<\/code><\/pre>\n\n<p>which goes on and on for the hundreds of iterations that I ran it for. Why are there so many generators? Shouldn't my first generator be re-used? Maybe I'm just misinterpreting the output but the fact is that each image takes a few milliseconds longer than the previous image in terms of computation time, and when I need to make a video it really starts to add up.<\/p>\n\n<p>Can someone help me decipher at what point in my generator am I adding operations to the graph that aren't being re-used? I feel like I am following most of the advice I've read about using placeholders and whatnot, but I am totally clueless how to fix my generator up. Evidently the operations in my generator are constantly adding to the graph, but I don't see how I fix this.<\/p>\n\n<p>Any advice would be greatly appreciated. I have already spend weeks looking at many other questions similar to mine and I haven't been able to apply it to my own code.<\/p>\n","answers":[]}
+{"QuestionId":57663233,"AnswerCount":1,"Tags":"<python><machine-learning><image-processing><keras><conv-neural-network>","CreationDate":"2019-08-26T18:40:53.047","AcceptedAnswerId":57663549.0,"OwnerUserId":11917598.0,"Title":"My Keras CNN return the same output value, how can I fix\/improve my code","Body":"<p>I've created a code to train my CNN which will determine if a image be one or another class (\"pdr or nonPdr\")<\/p>\n\n<p>This is my keras model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential()\nmodel.add(Conv2D(input_shape=(605,700,3), filters=64, kernel_size=(3,3), padding=\"same\",activation=\"tanh\"))\nmodel.add(MaxPooling2D((2, 2)))\nmodel.add(Flatten())\nmodel.add(Dense(32, activation='tanh', input_dim=100))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(optimizer='rmsprop', loss='binary_crossentropy', metrics=['accuracy'])\n\ndata, labels = ReadImages(TRAIN_DIR)\n\nmodel.fit(np.array(data), np.array(labels), epochs=10, batch_size=16)\n\nmodel.save('model.h5')\n<\/code><\/pre>\n\n<p>And this is my test_predict file:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = load_model('model.h5')\n\nmodel.compile(optimizer='rmsprop', loss='binary_crossentropy', metrics=['accuracy'])\n\ntest_image = cv2.imread(\"test_data\/im0203.ppm\") \ntest_image = test_image.reshape(1,605,700,3).astype('float')\ntest_image = np.array(test_image)\n\ntest_image2 = cv2.imread(\"test_data\/im0001.ppm\") \ntest_image2 = test_image2.reshape(1,605,700,3).astype('float')\n\n#predict the result\nprint(model.predict(test_image))\nprint(model.predict(test_image2))\n<\/code><\/pre>\n\n<p>After run the code below I get the same value for my 2 images (that are different, one is pdr and the other is nonPdr)<\/p>\n\n<pre><code>[[0.033681]]\n[[0.033681]]\n<\/code><\/pre>\n\n<p>How can I fix it and improve my CNN. I appreciate your help.<\/p>\n\n<p><strong>Update<\/strong>\nI tried to remove one dense layer and change the last one in a (2, activation='sigmoid') but it not works too.. I really don't have any ideia what to do**<\/p>\n","answers":[{"AnswerId":"57663549","CreationDate":"2019-08-26T19:05:57.610","ParentId":null,"OwnerUserId":"11309234","Title":null,"Body":"<p>This could be because of multiple reasons. Most plausible ones are:<\/p>\n\n<ol>\n<li><p>The input fed to network and the image fed during prediction may not have the same preprocessing steps. <\/p><\/li>\n<li><p>The model is over-fitted and when it finds a little different prediction image, it tends to bias its decision based on its training. <\/p><\/li>\n<li><p>Model is not good<\/p><\/li>\n<\/ol>\n\n<p>To avoid such complications, possible future solutions could be:\n1. Try feeding a normalized or scaled image only during training and testing. <\/p>\n\n<ol start=\"2\">\n<li>Do not use too many Dense layers (Especially for image classification) because dense layers are good at learning spatial information and for an image to classify correctly, the model needs to look the image in 2d matrix rather than 1d spatial matrix (since dense layers flatten the 2d matrix to 1d matrix).<\/li>\n<\/ol>\n\n<p>Even if you are trying to use a dense layer, I would recommend using a dense layer only at the last layer, it, for the final number_of_classes layers. \nA better architecture would be<\/p>\n\n<p>[ Conv followed by few transition blocks and bottleneck layers ] x 2\nflatten\nDense(number_of_classes)<\/p>\n"}]}
+{"QuestionId":57663688,"AnswerCount":0,"Tags":"<python><tensorflow><keras><nlp><word-embedding>","CreationDate":"2019-08-26T19:16:12.377","AcceptedAnswerId":null,"OwnerUserId":11876924.0,"Title":"Preprecessing for Embedding: transforming word tokens to integer vectors","Body":"<p>I am trying to preprocess a corpus of text for input to a word embedding layer, which takes padded vectors of integers. <\/p>\n\n<p>I know Keras\/Tensorflow already has a list of functions that can do this (e.g., <a href=\"https:\/\/keras.io\/preprocessing\/text\/#one_hot\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/preprocessing\/text\/#one_hot<\/a>). But I would like to do my custom tokenization. This will be useful if the language is not English, for example.<\/p>\n\n<p>Does anyone know of examples of code that can take in a list of tokens, and transform them to vectors of integers? I would think this is a fairly common NLP task, so I wanted to check before reinventing the wheel.<\/p>\n","answers":[]}
+{"QuestionId":57663943,"AnswerCount":0,"Tags":"<tensorflow><keras><deep-learning><loss-function><imbalanced-data>","CreationDate":"2019-08-26T19:37:09.687","AcceptedAnswerId":null,"OwnerUserId":6184504.0,"Title":"Combating class imbalance with the right loss function: IoU, Dice or 2-class Dice?","Body":"<p>I am currently working on my Bachelor's thesis and facing some difficulties while trying to understand differences in loss functions regarding class imbalance, and class imbalance itself.<\/p>\n\n<p>I am working on a Segmentation Problem with a variation of the U-Net (<a href=\"https:\/\/arxiv.org\/pdf\/1505.04597.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1505.04597.pdf<\/a>). <\/p>\n\n<p>My dataset contains 11 classes, wherein one is far less frequent than most others. The input to my network is a np-array of shape (700,512,512,3) for the input rgb-images, and shape (700,512,512,11) for the ground truth labels. Each of the 11 channels of the ground truth are binary encoded (1 for specific class present, 0 for lack thereof). So there is one mask per class per image. In most masks, the amount of 1s are far less than those of 0s.<\/p>\n\n<p>First regarding the class imbalance:<\/p>\n\n<p>I have not found any clear definition of the termn \"class imbalance\". In my opinion, three interpretations are possible:<\/p>\n\n<ol>\n<li>Over all 700 input images, one class is far less frequent than most others.<\/li>\n<li>Within one mask of one image, the number of 0s is far higher than that of the 1s<\/li>\n<li>Over all input images, for one class, the (mean) number of 0s is far higher than that of the 1s<\/li>\n<\/ol>\n\n<p>For each of those interpretations, I can imagine problems. Which one is commonly meant when talking about class imbalance?\nAnd how would that one lead to problems?<\/p>\n\n<p>Now the loss functions. The loss functions I am trying to understand are:<\/p>\n\n<ul>\n<li>IoU approximation Loss (<a href=\"http:\/\/www.cs.umanitoba.ca\/~ywang\/papers\/isvc16.pdf\" rel=\"nofollow noreferrer\">http:\/\/www.cs.umanitoba.ca\/~ywang\/papers\/isvc16.pdf<\/a>)<\/li>\n<li>Dice Loss (<a href=\"https:\/\/arxiv.org\/pdf\/1707.00478.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1707.00478.pdf<\/a>)<\/li>\n<li>2-class Dice Loss (<a href=\"https:\/\/arxiv.org\/pdf\/1707.03237.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1707.03237.pdf<\/a>)<\/li>\n<\/ul>\n\n<p>In my research I have read that IoU Loss has problems when facing imbalanced datasets. I have also read that Dice Loss performes better in that case, but I could not find an explanation. It was just discounted as \"common knowledge\". I assume 2-class Dice loss would be even better, since it regards the binary problem from both perspectives, compensating the class imbalance somewhat.<\/p>\n\n<p>My assumption based mainly on an example I exercised (see below) is:<\/p>\n\n<p>With IoU, false positive predictions have a smaller error compared to that for false negatives. However, the probability of receiving a false positive prediction are higher for a category with few pixels, so that the network might settle for false negatives. The Dice Loss punishes false negatives and false positives less than the IoU Loss and the difference between false positive error and false negative error is smaller for the Dice loss. This should lead to improved performance, because the imbalance in probabilities is compensated slightly. With the 2-class Dice Loss, all errors are punished from both perspectives, resulting in less difference between false negative and false positive error terms.<\/p>\n\n<p>Is any of this true? Where am I wrong? At this point, I am really confused and would appreciate any indication. Some sources regarding my questions would also help a lot.<\/p>\n\n<p>And the last question: What loss function would be most suitable to my case with regard to class imbalance?<\/p>\n\n<p>Thanks in advance!<\/p>\n\n<pre><code>Example with 9 pixels, two 1s, seven 0s:\n\n         all correct | 1 false positive | 1 false negative\n\nIoU     |     1      |        0.33      |      0.5\n\nDice    |     1      |        0.2       |      0.33\n\nDice2   |     1      |        0.14      |      0.2\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57664006,"AnswerCount":1,"Tags":"<python><keras><mnist>","CreationDate":"2019-08-26T19:41:41.700","AcceptedAnswerId":57667267.0,"OwnerUserId":7253901.0,"Title":"Images not standardizing in Keras","Body":"<p>I am following a 2017 tutorial on deep learning in Python, and I am trying to standardize some images from the well-known MNIST dataset. The author has the following to say about the piece of code below (Decided to include it all, so anyone who might want to help can simply copy\/paste and reproduce the issue). <\/p>\n\n<blockquote>\n  <p>It is also possible to standardize pixel values across the entire dataset. This is called feature standardization and mirrors the type of standardization often performed for each column in a tabular\n  dataset. This is different to sample standardization described in the previous section as pixel values are standardized across all samples (all images in the dataset). In this case each image is considered a feature. You can perform feature standardization by setting the featurewise center\n  and featurewise std normalization arguments on the ImageDataGenerator class.<\/p>\n<\/blockquote>\n\n<p>Here is the code:<\/p>\n\n<pre><code>from keras.datasets import mnist\nfrom keras.preprocessing.image import ImageDataGenerator\nfrom matplotlib import pyplot as plt\nfrom keras import backend as K\nK.set_image_dim_ordering('th')\n\n#Load data\n(X_train, y_train), (X_test, y_test) = mnist.load_data()\n\n# Reshape to be [samples][pixels][width][height]\nX_train = X_train.reshape(X_train.shape[0], 1, 28, 28).astype('float32')\nX_test = X_test.reshape(X_test.shape[0], 1, 28, 28).astype('float32')\n\ndatagen = ImageDataGenerator(featurewise_center=True, featurewise_std_normalization=True) #This standardizes pixel values across the entire dataset.\n\n# Fit parameters from data\ndatagen.fit(X_train)\n\n#Configure batch size and retrieve one batch of images\nfor X_batch, y_batch in datagen.flow(X_train, y_train, batch_size=9, shuffle=False):\n    #Create grid of 3x3 images\n    for i in range(0,9):\n        plt.subplot(330 + 1 + i)\n        plt.imshow(X_batch[i].reshape(28,28), cmap=plt.get_cmap('gray'))\n    plt.show()\n    break\n<\/code><\/pre>\n\n<p><strong>I think that something has changed since 2017 in Keras especially in the ImageDataGenerator class, or in the flow function of that class, so that I now have to do this somewhat differently. The problem is: I can not find how. Any help would be appreciated<\/strong> <\/p>\n\n<p>Current output: <a href=\"https:\/\/i.stack.imgur.com\/mVc9A.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/mVc9A.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Expected output: <a href=\"https:\/\/i.stack.imgur.com\/rMxHb.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rMxHb.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"57667267","CreationDate":"2019-08-27T03:16:02.217","ParentId":null,"OwnerUserId":"3595291","Title":null,"Body":"<p><code>ImageDataGenerator<\/code> in Keras performs normalization across all images in your code example. You can check the actual <code>X_batch<\/code> values using the following code:<\/p>\n\n<pre><code>for X_batch, y_batch in datagen.flow(X_train, y_train, batch_size=9, shuffle=False):\n    print(X_batch[0])\n    break \n<\/code><\/pre>\n\n<p>The problem is related to the behavior of <code>imshow<\/code>, which is by default <a href=\"https:\/\/stackoverflow.com\/questions\/22121239\/matplotlib-imshow-default-colour-normalisation\">use color normalization<\/a>. Hence, <code>imshow<\/code> visualizes the normalized images as non-normalized.<\/p>\n"}]}
+{"QuestionId":57664097,"AnswerCount":2,"Tags":"<python><tensorflow><anaconda><ml-agent>","CreationDate":"2019-08-26T19:48:00.587","AcceptedAnswerId":57672391.0,"OwnerUserId":11980310.0,"Title":"ModuleNotFoundError for Tensorflow and Anaconda (Windows 10)","Body":"<p>I can't get tensorflow to work on two different Windows PCs and on both I get \"ModuleNotFoundError: No module named 'tensorflow' when try to import them in python. <\/p>\n\n<p>Numpy for example works on one machine not the other. Checked a dozen of posts on stackoverflow, nothing really helps. Clearly somethings wrong with the paths, can't really figure out what and how to trouble check it. I followed this set-up  <a href=\"https:\/\/github.com\/Unity-Technologies\/ml-agents\/blob\/master\/docs\/Installation-Windows.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Unity-Technologies\/ml-agents\/blob\/master\/docs\/Installation-Windows.md<\/a>\nI need the 1.7.1 version of TF for Unity ML-agents to work<\/p>\n\n<pre><code>                  OS: windows 10\n      conda version : 4.4.11\nconda-build version : 3.4.1\n     python version : 3.6.4.final.0\n<\/code><\/pre>\n\n<p>Even a clean install of Anaconda didn't help.<\/p>\n","answers":[{"AnswerId":"57664223","CreationDate":"2019-08-26T19:58:25.247","ParentId":null,"OwnerUserId":"7253901","Title":null,"Body":"<p>These problems often happen when you are on the wrong environment, or when you do not activate your environment. Please try the following steps and see if the problem persists:<\/p>\n\n<ol>\n<li>Open Anaconda prompt<\/li>\n<li>Activate the environment in which you installed tensorflow by using <code>conda activate YOUR_ENV_NAME<\/code><\/li>\n<li>Type <code>python<\/code> to launch python<\/li>\n<li>try to <code>import tensorflow<\/code><\/li>\n<\/ol>\n\n<p>And see if the error persists.<\/p>\n"},{"AnswerId":"57672391","CreationDate":"2019-08-27T10:10:32.307","ParentId":null,"OwnerUserId":"11980310","Title":null,"Body":"<p>I finally managed to fix this (not with ML-agents, but that's a different story). I uninstalled Anaconda and Python as well as cleared the registry from all Python information. I had a previous installation of Python 3.7 and I'm guessing this might have caused some issues. Clean install of Anaconda with a 3.6 env and TF is working both with cpu and gpu support. <\/p>\n"}]}
+{"QuestionId":57664152,"AnswerCount":0,"Tags":"<tensorflow><object-detection-api><k-fold>","CreationDate":"2019-08-26T19:52:47.543","AcceptedAnswerId":null,"OwnerUserId":4231571.0,"Title":"How to use k-fold cross validation on object-detection API","Body":"<p>I would like to know how I can perform k-fold cross validation with the Tensorflow Object-Detection API. In this Stack Overflow thread: <a href=\"https:\/\/stackoverflow.com\/questions\/51989259\/k-fold-cross-validation-tensorflow-object-detection\">K fold cross validation Tensorflow Object Detection<\/a> a comment mentions how using that is a bad idea, but doesn't really explain why.<\/p>\n\n<p>Thank you for your help.<\/p>\n","answers":[]}
+{"QuestionId":57664289,"AnswerCount":1,"Tags":"<python><pytorch><dataloader>","CreationDate":"2019-08-26T20:04:20.740","AcceptedAnswerId":null,"OwnerUserId":11913500.0,"Title":"DataLoader crashes when shuffling","Body":"<p>I'm using DataLoader to read from a custom Dataset object based on numpy memmap.\nAs long as I read the data without shuffling everything works fine but, as I set <code>shuffle=True<\/code>, the runtime crash.<\/p>\n\n<p>I tried implementing the shuffling mechanism in the Dataset class by using a permutation vector and setting <code>shuffle=False<\/code> in the DataLoader but the issue persists.\nI also noticed that, when shuffling, the <code>__getitem__()<\/code> function of the Dataset object is called n times, where n is the batch_size.<\/p>\n\n<p>Here's the Dataset code:<\/p>\n\n<pre><code>class CustomDataset(Dataset):\n  num_pattern = 60112\n  base_folder = 'dataset'\n\n  def __init__(self, root):\n    self.root = os.path.expanduser(root)\n\n    self.output_ = np.memmap('{0}\/output'.format(root), 'int64', 'r', shape=(60112, 62))\n    self.out_len = np.memmap('{0}\/output-lengths'.format(root), 'int32', 'r', shape=(60112))\n    self.input_ = np.memmap('{0}\/input'.format(root), 'float32', 'r', shape=(60112, 512, 1024))\n    self.in_len = np.memmap('{0}\/input-lengths'.format(root), 'int32', 'r', shape=(60112))\n\n\n  def __len__(self):\n    return self.num_pattern\n\n  def __getitem__(self, index):\n    return (self.in_len[index], torch.from_numpy(self.input_[index])), (self.out_len[index], torch.from_numpy(self.output_[index]))\n\nif __name__ == '__main__':\n  dataset = CustomDataset(root='\/content\/')\n  data_loader = data.DataLoader(dataset, batch_size=32, shuffle=False, num_workers=1)\n  for i, data in enumerate(data_loader, 0):\n    # training\n<\/code><\/pre>\n\n<p>The error stack is the following:<\/p>\n\n<pre><code>RuntimeError                              Traceback (most recent call last)\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/utils\/data\/dataloader.py in _try_get_batch(self, timeout)\n    510         try:\n--&gt; 511             data = self.data_queue.get(timeout=timeout)\n    512             return (True, data)\n\n9 frames\n\/usr\/lib\/python3.6\/multiprocessing\/queues.py in get(self, block, timeout)\n    103                     timeout = deadline - time.monotonic()\n--&gt; 104                     if not self._poll(timeout):\n    105                         raise Empty\n\n\/usr\/lib\/python3.6\/multiprocessing\/connection.py in poll(self, timeout)\n    256         self._check_readable()\n--&gt; 257         return self._poll(timeout)\n    258 \n\n\/usr\/lib\/python3.6\/multiprocessing\/connection.py in _poll(self, timeout)\n    413     def _poll(self, timeout):\n--&gt; 414         r = wait([self], timeout)\n    415         return bool(r)\n\n\/usr\/lib\/python3.6\/multiprocessing\/connection.py in wait(object_list, timeout)\n    910             while True:\n--&gt; 911                 ready = selector.select(timeout)\n    912                 if ready:\n\n\/usr\/lib\/python3.6\/selectors.py in select(self, timeout)\n    375             try:\n--&gt; 376                 fd_event_list = self._poll.poll(timeout)\n    377             except InterruptedError:\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/utils\/data\/_utils\/signal_handling.py in handler(signum, frame)\n     62         # Python can still get and update the process status successfully.\n---&gt; 63         _error_if_any_worker_fails()\n     64         if previous_handler is not None:\n\nRuntimeError: DataLoader worker (pid 3978) is killed by signal: Bus error. \n\nDuring handling of the above exception, another exception occurred:\n\nRuntimeError                              Traceback (most recent call last)\n&lt;ipython-input-8-b407a8532808&gt; in &lt;module&gt;()\n      5   data_loader = data.DataLoader(dataset, batch_size=4, shuffle=True, num_workers=1)\n      6 \n----&gt; 7   for i, data in enumerate(data_loader, 0):\n      8     print(i)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/utils\/data\/dataloader.py in __next__(self)\n    574         while True:\n    575             assert (not self.shutdown and self.batches_outstanding &gt; 0)\n--&gt; 576             idx, batch = self._get_batch()\n    577             self.batches_outstanding -= 1\n    578             if idx != self.rcvd_idx:\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/utils\/data\/dataloader.py in _get_batch(self)\n    551         else:\n    552             while True:\n--&gt; 553                 success, data = self._try_get_batch()\n    554                 if success:\n    555                     return data\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/utils\/data\/dataloader.py in _try_get_batch(self, timeout)\n    517             if not all(w.is_alive() for w in self.workers):\n    518                 pids_str = ', '.join(str(w.pid) for w in self.workers if not w.is_alive())\n--&gt; 519                 raise RuntimeError('DataLoader worker (pid(s) {}) exited unexpectedly'.format(pids_str))\n    520             if isinstance(e, queue.Empty):\n    521                 return (False, None)\n\nRuntimeError: DataLoader worker (pid(s) 3978) exited unexpectedly\n<\/code><\/pre>\n","answers":[{"AnswerId":"57671814","CreationDate":"2019-08-27T09:40:29.513","ParentId":null,"OwnerUserId":"10232932","Title":null,"Body":"<p>This is a shared memory error, your dataloader might be requiring more memory for this specific task<\/p>\n"}]}
+{"QuestionId":57664937,"AnswerCount":0,"Tags":"<python><keras><slice>","CreationDate":"2019-08-26T21:05:53.520","AcceptedAnswerId":null,"OwnerUserId":611660.0,"Title":"Slicing in keras Lambda doesn't work as expected","Body":"<p>I'm trying to pick out different parts of the input to be fed into different streams. Slicing works like this in numpy:<\/p>\n\n<pre><code>x = numpy.arange(10).reshape(-1,10)\nprint(x[:, [1,2,3]])\n# &gt;&gt;&gt; [[1 2 3]]\nprint(x[:, [2,5,9]])\n#&gt;&gt;&gt; [[2 5 9]]\n<\/code><\/pre>\n\n<p>But not for keras:<\/p>\n\n<pre><code>features = numpy.arange(10).reshape(-1, 10)\n\ninputs = keras.layers.Input(shape=(10,))\npredictions = keras.layers.Lambda(lambda x: x)(inputs)\nmodel = keras.models.Model(inputs=inputs, outputs=predictions)\nmodel.compile(loss='mean_squared_error', optimizer='adam')\nprint(model.predict(features))\n# &gt;&gt;&gt; [[0. 1. 2. 3. 4. 5. 6. 7. 8. 9.]]\n\n\ninputs = keras.layers.Input(shape=(10,))\npredictions = keras.layers.Lambda(lambda x: x[:,1:3])(inputs)\nmodel = keras.models.Model(inputs=inputs, outputs=predictions)\nmodel.compile(loss='mean_squared_error', optimizer='adam')\nprint(model.predict(features))\n# &gt;&gt;&gt; [[1. 2.]]\n\ninputs = keras.layers.Input(shape=(10,))\npredictions = keras.layers.Lambda(lambda x: x[:,[1,2,3]])(inputs)\nmodel = keras.models.Model(inputs=inputs, outputs=predictions)\nmodel.compile(loss='mean_squared_error', optimizer='adam')\nprint(model.predictsn(features))\n# &gt;&gt;&gt; Traceback (most recent call last):\n#  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n#  File \"\/home\/user\/test.py\", line 589, in &lt;module&gt;\n#    predictions = keras.layers.Lambda(lambda x: x[:,[1,2,3]])(inputs)\n#  File \"\/home\/user\/.local\/lib\/python3.6\/site-packages\/keras\/engine\/base_layer.py\", line 451, in __call__\n#    output = self.call(inputs, **kwargs)\n#  File \"\/home\/user\/.local\/lib\/python3.6\/site-packages\/keras\/layers\/core.py\", line 716, in call\n#    return self.function(inputs, **arguments)\n#  File \"\/home\/user\/test.py\", line 589, in &lt;lambda&gt;\n#    predictions = keras.layers.Lambda(lambda x: x[:,[1,2,3]])(inputs)\n#  File \"\/home\/user\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/array_ops.py\", line 491, in _slice_helper\n#    end.append(s + 1)\n#TypeError: can only concatenate list (not \"int\") to list\n\ninputs = keras.layers.Input(shape=(10,))\npredictions = keras.layers.Lambda(lambda x: tf.gather_nd(x, [1,2,3]))(inputs)\nmodel = keras.models.Model(inputs=inputs, outputs=predictions)\nmodel.compile(loss='mean_squared_error', optimizer='adam')\nprint(model.predict(features))\n#&gt;&gt;&gt; Traceback (most recent call last):\n#  File \"\/home\/user\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1628, in _create_c_op\n#    c_op = c_api.TF_FinishOperation(op_desc)\n#tensorflow.python.framework.errors_impl.InvalidArgumentError: indices.shape[-1] must be &lt;= params.rank, but saw indices shape: [3] and params shape: [?,10] for 'lambda_28\/GatherNd' (op: 'GatherNd') with input shapes: [?,10], [3].\n\n\n&gt;&gt;&gt; tf.__version__\n'1.12.0'\n<\/code><\/pre>\n\n<p>What sort of details would the stackoverflow robot like me to add to make this question contain less code? Should I remove the parts where I tried things and the errors that I got so that somebody can ask me for them as clarifications?<\/p>\n","answers":[]}
+{"QuestionId":57665342,"AnswerCount":1,"Tags":"<python><tensorflow><keras><concatenation><conv-neural-network>","CreationDate":"2019-08-26T21:45:57.900","AcceptedAnswerId":57671442.0,"OwnerUserId":8510077.0,"Title":"Concatenate two layers","Body":"<p>I have met an error message when I tried to concatenate outcomes from two layers.<\/p>\n\n<pre><code>def cnn_model_fn(learning_rate):\n    \"\"\"Model function for CNN.\"\"\"\n    model1=Sequential()\n\n      # Convolutional Layer #1\n    model1.add(tf.keras.layers.Conv2D(\n          filters=20,\n          kernel_size=[10, 1],\n          kernel_initializer='he_uniform',\n          bias_initializer=keras.initializers.Constant(value=0),\n          padding=\"same\",\n          activation=tf.nn.relu, input_shape=(410,1,3)))\n    model1.add(Flatten())\n\n    model2=Sequential()\n\n    model2.add(tf.keras.layers.Conv2D(\n          filters=20,\n          kernel_size=[10, 1],\n          kernel_initializer='he_uniform',\n          bias_initializer=keras.initializers.Constant(value=0),\n          padding=\"same\",\n          activation=tf.nn.relu, input_shape=(410,1,3)))\n    model2.add(Flatten())\n\n    model4=Sequential()\n    model4.add(keras.layers.Concatenate(axis=-1)([model1, model2]))\n\n    optimizer = tf.train.AdamOptimizer(learning_rate)\n    model4.compile(loss='mean_squared_error',\n                optimizer=optimizer,\n                metrics=['mean_absolute_error', 'mean_squared_error'])\n\n    return model4\n\nmodel4=cnn_model_fn(0.1) \nmodel4.summary()\n<\/code><\/pre>\n\n<blockquote>\n  <p>\"\/usr\/local\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/layers\/merge.py\n  in build(self, input_shape)\n      377     # Used purely for shape validation.\n      378     if not isinstance(input_shape, list) or len(input_shape) &lt; 2:\n  --> 379       raise ValueError('A <code>Concatenate<\/code> layer should be called '\n      380                        'on a list of at least 2 inputs')\n      381     if all([shape is None for shape in input_shape]):<\/p>\n  \n  <p>ValueError: A <code>Concatenate<\/code> layer should be called on a list of at\n  least 2 inputs\"<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"57671442","CreationDate":"2019-08-27T09:19:37.097","ParentId":null,"OwnerUserId":"8563649","Title":null,"Body":"<p>You are trying to concatenate 2 models but what you want is to concatenate 2 layers. Try the following code.<\/p>\n\n<pre><code>from tensorflow.keras.models import Sequential, Model\nfrom tensorflow.keras.layers import Flatten, Input\n\ndef cnn_model_fn(learning_rate):\n    \"\"\"Model function for CNN.\"\"\"\n    input_layer=Input(shape=(410,1,3))\n\n    x1 = (tf.keras.layers.Conv2D(\n          filters=20,\n          kernel_size=[10, 1],\n          kernel_initializer='he_uniform',\n          bias_initializer=keras.initializers.Constant(value=0),\n          padding=\"same\",\n          activation=tf.nn.relu ))(input_layer)\n    x1 = Flatten()(x1)\n\n    x2 = (tf.keras.layers.Conv2D(\n          filters=20,\n          kernel_size=[10, 1],\n          kernel_initializer='he_uniform',\n          bias_initializer=keras.initializers.Constant(value=0),\n          padding=\"same\",\n          activation=tf.nn.relu))(input_layer)\n    x2 = Flatten()(x2)\n\n    x = (keras.layers.Concatenate(axis=-1)([x1,x2]))\n\n    model = Model(input_layer, x)\n    optimizer = tf.train.AdamOptimizer(learning_rate)\n    model.compile(loss='mean_squared_error',\n                optimizer=optimizer,\n                metrics=['mean_absolute_error', 'mean_squared_error'])\n\n    return model\n<\/code><\/pre>\n"}]}
+{"QuestionId":57665499,"AnswerCount":1,"Tags":"<keras><keras-layer><transfer-learning>","CreationDate":"2019-08-26T22:06:47.917","AcceptedAnswerId":null,"OwnerUserId":11978986.0,"Title":"ValueError: Layer conv2d_41 was called with an input that isn't a symbolic tensor. All inputs to the layer should be tensors","Body":"<p>I try transfer learning with custom input of backbone:<\/p>\n\n<p>(I can not transfer learning normally because my input shape is N*N*8, so I need add small network_1 to reach N*N*3) <\/p>\n\n<pre><code>model_1 \n   |\n   |\nadd model_2\n   |\nadd some layer\n<\/code><\/pre>\n\n<p>My code:<\/p>\n\n<p>model_1.add(model_2)<\/p>\n\n<p>model_1 is my small network:\nmodel_2 is Mobilenet or VGG16, or Densenet .....<\/p>\n\n<pre><code>model_1 = Sequential()\nmodel_1.add(InputLayer(input_shape=(size, size, F), name=\"InputLayer\"))\nmodel_1.add(Convolution2D(3, 128, padding = 'same'))\n\nfrom keras.applications.densenet import DenseNet169\nmodel_2=DenseNet169(weights='imagenet',include_top=False)\n\nmodel_2.layers.pop(0) # remove input_layer of model_2 \nmodel_1.add(model_2) # output model_1 is input model_2?\n\nmodel_1 = GlobalAveragePooling2D()(model_1)\nmodel_1 = Dropout(0.2)(model_1)\nmodel_1 = Dense(256*256, activation='softmax')(model_1)\nmodel_1 = Reshape(256, 256)(model_1)\n\n<\/code><\/pre>\n\n<p>I got errors:<\/p>\n\n<pre><code>ValueError: Layer global_average_pooling2d_3 was called with an input that isn't a symbolic tensor. Received type: &lt;class 'keras.engine.sequential.Sequential'&gt;. Full input: [&lt;keras.engine.sequential.Sequential object at 0x7f74f6621d68&gt;]. All inputs to the layer should be tensors.\n<\/code><\/pre>\n\n<p>What wrong in my code?<\/p>\n","answers":[{"AnswerId":"57665601","CreationDate":"2019-08-26T22:22:13.133","ParentId":null,"OwnerUserId":"10023182","Title":null,"Body":"<p>Global Average Pooling 2D is a layer that performs an operation on a tensor, or multi-dimensional array. Therefore, passing a model architecture like DenseNet throws an error because the model has no idea what it's looking at. Model Architecture files are totally different from tensors.<\/p>\n\n<p>To achieve what I think you're trying to achieve, run DenseNet and then pass the <em>output<\/em> of DenseNet into the model you're creating, instead of passing the model itself. Good luck!<\/p>\n"}]}
+{"QuestionId":57665770,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-26T22:46:23.043","AcceptedAnswerId":null,"OwnerUserId":7670875.0,"Title":"Why tf.gradients doesn't work with integer constants?","Body":"<p>Why the function below prints <code>None<\/code>?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>a = tf.constant(4)\nb = tf.constant(2)\n\ngr = tf.gradients(a + b, [a, b])\n\nprint(sess.run(gr))\n<\/code><\/pre>\n\n<p>But when I change <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>-a = tf.constant(4)\n-b = tf.constant(2)\n+a = tf.constant(4.0)\n+b = tf.constant(2.0)\n<\/code><\/pre>\n\n<p>It prints gradients. What is the reason for that?<\/p>\n","answers":[{"AnswerId":"57677851","CreationDate":"2019-08-27T15:18:59.177","ParentId":null,"OwnerUserId":"7670875","Title":null,"Body":"<p>According to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/20524\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/20524<\/a> \ntensorflow team made <code>tf.gradients<\/code> incompatible with integer tensors for this reason:<\/p>\n\n<blockquote>\n  <p>In effect, allowing gradients on integer tensors was causing\n  incorrectness in tf.while_loop, and there was no satisfactory way to\n  resolve them without this change.<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":57665904,"AnswerCount":0,"Tags":"<python><keras><scikit-learn>","CreationDate":"2019-08-26T23:09:46.607","AcceptedAnswerId":null,"OwnerUserId":6426203.0,"Title":"GridSearchCV scores are all the same","Body":"<p>I get the same negative mean squared error with different grid search parameters. In effect, this means my network training is <em>parameter-independent<\/em>, which I'm sure is wrong.<\/p>\n\n<p>I've gone through the GridSearchCV documentation as well as many SO posts. Perhaps the error function is minimized with the exact same precision every time but I doubt it. The loss values outputted during training vary between folds which tells me that each iteration of training is unique.<\/p>\n\n<p>Code:<\/p>\n\n<pre><code>def create_model(lr=0.001, dropout=0.2):\n    NN_model = Sequential()\n\n    NN_model.add(Dense(128, kernel_initializer='normal',input_dim = X_train.shape[1], activation='relu'))\n\n    NN_model.add(Dense(256, kernel_initializer='normal',activation='relu'))\n    if dropout != 0:\n        NN_model.add(Dropout(dropout))\n    NN_model.add(Dense(256, kernel_initializer='normal',activation='relu'))\n    if dropout != 0:\n        NN_model.add(Dropout(dropout))\n    NN_model.add(Dense(256, kernel_initializer='normal',activation='relu'))\n    if dropout != 0:\n        NN_model.add(Dropout(dropout))\n\n    NN_model.add(Dense(1, kernel_initializer='normal',activation='linear'))\n\n    Adam = optimizers.Adam(lr)\n\n    NN_model.compile(loss='mse', optimizer=Adam) # don't specify metric\n    return NN_model\n\ncheckpoint_name = 'Weights-{epoch:03d}--{loss:.5f}.hdf5' \ncheckpoint = ModelCheckpoint(checkpoint_name, monitor='val_loss', verbose = 1, save_best_only = True, mode ='auto')\ncallbacks_list = [checkpoint]\n\nmodel = KerasClassifier(build_fn=create_model)\n\n\nparam_grid = {\"epochs\": [300], \"batch_size\": [32], \"lr\": [0.0001, 0.001, 0.01], \"dropout\": [0.02, 0.2]}\n\ngrid = GridSearchCV(estimator=model, \n                param_grid=param_grid, \n                n_jobs=1, \n                scoring=[\"neg_mean_squared_error\", \"r2\"],\n                cv=2,\n                refit=\"neg_mean_squared_error\")\n\ngrid_result = grid.fit(X_train, y_train, callbacks=callbacks_list)\n<\/code><\/pre>\n\n<p>Results:<\/p>\n\n<p><img src=\"https:\/\/user-images.githubusercontent.com\/11367325\/63728453-f6678980-c820-11e9-9bb7-2461c221020e.png\" alt=\"results\"><\/p>\n\n<p>The loss curve looks good. Again, not sure what it's calculating. I set loss='mse', which typically for my data produces values in the 1e5\/1e6 range.<\/p>\n\n<p><img src=\"https:\/\/user-images.githubusercontent.com\/11367325\/63729103-62e38800-c823-11e9-81b6-3535811ac0d7.png\" alt=\"image\"><\/p>\n","answers":[]}
+{"QuestionId":57666205,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-08-27T00:06:16.260","AcceptedAnswerId":null,"OwnerUserId":11377572.0,"Title":"question about validation split in keras model fit function","Body":"<p>The default value of validation_split argument of keras model fit function is 0.0. Does it means if don't set validation_split, the model will not evaluate on validation data set? If model doesn't evaluate on validation data set, then during training, where does the loss value like mean absolute error comes from? does it calculated with predicted batch of training data and actual batch of training data?<\/p>\n\n<p>What's the benefit and difference of setting validation_split argument?<\/p>\n","answers":[]}
+{"QuestionId":57666226,"AnswerCount":0,"Tags":"<python><performance><tensorflow><ubuntu-18.04>","CreationDate":"2019-08-27T00:10:01.967","AcceptedAnswerId":null,"OwnerUserId":3444294.0,"Title":"Counter-Intuitive Tensorflow Performance Variance Between Platforms","Body":"<p>I have two Ubuntu 18.04 machines with identical GPUs but different CPU and RAM. The smaller machine consistently runs Tensorflow ~1.5x faster than the bigger machine. Both machines have: GTX 1080 Ti, conda 4.7.11, tensorflow 1.14.0, same\/latest versions of python (3.7.4) and packages, all updates, CUDA, etc. The only software difference is that the <a href=\"https:\/\/www.tensorflow.org\/install\/gpu\" rel=\"nofollow noreferrer\">tensorflow GPU install procedure<\/a> installs Nvidia driver v430.26 on the small machine and v430.40 on the big machine (why??).<\/p>\n\n<ul>\n<li>Small machine: i7-6700K 4.3 GHz, 32 GB RAM, ASUS Z170K motherboard.<\/li>\n<li>Big machine: i9-9900K 5 GHz, 64 GB RAM, Gigabyte Z390UD motherboard.<\/li>\n<\/ul>\n\n<p>Both machines use XMP memory profile 1, identical SSDs, have the GPU in a PCIEX16 slot, and no other PCI cards.<\/p>\n\n<p>Benchmark Tests:<\/p>\n\n<ul>\n<li><p>CUDA-Z tests showed GPU performance virtually the same between the 2 machines, with the big machine slightly faster in GPU memory copy.<\/p><\/li>\n<li><p>Geekbench tests showed the big machine faster in all CPU and memory tests. <\/p><\/li>\n<\/ul>\n\n<p>I even swapped the drives between the machines (i.e. booted and tested each machine using the other machine's SSD) and still the smaller machine was consistently 50% faster in Tensorflow. That would seem to indicate that the Nvidia driver version is not the cause. I also swapped the GPUs between the machines, and still the small machine is faster.<\/p>\n\n<p>What evil magic is this?<\/p>\n\n<p>Could it be the motherboard? that would be strange, since the benchmarks show the larger machine faster all around.<\/p>\n\n<p>Could it be something in the TF binary that does better on an i7 vs i9? I have not compiled from source (I don't like pain) but could do so if necessary. I will try to run the <a href=\"https:\/\/github.com\/tensorflow\/benchmarks\/tree\/master\/perfzero\" rel=\"nofollow noreferrer\">TF benchmarks<\/a> to see what that uncovers.<\/p>\n","answers":[]}
+{"QuestionId":57666249,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-08-27T00:14:53.193","AcceptedAnswerId":57668000.0,"OwnerUserId":710942.0,"Title":"How can I combine and overlap two tensors?","Body":"<p>I have two tensors that should together overlap each other to form a larger tensor. To illustrate:<\/p>\n\n<pre><code>a = torch.Tensor([[1, 2, 3], [1, 2, 3]])\nb = torch.Tensor([[5, 6, 7], [5, 6, 7]])\n\na = [[1 2 3]    b = [[5 6 7]\n     [1 2 3]]        [5 6 7]]\n<\/code><\/pre>\n\n<p>I want to combine the two tensors and have them partially overlap by a single column, with the average being taken for those elements that overlap.<\/p>\n\n<p>e.g.<\/p>\n\n<pre><code>result = [[1 2 4 6 7]\n          [1 2 4 6 7]]\n<\/code><\/pre>\n\n<p>The first two columns are the first two columns of  'a'. The last two columns are the last two columns of 'b'. The middle column is the average of 'a's last column and 'b's first column.<\/p>\n\n<p>I know how to merge two tensors side by side or in a new dimension. But doing this eludes me.<\/p>\n\n<p>Can anyone help?<\/p>\n","answers":[{"AnswerId":"57668000","CreationDate":"2019-08-27T05:09:44.267","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>This is not a trivial operation, and this solution is not very trivial or intuitive either.<\/p>\n\n<p>Looking at <code>result<\/code> with shape=<code>(2, 5)<\/code>, you can think of <code>a<\/code> and <code>b<\/code> as two <code>2x3<\/code> patches of <code>result<\/code> taken with <code>stride=2<\/code>. Like this illustration:<br>\n<a href=\"https:\/\/i.stack.imgur.com\/kRaiY.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/kRaiY.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>We can use pytorch's <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/nn.html#unfold\" rel=\"nofollow noreferrer\"><code>unfold<\/code><\/a> to \"recover\" the green (<code>a<\/code>) and blue (<code>b<\/code>) patches from <code>result<\/code> (\"recover\" up to the averaged values):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from torch.nn import functional as nnf\n\nrecovered = nnf.unfold(result, kernel_size=(2,3), stride=2)\n<\/code><\/pre>\n\n<p>The result is:  <\/p>\n\n<blockquote>\n<pre><code>tensor([[[1., 4.],\n         [2., 6.],\n         [4., 7.],\n         [1., 4.],\n         [2., 6.],\n         [4., 7.]]])\n<\/code><\/pre>\n<\/blockquote>\n\n<p>The patches were recovered (as column vectors).<\/p>\n\n<p>Now that we understand how to get <code>a<\/code> and <code>b<\/code> from <code>result<\/code>, we can use <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/nn.html#fold\" rel=\"nofollow noreferrer\"><code>fold<\/code><\/a> to perform the \"inverse\" operation and go from <code>b<\/code> and <code>b<\/code> to <code>result<\/code>.<br>\nFirst, we need to flatten concatenate <code>a<\/code> and <code>b<\/code> to the shape <code>fold<\/code> expects (mimicking the output of <code>unfold<\/code>, two \"flatten\" patches of 3x2 elements):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>uf = torch.cat((a.view(1, 6, 1), b.view(1, 6, 1)), dim=2)\n<\/code><\/pre>\n\n<p>We can now \"fold\" the patches <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>raw = nnf.fold(uf, (2,5), kernel_size=(2,3), stride=2)\n<\/code><\/pre>\n\n<p>We are not there yet, when there are overlapping elements <code>fold<\/code> <a href=\"https:\/\/stackoverflow.com\/a\/53972525\/1714410\">sums up<\/a> the overlapping elements, resulting with<\/p>\n\n<blockquote>\n<pre><code>tensor([[[[1., 2., 8., 6., 7.],\n          [1., 2., 8., 6., 7.]]]])\n<\/code><\/pre>\n<\/blockquote>\n\n<p>To count how many elements were summed for each entry in result, we can simply \"fold\" an all ones tensor<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>counter = nnf.fold(torch.ones_like(uf), (2, 5), kernel_size=(2, 3), stride=2)\n<\/code><\/pre>\n\n<p>And finally, we can recover <code>result<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>result = raw \/ counter\n<\/code><\/pre>\n\n<blockquote>\n<pre><code>tensor([[[[1., 2., 4., 6., 7.],\n          [1., 2., 4., 6., 7.]]]])\n<\/code><\/pre>\n<\/blockquote>\n\n<p>Piece of cake.<\/p>\n"}]}
+{"QuestionId":57666268,"AnswerCount":0,"Tags":"<python><tensorflow><keras><memory-leaks><model>","CreationDate":"2019-08-27T00:18:22.863","AcceptedAnswerId":null,"OwnerUserId":10792140.0,"Title":"Leaky of Memory in Keras\/Tensorflow","Body":"<p>I am trying to run several models in sequence in python using keras \/ tensorflow (version 1.14.0). The problem is that a memory leak is occurring. =\/<\/p>\n\n<p>I tried to clean up the session and the graph generated by tensorflow.\nI tried gc and it didn't work<\/p>\n\n<pre><code>    import gc\n    import tensorflow as tf\n    tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)\n    from sklearn.datasets import load_iris\n    from sklearn.model_selection import train_test_split\n\n    @profile\n    def train_model(x_train, x_test, y_train, y_test):\n\n        model = 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\n        model.compile(optimizer='adam',\n                    loss='sparse_categorical_crossentropy',\n                    metrics=['accuracy'])\n\n        model.fit(x_train, y_train, epochs=1000, verbose=0)\n        model.evaluate(x_test, y_test)\n\n        tf.keras.backend.clear_session()\n        tf.compat.v1.reset_default_graph()\n        gc.collect()\n\n    @profile\n    def main():\n        n = 500 \n        data = load_iris()\n        x_train, x_test, y_train, y_test = train_test_split(data.data, data.target, test_size=0.33, random_state=42)\n\n        for i in range(n):\n            train_model(x_train, x_test, y_train, y_test)\n\n<\/code><\/pre>\n\n<p>I change the value of n 3 times... n=2, n=100 and n=500.\nThis is the result of memory used:<\/p>\n\n<p>n=2<\/p>\n\n<pre><code>    Line #    Mem usage    Increment   Line Contents\n    ================================================\n        12  238.602 MiB  238.602 MiB   @profile\n        13                             def train_model(x_train, x_test, y_train, y_test):\n        14\n        15  238.602 MiB    0.000 MiB       model = tf.keras.models.Sequential([\n        16  238.602 MiB    2.852 MiB           tf.keras.layers.Flatten(),\n        17  238.602 MiB    0.051 MiB           tf.keras.layers.Dense(512, activation=tf.nn.relu),\n        18  238.602 MiB    0.000 MiB           tf.keras.layers.Dropout(0.2),\n        19  238.602 MiB    0.043 MiB           tf.keras.layers.Dense(10, activation=tf.nn.softmax)\n        20                                 ])\n        21\n        22  238.602 MiB    0.000 MiB       model.compile(optimizer='adam',\n        23  238.602 MiB    0.000 MiB                   loss='sparse_categorical_crossentropy',\n        24  238.602 MiB    0.055 MiB                   metrics=['accuracy'])\n        25\n        26  253.176 MiB   29.328 MiB       model.fit(x_train, y_train, epochs=1000, verbose=0)\n        27  253.176 MiB    0.031 MiB       model.evaluate(x_test, y_test)\n        28\n        29  253.176 MiB    0.027 MiB       tf.keras.backend.clear_session()\n        30  253.176 MiB    0.000 MiB       tf.compat.v1.reset_default_graph()\n        31  253.238 MiB    0.062 MiB       gc.collect()\n\n\n    Filename: main.py\n\n    Line #    Mem usage    Increment   Line Contents\n    ================================================\n        33  206.164 MiB  206.164 MiB   @profile\n        34                             def main():\n        35  206.164 MiB    0.000 MiB       data = load_iris()\n        36  206.215 MiB    0.051 MiB       x_train, x_test, y_train, y_test = train_test_split(data.data, data.target, test_size=0.33, random_state=42)\n        37\n        38  253.238 MiB    0.000 MiB       for i in range(2):\n        39  253.238 MiB  253.238 MiB           train_model(x_train, x_test, y_train, y_test)\n<\/code><\/pre>\n\n<p>n=100<\/p>\n\n<pre><code>    Line #    Mem usage    Increment   Line Contents\n    ================================================\n        12 1040.012 MiB 1040.012 MiB   @profile\n        13                             def train_model(x_train, x_test, y_train, y_test):\n        14\n        15 1040.012 MiB    0.000 MiB       model = tf.keras.models.Sequential([\n        16 1040.012 MiB    2.855 MiB           tf.keras.layers.Flatten(),\n        17 1040.012 MiB    0.051 MiB           tf.keras.layers.Dense(512, activation=tf.nn.relu),\n        18 1040.012 MiB    0.000 MiB           tf.keras.layers.Dropout(0.2),\n        19 1040.012 MiB    0.043 MiB           tf.keras.layers.Dense(10, activation=tf.nn.softmax)\n        20                                 ])\n        21\n        22 1040.012 MiB    0.000 MiB       model.compile(optimizer='adam',\n        23 1040.012 MiB    0.000 MiB                   loss='sparse_categorical_crossentropy',\n        24 1040.012 MiB    0.051 MiB                   metrics=['accuracy'])\n        25\n        26 1048.164 MiB   28.973 MiB       model.fit(x_train, y_train, epochs=1000, verbose=0)\n        27 1048.164 MiB    0.258 MiB       model.evaluate(x_test, y_test)\n        28\n        29 1048.164 MiB    0.027 MiB       tf.keras.backend.clear_session()\n        30 1048.164 MiB    0.000 MiB       tf.compat.v1.reset_default_graph()\n        31 1048.164 MiB    0.062 MiB       gc.collect()\n\n\n    Filename: main.py\n\n    Line #    Mem usage    Increment   Line Contents\n    ================================================\n        33  206.164 MiB  206.164 MiB   @profile\n        34                             def main():\n        35  206.164 MiB    0.000 MiB       data = load_iris()\n        36  206.215 MiB    0.051 MiB       x_train, x_test, y_train, y_test = train_test_split(data.data, data.target, test_size=0.33, random_state=42)\n        37\n        38 1048.164 MiB    0.000 MiB       for i in range(100):\n        39 1048.164 MiB 1048.164 MiB           train_model(x_train, x_test, y_train, y_test)\n<\/code><\/pre>\n\n<p>n=500<\/p>\n\n<pre><code>    Line #    Mem usage    Increment   Line Contents\n    ================================================\n        12 4279.691 MiB 4279.691 MiB   @profile\n        13                             def train_model(x_train, x_test, y_train, y_test):\n        14\n        15 4279.691 MiB    0.062 MiB       model = tf.keras.models.Sequential([\n        16 4279.691 MiB    2.797 MiB           tf.keras.layers.Flatten(),\n        17 4279.691 MiB    0.051 MiB           tf.keras.layers.Dense(512, activation=tf.nn.relu),\n        18 4279.691 MiB    0.000 MiB           tf.keras.layers.Dropout(0.2),\n        19 4279.691 MiB    0.043 MiB           tf.keras.layers.Dense(10, activation=tf.nn.softmax)\n        20                                 ])\n        21\n        22 4279.691 MiB    0.000 MiB       model.compile(optimizer='adam',\n        23 4279.691 MiB    0.000 MiB                   loss='sparse_categorical_crossentropy',\n        24 4279.691 MiB    0.055 MiB                   metrics=['accuracy'])\n        25\n        26 4287.664 MiB   31.020 MiB       model.fit(x_train, y_train, epochs=1000, verbose=0)\n        27 4287.664 MiB    0.254 MiB       model.evaluate(x_test, y_test)\n        28\n        29 4287.664 MiB    0.027 MiB       tf.keras.backend.clear_session()\n        30 4287.664 MiB    0.000 MiB       tf.compat.v1.reset_default_graph()\n        31 4287.664 MiB    0.062 MiB       gc.collect()\n\n\n    Filename: main.py\n\n    Line #    Mem usage    Increment   Line Contents\n    ================================================\n        33  206.164 MiB  206.164 MiB   @profile\n        34                             def main():\n        35  206.164 MiB    0.000 MiB       data = load_iris()\n        36  206.215 MiB    0.051 MiB       x_train, x_test, y_train, y_test = train_test_split(data.data, data.target, test_size=0.33, random_state=42)\n        37\n        38 4287.664 MiB    0.000 MiB       for i in range(500):\n        39 4287.664 MiB 4287.664 MiB           train_model(x_train, x_test, y_train, y_test)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57666392,"AnswerCount":0,"Tags":"<tensorflow><keras><image-resizing><volume>","CreationDate":"2019-08-27T00:39:54.237","AcceptedAnswerId":null,"OwnerUserId":11354575.0,"Title":"How to write a custom keras layer to resample a 3D volume?","Body":"<p>I have a tensor <code>(batch_size, height, width, depth, channel)<\/code>(representing 3D volumes) inside the network, which needs to be resampled. I am considering writing a custom keras layer to do this. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>input = Input()\nx = net(input) # another network\nx = custom_resample_3d(...)(x)\nx = ...\n<\/code><\/pre>\n\n<p>The output of <code>net<\/code> is let's say <code>(batch_size, 64, 64, 64, 1)<\/code>. I want to resize the tensor to <code>(batch_size, 256, 256, 32, 1)<\/code> using interpolation.<\/p>\n\n<p>So my initial thought would be like customizing a keras layer with <code>scipy.ndimage.interpolation.zoom<\/code>.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class MyLayer(Layer):\n\n    def __init__(self, output_dim, **kwargs):\n        self.output_dim = output_dim\n        super(MyLayer, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        super(MyLayer, self).build(input_shape)  # Be sure to call this at the end\n\n    def call(self, x):\n        return scipy.ndimage.interpolation.zoom(x, ...)\n\n    def compute_output_shape(self, input_shape):\n        return ...\n<\/code><\/pre>\n\n<p>Is there any better way for this? Or do I need to convert the <code>keras tensor<\/code> to numpy array and then convert it back as a tensor?<\/p>\n","answers":[]}
+{"QuestionId":57666449,"AnswerCount":0,"Tags":"<python><machine-learning><keras><neural-network>","CreationDate":"2019-08-27T00:50:02.633","AcceptedAnswerId":null,"OwnerUserId":11976935.0,"Title":"Changing learning rate of keras sequential model sgd optimizer not having expected result","Body":"<p>Changing the learning rate of stochastic gradient descent optimizer for keras sequential model does not have the expected effect on the weights after training.<\/p>\n\n<p>I have been trying to update the weights of a Keras sequential model based on an extremely small (4 samples and 3 features with binary labels) dataset after just one iteration over the whole dataset. Im doing this exercise because I am going to later manually perform backpropagation and show that my results match the computer. I am using the train_on_batch class to accomplish this. I want to emphasize the changes in weights after the one iteration done by train_on_batch but changing the learning rate does not seem to have a predictable effect on how big of a change the weights and biases experience. I tested what would happen if I set the learning rate to 0 and indeed I got an answer which makes no sense since the weights and biases should have just not changed. Shouldn't a change in learning rate be proportional to the changes in weight and bias values? Why are the weights changing even though I set the learning rate to zero?<\/p>\n\n<pre><code>model = Sequential([\n    Dense(2, input_shape=(3,), activation='sigmoid', kernel_initializer='ones', bias_initializer='zeros'), \n    Dense(1, activation='sigmoid', kernel_initializer='ones', bias_initializer='zeros')\n\n])\n\nsgd = keras.optimizers.SGD(lr=0)\n\nmodel.compile(\n  optimizer='sgd',\n  loss='mean_squared_error',\n  metrics=['accuracy'],\n)\n\nmodel.train_on_batch(\n  features, # training data\n  labels, # training targets\n)\n\n\nfor layer in model.layers:\n    weights = layer.get_weights()[0]\n    print(weights)\n\n[[1.        1.       ]\n [1.        1.       ]\n [0.9999999 0.9999999]]\n[[0.99839985]\n [0.99839985]]\n\nfor layer in model.layers:\n    biases = layer.get_weights()[1]\n    print(biases)\n\n[-7.547273e-09 -7.547273e-09]\n[-0.00160019]\n<\/code><\/pre>\n\n<p>Since weights are initialized at 1 and biases at zero, I would expect them to stay that way after I run train_on_batch since learn rate = 0. The fact that this is not the case reveals some underlying problem which I am not aware of.<\/p>\n","answers":[]}
+{"QuestionId":57666618,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-27T01:22:51.173","AcceptedAnswerId":null,"OwnerUserId":3654891.0,"Title":"InvalidArgumentError when serving model on CPU, no issues on GPU","Body":"<p>I'm trying to import a model I trained on a GPU ec2 instance to an ec2 instance that <strong>doesn't<\/strong> have a GPU for inference, and I'm having some issues when calling the <code>model.predict()<\/code> function.<\/p>\n\n<p>I've tried setting the device to only using the available CPU on the instance and that didn't help, and I'm pretty sure this is the problem as I've recreated my pipenv environment on a different ec2 instance with a GPU, ran the same code I have for inference, and had no issues. <\/p>\n\n<p>Here's my code for the model inference:<\/p>\n\n<pre><code>import keras\nfrom keras import backend as K\nimport tensorflow as tf\nimport pandas as pd\nimport numpy as np\n\n# A preprocessed pandas dataframe I read in\nf_listing_data = pd.read_csv(\"f_listing_data.csv\")\n\ndef df_to_array(df):\n    \"\"\"Converts a pandas dataframe to a list of numpy arrays\"\"\"\n    array_list = []\n    for col in df.columns:\n        array_list.append(df[col].values)\n    return array_list\n\ndef r2(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\ndef rmse(y_true, y_pred):\n        return K.sqrt(K.mean(K.square(y_pred - y_true)))\n\n# loading in my pretrained model\ninf_model = keras.models.load_model('keras_instant_price_no_upload.h5', custom_objects={'rmse':rmse, 'r2':r2})\n\n# dropping target label and getting predictions \npreds = inf_model.predict(df_to_array(f_listing_data.drop(\"price\", axis=1)))\n<\/code><\/pre>\n\n<p>Here's what my pipenv Pipfile looks like for my environment setup (I'm using packages downloaded from a private Pypi server):<\/p>\n\n<pre><code>[[source]]\nname = \"pypi\"\nurl = \"http:\/\/$PYPI_USERNAME:${PYPI_PASSWORD}@{SERVER_IP}:{SERVER_PORT}\/\"\nverify_ssl = true\n\n[dev-packages]\n\n[packages]\nscikit-learn = \"*\"\nxgboost = \"*\"\npandas = \"*\"\nnumpy = \"*\"\ntensorflow = \"*\"\nipython = \"*\"\nkeras = \"*\"\nloftydatafetcher = \"*\"  # a custom package for downloading\/uploading data\n\n[requires]\npython_version = \"3.6\"\n<\/code><\/pre>\n\n<p>My Pipenv graph:<\/p>\n\n<pre><code>ipython==7.7.0\n  - backcall [required: Any, installed: 0.1.0]\n  - decorator [required: Any, installed: 4.4.0]\n  - jedi [required: &gt;=0.10, installed: 0.15.1]\n    - parso [required: &gt;=0.5.0, installed: 0.5.1]\n  - pexpect [required: Any, installed: 4.7.0]\n    - ptyprocess [required: &gt;=0.5, installed: 0.6.0]\n  - pickleshare [required: Any, installed: 0.7.5]\n  - prompt-toolkit [required: &gt;=2.0.0,&lt;2.1.0, installed: 2.0.9]\n    - six [required: &gt;=1.9.0, installed: 1.12.0]\n    - wcwidth [required: Any, installed: 0.1.7]\n  - pygments [required: Any, installed: 2.4.2]\n  - setuptools [required: &gt;=18.5, installed: 41.2.0]\n  - traitlets [required: &gt;=4.2, installed: 4.3.2]\n    - decorator [required: Any, installed: 4.4.0]\n    - ipython-genutils [required: Any, installed: 0.2.0]\n    - six [required: Any, installed: 1.12.0]\nKeras==2.2.5\n  - h5py [required: Any, installed: 2.9.0]\n    - numpy [required: &gt;=1.7, installed: 1.17.0]\n    - six [required: Any, installed: 1.12.0]\n  - keras-applications [required: &gt;=1.0.8, installed: 1.0.8]\n    - h5py [required: Any, installed: 2.9.0]\n      - numpy [required: &gt;=1.7, installed: 1.17.0]\n      - six [required: Any, installed: 1.12.0]\n    - numpy [required: &gt;=1.9.1, installed: 1.17.0]\n  - keras-preprocessing [required: &gt;=1.1.0, installed: 1.1.0]\n    - numpy [required: &gt;=1.9.1, installed: 1.17.0]\n    - six [required: &gt;=1.9.0, installed: 1.12.0]\n  - numpy [required: &gt;=1.9.1, installed: 1.17.0]\n  - pyyaml [required: Any, installed: 5.1.2]\n  - scipy [required: &gt;=0.14, installed: 1.3.1]\n    - numpy [required: &gt;=1.13.3, installed: 1.17.0]\n  - six [required: &gt;=1.9.0, installed: 1.12.0]\nloftydatafetcher==0.2.3\n  - boto3 [required: Any, installed: 1.9.216]\n    - botocore [required: &gt;=1.12.216,&lt;1.13.0, installed: 1.12.216]\n      - docutils [required: &gt;=0.10,&lt;0.16, installed: 0.15.2]\n      - jmespath [required: &gt;=0.7.1,&lt;1.0.0, installed: 0.9.4]\n      - python-dateutil [required: &gt;=2.1,&lt;3.0.0, installed: 2.8.0]\n        - six [required: &gt;=1.5, installed: 1.12.0]\n      - urllib3 [required: &gt;=1.20,&lt;1.26, installed: 1.25.3]\n    - jmespath [required: &gt;=0.7.1,&lt;1.0.0, installed: 0.9.4]\n    - s3transfer [required: &gt;=0.2.0,&lt;0.3.0, installed: 0.2.1]\n      - botocore [required: &gt;=1.12.36,&lt;2.0.0, installed: 1.12.216]\n        - docutils [required: &gt;=0.10,&lt;0.16, installed: 0.15.2]\n        - jmespath [required: &gt;=0.7.1,&lt;1.0.0, installed: 0.9.4]\n        - python-dateutil [required: &gt;=2.1,&lt;3.0.0, installed: 2.8.0]\n          - six [required: &gt;=1.5, installed: 1.12.0]\n        - urllib3 [required: &gt;=1.20,&lt;1.26, installed: 1.25.3]\n  - mysql-connector [required: Any, installed: 2.2.9]\n  - pandas [required: Any, installed: 0.25.1]\n    - numpy [required: &gt;=1.13.3, installed: 1.17.0]\n    - python-dateutil [required: &gt;=2.6.1, installed: 2.8.0]\n      - six [required: &gt;=1.5, installed: 1.12.0]\n    - pytz [required: &gt;=2017.2, installed: 2019.2]\n  - sqlalchemy [required: Any, installed: 1.3.7]\nscikit-learn==0.21.3\n  - joblib [required: &gt;=0.11, installed: 0.13.2]\n  - numpy [required: &gt;=1.11.0, installed: 1.17.0]\n  - scipy [required: &gt;=0.17.0, installed: 1.3.1]\n    - numpy [required: &gt;=1.13.3, installed: 1.17.0]\ntensorflow==1.14.0\n  - absl-py [required: &gt;=0.7.0, installed: 0.7.1]\n    - six [required: Any, installed: 1.12.0]\n  - astor [required: &gt;=0.6.0, installed: 0.8.0]\n  - gast [required: &gt;=0.2.0, installed: 0.2.2]\n  - google-pasta [required: &gt;=0.1.6, installed: 0.1.7]\n  - grpcio [required: &gt;=1.8.6, installed: 1.23.0]\n    - six [required: &gt;=1.5.2, installed: 1.12.0]\n  - keras-applications [required: &gt;=1.0.6, installed: 1.0.8]\n    - h5py [required: Any, installed: 2.9.0]\n      - numpy [required: &gt;=1.7, installed: 1.17.0]\n      - six [required: Any, installed: 1.12.0]\n    - numpy [required: &gt;=1.9.1, installed: 1.17.0]\n  - keras-preprocessing [required: &gt;=1.0.5, installed: 1.1.0]\n    - numpy [required: &gt;=1.9.1, installed: 1.17.0]\n    - six [required: &gt;=1.9.0, installed: 1.12.0]\n  - numpy [required: &gt;=1.14.5,&lt;2.0, installed: 1.17.0]\n  - protobuf [required: &gt;=3.6.1, installed: 3.9.1]\n    - setuptools [required: Any, installed: 41.2.0]\n    - six [required: &gt;=1.9, installed: 1.12.0]\n  - six [required: &gt;=1.10.0, installed: 1.12.0]\n  - tensorboard [required: &gt;=1.14.0,&lt;1.15.0, installed: 1.14.0]\n    - absl-py [required: &gt;=0.4, installed: 0.7.1]\n      - six [required: Any, installed: 1.12.0]\n    - grpcio [required: &gt;=1.6.3, installed: 1.23.0]\n      - six [required: &gt;=1.5.2, installed: 1.12.0]\n    - markdown [required: &gt;=2.6.8, installed: 3.1.1]\n      - setuptools [required: &gt;=36, installed: 41.2.0]\n    - numpy [required: &gt;=1.12.0, installed: 1.17.0]\n    - protobuf [required: &gt;=3.6.0, installed: 3.9.1]\n      - setuptools [required: Any, installed: 41.2.0]\n      - six [required: &gt;=1.9, installed: 1.12.0]\n    - setuptools [required: &gt;=41.0.0, installed: 41.2.0]\n    - six [required: &gt;=1.10.0, installed: 1.12.0]\n    - werkzeug [required: &gt;=0.11.15, installed: 0.15.5]\n    - wheel [required: &gt;=0.26, installed: 0.33.6]\n  - tensorflow-estimator [required: &gt;=1.14.0rc0,&lt;1.15.0rc0, installed: 1.14.0]\n  - termcolor [required: &gt;=1.1.0, installed: 1.1.0]\n  - wheel [required: &gt;=0.26, installed: 0.33.6]\n  - wrapt [required: &gt;=1.11.1, installed: 1.11.2]\nxgboost==0.90\n  - numpy [required: Any, installed: 1.17.0]\n  - scipy [required: Any, installed: 1.3.1]\n    - numpy [required: &gt;=1.13.3, installed: 1.17.0]\n<\/code><\/pre>\n\n<p>I expect the results of the predictions to be an array 1 float value per row in the data I pass into the <code>model.predict()<\/code> function. These are the results I get on my GPU instance (I should mention that I'm running this code from a Jupyter Notebook):<\/p>\n\n<pre><code>&gt;&gt;&gt; preds = inf_model.predict(df_to_array(f_listing_data.drop(\"price\", axis=1)))\n&gt;&gt;&gt; preds[:5], len(preds)\n\n(array([[0.18931788],\n        [0.08384943],\n        [0.48033524],\n        [0.19470096],\n        [0.16680539]], dtype=float32), 340)\n<\/code><\/pre>\n\n<p>This the error stack trace I get when running this same line of code on my CPU instance (running in ipython):<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nInvalidArgumentError                      Traceback (most recent call last)\n&lt;ipython-input-3-6559eecde46a&gt; in &lt;module&gt;\n     27 \n     28 # dropping target label and getting predictions\n---&gt; 29 preds = inf_model.predict(df_to_array(f_listing_data.drop(\"price\", axis=1)))\n\n~\/.local\/share\/virtualenvs\/xgb_listing_pred-eH2Cah04\/lib\/python3.6\/site-packages\/keras\/engine\/training.py in predict(self, x, batch_size, verbose, steps, callbacks, max_queue_size, workers, use_multiprocessing)\n   1399                                             verbose=verbose,\n   1400                                             steps=steps,\n-&gt; 1401                                             callbacks=callbacks)\n   1402 \n   1403     def train_on_batch(self, x, y,\n\n~\/.local\/share\/virtualenvs\/xgb_listing_pred-eH2Cah04\/lib\/python3.6\/site-packages\/keras\/engine\/training_arrays.py in predict_loop(model, f, ins, batch_size, verbose, steps, callbacks)\n    330             batch_logs = {'batch': batch_index, 'size': len(batch_ids)}\n    331             callbacks._call_batch_hook('predict', 'begin', batch_index, batch_logs)\n--&gt; 332             batch_outs = f(ins_batch)\n    333             batch_outs = to_list(batch_outs)\n    334             if batch_index == 0:\n\n~\/.local\/share\/virtualenvs\/xgb_listing_pred-eH2Cah04\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py in __call__(self, inputs)\n   2977                     return self._legacy_call(inputs)\n   2978 \n-&gt; 2979             return self._call(inputs)\n   2980         else:\n   2981             if py_any(is_tensor(x) for x in inputs):\n\n~\/.local\/share\/virtualenvs\/xgb_listing_pred-eH2Cah04\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py in _call(self, inputs)\n   2935             fetched = self._callable_fn(*array_vals, run_metadata=self.run_metadata)\n   2936         else:\n-&gt; 2937             fetched = self._callable_fn(*array_vals)\n   2938         return fetched[:len(self.outputs)]\n   2939 \n\n~\/.local\/share\/virtualenvs\/xgb_listing_pred-eH2Cah04\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py in __call__(self, *args, **kwargs)\n   1456         ret = tf_session.TF_SessionRunCallable(self._session._session,\n   1457                                                self._handle, args,\n-&gt; 1458                                                run_metadata_ptr)\n   1459         if run_metadata:\n   1460           proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\nInvalidArgumentError: indices[0,0] = 1490 is not in [0, 38)\n         [[{{node embedding_22\/embedding_lookup}}]]\n\n<\/code><\/pre>\n\n<p>Any help\/advice is greatly appreciated, thank you for your time.<\/p>\n","answers":[]}
+{"QuestionId":57666829,"AnswerCount":1,"Tags":"<python-3.x><keras><google-colaboratory>","CreationDate":"2019-08-27T01:59:40.793","AcceptedAnswerId":null,"OwnerUserId":11955299.0,"Title":"FileNotFoundError: [Errno 2] No such file or directory -- even if I'm using the full path","Body":"<p>I'm using Python 3 in Google Colab. I keep getting the error \"FileNotFoundError: [Errno 2] No such file or directory\" even if I'm pretty sure that the path I placed in was correct and verified that the folders are there.<\/p>\n\n<p>I've used \"pip install keras\" and \"pip install tensorflow\" initially. I've tried using \\ instead of \\ for the path.<\/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 = '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 = ['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('E:\\Project\\dataset\\training_set',\ntarget_size = (64, 64),\nbatch_size = 32,\nclass_mode = 'binary')\ntest_set = test_datagen.flow_from_directory('E:\\Project\\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('E:\\Project\\dataset\\single_prediction\\cat_or_dog_1.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:\n    prediction = 'dog'\nelse:\n    prediction = 'cat'\n\nFileNotFoundError                         Traceback (most recent call last)\n&lt;ipython-input-7-757ca18e13d5&gt; in &lt;module&gt;()\n     30 target_size = (64, 64),\n     31 batch_size = 32,\n---&gt; 32 class_mode = 'binary')\n     33 test_set = test_datagen.flow_from_directory('E:\\Project\\dataset\\test_set',\n     34 target_size = (64, 64),\n\n1 frames\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras_preprocessing\/image\/directory_iterator.py in __init__(self, directory, image_data_generator, 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    104         if not classes:\n    105             classes = []\n--&gt; 106             for subdir in sorted(os.listdir(directory)):\n    107                 if os.path.isdir(os.path.join(directory, subdir)):\n    108                     classes.append(subdir)\n\nFileNotFoundError: [Errno 2] No such file or directory: 'E:\\\\Project\\\\dataset\\training_set'\n<\/code><\/pre>\n","answers":[{"AnswerId":"57666861","CreationDate":"2019-08-27T02:05:15.963","ParentId":null,"OwnerUserId":"11945766","Title":null,"Body":"<p>I think you cannot directly access files from your local machine in Google colab notebook. You have to upload your files to Colab either through Google Drive or direct Upload from your local computer.<\/p>\n\n<p>You can refer <a href=\"https:\/\/towardsdatascience.com\/3-ways-to-load-csv-files-into-colab-7c14fcbdcb92\" rel=\"nofollow noreferrer\">this link<\/a> for more details.<\/p>\n"}]}
+{"QuestionId":57667018,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><neural-network><linear-regression>","CreationDate":"2019-08-27T02:32:12.350","AcceptedAnswerId":null,"OwnerUserId":7866576.0,"Title":"\"InvalidArgumentError\" when trying to do linear regression with Tensorflow in google colab py 3 notebook","Body":"<p>Trying to run linear regression on california housing dataset with Tensorflow. When working with placeholders I get an \"InvalidArgumentError: You must feed a value for placeholder tensor 'y' with dtype float and shape [?,1]<br>\n[[{{node y}}]]\". Code is below, please help! <\/p>\n\n<p>I assumed that for some reason my y_batch or x_batch isn't a float32. So I added a tf.cast(..., tf.float32). That hasn't solved the problem. Also added a tf.reset_default_graph. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Construction\ntf.reset_default_graph()\nx = tf.placeholder(dtype = tf.float32, shape = (None, n + 1), name = \"x\")\ny = tf.placeholder(dtype = tf.float32, shape = (None, 1), name = \"y\")\ntheta = tf.Variable(tf.random_uniform(shape=[n + 1, 1], minval=-1.0, maxval=1.0), dtype = tf.float32, name = \"theta\")\ny_pred = tf.matmul(x, theta, name = \"predictions\")\nerror = y_pred - y\nmse = tf.reduce_mean(tf.square(error), name=\"mse\")\ngrad = tf.gradients(mse, [theta])[0]\noptimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate)\ntraining_op = optimizer.minimize(mse)\ninit = tf.global_variables_initializer()\n\nn_epochs = 1000\nbatch_size = 100\nn_batches = np.int(np.ceil(m\/ batch_size))\na = 0\nb = 0\n\ndef fetch_batch(epoch, batch_index, batch_size):\n  a = batch_size * batch_index\n  b = a + batch_size\n  x_batch = tf.cast(scaled_housing_data_bias[a:b,:], tf.float32)\n  y_batch = tf.cast(housing_target[a:b,:], tf.float32)\n  return x_batch, y_batch\n\n\n# dtype and shape checking\na, b = fetch_batch(1, 3, 100)\nprint(a.shape)\nprint(b.shape)\nprint(a.dtype)\nprint(b.dtype)\n\n# (100, 9)\n# (100, 1)\n# &lt;dtype: 'float32'&gt;\n# &lt;dtype: 'float32'&gt;\n\n\n# execution\nwith tf.Session() as sess:\n  sess.run(init)\n  for epoch in range(n_epochs):\n    rmse = mse.eval()\n    if epoch_num % 100 == 0: \n      print(\"\\n Epoch: \", epoch_num, \"\\t MSE: \", rmse)\n    for batch_index in range(n_batches):\n      x_batch, y_batch = fetch_batch(epoch, batch_index, batch_size)\n      sess.run(training_op, feed_dict = {x: x_batch, y: y_batch})\n      best_theta = theta.eval()\nprint(\"Best_theta: \\n\", best_theta)\nprint(\"\\n Final rmse: \", rmse)\n<\/code><\/pre>\n\n<hr>\n\n<pre class=\"lang-py prettyprint-override\"><code>InvalidArgumentError                      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   1355     try:\n-&gt; 1356       return fn(*args)\n   1357     except errors.OpError as e:\n\n8 frames\nInvalidArgumentError: You must feed a value for placeholder tensor 'y' with dtype float and shape [?,1]\n     [[{{node y}}]]\n\nDuring handling of the above exception, another exception occurred:\n\nInvalidArgumentError                      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   1368           pass\n   1369       message = error_interpolation.interpolate(message, self._graph)\n-&gt; 1370       raise type(e)(node_def, op, message)\n   1371 \n   1372   def _extend_graph(self):\n\nInvalidArgumentError: You must feed a value for placeholder tensor 'y' with dtype float and shape [?,1]\n     [[node y (defined at &lt;ipython-input-16-ccae920efe43&gt;:3) ]]\n\nOriginal stack trace for 'y':\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\/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 450, in _handle_events\n    self._handle_recv()\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/zmq\/eventloop\/zmqstream.py\", line 480, in _handle_recv\n    self._run_callback(callback, msg)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/zmq\/eventloop\/zmqstream.py\", line 432, 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-16-ccae920efe43&gt;\", line 3, in &lt;module&gt;\n    y = tf.placeholder(dtype = tf.float32, shape = (None, 1), name = \"y\")\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/array_ops.py\", line 2143, in placeholder\n    return gen_array_ops.placeholder(dtype=dtype, shape=shape, name=name)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/gen_array_ops.py\", line 6262, in placeholder\n    \"Placeholder\", dtype=dtype, shape=shape, name=name)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 788, in _apply_op_helper\n    op_def=op_def)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/util\/deprecation.py\", line 507, in new_func\n    return func(*args, **kwargs)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 3616, in create_op\n    op_def=op_def)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 2005, in __init__\n    self._traceback = tf_stack.extract_stack()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57667103,"AnswerCount":0,"Tags":"<python><numpy><pytorch>","CreationDate":"2019-08-27T02:49:11.913","AcceptedAnswerId":null,"OwnerUserId":11981370.0,"Title":"why tensor.numpy() outputs different value?","Body":"<p>I used tensor.numpy() on a feature matrix and a row of this matrix for comparison. I found the value is not the same for the row output and the corresponding row of the matrix. Anyone knows why?<\/p>\n\n<pre><code>def consine_distance_byrow(f1, f2):\n    feat1 = f1.view(-1, 1)\n    dists = torch.mm(f2, feat1)\n    dists = dists.numpy()\n    return dists\n\ndef consine_distance(f1, f2):\n    feat1 = torch.transpose(f1, 0, 1)\n    dists = torch.mm(f2, feat1)\n    dists = dists.numpy()\n    return dists\n\n# load trainingset data\nfeature_path = '..\/..\/cnn\/pytorch_result_train_val_' + str(self.feature_nums) + '.mat'\nresult = scipy.io.loadmat(feature_path)\ntrain_feature = torch.FloatTensor(result['train_f']) # M x 512d\ndists = consine_distance(train_feature, train_feature) # way1\nfor i in range(len(train_label)):\n    distsT = consine_distance_byrow(train_feature[i], train_feature) # way2\n    # here, dists[:,i] is slightly different with distsT\n<\/code><\/pre>\n\n<p>e.g.<\/p>\n\n<pre><code>dists[:0] = [ 1.0000001  -0.15086517 -0.08085391 ... -0.05950543 -0.03058994\n  0.10164267]\n\n\nwhile i=0, distT = [[ 0.99999964]\n [-0.15086518]\n [-0.0808539 ]\n ...\n [-0.05950541]\n [-0.03058996]\n [ 0.1016427 ]]\n<\/code><\/pre>\n\n<p>apparently, there is a slightly difference.<\/p>\n\n<p>and this is caused after tensor.numpy(), before this step the two tensors are the same<\/p>\n","answers":[]}
+{"QuestionId":57667146,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-27T02:56:04.893","AcceptedAnswerId":null,"OwnerUserId":11354575.0,"Title":"keras custom layer error: InvalidArgumentError: You must feed a value for placeholder tensor","Body":"<p>I am writing a custom layer with <code>keras<\/code><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nfrom scipy.ndimage.interpolation import zoom\nfrom keras.layers import Layer\nimport keras.backend as K\n\nclass resample3d(Layer):\n    def __init__(self, output_dim, org_shape, **kwargs):\n        self.x_factor = output_dim[0] \/ org_shape[0]\n        self.y_factor = output_dim[1] \/ org_shape[1]\n        self.z_factor = output_dim[2] \/ org_shape[2]\n        super(resample3d, self).__init__(**kwargs)\n\n    def call(self, x):\n        np_array = K.eval(x)\n        np_array = np.squeeze(np_array, axis=-1)\n        resampled_array = zoom(np_array, (x_factor, y_factor, z_factor))\n        resampled_array = np.expand_dims(resampled_array, axis=-1)\n        return K.variable(resampled_array)\n<\/code><\/pre>\n\n<p>But when I tried to instantialize it,<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>inputs = Input(shape=(3, 3, 3, 1))\nx = resample3d(output_dim=(6, 6, 6, 1), org_shape=(3, 3, 3, 1))(inputs)\n<\/code><\/pre>\n\n<p>An error raises,<\/p>\n\n<pre><code>Caused by op 'input_1', defined at:\n  File \"custom_layer.py\", line 28, in &lt;module&gt;\n    inputs = Input(shape=(3, 3, 3, 1))\n  File \"D:\\anaconda\\lib\\site-packages\\keras\\engine\\input_layer.py\", line 178, in Input\n    input_tensor=tensor)\n  File \"D:\\anaconda\\lib\\site-packages\\keras\\legacy\\interfaces.py\", line 91, in wrapper\n    return func(*args, **kwargs)\n  File \"D:\\anaconda\\lib\\site-packages\\keras\\engine\\input_layer.py\", line 87, in __init__\n    name=self.name)\n  File \"D:\\anaconda\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py\", line 517, in placeholder\n    x = tf.placeholder(dtype, shape=shape, name=name)\n  File \"D:\\anaconda\\lib\\site-packages\\tensorflow\\python\\ops\\array_ops.py\", line 2077, in placeholder\n    return gen_array_ops.placeholder(dtype=dtype, shape=shape, name=name)\n  File \"D:\\anaconda\\lib\\site-packages\\tensorflow\\python\\ops\\gen_array_ops.py\", line 6833, in placeholder\n    \"Placeholder\", dtype=dtype, shape=shape, name=name)\n  File \"D:\\anaconda\\lib\\site-packages\\tensorflow\\python\\framework\\op_def_library.py\", line 788, in _apply_op_helper\n    op_def=op_def)\n  File \"D:\\anaconda\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py\", line 507, in new_func\n    return func(*args, **kwargs)\n  File \"D:\\anaconda\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3300, in create_op\n    op_def=op_def)\n  File \"D:\\anaconda\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 1801, in __init__\n    self._traceback = tf_stack.extract_stack()\n\nInvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'input_1' with dtype float and shape [?,3,3,3,1]\n         [[node input_1 (defined at D:\\anaconda\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py:517) ]]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57667265,"AnswerCount":1,"Tags":"<python><tensorflow><keras><image-segmentation><loss-function>","CreationDate":"2019-08-27T03:15:58.550","AcceptedAnswerId":null,"OwnerUserId":8262535.0,"Title":"How to properly use CategoricalCrossentropy loss for image segmentation in Tensorflow 2.0\/Keras?","Body":"<p>I am trying to train a U-Net derivative to do single-class image segmentation but am having problems using the <code>tf.keras.losses.SparseCategoricalCrossentropy()<\/code> and <code>tf.keras.losses.CategoricalCrossentropy()<\/code> functions in Keras. Which is the more appropriate and how to use it properly?<\/p>\n\n<p>If I try to use <code>SpareCategoricalCrossentropy<\/code>, I get the error:<\/p>\n\n<blockquote>\n  <p>Received a label value of 1 which is outside the valid range of [0, 1) <\/p>\n<\/blockquote>\n\n<p>If I try to use <code>CategoricalCrossentropy<\/code>, I get:<\/p>\n\n<blockquote>\n  <p>You are passing a target array of shape <code>(3600, 64, 64, 1)<\/code> while using as loss <code>categorical_crossentropy<\/code>. <code>categorical_crossentropy<\/code> expects targets to be binary matrices (1s and 0s) of shape (samples, classes). If your targets are integer classes, you can convert them to the expected format via: <code>y_binary = tf.keras.utils.to_categorical(y_int)<\/code><\/p>\n<\/blockquote>\n\n<p>Using <code>to_categorical<\/code> for my mask vs background segmentation problem, it increases the last dimension to 2, which should not be necessary. My prediction should be a number between 0 and 1 in a single \"channel\". <\/p>\n\n<p>Model definition snippet:<\/p>\n\n<pre><code>input_x = tf.keras.Input(batch_shape=(batch_size, xsze, ysze, 3), name='input_x') \npredictions = tf.keras.layers.Conv2D(1, [1, 1], activation='linear', name='output_x')(drop11) \n\nloss = tf.keras.losses.SparseCategoricalCrossentropy()\nmodel.compile(optimizer=tf.keras.optimizers.Adam(),  # Optimizer\n              loss=loss, \n              metrics=[tf.keras.metrics.SparseCategoricalAccuracy()]) \n\ncheckpointer = tf.keras.callbacks.ModelCheckpoint(session_name + '_backup.h5', save_best_only=True, monitor = 'acc', verbose = 0)\nearly_stopper = tf.keras.callbacks.EarlyStopping(monitor='val_loss',patience=5, verbose=1,min_delta=0.005)\nhistory = model.fit(data_train, roi_train,\n                    batch_size=batch_size,\n                    epochs = 10,\n                    validation_data=(data_val, roi_zoom_val),callbacks=[checkpointer,early_stopper]) \n<\/code><\/pre>\n\n<p>My <code>roi_train<\/code> is a numpy array with 0's and 1's of type <code>float32<\/code>.<\/p>\n","answers":[{"AnswerId":"57688677","CreationDate":"2019-08-28T09:02:09.877","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Since you only have one class and you want to each value in the segmentation map to be between 0 and 1, then you should use <code>sigmoid<\/code> as the activation of last layer and <code>binary_crossentropy<\/code> as the loss function. That's because for each pixel you are facing a binary decision: is this pixel belong to foreground or background?<\/p>\n"}]}
+{"QuestionId":57667304,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-08-27T03:23:02.303","AcceptedAnswerId":null,"OwnerUserId":10855529.0,"Title":"What is the difference between feature and classifier?","Body":"<p>I saw code like this<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>self.feature = model_func()\nif loss_type == 'softmax':\n    self.classifier = nn.Linear(self.feature.final_feat_dim, num_class)\n    self.classifier.bias.data.fill_(0)\nelif loss_type == 'dist': #Baseline ++\n    self.classifier = backbone.distLinear(self.feature.final_feat_dim, num_class)\n<\/code><\/pre>\n\n<p>where <code>model_func<\/code> is a ConvNet 4\/6 or ResNet 10\/18\/34\/101<\/p>\n\n<p>What here is classifier?<\/p>\n\n<p>I know that in neural networks we have parameters that we learn, buffers that are used to store something that gets updated during training, activations that are the results after each layer.<\/p>\n\n<p>Is feature same as activation, and what is a classifier, where is the end of features, and the beginning of classifier in a neural network? Is the result of a classifier also activation?<\/p>\n","answers":[{"AnswerId":"57667433","CreationDate":"2019-08-27T03:47:15.447","ParentId":null,"OwnerUserId":"5386951","Title":null,"Body":"<p>I find the question a little messy, but I'll give my best from what I understand you're asking.<\/p>\n\n<blockquote>\n  <p>What here is classifier?<\/p>\n<\/blockquote>\n\n<p>The classifier would be the model itself. The model is the one who will, after being trained, be able to classify new data. <\/p>\n\n<blockquote>\n  <p>Is feature same as activation<\/p>\n<\/blockquote>\n\n<p>I don't know what kind of feature you have in mind. In data science context, a feature is understood to be one of the variables of the data one has. For example, if you have a dataset about houses, you may have features such as latitude, long., if it has a pool, how many bedrooms it has, etc.<\/p>\n\n<p>Activation functions are mathematical equations that determine the output of a neural network. The function is attached to each neuron in the network, and determines whether it should be activated (\u201cfired\u201d) or not, based on whether each neuron\u2019s input is relevant for the model\u2019s prediction. [1]<\/p>\n\n<p>I'm not sure I'm truly understanding what you're asking. <\/p>\n\n<blockquote>\n  <p>Is the result of a classifier also activation?<\/p>\n<\/blockquote>\n\n<p>The result of a classifier is the label, the class to which each data point belongs. Activation functions are used by neural networks in the process of classifying.<\/p>\n\n<hr>\n\n<p>Hope this helps!<\/p>\n\n<p>[1] <a href=\"https:\/\/missinglink.ai\/guides\/neural-network-concepts\/7-types-neural-network-activation-functions-right\/\" rel=\"nofollow noreferrer\">https:\/\/missinglink.ai\/guides\/neural-network-concepts\/7-types-neural-network-activation-functions-right\/<\/a><\/p>\n"}]}
+{"QuestionId":57668052,"AnswerCount":1,"Tags":"<python-3.x><keras>","CreationDate":"2019-08-27T05:16:44.907","AcceptedAnswerId":57668311.0,"OwnerUserId":7215349.0,"Title":"ModelCheckpoint - save_best_only=True","Body":"<p>I need explanation of save best only option of ModelCheckpoint. If I have a code like this<\/p>\n\n<pre><code>model.compile(optimizer='adam', loss='mse', metrics=['accuracy'])\ncp = [ModelCheckpoint(filepath=path+\"\/model-lstmMulti\", verbose=1, save_best_only=True)]        \nhistory_callback = model.fit(X, y, epochs=350, verbose=1, callbacks=cp)\n<\/code><\/pre>\n\n<p>and then I want to see the accuracy of that best model:<\/p>\n\n<pre><code>acc_history = history_callback.history[\"acc\"]\nnp.savetxt(path+\"\/acc_history.txt\", np.asarray(acc_history)) \n<\/code><\/pre>\n\n<p>I got the array ie. accuracy of the models of all epochs. Why I don't get only one value - accuracy of the best model?<\/p>\n","answers":[{"AnswerId":"57668311","CreationDate":"2019-08-27T05:44:48.207","ParentId":null,"OwnerUserId":"10371987","Title":null,"Body":"<p><code>ModelCheckpoint<\/code> is a callback function used to save model file (h5) after epochs. It doesn't affect the return history of <code>fit()<\/code> method. Just use <code>np.max<\/code> to get the best acc from acc history will do your job.<\/p>\n"}]}
+{"QuestionId":57668368,"AnswerCount":1,"Tags":"<python><pytorch><tensor><gradient-descent>","CreationDate":"2019-08-27T05:51:31.357","AcceptedAnswerId":57668669.0,"OwnerUserId":5631202.0,"Title":"How to create torch.tensor object and to update only some of its elements?","Body":"<p>Let's say I want to create torch.tensor object of size [2,3] filled with random elements, and I intend to use this matrix in the network and optimize it's values. However, I want to update only some of the the values in the matrix.<\/p>\n\n<p>I know that it can be done for a tensor by setting up parameter <code>requires_grad<\/code> To <code>True<\/code> or <code>False<\/code>. However, the following code <\/p>\n\n<pre><code>z = torch.rand([2,3], requires_grad=True)\nz[-1][-1].requires_grad=False\n<\/code><\/pre>\n\n<p>does not work as expected<\/p>\n\n<pre><code>RuntimeError: you can only change requires_grad flags of leaf variables. If you want to use a computed variable in a subgraph that doesn't require differentiation use var_no_grad = var.detach().\n<\/code><\/pre>\n\n<p>How to fix this RuntimeError? How to initialize torch tensor and then define which elements there would have <code>requires_grad =True<\/code>?<\/p>\n\n<p>If I write code in a similar manner:<\/p>\n\n<pre><code>z = torch.rand([2,3], requires_grad=False)\nz[-1][-1].requires_grad=True\n<\/code><\/pre>\n\n<p>There will be no error, but no change of the requires_grad as well.<\/p>\n","answers":[{"AnswerId":"57668669","CreationDate":"2019-08-27T06:20:10.580","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It does not really make much sense to have a <em>single<\/em> tensor which <code>requires_grad<\/code> for only part of its entries.<br>\nWhy not have two <em>separate<\/em> tensors one that us updated (<code>requires_grad=True<\/code>) and another one fixed (<code>requires_grad=False<\/code>)? You can then merge them for computational ease:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>fixed = torch.rand([2, 3], require_grad=False)\nupd = torch.rand([2, 3], require_grad=True)\nmask = torch.tensor([[0, 1, 0], [1, 0, 1]], require_grad=False)  # how to combine the two\n# combine them using fixed \"mask\":\nz = mask * fixed + (1-mask) * upd\n<\/code><\/pre>\n\n<p>You can obviously have other methods of combining <code>fixed<\/code> and <code>upd<\/code> other than using a binary <code>mask<\/code>.<br>\nFor example, if <code>upd<\/code> occupies the first two columns of <code>z<\/code> and <code>fixed<\/code> the rest, then:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>fixed = torch.rand([2, 1], require_grad=False)\nupd = torch.rand([2, 2], require_grad=True)\n# combine them using concatination\nz = torch.cat((upd, fixed),dim=1)\n<\/code><\/pre>\n\n<p>Or, if you know the indices<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>fidx = torch.tensor([0, 2], dtype=torch.long)\nuidx = torch.tensor([1, 3, 4, 5], dtype=torch.long)\nfixed = torch.rand([2,], require_grad=False)\nupd = torch.rand([4,], require_grad=True)\nz = torch.empty([2, 3])\nz[fidx] = fixed \nz[uidx] = upd\n<\/code><\/pre>\n"}]}
+{"QuestionId":57668390,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-08-27T05:54:00.020","AcceptedAnswerId":null,"OwnerUserId":6703176.0,"Title":"Getting \"InvalidArgumentError\" while executing the code","Body":"<p>I am new to tensorflow. I am writing a code to execute a simple Multi Layer Perceptron code in python using tensorflow. While executing the code in python, I am getting the \"InvalidArgumentError\". Please find the code which I have written in the python and the error which the code in generating below.<\/p>\n\n<p>Plus in am using UBUNTU 19.04 OS, Python 3.7.4, TensorFlow version 1.14(GPU version) and Conda Version 4.7.11<\/p>\n\n<p>Thanks in Advance<\/p>\n\n<p>This is the code I have written. I am randomly generating data using numpy.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nnp.random.seed(0)\n\n# setting my data for experiment\nx1 = np.random.uniform(-100,100, 10000)\nx2 = np.random.uniform(-100,100, 10000)\nx3 = np.random.uniform(-100,100, 10000)\nx_data = np.append(x1.reshape((10000,1)),x2.reshape((10000,1)),axis = 1)\nx_data = np.append(x_data,x3.reshape((10000,1)),axis = 1)\ny_data = (8*x1 - 3*x2 + x3 + np.random.uniform(-5,5,10000)).reshape((10000,1))\ndel x1,x2,x3\n\n# Tensorflow code\n\nX = tf.placeholder(dtype = tf.float64, shape = (10, 3), name = 'X')\ny = tf.placeholder(dtype = tf.float64, shape = (10,1), name = 'y')\n\nW1 = tf.Variable(np.random.uniform(-0.5,0.5, size = (3,5)), name = 'W1')\nb1 = tf.Variable(np.random.uniform(-0.5,0.5, size = (1,5)), name = 'b1')\n\nh1 = tf.add(tf.matmul(X,W1) , b1)\n\nW2 = tf.Variable(np.random.uniform(-0.5,0.5, size = (5,1)), name = 'W2')\nb2 = tf.Variable(np.random.uniform(-0.5,0.5, size = (1,1)), name = 'b2')\n\npredict = tf.add(tf.matmul(h1,W2) , b2)\n\nloss = tf.reduce_sum(tf.square(y - predict))\n\noptimizer = tf.train.GradientDescentOptimizer(0.1).minimize(loss)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n\n    total = 10000\n    k = 10\n    for i in range(total\/\/k):\n        feed_dict = {X : x_data[i*k:(i+1)*k], y : y_data[i*k: (i+1)*k] }\n        sess.run(optimizer, feed_dict)\n        print(sess.run(loss))\n\n\n<\/code><\/pre>\n\n<pre><code>Traceback (most recent call last):\n\n  File \"&lt;ipython-input-1-859d9ba5b0e9&gt;\", line 1, in &lt;module&gt;\n    runfile('\/home\/devesh\/Project Work\/Code\/Tensorflow\/MLP_1.py', wdir='\/home\/devesh\/Project Work\/Code\/Tensorflow')\n\n  File \"\/home\/devesh\/anaconda3\/lib\/python3.7\/site-packages\/spyder_kernels\/customize\/spydercustomize.py\", line 827, in runfile\n    execfile(filename, namespace)\n\n  File \"\/home\/devesh\/anaconda3\/lib\/python3.7\/site-packages\/spyder_kernels\/customize\/spydercustomize.py\", line 110, in execfile\n    exec(compile(f.read(), filename, 'exec'), namespace)\n\n  File \"\/home\/devesh\/Project Work\/Code\/Tensorflow\/MLP_1.py\", line 42, in &lt;module&gt;\n    print(sess.run(loss))\n\n  File \"\/home\/devesh\/anaconda3\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 950, in run\n    run_metadata_ptr)\n\n  File \"\/home\/devesh\/anaconda3\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1173, in _run\n    feed_dict_tensor, options, run_metadata)\n\n  File \"\/home\/devesh\/anaconda3\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1350, in _do_run\n    run_metadata)\n\n  File \"\/home\/devesh\/anaconda3\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1370, in _do_call\n    raise type(e)(node_def, op, message)\n\nInvalidArgumentError: 2 root error(s) found.\n  (0) Invalid argument: You must feed a value for placeholder tensor 'y' with dtype double and shape [10,1]\n     [[node y (defined at \/home\/devesh\/Project Work\/Code\/Tensorflow\/MLP_1.py:18) ]]\n     [[Sum\/_5]]\n  (1) Invalid argument: You must feed a value for placeholder tensor 'y' with dtype double and shape [10,1]\n     [[node y (defined at \/home\/devesh\/Project Work\/Code\/Tensorflow\/MLP_1.py:18) ]]\n0 successful operations.\n0 derived errors ignored.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57668538,"AnswerCount":3,"Tags":"<python><pytorch><ubuntu-18.04>","CreationDate":"2019-08-27T06:08:38.913","AcceptedAnswerId":57680575.0,"OwnerUserId":653397.0,"Title":"Trouble installing PyTorch for CUDA 9.0","Body":"<p>I am trying to install <code>PyTorch<\/code> for <code>CUDA 9.0<\/code> (Ubuntu 18.04 LTS). I followed instructions as per <a href=\"https:\/\/pytorch.org\/get-started\/previous-versions\/\" rel=\"nofollow noreferrer\">official documentation<\/a>, which says <\/p>\n\n<p>Download the <code>whl<\/code> file with the desired version from the following html page:\n<a href=\"https:\/\/download.pytorch.org\/whl\/cu90\/torch_stable.html\" rel=\"nofollow noreferrer\">https:\/\/download.pytorch.org\/whl\/cu90\/torch_stable.html<\/a><\/p>\n\n<pre><code>$ pip install [downloaded file]\n<\/code><\/pre>\n\n<p>But above link is broken, I am unable to open it.<\/p>\n","answers":[{"AnswerId":"57669781","CreationDate":"2019-08-27T07:35:47.957","ParentId":null,"OwnerUserId":"5816293","Title":null,"Body":"<p>Use the link from <\/p>\n\n<p><a href=\"https:\/\/pytorch.org\/\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/<\/a><\/p>\n\n<p>Its the easiest and best way to get what you want.<\/p>\n\n<p>If cuda 9.0 is not working then try non cuda or cuda 9.2<\/p>\n"},{"AnswerId":"57676503","CreationDate":"2019-08-27T14:07:23.697","ParentId":null,"OwnerUserId":"8234169","Title":null,"Body":"<p>According to the <a href=\"https:\/\/pytorch.org\/get-started\/locally\/\" rel=\"nofollow noreferrer\">official documentation<\/a> you can install pytorch with command<\/p>\n\n<pre><code># for Python 3.x\npip3 install torch torchvision\n<\/code><\/pre>\n\n<p>or if you prefer Python 2<\/p>\n\n<pre><code># for Python 2.x\npip install torch torchvision\n<\/code><\/pre>\n"},{"AnswerId":"57680575","CreationDate":"2019-08-27T18:44:59.747","ParentId":null,"OwnerUserId":"653397","Title":null,"Body":"<p>It turns out <a href=\"https:\/\/download.pytorch.org\/whl\/cu90\/torch_stable.html\" rel=\"nofollow noreferrer\">https:\/\/download.pytorch.org\/whl\/cu90\/torch_stable.html<\/a> is a list of <code>&lt;a href&gt;<\/code> tags, which we can see by going into the source from the browser. We can then download the appropriate version and install via <code>pip install [downloaded file]<\/code>.<\/p>\n"}]}
+{"QuestionId":57668675,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-27T06:20:28.223","AcceptedAnswerId":null,"OwnerUserId":1733587.0,"Title":"Tensorflow Python: How to apply a dense layer consecutively across a tensor","Body":"<p>I have an input tensor shaped [?, 6, 2] and a window size of 3. I want to slide the window along my inputs, concatenate them, and run them through a dense layer that takes 6 inputs and has 1 output.<\/p>\n\n<p>It would run concatenated input values at indexes [0-2], [1-3], [2-4], [3-5] and the dense network outputs would be concatenated to create a new tensor of size 4.<\/p>\n\n<p>So [?, 6, 2] -> [?, 4]<\/p>\n\n<p>What is the best way to accomplish this?<\/p>\n\n<p>I feel like this can maybe be done with conv2d, but not really sure how to formulate it.<\/p>\n","answers":[]}
+{"QuestionId":57668913,"AnswerCount":0,"Tags":"<python><linux><numpy><tensorflow><pip>","CreationDate":"2019-08-27T06:39:10.797","AcceptedAnswerId":null,"OwnerUserId":9533336.0,"Title":"tensorflow does not work after doing several numpy uninstalls in pip3","Body":"<p>I have applied <code>pip3 uninstall numpy<\/code> multiple times, during uninstalling I got messages as<\/p>\n\n<pre><code>botocore 1.12.80 has requirement urllib3&lt;1.25,&gt;=1.20; python_version &gt;= \"3.4\", but you'll have urllib3 1.25.3 which is incompatible.                                                  \ntensorflow 1.10.0 has requirement numpy&lt;=1.14.5,&gt;=1.13.3, but you'll have numpy 1.17.0 which is incompatible.                                                                         \ntensorflow 1.10.0 has requirement setuptools&lt;=39.1.0, but you'll have setuptools 40.4.3 which is incompatible. \n<\/code><\/pre>\n\n<p>I continue uninstalling, after that I tried to install many numpy versions (1.13.3, 1.14.5, 1.16.2,), to solve the issue, but installing a version removes the last installed one. I have many tensorflow models and now they doesn't work.<\/p>\n\n<p>Here is the packages that I uninstalled: <\/p>\n\n<pre><code>Uninstalling numpy-1.13.3:                                                                                                                                                            \n  Would remove:                                                                                                                                                                       \n    \/usr\/lib\/python3\/dist-packages\/numpy                                                                                                                                              \n    \/usr\/lib\/python3\/dist-packages\/numpy-1.13.3.egg-info                                                                                                                              \nProceed (y\/n)? y                                                                                                                                                                      \n  Successfully uninstalled numpy-1.13.3   \n\nUninstalling numpy-1.14.5:                                                                                                                                                            \n  Would remove:                                                                                                                                                                       \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy-1.14.5.dist-info\/*                                                                                                                   \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/.libs\/libopenblasp-r0-39a31c03.2.18.so                                                                                               \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/multiarray_tests.cpython-36m-x86_64-linux-gnu.so                                                                                \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/operand_flag_tests.cpython-36m-x86_64-linux-gnu.so                                                                              \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/struct_ufunc_test.cpython-36m-x86_64-linux-gnu.so                                                                               \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/test_rational.cpython-36m-x86_64-linux-gnu.so                                                                                   \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/umath_tests.cpython-36m-x86_64-linux-gnu.so                                                                                     \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/testing\/nose_tools\/*                                                                                                                 \nProceed (y\/n)? y                                                                                                                                                                      \n  Successfully uninstalled numpy-1.14.5  \n\nUninstalling numpy-1.14.5:                                                                                                                                                            \n  Would remove:                                                                                                                                                                       \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy-1.14.5.dist-info\/*                                                                                                                   \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/.libs\/libopenblasp-r0-39a31c03.2.18.so                                                                                               \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/multiarray_tests.cpython-36m-x86_64-linux-gnu.so                                                                                \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/operand_flag_tests.cpython-36m-x86_64-linux-gnu.so                                                                              \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/struct_ufunc_test.cpython-36m-x86_64-linux-gnu.so                                                                               \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/test_rational.cpython-36m-x86_64-linux-gnu.so                                                                                   \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/umath_tests.cpython-36m-x86_64-linux-gnu.so                                                                                     \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/testing\/nose_tools\/*                                                                                                                 \nProceed (y\/n)? y \n\nUninstalling numpy-1.16.2:                                                                                                                                                            \n  Would remove:                                                                                                                                                                       \n    \/usr\/bin\/f2py                                                                                                                                                                     \n    \/usr\/bin\/f2py3                                                                                                                                                                    \n    \/usr\/bin\/f2py3.6                                                                                                                                                                  \n    \/usr\/local\/bin\/f2py                                                                                                                                                               \n    \/usr\/local\/bin\/f2py3                                                                                                                                                              \n    \/usr\/local\/bin\/f2py3.6                                                                                                                                                            \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy-1.16.2.dist-info\/*                                                                                                                   \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/*                                                                                                                                    \n  Would not remove (might be manually added):                                                                                                                                         \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/.libs\/libopenblasp-r0-39a31c03.2.18.so                                                                                               \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/multiarray_tests.cpython-36m-x86_64-linux-gnu.so                                                                                \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/operand_flag_tests.cpython-36m-x86_64-linux-gnu.so                                                                              \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/struct_ufunc_test.cpython-36m-x86_64-linux-gnu.so                                                                               \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/test_rational.cpython-36m-x86_64-linux-gnu.so                                                                                   \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/umath_tests.cpython-36m-x86_64-linux-gnu.so                                                                                     \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/testing\/nose_tools\/__init__.py                                                                                                       \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/testing\/nose_tools\/decorators.py                                                                                                     \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/testing\/nose_tools\/noseclasses.py                                                                                                    \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/testing\/nose_tools\/nosetester.py                                                                                                     \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/testing\/nose_tools\/parameterized.py                                                                                                  \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/testing\/nose_tools\/utils.py                                                                                                          \nProceed (y\/n)? y                                                                                                                                                                      \n  Successfully uninstalled numpy-1.16.2 \n\n  Uninstalling numpy-1.16.2:                                                                                                                                                            \n  Would remove:                                                                                                                                                                       \n    \/usr\/bin\/f2py                                                                                                                                                                     \n    \/usr\/bin\/f2py3                                                                                                                                                                    \n    \/usr\/bin\/f2py3.6                                                                                                                                                                  \n    \/usr\/local\/bin\/f2py                                                                                                                                                               \n    \/usr\/local\/bin\/f2py3                                                                                                                                                              \n    \/usr\/local\/bin\/f2py3.6                                                                                                                                                            \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy-1.16.2.dist-info\/*                                                                                                                   \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/*                                                                                                                                    \n  Would not remove (might be manually added):                                                                                                                                         \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/.libs\/libopenblasp-r0-39a31c03.2.18.so                                                                                               \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/multiarray_tests.cpython-36m-x86_64-linux-gnu.so                                                                                \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/operand_flag_tests.cpython-36m-x86_64-linux-gnu.so                                                                              \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/struct_ufunc_test.cpython-36m-x86_64-linux-gnu.so                                                                               \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/test_rational.cpython-36m-x86_64-linux-gnu.so                                                                                   \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/core\/umath_tests.cpython-36m-x86_64-linux-gnu.so                                                                                     \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/testing\/nose_tools\/__init__.py                                                                                                       \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/testing\/nose_tools\/decorators.py                                                                                                     \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/testing\/nose_tools\/noseclasses.py                                                                                                    \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/testing\/nose_tools\/nosetester.py                                                                                                     \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/testing\/nose_tools\/parameterized.py                                                                                                  \n    \/usr\/local\/lib\/python3.6\/dist-packages\/numpy\/testing\/nose_tools\/utils.py                                                                                                          \nProceed (y\/n)? y  \n\n\nUninstalling numpy-1.16.4:                                                                                                                                                            \n  Would remove:                                                                                                                                                                       \n    \/home\/alphamale\/.local\/bin\/f2py                                                                                                                                                   \n    \/home\/alphamale\/.local\/bin\/f2py3                                                                                                                                                  \n    \/home\/alphamale\/.local\/bin\/f2py3.6                                                                                                                                                \n    \/home\/alphamale\/.local\/lib\/python3.6\/site-packages\/numpy-1.16.4.dist-info\/*                                                                                                       \n    \/home\/alphamale\/.local\/lib\/python3.6\/site-packages\/numpy\/*                                                                                                                        \n  Would not remove (might be manually added):                                                                                                                                         \n    \/home\/alphamale\/.local\/lib\/python3.6\/site-packages\/numpy\/.libs\/libopenblasp-r0-382c8f3a.3.5.dev.so                                                                                \nProceed (y\/n)? y                                                                                                                                                                      \n  Successfully uninstalled numpy-1.16.4 \n\n<\/code><\/pre>\n\n<p>Is there a way to restore pip3 as it was before I did these uninstalls?<\/p>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>the issue was fixed as follows:\ninstalling <code>numpy==1.16.4<\/code>, then upgrading as <code>pip3 install --upgrade numpy==1.16.4<\/code><\/p>\n","answers":[]}
+{"QuestionId":57669168,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning><speech-recognition>","CreationDate":"2019-08-27T06:56:40.893","AcceptedAnswerId":null,"OwnerUserId":11982099.0,"Title":"I trained a neural network but I cannot find where it got saved, I cannot find any .meta , .index, .data files","Body":"<p>I was trying to follow this page <a href=\"https:\/\/www.tensorflow.org\/tutorials\/sequences\/audio_recognition\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/sequences\/audio_recognition<\/a><\/p>\n\n<p>I successfully executed the following command:<\/p>\n\n<p>python tensorflow\/examples\/speech_commands\/train.py<\/p>\n\n<p>I used a virtual environment in Anaconda. Used Tensorflow 14 and Python 3.6<\/p>\n\n<p>It took about about 22 hours to train it. it said \"\/tmp\/speech_commands_train\/conv.ckpt-100\" after every 100 iterations \n (there were 18000 in total) <\/p>\n\n<p>but now when I try to find conv.ckpt-18000.meta   or just speech_commands_train I cannot find it.<\/p>\n\n<p>I am very new to this. This is my first effort in deep learning.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/dJjkN.png\" rel=\"nofollow noreferrer\">how the terminal looked when training ended <\/a><\/p>\n","answers":[{"AnswerId":"57707215","CreationDate":"2019-08-29T09:33:39.577","ParentId":null,"OwnerUserId":"1616773","Title":null,"Body":"<p>Firstly, what you mean by \" Where It saved\", by it you mean logs, the trained model or weights.<\/p>\n\n<p>In your case, you are just storing the weights at given checkpoints hence you can acess them at given paths said in the tutorial <\/p>\n\n<pre><code>I0730 16:54:41.813438 55030 train.py:252] Saving to \"\/tmp\/speech_commands_train\/conv.ckpt-100\"\n\n*This is saving out the current trained weights to a checkpoint file. If your training script gets interrupted, you can look for the last saved checkpoint and then restart the script with -*\n<\/code><\/pre>\n\n<p>Also you can store logs using file writer and model using save_model or tensorboard callback with logdir.<\/p>\n\n<p>Don't forget to upvote if found it useful <\/p>\n"}]}
+{"QuestionId":57669222,"AnswerCount":1,"Tags":"<python><machine-learning><deep-learning><conv-neural-network><pytorch>","CreationDate":"2019-08-27T06:59:34.467","AcceptedAnswerId":null,"OwnerUserId":2611189.0,"Title":"How to improve cats and dogs classification using CNN with pytorch","Body":"<p>I tried to follow the architecture of CNN in this paper, ImageNet Classification with Deep Convolutional Neural Networks (<a href=\"https:\/\/papers.nips.cc\/paper\/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf\" rel=\"nofollow noreferrer\">https:\/\/papers.nips.cc\/paper\/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf<\/a>). In this paper, they tried to classify 1000 classes, whereas I am just trying to classify 2 classes.<\/p>\n\n<p>But, my test accuracy got stuck at 50%, and the model is not learning.<\/p>\n\n<p>I am training with 23K images of cats and dogs, and test with 2500 images.<\/p>\n\n<p>This is URL to my notebook <a href=\"https:\/\/github.com\/jinglescode\/workspace\/blob\/master\/my-journey-computer-vision\/codes\/Cats_and_Dogs.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/jinglescode\/workspace\/blob\/master\/my-journey-computer-vision\/codes\/Cats_and_Dogs.ipynb<\/a><\/p>\n\n<p>Could anyone advise what's wrong? What have I missed out? Willing to learn.<\/p>\n","answers":[{"AnswerId":"57670034","CreationDate":"2019-08-27T07:53:12.333","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<h1><a href=\"https:\/\/stackoverflow.com\/a\/57252898\/1714410\">Normalize your data<\/a>!!<\/h1>\n\n<p>For image data, you can use the <a href=\"https:\/\/github.com\/pytorch\/examples\/blob\/master\/imagenet\/main.py#L197-L198\" rel=\"nofollow noreferrer\">recommended <code>transform<\/code><\/a> to your train and test datasets<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],\n                                     std=[0.229, 0.224, 0.225])\ntrain_transforms = transforms.Compose([transforms.RandomRotation(30),\n                                       transforms.RandomResizedCrop(224),\n                                       transforms.RandomHorizontalFlip(),\n                                       transforms.ToTensor(),\n                                       normalize]) \n\ntest_transforms = transforms.Compose([transforms.Resize(255),\n                                      transforms.CenterCrop(224),\n                                      transforms.ToTensor(),\n                                      normalize])\n<\/code><\/pre>\n\n<hr>\n\n<p>Additional comments:<\/p>\n\n<ol>\n<li><p>If you want to \"peek\" into a dataloader, instantiating <code>next(iter(dataloader))<\/code> is <a href=\"https:\/\/stackoverflow.com\/q\/53280967\/1714410\">not a good idea<\/a>. Instead, you can access the dataset stored inside the dataloader and use its <code>__getitem__<\/code>:  <\/p>\n\n<pre><code>images, labels = dataloader.dataset[0]\n<\/code><\/pre><\/li>\n<li>If your training is \"stuck\", the usual first reaction is to change the learning rate<\/li>\n<\/ol>\n"}]}
+{"QuestionId":57669234,"AnswerCount":0,"Tags":"<python-3.x><pytorch><tensorboard><tensorboardx>","CreationDate":"2019-08-27T07:00:12.667","AcceptedAnswerId":null,"OwnerUserId":2913106.0,"Title":"How to display more than 10 images in Tensorboard?","Body":"<p>I noticed that it doesn't matter how many image I save to the tensorboard log file, tensorboard will only ever show 10 of them (per tag).<\/p>\n\n<blockquote>\n  <p>How can we increase the number of images or at least select which ones are displayed?<\/p>\n<\/blockquote>\n\n<hr>\n\n<p>To reproduce what I mean run following MCVE:<\/p>\n\n<pre><code>import torch\nfrom torch.utils.tensorboard import SummaryWriter\ntb = SummaryWriter(comment=\"test\")\n\nfor k in range(100):\n    # create an image with some funny pattern\n    b = [n for (n, c) in enumerate(bin(k)) if c == '1']\n    img = torch.zeros((1,10,10))\n    img[0, b, :] = 0.5\n    img =img + img.permute([0, 2, 1])\n\n    # add the image to the tensorboard file\n    tb.add_image(tag=\"test\", img_tensor=img, global_step=k)\n<\/code><\/pre>\n\n<p>This creates a folder <code>runs<\/code> in which the data is saved. From the same folder execute <code>tensorboard --logdir runs<\/code>, open the browser and go to <code>localhost:6006<\/code> (or replace <code>6006<\/code> with whatever port tensorboard happens to display after starting it). Then go to the tab called \"images\" and move the slider above the grayscale image.<\/p>\n\n<p>In my case it only displayed the images from steps <\/p>\n\n<pre><code>k = 3, 20, 24, 32, 37, 49, 52, 53, 67, 78\n<\/code><\/pre>\n\n<p>which isn't even an nice even spacing, but looks pretty random. I'd prefer to have <\/p>\n\n<ol>\n<li>see more than just 10 of the images I saved, and<\/li>\n<li>have a more even spacing of number of steps between each image displayed.<\/li>\n<\/ol>\n\n<p>How can I achieve this?<\/p>\n\n<p><strong>EDIT:<\/strong> I just found the option <code>--samples_per_plugin<\/code> and tried <code>tensorboard --logdir runs --samples_per_plugin \"images=100\"<\/code>. This indeed increased the number of images, but it only showed the images from steps <code>k = 0,1,2,3....,78<\/code>, but none from above <code>78<\/code>.<\/p>\n","answers":[]}
+{"QuestionId":57669258,"AnswerCount":0,"Tags":"<python><machine-learning><keras><deep-learning><lstm>","CreationDate":"2019-08-27T07:01:45.437","AcceptedAnswerId":null,"OwnerUserId":8513040.0,"Title":"How can I reshape my data for multi-step output model for CNN-LSTM and ConvLSTM in keras?","Body":"<p>I am building a deep learning lstm time series forecasting model. I have read this article: <a href=\"https:\/\/machinelearningmastery.com\/how-to-develop-lstm-models-for-time-series-forecasting\/\" rel=\"nofollow noreferrer\">https:\/\/machinelearningmastery.com\/how-to-develop-lstm-models-for-time-series-forecasting\/<\/a><\/p>\n\n<p>I am using a univariate data for multi-step output forecasting.  <\/p>\n\n<p>For vanilla lstm and encoder model I have tried this code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>raw_seq = [10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150]\n\n# split a univariate sequence into samples\ndef split_sequence(sequence, n_steps_in, n_steps_out):\n    X, y = list(), list()\n    for i in range(len(sequence)):\n        # find the end of this pattern\n        end_ix = i + n_steps_in\n        out_end_ix = end_ix + n_steps_out\n        # check if we are beyond the sequence\n        if out_end_ix &gt; len(sequence):\n            break\n        # gather input and output parts of the pattern\n        seq_x, seq_y = sequence[i:end_ix], sequence[end_ix:out_end_ix]\n        X.append(seq_x)\n        y.append(seq_y)\n    return array(X), array(y)\n\n# choose a number of time steps\nn_steps_in, n_steps_out = 5, 5\n\n# split into samples\nX, y = split_sequence(raw_seq, n_steps_in, n_steps_out)\nprint(X.shape)\nprint(y.shape)\n\n# reshape from [samples, timesteps] into [samples, timesteps, features]\nn_features = 1\nX = X.reshape((X.shape[0], X.shape[1], n_features))\n<\/code><\/pre>\n\n<p>For example, for a ConvLSTM model I need to  convert my series shape according to the article:<\/p>\n\n<pre><code># reshape from [samples, timesteps] into [samples, timesteps, rows, columns, features]\n<\/code><\/pre>\n\n<p>And model needs an input like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.add(ConvLSTM2D(filters=64, kernel_size=(1,2), activation='relu', input_shape=(n_seq, 1, n_steps, n_features)))\n<\/code><\/pre>\n\n<p>I can convert my univariate series for vanilla lstm and encoder-decoder model for multi-output model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># reshape from [samples, timesteps] into [samples, timesteps, features]\n<\/code><\/pre>\n\n<p>But I cannot convert my series for CNN-LSTM and ConvLSTM for multi-output model:<\/p>\n\n<pre><code># reshape from [samples, timesteps] into [samples, timesteps, rows, columns, features]\n<\/code><\/pre>\n\n<p>How can I reshape my series for CNN-LSTM and ConvLSTM for multi-output model?<\/p>\n","answers":[]}
+{"QuestionId":57669621,"AnswerCount":1,"Tags":"<shared-libraries><pytorch><yocto><bitbake><precompiled>","CreationDate":"2019-08-27T07:25:11.820","AcceptedAnswerId":null,"OwnerUserId":6485320.0,"Title":"Installing a precompiled library that depends on another in a bitbake recipe","Body":"<p>I am trying to include pytorch\/libtorch in our Yocto-based distribution and running into troubles when it comes to installing precompiled libraries with dependencies among them onto the target.<\/p>\n\n<p>After trying to compile <code>pytorch<\/code> from source <a href=\"https:\/\/stackoverflow.com\/questions\/56956555\/yocto-warrior-bitbake-recipe-for-pytorch-for-nvidia-jetson-nano\">to no avail<\/a> I thought I could use the <a href=\"https:\/\/download.pytorch.org\/libtorch\/cpu\/libtorch-cxx11-abi-shared-with-deps-1.2.0.zip\" rel=\"nofollow noreferrer\">pre-compiled shared libraries<\/a> and move onto the next problem. This however proves so far more difficult than initially anticipated.<\/p>\n\n<p>Specifically, one of the shared libraries I'd like to install, <code>libtorch.so<\/code> depends on another shared library distributed along it <code>libgomp-753e6e92.so.1<\/code>. When running <code>ldd<\/code> it looks something like this:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>$ ldd libtorch.so \n    linux-vdso.so.1 (0x00007ffc27bcb000)\n    libgomp-753e6e92.so.1 =&gt; \/home\/user\/Desktop\/libtorch-1.2.0\/libtorch\/lib\/.\/libgomp-753e6e92.so.1 (0x00007f3052954000)\n    libpthread.so.0 =&gt; \/lib\/x86_64-linux-gnu\/libpthread.so.0 (0x00007f3052735000)\n    librt.so.1 =&gt; \/lib\/x86_64-linux-gnu\/librt.so.1 (0x00007f305252d000)\n    libgcc_s.so.1 =&gt; \/lib\/x86_64-linux-gnu\/libgcc_s.so.1 (0x00007f3052315000)\n    libdl.so.2 =&gt; \/lib\/x86_64-linux-gnu\/libdl.so.2 (0x00007f3052111000)\n    libm.so.6 =&gt; \/lib\/x86_64-linux-gnu\/libm.so.6 (0x00007f3051d73000)\n    libc10.so =&gt; \/home\/user\/Desktop\/libtorch-1.2.0\/libtorch\/lib\/.\/libc10.so (0x00007f3051b2c000)\n    libstdc++.so.6 =&gt; \/usr\/lib\/x86_64-linux-gnu\/libstdc++.so.6 (0x00007f30517a3000)\n    libc.so.6 =&gt; \/lib\/x86_64-linux-gnu\/libc.so.6 (0x00007f30513b2000)\n    \/lib64\/ld-linux-x86-64.so.2 (0x00007f305ef4a000)\n<\/code><\/pre>\n\n<p>The bitbake recipe that tries to put everything into place looks like this:<\/p>\n\n<pre><code>SUMMARY = \"Facebook PyTorch AI\"\nDESCRIPTION = \"Facebook PyTorch AI\"\nHOMEPAGE = \"https:\/\/pytorch.org\/\"\nLICENSE = \"MIT\"\nLIC_FILES_CHKSUM = \"file:\/\/LICENSE;md5=acf4d595f99e159bf31797aa872aef57\"\n\nS = \"${WORKDIR}\/libtorch\"\nLOCAL_INCLUDE = \"${S}\/include\"\nLOCAL_LIB = \"${S}\/lib\"\nTARGET_INCLUDE = \"${D}${includedir}\"\nTARGET_LIB = \"${D}${libdir}\"\n\nSRC_URI = \"\\\n    https:\/\/download.pytorch.org\/libtorch\/cpu\/libtorch-cxx11-abi-shared-with-deps-${PV}.zip \\\n    file:\/\/LICENSE \\\n\"\nSRC_URI[md5sum] = \"f34c5b6e46331f79100cd93522bad3ae\"\n\nINSANE_SKIP_${PN} = \"ldflags\"\nINSANE_SKIP_${PN} += \"already-stripped\"\nINHIBIT_PACKAGE_STRIP = \"1\"\nINHIBIT_SYSROOT_STRIP = \"1\"\nSOLIBS = \".so\"\nFILES_SOLIBSDEV = \"\"\n\n# RDEPENDS_${PN} += \"libgomp\"\n\ndo_install() {\n    # Create include directory\n    install -d ${TARGET_INCLUDE}\n    # Copy all header files\n    cp -R --no-dereference --preserve=mode,links -v ${LOCAL_INCLUDE}\/* ${TARGET_INCLUDE}\n\n    # Create library directory\n    install -d ${TARGET_LIB}\n    # Install shared libraries\n    install -m 0755 ${LOCAL_LIB}\/*.so* ${TARGET_LIB}\n}\n\nFILES_${PN} += \"\\\n    ${includedir}\/* \\\n    ${libdir}\/* \\\n\"\n<\/code><\/pre>\n\n<p>Unfortunately, I keep getting an error telling me that nothing ir <em>rproviding<\/em> the <code>libgomp<\/code> library which is required by <code>libtorch<\/code>.\nI have tried multiple things including putting the library in question in its own recipe, explicitly installing it or using the already provided <code>libgomp<\/code> runtime dependency but without any luck so far. The error is always something in the likes of:<\/p>\n\n<pre><code>ERROR: pytorch-1.2.0-r0 do_package_qa: QA Issue: \/usr\/lib\/libtorch.so contained in package pytorch requires libgomp-753e6e92.so.1(GOMP_4.0)(64bit), but no providers found in RDEPENDS_pytorch? [file-rdeps]\nERROR: pytorch-1.2.0-r0 do_package_qa: QA Issue: \/usr\/lib\/libtorch.so contained in package pytorch requires libgomp-753e6e92.so.1(OMP_1.0)(64bit), but no providers found in RDEPENDS_pytorch? [file-rdeps]\nERROR: pytorch-1.2.0-r0 do_package_qa: QA Issue: \/usr\/lib\/libtorch.so contained in package pytorch requires libgomp-753e6e92.so.1()(64bit), but no providers found in RDEPENDS_pytorch? [file-rdeps]\nERROR: pytorch-1.2.0-r0 do_package_qa: QA Issue: \/usr\/lib\/libtorch.so contained in package pytorch requires libgomp-753e6e92.so.1(GOMP_1.0)(64bit), but no providers found in RDEPENDS_pytorch? [file-rdeps]\nERROR: pytorch-1.2.0-r0 do_package_qa: QA run found fatal errors. Please consider fixing them.\nERROR: pytorch-1.2.0-r0 do_package_qa: Function failed: do_package_qa\n<\/code><\/pre>\n\n<p>Any clues on what I am missing?<\/p>\n","answers":[{"AnswerId":"57689942","CreationDate":"2019-08-28T10:08:39.793","ParentId":null,"OwnerUserId":"6485320","Title":null,"Body":"<p>Due to a lack of a better solution, we compiled our own <code>libtorch<\/code> from source while setting the <code>USE_OPENMP<\/code> environment variable to <code>OFF<\/code>. We still cannot determine why this dependency to a versioned library gave us so much trouble.<\/p>\n"}]}
+{"QuestionId":57669786,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow><anaconda>","CreationDate":"2019-08-27T07:36:02.707","AcceptedAnswerId":null,"OwnerUserId":11736192.0,"Title":"A lot of Warnings when importing tensorflow","Body":"<p>I have installed TensorFlow (by creating an environment on Conda and then installing tensor flow on it using this tutorial video (<a href=\"https:\/\/www.youtube.com\/watch?v=ujTCoH21GlA&amp;list=PLzMcBGfZo4-mP7qA9cagf68V06sko5otr&amp;index=1\" rel=\"nofollow noreferrer\">https:\/\/www.youtube.com\/watch?v=ujTCoH21GlA&amp;list=PLzMcBGfZo4-mP7qA9cagf68V06sko5otr&amp;index=1<\/a>)<\/p>\n\n<p>but when I run the simple code:<\/p>\n\n<pre><code>     import tensorflow\n<\/code><\/pre>\n\n<p>it returns a lot of warnings like this:<\/p>\n\n<pre><code>    C:\\Users\\safa\\Anaconda3\\envs\\tensor\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:550: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.np_resource = np.dtype([(\"resource\", np.ubyte, 1)])\n<\/code><\/pre>\n\n<p>However, the process finishes with exit code 0. What are these warning for?<\/p>\n","answers":[{"AnswerId":"57669920","CreationDate":"2019-08-27T07:45:08.227","ParentId":null,"OwnerUserId":"8638212","Title":null,"Body":"<p>Deprecated means that the function is going to be removed in the future version. Basically the warning is for the tensorflow itself from numpy since tensorflow uses a deprecated function from numpy. <\/p>\n\n<pre><code>Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,))\n<\/code><\/pre>\n\n<p>And tensorflow uses (type,1) in <\/p>\n\n<pre><code>'(1,)type'.np_resource = np.dtype([(\"resource\", np.ubyte, 1)])\n<\/code><\/pre>\n\n<p>It's actually not a problem for you but if you are concerned about it maybe try to update tensorflow to the newest version.<\/p>\n"},{"AnswerId":"57670295","CreationDate":"2019-08-27T08:09:26.110","ParentId":null,"OwnerUserId":"11736192","Title":null,"Body":"<p>I think since the Python=3.6 is the latest python version which supports TensorFlow, and the newest version of NumPy for Python 3.6 is NumPy 1.16.4; I had to downgrade my NumPy package from 1.17 to 1.16.4 to solve this problem. The warnings disappeared after doing this.<\/p>\n"}]}
+{"QuestionId":57669879,"AnswerCount":0,"Tags":"<pytorch><tensorboard><tensorboardx>","CreationDate":"2019-08-27T07:41:46.850","AcceptedAnswerId":null,"OwnerUserId":11982359.0,"Title":"How to understand the meaning of 'input' in tensorboardX-GRAPHS?","Body":"<p>When I used tensorboardX in pytorch, I felt confused about the 'input' in GRAPHS.Here is my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>writer = SummaryWriter(log_dir='runs\/testRes18\/Restest18')\nnet = nn.Linear(512,10,bias=False)\ninput_tensor = torch.Tensor(10, 512)\nwriter.add_graph(net, input_tensor)\n<\/code><\/pre>\n\n<p>It was expected that one input(10*512) and one branch are in Linear in my GRAPHS, but I found another constant 5 in the input,and two branches in the Linear.\nCan anybody explain this? How to understand the meaning of 'input'?Is it equals to the input of network?<\/p>\n\n<p>\u2193the Screecut<br>\n<a href=\"https:\/\/i.stack.imgur.com\/uy9hu.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n","answers":[]}
+{"QuestionId":57670240,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-08-27T08:05:55.313","AcceptedAnswerId":null,"OwnerUserId":9454845.0,"Title":"About TensorFlow import","Body":"<p>Why do I import TensorFlow error report the first time and not the second time<\/p>\n\n<p>Python version 3.7.4\nTensorFlow Version 1.14.0<\/p>\n\n<pre><code>In [1]: import tensorflow\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/dtypes.py:516: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint8 = np.dtype([(\"qint8\", np.int8, 1)])\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/dtypes.py:517: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint8 = np.dtype([(\"quint8\", np.uint8, 1)])\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/dtypes.py:518: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint16 = np.dtype([(\"qint16\", np.int16, 1)])\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/dtypes.py:519: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint16 = np.dtype([(\"quint16\", np.uint16, 1)])\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/dtypes.py:520: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint32 = np.dtype([(\"qint32\", np.int32, 1)])\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/dtypes.py:525: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  np_resource = np.dtype([(\"resource\", np.ubyte, 1)])\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorboard\/compat\/tensorflow_stub\/dtypes.py:541: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint8 = np.dtype([(\"qint8\", np.int8, 1)])\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorboard\/compat\/tensorflow_stub\/dtypes.py:542: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint8 = np.dtype([(\"quint8\", np.uint8, 1)])\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorboard\/compat\/tensorflow_stub\/dtypes.py:543: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint16 = np.dtype([(\"qint16\", np.int16, 1)])\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorboard\/compat\/tensorflow_stub\/dtypes.py:544: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint16 = np.dtype([(\"quint16\", np.uint16, 1)])\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorboard\/compat\/tensorflow_stub\/dtypes.py:545: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint32 = np.dtype([(\"qint32\", np.int32, 1)])\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorboard\/compat\/tensorflow_stub\/dtypes.py:550: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  np_resource = np.dtype([(\"resource\", np.ubyte, 1)])\n\nIn [2]: import tensorflow\n\nIn [3]:\n<\/code><\/pre>\n","answers":[{"AnswerId":"57670417","CreationDate":"2019-08-27T08:16:49.257","ParentId":null,"OwnerUserId":"1849975","Title":null,"Body":"<p>This is a known problem with NumPy <code>1.17.0<\/code> and TensorFlow. Downgrading your NumPy to the <code>1.16.4<\/code> should fix the warning messages.<\/p>\n\n<p>You can check your NumPy version with the following command:<\/p>\n\n<pre><code>python -c \"import numpy; print(numpy.__version__)\"\n<\/code><\/pre>\n"}]}
+{"QuestionId":57670482,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><tensorflow-lite>","CreationDate":"2019-08-27T08:21:16.187","AcceptedAnswerId":null,"OwnerUserId":8287204.0,"Title":"Can we run multiple deep learning models on Raspberry Pi3 with TensorflowLite for inferenc","Body":"<p>I have multiple deep learning model (say 2) and output of those is input of another deep learning model. I wannna to run this on Raspberry Pi3 for inferencing. <\/p>\n\n<p>Can its possible to run multiple model simultaneously on device. I am using TensorFlow Lite for optimization and inferencing.<\/p>\n\n<p>Please suggest solutions.<\/p>\n","answers":[{"AnswerId":"57782052","CreationDate":"2019-09-04T05:34:25.750","ParentId":null,"OwnerUserId":"366806","Title":null,"Body":"<p>The short answer is YES. For example, in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r2.0\/tensorflow\/lite\/examples\/label_image\/label_image.md\" rel=\"nofollow noreferrer\">lable_image for TFLite<\/a>, your input bitmap image is feed into a simple op (resize bilinear) to resize it, and the output of the simple op is then set as the output of the input of the model you specified.<\/p>\n"}]}
+{"QuestionId":57671009,"AnswerCount":0,"Tags":"<keras><autoencoder>","CreationDate":"2019-08-27T08:54:18.300","AcceptedAnswerId":null,"OwnerUserId":8167827.0,"Title":"Which are the random values in autoencoder?","Body":"<p>I create a simple autoencoder and to have a reproducible result (on CPU and GPU) I set:<\/p>\n\n<pre><code>---------------------------------------------------------\nfrom numpy.random import seed\nseed(1)\nfrom tensorflow import set_random_seed\nset_random_seed(1)\nimport os\nos.environ['PYTHONHASHSEED']=str(1)\nos.environ['TF_CPP_MIN_LOG_LEVEL'] = '1'\n---------------------------------------------------------\n<\/code><\/pre>\n\n<p>And I use a kernel initializer: <code>kernel_initializer='glorot_normal'<\/code><\/p>\n\n<p>Effectively I have the same result at the end in CPU and GPU environment (I run the code many times on CPU and many times on GPU):<\/p>\n\n<pre><code>Confusion matrix: \n[[653  14]\n [  0  28]]\nAccuracy:0.9895052473763118\nPrecision:0.9790104947526237\nRecall:1.0\nF1:0.9893939393939394\nROC AUC is: 0.989505\n<\/code><\/pre>\n\n<p>But today I decide to run the code again and I have:<\/p>\n\n<pre><code>Confusion matrix: \n[[667   0]\n [  0  28]]\nAccuracy:1.0\nPrecision:1.0\nRecall:1.0\nF1:1.0\nROC AUC is: 1.000000\n<\/code><\/pre>\n\n<p>How can I maximize the result in stable way? I want have the last result every time I run the code!<\/p>\n","answers":[]}
+{"QuestionId":57671282,"AnswerCount":1,"Tags":"<python><tensorflow><regression><linear-regression>","CreationDate":"2019-08-27T09:11:04.747","AcceptedAnswerId":57672030.0,"OwnerUserId":289970.0,"Title":"Tensorflow Linear Regression predictions returning [nan]","Body":"<p>I am trying to create my first linear regressor using Tensor Flow (without the help of estimators), and in each iteration, I only see a <code>cost<\/code> value of <code>NaN<\/code>. I think I am not doing something right, but unable to zero in on the issue. Can someone please help me troubleshoot the problem?<\/p>\n\n<p>I am using the CA housing dataset<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Common imports\nimport math\nimport numpy as np\nimport tensorflow as tf\nimport pandas as pd\nfrom sklearn import metrics\n\ncalifornia_housing_dataframe = pd.read_csv(\"https:\/\/download.mlcc.google.com\/mledu-datasets\/california_housing_train.csv\", sep=\",\")\n\n<\/code><\/pre>\n\n<p>I am predicting the <code>median_house_value<\/code> column<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>data_X = california_housing_dataframe.iloc[:, :8]\ndata_y = california_housing_dataframe.iloc[:, 8]\nprint('Features (X):\\n', data_X.head(), '\\n')\nprint('Target (y):\\n', data_y.head(), '\\n')\n<\/code><\/pre>\n\n<p>Create training and validation sets<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from sklearn.model_selection import train_test_split\n\ndata_X_train, data_X_validate = train_test_split(data_X, test_size=0.2, random_state=42)\ndata_y_train, data_y_validate = train_test_split(data_y, test_size=0.2, random_state=42)\n<\/code><\/pre>\n\n<p>Setup the hyperspace parameters and TensorFlow variables<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Hyperspace Params\nlearning_rate = 0.01\ntraining_epochs = 1 #40\nbatch_size = 500 #50\ntotalBatches = len(data_X_train)\/batch_size\n\nn, m = data_X_train.shape # 17,000 Rows + 9 Features\nprint('n=', n, ', m=', m)\n\nW = tf.Variable(tf.random_uniform([m, 1], -1.0, 1.0, dtype = tf.float64), name=\"theta\") # Random initialization\nb = tf.Variable(np.random.randn(), name = \"b\", dtype = tf.float64)\nX = tf.placeholder(tf.float64, shape=(None, m), name=\"X\")\ny = tf.placeholder(tf.float64, shape=(None, 1), name=\"y\")\n\nprint('X.shape :\\n', X.shape, '\\n')\nprint('y.shape :\\n', y.shape, '\\n')\nprint('b.shape :\\n', b.shape, '\\n')\nprint('Thetha.shape (W):\\n', W.shape, '\\n')\n\ny_pred = tf.add(tf.matmul(X, W), b, name=\"predictions\")\nerror = y_pred - y\ncost = tf.reduce_mean(tf.square(error), name=\"mse\")\noptimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)\n\n# Global Variables Initializer\ninit = tf.global_variables_initializer()\n<\/code><\/pre>\n\n<p>Now, training the model returns me <code>NaN<\/code> values only<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def get_batch(X, y, batch_size):\n  rnd_idx = np.random.permutation(len(X))\n  n_batches = len(X) \/\/ batch_size\n  for batch_idx in np.array_split(rnd_idx, n_batches):\n    X_batch, y_batch = X.iloc[batch_idx, :], y[batch_idx]\n    yield X_batch, y_batch\n\n# Global Variables Initializer\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n  sess.run(init)\n  for epoch in range(training_epochs):\n    for X_batch, y_batch in get_batch(data_X_train, data_y_train, batch_size):\n      y_batch = np.array(y_batch).reshape(-1, 1)\n      sess.run(optimizer, feed_dict={X: X_batch, y: y_batch})\n      curr_y_pred, curr_error, curr_cost = sess.run([y_pred, error, cost], {X: X_batch, y: y_batch})\n      print('Training... batch.shape: ', X_batch.shape,'curr_error:', curr_error)\n<\/code><\/pre>\n\n<p>Result looks like<\/p>\n\n<pre><code>Training... batch.shape:  (504, 8) curr_error: [[nan]\n [nan]\n [nan]\n [nan]\n [nan]\n [nan]\n [nan]\n [nan]\n [nan]\n ...\n<\/code><\/pre>\n","answers":[{"AnswerId":"57672030","CreationDate":"2019-08-27T09:51:29.417","ParentId":null,"OwnerUserId":"1849975","Title":null,"Body":"<p>Your issue comes from the <code>pd.read_csv(...)<\/code> function. I swapped it for the <code>NumPy<\/code> version (I am not familiar with <code>Pandas<\/code>) and it works like a charm. Here is the whole snippet:<\/p>\n\n<pre><code>import math\nimport numpy as np\nimport tensorflow as tf\nfrom sklearn import metrics\n\ncalifornia_housing_dataframe = np.genfromtxt('https:\/\/download.mlcc.google.com\/mledu-datasets\/california_housing_train.csv', delimiter=',', skip_header=1)\n\ndata_X = california_housing_dataframe[:, :8]\ndata_y = california_housing_dataframe[:, 8]\n\nfrom sklearn.model_selection import train_test_split\n\ndata_X_train, data_X_validate = train_test_split(data_X, test_size=0.2, random_state=42)\ndata_y_train, data_y_validate = train_test_split(data_y, test_size=0.2, random_state=42)\n\n# Hyperspace Params\nlearning_rate = 0.01\ntraining_epochs = 1 #40\nbatch_size = 500 #50\ntotalBatches = len(data_X_train)\/batch_size\n\nn, m = data_X_train.shape # 17,000 Rows + 9 Features\nprint('n=', n, ', m=', m)\n\nW = tf.Variable(tf.random_uniform([m, 1], -1.0, 1.0, dtype = tf.float64), name=\"theta\") # Random initialization\nb = tf.Variable(np.random.randn(), name = \"b\", dtype = tf.float64)\nX = tf.placeholder(tf.float64, shape=(None, m), name=\"X\")\ny = tf.placeholder(tf.float64, shape=(None, 1), name=\"y\")\n\nprint('X.shape :\\n', X.shape, '\\n')\nprint('y.shape :\\n', y.shape, '\\n')\nprint('b.shape :\\n', b.shape, '\\n')\nprint('Thetha.shape (W):\\n', W.shape, '\\n')\n\ny_pred = tf.add(tf.matmul(X, W), b, name=\"predictions\")\nerror = y_pred - y\ncost = tf.reduce_mean(tf.square(error), name=\"mse\")\noptimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)\n\n# Global Variables Initializer\ninit = tf.global_variables_initializer()\n\ndef get_batch(X, y, batch_size):\n  rnd_idx = np.random.permutation(len(X))\n  n_batches = len(X) \/\/ batch_size\n  for batch_idx in np.array_split(rnd_idx, n_batches):\n    X_batch, y_batch = X[batch_idx, :], y[batch_idx]\n    yield X_batch, y_batch\n\nwith tf.Session() as sess:\n  sess.run(init)\n  for epoch in range(training_epochs):\n    for X_batch, y_batch in get_batch(data_X_train, data_y_train, batch_size):\n      y_batch = np.array(y_batch).reshape(-1, 1)\n      sess.run(optimizer, feed_dict={X: X_batch, y: y_batch})\n      curr_y_pred, curr_error, curr_cost = sess.run([y_pred, error, cost], {X: X_batch, y: y_batch})\n      print('Training... batch.shape: ', X_batch.shape,'curr_error:', curr_error)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57671550,"AnswerCount":0,"Tags":"<python><tensorflow><pytorch><k-means><embedding>","CreationDate":"2019-08-27T09:25:16.393","AcceptedAnswerId":null,"OwnerUserId":2354908.0,"Title":"k-means on BERT's sentence embeddings","Body":"<p>I'm trying to do a k-means clustering on tensors (sentence-embeddings) obtained from pre-trained BERT models.<\/p>\n\n<pre><code>from sklearn.cluster import KMeans\n\nembedding = BERTembeddingGenerator.generateSentenceEmbedding(sentence)\nembeddingMapping[embedding] = sentence\nsentenceEmbeddingsList = list(embeddingMapping.keys())\nmodel = KMeans(n_clusters=10, init='k-means++', max_iter=100, n_init=1)\nlabels = model.fit_predict(sentenceEmbeddingsList)\n<\/code><\/pre>\n\n<p>This throws an error <\/p>\n\n<pre><code>ValueError: only one element tensors can be converted to Python scalars\n<\/code><\/pre>\n\n<p>The sentenceEmbedding is of the size = 768<\/p>\n\n<p>I tried to convert the tensors to numpy array. Any pointers as to where I am getting it wrong<\/p>\n","answers":[]}
+{"QuestionId":57671754,"AnswerCount":1,"Tags":"<tensorflow><object-detection-api>","CreationDate":"2019-08-27T09:37:16.023","AcceptedAnswerId":null,"OwnerUserId":5291590.0,"Title":"How to evaluate all checkpoints using model_main.py","Body":"<p>Currently I am running the following command to evaluate the latest checkpoint on my dataset:<\/p>\n\n<pre><code>python model_main.py --pipeline_config_path=training\/faster_rcnn_inception_v2_pets.config --model_dir=training\/ --sample_1_of_n_eval_examples=1 --eval_training_data=True  --sample_1_of_n_eval_on_train_examples=1 --checkpoint_dir=.\/training\/  --run_once=True --alsologtostderr\n<\/code><\/pre>\n\n<p>However, I would like to evaluate all the checkpoints that I have saved. I am not able to figure out how to do this with model_main.py. I don't want to use legacy\/eval.py as there are a lot of issues popping up with that script.<\/p>\n\n<p>EDIT:\nmy checkpoint file looks like this:<\/p>\n\n<pre><code>model_checkpoint_path: \"model.ckpt-900\"\nall_model_checkpoint_paths: \"model.ckpt-0\"\nall_model_checkpoint_paths: \"model.ckpt-100\"\nall_model_checkpoint_paths: \"model.ckpt-200\"\nall_model_checkpoint_paths: \"model.ckpt-300\"\nall_model_checkpoint_paths: \"model.ckpt-400\"\nall_model_checkpoint_paths: \"model.ckpt-500\"\nall_model_checkpoint_paths: \"model.ckpt-600\"\nall_model_checkpoint_paths: \"model.ckpt-700\"\nall_model_checkpoint_paths: \"model.ckpt-800\"\nall_model_checkpoint_paths: \"model.ckpt-900\"\n<\/code><\/pre>\n\n<p>Now instead of the latest checkpoint which is denoted by model_checkpoint_path, I want to evaluate all_mode_checkpoint_paths on a particular set.<\/p>\n","answers":[{"AnswerId":"57672453","CreationDate":"2019-08-27T10:14:29.787","ParentId":null,"OwnerUserId":"1849975","Title":null,"Body":"<p>For your answer a bit more detail is required so I am just going to make a guess here. <\/p>\n\n<p>In the checkpoint directory <code>.\/training\/<\/code>, there should be a <code>checkpoint<\/code> file. This is a textual file and if you open it you will see something like this:<\/p>\n\n<pre><code>model_checkpoint_path: \"model_10.ckpt\"\nall_model_checkpoint_paths: \"model_01.ckpt\"\nall_model_checkpoint_paths: \"model_02.ckpt\"\nall_model_checkpoint_paths: \"model_03.ckpt\"\n...\n<\/code><\/pre>\n\n<p>You just need to edit the first line in order to evaluate a different checkpoint. For example, if you want to evaluate first checkpoint, just exchange first line for <code>model_checkpoint_path: \"model_01.ckpt\"<\/code>.<\/p>\n\n<p>You can then write small Python or Shell script which would evaluate all checkpoints.<\/p>\n"}]}
+{"QuestionId":57672006,"AnswerCount":0,"Tags":"<tensorflow><keras><keras-layer><tensor><tf.keras>","CreationDate":"2019-08-27T09:50:17.130","AcceptedAnswerId":null,"OwnerUserId":4806177.0,"Title":"ValueError: Layer Weight shape(129,1) not compartible with provided weight shape(128,1)","Body":"<p>I'm trying to run the code from the repository <a href=\"https:\/\/github.com\/familyld\/DeepCF\/blob\/master\/CFNet.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/familyld\/DeepCF\/blob\/master\/CFNet.py<\/a>. <\/p>\n\n<p>Here is the code for the function where the error occurs.<\/p>\n\n<pre><code>def load_pretrain_model1(model, dmf_model, dmf_layers,):\n# MF embeddings\ndmf_user_embeddings = dmf_model.get_layer('user_layer0').get_weights()\ndmf_item_embeddings = dmf_model.get_layer('item_layer0').get_weights()\nmodel.get_layer('user_layer0').set_weights(dmf_user_embeddings)\nmodel.get_layer('item_layer0').set_weights(dmf_item_embeddings)\n\n# DMF layers\nfor i in range(1, len(dmf_layers)):\n    dmf_user_layer_weights = dmf_model.get_layer('user_layer%d' % i).get_weights()\n    model.get_layer('user_layer%d' % i).set_weights(dmf_user_layer_weights)\n    dmf_item_layer_weights = dmf_model.get_layer('item_layer%d' % i).get_weights()\n    model.get_layer('item_layer%d' % i).set_weights(dmf_item_layer_weights)\n\n# Prediction weights\ndmf_prediction = dmf_model.get_layer('prediction').get_weights()\nnew_weights = np.concatenate((dmf_prediction[0], np.array([[0,]] * dmf_layers[-1])), axis=0)\nnew_b = dmf_prediction[1]\nmodel.get_layer('prediction').set_weights([new_weights, new_b]) \nreturn model\n<\/code><\/pre>\n\n<p>When I try to run it I get the following error message.<\/p>\n\n<pre><code>File \"\/home\/ade-lab\/PycharmProjects\/deepcf\/CFNet.py\" line 219, in &lt;module&gt;\nmodel=load_pretrain_model(model,dmf_model,userlayers)\n<\/code><\/pre>\n\n<p><br><\/p>\n\n<pre><code>File \"\/home\/ade-lab\/PycharmProjects\/deepcf\/CFNet.py\" line 128, in load_pretrain_model1\nmodel.get_layer('prediction').set_weights([new_weights,new_b])\n<\/code><\/pre>\n\n<blockquote>\n  <p>ValueError:Layer weight shape(129,1) not compatible with provided weight shape(128,1)<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":57672444,"AnswerCount":1,"Tags":"<python><tensorflow><nlp>","CreationDate":"2019-08-27T10:13:52.477","AcceptedAnswerId":57672841.0,"OwnerUserId":9306691.0,"Title":"Calling tf.session.run gets slower","Body":"<p>I am having trouble with my performance doing nlp tasks. I want to use <a href=\"https:\/\/tfhub.dev\/google\/nnlm-en-dim128-with-normalization\/1\" rel=\"nofollow noreferrer\">this<\/a> module for word embeddings and it produces output, but its runtime increases with each iterative call. I have already read about different solutions, but i cant get them to work. I suspect using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/placeholder\" rel=\"nofollow noreferrer\">tf.placeholders<\/a> would be the a good solution, but i dont know how to use them in this instance.<\/p>\n\n<p>Example code for my problem:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>embedder = hub.Module(\"https:\/\/tfhub.dev\/google\/nnlm-en-dim128-with-normalization\/1\")\nsession = tf.Session()\nsession.run(tf.global_variables_initializer())\nsession.run(tf.tables_initializer())\n\ndoc = [[\"Example1\", \"Example2\", \"Example3\", \"Example4\", ...], [...], ...]\n\nfor paragraph in doc:\n   vectors = session.run(embedder(paragraph))\n   #do something with vectors\n<\/code><\/pre>\n\n<p>Note, that doc cant be fed to the embedder all at once.<\/p>\n\n<p>Thank you in advance.<\/p>\n","answers":[{"AnswerId":"57672841","CreationDate":"2019-08-27T10:36:42.417","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You are recreating the whole model on each iteration, so the TensorFlow graph is growing constantly. You should instead have a single model with a placeholder for your input, then feed the different paragraphs.<\/p>\n\n<pre><code>embedder = hub.Module(\"https:\/\/tfhub.dev\/google\/nnlm-en-dim128-with-normalization\/1\")\nsession = tf.Session()\nsession.run(tf.global_variables_initializer())\nsession.run(tf.tables_initializer())\n\ndoc = [[\"Example1\", \"Example2\", \"Example3\", \"Example4\", ...], [...], ...]\n\nmy_input = tf.placeholder(tf.string, [None])\nmy_embedder = embedder(my_input)\nfor paragraph in doc:\n   vectors = session.run(my_embedder, feed_dict={my_input: paragraph})\n   #do something with vectors\n<\/code><\/pre>\n"}]}
+{"QuestionId":57672642,"AnswerCount":1,"Tags":"<python><keras><python-imaging-library>","CreationDate":"2019-08-27T10:25:11.157","AcceptedAnswerId":58800076.0,"OwnerUserId":3937131.0,"Title":"How to fix IOError in _load_libtiff function?","Body":"<p>I use flow_from_dataframe function in order to augment my dataset.<\/p>\n\n<p>however when, <\/p>\n\n<pre><code>train= data_gen.flow_from_dataframe(dataframe=train_df,\n                                               directory=None,\n                                               x_col='image_name',\n                                               y_col='label',\n                                               class_mode=\"raw\",\n                                               target_size=(img_width, img_height),\n                                               batch_size=2048)\nfor i in train:\n    break\n<\/code><\/pre>\n\n<p><strong>I iterate over the output in order to save my images I found IOError.<\/strong><\/p>\n\n<blockquote>\n  <p>for i in train:<\/p>\n<\/blockquote>\n\n<pre><code>  File \"C:\\Users\\m.seifikar\\PycharmProjects\\firmfile\\venv\\lib\\site-packages\\keras_preprocessing\\image\\iterator.py\", line 104, in __next__\n    return self.next(*args, **kwargs)\n  File \"C:\\Users\\m.seifikar\\PycharmProjects\\firmfile\\venv\\lib\\site-packages\\keras_preprocessing\\image\\iterator.py\", line 116, in next\n    return self._get_batches_of_transformed_samples(index_array)\n  File \"C:\\Users\\m.seifikar\\PycharmProjects\\firmfile\\venv\\lib\\site-packages\\keras_preprocessing\\image\\iterator.py\", line 230, in _get_batches_of_transformed_samples\n    interpolation=self.interpolation)\n  File \"C:\\Users\\m.seifikar\\PycharmProjects\\firmfile\\venv\\lib\\site-packages\\keras_preprocessing\\image\\utils.py\", line 119, in load_img\n    img = img.convert('RGB')\n  File \"C:\\Users\\m.seifikar\\PycharmProjects\\firmfile\\venv\\lib\\site-packages\\PIL\\Image.py\", line 912, in convert\n    self.load()\n  File \"C:\\Users\\m.seifikar\\PycharmProjects\\firmfile\\venv\\lib\\site-packages\\PIL\\TiffImagePlugin.py\", line 1077, in load\n    return self._load_libtiff()\n  File \"C:\\Users\\m.seifikar\\PycharmProjects\\firmfile\\venv\\lib\\site-packages\\PIL\\TiffImagePlugin.py\", line 1168, in _load_libtiff\n    raise IOError(err)\nOSError: -9\n<\/code><\/pre>\n","answers":[{"AnswerId":"58800076","CreationDate":"2019-11-11T11:01:29.033","ParentId":null,"OwnerUserId":"3937131","Title":null,"Body":"<p>Finally, I solved this problem. when we use flow_from_dataframe we should specify the color mode(the default is RGB). if our image is in grey mode and we use the default setting this error will be raised.<\/p>\n"}]}
+{"QuestionId":57673619,"AnswerCount":0,"Tags":"<caffe><caffe2>","CreationDate":"2019-08-27T11:23:51.430","AcceptedAnswerId":null,"OwnerUserId":4417582.0,"Title":"How is caffe resizing my images for training in transform_param","Body":"<p>So, I want to understand is caffe loosing detail of my high resolution image while resizing to 300x300?<\/p>\n\n<p>There could be two cases while resizing image to 300x300 my 1200x990 image:\n1) Cropping\n2) Squishing <\/p>\n\n<p>In first case i.e. cropping it's loosing details of my actual labelled image<\/p>\n\n<p>In second case i.e. Squishing my original high resolution image is squished to small size which also means it's a waste to pass high resolution image<\/p>\n\n<p>Now, I saw the source and <strong><em>train.prototxt<\/em><\/strong><\/p>\n\n<pre><code>transform_param {\n    resize_param {\n          resize_mode: WARP\n          height: 300\n          width: 300\n    }\n}\n<\/code><\/pre>\n\n<p>then I saw cpp code for wrap &amp; found<\/p>\n\n<p>Link to cpp code <a href=\"https:\/\/github.com\/intel\/caffe\/blob\/master\/src\/caffe\/util\/im_transforms.cpp\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<pre><code>void UpdateBBoxByResizePolicy(const ResizeParameter&amp; param,\n                              const int old_width, const int old_height,\n                              NormalizedBBox* bbox) {\n  float new_height = param.height();\n  float new_width = param.width();\n  float orig_aspect = static_cast&lt;float&gt;(old_width) \/ old_height;\n  float new_aspect = new_width \/ new_height;\n\n  float x_min = bbox-&gt;xmin() * old_width;\n  float y_min = bbox-&gt;ymin() * old_height;\n  float x_max = bbox-&gt;xmax() * old_width;\n  float y_max = bbox-&gt;ymax() * old_height;\n.....\n.....\n<\/code><\/pre>\n\n<p>and the actual logic<\/p>\n\n<pre><code>  switch (param.resize_mode()) {\n    case ResizeParameter_Resize_mode_WARP:\n      x_min = std::max(0.f, x_min * new_width \/ old_width);\n      x_max = std::min(new_width, x_max * new_width \/ old_width);\n      y_min = std::max(0.f, y_min * new_height \/ old_height);\n      y_max = std::min(new_height, y_max * new_height \/ old_height);\n      break;\n<\/code><\/pre>\n\n<p>now I understand Cplusplus but what I don't understand is what is bbox? and what's the logic for WRAP mode?<\/p>\n\n<p>Can someone please explain me what's happening with my 1200x900 image. Thanks<\/p>\n","answers":[]}
+{"QuestionId":57673826,"AnswerCount":0,"Tags":"<python><python-3.x><windows><tensorflow>","CreationDate":"2019-08-27T11:35:35.570","AcceptedAnswerId":null,"OwnerUserId":10490382.0,"Title":"tensorflow is fail on import windows","Body":"<p>ok so i installed tensorflow 1.12.0 on python 3.6.4\nand i wanted to import tensor to see if everthing is ok and then that happened<\/p>\n\n<p>i expected successfull install and that happened i have cuda and cudnn installed i tried alot of stuff please help<\/p>\n\n<p>uninstall reinstall force install <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>PS H:\\Program Files (x86)\\Python36-32&gt; python\nPython 3.6.4 (v3.6.4:d48eceb, Dec 19 2017, 06:04:45) [MSC v.1900 32 bit (Intel)] on win32\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n&gt;&gt;&gt; import tensorflow\nTraceback (most recent call last):\n  File \"H:\\Program Files (x86)\\Python36-32\\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 \"H:\\Program Files (x86)\\Python36-32\\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 \"H:\\Program Files (x86)\\Python36-32\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"H:\\Program Files (x86)\\Python36-32\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"H:\\Program Files (x86)\\Python36-32\\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 \"H:\\Program Files (x86)\\Python36-32\\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 \"H:\\Program Files (x86)\\Python36-32\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"H:\\Program Files (x86)\\Python36-32\\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 \"H:\\Program Files (x86)\\Python36-32\\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 \"H:\\Program Files (x86)\\Python36-32\\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 \"H:\\Program Files (x86)\\Python36-32\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"H:\\Program Files (x86)\\Python36-32\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"H:\\Program Files (x86)\\Python36-32\\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==============================================================================\nand that happen if you do the import tensorflow as tf in side his own directory\n==============================================================================\n```python\nPS H:\\Program Files (x86)\\Python36-32\\Lib\\site-packages\\tensorflow\\python&gt; python test.py\nTraceback (most recent call last):\n  File \"test.py\", line 1, in &lt;module&gt;\n    import tensorflow as tf\n  File \"H:\\Program Files (x86)\\Python36-32\\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 \"H:\\Program Files (x86)\\Python36-32\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 47, in &lt;module&gt;\n    import numpy as np\n  File \"C:\\Users\\Sagi\\AppData\\Roaming\\Python\\Python36\\site-packages\\numpy\\__init__.py\", line 142, in &lt;module&gt;\n    from . import core\n  File \"C:\\Users\\Sagi\\AppData\\Roaming\\Python\\Python36\\site-packages\\numpy\\core\\__init__.py\", line 100, in &lt;module&gt;\n    from . import _internal\n  File \"C:\\Users\\Sagi\\AppData\\Roaming\\Python\\Python36\\site-packages\\numpy\\core\\_internal.py\", line 20, in &lt;module&gt;\n    IS_PYPY = platform.python_implementation() == 'PyPy'\nAttributeError: module 'platform' has no attribute 'python_implementation'```\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57673862,"AnswerCount":0,"Tags":"<python><tensorflow><resuming-training>","CreationDate":"2019-08-27T11:37:18.577","AcceptedAnswerId":null,"OwnerUserId":11983366.0,"Title":"How to increase training steps in Tensorflow?","Body":"<p>I followed the following Tensorflow tutorial to retrain the Inception V3 on my own classes.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/hub\/tutorials\/image_retraining\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/hub\/tutorials\/image_retraining<\/a><\/p>\n\n<p>Everything worked well so far and I got an acceptable final test accuracy. However, I want to improve the result by increasing the training the steps. I trained the model for 4000 steps and I want to increase it to 8000 steps. How can I do that without starting the training all over again?<\/p>\n\n<p>I have read so many documents about saving and restoring checkpoints but I could not understand how to use them. Should I modify the retain.py to allow continuation of the training? If yes, how can I do so?<\/p>\n\n<p>Thank you for your help!<\/p>\n","answers":[]}
+{"QuestionId":57674274,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-08-27T12:03:12.527","AcceptedAnswerId":null,"OwnerUserId":4354355.0,"Title":"Poor performance transfer learning ResNet50","Body":"<p>I have a dataset of 11k images labeled for semantic segmentation. About 8.8k belong to 'group 1' and the rest to 'group 2'<\/p>\n\n<p>I am trying to simulate what would happen if we lost access to 'group 1' imagery but not a network trained from them.<\/p>\n\n<p>So I trained ResNet50 on group 1 only. Then used that network as a starting point for training group 2 only. <\/p>\n\n<p>Results are essentially slightly better than not training with group 2 imagery (3% in average per class accuracy) but less than 1% better than if I just started with imagenet weights. I tested freezing blocks of resnet50 and a range of learning rates.<\/p>\n\n<p>Group 1 and 2 are part of the same problem domain but are a bit different. They are taken at different regions (in fact the whole set covers a bunch of areas but group 1 and 2 are disjoint in this regard) and a different camera\/resolution. They are resized to a fixed size though this fixed size is closer to group 1 average size.<\/p>\n\n<p>They are very different to imagenet images. They are monochrome, rectangular and are essentially one type of object that I'm segmenting.<\/p>\n\n<p>I'm not seeking to get the same result as training on all the images at once but surely there must be a bump in doing this over just training from imagenet. <\/p>\n","answers":[]}
+{"QuestionId":57674320,"AnswerCount":0,"Tags":"<python><keras><text-classification>","CreationDate":"2019-08-27T12:05:31.597","AcceptedAnswerId":null,"OwnerUserId":3261292.0,"Title":"modeling the context with LSTM","Body":"<p>I have around 500 documents, these documents contain \"malicious\" sentences (positive class).  What I want to do is to identify these sentences from the rest of \"legitimate\" sentences.<\/p>\n\n<p>What I did is: I built a LSTM network (many-to-one structure) which takes each sentence with the label (malicious or not). I noticed that the results are weak. <\/p>\n\n<p>The identification of these sentences needs context information to decide if they are malicious.\nThus, I build another bi-LSTM (many-to-many structure) to model the context. \nI feed to the network sentence i, i-1, and i+1. I will use the embeddings from the last_layer-1 to feed them to a classifier.<\/p>\n\n<pre><code>inp = Input(shape=(3, 512))\nx = Bidirectional(LSTM(256, dropout=0.2, return_sequences=True))(inp)\nx = TimeDistributed(Dense(128, activation='relu'))(x)\nx = Dropout(0.3)(x)\nout = TimeDistributed(Dense(2, activation=\"softmax\"))(x)\nmodel = Model(inputs=inp, outputs=out)\n<\/code><\/pre>\n\n<p>Each sentence is represented with 512 embeddings, thus, I have 3 steps and each has 512 dimension.<\/p>\n\n<p>What confuses me is that I'm not sure if this approach will utilize the context, does it?<\/p>\n","answers":[]}
+{"QuestionId":57674698,"AnswerCount":0,"Tags":"<android><tensorflow><mobile><tensorflow-lite>","CreationDate":"2019-08-27T12:27:32.507","AcceptedAnswerId":null,"OwnerUserId":11983622.0,"Title":"Tensorflow-lite large model run faster on Kirin 970 CPU than it's GPU","Body":"<p>I'm testing tensorflow-lite benchmark on Kirin 970, and notice that the small model like mobilenet and squeezenet run faster on GPU than CPU and the large model like VGG-16 run faster on CPU. It's sort of weird. What are the causes of these phenomena?<\/p>\n\n<p>Test using command:<\/p>\n\n<pre><code>.\/benchmark_model --graph=convert_tflite_model\/vgg16.tflite  --num_threads=8 --warmup_runs=1 --num_runs=10 --use_gpu=true --run_delay=1\n<\/code><\/pre>\n\n<p>Here are my test results:<\/p>\n\n<p><em>\"GPU default\" means using default setting of GPU.<\/em><\/p>\n\n<p><em>\"GPU 767MHz\" means setting GPU frequency to 767MHz(Max frequency of GPU on Kirin 970)<\/em><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Xe7Gu.png\" rel=\"nofollow noreferrer\">Table: models run time on CPU and GPU<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/hEV2h.png\" rel=\"nofollow noreferrer\">Histogram: CPU time\/GPU time<\/a><\/p>\n\n<hr>\n\n<p>BTW, how to shut down Kirin 970 thermal throttling? It will reduce frequency just in few seconds during test because of overheating. Frequency reduction makes it harder to get solid test results.<\/p>\n","answers":[]}
+{"QuestionId":57674778,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-27T12:31:33.357","AcceptedAnswerId":null,"OwnerUserId":7697890.0,"Title":"Custom \"binary\" accuracy in a multi class model","Body":"<p>I'm working with a classification model that outputs 10 classes. <\/p>\n\n<p>The objective is to \"regroup\" the classes 1-5 and 6-10, the idea is to extract the argmax of both y_true and y_pred, then to inspect if they are in the same \"super-group\". This result is the accuracy<\/p>\n\n<p>I tried this code but the result is 0 for the accuracy (the model is pre-trained and has no other problem)<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def test(y_true, y_pred):\n\n    y_true_index = K.argmax(y_true, axis=-1)\n    y_pred_index = K.argmax(y_pred, axis=-1)\n\n    #I tried to set threshold to just 4 but the result is the same\n    #I just tried it this way\n    threshold = np.full(K.int_shape(y_true_index)[0], 4)\n\n    y_true_bool = K.greater(y_true_index, threshold)\n    y_pred_bool = K.greater(y_pred_index, threshold)\n\n    matches = K.cast(K.equal(y_true_bool, y_pred_bool), 'int32')\n\n    return K.mean(matches)\n\n\n# after model is loaded\n\nmodel.compile(opt, \n              loss=loss, \n              metrics=[test])\n\nmodel.evaluate_generator(generator,\n                         steps=steps,\n                         verbose=1)\n<\/code><\/pre>\n\n<p>I'm not a real expert with axis, maybe the problem is there<\/p>\n","answers":[]}
+{"QuestionId":57674908,"AnswerCount":0,"Tags":"<python><tensorflow><keras><batch-normalization>","CreationDate":"2019-08-27T12:39:09.140","AcceptedAnswerId":null,"OwnerUserId":4084900.0,"Title":"TensorFlow Keras use BatchNormalization moving statistics during training phase?","Body":"<p>By default, TensorFlow's Keras' BatchNormalization will train by using the mean and variance of the current batch. At the same time, it will incorporate these values into the moving average statistics. Then, during testing, the moving average statistics will be used rather than the statistics of the test data. This is documented <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/layers\/BatchNormalization#call_arguments\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>Is there a way to use the moving statistics for the normalization both during training and testing? Specifically, using them while still updating the statistics during training, but not updating the statistics during testing?<\/p>\n","answers":[]}
+{"QuestionId":57675242,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow2.0>","CreationDate":"2019-08-27T12:58:40.173","AcceptedAnswerId":57704523.0,"OwnerUserId":10449967.0,"Title":"How does tf.function works when two functions call each other","Body":"<p>I build my model using tensorflow==1.14 with tf.enable_eager_execution() like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class Model:\n  def __init__(self):\n    self.embedding = tf.keras.layers.Embedding(10, 15)\n    self.dense = tf.keras.layers.Dense(10)\n\n  @tf.function\n  def inference(self, inp):\n    print('call function: inference')\n    inp_em = self.embedding(inp)\n    inp_enc = self.dense(inp_em)\n\n    return inp_enc\n\n  @tf.function\n  def fun(self, inp):\n    print('call function: fun')\n    return self.inference(inp)\n\nmodel = Model()\n<\/code><\/pre>\n\n<p>when I ran the following code for first time:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>a = model.fun(np.array([1, 2, 3]))\nprint('=' * 20)\na = model.inference(np.array([1, 2, 3]))\n<\/code><\/pre>\n\n<p>the output is<\/p>\n\n<pre><code>call function: fun\ncall function: inference\ncall function: inference\n====================\ncall function: inference\n<\/code><\/pre>\n\n<p>it seems like tensorflow build three graphs for inference function, how can I just build just one graph for inference function.\nAnd I also want to know how tf.function woks when two functions call each other. Is this the right way to build my model?<\/p>\n","answers":[{"AnswerId":"57704523","CreationDate":"2019-08-29T06:49:22.283","ParentId":null,"OwnerUserId":"2455494","Title":null,"Body":"<p>Sometimes the way <code>tf.function<\/code> executes can cause us a little confusion - particularly when we mix in vanilla python operations such as <code>print()<\/code>. <\/p>\n\n<p>We should remember that when we decorate a function with <code>tf.function<\/code> it's no longer <em>just<\/em> a python function. It behaves a little differently in order to enable fast and efficient use in TF. The vast majority of the time, the changes in behaviour are pretty much unnoticeable (except for the increased speed!) but occasionally we can encounter a little nuance like this. <\/p>\n\n<p>The first thing to note is that if we use <code>tf.print()<\/code> in place of <code>print()<\/code> then we get the expected output:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class Model:\n  def __init__(self):\n    self.embedding = tf.keras.layers.Embedding(10, 15)\n    self.dense = tf.keras.layers.Dense(10)\n\n  @tf.function\n  def inference(self, inp):\n    tf.print('call function: inference')\n    inp_em = self.embedding(inp)\n    inp_enc = self.dense(inp_em)\n\n    return inp_enc\n\n  @tf.function\n  def fun(self, inp):\n    tf.print('call function: fun')\n    return self.inference(inp)\n\nmodel = Model()\n\na = model.fun(np.array([1, 2, 3]))\nprint('=' * 20)\na = model.inference(np.array([1, 2, 3]))\n<\/code><\/pre>\n\n<p>outputs:<\/p>\n\n<pre><code>call function: fun\ncall function: inference\n====================\ncall function: inference\n<\/code><\/pre>\n\n<p>If your question is the symptom of a real world problem this is probably the fix!<\/p>\n\n<p>So what's going on? <\/p>\n\n<p>Well the first time we call a function decorated with <code>tf.function<\/code> tensorflow will build an execution graph. In order to do so it \"traces\" the tensorflow operations executed by the python function. <\/p>\n\n<p><strong>In order to do this tracing it is possible that tensorflow will call the decorated function more than once<\/strong>! <\/p>\n\n<p>This means that the python only operations (such as <code>print()<\/code> could get executed more than once) but tf operations such as <code>tf.print()<\/code> will behave as you would normally expect. <\/p>\n\n<p>A side effect of this nuance is that we ought to be aware of how <code>tf.function<\/code> decorated functions handle state but this is outside of the scope of your question. See the <a href=\"https:\/\/github.com\/tensorflow\/community\/blob\/master\/rfcs\/20180918-functions-not-sessions-20.md\" rel=\"nofollow noreferrer\">original RFC<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/26812#issuecomment-474595550\" rel=\"nofollow noreferrer\">this github issue<\/a> for more info.<\/p>\n\n<blockquote>\n  <p>And I also want to know how tf.function woks when two functions call each other. Is this the right way to build my model?<\/p>\n<\/blockquote>\n\n<p>In general, we <em>need<\/em> only decorate the \"outer\" function with <code>tf.function<\/code> (<code>.fun()<\/code> in your example) but if you could call the inner function directly too then you are free to decorate that too.<\/p>\n"}]}
+{"QuestionId":57676349,"AnswerCount":0,"Tags":"<python><numpy><tensorflow><neural-network>","CreationDate":"2019-08-27T13:59:18.277","AcceptedAnswerId":null,"OwnerUserId":9954063.0,"Title":"How to pass a tensor variable to feed_dict?","Body":"<p>I have basic neural network which has a function run(). which takes the input and provides the output of the neural network. In the function body, I would like to <code>feed_dict<\/code> weights when running a session to calculate the output. Since these weights are <code>tf.compat.v1.variable<\/code> I am having an error <code>ValueError: Setting an array element with sequence<\/code><\/p>\n\n<p>I have initially thought it was my input that is causing the error and hence, I have made sure it has same dimensions as weights to use <code>tf.multiply<\/code> and since the error continued, I made sure to convert the numpy array into a tensor object hoping that might be causing the error. But ti still does not work.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def run(self, carollis_input):\n        self.normalization(carollis_input)\n        c_in = np.array(carollis_input)\n        c_input = tf.compat.v1.convert_to_tensor(c_in, tf.float64)\n        #'finding the output of the input layer'\n        #with tf.Session() as sess1_2:\n        input_weight = tf.compat.v1.placeholder(tf.float64, shape=(self.neurons, 1))\n        input_bias = tf.compat.v1.placeholder(tf.float64, shape=(self.neurons, 1))\n        hidden_weight = tf.compat.v1.placeholder(tf.float64, shape=(self.neurons, 1))\n        hidden_bias =tf.compat.v1.placeholder(tf.float64, shape=(self.neurons, 1))\n        output_weight = tf.compat.v1.placeholder(tf.float64, shape=(4, 1))\n        output_bias = tf.compat.v1.placeholder(tf.float64, shape=(self.neurons, 1))\n\n        knowledge_input = tf.add(tf.multiply(c_input, input_weight), input_bias)\n        with tf.Session() as ses1:\n            ses1.run(knowledge_input, feed_dict={input_weight: self.weight_in, input_bias:self.bias_in})\n        knowledge_hidden = tf.nn.leaky_relu(knowledge_input, alpha=0.01) \n        #'calculating the output of hidden layer'\n        knowledge_hidden_output = 3.14*(tf.add(tf.multiply(knowledge_hidden, hidden_weight), hidden_bias))#input function of hidden layer\n        knowledge_hidden_out = tf.nn.leaky_relu(knowledge_hidden_output, alpha=0.01, name='leaky_relu')\n\n        with tf.Session() as ses2:\n            ses2.run(knowledge_hidden_out, feed_dict={hidden_weight: self.weight_hid, hidden_bias:self.bias_hid})\n        #'calculating the input of output layer'\n        tf.reshape(knowledge_hidden_out, [4, 2])#for quadrant method\n        in_out = tf.add(tf.multiply(knowledge_hidden_out, output_weight), output_bias)\n        with tf.Session() as ses3:\n            ses3.run(in_out, feed_dict={output_weight: self.weight_out, output_bias: self.bias_out})\n        #'finding the softmax output of the neurons'\n        softmax_output = np.array(4)\n        softmax_output = self.out_softmax(in_out)  # this gives the softmax output and stores it in the newly created array\n        return softmax_output\n<\/code><\/pre>\n\n<p>After going through other answers , I believe that error might be cause because I was trying to feed in a tensor variable i.e. weights and biases to the feed_dict. If thats the problem, how can i solve it, please kindly help me. The exact error is <\/p>\n\n<pre><code>File \"\/home\/microbot\/catkin_ws\/src\/spider\/spider_control\/knowledge_transfer.py\", line 85, in run\n    ses1.run(knowledge_input, feed_dict={input_weight: self.weight_in, input_bias:self.bias_in})\n  File \"\/home\/microbot\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 950, in run\n    run_metadata_ptr)\n  File \"\/home\/microbot\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1142, in _run\n    np_val = np.asarray(subfeed_val, dtype=subfeed_dtype)\n  File \"\/home\/microbot\/.local\/lib\/python3.6\/site-packages\/numpy\/core\/_asarray.py\", line 85, 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>EDIT:\nThe weights are initialized inside the class initialiser i.e. <\/p>\n\n<pre><code>def __init__(self, neurons, layers):\n        self.neurons = neurons\n        self.layers =  layers\n        self.weight_initer = tf.truncated_normal_initializer(mean=0.0, stddev=0.01)\n        self.weight_in = tf.get_variable(name=\"Weight_input\", dtype=tf.float64, shape=[self.neurons, 1], initializer=self.weight_initer)\n        self.weight_initer1 = tf.truncated_normal_initializer(mean=1.0, stddev=0.01)\n        self.weight_hid = tf.get_variable(name=\"Weight_hidden\", dtype=tf.float64, shape=[self.neurons, 1], initializer=self.weight_initer1)\n        self.weight_initer2 = tf.truncated_normal_initializer(mean=2.0, stddev=0.01)\n        self.weight_out = tf.get_variable(name=\"Weight_output\", dtype=tf.float64, shape=[4, 2], initializer=self.weight_initer2)\n        self.bias_initer =tf.truncated_normal_initializer(mean=0.1, stddev=0.01)\n        self.bias_in  =tf.get_variable(name=\"Bias_input\", dtype=tf.float64, shape=[self.neurons, 1], initializer=self.bias_initer)\n        self.bias_initer1 =tf.truncated_normal_initializer(mean=0.2, stddev=0.01)\n        self.bias_hid = tf.get_variable(name=\"Bias_hidden\", dtype=tf.float64, shape=[self.neurons, 1], initializer=self.bias_initer1)\n        self.bias_initer2 =tf.truncated_normal_initializer(mean=0.3, stddev=0.01)\n        self.bias_out = tf.get_variable(name=\"Bias_output\", dtype=tf.float64, shape=[4, 1], initializer=self.bias_initer2)\n\n<\/code><\/pre>\n\n<p>Although now I have been having two different confusing, one with respective to scope of the graph and other to the weight initialising. Since I have defined the weight tensors in the initialiser, would this be the defaul graph and the graph that i am trying to build inside the run function be secondary ? <\/p>\n","answers":[]}
+{"QuestionId":57676566,"AnswerCount":0,"Tags":"<tensorflow><keras><tensorboard><tf.keras>","CreationDate":"2019-08-27T14:10:54.397","AcceptedAnswerId":null,"OwnerUserId":1616773.0,"Title":"Switching my tensorboard plugin from TF.Layers to TF.keras ( Tf. Sequential)","Body":"<p>I have a working model of Tensor flow and on top of it, I have developed a plugin for tensorboard which is working just fine with Tf.Layers model but I am facing some problems when I am trying to replace existing mode with TF.keras. The core features I was using like - session, summary and plugin_summary are missing.\nRefer to the official example from TF - <a href=\"https:\/\/github.com\/tensorflow\/tensorboard-plugin-example\/tree\/master\/greeter_plugin\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorboard-plugin-example\/tree\/master\/greeter_plugin<\/a> <\/p>\n\n<p>I tried calling the same summary function from a callback but when I tried storing values returned by the summary function, it throws an error \"tensor doesn't have value\"<\/p>\n\n<pre><code>In grreter_summary.py we add following details to nodes \n  summary_metadata = tf.SummaryMetadata(\n      display_name=display_name,\n      summary_description=description,\n      plugin_data=tf.SummaryMetadata.PluginData(\n          plugin_name=PLUGIN_NAME))\n<\/code><\/pre>\n\n<p>I am unable to find alternate of this in Keras model.<\/p>\n\n<p>I tried callbacks( code below), they store weights but not plugin name <\/p>\n\n<p>( PLugin name is used  bt tensor board to recognize the relevant data for a plugin )<\/p>\n\n<pre><code>cp_callback = tf.keras.callbacks.ModelCheckpoint(\n    logdir, verbose=1, save_weights_only=True,\n    # Save weights, every 5-epochs.\n    period=5)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57677060,"AnswerCount":0,"Tags":"<python><tensorflow><faster-rcnn>","CreationDate":"2019-08-27T14:36:43.743","AcceptedAnswerId":null,"OwnerUserId":6475103.0,"Title":"Tensorflow GPU gettting stop while training the model","Body":"<p>I am trying to train model for custom object detection following the tutorial <a href=\"https:\/\/www.youtube.com\/watch?v=Rgpfk6eYxJA\" rel=\"nofollow noreferrer\">Youtube tutorial<\/a><\/p>\n\n<p>I am trying to train the model.<\/p>\n\n<p>Before step execution their are multiple warnings like- \"Allocator (GPU_0_bfc) ran out of memory trying to allocate 2.06GiB with freed_by_count=0.  The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.\"\n(see the screenshot)\n<a href=\"https:\/\/i.stack.imgur.com\/lKX7I.png\" rel=\"nofollow noreferrer\">warnings before step executions<\/a><\/p>\n\n<p>then steps get excutes like-<\/p>\n\n<p>global step 1901: loss = 0.2125<\/p>\n\n<p>..<\/p>\n\n<p>..<\/p>\n\n<p>and it step execution stops after some steps with error(following screenshot)\n<a href=\"https:\/\/i.stack.imgur.com\/QeNrn.png\" rel=\"nofollow noreferrer\">steps get stoped with this errors<\/a><\/p>\n\n<p>I am using faster_rcnn_inception_v2_pets.config model for training\nTensorflow - 1.14(lastest)\ncuda - 10.1(lastest)\ncucnn - 10.1cuda - v7.63.30(lastest)<\/p>\n\n<p>The config file as below<\/p>\n\n<p>my file is:-<\/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># Faster R-CNN with Inception v2, configured for Oxford-IIIT Pets Dataset.\r\n# Users should configure the fine_tune_checkpoint field in the train config as\r\n# well as the label_map_path and input_path fields in the train_input_reader and\r\n# eval_input_reader. Search for \"PATH_TO_BE_CONFIGURED\" to find the fields that\r\n# should be configured.\r\n\r\nmodel {\r\n  faster_rcnn {\r\n    num_classes: 5\r\n    image_resizer {\r\n      keep_aspect_ratio_resizer {\r\n        min_dimension: 600\r\n        max_dimension: 1024\r\n      }\r\n    }\r\n    feature_extractor {\r\n      type: 'faster_rcnn_inception_v2'\r\n      first_stage_features_stride: 16\r\n    }\r\n    first_stage_anchor_generator {\r\n      grid_anchor_generator {\r\n        scales: [0.25, 0.5, 1.0, 2.0]\r\n        aspect_ratios: [0.5, 1.0, 2.0]\r\n        height_stride: 16\r\n        width_stride: 16\r\n      }\r\n    }\r\n    first_stage_box_predictor_conv_hyperparams {\r\n      op: CONV\r\n      regularizer {\r\n        l2_regularizer {\r\n          weight: 0.0\r\n        }\r\n      }\r\n      initializer {\r\n        truncated_normal_initializer {\r\n          stddev: 0.01\r\n        }\r\n      }\r\n    }\r\n    first_stage_nms_score_threshold: 0.0\r\n    first_stage_nms_iou_threshold: 0.7\r\n    first_stage_max_proposals: 300\r\n    first_stage_localization_loss_weight: 2.0\r\n    first_stage_objectness_loss_weight: 1.0\r\n    initial_crop_size: 14\r\n    maxpool_kernel_size: 2\r\n    maxpool_stride: 2\r\n    second_stage_box_predictor {\r\n      mask_rcnn_box_predictor {\r\n        use_dropout: false\r\n        dropout_keep_probability: 1.0\r\n        fc_hyperparams {\r\n          op: FC\r\n          regularizer {\r\n            l2_regularizer {\r\n              weight: 0.0\r\n            }\r\n          }\r\n          initializer {\r\n            variance_scaling_initializer {\r\n              factor: 1.0\r\n              uniform: true\r\n              mode: FAN_AVG\r\n            }\r\n          }\r\n        }\r\n      }\r\n    }\r\n    second_stage_post_processing {\r\n      batch_non_max_suppression {\r\n        score_threshold: 0.0\r\n        iou_threshold: 0.6\r\n        max_detections_per_class: 100\r\n        max_total_detections: 300\r\n      }\r\n      score_converter: SOFTMAX\r\n    }\r\n    second_stage_localization_loss_weight: 2.0\r\n    second_stage_classification_loss_weight: 1.0\r\n  }\r\n}\r\n\r\ntrain_config: {\r\n  batch_size: 1\r\n  optimizer {\r\n    momentum_optimizer: {\r\n      learning_rate: {\r\n        manual_step_learning_rate {\r\n          initial_learning_rate: 0.0002\r\n          schedule {\r\n            step: 1\r\n            learning_rate: .0002\r\n          }\r\n          schedule {\r\n            step: 5000\r\n            learning_rate: .00002\r\n          }\r\n          schedule {\r\n            step: 10000\r\n            learning_rate: .000002\r\n          }\r\n          schedule {\r\n            step: 30000\r\n            learning_rate: .000002\r\n          }\r\n        }\r\n      }\r\n      momentum_optimizer_value: 0.9\r\n    }\r\n    use_moving_average: false\r\n  }\r\n  gradient_clipping_by_norm: 10.0\r\n  fine_tune_checkpoint: \"C:\/tensorflow1\/models\/research\/object_detection\/faster_rcnn_inception_v2_coco_2018_01_28\/model.ckpt\"\r\n  from_detection_checkpoint: true\r\n  # Note: The below line limits the training process to 200K steps, which we\r\n  # empirically found to be sufficient enough to train the pets dataset. This\r\n  # effectively bypasses the learning rate schedule (the learning rate will\r\n  # never decay). Remove the below line to train indefinitely.\r\n  num_steps: 200000\r\n  data_augmentation_options {\r\n    random_horizontal_flip {\r\n    }\r\n  }\r\n}\r\n\r\n\r\ntrain_input_reader: {\r\n  tf_record_input_reader {\r\n    input_path: \"data\/train.record\"\r\n  }\r\n  label_map_path: \"data\/object-detection.pbtxt\"\r\n}\r\n\r\neval_config: {\r\n  num_examples: 67\r\n  # Note: The below line limits the evaluation process to 10 evaluations.\r\n  # Remove the below line to evaluate indefinitely.\r\n  max_evals: 10\r\n}\r\n\r\neval_input_reader: {\r\n  tf_record_input_reader {\r\n    input_path: \"data\/test.record\"\r\n  }\r\n  label_map_path: \"data\/object-detection.pbtxt\"\r\n  shuffle: false\r\n  num_readers: 1\r\n}<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n","answers":[]}
+{"QuestionId":57677083,"AnswerCount":1,"Tags":"<python><tensorflow><keras><cuda><nvidia>","CreationDate":"2019-08-27T14:38:32.343","AcceptedAnswerId":57696573.0,"OwnerUserId":3630804.0,"Title":"CUDA driver version is insufficient for CUDA runtime version , CUDA version 10.0","Body":"<p>I get this error when I run <code>keras<\/code> on GPU. <\/p>\n\n<pre><code>  File \"modeltra.py\", line 165, in &lt;module&gt;\n    x_train = a_model.predict(y_training)\n  File \"\/home\/hgam\/anaconda3\/envs\/gp_ker_pip\/lib\/python3.7\/site-packages\/keras\/engine\/training.py\", line 1169, in predict\n    steps=steps)\n  File \"\/home\/hgam\/anaconda3\/envs\/gp_ker_pip\/lib\/python3.7\/site-packages\/keras\/engine\/training_arrays.py\", line 294, in predict_loop\n    batch_outs = f(ins_batch)\n  File \"\/home\/hgam\/anaconda3\/envs\/gp_ker_pip\/lib\/python3.7\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 2697, in __call__\n    if hasattr(get_session(), '_make_callable_from_options'):\n  File \"\/home\/hgam\/anaconda3\/envs\/gp_ker_pip\/lib\/python3.7\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 186, in get_session\n    _SESSION = tf.Session(config=config)\n  File \"\/home\/hgam\/anaconda3\/envs\/gp_ker_pip\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1570, in __init__\n    super(Session, self).__init__(target, graph, config=config)\n  File \"\/home\/hgam\/anaconda3\/envs\/gp_ker_pip\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 693, in __init__\n    self._session = tf_session.TF_NewSessionRef(self._graph._c_graph, opts)\ntensorflow.python.framework.errors_impl.InternalError: cudaGetDevice() failed. Status: CUDA driver version is insufficient for CUDA runtime version\n<\/code><\/pre>\n\n<p>My GPU information is as follows<\/p>\n\n<pre><code>(gp_ker_pip) [hgam@xx]$ nvidia-smi\nTue Aug 27 17:29:34 2019\n+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 410.79       Driver Version: 410.79       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  Tesla P100-PCIE...  On   | 00000000:03:00.0 Off |                    0 |\n| N\/A   34C    P0    33W \/ 250W |    855MiB \/ 16280MiB |      0%      Default |\n+-------------------------------+----------------------+----------------------+\n|   1  Tesla P100-PCIE...  On   | 00000000:82:00.0 Off |                    0 |\n| N\/A   32C    P0    31W \/ 250W |    855MiB \/ 16280MiB |      0%      Default |\n+-------------------------------+----------------------+----------------------+\n<\/code><\/pre>\n\n<pre><code>(gp_ker_pip) [hgam@xx]$ nvcc --version\nnvcc: NVIDIA (R) Cuda compiler driver\nCopyright (c) 2005-2018 NVIDIA Corporation\nBuilt on Sat_Aug_25_21:08:01_CDT_2018\nCuda compilation tools, release 10.0, V10.0.130\n<\/code><\/pre>\n\n<p>First installed tensorflow using \n<code>pip install tensorflow-gpu<\/code><\/p>\n\n<p>and installed keras with\n<code>conda install keras-gpu<\/code><\/p>\n\n<p>What should I follow to make this work? Also how can I check the CUDA runtime version and the driver version?<\/p>\n\n<p>PS: \nI also downgraded the cuda version using <\/p>\n\n<pre><code>&gt;    export PATH=$PATH:\/usr\/local\/cuda-9.0\/bin\n&gt;    export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:\/usr\/local\/cuda-9.0\/lib64\n<\/code><\/pre>\n\n<p>Still I get the error.<\/p>\n","answers":[{"AnswerId":"57696573","CreationDate":"2019-08-28T16:22:51.513","ParentId":null,"OwnerUserId":"3630804","Title":null,"Body":"<p>This issue was solved after installing CUDA 10.1 as @Robert suggested.<\/p>\n\n<pre><code>+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 418.87.00    Driver Version: 418.87.00    CUDA Version: 10.1     |\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 T4            Off  | 00000000:03:00.0 Off |                    0 |\n| N\/A   39C    P0    64W \/  70W |  14503MiB \/ 15079MiB |     98%      Default |\n+-------------------------------+----------------------+----------------------+\n|   1  Tesla T4            Off  | 00000000:82:00.0 Off |                    0 |\n| N\/A   35C    P0    25W \/  70W |    111MiB \/ 15079MiB |      0%      Default |\n+-------------------------------+----------------------+----------------------+\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":57677119,"AnswerCount":0,"Tags":"<apache-spark><keras><nlp>","CreationDate":"2019-08-27T14:40:25.207","AcceptedAnswerId":null,"OwnerUserId":10983663.0,"Title":"my spark dataframe don't want to be exported to JSON because of my keras model","Body":"<p>I have a sentiment analysis model for tweets that I apply to spark<\/p>\n\n<p>When I apply my model and I visualize my dataframe everything goes well but when I tried to export it in JSON format I have the following error<\/p>\n\n<pre><code>Caused by: org.apache.spark.api.python.PythonException: Traceback (most recent call last):\n  File \"\/home\/master\/anaconda3\/lib\/python3.7\/site-packages\/pyspark\/python\/lib\/pyspark.zip\/pyspark\/worker.py\", line 377, in main\n    process()\n  File \"\/home\/master\/anaconda3\/lib\/python3.7\/site-packages\/pyspark\/python\/lib\/pyspark.zip\/pyspark\/worker.py\", line 372, in process\n    serializer.dump_stream(func(split_index, iterator), outfile)\n  File \"\/home\/master\/anaconda3\/lib\/python3.7\/site-packages\/pyspark\/python\/lib\/pyspark.zip\/pyspark\/serializers.py\", line 345, in dump_stream\n    self.serializer.dump_stream(self._batched(iterator), stream)\n  File \"\/home\/master\/anaconda3\/lib\/python3.7\/site-packages\/pyspark\/python\/lib\/pyspark.zip\/pyspark\/serializers.py\", line 141, in dump_stream\n    for obj in iterator:\n  File \"\/home\/master\/anaconda3\/lib\/python3.7\/site-packages\/pyspark\/python\/lib\/pyspark.zip\/pyspark\/serializers.py\", line 334, in _batched\n    for item in iterator:\n  File \"&lt;string&gt;\", line 1, in &lt;lambda&gt;\n  File \"\/home\/master\/anaconda3\/lib\/python3.7\/site-packages\/pyspark\/python\/lib\/pyspark.zip\/pyspark\/worker.py\", line 85, in &lt;lambda&gt;\n    return lambda *a: f(*a)\n  File \"\/home\/master\/anaconda3\/lib\/python3.7\/site-packages\/pyspark\/python\/lib\/pyspark.zip\/pyspark\/util.py\", line 99, in wrapper\n    return f(*args, **kwargs)\n  File \"&lt;ipython-input-18-8616f0dc6e22&gt;\", line 6, in pred\n  File \"\/home\/master\/anaconda3\/lib\/python3.7\/site-packages\/keras_preprocessing\/text.py\", line 279, in texts_to_sequences\n    return list(self.texts_to_sequences_generator(texts))\n  File \"\/home\/master\/anaconda3\/lib\/python3.7\/site-packages\/keras_preprocessing\/text.py\", line 310, in texts_to_sequences_generator\n    self.split)\n  File \"\/home\/master\/anaconda3\/lib\/python3.7\/site-packages\/keras_preprocessing\/text.py\", line 43, in text_to_word_sequence\n    text = text.lower()\nAttributeError: 'NoneType' object has no attribute 'lower'\n<\/code><\/pre>\n\n<p>Here is my code<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>sc = SparkContext('local')\nspark = SparkSession(sc)\nreal =spark.read.json('data.txt')\nimport pickle\nwith open('tokenizer.pickle', 'rb') as handle:\n    tokenizer = pickle.load(handle)\nfrom keras.models import load_model\nlstm = load_model('lstm_best_weights.03-0.8345.hdf5')\nfrom keras.preprocessing.sequence import pad_sequences\n\ndef pred (st):\n    str(st)\n    st = [st]\n    sequences_test = tokenizer.texts_to_sequences(st)\n    x = pad_sequences(sequences_test, maxlen=45)\n    result = lstm.predict(x.reshape(1,45),batch_size=1,verbose = 2)[0]\n    if (result[0] &gt; 0.5):\n        result = 1\n    else:\n        result = 0 \n    return result\ndef parse_datetime(date):\n    date = time.strftime('%Y-%m-%d %H:%M:%S', time.strptime(date,'%a %b %d %H:%M:%S +0000 %Y'))\n    return date\nudf_pred = F.udf(pred,IntegerType())\nudf_date = F.udf(parse_datetime,StringType())\ndef clean_tweet(df):\n    # Supprime les chants inuiles\n    to_drop = ['contributors','coordinates','display_text_range','favorited','geo','id_str',\n              'n_reply_to_screen_name','in_reply_to_status_id','in_reply_to_status_id_str',\n              'in_reply_to_user_id','in_reply_to_status_id_str','in_reply_to_user_id',\n              'in_reply_to_user_id_str','is_quote_status','lang','metadata',\n              'place','retweeted','source','truncated','in_reply_to_screen_name',\n              'possibly_sensitive','quoted_status_id','quoted_status_id_str',\n              'source','text','truncated','withheld_copyright','withheld_in_countries',\n              'withheld_scope']\n    d = df.drop(*to_drop)\n    # Aplati le json pour r\u00e9cup\u00e9rer les champs importants\n    a =d.withColumn('user_create',d.user.created_at)\\\n    .withColumn('created_at',d.created_at)\\\n    .withColumn('full_text',d.full_text)\\\n    .withColumn('user_desc',d.user.description)\\\n    .withColumn('user_fav_count',d.user.favourites_count)\\\n    .withColumn('user_followers_count',d.user.followers_count)\\\n    .withColumn('user_statuses_count',d.user.statuses_count)\\\n    .withColumn('user_friends_count',d.user.friends_count)\\\n    .withColumn('user_retweet_count',d.retweet_count)\\\n    .withColumn('favorite_count',d.favorite_count)\\\n    .withColumn('RT_created_at',d.retweeted_status.created_at)\\\n    .withColumn('RT_favorite_count',d.retweeted_status.favorite_count)\\\n    .withColumn('RT_retweet_count',d.retweeted_status.retweet_count)\\\n    .withColumn('RT_retweet_full_text',d.retweeted_status.full_text)\\\n    .withColumn('RT_QT_created_at',d.retweeted_status.quoted_status.created_at)\\\n    .withColumn('RT_QT_favorite_count',d.retweeted_status.quoted_status.favorite_count)\\\n    .withColumn('RT_QT_full_text',d.retweeted_status.quoted_status.full_text)\\\n    .withColumn('RT_QT_retweet_count',d.retweeted_status.quoted_status.retweet_count)\\\n    .withColumn('RT_QT_favorite_count',d.retweeted_status.quoted_status.favorite_count)\\\n    .withColumn('QT_created_at',d.quoted_status.created_at)\\\n    .withColumn('QT_favorite_count',d.quoted_status.favorite_count)\\\n    .withColumn('QT_full_text',d.quoted_status.created_at)\\\n    .withColumn('QT_retweet_count',d.quoted_status.retweet_count)\\\n    .drop('quoted_status')\\\n    .drop('user')\\\n    .drop('retweeted_status')\n    # Nettoyage du texte\n    b = a.withColumn('text_sans_http',F.regexp_replace(F.col(\"full_text\"), \"https?:\/\/[A-Za-z0-9.\/]+\", \"\"))\\\n    .withColumn('text_sans_@',F.regexp_replace(F.col(\"text_sans_http\"), \"@[A-Za-z0-9]+\", \"\"))\\\n    .withColumn('text_sans_RT',F.regexp_replace(F.col(\"text_sans_@\"), \"^RT\", \"\"))\\\n    .withColumn('cleaned',F.regexp_replace(F.col(\"text_sans_RT\"), \"^&amp;amp\", \"&amp;\"))\\\n    .withColumn('clean',F.regexp_replace(F.col(\"cleaned\"), \"(@[A-Za-z0-9]+)|([^0-9A-Za-z \\t])|(\\w+:\\\/\\\/\\S+)\", \" \"))\\\n    .drop('cleaned','text_sans_RT','text_sans_@','text_sans_http','full_text')\\\n    .withColumn('created_date',udf_date(F.col(\"created_at\")))\\\n    .drop('created_at')\\\n    .withColumn('clean',F.lower(F.col('clean')))\\\n    .withColumn('sentiment',udf_pred(F.col(\"clean\")))\n    return b\ndf = clean_tweet(real)\ndf.repartition(1).write.json('result')\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57677160,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2019-08-27T14:42:35.613","AcceptedAnswerId":null,"OwnerUserId":7805642.0,"Title":"cannot import name 'keras_export' from 'tensorflow.python.util.tf_export'","Body":"<p>I am trying to use the BERT in Keras with keras_bert and tensorflow, the complet code in <a href=\"https:\/\/colab.research.google.com\/github\/CyberZHG\/keras-bert\/blob\/master\/demo\/tune\/keras_bert_classification_tpu.ipynb#scrollTo=gGzsxkLTpRrs&amp;uniqifier=1\" rel=\"nofollow noreferrer\">her<\/a>, but I get this error:<\/p>\n\n<p>because of this lien:<\/p>\n\n<pre><code>from tensorflow.python.util.tf_export import keras_export\n<\/code><\/pre>\n\n<p>the error is:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nImportError                               Traceback (most recent call last)\n&lt;ipython-input-18-eea03a4a6625&gt; in &lt;module&gt;\n----&gt; 1 from tensorflow.python.util.tf_export import keras_export\n\nImportError: cannot import name 'keras_export' from 'tensorflow.python.util.tf_export' (C:\\Users\\DILAW\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\util\\tf_export.py)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57677207,"AnswerCount":0,"Tags":"<python><tensorflow><recurrent-neural-network>","CreationDate":"2019-08-27T14:45:00.747","AcceptedAnswerId":null,"OwnerUserId":10291100.0,"Title":"Tensorflow RNN Predicting stock prices - weird prediction plot","Body":"<p><br>\nI have dataset with 1212 samples every value is stock price. I created RNN model to predict them but when I generated prediction and original plot i saw that prediction plot look very weird, so here I am. And i dont know how to make predictions to more look like original plot<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/PCxQZ.png\" rel=\"nofollow noreferrer\">PLOT Last-Original<\/a> <br><\/p>\n\n<p>Config:<\/p>\n\n<pre><code>n_inputs=1 \nn_outputs=1\nn_neurons=48,\nn_iterations=1212\nn_time_steps=12\nlearning_rate=.03\n<\/code><\/pre>\n\n<p>LSTM:<\/p>\n\n<pre><code>lstm_cell = tf.contrib.rnn.OutputProjectionWrapper(tf.contrib.rnn.BasicLSTMCell(num_units=self.n_neurons, activation=tf.nn.sigmoid), output_size=self.n_outputs)\noutputs, sates = tf.nn.dynamic_rnn(lstm_cell, self.x, dtype=tf.float32)\n<\/code><\/pre>\n\n<p>Loss and optimizer:<\/p>\n\n<pre><code>loss = tf.reduce_mean(tf.square(outputs - self.y))\noptimizer = tf.train.AdamOptimizer(learning_rate=self.learning_rate).minimize(loss)\n<\/code><\/pre>\n\n<p>Batch system:<\/p>\n\n<pre><code>def next_batch(self, data):\n    rand_start = np.random.randint(len(data) - self.n_time_steps)\n    y_batch = np.array(data[rand_start:rand_start + self.n_time_steps + 1]).reshape(1, self.n_time_steps + 1)\n\n    return y_batch[: ,:-1].reshape(-1, self.n_time_steps, 1), y_batch[:, 1:].reshape(-1, self.n_time_steps, 1)      \n<\/code><\/pre>\n\n<p>Training:<\/p>\n\n<pre><code>for i in range(self.n_iterations):    \n    x_batch, y_batch = self.next_batch(data)\n\n    sess.run(optimizer, feed_dict={\n        self.x:x_batch,\n        self.y:y_batch\n    })\n<\/code><\/pre>\n\n<p>Here's how i predict values - I restore saved model<\/p>\n\n<pre><code>graph = tf.get_default_graph()\n\nx = graph.get_tensor_by_name('x_input:0')\noutputs_tensor = graph.get_tensor_by_name('outputs:0')\nn_time_steps = graph.get_tensor_by_name('n_time_steps:0')\n\nn_time_steps = sess.run(n_time_steps)\ndata = data.reshape(-1, n_time_steps, 1)\n\noutputs_values = sess.run(outputs_tensor, feed_dict={\n    x: data\n})\n<\/code><\/pre>\n\n<p>And in other file:<\/p>\n\n<pre><code>selected = list(data[-48:])\nn_time_steps = 48\nfor _ in range(48):\n    part = selected[-n_time_steps:]\n    val = predictor.predict(part)\n\nselected.append(val[0, -1, 0])\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57677648,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-27T15:07:41.310","AcceptedAnswerId":null,"OwnerUserId":7788938.0,"Title":"Why tf.gradient with tf.where returns None?","Body":"<p>I would like to do change input or\/and output according to some condition. But the grad is None. <\/p>\n\n<p>How can I get the gradient and keep the selection like the code below?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>input1, input2 = ..., ...\noutput1, output2 = model(input1), model(input2)\n\ninput = tf.where(tf.less(output1, output2), input1, input2)\noutput = tf.where(tf.less(output1, output2), output1, output2)\n\ngrad, = tf.gradient(output, input)\n<\/code><\/pre>\n\n<p>I print the type of input and output, their dimensions are the same as input1\/output1(input2\/output2). And if I only compute tf.gradient(output1,input1), it is no problem. What's the difference between them?<\/p>\n","answers":[{"AnswerId":"57678696","CreationDate":"2019-08-27T16:15:55.283","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You can just do this:<\/p>\n\n<pre><code>input1, input2 = ..., ...\noutput1, output2 = model(input1), model(input2)\nmask = tf.less(output1, output2)\ninput = tf.where(mask, input1, input2)\noutput = tf.where(mask, output1, output2)\ngrad = tf.add(*tf.gradients(output, [input1, input2]))\n<\/code><\/pre>\n\n<p><code>tf.gradient<\/code> will return two tensors with zeros on the places where the corresponding input has not been selected. Hence, <code>grad<\/code> will hold the correct aggregated gradient for <code>input<\/code>.<\/p>\n\n<p>Your original approach does not work because, from the point of view of TensorFlow, there is no dependency between <code>input<\/code> and <code>output<\/code>. <code>input<\/code> is computed from <code>input1<\/code> and <code>input2<\/code> and <code>output<\/code> is also computed from <code>input1<\/code> and <code>input2<\/code>, but there is no path in the graph from <code>input<\/code> to <code>output<\/code>, so there is no gradient.<\/p>\n"}]}
+{"QuestionId":57677834,"AnswerCount":2,"Tags":"<python><tensorflow><keras><recurrent-neural-network><tensor>","CreationDate":"2019-08-27T15:17:44.883","AcceptedAnswerId":57685279.0,"OwnerUserId":9372721.0,"Title":"How to combine input with output tensors to create a recurrent layer?","Body":"<p>I'm trying to change a layer that calculates an output with ny outputs to a layer that calculates a recurrent output so the output has the same shape as the input. For example, consider the following<\/p>\n\n<pre><code>nt = 1000\nnx_in = 8\nny = 2\n\nx_train = np.array(shape=(nt, nx_in))\ninput = keras.Input(shape=(1, None, x_train.shape[1]), name='x_input')\noutput_ny = layers.Dense(ny)(x_input)\n<\/code><\/pre>\n\n<p>The above generates the expected results.  Now I would like to create a recurrent layer by generating a new output tensor that has the same shape as the input tensor and is created by taking one value from tensor output_ny, and (nx_in\/ny-1), or 3, values from tensor input<\/p>\n\n<pre><code>print('x_input.shape: ', x_input.shape)\nprint('ny_output.shape:', ny_output.shape)\nprint('max_lag: ', max_lag)\n\noutput_list = list()\nky_start = 0\nmax_lag = 5\nstored_lags = max_lag - 1\nfor iy in range(ny):\n    ky_end = ky_start + stored_lags - 1\n    print('append output, {}:{}'.format(iy, iy+1))\n    output_list.append(ny_output[:, :, :, iy:(iy+1)]) \n    print('append input, {}:{}'.format(ky_start, ky_end))\n    output_list.append(x_input[:, :, :, ky_start:ky_end])\n    ky_start = ky_end + 1 \n\noutputs = tf.unstack(output_list, axis=3)\n<\/code><\/pre>\n\n<p>The printed output is<\/p>\n\n<pre><code>x_input.shape:  (?, 1, ?, 8)\nny_output.shape: (?, 1, ?, 2)\nmax_lag:  5\nappend output, 0:1\nappend input, 0:3\nappend output, 1:2\nappend input, 4:7\n<\/code><\/pre>\n\n<p>This generates the following error message<\/p>\n\n<pre><code>ValueError: Dimension 3 in both shapes must be equal, but are 1 and 3. Shapes are [?,1,?,1] and [?,1,?,3].\n    From merging shape 2 with other shapes. for 'packed' (op: 'Pack') with input shapes: [?,1,?,1], [?,1,?,3], [?,1,?,1], [?,1,?,3].\n<\/code><\/pre>\n\n<p>How to generate a new output tensor that has the same shape as the input tensor, and is built by appending one element from tensor output_ny, and 3 elements from tensor input, for each ny?<\/p>\n","answers":[{"AnswerId":"57684441","CreationDate":"2019-08-28T02:50:40.377","ParentId":null,"OwnerUserId":"1494821","Title":null,"Body":"<p>The Tensorflow \"graph\" is a Directed Acyclic Graph of computations. The backpropagation algorithm walks this graph backwards, and prediction walks it forwards.<\/p>\n\n<p>My understanding is that you are trying to introduce a cycle into the graph. This will not work.<\/p>\n\n<p>If you start from a basic implementation of neural networks, you can add your own recurrent cell. <\/p>\n"},{"AnswerId":"57685279","CreationDate":"2019-08-28T04:56:08.173","ParentId":null,"OwnerUserId":"7164381","Title":null,"Body":"<p>To concatenate your layers, you can use keras.layers.Concatenate:<\/p>\n\n<p><code>outputs = keras.layers.Concatenate(axis=3)(output_list)<\/code><\/p>\n\n<p>Thus, if you print outputs shape, it gives the expected result:<\/p>\n\n<pre><code>&gt; print(outputs.shape)\n(?, 1, ?, 8)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57677842,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-08-27T15:18:07.740","AcceptedAnswerId":null,"OwnerUserId":11984636.0,"Title":"heatmap refinement by using ConvLstm2D","Body":"<p>I'm going to refine a sequential heatmaps which are related to frames of a video by considering their time correlation.\nI used ConvLSTM2D, but the output looks like the input average! and the accuracy is less than input.\nEach heatmap shows spatial probability of each joint in each frame.\nwould you help me to find the problem?<\/p>\n\n<pre><code>ModelInputSize = 300\nTimeStep = 7\nJoint= 16\ndef get_model(): \n\nX = Input(shape=(TimeStep, ModelInputSize, ModelInputSize, Joint),\n                 batch_shape=(BatchSize,TimeStep, ModelInputSize, ModelInputSize, Joint),name='X')\n\n\n\ntemp = ConvLSTM2D(filters=Joint, kernel_size=(3, 3),  # 16\n           padding='same', return_sequences=True)(X)\n\n\n\ntemp = BatchNormalization()(temp)\n\nY = Conv3D(filters=Joint, kernel_size=(3, 3, 3),  # 16\n                 activation='sigmoid',  # 'sigmoid',\n                 padding='same', data_format='channels_last')(temp)\n\nY = Conv3D(filters=Joint, kernel_size=(1, 1, 1),  # 16\n           activation='sigmoid',  # 'sigmoid',\n           padding='same', data_format='channels_last')(Y)\n\ntemp = Conv3D(filters=Joint, kernel_size=(1,1,1),  # 16\n           activation='sigmoid',  # 'sigmoid',\n           padding='same', data_format='channels_last')(X)\n\nY = keras.layers.Add()([temp, Y])\n\nmodel = Model(input=X, output=Y)\nmodel.compile(loss='mean_squared_error', optimizer='sgd')  # 'sgd'\nmodel.summary()\nreturn model\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57677980,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-27T15:27:04.340","AcceptedAnswerId":null,"OwnerUserId":9120831.0,"Title":"Feed Dict of Multiple Inputs, Correct Way to Think About Tensorflow Batching","Body":"<p>I am attempting to run some matrix computations (non neural network related) using tensorflow. I would like to evaluate my computational graph for several inputs and am wondering about the most \"tensorflow\" like way to do this.<\/p>\n\n<p>I plan to eventually be running the below computations with a 3000 by 3000 matrix A and about ~1000 of a 3000 by 1 vector v, each several hundred of thousand times. I could run the evals in a for loop over individual vectors v but that seems like the wrong way to do it.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tf.reset_default_graph()\n\nA = tf.Variable([[2.,3.],[4.,5.]],dtype=tf.float32)\nv = tf.placeholder(dtype=tf.float32)\n\nres1 = v @ A @ tf.transpose(v)\n\ntest1 = [[6.,7.]]\ntest2 = [[8.,9.]]\ntest12 = [[6.,7.],[8.,9.]]\n\nres2 = tf.diag_part(v @ A @ tf.transpose(v))\nres3 = tf.reshape(v,[2,1,2]) @ (tf.ones([2, 2, 2])*A) @ tf.transpose(tf.reshape(v,[2,1,2]),perm=[0,2,1]) \n\nwith tf.Session() as sess:\n    init = tf.global_variables_initializer()\n    sess.run(init)\n\n    print('res1: test1','\\n',res1.eval(feed_dict={v:test1}),'\\n')\n    print('res1: test2','\\n',res1.eval(feed_dict={v:test2}),'\\n')\n    print('res1: test12','\\n',res1.eval(feed_dict={v:test12}),'\\n')\n\n    # proposals\n    print('res2: test12','\\n',res2.eval(feed_dict={v:test12}),'\\n')\n    print('res3: test12','\\n',res3.eval(feed_dict={v:test12}),'\\n')\n<\/code><\/pre>\n\n<p>The outcome I would like are the ones corresponding to res2 or res3.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>res1: test1 \n [[611.]] \n\nres1: test2 \n [[1037.]] \n\nres1: test12 \n [[ 611.  797.]\n [ 795. 1037.]] \n\nres2: test12 \n [ 611. 1037.] \n\nres3: test12 \n [[[ 611.]]\n\n [[1037.]]] \n<\/code><\/pre>\n\n<p>res2, especially for large NxN, seems like an odd way to do it because it will discard most of the computations (off diagonals). <\/p>\n\n<p>res3 seems memory inefficient as it copies matrix A many many times. <\/p>\n\n<p>Is there another way to do this or does tensorflow's graph optimization make it unnecessary to worry about this?<\/p>\n","answers":[]}
+{"QuestionId":57678712,"AnswerCount":1,"Tags":"<c++><tensorflow><build><bazel>","CreationDate":"2019-08-27T16:16:47.857","AcceptedAnswerId":null,"OwnerUserId":11984771.0,"Title":"Issue building Tensorflow using Bazel (Windows) => to use Tensorflow C++ API","Body":"<p>My initial goal was simply to use Tensorflow's C++ API to read a \".pb\" frozen model and make predictions with it. Was using Keras to prototype it, managed to freeze the model and try it with Tensorflow on Python ; everything goes smoothly.\nSo i basically JUST need the C++ API to work.<\/p>\n\n<p>I'm using an i5-8400T computer, no GPU, 8GB ram. All versions of bazel and tensorflow were downloaded today (27\/08\/2019).<\/p>\n\n<p>First thing I did was to follow <a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/cc\" rel=\"nofollow noreferrer\">tensorflow's guide<\/a> of course. Ran the little <code>bazel run -c opt \/\/tensorflow\/cc\/example:example<\/code>, and that... several times.<\/p>\n\n<p>First time, needed to downgrade my bazel version to <code>0.26.1<\/code> to make tensorflow's <code>configure.py<\/code> happy.<\/p>\n\n<p>Then I had the same problem as a few people had apparently, which is an error fetching <code>io_bazel_rules_docker<\/code>, which then stops the rest of the build of course. (what follows is what i got as an output here)<\/p>\n\n<hr>\n\n<pre><code>\nbazel : INFO: Options provided by the client:\nAu caract\u00e8re Ligne:1 : 1\n+ bazel run -c opt \/\/tensorflow\/cc\/example:example &gt; error.txt 2&gt;&amp;1\n+ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n    + CategoryInfo          : NotSpecified: (INFO: Options provided by the client::String) [], RemoteException\n    + FullyQualifiedErrorId : NativeCommandError\n\n  Inherited 'common' options: --isatty=0 --terminal_columns=80\n\nINFO: Options provided by the client:\n  Inherited 'build' options: --python_path=C:\/Users\/IMEDSYS\/AppData\/Local\/Programs\/Python\/Python37\/python.exe\n\nINFO: Reading rc options for 'run' from c:\\users\\imedsys\\bin\\tensorflow\\.bazelrc:\n  Inherited 'build' options: --apple_platform_type=macos --define framework_shared_object=true --define \nopen_source_build=true --define=use_fast_cpp_protos=true --define=allow_oversize_protos=true \n--spawn_strategy=standalone --strategy=Genrule=standalone -c opt --cxxopt=-std=c++14 --host_cxxopt=-std=c++14 \n--announce_rc --define=grpc_no_ares=true --define=PREFIX=\/usr --define=LIBDIR=$(PREFIX)\/lib \n--define=INCLUDEDIR=$(PREFIX)\/include\n\nINFO: Reading rc options for 'run' from c:\\users\\imedsys\\bin\\tensorflow\\.tf_configure.bazelrc:\n  Inherited 'build' options: --action_env \nPYTHON_BIN_PATH=C:\/Users\/IMEDSYS\/AppData\/Local\/Programs\/Python\/Python37\/python.exe --action_env \nPYTHON_LIB_PATH=C:\/Users\/IMEDSYS\/AppData\/Local\/Programs\/Python\/Python37\/lib\/site-packages \n--python_path=C:\/Users\/IMEDSYS\/AppData\/Local\/Programs\/Python\/Python37\/python.exe --config monolithic --copt=-w \n--host_copt=-w --copt=-DWIN32_LEAN_AND_MEAN --host_copt=-DWIN32_LEAN_AND_MEAN --copt=-DNOGDI --host_copt=-DNOGDI \n--verbose_failures --distinct_host_configuration=false --action_env TF_CONFIGURE_IOS=0\n\nINFO: Found applicable config definition build:monolithic in file c:\\users\\imedsys\\bin\\tensorflow\\.bazelrc: --define \nframework_shared_object=false\n\nLoading: \n\nLoading: 0 packages loaded\n\nINFO: Call stack for the definition of repository 'io_bazel_rules_docker' which is a git_repository (rule definition \nat C:\/users\/imedsys\/_bazel_imedsys\/wmdtvioi\/external\/bazel_tools\/tools\/build_defs\/repo\/git.bzl:252:18):\n - C:\/users\/imedsys\/_bazel_imedsys\/wmdtvioi\/external\/bazel_toolchains\/repositories\/repositories.bzl:37:9\n - C:\/users\/imedsys\/bin\/tensorflow\/WORKSPACE:35:1\n\nERROR: An error occurred during the fetch of repository 'io_bazel_rules_docker':\n   Traceback (most recent call last):\n    File \"C:\/users\/imedsys\/_bazel_imedsys\/wmdtvioi\/external\/bazel_tools\/tools\/build_defs\/repo\/git.bzl\", line 234\n        _clone_or_update(ctx)\n    File \"C:\/users\/imedsys\/_bazel_imedsys\/wmdtvioi\/external\/bazel_tools\/tools\/build_defs\/repo\/git.bzl\", line 74, in \n_clone_or_update\n        fail((\"error cloning %s:\\n%s\" % (ctx....)))\nerror cloning io_bazel_rules_docker:\nfatal: not a git repository (or any parent up to mount point \/mnt)\nStopping at filesystem boundary (GIT_DISCOVERY_ACROSS_FILESYSTEM not set).\n\/bin\/bash: line 1: cd: C:\/users\/imedsys\/_bazel_imedsys\/wmdtvioi\/external: No such file or directory\n+ cd C:\/users\/imedsys\/_bazel_imedsys\/wmdtvioi\/external\n\/bin\/bash: line 3: cd: C:\/users\/imedsys\/_bazel_imedsys\/wmdtvioi\/external: No such file or directory\n+ git -C C:\/users\/imedsys\/_bazel_imedsys\/wmdtvioi\/external\/io_bazel_rules_docker reset --hard \n251f6a68b439744094faff800cd029798edf9faa\nfatal: cannot change to 'C:\/users\/imedsys\/_bazel_imedsys\/wmdtvioi\/external\/io_bazel_rules_docker': No such file or \ndirectory\n+ git -C C:\/users\/imedsys\/_bazel_imedsys\/wmdtvioi\/external\/io_bazel_rules_docker fetch '' origin \n251f6a68b439744094faff800cd029798edf9faa:251f6a68b439744094faff800cd029798edf9faa\nfatal: cannot change to 'C:\/users\/imedsys\/_bazel_imedsys\/wmdtvioi\/external\/io_bazel_rules_docker': No such file or \ndirectory\n+ git -C C:\/users\/imedsys\/_bazel_imedsys\/wmdtvioi\/external\/io_bazel_rules_docker fetch origin \n251f6a68b439744094faff800cd029798edf9faa:251f6a68b439744094faff800cd029798edf9faa\nfatal: cannot change to 'C:\/users\/imedsys\/_bazel_imedsys\/wmdtvioi\/external\/io_bazel_rules_docker': No such file or \ndirectory\n\nERROR: ... *idem*\n\nERROR: ... *idem*\n\nINFO: Elapsed time: 2.346s\n\nINFO: 0 processes.\n\nFAILED: Build did NOT complete successfully (0 packages loaded)\n\nERROR: Build failed. Not running target\n\nFAILED: Build did NOT complete successfully (0 packages loaded)\n\n***\n<\/code><\/pre>\n\n<p>I so tried applying what <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/28824\" rel=\"nofollow noreferrer\">this issue<\/a> was offering, which involves going to the <a href=\"https:\/\/github.com\/bazelbuild\/rules_docker\" rel=\"nofollow noreferrer\">rules_docker<\/a> repository and replace targets with the latest ones.<\/p>\n\n<p>The thing is even after this you get the same kind of errors with the protobuf targets, which I get hash errors for if i apply the same strategy.<\/p>\n\n<p>I then tried to do it with CMake as many were saying last year that bazel building on windows was not yet usable. I used two similar guides, one of them <a href=\"http:\/\/mc.ai\/building-tensorflow-c-shared-library-on-windows\/\" rel=\"nofollow noreferrer\">here<\/a>, and so tried to build with <\/p>\n\n<pre><code>cmake .. -A x64 -DCMAKE_BUILD_TYPE=Release -DSWIG_EXECUTABLE=C:\\swigwin-3.0.12\\swig.exe -DPYTHON_EXECUTABLE=C:\\Anaconda3\\envs\\tensorflow\\python.exe -DPYTHON_LIBRARIES=C:\\Anaconda3\\envs\\tensorflow\\libs\\python35.lib\n<\/code><\/pre>\n\n<p>at first, and then with <\/p>\n\n<pre><code>`cmake .. -A x64 -DCMAKE_BUILD_TYPE=Release -DSWIG_EXECUTABLE=C:\\swigwin-3.0.12\\swig.exe -DPYTHON_EXECUTABLE=C:\\Anaconda3\\envs\\tensorflow\\python.exe -DPYTHON_LIBRARIES=C:\\Anaconda3\\envs\\tensorflow\\libs\\python35.lib -Dtensorflow_BUILD_PYTHON_BINDINGS=OFF -Dtensorflow_ENABLE_GRPC_SUPPORT=OFF -Dtensorflow_BUILD_SHARED_LIB=ON`\n<\/code><\/pre>\n\n<p>and i get those errors : <\/p>\n\n<hr>\n\n<pre><code>cmake : CMake Warning at CMakeLists.txt:9 (message):\nAu caract\u00e8re Ligne:1 : 1\n+ cmake .. -A x64 -DCMAKE_BUILD_TYPE=Release -DSWIG_EXECUTABLE=C:\\Users ...\n+ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n    + CategoryInfo          : NotSpecified: (CMake Warning a...xt:9 (message)::String) [], RemoteException\n    + FullyQualifiedErrorId : NativeCommandError\n\n  Your current cmake generator is set to use 32 bit toolset architecture.\n\n  This may cause \"compiler out of heap space\" errors when building.  Consider\n\n  using the flag -Thost=x64 when running cmake.  Ignore this if you are on\n\n  CMake GUI.\n\n\n\n-- Selecting Windows SDK version  to target Windows 10.0.17763.\n-- C:\/Users\/IMEDSYS\/bin\/tensorflow\/tensorflow\/contrib\/cmake\/build\/abseil_cpp\/src\/abseil_cpp_build\n-- Configuring done\nCMake Error at tf_core_ops.cmake:77 (add_library):\n\n  Cannot find source file:\n\n    C:\/Users\/IMEDSYS\/bin\/tensorflow\/tensorflow\/contrib\/factorization\/ops\/clustering_ops.cc\n\n  Tried extensions .c .C .c++ .cc .cpp .cxx .cu .m .M .mm .h .hh .h++ .hm\n\n  .hpp .hxx .in .txx\n\nCall Stack (most recent call first):\n\n  tf_core_ops.cmake:95 (GENERATE_CONTRIB_OP_LIBRARY)\n\n  CMakeLists.txt:594 (include)\n\n\n\nCMake Error at tf_core_ops.cmake:77 (add_library):\n\n  Cannot find source file:\n\n    C:\/Users\/IMEDSYS\/bin\/tensorflow\/tensorflow\/contrib\/rnn\/ops\/gru_ops.cc\n\n  Tried extensions .c .C .c++ .cc .cpp .cxx .cu .m .M .mm .h .hh .h++ .hm\n\n  .hpp .hxx .in .txx\n\nCall Stack (most recent call first):\n\n  tf_core_ops.cmake:107 (GENERATE_CONTRIB_OP_LIBRARY)\n\n  CMakeLists.txt:594 (include)\n\n\n\nCMake Error at tf_core_ops.cmake:77 (add_library):\n\n  Cannot find source file:\n\n    C:\/Users\/IMEDSYS\/bin\/tensorflow\/tensorflow\/contrib\/rnn\/ops\/lstm_ops.cc\n  Tried extensions .c .C .c++ .cc .cpp .cxx .cu .m .M .mm .h .hh .h++ .hm\n\n  .hpp .hxx .in .txx\n\nCall Stack (most recent call first):\n\n  tf_core_ops.cmake:108 (GENERATE_CONTRIB_OP_LIBRARY)\n\n  CMakeLists.txt:594 (include)\n\n\n\nCMake Error at tf_core_kernels.cmake:205 (add_library):\n\n  Cannot find source file:\n\n    C:\/Users\/IMEDSYS\/bin\/tensorflow\/tensorflow\/contrib\/factorization\/kernels\/clustering_ops.cc\n\n  Tried extensions .c .C .c++ .cc .cpp .cxx .cu .m .M .mm .h .hh .h++ .hm\n\n  .hpp .hxx .in .txx\n\nCall Stack (most recent call first):\n\n  CMakeLists.txt:596 (include)\n\n\n\nCMake Error at tf_core_ops.cmake:77 (add_library):\n\n  No SOURCES given to target: tf_contrib_factorization_clustering_ops\n\nCall Stack (most recent call first):\n\n  tf_core_ops.cmake:95 (GENERATE_CONTRIB_OP_LIBRARY)\n\n  CMakeLists.txt:594 (include)\n\n\n\nCMake Error at tf_core_ops.cmake:77 (add_library):\n\n  No SOURCES given to target: tf_contrib_rnn_gru_ops\nCall Stack (most recent call first):\n\n  tf_core_ops.cmake:107 (GENERATE_CONTRIB_OP_LIBRARY)\n\n  CMakeLists.txt:594 (include)\n\n\n\nCMake Error at tf_core_ops.cmake:77 (add_library):\n\n  No SOURCES given to target: tf_contrib_rnn_lstm_ops\n\nCall Stack (most recent call first):\n\n  tf_core_ops.cmake:108 (GENERATE_CONTRIB_OP_LIBRARY)\n\n  CMakeLists.txt:594 (include)\n\n\n\nCMake Error at tf_core_ops.cmake:77 (add_library):\n\n  No SOURCES given to target: tf_contrib_tpu_ops\n\nCall Stack (most recent call first):\n\n  tf_core_ops.cmake:115 (GENERATE_CONTRIB_OP_LIBRARY)\n\n  CMakeLists.txt:594 (include)\n\n\n\nCMake Error at tf_core_kernels.cmake:205 (add_library):\n\n  No SOURCES given to target: tf_core_kernels\n\nCall Stack (most recent call first):\n\n  CMakeLists.txt:596 (include)\n\n\n\nCMake Generate step failed.  Build files cannot be regenerated correctly.\n<\/code><\/pre>\n\n<hr>\n\n<p>This time i feel like it's simply that CMake looks for files that are either depreciated, either relocated.<\/p>\n\n<h1>SO,<\/h1>\n\n<p>my questions is rather simple : <\/p>\n\n<ul>\n<li>does someone has 2 version numbers of both bazel and tensorflow that actually work together (because it seems that it does work for some people => I consider trying voodoo)?<\/li>\n<li>does someone has a very straightforward solution to use Tensorflow's c_api or C++ API ?<\/li>\n<li>does someone knows another way to make a prediction with either Keras or Tensorflow outputs (.hdf5 or .pb) from a C++ built software, without doing as disgusting as \"system(\"python predict.py\")\"?<\/li>\n<\/ul>\n\n<hr>\n\n<p>Thank you in advance for your quick answers! :)<\/p>\n","answers":[{"AnswerId":"57679047","CreationDate":"2019-08-27T16:40:13.843","ParentId":null,"OwnerUserId":"1814498","Title":null,"Body":"<p>I  had done similar work building  older Tensorflow version (not 2.) for VS 2017 using Cmake.  <\/p>\n\n<p>Cuda 9 . Tensorflow 1.14  VS 2017  . <\/p>\n"}]}
+{"QuestionId":57678820,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><keypoint>","CreationDate":"2019-08-27T16:24:14.513","AcceptedAnswerId":null,"OwnerUserId":10667925.0,"Title":"can i retrain exist keypoints detection model like i do for object detection model","Body":"<p>Earlier i used tensorflow object detection api for retrain exist model from <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/detection_model_zoo.md\" rel=\"nofollow noreferrer\">zoo<\/a> for my targets by this <a href=\"https:\/\/tensorflow-object-detection-api-tutorial.readthedocs.io\/en\/latest\/\" rel=\"nofollow noreferrer\">tutorial<\/a><\/p>\n\n<p>So can I do the same for <strong>keypoint<\/strong> detection task?<\/p>\n","answers":[]}
+{"QuestionId":57678886,"AnswerCount":0,"Tags":"<python><keras><neural-network>","CreationDate":"2019-08-27T16:29:10.467","AcceptedAnswerId":null,"OwnerUserId":2208967.0,"Title":"keras fit with y=None with embedding layer","Body":"<p>I tried the following keras Sequential model:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Embedding(vocab_size, neurons, input_length=max_length))\nmodel.add(Flatten())\nmodel.compile(optimizer=Adam(lr=lr), loss='binary_crossentropy', metrics=['acc'])\nprint(model.summary())\nreturn model\n<\/code><\/pre>\n\n<p>i.e.<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nembedding (Embedding)        (None, 44, 300)           8100000   \n_________________________________________________________________\nflatten (Flatten)            (None, 13200)             0         \n=================================================================\nTotal params: 8,100,000\nTrainable params: 8,100,000\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>which I fit like this:<\/p>\n\n<pre><code>encoded_lines = [keras.preprocessing.text.one_hot(d, vocab_size) for d in text_lines]\npadded_lines = keras.preprocessing.sequence.pad_sequences(encoded_lines, maxlen=mymaxlength, padding='post')\nmodel.fit(padded_lines, epochs=10)\n<\/code><\/pre>\n\n<p>where text_lines is an array of lines of text and I'm passing the padded one_hot encoded (I'm using this <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/preprocessing\/text\/one_hot\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/preprocessing\/text\/one_hot<\/a>) as training data and I pass nothing as target data. This trains happily, so my question is what does Keras use as target data? Against what target value is the accuracy computed? I'd guess the only possibility is simply that the target data is the same as the training data.<\/p>\n\n<p>From the documentation of model.fit (<a href=\"https:\/\/keras.io\/models\/model\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/models\/model\/<\/a>):<\/p>\n\n<blockquote>\n  <p>y: Numpy array of target (label) data (if the model has a single\n  output), or list of Numpy arrays (if the model has multiple outputs).\n  If output layers in the model are named, you can also pass a\n  dictionary mapping output names to Numpy arrays. y can be None\n  (default) if feeding from framework-native tensors (e.g. TensorFlow\n  data tensors).<\/p>\n<\/blockquote>\n\n<p>but this doesn't help me much.<\/p>\n","answers":[]}
+{"QuestionId":57678943,"AnswerCount":0,"Tags":"<python><deep-learning><conv-neural-network><pytorch>","CreationDate":"2019-08-27T16:32:50.627","AcceptedAnswerId":null,"OwnerUserId":11940717.0,"Title":"CNN is overfitting to minority class","Body":"<p>I'm new to developing CNNs and currently making a binary image classifier using PyTorch. My dataset was heavily unbalanced, and I've manually augmented my a testing split and training split to be balanced. I have a class 0 (Training set has 6500 images), and class 1 (training set has 5200ish images). When I try to use skorch's fit function, I get a validation accuracy equivalent to the percent of class 1 images in the set, and my prediction function only outputs 1 for all images. <\/p>\n\n<p>This is the tutorial I adapted my CNN from: <a href=\"https:\/\/colab.research.google.com\/github\/dnouri\/skorch\/blob\/master\/notebooks\/Transfer_Learning.ipynb#scrollTo=cane7VRWW3dO\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/github\/dnouri\/skorch\/blob\/master\/notebooks\/Transfer_Learning.ipynb#scrollTo=cane7VRWW3dO<\/a><\/p>\n\n<p>Here's my CNN: (it's adapted from a tutorial)<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nimport torch.nn as nn\nimport torch.optim as optim\nimport torchvision.transforms as transforms\nimport torchvision.datasets as datasets\nimport torchvision.models as models\nimport numpy as np\nfrom numpy import array\nimport pandas as pd\nfrom skorch import NeuralNetClassifier\nfrom skorch.helper import predefined_split\nfrom skorch.callbacks import LRScheduler\nfrom skorch.callbacks import Checkpoint\nfrom skorch.callbacks import Freezer\n\nfrom PIL import Image\nimport skorch\nimport os\nimport cv2\nimport glob\n\nfrom sklearn.pipeline import FeatureUnion, Pipeline\nfrom sklearn.preprocessing import StandardScaler, MinMaxScaler, Normalizer\n#Define transforms (will be same w\/o online transforms)\n#Manually augmented earlier\ntrain_transforms = transforms.Compose([\n    #transforms.RandomResizedCrop(224),\n    #transforms.RandomHorizontalFlip(),\n    transforms.CenterCrop(224),\n    transforms.ToTensor(),\n    transforms.Normalize([0.485, 0.456, 0.406], \n                         [0.229, 0.224, 0.225])\n])\n\nval_transforms = transforms.Compose([\n    #transforms.Resize(256),\n    transforms.CenterCrop(224),\n    transforms.ToTensor(),\n    transforms.Normalize([0.485, 0.456, 0.406], \n                         [0.229, 0.224, 0.225])\n])\n\ntrain_ds = datasets.ImageFolder(train_split_path, train_transforms)\nval_ds = datasets.ImageFolder(test_aug_split_path, val_transforms)\n\ncheckpoint = Checkpoint(\n    f_params='best_model.pt', monitor='valid_acc_best')\n\n#Using ResNet with some layers for now\nclass PretrainedModel(nn.Module):\n\n    def __init__(self, output_features):\n        super().__init__()\n        model = models.resnet152(pretrained=True)\n        #Don't want to change pretrained weights\n        for param in model.parameters():\n            param.requires_grad = False\n\n        num_features = model.fc.in_features\n        fc_layers = nn.Sequential(\n                nn.Linear(num_features, 4096),\n                nn.ReLU(inplace=True),\n                nn.Dropout(p=0.3),\n                nn.Linear(4096, output_features),\n                nn.ReLU(inplace=True),\n                nn.Dropout(p=0.3),\n            )\n        model.fc = fc_layers\n\n        self.model = model\n\n    def forward(self, x):\n        return self.model(x)\n\nuse_cuda = torch.cuda.is_available()\n\nnet = NeuralNetClassifier(\n    module=PretrainedModel,\n    module__output_features = 2,\n    criterion=nn.CrossEntropyLoss,\n    batch_size = 16,\n    lr=0.0001,\n    max_epochs=3,\n    optimizer=optim.Adam,\n    train_split=predefined_split(val_ds),\n    callbacks=[checkpoint],\n    device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n)\n\nnet.fit(train_ds, y=None)\n<\/code><\/pre>\n\n<p>Here are the results of the fit function:<\/p>\n\n<p>Epoch 1: Train loss = .46 Valid_acc = .4301 Valid Loss = .6931<\/p>\n\n<p>Epoch 2: Train loss = .6931 Valid_acc = .4301 Valid Loss = .6931<\/p>\n\n<p>Epoch 3: Train loss = .6931 Valid_acc = .4301 Valid Loss = .6931<\/p>\n\n<p>As it turns out for this particular dataset, exactly 43% of my validation images are Class 1 images. <\/p>\n\n<p>y_pred = net.predict(val_ds) gives me the following:<\/p>\n\n<p>array([1, 1, 1, ..., 1, 1, 1], dtype=int64)<\/p>\n\n<p>I guess I have two questions:<\/p>\n\n<p>1) Is there anything that I have done incorrectly in my initialization of my CNN that would cause this? <\/p>\n\n<p>2) What would cause this, and is there anything I can do to correct for it?<\/p>\n","answers":[]}
+{"QuestionId":57679070,"AnswerCount":0,"Tags":"<python><tensorflow><keras><loss>","CreationDate":"2019-08-27T16:42:26.030","AcceptedAnswerId":null,"OwnerUserId":8262535.0,"Title":"How to properly use custom loss (e.g. dice coefficient) with tensorflow.keras model?","Body":"<p>I am trying to input a custom loss or metric function into tensorflow.keras model compile but am getting an error. It looks like the inputs are not being passed correctly. <\/p>\n\n<p>I have been able to run model compile\/fit with the following tensorflow losses: <\/p>\n\n<pre><code>loss = tf.keras.losses.MeanSquaredError() # Breaks if I remove ()\nloss = tf.nn.sigmoid_cross_entropy_with_logits # Breaks if I add ()\n<\/code><\/pre>\n\n<p>I am confused by the difference in syntax which makes these work (maybe incorrectly?) and if I use a custom dice loss, it doesn't work with either syntax. <\/p>\n\n<p>When I run the custom dice loss below, the input labels is passed correctly as <code>batch_size*height*width<\/code> but the input logits is passed as <code>None,None,None,None<\/code> (does not seem correct?) and the dice loss function errors out. I am looking to do batch optimization, so the loss should be calculated on each batch by the <code>model.fit<\/code>. <\/p>\n\n<pre><code>def generalized_dice(labels, logits):\n    smooth = 1e-17\n    shape = tf.TensorShape(logits.shape).as_list()\n    depth = int(shape[-1])\n    labels = tf.one_hot(labels, depth, dtype=tf.float32)\n    logits = tf.nn.softmax(logits)\n    weights = 1.0 \/ (tf.reduce_sum(labels, axis=[0, 1, 2])**2)\n\n    numerator = tf.reduce_sum(labels * logits, axis=[0, 1, 2])\n    numerator = tf.reduce_sum(weights * numerator)\n\n    denominator = tf.reduce_sum(labels + logits, axis=[0, 1, 2])\n    denominator = tf.reduce_sum(weights * denominator)\n\n    loss = 2.0*(numerator + smooth)\/(denominator + smooth)\n    return loss\n\n\ndef generalized_dice_loss(dice):\n    return 1-dice \n\n\nmodel = tf.keras.Model(inputs=[input_x], outputs=[predictions])\n\nloss = tf.keras.losses.MeanSquaredError() # Breaks if I remove (), with inputs being passed as None's \n# loss = tf.nn.sigmoid_cross_entropy_with_logits \nmetric = generalized_dice # If no parantehses, labels passed as None's; \n# Both tf.nn.sigmoid_cross_entropy_with_logits and generalized_dice don't work if I add () saying inputs and labels must be provided; since I am training in batches, these need to be provided by the model during training, not when I am defining loss?  \n\nmodel.compile(optimizer=tf.keras.optimizers.Adam(),  \n              loss=loss, \n              metrics=[metric]) \n<\/code><\/pre>\n\n<p>Due to the input to dice loss being passed as None's, <\/p>\n\n<pre><code>labels = tf.one_hot(labels, depth, dtype=tf.float32) \n<\/code><\/pre>\n\n<p>Results in: <\/p>\n\n<pre><code>TypeError: Value passed to parameter 'indices' has DataType float32 not in list of allowed values: uint8, int32, int64\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57679120,"AnswerCount":0,"Tags":"<tensorflow><keras><hdf5><google-colaboratory>","CreationDate":"2019-08-27T16:45:30.990","AcceptedAnswerId":null,"OwnerUserId":5838180.0,"Title":"How to speed up neural network in google-colab with hdf5 files?","Body":"<p>I am training a keras neural network in colab, that I want to speed it up. Here is my <a href=\"https:\/\/colab.research.google.com\/drive\/1Wjzf9yZSW92Rr1vqcNgspVqm969b7NF4\" rel=\"nofollow noreferrer\">code<\/a>.  Training takes ~145sec\/epoch on colab, while the same neural net trains for 5sec\/epoch on a server with a k20 GPU. From what I <a href=\"https:\/\/medium.com\/@oribarel\/getting-the-most-out-of-your-google-colab-2b0585f82403\" rel=\"nofollow noreferrer\">read<\/a>, the most probable cause for the slow training is that I have my training files in gdrive and accessing them from colab takes a lot of time. Supposedly, the solution is to do \"memory mapping\" of my data with the file format hdf5. The slow speed of training in colab seems to be a common problem, but I can't find any code solutions utilizing hdf5.<\/p>\n\n<p><strong>I never used hdf5 before so can you give me some code examples about how to do the \"mapping\" in my code?<\/strong> Any other suggestions for speeding up training are appreciated. I have also tried to:<\/p>\n\n<ul>\n<li>use larger <code>batch_size=<\/code> 256 or 512, but the speed didn't change<\/li>\n<li>skip the image augmentation in <code>ImageDataGenerator()<\/code>, but this saved only 25% of time<\/li>\n<li>use <code>tf.distribute.cluster_resolver<\/code> to distribute over multiple GPUs\/TPUs, but I receive the error message \"fit_generator is not supported for models compiled with tf.distribute.Strategy.\"<\/li>\n<\/ul>\n\n<p>Thanks in advance, an answer should be useful to many more in the community!<\/p>\n","answers":[]}
+{"QuestionId":57679829,"AnswerCount":0,"Tags":"<tensorflow><image-processing><aws-lambda><endpoint><tf.keras>","CreationDate":"2019-08-27T17:45:48.963","AcceptedAnswerId":null,"OwnerUserId":3261772.0,"Title":"Combine Multiple TensorFlow (TF) models and pre-processing step in a TF model","Body":"<p>I have created three different Convnets models using Tensorflow Keras API.  The client post an image to my API gateway URL, and the lambda receive an image, I do the pre-processing step and then I invoke the first endpoint and fix the image orientation and then I invoke the second model to get some predictions and finally if the second model predicts a class,  I would invoke another endpoint.  All this logic is in my lambda function and I have passed three endpoint names as an environment variables.   Now,  I also know how to put all models in one endpoint and invoke the models only on one endpoint through boto3, invoke_endpoint method with CustomAttributes argument.  <\/p>\n\n<p>My question is:  Is there a way to have a Tensorflow model which includes (serialize) all these logic (preprocessing -loading TF models and get predictions) and return the json output?  This way my lambda would be much smaller.  I just need to create an endpoint for \"the mTF model\" and then invoke it.<\/p>\n","answers":[]}
+{"QuestionId":57679975,"AnswerCount":2,"Tags":"<python><keras><conv-neural-network><tf.keras>","CreationDate":"2019-08-27T17:56:36.843","AcceptedAnswerId":57689475.0,"OwnerUserId":7566673.0,"Title":"ValueError: Error when checking target: expected dense_35 to have 4 dimensions, but got array with shape (1157, 1)","Body":"<p>I have train &amp; test image data for which Shapes are given below.<\/p>\n\n<pre><code>X_test.shape , y_test.shape , X_train.shape , y_train.shape\n    ((277, 128, 128, 3), (277, 1), (1157, 128, 128, 3), (1157, 1))\n<\/code><\/pre>\n\n<p>I am training a model <\/p>\n\n<pre><code>def baseline_model():\n    filters = 100\n    model = Sequential()\n    model.add(Conv2D(filters, (3, 3), input_shape=(128, 128, 3), padding='same', activation='relu'))\n    #model.add(Dropout(0.2))\n    model.add(BatchNormalization())\n    model.add(Conv2D(filters, (3, 3), activation='relu', padding='same'))\n    model.add(BatchNormalization())\n\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n    #model.add(Flatten())\n\n    model.add(Conv2D(filters, (3, 3), activation='relu', padding='same'))\n    model.add(BatchNormalization())\n    model.add(Conv2D(filters, (3, 3), activation='relu', padding='same'))\n    model.add(Activation('linear'))\n    model.add(BatchNormalization())\n\n    model.add(Dense(512, activation='relu'))\n    model.add(Dense(num_classes, activation='softmax'))\n    # Compile model\n    lrate = 0.01\n    epochs = 10\n    decay = lrate\/epochs\n    sgd = SGD(lr=lrate, momentum=0.9, decay=decay, nesterov=False)\n    model.compile(loss='sparse_categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])\n\n    print(model.summary())\n    return model\n<\/code><\/pre>\n\n<p>But I am getting an error Given below<\/p>\n\n<blockquote>\n  <p>Error when checking target: expected dense_35 to have 4 dimensions,\n  but got array with shape (1157, 1)<\/p>\n<\/blockquote>\n\n<p>Please tell me what mistake I am making and how to fix this. I have attached snapshot of model summary \n<a href=\"https:\/\/i.stack.imgur.com\/aXg9c.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/aXg9c.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"57682152","CreationDate":"2019-08-27T20:55:28.513","ParentId":null,"OwnerUserId":"6180759","Title":null,"Body":"<p>Although dense_35 needs to feed with 4 dimension data, according to the error, the network feed with 2 dimension data which is the label vector.<\/p>\n"},{"AnswerId":"57689475","CreationDate":"2019-08-28T09:45:49.230","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>One thing you have probably forgotten to do is adding a <code>Flatten<\/code> layer right before the first <code>Dense<\/code> layer:<\/p>\n\n<pre><code>model.add(BatchNormalization())\n\nmodel.add(Flatten()) # flatten the output of previous layer before feeding it to Dense layer\nmodel.add(Dense(512, activation='relu'))\n<\/code><\/pre>\n\n<p>You need it because <code>Dense<\/code> layer does not flatten its input; rather, <a href=\"https:\/\/stackoverflow.com\/a\/52092176\/2099607\">it is applied on the last dimension<\/a>.<\/p>\n"}]}
+{"QuestionId":57680010,"AnswerCount":0,"Tags":"<python><keras><deep-learning>","CreationDate":"2019-08-27T17:59:56.523","AcceptedAnswerId":null,"OwnerUserId":5848690.0,"Title":"'TypeError: Not JSON Serilizable' when trying to checkpoint Keras model","Body":"<p>I have implemented a residual network consisting of convolutional layers followed by a few LSTM layers for binary classification of sequences of snapshots.  In order to use this hybrid network, I make extensive use of the <code>TimeDistributed<\/code> layer wrapper on the layers prior to the recurrent layers.  Also, my input data comes in the form of very long sequences of snapshots.  I manually chop up those long sequences into smaller, more manageable sequences by reshaping the data.  However, when I try to use the Keras <code>ModelCheckpoint<\/code> callback to checkpoint the model, I get the following stack trace:<\/p>\n\n<pre><code>8\/8 [==============================] - 6s 728ms\/step - loss: 0.6802 - acc: 0.1250 - val_loss: 2.2920 - val_acc: 0.5000\n\nEpoch 00001: val_acc improved from -inf to 0.50000, saving model to epoch_01.h5\nTraceback (most recent call last):\n  File \"SO_post.py\", line 174, in &lt;module&gt;\n    callbacks=callbacks_list)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/Keras-2.2.4-py3.5.egg\/keras\/engine\/training.py\", line 1178, in fit\n    validation_freq=validation_freq)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/Keras-2.2.4-py3.5.egg\/keras\/engine\/training_arrays.py\", line 224, in fit_loop\n    callbacks.on_epoch_end(epoch, epoch_logs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/Keras-2.2.4-py3.5.egg\/keras\/callbacks.py\", line 152, in on_epoch_end\n    callback.on_epoch_end(epoch, logs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/Keras-2.2.4-py3.5.egg\/keras\/callbacks.py\", line 719, in on_epoch_end\n    self.model.save(filepath, overwrite=True)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/Keras-2.2.4-py3.5.egg\/keras\/engine\/network.py\", line 1139, in save\n    save_model(self, filepath, overwrite, include_optimizer)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/Keras-2.2.4-py3.5.egg\/keras\/engine\/saving.py\", line 415, in save_wrapper\n    save_function(obj, filepath, overwrite, *args, **kwargs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/Keras-2.2.4-py3.5.egg\/keras\/engine\/saving.py\", line 507, in save_model\n    _serialize_model(model, h5dict, include_optimizer)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/Keras-2.2.4-py3.5.egg\/keras\/engine\/saving.py\", line 101, in _serialize_model\n    model_config = json.dumps(model_config, default=get_json_type)\n  File \"\/usr\/lib\/python3.5\/json\/__init__.py\", line 237, in dumps\n    **kw).encode(obj)\n  File \"\/usr\/lib\/python3.5\/json\/encoder.py\", line 198, in encode\n    chunks = self.iterencode(o, _one_shot=True)\n  File \"\/usr\/lib\/python3.5\/json\/encoder.py\", line 256, in iterencode\n    return _iterencode(o, 0)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/Keras-2.2.4-py3.5.egg\/keras\/engine\/saving.py\", line 91, in get_json_type\n    raise TypeError('Not JSON Serializable: %s' % (obj,))\nTypeError: Not JSON Serializable: 4\n<\/code><\/pre>\n\n<p>When I change <code>t = 4<\/code> to <code>t = 8<\/code>, I get the same trace, but with 8 not being JSON Serializable rather than 4.  It just doesn't make sense to me.<\/p>\n\n<p>Most of the SO posts that have anything to do with this error have been resolved by fixing the data type of the new dimension.  I have checked the data types in every place I can think of and they all show 'int' as the data type.<\/p>\n\n<p>I've tried to reduce this as much as possible and here's what I've got.  I'm pretty sure the error is somehow introduced when I reshape the data in the \"Work with data\" section.  Can anyone shed some light on this for me?<\/p>\n\n<p><strong>SO_post.py<\/strong><\/p>\n\n<pre><code>from keras.backend import clear_session\nfrom keras.backend.tensorflow_backend import set_session\nfrom keras.callbacks import ModelCheckpoint\nfrom keras.layers import add, Activation, AveragePooling1D, BatchNormalization, Conv1D, Dense, Flatten, Input, LSTM, MaxPooling1D, ReLU, TimeDistributed\nfrom keras.models import Model\nfrom keras.optimizers import Adam\nfrom keras.regularizers import l2\nimport numpy as np\nimport tensorflow as tf\n\n# Define some parameters to use in the functions\nbatchsize=32\nnum_epochs = 50\nGPU_mem_fraction = 0.5\nadam_lr = 5e-3\ncheckpoint_freq = 1\nN = 32\nsnapshot_len = 128\nchannels = 2\nclasses = 2\n\n# ==============================================================================\n#  Convenience functions\n# ==============================================================================\ndef create_model(input_sh):\n\n  clear_session()\n  num_stacks = 2\n  num_res_blocks = 2\n  num_rec = 2\n  rec_size = 128\n  model = recurrent_resnet(input_sh, num_stacks, num_res_blocks, num_rec, rec_size)\n\n  # Use an Adam optimizer to backpropagate error.\n  adam = Adam(lr=adam_lr)\n  model.compile(loss='categorical_crossentropy', optimizer=adam, metrics=['accuracy'])\n  return model\n\ndef resnet_layer(inputs, num_filters=16, activation='relu', batchnorm=True, conv_first=True):\n\n  conv = Conv1D(num_filters, kernel_size=3, strides=1, padding='same', \\\n                kernel_initializer='he_normal',\n                kernel_regularizer=l2(1e-4))\n\n  x = inputs\n  if conv_first:\n      x = TimeDistributed(conv, input_shape=x.shape[1:])(x)\n      if batchnorm:\n          x = TimeDistributed(BatchNormalization())(x)\n      if activation is not None:\n          x = TimeDistributed(Activation(activation))(x)\n  else:\n      if batchnorm:\n          x = TimeDistributed(BatchNormalization(), input_shape=x.shape[1:])(x)\n      if activation is not None:\n          x = TimeDistributed(Activation(activation))(x)\n      x = TimeDistributed(conv)(x)\n  return x\n\ndef residual_block(x, num_filters, downsample=False):\n\n  strides = 1\n  y = resnet_layer(inputs=x,\n                   num_filters=num_filters)\n  y = resnet_layer(inputs=y,\n                   num_filters=num_filters,\n                   activation=None)\n\n  if downsample:\n      x = resnet_layer(inputs=x,\n                       num_filters=num_filters,\n                       activation=None,\n                       batchnorm=False)\n\n  x = add([x, y])\n\n  x = TimeDistributed(Activation('relu'))(x)\n  return x\n\ndef recurrent_resnet(input_shape, num_stacks=3, num_res_blocks=7,\n                     num_recurrent=2, recurrent_size=256):\n\n  num_filters = 16\n\n  inputs = Input(shape=input_shape)\n  my_x = resnet_layer(inputs=inputs)\n\n  # Instantiate the stack of residual units\n  for stack in range(num_stacks):\n      for res_block in range(num_res_blocks):\n        my_x = residual_block(my_x, \\\n                                  num_filters, \\\n                                  ((stack &gt; 0) and (res_block == 0)))\n      num_filters *= 2\n\n  my_x = TimeDistributed(AveragePooling1D(pool_size=8))(my_x)\n  my_y = TimeDistributed(Flatten())(my_x)\n\n  # Add recurrent layers to keep some memory through time.\n  for i in range(num_recurrent):\n    # The last recurrent layer should just return the last output rather\n    # than the whole time sequence.\n    if (i == (num_recurrent - 1)):\n      my_y = LSTM(recurrent_size, return_sequences=False)(my_y)\n    else:\n      my_y = LSTM(recurrent_size, return_sequences=True)(my_y)\n\n  # Classify the output of the LSTM layers.\n  outputs = Dense(2, activation='softmax', kernel_initializer='he_normal')(my_y)\n\n  # Instantiate model.\n  model = Model(inputs=inputs, outputs=outputs)\n  return model\n\n# ==============================================================================\n#  Work with data\n# ==============================================================================\n\n# This mimics what I'm given in the data set (I can't change this).\norig_X_train = np.random.randn(N, snapshot_len, channels)\norig_X_test = np.random.randn(N, snapshot_len, channels)\norig_y_train = np.random.randint(0, 2, size=(N, classes))\norig_y_test = np.random.randint(0, 2, size=(N, classes))\n\n# -----&gt;  This next little bit seems to be the issue! &lt;-----\n# Re-shape inputs to create sequences of snapshots.\n# Shape before:  (N, snapshot_len, channels)\n# Shape after:   (N\/t, t, snapshot_len, channels), in this case try t = 4\nt = int(4)\nX_train = np.reshape(orig_X_train, (int(orig_X_train.shape[0] \/ t), \\\n                     t, orig_X_train.shape[1], orig_X_train.shape[2]))\nX_test = np.reshape(orig_X_test, (int(orig_X_test.shape[0] \/ t), t, \\\n                    orig_X_test.shape[1], orig_X_test.shape[2]))\n\n# Reshape labels to have one per sequence.\n# Shape before: (N, classes)\n# Shape after:  (N\/t, classes)\ny_train = np.zeros((int(orig_y_train.shape[0]\/t), orig_y_train.shape[1]))\ny_test = np.zeros((int(orig_y_test.shape[0]\/t), orig_y_test.shape[1]))\nnew_ind = 0\nfor i in range(len(orig_y_train)):\n  if (i % t == 0):\n    y_train[new_ind] = orig_y_train[i]\n    new_ind += 1\n\nnew_ind = 0\nfor i in range(len(orig_y_test)):\n  if (i % t == 0):\n    y_test[new_ind] = orig_y_test[i]\n    new_ind += 1\n\n# ==============================================================================\n#  Train the model\n# ==============================================================================\n\nconfig = tf.ConfigProto()\nconfig.gpu_options.per_process_gpu_memory_fraction = GPU_mem_fraction\nset_session(tf.Session(config=config))\n\nmodel = create_model(X_train.shape[1:])\n\ncheckpoint_cb = ModelCheckpoint('epoch_{epoch:02d}.h5', \\\n                                monitor='val_acc', \\\n                                verbose=1, \\\n                                save_best_only=True, \\\n                                save_weights_only=False, \\\n                                period=checkpoint_freq)\ncallbacks_list = [checkpoint_cb]\n\nhistory = model.fit(X_train, y_train, \\\n                    epochs=num_epochs, \\\n                    batch_size=batchsize, \\\n                    verbose=1, \\\n                    validation_data=(X_test, y_test), \\\n                    callbacks=callbacks_list)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57680025,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2019-08-27T18:00:50.213","AcceptedAnswerId":57680939.0,"OwnerUserId":1893354.0,"Title":"Using the TensorFlow Dataset api to import input\/output pairs of variable lengths","Body":"<p>I have a collection of input\/output pairs that I want to use as training examples for an RNN model.   So each input\/output is a list of the same length consisting of integers id's.  So 3 training examples might look like<\/p>\n\n<pre><code>[\n[[1,5,3,2],[22,5,3,8]],\n[[2,3],[4,7]],\n[[5,4,8,9,2,1,3],[32,4,7,4,5,21,33]]\n]\n<\/code><\/pre>\n\n<p>My question is how to load this kind of data into a TensorFlow Dataset such that it can be fed to the model downstream?<\/p>\n\n<p>I have experimented with the <code>tf.data.Dataset.from_tensor_slices()<\/code> method however it seems like all of the inputs and outputs must be of the same length in order to use this method.  However, as you can see from the above example, my inputs and outputs are of variable lengths.  <\/p>\n\n<p>Is the answer to pad the examples such that they are all the same length then use  <code>tf.data.Dataset.from_tensor_slices()<\/code>? If so is there a TensorFlow helper function that does this or should I do it manually?<\/p>\n","answers":[{"AnswerId":"57680939","CreationDate":"2019-08-27T19:14:14.380","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>Make use of generator and padded_batch concept of tf.data to overcome the variably sized input.<\/p>\n\n<pre><code>import tensorflow as tf\ntf.enable_eager_execution()\n\ndata = [\n[[1,5,3,2],[22,5,3,8]],\n[[2,3],[4,7]],\n[[5,4,8,9,2,1,3],[32,4,7,4,5,21,33]]\n]\n\ndata_in = [x for x, y in data]\ndata_out = [y for x, y in data]\n\ndef gen_series():\n    index_at = 0\n    while True:\n        yield data_in[index_at], data_out[index_at]\n        index_at += 1\n        if index_at &gt;= len(data):\n            index_at = 0\n\nds_series = tf.data.Dataset.from_generator(\n    gen_series,\n    output_types=(tf.int32, tf.int32),\n    output_shapes = ((None, None)))\n\nBATCH_SIZE = 2\nds_series_batch = ds_series.padded_batch(BATCH_SIZE, padded_shapes=([None], [None]))\n\nfor input_tensor, output_tensor in ds_series_batch.take(2):\n    print(input_tensor)\n    print(output_tensor)\n    print()\n<\/code><\/pre>\n\n<p>Padding will take place with maximum size of the input in the batch.<\/p>\n\n<p>Output:<\/p>\n\n<pre><code>tf.Tensor(\n[[1 5 3 2]\n [2 3 0 0]], shape=(2, 4), dtype=int32)\ntf.Tensor(\n[[22  5  3  8]\n [ 4  7  0  0]], shape=(2, 4), dtype=int32)\n\ntf.Tensor(\n[[5 4 8 9 2 1 3]\n [1 5 3 2 0 0 0]], shape=(2, 7), dtype=int32)\ntf.Tensor(\n[[32  4  7  4  5 21 33]\n [22  5  3  8  0  0  0]], shape=(2, 7), dtype=int32)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57680038,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-27T18:02:07.330","AcceptedAnswerId":null,"OwnerUserId":5809485.0,"Title":"Getting warning while importing keras with Tensorflow backend","Body":"<p>I have just installed Keras 2.2.5 and Tensorflow 1.14 on my Windows 10 machine. I am getting some warning as follows<\/p>\n\n<pre><code>WARNING: Logging before flag parsing goes to stderr.\nW0827 23:30:59.561793  9776 deprecation_wrapper.py:119] From C:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py:74: The name tf.get_default_graph is deprecated. Please use tf.compat.v1.get_default_graph instead.\n\nW0827 23:30:59.579783  9776 deprecation_wrapper.py:119] From C:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py:517: The name tf.placeholder is deprecated. Please use tf.compat.v1.placeholder instead.\n\nW0827 23:30:59.582780  9776 deprecation_wrapper.py:119] From C:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py:4138: The name tf.random_uniform is deprecated. Please use tf.random.uniform instead.\n<\/code><\/pre>\n\n<p>Although, this warning might not hurt my program but I found it annoying. Can you please help me to reduce this warning?<\/p>\n","answers":[{"AnswerId":"57680565","CreationDate":"2019-08-27T18:43:47.960","ParentId":null,"OwnerUserId":"10908375","Title":null,"Body":"<p>Tensorflow 1.14 is not the latest version, and most likely isn't updated anymore. That means it will have more and more dependency issues.<\/p>\n"}]}
+{"QuestionId":57680125,"AnswerCount":0,"Tags":"<python-3.x><debugging><keras><cluster-computing><centos7>","CreationDate":"2019-08-27T18:08:29.887","AcceptedAnswerId":null,"OwnerUserId":9526214.0,"Title":"A CNN while working for a while crashes with no error in Keras and on a Centos 7 Cluster","Body":"<p>I have written a Keras code with using the ResNet50 architecture which is available in Keras pretrained models. My training dataset is huge (70K 512*512 images). The network can be trained with no problem for a while (sometimes a week, sometimes for two weeks, and it can happen on different epochs); but it suddenly crashes with no error.\nThe code, dataset, and system remain intact while the code working.\nAlso, I am saving the weights for each epoch and printing the log of training for each batch which does not show any failure.<\/p>\n\n<p>I run everything on a reliable cluster and I have talked with our cluster admin; he mentioned there is no error recorded on the cluster side. Since this has happened a couple of time, he suggested me to contact you to check if you have seen such behavior from Keras models?\nIf not, can you please help us if there is any way in the Keras training procedure to check the followings:<\/p>\n\n<p>stack size\nworking set size\nnumber of open files\nthe name of each function or sub-routine as it is called -- to see if the program always exits from the same place\ninternal information about the variables &amp; data structures within your program\nI can print all these from python but I do not know how to do that while my network is being trained.<\/p>\n\n<p>Can anyone help with this?<\/p>\n\n<p>The code works and it suddenly crashes with no error, neither on the cluster side and on the code log or saved files.<\/p>\n\n<p>This is my system info:\nHave I written custom code (as opposed to using example directory): The code does training a Resnet using Keras pretrained model.\nOS Platform and Distribution (e.g., Linux Ubuntu 16.04): Centos (Cluster)\nTensorFlow backend (yes \/ no): Yes\nTensorFlow version: tensorflow-gpu==1.12.0\nKeras version: Keras==2.2.4\nPython version: 3.6\nCUDA\/cuDNN version: 9.0\nGPU model and memory: P100, 12GB<\/p>\n","answers":[]}
+{"QuestionId":57680701,"AnswerCount":0,"Tags":"<python><tensorflow><recurrent-neural-network><cudnn><tf.keras>","CreationDate":"2019-08-27T18:54:13.190","AcceptedAnswerId":null,"OwnerUserId":11756222.0,"Title":"Sentiment Analysis RNN not learning","Body":"<p>I am trying to create an RNN for sentiment analysis. My training data is the 'Sentiment140' dataset.<\/p>\n\n<p>The problem is that although the code runs, it is very slow (although this may be due to my computer) and the accuracy seems to remain stagnant at near 50%, which is more or less random.<\/p>\n\n<p>I am preprocessing the data using a 'word2vec' system where each word is represented as one-hot array. These arrays are then placed in another array which then wholly represents a single tweet. Each word array is also padded so they are all the same length.<\/p>\n\n<pre><code>class Tweet:\n\n    def __init__(self, line):\n        self.line = line\n\n    def Process(self):\n\n        #using file structure\n\n        # polarity = int(self.line.split(\",\")[0][1:])\n\n        polarity = int(self.line[0])\n\n        if polarity == 0:\n            polarity = 0\n        elif polarity == 2:\n            polarity = 1\n        elif polarity == 4:\n            polarity = 2\n\n        words = self.line[-1].lower()\n\n        #removes punctuation etc\n        words = clean(words)\n\n        TokenisedText = nltk.word_tokenize(words)\n\n        # lemmatised, tokenised text for entire line\n        TokenisedText = [WordNetLemmatizer().lemmatize(word) for word in TokenisedText]\n\n        WordVectors = np.zeros(shape=(MaxWordsLength, LexiconLength))\n\n        for i in range(0,len(TokenisedText)):\n            word = TokenisedText[i]\n\n            WordVector = np.zeros(LexiconLength)\n\n            index = np.where(lexicon == word)\n            WordVector[index] = 1\n            # same as += 1 here as max value for any one element is 1 - one hot array\n\n            # may be inefficient to append\n            WordVectors[i] = WordVector\n\n        return WordVectors, polarity\n\n# tweet structure [ [1,0,0,...], ... (incl. padding) ], ], [0,1,0,...], ... (incl. padding) ], ... ]\n# polarity - 0, 1 or 2\n\n#preprocesses the data and yields in batches\ndef GrabTrainingData():\n\n    with open(ShuffledTrainingPath, \"r\", buffering=200000, encoding=\"latin-1\") as InputFile:\n        i = 0\n        while True:\n            # BatchSize: how many tweets are you passing at one time?\n            # MaxWordsLength: how many word arrays in each tweet? ADDS PADDING TO ENSURE UNIFORM TENSOR SIZES\n\n            # DOES THIS MEAN CAN'T ANALYSE ANYTHING LONGER THAN MaxWordsLength?!?!\n\n            # LexiconLength: how long is each word array?\n            BatchTrain_x = np.zeros(shape=(BatchSize, MaxWordsLength, LexiconLength))\n            BatchTrain_y = np.zeros(BatchSize)\n\n            reader = csv.reader(InputFile)\n            for line in reader:\n                WordVectors, polarity = Tweet(line).Process()\n\n                BatchTrain_x[i] = WordVectors\n                BatchTrain_y[i] = polarity\n\n                i+=1\n\n                if i&gt;= BatchSize:\n                    i = 0\n                    yield BatchTrain_x, BatchTrain_y\n\n            # loops file again\n            i=0\n\n\n#neural network model\ndef RNN():\n    model = keras.models.Sequential()\n\n    NumberOfClassifications = 3\n\n    model.add(keras.layers.CuDNNLSTM(128, input_shape=(MaxWordsLength, LexiconLength), return_sequences=True))\n    model.add(keras.layers.Dropout(0.2))\n\n    model.add(keras.layers.CuDNNLSTM(128))\n    model.add(keras.layers.Dropout(0.2))\n\n    model.add(keras.layers.Dense(32, activation=\"relu\"))\n    model.add(keras.layers.Dropout(0.2))\n\n    model.add(keras.layers.Dense(NumberOfClassifications, activation=\"softmax\"))\n\n    MyOptimiser = keras.optimizers.Adam(lr=1e-3, decay=1e-5)\n\n    model.compile(loss=\"sparse_categorical_crossentropy\", optimizer = MyOptimiser, metrics=[\"accuracy\"] )\n\n    x_test, y_test = ProcessTestingData(ShuffledTestingPath)\n    model.fit_generator(generator=GrabTrainingData(), epochs=3, steps_per_epoch=math.ceil(TrainingNumberOfLines \/ BatchSize), validation_data=(x_test, y_test))\n\n    model.save(os.path.join(sys.path[0], \"MyRNN.h5\"))\n<\/code><\/pre>\n\n<p>P.S. As a side note, could someone also explain how I could correctly use use_multiprocessing in the data generator? I know I have to make use of keras.layers.Sequence() however I am not exactly sure how.<\/p>\n\n<p>Any help would be much appreciated.<\/p>\n","answers":[]}
+{"QuestionId":57680822,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-08-27T19:03:15.070","AcceptedAnswerId":57680958.0,"OwnerUserId":3288004.0,"Title":"Cannot predict single instance in Keras using a loop","Body":"<p>I'm playing a little with deep learning and Keras has been my choice due to its simplicity.<\/p>\n\n<p>I've built a simple multilayer perceptron model for binary classification and fitted it on input data (the same that I'm using for other ML models and which are working ok).<\/p>\n\n<p>The Following picture displays the Model summary:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vhFFU.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/vhFFU.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The first dense layer was defined as such:<\/p>\n\n<pre><code>model.add(Dense(18, input_dim=len(X_encoded.columns), activation = \"relu\", kernel_initializer=\"uniform\"))\n<\/code><\/pre>\n\n<p>When I attempt to predict over a loop like so:<\/p>\n\n<pre><code>for vals in X_encoded.values:\n    print(\"Survives?\", model.predict([vals], batch_size=1))\n<\/code><\/pre>\n\n<p>I get the following error:<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking input: expected dense_90_input to have shape (35,) but got array with shape (1,)<\/p>\n<\/blockquote>\n\n<p>These are my variable sizes:<\/p>\n\n<pre><code>print(\"Shape of vals:\", vals.shape, \"Number of Columns and First Layer Dimension:\", len(X_encoded.columns))\n<\/code><\/pre>\n\n<p>Result:<\/p>\n\n<blockquote>\n  <p>Shape of vals: (35,) Number of Columns and First Layer Dimension: 35<\/p>\n<\/blockquote>\n\n<p>As you can see, these match in size which is the expected input.<\/p>\n\n<p>What is going on? When I pass the entire dataframe \"predict\" it works correctly, but not when I pass a single value...<\/p>\n","answers":[{"AnswerId":"57680958","CreationDate":"2019-08-27T19:16:17.663","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>You need an array, not a list. You only use a list for multiple input tensors. <\/p>\n\n<pre><code>model.predict(np.array([vals]), batch_size=1)    \n<\/code><\/pre>\n\n<p>But why not:<\/p>\n\n<pre><code>model.predict(X_encoded.values, batch_size=1)    \n<\/code><\/pre>\n"}]}
+{"QuestionId":57680916,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-27T19:12:17.573","AcceptedAnswerId":57681530.0,"OwnerUserId":249341.0,"Title":"Fix part of a variable during eager execution training in tensorflow","Body":"<p>Is there a way to only update <em>some<\/em> of the variables during an eager execution update step? Consider this minimal working example:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\ntf.enable_eager_execution()\noptimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01)\n\nx = tf.Variable([1.0, 2.0])\n\ndef train(x):\n    with tf.GradientTape() as tape:\n        loss = x[0]**2 + x[1]**2 + 1\/(x[0]+x[1])\n        variables = [x]\n        grads = tape.gradient(loss, variables)\n        optimizer.apply_gradients(zip(grads, variables))\n\nfor _ in range(2000):\n    train(x)\n    print(x.numpy())\n<\/code><\/pre>\n\n<p>Which converges to <code>[0.5, 0.5]<\/code>. I'd like to fix the value of <code>x[0]<\/code> to it's initial value, while keeping everything else the way it is. What I've tried so far:<\/p>\n\n<ul>\n<li>Adding a <code>x[0].assign(1.0)<\/code> operation to the training step which grows the graph unnecessarily<\/li>\n<li>Changing <code>variables = [x[:-1]]<\/code> which gives <code>ValueError: No gradients provided for any variable: ['tf.Tensor([1.], shape=(1,), dtype=float32)']<\/code><\/li>\n<li>Adding <code>grads = [grads[0][1:]]<\/code> which gives <code>tensorflow.python.framework.errors_impl.InvalidArgumentError: var and delta do not have the same shape[2] [1] [Op:ResourceApplyGradientDescent]<\/code><\/li>\n<li>Doing both, which gives <code>TypeError: 'NoneType' object is not subscriptable<\/code><\/li>\n<\/ul>\n\n<p>For this MWE I can easily use two separate variables, but I'm interested in the generic case where I only want to update a known slice of an array.<\/p>\n","answers":[{"AnswerId":"57681530","CreationDate":"2019-08-27T20:03:19.997","ParentId":null,"OwnerUserId":"3987085","Title":null,"Body":"<p>You can set the gradient of the index you don't want to update to 0. In the code snippet bellow, the <code>mask<\/code> tensor indicates which elements we want to update (values <code>1<\/code>), and which elements we don't want to update (values <code>0<\/code>).<\/p>\n\n<pre><code>import tensorflow as tf\ntf.enable_eager_execution()\n\noptimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01)\n\nx = tf.Variable([1.0, 2.0])\nmask = tf.constant([0.0, 1.0])\n\ndef train(x):\n    with tf.GradientTape() as tape:\n        loss = x[0]**2 + x[1]**2 + 1\/(x[0]+x[1])\n        variables = [x]\n\n        grads = tape.gradient(loss, variables) * mask\n        optimizer.apply_gradients(zip(grads, variables))\n\nfor _ in range(100):\n    train(x)\n    print(x.numpy())\n<\/code><\/pre>\n\n<p>Another possible solution for your problem can be to stop the gradient on operations that <code>x[0]<\/code> is dependent on. For example:<\/p>\n\n<pre><code>optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01)\n\nx = tf.Variable([1.0, 2.0])\n\ndef train(x):\n    with tf.GradientTape() as tape:\n        loss = tf.stop_gradient(x[0])**2 + x[1]**2 + 1\/(tf.stop_gradient(x[0])+x[1])\n        variables = [x]\n\n        grads = tape.gradient(loss, variables)\n        optimizer.apply_gradients(zip(grads, variables))\n\nfor _ in range(100):\n    train(x)\n    print(x.numpy())\n<\/code><\/pre>\n"}]}
+{"QuestionId":57680948,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-08-27T19:15:16.407","AcceptedAnswerId":null,"OwnerUserId":915783.0,"Title":"Adam with learning rate decay vs ReduceLROnPlateau keras","Body":"<p>can anyone explain in detail how is Learning rate decay with Adam optimizer different\/similar from using <em>ReduceLROnPlateau<\/em>(Keras) with Adam optimizer.\nAlso, since Adam is variable learning rate algorithm, does it make sense to use \n<em>ReduceLROnPlateau<\/em> or any other rate variation technique with it?<\/p>\n\n<p>Thanks<br>\nsedy<\/p>\n","answers":[]}
+{"QuestionId":57681053,"AnswerCount":0,"Tags":"<python><keras><scikit-learn>","CreationDate":"2019-08-27T19:24:03.153","AcceptedAnswerId":null,"OwnerUserId":10007302.0,"Title":"How to access the Keras layers attribute from within sklearn's wrapper","Body":"<p>I'm trying to make use of the <code>layers<\/code> <code>get_weights()<\/code> from this post here:\n<a href=\"https:\/\/stackoverflow.com\/questions\/54009036\/how-to-extract-weights-from-input-layer-to-hidden-layer-and-from-hidden-layer\/54009163#54009163\">How to extract weights &quot;from input layer to hidden layer&quot; and &quot;from hidden layer to output layer&quot; with Keras API?<\/a><\/p>\n\n<p>However, I'm using <code>sklearn<\/code>'s <code>KerasRegressor<\/code> wrapper so I can make use of the <code>GridSearch<\/code>. I was wondering if there was any way to access the <code>get_weights<\/code> from within the <code>GridSearch<\/code> so I can track the weights used for each CV split. <\/p>\n\n<p>Here is my code. I tried using the final line in the code block but got the following error:<\/p>\n\n<blockquote>\n  <p>GridsearchCV object has no attribute 'layers'<\/p>\n<\/blockquote>\n\n<pre><code>X = data.iloc[:, 0:8]\ny = data.iloc[:, 8:9]\n\nX_train, X_test, y_train, y_test = train_test_split(X, y, test_size=.25, shuffle=False)\n\n\ntest_index = X_test.index.values.tolist()\n\nscalarX = MinMaxScaler()\nscalary = MinMaxScaler()\n\nscalarX.fit(X_train)\nscalary.fit(y_train)\n\nX_train, X_test = scalarX.transform(X_train), scalarX.transform(X_test)\ny_train, y_test = scalary.transform(y_train), scalary.transform(y_test)\n\ndef build_regressor(units1):\n    regressor = Sequential()\n    regressor.add(Dense(units=units1, input_dim=8, activation='relu'))\n    regressor.add(Dense(units=1))\n    regressor.compile(loss='mean_squared_error', optimizer='adam', \nmetrics=['mae', 'accuracy'])\n    return regressor\n\nregressor = KerasRegressor(build_fn=build_regressor)\n\nann_params = {'batch_size': [10, 25, 35],\n              'nb_epoch': [20, 50],\n              'units1': [8, 16, 32, 64]}\n\n\ngsc = GridSearchCV(regressor, param_grid=ann_params,\n      cv=TimeSeriesSplit(n_splits=5).split(X_train), verbose=10, n_jobs=-1, refit=True)\n\n\ngsc.fit(X_train, y_train)\nlayers = gsc.layers[1].get_weights() #doesn't work -returns error above\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57681345,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><deep-learning><artificial-intelligence>","CreationDate":"2019-08-27T19:48:04.700","AcceptedAnswerId":57681520.0,"OwnerUserId":10490382.0,"Title":"why tensorflow reshape array is out of range","Body":"<p>I have array reshape and sizes issue<\/p>\n\n<p>I haven't try anything due to the reason I am still new in this and I dont want to mess up things that are unreleated to the issue<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\n\nmnist = tf.keras.datasets.mnist\n(x_train, y_train),(x_test, y_test) = mnist.load_data()\n\nx_train = tf.keras.utils.normalize(x_train, axis=1)  # scales data between 0 and 1\nx_test = tf.keras.utils.normalize(x_test, axis=1) \n\nmodel = tf.keras.models.Sequential()\nmodel.add(tf.keras.layers.Flatten(input_shape=(32,)))\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\nx_train = np.reshape(x_train, (x_train.shape[0], 1, x_train.shape[1]))\nx_test = np.reshape(x_test, (x_test.shape[0], 1, x_test.shape[1]))\n\nmodel.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy', \n              metrics=['accuracy'])\n\nmodel.fit(x_train, y_train, epochs=3)\n\nval_loss, val_acc = model.evaluate(x_test, y_test)\nprint(val_loss)\nprint(val_acc)\n<\/code><\/pre>\n\n<pre><code>  File \"t1.py\", line 17, in &lt;module&gt;\n    x_train = np.reshape(x_train, (x_train.shape[0], 1, x_train.shape[1]))\n  File \"&lt;__array_function__ internals&gt;\", line 6, in reshape\n  File \"H:\\Program Files\\Python36\\lib\\site-packages\\numpy\\core\\fromnumeric.py\", line 301, in reshape\n    return _wrapfunc(a, 'reshape', newshape, order=order)\n  File \"H:\\Program Files\\Python36\\lib\\site-packages\\numpy\\core\\fromnumeric.py\", line 61, in _wrapfunc\n    return bound(*args, **kwds)\nValueError: cannot reshape array of size 47040000 into shape (60000,1,28)```\n<\/code><\/pre>\n","answers":[{"AnswerId":"57681520","CreationDate":"2019-08-27T20:02:42.337","ParentId":null,"OwnerUserId":"10490382","Title":null,"Body":"<p><code>model.add(tf.keras.layers.Flatten(input_shape=(28,28)))<\/code><\/p>\n\n<p>it is an 28x28 image not a 32 vector\nsoo there we know it is not should be a vector of 32\nby lefting an argument<\/p>\n"}]}
+{"QuestionId":57681458,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-27T19:58:15.073","AcceptedAnswerId":null,"OwnerUserId":11876924.0,"Title":"tf.Print() output reduction","Body":"<p>I am using <code>tf.Print()<\/code> to debug a model, but <code>tf.Print()<\/code> is printing too often. What are some good strategies to limit the output of this function, perhaps to be invoked on every nth epoch?<\/p>\n","answers":[{"AnswerId":"57681807","CreationDate":"2019-08-27T20:26:07.417","ParentId":null,"OwnerUserId":"6180759","Title":null,"Body":"<p>You can print output on every nth epoch as you mentioned in the question. Here is the basic example:<\/p>\n\n<pre><code>if epoch % n == 0:\n    ...\n<\/code><\/pre>\n"},{"AnswerId":"57681996","CreationDate":"2019-08-27T20:42:06.160","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>You could hold a counter for the steps\/epochs in a TF variable, and then use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/cond\" rel=\"nofollow noreferrer\">tf.cond<\/a> in combination with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/math\/floormod\" rel=\"nofollow noreferrer\">tf.math.mod<\/a> to print the tensor when the counter is a multiple of <code>n<\/code>.<\/p>\n\n<p>Here is an example:<\/p>\n\n<pre><code>import tensorflow as tf\n\n\ndef print_step(step):\n    print_op = tf.print(step)\n    with tf.control_dependencies([print_op]):\n        out = tf.identity(step)\n    return out\n\n\nstep = tf.Variable(0)\nprint_freq = 100\n\nprint_flag = tf.equal(tf.math.mod(step, print_freq), 0)\nupdate_step = tf.assign(step, step + 1)\ncond_print_op = tf.cond(print_flag,\n                        lambda: print_step(update_step),\n                        lambda: update_step)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    for i in range(1000):\n        sess.run(cond_print_op)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57681850,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2019-08-27T20:30:30.070","AcceptedAnswerId":57682030.0,"OwnerUserId":11309950.0,"Title":"Is there a way to reduce execution time while reducing batch_size?","Body":"<p>Here is my code<\/p>\n\n<pre><code>for i in range(10):\n    model = Sequential()\n    model.add(CuDNNLSTM(1))\n    model.add(Dense(1))\n\n    model.compile(loss='mean_squared_error', optimizer=optimizers.adam())\n\n    model.fit(x_train, y_train, epochs=20, batch_size=16, verbose=0)\n\n    p = model.predict(x_test)\n    mse = mean_squared_error(y_test, p)\n    if mse &lt; minval:\n        model.save('best.h5')\n<\/code><\/pre>\n\n<p>If I increase <code>batch_size<\/code>, MSE goes up as well. Therefore, to reduce MSE, I must reduce <code>batch_size<\/code>. <\/p>\n\n<p>The problem with increasing <code>batch_size<\/code> is time. As <code>batch_size<\/code> decreases, the execution time increases significantly.<\/p>\n\n<p>I am wondering how to increase execution speed by reducing <code>batch_size<\/code> without changing hardware. <\/p>\n\n<p>Or how can I reduce MSE while keeping <code>batch_size<\/code> the same? In this code, <code>batch_size<\/code> has the biggest impact on MSE.<\/p>\n\n<p>for one\u00b4s information, computer resources remain. I want to save time by taking advantage of all the remaining resources, but I'm curious what I can do if I can. (Not half of the GPU memory is used, and the remaining memory is idle. The rest of the resources are the same.)<\/p>\n","answers":[{"AnswerId":"57682030","CreationDate":"2019-08-27T20:45:29.693","ParentId":null,"OwnerUserId":"9000214","Title":null,"Body":"<p>If you decrease the batch size, then the time is going to increase, and this is basically unavoidable without tuning other parameters. <\/p>\n\n<p>Please note that while decreasing the batch size to an extent reduce your MSE, reducing it to an absurdly low value is going to raise it, so there's a balance you need to keep. <\/p>\n\n<p>I do note that you are currently not manually tuning your learning rate, and increasing the learning rate could not only speed up your results but help prevent falling into a local minima. Of course, a learning rate that is also too high might make it less likely for you to reach the global minimum as well.<\/p>\n\n<p>You can play around with your parameters though overall it's gonna depend on your computer. For example, if you have an i7 instead of an i5, you have hyperthreading which could speed up the process. <\/p>\n"}]}
+{"QuestionId":57681910,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras><conv-neural-network>","CreationDate":"2019-08-27T20:35:40.890","AcceptedAnswerId":null,"OwnerUserId":11971761.0,"Title":"unable to execute plot_model in keras","Body":"<p>i'm trying to visualise my model, but when i'm using plot_model function of keras it's giving me error saying \"'InputLayer' object is not iterable\"\ni'm attaching my code as well as error. please help<\/p>\n\n<pre><code>model = tf.keras.models.Sequential([\n    tf.keras.layers.Conv2D(96, (5, 5), activation='relu', input_shape=(28, 28, 3), padding = 'same'),\n    tf.keras.layers.Conv2D(96, (5, 5), activation='relu', padding = 'same'),\n    tf.keras.layers.MaxPooling2D(2, 2),\n    tf.keras.layers.Conv2D(256, (5, 5), activation='relu', padding = 'same'),\n    tf.keras.layers.Conv2D(256, (5, 5), activation='relu', padding = 'same'),\n    tf.keras.layers.MaxPooling2D(2, 2),\n    tf.keras.layers.Conv2D(384, (3, 3), activation='relu', padding = 'same'),\n    tf.keras.layers.Conv2D(384, (3, 3), activation='relu', padding = 'same'),\n    tf.keras.layers.MaxPooling2D(2, 2),\n    tf.keras.layers.Conv2D(256, (3, 3), activation='relu', padding = 'same'),\n    tf.keras.layers.Conv2D(256, (3, 3), activation='relu', padding = 'same'),\n    tf.keras.layers.Flatten(),\n    tf.keras.layers.Dense(2304, activation='relu'),\n    tf.keras.layers.Dense(2304, activation='relu'),\n    tf.keras.layers.Dense(10, activation=tf.nn.softmax)\n])\n\nmodel.compile(optimizer=Adam(lr=0.001), loss='sparse_categorical_crossentropy', metrics=['acc'])\n\nplot_model(model, to_file='model_plot.png', show_shapes=True, show_layer_names=True)\n<\/code><\/pre>\n\n<pre><code>---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-92-2aa57a1383be&gt; in &lt;module&gt;()\n----&gt; 1 plot_model(model, to_file='model_plot.png', show_shapes=True, show_layer_names=True)\n\n1 frames\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    130             'LR' creates a horizontal plot.\n    131     \"\"\"\n--&gt; 132     dot = model_to_dot(model, show_shapes, show_layer_names, rankdir)\n    133     _, extension = os.path.splitext(to_file)\n    134     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    107             node_key = layer.name + '_ib-' + str(i)\n    108             if node_key in model._network_nodes:\n--&gt; 109                 for inbound_layer in node.inbound_layers:\n    110                     inbound_layer_id = str(id(inbound_layer))\n    111                     dot.add_edge(pydot.Edge(inbound_layer_id, layer_id))\n\nTypeError: 'InputLayer' object is not iterable\n<\/code><\/pre>\n","answers":[{"AnswerId":"57749946","CreationDate":"2019-09-01T22:37:40.293","ParentId":null,"OwnerUserId":"1673784","Title":null,"Body":"<p>Try importing from tensorflow directly:<\/p>\n\n<p><code>from tensorflow.keras.utils import plot_model<\/code><\/p>\n"},{"AnswerId":"57682510","CreationDate":"2019-08-27T21:30:53.380","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You are mixing the usage\/imports of the <code>keras<\/code> and <code>tf.keras<\/code> packages, these packages are not compatible with each other, you must make all relevant imports from one package only.<\/p>\n"}]}
+{"QuestionId":57682182,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-08-27T20:57:58.457","AcceptedAnswerId":null,"OwnerUserId":915783.0,"Title":"how to use keras min_delta and cooldown in ReduceLROnPlateau","Body":"<p>Lets say we set   <strong>ReduceLROnPlateau<\/strong> with starting learning rate of <em>lr1<\/em> as follows:    <\/p>\n\n<p>ReduceLROnPlateau(monitor='val_acc', factor=0.1, patience=10, verbose=0, mode='auto', min_delta=0.1, cooldown=0, min_lr=1e-8)<\/p>\n\n<p>is this how algorithm proceeds?  <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/NjZQt.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/NjZQt.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>if yes, then for validation accuracy as metric, as we lower the min_delta(finer), we\nshould get more frequent learning rate changes  and vice versa?<\/p>\n\n<p>Also, how does the above change with lets say a cool down of 2?(is it useful to have non zero cooldown?)<\/p>\n","answers":[]}
+{"QuestionId":57682485,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2019-08-27T21:28:47.140","AcceptedAnswerId":57810004.0,"OwnerUserId":3878186.0,"Title":"Tensorflow classifier.evaluate running indefinitely?","Body":"<p>I've started trying out some of the Tensorflow API's. I am using the iris data set to experiment with Tensorflows Estimator's. I'm loosely following this tutorial except that I load my data in a little differently: <a href=\"https:\/\/www.tensorflow.org\/guide\/premade_estimators#top_of_page\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/premade_estimators#top_of_page<\/a><\/p>\n\n<p>My problem is that when the code below executes and I get to the section with:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>   # Evaluate the model.\n   eval_result = classifier.evaluate(\n<\/code><\/pre>\n\n<p>My computer just runs seemingly without end. I've been waiting for my jupyter notebook to complete this step now for an hour and a half but no end in sight. The lastoutput of the notebook is:<\/p>\n\n<p>INFO:tensorflow:Running local_init_op.<\/p>\n\n<p>INFO:tensorflow:Done running local_init_op.<\/p>\n\n<p>Problem statement: How can I adjust my code to make this evaluation more efficient I'm obviously making it do much more work than I anticipated.<\/p>\n\n<p>So far I have tried adjusting the batch size and the number or neurons in the layers but with no luck.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#First we want to import what we need. Typically this will be some combination of:\nimport tensorflow as tf\nimport pandas as pd\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.preprocessing import LabelEncoder \n\n%matplotlib inline\n\n#Extract the data from the iris dataset.\ndf = pd.read_csv('IRIS.csv')\nle = LabelEncoder()\ndf['species'] = le.fit_transform(df['species'])\n\n#Extract both into features and labels.\n#features should be a dictionary.\n#label can just be an array\ndef extract_features_and_labels(dataframe):\n\n    #features and label for training\n    x = dataframe.copy()\n    y = dataframe.pop('species')\n    return dict(x), y\n\n#break the data up into train and test.\n#split the overall df into training and testing data\ntrain, test = train_test_split(df, test_size=0.2)\n\ntrain_x, train_y = extract_features_and_labels(train)\ntest_x, test_y = extract_features_and_labels(test)\n\nprint(len(train_x), 'training examples')\nprint(len(train_y), 'testing examples')\n\nmy_feature_columns = []\nfor key in train_x.keys():\n    my_feature_columns.append(tf.feature_column.numeric_column(key=key))\n\ndef train_input_fn(features, labels, batch_size):\n    \"\"\"An input function for training\"\"\"\n    # Convert the inputs to a Dataset.\n    dataset = tf.data.Dataset.from_tensor_slices((dict(features), labels))\n\n    # Shuffle, repeat, and batch the examples.\n    return dataset.shuffle(1000).repeat().batch(batch_size)\n\n#Build the classifier!!!!\nclassifier = tf.estimator.DNNClassifier(\n    feature_columns=my_feature_columns,\n    # Two hidden layers of 10 nodes each.\n    hidden_units=[4, 4],\n    # The model must choose between 3 classes.\n    n_classes=3)\n\n# Train the Model.\nclassifier.train(\n    input_fn=lambda:train_input_fn(train_x, train_y, 10), steps=1000)\n\n# Evaluate the model.\neval_result = classifier.evaluate(\n    input_fn=lambda:train_input_fn(test_x, test_y, 100))\n\nprint('\\nTest set accuracy: {accuracy:0.3f}\\n'.format(**eval_result))\n<\/code><\/pre>\n","answers":[{"AnswerId":"57810004","CreationDate":"2019-09-05T16:58:28.050","ParentId":null,"OwnerUserId":"3878186","Title":null,"Body":"<p>Had to switch a lot of things up but finally got the estimator working on the IRIS data set. Here is the code below for any who may find it useful in the future. Cheers.<\/p>\n\n<pre><code>#First we want to import what we need. Typically this will be some combination of:\nimport tensorflow as tf\nimport pandas as pd\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom sklearn.model_selection import train_test_split\n\nfrom sklearn.preprocessing import LabelEncoder  \nfrom sklearn import preprocessing\n\n%matplotlib inline\n\n#Extract the data from the iris dataset.\ndf = pd.read_csv('IRIS.csv')\n\n#Grab only our categorial data\n#categories = df.select_dtypes(include=[object])\nle = LabelEncoder()\ndf['species'] = le.fit_transform(df['species'])\n# use df.apply() to apply le.fit_transform to all columns\n#X_2 = X.apply(le.fit_transform)\n\n#Reshape as the one hot encoder doesnt like one row\/column\n#X_2 = X.reshape(-1, 1)\n\n#features is everything but our label so makes sense to simply....\nfeatures = df.drop('species', axis=1)\nprint(features.head())\ntarget = df['species']\nprint(target.head())\n\n#trains and test\nX_train, X_test, y_train, y_test = train_test_split(\n     features, target, test_size=0.33, random_state=42)\n\n#Introduce Tensorflow feature column (numeric column)\nnumeric_column = ['sepal_length','sepal_width','petal_length','petal_width']\n\nnumeric_features = [tf.feature_column.numeric_column(key = column)\n                   for column in numeric_column]\nprint(numeric_features[0])\n\n#Build the input function for training\ntraining_input_fn = tf.estimator.inputs.pandas_input_fn(x = X_train,\n                                                       y=y_train,\n                                                       batch_size=10,\n                                                       shuffle=True,\n                                                       num_epochs=None)\n\n#Build the input function for testing input\neval_input_fn = tf.estimator.inputs.pandas_input_fn(x=X_test,\n                                                   y=y_test,\n                                                   batch_size=10,\n                                                   shuffle=False,\n                                                   num_epochs=1)\n\n#Instansiate the model\ndnn_classifier = tf.estimator.DNNClassifier(feature_columns=numeric_features,\n                                              hidden_units=[3,3],\n                                              optimizer=tf.train.AdamOptimizer(1e-4),\n                                              n_classes=3,\n                                              dropout=0.1,\n                                              model_dir = \"dnn_classifier\")\n\ndnn_classifier.train(input_fn = training_input_fn,steps=2000)\n\n#Evaluate the trained model\ndnn_classifier.evaluate(input_fn = eval_input_fn)\n# print(\"Loss is \" + str(loss))\n\npred = list(dnn_classifier.predict(input_fn = eval_input_fn))\n\nfor e in pred:\n    print(e)\n    print(\"\\n\")\n#pred = [p['species'][0] for p in pred]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57682831,"AnswerCount":1,"Tags":"<python><tensorflow><keras><autoencoder>","CreationDate":"2019-08-27T22:09:16.737","AcceptedAnswerId":57688001.0,"OwnerUserId":7154924.0,"Title":"What happens if Keras Tensorflow's `Model.fit`'s target\/output is `None`?","Body":"<p>I was looking at this code: <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/variational_autoencoder.py#L198\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/variational_autoencoder.py#L198<\/a>, where <code>Model.fit()<\/code> is called without an output or target Tensor. At first, I thought the behavior of <code>Model.fit()<\/code> is to use the input as the output (which would make sense for this autoencoder implementation). But then I looked into the documentation, and that's not what it says: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/Model#fit\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/Model#fit<\/a><\/p>\n\n<p>It implies when <code>y<\/code>, the target, is <code>None<\/code>, <code>x<\/code> should be some kind of structure that contains both the input and target. <\/p>\n\n<p>But it's clear in this autoencoder implementation, that's not what happens (<code>x<\/code> only contains the input). Could someone explain what happens in this case?<\/p>\n","answers":[{"AnswerId":"57688001","CreationDate":"2019-08-28T08:25:07.327","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>In the Keras documentation for <a href=\"https:\/\/keras.io\/models\/model\/#fit\" rel=\"nofollow noreferrer\">model.fit<\/a> the following is stated:<\/p>\n\n<blockquote>\n  <ul>\n  <li>y: Numpy array of target (label) data (if the model has a single output), or list of Numpy arrays (if the model has multiple outputs).\n  If output layers in the model are named, you can also pass a\n  dictionary mapping output names to Numpy arrays.  <strong>y can be None\n  (default) if feeding from framework-native tensors (e.g. TensorFlow\n  data tensors).<\/strong><\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>Now, notice that in the variational autoencoder example, the argument <code>outputs<\/code> of the model <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/variational_autoencoder.py#L161\" rel=\"nofollow noreferrer\"><code>vae<\/code><\/a> is a TensorFlow native tensor, since it is given by the output of another model <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/variational_autoencoder.py#L155\" rel=\"nofollow noreferrer\"><code>decoder<\/code><\/a> (a TensorFlow native tensor) whose <code>inputs<\/code> argument is <em>independent<\/em> of the <code>vae<\/code>'s input.<\/p>\n"}]}
+{"QuestionId":57683209,"AnswerCount":0,"Tags":"<python><tensorflow><keras><autoencoder>","CreationDate":"2019-08-27T23:10:22.187","AcceptedAnswerId":null,"OwnerUserId":7154924.0,"Title":"Tensor fails to print in Keras Tensorflow autoencoder implementation","Body":"<p>This is my second question about Keras autoencoder implementation in the official repo: <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/variational_autoencoder.py#L142\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/variational_autoencoder.py#L142<\/a><\/p>\n\n<p>The problem I am having is that I can't seem to print any Tensors after Line 142. Namely, I inserted <code>z = K.print_tensor(z, \"zzz\")<\/code> after Line 142, and I don't see any output. Similar print statements also do not work for decoder Tensors. <\/p>\n\n<p>Anyone know what's going on?<\/p>\n","answers":[]}
+{"QuestionId":57683482,"AnswerCount":1,"Tags":"<python><tensorflow><keras><callback>","CreationDate":"2019-08-28T00:01:56.750","AcceptedAnswerId":57693784.0,"OwnerUserId":11986406.0,"Title":"Is it possible to log the epoch results in the tf.keras model using a tensorflow callback, in order to save at the end of training?","Body":"<p>I am trying to save the <code>logs<\/code> value that gets passed into a tf.keras callback at the end of every epoch to track how the model did along the way. <\/p>\n\n<p>I have tried writing a custom tf.keras callback that appends the <code>logs<\/code> value that is passed to the callback to an array that I initialized at the beginning of training using the same callback. However upon debugging after the first epoch the array that I initialized in the model is <code>None<\/code> at the end of the second epoch. Below is the custom callback I created. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class LogEpochScores(tf.keras.callbacks.Callback):\n    def __init__(self):\n        super(LogEpochScores, self).__init__()\n\n    def on_train_begin(self, logs=None):\n        self.model.epoch_log = []\n\n    def on_epoch_end(self, epoch, logs=None):\n        self.model.epoch_log = self.model.epoch_log.append(logs)\n<\/code><\/pre>\n\n<p>I expected the following array at the end of training <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>[\n{loss: 1, acc:1, val_loss:1, val_acc:1},\n{loss: 2, acc:2, val_loss:2, val_acc:2},\n{loss: 3, acc:1, val_loss:3, val_acc:3},\n{loss: 4, acc:1, val_loss:4, val_acc:4},\n{loss: 5, acc:1, val_loss:5, val_acc:5},\n{loss: 6, acc:1, val_loss:6, val_acc:6}\n]\n<\/code><\/pre>\n\n<p>Each entry being the results from the respective epoch.<\/p>\n\n<p>The output actually is <code>None<\/code><\/p>\n\n<p>Edit: Formatting<\/p>\n","answers":[{"AnswerId":"57693784","CreationDate":"2019-08-28T13:37:31.057","ParentId":null,"OwnerUserId":"11986406","Title":null,"Body":"<p>Apparently I am not intelligent. I was using append the wrong way. Its destructive so you don't set it as a variable after use. The corrected code is as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class LogEpochScores(tf.keras.callbacks.Callback):\n    def __init__(self):\n        super(LogEpochScores, self).__init__()\n\n    def on_train_begin(self, logs=None):\n        self.model.epoch_log = []\n\n    def on_epoch_end(self, epoch, logs=None):\n        self.model.epoch_log.append(logs)\n<\/code><\/pre>\n\n<p>Notice in <code>on_epoch_end<\/code> I change <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>self.model.epoch_log = self.model.epoch_log.append(logs)\n<\/code><\/pre>\n\n<p>to <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>self.model.epoch_log.append(logs)\n<\/code><\/pre>\n\n<p>Additionally, in the tf.kera model object, there is an attribute called history which is exactly what I needed. So in the end there was no need for this callback<\/p>\n"}]}
+{"QuestionId":57683984,"AnswerCount":1,"Tags":"<python><ios><tensorflow><machine-learning><coreml>","CreationDate":"2019-08-28T01:31:44.543","AcceptedAnswerId":57684015.0,"OwnerUserId":8141402.0,"Title":"How to train an intelligent chatbot using a database to then add to an iOS app","Body":"<p>I want to create an intelligent chatbot-type app for iOS that uses the <a href=\"https:\/\/www.cs.cornell.edu\/~cristian\/Cornell_Movie-Dialogs_Corpus.html\" rel=\"nofollow noreferrer\">Cornell movie database<\/a> and some custom training data I create to train a machine learning model that can talk to the user, preferably keeping track of the conversation and performing a few simple tasks.<\/p>\n\n<p>I'll go more in depth.<\/p>\n\n<p>Recently I came across a way to <a href=\"https:\/\/medium.com\/towards-artificial-intelligence\/tensorflow-to-coreml-conversion-and-model-inspection-8a30123b1289\" rel=\"nofollow noreferrer\">convert TensorFlow models to CoreML models<\/a> so I tried working with TensorFlow and realized that the conversion method is limited and working with TensorFlow would require me to learn a lot more than I already knew and I simply do not have the time to do that. I am an iOS developer and am therefore familiar with Swift and Objective-C, so I was wondering if there was a way to do the same thing using them.<\/p>\n\n<p>I would like to be able to train a model with the <a href=\"https:\/\/www.cs.cornell.edu\/~cristian\/Cornell_Movie-Dialogs_Corpus.html\" rel=\"nofollow noreferrer\">Cornell movie database<\/a> I tried using when training TensorFlow, plus whatever contributions I make to it. What I mean by that is I might want my chatbot to do some specific tasks or reply to certain abnormal questions in a specific way. I think I could do this by adding my own \"conversations\" to the Cornell database that contain my specific needs. If there is some other way I could inject my own data into this, please let me know.<\/p>\n\n<p>I also want the chatbot to be able to do specific tasks. Therefore I believe I need to have a more complicated model that can recognize when the chatbot is told to do something such as turn on a light here or send a text to someone (very small things that would just need a trigger and some parameters). <\/p>\n\n<p>It would also be great to have the model remember certain aspects of the conversation. For instance, if the chatbot asks for my name, and I reply, it should remember what my name is later if I ask it what my name is. This would also apply to more complicated things such as a scenario where I ask it what the weather is in Toronto tomorrow, it replies, and then I ask \"What about next Tuesday?\", and it replies with the weather in Toronto for next Tuesday.<\/p>\n\n<p>Overall, I would like to use CoreML (and possibly its seemingly very good Natural Language processing capabilities) to create a complex model for a chatbot for iOS applications.<\/p>\n\n<p>I don't know if this is at all possible but I would really like to try to get the most advanced and conversational chatbot for iOS as possible.<\/p>\n","answers":[{"AnswerId":"57684015","CreationDate":"2019-08-28T01:36:43.003","ParentId":null,"OwnerUserId":"9000214","Title":null,"Body":"<p>If you\u2019re fine with making calls to an API (which would make it useless without WiFi) and paying for it, Google\u2019s Dialogue Flow might be a good tool for you, as it already has great NLP. You\u2019d still have to do a lot of the information storing by yourself, and write your own functionality for other parts (like actually sending requests to a weather API) but it\u2019s very good for understanding the intent behind someone\u2019s speech.<\/p>\n\n<p>If you don\u2019t want to pay or need it to work without WiFi, There is a 40+ hour long course called Natural Language Processing A-Z on Udemy which is very useful for learning NLP with tensorflow. Takes more work but google doesn\u2019t take a cut and you might be able to work without WiFi.<\/p>\n"}]}
+{"QuestionId":57684172,"AnswerCount":2,"Tags":"<tensorflow><machine-learning><scikit-learn><linear-regression>","CreationDate":"2019-08-28T02:01:58.423","AcceptedAnswerId":57687978.0,"OwnerUserId":11970084.0,"Title":"What is the advantage of using tensorflow instead of scikit-learn for doing regression?","Body":"<p>I am new to machine learning and I want to start doing basic regression analysis. I saw that scikit-learn provides a simple way to do this. But  why people use tensorflow for regression instead? Thanks!<\/p>\n","answers":[{"AnswerId":"57687978","CreationDate":"2019-08-28T08:23:30.090","ParentId":null,"OwnerUserId":"9451306","Title":null,"Body":"<p>If the only thing you are doing is regression, scikit-learn is good enough and will definitely do you job. Tensorflow is more a deep learning framework for building deep neural networks.<\/p>\n\n<p>There're people using Tensorflow to do regression maybe just out of personal interests or they think Tensorflow is more famous or \"advanced\".<\/p>\n"},{"AnswerId":"57687656","CreationDate":"2019-08-28T08:04:42.670","ParentId":null,"OwnerUserId":"2672299","Title":null,"Body":"<p>Tensorflow is a deep learning framework and involves far more complex decisions concerning algorithm design. <\/p>\n\n<p>In the first step, it is recommended to use sklearn, because you will get a first ml model with scikit-learn faster. Later you can use a dl model with tensorflow. :-)<\/p>\n"}]}
+{"QuestionId":57684516,"AnswerCount":0,"Tags":"<java><android><tensorflow>","CreationDate":"2019-08-28T03:05:48.100","AcceptedAnswerId":null,"OwnerUserId":9780088.0,"Title":"MAPPEDBYTEBUFFER IS NOT VALID FLATBUFFER MODEL when loading tflite model file into project","Body":"<p>The whole android project origins from tensorflow official's example called Camera2BasicFragment.\nWe edit it to meet our needs which is detect the people body motion.\nHowever , when we tried to load the tflite file under the Asset folder, it arose some question.<\/p>\n\n<blockquote>\n  <p>Have I written custom code : No <\/p>\n  \n  <p>OS Platform and Distribution (e.g., Linux Ubuntu 16.04): OSX 10.12.3<\/p>\n  \n  <p>Mobile device: ASUS_X00ID<\/p>\n  \n  <p>TensorFlow installed from (source or binary):<\/p>\n  \n  <p>TensorFlow version (use command below): 1.14.0<\/p>\n  \n  <p>Python version: 3.6.9<\/p>\n  \n  <p>Bazel version (if compiling from source): None<\/p>\n  \n  <p>GCC\/Compiler version (if compiling from source):None<\/p>\n  \n  <p>CUDA\/cuDNN version: None<\/p>\n  \n  <p>GPU model and memory: None<\/p>\n<\/blockquote>\n\n<p>this is the code we try to load the model\nthe error arrise at \nthis line\nwhen we trying to put it on our phone it throw into the catch and show the error message \"JAVA.LANG.ILLEGALARGUMENTEXCEPTION:MAPPEDBYTEBUFFER IS NOT A VALID FLATBUFFER MODEL\"\nWe assume it has something to do with the \"MAPPREDBYTEBUFFER\",but the problem is we did create a pretty simple project to test the load model process and showed no error.<\/p>\n\n<pre><code>private void trymodel(){\n        String modelFile=\"LSTM_1.tflite\";\n        Interpreter tflite = null;\n        try {\n            tflite=new Interpreter(loadModelFile(CameraActivity.this));\n        } catch (Exception e) {\n            testmodel_but.setText(String.valueOf(e));\n        }\n\n        float[][][] inputX;\n        inputX = new float[1][1][22];\n        \/\/random the array\n        Random random=new Random();\n        for( int i = 0 ; i &lt; inputX.length ; i++ ) {\n            for ( int j = 0 ; j &lt; inputX[i].length ; j++ ) {\n                for (int k=0;k &lt; inputX[j].length ; k++){\n                    inputX[i][j][k]= random.nextInt(101);\n                }\n            }\n        }\n<\/code><\/pre>\n\n<p>the function we loadModelFile<\/p>\n\n<pre><code> private MappedByteBuffer loadModelFile(Activity activity) throws Exception {\n            AssetFileDescriptor fileDescriptor = activity.getAssets().openFd(\"LSTM_1.tflite\");\n            FileInputStream inputStream = new FileInputStream(fileDescriptor.getFileDescriptor());\n            FileChannel fileChannel = inputStream.getChannel();\n            long startOffset = fileDescriptor.getStartOffset();\n            long declaredLength = fileDescriptor.getDeclaredLength();\n            return fileChannel.map(FileChannel.MapMode.READ_ONLY, startOffset, declaredLength);\n\n    }\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57684638,"AnswerCount":1,"Tags":"<tensorflow><distributed-computing>","CreationDate":"2019-08-28T03:26:48.250","AcceptedAnswerId":null,"OwnerUserId":4515762.0,"Title":"Is there a way to use keras flow_from_directory with distributed training (mirrored strategy)?","Body":"<p>I cannot find any documentation or example code on how to do dynamic file loading with distribute strategies in tensorflow. I see many examples using keras flow_from_directory which seems to be very nice because it relies on generator and does scaling etc on the fly so doesn't require loading hundreds of thousands of images at once. But the distribute strategy examples (mirrored strategy etc) all show pre-loading the entire dataset and then using dataset.from_tensor_slices() and loading that to strategy.experimental_distribute_dataset(). This is not feasable with large data that exceeds the memory. I've tried to combine the above dynamic loading method from keras with distributed batching but it looks like there is no trivial way to convert some flow_from_directory output to be compatible with what strategy.experimental_distribute_dataset() expects so these two features are probably not compatible with one another. <\/p>\n\n<p>Is there some other way to do this with distributed training? <\/p>\n\n<p>I could probably hand-code dynamic file loading during training but it would be pretty slow and basic compared to what generators and flow offer. I would be very surprised if TF2 didn't include this functionality for distributed training yet.<\/p>\n","answers":[{"AnswerId":"58607264","CreationDate":"2019-10-29T12:13:02.347","ParentId":null,"OwnerUserId":"11270789","Title":null,"Body":"<p>dataset.from_tensor_slices() does not load the whole data to your memory at once.\nit may look that way, but when configured the right way, it uses batches, prefetching, caching etc. <\/p>\n\n<p>look at my small example of how I am using the tf.data pipeline. the map function gets executed during the training or prefetching and not in the beginning. so not all the data gets loaded to your memory at the same time.  <\/p>\n\n<pre><code>def parse_img(label, path):\n    img = tf.io.read_file(path)\n    img = tf.image.decode_png(img, channels=0, dtype=tf.uint8)\n    img = tf.image.convert_image_dtype(img, tf.float32)\n    return img, label\n\nds_train = tf.data.Dataset.from_tensor_slices((list_labels_train,\n                                               list_paths_train))\n\nds_train = ds_train.map(parse_img, num_parallel_calls=AUTOTUNE)\nds_train = ds_train.cache()\nds_train = ds_train.shuffle(len(list_paths_train), seed=42,\n                            reshuffle_each_iteration=True)\nds_train = ds_train.batch(BATCH_SIZE)\nds_train = ds_train.repeat()\nds_train = ds_train.prefetch(buffer_size=AUTOTUNE)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57685158,"AnswerCount":1,"Tags":"<protocol-buffers><pytorch><recurrent-neural-network>","CreationDate":"2019-08-28T04:39:23.977","AcceptedAnswerId":null,"OwnerUserId":8119473.0,"Title":"Dataloader on top of protobuf file using pytorch's torch.utils.data.IterableDataset","Body":"<p>I am building a RNN network using pytorch.\nThe data is stored in various protobuf file.\nEach record in protobuf represents one training example with multiple timestamp.<\/p>\n\n<p>As this is very large dataset, reading the whole data in memory or random read by extending torch.utils.data.Dataset class isn't feasible.<\/p>\n\n<p>As per the docs using the torch.utils.data.IterableDataset is recommended.<\/p>\n\n<p>DataLoader on top of IterableDataset would be able to achieve parallelism<\/p>\n\n<p>However I am not able to find an implementation of this on custom data,  docs only talk about a simple range iterator.<\/p>\n\n<pre><code>import math\n\nimport stream\nfrom src import record_pb2\nimport torch\n\nclass MyIterableDataset(torch.utils.data.IterableDataset):\n    def __init__(self, pb_file):\n        self.pb_file = pb_file\n        self.start = 0\n        self.end = 0\n        # One time read of the data to get the total count of records in the dataset\n        with stream.open(self.pb_file, 'rb') as data_stream:\n            for _ in data_stream:\n                self.end += 1\n\n    def __iter__(self):\n        worker_info = torch.utils.data.get_worker_info()\n        if worker_info is None:  # Single-process data loading, return the full iterator\n            iter_start = self.start\n            iter_end = self.end\n        else:\n            # in a worker process, split the workload\n            per_worker = int(math.ceil((self.end - self.start))\/float(worker_info.num_workers))\n            worker_id = worker_info.id\n            iter_start = self.start + worker_id * per_worker\n            iter_end = min(iter_start + per_worker, self.end)\n\n        data_stream = stream.open(self.pb_file, 'rb')\n\n        # Block to skip the streaming data till the iter start for the current worker process\n        i = 0\n        for _ in data_stream:\n            i += 1\n            if i &gt;= iter_start:\n                break\n\n        return iter(self.pb_stream)\n<\/code><\/pre>\n\n<p>I am expecting a mechanism by which a parallel data feeder could be designed on top of a large streaming data (protobuf)<\/p>\n","answers":[{"AnswerId":"57685685","CreationDate":"2019-08-28T05:38:51.537","ParentId":null,"OwnerUserId":"1849998","Title":null,"Body":"<p>The <code>__iter__<\/code> method of the <code>IterableDataset<\/code> would <code>yield<\/code> your data samples one at a time. In a parallel setup, you have to choose the samples based on worker_id. And with respect to the <code>DataLoader<\/code> using this dataset, <code>shuffle<\/code> and <code>sampler<\/code> options would not work, as an <code>IterableDataset<\/code> is not going to have any indices. In other words, have your dataset yield one sample at a time and the data loader will take care of loading them. Does this answer?<\/p>\n"}]}
+{"QuestionId":57685633,"AnswerCount":1,"Tags":"<keras><neural-network><word-embedding><glove>","CreationDate":"2019-08-28T05:33:07.313","AcceptedAnswerId":57787400.0,"OwnerUserId":205162.0,"Title":"Should the vocabulary be restricted to the training-set vocabulary when training an NN model with pretrained word2vec like GLOVE?","Body":"<p>I wanted to use word embeddings for the embedding Layer in my neural network using pre-trained vectors from GLOVE. Do I need to restrict the vocabulary to the training-set when constructing the word2index dictionary? \nWouldn't that lead to a limited non-generalizable model? \nIs considering all the vocabulary of GLOVE a recommended practice?<\/p>\n","answers":[{"AnswerId":"57787400","CreationDate":"2019-09-04T11:25:47.420","ParentId":null,"OwnerUserId":"7339624","Title":null,"Body":"<p>Yes, it is better to restrict your vocab size. Because pre-trained embeddings (like GLOVE) have many words in them that are not very useful (and so Word2Vec) and the bigger vocab size the more RAM you need and other problems.<\/p>\n\n<p>Select your tokens from all of your data. it won't lead to a limited non-generalizable model if your data is big enough. if you think that your data does not have as many tokens as are needed, then you should know 2 things:<\/p>\n\n<ol>\n<li>Your data is not good enough and you have to gather more.<\/li>\n<li>Your model can't generate well on the tokens that it hasn't seen at training! so it has no point to having many unused words on your embedding and better to gather more data to cover those words.<\/li>\n<\/ol>\n\n<p>I have an answer to show how you can select a minor set of word vectors from a pre-trained model <a href=\"https:\/\/stackoverflow.com\/a\/55725093\/7339624\">in here<\/a><\/p>\n"}]}
+{"QuestionId":57685689,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2019-08-28T05:39:14.153","AcceptedAnswerId":null,"OwnerUserId":11158631.0,"Title":"ENet sementic segmentation model is not working for smaller images","Body":"<p>I am trying to segment road and non-road part using ENet deep learning model. I uses this github link: <a href=\"https:\/\/github.com\/kwotsin\/TensorFlow-ENet\" rel=\"nofollow noreferrer\">https:\/\/github.com\/kwotsin\/TensorFlow-ENet<\/a> which has original image size of 340X480 and it's working fine for images of 340X480 or above but as soon as I reduxe the size of the image  it's not working. It's showing a ruined image of random black and white pixel. Even I try to reduce the size in same aspect ratio but it still not working.   <\/p>\n\n<p>Here is my ENet model structure code:<\/p>\n\n\n\n<pre class=\"lang-html prettyprint-override\"><code>#Now actually start building the network\ndef ENet(inputs,\n         num_classes,\n         batch_size,\n         num_initial_blocks=1,\n         stage_two_repeat=2,\n         skip_connections=True,\n         reuse=None,\n         is_training=True,\n         scope='ENet'):\n    '''\n    The ENet model for real-time semantic segmentation!\n    INPUTS:\n    - inputs(Tensor): a 4D Tensor of shape [batch_size, image_height, image_width, num_channels] that represents one batch of preprocessed images.\n    - num_classes(int): an integer for the number of classes to predict. This will determine the final output channels as the answer.\n    - batch_size(int): the batch size to explictly set the shape of the inputs in order for operations to work properly.\n    - num_initial_blocks(int): the number of times to repeat the initial block.\n    - stage_two_repeat(int): the number of times to repeat stage two in order to make the network deeper.\n    - skip_connections(bool): if True, add the corresponding encoder feature maps to the decoder. They are of exact same shapes.\n    - reuse(bool): Whether or not to reuse the variables for evaluation.\n    - is_training(bool): if True, switch on batch_norm and prelu only during training, otherwise they are turned off.\n    - scope(str): a string that represents the scope name for the variables.\n    OUTPUTS:\n    - net(Tensor): a 4D Tensor output of shape [batch_size, image_height, image_width, num_classes], where each pixel has a one-hot encoded vector\n                      determining the label of the pixel.\n    '''\n    #Set the shape of the inputs first to get the batch_size information\n    inputs_shape = inputs.get_shape().as_list()\n    inputs.set_shape(shape=(batch_size, inputs_shape[1], inputs_shape[2], inputs_shape[3]))\n\n    with tf.variable_scope(scope, reuse=reuse):\n        #Set the primary arg scopes. Fused batch_norm is faster than normal batch norm.\n        with slim.arg_scope([initial_block, bottleneck], is_training=is_training),\\\n             slim.arg_scope([slim.batch_norm], fused=True), \\\n             slim.arg_scope([slim.conv2d, slim.conv2d_transpose], activation_fn=None): \n            #=================INITIAL BLOCK=================\n            net = initial_block(inputs, scope='initial_block_1')\n            for i in xrange(2, max(num_initial_blocks, 1) + 1):\n                net = initial_block(net, scope='initial_block_' + str(i))\n\n            #Save for skip connection later\n            if skip_connections:\n                net_one = net\n\n            #===================STAGE ONE=======================\n            net, pooling_indices_1, inputs_shape_1 = bottleneck(net, output_depth=64, filter_size=3, regularizer_prob=0.01, downsampling=True, scope='bottleneck1_0')\n            net = bottleneck(net, output_depth=64, filter_size=3, regularizer_prob=0.01, scope='bottleneck1_1')\n            net = bottleneck(net, output_depth=64, filter_size=3, regularizer_prob=0.01, scope='bottleneck1_2')\n            net = bottleneck(net, output_depth=64, filter_size=3, regularizer_prob=0.01, scope='bottleneck1_3')\n            net = bottleneck(net, output_depth=64, filter_size=3, regularizer_prob=0.01, scope='bottleneck1_4')\n\n            #Save for skip connection later\n            if skip_connections:\n                net_two = net\n\n            #regularization prob is 0.1 from bottleneck 2.0 onwards\n            with slim.arg_scope([bottleneck], regularizer_prob=0.1):\n                net, pooling_indices_2, inputs_shape_2 = bottleneck(net, output_depth=128, filter_size=3, downsampling=True, scope='bottleneck2_0')\n\n                #Repeat the stage two at least twice to get stage 2 and 3:\n                for i in xrange(2, max(stage_two_repeat, 2) + 2):\n                    net = bottleneck(net, output_depth=128, filter_size=3, scope='bottleneck'+str(i)+'_1')\n                    net = bottleneck(net, output_depth=128, filter_size=3, dilated=True, dilation_rate=2, scope='bottleneck'+str(i)+'_2')\n                    net = bottleneck(net, output_depth=128, filter_size=5, asymmetric=True, scope='bottleneck'+str(i)+'_3')\n                    net = bottleneck(net, output_depth=128, filter_size=3, dilated=True, dilation_rate=4, scope='bottleneck'+str(i)+'_4')\n                    net = bottleneck(net, output_depth=128, filter_size=3, scope='bottleneck'+str(i)+'_5')\n                    net = bottleneck(net, output_depth=128, filter_size=3, dilated=True, dilation_rate=8, scope='bottleneck'+str(i)+'_6')\n                    net = bottleneck(net, output_depth=128, filter_size=5, asymmetric=True, scope='bottleneck'+str(i)+'_7')\n                    net = bottleneck(net, output_depth=128, filter_size=3, dilated=True, dilation_rate=16, scope='bottleneck'+str(i)+'_8')\n\n            with slim.arg_scope([bottleneck], regularizer_prob=0.1, decoder=True):\n                #===================STAGE FOUR========================\n                bottleneck_scope_name = \"bottleneck\" + str(i + 1)\n\n                #The decoder section, so start to upsample.\n                net = bottleneck(net, output_depth=64, filter_size=3, upsampling=True,\n                                 pooling_indices=pooling_indices_2, output_shape=inputs_shape_2, scope=bottleneck_scope_name+'_0')\n\n                #Perform skip connections here\n                if skip_connections:\n                    net = tf.add(net, net_two, name=bottleneck_scope_name+'_skip_connection')\n\n                net = bottleneck(net, output_depth=64, filter_size=3, scope=bottleneck_scope_name+'_1')\n                net = bottleneck(net, output_depth=64, filter_size=3, scope=bottleneck_scope_name+'_2')\n\n                #===================STAGE FIVE========================\n                bottleneck_scope_name = \"bottleneck\" + str(i + 2)\n\n                net = bottleneck(net, output_depth=16, filter_size=3, upsampling=True,\n                                 pooling_indices=pooling_indices_1, output_shape=inputs_shape_1, scope=bottleneck_scope_name+'_0')\n\n                #perform skip connections here\n                if skip_connections:\n                    net = tf.add(net, net_one, name=bottleneck_scope_name+'_skip_connection')\n\n                net = bottleneck(net, output_depth=16, filter_size=3, scope=bottleneck_scope_name+'_1')\n\n            #=============FINAL CONVOLUTION=============\n            logits = slim.conv2d_transpose(net, num_classes, [2,2], stride=2, scope='fullconv')\n            probabilities = tf.nn.softmax(logits, name='logits_to_softmax')\n\n        return logits, probabilities\n<\/code><\/pre>\n\n<p>and here is the full link of the code: <a href=\"https:\/\/github.com\/kwotsin\/TensorFlow-ENet\/blob\/master\/enet.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/kwotsin\/TensorFlow-ENet\/blob\/master\/enet.py<\/a><\/p>\n\n<p>Prediction segmentation code:<\/p>\n\n\n\n<pre class=\"lang-html prettyprint-override\"><code>image_dir = '.\/dataset\/test\/'\nimages_list = sorted([os.path.join(image_dir, file) for file in os.listdir(image_dir) if file.endswith('.png')])\n\ncheckpoint_dir = \"log\/original\/\"\nlisti = os.listdir(checkpoint_dir)\nprint(images_list)\ncheckpoint = tf.train.latest_checkpoint(\"\/home\/nikhil_m\/TensorFlow-ENet\/log\/original\")\nprint(tf.train.latest_checkpoint(\"\/home\/nikhil_m\/TensorFlow-ENet\/log\/original\"),'-DDD--------------------------------------++++++++++++++++++++++++++++++++++++++++++++++++++++')\nnum_initial_blocks = 1\nskip_connections = False\nstage_two_repeat = 2\n'''\n#Labels to colours are obtained from here:\nhttps:\/\/github.com\/alexgkendall\/SegNet-Tutorial\/blob\/c922cc4a4fcc7ce279dd998fb2d4a8703f34ebd7\/Scripts\/test_segmentation_camvid.py\n\nHowever, the road_marking class is collapsed into the road class in the dataset provided.\n\nClasses:\n------------\nSky = [128,128,128]\nBuilding = [128,0,0]\nPole = [192,192,128]\nRoad_marking = [255,69,0]\nRoad = [128,64,128]\nPavement = [60,40,222]\nTree = [128,128,0]\nSignSymbol = [192,128,128]\nFence = [64,64,128]\nCar = [64,0,128]\nPedestrian = [64,64,0]\nBicyclist = [0,128,192]\nUnlabelled = [0,0,0]\n'''\nlabel_to_colours =    {0: [128,128,128],\n                       1: [0, 0, 0]}\n\n#Create the photo directory\nphoto_dir = checkpoint_dir + \"\/test_images\"\nif not os.path.exists(photo_dir):\n    os.mkdir(photo_dir)\n\n#Create a function to convert each pixel label to colour.\ndef grayscale_to_colour(image):\n    print 'Converting image...'\n    image = image.reshape((256, 256, 1))\n    image = np.repeat(image, 3, axis=-1)\n    for i in xrange(image.shape[0]):\n        for j in xrange(image.shape[1]):\n            label = int(image[i][j][0])\n            image[i][j] = np.array(label_to_colours[label])\n\n    return image\n\n\nwith tf.Graph().as_default() as graph:\n    images_tensor = tf.train.string_input_producer(images_list, shuffle=False)\n    reader = tf.WholeFileReader()\n    key, image_tensor = reader.read(images_tensor)\n    image = tf.image.decode_png(image_tensor, channels=3)\n    print(image.shape, 'newwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwww shapeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee')\n    # image = tf.image.resize_image_with_crop_or_pad(image, 360, 480)\n    # image = tf.cast(image, tf.float32)\n    image = preprocess(image)\n    images = tf.train.batch([image], batch_size = 10, allow_smaller_final_batch=True)\n\n    #Create the model inference\n    with slim.arg_scope(ENet_arg_scope()):\n        logits, probabilities = ENet(images,\n                                     num_classes=2,\n                                     batch_size=10,\n                                     is_training=True,\n                                     reuse=None,\n                                     num_initial_blocks=num_initial_blocks,\n                                     stage_two_repeat=stage_two_repeat,\n                                     skip_connections=skip_connections)\n\n    variables_to_restore = slim.get_variables_to_restore()\n    saver = tf.train.Saver(variables_to_restore)\n    def restore_fn(sess):\n        return saver.restore(sess, checkpoint)\n\n    predictions = tf.argmax(probabilities, -1)\n    predictions = tf.cast(predictions, tf.float32)\n    print 'HERE', predictions.get_shape()\n\n    sv = tf.train.Supervisor(logdir=None, init_fn=restore_fn)\n\n    with sv.managed_session() as sess:\n        for i in xrange(len(images_list) \/ 10):\n            segmentations = sess.run(predictions)\n            # print segmentations.shape\n            print(segmentations.shape, 'shape') \n            for j in xrange(segmentations.shape[0]):\n\n                converted_image = grayscale_to_colour(segmentations[j])\n                print 'Saving image %s\/%s' %(i*10 + j, len(images_list))\n                #plt.axis('off')\n                #plt.imshow(converted_image)\n                imsave(photo_dir + \"\/image_%s.png\" %(i*10 + j), converted_image)\n                # plt.show()\n<\/code><\/pre>\n\n<p>Here is the full code link: <a href=\"https:\/\/github.com\/kwotsin\/TensorFlow-ENet\/blob\/master\/predict_segmentation.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/kwotsin\/TensorFlow-ENet\/blob\/master\/predict_segmentation.py<\/a><\/p>\n","answers":[{"AnswerId":"59031658","CreationDate":"2019-11-25T12:16:43.377","ParentId":null,"OwnerUserId":"8551949","Title":null,"Body":"<p>You can try this model. Its written in tf.keras<\/p>\n\n<pre><code>    import tensorflow as tf\n    from tensorflow.keras.layers import *\n    from tensorflow.keras.models import Model\n\n    print('Tensorflow', tf.__version__)\n\n    def initial_block(inp):\n        inp1 = inp\n        conv = Conv2D(filters=13, kernel_size=3, strides=2, padding='same', kernel_initializer='he_normal')(inp)\n        pool = MaxPool2D(2)(inp1)\n        concat = concatenate([conv, pool])\n        return concat\n\n\n    def encoder_bottleneck(inp, filters, name, dilation_rate=2, downsample=False, dilated=False, asymmetric=False, drop_rate=0.1):\n        reduce = filters \/\/ 4\n        down = inp\n        kernel_stride = 1\n\n        #Downsample\n        if downsample:\n            kernel_stride = 2\n            pad_activations = filters - inp.shape.as_list()[-1]\n            down = MaxPool2D(2)(down)\n            down = Permute(dims=(1, 3, 2))(down)\n            down = ZeroPadding2D(padding=((0, 0), (0, pad_activations)))(down)\n            down = Permute(dims=(1, 3, 2))(down)\n\n        #1*1 Reduce\n        x = Conv2D(filters=reduce, kernel_size=kernel_stride, strides=kernel_stride, padding='same', use_bias=False, kernel_initializer='he_normal', name=f'{name}_reduce')(inp)\n        x = BatchNormalization(momentum=0.1)(x)\n        x = PReLU(shared_axes=[1, 2])(x)\n\n        #Conv\n        if not dilated and not asymmetric:\n            x = Conv2D(filters=reduce, kernel_size=3, padding='same', kernel_initializer='he_normal', name=f'{name}_conv_reg')(x)\n        elif dilated:\n            x = Conv2D(filters=reduce, kernel_size=3, padding='same', dilation_rate=dilation_rate, kernel_initializer='he_normal', name=f'{name}_reduce_dilated')(x)\n        elif asymmetric:\n            x = Conv2D(filters=reduce, kernel_size=(1,5), padding='same', use_bias=False, kernel_initializer='he_normal', name=f'{name}_asymmetric')(x)\n            x = Conv2D(filters=reduce, kernel_size=(5,1), padding='same', kernel_initializer='he_normal', name=name)(x)\n        x = BatchNormalization(momentum=0.1)(x)\n        x = PReLU(shared_axes=[1, 2])(x)\n\n        #1*1 Expand\n        x = Conv2D(filters=filters, kernel_size=1, padding='same', use_bias=False, kernel_initializer='he_normal', name=f'{name}_expand')(x)\n        x = BatchNormalization(momentum=0.1)(x)\n        x = SpatialDropout2D(rate=drop_rate)(x)\n\n        concat = Add()([x, down])\n        concat = PReLU(shared_axes=[1, 2])(concat)\n        return concat\n\n\n    def decoder_bottleneck(inp, filters, name, upsample=False):\n        reduce = filters \/\/ 4\n        up = inp\n\n        #Upsample\n        if upsample:\n            up = Conv2D(filters=filters, kernel_size=1, strides=1, padding='same', use_bias=False, kernel_initializer='he_normal', name=f'{name}_upsample')(up)\n            up = UpSampling2D(size=2)(up)\n\n        #1*1 Reduce\n        x = Conv2D(filters=reduce, kernel_size=1, strides=1, padding='same', use_bias=False, kernel_initializer='he_normal', name=f'{name}_reduce')(inp)\n        x = BatchNormalization(momentum=0.1)(x)\n        x = PReLU(shared_axes=[1, 2])(x)\n\n        #Conv\n        if not upsample:\n            x = Conv2D(filters=reduce, kernel_size=3, strides=1, padding='same', kernel_initializer='he_normal', name=f'{name}_conv_reg')(x)\n        else:\n            x = Conv2DTranspose(filters=reduce, kernel_size=3, strides=2, padding='same', kernel_initializer='he_normal', name=f'{name}_transpose')(x)\n        x = BatchNormalization(momentum=0.1)(x)\n        x = PReLU(shared_axes=[1, 2])(x)\n\n        #1*1 Expand\n        x = Conv2D(filters=filters, kernel_size=1, strides=1, padding='same', use_bias=False, kernel_initializer='he_normal', name=f'{name}_expand')(x)\n        x = BatchNormalization(momentum=0.1)(x)\n\n        concat = Add()([x, up])\n        concat = ReLU()(concat)\n\n        return concat\n\n\n    def ENet(H, W, nclasses):\n        '''\n        Args:\n            H: Height of the image\n            W: Width of the image\n            nclasses: Total no of classes\n\n        Returns:\n            model: Keras model in .h5 format\n        '''\n\n        inp = Input(shape=(H, W, 3))\n        enc = initial_block(inp)\n\n        #Bottleneck 1.0\n        enc = encoder_bottleneck(enc, 64, name='enc1', downsample=True, drop_rate=0.001)\n\n        enc = encoder_bottleneck(enc, 64, name='enc1.1', drop_rate=0.001)\n        enc = encoder_bottleneck(enc, 64, name='enc1.2', drop_rate=0.001)\n        enc = encoder_bottleneck(enc, 64, name='enc1.3', drop_rate=0.001)\n        enc = encoder_bottleneck(enc, 64, name='enc1.4', drop_rate=0.001)\n\n        #Bottleneck 2.0\n        enc = encoder_bottleneck(enc, 128, name='enc2.0', downsample=True)\n        enc = encoder_bottleneck(enc, 128, name='enc2.1')\n        enc = encoder_bottleneck(enc, 128, name='enc2.2', dilation_rate=2, dilated=True)\n        enc = encoder_bottleneck(enc, 128, name='enc2.3', asymmetric=True)\n        enc = encoder_bottleneck(enc, 128, name='enc2.4', dilation_rate=4, dilated=True)\n        enc = encoder_bottleneck(enc, 128, name='enc2.5')\n        enc = encoder_bottleneck(enc, 128, name='enc2.6', dilation_rate=8, dilated=True)\n        enc = encoder_bottleneck(enc, 128, name='enc2.7', asymmetric=True)\n        enc = encoder_bottleneck(enc, 128, name='enc2.8', dilation_rate=16, dilated=True)\n\n        #Bottleneck 3.0\n        enc = encoder_bottleneck(enc, 128, name='enc3.0')\n        enc = encoder_bottleneck(enc, 128, name='enc3.1', dilation_rate=2, dilated=True)\n        enc = encoder_bottleneck(enc, 128, name='enc3.2', asymmetric=True)\n        enc = encoder_bottleneck(enc, 128, name='enc3.3', dilation_rate=4, dilated=True)\n        enc = encoder_bottleneck(enc, 128, name='enc3.4')\n        enc = encoder_bottleneck(enc, 128, name='enc3.5', dilation_rate=8, dilated=True)\n        enc = encoder_bottleneck(enc, 128, name='enc3.6', asymmetric=True)\n        enc = encoder_bottleneck(enc, 128, name='enc3.7', dilation_rate=16, dilated=True)\n\n        #Bottleneck 4.0\n        dec = decoder_bottleneck(enc, 64, name='dec4.0', upsample=True)\n        dec = decoder_bottleneck(dec, 64, name='dec4.1')\n        dec = decoder_bottleneck(dec, 64, name='dec4.2')\n\n        #Bottleneck 5.0\n        dec = decoder_bottleneck(dec, 16, name='dec5.0', upsample=True)\n        dec = decoder_bottleneck(dec, 16, name='dec5.1')\n\n        dec = Conv2DTranspose(filters=nclasses, kernel_size=2, strides=2, padding='same', kernel_initializer='he_normal', name='fullconv')(dec)\n        dec = Activation('softmax')(dec)\n\n        model = Model(inputs=inp, outputs=dec, name='Enet')\n        model.save(f'enet_{nclasses}.h5')\n        return model\n<\/code><\/pre>\n"}]}
+{"QuestionId":57685731,"AnswerCount":0,"Tags":"<python><tensorflow><keras><out-of-memory><gpu>","CreationDate":"2019-08-28T05:42:48.727","AcceptedAnswerId":null,"OwnerUserId":4849439.0,"Title":"Keras small network takes a lot of GPU memory","Body":"<p>I have a really small neural network - <\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nscores_input (InputL (None, 2)                 0         \n_________________________________________________________________\ndense1 (Dense)       (None, 1)                 2         \n_________________________________________________________________\nbn (BatchNormalizati (None, 1)                 4         \n_________________________________________________________________\nsigmoid (Activation) (None, 1)                 0         \n=================================================================\nTotal params: 6\nTrainable params: 4\nNon-trainable params: 2\n<\/code><\/pre>\n\n<p>However, it takes about 1.3 GB in the gpu memory as shown by nvidia-smi<\/p>\n\n<pre><code>Wed Aug 28 08:41:38 2019       \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 960     Off  | 00000000:01:00.0  On |                  N\/A |\n| 30%   42C    P2    29W \/ 120W |   1763MiB \/  1988MiB |      0%      Default |\n+-------------------------------+----------------------+----------------------+\n\n+-----------------------------------------------------------------------------+\n| Processes:                                                       GPU Memory |\n|  GPU       PID   Type   Process name                             Usage      |\n|=============================================================================|\n|    0      1258      G   \/usr\/lib\/xorg\/Xorg                           251MiB |\n|    0      2336      G   compiz                                       231MiB |\n|    0     25537      G   ...\/innereye\/qtcreator-4.9.2\/bin\/qtcreator     2MiB |\n|    0     30098      G   ...-token=DA8EE4CD7070EDEBCD3537BAAD982629    37MiB |\n|    0     30436      C   python                                      1227MiB |\n+-----------------------------------------------------------------------------+\n<\/code><\/pre>\n\n<p>I need to load another (larger) network, but they don't fit together in the GPU memory. Any help?<\/p>\n","answers":[]}
+{"QuestionId":57686201,"AnswerCount":0,"Tags":"<r><keras>","CreationDate":"2019-08-28T06:25:19.393","AcceptedAnswerId":null,"OwnerUserId":521776.0,"Title":"ImageDataGenerator object has no attribute shape","Body":"<p>I am training a simple CNN using the ImageDataGenerator, nothing fancy. The model trains for the first epoch, and at the end I get this error message:<\/p>\n\n<pre><code>Error in py_call_impl(callable, dots$args, dots$keywords) :\nAttributeError: 'ImageDataGenerator' object has no attribute 'shape'\n<\/code><\/pre>\n\n<p>The funny thing is, if I run an equivalent example using python from within the <code>r-reticulate<\/code> environment, everything works fine.<\/p>\n\n<p>Here are some info from jupyter:<\/p>\n\n<pre><code>sys.executable C:\\Users\\edo\\Anaconda3\\envs\\r-reticulate\\python.exe\nsys.version 3.6.7 (default, Jul  2 2019, 02:21:41) [MSC v.1900 64 bit (AMD64)]\nsys.version_info sys.version_info(major=3, minor=6, micro=7, releaselevel='final', serial=0)\ntf.__version__ 1.13.1\ntf.__file__ C:\\Users\\edo\\Anaconda3\\envs\\r-reticulate\\lib\\site-packages\\tensorflow\\__init__.py\nnp.__version__ 1.17.1\nnp.__file__ C:\\Users\\edo\\Anaconda3\\envs\\r-reticulate\\lib\\site-packages\\numpy\\__init__.py\n<\/code><\/pre>\n\n<p>And here are info from Rstudio:<\/p>\n\n<pre><code>&gt; reticulate::py_config()\npython:         C:\\Users\\edo\\Anaconda3\\envs\\r-reticulate\\python.exe\nlibpython:      C:\/Users\/edo\/Anaconda3\/envs\/r-reticulate\/python36.dll\npythonhome:     C:\\Users\\edo\\ANACON~1\\envs\\R-RETI~1\nversion:        3.6.7 (default, Jul  2 2019, 02:21:41) [MSC v.1900 64 bit (AMD64)]\nArchitecture:   64bit\nnumpy:          C:\\Users\\edo\\ANACON~1\\envs\\R-RETI~1\\lib\\site-packages\\numpy\nnumpy_version:  1.17.1\ntensorflow:     C:\\Users\\edo\\ANACON~1\\envs\\R-RETI~1\\lib\\site-packages\\tensorflow\\__init__.p\n\npython versions found: \n C:\\Users\\edo\\Anaconda3\\envs\\r-reticulate\\python.exe\n C:\\Users\\edo\\ANACON~1\\python.exe\n C:\\PYTHON~1\\\\python.exe\n C:\\PROGRA~2\\MICROS~1\\Shared\\PYTHON~1\\\\python.exe\n C:\\Users\\edo\\Anaconda3\\python.exe\n C:\\Users\\edo\\Anaconda3\\envs\\py2\\python.exe\n C:\\Users\\wangk\\Anaconda3\\python.exe\n&gt; tf_config()\nTensorFlow v1.13.1 (C:\\Users\\edo\\ANACON~1\\envs\\R-RETI~1\\lib\\site-packages\\tensorflow\\__init__.p)\nPython v3.6 (C:\\Users\\edo\\Anaconda3\\envs\\r-reticulate\\python.exe)\n&gt; devtools::session_info()\n- Session info -----------------------------------------------------------------------------------------------------------------------------\n setting  value                       \n version  R version 3.6.0 (2019-04-26)\n os       Windows 10 x64              \n system   x86_64, mingw32             \n ui       RStudio                     \n language (EN)                        \n collate  English_United States.1252  \n ctype    English_United States.1252  \n tz       Europe\/Berlin               \n date     2019-08-27                  \n\n- Packages ---------------------------------------------------------------------------------------------------------------------------------\n package     * version date       lib source        \n assertthat    0.2.1   2019-03-21 [1] CRAN (R 3.6.0)\n backports     1.1.4   2019-04-10 [1] CRAN (R 3.6.0)\n base64enc     0.1-3   2015-07-28 [1] CRAN (R 3.6.0)\n callr         3.3.1   2019-07-18 [1] CRAN (R 3.6.1)\n cli           1.1.0   2019-03-19 [1] CRAN (R 3.6.0)\n crayon        1.3.4   2017-09-16 [1] CRAN (R 3.6.0)\n desc          1.2.0   2018-05-01 [1] CRAN (R 3.6.1)\n devtools      2.1.0   2019-07-06 [1] CRAN (R 3.6.1)\n digest        0.6.20  2019-07-04 [1] CRAN (R 3.6.1)\n fs            1.3.1   2019-05-06 [1] CRAN (R 3.6.1)\n generics      0.0.2   2018-11-29 [1] CRAN (R 3.6.1)\n glue          1.3.1   2019-03-12 [1] CRAN (R 3.6.0)\n jsonlite      1.6     2018-12-07 [1] CRAN (R 3.6.0)\n keras       * 2.2.4.1 2019-04-05 [1] CRAN (R 3.6.1)\n knitr         1.23    2019-05-18 [1] CRAN (R 3.6.0)\n lattice       0.20-38 2018-11-04 [2] CRAN (R 3.6.0)\n magrittr      1.5     2014-11-22 [1] CRAN (R 3.6.0)\n Matrix        1.2-17  2019-03-22 [2] CRAN (R 3.6.0)\n memoise       1.1.0   2017-04-21 [1] CRAN (R 3.6.1)\n pkgbuild      1.0.5   2019-08-26 [1] CRAN (R 3.6.0)\n pkgload       1.0.2   2018-10-29 [1] CRAN (R 3.6.1)\n prettyunits   1.0.2   2015-07-13 [1] CRAN (R 3.6.1)\n processx      3.4.1   2019-07-18 [1] CRAN (R 3.6.1)\n ps            1.3.0   2018-12-21 [1] CRAN (R 3.6.1)\n R6            2.4.0   2019-02-14 [1] CRAN (R 3.6.0)\n Rcpp          1.0.1   2019-03-17 [1] CRAN (R 3.6.0)\n remotes       2.1.0   2019-06-24 [1] CRAN (R 3.6.1)\n reticulate    1.13    2019-07-24 [1] CRAN (R 3.6.1)\n rlang         0.4.0   2019-06-25 [1] CRAN (R 3.6.0)\n rprojroot     1.3-2   2018-01-03 [1] CRAN (R 3.6.0)\n rstudioapi    0.10    2019-03-19 [1] CRAN (R 3.6.0)\n sessioninfo   1.1.1   2018-11-05 [1] CRAN (R 3.6.1)\n tensorflow  * 1.14.0  2019-08-01 [1] CRAN (R 3.6.1)\n testthat      2.2.1   2019-07-25 [1] CRAN (R 3.6.1)\n tfruns        1.4     2018-08-25 [1] CRAN (R 3.6.1)\n usethis       1.5.1   2019-07-04 [1] CRAN (R 3.6.1)\n whisker       0.3-2   2013-04-28 [1] CRAN (R 3.6.1)\n withr         2.1.2   2018-03-15 [1] CRAN (R 3.6.0)\n xfun          0.8     2019-06-25 [1] CRAN (R 3.6.0)\n zeallot       0.1.0   2018-01-28 [1] CRAN (R 3.6.0)\n\n[1] C:\/Users\/edo\/R\/win-library\/3.6\n[2] C:\/Program Files\/R\/R-3.6.0\/library\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57686486,"AnswerCount":0,"Tags":"<keras><lstm>","CreationDate":"2019-08-28T06:50:19.587","AcceptedAnswerId":null,"OwnerUserId":7687835.0,"Title":"Use checkpoint model and restore it in memory,but when I user `predict` is very slow","Body":"<p>When I train a LSTM model,I use <code>checkpoint<\/code> to save it ,then I use <code>load_model<\/code> to restore it(this step also takes some times),but when i called <code>predict<\/code>,it is very slow,just as I was trainning it.<\/p>\n\n<p>I run my code with CPU in my PC,and the size of my model's <code>.h5<\/code> file is 63M.But,when i use some ML models it always very fast(the ml model size about 8M)<\/p>\n\n<p>fit code<\/p>\n\n<pre><code> model.fit(x=[X_add_train_fold_s1, X_add_train_fold_s2], y=y_add_train_fold,\n           validation_data=([X_add_val_fold_s1, X_add_val_fold_s2], y_add_val_fold),\n           batch_size=batch_size,\n           epochs=nepoch,\n           verbose=1,\n           class_weight=project.class_wight,\n           callbacks=[EarlyStopping(\n                          monitor='val_loss',\n                          min_delta=0.002,\n                          patience=3,\n                          verbose=1,\n                          mode='auto'),\n                      reduce_lr,\n                      ModelCheckpoint(\n                          model_checkpoint_path,\n                          monitor='val_loss',\n                          save_best_only=True,\n                          save_weights_only=False,\n                          verbose=1),                                           \n                      TensorBoard(\n                          log_dir='.\/log\/nn\/mandist\/{}\/{}'.format(fold_num + 1,project.batch_size))])\n<\/code><\/pre>\n\n<p>restore code:<\/p>\n\n<pre><code>models = []\nfor fold_num in range(self.num_folds):\n    model_checkpoint_path = project.trained_model_dir + 'dl_siamese_lstm_dssm_model{}.h5'.format(fold_num)\n    model = load_model(model_checkpoint_path, custom_objects=self.custom_objects)\n    models.append(model)\nself.dssm_models = models\n<\/code><\/pre>\n\n<p>predict code:<\/p>\n\n<pre><code>for model in self.dssm_models:\n    y_train_oofp += predict1(model, s1_word_ids_pad, s2_word_ids_pad, s1_char_ids_pad, s2_char_ids_pad)\n    col_names = ['{}_{}'.format(feature_name, index) for index in range(class_num)]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57687200,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><sparse-matrix>","CreationDate":"2019-08-28T07:36:47.633","AcceptedAnswerId":null,"OwnerUserId":9427280.0,"Title":"How to multiply very large sparse matrix in tensorflow?","Body":"<p>I have a very large sparse tensor <code>input_a<\/code> with shape: [20k, 20k], and a dense tensor <code>input_b<\/code> with shape: [20k, 512]. nnz(input_a) is about 400k (There are 400k nonzero values in input_a). I would like to multiply these two tensor, to get the output tensor <code>output<\/code> with shape: [20k, 512]. If I use <code>tf.sparse_tensor_dense_matmul<\/code> to multiply input_a and input_b, there will be an error: (with tensorflow version v1.8)<\/p>\n\n<blockquote>\n  <p>Cannot use GPU when output.shape[1] * nnz(a) > 2^31\"<\/p>\n<\/blockquote>\n\n<p>If output.shape[1] is 20k, it might make sense, because 20k * 400k = 8000000000 > 2^31 = 2147483648. But, I think output.shape[1] should be 512 in this case?<\/p>\n\n<p>My code: <\/p>\n\n<p><code>input_a<\/code>: shape [20k, 20k]<\/p>\n\n<p><code>input_b<\/code>: shape [20k, 512]<\/p>\n\n<p><code>output<\/code>: shape [20k, 512]<\/p>\n\n<pre><code>output = tf.sparse_tensor_dense_matmul(input_a, input_b)\n<\/code><\/pre>\n\n<p>Then I will get the error message.<\/p>\n\n<p>If I transform input_a to a dense tensor and then use <code>tf.matmul<\/code>, it works, but it might cost a lot of memory and time?<\/p>\n\n<pre><code>dense_input_a = tf.sparse_tensor_to_dense(dense_input_a)\noutput = tf.matmul(dense_input_a, input_b, a_is_sparse = True)\n<\/code><\/pre>\n\n<p>Why the error occurred and how should I multiply these two tensor in a efficient way? Thanks!<\/p>\n","answers":[]}
+{"QuestionId":57687251,"AnswerCount":1,"Tags":"<tensorflow><keras><tensorflow-lite>","CreationDate":"2019-08-28T07:39:53.453","AcceptedAnswerId":null,"OwnerUserId":2604291.0,"Title":"What is the 'index' in TFLite interpreter.get_input_details referring to?","Body":"<p>I'm working on doing some inference with Keras\/TensorFlow models but the documenation seems a little sparse so I'm trying to learn and document as much as possible as I go and not just rely on copied code examples. Examples include this line:<\/p>\n\n<p><code>interpreter.get_input_details()<\/code><\/p>\n\n<p>Which returns a list containing a single dictionary with the following info:<\/p>\n\n<p><code>[{'name': 'conv2d_input', 'index': 8, 'shape': array([ 1, 28, 28,  1]), 'dtype': &lt;class 'numpy.float32'&gt;, 'quantization': (0.0, 0)}]<\/code><\/p>\n\n<p><code>get_output_details<\/code> returns something very similar. Most of this is self-explanatory but I'm not clear what the index refers to. Examples I've found show that the number is needed by the set_tensor and get_tensor methods. All I've found so far is a comment somewhere saying that this refers to a tensor index which seems like a tautology to me!<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"57781992","CreationDate":"2019-09-04T05:28:11.597","ParentId":null,"OwnerUserId":"366806","Title":null,"Body":"<p>In TFLite interpreter, all tensors are put into a tensor list (see the <code>TfLiteTensor* tensors;<\/code> in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r2.0\/tensorflow\/lite\/c\/c_api_internal.h#L434\" rel=\"nofollow noreferrer\">TfLiteContext<\/a>), the index is the index of tensor in the tensor list.<\/p>\n"}]}
+{"QuestionId":57687429,"AnswerCount":0,"Tags":"<r><tensorflow><anaconda><rstudio>","CreationDate":"2019-08-28T07:51:05.460","AcceptedAnswerId":null,"OwnerUserId":6911037.0,"Title":"Can't find conda at rstudio terminal, while it is workable outside?","Body":"<p>Just trying to use tensorflow-gpu on my notebook, after installing anaconda, VSC++, CUDA, cdnn, and then tensorflow-gpu on conda base environment.\nBefore the gpu version, I'd using keras(cpu) for a while, so the cpu version is installed, and running well.<\/p>\n\n<p>My question is, why I can't find r-tensorflow environment, even the <code>conda<\/code> commend in the rstudio terminal does not exist.<\/p>\n\n<p>the rstudio terminal windows<\/p>\n\n<pre><code>eric1@NB  \/g\/work\/R_Prj\/KEK\/2019-08-TEST (master)\n$ conda\nbash: conda: command not found \n<\/code><\/pre>\n\n<p>the Anaconda prompt window<\/p>\n\n<pre><code>(base) C:\\Users\\eric1&gt;conda info --envs\n# conda environments:\n#\nbase                  *  C:\\Users\\eric1\\Anaconda3\n<\/code><\/pre>\n\n<p>and both on the same notebook at the same time.\nPlease advise how I can see conda in the rstudio, and how I can install tensorflow-gpu version in R. Many thanks.<\/p>\n","answers":[]}
+{"QuestionId":57687465,"AnswerCount":0,"Tags":"<tensorflow><python-3.7>","CreationDate":"2019-08-28T07:53:06.083","AcceptedAnswerId":null,"OwnerUserId":10150715.0,"Title":"Tensorflow `categorical_crossentropy` is creating dimension mismatch error in model.fit","Body":"<pre><code>IMG_SHAPE = (192, 192, 3)\n\n# Create the base model from the pre-trained model MobileNet V2\n\nbase_model = tf.keras.applications.MobileNetV2(input_shape=IMG_SHAPE,\n                                               include_top=False`\n                                               weights='imagenet')\n\nbase_model.trainable = False\nglobal_average_layer = tf.keras.layers.GlobalAveragePooling2D()\nprediction_layer = tf.keras.layers.Dense(3)\n\n# adding the custom layers\n\nmodel = tf.keras.Sequential([\n    base_model,\n    global_average_layer,\n    prediction_layer\n])\n\n# compiling the model\n\nbase_learning_rate = 0.0001\nmodel.compile(optimizer=tf.keras.optimizers.RMSprop(lr=base_learning_rate),\n              loss='categorical_crossentropy',\n              metrics=['accuracy'])\n\n\n# Training\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    filenames = tf.placeholder(tf.string)\n    evaluated_filenames = sess.run(filenames, feed_dict={filenames:'train.tfrecords'})\n    train = input_fn(evaluated_filenames, train=True, buffer_size=512, batch_size = 8) \n    iterator = train.make_initializable_iterator()\n    sess.run(iterator.initializer)\n    one_batch = iterator.get_next()\n    print(one_batch)\n    history = model.fit(one_batch,\n                    epochs=2,\n                    steps_per_epoch = 3\n                   )\n<\/code><\/pre>\n\n<p>I expected to see the training loss and accuracy loss printed out after each epoch and <code>history<\/code> as a dictionary for the metrics . But rather the block threw a dimension mismatch error: <\/p>\n\n<pre><code>InvalidArgumentError: Expected dimension in the range [0, 0), but got -1\n     [[{{node metrics_4\/acc\/ArgMax}}]]\n<\/code><\/pre>\n\n<p>Meanwhile, it ran each time I did one of the following:<\/p>\n\n<p>Changed the <code>cost function<\/code> to <code>binary_crossentropy<\/code>(But this is a 3 class problem).<\/p>\n\n<p>Commented out the <code>metrics=['accuracy']<\/code>in the <code>.compile<\/code> line<\/p>\n\n<p>Enabled <code>eager_mode<\/code>and took out the place holder. What could be wrong here?<\/p>\n","answers":[]}
+{"QuestionId":57687784,"AnswerCount":0,"Tags":"<python><keras><gan>","CreationDate":"2019-08-28T08:11:09.490","AcceptedAnswerId":null,"OwnerUserId":4845195.0,"Title":"How To Test That The Discriminator Of GAN Has Trained Correctly Or Not?","Body":"<p>I am using CCGAN(Complementary Conditional GAN). I have trained both the generator and the discriminator. but at the testing time, I only want to use the discriminator. but I am not able to do that. predict function is working correctly but I want to evaluate the discriminator.<\/p>\n\n<p>Training is working fine the generated images are pretty good. Nut I want to test the discriminator on different images other than trained images.<\/p>\n\n<p>I have used test_on_batch at the testing time.<\/p>\n\n<pre><code>def build_discriminator(self):\n\n    img = Input(shape=self.img_shape)\n\n    model = Sequential()\n    model.add(Conv2D(64, kernel_size=4, strides=2, padding='same', \n    input_shape=self.img_shape))\n    model.add(LeakyReLU(alpha=0.8))\n    model.add(Conv2D(128, kernel_size=4, strides=2, padding='same'))\n    model.add(LeakyReLU(alpha=0.2))\n    # model.add(InstanceNormalization())\n    model.add(Conv2D(256, kernel_size=4, strides=2, padding='same'))\n    model.add(LeakyReLU(alpha=0.2))\n    # model.add(InstanceNormalization())\n\n\n\n    img = Input(shape=self.img_shape)\n    features = model(img)\n    # print(\"Shape of the features of descriminator is::\",features.shape)\n    validity = Conv2D(1, kernel_size=4, strides=1, padding='same')(features)\n\n    label = Flatten()(features)\n    label = Dense(self.num_classes+1, activation=\"softmax\")(label)\n    modelDes= Model(img, [validity, label])\n    # print(\"Descriminator summary\")\n    # modelDes.summary()\n    return modelDes\n<\/code><\/pre>\n\n<p>At testing time i am using:<\/p>\n\n<pre><code> idx = np.random.randint(0, X_test.shape[0],256)\n imgs = X_test[idx]\n labels = y_test[idx]    \n score_loaded_model=loaded_model.test_on_batch(imgs,[valid,labels])\n<\/code><\/pre>\n\n<p>But i am unable to understand the problem. When the size of the test images are same as the trained image why it is showing error \"dense_1 requires (11,) but got (1,)\"<\/p>\n","answers":[]}
+{"QuestionId":57687975,"AnswerCount":0,"Tags":"<tensorflow><neural-network><deep-learning><lstm><deeplearning4j>","CreationDate":"2019-08-28T08:23:24.910","AcceptedAnswerId":null,"OwnerUserId":11132972.0,"Title":"What is the equations tensorflow\/deeplearning4j using for LSTM network?","Body":"<p>This is a simple question. Deeplearning4J and Tensorflow uses the same equations to compute and train a neural network. <\/p>\n\n<p>When I create a feed forward neural network, e.g <code>DenseLayer<\/code>, and then I extract the weight matrices and vectors, I get the following:<\/p>\n\n<pre><code>0_W, 0_b\n<\/code><\/pre>\n\n<p>Where the zero (0) tells me that it's the first layer and <code>W<\/code> is the weight matrix and <code>b<\/code> is the bias vector. The equation for that layer are:<\/p>\n\n<pre><code>output = 0_W * input + 0_b\n<\/code><\/pre>\n\n<p>Easy! But when I create a <code>LSTM<\/code> layer, and then I extract the following matrices and vector:<\/p>\n\n<pre><code>0_W, 0_WR, 0_b\n<\/code><\/pre>\n\n<p>So what is the equation for that layer? Is it:<\/p>\n\n<pre><code>output = 0_W * input + 0_WR * past_input + 0_b\n<\/code><\/pre>\n\n<p>? According to the equations for LSTM, it looks more like this:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ktZX7.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ktZX7.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57688066,"AnswerCount":0,"Tags":"<python><tensorflow><object-detection>","CreationDate":"2019-08-28T08:28:50.203","AcceptedAnswerId":null,"OwnerUserId":11659030.0,"Title":"Xml ElementTree ParseError how to fix?","Body":"<blockquote>\n  <p>I accidentally brought back the xml files that I completely deleted, but now I'm getting this error what I did to convert them to csv please help\n  When i try run xml to csv \u0131 got this error :<\/p>\n<\/blockquote>\n\n<p>Traceback (most recent call last):<\/p>\n\n<pre><code> File \"xml_to_csv.py\", line 36, in &lt;module&gt;\n main()\n File \"xml_to_csv.py\", line 31, in main\n xml_df = xml_to_csv(image_path)\nFile \"xml_to_csv.py\", line 10, in xml_to_csv\ntree = ET.parse(xml_file)\nFile \"C:\\Users\\MRX\\Anaconda3\\envs\\tensorflow2\\lib\\xml\\etree\\ElementTree.py\", \nline 1197, in parse\ntree.parse(source, parser)\nFile \"C:\\Users\\MRX\\Anaconda3\\envs\\tensorflow2\\lib\\xml\\etree\\ElementTree.py\", \nline 598, in parse\nself._root = parser._parse_whole(source)\nxml.etree.ElementTree.ParseError: not well-formed (invalid token): line 1, \ncolumn 0\n<\/code><\/pre>\n\n<blockquote>\n  <p>Code block is here<\/p>\n<\/blockquote>\n\n<pre><code>import os\nimport glob\nimport pandas as pd\nimport xml.etree.ElementTree as ET\n\n\ndef xml_to_csv(path):\nxml_list = []\nfor xml_file in glob.glob(path + '\/*.xml'):\n    tree = ET.parse(xml_file)\n    root = tree.getroot()\n    for member in root.findall('object'):\n        value = (root.find('filename').text,\n                 int(root.find('size')[0].text),\n                 int(root.find('size')[1].text),\n                 member[0].text,\n                 int(member[4][0].text),\n                 int(member[4][1].text),\n                 int(member[4][2].text),\n                 int(member[4][3].text)\n                 )\n        xml_list.append(value)\ncolumn_name = ['filename', 'width', 'height', 'class', 'xmin', 'ymin', 'xmax', 'ymax']\nxml_df = pd.DataFrame(xml_list, columns=column_name)\nreturn xml_df\n\n\ndef main():\n  for folder in ['train', 'test']:\n     image_path = os.path.join(os.getcwd(), ('images\/' + folder))\n     xml_df = xml_to_csv(image_path)\n     xml_df.to_csv(('images\/' + folder + '_labels.csv'), index=None)\n     print('Successfully converted xml to csv.')\n\n\nmain()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57688191,"AnswerCount":2,"Tags":"<python><numpy><tensorflow><deep-learning><pytorch>","CreationDate":"2019-08-28T08:36:24.497","AcceptedAnswerId":57688877.0,"OwnerUserId":470184.0,"Title":"Neural network from scratch - predict single example","Body":"<p>Here is a neural network I've modified from Coursera Deep Learning Specialization to train on a dataset containing a flattened array of training data : <\/p>\n\n<pre><code>%reset -s -f\n\nimport numpy as np\nimport math\n\ndef sigmoid(x):\n    return 1 \/ (1 + np.exp(-x))\n\ndef initialize_with_zeros(dim):\n\n    w = np.zeros(shape=(dim, 1))\n    b = 0\n\n    return w, b\n\nX = np.array([[1,1,1,1],[1,0,1,0] , [1,1,1,0], [0,0,0,0], [0,1,0,0], [0,1,0,1]])\nY = np.array([[1,0,1,1,1,1]])\n\nX = X.reshape(X.shape[0], -1).T\nY = Y.reshape(Y.shape[0], -1).T\n\nprint('X shape' , X.shape)\nprint('Y shape' , Y.shape)\n\nb = 1\nw, b = initialize_with_zeros(4)\n\ndef propagate(w, b, X, Y) : \n\n    m = X.shape[1]\n\n    A = sigmoid(np.dot(w.T, X) + b)  # compute activation\n    cost = (- 1 \/ m) * np.sum(Y * np.log(A) + (1 - Y) * (np.log(1 - A)))  # compute cost\n    dw = (1.\/m)*np.dot(X,((A-Y).T))\n    db = (1.\/m)*np.sum(A-Y, axis=1)\n    cost = np.squeeze(cost)\n\n    grads = {\"dw\": dw,\n             \"db\": db}\n\n    return grads, cost\n\npropagate(w , b , X , Y)\n\nlearning_rate = .001\ncosts = []\n\ndef optimize(w , b, X , Y) :\n    for i in range(2):\n\n        grads, cost = propagate(w=w, b=b, X=X, Y=Y)\n\n        dw = grads[\"dw\"]\n        db = grads[\"db\"]\n\n        w = w - learning_rate*dw\n        b = b -  learning_rate*db\n\n        if i % 100 == 0:\n            costs.append(cost)\n\n    return w , b\n\nw , b = optimize(w , b , X , Y)\n\ndef predict(w, b, X):\n\n    m = 6\n    Y_prediction = np.zeros((1,m))\n#     w = w.reshape(X.shape[0], 1)\n\n    A = sigmoid(np.dot(w.T, X) + b)\n    for i in range(A.shape[1]):\n\n        if A[0, i] &gt;= 0.5:\n            Y_prediction[0, i] = 1\n\n        else:\n            Y_prediction[0, i] = 0\n\n    return Y_prediction\n\npredict(w , b, X)\n<\/code><\/pre>\n\n<p>This works as expected but I'm struggling to predict a single example.<\/p>\n\n<p>If I use : <\/p>\n\n<pre><code>predict(w , b, X[0])\n<\/code><\/pre>\n\n<p>returns error : <\/p>\n\n<pre><code>ValueError: shapes (6,4) and (6,) not aligned: 4 (dim 1) != 6 (dim 0)\n<\/code><\/pre>\n\n<p>How to re-arrange matrix operation in order to predict a single instance ?<\/p>\n","answers":[{"AnswerId":"57688877","CreationDate":"2019-08-28T09:12:23.193","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Try <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>predict(w, b, X[:1])\n<\/code><\/pre>\n\n<p>It seems like you <code>predict<\/code> function expects <code>X<\/code> to be 2-d, when passing only one <code>X<\/code> it should have a singleton second dimension (i.e., shape=(6,1)) rather than being a single dimension (i.e., shape=(6,)).<\/p>\n"},{"AnswerId":"57688972","CreationDate":"2019-08-28T09:17:42.640","ParentId":null,"OwnerUserId":"10511793","Title":null,"Body":"<p>The error comes from the fact that predict expect to be called on a batch of data of shape <code>... * bs<\/code>. In order to predict on a single element you can create a batch of size 1 using <code>np.expand_dims<\/code> :<\/p>\n\n<pre><code>predict(w, b, np.expand_dims(X[0], axis=1)\n<\/code><\/pre>\n\n<p>should work.<\/p>\n"}]}
+{"QuestionId":57688594,"AnswerCount":0,"Tags":"<r><keras>","CreationDate":"2019-08-28T08:57:26.907","AcceptedAnswerId":null,"OwnerUserId":3572543.0,"Title":"tfruns failed to create a directory","Body":"<p>currently I am using keras in R for lstm networks. I would like to conduct hyperparameter tuning using the tfruns package. When I try to use the training_run() function an error occurs which is:<\/p>\n\n<blockquote>\n  <p>NotFoundError: Failed to create a directory:\n  runs\/2019-08-28T08-49-21Z\/...<\/p>\n<\/blockquote>\n\n<p>I already tried to resolve by updating R and all packages and executing R as an administrator, but this did not work. Has anyone an idea how to fix this? Any help would be highly appreciated!<\/p>\n\n<p>Cheers<\/p>\n","answers":[]}
+{"QuestionId":57689006,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2019-08-28T09:19:07.150","AcceptedAnswerId":57689231.0,"OwnerUserId":11325381.0,"Title":"How to get class of different image in a folder using 'for' loop. I already have a trained model","Body":"<p>I have a trained model and i want to use this model to predict the class of different images which are kept in a folder. I already tried doing this with a for loop but it is showing me file not found error.<\/p>\n\n<p>I have already tried doing this with a for loop but it is showing me file not found error. I am using the exact path of the folder location.<\/p>\n\n<pre><code># image folder\nfolder_path = '\/path\/to\/folder\/'\n# path to model\nmodel_path = '\/path\/to\/saved\/model.h5'\n# dimensions of images\nimg_width, img_height = 320, 240\n\n# load the trained model\nmodel = load_model(model_path)\nmodel.compile(loss='binary_crossentropy',\n              optimizer='rmsprop',\n              metrics=['accuracy'])\n\n# load all images into a list\nimages = []\nfor img in os.listdir(folder_path):\n    img = image.load_img(img, target_size=(img_width, img_height))\n    img = img.img_to_array(img)\n    img = np.expand_dims(img, axis=0)\n    images.append(img)\n\n# stack up images list to pass for prediction\nimages = np.vstack(images)\nclasses = model.predict_classes(images, batch_size=10)\nprint(classes)\n<\/code><\/pre>\n\n<p>Expected result- The code should predict to which class images belong to.\nCurrent result- Showing error \"File not found\" when I run the code, despite using exact path of the folder.<\/p>\n","answers":[{"AnswerId":"57689231","CreationDate":"2019-08-28T09:32:41.677","ParentId":null,"OwnerUserId":"1029324","Title":null,"Body":"<p>Try to change for loop to os.walk<\/p>\n\n<pre><code>for root, dirs, files in os.walk(folder_path)\n    for filename in files:\n        img= os.path.join(root, filename)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57689041,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-28T09:21:12.097","AcceptedAnswerId":null,"OwnerUserId":9954063.0,"Title":"How to properly initialize all the variables in tensorflow?","Body":"<p>I have a simple neural network class, where I have member methods to create reuse weight variables. in the run member method, I am calling those functions to create weight variables and then initializing them with <code>tf.global_variables_initializer````to initialize all those weight and bias variables. But when I run this program, I am getting an error. When I run it multiple times, the error varies, sometimes its about<\/code>failedpreconditionederror: attempting to use uninitialized variables``` and sometimes the error is at much later steps where i am using the initialized weights to calculate the input of the neuron. Can anyone help me with what is going on with this problem ?<\/p>\n\n<p>I have tried to create an object with <code>tf.Session()<\/code> and used that object to run the initialiser object. <\/p>\n\n<pre><code>def run(self, carollis_input):\n        c_in=normalize(carollis_input, axis = 0)\n        print('normalized carollis input is \\n')\n        print(c_in)\n        c_in = np.array(c_in)\n        c_input = tf.compat.v1.convert_to_tensor(c_in, tf.float64)\n        #'finding the output of the input layer'\n        weight_i = self.input_weight()\n        weight_h = self.hid_weight()\n        weight_o = self.out_weight()\n        bias_i = self.bias_in()\n        bias_h = self.bias_hid()\n        bias_o = self.bias_out()\n        init = tf.global_variables_initializer()\n        sess = tf.compat.v1.Session()\n        sess.run(init)\n        knowledge_input = tf.add(tf.multiply(c_input, weight_i), bias_i)\n        sess.run(knowledge_input)\n        knowledge_hidden = tf.nn.leaky_relu(knowledge_input, alpha=0.01) \n        #'calculating the output of hidden layer'\n        knowledge_hidden_output = 3.14*(tf.add(tf.multiply(knowledge_hidden, weight_h), bias_h))#input function of hidden layer\n        knowledge_hidden_out = tf.nn.leaky_relu(knowledge_hidden_output, alpha=0.01, name='leaky_relu')\n\n        sess.run(knowledge_hidden_out)\n        #'calculating the input of output layer'\n        knowledge_hidden_out = tf.reshape(knowledge_hidden_out, [4, 2])#for quadrant method\n        in_out = tf.add(tf.multiply(knowledge_hidden_out, weight_o), bias_o)\n        sess.run(in_out)\n        #'finding the softmax output of the neurons'\n        softmax_output = np.array(4)\n        softmax_output = self.out_softmax(in_out)  # this gives the softmax output and stores it in the newly created array\n        return softmax_output\n<\/code><\/pre>\n\n<p>The function to intialize variables is as follows :<\/p>\n\n<pre><code>def input_weight(self):\n        with tf.compat.v1.variable_scope(\"weight_in\", reuse=tf.compat.v1.AUTO_REUSE):\n            v = tf.compat.v1.get_variable(\"weight_input\", dtype=tf.float64, shape=[self.neurons, 1], initializer=self.weight_initer)\n        return v\n<\/code><\/pre>\n\n<p>The error sometimes is ,<\/p>\n\n<pre><code>[ERROR] [1566983446.912849, 15.130000]: bad callback: &lt;function joint_callback at 0x7f0722d10510&gt;\nTraceback (most recent call last):\n  File \"\/home\/microbot\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1356, in _do_call\n    return fn(*args)\n  File \"\/home\/microbot\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1341, in _run_fn\n    options, feed_dict, fetch_list, target_list, run_metadata)\n  File \"\/home\/microbot\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1429, in _call_tf_sessionrun\n    run_metadata)\ntensorflow.python.framework.errors_impl.FailedPreconditionError: Attempting to use uninitialized value bias_in\/bias_input\n     [[{{node bias_in\/bias_input\/read}}]]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57689245,"AnswerCount":1,"Tags":"<java><spring-boot><caching><java-8><caffeine>","CreationDate":"2019-08-28T09:33:08.483","AcceptedAnswerId":null,"OwnerUserId":11679316.0,"Title":"How to test inmemory caching using Caffeine cache manager? (i.e. getting the count of entries in cache after caching, cache evict)","Body":"<p>I want to get the details of entries in my custom cache after caching data or eviction of data.<\/p>\n\n<p>I tried using actuator dependency to get 'actuator\/metrics' path to get details but I'm getting empty tomcat server cache. There is no sign of my custom cache say myCache (the name which I passed into <code>@Cacheable<\/code> annotation value argument).<\/p>\n","answers":[{"AnswerId":"57691054","CreationDate":"2019-08-28T11:10:41.813","ParentId":null,"OwnerUserId":"2594725","Title":null,"Body":"<p>You tagged your question with Caffeine and Spring Boot, so I assume you use those two products.<\/p>\n\n<p>If you use a recent Spring Boot and Caffeine, statistics will be automatically available at <code>actuator\/caches<\/code>. If not, double check you have the needed libraries on your classpath and no configuration that enables another cache or disables caching at all, like <code>spring.cache.type=none<\/code>.<\/p>\n\n<p>If you don't use Spring Boot, but just Spring, you need to add a <code>CacheManager<\/code> to your configuration, otherwise Spring defaults to the <code>ConcurrentHashMap<\/code> which does not have cache statistics.<\/p>\n"}]}
+{"QuestionId":57689788,"AnswerCount":1,"Tags":"<tensorflow><machine-learning>","CreationDate":"2019-08-28T10:01:04.313","AcceptedAnswerId":57692272.0,"OwnerUserId":11840960.0,"Title":"Want to know the list of the inputs and outputs along with the shape of a tensorflow model checkpoint","Body":"<p>I have the model checkpoint file along with the <code>.index<\/code> and <code>.meta<\/code> files but i want to know what are the inputs and outputs for the model along with the dimensions and shape for each of them.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/xeekf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/xeekf.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"57692272","CreationDate":"2019-08-28T12:18:59.773","ParentId":null,"OwnerUserId":"1849975","Title":null,"Body":"<p>As mentioned <a href=\"https:\/\/stackoverflow.com\/questions\/55757380\/get-input-and-output-node-name-from-ckpt-and-meta-files-tensorflow\">here<\/a>, the checkpoint save doesn't have strictly defined inputs and outputs (in contrast to the frozen graph). Still, you can inspect your graph with <code>TensorBoard<\/code> and identify input and output tensors:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.summary import FileWriter\n\nsess = tf.Session()\ntf.train.import_meta_graph(\"your-meta-graph-file.meta\")\nFileWriter(\"__tb\", sess.graph)\n<\/code><\/pre>\n\n<p>and inspect it with <code>tensorboard --logdir __tb<\/code>.<\/p>\n"}]}
+{"QuestionId":57690048,"AnswerCount":1,"Tags":"<tensorflow><keras><neural-network><deep-learning><keras-layer>","CreationDate":"2019-08-28T10:14:21.600","AcceptedAnswerId":57693306.0,"OwnerUserId":10227958.0,"Title":"Understanding output of Dense layer for higher dimension","Body":"<p>I don't have problem in understanding output shape of a Dense layer followed by a Flatten layer. Output shape is in accordance of my understanding i.e (Batch size, unit).<\/p>\n\n<pre><code>nn= keras.Sequential()\nnn.add(keras.layers.Conv2D(8,kernel_size=(2,2),input_shape=(4,5,1)))\nnn.add(keras.layers.Conv2D(1,kernel_size=(2,2)))\nnn.add(keras.layers.Flatten())\nnn.add(keras.layers.Dense(5))\nnn.add(keras.layers.Dense(1))\n\nnn.summary()\n<\/code><\/pre>\n\n<p>Output is:<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nconv2d_1 (Conv2D)            (None, 3, 4, 8)           40        \n_________________________________________________________________\nconv2d_2 (Conv2D)            (None, 2, 3, 1)           33        \n_________________________________________________________________\nflatten_1 (Flatten)          (None, 6)                 0         \n_________________________________________________________________\ndense_1 (Dense)              (None, 5)                 35        \n_________________________________________________________________\ndense_2 (Dense)              (None, 1)                 6         \n=================================================================\nTotal params: 114\nTrainable params: 114\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>But I am having trouble in understanding the output shape of a dense layer for multidimensional input .So for following code<\/p>\n\n<pre><code>nn= keras.Sequential()\nnn.add(keras.layers.Conv2D(8,kernel_size=(2,2),input_shape=(4,5,1)))\nnn.add(keras.layers.Conv2D(1,kernel_size=(2,2)))\n#nn.add(keras.layers.Flatten())\nnn.add(keras.layers.Dense(5))\nnn.add(keras.layers.Dense(1))\n\nnn.summary()\n<\/code><\/pre>\n\n<p>output is <\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nconv2d_1 (Conv2D)            (None, 3, 4, 8)           40        \n_________________________________________________________________\nconv2d_2 (Conv2D)            (None, 2, 3, 1)           33        \n_________________________________________________________________\ndense_1 (Dense)              (None, 2, 3, 5)           10        \n_________________________________________________________________\ndense_2 (Dense)              (None, 2, 3, 1)           6         \n=================================================================\nTotal params: 89\nTrainable params: 89\n<\/code><\/pre>\n\n<p>I am unable to make intuition for output shape of <code>dense_1<\/code> and <code>dense_2<\/code> layer. Shouldn't the final output be a scalar or (batch,unit)?\nFollowing <a href=\"https:\/\/stackoverflow.com\/a\/50311243\/10227958\">answer<\/a> to similar question tries to explain the intuition but I can not fully grasp the concept.\nFrom the same <a href=\"https:\/\/stackoverflow.com\/a\/50311243\/10227958\">answer<\/a>:<\/p>\n\n<blockquote>\n  <p>That is, each output \"pixel\" (i, j) in the 640x959 grid is calculated as a dense combination of the 8 different convolution channels at point (i, j) from the previous layer.<\/p>\n<\/blockquote>\n\n<p>May be some explanation with pictures will be useful .<\/p>\n","answers":[{"AnswerId":"57693306","CreationDate":"2019-08-28T13:13:55.740","ParentId":null,"OwnerUserId":"567620","Title":null,"Body":"<p>This is tricky but it does fit with the documentation from Keras on dense layers,<\/p>\n\n<blockquote>\n  <p>Output shape<\/p>\n  \n  <p><code>nD<\/code> tensor with shape: <code>(batch_size, ..., units)<\/code>. For instance, for a 2D input with shape <code>(batch_size, input_dim)<\/code>, the output would have shape <code>(batch_size, units)<\/code><\/p>\n<\/blockquote>\n\n<p>Note it is not the clearest, but they are saying with the <code>...<\/code> that the final dimension of the input shape will be elided by the number of dense connections. Basically, for each item of the final dimension, create a connection to each of the requested dense nodes in the coming dense layer.<\/p>\n\n<p>In your case you have something which is 2 x 3 x 1. So there is \"one thing\" (the 2 x 3 thing) to be connected to each of the 5 dense layer nodes, hense 2 x 3 x 5. You can think of it like channels of a CNN layer in this particular case. There is a distinct 2 x 3 sheet of outputs for each of the 5 output \"nodes\".<\/p>\n\n<p>In a purely 2-D case <code>(batch_size, units)<\/code> ... then each item iterated by the final dimension <code>units<\/code> is itself a scalar value, so you end up with something of exactly the size of the number of dense nodes requested.<\/p>\n\n<p>But in a higher-dimensional case, each item you iterate along the final dimension of the input will itself still be a higher-dimensional thing, and so the output is k distinct \"clones\" of those higher-dimensional things, where k is the dense layer size requested, and by \"clone\" we mean the output for a single dense connection has the same shape as the the items in the final dimension of the input. <\/p>\n\n<p>Then the Dense-ness means that each specific element of that output has a connection to each element of the corresponding set of inputs. But be careful about this. Dense layers are defined by having \"one\" connection between each item of the output and each item of the input. So even though you have 5 \"2x3 things\" in your output, they each just have one solitary weight associated with them about how they are connected to the 2x3 thing that is the input. Keras also defaults to using a bias <em>vector<\/em> (not bias tensor), so if the dense layer has dimension <code>k<\/code> and the final dimension of the previous layer is <code>n<\/code> you should expect <code>(n+1)k<\/code> trainable parameters. These will always be used with numpy-like broadcasting to make the lesser dimensional shape of the weight and bias vectors conformable to the actual shapes of the input tensors. <\/p>\n\n<p>It is customary to use Flatten as in your first example if you want to enforce the exact size of the coming dense layer. You would use multidimensional Dense layer when you want different \"(n - 1)D\" groups of connections to each Dense node. This is probably extremely rare for higher dimensional inputs because you'd typically want a CNN type of operation, but I could imagine maybe in some cases where a model predicts pixel-wise values or if you are generating a full nD output, like from the decoder portion of an encoder-decoder network, you might want a dense array of cells that match the dimensions of some expected structured output type like an image or video. <\/p>\n"}]}
+{"QuestionId":57690122,"AnswerCount":1,"Tags":"<python><caffe><lmdb>","CreationDate":"2019-08-28T10:18:14.303","AcceptedAnswerId":57702975.0,"OwnerUserId":10973854.0,"Title":"how to convert the annotation.xml files into train.txt and val.txt for object detection?","Body":"<p>I'm training custom image data sets for object detection using caffe. Here i have train and validation directory. By using LabelImg , the object annotation created in a .xml file for each image separately (image1.xml,image2.xml etc.). Now i want to convert all this xml files into train.txt and val.txt. How can i do it? I stuck in this step since 3 weeks<\/p>\n\n<p>218.xml<\/p>\n\n<pre><code>`&lt;annotation&gt;\n    &lt;folder&gt;train&lt;\/folder&gt;\n    &lt;filename&gt;218&lt;\/filename&gt;\n    &lt;path&gt;C:\\kp\\images\\train\\218.jpg&lt;\/path&gt;\n    &lt;source&gt;\n        &lt;database&gt;Unknown&lt;\/database&gt;\n    &lt;\/source&gt;\n    &lt;size&gt;\n        &lt;width&gt;352&lt;\/width&gt;\n        &lt;height&gt;288&lt;\/height&gt;\n        &lt;depth&gt;3&lt;\/depth&gt;\n    &lt;\/size&gt;\n    &lt;segmented&gt;0&lt;\/segmented&gt;\n    &lt;object&gt;\n        &lt;name&gt;person&lt;\/name&gt;\n        &lt;pose&gt;Unspecified&lt;\/pose&gt;\n        &lt;truncated&gt;0&lt;\/truncated&gt;\n        &lt;difficult&gt;0&lt;\/difficult&gt;\n        &lt;bndbox&gt;\n            &lt;xmin&gt;79&lt;\/xmin&gt;\n            &lt;ymin&gt;56&lt;\/ymin&gt;\n            &lt;xmax&gt;262&lt;\/xmax&gt;\n            &lt;ymax&gt;283&lt;\/ymax&gt;\n        &lt;\/bndbox&gt;\n    &lt;\/object&gt;\n&lt;\/annotation&gt;`\n<\/code><\/pre>\n\n<p>For each image i have created annotation.xml file (image_name.xml) by using the tool. Now i need to convert it into train.txt and val.txt for input to the create LMDB.  I have already gone through <a href=\"https:\/\/stackoverflow.com\/questions\/45446829\/caffe-lmdb-train-and-val-txt\">Caffe LMDB train and val.txt<\/a> , <a href=\"https:\/\/stackoverflow.com\/questions\/41586449\/how-to-make-train-txt-file-in-caffe\">how to make train.txt file in caffe<\/a> , <a href=\"https:\/\/stackoverflow.com\/questions\/31427094\/a-guide-to-convert-imageset-cpp\/31431716#31431716\">A guide to convert_imageset.cpp<\/a> , <a href=\"https:\/\/github.com\/chuanqi305\/MobileNet-SSD\/issues\/90\" rel=\"nofollow noreferrer\">https:\/\/github.com\/chuanqi305\/MobileNet-SSD\/issues\/90<\/a><\/p>\n\n<p>But everything is for object classification and and others didn't explain it too.. Any suggestions or help will be appreciated<\/p>\n","answers":[{"AnswerId":"57702975","CreationDate":"2019-08-29T04:27:05.673","ParentId":null,"OwnerUserId":"11976900","Title":null,"Body":"<p>train.txt<\/p>\n\n<p>imagepath xmlpath<\/p>\n\n<pre><code>image\/image1.jpg label\/image1.xml\nimage\/image2.jpg label\/image2.xml\nimage\/image3.jpg label\/image3.xml\nimage\/image4.jpg label\/image4.xml\n...\n<\/code><\/pre>\n\n<p>this is train.txt or test.txt<\/p>\n"}]}
+{"QuestionId":57690495,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-estimator><distributed-tensorflow>","CreationDate":"2019-08-28T10:37:57.527","AcceptedAnswerId":null,"OwnerUserId":6098261.0,"Title":"Tensorflow Distributed training is not evaluating the model properly","Body":"<p><strong>I'm running asynchronous distributed training in Tensorflow using Parameter server strategy. Multi worker on multiple CPUs with evaluator as a separate node.<\/strong><\/p>\n\n<p><strong>Sample tf_config of Parameter server:<\/strong> Index and type varies for chief, worker and evaluator on other TF_CONFIGs.<\/p>\n\n<pre><code>TF_CONFIG={\n\"task\": {\n    \"type\": \"ps\",\n    \"index\": 0\n},\n\"cluster\": {\n    \"chief\": [\"machine2:2222\"],\n    \"worker\": [\"machine3:2223\",\"machine4:2224\"],\n    \"evaluator\": [\"machine5:2225\"],\n    \"ps\": [\"machine1:2218\"]\n}\n<\/code><\/pre>\n\n<p>}<\/p>\n\n<pre><code>def main(unused_argv):\nflags.mark_flag_as_required('model_dir')\nflags.mark_flag_as_required('pipeline_config_path')\nconfig = tf.estimator.RunConfig(model_dir=FLAGS.model_dir)\n\ntrain_and_eval_dict = model_lib.create_estimator_and_inputs(\n    run_config=config,\n    hparams=model_hparams.create_hparams(FLAGS.hparams_overrides),\n    pipeline_config_path=FLAGS.pipeline_config_path,\n    train_steps=FLAGS.num_train_steps,\n    sample_1_of_n_eval_examples=FLAGS.sample_1_of_n_eval_examples,\n    sample_1_of_n_eval_on_train_examples=(\n        FLAGS.sample_1_of_n_eval_on_train_examples))\nestimator = train_and_eval_dict['estimator']\ntrain_input_fn = train_and_eval_dict['train_input_fn']\neval_input_fns = train_and_eval_dict['eval_input_fns']\neval_on_train_input_fn = train_and_eval_dict['eval_on_train_input_fn']\npredict_input_fn = train_and_eval_dict['predict_input_fn']\ntrain_steps = train_and_eval_dict['train_steps']\n\nif FLAGS.checkpoint_dir:\n    if FLAGS.eval_training_data:\n        name = 'training_data'\n        input_fn = eval_on_train_input_fn\n    else:\n        name = 'validation_data'\n\n    # The first eval input will be evaluated.\n    input_fn = eval_input_fns[0]\n\n    if FLAGS.run_once:\n        estimator.evaluate(input_fn,\n            num_eval_steps=None,\n            checkpoint_path=tf.train.latest_checkpoint(\n                FLAGS.checkpoint_dir))\n    else:\n        model_lib.continuous_eval(estimator, FLAGS.checkpoint_dir, input_fn, train_steps, name)\nelse:\n    train_spec, eval_specs = model_lib.create_train_and_eval_specs(\n        train_input_fn,\n        eval_input_fns,\n        eval_on_train_input_fn,\n        predict_input_fn,\n        train_steps,\n        eval_on_train_data=False)\n\n  # Currently only a single Eval Spec is allowed.\n  tf.estimator.train_and_evaluate(estimator, train_spec, eval_specs[0])\n\n\nif __name__ == '__main__':\n    tf.app.run()\n<\/code><\/pre>\n\n<p><strong>I'm getting multiple warnings:<\/strong><\/p>\n\n<blockquote>\n  <p>W0828 00:03:55.229441 140490069309248 estimator.py:1924] Estimator's model_fn (.model_fn at 0x7fc5da9b5268>) includes params argument, but params are not passed to Estimator.<\/p>\n  \n  <p>.\/tensorflow\/core\/grappler\/optimizers\/graph_optimizer_stage.h:241] Failed to run optimizer ArithmeticOptimizer, stage RemoveStackStridedSliceSameAxis node Preprocessor\/ResizeToRange\/strided_slice_3. Error: Pack node (Preprocessor\/ResizeToRange\/stack_2) axis attribute is out of bounds: 0<\/p>\n<\/blockquote>\n\n<hr>\n\n<p><strong>But the training runs fine and evaluation happens. But my evaluation results are 0 all the time.<\/strong><\/p>\n\n<blockquote>\n  <p>creating index...\n  index created!\n  creating index...\n  index created!\n  Running per image evaluation...\n  Evaluate annotation type <em>bbox<\/em>\n  DONE (t=1.66s).\n  Accumulating evaluation results...\n  DONE (t=0.52s).<\/p>\n  \n  <p>Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.000<\/p>\n  \n  <p>Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.000<\/p>\n  \n  <p>Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.000<\/p>\n  \n  <p>Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000<\/p>\n  \n  <p>Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.000<\/p>\n  \n  <p>Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.000<\/p>\n  \n  <p>Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.000<\/p>\n  \n  <p>Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.000<\/p>\n  \n  <p>Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.000<\/p>\n  \n  <p>Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000<\/p>\n  \n  <p>Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.000<\/p>\n  \n  <p>Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.000<\/p>\n<\/blockquote>\n\n<p>Any help would be much appreciated. Thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":57690658,"AnswerCount":1,"Tags":"<machine-learning><keras><neural-network><deep-learning><conv-neural-network>","CreationDate":"2019-08-28T10:47:15.893","AcceptedAnswerId":57691643.0,"OwnerUserId":2348704.0,"Title":"Keras target dimensions mismatch","Body":"<p>Attempting a single-label classification problem with <code>num_classes = 73<\/code><\/p>\n\n<p>Here's my simplified Keras model:<\/p>\n\n<pre><code>num_classes = 73\nbatch_size = 4\n\ntrain_data_list = [training_file_names list here..]\nvalidation_data_list = [ validation_file_names list here..]\n\ntraining_generator = DataGenerator(train_data_list, batch_size, num_classes)\nvalidation_generator = DataGenerator(validation_data_list, batch_size, num_classes)\n\nmodel = Sequential()\nmodel.add(Conv1D(32, 3, strides=1, input_shape=(15,120), activation=\"relu\"))\nmodel.add(Conv1D(16, 3, strides=1, activation=\"relu\"))\nmodel.add(Flatten())\nmodel.add(Dense(n_classes, activation='softmax'))\n\nsgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)\nmodel.compile(loss=\"categorical_crossentropy\",optimizer=sgd,metrics=['accuracy'])\n\nmodel.fit_generator(generator=training_generator, epochs=100,\n                    validation_data=validation_generator)\n<\/code><\/pre>\n\n<p>Here's my <code>DataGenerator<\/code>'s <code>__get_item__<\/code> method:<\/p>\n\n<pre><code>def __get_item__(self):\n    X = np.zeros((self.batch_size,15,120))\n    y = np.zeros((self.batch_size, 1 ,self.n_classes))\n    for i in range(self.batch_size):\n      X_row = some_method_that_gives_X_of_15x20_dim()   \n      target = some_method_that_gives_target()    \n      one_hot = keras.utils.to_categorical(target, num_classes=self.n_classes)\n      X[i] = X_row\n      y[i] = one_hot\n    return X, y\n<\/code><\/pre>\n\n<p>Since my <code>X<\/code> values are correctly returned with dimension <code>(batch_size, 15, 120)<\/code>, I am not showing it here. My issue is with the y value returned.<\/p>\n\n<p><code>y<\/code> returned from this generator method has a shape of <code>(batch_size, 1, 73)<\/code> as one hot encoded label for the 73 classes, which I think is the correct shape to return.<\/p>\n\n<p>However Keras gives the following error for the last layer:<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking target: expected dense_1 to have 2\n  dimensions, but got array with shape (4, 1, 73)<\/p>\n<\/blockquote>\n\n<p>Since the batch size is 4, I think the target batch should also be 3 dimensional (4,1,73). Why is then Keras expecting the last layer to be 2 dimensions ?<\/p>\n","answers":[{"AnswerId":"57691643","CreationDate":"2019-08-28T11:44:04.153","ParentId":null,"OwnerUserId":"8563649","Title":null,"Body":"<p>you model' s summary shows that in the output layer there should be only 2 dimensions, (None, 73) <\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nconv1d_7 (Conv1D)            (None, 13, 32)            11552     \n_________________________________________________________________\nconv1d_8 (Conv1D)            (None, 11, 16)            1552      \n_________________________________________________________________\nflatten_5 (Flatten)          (None, 176)               0         \n_________________________________________________________________\ndense_4 (Dense)              (None, 73)                12921     \n=================================================================\nTotal params: 26,025\nTrainable params: 26,025\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>Since dimension of your target is (batch_size, 1, 73), you can just change to (batch_size, 73) in order for your model to run<\/p>\n"}]}
+{"QuestionId":57690858,"AnswerCount":2,"Tags":"<python><tensorflow><keras><loss-function>","CreationDate":"2019-08-28T10:57:53.660","AcceptedAnswerId":57709479.0,"OwnerUserId":7131919.0,"Title":"Multiple losses in Tensorflow and Keras","Body":"<p>In my simple Variational Autoencoder code, I want to see both reconstruction and KL divergance loss values when the model is running. I used the <a href=\"https:\/\/keras.io\/examples\/variational_autoencoder\/\" rel=\"nofollow noreferrer\">Example of VAE on MNIST dataset using MLP<\/a> in Keras documentation page, and modified the loss part as below:<\/p>\n\n<pre><code>reconstruction_loss = original_dim*binary_crossentropy(inputs,outputs)                                              \nkl_loss = -0.5*K.sum(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=-1)\nvae_loss = K.mean(reconstruction_loss + kl_loss)\n\n\n\nlosses = {  'recon_loss': reconstruction_loss,\n             'latent_loss': kl_loss,\n             'total_loss': vae_loss,}\n\n\nvae.add_loss(losses)\nvae.compile(optimizer='adam')\nhistory = vae.fit(x_train, epochs=epochs, batch_size=batch_size,\n                validation_data=(x_test, None))\n<\/code><\/pre>\n\n<p>But when I run the model, the <strong>Compile<\/strong> command returns this error message:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-21-b668ab0f8437&gt; in &lt;module&gt;\n----&gt; 1 vae.compile(optimizer='adam')\n      2 \n\n~\\Anaconda3\\lib\\site-packages\\keras\\engine\\training.py in compile(self, optimizer, loss, metrics, loss_weights, sample_weight_mode, weighted_metrics, target_tensors, **kwargs)\n    349                     total_loss += loss_weight * output_loss\n    350             if total_loss is None:\n--&gt; 351                 if not self.losses:\n    352                     raise ValueError('The model cannot be compiled '\n    353                                      'because it has no loss to optimize.')\n\n~\\Anaconda3\\lib\\site-packages\\keras\\engine\\network.py in losses(self)\n    423 \n    424         unique_tensors = list(\n--&gt; 425             set(x for x in losses if not isinstance(x, (float, int))))\n    426         non_tensors = [x for x in losses if isinstance(x, (float, int))]\n    427         return unique_tensors + non_tensors\n\nTypeError: unhashable type: 'dict'\n<\/code><\/pre>\n\n<p>I have seen some other questions about <strong><em>TypeError: unhashable type: 'dict'<\/em><\/strong>, but I couldn't solve my problem. Are there any idea to adopt the Compile command in Keras and work with multiple losses?<\/p>\n","answers":[{"AnswerId":"57690897","CreationDate":"2019-08-28T11:00:07.577","ParentId":null,"OwnerUserId":"9374613","Title":null,"Body":"<p>Please check the type of each loss in your losses dictionary.<\/p>\n\n<pre><code>print (type(losses['recon_loss']))\n<\/code><\/pre>\n"},{"AnswerId":"57709479","CreationDate":"2019-08-29T11:44:45.857","ParentId":null,"OwnerUserId":"7131919","Title":null,"Body":"<p>I modified the loss functions and used the <em>metrics<\/em> in compile finction.<\/p>\n\n<pre><code>def recon_loss(inputs,outputs):\n    reconstruction_loss = original_dim*binary_crossentropy(inputs,outputs)                                              \n    return K.mean(reconstruction_loss)\n\ndef latent_loss(inputs,outputs):\n    kl_loss =  -0.5*K.sum(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=-1)    \n    return K.mean(kl_loss)\n\ndef total_loss(inputs,outputs):\n    reconstruction_loss = original_dim*binary_crossentropy(inputs,outputs)                                              \n    kl_loss =  -0.5*K.sum(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=-1)   \n    return K.mean(reconstruction_loss + kl_loss)\nvae.compile(optimizer='adam',loss=total_loss,metrics=[recon_loss, latent_loss])\n<\/code><\/pre>\n\n<p>Now, the model returns reconstruction, latent and total losses for both training and validation data sets.<\/p>\n"}]}
+{"QuestionId":57690907,"AnswerCount":0,"Tags":"<tensorflow><tensorboard>","CreationDate":"2019-08-28T11:00:46.153","AcceptedAnswerId":null,"OwnerUserId":3565061.0,"Title":"Remove graph from existing TensorFlow\/TensorBoard log files","Body":"<p>I trained and did validation on my checkpoints which resulted in about 50000 logs. For training, I took care of the graph in my logs in such a way that only few logs actually contain the graph, while during validation, I accidentally added the graph to more or less every single log.<\/p>\n\n<p>While the logs without a graph have around 10kb, those with a graph can have up to 8mb, resulting in 1gb of logs for training, and almost 80gb for validation. <\/p>\n\n<p>This not only takes up way to much space, this also makes it super slow to load it into tensorboard and causes issues with my limited RAM.<\/p>\n\n<p>Is there a way to remove the graph from my log files, after they've been already created?<\/p>\n","answers":[]}
+{"QuestionId":57691341,"AnswerCount":0,"Tags":"<tensorflow><object-detection>","CreationDate":"2019-08-28T11:27:34.697","AcceptedAnswerId":null,"OwnerUserId":9010109.0,"Title":"Add new classifier to existing tensorflow models","Body":"<p>I want to customize existing tensorflow object detection model ssd mobilenetv2 to solve multitask problem. How i can get copy output tensors in feature extractor and use it for my own classifier input layers.<\/p>\n","answers":[]}
+{"QuestionId":57692001,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-08-28T12:04:21.770","AcceptedAnswerId":null,"OwnerUserId":11988902.0,"Title":"How do we clip the derived variables so that values of trainable values are also clipped accordingly","Body":"<p>When I create a variable (not used in calculating predicted output) that is being derived from the trainable variables and apply a clip-on that, are the actual trainable variables also getting clipped according to the values that the derived variables can take or not? If not, is there a way to do that.<\/p>\n\n<p>If you take a simple linear regression example for the ease of explanation:<\/p>\n\n<p>Y = Ax + B (where A and B are the trainable variables and x is our input)<\/p>\n\n<p>if we create a variable:<\/p>\n\n<pre><code>C = tf.Variable(B\/A, dtype ='float64')\n<\/code><\/pre>\n\n<p>and apply a clip on C:<\/p>\n\n<pre><code>C = tf.clip_by_value(C, c_min, c_max)\n<\/code><\/pre>\n\n<p>And now optimize the mean_square_error of the predicted Y and actual Y on a set of inputs, will we get B and A such that it satisfies values that C can take? \nIf not, is there a way to do that?<\/p>\n\n<p>In my case I have complex derived variables whose ratio of sums are to be clipped but if I have the answer to the above problem I think I will have the foundation and will be able to build on that.<\/p>\n","answers":[]}
+{"QuestionId":57692152,"AnswerCount":1,"Tags":"<python><tensorflow><keras><convolution>","CreationDate":"2019-08-28T12:12:42.480","AcceptedAnswerId":null,"OwnerUserId":8447424.0,"Title":"How to perform element convolution between two tensors?","Body":"<p>I have two tensors, both with batch size N of images and same resolution. I would like to convolve the first image in tensor 1 with the first image of tensor 2, second image of tensor 1 with tensor 2, and so on. I want the output to be a tensor with N images of the same size. \nI looked into using tf.nn.conv2d, but it seems like this command will take in a batch of N images and convolve them with a single filter.<\/p>\n\n<p>I looked into examples like <a href=\"https:\/\/stackoverflow.com\/questions\/34619177\/what-does-tf-nn-conv2d-do-in-tensorflow?answertab=active#tab-top\">What does tf.nn.conv2d do in tensorflow?<\/a>\nbut they do not talk about multiple images and multiple filters.<\/p>\n","answers":[{"AnswerId":"57707273","CreationDate":"2019-08-29T09:36:09.970","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You can manage to do something like that using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/separable_conv2d\" rel=\"nofollow noreferrer\"><code>tf.nn.separable_conv2d<\/code><\/a>, using the batch dimension as the separable channels and the actual input channels as batch dimension. I am not sure if it is going to be perform very well, though, as it involves several transpositions (which are not free in TensorFlow) and a convolution through a large number of channels, which is not really the optimized use case. Here is how it could work:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport scipy.signal\n\n# Expects imgs with shape (B, H, W, C) and filters with shape (B, H, W, 1)\ndef batch_conv(imgs, filters, strides, padding, rate=None):\n    imgs = tf.convert_to_tensor(imgs)\n    filters = tf.convert_to_tensor(filters)\n    b = tf.shape(imgs)[0]\n    imgs_t = tf.transpose(imgs, [3, 1, 2, 0])\n    filters_t = tf.transpose(filters, [1, 2, 0, 3])\n    strides = [strides[3], strides[1], strides[2], strides[0]]\n    # \"do-nothing\" pointwise filter\n    pointwise = tf.eye(b, batch_shape=[1, 1])\n    conv = tf.nn.separable_conv2d(imgs_t, filters_t, pointwise, strides, padding, rate)\n    return tf.transpose(conv, [3, 1, 2, 0])\n\n# Slow, loop-based version using SciPy's correlate to check result\ndef batch_conv_np(imgs, filters, padding):\n    return np.stack(\n        [np.stack([scipy.signal.correlate2d(img[..., i], filter[..., 0], padding.lower())\n                   for i in range(img.shape[-1])], axis=-1)\n         for img, filter in zip(imgs, filters)], axis=0)\n\n# Make random input\nnp.random.seed(0)\nimgs = np.random.rand(5, 20, 30, 3).astype(np.float32)\nfilters = np.random.rand(5, 20, 30, 1).astype(np.float32)\npadding = 'SAME'\n# Test\nres_np = batch_conv_np(imgs, filters, padding)\nwith tf.Graph().as_default(), tf.Session() as sess:\n    res_tf = batch_conv(imgs, filters, [1, 1, 1, 1], padding)\n    res_tf_val = sess.run(res_tf)\nprint(np.allclose(res_np, res_tf_val))\n# True\n<\/code><\/pre>\n"}]}
+{"QuestionId":57692323,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-28T12:22:10.703","AcceptedAnswerId":57693026.0,"OwnerUserId":519882.0,"Title":"non trainable parameters params in keras model is calculated","Body":"<p>I have following program taken from Internet<\/p>\n\n<pre><code>def my_model(input_shape):\n    # Define the input placeholder as a tensor with shape input_shape. Think of this as your input image!\n    X_input = Input(input_shape)\n\n    # Zero-Padding: pads the border of X_input with zeroes\n    X = ZeroPadding2D((3, 3))(X_input)\n\n    # CONV -&gt; BN -&gt; RELU Block applied to X\n    X = Conv2D(32, (7, 7), strides = (1, 1), name = 'conv0')(X)\n    X = BatchNormalization(axis = 3, name = 'bn0')(X)\n    X = Activation('relu')(X)\n\n    # MAXPOOL\n    X = MaxPooling2D((2, 2), name='max_pool')(X)\n\n    # FLATTEN X (means convert it to a vector) + FULLYCONNECTED\n    X = Flatten()(X)\n    X = Dense(1, activation='sigmoid', name='fc')(X)\n\n    # Create model. This creates your Keras model instance, you'll use this instance to train\/test the model.\n    model = Model(inputs = X_input, outputs = X, name='myModel')\n\n    return model\n\nmymodel = my_model((64,64,3))\nmymodel.summary()\n<\/code><\/pre>\n\n<p>Here output of summary is shown as below<\/p>\n\n<pre><code>Layer (type)                 Output Shape              Param #   \n=================================================================\ninput_3 (InputLayer)         (None, 64, 64, 3)         0         \n_________________________________________________________________\nzero_padding2d_3 (ZeroPaddin (None, 70, 70, 3)         0         \n_________________________________________________________________\nconv0 (Conv2D)               (None, 64, 64, 32)        4736      \n_________________________________________________________________\nbn0 (BatchNormalization)     (None, 64, 64, 32)        128       \n_________________________________________________________________\nactivation_2 (Activation)    (None, 64, 64, 32)        0         \n_________________________________________________________________\nmax_pool (MaxPooling2D)      (None, 32, 32, 32)        0         \n_________________________________________________________________\nflatten_2 (Flatten)          (None, 32768)             0         \n_________________________________________________________________\nfc (Dense)                   (None, 1)                 32769     \n=================================================================\nTotal params: 37,633\nTrainable params: 37,569\nNon-trainable params: 64\n<\/code><\/pre>\n\n<hr>\n\n<p>My question is from which layer this non-trainable params are taken i.e., 64. \nAnother question is how batch normalization has parameters 128?<\/p>\n\n<p>Request your help how above numbers we got from model defined above. Thanks for the time and help.<\/p>\n","answers":[{"AnswerId":"57693026","CreationDate":"2019-08-28T12:59:33.060","ParentId":null,"OwnerUserId":"11669446","Title":null,"Body":"<p><code>BatchNormalization<\/code> layer is composed of <code>[gamma weights, beta weights, moving_mean(non-trainable), moving_variance(non-trainable)]<\/code> and for each parameter there is one value for each element in the last axis (by default in keras, but you can change the axis if you want to). <\/p>\n\n<p>In your code you have a size 32 in the last dimension before the BatchNormalization layer, so <strong>32*4=128 parameters<\/strong> and since there are <strong>2 non-trainable parameters there are 32*2=64 non-trainable parameters<\/strong><\/p>\n"}]}
+{"QuestionId":57692934,"AnswerCount":0,"Tags":"<tensorflow><keras><loss-function>","CreationDate":"2019-08-28T12:54:07.903","AcceptedAnswerId":null,"OwnerUserId":1619428.0,"Title":"Can I implement a multilabel loss function that selectively ignores coordinates per sample, and how will backpropagation react to this?","Body":"<p>I'm trying to write a multilabel loss function (in keras) which is based on binary crossentropy, but for each sample I attach an additional vector specifying which coordinates to ignore for that sample (see code below). <em>i.e.<\/em>, <strong>y_true<\/strong> is a 2d array of dimensions (batch_size, label_dim * 2). For a given sample <strong>sample = y_true[0]<\/strong> for example, <strong>sample[:y_split_point]<\/strong> is a vector with the true labels (0 or 1 per coordinate), while <strong>sample[:y_split_point]<\/strong> tells us which respective coordinate in said vector should count and which should be ignored. In other words, if <strong>sample[y_split_point + i] == 0<\/strong>, we should ignore the label in <strong>sample[i]<\/strong> as it is unreliable. <\/p>\n\n<p>I assumed that since no gradient can backpropagate for zeroed-out coordinates in a sample, the model would converge \"normally\" (without the special loss function, the data is noisy on those coordinates but the model does converge). In practice, I get a model that predicts 0 on everything. Is there something wrong in my understanding of keras backpropagation?<\/p>\n\n<pre><code>def weighted_loss(self, loss_function=keras.losses.binary_crossentropy,\n                  y_split_point):   \n\n    def weighted_loss_in_output(y_true_batch, y_pred_batch):\n        y_true_labels = y_true_batch[:, :y_split_point]\n        y_weights = y_true_batch[:, y_split_point:]\n\n        y_weighted_pred = y_pred_batch * y_weights\n        y_weighted_true = y_true_labels * y_weights\n        loss = loss_function(y_weighted_true, y_weighted_pred)\n\n        return loss\n\n    return weighted_loss_in_output\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57692974,"AnswerCount":1,"Tags":"<tensorflow><keras><recurrent-neural-network>","CreationDate":"2019-08-28T12:56:48.073","AcceptedAnswerId":null,"OwnerUserId":1673372.0,"Title":"Use pretrained model and create another sequential model","Body":"<p>I have a pre-trained model, as below<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/5yGQN.png\" rel=\"nofollow noreferrer\">Image 1<\/a><\/p>\n\n<p>Code the above model is as below,which is working as expected and I have saved this model separately in \".h5\" keras format <\/p>\n\n<pre><code>data= pd.read_csv('TrainPWM.csv')\ndata.values[:,0]\nactual_pwm = data.values[:,0]\nactual_rpm = data.values[:,1]\nactual_pwm = np.reshape(actual_pwm,(36954,1,1))\/10000\nactual_rpm = np.reshape(actual_rpm,(36954,1,1))\/10000\nx_train,x_test,y_train,y_test=train_test_split(actual_pwm,actual_rpm,test_size=0.3,random_state=50)\nmodel = models.Sequential()\nmodel.add(layers.LSTM(units=1,return_sequences=True))\nmodel.add(layers.LSTM(units=20,return_sequences=True))\nmodel.add(layers.LSTM(units=20,return_sequences=True))\nmodel.add(layers.LSTM(units=1,return_sequences=True))\nmodel.add(layers.Dense(units=1))\nmodel.compile(optimizer='adam',\n          loss='mean_absolute_error',\n          metrics=['accuracy'])\nmodel.fit(x_train,y_train,epochs=1)\n<\/code><\/pre>\n\n<p>Now I want to create another sequential model, which must be trained with the above pretrained model(rnn2), below some explanation<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/WBpAu.png\" rel=\"nofollow noreferrer\">Image2<\/a><\/p>\n\n<p>I must create a sequential training model based on this where,When I give input x, first rnn1 must train itself based on the output y using rnn2<\/p>\n\n<p>I have trouble visualizing the way to do this<\/p>\n","answers":[{"AnswerId":"57697628","CreationDate":"2019-08-28T17:46:04.330","ParentId":null,"OwnerUserId":"2874139","Title":null,"Body":"<p>Assuming that you saved your model weights like <code>model.save('my_model.h5')<\/code>. <\/p>\n\n<p>Then you can construct your second model and embed the first model in there:<\/p>\n\n<pre><code>new_model = models.Sequential()\nnew_model.add(...)  # Some layers\n\n# Here you add the model you previously trained\n# In case you need to use another head, use model.layers[-1].output\nconvnet = KerasModel(inputs=model.inputs, outputs=model.output)\nconvnet.set_weights(WEIGHTS_FILE)  # The path to the file where you saved your model\nconvnet.trainable = False  # Lock in its weights\nnew_model.add(convnet)\n\n# Add a new head if you got rid of it\n<\/code><\/pre>\n"}]}
+{"QuestionId":57693124,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-28T13:04:16.273","AcceptedAnswerId":null,"OwnerUserId":10957845.0,"Title":"Unable to get keras flow_from_directory to work with .gif data","Body":"<p>I am trying to build an Image classifier with keras and tensorflow. However currently <code>flow_from_directory<\/code> does not see my images due to them being in .gif format (I checked this with .jpg and it works here). How can I fix this?<\/p>\n\n<p>This old <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/3946\" rel=\"nofollow noreferrer\">github page<\/a> claims that I should be able to put .gif on my <code>white_list_formats<\/code> in the keras\/preprocessing\/image.py file. But after opening it there seems to be no <code>white_list_formats<\/code> in my version of image.py. Did keras change anything here?<\/p>\n\n<p>I'm on Windows using anaconda3 distribution in case that matters.<\/p>\n\n<p>Thanks for your help!<\/p>\n","answers":[{"AnswerId":"58556227","CreationDate":"2019-10-25T09:55:16.563","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>The github post you are referring to was posted for keras version 1.1.0. In the latest version they have removed that variable. <\/p>\n\n<p>In the latest version of Keras, below formats are only supported.<\/p>\n\n<p><strong>PNG, JPG, BMP, PPM, TIF<\/strong><\/p>\n\n<p>Please refer to the <a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"nofollow noreferrer\">this<\/a> for documentation.<\/p>\n\n<p>If you still want to try and edit white_list_format, please install keras 1.1.0 and try it.<\/p>\n"}]}
+{"QuestionId":57693214,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-08-28T13:09:08.610","AcceptedAnswerId":null,"OwnerUserId":9561241.0,"Title":"How to decide convolution layers' filter?","Body":"<p>When you guys make a convolution layer includes hidden layers, how to decide parameters? like filter, stride and even the number of convolution layers? I know the meaning of each parameter, but if I have to make from the start, how can I? <\/p>\n\n<p>Thanks in advanced.<\/p>\n","answers":[{"AnswerId":"58555931","CreationDate":"2019-10-25T09:37:12.670","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>Please refer to below links to have better understanding of CNNs and how to make use of them.<\/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:\/\/medium.com\/@RaghavPrabhu\/understanding-of-convolutional-neural-network-cnn-deep-learning-99760835f148\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@RaghavPrabhu\/understanding-of-convolutional-neural-network-cnn-deep-learning-99760835f148<\/a><\/p>\n\n<p><a href=\"https:\/\/www.analyticsvidhya.com\/blog\/2018\/12\/guide-convolutional-neural-network-cnn\/\" rel=\"nofollow noreferrer\">https:\/\/www.analyticsvidhya.com\/blog\/2018\/12\/guide-convolutional-neural-network-cnn\/<\/a><\/p>\n\n<p><a href=\"https:\/\/towardsdatascience.com\/deciding-optimal-filter-size-for-cnns-d6f7b56f9363\" rel=\"nofollow noreferrer\">https:\/\/towardsdatascience.com\/deciding-optimal-filter-size-for-cnns-d6f7b56f9363<\/a><\/p>\n"}]}
+{"QuestionId":57693333,"AnswerCount":3,"Tags":"<keras><scikit-learn><nlp><tokenize><train-test-split>","CreationDate":"2019-08-28T13:15:21.420","AcceptedAnswerId":57693550.0,"OwnerUserId":139150.0,"Title":"processing before or after train test split","Body":"<p>I am using this excellent article to learn Machine learning.<\/p>\n\n<p><a href=\"https:\/\/stackabuse.com\/python-for-nlp-multi-label-text-classification-with-keras\/\" rel=\"nofollow noreferrer\">https:\/\/stackabuse.com\/python-for-nlp-multi-label-text-classification-with-keras\/<\/a><\/p>\n\n<p>The author has tokenized the X and y data after splitting it up.<\/p>\n\n<pre><code>X_train, X_test, y_train, y_test = train_test_split(\n    X, y, test_size=0.20, random_state=42\n)\n\ntokenizer = Tokenizer(num_words=5000)\ntokenizer.fit_on_texts(X_train)\n\nX_train = tokenizer.texts_to_sequences(X_train)\nX_test = tokenizer.texts_to_sequences(X_test)\n\nvocab_size = len(tokenizer.word_index) + 1\n\nmaxlen = 200\n\nX_train = pad_sequences(X_train, padding=\"post\", maxlen=maxlen)\nX_test = pad_sequences(X_test, padding=\"post\", maxlen=maxlen)\n<\/code><\/pre>\n\n<p>If I tokenize it before using train_test_split class, I can save a few lines of code.<\/p>\n\n<pre><code>tokenizer = Tokenizer(num_words=5000)\ntokenizer.fit_on_texts(X)\n\nX_t = tokenizer.texts_to_sequences(X)\nvocab_size = len(tokenizer.word_index) + 1\nmaxlen = 200\n\nX = pad_sequences(X_t, padding=\"post\", maxlen=maxlen)\n<\/code><\/pre>\n\n<p>I just want to confirm that my approach is correct and I do not expect any surprises later in the script.<\/p>\n","answers":[{"AnswerId":"57693733","CreationDate":"2019-08-28T13:34:31.080","ParentId":null,"OwnerUserId":"5268235","Title":null,"Body":"<p>Agreed with @desertnaut's comment that the question is better suited for \"<a href=\"https:\/\/stats.stackexchange.com\/help\/on-topic\">Cross Validated<\/a>\", you'll get a better response there. But I'd still like to make a remark.<\/p>\n\n<blockquote>\n  <p><strong><em>TL;DR:<\/em><\/strong> Don't do it, it's generally not a good idea to cross contaminate your training and test set. It's not statistically correct to do so.<\/p>\n<\/blockquote>\n\n<p>The <code>Tokenizer.fit_to_texts(dictionary)<\/code> does the word indexing, i.e. it builds a translation of your any sequence of words to numbers(vector representation), so it might be that vocabulary difference between the training and test set is not a <strong>null set<\/strong>, i.e. some of the words in test are not present in the word indexer built by the <code>Tokenizer<\/code> object if it used only the train data. Which could result in some test set generating different vector if you'd have <strong>trained<\/strong> your tokenizer only on training set.<\/p>\n\n<p>Since the test sets in a learning problem are supposed to be hidden, using it during any process of training the model is statistically not correct.<\/p>\n"},{"AnswerId":"57693615","CreationDate":"2019-08-28T13:28:39.853","ParentId":null,"OwnerUserId":"10232932","Title":null,"Body":"<p>To add to Simons post, I would say it is even forbidden to tokenize before the splitting. <\/p>\n\n<p>The algorithm would learn  with data from the tokenizer, which is strictly for testing the algorithm. This is the main approach between train and test set.<\/p>\n"},{"AnswerId":"57693550","CreationDate":"2019-08-28T13:25:23.910","ParentId":null,"OwnerUserId":"11669446","Title":null,"Body":"<p>Both approaches will work in practice. But fitting the tokenizer on the train set and the applied it to both train and test set is better than fitting on the whole dataset. Indeed with the first method you are mimicking the fact that unseen words by the model will appear at some point after deploying your model. Thus your model evaluation will be closer to what will happen in a production environnement.<\/p>\n"}]}
+{"QuestionId":57693793,"AnswerCount":1,"Tags":"<python><tensorflow><optimization><loss-function>","CreationDate":"2019-08-28T13:38:01.220","AcceptedAnswerId":null,"OwnerUserId":1860372.0,"Title":"Optimizing a loss function that ignores certain label values","Body":"<p>I'm writing a binary classifier for sparse input data, and I would like to regard inputs of 0 as an indication that data does not exist rather than an indication that the value is affirmatively 0. I was originally using <code>tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(...))<\/code>, but it penalized false positives too harshly.<\/p>\n\n<p>I successfully wrote a loss function, shown below, that provides the behavior I would like, but it is multiple orders of magnitude slower and I need to find a way to steal some performance back.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def loss(labels, logits):\n    labels = tf.reshape(labels, shape=(-1,))\n    logits = tf.reshape(logits, shape=(-1,))\n    pairs = tf.stack([labels, logits], axis=1)\n\n    return tf.reduce_mean(tf.map_fn(\n        lambda x: tf.cond(\n            x[0] &lt; x[1],  # x[0] is in {0,1} and x[1] is in (0,1)\n            lambda: 0.0,  # thus the condition is true iff x[0] == 0\n            lambda: tf.nn.sigmoid_cross_entropy_with_logits(\n                labels=x[0],\n                logits=x[1])),\n        pairs))\n<\/code><\/pre>\n","answers":[{"AnswerId":"57694327","CreationDate":"2019-08-28T14:06:39.633","ParentId":null,"OwnerUserId":"1860372","Title":null,"Body":"<p>I believe I've come up with a solution. It occurred to me that I didn't actually need to normalize against the number of labels to get the mean rather than just the sum, so created a <code>{0,1}<\/code> mask out of the condition by running the comparison and casting the result to a float. I then used matrix multiplication to take the dot product of the <code>tf.nn.sigmoid_cross_entropy_with_logits(...)<\/code> result with this mask, so that values where the condition is true are added to the sum and values where the condition is false are canceled by multiplying them with zero.<\/p>\n\n<p>This may not be the best solution, and I'll be leaving the question open for a while in case someone can suggest a better alternative in terms of being more idiomatic or performance benefits, but it offers sufficient performance for my current needs.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def loss(labels, logits):\n    labels = tf.reshape(labels, shape=(-1,))\n    logits = tf.reshape(logits, shape=(-1,))\n    output = tf.reshape(\n        tf.nn.sigmoid_cross_entropy_with_logits(\n            labels=labels,\n            logits=logits),\n        shape=(1, -1))\n\n    mask = tf.reshape(\n        tf.cast(labels &gt; logits, tf.float32),  # True iff label==1\n        shape=(-1, 1))\n\n    return tf.matmul(output, mask)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57694137,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2019-08-28T13:55:34.527","AcceptedAnswerId":null,"OwnerUserId":8214268.0,"Title":"How to get input index in a Keras Custom Layer","Body":"<p>I am trying to implement a custom Layer in Keras, which is the last layer in this network : <\/p>\n\n<p><a href=\"https:\/\/i.imgur.com\/p42Saw3.jpg\" rel=\"nofollow noreferrer\">https:\/\/i.imgur.com\/p42Saw3.jpg<\/a><\/p>\n\n<p>And this layers takes as input the 36*(2*33) feature map of the previous convolution, and outputs a fixed number of k-uplets as stated in this text : <\/p>\n\n<p><a href=\"https:\/\/i.imgur.com\/MLwWFnQ.jpg\" rel=\"nofollow noreferrer\">https:\/\/i.imgur.com\/MLwWFnQ.jpg<\/a><\/p>\n\n<p>As you see, to compute the k-uplets, I need to have access to the index i, j of the input feature that is used for the current call(). I don't know how to do that.<\/p>\n\n<p>I am writing a custom layer with tf.keras.Layer, and I know that the inputs of the layer goes into the call() method, but I don't know how to get the index i and j of the input.<\/p>\n\n<p>Here is my custom layer : <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class SDL_Layer(Layer):\n\n  def __init__(self, units=32):\n    super(SDL_Layer, self).__init__()\n    self.units = units\n\n  def build(self, input_shape):\n    self.DX = 0\n    self.dX = 0\n    self.DY = 0\n    self.dY = 0\n\n    self.rn = self.add_weight(shape=input_shape,\n                             initializer='random_normal',\n                             trainable=True)\n    self.sn = self.add_weight(shape=input_shape,\n                              initializer='random_normal',\n                              trainable=True)\n    self.tn = self.add_weight(shape=input_shape,\n                              initializer='random_normal',\n                              trainable=True)\n\n    self.bn = self.add_weight(initializer='random_normal',\n                             trainable=True)\n    self.cn = self.add_weight(initializer='random_normal',\n                              trainable=True)\n    self.dn = self.add_weight(initializer='random_normal',\n                              trainable=True)\n\n\n  def call(self, inputs, i, j):\n    rho = tf.math.sigmoid(tf.matmul(self.rn, inputs, transpose_a=True) + self.bn)\n    x = tf.math.sigmoid(tf.matmul(self.sn, inputs, transpose_a=True) + self.cn)*self.DX + (i-1)*self.dX\n    y = tf.math.sigmoid(tf.matmul(self.tn, inputs, transpose_a=True) + self.dn)*self.DX + (j-1)*self.dY\n    return tf.matmul(inputs, self.w) + self.b\n\n\n  def compute_output_shape(self, input_shape):\n    return (input_shape[0], self.output_dim)\n<\/code><\/pre>\n\n<p>Thank you.<\/p>\n","answers":[]}
+{"QuestionId":57694255,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-08-28T14:02:18.607","AcceptedAnswerId":null,"OwnerUserId":8909870.0,"Title":"Why does Tensoflow 2.0 and tf.function process some function by default on CPU","Body":"<p>I use tf.function with a GPU version of tensorflow. However for some functions I have to explicitly state the GPU as device (<code>with tf.device('\/device:GPU:0'):<\/code>) for tensorflow or otherwise the functions is executed on the CPU.<\/p>\n\n<p>For example this function of introducing k-sparsity to a given tensor is executed by default on only one CPU core:<\/p>\n\n<pre><code>@tf.function\ndef introdcue_k_sparsity(self, inputs; k):\n    k_real = round(k * inputs.shape[-1])\n    if k_real == 0:\n        k_real = 1\n\n    with tf.name_scope(\"sparsity_within_filter\"):\n        # Reorder pixel\n        winners, _ = tf.nn.top_k(inputs, k_real)\n\n        # Get all pixels that were sorted out and multiply with 0\n        winners_inverse_shape, _ = tf.nn.top_k(inputs, inputs.shape[-1]-k_real)\n        zeros = tf.scalar_mul(0, winners_inverse_shape)\n\n        k_sparse_outputs = tf.concat([winners, zeros], axis=-1)\n\n    return k_sparse_outputs\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57694732,"AnswerCount":1,"Tags":"<python><tensorflow><pip><object-detection>","CreationDate":"2019-08-28T14:29:22.320","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"Python uses wrong Package version","Body":"<p>Hi i try to launch the <code>object_detection_tutorial<\/code> on my PC. When i run the following code to load a (frozen) <code>Tensorflow<\/code> model into memory. <\/p>\n\n<pre><code>detection_graph = tf.Graph()\nwith 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<\/code><\/pre>\n\n<blockquote>\n  <p>ValueError: No op named NonMaxSuppressionV2 in defined operations.<\/p>\n<\/blockquote>\n\n<p>I googled the error and upgrading the tensorflow version to 1.4 should fix the bug. In my code i used tensorflow 1.13 and it worked in google cloud. But even after uninstalling and installing eg. tensorflow 1.4, python uses 1.2.1<\/p>\n\n<p>Picture of my code: <a href=\"https:\/\/ibb.co\/VYkq2rF\" rel=\"nofollow noreferrer\">https:\/\/ibb.co\/VYkq2rF<\/a><\/p>\n","answers":[{"AnswerId":"58553840","CreationDate":"2019-10-25T07:11:51.560","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>Looks like the module is not installed correctly. Try creating a new environment using conda and set up Object Detection in it. It should resolve your issue. <\/p>\n\n<p>Also, as best practices, it is always better to work on a conda environment than working on the base environment. <\/p>\n\n<p>Please use below command to create an empty new environment and then install packages on it.<\/p>\n\n<pre><code>conda create --no-default-packages --name &lt;env_name&gt; python=&lt;version&gt;\n<\/code><\/pre>\n"}]}
+{"QuestionId":57694957,"AnswerCount":0,"Tags":"<python><numpy><tensorflow><tensorflow-probability>","CreationDate":"2019-08-28T14:41:38.900","AcceptedAnswerId":null,"OwnerUserId":5370186.0,"Title":"tensor conversion requested dtype float32 for tensor with dtype float64 gradients\/BiasAdd_grad\/BiasAddGrad:0","Body":"<p>I am trying to create a variational gaussian output layer in a tensorflow probability model, but backprop is unable to add the bias as the dtype is mismatched - I am not sure how to convert though, as it seems that the float32 is coming from within tfp itself...<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tfk.backend.set_floatx('float64')\ndtype = x_white.dtype # dtype('float64')\n\nmodel = tfk.Sequential([\n    tfkl.InputLayer(input_shape=input_shape, dtype=dtype),\n    tfkl.LSTM(15, activation=tf.nn.leaky_relu, return_sequences=True),\n    tfkl.LSTM(15, activation=tf.nn.leaky_relu, return_sequences=True),\n    tfkl.LSTM(15, activation=tf.nn.leaky_relu, return_sequences=False),\n    tfkl.Dense(1, kernel_initializer='ones', dtype=dtype, use_bias=False),\n    tfpl.VariationalGaussianProcess(\n        num_inducing_points=ni,\n        kernel_provider=KernelFn(dtype=dtype),\n        event_shape=[1],\n        inducing_index_points_initializer=tf.constant_initializer(np.linspace(*x_range, num=ni, dtype=dtype)[..., np.newaxis]),\n    ),\n])\n\n\ndef loss(y, py):\n    return py.variational_loss(tf.cast(y, dtype), kl_weight=np.array(batch_size, dtype)\/samples)\n\nmodel.compile(optimizer=tf.optimizers.Adam(learning_rate=1e-3), loss=loss, metrics=['mse', 'acc'])\n\nprint(input_labels.dtype) # dtype('int64')\ninput_labels = input_labels.astype(dtype)\n\nhistory=model.fit(x_white, input_labels, validation_split=0.2, epochs=50, verbose=1, batch_size=batch_size, callbacks=[checkpoint, reducelrplateau])\n\n<\/code><\/pre>\n\n<pre><code>ValueError: Tensor conversion requested dtype float32 for Tensor with dtype float64: \n&lt;tf.Tensor 'gradients\/BiasAdd_grad\/BiasAddGrad:0', shape=(None,), dtype=float64&gt;\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57695308,"AnswerCount":0,"Tags":"<python><keras><recurrent-neural-network>","CreationDate":"2019-08-28T15:01:54.530","AcceptedAnswerId":null,"OwnerUserId":9372721.0,"Title":"How to correct a keras error generated by an unconnected input?","Body":"<p>I am estimating the values of the next time step based on a vector of lagged values in the current time step.  I would like to integrate the values predicted for the next time step with the lagged vector current time step.  The values predicted need to occupy columns 0 and 4 in the next lagged vector, or lagged[i+1][[0,4]] = ny_output.  The current lagged vector needs to be shifted so that lagged[i+1][1:4] = lagged[i][0:3] and lagged[i+1][5:7] = lagged[i+1][4:6].  The error message is<\/p>\n\n<p>AttributeError: 'NoneType' object has no attribute '_inbound_nodes'<\/p>\n\n<p>I believe this error is due to not using a keras layer.<\/p>\n\n<p>The minimum code to generate the error is<\/p>\n\n<pre><code>import keras\nimport numpy as np\n\n\ndef generate_data(a, nt):\n    ny = len(a)\n    max_lag = -1\n    for iy in range(ny):\n        for jy in range(ny):\n            max_lag = max(max_lag, len(a[iy][jy]))\n    d2 = np.zeros(shape=(nt, ny, max_lag), dtype=float)\n    # seed random number generator\n    np.random.seed(1)\n    d2[0] = np.random.rand(ny, max_lag)  # initalize lagged values for time 0\n    d2[:, :, 0] = np.nan  # Set all the y values to nan\n    for it in range(nt):\n        for iy in range(ny):\n            da = 0\n            for jy in range(ny):\n                lag_coeff = a[iy][jy][1:]\n                n_terms = len(lag_coeff)\n                if n_terms &gt; 0:\n                    lags = d2[it][jy][1:(n_terms + 1)]\n                    this_term = np.matmul(lags, lag_coeff)\n                    da = da - this_term\n            d2[it][iy][0] = da\n            if it &lt; (nt - 1):\n                d2[it + 1][iy][1:max_lag] = d2[it][iy][:(max_lag - 1)]\n    x2 = d2[:, :, 1:]\n    x2_1 = x2.shape[1]\n    x2_2 = x2.shape[2]\n    x2_product = x2_1 * x2_2\n    x2 = x2.reshape((nt, x2_product))\n    y2 = d2[:, :, 0].reshape((nt, ny))\n    return max_lag, ny, x2, y2\n\n\ndef data_split(fraction_train, fraction_validate, x2, y2):\n    nt = y2.shape[0]\n    n_train = round(fraction_train * nt)\n    n_validation = round(fraction_validate * n_train)\n    x_train_val = x2[:(n_train + n_validation)]\n    y_train_val = y2[:(n_train + n_validation)]\n    x_train = x_train_val[:n_train]\n    x_validation = x_train_val[n_train:]\n    y_train = y_train_val[:n_train]\n    y_validation = y_train_val[n_train:]\n    x_test = x2[(n_train + n_validation):]\n    y_test = y2[(n_train + n_validation):]\n    return x_train, y_train, x_test, y_test, x_validation, y_validation\n\n\nnt = 1000\na = [[[1, -1.1, 0.7, .169], [0, -.5, .1]],\n     [[0, -.04, .104], [1, -1.2, .6, .1, -.531]]]\n\nmax_lag, ny, x_data, y_data = generate_data(a, nt)\n\nfraction_train = 0.8\nfraction_validate = 0.1\nx_train, y_train, x_test, y_test, x_validation, y_validation = \\\n    data_split(fraction_train, fraction_validate, x_data, y_data)\n\nx_input = keras.Input(shape=(1, None, x_train.shape[1]), name='x_input')\nprint('x_input.shape: ', x_input.shape)\nny_output = keras.layers.Dense(ny, name='dense_a', use_bias=False)(x_input)\nprint('ny_output.shape:', ny_output.shape)\nprint('max_lag: ', max_lag)\noutput_list = list()\nky_start = 0\nstored_lags = max_lag - 1\nfor iy in range(ny):\n    ky_end = ky_start + stored_lags - 1\n    print('append output, {}:{}'.format(iy, iy + 1))\n    output_list.append(ny_output[:, :, :, iy:(iy + 1)])  # Add the next time step\n    print('append input, {}:{}'.format(ky_start, ky_end))\n    output_list.append(x_input[:, :, :, ky_start:ky_end])\n    ky_start = ky_end + 1  # Move to the next set of lags\n\noutputs = keras.layers.Concatenate(axis=3)(output_list)\nmodel = keras.Model(inputs=x_input, outputs=outputs)\nprint(model.summary())\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57695441,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-08-28T15:09:57.500","AcceptedAnswerId":null,"OwnerUserId":11989457.0,"Title":"Add a gaussian noise layer between pretrained network layers in Keras","Body":"<p>I would like to add a gaussian layer model to a pretrained vgg16 network (after convolutional layer 1) and have the noisy layer active when I pass the images (I would like to obtain activation of layers after passing through a noise layer during inference phase).<\/p>\n\n<p>I have managed to compile a 'new' model which has the gaussian layer added to it. I also have code for passing images and obtaining activations to them (using model.predict(img)). My problem is that now when I am passing the images through the gaussian noise layer the activation does not change from passing through the layer prior. I have read that gaussian noise layers are usually used in training (not testing - which is relevant for me), but I've also read that you can 'trick' a layer into believing you are in a training state and still have the gaussian noise active. I don't know how to practically implement this. <\/p>\n\n<p>Here I am creating my new model with gaussian noise layer:<\/p>\n\n<pre><code># load the model\nmodel = VGG16(weights='imagenet')\n# summarize the model\nmodel.summary()\n# loop over layers and add noise - create a new model\nfor i in range(1,len(model.layers)): # loop over all layers\n    #print ('Working on '+layer_names[i])\n    layer1 = model.layers[i]\n    if i==1: # adding the noise\n        noise = GaussianNoise(0.85)\n        x = noise(layer1.output)\n    if i&lt;len(model.layers)-1: # reconnecting the layers\n        layer2 = model.layers[i+1]\n        x = layer2(x) \n    if i==len(model.layers)-1:\n        predictors = x # this is the last layer\n# Create a new model \nmodel2 = Model(input=model.input, output=predictors)\nmodel2.summary() \n<\/code><\/pre>\n\n<p>Here is an example of how I would like to obtain activation of all layers (including Gaussian layer).<\/p>\n\n<pre><code>layer_names = ['block1_conv1',model2.layers[2].name,'block1_conv2','block2_conv1','block2_conv2','block3_conv1','block3_conv2','block3_conv3',\\\n               'block4_conv1','block4_conv2','block4_conv3','block5_conv1','block5_conv2','block5_conv3','fc1','fc2','predictions'];\nfor i in range(1,len(layer_names)): # loop over all layers\n    print ('Working on '+layer_names[i])\n    layer_name = layer_names[i]\n    intermediate_layer_model = Model(inputs=model2.input,outputs=model2.get_layer(layer_name).output)\n    # load the image with the required shape\n    img = image.load_img('test.BMP', target_size=(224,224)) # plot with img\n    # convert the image to an array\n    img = image.img_to_array(img)\n    # expand dimensions so that it represents a single 'sample'\n    img = np.expand_dims(img, axis=0)\n    # prepare the image (e.g. scale pixel values for the vgg)\n    img = preprocess_input(img)\n    # Now get the activations to this image\n    intermediate_output = intermediate_layer_model.predict(img)\n<\/code><\/pre>\n\n<p>The output (intermediate_output) should be different after receiving output from block1_cov1 layer vs. gaussian_noise layer, but it isn't. I guess that's because the Gaussian noise layer is not active, but I don't know how to change that. I know my code is suboptimal (just starting to work with Keras), but any help on how to get a noise to a pretrained network and let it 'propagate' through a network would be highly appreciated.<\/p>\n","answers":[]}
+{"QuestionId":57695689,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2019-08-28T15:23:45.717","AcceptedAnswerId":null,"OwnerUserId":11989989.0,"Title":"Tensor Shape Not Recognized in Tensorflow 2.0","Body":"<p>I cant get my RNN classifier to work with my input data. I am using TF 2.0 pre-release with a sliding window.  <\/p>\n\n<p>I am trying to build an RNN which I am feeding 5 timesteps with 6 features each and having it produce the 6th timestep as the target. When I run my code it is giving me an error saying that the input is (None,6) where as when I print out my training data it clearly says the shape is (5,6). I am very confused as to how to fix this. <\/p>\n\n<p>Error: <\/p>\n\n<pre><code>File \"C:\\Users\\employee\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training.py\", line 734, in fit\n    use_multiprocessing=use_multiprocessing)\n  File \"C:\\Users\\employee\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2.py\", line 224, in fit\n    distribution_strategy=strategy)\n  File \"C:\\Users\\employee\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2.py\", line 547, in _process_training_inputs\n    use_multiprocessing=use_multiprocessing)\n  File \"C:\\Users\\employee\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2.py\", line 593, in _process_inputs\n    steps=steps)\n  File \"C:\\Users\\employee\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training.py\", line 2384, in _standardize_user_data\n    all_inputs, y_input, dict_inputs = self._build_model_with_inputs(x, y)\n  File \"C:\\Users\\employee\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training.py\", line 2587, in _build_model_with_inputs\n    self._set_inputs(cast_inputs)\n  File \"C:\\Users\\employee\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training.py\", line 2674, in _set_inputs\n    outputs = self(inputs, **kwargs)\n  File \"C:\\Users\\employee\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\base_layer.py\", line 772, in __call__\n    self.name)\n  File \"C:\\Users\\employee\\AppData\\Local\\Programs\\Python\\Python36\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\input_spec.py\", line 177, in assert_input_compatibility\n    str(x.shape.as_list()))\nValueError: Input 0 of layer sequential is incompatible with the layer: expected ndim=3, found ndim=2. Full shape received: [None, 6]\n<\/code><\/pre>\n\n<p>Print readout:<\/p>\n\n<pre><code>********************tf.Tensor(\n[[0.07812838 0.08639083 0.07809999 0.08601701 0.6974719  0.6974719 ]\n [0.06794664 0.06995372 0.06220453 0.06934043 0.70064694 0.70064694]\n [0.08323035 0.08651368 0.07691107 0.08147305 0.69750804 0.69750804]\n [0.09781507 0.10009027 0.08847085 0.08919457 0.6944895  0.6944895 ]\n [0.12235662 0.12269666 0.11316498 0.11738694 0.6868     0.6868    ]], shape=(5, 6), dtype=float32)********************tf.Tensor([[0.08238748 0.09074993 0.07986343 0.09017278 0.6965872  0.6965872 ]], shape=(1, 6), dtype=float32)********************\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>\/data comes in as an array of shape [737,6]\n\ntrain=tf.data.Dataset.from_tensor_slices(features).window(6,1,1,drop_remainder=True).flat_map(lambda x: x.batch(6)).map(lambda window: (window[:-1],window[-1:]))\n\nvalid=train.take(200).shuffle(1000).repeat()\ntrain=train.shuffle(3000).repeat()\n\nfor x,y in valid:\n    print('*'*20+str(x)+\"*\"*20+str(y)+\"*\"*20)\n\nprint(train)\n\n\nmodel = tf.keras.Sequential()\n\nmodel.add(layers.SimpleRNN(128,batch_size=10))\nmodel.add(layers.Dense(124,kernel_initializer='he_uniform',activation='softmax'))\n\nmodel.compile(optimizer='adagrad', batch_size=10,step_size=.01, loss=tf.keras.losses.MeanAbsoluteError(), metrics=['accuracy'])\n\n\n\nhistory = model.fit(train,epochs=100, validation_data=valid,steps_per_epoch=3000,validation_steps=1000)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57701509","CreationDate":"2019-08-29T00:25:12.973","ParentId":null,"OwnerUserId":"8748870","Title":null,"Body":"<p>The model expects an input with rank 3, but is passed an input with rank 2.<\/p>\n\n<p>The first layer is a SimpleRNN, which expects data in the form (batch_size, timesteps, features), i.e. rank 3. The shape of the data passed by the user is (5, 6), i.e. rank 2.<\/p>\n\n<p>Passing rank-3 data (including the batch dimension) will fix the issue.<\/p>\n"},{"AnswerId":"58552657","CreationDate":"2019-10-25T05:26:46.017","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>You need to reshape your data or train variable into 3 dimensions i.e. \"[batch, timesteps, features]\".<\/p>\n\n<p>The model is expecting 3 dimensionsal input and your data is 2 dimensional.<\/p>\n\n<p>You can reshape your data like this : <\/p>\n\n<pre><code>data = tf.reshape(data, [-1,5,6])\n<\/code><\/pre>\n\n<p>And it should solve your issue.<\/p>\n"}]}
+{"QuestionId":57695692,"AnswerCount":0,"Tags":"<spring-boot><caching><caffeine><caffeine-cache>","CreationDate":"2019-08-28T15:23:47.850","AcceptedAnswerId":null,"OwnerUserId":3457177.0,"Title":"CaffeineCache is not refreshed immediately after put operation in Spring Boot application","Body":"<p>I have a random issue with <code>CaffeineCache<\/code> in my Spring Boot application. The problem is with an integration test that makes next<\/p>\n\n<ul>\n<li>find user<\/li>\n<li>delete it<\/li>\n<li>find it again <\/li>\n<\/ul>\n\n<p>It seems that sometimes cache doesn't not refreshes on time before the second call of <code>find<\/code> that comes immediately after delete. <\/p>\n\n<p>Here is a simplified signature of <code>find<\/code> method<\/p>\n\n<pre><code>@Cacheable(cacheNames = \"name\", key = \"{#config.id, #userId}\", unless = \"#result == null\")\npublic User find(SomeConfig config, String userId) { \n     \/\/ ... \n}\n<\/code><\/pre>\n\n<p>And a simplified signature of delete <\/p>\n\n<pre><code>@Caching(put = {\n        @CachePut(cacheNames = \"someOtherCache\", key = \"#userId.technicalId\"),\n        @CachePut(cacheNames = \"name\", key = \"{#config.id, #userId}\")\n})\npublic User delete(SomeConfig config, String userId)  { \n     \/\/ ... \n}\n<\/code><\/pre>\n\n<p>I suppose that after call <code>delete<\/code> cache is not updated immediately and that's why method <code>find<\/code> is not called the second time. It happens 1 times from 10. <\/p>\n\n<p>Any ideas about fix?<\/p>\n","answers":[]}
+{"QuestionId":57695745,"AnswerCount":0,"Tags":"<python><airflow><theano><pymc3><google-cloud-composer>","CreationDate":"2019-08-28T15:26:25.353","AcceptedAnswerId":null,"OwnerUserId":10791083.0,"Title":"Theano compile fails in composer","Body":"<p>My calculations using PyMC3 fail after pushing my operators to Google Cloud Composer. Errors provided below.<\/p>\n\n<p>We've tried updating the PyPI packages of the compute environment and looked at the logs. This shows some clear errors, which seem to indicate that a compile job is failing on the cloud.<\/p>\n\n<p>This is happens when airflow tries to parse our files - it is a fail on import.<\/p>\n\n<pre><code>Traceback (most recent call last): \n\nFile \"\/opt\/python3.6\/lib\/python3.6\/site-packages\/theano\/gof\/lazylinker_c.py\", line 81, in &lt;module&gt; \n    actual_version, force_compile, _need_reload)) \n\nImportError: Version check of the existing lazylinker compiled file. Looking for version 0.211, but found None. \n\nExtra debug information: force_compile=False, _need_reload=True\n<\/code><\/pre>\n\n<pre><code>Traceback (most recent call last): \nFile \"\/usr\/local\/lib\/airflow\/airflow\/models.py\", line 374, in \n    process_file m = imp.load_source(mod_name, filepath) \n\nFile \"\/opt\/python3.6\/lib\/python3.6\/imp.py\", line 172, in\n    load_source module = _load(spec) \n\nFile \"&lt;frozen importlib._bootstrap&gt;\", line 684, in \n    _load \n\nFile \"&lt;frozen importlib._bootstrap&gt;\", line 665, in \n    _load_unlocked \n\nFile \"&lt;frozen importlib._bootstrap_external&gt;\", line 678, in \n    exec_module \n\nFile \"&lt;frozen importlib._bootstrap&gt;\", line 219, in \n    _call_with_frames_removed \n\nFile \"\/home\/airflow\/gcs\/dags\/dag_abtesting\/abtesting_orchestration.py\", line 10, in &lt;module&gt; \n    from dag_abtesting.abtesting_tasks import task_analysis \n\nFile \"\/home\/airflow\/gcs\/dags\/dag_abtesting\/abtesting_tasks.py\", line 16, in &lt;module&gt; \n    from dag_abtesting.task_bernoulli import task_analyse_bernoulli \n\nFile \"\/home\/airflow\/gcs\/dags\/dag_abtesting\/task_bernoulli.py\", line 3, in &lt;module&gt; \n    from dag_abtesting.template_distributions import abstract_template_distribution, bernoulli_distribution \n\nFile \"\/home\/airflow\/gcs\/dags\/dag_abtesting\/template_distributions\/bernoulli_distribution.py\", line 6, in &lt;module&gt; \n    import pymc3 as pm \n\nFile \"\/opt\/python3.6\/lib\/python3.6\/site-packages\/pymc3\/__init__.py\", line 5, in &lt;module&gt; \n    from .distributions import * \n\nFile \"\/opt\/python3.6\/lib\/python3.6\/site-packages\/pymc3\/distributions\/__init__.py\", line 1, in &lt;module&gt; \n    from . import timeseries \n\nFile \"\/opt\/python3.6\/lib\/python3.6\/site-packages\/pymc3\/distributions\/timeseries.py\", line 1, in &lt;module&gt; \n    import theano.tensor as tt \n\nFile \"\/opt\/python3.6\/lib\/python3.6\/site-packages\/theano\/__init__.py\", line 110, in &lt;module&gt; \n    from theano.compile import ( \n\nFile \"\/opt\/python3.6\/lib\/python3.6\/site-packages\/theano\/compile\/__init__.py\", line 12, in &lt;module&gt; \n    from theano.compile.mode import * \n\nFile \"\/opt\/python3.6\/lib\/python3.6\/site-packages\/theano\/compile\/mode.py\", line 11, in &lt;module&gt; \n    import theano.gof.vm \n\nFile \"\/opt\/python3.6\/lib\/python3.6\/site-packages\/theano\/gof\/vm.py\", line 674, in &lt;module&gt; \n    from . import lazylinker_c \n\nFile \"\/opt\/python3.6\/lib\/python3.6\/site-packages\/theano\/gof\/lazylinker_c.py\", line 140, in &lt;module&gt; \n    preargs=args) \n\nFile \"\/opt\/python3.6\/lib\/python3.6\/site-packages\/theano\/gof\/cmodule.py\", line 2396, in \n    compile_str (status, compile_stderr.replace('\\n', '. '))) \n\nException: Compilation failed (return status=1): \/usr\/bin\/ld: cannot find -lpython3.6m. collect2: error: ld returned 1 exit status.\n<\/code><\/pre>\n\n<p>It looks to me as if I'm missing something -likely a compiler - on my hosted environment? Can someone advise on a next action to take?<\/p>\n","answers":[]}
+{"QuestionId":57695998,"AnswerCount":1,"Tags":"<python><tensorflow><keras><profiling><fft>","CreationDate":"2019-08-28T15:43:08.580","AcceptedAnswerId":57706197.0,"OwnerUserId":4332585.0,"Title":"Weird profiling result on keras models: the more complex the faster","Body":"<p>So I am currently trying to figure out which deep learning framework is best to handle fourier transforms. So far I was using <code>keras<\/code> with a <code>tensorflow<\/code> backend, but I noticed the fft was kinda slow (cf <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6541#issuecomment-524371619\" rel=\"nofollow noreferrer\">this issue on Github<\/a>).<\/p>\n\n<p>So recently I tried to compare the speed directly with <code>pytorch<\/code>. Since I want to do more than simply do a fourier transform, I tried adding some operations to do a more comprehensive benchmark, and I noticed that for <code>keras<\/code>, adding operations was decreasing the compute time.<\/p>\n\n<p>Here is the minimal working example (basically doing an inverse fourier transform in 2D, finished by taking the module of the image, and in between potentially \"decomplexification\" and \"recomplexification\"):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import os\nos.environ[\"CUDA_VISIBLE_DEVICES\"]=\"-1\"\nfrom keras.layers import Input, Lambda, concatenate\nfrom keras.models import Model\nimport numpy as np\nimport tensorflow as tf\nfrom tensorflow.signal import ifft2d\nimport time\n\ndef concatenate_real_imag(x):\n    x_real = Lambda(tf.math.real)(x)\n    x_imag = Lambda(tf.math.imag)(x)\n    return concatenate([x_real, x_imag])\n\ndef to_complex(x):\n    return tf.complex(x[0], x[1])\n\ndef complex_from_half(x, n, output_shape):\n    return Lambda(lambda x: to_complex([x[..., :n], x[..., n:]]), output_shape=output_shape)(x)\n\ndef weird_model(conc_then_com=False):\n    input_size = (320, None, 1)\n    kspace_input = Input(input_size, dtype='complex64', name='kspace_input')\n    inv_kspace = Lambda(ifft2d, output_shape=input_size)(kspace_input)\n    if conc_then_com:\n        inv_kspace = concatenate_real_imag(kspace_input)\n        inv_kspace = complex_from_half(inv_kspace, 1, input_size)\n    abs_inv_kspace = Lambda(tf.math.abs)(inv_kspace)\n    model = Model(inputs=kspace_input, outputs=abs_inv_kspace)\n    model.compile(\n        optimizer='adam',\n        loss='mse',\n    )\n    return model\n\n# fake data\ndata_x = np.random.rand(35, 320, 320, 1) + 1j * np.random.rand(35, 320, 320, 1)\ndata_y = np.random.rand(35, 320, 320, 1)\n\nstart = time.time()\nr = weird_model(conc_then_com=True).predict_on_batch(data_x)\nend = time.time()\nduration = end - start\nprint(f'For the prediction with the complex model it took {duration}')\n\nstart = time.time()\nr = weird_model(conc_then_com=False).predict_on_batch(data_x)\nend = time.time()\nduration = end - start\nprint(f'For the prediction with the simple model it took {duration}')\n\nstart = time.time()\nweird_model(conc_then_com=True).fit(\n    x=data_x,\n    y=data_y,\n    batch_size=35,\n    epochs=1,\n    verbose=2,\n    shuffle=False,\n)\nend = time.time()\nduration = end - start\nprint(f'For the fitting with the complex model it took {duration}')\n\nstart = time.time()\nweird_model(conc_then_com=False).fit(\n    x=data_x,\n    y=data_y,\n    batch_size=35,\n    epochs=1,\n    verbose=2,\n    shuffle=False,\n)\nend = time.time()\nduration = end - start\nprint(f'For the fitting with the simple model it took {duration}')\n<\/code><\/pre>\n\n<p>Which gives the following times (more or less):<\/p>\n\n<pre><code>For the prediction with the complex model it took 0.24\nFor the prediction with the simple model it took 3.98\nFor the fitting with the complex model it took 0.28\nFor the fitting with the simple model it took 4.01\n<\/code><\/pre>\n\n<p>I have no idea of what's going on.<\/p>\n","answers":[{"AnswerId":"57706197","CreationDate":"2019-08-29T08:33:06.540","ParentId":null,"OwnerUserId":"4332585","Title":null,"Body":"<p>Actually, this was just a typo:\n<code>inv_kspace = concatenate_real_imag(kspace_input)<\/code> should be <code>inv_kspace = concatenate_real_imag(inv_kspace)<\/code><\/p>\n"}]}
+{"QuestionId":57696071,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-08-28T15:47:56.250","AcceptedAnswerId":null,"OwnerUserId":11410664.0,"Title":"Numerically Stable Tensorflow","Body":"<p>Tensorflow has been used to compute images but I want to use Tensorflow to compute Biological Models. However, the biological model requires big division and this causes numerical instability. I want to have TensorFlow that supports more numerically stability. Are there any hacks to allow Tensorflow to be more numerically stable? I will follow up with more codes in the near feature but if there are any options please tell me. <\/p>\n","answers":[{"AnswerId":"58553481","CreationDate":"2019-10-25T06:43:42.283","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>Please refer to the below paper on Tensorflow Distributions and see if it helps answer your question. <\/p>\n\n<p><a href=\"https:\/\/arxiv.org\/pdf\/1711.10604.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1711.10604.pdf<\/a><\/p>\n\n<p>If not, please elaborate your issue. <\/p>\n\n<p>Mention any code snippet of what you have tried, or any limitations you faced while implementing your model.<\/p>\n"}]}
+{"QuestionId":57696200,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><tensorboard>","CreationDate":"2019-08-28T15:56:19.373","AcceptedAnswerId":null,"OwnerUserId":7327257.0,"Title":"Tensorboard. Duplicated graph node: one unconnected with placeholders and one connected","Body":"<p>I'm trying to visualize my model graph on TensorBoard. I'm using Keras 2.1.5 and tensorflow-gpu 1.13.1. My model is a concatenation of convolutional layers and, at the end, a custom layer where I make some operations with tensors. \nEverything works fine, although I defined some prints at the end and at the beggining of my custom layer and then check that Python enters two time my code when I just call it once. <\/p>\n\n<p>I'm running some trainings and check the model graph on TensorBoard and then found out something I haven't seen in any other example on the web:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/TqlGH.png\" rel=\"nofollow noreferrer\">Graph of custom layer<\/a><\/p>\n\n<p>There is a connected graph of my custom layer (trans2img) and another one unconnected and with empty placeholders as inputs. I don't understand the reason. <\/p>\n\n<p>Here is a simple example as my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def custom_layer(inputs):\n    with tf.name_scope('trans2img'):\n        a = inputs[0]\n        def some_operation(a):\n            with tf.name_scope('op1'):\n                b = 2*a\n                return b\n        def some_other_op(b, c):\n            with tf.name_scope('op2'):\n                d = b\/c\n                return d\n\n        b = some_operation(a)\n        d = some_other_op(b, inputs[1])\n        return d\n\n<\/code><\/pre>\n\n<p>After that, in my network definition file, I load this custom layer as <code>from custom_layer import custom_layer<\/code>, and then use it as a Lambda layer:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>net = Lambda(custom_layer)[branch1, branch2]\n<\/code><\/pre>\n\n<p>I don't know if it is because the way I define the inner operations in my custom_layer or the way I call them. I would like to know how to interpret this second unconnected graph I get and if it's an indicator of unnineficientcode. I would appreciate any clue and help.<\/p>\n","answers":[{"AnswerId":"58553356","CreationDate":"2019-10-25T06:33:13.053","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>This issue happens when you are using multiple tf scope. <\/p>\n\n<p>For each scope it creates a new op with \"op_{integer}\". <\/p>\n\n<p>You need to make use of \"absolute_name_scope(scope)\" to resolve your issue. <\/p>\n\n<p>Please refer to below link on how to make use of it. <\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/9545\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/9545<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/23250\/commits\/1169eaca048b660905fc5776e617cc824a19b125\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/pull\/23250\/commits\/1169eaca048b660905fc5776e617cc824a19b125<\/a><\/p>\n"}]}
+{"QuestionId":57696283,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-08-28T16:01:45.140","AcceptedAnswerId":null,"OwnerUserId":4522274.0,"Title":"TensorFlow: embedding model uses too much GPU memory unnecessarily","Body":"<p>I am working with knowledge graph embeddings in TensorFlow and I am having problems scaling it up: <\/p>\n\n<p>TensorFlow maintains the gradient of each batch sample with respect to <em>all embeddings<\/em>, even though each sample is only directly associated with three embeddings.<\/p>\n\n<p>In theory, for each batch it should only be necessary to bring to GPU memory the gradients of the embeddings associated with the present sample. This would massively decrease memory usage, with more memory overhead.<\/p>\n\n<p>Is there a way to be more memory efficient when batching the embeddings? <\/p>\n\n<p>I am using <code>tf.nn.embedding_lookup<\/code> without partitioning. I have tried paritioning with <code>tf.fixed_size_partitioner<\/code> but the memory usage was similar.<\/p>\n","answers":[]}
+{"QuestionId":57696905,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><windows-10>","CreationDate":"2019-08-28T16:49:46.237","AcceptedAnswerId":null,"OwnerUserId":11990399.0,"Title":"Error in python after 'import tensorflow' on windows","Body":"<p>After install TensorFlow for Python3 and type \"import tensorflow\" it wrote this long error.<\/p>\n\n<p>Any ideas how to fix it?<\/p>\n\n<pre><code>C:\\Users\\marsa&gt;python\nPython 3.7.3 (v3.7.3:ef4ec6ed12, Mar 25 2019, 22:22:05) [MSC v.1916 64 bit (AMD64)] on win32\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n&gt;&gt;&gt; import tensorflow\nTraceback (most recent call last):\n  File \"C:\\Users\\marsa\\AppData\\Local\\Programs\\Python\\Python37\\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\\marsa\\AppData\\Local\\Programs\\Python\\Python37\\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\\marsa\\AppData\\Local\\Programs\\Python\\Python37\\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\\marsa\\AppData\\Local\\Programs\\Python\\Python37\\lib\\imp.py\", line 242, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\Users\\marsa\\AppData\\Local\\Programs\\Python\\Python37\\lib\\imp.py\", line 342, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: Proces inicializace dynamicky p\u0159ipojovan\u00e9 knihovny (DLL) se nezda\u0159il.\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\\marsa\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\__init__.py\", line 28, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"C:\\Users\\marsa\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"C:\\Users\\marsa\\AppData\\Local\\Programs\\Python\\Python37\\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\\marsa\\AppData\\Local\\Programs\\Python\\Python37\\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\\marsa\\AppData\\Local\\Programs\\Python\\Python37\\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\\marsa\\AppData\\Local\\Programs\\Python\\Python37\\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\\marsa\\AppData\\Local\\Programs\\Python\\Python37\\lib\\imp.py\", line 242, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\Users\\marsa\\AppData\\Local\\Programs\\Python\\Python37\\lib\\imp.py\", line 342, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: Proces inicializace dynamicky p\u0159ipojovan\u00e9 knihovny (DLL) se nezda\u0159il.\n<\/code><\/pre>\n\n<p>Failed to load the native TensorFlow runtime.<\/p>\n","answers":[{"AnswerId":"57698289","CreationDate":"2019-08-28T18:34:11.110","ParentId":null,"OwnerUserId":"11978536","Title":null,"Body":"<p>What operating system are you using? Are you in a virtualenv? Please make sure the same python interpreter used for installing is also used for testing. For example, you could do:<\/p>\n\n<pre><code>python -m pip install tensorflow\npython -c \"import tensorflow\"\n<\/code><\/pre>\n\n<p>And this should not fail.<\/p>\n"}]}
+{"QuestionId":57697301,"AnswerCount":0,"Tags":"<tensorflow><networkx>","CreationDate":"2019-08-28T17:21:15.627","AcceptedAnswerId":null,"OwnerUserId":11591848.0,"Title":"TensorFlow Graph Using NetworkX Graph Data","Body":"<p>Can Anyone Help me in Making A TensorFlow Graph form The Below Kind of NetworkX Graph Link Data in Python<\/p>\n\n<pre><code>NetworkX Graph Data :\n{\n    \"nodes\": [\n            {\"name\": Peter },\n            {\"name\": Michael},\n            {\"name\": Neo4j},\n            {\"name\": Graph Database}],\n    \"links\": [\n        {\"source\": Peter,\"target\": Michael}, \n        {\"source\": Peter,\"target\": Neo4j},\n        {\"source\": Michael,\"target\": Neo4j},\n        {\"source\": Neo4j,\"target\": Graph Database}]\n}\n<\/code><\/pre>\n\n<p>This Kind of Tensor Flow Graph Required\nVisit <a href=\"https:\/\/www.tensorflow.org\/images\/graph_vis_animation.gif\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/images\/graph_vis_animation.gif<\/a><\/p>\n","answers":[]}
+{"QuestionId":57698091,"AnswerCount":0,"Tags":"<tensorflow><conv-neural-network><tf.keras>","CreationDate":"2019-08-28T18:19:46.160","AcceptedAnswerId":null,"OwnerUserId":10192534.0,"Title":"How to mask weights of convolutional layer in keras to keep some of them zero throughout training","Body":"<p>I am reproducing a pixelcnn. I am facing a problem in masking the kernel in keras. I make a tensor of the same shape as kernel, apply the mask and then set the weights by multiplying the with my mask. But during training all the weights change and the zeroed weights become non-zero. I want to keep them zero. How can I do this in keras?<\/p>\n","answers":[]}
+{"QuestionId":57698878,"AnswerCount":0,"Tags":"<python><tensorflow><gan>","CreationDate":"2019-08-28T19:24:43.510","AcceptedAnswerId":null,"OwnerUserId":11990893.0,"Title":"How to use trained generator to generate different samples","Body":"<p>I am playing with <code>Gan<\/code> on <code>MNIST<\/code>, the code is from a previous <code>tensorflow<\/code> tutorial. The problem is after I trained the <code>GAN<\/code>, I can only generate fake images reproduced the same training input sample. I think <code>GAN<\/code> is able to generate fake images reproduced all the training images. I am not sure if there are some extra codes I should add or there is something wrong with this code. <\/p>\n\n<p>After training, I run sampling from Generator several times. It gave me some similar images looks like reproduced the same certain training image. If I retrain it, it can give me a different type of image compared to the previous one but still suffer the previous issue<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport random\nimport numpy as np\nimport matplotlib.pyplot as plt\n%matplotlib inline\nfrom tensorflow.examples.tutorials.mnist import input_data\n\nmnist = input_data.read_data_sets(\"MNIST_data\/\")\nx_train = mnist.train.images[:55000,:]\nx_train.shape\nrandomNum = random.randint(0,55000)\nimage = x_train[randomNum].reshape([28,28])\nplt.imshow(image, cmap=plt.get_cmap('gray_r'))\nplt.show()\ndef conv2d(x, W):\n    return tf.nn.conv2d(input=x, filter=W, strides=[1, 1, 1, 1], padding='SAME')\n\ndef avg_pool_2x2(x):\n    return tf.nn.avg_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')\ndef discriminator(x_image, reuse=False):\n    with tf.variable_scope('discriminator') as scope:\n        if (reuse):\n            tf.get_variable_scope().reuse_variables()\n        #First Conv and Pool Layers\n        W_conv1 = tf.get_variable('d_wconv1', [5, 5, 1, 8], initializer=tf.truncated_normal_initializer(stddev=0.02))\n        b_conv1 = tf.get_variable('d_bconv1', [8], initializer=tf.constant_initializer(0))\n        h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)\n        h_pool1 = avg_pool_2x2(h_conv1)\n\n        #Second Conv and Pool Layers\n        W_conv2 = tf.get_variable('d_wconv2', [5, 5, 8, 16], initializer=tf.truncated_normal_initializer(stddev=0.02))\n        b_conv2 = tf.get_variable('d_bconv2', [16], initializer=tf.constant_initializer(0))\n        h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)\n        h_pool2 = avg_pool_2x2(h_conv2)\n\n        #First Fully Connected Layer\n        W_fc1 = tf.get_variable('d_wfc1', [7 * 7 * 16, 32], initializer=tf.truncated_normal_initializer(stddev=0.02))\n        b_fc1 = tf.get_variable('d_bfc1', [32], initializer=tf.constant_initializer(0))\n        h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*16])\n        h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)\n\n        #Second Fully Connected Layer\n        W_fc2 = tf.get_variable('d_wfc2', [32, 1], initializer=tf.truncated_normal_initializer(stddev=0.02))\n        b_fc2 = tf.get_variable('d_bfc2', [1], initializer=tf.constant_initializer(0))\n\n        #Final Layer\n        y_conv=(tf.matmul(h_fc1, W_fc2) + b_fc2)\n    return y_conv\ndef generator(z, batch_size, z_dim, reuse=False):\n    with tf.variable_scope('generator') as scope:\n        if (reuse):\n            tf.get_variable_scope().reuse_variables()\n        g_dim = 64 #Number of filters of first layer of generator \n        c_dim = 1 #Color dimension of output (MNIST is grayscale, so c_dim = 1 for us)\n        s = 28 #Output size of the image\n        s2, s4, s8, s16 = int(s\/2), int(s\/4), int(s\/8), int(s\/16) #We want to slowly upscale the image, so these values will help\n                                                                  #make that change gradual.\n\n        h0 = tf.reshape(z, [batch_size, s16+1, s16+1, 25])\n        h0 = tf.nn.relu(h0)\n        #Dimensions of h0 = batch_size x 2 x 2 x 25\n\n        #First DeConv Layer\n        output1_shape = [batch_size, s8, s8, g_dim*4]\n        W_conv1 = tf.get_variable('g_wconv1', [5, 5, output1_shape[-1], int(h0.get_shape()[-1])], \n                                  initializer=tf.truncated_normal_initializer(stddev=0.1))\n        b_conv1 = tf.get_variable('g_bconv1', [output1_shape[-1]], initializer=tf.constant_initializer(.1))\n        H_conv1 = tf.nn.conv2d_transpose(h0, W_conv1, output_shape=output1_shape, \n                                         strides=[1, 2, 2, 1], padding='SAME') + b_conv1\n        H_conv1 = tf.contrib.layers.batch_norm(inputs = H_conv1, center=True, scale=True, is_training=True, scope=\"g_bn1\")\n        H_conv1 = tf.nn.relu(H_conv1)\n        #Dimensions of H_conv1 = batch_size x 3 x 3 x 256\n\n        #Second DeConv Layer\n        output2_shape = [batch_size, s4 - 1, s4 - 1, g_dim*2]\n        W_conv2 = tf.get_variable('g_wconv2', [5, 5, output2_shape[-1], int(H_conv1.get_shape()[-1])], \n                                  initializer=tf.truncated_normal_initializer(stddev=0.1))\n        b_conv2 = tf.get_variable('g_bconv2', [output2_shape[-1]], initializer=tf.constant_initializer(.1))\n        H_conv2 = tf.nn.conv2d_transpose(H_conv1, W_conv2, output_shape=output2_shape, \n                                         strides=[1, 2, 2, 1], padding='SAME') + b_conv2\n        H_conv2 = tf.contrib.layers.batch_norm(inputs = H_conv2, center=True, scale=True, is_training=True, scope=\"g_bn2\")\n        H_conv2 = tf.nn.relu(H_conv2)\n        #Dimensions of H_conv2 = batch_size x 6 x 6 x 128\n\n        #Third DeConv Layer\n        output3_shape = [batch_size, s2 - 2, s2 - 2, g_dim*1]\n        W_conv3 = tf.get_variable('g_wconv3', [5, 5, output3_shape[-1], int(H_conv2.get_shape()[-1])], \n                                  initializer=tf.truncated_normal_initializer(stddev=0.1))\n        b_conv3 = tf.get_variable('g_bconv3', [output3_shape[-1]], initializer=tf.constant_initializer(.1))\n        H_conv3 = tf.nn.conv2d_transpose(H_conv2, W_conv3, output_shape=output3_shape, \n                                         strides=[1, 2, 2, 1], padding='SAME') + b_conv3\n        H_conv3 = tf.contrib.layers.batch_norm(inputs = H_conv3, center=True, scale=True, is_training=True, scope=\"g_bn3\")\n        H_conv3 = tf.nn.relu(H_conv3)\n        #Dimensions of H_conv3 = batch_size x 12 x 12 x 64\n\n        #Fourth DeConv Layer\n        output4_shape = [batch_size, s, s, c_dim]\n        W_conv4 = tf.get_variable('g_wconv4', [5, 5, output4_shape[-1], int(H_conv3.get_shape()[-1])], \n                                  initializer=tf.truncated_normal_initializer(stddev=0.1))\n        b_conv4 = tf.get_variable('g_bconv4', [output4_shape[-1]], initializer=tf.constant_initializer(.1))\n        H_conv4 = tf.nn.conv2d_transpose(H_conv3, W_conv4, output_shape=output4_shape, \n                                         strides=[1, 2, 2, 1], padding='VALID') + b_conv4\n        H_conv4 = tf.nn.tanh(H_conv4)\n        #Dimensions of H_conv4 = batch_size x 28 x 28 x 1\n\n    return H_conv4\n\nbatch_size = 16\nz_dimensions = 100\ntf.reset_default_graph()\n #Since we changed our batch size (from 1 to 16), we need to reset our Tensorflow graph\nsess = tf.Session()\nx_placeholder = tf.placeholder(\"float\", shape = [None,28,28,1]) #Placeholder for input images to the discriminator\nz_placeholder = tf.placeholder(tf.float32, [None, z_dimensions]) #Placeholder for input noise vectors to the generator\nDx = discriminator(x_placeholder) #Dx will hold discriminator prediction probabilities for the real MNIST images\nGz = generator(z_placeholder, batch_size, z_dimensions) #Gz holds the generated images\nDg = discriminator(Gz, reuse=True) #Dg will hold discriminator prediction probabilities for generated images\ng_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits = Dg, labels = tf.ones_like(Dg))) \n# ensure forward compatibility: function needs to have logits and labels args explicitly used\nd_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits = Dx, labels = tf.ones_like(Dx)))\nd_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits = Dg, labels = tf.zeros_like(Dg)))\nd_loss = d_loss_real + d_loss_fake\ntvars = tf.trainable_variables()\nd_vars = [var for var in tvars if 'd_' in var.name]\ng_vars = [var for var in tvars if 'g_' in var.name]\nadam = tf.train.AdamOptimizer()\ntrainerD = adam.minimize(d_loss, var_list=d_vars)\ntrainerG = adam.minimize(g_loss, var_list=g_vars)\nsess.run(tf.global_variables_initializer())\niterations = 500\nfor i in range(iterations):\n    z_batch = np.random.uniform(-1, 1, size=[batch_size, z_dimensions])\n    real_image_batch = mnist.train.next_batch(batch_size)\n    real_image_batch = np.reshape(real_image_batch[0],[batch_size,28,28,1])\n    _,dLoss = sess.run([trainerD, d_loss],feed_dict={z_placeholder:z_batch,x_placeholder:real_image_batch}) #Update the discriminator\n    if i%100 ==0:\n        print(\"dLoss:\",dLoss)\n    _,gLoss = sess.run([trainerG,g_loss],feed_dict={z_placeholder:z_batch}) #Update the generator\n    if i%100 ==0:\n        print(\"gLoss:\",gLoss)\nsample_image = generator(z_placeholder, 1, z_dimensions, reuse=True)#Gz\nz_batch = np.random.uniform(-1, 1, size=[1, z_dimensions])\ntemp = (sess.run(sample_image, feed_dict={z_placeholder: z_batch}))\nmy_i = temp.squeeze()\nplt.imshow(my_i, cmap='gray_r')\nplt.show()\nim = temp[0].reshape([1,28,28,1])\nresult = discriminator(x_placeholder, reuse=True)#D\nestimate = sess.run(result,{x_placeholder: im})#Dg discriminator prediction probability\nprint(\"Estimate:\",np.squeeze(estimate))\n<\/code><\/pre>\n\n<p>Could you help me to check the code attached? As it is a toy model, it can run quickly.<\/p>\n","answers":[]}
+{"QuestionId":57699309,"AnswerCount":1,"Tags":"<numpy><computer-vision><pytorch><image-segmentation><semantic-segmentation>","CreationDate":"2019-08-28T19:59:43.393","AcceptedAnswerId":57703364.0,"OwnerUserId":6908509.0,"Title":"Ground Truth Image to One Hot Encoded Array (Semantic Segmentation)","Body":"<p>I'm creating my own Dataset for People and Street Segmentation. Below, you see a labled Ground Truth (GT) Image. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/esYY4.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/esYY4.png\" alt=\"Ground Truth\"><\/a><\/p>\n\n<p>In the past I did a simple Regression between the model output and the GT Image (in the past I only used Streets). Now I read, Cross Entropy Loss is more common in that case. Since, my GT and also the model output Image has the same width <em>w<\/em> and height <em>h<\/em> as the input image, I have to create an array of size <em>h<\/em> x <em>w<\/em> x <em>c<\/em>, where <em>c<\/em> is the number of classes (in my case 3, background, street, people). I think, this is called One-Hot-Encoded Array.<\/p>\n\n<p>I solved this as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>        for height in range(len(img_as_np_array)):\n            for width in range(len(img_as_np_array[0])):\n                temp = np.zeros(classes)\n                if get_class(img_as_np_array[height,width]) == 1:\n                    temp[1] = 1\n                    one_hot_label[height,width] = temp\n                if get_class(img_as_np_array[height,width]) == 2:\n                    temp[2] = 1\n                    one_hot_label[height,width] = temp\n<\/code><\/pre>\n\n<p>where the method get_class(channels) decides the pixel class by the color of the pixel.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def get_class(channels):\n    threshold = 40\n    # Class 1 corresponds to streets, roads\n    if channels[0] in np.arange(243-threshold,243+threshold,1) and \\\n        channels[1] in np.arange(169-threshold,169+threshold,1) and \\\n        channels[2] in np.arange(0,threshold,1):\n        return 1\n\n    # Class 2 corresponds to people\n    if channels[0] in np.arange(0,threshold,1) and \\\n        channels[1] in np.arange(163-threshold,163+threshold,1) and \\\n        channels[2] in np.arange(232-threshold,232+threshold,1):\n        return 2\n\n    # Class 0 corresponds to background respectively other things\n    return 0\n<\/code><\/pre>\n\n<p>I have two questions:<\/p>\n\n<ol>\n<li><p>My approach is very slow (about 3 minutes for a Full HD Image), is there a way to speed this up?<\/p><\/li>\n<li><p>I noticed that the colors differ in the sense of channel values. For example, orange should be [243,169,0] (RGB), but I found entries like this [206,172,8] or even this [207,176,24] could that happen, because I store my labels as jpg? Is there a better way to find the orange and blue pixels than my idea above with the threshold?<\/p><\/li>\n<\/ol>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>I solved the first question by my self. This takes 2 or 3 seconds for an Full HD Image:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>threshold = 40\nclass_1_shape_cond_1 = (img_as_array[:, :, 0] &gt;= 243 - threshold) * (img_as_array[:, :, 0] &lt;= 243 + threshold)\nclass_1_shape_cond_2 = (img_as_array[:, :, 1] &gt;= 171 - threshold) * (img_as_array[:, :, 1] &lt;= 171 + threshold)\nclass_1_shape_cond_3 = (img_as_array[:, :, 2] &gt;= 0) * (img_as_array[:, :, 2] &lt;= threshold)\nclass_1_shape = (class_1_shape_cond_1 * class_1_shape_cond_2 * class_1_shape_cond_3)\n<\/code><\/pre>\n\n<p>Then I do the same for class 2 and for class 3 (everything else) I can do:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class_3_shape = 1 - (class_1_shape + class_2_shape)\n<\/code><\/pre>\n\n<p>After that I have to adjust the type with:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class_1_shape = class_1_shape.astype(np.uint8)\nclass_2_shape = class_2_shape.astype(np.uint8)\nclass_3_shape = class_3_shape.astype(np.uint8)\n<\/code><\/pre>\n\n<p>Question 2 is still an open.<\/p>\n","answers":[{"AnswerId":"57703364","CreationDate":"2019-08-29T05:12:19.573","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<h1>Do NOT save labels as JPEG images!<\/h1>\n\n<p>jpeg is a <a href=\"https:\/\/en.wikipedia.org\/wiki\/Lossy_compression\" rel=\"nofollow noreferrer\">lossy compression<\/a> method - that is, it is designed to save images using less bits even if it changes a bit the pixel values as long as it \"look good\" to a human observer.<br>\nThis is NOT the case for training labels stored as images! You cannot afford inaccuracies in the labels. You must use a <a href=\"https:\/\/en.wikipedia.org\/wiki\/Lossless_compression\" rel=\"nofollow noreferrer\">lossless compression<\/a> method, e.g., <a href=\"https:\/\/en.wikipedia.org\/wiki\/Portable_Network_Graphics\" rel=\"nofollow noreferrer\">png<\/a>.<\/p>\n\n<p>Better still, store your lables as <a href=\"https:\/\/stackoverflow.com\/a\/54929757\/1714410\">indexed RGB<\/a> images to begin with and save you all the trouble of inferring the discrete labels from the RGB values. <\/p>\n"}]}
+{"QuestionId":57699374,"AnswerCount":2,"Tags":"<python-3.x><keras><deep-learning><concatenation>","CreationDate":"2019-08-28T20:04:33.223","AcceptedAnswerId":null,"OwnerUserId":3261292.0,"Title":"Error while concatenating keras layers: A `Concatenate` layer requires inputs with matching shapes","Body":"<p>I have this model, which called <a href=\"https:\/\/medium.com\/analytics-vidhya\/hierarchical-attention-networks-d220318cf87e\" rel=\"nofollow noreferrer\">Hierarchical Attention Networks<\/a>: \n<a href=\"https:\/\/i.stack.imgur.com\/fbkaV.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/fbkaV.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Which proposed for document classification. I use word2vec embedding for the sentences words, and I want to concatenate another sentence-level embedding at point A (see the figure).<\/p>\n\n<p>I used it with documents contain 3 sentences; the model summary:\n<a href=\"https:\/\/i.stack.imgur.com\/l5CPm.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/l5CPm.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><code>word_input = Input(shape=(self.max_senten_len,), dtype='float32')\nword_sequences = self.get_embedding_layer()(word_input)\nword_lstm = Bidirectional(self.hyperparameters['rnn'](self.hyperparameters['rnn_units'], return_sequences=True, kernel_regularizer=kernel_regularizer))(word_sequences)\nword_dense = TimeDistributed(Dense(self.hyperparameters['dense_units'], kernel_regularizer=kernel_regularizer))(word_lstm)\nword_att = AttentionWithContext()(word_dense)\nwordEncoder = Model(word_input, word_att)\nsent_input = Input(shape=(self.max_senten_num, self.max_senten_len), dtype='float32')\nsent_encoder = TimeDistributed(wordEncoder)(sent_input)\n\n\"\"\" I added these following 2 lines. The dimension of self.training_features is (number of training rows, 3, 512). 512 is the dimension of the sentence-level embedding.  \"\"\"\nUSE = Input(shape=(self.training_features.shape[1], self.training_features.shape[2]), name='USE_branch')\nmerge = concatenate([sent_encoder, USE], axis=1)\n\nsent_lstm = Bidirectional(self.hyperparameters['rnn'](self.hyperparameters['rnn_units'], return_sequences=True, kernel_regularizer=kernel_regularizer))(merge)\nsent_dense = TimeDistributed(Dense(self.hyperparameters['dense_units'], kernel_regularizer=kernel_regularizer))(sent_lstm)\nsent_att = Dropout(dropout_regularizer)(AttentionWithContext()(sent_dense))\npreds = Dense(len(self.labelencoder.classes_))(sent_att)\nself.model = Model(sent_input, preds)\n<\/code><\/pre>\n\n<p>When I compile the above code, I get the following error:<\/p>\n\n<blockquote>\n  <p>ValueError: A <code>Concatenate<\/code> layer requires inputs with matching shapes\n  except for the concat axis. Got inputs shapes: [(None, 3, 128), (None,\n  3, 514)]<\/p>\n<\/blockquote>\n\n<p>I specified the concatenation axis=1, to concatenate on (3) the number of sentences, but I don't know why I'm still getting the error.<\/p>\n","answers":[{"AnswerId":"57700125","CreationDate":"2019-08-28T21:15:01.583","ParentId":null,"OwnerUserId":"8382707","Title":null,"Body":"<p>This is because the shape don't match if you specify that axis. This will work if you do this instead:<\/p>\n\n<pre><code>merge = concatenate([sent_encoder, USE], axis=-1)\n<\/code><\/pre>\n\n<p>Now there is no conflict of shapes on the remaining axes<\/p>\n"},{"AnswerId":"57700430","CreationDate":"2019-08-28T21:46:10.787","ParentId":null,"OwnerUserId":"3261292","Title":null,"Body":"<p>The error was due to two lines:<\/p>\n\n<pre><code>merge = concatenate([sent_encoder, USE], axis=1)\n# should be:\nmerge = concatenate([sent_encoder, USE], axis=2) # or -1 as @mlRocks suggested\n<\/code><\/pre>\n\n<p>and the line:<\/p>\n\n<pre><code>self.model = Model(sent_input, preds)\n# should be:\nself.model = Model([sent_input, USE], preds) # to define both inputs\n<\/code><\/pre>\n"}]}
+{"QuestionId":57699510,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2019-08-28T20:17:28.597","AcceptedAnswerId":null,"OwnerUserId":11991205.0,"Title":"tensorflow-serving gives me a \"is fed and fetched\" error for the same model that works from the command line","Body":"<p>When I run a prediction with <code>saved_model_cli<\/code> it executes properly, but the same inputs give me an error when I try to run it through tensorflow-serving.<\/p>\n\n<p>It tells me that an item is being both fed and fetched (because it is in the inputs and the outputs) but it works fine from the command line. Does anybody know why it works with one and not the other or how to go about fixing it?<\/p>\n","answers":[{"AnswerId":"58552866","CreationDate":"2019-10-25T05:50:43.350","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>Looks similar to the issue mentioned in the below link : <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/39307108\/placeholder-20-is-both-fed-and-fetched\">Placeholder_2:0 is both fed and fetched<\/a><\/p>\n\n<p>Please refer to this and see if it resolves your issue. If not, please share the code snippet to further debug the issue.<\/p>\n"}]}
+{"QuestionId":57699599,"AnswerCount":1,"Tags":"<tensorflow><keras><tf.keras>","CreationDate":"2019-08-28T20:26:20.463","AcceptedAnswerId":null,"OwnerUserId":2364295.0,"Title":"Tensorflow keras, shuffle not shuffling sample_weight?","Body":"<p>I have a sample weight array I am feeding in to fit().  It is of equal size to the number of training examples in my input.  If I turn my learning rate to zero and train for a number of epochs, I see different loss results for each epoch.  If I turn shuffle=False, results stays constant and match the result from evaluate.  It appears to me that the sample weight array must not be getting shuffled.  This is with tensorflow.keras. I assume this is not the expected behavior.  Is it a known bug, or might I be interpreting what I am seeing incorrectly?<\/p>\n","answers":[{"AnswerId":"58552754","CreationDate":"2019-10-25T05:39:12.860","ParentId":null,"OwnerUserId":"12198627","Title":null,"Body":"<p>You need to make <code>Shuffle=True<\/code> and also it shuffles only the training data and not the <a href=\"https:\/\/keras.io\/getting-started\/faq\/#is-the-data-shuffled-during-training\" rel=\"nofollow noreferrer\">validation data<\/a>.<\/p>\n\n<p>If you want to explicitly shuffle the weights in keras you can use <a href=\"https:\/\/gist.github.com\/jkleint\/eb6dc49c861a1c21b612b568dd188668\" rel=\"nofollow noreferrer\">this<\/a><\/p>\n"}]}
+{"QuestionId":57699641,"AnswerCount":1,"Tags":"<python><tensorflow><anaconda>","CreationDate":"2019-08-28T20:29:40.460","AcceptedAnswerId":null,"OwnerUserId":6328478.0,"Title":"AttributeError: module 'tensorflow' has no attribute 'executing_eagerly'","Body":"<p>Hello everyone I am new to machine learning, and I am attempting to build a custom object detection model on the CPU. However, I am getting this error below after I type:<\/p>\n\n<pre><code>python train.py --logtostderr --train_dir=training\/ -pipeline_config_path=training\/ssd_mobilenet_v1_coco.config. \n<\/code><\/pre>\n\n<p>Any input would be most appreciated. My path has been config but the problem occurs the seem.<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"train.py\", line 184, in &lt;module&gt;\n    tf.app.run()\n  File \"C:\\Users\\Thremane\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\", line 124, in run\n    _sys.exit(main(argv))\n  File \"C:\\Users\\Thremane\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py\", line 136, 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\\Thremane\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\object_detection-0.1-py3.5.egg\\object_detection\\legacy\\trainer.py\", line 304, in train\n    train_config.optimizer)\n  File \"C:\\Users\\Thremane\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\object_detection-0.1-py3.5.egg\\object_detection\\builders\\optimizer_builder.py\", line 45, in build\n    global_step=global_step)\n  File \"C:\\Users\\Thremane\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\object_detection-0.1-py3.5.egg\\object_detection\\builders\\optimizer_builder.py\", line 113, in _create_learning_rate\n    staircase=config.staircase)\n  File \"C:\\Users\\Thremane\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\object_detection-0.1-py3.5.egg\\object_detection\\utils\\learning_schedules.py\", line 79, in exponential_decay_with_burnin\n    if tf.executing_eagerly():\nAttributeError: module 'tensorflow' has no attribute 'executing_eagerly'\n<\/code><\/pre>\n","answers":[{"AnswerId":"57735878","CreationDate":"2019-08-31T06:55:47.777","ParentId":null,"OwnerUserId":"6328478","Title":null,"Body":"<p>I figured out what the issue was. The version of TensorFlow and the pre-train model must be compatible. I installed Tensorflow v1.5 in my environment and the appropriate comit package. Now I\u2019m training my model.<\/p>\n"}]}
+{"QuestionId":57699968,"AnswerCount":0,"Tags":"<python><loops><tensorflow><keras><tensor>","CreationDate":"2019-08-28T20:59:06.843","AcceptedAnswerId":null,"OwnerUserId":3737091.0,"Title":"Is there any way to iterate and take out 3d slices of a tensor?","Body":"<p>Is there a way to iterate over values of a 3D tensor for multiplication?\nI have a 3x3 mask that I want to multiply with each 3 dimensional pixels of an rgb image<\/p>\n\n<p>AN equivalent implementation is, say I have a 3x3 tensor which is a mask and a nonex200x200x3 rgb image. I want to multiply the 3x3 mask with the (r,g,b) value at (0,0,:) and then again with (0,1,:) and so on until (0,199,:) and then on to (1,0,:) all the way upto (199,199,:) to produce a (200,200,3) array.\nThis is because each of 3x3 mask by 3x1 element is supposed to yield a 3x1 element (or a 3D pixel) for each of 200x200 pixels in the input image.<\/p>\n\n<p>slicing is easy enough in a numpy array since I can iterate using indices, but is there any way to do so in Keras?\nAny help is highly appreciated\nThanks<\/p>\n\n<p>Keras.backend.dot is able to carry out the actual multiplication, but how to go about actually taking the slice?<\/p>\n","answers":[]}
+{"QuestionId":57700365,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-08-28T21:37:15.777","AcceptedAnswerId":57702559.0,"OwnerUserId":8622314.0,"Title":"Being confused about tensor dimensions in pytorch","Body":"<p>I'm new to pytorch and machine learning in general and I'm trying to create a simple convolutional neural net that classifies the MNIST handwritten digits.\nUnfortunately when I'm trying to train it, I get the following error:<\/p>\n\n<pre><code>ValueError: Expected input batch_size (288) to match target batch_size (64).\n<\/code><\/pre>\n\n<p>Here is the neural network code.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from torch import nn\nfrom torch.nn.functional import relu, log_softmax\n\nclass MNIST_SimpleConv(nn.Module):\n    def __init__(self):\n        super(MNIST_SimpleConv, self).__init__()\n\n        self.conv1 = nn.Conv2d(in_channels=3,  out_channels=32, kernel_size=3, stride=1)\n        self.conv2 = nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3, stride=1)\n\n        self.pool1 = nn.MaxPool2d(2, 2)\n\n        self.dense1 = nn.Linear(4*4*64, 100)\n        self.dense2 = nn.Linear(100, 10)\n\n    def forward(self, x):\n        x = relu(self.conv1(x))\n        x = relu(self.conv2(x))\n        x = self.pool1(x)\n\n        x = x.view(-1, 4*4*64)\n\n        x = relu(self.dense1(x))\n        return log_softmax(self.dense2(x), dim=1)\n<\/code><\/pre>\n\n<p>And the training code is as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from nets.conv import MNIST_SimpleConv\nfrom torchvision import datasets, transforms\nfrom torch.utils.data import DataLoader\nfrom torch.nn.functional import nll_loss\nimport torch.optim as optim\nimport torch\nfrom torch import nn\n\nMNIST_ROOT = \"data\/MNIST\"\n\n#prepare dataset\nmnist_train_ds = datasets.ImageFolder(root=MNIST_ROOT+\"\/train\", transform=transforms.Compose([\n                           transforms.ToTensor()]))\nmnist_test_ds  = datasets.ImageFolder(root=MNIST_ROOT+\"\/test\", transform=transforms.Compose([\n                           transforms.ToTensor()]))\nmnist_train = DataLoader(mnist_train_ds, batch_size=64, shuffle=True, num_workers=6)\nmnist_test  = DataLoader(mnist_test_ds, batch_size=64, shuffle=True, num_workers=6)\n\ncriterion = nn.CrossEntropyLoss()\n\ndef train(model, device, train_loader, optimizer, epoch):\n    model.train()\n\n    for batch_idx, (data, target) in enumerate(train_loader, 0):\n        data, target = data.to(device), target.to(device)\n        optimizer.zero_grad()\n        output = model(data)\n        loss = criterion(output, target)\n        loss.backward()\n        optimizer.step()\n        print('Train Epoch: {} [{}\/{} ({:.0f}%)]\\tLoss: {:.6f}'.format(\n                epoch, batch_idx * len(data), len(train_loader.dataset),\n                100. * batch_idx \/ len(train_loader), loss.item()))\n\n\ndevice = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n\n\nmodel = MNIST_SimpleConv().to(device)\noptimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.5)\n\nfor epoch in range(1, 10):\n    train(model, device, mnist_train , optimizer, epoch)\n\n<\/code><\/pre>\n\n<p>So far I have investigated how the dimensions of 'x' change while x is forwarded through the network.<\/p>\n\n<p>Input:\ntorch.Size([64, 3, 28, 28])<\/p>\n\n<p>After <code>x = relu(self.conv1(x))<\/code>:\ntorch.Size([64, 32, 26, 26])<\/p>\n\n<p>After <code>x = relu(self.conv2(x))<\/code>:\ntorch.Size([64, 64, 24, 24])<\/p>\n\n<p>After <code>x = self.pool1(x)<\/code>:\ntorch.Size([64, 64, 12, 12])<\/p>\n\n<p>After <code>x = x.view(-1, 4*4*64)<\/code>\ntorch.Size([576, 1024])<\/p>\n\n<p>After <code>x = relu(self.dense1(x))<\/code>\ntorch.Size([576, 100])<\/p>\n\n<p>After <code>x = log_softmax(self.dense2(x), dim=1)<\/code>\ntorch.Size([576, 10])<\/p>\n\n<p>The error is probably caused by x = x.view(-1, 4*4*64) for some reason producing a tensor with the shape [576, 1024] instead of [64, 1024].\n(If I understand this correctly the first dimension should be equal to the batch size which in my case is 64.)<\/p>\n\n<p>What am I doing wrong?<\/p>\n","answers":[{"AnswerId":"57702559","CreationDate":"2019-08-29T03:21:44.900","ParentId":null,"OwnerUserId":"5027315","Title":null,"Body":"<p>Passing a value of -1 to any dimension in <code>view<\/code> means that value of that particular dimension will be determined by other dimensions. for example:<\/p>\n\n<pre><code>x = torch.rand(1,10) # x.shape = [1,10]\nx = x.view(-1, 5)  # x.shape = [2, 5]\n<\/code><\/pre>\n\n<p>In your case, if you want to merge all dimensions of output of <code>pool1<\/code>, then it should be something like this:<\/p>\n\n<pre><code>x = x.view(-1, 64*12,*12) # x.shape = [64, 9216]\n<\/code><\/pre>\n\n<p>Also, we have to update the input channels for  <code>self.dense1<\/code> in this case:<\/p>\n\n<pre><code>self.dense1 = nn.Linear(64*12*12, 100)\n<\/code><\/pre>\n\n<p>However, one thing we need to make sure is that output dimensions of <code>self.pool1<\/code> are always going to be <code>batch_size x 64 x 12 x 12<\/code>, specifically the last two dimensions should stay 12 for the whole process. This can be made sure by fixing the input image dimensions across dataset.<\/p>\n"}]}
+{"QuestionId":57700596,"AnswerCount":0,"Tags":"<machine-learning><deep-learning><pytorch><resnet><cuda-streams>","CreationDate":"2019-08-28T22:05:12.720","AcceptedAnswerId":null,"OwnerUserId":11991273.0,"Title":"Execute another model in parallel to a model's forward pass with PyTorch","Body":"<p>I am trying to make some changes to the ResNet-18 model in PyTorch to invoke the execution of another auxiliary trained model which takes in the ResNet intermediate layer output at the end of each ResNet block as an input and makes some <strong>auxiliary predictions during the inference phase<\/strong>.<\/p>\n\n<p>I want to be able to do the <strong>auxiliary computation after the computation of a block in parallel to the computation of the next ResNet block<\/strong> so as to reduce the end-to-end latency of the entire pipeline execution on GPU.<\/p>\n\n<p>I have a base code that works correctly from the functionality perspective, but the execution of the auxiliary model is serial to the computation of the ResNet block. I verified this in two ways - <\/p>\n\n<ol>\n<li><p>By adding print statements and verifying the order of execution.<\/p><\/li>\n<li><p>By instrumenting the running time of the original ResNet model (say time t1) and the auxiliary model (say time t2). My execution time is currently t1+t2.<\/p><\/li>\n<\/ol>\n\n<p>The original ResNet block code (This is the BasicBlock since I am experimenting on ResNet-18). The entire code is available <a href=\"https:\/\/github.com\/pytorch\/vision\/blob\/master\/torchvision\/models\/resnet.py\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<pre><code>  class BasicBlock(nn.module):\n  ...\n\n      def forward(self, x):\n\n        residual = x\n        out = self.conv1(x)\n        out = self.bn1(out)\n        out = self.relu(out)\n        out = self.conv2(out)\n        out = self.bn2(out)\n\n        if self.downsample is not None:\n          residual = self.downsample(x)\n\n        out += residual\n        out = self.relu(out)\n        return out\n<\/code><\/pre>\n\n<p>This is my modification which works in a serial fashion -<\/p>\n\n<pre><code>def forward(self, x):\n    if len(x[0]) == self.auxiliary_prediction_size: # Got an Auxiliary prediction earlier\n        return x\n\n    # Do usual block computation\n    residual = x\n    out = self.conv1(x)\n    out = self.bn1(out)\n    out = self.relu(out)\n    out = self.conv2(out)\n    out = self.bn2(out)\n\n    if self.downsample is not None:\n        residual = self.downsample(x)\n\n    out += residual\n    out = self.relu(out)\n\n    # Try to make an auxiliary prediction\n    # First flatten the tensor (also assume for now that batch size is 1)\n    batchSize = x.shape[0]\n    intermediate_output = out.view(batchSize, -1)\n    # Place the flattened on GPU\n    device = torch.device(\"cuda:0\")\n    input = intermediate_output.to(device)\n    # Make auxiliary prediction\n    auxiliary_input = out.float()\n    auxiliary_prediction = self.auxiliary_model(auxiliary_input)\n    if auxiliary_prediction meets some condition:\n      return auxiliary_prediction\n\n    # If no auxiliary prediction, then return intermediate output\n    return out\n<\/code><\/pre>\n\n<p>Understandably, the above code causes a data dependency between the execution of the auxiliary model and the next block and hence things to happen serially. The first solution I tried was to check if breaking this data dependency reduces latency. I tried doing so by allowing the auxiliary model to execute but not having the auxiliary_prediction return if the condition is met (Note that this would break functionality but this experiment was purely to understand the behavior). Essentially, what I did was -<\/p>\n\n<pre><code>    batchSize = x.shape[0]\n    intermediate_output = out.view(batchSize, -1)\n    # Place the flattened on GPU\n    device = torch.device(\"cuda:0\")\n    input = intermediate_output.to(device)\n    # Make auxiliary prediction\n    auxiliary_input = out.float()\n    auxiliary_prediction = self.auxiliary_model(auxiliary_input)\n    if auxiliary_prediction meets some condition:\n      # Comment out return to break data dependency\n      #return auxiliary_prediction\n\n    # If no auxiliary prediction, then return intermediate output\n    return out\n<\/code><\/pre>\n\n<p>However, this did not work and upon researching further, I stumbled upon CUDA streams at <a href=\"https:\/\/stackoverflow.com\/questions\/52498690\/how-to-use-cuda-stream-in-pytorch\">this Stack Overflow link<\/a>. I tried incorporating the idea of CUDA streams to solve my problem in the below way -<\/p>\n\n<pre><code>def forward(self, x):\n    if len(x[0]) == self.auxiliary_prediction_size: # Got an Auxiliary prediction earlier\n        return x\n\n    s1 = torch.cuda.Stream()\n    s2 = torch.cuda.Stream()\n\n    with torch.cuda.Stream(s1):\n      # Do usual block computation\n      residual = x\n      out = self.conv1(x)\n      out = self.bn1(out)\n      out = self.relu(out)\n      out = self.conv2(out)\n      out = self.bn2(out)\n\n      if self.downsample is not None:\n        residual = self.downsample(x)\n\n      out += residual\n      out = self.relu(out)\n\n    with torch.cuda.Stream(s2):\n      # Try to make an auxiliary prediction\n      # First flatten the tensor (also assume for now that batch size is 1)\n      out_detach = out.detach() # Detach from backprop flow and from computational graph dependency\n      batchSize = x.shape[0]\n      intermediate_output = out_detach.view(batchSize, -1)\n      # Place the flattened on GPU\n      device = torch.device(\"cuda:0\")\n      input = intermediate_output.to(device)\n      # Make auxiliary prediction\n      auxiliary_input = out_detach.float()\n      auxiliary_prediction = self.auxiliary_model(auxiliary_input)\n      if auxiliary_prediction meets some condition:\n        return auxiliary_prediction\n\n    # If no auxiliary prediction, then return intermediate output\n    return out\n<\/code><\/pre>\n\n<p>However, the output from Nvidia Visual Profiler still indicates that all work is still being done on the default stream and still serialized. Note that I did verify with a small CUDA program that CUDA streams is supported by the CUDA version I am using.<\/p>\n\n<p>My questions -<\/p>\n\n<ol>\n<li><p>Why does breaking the data dependency not cause PyTorch to schedule the computations in parallel? I thought this was the point of the dynamic computation graphs in PyTorch.<\/p><\/li>\n<li><p>Why does using CUDA streams not delegate the computation to non-default streams?<\/p><\/li>\n<li><p>Are there alternative approaches to execute the auxiliary model asynchronously\/parallelly to the ResNet block computation?<\/p><\/li>\n<\/ol>\n","answers":[]}
+{"QuestionId":57700632,"AnswerCount":1,"Tags":"<python><deep-learning><anaconda><conda><caffe>","CreationDate":"2019-08-28T22:09:18.160","AcceptedAnswerId":null,"OwnerUserId":3204706.0,"Title":"Install and import Caffe with Anaconda (on Windows 10)","Body":"<p>I'm struggling to just install and import <code>caffe<\/code> on my windows 10's Anaconda (<code>conda 4.7.11<\/code>). I really don't know why such a well-known deep-learning package is that hard to just install!<\/p>\n\n<p>First, I tried to install it. With huge difficulties, I did the following and at last, apparently installation was successful:<\/p>\n\n<pre><code>conda create -n caffe python=3.7\n\/\/ activate the environment REM Depending on your conda version you may have to use activate caffe\n\nconda activate caffe\n\/\/ add the anaconda channel to resolve all dependencies\n\nconda config --add channels anaconda\n\/\/ Install caffe with cuda support or install caffe-cpu if you do not want CUDA support\n\nconda install caffe-cpu -c willyd\n<\/code><\/pre>\n\n<p>And now, just trying to <code>import caffe<\/code>, I get this error:<\/p>\n\n<pre><code>&gt;&gt;&gt; import caffe\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"C:\\Users\\user\\AppData\\Local\\Continuum\\anaconda3\\envs\\caffe\\lib\\site-packages\\caffe\\__init__.py\", line 1, in &lt;module&gt;\n    from .pycaffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, RMSPropSolver, AdaDeltaSolver, AdamSolver, NCCL, Timer\n  File \"C:\\Users\\user\\AppData\\Local\\Continuum\\anaconda3\\envs\\caffe\\lib\\site-packages\\caffe\\pycaffe.py\", line 13, in &lt;module&gt;\n    from ._caffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, \\\nImportError: DLL load failed: The specified module could not be found.\n<\/code><\/pre>\n\n<p>What should I do? How to fix it?<\/p>\n\n<p>BTW, is <code>Caffe2<\/code> different from <code>Caffe<\/code>?<\/p>\n","answers":[{"AnswerId":"58380719","CreationDate":"2019-10-14T16:29:26.437","ParentId":null,"OwnerUserId":"8462345","Title":null,"Body":"<p>Caffe2 is built on caffe but yes it is different. This link : <a href=\"http:\/\/gianni.rosagallina.com\/en\/posts\/2018\/10\/04\/caffe2-gpu-windows-1.html#step-4-build\" rel=\"nofollow noreferrer\">http:\/\/gianni.rosagallina.com\/en\/posts\/2018\/10\/04\/caffe2-gpu-windows-1.html#step-4-build<\/a> provides a good overview of what steps need to be taken to install caffe2. Incase you hit and error you cannot resolve, you can report the problems on the pytorch git and the team will help you out.<\/p>\n"}]}
+{"QuestionId":57700798,"AnswerCount":0,"Tags":"<python><tensorflow><keras><eager-execution>","CreationDate":"2019-08-28T22:31:24.967","AcceptedAnswerId":null,"OwnerUserId":4744283.0,"Title":"Manual weights regularization in TF+eager+Keras","Body":"<p>Slightly related to <a href=\"https:\/\/stackoverflow.com\/q\/57565019\/4744283\">Train complicated nn models with tf.eager (better with TF2 symbolic support)<\/a><\/p>\n\n<p>I want to train a network with a custom train process. For example I want to apply only gradients from current-hardest-samples-from-batch. Or I want to train multiple models with shared loss. I want to split train process into the forward pass, calculation of the gradients and the application of the gradients because I'm not sure whether such methods can be used within usual <code>model.compile<\/code> -> <code>model.fit<\/code> approach of <code>Keras<\/code>.<\/p>\n\n<p>I'm not sure that I will do it in a manual way, may be there exists a way to handle gradients with <code>model.compile<\/code> and <code>model.fit<\/code>, but I'm still interested in this topic. <\/p>\n\n<p>I want to ask following questions<\/p>\n\n<ol>\n<li>Does a code like the code below calculates gradients for regularization loss correctly?<\/li>\n<\/ol>\n\n<pre class=\"lang-py prettyprint-override\"><code>def loss_fun(model, image_batch, label_batch, reg_loss_coef=1e-5):\n    main_loss = my_another_loss_fun(image_batch, label_batch)\n\n    reg_sum = 1e-5\n    for layer in model.layer:\n        is_trainable = layer.trainable\n\n        is_weight = is_trainable and (type(layer) == tf.keras.layers.Conv2D) #model-dependent check here\n        if is_weight:\n            for var in layer.variables:\n                reg_sum += tf.math.reduce_sum(tf.math.square(var)) * reg_loss_coef\n\n    return main_loss + reg_sum\n\nmodel = tf.keras.applications.inception_resnet_v2.InceptionResNetV2(weights=None)\nopt = tf.train.AdamOptimizer(0.01, epsilon=1e-5)    \nds_tf = tf.data.Dataset.something()\nfor im_b, label_b in ds_tf:\n    with tf.GradientTape() as tape:\n        curr_loss = loss_fun(model, im_b, label_b)\n        with tape.stop_recording():\n            grads = tape.gradient(curr_loss, model.trainable_variables)\n            opt.apply_gradients(zip(grads, model_vars),\n                        global_step=tf.train.get_or_create_global_step())\n<\/code><\/pre>\n\n<ol start=\"2\">\n<li>Can I somehow check gradients calculation (especially regularization grads)?<\/li>\n<\/ol>\n","answers":[]}
+{"QuestionId":57701360,"AnswerCount":1,"Tags":"<python><tensorflow><object-detection-api>","CreationDate":"2019-08-29T00:01:12.013","AcceptedAnswerId":null,"OwnerUserId":11991681.0,"Title":"Fatal Python error: Segmentation fault on Win 10 while training the model (faster_rcnn_inception_v2_coco) for object detection","Body":"<p>I was trying to use the faster_rcnn_inception_v2_coco_2018_01_28 for my first object detection project. I was trying to train the model, and got this error. Before this error occurred, I was struggling with 'abs' being not found. I reinstalled tensorflow, and somehow it worked. <\/p>\n\n<p>This is the code that I ran on cmd. If I need to post what's in train.py or any code that I should have posted here, let me know.<\/p>\n\n<p><code>(tf1.12) C:\\Users\\EuisanKim\\Desktop\\tf1.12\\model\\research\\object_detection&gt;python legacy\/train.py --logtostderr --train_dir=training\/ --pipeline_config=training\/faster_rcnn_inception_v2_pets.config<\/code><\/p>\n\n<p>C:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:516: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint8 = np.dtype([(\"qint8\", np.int8, 1)])\nC:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:517: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint8 = np.dtype([(\"quint8\", np.uint8, 1)])\nC:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:518: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint16 = np.dtype([(\"qint16\", np.int16, 1)])\nC:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:519: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint16 = np.dtype([(\"quint16\", np.uint16, 1)])\nC:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:520: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint32 = np.dtype([(\"qint32\", np.int32, 1)])\nC:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:525: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  np_resource = np.dtype([(\"resource\", np.ubyte, 1)])\nC:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:541: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint8 = np.dtype([(\"qint8\", np.int8, 1)])\nC:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:542: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint8 = np.dtype([(\"quint8\", np.uint8, 1)])\nC:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:543: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint16 = np.dtype([(\"qint16\", np.int16, 1)])\nC:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:544: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint16 = np.dtype([(\"quint16\", np.uint16, 1)])\nC:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:545: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint32 = np.dtype([(\"qint32\", np.int32, 1)])\nC:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:550: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  np_resource = np.dtype([(\"resource\", np.ubyte, 1)])\nWARNING: Logging before flag parsing goes to stderr.\nW0829 08:28:04.658330 10092 lazy_loader.py:50]\nThe TensorFlow contrib module will not be included in TensorFlow 2.0.\nFor more information, please see:\n  * <a href=\"https:\/\/github.com\/tensorflow\/community\/blob\/master\/rfcs\/20180907-contrib-sunset.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/community\/blob\/master\/rfcs\/20180907-contrib-sunset.md<\/a>\n  * <a href=\"https:\/\/github.com\/tensorflow\/addons\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/addons<\/a>\n  * <a href=\"https:\/\/github.com\/tensorflow\/io\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/io<\/a> (for I\/O related ops)\nIf you depend on functionality not listed there, please file an issue.<\/p>\n\n<p>W0829 08:28:04.851034 10092 deprecation_wrapper.py:119] From C:\\Users\\EuisanKim\\Desktop\\tf1.12\\model\\research\\slim\\nets\\inception_resnet_v2.py:373: The name tf.GraphKeys is deprecated. Please use tf.compat.v1.GraphKeys instead.<\/p>\n\n<p>W0829 08:28:04.878998 10092 deprecation_wrapper.py:119] From C:\\Users\\EuisanKim\\Desktop\\tf1.12\\model\\research\\slim\\nets\\mobilenet\\mobilenet.py:397: The name tf.nn.avg_pool is deprecated. Please use tf.nn.avg_pool2d instead.<\/p>\n\n<p>W0829 08:28:04.907918 10092 deprecation_wrapper.py:119] From legacy\/train.py:55: The name tf.logging.set_verbosity is deprecated. Please use tf.compat.v1.logging.set_verbosity instead.<\/p>\n\n<p>W0829 08:28:04.908879 10092 deprecation_wrapper.py:119] From legacy\/train.py:55: The name tf.logging.INFO is deprecated. Please use tf.compat.v1.logging.INFO instead.<\/p>\n\n<p>W0829 08:28:04.913866 10092 deprecation_wrapper.py:119] From legacy\/train.py:184: The name tf.app.run is deprecated. Please use tf.compat.v1.app.run instead.<\/p>\n\n<p>W0829 08:28:04.916857 10092 deprecation.py:323] From C:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\absl\\app.py:251: main (from <strong>main<\/strong>) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse object_detection\/model_main.py.\nW0829 08:28:04.917856 10092 deprecation_wrapper.py:119] From legacy\/train.py:90: The name tf.gfile.MakeDirs is deprecated. Please use tf.io.gfile.makedirs instead.<\/p>\n\n<p>W0829 08:28:04.920847 10092 deprecation_wrapper.py:119] From legacy\/train.py:107: The name tf.gfile.Copy is deprecated. Please use tf.io.gfile.copy instead.<\/p>\n\n<p>Fatal Python error: Segmentation fault<\/p>\n\n<p>Current thread 0x0000276c (most recent call first):\n  File \"C:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\tensorflow\\python\\lib\\io\\file_io.py\", line 486 in copy_v2\n  File \"C:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\tensorflow\\python\\lib\\io\\file_io.py\", line 469 in copy\n  File \"legacy\/train.py\", line 108 in main\n  File \"C:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py\", line 324 in new_func\n  File \"C:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\absl\\app.py\", line 251 in _run_main\n  File \"C:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\absl\\app.py\", line 300 in run\n  File \"C:\\Users\\EuisanKim\\Anaconda3\\envs\\tf1.12\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\", line 40 in run\n  File \"legacy\/train.py\", line 184 in <\/p>\n\n<p>What I did before running the code above are:<\/p>\n\n<p>1) Annotating the data\n  - I was using different sizes of pictures (Could this lead to this error?)<\/p>\n\n<p>2) Creating the record files, label map file, and etc that are necessary for training<\/p>\n\n<p>I just started to learn programming. I don't know much about what's happening with this problem. Please, help me. Thank you.<\/p>\n","answers":[{"AnswerId":"58546911","CreationDate":"2019-10-24T18:03:42.950","ParentId":null,"OwnerUserId":"12198627","Title":null,"Body":"<p>It is due to . <\/p>\n\n<ol>\n<li>An import issue is there<\/li>\n<li>Code being written in python2 but running in python3 <\/li>\n<\/ol>\n\n<p>Do these steps to remove this error,<\/p>\n\n<ol>\n<li>Put <code>from object_detection import eval_util<\/code> as last import in file <code>model_lib.py<\/code> under <code>research\/object_detection<\/code><\/li>\n<li>Wrap range() objects with list() in API code (to support python3)<\/li>\n<\/ol>\n"}]}
+{"QuestionId":57701402,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2019-08-29T00:07:56.187","AcceptedAnswerId":null,"OwnerUserId":11991687.0,"Title":"problems of training RCF model using caffe","Body":"<p>I'm using caffe to train a model. I am sure I have connected the data layer with the train.txt in the source of image_data_param.But when I try .\/train.sh. It always prompt: can not find image.<\/p>\n\n<p>Ubuntu 18.04 openCV3 python2<\/p>\n\n<pre><code>layer {\n  name: \"data\"\n  type: \"ImageLabelmapData\"\n  top: \"data\"\n  top: \"label\"\n  include {\n    phase: TRAIN\n  }\n  transform_param {\n    mirror: false\n    mean_value: 104.00699\n    mean_value: 116.66877\n    mean_value: 122.67892\n  }\n  image_data_param {\n    root_folder: \"\/home\/yogang\/Desktop\/rawdata\/train\/\"\n    source: \"\/home\/yogang\/Desktop\/rawdata\/train.txt\"\n    batch_size: 1\n    shuffle: true\n    new_height:0\n    new_width: 0\n  }\n}\n\nI0828 19:29:13.834946 14079 layer_factory.hpp:77] Creating layer data\nI0828 19:29:13.835011 14079 net.cpp:101] Creating Layer data\nI0828 19:29:13.835031 14079 net.cpp:409] data -&gt; data\nI0828 19:29:13.835059 14079 net.cpp:409] data -&gt; label\nI0828 19:29:13.835124 14079 image_labelmap_data_layer.cpp:42] Opening file \/home\/yogang\/Desktop\/rawdata\/train.txt\nI0828 19:29:13.835505 14079 image_labelmap_data_layer.cpp:52] Shuffling data\nI0828 19:29:13.835677 14079 image_labelmap_data_layer.cpp:57] A total of 242 images.\nE0828 19:29:13.836748 14079 io.cpp:80] Could not open or find file \/home\/yogang\/Desktop\/rawdata\/train\/\/home\/yogang\/Desktop\/rawdata\/train\/satellite144.jpg\nE0828 19:29:13.836797 14079 io.cpp:80] Could not open or find file \/home\/yogang\/Desktop\/rawdata\/train\/\/home\/yogang\/Desktop\/rawdata\/train\/400.jpg\nF0828 19:29:13.836818 14079 image_labelmap_data_layer.cpp:86] Check failed: cv_img.data Could not load \/home\/yogang\/Desktop\/rawdata\/train\/satellite144.jpg\n*** Check failure stack trace: ***\n.\/train.sh: line 8: 14079 Aborted                 (core dumped) .\/solve.py\n<\/code><\/pre>\n","answers":[{"AnswerId":"57703097","CreationDate":"2019-08-29T04:44:15.450","ParentId":null,"OwnerUserId":"11976900","Title":null,"Body":"<p>i think image path is wrong\ncheck your image path<\/p>\n"}]}
+{"QuestionId":57701511,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras><nlp>","CreationDate":"2019-08-29T00:25:29.853","AcceptedAnswerId":null,"OwnerUserId":9922891.0,"Title":"Language model computes incorrect perplexity using keras.backend.pow()","Body":"<p>I implement a <code>perplexity()<\/code> function to evaluate an <code>LSTM<\/code> language model in <code>Keras<\/code>. I am using the <code>cross entropy loss<\/code> and <code>one-hot encoding<\/code>. The resulting perplexity is not <code>exp(loss)<\/code>.<\/p>\n\n<p>I adopted the perplexity code from <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/2317#issuecomment-307605458\" rel=\"nofollow noreferrer\">this github discussion<\/a> and change the base from <code>2<\/code> to <code>e<\/code> as suggested in <a href=\"https:\/\/stackoverflow.com\/a\/46606291\/9922891\">this stackoverflow post<\/a>.<\/p>\n\n<p>I also wrote a callback to compute perplexity. This callback computes the correct perplexity (<code>exp(loss)<\/code>).<\/p>\n\n<p>Some additional background information:<br>\n- <code>tensorflow: 1.14.0<\/code><br>\n- <code>keras: 2.2.4<\/code><br>\n- dataset: Penn Treebank (PTB)<br>\n- <code>loss: categorical_crossentropy()<\/code><\/p>\n\n<p>The <code>perplexity()<\/code> function.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras import backend as K\n\ndef perplexity(true_label, pred_label):\n  cross_entropy = K.categorical_crossentropy(true_label, pred_label)\n  return K.exp(cross_entropy)\n<\/code><\/pre>\n\n<p>The callback class.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class Perplexity(Callback):\n  def on_epoch_end(self, epoch, logs={}):\n    super().on_epoch_end(epoch,logs)\n    print(f\"Callback Perplexity: {K.eval(K.exp(logs['loss']))}\")\n<\/code><\/pre>\n\n<p>Here is the model. Irrelevant variables are omitted.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential()\nmodel.add(Embedding(...))\nmodel.add(LSTM(...))\nmodel.add(Dense(activation='softmax', ...))\nmodel.compile(loss='categorical_crossentropy', metrics=[perplexity], ...)\nmodel.summary()\n\nmodel.fit_generator(callbacks=[Perplexity()], ...)\n<\/code><\/pre>\n\n<p>For example, if the <code>cross entropy loss<\/code> is <code>7.0554<\/code>,<br>\n- the <code>perplexity()<\/code> function gives <code>9555.6286<\/code><br>\n- the callback gives <code>1158.995545671746<\/code><br>\nThe callback's result is close to the expected perplexity <code>exp(7.0554)=1159.101019<\/code>.  <\/p>\n\n<p>How can I use the <code>perplexity()<\/code> function to compute the correct perplexity? Are there any bugs in my code? I can provide more code if necessary. Thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":57701525,"AnswerCount":2,"Tags":"<python><tensorflow><tensorflow-probability>","CreationDate":"2019-08-29T00:29:13.910","AcceptedAnswerId":57702786.0,"OwnerUserId":5910685.0,"Title":"How to bound input dimension for differential_evolution_minimize in tensorflow-probability?","Body":"<p>Unlike in <a href=\"https:\/\/docs.scipy.org\/doc\/scipy-0.15.1\/reference\/generated\/scipy.optimize.differential_evolution.html\" rel=\"nofollow noreferrer\">scipy's implementation<\/a> of differential evolution (DE), there is no direct way to define bounds for my inputs\nin <a href=\"https:\/\/www.tensorflow.org\/probability\/api_docs\/python\/tfp\/optimizer\/differential_evolution_minimize\" rel=\"nofollow noreferrer\">tensorflow-probability<\/a>. <\/p>\n\n<p>Since the inputs to my function are 5-tuples defining a pixel in an image\n(x,y,R,G,B), the values would need to be integers and bounded to my image dimension as well as RGB values of 0 to 255 during optimization.<\/p>\n\n<p>To paint the bigger picture:\nI want to minimize the confidence of an image classifier which itself is part of a reinforcement learning agent that needs to decide on an action depending on its observation of the environment. The agent is fully trained but now I want to throw a few stones by perturbing a single pixel in its observation and monitor its performance.<\/p>\n\n<p>I want to use the differential evolution algorithm to find the pixel which would decrease the confidence of the agent in its actions the most.<\/p>\n\n<p>Currently I have an action-prediction function, which takes as an argument the perturbation pixel, runs the perturbed observation through the classifier and returns the resulting confidence in the action which the agent would have chosen without the perturbation:<\/p>\n\n<p>Code block I<\/p>\n\n<pre><code>#random perturbation pixel as an example of my input:\npixel = tf.constant([36,48,255,255,255]) # (x,y,R,G,B)\n\ndef predict_action(pixel):\n   perturbed_obs = perturb_obs(pixel, observation)\n   confidence = classifier(perturbed_obs)\n   return confidence\n<\/code><\/pre>\n\n<p>Now I would want to give this function to the optimizer with an initial population:<\/p>\n\n<p>Code block II<\/p>\n\n<pre><code>popsize=80\n\ninit_pop = generate_population(popsize)\n# returns Tensor(\"scan\/while\/Squeeze:0\", shape=(80, 5), dtype=int64)\n# i.e. 80 random perturbation pixels\n\nresults = tfp.optimizer.differential_evolution_minimize(\n    predict_action, initial_population=init_pop, seed=42)\n<\/code><\/pre>\n\n<p>Yet, how do I define the bounds of my input, such that the population will always be valid pixels?<\/p>\n\n<p>I asked about this on <a href=\"https:\/\/github.com\/tensorflow\/probability\/issues\/534\" rel=\"nofollow noreferrer\">GitHub<\/a> and a possible way of achieving this is through the use of their bijector functionality:<\/p>\n\n<p>Code block III<\/p>\n\n<pre><code># First we squash `pixel_logits` to (0, 1), then scale it to (0, 255).\nbijector = tfb.Affine(scale=255.)(tfb.Sigmoid())\n\ndef unconstrained_objective_fn(pixel_logits):\n    return objective_fn(bijector.forward(pixel_logits))\n\nresults = minimize(unconstrained_objective_fn, initial_position=bijector.inverse(initial_pixels))\npixels = bijector.forward(results.position)\n<\/code><\/pre>\n\n<p>While I understand the approach in principle, I fail to apply this to my current situation\/understanding of my problem.<\/p>\n\n<p>EDIT: Removed information that was unrelated to the main question. Always mind your types and dimensions!<\/p>\n","answers":[{"AnswerId":"57718947","CreationDate":"2019-08-29T23:57:22.337","ParentId":null,"OwnerUserId":"6716334","Title":null,"Body":"<p>It seems like a regular dtype error: you appear to be passing a <code>tf.int64<\/code> type into the initial position\/population argument of differential evolution, while it expects (understandably) floating point tensors. Here's a version of your final code snippet that works for me:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>initial_pixels = 0.5\n\nbijector = tfb.AffineScalar(scale=255.)(tfb.Sigmoid())\n\ndef objective_fn(pixels):\n  return tf.square(pixels - 0.75)\n\ndef unconstrained_objective_fn(pixel_logits):\n    return objective_fn(bijector.forward(pixel_logits))\n\nresults = tfp.optimizer.differential_evolution_minimize(unconstrained_objective_fn, initial_position=bijector.inverse(initial_pixels))\npixels = bijector.forward(results.position)\nprint(pixels)\n<\/code><\/pre>\n"},{"AnswerId":"57702786","CreationDate":"2019-08-29T03:58:54.093","ParentId":null,"OwnerUserId":"8935282","Title":null,"Body":"<p>Based on my experience with tfp-0.6 and tf-1.13.1, Code block III could be rewritten as follows:<\/p>\n\n<pre><code>width = ... #some Python float\nheight = ... #some other Python float\n\nbijectors = [\n    tfb.Chain([tfb.AffineScalar(scale=width), tfb.Sigmoid()]),\n    tfb.Chain([tfb.AffineScalar(scale=height), tfb.Sigmoid()]),\n    tfb.Chain([tfb.AffineScalar(scale=255.), tfb.Sigmoid()]),\n    tfb.Chain([tfb.AffineScalar(scale=255.), tfb.Sigmoid()]),\n    tfb.Chain([tfb.AffineScalar(scale=255.), tfb.Sigmoid()])\n  ]\n\ndef constrained_objective_fn(pixel_logits):\n    constrained_pixel_logits = [b.forward(p) for b, p in zip(bijectors, pixel_logits)]\n    return objective_fn(constrained_pixel_logits)\n\nresults = tfp.optimizer.differential_evolution_minimize(\nconstrained_objective_fn, \ninitial_population=init_pop, \nseed=42)\n\npixels = [b.forward(p) for b, p in zip(bijectors, results.position)]\n<\/code><\/pre>\n\n<p>Note that the initial population from Code Block II is preserved and bijector.inverse is not applied. <\/p>\n"}]}
+{"QuestionId":57701550,"AnswerCount":0,"Tags":"<c++><tensorflow><ram><tensorflow-datasets>","CreationDate":"2019-08-29T00:33:31.063","AcceptedAnswerId":null,"OwnerUserId":8215939.0,"Title":"Can you train a neural network on GPUs while generating the next batch of data on CPUs concurrently?","Body":"<p>Currently, we have a data pipeline that takes about 1 week to compute in parallel (written in python) and occupies about 4 TB. We would like to research if the data representation we chose is optimal or not and how different architectures perform on the different data representation. Thus, we would like to be able change the data representation and build models on these new data representations. However, to make a new representation we would have to wait 7 days maybe longer depending on the changes we made and then begin training our model on the data (3 days of training as of the current model). Thus, our current cycle requires 10 days to generate and then train our neural net.  This is very time and space expensive if we would like to explore 50 different data representations. <\/p>\n\n<p>Thus, we have began rewriting the entire data pipeline in C++ to speed up the data generation. While this alleviates the time cost of generating the data, it does not solve our issue of storing that much data. We would much rather save the metadata used to generate the data. <\/p>\n\n<p>Is it possible for us to generate the data and pass it straight to tensorflow without ever writing it to disk?<\/p>\n\n<p>The supercomputers we have access to have 16 CPUs and 4 GPUs per node. So we were wondering if we could train a model on the GPUs and generate the next batch of data on the CPUs while the net is training. <\/p>\n\n<p>I have not been able to find anything on the internet that addresses this specific question. I am wondering if this functionality is built into tensorflow or pytorch? We are almost done rewriting the data pipeline in C++ and are a few weeks away from trying to access the data from tensorflow. Again, our whole goal is never write the data to disk.<\/p>\n","answers":[]}
+{"QuestionId":57701568,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-29T00:36:51.943","AcceptedAnswerId":null,"OwnerUserId":11354575.0,"Title":"The output of train_on_batch is batch_wise average or global average","Body":"<p>The output or <code>train_on_batch<\/code> of keras is the average of the current batch or global average of all trained samples?<\/p>\n","answers":[{"AnswerId":"58546345","CreationDate":"2019-10-24T17:22:15.113","ParentId":null,"OwnerUserId":"12198627","Title":null,"Body":"<p>The <code>train_on_batch<\/code> function accepts a single batch of data, performs backpropagation, and then updates the model parameters. It uses single batch only and once.<\/p>\n\n<p>Now to answer your question, Output of <code>train_on_batch<\/code> provides the training loss for that <strong>current batch<\/strong>.<\/p>\n\n<p>You can refer below links for more description about same.<\/p>\n\n<ol>\n<li><a href=\"https:\/\/keras.io\/models\/model\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/models\/model\/<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/10364\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/10364<\/a><\/li>\n<\/ol>\n"}]}
+{"QuestionId":57701649,"AnswerCount":0,"Tags":"<c++><tensorflow>","CreationDate":"2019-08-29T00:52:22.700","AcceptedAnswerId":null,"OwnerUserId":11991802.0,"Title":"Tensorflow frozen graph producing empty output in C++ API","Body":"<p>I am using the Tensorflow 1.11.0 C++ API to load a frozen graph of the DeepLab mobilenetv2_coco_voc_trainval model (<a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/deeplab\/g3doc\/model_zoo.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/deeplab\/g3doc\/model_zoo.md<\/a>).<\/p>\n\n<p>I am able to load the model without complaints like so:<\/p>\n\n<pre class=\"lang-cpp prettyprint-override\"><code>      \/\/ Read in the protobuf graph we exported\n      tensorflow::GraphDef graph_def;\n      status = ReadBinaryProto(tensorflow::Env::Default(), graph_fn, &amp;graph_def);\n      if (status != tensorflow::Status::OK()) return status;\n\n      \/\/ create the graph in the current session\n      status = sess-&gt;Create(graph_def);\n      if (status != tensorflow::Status::OK()) return status;\n\n      return tensorflow::Status::OK();\n<\/code><\/pre>\n\n<p>but when I run an image through the model like below, the output is all zeroes no matter what image I pass in. The output is of the correct shape, however.<\/p>\n\n<pre class=\"lang-cpp prettyprint-override\"><code>      tensor_dict feed_dict = { \n          {\"ImageTensor:0\", data},\n      };  \n\n      std::vector&lt;tensorflow::Tensor&gt; outputs;\n      TF_CHECK_OK(sess-&gt;Run(feed_dict, {\"SemanticPredictions:0\"},\n                            {}, &amp;outputs));\n<\/code><\/pre>\n\n<p>I load the image using opencv like this:<\/p>\n\n<pre class=\"lang-cpp prettyprint-override\"><code>      \/\/ Load input image and convert to uint8 for model\n      cv::Mat input_image = cv::imread(image_filename);\n      cv::cvtColor(image,image,cv::COLOR_BGR2RGB);\n      input_image.convertTo(image,CV_8U);\n\n      \/\/ Define the input tensor\n      tensorflow::TensorShape data_shape({1,height,width,3});\n      tensorflow::Tensor data(tensorflow::DT_UINT8, data_shape);\n\n      \/\/ Pointer to the underlying data structure\n      uint8_t *tensorPointer = data.flat&lt;uint8_t&gt;().data();\n\n      \/\/Copy into tensor\n      memcpy(tensorPointer, input_image.data, height * width * 3 * sizeof(uint8_t));\n\n<\/code><\/pre>\n\n<p>I need to stay at Tensorflow 1.11.0 for another model I'm using. Is there something I missed?<\/p>\n","answers":[]}
+{"QuestionId":57701959,"AnswerCount":1,"Tags":"<tensorflow><keras><neural-network><keras-layer>","CreationDate":"2019-08-29T01:48:57.060","AcceptedAnswerId":null,"OwnerUserId":7655096.0,"Title":"Adapt Text Classifier Neural Net To Accept Multiple Categories","Body":"<p>I'm trying to adjust the text classifier neural net in this Keras\/Tensorflow tutorial to output multiple (more than 2 categories). I think I can change the output layer to use a 'softmax' activation but I'm not sure how to adjust the input layer.<\/p>\n\n<p>Tutorial Link:\n<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<p>The tutorial is using movie review data and the only two categories are positive or negative so the model only uses an output layer with activation set to 'sigmoid'.<\/p>\n\n<p>I have 16 categories represented using one-hot encoding.<\/p>\n\n<p><strong>Tutorial Example:<\/strong><\/p>\n\n<pre><code>vocab_size = 10000\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))\n<\/code><\/pre>\n\n<p><strong>My Attempt:<\/strong><\/p>\n\n<pre><code>model.add(keras.layers.Embedding(10000, 16))\nmodel.add(keras.layers.GlobalAveragePooling1D())\nmodel.add(keras.layers.Dense(16, activation=tf.nn.relu))\nmodel.add(keras.layers.Dense(16, activation='softmax'))\n\n\nmodel.compile(optimizer='adam',\n              loss='binary_crossentropy',\n              metrics=['acc'])\n\nhistory = model.fit(x_train[:5000],\n                y_train[:5000],\n                epochs=1,\n                batch_size=256,\n                validation_data=(x_train[5000:], y_train[5000:]),\n                verbose=1)\n<\/code><\/pre>\n\n<p><strong>Error:<\/strong>\nValueError: A target array with shape (5000, 1) was passed for an output of shape (None, 16) while using as loss <code>binary_crossentropy<\/code>. This loss expects targets to have the same shape as the output.<\/p>\n\n<p><strong>Data Shapes:<\/strong><\/p>\n\n<pre><code>x_train[:5000] (5000, 2000)\ny_train[:5000] (5000,16)\nx_train[5000:] (1934, 2000)\ny_train[5000:] (1934,16)\n<\/code><\/pre>\n\n<p><strong>Model Summary:<\/strong>\nModel: \"sequential_16\"<\/p>\n\n<pre><code>Layer (type)                 Output Shape              Param #   \nembedding_15 (Embedding)     (None, None, 16)          160000    \nglobal_average_pooling1d_15  (None, 16)                0         \ndense_30 (Dense)             (None, 16)                272       \ndense_31 (Dense)             (None, 16)                272\n<\/code><\/pre>\n\n<p>Total params: 160,544<\/p>\n\n<p>Trainable params: 160,544<\/p>\n\n<p>Non-trainable params: 0<\/p>\n","answers":[{"AnswerId":"58507349","CreationDate":"2019-10-22T15:12:50.717","ParentId":null,"OwnerUserId":"6410450","Title":null,"Body":"<p>Binary_crossentropy is for binary classification, but what you're looking for is categorical_crossentropy.  Binary_crossentropy expects your y matrix to be (n_samples x 1), with values of 0 or 1.  Categorical_crossentropy expects your y matrix to be (n_samples x n_categories), with the correct category labeled as 1 and the other categories labeled as 0.  It sounds like the way you did one-hot-encoding would be correct, so you probably just need to change your loss function.<\/p>\n\n<pre><code>model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['acc'])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57702016,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning>","CreationDate":"2019-08-29T02:00:13.107","AcceptedAnswerId":57703571.0,"OwnerUserId":11920225.0,"Title":"how to change input_shape to one channel like (64,64,1) from my code?","Body":"<p>so I have a dataset in the form of a grayscale image, that means one channel right?<\/p>\n\n<p>I changed input_shape to 64.64.1 from this code<\/p>\n\n<pre><code>classifier.add(ZeroPadding2D(padding=(2,2), input_shape = (64, 64, 1)))\n<\/code><\/pre>\n\n<p>but i got error<\/p>\n\n<pre><code>Error when checking input: expected zero_padding2d_1_input to have shape (64, 64, 1) but got array with shape (64, 64, 3)\n<\/code><\/pre>\n\n<p>what's wrong with that and how can I run training with one channel input_shape (64,64,1)?<\/p>\n\n<p>full code can you see on this <a href=\"https:\/\/stackoverflow.com\/questions\/57635645\/why-my-acc-always-higher-but-my-val-acc-is-very-small\">Why my acc always higher but my val_acc is very small?<\/a><\/p>\n","answers":[{"AnswerId":"57703571","CreationDate":"2019-08-29T05:33:30.970","ParentId":null,"OwnerUserId":"4578111","Title":null,"Body":"<p>The primary reason is all your images are RGB and are read as RGB images by keras flow from directory by default. So set the flow_from_directory to read your images as grayscale and set the input_shape in the first layer to (64,64,1)<\/p>\n\n<p>code sample for flow_from_directory<\/p>\n\n<pre><code> train_datagen.flow_from_directory('dataset\/training_set',\n                                                 target_size = (64, 64),\n                                                 batch_size = 32,\n                                                 class_mode = 'binary',\n                                                 color_mode = \"grayscale\")\n<\/code><\/pre>\n\n<p>)<\/p>\n"}]}
+{"QuestionId":57702112,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><recurrent-neural-network><seq2seq>","CreationDate":"2019-08-29T02:16:11.260","AcceptedAnswerId":null,"OwnerUserId":11839345.0,"Title":"Can I provide multiple targets to a seq2seq model?","Body":"<p>I'm doing video captioning on MSR-VTT dataset<\/p>\n\n<p>In this dataset, I've got 10,000 videos and, <strong>for each videos, I've got 20 different captions<\/strong>.<\/p>\n\n<p>My model consists of a seq2seq RNN. Encoder's inputs are the videos features, decoder's inputs are embedded target captions and decoder's output are predicted captions.<\/p>\n\n<p><strong>I'm wondering if using several time the same videos with different captions is useful, or not.<\/strong><\/p>\n\n<hr>\n\n<p>Since I couldn't find explicit info, I tried to benchmark it<\/p>\n\n<h2>Benchmark:<\/h2>\n\n<h3>Model 1: One caption for each video<\/h3>\n\n<p>I trained it on 1108 sport videos, with a batch size of 5, over 60 epochs.  This configuration takes about 211 seconds per epochs.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Epoch 1\/60 ; Batch loss: 5.185806 ; Batch accuracy: 14.67% ; Test accuracy: 17.64%\nEpoch 2\/60 ; Batch loss: 4.453338 ; Batch accuracy: 18.51% ; Test accuracy: 20.15%\nEpoch 3\/60 ; Batch loss: 3.992785 ; Batch accuracy: 21.82% ; Test accuracy: 54.74%\n...\nEpoch 10\/60 ; Batch loss: 2.388662 ; Batch accuracy: 59.83% ; Test accuracy: 58.30%\n...\nEpoch 20\/60 ; Batch loss: 1.228056 ; Batch accuracy: 69.62% ; Test accuracy: 52.13%\n...\nEpoch 30\/60 ; Batch loss: 0.739343; Batch accuracy: 84.27% ; Test accuracy: 51.37%\n...\nEpoch 40\/60 ; Batch loss: 0.563297 ; Batch accuracy: 85.16% ; Test accuracy: 48.61%\n...\nEpoch 50\/60 ; Batch loss: 0.452868 ; Batch accuracy: 87.68% ; Test accuracy: 56.11%\n...\nEpoch 60\/60 ; Batch loss: 0.372100 ; Batch accuracy: 91.29% ; Test accuracy: 57.51%\n<\/code><\/pre>\n\n<h3>Model 2: 12 captions for each video<\/h3>\n\n<p>Then I trained the <strong>same<\/strong> 1108 sport videos, with a batch size of 64.<br>\nThis configuration takes about 470 seconds per epochs.<\/p>\n\n<p>Since I've 12 captions for each videos, the total number of samples in my dataset is 1108*12.<br>\nThat's why I took this batch size (64 ~= 12*old_batch_size). So the two models launch the optimizer the same number of times.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Epoch 1\/60 ; Batch loss: 5.356736 ; Batch accuracy: 09.00% ; Test accuracy: 20.15%\nEpoch 2\/60 ; Batch loss: 4.435441 ; Batch accuracy: 14.14% ; Test accuracy: 57.79%\nEpoch 3\/60 ; Batch loss: 4.070400 ; Batch accuracy: 70.55% ; Test accuracy: 62.52%\n...\nEpoch 10\/60 ; Batch loss: 2.998837 ; Batch accuracy: 74.25% ; Test accuracy: 68.07%\n...\nEpoch 20\/60 ; Batch loss: 2.253024 ; Batch accuracy: 78.94% ; Test accuracy: 65.48%\n...\nEpoch 30\/60 ; Batch loss: 1.805156 ; Batch accuracy: 79.78% ; Test accuracy: 62.09%\n...\nEpoch 40\/60 ; Batch loss: 1.449406 ; Batch accuracy: 82.08% ; Test accuracy: 61.10%\n...\nEpoch 50\/60 ; Batch loss: 1.180308 ; Batch accuracy: 86.08% ; Test accuracy: 65.35%\n...\nEpoch 60\/60 ; Batch loss: 0.989979 ; Batch accuracy: 88.45% ; Test accuracy: 63.45%\n<\/code><\/pre>\n\n<p>Here is the intuitive representation of my datasets:<\/p>\n\n<p><img src=\"https:\/\/zupimages.net\/up\/19\/35\/u6wj.png\" alt=\"Model 1 and Model 2\" title=\"Dataset illustration\"><\/p>\n\n<hr>\n\n<h2>How can I interprete this results ?<\/h2>\n\n<p>When I manually looked at the test predictions,  Model 2 predictions looked more accurate than Model 1 ones.<\/p>\n\n<p>In addition, I used a batch size of 64 for Model 2. That means that I could obtain even more good results by choosing a smaller batch size. It seems I can't have better training method for Mode 1 since batch size is already very low<\/p>\n\n<p>On the other hand, Model 1 have better loss and training accuracy results...<\/p>\n\n<p><strong>What should I conclude ?<\/strong><br>\n<strong>Does the Model 2 constantly overwrites the previously trained captions with the new ones instead of adding new possible captions ?<\/strong><\/p>\n","answers":[{"AnswerId":"57734595","CreationDate":"2019-08-31T01:30:56.407","ParentId":null,"OwnerUserId":"4607955","Title":null,"Body":"<blockquote>\n  <p>I'm wondering if using several time the same videos with different captions is useful, or not.<\/p>\n<\/blockquote>\n\n<p>I think it is definately. It can be interpreted as video to captions is <strong>not one-to-one mapped<\/strong>. And thus weights gets trained more based on the video context.<\/p>\n\n<p>Since the video to caption is not one to one mapped. So even if the neural network is indefinitely dense it should <strong>never<\/strong> achieve 100% training accuracy(or loss as zero) thus reducing overfitting significantly.<\/p>\n\n<blockquote>\n  <p>When I manually looked at the test predictions, Model 2 predictions looked more accurate than Model 1 ones.<\/p>\n<\/blockquote>\n\n<p>Nice! Same is visible here:<\/p>\n\n<pre><code>Model1; Batch accuracy: 91.29% ; Test accuracy: 57.51%\nModel2; Batch accuracy: 88.45% ; Test accuracy: 63.45%\n        Increasing Generalization!!\n<\/code><\/pre>\n\n<blockquote>\n  <p>In addition, I used a batch size of 64 for Model 2. That means that I could obtain even more good results by choosing a smaller batch size. It seems I can't have better training method for Mode 1 since batch size is already very low.<\/p>\n<\/blockquote>\n\n<p>I might not be the right person to comment on the value of the batch_size here, but increasing it a bit more should be worth a try.<\/p>\n\n<p>batch_size is a balance between moving the previous knowledge towards current batch(trying to converge in different directions after some time based on the learning rate) vs trying to learn similar knowledge again and again(converging in almost same direction).<\/p>\n\n<p>And remember there are lot of other ways to improve the results.<\/p>\n\n<blockquote>\n  <p>On the other hand, Model 1 have better loss and training accuracy results...\n  What should I conclude ? ?<\/p>\n<\/blockquote>\n\n<p>Training accuracy results and loss value tells about how the model is performing on the training data not on the validation\/test data. In other words, having very small value of loss might mean memorization.<\/p>\n\n<blockquote>\n  <p>Does the Model 2 constantly overwrites the previously trained captions with the new ones instead of adding new possible captions.<\/p>\n<\/blockquote>\n\n<p>Depends on how data is being splitted into batches.\nIs multiple caption of the same video in same batch or spreaded over multiple batches.<\/p>\n\n<p>Remember, Model 2 has multiple caption which might be a major factor behind generalization. Thus increasing the training loss value.<\/p>\n\n<p>Thanks!<\/p>\n"}]}
+{"QuestionId":57702154,"AnswerCount":0,"Tags":"<python><database><parsing><tensorflow><tfrecord>","CreationDate":"2019-08-29T02:22:38.827","AcceptedAnswerId":null,"OwnerUserId":11058721.0,"Title":"How do I parse tfrecords?","Body":"<p>Following this <a href=\"https:\/\/colab.research.google.com\/github\/GoogleCloudPlatform\/training-data-analyst\/blob\/master\/courses\/fast-and-lean-data-science\/07_Keras_Flowers_TPU_squeezenet.ipynb#scrollTo=qGYZ2uOxLgkr\" rel=\"nofollow noreferrer\">tutorial<\/a> on parsing <code>tfrecords<\/code> I stumbled upon an error while running <code>model.fit<\/code>:<br>\n<code>AbortedError: Session 581cc3040540fb5e is not found.<\/code><br>\nFollowing <a href=\"https:\/\/stackoverflow.com\/questions\/56608850\/session-lost-with-keras-and-tpus-in-google-colab\">this<\/a> Stack question I think there is a problem with my <code>tfrecords<\/code> parsing.<br>\nMy current code to view the records:  <\/p>\n\n<pre><code>count = 0\nfor example in tf.python_io.tf_record_iterator(train_filenames[0]):\n    print(tf.train.Example.FromString(example))\n    count +=1\n    if count &gt;5:\n        break\n<\/code><\/pre>\n\n<p>Outputs:<\/p>\n\n<pre><code>features {\n  feature {\n    key: \"filename\"\n    value {\n      bytes_list {\n        value: \"BM.jpg\"\n      }\n    }\n  }\n  feature {\n    key: \"f\"\n    value {\n      bytes_list {\n        value: \"\\211P\n\n}\n    }\n  }\n  feature {\n    key: \"label\"\n    value {\n      int64_list {\n        value: 0\n      }\n    }\n  }\n}\n<\/code><\/pre>\n\n<p>The code that I'm running to parse it is:  <\/p>\n\n<pre><code>AUTO = tf.data.experimental.AUTOTUNE\nimage_size = [224,224]\ndef read_tfrecord(tf_record):\n    features = {\n        \"filename\": tf.io.FixedLenFeature([], tf.string), # tf.string means bytestring\n        \"fundus\": tf.io.FixedLenFeature([], tf.string),  \n        \"label\": tf.io.VarLenFeature(tf.float32),\n    }\n    tf_record = tf.parse_single_example(tf_record, features)\n\n    filename = tf.image.decode_jpeg(tf_record['filename'], channels=3)\n    filename = tf.cast(filename, tf.float32) \/ 255.0  # convert image to floats in [0, 1] range\n    filename = tf.reshape(filename, [*image_size, 3]) # explicit size will be needed for TPU\n\n    label = tf.cast(tf_record['label'],tf.float32)\n\n    return filename, label\n\ndef load_dataset(filenames):\n    option_no_order = tf.data.Options()\n    option_no_order.experimental_deterministic = False\n\n    dataset = tf.data.Dataset.from_tensor_slices(filenames)\n    dataset = dataset.with_options(option_no_order)\n    #dataset = tf.data.TFRecordDataset(filenames, num_parallel_reads=16)\n    dataset = dataset.interleave(tf.data.TFRecordDataset, cycle_length=32, num_parallel_calls=AUTO) # faster\n    dataset = dataset.map(read_tfrecord, num_parallel_calls=AUTO)\n    return dataset\n<\/code><\/pre>\n\n<p>Is there a problem with the way I parse it?<\/p>\n","answers":[]}
+{"QuestionId":57702247,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2019-08-29T02:36:37.533","AcceptedAnswerId":57702783.0,"OwnerUserId":8176285.0,"Title":"Fintuning custom keras model","Body":"<p>I have a keras model which is trained on 5 classes,The final layers of the model look like so<\/p>\n\n<pre><code>dr_steps = Dropout(0.25)(Dense(128, activation = 'relu')(gap_dr))\nout_layer = Dense(5, activation = 'softmax')(dr_steps)\nmodel = Model(inputs = [in_lay], outputs = [out_layer])\n<\/code><\/pre>\n\n<p>What i want to do is fine tune this model on an 8 class multilabel problem but i am not sure how to achieve this,this is what i have tried<\/p>\n\n<pre><code>dr_steps = Dropout(0.25)(Dense(128, activation = 'relu')(gap_dr))\nout_layer = Dense(t_y.shape[-1], activation = 'softmax')(dr_steps)\nmodel = Model(inputs = [in_lay], outputs = [out_layer])\nweights_path = 'weights.best.hdf5'\nretina_model.load_weights(weights_path)\nmodel.layers.pop()\noutput = Dense(8, activation = 'sigmoid')(model.layers[-1].output)\nmodel = Model(inputs = [in_lay], outputs = [output])\nloss = 'binary_crossentropy'\nmodel.compile(optimizer = RAdam(), loss = FocalLoss,\n                         metrics = [\"binary_accuracy\",precision, recall,auc])\n<\/code><\/pre>\n\n<p>but this will raise an error like this<\/p>\n\n<pre><code>raise ValueError(str(e))\nValueError: Dimension 1 in both shapes must be equal, but are 8 and 5. Shapes are [128,8] and [128,5]. for 'Assign_390' (op: 'Assign') with input shapes: [128,8], [128,5].\n<\/code><\/pre>\n\n<p>Any suggestions on how to fine tune this model will be very helpful,Thanks in advance.<\/p>\n","answers":[{"AnswerId":"57702783","CreationDate":"2019-08-29T03:58:33.770","ParentId":null,"OwnerUserId":"4257991","Title":null,"Body":"<p>Here,<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Model(inputs = [in_lay], outputs = [out_layer])\nweights_path = 'weights.best.hdf5'\n<\/code><\/pre>\n\n<p>this <strong>out_layer<\/strong> should have the same dimension(5 classes) described inside <strong>weights.best.hdf5<\/strong>. <\/p>\n\n<p>So, <code>t_y.shape[-1]<\/code> should be <code>5<\/code> dimensional, <em>not 8<\/em>. <\/p>\n"}]}
+{"QuestionId":57702915,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2019-08-29T04:18:40.540","AcceptedAnswerId":null,"OwnerUserId":11813916.0,"Title":"ValueError: Output of generator should be a tuple `(x, y, sample_weight)` or `(x, y)`. Found: [array([[[[255, 255, 255],","Body":"<p>we want to Convert an existing Keras SSD7 code to tf.keras. We currently found the mentioned error. We used a custom data_generator in this. Also we want to know what changes we have to do when converting from keras to tf.keras. \nGithub repository code taken from : <a href=\"https:\/\/github.com\/pierluigiferrari\/ssd_keras\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pierluigiferrari\/ssd_keras<\/a><\/p>\n\n<p>We changed the following lines in necessary places.<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python.keras.optimizers import Adam\nfrom tensorflow.python.keras.callbacks import ModelCheckpoint, EarlyStopping, ReduceLROnPlateau, TerminateOnNaN, CSVLogger\nfrom tensorflow.python.keras import backend as K\nfrom tensorflow.python.keras.models import load_model\n<\/code><\/pre>\n\n<p>A custom data class is used as follow : <a href=\"https:\/\/github.com\/pierluigiferrari\/ssd_keras\/blob\/master\/data_generator\/object_detection_2d_data_generator.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pierluigiferrari\/ssd_keras\/blob\/master\/data_generator\/object_detection_2d_data_generator.py<\/a> <\/p>\n\n<p>We found this error in the process of converting<\/p>\n\n<pre><code>ValueError: Output of generator should be a tuple (x, y, sample_weight) or (x, y). Found: [array([[[[121, 135, 161],\n        [114, 130, 155],\n        [109, 125, 151],\n        ...,\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57703276,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-29T05:03:01.087","AcceptedAnswerId":null,"OwnerUserId":10083113.0,"Title":"I can't use someone else's model to generate inputs to mine, in a generator. How do I fix this?","Body":"<p>I am trying to train a neural network on the SQuAD v1.1 dataset, using a pretrained BERT model. Someone suggested that I first grab the output of the BERT model and then feed those into my neural network as inputs. Due to the large amount of data, I feel that I need to create a generator, on which my neural network can then fit:<\/p>\n\n<pre><code># @title Preparation\n!pip install -q keras-bert\n!wget -q https:\/\/storage.googleapis.com\/bert_models\/2018_10_18\/uncased_L-12_H-768_A-12.zip\n!unzip -o uncased_L-12_H-768_A-12.zip\n<\/code><\/pre>\n\n<pre><code>import os\n<\/code><\/pre>\n\n<pre><code>pretrained_path = 'uncased_L-12_H-768_A-12'\nconfig_path = os.path.join(pretrained_path, 'bert_config.json')\ncheckpoint_path = os.path.join(pretrained_path, 'bert_model.ckpt')\nvocab_path = os.path.join(pretrained_path, 'vocab.txt')\n<\/code><\/pre>\n\n<pre><code># TF_KERAS must be added to environment variables in order to use TPU\nos.environ['TF_KERAS'] = '1'\nimport codecs\nfrom keras_bert import load_trained_model_from_checkpoint\n<\/code><\/pre>\n\n<pre><code>token_dict = {}\nwith codecs.open(vocab_path, 'r', 'utf8') as reader:\n    for line in reader:\n        token = line.strip()\n        token_dict[token] = len(token_dict)\n<\/code><\/pre>\n\n<pre><code>model = load_trained_model_from_checkpoint(config_path, checkpoint_path)\nimport numpy as np\nfrom keras_bert import Tokenizer\ntokenizer = Tokenizer(token_dict)\ndef tokenize(text):\n  tokens = tokenizer.tokenize(text)\n  indices, segments = tokenizer.encode(first=text, max_len=512)\n  return indices,segments\ndef feature_extraction(texts):\n  return_values = []\n  for text_ in texts:\n    try:\n      text_.split(\" \")\n    except AttributeError as e:\n      raise TypeError(\"Expected array of strings.\")\n    try:\n      indices,segments = tokenize(text_)\n      predicts = model.predict([np.array([indices] * 8), np.array([segments] * 8)])[0]\n      return_values.append(predicts)\n    except ValueError as v:\n      print(v)\n  return_values = np.array(return_values)\n  return return_values\nprint(feature_extraction(text_array).shape)\ndef batch_generator(dataframe,batch_size):\n  while True:\n    batch = dataframe.sample(n=batch_size)\n    try:\n      batch_features = feature_extraction(batch[\"question\"].values)\n    except ValueError as v:\n      print(\"Oops, I'm getting a ValueError for batch_features.\")\n      print(v)\n    try:\n      batch_targets = batch[\"answer_start\"]\n    except ValueError as v:\n      print(\"Oops, I'm getting a ValueError for batch_targets.\")\n      print(v)\n\n    yield batch_features,batch_targets\n<\/code><\/pre>\n\n<p>This works when I feed it test data:<\/p>\n\n<pre><code>def batch_generator(dataframe,batch_size):\n  while True:\n    batch = dataframe.sample(n=batch_size)\n    try:\n      batch_features = feature_extraction(batch[\"question\"].values)\n    except ValueError as v:\n      print(\"Oops, I'm getting a ValueError for batch_features.\")\n      print(v)\n    try:\n      batch_targets = batch[\"answer_start\"]\n    except ValueError as v:\n      print(\"Oops, I'm getting a ValueError for batch_targets.\")\n      print(v)\n\n    yield batch_features,batch_targets\n<\/code><\/pre>\n\n<p>This works when I use this test code:  <\/p>\n\n<pre><code>testDataframe = pd.DataFrame({\"question\":[\"Does she sell seashells by the seashore?\"],\"answer\":[\"She sells seashells by the seashore\"],\"answer_start\":[0]})\nfor x,y in batch_generator(testDataframe,1):\n  print (x)\n  print (y)\n  break\n<\/code><\/pre>\n\n<p><strong>Output:<\/strong><\/p>\n\n<blockquote>\n  <p>[[[-0.11251544 -0.09277309  0.04996187 ... -0.43535435  0.23852573\n      0.3206718 ]   [ 0.35688528  0.43881682 -0.1390086  ... -0.32458037  0.64422214    -0.11743623]   [ 0.6213926  -0.9945548   0.07564903 ... -0.87357795  0.2069801    -0.25303575]   ...   [-0.06796454 -0.24819699 -0.25508618 ...  0.20477912  0.36703664\n      0.04691853]   [ 0.15030818 -0.05989693  0.17198643 ...  0.19960165  0.0324061    -0.31075317]   [ 0.05091426 -0.14167279  0.18194658 ...  0.12112649  0.05029908    -0.15253511]]] 0    0 Name: answer_start, dtype: int64<\/p>\n<\/blockquote>\n\n<p>I create and compile my neural network and inputs like so:<\/p>\n\n<pre><code>import json\nimport re\n#regex = re.compile(r'\\W+')\nimport json\nimport re\n#regex = re.compile(r'\\W+')\ndef readFile(filename):\n  with open(filename) as file:\n    fields = []\n    JSON = json.loads(file.read())\n    articles = []\n    for article in JSON[\"data\"]:\n      articleTitle = article[\"title\"]\n      article_body = []\n      for paragraph in article[\"paragraphs\"]:\n        paragraphContext = paragraph[\"context\"]\n        article_body.append(paragraphContext)\n        for qas in paragraph[\"qas\"]:\n          question = qas[\"question\"]\n          answer = qas[\"answers\"][0]\n          fields.append({\"question\":question,\"answer_text\":answer[\"text\"],\"answer_start\":answer[\"answer_start\"],\"paragraph_context\":paragraphContext,\"article_title\":articleTitle})\n      article_body = \"\\\\n\".join(article_body)\n      article = {\"title\":articleTitle,\"body\":article_body}\n      articles.append(article)\n  fields = pd.DataFrame(fields)\n  #fields[\"question\"] = fields[\"question\"].str.replace(regex,\" \")\n  assert not (fields[\"question\"].str.contains(\"catalanswhat\").any())\n  #fields[\"paragraph_context\"] = fields[\"paragraph_context\"].str.replace(regex,\" \")\n  #fields[\"answer_text\"] = fields[\"answer_text\"].str.replace(regex,\" \")\n  assert not (fields[\"paragraph_context\"].str.contains(\"catalanswhat\").any())\n  fields[\"article_title\"] = fields[\"article_title\"].str.replace(\"_\",\" \")\n  assert not (fields[\"article_title\"].str.contains(\"catalanswhat\").any())\n  return fields,articles\n<\/code><\/pre>\n\n<pre><code>trainingData,trainingArticles = readFile(\"train-v1.1.json\")\n<\/code><\/pre>\n\n<pre><code>answers_network = Sequential()\nanswers_network.add(Dense(32,input_shape=(512,768)))\nanswers_network.summary()\nanswers_network.compile(\"rmsprop\",\"categorical_crossentropy\")\nanswers_network_checkpoint = ModelCheckpoint('answers_network-rnn-best.h5', verbose=1, monitor='val_loss',save_best_only=True, mode='auto')\nanswers_network.fit_generator(batch_generator(trainingData[[\"question\",\"paragraph_context\",\"answer_start\"]],100),steps_per_epoch=8)\n<\/code><\/pre>\n\n<p>This fails with an error:<\/p>\n\n<blockquote>\n  <p>Tensor Input-Token:0, specified in either feed_devices or\n  fetch_devices was not found in the Graph<\/p>\n<\/blockquote>\n\n<p>Now, Input-Token is the name of one of the input layers in the BERT model. <\/p>\n\n<p>I think TensorFlow is suggesting that the BERT model is using a different graph than my model.<\/p>\n\n<p>Apparently, the BERT model uses custom layers and activation functions, so making a deep copy of the model may not be the best course of action. <\/p>\n\n<p>What should I do?<\/p>\n\n<p>EDIT: A copy of my train-v1.1.json dataset is available here: <a href=\"https:\/\/drive.google.com\/file\/d\/1qQbrQnH3WkibtXIHFA88gJuGESvyz-Ag\/view?usp=sharing\" rel=\"nofollow noreferrer\">https:\/\/drive.google.com\/file\/d\/1qQbrQnH3WkibtXIHFA88gJuGESvyz-Ag\/view?usp=sharing<\/a><\/p>\n","answers":[{"AnswerId":"57764306","CreationDate":"2019-09-03T02:20:41.463","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<h2>Option 1<\/h2>\n\n<p>Generate the data first and save it. \nLater you train with the saved data.<\/p>\n\n<p>It seems like doing this:<\/p>\n\n<pre><code>features = feature_extraction(text_array)\nnp.save('features.npy', features) #or not... \n<\/code><\/pre>\n\n<p>Then just fit with this array later:<\/p>\n\n<pre><code>features = np.load('features.npy')    \nnew_model.fit(features, targets, ...)\n<\/code><\/pre>\n\n<hr>\n\n<p>If the data is too big for your memory (there is no problem with the model, just set a proper <code>batch_size<\/code>. This is also true for the bert model when generating features), I mean, the whole data can't fit your memory, then you could save the batches individually:<\/p>\n\n<pre><code>for i in range(batches):\n    batch = text_array[i*batch_size : (i+1)*batch_size)\n    features = feature_extraction(batch)\n    np.save('batch' + str(i) + '.npy', features)\n<\/code><\/pre>\n\n<p>Your generator then loads these batches:<\/p>\n\n<pre><code>while True:\n    for i in range(batches):\n        batch = np.load('batch' + str(i) + '.npy')\n<\/code><\/pre>\n\n<h2>Option 2<\/h2>\n\n<p>Make the models use the same graph, creating a single big model:<\/p>\n\n<pre><code>bertInputs = Input(shape_for_bert_input)\nbertOutputs = bert_model(bertInputs)\nyourOutputs = your_model(bertOutputs)\n\nfullModel = Model(bertInputs, yourOutputs)\n<\/code><\/pre>\n\n<p>Train using the generator directly from the dataframe, the bert prediction will happen automatically as a part of your model. <\/p>\n\n<hr>\n\n<p>Alternatively, if you want to see all layers explicitly in the summary:<\/p>\n\n<pre><code>bertOutputs = bert_model.output\nyourOutputs = Dense(....)(bertOutputs)\n\nfullModel = Model(bert_model.input, yourOutputs)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57703495,"AnswerCount":0,"Tags":"<python><machine-learning><keras><classification><lstm>","CreationDate":"2019-08-29T05:25:11.687","AcceptedAnswerId":null,"OwnerUserId":10704050.0,"Title":"How to use cross validation in LSTM in python","Body":"<p>I have about <code>1000 nodes<\/code> dataset where each node has <code>4 time-series<\/code>. Each time series is exactly 6 length long.The label is <code>0<\/code> or <code>1<\/code>.<\/p>\n\n<p>More precisely my dataset looks as follows.<\/p>\n\n<pre><code>node, time-series1, time_series2, time_series_3, time_series4, Label\nn1, [1.2, 2.5, 3.7, 4.2, 5.6, 8.8], [6.2, 5.5, 4.7, 3.2, 2.6, 1.8], \u2026, 1\nn2, [5.2, 4.5, 3.7, 2.2, 1.6, 0.8], [8.2, 7.5, 6.7, 5.2, 4.6, 1.8], \u2026, 0\nand so on.\n<\/code><\/pre>\n\n<p>I wanted to use LSTM model to make the prediction. I am new to deep learning. So, I am making use of the most basic LSTM model as follows using keras.<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(10, input_shape=(6,4)))\nmodel.add(Dense(32))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n\nprint(data.shape) # (1000, 6, 4)\nmodel.fit(data, target)\n<\/code><\/pre>\n\n<p>However, since my dataset is small, I would like to try <code>10-fold cross validation<\/code> to my LSTM model and get the prediction probability for <code>label 1<\/code> for each <code>node<\/code> and rank the prediction probability in descending order. In other words, I am expecting an output as follows.<\/p>\n\n<pre><code>sorted_probability_of_class_1 = [[n888, 1], [n512, 1], ........, [n99, 0.9], [n112, 0.8], ........, [n882, 0], [n9, 0]]\n<\/code><\/pre>\n\n<hr>\n\n<p>I was able to do a similar task using <code>randomforest<\/code> classifier in <code>sklearn<\/code>. My code is as follows using <code>iris<\/code> dataset which has 150 data points.<\/p>\n\n<pre><code>from sklearn import datasets\n\niris = datasets.load_iris()\niris_names = ['iris_0', 'iris_1', 'iris_2', 'iris_3', 'iris_4', 'iris_5', 'iris_6', 'iris_7', 'iris_8', 'iris_9', 'iris_10', 'iris_11', 'iris_12', 'iris_13', 'iris_14', 'iris_15', 'iris_16', 'iris_17', 'iris_18', 'iris_19', 'iris_20', 'iris_21', 'iris_22', 'iris_23', 'iris_24', 'iris_25', 'iris_26', 'iris_27', 'iris_28', 'iris_29', 'iris_30', 'iris_31', 'iris_32', 'iris_33', 'iris_34', 'iris_35', 'iris_36', 'iris_37', 'iris_38', 'iris_39', 'iris_40', 'iris_41', 'iris_42', 'iris_43', 'iris_44', 'iris_45', 'iris_46', 'iris_47', 'iris_48', 'iris_49', 'iris_50', 'iris_51', 'iris_52', 'iris_53', 'iris_54', 'iris_55', 'iris_56', 'iris_57', 'iris_58', 'iris_59', 'iris_60', 'iris_61', 'iris_62', 'iris_63', 'iris_64', 'iris_65', 'iris_66', 'iris_67', 'iris_68', 'iris_69', 'iris_70', 'iris_71', 'iris_72', 'iris_73', 'iris_74', 'iris_75', 'iris_76', 'iris_77', 'iris_78', 'iris_79', 'iris_80', 'iris_81', 'iris_82', 'iris_83', 'iris_84', 'iris_85', 'iris_86', 'iris_87', 'iris_88', 'iris_89', 'iris_90', 'iris_91', 'iris_92', 'iris_93', 'iris_94', 'iris_95', 'iris_96', 'iris_97', 'iris_98', 'iris_99', 'iris_100', 'iris_101', 'iris_102', 'iris_103', 'iris_104', 'iris_105', 'iris_106', 'iris_107', 'iris_108', 'iris_109', 'iris_110', 'iris_111', 'iris_112', 'iris_113', 'iris_114', 'iris_115', 'iris_116', 'iris_117', 'iris_118', 'iris_119', 'iris_120', 'iris_121', 'iris_122', 'iris_123', 'iris_124', 'iris_125', 'iris_126', 'iris_127', 'iris_128', 'iris_129', 'iris_130', 'iris_131', 'iris_132', 'iris_133', 'iris_134', 'iris_135', 'iris_136', 'iris_137', 'iris_138', 'iris_139', 'iris_140', 'iris_141', 'iris_142', 'iris_143', 'iris_144', 'iris_145', 'iris_146', 'iris_147', 'iris_148', 'iris_149']\n\nX = iris.data\ny = iris.target\n\nclf=RandomForestClassifier(n_estimators=10, random_state = 42, class_weight=\"balanced\")\n\nk_fold = StratifiedKFold(n_splits=10, shuffle=True, random_state=0)\nproba = cross_val_predict(clf, X, y, cv=k_fold, method='predict_proba')\n\n#print probability of class 1\nprint(proba[:,1])\n<\/code><\/pre>\n\n<p>My output is as follows.<\/p>\n\n<pre><code>[0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.1 0.  0.  0.\n 0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.\n 0.2 0.  0.  0.  0.  0.1 0.  0.  0.  0.  0.  0.  0.  0.  0.9 1.  0.7 1.\n 1.  1.  1.  0.7 1.  1.  1.  1.  1.  1.  1.  1.  1.  1.  0.9 0.9 0.1 1.\n 0.6 1.  1.  1.  0.9 0.  1.  1.  1.  1.  1.  0.4 0.9 0.9 1.  1.  1.  0.9\n 1.  1.  1.  1.  1.  1.  1.  1.  1.  1.  0.  0.  0.  0.  0.  0.  0.9 0.\n 0.1 0.  0.  0.  0.  0.  0.  0.  0.1 0.  0.  0.8 0.  0.1 0.  0.1 0.  0.1\n 0.3 0.2 0.  0.6 0.  0.  0.  0.6 0.4 0.  0.  0.  0.8 0.  0.  0.  0.  0.\n 0.  0.  0.  0.  0.  0. ]\n<\/code><\/pre>\n\n<p>This is how I organised the sorted probability of class 1 in the above code.<\/p>\n\n<pre><code>probability_of_class_1 = zip(proba[:, 1], iris_names)\nsorted(merged_list, key=itemgetter(1))\n<\/code><\/pre>\n\n<p>However, I am not clear how to do the same process with my LSTM model above.<\/p>\n\n<p>I am happy to provide more details if needed.<\/p>\n","answers":[]}
+{"QuestionId":57703764,"AnswerCount":0,"Tags":"<python><c++><pytorch><convolution><aten>","CreationDate":"2019-08-29T05:53:11.167","AcceptedAnswerId":null,"OwnerUserId":4615946.0,"Title":"How to implement a variant of convolution operation using low-level pytorch APIs?","Body":"<p>My target is to implement a variant of convolution operation in pytorch which requires high inference speed.<\/p>\n\n<p>I have implemented one in C++ and CUDA through pytorch CPP extension. However, the speed is not satisfactory, I guess I\u2019ll have to implement it with lower level APIs. It seems <code>aten\/src\/ATen\/native\/Convolution.cpp<\/code> is the right place to do so, and it further calls <code>aten\/src\/THNN\/generic\/SpatialConvolutionMM.c<\/code>, so I modify <code>SpatialConvolutionMM.c<\/code>. The problem is after compiling, I cannot see any changes on the outputs. Maybe the files in THNN are wrapped in <code>ATen\/native<\/code> in some way, I just cannot figure it out. Could anyone give me some suggestions? Thank you.<\/p>\n","answers":[]}
+{"QuestionId":57703808,"AnswerCount":2,"Tags":"<python><callback><pytorch><tensor>","CreationDate":"2019-08-29T05:56:30.757","AcceptedAnswerId":57837526.0,"OwnerUserId":610569.0,"Title":"How do I pass a keyword argument to the forward used by a pre-forward hook?","Body":"<p>Given a torch's <code>nn.Module<\/code> with a pre-forward hook, e.g. <\/p>\n\n<pre><code>import torch\nimport torch.nn as nn\n\nclass NeoEmbeddings(nn.Embedding):\n    def __init__(self, num_embeddings:int, embedding_dim:int, padding_idx=-1):\n        super().__init__(num_embeddings, embedding_dim, padding_idx)\n        self.register_forward_pre_hook(self.neo_genesis)\n\n    @staticmethod\n    def neo_genesis(self, input, higgs_bosson=0):\n        if higgs_bosson:\n            input = input + higgs_bosson\n        return input\n<\/code><\/pre>\n\n<p>It's possible to let an input tensor go through some manipulation before going to the actual <code>forward()<\/code> function, e.g. <\/p>\n\n<pre><code>&gt;&gt;&gt; x = NeoEmbeddings(10, 5, 1)\n&gt;&gt;&gt; x.forward(torch.tensor([0,2,5,8]))\ntensor([[-1.6449,  0.5832, -0.0165, -1.3329,  0.6878],\n        [-0.3262,  0.5844,  0.6917,  0.1268,  2.1363],\n        [ 1.0772,  0.1748, -0.7131,  0.7405,  1.5733],\n        [ 0.7651,  0.4619,  0.4388, -0.2752, -0.3018]],\n       grad_fn=&lt;EmbeddingBackward&gt;)\n\n&gt;&gt;&gt; print(x._forward_pre_hooks)\nOrderedDict([(25, &lt;function NeoEmbeddings.neo_genesis at 0x1208d10d0&gt;)])\n<\/code><\/pre>\n\n<p><strong>How could we pass the arguments (<code>*args<\/code> or <code>**kwargs<\/code>) that the pre-forward hook needs but not accepted by the default <code>forward()<\/code> function?<\/strong><\/p>\n\n<p>Without modification\/overriding the <code>forward()<\/code> function, this is not possible:<\/p>\n\n<pre><code>&gt;&gt;&gt; x = NeoEmbeddings(10, 5, 1)\n&gt;&gt;&gt; x.forward(torch.tensor([0,2,5,8]), higgs_bosson=2)\n\n----------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-102-8705a40a3cc2&gt; in &lt;module&gt;\n      1 x = NeoEmbeddings(10, 5, 1)\n----&gt; 2 x.forward(torch.tensor([0,2,5,8]), higgs_bosson=2)\n\nTypeError: forward() got an unexpected keyword argument 'higgs_bosson'\n<\/code><\/pre>\n","answers":[{"AnswerId":"57749944","CreationDate":"2019-09-01T22:36:27.137","ParentId":null,"OwnerUserId":"3767239","Title":null,"Body":"<p>Since a <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.Module.register_forward_pre_hook\" rel=\"nofollow noreferrer\">forward pre-hook<\/a> is called with only the tensor by definition, a keyword argument doesn't make much sense here. What would make more sense is to use an instance attribute for example:<\/p>\n\n<pre><code>def neo_genesis(self, input):\n    if self.higgs_bosson:\n        input = input + self.higgs_bosson\n    return input\n<\/code><\/pre>\n\n<p>Then you can switch that attribute as appropriate. You could also use a context manager for that:<\/p>\n\n<pre><code>from contextlib import contextmanager\n\n@contextmanager\ndef HiggsBoson(module):\n    module.higgs_boson = 1\n    yield\n    module.higgs_boson = 0\n\nwith HiggsBoson(x):\n    x.forward(...)\n<\/code><\/pre>\n\n<p>If you have that function already and you really need to change that parameter you can still replace the function's <code>__defaults__<\/code> attribute:<\/p>\n\n<pre><code>x.neo_genesis.__defaults__ = (1,)  # this corresponds to `higgs_boson` parameter\nx.forward(...)\nx.neo_genesis.__defaults__ = (0,)  # reset to default\n<\/code><\/pre>\n"},{"AnswerId":"57837526","CreationDate":"2019-09-07T21:16:37.677","ParentId":null,"OwnerUserId":"10886420","Title":null,"Body":"<h1>Torchscript incompatible (as of <code>1.2.0<\/code>)<\/h1>\n\n<p>First of all, your example <code>torch.nn.Module<\/code> has some minor mistakes (probably by an accident).<\/p>\n\n<p>Secondly, you can pass <strong>anything<\/strong> to forward and <code>register_forward_pre_hook<\/code> will just get the argument that will be passed your your <code>torch.nn.Module<\/code> (be it layer or model or anything) else. You indeed <strong>cannot do it<\/strong> without modifying <code>forward<\/code> call, but why would you want to avoid that? You could simply forward the arguments to base function as can be seen below:<\/p>\n\n<pre><code>import torch\n\n\nclass NeoEmbeddings(torch.nn.Embedding):\n    def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx=-1):\n        super().__init__(num_embeddings, embedding_dim, padding_idx)\n        self.register_forward_pre_hook(NeoEmbeddings.neo_genesis)\n\n    # First argument should be named something like module, as that's what \n    # you are registering this hook to\n    @staticmethod\n    def neo_genesis(module, inputs):  # No need for self as first argument\n        net_input, higgs_bosson = inputs  # Simply unpack tuple here\n        return net_input\n\n    def forward(self, inputs, higgs_bosson):\n        # Do whatever you want here with both arguments, you can ignore \n        # higgs_bosson if it's only needed in the hook as done here\n        return super().forward(inputs)\n\n\nif __name__ == \"__main__\":\n    x = NeoEmbeddings(10, 5, 1)\n    # You should call () instead of forward so the hooks register appropriately\n    print(x(torch.tensor([0, 2, 5, 8]), 1))\n<\/code><\/pre>\n\n<p>You can't do it in more succinct way, but the limitation is base's class <code>forward<\/code> method, not the hook itself (and tbh I wouldn't want it to be more succinct as it would become unreadable IMO).<\/p>\n\n<h2>Torchscript compatible<\/h2>\n\n<p>If you want to use torchscript (tested on <code>1.2.0<\/code>) you could use composition instead of inheritance. All you have to change are merely two lines and your code may look something like this:<\/p>\n\n<pre><code>import torch\n\n# Inherit from Module and register embedding as submodule\nclass NeoEmbeddings(torch.nn.Module):\n    def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx=-1):\n        super().__init__()\n        # Just use it as a container inside your own class\n        self._embedding = torch.nn.Embedding(num_embeddings, embedding_dim, padding_idx)\n        self.register_forward_pre_hook(NeoEmbeddings.neo_genesis)\n\n    @staticmethod\n    def neo_genesis(module, inputs):\n        net_input, higgs_bosson = inputs\n        return net_input\n\n    def forward(self, inputs: torch.Tensor, higgs_bosson: torch.Tensor):\n        return self._embedding(inputs)\n\n\nif __name__ == \"__main__\":\n    x = torch.jit.script(NeoEmbeddings(10, 5, 1))\n    # All arguments must be tensors in torchscript\n    print(x(torch.tensor([0, 2, 5, 8]), torch.tensor([1])))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57703853,"AnswerCount":1,"Tags":"<python><tensorflow><cuda><gpu><nvidia>","CreationDate":"2019-08-29T05:59:35.703","AcceptedAnswerId":null,"OwnerUserId":11986956.0,"Title":"GPU control using CUDA in Python","Body":"<p>I want to use Tensorflow-gpu in Python.\nMy PC has two GPUs. (GPU-A and GPU-B)<\/p>\n\n<p>I want to specify the GPU to be used.<\/p>\n\n<p>So when I run the program, I used the following command:<\/p>\n\n<pre><code>CUDA_VISIBLE_DEVICES = 1\n<\/code><\/pre>\n\n<p>The GPU to be used can be specified according to the value.\nSpecifically, it was assigned as follows.<\/p>\n\n<pre><code>CUDA_VISIBLE_DEVICES = 0\n<\/code><\/pre>\n\n<p>I was able to use GPU-B.<\/p>\n\n<p>On the other hand<\/p>\n\n<pre><code>CUDA_VISIBLE_DEVICES = 1\n<\/code><\/pre>\n\n<p>At that time, I was able to use GPU-A.<\/p>\n\n<p>However, I cannot know in advance whether GPU-A or GPU-B corresponds to the value 0 or 1.\nI found a statement that \"this index is assigned in descending order of computing power\".\nHowever, GPU-A can calculate faster than GPU-B.\nI think that \"indexes are assigned in order of the revision number (Compute Capability), not the calculation speed\".\nIn fact, GPU-B has higher compute capability than GPU-A.<\/p>\n\n<p>If the hypothesis is correct, I have another question.\nIf multiple GPUs with the same Compute Capability are used, how will they be assigned?<\/p>\n","answers":[{"AnswerId":"57704393","CreationDate":"2019-08-29T06:40:49.860","ParentId":null,"OwnerUserId":"1849975","Title":null,"Body":"<p>You can try running <code>nvidia-smi<\/code> in your terminal. The output you get shows you GPU ordering. For example:<\/p>\n\n<pre><code>+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 410.104      Driver Version: 410.104      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 Off |                  N\/A |\n|  0%   52C    P8    21W \/ 250W |      2MiB \/ 11176MiB |      0%      Default |\n+-------------------------------+----------------------+----------------------+\n|   1  GeForce GT 1030     Off  | 00000000:02:00.0  On |                  N\/A |\n| 41%   54C    P0    N\/A \/  30W |   1055MiB \/  2001MiB |      6%      Default |\n+-------------------------------+----------------------+----------------------+\n<\/code><\/pre>\n\n<p>Here, I have 2 GPUs: <\/p>\n\n<ol>\n<li>GPU 0: GTX 1080Ti<\/li>\n<li>GPU 1: GT 1030<\/li>\n<\/ol>\n\n<p>You can then further use indexes in <code>TensorFlow<\/code>.<\/p>\n"}]}
+{"QuestionId":57704412,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><ubuntu><terminal>","CreationDate":"2019-08-29T06:42:22.610","AcceptedAnswerId":null,"OwnerUserId":11737390.0,"Title":"How to suppress \"Future warning\" tensorflow?","Body":"<p><strong>I am running \".\/buildTF.sh\" which uses TensorFlow, on ubuntu terminal. And getting the error as:<\/strong><\/p>\n\n<pre><code>\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/dtypes.py:516:\nFutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated;\nin a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n<\/code><\/pre>\n","answers":[{"AnswerId":"58546467","CreationDate":"2019-10-24T17:30:33.693","ParentId":null,"OwnerUserId":"12198627","Title":null,"Body":"<p>This is a warning message which comes because of numpy version, uninstall the current numpy version and update it to <code>1.16.4<\/code>.<\/p>\n\n<pre><code># pip uninstall numpy \n# pip install numpy==1.16.4\n<\/code><\/pre>\n\n<p>Thanks to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/31249\" rel=\"nofollow noreferrer\">ymodak<\/a><\/p>\n"}]}
+{"QuestionId":57704413,"AnswerCount":1,"Tags":"<loops><tensorflow>","CreationDate":"2019-08-29T06:42:24.417","AcceptedAnswerId":null,"OwnerUserId":10561436.0,"Title":"How to implement loops in tensorflow","Body":"<p>I am learning tesnsorflow currently. it is possible to implement such a code in tensorflow when eager execution is not enabled ?.<\/p>\n\n<pre><code>for param, dparam, mem in zip([self.W1,self.W2,self.W3,self.W4,self.W5,self.b1,self.b2,self.b3,self.b4,self.b5],[dW1,dW2,dW3,dW4,dW5,db1,db2,db3,db4,db5],[mW1,mW2,mW3,mW4,mWO,mb1,mb2,mb3,mb4,mbO]):\n                        mem += dparam * dparam\n                        param += -(self.learning_rate * dparam\/(np.sqrt(mem+1e-8)))\n<\/code><\/pre>\n","answers":[{"AnswerId":"58553057","CreationDate":"2019-10-25T06:07:42.627","ParentId":null,"OwnerUserId":"12198627","Title":null,"Body":"<p>Tensor objects are not iterable. So, to answer your question, there is no zip-like function in TensorFlow.<\/p>\n\n<p>You can indeed use <code>tf.map_fn<\/code> to apply a function to a sequence of tensors.<\/p>\n\n<p>Your can use following sample program for that,<\/p>\n\n<pre><code># declare variables \ndparam  = tf.constant([1, 2, 3, 4])\n\n# NOTE: use stack because map_tf only takes one input tensor\nab = tf.stack([dparam , dparam ], 1)\n\n\ndef map_operation(value_dparam ):\n    # iterates each value_dparam \n    value_a = value_dparam [0]\n    value_b = value_dparam [1]\n    return value_a*value_b\n\n# iterates each value_ab at map_operation()\nmap_result = tf.map_fn(map_operation, dparam , dtype=tf.int32)\n\nwith tf.Session() as sess:\n    tf.initialize_all_variables().run()\n    print(sess.run(map_result))         # [1*1 2*2 3*3 4*4]\n\n''' output\n[1 4 9 16]\n'''\n<\/code><\/pre>\n"}]}
+{"QuestionId":57704579,"AnswerCount":1,"Tags":"<python><keras><syntax-error>","CreationDate":"2019-08-29T06:52:07.080","AcceptedAnswerId":null,"OwnerUserId":11597855.0,"Title":"keras -SyntaxError: invalid syntax","Body":"<p>how to fix error?`<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Flatten, Dense, Embedding\nfrom keras.preprocessing.text import Tokenizer\n....\n<\/code><\/pre>\n\n<p>Error:<\/p>\n\n<pre><code>Using TensorFlow backend.\nTraceback (most recent call last):\nFile \"C:\\Users\\Desktop\\a.py\", line 3, in &lt;module&gt;\nfrom keras.models import Sequential\nFile \"C:\\Users\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site- \npackages\\keras-2.2.4-py3.7.egg\\keras\\__init__.py\", line 3, in &lt;module&gt;\nfrom . import utils\nFile \"C:\\Users\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site- \npackages\\keras-2.2.4-py3.7.egg\\keras\\utils\\__init__.py\", line 6, in \n&lt;module&gt;\nfrom . import conv_utils\nFile \"C:\\Users\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site- \npackages\\keras-2.2.4-py3.7.egg\\keras\\utils\\conv_utils.py\", line 9, in \n&lt;module&gt;\nfrom .. import backend as K\nFile \"C:\\Users\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site- \npackages\\keras-2.2.4-py3.7.egg\\keras\\backend\\__init__.py\", line 89, in \n&lt;module&gt;\nfrom .tensorflow_backend import *\nFile \"C:\\Users\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site- \npackages\\keras-2.2.4-py3.7.egg\\keras\\backend\\tensorflow_backend.py\", \nline 5, in &lt;module&gt;\nimport tensorflow as tf\nFile \"C:\\Users\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site- \npackages\\tensorflow-1.8.0-py3.7.egg\\tensorflow\\__init__.py\", line 24, in \n&lt;module&gt;\nfrom tensorflow.python import pywrap_tensorflow  # pylint: \ndisable=unused-import\nFile \"C:\\UsersAppData\\Local\\Programs\\Python\\Python37-32\\lib\\site- \npackages\\tensorflow-1.8.0-py3.7.egg\\tensorflow\\python\\__init__.py\", line \n49, in &lt;module&gt;\nfrom tensorflow.python import pywrap_tensorflow\nFile \"C:\\Users\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site- \npackages\\tensorflow-1.8.0- \npy3.7.egg\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\nfrom tensorflow.python.pywrap_tensorflow_internal import *\nFile \"C:\\Users\\AppData\\Local\\Programs\\Python\\Python37-32\\lib\\site- \npackages\\tensorflow-1.8.0- \npy3.7.egg\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 114\ndef TFE_ContextOptionsSetAsync(arg1, async):\nSyntaxError: invalid syntax\n<\/code><\/pre>\n\n<p>(I use Windows 7)<\/p>\n","answers":[{"AnswerId":"57704945","CreationDate":"2019-08-29T07:16:06.663","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>That version of TensorFlow does not support python 3.7, this support was added later, so you should update TensorFlow to the latest version (say 1.14).<\/p>\n"}]}
+{"QuestionId":57704653,"AnswerCount":1,"Tags":"<javascript><tensorflow><machine-learning><tensor>","CreationDate":"2019-08-29T06:56:22.793","AcceptedAnswerId":57851651.0,"OwnerUserId":11796570.0,"Title":"Accuracy reduced when predicting using tfjs from model trained by ML5","Body":"<p>I am using <strong>tfjs 1.0.0<\/strong> on <strong>Google Chrome | 76.0.3809.132 (Official Build) (64-bit)<\/strong><\/p>\n\n<p>I was using ML5 to train models for image classification in my project. I used the Feature Extractor for transfer learning. I was using <code>mobilenet_v1_0.25<\/code> as a base model. I wanted to integrate it such that it performs predictions from chrome extension. I had to use tfjs because I found that ML5 does not run from the background page of the extension. I used tfjs to load models trained by ML5 and then start predicting. However, the prediction accuracy was very low in tfjs as compared to when predicting with the same model in ML5 itself.<\/p>\n\n<p>I tried reproducing the prediction from ML5 in tfjs by scrapping <strong>ML5 Feature Extractor<\/strong> source code, but still, the prediction accuracy is highly decreased while predicting from tfjs.<\/p>\n\n<p>I am firstly loading the mobilenet and the custom model to make a joint model<\/p>\n\n<pre><code>load() {\n    console.log(\"ML Data loading..\");\n    \/\/ ! ==========================================\n    \/\/ ! This is a work around and will only work for default version and alpha values that were used while training model.\n    this.mobilenet = await tf.loadLayersModel(\"https:\/\/storage.googleapis.com\/tfjs-models\/tfjs\/mobilenet_v1_0.25_224\/model.json\");\n    const layer = this.mobilenet.getLayer('conv_pw_13_relu');\n    \/\/ ! ==========================================\n\n    this.mobilenetFeatures = await tf.model({ inputs: this.mobilenet.inputs, outputs: layer.output });\n    this.customModel = await tf.loadLayersModel(\".\/model.json\");\n    this.model.add(this.mobilenetFeatures);\n    this.model.add(this.customModel);\n}\n<\/code><\/pre>\n\n<p>I am then passing the image to a function that gets top classes after predicting<\/p>\n\n<p><code>let result = this.getTopKClasses(this.predict(image), 5)<\/code><\/p>\n\n<pre><code>getTopKClasses(logits, topK) {\n    const predictions = logits;\n    const values = predictions.dataSync();\n    predictions.dispose();\n    let predictionList = [];\n    for (let i = 0; i &lt; values.length; i++) {\n        predictionList.push({ value: values[i], index: i });\n    }\n    predictionList = predictionList\n        .sort((a, b) =&gt; {\n            return b.value - a.value;\n        })\n        .slice(0, topK);\n    console.log(predictionList);\n    let site = predictionList[0];\n    let result = { type: 'custom', site: IMAGENET_CLASSES[site.index] }\n    console.log('ML Result: Site: %s, Probability: %i%', result.site, (site.value * 100));\n    if (site.value &gt; ML_THRESHOLD) {\n        return result;\n    } else {\n        return null;\n    }\n}\n\npredict(image) {\n    const preprocessed = this.imgToTensor(image, [224, 224])\n    console.log(preprocessed);\n    var result = this.model.predict(preprocessed);\n    return result;\n}\n<\/code><\/pre>\n\n<p>Helper functions:<\/p>\n\n<pre><code>imgToTensor(input, size = null) {\n    return tf.tidy(() =&gt; {\n        let img = tf.browser.fromPixels(input);\n        if (size) {\n            img = tf.image.resizeBilinear(img, size);\n        }\n        const croppedImage = this.cropImage(img);\n        const batchedImage = croppedImage.expandDims(0);\n        return batchedImage.toFloat().div(tf.scalar(127)).sub(tf.scalar(1));\n    });\n}\n\ncropImage(img) {\n    const size = Math.min(img.shape[0], img.shape[1]);\n    const centerHeight = img.shape[0] \/ 2;\n    const beginHeight = centerHeight - (size \/ 2);\n    const centerWidth = img.shape[1] \/ 2;\n    const beginWidth = centerWidth - (size \/ 2);\n    return img.slice([beginHeight, beginWidth, 0], [size, size, 3]);\n};\n<\/code><\/pre>\n","answers":[{"AnswerId":"57851651","CreationDate":"2019-09-09T09:54:18.047","ParentId":null,"OwnerUserId":"11796570","Title":null,"Body":"<p>In order to predict from a model that was trained using transfer learning (i.e. trained on another, pre-trained model), you need to firstly predict from the base model and then predict from the custom model by passing the base-model predicted tensors to the input of the custom model prediction.<\/p>\n\n<pre><code>async load(options = {}) {\n\n    \/\/ Loading the layer from the base model.\n    this.mobilenet = await tf.loadLayersModel(`${BASE_URL}${this.config.version}_${this.config.alpha}_${IMAGE_SIZE}\/model.json`);\n    const layer = this.mobilenet.getLayer(this.config.layer);\n\n    \/\/Converting the base-model layer to a model.\n    this.mobilenetFeatures = await tf.model({ inputs: this.mobilenet.inputs, outputs: layer.output });\n\n    \/\/ Loading the custom model that was trained by us.\n    this.customModel = await tf.loadLayersModel(CUSTOM_MODEL_FILE_URL);\n}\n<\/code><\/pre>\n\n<p>Now to predict from these models:<\/p>\n\n<pre><code>predict(image) {\n    \/\/ Converting image to tensor\n    const preprocessed = this.imgToTensor(image, [224, 224])\n\n    \/\/ * Make predictions about the image firstly, from the Mobilenet (base) Model.\n    const embeddings = this.mobilenetFeatures.predict(preprocessed);\n\n    \/\/ * Filter predictions from Mobilenet Model using custom trained Model.\n    const result = this.customModel.predict(embeddings);\n\n    return result;\n}\n<\/code><\/pre>\n"}]}
+{"QuestionId":57704771,"AnswerCount":2,"Tags":"<python><tensorflow><keras><deep-learning><eager-execution>","CreationDate":"2019-08-29T07:05:42.780","AcceptedAnswerId":57707690.0,"OwnerUserId":880783.0,"Title":"Inputs to eager execution function cannot be Keras symbolic tensors","Body":"<p>I am trying to implement sample- and pixel-dependent dependent loss weighting in <code>tf.Keras<\/code> (TensorFlow 2.0.0rc0) for a 3-D U-Net with sparse annotation data (Cicek 2016, arxiv:1606.06650).<\/p>\n\n<p>This is my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport tensorflow as tf\nfrom tensorflow.keras import layers, losses, models\n\n# disabling eager execution makes this example work:\n# tf.python.framework_ops.disable_eager_execution()\n\n\ndef get_loss_fcn(w):\n    def loss_fcn(y_true, y_pred):\n        loss = w * losses.mse(y_true, y_pred)\n        return loss\n    return loss_fcn\n\n\ndata_x = np.random.rand(5, 4, 1)\ndata_w = np.random.rand(5, 4)\ndata_y = np.random.rand(5, 4, 1)\n\nx = layers.Input([4, 1])\nw = layers.Input([4])\ny = layers.Activation('tanh')(x)\nmodel = models.Model(inputs=[x, w], outputs=y)\nloss = get_loss_fcn(model.input[1])\n\n# using another loss makes it work, too:\n# loss = 'mse'\n\nmodel.compile(loss=loss)\nmodel.fit((data_x, data_w), data_y)\n\nprint('Done.')\n<\/code><\/pre>\n\n<p>This runs fine when disabling eager execution, but one of the points of TensorFlow 2 is to have eager execution by default. What stands between me and that goal is the custom loss function, as you can see (using <code>'mse'<\/code> as a loss removes that error, too):<\/p>\n\n<pre><code>  File \"MWE.py\", line 30, in &lt;module&gt;\n    model.fit((data_x, data_w), data_y)\n[...]\ntensorflow.python.eager.core._SymbolicException: Inputs to eager execution function cannot be Keras symbolic tensors, but found [&lt;tf.Tensor 'input_2:0' shape=(None, 4) dtype=float32&gt;]\n<\/code><\/pre>\n\n<p>What can I do to make this kind of structure work with eager execution?<\/p>\n\n<p>One idea that I had was to concatenate <code>w<\/code> to the output <code>y<\/code> and separate <code>y_pred<\/code> into the original <code>y_pred<\/code> and <code>w<\/code> in the loss function, but this is a hack I'd like to avoid. It works, though, with changes marked by <code># HERE<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport tensorflow as tf\nfrom tensorflow.keras import layers, losses, models\n\n\n# HERE\ndef loss_fcn(y_true, y_pred):\n    w = y_pred[:, :, -1]  # HERE\n    y_pred = y_pred[:, :, :-1]  # HERE\n    loss = w * losses.mse(y_true, y_pred)\n    return loss\n\n\ndata_x = np.random.rand(5, 4, 1)\ndata_w = np.random.rand(5, 4, 1)  # HERE\ndata_y = np.random.rand(5, 4, 1)\n\nx = layers.Input([4, 1])\nw = layers.Input([4, 1])  # HERE\ny = layers.Activation('tanh')(x)\noutput = layers.Concatenate()([y, w])  # HERE\nmodel = models.Model(inputs=[x, w], outputs=output)  # HERE\nloss = loss_fcn  # HERE\n\nmodel.compile(loss=loss)\nmodel.fit((data_x, data_w), data_y)\n\nprint('Done.')\n<\/code><\/pre>\n\n<p>Any other ideas?<\/p>\n","answers":[{"AnswerId":"57707690","CreationDate":"2019-08-29T09:59:16.277","ParentId":null,"OwnerUserId":"880783","Title":null,"Body":"<p>One alternative solution is to pass weights as additional output features rather than input features.<\/p>\n\n<p>This keeps the model completely free of anything weights related, and the weights appear only in the loss function and the <code>.fit()<\/code> call:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nfrom tensorflow.keras import layers, losses, models\n\ndata_x = 2 * np.ones((7, 11, 15, 3), dtype=float)\ndata_y = 5 * np.ones((7, 9, 13, 5), dtype=float)\n\nx = layers.Input(data_x.shape[1:])\ny = layers.Conv2D(5, kernel_size=3)(x)\nmodel = models.Model(inputs=x, outputs=y)\n\n\ndef loss(y_true, y_pred):\n    (y_true, w) = tf.split(y_true, num_or_size_splits=[-1, 1], axis=-1)\n    loss = tf.squeeze(w, axis=-1) * losses.mse(y_true, y_pred)\n\n    tf.print(tf.math.reduce_mean(y_true), \"== 5\")\n    tf.print(tf.math.reduce_mean(w), \"== 3\")\n\n    return loss\n\n\nmodel.compile(loss=loss)\n\ndata_w = 3 * np.ones((7, 9, 13, 1), dtype=float)\ndata_yw = np.concatenate((data_y, data_w), axis=-1)\nmodel.fit(data_x, data_yw)\n<\/code><\/pre>\n\n<p>One drawback still is that you need to manipulate (potentially) large arrays when merging <code>y<\/code> and <code>w<\/code> in <code>numpy.stack()<\/code>, so anymore more TensorFlow-like will be appreciated.<\/p>\n"},{"AnswerId":"58730735","CreationDate":"2019-11-06T12:57:57.543","ParentId":null,"OwnerUserId":"11054824","Title":null,"Body":"<p>Another way:<\/p>\n\n<pre><code>from tensorflow.keras import layers, models, losses\nimport numpy as np\n\ndef loss_fcn(y_true, y_pred, w):\n    loss = w * losses.mse(y_true, y_pred)\n    return loss\n\n\ndata_x = np.random.rand(5, 4, 1)\ndata_w = np.random.rand(5, 4)\ndata_y = np.random.rand(5, 4, 1)\n\nx = layers.Input([4, 1])\ny_true = layers.Input([4, 1])\nw = layers.Input([4])\ny = layers.Activation('tanh')(x)\n\n\nmodel = models.Model(inputs=[x, y_true, w], outputs=y)\nmodel.add_loss(loss_fcn(y, y_true, w))\n\n\nmodel.compile()\nmodel.fit((data_x, data_y, data_w))\n<\/code><\/pre>\n\n<p>I think this is the most elegant solution.<\/p>\n"}]}
+{"QuestionId":57704850,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2019-08-29T07:10:46.187","AcceptedAnswerId":57705480.0,"OwnerUserId":7709727.0,"Title":"How to convert SparseTensorValue to numpy array?","Body":"<p>I have a Tensorflow network and can get the values of the graph after I call <code>Session().run()<\/code>. However, I have some trouble converting <code>SparseTensorValue<\/code> to other types. <\/p>\n\n<p>For example, the following program creates a <code>SparseTensorValue<\/code>. <\/p>\n\n<pre><code>&gt;&gt;&gt; import tensorflow as tf\n&gt;&gt;&gt; t = tf.Session().run(tf.SparseTensor([[0,1], [0,0], [1,1], [1,0]],[1,2,3,4],[2,2]))\n&gt;&gt;&gt; print(t)\nSparseTensorValue(indices=array([[0, 1],\n       [0, 0],\n       [1, 1],\n       [1, 0]]), values=array([1, 2, 3, 4], dtype=int32), dense_shape=array([2, 2]))\n&gt;&gt;&gt; \n<\/code><\/pre>\n\n<p>What I want is some way to convert <code>t<\/code> to a <code>np.array<\/code> or <code>np.matrix<\/code>, for example, <code>np.array([[2., 1.], [4., 3.]])<\/code>. <\/p>\n\n<p>What I have currently is the following<\/p>\n\n<pre><code>&gt;&gt;&gt; import numpy as np\n&gt;&gt;&gt; a = np.zeros(t.dense_shape)\n&gt;&gt;&gt; for i, v in zip(t.indices, t.values) :\n...     a[tuple(i)] = v\n... \n&gt;&gt;&gt; print(a)\n[[2. 1.]\n [4. 3.]]\n&gt;&gt;&gt; \n<\/code><\/pre>\n\n<p>Is there a better way to perform the conversion? Especially, I want to eliminate the for-loop. <\/p>\n","answers":[{"AnswerId":"57705480","CreationDate":"2019-08-29T07:48:10.323","ParentId":null,"OwnerUserId":"7709727","Title":null,"Body":"<p>Thanks to <a href=\"https:\/\/stackoverflow.com\/users\/901925\/hpaulj\">hpaulj<\/a>'s hint, I found the way to convert from Tensorflow's website. <\/p>\n\n<pre><code>tf.Session().run(tf.sparse.to_dense(tf.sparse.reorder(t)))\n<\/code><\/pre>\n\n<p>First reorder the values to lexicographical order, then use <code>to_dense<\/code> to make it dense, and finally feed the tensor to <code>Session().run()<\/code>. <\/p>\n"}]}
+{"QuestionId":57704914,"AnswerCount":1,"Tags":"<machine-learning><keras><lstm>","CreationDate":"2019-08-29T07:14:37.617","AcceptedAnswerId":null,"OwnerUserId":11992840.0,"Title":"LSTM fit_generator steps_per_epoch","Body":"<p>Because of the large data, I'm using fit_generator with a custom generator to train LSTM model.<\/p>\n\n<p>I haven't used LSTM with fit_generator before, so I don't know whether my code is correct.<\/p>\n\n<pre><code>def generator_v2(trainDir,nb_classes,batch_size):\nprint('start generator')\nclasses = [\"G11\",\"G15\",\"G17\",\"G19\",\"G32\",\"G34\",\"G48\",\"G49\"]\nwhile 1:\n    print('loop generator')\n    for root, subdirs, files in os.walk(trainDir):\n        for file in files:\n            try:\n\n                label = root.split(\"\\\\\")[-1]\n                label = classes.index(label)\n                label = to_categorical(label,num_classes=nb_classes).reshape(1,nb_classes)\n                df = pd.read_csv(root +\"\\\\\"+ file)\n                batches = int(np.ceil(len(df) \/ batch_size))\n                for i in range(0, batches):\n                    x_batch = df[i * batch_size:min(len(df), i * batch_size + batch_size)].values\n                    x_batch = x_batch.reshape(1, x_batch.shape[0], x_batch.shape[1])\n                    yield x_batch, label\n\n                del df\n\n            except EOFError:\n                print(\"error\" + file)\n\ntrainDir = \"data_diff_level2_statistics\"\nnb_classes = 8\nbatch_size = 128\nMaxLen = 449    # each csv file has 449 rows,\nbatches = int(np.ceil(MaxLen \/ batch_size))\nfilesCount = sum([len(files) for r, d, files in os.walk(trainDir)])  # the number of all files\n\nsteps_per_epoch = batches*filesCount\n\nmodel = Sequential()\n\nmodel.add(LSTM(4,input_shape=(None,5)))\nmodel.add(Dense(128))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(nb_classes))\nmodel.add(Activation('softmax'))\nmodel.compile(loss='categorical_crossentropy', optimizer='adadelta',metrics=['acc'])\n\nmodel.fit_generator(generator_v2(trainDir,nb_classes,batch_size),steps_per_epoch=steps_per_epoch, nb_epoch = 100, verbose=1)\n<\/code><\/pre>\n\n<p>Do I set the correct number of steps_per_epoch?<\/p>\n\n<p>My all training data shape is : (230,449,5)<\/p>\n\n<p>So, I set steps_per_epoch with 230 * (449\/batch_size).<\/p>\n\n<p>(449\/batch_size) means I read a csv file 128 rows at a time.<\/p>\n","answers":[{"AnswerId":"57705027","CreationDate":"2019-08-29T07:21:26.210","ParentId":null,"OwnerUserId":"6117017","Title":null,"Body":"<p>The argument  <code>steps_per_epoch<\/code> should be equal to the total number of samples (length of your training set) divided by batch_size(the same is available for <code>validation_steps<\/code>.<\/p>\n\n<p>In your example, the length of the dataset is given by <code>dataset_length = number_of_csv_files * length_of_csv_file<\/code>.<\/p>\n\n<p>Therefore, your code is correct since you have 230 * (449\/batch_size), which is similar to what I have written above.<\/p>\n"}]}
+{"QuestionId":57705470,"AnswerCount":1,"Tags":"<python><keras><regression>","CreationDate":"2019-08-29T07:47:39.787","AcceptedAnswerId":57706515.0,"OwnerUserId":11548429.0,"Title":"CNN with keras to predict energy","Body":"<p>I am new to machine learning and keras library and I made a CNN code for regression like below.<\/p>\n\n<pre><code>%matplotlib inline\nfrom __future__ import division\nimport numpy as np\nfrom numpy.random import rand\nimport matplotlib.pyplot as plt\n\n\ndef initial_spin_state(N):\n    state = np.random.choice((0.11111, 0.99999), (N, N))\n    return state\n\n\ndef metropolis_algorithm(config, beta):\n\n    N = len(config)\n\n    for i in range(N):\n        for j in range(N):\n\n            a = np.random.randint(0, N)\n            b = np.random.randint(0, N)\n            s = config[a, b]\n            near=config[(a+1)%N,b] + config[a,(b+1)%N] + config[(a-1)%N,b] + config[a,(b-1)%N]\n\n            delta = 2 * s *near\n            if delta &lt; 0:\n                s *= -1\n            elif rand() &lt; np.exp(-delta * beta):\n                s *= -1\n            config[a, b] = s\n\n    return config\n\ndef get_energy(config):\n    energy = 0\n    N = len(config)\n    for i in range(N):\n        for j in range(N):\n            S = config[i, j]\n            near = config[(i+1)%N, j] + config[i,(j+1)%N] + config[(i-1)%N, j] + config[i,(j-1)%N]\n            energy += near*S\n\n    return energy\n\n\nx_train = []\ny_train = []\nfor i in range(50000):\n\n    config = initial_spin_state(16)\n    energy = get_energy(config)\n\n    x_train.append(config)\n    y_train.append(energy)\n\nx_train = np.array(x_train)\ny_train = np.array(y_train)\n\nprint(x_train.shape)\nprint(y_train.shape)\n\n\nx_test = []\ny_test = []\nfor j in range(20000):\n\n    config = initial_spin_state(16)\n    energy = get_energy(config)\n\n    x_test.append(config)\n    y_test.append(energy)\n\nx_test = np.array(x_test)\ny_test = np.array(y_test)\n\nprint(x_test.shape)\nprint(y_test.shape)\n\n\nx_train = x_train.reshape(50000, 16, 16, 1)\nx_test = x_test.reshape(20000, 16, 16, 1)\nprint(x_train.shape)\nprint(x_test.shape)\n\nimport numpy as np\nimport tensorflow as tf\ntf.logging.set_verbosity(tf.logging.ERROR)\nfrom keras.datasets import mnist\nfrom keras.utils import np_utils\nfrom keras.models import Sequential\nfrom keras.layers import Conv2D, MaxPooling2D, Flatten, Dense, Dropout\nfrom keras.optimizers import Adadelta\nfrom keras.regularizers import l2\n\n\nmodel = Sequential()\nmodel.add(Conv2D(32, (2, 2), input_shape = (16, 16, 1), activation = 'relu'))\nmodel.add(MaxPooling2D(2,2))\n\n\nmodel.add(Conv2D(16, (2, 2), activation = 'relu'))\nmodel.add(MaxPooling2D(2,2))\n\nmodel.add(Flatten())\nmodel.add(Dense(512, activation = 'relu'))\n#model.add(Dense(1024, activation = 'relu'))\nmodel.add(Dense(1))\n\nnp.random.seed(0)\nmodel.summary()\nmodel.compile(loss = \"mse\", metrics = ['accuracy'], optimizer = 'adam')\n\n\n\n%%time\nhist = model.fit(x_train, y_train, epochs = 200, batch_size = 500,\n                validation_data = (x_test, y_test), verbose = 2)\n\n\nimport matplotlib.pyplot as plt\nplt.plot(hist.history['acc'], '_b', label = \"training\")\nplt.plot(hist.history['val_acc'], 'r:', label = \"test\")\nplt.legend()\nplt.grid(\"on\")\nplt.show()\n\n\n<\/code><\/pre>\n\n<p>this code is for image input, and contiuous energy value output.<\/p>\n\n<p>so if I put a image(ising configuration) to CNN, it should predict a energy for the configuration.<\/p>\n\n<p>the problem is..<\/p>\n\n<p>when I train CNN, training loss and validation loss is decreased very slowly.<\/p>\n\n<p>of course, training accuracy and validation accuracy is incread very slowly.<\/p>\n\n<p>and, sometimes only training accuracy is increasd, val-accuracy is not increased.<\/p>\n\n<p>genious guys.. what`s wrong with my code??<\/p>\n\n<p>teach me plz <\/p>\n","answers":[{"AnswerId":"57706515","CreationDate":"2019-08-29T08:53:32.627","ParentId":null,"OwnerUserId":"10371987","Title":null,"Body":"<p>First of all, since you are doing regression problem, I dont think its a good idea to use <code>acc<\/code> as your metric, instead, you might consider using mean absolute error <code>mae<\/code> as your matric.  <\/p>\n\n<p>The loss you are using is <code>mse<\/code> (mean <strong>squared<\/strong> error), so the value would be pretty large especially when you are not normalising your <code>y<\/code> values. However, after running ~25 epochs of your provided code, the <code>validation loss<\/code> dropped to <code>290.xx<\/code> with <code>13.xx<\/code> of <code>mae<\/code> (And it is <strong>not<\/strong> converging yet). And I tried to use your model to predict some validation data, it works fine. Maybe you should test your model before you assume there is something went wrong.  <\/p>\n"}]}
+{"QuestionId":57705654,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning><lstm>","CreationDate":"2019-08-29T07:59:07.890","AcceptedAnswerId":null,"OwnerUserId":10704050.0,"Title":"How to improve LSTM model in python","Body":"<p>I have about 1000 nodes dataset where each node has 4 time-series. Each time series is exactly 6 length long.The label is 0 or 1 (i.e. binary classification).<\/p>\n\n<p>More precisely my dataset looks as follows.<\/p>\n\n<pre><code>node, time-series1, time_series2, time_series_3, time_series4, Label\nn1, [1.2, 2.5, 3.7, 4.2, 5.6, 8.8], [6.2, 5.5, 4.7, 3.2, 2.6, 1.8], \u2026, 1\nn2, [5.2, 4.5, 3.7, 2.2, 1.6, 0.8], [8.2, 7.5, 6.7, 5.2, 4.6, 1.8], \u2026, 0\nand so on.\n<\/code><\/pre>\n\n<p>I am currently using the below mentioned LSTM model to make the classification. <\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(10, input_shape=(6,4)))\nmodel.add(Dense(32))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n\nprint(data.shape) # (1000, 6, 4)\nmodel.fit(data, target)\n<\/code><\/pre>\n\n<p>I am new to keras and deep learning. That is why I am using the above mentioned simple LSTM model. However, when I use it for my dataset, it gives me about ~50% of accuracy. <\/p>\n\n<p>I hope my results are bad becuase I am using the most basic LSTM model. In that case, I would like to improve my current model to get better results.<\/p>\n\n<p>Note: I am not limited to LSTM and happy to explore other suitable models that fits my dataset.<\/p>\n\n<p>I am happy to provide more details if needed.<\/p>\n","answers":[]}
+{"QuestionId":57705869,"AnswerCount":0,"Tags":"<python><keras><deep-learning><conv-neural-network><autoencoder>","CreationDate":"2019-08-29T08:12:45.840","AcceptedAnswerId":null,"OwnerUserId":10844131.0,"Title":"tensorflow.python.framework.errors_impl.InvalidArgumentError:Requested Tensor connection between nodes \"input_1\" and \"input_1\" would create a cycle","Body":"<p>I work with keras and deep learning in python.\nMy network is an autoencoder That's inside a CNN 1D<\/p>\n\n<p>Part of my code is:<\/p>\n\n<pre><code>history = autoencoder.fit(input_img,input_img,\n                                 epochs= 100,\n                                 steps_per_epoch=64,\n                                 validation_steps=1,\n                                 validation_split = 0.1 ,\n                                 verbose=1,                              \n                                 shuffle=True,                                    \n                                 validation_data=(x_test, x_test),\n                                 callbacks=[TensorBoard(log_dir='\/tmp\/autoencoder')]\n                          )\n<\/code><\/pre>\n\n<p>I obtained the following error message:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError:Requested \nTensor connection between nodes \"input_1\" and \"input_1\" would create a cycle\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57706067,"AnswerCount":1,"Tags":"<tensorflow><tensorrt>","CreationDate":"2019-08-29T08:26:08.233","AcceptedAnswerId":57782673.0,"OwnerUserId":2467772.0,"Title":"Data input format to Plugin","Body":"<p>I have tensorflow model as shown in the picture. And trying to convert to TensorRT.\n<a href=\"https:\/\/i.stack.imgur.com\/RMWex.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/RMWex.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Need a plugin for <code>upsample<\/code> node in converting to TensorRT.<\/p>\n\n<p>For Tensorflow implementation, say input size is 1x3x4x19 shape and upsampled to 1x12x14x19 tensor.<\/p>\n\n<p>Implemented the same thing in TensorRT plugin and the considered process flow is as follow.<\/p>\n\n<p>1x3x4x19 tensor in tensorflow is<\/p>\n\n<pre><code>[[[...19channel data ...],[...19channel data ...],[...19channel data ...],[...19channel data ...]],\n\n [[...19channel data ...],[...19channel data ...],[...19channel data ...],[...19channel data ...]],\n\n [[...19channel data ...],[...19channel data ...],[...19channel data ...],[...19channel data ...]]]\n<\/code><\/pre>\n\n<p>flattened first into<\/p>\n\n<pre><code>[...19channel data ..., ...19channel data ..., ...19channel data ..., etc.,...19channel data ...]\n<\/code><\/pre>\n\n<p>the flattened datalength is 228.<\/p>\n\n<p>19 channels data is quite difficult to visualize.<\/p>\n\n<p>So 3 channels data is used as example again for the flattened data.<\/p>\n\n<pre><code>[[[1,2,3],[4,5,6],[7,8,9],[10,11,12]],\n [[1,2,3],[4,5,6],[7,8,9],[10,11,12]],\n [[1,2,3],[4,5,6],[7,8,9],[10,11,12]]]\n<\/code><\/pre>\n\n<p>The flattened array for 3 channels data is<\/p>\n\n<pre><code>[1,2,3,4,5,6,7,8,9,10,11,12,1,2,3,4,5,6,7,8,9,10,11,12,1,2,3,4,5,6,7,8,9,10,11,12]\n<\/code><\/pre>\n\n<p>That flattened data is input into plugin for upsample.\nMy CUDA code in plugin expected the flattened data as shown above.<\/p>\n\n<p>But the plugin output is weired and not same as Tensorflow's upsample output.\nI checked using Openpose1(shown in picture) data to Tensorflow's upsample operation.<\/p>\n\n<p>Is the correct data format to plugin in TensorRT engine?\nIf not, how input data is fed to plugin?<\/p>\n\n<p>The plugin output looks like input is vertically flattened like<\/p>\n\n<pre><code>[1,1,1,2,2,2,3,3,3...etc.]\n<\/code><\/pre>\n","answers":[{"AnswerId":"57782673","CreationDate":"2019-09-04T06:30:47.820","ParentId":null,"OwnerUserId":"2467772","Title":null,"Body":"<p>After taking sometime, the problem is solved. \nThe plugin input is NCHW format.\nIt is like 1,4,7,10,1,4,7,10,1,4,7,10,2,5,8,11,2,5,8,11,2,5,8,11,3,6,9,12,3,6,9,12,3,6,9,12.<\/p>\n\n<p>So need to work on that data and reformat back to NHWC format if match Tensorflow's NHWC format.<\/p>\n"}]}
+{"QuestionId":57706171,"AnswerCount":1,"Tags":"<python><linux><tensorflow><faster-rcnn><distributed-tensorflow>","CreationDate":"2019-08-29T08:32:02.267","AcceptedAnswerId":null,"OwnerUserId":4741060.0,"Title":"implementing mask-r-cnn with tensorflow-distributed","Body":"<p>I'm training a <a href=\"https:\/\/github.com\/matterport\/Mask_RCNN\" rel=\"nofollow noreferrer\" title=\"Link to the git of mask-r-cnn\">mask-r-cnn network<\/a>, which is built on tensorflow and keras. I'm searching for a way to reduce training time, so I thought implementing it with tensorflow-distributed.<\/p>\n\n<p>I've been working with mask-r-cnn for some time, but it seems what I'm trying to do will require me to modify the source code of mask-r-cnn, which is above my current skills.<\/p>\n\n<p>So, my question is, has someone ever done it, or something similar? is it possible at all, or am I misunderstand the use of tensorflow-distributed.<\/p>\n\n<p>Thanks ahead.<\/p>\n","answers":[{"AnswerId":"57736469","CreationDate":"2019-08-31T08:43:29.083","ParentId":null,"OwnerUserId":"1601927","Title":null,"Body":"<p>Even without distributed training, the tensorpack implementation of Mask R-CNN in tensorflow <a href=\"https:\/\/github.com\/tensorpack\/benchmarks\/tree\/master\/MaskRCNN\" rel=\"nofollow noreferrer\">runs 5x faster<\/a> (and more accurate as well) than the one you linked to.<\/p>\n\n<p>It also supports distributed training with MPI.<\/p>\n"}]}
+{"QuestionId":57706179,"AnswerCount":0,"Tags":"<tensorflow><conv-neural-network>","CreationDate":"2019-08-29T08:32:20.527","AcceptedAnswerId":null,"OwnerUserId":4799752.0,"Title":"Does decoding library affects cnn's performance?","Body":"<p>In training, I use opencv to decode jpg files and then use them to train a cnn image classifier. <\/p>\n\n<p>During inference, I use both nvjpeg and opencv to decode jpg files and evaluate the trained cnn model. <\/p>\n\n<p>evaluating with opencv decoding is much better than nvjpeg.<\/p>\n\n<p>Is this normal or I am missing something here?<\/p>\n","answers":[]}
+{"QuestionId":57706256,"AnswerCount":1,"Tags":"<python><pytorch><tensorboard>","CreationDate":"2019-08-29T08:36:50.010","AcceptedAnswerId":null,"OwnerUserId":8119473.0,"Title":"Not able to load the saved graph using torch.utils.tensorboard.SummaryWriter.add_graph method","Body":"<p>I am saving the scalar summary along with model graph using  <code>add_scalar<\/code> and <code>add_graph<\/code> methods from <code>torch.utils.tensorboard.SummaryWriter<\/code>. <\/p>\n\n<p>While running tensorboard on the summary file, it doesnt show the model graph. Just 2 small rectangle at the bottom right, However, it is able to show the scalar variable and images.<\/p>\n\n<p>Sample code from the pytorch documentation attached <\/p>\n\n<pre><code>import torch\nimport torchvision\nfrom torch.utils.tensorboard import SummaryWriter\nfrom torchvision import datasets, transforms\n\n# Writer will output to .\/runs\/ directory by default\nwriter = SummaryWriter()\n\ntransform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))])\ntrainset = datasets.MNIST('mnist_train', train=True, download=True, transform=transform)\ntrainloader = torch.utils.data.DataLoader(trainset, batch_size=64, shuffle=True)\nmodel = torchvision.models.resnet50(False)\n\nprint(model)\n\n# Have ResNet model take in grayscale rather than RGB\nmodel.conv1 = torch.nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3, bias=False)\nimages, labels = next(iter(trainloader))\n\ngrid = torchvision.utils.make_grid(images)\nwriter.add_image('images', grid, 0)\nwriter.add_graph(model, images)\nwriter.close()\n<\/code><\/pre>\n\n<p>Any pointer to the solution would be appreciated<\/p>\n\n<blockquote>\n  <p>PyTorch version - 1.2.0 Tensorboard version - 1.15.0a20190828<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/8v9bB.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/8v9bB.png\" alt=\"Attaching the bottom right crop of the tensorboard\"><\/a><\/p>\n","answers":[{"AnswerId":"58980979","CreationDate":"2019-11-21T17:54:43.473","ParentId":null,"OwnerUserId":"12211228","Title":null,"Body":"<p>I had the same problem. \nCheck out this thread:\n<a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/24157\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/pytorch\/issues\/24157<\/a><\/p>\n\n<p>TLDR: Update PyTorch to PyTorch-nightly and the problem should be solved.\n<a href=\"https:\/\/pytorch.org\/get-started\/locally\/\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/get-started\/locally\/<\/a><\/p>\n"}]}
+{"QuestionId":57706271,"AnswerCount":0,"Tags":"<tensorflow><keras><yolo><coremltools>","CreationDate":"2019-08-29T08:37:40.200","AcceptedAnswerId":null,"OwnerUserId":3686137.0,"Title":"ValueError: Invalid input shape for image. in keras model conversion to core ml","Body":"<p>I've trained my YOLOv3 network with darknet to recognize some obj in an image. All is OK.\nI want to use weight file in iOS app, so following some tutorials I obtained keras h5 model from darknet weight file ancd I check that also h5 model works fine. OK.\nLast step, using coremltools I tried to convert h5 model in coreml model usable under xcode.\nHere I have the issue... last conversion is performed with this little py script:<\/p>\n\n<pre><code>import coremltools\n\n....\n\ncoreml_model = coremltools.converters.keras.convert('yolorcgz.h5', input_names='image', class_labels=output_labels, image_input_names='image', input_name_shape_dict={'image': [1, 416, 416, 3]})\n\ncoreml_model.input_description['image'] = 'Takes a photo'\ncoreml_model.output_description['output'] = 'Prediction of obj in the photo'\n\ncoreml_model.author = 'SW Team'\ncoreml_model.license = 'Public Domain'\ncoreml_model.short_description = \"YOLOv3 network trained for obj recognition\"\n\ncoreml_model.save('yolorcgz.mlmodel')\n<\/code><\/pre>\n\n<p>when I run the script I've always have this error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"coreml.py\", line 9, in &lt;module&gt;\ncoreml_model =    coremltools.converters.keras.convert('yolorcgz.h5',input_name_shape_dict={'input1': [1, 416, 416, 3]})\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/coremltools\/converters\/keras\/_keras_converter.py\", line 760, in convert\ncustom_conversion_functions=custom_conversion_functions)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/coremltools\/converters\/keras\/_keras_converter.py\", line 556, in convertToSpec\ncustom_objects=custom_objects)\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/coremltools\/converters\/keras\/_keras2_converter.py\", line 305, in _convert\nraise ValueError(errMsg)\nValueError: Invalid input shape for image.\nPlease provide a finite height (H), width (W) &amp; channel value (C) using input_name_shape_dict arg with key = 'image' and value = [None,H,W,C]\nConverted .mlmodel can be modified to have flexible input shape using     coremltools.models.neural_network.flexible_shape_utils\n<\/code><\/pre>\n\n<p>Any ideas on wht could goes wrong?\nThank's a lot<\/p>\n","answers":[]}
+{"QuestionId":57706502,"AnswerCount":0,"Tags":"<optimization><keras><reinforcement-learning><q-learning><adam>","CreationDate":"2019-08-29T08:52:38.487","AcceptedAnswerId":null,"OwnerUserId":5946606.0,"Title":"Does Keras Adam Optimizer and other momemtum-based optimizers retain its past update information over different fit calls?","Body":"<p>The Adam optimizer uses a momentum-like approach to train a neural network in fewer iterations of gradient descent than vanilla gradient descent. I'm trying to figure out whether the Adam optimizer works in a Q-learning situation where you have a non-stationary dataset and numerous model.fit() calls. Does Adam and other optimizers retain their momentum over the different calls or is the state of the optimizer reset at each fit call?\nI've tried searching the code on <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/optimizers.py#L436\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/optimizers.py#L436<\/a> but can't find where this information is stored and whether it's retained.<\/p>\n","answers":[]}
+{"QuestionId":57706675,"AnswerCount":1,"Tags":"<tensorflow><tensorrt>","CreationDate":"2019-08-29T09:02:55.637","AcceptedAnswerId":57782612.0,"OwnerUserId":2467772.0,"Title":"How TensorRT process data?","Body":"<p>In Tensorflow, Tensor format is NxWxHxC and go through the network.<\/p>\n\n<p>When Tensorflow model is converted to TensorRT engine, how Tensor data is processed.\nSame shape as in Thensoflow or flattened.<\/p>\n\n<p>Flattened means, for example, 1x3x4x3 Tensor<\/p>\n\n<pre><code>[[[1,2,3],[4,5,6],[7,8,9],[10,11,12]],\n     [[1,2,3],[4,5,6],[7,8,9],[10,11,12]],\n     [[1,2,3],[4,5,6],[7,8,9],[10,11,12]]]\n<\/code><\/pre>\n\n<p>is flattened into<\/p>\n\n<pre><code>[1,2,3,4,5,6,7,8,9,10,11,12,1,2,3,4,5,6,7,8,9,10,11,12,1,2,3,4,5,6,7,8,9,10,11,12]\n<\/code><\/pre>\n\n<p>single array and processed. <\/p>\n","answers":[{"AnswerId":"57782612","CreationDate":"2019-09-04T06:25:57.473","ParentId":null,"OwnerUserId":"2467772","Title":null,"Body":"<p>TensorRT processes data in NCHW format. When the data is flattened, it is in NCHW format. That means, data is flattened as<\/p>\n\n<pre><code>1,4,7,10,1,4,7,10,1,4,7,10,2,5,8,11,2,5,8,11,2,5,8,11,3,6,9,12,3,6,9,12,3,6,9,12\n<\/code><\/pre>\n"}]}
+{"QuestionId":57706793,"AnswerCount":0,"Tags":"<python><tensorflow><multi-gpu>","CreationDate":"2019-08-29T09:09:30.020","AcceptedAnswerId":null,"OwnerUserId":8578595.0,"Title":"how to train\/finetune a model build graph by import_meta_graph when using tf.estimator","Body":"<p>I am finetuning a model trained by tf.estimator, but build graph by import_meta_graph stead of by code.\nhow to use multi gpus to finetune it?<\/p>\n\n<p>What I want:<\/p>\n\n<ul>\n<li>1- I have a model trained by <code>tf.estimator<\/code>.<\/li>\n<\/ul>\n\n<pre><code>def model_fn_original(feature, labels, mode, params):\n  input = tf.layers.xxx\n  input = tf.layers.xxx\n  logits = tf.layers.dense(input)\n  if mode == tf.estimator.ModeKeys.TRAIN:\n    return tf.estimator.EstimatorSpec()\n<\/code><\/pre>\n\n<ul>\n<li>2- use <code>export_meta_graph<\/code> to save forward graph<\/li>\n<\/ul>\n\n<pre><code>def export_meta_graph():\n  input = tf.layers.xxx\n  input = tf.layers.xxx\n  logits = tf.layers.dense(input)\n  tf.export_meta_graph(meta_file_name)\n\n<\/code><\/pre>\n\n<ul>\n<li>3- use <code>import_meta_graph<\/code> way to replace build graph code in model_fn to finetune model with multi gpus<\/li>\n<\/ul>\n\n<pre><code>def model_fn_imported(feature, labels, mode, params):\n  tf.import_meta_graph(meta_file_name)\n  logits = tf.get_defalt_graph().get_tensor_by_name(\"dense\/BiasAdd:0\")\n  if mode == tf.estimator.ModeKeys.TRAIN:\n    return tf.estimator.EstimatorSpec()\n\nestimator = tf.estimator.Estiator(model_fn=model_fn_imported, ...)\nestimator.train()\n<\/code><\/pre>\n\n<p>I have test that export meta_graph(mode=eval) first build graph by <code>import_meta_graph<\/code>. \nEvaluate it successfully and accuracy is the same as original accuracy.<\/p>\n\n<p>When finetune it, if I only use one GPU, everything go correctly.<\/p>\n\n<p>but when I use multi GPUs:<\/p>\n\n<ul>\n<li><p>1- using <code>tf.contrib.estimator.replicate_model_fn<\/code>:<\/p>\n\n<p>some errors occurs:\ntower_1\/model\/xxx\/xxx\/xxx not found in checkpoints<\/p><\/li>\n<\/ul>\n\n<p>After investigating online, this should be caused by that variables from <code>import_meta_graph<\/code> can not be set reuse. So when run <code>replicate_model_fn<\/code>, and run <code>model_fn many<\/code> times, and create new variable every time.<\/p>\n\n<ul>\n<li><p>2- using tf.contrib.distribute.MirroredStrategy :<\/p>\n\n<p>some errors occurs: destinations must be one of a <code>DistributedValues<\/code> object  a tf.Variable object  a device string a list of device string<\/p><\/li>\n<\/ul>\n\n<p><strong>So question is that is there a way to finetune\/train model using multi GPUs when using import_meta_graph to build graph.<\/strong><\/p>\n\n<p><strong>Is it theoretically ok?<\/strong><\/p>\n","answers":[]}
+{"QuestionId":57706942,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><keras-layer>","CreationDate":"2019-08-29T09:17:32.170","AcceptedAnswerId":57707430.0,"OwnerUserId":11993310.0,"Title":"Keras error: expected input_1 to have 3 dimensions, but got array with shape (256326, 3)","Body":"<p>I was trying to modeling multiple-output dnn. also using kaggle creditcard <a href=\"https:\/\/www.kaggle.com\/mlg-ulb\/creditcardfraud\" rel=\"nofollow noreferrer\">data<\/a>. cause I was just trying to test, my code learn from only three dimensions.<\/p>\n\n<p>my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>df = pd.read_csv('creditcard.csv')\n\nX = df.iloc[:, :-1].values\ny = df.iloc[:, -1].values\n\nX_train, X_test, Y_train, Y_test = train_test_split(X, y, test_size=0.1, random_state=1)\ntemp = []\nfor x in X_train:\n    temp.append(x[:3])\nX_train = temp\ntemp = []\nfor x in X_test:\n    temp.append(x[:3])\nX_test = temp\n\nsc = StandardScaler()\nX_train = sc.fit_transform(X_train)\nX_test = sc.transform(X_test)\n\ninputs = keras.layers.Input(shape=(None, 3))\nx = layers.Dense(16, activation='relu')(inputs)\nx = layers.Dense(20, activation='relu')(x)\nx = layers.Dropout(0.25)(x)\nx = layers.Dense(16, activation='relu')(x)\na_prediction = layers.Dense(1, name='a')(x)\nb_prediction = layers.Dense(16, activation='softmax', name='b')(x)\nc_prediction = layers.Dense(1, activation='sigmoid', name='c')(x)\nmodel = Model(inputs, [a_prediction, b_prediction, c_prediction])\n\nmodel.compile(optimizer='rmsprop', loss={'a': mean_squared_error, 'b': categorical_crossentropy, 'c': binary_crossentropy}, loss_weights={'a': 0.25, 'b': 1., 'c': 10.})\n\nmodel.fit(X_train, {'a': Y_train, 'b': Y_train, 'c': Y_train}, epochs=10, batch_size=64)\n<\/code><\/pre>\n\n<p>error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Traceback (most recent call last):\n  File \"C:\/Users\/Develop\/PycharmProjects\/reinforcement recommandation system\/test2.py\", line 44, in &lt;module&gt;\n    model.fit(X_train, {'a': Y_train, 'b': Y_train, 'c': Y_train}, epochs=10, batch_size=64)\n  File \"C:\\Users\\Develop\\PycharmProjects\\reinforcement recommandation system\\lib\\site-packages\\keras\\engine\\training.py\", line 1089, in fit\n    batch_size=batch_size)\n  File \"C:\\Users\\Develop\\PycharmProjects\\reinforcement recommandation system\\lib\\site-packages\\keras\\engine\\training.py\", line 757, in _standardize_user_data\n    exception_prefix='input')\n  File \"C:\\Users\\Develop\\PycharmProjects\\reinforcement recommandation system\\lib\\site-packages\\keras\\engine\\training_utils.py\", line 131, in standardize_input_data\n    'with shape ' + str(data_shape))\nValueError: Error when checking input: expected input_1 to have 3 dimensions, but got array with shape (256326, 3)\n<\/code><\/pre>\n\n<p>How can I solve this problem?<\/p>\n","answers":[{"AnswerId":"57707430","CreationDate":"2019-08-29T09:43:49.330","ParentId":null,"OwnerUserId":"4453582","Title":null,"Body":"<p>The shape parameter for the Input layer should not include the batch size (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/Input\" rel=\"nofollow noreferrer\">link to doc<\/a>). Switching that line to <code>inputs = keras.layers.Input(shape=(3, ))<\/code> should solve your issue.<\/p>\n\n<p>In future, you can use the <code>model.summary()<\/code> method to see the internal names of your layers and the expected output shape of each. For <strong>your current code<\/strong> the following is printed:<\/p>\n\n<pre><code>__________________________________________________________________________________________________\nLayer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput_1 (InputLayer)            (None, None, 3)      0                                            \n__________________________________________________________________________________________________\ndense_1 (Dense)                 (None, None, 16)     64          input_1[0][0]                    \n__________________________________________________________________________________________________\ndense_2 (Dense)                 (None, None, 20)     340         dense_1[0][0]                    \n__________________________________________________________________________________________________\ndropout_1 (Dropout)             (None, None, 20)     0           dense_2[0][0]                    \n__________________________________________________________________________________________________\ndense_3 (Dense)                 (None, None, 16)     336         dropout_1[0][0]                  \n__________________________________________________________________________________________________\na (Dense)                       (None, None, 1)      17          dense_3[0][0]                    \n__________________________________________________________________________________________________\nb (Dense)                       (None, None, 16)     272         dense_3[0][0]                    \n__________________________________________________________________________________________________\nc (Dense)                       (None, None, 1)      17          dense_3[0][0]                    \n==================================================================================================\nTotal params: 1,046\nTrainable params: 1,046\nNon-trainable params: 0\n__________________________________________________________________________________________________\n<\/code><\/pre>\n\n<p>We can see that the input layer (<code>input_1<\/code>, the same as mentioned in the stack trace) incorrectly has three dimensions.<\/p>\n"}]}
+{"QuestionId":57707154,"AnswerCount":1,"Tags":"<python-3.x><matplotlib><deep-learning><pycharm><pytorch>","CreationDate":"2019-08-29T09:30:15.610","AcceptedAnswerId":57707727.0,"OwnerUserId":11480928.0,"Title":"plot function does not plot curves just plotted legend in PyCharm","Body":"<p>I am doing some homework about image classification using PyTorch. For this purpose I have visited the following link: <a href=\"https:\/\/github.com\/Bjarten\/early-stopping-pytorch\/blob\/master\/MNIST_Early_Stopping_example.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Bjarten\/early-stopping-pytorch\/blob\/master\/MNIST_Early_Stopping_example.ipynb<\/a> and obtained some results like this related to training losses and validation losses<\/p>\n\n<pre><code>[ 1\/10] train_loss: 1.56952 valid_loss: 1.54557\nValidation loss decreased (inf --&gt; 1.545571).  Saving model ...\n[ 2\/10] train_loss: 1.52198 valid_loss: 1.50932\nValidation loss decreased (1.545571 --&gt; 1.509324).  Saving model ...\n[ 3\/10] train_loss: 1.48958 valid_loss: 1.48514\nValidation loss decreased (1.509324 --&gt; 1.485136).  Saving model ...\n[ 4\/10] train_loss: 1.46960 valid_loss: 1.46938\nValidation loss decreased (1.485136 --&gt; 1.469381).  Saving model ...\n<\/code><\/pre>\n\n<p>But the problem is that whenever I try to plot loss curves it does not show anything. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/tthbt.png\" rel=\"nofollow noreferrer\">my plotted figure<\/a><\/p>\n\n<p>I would expect my curve to be like <a href=\"https:\/\/i.stack.imgur.com\/uKcbH.png\" rel=\"nofollow noreferrer\">the one in this picture<\/a>.\nHere is the code that I used to draw this curve:<\/p>\n\n<pre><code>  # visualize the loss as the network trained\n    fig = plt.figure(figsize=(10, 8))\n    plt.plot(range(1, len(train_loss) + 1), train_loss, label='Training Loss')\n    plt.plot(range(1, len(valid_loss) + 1), valid_loss, label='Validation Loss')\n\n        # find position of lowest validation loss\n    minposs = valid_loss.index(min(valid_loss)) + 1\n    plt.axvline(minposs, linestyle='--', color='r', label='Early Stopping Checkpoint')\n\n    plt.xlabel('epochs')\n    plt.ylabel('loss')\n    plt.ylim(0, 0.5)  # consistent scale\n    plt.xlim(0, len(train_loss) + 1)  # consistent scale\n    plt.grid(True)\n    plt.legend()\n    plt.tight_layout()\n    plt.show()\n    fig.savefig('loss_plot.png', bbox_inches='tight')\n<\/code><\/pre>\n\n<p>Here is the value of length of the training loss and the testing loss<\/p>\n\n<pre><code>len of train loss: 1\ntrain loss :[1.6059992909431458]\n<\/code><\/pre>\n","answers":[{"AnswerId":"57707727","CreationDate":"2019-08-29T10:01:02.457","ParentId":null,"OwnerUserId":"4280242","Title":null,"Body":"<p>Since <code>len(train_loss) == 1<\/code> and <code>train_loss = [1.6059992909431458]<\/code>, your plot is exactly the same as if you called <code>plt.plot(range(1, 2), [1.6059992909431458])<\/code>, which, as you can check, is empty (this is because <code>plt.plot<\/code> draws lines between data points, and as there is a single data point, there can be no line drawn). You can change <code>plt.plot<\/code> to <code>plt.scatter<\/code> in order to see that single data point.<\/p>\n\n<p>So this problem has nothing to do with plotting really, and you should rather figure out why your algorithm only runs for a single epoch and troubleshoot that, rather than plotting.<\/p>\n"}]}
+{"QuestionId":57707358,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2019-08-29T09:40:37.910","AcceptedAnswerId":null,"OwnerUserId":11088375.0,"Title":"k-fold cross validation shows higher accuracy than i expect","Body":"<p>I'm training Facial expression recognition data(CK+)\nI use  KFold from sklearn.model_selection on Keras.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def load_data_kfold(X, Y, k):\n    X_train = X\n    y_train = Y\n    print(\"*********\", y_train.shape)\n    folds = list(KFold(n_splits=k, shuffle=True, random_state=1).split(X_train, y_train))\n    return folds, X_train, y_train\n<\/code><\/pre>\n\n<p>here is code of loading data for kfold <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    for j, (train_idx, val_idx) in enumerate(folds):\n        print('\\nFOld ' , j)\n        X_train_cv = X_train[train_idx]\n        y_train_cv = y_train[train_idx]\n        X_valid_cv = X_train[val_idx]\n        y_valid_cv = y_train[val_idx]\n<\/code><\/pre>\n\n<p>by the way, I set 10 k-fold validation\nand accuracy result shows like this<\/p>\n\n<pre><code>\nFOld  9\n\n  Training set =  (1581, 3, 128, 128, 3)\nValidation set =  (175, 3, 128, 128, 3)\nTrain on 1756 samples, validate on 175 samples\nEpoch 1\/10\n - 27s - loss: 0.0208 - acc: 0.9932 - val_loss: 0.0016 - val_acc: 1.0000\n\nEpoch 00001: saving model to app_cnn6\/model\/appearance_lbp_20190829T1710\/3dconv_appearance_9fold_0001.h5\nEpoch 2\/10\n - 27s - loss: 0.0511 - acc: 0.9943 - val_loss: 0.0486 - val_acc: 0.9943\n\nEpoch 00002: saving model to app_cnn6\/model\/appearance_lbp_20190829T1710\/3dconv_appearance_9fold_0002.h5\nEpoch 3\/10\n - 27s - loss: 0.0492 - acc: 0.9909 - val_loss: 0.0029 - val_acc: 1.0000\n\nEpoch 00003: saving model to app_cnn6\/model\/appearance_lbp_20190829T1710\/3dconv_appearance_9fold_0003.h5\nEpoch 4\/10\n - 27s - loss: 0.0269 - acc: 0.9937 - val_loss: 6.4302e-04 - val_acc: 1.0000\n\nEpoch 00004: saving model to app_cnn6\/model\/appearance_lbp_20190829T1710\/3dconv_appearance_9fold_0004.h5\nEpoch 5\/10\n - 27s - loss: 0.0366 - acc: 0.9954 - val_loss: 0.0215 - val_acc: 0.9943\n\nEpoch 00005: saving model to app_cnn6\/model\/appearance_lbp_20190829T1710\/3dconv_appearance_9fold_0005.h5\nEpoch 6\/10\n - 27s - loss: 0.0284 - acc: 0.9932 - val_loss: 0.0293 - val_acc: 0.9886\n\nEpoch 00006: saving model to app_cnn6\/model\/appearance_lbp_20190829T1710\/3dconv_appearance_9fold_0006.h5\nEpoch 7\/10\n - 26s - loss: 0.0442 - acc: 0.9915 - val_loss: 0.0025 - val_acc: 1.0000\n\nEpoch 00007: saving model to app_cnn6\/model\/appearance_lbp_20190829T1710\/3dconv_appearance_9fold_0007.h5\nEpoch 8\/10\n - 27s - loss: 0.0644 - acc: 0.9886 - val_loss: 9.6634e-04 - val_acc: 1.0000\n\nEpoch 00008: saving model to app_cnn6\/model\/appearance_lbp_20190829T1710\/3dconv_appearance_9fold_0008.h5\nEpoch 9\/10\n - 27s - loss: 0.0900 - acc: 0.9863 - val_loss: 0.0129 - val_acc: 0.9943\n\nEpoch 00009: saving model to app_cnn6\/model\/appearance_lbp_20190829T1710\/3dconv_appearance_9fold_0009.h5\nEpoch 10\/10\n - 27s - loss: 0.0382 - acc: 0.9915 - val_loss: 0.0013 - val_acc: 1.0000\n\nEpoch 00010: saving model to app_cnn6\/model\/appearance_lbp_20190829T1710\/3dconv_appearance_9fold_0010.h5\n175\/175 [==============================] - 1s 5ms\/step\nTest Accuracy: 100.00%\nTest Loss:0.00%\n\n<\/code><\/pre>\n\n<p>include this fold 9,\nother fold's accuracy are also higher than i expected.\ndo yo have any idea about this happening?\nwould it be possible?<\/p>\n","answers":[]}
+{"QuestionId":57707368,"AnswerCount":0,"Tags":"<tensorflow><gpu><google-colaboratory><finite-element-analysis>","CreationDate":"2019-08-29T09:41:15.220","AcceptedAnswerId":null,"OwnerUserId":11993565.0,"Title":"Slow computation on google colab while solving partial differential equation","Body":"<p>I 'm using google colab to solve the homogeneous heat equation. I had made a program earlier with scipy using sparse matrices which worked upto N = 10(hyperparameter) but I need to run it for like N = 4... 1000 and thus it won't work on my pc. I therefore converted the code to tensorflow and here I 'm unable to use sparse matrices like I could in sympy but even the GPU\/TPU computation is also slow and slower than my pc. Problems that I'm facing in the code and require solution for<\/p>\n\n<p>1) tf.contrib is removed and thus I 've to use an older version of tensorflow for odeint function. Where is it in 2.0?\n2)If the computation can be computed with sparse matrices it could be good since matrices are tridiagonal.I know about sparse_dense_mul() function but that returns dense tensor and it wouldn't do the job. The \"func\" function applies time independent boundary conditions and then requires matrix multiplication of (nxn) with (nX1) which gives (nX1) with multiple matrices. <\/p>\n\n<p>Also the program was running faster without I created the class.<\/p>\n\n<p>Also it's giving this <\/p>\n\n<pre><code>WARNING: Logging before flag parsing goes to stderr.\nW0829 09:12:24.415445 139855355791232 lazy_loader.py:50] \nThe TensorFlow contrib module will not be included in TensorFlow 2.0.\nFor more information, please see:\n  * https:\/\/github.com\/tensorflow\/community\/blob\/master\/rfcs\/20180907-contrib-sunset.md\n  * https:\/\/github.com\/tensorflow\/addons\n  * https:\/\/github.com\/tensorflow\/io (for I\/O related ops)\nIf you depend on functionality not listed there, please file an issue.\n\nW0829 09:12:24.645356 139855355791232 deprecation.py:323] From \/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/contrib\/integrate\/python\/ops\/odes.py:233: div (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nDeprecated in favor of operator or tf.math.divide.\n<\/code><\/pre>\n\n<p>when I run code for loop in range(2, 10) and tqdm does not display and cell keeps running forever but it works fine for in (2, 5) and tqdm bar does appears.<\/p>\n\n<pre><code>#find a way to use sparse matrices\nclass Heat:\n\n    def __init__(self, N):\n\n        self.N = N\n        self.H = 1\/N\n\n        self.A = ts.to_dense(ts.SparseTensor(indices=[[0, 0], [0, 1]] + \\\n                 [[i, i+j] for i in range(1, N) for j in [-1, 0, 1]] +[[N, N-1], [N, N]],\n                  values=self.H*np.array([1\/3, 1\/6] + [1\/6, 2\/3, 1\/6]*(N-1) + [1\/6, 1\/3], dtype=np.float32),\n                  dense_shape=(N+1, N+1 )))\n\n        self.D = ts.to_dense(ts.SparseTensor(indices=[[0, 0], [0, 1]] + [[i, i+j] \\\n                 for i in range(1, N) for j in [-1, 0, 1]] +[[N, N-1], [N, N]],\n                 values=N*np.array([1-(1), -1 -(-1)] + [-1, 2, -1]*(N-1) + [-1-(-1), 1-(1)], dtype=np.float32),\n                 dense_shape=(N+1, N+1)))\n\n        self.domain = tf.linspace(0.0, 1.0, N+1)\n\n\n\n        def f(k):\n\n            if k == 0:\n\n                return (1 + math.pi**2)*(math.pi*self.H - math.sin(math.pi*self.H))\/(math.pi**2*self.H)\n\n            elif k == N:\n\n                return -(1 + math.pi**2)*(-math.pi*self.H + math.sin(math.pi*self.H))\/(math.pi**2*self.H)\n\n            else:\n\n                return -2*(1 + math.pi**2)*(math.cos(math.pi*self.H) - 1)*math.sin(math.pi*self.H*k)\/(math.pi**2*self.H)\n\n\n        self.F = tf.constant([f(k) for k in range(N+1)], shape=(N+1,), dtype=tf.float32)  #caution! shape changed caution caution 1, N+1(problem) is different from N+1,\n\n        self.exact = tm.scalar_mul(scalar=np.exp(1), x=tf.sin(math.pi*self.domain))\n\n\n\n    def error(self):\n\n        return np.linalg.norm(self.exact.numpy() -  self.approx, 2)\n\n\n    def func (self, y, t):\n        y = tf.Variable(y)\n        y = y[0].assign(0.0)\n        y = y[self.N].assign(0.0)\n        if self.N**2&gt; 100:\n            y_dash = tl.matvec(tf.linalg.inv(self.A), tl.matvec(a=tm.negative(self.D), b=y, a_is_sparse=True) + tm.scalar_mul(scalar=math.exp(t), x=self.F))  #caution! shape changed F is (1, N+1) others too\n        else:\n            y_dash = tl.matvec(tf.linalg.inv(self.A), tl.matvec(a=tm.negative(self.D), b=y) + tm.scalar_mul(scalar=math.exp(t), x=self.F))  #caution! shape changed F is (1, N+1) others too\n\n        y_dash = tf.Variable(y_dash) #!!y_dash performs Hadamard product like multiplication not matrix-like multiplication;returns 2-D\n\n        y_dash = y_dash[0].assign(0.0)\n        y_dash = y_dash[self.N].assign(0.0)\n\n        return y_dash    \n\n    def algo_1(self):\n\n        self.approx = tf.contrib.integrate.odeint(\n            func=self.func,\n            y0=tf.sin(tm.scalar_mul(scalar=math.pi, x=self.domain)),\n            t=tf.constant([0.0, 1.0]),\n            rtol=1e-06,\n            atol=1e-12,\n            method='dopri5',\n            options={\"max_num_steps\":10**10},\n            full_output=False,\n            name=None\n           ).numpy()[1]\n\n    def algo_2(self):\n\n        self.approx = tf.contrib.integrate.odeint_fixed(\n                      func=self.func,\n                      y0=tf.sin(tm.scalar_mul(scalar=math.pi, x=self.domain)),\n                      t=tf.constant([0.0, 1.0]),\n                      dt=tf.constant([self.H**2], dtype=tf.float32),\n                      method='rk4',\n                      name=None\n                    ).numpy()[1]\n\n\ndf = pd.DataFrame(columns=[\"NumBasis\", \"Errors\"])\nNs = [2**r for r in range(2, 10)]\nl =[]\nfor i in tqdm_notebook(Ns):\n    heateqn = Heat(i)\n    heateqn.algo_1()\n    l.append([i, heateqn.error()])\n    df.append({\"NumBasis\":i, \"Errors\":heateqn.error()}, ignore_index=True)\n    tf.keras.backend.clear_session()\n\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57707438,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2019-08-29T09:44:23.170","AcceptedAnswerId":null,"OwnerUserId":7820400.0,"Title":"Keras embedding layer: indices[4,0] = 179819 is not in [0, 8984)","Body":"<p>I'm trying to train a simple movie recommendation system using the latest-small movie lens dataset, but I keep getting an error saying that:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"D:\\AI\\Python projects\\anotherone.py\", line 48, in &lt;module&gt;\n    history = model.fit([train.userId,train.movieId], train.rating,epochs=10, verbose=1)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 780, in fit\n    steps_name='steps_per_epoch')\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_arrays.py\", line 363, in model_iteration\n    batch_outs = f(ins_batch)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py\", line 3292, in _call_\n    run_metadata=self.run_metadata)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1458, in _call_\n    run_metadata_ptr)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: indices[4,0] = 179819 is not in [0, 8984)\n     [[{{node Movie-Embedding\/embedding_lookup}}]]\n<\/code><\/pre>\n\n<p>Code:<\/p>\n\n<pre><code>#  importing libraries\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport os\nimport warnings\nimport tensorflow as tf\nfrom tensorflow.keras.layers import Input, Embedding, Flatten, Dot, Dense, concatenate\nfrom tensorflow.keras.models import Model\nfrom tensorflow.keras.models import load_model\nfrom sklearn.model_selection import train_test_split\n# ignoring warnings\nwarnings.filterwarnings('ignore')\nos.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'\n\n# importing our data\ndf = pd.read_csv('D:\/AI\/Data sets\/ml-latest-small\/ratings.csv')\ntrain, test = train_test_split(df, test_size=0.2, random_state=42)\n\n#  extracting the unique number of users and movies\nn_users = len(df.userId.unique())\nn_movies = len(df.movieId.unique())\n\n\n# creating the embedding\nmovie_input = Input(shape=[1], name=\"Movie-Input\")\nmovie_embedding = Embedding(n_movies+1, 5, name=\"Movie-Embedding\")(movie_input)\nmovie_vec  =  Flatten(name=\"Flatten-Movie\")(movie_embedding)\n\nuser_input = Input(shape=[1], name = \"User-Input\")\nuser_embedding = Embedding(n_users+1, 5, name=\"User-Embedding\")(user_input)\nuser_vec = Flatten(name=\"Flatten-user\")(user_embedding)\n\n# concatinating the features\nconcat = concatenate([movie_vec,user_vec])\n\n# creating our model\nlayer1 = Dense(128,activation=\"relu\")(concat)\nlayer2 = Dense(32,activation=\"relu\")(layer1)\noutputLayer = Dense(1)(layer2)\nmodel = Model([user_input,movie_input],outputLayer)\nmodel.compile('adam','mean_squared_error')\n\n#  tranning the model\nif os.path.exists('multiParam.h5'):\n    model = load_model('multiParam.h5')\nelse:\n    history = model.fit([train.userId,train.movieId], train.rating,epochs=10, verbose=1)\n    model.save('multiParam.h5')\n    plt.plot(history.history['loss'])\n    plt.xlabel(\"Epochs\")\n    plt.ylabel(\"Training Error\")\n\n# testing the model\nprint(model.evaluate([test.userId, test.movieId], test.rating))\n\n# running some predictions\npredictions = model.predict([test.userId.head(10), test.movieId.head(10)])\n[print(predictions[i], test.rating.iloc[i]) for i in range(0,10)]\n<\/code><\/pre>\n\n<p>I'm still new to machine learning, but from the research I understood, I need to provide the number of unique values +1 to the embedding layer, which I'm doing but It's still not working, any help can be appreciated, thank you :)<\/p>\n","answers":[]}
+{"QuestionId":57707445,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-29T09:44:51.210","AcceptedAnswerId":null,"OwnerUserId":133374.0,"Title":"How to add control input to an op after it was run by a session","Body":"<p>Demo code:<\/p>\n\n<pre><code>y = tf.sqrt(42.)\nsession.run(y)\n\nanother_debug_test = tf.Print(y.op.inputs[0], [\"debug print:\"] + list(y.op.inputs))\nadd_control_input(y.op, another_debug_test.op)\ny.op._add_control_input(another_debug_test.op)\n\nsession.run(y)\n<\/code><\/pre>\n\n<p>This results in this warning:<\/p>\n\n<p><code>W tensorflow\/c\/c_api.cc:728] Operation ... was changed by adding control input after it was run by a session. This mutation will have no effect, and will trigger an error in the future. Either don't modify nodes after running them or create a new session.<\/code><\/p>\n\n<p>And the <code>tf.Print<\/code> op is not executed, as the warning says.<\/p>\n\n<p>If you remove the first <code>session.run<\/code>, it works as expected.<\/p>\n\n<p>The question is, how to make it work including the first <code>session.run<\/code>?<\/p>\n\n<p>And also, why does it not work currently? I guess the reason is that the session somehow has cached the existing graph up to that point, and will not automatically redo that when some op of the graph changed afterwards. But if that is the case, can't I just invalidate this cache? How?<\/p>\n\n<p>Creating a new session can be costly in practice. It means that I would need to copy over all the variables. (Or not? But this is a <a href=\"https:\/\/stackoverflow.com\/questions\/57707501\/create-new-session-which-shares-memory-and-variables-of-an-existing-session\">separate question<\/a>.)<\/p>\n\n<p>(To answer the obligatory question, why I want to do this: 1. For debugging purpose. 2. For just playing around.)<\/p>\n","answers":[]}
+{"QuestionId":57707501,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-29T09:47:45.597","AcceptedAnswerId":null,"OwnerUserId":133374.0,"Title":"Create new session which shares memory and variables of an existing session","Body":"<p>I already have a TensorFlow session initialized, with a couple of variables in it.<\/p>\n\n<p>I want to create a new session as cheaply as possible, and reuse the existing variables (such that I don't need to allocate new memory, and I don't need to initialize the variables again), and maybe also reuse the same memory allocator.<\/p>\n\n<p>Is that possible? How?<\/p>\n\n<p>(To answer the obligatory question, why I want to do this: Because of <a href=\"https:\/\/stackoverflow.com\/questions\/57707445\/how-to-add-control-input-to-an-op-after-it-was-run-by-a-session\">this<\/a>. Also because I just want to know.)<\/p>\n","answers":[]}
+{"QuestionId":57707628,"AnswerCount":1,"Tags":"<tensorflow><tensorrt>","CreationDate":"2019-08-29T09:54:51.953","AcceptedAnswerId":57783996.0,"OwnerUserId":2467772.0,"Title":"TensorRT's PluginFormat has to be kNCHW?","Body":"<p>My Tensorflow model is follow (a part of the model).\n<a href=\"https:\/\/i.stack.imgur.com\/kHY2f.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/kHY2f.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Tensorflow model needs to be in NHWC format for input image and its processing.\nIn converting the Tensorflow model to TensorRT engine, upsample needs to implement Plugin.<\/p>\n\n<p>But TensorRT plugin needs format has to be in <code>PluginFormat::kNCHW<\/code>.\nIf set to PluginFormat::kNHWC, plugin can't be compiled.<\/p>\n\n<p>So how to create plugin for such Tensorflow model?<\/p>\n","answers":[{"AnswerId":"57783996","CreationDate":"2019-09-04T08:02:56.487","ParentId":null,"OwnerUserId":"2467772","Title":null,"Body":"<p>Yes TensorRT plugin needs to be NCHW format. To work with Tensorflow model in NHWC format, the processing part, for example the part running on CUDA code needs to design to work on the input array in NCHW format. Then if the Tensorflow model is NHWC format, the reformat back to NHWC at plugin's output.<\/p>\n"}]}
+{"QuestionId":57707760,"AnswerCount":0,"Tags":"<tensorflow><keras><neural-network><tensorboard><tf.keras>","CreationDate":"2019-08-29T10:03:17.697","AcceptedAnswerId":null,"OwnerUserId":1616773.0,"Title":"How to prune weights in Tensorflow CNN model (Without migrating model to Keras)","Body":"<p>I have a TF CNN model and now I want to use the weight pruning API of tensor flow with this but all the examples I checked online, it works only with Keras model <\/p>\n\n<p>I want to prune wights of my existing model<\/p>\n\n<pre><code>def mnist_cnn(inputs):\n    input_layer = tf.reshape(inputs, [-1, 28, 28, 3])\n\n    # Convolutional Layer #1\n    conv1 = tf.layers.conv2d(\n          inputs=input_layer,\n          filters=32,\n          kernel_size=[5, 5],\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 and Pooling Layer #2\n    conv2 = tf.layers.conv2d(\n          inputs=pool1,\n          filters=64,\n          kernel_size=[5, 5],\n          padding=\"same\",\n          activation=tf.nn.relu)\n    pool2 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2], strides=2)\n\n    # Dense Layer\n    pool2_flat = tf.reshape(pool2, [-1, 7 * 7 * 64])\n    dense = tf.layers.dense(inputs=pool2_flat, units=1024, activation=tf.nn.relu)\n    dropout = tf.layers.dropout(inputs=dense, rate=0.4)\n\n    # Logits Layer\n    outputs = tf.layers.dense(inputs=dropout, units=10)\n\n    return outputs\n<\/code><\/pre>\n\n<p>This tutorial defines weight pruning\n<a href=\"https:\/\/www.tensorflow.org\/model_optimization\/guide\/pruning\/pruning_with_keras\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/model_optimization\/guide\/pruning\/pruning_with_keras<\/a><\/p>\n\n<p>but I can't figure out how to use Pruning APIs with my exiting model ( without upgrading it to keras sequential model )<\/p>\n","answers":[]}
+{"QuestionId":57707839,"AnswerCount":0,"Tags":"<keras><pytorch>","CreationDate":"2019-08-29T10:07:45.813","AcceptedAnswerId":null,"OwnerUserId":2999110.0,"Title":"Error in Converting Pytorch model to Keras using Pytorch2Keras","Body":"<p>I am using Pytorch2Keras to convert Signature Verification \n(<a href=\"https:\/\/github.com\/luizgh\/sigver\" rel=\"nofollow noreferrer\">https:\/\/github.com\/luizgh\/sigver<\/a>) available in Pytorch to Keras using following code<\/p>\n\n<pre><code>from pytorch2keras.converter import pytorch_to_keras\nstate_dict, _, _ = torch.load('models\/signet_f_lambda_0.95.pth')\nbase_model = SigNet().to(device).eval()\ninput_np = np.random.uniform(0, 1, (1,1,150,220))\ninput_var = Variable(torch.FloatTensor(input_np))\nk_model = pytorch_to_keras(base_model, input_var, [(1, 150, 220)], verbose=True)  \n<\/code><\/pre>\n\n<p>I am getting the following Error<\/p>\n\n<pre><code>TypeError                                 Traceback (most recent call last)\n----&gt; 6 k_model = pytorch_to_keras(base_model, input_var, [(1, 150, 220)], verbose=True)\n\n\/pytorch2keras\/converter.py in pytorch_to_keras(model, args, input_shapes, change_ordering, verbose, name_policy)\n 71     k_model = onnx_to_keras(onnx_model=onnx_model, input_names=input_names,\n 72                             input_shapes=input_shapes, name_policy=name_policy,\n---&gt; 73                             verbose=verbose, change_ordering=change_ordering)\n 74 \n 75     return k_model\n\n\/onnx2keras\/converter.py in onnx_to_keras(onnx_model, input_names, input_shapes, name_policy, verbose, change_ordering)\n144             node_params,\n145             layers,\n--&gt; 146             node_name\n147         )\n148 \n\n\/onnx2keras\/reshape_layers.py in convert_unsqueeze(node, params, layers, node_name)\n151         layers[node_name] = np.expand_dims(layers[node.input[0]], params['axes'][0])\n152     else:\n--&gt; 153         if len(params['axes'][0]) != 0:\n154             raise AttributeError('Axes is not 0. Cannot unsqueeze')\n155 \n\n TypeError: object of type 'int' has no len()\n<\/code><\/pre>\n\n<p>Help me to solve the issue<\/p>\n","answers":[]}
+{"QuestionId":57707887,"AnswerCount":1,"Tags":"<python-3.x><keras><deep-learning><conv-neural-network><dimension>","CreationDate":"2019-08-29T10:10:12.743","AcceptedAnswerId":57720616.0,"OwnerUserId":8064813.0,"Title":"Keras CNN's input dimension error: expected 4-dim, but found 3-dim","Body":"<p>I have a question about coding CNN using Keras.<\/p>\n\n<p>The shape of input data(adj) is (20000, 50, 50); 20000 is the number of samples, 50 x 50 are 2-D data (like images). Batch size is 100.\n(actually, there are two inputs: adj=(20000, 50, 50), features=(20000, 50, 52). <\/p>\n\n<p>The issued part is like below:<\/p>\n\n<pre><code>from keras.layers import Conv2D, MaxPool2D, Flatten\n\nadj_visible1 = Input(shape=(50, 50, 1))\nconv11 = Conv2D(16, kernel_size=5, activation='relu')(adj_visible1)\npool11 = MaxPool2D(pool_size=(2, 2))(conv11)\nconv12 = Conv2D(8, kernel_size=5, activation='relu')(pool11)\npool12 = MaxPool2D(pool_size=(2, 2))(conv12)\nflat1 = Flatten()(pool12)\n<\/code><\/pre>\n\n<p>But an error message occurred like below:<\/p>\n\n<pre><code>ValueError: Input 0 is incompatible with layer conv2d_1: expected ndim=4, found ndim=3\n<\/code><\/pre>\n\n<p>I found similar cases that print the same message, however, most of the reason is that they didn't consider the filter like (50, 50), not (50, 50, \"1\") for input shape.<\/p>\n\n<p>In my case, I used the shape (50, 50, 1) not (50, 50). However, it still prints the same error message.<\/p>\n\n<p>What should I do?<\/p>\n\n<p>I'm attaching the full code as follows:<\/p>\n\n<pre><code>from sklearn.cross_validation import train_test_split\nfrom keras.models import Sequential\nfrom keras.layers.core import Dense, Dropout\nfrom keras.optimizers import RMSprop, Adam, Adadelta\nfrom keras.utils import plot_model\nfrom keras.models import Model\nfrom keras.layers import Input, Flatten, MaxPool2D\nfrom keras.layers.convolutional import Conv2D\nfrom keras.layers.merge import concatenate\nfrom keras.callbacks import CSVLogger\n\n#Settings\nepoch = 100\nbatch_size = 100\ntest_size = 10000\n\n# Load data\nadj = np.load('adj.npy')   #(20000, 50, 50)\nfeatures = np.load('features.npy')    #(20000, 50, 52)\nProp = np.load('Properties.npy')     #(20000, 1)\n\ndatabase = np.dstack((adj, features))    #(20000, 50, 102)\n\n\n#Train\/Test split\nX_tr, X_te, Y_tr, Y_te = train_test_split(database, Prop, test_size=test_size)\n\nX_tr_adj, X_tr_features = X_tr[:, :, 0:50], X_tr[:, :, 50:]\nX_te_adj, X_te_features = X_te[:, :, 0:50], X_te[:, :, 50:]\n\n\ndef create_model():\n\n    # first input model\n    adj_visible1 = Input(shape=(50, 50, 1))\n    conv11 = Conv2D(16, kernel_size=5, activation='relu')(adj_visible1)\n    pool11 = MaxPool2D(pool_size=(2, 2))(conv11)\n    conv12 = Conv2D(8, kernel_size=5, activation='relu')(pool11)\n    pool12 = MaxPool2D(pool_size=(2, 2))(conv12)\n    flat1 = Flatten()(pool12)\n\n    # second input model\n    features_visible2 = Input(shape=(50, 52, 1))\n    conv21 = Conv2D(16, kernel_size=5, activation='relu')(features_visible2)\n    pool21 = MaxPool2D(pool_size=(2, 2))(conv21)\n    conv22 = Conv2D(8, kernel_size=5, activation='relu')(pool21)\n    pool22 = MaxPool2D(pool_size=(2, 2))(conv22)\n    flat2 = Flatten()(pool22)\n\n    # merge input models\n    merge = concatenate([flat1, flat2])\n\n    # interpretation model\n    hidden1 = Dense(128, activation='relu')(merge)\n    hidden2 = Dense(32, activation='relu')(hidden1)\n    output = Dense(1, activation='linear')(hidden2)\n    model = Model(inputs=[adj_visible1, features_visible2], outputs=output)\n    model.compile(loss='mean_squared_error', optimizer=Adam())\n    # summarize layers\n    print(model.summary())\n\n    return model\n\ndef train_model(batch_size = 100, nb_epoch = 20):\n    model = create_model()\n    csv_logger = CSVLogger('CNN trial.csv')\n    history = model.fit([X_tr_adj, X_tr_features], Y_tr,\n                        batch_size=batch_size,\n                        epochs=nb_epoch,\n                        verbose=1,\n                        validation_data=([X_te_adj, X_te_features], Y_te),\n                        callbacks=[csv_logger])\n\n    predictions_valid = model.predict(X_te_adj, X_te_features, batch_size=batch_size, verbose=1)\n\n    return model\n\ntrain_model(nb_epoch = epoch)\n<\/code><\/pre>\n\n<p>I wrote the code with reference to the following material: <a href=\"https:\/\/machinelearningmastery.com\/keras-functional-api-deep-learning\/\" rel=\"nofollow noreferrer\">https:\/\/machinelearningmastery.com\/keras-functional-api-deep-learning\/<\/a><\/p>\n","answers":[{"AnswerId":"57720616","CreationDate":"2019-08-30T04:45:35.407","ParentId":null,"OwnerUserId":"7331631","Title":null,"Body":"<p>You have to use <strong>conv1D<\/strong> and <strong>MaxPool1D<\/strong> instead of <em>conv2D<\/em> and <em>MaxPool2D<\/em> cause your dataset is a single-channel image instead of 3 channel image. In conv1D layer expect the input to be in format Batch X Height X Width,  while in conv2D it expects the input to dimension = 4 i.e Batch X Height X Width X Channels.<\/p>\n\n<pre><code>from sklearn.model_selection import train_test_split\nfrom keras.models import Sequential\nfrom keras.layers.core import Dense, Dropout\nfrom keras.optimizers import RMSprop, Adam, Adadelta\nfrom keras.utils import plot_model\nfrom keras.models import Model\nfrom keras.layers import Input, Flatten, MaxPool1D\nfrom keras.layers.convolutional import Conv1D\nfrom keras.layers.merge import concatenate\nfrom keras.callbacks import CSVLogger\nimport numpy as np\nepoch = 100\nbatch_size = 100\ntest_size = 10000\n\n\nadj = np.random.randint(0,high=100, size=(20000, 50, 50))      #(20000, 50, 50)\nfeatures = np.random.randint(0,high=100, size=(20000, 50, 52))    #(20000, 50, 52)\nProp = np.random.randint(0,high=100, size=(20000,))     #(20000, 1)\n\ndatabase = np.dstack((adj, features))    #(20000, 50, 102)\n\nprint( \" shape of database :\", database.shape)\n\n\nt\nX_tr, X_te, Y_tr, Y_te = train_test_split(database, Prop, test_size=test_size)\n\nX_tr_adj, X_tr_features = X_tr[:, :, 0:50], X_tr[:, :, 50:]\nX_te_adj, X_te_features = X_te[:, :, 0:50], X_te[:, :, 50:]\n\n\ndef create_model():\n\n    # first input model\n    adj_visible1 = Input(shape=(50, 50))\n    conv11 = Conv1D(16, kernel_size=5, activation='relu')(adj_visible1)\n    pool11 = MaxPool1D(pool_size=2)(conv11)\n    conv12 = Conv1D(8, kernel_size=5, activation='relu')(pool11)\n    pool12 = MaxPool1D(pool_size=2)(conv12)\n    flat1 = Flatten()(pool12)\n\n    # second input model\n    features_visible2 = Input(shape=(50, 52))\n    conv21 = Conv1D(16, kernel_size=5, activation='relu')(features_visible2)\n    pool21 = MaxPool1D(pool_size=2)(conv21)\n    conv22 = Conv1D(8, kernel_size=5, activation='relu')(pool21)\n    pool22 = MaxPool1D(pool_size=2)(conv22)\n    flat2 = Flatten()(pool22)\n\n    # merge input models\n    merge = concatenate([flat1, flat2])\n\n    # interpretation model\n    hidden1 = Dense(128, activation='relu')(merge)\n    hidden2 = Dense(32, activation='relu')(hidden1)\n    output = Dense(1, activation='linear')(hidden2)\n    model = Model(inputs=[adj_visible1, features_visible2], outputs=output)\n    model.compile(loss='mean_squared_error', optimizer=Adam())\n    # summarize layers\n    print(model.summary())\n\n    return model\n\ndef train_model(batch_size = 100, nb_epoch = 20):\n    model = create_model()\n    csv_logger = CSVLogger('CNN trial.csv')\n    history = model.fit([X_tr_adj, X_tr_features], Y_tr,\n                        batch_size=batch_size,\n                        epochs=nb_epoch,\n                        verbose=1,\n                        validation_data=([X_te_adj, X_te_features], Y_te),\n                        callbacks=[csv_logger])\n\n    return model\n\ntrain_model(nb_epoch = 10)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57708166,"AnswerCount":0,"Tags":"<python><keras><deep-learning>","CreationDate":"2019-08-29T10:26:56.337","AcceptedAnswerId":null,"OwnerUserId":1790744.0,"Title":"Problem about Keras custom layers while implementing VQ-VAE","Body":"<p>I'm writing code about vq-vae using Keras. However, I've encountered a strange situation that if implement the encoder layers without custom layer, it works well. But once I put the same code in the custom layer, it gave me a different result.  <\/p>\n\n<p>I already write a custom encoder layer, and it works fine. Both custom layers are written in the same template from Keras documents. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def residual_stack(h, num_hiddens, num_residual_layers, num_residual_hiddens):\n    for i in range(num_residual_layers):\n        h_i = layers.ReLU()(h)\n        h_i = layers.Conv2D(filters=num_residual_hiddens, kernel_size=(3, 3), strides=(1, 1), padding=\"same\")(h_i)\n        h_i = layers.ReLU()(h_i)\n        h_i = layers.Conv2D(filters=num_hiddens, kernel_size=(1, 1), strides=(1, 1), padding=\"same\")(h_i)\n        h = layers.Add()([h, h_i])\n        h = layers.ReLU()(h)\n    return h\n\n\nclass Encoder(layers.Layer):\n    def __init__(self, num_hiddens, num_residual_layers, num_residual_hiddens, **kwargs):\n        self.num_hiddens = num_hiddens\n        self.num_residual_layers = num_residual_layers\n        self.num_residual_hiddens = num_residual_hiddens\n        super(Encoder, self).__init__(**kwargs)\n\n    def call(self, x):\n        h = layers.Conv2D(\n            filters=int(self.num_hiddens \/ 2),\n            kernel_size=(4, 4),\n            strides=(2, 2),\n            padding=\"same\")(x)\n        h = layers.ReLU()(h)\n\n        h = layers.Conv2D(\n            filters=self.num_hiddens,\n            kernel_size=(4, 4),\n            strides=(2, 2),\n            padding=\"same\")(h)\n        h = layers.ReLU()(h)\n\n        h = layers.Conv2D(\n            filters=self.num_hiddens,\n            kernel_size=(3, 3),\n            strides=(1, 1),\n            padding=\"same\")(h)\n\n        h = residual_stack(\n            h,\n            self.num_hiddens,\n            self.num_residual_layers,\n            self.num_residual_hiddens)\n        return h\n\n    def compute_output_shape(self, input_shape):\n        space = input_shape[1:-1]\n        new_space = []\n        for i in range(len(space)):\n            new_space.append(int(space[i] \/\/ 4))\n        return (input_shape[0],) + tuple(new_space) + (self.num_hiddens,)\n\n\nclass Decoder(layers.Layer):\n    def __init__(self, num_hiddens, num_residual_layers, num_residual_hiddens, **kwargs):\n        self.num_hiddens = num_hiddens\n        self.num_residual_layers = num_residual_layers\n        self.num_residual_hiddens = num_residual_hiddens\n        super(Decoder, self).__init__(**kwargs)\n\n    def call(self, x):\n        h = layers.Conv2D(\n            filters=self.num_hiddens,\n            kernel_size=(3, 3),\n            strides=(1, 1),\n            padding=\"same\")(x)\n\n        h = residual_stack(\n            h,\n            self.num_hiddens,\n            self.num_residual_layers,\n            self.num_residual_hiddens)\n\n        h = layers.Conv2DTranspose(\n            filters=int(self.num_hiddens \/ 2),\n            kernel_size=(4, 4),\n            strides=(2, 2),\n            padding=\"same\")(h)\n        h = layers.ReLU()(h)\n\n        x_recon = layers.Conv2DTranspose(\n            filters=1,\n            kernel_size=(4, 4),\n            strides=(2, 2),\n            padding=\"same\")(h)\n\n        return x_recon\n\n    def compute_output_shape(self, input_shape):\n        space = input_shape[1:-1]\n        new_space = []\n        for i in range(len(space)):\n            new_space.append(int(space[i] * 4))\n        return (input_shape[0],) + tuple(new_space) + (1,)\n\nef build_vqvae(x_train, input_shape, embedding_dim=64, num_embeddings=128, commitment_cost=0.25):\n    '''\n    VQ-VAE Hyper Parameters.\n    embedding_dim = 64 # Length of embedding vectors.\n    num_embeddings = 128 # Number of embedding vectors (high value = high bottleneck capacity).\n    commitment_cost = 0.25 # Controls the weighting of the loss terms.\n    '''\n    num_hiddens = 128\n    num_residual_hiddens = 32\n    num_residual_layers = 2\n    assert(x_train.ndim == 4)\n    # Build modules\n    encoder = Encoder(num_hiddens, num_residual_layers, num_residual_hiddens)\n    decoder = Decoder(num_hiddens, num_residual_layers, num_residual_hiddens)\n    pre_vq_conv1 = layers.Conv2D(embedding_dim, kernel_size=(1, 1), strides=(1, 1), padding=\"same\", name=\"to_vq\")\n    input_img = Input(shape=input_shape)\n    h = encoder(input_img)\n\n    # VQVAELayer.\n    enc = Conv2D(embedding_dim, kernel_size=(1, 1), strides=(1, 1), name=\"pre_vqvae\")(h)\n    enc_inputs = enc\n    enc, perplexity = VectorQuantizer(embedding_dim, num_embeddings, commitment_cost, name=\"vqvae\")(enc)\n    x = Lambda(lambda enc: enc_inputs + K.stop_gradient(enc - enc_inputs), name=\"encoded\")(enc)\n    data_variance = np.var(x_train)\n    loss = vq_vae_loss_wrapper(data_variance, commitment_cost, enc, enc_inputs)\n    perplexity = metrices_perplexity_wrapper(perplexity)\n\n    x_recon = decoder(x)\n    vqvae = Model(input_img, x_recon)\n\n<\/code><\/pre>\n\n<p>After 50 epochs, the model learned nothing:<\/p>\n\n<pre><code>mean_squared_error: 0.0013 - metrices: 0.0078\n 9728\/13090 [=====================&gt;........] - ETA: 3s - loss: 0.3939 - mean_squared_error: 0.0013 - metrices: 0.0078\n 9856\/13090 [=====================&gt;........] - ETA: 2s - loss: 0.3940 - mean_squared_error: 0.0013 - metrices: 0.0078\n 9984\/13090 [=====================&gt;........] - ETA: 2s - loss: 0.3938 - mean_squared_error: 0.0013 - metrices: 0.0078\n10112\/13090 [======================&gt;.......] - ETA: 2s - loss: 0.3937 - mean_squared_error: 0.0013 - metrices: 0.0078\n10240\/13090 [======================&gt;.......] - ETA: 2s - loss: 0.3935 - mean_squared_error: 0.0013 - metrices: 0.0078\n10368\/13090 [======================&gt;.......] - ETA: 2s - loss: 0.3939 - mean_squared_error: 0.0013 - metrices: 0.0078\n10496\/13090 [=======================&gt;......] - ETA: 2s - loss: 0.3956 - mean_squared_error: 0.0013 - metrices: 0.0078\n10624\/13090 [=======================&gt;......] - ETA: 2s - loss: 0.3950 - mean_squared_error: 0.0013 - metrices: 0.0078\n10752\/13090 [=======================&gt;......] - ETA: 2s - loss: 0.3950 - mean_squared_error: 0.0013 - metrices: 0.0078\n10880\/13090 [=======================&gt;......] - ETA: 2s - loss: 0.3950 - mean_squared_error: 0.0013 - metrices: 0.0078\n11008\/13090 [========================&gt;.....] - ETA: 1s - loss: 0.3950 - mean_squared_error: 0.0013 - metrices: 0.0078\n11136\/13090 [========================&gt;.....] - ETA: 1s - loss: 0.3951 - mean_squared_error: 0.0013 - metrices: 0.0078\n11264\/13090 [========================&gt;.....] - ETA: 1s - loss: 0.3951 - mean_squared_error: 0.0013 - metrices: 0.0078\n11392\/13090 [=========================&gt;....] - ETA: 1s - loss: 0.3950 - mean_squared_error: 0.0013 - metrices: 0.0078\n11520\/13090 [=========================&gt;....] - ETA: 1s - loss: 0.3956 - mean_squared_error: 0.0013 - metrices: 0.0078\n11648\/13090 [=========================&gt;....] - ETA: 1s - loss: 0.3957 - mean_squared_error: 0.0013 - metrices: 0.0078\n11776\/13090 [=========================&gt;....] - ETA: 1s - loss: 0.3957 - mean_squared_error: 0.0013 - metrices: 0.0078\n11904\/13090 [==========================&gt;...] - ETA: 1s - loss: 0.3973 - mean_squared_error: 0.0013 - metrices: 0.0078\n12032\/13090 [==========================&gt;...] - ETA: 0s - loss: 0.3970 - mean_squared_error: 0.0013 - metrices: 0.0078\n12160\/13090 [==========================&gt;...] - ETA: 0s - loss: 0.3978 - mean_squared_error: 0.0013 - metrices: 0.0078\n12288\/13090 [===========================&gt;..] - ETA: 0s - loss: 0.3978 - mean_squared_error: 0.0013 - metrices: 0.0078\n12416\/13090 [===========================&gt;..] - ETA: 0s - loss: 0.3973 - mean_squared_error: 0.0013 - metrices: 0.0078\n12544\/13090 [===========================&gt;..] - ETA: 0s - loss: 0.3972 - mean_squared_error: 0.0013 - metrices: 0.0078\n12672\/13090 [============================&gt;.] - ETA: 0s - loss: 0.3969 - mean_squared_error: 0.0013 - metrices: 0.0078\n12800\/13090 [============================&gt;.] - ETA: 0s - loss: 0.3970 - mean_squared_error: 0.0013 - metrices: 0.0078\n12928\/13090 [============================&gt;.] - ETA: 0s - loss: 0.3967 - mean_squared_error: 0.0013 - metrices: 0.0078\n13056\/13090 [============================&gt;.] - ETA: 0s - loss: 0.3964 - mean_squared_error: 0.0013 - metrices: 0.0078\n13090\/13090 [==============================] - 13s 966us\/step - loss: 0.3965 - mean_squared_error: 0.0013 - metrices: 0.0078 - val_loss: 0.4074 - val_mean_squared_error: 0.0013 - val_metrices: 0.0078\n<\/code><\/pre>\n\n<p>However, if I put the code block in the build_vae() method like this, it works well. How does this happen?<\/p>\n\n<pre><code>def build_vqvae(x_train, input_shape, embedding_dim=64, num_embeddings=128, commitment_cost=0.25):\n    '''\n    VQ-VAE Hyper Parameters.\n    embedding_dim = 64 # Length of embedding vectors.\n    num_embeddings = 128 # Number of embedding vectors (high value = high bottleneck capacity).\n    commitment_cost = 0.25 # Controls the weighting of the loss terms.\n    '''\n    num_hiddens = 128\n    num_residual_hiddens = 32\n    num_residual_layers = 2\n    assert(x_train.ndim == 4)\n    # Build modules\n    encoder = Encoder(num_hiddens, num_residual_layers, num_residual_hiddens)\n    decoder = Decoder(num_hiddens, num_residual_layers, num_residual_hiddens)\n    pre_vq_conv1 = layers.Conv2D(embedding_dim, kernel_size=(1, 1), strides=(1, 1), padding=\"same\", name=\"to_vq\")\n    input_img = Input(shape=input_shape)\n    h = encoder(input_img)\n\n    # VQVAELayer.\n    enc = Conv2D(embedding_dim, kernel_size=(1, 1), strides=(1, 1), name=\"pre_vqvae\")(h)\n    enc_inputs = enc\n    enc, perplexity = VectorQuantizer(embedding_dim, num_embeddings, commitment_cost, name=\"vqvae\")(enc)\n    x = Lambda(lambda enc: enc_inputs + K.stop_gradient(enc - enc_inputs), name=\"encoded\")(enc)\n    data_variance = np.var(x_train)\n    loss = vq_vae_loss_wrapper(data_variance, commitment_cost, enc, enc_inputs)\n    perplexity = metrices_perplexity_wrapper(perplexity)\n\n    h = layers.Conv2D(\n        filters=num_hiddens,\n        kernel_size=(3, 3),\n        strides=(1, 1),\n        padding=\"same\")(x)\n\n    h = residual_stack(\n        h,\n        num_hiddens,\n        num_residual_layers,\n        num_residual_hiddens)\n\n    h = layers.Conv2DTranspose(\n        filters=int(num_hiddens \/ 2),\n        kernel_size=(4, 4),\n        strides=(2, 2),\n        padding=\"same\")(h)\n    h = layers.ReLU()(h)\n\n    x_recon = layers.Conv2DTranspose(\n        filters=1,\n        kernel_size=(4, 4),\n        strides=(2, 2),\n        padding=\"same\")(h)\n    vqvae = Model(input_img, x_recon)\n\n    return vqvae, loss, perplexity\n<\/code><\/pre>\n\n<p>The result is good for only a few epochs of training: <\/p>\n\n<pre><code> 4224\/13090 [========&gt;.....................] - ETA: 10s - loss: 0.0413 - mean_squared_error: 1.4056e-04 - metrices: 0.0078\n 4352\/13090 [========&gt;.....................] - ETA: 10s - loss: 0.0412 - mean_squared_error: 1.4039e-04 - metrices: 0.0078\n 4480\/13090 [=========&gt;....................] - ETA: 10s - loss: 0.0413 - mean_squared_error: 1.4049e-04 - metrices: 0.0078\n 4608\/13090 [=========&gt;....................] - ETA: 9s - loss: 0.0411 - mean_squared_error: 1.3994e-04 - metrices: 0.0078 \n 4736\/13090 [=========&gt;....................] - ETA: 9s - loss: 0.0409 - mean_squared_error: 1.3917e-04 - metrices: 0.0078\n 4864\/13090 [==========&gt;...................] - ETA: 9s - loss: 0.0407 - mean_squared_error: 1.3861e-04 - metrices: 0.0078\n 4992\/13090 [==========&gt;...................] - ETA: 9s - loss: 0.0406 - mean_squared_error: 1.3821e-04 - metrices: 0.0078\n 5120\/13090 [==========&gt;...................] - ETA: 9s - loss: 0.0405 - mean_squared_error: 1.3785e-04 - metrices: 0.0078\n 5248\/13090 [===========&gt;..................] - ETA: 9s - loss: 0.0403 - mean_squared_error: 1.3733e-04 - metrices: 0.0078\n 5376\/13090 [===========&gt;..................] - ETA: 8s - loss: 0.0401 - mean_squared_error: 1.3672e-04 - metrices: 0.0078\n 5504\/13090 [===========&gt;..................] - ETA: 8s - loss: 0.0401 - mean_squared_error: 1.3657e-04 - metrices: 0.0078\n 5632\/13090 [===========&gt;..................] - ETA: 8s - loss: 0.0401 - mean_squared_error: 1.3655e-04 - metrices: 0.0078\n 5760\/13090 [============&gt;.................] - ETA: 8s - loss: 0.0400 - mean_squared_error: 1.3619e-04 - metrices: 0.0078\n 5888\/13090 [============&gt;.................] - ETA: 8s - loss: 0.0400 - mean_squared_error: 1.3606e-04 - metrices: 0.0078\n 6016\/13090 [============&gt;.................] - ETA: 8s - loss: 0.0399 - mean_squared_error: 1.3579e-04 - metrices: 0.0078\n 6144\/13090 [=============&gt;................] - ETA: 8s - loss: 0.0398 - mean_squared_error: 1.3544e-04 - metrices: 0.0078\n 6272\/13090 [=============&gt;................] - ETA: 7s - loss: 0.0397 - mean_squared_error: 1.3532e-04 - metrices: 0.0078\n 6400\/13090 [=============&gt;................] - ETA: 7s - loss: 0.0397 - mean_squared_error: 1.3505e-04 - metrices: 0.0078\n 6528\/13090 [=============&gt;................] - ETA: 7s - loss: 0.0403 - mean_squared_error: 1.3727e-04 - metrices: 0.0078\n 6656\/13090 [==============&gt;...............] - ETA: 7s - loss: 0.0401 - mean_squared_error: 1.3673e-04 - metrices: 0.0078\n 6784\/13090 [==============&gt;...............] - ETA: 7s - loss: 0.0401 - mean_squared_error: 1.3641e-04 - metrices: 0.0078\n 6912\/13090 [==============&gt;...............] - ETA: 7s - loss: 0.0403 - mean_squared_error: 1.3722e-04 - metrices: 0.0078\n 7040\/13090 [===============&gt;..............] - ETA: 7s - loss: 0.0402 - mean_squared_error: 1.3697e-04 - metrices: 0.0078\n 7168\/13090 [===============&gt;..............] - ETA: 6s - loss: 0.0401 - mean_squared_error: 1.3670e-04 - metrices: 0.0078\n 7296\/13090 [===============&gt;..............] - ETA: 6s - loss: 0.0401 - mean_squared_error: 1.3665e-04 - metrices: 0.0078\n 7424\/13090 [================&gt;.............] - ETA: 6s - loss: 0.0401 - mean_squared_error: 1.3666e-04 - metrices: 0.0078\n 7552\/13090 [================&gt;.............] - ETA: 6s - loss: 0.0401 - mean_squared_error: 1.3655e-04 - metrices: 0.0078\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57708352,"AnswerCount":0,"Tags":"<protocol-buffers><caffe><ubuntu-18.04><protoc>","CreationDate":"2019-08-29T10:37:46.557","AcceptedAnswerId":null,"OwnerUserId":9137206.0,"Title":"Error \"make all\" \/usr\/bin\/ld: cannot find -lprotobuf when installing caffe","Body":"<p>I want to install caffe with conda. In the tutorial I followed it was told to install these stuff <\/p>\n\n<pre><code>sudo apt-get install protobuf-compiler libprotobuf-dev\n<\/code><\/pre>\n\n<p>Soon I found out that is was out-of-date and removed both of them and got latest protoc 3.9.1 . But I'm facing an error that said there is no protobuff in the folder.\nI got confused about protobuf lib and protoc.\nI have only one protoc in this folder<\/p>\n\n<pre><code>protoc: \/usr\/bin\/protoc\n<\/code><\/pre>\n\n<p>And here's what I have after performing<\/p>\n\n<pre><code>sudo make all -j4\n\nCXX\/LD -o .build_release\/tools\/upgrade_net_proto_text.bin\nCXX\/LD -o .build_release\/tools\/convert_imageset.bin\nCXX\/LD -o .build_release\/tools\/upgrade_solver_proto_text.bin\nCXX\/LD -o .build_release\/tools\/extract_features.bin\n\/usr\/bin\/ld: cannot find -lprotobuf\ncollect2: error: ld returned 1 exit status\nMakefile:635: recipe for target '.build_release\/tools\n\/convert_imageset.bin' failed\nmake: *** [.build_release\/tools\/convert_imageset.bin] Error 1\nmake: *** \u041e\u0436\u0438\u0434\u0430\u043d\u0438\u0435 \u0437\u0430\u0432\u0435\u0440\u0448\u0435\u043d\u0438\u044f \u0437\u0430\u0434\u0430\u043d\u0438\u0439\u2026\n\/usr\/bin\/ld: cannot find -lprotobuf\ncollect2: error: ld returned 1 exit status\nMakefile:635: recipe for target '.build_release\/tools\n\/upgrade_solver_proto_text.bin' failed\nmake: *** [.build_release\/tools\/upgrade_solver_proto_text.bin]\nError 1\n\/usr\/bin\/ld: cannot find -lprotobuf\ncollect2: error: ld returned 1 exit status\nMakefile:635: recipe for target '.build_release\/tools\/upgrade_net_proto_text.bin' failed\nmake: *** [.build_release\/tools\/upgrade_net_proto_text.bin] Error 1\n\/usr\/bin\/ld: cannot find -lprotobuf\ncollect2: error: ld returned 1 exit status\nMakefile:635: recipe for target '.build_release\/tools\/extract_features.bin' failed\nmake: *** [.build_release\/tools\/extract_features.bin] Error 1\n<\/code><\/pre>\n\n<p>Can you help me solving this issue? And also it would be great if you explain me the matter with protoc and protobuff<\/p>\n","answers":[]}
+{"QuestionId":57708435,"AnswerCount":2,"Tags":"<keras><deep-learning>","CreationDate":"2019-08-29T10:42:56.277","AcceptedAnswerId":null,"OwnerUserId":5079088.0,"Title":"What are inputs of Keras layers and custom functions?","Body":"<p>Sorry for a nub's question:<\/p>\n\n<p>Having the NN that is trained in fit_generator mode, say something like:<\/p>\n\n<pre><code>Lambda(...)\n<\/code><\/pre>\n\n<p>or <\/p>\n\n<pre><code>Dense(...)\n<\/code><\/pre>\n\n<p>and the custom loss function, what are input tensors?\nAm I correct expecting (batch size, previous layer's output) in case of a Lambda layer?\nIs it going to be the same (batch size, data) in case of a custom loss function that looks like:<\/p>\n\n<pre><code>triplet_loss(y_true, y_pred)\n<\/code><\/pre>\n\n<p>Are y_true, y_pred in format (batch,previous layer's output) and (batch, true 'expected' data we fed to NN)? <\/p>\n","answers":[{"AnswerId":"57710988","CreationDate":"2019-08-29T13:11:19.567","ParentId":null,"OwnerUserId":"9318797","Title":null,"Body":"<p>I would probaly duplicate the dense layers. Instead of having 2 layers with 128 units, have 4 layers with 64 units. The result is the same, but you will be able to perform the cross products better.<\/p>\n\n<pre><code>from keras.models import Model\n\n#create dense layers and store their output tensors, they use the output of models 1 and to as input    \nd1 = Dense(64, ....)(Model_1.output)   \nd2 = Dense(64, ....)(Model_1.output)   \nd3 = Dense(64, ....)(Model_2.output)   \nd4 = Dense(64, ....)(Model_2.output)   \n\ncross1 = Lambda(myFunc, output_shape=....)([d1,d4])\ncross2 = Lambda(myFunc, output_shape=....)([d2,d3])\n\n#I don't really know what kind of \"merge\" you want, so I used concatenate, there are \nAdd, Multiply and others....\noutput = Concatenate()([cross1,cross2])\n#use the \"axis\" attribute of the concatenate layer to define better which axis will \nbe doubled due to the concatenation    \n\nmodel = Model([Model_1.input,Model_2.input], output)\n<\/code><\/pre>\n\n<p>Now, for the lambda function:<\/p>\n\n<pre><code>import keras.backend as K\n\ndef myFunc(x):\n    return x[0] * x[1]\n<\/code><\/pre>\n"},{"AnswerId":"57712778","CreationDate":"2019-08-29T14:51:29.857","ParentId":null,"OwnerUserId":"4257991","Title":null,"Body":"<blockquote>\n  <p>custom loss function, what are input tensors?<\/p>\n<\/blockquote>\n\n<p>It depends on how you define your model <code>outputs<\/code>.<\/p>\n\n<p>For example, let's define a simple model that returns the input unchanged.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential([Lambda(lambda x: x, input_shape=(1,))])\n<\/code><\/pre>\n\nLet's use dummy input X and label Y\n\n<pre class=\"lang-py prettyprint-override\"><code>x = [[0]]\nx = np.array(x)\n\ny = [[4]]\ny = np.array(y)\n<\/code><\/pre>\n\nIf our <strong>custom loss<\/strong> function looks like this\n\n<pre class=\"lang-py prettyprint-override\"><code>def mce(y_true, y_pred):\n  print(y_true.shape)\n  print(y_pred.shape)\n  return K.mean(K.pow(K.abs(y_true - y_pred), 3))\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.compile('sgd', mce)\n<\/code><\/pre>\n\nand then we can see the <strong>shape<\/strong> of <code>y_true<\/code> and <code>y_pred<\/code> will be\n\n<pre class=\"lang-py prettyprint-override\"><code>y_true:  (?, ?)\ny_pred:  (?, 1)\n<\/code><\/pre>\n\n<hr>\n\n<p>However, for <code>triplet loss<\/code> the input for the loss function also can be received like this- <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>ALPHA = 0.2\ndef triplet_loss(x):\n    anchor, positive, negative = x\n\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), ALPHA)\n    loss = tf.reduce_mean(tf.maximum(basic_loss, 0.0), 0)\n\n    return loss\n# Source: https:\/\/github.com\/davidsandberg\/facenet\/blob\/master\/src\/facenet.py\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>def build_model(input_shape):\n    # Standardizing the input shape order\n    K.set_image_dim_ordering('th')\n\n    positive_example = Input(shape=input_shape)\n    negative_example = Input(shape=input_shape)\n    anchor_example = Input(shape=input_shape)\n\n    # Create Common network to share the weights along different examples (+\/-\/Anchor)\n    embedding_network = faceRecoModel(input_shape)\n\n    positive_embedding = embedding_network(positive_example)\n    negative_embedding = embedding_network(negative_example)\n    anchor_embedding = embedding_network(anchor_example)\n\n    loss = merge([anchor_embedding, positive_embedding, negative_embedding],\n                 mode=triplet_loss, output_shape=(1,))\n    model = Model(inputs=[anchor_example, positive_example, negative_example],\n                  outputs=loss)\n    model.compile(loss='mean_absolute_error', optimizer=Adam())\n\n    return model\n<\/code><\/pre>\n"}]}
+{"QuestionId":57708753,"AnswerCount":0,"Tags":"<tensorflow><mask><faster-rcnn>","CreationDate":"2019-08-29T11:02:23.777","AcceptedAnswerId":null,"OwnerUserId":11994074.0,"Title":"MASK RCNN train.py error = assertion failed: [] [Condition x == y did not hold element-wise:]","Body":"<p>I am trying to perform a training job with the help of the model (specified below) on my own dataset which contains only 2 classes.<\/p>\n\n<p>System Info and things completed so far:<\/p>\n\n<ol>\n<li><p>Windows 10 64 bit (using python 3.6)<\/p><\/li>\n<li><p>I generated tfrecords using instructions provided in this repo:- <a href=\"https:\/\/github.com\/vijendra1125\/Custom-Mask-RCNN-using-Tensorfow-Object-detection-API\" rel=\"nofollow noreferrer\">https:\/\/github.com\/vijendra1125\/Custom-Mask-RCNN-using-Tensorfow-Object-detection-API<\/a><\/p><\/li>\n<li><p>Model used: mask_rccn_v2_coco<\/p><\/li>\n<li><p>My images have any resolution from high to low, created bounding box classes with the help of labelImg tool available here:- <a href=\"https:\/\/github.com\/tzutalin\/labelImg\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tzutalin\/labelImg<\/a>, and each image has more than 2 instances of the classes. <\/p>\n\n<p>for example, my xmls have two or more instances of s'.<\/p><\/li>\n<li><p>I create my mask image with : <a href=\"https:\/\/github.com\/abreheret\/PixelAnnotationTool\" rel=\"nofollow noreferrer\">https:\/\/github.com\/abreheret\/PixelAnnotationTool<\/a><\/p><\/li>\n<\/ol>\n\n<p>I am currently do the training job on a gpu.\nError Description: (sorry, its a long post)\nAfter step 26 i found error<\/p>\n\n<pre><code>INFO:tensorflow:global step 25: loss = 3.4984 (0.626 sec\/step)\n\nINFO:tensorflow:global step 25: loss = 3.4984 (0.626 sec\/step)\n\nINFO:tensorflow:global step 26: loss = 3.9062 (0.658 sec\/step)\n\nINFO:tensorflow:global step 26: loss = 3.9062 (0.658 sec\/step)\n\nINFO:tensorflow:Error reported to Coordinator: &lt;class 'tensorflow.python.framework.errors_impl.InvalidArgumentError'&gt;, assertion failed: [] [Condition x == y did not hold element-wise:] [x (Loss\/BoxClassifierLoss\/assert_equal_1\/x:0) = ] [800 1166] [y (Loss\/BoxClassifierLoss\/assert_equal_1\/y:0) = ] [1166 800]\n\n[[Node: Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert = Assert[T=[DT_STRING, DT_STRING, DT_STRING, DT_INT32, DT_STRING, DT_INT32], summarize=3, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]\n\n(Loss\/BoxClassifierLoss\/assert_equal_1\/All\/_221, Loss\/RPNLoss\/assert_equal\/Assert\/Assert\/data_0, Loss\/RPNLoss\/assert_equal\/Assert\/Assert\/data_1, Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\/data_2, Loss\/BoxClassifierLoss\/assert_equal_1\/x\/_223, Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\/data_4, Loss\/BoxClassifierLoss\/assert_equal_1\/y\/_225)]]\n[[Node: Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\/_1484 = _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_1094_Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]]\n\nCaused by op 'Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert', defined at:\nFile \"train.py\", line 184, in \ntf.app.run()\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\", line 125, in run\n_sys.exit(main(argv))\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py\", line 250, in new_func\nreturn func(*args, **kwargs)\nFile \"train.py\", line 180, in main\ngraph_hook_fn=graph_rewriter_fn)\nFile \"C:\\tensorflow1\\models\\research\\object_detection\\legacy\\trainer.py\", line 291, in train\nclones = model_deploy.create_clones(deploy_config, model_fn, [input_queue])\nFile \"C:\\tensorflow1\\models\\research\\slim\\deployment\\model_deploy.py\", line 193, in create_clones\noutputs = model_fn(*args, **kwargs)\nFile \"C:\\tensorflow1\\models\\research\\object_detection\\legacy\\trainer.py\", line 206, in _create_losses\nlosses_dict = detection_model.loss(prediction_dict, true_image_shapes)\nFile \"C:\\tensorflow1\\models\\research\\object_detection\\meta_architectures\\faster_rcnn_meta_arch.py\", line 1711, in loss\nfields.DetectionResultFields.num_detections)))\nFile \"C:\\tensorflow1\\models\\research\\object_detection\\meta_architectures\\faster_rcnn_meta_arch.py\", line 1983, in _loss_box_classifier\ngt_weights_batch=groundtruth_weights_list)\nFile \"C:\\tensorflow1\\models\\research\\object_detection\\core\\target_assigner.py\", line 480, in batch_assign_targets\nanchors, gt_boxes, gt_class_targets, unmatched_class_label, gt_weights)\nFile \"C:\\tensorflow1\\models\\research\\object_detection\\core\\target_assigner.py\", line 156, in assign\nshape_utils.combined_static_and_dynamic_shape(unmatched_class_label))\nFile \"C:\\tensorflow1\\models\\research\\object_detection\\utils\\shape_utils.py\", line 315, in assert_shape_equal\nreturn tf.assert_equal(shape_a, shape_b)\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\ops\\check_ops.py\", line 382, in assert_equal\nreturn control_flow_ops.Assert(condition, data, summarize=summarize)\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\util\\tf_should_use.py\", line 118, in wrapped\nreturn _add_should_use_warning(fn(*args, **kwargs))\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\ops\\control_flow_ops.py\", line 143, in Assert\nreturn gen_logging_ops._assert(condition, data, summarize, name=\"Assert\")\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\ops\\gen_logging_ops.py\", line 51, in _assert\nname=name)\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\framework\\op_def_library.py\", line 787, in _apply_op_helper\nop_def=op_def)\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3414, in create_op\nop_def=op_def)\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\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): assertion failed: [] [Condition x == y did not hold element-wise:] [x (Loss\/BoxClassifierLoss\/assert_equal_1\/x:0) = ] [800 1166] [y (Loss\/BoxClassifierLoss\/assert_equal_1\/y:0) = ] [1166 800]\n[[Node: Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert = Assert[T=[DT_STRING, DT_STRING, DT_STRING, DT_INT32, DT_STRING, DT_INT32], summarize=3, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](Loss\/BoxClassifierLoss\/assert_equal_1\/All\/_221, Loss\/RPNLoss\/assert_equal\/Assert\/Assert\/data_0, Loss\/RPNLoss\/assert_equal\/Assert\/Assert\/data_1, Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\/data_2, Loss\/BoxClassifierLoss\/assert_equal_1\/x\/_223, Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\/data_4, Loss\/BoxClassifierLoss\/assert_equal_1\/y\/_225)]]\n[[Node: Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\/_1484 = _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_1094_Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]]\n\nINFO:tensorflow:Error reported to Coordinator: &lt;class 'tensorflow.python.framework.errors_impl.InvalidArgumentError'&gt;, assertion failed: [] [Condition x == y did not hold element-wise:] [x (Loss\/BoxClassifierLoss\/assert_equal_1\/x:0) = ] [800 1166] [y (Loss\/BoxClassifierLoss\/assert_equal_1\/y:0) = ] [1166 800]\n[[Node: Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert = Assert[T=[DT_STRING, DT_STRING, DT_STRING, DT_INT32, DT_STRING, DT_INT32], summarize=3, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](Loss\/BoxClassifierLoss\/assert_equal_1\/All\/_221, Loss\/RPNLoss\/assert_equal\/Assert\/Assert\/data_0, Loss\/RPNLoss\/assert_equal\/Assert\/Assert\/data_1, Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\/data_2, Loss\/BoxClassifierLoss\/assert_equal_1\/x\/_223, Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\/data_4, Loss\/BoxClassifierLoss\/assert_equal_1\/y\/_225)]]\n[[Node: Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\/_1484 = _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_1094_Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]]\n\nCaused by op 'Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert', defined at:\nFile \"train.py\", line 184, in \ntf.app.run()\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\", line 125, in run\n_sys.exit(main(argv))\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py\", line 250, in new_func\nreturn func(*args, **kwargs)\nFile \"train.py\", line 180, in main\ngraph_hook_fn=graph_rewriter_fn)\nFile \"C:\\tensorflow1\\models\\research\\object_detection\\legacy\\trainer.py\", line 291, in train\nclones = model_deploy.create_clones(deploy_config, model_fn, [input_queue])\nFile \"C:\\tensorflow1\\models\\research\\slim\\deployment\\model_deploy.py\", line 193, in create_clones\noutputs = model_fn(*args, **kwargs)\nFile \"C:\\tensorflow1\\models\\research\\object_detection\\legacy\\trainer.py\", line 206, in _create_losses\nlosses_dict = detection_model.loss(prediction_dict, true_image_shapes)\nFile \"C:\\tensorflow1\\models\\research\\object_detection\\meta_architectures\\faster_rcnn_meta_arch.py\", line 1711, in loss\nfields.DetectionResultFields.num_detections)))\nFile \"C:\\tensorflow1\\models\\research\\object_detection\\meta_architectures\\faster_rcnn_meta_arch.py\", line 1983, in _loss_box_classifier\ngt_weights_batch=groundtruth_weights_list)\nFile \"C:\\tensorflow1\\models\\research\\object_detection\\core\\target_assigner.py\", line 480, in batch_assign_targets\nanchors, gt_boxes, gt_class_targets, unmatched_class_label, gt_weights)\nFile \"C:\\tensorflow1\\models\\research\\object_detection\\core\\target_assigner.py\", line 156, in assign\nshape_utils.combined_static_and_dynamic_shape(unmatched_class_label))\nFile \"C:\\tensorflow1\\models\\research\\object_detection\\utils\\shape_utils.py\", line 315, in assert_shape_equal\nreturn tf.assert_equal(shape_a, shape_b)\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\ops\\check_ops.py\", line 382, in assert_equal\nreturn control_flow_ops.Assert(condition, data, summarize=summarize)\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\util\\tf_should_use.py\", line 118, in wrapped\nreturn _add_should_use_warning(fn(*args, **kwargs))\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\ops\\control_flow_ops.py\", line 143, in Assert\nreturn gen_logging_ops._assert(condition, data, summarize, name=\"Assert\")\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\ops\\gen_logging_ops.py\", line 51, in _assert\nname=name)\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\framework\\op_def_library.py\", line 787, in _apply_op_helper\nop_def=op_def)\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3414, in create_op\nop_def=op_def)\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\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): assertion failed: [] [Condition x == y did not hold element-wise:] [x (Loss\/BoxClassifierLoss\/assert_equal_1\/x:0) = ] [800 1166] [y (Loss\/BoxClassifierLoss\/assert_equal_1\/y:0) = ] [1166 800]\n[[Node: Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert = Assert[T=[DT_STRING, DT_STRING, DT_STRING, DT_INT32, DT_STRING, DT_INT32], summarize=3, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](Loss\/BoxClassifierLoss\/assert_equal_1\/All\/_221, Loss\/RPNLoss\/assert_equal\/Assert\/Assert\/data_0, Loss\/RPNLoss\/assert_equal\/Assert\/Assert\/data_1, Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\/data_2, Loss\/BoxClassifierLoss\/assert_equal_1\/x\/_223, Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\/data_4, Loss\/BoxClassifierLoss\/assert_equal_1\/y\/_225)]]\n[[Node: Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\/_1484 = _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_1094_Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]]\n\nTraceback (most recent call last):\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1322, in _do_call\nreturn fn(*args)\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1307, in _run_fn\noptions, feed_dict, fetch_list, target_list, run_metadata)\nFile \"C:\\Users\\Irfan Prabaswara\\Anaconda3\\envs\\tensorflow1\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1409, in _call_tf_sessionrun\nrun_metadata)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: assertion failed: [] [Condition x == y did not hold element-wise:] [x (Loss\/BoxClassifierLoss\/assert_equal_1\/x:0) = ] [800 1166] [y (Loss\/BoxClassifierLoss\/assert_equal_1\/y:0) = ] [1166 800]\n[[Node: Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert = Assert[T=[DT_STRING, DT_STRING, DT_STRING, DT_INT32, DT_STRING, DT_INT32], summarize=3, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](Loss\/BoxClassifierLoss\/assert_equal_1\/All\/_221, Loss\/RPNLoss\/assert_equal\/Assert\/Assert\/data_0, Loss\/RPNLoss\/assert_equal\/Assert\/Assert\/data_1, Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\/data_2, Loss\/BoxClassifierLoss\/assert_equal_1\/x\/_223, Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\/data_4, Loss\/BoxClassifierLoss\/assert_equal_1\/y\/_225)]]\n[[Node: Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\/_1484 = _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_1094_Loss\/BoxClassifierLoss\/assert_equal_1\/Assert\/Assert\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]]\n<\/code><\/pre>\n\n<p>I would be glad if someone provides a workaround for the above issue.<\/p>\n","answers":[]}
+{"QuestionId":57709129,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2019-08-29T11:25:13.683","AcceptedAnswerId":null,"OwnerUserId":7327257.0,"Title":"Keras DataGenerator: how can I return two outputs but use just one for training?","Body":"<p>I have defined a DataGenerator in Keras using <code>keras.utils.Sequence<\/code>. I need to get as output two different tensors, but I want only one of them to be used in the trainig phase (<code>y1<\/code>), and use the other (<code>y2<\/code>) as input to a custom metric. I have several doubts about doing it with the same DataGenerator, but I couldn't figure out other way.<\/p>\n\n<p>What I actually need is to organize data in <code>y2<\/code> in the same way as input and <code>y1<\/code> tensors: in batches with same data order. I thought I could do it giving as output two tensors in DataGenerator but as long as I know, Keras take as output for training phase all the second part of what the generator returns. <\/p>\n\n<p>So my questions are the following:<\/p>\n\n<ol>\n<li><p>I need to organize some data in exactly the same way as input and output to the network, to use it as input to a custom metric but not use it at all in the training phase: is there a way to do it without using DataGenerator?<\/p><\/li>\n<li><p>If answer to question 1 is \"no\", how could I use DataGenerator to separately work with <code>y2<\/code>? I mean to prevent it to influence in training and to acces later as a tensor to work with it. <\/p><\/li>\n<\/ol>\n\n<p>Thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":57709398,"AnswerCount":1,"Tags":"<python><pandas><tensorflow><keras><google-trends>","CreationDate":"2019-08-29T11:40:05.910","AcceptedAnswerId":57713462.0,"OwnerUserId":4227746.0,"Title":"Using simple models on Google Trends data to predict something doesn't work as expected","Body":"<p>I'm using Google Trends to develop a simple model to predict the future trend for a set of search terms. I took inspiration from <a href=\"https:\/\/www.statworx.com\/de\/blog\/forecasting-last-christmas-search-volume-on-google-trends-using-deep-learning\/\" rel=\"nofollow noreferrer\">this blog post<\/a> and tried to do basically the same thing to other search terms, trying to find the best models for this kind of task.<\/p>\n\n<hr>\n\n<p>The problem is: the predictions for other search terms are completely wrong. I only used terms with a regular pattern, sometimes less regular than the pattern in the example of the blog. Here is my adapted code:<\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nfrom datetime import date\nfrom matplotlib import pyplot as plt\nfrom keras.models import Sequential\nfrom keras.layers import InputLayer, Reshape, Conv1D, MaxPool1D, Flatten, Dense, LSTM\nfrom keras.callbacks import EarlyStopping, ModelCheckpoint\nfrom pandas.plotting import register_matplotlib_converters\nregister_matplotlib_converters()\n\n\n\ndef prepare_data(target, window_X, window_y):\n    \"\"\" Data preprocessing for multistep forecast \"\"\"\n    X, y = [], []\n    start_X = 0\n    end_X = start_X + window_X\n    start_y = end_X\n    end_y = start_y + window_y\n    for _ in range(len(target)):\n        if end_y &lt; len(target):\n            X.append(target[start_X:end_X])\n            y.append(target[start_y:end_y])\n        start_X += 1\n        end_X = start_X + window_X\n        start_y += 1\n        end_y = start_y + window_y\n    X = np.array(X)\n    y = np.array(y)\n    return np.array(X), np.array(y)\n\n\ndef fit_model(type, X_train, y_train, X_test, y_test, batch_size, epochs):\n    \"\"\" Training function for network \"\"\"\n    # Model input\n    model = Sequential()\n    model.add(InputLayer(input_shape=(X_train.shape[1], )))\n\n    if type == 'mlp':\n        model.add(Reshape(target_shape=(X_train.shape[1], )))\n        model.add(Dense(units=64, activation='relu'))\n\n    if type == 'cnn':\n        model.add(Reshape(target_shape=(X_train.shape[1], 1)))\n        model.add(Conv1D(filters=64, kernel_size=4, activation='relu'))\n        model.add(MaxPool1D())\n        model.add(Flatten())\n\n    if type == 'lstm':\n        model.add(Reshape(target_shape=(X_train.shape[1], 1)))\n        model.add(LSTM(units=64, return_sequences=False))\n\n    # Output layer\n    model.add(Dense(units=64, activation='relu'))\n    model.add(Dense(units=y_train.shape[1], activation='sigmoid'))\n\n    # Compile\n    model.compile(optimizer='adam', loss='mse')\n\n    # Callbacks\n    early_stopping = EarlyStopping(monitor='val_loss', patience=10)\n    model_checkpoint = ModelCheckpoint(filepath='model.h5', save_best_only=True)\n    callbacks = [early_stopping, model_checkpoint]\n\n    # Fit model\n    model.fit(x=X_train, y=y_train, validation_data=(X_test, y_test),\n              batch_size=batch_size, epochs=epochs, callbacks=callbacks, verbose=2)\n\n    # Load best model\n    model.load_weights('model.h5')\n\n    # Return\n    return model\n\n\n# Define windows\nwindow_X = 12\nwindow_y = 6\n\n# Load data\ndata = pd.read_csv('data\/holocaust-world.csv', sep=',')\ndata = data.set_index(keys=pd.to_datetime(data['month']), drop=True).drop('month', axis=1)\n\n# Scale data\ndata['y'] = data['y'] \/ 100.\n\n# Prepare tensors\nX, y = prepare_data(target=data['y'].values, window_X=window_X, window_y=window_y)\n\n# Training and test\ntrain = 100\nX_train = X[:train]\ny_train = y[:train]\nX_valid = X[train:]\ny_valid = y[train:]\n\n# Train models\nmodels = ['mlp', 'cnn', 'lstm']\n\n# Test data\nX_test = data['y'].values[-window_X:].reshape(1, -1)\n\n# Predictions\npreds = pd.DataFrame({'mlp': [np.nan]*6, 'cnn': [np.nan]*6, 'lstm': [np.nan]*6})\npreds = preds.set_index(pd.date_range(start=date(2018, 11, 1), end=date(2019, 4, 1), freq='MS'))\n\n# Fit models and plot\nfor mod in models:\n\n    # Train models\n    model = fit_model(type=mod, X_train=X_train, y_train=y_train, X_test=X_valid, y_test=y_valid, batch_size=16, epochs=1000)\n\n    # Predict\n    p = model.predict(x=X_test)\n\n    # Fill\n    preds[mod] = p[0]\n\n# Plot the entire timeline, including the predicted segment\nidx = pd.date_range(start=date(2004, 1, 1), end=date(2019, 4, 1), freq='MS')\ndata = data.reindex(idx)\nplt.plot(data['y'], label='Google')\n\n# Plot\nplt.plot(preds['mlp'], label='MLP')\nplt.plot(preds['cnn'], label='CNN')\nplt.plot(preds['lstm'], label='LSTM')\nplt.legend()\nplt.show()\n<\/code><\/pre>\n\n<p>Here i tried evaluating the interest in the theme of the holocaust, which is also periodic (peak in january, you can grab the csv from Google Trends site obviously). Here are the results:\n<a href=\"https:\/\/i.stack.imgur.com\/joU8C.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/joU8C.png\" alt=\"Results\"><\/a> <\/p>\n\n<hr>\n\n<p>So the questions are:<\/p>\n\n<ul>\n<li><p>how can i <strong>adapt this model to use every month available<\/strong> (at the time of writing, until august 2019)? When i try to do that, i have weird behaviors, so i manually deleted everything after october 2018 in the csv for now.<\/p><\/li>\n<li><p>how can i <strong>improve those simple models to actually give useful and meaningful results<\/strong>? I wonder why the example in the blog post works perfectly, while my attempts fail miserably.<\/p><\/li>\n<\/ul>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"57713462","CreationDate":"2019-08-29T15:30:42.417","ParentId":null,"OwnerUserId":"3190076","Title":null,"Body":"<p>Increase the number of predictions you test and you should get better results.<\/p>\n\n<pre><code>window_y = 49\n...\n# Predictions\npreds = pd.DataFrame({'mlp': [np.nan]*49, 'cnn': [np.nan]*49, 'lstm': [np.nan]*49})\npreds = preds.set_index(pd.date_range(start=date(2015, 1, 1), end=date(2019, 1, 1), freq='MS'))\n<\/code><\/pre>\n\n<p>Playing with the training\/test set will also help:<\/p>\n\n<pre><code># Training and test\ntrain = 50\nX_train = X[:train]\ny_train = y[:train]\nX_valid = X[train:]\ny_valid = y[train:]\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/gabqk.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/gabqk.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>However, this particular trend is periodic but also affected by other factors. <a href=\"https:\/\/facebook.github.io\/prophet\/docs\/seasonality,_holiday_effects,_and_regressors.html#seasonalities-that-depend-on-other-factors\" rel=\"nofollow noreferrer\">Phrophet can help you dealing with this kind of trends<\/a> better than a simple machine learning model.<\/p>\n"}]}
+{"QuestionId":57709758,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-08-29T12:00:03.627","AcceptedAnswerId":57743344.0,"OwnerUserId":5922769.0,"Title":"using transforms.LinearTransformation to apply whitening in PyTorch","Body":"<p>I need to apply ZCA whitening in PyTorch. I think I have found a way this can be done by using transforms.LinearTransformation and I have found a test in the PyTorch repo which gives some insight into how this is done (see final code block or link below)<\/p>\n\n<p><a href=\"https:\/\/github.com\/pytorch\/vision\/blob\/master\/test\/test_transforms.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/vision\/blob\/master\/test\/test_transforms.py<\/a><\/p>\n\n<p>I am struggling to work out how I apply something like this myself.<\/p>\n\n<p>Currently I have transforms along the lines of:<\/p>\n\n<pre><code>    transform_test = transforms.Compose([\n    transforms.ToTensor(),\n    transforms.Normalize(np.array([125.3, 123.0, 113.9]) \/ 255.0,\n                         np.array([63.0, 62.1, 66.7]) \/ 255.0),\n])\n<\/code><\/pre>\n\n<p>The documents say they way to use LinearTransformation is as follows:<\/p>\n\n<pre><code>torchvision.transforms.LinearTransformation(transformation_matrix, mean_vector) \n<\/code><\/pre>\n\n<blockquote>\n  <p>whitening transformation: Suppose X is a column vector zero-centered\n  data. Then compute the data covariance matrix [D x D] with\n  torch.mm(X.t(), X), perform SVD on this matrix and pass it as\n  transformation_matrix.<\/p>\n<\/blockquote>\n\n<p>I can see from the tests I linked above and copied below that they are using torch.mm to calculate what they call a principal_components:<\/p>\n\n<pre><code>def test_linear_transformation(self):\n    num_samples = 1000\n    x = torch.randn(num_samples, 3, 10, 10)\n    flat_x = x.view(x.size(0), x.size(1) * x.size(2) * x.size(3))\n    # compute principal components\n    sigma = torch.mm(flat_x.t(), flat_x) \/ flat_x.size(0)\n    u, s, _ = np.linalg.svd(sigma.numpy())\n    zca_epsilon = 1e-10  # avoid division by 0\n    d = torch.Tensor(np.diag(1. \/ np.sqrt(s + zca_epsilon)))\n    u = torch.Tensor(u)\n    principal_components = torch.mm(torch.mm(u, d), u.t())\n    mean_vector = (torch.sum(flat_x, dim=0) \/ flat_x.size(0))\n    # initialize whitening matrix\n    whitening = transforms.LinearTransformation(principal_components, mean_vector)\n    # estimate covariance and mean using weak law of large number\n    num_features = flat_x.size(1)\n    cov = 0.0\n    mean = 0.0\n    for i in x:\n        xwhite = whitening(i)\n        xwhite = xwhite.view(1, -1).numpy()\n        cov += np.dot(xwhite, xwhite.T) \/ num_features\n        mean += np.sum(xwhite) \/ num_features\n    # if rtol for std = 1e-3 then rtol for cov = 2e-3 as std**2 = cov\n    assert np.allclose(cov \/ num_samples, np.identity(1), rtol=2e-3), \"cov not close to 1\"\n    assert np.allclose(mean \/ num_samples, 0, rtol=1e-3), \"mean not close to 0\"\n\n    # Checking if LinearTransformation can be printed as string\n    whitening.__repr__()\n<\/code><\/pre>\n\n<p>How do I apply something like this? do I use it where I define my transforms or apply it in my training loop where I am iterating over my training loop? <\/p>\n\n<p>Thanks in advance<\/p>\n","answers":[{"AnswerId":"57743344","CreationDate":"2019-09-01T05:06:30.843","ParentId":null,"OwnerUserId":"10935717","Title":null,"Body":"<p>ZCA whitening is typically a preprocessing step, like center-reduction, which basically aims at making your data more NN-friendly (additional info below). As such, it is supposed to be applied once, right before training. <\/p>\n\n<p>So right before you starts training your model with a given dataset <code>X<\/code>, compute the whitened dataset <code>Z<\/code>, which is simply the multiplication of <code>X<\/code> with the ZCA matrix <code>W_zca<\/code> that you can learn to compute <a href=\"https:\/\/stackoverflow.com\/questions\/31528800\/how-to-implement-zca-whitening-python\">here<\/a>. Then train your model on the whitened dataset.\nFinally, you should have something that looks like this<\/p>\n\n<pre><code>class MyModule(torch.nn.Module):\n    def __init__(self):\n        super(MyModule,self).__init__()\n        # Feel free to use something more useful than a simple linear layer\n        self._network = torch.nn.Linear(...)\n        # Do your stuff\n        ...\n\n    def fit(self, inputs, labels):\n    \"\"\" Trains the model to predict the right label for a given input \"\"\"\n        # Compute the whitening matrix and inputs\n        self._zca_mat = compute_zca(inputs)\n        whitened_inputs = torch.mm(self._zca_mat, inputs)\n\n        # Apply training on the whitened data\n        outputs = self._network(whitened_inputs)\n        loss = torch.nn.MSEloss()(outputs, labels)\n        loss.backward()\n        optimizer.step()\n\n     def forward(self, input):\n         # You always need to apply the zca transform before forwarding, \n         # because your network has been trained with whitened data\n         whitened_input = torch.mm(self._zca_mat, input)\n         predicted_label = self._network.forward(whitened_input)\n         return predicted_label\n<\/code><\/pre>\n\n<h1>Additional info<\/h1>\n\n<p>Whitening your data means decorrelating its dimensions so that the correlation matrix of the whitened data is the identity matrix. It is a rotation-scaling operation (thus linear), and there are actually an infinity of possible ZCA transforms. To understand the maths behind ZCA, read <a href=\"https:\/\/arxiv.org\/pdf\/1512.00809.pdf\" rel=\"nofollow noreferrer\">this<\/a><\/p>\n"}]}
+{"QuestionId":57710588,"AnswerCount":1,"Tags":"<python-3.x><deep-learning><pytorch>","CreationDate":"2019-08-29T12:48:18.670","AcceptedAnswerId":57713634.0,"OwnerUserId":11571669.0,"Title":"Understanding deep Markov model code on pyro","Body":"<p>I am going through an a deep Markov model code given on pyro's website:<\/p>\n\n<p><a href=\"https:\/\/pyro.ai\/examples\/dmm.html\" rel=\"nofollow noreferrer\">https:\/\/pyro.ai\/examples\/dmm.html<\/a><\/p>\n\n<p>I am really confused about combiner module that they've implemented. You an find combiner module on line 104 on this GitHub page:<\/p>\n\n<p><a href=\"https:\/\/github.com\/pyro-ppl\/pyro\/blob\/dev\/examples\/dmm\/dmm.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pyro-ppl\/pyro\/blob\/dev\/examples\/dmm\/dmm.py<\/a><\/p>\n\n<p>The article they are following is:<\/p>\n\n<p><a href=\"https:\/\/arxiv.org\/abs\/1609.09869\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1609.09869<\/a><\/p>\n\n<p>They explain combiner module in section 4 (Structured inference networks). I am really confused about why they are doing three linear transformations in the code starting from line 104 on GitHub. Aren't they just suppose to use RNN to produce the distribution or am I missing something? Insights would be appreciated.<\/p>\n","answers":[{"AnswerId":"57713634","CreationDate":"2019-08-29T15:42:17.623","ParentId":null,"OwnerUserId":"819544","Title":null,"Body":"<p>The <code>Combiner<\/code> module implementation corresponds to the formulas described at the bottom of page 4, \"Combiner Function for Structured Approximations (for DKS)\". mu_t is <code>loc<\/code> and sigma_t**2 is <code>scale<\/code>. <\/p>\n\n<p>The RNN state is used to parameterize the distribution, but that distribution is parameterized by two variables. These variables are extracted from the RNN state via the transformations in question.<\/p>\n"}]}
+{"QuestionId":57710920,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-08-29T13:07:16.993","AcceptedAnswerId":null,"OwnerUserId":10203273.0,"Title":"How can I find a solution for the \"FileNotFoundError\"","Body":"<p>I'm currently working on an image classifier project. During the testing of the predict function, I receive the error: FileNotFoundError: [Errno 2] No such file or directory: 'flowers\/test\/1\/image_06760'<\/p>\n\n<p>the path of the file is correct.\nYou can find the whole notebook here: \n<a href=\"https:\/\/github.com\/MartinTschendel\/image-classifier-1\/blob\/master\/190829-0510-Image%20Classifier%20Project.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/MartinTschendel\/image-classifier-1\/blob\/master\/190829-0510-Image%20Classifier%20Project.ipynb<\/a><\/p>\n\n<p>Here is the respective predict function and the test of this function:<\/p>\n\n<pre><code>def predict(image_path, model, topk=5):\n    ''' Predict the class (or classes) of an image using a trained deep learning model.\n    '''\n\n    #loading model\n    loaded_model = load_checkpoint(model)\n\n    # implement the code to predict the class from an image file\n    img = Image.open(image_name)\n    img = process_image(img)\n\n    # convert 2D image to 1D vector\n    img = np.expand_dims(img, 0)\n\n    img = torch.from_numpy(img)\n\n    #set model to evaluation mode and turn off gradients\n    loaded_model.eval()\n    with torch.no_grad():\n        #run image through network\n        output = loaded_model.forward(img)\n\n    #calculate probabilities\n    probs = torch.exp(output)\n    probs_top = probs.topk(topk)[0]\n    index_top = probs.topk(topk)[1]\n\n    # Convert probabilities and outputs to lists\n    probs_top_list = np.array(probs_top)[0]\n    index_top_list = np.array(index_top[0])\n\n    #load index and class mapping\n    loaded_model.class_to_idx = train_data.class_to_idx\n    #invert index-class dictionary\n    idx_to_class = {x: y for y, x in class_to_idx.items()}\n\n    #convert index list to class list\n    classes_top_list = []\n    for index in index_top_list:\n        classes_top_list += [idx_to_class[index]]\n\n    return probs_top_list, classes_top_list\n<\/code><\/pre>\n\n<pre><code># test predict function\n\n# inputs equal paths to saved model and test image\nmodel_path = '190824_checkpoint.pth'\nimage_name = data_dir + '\/test' + '\/1\/' + 'image_06760'\n\nprobs, classes = predict(image_name, model_path, topk=5)\n\nprint(probs)\nprint(classes)\n<\/code><\/pre>\n\n<p>This is the error I receive: <\/p>\n\n<pre><code>FileNotFoundError: [Errno 2] No such file or directory: 'flowers\/test\/1\/image_06760'\n<\/code><\/pre>\n","answers":[{"AnswerId":"57711126","CreationDate":"2019-08-29T13:18:20.890","ParentId":null,"OwnerUserId":"10295265","Title":null,"Body":"<p>These are the images you have <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/YZjME.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/YZjME.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>You should have it as <\/p>\n\n<p><code>image_name = data_dir + '\/test' + '\/1\/' + 'image_06760.jpg'<\/code> <\/p>\n\n<p>for it to work as you were not specifying the image extension.<\/p>\n"}]}
+{"QuestionId":57711103,"AnswerCount":1,"Tags":"<python><tensorflow><repeat>","CreationDate":"2019-08-29T13:17:10.610","AcceptedAnswerId":57712269.0,"OwnerUserId":10561436.0,"Title":"Difference between tf.data.Dataset.repeat() vs iterator.initializer","Body":"<p>Tensorflow has <code>tf.data.Dataset.repeat(x)<\/code> that iterates through the data <code>x<\/code> number of times. It also has <code>iterator.initializer<\/code> which when <code>iterator.get_next()<\/code> is exhausted, <code>iterator.initializer<\/code> can be used to restart the iteration. My question is is there difference when using <code>tf.data.Dataset.repeat(x)<\/code> technique vs <code>iterator.initializer<\/code>?<\/p>\n","answers":[{"AnswerId":"57712269","CreationDate":"2019-08-29T14:22:40.923","ParentId":null,"OwnerUserId":"10878733","Title":null,"Body":"<p>As we know, each epoch in the training process of a model takes in the whole dataset and breaks it into batches. This happens on every epoch.\nSuppose, we have a dataset with 100 samples. On every epoch, the 100 samples are broken into 5 batches ( of 20 each ) for feeding them to the model. But, if I have to train the model for say 5 epochs then, I need to <strong>repeat<\/strong> the dataset 5 times. Meaning, the total elements in the repeated dataset will have 500 samples ( 100 samples multipled 5 times ).<\/p>\n\n<p>Now, this job is done by the <code>tf.data.Dataset.repeat()<\/code> method. Usually we pass the <code>num_epochs<\/code> argument to the method.<\/p>\n\n<p>The <code>iterator.get_next()<\/code> is just a way of getting the next batch of data from the <code>tf.data.Dataset<\/code>. You are iterating the dataset batch by batch.<\/p>\n\n<p>That's the difference. The <code>tf.data.Dataset.repeat()<\/code> repeats the samples in the dataset whereas <code>iterator.get_next()<\/code> one-by-one fetches the data in the form of batches.<\/p>\n"}]}
+{"QuestionId":57711285,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-estimator>","CreationDate":"2019-08-29T13:28:49.760","AcceptedAnswerId":57712129.0,"OwnerUserId":10846167.0,"Title":"How can I get global step in tensorflow.estimator?","Body":"<p>Does anyone know, how can I get the global step count in estimator definition?<br>\nI need it to adjust the learning rate, when the optimizer created.<br>\nlike the example below: <\/p>\n\n<pre><code>def estimator_fn(features, labels, mode):\n  if mode == tf.estimator.ModeKeys.TRAIN:\n    optimizer=xxx(learning_rate=GLOBAL_STEP*some process)\n<\/code><\/pre>\n\n<p>And does <code>tf.train.get_global_step<\/code> work?<\/p>\n","answers":[{"AnswerId":"57712129","CreationDate":"2019-08-29T14:14:58.800","ParentId":null,"OwnerUserId":"9637085","Title":null,"Body":"<p>You can also use this:<\/p>\n\n<pre><code>if mode == tf.estimator.ModeKeys.TRAIN:\n   global_step = tf.train.get_or_create_global_step()\n   learning_rate = learning_rate_fn(global_step)\n<\/code><\/pre>\n\n<p>where learning_rate_fn is a function that you can modify your learning rate.<\/p>\n\n<p>For more info look at <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/57e075203f8fba8d85e6b74f17f63d0a07da233a\/official\/resnet\/resnet_run_loop.py#L367\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":57711763,"AnswerCount":0,"Tags":"<tensorflow><keras><google-cloud-platform><hyperparameters>","CreationDate":"2019-08-29T13:54:39.933","AcceptedAnswerId":null,"OwnerUserId":3411238.0,"Title":"GCP: Hyperparameter tuning does not retrieve the defined evaluation metric","Body":"<p>I have built a machine learning model using keras and then I use the method <code>tf.keras.estimator.model_to_estimator<\/code> to convert that model to an estimator object and train it on the cloud. Below is the relevant code:<\/p>\n\n<pre><code>def categorical_accuracy(labels, predictions):\n    pred_values = predictions['2_dense']\n    return {'accuracy': tf.compat.v1.metrics.accuracy(labels, pred_values, name=\"accuracy\")}\n\n\ndef train(model_dir, data_dir, hyperparams):\n    keras_model = model.create_keras_model(hyperparams)\n    my_checkpointing_config = tf.estimator.RunConfig(\n        save_checkpoints_secs=10,  # Save checkpoints every 20 minutes.\n        keep_checkpoint_max=10,  # Retain the 10 most recent checkpoints.\n        tf_random_seed=0\n    )\n    estimator =  tf.keras.estimator.model_to_estimator(keras_model=keras_model, \n                                                       model_dir=model_dir,                                              \n                                                       config=my_checkpointing_config)\n\n\n    estimator = tf.contrib.estimator.add_metrics(estimator, categorical_accuracy)\n\n    train_spec = tf.estimator.TrainSpec(\n                     get_input_fn(data_dir, hyperparams['train_filename'], \n                     hyperparams)\n                 )\n    exporter = tf.estimator.BestExporter(name=\"best_exporter\",\n                   serving_input_receiver_fn=get_serving_input_fn(hyperparams),\n                   exports_to_keep=1\n               )\n\n    eval_spec = tf.estimator.EvalSpec(\n                    get_input_fn(data_dir, hyperparams['eval_filename'], hyperparams),\n                    exporters=exporter, throttle_secs=10\n                )\n\n    tf.estimator.train_and_evaluate(estimator=estimator,\n                                    train_spec=train_spec, eval_spec=eval_spec)\n<\/code><\/pre>\n\n<p>So essentially what I have done is that I have build some kind of keras model using a function I have defined named <code>create_keras_model<\/code>. Then, I convert this model to an estimator object and I add the custom <code>categorical_accuracy<\/code> metric to the model.<\/p>\n\n<p>I added the custom \"accuracy\" metric according to the GCP guide for hyperparameter tuning on <a href=\"https:\/\/cloud.google.com\/ml-engine\/docs\/using-hyperparameter-tuning\" rel=\"nofollow noreferrer\">https:\/\/cloud.google.com\/ml-engine\/docs\/using-hyperparameter-tuning<\/a>. I guess the only significant difference between my code and this of the example is that I convert the keras model to an estimator.<\/p>\n\n<p>Could that be the reason why I get no accuracy values when I do hyperparameter tuning on cloud? Anybody knows why and how I can solve my problem?<\/p>\n\n<p>I should also note that the \"accuracy\" node does have an output value when I train the network and monitor training in tensorboard. So it is just the hyperparameter tuning framework that doesn't work as intended.<\/p>\n","answers":[]}
+{"QuestionId":57712087,"AnswerCount":0,"Tags":"<python><tensorflow><shapes><uniform>","CreationDate":"2019-08-29T14:12:37.893","AcceptedAnswerId":null,"OwnerUserId":6601225.0,"Title":"tensorflow.random.uniform not accepting tuple","Body":"<p>I'm learning TensorFlow 2.0. I'm trying to use the function or method<\/p>\n\n<pre><code>uniform = tf.random.uniform(shape=(5), minval=5, maxval=10)\n<\/code><\/pre>\n\n<p>but Python throws this error:<\/p>\n\n<pre><code>Truncated Traceback (Use C-c C-$ to view full TB):\n\/usr\/lib\/python3.7\/site-packages\/six.py in raise_from(value, from_value)\n\nInvalidArgumentError: shape must be a vector of {int32,int64}, got shape [] [Op:RandomUniform]\n<\/code><\/pre>\n\n<p>When I try to execute this line using <code>shape=[5]<\/code>, it executes without problem. It's a TensorFlow problem, or I'm not understanding something?<\/p>\n","answers":[]}
+{"QuestionId":57712437,"AnswerCount":2,"Tags":"<tensorflow><keras><keras-layer>","CreationDate":"2019-08-29T14:32:25.167","AcceptedAnswerId":null,"OwnerUserId":11927480.0,"Title":"how to create a keras layer that takes effect only during evaluation phase (and that is transparent during training)?","Body":"<p>I want to add to my model a layer that, during evaluation, takes the input, applies some transformations (a quantization in this case, but can be whatever) and return it as the output. This layer must, however, be completely transparent during training, meaning that it must return the same input tensor.<\/p>\n\n<p>I have written the following function<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.layers import Lambda\nimport keras.backend as K\n\ndef myquantize(x):\n    return K.in_test_phase( K.clip(K.round(x*(2**5))\/(2**5),-3.9,3.9) , x)\n<\/code><\/pre>\n\n<p>which I then use via a Lambda layer<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>y = keras.layers.Conv1D(**args1)\ny = keras.layers.AveragePooling1D(pool_size=2)(y)\ny = keras.layers.Lambda(myquantize)(y)\ny = keras.layers.Conv1D(**args2)\n#...\n<\/code><\/pre>\n\n<p>Now, in principle the K.in_test_phase should return x during training, and that expression during test.\nHowever, training the network with such layer prevent the network from learning (i.e. the train loss stops decreasing after 3 epochs), while if I remove it the network keeps training normally. I assume this layer is not actually transparent during training as expected.<\/p>\n","answers":[{"AnswerId":"57716345","CreationDate":"2019-08-29T18:57:44.890","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p><a href=\"https:\/\/keras.io\/backend\/#in_test_phase\" rel=\"nofollow noreferrer\"><code>in_test_phase<\/code><\/a> has a parameter of <code>training<\/code> which you can explicitly set to indicate whether you are training or not. If you don't set it explicitly, then the value of <a href=\"https:\/\/keras.io\/backend\/#learning_phase\" rel=\"nofollow noreferrer\"><code>learning_phase<\/code><\/a> is used. This value keeps changing when you reset the graph or when you call different types of fit\/predict\/evaluate functions of model.<\/p>\n\n<p>Since your full code isn't present, you can make use of <code>training<\/code> parameter. Set it to True during training. Then save the weights of the model using <a href=\"https:\/\/keras.io\/models\/about-keras-models\/#about-keras-models\" rel=\"nofollow noreferrer\"><code>save_weights<\/code><\/a> function of model. When you wish to test your model, set the <code>training<\/code> parameter to False. Then load the weights using <a href=\"https:\/\/keras.io\/models\/about-keras-models\/#about-keras-models\" rel=\"nofollow noreferrer\"><code>load_weights<\/code><\/a> function and then you can proceed accordingly.<\/p>\n"},{"AnswerId":"57754832","CreationDate":"2019-09-02T09:52:37.447","ParentId":null,"OwnerUserId":"11927480","Title":null,"Body":"<p>For those who are in a similar situation, I created a custom layer like the following, which I only use during training:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class MyLayer(keras.layers.Layer):\n\n    def __init__(self, **kwargs):\n        super(MyLayer, self).__init__(**kwargs)\n\n    def compute_output_shape(self, input_shape):\n        return input_shape\n\n    def call(self, inputs, **kwargs):\n        x=inputs\n        return K.identity(x)\n<\/code><\/pre>\n\n<p>note that this layer always returns the input tensor, but it serves as 'placeholder' for the next step. On the evaluation part of the code, I wrote the following code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class MyLayer(keras.layers.Layer):\n\n    def __init__(self, **kwargs):\n        super(MyLayer, self).__init__(**kwargs)\n\n    def compute_output_shape(self, input_shape):\n        return input_shape\n\n    def call(self, inputs, **kwargs):\n        x=inputs\n        return #Your actual processing here\n<\/code><\/pre>\n\n<p>Here, the only difference is that you actually perform the desired processing steps on your tensor. When I load my stored model, I pass this class as custom object<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = keras.models.load_model(model_file,custom_objects={'MyLayer':MyLayer})\n<\/code><\/pre>\n\n<p>be careful to pass as MyLayer the one where the actual processing is performed.<\/p>\n\n<p>This is my solution, other suggestions are welcome<\/p>\n"}]}
+{"QuestionId":57712653,"AnswerCount":1,"Tags":"<python><tensorflow><jupyter-notebook><google-colaboratory>","CreationDate":"2019-08-29T14:44:59.053","AcceptedAnswerId":57714797.0,"OwnerUserId":11922064.0,"Title":"Tensorflow import error: libcudnn.so.5 No such file or directory","Body":"<p>I am trying to work with <code>tensorflow-gpu==1.2.0<\/code> in google-colab.\nI did install the package, cuda-8.0 and download and move libcudnn.so.5 into \/usr\/local\/cuda\/lib64 and still I get this error:<\/p>\n\n<pre><code>ImportError: Traceback (most recent call last):\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 41, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 29, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 25, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\nImportError: libcudnn.so.5: cannot open shared object file: No such file or directory\n<\/code><\/pre>\n\n<p>I tried to follow the next solution:\n<a href=\"https:\/\/stackoverflow.com\/questions\/41991101\/importerror-libcudnn-when-running-a-tensorflow-program\">ImportError: libcudnn when running a TensorFlow program<\/a>\nBut it does not work for me. \nI would be happy to get some help.<\/p>\n","answers":[{"AnswerId":"57714797","CreationDate":"2019-08-29T16:58:06.597","ParentId":null,"OwnerUserId":"196964","Title":null,"Body":"<p>If the file cannot be found, it's not in a place the linker is looking.<\/p>\n\n<p>From the <a href=\"https:\/\/linux.die.net\/man\/1\/ld\" rel=\"nofollow noreferrer\">ld man page<\/a>:<\/p>\n\n<blockquote>\n  <p>The linker uses the following search paths to locate required shared libraries: <\/p>\n  \n  <ol>\n  <li>Any directories specified by -rpath-link options.<\/li>\n  <li>Any directories specified by -rpath options. The difference between -rpath and -rpath-link is that directories specified by -rpath options are included in the executable and used at runtime, whereas the -rpath-link option is only effective at link time. Searching -rpath in this way is only supported by native linkers and cross linkers which have been configured with the --with-sysroot option.<\/li>\n  <li>On an ELF system, for native linkers, if the -rpath and -rpath-link options were not used, search the contents of the environment variable \"LD_RUN_PATH\".  <\/li>\n  <li>On SunOS, if the -rpath option was not used, search any directories specified using -L options.<\/li>\n  <li>For a native linker, the search the contents of the environment variable \"LD_LIBRARY_PATH\".<\/li>\n  <li>For a native ELF linker, the directories in \"DT_RUNPATH\" or \"DT_RPATH\" of a shared library are searched for shared libraries needed by it. The \"DT_RPATH\" entries are ignored if \"DT_RUNPATH\" entries exist.<\/li>\n  <li>The default directories, normally \/lib and \/usr\/lib.<\/li>\n  <li>For a native linker on an ELF system, if the file \/etc\/ld.so.conf exists, the list of directories found in that file. <\/li>\n  <\/ol>\n  \n  <p>If the required shared library is not found, the linker will issue a warning and continue with the link. <\/p>\n<\/blockquote>\n\n<p>Also see <a href=\"https:\/\/linux.die.net\/man\/8\/ldconfig\" rel=\"nofollow noreferrer\">ldconfig man page<\/a>:<\/p>\n\n<blockquote>\n  <p>ldconfig creates the necessary links and cache to the most recent shared libraries found in the directories specified on the command line, in the file \/etc\/ld.so.conf, and in the trusted directories (\/lib and \/usr\/lib). The cache is used by the run-time linker, ld.so or ld-linux.so. ldconfig checks the header and filenames of the libraries it encounters when determining which versions should have their links updated. <\/p>\n  \n  <p>...<\/p>\n  \n  <h1>Options<\/h1>\n  \n  <h2>-v<\/h2>\n  \n  <p>Verbose mode. Print current version number, the name of each directory as it is scanned, and any links that are created. Overrides quiet mode.<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":57712963,"AnswerCount":2,"Tags":"<machine-learning><keras><classification><lstm><multilabel-classification>","CreationDate":"2019-08-29T15:01:55.330","AcceptedAnswerId":null,"OwnerUserId":8587782.0,"Title":"How can I add specific numbers of each classification that my NN can predict?","Body":"<p>I have an np-array something like one-hot-encoding with 0-1. For each sample I always have 15 zeros and 5 ones. What can I do to make it predict 5 ones and 15 zeros only? I am using the keras library is there a setting I can apply so that my model have to predict exactly 15 zeros and 5 ones?<\/p>\n\n<p>-<\/p>\n\n<p>example of input = [0 0 0 1 0 1 0 0 1 0 0 1 0 0 0 1 0 0 0 0]<\/p>\n\n<pre><code>#Building RNN\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.layers import LSTM\nfrom keras.layers import Dropout\n\nregressor = Sequential()\n\nregressor.add(LSTM(units = 50, return_sequences = True, input_shape = (X_train.shape[1],20)))\nregressor.add(Dropout(0.2))\n\nregressor.add(LSTM(units = 50, return_sequences = True))\nregressor.add(Dropout(0.2))\n\nregressor.add(LSTM(units = 50, return_sequences = True))\n\nregressor.add(LSTM(units = 50))\nregressor.add(Dropout(0.2))\n\nregressor.add(Dense(units = 20, activation='sigmoid'))\n\n# Compiling RNN\nregressor.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>I would like my model always predicting 15 zeros and 5 ones.<\/p>\n","answers":[{"AnswerId":"57715527","CreationDate":"2019-08-29T17:52:52.253","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>You can do this with rigorous training.  Heavily penalize anything that has any other distribution. Experiment with heavier penalties for the difference between the prediction and exactly 5 <code>1<\/code>s.<\/p>\n\n<p>You can do most or all of this by writing an appropriate loss function.<\/p>\n"},{"AnswerId":"57714448","CreationDate":"2019-08-29T16:34:19.467","ParentId":null,"OwnerUserId":"4257991","Title":null,"Body":"<p>You may also try to define your custom loss function which returns according your format <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def custom_loss(y_true, y_pred):\n\n  # calculate binary_crossentropy and reshape the result according your need\n  ...\n\n  return K.variable(...)\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>regressor.compile(loss=custom_loss, optimizer='adam', metrics=['accuracy'])\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":57713020,"AnswerCount":1,"Tags":"<deep-learning><conv-neural-network><pytorch><transformation><image-segmentation>","CreationDate":"2019-08-29T15:05:36.600","AcceptedAnswerId":null,"OwnerUserId":6642287.0,"Title":"Pytorch : Applying lambda transformations in case of segmentation tasks in custom transformation function","Body":"<p>I am trying to perform multiclass image segmentation. There are some transformations which should be applied to both image and masks like rotations and flips etc. I followed <a href=\"https:\/\/discuss.pytorch.org\/t\/torchvision-transfors-how-to-perform-identical-transform-on-both-image-and-target\/10606\/7?u=rishav_sapahia\" rel=\"nofollow noreferrer\">this<\/a>\npost and applied it but the issue is I have certain lambda transformations also which I need to apply but I can't apply using the above method.How to apply them.\nBelow are the transformations I need-<\/p>\n\n<pre><code>data_transforms = transforms.Compose([transforms.RandomCrop((512,512)),\n                             transforms.Lambda(gaussian_blur),\n                             transforms.Lambda(elastic_transform),\n                             transforms.RandomRotation([+90,+180]),\n                             transforms.RandomRotation([+180,+270]),\n                             transforms.RandomHorizontalFlip(),\n                             transforms.ToTensor(),\n                             transforms.Normalize(mean=train_mean, std=train_std)\n                           ])\n<\/code><\/pre>\n\n<p>My Dataloader for identical transformations for the image and mask is below, I wasn't able to apply my lambda transformations in it, it was throwing an error \n<code>AttributeError: module 'torchvision.transforms.functional' has no attribute 'Lambda'\n<\/code><\/p>\n\n<pre><code>class DataLoaderSegmentation(data.Dataset):\ndef __init__(self,folder_path,transform = None):\n    super(DataLoaderSegmentation, self).__init__()\n    self.img_files = glob.glob(os.path.join(folder_path,'images','*.tif'))\n    self.mask_files = glob.glob(os.path.join(folder_path,'mask','*.bmp'))\n    # self.transform = transform\n\n\ndef mask_to_class(self,mask):\n    target = torch.from_numpy(mask)\n    h,w = target.shape[0],target.shape[1]\n    masks = torch.empty(h, w, dtype=torch.long)\n    colors = torch.unique(target.view(-1,target.size(2)),dim=0).numpy()\n    target = target.permute(2, 0, 1).contiguous()\n    mapping = {tuple(c): t for c, t in zip(colors.tolist(), range(len(colors)))}\n    for k in mapping:\n        idx = (target==torch.tensor(k, dtype=torch.uint8).unsqueeze(1).unsqueeze(2))\n        validx = (idx.sum(0) == 3) \n        masks[validx] = torch.tensor(mapping[k], dtype=torch.long)\n    return masks\n\ndef transform(self,image,mask):\n    i, j, h, w = transforms.RandomCrop.get_params(\n    image, output_size=(512, 512))\n    image = TF.crop(image, i, j, h, w)\n    mask = TF.crop(mask, i, j, h, w)\n\n    #image = TF.Lambda(gaussian_blur),\n\n    #image = TF.Lambda(elastic_transform)\n    # Random horizontal flipping\n    if random.random() &gt; 0.5:\n        image = TF.hflip(image)\n        mask = TF.hflip(mask)\n\n    image = TF.rotate(image,90)\n    mask = TF.rotate(mask,90)\n    image = TF.rotate(image,180)\n    mask = TF.rotate(mask,180)\n    image = TF.rotate(image,270)\n    mask = TF.rotate(mask,270)\n\n    # Transform to tensor\n    image = TF.to_tensor(image)\n    return image, mask\n\ndef __getitem__(self, index):\n    img_path = self.img_files[index]\n    mask_path = self.mask_files[index]\n    data = Image.open(img_path)\n    label = Image.open(mask_path)\n    data,label = self.transform(data,label)\n    label = np.array(label)  \n    mask = self.mask_to_class(label)\n    return data,mask\n\n\ndef __len__(self):\n    return len(self.img_files)\n<\/code><\/pre>\n\n<p>How to do it, or is there any other way to do it?<\/p>\n\n<p>Edit---\nFollowing is my solution for it-<\/p>\n\n<pre><code>class DataLoaderSegmentation(data.Dataset):\ndef __init__(self,folder_path,transform = None):\n    super(DataLoaderSegmentation, self).__init__()\n    self.img_files = glob.glob(os.path.join(folder_path,'images','*.tif'))\n    self.mask_files = glob.glob(os.path.join(folder_path,'mask','*.bmp'))\n    self.transforms = transform\n    #for img_path in img_files:\n     #   self.mask_files.append(os.path.join(folder_path,'masks',os.path.basename(img_path))\n\ndef mask_to_class(self,mask):\n    target = torch.from_numpy(mask)\n    h,w = target.shape[0],target.shape[1]\n    masks = torch.empty(h, w, dtype=torch.long)\n    colors = torch.unique(target.view(-1,target.size(2)),dim=0).numpy()\n    target = target.permute(2, 0, 1).contiguous()\n    mapping = {tuple(c): t for c, t in zip(colors.tolist(), range(len(colors)))}\n    for k in mapping:\n        idx = (target==torch.tensor(k, dtype=torch.uint8).unsqueeze(1).unsqueeze(2))\n        validx = (idx.sum(0) == 3) \n        masks[validx] = torch.tensor(mapping[k], dtype=torch.long)\n    return masks\n\ndef transform(self,image,mask):\n    i, j, h, w = transforms.RandomCrop.get_params(\n    image, output_size=(512, 512))\n    image = TF.crop(image, i, j, h, w)\n    mask = TF.crop(mask, i, j, h, w)\n\n    #image = TF.Lambda(gaussian_blur),\n   # mask = \n    #image = TF.Lambda(elastic_transform)\n    # Random horizontal flipping\n    image = F.Lambda(gaussian_blur)\n    if random.random() &gt; 0.5:\n        image = TF.hflip(image)\n        mask = TF.hflip(mask)\n\n    image = TF.rotate(image,90)\n    mask = TF.rotate(mask,90)\n    image = TF.rotate(image,180)\n    mask = TF.rotate(mask,180)\n    image = TF.rotate(image,270)\n    mask = TF.rotate(mask,270)\n\n    # Transform to tensor\n    #image = TF.to_tensor(image)\n    #mask = TF.to_tensor(mask)\n    return image, mask\n\ndef __getitem__(self, index):\n    img_path = self.img_files[index]\n    mask_path = self.mask_files[index]\n    data = Image.open(img_path)\n    label = Image.open(mask_path)\n   # label = np.array(label)\n    data,label = self.transform(data,label)\n    label = np.array(label)\n    data = np.array(data)\n    #label = np.transpose(label,(2,0,1))\n    mask = self.mask_to_class(label)\n    if transforms is not None:\n        data = self.transforms(data)\n    return data,mask\n   # return data, torch.from_numpy(label).long()\n\ndef __len__(self):\n    return len(self.img_files)\n<\/code><\/pre>\n\n<p>and the transformation are below-<\/p>\n\n<pre><code>data_transforms = transforms.Compose([\n                             transforms.Lambda(gaussian_blur),\n                             transforms.Lambda(elastic_transform),\n                             transforms.ToTensor(),\n                             transforms.Normalize(mean=train_mean, std=train_std)\n                           ])\n<\/code><\/pre>\n","answers":[{"AnswerId":"57713856","CreationDate":"2019-08-29T15:54:34.130","ParentId":null,"OwnerUserId":"819544","Title":null,"Body":"<p>Looks like the class was moved from <code>torchvision.transforms.functional<\/code> to <a href=\"https:\/\/github.com\/pytorch\/vision\/blob\/3483342733673c3182bd5f8a4de3723a74ce5845\/torchvision\/transforms\/transforms.py#L301\" rel=\"nofollow noreferrer\"><code>torchvision.transforms.transforms<\/code><\/a>. I think you can probably just import it directly from <a href=\"https:\/\/github.com\/pytorch\/vision\/blob\/3483342733673c3182bd5f8a4de3723a74ce5845\/torchvision\/transforms\/__init__.py\" rel=\"nofollow noreferrer\"><code>torchvision.transforms<\/code><\/a><\/p>\n"}]}
+{"QuestionId":57713231,"AnswerCount":1,"Tags":"<tensorflow><keras><conv-neural-network><image-segmentation>","CreationDate":"2019-08-29T15:17:31.497","AcceptedAnswerId":57715798.0,"OwnerUserId":11515668.0,"Title":"Keras gives 'Not JSON Serializable' error when saving the model","Body":"<p>I'm implementing a fully convolutional neural network for image segmentation by using unet defined here<\/p>\n\n<p><a href=\"https:\/\/github.com\/zhixuhao\" rel=\"nofollow noreferrer\">https:\/\/github.com\/zhixuhao<\/a><\/p>\n\n<p>To give different weights to the pixels of different classes I defined an extra Lambda layer, as suggested here<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/48555820\/keras-binary-segmentation-add-weight-to-loss-function\/48577360#48577360\">Keras, binary segmentation, add weight to loss function<\/a> <\/p>\n\n<p>The problem is that Keras raises this error when saving the model<\/p>\n\n<pre><code>.....\nself.model.save(filepath, overwrite=True)\n.....\nTypeError: ('Not JSON Serializable:', b'\\n\\x15clip_by_value\/Minimum\\x12\\x07Minimum\\x1a\\x12conv2d_23\/Identity\\x1a\\x17clip_by_value\/Minimum\/y*\\x07\\n\\x01T\\x12\\x020\\x01')\n<\/code><\/pre>\n\n<p>My network is defined in an external function<\/p>\n\n<pre><code>def weighted_binary_loss(X):\n    y_pred, y_true, weights = X\n    loss = binary_crossentropy(y_true, y_pred)\n    weights_mask = y_true*weights[0] + (1.-y_true)*weights[1]\n    loss = multiply([loss, weights_mask])\n    return loss    \n\ndef identity_loss(y_true, y_pred):\n    return y_pred\n\n\ndef net()\n.....\n....\nconv10 = Conv2D(1, 1, activation = 'sigmoid')(conv9)\nw_loss = Lambda(weighted_binary_loss, output_shape=input_size, name='loss_output')([conv10, inputs, weights])\nmodel = Model(inputs = inputs, outputs = w_loss)\nmodel.compile(optimizer = Adam(lr = 1e-5), loss = identity_loss, metrics = ['accuracy'])\n\n<\/code><\/pre>\n\n<p>that I call in my main function<\/p>\n\n<pre><code>...\nmodel_checkpoint = ModelCheckpoint('temp_model.hdf5', monitor='loss',verbose=1, save_best_only=True)\nmodel.fit_generator(imgs,steps_per_epoch=20,epochs=1,callbacks=[model_checkpoint])\n<\/code><\/pre>\n\n<p>When I erase the Lambda layer, the error desappears<\/p>\n\n<pre><code>...\nconv10 = Conv2D(1, 1, activation = 'sigmoid')(conv9)\nmodel = Model(inputs = inputs, outputs = conv10)\nmodel.compile(optimizer = Adam(lr = 1e-5), loss = 'binary_crossentropy', metrics = ['accuracy'])\n<\/code><\/pre>\n\n<p>I'm using \nKeras==2.2.4, tensorflow-gpu==2.0.0b1<\/p>\n","answers":[{"AnswerId":"57715798","CreationDate":"2019-08-29T18:15:44.180","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>It appears that you are computing the loss in the layer of a model. It is not a good practice to accomodate the loss function as a layer. You can compute your weighted loss using custom loss function.<\/p>\n\n<p>So your code can be rewritten as follows:<\/p>\n\n<pre><code>def weighted_binary_loss(y_true, y_pred):\n    weights = [0.5, 0.6]  # Define your weights here\n    loss = binary_crossentropy(y_true, y_pred)\n    weights_mask = y_true*weights[0] + (1.-y_true)*weights[1]\n    loss = multiply([loss, weights_mask])\n    return loss  \n\nconv10 = Conv2D(1, 1, activation = 'sigmoid')(conv9)\nmodel = Model(inputs = inputs, outputs = conv10)\nmodel.compile(optimizer = Adam(lr = 1e-5), loss = weighted_binary_loss, metrics = ['accuracy'])\n<\/code><\/pre>\n\n<p>If it is needed that <code>weights<\/code> is a dynamic property and you have to send it as a separate parameter in loss function, you can follow this <a href=\"https:\/\/stackoverflow.com\/questions\/45961428\/make-a-custom-loss-function-in-keras\">question<\/a>.<\/p>\n"}]}
+{"QuestionId":57713676,"AnswerCount":0,"Tags":"<numpy><pyinstaller><theano><python-2.x><lasagne>","CreationDate":"2019-08-29T15:45:21.377","AcceptedAnswerId":null,"OwnerUserId":10881028.0,"Title":"Pyinstaller generating exe errors","Body":"<p>Background: I am working on a Mac. I made a program which scrapes and analyzes tweets using the following modules: tweet2vec, theanos, numpy, lasagne, openpyxl. The code runs fine by itself on my machine, obviously because i've installed the correct packages. However, I am trying to use pyinstaller so that the program can act as a standalone exe.<\/p>\n\n<p>Problem: The pyinstaller command will sometimes run and generate a faulty exe(doesnt actually run my code). Other times it will produce this error:<\/p>\n\n<blockquote>\n  <p>ImportError: cannot import name ccompiler\n  pre-safe-import-module hook failed, needs fixing.<\/p>\n<\/blockquote>\n\n<p>What I've done: I have run the command below several times with different pathing options and hooks. Sometimes I'll get the error mentioned above. Sometimes I'll get an exe which says it cant find certain modules (Am not currently able to generate an exe)<\/p>\n\n<p><code>pyinstaller --onefile --windowed --paths=\/Users\/NAME\/Q3Crawler\/Q3Crawler --paths=\/Users\/NAME\/Q3Crawler\/tweet2vec --path=\/Users\/NAME\/anaconda2 --paths=\/Users\/NAME\/anaconda2\/lib\/python2.7\/site-packages\/lasagne --paths=\/Users\/NAME\/anaconda2\/lib\/python2.7\/site-packages\/numpy --paths=\/Users\/NAME\/anaconda2\/pkgs\/theano-1.0.3-py27h04f5b5a_0\/lib\/python2.7\/site-packages\/theano --additional-hooks-dir=hooks\/ q3crawler.py<\/code><\/p>\n\n<p>I would like a standalone exe of my program that I can distribute to others.<\/p>\n\n<p>===========================EDIT============================================\nAfter some more research it seems like this is a common error involving numpy 1.16.4 and pyinstaller. I tried again with numpy 1.15.4 but unfortunately my code is dependent on the latest version of numpy. Is there a known workaround for this error involving numpy and pyinstaller? I am working on a Mac.<\/p>\n","answers":[]}
+{"QuestionId":57713804,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2019-08-29T15:51:39.493","AcceptedAnswerId":null,"OwnerUserId":9568153.0,"Title":"Can't interpret prediction with neural network use","Body":"<p>I'm trying to use TensorFlow in python, to make some prediction with cryptocurrency data. The problem is that the output of the prediction is like a 0.1-0.9 number whereas the cryptocurrency data should be a 10000-10100 format, and I don't find a solution to convert the 0.* number to the real one. <\/p>\n\n<p>I've try to create a ratio, with substrat max - min from predicted values, and max-min from tested data, and divide to have a ratio but when I multiply this ratio with prediction there is a big rate of error ( found a 14000 number instead of a 10000 one ) <\/p>\n\n<p>Here some code : <\/p>\n\n<pre><code>train_start = 0\ntrain_end = int(np.floor(0.7*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\nScale data:\n\nfrom sklearn.preprocessing import MinMaxScaler\nscaler = MinMaxScaler()\ndata_train = scaler.fit_transform(data_train)\ndata_test = scaler.transform(data_test)\n\nBuild X and y:\n\nX_train = data_train[:, 1:]\ny_train = data_train[:, 0]\nX_test = data_test[:, 1:]\ny_test = data_test[:, 0]\n\n.\n.\n.\n\nn_data = 10\nn_neurons_1 = 1024\nn_neurons_2 = 512\nn_neurons_3 = 256\nn_neurons_4 = 128\nn_target = 1\nX = tf.compat.v1.placeholder(dtype=tf.compat.v1.float32, shape=[None, n_data])\nY = tf.compat.v1.placeholder(dtype=tf.compat.v1.float32, shape=[None])\n\nHidden layer\n..\nOutput layer (must be transposed)\n..\nCost function\n..\nOptimizer\n..\n\nMake Session:\n\nsess = tf.compat.v1.Session()\n\nRun initializer:\n\nsess.run(tf.compat.v1.global_variables_initializer())\n\nSetup interactive plot:\n\nplt.ion()\nfig = plt.figure()\nax1 = fig.add_subplot(111)\nline1, = ax1.plot(y_test)\nline2, = ax1.plot(y_test*0.5)\nplt.show()\n\n\nepochs = 10\nbatch_size = 256\nfor e in range(epochs):\n\n# Shuffle training data\nshuffle_indices = np.random.permutation(np.arange(len(y_train)))\nX_train = X_train[shuffle_indices]\ny_train = y_train[shuffle_indices]\n\n# Minibatch training\nfor 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    sess.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 = sess.run(out, feed_dict={X: X_test})\n\n        #This pred var is the output of the prediction \n<\/code><\/pre>\n\n<p>I persiste my result in a file and this is what its looks like : <\/p>\n\n<p>2019-08-21 06-AM;15310.444858356934;0.50021994;<\/p>\n\n<p>2019-08-21 12-PM;14287.717187390663;0.46680558;<\/p>\n\n<p>2019-08-21 06-PM;14104.63871795706;0.46082407;<\/p>\n\n<p>For example, the last prediction is 0,46 but when I try to convert it I found 14104 whereas it should be nearer a 10000 value<\/p>\n\n<p>Does anyone have an idea how to convert those predictions? <\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"57715339","CreationDate":"2019-08-29T17:39:09.303","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>You will have to make use of <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.MinMaxScaler.html#sklearn.preprocessing.MinMaxScaler.inverse_transform\" rel=\"nofollow noreferrer\"><code>inverse_transform<\/code><\/a> of <code>MinMaxScaler<\/code> to convert back the output you are getting in range of 0-1. <\/p>\n\n<p>You have not given your model, but I believe you are making use of regression task with few dense layers. You will have to keep minimizing your loss. If you are using mean squared error, the larger the loss, more is the likelihood your output will be far away from the desired set of results.<\/p>\n\n<p>Even after your loss is a small number and the result is coming good for train samples, but the prediction is bad for test dataset, you may have to consider increasing your train dataset so that more possibilities are covered. If that is not possible, consider reducing the number of neurons in your neural network so that it stops over-fitting.<\/p>\n\n<p>You can do some postprocessing to restrict the output to some desired range.<\/p>\n"}]}
+{"QuestionId":57713991,"AnswerCount":0,"Tags":"<keras><neural-network><conv-neural-network><layer>","CreationDate":"2019-08-29T16:04:10.557","AcceptedAnswerId":null,"OwnerUserId":1019129.0,"Title":"Multiple input and multiple output layers?","Body":"<p>How do I configure two (or more) input layers and two (or more) output layers.<\/p>\n\n<p>What I have in mind ? Let say I have a sentence and tags Ex: <\/p>\n\n<pre><code>The apple is red \n DT   NN  VBZ JJ\n<\/code><\/pre>\n\n<p>then I encode them as vocabulary-based-vector (for the sentence) and category-vector (for the tags):<\/p>\n\n<pre><code> 55 12 2 41\n 0 1 0 0 0 1 0 1 0 0 1\n<\/code><\/pre>\n\n<p>As you see I have two separate vectors. I imagine I will pass sent-vec to embeeding-layer and the category-vec to a normal layer. <\/p>\n\n<p>The question is how do I do that ?<\/p>\n\n<p>Second I want also the same arrangement for the output ?<\/p>\n\n<p>What do I use in the middle ? normal, CONV layers !<\/p>\n\n<p>Should I use category-vector for the tags OR vocab-vec would be better ?<\/p>\n\n<p>Examples even in pseudo code are welcome ..<\/p>\n","answers":[]}
+{"QuestionId":57714233,"AnswerCount":1,"Tags":"<python><tensorflow><keras><image-generation>","CreationDate":"2019-08-29T16:20:07.180","AcceptedAnswerId":null,"OwnerUserId":5838180.0,"Title":"Why custom image generator in keras gives error \"object cannot be interpreted as an integer\"?","Body":"<p>I used a <a href=\"https:\/\/stackoverflow.com\/a\/56252273\/5838180\">template<\/a> of a custom image generator in keras, so that I can use hdf5 files as input. Initially, the code was giving a \"shape\" error, so I only included <code>from tensorflow.python.keras.utils.data_utils import Sequence<\/code> following <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/12586\" rel=\"nofollow noreferrer\">this post<\/a>. Now I use it in this form, as you can also see in my <a href=\"https:\/\/colab.research.google.com\/drive\/1vJ-pJYY36WVpx7PlmpWktJrk6MqpavXT\" rel=\"nofollow noreferrer\">colab notebook<\/a>:<\/p>\n\n<pre><code>from numpy.random import uniform, randint\nfrom tensorflow.python.keras.preprocessing.image import ImageDataGenerator\nimport numpy as np\nfrom tensorflow.python.keras.utils.data_utils import Sequence\n\nclass CustomImagesGenerator(Sequence):\n    def __init__(self, x, zoom_range, shear_range, rescale, horizontal_flip, batch_size):\n        self.x = x\n        self.zoom_range = zoom_range\n        self.shear_range = shear_range\n        self.rescale = rescale\n        self.horizontal_flip = horizontal_flip\n        self.batch_size = batch_size\n        self.__img_gen = ImageDataGenerator()\n        self.__batch_index = 0\n\n    def __len__(self):\n        # steps_per_epoch, if unspecified, will use the len(generator) as a number of steps.\n        # hence this\n        return np.floor(self.x.shape[0]\/self.batch_size)\n\n    # @property\n    # def shape(self):\n    #     return self.x.shape\n\n    def next(self):\n        return self.__next__()\n\n    def __next__(self):\n        start = self.__batch_index*self.batch_size\n        stop = start + self.batch_size\n        self.__batch_index += 1\n        if stop &gt; len(self.x):\n            raise StopIteration\n        transformed = np.array(self.x[start:stop])  # loads from hdf5\n        for i in range(len(transformed)):\n            zoom = uniform(self.zoom_range[0], self.zoom_range[1])\n            transformations = {\n                'zx': zoom,\n                'zy': zoom,\n                'shear': uniform(-self.shear_range, self.shear_range),\n                'flip_horizontal': self.horizontal_flip and bool(randint(0,2))\n            }\n            transformed[i] = self.__img_gen.apply_transform(transformed[i], transformations)\n        import pdb;pdb.set_trace()\n        return transformed * self.rescale\n<\/code><\/pre>\n\n<p>And I call the generator with:<\/p>\n\n<pre><code>import h5py\nimport tables \n\nin_hdf5_file = tables.open_file(\"gdrive\/My Drive\/Colab Notebooks\/dataset.hdf5\", mode='r')\nimages = in_hdf5_file.root.train_img\n\nmy_gen = CustomImagesGenerator(\n    images,\n    zoom_range=[0.8, 1],\n    batch_size=32,\n    shear_range=6, \n    rescale=1.\/255, \n    horizontal_flip=False\n)\n\nclassifier.fit_generator(my_gen, steps_per_epoch=100, epochs=1, verbose=1)\n<\/code><\/pre>\n\n<p>The import of <code>Sequence<\/code> resolved the \"shape\" error, but now I am getting the error:<\/p>\n\n<blockquote>\n  <p>Exception in thread Thread-5: Traceback (most recent call last):<br>\n  File \"\/usr\/lib\/python3.6\/threading.py\", line 916, in _bootstrap_inner\n      self.run()   File \"\/usr\/lib\/python3.6\/threading.py\", line 864, in run\n      self._target(*self._args, **self._kwargs)   File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/utils\/data_utils.py\",\n  line 742, in _run\n      sequence = list(range(len(self.sequence))) TypeError: 'numpy.float64' object cannot be interpreted as an integer<\/p>\n<\/blockquote>\n\n<p>How can I resolve this? I suspect it might be again a conflict in the keras packages and don't know how to tackle it.<\/p>\n","answers":[{"AnswerId":"57747545","CreationDate":"2019-09-01T16:06:54.533","ParentId":null,"OwnerUserId":"10190971","Title":null,"Body":"<p>Example usage with <code>model.fit()<\/code> in your case:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.keras.utils import to_categorical\nimport tensorflow as tf\nimport tables\n\n#define your model\n\n...\n\n#load your data from an hdf5 file\nin_hdf5_file = tables.open_file(\"path\/to\/your\/dataset.hdf5\", mode='r')\nx = in_hdf5_file.root.train_img[:]\ny = in_hdf5_file.root.train_labels[:]\n\nyourModel.fit(x, to_categorical(y, 3), epochs=2, batch_size=5)\n<\/code><\/pre>\n\n<p>For more info read my comments to your original post, or feel free to ask.<\/p>\n\n<p>EDIT: I fixed your generator, now it only needs a path to your hdf5 file<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import *\nfrom tensorflow.keras.utils import to_categorical\n\nimport numpy as np\nfrom tensorflow.python.keras.utils.data_utils import Sequence\nimport tensorflow as tf\n\nimport tables\n\n#define your model\n\n...\n\n#training\ndef h5data_generator(path, batch_size=1):\n    batch_index = 0\n    while 1:\n        with tables.open_file(path, mdoe='r') as f:\n            x = f.root.train_img[batch_index:batch_index + batch_size]\n            y = f.root.train_labels[batch_index:batch_index + batch_size]\n\n            if batch_index &gt;= x.shape[0]:\n                batch_index = 0\n\n            batch_index += 1\n\n            yield (x, to_categorical(y, 3))\n\n            del x\n            del y\n\n\nmy_gen = h5data_generator(\"path\/to\/your\/dataset.hdf5\")\n\nyourModel.fit_generator(my_gen, steps_per_epoch=100, epochs=20, verbose=1)\n<\/code><\/pre>\n\n<p>The problem with your generator was bad data output on step, it wasn't outputting <code>(x, y)<\/code>, there was no way it could, it was outputting <code>x<\/code>(image in your case), also since it was using <code>Sequential<\/code> keras tried to interpret it as an object that used it's api(not the case in your generator). Also it doesn't have to be a <code>class<\/code>, it needs to be <a href=\"https:\/\/wiki.python.org\/moin\/Generators\" rel=\"nofollow noreferrer\">a python generator<\/a>, as indicated by an example in keras it self(doc string of <code>fit_generator()<\/code>),<\/p>\n\n<p><code>fit_generator.__doc__<\/code>:<\/p>\n\n<pre><code>Fits the model on data yielded batch-by-batch by a Python generator.\n\n    The generator is run in parallel to the model, for efficiency.\n    For instance, this allows you to do real-time data augmentation\n    on images on CPU in parallel to training your model on GPU.\n\n    The use of `keras.utils.Sequence` guarantees the ordering\n    and guarantees the single use of every input per epoch when\n    using `use_multiprocessing=True`.\n\n    Arguments:\n        generator: A generator or an instance of `Sequence`\n          (`keras.utils.Sequence`)\n            object in order to avoid duplicate data\n            when using multiprocessing.\n            The output of the generator must be either\n            - a tuple `(inputs, targets)`\n            - a tuple `(inputs, targets, sample_weights)`.\n            This tuple (a single output of the generator) makes a single batch.\n            Therefore, all arrays in this tuple must have the same length (equal\n            to the size of this batch). Different batches may have different\n              sizes.\n            For example, the last batch of the epoch is commonly smaller than\n              the\n            others, if the size of the dataset is not divisible by the batch\n              size.\n            The generator is expected to loop over its data\n            indefinitely. An epoch finishes when `steps_per_epoch`\n            batches have been seen by the model.\n        steps_per_epoch: Total number of steps (batches of samples)\n            to yield from `generator` before declaring one epoch\n            finished and starting the next epoch. It should typically\n            be equal to the number of samples of your dataset\n            divided by the batch size.\n            Optional for `Sequence`: if unspecified, will use\n            the `len(generator)` as a number of steps.\n        epochs: Integer, total number of iterations on the data.\n        verbose: Verbosity mode, 0, 1, or 2.\n        callbacks: List of callbacks to be called during training.\n        validation_data: This can be either\n            - a generator for the validation data\n            - a tuple (inputs, targets)\n            - a tuple (inputs, targets, sample_weights).\n        validation_steps: Only relevant if `validation_data`\n            is a generator. Total number of steps (batches of samples)\n            to yield from `generator` before stopping.\n            Optional for `Sequence`: if unspecified, will use\n            the `len(validation_data)` as a number of steps.\n        validation_freq: Only relevant if validation data is provided. Integer\n            or `collections.Container` instance (e.g. list, tuple, etc.). If an\n            integer, specifies how many training epochs to run before a new\n            validation run is performed, e.g. `validation_freq=2` runs\n            validation every 2 epochs. If a Container, specifies the epochs on\n            which to run validation, e.g. `validation_freq=[1, 2, 10]` runs\n            validation at the end of the 1st, 2nd, and 10th epochs.\n        class_weight: Dictionary mapping class indices to a weight\n            for the class.\n        max_queue_size: Integer. Maximum size for the generator queue.\n            If unspecified, `max_queue_size` will default to 10.\n        workers: Integer. Maximum number of processes to spin up\n            when using process-based threading.\n            If unspecified, `workers` will default to 1. If 0, will\n            execute the generator on the main thread.\n        use_multiprocessing: Boolean.\n            If `True`, use process-based threading.\n            If unspecified, `use_multiprocessing` will default to `False`.\n            Note that because this implementation relies on multiprocessing,\n            you should not pass non-picklable arguments to the generator\n            as they can't be passed easily to children processes.\n        shuffle: Boolean. Whether to shuffle the order of the batches at\n            the beginning of each epoch. Only used with instances\n            of `Sequence` (`keras.utils.Sequence`).\n            Has no effect when `steps_per_epoch` is not `None`.\n        initial_epoch: Epoch at which to start training\n            (useful for resuming a previous training run)\n\n    Returns:\n        A `History` object.\n\n    Example:\n\n    ```python\n        def generate_arrays_from_file(path):\n            while 1:\n                f = open(path)\n                for line in f:\n                    # create numpy arrays of input data\n                    # and labels, from each line in the file\n                    x1, x2, y = process_line(line)\n                    yield ({'input_1': x1, 'input_2': x2}, {'output': y})\n                f.close()\n\n        model.fit_generator(generate_arrays_from_file('\/my_file.txt'),\n                            steps_per_epoch=10000, epochs=10)\n    ```\n    Raises:\n        ValueError: In case the generator yields data in an invalid format.\n<\/code><\/pre>\n\n<p>For more info check out the github page of keras, <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/engine\/training.py#L1510\" rel=\"nofollow noreferrer\"><code>fit_generator()<\/code> to be exact<\/a>, or again feel free to ask.<\/p>\n\n<p>EDIT 2: You can also pass <code>batch_size<\/code> to the <code>h5data_generator()<\/code>, which will set the batch size of data getting pulled from your dataset on single step.<\/p>\n"}]}
+{"QuestionId":57714476,"AnswerCount":1,"Tags":"<python><machine-learning><keras><conv-neural-network>","CreationDate":"2019-08-29T16:37:04.643","AcceptedAnswerId":57715051.0,"OwnerUserId":1290309.0,"Title":"Incompatible input data error in keras, dimensions mismatch ValueError","Body":"<p>I'm trying to understand how to write a Conv1D model on Keras but I keep running into a dimensionality mismatch error. Right now x_train = [1000,294] (1000 items and 294 features) and y_train = [1000,9] (1000 items and 9 labels). I keep getting the error message expected 3 dimensions but got 2 dimensions. But I try to fix it, it appears again. Some issues on github has suggested <code>Flatten()<\/code> but it is already there and no luck. Any idea what I'm missing? Thank you<\/p>\n\n<p>The error I got, <\/p>\n\n<ol>\n<li>ValueError: Input 0 is incompatible with layer conv1d_1: expected\nndim=3, found ndim=2<\/li>\n<\/ol>\n\n<p>Code is included below<\/p>\n\n<pre><code>tb = [keras.callbacks.TensorBoard(log_dir='.\/dllogs')]\n\nmain_input = Input(shape=(294, ), dtype='float32')\n\nx_train = np.array(x_train)\ny_train = np.array(y_train)\n\nx_dev = np.array(x_dev)\ny_dev = np.array(y_dev)\nx = Conv1D(128, 5, activation='relu')(main_input)\nx = MaxPooling1D(5)(x)\nx = Conv1D(128, 5, activation='relu')(x)\nx = MaxPooling1D(5)(x)\nx = Conv1D(128, 5, activation='relu')(x)\nx = MaxPooling1D(35)(x)\nx = Flatten()(x)\nx = Dropout(0.25)(x)\nx = Dense(128, activation='relu')(x)\n\npreds = Dense(pred_dim, activation='softmax')(x)\nmodel = Model(inputs=main_input, outputs=preds)\nmodel.compile(optimizer='adam', loss='kullback_leibler_divergence', metrics=['accuracy'])\nprint(model.summary())\nhistory_NN = model.fit(x_train, y_train, batch_size=BATCHSIZE, epochs=EPOCHS, callbacks=tb, validation_data=(x_dev, y_dev))\n<\/code><\/pre>\n\n<p><strong>Answer<\/strong><\/p>\n\n<ol>\n<li>The dimensions was an issue that was causing trouble, this was address by the comment by Today. \n\n<ol start=\"2\">\n<li>Next was the pooling dimensions, this is the dimensions reduce after pooling so after the whole thing, it didn't have 35 slides for the final pooling, when I changed it. Fixed. <\/li>\n<\/ol><\/li>\n<\/ol>\n\n<p>Resulting code, <\/p>\n\n<pre><code>tb = [keras.callbacks.TensorBoard(log_dir='.\/dllogs')]\nmain_input = Input(shape=(294, 1))#changed\n\nx_train = np.array(x_train)\ny_train = np.array(y_train)\n\nx_dev = np.array(x_dev)\ny_dev = np.array(y_dev)\nx_train = np.expand_dims(x_train, axis=-1) #changed\nx_dev = np.expand_dims(x_dev, axis=-1) #changed\nx = Conv1D(128, 5, activation='relu')(main_input)\nx = MaxPooling1D(5)(x)\nx = Conv1D(128, 5, activation='relu')(x)\nx = MaxPooling1D(5)(x)\nx = Conv1D(128, 5, activation='relu')(x)\nx = MaxPooling1D(11)(x)\nx = Flatten()(x)\nx = Dropout(0.25)(x)\nx = Dense(128, activation='relu')(x)\n\npreds = Dense(pred_dim, activation='softmax')(x)\nmodel = Model(inputs=main_input, outputs=preds)\nmodel.compile(optimizer='adam', loss='kullback_leibler_divergence', metrics=['accuracy'])\nprint(model.summary())\nhistory_NN = model.fit(x_train, y_train, batch_size=BATCHSIZE, epochs=EPOCHS, callbacks=tb, validation_data=(x_dev, y_dev))\n<\/code><\/pre>\n","answers":[{"AnswerId":"57715051","CreationDate":"2019-08-29T17:16:02.117","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p><code>Conv1D<\/code> layer expects <strong>sequence<\/strong> inputs of shape <code>(sequence_length, num_features)<\/code>. It seems you have sequences of length 294 with <strong>one<\/strong> feature; therefore, each input sample needs to have a shape of <code>(294,1)<\/code> (and not <code>(294,)<\/code>). To fix it, you can use <code>np.expand_dims<\/code> to add a third dimension of size 1 to your input data:<\/p>\n\n<pre><code>x_train = np.expand_dims(x_train, axis=-1)\nx_dev = np.expand_dims(x_dev, axis=-1)\n\nmain_input = Input(shape=(294, 1)) # fix the input shape here as well \n<\/code><\/pre>\n"}]}
+{"QuestionId":57715248,"AnswerCount":2,"Tags":"<python><tensorflow><tensorflow-serving>","CreationDate":"2019-08-29T17:30:15.723","AcceptedAnswerId":57930130.0,"OwnerUserId":6249001.0,"Title":"TensorFlow: Help creating a serving input function","Body":"<p>I'm new to TensorFlow Serving. I trained a wide and deep model using an estimator. Now I want to serve my model. I create my serving input receiver function and save the model. When I try to predict using the saved model I always receive <code>InternalError: Unable to get element as bytes<\/code>. I don't really understand what should be inside the serving function or what format type should I send. Could someone first explain me the concept of serving function and also how to properly create the function.<\/p>\n\n<p>Reproducible example: <a href=\"https:\/\/github.com\/dangz90\/wide_and_deep_debugging\/blob\/master\/wide%20and%20deep%20debug.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/dangz90\/wide_and_deep_debugging\/blob\/master\/wide%20and%20deep%20debug.ipynb<\/a><\/p>\n\n<p>Serving function:<\/p>\n\n<pre><code>def serving_input_receiver_fn():\n  serialized_tf_example = tf.placeholder(dtype=tf.string, \n                                         shape=[],\n                                         name='input_tensor')\n  receiver_tensors = {'inputs': serialized_tf_example}\n\n  parsed_features = tf.parse_single_example(\n            serialized_tf_example,\n            # Defaults are not specified since both keys are required.\n            features={\n                'var1': tf.FixedLenFeature(shape=[1], dtype=tf.string),\n                'var2': tf.FixedLenFeature(shape=[1], dtype=tf.string),\n                'var3': tf.FixedLenFeature(shape=[1], dtype=tf.string),\n                'var4': tf.VarLenFeature(dtype=tf.string),\n            })\n\n  return tf.estimator.export.ServingInputReceiver(parsed_features, receiver_tensors)\n\nestimator_predictor = tf.contrib.predictor.from_estimator(m, serving_input_receiver_fn)\nestimator_predictor({ 'inputs': examples.SerializeToString() })\n<\/code><\/pre>\n\n<p>I've tried sending a pandas, example, etc. But really have no idea what format should the data be. For my training the data fed was saved as tfrecord then loaded as dataset. <em>Note: I if use the model directly <code>m.predict()<\/code> I'm able to get the prediction correctly<\/em><\/p>\n\n<p>Full error:<\/p>\n\n<pre><code>--------------------------------------------------------------------------- InternalError                             Traceback (most recent call last) \/databricks\/python\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py in _do_call(self, fn, *args)    1333     try:\n-&gt; 1334       return fn(*args)    1335     except errors.OpError as e:\n\n\/databricks\/python\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py in _run_fn(feed_dict, fetch_list, target_list, options, run_metadata)  1318       return self._call_tf_sessionrun(\n-&gt; 1319           options, feed_dict, fetch_list, target_list, run_metadata)    1320 \n\n\/databricks\/python\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py in _call_tf_sessionrun(self, options, feed_dict, fetch_list, target_list, run_metadata)    1406         self._session, options, feed_dict, fetch_list, target_list,\n-&gt; 1407         run_metadata)    1408 \n\nInternalError: Unable to get element as bytes.\n\nDuring handling of the above exception, another exception occurred:\n\nInternalError                             Traceback (most recent call last) &lt;command-364073753108128&gt; in &lt;module&gt;()\n      3 \n      4 estimator_predictor = tf.contrib.predictor.from_estimator(m, serving_input_receiver_fn)\n----&gt; 5 estimator_predictor({ 'inputs': examples_ })\n\n\/databricks\/python\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/predictor\/predictor.py in __call__(self, input_dict)\n     75       if value is not None:\n     76         feed_dict[self.feed_tensors[key]] = value\n---&gt; 77     return self._session.run(fetches=self.fetch_tensors, feed_dict=feed_dict)\n\n\/databricks\/python\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py in run(self, fetches, feed_dict, options, run_metadata)\n    674                           feed_dict=feed_dict,\n    675                           options=options,\n--&gt; 676                           run_metadata=run_metadata)\n    677 \n    678   def run_step_fn(self, step_fn):\n\n\/databricks\/python\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py in run(self, fetches, feed_dict, options, run_metadata)    1169        feed_dict=feed_dict,    1170                               options=options,\n-&gt; 1171                               run_metadata=run_metadata)    1172       except _PREEMPTION_ERRORS as e:    1173         logging.info('An error was raised. This may be due to a preemption in '\n\n\/databricks\/python\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py in run(self, *args, **kwargs)    1268         raise six.reraise(*original_exc_info)    1269       else:\n-&gt; 1270         raise six.reraise(*original_exc_info)    1271     1272 \n\n\/databricks\/python\/lib\/python3.6\/site-packages\/six.py in reraise(tp, value, tb)\n    691             if value.__traceback__ is not tb:\n    692                 raise value.with_traceback(tb)\n--&gt; 693             raise value\n    694         finally:\n    695             value = None\n\n\/databricks\/python\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py in run(self, *args, **kwargs)    1253   def run(self, *args,\n**kwargs):    1254     try:\n-&gt; 1255       return self._sess.run(*args, **kwargs)    1256     except _PREEMPTION_ERRORS:    1257       raise\n\n\/databricks\/python\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py in run(self, fetches, feed_dict, options, run_metadata)    1325        feed_dict=feed_dict,    1326                                   options=options,\n-&gt; 1327                                   run_metadata=run_metadata)    1328     1329     for hook in self._hooks:\n\n\/databricks\/python\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/monitored_session.py in run(self, *args, **kwargs)    1089     1090   def run(self, *args,\n**kwargs):\n-&gt; 1091     return self._sess.run(*args, **kwargs)    1092     1093   def run_step_fn(self, step_fn, raw_session, run_with_hooks):\n\n\/databricks\/python\/lib\/python3.6\/site-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\/databricks\/python\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py in _run(self, handle, fetches, feed_dict, options, run_metadata)    1150     if final_fetches or final_targets or (handle and feed_dict_tensor):    1151       results = self._do_run(handle, final_targets, final_fetches,\n-&gt; 1152                              feed_dict_tensor, options, run_metadata)    1153     else:    1154       results = []\n\n\/databricks\/python\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py in _do_run(self, handle, target_list, fetch_list, feed_dict, options, run_metadata)    1326     if handle is None:    1327       return self._do_call(_run_fn, feeds, fetches, targets, options,\n-&gt; 1328                            run_metadata)    1329     else:    1330       return self._do_call(_prun_fn, handle, feeds, fetches)\n\n\/databricks\/python\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py in _do_call(self, fn, *args)    1346           pass    1347       message = error_interpolation.interpolate(message, self._graph)\n-&gt; 1348       raise type(e)(node_def, op, message)    1349     1350   def _extend_graph(self):\n\nInternalError: Unable to get element as bytes.\n<\/code><\/pre>\n\n<p>Here's how examples look like before being serialize.<\/p>\n\n<pre><code>features {\n  feature {\n    key: \"var4\"\n    value {\n      bytes_list {\n        value: \"43\"\n        value: \"65\"\n        value: \"89\"\n        value: \"02\"\n      }\n    }\n  }\n  feature {\n    key: \"var3\"\n    value {\n      bytes_list {\n        value: \"0123194\"\n      }\n    }\n  }\n  feature {\n    key: \"var2\"\n    value {\n      bytes_list {\n        value: \"1243\"\n      }\n    }\n  }\n  feature {\n    key: \"var1\"\n    value {\n      bytes_list {\n        value: \"54\"\n      }\n    }\n  }\n}\n<\/code><\/pre>\n\n<p>To get example I run the following script:<\/p>\n\n<pre><code>def serialise_input(data):\n  dict_feature = {}\n\n  for e in data.items():\n    if e[0] == \"var4\":\n      dict_feature.update({e[0]: Feature(bytes_list=BytesList(value=[m.encode('utf-8') for m in e[1]]))})\n    else:\n      dict_feature.update({e[0]: Feature(bytes_list=BytesList( value=[e[1].encode()] ))})\n\n  example = Example(features=Features(feature=dict_feature))        \n\n  return example.SerializeToString()\n\n# Serialize Input\nraw_data = test.toPandas().dropna().iloc[0,:-1]\nexamples_ = serialise_input(raw_data)\n<\/code><\/pre>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"57930130","CreationDate":"2019-09-13T20:43:58.347","ParentId":null,"OwnerUserId":"6249001","Title":null,"Body":"<p>The problem was with how I was constructing the input and the input serving function. Here is the solution:<\/p>\n\n<pre><code>feat_name_type = {'var1':str, 'var2':str, 'var3':str, 'var4':list}\n\ndef input_fn(df):\n    examples = [None] * len(df)\n    for i, sample in df.iterrows():\n        ex = tf.train.Example()\n        for feat_name, feat_type in feat_name_type.items():\n            feat = ex.features.feature[feat_name]\n            if feat_type == int:\n                feat.int64_list.value.extend([sample[feat_name]])\n            elif feat_type == float:\n                feat.float_list.value.extend([sample[feat_name]])\n            elif feat_type == str:\n                feat.bytes_list.value.extend([sample[feat_name].encode()])                \n            elif feat_type == list:\n                feat.bytes_list.value.extend([s.encode() for s in sample[feat_name]])\n        examples[i] = ex.SerializeToString()\n    return {\"inputs\": examples}\n\n'''Service input function'''\ntf_feat_cols = deep_columns + wide_columns #The feature columns created for feeding into the model\n\nserve_rcvr_fn = tf.estimator.export.build_parsing_serving_input_receiver_fn(\n    tf.feature_column.make_parse_example_spec(tf_feat_cols)\n)\nrcvr_fn_map = {\n    tf.estimator.ModeKeys.PREDICT: serve_rcvr_fn,\n}\n<\/code><\/pre>\n"},{"AnswerId":"57776691","CreationDate":"2019-09-03T18:01:29.000","ParentId":null,"OwnerUserId":"3325672","Title":null,"Body":"<p>It looks like your <code>serialized_tf_example<\/code> placeholder has <code>shape=[]<\/code> which is a single example. You should pass a single <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Example\" rel=\"nofollow noreferrer\">tf.train.Example<\/a>, serialized as a string:<\/p>\n\n<pre><code># I assume here examples_ is a list of tf.train.Examples\nexample = examples_[0].SerializeToString()\nestimator_predictor({ 'inputs': example })\n<\/code><\/pre>\n\n<p>If you want to feed a batch of examples instead of a single serialized example, you need to use <code>shape=[None]<\/code> and <code>tf.io.parse_example<\/code> instead.<\/p>\n\n<p>Additionally, the example that you feed should have features defined on it that you reference in your features dictionary (e.g. <code>var1<\/code>) so that they can be properly parsed.<\/p>\n\n<hr>\n\n<p>A serving function specifies how your model should receive its' input at prediction time - when you export your trained TensorFlow graph \/ model as a SavedModel. Instead of having an active TF session like in training where you normally use placeholder or tf.data.Dataset and requires stateful python \/ TF runtime execution you want your SavedModel to be able to be serialized and written to disk so that you can deploy with TF Serving or run it on a mobile device, etc.<\/p>\n\n<p>So you export this SavedModel - the serving input is the piece inside that defines how to parse the requests it gets sent and then connect that into your model The way you probably want to send them is as tf.train.Examples protocol buffers (serialized as strings so that can be sent in an RPC). And then you parse them into a feature dict so your estimator can understand them in the same way it understood your training data.<\/p>\n"}]}
+{"QuestionId":57716116,"AnswerCount":0,"Tags":"<python><tensorflow><tf-slim>","CreationDate":"2019-08-29T18:39:14.200","AcceptedAnswerId":null,"OwnerUserId":5314757.0,"Title":"What does shuffling do in dataset_data_provider from tensorflow-slim","Body":"<p>I have a question a shuffling in Tensorflow-slim data provider code.<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/slim\/python\/slim\/data\/dataset_data_provider.py#L72\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/slim\/python\/slim\/data\/dataset_data_provider.py#L72<\/a><\/p>\n\n<p>Does anyone know what does shuffling do? It is indicated it shuffles data from data sources. Let's say, I have 10 TF records, so it means it shuffles tf records and use images in TF, based on batch size and train the model? Or it shuffles TF records AND shuffles images within TF records and take batch size number of images and starts training?<\/p>\n\n<p>Thanks<\/p>\n","answers":[]}
+{"QuestionId":57716256,"AnswerCount":1,"Tags":"<function><tensorflow>","CreationDate":"2019-08-29T18:51:17.600","AcceptedAnswerId":57716655.0,"OwnerUserId":10561436.0,"Title":"How to properly call functions in tensorflow","Body":"<p>Given the following code, where <code>begin<\/code> is an input to the function <code>sample_text()<\/code> at the same time  it is its output. I want to call the function <code>sample_text()<\/code> 200 times where its output(<code>begin<\/code>) is the input in the next iteration. However my initial <code>begin<\/code> is <code>begin=next_element[0]<\/code> is from the <code>tf.data.Dataset<\/code> while subsequent begin is from the function <code>sample_text()<\/code>. My question  is: if I do <code>sess.run(begin)<\/code>, which <code>begin<\/code> is executed ? this <code>begin=next_element[0]<\/code> line  or <code>begin=self.sample_text(hprev,begin,p_s)<\/code>,please help <\/p>\n\n<pre><code>begin=next_element[0]\nbegin=self.sample_text(hprev,begin,p_s)\nwith tf.Session() as sess:   \n           for j in range(200):\n                  sample=sess.run(begin)                                   \n<\/code><\/pre>\n","answers":[{"AnswerId":"57716655","CreationDate":"2019-08-29T19:25:02.890","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>If your <code>sample_text<\/code> function builds on top of your already existing <code>begin<\/code> node in your TensorFlow graph, then when you assign the Python variable named <code>begin<\/code> to the output of this <code>sample_text<\/code> function, there is nothing to worry. When you run <code>session.run(begin)<\/code>, both the nodes in your graph are run.<\/p>\n\n<p>If however, you don't build on top of the first <code>begin<\/code> variable in <code>sample_text<\/code> function, the initially assigned node will become unreachable for your subsequent code. And when you run <code>session.run(begin)<\/code>, only the output node of your <code>sample_text<\/code> will be run.<\/p>\n"}]}
+{"QuestionId":57716276,"AnswerCount":0,"Tags":"<machine-learning><keras><recurrent-neural-network>","CreationDate":"2019-08-29T18:52:35.227","AcceptedAnswerId":null,"OwnerUserId":7519842.0,"Title":"Inputting data sequentially to Keras RNN","Body":"<p>So, say I have a very simple Keras RNN model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential()\nmodel.add(GRU(50, input_shape = (300,1), return_sequences = True))\nmodel.add(Activation('sigmoid'))\n<\/code><\/pre>\n\n<p>In production, I want to be able to input data continuously. In other words, I want to be able to use <code>a&lt;t-1&gt;<\/code> and <code>c&lt;t-1&gt;<\/code> to be able to input a new <code>x&lt;t&gt;<\/code> without having to recompute all of the previous <code>a&lt;t-2&gt;<\/code>, <code>a&lt;t-3&gt;<\/code>... etc. values. I know I would be able to do this with some more manual code work, but is it possible natively in Keras?<\/p>\n","answers":[]}
+{"QuestionId":57716363,"AnswerCount":1,"Tags":"<python><keras><tensorflow2.0><tf.keras>","CreationDate":"2019-08-29T18:59:21.340","AcceptedAnswerId":57720708.0,"OwnerUserId":8242607.0,"Title":"Explicit broadcasting of variable batch-size tensor","Body":"<p>I'm trying to implement a custom Keras <code>Layer<\/code> in Tensorflow 2.0RC and need to concatenate a <code>[None, Q]<\/code> shaped tensor onto a <code>[None, H, W, D]<\/code> shaped tensor to produce a <code>[None, H, W, D + Q]<\/code> shaped tensor. It is assumed that the two input tensors have the same batch size even though it is not known beforehand. Also, none of H, W, D, and Q are known at write-time but are evaluated in the layer's <code>build<\/code> method when the layer is first called. The issue that I'm experiencing is when broadcasting the <code>[None, Q]<\/code> shaped tensor up to a <code>[None, H, W, Q]<\/code> shaped tensor in order to concatenate.<\/p>\n\n<p>Here is an example of trying to create a Keras <code>Model<\/code> using the Functional API that performs variable-batch broadcasting from shape <code>[None, 3]<\/code> to shape <code>[None, 5, 5, 3]<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport tensorflow.keras.layers as kl\nimport numpy as np\n\nx = tf.keras.Input([3])  # Shape [None, 3]\ny = kl.Reshape([1, 1, 3])(x)  # Need to add empty dims before broadcasting\ny = tf.broadcast_to(y, [-1, 5, 5, 3])  # Broadcast to shape [None, 5, 5, 3]\n\nmodel = tf.keras.Model(inputs=x, outputs=y)\n\nprint(model(np.random.random(size=(8, 3))).shape)\n<\/code><\/pre>\n\n<p>Tensorflow produces the error:<\/p>\n\n<pre><code>InvalidArgumentError:  Dimension -1 must be &gt;= 0\n<\/code><\/pre>\n\n<p>And then when I change <code>-1<\/code> to <code>None<\/code> it gives me:<\/p>\n\n<pre><code>TypeError: Failed to convert object of type &lt;class 'list'&gt; to Tensor. Contents: [None, 5, 5, 3]. Consider casting elements to a supported type.\n<\/code><\/pre>\n\n<p>How can I perform the specified broadcasting?<\/p>\n","answers":[{"AnswerId":"57720708","CreationDate":"2019-08-30T04:55:34.533","ParentId":null,"OwnerUserId":"8242607","Title":null,"Body":"<p>You need to use the dynamic shape of <code>y<\/code> to determine the batch size. The dynamic shape of a tensor <code>y<\/code> is given by <code>tf.shape(y)<\/code> and is a tensor op representing the shape of <code>y<\/code> evaluated at runtime. The modified example demonstrates this by selecting between the old shape, <code>[None, 1, 1, 3]<\/code>, and the new shape using <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/where\" rel=\"nofollow noreferrer\"><code>tf.where<\/code><\/a>.<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorflow.keras.layers as kl\nimport numpy as np\n\nx = tf.keras.Input([3])  # Shape [None, 3]\ny = kl.Reshape([1, 1, 3])(x)  # Need to add empty dims before broadcasting\n# Retain the batch and depth dimensions, but broadcast along H and W\nbroadcast_shape = tf.where([True, False, False, True],\n                           tf.shape(y), [0, 5, 5, 0])\ny = tf.broadcast_to(y, broadcast_shape)  # Broadcast to shape [None, 5, 5, 3]\n\nmodel = tf.keras.Model(inputs=x, outputs=y)\n\nprint(model(np.random.random(size=(8, 3))).shape)\n# prints: \"(8, 5, 5, 3)\"\n<\/code><\/pre>\n\n<p>References:<\/p>\n\n<p><a href=\"https:\/\/blog.metaflow.fr\/shapes-and-dynamic-dimensions-in-tensorflow-7b1fe79be363\" rel=\"nofollow noreferrer\">\"TensorFlow: Shapes and dynamic dimensions\"<\/a><\/p>\n"}]}
+{"QuestionId":57716410,"AnswerCount":0,"Tags":"<python><tensorflow><dll><load>","CreationDate":"2019-08-29T19:02:54.420","AcceptedAnswerId":null,"OwnerUserId":11996190.0,"Title":"Python**** ImportError: DLL load failed I cant start tensorflow","Body":"<p>When I try start these codes I take some errors.<\/p>\n\n<p>I tried almost everything.<\/p>\n\n<p>import tensorflow<\/p>\n\n<p>I didnt expect any errors but when it start this errors appears.  <\/p>\n\n<p>File \"C:\/Users\/PC\/.PyCharmCE2019.1\/config\/scratches\/scratch_2.py\", line 1, in \n    import  tensorflow\n  File \"C:\\Users\\PC\\PycharmProjects\\untitled\\ven\\lib\\site-packages\\tensorflow__init__.py\", line 28, in \n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"C:\\Users\\PC\\PycharmProjects\\untitled\\ven\\lib\\site-packages\\tensorflow\\python__init__.py\", line 52, in \n    from tensorflow.core.framework.graph_pb2 import *\n  File \"C:\\Users\\PC\\PycharmProjects\\untitled\\ven\\lib\\site-packages\\tensorflow\\core\\framework\\graph_pb2.py\", line 7, in \n    from google.protobuf import descriptor as _descriptor\n  File \"C:\\Users\\PC\\PycharmProjects\\untitled\\ven\\lib\\site-packages\\google\\protobuf\\descriptor.py\", line 47, in \n    from google.protobuf.pyext import _message\nImportError: DLL load failed: Belirtilen yordam bulunamad\u0131.<\/p>\n","answers":[]}
+{"QuestionId":57716511,"AnswerCount":0,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2019-08-29T19:10:47.617","AcceptedAnswerId":null,"OwnerUserId":190894.0,"Title":"How to monitor gradient flow in tensorflow 2.0?","Body":"<p>I would like to monitor the gradient flow as my models are being trained in Tensorflow 2.0. Ideally, I would like to store the gradients in an array as the model is being trained, and then just view them via matplotlib after training is complete. How would I do this?<\/p>\n\n<p>Here's a toy model that I would like to run:<\/p>\n\n<pre><code>tf.keras.backend.clear_session()\n\ninput_sequence = tf.keras.layers.Input(shape=[number_timesteps, number_features])\nx = tf.keras.layers.Conv1D(2**5, 3, padding='same', activation='relu')(input_sequence)\nx = tf.keras.layers.Flatten()(x)\nx = tf.keras.layers.Dropout(0.5)(x)\noutput_sequence = tf.keras.layers.Dense(number_classes, activation=\"softmax\")(x)\n\nmodel = tf.keras.Model(input_sequence, output_sequence)\nmodel.compile(loss=tf.keras.losses.categorical_crossentropy, metrics=[\"accuracy\"], optimizer=tf.keras.optimizers.Adam())\nmodel.summary()\n\nmodel.fit(...)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57716582,"AnswerCount":2,"Tags":"<tensorflow><tensorboard>","CreationDate":"2019-08-29T19:17:44.820","AcceptedAnswerId":null,"OwnerUserId":8894647.0,"Title":"i wanted to ask about my training model which i have trained so far i,e is it tending toward accuracy?","Body":"<h1>summary<\/h1>\n\n<p>I am new to deep learning and want to know about epoch accuracy regarding tensorboard. I have trained deep learning model with 400 images each about 960x1280 pixels and I have visualize it on tensorboard. Can anyone help me by telling me that what is meant by epoch accuracy and in my case is it good enough in the picture or what else should I do to match the training and validation accuracy? The picture is given below:\n<a href=\"https:\/\/i.stack.imgur.com\/2QHaX.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n","answers":[{"AnswerId":"57720611","CreationDate":"2019-08-30T04:44:20.023","ParentId":null,"OwnerUserId":"4072504","Title":null,"Body":"<p>While <a href=\"https:\/\/stackoverflow.com\/a\/57720035\/4072504\">Uzzal Podder's answer<\/a> is correct, I think it is likely that in your code the <strong>epoch _acc<\/strong> is representing the accuracy of the model after completing a full epoch of training.<\/p>\n\n<p>An <strong>epoch<\/strong> usually represents a full iteration over the training dataset. Therefore, for your dataset, it would mean a full training cycle over the 400 images.<\/p>\n\n<p>If this assumption is correct, the <strong>epoch_acc<\/strong> would reflect your training accuracy. To get a testing accuracy, you would want to use the trained model to evaluate a <a href=\"https:\/\/towardsdatascience.com\/train-test-split-and-cross-validation-in-python-80b61beca4b6\" rel=\"nofollow noreferrer\">holdout set of data<\/a>.<\/p>\n\n<p>The methodologies to close the gap between training and testing accuracies for CNNs is something that has been answered a number of times on this site and I would recommend looking into those answers first.<\/p>\n"},{"AnswerId":"57720035","CreationDate":"2019-08-30T03:12:22.970","ParentId":null,"OwnerUserId":"4257991","Title":null,"Body":"<p>The <strong>title<\/strong> <code>epoch_acc<\/code> depends on your code. <\/p>\n\n<p>If you use any predefined model(for example, <a href=\"https:\/\/keras.io\/applications\/\" rel=\"nofollow noreferrer\">keras ResNet50<\/a> or ready-made model from github) then the <strong>epoch_acc<\/strong> title will be mentioned somewhere in the code.<\/p>\n\n<p>In your description no source code is provided. Also, it is not clear which tensorflow api(<em>tensorflow core<\/em> or <em>keras backed tensorflow<\/em>) you are using. It is little difficult to tell the meaning of <code>epoch_acc<\/code>. <\/p>\n\n<p>Here <code>epoch_acc<\/code> might be <strong><em>total accuracy<\/strong> with respect to the incremental epoch number<\/em>. I mean, <em>Y<\/em> axis is for <strong>accuracy<\/strong> and <em>X<\/em> axis for simple <strong>epoch count<\/strong> <em>(how many times you are repeating your entire train dataset.)<\/em> <\/p>\n\n<p>However, please let me give an example what I am telling, please look at this code.<\/p>\n\n<p>The accuracy metrics <em>title<\/em> will be <code>softmax_out_acc<\/code> and <code>ring_loss_acc<\/code> \nif I use - <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.compile(loss = {'softmax_out' : 'categorical_crossentropy', \n                     'ring_loss' : identity_loss},\n               metrics = ['accuracy'], \n               ... )\n<\/code><\/pre>\n\n<pre><code>32\/32 [==============================] - 116s 4s\/step - loss: 3.1720 - softmax_out_loss: 0.8834 - ring_loss_loss: 22.8862 - softmax_out_acc: 0.5391 - ring_loss_acc: 0.0000e+00 - val_loss: 3.1074 - val_softmax_out_loss: 0.9071 - val_ring_loss_loss: 22.0037 - val_softmax_out_acc: 0.4590 - val_ring_loss_acc: 0.0000e+00\n<\/code><\/pre>\n\n<p>Or, if you use <code>cross_entropy<\/code> <em>string<\/em> in <code>tf.summary.scalar<\/code> of tensorboard writer <a href=\"https:\/\/www.tensorflow.org\/guide\/summaries_and_tensorboard\" rel=\"nofollow noreferrer\">like this:<\/a>  <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with tf.name_scope('cross_entropy'):\n  with tf.name_scope('total'):\n    cross_entropy = tf.losses.sparse_softmax_cross_entropy(labels=y_, logits=y)\ntf.summary.scalar('cross_entropy', cross_entropy)\n\n<\/code><\/pre>\n\nThe output will look like this\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/XjFo1.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/XjFo1.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><strong>so, there is nothing well defined <code>epoch_acc<\/code> in deep learning, or <code>epoch_acc<\/code> is not a standards metrics, it entirely depends on your code.<\/strong> <\/p>\n"}]}
+{"QuestionId":57716751,"AnswerCount":0,"Tags":"<python-2.7><dictionary><tensorflow><tensor>","CreationDate":"2019-08-29T19:31:50.360","AcceptedAnswerId":null,"OwnerUserId":11995873.0,"Title":"I'm using Tensors as keys of a dictionary but I cannot access them","Body":"<p>I am new to Tensorflow and I am trying to create a dictionary using integer indices in the form of Tensors (necessary for later in my code) as keys. I am able to iterate and print the dictionary keys and values after creating it but I cannot seem to access an element of the dictionary individually. For example:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport tensorflow as tf\n\nA=np.random.randn(320,320)\nA=tf.convert_to_tensor(A)\n\ndict_nid_neighbors = {}\nfor nid in range(A.get_shape()[0]):\n\n  zero = tf.constant(0, dtype=tf.float64)\n  where = tf.not_equal(A[nid], zero)\n  neighbors=tf.where(where)\n  neighbors = tf.reshape(tf.where(where),[-1])\n  nid=tf.constant(nid)\n  dict_nid_neighbors[nid] = neighbors\n<\/code><\/pre>\n\n<p>When I try this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>sess=tf.Session()\nfor x,y in dict_nid_neighbors.items():\n  print sess.run(x),sess.run(y) \n<\/code><\/pre>\n\n<p>it prints the expected result, but when I try to access the first element:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>print 'Tensor key 0: ',dict_nid_neighbors[tf.constant(0)]\n<\/code><\/pre>\n\n<p>I get the error: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>KeyError: &lt;tf.Tensor 'Const_18122:0' shape=() dtype=int32&gt;\n<\/code><\/pre>\n\n<p>Are you not allowed to do this, am I missing something obvious? Thank you.<\/p>\n","answers":[]}
+{"QuestionId":57716770,"AnswerCount":1,"Tags":"<python><tensorflow><object-detection>","CreationDate":"2019-08-29T19:33:40.467","AcceptedAnswerId":null,"OwnerUserId":11346798.0,"Title":"Tensorflow Can not train Model","Body":"<p>I have followed this Tutorial to train my Custom Object but at the Training the Model step by executing this command :<\/p>\n\n<p>python train.py --logtostderr --train_dir=training\/ --pipeline_config_path=training\/ssd_inception_v2_coco.config<\/p>\n\n<p>I get this result:<\/p>\n\n<pre><code>(tensorflow_cpu) C:\\Users\\sid\\Documents\\Tensorflow\\models\\research\\object_detection\\legacy&gt;python train.py --logtostderr --train_dir=training\/ --pipeline_config_path=training\/ssd_inception_v2_coco.config\n\nTraceback (most recent call last):\n  File \"train.py\", line 51, in &lt;module&gt;\n    from object_detection.builders import model_builder\n  File \"C:\\Users\\sid\\Anaconda3\\envs\\tensorflow_cpu\\lib\\site-packages\\object_detection-0.1-py3.6.egg\\object_detection\\builders\\model_builder.py\", line 35, in &lt;module&gt;\n    from object_detection.models import faster_rcnn_inception_resnet_v2_feature_extractor as frcnn_inc_res\n  File \"C:\\Users\\sid\\Anaconda3\\envs\\tensorflow_cpu\\lib\\site-packages\\object_detection-0.1-py3.6.egg\\object_detection\\models\\faster_rcnn_inception_resnet_v2_feature_extractor.py\", line 29, in &lt;module&gt;\n    from nets import inception_resnet_v2\nModuleNotFoundError: No module named 'nets'\n<\/code><\/pre>\n","answers":[{"AnswerId":"57716820","CreationDate":"2019-08-29T19:38:45.703","ParentId":null,"OwnerUserId":"4966536","Title":null,"Body":"<p>You probably need to install the package <strong>\u201cnets\u201d<\/strong>.<\/p>\n\n<p>In Linux this means typing this into the terminal:<\/p>\n\n<pre><code>pip install nets\n<\/code><\/pre>\n\n<p>Of course, then you need to add an import line into the python script:<\/p>\n\n<pre><code>import nets\n<\/code><\/pre>\n"}]}
+{"QuestionId":57716809,"AnswerCount":0,"Tags":"<python><tensorflow><fetch>","CreationDate":"2019-08-29T19:37:33.017","AcceptedAnswerId":null,"OwnerUserId":11996287.0,"Title":"How can I understand and interpret the working of tensorflow fetch (for the given code)?","Body":"<p>I have been trying to learn how to use Tensorflow and I don't know how fetch in Tensorflow works. Here I have just created a simple linear regression and don't know how it is fetching the result <\/p>\n\n<pre><code>m = tf.Variable(0.39)\nb = tf.Variable(0.2)\nerror = 0\nfor x,y in zip(x_data,y_label):\ny_hat = m*x + b  #Our predicted value\nerror += (y-y_hat)**2\noptimizer = tf.train.GradientDescentOptimizer(learning_rate=0.001)\ntrain = optimizer.minimize(error)\ninit = tf.global_variables_initializer()\n<\/code><\/pre>\n\n<p>with tf.Session() as sess:<\/p>\n\n<pre><code>sess.run(init)\n\n\nepochs = 100\n\nfor i in range(epochs):\n\n    sess.run(train)\n\n\n# Fetch Back Results\nfinal_slope , final_intercept = sess.run([m,b])\n<\/code><\/pre>\n\n<p>i dont how the the variable m and b got updated using fetch<\/p>\n","answers":[]}
+{"QuestionId":57717004,"AnswerCount":1,"Tags":"<python><tensorflow><tf.keras><tfrecord>","CreationDate":"2019-08-29T19:57:26.117","AcceptedAnswerId":null,"OwnerUserId":9988487.0,"Title":"Tensorflow: Modern way to load large data","Body":"<p>I want to train a convolutional neural network (using tf.keras from Tensorflow version 1.13) using numpy arrays as input data. The training data (which I currently store in a single >30GB '.npz' file) does not fit in RAM all at once. <strong>What is the best way to save and load large data-sets into a neural network for training?<\/strong> Since I didn't manage to find a good answer to this (surely ubiquitous?) problem, I'm hoping to hear one here. Thank you very much in advance for any help!<\/p>\n\n<h3>Sources<\/h3>\n\n<p>Similar questions seem to have been asked many times (e.g. <a href=\"https:\/\/stackoverflow.com\/questions\/49169016\/training-classifier-from-tfrecords-in-tensorflow\">training-classifier-from-tfrecords-in-tensorflow<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/40467990\/tensorflow-synchronize-readings-from-tfrecord\">tensorflow-synchronize-readings-from-tfrecord<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/51357223\/how-to-load-data-parallelly-in-tensorflow\">how-to-load-data-parallelly-in-tensorflow<\/a>) but are several years old and usually contain no conclusive answer.<\/p>\n\n<p>My current understanding is that using TFRecord files is a good way to approach this problem. The most promising tutorial I found so far explaining how to use TFRecord files with keras is <a href=\"https:\/\/medium.com\/@moritzkrger\/speeding-up-keras-with-tfrecord-datasets-5464f9836c36\" rel=\"nofollow noreferrer\">medium.com<\/a>. Other helpful sources were <a href=\"http:\/\/machinelearninguru.com\/deep_learning\/data_preparation\/tfrecord\/tfrecord.html\" rel=\"nofollow noreferrer\">machinelearninguru.com<\/a> and <a href=\"https:\/\/medium.com\/mostly-ai\/tensorflow-records-what-they-are-and-how-to-use-them-c46bc4bbb564\" rel=\"nofollow noreferrer\">medium.com_source2<\/a> and sources therin.<\/p>\n\n<p>The official tensorflow documentation and tutorials (on <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset\" rel=\"nofollow noreferrer\">tf.data.Dataset<\/a>, <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets\" rel=\"nofollow noreferrer\">Importing Data<\/a>, <a href=\"https:\/\/www.tensorflow.org\/tutorials\/load_data\/tf_records\" rel=\"nofollow noreferrer\">tf_records<\/a> etc.) did not help me. In particular, several of the examples given there didn't work for me even without modifications.<\/p>\n\n<h3>My Attempt at using TFRecord files<\/h3>\n\n<p>I'm assuming TFRecords are a good way to solve my problem but I'm having a hard time using them. Here is an example I made based on the tutorial <a href=\"https:\/\/medium.com\/@moritzkrger\/speeding-up-keras-with-tfrecord-datasets-5464f9836c36\" rel=\"nofollow noreferrer\">medium.com<\/a>. I stripped down the code as much as I could.<\/p>\n\n<pre><code># python 3.6, tensorflow 1.13.\n# Adapted from https:\/\/medium.com\/@moritzkrger\/speeding-up-keras-with-tfrecord-datasets-5464f9836c36\nimport tensorflow as tf\nimport numpy as np\nfrom tensorflow.python import keras as keras\n\n\n# Helper functions (see also https:\/\/www.tensorflow.org\/tutorials\/load_data\/tf_records)\ndef _int64_feature(value):\n    return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))\n\n\ndef _bytes_feature(value):\n    return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))\n\n\ndef writeTFRecords():\n    number_of_samples = 100  # create some random data to play with\n    images, labels = (np.random.sample((number_of_samples, 256, 256, 1)), np.random.randint(0, 30, number_of_samples))\n\n    writer = tf.python_io.TFRecordWriter(\"bla.tfrecord\")\n\n    for index in range(images.shape[0]):\n        image = images[index]\n        label = labels[index]\n\n        feature = {'image':  _bytes_feature(tf.compat.as_bytes(image.tostring())),\n                   'label':  _int64_feature(int(label))}\n\n        example = tf.train.Example(features=tf.train.Features(feature=feature))\n        writer.write(example.SerializeToString())\n    writer.close()\n\n\ndef loadTFRecord(data_path):\n    with tf.Session() as sess:\n        feature = {'train\/image': tf.FixedLenFeature([], tf.string),\n                   'train\/label': tf.FixedLenFeature([], tf.int64)}\n        # Create a list of filenames and pass it to a queue\n        filename_queue = tf.train.string_input_producer([data_path], num_epochs=1)\n        # Define a reader and read the next record\n        reader = tf.TFRecordReader()\n        _, serialized_example = reader.read(filename_queue)\n        # Decode the record read by the reader\n        features = tf.parse_single_example(serialized_example, features=feature)\n        # Convert the image data from string back to the numbers\n        image = tf.decode_raw(features['train\/image'], tf.float32)\n\n        # Cast label data into int32\n        label = tf.cast(features['train\/label'], tf.int32)\n        # Reshape image data into the original shape\n        image = tf.reshape(image, [256, 256, 1])\n\n        return image, label  # I'm not 100% sure that's how this works...\n\n\n# ######### generate a TFRecords file in the working directory containing random data. #################################\nwriteTFRecords()\n# ######## Load the TFRecords file and use it to train a simple example neural network. ################################\nimage, label = loadTFRecord(\"bla.tfrecord\")\n\nmodel_input = keras.layers.Input(tensor=image)\nmodel_output = keras.layers.Flatten(input_shape=(-1, 256, 256, 1))(model_input)\nmodel_output = keras.layers.Dense(16, activation='relu')(model_output)\n\ntrain_model = keras.models.Model(inputs=model_input, outputs=model_output)\ntrain_model.compile(optimizer=keras.optimizers.RMSprop(lr=0.0001),  \n                    loss='mean_squared_error',\n                    target_tensors=[label])\n\nprint(\"\\n \\n start training \\n \\n\") # Execution gets stuck on fitting\ntrain_model.fit(epochs=1, steps_per_epoch=10)  # no output or error messages.\n\n<\/code><\/pre>\n\n<p>The code creates a TFRecord file and starts fitting, then just gets stuck with no output or error messages. I don't know what the problem is or how I could try to fix it.<\/p>\n","answers":[{"AnswerId":"58250849","CreationDate":"2019-10-05T17:47:05.293","ParentId":null,"OwnerUserId":"9988487","Title":null,"Body":"<p>While this is no real answer to the original question (i.e. \"what is the optimal way to train on large datasets\"), I managed to get tfrecords and datasets to work. Of particular help was this <a href=\"https:\/\/www.youtube.com\/watch?v=oxrcZ9uUblI&amp;list=PL9Hr9sNUjfsmEu1ZniY0XpHSzl5uihcXZ&amp;index=25\" rel=\"nofollow noreferrer\">tutorial on YouTube<\/a>. I include a minimal example with working code for anyone struggling with the same problem.<\/p>\n\n<pre><code># Developed using python 3.6, tensorflow 1.14.0.\n# This code writes data (pairs (label, image) where label is int64 and image is np.ndarray) into .tfrecord files and\n# uses them for training a simple neural network. It is meant as a minimal working example of how to use tfrecords. This\n# solution is likely not optimal. If you know how to improve it, please comment on\n# https:\/\/stackoverflow.com\/q\/57717004\/9988487. Refer to links therein for further information.\nimport tensorflow as tf\nimport numpy as np\nfrom tensorflow.python import keras as keras\n\n\n# Helper functions (see also https:\/\/www.tensorflow.org\/tutorials\/load_data\/tf_records)\ndef _int64_feature(value):\n    return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))\n\n\ndef _bytes_feature(value):\n    return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))\n\n\ndef write_tfrecords_file(out_path: str, images: np.ndarray, labels: np.ndarray) -&gt; None:\n    \"\"\"Write all image-label pairs into a single .tfrecord file.\n    :param out_path: File path of the .tfrecord file to generate or overwrite.\n    :param images: array with first dimension being the image index. Every images[i].tostring() is\n        serialized and written into the file as 'image': wrap_bytes(img_bytes)\n    :param labels: 1d array of integers. labels[i] is the label of images[i]. Written as 'label': wrap_int64(label)\"\"\"\n    assert len(images) == len(labels)\n    with tf.io.TFRecordWriter(out_path) as writer:  # could use writer_options parameter to enable compression\n        for i in range(len(labels)):\n            img_bytes = images[i].tostring()  # Convert the image to raw bytes.\n            label = labels[i]\n            data = {'image': _bytes_feature(img_bytes), 'label': _int64_feature(label)}\n            feature = tf.train.Features(feature=data)  # Wrap the data as TensorFlow Features.\n            example = tf.train.Example(features=feature)  # Wrap again as a TensorFlow Example.\n            serialized = example.SerializeToString()  # Serialize the data.\n            writer.write(serialized)  # Write the serialized data to the TFRecords file.\n\n\ndef parse_example(serialized, shape=(256, 256, 1)):\n    features = {'image': tf.io.FixedLenFeature([], tf.string), 'label': tf.io.FixedLenFeature([], tf.int64)}\n    # Parse the serialized data so we get a dict with our data.\n    parsed_example = tf.io.parse_single_example(serialized=serialized, features=features)\n    label = parsed_example['label']\n    image_raw = parsed_example['image']  # Get the image as raw bytes.\n    image = tf.decode_raw(image_raw, tf.float32)  # Decode the raw bytes so it becomes a tensor with type.\n    image = tf.reshape(image, shape=shape)\n    return image, label  # this function will be called once (to add it to tf graph; then parse images individually)\n\n\n# create some arbitrary data to play with: 1000 images sized 256x256 with one colour channel. Use your custom np-arrays\nIMAGE_WIDTH, NUM_OF_IMAGES, NUM_OF_CLASSES, COLOUR_CHANNELS = 256, 10_000, 10, 1\n# using float32 to save memory. Must match type in parse_example(), tf.decode_raw(image_raw, tf.float32)\nfeatures_train = np.random.sample((NUM_OF_IMAGES, IMAGE_WIDTH, IMAGE_WIDTH, COLOUR_CHANNELS)).astype(np.float32)\nlabels_train = np.random.randint(low=0, high=NUM_OF_CLASSES, size=NUM_OF_IMAGES)  # one random label for each image\nfeatures_eval = features_train[:200]  # use the first 200 images as evaluation data for simplicity.\nlabels_eval = labels_train[:200]\nwrite_tfrecords_file(\"train.tfrecord\", features_train, labels_train)  # normal: split the data files of several GB each\nwrite_tfrecords_file(\"eval.tfrecord\", features_eval, labels_eval)  # this may take a while. Consider a progressbar\n# The files are complete. Now define a model and use datasets to feed the data from the .tfrecord files into the model.\nmodel = keras.Sequential([keras.layers.Flatten(input_shape=(256, 256, 1)),\n                          keras.layers.Dense(128, activation='relu'),\n                          keras.layers.Dense(10, activation='softmax')])\nmodel.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\n# Check docs for parameters (compression, buffer size, thread count. Also www.tensorflow.org\/guide\/performance\/datasets\n\ntrain_dataset = tf.data.TFRecordDataset(\"train.tfrecord\")  # specify a list (or dataset) of file names for large data\ntrain_dataset = train_dataset.map(parse_example)  # parse tfrecords. Parameter num_parallel_calls may help performance.\ntrain_dataset = train_dataset.shuffle(buffer_size=1024).batch(64)\n\nvalidation_dataset = tf.data.TFRecordDataset(\"eval.tfrecord\")\nvalidation_dataset = validation_dataset.map(parse_example).batch(64)\n\nmodel.fit(train_dataset, epochs=3)\n# evaluate the results\nresults = model.evaluate(validation_dataset)\nprint('\\n\\nvalidation loss, validation acc:', results)\n\n<\/code><\/pre>\n\n<p>Note that it's tricky to use some_keras_model.fit(..., validation_data=some_dataset) with dataset objects. It may result in\n<code>TypeError: 'DatasetV1Adapter' object does not support indexing<\/code>.\nThis seems to be a bug (see github.com\/tensorflow\/tensorflow\/issues\/28995) and is supposedly fixed as of  tf-nightly version '1.15.0-dev20190808'; The <a href=\"https:\/\/www.tensorflow.org\/guide\/keras\/train_and_evaluate\" rel=\"nofollow noreferrer\">official tutorial<\/a> uses this too, although it doesn't work in most versions. An easy but dirty-ish fix is to use verbose=0 (which only suppresses program output) and plot the validation results using tensorboard. Also see <a href=\"https:\/\/stackoverflow.com\/questions\/50955798\/keras-model-fit-with-tf-dataset-api-validation-data#50979587\">Keras model.fit() with tf.dataset API + validation_data<\/a>.<\/p>\n"}]}
+{"QuestionId":57717142,"AnswerCount":1,"Tags":"<tensorflow><keras><neural-network><tensorboard>","CreationDate":"2019-08-29T20:09:11.683","AcceptedAnswerId":null,"OwnerUserId":140100.0,"Title":"why biases change more frequent then weights?","Body":"<p>I am visualizing biases and weights of fully connected part of my network.\nI see that biases change more frequent then weights and I am trying to understand what does it mean.\nIn both cases I just took the weights and biases and presented them as histogram.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ZtE9r.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ZtE9r.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>here we see:\n bias which is 1x256 parameters.\n weights which is matrix of 256X512 parameters.<\/p>\n\n<p>Can be the reason that weight hardly change in visualization due to high number of numbers that just fall at the same bins ? can be better way to visualize the weight learning? or there some other reason.<\/p>\n\n<p>Another part of the network that connects 16 neurons 1 neuron.\n<a href=\"https:\/\/i.stack.imgur.com\/wEsgK.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/wEsgK.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"57720444","CreationDate":"2019-08-30T04:17:15.577","ParentId":null,"OwnerUserId":"10371987","Title":null,"Body":"<p>I guess this could be caused by gradient vanishing problem. If you look at the back-propagation algorithm you would see the gradient of weights is calculated by chain rule, for example, something like <a href=\"https:\/\/mattmazur.com\/2015\/03\/17\/a-step-by-step-backpropagation-example\/\" rel=\"nofollow noreferrer\">this<\/a>:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/lSiou.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lSiou.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>where  <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/dpvha.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/dpvha.png\" alt=\"enter image description here\"><\/a><br>\n<a href=\"https:\/\/i.stack.imgur.com\/i9SUu.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/i9SUu.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>So, if the output values of last layer is very <strong>close to 0<\/strong>, the gradient of weights would also be closed to 0. As a result, the weights would not be updated with large magnitude. On the other hand, however, bias would not suffer from this problem, since the bias is not affected by previous layers during back-propagation.  <\/p>\n\n<p>You might search online to see the causes of gradient vanishing problem if you aren't familiar with it.<\/p>\n"}]}
+{"QuestionId":57717362,"AnswerCount":1,"Tags":"<pytorch><backpropagation><autograd>","CreationDate":"2019-08-29T20:28:22.557","AcceptedAnswerId":57717852.0,"OwnerUserId":11948285.0,"Title":"Encounter the RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation","Body":"<p>I'm getting the following error when calling .backward():<\/p>\n\n<p>Encounter the RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation<\/p>\n\n<p>Here's the code:<\/p>\n\n<pre><code>for i, j, k in zip(X, Y, Z):\n    A[:, i, j] = A[:, i, j] + k\n<\/code><\/pre>\n\n<p>I've tried .clone(), torch.add(), and so on.<\/p>\n\n<p>Please help!<\/p>\n","answers":[{"AnswerId":"57717852","CreationDate":"2019-08-29T21:20:22.627","ParentId":null,"OwnerUserId":"2790047","Title":null,"Body":"<p>After the comments I'm a bit confused about what you want to accomplish. The code you gave gives me an error using the dimensions you provided in the comments<\/p>\n\n<pre><code>Traceback (most recent call last):\n    A[:, i, j] = A[:, i, j] + k\nRuntimeError: The size of tensor a (32) must match the size of tensor b (200) at non-singleton dimension 0\n<\/code><\/pre>\n\n<p>But here's what I think you want to do, please correct me in the comments if this is wrong...<\/p>\n\n<p>Given tensors <code>X<\/code>, <code>Y<\/code>, and <code>Z<\/code>, each entry of <code>X<\/code>, <code>Y<\/code>, and <code>Z<\/code> correspond to a coordinate (x,y) and a value z. What you want is to add z to <code>A<\/code> at coordinate (x,y). For most cases the batch dimension is kept independent, although its not clear that's the case in the code you posted. For now that's what I'll assume you want to do.<\/p>\n\n<p>For example lets say <code>A<\/code> contains all zeros and has shape 3x4x5 and <code>X<\/code>,<code>Y<\/code> are shape 3x3 and <code>Z<\/code> is shape 3x3x1. For this example let's assume <code>A<\/code> contains all zeros to start, and <code>X<\/code>, <code>Y<\/code>, and <code>Z<\/code> have the following values<\/p>\n\n<pre><code>X = tensor([[1, 2, 3],\n            [1, 2, 3],                                \n            [2, 2, 2]])                               \nY = tensor([[1, 2, 3],                                \n            [1, 2, 3],                                                        \n            [1, 1, 1]])             \nZ = tensor([[[0.1], [0.2], [0.3]],\n            [[0.4], [0.5], [0.6]],\n            [[0.7], [0.8], [0.9]]])\n<\/code><\/pre>\n\n<p>Then we would expect <code>A<\/code> to have the following values after the operation<\/p>\n\n<pre><code>A = tensor([[[0,   0,   0,   0,   0],                                                                                                                                                                          \n             [0,   0.1, 0,   0,   0],                                                                                                                                                                      \n             [0,   0,   0.2, 0,   0],                                                                                                                                                                      \n             [0,   0,   0,   0.3, 0]],                                                                                                                                                                   \n\n            [[0,   0,   0,   0,   0],                                                                                                                                                                      \n             [0,   0.4, 0,   0,   0],                                                                                                                                                                      \n             [0,   0,   0.5, 0,   0],                                                                                                                                                                      \n             [0,   0,   0,   0.6, 0]],                                                                                                                                                                   \n\n            [[0,   0,   0,   0,   0],                                                                                                                                                                    \n             [0,   0,   0,   0,   0],                                                                                                                                                                    \n             [0,   2.4, 0,   0,   0],                                                                                                                                                                    \n             [0,   0,   0,   0,   0]]])\n<\/code><\/pre>\n\n<p>In order to accomplish this we can make use to the <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.Tensor.index_add\" rel=\"nofollow noreferrer\"><code>index_add<\/code><\/a> function which allows us to add to a list of indices. Since this only supports 1-dimensional operations we first need to convert <code>X<\/code>,<code>Y<\/code> to a linear index for flattened tensor <code>A<\/code>. Afterwards we can un-flatten to the original shape.<\/p>\n\n<pre><code>layer_size = A.shape[1] * A.shape[2]                                                                                                                                                                       \nindex_offset = torch.arange(0, A.shape[0] * layer_size, layer_size).unsqueeze(1)                                                                                                                           \nindices = (X * A.shape[2] + Y) + index_offset                                                                                                                                                              \nA = A.view(-1).index_add(0, indices.view(-1), Z.view(-1)).view(A.shape)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57717569,"AnswerCount":0,"Tags":"<tensorflow><tensorboard>","CreationDate":"2019-08-29T20:48:53.093","AcceptedAnswerId":null,"OwnerUserId":6361137.0,"Title":"Increase number of points in Tensorboard Projector","Body":"<p>How do I increase the number of points in Tensorboard Projector?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/rLRpW.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rLRpW.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I am stuck with 100,000 points, and I'd like to increase it up to 800,000<\/p>\n","answers":[]}
+{"QuestionId":57717892,"AnswerCount":1,"Tags":"<linux><python-2.7><keras><protocol-buffers><azure-web-app-service>","CreationDate":"2019-08-29T21:24:23.943","AcceptedAnswerId":57899445.0,"OwnerUserId":304319.0,"Title":"Package import error when running Python 2 app on Azure App Services","Body":"<p>The application works fine on my local linux server but I have trouble of running it on Azure.<\/p>\n\n<p>I checked the Files folder and all python libraries are installed correctly under <em>\/antenv2.7\/lib\/python2.7\/site-packages<\/em>.<\/p>\n\n<p>Here is my pip requirements file:<\/p>\n\n<pre><code>bs4==0.0.1\nFlask==1.0.2\nFlask-Cors==3.0.7\nKeras==2.2.0\nnltk==3.2.5\nnumpy==1.14.5\npandas==0.23.3\nscikit-learn==0.19.1\nscipy==1.1.0\ntensorflow==1.9.0\nprotobuf==3.9.1\n<\/code><\/pre>\n\n<p>But when <strong><em>gunicorn<\/em><\/strong> starts, it generates this error. I don't know why protobuf can't be imported while other libraries like Keras can be imported without issue.<\/p>\n\n<p>Here is the full stacktrace. Thank you!<\/p>\n\n<pre><code>2019-08-29T21:09:37.697162031Z \n2019-08-29T21:09:37.697200631Z   _____                               \n2019-08-29T21:09:37.697207831Z   \/  _  \\ __________ _________   ____  \n2019-08-29T21:09:37.697212431Z  \/  \/_\\  \\___   \/  |  \\_  __ \\_\/ __ \\ \n2019-08-29T21:09:37.697216931Z \/    |    \\\/    \/|  |  \/|  | \\\/\\  ___\/ \n2019-08-29T21:09:37.697221431Z \\____|__  \/_____ \\____\/ |__|    \\___  &gt;\n2019-08-29T21:09:37.697226031Z         \\\/      \\\/                  \\\/ \n2019-08-29T21:09:37.697230631Z \n2019-08-29T21:09:37.697234631Z A P P   S E R V I C E   O N   L I N U X\n2019-08-29T21:09:37.697238731Z \n2019-08-29T21:09:37.697242831Z Documentation: http:\/\/aka.ms\/webapp-linux\n2019-08-29T21:09:37.697247031Z \n2019-08-29T21:09:38.233956948Z Starting OpenBSD Secure Shell server: sshd.\n2019-08-29T21:09:38.323527534Z Site's appCommandLine: startup.txt\n2019-08-29T21:09:38.329445366Z Checking of startup.txt is a file\n2019-08-29T21:09:38.331698079Z App command line is a file on disk\n2019-08-29T21:09:38.342990940Z startup.txt file exists on disk, reading its contents to run as startup arguments\n2019-08-29T21:09:38.343698744Z Contents of startupScript: gunicorn --bind=0.0.0.0 --timeout 600 server:app\n2019-08-29T21:09:38.344176546Z Launching oryx with: -appPath \/home\/site\/wwwroot -output \/opt\/startup\/startup.sh -virtualEnvName antenv2.7 -defaultApp \/opt\/defaultsite -bindPort 8000 -userStartupCommand 'gunicorn --bind=0.0.0.0 --timeout 600 server:app'\n2019-08-29T21:09:38.469155825Z Oryx Version : 0.2.20190518.2, Commit: 5e1ddd1855bcb53ce686e2124ed6e9603cb0587a\n2019-08-29T21:09:38.469591228Z \n2019-08-29T21:09:39.415029965Z Writing output script to '\/opt\/startup\/startup.sh'\n2019-08-29T21:09:39.467672451Z Using 'oryx-manifest.toml'.\n2019-08-29T21:09:39.472390477Z Checking if virtual environment was compressed...\n2019-08-29T21:09:39.729331673Z Using packages from virtual environment 'antenv2.7' located at '\/home\/site\/wwwroot\/antenv2.7'.\n2019-08-29T21:09:43.204792157Z [2019-08-29 21:09:43 +0000] [35] [INFO] Starting gunicorn 19.9.0\n2019-08-29T21:09:43.206124364Z [2019-08-29 21:09:43 +0000] [35] [INFO] Listening at: http:\/\/0.0.0.0:8000 (35)\n2019-08-29T21:09:43.206716767Z [2019-08-29 21:09:43 +0000] [35] [INFO] Using worker: sync\n2019-08-29T21:09:43.224584364Z [2019-08-29 21:09:43 +0000] [39] [INFO] Booting worker with pid: 39\n2019-08-29T21:10:08.320887834Z Using TensorFlow backend.\n2019-08-29T21:10:09.701427737Z [2019-08-29 21:10:09 +0000] [39] [ERROR] Exception in worker process\n2019-08-29T21:10:09.701474537Z Traceback (most recent call last):\n2019-08-29T21:10:09.701479937Z   File \"\/usr\/local\/lib\/python2.7\/site-packages\/gunicorn\/arbiter.py\", line 583, in spawn_worker\n2019-08-29T21:10:09.701484237Z     worker.init_process()\n2019-08-29T21:10:09.701487937Z   File \"\/usr\/local\/lib\/python2.7\/site-packages\/gunicorn\/workers\/base.py\", line 129, in init_process\n2019-08-29T21:10:09.701498537Z     self.load_wsgi()\n2019-08-29T21:10:09.701502237Z   File \"\/usr\/local\/lib\/python2.7\/site-packages\/gunicorn\/workers\/base.py\", line 138, in load_wsgi\n2019-08-29T21:10:09.701505937Z     self.wsgi = self.app.wsgi()\n2019-08-29T21:10:09.701509337Z   File \"\/usr\/local\/lib\/python2.7\/site-packages\/gunicorn\/app\/base.py\", line 67, in wsgi\n2019-08-29T21:10:09.701512937Z     self.callable = self.load()\n2019-08-29T21:10:09.701516437Z   File \"\/usr\/local\/lib\/python2.7\/site-packages\/gunicorn\/app\/wsgiapp.py\", line 52, in load\n2019-08-29T21:10:09.701520037Z     return self.load_wsgiapp()\n2019-08-29T21:10:09.701523337Z   File \"\/usr\/local\/lib\/python2.7\/site-packages\/gunicorn\/app\/wsgiapp.py\", line 41, in load_wsgiapp\n2019-08-29T21:10:09.701527037Z     return util.import_app(self.app_uri)\n2019-08-29T21:10:09.701530537Z   File \"\/usr\/local\/lib\/python2.7\/site-packages\/gunicorn\/util.py\", line 350, in import_app\n2019-08-29T21:10:09.701579737Z     __import__(module)\n2019-08-29T21:10:09.701612538Z   File \"\/home\/site\/wwwroot\/antenv2.7\/lib\/python2.7\/site-packages\/keras\/__init__.py\", line 3, in &lt;module&gt;\n2019-08-29T21:10:09.701616338Z     from . import utils\n2019-08-29T21:10:09.701619738Z   File \"\/home\/site\/wwwroot\/antenv2.7\/lib\/python2.7\/site-packages\/keras\/utils\/__init__.py\", line 6, in &lt;module&gt;\n2019-08-29T21:10:09.701623538Z     from . import conv_utils\n2019-08-29T21:10:09.701626838Z   File \"\/home\/site\/wwwroot\/antenv2.7\/lib\/python2.7\/site-packages\/keras\/utils\/conv_utils.py\", line 9, in &lt;module&gt;\n2019-08-29T21:10:09.701630638Z     from .. import backend as K\n2019-08-29T21:10:09.701633938Z   File \"\/home\/site\/wwwroot\/antenv2.7\/lib\/python2.7\/site-packages\/keras\/backend\/__init__.py\", line 87, in &lt;module&gt;\n2019-08-29T21:10:09.701637638Z     from .tensorflow_backend import *\n2019-08-29T21:10:09.701647638Z   File \"\/home\/site\/wwwroot\/antenv2.7\/lib\/python2.7\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 5, in &lt;module&gt;\n2019-08-29T21:10:09.701652338Z     import tensorflow as tf\n2019-08-29T21:10:09.701659738Z   File \"\/home\/site\/wwwroot\/antenv2.7\/lib\/python2.7\/site-packages\/tensorflow\/__init__.py\", line 22, in &lt;module&gt;\n2019-08-29T21:10:09.701663538Z     from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n2019-08-29T21:10:09.701667138Z   File \"\/home\/site\/wwwroot\/antenv2.7\/lib\/python2.7\/site-packages\/tensorflow\/python\/__init__.py\", line 52, in &lt;module&gt;\n2019-08-29T21:10:09.701670938Z     from tensorflow.core.framework.graph_pb2 import *\n2019-08-29T21:10:09.701674438Z   File \"\/home\/site\/wwwroot\/antenv2.7\/lib\/python2.7\/site-packages\/tensorflow\/core\/framework\/graph_pb2.py\", line 6, in &lt;module&gt;\n2019-08-29T21:10:09.701678038Z     from google.protobuf import descriptor as _descriptor\n2019-08-29T21:10:09.701681538Z ImportError: No module named google.protobuf\n2019-08-29T21:10:09.714154906Z [2019-08-29 21:10:09 +0000] [39] [INFO] Worker exiting (pid: 39)\n2019-08-29T21:10:10.020760372Z [2019-08-29 21:10:10 +0000] [35] [INFO] Shutting down: Master\n2019-08-29T21:10:10.022103379Z [2019-08-29 21:10:10 +0000] [35] [INFO] Reason: Worker failed to boot.\n<\/code><\/pre>\n","answers":[{"AnswerId":"57899445","CreationDate":"2019-09-12T03:45:53.210","ParentId":null,"OwnerUserId":"304319","Title":null,"Body":"<p>I have to add <code>__init__.py<\/code> under <code>site-packages\/google\/<\/code> manually after the deployment is complete.<\/p>\n\n<p>Azure App Services runs <em>Debian<\/em> linux. Maybe python runs a bit differently on this variety of linux. However my application runs on my <em>Ubuntu<\/em> vm properly without the need of doing this.<\/p>\n\n<p>Azure App Service is trash. Pretty much.<\/p>\n"}]}
+{"QuestionId":57718073,"AnswerCount":0,"Tags":"<keras><typeerror>","CreationDate":"2019-08-29T21:47:25.173","AcceptedAnswerId":null,"OwnerUserId":11996678.0,"Title":"keras 2.2.4 fit_generator crush with Error: TypeError: unsupported operand type(s) for \/: 'NoneType' and 'int' after two epochs","Body":"<p>I run my code with Keras 2.2.4. It got crushed exactly at the end of the second epoch with TypeError. To be more specific, the second epoch training is normal because the training loss is shown, but the second epoch validation is in trouble.<\/p>\n\n<p>I haven't tried anything about this error because I don't know how to check where it comes from.<\/p>\n\n<pre><code>Epoch 1\/80\n30\/30 [==============================] - 44s 1s\/step - loss: 0.6559 - acc: 0.6533 - val_loss: 0.6532 - val_acc: 0.6480\n\nEpoch 2\/80\n29\/30 [============================&gt;.] - ETA: 0s - loss: 0.6517 - acc: 0.6552\n<\/code><\/pre>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File \"blended_learning.py\", line\n  178, in \n      validation_steps=50,)   File \"build\/bdist.linux-x86_64\/egg\/keras\/legacy\/interfaces.py\", line 91, in\n  wrapper   File\n  \"build\/bdist.linux-x86_64\/egg\/keras\/engine\/training.py\", line 1418, in\n  fit_generator   File\n  \"build\/bdist.linux-x86_64\/egg\/keras\/engine\/training_generator.py\",\n  line 234, in fit_generator   File\n  \"build\/bdist.linux-x86_64\/egg\/keras\/legacy\/interfaces.py\", line 91, in\n  wrapper   File\n  \"build\/bdist.linux-x86_64\/egg\/keras\/engine\/training.py\", line 1472, in\n  evaluate_generator   File\n  \"build\/bdist.linux-x86_64\/egg\/keras\/engine\/training_generator.py\",\n  line 330, in evaluate_generator   File\n  \"build\/bdist.linux-x86_64\/egg\/keras\/utils\/data_utils.py\", line 601, in\n  get   File \"build\/bdist.linux-x86_64\/egg\/keras\/utils\/data_utils.py\",\n  line 595, in get   File \"\/usr\/lib\/python2.7\/multiprocessing\/pool.py\",\n  line 567, in get\n      raise self._value TypeError: unsupported operand type(s) for \/: 'NoneType' and 'int'<\/p>\n<\/blockquote>\n\n<p>I am combining an image matrix and a metadata vector as input.\nAs suggested, I attached the generator code below:<\/p>\n\n<pre><code>def __data_generation(self, list_IDs_temp, meta_df):\n    'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels)\n\n    # Initialization\n    X_image = np.empty((self.batch_size, self.dim[0], self.dim[1], self.n_channels))\n    X_meta = np.empty((self.batch_size, self.meta_len), dtype=float)\n    y = np.empty((self.batch_size), dtype=int)\n\n\n    # Generate data\n    for i, ID in enumerate(list_IDs_temp):\n        # Store sample\n        X_image[i,] = cv2.imread('.\/6_Batch1To6And8Mixed\/' + ID)\/255\n        X_meta[i,] = meta_df[meta_df.ID == ID].values[0][1:]\n\n        # Store class\n        y[i] = self.labels[ID]\n\n    if np.isnan(X_image).any():\n        print ('X_image has nan')\n    if np.isnan(X_meta).any():\n        print ('X_meta has nan')\n    if np.isnan(y).any():\n        print ('y has nan')\n\n    return [X_image, X_meta], y\n<\/code><\/pre>\n\n<p>steps_per_epoch is 30.\nAny ideas are appreciated!<\/p>\n","answers":[]}
+{"QuestionId":57718155,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><gpu>","CreationDate":"2019-08-29T21:55:15.147","AcceptedAnswerId":null,"OwnerUserId":2462485.0,"Title":"Tensorflow is not using the GPU for python_io library","Body":"<p>I am really new to tensorflow and this might be a simple question. I was wondering what are the correct mechanism for assignment of the GPU devices in the code. \nSpecifically I want to transfer this part of the code to the GPU:<\/p>\n\n<pre><code>tfr_opt = tf.python_io.TFRecordOptions(tf.python_io.TFRecordCompressionType.NONE)\n            for lod in range(self.resolution_log2 - 1):\n                tfr_file = self.tfr_prefix + '-r%02d.tfrecords' % (self.resolution_log2 - lod)\n                self.tfr_writers.append(tf.python_io.TFRecordWriter(tfr_file, tfr_opt))\n<\/code><\/pre>\n\n<p>But using line:<\/p>\n\n<pre><code>with tf.device('\/GPU:0'):\n<\/code><\/pre>\n\n<p>Doesn't seem to be helping and when I do nvidia-smi I am seeing that memory and GPU usage is 0.<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"58539644","CreationDate":"2019-10-24T10:45:06.487","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>Can you try to execute the below code and see which device it uses. <\/p>\n\n<pre><code>tf.debugging.set_log_device_placement(True)\n\n# Place tensors on the GPU\nwith tf.device('\/GPU:0'):\n  a = tf.constant([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])\n  b = tf.constant([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]])\n  c = tf.matmul(a, b)\n  print(c)\n<\/code><\/pre>\n\n<p>and then place the same \"tf.debugging.set_log_device_placement(True)\" and try the code you have mentioned and see if it works. <\/p>\n\n<p>It should log the device while executing the code.<\/p>\n"}]}
+{"QuestionId":57718447,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><pytorch><albumentations>","CreationDate":"2019-08-29T22:34:57.977","AcceptedAnswerId":null,"OwnerUserId":11468164.0,"Title":"Getting truth value of array with more than one element in ambigous for albuementation transform","Body":"<p>I am using albumentations for applying transform to a Pytorch model but getting this error and I m not getting any clue of what this error is about. Only thing I know is this is occuring due to transform which is being applied but not sure what is wrong with that.<\/p>\n\n<pre><code>ValueError: Traceback (most recent call last):\n  File \"\/opt\/conda\/lib\/python3.6\/site-packages\/torch\/utils\/data\/_utils\/worker.py\", line 99, in _worker_loop\n    samples = collate_fn([dataset[i] for i in batch_indices])\n  File \"\/opt\/conda\/lib\/python3.6\/site-packages\/torch\/utils\/data\/_utils\/worker.py\", line 99, in &lt;listcomp&gt;\n    samples = collate_fn([dataset[i] for i in batch_indices])\n  File \"&lt;ipython-input-23-119ea6bc360e&gt;\", line 24, in __getitem__\n    image = self.transform(image)\n  File \"\/opt\/conda\/lib\/python3.6\/site-packages\/albumentations\/core\/composition.py\", line 164, in __call__\n    need_to_run = force_apply or random.random() &lt; self.p\nValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()\n<\/code><\/pre>\n\n<p>This is are the snippets of code. \nDataloader <strong>getitem<\/strong>() method:<\/p>\n\n<pre><code>        image = cv2.imread(p_path)\n        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n        image = crop_image_from_gray(image)\n        image = cv2.resize(image, (IMG_SIZE, IMG_SIZE))\n        image = cv2.addWeighted ( image,4, cv2.GaussianBlur( image , (0,0) , 10) ,-4 ,128)\n        print(image.shape)\n        image = self.transform(image)\n<\/code><\/pre>\n\n<p>transforms applied :<\/p>\n\n<pre><code>val_transform = albumentations.Compose([\n            Normalize(\n                mean=[0.485, 0.456, 0.406],\n                std=[0.229, 0.224, 0.225],\n            ),\n            ToTensor()\n        ])\n<\/code><\/pre>\n\n<p>The class is called by:<\/p>\n\n<pre><code>valset       = MyDataset(val_df, transform = val_transform)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57724159","CreationDate":"2019-08-30T09:34:46.907","ParentId":null,"OwnerUserId":"8563649","Title":null,"Body":"<p>From official <a href=\"https:\/\/github.com\/albu\/albumentations\/blob\/master\/notebooks\/migrating_from_torchvision_to_albumentations.ipynb\" rel=\"nofollow noreferrer\">albumentation<\/a> documentation, you can apply transformation to image <\/p>\n\n<pre><code>from PIL import Image\nimport cv2\nimport numpy as np\nfrom torch.utils.data import Dataset\nfrom torchvision import transforms\nfrom albumentations import Compose, RandomCrop, Normalize, HorizontalFlip, Resize\nfrom albumentations.pytorch import ToTensor\n\n\nclass AlbumentationsDataset(Dataset):\n    \"\"\"__init__ and __len__ functions are the same as in TorchvisionDataset\"\"\"\n    def __init__(self, file_paths, labels, transform=None):\n        self.file_paths = file_paths\n        self.labels = labels\n        self.transform = transform\n\n    def __len__(self):\n        return len(self.file_paths)\n\n    def __getitem__(self, idx):\n        label = self.labels[idx]\n        file_path = self.file_paths[idx]\n\n        # Read an image with OpenCV\n        image = cv2.imread(file_path)\n\n        # By default OpenCV uses BGR color space for color images,\n        # so we need to convert the image to RGB color space.\n        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n\n        image = crop_image_from_gray(image)\n        image = cv2.resize(image, (IMG_SIZE, IMG_SIZE))\n        image = cv2.addWeighted ( image,4, cv2.GaussianBlur( image , (0,0) , 10) ,-4 ,128)\n\n        image = Img.fromarray(image, mode='RGB')  \n        if self.transform:\n            augmented = self.transform(image=np.array(image))\n            image = augmented['image']\n\n        image = np.transpose(image, (2, 0, 1))\n\n        return image, label\n\n\nalbumentations_transform = Compose([\n\n    Normalize(\n        mean=[0.485, 0.456, 0.406],\n        std=[0.229, 0.224, 0.225],\n    ),\n    ToTensor()\n])\n\n\nalbumentations_dataset = AlbumentationsDataset(\n    file_paths=['.\/images\/image_1.jpg', '.\/images\/image_2.jpg', '.\/images\/image_3.jpg'],\n    labels=[1, 2, 3],\n    transform=albumentations_transform,\n)\n\ntest_loader = DataLoader(dataset = albumentations_dataset, batch_size=4, drop_last=False, shuffle=False).\n<\/code><\/pre>\n"}]}
+{"QuestionId":57718512,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras><python-3.7><tensorboard>","CreationDate":"2019-08-29T22:44:15.183","AcceptedAnswerId":null,"OwnerUserId":7858253.0,"Title":"AttributeError: 'Sequential' object has no attribute 'run_eagerly'","Body":"<p>I'm trying to try using this model to train on rock, paper, scissor pictures. However, it was trained on 1800 pictures and only has an accuracy of 30-40%. I was then trying to use TensorBoard to see whats going on, but the error in the title appears.  <\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense, Dropout, Flatten\nfrom keras.layers import Conv2D, MaxPooling2D\nfrom tensorflow.python.keras.callbacks import TensorBoard\n\nmodel = Sequential()\nmodel.add(Conv2D(256, kernel_size=(4, 4),\n            activation='relu',\n            input_shape=(64,64,3)))\nmodel.add(Conv2D(196, (4, 4), activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n\nmodel.add(Conv2D(196, (4, 4), activation='relu'))\nmodel.add(Conv2D(196, (4, 4), activation='relu'))\nmodel.add(Dropout(0.25))\n\nmodel.add(Conv2D(128, (4, 4), activation='relu'))\nmodel.add(Conv2D(128, (4, 4), activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n\nmodel.add(Conv2D(96, (4, 4), activation='relu'))\nmodel.add(Conv2D(96, (4, 4), activation='relu'))\nmodel.add(Dropout(0.25))\n\nmodel.add(Flatten())\nmodel.add(Dense(128, activation='relu'))\nmodel.add(Dropout(0.5))\n\nmodel.add(Dense(3, activation='softmax'))\n\n''' here it instantiates the tensorboard '''\ntensorboard = TensorBoard(log_dir=\"C:\/Users\/bamla\/Desktop\/RPS project\/Logs\")\n\nmodel.compile(loss=\"sparse_categorical_crossentropy\",\n        optimizer=\"SGD\",\n        metrics=['accuracy'])\n\nmodel.summary()\n\n''' Here its fitting the model '''\nmodel.fit(x_train, y_train, batch_size=50, epochs = 3, callbacks= \n[tensorboard])\n<\/code><\/pre>\n\n<p>This outputs:<\/p>\n\n<pre><code>Traceback (most recent call last):\n\nFile \"c:\/Users\/bamla\/Desktop\/RPS project\/Testing.py\", line 82, in &lt;module&gt;\nmodel.fit(x_train, y_train, batch_size=50, epochs = 3, callbacks= \n[tensorboard])\n\nFile \"C:\\Users\\bamla\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site- \npackages\\keras\\engine\\training.py\", line 1178, in fit\nvalidation_freq=validation_freq)\n\nFile \"C:\\Users\\bamla\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site- \n packages\\keras\\engine\\training_arrays.py\", line 125, in fit_loop\ncallbacks.set_model(callback_model)\n\n File \"C:\\Users\\bamla\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site- \npackages\\keras\\callbacks.py\", line 68, in set_model\ncallback.set_model(model)\n\nFile \"C:\\Users\\bamla\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site- \npackages\\tensorflow\\python\\keras\\callbacks.py\", line 1509, in set_model\nif not model.run_eagerly:\n\nAttributeError: 'Sequential' object has no attribute 'run_eagerly'\n<\/code><\/pre>\n\n<p>Also, if you have any tips on how to improve the accuracy it would be appreciated!<\/p>\n","answers":[{"AnswerId":"57721577","CreationDate":"2019-08-30T06:30:09.697","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>The problem is here:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense, Dropout, Flatten\nfrom keras.layers import Conv2D, MaxPooling2D\nfrom tensorflow.python.keras.callbacks import TensorBoard\n<\/code><\/pre>\n\n<p>Do not mix <code>keras<\/code> and <code>tf.keras<\/code> imports, these are <strong>not compatible with each other<\/strong>, and produce weird errors as the ones you are seeing.<\/p>\n"}]}
+{"QuestionId":57718525,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><tensorboard>","CreationDate":"2019-08-29T22:46:03.080","AcceptedAnswerId":null,"OwnerUserId":11996844.0,"Title":"Running into an issue with tensorboard returning \"tensorboard: error: unrecognized arguments:\"","Body":"<p>I want to call <code>tensorboard -- logdir = 'logs'<\/code> in the terminal but I get an error saying tensorboard: error: unrecognized arguments: -- logdir = 'logs'<\/p>\n\n<p>Can someone help me out?<\/p>\n","answers":[]}
+{"QuestionId":57718653,"AnswerCount":0,"Tags":"<python><numpy><matplotlib><pytorch>","CreationDate":"2019-08-29T23:04:32.823","AcceptedAnswerId":null,"OwnerUserId":710942.0,"Title":"Why does one Python script raise a qt.qpa.plugin error, while another identical script does not?","Body":"<p>I have two virtually identical scripts running in the same PyCharm IDE. They both call into a third script that uses matplotlib to output a Numpy array to a PNG. One of the scripts works fine and outputs the PNG. The other script raises the following error:<\/p>\n\n<blockquote>\n  <p>qt.qpa.plugin: Could not load the Qt platform plugin \"windows\" in \"\" even though it was found.<\/p>\n<\/blockquote>\n\n<p>The differences between the scripts are minimal - they only vary in that they each import a different pytorch model (both models created by me). This leads me to think there might be some sort of environmental difference, but they are both launched from the same IDE and Conda environment.<\/p>\n\n<p>Does anyone have any suggestions?<\/p>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>Here is a stripped down version of the two files. The bits I have removed are identical to each other down to the character and whitespace.<\/p>\n\n<p>File 1:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import argparse\nfrom average_meter import AverageMeter\nimport criteria  # -&gt; NOTE: This is not used, but was not removed in this file\nimport dataset\nfrom densenet_mapmaker_model import DensenetMapMakerModel  # -&gt; Different model imported\nimport json\nimport math\nimport os\nimport time\nimport torch\nfrom torch.autograd import Variable\nimport torch.nn as nn\nfrom torchvision import transforms\nimport visualizer\n\nparser = argparse.ArgumentParser(description='Densenet Map Maker Trainer for Steel')\n\n# ...Define args...\n\n##########################################################################################\n# These args were defined here, but not in the other file. They are not actually used here, just defined and ignored.\nparser.add_argument('--shrinkage_coef', dest='shrinkage_coef', type=int,\n                    help='Use a shrinking gaussian loss and multiply its result by this amount.')\n\nparser.add_argument('--multiply', dest='multiply', type=int,\n                    help='If 1, in the shrinking gaussian loss, ' +\n                         'multiply by 1000 before comparing and then divide the loss by 1000.')\n\nparser.add_argument('--diff_coef', dest='diff_coef', type=int,\n                    help='Use the count difference loss and multiply its result by this amount.')\n\nparser.add_argument('--clustering_coef', dest='clustering_coef', type=int,\n                    help='Use the clustering loss and multiply its result by this amount.')\n##########################################################################################\n\nuse_cuda = torch.cuda.is_available()\n\n\ndef main():\n    best_mse = 1e6\n\n    version = 1\n\n    args = parser.parse_args()\n    # ...set up default arg values...\n    args.loss_calc_method = 'DenseNet'  # Defined here, but not in the other file. Not used here and could be removed.\n\n    # ...load dataset lists...\n\n    model = DensenetMapMakerModel()  # -&gt; Note: Different model used\n    if use_cuda:\n        model = model.cuda()\n\n    # ...Create criteria and optimizer...\n\n    # ...Load weights from last training run...\n\n    for epoch in range(args.start_epoch, args.epochs):\n        train(train_list, model, criterion, optimizer, epoch, args)\n        mse = validate(val_list, model, criterion, epoch, args)\n\n        # ...Check result and save weights...\n\n\ndef train(...):\n\n    # ...Set up counters and dataset...\n\n    model.train()\n    end = time.time()\n\n    # img: torch.Tensor      # Image data\n    # location : torch.Tensor. Target annotation locations for the image.\n    # image_id : str. Unique ID of image.\n    for i, (img, location, image_id) in enumerate(train_loader):\n\n        # ...Run input through model, check result and backpropagate...\n\n        if i % args.print_freq == 0:\n            # Note: Print statement has different whitespace formatting from other file...\n            print('Epoch: [{0}][{1}\/{2}]\\t'\n                  'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\\t'\n                  'Data {data_time.val:.3f} ({data_time.avg:.3f})\\t'\n                  'Loss {loss.val:.4f} ({loss.avg:.4f})\\t'\n                .format(\n                epoch, i, len(train_loader), batch_time=batch_time,\n                data_time=data_time, loss=losses))\n\n\ndef validate(...) -&gt; float:\n\n    # ...Set up dataset...\n\n    # img: torch.Tensor      # Image data\n    # location : torch.Tensor. Target annotation locations for the image.\n    # image_id : str. Unique ID of image.\n    for i, (img, location, image_id) in enumerate(test_loader):\n\n        # ...Run input through model and check result...\n\n        # Output target and source the first time...\n        output_dir = args.output_dir\n        if args.output_images == 1 and epoch % 1 == 0 and i == 2:  # -&gt; NOTE: Other file does NOT have additional check i == 2\n            visualizer.display_tensor(output, output_dir + \"\/epoch_\" + str(epoch) + \"_\" + str(i) + \"_output\")\n\n\nif __name__ == '__main__':\n    main()\n<\/code><\/pre>\n\n<p>File 2:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import argparse\nfrom average_meter import AverageMeter\nimport dataset\nfrom steel_scanner_model import SteelScanner  # -&gt; Different model imported\nimport json\nimport math\nimport os\nimport time\nimport torch\nfrom torch.autograd import Variable\nimport torch.nn as nn\nfrom torchvision import transforms\nimport visualizer\n\nparser = argparse.ArgumentParser(description='Densenet Map Maker Trainer for Steel')\n\n# ...Define args...\n\nuse_cuda = torch.cuda.is_available()\n\n\ndef main():\n    best_mse = 1e6\n\n    version = 1\n\n    args = parser.parse_args()\n    # ...set up default arg values...\n\n    # ...load dataset lists...\n\n    model = SteelScanner()  # -&gt; Note: Different model used\n    if use_cuda:\n        model = model.cuda()\n\n    # ...Create criteria and optimizer...\n\n    # ...Load weights from last training run...\n\n    for epoch in range(args.start_epoch, args.epochs):\n        train(train_list, model, criterion, optimizer, epoch, args)\n        mse = validate(val_list, model, criterion, epoch, args)\n\n        # ...Check result and save weights...\n\n\ndef train(...):\n\n    # ...Set up counters and dataset...\n\n    model.train()\n    end = time.time()\n\n    # img: torch.Tensor      # Image data\n    # location : torch.Tensor. Target annotation locations for the image.\n    # image_id : str. Unique ID of image.\n    for i, (img, location, image_id) in enumerate(train_loader):\n\n        # ...Run input through model, check result and backpropagate...\n\n        if i % args.print_freq == 0:\n            # Note: Print statement has different whitespace formatting from other file...\n            print('Epoch: [{0}][{1}\/{2}]\\t'\n                  'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\\t'\n                  'Data {data_time.val:.3f} ({data_time.avg:.3f})\\t'\n                  'Loss {loss.val:.4f} ({loss.avg:.4f})\\t'\n                  .format(epoch, i, len(train_loader), batch_time=batch_time, data_time=data_time, loss=losses))\n\n\ndef validate(...) -&gt; float:\n\n    # ...Set up dataset...\n\n    # img: torch.Tensor      # Image data\n    # location : torch.Tensor. Target annotation locations for the image.\n    # image_id : str. Unique ID of image.\n    for i, (img, location, image_id) in enumerate(test_loader):\n\n        # ...Run input through model and check result...\n\n        # Output target and source the first time...\n        output_dir = args.output_dir\n        if args.output_images == 1 and epoch % 1 == 0:  # -&gt; NOTE: Other file has additional check i == 2\n            # NOTE: The error occurs inside 'display_tensor'\n            visualizer.display_tensor(output, output_dir + \"\/epoch_\" + str(epoch) + \"_\" + str(i) + \"_output\")\n\n\nif __name__ == '__main__':\n    main()\n<\/code><\/pre>\n\n<p>The error occurs inside <code>visualizer.display_tensor<\/code> when called from the second file.<\/p>\n","answers":[]}
+{"QuestionId":57718883,"AnswerCount":1,"Tags":"<python><tensorflow><google-colaboratory>","CreationDate":"2019-08-29T23:45:25.847","AcceptedAnswerId":57750712.0,"OwnerUserId":11996941.0,"Title":"InvalidArgumentError: ValueError: `generator` yielded an element of shape (59, 224, 224, 3)","Body":"<p>I tried to follow this popular Tensorflow tutorial.<\/p>\n\n<p>Which basically should be a copy-paste, but I still get an error in my code.<\/p>\n\n<p>I have tried use None as a shape but that does not work. And since I am very new at this I am not really sure what to google for.<\/p>\n\n<p>This is the code where the error occurs;<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>epochs = 10\n\nhistory = model.fit(train_generator, \n                    epochs=epochs, \n                    validation_data=val_generator)\n<\/code><\/pre>\n\n<pre><code>InvalidArgumentError  Traceback (most recent call last)\n&lt;ipython-input-15-238fbae26d39&gt; in &lt;module&gt;()\n      3 history = model.fit(train_generator, \n      4                     epochs=epochs,\n----&gt; 5                     validation_data=val_generator)\n\n10 frames\n\/usr\/local\/lib\/python3.6\/dist-packages\/six.py in raise_from(value, from_value)\n\nInvalidArgumentError: ValueError: `generator` yielded an element of shape (59, 224, 224, 3) where an element of shape (64, 224, 224, 3) was expected.\nTraceback (most recent call last):\n\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow_core\/python\/ops\/script_ops.py\", line 235, in __call__\n    ret = func(*args)\n\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow_core\/python\/data\/ops\/dataset_ops.py\", line 630, in generator_py_func\n    \"of shape %s was expected.\" % (ret_array.shape, expected_shape))\n\nValueError: `generator` yielded an element of shape (59, 224, 224, 3) where an element of shape (64, 224, 224, 3) was expected.\n<\/code><\/pre>\n","answers":[{"AnswerId":"57750712","CreationDate":"2019-09-02T02:00:56.940","ParentId":null,"OwnerUserId":"7265511","Title":null,"Body":"<p>I think that it could be a bug from tf-nightly(-gpu)-2.0 since I create a virtualenv with standard tensorflow 1.14 and it works.<\/p>\n"}]}
+{"QuestionId":57718903,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-08-29T23:49:03.130","AcceptedAnswerId":null,"OwnerUserId":2281101.0,"Title":"How to improve Volatile GPU-Util?","Body":"<p>Using Keras to train model, Volatile GPU-Util of two GPU is too low.\n<a href=\"https:\/\/i.stack.imgur.com\/aPHSG.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/aPHSG.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The block of code:<\/p>\n\n<pre><code>%%time\nnp.random.seed(seed)\n\nmodel_d2v_01 = Sequential()\nmodel_d2v_01.add(Dense(64, activation='relu', input_dim=400))\nmodel_d2v_01.add(Dense(1, activation='sigmoid'))\n\nmodel_d2v_01 = multi_gpu_model(model_d2v_01, gpus=2)\nmodel_d2v_01.compile(loss='binary_crossentropy',\n                       optimizer='adam',\n                       metrics=['accuracy'])\n\nmodel_d2v_01.fit(train_vecs_dbow_dmm, y_train, validation_data=(validation_vecs_dbow_dmm, y_validation), epochs=5, batch_size=32*2, verbose=2)\n<\/code><\/pre>\n\n<p>How to modify this code? some suggestion, please.<\/p>\n","answers":[{"AnswerId":"57724002","CreationDate":"2019-08-30T09:24:40.090","ParentId":null,"OwnerUserId":"2281101","Title":null,"Body":"<p>I found that this situation was normal. The cause was from few layers, so the model was too simple. when add more model layer, this situation will improve. for example:<\/p>\n\n<pre><code>%%time\nnp.random.seed(seed)\n# with tf.device('\/cpu:0'):\nmodel_d2v_12 = Sequential()\nmodel_d2v_12.add(Dense(512, activation='relu', input_dim=400))\nmodel_d2v_12.add(Dense(512, activation='relu'))\nmodel_d2v_12.add(Dense(512, activation='relu'))\nmodel_d2v_12.add(Dense(1, activation='sigmoid'))\n\nmodel_d2v_12 = multi_gpu_model(model_d2v_12, gpus=2)\nmodel_d2v_12.compile(loss='binary_crossentropy',\n                       optimizer='adam',\n                       metrics=['accuracy'])\n\nmodel_d2v_12.fit(train_vecs_dbow_dmm, y_train, validation_data=(validation_vecs_dbow_dmm, y_validation), epochs=10, batch_size=2048*2, verbose=2)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57718976,"AnswerCount":0,"Tags":"<tensorflow><multi-gpu>","CreationDate":"2019-08-30T00:03:51.627","AcceptedAnswerId":null,"OwnerUserId":6450489.0,"Title":"Unexpected run times for 0, 1, and 2 GPUs","Body":"<p>I tried <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/image\/cifar10_estimator\" rel=\"nofollow noreferrer\">tensorflow's cifar10-estimator model<\/a> on two systems to evaluate their gpu performance. The configuration was:<\/p>\n\n<ul>\n<li>PC#1 is a bit older system with 1 GTX 1080 Ti GPU and is a Windows 10 OS configured with Tensorflow 1.12.2, Cuda Toolkit 9.0.176, CuDNN 7.3.0, and Nvidia display driver 385.54. Python version is 3.6.5.  <\/li>\n<li>PC#2 is a newer system with 2 RTX 2080 Ti GPUs and is a Unix Centos 7 OS configured with Tensorflow 1.14.0, Cuda Toolkit 10.0.130, CuDNN 7.6.3, Nvidia display driver 430.34, and Python version is 2.7.5.  <\/li>\n<\/ul>\n\n<p>I followed the instructions and command details as specified in the readme page of the above link. For only cpu setting (--num_gpus=0), training speed has stated from about 26 examples\/sec for PC#1 but just 9 examples\/sec for PC#2 (which I think is problematic). For one GPU setting (--num_gpus=1), training speed at the end (1000 steps) was 1940 examples\/sec for PC#1 and 2400 examples\/sec for PC#2 (which is good!). But for two GPU (only tested on PC#2), performance was more or less the same as one GPU (which I didn't expect).<\/p>\n\n<p>Any guide or explanation to understand the differences and possibly resolve the problems?<\/p>\n","answers":[]}
+{"QuestionId":57719106,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><machine-learning><computer-vision>","CreationDate":"2019-08-30T00:28:50.687","AcceptedAnswerId":null,"OwnerUserId":10002608.0,"Title":"How to Implement Hard Example Miner with Faster-RCNN-Inception on Tensorflow Object Detection API?","Body":"<p>I am using Tensorflow API for Object Detection. How do we actually implement the Hard Example Miner in the API? Is it the same way as on SSD Mobile Net's CONFIG file?<\/p>\n\n<p>I see this as an example on what to add to the SSDMobilenet CONFIG file, but how do we implement this on FasterRCNNInception?<\/p>\n\n<pre><code>n = 3000 # Number of images with no bounding boxes\nhard_example_miner {\nnum_hard_examples: n\niou_threshold: 0.99\nloss_type: CLASSIFICATION\nmax_negatives_per_positive: 3\nin_negatives_per_image: 3\n}\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57719272,"AnswerCount":0,"Tags":"<python><keras><deep-learning><conv-neural-network>","CreationDate":"2019-08-30T01:01:14.573","AcceptedAnswerId":null,"OwnerUserId":4646458.0,"Title":"Incrementing validation_steps improves my model (Keras)","Body":"<p>I have this <code>fit_generator<\/code> function:<\/p>\n\n<pre><code>model.fit_generator(batch_gen(xtrain, ytrain, 150), \n epochs=15,\n steps_per_epoch=200,\n validation_data=batch_gen(xtest, y_test, 150),\n validation_steps=100)\n<\/code><\/pre>\n\n<p>Well, a batch size of 150 for the validation data generator and a <code>validation_steps<\/code> of 100 would result in <code>150*100 = 15000<\/code> validation elements used in a epoch. My test set size is 3000, so it should be more than enough.<\/p>\n\n<p>When I try this model, it doesn't work well. I change <code>validation_steps<\/code> to 200 and then it works successfully. Why? <\/p>\n\n<p>Theoretically it should have <code>len(test_data)\/batch_size<\/code>(3000\/150 = 20, ten times less) to work properly, since validation data is not changing the training...<\/p>\n","answers":[]}
+{"QuestionId":57719398,"AnswerCount":1,"Tags":"<tensorflow><keras><model><neural-network>","CreationDate":"2019-08-30T01:23:49.203","AcceptedAnswerId":null,"OwnerUserId":4515762.0,"Title":"Unable to save model with tensorflow 2.0.0 beta1","Body":"<p>I have tried all the options described in the documentation but none of them allowed me to save my model in tensorflow 2.0.0 beta1. I've also tried to upgrade to the (also unstable) TF2-RC but that ruined even the code I had working in beta so I quickly rolled back for now to beta.<\/p>\n\n<p>See a minimal reproduction code below.<\/p>\n\n<p>What I have tried: <\/p>\n\n<ol>\n<li><pre><code>model.save(\"mymodel.h5\") \n<\/code><\/pre><\/li>\n<\/ol>\n\n<blockquote>\n  <p>NotImplementedError: Saving the model to HDF5 format requires the\n  model to be a Functional model or a Sequential model. It does not work\n  for subclassed models, because such models are defined via the body of\n  a Python method, which isn't safely serializable. Consider saving to\n  the Tensorflow SavedModel format (by setting save_format=\"tf\") or\n  using <code>save_weights<\/code>.<\/p>\n<\/blockquote>\n\n<ol start=\"2\">\n<li><pre><code>model.save(\"mymodel\", format='tf')\n<\/code><\/pre><\/li>\n<\/ol>\n\n<blockquote>\n  <p>ValueError: Model &lt;<strong>main<\/strong>.CVAE object at 0x7f1cac2e7c50> cannot be\n  saved because the input shapes have not been set. Usually, input\n  shapes are automatically determined from calling .fit() or .predict().\n  To manually set the shapes, call model._set_inputs(inputs).<\/p>\n<\/blockquote>\n\n<p>3.<\/p>\n\n<pre><code>model._set_input(input_sample)\nmodel.save(\"mymodel\", format='tf') \n<\/code><\/pre>\n\n<blockquote>\n  <p>AssertionError: tf.saved_model.save is not supported inside a traced\n  @tf.function. Move the call to the outer eagerly-executed context.<\/p>\n<\/blockquote>\n\n<p>And this is where I am stuck now because it gives me no reasonable hint whatsoever. That's because I am NOT calling the save() function from a @tf.function, I'm already calling it from the outermost scope possible. In fact, I have no @tf.function at all in this minimal reproduction script below and still getting the same error.<\/p>\n\n<p>So I really have no idea how to save my model, I've tried every options and they all throw errors and provide no hints.<\/p>\n\n<p>The minimal reproduction example below works fine if you set save_model=False and it reproduces the error when save_model=True. <\/p>\n\n<p>It may seem unnecessary in this simplified auto-encoder code example to use a subclassed model but I have lots of custom functions added to it in my original VAE code that I need it for.<\/p>\n\n<p>Code:<\/p>\n\n<pre><code>import tensorflow as tf\n\nsave_model = True\n\nlearning_rate = 1e-4\nBATCH_SIZE = 100\nTEST_BATCH_SIZE = 10\ncolor_channels = 1\nimsize = 28\n\n(train_images, _), (test_images, _) = tf.keras.datasets.mnist.load_data()\n\ntrain_images = train_images[:5000, ::]\ntest_images = train_images[:1000, ::]\ntrain_images = train_images.reshape(-1, imsize, imsize, 1).astype('float32')\ntest_images = test_images.reshape(-1, imsize, imsize, 1).astype('float32')\ntrain_images \/= 255.\ntest_images \/= 255.\ntrain_dataset = tf.data.Dataset.from_tensor_slices(train_images).batch(BATCH_SIZE)\ntest_dataset = tf.data.Dataset.from_tensor_slices(test_images).batch(TEST_BATCH_SIZE)\n\nclass AE(tf.keras.Model):\n    def __init__(self):\n        super(AE, self).__init__()\n        self.network = tf.keras.Sequential([\n            tf.keras.layers.InputLayer(input_shape=(imsize, imsize, color_channels)),\n            tf.keras.layers.Flatten(),\n            tf.keras.layers.Dense(50),\n            tf.keras.layers.Dense(imsize**2 * color_channels),\n            tf.keras.layers.Reshape(target_shape=(imsize, imsize, color_channels)),\n        ])\n    def decode(self, input):\n        logits = self.network(input)\n        return logits\n\noptimizer = tf.keras.optimizers.Adam(learning_rate)\nmodel = AE()\n\ndef compute_loss(data):\n    logits = model.decode(data)\n    loss = tf.reduce_mean(tf.losses.mean_squared_error(logits, data))\n    return loss\n\ndef train_step(data):\n    with tf.GradientTape() as tape:\n        loss = compute_loss(data)\n    gradients = tape.gradient(loss, model.trainable_variables)\n    optimizer.apply_gradients(zip(gradients, model.trainable_variables))\n    return loss, 0\n\ndef test_step(data):\n    loss = compute_loss(data)\n    return loss\n\ninput_shape_set = False\nepoch = 0\nepochs = 20\nfor epoch in range(epochs):\n    for train_x in train_dataset:\n        train_step(train_x)\n    if epoch % 1 == 0:\n        loss = 0.0\n        num_batches = 0\n        for test_x in test_dataset:\n            loss += test_step(test_x)\n            num_batches += 1\n        loss \/= num_batches\n        print(\"Epoch: {}, Loss: {}\".format(epoch, loss))\n\n        if save_model:\n            print(\"Saving model...\")\n            if not input_shape_set:\n                # Note: Why set input shape manually and why here:\n                # 1. If I do not set input shape manually: ValueError: Model &lt;main.CVAE object at 0x7f1cac2e7c50&gt; cannot be saved because the input shapes have not been set. Usually, input shapes are automatically determined from calling .fit() or .predict(). To manually set the shapes, call model._set_inputs(inputs).\n                # 2. If I set input shape manually BEFORE the first actual train step, I get: RuntimeError: Attempting to capture an EagerTensor without building a function.\n                model._set_inputs(train_dataset.__iter__().next())\n                input_shape_set = True\n            # Note: Why choose tf format: model.save('MNIST\/Models\/model.h5') will return NotImplementedError: Saving the model to HDF5 format requires the model to be a Functional model or a Sequential model. It does not work for subclassed models, because such models are defined via the body of a Python method, which isn't safely serializable. Consider saving to the Tensorflow SavedModel format (by setting save_format=\"tf\") or using save_weights.\n            model.save('MNIST\/Models\/model', save_format='tf')\n<\/code><\/pre>\n","answers":[{"AnswerId":"57729976","CreationDate":"2019-08-30T16:00:40.517","ParentId":null,"OwnerUserId":"4515762","Title":null,"Body":"<p>I have tried the same minimal reproduction example in tensorflow-gpu 2.0.0-rc0 and the error was more revealing than what the beta version gave me. The error in RC says:<\/p>\n\n<blockquote>\n  <p>NotImplementedError: When subclassing the Model class, you should\n  implement a call method.<\/p>\n<\/blockquote>\n\n<p>This got me read through <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/keras\/custom_layers_and_models\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/guide\/keras\/custom_layers_and_models<\/a> where I found examples of how to do subclassing in TF2 in a way that allows saving. I was able to resolve the error and have the model saved by replacing my 'decode' method by 'call' in the above example (although this will be more complicated with my actual code where I had various methods defined for the class). This solved the error both in beta and in rc. Strangely, the training (or the saving) got also much faster in rc.<\/p>\n"}]}
+{"QuestionId":57719499,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-30T01:40:11.157","AcceptedAnswerId":57721028.0,"OwnerUserId":2467772.0,"Title":"NCHW and NHWC netowrk format conversion Tensorflow model","Body":"<p>For this Humanpose Tensorflow network, <a href=\"https:\/\/github.com\/ildoonet\/tf-pose-estimation\/blob\/master\/tf_pose\/network_cmu.py\" rel=\"nofollow noreferrer\">network_cmu<\/a> and <a href=\"https:\/\/github.com\/ildoonet\/tf-pose-estimation\/blob\/master\/tf_pose\/network_base.py\" rel=\"nofollow noreferrer\">base<\/a>, it accepts only NHWC input format.\nIf I construct the network in NCHW format, there is error as<\/p>\n\n<pre><code>Depth of input (32) is not a multiple of input depth of filter (3) for 'conv1_1\/Conv2D' (op: 'Conv2D') with input shapes: [1,3,24,32], [3,3,3,64]. \n<\/code><\/pre>\n\n<p>My code to construct the network is<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nfrom network_cmu import CmuNetwork\n\n\ndef main():\n    #print(tensor_util.MakeNdarray(n.attr['value'].tensor))\n    placeholder_input = tf.placeholder(dtype=tf.float32, shape=(1, 3, 24, 32), name=\"image\") \n    net = CmuNetwork({'image': placeholder_input}, trainable=False)\n    # Add an op to initialize the variables.\n    init_op = tf.global_variables_initializer()\n    saver = tf.train.Saver()\n    init_op = tf.global_variables_initializer()\n    with tf.Session() as sess:\n        sess.run(init_op)\n        #for n in tf.get_default_graph().as_graph_def().node:\n        #   print(n.name)\n        save_path = saver.save(sess, \"cmuThreeOutputs\/model.ckpt\") \n\nif __name__ == '__main__':\n   main()\n<\/code><\/pre>\n\n<p>What should I change to have network in NCHW format?<\/p>\n","answers":[{"AnswerId":"57721028","CreationDate":"2019-08-30T05:39:02.103","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>You can make use of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/transpose\" rel=\"nofollow noreferrer\"><code>tf.transpose<\/code><\/a> to shift your axis from NHWC to NCHW<\/p>\n\n<pre><code>input_ = tf.convert_to_tensor(np.random.rand(1, 3, 24, 32))\na1 = tf.transpose(input_, perm=[0, 2, 3, 1])\nprint(a1.shape)  # 1, 24, 32, 3\n<\/code><\/pre>\n\n<p>You may even make use of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reshape\" rel=\"nofollow noreferrer\"><code>tf.reshape<\/code><\/a><\/p>\n\n<pre><code>a2 = tf.reshape(input_, (-1, input_.shape[2], input_.shape[3], input_.shape[1]))\nprint(a2.shape)  # 1, 24, 32, 3\n<\/code><\/pre>\n"}]}
+{"QuestionId":57719531,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras><tensorflow-datasets>","CreationDate":"2019-08-30T01:46:06.603","AcceptedAnswerId":57731042.0,"OwnerUserId":5510818.0,"Title":"TensorFlow training - \"Batch size\" and tf.unpack - unpacking non \"batch sized\" dynamic values?","Body":"<p>Hello TensorFlow friends.<\/p>\n\n<p><strong><em>TL-DR:<\/em><\/strong><\/p>\n\n<p>Im using Tensorflow 1.1.4 on Google collab with eager execution.<\/p>\n\n<p>My training fails on the last batch of the first epoch because I have a forced batch size tf.unpack in my training loop which sometimes does not get a full batch, and fails on :<\/p>\n\n<pre><code>Epoch 1\/5\n147\/148 [============================&gt;.] - ETA: 0s - loss: 0.0133 - acc: 0.3863\n---------------------------------------------------------------------------\nInvalidArgumentError                      Traceback (most recent call last)\n&lt;ipython-input-133-bba4aeb63f3b&gt; in &lt;module&gt;()\n      1 steps_per_epoch=tf.ceil(len(all_image_paths)\/BATCH_SIZE).numpy()\n      2 \n----&gt; 3 history = model.fit(ds, epochs=5, steps_per_epoch=steps_per_epoch)\n\n8 frames\n\/usr\/local\/lib\/python3.6\/dist-packages\/six.py in raise_from(value, from_value)\n\nInvalidArgumentError: Input shape axis 0 must equal 32, got shape [30]\n     [[{{node unstack}}]] [Op:IteratorGetNextSync]\n<\/code><\/pre>\n\n<p>I can't figure out how to unpack a tensor containing <code>BATCH_SIZE<\/code> of file paths (string tensors) which I unpack, load and pre-process into <code>BATCH_SIZE<\/code> of image tensors:<\/p>\n\n<pre><code>def load_and_preprocess_image(batch_of_paths, batch_of_labels):\n\n  # unpack our images from our batch\n  list_of_paths = tf.unstack(batch_of_paths, num=BATCH_SIZE)\n  list_of_images = []\n  for path in list_of_paths:\n      image = tf.read_file(path)\n      if image == None:\n        print(\"Unable to load image at path:\" + path )\n      image = preprocess_image(image)\n      list_of_images.append(image)\n\n\n  # repack our now loaded and processed images into a batch\n  batch_of_images = tf.stack(list_of_images, 0)\n\n  return batch_of_images,batch_of_labels\n<\/code><\/pre>\n\n<p>Is there a way to have my <code>BATCH_SIZE<\/code> be dynamically read from a variable that the dataset passes when a batch is not full? In my case, I request <code>BATCH_SIZE = 32<\/code>, but get <code>30<\/code> on my last pass - and then the abo<\/p>\n\n<p><strong><em>LONGER \/ FULL CONTEXT<\/em><\/strong><\/p>\n\n<p>Im using Tensorflow 1.1.4 on Google collab with eager execution.<\/p>\n\n<p>I am attempting to train a model (multi class, multi label) against CSV data set which lists image file paths and then columns of labels being positively present, negatively present, or unknown.<\/p>\n\n<p>Im also aware ,y data set also contains lots of unlabeled data (we'll address that later :)<\/p>\n\n<p>The header and the first row looks like:<\/p>\n\n<pre><code>filepath,color_key_blue,color_key_green,color_key_luma,color_key_matte,color_saturation_desaturated,color_saturation_neutral,color_saturation_pastel,color_saturation_saturated,color_theory_analagous,color_theory_complementary,color_theory_monochrome,color_tones_blackwhite,color_tones_cool,color_tones_warm,exterior,exterior_airplane,exterior_airport,exterior_apartment,exterior_auto_body,exterior_beach,exterior_bicycle,exterior_boat,exterior_bridge,exterior_bus,exterior_bus_stop,exterior_canyon,exterior_car,exterior_castle,exterior_cathedral,exterior_cave,exterior_church,exterior_city,exterior_desert,exterior_farm,exterior_forest,exterior_glacier,exterior_helicopter,exterior_hospital,exterior_house,exterior_industrial,exterior_lake,exterior_library,exterior_mall,exterior_mansion,exterior_monastery,exterior_mosque,exterior_motorcycle,exterior_mountains,exterior_ocean,exterior_office,exterior_palace,exterior_park,exterior_parkinglot,exterior_pier,exterior_plains,exterior_playground,exterior_polar,exterior_port,exterior_restaurant,exterior_river,exterior_road,exterior_ruins,exterior_school,exterior_sidewalk,exterior_sky,exterior_skyscraper,exterior_space,exterior_spacecraft,exterior_stadium,exterior_station_gas,exterior_station_subway,exterior_station_train,exterior_store,exterior_suburb,exterior_synagogue,exterior_temple,exterior_theater,exterior_town,exterior_train,exterior_truck,exterior_tunnel,exterior_warehouse,exterior_wetlands,interior,interior_airplane_cabin,interior_airplane_cockpit,interior_airport,interior_arena,interior_auditorium,interior_auto_repair_shop,interior_bar,interior_barn,interior_bathroom,interior_bedroom,interior_boat,interior_bus,interior_cafe,interior_cafeteria,interior_car,interior_cave,interior_classroom,interior_cloister,interior_closet,interior_command_center,interior_commercialkitchen,interior_conferenceroom,interior_courtroom,interior_crypt,interior_dancefloor,interior_diningroom,interior_dungeon,interior_elevator,interior_factory,interior_foyer,interior_gym,interior_hallway,interior_helicopter,interior_hospital,interior_kitchen,interior_livingroom,interior_lobby,interior_mall,interior_meditation,interior_nave,interior_office,interior_office_cubicle,interior_office_open,interior_prayer_hall,interior_prison,interior_pulpit,interior_restaurant,interior_spacecraft,interior_stage,interior_stairwell,interior_station_bus,interior_station_fire,interior_station_police,interior_station_subway,interior_station_train,interior_store,interior_store_aisle,interior_store_checkout,interior_study,interior_subway,interior_synagogue,interior_throneroom,interior_train,interior_truck,interior_warehouse,shot_angle_aerial,shot_angle_eyelevel,shot_angle_high,shot_angle_low,shot_focus_deep,shot_focus_out,shot_focus_shallow,shot_framing_closeup,shot_framing_extemelong,shot_framing_extremecloseup,shot_framing_long,shot_framing_medium,shot_level_level,shot_level_tilted,shot_lighting_hard,shot_lighting_key_high,shot_lighting_key_low,shot_lighting_silhouette,shot_lighting_soft,shot_subject_animal,shot_subject_location,shot_subject_object,shot_subject_person,shot_subject_person_body,shot_subject_person_face,shot_subject_person_feet,shot_subject_person_hands,shot_subject_text,shot_timeofday_day,shot_timeofday_night,shot_timeofday_twilight,shot_type_master,shot_type_overtheshoulder,shot_type_portrait,shot_type_twoshot,texture_banded,texture_blotchy,texture_braided,texture_bubbly,texture_bumpy,texture_chequered,texture_cobwebbed,texture_cracked,texture_crosshatched,texture_crystalline,texture_dotted,texture_fibrous,texture_flecked,texture_frilly,texture_gauzy,texture_grid,texture_grooved,texture_honeycombed,texture_interlaced,texture_knitted,texture_lacelike,texture_lined,texture_marbled,texture_matted,texture_meshed,texture_paisley,texture_perforated,texture_pitted,texture_pleated,texture_porous,texture_potholed,texture_scaly,texture_smeared,texture_spiralled,texture_sprinkled,texture_stained,texture_stratified,texture_striped,texture_studded,texture_swirly,texture_veined,texture_waffled,texture_woven,texture_wrinkled,texture_zigzagged\npath\/to\/image\/-5.jpg,1,0,0,0,-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,-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,-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,-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,-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,-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,-1,-1,-1,-1,-1,-1,-1,-1,-1\n<\/code><\/pre>\n\n<p>I am able to load the csv into a data set via <code>tf.data.experimental.make_csv_dataset<\/code> like so:<\/p>\n\n<pre><code># make our data set\nBATCH_SIZE = 32\nFILE_PATH = [\"filepath\"]\nCOLUMN_NAMES = [\"filepath\", \"color_key_blue\",\"color_key_green\",\"color_key_luma\",\"color_key_matte\",\"color_saturation_desaturated\",\"color_saturation_neutral\",\"color_saturation_pastel\",\"color_saturation_saturated\",\"color_theory_analagous\",\"color_theory_complementary\",\"color_theory_monochrome\",\"color_tones_blackwhite\",\"color_tones_cool\",\"color_tones_warm\",\"exterior\",\"exterior_airplane\",\"exterior_airport\",\"exterior_apartment\",\"exterior_auto_body\",\"exterior_beach\",\"exterior_bicycle\",\"exterior_boat\",\"exterior_bridge\",\"exterior_bus\",\"exterior_bus_stop\",\"exterior_canyon\",\"exterior_car\",\"exterior_castle\",\"exterior_cathedral\",\"exterior_cave\",\"exterior_church\",\"exterior_city\",\"exterior_desert\",\"exterior_farm\",\"exterior_forest\",\"exterior_glacier\",\"exterior_helicopter\",\"exterior_hospital\",\"exterior_house\",\"exterior_industrial\",\"exterior_lake\",\"exterior_library\",\"exterior_mall\",\"exterior_mansion\",\"exterior_monastery\",\"exterior_mosque\",\"exterior_motorcycle\",\"exterior_mountains\",\"exterior_ocean\",\"exterior_office\",\"exterior_palace\",\"exterior_park\",\"exterior_parkinglot\",\"exterior_pier\",\"exterior_plains\",\"exterior_playground\",\"exterior_polar\",\"exterior_port\",\"exterior_restaurant\",\"exterior_river\",\"exterior_road\",\"exterior_ruins\",\"exterior_school\",\"exterior_sidewalk\",\"exterior_sky\",\"exterior_skyscraper\",\"exterior_space\",\"exterior_spacecraft\",\"exterior_stadium\",\"exterior_station_gas\",\"exterior_station_subway\",\"exterior_station_train\",\"exterior_store\",\"exterior_suburb\",\"exterior_synagogue\",\"exterior_temple\",\"exterior_theater\",\"exterior_town\",\"exterior_train\",\"exterior_truck\",\"exterior_tunnel\",\"exterior_warehouse\",\"exterior_wetlands\",\"interior\",\"interior_airplane_cabin\",\"interior_airplane_cockpit\",\"interior_airport\",\"interior_arena\",\"interior_auditorium\",\"interior_auto_repair_shop\",\"interior_bar\",\"interior_barn\",\"interior_bathroom\",\"interior_bedroom\",\"interior_boat\",\"interior_bus\",\"interior_cafe\",\"interior_cafeteria\",\"interior_car\",\"interior_cave\",\"interior_classroom\",\"interior_cloister\",\"interior_closet\",\"interior_command_center\",\"interior_commercialkitchen\",\"interior_conferenceroom\",\"interior_courtroom\",\"interior_crypt\",\"interior_dancefloor\",\"interior_diningroom\",\"interior_dungeon\",\"interior_elevator\",\"interior_factory\",\"interior_foyer\",\"interior_gym\",\"interior_hallway\",\"interior_helicopter\",\"interior_hospital\",\"interior_kitchen\",\"interior_livingroom\",\"interior_lobby\",\"interior_mall\",\"interior_meditation\",\"interior_nave\",\"interior_office\",\"interior_office_cubicle\",\"interior_office_open\",\"interior_prayer_hall\",\"interior_prison\",\"interior_pulpit\",\"interior_restaurant\",\"interior_spacecraft\",\"interior_stage\",\"interior_stairwell\",\"interior_station_bus\",\"interior_station_fire\",\"interior_station_police\",\"interior_station_subway\",\"interior_station_train\",\"interior_store\",\"interior_store_aisle\",\"interior_store_checkout\",\"interior_study\",\"interior_subway\",\"interior_synagogue\",\"interior_throneroom\",\"interior_train\",\"interior_truck\",\"interior_warehouse\",\"shot_angle_aerial\",\"shot_angle_eyelevel\",\"shot_angle_high\",\"shot_angle_low\",\"shot_focus_deep\",\"shot_focus_out\",\"shot_focus_shallow\",\"shot_framing_closeup\",\"shot_framing_extemelong\",\"shot_framing_extremecloseup\",\"shot_framing_long\",\"shot_framing_medium\",\"shot_level_level\",\"shot_level_tilted\",\"shot_lighting_hard\",\"shot_lighting_key_high\",\"shot_lighting_key_low\",\"shot_lighting_silhouette\",\"shot_lighting_soft\",\"shot_subject_animal\",\"shot_subject_location\",\"shot_subject_object\",\"shot_subject_person\",\"shot_subject_person_body\",\"shot_subject_person_face\",\"shot_subject_person_feet\",\"shot_subject_person_hands\",\"shot_subject_text\",\"shot_timeofday_day\",\"shot_timeofday_night\",\"shot_timeofday_twilight\",\"shot_type_master\",\"shot_type_overtheshoulder\",\"shot_type_portrait\",\"shot_type_twoshot\",\"texture_banded\",\"texture_blotchy\",\"texture_braided\",\"texture_bubbly\",\"texture_bumpy\",\"texture_chequered\",\"texture_cobwebbed\",\"texture_cracked\",\"texture_crosshatched\",\"texture_crystalline\",\"texture_dotted\",\"texture_fibrous\",\"texture_flecked\",\"texture_frilly\",\"texture_gauzy\",\"texture_grid\",\"texture_grooved\",\"texture_honeycombed\",\"texture_interlaced\",\"texture_knitted\",\"texture_lacelike\",\"texture_lined\",\"texture_marbled\",\"texture_matted\",\"texture_meshed\",\"texture_paisley\",\"texture_perforated\",\"texture_pitted\",\"texture_pleated\",\"texture_porous\",\"texture_potholed\",\"texture_scaly\",\"texture_smeared\",\"texture_spiralled\",\"texture_sprinkled\",\"texture_stained\",\"texture_stratified\",\"texture_striped\",\"texture_studded\",\"texture_swirly\",\"texture_veined\",\"texture_waffled\",\"texture_woven\",\"texture_wrinkled\",\"texture_zigzagged\"]\nLABEL_NAMES = [\"color_key_blue\",\"color_key_green\",\"color_key_luma\",\"color_key_matte\",\"color_saturation_desaturated\",\"color_saturation_neutral\",\"color_saturation_pastel\",\"color_saturation_saturated\",\"color_theory_analagous\",\"color_theory_complementary\",\"color_theory_monochrome\",\"color_tones_blackwhite\",\"color_tones_cool\",\"color_tones_warm\",\"exterior\",\"exterior_airplane\",\"exterior_airport\",\"exterior_apartment\",\"exterior_auto_body\",\"exterior_beach\",\"exterior_bicycle\",\"exterior_boat\",\"exterior_bridge\",\"exterior_bus\",\"exterior_bus_stop\",\"exterior_canyon\",\"exterior_car\",\"exterior_castle\",\"exterior_cathedral\",\"exterior_cave\",\"exterior_church\",\"exterior_city\",\"exterior_desert\",\"exterior_farm\",\"exterior_forest\",\"exterior_glacier\",\"exterior_helicopter\",\"exterior_hospital\",\"exterior_house\",\"exterior_industrial\",\"exterior_lake\",\"exterior_library\",\"exterior_mall\",\"exterior_mansion\",\"exterior_monastery\",\"exterior_mosque\",\"exterior_motorcycle\",\"exterior_mountains\",\"exterior_ocean\",\"exterior_office\",\"exterior_palace\",\"exterior_park\",\"exterior_parkinglot\",\"exterior_pier\",\"exterior_plains\",\"exterior_playground\",\"exterior_polar\",\"exterior_port\",\"exterior_restaurant\",\"exterior_river\",\"exterior_road\",\"exterior_ruins\",\"exterior_school\",\"exterior_sidewalk\",\"exterior_sky\",\"exterior_skyscraper\",\"exterior_space\",\"exterior_spacecraft\",\"exterior_stadium\",\"exterior_station_gas\",\"exterior_station_subway\",\"exterior_station_train\",\"exterior_store\",\"exterior_suburb\",\"exterior_synagogue\",\"exterior_temple\",\"exterior_theater\",\"exterior_town\",\"exterior_train\",\"exterior_truck\",\"exterior_tunnel\",\"exterior_warehouse\",\"exterior_wetlands\",\"interior\",\"interior_airplane_cabin\",\"interior_airplane_cockpit\",\"interior_airport\",\"interior_arena\",\"interior_auditorium\",\"interior_auto_repair_shop\",\"interior_bar\",\"interior_barn\",\"interior_bathroom\",\"interior_bedroom\",\"interior_boat\",\"interior_bus\",\"interior_cafe\",\"interior_cafeteria\",\"interior_car\",\"interior_cave\",\"interior_classroom\",\"interior_cloister\",\"interior_closet\",\"interior_command_center\",\"interior_commercialkitchen\",\"interior_conferenceroom\",\"interior_courtroom\",\"interior_crypt\",\"interior_dancefloor\",\"interior_diningroom\",\"interior_dungeon\",\"interior_elevator\",\"interior_factory\",\"interior_foyer\",\"interior_gym\",\"interior_hallway\",\"interior_helicopter\",\"interior_hospital\",\"interior_kitchen\",\"interior_livingroom\",\"interior_lobby\",\"interior_mall\",\"interior_meditation\",\"interior_nave\",\"interior_office\",\"interior_office_cubicle\",\"interior_office_open\",\"interior_prayer_hall\",\"interior_prison\",\"interior_pulpit\",\"interior_restaurant\",\"interior_spacecraft\",\"interior_stage\",\"interior_stairwell\",\"interior_station_bus\",\"interior_station_fire\",\"interior_station_police\",\"interior_station_subway\",\"interior_station_train\",\"interior_store\",\"interior_store_aisle\",\"interior_store_checkout\",\"interior_study\",\"interior_subway\",\"interior_synagogue\",\"interior_throneroom\",\"interior_train\",\"interior_truck\",\"interior_warehouse\",\"shot_angle_aerial\",\"shot_angle_eyelevel\",\"shot_angle_high\",\"shot_angle_low\",\"shot_focus_deep\",\"shot_focus_out\",\"shot_focus_shallow\",\"shot_framing_closeup\",\"shot_framing_extemelong\",\"shot_framing_extremecloseup\",\"shot_framing_long\",\"shot_framing_medium\",\"shot_level_level\",\"shot_level_tilted\",\"shot_lighting_hard\",\"shot_lighting_key_high\",\"shot_lighting_key_low\",\"shot_lighting_silhouette\",\"shot_lighting_soft\",\"shot_subject_animal\",\"shot_subject_location\",\"shot_subject_object\",\"shot_subject_person\",\"shot_subject_person_body\",\"shot_subject_person_face\",\"shot_subject_person_feet\",\"shot_subject_person_hands\",\"shot_subject_text\",\"shot_timeofday_day\",\"shot_timeofday_night\",\"shot_timeofday_twilight\",\"shot_type_master\",\"shot_type_overtheshoulder\",\"shot_type_portrait\",\"shot_type_twoshot\",\"texture_banded\",\"texture_blotchy\",\"texture_braided\",\"texture_bubbly\",\"texture_bumpy\",\"texture_chequered\",\"texture_cobwebbed\",\"texture_cracked\",\"texture_crosshatched\",\"texture_crystalline\",\"texture_dotted\",\"texture_fibrous\",\"texture_flecked\",\"texture_frilly\",\"texture_gauzy\",\"texture_grid\",\"texture_grooved\",\"texture_honeycombed\",\"texture_interlaced\",\"texture_knitted\",\"texture_lacelike\",\"texture_lined\",\"texture_marbled\",\"texture_matted\",\"texture_meshed\",\"texture_paisley\",\"texture_perforated\",\"texture_pitted\",\"texture_pleated\",\"texture_porous\",\"texture_potholed\",\"texture_scaly\",\"texture_smeared\",\"texture_spiralled\",\"texture_sprinkled\",\"texture_stained\",\"texture_stratified\",\"texture_striped\",\"texture_studded\",\"texture_swirly\",\"texture_veined\",\"texture_waffled\",\"texture_woven\",\"texture_wrinkled\",\"texture_zigzagged\"]\n\ndef get_dataset(file_path, **kwargs):\n  dataset = tf.data.experimental.make_csv_dataset(\n      file_path,\n      batch_size=BATCH_SIZE, # Artificially small to make examples easier to show.\n      na_value=\"?\",\n      num_epochs=1,\n      column_names=COLUMN_NAMES,\n      ignore_errors=True, \n      shuffle=True, #TEMPORARY\n      **kwargs)\n  return dataset\n\ncsv_dataset = get_dataset(csv_file_path)\n<\/code><\/pre>\n\n<p>And then I transform the dataset to return a tuple of <code>string tensors<\/code> and <code>tensors<\/code> containing single dimensional vector of int32's containing my 1, 0 or -1's for each label:<\/p>\n\n<pre><code># we need to split out our data set to matching file paths and the labels as a sparse vector for each label,\n# containing 1, 0, -1 values for label concept present, label concept not present, dont know if present \n\n# for example, a CSV row like \n# FOLD_0\/color_key_blue\/1-5.jpg,1,0,0,0,-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,-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,-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,-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,-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,-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,-1,-1,-1,-1,-1,-1,-1,-1,-1\n# should return:\n# * a tensor containing a path like data_root + FOLD_0\/color_key_blue\/1-5.jpg, \n# * a tensor of values [1,0,0,0, -1 etc]\n\n# a function that returns a path, ordered dict of only the \n\nimport tensorflow.python.util\ndef split_csv_to_path_and_labels(csv_row_Dict):\n  print(\"Calling split_csv_to_path_and_label\")\n\n  filepath = csv_row_Dict.pop('filepath')\n  data_root_tensor = tf.constant(data_root)\n\n  filepath = tf.strings.join([data_root_tensor, filepath], separator='')\n\n  #make a new tensor with the values of the LABEL_NAMES keys but packed into a 0, len(LABEL_NAMES) array\n  labels = tf.stack(list(csv_row_Dict.values()), axis=1)\n\n  return filepath, labels\n\n#make a new data set from our csv by mapping every value to the above function\nsplit_dataset = csv_dataset.map(split_csv_to_path_and_labels)\n<\/code><\/pre>\n\n<p>Finally, I need to make a data set that loads images for each string tensor in my batch and returns a batch of image \/ label tuples. I define a function to load the images like so:<\/p>\n\n<pre><code>def load_and_preprocess_image(batch_of_paths, batch_of_labels):\n\n  # unpack our images from our batch\n  list_of_paths = tf.unstack(batch_of_paths, num=BATCH_SIZE)\n  list_of_images = []\n  for path in list_of_paths:\n      image = tf.read_file(path)\n      if image == None:\n        print(\"Unable to load image at path:\" + path )\n      image = preprocess_image(image)\n      list_of_images.append(image)\n\n\n  # repack our now loaded and processed images into a batch\n  batch_of_images = tf.stack(list_of_images, 0)\n\n  return batch_of_images,batch_of_labels\n<\/code><\/pre>\n\n<p>And then make a new data set by mapping the dataset to this function:<\/p>\n\n<pre><code>image_and_labels_ds = split_dataset.map(load_and_preprocess_image, num_parallel_calls=AUTOTUNE)\n<\/code><\/pre>\n\n<p>Since the above function has a hard coded <code>BATCH_SIZE<\/code> it works until the last batch is smaller than the batch size and TF errors out which the error above. <\/p>\n\n<p>How can I resolve this issue!?<\/p>\n\n<p>Ive looked at <a href=\"https:\/\/stackoverflow.com\/questions\/53422311\/dynamic-batch-size-in-tensorflow\">Dynamic batch size in tensorflow<\/a> but its unclear to me how this <\/p>\n","answers":[{"AnswerId":"57731042","CreationDate":"2019-08-30T17:34:47.003","ParentId":null,"OwnerUserId":"5510818","Title":null,"Body":"<p>I was able to resolve this issue by using <code>tf.map_fn<\/code> and mapping a function that processes a single string tensor and returns a preprocessed (normalized) image tensor:<\/p>\n\n<p>The main <code>tf.map_fn<\/code> call:<\/p>\n\n<pre><code>def load_and_preprocess_image_batch(batch_of_paths, batch_of_labels):\n    batch_of_images = tf.map_fn(load_and_preprocess_single_image_from_tensor, batch_of_paths, dtype=tf.float32)\n    return batch_of_images, batch_of_labels\n<\/code><\/pre>\n\n<p>And the declaration of <code>load_and_preprocess_single_image_from_tensor<\/code> and my preprocess code for anyone who cares:<\/p>\n\n<pre><code>IMG_SIZE = 224\n\n  def preprocess_image(image):\n  image = tf.image.decode_jpeg(image, channels=3)\n  image = tf.image.resize(image, [IMG_SIZE, IMG_SIZE])\n  image \/= 255.0  # normalize to [0,1] range\n\n  return image\n\ndef load_and_preprocess_single_image_from_tensor(path):\n   image = tf.read_file(path)\n   if image == None:\n    print(\"Unable to load image at path:\" + path )\n   return preprocess_image(image) \n<\/code><\/pre>\n\n<p>And finally, ensure my data set uses the <code>load_and_preprocess_image_batch<\/code> function:<\/p>\n\n<pre><code>image_and_labels_ds = split_dataset.map(load_and_preprocess_image_batch, num_parallel_calls=AUTOTUNE)\n<\/code><\/pre>\n"},{"AnswerId":"57720959","CreationDate":"2019-08-30T05:31:04.557","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<pre><code>batch_of_paths = tf.convert_to_tensor(np.random.rand(30))\n\n# SOLUTION 1\nlist_of_paths = tf.unstack(batch_of_paths, num=batch_of_paths.shape[0])\nprint(len(list_of_paths))\n\n# SOLUTION 2\nlist_of_paths = tf.unstack(batch_of_paths, axis=0)\nprint(len(list_of_paths))\n<\/code><\/pre>\n\n<p>It is preferable ot use SOLUTION 2 than SOLUTION 1 as it will make independent of extracting shape as well.<\/p>\n"}]}
+{"QuestionId":57719670,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-08-30T02:09:03.573","AcceptedAnswerId":null,"OwnerUserId":11758555.0,"Title":"why 'is_keras_tensor' function raising ValueError?","Body":"<p>I create a numpy array like np.array([1,2]) and try keras.backend.is_keras_tensor, it should return True or False according to official document, however, a ValueError raises.<\/p>\n\n<p>from keras import backend as K\nimport numpy as np\na=np.array([1,2])\nprint(K.is_keras_tensor(a))<\/p>\n\n<p>ValueError: Unexpectedly found an instance of type <code>&lt;class 'numpy.ndarray'&gt;<\/code>. Expected a symbolic tensor instance.<\/p>\n","answers":[]}
+{"QuestionId":57719697,"AnswerCount":0,"Tags":"<python><tensorflow><object-detection>","CreationDate":"2019-08-30T02:13:51.063","AcceptedAnswerId":null,"OwnerUserId":11997182.0,"Title":"Getting \"ValueError: Entries in boundaries must be strictly increasing.\" , during training using tensorflow object detection API","Body":"<p>I was trying to train my own dataset using tensorflow object detection API. So for this, at first I labeled my dataset using <a href=\"https:\/\/github.com\/tzutalin\/labelImg\" rel=\"nofollow noreferrer\">labelImg<\/a>. From that I got a xml file with the labeling information. Then I convert those data to tfrecord using a script. As I was trying to do transfer learning using faster rasnet inception v2 model, I changes the faster_rcnn_inception_v2_pets.config based on my information. Finally I run the model_main.py to train my data. But getting the following value error.<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"model_main.py\", line 109, in &lt;module&gt;\n    tf.app.run()\n  File \"\/home\/mahbubcseju\/anaconda3\/envs\/kashem\/lib\/python3.7\/site-packages\/tensorflow\/python\/platform\/app.py\", line 40, in run\n    _run(main=main, argv=argv, flags_parser=_parse_flags_tolerate_undef)\n  File \"\/home\/mahbubcseju\/anaconda3\/envs\/kashem\/lib\/python3.7\/site-packages\/absl\/app.py\", line 300, in run\n    _run_main(main, args)\n  File \"\/home\/mahbubcseju\/anaconda3\/envs\/kashem\/lib\/python3.7\/site-packages\/absl\/app.py\", line 251, in _run_main\n    sys.exit(main(argv))\n  File \"model_main.py\", line 105, in main\n    tf.estimator.train_and_evaluate(estimator, train_spec, eval_specs[0])\n  File \"\/home\/mahbubcseju\/anaconda3\/envs\/kashem\/lib\/python3.7\/site-packages\/tensorflow_estimator\/python\/estimator\/training.py\", line 473, in train_and_evaluate\n    return executor.run()\n  File \"\/home\/mahbubcseju\/anaconda3\/envs\/kashem\/lib\/python3.7\/site-packages\/tensorflow_estimator\/python\/estimator\/training.py\", line 613, in run\n    return self.run_local()\n  File \"\/home\/mahbubcseju\/anaconda3\/envs\/kashem\/lib\/python3.7\/site-packages\/tensorflow_estimator\/python\/estimator\/training.py\", line 714, in run_local\n    saving_listeners=saving_listeners)\n  File \"\/home\/mahbubcseju\/anaconda3\/envs\/kashem\/lib\/python3.7\/site-packages\/tensorflow_estimator\/python\/estimator\/estimator.py\", line 367, in train\n    loss = self._train_model(input_fn, hooks, saving_listeners)\n  File \"\/home\/mahbubcseju\/anaconda3\/envs\/kashem\/lib\/python3.7\/site-packages\/tensorflow_estimator\/python\/estimator\/estimator.py\", line 1158, in _train_model\n    return self._train_model_default(input_fn, hooks, saving_listeners)\n  File \"\/home\/mahbubcseju\/anaconda3\/envs\/kashem\/lib\/python3.7\/site-packages\/tensorflow_estimator\/python\/estimator\/estimator.py\", line 1188, in _train_model_default\n    features, labels, ModeKeys.TRAIN, self.config)\n  File \"\/home\/mahbubcseju\/anaconda3\/envs\/kashem\/lib\/python3.7\/site-packages\/tensorflow_estimator\/python\/estimator\/estimator.py\", line 1146, in _call_model_fn\n    model_fn_results = self._model_fn(features=features, **kwargs)\n  File \"\/home\/mahbubcseju\/Desktop\/Astro\/models\/research\/object_detection\/model_lib.py\", line 374, in model_fn\n    train_config.optimizer)\n  File \"\/home\/mahbubcseju\/Desktop\/Astro\/models\/research\/object_detection\/builders\/optimizer_builder.py\", line 56, in build\n    global_step=global_step)\n  File \"\/home\/mahbubcseju\/Desktop\/Astro\/models\/research\/object_detection\/builders\/optimizer_builder.py\", line 124, in _create_learning_rate\n    learning_rate_sequence, config.warmup)\n  File \"\/home\/mahbubcseju\/Desktop\/Astro\/models\/research\/object_detection\/utils\/learning_schedules.py\", line 187, in manual_stepping\n    raise ValueError('Entries in boundaries must be strictly increasing.')\nValueError: Entries in boundaries must be strictly increasing\n<\/code><\/pre>\n\n<p>.<\/p>\n\n<p>How to get rid of this error. Badly need some suggestion on this.<\/p>\n","answers":[]}
+{"QuestionId":57719800,"AnswerCount":0,"Tags":"<python-3.x><tensorflow>","CreationDate":"2019-08-30T02:31:34.877","AcceptedAnswerId":null,"OwnerUserId":11987146.0,"Title":"ValueError: Unknown loss function:triplet_loss","Body":"<p>Getting below error while trying to convert the model into tflite or .pb file:<\/p>\n\n<blockquote>\n  <p>ValueError: Unknown loss function:triplet_loss<\/p>\n<\/blockquote>\n\n<p>I have checked different solutions posted in StackOverflow and GitHub but none of them worked for my code.<\/p>\n\n<pre><code>#calculates triplet loss\ndef triplet_loss(y_true, y_pred, alpha = 0.2):\n\n    anchor, positive, negative = y_pred[0], y_pred[1], y_pred[2]\n\n    # triplet loss formula \n    pos_dist = tf.reduce_sum( tf.square(tf.subtract(anchor, positive)) )\n    neg_dist = tf.reduce_sum( tf.square(tf.subtract(anchor, negative)) )\n    basic_loss = pos_dist - neg_dist + alpha\n    loss = tf.maximum(basic_loss, 0.0)\n    return loss\n\n# load the model    \nmodel = load_model('facenet_model\/model.h5', custom_objects={'triplet_loss': triplet_loss})\n<\/code><\/pre>\n\n<p>I believe I have to make some changes in the triplet loss function to resolve value error.<\/p>\n","answers":[]}
+{"QuestionId":57719879,"AnswerCount":1,"Tags":"<numpy><tensorflow><matplotlib><keras><tfrecord>","CreationDate":"2019-08-30T02:45:47.100","AcceptedAnswerId":null,"OwnerUserId":11058721.0,"Title":"InvalidArgumentError: Expected image (JPEG, PNG, or GIF), got unknown format starting with 'B2.jpg'","Body":"<p>I have a dataset of <code>tfrecords<\/code> that I'm trying to parse.<br>\nI am using this code to parse it:<\/p>\n\n<pre><code>image_size = [224,224]\ndef read_tfrecord(tf_record):\n    features = {\n        \"filename\": tf.io.FixedLenFeature([], tf.string), # tf.string means bytestring\n        \"fun\": tf.io.FixedLenFeature([], tf.string),  \n        \"label\": tf.io.VarLenFeature(tf.int64),\n    }\n    tf_record = tf.parse_single_example(tf_record, features)\n\n    filename = tf.image.decode_jpeg(tf_record['filename'], channels=3)\n    filename = tf.cast(filename, tf.float32) \/ 255.0  # convert image to floats in [0, 1] range\n    filename = tf.reshape(filename, [*image_size, 3]) # explicit size will be needed for TPU\n\n    label = tf.cast(tf_record['label'],tf.float32)\n\n    return filename, label\n\ndef load_dataset(filenames):\n    option_no_order = tf.data.Options()\n    option_no_order.experimental_deterministic = False\n\n    dataset = tf.data.Dataset.from_tensor_slices(filenames)\n    dataset = dataset.with_options(option_no_order)\n    #dataset = tf.data.TFRecordDataset(filenames, num_parallel_reads=16)\n    dataset = dataset.interleave(tf.data.TFRecordDataset, cycle_length=32, num_parallel_calls=AUTO) # faster\n    dataset = dataset.map(read_tfrecord, num_parallel_calls=AUTO)\n    return dataset\n\ntrain_data=load_dataset(train_filenames)\nval_data=load_dataset(val_filenames)\ntest_data=load_dataset(test_filenames)\n<\/code><\/pre>\n\n<p>After running this code I get:\n<code>train_data<br>\n&lt;DatasetV1Adapter shapes: ((224, 224, 3), (?,)), types: (tf.float32, tf.float32)&gt;\n<\/code> \nI was trying to see the images in the dataset with:  <\/p>\n\n<pre><code>def display_9_images_from_dataset(dataset):\n    subplot=331\n    plt.figure(figsize=(13,13))\n    images, labels = dataset_to_numpy_util(dataset, 9)\n    for i, image in enumerate(images):\n        title = CLASSES[np.argmax(labels[i], axis=-1)]\n        subplot = display_one_flower(image, title, subplot)\n        if i &gt;= 8:\n            break;\n\n    plt.tight_layout()\n    plt.subplots_adjust(wspace=0.1, hspace=0.1)\n    plt.show()\n\n\n\ndef dataset_to_numpy_util(dataset, N):\n    dataset = dataset.batch(N)\n\n    if tf.executing_eagerly():\n    # In eager mode, iterate in the Datset directly.\n        for images, labels in dataset:\n            numpy_images = images.numpy()\n            numpy_labels = labels.numpy()\n            break;\n\n    else: # In non-eager mode, must get the TF note that \n        # yields the nextitem and run it in a tf.Session.\n        get_next_item = dataset.make_one_shot_iterator().get_next()\n        with tf.Session() as ses:\n            numpy_images, numpy_labels = ses.run(get_next_item)\n\n    return numpy_images, numpy_labels\n\n\ndisplay_9_images_from_dataset(train_data)\n<\/code><\/pre>\n\n<p>But I get the error:\n<code>InvalidArgumentError:  Expected image (JPEG, PNG, or GIF), got unknown format starting with 'B2.jpg'\n     [[{{node DecodeJpeg}}]]\n     [[IteratorGetNext_3]]<\/code>  <\/p>\n\n<p>I'm a bit confused, one because it says that the file is <code>jpg<\/code> format and it asks for <code>jpeg<\/code>, which from my understanding are the same.<br>\nAnd also because I'm not sure how to view the images or even know if I parsed it correctly.<\/p>\n","answers":[{"AnswerId":"58537798","CreationDate":"2019-10-24T08:59:16.560","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>Extensions \".jpg\" and \".jpeg\" are different in terms of the validation check done by the API which is consuming it. <\/p>\n\n<p>tf.image.decode_jpeg takes images with \".jpeg\" extensions. <\/p>\n\n<p>Try renaming your .jpg images with .jpeg extensions and it should start working. <\/p>\n"}]}
+{"QuestionId":57719942,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-30T02:58:09.700","AcceptedAnswerId":58536590.0,"OwnerUserId":11421418.0,"Title":"What's the problem of 'error code 3221225501' during running tensorflow codes","Body":"<p>I try to run the keras codes with tensorflow backend. But it caused \"ImportError: DLL load failed with error code 3221225501\". It seems not sloved by others..<\/p>\n\n<p>I just install python 3.7.4  in CPU enviroment and install the tensorflow with edition:tensorflow-1.13.1-cp37-cp37m-win_amd64.whl. And it is sucessfully installed. When I run my code, errors occours: ImportError: DLL load failed with error code 3221225501. I search the problem on internet, and lower down the tensorflow edition with 1.12.0 and download the files from github. It seems useless. So I come here for help. \nHope someone can provide a good suggestion, thanks in advance!<\/p>\n","answers":[{"AnswerId":"58536590","CreationDate":"2019-10-24T07:49:13.470","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>Please check if your CPU supports AVX Instructions. <\/p>\n\n<p>If it does not support AVX instructions, then you should use TF version 1.5 or earlier versions. <\/p>\n\n<p>From TF 1.6 and above, TF binaries uses AVX instructions. <\/p>\n\n<p>Please refer to below link for more information. <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/50430086\/issue-installing-tensorflow-not-a-cuda-cudnn-issue?answertab=votes#answer-50434293\">Issue installing Tensorflow -- not a CUDA\/CuDNN issue<\/a><\/p>\n"}]}
+{"QuestionId":57720676,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-30T04:52:13.810","AcceptedAnswerId":57837768.0,"OwnerUserId":7140567.0,"Title":"How can i implement orthogonality constraint on the columns of weight matrix in the hidden layers of a neural network?","Body":"<p>I want to train a neural network with orthogonality constraint on the columns of the weight matrix of the hidden layers i.e the weight matrix learned should have orthogonal columns. How to do this in keras ?<\/p>\n","answers":[{"AnswerId":"57837768","CreationDate":"2019-09-07T22:03:39.863","ParentId":null,"OwnerUserId":"11952444","Title":null,"Body":"<pre><code>class WeightsOrthogonalityConstraint(tf.keras.constraints.Constraint):\n    def __init__(self, encoding_dim, weightage = 1.0, axis = 0):\n        self.encoding_dim = encoding_dim\n        self.weightage = weightage\n        self.axis = axis\n\n    def weights_orthogonality(self, w):\n        if(self.axis==1):\n            w = tf.keras.backend.transpose(w)\n        if(self.encoding_dim &gt; 1):\n            m = tf.keras.backend.dot(tf.keras.backend.transpose(w), tf.Variable(w)) - tf.keras.backend.eye(self.encoding_dim)\n            return self.weightage * tf.keras.backend.sqrt(tf.keras.backend.sum(tf.keras.backend.square(m)))\n        else:\n            m = tf.keras.backend.sum(w ** 2) - 1.\n            return m\n\n    def __call__(self, w):\n        return self.weights_orthogonality(w)\n<\/code><\/pre>\n\n<p>Usage:<\/p>\n\n<pre><code>x = tf.keras.layers.Dense(2, 'relu', input_shape=(4,), kernel_regularizer=WeightsOrthogonalityConstraint(2))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57721324,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><tensorboard>","CreationDate":"2019-08-30T06:08:27.067","AcceptedAnswerId":57721344.0,"OwnerUserId":11768065.0,"Title":"Is it possible to specify scalar plot colours on Tensorboard?","Body":"<p>Can you specify colours for scalar plots on Tensorboard? Either in code or using the UI. It's a bit annoying to have multiple plots of the same colour that I want to compare. The only way around it is to flick on\/off the plot I'm looking at.<\/p>\n","answers":[{"AnswerId":"57721344","CreationDate":"2019-08-30T06:10:14.377","ParentId":null,"OwnerUserId":"9273406","Title":null,"Body":"<p>I don't think it's possible. Tensorboard doesn't have that option<\/p>\n\n<p>Have a look in here for more information:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorboard\/issues\/893\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorboard\/issues\/893<\/a><\/p>\n"}]}
+{"QuestionId":57721451,"AnswerCount":1,"Tags":"<python><numpy><pytorch>","CreationDate":"2019-08-30T06:18:20.993","AcceptedAnswerId":57736548.0,"OwnerUserId":11452683.0,"Title":"What is the PyTorch equivalent to Numpy's linalg.solve?","Body":"<p>I have looked it up and there are some posts, for example <a href=\"https:\/\/discuss.pytorch.org\/t\/solving-linear-equations-is-very-slow\/17774\" rel=\"nofollow noreferrer\">this one<\/a>, that suggest to use <code>torch.gesv<\/code> but I can't seem to find it in the PyTorch documentation.<\/p>\n","answers":[{"AnswerId":"57736548","CreationDate":"2019-08-31T08:55:04.637","ParentId":null,"OwnerUserId":"10935717","Title":null,"Body":"<p>Pytorch provides the function <code>torch.solve<\/code>, which behaves like <code>numpy.linalg.solve<\/code>. It will output the solution of the the linear system and the LU factorization that has been used to compute it. \nMore information and example codes <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.solve\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":57721524,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2019-08-30T06:25:41.817","AcceptedAnswerId":null,"OwnerUserId":3496060.0,"Title":"Overwrite checkpoint files from tf.train.Checkpoint.save","Body":"<p>In the example below, I subsequently call <code>root.save(path_to_checkpointfile)<\/code> and it auto-increments the filename each time. Each file is huge, and I want to just overwrite with the latest. I don't see any <code>kwarg<\/code> or property I can set to achieve this. Does anyone know of a way?<\/p>\n\n<pre><code>root = tf.train.Checkpoint(optimizer=optimizer,\n                               model=model,\n                               optimizer_step=tf.compat.v1.train.get_global_step())\n<\/code><\/pre>\n","answers":[{"AnswerId":"58536430","CreationDate":"2019-10-24T07:40:23.740","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>Using tf.train.CheckpointManager, you can specify how many checkpoints it needs to keep at any given point of time using the argument \"max_to_keep\". <\/p>\n\n<p>This keeps replacing the oldest checkpoint and saves the new one once it reaches the maximum number of checkpoints to be saved. <\/p>\n\n<p>Please refer below link for details:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/guide\/checkpoint\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/checkpoint<\/a><\/p>\n"}]}
+{"QuestionId":57721665,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-30T06:37:03.313","AcceptedAnswerId":null,"OwnerUserId":11997840.0,"Title":"Tensorflow frozen model only contains output placeholder variable","Body":"<p>I am trying to freeze a pre-trained model to then convert it in TF Lite and deploy it into an Android device. <\/p>\n\n<p>By inspecting the resulting .pb file with xxd I see that it only contains the placeholder output variable. The size of the .pb is a few Bytes.<\/p>\n\n<p>Why all the graph and variables are not included in the model? <\/p>\n\n<p>I used the code below derived from <a href=\"https:\/\/github.com\/sankit1\/cv-tricks.com\/tree\/master\/Tensorflow-tutorials\/freeze_model_and_deploy\" rel=\"nofollow noreferrer\">https:\/\/github.com\/sankit1\/cv-tricks.com\/tree\/master\/Tensorflow-tutorials\/freeze_model_and_deploy<\/a>. It works fine with other models but not with mine.<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python.framework import graph_util\nimport os,sys\n\npath = '.\/model-gaze\/'\noutput_node_names = \"pos\"\nmodel_name = 'model-23'\n\nsaver = tf.train.import_meta_graph(path+model_name+'.meta', clear_devices=True)\n\ngraph = tf.get_default_graph()\ninput_graph_def = graph.as_graph_def()\nsess = tf.Session()\nsaver.restore(sess, path+model_name)\noutput_graph_def = graph_util.convert_variables_to_constants(\n            sess, # The session is used to retrieve the weights\n            input_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) \noutput_graph=path+model_name+\".pb\"\nwith tf.gfile.GFile(output_graph, \"wb\") as f:\n    f.write(output_graph_def.SerializeToString())\n\nsess.close()\n<\/code><\/pre>\n\n<p>I would expect that all the weights and graph data are included inside the .pb  but cannot manage to get them there.<\/p>\n","answers":[{"AnswerId":"58536331","CreationDate":"2019-10-24T07:33:48.703","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>The link which you are following is the right procedure to freeze a tensorflow model. <\/p>\n\n<p>Freezing a model reduces the size of the model as it only saves the \"\"output_node_names\"\" which you give it to store. <\/p>\n\n<p>Please refer to the below link on the entire process. \n<a href=\"https:\/\/cv-tricks.com\/how-to\/freeze-tensorflow-models\/\" rel=\"nofollow noreferrer\">https:\/\/cv-tricks.com\/how-to\/freeze-tensorflow-models\/<\/a><\/p>\n\n<p>Here,  can you please elaborate on what \"\"pos\"\" is ? <\/p>\n\n<p>Also, here if you pass the prediction op as that is the required final op for predictions, it should work fine.<\/p>\n\n<p>And if this does not help, please share your model and the code from where you have saved the model, to further debug the issue.<\/p>\n"}]}
+{"QuestionId":57721676,"AnswerCount":0,"Tags":"<python><keras><neural-network>","CreationDate":"2019-08-30T06:37:38.117","AcceptedAnswerId":null,"OwnerUserId":11997825.0,"Title":"Keras Model with Meta Output Node without Loss or Gradient","Body":"<p>Is it possible to get a meta Node into a Model and train it, which is not connected to the output? So that I can access the layer later let's say via model.layers ?<\/p>\n\n<p>I am successfully combining two models with keras functional API.\nSo I use one pretrained Model as Feature Extractor and use for exampe layer x and y as features and Input for the second Model.<\/p>\n\n<p>What I can't get to work is following:\nAlso I am not using let's say layer z as extracted information in Model 2 i want to make the Tensor available in the model.<\/p>\n\n<p>I can just put it as additional Input to Model 2 and define the same Tensor as output. This yields several problems:<\/p>\n\n<ol>\n<li><p>I have to define a loss function for each output. The loss weight could be set to 0.<\/p><\/li>\n<li><p>I need to provide Trainingdata for each output for .fit() oder .fitGenerator(). I don't want to hack any code just to give empty training data or something else.<\/p><\/li>\n<\/ol>\n\n<pre class=\"lang-py prettyprint-override\"><code>def combine_models():\n\n   ## Model one will be usually be loaded with load_model()\n   in1 = Input()\n   x = Conv2D()(in1 )\n   y = Conv2D()(x)\n   z = Conv2d()(y)\n   model1 = Model(inputs=in1, outputs=z)  \n   ##model 2 uses part of own and model1 input\n   in2 = Input()\n   inX = Input() # with shape of x\n   inY = Input() # with shape o y\n   inZ = Input() # I know i have to declare z as Input for being available in model 2\n   conc = Concatenate()([in2,inX,inY])#just for example\n   conv= Conv2D()(conc)\n   model2 = Model(inputs=in2, outputs=conv)\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57721692,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><conv-neural-network>","CreationDate":"2019-08-30T06:38:51.990","AcceptedAnswerId":null,"OwnerUserId":11701704.0,"Title":"How to use a trained model on new inputs?","Body":"<p>I have created a CNN model that can be used to differentiate DOGS and CATS. During the training process my model was showing an training accuracy of 99% and testing accuracy of 81% by the end of 4\/25 epoch. \nIs this normal? or is there any problem that might occur after completion of all the epoch's?<\/p>\n\n<p>So I need to use this CNN model to my new inputs that do not belong to my training of test set. How do I use my model to predict some new photos? <\/p>\n\n<p>I have not used <code>classifier.save( )<\/code>, so after the training can I just use that command so that model gets saved? or do I have to recompile everything with <code>clssifier.save()<\/code> at the end?<\/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 Conv2D\nfrom keras.layers import MaxPooling2D\nfrom keras.layers import Flatten\nfrom keras.layers import Dense\n\n# Initialising the CNN\nclassifier = Sequential()\n\n# Step 1 - Convolution\nclassifier.add(Conv2D(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(Conv2D(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(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                                   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                         steps_per_epoch = 8000,\n                         epochs = 25,\n                         validation_data = test_set,\n                         validation_steps = 2000)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57722010","CreationDate":"2019-08-30T07:02:49.670","ParentId":null,"OwnerUserId":"7645295","Title":null,"Body":"<p>The model has a <code>save<\/code> method that exports the architecture and training configuration of the model to a file which can be later extracted and used. The documentation for the same can be found <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-save-a-keras-model\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>After importing the model, you can use the model on any data sets that you want to. About the accuracy of the model, it is possible to achieve the same. There is still a huge difference between the train and test accuracy so at the moment it is over-fitting the data. Also, try to randomize the data and train using them to make sure it is not an exceptional case.<\/p>\n"}]}
+{"QuestionId":57721797,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets><tensorflow2.0><data-augmentation>","CreationDate":"2019-08-30T06:46:06.917","AcceptedAnswerId":null,"OwnerUserId":6048112.0,"Title":"How to perform 10 Crop Image Augmentation at training time using Tensorflow 2.0 Dataset","Body":"<p>I am using Tensorflow Dataset API and reading data from TFRecord files. I can use the map function and use method like random_flip_left_right, random_crop for data augmentation.<\/p>\n\n<p>However when I am trying to replicate AlexNet paper I am facing an issue. I need to flip each image and then take 5 crops ( left, top, bottom, right &amp; middle).<\/p>\n\n<p>So the input dataset size will increase by 10 times. Is there anyway to do this using tensorflow dataset API? The map() function just returns the one image and I am not able to increase the number of images.<\/p>\n\n<p>Please see the code I have now.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>dataset = dataset.map(parse_image, num_parallel_calls=tf.data.experimental.AUTOTUNE) \\\n    .map(lambda image, label: (tf.image.random_flip_left_right(image), label), num_parallel_calls=tf.data.experimental.AUTOTUNE) \\\n    .map(lambda image, label: (tf.image.random_crop(image, size=[227, 227, 3]), label), num_parallel_calls=tf.data.experimental.AUTOTUNE) \\\n    .shuffle(buffer_size=1000) \\\n    .repeat() \\\n    .batch(256) \\\n    .prefetch(tf.data.experimental.AUTOTUNE)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57833337","CreationDate":"2019-09-07T11:41:40.150","ParentId":null,"OwnerUserId":"11478478","Title":null,"Body":"<pre><code>def tile_crop(img, label):\n    img_shape = tf.shape(img)\n    crop_left = lambda img: tf.image.random_crop(img[:,:img_shape[1]\/\/2,:], size=[227,227,3])\n    crop_top = lambda img: tf.image.random_crop(img[:img_shape[0]\/\/2,:,:], size=[227,227,3])\n    ...\n    img = tf.image.random_flip_left_right(img)\n    img = tf.stack([crop_left(img), crop_top(img),...], axis=0])\n    label = tf.reshape(label, [1,1]) #size: (,) -&gt; (1,1)\n    label = tf.tile(label, [5, 1]) #size: (1,1) -&gt; (5,1)\n    return img, label\ndt = parsed_dataset.map(tile_crop) #size: ((5,height,width,channels), (5, 1))\ndt = dt.unbatch() #size: ((height,width,channels), (1))\n<\/code><\/pre>\n\n<p>You can then use shuffle\/repeat\/batch\/prefetch as u like. Make sure that every cropped image has the same size.<\/p>\n"}]}
+{"QuestionId":57721804,"AnswerCount":0,"Tags":"<python><tensorflow><keras><lstm><recurrent-neural-network>","CreationDate":"2019-08-30T06:46:54.943","AcceptedAnswerId":null,"OwnerUserId":9961790.0,"Title":"setting sample_weight_mode=\"temporal\" doesn't seem to work","Body":"<p>I am working with LSTM in tensorflow 2.0 and I am trying to assign a weight to the training samples (I've already tried with class_weights dictionary but it complains that 3-d arrays are not supported for it. my array shape is (26000, 7, 1)).<\/p>\n\n<p>As suggested in the documentation, I am setting sample_weight_mode to \"temporal\" in compile, but after that when I try to fit the model I still get the error below:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-100-1bb94008291c&gt; in &lt;module&gt;\n      1 with tf.device(\"\/device:GPU:0\"):\n----&gt; 2     model.fit(X, y, epochs=500, batch_size=4096, verbose=1, validation_split=0.2, sample_weight=cw.reshape(26000,7,1))\n\nC:\\ProgramData\\Anaconda3\\envs\\thesis-gpu\\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, validation_freq, max_queue_size, workers, use_multiprocessing, **kwargs)\n    641         max_queue_size=max_queue_size,\n    642         workers=workers,\n--&gt; 643         use_multiprocessing=use_multiprocessing)\n    644 \n    645   def evaluate(self,\n\nC:\\ProgramData\\Anaconda3\\envs\\thesis-gpu\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_arrays.py in fit(self, model, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_freq, **kwargs)\n    630         steps=steps_per_epoch,\n    631         validation_split=validation_split,\n--&gt; 632         shuffle=shuffle)\n    633 \n    634     if validation_data:\n\nC:\\ProgramData\\Anaconda3\\envs\\thesis-gpu\\lib\\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, shuffle, extract_tensors_from_dataset)\n   2459           training_utils.standardize_weights(ref, sw, cw, mode)\n   2460           for (ref, sw, cw, mode) in zip(y, sample_weights, class_weights,\n-&gt; 2461                                          feed_sample_weight_modes)\n   2462       ]\n   2463       # Check that all arrays have the same length.\n\nC:\\ProgramData\\Anaconda3\\envs\\thesis-gpu\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py in &lt;listcomp&gt;(.0)\n   2458       sample_weights = [\n   2459           training_utils.standardize_weights(ref, sw, cw, mode)\n-&gt; 2460           for (ref, sw, cw, mode) in zip(y, sample_weights, class_weights,\n   2461                                          feed_sample_weight_modes)\n   2462       ]\n\nC:\\ProgramData\\Anaconda3\\envs\\thesis-gpu\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_utils.py in standardize_weights(y, sample_weight, class_weight, sample_weight_mode)\n    839     if sample_weight is not None and len(sample_weight.shape) != 1:\n    840       raise ValueError('Found a sample_weight array with shape ' +\n--&gt; 841                        str(sample_weight.shape) + '. '\n    842                        'In order to use timestep-wise sample weights, '\n    843                        'you should specify '\n\nValueError: Found a sample_weight array with shape (26000, 7, 1). In order to use timestep-wise sample weights, you should specify sample_weight_mode=\"temporal\" in compile(). If you just mean to use sample-wise weights, make sure your sample_weight array is 1D.\n<\/code><\/pre>\n\n<p>I have already tried changing the shape of the sample_weights passed in fit (also tried to flatten it), setting shuffle to False and removing the validation split without success.<\/p>\n\n<p>Here's the code I am using<\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nimport seaborn as sns\nimport matplotlib.pyplot as plt\nimport math\n\n# Import TensorFlow\nimport tensorflow as tf\nfrom tensorflow.keras.optimizers import RMSprop, Adam, SGD\nfrom tensorflow.keras.layers import SimpleRNN, LSTM, Dense\nfrom tensorflow.keras.models import Sequential\n\nfrom sklearn.utils import class_weight\n\nclass_weights = class_weight.compute_class_weight('balanced',\n                                                 np.unique(y_train),\n                                                 y_train)\n\ncw = class_weight.compute_sample_weight({False:1, True:50},#'balanced',\n                                                 #np.unique(y_train),\n                                                 y.flatten())\n\nmodel = Sequential()\nmodel.add(LSTM(160, return_sequences=True, dropout=0.05))\nmodel.add(LSTM(80, return_sequences=True, dropout=0.05, activation='relu'))\nmodel.add(LSTM(40, return_sequences=True, dropout=0.05, activation='relu'))\nmodel.add(LSTM(10, return_sequences=True, dropout=0.05))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(optimizer='adam', loss='binary_crossentropy', metrics=['acc', tf.keras.metrics.AUC()], sample_weight_mode=\"temporal\")\n\nwith tf.device(\"\/device:GPU:0\"):\n    model.fit(X, y, epochs=500, batch_size=4096, verbose=1, validation_split=0.2, sample_weight=cw.reshape(26000,7,1))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57721905,"AnswerCount":0,"Tags":"<keras><systemml>","CreationDate":"2019-08-30T06:55:33.023","AcceptedAnswerId":null,"OwnerUserId":11366367.0,"Title":"Traing a keras model on spark by uisng a 3d-shape spark dataframe","Body":"<p>I am training to write a distributed keras model in which both data and the model itself could be run on spark in parallel. The problem is that, I am working with 3-d shape image data as input, but I am not sure how to make dataframe from my 3d shape numpy array to train my keras model by system ml.\nI am using a pretrained xception keras models so my input need to be in the shape of (300,300,3)<\/p>\n\n<p>I am following the link<\/p>\n\n<pre><code>from sklearn import datasets, neighbors\nfrom pyspark.sql import DataFrame, SQLContext\nimport systemml as sml\nimport pandas as pd\nimport os, imp\nsqlCtx = SQLContext(sc)\ndigits = datasets.load_digits()\nX_digits = digits.data\ny_digits = digits.target + 1\nn_samples = len(X_digits)\n# Split the data into training\/testing sets and convert to PySpark DataFrame\nX_df = sqlCtx.createDataFrame(pd.DataFrame(X_digits[:.9 * n_samples]))\ny_df = sqlCtx.createDataFrame(pd.DataFrame(y_digits[:.9 * n_samples]))\nml = sml.MLContext(sc)\n# Get the path of MultiLogReg.dml\nscriptPath = os.path.join(imp.find_module(\"systemml\")[1], 'systemml-java', 'scripts', 'algorithms', 'MultiLogReg.dml')\nscript = sml.dml(scriptPath).input(X=X_df, Y_vec=y_df).output(\"B_out\")\nbeta = ml.execute(script).get('B_out').toNumPy()\n<\/code><\/pre>\n\n<p>but the problem when I use a 3d-shape numpy array, I cant convert it to spark dataframe, it ask me to give a 2d shape array.<\/p>\n\n<pre><code>l=np.array([[[1,2],[3,4],[5,6]],[[1,2],[3,4],[5,6]]])\nl.shape\n(2, 3, 2)\nX_df = sqlCtx.createDataFrame(pd.DataFrame(l))\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"C:\\Users\\KMOB\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\pandas\\core\\frame.py\", line 424, in __init__\n    copy=copy)\n  File \"C:\\Users\\KMOB\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\pandas\\core\\internals\\construction.py\", line 146, in init_ndarray\n    values = prep_ndarray(values, copy=copy)\n  File \"C:\\Users\\KMOB\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\pandas\\core\\internals\\construction.py\", line 249, in prep_ndarray\n    raise ValueError('Must pass 2-d input')\nValueError: Must pass 2-d input\n<\/code><\/pre>\n\n<p>Does anyone have any solution for this?<\/p>\n","answers":[]}
+{"QuestionId":57722072,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-hub>","CreationDate":"2019-08-30T07:06:54.563","AcceptedAnswerId":57723603.0,"OwnerUserId":11924886.0,"Title":"Colab Kernel Restarts Whenever Loading a Model From Tensorflow-hub","Body":"<p>I wanted to try out the embeddings provided in tensorflow-hub, the 'universal-sentence-encoder' to be specific. I tried the examples provided (<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>)\nand it worked fine. So I tried to do the same with 'multilingual' model but every time the multilingual model is loaded, the colab kernel fails and restarts. What is the problem and how can I get around this?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport tensorflow_hub as hub\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport os\nimport pandas as pd\nimport re\nimport seaborn as sns\nimport tf_sentencepiece\nimport sentencepiece\n\n# Import the Universal Sentence Encoder's TF Hub module\nembed = hub.Module(\"https:\/\/tfhub.dev\/google\/universal-sentence-encoder-multilingual\/1\") \/\/ This is where the kernel dies.\nprint(\"imported model\")\n# Compute a representation for each message, showing various lengths supported.\nword = \"\ucf54\ub07c\ub9ac\"\nsentence = \"\ub098\ub294 \ud55c\uad6d\uc5b4\ub85c \uc4f0\uc5ec\uc9c4 \ubb38\uc7a5\uc774\uc57c.\"\nparagraph = (\n    \"\ub3d9\ud574\ubb3c\uacfc \ubc31\ub450\uc0b0\uc774 \ub9c8\ub974\uace0 \ub2f3\ub3c4\ub85d. \"\n    \"\ud558\ub290\ub2d8\uc774 \ubcf4\uc6b0\ud558\uc0ac \uc6b0\ub9ac\ub098\ub77c \ub9cc\uc138~\")\nmessages = [word, sentence, paragraph]\n\n# Reduce logging output.\ntf.logging.set_verbosity(tf.logging.ERROR)\n\nwith tf.Session() as session:\n  session.run([tf.global_variables_initializer(), tf.tables_initializer()])\n  message_embeddings = session.run(embed(messages))\n\n  for i, message_embedding in enumerate(np.array(message_embeddings).tolist()):\n    print(\"Message: {}\".format(messages[i]))\n    print(\"Embedding size: {}\".format(len(message_embedding)))\n    message_embedding_snippet = \", \".join(\n        (str(x) for x in message_embedding[:3]))\n    print(\"Embedding: [{}, ...]\\n\".format(message_embedding_snippet))\n<\/code><\/pre>\n","answers":[{"AnswerId":"57723603","CreationDate":"2019-08-30T08:57:29.827","ParentId":null,"OwnerUserId":"6167591","Title":null,"Body":"<p>I had similar issues with the multilingual sentence encoder. I resolved it by specifying tensorflow version to 1.14.0 and tf-sentencepiece to 0.1.83, so before running your code in colab try:<\/p>\n\n<pre><code>!pip3 install tensorflow==1.14.0\n!pip3 install tensorflow-hub\n!pip3 install sentencepiece\n!pip3 install tf-sentencepiece==0.1.83\n<\/code><\/pre>\n\n<p>I was able to replicate your problem in colab and this solution loaded the model correctly:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/KU2TV.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/KU2TV.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>It seems to be a compatibility problem between sentencepiece and tensorflow, check for updates on this issue <a href=\"https:\/\/github.com\/tensorflow\/hub\/issues\/345\" rel=\"nofollow noreferrer\">here<\/a>.\nLet us know how it goes. Best of luck and I hope this helps.<\/p>\n\n<p>EDIT: <strong>If tensorflow version 1.14.0 does not work, change it to 1.13.1<\/strong>. This problem should be resolved once compatibility between tensorflow and sentencepiece is figured it out.<\/p>\n"}]}
+{"QuestionId":57722503,"AnswerCount":0,"Tags":"<python><tensorflow><split>","CreationDate":"2019-08-30T07:38:46.330","AcceptedAnswerId":null,"OwnerUserId":11995832.0,"Title":"Tensorflow: split tensor of unknown size into chunks of given size","Body":"<p>I would like to split a tensor into multiple sub-tensors of equal or near-equal size. Since I don't know the size of the tensor beforehand the splitting into even-sized chunks is not always guaranteed. However, <code>tf.split<\/code> seems to expect an even splitting and can therefore fail sometimes!<\/p>\n\n<p>In numpy, there is <code>np.split<\/code>, which also raises an exception if the array cannot be divided into evenly sized chunks. To avoid this issue, one can use <code>np.array_split<\/code>, which allows a last chunk of a different size. This is exactly the behavior I am looking for in tensorflow.<\/p>\n\n<p><strong>Summarized: What is the best way in tensorflow to split a tensor into multiple chunks if I don't now the size of the tensor but only the desired chunk size? Is there a similar function to <code>np.array_split<\/code> which I have not found yet?<\/strong><\/p>\n","answers":[]}
+{"QuestionId":57722852,"AnswerCount":3,"Tags":"<javascript><laravel><tensorflow>","CreationDate":"2019-08-30T08:03:42.627","AcceptedAnswerId":null,"OwnerUserId":11656696.0,"Title":"where to add the import command","Body":"<p>i am trying to implement a toxicity js in my laravel project \ni am using this library \n<a href=\"https:\/\/www.npmjs.com\/package\/@tensorflow-models\/toxicity\" rel=\"nofollow noreferrer\">https:\/\/www.npmjs.com\/package\/@tensorflow-models\/toxicity<\/a><\/p>\n\n<p>but when ever i run it i get this error in console<\/p>\n\n<pre><code>Uncaught ReferenceError: toxicity is not defined\n<\/code><\/pre>\n\n<p>my code looks like this atm <\/p>\n\n<p>View<\/p>\n\n<pre><code>&lt;div id=\"result\"&gt;&lt;\/div&gt;\n<\/code><\/pre>\n\n<p>myjs.js<\/p>\n\n<pre><code>\/\/import * as toxicity from '..\/..\/node_modules\/@tensorflow-models\/toxicity'; \/\/doesnt work\n\/\/import * as toxicity from '@tensorflow-models\/toxicity'; \/\/doesnt work\n$(document).ready(function() {\n    const threshold = 0.9;\n\n\/\/ Load the model. Users optionally pass in a threshold and an array of\n\/\/ labels to include.\n    toxicity.load(threshold).then(model =&gt; {\n        const sentences = ['you suck'];\n\n        model.classify(sentences).then(predictions =&gt; {\n            \/\/ `predictions` is an array of objects, one for each prediction head,\n            \/\/ that contains the raw probabilities for each input along with the\n            \/\/ final prediction in `match` (either `true` or `false`).\n            \/\/ If neither prediction exceeds the threshold, `match` is `null`.\n\n            console.log(predictions);\n            \/*\n            prints:\n            {\n              \"label\": \"identity_attack\",\n              \"results\": [{\n                \"probabilities\": [0.9659664034843445, 0.03403361141681671],\n                \"match\": false\n              }]\n            },\n            {\n              \"label\": \"insult\",\n              \"results\": [{\n                \"probabilities\": [0.08124706149101257, 0.9187529683113098],\n                \"match\": true\n              }]\n            },\n            ...\n             *\/\n        });\n    });\n});\n<\/code><\/pre>\n\n<p>i think the problem is that i do not know where should i include this line in laravel<\/p>\n\n<pre><code>import * as toxicity from '@tensorflow-models\/toxicity';\n<\/code><\/pre>\n\n<p>if i include this in the js file it gives an error that says:<\/p>\n\n<pre><code>Uncaught SyntaxError: Unexpected token *\n\n<\/code><\/pre>\n","answers":[{"AnswerId":"57723198","CreationDate":"2019-08-30T08:28:35.417","ParentId":null,"OwnerUserId":"11656696","Title":null,"Body":"<p>okay so i managed to fix this error by going into the \nbootstrap.js file and adding this<\/p>\n\n<pre><code>window.toxicity = require('@tensorflow-models\/toxicity');\n<\/code><\/pre>\n\n<p>and then running <\/p>\n\n<pre><code>npm run dev\n<\/code><\/pre>\n"},{"AnswerId":"57723017","CreationDate":"2019-08-30T08:14:07.727","ParentId":null,"OwnerUserId":"11995193","Title":null,"Body":"<p>To check, are you using something like Laravel Mix \/ webpack to compiler your code? So with Laravel Mix you'd run your file through the <code>js()<\/code> helper, see <a href=\"https:\/\/laravel.com\/docs\/5.8\/mix#working-with-scripts\" rel=\"nofollow noreferrer\">https:\/\/laravel.com\/docs\/5.8\/mix#working-with-scripts<\/a><\/p>\n\n<p>If you're using the default Laravel setup, you could add this into the file at <a href=\"https:\/\/github.com\/laravel\/laravel\/blob\/master\/resources\/js\/bootstrap.js\" rel=\"nofollow noreferrer\">https:\/\/github.com\/laravel\/laravel\/blob\/master\/resources\/js\/bootstrap.js<\/a><\/p>\n\n<p>If you look at the highlighted lines at <a href=\"https:\/\/github.com\/laravel\/laravel\/blob\/master\/resources\/js\/bootstrap.js#L32-L41\" rel=\"nofollow noreferrer\">https:\/\/github.com\/laravel\/laravel\/blob\/master\/resources\/js\/bootstrap.js#L32-L41<\/a> there is an example commented out where Laravel \"Echo\" library is imported in a similar manner to your own.<\/p>\n"},{"AnswerId":"57723187","CreationDate":"2019-08-30T08:27:27.647","ParentId":null,"OwnerUserId":"8981267","Title":null,"Body":"<p>Try adding it to the app.js file in resource\/js folder\nThen in your terminal run the command<\/p>\n\n<pre><code>npm run dev\n<\/code><\/pre>\n\n<p>then try using it again. Hope this helps<\/p>\n"}]}
+{"QuestionId":57723038,"AnswerCount":1,"Tags":"<python><tensorflow><conda><tensorboard>","CreationDate":"2019-08-30T08:15:05.003","AcceptedAnswerId":57727108.0,"OwnerUserId":6583936.0,"Title":"Update Tensorboard while keeping Tensorflow old with conda","Body":"<p>I have some legacy Keras\/Tensorflow code, which is unstable using latest Tensorflow versions (1.13+). It works just fine with previous versions. However i want to use Pytorch's Tensorboard support which requires it to be 1.14+. I've installed all Tensorflow-related packages to 1.10 and wanted to do just <code>conda install tensorboard=1.14<\/code> but it removes tensorflow=1.10 as a requirement. I know that these packages are generally independent. How to upgrade tensorboard while keeping tensorflow old? Preferably i would like to use a single conda environment.<\/p>\n","answers":[{"AnswerId":"57727108","CreationDate":"2019-08-30T12:50:56.747","ParentId":null,"OwnerUserId":"2160809","Title":null,"Body":"<p>as @jdehesa suggested in the comments, it's better to have a different conda environment for pytorch and then install just the tb there<br>\n<code>!pip install tb-nightly<\/code><\/p>\n"}]}
+{"QuestionId":57723209,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2019-08-30T08:29:00.453","AcceptedAnswerId":null,"OwnerUserId":11819060.0,"Title":"SDD Keras and Data Augmentation","Body":"<p>I'm using SDD by Keras to detect objects, and I'd like to do some data augmentation, but I have no ideia how.<\/p>\n\n<p>I'm following this rep:\n<a href=\"https:\/\/github.com\/pierluigiferrari\/ssd_keras\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pierluigiferrari\/ssd_keras<\/a><\/p>\n\n<p>And this notebook:<\/p>\n\n<pre><code>ssd300_training.ipynb   \n<\/code><\/pre>\n\n<p>I've read lots of thing about it, but I'm still confused. \nCan anyone help me?<\/p>\n\n<p>There are parameters to add? New code? What code?\nThanks <\/p>\n","answers":[]}
+{"QuestionId":57723264,"AnswerCount":1,"Tags":"<tensorflow><keras><keras-layer><tf.keras><keras-2>","CreationDate":"2019-08-30T08:32:56.007","AcceptedAnswerId":null,"OwnerUserId":996366.0,"Title":"How to freeze selected weights in keras or tf.keras?","Body":"<p>I am trying out something where I required to freeze some selected weights. Take this example <\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense,Input\n\nmodel = Sequential()\nmodel.add(Dense(4, input_shape=(4,),activation='relu'))\nmodel.add(Dense(3,name=\"hidden\",activation='relu'))\nmodel.add(Dense(2,activation='sigmoid'))\nmodel.compile(loss='mse', optimizer='adam')\n\nprint(model.layers[1].get_weights()[0])\n<\/code><\/pre>\n\n<p>This will print the input to hidden layer weights. <\/p>\n\n<pre><code># Weights input x hidden\n# Freeze 2Rx3C and 4Rx2C\n# 2Rx3C=0.14362943; 4Rx2C=-0.23868048\narray([[-0.05557871,  0.10941017, -0.59108734],\n       [ 0.37056673,  0.2968588 ,  0.14362943],\n       [-0.05471832, -0.21425706,  0.6455065 ],\n       [-0.7883829 , -0.23868048, -0.517396  ]], dtype=float32)\n<\/code><\/pre>\n\n<p>From this weight matrix I want to freeze values in <code>(2nd Row, 3rd Column)<\/code> and <code>(4th Row, 2nd Column)<\/code>, that is, <strong>0.14362943<\/strong> and <strong>-0.23868048<\/strong> respectively. I dont want to update these values on backprop. How can I freeze these selected weights?<\/p>\n","answers":[{"AnswerId":"58534659","CreationDate":"2019-10-24T05:26:58.450","ParentId":null,"OwnerUserId":"12198627","Title":null,"Body":"<p>You need to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/identity\" rel=\"nofollow noreferrer\">tf.identity<\/a>, it is used when you want to explicitly transport tensor between devices.<\/p>\n\n<pre><code>matrixVariable = tf.Variable(&lt;your matrix&gt;)\nmatrixVariableSliced = matrixVariable[&lt;sliced matrix&gt;] #take out your required weights\nmatrixVariable_stop = tf.stop_gradient(tf.identity(matrixVariableSliced)) \nmatrixVariable = tf.concat((matrixVariableSliced, matrixVariable_stop), axis=1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57723302,"AnswerCount":0,"Tags":"<keras><deep-learning><conv-neural-network><autoencoder><anomaly-detection>","CreationDate":"2019-08-30T08:35:32.573","AcceptedAnswerId":null,"OwnerUserId":11998215.0,"Title":"Deep Learning - classifying whether a given image is an anomaly\/novelty\/outlier?","Body":"<p>I have only \"Positive\" class image data from an experiment and the task is  to train them using a deep learning network. Anything which is slightly different from the above category should be classified as an anomaly\/outlier? [Taking negative class data from the same experiment is quite a difficult task]<\/p>\n\n<p>P.S : There will be no huge difference between the two classes of data. And the task is to use Deep learning  without having to define or give any features from the images. Also, only a limited number of images (around 300-400) for positive class data are available.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/mpkt6.png\" rel=\"nofollow noreferrer\">Positive class Image<\/a> <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/4oAG8.png\" rel=\"nofollow noreferrer\">A simple guess of what an anomalous image would look like<\/a><\/p>\n\n<p>I believe that this is entirely not an unsupervised learning because I know the labels of one of the classes. I tried using OneClassSVM for this without defining any features (Just tried - not much knowledge about SVM's). \nAlso, tried using Convolutional Autoencoders from this \"<a href=\"https:\/\/github.com\/artem-hryb\/cnn-anomaly-detection-keras\" rel=\"nofollow noreferrer\">https:\/\/github.com\/artem-hryb\/cnn-anomaly-detection-keras<\/a>\" link and Deep Autoencoders from \"<a href=\"https:\/\/github.com\/otenim\/AnomalyDetectionUsingAutoencoder\" rel=\"nofollow noreferrer\">https:\/\/github.com\/otenim\/AnomalyDetectionUsingAutoencoder<\/a>\". Some threshold value was set based on training the above models and then it was used to predict if the given image is an anomaly or not. The predictions were not correct. <\/p>\n\n<p>If anyone could point me in the right direction, it would be of great help. <\/p>\n","answers":[]}
+{"QuestionId":57723831,"AnswerCount":0,"Tags":"<android><tensorflow><assets><aar>","CreationDate":"2019-08-30T09:13:58.560","AcceptedAnswerId":null,"OwnerUserId":1498586.0,"Title":"Using TensorFlow Lite in aar library, Error in read files from aar assets","Body":"<p>I'm using Tensor-flow Lite on a example project to use it lib output as <strong>aar<\/strong> in another project. Every things fine in example app but when i build <strong>aar<\/strong> file and put on another project i have error on reading <code>tFlite<\/code> files from assets. I get this error:<\/p>\n\n<pre><code>This file can not be opened as a file descriptor; it is probably compressed\n        at android.content.res.AssetManager.openAssetFd(Native Method)\n        at android.content.res.AssetManager.openFd(AssetManager.java:380)\n<\/code><\/pre>\n\n<p>I googled and find out may it is not good to use assets on library but why it works on simple example( not build as aar one!) and if it should use as RAW folder there is some technical issue that Tensorflow needs Interpreter<\/p>\n\n<pre><code>Interpreter(@NonNull ByteBuffer byteBuffer, Interpreter.Options options)\n<\/code><\/pre>\n\n<p>I pass MappedByteBuffer with below method to <strong>Interpreter<\/strong>. But i could not change it to use RAW resources.<\/p>\n\n<pre><code> private MappedByteBuffer loadModelFile(AssetManager assetManager, String modelPath)\n      throws IOException {\n      AssetFileDescriptor fileDescriptor = assetManager.openFd(modelPath);\n      FileInputStream inputStream = new FileInputStream(fileDescriptor.getFileDescriptor());\n      FileChannel fileChannel = inputStream.getChannel();\n      long startOffset = fileDescriptor.getStartOffset();\n      long declaredLength = fileDescriptor.getDeclaredLength();\n      return fileChannel.map(FileChannel.MapMode.READ_ONLY, startOffset, declaredLength);\n }\n<\/code><\/pre>\n\n<p>I think i will have two apporache:<\/p>\n\n<ol>\n<li>Change configuration to load file properly from assets.<\/li>\n<li>Move from assets to RAW.<\/li>\n<\/ol>\n\n<p>Any body have any idea how to handle this?<\/p>\n","answers":[]}
+{"QuestionId":57724134,"AnswerCount":2,"Tags":"<python><pytorch>","CreationDate":"2019-08-30T09:32:38.867","AcceptedAnswerId":null,"OwnerUserId":11452683.0,"Title":"TypeError although same shape: if not (target.size() == input.size()): 'int' object is not callable","Body":"<p>This is the error message I get. In the first line, I output the shapes of <code>predicted<\/code> and  <code>target<\/code>. From my understanding, the error arises from those shapes not being the same but here they clearly are.<\/p>\n\n<pre><code>torch.Size([6890, 3]) torch.Size([6890, 3])\n\nTraceback (most recent call last):\n  File \"train.py\", line 251, in &lt;module&gt;\n    main()\n  File \"train.py\", line 230, in main\n    train(net, training_dataset, targets, device, criterion, optimizer, epoch, args.epochs)\n  File \"train.py\", line 101, in train\n    loss = criterion(predicted, target.detach().cpu().numpy())\n  File \"\/home\/hb119056\/.local\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/module.py\", line 493, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/home\/hb119056\/.local\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/loss.py\", line 443, in forward\n    return F.mse_loss(input, target, reduction=self.reduction)\n  File \"\/home\/hb119056\/.local\/lib\/python3.6\/site-packages\/torch\/nn\/functional.py\", line 2244, in mse_loss\n    if not (target.size() == input.size()):\nTypeError: 'int' object is not callable\n\n<\/code><\/pre>\n\n<p>I hope all the relevant context information is provided and if not, please let me know. Thanks for any suggestions!<\/p>\n\n<p>EDIT: This is the part of the code where this error occurs:<\/p>\n\n<pre><code>        target = torch.from_numpy(np.load(file_dir + '\/points\/points{:03}.npy'.format(i))).to(device)\n        rv = torch.zeros(12 * outputs.shape[0])\n\n        for j in [x for x in range(10) if x != i]:\n            source = torch.from_numpy(np.load(file_dir + '\/points\/points{:03}.npy'.format(j))).to(device)\n            rv = factor.ransac(source, target, prob, n_iter, tol, device) # some self-written RANSAC-like method\n\n        predicted = factor.predict(source, rv, outputs)\n        print(target.shape, predicted.shape)\n        loss = criterion(predicted, target.detach().cpu().numpy()) ## error occurs here\n<\/code><\/pre>\n\n<p><code>criterion<\/code> is <code>nn.MSELoss()<\/code>.<\/p>\n","answers":[{"AnswerId":"57724253","CreationDate":"2019-08-30T09:39:56.197","ParentId":null,"OwnerUserId":"7645295","Title":null,"Body":"<p>It probably means that you are trying to call a method when a property with the same name is available. If this is indeed the problem, the solution is easy. Simply change the method call into a property access.<\/p>\n\n<p>If you are comparing in the following way:<\/p>\n\n<pre><code>compare = (X.method() == Y.method())\n<\/code><\/pre>\n\n<p>Change it to:<\/p>\n\n<pre><code>compare = (X.method == Y.method)\n<\/code><\/pre>\n\n<p>If this does not answer your question, kindly share the code which you have used to compare the shapes.<\/p>\n"},{"AnswerId":"57724308","CreationDate":"2019-08-30T09:42:57.073","ParentId":null,"OwnerUserId":"7701908","Title":null,"Body":"<p>that's because your <code>target<\/code> is a <code>numpy<\/code> object<\/p>\n\n<pre><code>File \"train.py\", line 101, in train:\ntarget.detach().cpu().numpy()\n<\/code><\/pre>\n\n<p>in your code change the <code>target<\/code> type to numpy.<\/p>\n\n<p>TLDR  try change<\/p>\n\n<pre><code>loss = criterion(predicted, target.detach().cpu().numpy()) ## error occurs here\n<\/code><\/pre>\n\n<p>to <\/p>\n\n<pre><code>loss = criterion(predicted, target) ## error occurs here\n<\/code><\/pre>\n\n<hr>\n\n<p>for example:<\/p>\n\n<pre><code>In [6]: b = np.ones(3)                                                                                                                                                                                      \n\nIn [7]: b.size                                                                                                                                                                                              \nOut[7]: 3\n\nIn [8]: b.size()                                                                                                                                                                                            \n---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-8-365705555409&gt; in &lt;module&gt;\n----&gt; 1 b.size()\n\nTypeError: 'int' object is not callable\n<\/code><\/pre>\n"}]}
+{"QuestionId":57724146,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning>","CreationDate":"2019-08-30T09:33:26.680","AcceptedAnswerId":null,"OwnerUserId":4052624.0,"Title":"How to customize a Deep Learning model if the output is one-hot vectors?","Body":"<p>I am trying to build a Deep Learning model with TensorFlow and Keras. This is a sequential model for tasks of <strong>Single-Instance Multi-Label<\/strong>, which is a simplified version of <a href=\"https:\/\/www.aaai.org\/ocs\/index.php\/AAAI\/AAAI17\/paper\/download\/14747\/14350\" rel=\"nofollow noreferrer\"><strong>Multi-Instance Multi-Label<\/strong><\/a>. <\/p>\n\n<p>Concretely, the input of my model is an array of fixed length, so it can be represented as a vector like this:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/hDnh1.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/hDnh1.png\" alt=\"Input\"><\/a><\/p>\n\n<p>The output of my model is a sequence of letters, which are from a alphabet with a fixed size. For example, an alphabet of {A,B,C,D} with only 4 possible members. So I can use a one-hot vector to represent each letter in a sequence.<\/p>\n\n<p>The length of the sequences is variable, but for simplicity, I use a fixed length(equals to that of the longest sequence) to store all sequences. <\/p>\n\n<p>If the length of a sequence is shorter than the fixed length, the sequence is represented by one-hot vectors(equal to the seuqence's actual length) and zero vectors(equal to the remaining length). For example, <em>CADB<\/em> is represented by a 4 * 5 matrix like this:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/6uiDC.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/6uiDC.png\" alt=\"Output\"><\/a><\/p>\n\n<p>Please note: the first 4 columns of this matrix are one-hot vectors, each of which has one and only one 1 entry, and all other entries are 0s.\nBut the entries of the last column are all 0s, which can be seen as a zero padding because the sequence of letters is not long enough.<\/p>\n\n<p>So in one word, the input is a vector and the output is a matrix. <\/p>\n\n<p>Different from the link posted above, the output matrix should be seen as a whole. So one input vector is assigned to a whole matrix, not to a row or column of this matrix. <\/p>\n\n<p>My question is : how to customize my deep learning model for this special output, for example:<\/p>\n\n<ol>\n<li>What loss function and accuracy metric should I choose or design?<\/li>\n<li>Do I need to customize a special layer at the very end of my model?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"57725383","CreationDate":"2019-08-30T10:51:57.057","ParentId":null,"OwnerUserId":"9883438","Title":null,"Body":"<p>You should use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/backend\/softmax\" rel=\"nofollow noreferrer\"><code>softmax<\/code><\/a> activation on the output layer and have <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/backend\/categorical_crossentropy\" rel=\"nofollow noreferrer\"><code>categorical_crossentropy<\/code><\/a> as the loss function.<\/p>\n\n<p>However, as you can see in the links above, the problem is that these two functions by default are applied on the last axis (<code>axis=-1<\/code>), while in you situation it is the second last axis (the columns of the matrix) that is one-hot encoded.<\/p>\n\n<p>To use the right axis, one option is to define your own versions of these functions like so:<\/p>\n\n<pre><code>def softmax_columns(x):\n    return tf.keras.backend.softmax(x, axis=-2)\n\ndef categorical_crossentropy_columns(target, output):\n    return tf.keras.backend.categorical_crossentropy(target, output, axis=-2)\n<\/code><\/pre>\n\n<p>Then, you can use these like so:<\/p>\n\n<pre><code>model.add(SomeLayer(..., activation=softmax_columns, ...)) # output layer\n\nmodel.compile(loss=categorical_crossentropy_columns, ...)\n<\/code><\/pre>\n\n<p>One good alternative (in general, not only here) is to make use of <code>from_logits=True<\/code> in the <code>categorical_crossentropy<\/code> call. This effectively makes the softmax built-in into the loss function, so that your model itself does not need (in fact: <em>must not<\/em> have) the final softmax activation anymore. This not only saves work, but is also more numerically stable.<\/p>\n"}]}
+{"QuestionId":57724388,"AnswerCount":2,"Tags":"<tensorflow><tensorrt>","CreationDate":"2019-08-30T09:47:38.360","AcceptedAnswerId":57782554.0,"OwnerUserId":2467772.0,"Title":"Upsampling Tensor for TensorRT","Body":"<p>The tensorflow model is converted to TensorRT and Tensorflow's ResizeArea(upsample in the picture) need to implement plugin.\n<a href=\"https:\/\/i.stack.imgur.com\/BGTnw.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/BGTnw.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>So ResizeArea is implemented in CUDA.\nMy TensorRT input is NCHW format.<\/p>\n\n<pre><code>uff_path = model_to_uff(model_path)\nparser.register_input(ModelData.INPUT_NAME, (3, height, width), trt.UffInputOrder.NCHW)\nparser.register_output(ModelData.OUTPUT_NAME)\nparser.parse(uff_path, network)\n<\/code><\/pre>\n\n<p>So my CUDA code is implemented NCHW resampling.\nI like to make sure my resampling format is correct.<\/p>\n\n<p><strong>Method_1<\/strong><\/p>\n\n<p>NCHW resizearea(4 times upsampling) sample.<\/p>\n\n<pre><code>channel_1      channel_2        channel_3\n3,1,2,0,       0,4,3,1,         2,0,2,3,\n3,0,1,2,       0,1,2,1,         2,0,4,2,\n4,1,2,2,       1,3,2,4,         2,3,4,2,\n\nchannel_1                         channel_2                         channel_3  \n3,3,3,3,1,1,1,1,2,2,2,2,0,0,0,0,  0,0,0,0,4,4,4,4,3,3,3,3,1,1,1,1,  2,2,2,2,0,0,0,0,2,2,2,2,3,3,3,3,\n3,3,3,3,1,1,1,1,2,2,2,2,0,0,0,0,  0,0,0,0,4,4,4,4,3,3,3,3,1,1,1,1,  2,2,2,2,0,0,0,0,2,2,2,2,3,3,3,3,\n3,3,3,3,1,1,1,1,2,2,2,2,0,0,0,0,  0,0,0,0,4,4,4,4,3,3,3,3,1,1,1,1,  2,2,2,2,0,0,0,0,2,2,2,2,3,3,3,3,\n3,3,3,3,1,1,1,1,2,2,2,2,0,0,0,0,  0,0,0,0,4,4,4,4,3,3,3,3,1,1,1,1,  2,2,2,2,0,0,0,0,2,2,2,2,3,3,3,3,\n3,3,3,3,0,0,0,0,1,1,1,1,2,2,2,2,  0,0,0,0,1,1,1,1,2,2,2,2,1,1,1,1,  2,2,2,2,0,0,0,0,4,4,4,4,2,2,2,2,\n3,3,3,3,0,0,0,0,1,1,1,1,2,2,2,2,  0,0,0,0,1,1,1,1,2,2,2,2,1,1,1,1,  2,2,2,2,0,0,0,0,4,4,4,4,2,2,2,2,\n3,3,3,3,0,0,0,0,1,1,1,1,2,2,2,2,  0,0,0,0,1,1,1,1,2,2,2,2,1,1,1,1,  2,2,2,2,0,0,0,0,4,4,4,4,2,2,2,2,\n3,3,3,3,0,0,0,0,1,1,1,1,2,2,2,2,  0,0,0,0,1,1,1,1,2,2,2,2,1,1,1,1,  2,2,2,2,0,0,0,0,4,4,4,4,2,2,2,2,\n4,4,4,4,1,1,1,1,2,2,2,2,2,2,2,2,  1,1,1,1,3,3,3,3,2,2,2,2,4,4,4,4,  2,2,2,2,3,3,3,3,4,4,4,4,2,2,2,2,\n4,4,4,4,1,1,1,1,2,2,2,2,2,2,2,2,  1,1,1,1,3,3,3,3,2,2,2,2,4,4,4,4,  2,2,2,2,3,3,3,3,4,4,4,4,2,2,2,2,\n4,4,4,4,1,1,1,1,2,2,2,2,2,2,2,2,  1,1,1,1,3,3,3,3,2,2,2,2,4,4,4,4,  2,2,2,2,3,3,3,3,4,4,4,4,2,2,2,2,\n4,4,4,4,1,1,1,1,2,2,2,2,2,2,2,2,  1,1,1,1,3,3,3,3,2,2,2,2,4,4,4,4,  2,2,2,2,3,3,3,3,4,4,4,4,2,2,2,2,\n<\/code><\/pre>\n\n<p>Each pixel is 4 times upsampled(for example, first pixel 3 is upsample 4 times horizontally and vertically). That is considered as NCHW format upsampling.<\/p>\n\n<p><strong>Method_2<\/strong><\/p>\n\n<p>Another way implemented is considered upsampling in NHWC format.\n<a href=\"https:\/\/i.stack.imgur.com\/fiYRK.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/fiYRK.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/fyf10.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/fyf10.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>3-channel data (83,86,77) is upsampled horizontally and vertically.<\/p>\n\n<p>Is Method_1 is correct way of NCHW upsampling?<\/p>\n","answers":[{"AnswerId":"57771284","CreationDate":"2019-09-03T11:56:19.727","ParentId":null,"OwnerUserId":"3411080","Title":null,"Body":"<p>It seems method 1 is ok, since TensorRT expects CHW as said in <a href=\"https:\/\/docs.nvidia.com\/deeplearning\/sdk\/tensorrt-developer-guide\/index.html#import_tf_c\" rel=\"nofollow noreferrer\">the docs<\/a>, NHWC is a TF format, are you considering <code>AlignCorners<\/code> in your plugin layer? Also note that the resizing is nearest neighbor, in pytorch I used onnx-trt to do bilinear interpolation, which gave better results (in the case of segmentation, maybe for your case nn is ok). <\/p>\n"},{"AnswerId":"57782554","CreationDate":"2019-09-04T06:22:23.730","ParentId":null,"OwnerUserId":"2467772","Title":null,"Body":"<p>After taking some time. The issue was solved. TensorRT works in NCHW format. Tensorflow model is in NHWC format. So in my plugin, need to work upsampling in NCHW format but output needs to change as NHWC format so that it can interface to next Tensorflow operations.<\/p>\n"}]}
+{"QuestionId":57724559,"AnswerCount":2,"Tags":"<python><tensorflow><pip>","CreationDate":"2019-08-30T09:56:47.850","AcceptedAnswerId":57724746.0,"OwnerUserId":10783786.0,"Title":"Unable to install tensorflow-compression","Body":"<p>I get this error every time I try to install <strong>tensorflow-compression<\/strong><\/p>\n\n<pre><code>ERROR: Could not find a version that satisfies the requirement tensorflow-compression (from versions: none)\nERROR: No matching distribution found for tensorflow-compression\n<\/code><\/pre>\n\n<p>I'm using <strong>python 3.6.8<\/strong> and <strong>tensorflow-gpu 1.14.0<\/strong>.<\/p>\n\n<p>This package does exist and was recently updated:\n<a href=\"https:\/\/pypi.org\/project\/tensorflow-compression\/\" rel=\"nofollow noreferrer\">https:\/\/pypi.org\/project\/tensorflow-compression\/<\/a><\/p>\n\n<p>I've tried to install it in a <strong>python 2.7<\/strong> environment, but it didn't work.<\/p>\n\n<p>Also tried to specify the latest version <code>pip install tensorflow-compression==1.2<\/code>, but it also didn't work.<\/p>\n","answers":[{"AnswerId":"57724746","CreationDate":"2019-08-30T10:08:11.807","ParentId":null,"OwnerUserId":"2650249","Title":null,"Body":"<p>At the time of writing this, prebuilt wheels of <code>tensorflow-compression<\/code> exist only for Linux\/MacOS. Quoting the <a href=\"https:\/\/github.com\/tensorflow\/compression#installation\" rel=\"nofollow noreferrer\">readme<\/a>:<\/p>\n\n<blockquote>\n  <h3>Installation<\/h3>\n  \n  <p>Note: Precompiled packages are currently only provided for Linux (Python 2.7, 3.3-3.6) and Darwin\/Mac OS (Python 2.7, 3.7). To use these packages on Windows, consider using a <a href=\"https:\/\/www.tensorflow.org\/install\/docker\" rel=\"nofollow noreferrer\">TensorFlow Docker image<\/a> and installing tensorflow-compression using pip inside the Docker container.<\/p>\n<\/blockquote>\n"},{"AnswerId":"57725168","CreationDate":"2019-08-30T10:37:36.823","ParentId":null,"OwnerUserId":"11998821","Title":null,"Body":"<p>Windows is currently not supported. Please use one of TensorFlow's Docker images (<a href=\"https:\/\/www.tensorflow.org\/install\/docker\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/docker<\/a>) and pip install tensorflow_compression there.<\/p>\n"}]}
+{"QuestionId":57725029,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-08-30T10:28:23.407","AcceptedAnswerId":null,"OwnerUserId":11998575.0,"Title":"How to change input type(image) to list or array when using PyTorch tutorial code","Body":"<p>I have searched the code that uses list or array input data for training DQN code. But I have could not find any code. <\/p>\n\n<p>Currently, I reference the reinforcement learning tutorial(DQN) of Pytorch.\nHowever, this code uses image input data.<\/p>\n\n<p>I want to know how to change the image input data to list or array input data. <\/p>\n\n<p>(I need help to resolve my research that uses list input data. List input data shape is 1 by 9. )<\/p>\n","answers":[{"AnswerId":"57747144","CreationDate":"2019-09-01T15:09:50.663","ParentId":null,"OwnerUserId":"764285","Title":null,"Body":"<p>In PyTorch, we deal with tensors. Images, text, even sounds can be transformed to tensors and then PyTorch models can learn on the data.<\/p>\n\n<p>In PyTorch image classifier examples, you often see something like this, to transform images to tensors:<\/p>\n\n<pre><code>train_transform = transforms.Compose([\n                                      transforms.Resize(x),\n                                      ...\n                                      transforms.ToTensor()\n                                     ])\n<\/code><\/pre>\n\n<p>If your input is a numpy array x, you can convert it to a tensor like this:<\/p>\n\n<pre><code>torch.from_numpy(x)\n<\/code><\/pre>\n\n<p>You also have to pay attention to tensor dimensions, your input data needs to match what the model expects in the first layer.<\/p>\n"}]}
+{"QuestionId":57725172,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2019-08-30T10:37:50.563","AcceptedAnswerId":null,"OwnerUserId":4892495.0,"Title":"Iterating over a Dataset TF 2.0 with for loop","Body":"<p>This problem is about how to iterate over a TF Dataset given that <code>make_initializable_iterator()<\/code> is deprecated.<\/p>\n\n<p>I read a data set with the function below:<\/p>\n\n<pre><code>def read_dataset_new(filename, target='delay'):\n    ds = tf.data.TFRecordDataset(filename)\n    ds = ds.map(lambda buf: parse(buf, target=target))\n    ds = ds.batch(1)\n    return ds\n<\/code><\/pre>\n\n<p>Then I want to iterate over the data set. I have been using:\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset#make_initializable_iterator\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset#make_initializable_iterator<\/a><\/p>\n\n<pre><code>with tf.compat.v1.Session() as sess:\n    data_set = tfr_utils.read_dataset_new(self.tf_rcrds_fl_nm)\n    itrtr = data_set.make_initializable_iterator()\n    sess.run(itrtr.initializer)\n    features, label = itrtr.get_next()\n    features_keys = features.keys()\n...\n<\/code><\/pre>\n\n<p>But \"Warning: THIS FUNCTION IS DEPRECATED. It will be removed in a future version. Instructions for updating: Use for ... in dataset:....\"<\/p>\n\n<p>Apart from the deprecation warning, my code works as expected.<\/p>\n\n<p>Given the deprecation warning, though, I am now trying this:<\/p>\n\n<pre><code>with tf.compat.v1.Session() as sess:\n    data_set = tfr_utils.read_dataset_new(self.tf_rcrds_fl_nm)\n    for features, label in data_set:\n        features_keys = features.keys()\n        ...\n<\/code><\/pre>\n\n<p>But that does not work. I get:<\/p>\n\n<pre><code>self = &lt;tensorflow.python.client.session.Session object at 0x12f2e57d0&gt;\nfn = &lt;function BaseSession._do_run.&lt;locals&gt;._run_fn at 0x12f270440&gt;\nargs = ({}, [&lt;tensorflow.python.pywrap_tensorflow_internal.TF_Output; proxy of &lt;Swig Object of type 'TF_Output *' at 0x12f3f75a0&gt; &gt;], [], None, None)\nmessage = 'Resource AnonymousIterator\/AnonymousIterator0\/N10tensorflow4data16IteratorResourceE does not exist.\\n\\t [[node Iterat...tNext_1 (defined at \/demo-routenet\/tests\/unit\/test_tfrecord_utils.py:376) ]]'\nm = &lt;re.Match object; span=(102, 130), match='[[{{node IteratorGetNext_1}}'&gt;\n<\/code><\/pre>\n\n<p>The code samples I have been able to find all explicitly create an iterator, which is apparently not what one is supposed to do. I can't find an example of what one is supposed to do though.<\/p>\n\n<p>I suspect that something has not been initialised. So, I also tried:<\/p>\n\n<pre><code>sess.run(data_set)\n<\/code><\/pre>\n\n<p>But that didn't work either (nor do I have any reason to suppose it should have, but just so you all know what I tried).<\/p>\n\n<p>So, how does one use a Dataset in a for loop as the deprecation comment suggests please?<\/p>\n","answers":[{"AnswerId":"58158402","CreationDate":"2019-09-29T19:13:20.360","ParentId":null,"OwnerUserId":"10491729","Title":null,"Body":"<p>It is not very clear what you want to get at your output.\nIf you want to get the values of the dataset output you should execute eagerly.\nExample:<\/p>\n\n<pre><code>tf.compat.v1.enable_eager_execution()\n\ndef read_dataset_new(filename, target='delay'):\n    ds = tf.data.TFRecordDataset(filename)\n    ds = ds.map(lambda buf: parse(buf, target=target))\n    ds = ds.batch(1)\n    return ds\n# This should return your key values for each example.\nfor features, labels in read_dataset_new(self.tf_rcrds_fl_nm):\n    features_keys = features.keys()\n# This should return your tensor values if they supposed to be numeric.\nfor features, labels in read_dataset_new(self.tf_rcrds_fl_nm):\n    features_array = numpy.array(features)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57725588,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-30T11:06:20.220","AcceptedAnswerId":null,"OwnerUserId":1179925.0,"Title":"Tensorflow: is it possible to print number of sessions?","Body":"<p>Is it possible to have multiple sessions in tensorflow? Is it possible to print number of sessions in tensorflow?<\/p>\n\n<pre><code>def test_print_number_of_sessions():\n    sess1 = tf.Session()\n    sess2 = tf.Session()\n\n    \/\/print_number_of_sessions\n<\/code><\/pre>\n","answers":[{"AnswerId":"57728207","CreationDate":"2019-08-30T14:00:50.297","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You can have multiple sessions per graph, but there is no direct way of getting all the open sessions in a graph. The internal C data structure for the graph does have <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.14.0\/tensorflow\/c\/c_api_internal.h#L93-L107\" rel=\"nofollow noreferrer\">a collection with all the existing sessions<\/a>, but unfortunately the Python counter-part of that (the <code>._c_graph<\/code> attribute of <code>tf.Graph<\/code> objects) is just an opaque pointer without type information.<\/p>\n\n<p>One possible solution is to use your own session wrapper that tracks the open sessions per graph. This is one possible way to do that.<\/p>\n\n<pre><code>import tensorflow as tf\nimport collections\n\nclass TrackedSession(tf.Session):\n    _sessions = collections.defaultdict(list)\n    def __init__(self, target='', graph=None, config=None):\n        super(tf.Session, self).__init__(target=target, graph=graph, config=config)\n        TrackedSession._sessions[self.graph].append(self)\n    def close(self):\n        super(tf.Session, self).close()\n        TrackedSession._sessions[self.graph].remove(self)\n    @classmethod\n    def get_open_sessions(cls, g=None):\n        g = g or tf.get_default_graph()\n        return list(cls._sessions[g])\n\nprint(TrackedSession.get_open_sessions())\n# []\nsess1 = TrackedSession()\nprint(TrackedSession.get_open_sessions())\n# [&lt;__main__.TrackedSession object at 0x000001D75B0C77F0&gt;]\nsess2 = TrackedSession()\nprint(TrackedSession.get_open_sessions())\n# [&lt;__main__.TrackedSession object at 0x000001D75B0C77F0&gt;, &lt;__main__.TrackedSession object at 0x000001D75B0C7A58&gt;]\nsess1.close()\nprint(TrackedSession.get_open_sessions())\n# [&lt;__main__.TrackedSession object at 0x000001D75B0C7A58&gt;]\nsess2.close()\nprint(TrackedSession.get_open_sessions())\n# []\n<\/code><\/pre>\n\n<p>This however restricts you to use this custom session type, which may not be good enough depending on the scenario (e.g. if the session is opened by some external code, for example when you use Keras).<\/p>\n"}]}
+{"QuestionId":57725608,"AnswerCount":2,"Tags":"<python><pytorch><backpropagation>","CreationDate":"2019-08-30T11:08:18.080","AcceptedAnswerId":null,"OwnerUserId":11452683.0,"Title":"PyTorch: RuntimeError: Function MulBackward0 returned an invalid gradient at index 0 - expected type torch.cuda.FloatTensor but got torch.FloatTensor","Body":"<p>I don't understand what this error is telling me. In <a href=\"https:\/\/github.com\/NVIDIA\/flownet2-pytorch\/issues\/139\" rel=\"nofollow noreferrer\">a different post<\/a> the same problem was also addressed but there was no useful solution for this.<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"train.py\", line 252, in &lt;module&gt;\n    main()\n  File \"train.py\", line 231, in main\n    train(net, training_dataset, targets, device, criterion, optimizer, epoch, args.epochs)\n  File \"train.py\", line 103, in train\n    loss.backward()\n  File \"\/home\/hb119056\/.local\/lib\/python3.6\/site-packages\/torch\/tensor.py\", line 107, in backward\n    torch.autograd.backward(self, gradient, retain_graph, create_graph)\n  File \"\/home\/hb119056\/.local\/lib\/python3.6\/site-packages\/torch\/autograd\/__init__.py\", line 93, in backward\n    allow_unreachable=True)  # allow_unreachable flag\nRuntimeError: Function MulBackward0 returned an invalid gradient at index 0 - expected type torch.cuda.FloatTensor but got torch.FloatTensor\n<\/code><\/pre>\n\n<p>This is the corresponding segment from my code. <\/p>\n\n<pre><code>outputs = net(x, indices)\noutputs = outputs.transpose(0, 1)\nprob = F.normalize(outputs, p=1, dim=1).detach()\ntarget = torch.from_numpy(np.load(file_dir + '\/points\/points{:03}.npy'.format(i))).to(device)\nrv = torch.zeros(12 * outputs.shape[0])\n\nfor j in [x for x in range(10) if x != i]:\n    source = torch.from_numpy(np.load(file_dir + '\/points\/points{:03}.npy'.format(j))).to(device)\n    rv = factor.ransac(source, target, prob, n_iter, tol, device) # self-written\n\npredicted = factor.predict(source, rv, outputs, device) # self-written\nloss = criterion(predicted, target.type(torch.FloatTensor).to(device))\nloss.backward() # error occurs here\noptimizer.step()\n<\/code><\/pre>\n\n<p>Any help is greatly appreciated, thank you in advance!<\/p>\n","answers":[{"AnswerId":"57725985","CreationDate":"2019-08-30T11:34:24.687","ParentId":null,"OwnerUserId":"7701908","Title":null,"Body":"<p>change this line:<\/p>\n\n<pre><code>loss = criterion(predicted, target.type(torch.FloatTensor).to(device))\n<\/code><\/pre>\n\n<p>to <\/p>\n\n<pre><code>predicted = predicted.to(device)\ntarget=target.type(predicted.type()).to(predicted.device)\nloss = criterion(predicted, target)\n<\/code><\/pre>\n"},{"AnswerId":"57727013","CreationDate":"2019-08-30T12:44:54.243","ParentId":null,"OwnerUserId":"2160809","Title":null,"Body":"<p>try changing \n<code>loss = criterion(predicted, target.type(torch.FloatTensor).to(device))<\/code>\nto\n<code>loss = criterion(predicted, target.to(device).float())<\/code><\/p>\n"}]}
+{"QuestionId":57725839,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><conv-neural-network><mobilenet>","CreationDate":"2019-08-30T11:23:55.420","AcceptedAnswerId":null,"OwnerUserId":6276393.0,"Title":"Sequential' object has no attribute '_ckpt_saved_epoch' error when trying to save my model using callback on Keras","Body":"<p>I got this error when i tried to save my MobileNet model.<\/p>\n\n<pre><code>Traceback (most recent call last): File \"..\/src\/script.py\", line 150, in &lt;module&gt; callbacks=[cb_checkpointer, cb_early_stopper] File \"\/opt\/conda\/lib\/python3.6\/site-packages\/keras\/legacy\/interfaces.py\", line 91, in wrapper\n\nreturn func(*args, **kwargs) File \"\/opt\/conda\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\", line 1418, in fit_generator\n\ninitial_epoch=initial_epoch) File \"\/opt\/conda\/lib\/python3.6\/site-packages\/keras\/engine\/training_generator.py\", line 264, in fit_generator\n\ncallbacks.on_train_end() File \"\/opt\/conda\/lib\/python3.6\/site-packages\/keras\/callbacks.py\", line 142, in on_train_end callback.on_train_end(logs) File \"\/opt\/conda\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/callbacks.py\", line 940, in on_train_end\n\nif self.model._ckpt_saved_epoch is not None: AttributeError: 'Sequential' object has no attribute '_ckpt_saved_epoch'\n<\/code><\/pre>\n\n<p>I am using callbacks for saving:<\/p>\n\n<pre><code>filepath=\"weights-improvement-{epoch:02d}-{val_acc:.2f}.hdf5\"\ncb_early_stopper = EarlyStopping(monitor = 'val_loss', mode='min', verbose=1, patience = EARLY_STOP_PATIENCE)\ncb_checkpointer = ModelCheckpoint(filepath = filepath, monitor = 'val_loss', save_best_only = True, mode = 'auto')\n<\/code><\/pre>\n\n<p>My model code:<\/p>\n\n<pre><code>import numpy as np\nimport cv2\nimport os\n\n#from tensorflow.python.keras.models import Sequential\nfrom tensorflow.python.keras.layers import Dense\nfrom tensorflow.python.keras import optimizers\n\nfrom keras import regularizers\nfrom keras.regularizers import l2\nfrom keras.layers import Dropout\nfrom keras.applications.mobilenet import MobileNet\nfrom keras.layers import GlobalAveragePooling2D, Dense, Dropout, Flatten, BatchNormalization\nfrom keras.models import Sequential\nfrom keras.applications.resnet50 import preprocess_input\nfrom keras.preprocessing.image import ImageDataGenerator\nfrom tensorflow.python.keras.callbacks import EarlyStopping, ModelCheckpoint\n\n\n\n#print(os.listdir(\"testset\/\"))\n\n# Fixed for our classes\nNUM_CLASSES = 3\n\n# Fixed for color images\nCHANNELS = 3\n\nIMAGE_RESIZE = 224\nRESNET50_POOLING_AVERAGE = 'avg'\nDENSE_LAYER_ACTIVATION = 'softmax'\nOBJECTIVE_FUNCTION = 'categorical_crossentropy'\n\n# Common accuracy metric for all outputs, but can use different metrics for different output\nLOSS_METRICS = ['accuracy']\n\n# EARLY_STOP_PATIENCE must be &lt; NUM_EPOCHS\nNUM_EPOCHS = 100\nEARLY_STOP_PATIENCE = 50\n\n# These steps value should be proper FACTOR of no.-of-images in train &amp; valid folders respectively\n# Training images processed in each step would be no.-of-train-images \/ STEPS_PER_EPOCH_TRAINING\nSTEPS_PER_EPOCH_TRAINING = 27\nSTEPS_PER_EPOCH_VALIDATION = 11\n\n# These steps value should be proper FACTOR of no.-of-images in train &amp; valid folders respectively\n# NOTE that these BATCH* are for Keras ImageDataGenerator batching to fill epoch step input\nBATCH_SIZE_TRAINING = 8\nBATCH_SIZE_VALIDATION = 8\n\nbase_mobilenet_model = MobileNet(include_top = False, weights = None)\nmodel = Sequential()\nmodel.add(BatchNormalization(input_shape = [224,224,3]))\nmodel.add(base_mobilenet_model)\nmodel.add(BatchNormalization())\nmodel.add(GlobalAveragePooling2D())\nmodel.add(Dropout(0.5))\n\n# 2nd layer as Dense for 3-class classification,\nmodel.add(Dense(NUM_CLASSES, activation = DENSE_LAYER_ACTIVATION,activity_regularizer=regularizers.l2(0.01)))\n\n\n\nmodel.summary()\n\nmodel.compile(optimizer = 'adam', loss = OBJECTIVE_FUNCTION,  metrics =  LOSS_METRICS)\n\nimage_size = IMAGE_RESIZE\nshift = 0.2\n# preprocessing_function is applied on each image but only after re-sizing &amp; augmentation (resize =&gt; augment =&gt; pre-process)\n# Each of the keras.application.resnet* preprocess_input MOSTLY mean BATCH NORMALIZATION (applied on each batch) stabilize the inputs to nonlinear activation functions\n# Batch Normalization helps in faster convergence\n                                                    #  featurewise_center=True,  featurewise_std_normalization=True,\ndata_generator = ImageDataGenerator(preprocessing_function=preprocess_input,\n                                   width_shift_range=shift,\n                                   height_shift_range=shift,\n                                   horizontal_flip=True,\n                                   vertical_flip=True,\n                                   rotation_range=45,\n                                   brightness_range=[0.2,1.0],\n                                   zoom_range=[0.5,1.0]\n                                   )\n\n# flow_From_directory generates batches of augmented data (where augmentation can be color conversion, etc)\n# Both train &amp; valid folders must have NUM_CLASSES sub-folders\ntrain_generator = data_generator.flow_from_directory(\n        '\/kaggle\/input\/grade-dataset\/trainset\/',\n        target_size=(image_size, image_size),\n        batch_size=BATCH_SIZE_TRAINING,\n        class_mode='categorical')\n\nvalidation_generator = data_generator.flow_from_directory(\n        '\/kaggle\/input\/grade-dataset\/testset\/',\n        target_size=(image_size, image_size),\n        batch_size=BATCH_SIZE_VALIDATION,\n        class_mode='categorical')\n\n# Max number of steps that these generator will have opportunity to process their source content\n# len(train_generator) should be 'no. of available train images \/ BATCH_SIZE_TRAINING'\n# len(valid_generator) should be 'no. of available train images \/ BATCH_SIZE_VALIDATION'\n(BATCH_SIZE_TRAINING, len(train_generator), BATCH_SIZE_VALIDATION, len(validation_generator))\n\n# Early stopping &amp; checkpointing the best model in ..\/working dir &amp; restoring that as our model for prediction\nfilepath=\"weights-improvement-{epoch:02d}-{val_acc:.2f}.hdf5\"\ncb_early_stopper = EarlyStopping(monitor = 'val_loss', mode='min', verbose=1, patience = EARLY_STOP_PATIENCE)\ncb_checkpointer = ModelCheckpoint(filepath = filepath, monitor = 'val_loss', save_best_only = True, mode = 'auto')\n\nfit_history = model.fit_generator(\n        train_generator,\n        steps_per_epoch=STEPS_PER_EPOCH_TRAINING,\n        epochs = NUM_EPOCHS,\n        validation_data=validation_generator,\n        validation_steps=STEPS_PER_EPOCH_VALIDATION,\n        callbacks=[cb_checkpointer, cb_early_stopper]\n)\n\n<\/code><\/pre>\n","answers":[{"AnswerId":"59082439","CreationDate":"2019-11-28T05:46:10.193","ParentId":null,"OwnerUserId":"11530462","Title":null,"Body":"<p>Mentioning the Solution here, for the benefit of the Community.<\/p>\n\n<p>Replacing the code, <\/p>\n\n<p><code>from tensorflow.python.keras.callbacks import EarlyStopping, ModelCheckpoint<\/code> <\/p>\n\n<p>with the code, <\/p>\n\n<p><code>from keras.callbacks import EarlyStopping, ModelCheckpoint<\/code> <\/p>\n\n<p>has resolved the error.<\/p>\n"}]}
+{"QuestionId":57726064,"AnswerCount":1,"Tags":"<python><pytorch><cross-entropy>","CreationDate":"2019-08-30T11:39:25.463","AcceptedAnswerId":null,"OwnerUserId":10331422.0,"Title":"Binary cross entropy Vs categorical cross entropy with 2 classes","Body":"<p>When considering the problem of classifying an input to one of 2 classes, 99% of the examples I saw used a NN with a single output and sigmoid as their activation followed by a binary cross-entropy loss.  Another option that I thought of is having the last layer produce 2 outputs and use a categorical cross-entropy with C=2 classes, but I never saw it in any example.\nIs there any reason for that?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"57726399","CreationDate":"2019-08-30T12:02:58.907","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you are using <code>softmax<\/code> on top of the two output network you get an output that is <em>mathematically<\/em> equivalent to using a single output with <code>sigmoid<\/code> on top.<br>\nDo the math and you'll see.<\/p>\n\n<p>In practice, from my experience, if you look at the raw \"logits\" of the two outputs net (before <code>softmax<\/code>) you'll see that one is exactly the negative of the other. This is a result of the gradients pulling <em>exactly<\/em> in the opposite direction each neuron.<\/p>\n\n<p>Therefore, since both approaches are equivalent, the single output configuration has less parameters and requires less computations, thus it is more advantageous to use a single output with a sigmoid ob top.<\/p>\n"}]}
+{"QuestionId":57726175,"AnswerCount":0,"Tags":"<python><tensorflow><gpflow>","CreationDate":"2019-08-30T11:47:01.280","AcceptedAnswerId":null,"OwnerUserId":8082086.0,"Title":"GPFlow -- Speed and Performance Tips","Body":"<p>I just recently started trying out the <code>gpflow<\/code> package to build a respective GP model for a given training set.\nThe data I am dealing with has a very high dimensionality:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>&gt;&gt;&gt; X.shape; Y.shape\n(3489, 416)\n(3489, 1)\n<\/code><\/pre>\n\n<p>As a starter, I was following the tutorials provided to reproduce a basic model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>k = gpflow.kernels.RBF(input_dim=X.shape[1], ARD=True) + gpflow.kernels.Linear(input_dim=X.shape[1], ARD=True) + gpflow.kernels.Matern12(input_dim=X.shape[1], ARD=True)\nm = gpflow.models.GPR(X, Y, kern=k, mean_function=None)\nopt = gpflow.train.ScipyOptimizer()\nopt.minimize(m)\n<\/code><\/pre>\n\n<p>The time to optimize this model is roughly one hour, when running on a CPU, while running on a GPU seems to be ~4 times slower (Mobile Workstation: Quadro P600, Xeon E3-1200 v5).<\/p>\n\n<p>My question is whether this time is what is to be expected from a data set with this dimension and sample size, or is there something that can be done to reduce the optimization time? Is there any particular reason the optimization does not benefit from GPU usage at all? \nI understand combining multiple kernels and setting <code>ADR=True<\/code> might be a big bottleneck. Other than that, is there something that can be tweaked?<\/p>\n\n<p>Ideally, I would like to retrain the model consecutively, once a number of new reference points has been added. <\/p>\n\n<p>Thank you very much!<\/p>\n\n<p><em>Edit:<\/em> After some further reading I found out that GP's in general scale with order N^3, where N is the number of samples. Reducing the sample size helps a lot with speed. The GPU is still a lot slower though.<\/p>\n","answers":[]}
+{"QuestionId":57726317,"AnswerCount":0,"Tags":"<machine-learning><neural-network><deep-learning><pytorch><attention-model>","CreationDate":"2019-08-30T11:58:20.170","AcceptedAnswerId":null,"OwnerUserId":4517807.0,"Title":"Pytorch Implementing Simple Attention using Dummy data","Body":"<p>Hi I am trying to implement  simple\/General attention in Pytorch , \nSo far the model seems to working , but what i am intersted in doing is getting the attention weights , so that i can visualize it .<\/p>\n\n<p>Here's what i am doing , creating a dummy sequence data , the 5th sequence is set as the target , so all the model needs to do is to understand that the 5th sequence in the data is the target and give a higher attention weight to the 5th sequence .<\/p>\n\n<p>Strangely when i try to plot my attention weight the highest weight is observed in the last sequence! .<\/p>\n\n<p>So i would like to know two things .<\/p>\n\n<p>1) Given the code , is my implementation of attention Correct ? \n2) If its right,  why is the attention weights shifted to the last sequence? <\/p>\n\n<p>Here's the entire code that i am playing with (Hope its readable and understandable)<\/p>\n\n<pre><code>import torch\nimport torch.nn as nn\nimport matplotlib.pyplot as plt\nimport pandas as pd\nimport numpy as np\nimport math\nimport random\nimport matplotlib.pyplot as plt\nimport torch.nn.functional as F\nimport pandas as pd\n\nINPUT_DIMS = 10\nTIME_STEPS = 10\nATTENTION_COL = 5\ndef get_data_recurrent(n, time_steps, input_dim, attention_column=10):\n    \"\"\"\n    Data generation. x is purely random except that it's first value equals the target y.\n    In practice, the network should learn that the target = x[attention_column].\n    Therefore, most of its attention should be focused on the value addressed by attention_column.\n    :param n: the number of samples to retrieve.\n    :param time_steps: the number of time steps of your series.\n    :param input_dim: the number of dimensions of each element in the series.\n    :param attention_column: the column linked to the target. Everything else is purely random.\n    :return: x: model inputs, y: model targets\n    \"\"\"\n    x = np.random.standard_normal(size=(n, time_steps, input_dim))\n    y = np.random.randint(low=0, high=2, size=(n, 1))\n    x[:, attention_column, :] = np.tile(y[:], (1, input_dim))\n    return x, y\n\nX_train , y_train = get_data_recurrent(300000 , \n                                       input_dim=INPUT_DIMS ,\n                                       time_steps=TIME_STEPS , \n                                       attention_column=ATTENTION_COL)\n\nclass simple_lstm( nn.Module):\n    def __init__(self , input_size , hidden_size , output_units):\n        super(simple_lstm, self).__init__()\n        self.lstm = nn.LSTM(input_size=10,hidden_size=hidden_size , \n                            batch_first = True)\n        self.dense1 = nn.Linear(hidden_size , 1)\n\n    def forward(self , x):\n        out , (hn,cn) = self.lstm(x)\n        hn = hn.squeeze(0)\n        hidden_state = hn\n        attention_scores = torch.bmm(out,\n                                     hidden_state.unsqueeze(2)).squeeze(2)\n        soft_attention_weights = F.softmax(attention_scores, 1) \n        attention_output = torch.bmm(out.transpose(1, 2), \n                                     soft_attention_weights.unsqueeze(2)).squeeze(2)\n        out = self.dense1(attention_output)\n        out = out\n        return out , soft_attention_weights\n\n#Model Training\ntorch_model = simple_lstm(input_size=INPUT_DIMS ,\n                          hidden_size=32, \n                          output_units=1)\n\noptimiser = torch.optim.Adam(params = torch_model.parameters())\ncriterion = nn.MSELoss()\ntorch_train = torch.utils.data.TensorDataset(\n                        torch.tensor(X_train , dtype = torch.float) , \n                        torch.tensor(y_train , dtype = torch.float))\ntorch_train_loader = torch.utils.data.DataLoader(torch_train , batch_size=64)\n\nnum_epochs = 1\nsaw = []\nfor epoch in range(num_epochs):\n    for i , (X_tr ,y_tr) in enumerate(torch_train_loader):\n        optimiser.zero_grad()\n        output , att = torch_model(X_tr)\n        saw.append(att.data.numpy().mean(axis=0))\n        loss = criterion(output , y_tr)\n        loss.backward()\n        optimiser.step()\n    print('Epoch [{}\/{}], Loss: {:.4f}'.format(epoch + 1,\n                                               num_epochs, loss.item()))\n\n#PLOT ATTENTION  Weights\natt = []\nfor i in range(1,300):\n    X_test , y_test  = get_data_recurrent(1,input_dim=INPUT_DIMS ,\n                                          time_steps=TIME_STEPS ,\n                                          attention_column=ATTENTION_COL)\n    preds , attention = torch_model(torch.tensor(X_test , dtype = torch.float))\n    att.append(attention.data.numpy())\narr = np.mean(np.array(att) , axis=0)\npd.DataFrame(arr.squeeze(0), columns=['attention (%)']).plot(kind='bar',\n                                                             title='Attention Mechanism as '\n                                                             'a function of input')\n<\/code><\/pre>\n\n<p>and here the attentions weight that was plotted <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/3eQtk.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3eQtk.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57726538,"AnswerCount":0,"Tags":"<tensorflow><neural-network><pytorch>","CreationDate":"2019-08-30T12:12:29.583","AcceptedAnswerId":null,"OwnerUserId":10986701.0,"Title":"Is there an implementation of forward mapping\/warping in a neural network framework?","Body":"<h1>Introduction<\/h1>\n\n<p>I want to train a neural network and require a differentiable implementation of a forward mapping to warp a source image into a destination image, see e.g. <a href=\"https:\/\/www.cs.princeton.edu\/courses\/archive\/fall00\/cs426\/lectures\/warp\/warp.pdf\" rel=\"nofollow noreferrer\">https:\/\/www.cs.princeton.edu\/courses\/archive\/fall00\/cs426\/lectures\/warp\/warp.pdf<\/a>. <\/p>\n\n<p>What I already found is backward mapping via pytorch (grid_sample), in which every pixel in destination image has only a single connection to one particular pixel in source image (+ neighbours).<\/p>\n\n<h1>Code<\/h1>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nimport torch.nn.functional as F\nimport numpy as np\nimport matplotlib.pyplot as plt\n\n# create artificial image\nim = np.zeros((9,9)) \nim[4,4] = 1\n\n# warping images\nu = np.zeros((9,9))\nv = np.zeros((9,9))\n\nu[4,3] += 1\n\n\n# shift to pytorch\nh,w = np.shape(im)\n\nim_t = torch.from_numpy(im).float().view((1,1,h,w)).cpu()\nu_t = torch.from_numpy(u).float().view((1,h,w,1)).cpu()\nv_t = torch.from_numpy(v).float().view((1,h,w,1)).cpu()\n\n\nn, _, h, w = im_t.size()\nbatchSize, _, _, nmbSamples = im_t.size()\n\n# warp source image\nU = torch.arange(start=0, end=w).expand(h, w).float().view((n, h, w, 1)).cpu()\nV = torch.arange(start=0, end=h).expand(w, h).t().float().view((n, h, w, 1)).cpu()\n\nU_u = (U + u_t) \/ ((w-1) \/ 2) - 1\nV_v = (V + v_t) \/ ((h-1) \/ 2) - 1\n\ngrid = torch.cat((U_u, V_v), dim=3).float().cpu()\n\nwarped_t = F.grid_sample(im_t, grid)\n\n# plot routine\nplt.figure()\nplt.subplot(121)\nplt.imshow(im_t.data.cpu().numpy()[0,0,:,:])\nplt.grid()\n\nplt.subplot(122)\nplt.imshow(warped_t.data.cpu().numpy()[0,0,:,:])\nplt.grid()\n\nplt.show(block=False)\nplt.draw()\n<\/code><\/pre>\n\n<h1>Problem<\/h1>\n\n<p>What I need, is to render the destination image from different, non neighbouring pixels in source image.<\/p>\n\n<p>Help appreciated.\nRegards<\/p>\n","answers":[]}
+{"QuestionId":57726750,"AnswerCount":0,"Tags":"<tensorflow><keras><neural-network><deep-learning>","CreationDate":"2019-08-30T12:26:31.487","AcceptedAnswerId":null,"OwnerUserId":10146664.0,"Title":"How to change the output of specific training examples by condition in a Keras model","Body":"<p>So.... I have a matrix as input to a keras model. The training matrix has 2721 examples and each example is of shape (84, 55, 26). So it's of shape (2721, 84, 55, 26) and the validation is (681, 84, 55, 26). The right way to look at it is that each training example is 26 matrices of shape (84,55)<\/p>\n\n<p>I want to create a keras model that takes these kind of matrices and return a single number. the loss will be MSE. So naturally, when creating an input for the model, I take as input a keras tensor of shape (?, 84, 55, 26). Here's the thing: the model is supposed to go something like this: input --> turns into a 26 intermediate vectors --> turns into a single output (where --> is various layers). <\/p>\n\n<p>But, if a single example is of shape (84, 55, 26), then sometimes one of the 26 vectors of shape (84, 55) is all zeros. When that happens I want the intermediate vector of that specific matrix to always be all zeros. How can I check when defining the model the specific places where everything is zeros When the shape has (?) training examples. I need to only take the specific examples and the specific indices out of the 26 available that are all zeros and turn their intermediate vector to all zeros. <\/p>\n\n<p>I'll add code and a bad drawing. In the code, third_sense_layer_results contains 26 vectors of size (?, 1000) meaning that for each training example we have 26 intermediate vectors of size 1000. <\/p>\n\n<p>Thanks!! <\/p>\n\n<pre><code>inputs = Input(input_shape)\nplayers_slices = []\nfor k in range(input_shape[-1]):\n    sliced_tensor = Lambda(lambda x: x[:, :, :, k], output_shape=input_shape[:-1])(inputs)\n    players_slices.append(Flatten()(sliced_tensor))\nfirst_dense_layer = Dense(5000, activation='relu') #10000\n\nfirst_dense_results = []\nfor k in range(len(players_slices)):\n    first_dense_results.append((first_dense_layer(players_slices[k])))\n\nsecond_dense_layer = Dense(5000, activation='relu') #10000\nsecond_dense_results = []\n\nfor k in range(len(first_dense_results)):\n    second_dense_results.append((second_dense_layer(first_dense_results[k])))\n\nthird_dense_layer = Dense(1000, activation='relu')\nthird_dense_results = []\n\nfor k in range(len(second_dense_results)):\n    third_dense_results.append((third_dense_layer(second_dense_results[k])))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57726800,"AnswerCount":0,"Tags":"<pytorch><hessian>","CreationDate":"2019-08-30T12:29:55.967","AcceptedAnswerId":null,"OwnerUserId":7027783.0,"Title":"One of the differentiated Tensors does not require grad","Body":"<p>I\u2019m trying to get a hessian vector product, but the following codes from <a href=\"https:\/\/github.com\/quark0\/darts\/blob\/master\/cnn\/architect.py\" rel=\"nofollow noreferrer\">DARTS<\/a> model don\u2019t work as expected while self.named_parameters() is iterating over pairs n,p. <\/p>\n\n<p>The error is: RuntimeError: One of the differentiated Tensors does not require grad\nHow can I fix this error? Thank you in advance.<\/p>\n\n<pre><code>   def _hessian_vector_product(self, vector, input, target, r=1e-2):\n       R = r \/ _concat(vector).norm()\n       for p, v in zip(self.model.parameters(), vector):\n         p.data.add_(R, v)\n       loss = self.model._loss(input, target)\n       grads_p = torch.autograd.grad(loss, self.model.arch_parameters())\n\n       for p, v in zip(self.model.parameters(), vector):\n         p.data.sub_(2*R, v)\n       loss = self.model._loss(input, target)\n       grads_n = torch.autograd.grad(loss, self.model.arch_parameters())\n\n       for p, v in zip(self.model.parameters(), vector):\n         p.data.add_(R, v)\n\n       return [(x-y).div_(2*R) for x, y in zip(grads_p, grads_n)]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57727372,"AnswerCount":2,"Tags":"<python><pytorch><tensor>","CreationDate":"2019-08-30T13:08:12.813","AcceptedAnswerId":57727559.0,"OwnerUserId":10855529.0,"Title":"How do I get value of a tensor in PyTorch?","Body":"<pre><code>x = torch.tensor([3])\n<\/code><\/pre>\n\n<p>I have a tensor object\nwhen I<\/p>\n\n<pre><code>print(x)\n<\/code><\/pre>\n\n<p>it gives\n<code>tensor([3])<\/code><\/p>\n\n<p><code>x.data[0]<\/code> gives <code>tensor(3)<\/code>\nHow do I get just <code>3<\/code>?<\/p>\n","answers":[{"AnswerId":"57727748","CreationDate":"2019-08-30T13:32:26.413","ParentId":null,"OwnerUserId":"2888303","Title":null,"Body":"<p>You can use <code>x.item()<\/code> to get a Python number from a tensor that has one element<\/p>\n"},{"AnswerId":"57727559","CreationDate":"2019-08-30T13:20:42.480","ParentId":null,"OwnerUserId":"8563649","Title":null,"Body":"<p>Convert tensor to numpy:<\/p>\n\n<pre><code>x.numpy()[0]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57727468,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-serving><tensorflow-estimator>","CreationDate":"2019-08-30T13:14:31.970","AcceptedAnswerId":null,"OwnerUserId":11765356.0,"Title":"TensorFlow Serving REST ':predict' call returns serving_default instead","Body":"<p>I am trying to get a LinearClassifier Estimator model running predictions on TensorFlow Serving but my requests keep directing to the 'serving default' signature def instead and I don't know why.<\/p>\n\n<p>When I do a 'GET [model_path]\/metadata' the model shows that I have 'predict', 'classify' and 'serving_default' signatures as one would expect, and when tested locally with the saved_model_cli, they work as intended, but whenever 'POST [model_path]:predict' is tested with Postman or curl, I get the serving default output instead. \nI'm not too familiar with how to work with it, but when I change the serving_input_reciever_fn to something that only has a 'predict' signature, the service then tells me the following error<\/p>\n\n<pre><code>{\n    \"error\": \"Serving signature name: \\\"serving_default\\\" not found in signature def\"\n}\n<\/code><\/pre>\n\n<p>The 2 serving_input_receiver_fn that I have tried that seem somewhat promising are:<\/p>\n\n<pre><code>supervised_input_receiver_fn = (\n    tf.estimator.experimental.build_raw_supervised_input_receiver_fn(\n        {'int1': tf.placeholder(dtype=tf.int64, shape=[None]),\n         'int2': tf.placeholder(dtype=tf.int64, shape=[None]),\n         'int3': tf.placeholder(dtype=tf.int64, shape=[None]),\n         'string1': tf.placeholder(dtype=tf.int64, shape=[None])},\n        tf.placeholder(dtype=tf.int64, shape=[None])))\n<\/code><\/pre>\n\n<p>This first one giving me inputs for my predict signature accord to int 1, int 2, int3, and string 1\nand<\/p>\n\n<pre><code>feature_spec = tf.feature_column.make_parse_example_spec(feat_cols)\n\nserving_input_receiver_fn = tf.estimator.export.build_parsing_serving_input_receiver_fn(feature_spec)\n<\/code><\/pre>\n\n<p>This one instead taking a tf.example string as input which should supposedly be parsed to have the data registered. I have made this work in the past, but only with the serving_default signatre, not predict and for some reason it would always return the same results, even if I changed the input. <\/p>\n\n<p>I just want to be able to request my model and get prediction results that change based on the inputs!<\/p>\n","answers":[]}
+{"QuestionId":57727618,"AnswerCount":1,"Tags":"<concatenation><pytorch><tensor>","CreationDate":"2019-08-30T13:23:35.200","AcceptedAnswerId":null,"OwnerUserId":311744.0,"Title":"torch.cat but create a new dimension","Body":"<p>I would like to concatenate tensors, not along a dimension, but by creating a new dimension.<\/p>\n\n<p>For example:<\/p>\n\n<pre><code>x = torch.randn(2, 3)\nx.shape # (2, 3)\n\ntorch.cat([x,x,x,x], 0).shape # (8, 3)\n# This concats along dim 0, not what I want\n\ntorch.cat([x,x,x,x], -1).shape # (2, 10)\n# This concats along dim 1, not what I want\n\ntorch.cat([x[None, :, :],x[None, :, :],x[None, :, :],x[None, :, :]], 0).shape \n# =&gt; (4, 2, 3)\n# This is what I want, but unwieldy\n<\/code><\/pre>\n\n<p>Is there a simpler way?<\/p>\n","answers":[{"AnswerId":"57727619","CreationDate":"2019-08-30T13:23:35.200","ParentId":null,"OwnerUserId":"311744","Title":null,"Body":"<p>Just use <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html?highlight=stack#torch.stack\" rel=\"nofollow noreferrer\">torch.stack<\/a>:<\/p>\n\n<pre><code>torch.stack([x,x,x,x]).shape # (4, 2, 3)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57727782,"AnswerCount":2,"Tags":"<python><tensorflow><pip><python-3.7>","CreationDate":"2019-08-30T13:34:33.797","AcceptedAnswerId":null,"OwnerUserId":10641073.0,"Title":"Environmental Error occurs while installing Tensorflow","Body":"<p>When installing Tensorflow it said I was missing a file but i dont know how to fix it and or get that file<\/p>\n\n<p>I have tried re-downloading python but I am still missing the file <\/p>\n\n<pre><code>pip install Tensorflow\n<\/code><\/pre>\n\n<p>What I want is for the installation to complete.\nHere is the error message:<\/p>\n\n<blockquote>\n  <p>C:\\Users\\MYUSER\\AppData\\Local\\Packages\\PythonSoftwareFoundation.Python.3.7_qbz5n2kfra8p0\\LocalCache\\local-packages\\Python37\\site-packages\\tensorflow_estimator\\python\\estimator\\canned\\linear_optimizer\\python\\utils\\__pycache__\\sharded_mutable_dense_hashtable.cpython-37.pyc'<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"57730538","CreationDate":"2019-08-30T16:49:20.997","ParentId":null,"OwnerUserId":"10878366","Title":null,"Body":"<p>When installing Tensorflow it said I was missing a file but i don't know how to fix it and or get that file.<\/p>\n\n<p>Create a environment using anaconda.<\/p>\n\n<p>This Problem can be solved by execute the Below commands:<\/p>\n\n<pre><code>conda update anaconda\nconda create -n tensorflow pip python=3.5\nconda install --force html5lib\nactivate tensorflow\npip install --ignore-installed --upgrade tensorflow\nconda install -c conda-forge tensorflow\nconda install jupyter\n<\/code><\/pre>\n\n<p>Refer this <a href=\"https:\/\/www.tensorflow.org\/tutorials\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials<\/a><\/p>\n\n<p>Finally you can able to install Tensorflow.<\/p>\n\n<p>I hope it will help you.<\/p>\n"},{"AnswerId":"57729336","CreationDate":"2019-08-30T15:13:26.750","ParentId":null,"OwnerUserId":"2650249","Title":null,"Body":"<p>This is a typical Windows problem with long file paths. Disable the <code>MAX_PATH<\/code> limitation; the instruction for that can e.g. be found <a href=\"https:\/\/docs.python.org\/3\/using\/windows.html#removing-the-max-path-limitation\" rel=\"nofollow noreferrer\">here<\/a>:<\/p>\n\n<blockquote>\n  <h3>3.1.2. Removing the MAX_PATH Limitation<\/h3>\n  \n  <p>Windows historically has limited path lengths to 260 characters. This meant that paths longer than this would not resolve and errors would result.<\/p>\n  \n  <p>In the latest versions of Windows, this limitation can be expanded to approximately 32,000 characters. Your administrator will need to activate the \u201cEnable Win32 long paths\u201d group policy, or set the registry value <code>HKEY_LOCAL_MACHINE\\SYSTEM\\CurrentControlSet\\Control\\FileSystem@LongPathsEnabled<\/code> to <code>1<\/code>.<\/p>\n<\/blockquote>\n\n<p>After that, reboot and retry the installation.<\/p>\n"}]}
+{"QuestionId":57727920,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-30T13:43:16.227","AcceptedAnswerId":null,"OwnerUserId":3584765.0,"Title":"Getting the value of a tensor inside a function cracks the program","Body":"<p>I am working with keras\/tensorflow and I have an Single Shot Detector (SSD) model where at some point it uses a loss function which takes a <code>y_true<\/code> and <code>y_pred<\/code> and calculates the loss value. <\/p>\n\n<p>I am interested in recording both localization and classification loss so I tried to debug the provided loss function. Since, both are tensors I tried getting the value by calling:<\/p>\n\n<pre><code>y_true_np = tf.keras.backend.get_value(y_true)\ny_pred_np = tf.keras.backend.get_value(y_pred)\n<\/code><\/pre>\n\n<p>which in my case produces a CUDA_ERROR_OUT_OF_MEMORY error. The peculiar here is that I run the under debugging program in CPU (and RAM) because I don't have enough GPU memory available (I am running another experiment there).<\/p>\n\n<p>The same thing happens when using <code>tf.keras.backend.eval(y_true)<\/code> (I am not sure what's the difference between those two to be honest).<\/p>\n\n<p>It seems that keras tries to return the values I asked by creating a new session (?) and evaluating it there? Is that the case? And if so, why is that? Also, if I try to force it to be evaluated in CPU it fails even though the same commands (through config file) works fine without the above command. <\/p>\n\n<p>Can someone explain why I get this behavior and if I can get the actual value of the tensor somehow.<\/p>\n\n<p>Edit:<\/p>\n\n<p>The actual code I am using is from <a href=\"https:\/\/github.com\/pierluigiferrari\/ssd_keras\" rel=\"nofollow noreferrer\">here<\/a> but I am not sure I can strip it down to a simple one. Maybe it could be possible but I am probably lacking some additional information how to do it. The actual issue I am facing does not seem project specific though and probably would work the same for every loss function in keras. Anyway the actual loss function is <a href=\"https:\/\/github.com\/pierluigiferrari\/ssd_keras\/blob\/master\/keras_loss_function\/keras_ssd_loss.py\" rel=\"nofollow noreferrer\">this one<\/a> (from the same project). <\/p>\n","answers":[]}
+{"QuestionId":57728052,"AnswerCount":2,"Tags":"<python><keras><tensorflow2.0>","CreationDate":"2019-08-30T13:51:34.260","AcceptedAnswerId":57993473.0,"OwnerUserId":7537870.0,"Title":"Tensorflow 2.0rc not detecting GPUs","Body":"<p>TF2 is currently not detecting GPUs, I migrated from TF1.14 where using <\/p>\n\n<pre><code>tf.keras.utils.multi_gpu_model(model=model, gpus=2)\n<\/code><\/pre>\n\n<p>is now returning an error <\/p>\n\n<pre><code>ValueError: To call `multi_gpu_model` with `gpus=2`, we expect the following devices to be available: ['\/cpu:0', '\/gpu:0', '\/gpu:1']. However this machine only has: ['\/cpu:0', '\/xla_cpu:0', '\/xla_gpu:0', '\/xla_gpu:1', '\/xla_gpu:2', '\/xla_gpu:3']. Try reducing `gpus`.\n<\/code><\/pre>\n\n<p>Running <code>nvidia-smi<\/code> returns the following information<\/p>\n\n<pre><code>+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 418.67       Driver Version: 418.67       CUDA Version: 10.1     |\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 K80           Off  | 00000000:09:00.0 Off |                    0 |\n| N\/A   46C    P0    62W \/ 149W |      0MiB \/ 11441MiB |      0%      Default |\n+-------------------------------+----------------------+----------------------+\n|   1  Tesla K80           Off  | 00000000:0A:00.0 Off |                    0 |\n| N\/A   36C    P0    71W \/ 149W |      0MiB \/ 11441MiB |      0%      Default |\n+-------------------------------+----------------------+----------------------+\n|   2  Tesla K80           Off  | 00000000:86:00.0 Off |                    0 |\n| N\/A   38C    P0    58W \/ 149W |      0MiB \/ 11441MiB |      0%      Default |\n+-------------------------------+----------------------+----------------------+\n|   3  Tesla K80           Off  | 00000000:87:00.0 Off |                    0 |\n| N\/A   31C    P0    82W \/ 149W |      0MiB \/ 11441MiB |     73%      Default |\n+-------------------------------+----------------------+----------------------+\n\n+-----------------------------------------------------------------------------+\n| Processes:                                                       GPU Memory |\n|  GPU       PID   Type   Process name                             Usage      |\n|=============================================================================|\n|  No running processes found                                                 |\n+-----------------------------------------------------------------------------+\n<\/code><\/pre>\n\n<p>Also my TF version and is built for cuda<\/p>\n\n<pre><code>2.0.0-rc0\n<\/code><\/pre>\n\n<p>Please let me know what I am doing wrong so I can fix it.<\/p>\n","answers":[{"AnswerId":"57734818","CreationDate":"2019-08-31T02:47:36.767","ParentId":null,"OwnerUserId":"11477328","Title":null,"Body":"<p>CUDA should be the version 10.0, not 10.1<\/p>\n"},{"AnswerId":"57993473","CreationDate":"2019-09-18T13:03:46.090","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>I would suggest you to-<\/p>\n\n<ol>\n<li><p>Please first check your Cuda version. Make sure it is 10.0.<\/p><\/li>\n<li><p>If that is 10.0, then check your TF version if it is for GPU or not.<\/p><\/li>\n<li><p>Check if TF can access the GPUs using the command<\/p><\/li>\n<\/ol>\n\n<pre><code>value = tf.test.is_gpu_available(\n    cuda_only=False,\n    min_cuda_compute_capability=None\n)\nprint ('***If TF can access GPU: ***\\n\\n',value) # MUST RETURN True IF IT CAN!!\n<\/code><\/pre>\n\n<ol start=\"4\">\n<li>I assume first 2 points are already taken care of by you. If TF can also access your GPUs then, as you can see in your <code>Value error<\/code>, it has names of GPUs actually. I can not say about <code>tf.keras.utils.multi_gpu_model()<\/code> function because I did not use it in TF. But I would suggest you to use <code>with tf.device('\/gpu:0'):<\/code>. Inside this you call your <code>model<\/code> or define the model.<\/li>\n<li>If point 4 also doesn't work, then just add following lines <\/li>\n<\/ol>\n\n<pre><code>import os\nos.environ[\"CUDA_VISIBLE_DEVICES\"] = \"0,1,2,3\" # 0,1,2,3 are number of GPUs\n<\/code><\/pre>\n\n<p>at the top of your file and remove <code>with tf.device('\/gpu:0')<\/code><\/p>\n"}]}
+{"QuestionId":57728158,"AnswerCount":1,"Tags":"<python-3.x><pip><caffe>","CreationDate":"2019-08-30T13:57:34.900","AcceptedAnswerId":57728853.0,"OwnerUserId":517073.0,"Title":"How do I install caffe on debian with pip3?","Body":"<p>I am trying to do <strong>pip3 install caffe<\/strong>, but I am getting this on Ubuntu x64 machine:<\/p>\n\n<pre><code>Exception:\nTraceback (most recent call last):\n  File \"\/usr\/lib\/python3\/dist-packages\/pip\/basecommand.py\", line 215, in main\n    status = self.run(options, args)\n  File \"\/usr\/lib\/python3\/dist-packages\/pip\/commands\/install.py\", line 353, in run\n    wb.build(autobuilding=True)\n  File \"\/usr\/lib\/python3\/dist-packages\/pip\/wheel.py\", line 749, in build\n    self.requirement_set.prepare_files(self.finder)\n  File \"\/usr\/lib\/python3\/dist-packages\/pip\/req\/req_set.py\", line 380, in prepare_files\n    ignore_dependencies=self.ignore_dependencies))\n  File \"\/usr\/lib\/python3\/dist-packages\/pip\/req\/req_set.py\", line 554, in _prepare_file\n    require_hashes\n  File \"\/usr\/lib\/python3\/dist-packages\/pip\/req\/req_install.py\", line 278, in populate_link\n    self.link = finder.find_requirement(self, upgrade)\n  File \"\/usr\/lib\/python3\/dist-packages\/pip\/index.py\", line 465, in find_requirement\n    all_candidates = self.find_all_candidates(req.name)\n  File \"\/usr\/lib\/python3\/dist-packages\/pip\/index.py\", line 423, in find_all_candidates\n    for page in self._get_pages(url_locations, project_name):\n  File \"\/usr\/lib\/python3\/dist-packages\/pip\/index.py\", line 568, in _get_pages\n    page = self._get_page(location)\n  File \"\/usr\/lib\/python3\/dist-packages\/pip\/index.py\", line 683, in _get_page\n    return HTMLPage.get_page(link, session=self.session)\n  File \"\/usr\/lib\/python3\/dist-packages\/pip\/index.py\", line 795, in get_page\n    resp.raise_for_status()\n  File \"\/usr\/share\/python-wheels\/requests-2.18.4-py2.py3-none-any.whl\/requests\/models.py\", line 935, in raise_for_status\n    raise HTTPError(http_error_msg, response=self)\nrequests.exceptions.HTTPError: 404 Client Error: Not Found for url: https:\/\/pypi.org\/simple\/caffe\/\nstiv@gaidar:~\/lpr$ \n<\/code><\/pre>\n\n<p>how can I fix it<\/p>\n","answers":[{"AnswerId":"57728853","CreationDate":"2019-08-30T14:40:43.203","ParentId":null,"OwnerUserId":"517073","Title":null,"Body":"<p>Caffe should be installed on Ubuntu\/Debian via <strong>apt-get install cafe-cpu<\/strong> or <strong>apt-get install cafe-gpu<\/strong>.<\/p>\n"}]}
+{"QuestionId":57728477,"AnswerCount":1,"Tags":"<python><keras><conv-neural-network><object-detection><bounding-box>","CreationDate":"2019-08-30T14:17:51.307","AcceptedAnswerId":null,"OwnerUserId":11644036.0,"Title":"Bounding box prediction on CNN multiple class image classification in python","Body":"<p>I have the training set and test of 4 types of specific objects. I also have the bound box conditions \/ Area of interest coordinates (x,y,w,h) in csv format.\nMain aim of the project is to predict the class of test image along with bounding box around the area of interest along with printing the name of the class on the image.<\/p>\n\n<p>I have applied CNN model based on keras library. which classifies the given images of test set. what should i change in order to predict the bounding box coordinates of the given test image ?<\/p>\n\n<pre><code>        from keras.models import Sequential\n        from keras.layers import Convolution2D\n        from keras.layers import MaxPooling2D\n        from keras.layers import Flatten\n        from keras.layers import Dense\n\n        #CNN initializing\n        classifier= Sequential()\n\n        #convolutional layer\n        classifier.add(Convolution2D(filters = 32, kernel_size=(3,3), data_format= \"channels_last\", input_shape=(64, 64, 3), activation=\"relu\"))\n\n        #Pooling\n        classifier.add(MaxPooling2D(pool_size=(2,2)))\n\n        #addition of second convolutional layer\n        classifier.add(Convolution2D(filters = 32, kernel_size=(3,3), data_format= \"channels_last\", activation=\"relu\"))\n        classifier.add(MaxPooling2D(pool_size=(2,2)))\n\n        #step 3 - FLatttening\n        classifier.add(Flatten())\n\n        #step 4 - Full connection layer\n        classifier.add(Dense(128, input_dim = 11, activation = 'relu'))\n        #output layer\n        classifier.add(Dense(units = 4, activation = 'sigmoid'))\n\n        #compiling the CNN\n        classifier.compile(optimizer='adam',loss=\"categorical_crossentropy\",metrics =[\"accuracy\"])\n\n        #part 2 -Fitting the CNN to the images\n\n\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\/Train',\n                                                         target_size = (64, 64),\n                                                         batch_size = 32,\n                                                         class_mode = 'categorical')\n\n        test_set = test_datagen.flow_from_directory('dataset\/Test',\n                                                    target_size = (64, 64),\n                                                    batch_size = 32,\n                                                    class_mode = 'categorical')\n\n        classifier.fit_generator(training_set,\n                                 steps_per_epoch =4286\/32,\n                                 epochs = 25,\n                                 validation_data = test_set,\n                                 validation_steps = 44\/32)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57795740","CreationDate":"2019-09-04T21:17:47.603","ParentId":null,"OwnerUserId":"10952270","Title":null,"Body":"<p>The task you described is object detection, which usually requires a more complicated CNN model. Check <a href=\"https:\/\/github.com\/fizyr\/keras-retinanet\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fizyr\/keras-retinanet<\/a> for one of the famous neural network architectures.  <\/p>\n"}]}
+{"QuestionId":57728956,"AnswerCount":1,"Tags":"<python><c><tensorflow><keras><mnist>","CreationDate":"2019-08-30T14:47:42.197","AcceptedAnswerId":57819217.0,"OwnerUserId":4045358.0,"Title":"How to provide training data to TF_SessionRun from [c_api.h]","Body":"<p>I am trying to implement the training part of the <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/keras\/basic_classification\" rel=\"nofollow noreferrer\">fashion-mnist<\/a> tutorial using TF 2.0beta C api(c_api.h).<\/p>\n\n<p>Briefly, model is created using python following the previously mentioned tutorial. The model is then saved as a .pb file which is then loaded in my C program together with the training data (28*28 grayscale images) and training labels (scalars ranging from [0-9]).<\/p>\n\n<p>As some of you may know, the TF C api has poor documentation (insert rant full of frustration <a href=\"https:\/\/www.reddit.com\/r\/tensorflow\/comments\/9rl7yh\/motherfucking_opensource_tensorflow_shit_makes_it\/\" rel=\"nofollow noreferrer\">here<\/a>) and after much reading and searching on the internet <a href=\"https:\/\/gist.github.com\/asimshankar\/7c9f8a9b04323e93bb217109da8c7ad2#file-train-c\" rel=\"nofollow noreferrer\">here<\/a>, <a href=\"https:\/\/medium.com\/@vladislavsd\/undocumented-tensorflow-c-api-b527c0b4ef6\" rel=\"nofollow noreferrer\">here<\/a>, <a href=\"https:\/\/medium.com\/jim-fleming\/loading-a-tensorflow-graph-with-the-c-api-4caaff88463f\" rel=\"nofollow noreferrer\">here<\/a>, and <a href=\"https:\/\/github.com\/Neargye\/hello_tf_c_api\" rel=\"nofollow noreferrer\">here<\/a> among many other places, I am still unable to achieve this task (more details down below).<\/p>\n\n<h1>So let's get started:<\/h1>\n\n<p>I don't want to drown this question in code, but I want you to be able to access what I have done already.<\/p>\n\n<p>The python <a href=\"https:\/\/gist.github.com\/7PintsOfCherryGarcia\/094e68585be228e42efc5764a97ab5d7\" rel=\"nofollow noreferrer\">script<\/a> that generates the model saves it via:\nkeras.experimental.export_saved_model.<\/p>\n\n<p>To execute the code simply:\npython3 <a href=\"https:\/\/gist.github.com\/7PintsOfCherryGarcia\/094e68585be228e42efc5764a97ab5d7\" rel=\"nofollow noreferrer\">createModel.py<\/a>.<\/p>\n\n<p>Saved model will be stored in a new folder called \"saved_model\". <strong>This is important as this name is hard coded in the C code<\/strong><\/p>\n\n<p><a href=\"https:\/\/gist.github.com\/7PintsOfCherryGarcia\/455370c9550a6c5c10f9135155a8b282\" rel=\"nofollow noreferrer\">Data<\/a> for training the model is obtained via python TF as well.<\/p>\n\n<p>Again: python <a href=\"https:\/\/gist.github.com\/7PintsOfCherryGarcia\/455370c9550a6c5c10f9135155a8b282\" rel=\"nofollow noreferrer\">getData.py<\/a>.<\/p>\n\n<p>Saving the data and corresponding labels as data.txt and labels.txt <strong>Again, these names are important as they are hard coded in my C code<\/strong><\/p>\n\n<h1>Now we can move into C code!!<\/h1>\n\n<p>All code resides in a <a href=\"https:\/\/gist.github.com\/7PintsOfCherryGarcia\/407a0c115dd1eccb4ae78dc935dd4b71\" rel=\"nofollow noreferrer\">single<\/a> file with as much commenting as I imagined was necessary. Compilation should be straight forward as long as you have libtensorflow.so.2 and libtensorflow_framework.so.2.0.0<\/p>\n\n<p><code>gcc -Wall -I \/path\/to\/TFlib\/include -L \/path\/to\/TFlib\/lib -o trainMnist trainMnist.c<\/code><\/p>\n\n<p>In an x86_64 system there should be no warnings (At least in the two computers I have tried) <\/p>\n\n<p>After compilation execute with:<\/p>\n\n<p><code>LD_LIBRARY_PATH=\/path\/to\/TFlib\/lib .\/trainMnist<\/code><\/p>\n\n<p>You can skip LD_LIBRARY_PATH if the .so files are already reachable.<\/p>\n\n<p>If you don't have the TF shared object files I would gladly share them with anybody if you dare to trust compiled code from some random person. You can generate it yourself with:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>git clone https:\/\/github.com\/tensorflow\/tensorflow.git\ncd tensorflow\ngit checkout r2.0\n.\/configure\nbazel build -c opt \/\/tensorflow\/tools\/lib_package:libtensorflow \n<\/code><\/pre>\n\n<p>header and .so files will be in bazel-bin\/tensorflow\/tensorflow\/tools\/lib_package\/libtensorflow.tar.gz<\/p>\n\n<p>Assuming <code>saved_model<\/code>, <code>data.txt<\/code>, and <code>labels.txt<\/code> are in the same directory as where .\/trainMnist is ran, only TF warnings should be printed and some verbose messages from the code itself.<\/p>\n\n<p>but the function of interest is:<\/p>\n\n<pre class=\"lang-c prettyprint-override\"><code>int Belly_ModelTrain(model_t *model,\n                     float **train_data,\n                     float **label_data,\n                     int numPoints)\n<\/code><\/pre>\n\n<p><code>model<\/code> is a struct that holds important members to start a TF session<\/p>\n\n<p><code>train_data<\/code> holds numPoints arrays of 28*28=784 floats (pixel values of images)<\/p>\n\n<p><code>label_data<\/code> holds numPoints arrays of 10 floats (all 0 except the true label for that image).<\/p>\n\n<p><code>numPoints<\/code> is the number of images that are loaded. The value of <code>numPoints<\/code> is hard coded in main.<\/p>\n\n<p>In this function training occurs in the following way:<\/p>\n\n<pre class=\"lang-c prettyprint-override\"><code>  \/\/ The actual tensors that will hold the data and be passed to TF_SessionRun\n  TF_Tensor *x, *y;\n\n  \/\/ Establish the dimentions of the input data\n  const int64_t dimTrain[] = {numPoints, 784}; \/\/pixel values\n  const int64_t dimLabel[] = {numPoints, 10};  \/\/labels\n\n  {\n  \/\/Boilerplate allocation code happens here \n  size_t nbytesT = (numPoints*784) * sizeof(float);\n  x = TF_AllocateTensor(TF_FLOAT, dimTrain, 2, nbytesT);\n\n  size_t nbytesL = (numPoints*10) * sizeof(float);\n  y = TF_AllocateTensor(TF_FLOAT, dimLabel, 2, nbytesL);\n  }\n\n  \/\/Copy the data in my arrays to the tensors\n  memcpy(TF_TensorData(x), train_data, nbytesT);\n  memcpy(TF_TensorData(y), label_data, nbytesL);\n\n  \/\/TBH I really don't know what is going on in these two lines\n  \/\/Establish model inputs and model targets(labels)\n  TF_Output inputs[2] = {model-&gt;input, model-&gt;target};\n  TF_Tensor* input_values[2] = {x, y};\n  \/\/training operation defined in model\n  const TF_Operation* train_op[1] = {model-&gt;train_op};\n\n  \/\/Run the model with the data\n  TF_SessionRun(model-&gt;session,\n                NULL,\n                inputs, input_values, 2,\n                NULL, NULL, 0,\n                train_op, 1,\n                NULL, model-&gt;status);\n\n  TF_DeleteTensor(x);\n  TF_DeleteTensor(y);\n  \/\/Return 0 on success\n  return Belly_CheckStatus(model-&gt;status);\n<\/code><\/pre>\n\n<p>I expect this function to run nicely and return 0 based on the status of the model which should be TF_OK. This of course fails and I get:<\/p>\n\n<h1>ERROR: Node 'dense_2_target' (type: 'Placeholder', num of outputs: 1) does not have output [some crazy batshit number]<\/h1>\n\n<p>From the error message I understand my <code>target<\/code> input is misshaped. I have tried changing the way I pass <code>label_data<\/code> as well as changing <code>dimL<\/code> without any success.<\/p>\n","answers":[{"AnswerId":"57819217","CreationDate":"2019-09-06T09:23:07.383","ParentId":null,"OwnerUserId":"4045358","Title":null,"Body":"<p>So it turns out I was still thinking in python while working with the C api.<\/p>\n\n<p>In <a href=\"https:\/\/gist.github.com\/7PintsOfCherryGarcia\/407a0c115dd1eccb4ae78dc935dd4b71\" rel=\"nofollow noreferrer\">trainMnist.c<\/a> data is being read into an array of arrays (float **points):<\/p>\n\n<pre class=\"lang-c prettyprint-override\"><code>int Belly_ReadData(float **points, int numSamples) {\n  ...\n  ...\n  ...\n  for(int i = 0; i &lt; numSamples; i++) {\n    \/\/Allocate memory for datapoint\n    points[i] = (float *)malloc(784*sizeof(float));\n  ...\n  ...\n  ...\n    \/\/Read other 783 pixels\n    for(int j = 0; j &lt; 783; j++) {\n      point = strtok(NULL,\"\\t\");\n      points[i][j+1] = strtof(point,&amp;err);\n      if(*err != 0) return -1;\n    }\n  ...\n  }\n}\n<\/code><\/pre>\n\n<p>Afterwards this array of arrays is fed into the input tensor \"x\" in:<\/p>\n\n<pre class=\"lang-c prettyprint-override\"><code>int Belly_ModelTrain(model_t *model,\n                     float **train_data,\n                     float **label_data,\n                     int numPoints) {\n  ...\n  ...\n  ...\n\n  \/\/ Allocate data for tensors\n  x = TF_AllocateTensor(TF_FLOAT, dimTrain, 2, nbytesT);\n  if(x==NULL) {\n    printf(\"ERROR allocate x\\n\");\n    return -1;\n  }\n\n  y = TF_AllocateTensor(TF_FLOAT, dimLabel, 2, nbytesL);\n  if(y==NULL) {\n    printf(\"ERROR allocate y\\n\");\n    return -1;\n  }\n\n  \/\/ Copy data from arrays into tensors\n  memcpy(TF_TensorData(x), train_data, nbytesT); \/\/&lt;===== \"train_data should be an array of floats\"\n  memcpy(TF_TensorData(y), label_data, nbytesL); \/\/&lt;===== \"label_data should be an array of floats\"\n  ...\n  ...\n  ...\n<\/code><\/pre>\n\n<p>So effectively the way I understand it, I was passing garbage data into both tensors x and y.<\/p>\n\n<p>Will post corrected code for anybody who wants to try it. <\/p>\n"}]}
+{"QuestionId":57729279,"AnswerCount":2,"Tags":"<pytorch><dataloader>","CreationDate":"2019-08-30T15:08:25.963","AcceptedAnswerId":null,"OwnerUserId":9597731.0,"Title":"How to ensure that a batch contains samples from all workers with PyTorch's DataLoader?","Body":"<p>I want to know how to use <code>torch.utils.data.DataLoader<\/code> in PyTorch, especially in a multi-worker case.<\/p>\n\n<p>I found that one batch output from <code>DataLoader<\/code> always comes from a single worker. \nI expected that there is a queue in the DataLoader which stores data from all of the workers and DataLoader shuffles them in the queue to output the random batch data. I think this is the way in <code>tf.data.Dataset<\/code> in Tensorflow.\nCan we implement a similar function in PyTorch? I want to load a dataset from big serialized files (like <code>Tfrecord<\/code>) by using multi workers. In this case, mixing the source file in one batch, which means mixing the source of the worker, is important.<\/p>\n\n<p>Please refer to following code:<\/p>\n\n<pre><code>import random\nimport time\n\nimport torch\n\n\nclass MyDataset(torch.utils.data.Dataset):\n    def __len__(self):\n        return 50\n\n    def __getitem__(self, idx):\n        info = torch.utils.data.get_worker_info()\n\n        time.sleep(random.uniform(0, 1))\n        print(\"[{}]:{}\".format(info.id, idx))\n        return idx, info.id\n\n\nif __name__ == '__main__':\n    dataset = MyDataset()\n    dataloader = torch.utils.data.DataLoader(dataset, batch_size=5, shuffle=False, num_workers=2)\n    for batch in dataloader:\n        print(batch)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>[0]:0\n[1]:5\n[0]:1\n[1]:6\n[0]:2\n[0]:3\n[1]:7\n[0]:4\n[tensor([0, 1, 2, 3, 4]), tensor([0, 0, 0, 0, 0])]\n[1]:8\n[1]:9\n[tensor([5, 6, 7, 8, 9]), tensor([1, 1, 1, 1, 1])]\n[0]:10\n[0]:11\n[1]:15\n[1]:16\n[0]:12\n[1]:17\n...\n<\/code><\/pre>\n\n<p>Here, <code>[0, 1, 2, 3, 4]<\/code> and <code>[0, 0, 0, 0, 0]<\/code> in <code>[tensor([0, 1, 2, 3, 4]), tensor([0, 0, 0, 0, 0])]<\/code> mean that this batch includes index 0-th to 4-th data came from worker id <code>0<\/code>.\nNote that <code>shuffle=True<\/code> does not solve this problem which only change the indices of data.<\/p>\n\n<p>In this case, I want to get a batch like: <code>[tensor([0, 5, 1, 6, 2]), tensor([0, 1, 0, 1, 0])]<\/code>.<\/p>\n","answers":[{"AnswerId":"57734106","CreationDate":"2019-08-30T23:27:41.917","ParentId":null,"OwnerUserId":"1849998","Title":null,"Body":"<p>Note that a multi-worker <code>DataLoader<\/code> with a batch_size specified is going to load multiple batches parallelly, so essentially one batch comes from a worker always. However, I have achieved something close to what you require by doing as follows:<\/p>\n\n<ol>\n<li><p>Make the batch size as 1, so every worker only yields one sample at a time<\/p><\/li>\n<li><p>Write a background process that iterates through the DataLoader, fetches 1 sample at a time and inserts it into a queue. This makes it possible to have the samples in a different order in the queue rather than having worker specific batches<\/p><\/li>\n<li><p>Have a batching mechanism, like the <code>collate_fn<\/code> which takes a number of samples equal to your batch size from the queue and feed it to the model<\/p><\/li>\n<\/ol>\n\n<p>If you want to be more specific in batch creation, say like picking particular samples from specific workers, you can have multiples queues. Your collate procedure should be modified to account for multiple queues and choose from them. But I doubt if that kind of a specificity is needed.<\/p>\n"},{"AnswerId":"57989655","CreationDate":"2019-09-18T09:38:08.667","ParentId":null,"OwnerUserId":"2650622","Title":null,"Body":"<p>I've implemented something simple to solve a similar problem, where I have large video files as training data and each worker is responsible for loading and preprocessing a single file and then yielding samples from it. Problem is that as OP describes, with Pytorch's default data loading mechanism, each batch contains samples only from a single video file.<\/p>\n\n<p>First, let's review the problem. In this simplified code example each worker yields a single Tensor containing its zero-indexed worker id. With a batch size of 32 and 4 workers, we want each batch to contain 8 zeros, 8 ones, 8 twos and 8 threes.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from collections import defaultdict\n\nimport torch as T\nimport torch.utils.data as tdata\n\n\nclass Dataset(tdata.IterableDataset):\n    def __init__(self, batch_size: int):\n        self._bs = batch_size\n\n    def __iter__(self):\n        worker_info = tdata.get_worker_info()\n        if not worker_info:\n            raise NotImplementedError('Not implemented for num_workers=0')\n        for _ in range(self._bs):\n            yield T.tensor([worker_info.id])\n\n\nbatch_size = 32\nnum_workers = 4\ndataset = Dataset(batch_size)\nloader = tdata.DataLoader(dataset,\n                          batch_size=batch_size,\n                          num_workers=num_workers)\n\n\nfor batch in loader:\n    counts = defaultdict(int)\n    for n in batch.numpy().flatten():\n        counts[n] += 1\n    print(dict(counts))\n<\/code><\/pre>\n\n<p>Instead the code prints:<\/p>\n\n<pre><code>{0: 32}\n{1: 32}\n{2: 32}\n{3: 32}\n<\/code><\/pre>\n\n<p>Meaning that the first batch contains samples only from worker 0, the second only from worker 1, etc. To remedy this, we will set the batch size in the <code>DataLoader<\/code> to <code>batch_size \/\/ num_workers<\/code> and use a simple wrapper over the <code>DataLoader<\/code> to pool samples from each worker for our batch:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def pooled_batches(loader):\n    loader_it = iter(loader)\n    while True:\n        samples = []\n        for _ in range(loader.num_workers):\n            try:\n                samples.append(next(loader_it))\n            except StopIteration:\n                pass\n        if len(samples) == 0:\n            break\n        else:\n            yield T.cat(samples, dim=0)\n\n\nbatch_size = 32\nnum_workers = 4\ndataset = Dataset(batch_size)\nper_worker = batch_size \/\/ num_workers\nloader = tdata.DataLoader(dataset,\n                          batch_size=per_worker,\n                          num_workers=num_workers)\n\nfor batch in pooled_batches(loader):\n    counts = defaultdict(int)\n    for n in batch.numpy().flatten():\n        counts[n] += 1\n    print(dict(counts))\n<\/code><\/pre>\n\n<p>And the code now prints<\/p>\n\n<pre><code>{0: 8, 1: 8, 2: 8, 3: 8}\n{0: 8, 1: 8, 2: 8, 3: 8}\n{0: 8, 1: 8, 2: 8, 3: 8}\n{0: 8, 1: 8, 2: 8, 3: 8}\n<\/code><\/pre>\n\n<p>as expected.<\/p>\n"}]}
+{"QuestionId":57729857,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-08-30T15:50:54.030","AcceptedAnswerId":null,"OwnerUserId":9883450.0,"Title":"UnboundLocalError Tensorflow","Body":"<p>I'm trying to run a code for training a neural network using tensorflow and python 3.7, but this error is appearing. I'm new to tensorflow could someone give me a hint how to solve?<\/p>\n\n<p>This is my error: UnboundLocalError: local variable 'train_diretorio' referenced before assignment<\/p>\n\n<p>This is part of the code, If anyone can help, I'm new to python and tensorflow<\/p>\n\n<p><code>..Imports<\/p>\n\n<pre>def train():\nmodel_config = configuration.ModelConfig()\nmodel_config.input_file_pattern = input_file_pattern\nmodel_config.inception_checkpoint_file = inception_checkpoint_file\ntraining_config = configuration.TrainingConfig()\n\n# Create training directory.\ntrain_diretorio = train_diretorio\nif not tf.gfile.IsDirectory(train_diretorio):\n    tf.logging.info(\"Creating training directory: %s\", train_diretorio)\n    tf.gfile.MakeDirs(train_diretorio)\n\n# Build the TensorFlow graph.\ng = tf.Graph()\nwith g.as_default():\n    # Build the model.\n    model = show_and_tell_model.ShowAndTellModel(\n            model_config, mode=\"train\", train_inception=train_inception)\n    model.build()\n\n    # Set up the learning rate.\n    learning_rate_decay_fn = None\n    if train_inception:\n        learning_rate = tf.constant(training_config.train_inception_learning_rate)\n    else:\n        learning_rate = tf.constant(training_config.initial_learning_rate)\n        if training_config.learning_rate_decay_factor &gt; 0:\n            num_batches_per_epoch = (training_config.num_examples_per_epoch \/\n                             model_config.batch_size)\n            decay_steps = int(num_batches_per_epoch *\n                      training_config.num_epochs_per_decay)\n\n            def _learning_rate_decay_fn(learning_rate, global_step):\n                return tf.train.exponential_decay(\n                                  learning_rate,\n                                  global_step,\n                                  decay_steps=decay_steps,\n                                  decay_rate=training_config.learning_rate_decay_factor,\n                                  staircase=True)\n\n            learning_rate_decay_fn = _learning_rate_decay_fn\n\n    # Set up the training ops.\n    train_op = tf.contrib.layers.optimize_loss(\n                                    loss=model.total_loss,\n                                    global_step=model.global_step,\n                                    learning_rate=learning_rate,\n                                    optimizer=training_config.optimizer,\n                                    clip_gradients=training_config.clip_gradients,\n                                    learning_rate_decay_fn=learning_rate_decay_fn)\n\n    # Set up the Saver for saving and restoring model checkpoints.\n    saver = tf.train.Saver(max_to_keep=training_config.max_checkpoints_to_keep)\n\n# Run training.\ntf.contrib.slim.learning.train(\n                            train_op,\n                            train_diretorio,\n                            log_every_n_steps=log_every_n_steps,\n                            graph=g,\n                            global_step=model.global_step,\n                            number_of_steps=number_of_steps,\n                            init_fn=model.init_fn,\n                            saver=saver)\n\n\ninput_file_pattern = 'im2txt\/data\/mscoco\/train-?????-of-00256'\n\n\ninception_checkpoint_file = 'im2txt\/data\/inception_v3.ckpt'\ntrain_diretorio = 'im2txt\/model'\n\n\ntrain_inception = False\nnumber_of_steps = 1000000\nlog_every_n_steps = 1\n\ntrain()\n\n\ntrain_inception = False\nnumber_of_steps = 1000000\nlog_every_n_steps = 1\n\ntrain()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57730078,"AnswerCount":2,"Tags":"<python><tensorflow><tpu>","CreationDate":"2019-08-30T16:08:33.090","AcceptedAnswerId":null,"OwnerUserId":1500986.0,"Title":"What results can I expect from a quantized Tensoflow model?","Body":"<p>I have a very simple Tensorflow model that outputs 3 value (left, straight, right) based in the input from 6 sensors (input range from 0.0 to 1.0)<\/p>\n\n<pre><code>inputSensor = Input(shape=(3))\nmodelSensor = Dense(32, activation=\"relu\")(inputSensor)\nmodelSensor = Dense(32, activation=\"relu\")(modelSensor)\nmodelSensor = Dense(numClasses, activation=\"softmax\")(modelSensor)\nmodel = Model(inputs=inputSensor, outputs=modelSensor)\n<\/code><\/pre>\n\n<p>After training the model I can convert it into a normal TFLite model using<\/p>\n\n<pre><code>converter = tf.compat.v1.lite.TFLiteConverter.from_keras_model_file(\"models\/modelpath.h5\")\nconverter.optimizations = [tf.lite.Optimize.DEFAULT]\n<\/code><\/pre>\n\n<p>tflite_model = converter.convert()<\/p>\n\n<p>Then I run some inference tests on the model I get the results that I expect. One of the 3 outsputs\nis at 99%, the others very low.<\/p>\n\n<p>TEST RIGHT\nInput [0.0,0.0,1.0]\nResult [[0.00089144 0.00154889 0.99755967]]<\/p>\n\n<p>TEST Straight\nInput [0.0,1.0,0.0]\nResult [[0.00087439 0.996424   0.00270158]]<\/p>\n\n<p>TEST LEFT\nInput [1.0,0.0,0.0]\nResult [[0.9948014  0.00256803 0.00263061]]<\/p>\n\n<p>Then I quantize the model (in order to use it on a edge TPU)<\/p>\n\n<pre><code>converter =tf.compat.v1.lite.TFLiteConverter. from_keras_model_file \n(\"models\/modelpath.h5\")\nconverter.representative_dataset = representativeDataset_gen\nconverter.target_spec.supported_ops = \n[tf.lite.OpsSet.TFLITE_BUILTINS_INT8]\nconverter.inference_input_type = tf.uint8\nconverter.inference_output_type = tf.uint8\nconverter.optimizations = [tf.lite.Optimize.DEFAULT]\n<\/code><\/pre>\n\n<p>and I run the same test as above. The results are still correct, but I get 132 on the max output and 62 on the others.<\/p>\n\n<p>TEST RIGHT\nInput [0.0,0.0,1.0]\n[[ 62  62 132]]<\/p>\n\n<p>TEST Straight\nInput [0.0,1.0,0.0]\n[[ 62 132  62]]<\/p>\n\n<p>TEST LEFT\nInput [1.0,0.0,0.0]\n[[132  62  62]]<\/p>\n\n<p>I would have expect to get a result in the range of [[255  0  0]]<\/p>\n\n<p>The same thing happens, when I use integer inputs like\nInput [0,0,255]<\/p>\n\n<p>Is there something wrong with the quantazion conversion?\nAnd should I use Integers of floats as input?<\/p>\n","answers":[{"AnswerId":"57731311","CreationDate":"2019-08-30T17:56:28.513","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>Small sized models will certainly suffer from large accuracy drop when you do post-training quantization. In fact, it has been advised by TensorFlow Team to perform the accuracy check after post-training quantization has been applied.<\/p>\n\n<p>In case the accuracy is suffering badly, it has been suggested to make use of <a href=\"https:\/\/www.tensorflow.org\/lite\/performance\/post_training_quantization#model_accuracy\" rel=\"nofollow noreferrer\"><code>quantization-aware-training<\/code><\/a>.<\/p>\n\n<p>At the point of writing this answer, the quantization aware training is yet to be launched.<\/p>\n"},{"AnswerId":"57741050","CreationDate":"2019-08-31T19:23:50.760","ParentId":null,"OwnerUserId":"1500986","Title":null,"Body":"<p>I found the problem myself.<\/p>\n\n<p>The representative dataset that i was using for the quant-conversion was not representative enough. It consisted of an array of zero's. Using a proper dataset (the same one as used for training) fixed the outputs -> [[255   0   0]]<\/p>\n"}]}
+{"QuestionId":57730278,"AnswerCount":0,"Tags":"<machine-learning><keras><neural-network><deep-learning>","CreationDate":"2019-08-30T16:26:35.327","AcceptedAnswerId":null,"OwnerUserId":5079088.0,"Title":"Separate processing of classes in Keras Lambda layer","Body":"<p>I am trying to implement a triplet-center loss based on an idea that I haven't seen so far. To update centers of clusters using moving averages (or better but harder to implement) exponential moving averages.<\/p>\n\n<p>I need as much Keras and as little TF as possible :)<\/p>\n\n<p>Here is the problem:<\/p>\n\n<p>My Keras NN is receiving data in batches: (batch_size, prev_layer's_output), that is pretty standard.<\/p>\n\n<p>The NN (as always) receives data in form data + labels, where labels can be one-hot encoded or just class numbers.<\/p>\n\n<p>I need a Lambda Layer to do calculations based on labels (classes) and history (prev. result of this same layer). <\/p>\n\n<p>\"History\" here means that I need to use previous value that layer produces. It is required to calculate moving average.<\/p>\n\n<p>\"Based on labels\" means that I want to have SEPARATE outputs for separate classes.<\/p>\n\n<p>I see two ways of doing it. Either to make Lambda work with one class in a time, and to keep all its results (one result per each class) separately, or to make Lambda keep all classes, even if only one class data is updated in a time. <\/p>\n\n<p>Example:<\/p>\n\n<p>Let's say, I have MNIST dataset (10 classes) and previous (before Lambda) layer have produces 3 outputs (think coordinates), then I want the Lambda layer to hake those 3 inputs and a label (where and how do I provide a label to Lambda layer?) to produce 10 * 3 outputs (stored in Lambda or to update a separate array\/tensor), each containing the moving average of corresponding coordinate (X, Y, or Z) for corresponding class.<\/p>\n\n<p>Note that data is fed to a network one image in a time, so for all other classes, an old value of average is returned. <\/p>\n\n<p>Any ideas?\nThank you.<\/p>\n","answers":[]}
+{"QuestionId":57730494,"AnswerCount":3,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-30T16:45:18.427","AcceptedAnswerId":null,"OwnerUserId":1920178.0,"Title":"Keras fully connected followed by convolution","Body":"<p>I'm not super amazing with keras yet, so please be gentle.<\/p>\n\n<p>My input data is a matrix of size 60000 x 784.<\/p>\n\n<p>I'm trying to add convolutional layers after my fully connected layers, something like this:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(784, input_dim=train_amplitudes.shape[1], activation='relu'))\nmodel.add(Dense(784, activation='relu'))\nmodel.add(Dense(784, activation='relu'))\nmodel.add(Conv2D(100, kernel_size=5, activation='relu', input_shape=(28, 28))\nmode.add(Conv2D(20, kernel_size = 3, activation = 'relu'))\nmodel.add(Dense(train_targets.shape[1], activation='linear'))\n<\/code><\/pre>\n\n<p>Notice that 28 * 28 = 784.\nI get the error \"Input 0 is incompatible with layer conv2d_1: expected ndim=4, found ndim=2\" at the first convolution layer.<\/p>\n\n<p>Why and how can I fix this?<\/p>\n","answers":[{"AnswerId":"57730598","CreationDate":"2019-08-30T16:55:38.387","ParentId":null,"OwnerUserId":"9161105","Title":null,"Body":"<p>What is the purpose of this specific network structure? Assuming that your original data was 28x28, you should leave the input with 28x28 and then apply conv2d. After that you can flatten the last output of the convolutional blocks to continue with the fully connected layers.<\/p>\n"},{"AnswerId":"57730628","CreationDate":"2019-08-30T16:57:13.143","ParentId":null,"OwnerUserId":"6149357","Title":null,"Body":"<p>In Keras, input shape argument is a 4D tensor with shape: (batch, channels, rows, cols) if data_format is \"channels_first\" or 4D tensor with shape: (batch, rows, cols, channels) if data_format is \"channels_last\". You're just passing rows and columns (what you think) but it also requires batch and channels. More information can be found <a href=\"https:\/\/keras.io\/layers\/convolutional\/\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"57733876","CreationDate":"2019-08-30T22:41:48.187","ParentId":null,"OwnerUserId":"1920178","Title":null,"Body":"<p>I think I've managed to fix it. This is the code that \"works\"<\/p>\n\n<pre><code>    model = Sequential()\n    model.add(Dense(784, input_dim=train_amplitudes.shape[1], activation='relu'))\n    model.add(Dense(784, activation='relu'))\n    model.add(Dense(784, activation='relu'))\nmode.add(Reshape((28, 28, 1))\n    model.add(Conv2D(100, kernel_size=5, activation='relu')\n    mode.add(Conv2D(20, kernel_size = 3, activation = 'relu'))\nmodel.add(Flatten())\n    model.add(Dense(train_targets.shape[1], activation='linear'))\n<\/code><\/pre>\n\n<p>it works in the sense that no error is produced. Whether it makes sense or produces a good output, that's another matter, but this is good enough for me.<\/p>\n"}]}
+{"QuestionId":57730934,"AnswerCount":1,"Tags":"<tensorflow><object-detection><tensorflow-lite>","CreationDate":"2019-08-30T17:23:30.887","AcceptedAnswerId":57730935.0,"OwnerUserId":5963300.0,"Title":"How can I install Tensorflow GPU with latest version of CUDA and cuDNN","Body":"<p>Installing TensorFlow for object detection is annoying sometimes, especially when wired errors happen after starting one's own object detection project by finetuning pre-trained model.<\/p>\n\n<p>How to install latest Tensorflow GPU support and latest CUDA\/CUDNN without any error?<\/p>\n","answers":[{"AnswerId":"57730935","CreationDate":"2019-08-30T17:23:30.887","ParentId":null,"OwnerUserId":"5963300","Title":null,"Body":"<h2>To install tensorflow-gpu object detection api on Ubuntu 16.04<\/h2>\n\n<h3>upgrade system<\/h3>\n\n<pre><code>sudo apt-get update\n\nsudo apt-get upgrade\n<\/code><\/pre>\n\n<h3>install basic packages<\/h3>\n\n<pre><code>sudo apt-get install vim curl python-dev gnupg-curl python-tk git\n\ncurl https:\/\/bootstrap.pypa.io\/get-pip.py -o get-pip.py\n\nsudo -H python get-pip.py\n<\/code><\/pre>\n\n<h3>install tensorflow-gpu (tensorflow version=1.14)<\/h3>\n\n<pre><code>sudo -H pip install tensorflow-gpu\n<\/code><\/pre>\n\n<h3>install CUDA (Final CUDA version = 10.1)<\/h3>\n\n<pre><code>\/\/ Please turn off your secure boot from BIOS\n\nsudo apt-get install gnupg-curl\n\n\/\/ Here we install version 10.0 to avoid other issues. Later we can upgrade it to version 10.1\nwget https:\/\/developer.download.nvidia.com\/compute\/cuda\/repos\/ubuntu1604\/x86_64\/cuda-repo-ubuntu1604_10.0.130-1_amd64.deb\n\nsudo dpkg -i cuda-repo-ubuntu1604_10.0.130-1_amd64.deb\n\nsudo apt-key adv --fetch-keys https:\/\/developer.download.nvidia.com\/compute\/cuda\/repos\/ubuntu1604\/x86_64\/7fa2af80.pub\n\nsudo apt-get update\n\nwget http:\/\/developer.download.nvidia.com\/compute\/machine-learning\/repos\/ubuntu1604\/x86_64\/nvidia-machine-learning-repo-ubuntu1604_1.0.0-1_amd64.deb\n\nsudo apt install .\/nvidia-machine-learning-repo-ubuntu1604_1.0.0-1_amd64.deb\n\nsudo apt-get update \n<\/code><\/pre>\n\n<h3>install cuDNN (cuDNN version = 7.6.3.30-1, compatible to CUDA 10.1)<\/h3>\n\n<pre><code>sudo apt-get install --no-install-recommends cuda-10-0\n\n\/\/ restart your computer here\n\nnvidia-smi.\n\nsudo apt-get install --no-install-recommends libcudnn7=7.6.3.30-1+cuda10.0 libcudnn7-dev=7.6.3.30-1+cuda10.0\n\nsudo apt-get install -y --no-install-recommends libnvinfer5=5.1.5-1+cuda10.0 libnvinfer-dev=5.1.5-1+cuda10.0\n\nsudo apt-get update\n\n\/\/ this upgrade command will upgrade your CUDA to version 10.1\nsudo apt-get upgrade\n\nsudo apt-get autoremove\n<\/code><\/pre>\n\n<h3>install Tensorflow object detection api<\/h3>\n\n<pre><code>sudo -H pip install Cython\n\nsudo -H pip install contextlib2\n\nsudo -H pip install pillow\n\nsudo -H pip install lxml\n\nsudo -H pip install jupyter\n\nsudo -H pip install matplotlib\n\n\nmkdir tensorflow\ncd tensorflow\ngit clone https:\/\/github.com\/tensorflow\/models\n\n\/\/ install protocbuf version 3.0.0\nwget -O protobuf.zip https:\/\/github.com\/google\/protobuf\/releases\/download\/v3.0.0\/protoc-3.0.0-linux-x86_64.zip\nunzip protobuf.zip\nsudo cp .\/bin\/protoc \/bin\/\nsudo cp -r .\/include\/google \/usr\/local\/include\/\ncd tensorflow\/models\/research\nprotoc object_detection\/protos\/*.proto --python_out=.\n\ngit clone https:\/\/github.com\/cocodataset\/cocoapi.git\ncd cocoapi\/PythonAPI\nmake\ncp -r pycocotools &lt;path_to_tensorflow&gt;\/models\/research\/\n<\/code><\/pre>\n\n<h3>Export Environment<\/h3>\n\n<pre><code>export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:\/usr\/local\/cuda-10.1\/lib64\nexport PYTHONPATH=$PYTHONPATH:~\/tensorflow\/models\/research:~\/tensorflow\/models\/research\/object_detection\/slim\n<\/code><\/pre>\n\n<h3>Other<\/h3>\n\n<pre><code>If we install CUDA 10.1 it may have a compatibility issue that libxxxx.so.10.0 is not found when you start training your project.\n\nTo solve them:\n\na. sudo ln -s \/usr\/lib\/x86_64-linux-gnu\/libcublas.so.10.1 \/usr\/local\/cuda-10.1\/lib64\/libcublas.so.10.0\n\nb. sudo ln -s \/usr\/local\/cuda-10.1\/lib64\/libcudart.so.10.1 \/usr\/local\/cuda-10.1\/lib64\/libcudart.so.10.0\n\nc. sudo ln -s \/usr\/local\/cuda-10.1\/lib64\/libcufft.so.10 \/usr\/local\/cuda-10.1\/lib64\/libcufft.so.10.0\n\nd. sudo ln -s \/usr\/local\/cuda-10.1\/lib64\/libcurand.so.10 \/usr\/local\/cuda-10.1\/lib64\/libcurand.so.10.0\n\ne. sudo ln -s \/usr\/local\/cuda-10.1\/lib64\/libcusolver.so.10 \/usr\/local\/cuda-10.1\/lib64\/libcusolver.so.10.0\n\nf. sudo ln -s \/usr\/local\/cuda-10.1\/lib64\/libcusparse.so.10 \/usr\/local\/cuda-10.1\/lib64\/libcusparse.so.10.0\n<\/code><\/pre>\n\n<h3>All Set<\/h3>\n"}]}
+{"QuestionId":57730978,"AnswerCount":1,"Tags":"<json><tensorflow><tensorflow-serving><amazon-sagemaker>","CreationDate":"2019-08-30T17:28:20.260","AcceptedAnswerId":null,"OwnerUserId":12000572.0,"Title":"Error in giving inputs to the tensorflow serving model on sagemaker. {'error': \"Missing 'inputs' or 'instances' key\"}","Body":"<p>I have a custom model built-in TensorFlow. I am trying to deploy this model on amazon sagemaker for inference. The model takes three inputs and gives five outputs.\nThe name of the inputs are:<\/p>\n\n<pre><code>1.    input_image\n2.    input_image_meta\n3.    input_anchors\n<\/code><\/pre>\n\n<p>and the name of outputs are:<\/p>\n\n<pre><code>1    output_detections\n2    output_mrcnn_class\n3    output_mrcnn_bbox\n4    output_mrcnn_mask\n5    output_rois\n<\/code><\/pre>\n\n<p>I have successfully created the model endpoint on sagemaker and when I am trying to hit the request for the results, I am getting {'error': \"Missing 'inputs' or 'instances' key\"} in return.<\/p>\n\n<p>The sagemaker endpoint gets created and the tensorflow server also starts(as shown in CloudWatch logs).\nOn the client side, I call the predictor using follwoing code:<\/p>\n\n<pre><code>request = {}\nrequest[\"img_link\"] = \"image.jpg\"\nresult = predictor.predict(request)\n<\/code><\/pre>\n\n<p>But when I print the result the following gets printed out, {'error': \"Missing 'inputs' or 'instances' key\"}\nAll the bucket connections for loading the image are in inference.py<\/p>\n","answers":[{"AnswerId":"57755243","CreationDate":"2019-09-02T10:20:24.870","ParentId":null,"OwnerUserId":"10340061","Title":null,"Body":"<p>Tensorflow serving prediction api should have a key \"instances\" or \"inputs\" to modify your input. Your request body should be like this<\/p>\n\n<pre><code>{\n    \"instances\": [{\n        \"input_image\": DATA1,\n        \"input_image_meta\": DATA2,\n        \"input_anchors\": DATA3}]\n}\n<\/code><\/pre>\n\n<p>More <a href=\"https:\/\/tensorflow.google.cn\/tfx\/serving\/api_rest\" rel=\"nofollow noreferrer\">information<\/a><\/p>\n"}]}
+{"QuestionId":57731214,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-30T17:48:21.677","AcceptedAnswerId":null,"OwnerUserId":1179925.0,"Title":"What tf.keras.backend.clear_session actually do?","Body":"<p>What is <code>tf.keras.backend.clear_session<\/code> actually do?<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/backend\/clear_session\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/backend\/clear_session<\/a><\/p>\n\n<p>How it's related to <code>tf.reset_default_graph()<\/code> and  <code>sess.close()<\/code> ?<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reset_default_graph\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reset_default_graph<\/a><\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#close\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#close<\/a><\/p>\n","answers":[{"AnswerId":"57732427","CreationDate":"2019-08-30T19:39:58.203","ParentId":null,"OwnerUserId":"7331631","Title":null,"Body":"<p>There are two main concepts in TensorFlow 1.0 <strong>Graphs<\/strong> and <strong>Sessions<\/strong>.<\/p>\n\n<ul>\n<li><p>Graph - It is a set of connected operations and placeholders which doesn't hold any tensor(numpy array) or values without a session. As an analogy, you can consider a food processing assembly line without any ingredients, but process and recipes are defined. <\/p><\/li>\n<li><p>Session - It takes the graph and initializes the variable with initial values and ready to take some to feed in the placeholder to start implementing the operations defined the graph to the feed values in placeholders, at last, it will deliver you the final output from you desired operation node (in neural network nodes of the last layer.)(like feeding tomatoes and getting ketchup as output.)<\/p>\n\n<p>coming back to your real question. <\/p><\/li>\n<li><p>If you use the <code>tf.keras.backend.clear_session<\/code> it will discard the values resides in the variable defined in the graph, leaving an empty vessel. (It will free up your RAM space.), now you can load weights from some other files.<\/p><\/li>\n<li><p>If you use the <code>tf.reset_default_graph()<\/code> it will reset the graph and it will remove all the defined operations and their inter-connection with corresponding weights. Now you have to load both models architecture and weights for the execution.<\/p><\/li>\n<\/ul>\n\n<p>practically it seems it is doing same stuff cause it is tf.reset_default_graph() will be called internally while calling k.clear_session() but clear_session will also intiate the fresh graph for the new operation you can check the source code <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.14\/tensorflow\/python\/keras\/backend.py#L217-L242\" rel=\"nofollow noreferrer\">here<\/a> <\/p>\n"}]}
+{"QuestionId":57731613,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-08-30T18:24:06.860","AcceptedAnswerId":null,"OwnerUserId":12000763.0,"Title":"ImportError: cannot import name 'weak_script_method'","Body":"<p>When I use PyTorch to import weak_script_method, the code throws an error:<\/p>\n\n<blockquote>\n  <p>from torch._jit_internal import weak_script_method ImportError: cannot\n  import name 'weak_script_method'<\/p>\n<\/blockquote>\n\n<p>The code is:<\/p>\n\n<pre><code>from torch._jit_internal import weak_script_method\n<\/code><\/pre>\n\n<p>Does anyone have an idea about this? How to fix this error?<\/p>\n","answers":[{"AnswerId":"57811739","CreationDate":"2019-09-05T19:15:28.140","ParentId":null,"OwnerUserId":"3014199","Title":null,"Body":"<p><code>weak_script_method<\/code> is in <code>_jit_internal.py<\/code> in version <code>v1.1.0<\/code><\/p>\n\n<p>As you can look up here:<br>\n<a href=\"https:\/\/github.com\/pytorch\/pytorch\/blob\/v1.1.0\/torch\/_jit_internal.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/pytorch\/blob\/v1.1.0\/torch\/_jit_internal.py<\/a><\/p>\n\n<p>In version <code>v1.2.0<\/code> is is no longer there:<br>\n<a href=\"https:\/\/github.com\/pytorch\/pytorch\/blob\/v1.1.0\/torch\/_jit_internal.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/pytorch\/blob\/v1.1.0\/torch\/_jit_internal.py<\/a><\/p>\n\n<p>by inspecting the history you can find this commit which has removed the function you want:<br>\n<a href=\"https:\/\/github.com\/pytorch\/pytorch\/commit\/10c4b98ade8349d841518d22f19a653a939e260c#diff-ee07db084d958260fd24b4b02d4f078d\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/pytorch\/commit\/10c4b98ade8349d841518d22f19a653a939e260c#diff-ee07db084d958260fd24b4b02d4f078d<\/a><\/p>\n\n<p>I've never needed <code>weak_script_method<\/code> myself and I don't know what you need it for. It seems the pytorch developers don't need it any longer...<br>\nBest of luck!<\/p>\n"}]}
+{"QuestionId":57731870,"AnswerCount":1,"Tags":"<tensorflow><object-detection><object-detection-api><cudnn>","CreationDate":"2019-08-30T18:48:30.113","AcceptedAnswerId":57731871.0,"OwnerUserId":5963300.0,"Title":"What is \"cudnn_status_internal_error\" meaning Tensorflow pipeline?","Body":"<p>If we start an object_detection project using TensorFlow Model API, we got an error \"cudnn_status_internal_error\". <\/p>\n\n<p>I'm using TensorFlow object detection pipeline to train the model.<\/p>\n\n<p>How to solve it?<\/p>\n","answers":[{"AnswerId":"57731871","CreationDate":"2019-08-30T18:48:30.113","ParentId":null,"OwnerUserId":"5963300","Title":null,"Body":"<p>In my case, it is because the GPU resource is not enough.<\/p>\n\n<p>To solve this issue in the pipeline, please add a piece of code in \"model_main.py\"<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>  session_config = tf.ConfigProto()\n  #session_config.gpu_options.allow_growth = True\n  session_config.gpu_options.per_process_gpu_memory_fraction = 0.8\n\n  config = tf.estimator.RunConfig(model_dir=FLAGS.model_dir, session_config=session_config)\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":57732296,"AnswerCount":0,"Tags":"<keras><lstm><stateful>","CreationDate":"2019-08-30T19:28:09.207","AcceptedAnswerId":null,"OwnerUserId":7305901.0,"Title":"Is there any way to use the EarlyStopping Keras function when I apply a stateful LSTM and I reset the states?","Body":"<p>I am using a stateful LSTM regression model and I would like to apply the EarlyStopping function. In stateful LSTMs as I was reading the states should be reset at each epoch. However, I noticed that when I was resetting the states the EarlyStopping method wasn't working at all. I attach also the code.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nmodel = Sequential()\nmodel.add(LSTM(256, batch_input_shape=(batch_size, timesteps, features), return_sequences=False, stateful=True))\nmodel.add(Dropout(rate=0.2))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='mean_squared_error', optimizer='adam')\nmc = ModelCheckpoint('best_model.h5', monitor='val_loss', mode='min', verbose=0, save_best_only=True)\nes = EarlyStopping(monitor='val_loss', mode='min', patience=1, restore_best_weights=True, verbose=1)\n\nfor epoch in range(epochs):\n    print(\"Epoch: \", epoch + 1)\n\n    hist = model.fit(train_x, train_y, epochs=1, batch_size, shuffle=False,\n                     validation_data=(validation_x, validation_y), verbose=2, callbacks=[mc, es])\n    model.reset_states()\n\n<\/code><\/pre>\n\n<p>If I run the above code without the for loop and without the reset of states then the EarlyStopping works fine. Is there any way to apply the EarlyStopping in a for loop? <\/p>\n\n<p>Thank you in advance<\/p>\n","answers":[]}
+{"QuestionId":57732694,"AnswerCount":0,"Tags":"<python><keras><lstm>","CreationDate":"2019-08-30T20:07:12.697","AcceptedAnswerId":null,"OwnerUserId":7024618.0,"Title":"LSTM many to many prediction","Body":"<p>I have 26 time series about batteries observations. I need to estimate the change in capacity after every charge\/discharge cycle. I modeled each cycle with 4 features, each battery has a variable number of cycles. After each cycle there is an estimation of the capacity in terms of decrease from the previous step. How can I predict the capacity decrease at each time step?<\/p>\n\n<p>I have 26 batteries representing 26 samples. Each sample has a variable number of time steps and each time step has a label which is the capacity decrease (delta Q) calculated at the end of the charge\/discharge cycle. I'm new at RNN, I'm trying to use a LSTM with batch size of 1 where each batch represents a single battery and each battery is independent from the other, so I set the stateful parameter to False. I'm using a single LSTM layer that outputs the n values of delta Q (capacity) which I want to predict. I removed metrics accuracy because I think this is more a regression task, given a sequence of time steps I want n output values. I train one batch at a time because the sequences are not of the same length. The purpose is to predict, given another battery time series, the n values of delta capacity at each one of the n time steps of that battery. I'm obtaining weird values of loss when I train the model, I'm at the beginning with this subject, so please be patient and supportive. I need help<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>batch_size = 1\nn_features = 4\nn_steps = None\nn_features = 4\n\n# define model\nmodel = Sequential()\nmodel.add(LSTM(units=1, batch_input_shape=(1, None, n_features), activation='linear',return_sequences=True, stateful=False))\nmodel.compile(loss = 'mean_squared_error', optimizer = 'adam', metrics=['mean_squared_error'])\nmodel.summary()\nepochs = 50\nhistory = []\n\nindexes = [index for index in range(len(X_train))]\n\nfor epoch in range(epochs):\n    random.shuffle(indexes)\n    for i in range(len(X_train)):\n        index = indexes[i]\n        x_train = np.array(X_train[index], dtype=np.float)\n        x_train = np.reshape(x_train, (1, x_train.shape[0], x_train.shape[1]))\n        y_train = np.array(Y_train[index], dtype=np.float)\n        y_train = np.reshape(y_train, (1, y_train.shape[0], 1))\n        metrics = model.train_on_batch(x_train, y_train)\n    history.append(metrics)\n    print(\"loss: \", metrics[0])\n<\/code><\/pre>\n\n<p>My loss results during training are totally random, I don't know why. Are 26 samples (each sample has between 11 and 39 arrays of 4 features) too few for an LSTM to work properly?<\/p>\n","answers":[]}
+{"QuestionId":57733053,"AnswerCount":0,"Tags":"<python><tensorflow><image-processing>","CreationDate":"2019-08-30T20:48:13.903","AcceptedAnswerId":null,"OwnerUserId":5910685.0,"Title":"How to replace pixel value as a tensorflow operation?","Body":"<p>I need to replace a pixel value in an image as an operation in the graph. Doing this beforehand is unfortunately not an option as it is part of an optimization process.<\/p>\n\n<p>As a fix until I come up with a solution, I am simply using <code>tf.py_func()<\/code> but since this operation has to be executed a lot it's very slow and inefficient.<\/p>\n\n<pre><code># numpy function to perturb a single pixel in an image\n\ndef perturb_image(pixel, img):\n\n  # At each pixel's x,y position, assign its rgb value\n\n  x_pos, y_pos, r, g, b = pixel\n  rgb = [r,g,b]\n  img[x_pos, y_pos] = rgb\n\n  return img\n\n# pixel is a 1D tensor like [x-dim,y-dim,R,G,B]\n# image is tensor with shape (x-dim,y-dim,3)\n\nimg_perturbed = tf.py_func(perturb_image,[pixel, image], tf.uint8)\n<\/code><\/pre>\n\n<p>One way I thought of is using <code>tf.add(perturbation, image)<\/code> where both have the same dimension and perturbation is all zeros except at the pixel location which needs its RGB-values changed to the same value as defined in <code>pixel<\/code> from the above code snippet. Unfortunately, I would need to rewrite a lot of code surrounding this operation which I am trying to avoid.<\/p>\n\n<p>Can you think of a solution to replace <code>py_func<\/code> with another tensorflow operation using the same inputs?<\/p>\n\n<p>Any help is much appreciated.<\/p>\n","answers":[]}
+{"QuestionId":57733316,"AnswerCount":0,"Tags":"<tensorflow><neural-network><gradient><layer><tensorboard>","CreationDate":"2019-08-30T21:17:57.803","AcceptedAnswerId":null,"OwnerUserId":11969270.0,"Title":"how to get gradient of last layer of CNN for calculating gradient_norm?","Body":"<p>I tried to calculate gradient norm of the last dense layer before output of a CNN. windows 10, jupyter notebook 6, tensorflow 1.13. Here are the codes.<\/p>\n\n<pre><code>optimizer=tf.train.AdamOptimizer(learning_rate)\n\ngradients = optimizer.compute_gradients(loss=loss)\n\nlast_grad_norm=0;\nfor g,v in gradients:\n      if 'last' in v.name :\n           last_grad_norm = tf.sqrt(tf.reduce_mean(g**2))\n      break\n\ntrain_op = optimizer.apply_gradients(gradients, global_step=global_step)\n\neval_metric_ops = None\n\nwith tf.name_scope('performance'):\n      tf.summary.scalar('gradient_norm', last_grad_norm)\n      tf.summary.scalar('learning_rate', learning_rate)\n      tf.summary.scalar('accuracy', accuracy[1])\n<\/code><\/pre>\n\n<p>I did name that layer as 'last'. on tensorboard, the graph is still showing it as 'dense' in the box.<\/p>\n\n<p>and the last_grad_norm is always 0 in the plot, which indicates that the if condition was never True. I checked the tensorflow doc. I have found how the var_list are named. <\/p>\n\n<p>anyone can provide a way to get that gradient out?<\/p>\n","answers":[]}
+{"QuestionId":57733453,"AnswerCount":0,"Tags":"<python><keras><neural-network>","CreationDate":"2019-08-30T21:34:08.473","AcceptedAnswerId":null,"OwnerUserId":1019129.0,"Title":"integer\/real based classification at output layer?","Body":"<p>Normally you would use softmax activation at the output layer if you have multiple-dependent classes and sigmoid if you have independent-classes (with one-hot encoding). By using multiple neurons , one per class.<\/p>\n\n<p>How would you classify on a single neuron with integer\/real value, where every value means a different class.<\/p>\n\n<p>F.e. let say we pass at the input a number (based on vocabulary) which encodes a word and get at the output another number (signifying a word from vocabulary). <\/p>\n\n<p>What would be the output layer configuration ? how would you that ? <\/p>\n\n<hr>\n\n<p>when I test the following network on different arithmetic operations it seems to work ! Where I use arithmetic operation to simulate a categorical-in-out relation.<\/p>\n\n<p>The question is will it preserve its behavior when mixed with category data, I mean in a mixed input data scenario, where this one is a branch of the full NN !?<\/p>\n\n<pre><code>class Arithmetic:\n\n    def tt_data(self, ins=[np.random.randint(100) for _ in xrange(500)], outs=lambda x: x \/ 3):\n        self.ins = np.array(ins)\n        self.outs = np.array([ outs(i) for i in self.ins])\n        self.in_size = 1\n        self.out_size = 1\n        return ( self.ins, self.outs )\n\n\n    def build_model(self, ins=None, outs=None) :\n        if isinstance(ins,tuple) : ins, outs = ins\n        if ins is None : ins = self.ins\n        if outs is None : outs = self.outs\n\n        inputs = Input(shape=(self.in_size,))\n        x = Dense(3,activation='relu')(inputs)\n        x = Dense(3,activation='relu')(x)\n        outputs = Dense(self.out_size,activation='linear')(x)\n        self.model = Model(inputs=inputs, outputs=outputs)\n        self.model.compile(optimizer='adam', loss='mse', metrics=['accuracy','mae'])\n        print self.model.summary()\n        self.history = self.model.fit(self.ins, self.outs, batch_size=10, epochs=200, verbose=1, validation_split=0.2)\n        self.score = self.model.evaluate(ins, outs, verbose=1)\n        print self.score\n\n<\/code><\/pre>\n\n<pre><code> In [29]: a.build_model(a.tt_data())               \n In [26]: a.model.predict(list(xrange(200,301,10)))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57733798,"AnswerCount":0,"Tags":"<tensorflow><keras><tensorflow2.0>","CreationDate":"2019-08-30T22:26:27.567","AcceptedAnswerId":null,"OwnerUserId":6082275.0,"Title":"Load model weights and optimizer from Keras callback in Tensorflow 2","Body":"<p>I am dealing with a strange situation, possibly a Tensorflow bug.\nI am using keras callback to save my model and my optimizer, the command is the following:<\/p>\n\n<pre><code>cp_callback = tf.keras.callbacks.ModelCheckpoint(\n    filepath=checkpoints_path,\n    verbose=1,\n    save_freq=args.save_checkpoint_steps)\n<\/code><\/pre>\n\n<p>After training for some epochs, I end up with a checkpoints folder with the following contents:<\/p>\n\n<p><code>checkpoint, cp-0001.ckpt.data-00000-of-00002, cp-0001.ckpt.data-00001-of-00002, cp-0001.ckpt.index<\/code><\/p>\n\n<p>When using <code>tf.train.latest_checkpoint(dir)<\/code> I am getting <code>cp-0001.ckpt<\/code> which is fine.<\/p>\n\n<p>Loading the model weights is straightforward: <\/p>\n\n<pre><code>latest = tf.train.latest_checkpoint(checkpoints_dir)\nmodel.load_weights(latest)\n<\/code><\/pre>\n\n<p>However, loading model and optimizer doesn't seem to work. Even worse, the Tensorflow manager can't even find that there are existing checkpoints.<\/p>\n\n<pre><code>ckpt = tf.train.Checkpoint(step=tf.Variable(1), net=model,\n                           optimizer=optimizer)\nmanager = tf.train.CheckpointManager(\n              ckpt, tf.train.latest_checkpoint(checkpoints_dir),\n              max_to_keep=20)\nckpt.restore(manager.latest_checkpoint)\n<\/code><\/pre>\n\n<p>This doesn't restore anything, I checked the status. Also, <code>manager.checkpoints<\/code> is <code>None<\/code>.<\/p>\n\n<p>How is this even possible? I can't find anything that explains this behavior in the docs. \nAny ideas that might help?\nThanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":57733802,"AnswerCount":1,"Tags":"<tensorflow><object-detection>","CreationDate":"2019-08-30T22:27:39.707","AcceptedAnswerId":null,"OwnerUserId":12001372.0,"Title":"Tensorflow C++ Object detection(video)","Body":"<p>its my first question in stackoverflow however \nmy Question how i can use my already trained frozen_graph and label_map from python object detection and use that in my c++ object detection.<\/p>\n\n<p>i used python for object detection in a video and it wasnt so fast at all 25 fps with mss.sct() and multithreaded<\/p>\n\n<p>so i want to try it in c++ cause i know that there are more possible and heared that python is little slow.<\/p>\n\n<p>Someone has an object detection example for an image in c++ or c# ?<\/p>\n\n<p>Would be awesome <\/p>\n","answers":[{"AnswerId":"57795682","CreationDate":"2019-09-04T21:10:27.360","ParentId":null,"OwnerUserId":"10952270","Title":null,"Body":"<p>Check <a href=\"https:\/\/medium.com\/jim-fleming\/loading-a-tensorflow-graph-with-the-c-api-4caaff88463f\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/jim-fleming\/loading-a-tensorflow-graph-with-the-c-api-4caaff88463f<\/a>, <\/p>\n\n<p>and you need to build C++ API like <a href=\"https:\/\/joe-antognini.github.io\/machine-learning\/build-windows-tf\" rel=\"nofollow noreferrer\">https:\/\/joe-antognini.github.io\/machine-learning\/build-windows-tf<\/a><\/p>\n"}]}
+{"QuestionId":57734203,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-08-30T23:50:23.763","AcceptedAnswerId":null,"OwnerUserId":915783.0,"Title":"Does Shuffling in Keras ImagedataGenerator cover all the samples in a epoch","Body":"<p>I wanted to check that in keras ImageDataGnerator , Shuffling don't cause duplication of data in certain batches, while missing samples in others, from original training data. here is my code:<\/p>\n\n<pre><code>from __future__ import print_function\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom sklearn.utils import class_weight\nfrom keras.preprocessing.image import ImageDataGenerator, image,load_img\nfrom keras import backend as K\nfrom keras import models\nfrom keras import layers\nfrom keras import optimizers\nfrom keras.models import load_model\nfrom keras.models import model_from_json\n# set image dimension for Conv layer etc based on tensor flow or theano\nK.set_image_dim_ordering('tf')\nfrom keras.applications import InceptionResNetV2\nfrom keras.applications.inception_resnet_v2 import preprocess_input\nfrom keras.callbacks import ModelCheckpoint\nimport os\nimport sys\nimport glob\nimport argparse\nfrom keras.utils import Sequence\nfrom keras.callbacks import ReduceLROnPlateau\nfrom collections import Counter\nimport pandas\n\n    #************************************************************\n    #************ create data generator, load data, and fit the model ********************\n\nbs_pth = '.\/data\/no_resize'\ntrain_dir = bs_pth + '\/train'\n\nW=299\nH=299\nbatch_size = 20\ndatagen_tr  = ImageDataGenerator(preprocessing_function=preprocess_input)\ntrain_gen = datagen_tr.flow_from_directory(train_dir,target_size=(W, H),batch_size=batch_size,class_mode='categorical',shuffle=False,interpolation=\"bilinear\")\ntrain_gen1 = datagen_tr.flow_from_directory(train_dir,target_size=(W, H),batch_size=batch_size,class_mode='categorical',shuffle=True,interpolation=\"bilinear\")\n\n\nimg_files_tr_true = []\n# find true filenames and combine into a list\nfor path, subdirs1_categ, files in os.walk(train_dir):\n    if len(files) != 0:\n        for nm in files:\n            img_files_tr_true = img_files_tr_true + [nm]\n\n#  without shuffle\nimg_files_tr = []\nbatches_per_epoch = train_gen.samples \/\/ train_gen.batch_size + (train_gen.samples % train_gen.batch_size &gt; 0)  # datagen.batch_size + (datagen.samples % datagen.batch_size &gt; 0)\nwith open ('.\/data\/saved_plots\/imageDatagen_data_analys_v1.txt',\"w\")as fp:\n    for i in range(batches_per_epoch):\n        batch = next(train_gen)\n        current_index = ((train_gen.batch_index-1) * train_gen.batch_size)\n        if current_index &lt; 0:\n            if train_gen.samples % train_gen.batch_size &gt; 0:\n                current_index = max(0,train_gen.samples - train_gen.samples % train_gen.batch_size)\n            else:\n                current_index = max(0,train_gen.samples - train_gen.batch_size)\n        index_array = train_gen.index_array[current_index:current_index + train_gen.batch_size].tolist()\n        cu_btch = \"batch\" + str(i)\n        fp.write(cu_btch + \"\\n\")\n        for idx in index_array:\n            cu_pth = train_gen.filepaths[idx]\n            bs = train_gen.filenames[idx]\n            cu_nm =  os.path.split(bs)[1]\n            img_files_tr = img_files_tr + [cu_nm]\n            fp.write(cu_pth + \"\\n\")\nimg_files_tr_diff1 = np.setdiff1d(img_files_tr_true,img_files_tr)\n\n# with shuffle\nimg_files_tr1 = []\nbatches_per_epoch = train_gen1.samples \/\/ train_gen1.batch_size + (train_gen1.samples % train_gen1.batch_size &gt; 0)  # datagen.batch_size + (datagen.samples % datagen.batch_size &gt; 0)\n\nwith open ('.\/data\/saved_plots\/imageDatagen_data_analys_v2.txt',\"w\")as fp:\n    for i in range(batches_per_epoch):\n        batch = next(train_gen1)\n        current_index = ((train_gen1.batch_index-1) * train_gen1.batch_size)\n        if current_index &lt; 0:\n            if train_gen1.samples % train_gen1.batch_size &gt; 0:\n                current_index = max(0,train_gen1.samples - train_gen1.samples % train_gen1.batch_size)\n            else:\n                current_index = max(0,train_gen1.samples - train_gen1.batch_size)\n        index_array = train_gen1.index_array[current_index:current_index + train_gen1.batch_size].tolist()\n        cu_btch = \"batch\" + str(i)\n        fp.write(cu_btch + \"\\n\")\n        for idx in index_array:\n            cu_pth = train_gen1.filepaths[idx]\n            bs = train_gen1.filenames[idx]\n            cu_nm = os.path.split(bs)[1]\n            img_files_tr1 = img_files_tr1 + [cu_nm]\n            fp.write(cu_pth + \"\\n\")\n\nimg_files_tr_diff2 = np.setdiff1d(img_files_tr_true,img_files_tr1)\n<\/code><\/pre>\n\n<p>As can be seen, <strong>img_files_tr_diff1<\/strong> and <strong>img_files_tr_diff2<\/strong> turn out to be empty lists confirming that Shuffle or not, Keras ImageDataGenerator does use all the data when using flow_from_directory. Is my assertion correct or am i missing something?<\/p>\n\n<p>sedy<\/p>\n","answers":[]}
+{"QuestionId":57734846,"AnswerCount":0,"Tags":"<python><tensorflow><lstm>","CreationDate":"2019-08-31T02:57:00.630","AcceptedAnswerId":null,"OwnerUserId":12001724.0,"Title":"Using lstm in TensorFlow for prediction by time sequence, while my result can be so bad and absurd","Body":"<p>I have some 20 GDPs in a group, and I wanted to use LSTM NN in TensorFlow to predict the next 10 years GDP performance.\nAs the original data shows, the GDPs are supposed to increase exponentially, but my result through the lstm suggest the totally opposite trend, I found it very very confused to debug!! Is there anything wrong with my network architecture?<\/p>\n\n<p>I tried to adjust the learning rate from 0.1 to 0.0001, and have changed the batch_size, time_step, and the run_unit_size, both in range of (2, 20), but none of those method works.<\/p>\n\n<pre class=\"lang-html prettyprint-override\"><code>raw data: \nyear      gdp\n1996    349600\n1997    422100\n1998    479100\n1999    389711\n2000    428654\n2001    486430\n2002    565532\n2003    677583\n2004    950951\n2005    1453209\n2006    1831073\n2007    2454598\n2008    3018032\n2009    4594015\n2010    5736350\n2011    7678600\n2012    8793200\n2013    10017593\n2014    11153849\n2015    11933430\n2016    12933462\n<\/code><\/pre>\n\n<p>getting batches of training data:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def get_train_data(normalized):\n    target = list()\n    label = list()\n    for i in range(len(a_np) - time_step - 1):\n        target.append(normalized[i:i + time_step])\n        label.append(normalized[i + 1:i + time_step + 1])\n    return target, label\n<\/code><\/pre>\n\n<p>the basic info of LSTM:<\/p>\n\n<pre><code>time_step = 4\nrnn_unit = 20\nbatch_size = 5\ninput_size = 1\noutput_size = 1\nlr = 0.008\nX = tf.placeholder(tf.float32, [None, time_step, input_size])\nY = tf.placeholder(tf.float32, [None, time_step, output_size])\nweights = {\n        'in': tf.Variable(tf.random_normal([input_size, rnn_unit])),\n        'out': tf.Variable(tf.random_normal([rnn_unit, 1]))\n    }\nbiases = {\n    'in': tf.Variable(tf.constant(0.1, shape=[rnn_unit, ])),\n    'out': tf.Variable(tf.constant(0.1, shape=[1, ]))\n}\n<\/code><\/pre>\n\n<p>3 layers lstm from input-hidden through to lstm finally to the output-hidden layer:<\/p>\n\n<pre><code>def lstm(batch):\n    w_in = weights['in']\n    b_in = biases['in']\n    input = tf.reshape(X, [-1, input_size])\n    input_rnn = tf.matmul(input, w_in) + b_in\n    input_rnn = tf.reshape(input_rnn, [-1, time_step, rnn_unit])\n    cell = tf.nn.rnn_cell.BasicLSTMCell(rnn_unit, forget_bias=0.2, state_is_tuple=True)\n    init_state = cell.zero_state(batch, dtype=tf.float32)\n    output_rnn, final_states = tf.nn.dynamic_rnn(cell, 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<\/code><\/pre>\n\n<p>loss function and optimizer:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>loss = tf.reduce_mean(tf.square(tf.reshape(Y, [-1]) - tf.reshape(pred, [-1])))\ntrain_op = tf.train.AdamOptimizer(lr).minimize(loss)\n<\/code><\/pre>\n\n<pre class=\"lang-html prettyprint-override\"><code>this is the predicted result\n2017    [11262389.]\n2018    [11572339.]\n2019    [11730970.]\n2020    [11825313.]\n2021    [11873324.]\n2022    [11906612.]\n2023    [11924690.]\n2024    [11935304.]\n2025    [11941413.]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57734929,"AnswerCount":1,"Tags":"<python><tensorflow><keras><conv-neural-network>","CreationDate":"2019-08-31T03:24:59.037","AcceptedAnswerId":null,"OwnerUserId":5313254.0,"Title":"3D Convolutional Autoencoder doesn't match the output layer with the input layer","Body":"<p>I'm trying to create a 3 Dimensional Convolutional Neural Network AutoEncoder. I'm unable to match my input dimension of the tensor with the output<\/p>\n\n<p>I have tried changing the layer shape and using Keras autoencoder.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>        padding = 'SAME'\n        stride = [1,1,1]\n\n        self.inputs_ = tf.placeholder(tf.float32, input_shape, name='inputs')\n        self.targets_ = tf.placeholder(tf.float32, input_shape, name='targets')\n\n        conv1 = tf.layers.conv3d(inputs= self.inputs_, filters=16, kernel_size=(3,3,3), padding= padding, strides = stride, activation=tf.nn.relu)  \n        maxpool1 = tf.layers.max_pooling3d(conv1, pool_size=(2,2,2), strides=(2,2,2), padding= padding)\n        conv2 = tf.layers.conv3d(inputs=maxpool1, filters=32, kernel_size=(3,3,3), padding= padding, strides = stride, activation=tf.nn.relu)\n        maxpool2 = tf.layers.max_pooling3d(conv2, pool_size=(3,3,3), strides=(3,3,3), padding= padding)\n        conv3 = tf.layers.conv3d(inputs=maxpool2, filters=96, kernel_size=(2,2,2), padding= padding , strides = stride, activation=tf.nn.relu)\n        maxpool3 = tf.layers.max_pooling3d(conv3, pool_size=(2,2,2), strides=(2,2,2), padding= padding)\n        #latent internal representation\n\n        #decoder\n#         tf.keras.layers.UpSampling3D()\n        unpool1 =K.resize_volumes(maxpool3,2,2,2,\"channels_last\")\n        deconv1 = tf.layers.conv3d_transpose(inputs=unpool1, filters=96, kernel_size=(2,2,2), padding= padding , strides = stride, activation=tf.nn.relu)\n        unpool2 = K.resize_volumes(deconv1,3,3,3,\"channels_last\")\n        deconv2 = tf.layers.conv3d_transpose(inputs=unpool2, filters=32, kernel_size=(3,3,3), padding= padding , strides = stride, activation=tf.nn.relu)\n        unpool3 = K.resize_volumes(deconv2,2,2,2,\"channels_last\")\n        deconv3 = tf.layers.conv3d_transpose(inputs=unpool3, filters=16, kernel_size=(3,3,3), padding= padding , strides = stride, activation=tf.nn.relu)\n        self.output = tf.layers.dense(inputs=deconv3, units=3)\n        self.output = tf.reshape(self.output, self.input_shape)\n\n<\/code><\/pre>\n\n<p>ValueError: Cannot reshape a tensor with 1850688 elements to shape [1,31,73,201,3] (1364589 elements) for 'Reshape' (op: 'Reshape') with input shapes: [1,36,84,204,3], [5] and with input tensors computed as partial shapes: input[1] = [1,31,73,201,3].<\/p>\n","answers":[{"AnswerId":"57737867","CreationDate":"2019-08-31T12:05:23.813","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>Your input shape is <code>[1, 31, 73, 201, 3]<\/code>. During the transpose convolution, you are performing upscaling of <code>[2,2,2]<\/code>, <code>[3,3,3]<\/code> and <code>[2,2,2]<\/code> in your three <code>resize_volumes<\/code> layers. If you multiply these numbers across the axis it will be <code>[12, 12, 12]<\/code> (each one 2*3*2). So, the output of the decoder will be multiple of 12 in each dimension. <\/p>\n\n<p>But your input dimension shape of <code>[x, 31, 73, 201, x]<\/code> are not multiples of 12. Nearest multiples greater than those dimensions are <code>[x, 36, 84, 204, x]<\/code>. So, the solution will be either after the decoding part you will have strip out the excess dimension and match it to your original dimension, or else better solution is to pad the original shape with zeros and make it a multiple of 12. After following the second solution, you will have to consider the new dimension of input.<\/p>\n\n<p>Updated code (only changed part)<\/p>\n\n<pre><code>self.inputs_ = tf.placeholder(tf.float32, input_shape, name='inputs')\npad_inputs = tf.pad(self.inputs_, [[0,0], [2, 3], [5, 6], [1, 2], [0, 0]]) # Pad at the edges\nprint(pad_inputs.shape)  # [1, 36, 84, 204, 3]\n\nconv1 = tf.layers.conv3d(inputs= pad_inputs, filters=16, kernel_size=(3,3,3), padding= padding, strides = stride, activation=tf.nn.relu)\n<\/code><\/pre>\n\n<p>And at the end,<\/p>\n\n<pre><code>self.output = tf.reshape(self.output, pad_inputs.shape)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57734955,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras><scikit-learn>","CreationDate":"2019-08-31T03:31:09.213","AcceptedAnswerId":null,"OwnerUserId":10704050.0,"Title":"How to use LSTM for sequence classification using KerasClassifier","Body":"<p>I have a <code>binary classification<\/code> problem where I need to predict the potential future trendy\/popular products based on customer interactions during <code>2010-2015<\/code>.<\/p>\n\n<p>Currently, my dataset includes <code>1000 products<\/code> and each product is labelled as <code>0<\/code> or <code>1<\/code> (i.e. binary classification). The label was decided based on customer interactions during <code>2016-2018<\/code>.<\/p>\n\n<p>I am calculating <code>how centrality measures changed over time for each product<\/code> during <code>2010-2015<\/code> as the features for my binary classification problem. For example, consider the below figure that shows how <code>degree centrality<\/code> changed over time for each product.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/mXslL.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/mXslL.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>More specifically, I analyse the change of following <code>centrality measures<\/code> as the features for my binary classification problem.<\/p>\n\n<ul>\n<li>how <code>degree centrality<\/code> of each good changed from 2010-2016 (see the above figure)<\/li>\n<li>how <code>betweenness centrality<\/code> of each good changed from 2010-2016<\/li>\n<li>how <code>closeness centrality<\/code> of each good changed from 2010-2016<\/li>\n<li>how <code>eigenvector centrality<\/code> of each good changed from 2010-2016<\/li>\n<\/ul>\n\n<p>In a nutshell, my data looks as follows.<\/p>\n\n<pre><code>product, change_of_degree_centrality, change_of_betweenness_centrality, change_of_closenss_centrality, change_of_eigenvector_centrality, Label\nitem_1, [1.2, 2.5, 3.7, 4.2, 5.6, 8.8], [8.8, 4.6, 3.2, 9.2, 7.8, 8.6], \u2026, 1\nitem_2, [5.2, 4.5, 3.7, 2.2, 1.6, 0.8], [1.5, 0, 1.2, 1.9, 2.5, 1.2], \u2026, 0\nand so on.\n<\/code><\/pre>\n\n<p>I wanted to use deep learning model to solve my issue. When reading tutorials, I realised that <code>LSTM<\/code> suits my problem.<\/p>\n\n<p>So, I am using the below mentioned model for my classification.<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(10, input_shape=(6,4))) #where 6 is length of centrality sequence and 4 is types of centrality (i.e. degree centrality, betweenness centrality, closeness centrality, and eigenvector centrality)\nmodel.add(Dense(32))\nmodel.add(Dense(1, activation=\u2019sigmoid\u2019))\nmodel.compile(loss=\u2019binary_crossentropy\u2019, optimizer=\u2019adam\u2019, metrics=[\u2018accuracy\u2019])\n<\/code><\/pre>\n\n<p>Since, I have a small dataset I wanted to perform 10-fold cross-validation. So, I am using <code>KerasClassifier<\/code> as follows by following <a href=\"https:\/\/machinelearningmastery.com\/use-keras-deep-learning-models-scikit-learn-python\/\" rel=\"nofollow noreferrer\">this tutorial<\/a>.<\/p>\n\n<pre><code>print(features.shape) #(1000,6,4)\nprint(target.shape) #(1000) \n\n# Create function returning a compiled network\ndef create_network():\n    model = Sequential()\n    model.add(LSTM(10, input_shape=(6,4)))\n    model.add(Dense(32))\n    model.add(Dense(1, activation='sigmoid'))\n    model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])     \n\n    return model\n\n# Wrap Keras model so it can be used by scikit-learn\nneural_network = KerasClassifier(build_fn=create_network, \n                                 epochs=10, \n                                 batch_size=100, \n                                 verbose=0)\n\nprint(cross_val_score(neural_network, features, target, cv=5))\n<\/code><\/pre>\n\n<p>However, I noted that it is wrong to use <code>cross validation<\/code> with LSTM (e.g., <a href=\"https:\/\/machinelearningmastery.com\/backtest-machine-learning-models-time-series-forecasting\/\" rel=\"nofollow noreferrer\">this tutorial<\/a>, <a href=\"https:\/\/stats.stackexchange.com\/questions\/326228\/cross-validation-with-time-series\">this question<\/a>). <\/p>\n\n<p>However, I am not clear if this is applicable to me as I am only doing a \n<code>binary classification<\/code> predition to identify products that would be trendy\/popular in future (not a forecasting).<\/p>\n\n<p>I think the data in my problem setting is divided by <em>point-wise<\/em> in the cross-validation, but not <em>time-wise<\/em>.<\/p>\n\n<p>i.e. (point-wise)<\/p>\n\n<pre><code>1st fold training:\nitem_1, item2, ........, item_799, item_800\n\n1st fold testing:\nitem 801, ........, item_1000\n<\/code><\/pre>\n\n<p>not (time-wise)<\/p>\n\n<pre><code>1st fold training:\n2010, 2011, ........, 2015\n\n1st fold testing:\n2016, ........, 2018\n<\/code><\/pre>\n\n<p>Due to this fact, I am assuming that using <code>cross validation<\/code> is correct in my problem.<\/p>\n\n<p>Please let me know a suitable way to use cross-validation according to my problem and dataset.<\/p>\n\n<p>NOTE: I am not limited to LSTM and happy to explore other models as well.<\/p>\n\n<p>I am happy to provide more details if needed.<\/p>\n","answers":[{"AnswerId":"57737501","CreationDate":"2019-08-31T11:10:43.630","ParentId":null,"OwnerUserId":"10606534","Title":null,"Body":"<p>Maybe you misunderstand the concept ,the KerasClassifier is suite for LSTM<\/p>\n\n<p>base on those link you give , it just say the cross-valid not suite for time-series <\/p>\n\n<p>row-grow <\/p>\n\n<p>but LSTM is clomn grow n time series<\/p>\n"},{"AnswerId":"57735357","CreationDate":"2019-08-31T05:11:41.873","ParentId":null,"OwnerUserId":"2255479","Title":null,"Body":"<p>There are many types of cross validation similar to how there are many types of neural networks. In your case you are trying to use kfold cross validation. <\/p>\n\n<p>In the question you linked, it correctly states that kfold cross validation should not be used with time series data. You can\u2019t accurately evaluate your model if you are training on data and then testing on data that occurred before the training data. <\/p>\n\n<p>However, other forms of cross validation (such as the mentioned sliding window or expanding window) will still work with your time series data. There is a function in sklearn that splits the data using the expanding window method. <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.model_selection.TimeSeriesSplit.html\" rel=\"nofollow noreferrer\">https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.model_selection.TimeSeriesSplit.html<\/a><\/p>\n\n<p>With all that said, I am not sure if you are really using time series data. If you simply have the centrality scores for each year as a separate feature, then the order of your data does not matter since each item is only one data point (assuming that the scores of one item don\u2019t impact another). In that case you can use kfold cross validation and other networks that work with iid data. You could even use non neural networks such as SVMs or decision trees. <\/p>\n"}]}
+{"QuestionId":57735388,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-08-31T05:18:55.497","AcceptedAnswerId":null,"OwnerUserId":9922891.0,"Title":"Pass input and target samples with different sizes in Keras","Body":"<p>I am writing an LSTM language model in Keras. Suppose<br>\n- the input samples have the shape of <code>(20, 5)<\/code><br>\n- the target samples (labels) have the shape of <code>(20 * 5)<\/code><br>\n- the output predictions have the shape of <code>(20 * 5, 100)<\/code><\/p>\n\n<p>When I pass the input-target pairs to the model, Keras throws the exception<br>\n<code>ValueError: Input arrays should have the same number of samples as target arrays.<\/code><\/p>\n\n<p>The input samples and the target samples do have different sizes, but eventually the shape of the output predictions and the target samples will have the same size (<code>(20 * 5, ...)<\/code>). I want to compute the <code>sparse_categorial_crossentropy<\/code> between the output predictions and the target samples with the shapes mentioned above.<\/p>\n\n<p>How can I pass the input and target samples with different sizes to the Keras model if the output samples will have the same size as the target samples? Thanks in advance.<\/p>\n\n<p>Here is some example code to create the model.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential()\n\n# input shape: (20, 5)\nmodel.add(Embedding(input_dim=100, output_dim=10))\n\n# shape after encoding: (20, 5, 10)\nmodel.add(LSTM(units=10, return_sequences=True))\nmodel.add(LSTM(units=10))\n\n# shape after lstm: (20 * 5, 10)\nmodel.add(Dense(units=100, activation='softmax'))\n# output shape: (20 * 5, 100)\n\nmodel.compile(loss='sparse_categorical_crossentropy')\n\n# train_batch is a batch generator. For each batch, it returns the \n# input and target samples with the shape (20, 5) and (20 * 5) respectively\nmodel.fit_generator(generator=train_batch)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57735481,"AnswerCount":0,"Tags":"<python><tensorflow><keras><tensorflow2.0><eager-execution>","CreationDate":"2019-08-31T05:41:26.903","AcceptedAnswerId":null,"OwnerUserId":11169701.0,"Title":"Tensorflow 2.0 : Combining model.add_loss and keras losses function in training doesn't work","Body":"<p>I read about TensorFlow 2.0 <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/keras\/custom_layers_and_models\" rel=\"nofollow noreferrer\">tutorial<\/a> in VAE section. I follow the tutorial but the model doesn't work as expected despite running the notebook directly from given <a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/docs\/blob\/r2.0rc\/site\/en\/r2\/guide\/keras\/custom_layers_and_models.ipynb\" rel=\"nofollow noreferrer\">Google Colab<\/a>. The result actually is the same as in the tutorial (i.e. loss value is very similar) but if you look at the output you'll see that the model can't reconstruct the input at all (i.e. output the same image for all inputs). This seems to be a mistake from the tutorial itself when combining <code>model.add_loss()<\/code> and <code>keras.losses<\/code>.<\/p>\n\n<p><a href=\"https:\/\/imgur.com\/3rTAOrZ\" rel=\"nofollow noreferrer\">Original code<\/a><\/p>\n\n<p>I changed MSE loss to BinaryCrossentropy but the result is still the same.<\/p>\n\n<p>Later I tried compute the BinaryCrossentropy loss explicitly in my forward pass then use <code>model.add_loss()<\/code> in addition with the KL-divergence loss<\/p>\n\n<p><a href=\"https:\/\/imgur.com\/VVOROCI\" rel=\"nofollow noreferrer\">Use only model.add_loss() to calculate the loss<\/a><\/p>\n\n<p>This way the model can actually learn the data and the output seems good enough.<\/p>\n\n<p>So I have a question about <code>model.add_loss()<\/code> and losses as a function that takes <code>(y_true, y_pred)<\/code> (i.e. <code>keras.losses<\/code>). The updated code works only if it can calculate losses in forward pass (e.g. kl-divergence or reconstruction loss), how can I combine <code>model.add_loss()<\/code> and <code>keras.losses<\/code> correctly in the case where the model need ground truth of the output (e.g. denoise VAE).<\/p>\n","answers":[]}
+{"QuestionId":57735575,"AnswerCount":0,"Tags":"<python-3.x><machine-learning><keras><initialization>","CreationDate":"2019-08-31T06:02:19.240","AcceptedAnswerId":null,"OwnerUserId":5894206.0,"Title":"Keras RandomUniform not random or uniform?","Body":"<p>I initialize each layer of my network with Keras' RandomUniform, but when I predict on it the outputs don't look random at all<\/p>\n\n<p>the network<\/p>\n\n<pre><code>    def __init__(self):\n        inputs = Input(shape=(6,7))\n        net = Reshape((6,7,1))(inputs)\n        for i in range(4):\n            net = Conv2D(64, kernel_size=4, strides=1, padding='same', \n                    kernel_initializer=RandomUniform() )(net)\n            net = BatchNormalization(axis=3)(net)\n            net = Activation('relu')(net)\n        net = Flatten()(net)\n        for i in range(2):\n            net = Dense(int(1024\/(4**i)), \n                    kernel_initializer=RandomUniform()  )(net)\n            net = BatchNormalization(axis=1)(net)\n            net = Activation('relu')(net)\n            net = Dropout(rate=0.3)(net)\n        Q = Dense(1, activation='sigmoid', \n                kernel_initializer=RandomUniform(minval=-1, maxval=1), name='Q')(net)\n        pi = Dense(7, activation='softmax', \n                kernel_initializer=RandomUniform(minval=0, maxval=1), name='pi')(net)\n\n        self.model = Model(inputs=inputs, outputs=[Q, pi])\n        self.model.compile(loss=['mean_squared_error','categorical_crossentropy'], \n                           optimizer=Adam(0.001))\n\n<\/code><\/pre>\n\n<p>with inputs like <\/p>\n\n<pre><code>array([[[2, 2, 2, 2, 2, 2, 2],\n        [2, 2, 2, 2, 2, 2, 2],\n        [2, 2, 2, 2, 2, 2, 2],\n        [2, 2, 2, 2, 2, 2, 2],\n        [2, 2, 2, 2, 2, 1, 2],\n        [1, 0, 2, 2, 1, 0, 2]]])\n<\/code><\/pre>\n\n<p>to represent a game state like <\/p>\n\n<pre><code>    |   |   |   |   |   |   |   |\n    |   |   |   |   |   |   |   |\n    |   |   |   |   |   |   |   |\n    |   |   |   |   |   |   |   |\n    |   |   |   |   |   | O |   |\n    | O | X |   |   | O | X |   |\n      _   _   _   _   _   _   _\n      0   1   2   3   4   5   6\n<\/code><\/pre>\n\n<p>and a bunch of predictions from an untrained model with board states of a game <code>Q, pi = net.model.predict_on_batch(board)<\/code> and printed like <code>print(f'Qr: {self.Q_net:.1f}, ')\n<\/code><\/p>\n\n<pre><code>Q:  0.53    0.52    0.53    0.52    0.53    0.52    0.52    policy:   0.132  0.150  0.135  0.150  0.141  0.143  0.149\nQ:  0.54    0.53    0.54    0.52    0.53    0.53    0.53    policy:   0.132  0.149  0.135  0.150  0.143  0.141  0.150\nQ:  0.53    0.53    0.53    0.53    0.52    0.53    0.53    policy:   0.131  0.151  0.136  0.150  0.142  0.141  0.149\nQ:  0.53    0.53    0.53    0.53    0.52    0.52    0.53    policy:   0.133  0.152  0.137  0.149  0.141  0.142  0.146\nQ:  0.52    0.52    0.52    0.53    0.53    0.52    0.52    policy:   0.133  0.153  0.136  0.150  0.142  0.141  0.146\nQ:  0.52    0.51    0.52    0.52    0.52    0.51    0.51    policy:   0.135  0.150  0.136  0.150  0.141  0.139  0.148\nQ:  0.52    0.52    0.52    0.52    0.52    0.52    0.52    policy:   0.136  0.149  0.137  0.151  0.142  0.140  0.146\nQ:  0.52    0.52    0.51    0.52    0.52    0.51    0.52    policy:   0.135  0.149  0.137  0.153  0.141  0.136  0.148\nQ:  0.52    0.52    0.52    0.52    0.52    0.52    0.52    policy:   0.135  0.152  0.136  0.152  0.141  0.136  0.147\nQ:  0.52    0.52    0.52    0.53    0.52    0.52    0.52    policy:   0.135  0.152  0.135  0.151  0.141  0.137  0.149\nQ:  0.00    0.00    0.00    0.00    0.00    0.00    0.00    policy:   0.000  0.000  0.000  0.000  0.000  0.000  0.000\nQ:  0.00    0.00    0.00    0.00    0.00    0.00    0.00    policy:   0.000  0.000  0.000  0.000  0.000  0.000  0.000\nQ:  0.00    0.00    0.00    0.00    0.00    0.00    0.00    policy:   0.000  0.000  0.000  0.000  0.000  0.000  0.000\n\n<\/code><\/pre>\n\n<p>I was expecting predictions on an untrained randomly initialized network to look uniform from (-1,1) or (0,1). is this not how you initialize a Keras network?<\/p>\n","answers":[]}
+{"QuestionId":57735701,"AnswerCount":1,"Tags":"<python><jupyter-notebook><pytorch>","CreationDate":"2019-08-31T06:25:34.277","AcceptedAnswerId":null,"OwnerUserId":10068494.0,"Title":"Can't import torch in jupyter notebook","Body":"<p>System: macOS 10.13.6\nPython: 3.7\nAnaconda3<\/p>\n\n<p>I have trouble when <code>import torch<\/code> in jupyter notebook.<\/p>\n\n<pre><code>ModuleNotFoundError: No module named 'torch'\n<\/code><\/pre>\n\n<p>Here is how I install pytorch:<\/p>\n\n<pre><code>conda install pytorch torchvision -c pytorch\n<\/code><\/pre>\n\n<p>I've checked PyTorch is installed in my anaconda environment:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/5cSxk.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5cSxk.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>When I command <code>python3<\/code> in my terminal and <code>import torch<\/code>, it works. But not work in jupyter notebook<\/p>\n\n<p>I've tried:<\/p>\n\n<pre><code>conda update conda\nconda install mkl=2018\n<\/code><\/pre>\n\n<p>But still the same error.<\/p>\n\n<p>Can anyone help me with this? Thanks!!<\/p>\n","answers":[{"AnswerId":"57753725","CreationDate":"2019-09-02T08:32:32.793","ParentId":null,"OwnerUserId":"10749432","Title":null,"Body":"<p>You have to install <code>jupyter<\/code> in addition to <code>pytorch<\/code> inside your activated conda env. Here is installation steps:<\/p>\n\n<h2>1. Create conda env<\/h2>\n\n<p>for example: <code>pytorch_p37<\/code> with <code>python 3.7<\/code>:<\/p>\n\n<pre><code>user@pc:~$ conda create -n pytorch_p37 python=3.7\n<\/code><\/pre>\n\n<h2>2. Activate it<\/h2>\n\n<pre><code>user@pc:~$ conda activate pytorch_p37\n<\/code><\/pre>\n\n<p>Or with (for older <code>conda<\/code> versions):<\/p>\n\n<pre><code>user@pc:~$ source activate pytorch_p37\n<\/code><\/pre>\n\n<p>Now you should see <code>(pytorch_p37)<\/code> before your shell prompt: <\/p>\n\n<pre><code>(pytorch_p37) user@pc:~$ \n<\/code><\/pre>\n\n<h2>3. Go to <a href=\"https:\/\/pytorch.org\/\" rel=\"nofollow noreferrer\">PyTorch website<\/a> and choose appropriate installation command via <code>conda<\/code>. Run it in your shell, for example<\/h2>\n\n<pre><code>(pytorch_p37) user@pc:~$ conda install pytorch torchvision -c pytorch\n<\/code><\/pre>\n\n<h2>4. Install <code>jupyter<\/code> inside your activated env as well<\/h2>\n\n<pre><code>(pytorch_p37) user@pc:~$ conda install jupyter\n<\/code><\/pre>\n\n<h2>5. Verify the installation<\/h2>\n\n<pre><code>(pytorch_p37) user@pc:~$ conda list\n\n# packages in environment at \/home\/user\/anaconda3\/envs\/pytorch_p37:\n#\n# Name\n...\njupyter                   1.0.0  \njupyter_client            5.3.1  \njupyter_console           6.0.0 \njupyter_core              4.5.0\n...\npython                    3.7.4\npytorch                   1.2.0\n...\n<\/code><\/pre>\n\n<h2>6. Run <code>jupyter<\/code><\/h2>\n\n<pre><code>(pytorch_p37) user@pc:~$ jupyter notebook\n<\/code><\/pre>\n"}]}
+{"QuestionId":57736145,"AnswerCount":0,"Tags":"<python><keras><autoencoder><loss-function>","CreationDate":"2019-08-31T07:44:45.603","AcceptedAnswerId":null,"OwnerUserId":12002289.0,"Title":"How to restrict or constrain the average \"energy\" of output of a layer?","Body":"<p>I'm trying to train my autoencoder's encoder output to be within a fixed energy, such that for a 2D plane, the output x1^2 + x2^2 is &lt;= 2. I couldn't think of how to do it through a normalization layer, because the two outputs are dependent, so I tried to create a custom loss function.<\/p>\n\n<pre><code>def custom_loss(signal):\n     def loss(y_true, y_pred):\n         sloss =  K.switch(K.greater((K.sum(K.square(signal))),K.constant(2)),K.sum(K.square(signal))-K.constant(n_channel),K.constant(0))\n         return sloss+keras.losses.categorical_crossentropy(y_true, y_pred)\n     return loss\n<\/code><\/pre>\n\n<p>However, when I tried to make a custom loss function, I expected my outputs for 16 different one-hot-vector inputs to be spaced throughout a radius of 2, but actually all took on values near 0 instead. What is wrong?<\/p>\n","answers":[]}
+{"QuestionId":57736407,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2019-08-31T08:28:13.257","AcceptedAnswerId":null,"OwnerUserId":11695259.0,"Title":"pip list cmd shows tensorflow installed but cant import it","Body":"<p>I have created a virtual environment with python 3.6 and installed various packages using pip, when i do <code>pip install tensorflow<\/code> it says \"Requirement already satisfied\". However when i import it, I get No module name tensorflow.python<\/p>\n","answers":[{"AnswerId":"57890422","CreationDate":"2019-09-11T13:34:24.227","ParentId":null,"OwnerUserId":"11695259","Title":null,"Body":"<p>And as every guy should know the very first solution of every problem...\nUninstall and reinstall\nYep, It worked.<\/p>\n"}]}
+{"QuestionId":57736416,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch><recurrent-neural-network>","CreationDate":"2019-08-31T08:30:12.787","AcceptedAnswerId":null,"OwnerUserId":12002420.0,"Title":"How to fix error while using DataParallel in GRU's network","Body":"<p>I'm trying to do dataparallel into GRU's network as explained in the docs and I keep getting the same error<\/p>\n\n<pre><code>\"\"\"Defines the neural network, losss function and metrics\"\"\"\n\nimport numpy as np\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\n\nclass Net(nn.Module):\n    def __init__(self, params, anchor_is_phrase):\n        \"\"\"\n        Simple LSTM, used to generate the LSTM for both the word and video\n        embeddings. \n        Args:\n            params: (Params) contains vocab_size, embedding_dim, lstm_hidden_dim\n            is_phrase: is word lstm or the vid lstm\n        \"\"\"\n        super(Net, self).__init__()\n        if anchor_is_phrase:\n            self.lstm = nn.DataParallel(nn.GRU(params.word_embedding_dim, params.hidden_dim, 1)).cuda()#, batch_first=True)\n        else:\n            self.lstm = nn.DataParallel(nn.GRU(params.vid_embedding_dim, params.hidden_dim, 1)).cuda() #, batch_first=True)\n\n    def forward(self, s, anchor_is_phrase = False):\n        \"\"\"\n        Forward prop. \n        \"\"\"\n        s, _ = self.lstm(s)\n        s.data.contiguous()\n        return s\n<\/code><\/pre>\n\n<p>The error happens at line  s, _ = self.lstm(s) in the previous code :<\/p>\n\n<pre><code> here:  s, _ = self.lstm(s)\n        s.data.contiguous()\n        return s\n<\/code><\/pre>\n\n<p>I get the following error message :<\/p>\n\n<pre><code>    s, _ = self.lstm(s)\n  File \"\/home\/pavelameen\/miniconda3\/envs\/TD2\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/module.py\", line 493, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/home\/pavelameen\/miniconda3\/envs\/TD2\/lib\/python3.6\/site-packages\/torch\/nn\/parallel\/data_parallel.py\", line 152, in forward\n    outputs = self.parallel_apply(replicas, inputs, kwargs)\n  File \"\/home\/pavelameen\/miniconda3\/envs\/TD2\/lib\/python3.6\/site-packages\/torch\/nn\/parallel\/data_parallel.py\", line 162, in parallel_apply\n    return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)])\n  File \"\/home\/pavelameen\/miniconda3\/envs\/TD2\/lib\/python3.6\/site-packages\/torch\/nn\/parallel\/parallel_apply.py\", line 83, in parallel_apply\n    raise output\n  File \"\/home\/pavelameen\/miniconda3\/envs\/TD2\/lib\/python3.6\/site-packages\/torch\/nn\/parallel\/parallel_apply.py\", line 59, in _worker\n    output = module(*input, **kwargs)\n  File \"\/home\/pavelameen\/miniconda3\/envs\/TD2\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/module.py\", line 493, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/home\/pavelameen\/miniconda3\/envs\/TD2\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/rnn.py\", line 193, in forward\n    max_batch_size = input.size(0) if self.batch_first else input.size(1)\nAttributeError: 'tuple' object has no attribute 'size'\n<\/code><\/pre>\n\n<p>the interesting part is i try to output the type of s in line 27 and i get <strong>PackedSequence<\/strong> , why it convert to tuple in lstm forward method?<\/p>\n","answers":[{"AnswerId":"57739728","CreationDate":"2019-08-31T16:19:39.163","ParentId":null,"OwnerUserId":"5027315","Title":null,"Body":"<p><code>nn.GRU<\/code> <a href=\"https:\/\/github.com\/pytorch\/pytorch\/blob\/master\/torch\/nn\/modules\/rnn.py\" rel=\"nofollow noreferrer\">expects<\/a> (line 181) either a <a href=\"https:\/\/github.com\/pytorch\/pytorch\/blob\/60c4e74e49770ca7204b812006926ac0778d30df\/torch\/nn\/utils\/rnn.py#L23\" rel=\"nofollow noreferrer\"><code>PackedSequence<\/code><\/a> or a tesnor as input. As mentioned in the error, you are passing a tuple object <code>s<\/code> intead.<\/p>\n"}]}
+{"QuestionId":57736488,"AnswerCount":1,"Tags":"<python><keras><neural-network><mnist>","CreationDate":"2019-08-31T08:46:23.930","AcceptedAnswerId":57740385.0,"OwnerUserId":6136725.0,"Title":"Error when checking input: expected dense_8_input to have 2 dimensions","Body":"<p>I am trying to run a simple network for classifying MNIST:<\/p>\n\n<pre><code>network = models.Sequential()\nnetwork.add(layers.Dense(512, activation = 'relu', input_shape=(28 * 28,)))\nnetwork.add(layers.Dense(10, activation='softmax'))\n<\/code><\/pre>\n\n<p>When I try to fit:<\/p>\n\n<pre><code>network.fit(train_images, train_labels, epochs =5, batch_size = 128)\n<\/code><\/pre>\n\n<p>I get this error:<\/p>\n\n<blockquote>\n  <p>Error when checking input: expected dense_8_input to have 2\n  dimensions, but got array with shape (60000, 28, 28)<\/p>\n<\/blockquote>\n\n<p>What am I doing worng?<\/p>\n","answers":[{"AnswerId":"57740385","CreationDate":"2019-08-31T17:50:00.247","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>Your model expects each input sample to have a shape of <code>(784,)<\/code> (i.e. <code>input_shape=(28 * 28,)<\/code>). However, as the error shows, the input array currently has a shape of <code>(num_samples, 28, 28)<\/code>. So you need to reshape it:<\/p>\n\n<pre><code>import numpy as np\n\ntrain_images = np.reshape(train_images, (-1, 28*28))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57736680,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><lstm>","CreationDate":"2019-08-31T09:17:02.003","AcceptedAnswerId":57759709.0,"OwnerUserId":11271299.0,"Title":"How to solve \" NameError: name 'model' is not defined \" error?","Body":"<p>While trying to predict the output I am facing the error <code>NameError: name 'model' is not defined<\/code>. How to solve this.<\/p>\n\n<pre><code>%%time\n# Lstm\nmodel = Sequential()\nmodel.add(LSTM(data_dim, input_shape=(95,data_dim), activation='relu'))\nmodel.add(Dense(data_dim))\nmodel.compile(loss='mse', optimizer='adam')\nmodel.fit(X_train, y_train, epochs=10, batch_size=96)\nmodel.summary()\n<\/code><\/pre>\n\n<p>The above model trained well. While trying <code>model.predict(X_test1)<\/code>, I am having the issue mentioned above.<\/p>\n","answers":[{"AnswerId":"57759709","CreationDate":"2019-09-02T15:43:42.883","ParentId":null,"OwnerUserId":"11394329","Title":null,"Body":"<p>The issue lies in the magic function <code>%%time<\/code>. In the latest version of IPython in Jupyter, running a cell with <code>time<\/code> magic function as a header runs the cell out of the global context. This is also true for <code>%%timeit<\/code>.<\/p>\n\n<p>Practically it means that all the new variables defined in the <code>%%time<\/code> cell do not exist in the main context, including your <code>model<\/code> variable, which is why you receive the NameError exception, since the interpreter can not find a variable named <code>model<\/code>. <\/p>\n\n<p>Removing the <code>%%time<\/code> line from your cell will do the trick.<\/p>\n"}]}
+{"QuestionId":57736792,"AnswerCount":1,"Tags":"<tensorflow><keras><dot-product>","CreationDate":"2019-08-31T09:31:39.553","AcceptedAnswerId":57738491.0,"OwnerUserId":6134679.0,"Title":"no attribute '_inbound_nodes' error even when using Lambda layer in Keras","Body":"<p>I have a <strong>(28,000 x 300)<\/strong> dimension matrix, let's call it <strong>label_embedding<\/strong>, which I want to do a dot product with the bottleneck layer of my model. I have created an architecture which gives a <strong>(batch_size x 300)<\/strong> at its bottleneck layer.<\/p>\n\n<p>I am using a generator function for input. The label_embedding matrix is taken as input in the following way:<\/p>\n\n<pre><code>inp7 = Input(tensor=labels_embeddings)\n<\/code><\/pre>\n\n<p>And for the dot product I am doing the following:<\/p>\n\n<pre><code>out = Lambda(dot_)([x1, K.transpose(inp7)])\n<\/code><\/pre>\n\n<p>where <strong>x1<\/strong> is the bottleneck layer and <strong>dot_<\/strong> is:<\/p>\n\n<pre><code>def dot_(tensors):\n    return K.dot(tensors[0], tensors[1])\n<\/code><\/pre>\n\n<p>The problem is that even though the shape of the <strong>out<\/strong> variable is correct, that is, <strong>(batch_size x 28000)<\/strong>, I get the following error:<\/p>\n\n<pre><code>AttributeError: 'NoneType' object has no attribute '_inbound_nodes'\n<\/code><\/pre>\n\n<p>P.S.: I am using tensorflow and keras<\/p>\n\n<p>P.S.: I have been using keras layers until the out variable where I use keras backend as K<\/p>\n","answers":[{"AnswerId":"57738491","CreationDate":"2019-08-31T13:39:35.243","ParentId":null,"OwnerUserId":"6134679","Title":null,"Body":"<p>Okay, I solved the problem. So all backend based functions need to be wrapped within the lambda layer. So instead of:<\/p>\n\n<pre><code>out = Lambda(dot_)([x1, K.transpose(inp7)])\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>def dot_(tensors):\n    return K.dot(tensors[0], tensors[1])\n<\/code><\/pre>\n\n<p>I wrote:<\/p>\n\n<pre><code>out = Lambda(dot_)([x1, inp7])\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>def dot_(tensors):\n    return K.dot(tensors[0], K.transpose(tensors[1]))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57736960,"AnswerCount":0,"Tags":"<python><tensorflow><ubuntu-18.04><bazel><avx>","CreationDate":"2019-08-31T09:55:48.793","AcceptedAnswerId":null,"OwnerUserId":6509922.0,"Title":"Trouble building Tensorflow 2.0 from source","Body":"<p>I'm trying to push some tensorflow 2.0 models onto an Ubuntu 18.04 server. The server's cpu's don't support AVX, so as per the community on github, I'll need to build from source to build tf 2.0 without AVX. I've spent a week trying to build from source, and the build complete's successfully, but I still keep getting the same \"Illegal Instruction\" error when trying to import tensorflow into a python notebook.<\/p>\n\n<p>I'm running python 3.6 on Ubuntu 18.04<\/p>\n\n<p>the bazel configuration is with XLA and JIT but and No to the remaining options.<\/p>\n\n<p>copt flags I'm using are:<\/p>\n\n<pre><code>-march=native -Wno-sign-compare -mssse3 -mcx16 -msse4.1 -msse4.2 -mpopcnt -mno-avx\n<\/code><\/pre>\n\n<p>The build command is:<\/p>\n\n<pre><code>bazel build --config=opt  \/\/tensorflow\/tools\/pip_package:build_pip_package --local_ram_resources=1024 --jobs=1 --define=tf_api_version=2\n<\/code><\/pre>\n\n<p>the server's CPU's are:<\/p>\n\n<pre><code>sudo lshw | grep -i cpu\n\n     *-cpu:0\n          description: CPU\n          product: QEMU Virtual CPU version 2.5+\n          bus info: cpu@0\n          slot: CPU 0\n          capabilities: fpu fpu_exception wp de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pse36 clflush mmx fxsr sse sse2 syscall nx x86-64 rep_good nopl xtopology cpuid pni cx16 x2apic hypervisor lahf_lm pti\n     *-cpu:1\n          description: CPU\n          product: QEMU Virtual CPU version 2.5+\n          bus info: cpu@1\n          slot: CPU 1\n          capabilities: fpu fpu_exception wp de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pse36 clflush mmx fxsr sse sse2 syscall nx x86-64 rep_good nopl xtopology cpuid pni cx16 x2apic hypervisor lahf_lm pti\n     *-cpu:2\n          description: CPU\n          product: QEMU Virtual CPU version 2.5+\n          bus info: cpu@2\n          slot: CPU 2\n          capabilities: fpu fpu_exception wp de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pse36 clflush mmx fxsr sse sse2 syscall nx x86-64 rep_good nopl xtopology cpuid pni cx16 x2apic hypervisor lahf_lm pti\n     *-cpu:3\n          description: CPU\n          product: QEMU Virtual CPU version 2.5+\n          bus info: cpu@3\n          slot: CPU 3\n          capabilities: fpu fpu_exception wp de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pse36 clflush mmx fxsr sse sse2 syscall nx x86-64 rep_good nopl xtopology cpuid pni cx16 x2apic hypervisor lahf_lm pti\n<\/code><\/pre>\n\n<p>This is the first time I've had to dig so deeply into questions of hardware and I'm at a loss. <\/p>\n\n<p>Firstly, is it possible to build tensorflow to run on a QEMU Virtual CPU version 2.5+? <\/p>\n\n<p>Secondly, if so, what am I doing wrong?<\/p>\n\n<p>Thanks<\/p>\n","answers":[]}
+{"QuestionId":57737176,"AnswerCount":0,"Tags":"<python><tensorflow><keras><tensorflow2.0><tensorflow-probability>","CreationDate":"2019-08-31T10:25:17.650","AcceptedAnswerId":null,"OwnerUserId":1106578.0,"Title":"Share Weights among layers in TensorFlow 2.0","Body":"<p>I am trying to ascertain how one would go about sharing weights among a number of copies of a particular type of layer that are for instance stacked in a sequential keras model, when using Tensorflow 2.0<\/p>\n\n<p>Specifically from tensorflow probability a tfp.layers.VariationalGaussianProcess layer. One of its inputs is a kernel_provider (which you make by subclassing a keras layer).<\/p>\n\n<p>What I would like to do is use a tfp.positive_semidefinite_kernels.MaternOneHalf kernel, which has amplitude and length_scale weights. Then for all the VariationalGaussianProcess layers in the model, for the amplitude and length_scale weights to be shared across all these layers.<\/p>\n","answers":[]}
+{"QuestionId":57737315,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><tensorboard><tf.keras>","CreationDate":"2019-08-31T10:45:07.527","AcceptedAnswerId":null,"OwnerUserId":1616773.0,"Title":"While creating a plugin for tensorboard with TF.keras , where to define a plugin name?","Body":"<p>I created a tensorboard plugin with Tf.layer CNN model and its quite simple, store summary in logs with \"plugin name\" and while running tensor board, fetch data from logs based on the \"plugin name\".<\/p>\n\n<p>When I try to do the same with Keras model, session and epochs and even log saving are handled in callbacks. I can not figure out how to \"save weights in logs along with plugin name \"<\/p>\n\n<p>In main file (<a href=\"https:\/\/github.com\/tensorflow\/tensorboard-plugin-example\/blob\/master\/greeter_plugin\/greeter_demo.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorboard-plugin-example\/blob\/master\/greeter_plugin\/greeter_demo.py<\/a>)<\/p>\n\n<pre><code>summary = greeter_summary.pb(\"greetings\", extra_character)\n<\/code><\/pre>\n\n<p>In summary file <\/p>\n\n<pre><code>tensor_summaries = list(map(lambda graph_node:\n                                tf.summary.tensor_summary(graph_node.name,\n                                                          graph.get_tensor_by_name(graph_node.name + \":0\"),\n                                                          summary_metadata=tf.SummaryMetadata(\n                                                              display_name=graph_node.name,\n                                                              summary_description=description,\n                                                              plugin_data=tf.SummaryMetadata.PluginData(\n                                                                  plugin_name=PLUGIN_NAME)),\n                                                          collections=collections),\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57737586,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><tensorflow-probability>","CreationDate":"2019-08-31T11:24:06.807","AcceptedAnswerId":57853720.0,"OwnerUserId":10675934.0,"Title":"Creating a custom gradient function for HMC for large datasets","Body":"<p>I'm trying to infer the parameters of a Gaussian process using HMC in tensorflow-probability. <\/p>\n\n<p>I have multiple independent data sequences that are generated from the same underlying process and I want to infer the kernel parameters that they all share. <\/p>\n\n<p>To calculate the likelihood I'm using eager mode and looping over each independent sequence. I'm able to calculate the likelihood but I run into resource exhausted errors when trying to calculate the gradient. <\/p>\n\n<p>I know this will be very slow but I would like to be able to use HMC with any size dataset without running out of memory.<\/p>\n\n<p>I use the following code to create synthetic data, this creates N samples from a GP with p data points.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>L = 5\nvariance=2\nm_noise=0.05\n\nkernel=psd_kernels.ExponentiatedQuadratic(np.float64(variance), np.float64(L))\n\n\ndef gram_matrix(xs):\n    return kernel.matrix(xs,xs).numpy() + m_noise*np.identity(xs.shape[0])\n\n\nobservation_index_points = []\nobservations = []\nN=200\np = 2000\nfor i in range(0, N):\n\n    xs = np.sort(np.random.uniform(0,100,p))[...,None]\n    mean = [0 for x in xs]\n    gram = gram_matrix(xs)\n\n    ys = np.random.multivariate_normal(mean, gram)\n\n    observation_index_points.append(xs)\n\n    observations.append(ys)\n\nfor i in range(0, N):\n\n    plt.plot(observation_index_points[i],observations[i])\nplt.show()\n<\/code><\/pre>\n\n<p>The following code to calculate the log likelihood will run with the sampler for small values of N but fails with larger values of N (resource exhausted). The error occurs when trying to calculate the gradient of the likelihood.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>@tf.function()\ndef gp_log_prob(amplitude, length_scale, seg_index_points, noise_variance, seg_observations):\n\n    kernel = psd_kernels.ExponentiatedQuadratic(amplitude, length_scale)\n    gp = tfd.GaussianProcess(kernel=kernel,\n                                index_points=seg_index_points,\n                                observation_noise_variance=noise_variance)\n\n\n    return gp.log_prob(seg_observations)\n\nrv_amplitude = tfd.LogNormal(np.float64(0.), np.float64(1))\nrv_length_scale = tfd.LogNormal(np.float64(0.), np.float64(1))\nrv_noise_variance = tfd.LogNormal(np.float64(0.), np.float64(1))\n\ndef joint_log_prob_no_grad(amplitude, length_scale, noise_variance):\n    ret_val = rv_amplitude.log_prob(amplitude) \\\n                + rv_length_scale.log_prob(length_scale) \\\n                + rv_noise_variance.log_prob(noise_variance)\n\n    for i in range(N):\n        ret_val = ret_val + gp_log_prob(amplitude, \n                                        length_scale, \n                                        observation_index_points[i], \n                                        noise_variance, \n                                        observations[i])\n\n    return ret_val\n\n<\/code><\/pre>\n\n<p>BUT I can calculate the gradient for large N using gradienttape inside the loop. This code runs for any N and returns the correct likelihood and gradient:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def joint_log_prob(amplitude, length_scale, noise_variance):\n\n    with tf.GradientTape() as tape:\n        tape.watch(amplitude)\n        tape.watch(length_scale)\n        tape.watch(noise_variance)\n\n        ret_val = rv_amplitude.log_prob(amplitude) \\\n                    + rv_length_scale.log_prob(length_scale) \\\n                    + rv_noise_variance.log_prob(noise_variance)\n\n    grads = tape.gradient(ret_val, [amplitude, length_scale, noise_variance])\n\n    for i in range(N):\n        with tf.GradientTape() as tape:\n            tape.watch([amplitude, length_scale, noise_variance])\n            gp_prob = gp_log_prob(amplitude, length_scale, \n                                  observation_index_points[i], noise_variance, observations[i])\n\n        gp_grads = tape.gradient(gp_prob, [amplitude, length_scale, noise_variance])\n\n\n        grads = [a+b for a,b in zip(grads,gp_grads)]\n        ret_val = ret_val + gp_prob\n\n    return ret_val, grads\n\nx = tf.convert_to_tensor(np.float64(1.0))\ny = tf.convert_to_tensor(np.float64(1.0))\nz = tf.convert_to_tensor(np.float64(0.1))\n\njoint_log_prob(x,y,z) # correct output even for large N\n<\/code><\/pre>\n\n<p>If I then turn this into a customgradient it fails again:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>@tf.custom_gradient\ndef joint_log_prob_cg(amplitude, length_scale, noise_variance):\n\n    with tf.GradientTape() as tape:\n        tape.watch(amplitude)\n        tape.watch(length_scale)\n        tape.watch(noise_variance)\n\n        ret_val = rv_amplitude.log_prob(amplitude) \\\n                    + rv_length_scale.log_prob(length_scale) \\\n                    + rv_noise_variance.log_prob(noise_variance)\n\n    grads = tape.gradient(ret_val, [amplitude, length_scale, noise_variance])\n\n    for i in range(N):\n        with tf.GradientTape() as tape:\n            tape.watch([amplitude, length_scale, noise_variance])\n            gp_prob = gp_log_prob(amplitude, length_scale, \n                                  observation_index_points[i], noise_variance, observations[i])\n\n        gp_grads = tape.gradient(gp_prob, [amplitude, length_scale, noise_variance])\n\n\n        grads = [a+b for a,b in zip(grads,gp_grads)]\n        ret_val = ret_val + gp_prob\n\n    def grad(dy):\n        return grads\n    return ret_val, grad\n\nwith tf.GradientTape() as t:\n    t.watch([x,y,z])\n    lp = joint_log_prob_cg(x,y,z)\n\nt.gradient(lp, [x,y,z]) # fails for large N\n<\/code><\/pre>\n\n<p>My question is how can I get <code>grads<\/code> from the <code>joint_log_prob<\/code> function above (which I know can be computed for any large dataset) into the HMC sampler? It seems that if the whole function is wrapped in the gradienttape call then the <code>for<\/code> loop is unrolled and it runs out of memory - but is there a way around this?<\/p>\n","answers":[{"AnswerId":"57853720","CreationDate":"2019-09-09T12:06:51.893","ParentId":null,"OwnerUserId":"10675934","Title":null,"Body":"<p>In case anyone is interested, I was able to solve this by using a custom gradient and stopping the tape recording in the <code>for<\/code> loop. I needed to import the tape utilities:<\/p>\n\n<pre><code>from tensorflow.python.eager import tape\n<\/code><\/pre>\n\n<p>then stop recording around the <code>for<\/code> loop<\/p>\n\n<pre><code>with tape.stop_recording():\n    for i in range(N):\n        ...\n<\/code><\/pre>\n\n<p>This stops tracing, I then have to compute the gradient in graph mode and add them manually to stop the OOM error.<\/p>\n"}]}
+{"QuestionId":57737674,"AnswerCount":2,"Tags":"<python-3.x><keras><neural-network>","CreationDate":"2019-08-31T11:35:14.637","AcceptedAnswerId":57763134.0,"OwnerUserId":7416201.0,"Title":"Loss is decreasing however so is accuracy for keras neural network model","Body":"<p>(I'm relatively new to neural networks and working with data so sorry if I'm doing something stupid.)<\/p>\n\n<p>I'm trying to evaluate deep_mojis ability to generalise by trying to use it on different datasets. I'm using it on stanford SST-fine and trying to map the emoji scores to a sentiment score using a simple softmax network with one hidden layer.<\/p>\n\n<p>From what I've read on the stanford SST, I should be converting a score to a vector with intervals of 0.2. for example 0.7 should go to [0,0,0,1,0]. <\/p>\n\n<p>However when I try to train the training loss and accuracy both decrease as well as the accuracy.<\/p>\n\n<p>NN model<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class sst_fine_predictor:\n    def build_model(self):\n        self.model = tf.keras.Sequential()\n        self.model.add(layers.Dense(128,input_shape=(64, ), activation='tanh', kernel_regularizer=tf.keras.regularizers.l2(0.01)))\n        self.model.add(\n            layers.Dense(128, activation='tanh', kernel_regularizer=tf.keras.regularizers.l2(0.01)))\n        self.model.add(layers.Dense(5, activation='softmax', kernel_regularizer=tf.keras.regularizers.l2(0.01)))\n        self.model.compile(optimizer=tf.keras.optimizers.Adagrad(lr=0.01),\n                           loss='categorical_crossentropy',\n                           metrics=['accuracy'])\n        return\n\n    def __init__(self):\n        self.build_model()\n        self.stopper_1 = EarlyStopping(monitor='val_loss', mode='min', verbose=1, patience=50)\n        # self.stopper_2 = EarlyStopping(monitor='val_acc', mode='min', verbose=1, patience=200)\n        self.checkpointer = ModelCheckpoint(SAVE_PATH, monitor='val_acc', mode='max', save_best_only=True)\n\n    def __call__(self):\n        return self.model\n\n    def train(self, train_ins, train_outs, test_ins, test_outs, max_epochs=4000):\n        history = self().fit(train_ins, train_outs, validation_data=(test_ins, test_outs), epochs=max_epochs, verbose=1,\n                             callbacks=[self.stopper_1, self.checkpointer], batch_size=8117)\n        return history\n\n[accuracy graph][1][loss graph][1]\n\n<\/code><\/pre>\n\n<p>training_code<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>training_data = pd.read_csv(SST_PATH + \"emoji_train.csv\", sep=\"|\")\ntest_data = pd.read_csv(SST_PATH + \"emoji_test.csv\", sep=\"|\")\n\ntrain_ins = np.delete(training_data.values, [0,1,2], axis=1)\ntrain_outs = categorize(training_data.values[:, 2], discrete)\n\ncorrectly classified\ntest_ins = np.delete(test_data.values, [0,1,2], axis=1)\ntest_outs = categorize(test_data.values[:, 2], discrete)\n\nmodel = sst_fine_predictor()\n\n\nprint(\"training model\")\nhistory = model.train(train_ins, train_outs, test_ins, test_outs, max_epochs=4000)\n<\/code><\/pre>\n\n<p>emoji_train.csv and test I'm quite certain has no issues and when loaded has these columns<\/p>\n\n<pre><code>Index(['Unnamed: 0', 'sentence', 'sentiment_score', '0', '1', '2', '3', '4',\n       '5', '6', '7', '8', '9', '10', '11', '12', '13', '14', '15', '16', '17',\n       '18', '19', '20', '21', '22', '23', '24', '25', '26', '27', '28', '29',\n       '30', '31', '32', '33', '34', '35', '36', '37', '38', '39', '40', '41',\n       '42', '43', '44', '45', '46', '47', '48', '49', '50', '51', '52', '53',\n       '54', '55', '56', '57', '58', '59', '60', '61', '62', '63'],\n      dtype='object')\n<\/code><\/pre>\n\n<p>with 0 to 63 representing the emoji scores<\/p>\n\n<p>accuracy graph\n<a href=\"https:\/\/i.stack.imgur.com\/rPnPS.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rPnPS.png\" alt=\"accuracy graph\"><\/a><\/p>\n\n<p>loss graph\n<a href=\"https:\/\/i.stack.imgur.com\/0jfzD.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0jfzD.png\" alt=\"loss graph\"><\/a><\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"57737764","CreationDate":"2019-08-31T11:48:00.357","ParentId":null,"OwnerUserId":"7339624","Title":null,"Body":"<p>When the loss is decreasing and so the accuracy you are facing with <a href=\"https:\/\/en.wikipedia.org\/wiki\/Overfitting\" rel=\"nofollow noreferrer\">overfitting<\/a>.<\/p>\n\n<p>you can avoid it by things like making your data bigger, <a href=\"https:\/\/machinelearningmastery.com\/dropout-for-regularizing-deep-neural-networks\/\" rel=\"nofollow noreferrer\">do dropout<\/a> or simplifying the model. here is a good post on that: <a href=\"https:\/\/towardsdatascience.com\/dont-overfit-how-to-prevent-overfitting-in-your-deep-learning-models-63274e552323\" rel=\"nofollow noreferrer\">How to prevent Overfitting in your Deep Learning Models<\/a>. <\/p>\n"},{"AnswerId":"57763134","CreationDate":"2019-09-02T22:15:22.900","ParentId":null,"OwnerUserId":"7416201","Title":null,"Body":"<p>(Just in case anyone does need this in the future), I'm pretty sure the issue is that I'd set the L2 regularization too high and therefore it benefited the model to be more accurate with outputs \/ labels that required low weights to be found than it was to be accurate with all the training labels therefore making it decrease the overall accuracy as well as decreasing loss.<\/p>\n\n<p>To put it simply having the regularization too high can make it so the loss function can benefit from getting certain answers wrong if they require high weights and then focus on certain answers being correct which require low weights.<\/p>\n\n<p>(I may be wrong and if I am then please correct me, but this is my best guess and I'm fairly certain it's correct)<\/p>\n"}]}
+{"QuestionId":57737852,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><classification>","CreationDate":"2019-08-31T12:03:29.187","AcceptedAnswerId":null,"OwnerUserId":12002941.0,"Title":"Could anyone help me some ideas how can I feed batches of patches of a single image to CNNs?","Body":"<p>In particular, I have a batch of images with a batch size of 10, namely, [10,256,256,3]. Then I extract 4 patches for each image of 50x50, so the new shape would be [40,50,50,3]. Each of 4 patches represents an image corresponding to its label. After that, I would feed the new batch of [40,50,50,3] to CNNs including 4 streams of shared weighted CNN with corresponding 4 patches to make the prediction.<\/p>\n\n<p>Unfortunately, I do not figure out how I can input 4 patches to 4 streams of CNNs from the batch [40,50,50,3].\nNote that the batch in here is not used placeholder as tensorflow. It is considered a feature with the shape of [40,50,50,3].<\/p>\n","answers":[]}
+{"QuestionId":57737884,"AnswerCount":1,"Tags":"<python><keras><neural-network>","CreationDate":"2019-08-31T12:07:35.533","AcceptedAnswerId":null,"OwnerUserId":11938775.0,"Title":"Keras - Image Classification with two classes Tensorflow Python","Body":"<p>i want to classify two classes with CNN. <\/p>\n\n<ul>\n<li><strong>first class<\/strong> are objects(car, building, and so go an)<\/li>\n<li><strong>secound class<\/strong> are NoObjects(for example Background without objects)<\/li>\n<\/ul>\n\n<p>So, the question is, how can I realize this? I want to have many objects images for the CNN. I saw a example with Cat and Dog.. But I have only images for one class. The other class are the background. It is possible to create a background class??<\/p>\n\n<p>I\u00b4m very new in CNN.. <\/p>\n","answers":[{"AnswerId":"57740122","CreationDate":"2019-08-31T17:13:34.643","ParentId":null,"OwnerUserId":"10807234","Title":null,"Body":"<p>It seems that you're talking about <a href=\"https:\/\/medium.com\/nanonets\/how-to-do-image-segmentation-using-deep-learning-c673cc5862ef\" rel=\"nofollow noreferrer\">semantic segmentation<\/a>, where you have a variable N number of classes to localize (in this case N=1) plus the background class. There's a lot of literature and baseline models out there that can help you out. I recommend these two since they're the most popular ones:<\/p>\n\n<ul>\n<li>Pretrained models: <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/detection_model_zoo.md\" rel=\"nofollow noreferrer\">Tensorflow's model zoo<\/a><\/li>\n<li>Datasets: Specifically <a href=\"http:\/\/cocodataset.org\/#home\" rel=\"nofollow noreferrer\">COCO<\/a><\/li>\n<\/ul>\n"}]}
+{"QuestionId":57738069,"AnswerCount":0,"Tags":"<tensorflow><matplotlib><keras><tensorboard>","CreationDate":"2019-08-31T12:35:28.147","AcceptedAnswerId":null,"OwnerUserId":8665962.0,"Title":"Using TensorBoard to visualize error values from CSV file","Body":"<p>I have a CSV file containing loss values after training a Keras neural network model, somthing like this:<\/p>\n\n<pre><code>7.019360989332198431e-01\n6.927479231357573930e-01\n6.899647492170333241e-01\n6.840079456567764726e-01\n6.766045588254929743e-01\n9.124663196504115281e-02\n8.141199145466089526e-02\n7.456063829362392925e-02\n6.891768645495176260e-02\n6.426200993359089397e-02\n6.076486252248287923e-02\n.\n.\n.\n<\/code><\/pre>\n\n<p>I can use MS Excel to plot it, but actually I want to visualize it using TensorBoard. I know I could have done it easily via TensorBoard class but I missed this chance and I really don't want to go back to train again. Is this possible?<\/p>\n\n<p>Thank you<\/p>\n","answers":[]}
+{"QuestionId":57738199,"AnswerCount":1,"Tags":"<python><parameters><deep-learning><pytorch>","CreationDate":"2019-08-31T12:56:10.230","AcceptedAnswerId":null,"OwnerUserId":7701908.0,"Title":"What does the `model.parameters()` include?","Body":"<p>In Pytorch,<\/p>\n\n<p>What will be registered into the <code>model.parameters()<\/code>.<\/p>\n\n<p>As far as now, what I know are as belows:<\/p>\n\n<pre><code>1.  Conv layer: weight  bias\n2.  BN layers: weight(gamma)  bias(beta)\n3.  nn.Parameter() \n    such as:   self.alpha = nn.Parameter(torch.rand(10))  defined in the model.\n<\/code><\/pre>\n\n<p>My <strong>question<\/strong> is:\nAnd are there some parameters else that are registered in the <code>model.parameters()<\/code> ?<\/p>\n\n<p><strong>PS<\/strong>. The most common case for the <code>model.parameters()<\/code> is in the optimizer,\ne.g.  <a href=\"https:\/\/github.com\/pytorch\/examples\/blob\/master\/imagenet\/main.py#L171-L173\" rel=\"nofollow noreferrer\">pytorch resnet example<\/a><\/p>\n\n<pre><code>optimizer = torch.optim.SGD(model.parameters(), args.lr,\n                                momentum=args.momentum,\n                                weight_decay=args.weight_decay)\n<\/code><\/pre>\n\n<p>Thank you in advance.<\/p>\n","answers":[{"AnswerId":"57738629","CreationDate":"2019-08-31T13:56:16.290","ParentId":null,"OwnerUserId":"8047535","Title":null,"Body":"<p>Like you wrote there, <code>model.parameters()<\/code> stores the weight and bias (if set to true) values of the model.\nIt is given as an argument to an optimizer to update the weight and bias values of the model with one line of code <code>optimizer.step()<\/code>, which you then use when next you go over your dataset.<\/p>\n"}]}
+{"QuestionId":57738755,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><nlp><word2vec><bert-language-model>","CreationDate":"2019-08-31T14:13:36.727","AcceptedAnswerId":null,"OwnerUserId":11906455.0,"Title":"BERT-as-Service crashes when beginning","Body":"<p>It is my first time to use Tensorflow and Bert-as-service. I failed to start the Bert-as-service server side with the use of VS code as editor. At the same time, the Bert-as-service client side, which is on Jupyter Notebook, failed, too, as presented in the following picture. <\/p>\n\n<p>I just want to start the BERT service and Use Client to Get Sentence Encodes as sampled here. Please help me.<\/p>\n\n<p>Env: Windows 7 with RAM 4G\/Anaconda 2019.07\/Python 3.7.3\/tensorflow 1.14\/Bert-as-Service CPU version<\/p>\n\n<pre><code>(base) C:\\Users\\lee&gt;bert-serving-start -model_dir E:\/MyDownloads\/uncased_L-12_H-768_A-12 -cpu -http_port 80 -max_seq_len 20\nusage: D:\\MyPrograms\\Anaconda\\Scripts\\bert-serving-start -model_dir E:\/MyDownloads\/uncased_L-12_H-768_A-12 -cpu -http_port 80 -max_seq_len 20\n                 ARG   VALUE\n__________________________________________________\n           ckpt_name = bert_model.ckpt\n         config_name = bert_config.json\n                cors = *\n                 cpu = True\n          device_map = []\n       do_lower_case = True\n  fixed_embed_length = False\n                fp16 = False\n gpu_memory_fraction = 0.5\n       graph_tmp_dir = None\n    http_max_connect = 10\n           http_port = 80\n        mask_cls_sep = False\n      max_batch_size = 256\n         max_seq_len = 20\n           model_dir = E:\/MyDownloads\/uncased_L-12_H-768_A-12\n          num_worker = 1\n       pooling_layer = [-2]\n    pooling_strategy = REDUCE_MEAN\n                port = 5555\n            port_out = 5556\n       prefetch_size = 10\n priority_batch_size = 16\nshow_tokens_to_client = False\n     tuned_model_dir = None\n             verbose = False\n                 xla = False\n\nI:?[35mVENTILATOR?[0m:freeze, optimize and export graph, could take a while...\nWARNING:tensorflow:From d:\\myprograms\\anaconda\\lib\\site-packages\\bert_serving\\server\\helper.py:180: The name tf.logging.set_verbosity is deprecated. Please use tf.compat.v1.logging.set_verbosity instead.\n\nWARNING:tensorflow:From d:\\myprograms\\anaconda\\lib\\site-packages\\bert_serving\\server\\helper.py:180: The name tf.logging.ERROR is deprecated. Please use tf.compat.v1.logging.ERROR instead.\n\nI:?[36mGRAPHOPT?[0m:model config: E:\/MyDownloads\/uncased_L-12_H-768_A-12\\bert_config.json\nI:?[36mGRAPHOPT?[0m:checkpoint: E:\/MyDownloads\/uncased_L-12_H-768_A-12\\bert_model.ckpt\n(Then it stopped here and nothing happened that seemed to clash.)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Qdj8z.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Qdj8z.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57738842,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-08-31T14:25:59.430","AcceptedAnswerId":57748202.0,"OwnerUserId":12003219.0,"Title":"How to save an operation node's output in a graph trained by Tensorflow?","Body":"<p>I have trained a deep learning model by Tensorflow. The model is saved as a saved_model.pb and also there are variables and checkpoints saved during training. The model's task is regression. The output is a single scalar. \nI use the saved model on some new data to test the model accuracy. Now, I want to use the saved model to save the vector just before the last layer which do regression. I'm going to use that vector by different regressors to take and improve the best result. I have also visualized the graph by Tensorboard, and I realized the vector I'm looking for is the output of an operation node in Tensorflow. I attach the screenshot of the desired part of graph. \nHow it is possible to use a saved model to do that? I mean to save the output vector just before the regression layer. <a href=\"https:\/\/i.stack.imgur.com\/7c9Cw.png\" rel=\"nofollow noreferrer\">https:\/\/drive.google.com\/file\/d\/1tn8kGO6KBYSaD-Yz5Xp5OYQmdAniY84d\/view?usp=sharing<\/a>\nFurthermore, I have some codes to get the predicted single scalar. <\/p>\n\n<pre><code>my_predictor = predictor.from_saved_model(export_dir)\n\nmae = []\nfor output in read_fn(file_references=file_names,\n                      mode=tf.estimator.ModeKeys.EVAL,\n                      params=READER_PARAMS):\n\n    img = output['features']['x']\n    lbl = output['labels']['y']\n    test_id = output['img_id']\n\n    num_crop_predictions = 4\n    crop_batch = extract_random_example_array(\n        image_list=img,\n        example_size=[32, 32, 32],\n        n_examples=num_crop_predictions)\n\n    y_ = my_predictor.session.run(\n        fetches=my_predictor._fetch_tensors['logits'],\n        feed_dict={my_predictor.feed_tensors['x']: crop_batch})\n\n    y_ = np.mean(y_)\n\n    mae.append(np.abs(y_ - lbl))\n<\/code><\/pre>\n\n<p>I have changed the part<\/p>\n\n<pre><code>y_ = my_predictor.session.run(\n        fetches=my_predictor._fetch_tensors['logits'],\n        feed_dict={my_predictor.feed_tensors['x']: crop_batch})\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>y_ = my_predictor.session.run(\n        fetches=my_predictor.graph,\n        feed_dict={my_predictor.feed_tensors['x']: crop_batch})\n<\/code><\/pre>\n\n<p>due to feed the 3D inputs, and fetch the desired node of graph. But, the output of an operation node (which is what I'm looking for (pool\/global_avg_pool)) is not in the graph.<\/p>\n","answers":[{"AnswerId":"57748202","CreationDate":"2019-09-01T17:41:00.077","ParentId":null,"OwnerUserId":"2851685","Title":null,"Body":"<p>This is how I did it on a DNN learning transfer.<\/p>\n\n<p><strong>Save Trained Model<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>self._saver.save(self._session, self._get_save_path(name))\n<\/code><\/pre>\n\n<p><strong>Restore Model<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>## restore the graph\nimported_meta = tf.train.import_meta_graph(\"%s.meta\"%self._get_save_path(name))\ngraph = tf.get_default_graph()\n\n## get reference of output of operation DNN\/hidden5_out and set it to variable\n## here \":0\" indicates the output\nself._frozen_out = self._graph.get_tensor_by_name(\"DNN\/hidden5_out:0)\n\n## restore the variables in session\nself._session = tf.Session(graph=graph)\nimported_meta.restore(self._session, self._get_save_path(name))\n<\/code><\/pre>\n\n<p><strong>Calculate Frozen Output Once<\/strong><\/p>\n\n<p>I need to do this because it needs for placeholders to calculate frozen output. I think it is not necessary if you don't need to recalculate tensor value by feeding feed_dict <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tensor = sess.eval(self._frozen_out, feed_dict=feed_dict)\n<\/code><\/pre>\n\n<p><strong>Reuse the Frozen Output<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>## here dnn is an operation which uses output of operation \"DNN\/hidden5_out\"\n## I am passing the value of it directly as feed_dict so it will not go down the \n## graph to calculate it\nsess.eval(dnn, feed_dict={self._frozen_out: tensor})\n<\/code><\/pre>\n\n<p>Hope this helps, comment if it is not clear.<\/p>\n"}]}
+{"QuestionId":57738890,"AnswerCount":2,"Tags":"<image-processing><pytorch><convolution>","CreationDate":"2019-08-31T14:32:55.247","AcceptedAnswerId":57739676.0,"OwnerUserId":2828086.0,"Title":"Display result of convolution in PyTorch","Body":"<p>PyTorch newbie here. I wrote a script (code below) that performs the following operations: load an image, perform a 2D convolution operation and then display the output and the input.<\/p>\n\n<p>At present I have the image below, which seems off. How can I plot the feature map correctly?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ngNq7.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ngNq7.png\" alt=\"ted \"><\/a><\/p>\n\n<pre><code>import numpy as np\nimport torch\nimport torchvision\nimport torchvision.transforms as transforms\nimport torch.nn as nn\nimport matplotlib.pyplot as plt\nimport imageio\nimport sys\n\nA = imageio.imread('LiT.png')\n# Define how the convolution operation works\nconv2 = nn.Conv2d(in_channels=3, out_channels=3, kernel_size=3, stride=1, padding=1)\n\nimage_d = torch.FloatTensor(np.asarray(A.reshape(1, 3, A.shape[0] , A.shape[1])))\nfc = conv2(image_d)\nfc1 = fc.permute(0, 2, 3, 1).reshape([516, 780, 3])\n\nplt.figure(figsize=(16,8))\nplt.subplot(1,2,1)\nplt.imshow(A)\nplt.subplot(1,2,2)\nplt.imshow(fc1.data.numpy())\n\nplt.show()\n<\/code><\/pre>\n","answers":[{"AnswerId":"57739639","CreationDate":"2019-08-31T16:08:02.887","ParentId":null,"OwnerUserId":"5027315","Title":null,"Body":"<p>To my understanding, the problem lies in how you are permuting channels position in the image by using reshape. Instead, <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.transpose.html\" rel=\"nofollow noreferrer\"><code>'np.transpose<\/code><\/a> or <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.Tensor.permute\" rel=\"nofollow noreferrer\"><code>tensor.permute<\/code><\/a> should be used. Using torch for permutation:<\/p>\n\n<pre><code>image_d  = torch.FloatTensor(np.asarray(A)).unsqueeze(0).permute(0,3,1,2)\n<\/code><\/pre>\n\n<p>Or, if we want to handle the permutation part in numpy:<\/p>\n\n<pre><code>image_d = np.transpose(np.asarray(A), (2,0,1))\nimage_d = torch.FloatTensor(image_d).unsqueeze(0)\n<\/code><\/pre>\n"},{"AnswerId":"57739676","CreationDate":"2019-08-31T16:12:40.377","ParentId":null,"OwnerUserId":"3165451","Title":null,"Body":"<p>The issue with your code is this line<\/p>\n\n<pre><code>image_d = torch.FloatTensor(np.asarray(A.reshape(1, 3, A.shape[0] , A.shape[1])))\n<\/code><\/pre>\n\n<p>You can't just reshape the image you need to transpose the channels. As a remark for the future, if you get a stripy result like you did it's most likely some permutation\/transposition or reshaping operation that's not correct.<\/p>\n\n<p>Other than that I also scaled the input image to <code>[0, 1]<\/code> to show it properly. Below is the working code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport torch\nimport torchvision\nimport torchvision.transforms as transforms\nimport torch.nn as nn\nimport matplotlib.pyplot as plt\nimport imageio\nimport sys\n\nA = imageio.imread('LiT.png')\n# Define how the convolution operation works\nconv2 = nn.Conv2d(in_channels=3, out_channels=3, kernel_size=3, stride=1, padding=1)\n\n# from [H, W, C] to [C, H, W]\ntransposed_image = A.transpose((2, 0, 1))\n# add batch dim\ntransposed_image = np.expand_dims(transposed_image, 0)\n\nimage_d = torch.FloatTensor(transposed_image)\nfc = conv2(image_d)\nfc1 = fc.permute(0, 2, 3, 1)[0]\nresult = fc1.data.numpy()\nmax_ = np.max(result)\nmin_ = np.min(result)\nresult -= min_\nresult \/= max_\n\nplt.figure(figsize=(16,8))\nplt.subplot(1,2,1)\nplt.imshow(A)\nplt.subplot(1,2,2)\nplt.imshow(result)\n\nplt.show()\n<\/code><\/pre>\n"}]}
+{"QuestionId":57739066,"AnswerCount":0,"Tags":"<python><opencv><tensorflow><python-3.6>","CreationDate":"2019-08-31T14:58:07.377","AcceptedAnswerId":null,"OwnerUserId":9316074.0,"Title":"How can I re-run an object detection program until there is a manual breakout?","Body":"<p>I have downloaded the TensorFlow Model Zoo and have transfer-learned the model network. The program works as expected, producing a picture with the objects and their detection classes and scores. <\/p>\n\n<p>Whenever I put this program into an infinte loop, the image is never loaded and causes a \"non-responding\" window. The program still runs in the background as I can see the CUDA engine running in the command prompt and the print statements are still being produced to the shell.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def detection():\n        #image detection here...\n\n\n        #Display the image and then clean up\n           cv2.imshow('Object Detection', image)\n        #Wait for clean up - 1 minute\n           time.sleep(60)\n        #Clean up    \n           cv2.destroyAllWindows\n           print(\"end of code\")\n\nwhile(True):\n        detection()\n<\/code><\/pre>\n\n<p>I expected the output to produce a window for 1 minute, clear the window after the time has passed and then re-run the object detection to produce the window again.<\/p>\n","answers":[]}
+{"QuestionId":57739225,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-08-31T15:16:24.090","AcceptedAnswerId":null,"OwnerUserId":10366671.0,"Title":"How to do fine-tuning with segnet","Body":"<p>I'm trying to use SegNet with my own class numbers. Because I want to change only class number, and I want to use weights and almost all network structures, I'm trying fine-tuning with keras.<\/p>\n\n<p>I did<br>\n<strong>1.<\/strong> use <code>load_model<\/code> and load SegNet's structure and weights<br>\n<strong>2.<\/strong> <code>pop(0)<\/code> the last conv, batch norm and activation layers<br>\n<strong>3.<\/strong> I create the Sequential instance, and put old model and new conv, batch norm and activation layers<br>\n<strong>4.<\/strong> compile the new model<br><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#1 load the model\nmodel = load_model(model_path) \n\n#2 pop the layers\nmodel.layers.pop() #conv\nmodel.layers.pop() #batch norm\nmodel.layers.pop() #activation\n\n#3 create new model\ntop_model = Sequential()\ntop_model.add(Conv2D(new_class_num, (1, 1), padding='same'))\ntop_model.add(Reshape((imagew * imageh, new_class_num)))\ntop_model.add(Activation('softmax'))\n\n#4 build the new model\ntop_model.compile(loss='categorical_crossentropy',optimizer=optimizers.RMSprop(lr=1e-5), metrics=['acc'])\n\nmodel=Model(inputs=model.input, outputs=top_model(model.output))\n\n<\/code><\/pre>\n\n<p>At the <code>top_model.add(Conv2D(new_class_num, (1, 1), padding='same'))<\/code> code, I cought <code>Input 0 is incompatible with layer conv2d_1: expected ndim=4, found ndim=3<\/code> error.\nHow can I fix this?<\/p>\n","answers":[]}
+{"QuestionId":57739331,"AnswerCount":1,"Tags":"<c><tensorflow><deeplab>","CreationDate":"2019-08-31T15:30:14.860","AcceptedAnswerId":null,"OwnerUserId":7095746.0,"Title":"Running Tensorflow session produces an empty tensor","Body":"<p>I am trying to use the Tensorflow C API to run a session with the Deeplab graph. The frozen graph of Deeplab, pre-trained on Cityscapes, was downloaded from here:\n<a href=\"http:\/\/download.tensorflow.org\/models\/deeplabv3_mnv2_cityscapes_train_2018_02_05.tar.gz\" rel=\"nofollow noreferrer\">http:\/\/download.tensorflow.org\/models\/deeplabv3_mnv2_cityscapes_train_2018_02_05.tar.gz<\/a><\/p>\n\n<p>When I run with python, I get this segmentation output:\n<a href=\"https:\/\/i.stack.imgur.com\/E8WKn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/E8WKn.png\" alt=\"Segmentation output using Python\"><\/a><\/p>\n\n<p>By printing out all of the graph's tensors via the python line: <code>tensors = [n.values() for n in tf.get_default_graph().get_operations()]\n<\/code>, I found out that the dimensions of the input tensor are {1,?,?,3}, and the output tensor are {1,?,?}, and the data types of the input and output tensors are uint8 and int64, respectively. I used this information to write a C++ method to run the graph session:<\/p>\n\n<pre><code>int Deeplab::run_segmentation(image_t* img, segmap_t* seg) {\n    using namespace std;\n\n    \/\/ Allocate the input tensor\n    TF_Tensor* const input = TF_NewTensor(TF_UINT8, img-&gt;dims, 4, img-&gt;data_ptr, img-&gt;bytes, &amp;free_tensor, NULL);\n    TF_Operation* oper_in = TF_GraphOperationByName(graph, \"ImageTensor\");\n    const TF_Output oper_in_ = {oper_in, 0};\n\n    \/\/ Allocate the output tensor\n    TF_Tensor* output = TF_NewTensor(TF_INT64, seg-&gt;dims, 3, seg-&gt;data_ptr, seg-&gt;bytes, &amp;free_tensor, NULL);\n    TF_Operation* oper_out = TF_GraphOperationByName(graph, \"SemanticPredictions\");\n    const TF_Output oper_out_ = {oper_out, 0};\n\n    \/\/ Run the session on the input tensor\n    TF_SessionRun(session, nullptr, &amp;oper_in_, &amp;input, 1, &amp;oper_out_, &amp;output, 1, nullptr, 0, nullptr, status);\n\n    return TF_GetCode(status); \/\/ https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/c\/tf_status.h#L42 \n}\n<\/code><\/pre>\n\n<p>Where the argument types <code>image_t<\/code> and <code>segmap_t<\/code> contain the parameters needed to call TF_NewTensor. They simply hold the pointers to the allocated buffer for the input\/output tensors, the dimensions of the tensors, and the size in bytes:<\/p>\n\n<pre><code>typedef struct segmap {\n    const int64_t* dims;\n    size_t bytes;\n    int64_t* data_ptr;\n} segmap_t;\n\ntypedef struct image {\n    const int64_t* dims;\n    size_t bytes;\n    uint8_t* data_ptr;\n} image_t;\n<\/code><\/pre>\n\n<p>Then, I used OpenCV to fill an array with the street scene image (same one as above), and passed the <code>image_t<\/code> and <code>segmap_t<\/code> structs into the session run method :<\/p>\n\n<pre><code>    \/\/ Allocate input image object\n    const int64_t dims_in[4] = {1, new_size.width, new_size.height, 3};\n    image_t* img_in = (image_t*)malloc(sizeof(image_t));\n    img_in-&gt;dims = &amp;dims_in[0];\n    \/\/img_in-&gt;data_ptr = (uint8_t*)malloc(new_size.width*new_size.height*3);\n    img_in-&gt;data_ptr = resized_image.data;\n    img_in-&gt;bytes = new_size.width*new_size.height*3*sizeof(uint8_t);\n\n    \/\/ Allocate output segmentation map object\n    const int64_t dims_out[3] = {1, new_size.width, new_size.height};\n    segmap_t* seg_out = (segmap_t*)malloc(sizeof(segmap_t));\n    seg_out-&gt;dims = &amp;dims_out[0];\n    seg_out-&gt;data_ptr = (int64_t*)calloc(new_size.width*new_size.height, sizeof(int64_t));\n    seg_out-&gt;bytes = new_size.width*new_size.height*sizeof(int64_t);\n<\/code><\/pre>\n\n<p>But the resulting tensor (<code>set_out-&gt;data_ptr<\/code>) consisted of all 0s. The graph seemed to execute for about 5 seconds, the same amount of time as the working python implementation. Somehow, the graph is failing to dump the output tensor data in the buffer I allocated. What am I doing wrong?<\/p>\n","answers":[{"AnswerId":"57740590","CreationDate":"2019-08-31T18:17:16.960","ParentId":null,"OwnerUserId":"7095746","Title":null,"Body":"<p>There were two mistakes:<\/p>\n\n<p>First, Deeplab's input tensor dimensions are {1, height, width, 3}, and the output tensor dimensions are {1, height, width}. So I had to swap height and width. <\/p>\n\n<p>Also, for some reason, you have to fetch the data from the tensor with the <code>TF_TensorData<\/code> method. Creating the output tensor by doing <code>TF_NewTensor(..., data_ptr, ...)<\/code>, then running <code>TF_SessionRun<\/code>, and finally accessing <code>data_ptr<\/code> does not work. You have to instead create the output tensor by calling <code>TF_AllocateTensor(...)<\/code>, run <code>TF_SessionRun<\/code>, and access the tensor data with <code>TF_TensorData(&amp;tensor)<\/code>. <\/p>\n"}]}
+{"QuestionId":57739512,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-08-31T15:52:25.060","AcceptedAnswerId":null,"OwnerUserId":10561436.0,"Title":"How to add after each iteration in tensorflow","Body":"<p>I am trying to achieve the following:<\/p>\n\n<p>compute the losses in the previous 25 predictions and sum them before\ncomputing the gradient. I have tried this:<\/p>\n\n<pre><code>loss_summation=tf.Variable(0,dtype=tf.dtypes.float32,name=\"loss\")\nxentropy=tf.nn.sparse_softmax_cross_entropy_with_logits(labels=next_element[1],logits=logits2,name=\"xentropy\")\nloss=tf.math.reduce_sum(tf.reduce_mean(xentropy,name=\"loss\"))\nloss_summation=tf.assign(loss_summation,loss_summation+loss)\noptimizer = tf.train.AdamOptimizer(learning_rate=self.learning_rate)\n gvs = optimizer.compute_gradients(loss_summation,[vars])\nwith tf.Session() as sess():\n      for i in range(25):\n                     b=sess.run([loss_summation])\n<\/code><\/pre>\n\n<p>However <code>optimizer.compute_gradients()<\/code> complains that\n<code>None values not supported<\/code>. How can go around this ?<\/p>\n\n<p>I am actually trying to implement the following function(feedforward of LSTM) in tensorflow to predict the next word given the previous ones<\/p>\n\n<pre><code>def feedforward(self,x_s,hpre,targets,p_s):\n\n\n    fts,its,gts,css,ots,output,inputs=[],[],[],[],[],[],[]\n        losses=[]\n        hprev=hpre\n        hts=[hprev]\n        loss=0\n        losses=[]\n        previous_state=p_s\n        css.append(previous_state)\n        for x,y in zip(x_s,targets):\n            k=np.zeros((self.vocab_size,1))\n            k[x]=1 \n            M_c=np.row_stack((hprev,k))                      \n            ft=self.sigmoid(np.dot(self.W1,M_c)+self.b1)\n            fts.append(ft)\n            it=self.sigmoid(np.dot(self.W2,M_c)+self.b2)\n            its.append(it)\n            gt=np.tanh(np.dot(self.W3,M_c)+self.b3)\n            gts.append(gt)\n            cs=(ft*previous_state)+(it*gt)\n            previous_state=cs\n            css.append(cs)\n            ot=self.sigmoid(np.dot(self.W4,M_c)+self.b4)\n            ots.append(ot)\n            ht=ot*np.tanh(cs)\n            hts.append(ht)\n            yt=self.softmax(np.dot(self.W5,ht)+self.b5)\n            hprev=ht\n            output.append(yt)\n            inputs.append(M_c)\n            loss+=-np.log(yt[y])\n            losses.append(loss)\n\n        return fts,its,gts,css,ots,output,hts,loss,hts[-1],css[-1],inputs\n<\/code><\/pre>\n\n<p><code>x_s<\/code> is a list of integers representing words. <\/p>\n\n<pre><code>x_s=[0,1,2,3,4,5,6,7,8....,24]\n<\/code><\/pre>\n\n<p>targets is the list of integers expected i.e if x_s=0 then next letter is 1<\/p>\n\n<pre><code>targets=[1,2,3,4,5,6,7,8,9...,25]\n<\/code><\/pre>\n\n<p>The loss which is a summation of 25 losses is what will be minimized.<\/p>\n","answers":[{"AnswerId":"57807008","CreationDate":"2019-09-05T13:49:14.727","ParentId":null,"OwnerUserId":"997378","Title":null,"Body":"<p>There are a few things you need to address here:<\/p>\n\n<ol>\n<li><p>Is there a good reason not to just use larger batches? Are you trying to implement the <a href=\"https:\/\/arxiv.org\/abs\/1907.08610\" rel=\"nofollow noreferrer\">lookahead optimizer<\/a> or something?<\/p><\/li>\n<li><p>You look like you're getting started with TensorFlow. Consider turning on eager execution with <code>tf.enable_eager_execution()<\/code>. TensorFlow 2.0 is coming soon, don't waste your time messing with <code>tf.Sessions<\/code>.<\/p><\/li>\n<li><p>Variables are not differentiable. So accumulating the losses in a variable doesn't make any sense. <\/p><\/li>\n<li><p>I would make a copy of all the model's variables, and accumulate new values there. Then, after <code>N<\/code> iterations assign those values back to the model. Something like:<\/p><\/li>\n<\/ol>\n\n<pre><code>model = tf.keras.Sequential(...)\nvars = model.trainable_variables\nweight_acc = [tf.Variable(var) for var in model.trainable_variables]\nfor n,(batch, label) in enumerate(dataset):\n  with tf.GradientTape() as tape:\n    pred = model(batch)\n    loss = cal_loss(batch, label)\n  grads = tape.gradients(loss, vars)\n  for g, a in zip(grad, weight_acc):\n    a.assign_add(learning_rate*g)\n  if n%25 == 0:\n    for a, v in zip(weight_acc, vars):\n      v.assign_add(lookahead_fraction*(a-v)) \n\n<\/code><\/pre>\n"}]}
+{"QuestionId":57739721,"AnswerCount":0,"Tags":"<python><tensorflow><lstm>","CreationDate":"2019-08-31T16:18:38.620","AcceptedAnswerId":null,"OwnerUserId":12003543.0,"Title":"How to reuse tf.nn.LSTMCell? I got an error when set resut=True in tf.variable_scope","Body":"<p>When I use TensorFLow 1.12.0 I got trouble in resuing tf.nn.rnn_cell.LSTMCell. First I declare a LSTMCell then I use it with tf.nn.dynamic_rnn it's ok. And when use the same LSTMCell twice in a while-loop structure and set the reuse=True I got an error<\/p>\n\n<pre><code>lstm_cell\/kernel does not exist, or was not created with tf.get_variable(). Did you mean to set reuse=tf.AUTO_REUSE in VarScope?\"\n<\/code><\/pre>\n\n<p>I'm really confused about it since I an using the same LSTMCell.\nSo could someone have some ideas? Thanks a lot!<\/p>\n\n<p>And when I use a tf.nn.dynamic_rnn at the second time it works. Similarly, I use while-loop structure at first time it works again.<\/p>\n\n<p>First time to use LSTMCell<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def official(LSTM_cell):\n    prev_state = LSTM_cell.zero_state(4, tf.float32) # 4 for batch size\n    outputs, state = tf.nn.dynamic_rnn(LSTM_cell, inputs, initial_state=prev_state)\n    return outputs\n<\/code><\/pre>\n\n<p>Second time a while-loop structure<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def diy(LSTM_cell):\n    inputs_array = tf.TensorArray(dtype=tf.float32, size=25)\n    inputs_array = inputs_array.unstack(tf.transpose(inputs, [1,0,2])) # inputs is a tensor with shape (batch size=4, sequence_length=25, dim)\n    outputs_array = tf.TensorArray(dtype=tf.float32, size=25)\n    prev_state = LSTM_cell.zero_state(4, tf.float32) # 4 for batch size\n    def cond(inputs, outputs, prev_state, i):\n        return i&lt;25\n    def body(inputs, outputs, prev_state, i):\n        x = inputs.read(i)\n        output, state = LSTM_cell(x, prev_state)\n        outputs = outputs.write(i, output)\n\n        return inputs, outputs, state, i+1\n    i = 0\n    lstm_outputs = tf.while_loop(cond, body, loop_vars=[inputs_array, outputs_array, prev_state, i])\n\n    final_outputs = tf.transpose(lstm_outputs[1].stack(), [1, 0, 2])\n    return final_outputs\n<\/code><\/pre>\n\n<p>And I implement a function to run<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def run(is_diy=False, reuse=False):\n    with tf.variable_scope(\"demo\", reuse=reuse):\n        LSTM_cell = tf.nn.rnn_cell.LSTMCell(128)\n        if is_diy:\n            return diy(LSTM_cell) # use while-loop\n        else:\n            return official(LSTM_cell) # use tf.nn.dynamic_rnn\n<\/code><\/pre>\n\n<p>Finally I run it twice and set the scond time with reuse=True<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>outputs1 = run(is_diy=False, reuse=False)\noutputs2 = run(is_diy=True, resue=True)\n<\/code><\/pre>\n\n<p>And I got the error. \nWhen implement like this, It works<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>outputs1 = run(is_diy=False, reuse=False)\noutputs2 = run(is_diy=False, resue=True) # use tf.nn.dynamic_rnn twice\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57739947,"AnswerCount":1,"Tags":"<python><tensorflow><keras><scikit-learn>","CreationDate":"2019-08-31T16:49:48.140","AcceptedAnswerId":null,"OwnerUserId":3443033.0,"Title":"y_true and y_pred have different number of output (10!=1)","Body":"<p>I am using Scikit learn wrapper KerasClassifier for hyperparameters tuning for my LSTM model using RandomizedSearchCV. Below is a summary of what I am doing:\n1. xtrain has the shape [355,5,10] and ytrain has the shape[355,10], There are 355 training samples and 10 features and labels.\n2. First I create the model using build_lstm_model function\n3. Define KerasClassifier\n4. Specify parameters that to be sued for fitting to determine the scoring\n5. Specify parameters to be searched using RandomizedSearchCV\n5. fit the model<\/p>\n\n<p>I am using 'neg_mean_squared_error' as the scoring metrics. When I run the code I get an error \"y_true and y_pred have different number of output (10!=1)\"<\/p>\n\n<p>I found that, if I do not specify any scoring metrics, then it works fine. But, I want to use neg_mean_squared_error, since its a regression problem. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># keras model\ndef build_lstm_model(n_blocks=6, n_cells=40, lr=0.001, lookback=lookback, n=n):\n    model = Sequential()\n\n    for i in range(n_blocks-1):\n        model.add(LSTM(n_cells, input_shape=(lookback, n), return_sequences=True, activation='tanh', kernel_initializer='uniform'))\n\n    model.add(LSTM(n_cells, input_shape=(lookback, n), activation='tanh', kernel_initializer='uniform'))\n    model.add(Dense(n))\n\n    adam = optimizers.Adam(lr=lr, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False)\n    model.compile(loss='mean_squared_error', optimizer=adam, metrics=['accuracy'])\n\n    return model\n\n# pass in fixed parameters n_input and n_class\nmodel_lstm = KerasClassifier(\n    build_fn = build_lstm_model,\n    lookback = lookback, n = n)\n\n# specify other extra parameters pass to the .fit\n# number of epochs is set to a large number\nkeras_fit_params = {   \n    'epochs': 10,\n    'batch_size': 16,\n    'validation_data': (xvalid, yvalid),\n    'verbose': 0\n}\n\n# random search parameters \n# specify the options and store them inside the dictionary\n# batch size and training method can also be hyperparameters, but it is fixed\nn_blocks_params = [3, 4, 5, 6, 7, 8]\nn_cells_params = [20, 30, 40, 50, 60]\nlr_params = [0.001, 0.0001]\n\nkeras_param_options = {\n    'n_blocks': n_blocks_params,\n    'n_cells': n_cells_params,  \n    'lr': lr_params\n}\n\n# `verbose` 2 will print the class info for every cross-validation, kind of too much\nrs_lstm = GridSearchCV( \n    model_lstm, \n    param_distributions = keras_param_options,\n    #fit_params = keras_fit_params,\n    scoring = 'neg_mean_squared_error',\n    n_iter = 3, \n    cv = 5,\n    n_jobs = -1\n    #verbose = 0\n)\n\nrs_lstm.fit(xtrain, ytrain)\n<\/code><\/pre>\n\n<p>Is there a way I can use mean_squared_error as the metrics in RandomizedSearchCV?<\/p>\n","answers":[{"AnswerId":"57741075","CreationDate":"2019-08-31T19:29:36.670","ParentId":null,"OwnerUserId":"3443033","Title":null,"Body":"<p>I was using KerasClassifier. I didn't know that there is another wrapper KerasRegressor in SKlearn. When I use KerasRegressor, I can use regression related metrics for finding a good model. Thank you. <\/p>\n"}]}
+{"QuestionId":57740150,"AnswerCount":0,"Tags":"<python><tensorflow><keras><neural-network>","CreationDate":"2019-08-31T17:18:43.083","AcceptedAnswerId":null,"OwnerUserId":7131919.0,"Title":"custom loss functions with multiple inputs and outputs","Body":"<p>I have a network with two inputs and two outputs as below:<\/p>\n\n<pre><code>image_inputs = Input(shape=(784,))\nA_inputs = Input(shape=(780,))\ninputs = keras.layers.concatenate([image_inputs, A_inputs])\nx = Dense(512, activation='relu')(inputs)\noutputs1 = Dense(784, activation='sigmoid')(x)\noutputs2 = Dense(780, activation='sigmoid')(x)\noutputs = keras.layers.concatenate([outputs1, outputs2])\n<\/code><\/pre>\n\n<p>My model is as:<\/p>\n\n<pre><code>model = Model([image_inputs, A_inputs], outputs, name='vae_mlp')\n<\/code><\/pre>\n\n<p>I want to define some loss functions:<\/p>\n\n<pre><code>def loss_X(inputs,outputs):\n    reconstruction_loss = original_dim*binary_crossentropy(inputs[::,0:-(n_d)], outputs[::,0:-(780)])                                              \n    return K.mean(reconstruction_loss)\n\ndef loss_A(inputs,outputs):\n    reconstruction_loss = original_dim*binary_crossentropy(inputs[::,-(n_d):], outputs[::,-(780):])                                              \n    return K.mean(reconstruction_loss)\n\ndef loss_KL(inputs,outputs):\n    A,A_T=triangle(n_z,A_inputs)\n    sum_A=tf.reduce_sum(A*A,0)    \n    kl_loss =  -0.5*K.sum(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var)*sum_A, axis=-1)    \n    return K.mean(kl_loss)\n\ndef loss_total(inputs,outputs):\n    reconstruction_loss = original_dim*binary_crossentropy(inputs,outputs)                                              \n    A,A_T=triangle(n_z,A_inputs)\n    sum_A=tf.reduce_sum(A*A,0)    \n    kl_loss =  -0.5*K.sum(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var)*sum_A, axis=-1)  \n    return K.mean(reconstruction_loss + kl_loss)\n<\/code><\/pre>\n\n<p>and at the end:<\/p>\n\n<pre><code>model.compile(optimizer='adam',loss=loss_total,metrics= [loss_X,loss_A,loss_KL])\nhistory = model.fit([x_train,A_vec],epochs=epochs, batch_size=batch_size, \n                    validation_data=([x_test,A_vec], None))\n<\/code><\/pre>\n\n<p>In fact, I want to test a loss function (loss_total) which has some parts and I need to check these individual parts too. But since the input and outputs of the model are multiple, when I want to run the model, I receive this error: <\/p>\n\n<pre><code>IndexError: list index out of range\n<\/code><\/pre>\n\n<p>The model works for single input and output case, but I do not know how I can run the model in this case?<\/p>\n","answers":[]}
+{"QuestionId":57740492,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-08-31T18:04:42.803","AcceptedAnswerId":null,"OwnerUserId":10314925.0,"Title":"Custom Training Keras Model","Body":"<p>I need to custom train a keras model : <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def base_model(): \n    inputs = Input(shape=(128,)) \n    x = Dense(128,activation='relu')(inputs)\n    outputs = Dense(13, activation='relu')(x)\n    return  Model(inputs=inputs, outputs=outputs)\n<\/code><\/pre>\n\n<p>I tried using <code>tf.GradientTape<\/code> for the same as :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code> 1. x = tf.random.uniform([1,128])\n 2. y = tf.random.uniform([13])\n 3. model = base_model()\n 4. \n 5. optimizer = keras.optimizers.Adam(learning_rate=0.001)\n 6. with tf.GradientTape() as t:\n 7.     logits = model(x)\n 8.     loss = keras.backend.mean(keras.losses.mean_squared_error(y, logits))\n 9. gradients = t.gradient(loss, model.trainable_weights)\n10. optimizer.apply_gradients(zip(gradients, model.trainable_weights))\n<\/code><\/pre>\n\n<p>The snippet raised the following error at line 9:<\/p>\n\n<pre><code>'RefVariable' object has no attribute '_id'\n<\/code><\/pre>\n\n<p>I'm using <code>Tensorflow 1.14<\/code> and have importedkeras as <\/p>\n\n<pre><code>import tensoflow.keras as keras\n<\/code><\/pre>\n\n<p>I'm new to tensorflow. I'm sorry if I said anything technically wrong. Also I wonder if I can train the same model in <strong>eager execution<\/strong> mode.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":57740570,"AnswerCount":0,"Tags":"<python><keras><jupyter-notebook>","CreationDate":"2019-08-31T18:14:05.227","AcceptedAnswerId":null,"OwnerUserId":12003851.0,"Title":"not able to import Keras in my Jupyter notebook","Body":"<p>I have been trying to import Keras into my Juypter notebook which is running inside a virtualenv titled 'ai'. All other imports (such as import numpy, tensorflow) work fine. But import keras constantly gives me the error \"module not found\". <\/p>\n\n<p>I have installed Jupyter, Keras, Tensorflow etc. inside my virtual environment and am invoking Jupyter from within it. When I list the virtual environments loaded in my notebook, I do list my virtual env listed. \nWhat could possibly be wrong?<\/p>\n","answers":[]}
+{"QuestionId":57741255,"AnswerCount":0,"Tags":"<keras><keras-layer>","CreationDate":"2019-08-31T20:00:27.653","AcceptedAnswerId":null,"OwnerUserId":12004063.0,"Title":"Lambda slice causes InvalidArgumentError: Incompatible shapes","Body":"<p>When I use Lambda layers to slice tensors, then concatenate them, my model ends up producing an error: \"InvalidArgumentError: Incompatible shapes: [4,2] vs [8]\"<\/p>\n\n<p>I've tried specifying input_shape and output_shape in various places. This does not seem to make a difference. I've also tried changing the axis parameter in the Concatenate layer. If I use -1 (the default), I get the same error. If I use 0, I get a different error indicating that the output shape is (2,) when (1,) is expected.  axis=1 gives an error indicating that the inputs are of insufficient rank.<\/p>\n\n<p>I have produced the short code below that reproduces the problem.<\/p>\n\n<p>The model takes input of the shape (None,3) and is supposed to produce the output of (None,2) as categories. The data is made up just to illustrate the problem.<\/p>\n\n<p>Even though this is a small model which ends up with no parameters, my larger ends up with the same error message.<\/p>\n\n<p>FYI, I'm using Google Colab.<\/p>\n\n<pre><code># repro of - InvalidArgumentError: Incompatible shapes\nimport pandas as pd\nimport tensorflow as tf\nfrom tensorflow import keras\nfrom keras import optimizers\nfrom keras.layers import Concatenate\nfrom keras.layers import Input\nfrom keras.layers import Lambda\nfrom keras.models import Model\n\nprint (\"tensorflow version %s\" % tf.__version__)\nprint (\"keras version %s\" % keras.__version__)\n\n# Mock data\nX = pd.DataFrame([[0,1,0], [0,2,1], [7,9,0], [4,5,6]])\nY = pd.DataFrame([[0,1], [0,1], [0,1], [1,0]])\nprint(X)\nprint(\"Multi-layer NN training. X size %s. Y size %s.\"\n      % (X.shape, Y.shape))\n\ninput_layer_X = Input(shape=(X.shape[1],))\na_slice = Lambda(lambda x: x[:, 0], output_shape=(1,), name='a_slice')(input_layer_X)\nb_slice = Lambda(lambda x: x[:, 1], output_shape=(1,), name='b_slice')(input_layer_X)\nconcat_layer = Concatenate(name='concat_example', axis=-1)([a_slice, b_slice])\n\nmodel = Model(input_layer_X, concat_layer)\nsgd = optimizers.SGD(lr=0.00001)\nmodel.compile(sgd, loss='categorical_crossentropy')\nprint(model.summary())\n\nmodel.fit(X, Y, epochs=3)\n<\/code><\/pre>\n\n<p>The model fit should run. Instead I get this error message from the model.fit call:<\/p>\n\n<pre><code>InvalidArgumentError: Incompatible shapes: [4,2] vs. [8]\n     [[{{node loss_8\/concat_example_loss\/mul}}]]\n<\/code><\/pre>\n\n<p>But as far as I can tell, the model summary looks like I've set it up right:<\/p>\n\n<pre><code>Model: \"model_9\"\n__________________________________________________________________________________________________\nLayer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput_9 (InputLayer)            (None, 3)            0                                            \n__________________________________________________________________________________________________\na_slice (Lambda)                (None, 1)            0           input_9[0][0]                    \n__________________________________________________________________________________________________\nb_slice (Lambda)                (None, 1)            0           input_9[0][0]                    \n__________________________________________________________________________________________________\nconcat_example (Concatenate)    (None, 2)            0           a_slice[0][0]                    \n                                                                 b_slice[0][0]                    \n==================================================================================================\n<\/code><\/pre>\n\n<p>TF and keras version info:<\/p>\n\n<pre><code>tensorflow version 1.14.0\nkeras version 2.2.4-tf\n<\/code><\/pre>\n\n<p>Any help is appreciated.<\/p>\n","answers":[]}
+{"QuestionId":57741595,"AnswerCount":0,"Tags":"<tensorflow><keras><neural-network>","CreationDate":"2019-08-31T21:06:30.190","AcceptedAnswerId":null,"OwnerUserId":5892689.0,"Title":"Radial Basis Network with a Keras custom Layer","Body":"<p>I'm trying to develop a custom Keras Layer to build a Radial Based Network.\nI started with this post: <a href=\"https:\/\/www.kaggle.com\/residentmario\/radial-basis-networks-and-custom-keras-layers\" rel=\"nofollow noreferrer\">RBN and Keras Layer<\/a><\/p>\n\n<p>I want to train the network using KMeans Clustering, so the Keras Layer is not required to be 'trainable'. So I pretrain the Centers of the RBN and then use them for the kernel operation. Instead of building a layer with add_weight method, I create a variable K.variable( centers ).<\/p>\n\n<p>In the code below, I create a random data X with 7 features. An RBN with 9 centers ( pretrained centers with KMeans ). The call method should kernelize the input data.<\/p>\n\n<p>PROBLEM\nI don't know how to operate the tensors 'inputs' with the centers ('self.mu').\nI'm getting an Error while fitting:<\/p>\n\n<pre><code>ValueError: Dimensions must be equal, but are 9 and 7 for 'dense_94\/MatMul' (op: 'MatMul') with input shapes: [1,9], [7,15].\n\ndef call(self, inputs):\n    diff = K.expand_dims(inputs,0) - K.expand_dims(self.mu,0)\n    l2 = K.sum(K.pow(diff, 2), axis=-1)\n    res = K.exp(-1 * self.gamma * l2)\n    return res\n<\/code><\/pre>\n\n<p>The code :<\/p>\n\n<pre><code>from keras.layers import Layer\nfrom keras import backend as K\nimport numpy as np\nfrom keras.layers import Dense\nfrom keras.models import Sequential\nfrom keras.losses import binary_crossentropy\nfrom keras.utils import to_categorical\nfrom sklearn.cluster import KMeans\n\n\nclass RBFLayer(Layer):\n    def __init__(self, centers, gamma, **kwargs):\n        super(RBFLayer, self).__init__(**kwargs)\n        self.units = centers.shape[0]\n        self.gamma = K.cast_to_floatx(gamma)\n        self.centers = centers\n    def build(self, input_shape):\n        self.mu = K.variable( self.centers, name='mu')\n#        self.mu = self.add_weight(name='mu',\n#                                  shape=(int(input_shape[1]), self.units),\n#                                  initializer='uniform',\n#                                  trainable=True)\n        super(RBFLayer, self).build(input_shape)\n    def call(self, inputs):\n        diff = K.expand_dims(inputs,0) - K.expand_dims(self.mu,0)\n        l2 = K.sum(K.pow(diff, 2), axis=-1)\n        res = K.exp(-1 * self.gamma * l2)\n        return res\n    def compute_output_shape(self, input_shape):\n        return (input_shape[0], self.units)\n\nX = np.random.randint(0,high=10, size=(50,7))\nC = KMeans(n_clusters=9).fit(X).cluster_centers_\n\ny = np.random.randint(0,high=4,size=(50,1))\ny = to_categorical(y,num_classes=4)\n\nmodel = Sequential()\nmodel.add(RBFLayer(C,0.5))\nmodel.add(Dense(15, activation='relu'))\nmodel.add(Dense(4, activation='softmax'))\nmodel.compile(optimizer='adam', loss='categorical_crossentropy')\nmodel.fit(X, y, epochs=3)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57741714,"AnswerCount":1,"Tags":"<python><machine-learning><keras><neural-network><recurrent-neural-network>","CreationDate":"2019-08-31T21:30:38.160","AcceptedAnswerId":57741836.0,"OwnerUserId":10934024.0,"Title":"How to define input_dim for Keras recurrent layers properly","Body":"<p>I am trying to train some NN to predict timeseries. I am using <code>Sequentional<\/code> model to define my NN structure. It looks like:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense, Activation, SimpleRNN, Embedding\nfrom keras import optimizers\nfrom keras import losses\nmodel = Sequential()\n#model.add(Dense(units=5, input_dim=3, activation = 'tanh'))\nmodel.add(SimpleRNN(units=5, input_dim = 3, activation = 'tanh'))\nmodel.add(Dense(units=16, activation='tanh'))\nmodel.add(Dense(1, activation='linear'))\nprop = optimizers.rmsprop(lr=0.01)\nsgd = optimizers.sgd(lr=0.01, nesterov=True, momentum=0.005)\nmodel.compile(optimizer=prop, loss='mean_squared_error')\n<\/code><\/pre>\n\n<p>It does not execute and returned error is:<\/p>\n\n<pre><code>ValueError: Error when checking input: expected simple_rnn_9_input to have 3 dimensions, but got array with shape (221079, 3)\n<\/code><\/pre>\n\n<p>When I use commented out <code>Dense<\/code> layer everything is just fine. I read Keras documentation and I see they are using <code>Embedding<\/code> layer. Although, I do not really understand why <code>Embedding<\/code> layer is necessary to use recurrent layers like <code>SimpleRNN<\/code> or <code>LSTM<\/code>.<\/p>\n\n<p><code>train_set<\/code> is 2D array with 4 columns - 4-th one is target column, rest are inputs.<\/p>\n\n<p>Is there any simple way how to use Keras' recurrent layers together with traditional <code>Dense<\/code> layers? I would appreciate explanation and some code examples.<\/p>\n\n<p>Best regards,\nMaks<\/p>\n","answers":[{"AnswerId":"57741836","CreationDate":"2019-08-31T21:57:26.847","ParentId":null,"OwnerUserId":"4357329","Title":null,"Body":"<p>I am no expert on this, but this may help<\/p>\n\n<pre><code>import numpy as  np\nimport numpy as  np\n\ndata = np.zeros((10,4))\nX = data[:,0:3].reshape(-1,1,3)\ny = data[:,3].reshape(-1,1)\nprint(X.shape)\nprint(y.shape)\n<\/code><\/pre>\n\n<p>prints:<\/p>\n\n<pre><code>(10, 1, 3)\n(10, 1)\n<\/code><\/pre>\n\n<p>then:<\/p>\n\n<pre><code>model.fit(X, y)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57741738,"AnswerCount":0,"Tags":"<python><tensorflow><nlp>","CreationDate":"2019-08-31T21:36:08.077","AcceptedAnswerId":null,"OwnerUserId":5865957.0,"Title":"tensorflow BERT - Issue with switching to multiclass from multilabel in","Body":"<p>I have been trying to learn BERT multiclass labeling through the example given here: <a href=\"https:\/\/towardsdatascience.com\/building-a-multi-label-text-classifier-using-bert-and-tensorflow-f188e0ecdc5d\" rel=\"nofollow noreferrer\">https:\/\/towardsdatascience.com\/building-a-multi-label-text-classifier-using-bert-and-tensorflow-f188e0ecdc5d<\/a><\/p>\n\n<p>I was able to get all the code to work with the sigmoid function(multilabel). However, I wanted to switch to a softmax approach(multiclass). There was already some existing code for multiclass that I uncommented and commented out the multilabel. However, I am getting an error(pasted below) with the matrix multiplication being done to calculate the loss in the create_model() function. Please let me know what I am doing wrong and how I should be approach multiclass.<\/p>\n\n<p>I have tried changing the line to this: per_example_loss = tf.nn.softmax_cross_entropy_with_logits(labels=labels, logits=logits)<\/p>\n\n<p>This works but I don't know if this changes the calculations being done. I want to make sure the model works in the correct way. <\/p>\n\n<p>def create_model(bert_config, is_training, input_ids, input_mask, segment_ids,\n                 labels, num_labels, use_one_hot_embeddings):\n    \"\"\"Creates a classification model.\"\"\"\n    model = modeling.BertModel(\n        config=bert_config,\n        is_training=is_training,\n        input_ids=input_ids,\n        input_mask=input_mask,\n        token_type_ids=segment_ids,\n        use_one_hot_embeddings=use_one_hot_embeddings)<\/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()\n# instead.\noutput_layer = model.get_pooled_output()\n\nhidden_size = output_layer.shape[-1].value\n\noutput_weights = tf.get_variable(\n    \"output_weights\", [num_labels, hidden_size],\n    initializer=tf.truncated_normal_initializer(stddev=0.02))\n\noutput_bias = tf.get_variable(\n    \"output_bias\", [num_labels], initializer=tf.zeros_initializer())\n\nwith tf.variable_scope(\"loss\"):\n    if is_training:\n        # I.e., 0.1 dropout\n        output_layer = tf.nn.dropout(output_layer, keep_prob=0.9)\n\n    logits = tf.matmul(output_layer, output_weights, transpose_b=True)\n    logits = tf.nn.bias_add(logits, output_bias)\n\n    #MULTILABEL\n    # probabilities = tf.nn.softmax(logits, axis=-1) ### multiclass case\n    # probabilities = tf.nn.sigmoid(logits)  #### multi-label case\n    #\n    # labels = tf.cast(labels, tf.float32)\n    # tf.logging.info(\"num_labels:{};logits:{};labels:{}\".format(num_labels, logits, labels))\n    # per_example_loss = tf.nn.sigmoid_cross_entropy_with_logits(labels=labels, logits=logits)\n    # loss = tf.reduce_mean(per_example_loss)\n\n    #MULTICLASS STUFF\n    probabilities = tf.nn.softmax(logits, axis=-1)\n    log_probs = tf.nn.log_softmax(logits, axis=-1)\n\n    one_hot_labels = tf.one_hot(labels, depth=num_labels, dtype=tf.float32)\n\n    per_example_loss = -tf.reduce_sum(one_hot_labels * log_probs, axis=-1)\n    loss = tf.reduce_mean(per_example_loss)\n\n    return (loss, per_example_loss, logits, probabilities)\n<\/code><\/pre>\n\n<p>ERROR MESSAGE<\/p>\n\n<p>File \"path\/Try.py\", line 571, in model_fn\n    num_labels, use_one_hot_embeddings)<\/p>\n\n<p>File \"path\/Try.py\", line 539, in create_model\n    per_example_loss = -tf.reduce_sum(one_hot_labels * log_probs, axis=-1)<\/p>\n\n<p>File \"path\/venv\/lib\/python3.7\/site-packages\/tensorflow\/python\/ops\/math_ops.py\", line 884, in binary_op_wrapper\n    return func(x, y, name=name)<\/p>\n\n<p>File \"path\/venv\/lib\/python3.7\/site-packages\/tensorflow\/python\/ops\/math_ops.py\", line 1180, in _mul_dispatch\n    return gen_math_ops.mul(x, y, name=name)<\/p>\n\n<p>File \"path\/venv\/lib\/python3.7\/site-packages\/tensorflow\/python\/ops\/gen_math_ops.py\", line 6490, in mul\n    \"Mul\", x=x, y=y, name=name)<\/p>\n\n<p>File \"path\/venv\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 788, in _apply_op_helper\n    op_def=op_def)<\/p>\n\n<p>File \"path\/venv\/lib\/python3.7\/site-packages\/tensorflow\/python\/util\/deprecation.py\", line 507, in new_func\n    return func(*args, **kwargs)<\/p>\n\n<p>File \"path\/venv\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 3616, in create_op\n    op_def=op_def)<\/p>\n\n<p>File \"path\/venv\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 2027, in <strong>init<\/strong>\n    control_input_ops)<\/p>\n\n<p>File \"path\/venv\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1867, in _create_c_op\n    raise ValueError(str(e))<\/p>\n\n<p>ValueError: Dimensions must be equal, but are 6 and 32 for 'loss\/mul' (op: 'Mul') with input shapes: [32,6,6], [32,6].<\/p>\n","answers":[]}
+{"QuestionId":57742038,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><neural-network>","CreationDate":"2019-08-31T22:39:40.260","AcceptedAnswerId":null,"OwnerUserId":11188381.0,"Title":"dummy coding with ColumTransformer in python","Body":"<p>I am trying to use the ColumnTransformer class of scikit as a real DataFrame transformer.<\/p>\n\n<pre><code>import numpy as np\nimport matplotlib.pyplot as plt\nimport pandas as pd\nfrom sklearn.preprocessing import LabelEncoder, OneHotEncoder\nfrom sklearn.compose import ColumnTransformer\n\ndataset = pd.read_csv('dataset.csv')\nX = dataset.iloc[:, 3:13].values\ny = dataset.iloc[:, 13].values\n\ntransformer = ColumnTransformer(transformers=[(\"OneHot\", OneHotEncoder(), [1]), (\"LabEnc\", LabelEncoder(), [2])], remainder=\"passthrough\")\n\nX = transformer.fit_transform(X)\n<\/code><\/pre>\n\n<p>but there is an error on the last line:<\/p>\n\n<pre><code>TypeError: fit_transform() takes 2 positional arguments but 3 were given\n<\/code><\/pre>\n\n<p>In addition, X is:<\/p>\n\n<pre><code>array([[619, 'France', 'Female', ..., 1, 1, 101348.88],\n       [608, 'Spain', 'Female', ..., 0, 1, 112542.58],\n       [502, 'France', 'Female', ..., 1, 0, 113931.57],\n       ...,\n       [709, 'France', 'Female', ..., 0, 1, 42085.58],\n       [772, 'Germany', 'Male', ..., 1, 0, 92888.52],\n       [792, 'France', 'Female', ..., 1, 0, 38190.78]], dtype=object)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57742393,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2019-09-01T00:15:38.700","AcceptedAnswerId":57755233.0,"OwnerUserId":8128271.0,"Title":"Local reduce with specified slices over a single axis in tensorflow","Body":"<p>I am trying to perform a local reduce with specified slices over a single axis on a 2D array. <\/p>\n\n<p>I achieved this using numpy's <code>numpy.ufunc.reduceat<\/code> or <code>numpy.add.reduceat<\/code> but I would like do the same in tensorflow as the input to this reduce operation is an output from tensorflow convolution.<\/p>\n\n<p>I came across <code>tf.math.reduce_sum<\/code> but I am not sure how this can be used in my case.<\/p>\n\n<p>It will be great if I can do the <code>reduceat<\/code> operation in tensorflow as I can take advantage of a GPU.<\/p>\n","answers":[{"AnswerId":"57755233","CreationDate":"2019-09-02T10:19:54.423","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You can do almost the same using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/math\/segment_sum\" rel=\"nofollow noreferrer\"><code>tf.math.segment_sum<\/code><\/a>:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ndef add_reduceat_tf(a, indices, axis=0):\n    a = tf.convert_to_tensor(a)\n    indices = tf.convert_to_tensor(indices)\n    # Transpose if necessary\n    transpose = not (isinstance(axis, int) and axis == 0)\n    if transpose:\n        axis = tf.convert_to_tensor(axis)\n        ndims = tf.cast(tf.rank(a), axis.dtype)\n        a = tf.transpose(a, tf.concat([[axis], tf.range(axis),\n                                       tf.range(axis + 1, ndims)], axis=0))\n    # Make segment ids\n    r = tf.range(tf.shape(a, out_type=indices.dtype)[0])\n    segments = tf.searchsorted(indices, r, side='right')\n    # Compute segmented sum and discard first unused segment\n    out = tf.math.segment_sum(a, segments)[1:]\n    # Transpose back if necessary\n    if transpose:\n        out = tf.transpose(out, tf.concat([tf.range(1, axis + 1), [0],\n                                           tf.range(axis + 1, ndims)], axis=0))\n    return out\n\n# Test\nnp.random.seed(0)\na = np.random.rand(5, 10).astype(np.float32)\nindices = [2, 4, 7]\naxis = 1\n# NumPy computation\nout_np = np.add.reduceat(a, indices, axis=axis)\n# TF computation\nwith tf.Graph().as_default(), tf.Session() as sess:\n    out = add_reduceat_tf(a, indices, axis=axis)\n    out_tf = sess.run(out)\n# Check result\nprint(np.allclose(out_np, out_tf))\n# True\n<\/code><\/pre>\n\n<p>You can replace <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/math\/segment_sum\" rel=\"nofollow noreferrer\"><code>tf.math.segment_sum<\/code><\/a> above with the reduction function you want to use. The only difference between this and the actual <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.ufunc.reduceat.html\" rel=\"nofollow noreferrer\"><code>np.ufunc.reduceat<\/code><\/a> is the special case where <code>indices[i] &gt;= indices[i + 1]<\/code>. The posted function requires <code>indices<\/code> to be sorted, and if there were a case where <code>indices[i] == indices[i + 1]<\/code> the corresponding <code>i<\/code> position in the output would be zero, not <code>a[indices[i]]<\/code>.<\/p>\n"}]}
+{"QuestionId":57742798,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-09-01T02:21:33.933","AcceptedAnswerId":null,"OwnerUserId":2460690.0,"Title":"Shrink or grow a tensor away from the origin in pytorch?","Body":"<p>I have a batch x ndim tensor.<\/p>\n\n<p>I want to add noise to each observation that pushes it either toward or away from the origin without flipping any of the signs.<\/p>\n\n<p>Can anyone recommend a good approach?<\/p>\n","answers":[{"AnswerId":"57743450","CreationDate":"2019-09-01T05:31:01.920","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It's not very clear what you are trying to achieve here, but if you want to actually shift values in a tensor, you might want to look at <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/torch.html#torch.roll\" rel=\"nofollow noreferrer\"><code>torch.roll<\/code><\/a><\/p>\n"}]}
+{"QuestionId":57742827,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><conv-neural-network><caffe><pycaffe>","CreationDate":"2019-09-01T02:32:41.787","AcceptedAnswerId":null,"OwnerUserId":2467772.0,"Title":"ROI Pooling layer for character recognition","Body":"<p>Trying to understand <a href=\"http:\/\/openaccess.thecvf.com\/content_ECCV_2018\/papers\/Zhenbo_Xu_Towards_End-to-End_License_ECCV_2018_paper.pdf\" rel=\"nofollow noreferrer\">this paper<\/a> for LP recognition.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/YDrtQ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/YDrtQ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Three ROI Pooling layers are used to have features for 7 classifiers. <\/p>\n\n<p>First one is for Province, second one is for alphabets and the rest are for five alphanumeric characters.\nThe plate sample is Chinese LP.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ret8r.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ret8r.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I am confused how these seven classifiers know which respective part of the plate to look at for recognition.<\/p>\n\n<p>Because all seven classifiers are fed with same concatenated three ROIPooling features as follows.<\/p>\n\n<pre><code>roi1 = roi_pooling_ims(_x1, boxNew.mm(p1), size=(16, 8))\nroi2 = roi_pooling_ims(_x3, boxNew.mm(p2), size=(16, 8))\nroi3 = roi_pooling_ims(_x5, boxNew.mm(p3), size=(16, 8))\nrois = torch.cat((roi1, roi2, roi3), 1)\n_rois = rois.view(rois.size(0), -1)\n\ny0 = self.classifier1(_rois)\ny1 = self.classifier2(_rois)\ny2 = self.classifier3(_rois)\ny3 = self.classifier4(_rois)\ny4 = self.classifier5(_rois)\ny5 = self.classifier6(_rois)\ny6 = self.classifier7(_rois)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57742865,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-09-01T02:44:57.650","AcceptedAnswerId":null,"OwnerUserId":11846982.0,"Title":"Access 3d tensor by 2d index","Body":"<p>I am trying to access a 3d tensor matrix by 2d matrix. The return should be an 2d matrix too. And the following is what i try to achieve.<\/p>\n\n<pre><code>dim1 = 2\ndim2 = 3\ndim3 = 4\nsource = torch.FloatTensor(dim1, dim2, dim3)\nsource.normal_()\n\ncheck = zn &gt; 0.0\nindex = torch.argmax(check, dim=0)\n\nfor i in range(dim2):\n    for j in range(dim3):\n        ret[i, j] = source[index[i, j], i, j]\n<\/code><\/pre>\n","answers":[{"AnswerId":"57743461","CreationDate":"2019-09-01T05:33:18.480","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>I think you are looking for <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/torch.html#torch.gather\" rel=\"nofollow noreferrer\"><code>torch.gather<\/code><\/a>, you might need to <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/tensors.html#torch.Tensor.permute\" rel=\"nofollow noreferrer\"><code>permute<\/code><\/a> your tensor to fit the requirements of `gather.<\/p>\n"}]}
+{"QuestionId":57743724,"AnswerCount":0,"Tags":"<tensorflow><python-3.6>","CreationDate":"2019-09-01T06:31:53.453","AcceptedAnswerId":null,"OwnerUserId":12005078.0,"Title":"I want to add the spp algorithm to my program so that the input and output images don't have to be fixed","Body":"<p><a href=\"https:\/\/github.com\/nickliqian\/cnn_captcha\" rel=\"nofollow noreferrer\">https:\/\/github.com\/nickliqian\/cnn_captcha<\/a>\nHere is the git address, which is the code I am using.<\/p>\n\n<p>Hello, I have a verification code recognition program. Now I can model the model and verify the recognition result, but I can only fix the image size of the input and output. Now I want to add the spp(Spatial Pyramid Pooling) algorithm into my program. I have tried many times and can't solve it normally. Can you help me? Thank you.<\/p>\n\n<p><a href=\"https:\/\/github.com\/nickliqian\/cnn_captcha\" rel=\"nofollow noreferrer\">https:\/\/github.com\/nickliqian\/cnn_captcha<\/a><\/p>\n\n<p>Let the input and output pictures not be in fixed size<\/p>\n","answers":[]}
+{"QuestionId":57743774,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-09-01T06:43:03.703","AcceptedAnswerId":57743827.0,"OwnerUserId":3907250.0,"Title":"Check the number of parameters from state_dict in pytorch","Body":"<p>SO has an answer about how to check the total # of params from the model:\n<code>pytorch_total_params = sum(p.numel() for p in model.parameters())<\/code><\/p>\n\n<p>However, how does one check the total # of params from\nthe <code>state_dict<\/code>?<\/p>\n\n<p><code>state_dict = torch.load(model_path, map_location='cpu')<\/code>?<\/p>\n","answers":[{"AnswerId":"57743827","CreationDate":"2019-09-01T06:55:00.063","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can count the number of saved entries in the state_dict:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>sum(p.numel() for p in state_dict.values())\n<\/code><\/pre>\n\n<p>However, there's a snag here: a state_dict stores both <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/nn.html#parameters\" rel=\"nofollow noreferrer\">parameters<\/a> <em>and<\/em> <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/nn.html#torch.nn.Module.register_buffer\" rel=\"nofollow noreferrer\">persistent buffers<\/a> (e.g., BatchNorm's running mean and var). There's no way (AFAIK) to tell them apart from the state_dict itself, you'll need to load them into the model and use <code>sum(p.numel() for p in model.parameters()<\/code> to count only the parameters.<\/p>\n\n<p>For instance, if you checkout <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/torchvision\/models.html#torchvision.models.resnet50\" rel=\"nofollow noreferrer\"><code>resnet50<\/code><\/a><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from torchvision.models import resnet50\nmodel = resnet50(pretrained=True)\nstate_dict = torch.load('~\/.torch\/models\/resnet50-19c8e357.pth')\nnum_parameters = sum(p.numel() for p in model.parameters())\nnum_state_dict = sum(p.numel() for p in state_dict.values())\nprint('num parameters = {}, stored in state_dict = {}, diff = {}'.format(num_parameters, num_state_dict, num_state_dict - num_parameters))\n<\/code><\/pre>\n\n<p>Resulting with<\/p>\n\n<blockquote>\n<pre><code>num parameters = 25557032, stored in state_dict = 25610152, diff = 53120\n<\/code><\/pre>\n<\/blockquote>\n\n<p>As you can see there can be quite a gap between the two values.<\/p>\n"}]}
+{"QuestionId":57743970,"AnswerCount":0,"Tags":"<tensorflow><object-detection>","CreationDate":"2019-09-01T07:20:53.610","AcceptedAnswerId":null,"OwnerUserId":9524309.0,"Title":"Getting ResourceExhaustedError when using a pretrained SSD Mobilenet V2 model","Body":"<p>I am new to Tensor Flow. I am trying to use MobileNet SSD model for training detect two classes. However,  I am running into ResourceExhausted Error after the training runs for about 300 steps. I am unable to figure out what needs to be done to resolve this issue. \nHere is the train.py file:<\/p>\n\n<pre><code># Copyright 2017 The TensorFlow Authors. All Rights Reserved.\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\nr\"\"\"Training executable for detection models.\n\nThis executable is used to train DetectionModels. There are two ways of\nconfiguring the training job:\n\n1) A single pipeline_pb2.TrainEvalPipelineConfig configuration file\ncan be specified by --pipeline_config_path.\n\nExample usage:\n    .\/train \\\n        --logtostderr \\\n        --train_dir=path\/to\/train_dir \\\n        --pipeline_config_path=pipeline_config.pbtxt\n\n2) Three configuration files can be provided: a model_pb2.DetectionModel\nconfiguration file to define what type of DetectionModel is being trained, an\ninput_reader_pb2.InputReader file to specify what training data will be used and\na train_pb2.TrainConfig file to configure training parameters.\n\nExample usage:\n    .\/train \\\n        --logtostderr \\\n        --train_dir=path\/to\/train_dir \\\n        --model_config_path=model_config.pbtxt \\\n        --train_config_path=train_config.pbtxt \\\n        --input_config_path=train_input_config.pbtxt\n\"\"\"\n\nimport functools\nimport json\nimport os\nimport tensorflow as tf\n\nfrom object_detection.builders import dataset_builder\nfrom object_detection.builders import graph_rewriter_builder\nfrom object_detection.builders import model_builder\nfrom object_detection.legacy import trainer\nfrom object_detection.utils import config_util\n\ntf.logging.set_verbosity(tf.logging.INFO)\n\nflags = tf.app.flags\nflags.DEFINE_string('master', '', 'Name of the TensorFlow master to use.')\nflags.DEFINE_integer('task', 0, 'task id')\nflags.DEFINE_integer('num_clones', 1, 'Number of clones to deploy per worker.')\nflags.DEFINE_boolean('clone_on_cpu', False,\n                     'Force clones to be deployed on CPU.  Note that even if '\n                     'set to False (allowing ops to run on gpu), some ops may '\n                     'still be run on the CPU if they have no GPU kernel.')\nflags.DEFINE_integer('worker_replicas', 1, 'Number of worker+trainer '\n                     'replicas.')\nflags.DEFINE_integer('ps_tasks', 0,\n                     'Number of parameter server tasks. If None, does not use '\n                     'a parameter server.')\nflags.DEFINE_string('train_dir', '',\n                    'Directory to save the checkpoints and training summaries.')\n\nflags.DEFINE_string('pipeline_config_path', '',\n                    'Path to a pipeline_pb2.TrainEvalPipelineConfig config '\n                    'file. If provided, other configs are ignored')\n\nflags.DEFINE_string('train_config_path', '',\n                    'Path to a train_pb2.TrainConfig config file.')\nflags.DEFINE_string('input_config_path', '',\n                    'Path to an input_reader_pb2.InputReader config file.')\nflags.DEFINE_string('model_config_path', '',\n                    'Path to a model_pb2.DetectionModel config file.')\n\nFLAGS = flags.FLAGS\n\n\n@tf.contrib.framework.deprecated(None, 'Use object_detection\/model_main.py.')\ndef main(_):\n  assert FLAGS.train_dir, '`train_dir` is missing.'\n  if FLAGS.task == 0: tf.gfile.MakeDirs(FLAGS.train_dir)\n  if FLAGS.pipeline_config_path:\n    configs = config_util.get_configs_from_pipeline_file(\n        FLAGS.pipeline_config_path)\n    if FLAGS.task == 0:\n      tf.gfile.Copy(FLAGS.pipeline_config_path,\n                    os.path.join(FLAGS.train_dir, 'pipeline.config'),\n                    overwrite=True)\n  else:\n    configs = config_util.get_configs_from_multiple_files(\n        model_config_path=FLAGS.model_config_path,\n        train_config_path=FLAGS.train_config_path,\n        train_input_config_path=FLAGS.input_config_path)\n    if FLAGS.task == 0:\n      for name, config in [('model.config', FLAGS.model_config_path),\n                           ('train.config', FLAGS.train_config_path),\n                           ('input.config', FLAGS.input_config_path)]:\n        tf.gfile.Copy(config, os.path.join(FLAGS.train_dir, name),\n                      overwrite=True)\n\n  model_config = configs['model']\n  train_config = configs['train_config']\n  input_config = configs['train_input_config']\n\n  model_fn = functools.partial(\n      model_builder.build,\n      model_config=model_config,\n      is_training=True)\n\n  def get_next(config):\n    return dataset_builder.make_initializable_iterator(\n        dataset_builder.build(config)).get_next()\n\n  create_input_dict_fn = functools.partial(get_next, input_config)\n\n  env = json.loads(os.environ.get('TF_CONFIG', '{}'))\n  cluster_data = env.get('cluster', None)\n  cluster = tf.train.ClusterSpec(cluster_data) if cluster_data else None\n  task_data = env.get('task', None) or {'type': 'master', 'index': 0}\n  task_info = type('TaskSpec', (object,), task_data)\n\n  # Parameters for a single worker.\n  ps_tasks = 0\n  worker_replicas = 1\n  worker_job_name = 'lonely_worker'\n  task = 0\n  is_chief = True\n  master = ''\n\n  if cluster_data and 'worker' in cluster_data:\n    # Number of total worker replicas include \"worker\"s and the \"master\".\n    worker_replicas = len(cluster_data['worker']) + 1\n  if cluster_data and 'ps' in cluster_data:\n    ps_tasks = len(cluster_data['ps'])\n\n  if worker_replicas &gt; 1 and ps_tasks &lt; 1:\n    raise ValueError('At least 1 ps task is needed for distributed training.')\n\n  if worker_replicas &gt;= 1 and ps_tasks &gt; 0:\n    # Set up distributed training.\n    server = tf.train.Server(tf.train.ClusterSpec(cluster), protocol='grpc',\n                             job_name=task_info.type,\n                             task_index=task_info.index)\n    if task_info.type == 'ps':\n      server.join()\n      return\n\n    worker_job_name = '%s\/task:%d' % (task_info.type, task_info.index)\n    task = task_info.index\n    is_chief = (task_info.type == 'master')\n    master = server.target\n\n  graph_rewriter_fn = None\n  if 'graph_rewriter_config' in configs:\n    graph_rewriter_fn = graph_rewriter_builder.build(\n        configs['graph_rewriter_config'], is_training=True)\n\n  trainer.train(\n      create_input_dict_fn,\n      model_fn,\n      train_config,\n      master,\n      task,\n      FLAGS.num_clones,\n      worker_replicas,\n      FLAGS.clone_on_cpu,\n      ps_tasks,\n      worker_job_name,\n      is_chief,\n      FLAGS.train_dir,\n      graph_hook_fn=graph_rewriter_fn)\n\n\nif __name__ == '__main__':\n  tf.app.run()\n<\/code><\/pre>\n\n<p>Below shows what I am currently using:<\/p>\n\n<p>-Online download ->ssd_mobilenet_v2_coco from <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<ul>\n<li>Tensorflow-gpu 1.13.1<\/li>\n<li>OS: Windows 10<\/li>\n<li>Python: 3.6<\/li>\n<li>RAM: 8GB<\/li>\n<li>ssd_mobilenet_v2_coco config details:<\/li>\n<li>BatchSize: 24<\/li>\n<li>num_steps: 200000<\/li>\n<li>fine_tune_checkpoint: \"C:\/tftrain2\/models\/research\/object_detection\/ssd_mobilenet_v2_coco_2018_03_29\/model.ckpt\"<\/li>\n<li>num_classes: 2<\/li>\n<li>Input Data Set:<\/li>\n<li>37 images in train directory<\/li>\n<li>11 images in test directory<\/li>\n<li>GPU: NVIDIA GeForce GTX 1650<\/li>\n<\/ul>\n\n<p>I have followed this <a href=\"https:\/\/www.youtube.com\/watch?v=Rgpfk6eYxJA\" rel=\"nofollow noreferrer\">Youtube<\/a> tutorial to train my own object detection model. Please advice in solving this issue. Thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":57744030,"AnswerCount":2,"Tags":"<python><machine-learning><keras>","CreationDate":"2019-09-01T07:32:40.427","AcceptedAnswerId":57744297.0,"OwnerUserId":7827820.0,"Title":"When I'm doing a regression problem and the metrics is 'accuracy', then how is accuracy calculated when it's a regression problem","Body":"<p>I want to somebody to tell me how is the accuracy calculated when the problem is a regression problem.<\/p>\n\n<p>Now I'm working on a regression problem to predict how much electricity each user USES each day,I use keras build a LSTM model to do this time series prediction. At the beginning, I use the 'accuracy' as the metrics, and when run <\/p>\n\n<pre><code>model.fit(...,verbose=2,...)\n<\/code><\/pre>\n\n<p><code>val_acc<\/code> has a value after every epoch. And in my result, the value doesn't change, it's always the same value.<\/p>\n\n<p>Then I realized that the regression problem was that there was no concept of accuracy, and then I started to wonder, how is that accuracy calculated?<\/p>\n\n<p>I have a guess that when metrics is 'accuracy' in the regression question, accuracy is also calculated in a similar way to the classification problem: the number of predicted values equal to true values divided by the total sample size.<\/p>\n\n<p>Am I right?<\/p>\n\n<p>Thank you for your answer~<\/p>\n","answers":[{"AnswerId":"57744310","CreationDate":"2019-09-01T08:15:01.400","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Yes, the accuracy is computed in exactly the same way as in classification, keras does not do any kind of adjustment. As you say it makes no sense to use the accuracy (which is a classification metric) for a regression problem.<\/p>\n"},{"AnswerId":"57744297","CreationDate":"2019-09-01T08:14:02.517","ParentId":null,"OwnerUserId":"8741356","Title":null,"Body":"<p>In regression you use metrics that measure how far your prediction is from the actual value, such as <em>Squared Error, Mean Squared Error<\/em>, etc.\nPlease see <a href=\"https:\/\/stats.stackexchange.com\/questions\/142873\/how-to-determine-the-accuracy-of-regression-which-measure-should-be-used\">How to determine the accuracy of regression? Which measure should be used?<\/a><\/p>\n\n<p>When building a keras LSTM model, you usually build a \"skeleton\" first, then you <em>compile<\/em>, <em>fit<\/em> and at the end, <em>predict<\/em>. During the <em>compile<\/em> step you need to define your loss function (see <a href=\"https:\/\/keras.io\/models\/sequential\/#sequential-model-methods\" rel=\"nofollow noreferrer\">Keras documentation on Sequential models<\/a>) and a metric, so you could do e.g.<\/p>\n\n<pre><code>model.compile(loss='mean_squared_error', optimizer='sgd', metrics=['mean_squared_error'])\n<\/code><\/pre>\n\n<p>(see <a href=\"https:\/\/keras.io\/metrics\/\" rel=\"nofollow noreferrer\">Keras documentation on metrics<\/a>). Therefore, if you put <code>accuracy<\/code> as a metric in a regression setting, you would not get reasonable results, as this metric is designed only for categorical tasks.<\/p>\n"}]}
+{"QuestionId":57744039,"AnswerCount":3,"Tags":"<python><tensorflow><python-3.7>","CreationDate":"2019-09-01T07:33:56.227","AcceptedAnswerId":null,"OwnerUserId":12005181.0,"Title":"How can I fix this error with Tensorflow?","Body":"<p>I am extremely new to coding with Python and in the tutorial I am following I have to install Tensorflow. However when I import Tensorflow to my python (2.6) I get an error message. I would like to fix this error before I get started actually coding.  <\/p>\n\n<p>I have been trying for 4 days to get rid of errors with Tensorflow in Miniconda3. I have tried installing select versions of Tensorflow but have had the best luck just doing: pip install tensorflow. I get the same error if I am in Command Prompt or PyCharm.<\/p>\n\n<p>This is what I type to get the error message every time:<\/p>\n\n<pre><code>import tensorflow\n<\/code><\/pre>\n\n<p>In videos I have watched this is not the what is supposed to happen, what is suppose to happen at least in PyCharm is a message saying:<\/p>\n\n<p><em>Process finished with exit code 0<\/em><\/p>\n\n<p><strong><em>Here is the exact error message:<\/em><\/strong><\/p>\n\n<pre><code>C:\\Users\\StRUT\\Miniconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:516: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint8 = np.dtype([(\"qint8\", np.int8, 1)])\nC:\\Users\\StRUT\\Miniconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:517: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint8 = np.dtype([(\"quint8\", np.uint8, 1)])\nC:\\Users\\StRUT\\Miniconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:518: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint16 = np.dtype([(\"qint16\", np.int16, 1)])\nC:\\Users\\StRUT\\Miniconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:519: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint16 = np.dtype([(\"quint16\", np.uint16, 1)])\nC:\\Users\\StRUT\\Miniconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:520: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint32 = np.dtype([(\"qint32\", np.int32, 1)])\nC:\\Users\\StRUT\\Miniconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:525: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  np_resource = np.dtype([(\"resource\", np.ubyte, 1)])\nC:\\Users\\StRUT\\Miniconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:541: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint8 = np.dtype([(\"qint8\", np.int8, 1)])\nC:\\Users\\StRUT\\Miniconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:542: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint8 = np.dtype([(\"quint8\", np.uint8, 1)])\nC:\\Users\\StRUT\\Miniconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:543: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint16 = np.dtype([(\"qint16\", np.int16, 1)])\nC:\\Users\\StRUT\\Miniconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:544: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint16 = np.dtype([(\"quint16\", np.uint16, 1)])\nC:\\Users\\StRUT\\Miniconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:545: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint32 = np.dtype([(\"qint32\", np.int32, 1)])\nC:\\Users\\StRUT\\Miniconda3\\envs\\tensorflow2\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:550: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  np_resource = np.dtype([(\"resource\", np.ubyte, 1)])\n<\/code><\/pre>\n\n<p>I hope I gave you guys enough info to help me out because I would really like to start learning to code python with a clean slate.<\/p>\n\n<p>Thanks in advance,\nBlayke<\/p>\n","answers":[{"AnswerId":"57747537","CreationDate":"2019-09-01T16:05:57.590","ParentId":null,"OwnerUserId":"12005181","Title":null,"Body":"<p>I fixed the warning by doing:<\/p>\n\n<pre><code>pip install --upgrade numpy==1.16.4\n<\/code><\/pre>\n\n<p>to which the warning went away completely and the problem was resolved. Thanks for the help!<\/p>\n"},{"AnswerId":"57744547","CreationDate":"2019-09-01T08:48:53.653","ParentId":null,"OwnerUserId":"2126254","Title":null,"Body":"<p>These are only warnings, that TensorFlow is using some stuff that will soon not be available, if you wish to suppress it, you can run the file with<\/p>\n\n<p>python -W ignore filename.py<\/p>\n"},{"AnswerId":"57744535","CreationDate":"2019-09-01T08:47:12.333","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>These are not error messages, but warnings about deprecated functionality in numpy, which is being used by tensorboard and tensorflow. Warnings are not errors, so do not worry about this.<\/p>\n\n<p>The only way to \"fix\" the warnings is by waiting for tensorflow developers to modify the code in order to not use deprecated functionality.<\/p>\n"}]}
+{"QuestionId":57744270,"AnswerCount":1,"Tags":"<image-processing><pytorch><normalization>","CreationDate":"2019-09-01T08:09:56.650","AcceptedAnswerId":null,"OwnerUserId":8919443.0,"Title":"Data Normalization using Pytorch","Body":"<p>I'm starting to work on the classification of images dataset, as many tutorials I followed; it starts by normalizing the data (train and test data)<\/p>\n\n<p>My question is: if I want to normalize the data by <strong>shifting<\/strong> and <strong>scaling<\/strong> it with a factor of 0.5<\/p>\n\n<p>What does this mean '<strong>the factor of something x<\/strong>'?<\/p>\n\n<p>I know that it will be used within <code>.Normalize()<\/code>:<\/p>\n\n<pre><code>transform_train = transforms.Compose([\n    transforms.ToTensor(),\n    transforms.Normalize(),\n])\n\ntransform_test = transforms.Compose([\n    transforms.ToTensor(),\n    transforms.Normalize(),\n])\n<\/code><\/pre>\n\n<p>But I'm a little bit confused about the meaning shift and scale (maybe it's like resize?) by the factor of X<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"57744798","CreationDate":"2019-09-01T09:29:40.130","ParentId":null,"OwnerUserId":"3165451","Title":null,"Body":"<p>Shifting and scaling refers to the color space. What you do is you subtract the mean (shifting to the mean of the pixel <em>values<\/em> of the whole dataset to 0) and divide by the standard deviation (scaling the pixel <em>values<\/em> to [0, 1]. <\/p>\n\n<p>It has nothing to do with modifying the size of the image or the like.<\/p>\n\n<p>In numy you would do something like:<\/p>\n\n<pre><code>mean, std = np.mean(image), np.std(image)\nimage = image - mean\nimage = image \/ std\n<\/code><\/pre>\n\n<p>Note: You wouldn't want to normalize the data bz just 0.5 but by its mean and standard deviation.<\/p>\n"}]}
+{"QuestionId":57744341,"AnswerCount":1,"Tags":"<python><tensorflow><model><export><tensorflow-serving>","CreationDate":"2019-09-01T08:20:31.003","AcceptedAnswerId":57744436.0,"OwnerUserId":7348956.0,"Title":"how to export tf model for serving directly from session (no creating of tf checkpoint) to minimize export time","Body":"<p>I wanted to share my findings on how to export a tf model for serving directly from session without creating model checkpoint. my use case requires minimum time to create a pb file, therefore I wanted to get a model.pb file directly from session without creating model checkpoint.<\/p>\n\n<p>most examples online (and documentation refers to the common case of creating a model checkpoint and loading it in order to create a tf-serving (pb) file. of course this use case is good in case export performance time is not an issue.<\/p>\n","answers":[{"AnswerId":"57744436","CreationDate":"2019-09-01T08:35:05.897","ParentId":null,"OwnerUserId":"7348956","Title":null,"Body":"<pre><code>import tensorflow as tf\nfrom tensorflow.python.framework import importer\noutput_path = '\/export_directory' # be sure to create it before export\ninput_ops = ['name\/s_of_model_input\/s']\noutput_ops = ['name\/s_of_model_output\/s']\nsession = tf.compat.v1.Session()\n\ndef get_ops_dict(ops, graph, name='op_'):\n    out_dict = dict()\n    for i, op in enumerate(ops):\n        out_dict[name + str(i)] = tf.compat.v1.saved_model.build_tensor_info(graph.get_tensor_by_name(op + ':0'))\n    return out_dict\n\n\ndef add_meta_graph(pbtxt_tmp_path, graph_def):\n    with tf.Graph().as_default() as graph:\n        importer.import_graph_def(graph_def, name=\"\")\n        os.unlink(pbtxt_tmp_path)\n\n        # used to rename model input\/outputs\n        inputs_dict = get_ops_dict(input_ops, graph, name='input_')\n        outputs_dict = get_ops_dict(output_ops, graph, name='output_')\n\n        prediction_signature = (\n            tf.compat.v1.saved_model.signature_def_utils.build_signature_def(\n            inputs=inputs_dict,\n            outputs=outputs_dict,\n            method_name=tf.saved_model.PREDICT_METHOD_NAME))\n\n        legacy_init_op = tf.group(tf.compat.v1.tables_initializer(), name='legacy_init_op')\n        builder = tf.compat.v1.saved_model.builder.SavedModelBuilder(output_path+'\/export')\n        builder.add_meta_graph_and_variables(\n            session,\n            tags=[tf.saved_model.SERVING],\n            signature_def_map={\n                tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY: prediction_signature},\n            legacy_init_op=legacy_init_op)\n        builder.save()\n    return prediction_signature\n\ndef export_model(session, output_path, output_ops):\n    graph_def = session.graph_def\n    tf.io.write_graph(graph_or_graph_def=graph_def, logdir=output_path, \n    name='model.pbtxt', as_text=False)\n\n    frozen_graph_def = tf.compat.v1.graph_util.convert_variables_to_constants(\n        session, graph_def, output_ops)\n\n    prediction_signature = add_meta_graph(output_path+'\/model.pbtxt', frozen_graph_def)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57744364,"AnswerCount":0,"Tags":"<neural-network><pytorch><torchvision>","CreationDate":"2019-09-01T08:23:38.570","AcceptedAnswerId":null,"OwnerUserId":4159376.0,"Title":"Pytorch loading images and ground truth images for for training neural network","Body":"<p>I am trying to use custom image data for training and testing a pytorch neural network. I have a folder of images as data and another folder images as ground truth. Files and ground truth images have same name but they are in a separate folder. I want to use them as image and ground truth as pair, for example, 1.jpg feed to network and result of htis picture should compare with other 1.jpg(its ground truth). I want to my neural network output to be image not classification(like image to image translation or pix2pix). <\/p>\n\n<p>I have tried pytorch dataloader and some other docs but they was not what I needed. Some docs I tried:<\/p>\n\n<p><a href=\"https:\/\/www.kaggle.com\/leighplt\/pytorch-tutorial-dataset-data-preparetion-stage\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/leighplt\/pytorch-tutorial-dataset-data-preparetion-stage<\/a><\/p>\n\n<p><a href=\"https:\/\/towardsdatascience.com\/building-efficient-custom-datasets-in-pytorch-2563b946fd9f\" rel=\"nofollow noreferrer\">https:\/\/towardsdatascience.com\/building-efficient-custom-datasets-in-pytorch-2563b946fd9f<\/a><\/p>\n\n<p><a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/data_loading_tutorial.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/tutorials\/beginner\/data_loading_tutorial.html<\/a><\/p>\n","answers":[]}
+{"QuestionId":57744633,"AnswerCount":2,"Tags":"<python><opencv><tensorflow><yolo><openvino>","CreationDate":"2019-09-01T09:02:39.183","AcceptedAnswerId":null,"OwnerUserId":11455772.0,"Title":"Interpretation script for Yolo's region based output to openvino","Body":"<p>Hello I am new to OPENCV\/CVAT, I use openvino to run auto annotation, I want to use YoloV3 for this mission.<\/p>\n\n<p>I need to convert Yolo model to OpenVINO format for opencv\/cvat\/auto_annotation.\n <a href=\"https:\/\/github.com\/opencv\/cvat\/tree\/develop\/cvat\/apps\/auto_annotation\" rel=\"nofollow noreferrer\">https:\/\/github.com\/opencv\/cvat\/tree\/develop\/cvat\/apps\/auto_annotation<\/a>.<\/p>\n\n<p>To annotate a task with a custom model I need to prepare 4 files:<\/p>\n\n<ol>\n<li>Model config (*.xml) - a text file with network configuration.<\/li>\n<li>Model weights (*.bin) - a binary file with trained weights.<\/li>\n<li>Label map (*.json) - a simple json file with label_map dictionary\nlike object with string values for label numbers.<\/li>\n<li>Interpretation script (*.py) - a file used to convert net output\nlayer to a predefined structure which can be processed by CVAT. This\ncode will be run inside a restricted python's environment, but it's\npossible to use some builtin functions like str, int, float, max,\nmin, range.<\/li>\n<\/ol>\n\n<p>I converted Yolo model to OpenVINO format and created xml and bin files. I write the mapping lson file.\nNow I need to write interpretation python script for Yolo's region based output. How can I do that?\nIs there an interrupt file from tensorflow models to openvino?<\/p>\n","answers":[{"AnswerId":"58079892","CreationDate":"2019-09-24T12:02:22.423","ParentId":null,"OwnerUserId":"9465768","Title":null,"Body":"<p>Recommend you to start with Yolo V3 C++ or Python sample.<\/p>\n\n<p>For C++:<\/p>\n\n<p><a href=\"https:\/\/github.com\/opencv\/open_model_zoo\/tree\/master\/demos\/object_detection_demo_yolov3_async\" rel=\"nofollow noreferrer\">https:\/\/github.com\/opencv\/open_model_zoo\/tree\/master\/demos\/object_detection_demo_yolov3_async<\/a><\/p>\n\n<p>For python sample:<\/p>\n\n<p><a href=\"https:\/\/github.com\/opencv\/open_model_zoo\/tree\/master\/demos\/python_demos\/object_detection_demo_yolov3_async\" rel=\"nofollow noreferrer\">https:\/\/github.com\/opencv\/open_model_zoo\/tree\/master\/demos\/python_demos\/object_detection_demo_yolov3_async<\/a><\/p>\n\n<p>Similar discussion can be found in <a href=\"https:\/\/software.intel.com\/en-us\/forums\/computer-vision\/topic\/816875\" rel=\"nofollow noreferrer\">https:\/\/software.intel.com\/en-us\/forums\/computer-vision\/topic\/816875<\/a><\/p>\n"},{"AnswerId":"58776691","CreationDate":"2019-11-09T05:28:20.867","ParentId":null,"OwnerUserId":"2701402","Title":null,"Body":"<p>This is now bundled with CVAT. The interpret script and json file can be found in the repo here: <a href=\"https:\/\/github.com\/opencv\/cvat\/tree\/develop\/utils\/open_model_zoo\/yolov3\" rel=\"nofollow noreferrer\">https:\/\/github.com\/opencv\/cvat\/tree\/develop\/utils\/open_model_zoo\/yolov3<\/a><\/p>\n"}]}
+{"QuestionId":57744975,"AnswerCount":0,"Tags":"<python><keras><attributes><classification>","CreationDate":"2019-09-01T09:57:01.630","AcceptedAnswerId":null,"OwnerUserId":3286607.0,"Title":"How can I save a CNNSequenceClassifier model?","Body":"<p>With other Keras models like <code>Sequential()<\/code>, saving process was easy. For <a href=\"https:\/\/machinelearningmastery.com\/save-load-keras-deep-learning-models\/\" rel=\"nofollow noreferrer\">example<\/a>:<\/p>\n\n<pre><code># 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<\/code><\/pre>\n\n<p>And later the model can be easily loaded from the disk.<\/p>\n\n<p>Now I am using a CNNSequenceClassifier from sequence_classifiers:<\/p>\n\n<pre><code>from sequence_classifiers import CNNSequenceClassifier\nclf = CNNSequenceClassifier(epochs=2)\nclf.fit(x_train, y_train)\n<\/code><\/pre>\n\n<p>But when using the above method to save the <code>clf<\/code> model I get <code>'CNNSequenceClassifier' object has no attribute 'save_weights' or 'to_json'<\/code>.<\/p>\n\n<p>A more general question here could be how can I find out what the attributes of a <code>CNNSequenceClassifier<\/code> (or any other new python thingy!!!) object actually are? The <a href=\"https:\/\/github.com\/aajanki\/sequence-classification\" rel=\"nofollow noreferrer\">documentation<\/a> does not provide enough information on this.<\/p>\n\n<p><strong>Edit 1:<\/strong> I have found out that <code>vars(clf)<\/code> gives me the attributes of this model which are the following. In other words the word \"attributes\" is not what I am looking for. I am looking for some command that finds fit() and other functions(?!) that can be applied on this model.<\/p>\n\n<pre><code>{'embedding_dim': 50,\n 'num_filters': 250,\n 'filter_size': 3,\n 'hidden_dim': 250,\n 'dropout_rates': (0.2, 0.2),\n 'epochs': 2,\n 'batch_size': 32,\n 'verbose': False,\n 'classes_': array(['0', '1', '2', '3'], dtype='&lt;U12'),\n 'net_': &lt;keras.engine.sequential.Sequential at 0x6bb46c4d68&gt;}\n<\/code><\/pre>\n\n<p><strong>Edit 2:<\/strong> I found my answer from previous edit too from this <a href=\"https:\/\/stackoverflow.com\/questions\/34439\/finding-what-methods-a-python-object-has\">SO question<\/a>'s accepted answer. The methods of model clf are as follows. But again there is no saving method among them:<\/p>\n\n<pre><code>['__class__',\n '__delattr__',\n '__dir__',\n '__eq__',\n '__format__',\n '__ge__',\n '__getattribute__',\n '__getstate__',\n '__gt__',\n '__hash__',\n '__init__',\n '__init_subclass__',\n '__le__',\n '__lt__',\n '__ne__',\n '__new__',\n '__reduce__',\n '__reduce_ex__',\n '__repr__',\n '__setattr__',\n '__setstate__',\n '__sizeof__',\n '__str__',\n '__subclasshook__',\n '_build_model',\n '_check_input',\n '_get_param_names',\n '_get_tags',\n 'fit',\n 'get_params',\n 'net_',\n 'predict',\n 'predict_log_proba',\n 'predict_proba',\n 'score',\n 'set_params']\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57745745,"AnswerCount":0,"Tags":"<python><numpy><keras><scikit-learn><deep-learning>","CreationDate":"2019-09-01T11:50:44.170","AcceptedAnswerId":null,"OwnerUserId":10537267.0,"Title":"ValueError while reshaping array to 4dim Keras API format","Body":"<p>DL beginner here. I'm trying to implement LeNet using Keras and apply it on good ol' MNIST. <\/p>\n\n<pre><code>class LeNet:\n@staticmethod\ndef build(height, width, depth, classes):\n    model= Sequential()\n    inputshape= (height,width, depth)\n\n    if K.image_data_format()== 'channels_first':\n        inputshape= (depth, height, width)\n\n    #build model\n    model.add(Conv2d(20, (5,5), padding= \"same\", input_shape= inputshape))\n    model.add(Activation('relu'))\n    model.add(MaxPooling2D(pool_size= (2,2), strides=(2,2)))\n\n    #replicate what's above again\n    model.add(Conv2d(50, (5,5), padding= \"same\"))\n    model.add(Activation('relu'))\n    model.add(MaxPooling2D(pool_size= (2,2), strides=(2,2)))\n\n    #Fully Connected Layers\n    model.add(Flatten())\n    model.add(Dense(500))\n    model.add(Activation('relu'))\n\n    #softmax\n    model.add(Dense(classes))\n    model.add(Activation('softmax'))\n\n\n    return model\n<\/code><\/pre>\n\n<p>Then I load the dataset<\/p>\n\n<pre><code>print('[INFO] Accessing MNIST...')\n\ndataset= loadmat('mnist-original.mat')\ndata=dataset['data']   \n<\/code><\/pre>\n\n<p>Now I'm trying to reshape the data whose current shape is (784, 70000) by<\/p>\n\n<pre><code>data= data.reshape(data.shape[0], 28, 28, 1)\n<\/code><\/pre>\n\n<p>But I get an error that says <strong>ValueError: cannot reshape array of size 54880000 into shape (784,28,28,1)<\/strong><\/p>\n\n<p>Where am I going wrong here? Please help. Thanks.<\/p>\n","answers":[]}
+{"QuestionId":57745893,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2019-09-01T12:11:05.783","AcceptedAnswerId":57746472.0,"OwnerUserId":2110356.0,"Title":"Multi variable\/feature Tensorflow","Body":"<p>i want to categorize the height of person using multilayer keras, if the height more than 170 should return 1 whereas less than 170 return 0<\/p>\n\n<pre><code>data_input = [[0,165], [0,166], [0,167], [0,172], [0,173]]\ndata_output = [0,0,0,1,1]\nmodel = keras.Sequential([\n  keras.layers.Flatten(input_shape=(1,)),\n  keras.layers.Dense(2, activation=tf.nn.relu),\n  keras.layers.Dense(2, activation=tf.nn.softmax)\n])\nmodel.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])\nmodel.fit(data_input, data_output, epochs=5)\npredictions = model.predict([0,184])\nprint(predictions)\n<\/code><\/pre>\n\n<p>but it gave me error \"ValueError: Input arrays should have the same number of samples as target arrays. Found 2 input samples and 5 target samples.\"<\/p>\n","answers":[{"AnswerId":"57746472","CreationDate":"2019-09-01T13:40:13.490","ParentId":null,"OwnerUserId":"10606534","Title":null,"Body":"<p>You Should Use Numpy to wrap your input  <\/p>\n\n<pre><code>data_input = np.array(data_input).reshape(5 ,2)\ndata_output = np.array(data_output).reshape(5)\n<\/code><\/pre>\n\n<p>by the way , why you use ML to calculate a simple condition classfy problem<\/p>\n"}]}
+{"QuestionId":57746073,"AnswerCount":0,"Tags":"<python><keras><regression>","CreationDate":"2019-09-01T12:40:58.720","AcceptedAnswerId":null,"OwnerUserId":6668992.0,"Title":"awful predictions where input test values out scale training input values","Body":"<p>using regression DL i get good prediction results only if test input values are within the range of train inputs values<\/p>\n\n<p>i have tried both minmax and standard scalers none helped achiving improvement, i did so for two regression systems: only continues values and mix of continues and categories input data, results are similar in both systems.  <\/p>\n\n<pre><code>import pandas as pd\nimport numpy as np\nimport tensorflow as tf\nfrom keras import models \nfrom keras.models import Sequential\nfrom keras.layers import Dense, Dropout\nfrom keras import optimizers\nfrom sklearn.model_selection import KFold\nfrom sklearn.preprocessing import MinMaxScaler\ntf.reset_default_graph()\nfrom keras.optimizers import SGD\nimport pickle\n\nsiz=100000\ninp0=np.random.randint(100, 1000000 , size=(siz,3))\nrand0=np.random.randint(-100, 100 , size=(siz,2))\na1=0.2;a2=0.8;a3=2.5;a4=2.6;a5=1.2;a6=0.3\noup1=np.dot(inp0[:,0],a1)+np.dot(inp0[:,1],a2)+np.dot(inp0[:,2],a3)\\\n+rand0[:,0]\noup2=np.dot(inp0[:,0],a4)+np.dot(inp0[:,1],a5)+np.dot(inp0[:,2],a6)\\\n+rand0[:,1]\noup0=np.concatenate((oup1.reshape(siz,1), \noup2.reshape(siz,1)),axis=1)\n\nnormzer_inp = MinMaxScaler()\ninp_norm = normzer_inp.fit_transform(inp0)\ninp_scale = 'normzer_inp_1.sav'\npickle.dump(normzer_inp, open(inp_scale, 'wb'))\n\nnormzer_oup = MinMaxScaler()\noup_norm = normzer_oup.fit_transform(oup0)\noup_scale = 'normzer_oup_1.sav'\npickle.dump(normzer_oup, open(oup_scale, 'wb'))\nX=inp_norm\nY=oup_norm\nopti_SGD = SGD(lr=0.1, momentum=0.9)\nmodel = Sequential()\nmodel.add(Dense(16, input_dim=X.shape[1], activation='relu'))\nmodel.add(Dense(16, activation='relu'))\n\nmodel.add(Dense(Y.shape[1], activation='linear'))\nmodel.compile(loss='mean_squared_error', optimizer=opti_SGD)\n\nhistory = model.fit(X, Y, epochs=100,batch_size=2048, verbose=2)\nmodel.save('DeepL_SelfTst_ContCtg_toCont_5.h5')\n\nsiz=10\ninpt=np.random.randint(100, 1000000 , size=(siz,3))\nrandt=np.random.randint(-100, 100 , size=(siz,2))\noupt1=np.dot(inpt[:,0],a1)+np.dot(inpt[:,1],a2)\\\n+np.dot(inpt[:,2],a3)+randt[:,0]\noupt2=np.dot(inpt[:,0],a4)+np.dot(inpt[:,1],a5)\\\n+np.dot(inpt[:,2],a6)+randt[:,1]\noupt0=np.concatenate((oupt1.reshape(siz,1), \noupt2.reshape(siz,1)),axis=1)\ninp_scale_2 = 'normzer_inp_1.sav' \ninp_norm_do = pickle.load(open(inp_scale_2, 'rb'))\ninp_norm_t = inp_norm_do.transform(inpt)\noup_scale_2 = 'normzer_oup_1.sav' \noup_norm_do = pickle.load(open(oup_scale_2, 'rb'))\n\nnew_model = models.load_model('DeepL_SelfTst_ContCtg_toCont_5.h5')\npredictions = new_model.predict(inp_norm_t)\npred_real=oup_norm_do.inverse_transform(predictions)\npre_sult=pred_real.copy()\npre_sult=np.concatenate((pre_sult, oupt0), axis=1)\ndf = pd.DataFrame({'pred 1': pre_sult[:, 0],\\\n 'real 1': pre_sult[:, 2],\\\n              'pred 2': pre_sult[:, 1],'real 2': pre_sult[:, 3],})\n\n\n        pred 1          real 1          pred 2          real 2\n0   2.889692e+06    2889419.3   2.048063e+06    2048469.7\n1   2.374380e+06    2374681.2   2.684326e+06    2683353.8\n2   2.754795e+06    2754609.5   1.796634e+06    1796445.9\n3   4.967822e+05    496659.8    1.372401e+06    1372884.8\n4   1.785036e+06    1785088.7   2.686990e+06    2687372.3\n5   8.210452e+05    821230.4    1.121044e+06    1120332.6\n6   2.124294e+06    2124235.6   9.910031e+05    990209.6\n7   1.994700e+06    1995039.0   2.210463e+06    2209663.8\n8   3.204532e+06    3204814.4   3.456536e+06    3455951.6\n9   2.889887e+06    2889767.8   1.388561e+06    1388345.4\n\ninpt=np.random.randint(1, 100 , size=(siz,3))\n\n    pred 1          real 1  pred 2          real 2\n0   49198.234375    -23.2   267707.34375    297.0\n1   49355.171875    186.6   267697.28125    114.8\n2   49330.609375    220.5   267652.18750    302.1\n3   49194.960938    104.3   267697.53125    203.5\n4   49245.285156    76.0    267710.71875    87.8\n5   49286.746094    141.3   267733.25000    55.7\n6   49195.906250    153.1   267700.15625    232.9\n7   49348.496094    162.7   267646.90625    404.5\n8   49375.519531    262.6   267653.40625    101.8\n9   49396.093750    275.7   267686.25000    267.5\n<\/code><\/pre>\n\n<p>how can i train the model to be more flexible, or any other way to solve that?<\/p>\n","answers":[]}
+{"QuestionId":57746616,"AnswerCount":0,"Tags":"<python><tensorflow><keras><tensorboard>","CreationDate":"2019-09-01T14:01:06.390","AcceptedAnswerId":null,"OwnerUserId":9477338.0,"Title":"Write value in Keras custom loss layers value to Tensorboard","Body":"<p>I want to monitor many parameters in my loss layers, but since Keras can only show the output of each layer and using <code>tf.Print<\/code> look really mess so I decide to use tensorboard for this.<\/p>\n\n<p>The problems is, when I put this code that copy from the <code>Tensorboard<\/code> callback.<\/p>\n\n<pre><code>summary = tf.Summary()\nsummary_value = summary.value.add()\nsummary_value.simple_value = np.sum(K.eval(loss_conf))\nsummary_value.tag = self.writername+'_loss_conf'\nself.writer.add_summary(summary, self.counter)\nsummary = tf.Summary()\nsummary_value = summary.value.add()\nsummary_value.simple_value = np.sum(K.eval(loss_class))\nsummary_value.tag = self.writername+'_loss_class'\nself.writer.add_summary(summary, self.counter)\n<\/code><\/pre>\n\n<p>It raise this <\/p>\n\n<pre><code>InvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'pred_yolo_1_target' with dtype float and shape [?,?,?,?]\n         [[node pred_yolo_1_target (defined at \/home\/palm\/miniconda3\/envs\/main36\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py:517)  = Placeholder[dtype=DT_FLOAT, shape=[?,?,?,?], _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]()]]\n         [[{{node loss\/pred_yolo_1_loss\/Sum_8\/_2581}} = _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_2683_loss\/pred_yolo_1_loss\/Sum_8\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n<\/code><\/pre>\n\n<p>It runs without any problem if I remove these line or change <code>np.sum(K.eval(x))<\/code> to an integer like 0 or 1.<\/p>\n","answers":[]}
+{"QuestionId":57746718,"AnswerCount":1,"Tags":"<keras><lstm><recurrent-neural-network>","CreationDate":"2019-09-01T14:14:17.990","AcceptedAnswerId":57747105.0,"OwnerUserId":6653602.0,"Title":"Keras lstm and dense layer","Body":"<p>How is dense layer changing the output coming from LSTM layer? How come that from 50 shaped output from previous layer i get output of size 1 from dense layer that is used for prediction?\nLets say i have this basic model:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(50,input_shape=(60,1)))\nmodel.add(Dense(1, activation=\"softmax\"))\n<\/code><\/pre>\n\n<p>Is the Dense layer taking the values coming from previous layer and assigning the probablity(using softmax function) of each of the 50 inputs and then taking it out as an output?<\/p>\n","answers":[{"AnswerId":"57747105","CreationDate":"2019-09-01T15:06:21.050","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>No, <code>Dense<\/code> layers do not work like that, the input has 50-dimensions, and the output will have dimensions equal to the number of neurons, one in this case. The output is a weighted linear combination of the input plus a bias.<\/p>\n\n<p>Note that with the softmax activation, it makes no sense to use it with a one neuron layer, as the softmax is normalized, the only possible output will be constant 1.0. That's probably now what you want.<\/p>\n"}]}
+{"QuestionId":57746905,"AnswerCount":0,"Tags":"<python><tensorflow><anaconda><jupyter>","CreationDate":"2019-09-01T14:38:20.037","AcceptedAnswerId":null,"OwnerUserId":11569633.0,"Title":"Cannot set up tensorflow gpu object detection","Body":"<p>I tried to follow a tutorial on how to setup tensorflow gpu (<a href=\"https:\/\/www.youtube.com\/watch?v=tPq6NIboLSc\" rel=\"nofollow noreferrer\">https:\/\/www.youtube.com\/watch?v=tPq6NIboLSc<\/a>) but on the jupyter step i am getting this error :<a href=\"https:\/\/pastebin.com\/gm3g8cC5\" rel=\"nofollow noreferrer\">https:\/\/pastebin.com\/gm3g8cC5<\/a>\n(ive set in in a pastebin because its very long)\nI dont know what to do, ive also tried others tutorial but none of the others works, they also got errors ( but they are differents) at the jupyter step.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><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 distutils.version import StrictVersion\nfrom collections import defaultdict\nfrom io import StringIO\nfrom matplotlib import pyplot as plt\nfrom PIL import Image\n\n# This is needed since the notebook 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('1.12.0'):\n  raise ImportError('Please upgrade your TensorFlow installation to v1.12.*.')\n<\/code><\/pre>\n\n<p>Here is the part where the errors pops.\nI dont know what is my tensorflow version since when i am trying to import it i am getting <code>OSError: [WinError 193] %1 is not a valid Win32 application<\/code><\/p>\n","answers":[]}
+{"QuestionId":57747129,"AnswerCount":0,"Tags":"<validation><tensorflow><keras><neural-network><backpropagation>","CreationDate":"2019-09-01T15:08:37.610","AcceptedAnswerId":null,"OwnerUserId":5838180.0,"Title":"Can I update weights of keras neural net only if validation improves?","Body":"<p>I am training a neural network in keras and I reach a classical limit - my training accuracy improves with increasing epochs, but my validation accuracy decreases after 9 epochs (see figure).<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/bdlgS.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/bdlgS.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I wonder if I can avoid the decrease of validation accuracy by doing the following: make the keras net only accept the changes to the weights after each epoch if the epoch led to an improvement of the validation accuracy, else reset to the state before the epoch? I assume that the validation is starting to diverge in a big part because after each epoch >9 the weights of the neural net diverge away from similarity to the validation data.<\/p>\n\n<p>So, is my suggestion a good practice and can I achieve it in keras (are there callbacks or options that allow me to update the net only if the validation improved)?<\/p>\n\n<p>Side question: Is my suggestion maybe violating the principle of \"don't use your validation data for training\"? Because I am making implicitly the performance of the neural net a function of my validation data.<\/p>\n","answers":[]}
+{"QuestionId":57747273,"AnswerCount":0,"Tags":"<keras><tensorflow2.0>","CreationDate":"2019-09-01T15:27:28.773","AcceptedAnswerId":null,"OwnerUserId":8164908.0,"Title":"add fully connected layer BEFORE each timestep of RNN","Body":"<p>I have a multivariate time series of 10 features and it is daily for 14 days. I would like to add a fully connected layer before each timestep of an LSTM. I would like to reduce the dimension to 5 which would then be fed to the LSTM.<\/p>\n\n<p>I've looked around stackoverflow and couldn't find any clear solutions. I am basically trying to recreate the same architecture from this paper <a href=\"http:\/\/hanj.cs.illinois.edu\/pdf\/kdd18_cyang.pdf\" rel=\"nofollow noreferrer\">http:\/\/hanj.cs.illinois.edu\/pdf\/kdd18_cyang.pdf<\/a> where it says Solution 2: LSTM with activity embedding.<\/p>\n\n<p>I provided code for what I have so far.<\/p>\n\n<p>I've tried adding an embedding layer before but don't even know if this makes sense because I'm not sure if the embedding layer is performed before EACH timestep of the LSTM and I couldn't get the dimensions to work. I looked at the documentation for the dimensions fo the embedding layer but I don't really understand it.<\/p>\n\n<p>I also tried to add a Dense layer before but this doesn't work anyway because my input data is 3-dimensional. <\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras\n\n# Fake data\nX = np.random.random(size=(500, 14, 10))\ny = np.random.randint(0,2, size=(500))\n\nm = X.shape[0]\ntimesteps = X.shape[1]\noriginal_num_features = X.shape[2]\nreduced_num_features = 5\n\nmodel = keras.models.Sequential([\n    #### not sure if I should be adding as embedding layer here. I want to reduce\n    #### the feature dimensions from 10 to 5 with a fully connected layer.\n    keras.layers.Embedding(input_dim = [timesteps,original_num_features],\n        output_dim = [timesteps, reduced_num_features]),\n    keras.layers.LSTM(64, input_shape=[timesteps, reduced_num_features]),\n    keras.layers.Dense(1, activation='sigmoid')\n])\n\nmodel.compile(loss='binary_crossentropy', \n          optimizer='adam', \n          metrics=['binary_accuracy'])\n\n\nmodel.fit(X, y, epochs=10, batch_size=32)\n<\/code><\/pre>\n\n<p>I got this error but I'm not even sure an embedding would reduce the dimensions like a fully connected layer.<\/p>\n\n<p>TypeError: int() argument must be a string, a bytes-like object or a number, not 'ListWrapper'<\/p>\n","answers":[]}
+{"QuestionId":57747405,"AnswerCount":0,"Tags":"<keras><tensorboard>","CreationDate":"2019-09-01T15:45:18.747","AcceptedAnswerId":null,"OwnerUserId":8894647.0,"Title":"what should be in my code so that when i try to visualize my data with tensorboard","Body":"<p>i am using keras model as for training my deep neural network and I want to integrate tensorboard with it for visulaization purpose but when I do it first it does not show me any data but then after 1 day or 2 my data is updated on tensorboard dashboard . My question is why is it happening like this <\/p>\n\n<p>when i used to access tensorboard via tensorboard --logdir=c:\\logsfiles it gives me port 6006 and when i go there it says no \"You haven\u2019t written any scalar data to your event files.\" but after 1 or 2 days when I access again it has the data<\/p>\n\n<pre><code>NAME = \"living-vs-nonliving-cnn-dense16-dense1-32x2- \n{}\".format(int(time.time()))\n\nmodel = Sequential()\n\nmodel.add(Conv2D(32, (3, 3), input_shape=X.shape[1:])) \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(16, (3,3)))\nmodel.add(MaxPooling2D(pool_size=(2,2)))\n\n\nmodel.add(Flatten()) \nmodel.add(Dense(1))\nmodel.add(Activation(\"sigmoid\"))\n\nmodel.summary()\nmodel.compile(loss=\"binary_crossentropy\",optimizer=\"adam\",metrics= \n['accuracy'])\n\ntboard_log_dir = os.path.join(\"logfiles\", NAME)\ntensorboard = TensorBoard(log_dir=tboard_log_dir)\n\nmodel.fit(X, Y, batch_size=8, epochs=12, validation_split=0.1, callbacks= \n[tensorboard])\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57747718,"AnswerCount":1,"Tags":"<python><keras><deep-learning><nlp><keras-layer>","CreationDate":"2019-09-01T16:33:28.597","AcceptedAnswerId":null,"OwnerUserId":6138564.0,"Title":"Working with multiple input (image,text) data in ResNet50 or any Deep Learning Models","Body":"<p>This is an abstract idea, I dont know the correct pipeline for implementing; I have used a RestNet50 architecture for training a model to classify image into 3 categories; one of the ways i was thinking of exploring was using the textual data of the image;<\/p>\n\n<pre><code>train_gen = image.ImageDataGenerator().flow_from_directory(dataset_path_train, target_size=input_shape[:2], batch_size=batch_size, class_mode='categorical', shuffle=True, seed=seed)\ntest_gen = image.ImageDataGenerator().flow_from_directory(dataset_path_valid, target_size=input_shape[:2], batch_size=batch_size, class_mode='categorical', shuffle=True, seed=seed)\n<\/code><\/pre>\n\n<p>Data prep for model; \nnow for each image i also have {text},{label} as key value pair for individual image;\nif i have to pass \n1. WordtoVec\n2. TFIDF <\/p>\n\n<p>I have read about embedding layer in Keras; I am not sure how to embed the text-data along with test_gen and train_gen in the model( in any intermediate layer or after Flatten().<\/p>\n\n<pre><code>base_model = ResNet50(weights='imagenet', include_top=False, \n\ninput_shape=input_shape) \nfrom keras.models import Model, load_model\nx = base_model.output\nx = Flatten(name='flatten')(x)\npredictions = Dense(3, activation='softmax', name='predictions')(x)\nmodel = Model(inputs=base_model.input, outputs=predictions)\n\nfor layer in model.layers[0:141]:\n    layer.trainable = True\nmodel.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])\nmodel.fit_generator(train_gen,steps_per_epoch=1000 , epochs=2,validation_steps=100, validation_data=test_gen,verbose=1)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57775437","CreationDate":"2019-09-03T16:18:13.460","ParentId":null,"OwnerUserId":"6138564","Title":null,"Body":"<p>So this is solved after reading through this <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">Multi-input and multi-output models<\/a><\/p>\n\n<p>The text input were TFIDF'ed using sklearn tfidf it resulted in 26038 features for matrix ;<\/p>\n\n<pre><code>x = base_model.output\nx = Flatten(name='flatten')(x)\nx2= Input(shape=(26078,), dtype='float32', name='tfidf_input')\ncombinedInput = concatenate([x2, x])\nx = Dense(1024, activation=\"relu\")(combinedInput)\n# adding a regularizer with x  sch as x=Dense(1024,activation=\"relu\",activity_regularizer=regularizers.l1(0.01))\n#x1= Dense(512, activation='softmax', name='predictions')(x)\n\npredictions = Dense(3, activation='softmax', name='predictions')(x)\nmodel = Model(inputs=[base_model.input,x2], outputs=predictions)\n\nfor layer in model.layers[0:141]:\n    layer.trainable = True\n<\/code><\/pre>\n\n<p>I have added an Input layer of dimension (26078,) which would be from the text after transforming it to TFIDF ( now this would add as text feature)<\/p>\n\n<p>While fitting the data on the model <\/p>\n\n<pre><code>model.fit([image_data, text_feature], label,epochs=50,validation_split=0.10,steps_per_epoch=100,validation_steps=8, shuffle = True) \n<\/code><\/pre>\n\n<p>It takes 2 inputs one is the ResNet input of image and the other an array of [datasize,26078].<\/p>\n"}]}
+{"QuestionId":57748144,"AnswerCount":1,"Tags":"<python><machine-learning><keras><lstm><recurrent-neural-network>","CreationDate":"2019-09-01T17:34:30.690","AcceptedAnswerId":null,"OwnerUserId":7876828.0,"Title":"RNN model predicting only one class?","Body":"<p>I am trying to use GloVe embeddings to train a rnn model based on <a href=\"https:\/\/github.com\/kk7nc\/Text_Classification#long-short-term-memory-lstm\" rel=\"nofollow noreferrer\">this article<\/a>.\nI have a labeled data: text(tweets) on one column, labels on another (hate, offensive or neither). \nHowever the model seems to predict only one class in the result.<\/p>\n\n<p>This is the LSTM model:<\/p>\n\n<pre><code>model = Sequential()\nhidden_layer = 3\ngru_node = 32\n\n# model embedding matrix here....\n\nfor i in range(0,hidden_layer):\n    model.add(GRU(gru_node,return_sequences=True, recurrent_dropout=0.2))\n    model.add(Dropout(dropout))\nmodel.add(GRU(gru_node, recurrent_dropout=0.2))\nmodel.add(Dropout(dropout))\nmodel.add(Dense(64, activation='softmax'))\nmodel.add(Dense(nclasses, activation='softmax'))\nstart=time.time()\nmodel.compile(loss='sparse_categorical_crossentropy',optimizer='adam',metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>fitting the 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)\nX_train_Glove,X_test_Glove, word_index, embeddings_index = loadData_Tokenizer(X_train, X_test)\n\nmodel_RNN = Build_Model_RNN_Text(word_index,embeddings_index, 20)    \nmodel_RNN.fit(X_train_Glove,y_train,\n                            validation_data=(X_test_Glove, y_test),\n                            epochs=4,\n                            batch_size=128,\n                            verbose=2)\ny_preds = model_RNN.predict_classes(X_test_Glove)\nprint(metrics.classification_report(y_test, y_preds))\n<\/code><\/pre>\n\n<p>Results:<\/p>\n\n<ol>\n<li>classification report<\/li>\n<\/ol>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/A7F1W.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/A7F1W.png\" alt=\"classifn report\"><\/a><\/p>\n\n<ol start=\"2\">\n<li>Confusion matrix<\/li>\n<\/ol>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/66yPG.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/66yPG.png\" alt=\"confus_mat\"><\/a><\/p>\n\n<p>Am I missing something here?<\/p>\n\n<p><strong>Update<\/strong>:\nthis is what the distribution looks like\n<a href=\"https:\/\/i.stack.imgur.com\/o5EJC.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/o5EJC.png\" alt=\"distribn\"><\/a><\/p>\n\n<p>and the model summary, more or less<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/OyV1T.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/OyV1T.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"57748495","CreationDate":"2019-09-01T18:25:51.983","ParentId":null,"OwnerUserId":"8609330","Title":null,"Body":"<p>How the distribution of your data looks like? The first suggestion is to stratify train\/test split (here is the <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.model_selection.train_test_split.html\" rel=\"nofollow noreferrer\">link<\/a> for the documentation).<\/p>\n\n<p>The second question is how much data do you have in comparison with the complexity of the model? Maybe, your model is so complex, that just do overfitting. You can use the command <a href=\"https:\/\/keras.io\/models\/about-keras-models\/\" rel=\"nofollow noreferrer\">model.summary()<\/a> to see the number of trainable parameters.<\/p>\n"}]}
+{"QuestionId":57748176,"AnswerCount":1,"Tags":"<python><tensorflow><pycharm>","CreationDate":"2019-09-01T17:38:28.723","AcceptedAnswerId":null,"OwnerUserId":2056201.0,"Title":"Cannot install Tensorflow with Pycharm because of Protobuf error","Body":"<p>I installed Pycharm for Windows 10<\/p>\n\n<p><code>pip install tensorflow<\/code> went fine without errors<\/p>\n\n<p>When trying to run a simple project<\/p>\n\n<pre><code>    # Import `tensorflow`\nimport tensorflow as tf\n\n# Initialize two constants\nx1 = tf.constant([1,2,3,4])\nx2 = tf.constant([5,6,7,8])\n\n# Multiply\nresult = tf.multiply(x1, x2)\n\n# Print the result\nprint(result)\n<\/code><\/pre>\n\n<p>I get the error message<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"C:\/Users\/\/PycharmProjects\/RiffRam\/RiffRam.py\", line 2, in &lt;module&gt;\n    import tensorflow as tf\n  File \"C:\\Users\\\\PycharmProjects\\RiffRam\\venv\\lib\\site-packages\\tensorflow\\__init__.py\", line 28, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"C:\\Users\\\\PycharmProjects\\RiffRam\\venv\\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\\\\PycharmProjects\\RiffRam\\venv\\lib\\site-packages\\tensorflow\\core\\framework\\graph_pb2.py\", line 7, in &lt;module&gt;\n    from google.protobuf import descriptor as _descriptor\n  File \"C:\\Users\\\\PycharmProjects\\RiffRam\\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>Can somebody please tell me how I can get a working installation of Tensorflow, it seems that all the versions are mismatched and pip makes no attempt to install correct versions of each package. I don't care if its outdated. <\/p>\n\n<p>What commands do I need to type in to get a working stable version from a clean installation of Pycharm so that the code above would run?\nThank you<\/p>\n","answers":[{"AnswerId":"57748334","CreationDate":"2019-09-01T18:01:50.433","ParentId":null,"OwnerUserId":"2056201","Title":null,"Body":"<p>I got it<\/p>\n\n<pre><code>pip install --upgrade --force-reinstall setuptools==39.1.0\npip install --upgrade --force-reinstall protobuf==3.6.0\npip install --upgrade --force-reinstall tensorflow==1.11\n<\/code><\/pre>\n"}]}
+{"QuestionId":57748399,"AnswerCount":1,"Tags":"<python-3.x><keras>","CreationDate":"2019-09-01T18:11:25.710","AcceptedAnswerId":57748502.0,"OwnerUserId":5208522.0,"Title":"python self keyword in a list","Body":"<p>I am reading a tutorial about keras and came a cross this class that is  inherited by an other class.<\/p>\n\n<pre><code>  class LearningRateDecay:\n  def plot(self, epochs, title=\"Learning Rate Schedule\"):\n    lrs = [self(i) for i in epochs]     \n    plt.style.use(\"ggplot\")\n    plt.figure()\n    plt.plot(epochs, lrs)\n    plt.title(title)\n    plt.xlabel(\"Epoch #\")\n    plt.ylabel(\"Learning Rate\")\n<\/code><\/pre>\n\n<p>in which epochs is defined like this :<\/p>\n\n<pre><code>    N = 100\n    epochs = np.arange(0, N)\n<\/code><\/pre>\n\n<p>and is passed to plot function like this :<\/p>\n\n<pre><code>    a = LearningRateDecay\n    a.plot(epochs,\"Learning Rate Schedule\")\n<\/code><\/pre>\n\n<p>I cant understand what is <strong>self(i)<\/strong> meaning? is this about accessing to self elements or some thing else?\nAny help is greatly appreciated.\nThank you<\/p>\n","answers":[{"AnswerId":"57748502","CreationDate":"2019-09-01T18:28:52.287","ParentId":null,"OwnerUserId":"3323394","Title":null,"Body":"<p><code>self(i)<\/code> refers to calling an object i.e. using their <code>__call__<\/code> method.<\/p>\n\n<p>For example, consider this class:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class MyClass:\n    def __call__(self, arg):\n        print(arg)\n    def method(self):\n        self('test')\n\na = MyClass()\na.method()  # Prints 'test'\na('other')  # Prints 'other'\n<\/code><\/pre>\n\n<p>In your example, this class probably have some defined behavior associated with <code>__call__<\/code> (it can also be inherited from an upper object). Refer to the class documentation to learn about it.<\/p>\n"}]}
+{"QuestionId":57748808,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-01T19:20:00.540","AcceptedAnswerId":57759122.0,"OwnerUserId":4910495.0,"Title":"How should I deal with variable sample length in neural networks during training and prediction?","Body":"<p>So my end goal is to build a model with Keras that can predict the gas or brake percentages based on the radar data fed in. However, my radar data can vary from anywhere to 1 point to 16 points (objects) all with their own metadata (velocity, lateral position, etc). An example of my input shape would be <code>[[4.5, -2.2], [5.6, 1.1]]<\/code>, being a list of the radar tracks, velocity and lateral position of that object inside each list.<\/p>\n\n<p>I know with classification problems we should set this missing data to a class never used except for missing data, though with regression I'm a little confused. If I set missing data to all zeros, how is the model supposed to differentiate between a zero being the velocity of a point, or that the point is missing?<\/p>\n\n<p>My first thought was to introduce a new data point in all the points and set it to either 0 to 1, corresponding to whether that point is 'real'\/to be considered or not. I'm just not sure if that's the best way of going about it.<\/p>\n","answers":[{"AnswerId":"57759122","CreationDate":"2019-09-02T14:59:50.480","ParentId":null,"OwnerUserId":"3694044","Title":null,"Body":"<p>I do not think an extra boolean would be necessary as there should only be few cases where the positional data is zero together with the velocity.<\/p>\n\n<p>However, if you do not want to use zero-padding you could also take advantage of \nRecurrent neural networks (RNNs), which accept a variable input length.<\/p>\n\n<p>Finally, you can also look at similar questions as answered <a href=\"https:\/\/ai.stackexchange.com\/questions\/2008\/how-can-neural-networks-deal-with-varying-input-sizes\">here<\/a><\/p>\n"}]}
+{"QuestionId":57748895,"AnswerCount":1,"Tags":"<python><windows><tensorflow>","CreationDate":"2019-09-01T19:31:17.043","AcceptedAnswerId":null,"OwnerUserId":2056201.0,"Title":"No module named Scimage in Pycharm when following tensorflow tutorial","Body":"<p>I'm following this tutorial using windows 10 and pycharm<\/p>\n\n<p><a href=\"https:\/\/www.datacamp.com\/community\/tutorials\/tensorflow-tutorial\" rel=\"nofollow noreferrer\">https:\/\/www.datacamp.com\/community\/tutorials\/tensorflow-tutorial<\/a><\/p>\n\n<p>Below is my code so far<\/p>\n\n<p>As suggested, I installed scimage using <code>pip install scikit-image<\/code>\nBecause <code>pip install scimage<\/code> doesnt work and tells me to install <code>scikit-image<\/code><\/p>\n\n<p>This worked for installing tensorflow, so I know its installing in the correct directory, but it wont run because it wont recognize <code>import scimage<\/code> <\/p>\n\n<p>Why wont it recognize this? Am I supposed to import something different?<\/p>\n\n<p>I tried <code>import scikit-image as scimage<\/code> it didn't find it either. <\/p>\n\n<pre><code># Import `tensorflow`\nimport tensorflow as tf\nimport numpy as np\nimport os\nimport scimage\n\n# Initialize two constants\nx1 = tf.constant([1,2,3,4])\nx2 = tf.constant([5,6,7,8])\n\n# Multiply\nresult = tf.multiply(x1, x2)\n\n# Print the result\nprint(result)\n\n\ndef load_data(data_directory):\n    directories = [d for d in os.listdir(data_directory)\n                   if os.path.isdir(os.path.join(data_directory, d))]\n    labels = []\n    images = []\n    for d in directories:\n        label_directory = os.path.join(data_directory, d)\n        file_names = [os.path.join(label_directory, f)\n                      for f in os.listdir(label_directory)\n                      if f.endswith(\".ppm\")]\n        for f in file_names:\n            images.append(skimage.data.imread(f))\n            labels.append(int(d))\n    return images, labels\n\nROOT_PATH = \"C:\\\\Users\\\\dm\\\\PycharmProjects\\\\test\"\ntrain_data_directory = os.path.join(ROOT_PATH, \"Training\")\ntest_data_directory = os.path.join(ROOT_PATH, \"Testing\")\n<\/code><\/pre>\n","answers":[{"AnswerId":"57748912","CreationDate":"2019-09-01T19:33:47.620","ParentId":null,"OwnerUserId":"2255479","Title":null,"Body":"<p>You need to use: <\/p>\n\n<pre><code>import skimage\n<\/code><\/pre>\n"}]}
+{"QuestionId":57750023,"AnswerCount":0,"Tags":"<python><tensorflow><keras><conv-neural-network><autoencoder>","CreationDate":"2019-09-01T22:54:08.973","AcceptedAnswerId":null,"OwnerUserId":10762045.0,"Title":"How to split the encoder and decoder in autoencoder?","Body":"<p>I'm trying to split my autoencoder but whenever I try I get this error and I have no Idea why I'm getting it. I am running this in python 3.7 64bit on windows 10.<\/p>\n\n<pre><code>inp = open('train.csv',\"rb\")\nX = pickle.load(inp)\nX = X\/255.0\nX = np.array(X)\nX = np.reshape(X,(-1,x,y,h))\n\nDATA = \"C:\/Users\/dalto\/Documents\/geo4\/dataset\"\ndataout = r\"C:\\Users\\dalto\\Documents\\geo4\\outset\\video\"\n\n\ninput_img = keras.layers.Input(shape = (x, y, h))\n\n#make encoder\nx = Conv2D(32, (3, 3), activation='relu', padding='same')(input_img)\nx = MaxPooling2D((2, 2), padding='same')(x)\nx = Conv2D(32, (3, 3), activation='relu', padding='same')(x)\nencoded = MaxPooling2D((2, 2), padding='same')(x)\n\nencoder = Model(input_img, encoded)\n\nencoder.summary()\n\n#make decoder\ndecoder_input= Input(shape = (32,16,32))\n\ndecoder = Conv2D(32, (3, 3), activation='relu', padding='same')(decoder_input)\nx = UpSampling2D((2, 2))(decoder)\nx = Conv2D(32, (3, 3), activation='relu', padding='same')(x)\nx = UpSampling2D((2, 2))(x)\ndecoded = Conv2D(1, (3, 3), activation='sigmoid', padding='same')(x)\n\ndecoder = Model(decoder_input, decoded)\n\n#combine them both\nauto_input = Input(shape=(x, y, h))\nencoded = encoder(auto_input)\ndecoded = decoder(encoded)\n\nautoencoder = Model(auto_input, decoded)\nautoencoder.compile(loss='mean_squared_error', optimizer = \nkeras.optimizers.RMSprop())\n#train\nautoencoder_train = autoencoder.fit(X, X, \nbatch_size=batch_size,epochs=epochs,verbose=1\n<\/code><\/pre>\n\n<p>I would like to be able to encoded something and save the encoded file and then run the decoder on it to get the enlarged file. This is the error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\eager\\execute.py\", line 146, in make_shape\n    shape = tensor_shape.as_shape(v)\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\framework\\tensor_shape.py\", line 1211, in as_shape\n    return TensorShape(shape)\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\framework\\tensor_shape.py\", line 781, in __init__\nself._dims = [as_dimension(d) for d in dims_iter]\nFile \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\framework\\tensor_shape.py\", line 781, in &lt;listcomp&gt;\n    self._dims = [as_dimension(d) for d in dims_iter]\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\framework\\tensor_shape.py\", line 723, in as_dimension\nreturn Dimension(value)\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\framework\\tensor_shape.py\", line 192, in __init__\nself._value = int(value)\nTypeError: int() argument must be a string, a bytes-like object or a number, not 'Tensor'\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n   File \"C:\\Users\\dalto\\Documents\\geo4\\train.py\", line 54, in &lt;module&gt;\n     auto_input = Input(shape=(x, y, h))\n   File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\input_layer.py\", line 256, in Input\ninput_tensor=tensor)\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\input_layer.py\", line 123, in __init__\n    name=self.name)\n   File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py\", line 1001, in placeholder\n     x = array_ops.placeholder(dtype, shape=shape, name=name)\n   File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\ops\\array_ops.py\", line 2129, in placeholder\nreturn gen_array_ops.placeholder(dtype=dtype, shape=shape, name=name)\n   File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\ops\\gen_array_ops.py\", line 7398, in placeholder\n     shape = _execute.make_shape(shape, \"shape\")\n  File \"C:\\Users\\dalto\\AppData\\Local\\Programs\\Python\\Python37\\lib\\site-packages\\tensorflow\\python\\eager\\execute.py\", line 148, in make_shape\nraise TypeError(\"Error converting %s to a TensorShape: %s.\" % (arg_name, e))\nTypeError: Error converting shape to a TensorShape: int() argument must be a string, a bytes-like object or a number, not 'Tensor'.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57751280,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-02T04:01:04.673","AcceptedAnswerId":null,"OwnerUserId":11954746.0,"Title":"how to interpret the timeline plot and the profiler output?","Body":"<p>I want to test the performance of my model in <code>pb<\/code> format, so I have used the profiler module in TensorFlow, its output includes requested bytes and peak bytes. Then I also use timeline module to produce a timeline JSON, but I found the timeline JSON show the CPU allocated memory is not the same as the profiler output peak bytes or requested bytes, what is the real meaning of this output?<\/p>\n\n<p>profiler output:<\/p>\n\n<pre><code>node name | requested bytes | peak bytes | total execution time | accelerator execution time | cpu execution time | # parameters\n_TFProfRoot (--\/400.04MB, --\/400.04MB, --\/279.18ms, --\/0us, --\/279.18ms, --\/1 params)\n<\/code><\/pre>\n\n<p>timeline(the peak value i select)<\/p>\n\n<pre><code>cpu cpu 12.321044921875(time)   136018540(value)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57751304,"AnswerCount":0,"Tags":"<python><tensorflow><nlp>","CreationDate":"2019-09-02T04:06:50.410","AcceptedAnswerId":null,"OwnerUserId":4544413.0,"Title":"Keep TensorFlow session open for API calls","Body":"<p>I'm working on the  GTP-2  library for NLP, specifically the nshepperd repository <a href=\"https:\/\/github.com\/nshepperd\/gpt-2\" rel=\"nofollow noreferrer\">https:\/\/github.com\/nshepperd\/gpt-2<\/a><\/p>\n\n<p>My objective here is to modify the <code>interactive_conditional_samples.py<\/code> so I can make calls to this function from a custom API.<\/p>\n\n<p>My issue is at this <strong>with tf.Session(graph=tf.Graph()) as sess<\/strong> secction of the code: <\/p>\n\n<pre><code>with tf.Session(graph=tf.Graph()) as sess:\n    context = tf.placeholder(tf.int32, [batch_size, None])\n    np.random.seed(seed)\n    tf.set_random_seed(seed)\n    output = sample.sample_sequence(\n        hparams=hparams, length=length,\n        context=context,\n        batch_size=batch_size,\n        temperature=temperature, top_k=top_k, top_p=top_p\n    )\n\n    saver = tf.train.Saver()\n    ckpt = tf.train.latest_checkpoint(os.path.join('models', model_name))\n    saver.restore(sess, ckpt)\n\n    for sample in range(1):\n        raw_text = input(\"Model prompt &gt;&gt;&gt; \")\n        while not raw_text:\n            print('Prompt should not be empty!')\n            raw_text = input(\"Model prompt &gt;&gt;&gt; \")\n        context_tokens = enc.encode(raw_text)\n        generated = 0\n        for _ in range(nsamples \/\/ batch_size):\n            out = sess.run(output, feed_dict={\n                context: [context_tokens for _ in range(batch_size)]\n            })[:, len(context_tokens):]\n            for i in range(batch_size):\n                generated += 1\n                text = enc.decode(out[i])\n                print(\"=\" * 40 + \" SAMPLE \" + str(generated) + \" \" + \"=\" * 40)\n                print(text)\n        print(\"=\" * 80)\n<\/code><\/pre>\n\n<p>Basically I want to separate the section 1 of the code, that loads the model from section to which makes the prediction, so I don't have that wait time for each predicction: <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/W132Z.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/W132Z.png\" alt=\"enter image description here\"><\/a>  <\/p>\n\n<p>I tried changing <\/p>\n\n<pre><code>tf.Session(graph=tf.Graph()) as sess:\n<\/code><\/pre>\n\n<p>for    <\/p>\n\n<pre><code>sess = tf.Session(graph=tf.Graph())     \n<\/code><\/pre>\n\n<p>But I get this error: <\/p>\n\n<pre><code>RuntimeError: The Session graph is empty.  Add operations to the graph before calling run().\n<\/code><\/pre>\n\n<p>The complete code to that file is here: <a href=\"https:\/\/github.com\/nshepperd\/gpt-2\/blob\/finetuning\/src\/interactive_conditional_samples.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/nshepperd\/gpt-2\/blob\/finetuning\/src\/interactive_conditional_samples.py<\/a><\/p>\n","answers":[]}
+{"QuestionId":57751417,"AnswerCount":3,"Tags":"<python><tensorflow><keras><tf.keras>","CreationDate":"2019-09-02T04:25:19.550","AcceptedAnswerId":null,"OwnerUserId":4339936.0,"Title":"What meant by sequential model in Keras","Body":"<p>I have recently started working Tensorflow for deep learning. I found this statement <code>model = tf.keras.models.Sequential()<\/code>  bit different. I couldn't understand what is actually meant and is there any other models as well for deep learning?\nI worked  a lot on MatconvNet (Matlab library for convolutional neural network). never saw any sequenstial defination in that.<\/p>\n","answers":[{"AnswerId":"57752374","CreationDate":"2019-09-02T06:35:57.220","ParentId":null,"OwnerUserId":"11007513","Title":null,"Body":"<p>There are two ways to build Keras models: sequential and functional.<\/p>\n\n<p>The sequential API allows you to create models layer-by-layer for most problems. It is limited in that it does not allow you to create models that share layers or have multiple inputs or outputs.<\/p>\n\n<p>Alternatively, the functional API allows you to create models that have a lot more flexibility as you can easily define models where layers connect to more than just the previous and next layers. In fact, you can connect layers to (literally) any other layer. As a result, creating complex networks such as siamese networks and residual networks become possible.<\/p>\n\n<p>for more details visit : <a href=\"https:\/\/machinelearningmastery.com\/keras-functional-api-deep-learning\/\" rel=\"nofollow noreferrer\">https:\/\/machinelearningmastery.com\/keras-functional-api-deep-learning\/<\/a><\/p>\n"},{"AnswerId":"57751512","CreationDate":"2019-09-02T04:41:55.677","ParentId":null,"OwnerUserId":"2991872","Title":null,"Body":"<blockquote>\n  <p>The <a href=\"https:\/\/keras.io\/getting-started\/sequential-model-guide\/\" rel=\"nofollow noreferrer\"><code>Sequential<\/code><\/a> model is a linear stack of layers.<\/p>\n<\/blockquote>\n\n<p>The common architecture of ConvNets is a sequential architecture. However, some architectures are not linear stacks. For example, siamese networks are two parallel neural networks with some shared layers. <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">More examples here<\/a>.<\/p>\n"},{"AnswerId":"57751447","CreationDate":"2019-09-02T04:30:55.947","ParentId":null,"OwnerUserId":"7484853","Title":null,"Body":"<p>From the definition of <strong>Keras<\/strong> documentation the Sequential model is a <strong>linear stack of layers<\/strong>.You can create a Sequential model by passing a list of layer instances to the constructor:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense, Activation\n\nmodel = Sequential([\n    Dense(32, input_shape=(784,)),\n    Activation('relu'),\n    Dense(10),\n    Activation('softmax'),\n])\n<\/code><\/pre>\n\n<p>You can also simply add layers via the .add() method:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(32, input_dim=784))\nmodel.add(Activation('relu'))\n<\/code><\/pre>\n\n<p>For More details click <a href=\"https:\/\/keras.io\/getting-started\/sequential-model-guide\/\" rel=\"nofollow noreferrer\">here<\/a> <\/p>\n"}]}
+{"QuestionId":57751787,"AnswerCount":0,"Tags":"<deep-learning><computer-vision><caffe>","CreationDate":"2019-09-02T05:23:53.970","AcceptedAnswerId":null,"OwnerUserId":8198304.0,"Title":"error when I run .\/data\/VOC0712\/create_list.sh","Body":"<p>I use <a href=\"https:\/\/github.com\/weiliu89\/caffe\" rel=\"nofollow noreferrer\">Caffe framework<\/a> on google colab (cpu &amp; python2).<\/p>\n\n<p>I got this error when I run <code>.\/data\/VOC0712\/create_list.sh<\/code><\/p>\n\n<pre><code> Create list for VOC2007 trainval...\ncp: cannot stat '\/root\/data\/VOCdevkit\/\/VOC2007\/ImageSets\/Main\/trainval.txt': No such file or directory\nsed: can't read \/content\/caffe\/data\/VOC0712\/trainval_img.txt: No such file or directory\nsed: can't read \/content\/caffe\/data\/VOC0712\/trainval_img.txt: No such file or directory\ncp: cannot stat '\/root\/data\/VOCdevkit\/\/VOC2007\/ImageSets\/Main\/trainval.txt': No such file or directory\nsed: can't read \/content\/caffe\/data\/VOC0712\/trainval_label.txt: No such file or directory\nsed: can't read \/content\/caffe\/data\/VOC0712\/trainval_label.txt: No such file or directory\npaste: \/content\/caffe\/data\/VOC0712\/trainval_img.txt: No such file or directory\nCreate list for VOC2012 trainval...\ncp: cannot stat '\/root\/data\/VOCdevkit\/\/VOC2012\/ImageSets\/Main\/trainval.txt': No such file or directory\nsed: can't read \/content\/caffe\/data\/VOC0712\/trainval_img.txt: No such file or directory\nsed: can't read \/content\/caffe\/data\/VOC0712\/trainval_img.txt: No such file or directory\ncp: cannot stat '\/root\/data\/VOCdevkit\/\/VOC2012\/ImageSets\/Main\/trainval.txt': No such file or directory\nsed: can't read \/content\/caffe\/data\/VOC0712\/trainval_label.txt: No such file or directory\nsed: can't read \/content\/caffe\/data\/VOC0712\/trainval_label.txt: No such file or directory\npaste: \/content\/caffe\/data\/VOC0712\/trainval_img.txt: No such file or directory\nCreate list for VOC2007 test...\ncp: cannot stat '\/root\/data\/VOCdevkit\/\/VOC2007\/ImageSets\/Main\/test.txt': No such file or directory\nsed: can't read \/content\/caffe\/data\/VOC0712\/test_img.txt: No such file or directory\nsed: can't read \/content\/caffe\/data\/VOC0712\/test_img.txt: No such file or directory\ncp: cannot stat '\/root\/data\/VOCdevkit\/\/VOC2007\/ImageSets\/Main\/test.txt': No such file or directory\nsed: can't read \/content\/caffe\/data\/VOC0712\/test_label.txt: No such file or directory\nsed: can't read \/content\/caffe\/data\/VOC0712\/test_label.txt: No such file or directory\npaste: \/content\/caffe\/data\/VOC0712\/test_img.txt: No such file or directory\n.\/data\/VOC0712\/create_list.sh: line 41: \/content\/caffe\/data\/VOC0712\/..\/..\/build\/tools\/get_image_size: No such file or directory \n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57751862,"AnswerCount":1,"Tags":"<tensorflow><object-detection-api>","CreationDate":"2019-09-02T05:34:36.103","AcceptedAnswerId":null,"OwnerUserId":12008153.0,"Title":"weight_quant\/min, max restored in the check point are different with the actual min\/max of the weight","Body":"<p>I'm doing some work with tensorflow object detection API. When I get the checkpoint (with quantization), I found that the weight_quant\/min and the weight_quant\/max are different with the actual min\/max of the weight. Since I want to transfer the model to pytorch, it makes me confused to handle the min\/max value in tf model and pytorch model. <\/p>\n\n<p>My environment is tf1.9 with object detection API, ubuntu 16.04. I trained a mobilenet v1 model, and checked some weights in the model, here is an example:<\/p>\n\n<blockquote>\n  <p>('FeatureExtractor\/MobilenetV1\/MobilenetV1\/Conv2d_1_depthwise\/weights_quant\/min:0',\n  -3.2073412) ('FeatureExtractor\/MobilenetV1\/MobilenetV1\/Conv2d_1_depthwise\/weights_quant\/max:0',\n  5.7885404)<\/p>\n  \n  <p>('FeatureExtractor\/MobilenetV1\/Conv2d_1_depthwise\/depthwise_weights:0',<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/mdfoa.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<blockquote>\n  <p>('actual min and max: ', -2.3553603, 2.1638136)<\/p>\n<\/blockquote>\n\n<p>As shown above, the actual min\/max value in 'FeatureExtractor\/MobilenetV1\/Conv2d_1_depthwise\/depthwise_weights:0' tensor is -2.3553603 and 2.1638136, while the weights_quant\/min and weights_quant\/max value are different.<\/p>\n\n<p>I tried assign the actual min\/max value to weights_quant\/min and weights_quant\/max, but the result model is failed to detect any object.<\/p>\n\n<p>So is there some other operation in recording the min\/max value? How the recorded min\/max work in inference?<\/p>\n","answers":[{"AnswerId":"57756566","CreationDate":"2019-09-02T11:57:05.900","ParentId":null,"OwnerUserId":"12008153","Title":null,"Body":"<p>Got it....\nThe weight_quant\/min and weight_quant\/max value are calculated from a conv layer that has merged the corresponding BN layer. <\/p>\n"}]}
+{"QuestionId":57751922,"AnswerCount":0,"Tags":"<tensorflow><ctc>","CreationDate":"2019-09-02T05:43:53.360","AcceptedAnswerId":null,"OwnerUserId":12008247.0,"Title":"What is the correct tensor values or shape for tf.nn.ctc_loss?","Body":"<p>I'm trying to do this <a href=\"https:\/\/stackoverflow.com\/questions\/44762631\/how-to-use-tf-nn-ctc-loss-in-cnnctc-network\">How to use tf.nn.ctc_loss in cnn+ctc network<\/a> and thus I tried modifying this repo <a href=\"https:\/\/github.com\/ugnelis\/tensorflow-rnn-ctc\/blob\/master\/train.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ugnelis\/tensorflow-rnn-ctc\/blob\/master\/train.py<\/a> .I tried to do what the approved answer says but I'm getting an incorrect tensor value.<\/p>\n\n<p>I have tried passing from the dense layer directly to the ctc without reshaping it to (?, ?, num_classes) but I got an error that read Invalid argument: sequence_length(2) &lt;= 463 instead.<\/p>\n\n<p>Here's the part where I change the RNN to CNN<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>            shape = tf.shape(inputs_placeholder)\n            batch_size, max_time_steps = shape[0], shape[1]\n\n            # Convolutional Layer #1\n            conv1 = layers.Conv2D(\n                    filters=32,\n                    kernel_size=(3, 5),\n                    activation='relu')(inputs_placeholder)\n\n            # Pooling Layer #1\n            pool1 = layers.MaxPooling2D((3, 1), strides=1)(conv1)\n\n            # Convolutional Layer #2\n            conv2 = layers.Conv2D(\n                    filters=64,\n                    kernel_size=(3, 5),\n                    padding=\"same\",\n                    activation='relu')(pool1)\n\n            # Dense Layer as per step 2 and 3\n            flat1 = tf.reshape(conv2, [batch_size, -1, 9*64])\n            dense = layers.TimeDistributed(layers.Dense(1024))(flat1)\n\n#             step 4\n            logits = tf.reshape(dense, [batch_size, -1, NUM_CLASSES])\n\n            # Time is major.\n            logits = tf.transpose(dense, (1, 0, 2))\n<\/code><\/pre>\n\n<p>Oh I don't know if this is related or not but I also added a new dimension to the input since I used Conv2D<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>                    batch_train_inputs = train_inputs[indexes]\n                    batch_train_inputs =  batch_train_inputs[:, :, :, np.newaxis]\n<\/code><\/pre>\n\n<p>This is the error I get when reshaping it during step 4 \"Invalid argument: Input to reshape is a tensor with 1896448 values, but the requested shape requires a multiple of 112\" which also confused me, like where is this 112 value came from? Since NUM_CLASSES is 28 shouldn't it request multiple of 28?<\/p>\n","answers":[]}
+{"QuestionId":57752017,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-02T05:55:56.073","AcceptedAnswerId":57968270.0,"OwnerUserId":9839954.0,"Title":"tensorflow deeplabv3+ train from scratch train.sh code to recurring 82.2% in VOC val","Body":"<p>DATA: VOC 2012 augmented datasets: I am training <a href=\"https:\/\/gist.githubusercontent.com\/sun11\/2dbda6b31acc7c6292d14a872d0c90b7\/raw\/5f5a5270089239ef2f6b65b1cc55208355b5acca\/trainaug.txt\" rel=\"nofollow noreferrer\">train_aug<\/a> with 10582 <a href=\"https:\/\/www.dropbox.com\/s\/oeu149j8qtbs1x0\/SegmentationClassAug.zip?dl=0&amp;file_subpath=%2FSegmentationClassAug\" rel=\"nofollow noreferrer\">annotations<\/a>.<\/p>\n\n<p>CODE: I am using official deeplabv3+ <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/deeplab\" rel=\"nofollow noreferrer\">code<\/a>.<strong>I didn't change the code<\/strong> except bash code.<\/p>\n\n<p>Pretrained weight: xception_65_imagenet_coco from <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/deeplab\/g3doc\/model_zoo.md\" rel=\"nofollow noreferrer\">official zoo<\/a><\/p>\n\n<p>So, this is my train.sh:<\/p>\n\n<pre><code>python \"${WORK_DIR}\"\/train.py \\\n  --logtostderr \\\n  --train_split=\"train_aug\" \\\n  --model_variant=\"xception_65\" \\\n  --atrous_rates=6 \\\n  --atrous_rates=12 \\\n  --atrous_rates=18 \\\n  --output_stride=16 \\\n  --decoder_output_stride=4 \\\n  --train_crop_size=\"513,513\" \\\n  --train_batch_size=15 \\\n  --training_number_of_steps=30000 \\\n  --fine_tune_batch_norm=False \\\n  --num_clones=5 \\\n  --base_learning_rate=0.007 \\\n  --tf_initial_checkpoint=\"${COCO_PRE}\/x65-b2u1s2p-d48-2-3x256-sc-cr300k_init.ckpt\" \\\n  --train_logdir=\"${TRAIN_LOGDIR}\" \\\n  --dataset_dir=\"${PASCAL_DATASET}\"\\\n  --initialize_last_layer=False\n<\/code><\/pre>\n\n<p>Result: I think it should be 82.2% with my configuration.But I got 80%on eval.OS=16 and 80.15% on eval.OS=8 in 30k steps<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/dgucJ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/dgucJ.png\" alt=\"train loss\"><\/a><\/p>\n\n<p>So my question is : How can I get 82.2%?<\/p>\n\n<p>edit:09.02.2019:-----------------<\/p>\n\n<p>I have notice that the <code>fine_tune_batch_norm=false<\/code>.<\/p>\n\n<p>and in train.py:<\/p>\n\n<blockquote>\n  <p>Set to True if one wants to fine-tune the batch norm parameters in\n  DeepLabv3<\/p>\n<\/blockquote>\n\n<p>So I decide to try <code>fine_tune_batch_norm=true<\/code> cause training from scratch need to change the BN parameters. \nedit------09\/07--------------------\nStill not working with:<\/p>\n\n<pre><code>python \"${WORK_DIR}\"\/train.py \\\n  --logtostderr \\\n  --train_split=\"train_aug\" \\\n  --model_variant=\"xception_65\" \\\n  --atrous_rates=6 \\\n  --atrous_rates=12 \\\n  --atrous_rates=18 \\\n  --output_stride=16 \\\n  --decoder_output_stride=4 \\\n  --train_crop_size=\"513,513\" \\\n  --train_batch_size=15 \\\n  --training_number_of_steps=100000 \\\n  --fine_tune_batch_norm=true \\\n  --num_clones=5 \\\n  --base_learning_rate=0.007 \\\n  --tf_initial_checkpoint=\"${COCO_PRE}\/x65-b2u1s2p-d48-2-3x256-sc-cr300k_init.ckpt\" \\\n  --train_logdir=\"${TRAIN_LOGDIR}\" \\\n  --dataset_dir=\"${PASCAL_DATASET}\"\\\n  --initialize_last_layer=False\n<\/code><\/pre>\n\n<p>This time the result is even worse.<\/p>\n","answers":[{"AnswerId":"57968270","CreationDate":"2019-09-17T06:08:50.690","ParentId":null,"OwnerUserId":"9839954","Title":null,"Body":"<p>I reproduce the result. I got 81.5%. I use 40000 steps for the first round, I think 30000 steps might better. <\/p>\n\n<p>Bash Code:<\/p>\n\n<pre><code>python \"${WORK_DIR}\"\/train.py \\\n  --logtostderr \\\n  --train_split=\"train_aug\" \\\n  --model_variant=\"xception_65\" \\\n  --atrous_rates=6 \\\n  --atrous_rates=12 \\\n  --atrous_rates=18 \\\n  --output_stride=16 \\\n  --decoder_output_stride=4 \\\n  --train_crop_size=\"513,513\" \\\n  --train_batch_size=24 \\\n  --base_learning_rate=0.007 \\\n  --training_number_of_steps=30000 \\\n  --fine_tune_batch_norm=true \\\n  --num_clones=8 \\\n  --tf_initial_checkpoint=\"${COCO_PRE}\/x65-b2u1s2p-d48-2-3x256-sc-cr300k_init.ckpt\" \\\n  --train_logdir=\"${TRAIN_LOGDIR}\" \\\n  --dataset_dir=\"${PASCAL_DATASET}\"\\\n  --initialize_last_layer=true\n\npython \"${WORK_DIR}\"\/train.py \\\n  --logtostderr \\\n  --train_split=\"train\" \\\n  --model_variant=\"xception_65\" \\\n  --atrous_rates=6 \\\n  --atrous_rates=12 \\\n  --atrous_rates=18 \\\n  --output_stride=16 \\\n  --decoder_output_stride=4 \\\n  --train_crop_size=\"513,513\" \\\n  --train_batch_size=24 \\\n  --training_number_of_steps=60000 \\\n  --fine_tune_batch_norm=false \\\n  --num_clones=8 \\\n  --base_learning_rate=0.01 \\\n  --tf_initial_checkpoint=\"${COCO_PRE}\/x65-b2u1s2p-d48-2-3x256-sc-cr300k_init.ckpt\" \\\n  --train_logdir=\"${TRAIN_LOGDIR}\" \\\n  --dataset_dir=\"${PASCAL_DATASET}\"\\\n  --initialize_last_layer=true\n<\/code><\/pre>\n\n<p>reference: <a href=\"https:\/\/github.com\/tensorflow\/models\/issues\/5116\" rel=\"nofollow noreferrer\">github<\/a><\/p>\n"}]}
+{"QuestionId":57752462,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><tensorflow-datasets>","CreationDate":"2019-09-02T06:43:47.897","AcceptedAnswerId":null,"OwnerUserId":2756531.0,"Title":"Tensorflow : Is it possible to identify the data is used for training?","Body":"<p>I have created text classification model(.pb) using tensorflow. Prediction is good.<\/p>\n\n<p>Is it possible to check the sentence using for prediction is already used to train the model or not. I need to retrain the model when new sentence is given to model to  predict.<\/p>\n","answers":[{"AnswerId":"57753158","CreationDate":"2019-09-02T07:45:51.857","ParentId":null,"OwnerUserId":"11519648","Title":null,"Body":"<p>I did some research and couldn't find a way to get the train data only with the pb file because that file only stores the features and not the actual train data(obviously),but if you have the dataset,then you can easily verify duh....<\/p>\n\n<p>I don't think you can ever find the exact train data with only the trained model,cause the model only contains the features and not the actual train data<\/p>\n"}]}
+{"QuestionId":57752765,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2019-09-02T07:14:04.727","AcceptedAnswerId":null,"OwnerUserId":5742328.0,"Title":"For tensorflow, What's the differences between ResourceApplyAdam and ApplyAdam?","Body":"<p>Is there any differences on features or performance between them, or it's just an alias for different version.<\/p>\n","answers":[{"AnswerId":"57752905","CreationDate":"2019-09-02T07:26:51.937","ParentId":null,"OwnerUserId":"9241362","Title":null,"Body":"<p>return types are different. see reference<\/p>\n\n<pre><code>tensorflow::ops::ResourceApplyAdam\nReturns: the created Operation\n\ntensorflow::ops::ApplyAdam\nOutput: Same as \"var\".\n<\/code><\/pre>\n\n<p>reference\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/cc\/class\/tensorflow\/ops\/resource-apply-adam\" rel=\"nofollow noreferrer\">ResourceApplyAdam<\/a> \n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/cc\/class\/tensorflow\/ops\/apply-adam\" rel=\"nofollow noreferrer\">ApplyAdam<\/a><\/p>\n"},{"AnswerId":"57754031","CreationDate":"2019-09-02T08:54:16.637","ParentId":null,"OwnerUserId":"11229126","Title":null,"Body":"<p>Read the Documentation carefully, along with return type, public attributes, public functions are also different\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/cc\/class\/tensorflow\/ops\/resource-apply-adam\" rel=\"nofollow noreferrer\">Resource apply adam<\/a>\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/cc\/class\/tensorflow\/ops\/apply-adam\" rel=\"nofollow noreferrer\">Apply Adam<\/a><\/p>\n"}]}
+{"QuestionId":57752969,"AnswerCount":1,"Tags":"<python><arrays><tensorflow><keras><crop>","CreationDate":"2019-09-02T07:31:52.610","AcceptedAnswerId":null,"OwnerUserId":10515022.0,"Title":"Cropping2D layer crops to wrong dimensions","Body":"<p>I have an 1920-by-1153-by-3 keras tensor that needs to be cropped to 1920-by-1080. However, creating a Cropping layer and passing the difference of the shapes gives it the wrong dimensions. This method gives the tensor the shape of 1920-by-1920.<\/p>\n\n<p>Code:<\/p>\n\n<pre><code>        def crop(x, size):\n            shape = K.int_shape(x)\n            shape_diff = shape[1] - size[1], shape[2] - size[2]\n            print(shape_diff)\n            y = tf.keras.layers.Cropping2D(shape_diff)(x)\n            return y\n<\/code><\/pre>\n\n<p>Partial output of model.summary and shape_diff:<\/p>\n\n<p>Shape_diff:\n<code>(0,72)<\/code><\/p>\n\n<p>model.summary:<\/p>\n\n<pre><code>conv2d_transpose_4 (Conv2DTr (None, 1920, 1152, 3)     195       \n_________________________________________________________________\ncropping2d (Cropping2D)      (None, 1920, 1008, 3)     0         \n=================================================================\n\n<\/code><\/pre>\n","answers":[{"AnswerId":"57756137","CreationDate":"2019-09-02T11:27:15.253","ParentId":null,"OwnerUserId":"10515022","Title":null,"Body":"<p>Solved this by dividing the sides by 2, since the Keras docs state:<\/p>\n\n<blockquote>\n  <p>If tuple of 2 ints: interpreted as <strong>two different symmetric<\/strong> cropping values for height and width.<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":57753145,"AnswerCount":0,"Tags":"<keras><generator><predict>","CreationDate":"2019-09-02T07:44:55.380","AcceptedAnswerId":null,"OwnerUserId":3898714.0,"Title":"keras predict_generator save predictions in batches","Body":"<p>I have a set of files for which i want to make predictions. I have written a generator function which reads the file and creates array for input to predict_generator function. Currently the prediction_generator function returns a single array with predictions for all the files. Is their a way i can save\/access the predictions of individual batches, because i later need to merge them with my original files.<\/p>\n\n<pre><code>def generate_prediction_array(pred_files):\n    for file in pred_files:\n        X = create_array(file)\n        yield X\n...\ny_pred_prob = model.predict_generator(generate_prediction_array(pred_files), \n                                      max_queue_size=2,\n                                      steps=len(pred_array_files))\n<\/code><\/pre>\n\n<p>When i look at y_pred_prob, it has predictions for all the data points in all the files. Instead i want to access prediction for each batch separately, so that i can merge it with my original file data.<\/p>\n","answers":[]}
+{"QuestionId":57753412,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-02T08:06:25.057","AcceptedAnswerId":null,"OwnerUserId":11709008.0,"Title":"Why keras code of mean squared error return K.mean and categorical crossentropy return reduce_sum","Body":"<p>Why<\/p>\n\n<ul>\n<li>keras code of mean squared error use return mean -> 1 \/ N -> is N batchsize?<\/li>\n<li>keras code of categorical crossentropy use return reduce_sum -> 1. I remember categorical crossentropy also need divided by batchsize.<\/li>\n<\/ul>\n\n<p>Please explain this.<\/p>\n\n<p>This is code of mean squared error:<\/p>\n\n<pre><code>def mean_squared_error(y_true, y_pred):\n    return K.mean(K.square(y_pred - y_true), axis=-1)\n<\/code><\/pre>\n\n<p>This is code of categorical cross entropy:<\/p>\n\n<pre><code>def categorical_crossentropy(target, output, from_logits=False, axis=-1):\n\"\"\"Categorical crossentropy between an output tensor and a target tensor.\n# Arguments\n    target: A tensor of the same shape as `output`.\n    output: A tensor resulting from a softmax\n        (unless `from_logits` is True, in which\n        case `output` is expected to be the logits).\n    from_logits: Boolean, whether `output` is the\n        result of a softmax, or is a tensor of logits.\n    axis: Int specifying the channels axis. `axis=-1`\n        corresponds to data format `channels_last`,\n        and `axis=1` corresponds to data format\n        `channels_first`.\n# Returns\n    Output tensor.\n# Raises\n    ValueError: if `axis` is neither -1 nor one of\n        the axes of `output`.\n\"\"\"\noutput_dimensions = list(range(len(output.get_shape())))\nif axis != -1 and axis not in output_dimensions:\n    raise ValueError(\n        '{}{}{}'.format(\n            'Unexpected channels axis {}. '.format(axis),\n            'Expected to be -1 or one of the axes of `output`, ',\n            'which has {} dimensions.'.format(len(output.get_shape()))))\n# Note: tf.nn.softmax_cross_entropy_with_logits\n# expects logits, Keras expects probabilities.\nif not from_logits:\n    # scale preds so that the class probas of each sample sum to 1\n    output \/= tf.reduce_sum(output, axis, True)\n    # manual computation of crossentropy\n    _epsilon = _to_tensor(epsilon(), output.dtype.base_dtype)\n    output = tf.clip_by_value(output, _epsilon, 1. - _epsilon)\n    return - tf.reduce_sum(target * tf.log(output), axis)\nelse:\n    return tf.nn.softmax_cross_entropy_with_logits(labels=target,\n                                                   logits=output)\ndef categorical_crossentropy(y_true, y_pred):\n    return K.categorical_crossentropy(y_true, y_pred)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57753640,"AnswerCount":1,"Tags":"<python><tensorflow><computer-vision><yolo>","CreationDate":"2019-09-02T08:25:33.117","AcceptedAnswerId":null,"OwnerUserId":11455772.0,"Title":"Convert YoloV3 output to coordinates of bounding box, label and confidence","Body":"<p>I run YoloV3 model and get detections - dictionary of 3 entries:<\/p>\n\n<ol>\n<li>\"detector\/yolo-v3\/Conv_22\/BiasAdd\/YoloRegion\" : numpy.ndarray with\nshape (1,255,52,52),<\/li>\n<li>\"detector\/yolo-v3\/Conv_6\/BiasAdd\/YoloRegion\" : numpy.ndarray with\nshape (1,255,13,13),<\/li>\n<li>\"detector\/yolo-v3\/Conv_14\/BiasAdd\/YoloRegion\" : numpy.ndarray with\nshape (1,255,26,26).<\/li>\n<\/ol>\n\n<p>I know that each entry in dictionary is other size of object detection.\nConv_22 is for small objects\nConv_14 is for medium objects\nConv_6 is for big objects<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/DQthF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DQthF.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>How can I convert this dictionary output to coordinates of bounding box, label and confidence?<\/p>\n","answers":[{"AnswerId":"57764063","CreationDate":"2019-09-03T01:36:48.663","ParentId":null,"OwnerUserId":"4578111","Title":null,"Body":"<p>Presuming you use python and opencv, <\/p>\n\n<p>Pelase find the below code with comments where ever required, to extract the output using cv2.dnn module. <\/p>\n\n<pre><code>net.setInput(blob)\n\nlayerOutputs = net.forward(ln)\n\nboxes = []\nconfidences = []\nclassIDs = []\nfor output in layerOutputs:\n# loop over each of the detections\n    for detection in output:\n        # extract the class ID and confidence (i.e., probability) of\n        # the current object detection\n        scores = detection[5:]\n        classID = np.argmax(scores)\n        confidence = scores[classID]\n\n        # filter out weak predictions by ensuring the detected\n        # probability is greater than the minimum probability\n        if confidence &gt; threshold:\n            # scale the bounding box coordinates back relative to the\n            # size of the image, keeping in mind that YOLO actually\n            # returns the center (x, y)-coordinates of the bounding\n            # box followed by the boxes' width and height\n            box = detection[0:4] * np.array([W, H, W, H])\n            (centerX, centerY, width, height) = box.astype(\"int\")\n\n            # use the center (x, y)-coordinates to derive the top and\n            # and left corner of the bounding box\n            x = int(centerX - (width \/ 2))\n            y = int(centerY - (height \/ 2))\n\n            # update our list of bounding box coordinates, confidences,\n            # and class IDs\n            boxes.append([x, y, int(width), int(height)])\n            confidences.append(float(confidence))\n            classIDs.append(classID)\nidxs = cv2.dnn.NMSBoxes(boxes, confidences, confidence, threshold)\n#results are stored in idxs,boxes,confidences,classIDs\n<\/code><\/pre>\n"}]}
+{"QuestionId":57753687,"AnswerCount":1,"Tags":"<python><keras><deep-learning><pytorch>","CreationDate":"2019-09-02T08:29:41.880","AcceptedAnswerId":null,"OwnerUserId":3933052.0,"Title":"Multiple outputs in Pytorch, Keras style","Body":"<p>How could you implement these 2 <code>Keras<\/code> models (inspired by the Datacamp course <a href=\"https:\/\/campus.datacamp.com\/courses\/advanced-deep-learning-with-keras-in-python\/multiple-outputs?ex=1\" rel=\"nofollow noreferrer\">'Advanced Deep Learning with Keras in Python'<\/a>) in <code>Pytorch<\/code>:<\/p>\n\n<p>Classification with 1 input, 2 outputs:<\/p>\n\n<pre><code>from keras.layers import Input, Concatenate, Dense\nfrom keras.models import Model\n\ninput_tensor = Input(shape=(1,))\noutput_tensor = Dense(2)(input_tensor)\n\nmodel = Model(input_tensor, output_tensor)\nmodel.compile(optimizer='adam', loss='categorical_crossentropy')\nX = ... # e.g. a pandas series\ny = ... # e.g. a pandas df with 2 columns\nmodel.fit(X, y, epochs=100)\n<\/code><\/pre>\n\n<p>A model with classification and regression:<\/p>\n\n<pre><code>from keras.layers import Input, Dense\nfrom keras.models import Model\n\ninput_tensor = Input(shape=(1,))\noutput_tensor_reg = Dense(1)(input_tensor)\noutput_tensor_class = Dense(1, activation='sigmoid')(output_tensor_reg)\n\nmodel.compile(loss=['mean_absolute_error','binary_crossentropy']\nX = ...\ny_reg = ...\ny_class = ...\nmodel.fit(X, [y_reg, y_class], epochs=100)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57754054","CreationDate":"2019-09-02T08:56:17.230","ParentId":null,"OwnerUserId":"3933052","Title":null,"Body":"<p><a href=\"https:\/\/discuss.pytorch.org\/t\/a-model-with-multiple-outputs\/10440\/15\" rel=\"nofollow noreferrer\">This ressource<\/a> was particularly helpful.<\/p>\n\n<p>Basically, the idea is that, contrary to Keras, you have to explicitly say where you're going to compute each output in your forward function and how the global loss is gonna be computed from them.<\/p>\n\n<p>For example, regarding the 1st example:<\/p>\n\n<pre><code>def __init__(self, ...):\n    ... # define your model elements\n\ndef forward(self, x):\n    # Do your stuff here\n    ...\n    x1 = F.sigmoid(x) # class probabilities\n    x2 = F.sigmoid(x) # bounding box calculation\n    return x1, x2\n<\/code><\/pre>\n\n<p>Then you compute the losses:<\/p>\n\n<pre><code>out1, out2 = model(data)\nloss1 = criterion1(out1, target1)\nloss2 = criterion2(out2, targt2)\nalpha = ... # define the weights of each sub-loss in the global loss\nloss = alpha * loss1 + (1-alpha) * loss2\nloss.backward()\n<\/code><\/pre>\n\n<p>For the 2nd one, it's almost the same but you compute the loss at different point in the forward pass:<\/p>\n\n<pre><code>def __init__(self, ...):\n    ... # define your model elements\n\ndef forward(self, main_input, aux_input):\n    aux = F.relu(self.dense_1(main_input))\n    x = F.relu(self.input_layer(aux))\n    x = F.sigmoid(x)\n    return x, aux\n<\/code><\/pre>\n"}]}
+{"QuestionId":57754012,"AnswerCount":1,"Tags":"<tensorflow><tensorflow2.0>","CreationDate":"2019-09-02T08:53:06.157","AcceptedAnswerId":null,"OwnerUserId":7779411.0,"Title":"How to make multiple fetches simultaneously on tf2.0?","Body":"<p>In tf1, I could define a list of the required fetches and use<\/p>\n\n<pre><code>sess.run(myList, feed_dict)\n<\/code><\/pre>\n\n<p>to get all the elements of the list computed simultaneously by tf1 through the graph. How to do it in tf2.0?<\/p>\n\n<p>Sample code in tf1:<\/p>\n\n<pre><code>import tensorflow as tf\na = [None]*5\nfor i in range(5):\n    a[i] = tf.Variable(tf.random.normal([3,3]))\nfetch_list = [None]*5\nfor i in range(5):\n    fetch_list[i] = tf.add(tf.gather(a, i), tf.ones([3,3]))\n\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\nsess.run(fetch_list)\n<\/code><\/pre>\n\n<p>Didn't check whether the above code runs or not but I hope you get the point of it. Thanks<\/p>\n","answers":[{"AnswerId":"57756263","CreationDate":"2019-09-02T11:36:36.783","ParentId":null,"OwnerUserId":"2455494","Title":null,"Body":"<p>because tf 2.x executes eagerly by default you <em>could<\/em> just do:<\/p>\n\n<pre><code>a = [None]*5\nfor i in range(5):\n    a[i] = tf.Variable(tf.random.normal([3,3]))\nfetch_list = [None]*5\nfor i in range(5):\n    fetch_list[i] = tf.add(tf.gather(a, i), tf.ones([3,3]))\n<\/code><\/pre>\n\n<p>then <code>fetch_list<\/code> is populated as before.<\/p>\n\n<p>Depending on the complexity of your real word example, you may also consider the use of the <code>@tf.function<\/code> which builds up an execution graph before pushing the data through to aid optimisation analogous to tf1 (this is a massive oversimplification but you get the point). <\/p>\n\n<p>You might consider simplifying\/reworking the code a little to make facilitate this. Probably best to just work with tensors where possible rather than a list of tensors. It's hard to advise exactly how to accomplish this as I don't know what you have simplified for your example. <\/p>\n\n<p>For example, if we consider your <code>fetch_list<\/code> to be a <code>(5,3,3)<\/code> tensor rather than a list of 5 <code>(3,3)<\/code> tensors then I'm sure you realise that your simplified example code (more or less) boils down to something like this:<\/p>\n\n<pre><code>@tf.function\ndef get_list(n):\n  return tf.random.normal((n,3,3))\n\nfetch_list = get_list(5)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57754497,"AnswerCount":1,"Tags":"<python><tensorflow><keras><neural-network>","CreationDate":"2019-09-02T09:27:41.230","AcceptedAnswerId":57754783.0,"OwnerUserId":10780811.0,"Title":"I can't load my trained h5 model with load.models(), how do I fix this error?","Body":"<p>So I think tensorflow.keras and the independant keras packages are in conflict and I can't load my model, which I have made with transfer learning.\nImport in the CNN ipynb:<\/p>\n\n<pre><code>!pip install tensorflow-gpu==2.0.0b1\n\nimport tensorflow as tf\nfrom tensorflow import keras\nprint(tf.__version__) \n<\/code><\/pre>\n\n<p>Loading this pretrained model<\/p>\n\n<pre><code>base_model = keras.applications.xception.Xception(weights=\"imagenet\",\n                                              include_top=False)\navg = keras.layers.GlobalAveragePooling2D()(base_model.output)\noutput = keras.layers.Dense(n_classes, activation=\"softmax\")(avg)\nmodel = keras.models.Model(inputs=base_model.input, outputs=output)\n<\/code><\/pre>\n\n<p>Saving with:<\/p>\n\n<pre><code>model.save('Leavesnet Model 2.h5')\n<\/code><\/pre>\n\n<p>Then in the new ipynb for the already trained model (the imports are the same as in the CNN ipynb:<\/p>\n\n<pre><code>from keras.models import load_model\n\nmodel =load_model('Leavesnet Model.h5')\n<\/code><\/pre>\n\n<p>I get the error: <\/p>\n\n<pre><code>AttributeError                            Traceback (most recent call last)\n&lt;ipython-input-4-77ca5a1f5f24&gt; in &lt;module&gt;()\n      2 from keras.models import load_model\n      3 \n----&gt; 4 model =load_model('Leavesnet Model.h5')\n\n13 frames\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/backend\/tensorflow_backend.py in placeholder(shape, ndim, dtype, sparse, name)\n    539         x = tf.sparse_placeholder(dtype, shape=shape, name=name)\n    540     else:\n--&gt; 541         x = tf.placeholder(dtype, shape=shape, name=name)\n    542     x._keras_shape = shape\n    543     x._uses_learning_phase = False\n\nAttributeError: module 'tensorflow' has no attribute 'placeholder' \n\n<\/code><\/pre>\n\n<p>I think there might be a conflict between tf.keras and the independant keras, can someone help me out?<\/p>\n","answers":[{"AnswerId":"57754783","CreationDate":"2019-09-02T09:48:58.663","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Yes, there is a conflict between <code>tf.keras<\/code> and <code>keras<\/code> packages, you trained the model using <code>tf.keras<\/code> but then you are loading it with the <code>keras<\/code> package. That is not supported, you should use only one version of this package.<\/p>\n\n<p>The specific problem is that you are using TensorFlow 2.0, but the standalone <code>keras<\/code> package does not support TensorFlow 2.0 yet.<\/p>\n"}]}
+{"QuestionId":57754926,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-02T09:58:34.790","AcceptedAnswerId":null,"OwnerUserId":10856137.0,"Title":"ERROR: Config value download_clang is not defined in any .rc file","Body":"<p>I tried building tensorflow from source and while doing so I encountered this error please help me.<\/p>\n\n<p>I have already tried using -c opt instead of --config=opt but it didnt work<\/p>\n\n<p>(base) rajkirito@rajkirito-Vostro-15-3568:~\/tensorflow$ bazel build --config=opt \/\/tensorflow\/tools\/pip_package:build_pip_package<\/p>\n\n<p>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\/home\/rajkirito\/tensorflow\/tools\/bazel.rc\nINFO: Options provided by the client:\n  Inherited 'common' options: --isatty=1 --terminal_columns=80\nINFO: Reading rc options for 'build' from \/home\/rajkirito\/tensorflow\/.tf_configure.bazelrc:\n  'build' options: --action_env PYTHON_BIN_PATH=\/home\/rajkirito\/anaconda3\/bin\/python --action_env PYTHON_LIB_PATH=\/home\/rajkirito\/anaconda3\/lib\/python3.7\/site-packages --python_path=\/home\/rajkirito\/anaconda3\/bin\/python --define with_jemalloc=true --define with_gcp_support=true --define with_hdfs_support=true --define with_aws_support=true --define with_kafka_support=true --action_env TF_NEED_OPENCL_SYCL=0 --action_env TF_NEED_CUDA=0 --action_env TF_DOWNLOAD_CLANG=1 --config=download_clang --define grpc_no_ares=true --strip=always\nERROR: Config value download_clang is not defined in any .rc file<\/p>\n","answers":[]}
+{"QuestionId":57755922,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><neural-network><autoencoder>","CreationDate":"2019-09-02T11:09:50.410","AcceptedAnswerId":null,"OwnerUserId":4265321.0,"Title":"How does training of Autoencoder work in Tensorflow (MSE calculation on input dataset)","Body":"<p>I am training an Autoencoder in Tensorflow:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\ninput = # some input\noutput = input\n\n\ntf.reset_default_graph()\n\nnum_inputs = np.shape(input)[1]\nnum_hid1=250\nnum_hid2=100\nnum_hid3=50\nnum_hid4=5\nnum_hid5=num_hid3\nnum_hid6=num_hid2\nnum_hid7=num_hid1\nnum_outputs=num_inputs\nlr=0.01\nactf=tf.nn.tanh\n\nX=tf.placeholder(tf.float32,shape=[None,num_inputs])\ninitializer=tf.variance_scaling_initializer()\n\nw1=tf.Variable(initializer([num_inputs,num_hid1]),dtype=tf.float32)\nw2=tf.Variable(initializer([num_hid1,num_hid2]),dtype=tf.float32)\nw3=tf.Variable(initializer([num_hid2,num_hid3]),dtype=tf.float32)\nw4=tf.Variable(initializer([num_hid3,num_hid4]),dtype=tf.float32)\nw5=tf.Variable(initializer([num_hid4,num_hid5]),dtype=tf.float32)\nw6=tf.Variable(initializer([num_hid5,num_hid6]),dtype=tf.float32)\nw7=tf.Variable(initializer([num_hid6,num_hid7]),dtype=tf.float32)\nw8=tf.Variable(initializer([num_hid7,num_outputs]),dtype=tf.float32)\n\nb1=tf.Variable(tf.zeros(num_hid1))\nb2=tf.Variable(tf.zeros(num_hid2))\nb3=tf.Variable(tf.zeros(num_hid3))\nb4=tf.Variable(tf.zeros(num_hid4))\nb5=tf.Variable(tf.zeros(num_hid5))\nb6=tf.Variable(tf.zeros(num_hid6))\nb7=tf.Variable(tf.zeros(num_hid7))\nb8=tf.Variable(tf.zeros(num_outputs))\n\nhid_layer1=actf(tf.matmul(X,w1)+b1)\nhid_layer2=actf(tf.matmul(hid_layer1,w2)+b2)\nhid_layer3=actf(tf.matmul(hid_layer2,w3)+b3)\nhid_layer4=actf(tf.matmul(hid_layer3,w4)+b4)\nhid_layer5=actf(tf.matmul(hid_layer4,w5)+b5)\nhid_layer6=tf.matmul(hid_layer5,w6)+b6\nhid_layer7=tf.matmul(hid_layer6,w7)+b7\noutput_layer=tf.matmul(hid_layer7,w8)+b8\n\nloss=tf.reduce_mean(tf.square(output_layer-X))\n\noptimizer=tf.train.AdamOptimizer(lr)\ntrain=optimizer.minimize(loss)\n\ninit=tf.global_variables_initializer()\n\nnum_epoch=1000\n\n\n# save trained Autoencoder\nsaver = tf.train.Saver()\nwith tf.Session() as sess:\n    sess.run(init)\n    for epoch in range(num_epoch):\n\n        sess.run(train,feed_dict={X:input})\n\n        train_loss=loss.eval(feed_dict={X:input})\n        print(\"epoch {} loss {}\".format(epoch,train_loss))\n\n\n    saver.save(sess, 'Autoencoder')\n<\/code><\/pre>\n\n<p>I am training the Autoencoder using the entire input dataset (<code>input<\/code>) all at once (i.e. I am not using batches). I would like to understand if the calculation of the Mean Squared Error is operated on ALL the data in the input dataset or if it is done in some other way.<\/p>\n\n<p>I am asking this question because I noticed that, once trained, my Autoencoder performs very well in compressing and decompressing the first data in the dataset, but not the ones at the end of the dataset. Secondly, the performance of the Autoencoder on the testing dataset is horrible.<\/p>\n\n<p>If relevant, my actual data consists of 2000 time series. Each time series has 501 time components (so in my case <code>num_inputs = 501<\/code>, and the Mean Squared Error should be calculated on <code>2000 * 501 = 1002000<\/code> components).<\/p>\n","answers":[]}
+{"QuestionId":57756173,"AnswerCount":2,"Tags":"<python><machine-learning><deep-learning><pytorch><imbalanced-data>","CreationDate":"2019-09-02T11:30:15.513","AcceptedAnswerId":57759365.0,"OwnerUserId":7003420.0,"Title":"Using imbalanced-learn with Pandas DataFrame","Body":"<p>My dataset is quite imbalanced. The two minority classes each contain half of the sample in the majority class. My RNN model is not able to learn anything about the least populated class.<\/p>\n\n<p>I'm trying to use the <code>imbalanced-learn<\/code> library. For instance:<\/p>\n\n<pre><code>sm = SMOTE(random_state=42, n_jobs=-1, k_neighbors=10)\nX_train, y_train = sm.fit_resample(train.drop(['label], axis=1), train['label'])\n<\/code><\/pre>\n\n<p>works if <code>train.drop(['label]<\/code> contains just the values of the used features. The problem is that my DataFrame contains one additional columns containing strings as values: I cannot drop it since those strings are the input for my RNN. And if I drop it, I would not be able to tell to which row of the oversampled dataset those strings belong to.<\/p>\n\n<p>Is there a way to keep all the columns and tell the function which columns to use for oversampling?<\/p>\n","answers":[{"AnswerId":"57759365","CreationDate":"2019-09-02T15:17:47.370","ParentId":null,"OwnerUserId":"7003420","Title":null,"Body":"<p>For those who need to do something similar, a co-author of the library suggested me to use <a href=\"https:\/\/imbalanced-learn.readthedocs.io\/en\/stable\/generated\/imblearn.over_sampling.SMOTENC.html#imblearn-over-sampling-smotenc\" rel=\"nofollow noreferrer\">SMOTENC<\/a>, which can handle also categorical variables (like strings).<\/p>\n"},{"AnswerId":"57756840","CreationDate":"2019-09-02T12:19:10.210","ParentId":null,"OwnerUserId":"4689878","Title":null,"Body":"<p>If the string column is an input to your RNN then, assuming that you plan to encode it somehow (one-hot encoding for example), then just encode that column before oversampling and then run the oversampling with your new encoded columns instead of the string column.<\/p>\n"}]}
+{"QuestionId":57756314,"AnswerCount":1,"Tags":"<python><tensorflow><optimization><gradient><dtype>","CreationDate":"2019-09-02T11:40:29.660","AcceptedAnswerId":57758301.0,"OwnerUserId":12009557.0,"Title":"optimizers minimize error: 'float' object has no attribute 'dtype'","Body":"<p>I am a beginner of tensorflow. There are some problems in gradient calculation with tensorflow 2.0. Can someone help me?<\/p>\n\n<p>Here's my code. The error prompt is:<\/p>\n\n<pre><code>if not t.dtype.is_floating:\nAttributeError: 'float' object has no attribute 'dtype'\n<\/code><\/pre>\n\n<p>I've tried:<\/p>\n\n<pre><code>w = tf.Variable([1.0,1.0],dtype = tf.float32)\n<\/code><\/pre>\n\n<p>the message changed into:<\/p>\n\n<pre><code>TypeError: 'tensorflow.python.framework.ops.EagerTensor' object is not callable\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\ntrain_X = np.linspace(-1, 1, 100)\ntrain_Y = 2 * train_X + np.random.randn(*train_X.shape) * 0.33 + 10\n\n# w = tf.Variable([1.0,1.0],dtype = tf.float32)\nw = [1.0,1.0]https:\/\/www.cybertec-postgresql.com\/en\/?p=9102&amp;preview=true\nopt=tf.keras.optimizers.SGD(0.1)\nmse=tf.keras.losses.MeanSquaredError()\nfor i in range(20):\n    print(\"epoch:\",i,\"w:\", w)\n    with tf.GradientTape() as tape:\n        logit = w[0] * train_X + w[1]\n        loss= mse(train_Y,logit)\n    w = opt.minimize(loss, var_list=w)\n<\/code><\/pre>\n\n<p>I dont know how to fix it.Thank you for any comments.<\/p>\n","answers":[{"AnswerId":"57758301","CreationDate":"2019-09-02T14:02:54.950","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>You have not used <code>GradientTape<\/code> correctly. I have demoed the code how you should be applying it.\nI have created a single unit dense layer model which mimics your <code>w<\/code> variable.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\ntrain_X = np.linspace(-1, 1, 100)\ntrain_X = np.expand_dims(train_X, axis=-1)\nprint(train_X.shape)    # (100, 1)\ntrain_Y = 2 * train_X + np.random.randn(*train_X.shape) * 0.33 + 10\nprint(train_Y.shape)    # (100, 1)\n\n# First create a  model with one unit of dense and one bias\ninput = tf.keras.layers.Input(shape=(1,))\nw = tf.keras.layers.Dense(1)(input)   # use_bias is True by default\nmodel = tf.keras.Model(inputs=input, outputs=w)\n\nopt=tf.keras.optimizers.SGD(0.1)\nmse=tf.keras.losses.MeanSquaredError()\n\nfor i in range(20):\n    print('Epoch: ', i)\n    with tf.GradientTape() as grad_tape:\n        logits = model(train_X, training=True)\n        model_loss = mse(train_Y, logits)\n        print('Loss =', model_loss.numpy())\n\n    gradients = grad_tape.gradient(model_loss, model.trainable_variables)\n    opt.apply_gradients(zip(gradients, model.trainable_variables))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57756345,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-09-02T11:42:14.080","AcceptedAnswerId":null,"OwnerUserId":7215349.0,"Title":"Bidirectional LSTM vs Multipleparallel accuracy","Body":"<p>I'm trying to predict money values with 2 different models. Mulitpleparallel model (model B) gives me over 95% acccuracy. At the same time bidirectional LSTM model (model A) gives me 0%. But, as you can see on the graph (<a href=\"https:\/\/drive.google.com\/file\/d\/151UUTdT8l1xQHf7QrbKjblaQOHYqpkbI\/view?usp=sharing\" rel=\"nofollow noreferrer\">https:\/\/drive.google.com\/file\/d\/151UUTdT8l1xQHf7QrbKjblaQOHYqpkbI\/view?usp=sharing<\/a>) both models are pretty accurate if we follow the real values. If you don't see on the graph, believe me that values are almost the same in predicting the real value although the multipleparallel LSTM use four values as an input.<\/p>\n\n<p>Model B (multipleparallel) looks like this:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(100, activation='relu', return_sequences=True, input_shape=(n_steps, n_features)))\nmodel.add(LSTM(100, activation='relu'))\nmodel.add(Dense(n_features))\n<\/code><\/pre>\n\n<p>and the model A (bidirectional) looks like this:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Bidirectional(LSTM(100, activation='relu'), input_shape=(n_steps, n_features)))\nmodel.add(Dense(1))\n<\/code><\/pre>\n\n<p>Code after is the same for both models:<\/p>\n\n<pre><code>cp = [ModelCheckpoint(filepath=path+\"\/epochBi.h5\", monitor='acc', verbose=1, save_best_only=True)]\nmodel.compile(optimizer='adam', loss='mse', metrics=['accuracy'])  \nhistory_callback = model.fit(X, y, epochs=200, verbose=1, callbacks=cp) \n\nmodel.load_weights(path+\"\/epoch.h5\")\nscore = model.evaluate(X, y, verbose=1)\n<\/code><\/pre>\n\n<p>Accuracy of the Bidirectional LSTM model: <a href=\"https:\/\/drive.google.com\/file\/d\/1YgX-8RwEqDucUxOv68a9EdTCQZat8p0w\/view?usp=sharing\" rel=\"nofollow noreferrer\">https:\/\/drive.google.com\/file\/d\/1YgX-8RwEqDucUxOv68a9EdTCQZat8p0w\/view?usp=sharing<\/a><\/p>\n\n<p>Accuracy of the MultipleParallel model: <a href=\"https:\/\/drive.google.com\/file\/d\/13ZdLjET0ofwrER4X6KkOqcadcZ5rmNmo\/view?usp=sharing\" rel=\"nofollow noreferrer\">https:\/\/drive.google.com\/file\/d\/13ZdLjET0ofwrER4X6KkOqcadcZ5rmNmo\/view?usp=sharing<\/a><\/p>\n","answers":[]}
+{"QuestionId":57756417,"AnswerCount":1,"Tags":"<tensorflow><keras><tf.keras>","CreationDate":"2019-09-02T11:46:42.217","AcceptedAnswerId":null,"OwnerUserId":2366125.0,"Title":"InvalidArgumentError: -1 is not between 0 and 3","Body":"<p>I have compiled the following model.<\/p>\n\n<pre><code>_________\nLayer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput1 (InputLayer)             (None, 2, 128)       0                                            \n__________________________________________________________________________________________________\ninput2 (InputLayer)             (None, 2, 128)       0                                            \n__________________________________________________________________________________________________\nselfattnlayer1 (SelfAttn)       (None, 2, 128)       49152       input1[0][0]                     \n__________________________________________________________________________________________________\nselfattnlayer2 (SelfAttn)       (None, 2, 128)       49152       input2[0][0]                     \n__________________________________________________________________________________________________\nlambda_24 (Lambda)              (None, 2, 768)       0           input1[0][0]                     \n                                                                 input2[0][0]                     \n                                                                 selfattnlayer1[0][0]             \n                                                                 selfattnlayer2[0][0]             \n__________________________________________________________________________________________________\nencoder_lstm (GRU)              (None, 2, 256)       787200      lambda_24[0][0]                  \n__________________________________________________________________________________________________\ntime_distributed_18 (TimeDistri (None, 2, 3)         771         encoder_lstm[0][0]               \n==================================================================================================\nTotal params: 886,275\nTrainable params: 886,275\nNon-trainable params: 0\n__________________________________________________________________________________________________\nNone\nTrain on 1350 samples, validate on 150 samples\nEpoch 1\/10\n<\/code><\/pre>\n\n<p>However, I'm getting the error while fitting the data.<\/p>\n\n<pre><code>   ---------------------------------------------------------------------------\nInvalidArgumentError                      Traceback (most recent call last)\n&lt;ipython-input-90-04168b02e9ca&gt; in &lt;module&gt;\n      5 y=np.array([[1,0,1] for _ in range(1500)])\n      6 \n----&gt; 7 decomp.fit(X,y)\n\n&lt;ipython-input-86-74a9247c4023&gt; in fit(self, X, y, entails)\n     76 \n     77 \n---&gt; 78         self.model.fit(x=[xg,xe], y=np.expand_dims(y, axis=1), validation_split = 0.1,batch_size=self.batch_size , epochs= self.epochs)\n\n~\/ckm\/py3\/lib\/python3.5\/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~\/ckm\/py3\/lib\/python3.5\/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    213           ins_batch[i] = ins_batch[i].toarray()\n    214 \n--&gt; 215         outs = f(ins_batch)\n    216         if not isinstance(outs, list):\n    217           outs = [outs]\n\n~\/ckm\/py3\/lib\/python3.5\/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~\/ckm\/py3\/lib\/python3.5\/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~\/ckm\/py3\/lib\/python3.5\/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: -1 is not between 0 and 3\n     [[{{node training_11\/RMSprop\/gradients\/selfattnlayer2_6\/transpose_2_grad\/InvertPermutation}} = InvertPermutation[T=DT_INT32, _class=[\"loc:@training_11\/RMSprop\/gradients\/selfattnlayer2_6\/transpose_2_grad\/transpose\"], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](selfattnlayer2_6\/transpose_2\/perm)]]\n<\/code><\/pre>\n\n<p>Data I'm feeding is <\/p>\n\n<pre><code>decomp = DecomposableAttention()\nd1 = np.random.randint(0,10, size=[1500,2,128])\nd2 = np.random.randint(0,4, size=[1500,2,128])\nX=np.array([d1,d2])\ny=np.array([[1,0,1] for _ in range(1500)])\n\ndecomp.fit(X,y)\n<\/code><\/pre>\n\n<p>Can anybody point me where to look for the error? If need be, I'll share the code (right now, the code is split in 3 files, that's why I've not attached the files). The problem I'm trying to solve is multilabel classification problem using some sort of self-attention.<\/p>\n\n<p><strong>Edit:<\/strong>\nI found the error. <\/p>\n\n<pre><code>K.permute_dimensions(K.dot(x, self.WK),(0,-1,1))\n<\/code><\/pre>\n\n<p>is offending line. We can not use <code>-1<\/code> while specifying dims. But use only positive integers like so:<\/p>\n\n<pre><code>K.permute_dimensions(K.dot(x, self.WK),(0,2,1))\n<\/code><\/pre>\n","answers":[{"AnswerId":"57758655","CreationDate":"2019-09-02T14:27:35.127","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p><code>permute_dimensions<\/code> basically refer to the ordering of your axis. So you have to specify all the axis in which order you wish to get it generated. <\/p>\n\n<p>You may be getting confused with operations like <code>reshape<\/code> where you can specify one axis as -1. In those operations, it refers to that you wish the TensorFlow to implicitly evaluate the dimension for you.<\/p>\n"}]}
+{"QuestionId":57756451,"AnswerCount":1,"Tags":"<python><machine-learning><keras><imbalanced-data>","CreationDate":"2019-09-02T11:49:01.943","AcceptedAnswerId":57756794.0,"OwnerUserId":7827820.0,"Title":"Why we use the loss to update our model but use the metrics to choose the model we need?","Body":"<blockquote>\n  <p>First of all,I am confused about why we use the loss to update the\n  model but use the metrics to choose the model we need.<\/p>\n  \n  <p>Maybe not all of code, but most of the code I've seen does,they use\n  EarlyStopping to monitor the metrics on the validation data to find\n  the best epoch(loss and metrics are different).<\/p>\n  \n  <p>Since you have chosen to use the loss to update the model, why not use\n  the loss to select the model? After all, the loss and the metrics are\n  not exactly the same. It gives me the impression that you do something\n  with this purpose, and then you evaluate it with another indicator,\n  which makes me feel very strange.Take the regression problem as an\n  example,when someone use the 'mse' as their loss, why they define\n  <code>metrics=['mae']<\/code> and monitor this to early stop or reduce learning\n  rate,I just can't understand and I want to know what is the advantages\n  of doing this?<\/p>\n  \n  <p>Secondly, when your training data is imbalance data and the problem is\n  a classfication problem, Some of the tutorial will tell you to use the\n  F1 or AUC as your metrics,and they say it will improve the problem\n  caused by imbalance data.I don't know why these metrics can improve\n  the problem caused by imbalance data.<\/p>\n  \n  <p>Thirdly,I am confused about when someone send more than one metric to\n  the parameter <code>metrics<\/code> in the function <code>compile<\/code>. I don't understand\n  why multiple, why not one. What is the advantage of defining multiple\n  metrics over one?<\/p>\n  \n  <p>I seem to have too many questions,and they have been bothering me for\n  a long time.<\/p>\n  \n  <p>Thank you for your kind answer.<\/p>\n<\/blockquote>\n\n<hr>\n\n<p>The content above is what I edited before. Some people think my questions are too broad, so I want to reorganize my language.<\/p>\n\n<blockquote>\n  <p>Now suppose that there is a binary classification problem, and the data is not balanced. The ratio of positive and negative classes is\n  <em>500:1<\/em>.<\/p>\n<\/blockquote>\n\n<p>I chose <code>DNN<\/code> as my classification model. I chose <code>cross entropy<\/code> as my <code>loss<\/code>. \nNow the question is whether I should choose <code>cross entropy<\/code> as my <code>metric<\/code>, or should I choose something else, why?<\/p>\n\n<p>I want to talk about the information I get from other people's answers, that is, when the problem is a regression problem, the general metric and loss are differentiable, so in fact, choosing the same metrice and loss, or different one, depends entirely on your own understanding of the problem. But if the problem is classification, the metric we want is not differentiable, so we will choose different loss and metric, such as <code>F1<\/code> and <code>AUC<\/code>, which are not differentiable. Why don't we choose <code>cross entropy<\/code> directly as the measure?<\/p>\n","answers":[{"AnswerId":"57756794","CreationDate":"2019-09-02T12:15:59.497","ParentId":null,"OwnerUserId":"4685471","Title":null,"Body":"<p>Question is arguably too broad for SO; nevertheless, here is a couple of things which you will hopefully find helpful...<\/p>\n\n<blockquote>\n  <p>Since you have chosen to use the loss to update the model, why not use the loss to select the model?<\/p>\n<\/blockquote>\n\n<p>Because, while the loss is the quantity we have to optimize from the <em>mathematical<\/em> perspective, the quantity of interest from the <em>business<\/em> perspective is the metric; in other words, at the end of the day, as <em>users<\/em> of the model, we are interested in the metric, and not in the loss (at least for settings where these two quantities are by default different, such as in classification problems).<\/p>\n\n<p>That said, selecting the model based on <em>loss<\/em> is also a perfectly valid strategy, too; as always, there is some <a href=\"https:\/\/datascience.stackexchange.com\/questions\/37186\/early-stopping-on-validation-loss-or-on-accuracy\">subjectivity<\/a>, and it depends on the specific problem.<\/p>\n\n<blockquote>\n  <p>Take the regression problem as an example, when someone use the 'mse' as their loss, why they define <code>metrics=['mae']<\/code><\/p>\n<\/blockquote>\n\n<p>This is not the norm, and is far from standard; normally, for <em>regression<\/em> problems, it is perfectly natural to use the loss as the metric, too. I agree with you that choices like the one you refer to seem unnatural, and in general do not seem to make much sense. Just keep in mind that because someone used it in a blog or something does not make it necessarily \"correct\" (or a good idea), but is is difficult to argue in general without taking into account possible arguments for the specific case.<\/p>\n\n<blockquote>\n  <p>I don't know why these metrics [F1 or AUC] can improve the problem caused by imbalance data.<\/p>\n<\/blockquote>\n\n<p>They don't \"improve\" anything - they are just more <em>appropriate<\/em> instead of accuracy, where a naive approach in a heavily imbalanced dataset (think of 99% majority class) will be simply to classify everything as the majority class, which would give a 99% accuracy without the model having learned anything.<\/p>\n\n<blockquote>\n  <p>I am confused about when someone send more than one metric to the parameter metrics in the function <code>compile<\/code>. I don't understand why multiple, why not one. What is the advantage of defining multiple metrics over one?<\/p>\n<\/blockquote>\n\n<p>Again, generally speaking, there is no advantage, neither this is the norm; but everything depends on possible specifics.<\/p>\n\n<hr>\n\n<p><strong>UPDATE<\/strong> (after comment): Limiting the discussion to classification settings (since in regression, the loss and the metric can be the same thing), similar questions pop up rather frequently, I guess because the subtle differences between the loss and the various available metrics (accuracy, precision, recall, F1 score etc) are not well understood; consider for example the inverse of your question:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/56498504\/optimizing-for-accuracy-instead-of-loss-in-keras-model?noredirect=1&amp;lq=1\">Optimizing for accuracy instead of loss in Keras model<\/a><\/p>\n\n<p>and the links therein. Quoting from one of my own <a href=\"https:\/\/stackoverflow.com\/questions\/47508874\/how-does-keras-evaluate-the-accuracy\/47515095#47515095\">linked answers<\/a>:<\/p>\n\n<blockquote>\n  <p>Loss and accuracy are different things; roughly speaking, the accuracy is what we are actually interested in from a <em>business<\/em> perspective, while the loss is the objective function that the learning algorithms (optimizers) are trying to minimize from a <em>mathematical<\/em> perspective. Even more roughly speaking, you can think of the loss as the \"translation\" of the business objective (accuracy) to the mathematical domain, a translation which is necessary in classification problems (in regression ones, usually the loss and the business objective are the same, or at least can be the same in principle, e.g. the RMSE)...<\/p>\n<\/blockquote>\n\n<p>You may also find the discussion in <a href=\"https:\/\/stackoverflow.com\/questions\/47891197\/cost-function-training-target-versus-accuracy-desired-goal\">Cost function training target versus accuracy desired goal<\/a> helpful.<\/p>\n"}]}
+{"QuestionId":57756635,"AnswerCount":0,"Tags":"<python><tensorflow><concatenation><mixture><multivariate-partition>","CreationDate":"2019-09-02T12:02:54.107","AcceptedAnswerId":null,"OwnerUserId":9481053.0,"Title":"Incompatible shape issue with a mixture density network in tensorflow\/keras","Body":"<p>I'm stuck on a problem while computing a mixture density network through keras (using tensorflow backends). The goal of this MDN is to learn a latent representation of an image (in order to implement the predictions of the MDN in an autoencoder). \nThen, I would like to model my input image as a multivariate normal distribution and get as network output a mu and sigma vector (of dimension 64 each), and a set of N weights alpha (where N is the number of components in the mixture).\nIf I consider an output having a shape of 64 for each parameter everything works well but it makes no sense to get more alpha factors than components (that is a 64 dimension alpha in my case).\nWhen I try to specify to alpha to get a different shape than mu and sigma, some issues appear.<\/p>\n\n<p>In order to get a covariance matrix to implement in the MixtureSameFamily module of tensorflow, I consider a diagonal matrix of the sigma vector. Then, I've found this loss function (a negative log-likelihood) on many forums and try to adapt it to my problem:<\/p>\n\n<pre><code>def slice_parameter_vectors(parameter_vector):\n    return tf.split(parameter_vector,[1*components,64*components,64*components],axis=1)\n\ndef gnll_loss(y, parameter_vector):\n    alpha, mu, sigma = slice_parameter_vectors(parameter_vector)  # Unpack parameter vectors\n\ngm = tfd.MixtureSameFamily(\n    mixture_distribution=tfd.Categorical(probs=alpha),\n    components_distribution=tfd.MultivariateNormalDiag(\n        loc=mu,      \n        scale_diag=  sigma))\n\n    log_likelihood = gm.log_prob(tf.transpose(y))                 # Evaluate log-probability of y\n\n    return -tf.reduce_mean(log_likelihood, axis=-1)\n<\/code><\/pre>\n\n<p>If I try to feed my network with some data and compile it, I alway have this error: <\/p>\n\n<pre><code>InvalidArgumentError: Incompatible shapes: [64,1,2] vs. [2,64]\n     [[{{node loss_20\/concatenate_6_loss\/MultivariateNormalDiag\/log_prob\/affine_linear_operator\/inverse\/sub}} = Sub[T=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"](loss_20\/concatenate_6_loss\/MixtureSameFamily\/log_prob\/pad_sample_dims\/Reshape, loss_20\/concatenate_6_loss\/split:1)]]\n     [[{{node loss_20\/mul\/_4221}} = _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_1846_loss_20\/mul\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"]()]]\n<\/code><\/pre>\n\n<p>64 is the dimension of the vector, and 2 is the batch-size. Using a batch-size of 1 works but giving me some NaN as a loss.<\/p>\n\n<p>Here is how alpha, mu and sigma layers are built:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    fc    = Dense((no_parameters-1) * components*64 + components, activation=\"tanh\", name=\"fc\")(layer)\n    alphas = Dense(1*components, activation=\"softmax\", name=\"alphas\")(fc)\n    mus    = Dense(64*components, name=\"mus\")(fc)\n    sigmas = Dense(64*components, activation=nnelu, name=\"sigmas\")(fc)\n    pvec   = Concatenate(axis=1)([alphas,mus,sigmas])\n    mdn    = Model(inputs=inputs,outputs=pvec)\n<\/code><\/pre>\n\n<p>Thus, the question is: is it possible to do that through tensorflow? Is there anybody else using this kind of networks here and could explain to me how to deal with the loss function? <\/p>\n\n<p>Kind regards,<\/p>\n\n<p>Adrien<\/p>\n","answers":[]}
+{"QuestionId":57757106,"AnswerCount":0,"Tags":"<c++><tensorflow><c++-winrt><onnx><windows-machine-learning>","CreationDate":"2019-09-02T12:40:00.437","AcceptedAnswerId":null,"OwnerUserId":11819032.0,"Title":"Binded model is not evaluating the image. What can be the possible reason behind it?","Body":"<p>I am following the cpp <a href=\"https:\/\/github.com\/microsoft\/Windows-Machine-Learning\/blob\/master\/Samples\/SqueezeNetObjectDetection\/Desktop\/cpp\/main.cpp\" rel=\"nofollow noreferrer\">tutorial<\/a>. I have replaced the model with my own and replaces the input and output tensors, <a href=\"https:\/\/github.com\/microsoft\/Windows-Machine-Learning\/blob\/5ddf3a215b21c48e2f8641275655b046402dc13d\/Samples\/SqueezeNetObjectDetection\/Desktop\/cpp\/main.cpp#L69\" rel=\"nofollow noreferrer\">here<\/a> and <a href=\"https:\/\/github.com\/microsoft\/Windows-Machine-Learning\/blob\/5ddf3a215b21c48e2f8641275655b046402dc13d\/Samples\/SqueezeNetObjectDetection\/Desktop\/cpp\/main.cpp#L72\" rel=\"nofollow noreferrer\">here<\/a>.\nThe Model is binding but on evaluation step abruptly aborts.<\/p>\n\n<p>As far as I know, there is nothing wrong with my code and model. I have tried to change the input image into various sizes and color modes(grayscale and rgb) but nothing works.<\/p>\n\n<p>Changes made:<\/p>\n\n<pre><code>int main(int argc, char* argv[])\n{\n    init_apartment();\n\n    \/\/ did they pass in the args \n    if (ParseArgs(argc, argv) == false)\n    {\n        printf(\"Usage: %s [imagefile] [cpu|directx]\", argv[0]);\n        return -1;\n    }\n\n    \/\/ Get model path\n    auto modelPath = GetModelPath();\n\n    \/\/ load the model\n    printf(\"Loading modelfile '%ws' on the '%s' device\\n\", modelPath.c_str(), deviceName.c_str());\n    DWORD ticks = GetTickCount();\n    auto model = LearningModel::LoadFromFilePath(modelPath);\n    ticks = GetTickCount() - ticks;\n    printf(\"model file loaded in %d ticks\\n\", ticks);\n\n    \/\/ load the image\n    printf(\"Loading the image...\\n\");\n    auto imageFrame = LoadImageFile(imagePath);\n    \/\/ now create a session and binding\n    LearningModelSession session(model, LearningModelDevice(deviceKind));\/\/Exception thrown at this line\n    LearningModelBinding binding(session);\n\n    \/\/ bind the intput image\n    printf(\"Binding...\\n\");\n    binding.Bind(L\"conv2d_1_input_01\", ImageFeatureValue::CreateFromVideoFrame(imageFrame));\n    \/\/ temp: bind the output (we don't support unbound outputs yet)\n    printf(\"Now...\\n\");\n    vector&lt;int64_t&gt; shape({ 1, 2 });\n    binding.Bind(L\"dense_1_Softmax_0\", TensorFloat::Create(shape));\n\n    \/\/ now run the model\n    printf(\"Running the model...\\n\");\n    ticks = GetTickCount();\n    printf(\"why\");\n    auto results = session.Evaluate(binding, L\"RunId\");\n    printf(\"why\");\n    ticks = GetTickCount() - ticks;\n    printf(\"model run took %d ticks\\n\", ticks);\n\n    \/\/ get the output\n    auto resultTensor = results.Outputs().Lookup(L\"dense_1_Softmax_0\").as&lt;TensorFloat&gt;();\n    auto resultVector = resultTensor.GetAsVectorView();\n    PrintResults(resultVector);\n}\n<\/code><\/pre>\n\n<p>Console output:\n<a href=\"https:\/\/imgur.com\/wKd9RXK\" rel=\"nofollow noreferrer\">https:\/\/imgur.com\/wKd9RXK<\/a><\/p>\n\n<p>Error:<\/p>\n\n<p><a href=\"https:\/\/imgur.com\/hl4osmI\" rel=\"nofollow noreferrer\">https:\/\/imgur.com\/hl4osmI<\/a><\/p>\n\n<p>[Console output in text form]<\/p>\n\n<p>Loading modelfile 'D:\\Projects\\Windows-Machine-Learning-master\\Samples\\SqueezeNetObjectDetection\\Desktop\\cpp\\Debug\\wifi3.onnx' on the 'default' device<\/p>\n\n<p>model file loaded in 531 ticks<\/p>\n\n<p>Loading the image...<\/p>\n\n<p>Binding...<\/p>\n\n<p>Now...<\/p>\n\n<p>Running the model...<\/p>\n\n<p>why<\/p>\n\n<p>Errormsg:<\/p>\n\n<p>Debug Error!<\/p>\n\n<p>Program: ...\/CPP.exe<\/p>\n\n<p>abort() has been called<\/p>\n\n<p>Model link: <a href=\"https:\/\/www.dropbox.com\/s\/fudgynpislpsyta\/wifi3.onnx?dl=0\" rel=\"nofollow noreferrer\">https:\/\/www.dropbox.com\/s\/fudgynpislpsyta\/wifi3.onnx?dl=0<\/a><\/p>\n\n<p>Output by Debugger:<\/p>\n\n<p>Exception thrown at 0x75013442 in SqueezeNetObjectDetectionCPP.exe: Microsoft C++ exception: winrt::hresult_error at memory location 0x006FF5C8. occurred<\/p>\n","answers":[]}
+{"QuestionId":57757515,"AnswerCount":0,"Tags":"<keras><face-detection><tensorflow2.0><tf.keras>","CreationDate":"2019-09-02T13:08:21.377","AcceptedAnswerId":null,"OwnerUserId":11160943.0,"Title":"Is there any implementation of mtcnn face detection in tensorflow 2.0?","Body":"<p>Recently I've moved to tensorflow==2.0.0-rc0 and now mtcnn for face detection is not working on my computer. Can I find tensorflow==2.0.0-rc0 version of mtcnn? Pure Keras implementation of mtcnn would also work in this situation. <\/p>\n\n<p>I've tried keras implementation of facenet at <a href=\"https:\/\/github.com\/nyoki-mtl\/keras-facenet\" rel=\"nofollow noreferrer\">https:\/\/github.com\/nyoki-mtl\/keras-facenet<\/a>. Its nice implementation of facenet in kears but the face detection part (mtcnn in keras) is missing.<\/p>\n","answers":[]}
+{"QuestionId":57758064,"AnswerCount":1,"Tags":"<r><input><keras><generator>","CreationDate":"2019-09-02T13:46:15.147","AcceptedAnswerId":null,"OwnerUserId":5541868.0,"Title":"How to write a Data Generator for a Keras model using multiple inputs","Body":"<p>I need some help with getting a Keras model working in RStudio. The problem arises when I have multiple inputs, and use a data generator.<\/p>\n\n<p>Keras reports an input error. The numpy arrays passed to the model are not what it was expecting.<\/p>\n\n<p>The following toy example reproduces the problem. When the model is fitted normally (no data generator) it runs OK, but when fitted with a generator it crashes.<\/p>\n\n<h1>Set up data<\/h1>\n\n<pre><code>library(magrittr)\nlibrary(keras)\n\n# Create 10 examples of input data and 10 labels\n\ninput1 &lt;- matrix(1:20,  ncol=2, nrow=10, byrow=T)   # [1,2; 3,4; 5,6 ... 19,20]\ninput2 &lt;- matrix(1:30,  ncol=3, nrow=10, byrow=T)    # [1,2,3; 4,5,6 5,6,7 ... 28,29,30]\nlabels &lt;- seq(0.1,1,0.1)                            # [0,1,0.2,0.3 ... 1.0]\n<\/code><\/pre>\n\n<h1>Build &amp; Run Model<\/h1>\n\n<pre><code># define input tensors for the two inputs\nin_a &lt;- layer_input(shape = c(2), name = \"input1\")\nin_b &lt;- layer_input(shape = c(3), name = \"input2\")\n\n# concatenate the inputs and follow them by an output layer\nout &lt;- layer_concatenate(c(in_a, in_b), axis=-1, name=\"concat\") %&gt;% \n  layer_dense(units = 1, activation = 'linear', name=\"output\")\n\n# build the model\nmodel &lt;- keras_model(inputs = list(in_a, in_b), outputs = out)\n\n#compile &amp; run\nmodel %&gt;% compile(loss = \"mse\", optimizer = \"adam\")\nmodel %&gt;% fit(list(input1, input2), labels, epochs = 5)\n<\/code><\/pre>\n\n<h1>Model with generator<\/h1>\n\n<pre><code># The generator will alternatively select the first five input rows and then the second five ad infinitum\ndata_sample_generator &lt;- function(input1, input2, labels) {\n\n  first_five &lt;- 1\n\n  function() {\n\n    first_five &lt;&lt;- ifelse(first_five == 0,1,0)\n\n    if (first_five==0) {\n      rows_to_return &lt;- 1:5 }\n    else {\n      rows_to_return &lt;- 6:10\n    }\n  return(list(input1[rows_to_return, ], input2[rows_to_return, ], labels[rows_to_return])) \n    }\n\n\n}\n# Examine generator output\nbatch &lt;- data_sample_generator(input1, input2, labels)\nbatch()\n\n# Examine generator output\nbatch &lt;- data_sample_generator(input1, input2, labels)\n\n<\/code><\/pre>\n\n<pre><code>batch()    # first sample\n[[1]]\n     [,1] [,2]\n[1,]    1    2\n[2,]    3    4\n[3,]    5    6\n[4,]    7    8\n[5,]    9   10\n\n[[2]]\n     [,1] [,2] [,3]\n[1,]    1    2    3\n[2,]    4    5    6\n[3,]    7    8    9\n[4,]   10   11   12\n[5,]   13   14   15\n\n[[3]]\n[1] 0.1 0.2 0.3 0.4 0.5\n<\/code><\/pre>\n\n<pre><code>batch()    # second sample\n[[1]]\n     [,1] [,2]\n[1,]   11   12\n[2,]   13   14\n[3,]   15   16\n[4,]   17   18\n[5,]   19   20\n\n[[2]]\n     [,1] [,2] [,3]\n[1,]   16   17   18\n[2,]   19   20   21\n[3,]   22   23   24\n[4,]   25   26   27\n[5,]   28   29   30\n\n[[3]]\n[1] 0.6 0.7 0.8 0.9 1.0\n\n<\/code><\/pre>\n\n<p>That's what I would have expected to see from the generator. Now to fit the model.<\/p>\n\n<pre><code>\nmodel %&gt;% \n  fit_generator(data_sample_generator(input1,input2,labels),\n                steps_per_epoch = 2,\n                epochs = 5)\n<\/code><\/pre>\n\n<pre><code>Error in py_call_impl(callable, dots$args, dots$keywords) :   \nValueError: Error when checking model input: the list of Numpy arrays that  \nyou are passing to your model is not the size the model expected.   \nExpected to see 2 array(s),   \nbut instead got the following list of 1 arrays: \n[array([[ 1,  2],\n       [ 3,  4],\n       [ 5,  6],\n       [ 7,  8],\n       [ 9, 10]])]...\n<\/code><\/pre>\n\n<p>I'm not sure what I'm doing wrong here. How can I fix the generator to provide the right shaped inputs? Thanks for any help.<\/p>\n\n<h1>Modify generator output<\/h1>\n\n<p>As @OIDor suggested, change output to return ([input1, input2], labels)<\/p>\n\n<pre><code>return(list(list(input1[rows_to_return, ], input2[rows_to_return, ]), labels[rows_to_return]))\n<\/code><\/pre>\n\n<p>Generator returns<\/p>\n\n<pre><code>[[1]]\n[[1]][[1]]\n     [,1] [,2]\n[1,]    1    2\n[2,]    3    4\n[3,]    5    6\n[4,]    7    8\n[5,]    9   10\n\n[[1]][[2]]\n     [,1] [,2] [,3]\n[1,]    1    2    3\n[2,]    4    5    6\n[3,]    7    8    9\n[4,]   10   11   12\n[5,]   13   14   15\n\n[[2]]\n[1] 0.1 0.2 0.3 0.4 0.5\n<\/code><\/pre>\n\n<p>This time, I get the error message ValueError: could not broadcast input array from shape (5,2) into shape (5)<\/p>\n","answers":[{"AnswerId":"57770314","CreationDate":"2019-09-03T10:58:54.737","ParentId":null,"OwnerUserId":"5165668","Title":null,"Body":"<p>Generator should return a tuple2d:<\/p>\n\n<pre><code> (X, y)\n<\/code><\/pre>\n\n<p>In your case X is a list of arrays, so it becomes:<\/p>\n\n<pre><code>([X1, X2], y)\n<\/code><\/pre>\n\n<p>Your generator yields:<\/p>\n\n<pre><code>list(input1[rows_to_return, ], input2[rows_to_return, ], labels[rows_to_return])\n<\/code><\/pre>\n\n<p>which is equivalent to:<\/p>\n\n<pre><code>([X1, X2, y])\n<\/code><\/pre>\n\n<p>I don't know <code>R<\/code>, but I think you should change your generator to this:<\/p>\n\n<pre><code>list(input1[rows_to_return, ], input2[rows_to_return, ]), labels[rows_to_return]\n<\/code><\/pre>\n\n<p><strong>UPDATE:<\/strong><\/p>\n\n<p>Now that you have updated your code, the Input shapes you pass to your model are:<\/p>\n\n<p>1st input: (5, 2)<\/p>\n\n<p>2nd input: (5, 3)<\/p>\n\n<p>output: (5)<\/p>\n\n<p>Keras error suggests that your model expects to see 1st input as (for example):<\/p>\n\n<pre><code>[1, 3, 5, 7, 9] \n<\/code><\/pre>\n\n<p>while you are passing:<\/p>\n\n<pre><code>[[1, 3, 5, 7, 9],\n[2, 4, 6, 8, 10]]\n<\/code><\/pre>\n\n<p>So you should either change your batch generator, or the model's Input shape<\/p>\n"}]}
+{"QuestionId":57758281,"AnswerCount":0,"Tags":"<pytorch><torch><channels><batch-normalization><momentum>","CreationDate":"2019-09-02T14:01:32.850","AcceptedAnswerId":null,"OwnerUserId":6199298.0,"Title":"pytorch - GroupNorm with momentum","Body":"<p>After looking at the documentation of <strong>pytorch<\/strong> I couldn't find the module of <strong>GroupNorm<\/strong> with <strong>momentum<\/strong>, there is only an implementation which doesn't use it (which is useless to me since I would want to use <strong>AdaBN<\/strong>, or <strong>AdaGN<\/strong> I should say). <\/p>\n\n<p>Therefore my idea was to <strong>take<\/strong> the <strong>BatchNorm<\/strong> class and <strong>extend<\/strong> it with <strong>num_groups<\/strong> and override the method that actually calculate mean and variance by firstly <strong>reshaping<\/strong> the <strong>input<\/strong> and <strong>grouping<\/strong> <strong>channels<\/strong> together, so to emulate the GroupNorm behavior.<\/p>\n\n<p>In order to do this I looked into the source code until I found a dead end in <a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/nn\/functional.html#batch_norm\" rel=\"nofollow noreferrer\">torch.nn.functional.batch_norm<\/a>. That function internally calls another function <code>torch.batch_norm(...)<\/code> but I couldn't find the source code for it.<\/p>\n\n<p>Do you know where to look for that part? Or do you think there is a better way to implement this? I tried looking online for other's works, but I only found unfinished ones.<\/p>\n","answers":[]}
+{"QuestionId":57758384,"AnswerCount":2,"Tags":"<python><keras><multiprocessing>","CreationDate":"2019-09-02T14:08:09.000","AcceptedAnswerId":57761943.0,"OwnerUserId":3154226.0,"Title":"Multiprocessing in Python: Is there a way to use pool.imap without accumulating memory?","Body":"<p>I am using the <code>multiprocessing<\/code> module in Python to train neural networks with <code>keras<\/code> in parallel, using a <code>Pool(processes = 4)<\/code> object with <code>imap<\/code>. This steadily uses more and more memory after every \"cycle\", i.e. every 4 processes, until it finally crashes.<\/p>\n\n<p>I used the <code>memory_profiler<\/code> module to track my memory usage over time, training 12 networks. Here's using vanilla <code>imap<\/code>:\n<img src=\"https:\/\/filedn.com\/lRBwPhPxgV74tO0rDoe8SpH\/mprof_imap_3cycles.png\" alt=\"vanilla\"><\/p>\n\n<p>If I put <code>maxtasksperchild = 1<\/code> in <code>Pool<\/code>:\n<img src=\"https:\/\/filedn.com\/lRBwPhPxgV74tO0rDoe8SpH\/mprof_imap1taskperchild_3cycles.png\" alt=\"1taskperchild\"><\/p>\n\n<p>If I use <code>imap(chunksize = 3)<\/code>:\n<img src=\"https:\/\/filedn.com\/lRBwPhPxgV74tO0rDoe8SpH\/mprof_imapchunks_3cycles.png\" alt=\"chunks\"><\/p>\n\n<p>In the latter case, where everything works out fine, I'm only sending off a single batch to every process in the pool, so it seems that the problem is that the processes carry information about previous batches. If so, can I force the pool to not do that?<\/p>\n\n<p>Even though the chunks solution seems to work I'd rather not use that, because<\/p>\n\n<ul>\n<li>I'd like to track progress using the <code>tqdm<\/code> module, and in the chunks case it will only update after every chunk, which effectively means it won't really track anything at all, as all the chunks finish at the same time (in this example)<\/li>\n<li>Currently all networks take exactly the same time to train, but I'd like to enable the possibility of them having separate training times, where the chunks solution would then potentially cause one process to get all the long training times.<\/li>\n<\/ul>\n\n<p>Here's a code snippet in the vanilla case. In the other two cases I just changed the <code>maxtasksperchild<\/code> parameter in <code>Pool<\/code>, and the <code>chunksize<\/code> parameter in <code>imap<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def train_network(network):\n    (...)\n    return score\n\npool = Pool(processes = 4)\nscores = pool.imap(train_network, networks)\nscores = tqdm(scores, total = networks.size)\n\nfor (network, score) in zip(networks, scores):\n    network.score = score\n\npool.close()\npool.join()\n<\/code><\/pre>\n","answers":[{"AnswerId":"57759522","CreationDate":"2019-09-02T15:29:15.330","ParentId":null,"OwnerUserId":"8282569","Title":null,"Body":"<p>Unfortunaly, <code>multiprocessing<\/code> module in python come with a great expense. data is mostly not shared between processes and need to be replicated. This will change starting from python 3.8.<\/p>\n\n<blockquote>\n  <p><a href=\"https:\/\/docs.python.org\/3.8\/library\/multiprocessing.shared_memory.html\" rel=\"nofollow noreferrer\">https:\/\/docs.python.org\/3.8\/library\/multiprocessing.shared_memory.html<\/a><\/p>\n<\/blockquote>\n\n<p>Although, the official release of python 3.8 is on 21 October 2019, you can already download it on <a href=\"https:\/\/github.com\/python\/cpython\/releases\" rel=\"nofollow noreferrer\">github<\/a><\/p>\n"},{"AnswerId":"57761943","CreationDate":"2019-09-02T19:39:17.743","ParentId":null,"OwnerUserId":"3154226","Title":null,"Body":"<p>I came up with a solution that seems to work. I ditched the pool and made my own simple queuing system. Aside from not increasing (it does increase ever so slightly though, but I think that's me storing some dictionaries as log), it even consumes <em>less<\/em> memory than the chunks solution above:<\/p>\n\n<p><img src=\"https:\/\/filedn.com\/lRBwPhPxgV74tO0rDoe8SpH\/mprof_imap_3cycles_queue.png\" alt=\"imapqueue\"><\/p>\n\n<p>I have no idea why that's the case. Perhaps the <code>Pool<\/code> objects just take up a lot of memory? Anyway, here's my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def train_network(network):\n    (...)\n    return score\n\n# Define queues to organise the parallelising\ntodo = mp.Queue(size = networks.size + 4)\ndone = mp.Queue(size = networks.size)\n\n# Populate the todo queue\nfor idx in range(networks.size):\n    todo.put(idx)\n\n# Add -1's which will be an effective way of checking\n# if all todo's are finished\nfor _ in range(4):\n    todo.put(-1)\n\ndef worker(todo, done):\n    ''' Network scoring worker. '''\n    from queue import Empty\n    while True:\n        try:\n            # Fetch the next todo\n            idx = todo.get(timeout = 1)\n        except Empty:\n            # The queue is never empty, so the silly worker has to go\n            # back and try again\n            continue\n\n        # If we have reached a -1 then stop\n        if idx == -1:\n            break\n        else:\n            # Score the network and store it in the done queue\n            score = train_network(networks[idx])\n            done.put((idx, score))\n\n# Construct our four processes\nprocesses = [mp.Process(target = worker,\n    args = (todo, done)) for _ in range(4)]\n\n# Daemonise the processes, which closes them when\n# they finish, and start them\nfor p in processes:\n    p.daemon = True\n    p.start()\n\n# Set up the iterable with all the scores, and set\n# up a progress bar\nidx_scores = (done.get() for _ in networks)\npbar = tqdm(idx_scores, total = networks.size)\n\n# Compute all the scores in parallel\nfor (idx, score) in pbar:\n    networks[idx].score = score\n\n# Join up the processes and close the progress bar\nfor p in processes:\n    p.join()\npbar.close()\n<\/code><\/pre>\n"}]}
+{"QuestionId":57758510,"AnswerCount":0,"Tags":"<keras><deep-learning><conv-neural-network>","CreationDate":"2019-09-02T14:16:54.997","AcceptedAnswerId":null,"OwnerUserId":9887569.0,"Title":"What is the proper way to determine numbers for batch size, steps_per_epoch, epochs, and validation steps?","Body":"<p>I'm just starting to study Deep Learning and CNNs \nand I'm having trouble getting used to setting parameter values<\/p>\n\n<p>I have 1713 training images and 165 test images<\/p>\n\n<p>how should I change the values in my code to appropriately train my model? (I ran the code and got 30 % accuracy :) ) \nI am a complete beginner and would appreciate any advice thanks<\/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('train',  \n                                             target_size = (64, 64),\n                                             batch_size = 32,\n                                             class_mode = 'binary')\n\ntest_set = test_datagen.flow_from_directory('test',\n                                        target_size = (64, 64),\n                                        batch_size = 32,\n                                        class_mode = 'binary')\n\nclassifier.fit_generator(training_set,\n                     steps_per_epoch = 54,\n                     epochs = 25,\n                     validation_data = test_set,\n                     validation_steps = 165)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57758785,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-datasets><tensorflow-serving><tensorflow-estimator>","CreationDate":"2019-09-02T14:35:44.330","AcceptedAnswerId":null,"OwnerUserId":11765356.0,"Title":"TensorFlow Serving crossed columns strange error","Body":"<p>I am receiving the following error when trying to send a prediction request to my saved model, running with TensorFlow Serving:<\/p>\n\n<pre><code>{\n    \"error\": \"Expected D2 of index to be 2 got 3 at position 0\\n\\t [[{{node linear\/linear_model\/linear_model\/linear_model\/int2Id_X_stringId\/SparseCross}}]]\"\n}\n<\/code><\/pre>\n\n<p>The problem appears to come from trying to use crossed columns in a linear model...?<\/p>\n\n<p>My model in service is a tf.estimator.LinearClassifier . My REST API request is a POST to 'model_directory\/model:predict' with the following intput:<\/p>\n\n<pre><code>{ \"signature_name\": \"predict\",\n    \"instances\": [{\n         \"intId\" : [2],\n        \"int2Id\": [847],\n        \"int3Id\": [0],\n        \"int4Id\":[3],\n        \"stringId\" : [\"STRING\"]}]\n}\n<\/code><\/pre>\n\n<p>Strangely, locally using the saved_model_cli with the following command to get a prediction works absolutely fine:<\/p>\n\n<pre><code>saved_model_cli run --dir Workspace\/saves\/1567433674 --tag_set serve --signature_def predict --input_exprs int4Id=[3];int1Id=[2];stringId=['STRING'];int3Id=[0];int2Id=[847]\n<\/code><\/pre>\n\n<p>The goal is to have a prediction output from my served model, an output similar to when I use saved_model_cli. Something like this:<\/p>\n\n<pre><code>Result for output key all_class_ids:\n[[0 1 2 3 4 5]]\nResult for output key all_classes:\n[[b'0' b'1' b'2' b'3' b'4' b'5']]\nResult for output key class_ids:\n[[2]]\nResult for output key classes:\n[[b'2']]\nResult for output key logits:\n[[ 0.11128154 -0.44881764  0.31520572 -0.08318427 -0.3479367  -0.08883157]]\nResult for output key probabilities:\n[[0.19719791 0.11263006 0.2418051  0.16234797 0.12458517 0.16143374]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"57768068","CreationDate":"2019-09-03T08:44:03.593","ParentId":null,"OwnerUserId":"10340061","Title":null,"Body":"<p>Change \"instances\" to \"inputs\" and have another try.<\/p>\n"}]}
+{"QuestionId":57758831,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-02T14:39:37.387","AcceptedAnswerId":57794812.0,"OwnerUserId":8176285.0,"Title":"On how to run custom keras code on TPU cloud","Body":"<p>I have my code written to run on GPU, but the turnaround time for experiments is a lot.I want to port this code so i can run it on TPU. how i can do so? This i what i have.<\/p>\n\n<pre><code>Bunch of datloading stuff\n!!!!!\n!!!!!\n\neffnet = efn.EfficientNetB5(weights=None, include_top=False)\neffnet.load_weights(eff_weights_path)\n\n# Replace all Batch Normalization layers by Group Normalization layers\n\nmodel = Sequential()\nmodel.add(effnet)\nmodel.add(GlobalAveragePooling2D())\nmodel.add(Dropout(0.5))\nmodel.add(Dense(8,name ='elu', activation=elu))\nmodel.load_weights('saved_models\/wieghts_ef5.h5',by_name = True)\nmodel.compile(loss='mse', optimizer='adam', metrics=['mse', 'acc'])\nprint(model.summary())        \n\nmodel_json = model.to_json()\nwith open(\".\/saved_models\/model_ef5_fn.json\", \"w\") as json_file:\n    json_file.write(model_json)\n\nBATCH_SIZE = 2\nIMG_WIDTH  = 456\nIMG_HEIGHT = 456\n# Add Image augmentation to our generator\ntrain_datagen = ImageDataGenerator(rotation_range=360,\n                                   horizontal_flip=True,\n                                   vertical_flip=True,\n                                   validation_split=0.15,\n                                   preprocessing_function=preprocess_image, \n                                   rescale=1 \/ 128.)\n\n# Use the dataframe to define train and validation generators\ntrain_generator = train_datagen.flow_from_dataframe(df, \n                                                    x_col='pic_id', \n                                                    y_col=Labels_list,\n                                                    directory = train_images_path,\n                                                    target_size=(IMG_WIDTH, IMG_HEIGHT),\n                                                    batch_size=BATCH_SIZE,\n                                                    class_mode='other', \n                                                    subset='training')\n\nval_generator = train_datagen.flow_from_dataframe(df, \n                                                  x_col='pic_id', \n                                                  y_col=Labels_list,\n                                                  directory = train_images_path,\n                                                  target_size=(IMG_WIDTH, IMG_HEIGHT),\n                                                  batch_size=BATCH_SIZE,\n                                                  class_mode='other',\n                                                  subset='validation')\nprint('Data was loaded')\n\n\n# For tracking Quadratic Weighted Kappa score\nkappa_metrics = Metrics()\n# Monitor MSE to avoid overfitting and save best model\nes = EarlyStopping(monitor='val_loss', mode='auto', verbose=1, patience=15)\nrlr = ReduceLROnPlateau(monitor='val_loss', \n                        factor=0.5, \n                        patience=4, \n                        verbose=1, \n                        mode='auto', \n                        min_delta=0.0001)\n\n# Begin training\nmodel.fit_generator(train_generator,\n                    # steps_per_epoch=60,\n                    steps_per_epoch=train_generator.samples \/\/ BATCH_SIZE,                    \n                    epochs=30,\n                    validation_data=val_generator,\n                    # validation_steps = 30,\n                    validation_steps = val_generator.samples \/\/ BATCH_SIZE,                    \n                    callbacks=[kappa_metrics, es, rlr])\n<\/code><\/pre>\n","answers":[{"AnswerId":"57794812","CreationDate":"2019-09-04T19:34:19.273","ParentId":null,"OwnerUserId":"4358570","Title":null,"Body":"<p>As discussed in chat here is the basic design to train your model on a TPU:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import os\nimport tensorflow as tf\nfrom tensorflow.keras import *\n\n!pip install -U git+https:\/\/github.com\/qubvel\/efficientnet\nimport efficientnet.tfkeras as efn\n\nbackend.clear_session()\n\nif \"COLAB_TPU_ADDR\" not in os.environ:\n    raise Exception(\"TPU not found\")\n\ntpu_addr = \"grpc:\/\/\" + os.environ[\"COLAB_TPU_ADDR\"]\nprint(\"TPU address:\", tpu_addr)\n\nresolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu = tpu_addr)\ntf.config.experimental_connect_to_host(resolver.master())\ntf.tpu.experimental.initialize_tpu_system(resolver)\nstrategy = tf.distribute.experimental.TPUStrategy(resolver)\n\ndef get_model():\n  effnet = efn.EfficientNetB5(weights = None, include_top = False)\n\n  # Replace all Batch Normalization layers by Group Normalization layers\n  for i, layer in enumerate(effnet.layers):\n      if \"batch_normalization\" in layer.name:\n          effnet.layers[i] = GroupNormalization(groups = 32, axis = -1, epsilon = 1e-5)\n\n  model = models.Sequential()\n  model.add(effnet)\n  # add other layers\n  return model\n\nwith strategy.scope():\n  model = get_model()\n  model.compile(optimizer = optimizers.Adagrad(learning_rate = 0.1), \n    loss = \"sparse_categorical_crossentropy\",\n    metrics = [\"sparse_categorical_crossentropy\"]\n  )\n\n  #model.load_weights(\"\/path\/to\/imagenet\/weights\")\n\nprint(model.summary())\n\n# model.fit(...)\nmodel.save_weights(\"weights.h5\")\n\ncpu_model = get_model()\ncpu_model.load_weights(\"weights.h5\")\n#cpu_model.predict(...)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57758965,"AnswerCount":0,"Tags":"<python-3.x><tensorflow>","CreationDate":"2019-09-02T14:49:53.027","AcceptedAnswerId":null,"OwnerUserId":11328779.0,"Title":"Tensorflow 2.0 (Python) - Parrallel Loop in Graph Mode","Body":"<p>Suppose there are variables in different shapes:<br><\/p>\n\n<pre><code>tvs = model.trainable_variables<\/code><\/pre>\n\n<p>In case one wants to apply some custom optimization operations to these variables, each variable should have its own graph:<\/p>\n\n<pre><code>ops = []\nfor idx, tv in enumerate(tvs):\n    @tf.function\n    def op(xs):\n        # Some operations\n        tv.assign_add(xs[idx])\n    ops.append(op)<\/code><\/pre>\n\n<p>and then apply these operations somewhere else (e.g. in training loop):<\/p>\n\n<pre><code>@tf.function\ndef step_ops(xs):\n    for op in ops : op(xs)<\/code><\/pre>\n\n<p>in which <code>step_ops<\/code> takes list of tensors (e.g., gradients of variables) as input.<\/p>\n\n<p>The question is, will op insides ops runs parallelly in step_ops?<\/p>\n\n<p>If not, how to get it run parallelly? Or are there any better approaches to achieve this?<\/p>\n\n<p>Many thanks.<\/p>\n","answers":[]}
+{"QuestionId":57759137,"AnswerCount":0,"Tags":"<python><pytorch>","CreationDate":"2019-09-02T15:00:39.060","AcceptedAnswerId":null,"OwnerUserId":5481214.0,"Title":"How nn.Linear works in shape(batch_size, seq_length, hidden_size)?","Body":"<p>In <code>self.classifier<\/code>, I assume there should be different weights to apply on tokens. Below is the huggingface's implementation. <\/p>\n\n<pre><code>def __init__(self, config, num_labels=2):\n    super(BertForTokenClassification, self).__init__(config)\n    self.num_labels = num_labels\n    self.bert = BertModel(config)\n    self.dropout = nn.Dropout(config.hidden_dropout_prob)\n    self.classifier = nn.Linear(config.hidden_size, num_labels)\n    self.apply(self.init_bert_weights)\n\ndef forward(self, input_ids, token_type_ids=None, attention_mask=None, labels=None):\n    sequence_output, _ = self.bert(input_ids, token_type_ids, attention_mask, output_all_encoded_layers=False)\n    sequence_output = self.dropout(sequence_output)\n    logits = self.classifier(sequence_output)\n\n    if labels is not None:\n        loss_fct = CrossEntropyLoss()\n        # Only keep active parts of the loss\n        if attention_mask is not None:\n            active_loss = attention_mask.view(-1) == 1\n            active_logits = logits.view(-1, self.num_labels)[active_loss]\n            active_labels = labels.view(-1)[active_loss]\n            loss = loss_fct(active_logits, active_labels)\n        else:\n            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))\n        return loss\n    else:\n        return logits\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57759563,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2019-09-02T15:32:49.633","AcceptedAnswerId":57760061.0,"OwnerUserId":10665023.0,"Title":"Minimize multivariate function in Tensorflow","Body":"<p>Suppose I have the following simple example of a function of several variables<\/p>\n\n<pre><code>@tf.function\ndef f(A, Y, X):\n  AX = tf.matmul(A, X)\n  norm = tf.norm(Y - AX)\n  return norm\n\nN = 2\nA = tf.Variable(np.array([[1, 2], [3, 4]]))\nY = tf.Variable(np.identity(N))\nX = tf.Variable(np.zeros((N, N)))\n<\/code><\/pre>\n\n<p>How do I find <code>X<\/code> that minimizes <code>f<\/code> with Tensorflow ?\nI would be interested in a generic solution that works with a function declared as above and when there are more than one variable to optimize.<\/p>\n","answers":[{"AnswerId":"57760061","CreationDate":"2019-09-02T16:13:23.883","ParentId":null,"OwnerUserId":"1561141","Title":null,"Body":"<p>Assuming Tensorflow 2, you can use a Keras optimizer:<\/p>\n\n<pre><code>@tf.function\ndef f(A, Y, X):\n    AX = tf.matmul(A, X)\n    norm = tf.norm(Y - AX)\n    return norm\n\nN = 2\nA = tf.Variable(np.array([[1., 2.], [3., 4.]]))\nY = tf.Variable(np.identity(N))\nX = tf.Variable(np.zeros((N, N)))\n\noptimizer = tf.keras.optimizers.SGD()\nfor iteration in range(0, 100):\n    with tf.GradientTape() as tape:\n        loss = f(X, Y, X)\n        print(loss)\n\n    grads = tape.gradient(loss, [A, Y, X])\n    optimizer.apply_gradients(zip(grads, [A, Y, X]))\n\nprint(A, Y, X)\n<\/code><\/pre>\n\n<p>That will work for any differentiable function. For non-differentiable functions you could look at other optimization techniques (such as Genetic Algorithms, or Swarm Optimization. NEAT has implementations of these <a href=\"https:\/\/neat-python.readthedocs.io\/en\/latest\/\" rel=\"nofollow noreferrer\">https:\/\/neat-python.readthedocs.io\/en\/latest\/<\/a>).<\/p>\n"},{"AnswerId":"57760158","CreationDate":"2019-09-02T16:21:57.617","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p><a href=\"https:\/\/stackoverflow.com\/a\/57760061\/1782792\">avanwyk<\/a> is essentially right, although note that: 1) you can directly use the <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/optimizers\/Optimizer#minimize\" rel=\"nofollow noreferrer\"><code>minimize<\/code><\/a> method of the optimizer for simplicity 2) if you only want to optimize <code>X<\/code> you should make sure that is the only variable you are updating.<\/p>\n\n<pre><code>import tensorflow as tf\n\n@tf.function\ndef f(A, Y, X):\n  AX = tf.matmul(A, X)\n  norm = tf.norm(Y - AX)\n  return norm\n\n# Input data\nN = 2\nA = tf.Variable([[1., 2.], [3., 4.]], tf.float32)\nY = tf.Variable(tf.eye(N, dtype=tf.float32))\nX = tf.Variable(tf.zeros((N, N), tf.float32))\n# Initial function value\nprint(f(A, Y, X).numpy())\n# 1.4142135\n\n# Setup a stochastic gradient descent optimizer\nopt = tf.keras.optimizers.SGD(learning_rate=0.01)\n# Define loss function and variables to optimize\nloss_fn = lambda: f(A, Y, X)\nvar_list = [X]\n# Optimize for a fixed number of steps\nfor _ in range(1000):\n    opt.minimize(loss_fn, var_list)\n\n# Optimized function value\nprint(f(A, Y, X).numpy())\n# 0.14933111\n\n# Optimized variable\nprint(X.numpy())\n# [[-2.0012102   0.98504114]\n#  [ 1.4754106  -0.5111093 ]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57759627,"AnswerCount":0,"Tags":"<python-3.x><keras><conv-neural-network><recurrent-neural-network>","CreationDate":"2019-09-02T15:37:24.567","AcceptedAnswerId":null,"OwnerUserId":7003616.0,"Title":"Must the input height of a 1D CNN be constant?","Body":"<p>I'm currently doing my honours research project on online\/dynamic signature verification. I am using the SVC 2004 dataset (Task 2). I have done the following data processing:<\/p>\n\n<pre><code>def load_dataset_normalized(path):\nfile_names = os.listdir(path)\n\nnum_of_persons = len(file_names)\n\ninitial_starting_point = np.zeros(np.shape([7]))\n\nx_dataset = []\ny_dataset = []\n\nfor infile in file_names:\n    full_file_name = os.path.join(path, infile)\n    file = open(full_file_name, \"r\")\n    file_lines = file.readlines()\n\n    num_of_points = int(file_lines[0])\n\n    x = []\n    y = []\n    time_stamp = []\n    button_status = []\n    azimuth_angles = []\n    altitude = []\n    pressure = []\n\n    for idx, line in enumerate(file_lines[1:]):\n        idx+=1\n        nums = line.split(' ')\n\n        if idx == 1:\n            nums[2] = 0\n            initial_starting_point = nums\n\n            x.append(int(nums[0]))\n            y.append(int(nums[1]))\n            time_stamp.append(0)\n            button_status.append(int(nums[3]))\n            azimuth_angles.append(int(nums[4]))\n            altitude.append(int(nums[5]))\n            pressure.append(int(nums[6]))\n\n        else:\n            x.append(int(nums[0]))\n            y.append(int(nums[1]))\n            time_stamp.append(10)\n            button_status.append(int(nums[3]))\n            azimuth_angles.append(int(nums[4]))\n            altitude.append(int(nums[5]))\n            pressure.append(int(nums[6]))\n\n    max_x = max(x)\n    max_y = max(y)\n    max_azimuth_angle = max(azimuth_angles)\n    max_altitude = max(altitude)\n    max_pressure = max(pressure)\n\n    min_x = min(x)\n    min_y = min(y)\n    min_azimuth_angle = min(azimuth_angles)\n    min_altitude = min(altitude)\n    min_pressure = min(pressure)\n\n    #Alignment normalization:\n    for i in range(num_of_points):\n        x[i] -= int(initial_starting_point[0])\n        y[i] -= int(initial_starting_point[1])\n        azimuth_angles[i] -= int(initial_starting_point[4])\n        altitude[i] -= int(initial_starting_point[5])\n        pressure[i] -= int(initial_starting_point[6])\n\n    #Size normalization\n    for i in range(num_of_points):\n        x[i] = ((x[i] - max_x) \/ (min_x - max_x))\n        y[i] = ((y[i] - max_y) \/ (min_y - max_y))\n        azimuth_angles[i] = ((azimuth_angles[i] - max_azimuth_angle) \/ (min_azimuth_angle - max_azimuth_angle))\n        altitude[i] = ((altitude[i] - max_altitude) \/ (min_altitude - max_altitude))\n        pressure[i] = ((pressure[i] - max_pressure) \/ (min_pressure - max_pressure))\n\n    #data points to dataset\n    x_line = []\n    for i in range (num_of_points):\n        x_line.append([x[i], y[i], time_stamp[i], button_status[i], azimuth_angles[i], altitude[i], pressure[i]])\n\n        if i == num_of_points-1:\n            x_dataset.append(x_line)\n\n    infile_without_extension = infile.replace('.TXT','')\n    index_of_s = infile_without_extension.find(\"S\")\n    index_of_num = index_of_s + 1\n    sig_ID = int(infile_without_extension[index_of_num:])\n    if sig_ID &lt; 21:\n        y_dataset.append([1,0])\n    else:\n        y_dataset.append([0,1])\n\nx_dataset = np.asarray(x_dataset)\ny_dataset = np.asarray(y_dataset)\nreturn x_dataset, y_dataset\n<\/code><\/pre>\n\n<p>I also have another method that takes the values as they are in the text file and created an \"original\" dataset.<\/p>\n\n<p>Now, the aim of my research is to create a CRNN (convolutional recurrent neural network) that can identify if a signature is authentic or forged. Here is the code for the model:<\/p>\n\n<pre><code>class crnn_model:\ndef __init__(self, trainX, trainy, testX, testy, optimizer_method):\n    self.trainX = trainX\n    self.trainy = trainy\n    self.testX = testX\n    self.testy = testy\n\n    self.evaluate_model(optimizer_method)\n\ndef evaluate_model(self, optimizer_method):\n    verbose, epochs, batch_size = 0, 40, 10\n    n_timesteps, n_features, n_outputs = len(self.trainX), 7, 2\n    print(n_timesteps)\n    model = keras.Sequential()\n    model.add(keras.layers.Conv1D(filters=64, kernel_size=3, activation='relu', input_shape=(n_timesteps, n_features), use_bias=True))\n    model.add(keras.layers.Conv1D(filters=64, kernel_size=3, activation='relu'))\n    model.add(keras.layers.Dropout(0.5))\n    model.add(keras.layers.MaxPooling1D(pool_size=2))\n    model.add(keras.layers.Flatten())\n    model.add(keras.layers.LSTM(2, input_shape=[30592,1], return_sequences=True))\n    model.summary()\n\n    # Compile the model\n    model.compile(optimizer=optimizer_method, loss='categorical_crossentropy', metrics=['accuracy'])\n\n    #fit model\n    model.fit(self.trainX, self.trainy, epochs=epochs, batch_size=batch_size, verbose=verbose)\n\n    # evaluate model\n    _, accuracy = model.evaluate(self.testX, self.testy, batch_size=batch_size, verbose=0)\n    return accuracy\n<\/code><\/pre>\n\n<p>Here is the problem I am having: the number of points used to store each signature is different, hence making the input height of the input matrix vary from one signature to the next. Must I now force the dataset to some uniform\/constant number of points? <\/p>\n\n<p>Much appreciated for your time.<\/p>\n","answers":[]}
+{"QuestionId":57759635,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorboard><tensorflow2.0>","CreationDate":"2019-09-02T15:37:55.490","AcceptedAnswerId":57834801.0,"OwnerUserId":3694044.0,"Title":"Get Gradients with Keras Tensorflow 2.0","Body":"<p>I would like to keep track of the <strong>gradients<\/strong> over tensorboard.\nHowever, since session run statements are not a thing anymore and the <em>write_grads<\/em> argument of <em>tf.keras.callbacks.TensorBoard<\/em> is <strong>depricated<\/strong>, I would like to know how to keep track of gradients during training with <strong>Keras<\/strong> or <strong>tensorflow 2.0<\/strong>.<\/p>\n\n<p>My current approach is to create a new callback class for this purpose, but without success. Maybe someone else knows how to accomplish this kind of advanced stuff.<\/p>\n\n<p>The code created for testing is shown below, but runs into errors independently of printing a gradient value to console or tensorboard.<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python.keras import backend as K\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(input_shape=(28, 28)),\n  tf.keras.layers.Dense(128, activation='relu', name='dense128'),\n  tf.keras.layers.Dropout(0.2),\n  tf.keras.layers.Dense(10, activation='softmax', name='dense10')\n])\n\nmodel.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])\n\n\nclass GradientCallback(tf.keras.callbacks.Callback):\n    console = True\n\n    def on_epoch_end(self, epoch, logs=None):\n        weights = [w for w in self.model.trainable_weights if 'dense' in w.name and 'bias' in w.name]\n        loss = self.model.total_loss\n        optimizer = self.model.optimizer\n        gradients = optimizer.get_gradients(loss, weights)\n        for t in gradients:\n            if self.console:\n                print('Tensor: {}'.format(t.name))\n                print('{}\\n'.format(K.get_value(t)[:10]))\n            else:\n                tf.summary.histogram(t.name, data=t)\n\n\nfile_writer = tf.summary.create_file_writer(\".\/metrics\")\nfile_writer.set_as_default()\n\n# write_grads has been removed\ntensorboard_cb = tf.keras.callbacks.TensorBoard(histogram_freq=1, write_grads=True)\ngradient_cb = GradientCallback()\n\nmodel.fit(x_train, y_train, epochs=5, callbacks=[gradient_cb, tensorboard_cb])\n<\/code><\/pre>\n\n<ul>\n<li>Priniting bias gradients to console (console parameter = True)\nleads to: <strong>AttributeError: 'Tensor' object has no attribute 'numpy'<\/strong><\/li>\n<li>Writing to tensorboard (console parameter = False) creates:\n<strong>TypeError: Using a <code>tf.Tensor<\/code> as a Python <code>bool<\/code> is not allowed. Use <code>if t is not None:<\/code> instead of <code>if t:<\/code> to test if a tensor is defined, and use TensorFlow ops such as tf.cond to execute subgraphs conditioned on the\nvalue of a tensor.<\/strong><\/li>\n<\/ul>\n","answers":[{"AnswerId":"57834801","CreationDate":"2019-09-07T14:54:30.930","ParentId":null,"OwnerUserId":"931303","Title":null,"Body":"<p>To compute the gradients of the loss against the weights, use <\/p>\n\n<pre><code>with tf.GradientTape() as tape:\n    loss = model(model.trainable_weights)\n\ntape.gradient(loss, model.trainable_weights)\n<\/code><\/pre>\n\n<p>This is (arguably poorly) documented on <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/GradientTape\" rel=\"nofollow noreferrer\">GradientTape<\/a>.<\/p>\n\n<p>We do not need to <code>tape.watch<\/code> the variable because trainable parameters are watched by default.<\/p>\n\n<p>As a function, it can be written as<\/p>\n\n<pre><code>def gradient(model, x):\n    x_tensor = tf.convert_to_tensor(x, dtype=tf.float32)\n    with tf.GradientTape() as t:\n        t.watch(x_tensor)\n        loss = model(x_tensor)\n    return t.gradient(loss, x_tensor).numpy()\n<\/code><\/pre>\n"}]}
+{"QuestionId":57759710,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2019-09-02T15:43:44.320","AcceptedAnswerId":null,"OwnerUserId":11880712.0,"Title":"convert grayscale image to RGB cifar10 by keras autoencoder","Body":"<p>I wanna create a model which predict the RGB image from grayscale one.\nI use binary cross-entropy, but the loss and val_loss are negative(~-1908)\nthen I use MSE as loss, now the loss is around 300.\nwhich one is correct or how can I improve my model?<\/p>\n\n<pre><code>input_layer = Input(shape=(32,32,1))\n\n\n# encoder\n\n\nconv1 = Conv2D(32, (3, 3), activation='elu', padding='same', name='convl1e')(input_layer)\nconv2 = Conv2D(64, (3, 3), activation='elu', padding='same', name='convl2e')(conv1)\n\ne1 = MaxPooling2D((2,2), padding='same')(conv2)\n\nconv3 = Conv2D(128, (3, 3), activation='elu', padding='same', name='convl3e')(e1)\nconv4 = Conv2D(256, (3, 3), activation='elu', padding='same', name='convl4e')(conv3)\n\ne2 = MaxPooling2D((2,2), padding='same')(conv4)\n\n# conv5 = Conv2D(16, (3, 3), activation='elu', padding='same', name='convl5e')(e2)\n# conv6 = Conv2D(16, (3, 3), activation='elu', padding='same', name='convl6e')(conv5)\n\n# e3 = MaxPooling2D((2,2), padding='same')(conv6)\n\n\n# bottle neck\nb = Conv2D(8, (3,3), activation='relu', padding='same',name='bottle_neck')(e2)\n\n\n# decoder\nd1 = UpSampling2D((2,2))(b)\ndeconv1 = Conv2D(128, (3, 3), activation='elu', padding='same', name='deconvl1d')(d1)\ndeconv2 = Conv2D(64, (3, 3), activation='elu', padding='same', name='deconvl2d')(deconv1)\n\nd2 = UpSampling2D((2,2))(deconv2)\n\ndeconv3 = Conv2D(32, (3, 3), activation='elu',padding='same', name='deconvl3d')(d2)\ndeconv4 = Conv2D(32, (3, 3), activation='elu',padding='same', name='deconvl4d')(deconv3)\n# deconv5 = Conv2D(16, (3, 3), activation='elu',padding='same', name='deconvl5d')(deconv4)\n\n# d3 = UpSampling2D((2,2))(deconv3)\n\n\n# output layer\no = Conv2D(3, (3,3),  activation='elu', padding='same', name='output')(deconv4)\n\nautoencoder = Model(input_layer, o)\n\nlearning_rate = 0.001\ndecay_rate = learning_rate \/ 50\nopt = optimizers.Adam(lr=learning_rate, beta_1=0.9, beta_2=0.999, epsilon=None, decay=decay_rate, amsgrad=False)\n\n\nautoencoder.compile(optimizer=opt, loss='mean_squared_error')\n\n# data preprocessing\n(X_train,_), (X_test, _)= cifar10.load_data()\nX_train_grayscale =  rgb2gray(X_train)\nX_test_grayscale =  rgb2gray(X_test)\nX_train_gray= X_train_grayscale.reshape(-1,32,32,1)\nX_test_gray= X_test_grayscale.reshape(-1,32,32,1)\nprint (X_train_gray.shape,X_test_gray.shape)\n\n\nhistory = autoencoder.fit(X_train_gray, X_train,\n               epochs = 50,\n               batch_size=32,\n               validation_data=(X_test_gray, X_test))\n<\/code><\/pre>\n\n<p>is this correct and how can I improve it<\/p>\n","answers":[]}
+{"QuestionId":57759752,"AnswerCount":0,"Tags":"<python><arrays><tensorflow><keras><deep-learning>","CreationDate":"2019-09-02T15:46:44.057","AcceptedAnswerId":null,"OwnerUserId":11877433.0,"Title":"ValueError: Error when checking target: expected category_output to have shape (2,) but got array with shape (1,)","Body":"<p>i made  a model with 2 sub branches, one to recognize the category and one to recognize color. it worked perfectly.\nnow i modified the code and added another 2 sub branches and i get the following error. note that when i used two branches it worked fine and i didnt have problems with array shape.<\/p>\n\n<p>this is the code: <\/p>\n\n<p>nothing worked<\/p>\n\n<pre><code># initialize the data, clothing category labels (i.e., shirts, jeans,\n# dresses, etc.) along with the color labels (i.e., red, blue, etc.)\ndata = []\ncategoryLabels = []\ncolorLabels = []\nattribute1Labels = []\nattribute2Labels = []\n# loop over the input images\nfor imagePath in imagePaths:\n    try:\n        # load the image, pre-process it, and store it in the data list\n        image = cv2.imread(imagePath)\n        image = cv2.resize(image, (IMAGE_DIMS[1], IMAGE_DIMS[0]))\n        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n        image = img_to_array(image)\n        data.append(image)\n\n        # extract the clothing color and category from the path and\n        # update the respective lists\n        (cat, attribute1 , attribute2 , color) = imagePath.split(os.path.sep)[-2].split(\"_\")\n        categoryLabels.append(cat)\n        attribute1Labels.append(attribute1)\n        attribute2Labels.append(attribute2)\n        colorLabels.append(color)\n    except Exception as e:\n            pass\n# scale the raw pixel intensities to the range [0, 1] and convert to\n# a NumPy array\ndata = np.array(data, dtype=\"float\") \/ 255.0\nprint(\"[INFO] data matrix: {} images ({:.2f}MB)\".format(\n    len(imagePaths), data.nbytes \/ (1024 * 1000.0)))\n\n# convert the label lists to NumPy arrays prior to binarization\ncategoryLabels = np.array(categoryLabels)\ncolorLabels = np.array(colorLabels)\nattribute1Labels = np.array(attribute1Labels)\nattribute2Labels = np.array(attribute2Labels)\n\n# binarize both sets of labels\nprint(\"[INFO] binarizing labels...\")\ncategoryLB = LabelBinarizer()\ncolorLB = LabelBinarizer()\natt1LB = LabelBinarizer()\natt2LB = LabelBinarizer()\ncategoryLabels = categoryLB.fit_transform(categoryLabels)\ncolorLabels = colorLB.fit_transform(colorLabels)\nattribute1Labels = att1LB.fit_transform(attribute1Labels)\nattribute2Labels = att2LB.fit_transform(attribute2Labels)\n\n# partition the data into training and testing splits using 80% of\n# the data for training and the remaining 20% for testing\nsplit = train_test_split(data, categoryLabels,attribute1Labels,attribute2Labels, colorLabels,\n    test_size=0.2, random_state=42)\n(trainX, testX, trainCategoryY, testCategoryY, trainAtt1, testAtt1 , trainAtt2,testAtt2,\n    trainColorY, testColorY) = split\n\n# initialize our FashionNet multi-output network\nmodel = FashionNet.build(96, 96,\n    numCategories=len(categoryLB.classes_), numAtts1 = len(att1LB.classes_) , numAtts2 = len(att2LB.classes_),\n    numColors=len(colorLB.classes_),\n    finalAct=\"softmax\")\n\n# define two dictionaries: one that specifies the loss method for\n# each output of the network along with a second dictionary that\n# specifies the weight per loss\nlosses = {\n    \"category_output\": \"categorical_crossentropy\", \"attribute1_output\": \"categorical_crossentropy\",\"attribute2_output\": \"categorical_crossentropy\",\n    \"color_output\": \"categorical_crossentropy\",\n}\nlossWeights = {\"category_output\": 1.0, \"color_output\": 1.0 , \"attribute1_output\":1.0 , \"attribute2_output\":1.0}\n\n# initialize the optimizer and compile the model\nprint(\"[INFO] compiling model...\")\nopt = Adam(lr=INIT_LR, decay=INIT_LR \/ EPOCHS)\nmodel.compile(optimizer=opt, loss=losses, loss_weights=lossWeights,\n    metrics=[\"accuracy\"])\nH = model.fit(trainX,\n    {\"category_output\": trainCategoryY , \"attribute1_output\":trainAtt1 ,\"attribute2_output\":trainAtt2, \"color_output\": trainColorY},\n    validation_data=(testX,\n        {\"category_output\": testCategoryY, \"attribute1_output\":testAtt1,\"attribute2_output\":testAtt2, \"color_output\": testColorY}),\n    epochs=EPOCHS,\n    verbose=1)\n\n# save the model to disk\nprint(\"[INFO] serializing network...\")\n<\/code><\/pre>\n\n<p>and this is the error : <\/p>\n\n<pre><code>ValueError                                Traceback (most recent call last)\n&lt;ipython-input-16-7d553dc48bc9&gt; in &lt;module&gt;\n      5         {\"category_output\": testCategoryY, \"attribute1_output\":testAtt1,\"attribute2_output\":testAtt2, \"color_output\": testColorY}),\n      6         epochs=EPOCHS,\n----&gt; 7     verbose=1)\n      8 \n      9 # save the model to disk\n\nc:\\users\\sokhen\\appdata\\local\\programs\\python\\python37\\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    970                 val_x, val_y,\n    971                 sample_weight=val_sample_weight,\n--&gt; 972                 batch_size=batch_size)\n    973             if self._uses_dynamic_learning_phase():\n    974                 val_ins = val_x + val_y + val_sample_weights + [0.]\n\nc:\\users\\sokhen\\appdata\\local\\programs\\python\\python37\\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\nc:\\users\\sokhen\\appdata\\local\\programs\\python\\python37\\lib\\site-packages\\keras\\engine\\training_utils.py in standardize_input_data(data, names, shapes, check_batch_axis, exception_prefix)\n    136                             ': expected ' + names[i] + ' to have shape ' +\n    137                             str(shape) + ' but got array with shape ' +\n--&gt; 138                             str(data_shape))\n    139     return data\n    140 \n\nValueError: Error when checking target: expected category_output to have shape (2,) but got array with shape (1,)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57759890,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras><deep-learning><tensorflow2.0>","CreationDate":"2019-09-02T15:59:15.973","AcceptedAnswerId":null,"OwnerUserId":11289239.0,"Title":"How to take the output from some layers within the model and use it to compute the final loss?","Body":"<p>I am trying to implement SFSnet(<a href=\"https:\/\/images.app.goo.gl\/pweUBKkrwyDVWeG48\" rel=\"nofollow noreferrer\">https:\/\/images.app.goo.gl\/pweUBKkrwyDVWeG48<\/a>) using tensorflow 2.0, in there the loss of the model is dependent on the output of the normal_conv, light_estimator, albedo_conv and recon_image as <\/p>\n\n<pre><code>final_loss = 0.5 * (normal_loss + albedo_loss + recon_loss) + 0.1 * light_loss \n<\/code><\/pre>\n\n<p>Now to calculate the normal_loss, albedo_loss, light_loss and recon_loss I need to call those layers in my custom loss function..<\/p>\n\n<pre><code>def custom_loss(lighting):\n\n    def loss(y_target,y_predicted):\n\n        # Defining our loss functions\n        mae = tf.keras.losses.MeanAbsoluteError()\n        mse = tf.keras.losses.MeanSquaredError()\n        normal_pred, albedo_pred, lighting_pred, recon_pred = y_predicted\n        normal_target, albedo_target, lighting_target, recon_target = y_target\n\n        # Get normal loss : L1 loss\n        normal_loss = mae(normal_target, normal_pred)\n        # Get albedo loss : L1 loss\n        albedo_loss = mae(albedo_target, albedo_pred)\n        # Get lighting loss : L2 Loss\n        light_loss = mse(lighting_target, lighting_pred)\n\n        # Get reconstrucion loss : L1 loss\n        recon_loss = mae(recon_target, recon_pred)\n\n\n        l = 0.5 * (normal_loss + albedo_loss + recon_loss) + 0.1 * light_loss \n        return l\n\n    return loss\n<\/code><\/pre>\n\n<p>And that did not work, so is there any way we can use the outputs of layers in between the model and compute their losses and use those losses to calculate the final loss and train the model with that loss.<\/p>\n","answers":[]}
+{"QuestionId":57759984,"AnswerCount":1,"Tags":"<python><tensorflow><keras><regression>","CreationDate":"2019-09-02T16:06:39.220","AcceptedAnswerId":null,"OwnerUserId":10699184.0,"Title":"Is it wrong to have the input shape as a sum of number of features and labels","Body":"<p>I try to make a regression and I have 3 features and 1 label (they all have 190 entries). I used at first <code>input_shape=[len(dataset.keys())]<\/code>, which is 4 and the results are very good. <\/p>\n\n<p>Now I've realized that I should have written input_shape=<code>[len(training_dataset.keys())]<\/code>, which is 3 and tried many combinations of layers of different sizes and activation functions, to find the best model.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def build_model():\n  model = keras.Sequential([\n    layers.Dense(64, activation='elu', input_shape=[len(dataset.keys())]),\n    layers.Dense(64, activation='relu'),\n    layers.Dense(1)\n  ])\n\n  optimizer = tf.keras.optimizers.RMSprop(0.001)\n\n  model.compile(loss='mse',\n                optimizer=optimizer,\n                metrics=['mae', 'mse'])\n  return model\n<\/code><\/pre>\n\n<p>The results are awful (mean error rate used to be 156 and now is above 4500). So can I let it be like it was in the beginning or is it wrong?<\/p>\n","answers":[{"AnswerId":"57760976","CreationDate":"2019-09-02T17:46:18.727","ParentId":null,"OwnerUserId":"2932946","Title":null,"Body":"<p>Depending on the type of your dataset, there are better ways to feed the amount of columns. The easiest and most straightforward (for me at least) is:<\/p>\n\n<pre><code>Dense(12, activation='relu', input_dim=training_dataset.shape[1])\n<\/code><\/pre>\n\n<p>That will help you in case your dataset is a Numpy array or a Pandas dataframe.<\/p>\n\n<p>However, since you're using a function for building the model, I'd encourage you to make it like<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def build_model(input_dim):\n  model = keras.Sequential([\n    layers.Dense(64, activation='relu', input_dim=input_dim),\n    layers.Dense(64, activation='relu'),\n    layers.Dense(1)\n  ])\n\n  optimizer = tf.keras.optimizers.RMSprop(0.001)\n\n  model.compile(loss='mse',\n                optimizer=optimizer,\n                metrics=['mae', 'mse'])\n  return model\n<\/code><\/pre>\n\n<p>So you can modify it and use it regardless of the scope of the dataset (suppose you want to build another model with that same function, it wouldn't work since it was linked to just one dataset).<\/p>\n\n<p>As for the performance drop, my bet is that you were including the labels as part of the training dataset (it has happened to me before) and that created a \"perfect\" model. May be the error is on the activation function for the last layer or something like that, but I wouldn't say it has to do with that shape.<\/p>\n\n<p>Edit: I thought it was a Classification task.<\/p>\n"}]}
+{"QuestionId":57760367,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-02T16:42:11.647","AcceptedAnswerId":null,"OwnerUserId":12010850.0,"Title":"How to omit passing through placeholder raw\/column filled with zeros?","Body":"<p>I want to train a data which varies array\/matrix size. Therefore I define placeholder, for example, as follows:<\/p>\n\n<p>x = tf.placeholder(tf.int32, [None, max_input_steps, additional_dimension])<\/p>\n\n<p>y = tf.placeholder(tf.int32, [None, max_output_steps])<\/p>\n\n<p>The \"additional_dimension\" is filled with zeros if the input has dimensionality smaller than max_input_steps. The problem is that I want to set the output of such \"not relevant\" inputs to zero without pushing the data through the network. Are there any solutions to this problem? Could I do something like boolean mask or tf.case? (I found no suitable solution using the latter.)<\/p>\n\n<p>I tried to introduce a \"scaling\" placeholder which I multiplied with the output in order to force not relevant outputs to be zero.<\/p>\n","answers":[]}
+{"QuestionId":57760510,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><keras>","CreationDate":"2019-09-02T16:56:41.930","AcceptedAnswerId":57761783.0,"OwnerUserId":379181.0,"Title":"Converting Array of Lists to Keras Input","Body":"<p>Given an array of the form<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>array([list([21603, 125, 737, 579, 2065, 10399, 1175, 0, 0, 0]),\n       ...\n       list([1896, 3917, 498, 296, 1452, 523, 754, 450, 3795, 341])],\n      dtype=object)\n<\/code><\/pre>\n\n<p>How do you prepare it to be consumed by a Keras model in TensorFlow 2.0 RC0? In its current form it throws the error<\/p>\n\n<blockquote>\n  <p>ValueError: Failed to convert a NumPy array to a Tensor (Unsupported object type list).<\/p>\n<\/blockquote>\n\n<p>and I can't seem to get it in the form I'm familiar with,<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>array([[21603, 125, 737, 579, 2065, 10399, 1175, 0, 0, 0],\n       ...\n       [1896, 3917, 498, 296, 1452, 523, 754, 450, 3795, 341]])\n<\/code><\/pre>\n","answers":[{"AnswerId":"57761783","CreationDate":"2019-09-02T19:19:54.883","ParentId":null,"OwnerUserId":"379181","Title":null,"Body":"<p>If you find yourself in this situation, you can correct your array(s) with the following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>new_array = np.array(list(x for x in old_array))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57761323,"AnswerCount":1,"Tags":"<python-3.x><keras><conv-neural-network><recurrent-neural-network>","CreationDate":"2019-09-02T18:24:05.257","AcceptedAnswerId":57762166.0,"OwnerUserId":7003616.0,"Title":"Is it possible to use the output of a flatten layer of a CNN to be the input of a RNN?","Body":"<p>I'm currently doing my honours research project on online\/dynamic signature verification. I am using the SVC 2004 dataset (Task 2). The aim of my research is to create a CRNN (convolutional recurrent neural network) that can identify if a signature is authentic or forged. Here is the code for the model: (my data preprocessing can be found here: <a href=\"https:\/\/stackoverflow.com\/questions\/57759627\/must-the-input-height-of-a-1d-cnn-be-constant\">Data preprocessing code<\/a><\/p>\n\n<pre><code>class crnn_model:\ndef __init__(self, trainX, trainy, testX, testy, optimizer_method):\n   self.trainX = trainX\n   self.trainy = trainy\n   self.testX = testX\n   self.testy = testy\n\n   self.evaluate_model(optimizer_method)\n\ndef evaluate_model(self, optimizer_method):\n    verbose, epochs, batch_size = 0, 40, 10\n    n_timesteps, n_features, n_outputs = len(self.trainX), 7, 2\n    model = keras.Sequential()\n    model.add(keras.layers.Conv1D(filters=64, kernel_size=3, activation='relu', input_shape=(n_timesteps, n_features), use_bias=True))\n    model.add(keras.layers.Conv1D(filters=64, kernel_size=3, activation='relu'))\n    model.add(keras.layers.Dropout(0.5))\n    model.add(keras.layers.MaxPooling1D(pool_size=2))\n    model.add(keras.layers.Flatten())\n    model.add(keras.layers.LSTM(2, input_shape=[30592,1], return_sequences=True))\n    model.summary()\n\n    # Compile the model\n    model.compile(optimizer=optimizer_method, loss='categorical_crossentropy', metrics=['accuracy'])\n\n    #fit model\n    model.fit(self.trainX, self.trainy, epochs=epochs, batch_size=batch_size, verbose=verbose)\n\n    #evaluate model\n    _, accuracy = model.evaluate(self.testX, self.testy, batch_size=batch_size, verbose=0)\n    return accuracy\n<\/code><\/pre>\n\n<p>This is not my final code, however I come across the following error:<\/p>\n\n<p>ValueError: Input 0 is incompatible with layer lstm_1: expected ndim=3, found ndim=2<\/p>\n\n<p>Much appreciated for your time and any tips on RNNs or CNNs.<\/p>\n","answers":[{"AnswerId":"57762166","CreationDate":"2019-09-02T20:03:57.930","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p><code>Flatten<\/code> converts a tensor of shape <code>(batch_size, timesteps, features)<\/code> to <code>(batch_size, timesteps*features)<\/code> which is why you are getting the error <code>found ndim=2<\/code>. Depending on what you are trying to achieve you might:<\/p>\n\n<ul>\n<li>Remove <code>Flatten<\/code> to pass the convolved learned features into an LSTM, or<\/li>\n<li><code>Reshape<\/code> the flattened tensor to <code>(batch_size, timesteps*features, 1)<\/code> essentially saying every timestep is a single feature.<\/li>\n<\/ul>\n\n<p>In either case the LSTM expects a tensor of rank 3. But wait, just because you reshape doesn't mean it is correct, it all depends on what you are trying to achieve and how the information flow \/ computation graph of the network should look like.<\/p>\n"}]}
+{"QuestionId":57761731,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-02T19:12:25.253","AcceptedAnswerId":null,"OwnerUserId":12003754.0,"Title":"Adding class_weight to .fit_generator() breaks to_categorical()","Body":"<p>Trying to train a CNN with a bunch of images using a DataGenerator class, model works perfectly fine normally. The problem is the training dataset is very skewed to a few classes so I want to add class_weights. However, every time I do this I get an index error in the part of the code that converts my labelled classes into one-hot arrays. <\/p>\n\n<p>This if for Keras running on top of tensorflow. The function that is having the problem is keras.utils.to_categorical()<\/p>\n\n<p>Here's the to catagorical function:<\/p>\n\n<pre><code>for i, pdb_id in enumerate(list_enzymes_temp):\n    mat = precomputed_distance_matrix(pdb_id, self.dim)\n\n    X[i,] = mat.distance_matrix.reshape(*self.dim)\n\n    y[i] = int(self.labels[pdb_id.upper()][1]) - 1\n\n    return X, keras.utils.to_categorical(y, num_classes=self.n_classes)\n<\/code><\/pre>\n\n<p>Here's the function I am using to generate the weights<\/p>\n\n<pre><code>def get_class_weights(dictionary, training_enzymes, mode):\n    'Gets class weights for Keras'\n    # Initialization\n    counter = [0 for i in range(6)]\n\n    # Count classes\n    for enzyme in training_enzymes:\n        counter[int(dictionary[enzyme.upper()][1])-1] += 1\n    majority = max(counter)\n\n    # Make dictionary\n    class_weights = {i: float(majority\/count) for i, count in enumerate(counter)}\n\n    # Value according to mode\n    if mode == 'unbalanced':\n        for key in class_weights:\n            class_weights[key] = 1\n    elif mode == 'balanced':\n        pass\n    elif mode == 'mean_1_balanced':\n        for key in class_weights:\n            class_weights[key] = (1+class_weights[key])\/2\n\n    return class_weights\n<\/code><\/pre>\n\n<p>and my fit_generator function:<\/p>\n\n<pre><code>model.fit_generator(generator=training_generator,\n                validation_data=validation_generator,\n                epochs=max_epochs,\n                max_queue_size=16,\n                class_weight=class_weights,\n                callbacks=[tensorboard])\n<\/code><\/pre>\n\n<p>Heres the IndexError message does not appear and model works perfectly without the class_weights added:<\/p>\n\n<pre><code>File \"C:\\Users\\Python\\DMCNN\\data_generator.py\", line 73, in __getitem__\nX, y = self.__data_generation(list_enzymes_temp)\nFile \"C:\\Users\\Python\\DMCNN\\data_generator.py\", line 59, in __data_generation\nreturn X, keras.utils.to_categorical(y, num_classes=self.n_classes)\nFile \"C:\\Users\\Python\\Anaconda3\\lib\\site-packages\\keras\\utils\\np_utils.py\", line 34, in to_categorical\ncategorical[np.arange(n), y] = 1\nIndexError: index 1065353216 is out of bounds for axis 1 with size 6\n<\/code><\/pre>\n","answers":[{"AnswerId":"57874704","CreationDate":"2019-09-10T16:16:54.050","ParentId":null,"OwnerUserId":"8286869","Title":null,"Body":"<p>I had the same error while using keras.utils.to_categorical.  The error I got is \"IndexError: index 1065353216 is out of bounds for axis 1 with size 2\" because I had 2 classes. <\/p>\n\n<p>I believed it is from converting 1.0 to 1.0f (32 bit float) because 1065353216 is the unsigned 32-bit integer representation of the 32-bit floating point value 1.0 (check here: <a href=\"https:\/\/stackoverflow.com\/questions\/24356579\/why-is-1-0f-in-c-code-represented-as-1065353216-in-the-generated-assembly\/24356601\">Why is 1.0f in C code represented as 1065353216 in the generated assembly?<\/a>).  In my case not all the batches have the same length, which ends up with some empty unfilled in X and y, which causes the issue.  You can check if there are some elements unfilled in your W (or even in X and Y) in advance.  You can also see that keras.utils.to_categorical has the default value dtype='float32'. You can try to specify dtype e.g. \"return X, keras.utils.to_categorical(y, num_classes=self.n_classes, dtype='uint8')\" in your case to see if it works.  <\/p>\n"}]}
+{"QuestionId":57761773,"AnswerCount":0,"Tags":"<conv-neural-network><pytorch>","CreationDate":"2019-09-02T19:18:29.553","AcceptedAnswerId":null,"OwnerUserId":11860162.0,"Title":"CNN Training Runs without error, but it does not display the results","Body":"<p>I am new to pytorch, and I am trying to train my model (CNN), using the following code:<\/p>\n\n<p>The program runs fine, but it does not display this Epoch\/Step\/Loss\/Accuracy part:<\/p>\n\n<pre><code>print(\u2018Epoch [{}\/{}], Step [{}\/{}], Loss: {:.4f}, Accuracy: {:.2f}%\u2019\n<\/code><\/pre>\n\n<p>as <code>if (i+1) % 100 == 0:<\/code> never turns to 0<\/p>\n\n<p>Training part: <\/p>\n\n<pre><code>iter = 0\n\nfor epoch in range(num_epochs):\n    for i, (images, labels) in enumerate(dataloaders['train']):\n        images = Variable(images)\n        labels = Variable(labels)\n\n        # Clear the gradients\n        optimizer.zero_grad()\n\n        # Forward propagation\n        outputs = model(images)\n\n        # Calculating loss with softmax to obtain cross entropy loss\n        loss = criterion(outputs, labels)\n\n        # Backward prop\n        loss.backward()\n\n        # Updating gradients\n        optimizer.step()\n\n        iter += 1\n\n        # Total number of labels\n        total = labels.size(0)\n\n\n        # Obtaining predictions from max value\n        _, predicted = torch.max(outputs.data, 1)\n\n\n\n        # Calculate the number of correct answers\n        correct = (predicted == labels).sum().item()\n\n        # Print loss and accuracy\n\n        if (i+1) % 100 == 0:\n            print('Epoch [{}\/{}], Step [{}\/{}], Loss: {:.4f}, Accuracy: {:.2f}%'\n                  .format(epoch + 1, num_epochs, i + 1, len(dataloaders['train']), loss.item(),\n                          (correct \/ total) * 100))\n<\/code><\/pre>\n\n<p>Full Code: <\/p>\n\n<p><a href=\"https:\/\/pastebin.com\/dshNmhRL\" rel=\"nofollow noreferrer\">https:\/\/pastebin.com\/dshNmhRL<\/a><\/p>\n","answers":[]}
+{"QuestionId":57761862,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-02T19:29:23.897","AcceptedAnswerId":null,"OwnerUserId":4820436.0,"Title":"Confused about the parameter type to DenseFeatures.__call__","Body":"<p>I am following <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/keras\/feature_columns\" rel=\"nofollow noreferrer\">this Tensorflow tutorial<\/a><\/p>\n\n<p>In the <code>demo<\/code> function, I'm confused about the <code>example_batch<\/code> parameter on the line <code>print(feature_layer(example_batch).numpy())<\/code>.  <code>example_batch<\/code> is a <code>dict<\/code>.  However, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/Layer#__call__\" rel=\"nofollow noreferrer\">based on Tensorflow's documentation<\/a>, I'm supposed to be passing a tensor as the argument to <code>Layer.__call__<\/code> (not a <code>dict<\/code>).<\/p>\n\n<p><code>feature_layer<\/code> is a <code>DenseFeatures<\/code> object.  Looking at <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/layers\/DenseFeatures\" rel=\"nofollow noreferrer\">Tensorflow's documentation<\/a>, I don't see a <code>__call__<\/code> method for <code>DenseFeatures<\/code>, but I'm guessing that <code>DenseFeatures<\/code> is a <code>Layer<\/code>, so that's why I'm looking at <code>Layer.__call__<\/code>.  Is that correct?<\/p>\n\n<p>Based on the documentation (without digging into Tensorflow's code), how would I know that I can pass a <code>dict<\/code> object to <code>DenseFeatures.__call__<\/code>?  Are <code>dict<\/code> objects able to be treated as tensors?<\/p>\n\n<pre><code>import pandas as pd\nimport tensorflow as tf\n\nfrom tensorflow import feature_column\nfrom tensorflow.keras import layers\nfrom sklearn.model_selection import train_test_split\n\nURL = 'https:\/\/storage.googleapis.com\/applied-dl\/heart.csv'\ndataframe = pd.read_csv(URL)\n\ntrain, test = train_test_split(dataframe, test_size=0.2)\n\n# A utility method to create a tf.data dataset from a Pandas Dataframe\ndef df_to_dataset(dataframe, shuffle=True, batch_size=32):\n    dataframe = dataframe.copy()\n    labels = dataframe.pop('target')\n    ds = tf.data.Dataset.from_tensor_slices((dict(dataframe), labels))\n    if shuffle:\n        ds = ds.shuffle(buffer_size=len(dataframe))\n    ds = ds.batch(batch_size)\n    return ds\n\ntrain_ds = df_to_dataset(train, batch_size=5)\n\nexample_batch = next(iter(train_ds))[0]    \n\n# A utility method to create a feature column\n# and to transform a batch of data\ndef demo(feature_column):\n    print('type(example_batch): ', type(example_batch))    \n    feature_layer = layers.DenseFeatures(feature_column)\n    print(feature_layer(example_batch).numpy())\n\n# Categorical columns\nthal = feature_column.categorical_column_with_vocabulary_list(\n      'thal', ['fixed', 'normal', 'reversible'])\nthal_one_hot = feature_column.indicator_column(thal)\ndemo(thal_one_hot)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57762183,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-02T20:05:39.847","AcceptedAnswerId":null,"OwnerUserId":9790914.0,"Title":"Custom loss for \"weighted\" rmse","Body":"<p>I'm doing a regression with Keras\/TF\nThe values to be predicted can range from -1 to 1.\nIn my use case, error on sign are a big problem, so I want to predict:<\/p>\n\n<ul>\n<li><p>Correct value<\/p><\/li>\n<li><p>But above, correct sign<\/p><\/li>\n<\/ul>\n\n<p>For example if I have to predict +0.01, predicting +0.1 is less a problem than predicting -0.01. \nSame if I have to predict -0.01, predicting -0.1 is less a problem than predicting +0.01<\/p>\n\n<p>I though I will use a custom loss to punish more error sign.\nSo from a rmse like function, I have added coefficient to sign errors like this:<\/p>\n\n<pre><code>def loss_rmse2(y_true, y_pred):\n    z = K.zeros_like(y_true)\n    tp = K.switch( tf.logical_and( K.greater(y_true, z), K.greater(y_pred, z)) , K.square(y_true-y_pred), z)\n    fp = K.switch( tf.logical_and( K.less(y_true, z), K.greater(y_pred, z)) , K.square(y_true-y_pred), z)\n    tn = K.switch( tf.logical_and( K.less(y_true, z), K.less(y_pred, z)) , K.square(y_true-y_pred), z)\n    fn = K.switch( tf.logical_and( K.greater(y_true, z), K.less(y_pred, z)) , K.square(y_true-y_pred), z)\n    return K.sqrt(K.mean(tp) + 2 * K.mean(fp) + K.mean(tn) + 2 * K.mean(fn) )\n<\/code><\/pre>\n\n<p>The problem is that my model does not converge to a solution.<\/p>\n\n<p>Any idea how to handle that?<\/p>\n\n<p>Thanks<\/p>\n\n<p>Loss function:\n<a href=\"https:\/\/i.stack.imgur.com\/YKoEu.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/YKoEu.png\" alt=\"Loss\"><\/a><\/p>\n","answers":[{"AnswerId":"57762686","CreationDate":"2019-09-02T21:09:03.640","ParentId":null,"OwnerUserId":"9683408","Title":null,"Body":"<p>If you are doing binary classification, I would recommend normalizing to between 0 and 1 (add 1, then divide by 2), and then using <code>keras.losses.binary_crossentropy(y_true, y_pred)<\/code><\/p>\n"}]}
+{"QuestionId":57762320,"AnswerCount":0,"Tags":"<tensorflow><gradient><reduce>","CreationDate":"2019-09-02T20:20:55.203","AcceptedAnswerId":null,"OwnerUserId":11979471.0,"Title":"How to conditionally reduce a tensor","Body":"<p>I have a custom gradient which calculates the inner product of two tensors: vector_conj_inner(a,b) over the last dimension. Tensor 'b' has a broadcasting axis, hence the gradient of b has to be reduced to 1 on that axis.<\/p>\n\n<p>To provide an example, take the following two tensors a,b. To make clear what happens in the forward and backward paths, I wrote the respective shapes in the comments.<\/p>\n\n<pre><code>\na.shape = (3,5,7)\nb.shape = (5,7)\n\nb = tf.expand_dims(b, axis=0)                      # shape = (1,5,7)\ns = vector_conj_inner(a,b)\n\ndef vector_conj_inner(a, b):\n    s = tf.reduce_sum(a*tf.conj(b), -1)            # shape = (3,5)\n\n    def grad(grad_s):\n        grad_s = tf.expand_dims(grad_s, axis=-1)   # shape = (3,5,1)\n        grad_a = grad_s*b                          # shape = (3,5,7)\n        grad_b = tf.conj(grad_s)*a                 # shape = (3,5,7)\n        grad_b = tf.reduce_sum(grad_b, axis=0, keepdims=True)   # shape = (1,5,7)\n        return grad_a, grad_b\n\n    return s, grad\n<\/code><\/pre>\n\n<p>You can see that the shapes work out when I reduce axis 0 of \"grad_b\" back to 1.<\/p>\n\n<p>My question now is, how can I make that more general, i.e. detecting automatically which axis in \"b\" is 1, and reduce \"grad_b\" according to that information? What I need for that is something like a conditional reduce_sum, where the condition is which axis of b is 1.<\/p>\n\n<p>Any ideas?<\/p>\n","answers":[]}
+{"QuestionId":57762326,"AnswerCount":1,"Tags":"<tensorflow><tensor>","CreationDate":"2019-09-02T20:21:30.917","AcceptedAnswerId":null,"OwnerUserId":4830432.0,"Title":"Cannot feed value of shape: Misunderstanding Tensor Dimensions","Body":"<p>I've been walking through some tensorflow tutorials and am cobbling together a pet experiment. However, I am running into some dimension errors and I can seem to figure them out.<\/p>\n\n<p>My goal: I have an input matrix for the shape 1xN. I have a training set of dimension 10xN. (1 and 10 were chosen arbitrarily). N is intended to represent N samples in a training set: 1 input value mapped to one vector of outputs. You can think of this as 1 input neuron and m output neurons. The training set is a set of these single values mapped to a 1d vector. I wish to train the network by running the set of these mapped inputs and outputs against it and reducing the error.  <\/p>\n\n<p>The simple algorithm that I am trying to accomplish:<\/p>\n\n<ol>\n<li>For each value in the input vector\n\n<ol start=\"2\">\n<li>Load the input neuron with that value<\/li>\n<li>Feed forward<\/li>\n<li>Evaluate against the corresponding vector<\/li>\n<\/ol><\/li>\n<\/ol>\n\n<p>Repeat to minimize error.<\/p>\n\n<p>However, I seem to be getting mixed up with how to format the data to feed to the network. I have a placeholder of 1 input neurons and one of n output neurons. I want to follow the above algorithm but I am not sure if I am doing it right:<\/p>\n\n<pre><code># Data parameters\n\nnum_frames = 10\n\nstimuli_value_low = .00001\nstimuli_value_high = 100\n\npixel_value_low = .00001\npixel_value_high = 256.0\n\nstimuli_dimension = 1\nframe_dimension = 10\n\nstimuli = np.random.uniform(stimuli_value_low, stimuli_value_high, (stimuli_dimension, num_frames))\nframes = np.random.uniform(pixel_value_low, pixel_value_high, (frame_dimension, num_frames))\n\n# Parameters\nlearning_rate = 0.01\ntraining_iterations = 1000\ndisplay_iteration = 10\n\n# Network Parameters\nn_hidden_1 = 100\nn_hidden_2 = 100\nnum_input_neurons = stimuli_dimension\nnum_output_neurons = frame_dimension\n\n# Create placeholders\ninput_placeholder = tf.placeholder(\"float\", [None, num_input_neurons])\noutput_placeholder = tf.placeholder(\"float\", [None, num_output_neurons])\n\n# Store layers weight &amp; bias\nweights = {\n    'h1': tf.Variable(tf.random_normal([num_input_neurons, n_hidden_1])),\n    'h2': tf.Variable(tf.random_normal([n_hidden_1, n_hidden_2])),\n    'out': tf.Variable(tf.random_normal([n_hidden_2, num_output_neurons]))\n}\nbiases = {\n    'b1': tf.Variable(tf.random_normal([n_hidden_1])),\n    'b2': tf.Variable(tf.random_normal([n_hidden_2])),\n    'out': tf.Variable(tf.random_normal([num_output_neurons]))\n}\n\n# Create model\ndef neural_net(input_placeholder):\n    # Hidden fully connected layer \n    layer_1 = tf.add(tf.matmul(input_placeholder, weights['h1']), biases['b1'])\n    # Hidden fully connected layer\n    layer_2 = tf.add(tf.matmul(layer_1, weights['h2']), biases['b2'])\n    # Output fully connected layer with a neuron for each pixel\n    out_layer = tf.matmul(layer_2, weights['out']) + biases['out']\n    return out_layer\n\n# Construct model\nlogits = neural_net(input_placeholder)\n\n# Define loss operation and optimizer\nloss_operation = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits = logits, labels = output_placeholder))\noptimizer = tf.train.AdamOptimizer(learning_rate = learning_rate)\ntrain_operation = optimizer.minimize(loss_operation)\n\n# Evaluate model (with test logits, for dropout to be disabled)\ncorrect_pred = tf.equal(tf.argmax(logits, 1), tf.argmax(output_placeholder, 1))\naccuracy_operation = 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, training_iterations + 1): \n\n    sess.run(train_operation, feed_dict = {X: stimuli, Y: frames})\n\n    if iteration % display_iteration == 0 or iteration == 1:\n\n      loss, accuracy = sess.run([loss_operation, accuracy_operation], feed_dict = {X: stimuli, Y: frames})\n\n      print(\"Step \" + str(iteration) + \n            \", Loss = \" + \"{:.4f}\".format(loss) + \n            \", Training Accuracy= \" + \\\n            \"{:.3f}\".format(acc))\n\n  print(\"Optimization finished!\")\n<\/code><\/pre>\n\n<p>I think it is something to do with how I am structuring my data or feeding it to the run function.<\/p>\n\n<p>Here is the error I am getting:<\/p>\n\n<pre><code>ValueError                                Traceback (most recent call last)\n\n&lt;ipython-input-420-7517598734d6&gt; in &lt;module&gt;()\n      6   for step in range(1, training_iterations + 1):\n      7 \n----&gt; 8     sess.run(train_operation, feed_dict = {X: stimuli, Y: frames})\n      9 \n     10     if iteration % display_iteration == 0 or iteration == 1:\n\n1 frames\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   1147                              'which has shape %r' %\n   1148                              (np_val.shape, subfeed_t.name,\n-&gt; 1149                               str(subfeed_t.get_shape())))\n   1150           if not self.graph.is_feedable(subfeed_t):\n   1151             raise ValueError('Tensor %s may not be fed.' % subfeed_t)\n\nValueError: Cannot feed value of shape (1, 10) for Tensor 'Placeholder_6:0', which has shape '(?, 1)'\n<\/code><\/pre>\n\n<p>How can I ensure I am formatting my input data correctly and forming my network corresponingly?<\/p>\n","answers":[{"AnswerId":"57764246","CreationDate":"2019-09-03T02:10:49.387","ParentId":null,"OwnerUserId":"4830432","Title":null,"Body":"<p>Turns out I had the dimensions of the arrays I was generating backwards:<\/p>\n\n<pre><code>stimuli = np.random.uniform(stimuli_value_low, stimuli_value_high, (stimuli_dimension, num_frames))\nframes = np.random.uniform(pixel_value_low, pixel_value_high, (frame_dimension, num_frames))\n<\/code><\/pre>\n\n<p>should be:<\/p>\n\n<pre><code>stimuli = np.random.uniform(stimuli_value_low, stimuli_value_high, (num_frames, stimuli_dimension))\nframes = np.random.uniform(pixel_value_low, pixel_value_high, (num_frames, frame_dimension))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57762493,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-02T20:39:58.067","AcceptedAnswerId":null,"OwnerUserId":10681422.0,"Title":"Saving tensor to file in each training step","Body":"<p>I am trying to learn TensorFlow to implement a DCGAN model from <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/generative\/dcgan\" rel=\"nofollow noreferrer\">this tutorial<\/a>. I followed the steps in the tutorial and edited the generator and discriminator models' layers to satisfy my architecture.<\/p>\n\n<p>I have <code>n<\/code> epochs n my training process, calling <code>train_step<\/code> in each epoch. I want to be able to either:<\/p>\n\n<ul>\n<li>Save the generated data by <code>generated_data = generator(noise, training=True)<\/code> to a file as numerical values, or<\/li>\n<li>Find a numerical value with <code>tf.losses.mean_squared_error(data, generated_data)<\/code> and save it to a file or append it into an array<\/li>\n<\/ul>\n\n<p>I tried creating a session and then running the session to get the numerical representation of the MSE, but could not do it.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>@tf.function\ndef train_step(data):\n    noise = tf.random.normal(noise_dim)\n\n    with tf.GradientTape() as gen_tape, tf.GradientTape() as disc_tape:\n        generated_data = generator(noise, training=True)\n\n        real_output = discriminator(data, training=True)\n        fake_output = discriminator(generated_data, training=True)\n\n        utility_losses.append(tf.losses.mean_squared_error(data, generated_data))\n\n        gen_loss = generator_loss(fake_output)\n        disc_loss = discriminator_loss(real_output, fake_output)\n\n    gradients_of_generator = gen_tape.gradient(gen_loss, generator.trainable_variables)\n    gradients_of_discriminator = disc_tape.gradient(disc_loss, discriminator.trainable_variables)\n\n    generator_optimizer.apply_gradients(zip(gradients_of_generator, generator.trainable_variables))\n    discriminator_optimizer.apply_gradients(zip(gradients_of_discriminator, discriminator.trainable_variables))\n<\/code><\/pre>\n\n<p>Here is the session part:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>sess = tf.Session()\nprint(sess.run(utility_losses[0]))\n<\/code><\/pre>\n\n<p>This failed with a long error trace containing <code>FailedPreconditionError<\/code><\/p>\n\n<p>I want to be able to save generated data into a file in each epoch, but I could not find how to do it in TensorFlow. I have been searching for hours to find a solution for this, but could not come up with one.<\/p>\n","answers":[]}
+{"QuestionId":57762851,"AnswerCount":0,"Tags":"<machine-learning><keras><neural-network><regression>","CreationDate":"2019-09-02T21:31:01.627","AcceptedAnswerId":null,"OwnerUserId":8598043.0,"Title":"How to add additional info to a existent keras NN model in order to improve accuracy","Body":"<p>So, I'll briefly describe the domain of the problem and I state that I am a beginner in neural networks and keras in general so the answer might be trivial for you.\nI have a dataset composed by the following rows:<\/p>\n\n<pre><code>v0;v1;v2;v3;err\n27;45;35;40;9.421857761389568e-17\n42;23;4;10;0.15046260426547653\n7;52;45;4;0.004021030533106114\n8;13;21;8;3.527736174640165e-05\n50;40;3;39;0.2594155190453719\n..........\n..........\n<\/code><\/pre>\n\n<p>I want to predict the errors given the variable values. As you can imagine, this is a fairly easy regression problem (not linear in my case) that I solved with different NN models (I also preprocessed data with normalization for the variables and standardization for the error first). In fact, I trained different models with different dense layers and I found out that the best one gave me an accuracy of 96.5% more or less.<\/p>\n\n<p>Supposed this is the best result I can achieve with keras NN models (maybe not but it does not matter), I have an additional info about the domain that I want to use to increase the accuracy. In fact, the error value depends in some way also on some relations between the different <code>v<\/code> values. I know that\nif <code>v0&lt;v3 and v1&lt;v3 and v3&lt;v2<\/code>, the error is the lowest possible, while if some of these relations are violated, the error increases in some way. \nSo, is there a way to let the NN model also learn those relationships and the consequent decreases or increases of the error if violations are absent or present? I read some papers and some stuff online and I found out that maybe I can represent these dependencies with a graph and then put a GNN layer (or layers) in the current NN model to try to improve the final result. This is just an idea, I don't know if this is possible to do and if it is the best thing to do. Any ideas?<\/p>\n\n<p>Thank you to anyone that can help me, really appreciated!<\/p>\n","answers":[]}
+{"QuestionId":57763066,"AnswerCount":2,"Tags":"<python><vectorization><pytorch>","CreationDate":"2019-09-02T22:02:39.593","AcceptedAnswerId":57771557.0,"OwnerUserId":11948285.0,"Title":"How to vectorize this pytorch code over (at least) the batch dimension?","Body":"<p>I want to implement a code to build an adjacency matrix such that (for example):<\/p>\n\n<p>If X[0] : [0, 1, 2, 0, 1, 0], then,<\/p>\n\n<pre><code>A[0, 1] = 1\nA[1, 2] = 1\nA[2, 0] = 1\nA[0, 1] = 1\nA[1, 0] = 1\n<\/code><\/pre>\n\n<p>The following code works fine, however, it's too slow! So, please help me to vectorize this code on the batch (first) dimension at least:<\/p>\n\n<pre><code>A = torch.zeros((3, 3, 3), dtype = torch.float)\nX = torch.tensor([[0, 1, 2, 0, 1, 0], [1, 0, 0, 2, 1, 1], [0, 0, 2, 2, 1, 1]])\nfor a, x in zip(A, X):\n    for i, j in zip(x, x[1:]):\n        a[i, j] = 1\n<\/code><\/pre>\n\n<p>Thanks! :)<\/p>\n","answers":[{"AnswerId":"57768537","CreationDate":"2019-09-03T09:13:56.203","ParentId":null,"OwnerUserId":"319584","Title":null,"Body":"<p>Make <code>X<\/code> a tuple instead of a tensor:<\/p>\n\n<pre><code>A = torch.zeros((3, 3, 3), dtype = torch.float)\nX = ([0, 1, 2, 0, 1, 0], [1, 0, 0, 2, 1, 1], [0, 0, 2, 2, 1, 1])\nA[X] = 1\n<\/code><\/pre>\n\n<p>For example, by casting it like this: <code>A[tuple(X)]<\/code><\/p>\n"},{"AnswerId":"57771557","CreationDate":"2019-09-03T12:14:33.823","ParentId":null,"OwnerUserId":"3607203","Title":null,"Body":"<p>I am pretty sure that there is a much simpler way of doing this, but I tried to keep within the realm of <code>torch<\/code> function calls, to make sure that any gradient operation could be properly tracked.<\/p>\n\n<p>In case this is not required for backpropagation, I strongly suggest you look into solution that maybe utilize some numpy functions, because I think there is a stronger guarantee to find something suitable here. But, without further ado, here is the solution I came up with.<\/p>\n\n<p>It essentially transforms your <code>X<\/code> vector into a series of tuple entries that correspond to the position in <code>A<\/code>. For this, we need to align some of the indices (specifically, the first dimension is only implicitly given in <code>X<\/code>, since the first list in <code>X<\/code> corresponds to <code>A[0,:,:]<\/code>, the second list to <code>A[1,:,:]<\/code>, and so on.\nThis is also probably where you can start optimizing the code, because I did not find a reasonable description of such a matrix, and therefore had to come up with my own way of creating it.<\/p>\n\n<pre><code># Start by \"aligning\" your shifted view of X\n# Essentially, take the all but the last element, \n# and put it on top of all but the first element.\nX_shift = torch.stack([X[:,:-1], X[:,1:]], dim=2)\n# X_shift.shape: (3,5,2) in your example\n\n# To assign this properly, we need to turn it into a \"concatenated\" list,\n# where each entry corresponds to a 2D tuple in the respective dimension of A.\ntemp_tuples = X_shift.view(-1,2).transpose(0,1)\n# temp_tuples.shape: (2,15) in your example. Below are the values:\ntensor([[0, 1, 2, 0, 1, 1, 0, 0, 2, 1, 0, 0, 2, 2, 1],\n        [1, 2, 0, 1, 0, 0, 0, 2, 1, 1, 0, 2, 2, 1, 1]])\n\n# Now we have to create a matrix do indicate the proper \"first dimension index\"\nfix_dims = torch.repeat_interleave(torch.arange(0,3,1), len(X[0])-1, 0).unsqueeze(dim=0)\n# fix_dims.shape: (1,15)\n# Long story short, this creates the following vector.\ntensor([[0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2]])\n# Note that the unsqueeze is necessary to properly concatenate the two matrices:\naccess_tuples = tuple(torch.cat([fix_dims, temp_tuples], dim=0))\nA[access_tuples] = 1\n<\/code><\/pre>\n\n<p>This further assumes that every dimension in X has the same number of tuples changed. If that is not the case, then you have to manually create a <code>fix_dims<\/code> vector, where each increment is repeated the length of <code>X[i]<\/code> times. If it is equal as in your example, you can safely use the proposed solution.<\/p>\n"}]}
+{"QuestionId":57763326,"AnswerCount":0,"Tags":"<python><tensorflow><keras><pytorch><keras-layer>","CreationDate":"2019-09-02T22:54:45.590","AcceptedAnswerId":null,"OwnerUserId":7131919.0,"Title":"updating the input layer of a neural networks in python","Body":"<p>I have a network with two input layers, the first one is an image data ( e.g. MNIST) and the second one is an auxiliary vector,<\/p>\n\n<pre><code>A_vec = np.random.normal(mu, sigma, (1, n_d))\nA_vec1=np.tile(A_vec, (size_1, 1))\n\nimage_inputs = Input(shape=(original_dim,),name='Image_Inputs')\nA_inputs = Input(shape=(n_d,),name='A_Inputs')\ninputs = keras.layers.concatenate([image_inputs, A_inputs],name='Inputs')\n<\/code><\/pre>\n\n<h3>.......intermediate layers......<\/h3>\n\n<pre><code>Image_Outputs= Dense(\noriginal_dim,\nactivation='sigmoid',\nname='Image_Outputs')(intermediate layers)\n\nA_Outputs= Dense(\n    n_d,\n    activation='sigmoid',\n    name='A_Outputs')(intermediate layers)\n\noutputs = keras.layers.concatenate([Image_Outputs, A_Outputs],name='Outputs')\n<\/code><\/pre>\n\n<p>Here, the <code>A_vec<\/code>in the initialization for <code>A_inputs<\/code>. But, I want to update it in each epoch by using the <code>A_Outputs<\/code>. I mean after the first iteration, I need to use the estimated values for <code>A_Outputs<\/code> as the updated values of the<code>A_inputs<\/code> in the second epochs and same prcedure for next epochs. Is there any way to do that in <em>Tensorflow<\/em>, <em>Keras<\/em> or <em>pytorch<\/em>?<\/p>\n","answers":[]}
+{"QuestionId":57763363,"AnswerCount":1,"Tags":"<python><machine-learning><keras><scikit-learn><classification>","CreationDate":"2019-09-02T23:01:14.997","AcceptedAnswerId":57772102.0,"OwnerUserId":10226602.0,"Title":"Validation accuracy metrics reported by Keras model.fit log and Sklearn.metrics.confusion_matrix don't match each other","Body":"<p>The problem is that the reported <code>validation accuracy<\/code> value I get from Keras <code>model.fit<\/code> history is significantly higher than the <code>validation accuracy<\/code> metric I get from <code>sklearn.metrics<\/code> functions.<\/p>\n\n<p>The results I get from <code>model.fit<\/code> are summarized below:<\/p>\n\n<pre><code>Last Validation Accuracy: 0.81\nBest Validation Accuracy: 0.84\n<\/code><\/pre>\n\n<p>The results (normalized) from <code>sklearn<\/code> are pretty different:<\/p>\n\n<pre><code>True Negatives: 0.78\nTrue Positives: 0.77\n\nValidation Accuracy = (TP + TN) \/ (TP + TN + FP + FN) = 0.775 \n\n(see confusion matrix below for reference)\n\nEdit: this calculation is incorrect, because one can not \nuse the normalized values to calculate the accuracy, since \nit does not account for differences in the total absolute \nnumber of points in the dataset. Thanks to the comment by desertnaut\n<\/code><\/pre>\n\n<ul>\n<li><p>Here is the graph of the validation accuracy data from model.fit history:\n<a href=\"https:\/\/i.stack.imgur.com\/Hxmgg.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Hxmgg.png\" alt=\"Validation accuracy from model.fit data history\"><\/a><\/p><\/li>\n<li><p>And here is the Confusion matrix generated from sklearn:<\/p><\/li>\n<\/ul>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ElSXB.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ElSXB.png\" alt=\"Confusion matrix from sklearn\"><\/a><\/p>\n\n<p>I think this question is somewhat similar as this one <a href=\"https:\/\/stackoverflow.com\/questions\/54580679\/sklearn-metrics-values-are-very-different-from-keras-values\">Sklearn metrics values are very different from Keras values<\/a> \nBut I've checked both methods are doing the validation on the same pool of data, so that answer is probably not adequate for my case.<\/p>\n\n<p>Also, this question <a href=\"https:\/\/stackoverflow.com\/questions\/46354182\/keras-binary-accuracy-metric-gives-too-high-accuracy\">Keras binary accuracy metric gives too high accuracy<\/a> seems to address some problems with the way that binary cross entropy affects a multiclass problem, but in my case it may not apply, since it is a true binary classification problem.<\/p>\n\n<p>Here are the commands used:<\/p>\n\n<p>Model definition:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>inputs = Input((Tx, ))\nn_e = 30\nembeddings = Embedding(n_x, n_e, input_length=Tx)(inputs)\nout = Bidirectional(LSTM(32, recurrent_dropout=0.5, return_sequences=True))(embeddings)\nout = Bidirectional(LSTM(16, recurrent_dropout=0.5, return_sequences=True))(out)\nout = Bidirectional(LSTM(16, recurrent_dropout=0.5))(out)\nout = Dense(3, activation='softmax')(out)\nmodelo = Model(inputs=inputs, outputs=out)\nmodelo.summary()\n<\/code><\/pre>\n\n<p>Model Summary:<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_1 (InputLayer)         (None, 100)               0         \n_________________________________________________________________\nembedding (Embedding)        (None, 100, 30)           86610     \n_________________________________________________________________\nbidirectional (Bidirectional (None, 100, 64)           16128     \n_________________________________________________________________\nbidirectional_1 (Bidirection (None, 100, 32)           10368     \n_________________________________________________________________\nbidirectional_2 (Bidirection (None, 32)                6272      \n_________________________________________________________________\ndense (Dense)                (None, 3)                 99        \n=================================================================\nTotal params: 119,477\nTrainable params: 119,477\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>Model compilation:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>mymodel.compile(optimizer='adam', loss='binary_crossentropy', metrics=['acc'])\n<\/code><\/pre>\n\n<p>Model fit call:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>num_epochs = 30\nmyhistory = mymodel.fit(X_pad, y, epochs=num_epochs, batch_size=50, validation_data=[X_val_pad, y_val_oh], shuffle=True, callbacks=callbacks_list)\n<\/code><\/pre>\n\n<p>Model fit log:<\/p>\n\n<pre><code>Train on 505 samples, validate on 127 samples\n\nEpoch 1\/30\n500\/505 [============================&gt;.] - ETA: 0s - loss: 0.6135 - acc: 0.6667\n[...]\nEpoch 10\/30\n500\/505 [============================&gt;.] - ETA: 0s - loss: 0.1403 - acc: 0.9633\nEpoch 00010: val_acc improved from 0.77953 to 0.79528, saving model to modelo-10-melhor-modelo.hdf5\n505\/505 [==============================] - 21s 41ms\/sample - loss: 0.1393 - acc: 0.9637 - val_loss: 0.5203 - val_acc: 0.7953\nEpoch 11\/30\n500\/505 [============================&gt;.] - ETA: 0s - loss: 0.0865 - acc: 0.9840\nEpoch 00011: val_acc did not improve from 0.79528\n505\/505 [==============================] - 21s 41ms\/sample - loss: 0.0860 - acc: 0.9842 - val_loss: 0.5257 - val_acc: 0.7953\nEpoch 12\/30\n500\/505 [============================&gt;.] - ETA: 0s - loss: 0.0618 - acc: 0.9900\nEpoch 00012: val_acc improved from 0.79528 to 0.81102, saving model to modelo-10-melhor-modelo.hdf5\n505\/505 [==============================] - 21s 42ms\/sample - loss: 0.0615 - acc: 0.9901 - val_loss: 0.5472 - val_acc: 0.8110\nEpoch 13\/30\n500\/505 [============================&gt;.] - ETA: 0s - loss: 0.0415 - acc: 0.9940\nEpoch 00013: val_acc improved from 0.81102 to 0.82152, saving model to modelo-10-melhor-modelo.hdf5\n505\/505 [==============================] - 21s 42ms\/sample - loss: 0.0413 - acc: 0.9941 - val_loss: 0.5853 - val_acc: 0.8215\nEpoch 14\/30\n500\/505 [============================&gt;.] - ETA: 0s - loss: 0.0443 - acc: 0.9933\nEpoch 00014: val_acc did not improve from 0.82152\n505\/505 [==============================] - 21s 42ms\/sample - loss: 0.0453 - acc: 0.9921 - val_loss: 0.6043 - val_acc: 0.8136\nEpoch 15\/30\n500\/505 [============================&gt;.] - ETA: 0s - loss: 0.0360 - acc: 0.9933\nEpoch 00015: val_acc improved from 0.82152 to 0.84777, saving model to modelo-10-melhor-modelo.hdf5\n505\/505 [==============================] - 21s 42ms\/sample - loss: 0.0359 - acc: 0.9934 - val_loss: 0.5663 - val_acc: 0.8478\n[...]\nEpoch 30\/30\n500\/505 [============================&gt;.] - ETA: 0s - loss: 0.0039 - acc: 1.0000\nEpoch 00030: val_acc did not improve from 0.84777\n505\/505 [==============================] - 20s 41ms\/sample - loss: 0.0039 - acc: 1.0000 - val_loss: 0.8340 - val_acc: 0.8110\n<\/code><\/pre>\n\n<p>Confusion matrix from sklearn:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from sklearn.metrics import confusion_matrix\nconf_mat = confusion_matrix(y_values, predicted_values)\n<\/code><\/pre>\n\n<p>The prediction values and gold values are determined as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>preds = mymodel.predict(X_val)\npreds_ints = [[el] for el in np.argmax(preds, axis=1)]\nvalues_pred = tokenizer_y.sequences_to_texts(preds_ints)\nvalues_gold = tokenizer_y.sequences_to_texts(y_val)\n<\/code><\/pre>\n\n<p>Finally, I'd like to add that I have printed out the data and all prediction errors and I believe the sklearn values are more reliable, since they seem to match the results I get from printing out the predictions for the saved \"best\" model. <\/p>\n\n<p>On the other hand, I can't understand how the metrics can be so different. Since they are both very well know softwares, I conclude I'm the one making the mistake here, but I can't pin down where or how.<\/p>\n","answers":[{"AnswerId":"57772102","CreationDate":"2019-09-03T12:48:05.200","ParentId":null,"OwnerUserId":"4685471","Title":null,"Body":"<p>Your question is ill-posed; as already commented, you have not computed the actual accuracy of your scikit-learn model, hence you seem to compare apples with oranges. The computation (TP + TN)\/2 from a normalized confusion matrix does <strong>not<\/strong> give the accuracy. Here is a simple deomonstration using toy data, adapting the <code>plot_confusion_matrix<\/code> from the <a href=\"https:\/\/scikit-learn.org\/stable\/auto_examples\/model_selection\/plot_confusion_matrix.html\" rel=\"nofollow noreferrer\">docs<\/a>:<\/p>\n\n<pre><code>import numpy as np\nimport matplotlib.pyplot as plt\nfrom sklearn.metrics import confusion_matrix\n\n# toy data\ny_true = [0, 1, 0, 1, 0, 0, 0, 1]\ny_pred =  [1, 1, 1, 0, 1, 1, 0, 1]\nclass_names=[0,1]\n\n# plot_confusion_matrix function\n\ndef plot_confusion_matrix(y_true, y_pred, classes,\n                          normalize=False,\n                          title=None,\n                          cmap=plt.cm.Blues):\n    \"\"\"\n    This function prints and plots the confusion matrix.\n    Normalization can be applied by setting `normalize=True`.\n    \"\"\"\n    if not title:\n        if normalize:\n            title = 'Normalized confusion matrix'\n        else:\n            title = 'Confusion matrix, without normalization'\n\n    # Compute confusion matrix\n    cm = confusion_matrix(y_true, y_pred)\n\n    if normalize:\n        cm = cm.astype('float') \/ cm.sum(axis=1)[:, np.newaxis]\n        print(\"Normalized confusion matrix\")\n    else:\n        print('Confusion matrix, without normalization')\n\n    print(cm)\n\n    fig, ax = plt.subplots()\n    im = ax.imshow(cm, interpolation='nearest', cmap=cmap)\n    ax.figure.colorbar(im, ax=ax)\n    # We want to show all ticks...\n    ax.set(xticks=np.arange(cm.shape[1]),\n           yticks=np.arange(cm.shape[0]),\n           # ... and label them with the respective list entries\n           xticklabels=classes, yticklabels=classes,\n           title=title,\n           ylabel='True label',\n           xlabel='Predicted label')\n\n    # Rotate the tick labels and set their alignment.\n    plt.setp(ax.get_xticklabels(), rotation=45, ha=\"right\",\n             rotation_mode=\"anchor\")\n\n    # Loop over data dimensions and create text annotations.\n    fmt = '.2f' if normalize else 'd'\n    thresh = cm.max() \/ 2.\n    for i in range(cm.shape[0]):\n        for j in range(cm.shape[1]):\n            ax.text(j, i, format(cm[i, j], fmt),\n                    ha=\"center\", va=\"center\",\n                    color=\"white\" if cm[i, j] &gt; thresh else \"black\")\n    fig.tight_layout()\n    return ax\n<\/code><\/pre>\n\n<p>Computing the <strong>normalized<\/strong> confusion matrix gives:<\/p>\n\n<pre><code>plot_confusion_matrix(y_true, y_pred, classes=class_names, normalize=True)\n# result:\nNormalized confusion matrix\n[[ 0.2         0.8       ]\n [ 0.33333333  0.66666667]]\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Af9TD.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Af9TD.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>and according to your <strong>incorrect<\/strong> rationale, the accuracy should be:<\/p>\n\n<pre><code>(0.67 + 0.2)\/2\n# 0.435\n<\/code><\/pre>\n\n<p>(Notice how in the normalized matrix the <em>rows<\/em> add to 100%, something that does not happen in the full confusion matrix)<\/p>\n\n<p>But let's now see what the <strong>real<\/strong> accuracy is from the <strong>un-normalized<\/strong> confusion matrix:<\/p>\n\n<pre><code>plot_confusion_matrix(y_true, y_pred, classes=class_names) # normalize=False by default\n# result\nConfusion matrix, without normalization\n[[1 4]\n [1 2]]\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/hImEx.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/hImEx.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>from which, by the definition of accuracy as (TP + TN) \/ (TP + TN + FP + FN), we get:<\/p>\n\n<pre><code>(1+2)\/(1+2+4+1)\n# 0.375\n<\/code><\/pre>\n\n<p>Of course, we don't need the confusion matrix to get something so elementary as the accuracy; as already advised in the comments, we can simply use the built-in <code>accuracy_score<\/code> method of scikit-learn:<\/p>\n\n<pre><code>from sklearn.metrics import accuracy_score\naccuracy_score(y_true, y_pred)\n# 0.375\n<\/code><\/pre>\n\n<p>which, rather unsurprisingly, agrees with our direct computation from the confusion matrix. <\/p>\n\n<hr>\n\n<p>Bottom line:<\/p>\n\n<ul>\n<li>where specific methods (like <code>accuracy_score<\/code>) exist, it is definitely preferable to use them instead of ad hoc inspirations, <em>especially<\/em> when something does not look right (like a discrepancy between Keras and scikit-learn reported accuracies)<\/li>\n<li>the fact that in this example the actual accuracy was lower than the one computed by your own way obviously does not say anything for the specific problem you report<\/li>\n<li>if the discrepancy with Keras still exists even after computing the correct accuracy for your data, please do <strong>not<\/strong> alter the question with the new situation, as this would make the answer invalid, despite the fact that it highlights a mistaken point in your method - open a new question instead<\/li>\n<\/ul>\n"}]}
+{"QuestionId":57763468,"AnswerCount":0,"Tags":"<python><tensorflow><keras><neural-network>","CreationDate":"2019-09-02T23:23:24.677","AcceptedAnswerId":null,"OwnerUserId":5071038.0,"Title":"Evaluate gradient for each input keras","Body":"<p>I want to be able to access the gradient function of a keras neural network. I want to be able to compute the gradient of the output with respect to the input at many input values. Essentially I want to be able to evaluate the gradient as a function.<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(500, activation='relu'))\nmodel.add(Dense(500, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(1, activation='sigmoid'))\n\nmodel.compile(loss='binary_crossentropy', optimizer='adam',  metrics=['accuracy'])\nmodel.fit(x_train, y_train, validation_data=(x_test, y_test), epochs=2, verbose = 1)\n<\/code><\/pre>\n\n<p>Now instead of evaluating the output at every input, I want to evaluate the gradient at every input. How can I do this?<\/p>\n\n<p>This is what I have tried.<\/p>\n\n<pre><code>grads = K.gradients(model.output, model.input)\ngradient_func = K.function([model.input], [grads[0]])\nprint(gradient_func([x]))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57763478,"AnswerCount":1,"Tags":"<tensorflow><keras><keras-layer>","CreationDate":"2019-09-02T23:25:48.060","AcceptedAnswerId":57770059.0,"OwnerUserId":7131919.0,"Title":"computing the norm of the output layer in Tensorflow","Body":"<p>How I can get the values of the output layer in a variational autoencoder network and calculating its norm? Assume I have the following networks in <code>Tensorflow<\/code>:<\/p>\n\n<pre><code>    inputs = Input(shape=(dim,))\n    x = Dense(intermediate_dim, activation='relu')(inputs)\n    z_mean = Dense(latent_dim, name='z_mean')(x)\n    z_log_var = Dense(latent_dim, name='z_log_var')(x)\n\n    z = Lambda(sampling, output_shape=(latent_dim,), name='Z')([z_mean, z_log_var])\n    encoder = Model(inputs, [z_mean, z_log_var, z], name='encoder')\n\n    latent_inputs = Input(shape=(latent_dim,), name='z_sampling')\n    x = Dense(intermediate_dim, activation='relu',name='Hidden_Layer')(latent_inputs)\n    outputs = Dense(dim, activation='sigmoid')(x)\n\n    decoder = Model(latent_inputs, outputs, name='decoder')\n    outputs = decoder(encoder(inputs)[2])\n\n    vae = Model(inputs, outputs, name='vae_mlp')\n<\/code><\/pre>\n\n<p>and I want to find the L_1 norm of the output layer after running the model, i.e. <code>tf.norm(outputs,ord=1)<\/code>. The available Tensorflow functions could not help me. For example after using this command <code>K.eval(tf.norm(decoder.layers[2].output,ord=2))<\/code>\nI got this error:<\/p>\n\n<pre><code>InvalidArgumentError: You must feed value for placeholder tensor 'z_sampling' with dtype float and shape [?,]\n<\/code><\/pre>\n\n<p>Do you have any idea?<\/p>\n","answers":[{"AnswerId":"57770059","CreationDate":"2019-09-03T10:43:19.980","ParentId":null,"OwnerUserId":"5597718","Title":null,"Body":"<p>The output of the network is a numpy array. So, you can use numpy functions.<\/p>\n\n<pre><code>import numpy as np\noutput = vae.predict(input)\nl1_norm = np.linalg.norm(output, 1)\n<\/code><\/pre>\n\n<p>Regarding this error: <\/p>\n\n<blockquote>\n  <p>InvalidArgumentError: You must feed value for placeholder tensor\n  'z_sampling' with dtype float and shape [?,]<\/p>\n<\/blockquote>\n\n<p>Evaluation of <code>tf.norm(decoder.layers[2].output,ord=2)<\/code>, requires the <code>latent_inputs<\/code> to be fed to the decoder which has the name <code>z_sampling<\/code>.<\/p>\n"}]}
+{"QuestionId":57763915,"AnswerCount":0,"Tags":"<tensorflow><indexing><keras>","CreationDate":"2019-09-03T01:06:50.870","AcceptedAnswerId":null,"OwnerUserId":5739514.0,"Title":"Does Keras perform any indexing of Tensors by default?","Body":"<p>I'm wondering if any indexing of tensors is provided by Keras by default or if this is a manual optimization.<\/p>\n\n<blockquote>\n  <p>tf.gather_nd\n  <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather_nd\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather_nd<\/a><\/p>\n<\/blockquote>\n\n<p>Also, can Keras read numpy indexes?<\/p>\n","answers":[]}
+{"QuestionId":57764355,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><tensor>","CreationDate":"2019-09-03T02:27:47.683","AcceptedAnswerId":null,"OwnerUserId":4830432.0,"Title":"Weird Loss and Training Accuracy behavior","Body":"<p>I am slowly cobbling together a toy neural network with Tensorflow. I've been adapting a tutorial to fit my needs but am perplexed at the results.<\/p>\n\n<p>In short, I am feeding arbitrarily generated inputs in to train the network to converge on more arbitrarily generated output values.<\/p>\n\n<p>Here is my code:<\/p>\n\n<pre><code># Imports\n\nimport tensorflow as tf\nimport numpy as np\nimport random as rand\n\n# Data parameters\n\nnum_frames = 1000\n\nstimuli_value_low = .001\nstimuli_value_high = 1.0\n\npixel_value_low = 1.0\npixel_value_high = 256.0\n\nstimuli_dimension = 1\nframe_dimension = 100\n\nstimuli = np.random.uniform(stimuli_value_low, stimuli_value_high, (num_frames, stimuli_dimension))\nframes = np.random.uniform(pixel_value_low, pixel_value_high, (num_frames, frame_dimension))\n\n# Parameters\nlearning_rate = 0.01\ntraining_iterations = 1000\ndisplay_iteration = 10\n\n# Network Parameters\nn_hidden_1 = 100\nn_hidden_2 = 100\nnum_input_neurons = stimuli_dimension\nnum_output_neurons = frame_dimension\n\n# Create placeholders\ninput_placeholder = tf.placeholder(\"float\", [None, num_input_neurons])\noutput_placeholder = tf.placeholder(\"float\", [None, num_output_neurons])\n\n# Store layers weight &amp; bias\nweights = {\n    'h1': tf.Variable(tf.random_normal([num_input_neurons, n_hidden_1])),\n    'h2': tf.Variable(tf.random_normal([n_hidden_1, n_hidden_2])),\n    'out': tf.Variable(tf.random_normal([n_hidden_2, num_output_neurons]))\n}\nbiases = {\n    'b1': tf.Variable(tf.random_normal([n_hidden_1])),\n    'b2': tf.Variable(tf.random_normal([n_hidden_2])),\n    'out': tf.Variable(tf.random_normal([num_output_neurons]))\n}\n\n# Create model\ndef neural_net(input_placeholder):\n    # Hidden fully connected layer \n    layer_1 = tf.add(tf.matmul(input_placeholder, weights['h1']), biases['b1'])\n    # Hidden fully connected layer\n    layer_2 = tf.add(tf.matmul(layer_1, weights['h2']), biases['b2'])\n    # Output fully connected layer with a neuron for each pixel\n    out_layer = tf.matmul(layer_2, weights['out']) + biases['out']\n    return out_layer\n\n# Construct model\nlogits = neural_net(input_placeholder)\n\n# Define loss operation and optimizer\nloss_operation = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits = logits, labels = output_placeholder))\noptimizer = tf.train.AdamOptimizer(learning_rate = learning_rate)\ntrain_operation = optimizer.minimize(loss_operation)\n\n# Evaluate mode\ncorrect_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(output_placeholder, 1))\naccuracy_operation = tf.reduce_mean(tf.cast(correct_prediction, 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 iteration in range(1, training_iterations + 1): \n\n    sess.run(train_operation, feed_dict = {input_placeholder: stimuli, output_placeholder: frames})\n\n    if iteration % display_iteration == 0 or iteration == 1:\n\n      loss = sess.run(loss_operation, feed_dict = {input_placeholder: stimuli, output_placeholder: frames})\n      accuracy = sess.run(accuracy_operation, feed_dict = {input_placeholder: stimuli, output_placeholder: frames})\n\n      print(\"Step \" + str(iteration) + \n            \", Loss = \" + \"{:.4f}\".format(loss) + \n            \", Training Accuracy= \" + \\\n            \"{:.3f}\".format(accuracy))\n\n  print(\"Optimization finished!\")\n<\/code><\/pre>\n\n<p>And here is a sample of the output:<\/p>\n\n<pre><code>Step 1, Loss = 3231716.2500, Training Accuracy= 0.012\nStep 10, Loss = 5028286.0000, Training Accuracy= 0.013\nStep 20, Loss = 8462586.0000, Training Accuracy= 0.013\nStep 30, Loss = 13730642.0000, Training Accuracy= 0.013\nStep 40, Loss = 21263566.0000, Training Accuracy= 0.013\nStep 50, Loss = 31451778.0000, Training Accuracy= 0.013\nStep 60, Loss = 44615720.0000, Training Accuracy= 0.013\nStep 70, Loss = 61044256.0000, Training Accuracy= 0.013\nStep 80, Loss = 80958576.0000, Training Accuracy= 0.013\nStep 90, Loss = 104527496.0000, Training Accuracy= 0.013\nStep 100, Loss = 131878944.0000, Training Accuracy= 0.013\nStep 110, Loss = 163108640.0000, Training Accuracy= 0.013\nStep 120, Loss = 198287152.0000, Training Accuracy= 0.013\nStep 130, Loss = 237464976.0000, Training Accuracy= 0.013\nStep 140, Loss = 280677600.0000, Training Accuracy= 0.013\nStep 150, Loss = 327947968.0000, Training Accuracy= 0.013\nStep 160, Loss = 379289344.0000, Training Accuracy= 0.013\nStep 170, Loss = 434707488.0000, Training Accuracy= 0.013\nStep 180, Loss = 494202976.0000, Training Accuracy= 0.013\nStep 190, Loss = 557770496.0000, Training Accuracy= 0.013\nStep 200, Loss = 625402304.0000, Training Accuracy= 0.013\nStep 210, Loss = 697086144.0000, Training Accuracy= 0.013\nStep 220, Loss = 772809856.0000, Training Accuracy= 0.013\nStep 230, Loss = 852556672.0000, Training Accuracy= 0.013\nStep 240, Loss = 936309952.0000, Training Accuracy= 0.013\nStep 250, Loss = 1024050176.0000, Training Accuracy= 0.013\nStep 260, Loss = 1115761920.0000, Training Accuracy= 0.013\nStep 270, Loss = 1211422336.0000, Training Accuracy= 0.013\nStep 280, Loss = 1311012096.0000, Training Accuracy= 0.013\nStep 290, Loss = 1414512896.0000, Training Accuracy= 0.013\nStep 300, Loss = 1521898112.0000, Training Accuracy= 0.013\nStep 310, Loss = 1633154560.0000, Training Accuracy= 0.013\nStep 320, Loss = 1748253184.0000, Training Accuracy= 0.013\nStep 330, Loss = 1867175680.0000, Training Accuracy= 0.013\nStep 340, Loss = 1989904384.0000, Training Accuracy= 0.013\nStep 350, Loss = 2116416384.0000, Training Accuracy= 0.013\nStep 360, Loss = 2246684416.0000, Training Accuracy= 0.013\nStep 370, Loss = 2380691712.0000, Training Accuracy= 0.013\nStep 380, Loss = 2518416384.0000, Training Accuracy= 0.013\nStep 390, Loss = 2659838208.0000, Training Accuracy= 0.013\nStep 400, Loss = 2804941312.0000, Training Accuracy= 0.013\nStep 410, Loss = 2953692672.0000, Training Accuracy= 0.013\nStep 420, Loss = 3106079232.0000, Training Accuracy= 0.013\nStep 430, Loss = 3262086400.0000, Training Accuracy= 0.013\nStep 440, Loss = 3421681152.0000, Training Accuracy= 0.013\nStep 450, Loss = 3584851456.0000, Training Accuracy= 0.013\nStep 460, Loss = 3751570688.0000, Training Accuracy= 0.013\nStep 470, Loss = 3921825280.0000, Training Accuracy= 0.013\nStep 480, Loss = 4095591168.0000, Training Accuracy= 0.013\nStep 490, Loss = 4272850432.0000, Training Accuracy= 0.013\nStep 500, Loss = 4453586432.0000, Training Accuracy= 0.013\nStep 510, Loss = 4637780480.0000, Training Accuracy= 0.013\nStep 520, Loss = 4825407488.0000, Training Accuracy= 0.013\nStep 530, Loss = 5016456704.0000, Training Accuracy= 0.013\nStep 540, Loss = 5210892800.0000, Training Accuracy= 0.013\nStep 550, Loss = 5408712192.0000, Training Accuracy= 0.013\nStep 560, Loss = 5609891840.0000, Training Accuracy= 0.013\nStep 570, Loss = 5814412800.0000, Training Accuracy= 0.013\nStep 580, Loss = 6022280704.0000, Training Accuracy= 0.013\nStep 590, Loss = 6233428480.0000, Training Accuracy= 0.013\nStep 600, Loss = 6447874048.0000, Training Accuracy= 0.013\nStep 610, Loss = 6665582592.0000, Training Accuracy= 0.013\nStep 620, Loss = 6886544896.0000, Training Accuracy= 0.013\nStep 630, Loss = 7110739968.0000, Training Accuracy= 0.013\nStep 640, Loss = 7338157056.0000, Training Accuracy= 0.013\nStep 650, Loss = 7568768512.0000, Training Accuracy= 0.013\nStep 660, Loss = 7802572288.0000, Training Accuracy= 0.013\nStep 670, Loss = 8039540224.0000, Training Accuracy= 0.013\nStep 680, Loss = 8279664128.0000, Training Accuracy= 0.013\nStep 690, Loss = 8522910720.0000, Training Accuracy= 0.013\nStep 700, Loss = 8769283072.0000, Training Accuracy= 0.013\nStep 710, Loss = 9018755072.0000, Training Accuracy= 0.013\nStep 720, Loss = 9271311360.0000, Training Accuracy= 0.013\nStep 730, Loss = 9526939648.0000, Training Accuracy= 0.013\nStep 740, Loss = 9785614336.0000, Training Accuracy= 0.013\nStep 750, Loss = 10047337472.0000, Training Accuracy= 0.013\nStep 760, Loss = 10312079360.0000, Training Accuracy= 0.013\nStep 770, Loss = 10579824640.0000, Training Accuracy= 0.013\nStep 780, Loss = 10850560000.0000, Training Accuracy= 0.013\nStep 790, Loss = 11124272128.0000, Training Accuracy= 0.013\nStep 800, Loss = 11400942592.0000, Training Accuracy= 0.013\nStep 810, Loss = 11680576512.0000, Training Accuracy= 0.013\nStep 820, Loss = 11963137024.0000, Training Accuracy= 0.013\nStep 830, Loss = 12248598528.0000, Training Accuracy= 0.013\nStep 840, Loss = 12536969216.0000, Training Accuracy= 0.013\nStep 850, Loss = 12828226560.0000, Training Accuracy= 0.013\nStep 860, Loss = 13122355200.0000, Training Accuracy= 0.013\nStep 870, Loss = 13419335680.0000, Training Accuracy= 0.013\nStep 880, Loss = 13719178240.0000, Training Accuracy= 0.013\nStep 890, Loss = 14021866496.0000, Training Accuracy= 0.013\nStep 900, Loss = 14327337984.0000, Training Accuracy= 0.013\nStep 910, Loss = 14635613184.0000, Training Accuracy= 0.013\nStep 920, Loss = 14946696192.0000, Training Accuracy= 0.013\nStep 930, Loss = 15260555264.0000, Training Accuracy= 0.013\nStep 940, Loss = 15577183232.0000, Training Accuracy= 0.013\nStep 950, Loss = 15896529920.0000, Training Accuracy= 0.013\nStep 960, Loss = 16218643456.0000, Training Accuracy= 0.013\nStep 970, Loss = 16543496192.0000, Training Accuracy= 0.013\nStep 980, Loss = 16871051264.0000, Training Accuracy= 0.013\nStep 990, Loss = 17201252352.0000, Training Accuracy= 0.013\nStep 1000, Loss = 17534134272.0000, Training Accuracy= 0.013\nOptimization finished!\n<\/code><\/pre>\n\n<p>My expectations:<\/p>\n\n<ol>\n<li>I would expect the loss to start high and slowly descend to a low value but instead it is all over the place and at weirdly high values. <\/li>\n<li>I would expect the training accuracy to be terrible at first and slowly get better but instead it is stable.<\/li>\n<\/ol>\n\n<p>Questions:<\/p>\n\n<ol>\n<li>Does the way I am trying to generate data and have the network converge as a mapping function over it make sense?<\/li>\n<li>Am I using sensible loss and accuracy operations?<\/li>\n<li>Is there a way I can inspect the math going on behind all of this? I find it hard to just keep poking around in the dark without seeing what is happening.<\/li>\n<\/ol>\n","answers":[]}
+{"QuestionId":57764449,"AnswerCount":0,"Tags":"<pytorch><image-generation><image-translation>","CreationDate":"2019-09-03T02:44:16.840","AcceptedAnswerId":null,"OwnerUserId":12012213.0,"Title":"In the U-Net architecture, h and w don't match, I don't know whether I misunderstood it or not","Body":"<p>I'm confused about the image-transformation structure in a <a href=\"https:\/\/www.mdpi.com\/2076-3417\/9\/13\/2668\" rel=\"nofollow noreferrer\">paper<\/a> that I read. The output dimensions of the layers they concatenate do not match.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/z615g.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/z615g.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Should it actually be layer 9 and layer 7?<\/p>\n\n<pre><code>class ImageTransformationNetwork(nn.Module):\ndef __init__(self):\n    super(ImageTransformationNetwork, self).__init__()\n    self.layer1 = nn.Sequential(nn.Conv2d(3, 64, padding=1, kernel_size=4, stride=2),\n                                nn.LeakyReLU())\n    self.layer2 = nn.Sequential(nn.Conv2d(64, 128, padding=1, kernel_size=4, stride=2),\n                                nn.LeakyReLU(),\n                                nn.BatchNorm2d(128))\n    self.layer3 = nn.Sequential(nn.Conv2d(128, 256, padding=1, kernel_size=4, stride=2),\n                                nn.LeakyReLU(),\n                                nn.BatchNorm2d(256))\n    self.layer4 = nn.Sequential(nn.Conv2d(256, 512, padding=1, kernel_size=4, stride=2),\n                                nn.LeakyReLU(),\n                                nn.BatchNorm2d(512))\n    self.layer5 = nn.Sequential(nn.Conv2d(512, 512, padding=1, kernel_size=4, stride=2),\n                                nn.LeakyReLU(),\n                                nn.BatchNorm2d(512))\n    self.layer6 = nn.Sequential(nn.Conv2d(512, 512, padding=1, kernel_size=4, stride=2),\n                                nn.LeakyReLU(),\n                                nn.BatchNorm2d(512))\n    self.layer7 = nn.Sequential(nn.Conv2d(512, 512, padding=1, kernel_size=4, stride=2),\n                                nn.LeakyReLU(),\n                                nn.BatchNorm2d(512))\n    self.layer8 = nn.Sequential(nn.Conv2d(512, 512, padding=1, kernel_size=4, stride=2),\n                                nn.LeakyReLU())\n    self.layer9 = nn.Sequential(nn.ConvTranspose2d(512, 1024, padding=1, kernel_size=4, stride=2),\n                                nn.ReLU(),\n                                nn.BatchNorm2d(1024))\n    self.layer10 = nn.Sequential(nn.ConvTranspose2d(1024+512, 1024, padding=1, kernel_size=4, stride=2),\n                                nn.ReLU(),\n                                nn.BatchNorm2d(1024))\n    self.layer11 = nn.Sequential(nn.ConvTranspose2d(1024+512, 1024, padding=1, kernel_size=4, stride=2),\n                                 nn.ReLU(),\n                                 nn.BatchNorm2d(1024))\n    self.layer12 = nn.Sequential(nn.ConvTranspose2d(1024+512, 1024, padding=1, kernel_size=4, stride=2),\n                                 nn.ReLU(),\n                                 nn.BatchNorm2d(1024))\n    self.layer13 = nn.Sequential(nn.ConvTranspose2d(1024+512, 512, padding=1, kernel_size=4, stride=2),\n                                 nn.ReLU(),\n                                 nn.BatchNorm2d(512))\n    self.layer14 = nn.Sequential(nn.ConvTranspose2d(512+256, 256, padding=1, kernel_size=4, stride=2),\n                                 nn.ReLU(),\n                                 nn.BatchNorm2d(256))\n    self.layer15 = nn.Sequential(nn.ConvTranspose2d(256+128, 128, padding=1, kernel_size=4, stride=2),\n                                 nn.Tanh(),\n                                 nn.BatchNorm2d(128))\n    self.layer16 = nn.Sequential(nn.ConvTranspose2d(128, 3, padding=1, kernel_size=4, stride=2))\n\ndef forward(self, x):\n    x1 = self.layer1(x)\n    x2 = self.layer2(x1)\n    x3 = self.layer3(x2)\n    x4 = self.layer4(x3)\n    x5 = self.layer5(x4)\n    x6 = self.layer6(x5)\n    x7 = self.layer7(x6)\n    x8 = self.layer8(x7)\n    x9 = self.layer9(x8)\n    x10 = self.layer10(torch.cat([x7, x9], dim=1))\n    x11 = self.layer11(torch.cat([x6, x10], dim=1))\n    x12 = self.layer12(torch.cat([x5, x11], dim=1))\n    x13 = self.layer13(torch.cat([x4, x12], dim=1))\n    x14 = self.layer14(torch.cat([x3, x13], dim=1))\n    x15 = self.layer15(torch.cat([x2, x14], dim=1))\n    x16 = self.layer16(x15)\n    return x16\n<\/code><\/pre>\n\n<p>I don't know if I understand it correctly. Could you provide some help about it?<\/p>\n","answers":[]}
+{"QuestionId":57764604,"AnswerCount":0,"Tags":"<deep-learning><gpu><pytorch><nvidia>","CreationDate":"2019-09-03T03:11:21.317","AcceptedAnswerId":null,"OwnerUserId":6366309.0,"Title":"Volatile gpu-util is 100% and no progress on PyTorch1.1 after infering lots of images","Body":"<p>I deploy a pretrained deep model on NVIDIA V100 GPU. After infering lots of images, I find Volatile GPU-Utils still shows 100%, but no infering is running now. How to fix this problem?<\/p>\n\n<p>When I type <code>nvidia-smi<\/code>, it shows:<\/p>\n\n<blockquote>\n  <p>+-----------------------------------------------------------------------------+<br>\n  | NVIDIA-SMI 384.183      Driver Version: 384.183      CUDA Version: 9.0      |<br>\n  |-------------------------------+----------------------+----------------------+<br>\n  | GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |<br>\n  | Fan  Temp  Perf  Pwr:Usage\/Cap|         Memory-Usage | GPU-Util  Compute M. |<br>\n  |=============+=============+===============|<br>\n  |   0  Tesla V100-PCIE...  On   | 00000000:3B:00.0 Off |                    0 |<br>\n  | N\/A   43C    P0    51W \/ 250W |   2530MiB \/ 32502MiB |    100%      Default |  <\/p>\n<\/blockquote>\n\n<p>I have tried <strong>sudo nvidia-smi -pm 1<\/strong> to enable Persistence Mode. But the problem still remains.<\/p>\n","answers":[]}
+{"QuestionId":57764685,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-03T03:26:13.833","AcceptedAnswerId":null,"OwnerUserId":1653558.0,"Title":"trouble with the .tflite model from TF 2.0 image classification","Body":"<p>I just built a <strong>TF 2.0<\/strong> RC0 image classification model using this guide <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/images\/hub_with_keras\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/tutorials\/images\/hub_with_keras<\/a>, then converted the model to .tflite <\/p>\n\n<pre><code>converter = tf.lite.TFLiteConverter.from_keras_model(model)\ntflite_model = converter.convert()\nopen(\"converted_model.tflite\", \"wb\").write(tflite_model)\n<\/code><\/pre>\n\n<p>When I try to use .tflite with the label_image script from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/lite\/examples\/python\/label_image.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/lite\/examples\/python\/label_image.py<\/a> I do not get images classified correctly.  The classification works fine with model.predict (i.e. before converting to .tflite)<\/p>\n\n<p>Any idea what I may have done wrong or how to troubleshoot this?<\/p>\n","answers":[]}
+{"QuestionId":57764859,"AnswerCount":0,"Tags":"<python><deep-learning><hook><pytorch>","CreationDate":"2019-09-03T03:55:25.770","AcceptedAnswerId":null,"OwnerUserId":9610081.0,"Title":"TypeError: get_features_hook() missing 1 required positional argument: 'output'","Body":"<p>I'm a beginner of Pytorch and i'm learning how to use hook. I'd like to see what's going on with hook. Below is what I've tried<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def get_features_hook(module,input,output):\n    print(input)\n    print(output)\nhandle_feat = alexnet.features[0].register_forward_hook(get_features_hook)\n\na = alexnet(input_data)\na.backward(torch.ones(1, num_class))\nhandle_feat.remove()\n<\/code><\/pre>\n\n<p>But I got this error:<\/p>\n\n<blockquote>\n  <p>TypeError: get_features_hook() missing 1 required positional argument: 'output'<\/p>\n<\/blockquote>\n\n<p>Also I tried this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def get_features_hook(input,output)\n\nand   \n\ndef get_features_hook(self,input,output)\n\n<\/code><\/pre>\n\n<p>And I got the same result,<\/p>\n\n<blockquote>\n  <p>TypeError: get_features_hook() missing 1 required positional argument: 'output'<\/p>\n<\/blockquote>\n\n<p>I was wondering if there's anyone could help me fix that and told me what's wrong with my codes.<\/p>\n","answers":[]}
+{"QuestionId":57764911,"AnswerCount":1,"Tags":"<python><multithreading><tensorflow><queue>","CreationDate":"2019-09-03T04:04:50.217","AcceptedAnswerId":null,"OwnerUserId":10969942.0,"Title":"Can't understand the result of TensorOverflow train.QueueRunner","Body":"<p>Sorry, I'm a beginner of TensorFlow. The following is just a TensorFlow's code using two threads to enqueue and dequeue independently. <\/p>\n\n<pre><code>import tensorflow as tf\n\nQ = tf.compat.v1.FIFOQueue(1000, tf.float32)\nvar = tf.Variable(0.0)\ndata = tf.compat.v1.assign_add(var, tf.constant(1.0))\nen_q = Q.enqueue(data)\nqr = tf.train.QueueRunner(Q, enqueue_ops=[en_q])\ninit_op = tf.compat.v1.global_variables_initializer()\nwith tf.compat.v1.Session() as sess:\n    sess.run(init_op)\n    coord = tf.train.Coordinator()\n    threads = qr.create_threads(sess, coord=coord, start=True)\n    for i in range(300):\n        print(sess.run(Q.dequeue()))\n    coord.request_stop()\n    coord.join(threads)\n<\/code><\/pre>\n\n<p>The result is like following:<\/p>\n\n<pre><code>3.0\n7.0\n10.0\n14.0\n18.0\n21.0\n26.0\n....\n<\/code><\/pre>\n\n<p>I'm quite confused about this result. Since <code>Q<\/code> is a FIFO queue, even though the dequeue and enqueue are in two different threads, the number enqueued still should be <code>1,2,3,4,5,6,...<\/code>. Why is it possible the dequeued number can be <code>3,7,10,14,....<\/code> Where are the numbers <code>1, 2, 4, 5, ...<\/code>?<\/p>\n","answers":[{"AnswerId":"57770934","CreationDate":"2019-09-03T11:37:44.050","ParentId":null,"OwnerUserId":"11427968","Title":null,"Body":"<p>The randomness has nothing to do with the <code>FIFOqueue<\/code> but is an expected behavior with the concurrency of thread execution. There are definitely more than just 2 threads involved.<\/p>\n\n<p>To simplify things, let's try to run your code without the <code>for<\/code> loop, add some <code>print<\/code> statements &amp; a <code>sleep<\/code> interval &amp; see what happens:<\/p>\n\n<pre><code>import tensorflow as tf\nimport time\n\nQ_size = 1000 # Q size\nt = 1 # t in sec\n\nQ = tf.compat.v1.FIFOQueue(Q_size, tf.float32)\nvar = tf.Variable(0.0)\ndata = tf.compat.v1.assign_add(var, tf.constant(1.0))\nen_q = Q.enqueue(data)\n\nqr = tf.train.QueueRunner(Q, enqueue_ops=[en_q])\n\ninit_op = tf.compat.v1.global_variables_initializer()\n\nwith tf.compat.v1.Session() as sess:\n    sess.run(init_op)\n    coord = tf.train.Coordinator()\n\n    print('*** before qr.create_threads ***')\n    print('Q.size()', sess.run(Q.size()))\n    print('var', var.eval())\n\n    threads = qr.create_threads(sess, coord=coord, start=True)\n\n    print('*** after qr.create_threads ***')\n    time.sleep(t) # sleep t sec in main to wait for all threads to finish running.\n    print('Q.size()', sess.run(Q.size()))\n    print('var', var.eval())\n\n    coord.request_stop()\n    coord.join(threads)    \n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>*** before qr.create_threads ***\nQ.size() 0\nvar 0.0\n\n*** after qr.create_threads ***\nQ.size() 1000\nvar 1001.0\n<\/code><\/pre>\n\n<p>Before calling <code>qr<\/code>, nothing happens. There's nothing in the <code>FIFOQueue<\/code> <code>Q<\/code> &amp; <code>var<\/code> stills equals 0. This is expected.<\/p>\n\n<p>After calling <code>qr.create_threads<\/code> &amp; setting <code>start=True<\/code>, <code>qr<\/code> does the following:<\/p>\n\n<ol>\n<li>Starts populating <code>Q<\/code> with your <code>enqueue_ops<\/code> which in this case, is <code>en_q<\/code>. It will try to pack as many <code>en_q<\/code> ops as possible which is decided by your <code>Q<\/code> size.<\/li>\n<li>Creating threads for each <code>en_q<\/code> in <code>Q<\/code>.<\/li>\n<li>Run the threads concurrently. (At this point, given the size of <code>Q<\/code>, it's clear that there can't be only just 2 threads involved as mentioned by the OP.)<\/li>\n<\/ol>\n\n<p>The <code>sleep<\/code> is introduce to allow all these threads to finish execution before we display the output. We want to know what is the size of <code>Q<\/code> &amp; value of <code>var<\/code> after all threads created by <code>qr<\/code> finish running.<\/p>\n\n<p>So as expected, <code>Q<\/code> is populated with 1000 <code>en_q<\/code> ops &amp; <code>var<\/code> is added 1001 times.<\/p>\n\n<p>Now, if we remove the <code>sleep<\/code> interval, we get random output like the one shown below.<\/p>\n\n<p>Random output without <code>sleep<\/code>:<\/p>\n\n<pre><code>*** before qr.create_threads ***\nQ.size() 0\nvar 0.0\n\n*** after qr.create_threads ***\nQ.size() 32\nvar 35.0\n<\/code><\/pre>\n\n<p>What the random output above is saying is that when we display the output in the main thread, <code>qr<\/code> is still populating the <code>Q<\/code>, creating &amp; running the threads concurrently in the background.<\/p>\n\n<p>Now, lets put your <code>for<\/code> loop back in while introducing another <code>sleep<\/code> interval in the loop:<\/p>\n\n<pre><code>import tensorflow as tf\nimport time\n\nQ_size = 1000 # Q size\nN =10 # range of for loop\nt = 1 # t in secs\n\nQ = tf.compat.v1.FIFOQueue(Q_size, tf.float32)\nvar = tf.Variable(0.0)\ndata = tf.compat.v1.assign_add(var, tf.constant(1.0))\nen_q = Q.enqueue(data)\n\nqr = tf.train.QueueRunner(Q, enqueue_ops=[en_q])\n\ninit_op = tf.compat.v1.global_variables_initializer()\n\nwith tf.compat.v1.Session() as sess:\n    sess.run(init_op)\n    coord = tf.train.Coordinator()\n\n    print('*** before qr.create_threads ***')\n    print('Q.size()', sess.run(Q.size()))\n    print('var', var.eval())\n\n    threads = qr.create_threads(sess, coord=coord, start=True)\n\n    print('*** after qr.create_threads ***')\n    time.sleep(t) # sleep t sec in main to wait for all threads to finish running.\n    print('Q.size()', sess.run(Q.size()))\n    print('var', var.eval())\n\n    print('*** for loop ***')\n    for i in range(N):\n        sess.run(Q.dequeue())\n        time.sleep(t) # sleep t sec in main to wait for all threads to finish running.\n        print('var', var.eval())\n        print('.size()', sess.run(Q.size()))\n\n    coord.request_stop()\n    coord.join(threads)    \n<\/code><\/pre>\n\n<p>Introducing <code>sleep<\/code> allows background threads to finish before we display the variables in the main thread.<\/p>\n\n<p>You can see in the output below that <code>var<\/code> increases seqentially now while <code>Q<\/code> size remains at 1000. Each <code>sess.run(Q.dequeue())<\/code> call now executes one <code>en_q<\/code> op from <code>Q<\/code>, adding 1 to <code>var<\/code>.<\/p>\n\n<p>I hope this clears up that the randomness that you observed has nothing to do with <code>Q<\/code> being <code>FIFOqueue<\/code> or not. The randomness is caused by the expected behavior of thread parallelism.<\/p>\n\n<p>Output:<\/p>\n\n<pre><code>*** before qr.create_threads ***\nQ.size() 0\nvar 0.0\n\n*** after qr.create_threads ***\nQ.size() 1000\nvar 1001.0\n\n*** for loop ***\nvar 1002.0\n.size() 1000\nvar 1003.0\n.size() 1000\nvar 1004.0\n.size() 1000\nvar 1005.0\n.size() 1000\nvar 1006.0\n.size() 1000\nvar 1007.0\n.size() 1000\nvar 1008.0\n.size() 1000\nvar 1009.0\n.size() 1000\nvar 1010.0\n.size() 1000\nvar 1011.0\n.size() 1000\n<\/code><\/pre>\n"}]}
+{"QuestionId":57765249,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-03T04:56:37.403","AcceptedAnswerId":null,"OwnerUserId":9803875.0,"Title":"How many times does tf.nn.rnn_cell.LSTMCell() run? Can I modify it or do I have to modify dynamic_rnn()?","Body":"<p>I want to create a modified RNN cell in TensorFlow, that will first compute the cell state c(t) and output state h(t) (like it usually does) and then perform certain modifications in a for loop to h(t) and return a new h'(t) instead.<\/p>\n\n<p>My understanding is that an RNN cell e.g. tf.nn.BasicLSTMCell() will run once for each time stamp. (I think I am wrong in this point).<\/p>\n\n<p>While debugging, I noticed that <code>__call__()<\/code> in my <code>OrthogonalLSTMCell_hard()<\/code> is only called once while declaring my new cell in my <code>build_rnn()<\/code> function, and then never gets called again during prediction. This results in <code>for h_t in self.previous_outputs<\/code> loop never being called again after build_model() and the cell is not doing what I want it to be doing. Is this because OrthogonalLSTMCell_hard() is only supposed to be instantiated once and added to the graph. <\/p>\n\n<p>Is dynamic_rnn() supposed to be modified instead to modify my h(t)s?<\/p>\n\n<pre><code>class OrthogonalLSTMCell_hard(RNNCell):\n  def __call__(self, inputs, state, scope=None):\n    \"\"\"Long short-term memory cell (LSTM).\"\"\"\n    with vs.variable_scope(scope or type(self).__name__):\n      print (scope)\n\n      eps = sys.float_info.min      \n      # Parameters of gates are concatenated into one multiply for efficiency.\n      if self._state_is_tuple:\n        c, h = state\n      else:\n        c, h = array_ops.split(state, 2, 1)\n\n      concat = _linear([inputs, h], 4 * self._num_units, True)\n\n      # i = input_gate, j = new_input, f = forget_gate, o = output_gate\n      i, j, f, o = array_ops.split(concat, 4, 1)\n\n      c = (c * sigmoid(f + self._forget_bias) + sigmoid(i) *\n               self._activation(j))\n\n      temp_h = self._activation(c)*sigmoid(o)\n\n      for h_t in self.previous_outputs:\n\n        # Step 1: find cosine sim between hid and temp_4, transpose and store as hid_temp\n        cos_factor = math_ops.div(math_ops.reduce_sum(math_ops.mul(temp_h, h_t),1),math_ops.reduce_sum(math_ops.mul(h_t,h_t),1) + eps)\n        h_t = array_ops.transpose(h_t)\n\n        h_t_tilda = math_ops.mul(h_t, cos_factor)\n        h_t_tilda = array_ops.transpose(h_t_tilda)\n\n        temp_h = temp_h - h_t_tilda\n\n\n\n\n\n      self.previous_outputs.append(temp_h)\n\n      if self._state_is_tuple:\n        new_state = LSTMStateTuple(c, temp_h)\n      else:\n        new_state = array_ops.concat(1, [c, temp_h])\n\n      return temp_h, new_state\n<\/code><\/pre>\n\n<p>Thanks in advance for any help.<\/p>\n","answers":[]}
+{"QuestionId":57765751,"AnswerCount":0,"Tags":"<python><deep-learning><hook><pytorch>","CreationDate":"2019-09-03T05:54:49.553","AcceptedAnswerId":null,"OwnerUserId":9610081.0,"Title":"TypeError: hook() takes 2 positional arguments but 3 were given","Body":"<p>I'm new to pytorch and I'm trying to use <code>hook()<\/code> and <code>register_forward_pre_hook<\/code> in my project<\/p>\n\n<p>What I've tried is <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def get_features_hook(module,input):\n    print(input)\n\nhandle_feat = alexnet.features[0].register_forward_pre_hook(get_features_hook)\n\n\na = alexnet(input_data)\n<\/code><\/pre>\n\n<p>And I got belows error at <code>a = alexnet(input_data)<\/code><\/p>\n\n<blockquote>\n  <p>TypeError: get_features_hook() takes 2 positional arguments but 3 were given<\/p>\n<\/blockquote>\n\n<p>I've lost few hours on this problem and I just can't able to figure it out.<\/p>\n\n<p>Anyone likes to help me?<\/p>\n\n<hr>\n\n<p>With Shai's help, I tried his codes, and I got this<\/p>\n\n<pre><code>Conv2d(3, 64, kernel_size=(11, 11), stride=(4, 4), padding=(2, 2))\nget_features_hook called with 2 args:\n    arg of type Conv2d\n    arg of type tuple\n\nFile \"&lt;input&gt;\", line 2, in get_features_hook\nNameError: name 'args' is not defined\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57766289,"AnswerCount":0,"Tags":"<tensorflow><lstm>","CreationDate":"2019-09-03T06:44:34.880","AcceptedAnswerId":null,"OwnerUserId":11007013.0,"Title":"how to save memory when training LSTM model with different data shape on tensorflow","Body":"<p>I have build a model with lstm cell as below code, and its inputs are data1 and data2, their shape is [None, 5] which means sometime [3, 5], sometime [13, 5], sometime [80, 5] and most more than [50, 5];<\/p>\n\n<p>when training, I find used memory is increasing continually with training step which means used memory is more when train more batches and finally memory error after train 200 or 300 batches;<\/p>\n\n<p>Now I know how to avoid this(used memory is continue increasing when training) by delete \"@tf.function\", but I don't know why and what is the difference between \"@tf.function\" and no \"@tf.function\".<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\ndatas1 = [\n        [[2.3,3.3,2,4.3,2.3],[3.,1.2,3.,22.1,23.2,]],\n        [[2.3,3.3,2,4.3,2.3],[3.,1.2,3.,22.1,23.2,],[2.3,3.3,2,4.3,2.3],[3.,1.2,3.,22.1,23.2,]],\n        [[2,3,2,3,2],[3,2,3,2,2,]],\n        [[2.3,3.3,2,4.3,2.3],[3.,1.2,3.,22.1,23.2,],[2.3,3.3,2,4.3,2.3]],\n        ...\n        ]\ndatas2 = [\n        [[2.3,3.3,2,4.3,2.3],[3.,1.2,3.,22.1,23.2,],[2.3,3.3,2,4.3,2.3]],\n        [[2.3,3.3,2,4.3,2.3],[3.,1.2,3.,22.1,23.2,],[3.,1.2,3.,22.1,23.2,]],\n        [[2,3,2,3,2],[3,2,3,2,2,]],\n        [[2.3,3.3,2,4.3,2.3],[3.,1.2,3.,22.1,23.2,],[2.3,3.3,2,4.3,2.3],[3,2,3,2,2,]],\n        ...\n        ]\nlabels = [[0.0], [1.0], [1.0], [0.0], ...]\nL = len(labels)\ndef generator():\n    for i in range(L):\n        yield datas1[i], datas2[i], labels[i]\ntrains = tf.data.Dataset.from_generator(generator, (tf.float32, tf.float32, tf.float16))\n\nclass ABC(tf.keras.Model):\n    def __init__(self):\n        super(ABC, self).__init__()\n        self.encoder1 = tf.keras.models.Sequential([\\\n                        tf.keras.layers.LSTM(32, return_sequences=True),\n                        tf.keras.layers.LSTM(16)])\n        self.encoder2 = tf.keras.models.Sequential([\\\n                        tf.keras.layers.LSTM(32, return_sequences=True),\n                        tf.keras.layers.LSTM(16)])\n        self.dense = tf.keras.layers.Dense(1, activation='sigmoid')\n    def call(self, data1, data2):\n        en1 = self.encoder1(data1)\n        en2 = self.encoder2(data2)\n        en12 = tf.keras.layers.concatenate([en1, en2])\n        return self.dense(en12)\nmodel = ABC()\nloss_object = tf.keras.losses.BinaryCrossentropy()\noptimizer = tf.keras.optimizers.Adam()\n\n@tf.function\ndef train_step(data1, data2, labels):\n    with tf.GradientTape() as tape:\n        predictions = model(data1, data2)\n        loss = loss_object(labels, predictions)\n    gradients = tape.gradient(loss, model.trainable_variables)\n    optimizer.apply_gradients(zip(gradients, model.trainable_variables))\n\nfor i in range(5):\n    for data1, data2, label in trains.batch(1):\n        train_step(data1, data2, label)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57766523,"AnswerCount":0,"Tags":"<tensorflow><keras><tpu>","CreationDate":"2019-09-03T07:02:12.000","AcceptedAnswerId":null,"OwnerUserId":8176285.0,"Title":"Trying to run effecientNet on a TPU","Body":"<p>Trying to run EffecientNet on a TPU,by running this <\/p>\n\n<pre><code>effnet = efn.EfficientNetB5(weights='imagenet', include_top=False)\n\n# Replace all Batch Normalization layers by Group Normalization layers\nfor i, layer in enumerate(effnet.layers):\n    if \"batch_normalization\" in layer.name:\n        effnet.layers[i] = GroupNormalization(groups=32, axis=-1, epsilon=0.00001)\n\n\nmodel = Sequential()\nmodel.add(effnet)\nmodel.add(GlobalAveragePooling2D())\nmodel.add(Dropout(0.5))\nmodel.add(Dense(8, activation=elu))\nmodel.compile(loss='mse', optimizer=RAdam(lr=0.00005), metrics=['mse', 'acc'])\nprint(model.summary())        \n\nmodel_json = model.to_json()\nwith open(\"model_ef7_fn.json\", \"w\") as json_file:\n    json_file.write(model_json)\n\n\ntpu_model = tf.contrib.tpu.keras_to_tpu_model(\n    model,\n    strategy=tf.contrib.tpu.TPUDistributionStrategy(\n        tf.contrib.cluster_resolver.TPUClusterResolver(\n            tpu='grpc:\/\/' + os.environ['COLAB_TPU_ADDR'])\n    )\n)\n<\/code><\/pre>\n\n<p>But i get this error<\/p>\n\n<p>ValueError: \n        Layer  has a variable shape in a non-batch dimension.  TPU models must\n        have constant shapes for all operations.<\/p>\n\n<pre><code>    You may have to specify `input_length` for RNN\/TimeDistributed layers.\n\n    Layer: &lt;keras.engine.training.Model object at 0x7f34e5c06f60&gt;\n    Input shape: (None, None, None, 3)\n    Output shape: (None, None, None, 2048)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57766591,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-03T07:06:36.990","AcceptedAnswerId":null,"OwnerUserId":12005078.0,"Title":"I want to achieve the size of the unfixed image in the tensorflow code","Body":"<p>I want to implement spp(Spatial Pyramid Pooling) in the project.\nBut it has not been successful, I want to ask for help, thank you.<\/p>\n\n<p>Implemented in the convolutional layer and the fully connected layer, I have seen a lot of code, but it has not been successful.<\/p>\n\n<p>This is the key code\uff1a<\/p>\n\n<pre><code>def model(self):\n    x = tf.reshape(self.X, shape=[-1, self.image_height, self.image_width, 1])\n    print(\"&gt;&gt;&gt; input x: {}\".format(x))\n\n    # Convolution layer  1\n    wc1 = tf.get_variable(name='wc1', shape=[3, 3, 1, 32], dtype=tf.float32,\n                          initializer=tf.contrib.layers.xavier_initializer())\n    bc1 = tf.Variable(self.b_alpha * tf.random_normal([32]))\n    conv1 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(x, wc1, strides=[1, 1, 1, 1], padding='SAME'), bc1))\n    conv1 = tf.nn.max_pool(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')\n    conv1 = tf.nn.dropout(conv1, self.keep_prob)\n\n    #  Convolution layer   2\n    wc2 = tf.get_variable(name='wc2', shape=[3, 3, 32, 64], dtype=tf.float32,\n                          initializer=tf.contrib.layers.xavier_initializer())\n    bc2 = tf.Variable(self.b_alpha * tf.random_normal([64]))\n    conv2 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(conv1, wc2, strides=[1, 1, 1, 1], padding='SAME'), bc2))\n    conv2 = tf.nn.max_pool(conv2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')\n    conv2 = tf.nn.dropout(conv2, self.keep_prob)\n\n    # Convolution layer  3\n    wc3 = tf.get_variable(name='wc3', shape=[3, 3, 64, 128], dtype=tf.float32,\n                          initializer=tf.contrib.layers.xavier_initializer())\n    bc3 = tf.Variable(self.b_alpha * tf.random_normal([128]))\n    conv3 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(conv2, wc3, strides=[1, 1, 1, 1], padding='SAME'), bc3))\n    conv3 = tf.nn.max_pool(conv3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')\n    conv3 = tf.nn.dropout(conv3, self.keep_prob)\n    print(\"&gt;&gt;&gt; convolution 3: \", conv3.shape)\n    next_shape = conv3.shape[1] * conv3.shape[2] * conv3.shape[3]\n\n    # Fully connected  1\n    wd1 = tf.get_variable(name='wd1', shape=[next_shape, 1024], dtype=tf.float32,\n                          initializer=tf.contrib.layers.xavier_initializer())\n    bd1 = tf.Variable(self.b_alpha * tf.random_normal([1024]))\n    dense = tf.reshape(conv3, [-1, wd1.get_shape().as_list()[0]])\n    dense = tf.nn.relu(tf.add(tf.matmul(dense, wd1), bd1))\n    dense = tf.nn.dropout(dense, self.keep_prob)\n\n    #  Fully connected  2\n    wout = tf.get_variable('name', shape=[1024, self.max_captcha * self.char_set_len], dtype=tf.float32,\n                           initializer=tf.contrib.layers.xavier_initializer())\n    bout = tf.Variable(self.b_alpha * tf.random_normal([self.max_captcha * self.char_set_len]))\n\n    with tf.name_scope('y_prediction'):\n        y_predict = tf.add(tf.matmul(dense, wout), bout)\n\n    return y_predict\n<\/code><\/pre>\n\n<p>all code\n<a href=\"https:\/\/github.com\/nickliqian\/cnn_captcha\" rel=\"nofollow noreferrer\">https:\/\/github.com\/nickliqian\/cnn_captcha<\/a><\/p>\n\n<p>table of Contents\uff1a\/cnnlib\/network.py<\/p>\n","answers":[]}
+{"QuestionId":57766761,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2019-09-03T07:18:43.577","AcceptedAnswerId":null,"OwnerUserId":11868407.0,"Title":"How to set batch scheduler in Tensorflow Serving","Body":"<p>I have seen there are 3 different schedulers for do batching in tensorflow:<\/p>\n\n<ul>\n<li><p>basic_batch_scheduler<\/p><\/li>\n<li><p>shared_batch_scheduler <\/p><\/li>\n<li><p>adaptive_shared_batch_scheduler<\/p><\/li>\n<\/ul>\n\n<p>Ref: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/core\/kernels\/batching_util\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/core\/kernels\/batching_util<\/a><\/p>\n\n<p>Is there any way to set wich one is running in Tensorflow Serving? The batching config file doesn't include any flag to do it.<\/p>\n","answers":[]}
+{"QuestionId":57766813,"AnswerCount":1,"Tags":"<python><tensorflow><keras><image-segmentation>","CreationDate":"2019-09-03T07:22:00.080","AcceptedAnswerId":null,"OwnerUserId":11806602.0,"Title":"Why using Binary accuracy as metric gives a higher accuracy(95%) while simple accuracy outputs low value(46%)","Body":"<p>I am doing semantic segmentation on DSTL kaggle satellite image. The masks for the classes given binary masks. when I use accuracy as metric the output is around 46% but loss (weighted binary cross-entropy loss) is round 0.073. when I use binary accuracy as metric then accuracy comes around 95%. why this kind of results are coming?<\/p>\n","answers":[{"AnswerId":"57784535","CreationDate":"2019-09-04T08:34:10.100","ParentId":null,"OwnerUserId":"8784825","Title":null,"Body":"<p>For semantic segmentation other metrics should be used. Since we are predicting areas and classes.<\/p>\n\n<p>Take a look at this <a href=\"https:\/\/towardsdatascience.com\/metrics-to-evaluate-your-semantic-segmentation-model-6bcb99639aa2\" rel=\"nofollow noreferrer\">post<\/a> for further understanding.<\/p>\n\n<p>Good luck!<\/p>\n"}]}
+{"QuestionId":57767113,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2019-09-03T07:42:55.700","AcceptedAnswerId":null,"OwnerUserId":9954063.0,"Title":"How to use tf.nn.softmax_cross_entropy_with_logits () function?","Body":"<p>I have a small neural network for which I am using <code>tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits())<\/code> to calculate the loss and then using gradient descent optimiser. Unfortunately I am having a value error with calculating the loss. <\/p>\n\n<p>I have tried to check the shape of input labels and logits and both of them have same shape.Where labels are the output of my neural network and logits are the preferred output which is a 4X1 hot encoded vector. Labels are also of same shape.<\/p>\n\n<p>this is how I am creating a hot encoded vector, where model_number is an int between 1 and 4.<\/p>\n\n<pre><code>\n<\/code><\/pre>\n\n<p>This is how learning function looks like in my code<\/p>\n\n<pre><code>def kt_learning(self, learning_rate, n_out , n_preferred):\n        'where n_out is the actual output and n_preffered is the prefered output of the network'\n        loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=n_out, logits=n_preferred))\n        optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss)\n        with tf.Session() as sess:\n            sess.run(optimizer)\n        return\n<\/code><\/pre>\n\n<p>and finally to have an idea of how ma NN output looks like, here it is,<\/p>\n\n<pre><code> def run(self, carollis_input):\n        c_in=normalize(carollis_input, axis = 0)\n        print('normalized carollis input is \\n')\n        print(c_in)\n        c_in = np.array(c_in)\n        c_input = tf.compat.v1.convert_to_tensor(c_in, tf.float64)\n        #'finding the output of the input layer'\n        weight_i = self.input_weight()\n        weight_h = self.hid_weight()\n        weight_o = self.out_weight()\n        bias_i = self.bias_in()\n        bias_h = self.bias_hid()\n        bias_o = self.bias_out()\n        init = tf.global_variables_initializer()\n        sess = tf.Session()\n        sess.run(init)\n        knowledge_input = tf.add(tf.multiply(c_input, weight_i), bias_i)\n        sess.run(knowledge_input)\n        knowledge_hidden = tf.nn.leaky_relu(knowledge_input, alpha=0.01) \n        #'calculating the output of hidden layer'\n        knowledge_hidden_output = 3.14*(tf.add(tf.multiply(knowledge_hidden, weight_h), bias_h))#input function of hidden layer\n        knowledge_hidden_out = tf.nn.leaky_relu(knowledge_hidden_output, alpha=0.01, name='leaky_relu')\n\n        sess.run(knowledge_hidden_out)\n        #'calculating the input of output layer'\n        knowledge_hidden_out = tf.reshape(knowledge_hidden_out, [4, 2])#for quadrant method\n        out_mult = tf.multiply(knowledge_hidden_out, weight_o)\n        out_add = tf.add(out_mult[:, 0], out_mult[:, 1])\n        in_out = tf.add(out_add, bias_o)\n        output = sess.run(in_out)\n        #'finding the softmax output of the neurons'\n        softmax_output = tf.nn.softmax(in_out)\n        output = sess.run(softmax_output)\n        return output\n<\/code><\/pre>\n\n<p>Here is the error <\/p>\n\n<pre><code>ValueError: Cannot reshape a tensor with 4 elements to shape [1] (1 elements) for 'softmax_cross_entropy_with_logits_sg_1\/Reshape_2' (op: 'Reshape') with input shapes: [4], [1] and with input tensors computed as partial shapes: input[1] = [1].\n\n``\n\ncan anyone please help me, I am kind of on deadline and these errors with tensorflow are eating up the time I have. Thank you.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57767679,"AnswerCount":0,"Tags":"<python><tensorflow><loss-function><tf.keras>","CreationDate":"2019-09-03T08:18:55.223","AcceptedAnswerId":null,"OwnerUserId":12010475.0,"Title":"Apply custom function to each vector in a batch in loss function in Tensorflow","Body":"<p>I am writing a custom loss function for neural network training. I am currently using Tensorflow 2.0-rc with <code>tf.keras<\/code> to define and train the model. The problem is that I need to do some computations with each vector in a batch and the code doing those operations can not handle tensors and is too complicated to be rewritten using Tensorflow.<\/p>\n\n<p>Is there a way to apply a function to each vector in a batch which extracts  the data as a Numpy array and uses that to produce another array?<\/p>\n\n<p>I know that I cannot simply get the batch size or convert the tensor arguments of the loss function to arrays. I also tried <code>tf.map_fn<\/code> like so:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def some_non_tf_function(t):\n    return np.mean(t)\n\n@tf.custom_gradient\ndef my_loss(y_true, y_pred):\n    def grad(dy):\n        # just a dummy\n        return tf.identity(y_true), dy*y_true\n\n    return tf.reduce_mean(tf.map_fn(some_non_tf_function, y_pred)), grad\n<\/code><\/pre>\n\n<p>But it complains saying <code>(TensorArrayV2Read\/TensorListGetItem:0) to a numpy array.<\/code>\nIs there any way to get this to work without writing a custom training routine?<\/p>\n","answers":[]}
+{"QuestionId":57767854,"AnswerCount":1,"Tags":"<tensorflow><keras><nlp><pytorch>","CreationDate":"2019-09-03T08:30:23.940","AcceptedAnswerId":null,"OwnerUserId":9135328.0,"Title":"keras.preprocessing.text.Tokenizer equivalent in Pytorch?","Body":"<p>Basically the title; is there any equivalent to<code>keras.preprocessing.text.Tokenizer<\/code> in Pytorch? I have yet to find any that gives all the utilities without handcrafting things.<\/p>\n","answers":[{"AnswerId":"57773051","CreationDate":"2019-09-03T13:44:49.947","ParentId":null,"OwnerUserId":"5652313","Title":null,"Body":"<p>PyTorch itself does not provide a function like this, you either need to it manually (which should be easy: use a tokenizer of your choice and do a dictionary lookup for the indices).<\/p>\n\n<p>Alternatively, you can use <a href=\"https:\/\/github.com\/pytorch\/text\" rel=\"nofollow noreferrer\">Torchtext<\/a>, which provides basic abstraction from text processing. All you need to do is create a <a href=\"https:\/\/torchtext.readthedocs.io\/en\/latest\/data.html#field\" rel=\"nofollow noreferrer\"><code>Field<\/code><\/a> object. You can use <code>string.split<\/code>, SpaCy or custom function for tokenization. You can provide a vocabulary or create it directly from data. Then you just call <a href=\"https:\/\/torchtext.readthedocs.io\/en\/latest\/data.html#torchtext.data.Field.process\" rel=\"nofollow noreferrer\">the <code>process<\/code> method<\/a> which tokenizes text and does the vocabulary lookup.<\/p>\n\n<p>If you want something more complex, you might consider using also <a href=\"https:\/\/allennlp.org\/\" rel=\"nofollow noreferrer\">AllenNLP<\/a>. In AllenNLP, you do separately <a href=\"https:\/\/allenai.github.io\/allennlp-docs\/api\/allennlp.data.tokenizers.html\" rel=\"nofollow noreferrer\">the tokenization<\/a> and <a href=\"https:\/\/allenai.github.io\/allennlp-docs\/api\/allennlp.data.vocabulary.html\" rel=\"nofollow noreferrer\">the vocabulary lookup<\/a>.<\/p>\n"}]}
+{"QuestionId":57767899,"AnswerCount":2,"Tags":"<python><scikit-learn><pytorch><amazon-sagemaker>","CreationDate":"2019-09-03T08:32:41.723","AcceptedAnswerId":null,"OwnerUserId":9584727.0,"Title":"How to create a pipeline in sagemaker with pytorch","Body":"<p>I am dealing with a classification problem with text data in sagemaker. Where, i first fit and transform it into structured format(say by using TFIDF in sklearn) then i kept the result in S3 bucket and i used it for training my pytorch model for which i have written the code in my entry point.<\/p>\n\n<p>if we notice, by the end of the above process, i have two models<\/p>\n\n<ol>\n<li>sklearn TFIDF model <\/li>\n<li>actual PyTorch model<\/li>\n<\/ol>\n\n<p>So, when every time i need to predict on a new text data, i need to separately process(transform) the text data with TFIDF model which i created during my training.<\/p>\n\n<p>How can i create a pipeline in sagemaker with sklearn's TFIDF and pytorch models.  <\/p>\n\n<p>if i fit and transform text data using TFIDF in my main method in entrypoint then if i train my pytorch model in my main method, i can return only one model which will be used in model_fn()<\/p>\n","answers":[{"AnswerId":"57772033","CreationDate":"2019-09-03T12:43:24.113","ParentId":null,"OwnerUserId":"6025065","Title":null,"Body":"<p>First, checkout the mnist example here:<\/p>\n\n<p><a href=\"https:\/\/github.com\/awslabs\/amazon-sagemaker-examples\/blob\/master\/sagemaker-python-sdk\/pytorch_mnist\/pytorch_mnist.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/awslabs\/amazon-sagemaker-examples\/blob\/master\/sagemaker-python-sdk\/pytorch_mnist\/pytorch_mnist.ipynb<\/a><\/p>\n\n<p>With script mode, you can run the code (in mnist.py) using the below estimator.<\/p>\n\n<pre><code>from sagemaker.pytorch import PyTorch\n\nestimator = PyTorch(entry_point='mnist.py',\n                    role=role,\n                    framework_version='1.1.0',\n                    train_instance_count=2,\n                    train_instance_type='ml.c4.xlarge',\n                    hyperparameters={\n                        'epochs': 6,\n                        'backend': 'gloo'\n                    })\n<\/code><\/pre>\n\n<p>Simply update the mnist.py script as per tfidf pipeline. Hope this helps.<\/p>\n"},{"AnswerId":"57772416","CreationDate":"2019-09-03T13:07:12.303","ParentId":null,"OwnerUserId":"9584727","Title":null,"Body":"<p>Apparently, We need to use inference pipelines.<\/p>\n\n<p>An inference pipeline is an Amazon SageMaker model that is <strong>composed of a linear sequence of two to five containers that process requests for inferences on data<\/strong>. You use an inference pipeline to define and deploy any combination of pretrained Amazon SageMaker built-in algorithms and your own custom algorithms packaged in Docker containers. You can use an inference pipeline to combine preprocessing, predictions, and post-processing data science tasks. Inference pipelines are fully managed.<\/p>\n\n<p>one can read the docs here - <\/p>\n\n<p><a href=\"https:\/\/docs.aws.amazon.com\/sagemaker\/latest\/dg\/inference-pipelines.html\" rel=\"nofollow noreferrer\">https:\/\/docs.aws.amazon.com\/sagemaker\/latest\/dg\/inference-pipelines.html<\/a><\/p>\n\n<p>Example -<\/p>\n\n<p><a href=\"https:\/\/github.com\/awslabs\/amazon-sagemaker-examples\/blob\/master\/sagemaker-python-sdk\/scikit_learn_inference_pipeline\/Inference%20Pipeline%20with%20Scikit-learn%20and%20Linear%20Learner.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/awslabs\/amazon-sagemaker-examples\/blob\/master\/sagemaker-python-sdk\/scikit_learn_inference_pipeline\/Inference%20Pipeline%20with%20Scikit-learn%20and%20Linear%20Learner.ipynb<\/a><\/p>\n"}]}
+{"QuestionId":57768036,"AnswerCount":0,"Tags":"<performance><tensorflow><tensorflow2.0>","CreationDate":"2019-09-03T08:42:05.120","AcceptedAnswerId":null,"OwnerUserId":997253.0,"Title":"How much slower is tensorflow 2.0 with eager execution?","Body":"<p>In Tensorflow 1 graph mode is faster than eager execution. However, Tensorflow 2.0 comes with <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/eager\" rel=\"nofollow noreferrer\">eager execution enabled by default<\/a>, and <a href=\"https:\/\/stackoverflow.com\/a\/55938423\/997253\">evidently<\/a>, you shouldn't disable it. I can't really seem to find any survey or benchmarking comparing eager execution to graph execution.<\/p>\n\n<p>How does performance compare between graph mode and eager execution Tensorflow 2.0? What, if any, are the advantages with using Graph mode in Tensorflow 2.0?<\/p>\n","answers":[]}
+{"QuestionId":57768083,"AnswerCount":0,"Tags":"<json><tensorflow>","CreationDate":"2019-09-03T08:44:52.813","AcceptedAnswerId":null,"OwnerUserId":7888603.0,"Title":"How to export a model as JSON in tensorflow 1.14","Body":"<p>I cant find any method to export tensorflow 1.14 models as JSON. <\/p>\n\n<p>I used one method-<\/p>\n\n<pre><code>import tensorflow as tf\nfrom google.protobuf import json_format\n\nwith tf.Graph().as_default() as graph:\n  # Add nodes to the graph...\n\ngraph_def = graph.as_graph_def()\n\njson_string = json_format.MessageToJson(graph_def)\n<\/code><\/pre>\n\n<p>But this gives the JSON representation of GraphDef but not the model definition as JSON<\/p>\n","answers":[]}
+{"QuestionId":57768369,"AnswerCount":0,"Tags":"<python><tensorflow><keras><lstm><embedding>","CreationDate":"2019-09-03T09:02:24.610","AcceptedAnswerId":null,"OwnerUserId":5929597.0,"Title":"Character Embedding in Keras","Body":"<p>Im trying to implement the embedding network of <a href=\"https:\/\/arxiv.org\/abs\/1903.00277\" rel=\"nofollow noreferrer\">this paper<\/a>\n(Sorry i am not allowed to post an image)<\/p>\n\n<p>The paper says following:\n\"The network first embeds each character of the sequence\ns in R128, then encodes it with a four-layer bidirectional\nLSTM recurrent network (with a hidden state of\nsize 128). '(s) is the output of the last bidirectional LSTM\nlayer.\"<\/p>\n\n<p>I have a numpy array with 2350 words of different lenght where every character is indexed by an integer. There are 96 different Charachters.<\/p>\n\n<p>Example:\nThe word \"asd\" is represented as <code>[[66], [84], [69]]<\/code><\/p>\n\n<p>My network looks as following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = Embedding(output_dim=128, input_dim=96, input_length=20)(text_input)\nx = Bidirectional(LSTM(64, return_sequences=True))(x)\nx = Bidirectional(LSTM(64, return_sequences=True))(x)\nx = Bidirectional(LSTM(64, return_sequences=True))(x)\nx = Bidirectional(LSTM(64))(x)\n<\/code><\/pre>\n\n<p>How can i get the embedded word form the network and use it in the other and what must be the input layer?<\/p>\n","answers":[]}
+{"QuestionId":57768405,"AnswerCount":0,"Tags":"<python><numpy><tensorflow><keras>","CreationDate":"2019-09-03T09:04:21.893","AcceptedAnswerId":null,"OwnerUserId":5210052.0,"Title":"Keras-vis for the saliency charts - shall we use softmax or linear activation","Body":"<p>We could start by training a simple model on MNIST data set using the code <a href=\"https:\/\/github.com\/raghakot\/keras-vis\/blob\/master\/examples\/mnist\/activation_maximization.ipynb\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>With the keras-vis package, if we could reload the MNIST model trained in the example and draw all the saliency charts for for one digit (e.g. '7') using filter index from zero to ten, we can see that it's very difficult to observe the contrast using the linear activation. <\/p>\n\n<p>Yet, the contrast can indeed be seen using softmax activation. Moreover, we need to manually set a consistent scale to see the contrast.<\/p>\n\n<p>It may be important, at least for defect detection topics, to know how and why an image is classified as one thing but not another by checking the significant contrasts in saliency visualizations.<\/p>\n\n<pre><code>from __future__ import print_function\nfrom keras.datasets import mnist\nimport numpy as np\nfrom matplotlib import pyplot as plt\nfrom vis.visualization import visualize_saliency\nfrom vis.utils import utils\nfrom keras import activations\nfrom keras.models import load_model\n\n(x_train, y_train), (x_test, y_test) = mnist.load_data()\n\n# Load once for a model with softmax for last dense layer, and load again for one with linear swap\n\nMODEL_PATH = \"model.h5\"\n\nmodel = load_model(MODEL_PATH)\nraw_model = load_model(MODEL_PATH)\n\n# check layers in the model\nNAMES = []\nfor index, layer in enumerate(model.layers):\n    NAMES.append(layer.name)\n    print(index, layer.name)\nprint('====================================================\\n\\n\\n')\n\n# swap softmax\nlayer_idx = utils.find_layer_idx(model, 'dense_2')\nmodel.layers[layer_idx].activation = activations.linear\nmodel = utils.apply_modifications(model)\n\n\n# prepare a sample image '7'\nimg = x_test[0]\/255\n\nseed = img.copy()\nseed = np.expand_dims(seed, 2)\nseed = np.expand_dims(seed, 0)\n\n# use absolute scale for the \nMAX_PIXEL_softmax = 0.01\nMAX_PIXEL_linear = 1\n\n\nfor index in range(10):\n    print('----------------------------------------------')\n    print('Digit: ', index)\n    f, ax = plt.subplots(1, 3)\n\n    grads_softmax = visualize_saliency(raw_model, layer_idx, filter_indices=index,\n                               seed_input=seed, backprop_modifier=\"guided\")\n    print('total:', round(grads_softmax.sum()*10000), '  max:', round(grads_softmax.max(),5), '  min:', round(grads_softmax.min(),5))\n    grads_softmax[0,0] = MAX_PIXEL_softmax\n    ax[0].set_title('Softmax ' + str(index))\n    ax[0].imshow(grads_softmax, cmap = 'jet')\n\n    grads_linear = visualize_saliency(model, layer_idx, filter_indices=index,\n                               seed_input=seed, backprop_modifier=\"guided\")\n    print('total:', round(grads_linear.sum()), '  max:', round(grads_linear.max(),5), '  min:', round(grads_linear.min(),5))\n    grads_linear[0,0] = MAX_PIXEL_linear\n    ax[1].set_title('Linear ' + str(index))\n    ax[1].imshow(grads_linear, cmap = 'jet')\n\n    ax[2].set_title('Raw image')\n    ax[2].imshow(img)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/UUyio.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/UUyio.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/Ijqws.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Ijqws.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/J8jCw.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/J8jCw.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>In the author's webpage, it's stated that <\/p>\n\n<blockquote>\n  <p>To visualize activation over final dense layer outputs, we need to\n  switch the softmax activation out for linear since gradient of output\n  node will depend on all the other node activations. Doing this in\n  keras is tricky, so we provide utils.apply_modifications to modify\n  network parameters and rebuild the graph.<\/p>\n  \n  <p>If this swapping is not done, the results might be suboptimal. We will\n  start by swapping out 'softmax' for 'linear' and compare what happens\n  if we dont do this at the end.<\/p>\n<\/blockquote>\n\n<p>But obviously, using linear activation will make the contrast disappear.\nIs there anything wrong in my code that makes linear activation behave like this? Or if we need to show the contrast, it's still better to use softmax activation?<\/p>\n","answers":[]}
+{"QuestionId":57768577,"AnswerCount":0,"Tags":"<python><machine-learning><keras>","CreationDate":"2019-09-03T09:16:30.513","AcceptedAnswerId":null,"OwnerUserId":515028.0,"Title":"To train in Keras two identical RNN but with different outputs","Body":"<p>I am training to solve a problem where I am using a RNN to track an object, learn about it, and then generate trajectories. Therefore the input of the RNN is stuff like x,y,speed,... and the output is the next x,y. So far this works fine.<\/p>\n\n<p>Now I am trying to use the same network not only to learn about the same as I did before, but also to learn how the shape of the object (just the radius so far) changes over time. Thefore the inputs of this RNN are x,y,speed,radius, and the outputs are x,y,radius. This is not working well, and I think it is because it is trying to learn two separate things.<\/p>\n\n<p>My first thought was to create two independent RNNs, one learns for x,y and the other for radius, but this would be very inefficient since they are learning over the same data, and probably both of them will have the same architecture.<\/p>\n\n<p>I wonder if there's a way of combining them into a single model in order to train them in a more efficient way.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def createRNNmodel(network_input):\n\n    model = Sequential()\n    model.add(LSTM(\n        128\/\/1,\n        input_shape=(network_input.shape[1], network_input.shape[2]),\n        return_sequences=True\n    ))\n    model.add(Dropout(0.5))\n    model.add(LSTM(128\/\/2, return_sequences=True))\n    model.add(Dropout(0.5))\n    model.add(LSTM(128\/\/2, return_sequences=True))\n    model.add(Dropout(0.5))\n    model.add(LSTM(128\/\/4, return_sequences=True))\n    model.add(Dropout(0.5))\n    model.add(LSTM(128\/\/8, return_sequences=False))\n    model.add(Dense(128\/\/2))\n    model.add(Dropout(0.5))\n    model.add(Dense(128\/\/8))\n    model.add(Dropout(0.5))\n    model.add(Dense(3))\n    model.compile(loss='mean_squared_error', optimizer='rmsprop')\n\n    return model\n\n\ndef trainRNNmodel(model, network_input, network_output):\n\n    folderpath = \"\/home\/data\/RNN_weights\/\"\n    filepath = folderpath+\"weights-{epoch:03d}-{loss:.4f}.hdf5\"    \n\n    checkpoint = ModelCheckpoint(\n        filepath, monitor='loss', \n        verbose=0,        \n        save_weights_only=False,\n        save_best_only=True,        \n        mode='min',\n        period=5\n    )    \n    callbacks_list = [checkpoint]     \n\n    model.fit(np.array(network_input), np.array(network_output), \n            validation_split=0.1, epochs=1000, batch_size=256, callbacks=callbacks_list)\n\n    return model\n<\/code><\/pre>\n\n<p>And the moment this is what I have, which is the RNN that tries to learn about x,y,radius (you can see the output is 3). I would like to create two identical networks, where the only difference is the output layer, one will have 2 (x,y) and the other 1 (radius), and they will be trained with the same data.<\/p>\n\n<p>I don't know if this is possible with Keras.<\/p>\n\n<p>Thanks for reading.<\/p>\n","answers":[]}
+{"QuestionId":57768685,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2019-09-03T09:21:57.757","AcceptedAnswerId":57769467.0,"OwnerUserId":11994808.0,"Title":"How to assign new values to a tensorflow constant?","Body":"<p>I am loading a TensorFlow model from a .pb file. I want to change the weights of all the layers. I am able to extract the weights but I am not able to change the weights.<\/p>\n\n<p>I converted the graph_def model to TensorFlow model but even then I cannot assign a new value to the weights as the weights are stored in a tensor of type \"Const\".<\/p>\n\n<pre><code>b = graph_tf.get_tensor_by_name(\"Variable_1:0\")      \ntf.assign(b, np.ones((1,1,64,64))) \n<\/code><\/pre>\n\n<p>I am getting the following error: <\/p>\n\n<pre><code>AttributeError: 'Tensor' object has no attribute 'assign'\n<\/code><\/pre>\n\n<p>Please provide a way to resolve this issue. Thanks in advance.<\/p>\n","answers":[{"AnswerId":"57769467","CreationDate":"2019-09-03T10:07:07.730","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Here is one way you can achieve something like that. You want to replace some constant operations with variables initialized to the value of those operations, so you can first extract those constant values, and then create the graph with the variables initalized to those. See the example below.<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Example graph\nwith tf.Graph().as_default():\n    inp = tf.placeholder(tf.float32, [None, 3], name='Input')\n    w = tf.constant([[1.], [2.], [3.]], tf.float32, name='W')\n    out = tf.squeeze(inp @ w, 1, name='Output')\n    gd = tf.get_default_graph().as_graph_def()\n\n# Extract weight values\nwith tf.Graph().as_default():\n    w, = tf.graph_util.import_graph_def(gd, return_elements=['W:0'])\n    # Get the constant weight values\n    with tf.Session() as sess:\n        w_val = sess.run(w)\n    # Alternatively, since it is a constant,\n    # you can get the values from the operation attribute directly\n    w_val = tf.make_ndarray(w.op.get_attr('value'))\n\n# Make new graph\nwith tf.Graph().as_default():\n    # Make variables initialized with stored values\n    w = tf.Variable(w_val, name='W')\n    init_op = tf.global_variables_initializer()\n    # Import graph\n    inp, out = tf.graph_util.import_graph_def(\n        gd, input_map={'W:0': w},\n        return_elements=['Input:0', 'Output:0'])\n    # Change value operation\n    w_upd = w[2].assign([5.])\n    # Test\n    with tf.Session() as sess:\n        sess.run(init_op)\n        print(sess.run(w))\n        # [[1.]\n        #  [2.]\n        #  [3.]]\n        sess.run(w_upd)\n        print(sess.run(w))\n        # [[1.]\n        #  [2.]\n        #  [5.]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57768800,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-03T09:28:17.687","AcceptedAnswerId":null,"OwnerUserId":9574181.0,"Title":"how to save custom trained model without full connect layer just like MobileNetV2 include_top=False","Body":"<p>i want to save my trained model to .h5 without last two layers, in order to transfer learning using my custom model in the furture, just like MobileNetV2 include_top=False, can someone help me, thanks!<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>base_model = tf.keras.applications.mobilenet_v2.MobileNetV2(\n    alpha=1.0,\n    input_shape=IMG_SHAPE,\n    include_top=False,\n    weights='imagenet')\n\nmodel = tf.keras.Sequential([\n    base_model,\n    tf.keras.layers.GlobalAveragePooling2D(),\n    tf.keras.layers.Dense(255, activation=tf.nn.softmax)\n])\n<\/code><\/pre>\n\n<p>trained model like this:<\/p>\n\n<pre><code>Model: \"sequential\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #\n=================================================================\nmobilenetv2_1.00_224 (Model) (None, 2, 2, 1280)        2257984\n_________________________________________________________________\nglobal_average_pooling2d (Gl (None, 1280)              0\n_________________________________________________________________\ndense (Dense)                (None, 205)               262605\n=================================================================\nTotal params: 2,520,589\nTrainable params: 2,486,477\nNon-trainable params: 34,112\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>when i try to using it for transfer learning<\/p>\n\n<pre><code>keras_model = loadModel(keras_model_path)\nkeras_model.summary()\n\ninput = keras_model.input\nhidden = tf.keras.layers.GlobalMaxPooling2D()(keras_model.layers[-3].output)\nout = tf.keras.layers.Dense(128, activation=tf.nn.softmax)(hidden)\nmodel2 = tf.keras.Model(input, out)\nmodel2.summary()\n<\/code><\/pre>\n\n<p>an error occurs<\/p>\n\n<pre><code>ValueError: Graph disconnected: cannot obtain value for tensor Tensor(\"input_1:0\", shape=(?, 64, 64, 3), dtype=float32) at layer \"input_1\". The following previous layers were accessed without issue: []\n<\/code><\/pre>\n","answers":[{"AnswerId":"57769224","CreationDate":"2019-09-03T09:52:46.577","ParentId":null,"OwnerUserId":"10509263","Title":null,"Body":"<blockquote>\n  <p>i want to save my trained model to .h5 without last two layers,<\/p>\n<\/blockquote>\n\n<p>why don't you save the full model with model.save() and when you reload it for transfer learning, just remove the layers using:<\/p>\n\n<pre><code>model.layers.pop()\n<\/code><\/pre>\n\n<p>You can also remove the layers before saving the model but I wouldn't do that<\/p>\n"}]}
+{"QuestionId":57769065,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-datasets><tensorflow-serving><tensorflow-estimator>","CreationDate":"2019-09-03T09:44:08.780","AcceptedAnswerId":null,"OwnerUserId":11765356.0,"Title":"TensorFlow Serving Error: 'StatelessIf has '_lower_using_switch_merge' attr set but it does not support lowering.'","Body":"<p>When attempting to serve a new model coded using TensorFlow 2.0 with TensorFlow serving, I get the following error from my Docker container logs:<\/p>\n\n<pre><code>2019-09-03 08:56:24.984824: I external\/org_tensorflow\/tensorflow\/cc\/saved_model\/reader.cc:31] Reading SavedModel from: \/models\/model_modeFact\/1567500955\n2019-09-03 08:56:24.989902: I external\/org_tensorflow\/tensorflow\/cc\/saved_model\/reader.cc:54] Reading meta graph with tags { serve }\n2019-09-03 08:56:25.002593: I external\/org_tensorflow\/tensorflow\/cc\/saved_model\/loader.cc:311] SavedModel load for tags { serve }; Status: fail. Took 17772 microseconds.\n2019-09-03 08:56:25.002658: E tensorflow_serving\/util\/retrier.cc:37] Loading servable: {name: model_modeFact version: 1567500955} failed: Internal: Node {{node zero_fraction\/total_zero\/zero_count\/else\/_1\/zero_fraction\/cond}} of type StatelessIf has '_lower_using_switch_merge' attr set but it does not support lowering.\n<\/code><\/pre>\n\n<p>Using the saved_model_cli, the model works fine and can make predictions.\nInitially I was getting this error: \"<a href=\"https:\/\/stackoverflow.com\/questions\/57758785\/tensorflow-serving-crossed-columns-strange-error\">TensorFlow Serving crossed columns strange error<\/a>\"<\/p>\n\n<p>I found that this error might be fixed by swapping to tf-nightly-2.0-preview==2.0.0.dev20190819\nBut instead I am now I can't even get my model to be served. <\/p>\n\n<p>The only changes I made to the code to compile my model in TF2 are:<\/p>\n\n<pre><code># Added this line to force eager execution, necessary for tf.placeholders\ntf.compat.v1.disable_eager_execution()\n\n# For every usage of tf.estimator...\ntf.compat.v1.estimator\n\n# For every usage of tf.placeholder...\ntf.compat.v1.placeholder\n<\/code><\/pre>\n\n<p>Like the previous problem, the goal is to have a prediction output from my served model, an output similar to when I use saved_model_cli. Something like this:<\/p>\n\n<pre><code>Result for output key all_class_ids:\n[[0 1 2 3 4 5]]\nResult for output key all_classes:\n[[b'0' b'1' b'2' b'3' b'4' b'5']]\nResult for output key class_ids:\n[[2]]\nResult for output key classes:\n[[b'2']]\nResult for output key logits:\n[[ 0.11128154 -0.44881764  0.31520572 -0.08318427 -0.3479367  -0.08883157]]\nResult for output key probabilities:\n[[0.19719791 0.11263006 0.2418051  0.16234797 0.12458517 0.16143374]]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57769108,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2019-09-03T09:46:48.173","AcceptedAnswerId":null,"OwnerUserId":11782908.0,"Title":"Invalid argument: Input to reshape is a tensor with 6000 values, but the requested shape has 6120","Body":"<p>I am working on an image to image regression task which requires me to develop a deep learning model that takes in a sequence of 5 images and return another image. The sequence of 5 images and the output images are conceptually and temporally related. In fact, the 5 images in the sequence each correspond to timestep in a simulation and the output, the one I am trying to predict, corresponds eventually to the 6th timestep of that sequence.<\/p>\n\n<p>To try complete this task I chose a U-net architecture that given an image of size 55 x 1636 (corresponding to the sequence of images) as input returns another 11 x 1636 image (corresponding to the 6th timestep image of that sequence). <\/p>\n\n<p>However when trying to train the model using \"model.fit_generator\" it is giving the error shown below:<\/p>\n\n<pre><code>Invalid argument: Input to reshape is a tensor with 6000 values, but the \nrequested shape has 6120\n<\/code><\/pre>\n\n<p>I am not figuring out why this error occurs, does anybody have any idea why this is happening?<\/p>\n\n<p>Down below is the code and the full error message.<\/p>\n\n<pre><code>#--------------- FUNCTIONS ---------------#\n\ndef down_sample(input_, filters, kernel_size, pool_size, padding=\"same\", \nstrides=1):\n    conv = keras.layers.Conv2D(filters, kernel_size, padding=padding, \n    strides=strides, activation=\"tanh\")(input_)\n\n    conv = keras.layers.Conv2D(filters, kernel_size, padding=padding, \n    strides=strides, activation=\"tanh\")(conv)\n\n    pool = keras.layers.MaxPool2D(pool_size)(conv)\n    return conv, pool\n\n\ndef up_sample(input_, skip, filters, kernel_size, pool_size, target_size, \npadding=\"same\", strides=1):\n    up_s = keras.layers.UpSampling2D(pool_size)(input_)\n    up_s_ = keras.layers.Reshape(target_size)(up_s)\n    concat = keras.layers.Concatenate(axis = -1)([up_s_, skip])\n    conv = keras.layers.Conv2D(filters, kernel_size, padding=padding, \n    strides=strides, activation=\"tanh\")(concat)\n\n    conv = keras.layers.Conv2D(filters, kernel_size, padding=padding, \n    strides=strides, activation=\"tanh\")(conv)\n    return conv\n\ndef bottleneck(input_, filters, kernel_size, padding=\"same\", strides=1):\n    conv = keras.layers.Conv2D(filters, kernel_size, padding=padding, \n    strides=strides, activation=\"tanh\")(input_)\n\n    conv = keras.layers.Conv2D(filters, kernel_size, padding=padding, \n    strides=strides, activation=\"tanh\")(conv)\n    return conv\n\ndef resize(input_, target_size):\n    l = keras.layers.Reshape(target_size)(input_)\n    return l\n\n\n#--------------- MODEL CONSTRUCTION ---------------#\n\ndef UNet():\n    f = [16, 8, 4, 2]\n    inputs = keras.layers.Input((55, 1636, 1))\n\n    p0 = inputs\n    print(p0.shape)\n    c1, p1 = down_sample(p0, f[0], (4,8), (4,8)) \n    c2, p2 = down_sample(p1, f[1], (2,4), (2,4)) \n    c3, p3 = down_sample(p2, f[2], (1,2), (1,2))\n\n    bn = bottleneck(p3, f[3], (3,3))\n\n    u1 = up_sample(bn, c3, f[2], (1,2), (1,2), (6, 51, 2))\n    u2 = up_sample(u1, c2, f[1], (2,4), (2,4), (13, 204, 4))\n    u3 = up_sample(u2, c1, f[0], (4,8), (4,8), (55, 1636, 8))\n\n    outputs = keras.layers.Conv2D(4, (1, 2), padding=\"same\", \n    activation=\"tanh\")(u3)\n    outputs_ = resize(outputs, (11, 1636, 1))\n    model = keras.models.Model(inputs, outputs_)\n\n    return model\n\n\n#--------------- MODEL SUMMARY ---------------#\n\nmodel = UNet()\n\ndef nrmse(y_true, y_pred):\nreturn backend.sqrt(backend.mean(backend.square(y_pred - y_true)))\/(2)\n\ndef rmse(y_true, y_pred):\nreturn backend.sqrt(backend.mean(backend.square(y_pred - y_true), \n    axis=-1))\n\nmodel.compile(loss = 'mean_squared_error',\n              optimizer = 'adam',\n              metrics = [nrmse])\n\nmodel.summary()\n\n\n#--------------- MODEL TRAINING ---------------#\n\nwith tf.device('\/gpu:0'):\nbatch_size = 10\nnb_epoch = 5\n\nresult = model.fit_generator(train_generator, \n                    steps_per_epoch = len(training_data) ,\n                    epochs = nb_epoch,\n                    validation_data = validation_generator,\n                    validation_steps = len(val_data))\n<\/code><\/pre>\n\n<p>Shown below are the model summary and the full error message:<\/p>\n\n<pre><code>Layer (type)                    Output Shape         Param #     Connected to\n\n==================================================================================================\n\ninput_36 (InputLayer)           [(None, 55, 1636, 1) 0\n\n\n__________________________________________________________________________________________________\n\nconv2d_384 (Conv2D)             (None, 55, 1636, 16) 528         input_36[0][0]        \n\n\n__________________________________________________________________________________________________\n\nconv2d_385 (Conv2D)             (None, 55, 1636, 16) 8208        conv2d_384[0][0]    \n\n\n__________________________________________________________________________________________________\n\nmax_pooling2d_105 (MaxPooling2D (None, 13, 204, 16)  0           conv2d_385[0][0] \n\n\n__________________________________________________________________________________________________\n\nconv2d_386 (Conv2D)             (None, 13, 204, 8)   1032        max_pooling2d_105[0][0]\n\n\n__________________________________________________________________________________________________\n\nconv2d_387 (Conv2D)             (None, 13, 204, 8)   520         conv2d_386[0][0] \n\n\n__________________________________________________________________________________________________\n\nmax_pooling2d_106 (MaxPooling2D (None, 6, 51, 8)     0           conv2d_387[0][0]\n\n\n__________________________________________________________________________________________________\n\nconv2d_388 (Conv2D)             (None, 6, 51, 4)     68          max_pooling2d_106[0][0] \n\n\n__________________________________________________________________________________________________\n\nconv2d_389 (Conv2D)             (None, 6, 51, 4)     36          conv2d_388[0][0] \n\n\n__________________________________________________________________________________________________\n\nmax_pooling2d_107 (MaxPooling2D (None, 6, 25, 4)     0           conv2d_389[0][0] \n\n\n__________________________________________________________________________________________________\n\nconv2d_390 (Conv2D)             (None, 6, 25, 2)     74          max_pooling2d_107[0][0]\n\n\n__________________________________________________________________________________________________\n\nconv2d_391 (Conv2D)             (None, 6, 25, 2)     38          conv2d_390[0][0] \n\n\n__________________________________________________________________________________________________\n\nup_sampling2d_66 (UpSampling2D) (None, 6, 50, 2)     0           conv2d_391[0][0]\n\n\n__________________________________________________________________________________________________\n\nreshape_99 (Reshape)            (None, 6, 51, 2)     0           up_sampling2d_66[0][0]  \n\n\n__________________________________________________________________________________________________\n\nconcatenate_61 (Concatenate)    (None, 6, 51, 6)     0           reshape_99[0][0]                 \n                                                             conv2d_389[0][0]\n\n\n__________________________________________________________________________________________________\n\nconv2d_392 (Conv2D)             (None, 6, 51, 4)     52          concatenate_61[0][0] \n\n\n__________________________________________________________________________________________________\n\nconv2d_393 (Conv2D)             (None, 6, 51, 4)     36          conv2d_392[0][0] \n\n\n__________________________________________________________________________________________________\n\nup_sampling2d_67 (UpSampling2D) (None, 12, 204, 4)   0           conv2d_393[0][0]\n\n\n__________________________________________________________________________________________________\n\nreshape_100 (Reshape)           (None, 13, 204, 4)   0           up_sampling2d_67[0][0]\n\n\n__________________________________________________________________________________________________\n\nconcatenate_62 (Concatenate)    (None, 13, 204, 12)  0           reshape_100[0][0]                \n                                                             conv2d_387[0][0]   \n\n\n__________________________________________________________________________________________________\n\nconv2d_394 (Conv2D)             (None, 13, 204, 8)   776         concatenate_62[0][0`]       \n\n\n__________________________________________________________________________________________________\n\nconv2d_395 (Conv2D)             (None, 13, 204, 8)   520         conv2d_394[0][0]   \n\n\n__________________________________________________________________________________________________\n\nup_sampling2d_68 (UpSampling2D) (None, 52, 1632, 8)  0           conv2d_395[0][0] \n\n\n__________________________________________________________________________________________________\n\nreshape_101 (Reshape)           (None, 55, 1636, 8)  0           up_sampling2d_68[0][0]  \n\n\n__________________________________________________________________________________________________\n\nconcatenate_63 (Concatenate)    (None, 55, 1636, 24) 0           reshape_101[0][0]                \n                                                             conv2d_385[0][0]\n\n\n__________________________________________________________________________________________________\n\nconv2d_396 (Conv2D)             (None, 55, 1636, 16) 12304       concatenate_63[0][0]  \n\n\n__________________________________________________________________________________________________\n\nconv2d_397 (Conv2D)             (None, 55, 1636, 16) 8208        conv2d_396[0][0] \n\n\n__________________________________________________________________________________________________\n\nconv2d_398 (Conv2D)             (None, 55, 1636, 4)  132         conv2d_397[0][0] \n\n\n__________________________________________________________________________________________________\n\nreshape_102 (Reshape)           (None, 11, 1636, 1)  0           conv2d_398[0][0]  \n\n==================================================================================================\n\nTotal params: 32,532\nTrainable params: 32,532\nNon-trainable params: 0\n<\/code><\/pre>\n\n<hr>\n\n<pre><code>InvalidArgumentError                      Traceback (most recent call \nlast)\n&lt;ipython-input-105-b6a4abb757fd&gt; in &lt;module&gt;\n      7                         epochs = nb_epoch,\n      8                         validation_data = validation_generator,\n----&gt; 9                         validation_steps = len(val_data))\n\nC:\\ProgramData\\Anaconda3\\lib\\site\npackages\\tensorflow\\python\\keras\\engine\\training.py in fit_generator(self, \ngenerator, steps_per_epoch, epochs, verbose, callbacks, validation_data, \nvalidation_steps, validation_freq, class_weight, max_queue_size, workers, \nuse_multiprocessing, shuffle, initial_epoch)\n   1431         shuffle=shuffle,\n   1432         initial_epoch=initial_epoch,\n-&gt; 1433         steps_name='steps_per_epoch')\n   1434 \n   1435   def evaluate_generator(self,\n\nC:\\ProgramData\\Anaconda3\\lib\\site- \npackages\\tensorflow\\python\\keras\\engine\\training_generator.py in \nmodel_iteration(model, data, steps_per_epoch, epochs, verbose, callbacks, \nvalidation_data, validation_steps, validation_freq, class_weight, \nmax_queue_size, workers, use_multiprocessing, shuffle, initial_epoch, \nmode, batch_size, steps_name, **kwargs)\n    262 \n    263       is_deferred = not model._is_compiled\n--&gt; 264       batch_outs = batch_function(*batch_data)\n    265       if not isinstance(batch_outs, list):\n    266         batch_outs = [batch_outs]\n\nC:\\ProgramData\\Anaconda3\\lib\\site- \npackages\\tensorflow\\python\\keras\\engine\\training.py in \ntrain_on_batch(self, x, y, sample_weight, class_weight, reset_metrics)\n   1173       \nself._update_sample_weight_modes(sample_weights=sample_weights)\n   1174       self._make_train_function()\n-&gt; 1175       outputs = self.train_function(ins)  # pylint: disable=not- \n   callable\n   1176 \n   1177     if reset_metrics:\n\nC:\\ProgramData\\Anaconda3\\lib\\site- \npackages\\tensorflow\\python\\keras\\backend.py in __call__(self, inputs)\n   3290 \n   3291     fetched = self._callable_fn(*array_vals,\n-&gt; 3292                                 run_metadata=self.run_metadata)\n   3293     self._call_fetch_callbacks(fetched[-len(self._fetches):])\n   3294     output_structure = nest.pack_sequence_as(\n\nC:\\ProgramData\\Anaconda3\\lib\\site- \npackages\\tensorflow\\python\\client\\session.py in __call__(self, *args, \n**kwargs)\n   1456         ret = \ntf_session.TF_SessionRunCallable(self._session._session,\n   1457                                                self._handle, args,\n-&gt; 1458                                                run_metadata_ptr)\n   1459         if run_metadata:\n   1460           proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\nInvalidArgumentError: 2 root error(s) found.\n  (0) Invalid argument: Input to reshape is a tensor with 6000 values, but \nthe requested shape has 6120\n     [[{{node reshape_99\/Reshape}}]]\n     [[loss_12\/mul\/_3253]]\n  (1) Invalid argument: Input to reshape is a tensor with 6000 values, but \nthe requested shape has 6120\n     [[{{node reshape_99\/Reshape}}]]\n0 successful operations.\n0 derived errors ignored.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57769713,"AnswerCount":1,"Tags":"<python><tensorflow><logging><prefix><verbosity>","CreationDate":"2019-09-03T10:21:18.363","AcceptedAnswerId":57930497.0,"OwnerUserId":5581893.0,"Title":"How to remove the prefix printed out by tf.logging.info","Body":"<p>Printing log lines in TensorFlow is done with functions in tf.logging like tf.logging.info, tf.logging.log, etc.<\/p>\n\n<pre><code>import tensorflow as tf;\n\ntf.logging.set_verbosity(tf.logging.INFO);\ntf.logging.info(tf.__version__);\n<\/code><\/pre>\n\n<p>However, when printing on <strong>Google Colab<\/strong>, there's a real long prefix before the printed value:<\/p>\n\n<pre><code>I0903 10:16:46.097078 140433498875776 &lt;ipython-input-9-a435f70e3dca&gt;:22] 1.14.0\n<\/code><\/pre>\n\n<p>How to change tf.logging to omit that prefix? It's long and unnecessary. And even when running the code directly with <code>python3<\/code> in terminal, there's still a shorter prefix (the version is 1.12.0 in terminal):<\/p>\n\n<pre><code>INFO:tensorflow: 1.12.0\n<\/code><\/pre>\n\n<p>And I want to remove this short prefix too.<\/p>\n","answers":[{"AnswerId":"57930497","CreationDate":"2019-09-13T21:22:18.120","ParentId":null,"OwnerUserId":"5581893","Title":null,"Body":"<p>I finally found the answer that it is impossible to remove the prefix by <code>tf.logging.*<\/code> functions. It is a done in the C\/C++ level of the library and no way to remove it.<\/p>\n\n<p>It is stated in there: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/12537\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/12537<\/a><\/p>\n"}]}
+{"QuestionId":57769714,"AnswerCount":0,"Tags":"<android><python-3.x><performance><tensorflow>","CreationDate":"2019-09-03T10:21:21.913","AcceptedAnswerId":null,"OwnerUserId":8675107.0,"Title":"Very high Inference time for custom object detector trained in SSDMobileNetV2 TensorFlow model in Object Detection Android Demo","Body":"<p>I have used <strong>SSDMobileNetV2<\/strong> for training a custom object detection.<\/p>\n\n<p>OS used : <strong>Windows 10<\/strong>\n<br> Tensorflow Version : <strong>1.13.0<\/strong>\n<br> Size of tflite mode = <strong>3Mb<\/strong><\/p>\n\n<p>The model is successfully deployed in android, it is showing good detection, the only problem is very High inference time. For detect.tflite which is already there, my phone is showing inference time of 50ms to 120ms but, for my model, it is giving up to 500ms.\nI want my app to speedup. <\/p>\n\n<p>I have trained the TensorFlow model in <code>tf version 1.13.0<\/code> .Then exported the graph using <code>export_tflite_ssd_graph.py<\/code> and then converted it using toco,\nFor conversion I have used:<\/p>\n\n<blockquote>\n  <p>toco\n  --graph_def_file=U:\\tensorflow-master\\tflite_lite_V2\\tflite_graph.pb --input_format=TENSORFLOW_GRAPHDEF --output_format=TFLITE --output_file=U:\\tensorflow-master\\tflite_lite_V2\\tflite_graphlite_V2.tflite\n  --inference_type=FLOAT --input_type=FLOAT --input_arrays=normalized_input_image_tensor --output_arrays=TFLite_Detection_PostProcess,TFLite_Detection_PostProcess:1,TFLite_Detection_PostProcess:2,TFLite_Detection_PostProcess:3\n  --input_shapes=1,300,300,3 --allow_custom_ops<\/p>\n<\/blockquote>\n\n<pre><code>\/\/ Configuration values for the prepackaged SSD model.\n  private static final int TF_OD_API_INPUT_SIZE = 300;\n  private static final boolean TF_OD_API_IS_QUANTIZED = true;\n  private static final String TF_OD_API_MODEL_FILE = \"tflite_graphlite_V2_quantized_inf.tflite\";\n  private static final String TF_OD_API_LABELS_FILE = \"file:\/\/\/android_asset\/labels.txt\";\n  private static final DetectorMode MODE = DetectorMode.TF_OD_API;\n<\/code><\/pre>\n\n<p>My script looks like as above.<\/p>\n\n<p>I did not get any error. The only thing coming my way is <strong>inference time<\/strong>. I have to reduce that.<\/p>\n\n<p>Would be thankful for any suggestions.<\/p>\n","answers":[]}
+{"QuestionId":57769905,"AnswerCount":0,"Tags":"<python><pandas><tensorflow><keras><pickle>","CreationDate":"2019-09-03T10:33:26.173","AcceptedAnswerId":null,"OwnerUserId":10268559.0,"Title":"Saving TensorFlow model (ensemble) to one pickle file, instead of multiple pickles, in Python","Body":"<p>I am looking to deploy a 50 model ensemble for a regression problem, with each model being a Keras.Sequential Neural Network.<\/p>\n\n<p>Below is a (simplified to 3 models) version of my code, that runs and works fine. <\/p>\n\n<p>However, I don't want to create a pickle file for every individual model, and so is there a way of creating a class with a list of all the models, resulting in me having to save\/load only one pickle file?<\/p>\n\n<pre><code>from __future__ import absolute_import, division, print_function\nimport tensorflow as tf\nfrom tensorflow import keras\nimport pandas as pd\nimport numpy as np\n\ntrain = pd.read_csv(\"Training Data.csv\").fillna(0)\nX_train = train.drop(['ID_NUMBER','DATE','X','Y','Z'],axis=1)\nY_train = train[['X','Y','Z']]\n\nEPOCHS = 1500\nBATCH_SIZE = 256\n\n#Defining the 3 layered Neural Network\ndef build_model():\n    model = keras.Sequential([\n    keras.layers.Dense(1000, activation=tf.nn.softplus,\n                       input_shape=(X_train.shape[1],)),\n    keras.layers.Dense(500, activation=tf.nn.softplus),\n    keras.layers.Dense(3)\n    ])\n\n    model.compile(loss='mse',optimizer='adam', metrics=['mse'])\n    return model\n\nmodel0 = build_model()\n# Store training stats\nhistory0 = model0.fit(X_train, Y_train, epochs=EPOCHS, batch_size=BATCH_SIZE,\n                validation_split=0.0, verbose=1)\n\nmodel1 = build_model()\n# Store training stats\nhistory1 = model1.fit(X_train, Y_train, epochs=EPOCHS, batch_size=BATCH_SIZE,\n                validation_split=0.0, verbose=1)\n\nmodel2 = build_model()\n# Store training stats\nhistory2 = model2.fit(X_train, Y_train, epochs=EPOCHS, batch_size=BATCH_SIZE,\n                validation_split=0.0, verbose=1)\n\nmodel0.save(\"model0.pkl\")\nmodel1.save(\"model1.pkl\")\nmodel2.save(\"model2.pkl\")\n<\/code><\/pre>\n\n<p>For making new predictions, my code would look something like this:<\/p>\n\n<pre><code>#Loading Models\nmodel0 = tf.keras.models.load_model(\"model0.pkl\")\nmodel1 = tf.keras.models.load_model(\"model1.pkl\")\nmodel2 = tf.keras.models.load_model(\"model2.pkl\")\n\n#Finding Weights (based on train score)\n\ntrain_nn_predictions = model0.predict(X_train)\ntrain['X'],train['Y'],train['Z'] = train_nn_predictions[:,0],train_nn_predictions[:,1],train_nn_predictions[:,2]     \nnn0 = #training score metric (irrelevant to show how it is calculated here)\nprint(\"Average Train Score for Model 0 is:\",nn0)\n\ntrain_nn_predictions = model1.predict(X_train)\ntrain['X'],train['Y'],train['Z'] = train_nn_predictions[:,0],train_nn_predictions[:,1],train_nn_predictions[:,2]     \nnn1 = #training score metric (irrelevant to show how it is calculated here)\nprint(\"Average Train Score for Model 1 is:\",nn1)\n\ntrain_nn_predictions = model2.predict(X_train)\ntrain['X'],train['Y'],train['Z'] = train_nn_predictions[:,0],train_nn_predictions[:,1],train_nn_predictions[:,2]     \nnn2 = #training score metric (irrelevant to show how it is calculated here)\nprint(\"Average Train Score for Model 2 is:\",nn2)\n\n#Apply the weightings for each of the models\nw0,w1,w2 = 1\/nn0,1\/nn1,1\/nn2\n\n#New Predictions\nnew_record = np.array([my variables])\ntarget_predictions = (w0*model0.predict(new_record)+w1*model1.predict(new_record)+w2*model2.predict(new_record))\/(w0+w1+w2)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57770047,"AnswerCount":0,"Tags":"<python><keras><overriding><tensor>","CreationDate":"2019-09-03T10:42:37.437","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"Overriding Keras dot function","Body":"<p>I am writting a new matrix multiplaction and need to override Keras dot funtion to see possible result in the code downloaded from <a href=\"https:\/\/keras.io\/examples\/mnist_cnn\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/examples\/mnist_cnn\/<\/a>.\nRight now, My code is just using tf to calculate resuly and printing type of input variables.But it does not workd very well,<\/p>\n\n<p>Would you be able to give me a clue how can I solve or any hint please?<\/p>\n\n<pre><code>from keras import backend as K\nimport tensorflow as tf\ndef mul(A,B,C = None):\n        print('A is :',A)\n        print('B is :',B)\n\n        C = tf.math.multiply (A,B)\n        print('C is :',C)\n        return  C\n\nK.dot = mul               #Moneky patch\n<\/code><\/pre>\n\n<p>My outputs are :<\/p>\n\n<blockquote>\n  <p>A is : <code>Tensor(\"flatten_12\/Reshape:0\", shape=(?, ?), dtype=float32)<\/code> <br>\n  B is : <code>&lt;tf.Variable 'dense_20\/kernel:0' shape=(9216, 128) &gt;dtype=float32_ref&gt;<\/code><br>\n  C is : <code>Tensor(\"dense_20\/Mul:0\", shape=(9216, 128), dtype=float32)<\/code><br>\n   <br>\n  A is : <code>Tensor(\"dropout_21\/cond\/Merge:0\", shape=(9216, 128), dtype=float32)<\/code> <br>\n  B is : <code>&lt;tf.Variable 'dense_21\/kernel:0' shape=(128, 10) dtype=float32_ref&gt;<\/code><\/p>\n<\/blockquote>\n\n<p>I get this error: <\/p>\n\n<blockquote>\n  <p>\"Dimensions must be equal, but are 9216 and 128 for 'dense_22\/mul' (op: 'Mul') with input shapes: [9216,128], [128,10]\"<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":57770082,"AnswerCount":0,"Tags":"<python><keras><coremltools><onnx>","CreationDate":"2019-09-03T10:44:28.870","AcceptedAnswerId":null,"OwnerUserId":12013917.0,"Title":"Variation in keras and mlmodel results","Body":"<p>I made a keras model for next word prediction. When i exported it to mlmodel and used it for text prediction, the results of the keras and the mlmodel varied. I used to Netron to visualize the model. I realized that the weights of some layers have changed during the conversion. But I had referred the official coreml documentation for this. Can you help me with this issue?. Thanks in advance.<\/p>\n\n<p>When I saw difference in weights and predictions, I tried using onnx to convert the keras to onnx and onnx to mlmodel. But there was this error: \"Unsupported ONNX ops of type: Cast,Gather,Shape,Identity\" while converting the onnx model to mlmodel.<\/p>\n\n<p>This is my LSTM model:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Embedding(vocab_size, 10, input_length=max_length-1))\nmodel.add(LSTM(100, return_sequences=True))\nmodel.add(LSTM(50))\nmodel.add(Dense(100, activation='relu'))\nmodel.add(Dense(vocab_size, activation='softmax'))\nprint(model.summary())\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', \n metrics=['accuracy'])\nmodel.fit(X, y, epochs=100, verbose=2)\n<\/code><\/pre>\n\n<p>The prediction function:<\/p>\n\n<pre><code>def generate_seq(model, tokenizer, seq_length, seed_text, n_words):\n result = list()\n #n_words is the number of predictions I want\n in_text = seed_text\n # generate a fixed number of words\n for _ in range(n_words):\n    # encode the text as integer\n    encoded = tokenizer.texts_to_sequences([in_text])[0]\n    # truncate sequences to a fixed length\n    encoded = pad_sequences([encoded], maxlen=seq_length, truncating='pre')\n    y_prob=model.predict(encoded)\n\n            #sort out the top n_words in order of the highest prob\n    top_indices = y_prob.argsort()[0][-n_words:][::-1]\n\n\n    # map predicted word index to word\n    out_word = ''\n    for top_index in top_indices:\n        for word, index in tokenizer.word_index.items():\n            if index == top_index:\n                out_word+= word+' '\n                        break\n\n return out_word\n<\/code><\/pre>\n\n<p>The keras to coreml conversion code:<\/p>\n\n<pre><code>from keras.models import load_model\nimport coremltools\n\nmodel = coremltools.converters.keras.convert(\"model_train.h5\")\nmodel.save(\"model_train.mlmodel\")\n<\/code><\/pre>\n\n<p>The results are totally variant. For e.g: For the input: \"I read\",the keras model predicts \"it,book,this\" whereas the mlmodel predicts \"thought,had,was\".<\/p>\n\n<p>These are the weights of the embedding layer of my keras and mlmodel.\n<a href=\"https:\/\/www.filehosting.org\/file\/details\/820874\/weightsKerasModel.npy\" rel=\"nofollow noreferrer\">Keras Model<\/a>\n<a href=\"https:\/\/www.filehosting.org\/file\/details\/820875\/weightsMLMODEL.npy\" rel=\"nofollow noreferrer\">ML Model<\/a><\/p>\n","answers":[]}
+{"QuestionId":57770090,"AnswerCount":0,"Tags":"<java><android><tensorflow><firebase-mlkit>","CreationDate":"2019-09-03T10:44:43.017","AcceptedAnswerId":null,"OwnerUserId":11856751.0,"Title":"How to implement the function of getYourInputImage()?","Body":"<p>I'm working on eye detection project and I have trained my own model using Tensorflow Library, I have now the Tensorflow lite model.<\/p>\n\n<p>I followed the documentation proposed by Tensorflow for how to integrate custom model MLKit firebase on the Android platform.<\/p>\n\n<p>This is <code>CustomModelActivity.java<\/code> that I followed<\/p>\n\n<p><a href=\"https:\/\/github.com\/firebase\/snippets-android\/blob\/a7be1d9b6208d112d1bacbfff937c32c219124b5\/mlkit\/app\/src\/main\/java\/com\/google\/firebase\/example\/mlkit\/CustomModelActivity.java#L127-L128\" rel=\"nofollow noreferrer\">https:\/\/github.com\/firebase\/snippets-android\/blob\/a7be1d9b6208d112d1bacbfff937c32c219124b5\/mlkit\/app\/src\/main\/java\/com\/google\/firebase\/example\/mlkit\/CustomModelActivity.java#L127-L128<\/a><\/p>\n\n<p>I tried to implement this function by setting the image on imageview and I have tried to convert it to bitmap \nBut I have encountered some issue ( false predictions ) on an image from the dataset.<\/p>\n\n<p>I have posted many questions about this problem but I didn't get any answer.<\/p>\n\n<p>I would like now to load the image from my gallery and get it, My question is:\nHow can I implement the last function <code>getYouInputImage()<\/code> ?<\/p>\n\n<pre class=\"lang-java prettyprint-override\"><code> private Bitmap getYourInputImage() {\n        \/\/ This method is just for show\n        return Bitmap.createBitmap(0, 0, Bitmap.Config.ALPHA_8);\n    }\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57770967,"AnswerCount":0,"Tags":"<python-3.x><deep-learning><conv-neural-network><pytorch>","CreationDate":"2019-09-03T11:39:24.427","AcceptedAnswerId":null,"OwnerUserId":11480928.0,"Title":"Problem in getting accuracy of different classes in image classification","Body":"<p>I am doing an assignment related to  image classification.My dataset has 5 different classes(labels).I have to calculate accuracy of each class(labels) separately.But my problem is that whenever I tested my CNN model I obtained 100% accuracy for one label and accuracy of other labels are 0%. I could not make sense where is my implementation problem.<\/p>\n\n<p>here is my CNN model <\/p>\n\n<pre><code>class ConvNet(nn.Module):\n    def __init__(self):\n        super(ConvNet, self).__init__()\n        self.layer1 = nn.Sequential(\n            nn.Conv2d(1, 16, kernel_size=3,stride=1,padding=1),\n            nn.ReLU(),\n            nn.MaxPool2d(kernel_size=2, stride=2))\n        self.layer2 = nn.Sequential(\n            nn.Conv2d(16, 32, kernel_size=3,stride=1,padding=1),\n            nn.ReLU(),\n            nn.MaxPool2d(kernel_size=2, stride=2))\n        self.drop_out = nn.Dropout()\n        self.fc1 = nn.Linear(in_features=16 * 16 * 32, out_features=256)\n        self.fc2 = nn.Linear(in_features=256, out_features=5)\n        #self.fc3 = nn.Linear(in_features=64, out_features=5)\ndef forward(self, x):\n    out = self.layer1(x)\n    out = self.layer2(out)\n    #out = out.view(out.size(0), -1)\n    out = out.reshape(-1,16 * 16 * 32)\n    out = self.drop_out(out)\n    out = self.fc1(out)\n    out = self.fc2(out)\n    #out = self.fc3(out)\n    return out\n<\/code><\/pre>\n\n<p>And here is the code that i used for calculating accuracy of class labels <\/p>\n\n<pre><code>classes = ['1', '2', '3', '4', '5']\n\nclass_correct = list(0. for i in range(5))\nclass_total = list(0. for i in range(5))\n\nnet.eval()  # prep model for evaluation\n\nfor batch in test_loader:\n    data, target = batch['image'], batch['grade']\n\n    if len(target.data) != BS:\n        break\n    # # forward pass: compute predicted outputs by passing inputs to the model\n    output = net(data)\n    # convert output probabilities to predicted class\n\n    _, pred = torch.max(output, 1)\n    # # compare predictions to true label\n    correct = np.squeeze(pred.eq(target.data.view_as(pred)))\n    # # calculate test accuracy for each object class\n    for i in range(BS):\n        label = target.data[i]\n        class_correct[label] += correct[i].item()\n        class_total[label] += 1\n\nfor i in range(5):\n    print('Test Accuracy of %5s: %2d%% (%2d\/%2d)' % (\n        classes[i], 100 * class_correct[i] \/ class_total[i],\n        np.sum(class_correct[i]), np.sum(class_total[i])))\n<\/code><\/pre>\n\n<p>This is my results that I obtained <\/p>\n\n<pre><code>Accuracy of the network (overall): 46 %\nTest Accuracy of     1: 100% (652\/652)\nTest Accuracy of     2:  0% ( 0\/279)\nTest Accuracy of     3:  0% ( 0\/459)\nTest Accuracy of     4:  0% ( 0\/263)\nTest Accuracy of     5:  0% ( 0\/47)\n<\/code><\/pre>\n\n<p>For evaluating my model I also plotted ROC curve and precision recall curve that I posted here<br>\n<a href=\"https:\/\/i.stack.imgur.com\/8M4vT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/8M4vT.png\" alt=\"ROC \"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/z9CvR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/z9CvR.png\" alt=\"Precision recall curve\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57771565,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-03T12:15:12.873","AcceptedAnswerId":null,"OwnerUserId":10013648.0,"Title":"Unexpected \"Stop Iteration\" when running tensorflow-gpu","Body":"<p>I've created a CNN model for the Galaxy Zoo project as part of a college project. I've created the training and validation data using a batch generator as the dataset is very large. I trained it over CPU with no problems. Recently, I switched over to Tensorflow-gpu and have been training it over GPU which is exponentially faster. But the Model stops unexpectedly when it reaches it's 33rd epoch. I've linked the code. Please help.<\/p>\n\n<p>I've tried with 10 epochs, 20 and so on. It works fine up to the 32nd epoch but it stops when it completes it 33rd epoch.<\/p>\n\n<pre><code>#Code for Image processor and Batch generator:\n\ndef img_processor(path):\n    img = cv2.imread(path)\n    img = img[106:106*3,106:106*3,:]\n    img = cv2.resize(img,(106,106), interpolation = cv2.INTER_CUBIC)\n    return img\n\ndef Batch_generator(DIR):\n    for img in os.listdir(DIR):\n        im = img_processor(os.path.join(DIR,img))\n        ind = int(img[:-4])\n        y_train = df.loc[ind].values\n        X_train = np.array(im)\n        X_train = X_train.reshape(1,106,106,3)\n        y_train = y_train.reshape(1,37)\n        yield(X_train,y_train)\n\ndef Validation_generator(DIR):\n    for img in os.listdir(DIR):\n        im = img_processor(os.path.join(DIR,img))\n        ind = int(img[:-4])\n        y_valid = df.loc[ind].values\n        X_valid = np.array(im)\n        X_valid = X_valid.reshape(1,106,106,3)\n        y_valid = y_valid.reshape(1,37)\n        yield(X_valid,y_valid)\n\n\n#Fitting the model:\nbatch_size = 32\n\nsteps_per_batch = int(len(os.listdir(train_path))\/batch_size)\nval_steps_per_batch = int(len(os.listdir(validate_path))\/batch_size)\n\nhistory = model.fit_generator(\n                    Batch_generator(train_path),\n                    samples_per_epoch=steps_per_batch,\n                    epochs = 70,\n                    validation_data=Validation_generator(validate_path),\n                    nb_val_samples=val_steps_per_batch,\n                    verbose = 1,\n                    callbacks = [tensorboard]\n                   )\n<\/code><\/pre>\n\n<h1>Problem:<\/h1>\n\n<pre><code>Epoch 31\/70\n1924\/1924 [==============================] - 173s 90ms\/step - loss: 0.0220 - acc: 0.5541 - val_loss: 0.0196 - val_acc: 0.4969\n\nEpoch 32\/70\n1924\/1924 [==============================] - 173s 90ms\/step - loss: 0.0217 - acc: 0.5660 - val_loss: 0.0187 - val_acc: 0.5825\n\nEpoch 33\/70\n  10\/1924 [..............................] - ETA: 2:20 - loss: 0.0270 - acc: 0.3000\n---------------------------------------------------------------------------\nStopIteration                             Traceback (most recent call last)\n&lt;ipython-input-9-fccdb670230f&gt; in &lt;module&gt;\n      6                     nb_val_samples=val_steps_per_batch,\n      7                     verbose = 1,\n----&gt; 8                     callbacks = [tensorboard]\n      9                    )\n\nc:\\users\\shinigami shrek\\appdata\\local\\programs\\python\\python36\\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\nc:\\users\\shinigami shrek\\appdata\\local\\programs\\python\\python36\\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\nc:\\users\\shinigami shrek\\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    179             batch_index = 0\n    180             while steps_done &lt; steps_per_epoch:\n--&gt; 181                 generator_output = next(output_generator)\n    182 \n    183                 if not hasattr(generator_output, '__len__'):\n\nStopIteration:\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57771578,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-09-03T12:16:14.267","AcceptedAnswerId":57772139.0,"OwnerUserId":7896153.0,"Title":"Python with Keras Dataset Error, \"X is not defined\"","Body":"<p>I am trying to load a dataset called cifar100 from Keras, and edit it with the function <code>X = X[y.flatten() == 8]<\/code>, but I keep getting the Error \"X is not defined\".<\/p>\n\n<pre><code>from keras.datasets import cifar100\n\n(x_train, y_train), (x_test, y_test) = cifar100.load_data(label_mode='fine')\n\nX = X[y.flatten() == 8]\n<\/code><\/pre>\n","answers":[{"AnswerId":"57772139","CreationDate":"2019-09-03T12:50:27.493","ParentId":null,"OwnerUserId":"6770027","Title":null,"Body":"<p>X is not defined by you.\nYou are using <code>X<\/code> instead of <code>x_train<\/code> and <code>y<\/code> instead of <code>y_train<\/code>\nmay be you want to use :<\/p>\n\n<pre><code>x_train[y_train.flatten() == 8]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57771706,"AnswerCount":2,"Tags":"<python><tensorflow><keras><neural-network><classification>","CreationDate":"2019-09-03T12:24:12.707","AcceptedAnswerId":57775320.0,"OwnerUserId":12010696.0,"Title":"Why is accuracy and loss staying exactly the same while training?","Body":"<p>So I have tried to modify the entry-tutorial from <a href=\"https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_classification\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/keras\/basic_classification<\/a>, to work with my own data. The goal is to classify images of dogs and cats. The code is very simple and given below. The problem is that the network does not seem to learn at all, training loss and accuracy stay the same after every epoch.<\/p>\n\n<p>The images (X_training) and the labels (y_training) seem to have the right format:\n<code>X_training.shape<\/code> returns: <code>(18827, 80, 80, 3)<\/code><\/p>\n\n<p><code>y_training<\/code> is a one dimensional list with entries in {0,1}<\/p>\n\n<p>I have checked several times, that the \"images\" in <code>X_training<\/code> are correctly labeled: \nLet's say <code>X_training[i,:,:,:]<\/code> represents a dog, then <code>y_training[i]<\/code> will return a 1, if <code>X_training[i,:,:,:]<\/code> represents a cat, then <code>y_training[i]<\/code> will return a 0.<\/p>\n\n<p>Shown below is the complete python file without the import statements.<\/p>\n\n<pre><code>#loading the data from 4 pickle files:\npickle_in = open(\"X_training.pickle\",\"rb\")\nX_training = pickle.load(pickle_in)\n\npickle_in = open(\"X_testing.pickle\",\"rb\")\nX_testing = pickle.load(pickle_in)\n\npickle_in = open(\"y_training.pickle\",\"rb\")\ny_training = pickle.load(pickle_in)\n\npickle_in = open(\"y_testing.pickle\",\"rb\")\ny_testing = pickle.load(pickle_in)\n\n\n#normalizing the input data:\nX_training = X_training\/255.0\nX_testing = X_testing\/255.0\n\n\n#building the model:\nmodel = keras.Sequential([\n    keras.layers.Flatten(input_shape=(80, 80,3)),\n    keras.layers.Dense(128, activation=tf.nn.relu),\n    keras.layers.Dense(1,activation='sigmoid')\n])\nmodel.compile(optimizer='adam',loss='mean_squared_error',metrics=['accuracy'])\n\n\n#running the model:\nmodel.fit(X_training, y_training, epochs=10)\n<\/code><\/pre>\n\n<p>The code compiles and trains for 10 epochs, but neither loss nor accuracy improve, they stay exactly the same after every epoch.\nThe code works fine with the MNIST-fashion dataset used in the tutorial with slight changes accounting for the difference in multiclass vs binary classification and input shape.<\/p>\n","answers":[{"AnswerId":"57772330","CreationDate":"2019-09-03T13:01:52.663","ParentId":null,"OwnerUserId":"8563649","Title":null,"Body":"<p>if you want to train a classification model you must have binary_crossentropy as you lost function and not <code>mean_squared_error<\/code> which is used for regression tasks<\/p>\n\n<p>replace<\/p>\n\n<pre><code>model.compile(optimizer='adam',loss='mean_squared_error',metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>with<\/p>\n\n<pre><code>model.compile(optimizer='adam',loss='binary_crossentropy',metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>Furthermore i would recommend not using <code>relu<\/code> activation on your dense layer but <code>linear<\/code><\/p>\n\n<p>replace<\/p>\n\n<pre><code>keras.layers.Dense(128, activation=tf.nn.relu),\n<\/code><\/pre>\n\n<p>with<\/p>\n\n<pre><code>keras.layers.Dense(128),\n<\/code><\/pre>\n\n<p>and of cource to better use the power of neural networks use some <code>convolutional layers<\/code> prior your <code>flatten layer<\/code><\/p>\n"},{"AnswerId":"57775320","CreationDate":"2019-09-03T16:10:31.730","ParentId":null,"OwnerUserId":"12010696","Title":null,"Body":"<p>I have found a different implementation with a slightly more complex model that works.\nHere is the complete code without the import statements:<\/p>\n\n<pre><code>#global variables:\nbatch_size = 32\nnr_of_epochs = 64\ninput_shape = (80,80,3)\n\n\n#loading the data from 4 pickle files:\npickle_in = open(\"X_training.pickle\",\"rb\")\nX_training = pickle.load(pickle_in)\n\npickle_in = open(\"X_testing.pickle\",\"rb\")\nX_testing = pickle.load(pickle_in)\n\npickle_in = open(\"y_training.pickle\",\"rb\")\ny_training = pickle.load(pickle_in)\n\npickle_in = open(\"y_testing.pickle\",\"rb\")\ny_testing = pickle.load(pickle_in)\n\n\n\n#building the model\ndef define_model():\n    model = Sequential()\n    model.add(Conv2D(32, (3, 3), activation='relu', input_shape=input_shape))\n    model.add(MaxPooling2D((2, 2)))\n    model.add(Flatten())\n    model.add(Dense(128, activation='relu'))\n    model.add(Dense(1, activation='sigmoid'))\n    # compile model\n    model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])\n    return model\nmodel = define_model()\n\n\n#Possibility for image data augmentation\ntrain_datagen = ImageDataGenerator(rescale=1.0\/255.0)\nval_datagen = ImageDataGenerator(rescale=1.\/255.) \ntrain_generator =train_datagen.flow(X_training,y_training,batch_size=batch_size)\nval_generator = val_datagen.flow(X_testing,y_testing,batch_size= batch_size)\n\n\n\n#running the model\nhistory = model.fit_generator(train_generator,steps_per_epoch=len(X_training) \/\/batch_size,\n                              epochs=nr_of_epochs,validation_data=val_generator,\n                              validation_steps=len(X_testing) \/\/batch_size)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57771739,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><keras><conv-neural-network>","CreationDate":"2019-09-03T12:26:39.337","AcceptedAnswerId":null,"OwnerUserId":11658889.0,"Title":"What is the difference between 'classifer.add' and 'keras.layers'?","Body":"<p>I am following a simple set of tutorials making convolutional neural networks in keras. I had two similar tutorials and wanted to make the code of the second a bit neater, so replaced the model creation style of the first tutorial with that of the second. It then gave me syntax errors. Is there a functional difference between adding a new layer in either of these styles? Do they have different use cases?<\/p>\n\n<p>Code A<\/p>\n\n<pre><code>model = keras.Sequential([\n    keras.layers.Dense(128, activation=tf.nn.relu),\n])\n<\/code><\/pre>\n\n<p>Code B<\/p>\n\n<pre><code>classifier = Sequential()\n\nclassifier.add(Dense(units = 128, activation = 'relu'))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57771825,"AnswerCount":1,"Tags":"<tensorflow><object-detection>","CreationDate":"2019-09-03T12:31:38.893","AcceptedAnswerId":null,"OwnerUserId":5061927.0,"Title":"How is IOU is calculated for multiple bounding box predictions in Tensorflow Object Detection API?","Body":"<p>How is IOU is calculated for multiple bounding box predictions in Tensorflow Object Detection API ?<\/p>\n","answers":[{"AnswerId":"57788546","CreationDate":"2019-09-04T12:36:46.513","ParentId":null,"OwnerUserId":"6129428","Title":null,"Body":"<p>Each bounding box around an object has an IoU (intersection over union) with the ground-truth box of that object. It is calculated by dividing the common area (overlap) between the predicted bounding box and the actual correct (ground-truth box) by the cumulative area of the two boxes. After calculating all the IoUs for the boxes around an object, the ones with the highest IoU are selected as the result. <a href=\"https:\/\/www.pyimagesearch.com\/2016\/11\/07\/intersection-over-union-iou-for-object-detection\/\" rel=\"nofollow noreferrer\">Here<\/a> it is explained better.<\/p>\n\n<p>Also you can print the IoU value after <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/fe748d4a4a1576b57c279014ac0ceb47344399c4\/research\/object_detection\/core\/post_processing.py#L74\" rel=\"nofollow noreferrer\">this line<\/a>.<\/p>\n"}]}
+{"QuestionId":57771959,"AnswerCount":0,"Tags":"<python><pytorch>","CreationDate":"2019-09-03T12:39:39.337","AcceptedAnswerId":null,"OwnerUserId":10855529.0,"Title":"How do I retrieve name of a layer inside hook function?","Body":"<p>I have a neural network<\/p>\n\n<pre><code>class ConvNet(nn.Module):\n  def __init__(self):\n    super().__init__()\n    self.trunk = nn.ModuleList()\n    self.trunk.add_module('conv1', nn.Conv2d(3, 10, 3))\n\n    self.classifier = nn.Linear(30, 2)\n  def forward(self, x):\n    out = self.classifier(self.trunk.conv1(x))\n    return out\n\nmodel = ConvNet()\n<\/code><\/pre>\n\n<p>I registered forward hook<\/p>\n\n<pre><code>def hook(module, input, output):\n    print(module, input[0].shape, output.shape)\n\nx =  model.trunk.conv1.register_forward_hook(hook)\n<\/code><\/pre>\n\n<p>How do I retrieve name of this layer that is 'conv1' inside hook function, <code>module._get_name<\/code> returns <code>Conv2d<\/code>, <code>module.__class__<\/code> returns <code>&lt;class 'torch.nn.modules.conv.Conv2d'&gt;<\/code>, how do I get just 'conv1'?<\/p>\n","answers":[]}
+{"QuestionId":57772268,"AnswerCount":0,"Tags":"<tensorflow><object-detection-api>","CreationDate":"2019-09-03T12:58:06.623","AcceptedAnswerId":null,"OwnerUserId":3730549.0,"Title":"Understanding the grid_anchor_generator strides for Faster RCNNs in TF Object Detection API","Body":"<p>I am experiencing a performance decrease of factor 10 when I am setting the grid_anchor_generator strides to<\/p>\n\n<pre><code>   height_stride: 8\n   width_stride: 8\n<\/code><\/pre>\n\n<p>So i tried to understand the anchor calculation and came to the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/9bbf8015dba2133ab2343ec6d6b5096033504e36\/research\/object_detection\/anchor_generators\/grid_anchor_generator.py\" rel=\"nofollow noreferrer\">grid_anchor_generator.py<\/a>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def _generate(self, feature_map_shape_list):\n\"\"\"Generates a collection of bounding boxes to be used as anchors.\n    Args:\n      feature_map_shape_list: list of pairs of convnet layer resolutions in the\n        format [(height_0, width_0)].  For example, setting\n        feature_map_shape_list=[(8, 8)] asks for anchors that correspond\n        to an 8x8 layer.  For this anchor generator, only lists of length 1 are\n        allowed.\n    Returns:\n      boxes_list: a list of BoxLists each holding anchor boxes corresponding to\n        the input feature map shapes.\n\n[\u2026]\n\n\"\"\"\n[\u2026]\n    grid_height, grid_width = feature_map_shape_list[0]\n    scales_grid, aspect_ratios_grid = ops.meshgrid(self._scales,\n                                                   self._aspect_ratios)\n\n    scales_grid = tf.reshape(scales_grid, [-1])\n    aspect_ratios_grid = tf.reshape(aspect_ratios_grid, [-1])\n    anchors = tile_anchors(grid_height,\n                           grid_width,\n                           scales_grid,\n                           aspect_ratios_grid,\n                           self._base_anchor_size,\n                           self._anchor_stride,\n                           self._anchor_offset)\n<\/code><\/pre>\n\n<p>The generate function takes the feature map, which is as I understand the input map for the RPN. The length and the width of this map forms a grid which will be used in <strong>tile_anchors<\/strong>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def tile_anchors(grid_height,\n                 grid_width,\n                 scales,\n                 aspect_ratios,\n                 base_anchor_size,\n                 anchor_stride,\n                 anchor_offset):\n  \"\"\"Create a tiled set of anchors strided along a grid in image space.\nThis op creates a set of anchor boxes by placing a \"basis\" collection of\n  boxes with user-specified scales and aspect ratios centered at evenly\n  distributed points along a grid.  The basis collection is specified via the\n  scale and aspect_ratios arguments.  For example, setting scales=[.1, .2, .2]\n  and aspect ratios = [2,2,1\/2] means that we create three boxes: one with scale\n  .1, aspect ratio 2, one with scale .2, aspect ratio 2, and one with scale .2\n  and aspect ratio 1\/2.  Each box is multiplied by \"base_anchor_size\" before\n  placing it over its respective center.\n  Grid points are specified via grid_height, grid_width parameters as well as\n  the anchor_stride and anchor_offset parameters.\n  Args:\n    grid_height: size of the grid in the y direction (int or int scalar tensor)\n    grid_width: size of the grid in the x direction (int or int scalar tensor)\n    scales: a 1-d  (float) tensor representing the scale of each box in the\n      basis set.\n    aspect_ratios: a 1-d (float) tensor representing the aspect ratio of each\n      box in the basis set.  The length of the scales and aspect_ratios tensors\n      must be equal.\n    base_anchor_size: base anchor size as [height, width]\n      (float tensor of shape [2])\n    anchor_stride: difference in centers between base anchors for adjacent grid\n                   positions (float tensor of shape [2])\n    anchor_offset: center of the anchor with scale and aspect ratio 1 for the\n                   upper left element of the grid, this should be zero for\n                   feature networks with only VALID padding and even receptive\n                   field size, but may need some additional calculation if other\n                   padding is used (float tensor of shape [2])\n  Returns:\n    a BoxList holding a collection of N anchor boxes\n\"\"\"\nratio_sqrts = tf.sqrt(aspect_ratios)\n  heights = scales \/ ratio_sqrts * base_anchor_size[0]\n  widths = scales * ratio_sqrts * base_anchor_size[1]\n\n  # Get a grid of box centers\n  y_centers = tf.cast(tf.range(grid_height), dtype=tf.float32)\n  y_centers = y_centers * anchor_stride[0] + anchor_offset[0]\n  x_centers = tf.cast(tf.range(grid_width), dtype=tf.float32)\n  x_centers = x_centers * anchor_stride[1] + anchor_offset[1]\n  x_centers, y_centers = ops.meshgrid(x_centers, y_centers)\n\n  widths_grid, x_centers_grid = ops.meshgrid(widths, x_centers)\n  heights_grid, y_centers_grid = ops.meshgrid(heights, y_centers)\n  bbox_centers = tf.stack([y_centers_grid, x_centers_grid], axis=3)\n  bbox_sizes = tf.stack([heights_grid, widths_grid], axis=3)\n  bbox_centers = tf.reshape(bbox_centers, [-1, 2])\n  bbox_sizes = tf.reshape(bbox_sizes, [-1, 2])\n  bbox_corners = _center_size_bbox_to_corners_bbox(bbox_centers, bbox_sizes)\n  return box_list.BoxList(bbox_corners)\n[\u2026]\n<\/code><\/pre>\n\n<p>If the strides are set smaller, as I did, the x_ and y_center values will also decrease. I have some rudimentary tensorflow skills and so I'm probably missing something: Where does the mapping between the x\/y_centers and the pixel values is caluclated? <\/p>\n\n<p>I tried to test it with some np equivalent<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\n    stride = 16 # or 8\n    scales = np.array([1])\n    aspect_ratios = np.array([1])\n    ratio_sqrts = math.sqrt(aspect_ratios)\n    heights = scales \/ ratio_sqrts * 255\n    widths = scales * ratio_sqrts * 255\n\n    x = np.array([0,1,2,3])\n    x = x * stride\n    y = np.array([0,1,2])\n    y = y * stride\n    x_centers, y_centers = np.meshgrid(x, y)\n    widths_grid, x_centers_grid = np.meshgrid(widths, x_centers)\n\n    heights_grid, y_centers_grid = np.meshgrid(heights, y_centers)\n    bbox_centers = np.stack([y_centers_grid, x_centers_grid], axis=2)\n\n    print(bbox_centers)\n\n<\/code><\/pre>\n\n<p>with stride=16, I get:<\/p>\n\n<pre><code>[[[ 0  0]]\n [[ 0 16]]\n [[ 0 32]]\n [[ 0 48]]\n [[16  0]]\n [[16 16]]\n [[16 32]]\n [[16 48]]\n [[32  0]]\n [[32 16]]\n [[32 32]]\n [[32 48]]\n<\/code><\/pre>\n\n<p>amd with stride=8, I get:<\/p>\n\n<pre><code>[[[ 0  0]]\n [[ 0  8]]\n [[ 0 16]]\n [[ 0 24]]\n [[ 8  0]]\n [[ 8  8]]\n [[ 8 16]]\n [[ 8 24]]\n [[16  0]]\n [[16  8]]\n [[16 16]]\n [[16 24]]]\n<\/code><\/pre>\n\n<p>(it should be 4times longer, shouldn't it?)<\/p>\n","answers":[]}
+{"QuestionId":57772572,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-09-03T13:16:20.973","AcceptedAnswerId":null,"OwnerUserId":8366805.0,"Title":"What is the benefit of using custom data generator in Keras?","Body":"<p>I learned three ways to feed data to a <code>keras<\/code> model for training. <\/p>\n\n<p>The first way:<\/p>\n\n<p><code>model1.fit(x_examples, y_examples, batch_size=10, epochs=2, shuffle=True)<\/code><\/p>\n\n<p>The second way, use a function:<\/p>\n\n<pre><code>def data_generator(batch_size=10):\n    indices = np.arange(len(x_examples)); n = len(x_examples)\n    i = 0\n    while True:\n        x_data, y_data = [], []\n        for b in range(batch_size):\n            if i==0: np.random.shuffle(indices)\n            x_data.append(x_examples[i]); y_data.append(y_examples[i])\n            i = (i+1) % n\n        x_data = np.array(x_data); y_data = np.array(y_data)\n        yield x_data, y_data\n\nmodel2.fit_generator(generator=data_generator(), \n    steps_per_epoch=len(x_examples)\/\/batch_size, epochs=2)\n<\/code><\/pre>\n\n<p>The third way, use a class:<\/p>\n\n<pre><code>class DataGenerator(keras.utils.Sequence):\n    def __init__(self, batch_size=10, shuffle=True):\n        super().__init__()\n        self.batch_size = batch_size\n        self.shuffle = shuffle\n        self.on_epoch_end()\n    def __len__(self):\n        return int(np.floor(len(x_examples) \/ self.batch_size))\n    def on_epoch_end(self):\n        self.indexes = np.arange(len(x_examples))\n        if self.shuffle == True:\n            np.random.shuffle(self.indexes)\n    def __getitem__(self, index):\n        indexes = self.indexes[index * self.batch_size: (index + 1) * self.batch_size]\n        X = x_examples[indexes]; y = y_examples[indexes]\n        return X, y\n\nmodel3.fit_generator(generator=DataGenerator(), epochs=2)\n<\/code><\/pre>\n\n<p>Are the three approaches equivalent? By running a simple test, where I have fixed the random states by<\/p>\n\n<pre><code>numpy.random.seed(0)\ntensorflow.set_random_seed(0)\n<\/code><\/pre>\n\n<p>The losses of the three identical models (<code>model1<\/code>, <code>model2<\/code>, <code>model3<\/code>) are different.<\/p>\n","answers":[]}
+{"QuestionId":57773636,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-03T14:20:06.893","AcceptedAnswerId":null,"OwnerUserId":3650983.0,"Title":"no attribute 'set_keras_submodules'","Body":"<p>getting:<\/p>\n\n<blockquote>\n  <p>AttributeError: module 'keras_applications' has no attribute\n  'set_keras_submodules'<\/p>\n<\/blockquote>\n\n<p>in:<\/p>\n\n<pre><code>keras_applications.set_keras_submodules(backend, layers, models, utils)\n<\/code><\/pre>\n\n<p>solutions of this issue\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/25062\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/25062<\/a>\ndon't solve the error<\/p>\n\n<p>on:<\/p>\n\n<ul>\n<li>python 3.7.3<\/li>\n<li>tensorflow 1.14.0<\/li>\n<li>keras 2.2.4<\/li>\n<li>keras_applications 1.0.8<\/li>\n<\/ul>\n","answers":[{"AnswerId":"57773637","CreationDate":"2019-09-03T14:20:06.893","ParentId":null,"OwnerUserId":"3650983","Title":null,"Body":"<p>downgrade keras_applications to 1.0.7 solve the problem for me<\/p>\n"}]}
+{"QuestionId":57774456,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-09-03T15:09:42.290","AcceptedAnswerId":null,"OwnerUserId":11452683.0,"Title":"Specify network to not backpropagate the gradients in a given function","Body":"<p>In my neural network, in order to compute the loss, I need to do some intermediate computations during training to first obtain some transformations <code>rv<\/code>.<\/p>\n\n<pre><code>rv = factor.ransac(source, target, prob, device)\npredicted = factor.predict(source, rv, outputs, device)\nloss = criterion(predicted, target)\n<\/code><\/pre>\n\n<p>I want to backpropagate the gradients only through <code>predicted<\/code> but not through <code>factor.ransac<\/code>. How do I do this?<\/p>\n","answers":[{"AnswerId":"57778280","CreationDate":"2019-09-03T20:22:38.707","ParentId":null,"OwnerUserId":"248522","Title":null,"Body":"<p>You can use <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.Tensor.detach\" rel=\"nofollow noreferrer\">detach()<\/a>.<\/p>\n\n<pre><code>rv = factor.ransac(source, target, prob, device).detach()\n<\/code><\/pre>\n\n<p>That way, gradients on <code>rv<\/code> will be discarded.<\/p>\n"}]}
+{"QuestionId":57774575,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><keras><tensorflow-lite>","CreationDate":"2019-09-03T15:17:20.047","AcceptedAnswerId":null,"OwnerUserId":8334261.0,"Title":"TFLite Converter: RandomStandardNormal implemented for keras model, but not for pure TensorFlow model","Body":"<p><strong>Task<\/strong><\/p>\n\n<p>I have two models which should be equivalent. The first one is built with keras, the second one with tensorflow. Both variational autoencoders use the <code>tf.random.normal<\/code> method in their model. Also, they produce similar results. Everything is tested with the nightly build (1.15).<\/p>\n\n<p>The confusion comes when I try to convert them into a tensorflow lite model with post-training quantization. I use the same command for both models:<\/p>\n\n<pre><code>converter = tf.compat.v1.lite.TFLiteConverter... # from respective save file\nconverter.representative_dataset = representative_dataset_gen\nconverter.optimizations = [tf.lite.Optimize.DEFAULT]\n\ntflite_model = converter.convert()\nopen('vae.tflite', 'wb').write(tflite_model)\n<\/code><\/pre>\n\n<p><strong>Error<\/strong><\/p>\n\n<p>For the keras model, everything goes well and I end up with a working tflite model. However, if I try to do this with the tensorflow version, I run into an error stating that <code>RandomStandardNormal<\/code> is not implemented:<\/p>\n\n<blockquote>\n  <p>Some of the operators in the model are not supported by the standard\n  TensorFlow Lite runtime. If those are native TensorFlow operators, you\n  might be able to use the extended runtime by passing\n  --enable_select_tf_ops, or by setting target_ops=TFLITE_BUILTINS,SELECT_TF_OPS when calling\n  tf.lite.TFLiteConverter(). Otherwise, if you have a custom\n  implementation for them you can disable this error with\n  --allow_custom_ops, or by setting allow_custom_ops=True when calling tf.lite.TFLiteConverter(). Here is a list of builtin operators you are\n  using: ADD, EXP, FULLY_CONNECTED, LEAKY_RELU, LOG, MUL. Here is a list\n  of operators for which you will need custom implementations:\n  RandomStandardNormal.<\/p>\n<\/blockquote>\n\n<p><strong>Question<\/strong><\/p>\n\n<p>This does not make sense to me because obviously this is already working with keras. Does keras know something that I have to tell my tensorflow model explicitly?<\/p>\n\n<p><strong>MODELS<\/strong><\/p>\n\n<p>tensorflow<\/p>\n\n<pre><code>fc_layer = tf.compat.v1.layers.dense\n\n# inputs have shape (90,)\nx = tf.identity(inputs, name='x')\n\n# encoder\noutputs = fc_layer(x, 40, activation=leaky_relu)\n\nself.z_mu = fc_layer(outputs, 10, activation=None)\nself.z_sigma = fc_layer(outputs, 10, activation=softplus)\n\n# latent space\neps = tf.random.normal(shape=tf.shape((10,)), mean=0, stddev=1, dtype=tf.float32)\noutputs = self.z_mu + eps * self.z_sigma\n\n# decoder\noutputs = fc_layer(outputs, 40, activation=leaky_relu)\n\n# prediction\nx_decoded = fc_layer(outputs, 90, activation=None)\n<\/code><\/pre>\n\n<p>keras<\/p>\n\n<pre><code>x = keras.layers.Input(shape=(90,))\n\nh = keras.layers.Dense(40, activation=leakyrelu)(x)\n\nz_mu = keras.layers.Dense(10)(h)\nz_sigma = keras.layers.Dense(10, activation=softplus)(h)\n\neps = tf.random.normal(shape=tf.shape((10,)), mean=0, stddev=1, dtype=tf.float32)\nz = z_mu + eps * z_sigma\n\nh_decoded = keras.layers.Dense(40, activation=leakyrelu)(z)\nx_decoded = keras.layers.Dense(90)(h_decoded)\n\ntrain_model = keras.models.Model(x, x_decoded)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57775058,"AnswerCount":0,"Tags":"<tensorflow><keras><deep-learning><tf.keras>","CreationDate":"2019-09-03T15:51:17.810","AcceptedAnswerId":null,"OwnerUserId":6913957.0,"Title":"Converting CNN-LSTM from keras to tf.keras brings down the accuracy","Body":"<p>First of all, I'm still learning to code in keras.\nI was able to convert regular keras into tf.keras with estimator for simple LSTM model. However when this conversion is done on CNN-LSTM, the converted tf.keras estimator produces unexpectedly high RMSE (taking average over 30 repetition):<\/p>\n\n<ul>\n<li>keras  avg RMSE = 1749.563 (+\/- 262.194)<\/li>\n<li>tf.keras avg RMSE = 4231.720 (+\/- 220.907)<\/li>\n<\/ul>\n\n<p>Pretty sure I must have missed something but am not sure what it is. Appreciate any help.<\/p>\n\n<p>I've tried switching to similar but different modules in tf.keras but it does not move a needle. I've done some tuning on batch and epoch but they do not bring down the loss.<\/p>\n\n<p>partial tf.keras code:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom math import sqrt\nfrom sklearn.metrics import mean_squared_error\nfrom tensorflow.keras import Sequential\nfrom tensorflow.keras.layers import Dense\nfrom tensorflow.keras.layers import TimeDistributed\nfrom tensorflow.keras.layers import Flatten\nfrom tensorflow.keras.layers import Conv1D\nfrom tensorflow.keras.layers import MaxPooling1D, MaxPool1D\n\nfrom tensorflow.keras.layers import LSTM\nfrom matplotlib import pyplot\nimport numpy as np\nimport pandas as pd\nimport warnings\n\ndef model_fit_cnnlstm(train_supv, config, first_round):\n    # train X shape = (48, 36)\n    train_x, train_y = train_supv[:, :-1], train_supv[:, -1]\n    # train X lstm shape = (48, 36, 1)\n    train_x_lstm = train_x.reshape(\n        train_x.shape[0], config['n_seq'], config['n_steps'], config['n_feature'])\n    if first_round.state:\n        # print('train X shape = {}'.format(train_x.shape))\n        # print('train X lstm shape = {}'.format(train_x_lstm.shape))\n        pass\n    #\n    # CuDNNLSTM so fast. But hardcoded to use tanh activation. No Relu support\n    # model.add(CuDNNLSTM(config['n_nodes'], input_shape=(config['n_input'], config['n_feature'])))\n    # LSTM slow not optimized for GPU\n    input_shape = (None, config['n_steps'], config['n_feature'])\n    # relu = tf.nn.relu\n    relu = 'relu'\n    model = Sequential()\n    # model.add(tf.keras.layers.InputLayer(input_shape=input_shape, name=config['model_name']))\n    model.add(tf.keras.layers.InputLayer(input_shape=input_shape))\n    model.add(TimeDistributed(\n        Conv1D(config['n_filter'], config['n_kernel'], activation=relu)))\n    model.add(TimeDistributed(\n        Conv1D(config['n_filter'], config['n_kernel'], activation=relu)))\n    model.add(TimeDistributed(MaxPool1D()))\n    model.add(TimeDistributed(Flatten()))\n    model.add(LSTM(config['n_nodes'], activation=relu))\n    model.add(Dense(config['n_nodes'], activation=relu))\n    model.add(Dense(1))\n    adam = tf.train.AdamOptimizer()\n    # adam = 'adam'\n    mse = tf.keras.losses.mean_squared_error\n    # mse = 'mse'\n    model.compile(optimizer=adam, loss=mse)\n    print('model {}'.format(model))\n    #\n    with warnings.catch_warnings():\n        warnings.filterwarnings(\"ignore\")\n        estimator = tf.keras.estimator.model_to_estimator(keras_model=model)\n    #\n    # train X lstm shape = (48, 36, 1)\n    # numpy_input_fn takes x: train_x_lstm shape\n    # print('model.input_names {}'.format(model.input_names))\n    train_input_fn = tf.estimator.inputs.numpy_input_fn(\n        x={model.input_names[0]: train_x_lstm.astype(np.float32)},\n        y=train_y.astype(np.float32),\n        num_epochs=config['n_epochs'],\n        batch_size=config['n_batch'],\n        shuffle=False\n    )\n    estimator.train(input_fn=train_input_fn)\n    #\n    return model, estimator\n\ndef model_predict_cnnlstm(model, estimator, test_supv, config, first_round):\n    test_x, test_y = test_supv[:, :-1], test_supv[:, -1]\n    test_x_lstm = test_x.reshape(\n        test_x.shape[0], config['n_seq'], config['n_steps'], config['n_feature'])\n    if first_round.state:\n        # print('test X shape = {}'.format(test_x.shape))\n        # print('test X lstm shape = {}'.format(test_x_lstm.shape))\n        pass\n    #\n    predict_input_fn = tf.estimator.inputs.numpy_input_fn(\n        x={model.input_names[0]: test_x_lstm.astype(np.float32)},\n        y=None,\n        batch_size=1,\n        num_epochs=1,\n        shuffle=False,\n    )\n    yhat_list = estimator.predict(input_fn=predict_input_fn)\n    # print('yhat_list {}'.format(yhat_list))\n    predictions = np.array([])\n    for yhat_dict in yhat_list:\n        for key, val in yhat_dict.items():\n            # print('yhat = {} {}'.format(key, val))\n            predictions = np.append(predictions, val)\n    #\n    return predictions\n<\/code><\/pre>\n\n<p>partial keras code:<\/p>\n\n<pre><code>from math import sqrt\nfrom sklearn.metrics import mean_squared_error\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.layers import LSTM\nfrom keras.layers import TimeDistributed\nfrom keras.layers import Flatten\nfrom keras.layers.convolutional import Conv1D\nfrom keras.layers.convolutional import MaxPooling1D\n\nfrom matplotlib import pyplot\nimport numpy as np\nimport pandas as pd\n\ndef model_fit_cnnlstm(train_supv, config, first_round):\n    train_x, train_y = train_supv[:, :-1], train_supv[:, -1]\n    train_x_lstm = train_x.reshape(\n        train_x.shape[0], config['n_seq'], config['n_steps'], config['n_feature'])\n    #\n    input_shape = (None, config['n_steps'], config['n_feature'])\n    model = Sequential()\n    model.add(TimeDistributed(Conv1D(\n        config['n_filter'], config['n_kernel'], activation='relu', input_shape=input_shape)))\n    model.add(TimeDistributed(\n        Conv1D(config['n_filter'], config['n_kernel'], activation='relu')))\n    model.add(TimeDistributed(MaxPooling1D()))\n    model.add(TimeDistributed(Flatten()))\n    model.add(LSTM(config['n_nodes'], activation='relu'))\n    model.add(Dense(config['n_nodes'], activation='relu'))\n    model.add(Dense(1))\n    model.compile(optimizer='adam', loss='mse')\n    #\n    # model fit\n    model.fit(train_x_lstm, train_y,\n              epochs=config['n_epochs'], batch_size=config['n_batch'], verbose=0)\n    return model\n\ndef model_predict_cnnlstm(model, test_supv, config, first_round):\n    test_x, test_y = test_supv[:, :-1], test_supv[:, -1]\n    test_x_lstm = test_x.reshape(\n        test_x.shape[0], config['n_seq'], config['n_steps'], config['n_feature'])\n    if first_round.state:\n        print('test X shape = {}'.format(test_x.shape))\n        print('test X cnn-lstm shape = {}'.format(test_x_lstm.shape))\n    #\n    predictions = np.array([])\n    for row in test_x_lstm:\n        test_x_lstm_row = row.reshape(\n            1, config['n_seq'], config['n_steps'], config['n_feature'])\n        yhat = model.predict(test_x_lstm_row, verbose=0)\n        predictions = np.append(predictions, yhat)\n    #\n    if first_round.state:\n        print('unconverted prediction shape {}'.format(predictions.shape))\n    # predictions = predictions.reshape(test_y.shape[0],)\n    return predictions\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57775106,"AnswerCount":1,"Tags":"<python><tensorflow><gpflow>","CreationDate":"2019-09-03T15:55:23.880","AcceptedAnswerId":57785972.0,"OwnerUserId":8655352.0,"Title":"Incorporate known data moments into GPFlow fitting","Body":"<p>I have recently been working with gpflow, in-particular Gaussian process regression, to model a process for which I have access to approximated moments for each input. I have a vector of input values X of size (N,1) and a vector of responses Y of size (N,1). However, I also know, for each (x,y) pair, an approximation of the associated variance, skewness, kurtosis and so on for the particular y value.<\/p>\n\n<p>From this, I know properties that inform me of appropriate likelihoods to use for each data point.\nIn the simplest case, I just assume all likelihoods are Gaussian, and specify the variance at each point. I've created a minimal example of my code by adapting the tutorial on: <a href=\"https:\/\/nbviewer.jupyter.org\/github\/GPflow\/GPflow\/blob\/develop\/doc\/source\/notebooks\/advanced\/varying_noise.ipynb#Demo-2:-grouped-noise-variances\" rel=\"nofollow noreferrer\">https:\/\/nbviewer.jupyter.org\/github\/GPflow\/GPflow\/blob\/develop\/doc\/source\/notebooks\/advanced\/varying_noise.ipynb#Demo-2:-grouped-noise-variances<\/a>.<\/p>\n\n<pre><code>import numpy as np\nimport gpflow \n\ndef generate_data(N=100):\n    X = np.random.rand(N)[:, None] * 10 - 5  # Inputs, shape N x 1\n    F = 2.5 * np.sin(6 * X) + np.cos(3 * X)  # Mean function values\n    groups = np.arange( 0, N, 1 ).reshape(-1,1)  \n    NoiseVar = np.array([i\/100.0 for i in range(N)])[groups]  \n    Y = F + np.random.randn(N, 1) * np.sqrt(NoiseVar)  # Noisy data\n    return X, Y, groups, NoiseVar\n\n# Get data\nX, Y, groups, NoiseVar = generate_data()\nY_data = np.hstack([Y, groups])\n# Generate one likelihood per data-point\nlikelihood = gpflow.likelihoods.SwitchedLikelihood( [gpflow.likelihoods.Gaussian(variance=NoiseVar[i]) for i in range(Y.shape[0])])\n\n# model construction (notice that num_latent is 1)\nkern  = gpflow.kernels.Matern52(input_dim=1, lengthscales=0.5)\nmodel = gpflow.models.VGP(X, Y_data, kern=kern, likelihood=likelihood, num_latent=1)\n# Specify the likelihood as non-trainable\nmodel.likelihood.set_trainable(False)\n\n# build the natural gradients optimiser\nnatgrad_optimizer = gpflow.training.NatGradOptimizer(gamma=1.)\nnatgrad_tensor    = natgrad_optimizer.make_optimize_tensor(model, var_list=[(model.q_mu, model.q_sqrt)])\nsession = model.enquire_session()\nsession.run(natgrad_tensor)\n# update the cache of the variational parameters in the current session\nmodel.anchor(session)\n\n# Stop Adam from optimising the variational parameters\nmodel.q_mu.trainable   = False\nmodel.q_sqrt.trainable = False\n# Create Adam tensor\nadam_tensor = gpflow.train.AdamOptimizer(learning_rate=0.1).make_optimize_tensor(model)\nfor i in range(200): \n    session.run(natgrad_tensor)\n    session.run(adam_tensor)\n# update the cache of the parameters in the current session\nmodel.anchor(session)\nprint(model)\n<\/code><\/pre>\n\n<p>The above code works for a gaussian likelihood, and known variances. Inspecting my real data, I see that it is skewed very often and as a result, I want to use non-gaussian likelihoods to model it, but am unsure how to specify these other likelihood parameters given what I know. <\/p>\n\n<p>So my question is: Given this setup, how can I adapt my code so far to include non-Gaussian likelihoods at each step, in-particular specifying and fixing their parameters based on my known variances, skewness, kurtosis and so on associated with each individual y value?<\/p>\n","answers":[{"AnswerId":"57785972","CreationDate":"2019-09-04T09:57:26.877","ParentId":null,"OwnerUserId":"1018464","Title":null,"Body":"<p>Firstly, you will need to choose which non-Gaussian likelihood you use. GPflow includes various ones in <code>likelihoods.py<\/code>. You then need to adapt the line<\/p>\n\n<pre><code>likelihood = gpflow.likelihoods.SwitchedLikelihood(\n    [gpflow.likelihoods.Gaussian(variance=NoiseVar[i]) for i in range(Y.shape[0])]\n)\n<\/code><\/pre>\n\n<p>to give a list of your non-Gaussian likelihoods.<\/p>\n\n<p>Which likelihood can take advantage of your skewness and kurtosis information is a statistical question. Depending on what you come up with, you may need to implement your own likelihood class, which can be done by inheriting from <code>Likelihood<\/code>. You should be able to follow some other examples from <code>likelihoods.py<\/code>.<\/p>\n"}]}
+{"QuestionId":57775212,"AnswerCount":0,"Tags":"<tensorflow><tensorboard><tensorflow-estimator>","CreationDate":"2019-09-03T16:02:39.537","AcceptedAnswerId":null,"OwnerUserId":386590.0,"Title":"Track metrics in Tensorboard with Tensorflow Estimator in TF 2.0","Body":"<p>I am trying to use Tensorflow Estimator in Tensorflow 2.0 with a Keras model and estimator but I can't find a way to track the metrics in Tensorboard without going through the classic Keras training loop.<\/p>\n\n<pre><code>def simple_model():\n  model = tf.keras.Sequential([\n  layers.Dense(64, activation='relu', name=\"dense_0\"),                \n  layers.Dense(64, activation='relu', name=\"dense_1\"),                \n  layers.Dense(10, activation='softmax', name=\"dense_2\"),                \n  ])\n  return model\n\nmodel = simple_model()\nmodel.compile(optimizer=keras.optimizers.Adam(lr=3e-4, beta_1=0.8, \n                                              beta_2=0.99),\n              loss='categorical_crossentropy',\n              metrics=['accuracy'])\n\nestimator = tf.keras.estimator.model_to_estimator(\n                keras_model = model,\n                model_dir = '\/tmp\/tf')\n\nestimator.train(input_fn=lambda : train_input_fn( 128), steps=10000)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57775393,"AnswerCount":0,"Tags":"<python-3.x><python-2.7><tensorflow>","CreationDate":"2019-09-03T16:15:40.833","AcceptedAnswerId":null,"OwnerUserId":9954063.0,"Title":"How to solve FailedPreconditionError?","Body":"<p>I have a small 3 layer network where I have created variables for weights and biases using initer, and I have used <code>tf.variables_initializer<\/code>to initialize variables as suggested by someone. Unfortunately despite initializin them, there still apears a Failed PreCondition Error. I am now out of ideas about how to solve it or what exactly is causing it.<\/p>\n\n<pre><code>with tf.compat.v1.variable_scope(\"weight_in\", reuse=tf.compat.v1.AUTO_REUSE):\n            weight_i = tf.compat.v1.get_variable(\"weight_input\", dtype=tf.float64, shape=[self.neurons, 1], initializer=self.weight_initer)\n        with tf.compat.v1.variable_scope(\"weight_hid\", reuse=tf.compat.v1.AUTO_REUSE):   \n            weight_h = tf.compat.v1.get_variable(\"weight_hidden\", dtype=tf.float64, shape=[self.neurons, 1], initializer=self.weight_initer1)\n        with tf.compat.v1.variable_scope(\"weight_out\", reuse=tf.compat.v1.AUTO_REUSE): \n            weight_o = tf.compat.v1.get_variable(\"weight_input\", dtype=tf.float64, shape=[4, 2], initializer=self.weight_initer2)\n        with tf.compat.v1.variable_scope(\"bias_in\", reuse=tf.compat.v1.AUTO_REUSE): \n            bias_i = tf.compat.v1.get_variable(\"bias_input\", dtype=tf.float64, shape=[self.neurons, 1], initializer=self.bias_initer)\n        with tf.compat.v1.variable_scope(\"bias_hid\", reuse=tf.compat.v1.AUTO_REUSE): \n            bias_h = tf.compat.v1.get_variable(\"bias_hidden\", dtype=tf.float64, shape=[self.neurons, 1], initializer=self.bias_initer1)\n        with tf.compat.v1.variable_scope(\"bias_out\", reuse=tf.compat.v1.AUTO_REUSE): \n            bias_o = tf.compat.v1.get_variable(\"bias_out\", dtype=tf.float64, shape=[1, 4], initializer=self.bias_initer2)\n       uninitialized_variables = []\n        sess = tf.Session()\n        print(\"printing the names of all variables\")\n        print(tf.all_variables())\n        for var in tf.all_variables():\n            try:\n                sess.run(var)\n            except tf.errors.FailedPreconditionError:\n                uninitialized_variables.append(var)\n        init = tf.variables_initializer(uninitialized_variables)\n        sess.run(init)\n<\/code><\/pre>\n\n<p>The error is <\/p>\n\n<pre><code>FailedPreconditionError: Attewmpting to use initialized value weight_in\/weight_input\n<\/code><\/pre>\n\n<p>sometimes the same error will be pointing another bias variable. I am wondering whether anyone can shed some light on how to properly create a variable and initialize it. Thank you.<\/p>\n","answers":[]}
+{"QuestionId":57775702,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><image-processing><deep-learning><pytorch>","CreationDate":"2019-09-03T16:37:45.430","AcceptedAnswerId":null,"OwnerUserId":6642287.0,"Title":"Tensorflow Equivalent of Elastic deformation for pytorch","Body":"<p>I am trying to implement elastic deformation by sampling control points on a regularly spaces 100*100 grid with \u03c3 = 20 with bilinear variant for the image and nearest neighbour for the masks.\nI found <a href=\"https:\/\/github.com\/deepmind\/multidim-image-augmentation\/blob\/master\/doc\/elastic_deformation_colab.md\" rel=\"nofollow noreferrer\">this<\/a> implementation for the TensorFlow by DeepMind.\nSo far, I was using the below implementation as a lambda transformation-<\/p>\n\n<pre><code>def elastic_transform(image, alpha=1000, sigma=20, spline_order=1, mode='nearest', random_state=np.random):\n\"\"\"Elastic deformation of image as described in [Simard2003]_.\n.. [Simard2003] Simard, Steinkraus and Platt, \"Best Practices for\n   Convolutional Neural Networks applied to Visual Document Analysis\", in\n   Proc. of the International Conference on Document Analysis and\n   Recognition, 2003.\n\"\"\"\n#assert image.ndim == 3\nimage = np.array(image)\nassert image.ndim == 3\nshape = image.shape[:2]\n\ndx = gaussian_filter((random_state.rand(*shape) * 2 - 1),\n                     sigma, mode=\"constant\", cval=0) * alpha\ndy = gaussian_filter((random_state.rand(*shape) * 2 - 1),\n                     sigma, mode=\"constant\", cval=0) * alpha\n\nx, y = np.meshgrid(np.arange(shape[0]), np.arange(shape[1]), indexing='ij')\nindices = [np.reshape(x + dx, (-1, 1)), np.reshape(y + dy, (-1, 1))]\nresult = np.empty_like(image)\nfor i in range(image.shape[2]):\n    result[:, :, i] = map_coordinates(\n        image[:, :, i], indices, order=spline_order, mode=mode).reshape(shape)\nresult = Image.fromarray(result)\nreturn result\n<\/code><\/pre>\n\n<p>But the issue with the above implementation is it uses spline interpolation.\nIs there any TensorFlow equivalent(kind of which i specified above) for PyTorch or is there any other way I can implement my requirement without switching to Tensorflow?<\/p>\n","answers":[]}
+{"QuestionId":57775938,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><machine-learning><keras><tf.keras>","CreationDate":"2019-09-03T16:56:04.013","AcceptedAnswerId":57784713.0,"OwnerUserId":6997665.0,"Title":"Evaluating pairwise distances between the output of a tf.keras.model","Body":"<p>I am trying to create a custom loss function in tensorflow. I am using tensorflow v2.0.rc0 for running the code. Following is the code and the function <code>min_dist_loss<\/code> computes the pairwise loss between the output of the neural network. Here's the code<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def min_dist_loss(_, y_pred):\n    distances = []\n    for i in range(0, 16):\n        for j in range(i + 1, 16):\n            distances.append(tf.linalg.norm(y_pred[i] - y_pred[j]))\n    return -tf.reduce_min(distances)\n<\/code><\/pre>\n\n<p>and the module is being initialized and compiled as follows<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nfrom tensorboard.plugins.hparams import api as hp\n\nHP_NUM_UNITS = hp.HParam('num_units', hp.Discrete([6, 7]))\nHP_OPTIMIZER = hp.HParam('optimizer', hp.Discrete(['adam', 'sgd']))\n\nMETRIC_ACCURACY = 'accuracy'\n\nwith tf.summary.create_file_writer('logs\\hparam_tuning').as_default():\n    hp.hparams_config(\n        hparams=[HP_NUM_UNITS, HP_OPTIMIZER],\n        metrics=[hp.Metric(METRIC_ACCURACY, display_name='Accuracy')]\n    )\n\ndef train_test_model(logdir, hparams):\n    weight1 = np.random.normal(loc=0.0, scale=0.01, size=[4, hparams[HP_NUM_UNITS]])\n    init1 = tf.constant_initializer(weight1)\n    weight2 = np.random.normal(loc=0.0, scale=0.01, size=[hparams[HP_NUM_UNITS], 7])\n    init2 = tf.constant_initializer(weight2)\n    model = tf.keras.models.Sequential([\n        # tf.keras.layers.Flatten(),\n        tf.keras.layers.Dense(hparams[HP_NUM_UNITS], activation=tf.nn.sigmoid, kernel_initializer=init1),\n        tf.keras.layers.Dense(7, activation=tf.nn.sigmoid, kernel_initializer=init2) if hparams[HP_NUM_UNITS] == 6 else\n        None,\n    ])\n    model.compile(\n        optimizer=hparams[HP_OPTIMIZER],\n        loss=min_dist_loss,\n        # metrics=['accuracy'],\n    )\n    x_train = [list(k) for k in itertools.product([0, 1], repeat=4)]\n    shuffle(x_train)\n    x_train = 2 * np.array(x_train) - 1\n    model.fit(\n        x_train, epochs=1, batch_size=16,\n        callbacks=[\n            tf.keras.callbacks.TensorBoard(logdir),\n            hp.KerasCallback(logdir, hparams)\n        ],\n    )\n<\/code><\/pre>\n\n<p>Now since the tensor object <code>y_pred<\/code> in the <code>min_dist_loss<\/code> is an object of shape <code>[?, 7]<\/code>, indexing with <code>i<\/code> is throwing the following error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/src\/try1.py\", line 95, in &lt;module&gt;\n    run('logs\\hparam_tuning' + run_name, hparams)\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/src\/try1.py\", line 78, in run\n    accuracy = train_test_model(run_dir, hparams)\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/src\/try1.py\", line 66, in train_test_model\n    hp.KerasCallback(logdir, hparams)\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/keras\/engine\/training.py\", line 734, in fit\n    use_multiprocessing=use_multiprocessing)\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/keras\/engine\/training_v2.py\", line 324, in fit\n    total_epochs=epochs)\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/keras\/engine\/training_v2.py\", line 123, in run_one_epoch\n    batch_outs = execution_function(iterator)\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/keras\/engine\/training_v2_utils.py\", line 86, in execution_function\n    distributed_function(input_fn))\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/eager\/def_function.py\", line 427, in __call__\n    self._initialize(args, kwds, add_initializers_to=initializer_map)\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/eager\/def_function.py\", line 370, in _initialize\n    *args, **kwds))\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/eager\/function.py\", line 1847, in _get_concrete_function_internal_garbage_collected\n    graph_function, _, _ = self._maybe_define_function(args, kwargs)\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/eager\/function.py\", line 2147, in _maybe_define_function\n    graph_function = self._create_graph_function(args, kwargs)\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/eager\/function.py\", line 2038, in _create_graph_function\n    capture_by_value=self._capture_by_value),\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/framework\/func_graph.py\", line 915, in func_graph_from_py_func\n    func_outputs = python_func(*func_args, **func_kwargs)\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/eager\/def_function.py\", line 320, in wrapped_fn\n    return weak_wrapped_fn().__wrapped__(*args, **kwds)\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/keras\/engine\/training_v2_utils.py\", line 73, in distributed_function\n    per_replica_function, args=(model, x, y, sample_weights))\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/distribute\/distribute_lib.py\", line 760, in experimental_run_v2\n    return self._extended.call_for_each_replica(fn, args=args, kwargs=kwargs)\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/distribute\/distribute_lib.py\", line 1787, in call_for_each_replica\n    return self._call_for_each_replica(fn, args, kwargs)\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/distribute\/distribute_lib.py\", line 2132, in _call_for_each_replica\n    return fn(*args, **kwargs)\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/autograph\/impl\/api.py\", line 292, in wrapper\n    return func(*args, **kwargs)\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/keras\/engine\/training_v2_utils.py\", line 264, in train_on_batch\n    output_loss_metrics=model._output_loss_metrics)\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/keras\/engine\/training_eager.py\", line 311, in train_on_batch\n    output_loss_metrics=output_loss_metrics))\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/keras\/engine\/training_eager.py\", line 252, in _process_single_batch\n    training=training))\n  File \"\/home\/pc\/Documents\/user\/code\/keras_tensorflow\/venv\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/keras\/engine\/training_eager.py\", line 166, in _model_loss\n    per_sample_losses = loss_fn.call(targets[i], outs[i])\nIndexError: list index out of range\n<\/code><\/pre>\n\n<hr>\n\n<p>How do I compute the minimum distance in this setting? Any help is appreciated. Also, if there are any errors in other parts of the code, please feel free to point it out. I am new to using <code>keras<\/code> on <code>tensorflow<\/code>.<\/p>\n","answers":[{"AnswerId":"57784713","CreationDate":"2019-09-04T08:45:31.300","ParentId":null,"OwnerUserId":"3337089","Title":null,"Body":"<p>Keras is expecting you to provide the true labels as well. Since you're defining your own loss function and you're not using the true labels, you can pass some garbage labels. Eg: <code>np.arange(16)<\/code>.<\/p>\n\n<p>Change your <code>model.fit<\/code> as below and it should work<\/p>\n\n<pre><code>model.fit(\n    x_train, np.arange(x_train.shape[0]), epochs=1, batch_size=16,\n    callbacks=[\n        tf.keras.callbacks.TensorBoard(logdir),\n        hp.KerasCallback(logdir, hparams)\n    ],\n)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57776215,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-estimator>","CreationDate":"2019-09-03T17:20:44.423","AcceptedAnswerId":null,"OwnerUserId":3390114.0,"Title":"How do I implement early stopping with Estimator API for distributed training?","Body":"<p>I'm using Tensorflow 1.4 Estimator and Dataset APIs for distributed training in Google Cloud Platform.  I want to implement early stopping to prevent overfitting during the training, and looked at early stopping hooks documentation below:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/experimental\/make_early_stopping_hook\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/experimental\/make_early_stopping_hook<\/a> <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/experimental\/stop_if_no_decrease_hook\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/experimental\/stop_if_no_decrease_hook<\/a><\/li>\n<\/ul>\n\n<p>But, none of these hooks support <em>distributed training<\/em>; so the question stands:<\/p>\n\n<p><strong>How do I implement early stopping with Estimator API that also supports distributed training?<\/strong> If that's not possible, what do you recommend? How do I prevent overfitting?<\/p>\n\n<p>There is a similar question<a href=\"https:\/\/stackoverflow.com\/questions\/47137061\/early-stopping-with-tf-estimator-how\/57775642#57775642\"> [here]<\/a>, but that doesn't address the distributed training concern.<\/p>\n","answers":[]}
+{"QuestionId":57776779,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-03T18:09:15.457","AcceptedAnswerId":null,"OwnerUserId":1067998.0,"Title":"TensorFlow 2.0: create replica-local variable under MirroredStrategy","Body":"<p>I'm implementing Memory Transformer and I need to keep the <code>memory<\/code> between calls to the model. I've tried to use <code>tf.Variable<\/code> for this task and it works perfectly on a single GPU.<\/p>\n\n<p>But under MirroredStrategy on multiple GPUs this approach fails because mirrored strategy wants to synchronize the variables written on multiple replicas. Which is not needed in my case, I need to create a personal set of memory variables on each 'tower' like in TransformerXL implementation.<\/p>\n\n<p>I think I can use <code>with tf.device():<\/code> to create those variables but I'm not sure how to get current replica's device inside <code>build<\/code> method.<\/p>\n","answers":[]}
+{"QuestionId":57777124,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-03T18:38:53.423","AcceptedAnswerId":null,"OwnerUserId":9005671.0,"Title":"How to apply stochastic variational inference in Bayesian Deep Learning with Tensorflow probability","Body":"<p>I am applying Bayesian Deep Learning in my work. I have the intention of applying Stochastic Variational approximation with the bid of comparing performance with the variational inference.\nCan I get clues on how I can go about this?<\/p>\n\n<p>Thanks <\/p>\n\n<p>Elvin<\/p>\n","answers":[]}
+{"QuestionId":57777574,"AnswerCount":1,"Tags":"<python><pytorch><tensor>","CreationDate":"2019-09-03T19:16:57.657","AcceptedAnswerId":57778205.0,"OwnerUserId":10023920.0,"Title":"How do I concatenate tensors along a given axis?","Body":"<pre class=\"lang-py prettyprint-override\"><code>\nimport torch\n\na = torch.Tensor(2,2,2)\n\nb = myfunction(a)\n\nprint(a)\n&gt;&gt;&gt; [[[1,2],\n      [5,6]],\n     [[7,8],\n      [9,10]]]\n\nprint(b)\n&gt;&gt;&gt; [[1,2,7,8],\n     [5,6,9,10]]\n\n<\/code><\/pre>\n\n<p>How do I code myfunction to get b from a?<\/p>\n\n<p>Is there some pytorch functions that transforms a in such way?<\/p>\n","answers":[{"AnswerId":"57778205","CreationDate":"2019-09-03T20:15:06.333","ParentId":null,"OwnerUserId":"5085211","Title":null,"Body":"<p>You can achieve this by using <code>transpose<\/code> to swap the first two axes (cf. e.g. <code>np.swapaxes<\/code>), and <code>reshape<\/code> to get your desired shape:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>In [12]: a\nOut[12]:\ntensor([[[  1.,   2.],\n         [  5.,   6.]],\n\n        [[  7.,   8.],\n         [  9.,  10.]]])\n\nIn [13]: a.transpose(0, 1).reshape(2, 4)\nOut[13]:\ntensor([[  1.,   2.,   7.,   8.],\n        [  5.,   6.,   9.,  10.]])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57777751,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-03T19:32:12.113","AcceptedAnswerId":null,"OwnerUserId":3771253.0,"Title":"Unweighted vs Weighted sample","Body":"<p>I have a classification task that I perform using DNN.<\/p>\n\n<p>Whenever I use a weighted training sample, I get an inverted ROC curve, which is not the case with unweighted sample.<\/p>\n\n<pre><code>model = tf.keras.models.Sequential([\ntf.keras.layers.Dense(150, input_dim=25, activation=tf.nn.relu),\ntf.keras.layers.Dense(150, activation=tf.nn.relu),\ntf.keras.layers.Dense(150, activation=tf.nn.relu),\ntf.keras.layers.Dense(1, activation=tf.nn.sigmoid)\n])\nmodel.compile(optimizer=tf.keras.optimizers.Adam(lr=0.001, \n['accuracy'],sample_weight_mode= None)\nmodel.fit(X_train, y_train, epochs=50,validation_split=0.3, \nbatch_size=250,verbose=0, sample_weight = weight_train)\n<\/code><\/pre>\n\n<h3>Weighted ROC<\/h3>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/oTj42.png\" alt=\"weightedROC\"><\/p>\n\n<h3>Unweighted ROC<\/h3>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/VGJtq.png\" alt=\"unWeightedROC\"><\/p>\n","answers":[]}
+{"QuestionId":57777839,"AnswerCount":0,"Tags":"<tensorflow><sparse-matrix>","CreationDate":"2019-09-03T19:42:01.050","AcceptedAnswerId":null,"OwnerUserId":383313.0,"Title":"Elementwise multiplication between a sparse matrix and a broadcasted vector in tensorflow?","Body":"<p>I'd like to multiply elementwise a sparse 2D matrix with a 1D dense vector broadcasted to the same dense_shape as the sparse matrix. (Sometimes the broadcasting operates on axis 0, sometimes on axis 1.)<\/p>\n\n<p>Without broadcasting, <a href=\"https:\/\/stackoverflow.com\/a\/45468794\/383313\">https:\/\/stackoverflow.com\/a\/45468794\/383313<\/a> solves the problem perfectly with the following nice line, which multiplies a sparse matrix with a dense matrix elementwise:<\/p>\n\n<pre><code>res = tf.SparseTensor(x.indices, tf.gather_nd(y, x.indices) * x.values, x.dense_shape)\n<\/code><\/pre>\n\n<p>Unfortunately, broadcasting is a non-negotiable requirement for me, as I don't have enough GPU memory to store the full dense matrix. I don't need a fully general solution, just this specific 2D-sparse-against-1D-dense case. Can I avoid writing my own op?<\/p>\n","answers":[]}
+{"QuestionId":57778020,"AnswerCount":0,"Tags":"<python><keras><jupyter-notebook><concurrent.futures><tf.keras>","CreationDate":"2019-09-03T19:58:29.160","AcceptedAnswerId":null,"OwnerUserId":4626234.0,"Title":"Using Keras with ProcessPoolExecutor gives \"TypeError: can't pickle _thread.RLock objects\" when retrieving result() from Futures object","Body":"<p>I am trying to parallelize the k-fold validation example for Boston Housing Prices regression in Francois Chollet's Deep Learning with Python book. This builds a small NN for regression using Keras.  The NN is trained on four different training examples obtained by changing the holdout examples for validation.  These are four independent training runs that should be easy to parallelize, or so I thought.  <\/p>\n\n<p>I am running Python 3.7.3 and fairly recent Keras under Anaconda in a Jupiter notebook.  I decided to try and use the concurrent.futures library for parallelization.  <\/p>\n\n<p>I have looked at other related StackOverflow questions and they don't solve my problem. <a href=\"https:\/\/stackoverflow.com\/questions\/47066635\/checkpointing-keras-model-typeerror-cant-pickle-thread-lock-objects\">Checkpointing keras model: TypeError: can&#39;t pickle _thread.lock objects<\/a>\n<a href=\"https:\/\/stackoverflow.com\/questions\/54421480\/cant-pickle-thread-rlock-objects-using-keras\">can&#39;t pickle _thread.RLock objects using keras<\/a>\netc...<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport tensorflow as tf\n\nboston_housing = tf.keras.datasets.boston_housing\nmodels = tf.keras.models\nlayers = tf.keras.layers\n\n# load data\n(train_data, train_targets), (test_data, test_targets) = boston_housing.load_data()\n\n# normalize the features (subtract mean and divide by standard deviation)\nmean = train_data.mean(axis=0)\nstd = train_data.std(axis=0)\ntrain_data -= mean\ntrain_data \/= std\ntest_data -= mean\ntest_data \/= std\n\n# model definition\ndef build_model():\n    model = models.Sequential()\n    model.add(layers.Dense(64, activation='relu', input_shape = (train_data.shape[1],)))\n    model.add(layers.Dense(64, activation='relu'))\n    model.add(layers.Dense(1))\n    model.compile(optimizer='rmsprop', loss='mse', metrics=['mae'])\n    return model\n\n# build and train model on one fold of training\/validation split\ndef build_and_train_model( fold, num_folds = 4, num_epochs = 100 ):\n    num_val_samples = len(train_data) \/\/ num_folds\n    start_idx = fold*num_val_samples\n    end_idx = (fold+1)*num_val_samples\n    partial_train_data = np.concatenate(\n        [train_data[:end_idx], train_data[end_idx:]], axis = 0)\n    partial_train_targets = np.concatenate(\n        [train_targets[:end_idx], train_targets[end_idx:]], axis = 0)\n    model = build_model()\n    history = model.fit(partial_train_data, partial_train_targets, epochs=num_epochs, \n                        batch_size=1, verbose=0)\n    return (model, history)\n\nfrom concurrent.futures import ProcessPoolExecutor\n\nk = 4\nnum_epochs = 10\nfutures = []\nwith ProcessPoolExecutor() as executor:\n    for i in range(k):\n        future = executor.submit(build_and_train_model, i, \n                                 num_folds = k, num_epochs = num_epochs )\n        futures.append(future)\n    models = []\n    histories = []\n    for future in futures:\n        model, history = future.result() \n        models.append(models)\n        histories.append(history)\n<\/code><\/pre>\n\n<pre><code>This is the error message:\n\n---------------------------------------------------------------------------\n_RemoteTraceback                          Traceback (most recent call last)\n_RemoteTraceback: \n\"\"\"\nTraceback (most recent call last):\n  File \"\/\/anaconda3\/envs\/tf\/lib\/python3.7\/concurrent\/futures\/process.py\", line 205, in _sendback_result\n    exception=exception))\n  File \"\/\/anaconda3\/envs\/tf\/lib\/python3.7\/multiprocessing\/queues.py\", line 358, in put\n    obj = _ForkingPickler.dumps(obj)\n  File \"\/\/anaconda3\/envs\/tf\/lib\/python3.7\/multiprocessing\/reduction.py\", line 51, in dumps\n    cls(buf, protocol).dump(obj)\nTypeError: can't pickle _thread.RLock objects\n\"\"\"\n\nThe above exception was the direct cause of the following exception:\n\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-6-499a12f73b9e&gt; in &lt;module&gt;\n     13     histories = []\n     14     for future in futures:\n---&gt; 15         model, history = future.result()\n     16         models.append(models)\n     17         histories.append(history)\n\n\/\/anaconda3\/envs\/tf\/lib\/python3.7\/concurrent\/futures\/_base.py in result(self, timeout)\n    433                 raise CancelledError()\n    434             elif self._state == FINISHED:\n--&gt; 435                 return self.__get_result()\n    436             else:\n    437                 raise TimeoutError()\n\n\/\/anaconda3\/envs\/tf\/lib\/python3.7\/concurrent\/futures\/_base.py in __get_result(self)\n    382     def __get_result(self):\n    383         if self._exception:\n--&gt; 384             raise self._exception\n    385         else:\n    386             return self._result\n\nTypeError: can't pickle _thread.RLock objects\n<\/code><\/pre>\n\n<p>Is there a way to get around this error?  Or is there a different way to achieve what I am trying to do -- namely launch multiple parallel training sessions?<\/p>\n","answers":[]}
+{"QuestionId":57778176,"AnswerCount":0,"Tags":"<pytorch>","CreationDate":"2019-09-03T20:12:32.497","AcceptedAnswerId":null,"OwnerUserId":3259896.0,"Title":"Creating Pytorch variables directly on GPU seems to still create variable on CPU first, judging by CPU RAM usage","Body":"<p>I need to create variables directly on the GPU because I am very limited in my CPU ram. <\/p>\n\n<p>I found the method to do this here <\/p>\n\n<p><a href=\"https:\/\/discuss.pytorch.org\/t\/how-to-create-a-tensor-on-gpu-as-default\/2128\/3\" rel=\"nofollow noreferrer\">https:\/\/discuss.pytorch.org\/t\/how-to-create-a-tensor-on-gpu-as-default\/2128\/3<\/a><\/p>\n\n<p>Which mentions using <\/p>\n\n<p><code>torch.set_default_tensor_type(\u2018torch.cuda.FloatTensor\u2019)<\/code><\/p>\n\n<p>However, when I tried <\/p>\n\n<pre><code>torch.set_default_tensor_type('torch.cuda.FloatTensor')\npytorchGPUDirectCreate = torch.FloatTensor(20000000, 128).uniform_(-1, 1).cuda()\n<\/code><\/pre>\n\n<p>It still seemed to take up mostly CPU RAM, before being transferred to GPU ram. <\/p>\n\n<p>I am using Google Colab. To view RAM usage during the variable creation process, after running the cell, go to Runtime -> Manage Sessions <\/p>\n\n<p>With and without using <code>torch.set_default_tensor_type('torch.cuda.FloatTensor')<\/code> , the CPU RAM bumps up to 11.34 GB while GPU ram stays low, and then GPU RAM goes to 9.85 and CPU ram goes back down.<\/p>\n\n<p>It seems that <code>torch.set_default_tensor_type(\u2018torch.cuda.FloatTensor\u2019)<\/code> didn't make a difference<\/p>\n\n<p>For convenience here's a direct link to a notebook anyone can directly run <\/p>\n\n<p><a href=\"https:\/\/colab.research.google.com\/drive\/1LxPMHl8yFAATH0i0PBYRURo5tqj0_An7\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/1LxPMHl8yFAATH0i0PBYRURo5tqj0_An7<\/a><\/p>\n","answers":[]}
+{"QuestionId":57778555,"AnswerCount":0,"Tags":"<python><tensorflow><session>","CreationDate":"2019-09-03T20:48:37.047","AcceptedAnswerId":null,"OwnerUserId":6188704.0,"Title":"FailedPreconditionError: Attempting to use uninitialized value conv2d_1_89\/kernel [[{{node _retval_conv2d_1_89\/kernel_0_1}}]]","Body":"<p>Trying to train a game (Sonic) with Tensorflow using Reinforcement Learning. After 43 Experiments\/runs of my first training i get the error. <\/p>\n\n<pre><code>FailedPreconditionError: Attempting to use uninitialized value conv2d_1_89\/kernel\n [[{{node _retval_conv2d_1_89\/kernel_0_1}}]]\n<\/code><\/pre>\n\n<p>I really do not know what it means and how to fix it. Help would be great.<\/p>\n\n<p>Our code can be seen here: <a href=\"https:\/\/github.com\/ofekluis\/sonic_project_ss19\/blob\/master\/sonic\/sonic_multi_input.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ofekluis\/sonic_project_ss19\/blob\/master\/sonic\/sonic_multi_input.py<\/a><\/p>\n\n<p>Starting tensorflow sessions at line 96 and 124. \nI would appreciate any kind of help.<\/p>\n","answers":[]}
+{"QuestionId":57779022,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-03T21:35:38.510","AcceptedAnswerId":57780517.0,"OwnerUserId":4496896.0,"Title":"ValueError: axes don't match array when loading previously saved model","Body":"<p><strong>System information<\/strong>  <\/p>\n\n<ul>\n<li>Have I written custom code (as opposed to using example directory):  Yes<\/li>\n<li>OS Platform and Distribution (e.g., Linux Ubuntu 16.04):  Linux Ubuntu 16.04<\/li>\n<li>TensorFlow backend (yes \/ no):  Yes<\/li>\n<li>TensorFlow version:  1.14.0<\/li>\n<li>Keras version:  2.2.5<\/li>\n<li>Python version:  3.6.8<\/li>\n<li>CUDA\/cuDNN version:  N\/A<\/li>\n<li>GPU model and memory:  N\/A<\/li>\n<\/ul>\n\n<p><strong>Describe the current behavior:<\/strong>  <\/p>\n\n<ul>\n<li>Error when loading model with <code>model = load_model(file.h5)<\/code><\/li>\n<\/ul>\n\n<pre><code>ValueError: axes don't match array\n<\/code><\/pre>\n\n<p><strong>Describe the expected behavior<\/strong>  <\/p>\n\n<ul>\n<li>Model is not loading again after saving with <code>model.save(file.h5)<\/code><\/li>\n<\/ul>\n\n<p><strong>Here is what I'm trying to do:<\/strong><\/p>\n\n<ul>\n<li>I have 15 models as a single <code>merged_model<\/code> which are basically just the classification models with 15 inputs and 15 outputs.<\/li>\n<li>I'm trying to combine the these 15 model's input into a single input model. So I don't have to provide 15 inputs!<\/li>\n<li>Here is how I'm doing that: <strong><em>(this is working without any issue)<\/em><\/strong><\/li>\n<\/ul>\n\n<pre><code>&gt;&gt; model_single_input = layers.Input((15,), dtype='int32', name='single.input')\n&gt;&gt; model_multiple_inputs = layers.Lambda(lambda x: [x] * 15, name='single.input.multiplier')(model_single_input)\n&gt;&gt; single_input_model = Model(inputs=model_single_input, outputs=model_multiple_inputs)\n&gt;&gt; single_input_model.input, single_input_model.output\n\n(&lt;tf.Tensor 'single.input:0' shape=(?, 15) dtype=int32&gt;,\n [&lt;tf.Tensor 'single.input.multiplier\/Identity:0' shape=(?, 15) dtype=int32&gt;,\n  &lt;tf.Tensor 'single.input.multiplier\/Identity_1:0' shape=(?, 15) dtype=int32&gt;,\n  &lt;tf.Tensor 'single.input.multiplier\/Identity_2:0' shape=(?, 15) dtype=int32&gt;,\n  &lt;tf.Tensor 'single.input.multiplier\/Identity_3:0' shape=(?, 15) dtype=int32&gt;,\n  &lt;tf.Tensor 'single.input.multiplier\/Identity_4:0' shape=(?, 15) dtype=int32&gt;,\n  &lt;tf.Tensor 'single.input.multiplier\/Identity_5:0' shape=(?, 15) dtype=int32&gt;,\n  &lt;tf.Tensor 'single.input.multiplier\/Identity_6:0' shape=(?, 15) dtype=int32&gt;,\n  &lt;tf.Tensor 'single.input.multiplier\/Identity_7:0' shape=(?, 15) dtype=int32&gt;,\n  &lt;tf.Tensor 'single.input.multiplier\/Identity_8:0' shape=(?, 15) dtype=int32&gt;,\n  &lt;tf.Tensor 'single.input.multiplier\/Identity_9:0' shape=(?, 15) dtype=int32&gt;,\n  &lt;tf.Tensor 'single.input.multiplier\/Identity_10:0' shape=(?, 15) dtype=int32&gt;,\n  &lt;tf.Tensor 'single.input.multiplier\/Identity_11:0' shape=(?, 15) dtype=int32&gt;,\n  &lt;tf.Tensor 'single.input.multiplier\/Identity_12:0' shape=(?, 15) dtype=int32&gt;,\n  &lt;tf.Tensor 'single.input.multiplier\/Identity_13:0' shape=(?, 15) dtype=int32&gt;,\n  &lt;tf.Tensor 'single.input.multiplier\/Identity_14:0' shape=(?, 15) dtype=int32&gt;])\n<\/code><\/pre>\n\n<ul>\n<li>Now here is I'm combining single input model with 15 models. <strong><em>(this is working without any issue)<\/em><\/strong><\/li>\n<\/ul>\n\n<pre><code>&gt;&gt; single_input_merged_output_model = Model(inputs  = single_input_model.input, outputs = merged_model(single_input_model.output))\n&gt;&gt; encoded_data = np.array([\n[12073, 14512, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],\n[336, 0, 744, 481, 13043, 118, 2563, 0, 0, 0, 0, 0, 0, 0, 0]\n])\n&gt;&gt; predictions = single_input_merged_output_model.predict(encoded_data)\n&gt;&gt; predictions\n\n[array([[ 0.        , 18.        ,  0.23679169],\n        [ 0.        , 13.        ,  0.5127094 ]], dtype=float32),\n array([[1.0000000e+00, 2.0700000e+02, 4.9950428e-02],\n        [1.0000000e+00, 9.2000000e+01, 3.4491304e-01]], dtype=float32),\n array([[  2.       , 229.       ,   0.9984485],\n        [  4.       ,  60.       ,   0.9372796]], dtype=float32),\n array([[2.000000e+00, 1.194000e+03, 9.985555e-01],\n        [3.000000e+00, 1.030000e+02, 9.584518e-01]], dtype=float32),\n array([[2.000000e+00, 1.558000e+03, 9.996946e-01],\n        [3.000000e+00, 8.800000e+01, 9.738545e-01]], dtype=float32),\n array([[2.000000e+00, 1.997000e+03, 9.998343e-01],\n        [7.000000e+00, 7.020000e+02, 9.954461e-01]], dtype=float32),\n array([[2.0000000e+00, 1.7690000e+03, 9.9997449e-01],\n        [3.0000000e+00, 1.7900000e+02, 9.9776447e-01]], dtype=float32),\n array([[2.000000e+00, 1.448000e+03, 9.999393e-01],\n        [3.000000e+00, 2.430000e+02, 9.982481e-01]], dtype=float32),\n array([[2.0000000e+00, 1.0770000e+03, 9.9984264e-01],\n        [3.0000000e+00, 2.0700000e+02, 9.9882430e-01]], dtype=float32),\n array([[  2.        , 754.        ,   0.9998847 ],\n        [  3.        , 493.        ,   0.99971205]], dtype=float32),\n array([[  2.       , 536.       ,   0.9996455],\n        [  3.       , 239.       ,   0.9998828]], dtype=float32),\n array([[  2.        , 444.        ,   0.99973446],\n        [  3.        ,  98.        ,   0.99974567]], dtype=float32),\n array([[8.0000000e+00, 1.0400000e+02, 1.3962857e-01],\n        [2.0000000e+00, 2.3600000e+02, 7.3362941e-01]], dtype=float32),\n array([[ 2.        , 34.        ,  0.06541887],\n        [ 2.        , 46.        ,  0.3399737 ]], dtype=float32),\n array([[ 2.        , 52.        ,  0.24562976],\n        [ 2.        ,  7.        ,  0.5339988 ]], dtype=float32)]\n<\/code><\/pre>\n\n<ul>\n<li>Here is how I'm saving the final model. <strong><em>(this is working without any issue)<\/em><\/strong><\/li>\n<\/ul>\n\n<pre><code>&gt;&gt; single_input_merged_output_model.save('file.h5', include_optimizer=False)\n<\/code><\/pre>\n\n<ul>\n<li>...and here is the issue when I try to load this above model which I just saved!<\/li>\n<\/ul>\n\n<pre><code>&gt;&gt; single_input_merged_output_model = load_model('file.h5', compile=False)\n<\/code><\/pre>\n\n<ul>\n<li><strong>Error:<\/strong><\/li>\n<\/ul>\n\n<pre><code>---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;timed exec&gt; in &lt;module&gt;\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/saving.py in load_wrapper(*args, **kwargs)\n    456                 os.remove(tmp_filepath)\n    457             return res\n--&gt; 458         return load_function(*args, **kwargs)\n    459 \n    460     return load_wrapper\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/saving.py in load_model(filepath, custom_objects, compile)\n    548     if H5Dict.is_supported_type(filepath):\n    549         with H5Dict(filepath, mode='r') as h5dict:\n--&gt; 550             model = _deserialize_model(h5dict, custom_objects, compile)\n    551     elif hasattr(filepath, 'write') and callable(filepath.write):\n    552         def load_function(h5file):\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/saving.py in _deserialize_model(h5dict, custom_objects, compile)\n    290                                                        original_keras_version,\n    291                                                        original_backend,\n--&gt; 292                                                        reshape=False)\n    293         if len(weight_values) != len(symbolic_weights):\n    294             raise ValueError('Layer #' + str(k) +\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/saving.py in preprocess_weights_for_loading(layer, weights, original_keras_version, original_backend, reshape)\n    821         weights = convert_nested_time_distributed(weights)\n    822     elif layer.__class__.__name__ in ['Model', 'Sequential']:\n--&gt; 823         weights = convert_nested_model(weights)\n    824 \n    825     if original_keras_version == '1':\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/saving.py in convert_nested_model(weights)\n    809                     weights=weights[:num_weights],\n    810                     original_keras_version=original_keras_version,\n--&gt; 811                     original_backend=original_backend))\n    812                 weights = weights[num_weights:]\n    813         return new_weights\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/saving.py in preprocess_weights_for_loading(layer, weights, original_keras_version, original_backend, reshape)\n    821         weights = convert_nested_time_distributed(weights)\n    822     elif layer.__class__.__name__ in ['Model', 'Sequential']:\n--&gt; 823         weights = convert_nested_model(weights)\n    824 \n    825     if original_keras_version == '1':\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/saving.py in convert_nested_model(weights)\n    797                     weights=weights[:num_weights],\n    798                     original_keras_version=original_keras_version,\n--&gt; 799                     original_backend=original_backend))\n    800                 weights = weights[num_weights:]\n    801 \n\n~\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/saving.py in preprocess_weights_for_loading(layer, weights, original_keras_version, original_backend, reshape)\n    940             weights[0] = np.reshape(weights[0], layer_weights_shape)\n    941         elif layer_weights_shape != weights[0].shape:\n--&gt; 942             weights[0] = np.transpose(weights[0], (3, 2, 0, 1))\n    943             if layer.__class__.__name__ == 'ConvLSTM2D':\n    944                 weights[1] = np.transpose(weights[1], (3, 2, 0, 1))\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/numpy\/core\/fromnumeric.py in transpose(a, axes)\n    637 \n    638     \"\"\"\n--&gt; 639     return _wrapfunc(a, 'transpose', axes)\n    640 \n    641 \n\n~\/anaconda3\/lib\/python3.6\/site-packages\/numpy\/core\/fromnumeric.py in _wrapfunc(obj, method, *args, **kwds)\n     54 def _wrapfunc(obj, method, *args, **kwds):\n     55     try:\n---&gt; 56         return getattr(obj, method)(*args, **kwds)\n     57 \n     58     # An AttributeError occurs if the object does not have\n\nValueError: axes don't match array\n<\/code><\/pre>\n\n<p><strong>Things that I have already tried:<\/strong><\/p>\n\n<ul>\n<li>Saving model architecture and weights separately and loading <strong>(Not Working)<\/strong><\/li>\n<\/ul>\n\n<p>Any suggestions on loading model? <\/p>\n","answers":[{"AnswerId":"57780517","CreationDate":"2019-09-04T01:49:03.767","ParentId":null,"OwnerUserId":"4496896","Title":null,"Body":"<p>Solved it! If you freeze weights of the model's layers before <code>model.save()<\/code>, and then save the model; <code>load_model()<\/code> works without any issue! This only works when you don't want to re-train your model further.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.models import Model\n\ndef freeze_layers(model):\n    for i in model.layers:\n        i.trainable = False\n        if isinstance(i, Model):\n            freeze_layers(i)\n    return model\n\n&gt;&gt; model_freezed = freeze_layers(model)\n&gt;&gt; model_freezed.save('file.tf')\n\n# refresh the notebook\nfrom keras.models import load_model\n&gt;&gt; model = load_model('file.tf', compile=False)\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":57779287,"AnswerCount":0,"Tags":"<python><numpy><tensorflow>","CreationDate":"2019-09-03T22:08:42.620","AcceptedAnswerId":null,"OwnerUserId":5085872.0,"Title":"Is there a faster way to find pairwise distances in tensorflow?","Body":"<p>I want to compute the pairwise distance between all points stored in some 2D tensor ('batch_size' points in 'd' dimensions).<\/p>\n\n<p>I currently am using the following tensorflow code, which is elegant, but I'm hoping there's a faster solution, as the time cost per iteration depends heavily on this step.<\/p>\n\n<pre><code>'''python\n Z = tf.Variable(np.empty((batch_size,d))\n Z = tf.reshape(Zy, (batch_size, 1, d)) - \n        tf.reshape(Z, (1, batch_size, d))\n norms = tf.norm(Z, axis=2)\n'''\n<\/code><\/pre>\n\n<p>Basically what this does is make an intermediate 3D tensor Z where Z(i,j,k) denotes the difference between point i and point j in their k'th coordinate. Any thoughts?<\/p>\n","answers":[]}
+{"QuestionId":57779517,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-03T22:43:55.400","AcceptedAnswerId":null,"OwnerUserId":880783.0,"Title":"How can this CNN have an odd number of parameters?","Body":"<p>I am planning to work with a 3-D U-Net, and for first steps I am following the paper by \u00c7i\u00e7ek et al, <a href=\"https:\/\/arxiv.org\/abs\/1606.06650\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1606.06650<\/a>, from which I reproduce the network structure here:\n<a href=\"https:\/\/i.stack.imgur.com\/T4iBN.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/T4iBN.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>They reportedly use a number 19,069,955 parameters. This number is odd (by which I mean, not divisible by 2). My own Keras implementations get quite close to that number (e.g., 19,067,648 parameters when using 2 instead of 3 input channels, or 19,068,448 when using 2 instead of 3 target classes). But whatever I try, my numbers of (trainable, non-trainable, total - in each layer) parameters remains even. How can this network have an odd number of parameters?<\/p>\n","answers":[]}
+{"QuestionId":57779967,"AnswerCount":2,"Tags":"<python><keras><tensorflow.js><transfer-learning><tensorflowjs-converter>","CreationDate":"2019-09-04T00:07:12.643","AcceptedAnswerId":null,"OwnerUserId":10826930.0,"Title":"Uncaught (in promise) Error: Error when checking : expected flatten_1_input to have shape [null,7,7,512] but got array with shape [1,224,224,3]","Body":"<p>I am using the concept of transfer learning to classify my image, I have reused the code mentioned at <a href=\"https:\/\/towardsdatascience.com\/transfer-learning-from-pre-trained-models-f2393f124751\" rel=\"nofollow noreferrer\">https:\/\/towardsdatascience.com\/transfer-learning-from-pre-trained-models-f2393f124751<\/a><\/p>\n\n<p>the model is working well for my data(<a href=\"https:\/\/www.dropbox.com\/s\/esirpr6q1lsdsms\/ricetransfer1.zip?dl=0\" rel=\"nofollow noreferrer\">https:\/\/www.dropbox.com\/s\/esirpr6q1lsdsms\/ricetransfer1.zip?dl=0<\/a>) in my jupyter notebook, but before testing the model, I am reshaping the image.<\/p>\n\n<p>But when I wanted to run the same model in the browser using TensorFlow, I used tfjs.converters.save_keras_model to save my model.<\/p>\n\n<pre><code>from keras.applications import VGG16\nimport tensorflowjs as tfjs\nimport tensorflow as tf\ntf.compat.v1.disable_eager_execution()\n\nimg_width, img_height = 224, 224  # Default input size for VGG16\n\nconv_base = VGG16(weights='imagenet', \n              include_top=False,\n              #input_shape=(224, 224, 3))\n              input_shape=(img_width, img_height, 3))\n\n# Extract features\nimport os, shutil\nfrom keras.preprocessing.image import ImageDataGenerator\nimport numpy as np\ntrain_size, validation_size, test_size = 148, 27, 31\n\ndatagen = ImageDataGenerator(rescale=1.\/255)\nbatch_size = 16\ntrain_dir = \"ricetransfer1\/train\"\nvalidation_dir = \"ricetransfer1\/validation\"\ntest_dir=\"ricetransfer1\/test\"\n#indices = np.random.choice(range(len(X_train)))\n\ndef extract_features(directory, sample_count):\n   #sample_count= X_train.ravel()\n\n    features = np.zeros(shape=(sample_count, 7, 7, 512))  # Must be \n    equal to the output of the convolutional base\n    labels = np.zeros(shape=(sample_count))\n    # Preprocess data\n    generator = datagen.flow_from_directory(directory,\n                                        target_size=(img_width,img_height),\n                                        batch_size = batch_size,\n                                        class_mode='binary')\n    # Pass data through convolutional base\n    i = 0\n    for inputs_batch, labels_batch in generator:\n        features_batch = conv_base.predict(inputs_batch)\n        features[i * batch_size: (i + 1) * batch_size] = features_batch\n        labels[i * batch_size: (i + 1) * batch_size] = labels_batch\n        i += 1\n        if i * batch_size &gt;= sample_count:\n            break\nreturn features, labels\n\ntrain_features, train_labels = extract_features(train_dir, train_size)  \nvalidation_features, validation_labels = extract_features(validation_dir, validation_size)\ntest_features, test_labels = extract_features(test_dir, test_size)\n\n\nfrom keras import models\nfrom keras import layers\nfrom keras import optimizers\n\nepochs = 1\n\n ricemodel = models.Sequential()\n ricemodel.add(layers.Flatten(input_shape=(7,7,512)),)\n ricemodel.add(layers.Dense(256, activation='relu', input_dim=(7*7*512)))\n ricemodel.add(layers.Dropout(0.5))\n ricemodel.add(layers.Dense(1, activation='sigmoid'))\n ricemodel.summary()\n\n ricemodel.compile(optimizer=optimizers.Adam(),\n          loss='binary_crossentropy',\n          metrics=['acc'])\n\n import os\n history=ricemodel.fit(train_features, train_labels,\n                epochs=epochs,\n                batch_size=batch_size, \n                validation_data=(validation_features, validation_labels))\n\npath='\\vgg'\ntfjs.converters.save_keras_model(ricemodel, path)\n<\/code><\/pre>\n\n<p>The TensorFlowjs code is <\/p>\n\n<pre><code> $(document).ready()\n{\n  $('.progress-bar').hide();\n}\n$(\"#image-selector\").change(function(){\nlet reader = new FileReader();\n\nreader.onload = function(){\n    let dataURL = reader.result;\n    $(\"#selected-image\").attr(\"src\",dataURL);\n    $(\"#prediction-list\").empty();\n}\nlet file = $(\"#image-selector\").prop('files')[0];\nreader.readAsDataURL(file);\n});\n\n\n$(\"#model-selector\").change(function(){\nloadModel($(\"#model-selector\").val());\n$('.progress-bar').show();\n})\n\nlet model;\nasync function loadModel(name){\nmodel=await tf.loadModel(`http:\/\/localhost:8081\/${name}\/model.json`);\n$('.progress-bar').hide();\n}\n\n\n$(\"#predict-button\").click(async function(){\nlet image= $('#selected-image').get(0);\nlet tensor = preprocessImage(image,$(\"#model-selector\").val());\n\nlet prediction = await model.predict(tensor).data();\nlet top5=Array.from(prediction)\n            .map(function(p,i){\nreturn {\n    probability: p,\n    className: IMAGENET_CLASSES[i]\n};\n}).sort(function(a,b){\n    return b.probability-a.probability;\n}).slice(0,5);\n\n$(\"#prediction-list\").empty();\ntop5.forEach(function(p){\n$(\"#prediction- \n list\").append(`&lt;li&gt;${p.className}:${p.probability.toFixed(6)}&lt;\/li&gt;`);\n});\n\n});\n\n\nfunction preprocessImage(image,modelName)\n{\n let tensor=tf.fromPixels(image)\n.resizeNearestNeighbor([224,224])\n.toFloat();\/\/.sub(meanImageNetRGB)\n\n if(modelName==undefined)\n {\n    return tensor.expandDims();\n}\nelse if(modelName==\"vgg\")\n{\n    let meanImageNetRGB= tf.tensor1d([123.68,116.779,103.939]);\n    return tensor.sub(meanImageNetRGB)\n                .reverse(2)\n                .expandDims();\n}\nelse if(modelName==\"mobilenet\")\n{\n    let offset=tf.scalar(127.5);\n    return tensor.sub(offset)\n                .div(offset)\n                .expandDims();\n}\nelse\n{\n    throw new Error(\"UnKnown Model error\");\n}\n}\n<\/code><\/pre>\n\n<p>Error Message in the browser after loading the model in the tensorflowjs, I get the below error(I can see the error message in the web development console)<\/p>\n\n<blockquote>\n  <p>Uncaught (in promise) Error: Error when checking : expected flatten_1_input to have shape [null,7,7,512] but got array with shape [1,224,224,3]<\/p>\n<\/blockquote>\n\n<p>Is there any way to solve the issue, can I reshape the image before loading to the browser?<\/p>\n\n<p>I tried all the possible option but I am stuck now.<\/p>\n\n<p>I have already checked the possible solutions on the stackoverflow. How can I run classification model on the browser?<\/p>\n","answers":[{"AnswerId":"57787350","CreationDate":"2019-09-04T11:22:14.373","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>As pointed out by the answer of <a href=\"https:\/\/stackoverflow.com\/questions\/57779967\/uncaught-in-promise-error-error-when-checking-expected-flatten-1-input-to-h\/57780279#57780279\">@andyPotato<\/a>, you need to convert the feature extractor model as well to js.<\/p>\n\n<pre><code>from tensorflow.python.keras.applications.vgg16 import VGG16\nfrom tensorflow.python.keras.models import save_model\n\n#download model \nmodel = VGG16(weights='imagenet', \n          include_top=False,\n          input_shape=(width, height, 3)) # tune parameters\n\n#save the model \nsave_model(\n    model,\n    \"vgg.h5\",\n    overwrite=True,\n)\n<\/code><\/pre>\n\n<p>Convert VGG feature extractor to js<\/p>\n\n<pre><code>tensorflowjs_converter --input_format keras \\\n                       path\/to\/my_model.h5 \\\n                       path\/to\/tfjs_target_dir\n<\/code><\/pre>\n\n<p>Now use both model for inference in js<\/p>\n\n<pre class=\"lang-js prettyprint-override\"><code>model1 = await tf.loadModel(`\/url\/of\/vgg\/converted\/model.json`);\n\nfeatureExtracted = await model1.predict(image)\n\nmodel2 = await tf.loadModel(`\/url\/of\/sequential\/model\/model.json`);\n\nprediction = await model2.predict(featureExtracted)\n<\/code><\/pre>\n"},{"AnswerId":"57780279","CreationDate":"2019-09-04T01:07:29.900","ParentId":null,"OwnerUserId":"11697360","Title":null,"Body":"<p>The issue is you are creating a completely new model <code>ricemodel<\/code> which looks like a fully connected layer and save it as a standalone model without the convolutional base below (VGG in your case). This is why your model has an input layer shaped [,7,7,512] (feature vector) and not [,224,224,3] (raw image data).<\/p>\n\n<p>To solve this you need to first load the VGG model with its weights pre-trained (for example \"imagenet\"), pop off the top layer and then add your <code>ricemodel<\/code> on top. Finally save the combined new model and export it to tfjs.<\/p>\n"}]}
+{"QuestionId":57779973,"AnswerCount":0,"Tags":"<conv-neural-network><pytorch><graph-databases><chemistry>","CreationDate":"2019-09-04T00:07:53.517","AcceptedAnswerId":null,"OwnerUserId":6392842.0,"Title":"Using edge features for GCN in DGL","Body":"<p>I'm trying to implement a graph convolutional network (GCN) in the Deep Graph Learning (DGL) package for Python. In many papers, edges have discrete features, and each possible value is associated with a different weight matrix or set of weight matrices. An example would be <a href=\"https:\/\/pubs.acs.org\/doi\/full\/10.1021\/acscentsci.8b00507\" rel=\"nofollow noreferrer\">here<\/a>. Is anyone familiar with how to implement a model like this in DGL? The DGL team's example of GCNs for <a href=\"https:\/\/docs.dgl.ai\/tutorials\/basics\/4_batch.html\" rel=\"nofollow noreferrer\">graph classification<\/a>, as does <a href=\"https:\/\/iwatobipen.wordpress.com\/2019\/02\/01\/try-gcn-qspr-with-pytorch-based-graph-library-rdkit-pytorch-dgl\/\" rel=\"nofollow noreferrer\">another example<\/a> I found online.<\/p>\n","answers":[]}
+{"QuestionId":57780053,"AnswerCount":1,"Tags":"<python><tensorflow><keras><keras-layer><autoencoder>","CreationDate":"2019-09-04T00:21:44.827","AcceptedAnswerId":58179673.0,"OwnerUserId":11861082.0,"Title":"How can I implement KL-divergence regularization for Keras?","Body":"<p>This is a follow-up question for this question <a href=\"https:\/\/stackoverflow.com\/questions\/57530823\/keras-backend-mean-function-float-object-has-no-attribute-dtype\">Keras backend mean function: &quot; &#39;float&#39; object has no attribute &#39;dtype&#39; &quot;?<\/a> <\/p>\n\n<p>I am trying to make a new regularizer for Keras. Here is my code<\/p>\n\n<pre><code>import keras\nfrom keras import initializers\nfrom keras.models import Model, Sequential\nfrom keras.layers import Input, Dense, Activation\nfrom keras import regularizers\nfrom keras import optimizers\nfrom keras import backend as K\n\nkullback_leibler_divergence = keras.losses.kullback_leibler_divergence\n\ndef kl_divergence_regularizer(inputs):\n        means = K.mean((inputs))\n        rho=0.05\n        down = 0.05 * K.ones_like(means)\n        up = (1 - 0.05) * K.ones_like(means)\n        return 0.5 *(0.01 * (kullback_leibler_divergence(down, means)\n                      + kullback_leibler_divergence(up, 1 - means)))\n\nmodel = Sequential([\n    Dense(900, input_shape=(x_train_s.shape[1],),kernel_initializer='random_uniform',kernel_regularizer=kl_divergence_regularizer),\n    Activation('elu'),\n    Dense(x_train_s.shape[1],kernel_initializer='random_uniform'),\n    Activation('tanh')\n])\n\nmodel.compile(optimizer='adam',loss='mean_squared_error')\n\nmodel.fit(x_train_s, y_train_s, epochs=5)\n<\/code><\/pre>\n\n<p>Here is the error:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nInvalidArgumentError                      Traceback (most recent call last)\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py in _create_c_op(graph, node_def, inputs, control_inputs)\n   1658   try:\n-&gt; 1659     c_op = c_api.TF_FinishOperation(op_desc)\n   1660   except errors.InvalidArgumentError as e:\n\nInvalidArgumentError: Invalid reduction dimension -1 for input with 0 dimensions. for 'dense_3\/weight_regularizer\/Sum' (op: 'Sum') with input shapes: [], [] and with computed input tensors: input[1] = &lt;-1&gt;.\n\nDuring handling of the above exception, another exception occurred:\n\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-4-9f4dfbe34659&gt; in &lt;module&gt;\n     39     Activation('elu'),\n     40     Dense(x_train_s.shape[1],kernel_initializer='random_uniform'),\n---&gt; 41     Activation('tanh')\n     42 ])\n     43 \n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\engine\\sequential.py in __init__(self, layers, name)\n     91         if layers:\n     92             for layer in layers:\n---&gt; 93                 self.add(layer)\n     94 \n     95     @property\n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\engine\\sequential.py in add(self, layer)\n    163                     # and create the node connecting the current layer\n    164                     # to the input layer we just created.\n--&gt; 165                     layer(x)\n    166                     set_inputs = True\n    167             else:\n\nC:\\ProgramData\\Anaconda3\\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\nC:\\ProgramData\\Anaconda3\\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\nC:\\ProgramData\\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\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\engine\\base_layer.py in add_weight(self, name, shape, dtype, initializer, regularizer, trainable, constraint)\n    253         if regularizer is not None:\n    254             with K.name_scope('weight_regularizer'):\n--&gt; 255                 self.add_loss(regularizer(weight))\n    256         if trainable:\n    257             self._trainable_weights.append(weight)\n\n&lt;ipython-input-4-9f4dfbe34659&gt; in kl_divergence_regularizer(inputs)\n     15         down = 0.05 * K.ones_like(means)\n     16         up = (1 - 0.05) * K.ones_like(means)\n---&gt; 17         return 0.5 *(0.01 * (kullback_leibler_divergence(down, means)\n     18                       + kullback_leibler_divergence(up, 1 - means)))\n     19 \n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\losses.py in kullback_leibler_divergence(y_true, y_pred)\n     81     y_true = K.clip(y_true, K.epsilon(), 1)\n     82     y_pred = K.clip(y_pred, K.epsilon(), 1)\n---&gt; 83     return K.sum(y_true * K.log(y_true \/ y_pred), axis=-1)\n     84 \n     85 \n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py in sum(x, axis, keepdims)\n   1286         A tensor with sum of `x`.\n   1287     \"\"\"\n-&gt; 1288     return tf.reduce_sum(x, axis, keepdims)\n   1289 \n   1290 \n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py in new_func(*args, **kwargs)\n    505                 'in a future version' if date is None else ('after %s' % date),\n    506                 instructions)\n--&gt; 507       return func(*args, **kwargs)\n    508 \n    509     doc = _add_deprecated_arg_notice_to_docstring(\n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\ops\\math_ops.py in reduce_sum_v1(input_tensor, axis, keepdims, name, reduction_indices, keep_dims)\n   1284   keepdims = deprecation.deprecated_argument_lookup(\"keepdims\", keepdims,\n   1285                                                     \"keep_dims\", keep_dims)\n-&gt; 1286   return reduce_sum(input_tensor, axis, keepdims, name)\n   1287 \n   1288 \n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\util\\dispatch.py in wrapper(*args, **kwargs)\n    178     \"\"\"Call target, and fall back on dispatchers if there is a TypeError.\"\"\"\n    179     try:\n--&gt; 180       return target(*args, **kwargs)\n    181     except (TypeError, ValueError):\n    182       # Note: convert_to_eager_tensor currently raises a ValueError, not a\n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\ops\\math_ops.py in reduce_sum(input_tensor, axis, keepdims, name)\n   1332       gen_math_ops._sum(\n   1333           input_tensor, _ReductionDims(input_tensor, axis), keepdims,\n-&gt; 1334           name=name))\n   1335 \n   1336 \n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\ops\\gen_math_ops.py in _sum(input, axis, keep_dims, name)\n   9607   _, _, _op = _op_def_lib._apply_op_helper(\n   9608         \"Sum\", input=input, reduction_indices=axis, keep_dims=keep_dims,\n-&gt; 9609                name=name)\n   9610   _result = _op.outputs[:]\n   9611   _inputs_flat = _op.inputs\n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\op_def_library.py in _apply_op_helper(self, op_type_name, name, **keywords)\n    786         op = g.create_op(op_type_name, inputs, output_types, name=scope,\n    787                          input_types=input_types, attrs=attr_protos,\n--&gt; 788                          op_def=op_def)\n    789       return output_structure, op_def.is_stateful, op\n    790 \n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py in new_func(*args, **kwargs)\n    505                 'in a future version' if date is None else ('after %s' % date),\n    506                 instructions)\n--&gt; 507       return func(*args, **kwargs)\n    508 \n    509     doc = _add_deprecated_arg_notice_to_docstring(\n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py in create_op(***failed resolving arguments***)\n   3298           input_types=input_types,\n   3299           original_op=self._default_original_op,\n-&gt; 3300           op_def=op_def)\n   3301       self._create_op_helper(ret, compute_device=compute_device)\n   3302     return ret\n\nC:\\ProgramData\\Anaconda3\\lib\\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   1821           op_def, inputs, node_def.attr)\n   1822       self._c_op = _create_c_op(self._graph, node_def, grouped_inputs,\n-&gt; 1823                                 control_input_ops)\n   1824 \n   1825     # Initialize self._outputs.\n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py in _create_c_op(graph, node_def, inputs, control_inputs)\n   1660   except errors.InvalidArgumentError as e:\n   1661     # Convert to ValueError for backwards compatibility.\n-&gt; 1662     raise ValueError(str(e))\n   1663 \n   1664   return c_op\n\nValueError: Invalid reduction dimension -1 for input with 0 dimensions. for 'dense_3\/weight_regularizer\/Sum' (op: 'Sum') with input shapes: [], [] and with computed input tensors: input[1] = &lt;-1&gt;.\n<\/code><\/pre>\n\n<p>How can I fix this? I need the KL divergence between 0.05 and mean calculate the following sum over i:<\/p>\n\n<p>KL=sum(0.05*\\log(0.05\/mean[i]))<\/p>\n","answers":[{"AnswerId":"58179673","CreationDate":"2019-10-01T07:08:30.130","ParentId":null,"OwnerUserId":"9298534","Title":null,"Body":"<p>In order to print means,<\/p>\n\n<pre><code>means = K.means((input), axis=1)\n... \nmeans_ = sess.run(means, feed_dict={x:  , y:   })\nprint(means_) \n<\/code><\/pre>\n"}]}
+{"QuestionId":57780150,"AnswerCount":0,"Tags":"<python><pytorch>","CreationDate":"2019-09-04T00:40:48.107","AcceptedAnswerId":null,"OwnerUserId":3259896.0,"Title":"Performing normal or uniform initializations on Float tensors result in only zeros","Body":"<pre><code>import torch\npytorchGPUDirectCreateWEmpty = torch.empty(size=(20000000, 128), dtype=torch.float, device='cuda', requires_grad=False, pin_memory=False).uniform_(-1, 1)\npytorchGPUDirectCreateWEmpty\n<\/code><\/pre>\n\n<p>results in <\/p>\n\n<pre><code>tensor([[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.]], device='cuda:0')\n<\/code><\/pre>\n\n<p>and <\/p>\n\n<pre><code>import torch\ntorch.set_default_tensor_type('torch.cuda.FloatTensor')\nu_embeddings = torch.nn.Embedding(20000000, 128, padding_idx=None, max_norm=None, norm_type=2.0, scale_grad_by_freq=False, sparse=False, _weight=None)\nu_embeddings.weight.data.uniform_(-1, 1)\nu_embeddings.weight.data\n<\/code><\/pre>\n\n<p>results in <\/p>\n\n<pre><code>tensor([[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<\/code><\/pre>\n\n<p>If I initialize with <code>double<\/code> instead of <code>float<\/code> and the initialization works fine. I could convert to float later, but I am working with limited memory and unable to initialize a double first before converting. <\/p>\n\n<p>Why is the initialization not working for float tensors? <\/p>\n","answers":[]}
+{"QuestionId":57780183,"AnswerCount":0,"Tags":"<c++><c><tensorflow>","CreationDate":"2019-09-04T00:49:03.140","AcceptedAnswerId":null,"OwnerUserId":4241727.0,"Title":"Custom Op in Tensorflow C","Body":"<p>I'm trying to create a TF custom op using Tensorflow C API. I've already read the <a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/op\" rel=\"nofollow noreferrer\">Adding a New Op<\/a> and <a href=\"https:\/\/www.tensorflow.org\/install\/lang_c\" rel=\"nofollow noreferrer\">Install TensorFlow for C<\/a> but there is almost no tutorial for creating a custom op using Tensorflow C. So, I was wondering if somebody can show me a small example code for it?<\/p>\n\n<p>I cannot use the bazel build system in my environment, so I assume I cannot use Tensorflow C++ API, since I only have to use a static or shared library for the custom op code. Please let me know if this task is still possible in Tensorflow C++ API.<\/p>\n","answers":[]}
+{"QuestionId":57780983,"AnswerCount":3,"Tags":"<python-3.x><tensorflow><jupyter-notebook>","CreationDate":"2019-09-04T03:08:11.957","AcceptedAnswerId":null,"OwnerUserId":12015469.0,"Title":"Error when verifying tensorflow installed correctly","Body":"<p>After installing tensorflow via this guide - <a href=\"https:\/\/github.com\/jeffheaton\/t81_558_deep_learning\/blob\/master\/t81_558_class_01_1_overview.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/jeffheaton\/t81_558_deep_learning\/blob\/master\/t81_558_class_01_1_overview.ipynb<\/a> - \nI get an error that I do not understand. I can get into  and then I active tf and get a  prompt. <\/p>\n\n<p>After install, I run the following code to verify installation:<\/p>\n\n<pre><code>import sys\n\n    import tensorflow.keras\n    import pandas as pd\n    import sklearn as sk\n    import tensorflow as tf\n\n    print(f\"Tensor Flow Version: {tf.__version__}\")\n    print(f\"Keras Version: {tensorflow.keras.__version__}\")\n    print()\n    print(f\"Python {sys.version}\")\n    print(f\"Pandas {pd.__version__}\")\n    print(f\"Scikit-Learn {sk.__version__}\")\n    print(\"GPU is\", \"available\" if tf.test.is_gpu_available() else \"NOT AVAILABLE\")\n<\/code><\/pre>\n\n<p>This is the error I get:<\/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            241         else:\n        --&gt; 242             return load_dynamic(name, filename, file)\n            243     elif type_ == PKG_DIRECTORY:\n\n        ~\\Anaconda3\\lib\\imp.py in load_dynamic(name, path, file)\n            341             name=name, loader=loader, origin=path)\n        --&gt; 342         return _load(spec)\n            343 \n\n        ImportError: DLL load failed: A dynamic link library (DLL) initialization routine failed.\n\n        During handling of the above exception, another exception occurred:\n\n        ImportError                               Traceback (most recent call last)\n        &lt;ipython-input-2-0b3eb45acbe5&gt; in &lt;module&gt;\n              1 import sys\n              2 \n        ----&gt; 3 import tensorflow.keras\n              4 import pandas as pd\n              5 import sklearn as sk\n\n        ~\\Anaconda3\\lib\\site-packages\\tensorflow\\__init__.py in &lt;module&gt;\n             38 import sys as _sys\n             39 \n        ---&gt; 40 from tensorflow.python.tools import module_util as _module_util\n             41 \n             42 from tensorflow._api.v2 import audio\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\n        ImportError: Traceback (most recent call last):\n          File \"C:\\Users\\madew\\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\\madew\\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\\madew\\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\\madew\\Anaconda3\\lib\\imp.py\", line 242, in load_module\n            return load_dynamic(name, filename, file)\n          File \"C:\\Users\\madew\\Anaconda3\\lib\\imp.py\", line 342, in load_dynamic\n            return _load(spec)\n        ImportError: DLL load failed: A dynamic link library (DLL) initialization routine failed.\n\n\n        Failed to load the native TensorFlow runtime.\n\n    See https:\/\/www.tensorflow.org\/install\/errors\n\n    for some common reasons and solutions.  Include the entire stack trace\n    above this error message when asking for help.\n<\/code><\/pre>\n\n<p>These are the packages I installed:<\/p>\n\n<pre><code>pip install --exists-action i --upgrade sklearn\npip install --exists-action i --upgrade pandas\npip install --exists-action i --upgrade pandas-datareader\npip install --exists-action i --upgrade matplotlib\npip install --exists-action i --upgrade pillow\npip install --exists-action i --upgrade tqdm\npip install --exists-action i --upgrade requests\npip install --exists-action i --upgrade h5py\npip install --exists-action i --upgrade pyyaml\npip install --exists-action i --upgrade tensorflow_hub\npip install --exists-action i --upgrade bayesian-optimization\npip install --exists-action i --upgrade spacy\npip install --exists-action i --upgrade gensim\npip install --exists-action i --upgrade flask\npip install --exists-action i --upgrade boto3\npip install --exists-action i --upgrade gym\npip install --exists-action i --upgrade tensorflow==2.0.0-beta1\npip install --exists-action i --upgrade keras-rl2 --user\n<\/code><\/pre>\n\n<p>As a noob, I have tried to google the error, but every installation that I have come across with errors is not similar to mine.<\/p>\n\n<p>The installation is via Juypter on a 3.6 kernel. All the packages installed correctly, but for some reason, verifying the install produces an error.<\/p>\n\n<p>edit: Sorry, I forgot to include the code!<\/p>\n","answers":[{"AnswerId":"57838523","CreationDate":"2019-09-08T01:18:59.810","ParentId":null,"OwnerUserId":"12015469","Title":null,"Body":"<p>My current installation was a mess. I installed a couple of ide's, then python. The path was messed up, and it was a royal pain to troubleshoot.<\/p>\n\n<p>I started with a fresh os install and I am doing it the right way. <\/p>\n"},{"AnswerId":"57890028","CreationDate":"2019-09-11T13:12:29.613","ParentId":null,"OwnerUserId":"3472735","Title":null,"Body":"<p>Installing tensorflow can be a pain, especially if you need GPU support. Luckily it's gotten easier recently, and it is much easier than that guide makes it seem. <\/p>\n\n<p>This one line should install the correct version of tensorflow-gpu, keras, and the appropriate cuda and cudnn toolkits:<\/p>\n\n<pre><code>conda install -c anaconda keras-gpu\n<\/code><\/pre>\n\n<p>I would just give it a go if you haven't already as it should just take a few minutes<\/p>\n"},{"AnswerId":"58111701","CreationDate":"2019-09-26T07:32:48.373","ParentId":null,"OwnerUserId":"12015469","Title":null,"Body":"<p>Too much of a headache getting tf installed on my laptop.<\/p>\n\n<p>Solution:\nAWS EC2 Ubuntu 18.04 (free tier), installed Tensorflow, doing remote development with VS Code locally.<\/p>\n\n<p>I learned a ton setting up tf on Linux.<\/p>\n"}]}
+{"QuestionId":57781188,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-04T03:40:14.590","AcceptedAnswerId":null,"OwnerUserId":12017479.0,"Title":"LARSOptimizer - Layer-wise decomposition","Body":"<p>LARSOptimizer provides the learning rate control used in LARS, but we need layer-wise decomposed parameters of a neural network. how do we do the layer-wise decomposition for a pre-defined architecture?<\/p>\n\n<p>LARS: <a href=\"https:\/\/arxiv.org\/abs\/1708.03888\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1708.03888<\/a> | <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/release_1.14.0\/tensorflow\/contrib\/opt\/python\/training\/lars_optimizer.py\" rel=\"nofollow noreferrer\">Release 1.14 code<\/a><\/p>\n\n<p>LAMB: <a href=\"https:\/\/arxiv.org\/abs\/1904.00962\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1904.00962<\/a> | <a href=\"https:\/\/github.com\/ymcui\/LAMB_Optimizer_TF\" rel=\"nofollow noreferrer\">Unofficial code<\/a><\/p>\n\n<p>NovoGrad: <a href=\"https:\/\/arxiv.org\/abs\/1905.11286\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1905.11286<\/a> | <a href=\"https:\/\/github.com\/NVIDIA\/OpenSeq2Seq\/blob\/master\/open_seq2seq\/optimizers\/novograd.py\" rel=\"nofollow noreferrer\">OpenSeq2Seq code<\/a><\/p>\n\n<p>e.g. <a href=\"https:\/\/docs.nvidia.com\/deeplearning\/frameworks\/caffe-user-guide\/index.html#larc\" rel=\"nofollow noreferrer\">Caffe's LARC<\/a><\/p>\n\n<p>A possible TensorFlow implementation for conv2d:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport tensorflow as tf\nfrom tensorflow.python.framework import ops\n\ndef lars_conv2d():\n    input = tf.placeholder(tf.float32, shape=[16,224,224,3])\n    with tf.name_scope('conv1_1') as scope:\n        kernel_size = 7*7*3*32\n        bias_size = 32\n        total_size = kernel_size + bias_size\n        params = tf.Variable(tf.constant(0.0, shape=[total_size], dtype=tf.float32),\n                             trainable=True, name='weights_and_biases')\n        kernel = tf.reshape(\n                    tf.slice(params, [0], [kernel_size], name='weights'),\n                    [7, 7, 3, 32])\n        conv = tf.nn.conv2d(input, kernel, [1, 1, 1, 1], padding='SAME')\n        biases = tf.slice(params, [kernel_size], [bias_size], name='biases')\n        bias = tf.nn.bias_add(conv, biases)\n        conv1 = tf.nn.relu(bias, name=scope)\n\n    init_op = tf.variables_initializer(var_list=[params])\n    with tf.Session() as sess:\n        sess.run(init_op)\n        print(np.sum(sess.run(conv1, {input: np.zeros([16,224,224,3])})))\n\nlars_conv2d()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/32125\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/32125<\/a><\/p>\n","answers":[]}
+{"QuestionId":57781441,"AnswerCount":1,"Tags":"<python><keras><deep-learning><conv-neural-network><recurrent-neural-network>","CreationDate":"2019-09-04T04:22:46.367","AcceptedAnswerId":null,"OwnerUserId":7876828.0,"Title":"CNN model overfitting on multi-class classification","Body":"<p>I am trying to use GloVe embeddings to train a cnn model based on <a href=\"https:\/\/github.com\/kk7nc\/Text_Classification#recurrent-convolutional-neural-networks-rcnn\" rel=\"nofollow noreferrer\">this article<\/a> (also a rnn, which has <a href=\"https:\/\/stackoverflow.com\/q\/57748144\/7876828\">this issue<\/a>). The dataset is a labeled data: text (tweets) with labels (hate, offensive or neither). <\/p>\n\n<p>The problem is that model performs well on train set but poorly on validation set.<\/p>\n\n<p>here is the model:<\/p>\n\n<pre><code>kernel_size = 2\nfilters = 256\npool_size = 2\ngru_node = 64\nmodel = Sequential()\nmodel.add(Embedding(len(word_index) + 1,\n                            EMBEDDING_DIM,\n                            weights=[embedding_matrix],\n                            input_length=MAX_SEQUENCE_LENGTH,\n                            trainable=True))\nmodel.add(Dropout(0.25))\nmodel.add(Conv1D(filters, kernel_size, activation='relu'))\nmodel.add(MaxPooling1D(pool_size=pool_size))\nmodel.add(Conv1D(filters, kernel_size, activation='softmax'))\nmodel.add(MaxPooling1D(pool_size=pool_size))\nmodel.add(LSTM(gru_node, return_sequences=True, recurrent_dropout=0.2))\nmodel.add(LSTM(gru_node, return_sequences=True, recurrent_dropout=0.2))\nmodel.add(LSTM(gru_node, return_sequences=True, recurrent_dropout=0.2))\nmodel.add(LSTM(gru_node, recurrent_dropout=0.2))\nmodel.add(Dense(1024,activation='relu'))\nmodel.add(Dense(nclasses))\nmodel.add(Activation('softmax'))\n\nmodel.compile(loss='sparse_categorical_crossentropy',\n              optimizer='adam',\n              metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>fitting the model:<\/p>\n\n<pre><code>X = df.tweet\ny = df['classifi']    # classes 0,1,2\n\nfrom sklearn.model_selection import train_test_split\nX_train, X_test, y_train, y_test = train_test_split(X, y, test_size = 0.2, shuffle=False)\nX_train_Glove,X_test_Glove, word_index,embeddings_index = loadData_Tokenizer(X_train,X_test)\n\nmodel_RCNN = Build_Model_RCNN_Text(word_index,embeddings_index, 20)\n\nmodel_RCNN.fit(X_train_Glove, y_train,validation_data=(X_test_Glove, y_test),\n                              epochs=15,batch_size=128,verbose=2)\n\npredicted = model_RCNN.predict(X_test_Glove)\n\npredicted = np.argmax(predicted, axis=1)\nprint(metrics.classification_report(y_test, predicted))\n<\/code><\/pre>\n\n<p>this is what the distribution looks like (0:hate, 1:offensive, 2:neither)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/o5EJC.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/o5EJC.png\" alt=\"distribn\"><\/a><\/p>\n\n<p>model summary<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/LztXj.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/LztXj.png\" alt=\"model\"><\/a><\/p>\n\n<p>Results:<\/p>\n\n<ol>\n<li>classification report\n<a href=\"https:\/\/i.stack.imgur.com\/0FuoR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0FuoR.png\" alt=\"enter image description here\"><\/a><\/li>\n<\/ol>\n\n<p>is this the correct approach or am I missing something here<\/p>\n","answers":[{"AnswerId":"57782032","CreationDate":"2019-09-04T05:31:58.637","ParentId":null,"OwnerUserId":"12016879","Title":null,"Body":"<p>Generally speaking there are two sides that you can tackle overfitting:<\/p>\n\n<ol>\n<li><p>Improving the data<\/p>\n\n<ul>\n<li>More unique data<\/li>\n<li>oversampling (to balance data)<\/li>\n<\/ul><\/li>\n<li><p>Limiting the network structure<\/p>\n\n<ul>\n<li>Dropout (You've implemented this)<\/li>\n<li>Less parameters (You might want to benchmark against a much smaller network)<\/li>\n<li>regularization (ex. L1 and L2)<\/li>\n<\/ul><\/li>\n<\/ol>\n\n<p>I'd suggest trying with significantly fewer parameters (because this is quick) and oversampling (because your data seems lopsided).<\/p>\n\n<p>Also, You can also try hyperparameter fitting. Making a large number of networks with different parameters than picking the best one. <\/p>\n\n<p>Note: if you do hyper parameter fitting make sure to have an extra validation set because you can easily overfit your test set this way.<\/p>\n\n<p>Side note: Sometimes when troubleshooting NN it is helpful to set the optimizer to a basic stochastic gradient descent. It slows the training down a bunch but makes the progression much clearer.<\/p>\n\n<p>Good luck!<\/p>\n"}]}
+{"QuestionId":57781533,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-04T04:33:08.157","AcceptedAnswerId":null,"OwnerUserId":12017171.0,"Title":"Chasing down Tensorflow 'InvalidArgumentError: Max scatter index must be < array size'","Body":"<p>I'm working on a CNN involving custom layers in Keras.  My goal is to have layers that allow a user to generate unique kernels for each image in a batch, and apply each kernel to the corresponding image.  I have the convolutional layer working using some code that I mostly stripped and refit from the Keras source code.  <\/p>\n\n<p>The problem comes when I try to use the <code>TimeDistributed<\/code> layer to apply this model to the convolutional layer.  The <code>TimeDistributed<\/code> layer only takes 1 tensor inputs, so I wrote my own <code>AuxTimeDistributed<\/code> (code below) that allows one to time distribute multiple tensors.  (I'm skipping anything like masking for now).<\/p>\n\n<p>The problem comes when I try to wrap the convolutional layer twice (which I need for my application). Final output is this error:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: Max scatter index must be &lt; array size (779 vs. 100)\n     [[{{node aux_time_distributed_3\/TensorArrayUnstack_1\/TensorArrayScatter\/TensorArrayScatterV3}}]]\n<\/code><\/pre>\n\n<p><strong>Here's my current setup:<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.layers import TimeDistributed, Input, Layer\nfrom keras.models import Model\nimport keras.backend as K\n\nimport numpy as np\n\nimport tensorflow as tf\nfrom tensorflow.python.keras.engine.input_spec import InputSpec\nfrom tensorflow.python.keras.utils import conv_utils\nfrom tensorflow.python.keras.utils.generic_utils import has_arg\nfrom tensorflow.python.keras.utils.generic_utils import is_all_none\nfrom tensorflow.python.keras.utils.generic_utils import object_list_uid\nfrom tensorflow.python.keras.utils.generic_utils import to_list\n\nfrom tensorflow.python.framework import constant_op\nfrom tensorflow.python.framework import ops\nfrom tensorflow.python.keras.utils.generic_utils import is_all_none\nfrom tensorflow.python.keras.utils.generic_utils import to_list\nfrom tensorflow.python.ops import array_ops\nfrom tensorflow.python.ops import control_flow_ops\nfrom tensorflow.python.ops import tensor_array_ops\nfrom tensorflow.python.util import nest\n\ndef batch_apply(step_function,\n        inputs,\n        aux_inputs,\n        batch_size=None,\n        constants=None,\n        input_length=None):\n    flatted_inputs = nest.flatten(inputs)\n    flatted_aux = nest.flatten(aux_inputs)\n    time_steps = flatted_inputs[0].shape[0]\n    time_steps_t = flatted_inputs[0].shape[1]\n    batch_steps_t = array_ops.shape(flatted_inputs[0])[0]\n\n    if constants is None:\n        constants = []\n\n    if batch_size is not None:\n        successive_outputs = []\n\n        inp_shape = inputs.shape.as_list()\n        aux_shape = aux_inputs.shape.as_list()\n        inp_shape[0] = batch_size\n        aux_shape[0] = batch_size\n        list_inputs = array_ops.unstack(tf.reshape(inputs, inp_shape))\n        list_aux_inputs = array_ops.unstack(tf.reshape(aux_inputs, aux_shape))\n\n        for i in range(batch_size):\n            inp = list_inputs[i]\n            aux = list_aux_inputs[i]\n            output = step_function(inp, aux, tuple(constants))\n            successive_outputs.append(output)\n        outputs = array_ops.stack(successive_outputs)\n\n          # static shape inference\n        def set_shape(output_):\n            if isinstance(output_, ops.Tensor):\n                shape = output_.shape.as_list()\n                shape[0] = batch_size\n                output_.set_shape(shape)\n            return output_\n\n        outputs = nest.map_structure(set_shape, outputs)\n    else:\n        def prep_input(flatted_inputs):\n            input_ta = tuple(\n                tensor_array_ops.TensorArray(\n                    dtype=inp.dtype, size=time_steps_t,\n                    tensor_array_name='input_ts_%s' % i)\n                for i, inp in enumerate(flatted_inputs))\n            input_ta = tuple(ta.unstack(input_)\n                for ta, input_ in zip(input_ta, flatted_inputs))\n            return input_ta\n        input_ta = prep_input(flatted_inputs)\n        input_au = prep_input(flatted_aux)\n\n        input_time_zero = nest.pack_sequence_as(inputs,\n            [inp[0] for inp in flatted_inputs])\n        aux_time_zero = nest.pack_sequence_as(aux_inputs,\n            [inp[0] for inp in flatted_aux])\n        output_time_zero = step_function(input_time_zero, aux_time_zero,\n            tuple(constants))\n\n        output_ta = tuple(\n            tensor_array_ops.TensorArray(dtype=out.dtype, size=batch_steps_t,\n                tensor_array_name='output_ta_%s' % i)\n            for i, out in enumerate(nest.flatten(output_time_zero)))\n\n\n        while_loop_kwargs = {'cond': lambda time, *_: time &lt; batch_steps_t,\n            'maximum_iterations': input_length, 'parallel_iterations': 32,\n            'swap_memory': True}\n\n        time = constant_op.constant(0, dtype='int32', name='time')\n\n        def _step(time, output_ta_t):\n            current_input = tuple(ta.read(time) for ta in input_ta)\n            current_input = nest.pack_sequence_as(inputs, current_input)\n            current_aux = tuple(au.read(time) for au in input_au)\n            current_aux = nest.pack_sequence_as(aux_inputs, current_aux)\n\n            output = step_function(current_input, current_aux, constants)\n            flat_output = nest.flatten(output)\n            output_ta_t = tuple(ta.write(time, out)\n                for ta, out in zip(output_ta_t, flat_output))\n            return (time + 1, output_ta_t)\n\n        final_outputs = control_flow_ops.while_loop(\n            body=_step, loop_vars=(time, output_ta),\n            **while_loop_kwargs\n        )\n\n        outputs = tuple(o.stack() for o in final_outputs[1])\n        outputs = nest.pack_sequence_as(output_time_zero, outputs)\n\n    return outputs\n\ndef unpack_singleton(x):\n    if len(x) == 1:\n        return x[0]\n    return x\n\ndef collect_previous_mask(input_tensors):\n    input_tensors = to_list(input_tensors)\n    masks = []\n    for x in input_tensors:\n        if hasattr(x, '_keras_history'):\n            inbound_layer, node_index, tensor_index = x._keras_history\n            node = inbound_layer._inbound_nodes[node_index]\n            mask = node.output_masks[tensor_index]\n            masks.append(mask)\n        else:\n            masks.append(None)\n    return unpack_singleton(masks)\n\ndef collect_input_shape(input_tensors):\n    input_tensors = to_list(input_tensors)\n    shapes = []\n    for x in input_tensors:\n        try:\n            shapes.append(K.int_shape(x))\n        except TypeError:\n            shapes.append(None)\n    return unpack_singleton(shapes)\n<\/code><\/pre>\n\n<p><strong>AuxTimeDistributed Layer:<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class AuxTimeDistributed(TimeDistributed):\n    def build(self, input_shape):\n        self.input_spec = [InputSpec(shape=input_shape[i])\n            for i in range(len(input_shape))]\n        child_input_shape = [(input_shape[i][0],) + tuple(input_shape[i][2:])\n            for i in range(len(input_shape))]\n        if not self.layer.built:\n            self.layer.build(child_input_shape)\n            self.layer.built = True\n        super(TimeDistributed, self).build()\n\n    def compute_output_shape(self, input_shapes):\n        child_input_shapes = [(input_shapes[i][0],) + tuple(input_shapes[i][2:])\n            for i in range(len(input_shapes))]\n        child_output_shape = self.layer.compute_output_shape(child_input_shapes)\n        timesteps = input_shapes[0][1]\n\n        output_shape = (child_output_shape[0], timesteps) + child_output_shape[1:]\n        return output_shape\n\n    def compute_mask(self, inputs, previous_mask):\n        return None\n\n    def call(self, inputs, training=None):\n        kwargs = {}\n        if has_arg(self.layer.call, 'training'):\n            kwargs['training'] = training\n        uses_learning_phase = False\n\n        def for_inputs(func):\n            return [func(i) for i in range(len(inputs))]\n\n        input_shapes = for_inputs(lambda i: K.int_shape(inputs[i]))\n\n        if input_shapes[0][0]:\n            # If given a batch size, use the tf rnn function\n            def step(x, _):\n                global uses_learning_phase\n                output = self.layer.call(x, **kwargs)\n                if hasattr(output, '_uses_learning_phase'):\n                    uses_learning_phase = (output._uses_learning_phase or\n                                           uses_learning_phase)\n                return output, []\n\n            _, outputs, _ = K.rnn(step, [inputs, *aux],\n                                  initial_states=[],\n                                  input_length=input_shape[1],\n                                  unroll=False)\n            y = outputs\n        else:\n            # No batch size specified, therefore the layer will be able\n            # to process batches of any size.\n            # We can go with reshape-based implementation for performance.\n            input_lengths = for_inputs(lambda i: input_shapes[i][1])\n\n            for i in range(len(input_lengths)):\n                if not input_lengths[i]:\n                    input_lengths[i] = K.shape(inputs[i])[1]\n\n            def get_shapes(i):\n                return self._get_shape_tuple((-1,), inputs[i], 2)\n            inner_shapes = for_inputs(get_shapes)\n\n            # Shape: (num_samples * timesteps, ...). And track the\n            # transformation in self._input_map.\n            input_uids = for_inputs(lambda i: object_list_uid(inputs[i]))\n            inputs = for_inputs(lambda i: K.reshape(inputs[i], inner_shapes[i]))\n\n            def assign_uid(i):\n                self._input_map[input_uids[i]] = inputs[i]\n            for_inputs(assign_uid)\n\n            # (num_samples * timesteps, ...)\n            y = self.layer.call(inputs, **kwargs)\n\n            if hasattr(y, '_uses_learning_phase'):\n                uses_learning_phase = y._uses_learning_phase\n            # Shape: (num_samples, timesteps, ...)\n            output_shape = self.compute_output_shape(input_shapes)\n            output_shape = self._get_shape_tuple(\n                (-1, input_lengths[0]), y, 1, output_shape[2:])\n            y = K.reshape(y, output_shape)\n\n        # Apply activity regularizer if any:\n        if (hasattr(self.layer, 'activity_regularizer') and\n           self.layer.activity_regularizer is not None):\n            regularization_loss = self.layer.activity_regularizer(y)\n            self.add_loss(regularization_loss, inputs)\n\n        if uses_learning_phase:\n            y._uses_learning_phase = True\n        return y\n\n<\/code><\/pre>\n\n<p><strong>KernelConv Layer:<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class KernelConv(Layer):\n    def __init__(self,\n                rank=2,\n                strides=1,\n                padding='valid',\n                data_format=None,\n                dilation_rate=1,\n                **kwargs):\n        super(KernelConv, self).__init__(**kwargs)\n        self.rank = rank\n        self.strides = conv_utils.normalize_tuple(strides, rank, 'strides')\n        self.padding = conv_utils.normalize_padding(padding)\n        self.data_format = K.normalize_data_format(data_format)\n        self.dilation_rate = conv_utils.normalize_tuple(dilation_rate, rank,\n            'dilation_rate')\n        self.input_spec = [InputSpec(ndim=self.rank+2),\n            InputSpec(ndim=self.rank+3)]\n\n    def build(self, input_shapes):\n        self.built=True\n\n    def call(self, inputs):\n        images, kernels = inputs\n\n        def conv_func(image_t, kernel_t, constants):\n            # Reshape image data to batch size 1\n            shape = (1,) + tuple(image_t.shape.as_list())\n            image_t = tf.reshape(image_t, shape)\n\n            # Use the backend conv2d function\n            return K.conv2d(\n                image_t,\n                kernel_t,\n                strides=self.strides,\n                padding=self.padding,\n                data_format=self.data_format,\n                dilation_rate=self.dilation_rate\n            )[0]\n\n        # Apply the convolution function to the batch of (image, aux) pairs\n        return batch_apply(conv_func, images, kernels)\n\n    def compute_output_shape(self, input_shapes):\n        image_shape, kernel_shape = input_shapes\n\n        if self.data_format == 'channels_last':\n            space = image_shape[1:-1]\n            filters = kernel_shape[-1]\n            k=1\n        elif self.data_format == 'channels_first':\n            space = image_shape[2:]\n            filters = kernel_shape[1]\n            k=2\n        new_space = []\n        for i in range(len(space)):\n            new_dim = conv_utils.conv_output_length(\n                space[i],\n                kernel_shape[i+k],\n                padding=self.padding,\n                stride=self.strides[i],\n                dilation=self.dilation_rate[i])\n            new_space.append(new_dim)\n        if self.data_format == 'channels_last':\n            output_shape = (image_shape[0],) + tuple(new_space) + (filters,)\n        elif self.data_format == 'channels_first':\n            output_shape = (image_shape[0], filters) + tuple(new_space)\n\n        return output_shape\n<\/code><\/pre>\n\n<p>The layer works as a simple convolutional layer:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>images = Input(shape=(100, 100, 5))\nkernels = Input(shape=(3, 3, 5, 32))\noutput = KernelConv(rank=2)([images, kernels])\nmodel = Model(inputs=[images, kernels], outputs=[output])\n\nimg = np.random.rand(6, 100, 100, 5)\nkrn = np.random.rand(6, 3, 3, 5, 32)\nout = np.random.rand(6, 98, 98, 32)\n\nmodel.compile('adam', loss='mape')\nprint(\"can compile\")\n\nmodel.predict([img, krn])\nprint(\"can predict\")\n\nmodel.fit([img, krn], out)\nprint(\"can fit\")\n<\/code><\/pre>\n\n<p>I can wrap it once like so:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>images = Input(shape=(10, 100, 100, 5))\nkernels = Input(shape=(10, 3, 3, 5, 32))\noutput = AuxTimeDistributed(KernelConv(rank=2))([images, kernels])\nmodel = Model(inputs=[images, kernels], outputs=[output])\n\nimg = np.random.rand(6, 10, 100, 100, 5)\nkrn = np.random.rand(6, 10, 3, 3, 5, 32)\nout = np.random.rand(6, 10, 98, 98, 32)\n\nmodel.compile('adam', loss='mape')\nprint(\"can compile\")\n\nmodel.predict([img, krn])\nprint(\"can predict\")\n\nmodel.fit([img, krn], out)\nprint(\"can fit\")\n<\/code><\/pre>\n\n<p><strong>BUT, when I wrap it twice I get this:<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>images = Input(shape=(13, 10, 100, 100, 5))\nkernels = Input(shape=(13, 10, 3, 3, 5, 32))\noutput = AuxTimeDistributed(AuxTimeDistributed(\n    KernelConv(rank=2)))([images, kernels])\nmodel = Model(inputs=[images, kernels], outputs=[output])\n\nimg = np.random.rand(6, 13, 10, 100, 100, 5)\nkrn = np.random.rand(6, 13, 10, 3, 3, 5, 32)\nout = np.random.rand(6, 13, 10, 98, 98, 32)\n\nmodel.compile('adam', loss='mape')\nprint(\"can compile\")\n\nmodel.predict([img, krn])\nprint(\"can predict\")\n\nmodel.fit([img, krn], out)\nprint(\"can fit\")\n<\/code><\/pre>\n\n<pre><code>can compile\nTraceback (most recent call last):\n  File \"minimal_time_dist.py\", line 360, in &lt;module&gt;\n    model.predict([img, krn])\n  File \"\/home\/parallels\/anaconda3\/lib\/python3.7\/site-packages\/keras\/engine\/training.py\", line 1169, in predict\n    steps=steps)\n  File \"\/home\/parallels\/anaconda3\/lib\/python3.7\/site-packages\/keras\/engine\/training_arrays.py\", line 294, in predict_loop\n    batch_outs = f(ins_batch)\n  File \"\/home\/parallels\/anaconda3\/lib\/python3.7\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 2715, in __call__\n    return self._call(inputs)\n  File \"\/home\/parallels\/anaconda3\/lib\/python3.7\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 2675, in _call\n    fetched = self._callable_fn(*array_vals)\n  File \"\/home\/parallels\/anaconda3\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1458, in __call__\n    run_metadata_ptr)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: Max scatter index must be &lt; array size (779 vs. 100)\n     [[{{node aux_time_distributed_3\/TensorArrayUnstack_1\/TensorArrayScatter\/TensorArrayScatterV3}}]]\n<\/code><\/pre>\n\n<p>The only source I can find on this error is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/tensor_array_ops.cc#L1115\" rel=\"nofollow noreferrer\">here, on line 1115<\/a>.<\/p>\n\n<p>Please help explain why this error arises, and how I could address it.  Thanks!<\/p>\n\n<p>Edit: I'm running in a linux (Ubuntu 16.04) VM on MacOs 10.14.6 via Parallels.<\/p>\n","answers":[]}
+{"QuestionId":57781610,"AnswerCount":1,"Tags":"<python-3.x><keras>","CreationDate":"2019-09-04T04:41:29.287","AcceptedAnswerId":57783607.0,"OwnerUserId":12017671.0,"Title":"How to use 3-block VGG to classify three classes?","Body":"<p>I was using 3-VGG block to classify images as two classes for example dog and cat and I was able to do it. But when I tried to use the same for classifying 3 classes for example- dog, cat and pandas, I am getting valuerror. I am suspecting I need to update Dense layer code but couldn't able to find exact issue. Please help me to know what I am missing here.<\/p>\n\n<p>I am using Python 3.7 and Keras for this classification.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def define_model():\n    model = Sequential()\n    model.add(Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same',\n                     input_shape=(200, 200, 3)))\n    model.add(MaxPooling2D((2, 2)))\n    model.add(Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))\n    model.add(MaxPooling2D((2, 2)))\n    model.add(Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))\n    model.add(MaxPooling2D((2, 2)))\n    model.add(Flatten())\n    model.add(Dense(128, activation='relu', kernel_initializer='he_uniform'))\n    model.add(Dense(1, activation='sigmoid'))\n    # compile model\n    opt = SGD(lr=0.001, momentum=0.9)\n    model.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['accuracy'])\n    return model\n\n\n# run the test harness for evaluating a model\ndef run_test_harness():\n    # define model\n    model = define_model()\n    # create data generator\n    datagen = ImageDataGenerator(rescale=1.0 \/ 255.0)\n    # prepare iterators\n    train_it = datagen.flow_from_directory('D:\/Proect\/images\/',\n                                           class_mode='categorical', batch_size=64, target_size=(200, 200))\n    label_map = train_it.class_indices\n    print(\"label_map: \", label_map)\n    # fit model\n    model.fit_generator(train_it, steps_per_epoch=len(train_it), epochs=20, verbose=0)\n    # save model\n    model.save('my_model.h5')\n\n<\/code><\/pre>\n\n<p>I am getting following error-<\/p>\n\n<blockquote>\n  <p>\"at line 41, in run_test_harness\n      model.fit_generator(train_it, steps_per_epoch=len(train_it), epochs=20, verbose=0) ValueError: Error when checking target: expected\n  dense_2 to have shape (1,) but got array with shape (3,)\"<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"57783607","CreationDate":"2019-09-04T07:37:56.370","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>The problem is in this line:<\/p>\n\n<pre><code>model.add(Dense(1, activation='sigmoid'))\n<\/code><\/pre>\n\n<p>This is configured to do binary classification, for 3 classes you need:<\/p>\n\n<pre><code>model.add(Dense(3, activation='softmax'))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57781740,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><confusion-matrix>","CreationDate":"2019-09-04T04:58:33.187","AcceptedAnswerId":null,"OwnerUserId":12017673.0,"Title":"How to create confusion matrix in Python","Body":"<p>I am trying to create a model that use Batch Normalization for a limited class and number of MNIST datasets. Here I am trying with 2 classes with 1 class has limited numbers of datasets. For example in my code, from MNIST class 4 and 7 are taken, and number of dataset of class 4 is limited to 39. However, this applies to training data only. Therefore total number of datasets of class 4 and 7 combined for training data is 6304, and for test data will be 2010.<\/p>\n\n<p>I am trying to create a confusion matrix of test data only, but I do not know how to do because I am quite new in Python.<\/p>\n\n<p>Could anyone show me how to create a confusion matrix that consists whole number of test data? <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># coding: utf-8\n\n\n\nimport gzip\nimport numpy as np\nfrom numpy.linalg import inv\n\ndef _load_testimage_filter(a,b,file):\n    x=np.loadtxt(file,delimiter=',')\n    idx=np.hstack((np.where(x[:,0]==a),np.where(x[:,0]==b)))\n    x=x[idx][0]\n    x0,x1=np.hsplit(x,[1])\n    return x1\n\n\ndef _load_image_filter(a,b,c,file):\n    x=np.loadtxt(file,delimiter=',')\n    idx=np.hstack((np.where(x[:,0]==a),np.where(x[:,0]==None)))\n    x=x[idx][0]\n    x=x[:c,:]\n    x0,x1=np.hsplit(x,[1])\n\n    y=np.loadtxt(file,delimiter=',')\n    idx=np.hstack((np.where(y[:,0]==b),np.where(y[:,0]==None)))\n    y=y[idx][0]\n    y0,y1=np.hsplit(y,[1])\n\n    z=np.vstack((x1,y1))\n    return z\n\ndef _load_image_filter2(a,b,c,file):\n    x=np.loadtxt(file,delimiter=',')\n    idx=np.hstack((np.where(x[:,0]==a),np.where(x[:,0]==None)))\n    x=x[idx][0]\n    x=x[:c,:]\n    x0,x1=np.hsplit(x,[1])\n\n    y=np.loadtxt(file,delimiter=',')\n    idx=np.hstack((np.where(y[:,0]==b),np.where(y[:,0]==None)))\n    y=y[idx][0]\n    y0,y1=np.hsplit(y,[1])\n\n    z=np.vstack((x0,y0))\n    return z,x0,y0\n\ndef _load_testlabel_filter(a,b,file):\n    x=np.loadtxt(file,delimiter=',')\n    idx=np.hstack((np.where(x[:,0]==a),np.where(x[:,0]==b)))\n    x=x[idx][0]\n    x0,x1=np.hsplit(x,[1])\n    label=x0[:,0]\n    label_1ofK=np.identity(10)[label.astype(np.int)]\n    return label_1ofK\n\ndef _load_label_filter(a,b,c,file):\n    x=np.loadtxt(file,delimiter=',')\n    idx=np.hstack((np.where(x[:,0]==a),np.where(x[:,0]==None)))\n    x=x[idx][0]\n    x=x[:c,:]\n    x0,x1=np.hsplit(x,[1])\n    y=np.loadtxt(file,delimiter=',')\n    idx=np.hstack((np.where(y[:,0]==b),np.where(y[:,0]==None)))\n    y=y[idx][0]\n    y0,y1=np.hsplit(y,[1])      \n    z=np.vstack((x0,y0))\n\n    label=z[:,0]\n    label_1ofK=np.identity(10)[label.astype(np.int)]\n    return label_1ofK\n\ndef load_mnist_filter(a,b,c):\n    return _load_image_filter(a,b,c,\".\/MNIST\/mnist_train.csv\"), _load_label_filter(a,b,c,\".\/MNIST\/mnist_train.csv\"), _load_testimage_filter(a,b,\".\/MNIST\/mnist_test.csv\"), _load_testlabel_filter(a,b,\".\/MNIST\/mnist_test.csv\")\n\n\n\n\n\n\n#BN\nimport numpy as np\nimport time\n\nfrom sklearn.metrics import confusion_matrix,accuracy_score\n\n\ndef learn(name,x_train,t_train,x_test,t_test,md=[200],weight_init_func=he_normal,weight_init_params={},bias_init_func=zeros_b,bias_init_params={},eta=0.01,batch_size=100,epoch=100,data_norm_func=min_max,data_norm_params={},middle_func=relu,middle_params={},output_func=softmax,output_params={},error_func2=cross_entropy_error2,optimizer_func=Adam,optimizer_params={\"beta1\":0.9,\"beta2\":0.999},regularization_params={},shuffle_flag=True):\n    print(name)\n    d=[x_train.shape[x_train.ndim-1]]+md+[t_train.shape[t_train.ndim-1]]\n    layer=len(d)-1\n    print(layer)\n    # \u91cd\u307f\u3001\u30d0\u30a4\u30a2\u30b9\u306e\u521d\u671f\u5316 Weight,Bias Initialization\n    W = {}\n    b = {}\n    for i in range(layer):\n        W[i+1]=weight_init_func(d[i], d[i+1], weight_init_params)\n    for i in range(layer):\n        b[i+1]=bias_init_func(d[i+1], bias_init_params)\n    # \u5165\u529b\u30c7\u30fc\u30bf\u306e\u6b63\u898f\u5316 Input data Normalization\n    stats={}\n    nx_train,train_stats=data_norm_func(x_train, stats, data_norm_params)\n    nx_test,test_stats=data_norm_func(x_test, train_stats, data_norm_params)\n    # \u6b63\u89e3\u7387\u3001\u8aa4\u5dee\u521d\u671f\u5316 Accuracy Rate and Train error initialization\n    train_rate=np.zeros(epoch+1)\n    test_rate=np.zeros(epoch+1)\n    train_err=np.zeros(epoch+1)\n    test_err=np.zeros(epoch+1)\n    # \u52fe\u914d\u95a2\u6570 Gradient Function\n    middle_back_func = eval(middle_func.__name__ + \"_back\")\n    output_error_back_func = eval(output_func.__name__ + \"_\" + error_func2.__name__ + \"_back\")\n    optimizer_statsW={}\n    optimizer_statsb={}\n    for i in range(layer):\n        optimizer_statsW[i+1]={}\n        optimizer_statsb[i+1]={}        \n    # \u6b63\u898f\u5316\u7528\u3001\u03b3\u3001\u03b2\u3001\u5b66\u7fd2\u7387\u306e\u521d\u671f\u5316\u3000BatchNormalization parameters\n    batch_norm_params = {}\n    batch_norm_params[\"batch_norm_gamma\"] = {}\n    batch_norm_params[\"batch_norm_beta\"]  = {}\n    for i in range(1, layer):\n        if regularization_params.get(\"batch_norm_node\"): \n            batch_norm_params[\"batch_norm_gamma\"][i] = np.ones((1, d[i]))\n            batch_norm_params[\"batch_norm_beta\"][i]  = np.zeros((1, d[i]))\n        else:\n            batch_norm_params[\"batch_norm_gamma\"][i] = np.ones(1)\n            batch_norm_params[\"batch_norm_beta\"][i]  = np.zeros(1)\n    batch_norm_eta = eta\n    if \"batch_norm_eta\" in regularization_params:\n        batch_norm_eta = regularization_params[\"batch_norm_eta\"]\n    optimizer_stats_gamma={}\n    optimizer_stats_beta={}\n    for i in range(1,layer):\n        optimizer_stats_gamma[i]={}\n        optimizer_stats_beta[i]={}\n\n    # \u5b9f\u884c(\u5b66\u7fd2\u30c7\u30fc\u30bf) Training(TrainData)\n    u_train, un_train, z_train, y_train=propagation(layer, nx_train, W, b, middle_func, middle_params, output_func, output_params, regularization_params, batch_norm_params)\n    train_rate[0]=accuracy_rate(y_train, t_train)\n    train_err[0]=calc_error2(y_train, t_train, W, error_func2, regularization_params)\n    # \u5b9f\u884c(\u30c6\u30b9\u30c8\u30c7\u30fc\u30bf) Training(TestData)\n    u_test, un_test, z_test, y_test=propagation(layer, nx_test, W, b, middle_func, middle_params, output_func, output_params, regularization_params, batch_norm_params)\n    test_rate[0]=accuracy_rate(y_test, t_test)\n    test_err[0]=calc_error2(y_test, t_test, W, error_func2, regularization_params)    \n    print(\"{0:3d} train_rate={1:6.2f}% test_rate={2:6.2f}% train_err={3:8.5f} test_err={4:8.5f}\".format((0), train_rate[0]*100, test_rate[0]*100, train_err[0], test_err[0]))\n\n    start_time = time.time()\n    for i in range(epoch):\n        nx = nx_train\n        t  = t_train\n        if shuffle_flag:\n            nx_t = np.concatenate([nx_train, t_train], axis=1)\n            np.random.shuffle(nx_t)\n            nx, t = np.split(nx_t, [nx_train.shape[1]], axis=1)       \n        for j in range(0, nx.shape[0], batch_size):\n            # \u5b9f\u884c (Training)\n            u, un, z, y = propagation(layer, nx[j:j+batch_size], W, b, middle_func, middle_params, output_func, output_params, regularization_params, batch_norm_params)\n            # \u52fe\u914d\u3092\u8a08\u7b97 Gradient Calculation\n            du, dz, dW, db, batch_norm_dparams = back_propagation(layer, u, un, z, y, t[j:j+batch_size], W, b, middle_back_func, middle_params, output_error_back_func, regularization_params, batch_norm_params)\n            # \u91cd\u307f\u3001\u30d0\u30a4\u30a2\u30b9\u306e\u8abf\u6574 Parameter adjustment\n            for k in range(1, layer+1):\n                W[k],optimizer_statsW[k] = optimizer_func(W[k],dW[k],eta,optimizer_params,optimizer_statsW[k])\n                b[k],optimizer_statsb[k] = optimizer_func(b[k],db[k],eta,optimizer_params,optimizer_statsb[k])\n            # \u30d0\u30c3\u30c1\u6b63\u898f\u5316\u306e\u8abf\u6574 Batch Normalization parameter adjustment\n            if regularization_params.get(\"batch_norm\"):\n                for k in range(1, layer):\n                    batch_norm_params[\"batch_norm_gamma\"][k],optimizer_stats_gamma[k]=optimizer_func(batch_norm_params[\"batch_norm_gamma\"][k],batch_norm_dparams[\"batch_norm_dgamma\"][k],batch_norm_eta,optimizer_params,optimizer_stats_gamma[k])\n                    batch_norm_params[\"batch_norm_beta\"][k],optimizer_stats_beta[k]=optimizer_func(batch_norm_params[\"batch_norm_beta\"][k],batch_norm_dparams[\"batch_norm_dbeta\"][k],batch_norm_eta,optimizer_params,optimizer_stats_beta[k])\n\n        # \u91cd\u307f\u3001\u30d0\u30a4\u30a2\u30b9\u8abf\u6574\u5f8c\u306e\u5b9f\u884c(\u5b66\u7fd2\u30c7\u30fc\u30bf) After adjustment training (train)\n        u_train, un_train, z_train, y_train = propagation(layer, nx_train, W, b, middle_func, middle_params, output_func, output_params, regularization_params, batch_norm_params)\n        train_rate[i+1] = accuracy_rate(y_train, t_train)\n        train_err[i+1] = calc_error2(y_train, t_train, W, error_func2, regularization_params)\n        # \u91cd\u307f\u3001\u30d0\u30a4\u30a2\u30b9\u8abf\u6574\u5f8c\u306e\u5b9f\u884c(\u30c6\u30b9\u30c8\u30c7\u30fc\u30bf) After adjustment training (test)\n        u_test, un_test, z_test, y_test = propagation(layer, nx_test, W, b, middle_func, middle_params, output_func, output_params, regularization_params, batch_norm_params)\n\n        #Confusion Matrix\n        test_predictions = np.argmax(y_test[1:10],1)\n        true_classes = np.argmax(t_test[1:10],1)\n        cm=confusion_matrix(true_classes, test_predictions)\n        print(cm)\n\n        test_rate[i+1] = accuracy_rate(y_test, t_test)\n        test_err[i+1] = calc_error2(y_test, t_test, W, error_func2, regularization_params)\n        # Show Accuracy rate, train error \n        print(\"{0:3d} train_rate={1:6.2f}% test_rate={2:6.2f}% train_err={3:8.5f} test_err={4:8.5f}\".format((i+1), train_rate[i+1]*100, test_rate[i+1]*100, train_err[i+1],test_err[i+1]))\n\n    end_time = time.time()\n    total_time = end_time - start_time\n    print(\"\u6240\u8981\u6642\u9593 = \" +str(int(total_time\/60))+\" \u5206 \"+str(int(total_time%60)) + \" \u79d2\")\n    #print(cm)\n\n    return y_train, y_test, W, b, train_rate, train_err, test_rate, test_err, total_time\n\n\n\n\n\n\nx_train,t_train,x_test,t_test=load_mnist_filter(4,7,39)\nprint(x_train.shape,t_train.shape,x_test.shape,t_test.shape)\n\n\n\n\n\n#BN\nregularization_params={\"batch_norm\":True}\n_, _, _, _,train_rate,train_err,test_rate,test_err,total_time=learn(\"batch_norm\",x_train,t_train,x_test,t_test,regularization_params=regularization_params)\n\n\n\n\n\n\n\n<\/code><\/pre>\n\n<p>I expect the output will be [[342,640],[0,1028]], something like this but the output is [[1,8],[0,0]]<\/p>\n","answers":[{"AnswerId":"57782039","CreationDate":"2019-09-04T05:32:55.873","ParentId":null,"OwnerUserId":"8868699","Title":null,"Body":"<p>Try this:<\/p>\n\n<pre><code>from sklearn.metrics import confusion_matrix\ny_true = [2, 0, 2, 2, 0, 1]\ny_pred = [0, 0, 2, 2, 0, 2]\nconfusion_matrix(y_true, y_pred)\n<\/code><\/pre>\n\n<p>You can also use matplotlib. This <a href=\"https:\/\/scikit-learn.org\/stable\/auto_examples\/model_selection\/plot_confusion_matrix.html#sphx-glr-auto-examples-model-selection-plot-confusion-matrix-py\" rel=\"nofollow noreferrer\">link<\/a> might help<\/p>\n"}]}
+{"QuestionId":57781990,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><nlp><pytorch>","CreationDate":"2019-09-04T05:27:52.770","AcceptedAnswerId":null,"OwnerUserId":10791002.0,"Title":"How to create Affix(Prefix + Suffix) embeddings in NLP","Body":"<p>I am working on a named entity recognition task. Traditional method is to concatenate word embeddings and character level embeddings for creating a word representation first. I want to also use affix embeddings to better understand the relation between the tags and the words. <\/p>\n\n<p>For example the words \"Afghanistan\" and \"Kajikistan\" are clear examples of Location. Here the suffix \"istan\" or \"tan\" will be useful to identify future \"location\" tags. So I want to extract the suffixes and prefixes of all the words and create embeddings for them, and then concatenate with the initial word representation. How to achieve this?<\/p>\n","answers":[{"AnswerId":"57783416","CreationDate":"2019-09-04T07:24:27.320","ParentId":null,"OwnerUserId":"11070254","Title":null,"Body":"<p>You may simply do this:<\/p>\n\n<p>1) Search for a surffixes vacabulary from google.   <br\/>\n 2) Write a simple max-backward segmentation scripte to generate all\n    words' surffixs, and add them as another item into your training and\n    testing data, just like words and characters.  <br\/>\n 3) Concate surfix embedding with words and character embedding  <br\/><\/p>\n"}]}
+{"QuestionId":57782061,"AnswerCount":0,"Tags":"<tensorflow><classification><tensorflow-serving><tf.keras><tf-hub>","CreationDate":"2019-09-04T05:35:09.220","AcceptedAnswerId":null,"OwnerUserId":1646099.0,"Title":"tensorflow image Re-training to get classify signaturedef hub with keras","Body":"<p>I need to retrain Inception V3 model with my own set of images and create a classify signature in saved_model<\/p>\n\n<p>I tried <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/images\/hub_with_keras\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/tutorials\/images\/hub_with_keras<\/a> to retrain my own set of images\nBut the saved_model generated only contains <strong>tensorflow\/serving\/predict<\/strong> signaturedef. What I need is the <strong>tensorflow\/serving\/classify<\/strong> signaturedef so that the model will return label and score as <\/p>\n\n<pre><code>{\n  \"result\": [\n    \/\/ List of class label\/score pairs for first Example (in request)\n    [ [&lt;label1&gt;, &lt;score1&gt;], [&lt;label2&gt;, &lt;score2&gt;], ... ],\n\n    \/\/ List of class label\/score pairs for next Example (in request)\n    [ [&lt;label1&gt;, &lt;score1&gt;], [&lt;label2&gt;, &lt;score2&gt;], ... ],\n    ...\n  ]\n}\n<\/code><\/pre>\n\n<p>see <a href=\"https:\/\/www.tensorflow.org\/tfx\/serving\/api_rest\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tfx\/serving\/api_rest<\/a> - tensorflow REST endpoints<\/p>\n\n<p>Basically I am trying the google colab notebook\n <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/images\/hub_with_keras\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/tutorials\/images\/hub_with_keras<\/a><\/p>\n\n<p>A slightly changed version of the notebook is available at \n<a href=\"https:\/\/colab.research.google.com\/drive\/1Vmb6h9Ps9eAP0UIH8pnzpxD_ac_pZRXK\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/1Vmb6h9Ps9eAP0UIH8pnzpxD_ac_pZRXK<\/a><\/p>\n\n<p>which contians a saved_model_cli at the end to query the signaturedefs<\/p>\n","answers":[]}
+{"QuestionId":57782112,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-04T05:41:27.143","AcceptedAnswerId":null,"OwnerUserId":847877.0,"Title":"How tf.image.resize_images fill the blank when preserve_aspect_ratio is True","Body":"<p>According to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/image\/resize_images\" rel=\"nofollow noreferrer\">tensorflow doc<\/a>, preserve_aspect_ratio argument in resize_images will keep the origin ratio, but how the blank will be filled?<\/p>\n","answers":[]}
+{"QuestionId":57782382,"AnswerCount":1,"Tags":"<tensorflow><tensorflow.js><tensorflowjs-converter>","CreationDate":"2019-09-04T06:07:03.647","AcceptedAnswerId":57797525.0,"OwnerUserId":2723463.0,"Title":"Tensorflowjs Operands could not be broadcast together with shapes","Body":"<p>I have a working model in Tensorflow that works fine with Python. It is now a saved model and attempted to convert to TensorflowJS.<\/p>\n\n<p>A converted model appears to not work correctly. The mean value from the TensorMap appears to be 0 and has a shape of [0], despite the input being [1,96,192,3].<\/p>\n\n<p>The conversion was done as follows....<\/p>\n\n<pre><code>tensorflowjs_converter --input_format=tf_saved_model ~\/Projects\/models\/model_0 ~\/Projects\/modelsjs\/model_0\n<\/code><\/pre>\n\n<p>This works fine and it also appears to load ok. However, when it comes to predicting, errors are thrown and any advice would be appreciated.<\/p>\n\n<pre><code>&lt;script&gt;\n    async function handleButtonClick(){\n\n       for(var i=0; i&lt;1;i++)\n       {\n           var t0 = performance.now();\n           console.log(\"Loading - model_\"+i);\n           var inputTensor = tf.tensor4d(input);\n           var model = await tf.loadGraphModel('\/modelsjs\/model_'+i+'\/model.json');\n           var poutput = model.predict(inputTensor);\n       }\n&lt;\/script&gt;\n<\/code><\/pre>\n\n<p>The error appears as follows.<\/p>\n\n<pre><code>graph_model.ts:213 Uncaught (in promise) Error: The model contains control flow or dynamic shape ops, please use executeAsync method\n    at t.execute_ (graph_model.ts:213)\n    at t.predict (graph_model.ts:169)\n    at (index):157\n    at engine.ts:156\n    at t.scopedRun (engine.ts:167)\n    at t.tidy (engine.ts:153)\n    at Object.t.tidy (environment.ts:186)\n    at handleButtonClick ((index):156)\n\n<\/code><\/pre>\n\n<p>As per the above error, the prediction was attempted using executeAsync but produces the error relevant to this question.<\/p>\n\n<pre><code>&lt;script&gt;\n    async function handleButtonClick(){\n\n       for(var i=0; i&lt;1;i++)\n       {\n           var t0 = performance.now();\n           console.log(\"Loading - model_\"+i);\n           var inputTensor = tf.tensor4d(input);\n           var model = await tf.loadGraphModel('\/modelsjs\/model_'+i+'\/model.json');\n           console.log(\"Load model_\" + i + \"took \" + (t1 - t0) + \" milliseconds.\");\n           const res = await model.executeAsync(inputTensor);\n       }\n&lt;\/script&gt;\n\n<\/code><\/pre>\n\n<p>The Error appears as follows. And appears to be related to the $mean value from the Tensormap. This value is [0]<\/p>\n\n<pre><code>broadcast_util.ts:81 Uncaught (in promise) Error: Operands could not be broadcast together with shapes 1,12,24,64 and 0.\n    at un (broadcast_util.ts:81)\n    at new kn (batchnorm_packed_gpu.ts:32)\n    at t.batchNormalization (backend_webgl.ts:869)\n    at Bt.engine.runKernel.$x (batchnorm.ts:344)\n    at engine.ts:206\n    at t.scopedRun (engine.ts:167)\n    at t.runKernel (engine.ts:202)\n    at $a (batchnorm.ts:343)\n    at batchNorm (operation.ts:46)\n    at xy (normalization_executor.ts:31)\n\n<\/code><\/pre>\n\n<p>With some digging in developer tools, it looks like the error begins here.....<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/H5AhV.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/H5AhV.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Setting 'strict' to true, the model still loads and does not throw and error.<\/p>\n\n<pre><code>var model = await \n        tf.loadGraphModel('\/modelsjs\/model_0\/model.json', {onProgress:onProgressCallback, strict:true});\n\n<\/code><\/pre>\n\n<p>Sadly, i am unable to share the model as it is proprietary.<\/p>\n","answers":[{"AnswerId":"57797525","CreationDate":"2019-09-05T02:01:36.947","ParentId":null,"OwnerUserId":"2723463","Title":null,"Body":"<p>The model is wrong.<\/p>\n\n<p>Originally a Keras model which was converted to a Tensorflow SavedModel.  Converting this model to Tensorflowjs did not work.<\/p>\n\n<p>However, converting the original model from Keras to Tensorflowjs worked.  <\/p>\n\n<p>Lesson learned, don't mix your models so much!<\/p>\n"}]}
+{"QuestionId":57782409,"AnswerCount":1,"Tags":"<python><tensorflow><nlp>","CreationDate":"2019-09-04T06:09:29.887","AcceptedAnswerId":57830166.0,"OwnerUserId":4544413.0,"Title":"Set the number of iterations gpt-2","Body":"<p>I'm fine tunning a gpt-2 model following this tutorial: <\/p>\n\n<p><a href=\"https:\/\/medium.com\/@ngwaifoong92\/beginners-guide-to-retrain-gpt-2-117m-to-generate-custom-text-content-8bb5363d8b7f\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@ngwaifoong92\/beginners-guide-to-retrain-gpt-2-117m-to-generate-custom-text-content-8bb5363d8b7f<\/a><\/p>\n\n<p>With is associated GitHub repository: <\/p>\n\n<p><a href=\"https:\/\/github.com\/nshepperd\/gpt-2\" rel=\"nofollow noreferrer\">https:\/\/github.com\/nshepperd\/gpt-2<\/a><\/p>\n\n<p>I have been able to replicate the examples, my issue is that I'm not fiding a parameter to set the number of iterations. \nBasically the training script shows a sample every 100 iterations and save a model version every 1000 iterations. But I'm not finding a parameter to train it for say, 5000 iterations and then close it. <\/p>\n\n<p>The script for training is here: \n<a href=\"https:\/\/github.com\/nshepperd\/gpt-2\/blob\/finetuning\/train.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/nshepperd\/gpt-2\/blob\/finetuning\/train.py<\/a><\/p>\n\n<p>EDIT: <\/p>\n\n<p>As suggested by cronoik I'm trying to replace the while for a for loop. <\/p>\n\n<p>I'm adding these changes: <\/p>\n\n<p>1) Adding one aditional argument: <\/p>\n\n<pre><code>parser.add_argument('--training_steps', metavar='STEPS', type=int, default=1000, help='a number representing how many training steps the model shall be trained for')\n<\/code><\/pre>\n\n<p>2) Changing the loop: <\/p>\n\n<pre><code>    try:\n        for iter_count in range(training_steps):\n            if counter % args.save_every == 0:\n                save()\n<\/code><\/pre>\n\n<p>3) Using the new argument: <\/p>\n\n<pre><code>python3 train.py --training_steps 300\n<\/code><\/pre>\n\n<p>But I'm getting this error: <\/p>\n\n<pre><code>  File \"train.py\", line 259, in main\n    for iter_count in range(training_steps):\nNameError: name 'training_steps' is not defined\n<\/code><\/pre>\n","answers":[{"AnswerId":"57830166","CreationDate":"2019-09-07T02:11:11.850","ParentId":null,"OwnerUserId":"6664872","Title":null,"Body":"<p>All you have to do is to modify the <code>while True<\/code> loop to a <code>for<\/code> loop:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>try:\n    #replaced\n    #while True:\n    for i in range(5000):\n        if counter % args.save_every == 0:\n            save()\n        if counter % args.sample_every == 0:\n            generate_samples()\n        if args.val_every &gt; 0 and (counter % args.val_every == 0 or counter == 1):\n            validation()\n\n        if args.accumulate_gradients &gt; 1:\n            sess.run(opt_reset)\n            for _ in range(args.accumulate_gradients):\n                sess.run(\n                    opt_compute, feed_dict={context: sample_batch()})\n            (v_loss, v_summary) = sess.run((opt_apply, summaries))\n        else:\n            (_, v_loss, v_summary) = sess.run(\n                (opt_apply, loss, summaries),\n                feed_dict={context: sample_batch()})\n\n        summary_log.add_summary(v_summary, counter)\n\n        avg_loss = (avg_loss[0] * 0.99 + v_loss,\n                    avg_loss[1] * 0.99 + 1.0)\n\n        print(\n            '[{counter} | {time:2.2f}] loss={loss:2.2f} avg={avg:2.2f}'\n            .format(\n                counter=counter,\n                time=time.time() - start_time,\n                loss=v_loss,\n                avg=avg_loss[0] \/ avg_loss[1]))\n\n        counter += 1\nexcept KeyboardInterrupt:\n    print('interrupted')\n    save()\n<\/code><\/pre>\n"}]}
+{"QuestionId":57782962,"AnswerCount":1,"Tags":"<python><tensorflow><keyerror>","CreationDate":"2019-09-04T06:53:10.403","AcceptedAnswerId":null,"OwnerUserId":12018044.0,"Title":"How to fix ' KeyError: 'accuracy' ' when running flowers_tf_lite.ipynb?","Body":"<p>When I run the <code>flowers_tf_lite.ipynb<\/code> script.<\/p>\n\n<p><a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/examples\/blob\/master\/community\/en\/flowers_tf_lite.ipynb\" rel=\"nofollow noreferrer\"><strong>Link to flowers_tf_lite.ipynb script<\/strong><\/a><\/p>\n\n<p>The getting <code>KeyError: 'accuracy'<\/code> error stopped the program.<\/p>\n\n<p>How can I solve this type of error?<\/p>\n\n<pre><code>acc = history.history['accuracy']\n\nval_acc = history.history['val_accuracy']\n\nloss = history.history['loss']\n\nval_loss = history.history['val_loss']\n\nplt.figure(figsize=(8, 8))\n\nplt.subplot(2, 1, 1)\n\nplt.plot(acc, label='Training Accuracy')\n\nplt.plot(val_acc, label='Validation Accuracy')\n\nplt.legend(loc='lower right')\n\nplt.ylabel('Accuracy')\n\nplt.ylim([min(plt.ylim()),1])\n\nplt.title('Training and Validation Accuracy')\n\n\nplt.subplot(2, 1, 2)\n\nplt.plot(loss, label='Training Loss')\n\nplt.plot(val_loss, label='Validation Loss')\n\nplt.legend(loc='upper right')\n\nplt.ylabel('Cross Entropy')\n\nplt.ylim([0,1.0])\n\nplt.title('Training and Validation Loss')\n\nplt.xlabel('epoch')\n\nplt.show()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/qdLVO.png\" rel=\"nofollow noreferrer\">Img of KeyError: 'accuracy' error<\/a><\/p>\n","answers":[{"AnswerId":"57784598","CreationDate":"2019-09-04T08:37:22.360","ParentId":null,"OwnerUserId":"5973377","Title":null,"Body":"<p>The accuracy metric is named as 'acc' and 'val_acc' after model compile, you can check that using<\/p>\n\n<pre><code>model.metrics_names\n<\/code><\/pre>\n\n<p>which will give you <\/p>\n\n<blockquote>\n  <p>['loss', 'acc', 'val_acc']<\/p>\n<\/blockquote>\n\n<p>So you just need to change first 2 lines is above code<\/p>\n\n<pre><code>acc = history.history['acc']\nval_acc = history.history['val_acc']\n<\/code><\/pre>\n"}]}
+{"QuestionId":57783029,"AnswerCount":2,"Tags":"<numpy><pytorch><numpy-slicing>","CreationDate":"2019-09-04T06:57:24.560","AcceptedAnswerId":null,"OwnerUserId":420867.0,"Title":"Is indexing with a bool array or index array faster in numpy\/pytorch?","Body":"<p>I can index my numpy array \/ pytorch tensor with a boolean array\/tensor of the same shape or an array\/tensor containing integer indexes of the elements I'm after. Which is faster?<\/p>\n","answers":[{"AnswerId":"57783030","CreationDate":"2019-09-04T06:57:24.560","ParentId":null,"OwnerUserId":"420867","Title":null,"Body":"<p>The following tests indicate that it's generally 3x to 20x faster with an index array in both numpy and pytorch:<\/p>\n\n<pre><code>In [1]: a = torch.arange(int(1e5))\nidxs = torch.randint(len(a), (int(1e4),))\nind = torch.zeros_like(a, dtype=torch.uint8)\nind[idxs] = 1\nac, idxsc, indc = a.cuda(), idxs.cuda(), ind.cuda()\n\nIn [2]: %timeit a[idxs]\n73.4 \u00b5s \u00b1 1 \u00b5s per loop (mean \u00b1 std. dev. of 7 runs, 10000 loops each)\n\nIn [3]: %timeit a[ind]\n622 \u00b5s \u00b1 8.99 \u00b5s per loop (mean \u00b1 std. dev. of 7 runs, 1000 loops each)\n\nIn [4]: %timeit ac[idxsc]\n9.51 \u00b5s \u00b1 475 ns per loop (mean \u00b1 std. dev. of 7 runs, 100000 loops each)\n\nIn [5]: %timeit ac[indc]\n59.6 \u00b5s \u00b1 313 ns per loop (mean \u00b1 std. dev. of 7 runs, 10000 loops each)\n\nIn [6]: idxs = torch.arange(len(a)-1, dtype=torch.long)\nind = torch.zeros_like(a, dtype=torch.uint8)\nind[idxs] = 1\nac, idxsc, indc = a.cuda(), idxs.cuda(), ind.cuda()\n\nIn [7]: %timeit a[idxs]\n146 \u00b5s \u00b1 14.2 \u00b5s per loop (mean \u00b1 std. dev. of 7 runs, 10000 loops each)\n\nIn [8]: %timeit a[ind]\n4.59 ms \u00b1 106 \u00b5s per loop (mean \u00b1 std. dev. of 7 runs, 100 loops each)\n\n\nIn [9]: %timeit ac[idxsc]\n33 \u00b5s \u00b1 15.1 ns per loop (mean \u00b1 std. dev. of 7 runs, 10000 loops each)\n\nIn [10]: %timeit ac[indc]\n85.9 \u00b5s \u00b1 56.9 ns per loop (mean \u00b1 std. dev. of 7 runs, 10000 loops each)\n<\/code><\/pre>\n"},{"AnswerId":"58926955","CreationDate":"2019-11-19T05:01:05.657","ParentId":null,"OwnerUserId":"10986701","Title":null,"Body":"<p>As the prior solution shows, I would expect the integer based indexing to be faster, since the dimension of the output tensor equals the indexing tensors dimension, which makes memory allocation easier.<\/p>\n"}]}
+{"QuestionId":57783468,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><deep-learning>","CreationDate":"2019-09-04T07:27:52.567","AcceptedAnswerId":null,"OwnerUserId":11694959.0,"Title":"How to custom train Face-Aging-with-Identity-Preserved-Conditional-Generative-Adversarial-Networks","Body":"<p>I am currently working on a project using GAN.\nSearching for several models, I found a model called Conditional GAN and found the Face-Aging-with-Identity-Preserved-Conditional-Generative-Adversarial-Networks project on github.\n<a href=\"https:\/\/github.com\/dawei6875797\/Face-Aging-with-Identity-Preserved-Conditional-Generative-Adversarial-Networks\" rel=\"nofollow noreferrer\">https:\/\/github.com\/dawei6875797\/Face-Aging-with-Identity-Preserved-Conditional-Generative-Adversarial-Networks<\/a><\/p>\n\n<p>The Readme.md for this project described how to download the data and how to train the model.\nViewing the results using the data provided by this project and the trained model was a success. But I failed to train from the data I wanted.\nI have tried many things about this, changing the path of the data, etc. but have not been successful. I'm not sure what I missed.<\/p>\n\n<p>this is my code<\/p>\n\n<pre><code>import os.path\nimport os\nos.environ['CUDA_DEVICES_ORDER'] = \"PCI_BUS_ID\"\nos.environ[\"CUDA_VISIBLE_DEVICES\"] = '0'\nimport numpy as np\nimport tensorflow as tf\nfrom datetime import datetime\nfrom models import FaceAging\nimport sys\nsys.path.append('.\/tools\/')\nfrom source_input import load_source_batch3\nfrom utils import save_images, save_source\nfrom data_generator import ImageDataGenerator\n\nflags = tf.app.flags\nflags.DEFINE_float(\"learning_rate\", 0.001, \"Learning rate\")\n\nflags.DEFINE_integer(\"batch_size\", 32, \"The size of batch images\")\n\nflags.DEFINE_integer(\"image_size\", 128, \"the size of the generated image\")\n\nflags.DEFINE_integer(\"noise_dim\", 256, \"the length of the noise vector\")\n\nflags.DEFINE_integer(\"feature_size\", 128, \"image size after stride 2 conv\")\n\nflags.DEFINE_integer(\"age_groups\", 5, \"the number of different age groups\")\n\nflags.DEFINE_integer('max_steps', 200000, 'Number of batches to run')\n\nflags.DEFINE_string(\"alexnet_pretrained_model\", \"pre_trained\/alexnet.model-292000\",\n                    \"Directory name to save the checkpoints\")\n\nflags.DEFINE_string(\"age_pretrained_model\", \"pre_trained\/age_classifier.model-300000\",\n                    \"Directory name to save the checkpoints\")\n\nflags.DEFINE_integer('model_index', None, 'the index of trained model')\n\nflags.DEFINE_float(\"gan_loss_weight\", 75, \"gan_loss_weight\")\n\nflags.DEFINE_float(\"fea_loss_weight\", 0.5e-4, \"fea_loss_weight\")\n\nflags.DEFINE_float(\"age_loss_weight\", 30, \"age_loss_weight\")\n\nflags.DEFINE_float(\"tv_loss_weight\", None, \"face_loss_weight\")\n\nflags.DEFINE_string(\"checkpoint_dir\", \"checkpoints\/age\/0_conv5_lsgan_transfer_g75_0.5f-4_a30\", \"Directory name to save the checkpoints\")\n\nflags.DEFINE_string(\"source_checkpoint_dir\", ' ', \"Directory name to save the checkpoints\")\n\nflags.DEFINE_string(\"sample_dir\", \"age\/0_conv5_lsgan_transfer_g75_0.5f-4_a30 \", \"Directory name to save the sample images\")\n\nflags.DEFINE_string(\"fea_layer_name\", 'conv5', \"which layer to use for fea_loss\")\n\nflags.DEFINE_string(\"source_file\", \"train\/\", \"source file path\")\n\nflags.DEFINE_string(\"root_folder\", \"CACD_cropped_400\/\", \"folder that contains images\")\n\nFLAGS = flags.FLAGS\n\n# How often to run a batch through the validation model.\nVAL_INTERVAL = 5000\n\n# How often to save a model checkpoint\nSAVE_INTERVAL = 10000\n\nd_iter = 1\ng_iter = 1\n\nconfig = tf.ConfigProto()\nconfig.gpu_options.allow_growth = True\n\n# Initalize the data generator seperately for the training and validation set\ntrain_generator = ImageDataGenerator(batch_size=FLAGS.batch_size, height=FLAGS.feature_size, width=FLAGS.feature_size,\n                                     z_dim=FLAGS.noise_dim, scale_size=(FLAGS.image_size, FLAGS.image_size), mode='train')\ndef my_train():\n    with tf.Graph().as_default():\n        sess = tf.Session(config=config)\n        model = FaceAging(sess=sess, lr=FLAGS.learning_rate, keep_prob=1., model_num=FLAGS.model_index, batch_size=FLAGS.batch_size,\n                        age_loss_weight=FLAGS.age_loss_weight, gan_loss_weight=FLAGS.gan_loss_weight,\n                        fea_loss_weight=FLAGS.fea_loss_weight, tv_loss_weight=FLAGS.tv_loss_weight)\n\n        imgs = tf.placeholder(tf.float32, [FLAGS.batch_size, FLAGS.image_size, FLAGS.image_size, 3])\n        true_label_features_128 = tf.placeholder(tf.float32, [FLAGS.batch_size, 128, 128, FLAGS.age_groups])\n        true_label_features_64 = tf.placeholder(tf.float32, [FLAGS.batch_size, 64, 64, FLAGS.age_groups])\n        false_label_features_64 = tf.placeholder(tf.float32, [FLAGS.batch_size, 64, 64, FLAGS.age_groups])\n        age_label = tf.placeholder(tf.int32, [FLAGS.batch_size])\n\n        source_img_227, source_img_128, face_label = load_source_batch3(FLAGS.source_file, FLAGS.root_folder, FLAGS.batch_size)\n\n        model.train_age_lsgan_transfer(source_img_227, source_img_128, imgs, true_label_features_128,\n                                       true_label_features_64, false_label_features_64, FLAGS.fea_layer_name, age_label)\n\n        ge_samples = model.generate_images(imgs, true_label_features_128, reuse=True, mode='train')\n\n        # Create a saver.\n        model.saver = tf.train.Saver(model.save_d_vars + model.save_g_vars, max_to_keep=200)\n        model.alexnet_saver = tf.train.Saver(model.alexnet_vars)\n        model.age_saver = tf.train.Saver(model.age_vars)\n\n        d_error = model.d_loss\/model.gan_loss_weight\n        g_error = model.g_loss\/model.gan_loss_weight\n        fea_error = model.fea_loss\/model.fea_loss_weight\n        age_error = model.age_loss\/model.age_loss_weight\n\n        # Start running operations on the Graph.\n        sess.run(tf.global_variables_initializer())\n        tf.train.start_queue_runners(sess)\n\n        model.alexnet_saver.restore(sess, FLAGS.alexnet_pretrained_model)\n        model.age_saver.restore(sess, FLAGS.age_pretrained_model)\n\n        if model.load(FLAGS.checkpoint_dir, model.saver):\n            print(\" [*] Load SUCCESS\")\n        else:\n            print(\" [!] Load failed...\")\n\n        print(\"{} Start training...\")\n\n        # Loop over max_steps\n        for step in range(FLAGS.max_steps):\n            images, t_label_features_128, t_label_features_64, f_label_features_64, age_labels = \\\n                train_generator.next_target_batch_transfer2()\n            dict = {imgs: images,\n                    true_label_features_128: t_label_features_128,\n                    true_label_features_64: t_label_features_64,\n                    false_label_features_64: f_label_features_64,\n                    age_label: age_labels\n                    }\n            for i in range(d_iter):\n                _, d_loss = sess.run([model.d_optim, d_error], feed_dict=dict)\n\n            for i in range(g_iter):\n                _, g_loss, fea_loss, age_loss = sess.run([model.g_optim, g_error, fea_error, age_error],\n                                                         feed_dict=dict)\n\n            format_str = ('%s: step %d, d_loss = %.3f, g_loss = %.3f, fea_loss=%.3f, age_loss=%.3f')\n            print(format_str % (datetime.now(), step, d_loss, g_loss, fea_loss, age_loss))\n\n            # Save the model checkpoint periodically.\n            if step % SAVE_INTERVAL == SAVE_INTERVAL-1 or (step + 1) == FLAGS.max_steps:\n                checkpoint_path = os.path.join(FLAGS.checkpoint_dir)\n                model.save(checkpoint_path, step, 'acgan')\n\n            if step % VAL_INTERVAL == VAL_INTERVAL-1:\n                if not os.path.exists(FLAGS.sample_dir):\n                    os.makedirs(FLAGS.sample_dir)\n                path = os.path.join(FLAGS.sample_dir, str(step))\n                if not os.path.exists(path):\n                    os.makedirs(path)\n\n                source = sess.run(source_img_128)\n                save_source(source, [4, 8], os.path.join(path, 'source.jpg'))\n                for j in range(train_generator.n_classes):\n                    true_label_fea = train_generator.label_features_128[j]\n                    dict = {\n                            imgs: source,\n                            true_label_features_128: true_label_fea\n                            }\n                    samples = sess.run(ge_samples, feed_dict=dict)\n                    save_images(samples, [4, 8], '.\/{}\/test_{:01d}.jpg'.format(path, j))\n\ndef main(argv=None):\n    my_train()\n\n\nif __name__ == '__main__':\n    tf.app.run()\n<\/code><\/pre>\n\n<p>this is error msg<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"C:\/Users\/admin\/Desktop\/Face\/Face-Aging-with-Identity-Preserved-Conditional-Generative-Adversarial-Networks-master\/age_lsgan_transfer.py\", line 167, in &lt;module&gt;\n    tf.app.run()\n  File \"C:\\Users\\admin\\PycharmProjects\\Deep\\venv\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\", line 40, in run\n    _run(main=main, argv=argv, flags_parser=_parse_flags_tolerate_undef)\n  File \"C:\\Users\\admin\\PycharmProjects\\Deep\\venv\\lib\\site-packages\\absl\\app.py\", line 300, in run\n    _run_main(main, args)\n  File \"C:\\Users\\admin\\PycharmProjects\\Deep\\venv\\lib\\site-packages\\absl\\app.py\", line 251, in _run_main\n    sys.exit(main(argv))\n  File \"C:\/Users\/admin\/Desktop\/Face\/Face-Aging-with-Identity-Preserved-Conditional-Generative-Adversarial-Networks-master\/age_lsgan_transfer.py\", line 163, in main\n    my_train()\n  File \"C:\/Users\/admin\/Desktop\/Face\/Face-Aging-with-Identity-Preserved-Conditional-Generative-Adversarial-Networks-master\/age_lsgan_transfer.py\", line 88, in my_train\n    source_img_227, source_img_128, face_label = load_source_batch3(FLAGS.source_file, FLAGS.root_folder, FLAGS.batch_size)\n  File \"C:\\Users\\admin\\Desktop\\Face\\Face-Aging-with-Identity-Preserved-Conditional-Generative-Adversarial-Networks-master\\source_input.py\", line 128, in load_source_batch3\n    img_list, label_list = get_imgAndlabel_list2(filename, img_folder)\n  File \"C:\\Users\\admin\\Desktop\\Face\\Face-Aging-with-Identity-Preserved-Conditional-Generative-Adversarial-Networks-master\\source_input.py\", line 173, in get_imgAndlabel_list2\n    f = open(filename, 'r')\nPermissionError: [Errno 13] Permission denied: 'train\/'\n<\/code><\/pre>\n\n<p>this code is source_input.py<\/p>\n\n<pre><code>\nfrom __future__ import absolute_import\nfrom __future__ import division\nfrom __future__ import print_function\n\nimport argparse\nimport os\nimport sys\n\nimport tensorflow as tf\n\n#from read_image import *\nfrom tensorflow.python.platform import gfile\nfrom tensorflow.python.platform import flags\nimport numpy as np\nimport scipy.io as scio\nfrom tensorflow.python.framework import ops\nfrom PIL import Image\nFLAGS = flags.FLAGS\n\nT = 1\nIM_HEIGHT = 400\nIM_WIDTH = 400\nIM_CHANNELS = 3\n\n\ndef _int64_feature(value):\n  return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))\n\n\ndef _bytes_feature(value):\n  return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))\n\ndef read_images(filename_queue, new_height=None, new_width=None):\n\n    reader = tf.WholeFileReader()\n    key, value = reader.read(filename_queue)\n    image = tf.image.decode_jpeg(value)  # use png or jpeg decoder based on your files\n    image = tf.reshape(image, [IM_HEIGHT, IM_WIDTH, IM_CHANNELS])\n\n    if new_height and new_width:\n        image = tf.image.resize_images(image, [new_height, new_width])\n\n    image = tf.cast(image, tf.float32) - np.array([104., 117., 124.])\n\n    return image\n\n\ndef read_images2(filename_queue):\n\n    reader = tf.WholeFileReader()\n    key, value = reader.read(filename_queue)\n    image = tf.image.decode_jpeg(value)  # use png or jpeg decoder based on your files\n    image = tf.reshape(image, [IM_HEIGHT, IM_WIDTH, IM_CHANNELS])\n\n    image_227 = tf.image.resize_images(image, [227, 227])\n    image_227 = tf.cast(image_227, tf.float32) - np.array([104., 117., 124.])\n\n    image_128 = tf.image.resize_images(image, [128, 128])\n    image_128 = tf.cast(image_128, tf.float32) - np.array([104., 117., 124.])\n\n    return image_227, image_128\n\ndef read_images3(input_queue):\n\n    label = input_queue[1]\n    file_contents = tf.read_file(input_queue[0])\n    image = tf.image.decode_image(file_contents, channels=3)\n    image = tf.reshape(image, [IM_HEIGHT, IM_WIDTH, IM_CHANNELS])\n\n    image_227 = tf.image.resize_images(image, [227, 227])\n    image_227 = tf.cast(image_227, tf.float32) - np.array([104., 117., 124.])\n\n    image_128 = tf.image.resize_images(image, [128, 128])\n    # image_128 = tf.cast(image_128, tf.float32)\n    image_128 = tf.cast(image_128, tf.float32) - np.array([104., 117., 124.])\n\n    return image_227, image_128, label\n\ndef load_source_batch(filename, img_folder, batch_size, img_size, shuffle=True):\n    filenames = get_imgAndlabel_list(filename, img_folder)\n    print('%d images to train' %(len(filenames)))\n    if not filenames:\n        raise RuntimeError('No data files found.')\n\n    with tf.name_scope('input'):\n        filename_queue = tf.train.string_input_producer(filenames, shuffle=shuffle)\n\n        # Even when reading in multiple threads, share the filename queue.\n        image = read_images(filename_queue, new_height=img_size, new_width=img_size)\n        image_batch = tf.train.shuffle_batch(\n            [image],\n            batch_size=batch_size,\n            num_threads=4,\n            capacity=1280,\n            min_after_dequeue=640)\n        # image_batch = tf.train.batch(\n        #     [image],\n        #     batch_size=batch_size,\n        #     num_threads=4,\n        #     capacity=1280)\n        #\n        return image_batch\n\ndef load_source_batch2(filename, img_folder, batch_size, shuffle=True):\n\n    filenames = get_imgAndlabel_list(filename, img_folder)\n    print('%d images to train' % (len(filenames)))\n    if not filenames:\n        raise RuntimeError('No data files found.')\n\n    with tf.name_scope('input'):\n        filename_queue = tf.train.string_input_producer(filenames, shuffle=shuffle)\n\n        # Even when reading in multiple threads, share the filename queue.\n        image_227, image_128 = read_images2(filename_queue)\n        image_227_batch, image_128_batch = tf.train.shuffle_batch(\n            [image_227, image_128],\n            batch_size=batch_size,\n            num_threads=4,\n            capacity=1280,\n            min_after_dequeue=640)\n\n        return image_227_batch, image_128_batch\n\ndef load_source_batch3(filename, img_folder, batch_size, shuffle=True):\n\n    img_list, label_list = get_imgAndlabel_list2(filename, img_folder)\n    print('%d images to train' % (len(img_list)))\n\n    images = ops.convert_to_tensor(img_list, dtype=tf.string)\n    labels = ops.convert_to_tensor(label_list, dtype=tf.int32)\n\n    # Makes an input queue\n    input_queue = tf.train.slice_input_producer([images, labels], shuffle=shuffle)\n\n    # Even when reading in multiple threads, share the filename queue.\n    image_227, image_128, label = read_images3(input_queue)\n    image_227_batch, image_128_batch, label_batch = tf.train.shuffle_batch(\n        [image_227, image_128, label],\n        batch_size=batch_size,\n        num_threads=4,\n        capacity=1280,\n        min_after_dequeue=640)\n\n    return image_227_batch, image_128_batch, label_batch\n\ndef get_imgAndlabel_list(filename, img_folder):\n    \"\"\"\n    :param filename:\n     each line in filename is img_name \\space label\n    :return:\n    img names list\n    label list\n    \"\"\"\n    f = open(filename, 'r')\n    lines = f.readlines()\n    f.close()\n    imgname_lists = []\n    for i in range(len(lines)):\n        img_name = lines[i].split()[0]\n        imgname_lists.append(os.path.join(img_folder, img_name))\n    return imgname_lists\n\ndef get_imgAndlabel_list2(filename, img_folder):\n    \"\"\"\n    :param filename:\n     each line in filename is img_name \\space label\n    :return:\n    img names list\n    label list\n    \"\"\"\n    f = open(filename, 'r')\n    lines = f.readlines()\n    f.close()\n    imgname_lists = []\n    label_lists = []\n    for i in range(len(lines)):\n        img_name, label = lines[i].split()\n        imgname_lists.append(os.path.join(img_folder, img_name))\n        label_lists.append(int(label))\n\n    return imgname_lists, label_lists\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57784115,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2019-09-04T08:10:16.320","AcceptedAnswerId":57786112.0,"OwnerUserId":11309950.0,"Title":"How can I use two-dimensional data as input to Conv1D?","Body":"<p>Here is my code<\/p>\n\n<pre><code>def create_dataset(signal_data, look_back=1):\n    dataX, dataY = [], []\n    for i in range(len(signal_data) - look_back):\n        dataX.append(signal_data[i:(i + look_back), :])\n        dataY.append(signal_data[i + look_back, -1])\n    return np.array(dataX), np.array(dataY)\n\n\nlook_back = 200\n\n\n\n\nstock = 'a.csv'\nscaler = MinMaxScaler()\n\ndf = pd.read_csv(stock)\n\n\nsignal_data = df[[\"close\"]].values.astype('float32')\n\n\n\nsignal_data = scaler.fit_transform(signal_data)\n\ntrain_size = int(len(signal_data) * 0.80)\ntest_size = len(signal_data) - train_size\n\ntrain = signal_data[0:train_size]\ntest = signal_data[train_size:]\n\nx_train, y_train = create_dataset(train, look_back)\nx_test, y_test = create_dataset(test, look_back)\n\n\nCNN = Sequential()\nCNN.add(Conv1D(filters=16, kernel_size=2, padding='same', activation='relu', input_shape=(None, 200)))\nCNN.add(MaxPooling1D(pool_size=1))\nCNN.add(Flatten())\n<\/code><\/pre>\n\n<p><code>x_train<\/code> and <code>y_train<\/code> have shapes (1399, 200, 1), (1399) by create_dataset.<\/p>\n\n<p>But Conv1D accepts input in three dimensions.<\/p>\n\n<p>To solve this, I tried to set <code>input_shape = (None, 200)<\/code> (since loock_back is 200).<\/p>\n\n<p>but, this method occurs a <code>ValueError: The shape of the input to \"Flatten\" is not fully defined (got (None, 16). Make sure to pass a complete \"input_shape\" or \"batch_input_shape\" argument to the first layer in your mode<\/code>.<\/p>\n\n<p>How can I solve this problem?<\/p>\n","answers":[{"AnswerId":"57786112","CreationDate":"2019-09-04T10:05:07.327","ParentId":null,"OwnerUserId":"10606534","Title":null,"Body":"<p>You should try 200 1 as input To first Layer<\/p>\n\n<p>If I understand correct , 1399 is the batch size right?<\/p>\n\n<p>When using Conv1D as the first layer in a model, provide an input_shape argument (tuple of integers or None, e.g. (10, 128) for sequences of 10 vectors of 128-dimensional vectors, or (None, 128) for variable-length sequences of 128-dimensional vectors.\nif your sequences  is 200 you can set (200 ,1) ignore the batch size<\/p>\n"}]}
+{"QuestionId":57785375,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-09-04T09:24:51.717","AcceptedAnswerId":null,"OwnerUserId":11377572.0,"Title":"keras train multiple models and create multiple sequential instances","Body":"<p>I want to create multiple sequential models with keras. If I create multiple instances of sequential class, and train each of them individually with for loop, train 1 model in 1 iteration, would the second, 3rd and following models automatically inherit the weights of previous trained models in the previous iterations?<\/p>\n","answers":[]}
+{"QuestionId":57785455,"AnswerCount":1,"Tags":"<python><machine-learning><keras><random-forest>","CreationDate":"2019-09-04T09:29:43.113","AcceptedAnswerId":null,"OwnerUserId":12018879.0,"Title":"accuracy of keras is not improving","Body":"<p>I am comparing Keras with Random forest. I follow research paper and it give keras model is high accuracy as compare to random forest model, but when i implemented it not give me. Accuracy and STD of RF* 0.997 0.0006\n     Accuracy of Keras 0.0079<\/p>\n\n<pre><code>#Importing dataset \n    dataset = pd.read_csv('KDD_Dataset.csv')\n    X = dataset.iloc[:, :-1].values\n    y = dataset.iloc[:, 41:42].values\n\n    from sklearn.preprocessing import LabelEncoder\n    labelencoder_X = LabelEncoder()\n    X[:,0] = labelencoder_X.fit_transform(X[:,0])\n    X[:,1] = labelencoder_X.fit_transform(X[:,1])\n    X[:,2] = labelencoder_X.fit_transform(X[:,2])\n    #\n    from sklearn.preprocessing import OneHotEncoder\n    onehotencoder_0 = OneHotEncoder(categorical_features=[0])\n    onehotencoder_1 = OneHotEncoder(categorical_features=[1])\n    onehotencoder_2 = OneHotEncoder(categorical_features=[2])\n    X = onehotencoder_0.fit_transform(X).toarray()\n    X = onehotencoder_1.fit_transform(X).toarray()\n    X = onehotencoder_2.fit_transform(X).toarray()\n     Encoding categorical data y\n    from sklearn.preprocessing import LabelEncoder\n    labelencoder_y = LabelEncoder()\n    y = labelencoder_y.fit_transform(y)\n    max(y)\n<\/code><\/pre>\n\n<p>Splitting the dataset into the Training set and Test set<\/p>\n\n<pre><code>        #from sklearn.cross_validation import train_test_split\n          from sklearn.model_selection import train_test_split\n    X_train, X_test, y_train, y_test = train_test_split(X, y, \n                                                        test_size = 0.2, \n                                                        random_state = 1)\n    \"\"\"sc_y = StandardScaler()\n    y_train = sc_y.fit_transform(y_train)\"\"\"\n\n#Feature Scaling\n       from sklearn.preprocessing import StandardScaler\n       sc = StandardScaler()\n       X_train = sc.fit_transform(X_train)\n       X_test = sc.transform(X_test)\n<\/code><\/pre>\n\n<p>Fitting Random Forest Classification to the Training set<\/p>\n\n<pre><code>        from sklearn.ensemble import RandomForestClassifier\n    classifier = RandomForestClassifier(n_estimators = 500, \n                                        criterion = 'entropy', \n                                        random_state = 0,\n                                        oob_score = True)\n    classifier.fit(X_train, y_train)\n\n    y_pred = classifier.predict(X_test)\n<\/code><\/pre>\n\n<p>Making the Confusion Matrix<\/p>\n\n<pre><code>      from sklearn.metrics import confusion_matrix\n    cm = confusion_matrix(y_test, y_pred)\n\n\n    from sklearn.model_selection import cross_val_score\n    accuracies = cross_val_score(estimator= classifier, \n                                 X = X_train,\n                                 y = y_train,\n                                 cv=10)\n\n    accuracies_mean = accuracies.mean()\n    accuracies_std = accuracies.std()\n\n    print(\"Accuracy and STD of RF\")\n    print(accuracies_mean)\n    print(accuracies_std)\n<\/code><\/pre>\n\n<p>Keras Model<\/p>\n\n<pre><code>    model = Sequential()\n    model.add(Dense(12, input_dim=45, activation='relu'))\n    model.add(Dense(8, activation='relu'))\n    model.add(Dense(1, activation='sigmoid'))\n\n\n    from keras import optimizers\n\n    numpy.random.seed(7)\n    import datetime, os\n    logdir = os.path.join(\"logs\", datetime.datetime.now().strftime(\"%Y%m%d-%H%M%S\"))\n    tensorboard_callback = tf.keras.callbacks.TensorBoard(logdir, histogram_freq=1)\n    sgd = optimizers.SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)\n\n    model.compile(loss='mean_squared_error', optimizer=sgd, metrics=['accuracy'])\n\n    sgd = optimizers.SGD(lr=0.01, clipnorm=1.)\n\n\n              model.fit(X_train, y_train,\n              batch_size=50000,\n              epochs=10,\n              verbose=1,\n              validation_data=(X_test, y_test),\n              callbacks=None)\n\n    y_pred = model.predict(X_test)\n\n    score = model.evaluate(X_test, y_test, verbose=1)\n<\/code><\/pre>\n\n<p>suggest me how i can improve accuracy of keras<\/p>\n","answers":[{"AnswerId":"57787415","CreationDate":"2019-09-04T11:26:37.420","ParentId":null,"OwnerUserId":"4685471","Title":null,"Body":"<p>You are in a <em>classification<\/em> (and not regression) setting, so you should <strong>not<\/strong> use MSE as your loss function in Keras (it is for regression problems); change your model compilation to<\/p>\n\n<pre><code>model.compile(loss='binary_crossentropy', optimizer=sgd, metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>See <a href=\"https:\/\/stackoverflow.com\/questions\/48775305\/what-function-defines-accuracy-in-keras-when-the-loss-is-mean-squared-error-mse\/48788577#48788577\">What function defines accuracy in Keras when the loss is mean squared error (MSE)?<\/a> for more details, albeit in an \"inverse\" setting (trying to use accuracy in a regression problem).<\/p>\n\n<p>Your <code>batch_size=50000<\/code> looks extremely high, but if you don't run into memory problems it could do.<\/p>\n"}]}
+{"QuestionId":57785752,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2019-09-04T09:45:43.863","AcceptedAnswerId":null,"OwnerUserId":12018753.0,"Title":"Creating a python algorithm to train a keras model to predict a large sequence of integers","Body":"<p>I'm new to machine learning but I'm trying to apply it to a project I have. I was able to train a model to convert words from one language to another using LSTM layers. Say I use A as input to my model and I get B as output. What I do is: <\/p>\n\n<blockquote>\n  <p>'original word' -> word embedding -> one-hot encode (A) -> MODEL -> one-hot encoded output (B) -> word embedding -> 'translated word'<\/p>\n<\/blockquote>\n\n<p>This is relatively simple as I'm using a character-level tokenizer to encode the words and that does not require much memory (small sequences, one for each word). <\/p>\n\n<p>However, I now have to train a model that takes B as input and gives me C (no longer a translation problem). C is later going to be used for different purposes. The difference is that C can have a length of say 315 numbers and each of them can be one of 5514 unique values i.e., shape(215, 5514). Generically what I want to do is, for example:<\/p>\n\n<blockquote>\n  <p>'banana' -> (some processing, word embedding or one-hot) -> MODEL -> [434, 434, 410, 321, 225, 146, 86, 43, 13, -8, -23, -32, -38, -41, -13, 101, 227, 332, 411, 470, 515, 550, 577, 597, 611, 622, 628, 622, 608, 593, 580, 570, 561, 554, 549, 547, 548, 548, 549, 555, 564, 572, 579, 584, 587, 589, 590, 591, 591, 591, 590, 590, 584, 567, 550, 535, 524, 516, 511, 506, 503, 503, 507, 511, 518, 530, 543, 553, 561, 568, 573, 577, 580, 582, 584, 585, 586, 586, 587, 587, 588, 588, 588, 588, 588, 586]<\/p>\n<\/blockquote>\n\n<p>So the <strong>problem<\/strong> is that I don't have <strong>enough memory<\/strong> to perform a one-hot encoding of the output sequences. I tried using generators to load each sequence from the disk instead of loading all of them from memory but It doesn't seem to be working.<\/p>\n\n<p>Do you have any suggestions as to how I should approach this problem?<\/p>\n\n<p>EDIT:\nThe dataset I'm using has the following format: n lines, each line contains 2 columns separated by a tab. The first column is the input word and the second column is the sequence I want to obtain if the input is that word.<\/p>\n","answers":[{"AnswerId":"57791481","CreationDate":"2019-09-04T15:22:29.710","ParentId":null,"OwnerUserId":"7121231","Title":null,"Body":"<p>One hot encoding increases number of columns according to unique categories in data set. I think you should check the performance of model with just using tokenizer not both. Because most of the time tokenizer alone performs very well.<\/p>\n"}]}
+{"QuestionId":57786253,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><keras><eager-execution>","CreationDate":"2019-09-04T10:12:31.260","AcceptedAnswerId":null,"OwnerUserId":9492673.0,"Title":"Equivalent of \"class_weight\" argument in eager execution training?","Body":"<p>I have a binary classification task in tensorflow. I've built a 3D-CNN with the tf.keras.Sequential class and called it <code>Net_1<\/code>. <\/p>\n\n<p>I have an <strong>unbalanced<\/strong> training dataset: class 0 has 171 negative samples and class 1 has 84 positive samples.<\/p>\n\n<p>Right now, given <code>epochs = 20<\/code>, I am training the model in <strong>eager execution<\/strong> with:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#... code where I define loss_function, grad_function and optimizer...\n\nfor epoch in range(1, epochs):  # make epochs start from 1\n\n    # define metrics\n    epoch_loss_avg = tf.keras.metrics.Mean(name='train_mean_loss')\n    epoch_bin_accuracy = tf.keras.metrics.BinaryAccuracy(name=\"train_binary_accuracy\")\n\n    for _train_patches, train_labels in batched_train_dataset:\n        # Optimize the model\n        loss_value, grads = grad(Net_1, _train_patches, train_labels)  # invoke function that computes the gradients\n        adam_optimizer.apply_gradients(zip(grads, Net_1.trainable_variables), global_step=global_step)  # update model's parameters through the optimizer\n\n        net_prediction = Net_1(_train_patches)  # compute model's output\n\n        # Track progress\n        epoch_loss_avg.update_state(loss_value)  # accumulate loss metric over current batch\n        epoch_bin_accuracy.update_state(y_true=train_labels, y_pred=net_prediction)  # accumulate accuracy metric over current batch\n\n    # end epoch: print loss and acc\n    print(\"Epoch {:03d}: Loss: {:.3f}, Bin_Accuracy: {:.3%}\".format(epoch, epoch_loss_avg.result(), epoch_bin_accuracy.result()))\n\n    # reset training metrics\n    epoch_loss_avg.reset_states()\n    epoch_bin_accuracy.reset_states()\n<\/code><\/pre>\n\n<p>Even though sometimes training works (loss decreases to zero and accuracy increases), some other times the model gets stuck and always outputs the majority class (in this case 0), and therefore loss and accuracy also get stuck.<\/p>\n\n<p>How can this be tackled in eager execution? I am looking for something equivalent to the <code>class_weight<\/code> attribute of the <code>fit<\/code> method, before resorting to minority class oversampling.<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[]}
+{"QuestionId":57786303,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2019-09-04T10:15:24.630","AcceptedAnswerId":58301761.0,"OwnerUserId":6878758.0,"Title":"Custom metric for Semantic segmentation","Body":"<p>I am working on a multi-class semantic segmentation task, and would like to define a custom, weighted metric for calculating how well my NN is performing. <\/p>\n\n<p>I am using U-net to segment my image into one of 8 classes, of which 1-7 are the particular classes and 0 is background. How do I use the standard custom metric template defined on the <a href=\"https:\/\/keras.io\/metrics\/\" rel=\"nofollow noreferrer\">Keras metrics<\/a> page, so that I only get the IoU of only channels 1-7, multiplied by a (1,7) weights array? I tried removing the background channel in the custom metric by using <\/p>\n\n<pre><code>y_true, y_pred = y_true[1:,:,:], y_pred[1:, :,:]\n<\/code><\/pre>\n\n<p>but it does not look like that's what I want. Any help will be appreciated.<\/p>\n","answers":[{"AnswerId":"58301761","CreationDate":"2019-10-09T10:06:59.813","ParentId":null,"OwnerUserId":"6878758","Title":null,"Body":"<p>The change that was necessary<\/p>\n\n<pre><code>def dice_coef_multilabel(y_true, y_pred, numLabels=CLASSES):\n    dice=0\n    for index in range(numLabels):\n        dice -= dice_coef(y_true[:,:,index], y_pred[:,:,index])\n    return dice\n<\/code><\/pre>\n\n<p>If needed, the dice coeff can be calcualted across channels by using two nested loops to loop over all the channel combinations. I'm also including the dice coefficient calculation.<\/p>\n\n<pre><code>def dice_coef(y_true, y_pred):\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) \/ (K.sum(y_true_f) + K.sum(y_pred_f) + smooth)\n<\/code><\/pre>\n\n<p>FWIW, <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/9395#issuecomment-379228094\" rel=\"nofollow noreferrer\">this github link<\/a> has various types of metrics were implemented in a channel-wise manner. <\/p>\n"}]}
+{"QuestionId":57787011,"AnswerCount":0,"Tags":"<python-3.x><keras><nlp><multiclass-classification>","CreationDate":"2019-09-04T11:00:49.037","AcceptedAnswerId":null,"OwnerUserId":12019297.0,"Title":"multi class labeling accuracy is very less?","Body":"<p>i have been working on kaggle competition.but the accuary of my model is very less plz help me<\/p>\n\n<p>i have tried lstm models but accuracy is still less. i have hot encoded my data set and then text to sequenced it<\/p>\n\n<p><strong>tokenzing<\/strong><\/p>\n\n<pre><code>tok = Tokenizer(num_words=max_features,filters='!\"#$%&amp;\\'()*+,-.\/:;&lt;=&gt;?@[\\\\]^_`{|}~',lower=True)\ntok.fit_on_texts(list(X_train) + list(X_test))\nx_train = tok.texts_to_sequences(X_train)\nx_test = tok.texts_to_sequences(X_test)\nprint(len(x_train), 'train sequences')\nprint(len(x_test), 'test sequences')\nword_index=tok.word_index\nword_index['PAD']=0\nword_index['START']=1\nword_index['UNX']=2\nprint('found %s unique tokens' %len(tok.word_index))\nprint('Average train sequence length: {}'.format(np.mean(list(map(len, x_train)), dtype=int)))\nprint('Average test sequence length: {}'.format(np.mean(list(map(len, x_test)), dtype=int)))\n\nx_train = sequence.pad_sequences(x_train, maxlen=maxlen)\nx_test = sequence.pad_sequences(x_test, maxlen=maxlen)\nprint('x_train shape:', x_train.shape)\nprint('x_test shape:', x_test.shape)\n<\/code><\/pre>\n\n<p><strong>Model<\/strong><\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Embedding(10000,128,input_length=9))\nmodel.add(SpatialDropout1D(0.4))\nmodel.add(Bidirectional(LSTM(128,dropout=0.4,recurrent_dropout=0.3)))\nmodel.add(Dense(128, input_dim=9, activation='relu'))\nmodel.add(Dense(128, activation='relu'))\nmodel.add(Dense(5, activation=\"softmax\"))\ntensorboard= TensorBoard(log_dir=\"log\\{}\".format(NAME))\nmodel.compile(optimizer=\"adam\",loss=\"categorical_crossentropy\", metrics=[\"accuracy\"])\n<\/code><\/pre>\n\n<p>the accuary is very less maximum is 0.51 when very is just one label<\/p>\n","answers":[]}
+{"QuestionId":57787201,"AnswerCount":1,"Tags":"<tensorflow><input><keras><preprocessor><tensorflow-serving>","CreationDate":"2019-09-04T11:13:19.743","AcceptedAnswerId":null,"OwnerUserId":8170648.0,"Title":"How preprocess and post-process at using TensorFlow Serving","Body":"<p>I am designing a sentence classification service using TensorFlow Serving. <\/p>\n\n<p>I succeeded boot up a docker server, request a prediction, and received a prediction results. <\/p>\n\n<p>My model's has two input layers and one output layer.<\/p>\n\n<pre><code>lexical_input (InputLayer)      [(None, 1000)]       0 \nphoneme_input (InputLayer)      [(None, 150)]        0                  \n... ... ... ... ... ... \ndense_2 (Dense)                 (None, 104)          43368       concatenate[0][0] \n<\/code><\/pre>\n\n<p>Now my service's input format and output format are vectors.<\/p>\n\n<p>Client(user) has to vectorize their input string, and they have to do argmax and map to get a label. <\/p>\n\n<p>I want my clients request with plane sentence not vector, and can get a label as a prediction results. <\/p>\n\n<p>I already have functions to vectorize input sentence and get a label from prediction result vector. <\/p>\n\n<p>I don't know how can I connect my functions and my model. <\/p>\n\n<p>As a result of searching, I had to do something when using tf.saved_model.simple_save().<\/p>\n","answers":[{"AnswerId":"57797028","CreationDate":"2019-09-05T00:29:06.977","ParentId":null,"OwnerUserId":"1871727","Title":null,"Body":"<p>One option is to use another webserver like flask to do the pre(post)-processing. You can do a http or gRPC call from flask to tf serving after pre-processing and use the response to do post processing.<\/p>\n"}]}
+{"QuestionId":57787386,"AnswerCount":1,"Tags":"<python><machine-learning><keras><neural-network><conv-neural-network>","CreationDate":"2019-09-04T11:24:47.247","AcceptedAnswerId":null,"OwnerUserId":10761064.0,"Title":"How to increase accuracy in Convolutional neural network","Body":"<p>I am implementing flower recognition - the dataset has few types of flowers.\nTotally the dataset contains about 4000 images<\/p>\n\n<p>My code - <\/p>\n\n<pre><code>model = Sequential()\n\nmodel.add(Conv2D(32, (3, 3), padding='same', input_shape=(32, 32, 3), activation=\"relu\"))\nmodel.add(Conv2D(32, (3, 3), activation=\"relu\"))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.5))\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(5, activation=\"softmax\"))\n\nmodel.compile(loss='categorical_crossentropy',optimizer='adam',metrics=['accuracy'])\nhistory=model.fit(x_train,y_train,batch_size=64,epochs=7,validation_data=(x_test, y_test),shuffle=True)\n<\/code><\/pre>\n\n<p>From this, I get the accuracy and loss<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/GiaWj.png\" rel=\"nofollow noreferrer\">Accuracy scale<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Z56zw.png\" rel=\"nofollow noreferrer\">Loss scale<\/a><\/p>\n\n<p>Then, I need more accuracy. So, I just increase the epochs=30 and the results are<\/p>\n\n<p>(FOR epochs = 30)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vVtpc.png\" rel=\"nofollow noreferrer\">Accuracy scale <\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/YVhfl.png\" rel=\"nofollow noreferrer\">Loss scale <\/a><\/p>\n\n<p>I understand that neural network set random weights so that we get different results. But how can I improve the accuracy. I am new to neural nets. Would greatly appreciate some explanation. Thank you<\/p>\n","answers":[{"AnswerId":"57787866","CreationDate":"2019-09-04T11:54:11.613","ParentId":null,"OwnerUserId":"11856284","Title":null,"Body":"<p>According to the training and validation metrics, the model seem to have over fitted. That is why training accuracy is increasing while validation accuracy seems constant. The reason for over-fitting is the lack of training samples.\nYou may have to increase the number of training samples. Try followings one by one.<\/p>\n\n<ul>\n<li>Reduce the dropout rate to ~0.2.<\/li>\n<li>Try reducing number of convolution layers<\/li>\n<\/ul>\n\n<p>Please update a link to data-set if you want us to try a model with that.<\/p>\n"}]}
+{"QuestionId":57787516,"AnswerCount":0,"Tags":"<tensorflow><keras><lstm><tensorflow2.0>","CreationDate":"2019-09-04T11:34:12.427","AcceptedAnswerId":null,"OwnerUserId":11698102.0,"Title":"LSTM Keras Internal Implementation","Body":"<p>I am implementing LSTM in TF2.0 from scratch. I am trying to have the same behaviour as in the <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/layers\/LSTM\" rel=\"nofollow noreferrer\">TF.keras.layers.LSTM <\/a>,<\/p>\n\n<p>Using the formulas from Wikipedia as below:<\/p>\n\n<p><img src=\"https:\/\/wikimedia.org\/api\/rest_v1\/media\/math\/render\/svg\/2db2cba6a0d878e13932fa27ce6f3fb71ad99cf1\" alt=\"here\"><\/p>\n\n<ul>\n<li><p>In my understanding, the input vector xt is the input in the current time step and <strong>NOT<\/strong> the concatenated version with [xt, ht-1] as per the below formula. <\/p><\/li>\n<li><p>Given the dimensions:\ndim(Wf) = dim(Wi) = dim(Wo) = dim(Wc) = <em>(n_input x n_hidden)<\/em><\/p><\/li>\n<\/ul>\n\n<p>dim(Uf) = dim(Ui) = dim(Uo) = dim(Uc) = <em>(n_hidden x n_hidden)<\/em><\/p>\n\n<p>The dimension of the input(xt) should be <em>(batch_size x n_hidden)<\/em> and output\/hidden (ht) dimension <em>(batch_size x n_input)<\/em><\/p>\n\n<p>Question : \nHow should the time steps be fed to such an implementation? \nI split the model input _X into multiple timesteps of dimension (batch_size x n_hidden).<\/p>\n\n<p>Is this right? With the implementation below, I am unable to get similar results like <em>tf.keras.layers.lstm(n_hidden)<\/em><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code> for inputs in _X:\n\n            input_gate = tf.sigmoid(tf.matmul(inputs, self.weights_input_gate) + tf.matmul(self.output_1, self.weights_input_hidden) + self.bias_input)\n\n            forget_gate = tf.sigmoid(tf.matmul(inputs, self.weights_forget_gate) + tf.matmul(self.output_1, self.weights_forget_hidden) + self.bias_forget)\n\n            output_gate = tf.sigmoid(tf.matmul(inputs, self.weights_output_gate) + tf.matmul(self.output_1, self.weights_output_hidden) + self.bias_output)\n\n            memory_cell = tf.tanh(\n                tf.matmul(inputs, self.weights_memory_cell) + tf.matmul(self.output_1, self.weights_memory_cell_hidden) + self.bias_memory_cell)\n\n            self.state = self.state * forget_gate + input_gate * memory_cell\n\n            self.output_1 = output_gate * tf.tanh(self.state)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57787706,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2019-09-04T11:44:58.610","AcceptedAnswerId":null,"OwnerUserId":11880284.0,"Title":"KeyError: u'BatchMatMulV2' while trying to generate .pb from .meta file in tensorflow 1.13.1","Body":"<p>Trying to generate a <code>.pb<\/code> file from <code>.meta<\/code> file but encountered <code>no op named batchmatmulv2 error<\/code> when running in <code>TensorFlow 1.13.1<\/code>.\nBut I need to generate the <code>.pb<\/code> file using TensorFlow 1.13.1 or below version due to environment restrictions<\/p>\n\n<p>I ran the same code on TensorFlow 1.14.0 and it was able to generate a pb file. <\/p>\n\n<pre><code>import tensorflow as tf\n\nmeta_path = 'fns.ckpt.meta' # Your .meta file\noutput_node_names = ['add_37']    # Output nodes\n\nwith tf.Session() as sess:\n\n    # Restore the graph\n    saver = tf.train.import_meta_graph(meta_path)\n\n    # Load weights\n    saver.restore(sess,\"checkpoints\/fns.ckpt\")\n\n    # Freeze the graph\n    frozen_graph_def = tf.graph_util.convert_variables_to_constants(\n        sess,\n        sess.graph_def,\n        output_node_names)\n\n    # Save the frozen graph\n    with open('output_graph.pb', 'wb') as f:\n      f.write(frozen_graph_def.SerializeToString())\n<\/code><\/pre>\n\n<blockquote>\n  <p>Traceback (most recent call last):<\/p>\n<\/blockquote>\n\n<pre><code> File \"&lt;ipython-input-3-e56112bc7d43&gt;\", line 4, in &lt;module&gt;\n    saver = tf.train.import_meta_graph(meta_path)\n\n  File \"\/home\/taif\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/training\/saver.py\", line 1435, in import_meta_graph\n    meta_graph_or_file, clear_devices, import_scope, **kwargs)[0]\n\n  File \"\/home\/taif\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/training\/saver.py\", line 1457, in _import_meta_graph_with_return_elements\n    **kwargs))\n\n  File \"\/home\/taif\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/meta_graph.py\", line 806, in import_scoped_meta_graph_with_return_elements\n    return_elements=return_elements)\n\n  File \"\/home\/taif\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/util\/deprecation.py\", line 507, in new_func\n    return func(*args, **kwargs)\n\n  File \"\/home\/taif\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/importer.py\", line 399, in import_graph_def\n    _RemoveDefaultAttrs(op_dict, producer_op_list, graph_def)\n\n  File \"\/home\/taif\/.local\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/importer.py\", line 159, in _RemoveDefaultAttrs\n    op_def = op_dict[node.op]\n\nKeyError: 'BatchMatMulV2'\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57787813,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><kaggle>","CreationDate":"2019-09-04T11:51:06.190","AcceptedAnswerId":null,"OwnerUserId":10911480.0,"Title":"How to use TensorFlow 2 GPU in a Kaggle kernel","Body":"<p>I would like to use TensorFlow 2.0 GPU in a Kaggle kernel.<\/p>\n\n<p>I have already activated the GPU on the Kaggle kernel settings and run those command lines:<\/p>\n\n<pre><code>!pip uninstall tensorflow-gpu -y\n!pip uninstall tensorflow -y\n!pip install tensorflow-gpu==2.0.0-rc0\n<\/code><\/pre>\n\n<p>When I run this command line <code>tf.__version__<\/code> I get <code>'2.0.0-rc0'<\/code>.\nAnd I run the following code to be sure I have access to a GPU device:<\/p>\n\n<pre><code>device_name = tf.test.gpu_device_name()\nif \"GPU\" not in device_name:\n    print(\"GPU device not found\")\nprint('Found GPU at: {}'.format(device_name))\n<\/code><\/pre>\n\n<p>\u2192 <code>Found GPU at: \/device:GPU:0<\/code><\/p>\n\n<p>Moreover when I run my model with CPU everything is fine.<\/p>\n\n<p>Here is the error I get:<\/p>\n\n<pre><code>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 conv_model\/conv1\/Conv2D (defined at \/opt\/conda\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/framework\/ops.py:1751) ]] [Op:__inference_train_step_862]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57788412,"AnswerCount":1,"Tags":"<python><pytorch><tensor>","CreationDate":"2019-09-04T12:28:52.857","AcceptedAnswerId":null,"OwnerUserId":1767774.0,"Title":"mask first k elements in a 3D tensor in PyTorch (different k for each row)","Body":"<p>I have a tensor <code>M<\/code> of dimensions <code>[NxQxD]<\/code> and a 1d tensor of indices <code>idx<\/code> (of size <code>N<\/code>). I want to efficiently create a tensor <code>mask<\/code> of dimensions <code>[NxQxD]<\/code> such that <code>mask[i,j,k] = 1 iff j &lt;= idx[i]<\/code>, i.e. I want to keep only the <code>idx[i]<\/code> first dimensions out of <code>Q<\/code> in the second dimension (dim=1) of <code>M<\/code>, for every row <code>i<\/code>. <\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"57790145","CreationDate":"2019-09-04T14:02:26.673","ParentId":null,"OwnerUserId":"1767774","Title":null,"Body":"<p>It turns out this can be done via a broadcasting trick:<\/p>\n\n<pre><code>mask_2d = torch.arange(Q)[None, :] &lt; idx[:, None] #(N,Q)\nmask_3d = mask[..., None] #(N,Q,1)\nmasked = mask.float() * data\n<\/code><\/pre>\n"}]}
+{"QuestionId":57788417,"AnswerCount":1,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2019-09-04T12:29:01.913","AcceptedAnswerId":57793594.0,"OwnerUserId":11208496.0,"Title":"Why tensors not connected to gradients in TensorBoard?","Body":"<p>For practice, I wanted to implement a model in tensorflow which gives me back the square of the input. My code works correctly, but when I have a look at the computation graph in TensorBoard, the LOSS operation is not connected to the Gradients subgraph and neither to Adam. Why is this? As I understand, the compute the gradients, tensorflow has to derivate the loss. <\/p>\n\n<p>Here is my code: <\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nnp_inp = np.array([3, 6, 4, 2, 9, 11, 0.48, 22, -2.3, -0.48])\nnp_outp = np.power(np_inp, 2)\n\ninputs = tf.Variable(np_inp, name='input', trainable=False)\noutputs = tf.Variable(np_outp, name='output', trainable=False)\n\nmultiplier = tf.Variable(0.1,                          \n                             dtype=tf.float64, trainable=True, name='multiplier')\n\nmul = inputs * multiplier\npredict = tf.square(mul, name='prediction')\n\nloss = tf.math.reduce_sum(tf.math.square(predict-outputs), name='LOSS')\noptimizer = tf.train.AdamOptimizer(0.1)\nto_minimize = optimizer.minimize(loss)\n\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\n\nlogs_path = \".\/logs\/unt\"  # path to the folder that we want to save the logs for Tensorboard\ntrain_writer = tf.summary.FileWriter(logs_path, sess.graph)\n\nfor i in range(100):\n  sess.run(to_minimize)\nprint(sess.run({'mult':multiplier}))\n<\/code><\/pre>\n\n<p>Tensorboard:\n<a href=\"https:\/\/gofile.io\/?c=jxbWiG\" rel=\"nofollow noreferrer\">https:\/\/gofile.io\/?c=jxbWiG<\/a><\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"57793594","CreationDate":"2019-09-04T17:55:10.193","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>This can be counter intuitive, but the actual value of the loss is not used for the training itself (although it can be useful to plot it to see its progress). What optimizers generally use is the gradient, that is, how each change in each variable would affect the loss value. To compute this, a tensor with the same shape as <code>LOSS<\/code> but filled with ones is created, and the gradient of each operation is computed through back-propagation. If you open the <code>gradients<\/code> box in the graph, you will see a <code>LOSS_grad<\/code> box representing this.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/WuDoK.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/WuDoK.png\" alt=\"LOSS_grad\"><\/a><\/p>\n\n<p>It is a couple of nodes making that tensor of ones, because the gradient of something with respect to itself is always one. From there, the rest of gradients are computed.<\/p>\n"}]}
+{"QuestionId":57789168,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-09-04T13:11:09.010","AcceptedAnswerId":null,"OwnerUserId":10033794.0,"Title":"How to fix inconsistency in last layer shape \"Error when checking target: expected dense_ to have shape xx but got array with shape yy\"","Body":"<p>I am a newbie to deep learning and TensorFlow and still have problem understanding the relationship between layers and their output. I am working on a binary classification for time series and get the following error. I wonder what is the reason for getting this error and what items should I check in my architecture to avoid getting this:<\/p>\n\n<pre><code>Error when checking target: expected dense_68 to have shape (1,) but got array with shape (24,)\n<\/code><\/pre>\n\n<p>I have a simple dataset with these sizes:<\/p>\n\n<pre><code>xTrain (32,24) # 32 training observations with 24 features\nyTrain (32,1) # 32 training binary label (0 or 1)\nxTest (16,24) \nyTest (16,1)\n<\/code><\/pre>\n\n<p>My architecture looks like this:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(100, activation='relu', input_dim=24))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\nprint(model.summary())\nmodel.fit(xTrain, yTrain, validation_data=(xTest, yTest), epochs=10, batch_size=2)\n<\/code><\/pre>\n\n<pre><code>\nLayer (type)                 Output Shape              Param #   \n=================================================================\ndense_1 (Dense)              (None, 100)               2500      \n_________________________________________________________________\ndense_2 (Dense)              (None, 1)                 101       \n=================================================================\nTotal params: 2,601\nTrainable params: 2,601\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>Since the size of <code>yTrain<\/code> and <code>yTest<\/code> is 1, and in the last dense layer, I set the unit size as 1, I wonder why I get the error that I get array with shape (24,), which is equal to my feature size as error.<\/p>\n","answers":[{"AnswerId":"57790128","CreationDate":"2019-09-04T14:01:23.800","ParentId":null,"OwnerUserId":"5432214","Title":null,"Body":"<p>I copy-pasted your code and created the datasets with the same shapes. The model correctly starts training so I think you just have a typo\/bug somewhere.<\/p>\n\n<p>My code is this: <\/p>\n\n<pre><code>import numpy as np\nfrom keras import Sequential\nfrom keras.layers import Dense\n\nxTrain = np.zeros((32,24)) # 32 training observations with 24 features\nyTrain = np.zeros((32,1)) # 32 training binary label (0 or 1)\nxTest = np.zeros((16,24))\nyTest = np.zeros((16,1))\n\nmodel = Sequential()\nmodel.add(Dense(100, activation='relu', input_dim=24))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\nprint(model.summary())\nmodel.fit(xTrain, yTrain, validation_data=(xTest, yTest), epochs=10, batch_size=2)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57789202,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning><shap>","CreationDate":"2019-09-04T13:13:05.747","AcceptedAnswerId":null,"OwnerUserId":2719362.0,"Title":"GradientExplainer with multi-input CNN model","Body":"<p>I'm trying to use the <a href=\"https:\/\/github.com\/slundberg\/shap\" rel=\"nofollow noreferrer\">SHAP library<\/a> to see what parts of my images are important in a convolutional neural network that I trained with Keras (backend=tensorflow). Specifically, I'm using the <strong>GradientExplainer<\/strong>. I've been able to follow the linked example on a one-input CNN that I trained, but I cannot get it to work with a more complex CNN.<\/p>\n\n<p>My CNN has three inputs, as the 0th, 1st and (-10)th layers (the last is scalar data). So I've modified map2layer thusly:<\/p>\n\n<pre><code>def map2layer(i1,i2,i3,layer):\n    feed_dict = dict(zip([model.layers[1].input,model.layers[0].input,model.layers[-10].input], [i1,i2,i3]))\n    for k in feed_dict.keys():\n      print(k,feed_dict[k].shape,feed_dict[k].dtype)\n    return K.get_session().run(model.layers[layer].input, feed_dict)\n<\/code><\/pre>\n\n<p>I've confirmed that the inputs corresond correctly to the layers' shapes and dtypes (see <code>for<\/code> loop in <code>map2layer<\/code>). Also note that the 'background' data sample is 1000 here. <\/p>\n\n<pre><code>Tensor(\"input_1:0\", shape=(?, 128, 128, 1), dtype=float32) (1000, 128, 128, 1) float32\nTensor(\"input_2:0\", shape=(?, 32, 32, 1), dtype=float32) (1000, 32, 32, 1) float32\nTensor(\"input_3:0\", shape=(?, 4), dtype=float32) (1000, 4) float32\n<\/code><\/pre>\n\n<p>One oddity here is that my <code>model.layers[0].input<\/code> is labeled <code>input_2<\/code> and <code>model.layers[1].input<\/code> is labeled <code>input_1<\/code> in my Keras model summary. I suspect this is perhaps due to a subsequent <code>up_sampling2d<\/code> layer. This may have no bearing on my issue, but wanted to mention it. <\/p>\n\n<p>I get an error on this line,<\/p>\n\n<pre><code>shap_values,indexes=e.shap_values(map2layer(i1_norm,i2_norm,i3_norm,layer) ranked_outputs=1)\n<\/code><\/pre>\n\n<p>Which gives me this message:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'input_3' with dtype float and shape [?,4]\n<\/code><\/pre>\n\n<p><code>i1_norm<\/code>, <code>i2_norm<\/code>, and <code>i3_norm<\/code> are normalized tensors with shapes of <code>(3, 128, 128, 1)<\/code>, <code>(3, 32, 32, 1)<\/code>, and <code>(3, 4)<\/code>, respectively. So I want SHAP images of 3 samples.<\/p>\n\n<p>It appears to me that I am giving it a tensor for <code>input_3<\/code> and it should just use that as a \"placeholder\". Admittedly, I only use Keras and not pure Tensorflow, so I'm not quite sure what it's expecting or how to provide it. Any solutions here are much appreciated!<\/p>\n","answers":[]}
+{"QuestionId":57789645,"AnswerCount":0,"Tags":"<deep-learning><pytorch>","CreationDate":"2019-09-04T13:37:17.070","AcceptedAnswerId":null,"OwnerUserId":9625519.0,"Title":"torch.utils.data.random_split() is not splitting the data","Body":"<p>I\u2019m not getting split when I use <code>torch.utils.data.random_split<\/code>.<\/p>\n\n<p>I get correct numbers for <code>train_size<\/code> and <code>val_size<\/code>, but when I do <code>random_split<\/code>, both <code>train_data<\/code> and <code>val_data<\/code> get <code>full_data<\/code>. There is no split happening.<\/p>\n\n<p>Please help me with this issue.<\/p>\n\n<pre><code>class DeviceLoader(Dataset):\n\ndef __init__(self, root_dir, train=True, transform=None):\n    self.file_path = root_dir\n    self.train = train\n    self.transform = transform\n    self.file_names = ['%s\/%s'%(root,file) for root,_,files in os.walk(root_dir) for file in files]\n    self.len = len(self.file_names)\n    self.labels = {'BP_Raw_Images':0, 'DT_Raw_Images':1, 'GL_Raw_Images':2, 'PO_Raw_Images':3, 'WS_Raw_Images':4}\n\ndef __len__(self):\n    return(len(self.file_names))\n\ndef __getitem__(self, idx):\n    file_name = self.file_names[idx]\n    device = file_name.split('\/')[5]\n    img = self.pil_loader(file_name)\n    if(self.transform):\n        img = self.transform(img)\n    cat = self.labels[device]            \n    if(self.train):\n        return(img, cat)\n    else:\n        return(img, file_name)\nfull_data = DeviceLoader(root_dir=\u2019\/kaggle\/input\/devices\/dataset\/\u2019, transform=transforms, train=True)\ntrain_size = int(0.7*len(full_data))\nval_size = len(full_data) - train_size\ntrain_data, val_data = torch.utils.data.random_split(full_data,[train_size,val_size])\n<\/code><\/pre>\n\n<p>Expected result is to split the <code>full_data<\/code> into <code>train_data<\/code>(2000) and <code>val_data<\/code>(500). But instead, I get <code>full_data<\/code>(2500) in both train and val.<\/p>\n","answers":[]}
+{"QuestionId":57789720,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-04T13:41:16.320","AcceptedAnswerId":null,"OwnerUserId":9318372.0,"Title":"TensorFlow get updates computed by optimizer","Body":"<p>In <code>tf<\/code>, the optimizer class only has two function:<\/p>\n\n<ul>\n<li><code>compute_gradients<\/code><\/li>\n<li><code>apply_gradients<\/code><\/li>\n<\/ul>\n\n<p>where <code>apply_gradients<\/code> returns an op that performs the update <code>w &lt;- w + \u0394w<\/code> via a <code>tf.assign_add<\/code> function.<\/p>\n\n<p>However I need direct access to the <code>\u0394w<\/code> itself. (or <code>w' = w+\u0394w<\/code>). I know that the optimizer adds nodes to the computational graph which compute this <code>\u0394w<\/code> for each variable. How can I access them? Or do I have to re-implement the optimizer myself?<\/p>\n\n<p>The reason is that I need to compute gradients <code>dw'\/dw<\/code>, as I am working on something related to gradient based hyperparameter optimization (cf. <a href=\"https:\/\/arxiv.org\/abs\/1703.01785\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1703.01785<\/a>)<\/p>\n","answers":[{"AnswerId":"57790935","CreationDate":"2019-09-04T14:48:53.370","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>The \"delta\" applied to each variable is not accessible through any common method or name. In fact, looking a bit into the source it seems rather difficult extract, as it varies from one optimizer to the other.<\/p>\n\n<p>What you can do, at least, is to compute the differences between variable values and their updates. For example it could work like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\nwith tf.Graph().as_default():\n    # Setup example model\n    x = tf.placeholder(tf.float32, [None, 1])\n    y = tf.placeholder(tf.float32, [None, 2])\n    w = tf.Variable([[1., 2.]], tf.float32)\n    pred = x @ w\n    loss = (tf.reduce_sum(tf.squared_difference(pred, y))\n            \/ tf.cast(tf.shape(x)[0], tf.float32))\n    # Record variable values before training step\n    # (tf.identity should work here but it does not so we use\n    # a trivial add operation to enforce the control dependency)\n    w_old = w + 0\n    # Train after having recorded variable values\n    with tf.control_dependencies([w_old]):\n        train_op = tf.train.GradientDescentOptimizer(0.1).minimize(loss)\n    # Compute deltas after training\n    with tf.control_dependencies([train_op]):\n        w_delta = w.read_value() - w_old\n    init_op = tf.global_variables_initializer()\n    # Test\n    with tf.Session() as sess:\n        sess.run(init_op)\n        print(sess.run(w))\n        # [[1. 2.]]\n        _, w_delta_val = sess.run(\n            [train_op, w_delta],\n            feed_dict={x: [[1.], [2.]], y: [[3., 4], [5., 6.]]})\n        print(w_delta_val)\n        # [[0.79999995 0.5999999 ]]\n        print(sess.run(w))\n        # [[1.8 2.6]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57790117,"AnswerCount":0,"Tags":"<amazon-web-services><tensorflow><aws-lambda><serverless><aws-lambda-layers>","CreationDate":"2019-09-04T14:00:50.333","AcceptedAnswerId":null,"OwnerUserId":11943909.0,"Title":"Deploying Tensroflow Library as AWS Lambda Layer","Body":"<p>I am trying to deploy <strong>Tensorflow<\/strong> (specifically: <code>libtensorflow.so.1, libtensorflow.so.1.14.0, libtensorflow_framework.so.1, libtensorflow_framework.so.1.14.0<\/code>) as <strong>AWS Lambda Layer<\/strong>. I am getting following error:<\/p>\n\n<pre><code>An error occurred: TensorflowLambdaLayer - Unzipped size must be smaller than 262144000 bytes (Service: AWSLambdaInternal; Status Code: 400; Error Code: InvalidParameterValueException; Request ID: d8668bfa-a61c-473c-aeb0-62bff3aa852f).\n<\/code><\/pre>\n\n<p>How can I change these .so files by deleting some of their content that is irrelevant for my Lambda function which utilizes Tensorflow Lambda Layer?<\/p>\n","answers":[]}
+{"QuestionId":57790722,"AnswerCount":1,"Tags":"<tensorflow><rasa>","CreationDate":"2019-09-04T14:36:22.487","AcceptedAnswerId":null,"OwnerUserId":12020314.0,"Title":"Tensorflow Built from Source CustomBuilt dont Work","Body":"<p>I tried to install Rasa on a server with an Intel Xeon E5620.\nRasa needs Tensorflow 14.0.0. I use an virtual Environment (Anaconda)<\/p>\n\n<p>I have to compile Tensorflow from source, because the CPU doesn't support AVX.\nI created a custom Built with <a href=\"https:\/\/github.com\/hadim\/docker-tensorflow-builder\" rel=\"nofollow noreferrer\">docker-tensorflow-builder<\/a> and install it.\nWhen I tested it, I got the message:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/anaconda3\/envs\/ki\/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 \"\/anaconda3\/envs\/ki\/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 \"\/anaconda3\/envs\/ki\/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 \"\/anaconda3\/envs\/ki\/lib\/python3.6\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/anaconda3\/envs\/ki\/lib\/python3.6\/imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: \/lib\/x86_64-linux-gnu\/libc.so.6: version `GLIBC_2.28' not found (required by \/anaconda3\/envs\/ki\/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;string&gt;\", line 1, in &lt;module&gt;\n  File \"\/anaconda3\/envs\/ki\/lib\/python3.6\/site-packages\/tensorflow\/__init__.py\", line 28, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"\/anaconda3\/envs\/ki\/lib\/python3.6\/site-packages\/tensorflow\/python\/__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"\/anaconda3\/envs\/ki\/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 \"\/anaconda3\/envs\/ki\/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 \"\/anaconda3\/envs\/ki\/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 \"\/anaconda3\/envs\/ki\/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 \"\/anaconda3\/envs\/ki\/lib\/python3.6\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/anaconda3\/envs\/ki\/lib\/python3.6\/imp.py\", line 343, in load_dynamic\n    return _load(spec)\nImportError: \/lib\/x86_64-linux-gnu\/libc.so.6: version `GLIBC_2.28' not found (required by \/anaconda3\/envs\/ki\/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\/errors\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 fix it by install <a href=\"https:\/\/anaconda.org\/neok.m4700\/glibc\" rel=\"nofollow noreferrer\">GlibC.2.30<\/a>.\nNow i get the Error \"Segmentation fault (core dumped) error\".\nI think I get the error because Tensorflow is not adapted to my CPU.\nI'm not sure how to find out the appropriate compiler options and finally set them. Has anyone here experience and can help me?<\/p>\n","answers":[{"AnswerId":"57968562","CreationDate":"2019-09-17T06:33:27.797","ParentId":null,"OwnerUserId":"10685378","Title":null,"Body":"<p>Here the problem is with gcc version. Please use gcc 6.4.0 for the installation of tensorflow.<\/p>\n\n<p>Below are the steps to install Intel optimized tensorflow in conda environment.<\/p>\n\n<pre><code>conda create -n tf_1.10 -c intel python=3.6\n\nsource activate tf_1.10\n<\/code><\/pre>\n\n<p>Make sure the following lines are there in the bash_profile and source it.<\/p>\n\n<pre><code>export CC=\/glob\/development-tools\/versions\/gcc-6.4.0\/bin\/gcc\n\nexport LD_LIBRARY_PATH=\/glob\/development-tools\/versions\/gcc-6.4.0\/lib64\/:$LD_LIBRARY_PATH\n\nexport PATH=\/glob\/development-tools\/versions\/gcc-6.4.0\/bin\/:$PATH\n<\/code><\/pre>\n\n<p>If the path setting was successful, you should be able to see the gcc version 6.4.0 on typing gcc --version<\/p>\n\n<p>If you get the error ImportError: \/lib\/x86_64-linux-gnu\/libc.so.6: version `GLIBC_2.28' not found, please follow the steps below:<\/p>\n\n<pre><code>mkdir ~\/lib\n\ncd ~\/lib\n\nln -s \/glob\/supplementary-software\/versions\/glibc\/glibc_2_28\/lib\/libm.so.6\n<\/code><\/pre>\n\n<p>Then export the LD_LIBRARY_PATH for it: <\/p>\n\n<pre><code>export LD_LIBRARY_PATH=~\/lib:$LD_LIBRARY_PATH\n<\/code><\/pre>\n\n<p>Then, install tensorflow 1.10 using: <\/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"}]}
+{"QuestionId":57791108,"AnswerCount":1,"Tags":"<tensorflow><google-colaboratory>","CreationDate":"2019-09-04T14:59:36.537","AcceptedAnswerId":null,"OwnerUserId":9770744.0,"Title":"How to downgrade Cuda and cuDNN Version in Google Colab?","Body":"<p>I require to run tensorflow-gpu of version 1.3.0. For that, I need to downgrade cuda to version 8. Can someone please share the code to downgrade cuda in google colab from 10.0 to 8.0.<\/p>\n\n<p>I got the code for downgrading to version 9 using this.<\/p>\n\n<pre><code>!wget https:\/\/developer.nvidia.com\/compute\/cuda\/9.0\/Prod\/local_installers\/cuda-repo-ubuntu1604-9-0-local_9.0.176-1_amd64-deb\n!dpkg -i cuda-repo-ubuntu1604-9-0-local_9.0.176-1_amd64-deb\n!apt-key add \/var\/cuda-repo-9-0-local\/7fa2af80.pub\n!apt-get update\n!apt-get install cuda=9.0.176-1\n<\/code><\/pre>\n\n<p>I tried to change 9 in the code to 8 but that didn't work.<\/p>\n","answers":[{"AnswerId":"57793556","CreationDate":"2019-09-04T17:52:43.243","ParentId":null,"OwnerUserId":"9770744","Title":null,"Body":"<p>I got the answer for downgrading cuda to version 8.0 using the following command line.<\/p>\n\n<pre><code>!wget https:\/\/developer.nvidia.com\/compute\/cuda\/8.0\/Prod2\/local_installers\/cuda-repo-ubuntu1604-8-0-local-ga2_8.0.61-1_amd64-deb\n!dpkg -i cuda-repo-ubuntu1604-8-0-local-ga2_8.0.61-1_amd64-deb\n!apt-get update\n!apt-get install cuda=8.0.61-1\n<\/code><\/pre>\n\n<p>But another requirement is to downgrade the cuDNN version to 6.0. Can someone please give me the set of codes to downgrade cuDNN version to 6.0 in Google colab.<\/p>\n"}]}
+{"QuestionId":57791153,"AnswerCount":0,"Tags":"<spring><caching><caffeine>","CreationDate":"2019-09-04T15:02:12.283","AcceptedAnswerId":null,"OwnerUserId":11337008.0,"Title":"How do we control cache error with spring cafeine cache?","Body":"<p>I use the <strong>cache caffeine spring<\/strong> for the cache of my web service. so when my webservice returns error 401, this error stay in the cache, so the next call to this web service automatically returns error 401.\nI wanted to know if the default configuration of  the cache caffeine cache errors?  and if the attribute (unless=\"# result.error\") can solve this problem ?<\/p>\n\n<pre><code>@Cacheable(value = \"testCache\", unless=\"# result.error\")\n    public Response getWebService(){\n\n    }\n}\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57791185,"AnswerCount":1,"Tags":"<python><tensorflow><keras><conv-neural-network>","CreationDate":"2019-09-04T15:03:48.807","AcceptedAnswerId":57801289.0,"OwnerUserId":12020433.0,"Title":"ValueError: Error when checking target: expected conv2d_19 to have shape (None, 320, 320, 1) but got array with shape (18, 320, 320, 2)","Body":"<p>I try to train a U-Net model in keras, with regular segmentation (not multui-classe), but I get the following error:<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking target: expected conv2d_19 to have shape (None, 320, 320, 1) but got array with shape (18, 320, 320, 2)<\/p>\n<\/blockquote>\n\n<p>I know this is related to the input Y_train, and I to the         keras.utils.to_categorical function, but to_categorical(Y_train) and to_categorical(Y_train, 2) returns this error, while to_categorical(Y_train, 1) returns IndexError: index 1 is out of bounds for axis 1 with size 1. <\/p>\n\n<p>I even tried to change Y_train by reshaping it into (None, 320, 320, 1) but it returned<\/p>\n\n<blockquote>\n  <p>'ValueError: You are passing a target array of shape (18, 320, 320, 1) while using as loss categorical_crossentropy. categorical_crossentropy expects targets to be binary matrices (1s and 0s) of shape (samples, classes). If your targets are integer classes, you can convert them to the expected format via:\n  from keras.utils import to_categorical\n  y_binary = to_categorical(y_int)\n  Alternatively, you can use the loss function <code>sparse_categorical_crossentropy<\/code> instead, which does expect integer targets.'<\/p>\n<\/blockquote>\n\n<p>The error is returned when I try to fit the model, using the training function.<\/p>\n\n<pre><code>    def preprocess_images(X_train, Y_train, X_test, Y_test):\n\n        X_train = X_train.reshape(X_train.shape[0], X_train.shape[1], X_train.shape[2], 1)\n        X_test = X_test.reshape(X_test.shape[0], X_train.shape[1], X_train.shape[2], 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\n        Y_train \/= 255\n        Y_test \/= 255\n        #Y_train = Y_train.reshape(Y_train.shape[0], Y_train.shape[1], Y_train.shape[2], 1)\n        #Y_test = Y_test.reshape(Y_test.shape[0], Y_train.shape[1], Y_train.shape[2], 1)\n        Y_train = keras.utils.to_categorical(Y_train, 1)\n        Y_test = keras.utils.to_categorical(Y_test, 1)\n\n        return ((X_train, Y_train), (X_test, Y_test))\n\n\n    def CNNs_layers(nb_conv_pool = 1, filters = 32, kernel_size = (3, 3), activation = 'relu', pool_size = (2 ,2), dropout_rate = 0.25, input_shape = (28, 28, 1), data_format = 'channels_last'):\n\n\n        inputs = tf.keras.layers.Input((IMG_HEIGHT, IMG_WIDTH, IMG_CHANNELS))\n        s = tf.keras.layers.Lambda(lambda x: x \/ 255)(inputs)\n\n        c1 = tf.keras.layers.Conv2D(16, (3, 3), activation=tf.keras.activations.elu, kernel_initializer='he_normal',\n                                padding='same')(s)\n        c1 = tf.keras.layers.Dropout(0.1)(c1)\n        c1 = tf.keras.layers.Conv2D(16, (3, 3), activation=tf.keras.activations.elu, kernel_initializer='he_normal',\n                                padding='same')(c1)\n        p1 = tf.keras.layers.MaxPooling2D((2, 2))(c1)\n\n        c2 = tf.keras.layers.Conv2D(32, (3, 3), activation=tf.keras.activations.elu, kernel_initializer='he_normal',\n                                padding='same')(p1)\n        c2 = tf.keras.layers.Dropout(0.1)(c2)\n        c2 = tf.keras.layers.Conv2D(32, (3, 3), activation=tf.keras.activations.elu, kernel_initializer='he_normal',\n                                padding='same')(c2)\n        p2 = tf.keras.layers.MaxPooling2D((2, 2))(c2)\n\n        c3 = tf.keras.layers.Conv2D(64, (3, 3), activation=tf.keras.activations.elu, kernel_initializer='he_normal',\n                                padding='same')(p2)\n        c3 = tf.keras.layers.Dropout(0.2)(c3)\n        c3 = tf.keras.layers.Conv2D(64, (3, 3), activation=tf.keras.activations.elu, kernel_initializer='he_normal',\n                                padding='same')(c3)\n\n        p3 = tf.keras.layers.MaxPooling2D((2, 2))(c3)\n\n        c4 = tf.keras.layers.Conv2D(128, (3, 3), activation=tf.keras.activations.elu, kernel_initializer='he_normal',\n                                padding='same')(p3)\n        c4 = tf.keras.layers.Dropout(0.2)(c4)\n        c4 = tf.keras.layers.Conv2D(128, (3, 3), activation=tf.keras.activations.elu, kernel_initializer='he_normal',\n                                padding='same')(c4)\n        p4 = tf.keras.layers.MaxPooling2D(pool_size=(2, 2))(c4)\n\n        c5 = tf.keras.layers.Conv2D(256, (3, 3), activation=tf.keras.activations.elu, kernel_initializer='he_normal',\n                                padding='same')(p4)\n        c5 = tf.keras.layers.Dropout(0.3)(c5)\n        c5 = tf.keras.layers.Conv2D(256, (3, 3), activation=tf.keras.activations.elu, kernel_initializer='he_normal',\n                                padding='same')(c5)\n\n        u6 = tf.keras.layers.Conv2DTranspose(128, (2, 2), strides=(2, 2), padding='same')(c5)\n        u6 = tf.keras.layers.concatenate([u6, c4])\n        c6 = tf.keras.layers.Conv2D(128, (3, 3), activation=tf.keras.activations.elu, kernel_initializer='he_normal',\n                                padding='same')(u6)\n        c6 = tf.keras.layers.Dropout(0.2)(c6)\n        c6 = tf.keras.layers.Conv2D(128, (3, 3), activation=tf.keras.activations.elu, kernel_initializer='he_normal',\n                                padding='same')(c6)\n\n        u7 = tf.keras.layers.Conv2DTranspose(64, (2, 2), strides=(2, 2), padding='same')(c6)\n        u7 = tf.keras.layers.concatenate([u7, c3])\n        c7 = tf.keras.layers.Conv2D(64, (3, 3), activation=tf.keras.activations.elu, kernel_initializer='he_normal',\n                                padding='same')(u7)\n        c7 = tf.keras.layers.Dropout(0.2)(c7)\n        c7 = tf.keras.layers.Conv2D(64, (3, 3), activation=tf.keras.activations.elu, kernel_initializer='he_normal',\n                                padding='same')(c7)\n\n        u8 = tf.keras.layers.Conv2DTranspose(32, (2, 2), strides=(2, 2), padding='same')(c7)\n        u8 = tf.keras.layers.concatenate([u8, c2])\n        c8 = tf.keras.layers.Conv2D(32, (3, 3), activation=tf.keras.activations.elu, kernel_initializer='he_normal',\n                                padding='same')(u8)\n        c8 = tf.keras.layers.Dropout(0.1)(c8)\n        c8 = tf.keras.layers.Conv2D(32, (3, 3), activation=tf.keras.activations.elu, kernel_initializer='he_normal',\n                                padding='same')(c8)\n\n        u9 = tf.keras.layers.Conv2DTranspose(16, (2, 2), strides=(2, 2), padding='same')(c8)\n        u9 = tf.keras.layers.concatenate([u9, c1], axis=3)\n        c9 = tf.keras.layers.Conv2D(16, (3, 3), activation=tf.keras.activations.elu, kernel_initializer='he_normal',\n                                padding='same')(u9)\n        c9 = tf.keras.layers.Dropout(0.1)(c9)\n        c9 = tf.keras.layers.Conv2D(16, (3, 3), activation=tf.keras.activations.elu, kernel_initializer='he_normal',\n                                padding='same')(c9)\n\n        outputs = tf.keras.layers.Conv2D(1, (1, 1), activation='sigmoid')(c9)\n\n        model = tf.keras.Model(inputs=[inputs], outputs=[outputs])\n         model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])\n        model.summary()\n\n        return (model)\n\n\n    def training(model, X_train, Y_train, loss = 'categorical_crossentropy', optimizer='adam', metrics=['accuracy'], batch_size = 32, nb_epoch = 100, verbose = 2, validation_split = 0.3):\n        model.compile(loss = loss, optimizer = optimizer, metrics = metrics)\n        history = model.fit(X_train, Y_train, batch_size = batch_size, epochs = nb_epoch, verbose = verbose, validation_split = validation_split)\n        return (history)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57801289","CreationDate":"2019-09-05T08:22:17.217","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Assuming your labels are one-hot encoded, then you should change the output layer to:<\/p>\n\n<pre><code>outputs = tf.keras.layers.Conv2D(2, (1, 1), activation='softmax')(c9)\n<\/code><\/pre>\n\n<p>And the loss function to <code>categorical_crossentropy<\/code>. If you want to use your current model and loss with no modifications, then you must <strong>not<\/strong> one-hot encode your labels, as <code>binary_crossentropy<\/code> takes 0-1 binary labels directly without one-hot encoding.<\/p>\n"}]}
+{"QuestionId":57791523,"AnswerCount":0,"Tags":"<pytorch><tf.keras><opticalflow><transfer-learning>","CreationDate":"2019-09-04T15:25:09.990","AcceptedAnswerId":null,"OwnerUserId":12020570.0,"Title":"Output difference by transferring weights(PWC-Net) from pytorch model to tf.keras model","Body":"<p>I implement a pwc net in tf keras with reference of this repos- <pre><a href=\"https:\/\/github.com\/sniklaus\/pytorch-pwc\" rel=\"nofollow noreferrer\">https:\/\/github.com\/sniklaus\/pytorch-pwc<\/a><\/pre>. The weights i uses to initialze the weights of conv layers of pwc net are the same as the weights of pytorch model in repos. I check my keras model many times, im sure that my model are the same as the pytorchmodel. So i would like to , is there a weight transferring problem between pytorch and tf keras 2.0?<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/t93cx.png\" alt=\"the result is\"><\/p>\n","answers":[]}
+{"QuestionId":57791826,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><tensorflow2.0>","CreationDate":"2019-09-04T15:43:24.267","AcceptedAnswerId":null,"OwnerUserId":8382950.0,"Title":"Tensorflow Transformer Colab: AttributeError: module 'tensorflow_core' has no attribute 'compiler'","Body":"<p>I tried to import the packages in the Transformer code in <a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/docs\/blob\/r2.0rc\/site\/en\/r2\/tutorials\/text\/transformer.ipynb\" rel=\"nofollow noreferrer\">Transformer Colab Notebook<\/a>. Is there a updated tensorflow core version?<\/p>\n\n<pre><code> from __future__ import absolute_import, division, print_function, unicode_literals\n\ntry:\n  !pip install tf-nightly-2.0-preview\nexcept Exception:\n  pass\nimport tensorflow_datasets as tfds\nimport tensorflow as tf\n\nimport time\nimport numpy as np\nimport matplotlib.pyplot as plt\n<\/code><\/pre>\n\n<p>I got the following error.<\/p>\n\n<pre><code> ---------------------------------------------------------------------------\nAttributeError                            Traceback (most recent call last)\n&lt;ipython-input-9-22c51fa8bb84&gt; in &lt;module&gt;()\n      5 except Exception:\n      6   pass\n----&gt; 7 import tensorflow_datasets as tfds\n      8 import tensorflow as tf\n      9 \n\n3 frames\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/__init__.py in &lt;module&gt;()\n     96 \n     97 # We still need all the names that are toplevel on tensorflow_core\n---&gt; 98 from tensorflow_core import *\n     99 \n    100 # These should not be visible in the main tf module.\n\nAttributeError: module 'tensorflow_core' has no attribute 'compiler'\n<\/code><\/pre>\n","answers":[{"AnswerId":"57991916","CreationDate":"2019-09-18T11:39:25.353","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>The error is from not having tensorflow datasets installed. Use <code>pip3 install tensorflow-datasets<\/code> for installing and run the script again. Here is a useful TF-datasets offical <a href=\"https:\/\/www.tensorflow.org\/datasets\/overview\" rel=\"nofollow noreferrer\">link<\/a>.<\/p>\n"}]}
+{"QuestionId":57791851,"AnswerCount":1,"Tags":"<python><tensorflow><casting>","CreationDate":"2019-09-04T15:44:30.107","AcceptedAnswerId":57794155.0,"OwnerUserId":6621346.0,"Title":"cast tensorflow 2.0 BatchDataset to numpy array","Body":"<p>I have this code:<\/p>\n\n<pre><code>(train_images, _), (test_images, _) = tf.keras.datasets.mnist.load_data()\n\ntrain_dataset = tf.data.Dataset.from_tensor_slices(train_images).shuffle(TRAIN_BUF).batch(BATCH_SIZE)\ntest_dataset = tf.data.Dataset.from_tensor_slices(test_images).shuffle(TRAIN_BUF).batch(BATCH_SIZE)\n\nprint(train_dataset, type(train_dataset), test_dataset, type(test_dataset))\n<\/code><\/pre>\n\n<p>And I want to cast these two <code>BatchDataset<\/code> variables to <code>numpy arrays<\/code>, can I do it easily? I am using <code>TF 2.0<\/code>, but I just found code to cast <code>tf.data<\/code> with <code>TF 1.0<\/code><\/p>\n","answers":[{"AnswerId":"57794155","CreationDate":"2019-09-04T18:42:24.417","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>After batching of dataset, the shape of last batch may not be same with that of rest of the batches. For example, if there are totally 100 elements in your dataset and you batch with size of 6, the last batch will have size of only 4. (100 = 6 * 16 + 4).<\/p>\n\n<p>So, in such cases, you will not be able to transform your dataset into numpy straight forward. For that reason, you will have to use <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/data\/Dataset#batch\" rel=\"nofollow noreferrer\"><code>drop_remainder<\/code><\/a> parameter to True in batch method. It will drop the last batch if it is not correctly sized. <\/p>\n\n<p>After that, I have enclosed the code on how to convert dataset to Numpy.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n(train_images, _), (test_images, _) = tf.keras.datasets.mnist.load_data()\n\nTRAIN_BUF=1000\nBATCH_SIZE=64\n\ntrain_dataset = tf.data.Dataset.from_tensor_slices(train_images).\n                          shuffle(TRAIN_BUF).batch(BATCH_SIZE, drop_remainder=True)\ntest_dataset = tf.data.Dataset.from_tensor_slices(test_images).\n                          shuffle(TRAIN_BUF).batch(BATCH_SIZE, drop_remainder=True)\n\n# print(train_dataset, type(train_dataset), test_dataset, type(test_dataset))\n\ntrain_np = np.stack(list(train_dataset))\ntest_np = np.stack(list(test_dataset))\nprint(type(train_np), train_np.shape)\nprint(type(test_np), test_np.shape)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>&lt;class 'numpy.ndarray'&gt; (937, 64, 28, 28)\n&lt;class 'numpy.ndarray'&gt; (156, 64, 28, 28)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57792085,"AnswerCount":1,"Tags":"<python><tensorflow><image-processing><yolo>","CreationDate":"2019-09-04T15:58:56.463","AcceptedAnswerId":57880895.0,"OwnerUserId":2737131.0,"Title":"YOLO : Either overfits or underfits, increase batch or increase sample image pool?","Body":"<p>I'm trying to train my yolo model to identify fire extinguishers and to label it as \"Fire Safety\". Currently is either I get a overfit or underfit images(see below). <\/p>\n\n<p>My sample images size with annotations is around ~1500<\/p>\n\n<p>yolo-new.cfg config of width=608 and height=608<\/p>\n\n<p>And I have trained using the following command:<\/p>\n\n<blockquote>\n  <p>python flow --model cfg\/yolo-new.cfg --labels one_label.txt  --train\n  --trainer adam --dataset \"C:\/\/Users\/\/G\/\/Desktop\/\/Development\/\/ML\/\/YOLO\/\/BBox-Label-Tool\/\/Images\/\/002\"\n  --annotation \"C:\/\/Users\/\/G\/\/Desktop\/\/Development\/\/ML\/\/YOLO\/\/BBox-Label-Tool\/\/AnnotationsXML\/\/002\"\n  --batch 4 --gpu 0.8<\/p>\n<\/blockquote>\n\n<p>So after 13000 steps:\n<a href=\"https:\/\/i.stack.imgur.com\/erFdB.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/erFdB.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>So I went to validate my results and this is what I get(Checkpoint 13000):\n<a href=\"https:\/\/i.stack.imgur.com\/LElaa.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/LElaa.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>So perhaps I thought this might be a case of severe overfitting, thus I iterate through the checkpoints to see which has the closest fit.<\/p>\n\n<p>This is what I get using checkpoint 6500\n<a href=\"https:\/\/i.stack.imgur.com\/482N2.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/482N2.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>This is what I get using checkpoint 6000<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/1KN45.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1KN45.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>This is what I get using checkpoint 5500\n<a href=\"https:\/\/i.stack.imgur.com\/usRdC.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/usRdC.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>So, as you can see checkpoint 6000 is the best possible result in my case but it isn't good enough. How do I improve on this? Increase batch size ?(My GPU 1070Ti cant handle. Cuda out of memory occurs) Any Ideas to solve this?<\/p>\n","answers":[{"AnswerId":"57880895","CreationDate":"2019-09-11T01:30:01.270","ParentId":null,"OwnerUserId":"2737131","Title":null,"Body":"<p>Using Yolov3 to train my imageset solved my issue. <a href=\"https:\/\/github.com\/AlexeyAB\/darknet\" rel=\"nofollow noreferrer\">https:\/\/github.com\/AlexeyAB\/darknet<\/a><\/p>\n\n<p>One thing to note is not to leave any blanks during annotation, perhaps this might be one of the reasons why the detection did not work as planned.<\/p>\n"}]}
+{"QuestionId":57792113,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2019-09-04T16:00:36.847","AcceptedAnswerId":null,"OwnerUserId":4735848.0,"Title":"Stratify batch in Tensorflow 2","Body":"<p>I have minibatches that I get from an sqlite database with data of integer and float type, <code>x<\/code>, and a binary label in 0 and 1, <code>y<\/code>. I am looking for something like <code>X_train, X_test, y_train, y_test = sklearn.model_selection.train_test_split(y, x, test_size=0.1, random_state=1, stratify=True)<\/code> from scikit-learn, where a keyword could stratify the data (i.e. the same number of class-0 and class-1 instances). <\/p>\n\n<p>In Tensorflow 2, stratification seems not straightforwardly possible. My very complicated solution works for me, but takes a lot of time because of all the reshaping and transposing:<\/p>\n\n<pre><code>def stratify(x, y):\n    # number of positive instances (the smaller class)\n    pos = np.sum(y).item() # how many positive bonds there are\n    x = np.transpose(x)\n\n    # number of features \n    f = np.shape(x)[1] \n\n    # filter only class 1\n    y = tf.transpose(y)\n    x_pos = tf.boolean_mask(x, \n    y_pos = tf.boolean_mask(y, y)\n\n    # filter only class 1\n    x_neg = tf.boolean_mask(x, tf.bitwise.invert(y)-254)\n    x_neg = tf.reshape(x_neg, [f,-1])\n    y_neg = tf.boolean_mask(y, tf.bitwise.invert(y)-254)\n\n    # just take randomy as many class-0 as there are class-1 \n    x_neg = tf.transpose(tf.random.shuffle(tf.transpose(x_neg)))\n    x_neg = x_neg[:,0:pos]\n    y_neg = y_neg[0:pos]\n\n    # concat the class-1 and class-0 together, then shuffle, and concat back together\n    x = tf.concat([x_pos,tf.transpose(x_neg)],0)\n    y = tf.concat([y_pos, tf.transpose(y_neg)],0)\n    xy = tf.concat([tf.transpose(x), tf.cast(np.reshape(y,[1, -1]), tf.float64)],0)\n    xy = tf.transpose((tf.random.shuffle(tf.transpose(xy)))) # because there is no axis arg in shuffle\n    x = xy[0:f,:]\n    x = tf.transpose(x)\n    y = xy[f,:]\n\n    return x, y\n<\/code><\/pre>\n\n<p>I am happy to see some feedback\/improvement on my own function or novel, easier ideas.<\/p>\n","answers":[{"AnswerId":"57793891","CreationDate":"2019-09-04T18:20:13.927","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>Data division is best if it is done in raw format only or before you transform it into tensors. If there is a strong requirement to do it in TensorFlow only, then I will suggest you to make use of <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/data\/Dataset\" rel=\"nofollow noreferrer\">tf.data.Dataset<\/a> class. I have added the demo code with relevant comments explaining the steps.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nTEST_SIZE = 0.1\nDATA_SIZE = 1000\n\n# Create data\nX_data = np.random.rand(DATA_SIZE, 28, 28, 1)\ny_data = np.random.randint(0, 2, [DATA_SIZE])\nsamples1 = np.sum(y_data)\nprint('Percentage of 1 = ', samples1 \/ len(y_data))\n\n# Create TensorFlow dataset\ndataset = tf.data.Dataset.from_tensor_slices((X_data, y_data))\n\n# Gather data with 0 and 1 labels separately\nclass0_dataset = dataset.filter(lambda x, y: y == 0)\nclass1_dataset = dataset.filter(lambda x, y: y == 1)\n\n# Shuffle them\nclass0_dataset = class0_dataset.shuffle(DATA_SIZE)\nclass1_dataset = class1_dataset.shuffle(DATA_SIZE)\n\n# Split them\nclass0_test_samples_len = int((DATA_SIZE - samples1) * TEST_SIZE)\nclass0_test = class0_dataset.take(class0_test_samples_len)\nclass0_train = class0_dataset.skip(class0_test_samples_len)\n\nclass1_test_samples_len = int(samples1 * TEST_SIZE)\nclass1_test = class1_dataset.take(class1_test_samples_len)\nclass1_train = class1_dataset.skip(class1_test_samples_len)\n\nprint('Train Class 0 = ', len(list(class0_train)), ' Class 1 = ', len(list(class1_train)))\nprint('Test Class 0 = ', len(list(class0_test)), ' Class 1 = ', len(list(class1_test)))\n\n# Gather datasets\ntrain_dataset = class0_train.concatenate(class1_train).shuffle(DATA_SIZE)\ntest_dataset = class0_test.concatenate(class1_test).shuffle(DATA_SIZE)\n\nprint('Train dataset size = ', len(list(train_dataset)))\nprint('Test dataset size = ', len(list(test_dataset)))\n<\/code><\/pre>\n\n<p>Sample output:<\/p>\n\n<pre><code>Percentage of 1 =  0.474\nTrain Class 0 =  474  Class 1 =  427\nTest Class 0 =  52  Class 1 =  47\nTrain dataset size =  901\nTest dataset size =  99\n<\/code><\/pre>\n"}]}
+{"QuestionId":57792140,"AnswerCount":1,"Tags":"<tensorflow><computer-vision><semantic-segmentation><deeplab>","CreationDate":"2019-09-04T16:02:04.600","AcceptedAnswerId":57904921.0,"OwnerUserId":10866154.0,"Title":"Why is the deeplab v3+ model confused about pixels outside image boundary?","Body":"<p>I'm using the google research github repository to run deeplab v3+ on my dataset to segment parts of a car. The crop size I've used is 513,513 (default) and the code adds a boundary to images smaller than that size (correct me if I'm wrong). <\/p>\n\n<p><a href=\"https:\/\/imgur.com\/Huuj3DX\" rel=\"nofollow noreferrer\">example<\/a>!<\/p>\n\n<p>The model seems to be performing poorly on the added boundary. Is there something I'm supposed to correct or will the model do fine with more training ?<\/p>\n\n<p>Update: <a href=\"https:\/\/imgur.com\/a\/L7akJLf\" rel=\"nofollow noreferrer\">Here's<\/a> the tensorboard graphs for training. Why is the regularization loss shooting like that? The output seems to be improving, can someone help me making inferences from these graphs?<\/p>\n","answers":[{"AnswerId":"57904921","CreationDate":"2019-09-12T10:37:21.883","ParentId":null,"OwnerUserId":"10478535","Title":null,"Body":"<blockquote>\n  <p>Is there something I'm supposed to correct or will the model do fine with more training ?<\/p>\n<\/blockquote>\n\n<p>its Ok, don't mind the boundary<\/p>\n\n<hr>\n\n<p>To inference you can use this code<\/p>\n\n<pre><code>import cv2\nimport tensorflow as tf\nimport numpy as np\nfrom PIL import Image\n\nfrom skimage.transform import resize\n\n\n\nclass DeepLabModel():\n    \"\"\"Class to load deeplab model and run inference.\"\"\"\n\n    INPUT_TENSOR_NAME = 'ImageTensor:0'\n    OUTPUT_TENSOR_NAME = 'SemanticPredictions:0'\n    INPUT_SIZE = 513\n\n\n    def __init__(self, path):\n        \"\"\"Creates and loads pretrained deeplab model.\"\"\"\n        self.graph = tf.Graph()\n\n        graph_def = None\n        # Extract frozen graph from tar archive.\n\n        with tf.gfile.GFile(path, 'rb')as file_handle:\n            graph_def = tf.GraphDef.FromString(file_handle.read())\n\n        if graph_def is None:\n            raise RuntimeError('Cannot find inference graph')\n\n        with self.graph.as_default():\n            tf.import_graph_def(graph_def, name='')\n\n        self.sess = tf.Session(graph=self.graph)\n\n\n    def run(self, image):\n        \"\"\"Runs inference on a single image.\n\n        Args:\n          image: A PIL.Image object, raw input image.\n\n        Returns:\n          seg_map: np.array. values of pixels are classes\n        \"\"\"\n\n        width, height = image.size\n\n        resize_ratio = 1.0 * self.INPUT_SIZE \/ max(width, height)\n        target_size = (int(resize_ratio * width), int(resize_ratio * height))\n\n        resized_image = image.convert('RGB').resize(target_size, Image.ANTIALIAS)\n\n        batch_seg_map = self.sess.run(\n           self.OUTPUT_TENSOR_NAME,\n           feed_dict={self.INPUT_TENSOR_NAME: [np.asarray(resized_image)]})\n\n        seg_map = batch_seg_map[0]        \n\n        seg_map = resize(seg_map.astype(np.uint8), (height, width), preserve_range=True, order=0, anti_aliasing=False)\n\n\n        return seg_map\n<\/code><\/pre>\n\n<p>the code is based on this file <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/deeplab\/deeplab_demo.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/deeplab\/deeplab_demo.ipynb<\/a><\/p>\n\n<pre><code>\nmodel = DeepLabModel(your_model_pb_path)\n\nimg = Image.open(img_path)\n\nseg_map = model.run(img)\n\n\n<\/code><\/pre>\n\n<hr>\n\n<p>to get <code>your_model_pb_path<\/code> you need to export your model to .pb file\nyou can do it using <code>export_model.py<\/code> file in Deeplab repo\n<a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/deeplab\/export_model.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/deeplab\/export_model.py<\/a><\/p>\n\n<p>if you were training <code>xception_65<\/code> version<\/p>\n\n<pre><code>python3 &lt;path to your deeplab folder&gt;\/export_model.py \\\n  --logtostderr \\\n  --checkpoint_path=&lt;your ckpt&gt; \\\n  --export_path=\".\/my_model.pb\" \\\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  --num_classes=&lt;NUMBER OF YOUR CLASSES&gt; \\\n  --crop_size=513 \\\n  --crop_size=513 \\\n  --inference_scales=1.0 \n<\/code><\/pre>\n\n<p><code>&lt;your ckpt&gt;<\/code> is a path to your trained model checkpoint you can find checkpoints in the folder that you passed as argument <code>--train_logdir<\/code> when training<\/p>\n\n<p>you need to include only model name and number of iterations in path, or in other words you will have in your training folder, for example, files \n<code>model-1500.meta<\/code>, <code>model-1500.index<\/code> and <code>model-1000.data-00000-of-00001<\/code> you need to discard everything that goes after <code>.<\/code>, so the ckpt path will be <code>model-1000<\/code><\/p>\n\n<p>please make sure that <code>atrous_rates<\/code> are the same as you used to train the model<\/p>\n\n<p>if you were training <code>mobilenet_v2<\/code> version<\/p>\n\n<pre><code>python3 &lt;path to your deeplab folder&gt;\/export_model.py \\\n  --logtostderr \\\n  --checkpoint_path=&lt;your ckpt&gt; \\\n  --export_path=\".\/my_model.pb\" \\\n  --model_variant=\"mobilenet_v2\" \\\n  --num_classes=&lt;NUMBER OF YOUR CLASSES&gt;  \\\n  --crop_size=513 \\\n  --crop_size=513 \\\n  --inference_scales=1.0\n<\/code><\/pre>\n\n<p>more you can find here\n<a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/deeplab\/local_test_mobilenetv2.sh\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/deeplab\/local_test_mobilenetv2.sh<\/a>\n<a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/deeplab\/local_test.sh\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/deeplab\/local_test.sh<\/a><\/p>\n\n<hr>\n\n<p>You can visualize results using this code<\/p>\n\n<pre><code>img_arr = np.array(img)\n\n# as may colors as you have classes\ncolors = [(255, 0, 0), (0, 255, 0), ...] \n\nfor c in range(0, N_CLASSES):\n    img_arr[seg_map == c] = 0.5 * img_arr[seg_map == c] + 0.5 * np.array(colors[c])\n\ncv2.imshow(img_arr)\ncv2.waitKey(0)        \n<\/code><\/pre>\n"}]}
+{"QuestionId":57792209,"AnswerCount":1,"Tags":"<machine-learning><keras><deep-learning><nlp><lstm>","CreationDate":"2019-09-04T16:06:42.953","AcceptedAnswerId":57792260.0,"OwnerUserId":9080239.0,"Title":"why my neural network sequential model reaches 0.9998 accuracy form the beginning?","Body":"<p>I am building a hashtag recommendation model for twitter media posts, which takes tweet text as input and does 300-dimensional word embedding on it and classifies it among 198 hashtags as classes. When I run my model I get 0.9998 accuracy from the beginning which does not change later! What is wrong in my model? <\/p>\n\n<pre><code>import numpy as np\nimport pickle\nfrom keras.layers.normalization import BatchNormalization\nfrom keras.models import Sequential, load_model\nfrom keras.layers import Dense, Dropout, Activation,LSTM, Embedding\nfrom keras.callbacks import ModelCheckpoint, ReduceLROnPlateau\nfrom keras import regularizers, initializers\npackage=\"2018_pickle\"\nwith open(path1,\"rb\") as f:\n    maxLen,l_h2i,l_w2i=pickle.load(f)\nwith open(path2,\"rb\") as f:\n    X_train,X_test,X_train_indices,X_test_indices=pickle.load(f)\nwith open(path3,\"rb\") as f:\n    Y_train,Y_test,Y_train_oh,Y_test_oh=pickle.load(f)\nwith open(path4,\"rb\") as f:\n    emd_matrix=pickle.load(f)\n\n\nif __name__ == '__main__':\nmodelname=\"model_1\"\ntrain=False\nvocab_size = len(emd_matrix)\nemd_dim=emd_matrix.shape[1]\nif train:\n    model = Sequential()\n    model.add(Embedding(vocab_size , emd_dim, weights=[emd_matrix]\n                        ,input_length=maxLen,trainable=False))\n    model.add(LSTM(256,return_sequences=True,activation=\"relu\",\n                   kernel_regularizer=regularizers.l2(0.01),\n                   kernel_initializer=initializers.glorot_normal(seed=None)))\n    model.add(LSTM(256,return_sequences=True,activation=\"relu\",\n                   kernel_regularizer=regularizers.l2(0.01),\n                   kernel_initializer=initializers.glorot_normal(seed=None)))\n    model.add(LSTM(256,return_sequences=False,activation=\"relu\",\n                   kernel_regularizer=regularizers.l2(0.01),\n                   kernel_initializer=initializers.glorot_normal(seed=None)))\n    model.add(Dense(198,activation='softmax'))\n    model.compile(loss='binary_crossentropy', optimizer='adam',\n                  metrics=['accuracy'])\n    checkpoint = ModelCheckpoint(filepath, monitor=\"loss\",\n                                 verbose=1, save_best_only=True, mode='min')\n    reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.5,\n                                  patience=2, min_lr=0.000001)\n    history=model.fit(X_train_indices, Y_train_oh, batch_size=2048,\n                      epochs=5, validation_split=0.1, shuffle=True,\n                      callbacks=[checkpoint, reduce_lr])\n\n\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nembedding_10 (Embedding)     (None, 54, 300)           22592100  \n_________________________________________________________________\nlstm_18 (LSTM)               (None, 54, 256)           570368    \n_________________________________________________________________\nlstm_19 (LSTM)               (None, 54, 256)           525312    \n_________________________________________________________________\nlstm_20 (LSTM)               (None, 256)               525312    \n_________________________________________________________________\ndense_7 (Dense)              (None, 198)               50886     \n=================================================================\nTotal params: 24,263,978\nTrainable params: 1,671,878\nNon-trainable params: 22,592,100\n_________________________________________________________________\n<\/code><\/pre>\n","answers":[{"AnswerId":"57792260","CreationDate":"2019-09-04T16:10:26.587","ParentId":null,"OwnerUserId":"4685471","Title":null,"Body":"<p>Most probably, this is due to the mistaken use of <code>loss='binary_crossentropy'<\/code> in a <strong>multi<\/strong>-class classification problem (see <a href=\"https:\/\/stackoverflow.com\/questions\/42081257\/keras-binary-crossentropy-vs-categorical-crossentropy-performance\/46038271#46038271\">Keras binary_crossentropy vs categorical_crossentropy performance?<\/a> for more details).<\/p>\n\n<p>You should change your model compilation to <\/p>\n\n<pre><code>model.compile(loss='categorical_crossentropy', optimizer='adam',\n                  metrics=['accuracy'])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57792591,"AnswerCount":0,"Tags":"<python-2.7><module><pyinstaller><theano><lasagne>","CreationDate":"2019-09-04T16:37:20.810","AcceptedAnswerId":null,"OwnerUserId":10881028.0,"Title":"Pyinstaller executable returning import error related to Theano module","Body":"<p>Background: I am trying to create a standalone Unix exe that runs the Python program I created. My program is dependent on Theano (1.0.4+unknown), Numpy (1.16.5), Scipy(1.2.1), Lasagne (0.2.dev1), Openpyxl, and Tweepy. When I run my program from the terminal everything works fine.<\/p>\n\n<p>Problem: After running a successful pyinstaller call, the executable returns an error saying <\/p>\n\n<blockquote>\n  <p>File \"\/Users\/user\/anaconda2\/lib\/python2.7\/site-packages\/PyInstaller-4.0.dev0+1085db707-py2.7.egg\/PyInstaller\/loader\/pyimod03_importers.py\", line 389, in load_module\n      exec(bytecode, module.<strong>dict<\/strong>)\n    File \"site-packages\/lasagne\/<strong>init<\/strong>.py\", line 14, in \n  ImportError: Could not import Theano.<\/p>\n<\/blockquote>\n\n<p>What I've Tried: I have tried including hidden imports for both Theano and Lasagne as well as copying and pasting the Theano module into the Dist folder generated by pyinstaller. None of these solutions seem to work for me. This is what my pyinstaller call looks like:<\/p>\n\n<blockquote>\n  <p>pyinstaller --onedir --windowed --paths=\/Users\/user\/Q3Crawler\/src:\/Users\/user\/Q3Crawler\/tweet2vec:\/Users\/user\/anaconda2:\/Users\/user\/anaconda2\/pkgs\/python-2.7.16-h97142e2_0\/lib:\/Users\/user\/anaconda2\/lib\/python2.7\/site-packages\/:\/Users\/user\/anaconda2\/pkgs\/ --hidden-import=lasagne --hidden-import=theano --add-data 'src\/:src' --add-data 'tweet2vec\/:tweet2vec' --add-data 'input files\/:input files' --add-data 'corpus\/:corpus' --add-data 'generated reports\/:generated reports' q3crawler.py<\/p>\n<\/blockquote>\n\n<p>It should also be noted that when I run pyinstaller this line appears <\/p>\n\n<blockquote>\n  <p>333964 INFO: Analyzing hidden import 'lasagne'<\/p>\n<\/blockquote>\n\n<p>but not for theano.<\/p>\n\n<p>Expected Results: A standalone executable that can be distributed to other users so they don't have to worry about dependencies.<\/p>\n","answers":[]}
+{"QuestionId":57792660,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2019-09-04T16:42:49.180","AcceptedAnswerId":57798873.0,"OwnerUserId":5623899.0,"Title":"Tensorflow 1.14+ Serialize Subclassed Keras Layers?","Body":"<p>I've gone through and re-re-re-read the Tensorflow Keras docs, e.g.<\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/keras\/functional\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/guide\/keras\/functional<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/keras\/custom_layers_and_models\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/guide\/keras\/custom_layers_and_models<\/a><\/li>\n<\/ul>\n\n<p>I have a simple subclassed layer:<\/p>\n\n<pre><code>class SimpleLayer(tf.keras.layers.Layer):\n    def __init__(self, filters, kernel_size, **kwargs):\n        super(SimpleLayer, self).__init__()\n        self.filters = filters\n        self.kernel_size = kernel_size\n        self.c1 = tf.keras.layers.Conv1D(filters, kernel_size, padding='same', activation='relu')\n        self.c2 = tf.keras.layers.Conv1D(filters, kernel_size, padding='same')\n\n    def call(self, inputs):\n        x = inputs\n        x = self.c1(x)\n        x = self.c2(x)\n        return x\n\n    def get_config(self):\n        # config = super(tf.keras.layers.Layer, self).get_config()\n        config = {}\n        config.update({\n            'filters': self.filters,\n            'kernel_size': self.kernel_size,\n        })\n        return config\n<\/code><\/pre>\n\n<p>and then have a functional model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nx = tf.keras.Inputs(...)\n\n# some keras layers\ny = tf.keras.layers... (x)\n\n# my keras layer\ny = SimpleLayer(...)(y)\n\n# some keras layers\ny = tf.keras.layers... (y)\ny = tf.keras.layers.Dense(1)(y)\n\nmodel = tf.keras.Model(inputs=x, outputs=y)\nmodel.compile(...)\n\nmodel.fit(...)\n\nmodel.save('model.h5')\n<\/code><\/pre>\n\n<p>and figure then I could load the model as:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tf.keras.models.load_model('model.h5')\n<\/code><\/pre>\n\n<p>but I get:<\/p>\n\n<pre><code>ValueError: Unknown layer: SimpleLayer\n<\/code><\/pre>\n\n<p>from the <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/keras\/custom_layers_and_models#you_can_optionally_enable_serialization_on_your_layers\" rel=\"nofollow noreferrer\">docs<\/a>:<\/p>\n\n<blockquote>\n  <p>If you need your custom layers to be serializable as part of a Functional model, you can optionally implement a get_config method<\/p>\n<\/blockquote>\n\n<p>which I have. <\/p>\n\n<p>What am I doing wrong?<\/p>\n","answers":[{"AnswerId":"57798873","CreationDate":"2019-09-05T05:21:34.877","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You need to tell keras about your custom layer during loading, you do this with the <code>custom_objects<\/code> parameter:<\/p>\n\n<pre><code>tf.keras.models.load_model('model.h5', custom_objects = {'SimpleLayer': SimpleLayer})\n<\/code><\/pre>\n"}]}
+{"QuestionId":57793430,"AnswerCount":0,"Tags":"<deep-learning><conv-neural-network><pytorch><relu>","CreationDate":"2019-09-04T17:43:21.640","AcceptedAnswerId":null,"OwnerUserId":9236891.0,"Title":"RuntimeError: size mismatch in PyTorch","Body":"<p>I have just started pyTorch and getting following error:\n<code>RuntimeError: size mismatch, m1: [128 x 256], m2: [400 x 120] at \/pytorch\/aten\/src\/THC\/generic\/THCTensorMathBlas.cu:268<\/code>\nI was trying LeNet architecture on MNIST dataset. I was trying various activation functions and while this error did not come for Tanh but it came for ReLU activation. So I'm guess that I have written code for ReLU wrong.\nModel: <\/p>\n\n<pre><code>class LeNet(nn.Module):\n    def __init__(self): \n        super(LeNet, self).__init__()\n        self.cnn_model = nn.Sequential(\n            nn.Conv2d(1, 6, 5),         \n            nn.ReLU(True),\n            nn.AvgPool2d(2, stride=2),  \n            nn.Conv2d(6, 16, 5),        \n            nn.ReLU(True),\n            nn.AvgPool2d(2, stride=2)  \n        )\n        self.fc_model = nn.Sequential(\n            nn.Linear(400,120),        \n            nn.ReLU(True),\n            nn.Linear(120,84),          \n            nn.ReLU(True),\n            nn.Linear(84,10)            \n        )\n\n    def forward(self, x):\n        print(x.shape)\n        x = self.cnn_model(x)\n        print(x.shape)\n        x = x.view(x.size(0), -1)\n        print(x.shape)\n        x = self.fc_model(x)\n        print(x.shape)\n        return x\n<\/code><\/pre>\n\n<p>I kept the batch size to 128:<\/p>\n\n<p>Data Import:<\/p>\n\n<pre><code>batch_size = 128\ntrainset = torchvision.datasets.MNIST(root='.\/data', train=True, download=True, transform=transforms.ToTensor())\ntrainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size, shuffle=True)\ntestset = torchvision.datasets.MNIST(root='.\/data', train=False, download=True, transform=transforms.ToTensor())\ntestloader = torch.utils.data.DataLoader(testset, batch_size=batch_size, shuffle=False)\n<\/code><\/pre>\n\n<p>Evaluation and Optimisation Model:<\/p>\n\n<pre><code> def evaluation(dataloader):\n        total, correct = 0, 0\n        for data in dataloader:\n            inputs, labels = data\n            inputs, labels = inputs.to(device), labels.to(device)\n            outputs = net(inputs)\n            _, pred = torch.max(outputs.data, 1)\n            total += labels.size(0)\n            correct += (pred == labels).sum().item()\n        return 100 * correct \/ total\n\n\nimport torch.optim as optim\n\nnet = LeNet().to(device)\nloss_fn = nn.CrossEntropyLoss()\nopt = optim.Adam(net.parameters())\n\nmax_epochs = 16\nfor epoch in range(max_epochs):\n\n    for i, data in enumerate(trainloader, 0):\n\n        inputs, labels = data\n        inputs, labels = inputs.to(device), labels.to(device)\n\n        opt.zero_grad()\n\n        outputs = net(inputs)\n        loss = loss_fn(outputs, labels)\n        loss.backward()\n        opt.step()\n\n    print('Epoch: %d\/%d' % (epoch, max_epochs))\n<\/code><\/pre>\n\n<p>I saw some earlier questions but was not able to understand how to modify ReLU<\/p>\n","answers":[]}
+{"QuestionId":57793574,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><matrix><neural-network><attention-model>","CreationDate":"2019-09-04T17:54:00.573","AcceptedAnswerId":null,"OwnerUserId":5865514.0,"Title":"Self-Attention Explainability of the Output Score Matrix","Body":"<p>I am learning about attention models, and following along with Jay Alammar's amazing blog tutorial on The <a href=\"http:\/\/jalammar.github.io\/illustrated-transformer\/\" rel=\"nofollow noreferrer\">Illustrated Transformer<\/a>. He gives a great walkthrough for how the attention scores are calculated, but I get a bit lost at a certain point, and am not seeing how the attention score <code>Z<\/code> matrix he explains is used to interpret strength of associations between different words within an input sequence.<\/p>\n\n<p>He mentions that given some input matrix <code>X<\/code>, with shape <code>N x D<\/code>, where <code>N<\/code> is the number of elements in an input sequence, and <code>D<\/code> is the input dimensionality, we multiply <code>X<\/code> with three separate weight matrices of shape <code>D x d<\/code>, where <code>d<\/code> is some lower dimensionality that represents the projected space of the query, key, and value matrices:\n<a href=\"https:\/\/i.stack.imgur.com\/0gV0C.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0gV0C.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The query and key matrices are dotted, and then divided by a scaling factor usually the square root of the projected dimensionality, and then run through a softmax function. This produces a weight matrix of size <code>N x N<\/code>, which is multiplied by the value matrix to get an output <code>Z<\/code> of shape <code>N x d<\/code>, which Jay says <\/p>\n\n<blockquote>\n  <p>That concludes the self-attention calculation. The resulting vector is\n  one we can send along to the feed-forward neural network.<\/p>\n<\/blockquote>\n\n<p>The screenshot from his blog for this calculation is below:\n<a href=\"https:\/\/i.stack.imgur.com\/4mhWz.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/4mhWz.png\" alt=\"enter image description here\"><\/a>\nHowever, this is where I'm confused. <code>Z<\/code> is <code>N x d<\/code>. However, I don't particularly understand what I'm supposed to do with this matrix from an interpretability sense, and as far as I understand, for a particular sequence element (ie. the word <code>cats<\/code> in the sequence <code>I love pets, especially cats<\/code>), self-attention is supposed to score other parts of the sequence high when it is relevant or strong associated with that word embedding. However, I'd expect then that <code>Z<\/code> is <code>N x N<\/code>, so I could say that I can select the <code>Z[i,j]<\/code> and say for the <code>i-th<\/code> word in the sequence, the <code>j-th<\/code> word relates or associates with it this or that much.<\/p>\n\n<p><strong>In fact, wouldn't it make much more sense to use only the softmax output of the weights (without multiplying them by the value matrix), since it already is <code>N x N<\/code>? In essence, how is Jay determining the strength of these associations in this particular sequence with the word <code>it<\/code>?<\/strong> \n<a href=\"https:\/\/i.stack.imgur.com\/UjNxJ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/UjNxJ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>This is an <code>N<\/code> by <code>1<\/code> relationship he is showing - there are <code>N<\/code> values that correspond with the strength of association to the word <code>it<\/code>.<\/p>\n","answers":[]}
+{"QuestionId":57793583,"AnswerCount":0,"Tags":"<python-3.x><keras><deep-learning><nlp><lstm>","CreationDate":"2019-09-04T17:54:14.577","AcceptedAnswerId":null,"OwnerUserId":9080239.0,"Title":"why my neural network sequential model gets low accuracy ( below than 0.0011)?","Body":"<p>I am building a hashtag recommendation model for twitter media posts, which takes tweet text as input and does 300-dimensional word embedding on it and classifies it among 198 hashtags as classes. When I run my model I get lower than 0.0011 accuracy which does not change later! What is wrong in my model? <\/p>\n\n<pre><code>import pickle\n\nimport numpy as np\n\nfrom keras import initializers, regularizers\nfrom keras.callbacks import ModelCheckpoint, ReduceLROnPlateau\nfrom keras.layers import LSTM, Activation, Dense, Dropout, Embedding\nfrom keras.layers.normalization import BatchNormalization\nfrom keras.models import Sequential, load_model\n\npackage = \"2018_pickle\"\nwith open(path1, \"rb\") as f:\n    maxLen, l_h2i, l_w2i = pickle.load(f)\nwith open(path2, \"rb\") as f:\n    X_train, X_test, X_train_indices, X_test_indices = pickle.load(f)\nwith open(path3, \"rb\") as f:\n    Y_train, Y_test, Y_train_oh, Y_test_oh = pickle.load(f)\nwith open(path4, \"rb\") as f:\n    emd_matrix = pickle.load(f)\n\n\nif __name__ == \"__main__\":\n    modelname = \"model_1\"\n    train = False\n    vocab_size = len(emd_matrix)\n    emd_dim = emd_matrix.shape[1]\n    if train:\n        model = Sequential()\n        model.add(\n            Embedding(\n                vocab_size,\n                emd_dim,\n                weights=[emd_matrix],\n                input_length=maxLen,\n                trainable=False,\n            )\n        )\n        model.add(\n            LSTM(\n                256,\n                return_sequences=True,\n                activation=\"relu\",\n                kernel_regularizer=regularizers.l2(0.01),\n                kernel_initializer=initializers.glorot_normal(seed=None),\n            )\n        )\n        model.add(\n            LSTM(\n                256,\n                return_sequences=True,\n                activation=\"relu\",\n                kernel_regularizer=regularizers.l2(0.01),\n                kernel_initializer=initializers.glorot_normal(seed=None),\n            )\n        )\n        model.add(\n            LSTM(\n                256,\n                return_sequences=False,\n                activation=\"relu\",\n                kernel_regularizer=regularizers.l2(0.01),\n                kernel_initializer=initializers.glorot_normal(seed=None),\n            )\n        )\n        model.add(Dense(198, activation=\"softmax\"))\n        model.compile(\n            loss=\"categorical_crossentropy\", optimizer=\"adam\", metrics=[\"accuracy\"]\n        )\n        checkpoint = ModelCheckpoint(\n            filepath, monitor=\"loss\", verbose=1, save_best_only=True, mode=\"min\"\n        )\n        reduce_lr = ReduceLROnPlateau(\n            monitor=\"val_loss\", factor=0.5, patience=2, min_lr=0.000001\n        )\n        history = model.fit(\n            X_train_indices,\n            Y_train_oh,\n            batch_size=2048,\n            epochs=5,\n            validation_split=0.1,\n            shuffle=True,\n            callbacks=[checkpoint, reduce_lr],\n        )\n\n<\/code><\/pre>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nembedding_10 (Embedding)     (None, 54, 300)           22592100  \n_________________________________________________________________\nlstm_18 (LSTM)               (None, 54, 256)           570368    \n_________________________________________________________________\nlstm_19 (LSTM)               (None, 54, 256)           525312    \n_________________________________________________________________\nlstm_20 (LSTM)               (None, 256)               525312    \n_________________________________________________________________\ndense_7 (Dense)              (None, 198)               50886     \n=================================================================\nTotal params: 24,263,978\nTrainable params: 1,671,878\nNon-trainable params: 22,592,100\n_________________________________________________________________\nNone\nTrain on 177278 samples, validate on 19698 samples\nEpoch 1\/5\n177278\/177278 [==============================] - 70s 396us\/step - loss:     3.3672 - acc: 8.7433e-04 - val_loss: 0.3103 - val_acc: 0.0000e+00\n\nEpoch 00001: loss improved from inf to 3.36719, saving model to     .\/checkpoints\/model_1\/lstm-01-3.367-0.001-0.310-0.000.hdf5\nEpoch 2\/5\n177278\/177278 [==============================] - 66s 371us\/step - loss: 0.1950 - acc: 2.4820e-04 - val_loss: 0.1616 - val_acc: 0.0016\n\nEpoch 00002: loss improved from 3.36719 to 0.19496, saving model to .\/checkpoints\/model_1\/lstm-02-0.195-0.000-0.162-0.002.hdf5\nEpoch 3\/5\n177278\/177278 [==============================] - 66s 370us\/step - loss: 0.1583 - acc: 0.0011 - val_loss: 0.1570 - val_acc: 0.0016\n\nEpoch 00003: loss improved from 0.19496 to 0.15826, saving model to     .\/checkpoints\/model_1\/lstm-03-0.158-0.001-0.157-0.002.hdf5\nEpoch 4\/5\n177278\/177278 [==============================] - 65s 369us\/step - loss: 0.1566 - acc: 0.0011 - val_loss: 0.1573 - val_acc: 0.0016\n\nEpoch 00004: loss improved from 0.15826 to 0.15660, saving model to     .\/checkpoints\/model_1\/lstm-04-0.157-0.001-0.157-0.002.hdf5\nEpoch 5\/5\n177278\/177278 [==============================] - 66s 374us\/step - loss: 0.1561 - acc: 0.0011 - val_loss: 0.1607 - val_acc: 0.0016\n\nEpoch 00005: loss improved from 0.15660 to 0.15610, saving model to     .\/checkpoints\/model_1\/lstm-05-0.156-0.001-0.161-0.002.hdf5\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57793812,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><machine-learning><lstm><simulation>","CreationDate":"2019-09-04T18:14:12.507","AcceptedAnswerId":null,"OwnerUserId":12021336.0,"Title":"Fixing reshaping\/shape problem of the LSTM","Body":"<p>Based on the work of <a href=\"https:\/\/github.com\/mgroncki\/DataScienceNotebooks\/blob\/master\/DeepHedging\/DeepHedging_Part1.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/mgroncki\/DataScienceNotebooks\/blob\/master\/DeepHedging\/DeepHedging_Part1.ipynb<\/a> and his RNN Network, I'm trying to extend this Model for Spread Options. The Payoff of a Spread Option is basicly max(Underlying1-Underlying2-Strike;0) at Time T.<\/p>\n\n<p>While extending and changing this Code I'm really getting trouble with shaping problems.<\/p>\n\n<p>I have written down all the Shapes of each Tensor, Variable and so on, but I didn't get the point. Also I've tried to put it in a smaller number an observe the whole process step by step without any positive reward.<\/p>\n\n<p>So here is the LSTM\/RNN code based on Matthias Github Code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class LSTM(object):\n    def __init__(self, time_steps, batch_size, features, nodes = [62,46,46,2], name='model'):\n        tf.reset_default_graph()\n        self.batch_size = batch_size\n        self.S_t_input = tf.placeholder(tf.float32, [time_steps, batch_size, features])\n        self.K = tf.placeholder(tf.float32, batch_size)\n        self.alpha = tf.placeholder(tf.float32)\n\n        S_T_0 = self.S_t_input[-1,:,0]\n        S_T_1 = self.S_t_input[-1,:,1]\n\n        # Returns from Underlying 1 and Underlying 2\n        dS_0 = self.S_t_input[1:, :, 0] - self.S_t_input[0:-1, :, 0]\n        dS_1 = self.S_t_input[1:, :, 1] - self.S_t_input[0:-1, :, 1]\n        #dS = tf.reshape(dS, (time_steps, batch_size))\n\n        #Prepare S_t for the use in the RNN remove the last time step (at T the portfolio is zero)\n        #S_t = tf.unstack(self.S_t_input[:-1, :,:], axis=0)\n        # S_t wird zu einer Liste bestehend aus Tensoren mit Shape\n\n        # Build the LSTM (Multi Zellen) !! Output in nodes = 1 !!!\n        lstm = tf.contrib.rnn.MultiRNNCell([tf.contrib.rnn.LSTMCell(n) for n in nodes])\n\n        # Output == Stragegy \n        self.strategy, state = tf.nn.dynamic_rnn(lstm, self.S_t_input[:-1,:,:], dtype=tf.float32)\n\n        #self.strategy = tf.reshape(self.strategy, (time_steps-1, batch_size, features))\n\n        # Optionvalue of the Spread option\n        self.option = tf.maximum(S_T_0-S_T_1-self.K, 0)\n\n        # Maybe Wrong but lets focus on the Shapes\n        self.Hedging_PnL = - self.option + tf.reduce_sum(dS_0*self.strategy[:,:,0], axis=0) + tf.reduce_sum(dS_1*self.strategy[:,:,1], axis=0)\n\n        print(tf.shape(self.option))\n        print(tf.shape(self.Hedging_PnL))\n\n        self.Hedging_PnL_Paths = - self.option + dS_0*self.strategy[:,:,0] + dS_1*self.strategy[:,:,1]\n\n        # Calculate the CVaR for a given confidence level alpha\n        # Take the 1-alpha largest losses (top 1-alpha negative PnLs) and calculate the mean\n        CVaR, idx = tf.nn.top_k(-self.Hedging_PnL, tf.cast((1-self.alpha)*batch_size, tf.int32))\n\n        #CVaR = tf.reduce_mean(CVaR)\n        CVaR = tf.math.reduce_sum(CVaR)\n\n        self.train = tf.train.AdamOptimizer().minimize(CVaR)\n        self.saver = tf.train.Saver()\n        self.modelname = name\n        print('--- DONE ---')\n<\/code><\/pre>\n\n<p>Python will build up the model, but when i run the trainroutine, i get the Errors from Python. The Train routine is exactly the same as by Matthias<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def _execute_graph_batchwise(self, paths, strikes, riskaversion, sess, epochs=1, train_flag=False):\n        sample_size = paths.shape[1]\n        batch_size=self.batch_size\n        idx = np.arange(sample_size)\n        start = dt.datetime.now()\n        for epoch in range(epochs):\n            # Save the hedging Pnl for each batch      \n            pnls = []\n            strategies = []\n            if train_flag:\n                np.random.shuffle(idx)\n            for i in range(int(sample_size\/batch_size)):\n                indices = idx[i*batch_size : (i+1)*batch_size]\n                batch = paths[:,indices,:]\n                if train_flag:\n                    _, pnl, strategy = sess.run([self.train, self.Hedging_PnL, self.strategy], {self.S_t_input: batch,\n                                                                                          self.K : strikes[indices],\n                                                                                          self.alpha: riskaversion})\n                else:\n                    pnl, strategy = sess.run([self.Hedging_PnL, self.strategy], {self.S_t_input: batch,\n                                                                                self.K : strikes[indices],\n                                                                                self.alpha: riskaversion})\n                pnls.append(pnl)\n                strategies.append(strategy)\n            #Calculate the option prive given the risk aversion level alpha\n            CVaR = np.mean(-np.sort(np.concatenate(pnls))[:int((1-riskaversion)*sample_size)])\n            if train_flag:\n                if epoch % 10 == 0:\n                    print('Time elapsed:', dt.datetime.now()-start)\n                    print('Epoch', epoch, 'CVaR', CVaR)\n                    self.saver.save(sess, r\"\/Users\/floriankonig\/LSTM\/%s\/model.ckpt\" % self.modelname)\n        self.saver.save(sess, r\"\/Users\/floriankonig\/LSTM\/%s\/model.ckpt\" % self.modelname)\n        return CVaR, np.concatenate(pnls), np.concatenate(strategies,axis=1)\n\n    def training(self, paths, strikes, riskaversion, epochs, session, init=True):\n        if init:\n            session.run(tf.global_variables_initializer())\n        self._execute_graph_batchwise(paths, strikes, riskaversion, session, epochs, train_flag=True)\n<\/code><\/pre>\n\n<p>The input data gets generated by a montecarlo simulation and has a shape (time_steps, batch_size, features).<\/p>\n\n<p>I want to get the CVaR calculated, but I get the following error.<\/p>\n\n<blockquote>\n  <p>Reshape cannot infer the missing input size for an empty tensor unless all specified input sizes are non-zero<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":57793970,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-estimator>","CreationDate":"2019-09-04T18:27:13.597","AcceptedAnswerId":null,"OwnerUserId":6249001.0,"Title":"Tensorflow: Error using label_vocabulary in estimator","Body":"<p>Whenever I try to train a classifier using <code>label_vocabulary<\/code> parameter I get an error such as: <code>TypeError: Value passed to parameter 'targets' has DataType string not in list of allowed values: int32, int64<\/code>. On the contrary, when I don't use it and label encode the target like mentioned in  TensorFlow API then the net is able to train. <\/p>\n\n<p>The model is created using:<\/p>\n\n<pre><code>tf.estimator.DNNClassifier(\n            # Config &amp; Params\n            n_classes=n_classes,\n            label_vocabulary=label_vocabulary['target'].tolist(),\n            config=tf.estimator.RunConfig(\n              model_dir=model_dir, \n              save_checkpoints_steps=100,\n              keep_checkpoint_max = 5,\n            ),\n            model_dir=model_dir,\n            feature_columns= feature_columns,      \n            hidden_units=[256, 128],            \n            activation_fn=tf.nn.relu,          \n            optimizer=tf.train.ProximalAdagradOptimizer(learning_rate=0.1,\n                                                            l1_regularization_strength=0.001),\n        )\n<\/code><\/pre>\n\n<p>Thanks<\/p>\n","answers":[]}
+{"QuestionId":57794598,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-04T19:17:21.600","AcceptedAnswerId":57795477.0,"OwnerUserId":7702081.0,"Title":"Best method to add noise on tf.dataset images \"on the fly\"","Body":"<p>After training a model (image classification) I would like to see how it performs differently when I evaluate a proper image and various noised versions of it.<\/p>\n\n<p>The type of noise I'm thinking is a random change in pixels value, I tried with this approach:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># --Inside the generator function that I provide to model.predict_generator--\n# dataset is a numpy array with denoised images path\ndt = tf.data.Dataset.from_generator(lambda: image_generator(dataset), output_types=(tf.float32))\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>def image_generator_(image_paths):\n    for path in image_paths:\n        # im is keras.preprocessing image\n        img = im.load_img(path,\n                          color_mode='rgb',\n                          target_size=(224,224))\n\n        img_to_numpy = np.array(img)\n\n        for _ in range (0, 5):\n            tmp_numpy_image = img_to_numpy.copy()\n            for i in range(tmp_numpy_image.shape[0]): \n                for j in range(tmp_numpy_image.shape[1]):\n                    # add noise\n                    tmp_numpy_image.shape[i][j] = ...\n            yield tmp_numpy_image\n<\/code><\/pre>\n\n<p>This process works fine but it is very slow. I also use dataset.batch and dataset.prefetch on <strong>dt<\/strong> and I didn't found a combination for their values that reduces the algorithm time<\/p>\n\n<p>Is there a smarter way to do it? I tried by yielding not noised images and to add the noise later inside dataset.map. The problem is that inside map I have to manipulate tensors and I didn't found a way to change each pixel value<\/p>\n\n<hr>\n\n<p><strong>SOLUTION<\/strong><\/p>\n\n<p>I used @Marat approach and it worked like a charm, the whole process went from 20-30 hours to minutes. My noise was a simple +-1 but I didn't want to go in overflow (255+1 = 0 in uint8) and therefore I only had to use numpy masks<\/p>\n\n<pre><code>...\ntmp_numpy_image = img_to_numpy.copy()\n\nnoise = np.random.randint(-1, 1, img_to_numpy.shape)\n\n# tmp_numpy_imag will become of type int32\ntmp_numpy_image = tmp_numpy_image + noise\n\nnp.putmask(tmp_numpy_image, tmp_numpy_image &lt; 0, 0)\nnp.putmask(tmp_numpy_image, tmp_numpy_image &gt; 255, 255)\n\ntmp_numpy_image = tmp_numpy_image.astype('uint8') \n\nyield tmp_numpy_image\n<\/code><\/pre>\n","answers":[{"AnswerId":"57795477","CreationDate":"2019-09-04T20:50:36.023","ParentId":null,"OwnerUserId":"271641","Title":null,"Body":"<p>The biggest overhead here is pixel operations (double for loop). Vectorizing it should result in substantial speedup:<\/p>\n\n<pre><code>noise_magnitude = 10\n...\n    img_max_value = img_to_numpy.max() * np.ones(img_to_numpy.shape)\n    for _ in range (0, 5):\n        # depending on range of values, you might want to adjust noise magnitude\n        noise = np.random.randint(0, noise_magnitude, img_to_numpy.shape)\n        # after adding noise, clip values exceeding max values\n        yield np.maximum(img_to_numpy + noise, img_max_value)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57794772,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-04T19:31:59.560","AcceptedAnswerId":null,"OwnerUserId":2317696.0,"Title":"Keras custom loss function (gradient error) with matrix multiplication","Body":"<p>I'm trying to recreate a model outlined a paper (<a href=\"https:\/\/arxiv.org\/pdf\/1811.06861.pdf\" rel=\"nofollow noreferrer\">Anomaly Detection using Deep Learning based\nImage Completion<\/a>) for image completion in Keras, but I'm currently stuck on the loss function:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/iUDLP.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/iUDLP.png\" alt=\"Loss function\"><\/a><\/p>\n\n<ul>\n<li>\u2299 denotes element-wise matrix multiplication<\/li>\n<li>\u2016\u22c5\u20161 denotes the L1 matrix norm<\/li>\n<li>In Keras, F is the predicted value, and X is the true value.<\/li>\n<\/ul>\n\n<p>What I have written is below.<\/p>\n\n<pre><code>import keras.backend as K\nimport tensorflow as tf\nimport numpy as np\n\n\ndef l1_reconstruction_loss(batch_size, mask_loss_weight=0.90, mask_size=32, img_size=128):\n    n_pixels = img_size * img_size\n    pad_size = int((img_size - mask_size) \/ 2)\n    mask = np.ones((batch_size, mask_size, mask_size, 3))\n    mask = np.pad(mask, ((0, 0), (pad_size, pad_size), (pad_size, pad_size), (0, 0)), 'constant')\n    inv_mask = np.logical_not(mask).astype(np.float32)\n\n    def _loss_func(y_true, y_pred):\n        mask_tf = tf.convert_to_tensor(mask, np.float32)\n        inv_mask_tf = tf.convert_to_tensor(inv_mask, np.float32)\n        inv_mask_true = tf.math.multiply(y_true, inv_mask_tf)\n        pred_dot_inv_mask_true = tf.matmul(y_pred, inv_mask_true)\n        true_sub_pred_dot_inv_mask_true = tf.math.subtract(y_true, pred_dot_inv_mask_true)\n\n        inner_expr_a = tf.math.multiply(mask_tf, true_sub_pred_dot_inv_mask_true)\n\n        inner_expr_b = tf.math.multiply(inv_mask_tf, true_sub_pred_dot_inv_mask_true)\n\n        mask_loss = mask_loss_weight * K.sum(K.abs(inner_expr_a)) \/ n_pixels + (1 - mask_loss_weight) * K.sum(K.abs(inner_expr_b)) \/ n_pixels\n        return mask_loss\n\n    def loss_func(y_true, y_pred):\n        return _loss_func(y_pred, y_true)\n\n    return loss_func\n<\/code><\/pre>\n\n<p>Edit: I revised the loss function code and no longer get an error related to the gradient. <\/p>\n\n<p>It results in the error:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: In[0] mismatch In[1] shape: 3 vs. 128: [128,128,128,3] [128,128,128,3] 0 0\n     [[{{node loss\/layer17_loss\/MatMul}}]]\n<\/code><\/pre>\n\n<p>Model compile step:<\/p>\n\n<pre><code>model.compile(loss=l1_reconstruction_loss(batch_size=128, mask_size=32, img_size=128),optimizer=optimizer)\n<\/code><\/pre>\n\n<p>Model:<\/p>\n\n<pre><code>img_input = layers.Input(shape=(img_size, img_size, 3))\n\n    # This will mask the central region (square), of image, replacing it with black.\n    # x = layers.Lambda(\n    #     function=lambda imgs: tf.map_fn(lambda x: mask_img(img=x, mask_size=mask_size, img_size=img_size), imgs)\n    # )(img_input)\n\n    # Layer 1\n    x = layers.Conv2D(32,\n                      kernel_size=5,\n                      strides=1,\n                      dilation_rate=(1, 1),\n                      kernel_initializer=initializers.RandomNormal(mean=0.0, stddev=1.0, seed=None),\n                      bias_initializer=initializers.Zeros(),\n                      padding='same',\n                      activation='elu',\n                      name='layer1')(img_input)\n    # Layer 2\n    x = layers.Conv2D(64,\n                      kernel_size=3,\n                      strides=1,\n                      dilation_rate=(1, 1),\n                      kernel_initializer=initializers.RandomNormal(mean=0.0, stddev=1.0, seed=None),\n                      bias_initializer=initializers.Zeros(),\n                      padding='same',\n                      activation='elu',\n                      name='layer2')(x)\n    # Layer 3\n    x = layers.Conv2D(64,\n                      kernel_size=3,\n                      strides=1,\n                      dilation_rate=(1, 1),\n                      kernel_initializer=initializers.RandomNormal(mean=0.0, stddev=1.0, seed=None),\n                      bias_initializer=initializers.Zeros(),\n                      padding='same',\n                      activation='elu',\n                      name='layer3')(x)\n    # Layer 4\n    x = layers.Conv2D(128,\n                      kernel_size=3,\n                      strides=2,\n                      dilation_rate=(1, 1),\n                      kernel_initializer=initializers.RandomNormal(mean=0.0, stddev=1.0, seed=None),\n                      bias_initializer=initializers.Zeros(),\n                      padding='same',\n                      activation='elu',\n                      name='layer4')(x)\n    # Layer 5\n    x = layers.Conv2D(128,\n                      kernel_size=3,\n                      strides=1,\n                      dilation_rate=(1, 1),\n                      kernel_initializer=initializers.RandomNormal(mean=0.0, stddev=1.0, seed=None),\n                      bias_initializer=initializers.Zeros(),\n                      padding='same',\n                      activation='elu',\n                      name='layer5')(x)\n\n    # Layer 6\n    x = layers.Conv2D(128,\n                      kernel_size=3,\n                      strides=1,\n                      dilation_rate=(1, 1),\n                      kernel_initializer=initializers.RandomNormal(mean=0.0, stddev=1.0, seed=None),\n                      bias_initializer=initializers.Zeros(),\n                      padding='same',\n                      activation='elu',\n                      name='layer6')(x)\n\n    # Layer 7\n    x = layers.Conv2D(128,\n                      kernel_size=3,\n                      strides=1,\n                      dilation_rate=(2, 2),\n                      kernel_initializer=initializers.RandomNormal(mean=0.0, stddev=1.0, seed=None),\n                      bias_initializer=initializers.Zeros(),\n                      padding='same',\n                      activation='elu',\n                      name='layer7')(x)\n\n    # Layer 8\n    x = layers.Conv2D(128,\n                      kernel_size=3,\n                      strides=1,\n                      dilation_rate=(4, 4),\n                      kernel_initializer=initializers.RandomNormal(mean=0.0, stddev=1.0, seed=None),\n                      bias_initializer=initializers.Zeros(),\n                      padding='same',\n                      activation='elu',\n                      name='layer8')(x)\n\n    # Layer 9\n    x = layers.Conv2D(128,\n                      kernel_size=3,\n                      strides=1,\n                      dilation_rate=(8, 8),\n                      kernel_initializer=initializers.RandomNormal(mean=0.0, stddev=1.0, seed=None),\n                      bias_initializer=initializers.Zeros(),\n                      padding='same',\n                      activation='elu',\n                      name='layer9')(x)\n\n    # Layer 10\n    x = layers.Conv2D(128,\n                      kernel_size=3,\n                      strides=1,\n                      dilation_rate=(16, 16),\n                      kernel_initializer=initializers.RandomNormal(mean=0.0, stddev=1.0, seed=None),\n                      bias_initializer=initializers.Zeros(),\n                      padding='same',\n                      activation='elu',\n                      name='layer10')(x)\n\n    # Layer 11\n    x = layers.Conv2D(128,\n                      kernel_size=3,\n                      strides=1,\n                      dilation_rate=(1, 1),\n                      kernel_initializer=initializers.RandomNormal(mean=0.0, stddev=1.0, seed=None),\n                      bias_initializer=initializers.Zeros(),\n                      padding='same',\n                      activation='elu',\n                      name='layer11')(x)\n\n    # Layer 12\n    x = layers.Conv2D(128,\n                      kernel_size=3,\n                      strides=1,\n                      dilation_rate=(1, 1),\n                      kernel_initializer=initializers.RandomNormal(mean=0.0, stddev=1.0, seed=None),\n                      bias_initializer=initializers.Zeros(),\n                      padding='same',\n                      activation='elu',\n                      name='layer12')(x)\n\n    # Bilinear Upscaling\/Upsampling\n    x = layers.UpSampling2D(size=2, interpolation='bilinear')(x)\n\n    # Layer 13\n    x = layers.Conv2D(64,\n                      kernel_size=3,\n                      strides=1,\n                      dilation_rate=(1, 1),\n                      kernel_initializer=initializers.RandomNormal(mean=0.0, stddev=1.0, seed=None),\n                      bias_initializer=initializers.Zeros(),\n                      padding='same',\n                      activation='elu',\n                      name='layer13')(x)\n\n    # Layer 14\n    x = layers.Conv2D(64,\n                      kernel_size=3,\n                      strides=1,\n                      dilation_rate=(1, 1),\n                      kernel_initializer=initializers.RandomNormal(mean=0.0, stddev=1.0, seed=None),\n                      bias_initializer=initializers.Zeros(),\n                      padding='same',\n                      activation='elu',\n                      name='layer14')(x)\n\n    # Layer 15\n    x = layers.Conv2D(32,\n                      kernel_size=3,\n                      strides=1,\n                      dilation_rate=(1, 1),\n                      kernel_initializer=initializers.RandomNormal(mean=0.0, stddev=1.0, seed=None),\n                      bias_initializer=initializers.Zeros(),\n                      padding='same',\n                      activation='elu',\n                      name='layer15')(x)\n\n    # Layer 16\n    x = layers.Conv2D(16,\n                      kernel_size=3,\n                      strides=1,\n                      dilation_rate=(1, 1),\n                      kernel_initializer=initializers.RandomNormal(mean=0.0, stddev=1.0, seed=None),\n                      bias_initializer=initializers.Zeros(),\n                      padding='same',\n                      activation='elu',\n                      name='layer16')(x)\n\n    # Layer 17\n    out = layers.Conv2D(3,\n                        kernel_size=3,\n                        strides=1,\n                        dilation_rate=(1, 1),\n                        kernel_initializer=initializers.RandomNormal(mean=0.0, stddev=1.0, seed=None),\n                        bias_initializer=initializers.Zeros(),\n                        padding='same',\n                        activation='elu',\n                        name='layer17')(x)\n<\/code><\/pre>\n\n<p>What am I doing wrong?<\/p>\n","answers":[]}
+{"QuestionId":57795697,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2019-09-04T21:12:17.243","AcceptedAnswerId":null,"OwnerUserId":5510818.0,"Title":"tf.keras h5 to Tensorflow pb - resulting pb lacks output node even though input clearly has it?","Body":"<p>I am using Google Collab, training a sequence model (NASNet) with a custom output. I export my model to an h5 using the <code>model.save()<\/code> method. I use a separate ipynb to load my h5 and convert it to a pb, however, my resulting pb lacks a named output node, and I cannot get predictions from my resulting model.<\/p>\n\n<p>My model sequence definition:<\/p>\n\n<pre><code>model = tf.keras.Sequential([\n  base_model,\n  tf.keras.layers.GlobalAveragePooling2D(),\n  tf.keras.layers.Dense(num_labels, activation = 'sigmoid', name=\"output\")])\n\nmodel.summary()\n<\/code><\/pre>\n\n<p>I am able to make compile my model, train, and make predictions in collab. <\/p>\n\n<p>I export my model like so:<\/p>\n\n<pre><code>model.save(model_path)\n<\/code><\/pre>\n\n<p>To convert to Tensorflow PB, I use a new ipynb and am able to 'successfully' run the conversion (ie, it doesn't fail).<\/p>\n\n<p>The conversion code seems widely used elsewhere. Here it is:<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorflow.keras.backend as K\n\nK.set_learning_phase(0)\n\nrestored_model = tf.keras.models.load_model(model_path)\nprint(restored_model.outputs)\nprint(restored_model.inputs)\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    from tensorflow.python.framework.graph_util import convert_variables_to_constants\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        # Graph -&gt; GraphDef ProtoBuf\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 = convert_variables_to_constants(session, input_graph_def,\n                                                      output_names, freeze_var_names)\n        return frozen_graph\n\n\n\n\nfrozen_graph = freeze_session(K.get_session(),\n                              output_names=[out.op.name for out in restored_model.outputs], \n                             clear_devices=True)\n\ntf.train.write_graph(frozen_graph, \"\/tmp\", model_name+\".pb\", as_text=False)\n\noutput_model_name = model_name + \".pb\"\noutput_model_path = \"\/tmp\/\" + output_model_name\n<\/code><\/pre>\n\n<p>This resulting pb appears well formed at first glance, but Netron and other tooling indicates it has no final output node in the graph.<\/p>\n\n<p>I am able to verify my tf.keras h5 model has inputs and outputs in my first ipynb via:<\/p>\n\n<pre><code>print(model.input.op.name)\nprint(model.output.op.name)\n<\/code><\/pre>\n\n<p>Which returns:<\/p>\n\n<pre><code>NASNet_input\noutput\/Identity\n<\/code><\/pre>\n\n<p>And when I load my saved h5 in prep to convert to Tensorflow protobuff, I also check the inout and output names:<\/p>\n\n<pre><code>print(restored_model.outputs)\nprint(restored_model.inputs)\n<\/code><\/pre>\n\n<p>Which returns in TF naming convention:<\/p>\n\n<pre><code>[&lt;tf.Tensor 'output\/Sigmoid:0' shape=(?, 229) dtype=float32&gt;]\n[&lt;tf.Tensor 'NASNet_input:0' shape=(?, 224, 224, 3) dtype=float32&gt;]\n<\/code><\/pre>\n\n<p>So clearly my h5 appears to have an output node named 'output' - as confirmed in Netron when I inspect the H5. <\/p>\n\n<p>Why does my conversion code appear to be removing the output node in the pb graph?<\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"57812490","CreationDate":"2019-09-05T20:20:04.073","ParentId":null,"OwnerUserId":"5510818","Title":null,"Body":"<p>So, it appears that tf 2.0 resolve this bug (see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/26809\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/26809<\/a>) , which is unhelpful because 1.14 can export to a valid protobuf (which I need, I can't use the SaveModel new format in 2.0 since coreML converters can't read it, and onnx converters die on this saved model for some other reason) - so this entire endeavor appears to be a dead end.<\/p>\n\n<p>Can one seriously not export TF 2.0 Keras model to a Protobuffers compatible with 1.14 runtime? <\/p>\n"}]}
+{"QuestionId":57795861,"AnswerCount":1,"Tags":"<python><machine-learning><pytorch><linear-regression>","CreationDate":"2019-09-04T21:31:55.930","AcceptedAnswerId":57797710.0,"OwnerUserId":11520760.0,"Title":"why Gradient Descent doesn't work as expected with pytorch","Body":"<p>so I'm starting with Pytorch and tried to start with an easy Linear Regression Example. Actually I made an easy Implementation of Linear Regression with Pytorch to calculate the equation 2*x+1 but the loss stay stuck at 120 and there is a Problem with Gradient Descent because it doesn't converge to a small loss value. I don't know why this is happening and it made me crazy because I don't see what's wrong. actually this example should be very easy to solve. this is the Code I'm using<\/p>\n\n<pre><code>import torch\nimport torch.nn.functional as F\nfrom torch.utils.data import TensorDataset, DataLoader\nimport numpy as np\n\nX = np.array([i for i in np.arange(1, 20)]).reshape(-1, 1)\nX = torch.tensor(X, dtype=torch.float32, requires_grad=True)\ny = np.array([2*i+1 for i in np.arange(1, 20)]).reshape(-1, 1)\ny = torch.tensor(y, dtype=torch.float32, requires_grad=True)\nprint(X.shape, y.shape)\n\nclass LR(torch.nn.Module):\n    def __init__(self, n_features, n_hidden1, n_out):\n        super(LR, self).__init__()\n        self.linear = torch.nn.Linear(n_features, n_hidden1)\n\n        self.predict = torch.nn.Linear(n_hidden1, n_out)\n\n    def forward(self, x):\n        x = F.relu(self.linear(x))\n        x = self.predict(x)\n        return x\n\nmodel = LR(1, 10, 1)\noptimizer = torch.optim.SGD(model.parameters(), lr=0.01)\nloss_fn = torch.nn.MSELoss()\n\ndef train(epochs=100):\n    for e in range(epochs):\n\n        pred = model(X)\n        loss = loss_fn(pred, y)\n        optimizer.zero_grad()\n        loss.backward()\n        optimizer.step()\n        print(f\"epoch: {e} and loss= {loss}\")\n<\/code><\/pre>\n\n<p>desired output is a small loss value and that the model train to give a good prediction later.<\/p>\n","answers":[{"AnswerId":"57797710","CreationDate":"2019-09-05T02:31:38.930","ParentId":null,"OwnerUserId":"4809029","Title":null,"Body":"<p>Your learning rate is too large. The model takes a few steps in the right direction, but it can't land on an actually good minimizer and henceforth zigzags around it. If you try <code>lr=0.001<\/code> instead,  your performance will be much better. This is why it's often useful to decay your learning rate over time when using first order optimizers.<\/p>\n"}]}
+{"QuestionId":57796221,"AnswerCount":0,"Tags":"<keras><deep-learning><rstudio><valueerror>","CreationDate":"2019-09-04T22:10:29.753","AcceptedAnswerId":null,"OwnerUserId":11760847.0,"Title":"Error in py_call_impl(callable, dots$args, dots$keywords) : ValueError: could not convert string to float: Null","Body":"<p>I am doing deep learning using Keras in Rstudio. I have some embedding layers at the beginning of the model. I checked continuous variables and there is no missing value, and also responding variable y is float.<\/p>\n\n<pre><code>df_cl_dl = df_cl %&gt;% filter(agency == \"FHA\") %&gt;% \n           select(lender, channel, fthb, region, credit_score, credit_score_null, source, ltv_uw, seasonality,\n                  current_ltv, loan_age, cash_incentive_a, hpas, loan_size, ur, risk,\n                  vsmm) %&gt;%\n           sample_n(100000)\n\ninp_lender = layer_input(shape = c(1), name = \"inp_lender\")\ninp_channel = layer_input(shape = c(1), name = \"inp_channel\")\ninp_fthb = layer_input(shape = c(1), name = \"inp_fthb\")\ninp_region = layer_input(shape = c(1), name = \"inp_region\")\ninp_cs_null = layer_input(shape = c(1), name = \"inp_cs_null\")\ninp_source = layer_input(shape = c(1), name = \"inp_source\")\ninp_season = layer_input(shape = c(1), name = \"inp_season\")\ninp_ltv_uw = layer_input(shape = c(1), name = \"inp_ltv_uw\")\ninp_continuous = layer_input(shape = c(8), name = \"inp_continuous\")\n\nembedding_out1 = inp_lender %&gt;% layer_embedding(input_dim = 3+1, output_dim = 2, input_length = 1, name = \"embedding_lender\") %&gt;% layer_flatten()\nembedding_out2 = inp_channel %&gt;% layer_embedding(input_dim = 3+1, output_dim = 2, input_length = 1, name = \"embedding_channel\") %&gt;% layer_flatten()\nembedding_out3 = inp_fthb %&gt;% layer_embedding(input_dim = 3+1, output_dim = 2, input_length = 1, name = \"embedding_fthb\") %&gt;% layer_flatten()\nembedding_out4 = inp_region %&gt;% layer_embedding(input_dim = 4+1, output_dim = 2, input_length = 1, name = \"embedding_region\") %&gt;% layer_flatten()\nembedding_out5 = inp_cs_null %&gt;% layer_embedding(input_dim = 2+1, output_dim = 2, input_length = 1, name = \"embedding_cs_null\") %&gt;% layer_flatten()\nembedding_out6 = inp_source %&gt;% layer_embedding(input_dim = 2+1, output_dim = 2, input_length = 1, name = \"embedding_source\") %&gt;% layer_flatten()\nembedding_out7 = inp_season %&gt;% layer_embedding(input_dim = 12+1, output_dim = 3, input_length = 1, name = \"embedding_season\") %&gt;% layer_flatten()\nembedding_out8 = inp_ltv_uw %&gt;% layer_embedding(input_dim = 2+1, output_dim = 2, input_length = 1, name = \"embedding_ltv_uw\") %&gt;% layer_flatten()\n\ncombined_model = layer_concatenate(c(embedding_out1, embedding_out2, embedding_out3, embedding_out4,\n                                     embedding_out5, embedding_out6, embedding_out7, embedding_out8, inp_continuous)) %&gt;%\n                                     layer_dense(units=32, activation = \"relu\") %&gt;%\n                                     layer_dropout(0.3) %&gt;%\n                                     layer_dense(units=10, activation = \"relu\") %&gt;%\n                                     layer_dropout(0.15) %&gt;%\n                                     layer_dense(units=1)\n\nmodel = keras::keras_model(inputs = c(inp_lender, inp_channel, inp_fthb, inp_region, inp_cs_null,\n                                      inp_source, inp_season, inp_ltv_uw, inp_continuous),\n                           outputs = combined_model)\n\nmodel %&gt;% compile(loss = \"mean_squared_error\", optimizer = \"sgd\", metric = \"accuracy\")\n\ninputVariables = list(as.matrix(df_cl_dl$lender),\n                      as.matrix(df_cl_dl$channel),\n                      as.matrix(df_cl_dl$fthb),\n                      as.matrix(df_cl_dl$region),\n                      as.matrix(df_cl_dl$credit_score_null),\n                      as.matrix(df_cl_dl$source),\n                      as.matrix(df_cl_dl$seasonality),\n                      as.matrix(df_cl_dl$ltv_uw),\n                      as.matrix(df_cl_dl[,c(\"credit_score\", \"current_ltv\", \"loan_age\", \"cash_incentive_a\", \"hpas\", \"loan_size\", \"ur\", \"risk\")]))\n\nmodel %&gt;% fit(x = inputVariables, y = as.matrix(df_cl_dl$vsmm), epochs = 10, batch_size = 2)\n<\/code><\/pre>\n\n<p>Error Massage:<\/p>\n\n<p>Error in py_call_impl(callable, dots$args, dots$keywords) : ValueError: could not convert string to float: Null<\/p>\n","answers":[]}
+{"QuestionId":57796262,"AnswerCount":0,"Tags":"<tensorflow><reinforcement-learning><tensor><softmax><ragged>","CreationDate":"2019-09-04T22:16:18.293","AcceptedAnswerId":null,"OwnerUserId":4986655.0,"Title":"How to mask logits for tf.softmax_cross_entropy_with_logits to implement valid actions","Body":"<p>I want to compute the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/softmax_cross_entropy_with_logits_v2\" rel=\"nofollow noreferrer\">softmax_cross_entropy_with_logits<\/a> of a batch tensor.\nI have a batch of logits tensor as input, however I want to mask this tensor before with a boolean mask. The boolean mask is also a batch of masks, in every mask there can be a different amount of <code>True<\/code>.\nThus applying this mask to the whole tensor will not be dense afterwards.<br>\nTrying this either flattens the tensor (<code>tf.boolean_mask<\/code>) or creates a ragged one (<code>tf.ragged.boolean_mask<\/code>), which both produce wrong results or don't work with the softmax function.<\/p>\n\n<p>So basically I want to make the following code work:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># logits.shape = (batch, outputs), e.g. (512,8)\n# mask.shape = (batch, valid), e.g. (512,8)\n# expected result shape (512,)\none_hot_actions = tf.one_hot(x, logits.get_shape().as_list()[-1])\nstopgradient = tf.stop_gradient(one_hot_actions)\nreturn tf.nn.softmax_cross_entropy_with_logits_v2(\n    logits=tf.boolean_mask(logits, mask),\n    labels=tf.boolean_mask(stopgradient, mask))\n<\/code><\/pre>\n\n<p>But with <code>tf.boolean_mask<\/code> this produces just one value, not four and with <code>tf.ragged.boolean_mask<\/code> this function does not work.<\/p>\n\n<p>I tried combing the two ragged tensors row wise (first masked logits row with first masked labels row) and compute the softmax rowwise. This did not work since the tf.map_fn that I used does not accept ragged tensors as inputs. I tried this same idea also with SparseTensors and list of Tensors (tf.split) but never got any working code out of it.<\/p>\n\n<p>The only idea I had to solve this issue is very ugly.\nReplace all masked values with <code>tf.where<\/code> to <code>NaN<\/code> and then use map_fn on these now dense tensors. Every row I can then mask again to exclude NaN and now can call the softmax function row-wise again.<br>\n<strong>EDIT<\/strong> This is what the code currently looks like:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>stopgradient = tf.stop_gradient(one_hot_actions)\nnan_logits = tf.where(mask, logits, float('NaN') + tf.zeros_like(self.logits))\nnan_labels = tf.where(mask, stopgradient, float('NaN') + tf.zeros_like(stopgradient))\nnan_lola = tf.stack([nan_logits, nan_labels], axis=1)\ndef fn(x):\n    nan_lo = x[0]\n    nan_la = x[1]\n    masked_lo = tf.boolean_mask(nan_lo, tf.logical_not(tf.math.is_nan(nan_lo)))\n    masked_la = tf.boolean_mask(nan_la, tf.logical_not(tf.math.is_nan(nan_la)))\n    return tf.nn.softmax_cross_entropy_with_logits_v2(\n        logits=masked_lo,\n        labels=masked_la\n    )\nresult = tf.map_fn(fn, nan_lola)\nreturn result\n<\/code><\/pre>\n\n<p>This works but is very slow (my training time almost doubles).<\/p>\n\n<p>To those interested: I stumbled upon this problem trying to mask valid actions in reinforcement learning.<\/p>\n\n<p>Do you know of any way to do this (faster)?\nCan you replace the masked values with a value that does not affect the softmax?\nThank you!<\/p>\n","answers":[]}
+{"QuestionId":57796344,"AnswerCount":0,"Tags":"<python><tensorflow><keras><tensorflow-datasets>","CreationDate":"2019-09-04T22:32:26.863","AcceptedAnswerId":null,"OwnerUserId":1273818.0,"Title":"Tensorflor 2.0 - from_generator shape mismatch","Body":"<p>I'm trying to use the <code>Dataset.from_generator<\/code> function to produce data for training and testing. Currently I'm facing a problem regarding data shape matching saying:<\/p>\n\n<pre><code>InvalidArgumentError:  Incompatible shapes: [7,3] vs. [7]\n     [[node metrics_16\/sparse_categorical_accuracy\/Equal (defined at &lt;ipython-input-13-d214206d5c0a&gt;:40) ]] [Op:__inference_keras_scratch_graph_9143]\n\nFunction call stack:\nkeras_scratch_graph\n<\/code><\/pre>\n\n<p>Here is the complete minimal code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport pandas as pd\nimport random\nimport tensorflow as tf\n\nINPUT_SHAPE=[3, 5]\nNUM_POINTS=20\nBATCH_SIZE=7\nEPOCHS=3\n\ndef data_gen(num=10, in_shape=[5, 3]):\n    print(\"IN_SHAPE: \", in_shape)\n    for i in range(num):\n        res = np.random.rand(in_shape[0], in_shape[1]), random.randint(0,2)\n        print(\"Output shape of el: \", res[0].shape, \" and training pair (x,y):\\n\", res)\n        yield res\n\ntrain = tf.data.Dataset.from_generator(\n    generator=data_gen,\n    output_types=(tf.float32, tf.int32),\n#     output_shapes=(tf.TensorShape([None, INPUT_SHAPE[1]]), tf.TensorShape(None)),\n#     output_shapes=(tf.TensorShape(INPUT_SHAPE), tf.TensorShape(None)),\n    args=([NUM_POINTS, INPUT_SHAPE])\n)\n\ndef create_model(input_shape):\n    model = tf.keras.models.Sequential([\n        tf.keras.layers.Dense(100, activation=\"tanh\",input_shape=input_shape),        \n        tf.keras.layers.Dense(3, activation=\"softmax\", kernel_regularizer= tf.keras.regularizers.l2(0.001))\n    ])\n    return model\n\nmodel = create_model(input_shape=INPUT_SHAPE)\nmodel.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=1e-4, clipvalue=1.0),\n          loss= tf.keras.losses.SparseCategoricalCrossentropy(), #'sparse_categorical_crossentropy',\n          metrics=[tf.keras.metrics.SparseCategoricalAccuracy()])\nprint(model.summary())\nmodel.fit(train.batch(BATCH_SIZE), epochs=EPOCHS, verbose=2)\nmodel.evaluate(train, steps=None, verbose=1)\n<\/code><\/pre>\n\n<p>I'm running the code on:  <\/p>\n\n<ul>\n<li>python=3.7.4 (default, Aug 17 2019, 20:42:51)\n[Clang 10.0.1\n(clang-1001.0.46.4)] <\/li>\n<li>tensorflow=2.0.0-beta1 <\/li>\n<li>numpy=1.17.0<\/li>\n<li>pandas=0.24.2<\/li>\n<\/ul>\n\n<p><strong>Later edit: made the example shorter<\/strong><\/p>\n","answers":[]}
+{"QuestionId":57796880,"AnswerCount":0,"Tags":"<python><tensorflow><loss-function>","CreationDate":"2019-09-04T23:59:38.963","AcceptedAnswerId":null,"OwnerUserId":6181351.0,"Title":"Custom loss function tensorflow","Body":"<p>I'm try to test custom loss function. <\/p>\n\n<p>i test how the network is learning base on tf loss function, and past it by me.\nthe result show that the network dont learn if i pass the function.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport matplotlib.pyplot as plt\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\ndef loss(y_true, y_pred):\n  return tf.keras.losses.sparse_categorical_crossentropy(y_true, y_pred)\n\nmodel = tf.keras.models.Sequential([\n  tf.keras.layers.Flatten(input_shape=(28, 28)),\n  tf.keras.layers.Dense(128, activation='relu'),\n  tf.keras.layers.Dropout(0.2),\n  tf.keras.layers.Dense(10, activation='softmax')\n])\n\nmodel.compile(optimizer='adam',\n              loss=loss,\n              metrics=['accuracy'])\n\nmodel.fit(x_train, y_train, epochs=7)\nmodel.evaluate(x_test, y_test)\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport matplotlib.pyplot as plt\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\ndef loss(y_true, y_pred):\n  return tf.keras.losses.sparse_categorical_crossentropy(y_true, y_pred)\n\nmodel = tf.keras.models.Sequential([\n  tf.keras.layers.Flatten(input_shape=(28, 28)),\n  tf.keras.layers.Dense(128, activation='relu'),\n  tf.keras.layers.Dropout(0.2),\n  tf.keras.layers.Dense(10, activation='softmax')\n])\n\nmodel.compile(optimizer='adam',\n              loss=tf.keras.losses.sparse_categorical_crossentropy,\n              metrics=['accuracy'])\n\nmodel.fit(x_train, y_train, epochs=7)\nmodel.evaluate(x_test, y_test)\n<\/code><\/pre>\n\n<p>for the first code:\nEpoch 1\/7<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>60000\/60000 [==============================] - 4s 66us\/sample - loss: 0.3035 - acc: 0.1015\nEpoch 2\/7\n60000\/60000 [==============================] - 4s 66us\/sample - loss: 0.1481 - acc: 0.0994\nEpoch 3\/7\n60000\/60000 [==============================] - 4s 69us\/sample - loss: 0.1097 - acc: 0.0990\nEpoch 4\/7\n60000\/60000 [==============================] - 4s 67us\/sample - loss: 0.0875 - acc: 0.0991\nEpoch 5\/7\n60000\/60000 [==============================] - 4s 66us\/sample - loss: 0.0770 - acc: 0.0989\nEpoch 6\/7\n60000\/60000 [==============================] - 4s 66us\/sample - loss: 0.0665 - acc: 0.0991\nEpoch 7\/7\n60000\/60000 [==============================] - 4s 66us\/sample - loss: 0.0598 - acc: 0.0988\n10000\/10000 [==============================] - 0s 36us\/sample - loss: 0.0692 - acc: 0.0993\n<\/code><\/pre>\n\n<p>for the second code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Epoch 1\/7\n60000\/60000 [==============================] - 4s 65us\/sample - loss: 0.3002 - acc: 0.9119\nEpoch 2\/7\n60000\/60000 [==============================] - 4s 63us\/sample - loss: 0.1458 - acc: 0.9561\nEpoch 3\/7\n60000\/60000 [==============================] - 4s 64us\/sample - loss: 0.1066 - acc: 0.9680\nEpoch 4\/7\n60000\/60000 [==============================] - 4s 64us\/sample - loss: 0.0875 - acc: 0.9727\nEpoch 5\/7\n60000\/60000 [==============================] - 4s 68us\/sample - loss: 0.0738 - acc: 0.9768\nEpoch 6\/7\n60000\/60000 [==============================] - 4s 64us\/sample - loss: 0.0662 - acc: 0.9787\nEpoch 7\/7\n60000\/60000 [==============================] - 4s 64us\/sample - loss: 0.0575 - acc: 0.9817\n10000\/10000 [==============================] - 0s 35us\/sample - loss: 0.0630 - acc: 0.9814\n<\/code><\/pre>\n\n<p>and its look like the loss is update but the acc is not learning from custom loss function<\/p>\n","answers":[]}
+{"QuestionId":57797113,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-05T00:44:02.160","AcceptedAnswerId":58966376.0,"OwnerUserId":4036855.0,"Title":"AttributeError: module 'keras.backend' has no attribute 'image_dim_ordering'","Body":"<p>I tried to execute some tutorial transfer learning project.\nBut I've got attribute error.<\/p>\n\n<p>I checked my tensorflow and keras version.<\/p>\n\n<p>tensorflow : 1.14.0\nkeras : 2.2.5<\/p>\n\n<p>and python 3.6.9 version.<\/p>\n\n<p>the code is here.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>if(K.image_dim_ordering() == 'th'):\n  input_tensor = Input(shape=(3, 299, 299))\n<\/code><\/pre>\n\n<p>error message here.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>AttributeError: module 'keras.backend' has no attribute 'image_dim_ordering'\n<\/code><\/pre>\n","answers":[{"AnswerId":"57797201","CreationDate":"2019-09-05T01:04:08.713","ParentId":null,"OwnerUserId":"4036855","Title":null,"Body":"<p>Self answer for who has same issue with me.\nimage_dim_ordering change to image_data_format from Keras 2.x.\nSince change image_dim_ordering to image_data_format.<\/p>\n\n<p>reference link<\/p>\n\n<ul>\n<li><p><a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/12649\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/12649<\/a><\/p><\/li>\n<li><p><a href=\"https:\/\/keras.io\/backend\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/backend\/<\/a><\/p><\/li>\n<\/ul>\n"},{"AnswerId":"58966376","CreationDate":"2019-11-21T02:30:14.550","ParentId":null,"OwnerUserId":"11635246","Title":null,"Body":"<p><code>keras.backend.common<\/code> module has <code>image_dim_ordering()<\/code> <\/p>\n\n<pre><code>if(K.common.image_dim_ordering() == 'th'):\n\n    input_tensor = Input(shape=(3, 299, 299))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57797242,"AnswerCount":0,"Tags":"<tensorflow><ubuntu><keras><installation>","CreationDate":"2019-09-05T01:10:34.250","AcceptedAnswerId":null,"OwnerUserId":6227035.0,"Title":"Tensorflow-gpu + Keras installation error: Failed to load the native TensorFlow runtime","Body":"<p>I have installed tensorflow-gpu==1.10.0 for GPU usage on Ubuntu but when I call Keras ( e.g. import keras) I get the error: Failed to load the native TensorFlow runtime<\/p>\n\n<p>I have been looking around the web but I could not find a clear answer. This is the error I get: <\/p>\n\n<pre><code>Using TensorFlow backend.\nTraceback (most recent call last):\n  File \"\/home\/boxx\/.local\/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\/boxx\/.local\/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\/boxx\/.local\/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 \"\/usr\/lib\/python3.6\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/usr\/lib\/python3.6\/imp.py\", line 343, in load_dynamic\n    return _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):\n  File \"sltmSV.py\", line 8, in &lt;module&gt;\n    from keras.models import Sequential\n  File \"\/home\/boxx\/.local\/lib\/python3.6\/site-packages\/keras\/__init__.py\", line 3, in &lt;module&gt;\n    from . import utils\n  File \"\/home\/boxx\/.local\/lib\/python3.6\/site-packages\/keras\/utils\/__init__.py\", line 6, in &lt;module&gt;\n    from . import conv_utils\n  File \"\/home\/boxx\/.local\/lib\/python3.6\/site-packages\/keras\/utils\/conv_utils.py\", line 9, in &lt;module&gt;\n    from .. import backend as K\n  File \"\/home\/boxx\/.local\/lib\/python3.6\/site-packages\/keras\/backend\/__init__.py\", line 1, in &lt;module&gt;\n    from .load_backend import epsilon\n  File \"\/home\/boxx\/.local\/lib\/python3.6\/site-packages\/keras\/backend\/load_backend.py\", line 89, in &lt;module&gt;\n    from .tensorflow_backend import *\n  File \"\/home\/boxx\/.local\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 5, in &lt;module&gt;\n    import tensorflow as tf\n  File \"\/home\/boxx\/.local\/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\/boxx\/.local\/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\/boxx\/.local\/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\/boxx\/.local\/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\/boxx\/.local\/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\/boxx\/.local\/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 \"\/usr\/lib\/python3.6\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/usr\/lib\/python3.6\/imp.py\", line 343, in load_dynamic\n    return _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>Can someone point me a method or some documentation to install keras + tensorFlow for GPU on Ubuntu?\nThank you!<\/p>\n","answers":[]}
+{"QuestionId":57797936,"AnswerCount":0,"Tags":"<keras><neural-network><attention-model>","CreationDate":"2019-09-05T03:10:31.407","AcceptedAnswerId":null,"OwnerUserId":12022850.0,"Title":"When the scale matrix is connected, the neural network starts learning again instead of checking the offer with the predict","Body":"<p>I trained the neural network, saved the weight in h5 format, and now that I want to test its performance with predict, it's starting to learn again. I have a custom Attention layer, at first I had a ValueError: Unknown layer: Attention bug, I took advice from <a href=\"https:\/\/stackoverflow.com\/questions\/50837728\/valueerror-unknown-layer-capsulelayer\">ValueError: Unknown layer: CapsuleLayer<\/a>, but after I did what I said, the NS started learning again. I don't know how to deal with this problem.<\/p>\n\n<pre><code>max_features = 100000\nmaxlen = 150\n\nmodel = load_model('C:\\\\Users\\\\\u041c\u0430\u0440\u0438\u044f\\\\PycharmProjects\\\\Attention.h5')\n\nmax_review_length = 250\ntext = numpy.array(['Cool day!'])\n\ntk = keras.preprocessing.text.Tokenizer(num_words=None, \nfilters='!\"#$%&amp;()*+,-.\/:;&lt;=&gt;?@[\\]^_`{|}~ ', lower=True, split=' ', char_level=False, oov_token=None, document_count=0)\ntk.fit_on_texts(text)\nprediction = model.predict(sequence.pad_sequences(tk.texts_to_sequences(text),maxlen=max_review_length))\nprint(prediction)\n<\/code><\/pre>\n\n<p>I want the predict function to give me an estimate of my text when I connect the scale.<\/p>\n","answers":[]}
+{"QuestionId":57798018,"AnswerCount":0,"Tags":"<tensorflow><keras><cross-entropy>","CreationDate":"2019-09-05T03:22:13.003","AcceptedAnswerId":null,"OwnerUserId":10038330.0,"Title":"Why does this not perform equivalently to keras\/tensorflows sparse categorical crossentropy?","Body":"<p>Im using <\/p>\n\n<pre><code>def celoss(y_true, y_pred):\n    bs = tf.shape(y_true)[0]\n    lin = tf.range(bs)\n    y_true = tf.squeeze(y_true, -1)\n    y_true = tf.cast(y_true, tf.int32)\n    y_true = tf.stack([lin, y_true], -1)\n    p = tf.gather_nd(y_pred, y_true)\n    return -tf.math.log(p + 1e-8) \n<\/code><\/pre>\n\n<p>and evaluation-wise it is always within 1e-3 of keras' sparse-categorical-crossentropy, but in training it does considerably worse on a variety of problems that I have tested on regardless that its forward pass is the same. So why isn't it's back-pass? <\/p>\n","answers":[]}
+{"QuestionId":57798033,"AnswerCount":1,"Tags":"<size><conv-neural-network><pytorch>","CreationDate":"2019-09-05T03:24:58.970","AcceptedAnswerId":null,"OwnerUserId":10240085.0,"Title":"ValueError: Target size (torch.Size([16])) must be the same as input size (torch.Size([16, 1]))","Body":"<pre><code>ValueError                                Traceback (most recent call last)\n&lt;ipython-input-30-33821ccddf5f&gt; in &lt;module&gt;\n     23         output = model(data)\n     24         # calculate the batch loss\n---&gt; 25         loss = criterion(output, target)\n     26         # backward pass: compute gradient of the loss with respect to model parameters\n     27         loss.backward()\n\nC:\\Users\\mnauf\\Anaconda3\\envs\\federated_learning\\lib\\site-packages\\torch\\nn\\modules\\module.py in __call__(self, *input, **kwargs)\n    487             result = self._slow_forward(*input, **kwargs)\n    488         else:\n--&gt; 489             result = self.forward(*input, **kwargs)\n    490         for hook in self._forward_hooks.values():\n    491             hook_result = hook(self, input, result)\n\nC:\\Users\\mnauf\\Anaconda3\\envs\\federated_learning\\lib\\site-packages\\torch\\nn\\modules\\loss.py in forward(self, input, target)\n    593                                                   self.weight,\n    594                                                   pos_weight=self.pos_weight,\n--&gt; 595                                                   reduction=self.reduction)\n    596 \n    597 \n\nC:\\Users\\mnauf\\Anaconda3\\envs\\federated_learning\\lib\\site-packages\\torch\\nn\\functional.py in binary_cross_entropy_with_logits(input, target, weight, size_average, reduce, reduction, pos_weight)\n   2073 \n   2074     if not (target.size() == input.size()):\n-&gt; 2075         raise ValueError(\"Target size ({}) must be the same as input size ({})\".format(target.size(), input.size()))\n   2076 \n   2077     return torch.binary_cross_entropy_with_logits(input, target, weight, pos_weight, reduction_enum)\n\nValueError: Target size (torch.Size([16])) must be the same as input size (torch.Size([16, 1]))\n<\/code><\/pre>\n\n<p>I am training a CNN. Working on the Horses vs humans dataset. <a href=\"https:\/\/github.com\/mnauf\/horses_humans\/blob\/master\/Untitled.ipynb\" rel=\"nofollow noreferrer\">This is my code<\/a>. I am using <code>criterion = nn.BCEWithLogitsLoss()<\/code> and <code>optimizer = optim.RMSprop(model.parameters(), lr=0.01<\/code>). My final layer is <code>self.fc2 = nn.Linear(512, 1)<\/code>. Out last neuron, will output 1 for horse and 0 for human, right? or should I choose 2 neurons for output?<\/p>\n\n<p><code>16<\/code> is the batch size. Since the error says <code>ValueError: Target size (torch.Size([16])) must be the same as input size (torch.Size([16, 1]))<\/code>. I don't understand, where do I need to make change, to rectify the error.<\/p>\n","answers":[{"AnswerId":"57802321","CreationDate":"2019-09-05T09:23:33.083","ParentId":null,"OwnerUserId":"10240085","Title":null,"Body":"<p><code>target = target.unsqueeze(1)<\/code>, before passing target to criterion, changed the target tensor size from <code>[16]<\/code> to <code>[16,1]<\/code>. Doing it solved the issue. Furthermore, I also needed to do <code>target = target.float()<\/code> before passing it to criterion, because our outputs are in float. Besides, there was another error in the code. I was using sigmoid activation function in the last layer, but I shouldn\u2019t because the criterion I am using already comes with sigmoid builtin.<\/p>\n"}]}
+{"QuestionId":57798047,"AnswerCount":0,"Tags":"<python><python-2.7><tensorflow>","CreationDate":"2019-09-05T03:27:27.323","AcceptedAnswerId":null,"OwnerUserId":12018583.0,"Title":"How to use multiple input pipelines for data parallelism of multi-gpu in tensorflow?","Body":"<p>I'm codeing for a GAN. Using the data parallelism over multiple GPUs to train.\nTo find more detail and the difference when training discriminator, I want to set multi indenpendent pipelines and each gpu getting data(every Image category is different, such as feed GPU0 with  dog pictures, and feed GPU\n1 with car pictures ) from each pipeline train synchronously.Finally, the model must do gradient-average successfully. <\/p>\n\n<p>I have just done with one pipeline to multi-GPU. Actually, I'm not sure if this question is  feasible in tensorflow. Could you please give me some code example and some thinking for my question?<\/p>\n","answers":[]}
+{"QuestionId":57798219,"AnswerCount":0,"Tags":"<python-2.7><pytorch><opennmt>","CreationDate":"2019-09-05T03:55:31.020","AcceptedAnswerId":null,"OwnerUserId":10716900.0,"Title":"How to take multi-GPU support to the OpenNMT-py (pytorch)?","Body":"<p>I used python-2.7 version to run the PyTorch with GPU support. I used this command to train the dataset using multi-GPU. \nCan someone please tell me how can I fix this error with PyTorch in OpenNMT-py or is there a way to take pytorch support for multi-GPU using python 2.7?<\/p>\n\n<p>Here is the command that I tried.<\/p>\n\n<blockquote>\n  <blockquote>\n    <p>CUDA_VISIBLE_DEVICES=1,2\n    python train.py -data data\/demo -save_model demo-model -world_size 2 -gpu_ranks 0 1<\/p>\n  <\/blockquote>\n<\/blockquote>\n\n<p>This is the error:<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):\n   File \"train.py\", line 200, in \n    main(opt)\n  File \"train.py\", line 60, in main\n  mp = torch.multiprocessing.get_context('spawn')\n  AttributeError: 'module' object has no attribute 'get_context'<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":57798388,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-05T04:21:00.717","AcceptedAnswerId":null,"OwnerUserId":1009508.0,"Title":"Keras model won't use predict correctly","Body":"<p>I have a Keras model that I have trained and evaluated and even tested. Now I am trying to use three test images into the model. <\/p>\n\n<p>I run the images through a preprocessor which is the same one I used to make the training data. I then do the exact same thing to the single images that I did for the testing data. But it gives me an error of <\/p>\n\n<blockquote>\n  <p>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 2 arrays:<\/p>\n<\/blockquote>\n\n<p>I don't know what is wrong with it. <\/p>\n\n<p>So this is how I test the model successfully.<\/p>\n\n<pre><code>y_pred = []\ny_true = []\nfor i in range(0, len(test_x1)):\n    x1 = test_x1[i]\n    x2 = test_x2[i]\n    x1 = np.expand_dims(x1, axis=0)\n    x2 = np.expand_dims(x2, axis=0)\n    y_true.append(np.argmax(test_y[i]))\n    pred = model.predict([x1, x2])\n    y_pred.append(make_binary(pred))\n<\/code><\/pre>\n\n<p>This is the preprocessing method I used for both images <\/p>\n\n<pre><code>def create_features(file, image_dir, base_model):\n    img_path = os.path.join(image_dir, file)\n    img = image.load_img(img_path, target_size=(224, 224))\n    img = image.img_to_array(img)\n    x = resnet50.preprocess_input(img)\n    x = np.array([x])\n    feature = base_model.predict(x)\n    return feature\n<\/code><\/pre>\n\n<p>And this is the way I am processing the new images:<\/p>\n\n<pre><code>IMAGE_DIR = 'Data'\nimg1 = 'test1.jpg'\nimg2 = 'test2.jpg'\nimg3 = 'test3.jpg'\n\nimg1_feat = create_features(img1, IMAGE_DIR, model)\nimg2_feat = create_features(img2, IMAGE_DIR, model)\nimg3_feat = create_features(img3, IMAGE_DIR, model)\n<\/code><\/pre>\n\n<p>Now when I look at the two features they are the same.<\/p>\n\n<pre><code>x1 = test_x1[0]\nx1 = np.expand_dims(x1, axis=0)\nprint(x1.shape)\nprint(type(x1))\n\nprint(img1_feat.shape)\nprint(type(img1_feat))\n(1, 1, 1000)\n&lt;class 'numpy.ndarray'&gt;\n(1, 1, 1000)\n&lt;class 'numpy.ndarray'&gt;\n<\/code><\/pre>\n\n<p>And then I try to make a prediction from it<\/p>\n\n<pre><code>pred1 = model.predict([img1_feat, img2_feat])\n<\/code><\/pre>\n\n<p>But that results in an error. <\/p>\n","answers":[{"AnswerId":"57803781","CreationDate":"2019-09-05T10:47:06.700","ParentId":null,"OwnerUserId":"1009508","Title":null,"Body":"<p>I figured out what was wrong thanks to @Matias Valdenegro and @Mukul\nI was doing this in an Ipython notebook and after a few epochs go around I found out that on occasion the model gets overwritten by an imported resent model from another class. \nThanks to everyone for the help. I didnt think about using the model.summary() as I didnt really think that it has changed.  <\/p>\n"}]}
+{"QuestionId":57798859,"AnswerCount":0,"Tags":"<tensorflow><tensorrt>","CreationDate":"2019-09-05T05:19:44.817","AcceptedAnswerId":null,"OwnerUserId":2467772.0,"Title":"Depthwise_conv2d in TensorRT","Body":"<p>In conversion to TensorRT from Tensorflow, there is issue at Depthwise_con2d.<\/p>\n\n<p>The implementation is as in the picture.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/C2ZTP.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/C2ZTP.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Tensorflow code is as follow.<\/p>\n\n<pre><code>    def gauss_kernel(self, kernlen=21, nsig=3, channels=1):\n        interval = (2*nsig+1.)\/(kernlen)\n        x = np.linspace(-nsig-interval\/2., nsig+interval\/2., kernlen+1)\n        kern1d = np.diff(st.norm.cdf(x))\n        kernel_raw = np.sqrt(np.outer(kern1d, kern1d))\n        kernel = kernel_raw\/kernel_raw.sum()\n        out_filter = np.array(kernel, dtype = np.float32)\n        out_filter = out_filter.reshape((kernlen, kernlen, 1, 1))\n        out_filter = np.repeat(out_filter, channels, axis = 2)\n        return out_filter\n\n    def make_gauss_var(self, name, size, sigma, c_i):\n        # with tf.device(\"\/cpu:0\"):\n        kernel = self.gauss_kernel(size, sigma, c_i)\n        var = tf.Variable(tf.convert_to_tensor(kernel), name=name)\n        return var\n    @layer\n    def smoother(self, input, name='gaussian_heatMat'):\n        # Get the number of channels in the input\n        c_i = input.get_shape().as_list()[3]\n        # Convolution for a given input and kernel\n        convolve = lambda i, k: tf.nn.depthwise_conv2d(i, k, [1, 1, 1, 1], padding='SAME')\n        with tf.variable_scope(name) as scope:\n            kernel = self.make_gauss_var('gauss_weight', 25, 3.0, c_i)#filter_size=25,sigma=3.0\n            output = convolve(input, kernel)\n        return output\n<\/code><\/pre>\n\n<p>I have errors in parsing to network from uff as<\/p>\n\n<pre><code>[TensorRT] ERROR: gaussian_heatMat\/depthwise: kernel weights has count 11875 but 7796358 was expected\n[TensorRT] ERROR: gaussian_heatMat\/depthwise: count of 11875 weights in kernel, but kernel dimensions (25, 25) with 1459200 input channels, 19 output channels and 19 groups were specified.\n[TensorRT] ERROR: UffParser: Parser error: MarkOutput_0: Order size is not matching the number dimensions of TensorRT\n<\/code><\/pre>\n\n<p>11875 = 25*25*19 so that is kernel. Why 7796358 was expected? Don't under the other errors as well. <\/p>\n","answers":[]}
+{"QuestionId":57799159,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-05T05:52:58.067","AcceptedAnswerId":null,"OwnerUserId":8154236.0,"Title":"Issues about Freezing the exported Graph","Body":"<p>I follow <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/slim\/README.md#freezing-the-exported-graph\" rel=\"nofollow noreferrer\">documentation<\/a> to build tool <strong>freeze_graph<\/strong>. But after building it successfully, I didn't find: <\/p>\n\n<pre><code>bazel-bin\/tensorflow\/python\/tools\/freeze_graph\n<\/code><\/pre>\n\n<p>However, I found:<\/p>\n\n<pre><code>bazel-out\/k8-py2-opt\/bin\/tensorflow\/python\/tools\/freeze_graph\n<\/code><\/pre>\n\n<p>It is different from the documentation. I'm not sure whether I did something wrong or the documentation is not correct?<\/p>\n","answers":[]}
+{"QuestionId":57799338,"AnswerCount":1,"Tags":"<keras><neural-network><deep-learning><constraints><regularized>","CreationDate":"2019-09-05T06:09:18.210","AcceptedAnswerId":57800524.0,"OwnerUserId":2511398.0,"Title":"output layer regularization implementation","Body":"<p>I\u2019m building a NN model using keras, and I wish to impose a constraint on it that doesn\u2019t (directly) have to do with the weights. Would be very grateful for some help \/ points me towards some relevant keywords to look up. The constraint I wish to impose is a bit complex, but it can be simplified in the following manner: I wish to impose a constraint on the output of certain inputs of the net. For the sake of simplicity, let\u2019s say the constraint looks like NN(3)+NN(4) &lt; 10, where NN is the neural net, which can be seen as a function. How can I impose such a constraint? Thank you very much in advance for any help on the subject!<\/p>\n\n<p>edit: A more detailed explanation of what I'm trying to do and why.\nThe theoretical model I'm building is this:\n<a href=\"https:\/\/i.stack.imgur.com\/J7L78.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/J7L78.png\" alt=\"Theoretical Model\"><\/a>\nI'm feeding the output of the first net into the input of the second net, along with an additive gaussian noise.\nThe constraint I wish to impose is on the output of the first NN (g). Why? Without a constraint, the net maps the inputs to outputs as high as it possibly can in order to make the additive noise as insignificant as possible. And rightly so, this <em>is<\/em> the optimal encoding function g, but it's not very interesting :) And so I wish to impose a constraint on the output of the first NN (g). More specifically, the constraint is on the total power of the function: integral{ fX(x) * g(x)^2 dx }. But this can be simplified more or less, to a function that looks something like what I described earlier - g(3)+g(4)&lt;10. More specifically, the function is sum { fX(x) * g(i)^2 * dx } &lt; max_power, for some sampled inputs i.\nThis is the problem, now here's how I attempted to implement it:<\/p>\n\n<pre><code>model = Sequential([\n        Dense(300, input_dim=1, activation='relu'),\n        Dense(300, activation='relu'),\n        Dense(1, activation='linear', name=encoder_output),\n        GaussianNoise(nvar, name='noise'),\n        Dense(300, activation='relu', name=decoder_input),\n        Dense(300, activation='relu'),\n        Dense(1, activation='linear', name=decoder_output),\n    ])\n<\/code><\/pre>\n\n<p>Mainly, this is supposedly a single neural net, and not really 2 (although there is no difference obviously).\nThe import things to note is the input dim 1, output dim 1 (x and y in the diagram), and the gaussian noise in the middle. The hidden layers are not very interesting right now, I'll optimize them at a later point.\nIn this model, I wish to impose a constraint on the output of a (supposedly) hidden layer named encoder_output. Hope this clarifies things.<\/p>\n","answers":[{"AnswerId":"57800524","CreationDate":"2019-09-05T07:34:08.307","ParentId":null,"OwnerUserId":"2042297","Title":null,"Body":"<p>You could use a multi input\/multi output model with shared weights layers. The model could for example look like this:<\/p>\n\n<pre><code>from keras.layers import Input, Dense, Add\nfrom keras.models import Model\n\n# Shared weights layers\nhidden = Dense(10, activation='relu')\nnn_output = Dense(1, activation='relu')\n\nx1 = Input(shape=(1,))\nh1 = hidden(x1)\ny1 = nn_output(h1)\n\nx2 = Input(shape=(1,))\nh2 = hidden(x2)\ny2 = nn_output(h2)\n\n# Your constraint\n# In case it should be more complicated, you can implement\n# a custom keras layer\nsum = Add()([y1, y2]) \n\nmodel = Model(inputs=[x1, x2], outputs=[y1, y2, sum])\nmodel.compile(optimizer='sgd', loss='mse')\n\nX_train_1 = [3,4]\nX_train_2 = [4,3]\ny_train_1 = [123,456] # your expected output\ny_train_2 = [456,123] # your expected output\ns = [10,10] # expected sums\n\nmodel.fit([X_train_1, X_train_2], [y_train_1, y_train_2, s], epochs=10)\n<\/code><\/pre>\n\n<p>If you have no exact value for your constraint that can be used as an expected output, you can remove it from the outputs and write a simple custom regularizer that would be used on it. There is a simple example for a custom regularizer in the Keras documentation.<\/p>\n"}]}
+{"QuestionId":57799367,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-09-05T06:12:25.733","AcceptedAnswerId":null,"OwnerUserId":8186207.0,"Title":"Can I disable image resizing when using ImageGenerator?","Body":"<p>The resize step in the Keras image preprocessing function flow_from_dataframe is done before the self-defined preprocessing function. However, I want to resize the image after I finish all the pre-processing. How can I disable the resize? How should I set the target_size parameter?<\/p>\n","answers":[]}
+{"QuestionId":57799794,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-09-05T06:44:21.150","AcceptedAnswerId":57799960.0,"OwnerUserId":11905830.0,"Title":"PyTorch dot product does not work for some cases","Body":"<p>I am struggling with an issue I cannot solve in PyTorch code.  My task is simple: just get the inner product of two tensors.  However, the output becomes zero in certain cases instead of returning the correct value.<\/p>\n\n<p>For example, the code below demonstrates the issue.<\/p>\n\n<pre><code>a = torch.tensor([3,1])\nb = torch.tensor([3.,1.])\na.dot(a)  # this returns 10, which is correct\nb.dot(b)  # this returns 0., which is not correct\n<\/code><\/pre>\n\n<p>I am using PyTorch version 1.2 and this has never happened to me before...  Is there something simple that I am missing around the way I define the tensors?<\/p>\n\n<p>Thank you for your time!<\/p>\n","answers":[{"AnswerId":"57799960","CreationDate":"2019-09-05T06:55:46.703","ParentId":null,"OwnerUserId":"848510","Title":null,"Body":"<p>I just ran the same code using torch version <code>'1.0.1.post2'<\/code> and saw the results <code>tensor(10)<\/code>and <code>tensor(10.)<\/code> respectively for <code>a.dot(a)<\/code> and <code>b.dot(b)<\/code>. Please confirm if you are missing something else in the code.<\/p>\n"}]}
+{"QuestionId":57800176,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-05T07:11:07.527","AcceptedAnswerId":null,"OwnerUserId":12023618.0,"Title":"I want to get the weights and bias of model,but i just get the bias of model","Body":"<p>I want to quantize the model, so I must get the weights and bias of the model that is someone who finish training the model give me, but when I use the <code>NewCheckpointReader<\/code> way to get the weights and bias, I just can get the bias, but I can not get the weights. What should I do to resolve the problem or tell if the model has some problem. Thanks for your time. <\/p>\n\n<pre><code>ubuntu 16.04\ntensorflow 1.14\n<\/code><\/pre>\n\n<p>code<\/p>\n\n<pre><code>#\u4ece\u8be5\u6a21\u578b\u4e2d\u8bfb\u53d6\u5185\u5bb9\nreader = pywrap_tensorflow.NewCheckpointReader(checkpoint_path) \nvar_to_shape_map = reader.get_variable_to_shape_map()\n\n#\u5b58\u653e\u6587\u4ef6\nf = open('\/home\/ubuntu\/GAP8\/gap_sdk_old\/tf2gap8\/examples\/portmodel\/data\/params.txt','w')\n#\u6253\u5370tensor\u540d\u53ca\u5176\u503c\nfor key in var_to_shape_map:\n    print(\"tensor_name: \", key)\n    w = reader.get_tensor(key) \n#    print (w)\n#    print(type(key))  \n#     print(key.shape) \n\n    print('\\n' + key + '\\n',file=f)\n    f.write(str(w))\nf.close()\n<\/code><\/pre>\n\n<p>log<\/p>\n\n<pre><code>tensor_name:  ssd300\/multibox_head\/loc_4\/kernel\/Momentum\ntensor_name:  ssd300\/multibox_head\/loc_4\/kernel\ntensor_name:  ssd300\/multibox_head\/loc_4\/bias\/Momentum\ntensor_name:  ssd300\/multibox_head\/loc_4\/bias\ntensor_name:  ssd300\/multibox_head\/loc_3\/kernel\/Momentum\ntensor_name:  ssd300\/multibox_head\/loc_3\/kernel\ntensor_name:  ssd300\/multibox_head\/loc_5\/kernel\ntensor_name:  ssd300\/multibox_head\/loc_3\/bias\/Momentum\ntensor_name:  ssd300\/multibox_head\/loc_5\/kernel\/Momentum\ntensor_name:  ssd300\/multibox_head\/loc_3\/bias\ntensor_name:  ssd300\/multibox_head\/loc_2\/kernel\/Momentum\ntensor_name:  ssd300\/multibox_head\/loc_2\/bias\ntensor_name:  ssd300\/multibox_head\/loc_1\/kernel\ntensor_name:  ssd300\/multibox_head\/loc_1\/bias\ntensor_name:  ssd300\/multibox_head\/loc_0\/kernel\/Momentum\ntensor_name:  ssd300\/multibox_head\/loc_1\/bias\/Momentum\ntensor_name:  ssd300\/multibox_head\/loc_0\/bias\/Momentum\ntensor_name:  ssd300\/multibox_head\/loc_0\/bias\ntensor_name:  ssd300\/multibox_head\/cls_5\/kernel\ntensor_name:  ssd300\/multibox_head\/cls_4\/bias\/Momentum\ntensor_name:  ssd300\/multibox_head\/cls_4\/bias\ntensor_name:  ssd300\/multibox_head\/loc_1\/kernel\/Momentum\ntensor_name:  ssd300\/multibox_head\/cls_3\/kernel\/Momentum\ntensor_name:  ssd300\/multibox_head\/cls_3\/kernel\ntensor_name:  ssd300\/multibox_head\/cls_3\/bias\ntensor_name:  ssd300\/multibox_head\/cls_2\/bias\ntensor_name:  ssd300\/multibox_head\/cls_0\/kernel\/Momentum\ntensor_name:  ssd300\/multibox_head\/cls_2\/kernel\ntensor_name:  ssd300\/multibox_head\/cls_0\/kernel\ntensor_name:  ssd300\/multibox_head\/cls_0\/bias\/Momentum\ntensor_name:  ssd300\/multibox_head\/loc_2\/kernel\ntensor_name:  ssd300\/multibox_head\/cls_1\/kernel\ntensor_name:  ssd300\/conv1\/conv1_2\/bias\ntensor_name:  ssd300\/conv3\/conv3_3\/kernel\ntensor_name:  ssd300\/conv1\/conv1_2\/kernel\ntensor_name:  ssd300\/conv1\/conv1_1\/kernel\ntensor_name:  ssd300\/conv1\/conv1_2\/bias\/Momentum\ntensor_name:  ssd300\/conv1\/conv1_1\/bias\ntensor_name:  ssd300\/additional_layers\/conv10\/conv10_2\/kernel\ntensor_name:  ssd300\/multibox_head\/cls_3\/bias\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv10\/conv10_2\/kernel\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv11\/conv11_1\/kernel\ntensor_name:  ssd300\/additional_layers\/conv9\/conv9_2\/bias\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv9\/conv9_1\/bias\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv9\/conv9_2\/kernel\ntensor_name:  ssd300\/additional_layers\/conv8\/conv8_1\/bias\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv8\/conv8_2\/kernel\ntensor_name:  ssd300\/conv2\/conv2_1\/bias\ntensor_name:  ssd300\/additional_layers\/conv8\/conv8_2\/kernel\/Momentum\ntensor_name:  ssd300\/conv2\/conv2_1\/bias\/Momentum\ntensor_name:  ssd300\/conv2\/conv2_1\/kernel\ntensor_name:  ssd300\/conv2\/conv2_2\/bias\ntensor_name:  ssd300\/additional_layers\/conv9\/conv9_1\/kernel\ntensor_name:  ssd300\/additional_layers\/conv11\/conv11_2\/kernel\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv8\/conv8_2\/bias\/Momentum\ntensor_name:  global_step\ntensor_name:  ssd300\/multibox_head\/cls_2\/bias\/Momentum\ntensor_name:  ssd300\/conv2\/conv2_2\/bias\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv10\/conv10_2\/bias\/Momentum\ntensor_name:  ssd300\/conv4\/conv4_2\/kernel\/Momentum\ntensor_name:  ssd300\/conv3\/conv3_3\/bias\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv11\/conv11_1\/kernel\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv10\/conv10_1\/kernel\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv10\/conv10_1\/bias\ntensor_name:  ssd300\/conv5\/conv5_2\/kernel\ntensor_name:  ssd300\/additional_layers\/conv9\/conv9_2\/bias\ntensor_name:  ssd300\/conv1\/conv1_2\/kernel\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv11\/conv11_1\/bias\ntensor_name:  ssd300\/conv2\/conv2_2\/kernel\ntensor_name:  ssd300\/additional_layers\/conv11\/conv11_2\/bias\/Momentum\ntensor_name:  ssd300\/multibox_head\/loc_0\/kernel\ntensor_name:  ssd300\/conv2\/conv2_1\/kernel\/Momentum\ntensor_name:  ssd300\/multibox_head\/cls_4\/kernel\/Momentum\ntensor_name:  ssd300\/multibox_head\/cls_4\/kernel\ntensor_name:  ssd300\/fc7\/bias\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv10\/conv10_1\/bias\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv11\/conv11_2\/bias\ntensor_name:  ssd300\/additional_layers\/conv8\/conv8_1\/kernel\ntensor_name:  ssd300\/conv5\/conv5_3\/bias\ntensor_name:  ssd300\/multibox_head\/loc_2\/bias\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv8\/conv8_2\/bias\ntensor_name:  ssd300\/additional_layers\/conv9\/conv9_1\/kernel\/Momentum\ntensor_name:  ssd300\/conv3\/conv3_2\/bias\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv10\/conv10_1\/kernel\ntensor_name:  ssd300\/multibox_head\/loc_5\/bias\/Momentum\ntensor_name:  ssd300\/multibox_head\/cls_5\/kernel\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv11\/conv11_1\/bias\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv9\/conv9_2\/kernel\/Momentum\ntensor_name:  ssd300\/conv5\/conv5_2\/bias\/Momentum\ntensor_name:  ssd300\/conv4_3_scale\/weights\ntensor_name:  ssd300\/conv5\/conv5_3\/kernel\/Momentum\ntensor_name:  ssd300\/multibox_head\/loc_5\/bias\ntensor_name:  ssd300\/additional_layers\/conv8\/conv8_1\/bias\ntensor_name:  ssd300\/multibox_head\/cls_5\/bias\/Momentum\ntensor_name:  ssd300\/multibox_head\/cls_1\/bias\/Momentum\ntensor_name:  ssd300\/conv5\/conv5_1\/kernel\ntensor_name:  ssd300\/conv2\/conv2_2\/kernel\/Momentum\ntensor_name:  ssd300\/conv3\/conv3_1\/kernel\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv8\/conv8_1\/kernel\/Momentum\ntensor_name:  ssd300\/conv3\/conv3_1\/bias\ntensor_name:  ssd300\/conv3\/conv3_1\/bias\/Momentum\ntensor_name:  ssd300\/conv4\/conv4_1\/bias\ntensor_name:  ssd300\/multibox_head\/cls_1\/bias\ntensor_name:  ssd300\/conv3\/conv3_1\/kernel\ntensor_name:  ssd300\/conv3\/conv3_2\/kernel\ntensor_name:  ssd300\/fc7\/bias\ntensor_name:  ssd300\/conv3\/conv3_2\/bias\ntensor_name:  ssd300\/conv3\/conv3_2\/kernel\/Momentum\ntensor_name:  ssd300\/conv4\/conv4_1\/bias\/Momentum\ntensor_name:  ssd300\/conv3\/conv3_3\/bias\ntensor_name:  ssd300\/additional_layers\/conv10\/conv10_2\/bias\ntensor_name:  ssd300\/conv5\/conv5_2\/kernel\/Momentum\ntensor_name:  ssd300\/multibox_head\/cls_1\/kernel\/Momentum\ntensor_name:  ssd300\/conv3\/conv3_3\/kernel\/Momentum\ntensor_name:  ssd300\/conv5\/conv5_3\/bias\/Momentum\ntensor_name:  ssd300\/multibox_head\/cls_2\/kernel\/Momentum\ntensor_name:  ssd300\/conv1\/conv1_1\/bias\/Momentum\ntensor_name:  ssd300\/conv4\/conv4_1\/kernel\ntensor_name:  ssd300\/additional_layers\/conv11\/conv11_2\/kernel\ntensor_name:  ssd300\/conv4\/conv4_1\/kernel\/Momentum\ntensor_name:  ssd300\/conv4\/conv4_2\/bias\ntensor_name:  ssd300\/conv4\/conv4_2\/bias\/Momentum\ntensor_name:  ssd300\/conv4\/conv4_2\/kernel\ntensor_name:  ssd300\/conv4\/conv4_3\/bias\ntensor_name:  ssd300\/fc7\/kernel\/Momentum\ntensor_name:  ssd300\/conv4\/conv4_3\/bias\/Momentum\ntensor_name:  ssd300\/conv4\/conv4_3\/kernel\ntensor_name:  ssd300\/conv5\/conv5_1\/bias\ntensor_name:  ssd300\/fc6\/kernel\ntensor_name:  ssd300\/conv4\/conv4_3\/kernel\/Momentum\ntensor_name:  ssd300\/conv5\/conv5_1\/kernel\/Momentum\ntensor_name:  ssd300\/additional_layers\/conv9\/conv9_1\/bias\ntensor_name:  ssd300\/conv5\/conv5_2\/bias\ntensor_name:  ssd300\/conv1\/conv1_1\/kernel\/Momentum\ntensor_name:  ssd300\/conv5\/conv5_1\/bias\/Momentum\ntensor_name:  ssd300\/conv5\/conv5_3\/kernel\ntensor_name:  ssd300\/fc6\/bias\ntensor_name:  ssd300\/fc6\/bias\/Momentum\ntensor_name:  ssd300\/multibox_head\/cls_0\/bias\ntensor_name:  ssd300\/multibox_head\/cls_5\/bias\ntensor_name:  ssd300\/fc6\/kernel\/Momentum\ntensor_name:  ssd300\/conv4_3_scale\/weights\/Momentum\ntensor_name:  ssd300\/fc7\/kernel\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57800189,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-05T07:11:54.157","AcceptedAnswerId":57800918.0,"OwnerUserId":12023546.0,"Title":"problem to calculate gradient using GradientTape() of tensorflow 2.0","Body":"<p>Using tensorflow 2.0 and GradientTape() function, the first tape.gradient() gives correct gradient tensor, But the second tape.gradient() gives 'None'.\nWhy the second value is 'None'? I expect the gradient computed respectively on a second. <\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nx = tf.constant([ [1.0, 2.0], [3.0, 4.0], [5.0, 6.0] ])\ny0 = tf.constant([ [4.0], [8.0], [12.0] ])\n\nw = tf.Variable( [[1.0], [1.0]] ) \n\nwith tf.GradientTape() as tape:\n    y = tf.matmul(x, w)\n    print(\"y : \", y.numpy())\n    loss = tf.reduce_sum(y-y0)\n    print(\"loss : \", loss.numpy())\n\ngrad = tape.gradient(loss, w)    # gradient calculation is correct\nprint(\"gradient : \", grad.numpy())\n\nmu = 0.01\nw = w - mu*grad\n\nwith tf.GradientTape() as tape:\n    y = tf.matmul(x, w)\n    print(\"y : \", y.numpy())\n    loss = tf.reduce_sum(y-y0)\n    print(\"loss : \", loss.numpy())\n\ngrad = tape.gradient(loss, w)    # gradient value go to 'None'\nprint(\"gradient : \", grad)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57800918","CreationDate":"2019-09-05T07:58:15.157","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>You are overwriting <code>w<\/code> by a <code>Tensor<\/code> (which is <em>not<\/em> a <code>Variable<\/code>) by assigning <code>w = w - mu*grad<\/code>. By default, <code>GradientTape<\/code> only tracks variables. You have two options.<\/p>\n\n<ol>\n<li>Recommended: Replace <code>w = w - mu*grad<\/code> by <code>w.assign(w - mu*grad)<\/code>. This keeps <code>w<\/code> as a <code>Variable<\/code> and is the way to update variable values.<\/li>\n<li>You can track non-variables explicitly in a <code>GradientTape<\/code>. In the second tape context, add <code>tape.watch(w)<\/code> in the very beginning (before the <code>matmul<\/code>).<\/li>\n<\/ol>\n"}]}
+{"QuestionId":57800920,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-lite><google-coral>","CreationDate":"2019-09-05T07:58:21.657","AcceptedAnswerId":null,"OwnerUserId":1925189.0,"Title":"Coral model returns corrupted data until the result is freed by python","Body":"<p>We are using the coral stick with a custom network we created (based on vgg as the backbone).\nThe edge-tpu version we're using is 2.11<\/p>\n\n<p>The network returns 2 outputs: shape 1*12*12*65 and 1*12*12*256.\nWe are calling the network like this from python:<\/p>\n\n<pre><code>def _run_net(self, crop : np.ndarray):\n\n    orig_w = crop.shape[1]\n    orig_h = crop.shape[0]\n\n    if orig_w != WIDTH or orig_h != HEIGHT:\n        crop = cv2.resize(crop, (WIDTH, HEIGHT))\n    image = crop.reshape(1, HEIGHT, WIDTH, 1)\n\n\n    all_data = self.engine.RunInference(image.flatten())[1]\n    tensor_sizes = self.engine.get_all_output_tensors_sizes()\n    num_points = int(math.sqrt(tensor_sizes[0] \/ 65))\n    raw_points = all_data[:tensor_sizes[0]]\n    raw_points = raw_points.reshape((1, num_points, num_points, 65))\n    raw_descriptors = all_data[tensor_sizes[0]:]\n    raw_descriptors = raw_descriptors.reshape((1, num_points, num_points, 256))\n    #Imprtant Note! the copy of raw points and desc here is a must, because as long as memory is held the net always return\n    #the same results!!! this seems to be a bug in coral\n\n    return (raw_points, raw_descriptors, orig_w, orig_h) #does not work\n\n    #return (np.copy(raw_points), np.copy(raw_descriptors), orig_w, orig_h) #works ok!\n<\/code><\/pre>\n\n<p>And it seems until python calls gc on raw_points and raw_descriptors\neach call to the same net (or even a new instance of the engine)\nwill result in the same output.<\/p>\n","answers":[]}
+{"QuestionId":57801222,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tf.keras>","CreationDate":"2019-09-05T08:17:28.200","AcceptedAnswerId":57801547.0,"OwnerUserId":514149.0,"Title":"Does the Keras implementation in TF2 support everything native Keras can do with TF1?","Body":"<p>We are working with Keras on top of TensorFlow 1.x. But now that TF 2.0 is coming we are thinking of switching to that update, using the Keras API implementation built into TF 2.0.<\/p>\n\n<p>But before we do so, I would like to ask you guys: Do you know whether the Keras implementation in TF 2.0 does support everything native Keras does with TF 1.0, or are there any features missing?<\/p>\n\n<p>Moreover, will I be able to use my Keras code 1:1 with the new TF 2.0 implementation of the Keras API, or do we need to re-write parts of our existing Keras code?<\/p>\n","answers":[{"AnswerId":"57801547","CreationDate":"2019-09-05T08:37:12.367","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>If you want to use TensorFlow, then I highly recommend you to switch and use the TensorFlow implementation of Keras (i.e. <code>tf.keras<\/code>) because it will support more features of TF and also would be much more efficient and optimized than native Keras.<\/p>\n\n<p>Actually, Keras maintainers released a new version (2.2.5) of Keras a few days ago (after more than 10 months with no new release!) and they also recommend to use <code>tf.keras<\/code>. Here are the <a href=\"https:\/\/github.com\/keras-team\/keras\/releases\/tag\/2.2.5\" rel=\"nofollow noreferrer\">release notes<\/a>:<\/p>\n\n<blockquote>\n  <p>Keras 2.2.5 is the last release of Keras that implements the 2.2.* API. It is the last release to only support TensorFlow 1 (as well as Theano and CNTK).<\/p>\n  \n  <p>The next release will be 2.3.0, which makes significant API changes and add support for TensorFlow 2.0. The 2.3.0 release will be the last major release of multi-backend Keras. Multi-backend Keras is superseded by <code>tf.keras<\/code>.<\/p>\n  \n  <p>At this time, we recommend that Keras users who use multi-backend Keras with the TensorFlow backend switch to <code>tf.keras<\/code> in TensorFlow 2.0. <code>tf.keras<\/code> is better maintained and has better integration with TensorFlow features.<\/p>\n<\/blockquote>\n\n<p>This: \"Multi-backend Keras is superseded by <code>tf.keras<\/code>\" is a strong indicator that it is better to switch to <code>tf.keras<\/code>, especially if you are still at the beginning of your project.<\/p>\n"}]}
+{"QuestionId":57801343,"AnswerCount":1,"Tags":"<apache><tensorflow><deep-learning><apache-tika>","CreationDate":"2019-09-05T08:24:55.850","AcceptedAnswerId":57801585.0,"OwnerUserId":11840960.0,"Title":"What model does apache tika use internally - TensorflowRESTCaptioner","Body":"<p>Iam working on an image captioning tool and came across the apache tika<\/p>\n\n<p><a href=\"https:\/\/tika.apache.org\/1.18\/api\/org\/apache\/tika\/parser\/captioning\/tf\/TensorflowRESTCaptioner.html\" rel=\"nofollow noreferrer\">TensorflowRESTCaptioner<\/a><\/p>\n\n<p>and would like to now which model does it use internally and how good are the results when compared with the state of the art right now in the market<\/p>\n\n<p><a href=\"https:\/\/github.com\/facebookresearch\/pythia\" rel=\"nofollow noreferrer\">pythia - BUTF - FacebookResearch<\/a><\/p>\n","answers":[{"AnswerId":"57801585","CreationDate":"2019-09-05T08:39:17.277","ParentId":null,"OwnerUserId":"1849975","Title":null,"Body":"<p>Following Apache Tika <a href=\"https:\/\/tika.apache.org\/1.17\/index.html\" rel=\"nofollow noreferrer\">changelog<\/a> I came to <a href=\"https:\/\/issues.apache.org\/jira\/browse\/TIKA-2262\" rel=\"nofollow noreferrer\">this<\/a> feature request for image captioning. There, the author stated that they used Google <\/p>\n\n<blockquote>\n  <p>'show and tell' neural network<\/p>\n<\/blockquote>\n\n<p>mentioned in <a href=\"https:\/\/ai.googleblog.com\/2016\/09\/show-and-tell-image-captioning-open.html\" rel=\"nofollow noreferrer\">this<\/a> blog post.<\/p>\n\n<p>Also, <a href=\"https:\/\/arxiv.org\/abs\/1609.06647\" rel=\"nofollow noreferrer\">here<\/a> is a link to the paper if you want to compare it to the current state of the art methods.<\/p>\n"}]}
+{"QuestionId":57801680,"AnswerCount":0,"Tags":"<python><tensorflow><interpreter><tensorflow-lite>","CreationDate":"2019-09-05T08:45:35.303","AcceptedAnswerId":null,"OwnerUserId":12024139.0,"Title":"How can I use GPU for running a tflite model (*.tflite) using tf.lite.Interpreter (in python)?","Body":"<p>I have converted a tensorflow inference graph to tflite model file (*.tflite), according to instructions from <a href=\"https:\/\/www.tensorflow.org\/lite\/convert\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/lite\/convert<\/a>.<\/p>\n\n<p>I tested the tflite model on my GPU server, which has 4 Nvidia TITAN GPUs. I used the tf.lite.Interpreter to load and run tflite model file. <\/p>\n\n<p>It works as the former tensorflow graph, however, the problem is that the inference became too slow. When I checked out the reason, I found that the GPU utilization is simply 0% when tf.lite.Interpreter is running.<\/p>\n\n<p>Is there any method that I can run tf.lite.Interpreter with GPU support?<\/p>\n","answers":[]}
+{"QuestionId":57802160,"AnswerCount":0,"Tags":"<tensorflow><pytorch><lstm>","CreationDate":"2019-09-05T09:13:46.260","AcceptedAnswerId":null,"OwnerUserId":4397868.0,"Title":"Is torch.nn.LSTM equal to tf.nn.rnn_cell.BasicLSTMCell?","Body":"<p>One paper has released a tensorflow code for the classification task. It relies on tf.nn.rnn_cell.BasicLSTMCell with default setting.<\/p>\n\n<p>I try to use PyTorch 1.0 to re-implement, and I use torch.nn.LSTM with default setting. But the testing accuracy is lower than the tensorflow codes. <\/p>\n\n<p>I check the usage of these two lstm, and I think the default settings are the same. So I am confused... <\/p>\n","answers":[]}
+{"QuestionId":57802632,"AnswerCount":1,"Tags":"<python><neural-network><deep-learning><pytorch><loss-function>","CreationDate":"2019-09-05T09:40:42.707","AcceptedAnswerId":null,"OwnerUserId":10240085.0,"Title":"Loss is not converging in Pytorch but does in Tensorflow","Body":"<pre><code>Epoch: 1    Training Loss: 0.816370     Validation Loss: 0.696534\nValidation loss decreased (inf --&gt; 0.696534).  Saving model ...\nEpoch: 2    Training Loss: 0.507756     Validation Loss: 0.594713\nValidation loss decreased (0.696534 --&gt; 0.594713).  Saving model ...\nEpoch: 3    Training Loss: 0.216438     Validation Loss: 1.119294\nEpoch: 4    Training Loss: 0.191799     Validation Loss: 0.801231\nEpoch: 5    Training Loss: 0.111334     Validation Loss: 1.753786\nEpoch: 6    Training Loss: 0.064309     Validation Loss: 1.348847\nEpoch: 7    Training Loss: 0.058158     Validation Loss: 1.839139\nEpoch: 8    Training Loss: 0.015489     Validation Loss: 1.370469\nEpoch: 9    Training Loss: 0.082856     Validation Loss: 1.701200\nEpoch: 10   Training Loss: 0.003859     Validation Loss: 2.657933\nEpoch: 11   Training Loss: 0.018133     Validation Loss: 0.593986\nValidation loss decreased (0.594713 --&gt; 0.593986).  Saving model ...\nEpoch: 12   Training Loss: 0.160197     Validation Loss: 1.499911\nEpoch: 13   Training Loss: 0.012942     Validation Loss: 1.879732\nEpoch: 14   Training Loss: 0.002037     Validation Loss: 2.399405\nEpoch: 15   Training Loss: 0.035908     Validation Loss: 1.960887\nEpoch: 16   Training Loss: 0.051137     Validation Loss: 2.226335\nEpoch: 17   Training Loss: 0.003953     Validation Loss: 2.619108\nEpoch: 18   Training Loss: 0.000381     Validation Loss: 2.746541\nEpoch: 19   Training Loss: 0.094646     Validation Loss: 3.555713\nEpoch: 20   Training Loss: 0.022620     Validation Loss: 2.833098\nEpoch: 21   Training Loss: 0.004800     Validation Loss: 4.181845\nEpoch: 22   Training Loss: 0.014128     Validation Loss: 1.933705\nEpoch: 23   Training Loss: 0.026109     Validation Loss: 2.888344\nEpoch: 24   Training Loss: 0.000768     Validation Loss: 3.029443\nEpoch: 25   Training Loss: 0.000327     Validation Loss: 3.079959\nEpoch: 26   Training Loss: 0.000121     Validation Loss: 3.578420\nEpoch: 27   Training Loss: 0.148478     Validation Loss: 3.297387\nEpoch: 28   Training Loss: 0.030328     Validation Loss: 2.218535\nEpoch: 29   Training Loss: 0.001673     Validation Loss: 2.934132\nEpoch: 30   Training Loss: 0.000253     Validation Loss: 3.215722\n<\/code><\/pre>\n\n<p>My loss is not converging. I am working on Horses vs humans dataset. There is an <a href=\"https:\/\/colab.research.google.com\/github\/lmoroney\/dlaicourse\/blob\/master\/Course%201%20-%20Part%208%20-%20Lesson%202%20-%20Notebook.ipynb\" rel=\"nofollow noreferrer\">official notebook<\/a> in tensorflow for that and it worked like a charm. When I am trying to replicate the same with pytorch, loss is not converging. Can you please have a look?<\/p>\n\n<p>I am using <code>criterion = nn.BCEWithLogitsLoss()<\/code> and <code>optimizer = optim.RMSprop(model.parameters(), lr=0.001)<\/code>. Although it seems to have some effect on Training Loss, but Validation losses look like random numbers and not forming any pattern. What could be the possible reasons for loss not converging?<\/p>\n\n<p>This is my CNN architecture:<\/p>\n\n<pre><code>class Net(nn.Module):\n    def __init__(self):\n        super(Net, self).__init__()\n        # convolutional layer (sees 298x298x3 image tensor)\n        self.conv1 = nn.Conv2d(3, 16, 3)\n        # convolutional layer (sees 147x147x16 tensor)\n        self.conv2 = nn.Conv2d(16, 32, 3)\n        # convolutional layer (sees 71x71x32 tensor)\n        self.conv3 = nn.Conv2d(32, 64, 3)\n        # convolutional layer (sees 33x33x64 tensor)\n        self.conv4 = nn.Conv2d(64, 64, 3)\n        # convolutional layer (sees 14x14x64 tensor)\n        self.conv5 = nn.Conv2d(64, 64, 3)\n        # max pooling layer\n        self.pool = nn.MaxPool2d(2, 2)\n        # linear layer (64 * 7 * 7 -&gt; 500)\n        self.fc1 = nn.Linear(3136, 512)\n        # linear layer (512 -&gt; 1)\n        self.fc2 = nn.Linear(512, 1)\n        # dropout layer (p=0.25)\n        self.dropout = nn.Dropout(0.25)\n\n    def forward(self, x):\n        # add sequence of convolutional and max pooling layers\n        x = self.pool(F.relu(self.conv1(x)))\n        x = self.pool(F.relu(self.conv2(x)))\n        x = self.pool(F.relu(self.conv3(x)))\n        x = self.pool(F.relu(self.conv4(x)))\n        x = self.pool(F.relu(self.conv5(x)))\n\n        # flatten image input\n        x = x.view(-1, 64 * 7 * 7)\n        # add dropout layer\n        x = self.dropout(x)\n        # add 1st hidden layer, with relu activation function\n        x = F.relu(self.fc1(x))\n        # add dropout layer\n        x = self.dropout(x)\n        # add 2nd hidden layer\n        x = self.fc2(x)\n        return x\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/github.com\/mnauf\/horsesvshumans\/blob\/master\/Untitled.ipynb\" rel=\"nofollow noreferrer\">This is the complete jupyter notebook<\/a>. Apologies for not being able to create a minimal reproduce-able example code.<\/p>\n","answers":[{"AnswerId":"57937126","CreationDate":"2019-09-14T15:59:55.710","ParentId":null,"OwnerUserId":"10749432","Title":null,"Body":"<p>I think the problem is in <code>dataloaders<\/code>, here I noticed, that you're not passing <code>samplers<\/code> to <code>loaders<\/code> here:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># define samplers for obtaining training and validation batches\ntrain_sampler = SubsetRandomSampler(train_idx)\nvalid_sampler = SubsetRandomSampler(valid_idx)\n\ntrain_loader = torch.utils.data.DataLoader(\n        train_dataset,\n        batch_size=16,\n        num_workers=0,\n        shuffle=True\n    )\n\ntest_loader = torch.utils.data.DataLoader(\n        test_dataset,\n        batch_size=16,\n        num_workers=0,\n        shuffle=True\n    )\n<\/code><\/pre>\n\n<p>I have never used <code>Samplers<\/code>, so I don't now how to correctly use them, but I suppose you wanted to do smth like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># define samplers for obtaining training and validation batches\ntrain_sampler = SubsetRandomSampler(train_idx)\nvalid_sampler = SubsetRandomSampler(valid_idx)\n\ntrain_loader = torch.utils.data.DataLoader(\n        train_dataset,\n        sampler=train_sampler,\n        batch_size=16,\n        num_workers=0,\n        shuffle=True\n    )\n\ntest_loader = torch.utils.data.DataLoader(\n        test_dataset,\n        sampler=valid_sampler,\n        batch_size=16,\n        num_workers=0,\n        shuffle=True\n    )\n<\/code><\/pre>\n\n<p>And according to docs:<\/p>\n\n<blockquote>\n  <p>sampler (Sampler, optional) \u2013 defines the strategy to draw samples from the dataset. If specified, shuffle must be False.<\/p>\n<\/blockquote>\n\n<p>if you are using samplers you should turn off shuffle.<\/p>\n"}]}
+{"QuestionId":57803005,"AnswerCount":1,"Tags":"<parallel-processing><nlp><pytorch>","CreationDate":"2019-09-05T10:01:44.210","AcceptedAnswerId":null,"OwnerUserId":7896124.0,"Title":"Pytorch : how to run code on several machines in cluster","Body":"<p>I am using a cluster to train a recurrent neural network developed using PyTorch. PyTorch automatically threads, which allows to use all the cores of a machine in parallel without having to explicitly program for it. This is great !<\/p>\n\n<p>Now when I try to use several nodes at the same time using a script like this one :<\/p>\n\n<pre><code>#$ -S \/bin\/bash\n#$ -N comparison_of_architecture\n#$ -pe mvapich2-rostam 32\n#4 -tc 4\n#$ -o \/scratch04.local\/cnelias\/Deep-Jazz\/logs\/out_comparison_training.txt\n#$ -e \/scratch04.local\/cnelias\/Deep-Jazz\/logs\/err_comparison_training.txt\n#$ -t 1\n#$ -cwd\n<\/code><\/pre>\n\n<p>I see that 4 nodes are being used but only one is actually doing work, so \"only\" 32 cores are in use.<\/p>\n\n<p>I have no knowledge of parallel programming and I don't understand a thing in the tutorial provided on PyTorch's website, I am afraid this is completely out of my scope.<\/p>\n\n<p>Are you aware of a simple way to let a PyTorch program run on several machines without having to explicitly program the exchanges of the messages and computation between these machines ?<\/p>\n\n<p>PS : I unfortunately don't have a GPU and the cluster I am using also doesn't, otherwise I would have tried it.<\/p>\n","answers":[{"AnswerId":"57805591","CreationDate":"2019-09-05T12:32:35.473","ParentId":null,"OwnerUserId":"5652313","Title":null,"Body":"<p><strong>tl;dr<\/strong> There is no easy solution.<\/p>\n\n<p>There are two ways how you can parallelize training of a deep learning model. The most commonly used is data parallelism (as opposed to model parallelism). In that case, you have a copy of the model on each device, run the model and back-propagation on each device independently and get the weight gradients. Now, the tricky part begins. You need to collect all gradients at a single place, sum them (differentiation is linear w.r.t. summation) and do the optimizer step. The optimizer computes the weight updates and you need to tell each copy of your model how to update the weights.<\/p>\n\n<p>PyTorch can somehow do this for multiple GPUs on a single machine, but as far as I know, there is no ready-made solution to do this on multiple machines.<\/p>\n"}]}
+{"QuestionId":57803130,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-05T10:08:58.587","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"Eager mode disable in Tensorflow.Keras automatically","Body":"<p>I am using <a href=\"https:\/\/keras.io\/examples\/mnist_cnn\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/examples\/mnist_cnn\/<\/a> code for Mnist Dataset.\nI need change matrix multiplication function. <br>\n1. instead of keras libraries , I imported tensorflow.keras libraries\n<br> <code>from tensorflow.keras.datasets import mnist<\/code> <br>\n2. enabled eager execution\n<br><code>tf.enable_eager_execution()<\/code><br>\n3. Tried to find matrix mutiplcation function, I get to gen_math_ops function and override it (Monkey patch)\n<br><code>from tensorflow.python.ops import gen_math_ops as  ops;ops.mat_mul = Smat<\/code><br>\nI that function, eager execution is disabled and I am not able to see all dimensions. <br>\n-Any Idea, how can I solve my problem please?<br>\n- If it is not possible to solve, where can I find user made keras layer for MNist database (I think I can solve my problem with it as well)<\/p>\n\n<pre><code>from tensorflow.python.ops import gen_math_ops as  ops\nA_mat =  ops.mat_mul\n\ndef Sparemat(a, b, transpose_a=False, transpose_b=False, name=None):\n    tf.enable_eager_execution()\n    print('a = ',a.shape)\n    print('b =',b.shape)\n    print(tf.executing_eagerly())\n    return A_mat(a,b,transpose_a,transpose_b,name)\n<\/code><\/pre>\n\n<p>My output is <\/p>\n\n<pre><code>a =  (?, 9216)\nb = (9216, 128)\nFalse\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57803236,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><deep-learning><deeplab>","CreationDate":"2019-09-05T10:14:58.363","AcceptedAnswerId":null,"OwnerUserId":12024544.0,"Title":"Training Script for DeepLabV3 for ADE20K type datasets","Body":"<p>I am attempting to fine-tune DeeplabV3 on a subset of ADE20K for indoor segmentation only. \nNow the demo training script for the MobileNet model is modified for the PASCAL VOC dataset, which has different training parameters than that for ADE20k. \nSpecifically, I am looking for the training script (or the training parameters) that I can use to train my customized ADE20K dataset. What values of <code>--train_crop_size<\/code>, <code>--fine_tune_batch_norm<\/code>, <code>--decoder_output_stride<\/code> and other relevant parameters I should use?\nAny help would be appreciated.<\/p>\n\n<p>I have successfully re-labeled the dataset into 6 classes (originally being 150), with all the unwanted objects as label 0, while the others are labelled in order of (1,2,3,4,5,6) as mentioned in the documentation.\nThe original dataset of ADE20K also has 0 as the ignore_label and 150 other classes, there by makin total number of classes to be 151.<\/p>\n\n<p>I also tried modifying parameters to suit my dataset but always get distorted masks as outputs. The training happens but the results are unacceptable.<\/p>\n\n<p>My training, eval and vis codes are as follows:<\/p>\n\n<pre><code>CKPT_NAME=\"deeplabv3_mnv2_ade20k_train\"\nCKPT_NAME11=\"deeplabv3_mnv2_ade20k_train_2018_12_03\"\nTF_INIT_CKPT=\"${CKPT_NAME}_2018_12_03.tar.gz\"\n\npython3 train.py \\\n  --logtostderr \\\n  --train_split=\"train\" \\\n  --model_variant=\"mobilenet_v2\" \\\n  --output_stride=16 \\\n  --decoder_output_stride=4 \\\n  --train_crop_size=\"513,513\" \\\n  --train_batch_size=16 \\\n  --training_number_of_steps=\"${NUM_ITERATIONS}\" \\\n  --fine_tune_batch_norm=true \\\n  --dataset='ade20k' \\\n  --tf_initial_checkpoint=\"${INIT_FOLDER}\/${CKPT_NAME11}\/model.ckpt\" \\\n  --train_logdir=\"${TRAIN_LOGDIR}\" \\\n  --dataset_dir=\"${PASCAL_DATASET}\"\n\n\n\npython3 eval.py \\\n  --logtostderr \\\n  --eval_split=\"val\" \\\n  --model_variant=\"mobilenet_v2\" \\\n  --dataset='ade20k' \\\n  --eval_crop_size=\"513,513\" \\\n  --checkpoint_dir=\"${TRAIN_LOGDIR}\" \\\n  --eval_logdir=\"${EVAL_LOGDIR}\" \\\n  --dataset_dir=\"${PASCAL_DATASET}\" \\\n  --max_number_of_evaluations=1\n\n\n\npython3 vis.py \\\n  --logtostderr \\\n  --vis_split=\"val\" \\\n  --model_variant=\"mobilenet_v2\" \\\n  --vis_crop_size=\"513,513\" \\\n  --decoder_output_stride=4 \\\n  --checkpoint_dir=\"${TRAIN_LOGDIR}\" \\\n  --dataset='ade20k' \\\n  --vis_logdir=\"${VIS_LOGDIR}\" \\\n  --dataset_dir=\"${PASCAL_DATASET}\" \\\n  --max_number_of_iterations=1\n<\/code><\/pre>\n\n<p>My training environment is Ubuntu 18.04\/Python3.6\/TF-1.13.0\/RTX-2080Ti.\nI have followed almost all the suggested modifications on Deeplab issues on GitHub, but nothing seems to work for me.<\/p>\n","answers":[]}
+{"QuestionId":57803287,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><pytorch><feed-forward>","CreationDate":"2019-09-05T10:18:08.430","AcceptedAnswerId":57804714.0,"OwnerUserId":3292636.0,"Title":"How to create custom neural network with custom weight initialization in tensorflow or pytorch","Body":"<p>I'm trying to create a small neural network with custom connections between neurons. The connections should exist over several layers and not be fully connected (sparse) as shown in the picture. I would also like to do the weight initialization manually and not completely randomly. My goal is to determine whether a connection is positive or negative. Is it possible to create such a neural net in tensorflow (python\/js) or pytorch?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/XSPPG.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/XSPPG.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"57804714","CreationDate":"2019-09-05T11:41:35.900","ParentId":null,"OwnerUserId":"3607203","Title":null,"Body":"<p>To summarize:<br\/>\nCan you do it? -- <strong>Yes<\/strong>, absolutely. <br\/>\nIs it going to be pretty? -- <strong>No<\/strong>, absolutely not.<\/p>\n\n<p>In my explanation, I will further focus on PyTorch, as this is the library that I am more comfortable with, and that is especially more useful if you have custom operations that you can easily express in a pythonic manner. Tensorflow also has eager execution mode (more serious integration from version 2, if I remember that correctly), but it is traditionally done with computational graphs, which make this whole thing a little uglier than it needs to be.<\/p>\n\n<p>As you hopefully know, backpropagation (the \"learning\" step in any ANN) is basically an inverse pass through the network, to calculate gradients, or at least close enough to the truth for our problem at hand. Importantly, <code>torch<\/code> functions store this \"reverse\" direction, which makes it trivial for the user to call backpropagation functions. <\/p>\n\n<p>To model a simple network as described in your image, we have only one major disadvantage:<br\/>\nThe available operations are usually excelling at what they are doing because they are <em>simply<\/em> and can be optimized quite heavily. In your case, you have to express different layers as custom operations, which generally scales incredibly poorly, unless you can express the functionals as some form of matrix operation, which I do not see straigt away in your example. I am further assuming that you are applying some form of non-linearity, as it would otherwise be a network that <a href=\"https:\/\/towardsdatascience.com\/perceptrons-logical-functions-and-the-xor-problem-37ca5025790a\" rel=\"nofollow noreferrer\">would fail for any non-linearly separable problem<\/a>.<\/p>\n\n<pre><code>import torch\nimport torch.nn as nn\nclass CustomNetwork(nn.module):\n    def __init__(self):\n        self.h_1_1 = nn.Sequential(nn.Linear(1,2), nn.ReLU) # top node in first layer\n        self.h_1_2 = nn.Sequential(nn.Linear(1,2), nn.ReLU) # bottom node in first layer\n        # Note that these nodes have no shared weights, which is why we\n        # have to initialize separately.\n        self.h_2_1 = nn.Sequential(nn.Linear(1,1), nn.ReLU) # top node in second layer\n        self.h_2_2 = nn.Sequential(nn.Linear(1,1), nn.ReLU) # bottom node in second layer\n\n        self.h_2_1 = nn.Sequential(nn.Linear(2,1), nn.ReLU) # top node in third layer\n        self.h_2_2 = nn.Sequential(nn.Linear(2,1), nn.ReLU) # bottom node in third layer\n       # out doesn't require activation function due to pairing with loss function\n        self.out = nn.Linear(2,1)\n\n    def forward(self, x):\n        # x.shape: (batch_size, 2)\n\n        # first layer. shape of (batch_size, 2), respectively\n        out_top = self.h_1_1(x[:,0])\n        out_bottom = self.h_1_2(x[:,1])\n\n        # second layer. shape of (batch_size, 1), respectively\n        out_top_2 = self.h_2_1(out_top[:,0])\n        out_bottom_2 = self.h_2_2(out_bottom[:,0])\n\n        # third layer. shape of (batch_size, 1), respectively\n        # additional concatenation of previous outputs required.\n        out_top_3 = self.h_3_1(torch.cat([out_top_2, -1 * out_top[:,1]], dim=1))\n        out_bottom_3 = self.h_3_2(torch.cat([out_bottom_2, -1 * out_bottom[:,1]], dim=1))\n        return self.out(torch.cat([out_top_3, out_bottom_3], dim=1))\n<\/code><\/pre>\n\n<p>As you can see, any computational step is (in this case rather explicitly) given, and very much possible. Again, once you want to scale your number of neurons for each layer, you are going to have to be a little more creative in how you process, but for-loops do very much work in PyTorch as well. Note that this will in any case be much slower than a vanilla linear layer, though.\nIf you can live with seperately trained weights, you can always also just define separate linear layers of smaller size and put them in a more convenient fashion.  <\/p>\n"}]}
+{"QuestionId":57803562,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2019-09-05T10:35:14.213","AcceptedAnswerId":null,"OwnerUserId":7896153.0,"Title":"Why am I getting this error: \"ModuleNotFoundError: No module named 'nst_utils'\"","Body":"<p>I am getting this error after running this line of codes to import libraries for implementing a neural style algorithm:<\/p>\n\n<pre><code>import os\nimport sys\nimport scipy.io\nimport scipy.misc\nimport matplotlib.pyplot as plt\nfrom matplotlib.pyplot import imshow\nfrom PIL import Image\nfrom nst_utils import *\nimport numpy as np\nimport tensorflow as tf\n\n%matplotlib inline\n<\/code><\/pre>\n","answers":[{"AnswerId":"57804133","CreationDate":"2019-09-05T11:07:57.953","ParentId":null,"OwnerUserId":"2745495","Title":null,"Body":"<p>Do you have a copy of <em>nst_utils.py<\/em>?<\/p>\n\n<p>It is part of the work done when going through <strong>deeplearning.ai<\/strong>'s Deep Learning specialization online course, specifically <a href=\"https:\/\/www.coursera.org\/learn\/convolutional-neural-networks\" rel=\"nofollow noreferrer\">Week 4 of Convolutional Neural Networks<\/a>. You need to have a copy of <em>nst_utils.py<\/em> in the same directory as that code you're running (or that it's added to your sys.path).<\/p>\n\n<p>If you do a search for <em>nst_utils.py<\/em>, you'll get some links to github repositories of people who went through the course, and posted their work online, like this: <a href=\"https:\/\/github.com\/JudasDie\/deeplearning.ai\" rel=\"nofollow noreferrer\">https:\/\/github.com\/JudasDie\/deeplearning.ai<\/a>. You could probably copy their <em>nst_utils.py<\/em> code (check their README for usage and license notes), though it would have been better if you went through the course yourself.<\/p>\n"}]}
+{"QuestionId":57803704,"AnswerCount":1,"Tags":"<python><arrays><numpy><tensorflow>","CreationDate":"2019-09-05T10:43:13.467","AcceptedAnswerId":null,"OwnerUserId":7266892.0,"Title":"How to convert array([array([ ....],type=float32)],[array([ ....],type=float32)]) to array([...],[...])?","Body":"<p>Here is my code:<\/p>\n\n<pre><code>def extract_doc2vec(doc, w2v):\nvecs = [w2v[word] for word in doc if word in w2v]\nif len(vecs) &gt; 0:\n    vecs = np.asarray(vecs).sum(axis=0)\n    return vecs\ny = newsgroups_all['target']\n\nx= []\n\nfor doc in newsgroups_all['data']:\n    edv = extract_doc2vec(list(gensim.utils.tokenize(doc)),model.wv)\n    x.append(edv)\n<\/code><\/pre>\n\n<p>I have to convert retrain_doc which can be fit into <\/p>\n\n<pre><code>for train_index, test_index in sssd.split(x, y):\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","answers":[{"AnswerId":"58046324","CreationDate":"2019-09-22T05:31:58.073","ParentId":null,"OwnerUserId":"5013853","Title":null,"Body":"<p>Convert the list x into data frame then convert it into numpy array and apply shuffle split<\/p>\n\n<pre><code>data = pd.DataFrame(x).to_numpy()\n<\/code><\/pre>\n"}]}
+{"QuestionId":57803706,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><conv-neural-network><pytorch>","CreationDate":"2019-09-05T10:43:21.440","AcceptedAnswerId":57827926.0,"OwnerUserId":9143069.0,"Title":"Why is my texture synthesis algorithm only producing blocky \/ noisy, non-sensible output?","Body":"<p>My end-goal is to create a script for neural style transfer, however, during writing code for said task, I stumbled upon a certain problem: the texture synthesis part of the algorithm seemed to have some problems with reproducing the artistic style. In order to solve this, I decided to create another script where I'd try to solve the task of texture synthesis using a neural network on its own.<\/p>\n\n<p>TL;DR ... even after tackling the problem on its own, my script still produced blocky \/ noisy, non-sensible output.<\/p>\n\n<p>I've tried having a look at how other people have solved this task, but most of what I found were more sophisticated solutions (\"fast neural-style-transfer\", etc.). Also, I couldn't find too many PyTorch implementations. <\/p>\n\n<p>Since I've already spent the past couple of days on trying to fix this issue and considering that I'm new to the PyTorch-Framework, I have decided to ask the StackOverflow community for help and advice. <\/p>\n\n<p>I use the VGG16 network for my model ... <\/p>\n\n<pre><code>class VGG16(nn.Module):\n    def __init__(self):\n        super(VGG16, self).__init__()\n        vgg_fs = models.vgg16(pretrained=True).features\n\n        self.sl1 = nn.Sequential()\n        self.sl2 = nn.Sequential()\n        self.sl3 = nn.Sequential()\n        self.sl4 = nn.Sequential()\n        self.sl5 = nn.Sequential()\n\n        for i in range(4):\n            self.sl1.add_module(str(i), vgg_fs[i])\n        for i in range(4, 9):\n            self.sl2.add_module(str(i), vgg_fs[i])\n        for i in range(9, 16):\n            self.sl3.add_module(str(i), vgg_fs[i])\n        for i in range(16, 23):\n            self.sl4.add_module(str(i), vgg_fs[i])\n        for i in range(23, 30):\n            self.sl5.add_module(str(i), vgg_fs[i])\n\n        for p in self.parameters():\n            p.requires_grad_(False)\n\n    def forward(self, x):\n        h = self.sl1(x)\n        h1 = h\n        h = self.sl2(h)\n        h2 = h\n        h = self.sl3(h)\n        h3 = h\n        h = self.sl4(h)\n        h4 = h\n        h = self.sl5(h)\n        h5 = h\n\n        return_tuple = namedtuple('hidden_states', ['h1', 'h2', 'h3', 'h4', 'h5'])\n        ret = return_tuple(h1, h2, h3, h4, h5)\n\n        return ret\n<\/code><\/pre>\n\n<p>Then, I have some functions for gram-matrix computation and normalization ...<\/p>\n\n<pre><code>def comp_gram(f):\n    (b, c, h, w) = f.shape\n    f = f.view(b, c, h * w)\n    g = f.bmm(f.transpose(1, 2)) \n    g = g \/ (c * h * w)\n    return g\n\ndef norm(b):\n    mean    = torch.Tensor([0.485, 0.456, 0.406]).cuda().view(3, 1, 1)\n    std     = torch.Tensor([0.229, 0.224, 0.225]).cuda().view(3, 1, 1)\n    return (b - mean) \/ std\n<\/code><\/pre>\n\n<p>And finally, my training-function ... <\/p>\n\n<pre><code>def train(model, style_img, w=224, h=224, iters=32, lr=1):\n    style = torch.from_numpy(np.array(Image.open(style_img))).cuda().float()\n    style = style \/ 255.\n    style = style.view(1, style.size()[2], *style.size()[:2])\n\n    style_fts = model(norm(style))\n    style_gms = [comp_gram(f) for f in style_fts]\n\n    img = torch.rand(*style.size()[:2], h, w, requires_grad=True, device='cuda')\n    optimizer = optim.Adam([img], lr=lr)\n    mse_loss = nn.MSELoss()\n\n    plt.ion()\n    for i in range(iters):\n        optimizer.zero_grad()\n        actvs = model(norm(img))\n\n        lss = 0.\n        for f, gm in zip(actvs, style_gms):\n            g = comp_gram(f)\n            lss += mse_loss(g, gm)\n\n        lss.backward()\n        optimizer.step()\n\n        if (i % 5 == 0) or (i == iters - 1):\n            plt.title('Iter#{:04d}'.format(i))\n            plt.imshow(img.detach().cpu().view(*img.shape[2:], img.shape[1]))\n            plt.pause(1e-3)\n\n        print('[iter#{:04d}]: Loss\\t-&gt; {:}'.format(i, lss.item()))\n\n    plt.ioff()\n\n    plt.subplot(121)\n    plt.title('Original')\n    plt.imshow(style.cpu().view(*style.size()[2:], style.size()[1]))\n\n    plt.subplot(122)\n    plt.title('Style')\n    plt.imshow(img.detach().cpu().view(*img.size()[2:], img.size()[1]))\n\n    plt.show()\n<\/code><\/pre>\n\n<p>I expected the algorithm to produce the style of the painting in some way, shape or form, but that didn't happen. \u200d\u2642\ufe0f<\/p>\n\n<p>(Since I sadly don't have enough reputation to post an image, you can find a picture of my non-sensible output here: <a href=\"https:\/\/mattmoony.github.io\/cdn\/conv-net_pytorch-style-transfer\/problem.png\" rel=\"nofollow noreferrer\">https:\/\/mattmoony.github.io\/cdn\/conv-net_pytorch-style-transfer\/problem.png<\/a>)<\/p>\n\n<p>I sincerely hope that you will be able to help me out and that I'm able to use this issue as a learning experience. Thank you! <\/p>\n","answers":[{"AnswerId":"57827926","CreationDate":"2019-09-06T20:00:53.450","ParentId":null,"OwnerUserId":"9143069","Title":null,"Body":"<p>Hurrah! <\/p>\n\n<p>After yet another day of researching and testing, I've finally discovered the bug in my code. <\/p>\n\n<p>The problem doesn't lie with the training process or the model itself, but rather with the lines responsible for loading the style image. (<a href=\"https:\/\/nextjournal.com\/gkoehler\/pytorch-neural-style-transfer\" rel=\"nofollow noreferrer\">this<\/a> article helped me discover the issue)<\/p>\n\n<p>So... I added the following two functions to my script ... <\/p>\n\n<pre><code>def load_img(p):\n    img = Image.open(p).convert('RGB')\n    im_transform = transforms.Compose([\n        transforms.ToTensor()\n    ])\n    img = im_transform(img).unsqueeze(0).cuda()\n\n    return img\n\ndef to_img(t):\n    img = t.cpu().clone().detach()\n    img = img.numpy().squeeze()\n    img = img.transpose(1, 2, 0)\n    img = img.clip(0, 1)\n\n    return img\n<\/code><\/pre>\n\n<p>And used them in my training function ... <\/p>\n\n<pre><code>def train(...):\n...\n    style = load_img(style_img)\n...\n            if i % 5 == 0:\n                plt.title('Iter#{:04d}'.format(i))\n                plt.imshow(to_img(img))\n                plt.pause(1e-3)\n...\n&lt;etc.&gt;\n...\n<\/code><\/pre>\n\n<p>I'm glad that I was able to figure out what the problematic part of my script was and I'll do my best to avoid similar bugs in the future. <\/p>\n\n<p>I hope that this helps anyone experiencing similar problems or that it at least inspires them to keep on chasing for a solution. <\/p>\n"}]}
+{"QuestionId":57803735,"AnswerCount":1,"Tags":"<python><tensorflow><keras><neural-network><generator>","CreationDate":"2019-09-05T10:44:59.020","AcceptedAnswerId":57804432.0,"OwnerUserId":1009508.0,"Title":"Keras image generator keep giving different number of lables","Body":"<p>So I am trying to make a simple fine turned Resnet50 model using the Market1501 dataset and keras. <\/p>\n\n<p>So the data set contains images (12000 or so) and 751 lables that I want to use (0-750)\nI can fit the data into a single go so I have to use a image generator for this. <\/p>\n\n<p>So my base model is like this<\/p>\n\n<pre><code>base_model = ResNet50(weights='imagenet', include_top=False,input_tensor=Input(shape=(224,224,3)))\nx = base_model.output\nx = Flatten(name=\"flatten\")(x)\nx = Dropout(0.5)(x)\nx = Dense(750, activation='softmax', name='fc8',kernel_initializer=RandomNormal(mean=0.0, stddev=0.001))(x)\nmodel = Model(input=base_model.input, output=x)\n<\/code><\/pre>\n\n<p>And my image generator is like this<\/p>\n\n<pre><code>def image_generator(image_array, batch_size):\n    # Define data generator arguments\n    datagen_args = dict(rotation_range=20,\n                    width_shift_range=0.2,\n                    height_shift_range=0.2,\n                    shear_range=0.1,\n                    zoom_range=0.1,\n                    horizontal_flip=True)\n\n    # Create different data generators for each image\n    # This gives each image a unique transformation wich will make it harder for the network\n    datagen = ImageDataGenerator(**datagen_args)\n    while True:\n        number_of_images = len(image_array)\n        indices = np.random.permutation(np.arange(number_of_images))\n        num_batches = number_of_images \/\/ batch_size\n        for bid in range(num_batches):\n            # loop once per batch\n            images = []\n            lables = []\n            batch_indices = indices[bid * batch_size: (bid + 1) * batch_size]\n            for i in batch_indices:\n                img, lbl = image_array[i]\n                # Process images\n                img = image.load_img(os.path.join(TRAIN, img), target_size=[224, 224])\n                img = image.img_to_array(img)\n                #img = np.expand_dims(img, axis=0)\n                img = preprocess_input(img)\n                img = datagen.random_transform(img)\n                images.append(img)\n                lables.append(lbl)\n            yield np.array(images), to_categorical(lables)\n<\/code><\/pre>\n\n<p>And I use it like this<\/p>\n\n<pre><code>batch_size = 64\nNUM_EPOCHS = 40\ntrain_gen = image_generator(image_array, batch_size)\nnum_train_steps = len(image_array)\n<\/code><\/pre>\n\n<p>The issue is it give me this error <\/p>\n\n<pre><code>Error when checking target: expected fc8 to have shape (751,) but got array with shape (742,)\n<\/code><\/pre>\n\n<p>And the bigger issue is the 2nd number keep changing so I know its something with the image generator not getting every label in for each iteration.  <\/p>\n\n<p><strong>EDIT<\/strong>\nHow the data is generated\nThere is a external list with the image and the label like this <\/p>\n\n<pre><code>['0002_451_03.jpg', '0']\nimg001.jpg, 0\nimg002.jpg, 0\n...\nimg1500.jpg, 750\n<\/code><\/pre>\n\n<p>This is read in and loaded into a array. The label is the number after the image<\/p>\n","answers":[{"AnswerId":"57804432","CreationDate":"2019-09-05T11:24:58.907","ParentId":null,"OwnerUserId":"8563649","Title":null,"Body":"<p>change <\/p>\n\n<pre><code>batch_indices = indices[bid * batch_size: (bid + 1) * batch_size]\n<\/code><\/pre>\n\n<p>with<\/p>\n\n<pre><code>batch_indices = indices[bid * batch_size: min((bid + 1) * batch_size, number_of_images)]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57804181,"AnswerCount":1,"Tags":"<tensorflow><nvidia><nvidia-docker>","CreationDate":"2019-09-05T11:10:29.897","AcceptedAnswerId":null,"OwnerUserId":3398900.0,"Title":"Tensorflow takes long time to connect to Nvidia-Drivers","Body":"<p>I have installed my Nvidia-drivers on my remote-server and <\/p>\n\n<p><code>$nvidia-smi<\/code> also returns the list of devices<\/p>\n\n<pre><code>+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 410.48                 Driver Version: 410.48                    |\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 208...  Off  | 00000000:01:00.0 Off |                  N\/A |\n| 36%   64C    P2    78W \/ 300W |   9813MiB \/ 10989MiB |      0%      Default |\n+-------------------------------+----------------------+----------------------+\n|   1  GeForce RTX 208...  Off  | 00000000:02:00.0 Off |                  N\/A |\n|  0%   45C    P8     8W \/ 300W |    774MiB \/ 10989MiB |      0%      Default |\n+-------------------------------+----------------------+----------------------+\n\n+-----------------------------------------------------------------------------+\n| Processes:                                                       GPU Memory |\n|  GPU       PID   Type   Process name                             Usage      |\n|=============================================================================|\n|    0       333      C   python                                      9039MiB |\n|    0     11773      C   python                                       191MiB |\n|    0     12111      C   python                                       191MiB |\n|    0     12591      C   python                                       191MiB |\n|    0     12999      C   python                                       191MiB |\n|    1     11773      C   python                                       191MiB |\n|    1     12111      C   python                                       191MiB |\n|    1     12591      C   python                                       191MiB |\n|    1     12999      C   python                                       191MiB |\n+-----------------------------------------------------------------------------+\n<\/code><\/pre>\n\n<p>But when I try to use.<\/p>\n\n<pre><code>&gt;&gt;&gt; from tensorflow.python.client import device_lib\n&gt;&gt;&gt; device_lib.list_local_devices()\n<\/code><\/pre>\n\n<p>It is taking too long time to give the list of devices<\/p>\n\n<pre><code>Python 3.7.3 (default, Mar 27 2019, 22:11:17)\n[GCC 7.3.0] :: Anaconda, Inc. on linux\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n&gt;&gt;&gt; from tensorflow.python.client import device_lib\n&gt;&gt;&gt; device_lib.list_local_devices()\n2019-09-05 15:56:27.568310: I tensorflow\/core\/platform\/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\n2019-09-05 15:56:27.596133: I tensorflow\/core\/platform\/profile_utils\/cpu_utils.cc:94] CPU Frequency: 3600000000 Hz\n2019-09-05 15:56:27.597408: I tensorflow\/compiler\/xla\/service\/service.cc:168] XLA service 0x55da9a2fe370 executing computations on platform Host. Devices:\n2019-09-05 15:56:27.597421: I tensorflow\/compiler\/xla\/service\/service.cc:175]   StreamExecutor device (0): &lt;undefined&gt;, &lt;undefined&gt;\n2019-09-05 15:56:27.598606: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcuda.so.1\n2019-09-05 15:59:27.755714: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-09-05 15:59:27.757588: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-09-05 15:59:27.758287: I tensorflow\/compiler\/xla\/service\/service.cc:168] XLA service 0x55da9c6d9140 executing computations on platform CUDA. Devices:\n2019-09-05 15:59:27.758299: I tensorflow\/compiler\/xla\/service\/service.cc:175]   StreamExecutor device (0): GeForce RTX 2080 Ti, Compute Capability 7.5\n2019-09-05 15:59:27.758303: I tensorflow\/compiler\/xla\/service\/service.cc:175]   StreamExecutor device (1): GeForce RTX 2080 Ti, Compute Capability 7.5\n2019-09-05 15:59:27.758508: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-09-05 15:59:27.758883: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1640] Found device 0 with properties:\nname: GeForce RTX 2080 Ti major: 7 minor: 5 memoryClockRate(GHz): 1.755\npciBusID: 0000:01:00.0\n2019-09-05 15:59:27.758919: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-09-05 15:59:27.759313: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1640] Found device 1 with properties:\nname: GeForce RTX 2080 Ti major: 7 minor: 5 memoryClockRate(GHz): 1.755\npciBusID: 0000:02:00.0\n2019-09-05 15:59:27.759441: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudart.so.10.0\n2019-09-05 15:59:27.760153: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcublas.so.10.0\n2019-09-05 15:59:27.760803: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcufft.so.10.0\n2019-09-05 15:59:27.760941: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcurand.so.10.0\n2019-09-05 15:59:27.761755: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcusolver.so.10.0\n2019-09-05 15:59:27.762395: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcusparse.so.10.0\n2019-09-05 15:59:27.764370: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudnn.so.7\n2019-09-05 15:59:27.764435: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-09-05 15:59:27.764846: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-09-05 15:59:27.765262: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-09-05 15:59:27.765638: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-09-05 15:59:27.766026: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1763] Adding visible gpu devices: 0, 1\n2019-09-05 15:59:27.766046: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudart.so.10.0\n2019-09-05 15:59:27.767185: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1181] Device interconnect StreamExecutor with strength 1 edge matrix:\n2019-09-05 15:59:27.767194: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1187]      0 1\n2019-09-05 15:59:27.767197: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1200] 0:   N Y\n2019-09-05 15:59:27.767200: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1200] 1:   Y N\n2019-09-05 15:59:27.767370: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-09-05 15:59:27.767755: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-09-05 15:59:27.768175: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-09-05 15:59:27.768546: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1326] Created TensorFlow device (\/device:GPU:0 with 950 MB memory) -&gt; physical GPU (device: 0, name: GeForce RTX 2080 Ti, pci bus id: 0000:01:00.0, compute capability: 7.5)\n2019-09-05 15:59:27.768811: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-09-05 15:59:27.769219: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1326] Created TensorFlow device (\/device:GPU:1 with 9703 MB memory) -&gt; physical GPU (device: 1, name: GeForce RTX 2080 Ti, pci bus id: 0000:02:00.0, compute capability: 7.5)\n[name: \"\/device:CPU:0\"\ndevice_type: \"CPU\"\nmemory_limit: 268435456\nlocality {\n}\nincarnation: 14451063918691325445\n, name: \"\/device:XLA_CPU:0\"\ndevice_type: \"XLA_CPU\"\nmemory_limit: 17179869184\nlocality {\n}\nincarnation: 3472161188084064797\nphysical_device_desc: \"device: XLA_CPU device\"\n, name: \"\/device:XLA_GPU:0\"\ndevice_type: \"XLA_GPU\"\nmemory_limit: 17179869184\nlocality {\n}\nincarnation: 1975372846861552523\nphysical_device_desc: \"device: XLA_GPU device\"\n, name: \"\/device:XLA_GPU:1\"\ndevice_type: \"XLA_GPU\"\nmemory_limit: 17179869184\nlocality {\n}\nincarnation: 8192574289833793917\nphysical_device_desc: \"device: XLA_GPU device\"\n, name: \"\/device:GPU:0\"\ndevice_type: \"GPU\"\nmemory_limit: 996802560\nlocality {\n  bus_id: 1\n  links {\n    link {\n      device_id: 1\n      type: \"StreamExecutor\"\n      strength: 1\n    }\n  }\n}\nincarnation: 9560317274260358344\nphysical_device_desc: \"device: 0, name: GeForce RTX 2080 Ti, pci bus id: 0000:01:00.0, compute capability: 7.5\"\n, name: \"\/device:GPU:1\"\ndevice_type: \"GPU\"\nmemory_limit: 10175270093\nlocality {\n  bus_id: 1\n  links {\n    link {\n      type: \"StreamExecutor\"\n      strength: 1\n    }\n  }\n}\nincarnation: 16877706551631715197\nphysical_device_desc: \"device: 1, name: GeForce RTX 2080 Ti, pci bus id: 0000:02:00.0, compute capability: 7.5\"\n<\/code><\/pre>\n\n<p>and when I try to check the info to use nvidia-docker<\/p>\n\n<pre><code>sudo nvidia-container-cli -k -d \/dev\/tty info \n<\/code><\/pre>\n\n<p>it gives the following output:<\/p>\n\n<pre><code>I0905 09:00:43.600569 9993 driver.c:233] driver service terminated with signal 15\nnvidia-container-cli: initialization error: driver error: timed out\n<\/code><\/pre>\n","answers":[{"AnswerId":"57848647","CreationDate":"2019-09-09T06:09:56.653","ParentId":null,"OwnerUserId":"3398900","Title":null,"Body":"<p>Seems like there is an issue with my nvlink,<\/p>\n\n<p>so if Incase if your training is taking too much Time to start.<\/p>\n\n<p>Please check if the links are up and not down.<\/p>\n\n<p>The following command is as follows.<\/p>\n\n<pre><code>$ nvidia-smi nvlink --status\n\nGPU 0: GeForce RTX 2080 Ti (UUID: GPU-797d7153-ea28-d678-dc38-859b914d6dd7)\n     Link 0: 25.781 GB\/s\n     Link 1: 25.781 GB\/s\nGPU 1: GeForce RTX 2080 Ti (UUID: GPU-8807c553-7571-582d-c2ee-02993527b0a6)\n     Link 0: 25.781 GB\/s\n     Link 1: 25.781 GB\/s\n<\/code><\/pre>\n\n<p>Thanks<\/p>\n"}]}
+{"QuestionId":57804431,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-05T11:24:51.293","AcceptedAnswerId":57806316.0,"OwnerUserId":9925589.0,"Title":"run multiple time a neural network and save only the best weight","Body":"<p>i'm making a convolutional neural network using keras for images classification.<\/p>\n\n<p>i'm making a script which will allow user to train again the model with new data.\nAs you know, model weight initialisation influence model performance, so I want to run automaticaly the same model (same architecture and same data) with different random value for weight initilisation and save the weight of the best model<\/p>\n\n<p>What is a bit tricky is that i'm creating a .exe file and I can't <\/p>\n\n<ul>\n<li>run all model and save their weight in different .h5 files<\/li>\n<li>delet all h5 files which are useless (worst performances)<\/li>\n<\/ul>\n\n<p>What I need is to create only one h5 file, correspoding to model with best performance<\/p>\n\n<p>But I have no idea how i could do that<\/p>\n\n<p>edit: about the code, even if my curent process is totaly normal, I add it for more visualisation<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv2D(24,kernel_size=3,padding='same',activation='relu',\n        input_shape=(n,n,1)))\nmodel.add(MaxPool2D())\nmodel.add(Flatten())\nmodel.add(Dense(128, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(X, activation='softmax'))\nmodel.compile(optimizer=\"adam\", loss=\"categorical_crossentropy\",metrics=[\"accuracy\"])   \n\natch_size=256\nes = EarlyStopping(monitor='val_loss', mode='min', verbose=1, patience=30) # Stop trainning si pas d'am\u00e9liorationapr\u00e8s n = patience\nmc = ModelCheckpoint(ROOT+\"\\\\model_weight_\"+name+'_'+str(n)+\".h5\", monitor='val_loss', mode='min', verbose=1, save_best_only=True) #Sauvegarder meilleur mod\u00e8le\ntrain = model.fit_generator(image_gen.flow(train_X, train_label, batch_size=batch_size),epochs=400,verbose=1,validation_data=(valid_X, valid_label),class_weight=class_weights,callbacks=[es,mc],steps_per_epoch=len(train_X)\/batch_size)\n\n<\/code><\/pre>\n","answers":[{"AnswerId":"57806316","CreationDate":"2019-09-05T13:12:55.390","ParentId":null,"OwnerUserId":"9477338","Title":null,"Body":"<p>I want to speak many things here but first, this should be what you ask for.<\/p>\n\n<pre><code>mc = ModelCheckpoint(ROOT+\"\\\\model_weight_\"+name+\"_{epoch:02d}-{val_loss:.2f}.h5\", monitor='val_loss', mode='min', verbose=1, save_best_only=True)\n<\/code><\/pre>\n\n<p>Where <code>val_loss<\/code> is metrics name Keras prints in training process(like <code>val_acc<\/code> if you put <code>metric=['accuracy']<\/code>).<\/p>\n\n<p>More information <a href=\"https:\/\/keras.io\/callbacks\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/callbacks\/<\/a>.<\/p>\n\n<p>And for deleting files. I didn't check the best weights in training part mainly because the <code>save_best_only=True<\/code>, a loop still needed to check every epoch in <code>train.history<\/code>.<\/p>\n\n<pre><code>import os\nmin_val_loss = float('inf')\nbest_file = ''\n# Find the best weights from file name\nfor weightsfile in os.listdir(ROOT):\n    if int(weightsfile[-7:-3]) &lt; min_val_loss:  # Since I use .2f above, I use -7:-3 here\n        min_val_loss = int(weightsfile[-7:-3])\n        best_file = weightsfile\n\n# delete everything in ROOT except the best.\nfor weightsfile in os.listdir(ROOT):\n    if weightsfile == best_file\n        continue  # do nothing in this loop\n    os.remove(os.path.join(ROOT, weightsfile))  # I recommend you to use `os.path.join` instead of `SOMTHING+\"\\\\\"+ANOTHER`. It might fuck you up later when you move to Linux machine(like Colab).\n\n<\/code><\/pre>\n\n<p>Now, for what I want to speak. Actually the weights initialization doesn't much matter in Deep CNN(where there're 10 or 100 layers). In fact, I trained the same network like <code>ResNet50<\/code> with same setting several times on the same data and it usually only differ for 2% or 3% for the accuracy(around 0.69 to 0.72). And the weights that perform the best on another validation set wasn't the 0.72 one. <\/p>\n\n<p>So, it's better to use the already fine-tuned weights of some good models like <code>ResNet<\/code>, <code>DenseNet<\/code>, <code>NasNet<\/code> or the most recent state-of-the-art <code>EfficientNet<\/code> and waste time on trying different setting like optimizer or learning rate schedule. And most importantly train on free GPU machine like Colab.<\/p>\n"}]}
+{"QuestionId":57804490,"AnswerCount":0,"Tags":"<pytorch><lstm>","CreationDate":"2019-09-05T11:28:36.417","AcceptedAnswerId":null,"OwnerUserId":10232223.0,"Title":"Is the whole hidden state or the last hidden state output is better for discrete level time prediction?","Body":"<p>I am doing a discrete time series prediction problem using the LSTM, I need to understand whether the whole LSTM output or only the last tensor of the LSTM output should be used.<\/p>\n\n<p>I am trying to model a time series containing discrete values like :- 0, 1, 2, 4, 0, 9, 1, 2, 4, 9.....\nI am using a LSTM connected to a Linear layer as the model. A sequence of 4 nos. is fed to the model to predict the 5th number in the sequence(sliding window logic). The inputs are one hot encoded so the dimensions of a single input is (4, 1, 11) where 11 is the cardinality of the series.<\/p>\n\n<p>I am using pytorch, the sequence is correct, the only problem is the output of the LSTM for every input is :-\nLSTM_Output, (LSTM_Hidden, LSTM_cell_state). \nThe LSTM_Output contains the hidden state output for all the four values fed per input and the LSTM_Hidden contains the hidden state output of the last value(among the four values fed per input).<\/p>\n\n<p>Now the Question is :- Which one among the LSTM_Output(containing the hidden state of all the four numbers) and LSTM_Hidden(containing the hidden state of the immediate previous number) should be fed to the Linear Layer.<\/p>\n\n<p>LSTM_Output, (LSTM_Hidden, LSTM_cell_state) = nn.LSTM(input_dim, output_dim)<\/p>\n\n<p><strong>output = F.relu(nn.Linear(LSTM_output.view(1, -1)))\nor\noutput = F.relu(nn.Linear(LSTM_Hidden.view(1, -1)))<\/strong><\/p>\n\n<p>Hence I need to know the input and output dimensions of the Linear layer too.<\/p>\n\n<p>P.S. - I know the logically only the last hidden state i.e. LSTM_Hidden should be fed, but most of the examples I see has the whole hidden state(LSTM_Output) fed to the next layer.<\/p>\n\n<p>Thanks in advance for any help!!<\/p>\n","answers":[]}
+{"QuestionId":57804630,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-05T11:36:26.753","AcceptedAnswerId":null,"OwnerUserId":5208088.0,"Title":"How to clear the keras backend and delete graphs?","Body":"<p>I am using tensorflow (1.14) and keras (2.2) and after some calculations I want to completely erase my session and backend, especially the graph. But even with the most rudimentary example I can't seem to do it. For example, when I run this code, shouldn't the second model.summary() not output anything meaningful because that graph is cleared from the session\/memory?<\/p>\n\n<pre><code>model = keras.models.load_model(\"orig_model.h5\")\n\nmodel.summary()\n\nK.clear_session()\n\n\nmodel.summary()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57804776,"AnswerCount":1,"Tags":"<python><machine-learning><keras><multilingual>","CreationDate":"2019-09-05T11:45:13.697","AcceptedAnswerId":null,"OwnerUserId":11050535.0,"Title":"How to Tokenize the word data column having multiple language with better accuracy.?","Body":"<p>Working on a multilingual classification data-set, Where one of the column have the text in multiple language, I have tried using the simple Tokenization but it has the lower accuracy rate . <\/p>\n\n<p>I have used simple tokenizer, But the accuracy rate is much lower. <\/p>\n\n<p>Dataset have a column with multilanguage text. <\/p>\n\n<pre><code>from keras.preprocessing.text import Tokenizer\nfrom keras.preprocessing.sequence import pad_sequences\ntokenizer = Tokenizer(num_words=18, lower=True)\ntokenizer.fit_on_texts(data.overview)\nsequences = tokenizer.texts_to_sequences(data.overview)\nx = pad_sequences(sequences, maxlen=18)\n<\/code><\/pre>\n\n<p>I have checked on the <strong>universal-sentence-encoder-multilingual-large<\/strong> on TF-Hub but it is not integrating with the keras model.<\/p>\n","answers":[{"AnswerId":"57812188","CreationDate":"2019-09-05T19:53:11.077","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p><code>num_words<\/code> in the <code>Tokenizer<\/code> constructor isn't the sequence length, it's the size of the vocabulary to use. So, you are setting the tokenizer to only keep the 18 most common words. A value around 10,000 to 100,000 is likely to work better, depending on what the dataset you're using looks like.<\/p>\n"}]}
+{"QuestionId":57804948,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><tensorflow-datasets>","CreationDate":"2019-09-05T11:55:07.853","AcceptedAnswerId":null,"OwnerUserId":3478580.0,"Title":"TF dataset of sequence of images as input for TimeDistributed CNN (CNN-LSTM)","Body":"<p>I have a set of images (frames of a video as jpg) and want to create a CNN - LSTM model to classify every image (many to many LSTM). How do I create the Tensorflow 2 dataset as an input for my model?<\/p>\n\n<p>One video contains about 10'000 to 40'000 frames, this sequence would be too long for a LSTM to still preserve information about previous states. I want to cut the videos into sequences of let's say 100 frames and classify them on their own.<\/p>\n\n<p>Here is my TF model (I am using TF hub for a pre-trained image embedding):<\/p>\n\n<pre><code>feature_extractor_url = \"https:\/\/tfhub.dev\/google\/tf2-preview\/inception_v3\/feature_vector\/4\"\n    model = tf.keras.Sequential([\n        tf.keras.layers.TimeDistributed(hub.KerasLayer(feature_extractor_url, input_shape=(width, height, 3), output_shape=[2048],\n                       trainable=True), batch_input_shape=(100, None, 299, 299, 3)),\n        tf.keras.layers.Bidirectional(tf.keras.layers.GRU(1024, return_sequences=True)),\n        tf.keras.layers.Bidirectional(tf.keras.layers.GRU(512, return_sequences=True)),\n        tf.keras.layers.Bidirectional(tf.keras.layers.GRU(256, return_sequences=True)),\n        tf.keras.layers.Dense(2, activation='softmax')\n    ])\n\n    model.compile(\n        optimizer=tf.keras.optimizers.Adam(lr=learning_rate),\n        loss=tf.keras.losses.CategoricalCrossentropy(label_smoothing=label_smoothing),\n        metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>Here is how I load my data:<\/p>\n\n<pre><code>def parse_image(file_path):\n    image = tf.io.read_file(file_path)\n    image = tf.image.decode_jpeg(image, channels=3)\n    image = tf.image.convert_image_dtype(image, tf.float32)\n    image = tf.image.resize(image, [constants.width, constants.height])\n    image \/= 255.0\n    return image\n\n\ndef get_image_ds(dir_name: str, sequence_length: int):\n    data_root = constants.data_root\n    data_root = pathlib.Path(str(data_root) + dir_name)\n    list_ds = tf.data.Dataset.list_files(str(data_root \/ 'all\/*'), shuffle=False)\n    image_data = list_ds.map(parse_image)\n    return image_data.batch(sequence_length, drop_remainder=True)\\\n\n\ndef get_label_ds(filename: str, label_to_index: Dict[str, int]):\n    with open(filename, 'r', encoding='utf8') as file:\n        labels = [line.split(' ')[1].replace('\\n', '') for line in file]\n    all_image_labels = [label_to_index[label_name] for label_name in labels]\n    label_ds = tf.data.Dataset.from_tensor_slices(all_image_labels)\n    return label_ds, len(all_image_labels)\n\n\ndef get_tf_dataset(sequence_length: int, batch_size: int):\n    label_names = ['selected', 'not_selected']\n    label_to_index = dict((name, index) for index, name in enumerate(label_names))\n\n    train_image_data = get_image_ds('\/train', sequence_length)\n    train_labels, n_train_data = get_label_ds('data\/train.txt', label_to_index)\n    train_ds = tf.data.Dataset.zip((train_image_data, train_labels))\\\n        .batch(batch_size)\\\n        .prefetch(buffer_size=AUTOTUNE)\n\n    test_image_data = get_image_ds('\/test', sequence_length)\n    test_labels, _ = get_label_ds('data\/test.txt', label_to_index)\n    test_ds = tf.data.Dataset.zip((test_image_data, test_labels))\\\n        .batch(batch_size)\\\n        .prefetch(buffer_size=AUTOTUNE)\n\n    val_image_data = get_image_ds('\/val', sequence_length)\n    val_labels, _ = get_label_ds('data\/val.txt', label_to_index)\n    val_ds = tf.data.Dataset.zip((val_image_data, val_labels))\\\n        .batch(batch_size)\\\n        .prefetch(buffer_size=AUTOTUNE)\n<\/code><\/pre>\n\n<p>Tensorflow now raises a NotImplementedError after I execute model.fit:<\/p>\n\n<pre><code>hist = model.fit(\n    train_image_data,\n    epochs=5,\n    steps_per_epoch=steps_per_epoch,\n    validation_data=val_image_data,\n    validation_steps=100,\n).history\n<\/code><\/pre>\n\n<p>Error message:<\/p>\n\n<pre><code>Train for 1266.0 steps, validate for 100 steps\nEpoch 1\/5\n\n   1\/1266 [..............................] - ETA: 10:50\n---------------------------------------------------------------------------\n\nNotImplementedError                       Traceback (most recent call last)\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2.py in on_epoch(self, epoch, mode)\n    679     try:\n--&gt; 680       yield epoch_logs\n    681     finally:\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2.py in fit(self, model, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_freq, **kwargs)\n    323                 training_context=training_context,\n--&gt; 324                 total_epochs=epochs)\n    325             cbks.make_logs(model, epoch_logs, training_result, ModeKeys.TRAIN)\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2.py in run_one_epoch(model, iterator, execution_function, dataset_size, batch_size, strategy, steps_per_epoch, num_samples, mode, training_context, total_epochs)\n    122       try:\n--&gt; 123         batch_outs = execution_function(iterator)\n    124       except (StopIteration, errors.OutOfRangeError):\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2_utils.py in execution_function(input_fn)\n     85     return nest.map_structure(_non_none_constant_value,\n---&gt; 86                               distributed_function(input_fn))\n     87 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\eager\\def_function.py in __call__(self, *args, **kwds)\n    426     initializer_map = object_identity.ObjectIdentityDictionary()\n--&gt; 427     self._initialize(args, kwds, add_initializers_to=initializer_map)\n    428     if self._created_variables:\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\eager\\def_function.py in _initialize(self, args, kwds, add_initializers_to)\n    369         self._stateful_fn._get_concrete_function_internal_garbage_collected(  # pylint: disable=protected-access\n--&gt; 370             *args, **kwds))\n    371 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\eager\\function.py in _get_concrete_function_internal_garbage_collected(self, *args, **kwargs)\n   1846       args, kwargs = None, None\n-&gt; 1847     graph_function, _, _ = self._maybe_define_function(args, kwargs)\n   1848     return graph_function\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\eager\\function.py in _maybe_define_function(self, args, kwargs)\n   2146         if graph_function is None:\n-&gt; 2147           graph_function = self._create_graph_function(args, kwargs)\n   2148           self._function_cache.primary[cache_key] = graph_function\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\eager\\function.py in _create_graph_function(self, args, kwargs, override_flat_arg_shapes)\n   2037             override_flat_arg_shapes=override_flat_arg_shapes,\n-&gt; 2038             capture_by_value=self._capture_by_value),\n   2039         self._function_attributes,\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\framework\\func_graph.py in func_graph_from_py_func(name, python_func, args, kwargs, signature, func_graph, autograph, autograph_options, add_control_dependencies, arg_names, op_return_value, collections, capture_by_value, override_flat_arg_shapes)\n    914 \n--&gt; 915       func_outputs = python_func(*func_args, **func_kwargs)\n    916 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\eager\\def_function.py in wrapped_fn(*args, **kwds)\n    319         # the function a weak reference to itself to avoid a reference cycle.\n--&gt; 320         return weak_wrapped_fn().__wrapped__(*args, **kwds)\n    321     weak_wrapped_fn = weakref.ref(wrapped_fn)\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2_utils.py in distributed_function(input_iterator)\n     72     outputs = strategy.experimental_run_v2(\n---&gt; 73         per_replica_function, args=(model, x, y, sample_weights))\n     74     # Out of PerReplica outputs reduce or pick values to return.\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\distribute\\distribute_lib.py in experimental_run_v2(self, fn, args, kwargs)\n    759                                 convert_by_default=False)\n--&gt; 760       return self._extended.call_for_each_replica(fn, args=args, kwargs=kwargs)\n    761 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\distribute\\distribute_lib.py in call_for_each_replica(self, fn, args, kwargs)\n   1786     with self._container_strategy().scope():\n-&gt; 1787       return self._call_for_each_replica(fn, args, kwargs)\n   1788 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\distribute\\distribute_lib.py in _call_for_each_replica(self, fn, args, kwargs)\n   2131         replica_id_in_sync_group=constant_op.constant(0, dtypes.int32)):\n-&gt; 2132       return fn(*args, **kwargs)\n   2133 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\autograph\\impl\\api.py in wrapper(*args, **kwargs)\n    291     with ag_ctx.ControlStatusCtx(status=ag_ctx.Status.DISABLED):\n--&gt; 292       return func(*args, **kwargs)\n    293 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2_utils.py in train_on_batch(model, x, y, sample_weight, class_weight, reset_metrics)\n    263       sample_weights=sample_weights,\n--&gt; 264       output_loss_metrics=model._output_loss_metrics)\n    265 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_eager.py in train_on_batch(model, inputs, targets, sample_weights, output_loss_metrics)\n    310           training=True,\n--&gt; 311           output_loss_metrics=output_loss_metrics))\n    312   if not isinstance(outs, list):\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_eager.py in _process_single_batch(model, inputs, targets, output_loss_metrics, sample_weights, training)\n    251               sample_weights=sample_weights,\n--&gt; 252               training=training))\n    253       if total_loss is None:\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_eager.py in _model_loss(model, inputs, targets, output_loss_metrics, sample_weights, training)\n    126 \n--&gt; 127   outs = model(inputs, **kwargs)\n    128   outs = nest.flatten(outs)\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\base_layer.py in __call__(self, inputs, *args, **kwargs)\n    806                 else:\n--&gt; 807                   outputs = call_fn(cast_inputs, *args, **kwargs)\n    808 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\sequential.py in call(self, inputs, training, mask)\n    254         self._init_graph_network(self.inputs, self.outputs, name=self.name)\n--&gt; 255       return super(Sequential, self).call(inputs, training=training, mask=mask)\n    256 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\network.py in call(self, inputs, training, mask)\n    696 \n--&gt; 697     return self._run_internal_graph(inputs, training=training, mask=mask)\n    698 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\network.py in _run_internal_graph(self, inputs, training, mask)\n    841           # Compute outputs.\n--&gt; 842           output_tensors = layer(computed_tensors, **kwargs)\n    843 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\base_layer.py in __call__(self, inputs, *args, **kwargs)\n    806                 else:\n--&gt; 807                   outputs = call_fn(cast_inputs, *args, **kwargs)\n    808 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\layers\\wrappers.py in call(self, inputs, training, mask)\n    255       # Shape: (num_samples, timesteps, ...)\n--&gt; 256       output_shape = self.compute_output_shape(input_shape).as_list()\n    257       output_shape = self._get_shape_tuple(\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\layers\\wrappers.py in compute_output_shape(self, input_shape)\n    209                                                  input_shape[2:])\n--&gt; 210     child_output_shape = self.layer.compute_output_shape(child_input_shape)\n    211     if not isinstance(child_output_shape, tensor_shape.TensorShape):\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\base_layer.py in compute_output_shape(self, input_shape)\n    598       return nest.map_structure(lambda t: t.shape, outputs)\n--&gt; 599     raise NotImplementedError\n    600 \n\nNotImplementedError: \n\n\nDuring handling of the above exception, another exception occurred:\n\nNotImplementedError                       Traceback (most recent call last)\n\n&lt;ipython-input-8-6bf7fd9bb8ab&gt; in &lt;module&gt;\n      5     callbacks=[tensorboard_callback, cm_callback],\n      6     validation_data=val_image_data,\n----&gt; 7     validation_steps=100,\n      8 ).history\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\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, validation_freq, max_queue_size, workers, use_multiprocessing, **kwargs)\n    732         max_queue_size=max_queue_size,\n    733         workers=workers,\n--&gt; 734         use_multiprocessing=use_multiprocessing)\n    735 \n    736   def evaluate(self,\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2.py in fit(self, model, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_freq, **kwargs)\n    370                       total_epochs=1)\n    371                   cbks.make_logs(model, epoch_logs, eval_result, ModeKeys.TEST,\n--&gt; 372                                  prefix='val_')\n    373 \n    374     return model.history\n\n~\\AppData\\Local\\Programs\\Python\\Python36\\Lib\\contextlib.py in __exit__(self, type, value, traceback)\n     97                 value = type()\n     98             try:\n---&gt; 99                 self.gen.throw(type, value, traceback)\n    100             except StopIteration as exc:\n    101                 # Suppress StopIteration *unless* it's the same exception that\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2.py in on_epoch(self, epoch, mode)\n    682       if mode == ModeKeys.TRAIN:\n    683         # Epochs only apply to `fit`.\n--&gt; 684         self.callbacks.on_epoch_end(epoch, epoch_logs)\n    685       self.progbar.on_epoch_end(epoch, epoch_logs)\n    686 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\callbacks.py in on_epoch_end(self, epoch, logs)\n    295     logs = logs or {}\n    296     for callback in self.callbacks:\n--&gt; 297       callback.on_epoch_end(epoch, logs)\n    298 \n    299   def on_train_batch_begin(self, batch, logs=None):\n\n&lt;ipython-input-6-ce1724da9329&gt; in log_confusion_matrix(epoch, logs)\n      6 def log_confusion_matrix(epoch, logs):\n      7     # Use the model to predict the values from the validation dataset.\n----&gt; 8     test_pred_raw = model.predict(test_image_data)\n      9     test_pred = np.argmax(test_pred_raw, axis=1)\n     10 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training.py in predict(self, x, batch_size, verbose, steps, callbacks, max_queue_size, workers, use_multiprocessing)\n    913         max_queue_size=max_queue_size,\n    914         workers=workers,\n--&gt; 915         use_multiprocessing=use_multiprocessing)\n    916 \n    917   def reset_metrics(self):\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2.py in predict(self, model, x, batch_size, verbose, steps, callbacks, **kwargs)\n    460     return self._model_iteration(\n    461         model, ModeKeys.PREDICT, x=x, batch_size=batch_size, verbose=verbose,\n--&gt; 462         steps=steps, callbacks=callbacks, **kwargs)\n    463 \n    464 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2.py in _model_iteration(self, model, mode, x, y, batch_size, verbose, sample_weight, steps, callbacks, **kwargs)\n    442               mode=mode,\n    443               training_context=training_context,\n--&gt; 444               total_epochs=1)\n    445           cbks.make_logs(model, epoch_logs, result, mode)\n    446 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2.py in run_one_epoch(model, iterator, execution_function, dataset_size, batch_size, strategy, steps_per_epoch, num_samples, mode, training_context, total_epochs)\n    121         step=step, mode=mode, size=current_batch_size) as batch_logs:\n    122       try:\n--&gt; 123         batch_outs = execution_function(iterator)\n    124       except (StopIteration, errors.OutOfRangeError):\n    125         # TODO(kaftan): File bug about tf function and errors.OutOfRangeError?\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2_utils.py in execution_function(input_fn)\n     84     # `numpy` translates Tensors to values in Eager mode.\n     85     return nest.map_structure(_non_none_constant_value,\n---&gt; 86                               distributed_function(input_fn))\n     87 \n     88   return execution_function\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\eager\\def_function.py in __call__(self, *args, **kwds)\n    425     # This is the first call of __call__, so we have to initialize.\n    426     initializer_map = object_identity.ObjectIdentityDictionary()\n--&gt; 427     self._initialize(args, kwds, add_initializers_to=initializer_map)\n    428     if self._created_variables:\n    429       try:\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\eager\\def_function.py in _initialize(self, args, kwds, add_initializers_to)\n    368     self._concrete_stateful_fn = (\n    369         self._stateful_fn._get_concrete_function_internal_garbage_collected(  # pylint: disable=protected-access\n--&gt; 370             *args, **kwds))\n    371 \n    372     def invalid_creator_scope(*unused_args, **unused_kwds):\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\eager\\function.py in _get_concrete_function_internal_garbage_collected(self, *args, **kwargs)\n   1845     if self.input_signature:\n   1846       args, kwargs = None, None\n-&gt; 1847     graph_function, _, _ = self._maybe_define_function(args, kwargs)\n   1848     return graph_function\n   1849 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\eager\\function.py in _maybe_define_function(self, args, kwargs)\n   2145         graph_function = self._function_cache.primary.get(cache_key, None)\n   2146         if graph_function is None:\n-&gt; 2147           graph_function = self._create_graph_function(args, kwargs)\n   2148           self._function_cache.primary[cache_key] = graph_function\n   2149         return graph_function, args, kwargs\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\eager\\function.py in _create_graph_function(self, args, kwargs, override_flat_arg_shapes)\n   2036             arg_names=arg_names,\n   2037             override_flat_arg_shapes=override_flat_arg_shapes,\n-&gt; 2038             capture_by_value=self._capture_by_value),\n   2039         self._function_attributes,\n   2040         # Tell the ConcreteFunction to clean up its graph once it goes out of\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\framework\\func_graph.py in func_graph_from_py_func(name, python_func, args, kwargs, signature, func_graph, autograph, autograph_options, add_control_dependencies, arg_names, op_return_value, collections, capture_by_value, override_flat_arg_shapes)\n    913                                           converted_func)\n    914 \n--&gt; 915       func_outputs = python_func(*func_args, **func_kwargs)\n    916 \n    917       # invariant: `func_outputs` contains only Tensors, CompositeTensors,\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\eager\\def_function.py in wrapped_fn(*args, **kwds)\n    318         # __wrapped__ allows AutoGraph to swap in a converted function. We give\n    319         # the function a weak reference to itself to avoid a reference cycle.\n--&gt; 320         return weak_wrapped_fn().__wrapped__(*args, **kwds)\n    321     weak_wrapped_fn = weakref.ref(wrapped_fn)\n    322 \n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2_utils.py in distributed_function(input_iterator)\n     71     strategy = distribution_strategy_context.get_strategy()\n     72     outputs = strategy.experimental_run_v2(\n---&gt; 73         per_replica_function, args=(model, x, y, sample_weights))\n     74     # Out of PerReplica outputs reduce or pick values to return.\n     75     all_outputs = dist_utils.unwrap_output_dict(\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\distribute\\distribute_lib.py in experimental_run_v2(self, fn, args, kwargs)\n    758       fn = autograph.tf_convert(fn, ag_ctx.control_status_ctx(),\n    759                                 convert_by_default=False)\n--&gt; 760       return self._extended.call_for_each_replica(fn, args=args, kwargs=kwargs)\n    761 \n    762   def reduce(self, reduce_op, value, axis):\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\distribute\\distribute_lib.py in call_for_each_replica(self, fn, args, kwargs)\n   1785       kwargs = {}\n   1786     with self._container_strategy().scope():\n-&gt; 1787       return self._call_for_each_replica(fn, args, kwargs)\n   1788 \n   1789   def _call_for_each_replica(self, fn, args, kwargs):\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\distribute\\distribute_lib.py in _call_for_each_replica(self, fn, args, kwargs)\n   2130         self._container_strategy(),\n   2131         replica_id_in_sync_group=constant_op.constant(0, dtypes.int32)):\n-&gt; 2132       return fn(*args, **kwargs)\n   2133 \n   2134   def _reduce_to(self, reduce_op, value, destinations):\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\autograph\\impl\\api.py in wrapper(*args, **kwargs)\n    290   def wrapper(*args, **kwargs):\n    291     with ag_ctx.ControlStatusCtx(status=ag_ctx.Status.DISABLED):\n--&gt; 292       return func(*args, **kwargs)\n    293 \n    294   if inspect.isfunction(func) or inspect.ismethod(func):\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2_utils.py in _predict_on_batch(***failed resolving arguments***)\n    160     def _predict_on_batch(model, x, y=None, sample_weights=None):\n    161       del y, sample_weights\n--&gt; 162       return predict_on_batch(model, x)\n    163 \n    164     func = _predict_on_batch\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2_utils.py in predict_on_batch(model, x)\n    368 \n    369   with backend.eager_learning_phase_scope(0):\n--&gt; 370     return model(inputs)  # pylint: disable=not-callable\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\base_layer.py in __call__(self, inputs, *args, **kwargs)\n    805                     outputs = base_layer_utils.mark_as_return(outputs, acd)\n    806                 else:\n--&gt; 807                   outputs = call_fn(cast_inputs, *args, **kwargs)\n    808 \n    809             except errors.OperatorNotAllowedInGraphError as e:\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\sequential.py in call(self, inputs, training, mask)\n    253       if not self.built:\n    254         self._init_graph_network(self.inputs, self.outputs, name=self.name)\n--&gt; 255       return super(Sequential, self).call(inputs, training=training, mask=mask)\n    256 \n    257     outputs = inputs  # handle the corner case where self.layers is empty\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\network.py in call(self, inputs, training, mask)\n    695                                 ' implement a `call` method.')\n    696 \n--&gt; 697     return self._run_internal_graph(inputs, training=training, mask=mask)\n    698 \n    699   def compute_output_shape(self, input_shape):\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\network.py in _run_internal_graph(self, inputs, training, mask)\n    840 \n    841           # Compute outputs.\n--&gt; 842           output_tensors = layer(computed_tensors, **kwargs)\n    843 \n    844           # Update tensor_dict.\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\base_layer.py in __call__(self, inputs, *args, **kwargs)\n    805                     outputs = base_layer_utils.mark_as_return(outputs, acd)\n    806                 else:\n--&gt; 807                   outputs = call_fn(cast_inputs, *args, **kwargs)\n    808 \n    809             except errors.OperatorNotAllowedInGraphError as e:\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\layers\\wrappers.py in call(self, inputs, training, mask)\n    254       y = self.layer(inputs, **kwargs)\n    255       # Shape: (num_samples, timesteps, ...)\n--&gt; 256       output_shape = self.compute_output_shape(input_shape).as_list()\n    257       output_shape = self._get_shape_tuple(\n    258           (-1, input_length), y, 1, output_shape[2:])\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\layers\\wrappers.py in compute_output_shape(self, input_shape)\n    208     child_input_shape = tensor_shape.TensorShape([input_shape[0]] +\n    209                                                  input_shape[2:])\n--&gt; 210     child_output_shape = self.layer.compute_output_shape(child_input_shape)\n    211     if not isinstance(child_output_shape, tensor_shape.TensorShape):\n    212       child_output_shape = tensor_shape.TensorShape(child_output_shape)\n\nc:\\users\\yanick\\code\\pycharmprojects\\avisec\\venv\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\base_layer.py in compute_output_shape(self, input_shape)\n    597                                       'layer (%s).' % self.__class__.__name__)\n    598       return nest.map_structure(lambda t: t.shape, outputs)\n--&gt; 599     raise NotImplementedError\n    600 \n    601   @doc_controls.for_subclass_implementers\n\nNotImplementedError: \n```\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57805113,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2019-09-05T12:05:27.563","AcceptedAnswerId":null,"OwnerUserId":3118805.0,"Title":"Making predictions from Keras Models","Body":"<p>I have a trained h5 model, the training data for that model required pre-processing to remove unwanted columns\/features.\nI want to run live inference on that model now, the live data however has lots of columns - how do I extract the exact features from the live data that were used to train the model..?\nDo I have to do this or will the keras model see that I have the various columns in my new data and automatically make the right selection and ignore the other columns..?<\/p>\n","answers":[]}
+{"QuestionId":57805332,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><python-3.7>","CreationDate":"2019-09-05T12:18:57.007","AcceptedAnswerId":null,"OwnerUserId":9318372.0,"Title":"Using tensorflow models both as standalone and subgraphs","Body":"<p>I want to do the following:<\/p>\n\n<p>Create a model that can be used in two ways;<\/p>\n\n<ol>\n<li>Use it as a standalone model (fit it using some data)<\/li>\n<li>Use it as a subgraph in another model<\/li>\n<\/ol>\n\n<p>Ideally I want to have a class <code>MyModel<\/code> that I can initialize <code>model=MyModel(hyperparameters)<\/code> and which supports <code>model.fit(X, Y)<\/code>, <code>model.predict(X)<\/code> and <code>model.evaluate(X, Y)<\/code>  methods.<\/p>\n\n<p>At the same time I want to be able to feed it <code>tf.Variables<\/code> in order to use it as part in  a bigger graph <code>latent_variable = model(x)<\/code> where <code>x<\/code> is a <code>tf.Variable<\/code><\/p>\n\n<p>Is there any way this is possible to do in <code>tensorflow<\/code>? Using tf.keras.Model doesn't work because it always returns a <code>tf.Tensor<\/code>, but not <code>tf.Variable<\/code> even if a <code>tf.Variable<\/code> is fed in.<\/p>\n","answers":[]}
+{"QuestionId":57805635,"AnswerCount":0,"Tags":"<tensorflow><text-to-speech>","CreationDate":"2019-09-05T12:35:01.863","AcceptedAnswerId":null,"OwnerUserId":778508.0,"Title":"Tacotron With TensorFlow Lite","Body":"<p>Is it possible to use tacotron implementation with TensorFlow Lite?\nI used keith ito's implementation of tacotron and I woud like to use TFLite.\nBut I don't know how to change the code. TFLite is about only the Inference?\nApparentley, I can't use Session, Variable and other things, righ?<\/p>\n","answers":[]}
+{"QuestionId":57805909,"AnswerCount":0,"Tags":"<keras><time-series><lstm>","CreationDate":"2019-09-05T12:50:55.843","AcceptedAnswerId":null,"OwnerUserId":1255603.0,"Title":"Advice on improving Human Activity model using Keras","Body":"<p>I am new to Deep Learning.  I am interested in determining whether my (older) parents fell based on accelerometer data.  I have built a model using open source data for walking, jogging, and such to understand timeseries w\/ Deep Learning.  <a href=\"https:\/\/github.com\/BitKnitting\/HAR_LSTM_COLAB\/blob\/master\/HAR_LSTM_FirstAttempt.ipynb\" rel=\"nofollow noreferrer\">Here is the python notebook I have created<\/a>  <\/p>\n\n<p>I'd be very grateful for any ideas on how I can improve the model, or how I do things.  I've learned through YouTube, reading the first four chapters of Deep Learning With Python.  <\/p>\n\n<p>Thank you.<\/p>\n\n<p>I've tried several models. Gotten to know the data, applied advice from several articles and YouTube videos.<\/p>\n","answers":[]}
+{"QuestionId":57806063,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><tensorrt>","CreationDate":"2019-09-05T12:59:40.870","AcceptedAnswerId":null,"OwnerUserId":12025393.0,"Title":"cannot import name '_pywrap_utils' from 'tensorflow.python'","Body":"<p>I am working on pose estimation using OpenPose. For that I installed <code>TensorFlow<\/code> GPU and installed all the requirements including CUDA development kit.<\/p>\n\n<p>While running the Python script: \n<code>C:\\Users\\abhi\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\ops\\resource_variable_ops.py<\/code>, I encountered the following error:<\/p>\n\n<blockquote>\n  <p>ImportError: cannot import name '_pywrap_utils' from\n  'tensorflow.python'\n  (C:\\Users\\abhi\\Anaconda3\\lib\\site-packages\\tensorflow\\python__init__.py)<\/p>\n<\/blockquote>\n\n<p>I tried searching for _pywrap_utils file but there was no such file.<\/p>\n","answers":[{"AnswerId":"57807062","CreationDate":"2019-09-05T13:51:49.800","ParentId":null,"OwnerUserId":"11787588","Title":null,"Body":"<p>Try <code>pip3 install pywrap<\/code> and <code>pip3 install tensorflow<\/code> pywrap utils should be included with tensorflow. If it is not found, that means TF was not installed correctly.<\/p>\n"}]}
+{"QuestionId":57806233,"AnswerCount":0,"Tags":"<tensorflow><tensorboard><tensorflow2.0>","CreationDate":"2019-09-05T13:08:28.197","AcceptedAnswerId":null,"OwnerUserId":10101844.0,"Title":"How to export summaries with the tensorboard backend in tensorflow 2.0?","Body":"<p>Before tensorflow 2.0 one could export data with tensorboard using:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorboard.backend.event_processing.event_accumulator import EventAccumulator\n\nevent_acc = EventAccumulator('\/...\/events.out.tfevents.1567583454.ragnar-System-Product-Name.7433.6241.v2').Reload()\nprint(event_acc.Tags())\n<\/code><\/pre>\n\n<p>This would give a list of tags to be accessed with <code>event_acc.Scalar(tag)<\/code>.\nIn tensorflow 2.0 however I get:<br>\n<code>{'images': [], 'audio': [], 'histograms': [], 'scalars': [], 'distributions': [], 'tensors': ['loss'], 'graph': False, 'meta_graph': False, 'run_metadata': []}\n<\/code>, that is the scalar tags have been replaced with <code>tensors<\/code> tags. These can be accessed with  <code>event_acc.Tensors(tag)<\/code> but unlike <code>event_acc.Scalar(tag)<\/code> I can't find documentation on how to translate these into a csv format. Also only a subset of the data from tensorboard seems to be contained in <code>event_acc.Tensors(tag)<\/code>.\nHow can I export csv data from tensorboard for tensorflow 2.0 event files?<\/p>\n","answers":[]}
+{"QuestionId":57806408,"AnswerCount":0,"Tags":"<python><keras><time-series><classification><lstm>","CreationDate":"2019-09-05T13:17:02.840","AcceptedAnswerId":null,"OwnerUserId":7045987.0,"Title":"Loss and accuracy remains constant in time series classification by LSTM","Body":"<p>I have a time series data with a classification label of 1 and 0. \nI am using a LSTM model to classify the series by taking 100 consecutive timestamps as input with a single label.<\/p>\n\n<p>Even after training for many epochs, the loss and accuracy remains constant without even a little variation.\nI have tried to change the optimizer(ADAM and SGD), learning rate and number of layers but still the issue remains.<\/p>\n\n<p>Any idea on why this could be happening? <\/p>\n\n<p>The problem is similar to <a href=\"https:\/\/stackoverflow.com\/questions\/51064649\/lstm-to-classify-long-time-series-sequences\">this<\/a> but since I couldn't find any solution, I am posting a question.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/WaBdf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/WaBdf.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57806642,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2019-09-05T13:30:53.013","AcceptedAnswerId":57806791.0,"OwnerUserId":2851685.0,"Title":"My DNN model to compare MNIST dataset using tensorflow not making progress","Body":"<p>I am doing exercise of DNN chapter from \"Hands on machine learning using scikit and tensorflow\" book. Below are the steps I am taking<\/p>\n\n<p><strong>Steps<\/strong><\/p>\n\n<ul>\n<li>Generate new data set from MNIST such that each row has two images and its label is 1 if the images represent same digit else 0. 50% rows are of same digit and other 50% are of different digits<\/li>\n<li>create 2 DNNs dnn1 and dnn2 with 5 hidden layers and not output layer. Each DNN takes one image as input<\/li>\n<li>Join output of dnn1 and dnn2 and feed it to a layer with 10 neurons<\/li>\n<li>add last layer with single layer <\/li>\n<\/ul>\n\n<p>Here is the link of <a href=\"https:\/\/colab.research.google.com\/drive\/16dR2ZL3BkF6EOXWcRo7R9ltZ6oUhqpJx#scrollTo=Okxup-QSOVyG&amp;uniqifier=4\" rel=\"nofollow noreferrer\">google colab<\/a> where I implemented it. It is not making any progress when training. But the solution provided by book is working, here is the <a href=\"https:\/\/colab.research.google.com\/drive\/1lMNYU4KD9U3t_X15ZOaiB_ZEhjl2XVPA\" rel=\"nofollow noreferrer\">google colab<\/a> link. My implementation and the books implementation is slightly different but they are similar with same steps. I have spent two days trying to figure out what is different but I am not able to find it.<\/p>\n\n<p>May be I have done a silly mistake, It would be great help to me if someone can point me what it is. I am new to tensorflow, someone with experience in it may easily see it. <\/p>\n\n<p><a href=\"https:\/\/colab.research.google.com\/drive\/16dR2ZL3BkF6EOXWcRo7R9ltZ6oUhqpJx#scrollTo=Okxup-QSOVyG&amp;uniqifier=4\" rel=\"nofollow noreferrer\">My implementation in google colab<\/a><br>\n<a href=\"https:\/\/colab.research.google.com\/drive\/1lMNYU4KD9U3t_X15ZOaiB_ZEhjl2XVPA\" rel=\"nofollow noreferrer\">Books solution in google colab<\/a><\/p>\n","answers":[{"AnswerId":"57806791","CreationDate":"2019-09-05T13:38:14.520","ParentId":null,"OwnerUserId":"10532452","Title":null,"Body":"<p>It doesn't look like you normalized your data anywhere. The book solution does this right at the beginning:<\/p>\n\n<pre><code>X_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\n<\/code><\/pre>\n"}]}
+{"QuestionId":57806665,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-05T13:31:34.133","AcceptedAnswerId":null,"OwnerUserId":826983.0,"Title":"Use tf.cond to implement pretraining","Body":"<p>I have a list of <code>tf.data.Dataset<\/code> objects and I would like to implement pretraining based ony a <code>tf.cond<\/code>.<\/p>\n\n<p>I don't know if this is possible but here is the main idea: Train on the first dataset for a certain number of steps (here <code>400<\/code>) and then switch over to another schedule:<\/p>\n\n<pre><code>def _pretrain(self, datasets) -&gt; tf.data.Dataset:\n\n    global_step = tf.train.get_global_step()\n\n    # output_types = self._const_schedule(datasets).output_types\n\n    pretrain_dataset = datasets[0].make_one_shot_iterator().get_next()\n    train_dataset = self._const_schedule(datasets).make_one_shot_iterator().get_next()\n\n    conditional_dataset = tf.cond(\n        global_step &lt; 400,\n        true_fn=lambda: pretrain_dataset,\n        false_fn=lambda: train_dataset\n    )\n\n    return tf.data.Dataset.from???(conditional_dataset)\n<\/code><\/pre>\n\n<p>This \"works\" in the sense of not crashing if I use <code>Dataset.from_tensor()<\/code> but it seems in this case the data is not getting presented correctly.. It looks as if only one sample gets presented.. I don't really know.<\/p>\n\n<p>The question is whether the above code can be changed s.t. the pretrain implementation works as intended. To be precise, I need to know whether I can transform <code>conditional_dataset<\/code> to an actual repeating <code>tf.data.Dataset<\/code>. <\/p>\n\n<p>The main problem here is that <code>tf.cond<\/code> expects tensors from the two callback functions.<\/p>\n","answers":[]}
+{"QuestionId":57806668,"AnswerCount":1,"Tags":"<python><swift><keras><coreml><mlmodel>","CreationDate":"2019-09-05T13:31:53.967","AcceptedAnswerId":57809350.0,"OwnerUserId":10171211.0,"Title":"How to make a MLModel updatable in Swift CoreML?","Body":"<p>I'm trying to implement the new CoreML 3 feature of updatable Models into my App, but I can't figure it out.<\/p>\n\n<p>Im creating a Neural Network with 2 layers in Keras and convert it to a core ML Model. Then I include the model into my iOS Project. \nHowever <code>self.testmodel.model.modelDescription.isUpdatable<\/code> is always false, and I can't make sense of the apple docs.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential([\n    Dense(10, activation=\"sigmoid\", input_shape=(2,)),\n    Dense(2, activation=\"relu\"),\n])\ncore_mlmodel = coremltools.converters.keras.convert(model)\ncore_mlmodel.save(\"FirstNN.mlmodel\")\n<\/code><\/pre>\n\n<p>Integrated in Swift I can now use the Model, but can't update it<\/p>\n\n<pre class=\"lang-swift prettyprint-override\"><code>let testmodel = FirstNN()\ntry testmodel.prediction(input: input) \/\/ works\ntestmodel.model.modelDescription.isUpdatable  \/\/ is false\n<\/code><\/pre>\n\n<p>Why isn't my model updatable and how can I change that?<\/p>\n","answers":[{"AnswerId":"57809350","CreationDate":"2019-09-05T16:09:37.717","ParentId":null,"OwnerUserId":"7501629","Title":null,"Body":"<p>You can pass the <code>respect_trainable=True<\/code> argument to <code>coremltools.converters.keras.convert()<\/code>.<\/p>\n\n<p>Alternatively, you can change the mlmodel file after the conversion to make the model updatable. Official documentation in the form of examples is here: <a href=\"https:\/\/github.com\/apple\/coremltools\/tree\/master\/examples\/updatable_models\" rel=\"nofollow noreferrer\">https:\/\/github.com\/apple\/coremltools\/tree\/master\/examples\/updatable_models<\/a><\/p>\n"}]}
+{"QuestionId":57806747,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2019-09-05T13:35:55.357","AcceptedAnswerId":null,"OwnerUserId":7820400.0,"Title":"model.evaluate returns different values on each run using the same saved model","Body":"<p>I'm trying to make a movie recommender system using keras and the movie lens 1m dataset, but every time i run the model.evaluate (after saving the model), I get different results, this is the first time this happens, I've made a few models before this one using the embedding layer and they worked fine, can someone explain to me why does this happen? Thanks =)<\/p>\n\n<pre><code>train, test = train_test_split(df, test_size=0.2, random_state=42)\nprint(test.head(10))\ntrain_rating = train.rating\ntest_rating = test.rating\ntrain.drop(['rating'], axis=1, inplace=True)\ntest.drop(['rating'], axis=1, inplace=True)\n\n# constructing the model\nmodel = Sequential()\nmodel.add(Dense(64, input_shape=(4,), activation='linear'))\nmodel.add(Dense(32, activation='linear'))\nmodel.add(Dense(16, activation='linear'))\nmodel.add(Dense(8, activation='linear'))\nmodel.add(Dense(1))\nmodel.compile(optimizer=optimizers.Adam(), loss='mean_absolute_error')\n\n#Training the predicting\nif(os.path.exists('reg.h5')):\n    load_model('reg.h5')\nelse:\n    # model.fit([train.age, train.gender, train.occupation, train.genre], train_rating)\n    model.fit(train, train_rating,  epochs=10, validation_split=0.2)\n    model.save('reg.h5')\n\neffeciency = model.evaluate(test, test_rating)\nprint(effeciency)\n\npredictions = model.predict(test.head(10))\nfor i in range(0, 10):\n    print(predictions[i], test_rating.iloc[i])\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57806915,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-05T13:44:29.373","AcceptedAnswerId":57807183.0,"OwnerUserId":6476080.0,"Title":"tensorflow - how to select all elements in an array except for a sequence of indices?","Body":"<p>The equivalent numpy operation can be done using <code>np.delete<\/code> as specified <a href=\"https:\/\/stackoverflow.com\/questions\/47540800\/numpy-how-to-select-all-elements-in-an-array-except-for-a-sequence-of-indices?noredirect=1&amp;lq=1\">here<\/a>. Since there's no <code>tf.delete<\/code>, I am not sure how to do this in <code>tensorflow<\/code>.<\/p>\n","answers":[{"AnswerId":"57807183","CreationDate":"2019-09-05T13:59:41.723","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>This is one way to do that:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef delete_tf(a, idx, axis=0):\n    n = tf.shape(a)[axis]\n    t = tf.ones_like(idx, dtype=tf.bool)\n    m = ~tf.scatter_nd(tf.expand_dims(idx, 1), t, [n])\n    return tf.boolean_mask(a, m, axis=axis)\n\nwith tf.Graph().as_default(), tf.Session() as sess:\n    data = tf.reshape(tf.range(12), [3, 4])\n    print(sess.run(delete_tf(data, [1], 0)))\n    # [[ 0  1  2  3]\n    #  [ 8  9 10 11]]\n    print(sess.run(delete_tf(data, [0, 2], 1)))\n    # [[ 1  3]\n    #  [ 5  7]\n    #  [ 9 11]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57806973,"AnswerCount":0,"Tags":"<keras><neural-network><transfer-learning><vgg-net>","CreationDate":"2019-09-05T13:47:32.177","AcceptedAnswerId":null,"OwnerUserId":3802483.0,"Title":"Keras Transfer Learning Issue","Body":"<p>I have trained &amp; saved a smaller network on my small dataset, and I want to use transfer learning. <\/p>\n\n<p>I want to use this saved network on top of the conv part of the pretrained VGG16, specifically I want to freeze some layers of VGG but not all then I want to use the fc that I have already trained on my smaller dataset, and learn a model which is a combination of both with transferred weights. <\/p>\n\n<p>I am following a mish and mash of tutorials: <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> and <a href=\"https:\/\/machinelearningmastery.com\/how-to-develop-a-convolutional-neural-network-to-classify-photos-of-dogs-and-cats\/\" rel=\"nofollow noreferrer\">https:\/\/machinelearningmastery.com\/how-to-develop-a-convolutional-neural-network-to-classify-photos-of-dogs-and-cats\/<\/a> I do not just want to use pretrained features, and I do not just want to add two new layers to the VGG's conv net, as I mentioned, I want to transfer the fc layers of the smaller network and freeze all blocks of conv layers but one of VGGs and train again. Below is my code but I get an error (no matter how I tried to change around the code, I get a similar error)<\/p>\n\n<pre><code>from keras import applications\nfrom keras.preprocessing.image import ImageDataGenerator\nfrom keras import optimizers\nfrom keras.models import Sequential, Model\nfrom keras.layers import Dropout, Flatten, Dense\n\n# path to the model weights files.\nweights_path = '\/home\/d\/Desktop\/s\/vgg16_weights.h5'\ntop_model_weights_path = '\/home\/d\/Desktop\/s\/model_weights.h5'\n# dimensions of our images.\nimg_width, img_height = 256, 256\n\n# build the VGG16 network\nbase_model = applications.VGG16(weights='imagenet', include_top=False, input_shape=(256, 256, 3))\nprint('Model loaded.')\n\n# set the first 25 layers (up to the last conv block)\n# to non-trainable (weights will not be updated)\nfor layer in base_model.layers[:15]:\n    layer.trainable = False\n\n# build a classifier model to put on top of the convolutional model\ntop_model = Sequential()\ntop_model.add(Flatten(input_shape=base_model.output_shape[1:]))\ntop_model.add(Dense(256, activation='relu'))\ntop_model.add(Dropout(0.5))\ntop_model.add(Dense(1, activation='sigmoid'))\n\n\n# note that it is necessary to start with a fully-trained\n# classifier, including the top classifier,\n# in order to successfully do fine-tuning\ntop_model.load_weights(top_model_weights_path)\n\nmodel= Model(inputs=base_model.input, outputs=top_model(base_model.output))\n\n\n\n# add the model on top of the convolutional base\n#model.add(top_model)\n\n\nprint(top_model.summary())\n\n# compile the model with a SGD\/momentum optimizer\n# and a very slow learning rate.\nmodel.compile(loss='binary_crossentropy',\n              optimizer=optimizers.SGD(lr=1e-4, momentum=0.9),\n              metrics=['accuracy'])\n\nimport matplotlib.pyplot as plt\nfrom keras.preprocessing.image import ImageDataGenerator\nimport os\n\nos.environ[\"CUDA_VISIBLE_DEVICES\"]=\"0\"\n\ntrain_dir = '\/home\/d\/Desktop\/s\/data\/train'\neval_dir = '\/home\/d\/Desktop\/s\/data\/eval'\ntest_dir = '\/home\/d\/Desktop\/s\/data\/test'\n\n\n\n# create a data generator\ntrain_datagen = ImageDataGenerator(rescale=1.\/255,   #Scale the image between 0 and 1\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\nval_datagen = ImageDataGenerator(rescale=1.\/255)  #We do not augment validation data. we only perform rescale\n\ntest_datagen = ImageDataGenerator(rescale=1.\/255)  #We do not augment validation data. we only perform rescale\n\n# load and iterate training dataset\ntrain_generator = train_datagen.flow_from_directory(train_dir, class_mode='binary', batch_size=16, shuffle='True', seed=42)\n# load and iterate validation dataset\nval_generator = val_datagen.flow_from_directory(eval_dir, class_mode='binary', batch_size=16, shuffle='True', seed=42)\n# load and iterate test dataset\ntest_generator = test_datagen.flow_from_directory(test_dir, class_mode=None, batch_size=1, shuffle='False', seed=42)\n\n#The training part\n#We train for 64 epochs with about 100 steps per epoch\nhistory = model.fit_generator(train_generator,\n                              steps_per_epoch=train_generator.n \/\/ train_generator.batch_size,\n                              epochs=6,\n                              validation_data=val_generator,\n                              validation_steps=val_generator.n \/\/ val_generator.batch_size)\n<\/code><\/pre>\n\n<p>The error I am getting is:<\/p>\n\n<pre><code>Model loaded.\nTraceback (most recent call last):\n  File \"\/home\/d\/Desktop\/s\/transferLearningS.py\", line 33, in &lt;module&gt;\n    top_model.load_weights(top_model_weights_path)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/engine\/network.py\", line 1166, in load_weights\n    f, self.layers, reshape=reshape)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/engine\/saving.py\", line 1030, in load_weights_from_hdf5_group\n    str(len(filtered_layers)) + ' layers.')\nValueError: You are trying to load a weight file containing 6 layers into a model with 2 layers.\n<\/code><\/pre>\n\n<p>And my smaller network is built this way:<\/p>\n\n<pre><code>model = models.Sequential()\nmodel.add(layers.Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', input_shape=(256, 256, 3)))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(Dropout(0.2))\nmodel.add(layers.Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform'))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(Dropout(0.2))\nmodel.add(layers.Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_uniform'))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(Dropout(0.2))\nmodel.add(layers.Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_uniform'))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(Dropout(0.2))\nmodel.add(layers.Flatten())\nmodel.add(layers.Dropout(0.5))  #Dropout for regularization\nmodel.add(layers.Dense(512, activation='relu'))\nmodel.add(layers.Dense(1, activation='sigmoid'))  #Sigmoid function at the end because we have just two classes\n<\/code><\/pre>\n\n<p>Any recommendations how I can fix this issue? <\/p>\n","answers":[]}
+{"QuestionId":57806980,"AnswerCount":1,"Tags":"<pytorch><pytorch-ignite>","CreationDate":"2019-09-05T13:47:44.293","AcceptedAnswerId":58059762.0,"OwnerUserId":3513738.0,"Title":"How can I use the LBFGS optimizer with pytorch ignite?","Body":"<p>I started using Ignite recently and i found it very interesting.\nI would like to train a model using as an optimizer the LBFGS algorithm from the <code>torch.optim<\/code> module.<\/p>\n\n<p>This is my code:<\/p>\n\n<pre><code>from ignite.engine import Events, Engine, create_supervised_trainer, create_supervised_evaluator\nfrom ignite.metrics import RootMeanSquaredError, Loss\nfrom ignite.handlers import EarlyStopping\n\n D_in, H, D_out = 5, 10, 1\n model = simpleNN(D_in, H, D_out) # a simple MLP with 1 Hidden Layer\n model.double()\n train_loader, val_loader = get_data_loaders(i)\n\n optimizer = torch.optim.LBFGS(model.parameters(), lr=1)\n loss_func = torch.nn.MSELoss()  \n\n #Ignite\n trainer = create_supervised_trainer(model, optimizer, loss_func)\n evaluator = create_supervised_evaluator(model, metrics={'RMSE': RootMeanSquaredError(),'LOSS': Loss(loss_func)})\n\n @trainer.on(Events.ITERATION_COMPLETED)\n def log_training_loss(engine):\n     print(\"Epoch[{}] Loss: {:.5f}\".format(engine.state.epoch, len(train_loader), engine.state.output))\n\ndef score_function(engine):\n    val_loss = engine.state.metrics['RMSE']\n    print(\"VAL_LOSS: {:.5f}\".format(val_loss))\n    return -val_loss\n\nhandler = EarlyStopping(patience=10, score_function=score_function, trainer=trainer)\nevaluator.add_event_handler(Events.COMPLETED, handler)\n\ntrainer.run(train_loader, max_epochs=100)\n<\/code><\/pre>\n\n<p>And the error that raises is:\n<code>TypeError: step() missing 1 required positional argument: 'closure'<\/code><\/p>\n\n<p>I know that is required to define a closure for the implementation of LBFGS, so my question is how can I do it using ignite? or is there another approach for doing this?<\/p>\n","answers":[{"AnswerId":"58059762","CreationDate":"2019-09-23T09:50:23.307","ParentId":null,"OwnerUserId":"3513738","Title":null,"Body":"<p>The way to do it is like this:<\/p>\n\n<pre><code>from ignite.engine import Engine\n\nmodel = ...\noptimizer = torch.optim.LBFGS(model.parameters(), lr=1)\ncriterion = \n\ndef update_fn(engine, batch):\n    model.train()\n    x, y = batch\n    # pass to device if needed as here: https:\/\/github.com\/pytorch\/ignite\/blob\/40d815930d7801b21acfecfa21cd2641a5a50249\/ignite\/engine\/__init__.py#L45\n    def closure():\n        y_pred = model(x)\n        loss = criterion(y_pred, y)\n        optimizer.zero_grad()\n        loss.backward()\n        return loss\n\n    optimizer.step(closure)\n\ntrainer = Engine(update_fn)\n\n# everything else is the same\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/github.com\/pytorch\/ignite\/issues\/610#issuecomment-528374297\" rel=\"nofollow noreferrer\">Source<\/a><\/p>\n"}]}
+{"QuestionId":57807050,"AnswerCount":1,"Tags":"<neural-network><classification><conv-neural-network><pytorch><transfer-learning>","CreationDate":"2019-09-05T13:51:22.917","AcceptedAnswerId":57811783.0,"OwnerUserId":3470496.0,"Title":"Can't replace classifier on Densenet121 in pytorch","Body":"<p>I am trying to do some transfer learning using this github DenseNet121 model (<a href=\"https:\/\/github.com\/gaetandi\/cheXpert.git\" rel=\"nofollow noreferrer\">https:\/\/github.com\/gaetandi\/cheXpert.git<\/a>). I'm running into issues resizing the classification layer from 14 to 2 outputs. <\/p>\n\n<p>Relevant part of the github code is:<\/p>\n\n<pre><code>class DenseNet121(nn.Module):\n    \"\"\"Model modified.\n    The architecture of our model is the same as standard DenseNet121\n    except the classifier layer which has an additional sigmoid function.\n    \"\"\"\n    def __init__(self, out_size):\n        super(DenseNet121, self).__init__()\n        self.densenet121 = torchvision.models.densenet121(pretrained=True)\n        num_ftrs = self.densenet121.classifier.in_features\n        self.densenet121.classifier = nn.Sequential(\n            nn.Linear(num_ftrs, out_size),\n            nn.Sigmoid()\n        )\ndef forward(self, x):\n    x = self.densenet121(x)\n    return x\n<\/code><\/pre>\n\n<p>I load and init with:<\/p>\n\n<pre><code># initialize and load the model\nmodel = DenseNet121(nnClassCount).cuda()\nmodel = torch.nn.DataParallel(model).cuda()\nmodeldict = torch.load(\"model_ones_3epoch_densenet.tar\")\nmodel.load_state_dict(modeldict['state_dict'])\n<\/code><\/pre>\n\n<p>It looks like DenseNet doesn't split layers up into children so <code>model = nn.Sequential(*list(modelRes.children())[:-1])<\/code> won't work. <\/p>\n\n<p><code>model.classifier = nn.Linear(1024, 2)<\/code> seems to work on default DenseNets, but with the modified classifier (additional sigmoid function) here it ends up just adding an additional classifier layer without replacing the original.<\/p>\n\n<p>I've tried <\/p>\n\n<pre><code>model.classifier = nn.Sequential(\n    nn.Linear(1024, dset_classes_number), \n    nn.Sigmoid()\n)\n<\/code><\/pre>\n\n<p>But am having the same added instead of replaced classifier issue:<\/p>\n\n<pre><code>...\n      )\n      (classifier): Sequential(\n        (0): Linear(in_features=1024, out_features=14, bias=True)\n        (1): Sigmoid()\n      )\n    )\n  )\n  (classifier): Sequential(\n    (0): Linear(in_features=1024, out_features=2, bias=True)\n    (1): Sigmoid()\n  )\n)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57811783","CreationDate":"2019-09-05T19:18:32.177","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you want to replace the <code>classifier<\/code> inside <code>densenet121<\/code> that is a member of your <code>model<\/code> you need to assign<\/p>\n\n<pre><code>model.densenet121.classifier = nn.Sequential(...)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57807175,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-05T13:59:08.490","AcceptedAnswerId":57907671.0,"OwnerUserId":3584765.0,"Title":"Use of 1-a weight in categorical focal loss","Body":"<p>I am trying to use focal loss in keras\/tensorflow with multiple classes which leads to use <em>Categorical focal loss<\/em> I guess. I have found some implementation <a href=\"https:\/\/github.com\/mkocabas\/focal-loss-keras\/blob\/master\/focal_loss.py\" rel=\"nofollow noreferrer\">here<\/a> and <a href=\"https:\/\/github.com\/umbertogriffo\/focal-loss-keras\/blob\/master\/losses.py\" rel=\"nofollow noreferrer\">there<\/a> or <a href=\"https:\/\/raw.githubusercontent.com\/aldi-dimara\/keras-focal-loss\/master\/focal_loss.py\" rel=\"nofollow noreferrer\">there<\/a>. <\/p>\n\n<p>As far as I get it the parameter <code>a<\/code> in focal loss is mainly used in the <em>Binary focal loss<\/em> case where 2 classes exist and the one get <code>a<\/code> as a weight and the other gets <code>1-a<\/code> as weight. In the case of the Categorical focal loss all implementations I found use only weight <code>a<\/code> in front of each class loss like:<\/p>\n\n<pre><code>    # Calculate weight that consists of  modulating factor and weighting factor\n    weight = alpha * y_true * K.pow((1-y_pred), gamma)\n    # Calculate focal loss\n    loss = weight * cross_entropy\n<\/code><\/pre>\n\n<p>or <\/p>\n\n<pre><code>    # Calculate Cross Entropy\n    cross_entropy = -y_true * K.log(y_pred)\n    # Calculate Focal Loss\n    loss = alpha * K.pow(1 - y_pred, gamma) * cross_entropy\n<\/code><\/pre>\n\n<p>So, my question is why such a weight factor would make any difference in the training process? Surely, loss is generally used to calculate the amount of weight added to (multiplied by the learning rate that is of course) after each iteration. But this just means that each class gets the same coefficient before it's loss part and so no big deal. This would mean that I could adjust the learning rate and have the same exactly effect? What am I missing here?\nWhat's the exact purpose of parameter <code>a<\/code>?<\/p>\n\n<p>I also found this implementation <a href=\"https:\/\/github.com\/fizyr\/keras-retinanet\/blob\/master\/keras_retinanet\/losses.py\" rel=\"nofollow noreferrer\">here<\/a> where they seem to also use <code>(1-a)<\/code> for the negatives examples although they pass the value to <code>keras.backend.binary_crossentropy<\/code> which make it even more complicated I guess.<\/p>\n\n<p>Any idea of which is the correct implementation?<\/p>\n","answers":[{"AnswerId":"57835068","CreationDate":"2019-09-07T15:23:23.343","ParentId":null,"OwnerUserId":"4028648","Title":null,"Body":"<p>I'm as puzzled as you are as to why they multiply the loss by a constant. The code you supplied has <code>alpha<\/code> as a parameter with the <em>default value<\/em> set to <code>0.25<\/code>, but perhaps when you call the function you're supposed to instead supply a tensor (with the same shape as <code>y_pred<\/code>\/<code>y_true<\/code>) with the weights to this parameter? That's the only explanation I can think of.<\/p>\n\n<p>However, I do think you might be able to simply omit those weights in the code altogether, and instead supply your weights to <code>tf.fit()<\/code>'s <code>class_weight<\/code> parameter, and then that'll do the weighting for you. Could you keep me updated on whether that works?<\/p>\n"},{"AnswerId":"57907671","CreationDate":"2019-09-12T13:20:11.993","ParentId":null,"OwnerUserId":"3584765","Title":null,"Body":"<p>TL; DR  <\/p>\n\n<p>They are the same. So, just use the simpler one.<\/p>\n\n<p>Longer version:<\/p>\n\n<p>(calculation in the toy example have been performed in <em>numpy<\/em> I think the same functionality applies to <em>tensorflow<\/em> also)<\/p>\n\n<p>After doing some research over the issue I concluded to the following:<\/p>\n\n<ul>\n<li>The methods that uses the fixed approach for <code>a<\/code> coefficient indeed apply (a kind of) focal loss in the sense that they inverse the confidence score and power to gamma but it's not exactly the focal loss mentioned in the original paper.<\/li>\n<li>The second approach seems more accurate to the original approach though BUT they are essentially the <em>same thing<\/em> (at least as it seems in those implementations).<\/li>\n<\/ul>\n\n<p>I used a toy example to make some experiments about these losses. For example using 8 samples in 2 foreground classes plus the background we get:<\/p>\n\n<pre><code>gamma = 2\nalpha2 = 0.25\npreds = np.array(\n    [[0.3, 0.2, 0.5], [0.1, 0.4, 0.4], [0.1, 0.8, 0.1], \n     [0.7, 0.2, 0.1], [0.75, 0.15, 0.1], [0.9, 0.05, 0.05],\n     [0.72, 0.18, 0.1], [0.8, 0.1, 0.1]])\ngt = np.array([[1, 0, 0], [0, 1, 0], [0, 1, 0], \n               [1, 0, 0], [1, 0, 0], [1, 0, 0], \n               [1, 0, 0], [1, 0, 0]]).astype(float)\n<\/code><\/pre>\n\n<p>And the two focal losses would be:<\/p>\n\n<pre><code>focal_log_loss_v1 = -np.sum(alpha2 * gt * np.power((1 - preds), gamma) * np.log(preds), axis=-1)\n\nalpha_factor = np.ones_like(gt) * alpha2\nalpha_factor = np.where(gt == 1, alpha_factor, 1 - alpha_factor)\nfocal_weight2 = np.where(gt == 1, 1 - preds, preds)\nfocal_weight2 = alpha_factor * np.power(focal_weight2, gamma)\nfocal_log_loss_v2 = -np.sum(focal_weight2 * gt * np.log(preds), axis=-1)\n<\/code><\/pre>\n\n<p>The results show that:<\/p>\n\n<pre><code>focal_log_loss_v1\n<\/code><\/pre>\n\n<blockquote>\n  <p>array([0.14748667, 0.08246617, 0.00223144, 0.00802519, 0.00449503,\n         0.0002634 , 0.00643868, 0.00223144])<\/p>\n<\/blockquote>\n\n<pre><code>focal_log_loss_v3\n<\/code><\/pre>\n\n<blockquote>\n  <p>array([0.14748667, 0.08246617, 0.00223144, 0.00802519, 0.00449503,\n         0.0002634 , 0.00643868, 0.00223144])<\/p>\n<\/blockquote>\n\n<p>the two methods are equivalent (!). That was unexpected (at least from my side).<br>\nAnyway, the explanation for this is in this command here:<\/p>\n\n<pre><code>gt * np.log(preds)\n<\/code><\/pre>\n\n<blockquote>\n  <p>array([[-1.2039728 , -0.        , -0.        ],<br>\n         [-0.        , -0.91629073, -0.        ],<br>\n         [-0.        , -0.22314355, -0.        ],<br>\n         [-0.35667494, -0.        , -0.        ],<br>\n         [-0.28768207, -0.        , -0.        ],<br>\n         [-0.10536052, -0.        , -0.        ],<br>\n         [-0.32850407, -0.        , -0.        ],<br>\n         [-0.22314355, -0.        , -0.        ]])<\/p>\n<\/blockquote>\n\n<p>which basically eliminates all contribution of all non-ground truth samples (meaning all confidence scores not belonging to the ground truth one). So, even if <code>focal_weight2<\/code> contains non-zero values to non-gt samples they are eliminated afterwards.<\/p>\n\n<pre><code>focal_weight2\n<\/code><\/pre>\n\n<blockquote>\n  <p>array([[0.1225  , 0.03    , 0.1875  ],<br>\n         [0.0075  , 0.09    , 0.12    ],<br>\n         [0.0075  , 0.01    , 0.0075  ],<br>\n         [0.0225  , 0.03    , 0.0075  ],<br>\n         [0.015625, 0.016875, 0.0075  ],<br>\n         [0.0025  , 0.001875, 0.001875],<br>\n         [0.0196  , 0.0243  , 0.0075  ],<br>\n         [0.01    , 0.0075  , 0.0075  ]])<\/p>\n<\/blockquote>\n\n<p>That's why this code produces the same loss:<\/p>\n\n<pre><code>alpha_factor = np.ones_like(gt) * alpha2\nalpha_factor = np.where(gt == 1, alpha_factor, 0)\nfocal_weight2 = np.where(gt == 1, 1 - preds, 0)\nfocal_weight2 = alpha_factor * np.power(focal_weight2, gamma)\nfocal_log_loss_v3 = -np.sum(focal_weight2 * gt * np.log(preds), axis=-1)\n\nfocal_log_loss_v3\n<\/code><\/pre>\n\n<blockquote>\n  <p>array([0.14748667, 0.08246617, 0.00223144, 0.00802519, 0.00449503,\n         0.0002634 , 0.00643868, 0.00223144])<\/p>\n<\/blockquote>\n\n<p>I am not sure if there should be any differentiation between background and foreground samples though (this is another topic question).<\/p>\n"}]}
+{"QuestionId":57807184,"AnswerCount":0,"Tags":"<python><tensorflow><imagenet><cleverhans>","CreationDate":"2019-09-05T13:59:46.180","AcceptedAnswerId":null,"OwnerUserId":6476080.0,"Title":"cleverhans - clip_min, clip_max when they vary by channels","Body":"<p><code>cleverhans<\/code> has <code>clip_min<\/code> and <code>clip_max<\/code> to make sure that the perturbed data is still within the valid range. This works fine, e.g. if I'm working with an image where the range of values in the matrix is <code>0-255<\/code> or <code>0-1<\/code>. However, when the range differs on different dimensions, a single value is not enough to constrain the image to be valid. For example, in most ImageNet models, the images are preprocessed by subtracting the mean and dividing by the standard deviation <em>per channel<\/em>, hence making a single value clip not ideal. Is there a smart workaround to this problem?<\/p>\n","answers":[]}
+{"QuestionId":57807362,"AnswerCount":0,"Tags":"<python-3.x><artificial-intelligence><pytorch><lstm>","CreationDate":"2019-09-05T14:09:19.873","AcceptedAnswerId":null,"OwnerUserId":10548439.0,"Title":"Loss and accuracy not changing of LSTM model","Body":"<p><a href=\"https:\/\/i.stack.imgur.com\/4YvoV.png\" rel=\"nofollow noreferrer\">Picture of the train and test loss<\/a>Here\u2019s a look at the LSTM model\n`<\/p>\n\n<pre><code># Create LSTM Model\nclass LSTMModel(nn.Module):\n  def __init__(self, input_dim, hidden_dim, layer_dim, output_dim):\n    super(LSTMModel, self).__init__()\n    # Number of hidden dimensions\n    self.hidden_dim = hidden_dim\n\n    # Number of hidden layers\n    self.layer_dim = layer_dim\n\n    # LSTM\n    self.lstm = nn.LSTM(input_dim, hidden_dim, layer_dim, batch_first=True, dropout=0.1)\n\n    # Readout layer       \n    self.f1 = nn.Linear(hidden_dim, output_dim)\n\n    self.dropout_layer = nn.Dropout(p=0.2)\n    self.softmax = nn.Softmax()\n    def forward(self, x):\n    # Initialize hidden state with zeros\n\n       h0 = Variable(torch.zeros(self.layer_dim, x.size(0), self.hidden_dim).type(torch.FloatTensor))\n\n    # Initialize cell state\n\n       c0 = Variable(torch.zeros(self.layer_dim, x.size(0), self.hidden_dim).type(torch.FloatTensor))\n\n    # One time step\n       out, (hn, cn) = self.lstm(x, (h0,c0))\n       out = self.dropout_layer(hn[-1])\n       out = self.f1(out)\n       out = self.softmax(out)    \n       return out\n\n    #LSTM Configuration\nbatch_size = 3000\nnum_epochs = 20\nlearning_rate = 0.001#Check this learning rate\n\n# Create LSTM\ninput_dim = 1       # input dimension\nhidden_dim = 30     # hidden layer dimension\nlayer_dim = 15  # number of hidden layers\noutput_dim = 1      # output dimension\nnum_layers = 10      #num_layers\n\nprint(\"input_dim = \", input_dim,\"\\nhidden_dim = \", hidden_dim,\"\\nlayer_dim = \", layer_dim,\"\\noutput_dim = \", output_dim)\n\n\nmodel = LSTMModel(input_dim, hidden_dim, layer_dim, output_dim)\nmodel.cuda()\nerror = nn.BCELoss()\noptimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)\n\n\ngraph_index = 0\ntest_loss = []\ntrain_loss = []\nplt_test_index = []\nplt_train_index = []\ntmp_index = []\ntmp_train = []\ntmp_test = []\n\n\n# model.init_hidden()\n\nfor epoch in range(num_epochs):\n\n  # Train\n  model.train()\n  loss_list_train = []\n  loss_list_test = []\n\n  total_train = 0\n  equals_train = 0\n  total_test = 0\n  num0_train = 0\n  num1_train = 0\n  num0_test = 0\n  num1_test = 0\n  equals_test = 0\n  TP_train = 0\n  FP_train = 0\n  TN_train = 0\n  FN_train = 0\n  TP_test = 0\n  FP_test = 0\n  TN_test = 0\n  FN_test = 0\n\n#   for i, (inputs, targets) in enumerate(train_loader):\n  for i, (inputs, targets) in enumerate(train_loader):\n    train  = Variable(inputs.type(torch.FloatTensor).cuda())\n    targets = Variable(targets.type(torch.FloatTensor).cuda())\n    optimizer.zero_grad()\n    outputs = model(train)\n    loss = error(outputs, targets)\n    loss_list_train.append(loss.item())\n    loss.backward()\n    # loss.backward(retain_graph=True)\n    optimizer.step()\n    t = np.where(targets.cpu().detach().numpy() &gt; 0.5, 1, 0)\n    o = np.where(outputs.cpu().detach().numpy() &gt; 0.5, 1, 0) \n    total_train += t.shape[0]\n    equals_train += np.sum(t == o)\n    num0_train += np.sum(t == 0)\n    num1_train += np.sum(t == 1)\n    TP_train += np.sum(np.logical_and(t == 1, o==1))\n    FP_train += np.sum(np.logical_and(t == 1, o==0))\n    TN_train += np.sum(np.logical_and(t == 0, o==0))\n    FN_train += np.sum(np.logical_and(t == 0, o==1))      \n    tb.save_value('Train Loss', 'train_loss', globaliter, loss.item())\n    globaliter += 1\n    tb.flush_line('train_loss')\n    print(i)\n\n  # Test\n  model.eval()\n  targets_plot = np.array([])\n  outputs_plot = np.array([])\n  inputs_plot = np.array([])\n\n\n  for inputs, targets in test_loader:\n\n    inputs = Variable(inputs.type(torch.FloatTensor).cuda())\n    targets = Variable(targets.type(torch.FloatTensor).cuda())  \n    outputs = model(inputs)\n    loss = error(outputs, targets)\n    loss_list_test.append(loss.item())\n    #print(outputs.cpu().detach().numpy())\n    t = np.where(targets.cpu().detach().numpy() &gt; 0.5, 1, 0)\n    o = np.where(outputs.cpu().detach().numpy() &gt; 0.5, 1, 0)\n    total_test += t.shape[0]\n    equals_test += np.sum(t == o)\n    num0_test += np.sum(t == 0)\n    num1_test += np.sum(t == 1)\n    TP_test += np.sum(np.logical_and(t == 1, o==1))\n    FP_test += np.sum(np.logical_and(t == 0, o==1))\n    TN_test += np.sum(np.logical_and(t == 0, o==0))\n    FN_test += np.sum(np.logical_and(t == 1, o==0))\n    tb.save_value('Test Loss', 'test_loss', globaliter2, loss.item())\n    globaliter2 += 1\n    tb.flush_line('test_loss')\n\n\n  # Save value in array    \n  graph_index += 1 \n  plt_train_index.append(graph_index)\n  plt_test_index.append(graph_index)\n  train_loss.append(np.mean(np.array(loss_list_train)))\n  test_loss.append(np.mean(np.array(loss_list_test)))\n\n  print(\"------------------------------\")\n  print(\"Epoch : \", epoch)\n  print(\"----- Train -----\")\n  print(\"Total =\", total_train, \" | Num 0 =\", num0_train, \" | Num 1 =\", num1_train)\n  print(\"Equals =\", equals_train)\n  print(\"Accuracy =\", (equals_train \/ total_train)*100, \"%\")\n  # print(\"TP =\", TP_train \/ total_train, \"% | TN =\", TN_train \/ total_train, \"% | FP =\", FP_train \/ total_train, \"% | FN =\", FN_train \/ total_train, \"%\")\n  print(\"----- Test -----\")\n  print(\"Total =\", total_test, \" | Num 0 =\", num0_test, \" | Num 1 =\", num1_test)\n  print(\"Equals =\", equals_test)\n  print(\"Accuracy =\", (equals_test \/ total_test)*100, \"%\")\n<\/code><\/pre>\n\n<p>I am using the model to do binary classification on the sequence length of 300. The accuracy and the loss are not changing over several epochs.I tried changing the no. of layers,no of hidden states, activation function, but all to no avail. I don\u2019t know what i am doing wrong, I am probably missing something fundamental. Any help is appreciated<\/p>\n","answers":[]}
+{"QuestionId":57807400,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><pygame><reinforcement-learning>","CreationDate":"2019-09-05T14:11:32.863","AcceptedAnswerId":null,"OwnerUserId":10500505.0,"Title":"How to create a reinforcement learning neural network to play flappy bird?","Body":"<p>I created a Flappy Bird clone using PyGame, and I want to teach a TensorFlow reinforcement learning neural network how to play it using the bird's y position, the bird's velocity, the distance to the next closest \"pipe\", and the y positions of the gap between pipes,top_y and bottom_y. So, 5 inputs to produce 1 output - a prediction of whether the bird should jump or not.<\/p>\n\n<p>Sorry for this formatting - StackOverflow thought it was code.\n    Here is my repo:\n    <a href=\"https:\/\/github.com\/joshuadargan\/PenguinoFly.ai\" rel=\"nofollow noreferrer\">https:\/\/github.com\/joshuadargan\/PenguinoFly.ai<\/a>\n    The key files for this problem are Game.py and NeuralNetwork.py:\n    <a href=\"https:\/\/github.com\/joshuadargan\/PenguinoFly.ai\/blob\/master\/NeuralNetwork.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/joshuadargan\/PenguinoFly.ai\/blob\/master\/NeuralNetwork.py<\/a>\n    <a href=\"https:\/\/github.com\/joshuadargan\/PenguinoFly.ai\/blob\/master\/Game.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/joshuadargan\/PenguinoFly.ai\/blob\/master\/Game.py<\/a><\/p>\n\n<p>I tried looking at a lot of different tutorials to figure things out, and I mostly based my neural network code from this:\n<a href=\"https:\/\/towardsdatascience.com\/reinforcement-learning-tutorial-part-3-basic-deep-q-learning-186164c3bf4\" rel=\"nofollow noreferrer\">https:\/\/towardsdatascience.com\/reinforcement-learning-tutorial-part-3-basic-deep-q-learning-186164c3bf4<\/a><\/p>\n\n<p>The tutorial talks about batching data, and that the input and output is two dimensional to allow for batching. I don't think batching the data would be possible. What changes should I make it to be able to train real-time with the 5 aforementioned variables and the 1 output.<\/p>\n\n<p>The tutorial also mentions one_hot, which I didn't think applied to my situation. Should I keep it in there?<\/p>\n\n<p>I have tried running it with the changes for my system (removing one_hot and adjusting the inputs and outputs), but I get this error: <\/p>\n\n<pre><code>ValueError: Cannot feed value of shape (5,) for Tensor 'Placeholder:0', \nwhich has shape '(?, 5)'\n<\/code><\/pre>\n\n<p>It starts from line 165 in Game.py where I call update. Then line 101 in \nNeuralNetwork.py where it calls train, which then calls get_Q on line 82, \nthen the error is produced on line 54 in the get_Q function.<\/p>\n\n<p>I am assuming that there are other errors in it too, but I can't bypass this error. I have tried to change the placeholder to be a rank 1 Tensor, but the hidden layer expects a rank 2 Tensor. I have also tried distributing the data across two arrays to satisfy it, but that didn't work either. <\/p>\n\n<p>I expect to input the 5 aforementioned variables as the game's \"state\" and I expect a number between 0-1 representing the percentage that I should jump or not jump in that state.<\/p>\n\n<p>What should I do to fix this problem, and what design suggestions do you have for my neural network?\nThank you!!<\/p>\n","answers":[]}
+{"QuestionId":57807716,"AnswerCount":1,"Tags":"<python><tensorflow><google-colaboratory>","CreationDate":"2019-09-05T14:29:28.403","AcceptedAnswerId":null,"OwnerUserId":12024899.0,"Title":"How to get weights from tensorflow fully_connected with Google Colab","Body":"<p>I'm trying to extract the weights from a model after training it. Here's a code.\n1- how to save the model in the previously uploaded directory path?\n2. how to get weights?\n3. I have some csv file as the same used for modeling but without class. how can I estimate the classes with this model?<\/p>\n\n<pre><code># Preprocess\nimport pandas as pd\nimport keras\nfrom sklearn.model_selection import train_test_split\n# Prepare data\nchurn = pd.read_csv('Churn_Modelling3.csv')\n# split data into train and test sets\nx_train, x_test, y_train, y_test = train_test_split(churn.iloc[:,0:10],churn.iloc[:, 10], test_size=0.2)                    \n# Create a Model\nimport tensorflow as tf\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import Conv2D\nfrom tensorflow.keras.layers import MaxPool2D\nfrom tensorflow.keras.layers import Flatten\nfrom tensorflow.keras.layers import Dense\nmodel = Sequential()\n# Input layer\nmodel.add(Dense(32, activation=tf.nn.sigmoid,input_dim=x_train.shape[1]))\n# Hidden layers\nmodel.add(Dense(64, activation=tf.nn.sigmoid))\nmodel.add(Dense(32, activation=tf.nn.sigmoid))\n# Output layer\nmodel.add(Dense(1, activation=tf.nn.sigmoid))\n# Compile the Model\nfrom keras.optimizers import SGD\nsgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)\nmodel.compile( optimizer= 'sgd', loss='binary_crossentropy', metrics=['accuracy'])\n# Fit the model\nhistory = model.fit(x_train, y_train, epochs=20, batch_size=128,validation_data=(x_test, y_test))\n# \u064dEvaluate the Model\nscore = model.evaluate(x_test, y_test, batch_size=128)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58055990","CreationDate":"2019-09-23T05:12:21.657","ParentId":null,"OwnerUserId":"11530462","Title":null,"Body":"<ol>\n<li><p>How to save the model in the previously uploaded directory path? => You can use any of the below commands, <\/p>\n\n<pre><code> `model.save_weights('.\/Weights\/Weights')` or \n `tf.keras.experimental.export_saved_model` or \n  `model.save`\n<\/code><\/pre>\n\n<ol start=\"2\">\n<li>How to get weights? => We can Load the Weights using any of the below commands:<\/li>\n<\/ol><\/li>\n<\/ol>\n\n<blockquote>\n<pre><code>model.load_weights('my_model_weights.h5') or         \nkeras.models.load_model('my_model.h5') or\n\ntf.keras.experimental.load_from_saved_model\n<\/code><\/pre>\n<\/blockquote>\n\n<ol start=\"3\">\n<li>I have some csv file as the same used for modeling but without class. how can I estimate the classes with this model? => One of the ways to do it is to iterate the below code for each instance of CSV data Label.<\/li>\n<\/ol>\n\n<blockquote>\n<pre><code>ynew = model.predict(Xnew)\n<\/code><\/pre>\n<\/blockquote>\n\n<p>Fore more information, you can refer to the below links:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/tutorials\/keras\/save_and_restore_models\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/keras\/save_and_restore_models<\/a><\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/saved_model<\/a><\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/Model\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/Model<\/a><\/p>\n\n<p>Please let me know if this is what you are looking for. Thanks!<\/p>\n"}]}
+{"QuestionId":57807728,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><conv-neural-network>","CreationDate":"2019-09-05T14:29:59.250","AcceptedAnswerId":null,"OwnerUserId":10736997.0,"Title":"How do I determine batch size from \"TFRecord shards?\"","Body":"<p>My preprocessing step involved converting 300k images into 1024 \"TFRecord shards.\" I am not sure how shards are different than TFRecords, but they are about 10 MB each in size. I have another step that takes 70k images and converts them each into a TFRecord. They all have a .tfrecord extension, where the shards do not.<\/p>\n\n<p>So if I have a batch size of 32 I am not sure what that means in the context of shards. Is it 32 shards which would be 320 MB per batch? Or 32 images and the size depends on the size of the original image file?<\/p>\n\n<p>Follow up question, I have a gpu with 16gb ram, it would seem I could fit the entire training data into one batch, but I am not sure if I am missing something here.<\/p>\n","answers":[]}
+{"QuestionId":57807816,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-05T14:35:25.853","AcceptedAnswerId":null,"OwnerUserId":11855471.0,"Title":"Unable to use 'pad_sequences' in multi columns","Body":"<p>For instance, I have a 3 x 3 data frame, and all the values are texts.\nAfter I used 'texts_to_sequences'\nI try to use 'pad_sequences', but it can't.<\/p>\n\n<pre><code>cols = df_train.columns[2:]\nall_cols = pd.Series(df_train[cols].values.ravel('F'))\n\nembed_size = 300 \nmaxlen = 150 \nembed_features = 50000 \ntokenizer = Tokenizer(num_words=embed_features)\ntokenizer.fit_on_texts(list(all_cols))\n\nfor col in cols: \n    df_train[col] = tokenizer.texts_to_sequences(df_train[col]) \n    df_test[col] = tokenizer.texts_to_sequences(df_test[col])\n<\/code><\/pre>\n\n<p>After the step above, I used 'pad_sequences'<\/p>\n\n<pre><code>df_train = pad_sequences(df_train, maxlen=maxlen)\n<\/code><\/pre>\n\n<p>I expected that it would show the values in the array with [0, \n 0, 0, 0, ..., 321, 421, 3]<\/p>\n\n<p>However, it shows [0]<\/p>\n","answers":[]}
+{"QuestionId":57808156,"AnswerCount":2,"Tags":"<python><pandas><tensorflow><machine-learning><keras>","CreationDate":"2019-09-05T14:56:27.857","AcceptedAnswerId":58183602.0,"OwnerUserId":3183808.0,"Title":"Attribute Error when calling 'predict_generator' from loaded Keras model","Body":"<p>I have trained a dual-input neural network for categorising images and saved the weights to an hdf5 file.<\/p>\n\n<p>I am now trying to load this network and use 'predict_generator' to see how well it does on my testing set. However, upon calling 'predict_generator' I get the following error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n   File \"load_cnn.py\", line 75, in &lt;module&gt;\n     pred = loaded_model.predict_generator(test_gen, steps=36, verbose=1)\n   File \n\"\/home\/ppxjm4\/anaconda3\/envs\/ML\/lib\/python3.7\/site-packages\/keras\/legacy\/interfaces.py\", \nline 91, in wrapper\n     return func(*args, **kwargs)\n   File \n\"\/home\/ppxjm4\/anaconda3\/envs\/ML\/lib\/python3.7\/site-packages\/keras\/engine\/training.py\", \nline 1772, in predict_generator\n     verbose=verbose)\n   File \n\"\/home\/ppxjm4\/anaconda3\/envs\/ML\/lib\/python3.7\/site-packages\/keras\/engine\/training_generator.py\", \nline 503, in predict_generator\n     batch_size = x[0].shape[0]\nAttributeError: 'list' object has no attribute 'shape'\n<\/code><\/pre>\n\n<p>I am confused about why it complains here, when during training and calling 'fit_generator()' seemed to work just fine.<\/p>\n\n<p>Here is my code that generates the error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import os\nimport pandas as pd\nimport numpy as np\n\nimport keras\nfrom keras import optimizers\nfrom keras.utils import to_categorical\nfrom keras.layers import Input, Dense, Flatten, Dropout, Conv2D, MaxPooling2D, GlobalAveragePooling2D, Activation, concatenate\nfrom keras.layers.advanced_activations import LeakyReLU\nfrom keras.layers.normalization import BatchNormalization\nfrom keras.models import Model\nfrom keras.callbacks import ModelCheckpoint, ReduceLROnPlateau\n\nfrom keras_preprocessing.image import ImageDataGenerator\n\ntrain_df = pd.read_pickle(\"collated_data.pkl\")\ntrain_df = train_df.sample(frac=1)\n\ndata_gen = ImageDataGenerator(rescale=1.\/255)\n\nfrom keras.models import model_from_json\n\njson_file = open('model.json', 'r')\nloaded_model_json = json_file.read()\njson_file.close()\n\nloaded_model = model_from_json(loaded_model_json)\nloaded_model.load_weights(\"weights.best.hdf5\")\n\nprint(\"loaded model from disk!\")\n\nloaded_model.compile(loss='categorical_crossentropy', optimizer='Adadelta', metrics=['accuracy'])\n\ndef generator_multiple(generator, dataframe, mode, batch_size, img_height, img_width):\n\n    if mode == 'train':\n\n        genX1 = generator.flow_from_dataframe(\n            dataframe=dataframe[:79000], x_col='reconstruction', y_col='label', class_mode='categorical',\n            batch_size=batch_size, shuffle=False, target_size=(img_height, img_width), color_mode='grayscale')\n\n        genX2 = generator.flow_from_dataframe(\n            dataframe=dataframe[:79000], x_col='observation', y_col='label', class_mode='categorical',\n            batch_size=batch_size, shuffle=False, target_size=(img_height, img_width), color_mode='grayscale')\n\n    if mode == 'validate':\n\n        genX1 = generator.flow_from_dataframe(\n            dataframe=dataframe[79000:85000], x_col='reconstruction', y_col='label', class_mode='categorical',\n            batch_size=batch_size, shuffle=False, target_size=(img_height, img_width), color_mode='grayscale')\n\n        genX2 = generator.flow_from_dataframe(\n            dataframe=dataframe[79000:85000], x_col='observation', y_col='label', class_mode='categorical',\n            batch_size=batch_size, shuffle=False, target_size=(img_height, img_width), color_mode='grayscale')\n\n    if mode == 'test':\n\n        genX1 = generator.flow_from_dataframe(\n            dataframe=dataframe[85000:], x_col='reconstruction', y_col=None, class_mode=None,\n            batch_size=batch_size, shuffle=False, target_size=(img_height, img_width), color_mode='grayscale')\n\n        genX2 = generator.flow_from_dataframe(\n            dataframe=dataframe[85000:], x_col='observation', y_col=None, class_mode=None,\n            batch_size=batch_size, shuffle=False, target_size=(img_height, img_width), color_mode='grayscale')\n\n    while True:\n        X1i = genX1.next()\n        X2i = genX2.next()\n        yield [[X1i[0], X2i[0]], X2i[1]]\n\n\ntest_gen = generator_multiple(data_gen, dataframe=train_df, mode='test', batch_size=36,\n                               img_height=100, img_width=100)\n\npred = loaded_model.predict_generator(test_gen, steps=36, verbose=1)\n\npredicted_class_indices = np.argmax(pred, axis=1)\nlabels = (train_gen.class_indices)\nlabels = dict((v,k) for k,v in labels.items())\npredictions = [labels[k] for k in predicted_class_indices]\n\nfilenames=test_gen.filenames\nresults=pd.DataFrame({\"Filename\":filenames,\n                      \"Predictions\":predictions})\nresults.to_csv(\"results.csv\",index=False)\n\n<\/code><\/pre>\n","answers":[{"AnswerId":"57821187","CreationDate":"2019-09-06T11:31:29.993","ParentId":null,"OwnerUserId":"10190971","Title":null,"Body":"<p>If you have more then 1 input to your model, i recommend you try yielding a dict from your generator rather then a list, here is an example from keras it self(doc string of <code>fit_generator()<\/code>):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def generate_arrays_from_file(path):\n    while True:\n        with open(path) as f:\n            for line in f:\n                # create numpy arrays of input data\n                # and labels, from each line in the file\n                x1, x2, y = process_line(line)\n                yield ({'input_1': x1, 'input_2': x2}, {'output': y})\n\nmodel.fit_generator(generate_arrays_from_file('\/my_file.txt'),\n                    steps_per_epoch=10000, epochs=10)\n<\/code><\/pre>\n\n<p>For more info, here is <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/engine\/training.py#L1510\" rel=\"nofollow noreferrer\">keras's github<\/a>.<\/p>\n"},{"AnswerId":"58183602","CreationDate":"2019-10-01T11:10:37.467","ParentId":null,"OwnerUserId":"3183808","Title":null,"Body":"<p>The solution to my problem was to modify the output of my generator when using it to feed testing data to the model. The flow_from_dataframe method produces data of different dimensions when y_col and class_mode are set to None.<\/p>\n\n<pre><code>def generator_multiple(generator, dataframe, batch_size, img_height, img_width):\n\n    genX1 = generator.flow_from_dataframe(\n        dataframe=dataframe, x_col=\"reconstruction\", y_col=None, class_mode=None,\n        batch_size=batch_size, shuffle=False, target_size=(img_height, img_width), color_mode='grayscale')\n\n    genX2 = generator.flow_from_dataframe(\n        dataframe=dataframe, x_col=\"observation\", y_col=None, class_mode=None,\n        batch_size=batch_size, shuffle=False, target_size=(img_height, img_width), color_mode='grayscale')\n\n    while True:\n        X1i = genX1.next()\n        X2i = genX2.next()\n        yield [X1i, X2i]\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":57808266,"AnswerCount":2,"Tags":"<python><tensorflow><keras><tensorflow2.0>","CreationDate":"2019-09-05T15:02:57.247","AcceptedAnswerId":57808454.0,"OwnerUserId":9280994.0,"Title":"Tensorflow2.0 Keras: Is dropout disabled during testing by default?","Body":"<p>I am wondering if in the following model, dropout will be disabled when I call <code>model.evaluate(...).<\/code><\/p>\n\n<pre><code>layers = [tf.keras.layers.Dense(size, activation='relu')\n                  for size in (20, 40, 20)]\nlayers.insert(1, tf.keras.layers.Dropout(0.2))\nlayers.append(tf.keras.layers.Dense(1, activation=\"sigmoid\"))\nmodel = tf.keras.models.Sequential(layers)\nmodel.compile(\n    optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001),\n    loss=tf.keras.losses.BinaryCrossentropy())\nmodel.fit(...)\nmodel.evaluate(...) #==&gt; will dropout be deactivated here?\n<\/code><\/pre>\n","answers":[{"AnswerId":"57808454","CreationDate":"2019-09-05T15:13:24.407","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>Yes, dropout is always disabled at inference (<code>evaluate<\/code>\/<code>predict<\/code>).<\/p>\n"},{"AnswerId":"57808705","CreationDate":"2019-09-05T15:29:09.910","ParentId":null,"OwnerUserId":"8942966","Title":null,"Body":"<p>As per the documentation:<\/p>\n\n<pre><code>Call arguments:\n    inputs: Input tensor (of any rank).\n    training: Python boolean indicating whether the layer should behave in\n      training mode (adding dropout) or in inference mode (doing nothing).\n  \"\"\"\n<\/code><\/pre>\n\n<p>So yes, dropout is disabled when testing, which is logically correct.\nThe same holds for <em>SpatialDropout<\/em> as well.<\/p>\n\n<p>Please find the link to the documentation below.\n<a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/layers\/Dropout\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/layers\/Dropout<\/a><\/p>\n\n<p>Tensorflow Keras API uses a learning phase flag to identify whether we are training or testing. The learning phase flag is a bool tensor (0 = test, 1 = train) to be passed as input to any Keras function that uses a different behavior at train time and test time.<\/p>\n"}]}
+{"QuestionId":57808624,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-05T15:23:54.467","AcceptedAnswerId":null,"OwnerUserId":12024545.0,"Title":"why am i getting this error 'reshape_and_normalize_image' is not defined'?","Body":"<p>I am trying to reshape and normalize a content_image using vgg-19 model<\/p>\n\n<pre><code>content_image = scipy.misc.imread(\"girl.jpg\")\ncontent_image = reshape_and_normalize_image(content_image)\n<\/code><\/pre>\n\n<p>and I also get the same error when performing the same action on a style_image<\/p>\n\n<pre><code>style_image = scipy.misc.imread(\"santorini.jpg\")\nstyle_image = reshape_and_normalize_image(style_image)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57808632,"AnswerCount":0,"Tags":"<tensorflow><object-detection><training-data><faster-rcnn><labelimg>","CreationDate":"2019-09-05T15:24:17.650","AcceptedAnswerId":null,"OwnerUserId":11164405.0,"Title":"Error in training my faster_rcnn_inception_v2_coco model","Body":"<p>Training an Object Detection Classifier for Multiple Objects Using TensorFlow (CPU) on Windows 8\nissuing following command gets me to this error:<\/p>\n\n<p>python train.py --logtostderr --train_dir=training\/ --pipeline_config_path=training\/faster_rcnn_inception_v2_pets.config<\/p>\n\n<pre><code>W0905 19:31:26.166561  7036 variables_helper.py:157] Variable [SecondStageFeatur\neExtractor\/InceptionV2\/Mixed_5c\/Branch_3\/Conv2d_0b_1x1\/BatchNorm\/gamma\/Momentum]\n is not available in checkpoint\nW0905 19:31:26.166561  7036 variables_helper.py:157] Variable [SecondStageFeatur\neExtractor\/InceptionV2\/Mixed_5c\/Branch_3\/Conv2d_0b_1x1\/weights\/Momentum] is not\navailable in checkpoint\nW0905 19:31:26.166561  7036 variables_helper.py:157] Variable [global_step] is n\not available in checkpoint\nW0905 19:31:28.854089  7036 deprecation.py:323] From C:\\Users\\AISHA\\Anaconda3\\li\nb\\site-packages\\tensorflow\\contrib\\slim\\python\\slim\\learning.py:742: Supervisor.\n__init__ (from tensorflow.python.training.supervisor) is deprecated and will be\nremoved in a future version.\nInstructions for updating:\nPlease switch to tf.train.MonitoredTrainingSession\n2019-09-05 19:31:52.299150: I tensorflow\/core\/platform\/cpu_feature_guard.cc:145]\n This TensorFlow binary is optimized with Intel(R) MKL-DNN to use the following\nCPU instructions in performance critical operations:  AVX AVX2\nTo enable them in non-MKL-DNN operations, rebuild TensorFlow with the appropriat\ne compiler flags.\n2019-09-05 19:31:53.046831: I tensorflow\/core\/common_runtime\/process_util.cc:115\n] Creating new thread pool with default inter op setting: 4. Tune using inter_op\n_parallelism_threads for best performance.\nW0905 19:31:53.057435  7036 deprecation.py:323] From C:\\Users\\AISHA\\Anaconda3\\li\nb\\site-packages\\tensorflow\\python\\training\\saver.py:1276: checkpoint_exists (fro\nm tensorflow.python.training.checkpoint_management) is deprecated and will be re\nmoved in a future version.\nInstructions for updating:\nUse standard file APIs to check for files with this prefix.\nI0905 19:31:53.073088  7036 saver.py:1280] Restoring parameters from training\/mo\ndel.ckpt-0\nW0905 19:32:09.260713  7036 deprecation.py:323] From C:\\Users\\AISHA\\Anaconda3\\li\nb\\site-packages\\tensorflow\\python\\training\\saver.py:1066: get_checkpoint_mtimes\n(from tensorflow.python.training.checkpoint_management) is deprecated and will b\ne removed in a future version.\n**Instructions for updating:\nUse standard file utilities to get mtimes.\nI0905 19:32:09.260713  7036 session_manager.py:500] Running local_init_op.\nI0905 19:32:12.729500  7036 session_manager.py:502] Done running local_init_op.\nI0905 19:33:21.745800  7036 learning.py:754] Starting Session.\nI0905 19:33:22.308308  2844 supervisor.py:1117] Saving checkpoint to path traini\nng\/model.ckpt\nI0905 19:33:22.323910  7036 learning.py:768] Starting Queues.\nI0905 19:33:35.167791  5528 supervisor.py:1099] global_step\/sec: 0\nI0905 19:36:05.575444  5528 supervisor.py:1099] global_step\/sec: 0\nFatal Python error: Aborted**\n\nThread 0x000015c8 (most recent call first):\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 296 in wait\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 552 in wait\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\co\nordinator.py\", line 311 in wait_for_stop\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\qu\neue_runner_impl.py\", line 293 in _close_on_stop\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 865 in run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 917 in _bootstrap_inner\n\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 885 in _bootstrap\n\nThread 0x00001b4c (most recent call first):\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1429 in _call_tf_sessionrun\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1279 in _single_operation_run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\qu\neue_runner_impl.py\", line 257 in _run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 865 in run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 917 in _bootstrap_inner\n\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 885 in _bootstrap\n\nThread 0x00000e28 (most recent call first):\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 296 in wait\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 552 in wait\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\co\nordinator.py\", line 311 in wait_for_stop\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\qu\neue_runner_impl.py\", line 293 in _close_on_stop\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 865 in run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 917 in _bootstrap_inner\n\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 885 in _bootstrap\n\nThread 0x00000c74 (most recent call first):\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1429 in _call_tf_sessionrun\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1279 in _single_operation_run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\qu\neue_runner_impl.py\", line 257 in _run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 865 in run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 917 in _bootstrap_inner\n\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 885 in _bootstrap\n\nThread 0x000015e8 (most recent call first):\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1429 in _call_tf_sessionrun\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1279 in _single_operation_run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\qu\neue_runner_impl.py\", line 257 in _run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 865 in run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 917 in _bootstrap_inner\n\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 885 in _bootstrap\n\nThread 0x00000038 (most recent call first):\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1429 in _call_tf_sessionrun\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1279 in _single_operation_run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\qu\neue_runner_impl.py\", line 257 in _run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 865 in run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 917 in _bootstrap_inner\n\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 885 in _bootstrap\n\nThread 0x00001ba8 (most recent call first):\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1429 in _call_tf_sessionrun\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1279 in _single_operation_run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\qu\neue_runner_impl.py\", line 257 in _run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 865 in run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 917 in _bootstrap_inner\n\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 885 in _bootstrap\n\nThread 0x000017a0 (most recent call first):\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1429 in _call_tf_sessionrun\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1279 in _single_operation_run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\qu\neue_runner_impl.py\", line 257 in _run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 865 in run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 917 in _bootstrap_inner\n\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 885 in _bootstrap\n\nThread 0x00001ab8 (most recent call first):\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1429 in _call_tf_sessionrun\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1279 in _single_operation_run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\qu\neue_runner_impl.py\", line 257 in _run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 865 in run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 917 in _bootstrap_inner\n\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 885 in _bootstrap\n\nThread 0x00000af4 (most recent call first):\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1429 in _call_tf_sessionrun\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1279 in _single_operation_run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\qu\neue_runner_impl.py\", line 257 in _run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 865 in run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 917 in _bootstrap_inner\n\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 885 in _bootstrap\n\nThread 0x00001038 (most recent call first):\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1429 in _call_tf_sessionrun\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1279 in _single_operation_run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\qu\neue_runner_impl.py\", line 257 in _run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 865 in run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 917 in _bootstrap_inner\n\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 885 in _bootstrap\n\nThread 0x00000b1c (most recent call first):\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 300 in wait\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 552 in wait\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\co\nordinator.py\", line 311 in wait_for_stop\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\co\nordinator.py\", line 493 in run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 917 in _bootstrap_inner\n\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 885 in _bootstrap\n\nThread 0x00001598 (most recent call first):\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1429 in _call_tf_sessionrun\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1341 in _run_fn\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1356 in _do_call\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1350 in _do_run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1173 in _run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 950 in run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\tr\naining_util.py\", line 68 in global_step\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\su\npervisor.py\", line 1081 in run_loop\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\co\nordinator.py\", line 495 in run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 917 in _bootstrap_inner\n\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 885 in _bootstrap\n\nThread 0x000018ac (most recent call first):\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1429 in _call_tf_sessionrun\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1341 in _run_fn\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1356 in _do_call\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1350 in _do_run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1173 in _run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 950 in run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\su\npervisor.py\", line 1045 in run_loop\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\training\\co\nordinator.py\", line 495 in run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 917 in _bootstrap_inner\n\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 885 in _bootstrap\n\nThread 0x000007d0 (most recent call first):\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 296 in wait\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\queue.py\", line 170 in get\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\summary\\wri\nter\\event_file_writer.py\", line 159 in run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 917 in _bootstrap_inner\n\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\threading.py\", line 885 in _bootstrap\n\nThread 0x00001b7c (most recent call first):\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1429 in _call_tf_sessionrun\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1341 in _run_fn\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1356 in _do_call\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1350 in _do_run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 1173 in _run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\client\\sess\nion.py\", line 950 in run\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\contrib\\slim\\pytho\nn\\slim\\learning.py\", line 490 in train_step\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\contrib\\slim\\pytho\nn\\slim\\learning.py\", line 775 in train\n  File \"C:\\Users\\AISHA\\Detection\\models\\research\\object_detection\\legacy\\trainer\n.py\", line 416 in train\n  File \"train.py\", line 180 in main\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\util\\deprec\nation.py\", line 324 in new_func\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\absl\\app.py\", line 251 in _ru\nn_main\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\absl\\app.py\", line 300 in run\n\n  File \"C:\\Users\\AISHA\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\platform\\ap\np.py\", line 40 in run\n  File \"train.py\", line 184 in &lt;module&gt;\n<\/code><\/pre>\n\n<p>please help me with this?<\/p>\n\n<p>Performed following changes to the faster_rcnn_inception_v2_pets.config file. Note: The paths must be entered with single forward slashes (NOT backslashes), or TensorFlow will give a file path error when trying to train the model! Also, the paths must be in double quotation marks ( \" ), not single quotation marks ( ' ).<\/p>\n\n<pre><code>Line 9. Change num_classes to the number of different objects you want the classifier to detect. For the above basketball, shirt, and shoe detector, it would be num_classes : 3 .\n\nLine 106. Change fine_tune_checkpoint to:\n    fine_tune_checkpoint : \"C:\/tensorflow1\/models\/research\/object_detection\/faster_rcnn_inception_v2_coco_2018_01_28\/model.ckpt\"\n\nLines 123 and 125. In the train_input_reader section, change input_path and label_map_path to:\n    input_path : \"C:\/tensorflow1\/models\/research\/object_detection\/train.record\"\n    label_map_path: \"C:\/tensorflow1\/models\/research\/object_detection\/training\/labelmap.pbtxt\"\n\nLine 130. Change num_examples to the number of images you have in the \\images\\test directory.\n\nLines 135 and 137. In the eval_input_reader section, change input_path and label_map_path to:\n    input_path : \"C:\/tensorflow1\/models\/research\/object_detection\/test.record\"\n    label_map_path: \"C:\/tensorflow1\/models\/research\/object_detection\/training\/labelmap.pbtxt\"\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57808828,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-05T15:36:19.700","AcceptedAnswerId":57810527.0,"OwnerUserId":866082.0,"Title":"Is it possible to map the dataset's batch size between layers?","Body":"<p>Consider this:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.keras.layers import Dense, LSTM\n\nmodel = tf.keras.models.Sequential([\n    Dense(10, batch_input_shape=(32, None, 100)),\n    LSTM(1, stateful=True)\n])\nmodel.summary()\n\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ndense (Dense)                (32, None, 10)            1010      \n_________________________________________________________________\nlstm  (LSTM)                 (32, 1)                   48        \n=================================================================\nTotal params: 1,058\nTrainable params: 1,058\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>Regardless of whether such a model makes sense or not, the batch size on the first layer (the Dense layer) is set only because the LSTM has <code>stateful=True<\/code> and it needs the batch size. And the way to provide the batch size to it is through the first layer. That's why the Dense layer specifies the batch size.<\/p>\n\n<p>I was wondering if there's a way to make this work:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.keras.layers import Dense, LSTM\n\nmodel = tf.keras.models.Sequential([\n    Dense(10, batch_input_shape=(None, 32, 100)),\n    #Going from (None, 32, 10) to (32, None, 10)\n    LSTM(1, stateful=True)\n])\n<\/code><\/pre>\n\n<p>I know that this is possible before I start the model using the Dataset class methods (map, window, batch). But I was wondering if there's a way to do this between layers?<\/p>\n","answers":[{"AnswerId":"57810527","CreationDate":"2019-09-05T17:35:51.040","ParentId":null,"OwnerUserId":"866082","Title":null,"Body":"<p>Apparently, you can do that using a Lambda layer:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.keras.layers import Dense, LSTM\n\nmodel = tf.keras.models.Sequential([\n    Dense(10, batch_input_shape=(None, 32, 100)),\n    tf.keras.layers.Lambda(lambda x: tf.reshape(x, (32, -1, 10))),\n    LSTM(1, stateful=True)\n])\nmodel.summary()\n\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ndense (Dense)                (None, 32, 10)            1010      \n_________________________________________________________________\nlambda (Lambda)              (32, None, 10)            0         \n_________________________________________________________________\nlstm   (LSTM)                (32, 1)                   48        \n=================================================================\nTotal params: 1,058\nTrainable params: 1,058\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>Who knew!?<\/p>\n"}]}
+{"QuestionId":57808870,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-09-05T15:38:41.113","AcceptedAnswerId":57809895.0,"OwnerUserId":3889510.0,"Title":"Optimize Custom Parameters in Pytorch","Body":"<p>I'm trying to create some custom parameters to optimize and came across this helpful link <a href=\"https:\/\/discuss.pytorch.org\/t\/defining-weights-of-a-custom-layer-as-parameters\/17687\" rel=\"nofollow noreferrer\">here<\/a>. However, I'm a bit confused as to why this code works.<\/p>\n\n<p>Here is the code with some slight modifications from the original post to clearly see the behavior of the optimization.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport numpy as np\n\nclass Mask(nn.Module):\n    def __init__(self):\n        super(Mask, self).__init__()\n        self.weight = torch.nn.Parameter(data=torch.Tensor(5, 5), requires_grad=True)\n\n        self.weight.data.uniform_(-1, 1)\n\n    def forward(self, x):\n        masked_wt = self.weight.mul(1)\n        return masked_wt\n\n\nclass Model(nn.Module):\n    def __init__(self):\n        super(Model, self).__init__()\n        self.mask = Mask()\n\n    def forward(self,x):\n        x = self.mask(x)\n        return x\n\nmodel = Model()\nindata = torch.ones(5,5)\noptimizer = torch.optim.Adam(model.parameters(),lr=0.001)\n\nwhile True:\n    x = model(indata)\n    optimizer.zero_grad()\n    loss = F.l1_loss(indata,x)\n    loss.backward()\n    optimizer.step()\n\n    print(model.mask.weight)\n<\/code><\/pre>\n\n<p>So two questions:<\/p>\n\n<ol>\n<li><p>In the forward function of the Mask class, why do I need to do self.weight.mul(1) and why doesn't x get used?<\/p><\/li>\n<li><p>x = model(indata) clearly does per element multiplication between the 5x5 matrix indata and the weight, but how does this happen if we didn't use x in the forward function of Mask?<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"57809895","CreationDate":"2019-09-05T16:48:20.620","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You are right, <code>Mask.forward<\/code> discard <code>x<\/code> completely. However, the output of your model \"sees\" <code>indata<\/code> when computing the loss.<br>\nWhat you are actually teaching your model is to have <code>mask.weight == indata<\/code>.<\/p>\n"}]}
+{"QuestionId":57808885,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-09-05T15:40:01.690","AcceptedAnswerId":null,"OwnerUserId":6676959.0,"Title":"Force a model to return a single output as a list of one element","Body":"<p>According to <a href=\"https:\/\/keras.io\/models\/model\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/models\/model\/<\/a>, when a model returns a multiple output, it is represented as a list of outputs.<\/p>\n\n<p>It is possible to force the model to return a list even with a single output? By default when building the model, if there is only one output, it is returned<\/p>\n\n<p>I am doing a generic utility and this inconsistency in the keras API (returning two output types) does not seem very elegant. I don't know what will the input be so I always assume multiple inputs\/outputs.<\/p>\n\n<p>Is it possible to return a list of one output?\nHow can I overcome this limitation?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nimport tensorflow as tf\nimport numpy as np\n\n\ndef get_model1():\n    in_layer = tf.keras.Input(\n        name='IN',\n        shape=(10), \n        dtype=tf.float32\n    )\n    out_layer = tf.keras.layers.Dense(\n        name=\"OUT\",\n        units=1,\n        activation=tf.keras.activations.sigmoid \n    )(in_layer)    \n    return tf.keras.Model(\n        name='Model_1',\n        inputs=[in_layer],\n        outputs=[out_layer]\n    )\n\ndef get_model2():\n    in_layer1 = tf.keras.Input(\n        name='IN1',\n        shape=(10),\n        dtype=tf.float32\n    )\n    in_layer2 = tf.keras.Input(\n        name='IN2',\n        shape=(10),\n        dtype=tf.float32\n    )\n    out_layer1 = tf.keras.layers.Dense(\n        name=\"OUT1\",\n        units=1,\n        activation=tf.keras.activations.sigmoid\n    )(in_layer1)    \n    out_layer2 = tf.keras.layers.Dense(\n        name=\"OUT2\",\n        units=1,\n        activation=tf.keras.activations.sigmoid\n    )(in_layer2) \n    return tf.keras.Model(\n        name='Model_1',\n        inputs=[in_layer1,in_layer2],\n        outputs=[out_layer1, out_layer2]\n    )\n\nm1 = get_model1()\nm2 = get_model2()\n\nprint(m1.input_shape)\nprint(m2.input_shape)\n\nprint(m1.output_shape)\nprint(m2.output_shape)\n\n<\/code><\/pre>\n\n<p>This is the result.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>(None, 10)\n[(None, 10), (None, 10)]\n(None, 1)\n[(None, 1), (None, 1)]\n<\/code><\/pre>\n\n<p>What I expect is the first model to have a list as an input and as an output, the same behavior as the second model but with one element.<\/p>\n","answers":[{"AnswerId":"57867678","CreationDate":"2019-09-10T09:19:15.827","ParentId":null,"OwnerUserId":"6676959","Title":null,"Body":"<p>At the end I have used the keras internal function <code>to_list<\/code><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.python.keras.utils.generic_utils import to_list\n\nto_list(model.predict(x))\n<\/code><\/pre>\n\n<p>This way I unify the expected output of my model by returning always a list and delegating the checks in the library that generates the inconsistency.<\/p>\n"}]}
+{"QuestionId":57808994,"AnswerCount":0,"Tags":"<python><keras><deep-learning><bots><reinforcement-learning>","CreationDate":"2019-09-05T15:47:26.600","AcceptedAnswerId":null,"OwnerUserId":12025987.0,"Title":"Keras high loss and high accuracy in gk bot with reinforcement learning?","Body":"<p>I'm making goal-keeper bot in haxball game. It worked well when i trained less but i worked worse when i trained more.<\/p>\n\n<p>Last reinforcement state: <strong>5160 episode - 4171281 steps - 0.05 epsilon<\/strong>:<\/p>\n\n<p>Last fit result= <strong>acc: 0.9905, loss: 293408940460887.5000<\/strong> <\/p>\n\n<p>(previous fit result like: <strong>acc: 0.7, loss: 0.012<\/strong>)<\/p>\n\n<p>Last game result image: <a href=\"https:\/\/i.ibb.co\/NmLL5b7\/hax2.gif\" rel=\"nofollow noreferrer\">https:\/\/i.ibb.co\/NmLL5b7\/hax2.gif<\/a><\/p>\n\n<p>This Keras model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>        NBACTIONS = 3\n        WINDOW_WIDTH = 630\n        WINDOW_HEIGHT = 400\n        IMGHEIGHT = int(WINDOW_HEIGHT\/5)\n        IMGWIDTH = int(WINDOW_WIDTH\/5)\n        IMGHISTORY = 4\n\n        model = Sequential()   \n        model.add(Conv2D(32, kernel_size=4, strides = (2,2), input_shape = (DCQL_Global.IMGHEIGHT,DCQL_Global.IMGWIDTH,DCQL_Global.IMGHISTORY),padding = \"same\"))\n        model.add(Activation(\"relu\"))\n        model.add(MaxPool2D(pool_size=(2, 2)))\n        model.add(Conv2D(64,kernel_size=4,strides=(2,2),padding=\"same\"))\n        model.add(Activation(\"relu\"))\n        model.add(MaxPool2D(pool_size=(2, 2)))\n        model.add(BatchNormalization())\n        model.add(Conv2D(64,kernel_size=3,strides=(1,1),padding=\"same\"))\n        model.add(Activation(\"relu\"))\n        model.add(MaxPool2D(pool_size=(2, 2)))\n        model.add(Flatten())\n        model.add(Dense(512))\n        model.add(Activation(\"relu\"))\n        model.add(Dense(units= DCQL_Global.NBACTIONS, activation=\"linear\"))\n        model.compile(loss = \"mse\", optimizer=Adam(lr=0.0001), metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>Reward rules= goal: <strong>-1<\/strong>, save(not missing goal): <strong>+1<\/strong>, after kicking the ball if gk follow the ball correctly <strong>+0.01<\/strong> else gk follow the ball incorrectly <strong>-0.01<\/strong> reward.<\/p>\n\n<p>I'm preprocessing the image like the following link:<\/p>\n\n<p><a href=\"https:\/\/i.ibb.co\/yPdrJKJ\/hax.gif\" rel=\"nofollow noreferrer\">https:\/\/i.ibb.co\/yPdrJKJ\/hax.gif<\/a><\/p>\n\n<p>Preprocessing code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def ProcessGameImage(I):    \n    GreyImage = skimage.color.rgb2gray(I)   \n    CroppedImage = GreyImage     \n    ReducedImage = skimage.transform.resize(CroppedImage,(DCQL_Global.IMGHEIGHT,DCQL_Global.IMGWIDTH))  \n    ReducedImage = skimage.exposure.rescale_intensity(ReducedImage, out_range = (0,255))     \n    ReducedImage = ReducedImage \/ 255.0\n\n    return ReducedImage.astype(np.float)\n<\/code><\/pre>\n\n<p>What's my fault?<\/p>\n","answers":[]}
+{"QuestionId":57809049,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-05T15:50:17.457","AcceptedAnswerId":null,"OwnerUserId":11427310.0,"Title":"tf.while_loop -> Number of inputs and outputs of body must match loop_vars: 1, 2","Body":"<p>I need to use a tf.while_loop but I got the following message : <\/p>\n\n<pre><code>ValueError: Number of inputs and outputs of body must match loop_vars: 1, 2\n<\/code><\/pre>\n\n<pre><code>total_loss, i = list(), tf.constant(0)\ndef sampled_softmax_body(i, total_loss):\n    loss = tf.nn.sampled_softmax_loss(weights=logits_weights, biases=logits_biases, labels=target, inputs=h2_decoder_outputs[:,i,:], num_sampled=n_sampled_softmax, num_classes=n_fra_words, partition_strategy=\"div\")     \n    total_loss.append(loss)\n    i = i + 1\n    return i, total_loss\ndef condition(i, total_loss):\n    return True and i &lt; max_words_per_sentence[\"fra\"]\ntf.while_loop(condition, sampled_softmax_body, [i, total_loss])\n<\/code><\/pre>\n","answers":[{"AnswerId":"57824035","CreationDate":"2019-09-06T14:35:17.543","ParentId":null,"OwnerUserId":"11427310","Title":null,"Body":"<p>I changed my strategy and come with this working :<\/p>\n\n<pre><code>total_loss, i = list(), tf.constant(0)\ndef sampled_softmax_body(i):\n    loss = tf.nn.sampled_softmax_loss(weights=logits_weights, biases=logits_biases, labels=target, inputs=h2_decoder_outputs[:,i,:], num_sampled=n_sampled_softmax, num_classes=n_fra_words, partition_strategy=\"div\")     \n    total_loss.append(loss)\n    return tf.add(i, 1)\ndef condition(i):\n    return i &lt; max_words_per_sentence[\"fra\"]\ntf.while_loop(condition, sampled_softmax_body, [i])\n<\/code><\/pre>\n\n<p>but how I am supposed to get the content of <code>total_list<\/code> ? \nbecause it is transform buy the loop. I have not the appended list I created.<\/p>\n\n<pre><code>In [8]: total_loss\nOut[8]: [&lt;tf.Tensor 'while\/softmax_cross_entropy_with_logits\/Reshape_2:0' shape=(?,) dtype=float32&gt;]\n<\/code><\/pre>\n\n<p>My goal is to use backpropagation using <code>total_loss<\/code><\/p>\n"}]}
+{"QuestionId":57809368,"AnswerCount":0,"Tags":"<python><keras><hdf5>","CreationDate":"2019-09-05T16:11:03.773","AcceptedAnswerId":null,"OwnerUserId":12026379.0,"Title":"How to load pretrained weights from hdf5 file for time distributed layer in keras?","Body":"<p>Normally we can load weights to a layer using hdf5 file by layer name as follow:<\/p>\n\n<pre><code>model.load_weights_from_hdf5_group_by_name(h5py_file, layers)\n<\/code><\/pre>\n\n<p>However, after applying the <code>TimeDistributed()<\/code> wrapper, the layer name becomes something like <em>time_distributed_#<\/em>. <\/p>\n\n<p>Then how can I load pre-trained weights for those layers?<\/p>\n","answers":[]}
+{"QuestionId":57809388,"AnswerCount":0,"Tags":"<python><tensorflow><pycharm>","CreationDate":"2019-09-05T16:12:36.813","AcceptedAnswerId":null,"OwnerUserId":12026323.0,"Title":"Tensorflow freeze_graph error list index out of range","Body":"<p>I m trying to implement one of tensorflow example to train images and save the model to use it on opencv.\nthe train part and saving .h5 file was perfect, even I test it on some of test image.\nwhen I was trying to freeze graph, I got an error of  list index out of range.\nI m new on tensorflow, please help<\/p>\n\n<p>the following is the code:<\/p>\n\n<pre><code>import os\nimport tensorflow as tf\nimport keras_preprocessing\nfrom keras_preprocessing import image\nfrom keras_preprocessing.image import ImageDataGenerator\nimport matplotlib.pyplot as plt\nfrom tensorflow.python.tools import freeze_graph\n\nrock_dir = os.path.join('tmp\/rps\/rock')\npaper_dir = os.path.join('tmp\/rps\/paper')\nscissors_dir = os.path.join('tmp\/rps\/scissors')\n\nprint('total training rock images:', len(os.listdir(rock_dir)))\nprint('total training paper images:', len(os.listdir(paper_dir)))\nprint('total training scissors images:', len(os.listdir(scissors_dir)))\n\nrock_files = os.listdir(rock_dir)\nprint(rock_files[:10])\n\npaper_files = os.listdir(paper_dir)\nprint(paper_files[:10])\n\nscissors_files = os.listdir(scissors_dir)\nprint(scissors_files[:10])\n\nTRAINING_DIR = \"tmp\/rps\/\"\ntraining_datagen = ImageDataGenerator(\n      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      fill_mode='nearest')\n\nVALIDATION_DIR = \"tmp\/rps-test-set\/\"\nvalidation_datagen = ImageDataGenerator(rescale = 1.\/255)\n\ntrain_generator = training_datagen.flow_from_directory(TRAINING_DIR,target_size=(150,150),class_mode='categorical')\n\nvalidation_generator = validation_datagen.flow_from_directory(VALIDATION_DIR,target_size=(150,150),class_mode='categorical')\ntf.compat.v1.reset_default_graph()\nwith tf.Graph().as_default():\n     with tf.compat.v1.Session() as sess:\n        model = tf.keras.models.Sequential([\n                # Note the input shape is the desired size of the image 150x150 with 3 bytes color\n                # This is the first convolution\n                tf.keras.layers.Conv2D(64, (3,3), activation='relu', input_shape=(150, 150, 3)),\n                 tf.keras.layers.MaxPooling2D(2, 2),\n                # The second convolution\n                tf.keras.layers.Conv2D(64, (3,3), activation='relu'),\n                tf.keras.layers.MaxPooling2D(2,2),\n                # The third convolution\n                tf.keras.layers.Conv2D(128, (3,3), activation='relu'),\n                tf.keras.layers.MaxPooling2D(2,2),\n                # The fourth convolution\n                tf.keras.layers.Conv2D(128, (3,3), activation='relu'),\n                tf.keras.layers.MaxPooling2D(2,2),\n                # Flatten the results to feed into a DNN\n                tf.keras.layers.Flatten(),\n                tf.keras.layers.Dropout(0.5),\n                # 512 neuron hidden layer\n                tf.keras.layers.Dense(512, activation='relu'),\n                tf.keras.layers.Dense(3, activation='softmax')\n                ])\n\n        model.summary()\n\n        model.compile(loss = 'categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])\n\n        history = model.fit_generator(\n            train_generator,\n            epochs=1,\n            validation_data = validation_generator,\n            verbose = 1)\n\n        model.save(\"rps.h5\")\n\n\n\n        init = tf.compat.v1.global_variables_initializer()\n        print(init)\n        sess.run(init)\n        print(sess.run(init))\n        saver = tf.compat.v1.train.Saver()\n        saver.save(sess, 'MODLE\/tensorflowModel.ckpt')\n        tf.io.write_graph(sess.graph.as_graph_def(), '.', 'MODLE\/tensorflowModel.pbtxt', as_text=True)\n         # tf.io.write_graph(sess.graph.as_graph_def(), '.', 'MODLE\/tensorflowModel.pb', as_text=False)\n\n        freeze_graph.freeze_graph(input_graph='MODLE\/tensorflowModel.pbtxt', input_saver='',\n                                   input_binary=False, input_checkpoint='MODLE\/tensorflowModel.ckpt',\n                                   output_node_names='softmax', restore_op_name='save\/restore_all',\n                                   filename_tensor_name='save\/Const:0', output_graph='MODLE\/tensorflowModel.pb',\n                                   clear_devices=True, initializer_nodes='')\n<\/code><\/pre>\n\n<p>this is the error<\/p>\n\n<pre><code>Use standard file APIs to check for files with this prefix.\nTraceback (most recent call last):\n  File \"C:\/Users\/Sufyan\/Desktop\/PYTHON\/TENSORFLOW\/rock paper.py\", line 95, in &lt;module&gt;\n    clear_devices=True, initializer_nodes='')\n  File \"C:\\Users\\Sufyan\\Desktop\\PYTHON\\TENSORFLOW\\venv\\lib\\site-packages\\tensorflow\\python\\tools\\freeze_graph.py\", line 363, in freeze_graph\n    checkpoint_version=checkpoint_version)\n  File \"C:\\Users\\Sufyan\\Desktop\\PYTHON\\TENSORFLOW\\venv\\lib\\site-packages\\tensorflow\\python\\tools\\freeze_graph.py\", line 190, in freeze_graph_with_def_protos\n    var_list=var_list, write_version=checkpoint_version)\n  File \"C:\\Users\\Sufyan\\Desktop\\PYTHON\\TENSORFLOW\\venv\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 831, in __init__\n    self.build()\n  File \"C:\\Users\\Sufyan\\Desktop\\PYTHON\\TENSORFLOW\\venv\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 843, in build\n    self._build(self._filename, build_save=True, build_restore=True)\n  File \"C:\\Users\\Sufyan\\Desktop\\PYTHON\\TENSORFLOW\\venv\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 880, in _build\n    build_save=build_save, build_restore=build_restore)\n  File \"C:\\Users\\Sufyan\\Desktop\\PYTHON\\TENSORFLOW\\venv\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 486, in _build_internal\n    names_to_saveables)\n  File \"C:\\Users\\Sufyan\\Desktop\\PYTHON\\TENSORFLOW\\venv\\lib\\site-packages\\tensorflow\\python\\training\\saving\\saveable_object_util.py\", line 341, in validate_and_slice_inputs\n    for converted_saveable_object in saveable_objects_for_op(op, name):\n  File \"C:\\Users\\Sufyan\\Desktop\\PYTHON\\TENSORFLOW\\venv\\lib\\site-packages\\tensorflow\\python\\training\\saving\\saveable_object_util.py\", line 207, in saveable_objects_for_op\n    variable, \"\", name)\n  File \"C:\\Users\\Sufyan\\Desktop\\PYTHON\\TENSORFLOW\\venv\\lib\\site-packages\\tensorflow\\python\\training\\saving\\saveable_object_util.py\", line 83, in __init__\n    self.handle_op = var.op.inputs[0]\n  File \"C:\\Users\\Sufyan\\Desktop\\PYTHON\\TENSORFLOW\\venv\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 2355, in __getitem__\n    return self._inputs[i]\nIndexError: list index out of range\n<\/code><\/pre>\n\n<p>I hope to get answer here<\/p>\n","answers":[]}
+{"QuestionId":57810025,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-05T17:00:29.547","AcceptedAnswerId":null,"OwnerUserId":9582951.0,"Title":"How to fix Tensorflow Error: Failed to create a NewWriteableFile","Body":"<p>I'm new with tensorflow, I'm facing an issue while executing a <a href=\"https:\/\/github.com\/hcmlab\/vadnet\/blob\/master\/train\/code\/main.py\" rel=\"nofollow noreferrer\">project<\/a> on window 10.<\/p>\n\n<p>Using :\npython-3.5.4\ntensorboard-gpu=1.14.0\ncuda-10<\/p>\n\n<blockquote>\n  <p>ERROR:\n  tensorflow.python.framework.errors_impl.NotFoundError: Failed to create a NewWriteableFile: C:\\Users\\Reverie-PC\\Desktop\\tools\\vadnet\\train\\nets\\vad_Model[48000,48000,24000,512,None]_Conv7[2,True,False,4,2]_SceAdam[1e-04,0.9,0.999,1e-08]\\ckpt\\model.ckpt-0_temp_df1af0f684d24212bc1ebde45848fe3f\/part-00000-of-00001.data-00000-of-00001.tempstate18107421248045726006 : The system cannot find the path specified.<\/p>\n<\/blockquote>\n\n<p>path is correct and already created \n\"nets\\vad_Model[48000,48000,24000,512,None]_Conv7[2,True,False,4,2]_SceAdam[1e-04,0.9,0.999,1e-08]\\ckpt\" by script but \"part-00000-of-00001.data-00000-of-00001.tempstate18107421248045726006\" not creating.<\/p>\n\n<p>I read somewhere, need to use an absolute path. I had tried that also but still facing the same error.\nthank you!<\/p>\n","answers":[]}
+{"QuestionId":57810157,"AnswerCount":0,"Tags":"<python><tensorflow><keras><multi-gpu>","CreationDate":"2019-09-05T17:10:45.203","AcceptedAnswerId":null,"OwnerUserId":6479070.0,"Title":"Custom loss function slows training of multi gpu model considerably","Body":"<p>I'm training variational autoencoders on protein structures using Keras' multi_gpu_model. When switching from normal AEs to VAEs, my model takes >3x longer per epoch to train. <\/p>\n\n<p>I identified the problem to be the loss function, changing it back to a built-in mse results in the same speed as seen before. <\/p>\n\n<p>I'm using more or less the same vae_loss implementation as seen in many tutorials:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def vae_loss(y_true, y_pred):\n    reconstruction_loss = recon_loss(y_true, y_pred)\n    kl_loss = beta * K.mean(\n             1 + K.flatten(z_log_var) - K.square(K.flatten(z_mean)) - K.exp(K.flatten(z_log_var)), axis=-1)\n    kl_loss \/= kl_loss_scaling # divide kl_loss by size of output dimension\n    total_loss = K.mean(reconstruction_loss + kl_loss)\n    return total_loss\n<\/code><\/pre>\n\n<p>When monitoring the GPU usage I realized, that they are well utilized and then drop to zero utilization after each epoch. The batchsize is adjusted for the number of GPUs and the exact same setup but with mse as loss works fine. It seems like the GPUs are waiting for the loss to be computed and hence have a considerable amount of downtime. (The effect is more strongly noticeable with smaller batchsizes, so increasing this parameter is somewhat of a solution, but I think this is far from optimal). <\/p>\n\n<p>Is this unavoidable since this loss is more expensive to compute or is there something I can adjust to achieve better performance?<\/p>\n","answers":[]}
+{"QuestionId":57810494,"AnswerCount":0,"Tags":"<tensorflow><threshold>","CreationDate":"2019-09-05T17:33:31.043","AcceptedAnswerId":null,"OwnerUserId":9010063.0,"Title":"What is default error threshold Tensorflow?","Body":"<p>My homework states that I am asked \"to report the default error threshold used in the TensorFlow default configuration for convergence. Usually it is documented for MNIST and CIFAR-10\".<\/p>\n\n<p>What's that? I am a beginner of BigData and this is my first project using Tensorflow<\/p>\n","answers":[]}
+{"QuestionId":57810732,"AnswerCount":0,"Tags":"<python><tensorflow><conv-neural-network>","CreationDate":"2019-09-05T17:51:44.697","AcceptedAnswerId":null,"OwnerUserId":7599739.0,"Title":"Convolutional neural network returns more than one output for each input","Body":"<p>I'm new to tensorflow and I'm working on a code that I have found on GitHub to find a car license plate on an input image. I changed the code for my needs but the problem now is I think my network structure is wrong, because when I feed an image to my network I get more than one desired output. My network input is <code>x = [?, 400, 400]<\/code> (question mark is for batch size) and the output should be a vector <code>y = [?, 208]<\/code>. When I set the batch <code>size = 10<\/code> I get a lot of errors because the output is more than 10, 208 vectors.<\/p>\n\n<pre><code>def 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, stride=(1, 1), padding='SAME'):\n  return tf.nn.conv2d(x, W, strides=[1, stride[0], stride[1], 1],\n                      padding=padding)\n\n\ndef max_pool(x, ksize=(2, 2), stride=(2, 2)):\n  return tf.nn.max_pool(x, ksize=[1, ksize[0], ksize[1], 1],\n                        strides=[1, stride[0], stride[1], 1], padding='SAME')\n\n\ndef avg_pool(x, ksize=(2, 2), stride=(2, 2)):\n  return tf.nn.avg_pool(x, ksize=[1, ksize[0], ksize[1], 1],\n                        strides=[1, stride[0], stride[1], 1], padding='SAME')\n\n\ndef convolutional_layers():\n    \"\"\"\n    Get the convolutional layers of the model.\n    \"\"\"\n    x = tf.placeholder(tf.float32, [None, None, None])\n\n    # First layer\n    W_conv1 = weight_variable([5, 5, 1, 48])\n    b_conv1 = bias_variable([48])\n    # x_expanded = tf.expand_dims(x, 3)\n    # h_conv1 = tf.nn.relu(conv2d(x_expanded, W_conv1) + b_conv1)\n    x_expanded = tf.expand_dims(x, 3)\n    h_conv1 = tf.nn.relu(conv2d(x_expanded, W_conv1) + b_conv1)\n    h_pool1 = max_pool(h_conv1, ksize=(2, 2), stride=(2, 2))\n    # print(h_pool1.shape)\n\n    # Second layer\n    W_conv2 = weight_variable([5, 5, 48, 64])\n    b_conv2 = bias_variable([64])\n\n    h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)\n    h_pool2 = max_pool(h_conv2, ksize=(2, 1), stride=(2, 1))\n    # print(h_pool2.shape)\n\n    # Third layer\n    W_conv3 = weight_variable([5, 5, 64, 104])\n    b_conv3 = bias_variable([104])\n\n    h_conv3 = tf.nn.relu(conv2d(h_pool2, W_conv3) + b_conv3)\n    h_pool3 = max_pool(h_conv3, ksize=(2, 2), stride=(2, 2))\n    # print(h_pool3.shape)\n\n    return x, h_pool3, [W_conv1, b_conv1,\n                        W_conv2, b_conv2,\n                        W_conv3, b_conv3]\n\n\ndef get_training_model():\n\n    x, conv_layer, conv_vars = convolutional_layers()\n\n    # Densely connected layer\n    W_fc1 = weight_variable([208, 512])\n    b_fc1 = bias_variable([512])\n\n    conv_layer_flat = tf.reshape(conv_layer, [-1, 208])\n\n    # print(conv_layer_flat.shape)\n\n    h_fc1 = tf.nn.relu(tf.matmul(conv_layer_flat, W_fc1) + b_fc1)\n\n    # Output layer\n    W_fc2 = weight_variable([512, 208])\n    b_fc2 = bias_variable([208])\n\n    y = tf.matmul(h_fc1, W_fc2) + b_fc2\n\n    return (x, y, conv_vars + [W_fc1, b_fc1, W_fc2, b_fc2])\n\n\ndef get_detect_model():\n\n    x, conv_layer, conv_vars = convolutional_layers()\n\n    # Densely connected layer\n    W_fc1 = weight_variable([208, 512])\n    b_fc1 = bias_variable([512])\n\n    conv_layer_flat = tf.reshape(conv_layer, [-1, 208])\n    h_fc1 = tf.nn.relu(tf.matmul(conv_layer_flat, W_fc1) + b_fc1)\n\n    # Output layer\n    W_fc2 = weight_variable([512, 208])\n    b_fc2 = bias_variable([208])\n\n    y = tf.matmul(h_fc1, W_fc2) + b_fc2\n\n    return (x, y, conv_vars + [W_fc1, b_fc1, W_fc2, b_fc2])\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57811445,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2019-09-05T18:49:21.490","AcceptedAnswerId":57812731.0,"OwnerUserId":4615389.0,"Title":"Tensorflow - use string labels to train neural network","Body":"<p>For an university project I have to implement a neural network for an OCR task using Tensorflow.\nThe training dataset consists of two files, <em>train-data.csv<\/em> and <em>train-target.csv<\/em>.\nIn <em>train-data<\/em> file every row is filled with bits of an 16x8 bitmap, in <em>train-target<\/em> file every row is a character [a-z] which is the label for the corresponding row in <em>train-data<\/em>.<\/p>\n\n<p>I'm having some issues with the label dataset, I've followed the tutorial with the MNIST dataset but here the difference is that I have string labels instead of a one-hot encoded vector.\nFollowing the tutorial I'm trying with the softmax function and the cross-entropy.<\/p>\n\n<pre><code># First y * tf.log(y_hat) computes the element-wise multiplication of the two resulting vectors\n\n# Second, tf.reduce_sum( , reduction_indices=[1]) computes the sum along the second dimension (the first one are the examples)\n# Finally, tf.reduce_mean() computes the mean over the first dimension, i.e. the examples\ncross_entropy = tf.reduce_mean(-tf.reduce_sum(tf.strings.to_number(y) * tf.math.log(y_hat), reduction_indices=[1]))\n\ntrain_step = tf.compat.v1.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)\n<\/code><\/pre>\n\n<p>In lines above I've used <code>tf.strings.to_number(y)<\/code> to convert the char to a numeric value.<\/p>\n\n<p>This conversion is causing issues when I run the session because the run() method does not accept tensor objects.<\/p>\n\n<pre><code>for _ in range(1000):\n    batch_xs, batch_ys = next_batch(100, raw_train_data, train_targets)\n    sess.run(train_step, feed_dict={x: batch_xs, y: tf.strings.to_number(batch_ys.reshape((100,1)))})\n<\/code><\/pre>\n\n<p>If I don't convert the char to a numeric value I got this error:<\/p>\n\n<pre><code>InvalidArgumentError: StringToNumberOp could not correctly convert string: e\n [[{{node StringToNumber}}]]\n<\/code><\/pre>\n\n<p>I'm trying to figure out how to solve this issue or how to train a neural network using character labels, it's the whole day that I'm working on this problem.\nDoes anyone know how to solve this?<\/p>\n","answers":[{"AnswerId":"57812731","CreationDate":"2019-09-05T20:41:07.947","ParentId":null,"OwnerUserId":"4615389","Title":null,"Body":"<p>Finally I've found the error.\nBecause I'm quite new to machine learning I've forgot that many algorithms does not handle categorical datasets.<\/p>\n\n<p>The solution has been to perform a one-hot encoding on the target labels and feed this new array to the newtork with this function:<\/p>\n\n<pre><code># define universe of possible input values\nalphabet = 'abcdefghijklmnopqrstuvwxyz'\n\n# define a mapping of chars to integers\nchar_to_int = dict((c, i) for i, c in enumerate(alphabet))\nint_to_char = dict((i, c) for i, c in enumerate(alphabet))\n\n\ndef one_hot_encode(data_array):\n    integer_encoded = [char_to_int[char] for char in data_array]\n\n    # one hot encode\n    onehot_encoded = list()\n    for value in integer_encoded:\n        letter = [0 for _ in range(len(alphabet))]\n        letter[value] = 1\n        onehot_encoded.append(letter)\n\n    return onehot_encoded\n<\/code><\/pre>\n"}]}
+{"QuestionId":57811473,"AnswerCount":0,"Tags":"<python><machine-learning><keras>","CreationDate":"2019-09-05T18:52:46.523","AcceptedAnswerId":null,"OwnerUserId":12026989.0,"Title":"Keras Validation Accuracy is Zero but other metrics are normal","Body":"<p>I am working on a computer vision problem in keras and I have run into a an interesting problem. My val_acc is 0.0000e+00. This is especially interesting as my other metrics such as loss, acc, and val_loss all are acting normally.<\/p>\n\n<p>This started happening when I switched from the Sequence data_generator to a custom one that I'm pretty sure is working as intended.<br><br> My issue is very similar to this one <a href=\"https:\/\/stackoverflow.com\/questions\/52197351\/validation-accuracy-is-0-with-keras-fit-generator\">validation accuracy is 0 with Keras fit_generator<\/a> but no answer was reached in that thread.<br><br>\nI have checked to make sure my activations and loss metrics are appropriate for my particular problem. I am using: loss='categorical_crossentropy' metrics=['accuracy'] and am attempting to predict the month that a certain spectrogram comes from.<br><br>The validation data is being loaded in the exact same way as the training data so I really can't figure out whats happening also even random guessing should give a 1\/12 val_acc right? It can't be zero.<\/p>\n\n<p>Here is my models architecture:<br><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = (Convolution2D(32,5,5,activation='relu',input_shape=(501,501,1)))(input_img)\nx = (MaxPooling2D(pool_size=(2,2)))(x)\nx = (Convolution2D(32,5,5,activation='relu'))(x)\nx = (MaxPooling2D(pool_size=(2,2)))(x)\nx = (Dropout(0.25))(x)\nx = (Flatten())(x)\nx = (Dense(128,activation='relu'))(x)\nx = (Dropout(0.5))(x)\nclassify = (Dense(12,activation='softmax', kernel_regularizer=regularizers.l1_l2(l1 = 0.001,l2 = 0.001)))(x)\n\nmodel = Model(input_img,classify)\nmodel.compile(loss='categorical_crossentropy',optimizer='nadam',metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>and here is my call to fit_generator:<br><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.fit_generator(generator = pd.data_generator(folder,'train'),\n                    validation_data = pd.data_generator(folder,'test'),\n                    steps_per_epoch=(120),\n                    validation_steps=(24),\n                    nb_epoch=20,\n                    verbose=1,\n                    shuffle=True,\n                    callbacks=[tensorboard_callback,early_stop_callback])\n<\/code><\/pre>\n\n<p>and finally here is the important part of my data generator: <br><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nif mode == 'test':\n        print('test')\n        while True:\n\n            for things in up.unpickle_batch(folder,50,6000,7200): #The last 1200 things in batches of 50\n                random.shuffle(things)\n                test_spect = []\n                test_months = []\n                for thing in things:\n                    test_spect.append(thing.spect) #GET BATCH DATA\n                    test_months.append(thing.month-1) #this is is here because the months go from 1-12 but should go from 0-11 for to_categorical   \n                x_test = np.asarray(test_spect) #PREPARE BATCH DATA\n                x_test =  x_test.astype('float32')\n                x_test \/= np.amax(x_test) #- 0.5\n                X_test = np.reshape(x_test, (-1,501, 501,1))\n\n\n                Y_test = np_utils.to_categorical(test_months,12)\n\n                yield X_test,Y_test #RETURN BATCH DATA\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57811573,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><tf.keras>","CreationDate":"2019-09-05T19:00:41.690","AcceptedAnswerId":57811752.0,"OwnerUserId":7643739.0,"Title":"Upon prediction, my loaded model is giving me an AttributeError","Body":"<p>I'm fairly new to Tensorflow and Machine Learning in general but I know enough that I've built a small model. Although, when loaded and I use <code>model.predict<\/code>, I get an attribute error: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\ncheckpoint_path = \"training_1\/cp.ckpt\"\n# Hyperparamters\nvocab_size = 2000\nembedding_dim = 16\nmax_length = 1\ntrunc_type = \"post\"\npadding_type = \"post\"\noov_tok = \"&lt;OOV&gt;\"\ntraining_size = 100\n\nmodel = tf.keras.Sequential([\n    tf.keras.layers.Embedding(\n        vocab_size, embedding_dim, input_length=max_length),\n    tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(32)),\n    tf.keras.layers.Dense(128, activation=\"relu\"),\n    tf.keras.layers.Dense(3, activation=\"softmax\")\n])\n\n# Compile the model\nmodel.compile(loss=\"sparse_categorical_crossentropy\",\n              optimizer=\"adam\", metrics=[\"accuracy\"])\n\nmodel.load_weights(checkpoint_path)\n\n\ntest = [\"Example of text here\"]\n\n\nprediction = model.predict(test)\nprint(prediction)\n<\/code><\/pre>\n\n<pre><code>Traceback (most recent call last):\n  File \".\/ModelTest.py\", line 36, in &lt;module&gt;\n    prediction = model.predict(test)\n  File \"\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/engine\/training.py\", line 1060, in predict\n    x, check_steps=True, steps_name='steps', steps=steps)\n  File \"\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/engine\/training.py\", line 2651, in _standardize_user_data\n    exception_prefix='input')\n  File \"\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/engine\/training_utils.py\", line 334, in standardize_input_data\n    standardize_single_array(x, shape) for (x, shape) in zip(data, shapes)\n  File \"\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/engine\/training_utils.py\", line 334, in &lt;listcomp&gt;\n    standardize_single_array(x, shape) for (x, shape) in zip(data, shapes)\n  File \"\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/engine\/training_utils.py\", line 265, in standardize_single_array\n    if (x.shape is not None and len(x.shape) == 1 and\nAttributeError: 'str' object has no attribute 'shape'\n<\/code><\/pre>\n","answers":[{"AnswerId":"57811752","CreationDate":"2019-09-05T19:16:13.650","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>Make sure the input you're giving is in the correct format for the model you build. In your case, an <code>Embedding<\/code> layer expects a 2D tensor. The data should be a numpy array that looks something like this: <code>[[0, 2, 64], [24, 6, 8]]<\/code>. Each number there represents a word, and each sequence of numbers represents a phrase. The whole tensor represents a batch of sequences. In my example, that's a batch of 2 sequences, each with 3 words.<\/p>\n\n<p>What you need to do is tokenize <code>\"Example of text here\"<\/code> using the correct vocabulary for the model you're loading. Once you do, you'll get an array like <code>[[3, 8, 4, 6]]<\/code>, where each of those numbers corresponds to one of the words in <code>\"Example of text here\"<\/code>. How to tokenize it correctly depends on how the data it was trained on was tokenized, and we don't know that without knowing where you got <code>training_1\/cp.ckpt<\/code> from.<\/p>\n"}]}
+{"QuestionId":57811901,"AnswerCount":0,"Tags":"<python><keras><neural-network>","CreationDate":"2019-09-05T19:28:18.323","AcceptedAnswerId":null,"OwnerUserId":3700020.0,"Title":"Weird Behavior of a Neural Network Implemented in Keras During Training","Body":"<p>I have implemented a complex neural network using Keras. The neural network trains normally and reaches up to 100% accuracy during training. However, during the evaluation, the results are very low and the output of the neural network is always 0.<\/p>\n\n<p>First, I investigated the weights of the neural network and found some nan values. I replaced these values and retrained the neural network. After the training is finished, no nan values exist in the weights anymore. However, the evaluation results are still the same.<\/p>\n\n<p>Out of desperation, I ran the following to train the network:<\/p>\n\n<pre><code>model.fit(dt[:200],labels[:200],epochs=100,batch_size=10,validation_data=(dt[:200],labels[:200]))\n<\/code><\/pre>\n\n<p>Despite all the nonsense in that code, I at least expected to receive identical values for the training and testing sets. But here is what I got:<\/p>\n\n<pre><code>Train on 200 samples, validate on 200 samples\nEpoch 1\/100\n200\/200 [==============================] - 21s 104ms\/step - loss: 0.2230 - acc: 0.9650 - val_loss: 7.9712 - val_acc: 0.5000\nEpoch 2\/100\n200\/200 [==============================] - 21s 104ms\/step - loss: 0.0302 - acc: 0.9850 - val_loss: 0.7086 - val_acc: 0.5000\nEpoch 3\/100\n200\/200 [==============================] - 21s 104ms\/step - loss: 0.0353 - acc: 0.9800 - val_loss: 0.7096 - val_acc: 0.5000\nEpoch 4\/100\n200\/200 [==============================] - 21s 104ms\/step - loss: 0.0349 - acc: 0.9900 - val_loss: 0.7088 - val_acc: 0.5000\nEpoch 5\/100\n200\/200 [==============================] - 21s 104ms\/step - loss: 0.0298 - acc: 0.9900 - val_loss: 0.7097 - val_acc: 0.5000\nEpoch 6\/100\n200\/200 [==============================] - 21s 103ms\/step - loss: 0.0457 - acc: 0.9800 - val_loss: 0.7065 - val_acc: 0.5000\nEpoch 7\/100\n200\/200 [==============================] - 21s 103ms\/step - loss: 0.0060 - acc: 0.9950 - val_loss: 0.7080 - val_acc: 0.5000\nEpoch 8\/100\n200\/200 [==============================] - 20s 102ms\/step - loss: 0.0337 - acc: 0.9900 - val_loss: 0.7045 - val_acc: 0.5000\nEpoch 9\/100\n200\/200 [==============================] - 21s 105ms\/step - loss: 0.0048 - acc: 1.0000 - val_loss: 0.7049 - val_acc: 0.5000\nEpoch 10\/100\n200\/200 [==============================] - 21s 104ms\/step - loss: 0.0020 - acc: 1.0000 - val_loss: 0.7059 - val_acc: 0.5000\n<\/code><\/pre>\n\n<p>The exact same sets are used for the training and evaluation but are returning different loss and accuracy values. Any ideas about what is going on!!!!?<\/p>\n","answers":[]}
+{"QuestionId":57812324,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorflow2.0><tf.keras>","CreationDate":"2019-09-05T20:05:02.783","AcceptedAnswerId":57832610.0,"OwnerUserId":4396475.0,"Title":"Saving and loading multiple models with the same graph in TensorFlow Functional API","Body":"<p>In the TensorFlow Functional API guide, there's an example shown where multiple models are created using the same graph of layers. (<a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/keras\/functional#using_the_same_graph_of_layers_to_define_multiple_models\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/beta\/guide\/keras\/functional#using_the_same_graph_of_layers_to_define_multiple_models<\/a>)<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>encoder_input = keras.Input(shape=(28, 28, 1), name='img')\nx = layers.Conv2D(16, 3, activation='relu')(encoder_input)\nx = layers.Conv2D(32, 3, activation='relu')(x)\nx = layers.MaxPooling2D(3)(x)\nx = layers.Conv2D(32, 3, activation='relu')(x)\nx = layers.Conv2D(16, 3, activation='relu')(x)\nencoder_output = layers.GlobalMaxPooling2D()(x)\n\nencoder = keras.Model(encoder_input, encoder_output, name='encoder')\nencoder.summary()\n\nx = layers.Reshape((4, 4, 1))(encoder_output)\nx = layers.Conv2DTranspose(16, 3, activation='relu')(x)\nx = layers.Conv2DTranspose(32, 3, activation='relu')(x)\nx = layers.UpSampling2D(3)(x)\nx = layers.Conv2DTranspose(16, 3, activation='relu')(x)\ndecoder_output = layers.Conv2DTranspose(1, 3, activation='relu')(x)\n\nautoencoder = keras.Model(encoder_input, decoder_output, name='autoencoder')\nautoencoder.summary()\n<\/code><\/pre>\n\n<p>Is it possible to save and load these two models while still sharing the same graph? If I save and load them in the following way:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Save\nencoder.save('encoder.h5')\nautoencoder.save('autoencoder.h5')\n\n# Load\nnew_encoder = keras.models.load_model('encoder.h5')\nnew_autoencoder = keras.models.load_model('autoencoder.h5')\n<\/code><\/pre>\n\n<p>the new encoder and autoencoder will no longer share the same graph, and therefore no longer train together.<\/p>\n","answers":[{"AnswerId":"57832610","CreationDate":"2019-09-07T09:58:25.743","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>That is a cool question. The encoder and autoencoder no longer share the same graph because they are being saved as disjoint models. In fact, <code>encoder<\/code> is being saved twice, as it is also embedded in <code>autoencoder<\/code>.<\/p>\n\n<p>To restore both models while still sharing the same graph, I would suggest the following approach:<\/p>\n\n<ol>\n<li><p>Name the <code>encoder<\/code>'s output layer. For example:<\/p>\n\n<pre><code>encoder_output = layers.GlobalMaxPooling2D(name='encoder_output')(x)\n<\/code><\/pre><\/li>\n<li><p>Save only the <code>autoencoder<\/code>:<\/p>\n\n<pre><code>autoencoder.save('autoencoder.h5')\n<\/code><\/pre><\/li>\n<li><p>Restore the <code>autoencoder<\/code>:<\/p>\n\n<pre><code>new_autoencoder = keras.models.load_model('autoencoder.h5')\n<\/code><\/pre><\/li>\n<li><p>Reconstruct the <code>encoder<\/code>'s graph from the restored <code>autoencoder<\/code> so that they share the common layers:<\/p>\n\n<pre><code>encoder_input = new_autoencoder.get_layer('img').input\nencoder_output = new_autoencoder.get_layer('encoder_output').output\nnew_encoder = keras.Model(encoder_input, encoder_output)\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>Alternatively, you could also save\/load the weights and reconstruct the graphs manually.<\/p>\n"}]}
+{"QuestionId":57812402,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras><deep-learning><loss-function>","CreationDate":"2019-09-05T20:11:28.197","AcceptedAnswerId":57895803.0,"OwnerUserId":2644955.0,"Title":"Keras custom loss function print tensor values","Body":"<p>I try to implement object detector like yolo. It uses complex custom loss function. So I need to print\/debug its tensors. I understand, that python code only builds computing graph so standard print won't work in not eager mode. \ntensorflow 1.12.0\nkeras 2.2.4<\/p>\n\n<p>I tried all methods from these posts <a href=\"https:\/\/stackoverflow.com\/questions\/55022966\/keras-custom-loss-function-not-printing-value-of-tensor\">Keras custom loss function not printing value of tensor<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/43448029\/debugging-keras-tensor-values\">Debugging keras tensor values<\/a> and nothing works. I tried tf.Print, tf.print, callback, K.tensor_print - the same result. In console i see only standart output messages. I`m not even sure if loss function been called. The answer from this post <a href=\"https:\/\/stackoverflow.com\/questions\/52023119\/keras-printing-intermediate-tensors-in-loss-function-tf-print-and-k-print-ten\/52023169\">Keras - printing intermediate tensors in loss function (tf.Print and K.print_tensor do not work...)<\/a> says that loss function sometimes doesn't even called! Ok, but how to use tf.contrib.eager.defun decorator then? The example is for pure tensorflow and dont't understand how to use it in keras.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nfrom keras.datasets import fashion_mnist\nimport matplotlib.pyplot as plt\nfrom keras.layers.convolutional import Conv2D, MaxPooling2D\nfrom keras.layers import Flatten, Dense, Dropout\nfrom keras.models import Sequential\nfrom keras import optimizers\nimport numpy as np\nfrom random import randrange\nfrom keras.callbacks import LambdaCallback\nimport keras.backend as K\nimport keras\n\nprint(tf.__version__)\nprint(keras.__version__)\n\n\nnum_filters = 64\n\n(x_train, y_train), (x_test, y_test) = fashion_mnist.load_data()\n\nx_train = x_train.astype('float32') \/ 255\nx_test = x_test.astype('float32') \/ 255\n\n#reshape\nx_train = x_train.reshape(60000,28,28,1)[:1000,...]\nx_test = x_test.reshape(10000,28,28,1)[:100,...]\n\n# One-hot encode the labels\ny_train = tf.keras.utils.to_categorical(y_train, 10)[:1000,...]\ny_test = tf.keras.utils.to_categorical(y_test, 10)[:100,]\n\n\nlabels = ['T-shirt\/top', 'Trouser', 'Pullover', 'Dress', 'Coat', 'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot']\n\n\nmodel = Sequential()\nmodel.add(Conv2D(input_shape=(28,28,1), filters=num_filters,kernel_size=3,strides=(1, 1),padding=\"valid\", activation='relu', use_bias=True))  \nmodel.add(MaxPooling2D(pool_size=2))\nmodel.add(Dropout(0.3))\n\nmodel.add(Conv2D(filters=num_filters,kernel_size=3,strides=(1, 1),padding=\"valid\", activation='relu', use_bias=True))  \nmodel.add(MaxPooling2D(pool_size=2))\nmodel.add(Dropout(0.3))\n\nmodel.add(Flatten())\nmodel.add(Dense(256, activation = 'relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(10, activation = 'softmax'))\n\n\n#model.summary()\n\n\n#loss 1\ndef customLoss(yTrue,yPred):\n    d = yPred-yTrue\n    d = K.print_tensor(d)\n    return K.mean(K.square(d), axis=-1)\n\n#loss 2\ndef cat_loss(y_true, y_pred):\n    d = y_true - y_pred\n    d = tf.Print(d, [d], \"Inside loss function\")\n    return tf.reduce_mean(tf.square(d))      \n\n\nmodel.compile(loss=customLoss,\n             optimizer='adam')\n\n\n\nimport keras.callbacks as cbks\n\n# 3 try to print with callback\nclass CustomMetrics(cbks.Callback):\n    def on_epoch_end(self, epoch, logs=None):\n        for k in logs:\n            if k.endswith('cat_loss'):\n               print(logs[k])   \n\n\n#checkpointer = ModelCheckpoint(filepath='model.weights.best.hdf5', verbose = 1, save_best_only=True)\nmodel.fit(x_train,\n         y_train,\n         #verbose=1,\n         batch_size=16,\n         epochs=10,\n         validation_data=(x_test, y_test),\n        callbacks=[CustomMetrics()])\n\n\n# Evaluate the model on test set\nscore = model.evaluate(x_test, y_test, verbose=0)\n\n# Print test accuracy\nprint('\\n', 'Test accuracy:', score)\n\n\nrand_img = randrange(100)\n\nresult = np.argmax(model.predict(x_test[rand_img].reshape(1,28,28,1)))\nplt.imshow(x_test[rand_img].reshape(28,28), cmap='gray')\nplt.title(labels[result])\n<\/code><\/pre>\n\n<pre><code>==========&gt;......] - ETA: 0s - loss: 0.0243\n832\/1000 [=======================&gt;......] - ETA: 0s - loss: 0.0242\nWarning (from warnings module):\n File \"C:\\Python36\\lib\\site-packages\\keras\\callbacks.py\", line 122\n   % delta_t_median)\nUserWarning: Method on_batch_end() is slow compared to the batch update (0.101474). Check your callbacks.\n\n976\/1000 [============================&gt;.] - ETA: 0s - loss: 0.0238\n992\/1000 [============================&gt;.] - ETA: 0s - loss: 0.0236\n1000\/1000 [==============================] - 3s 3ms\/step - loss: 0.0239 - val_loss: 0.0352\n\nTest accuracy: 0.035189545452594756```\n<\/code><\/pre>\n","answers":[{"AnswerId":"57895803","CreationDate":"2019-09-11T19:41:39.200","ParentId":null,"OwnerUserId":"2644955","Title":null,"Body":"<p>The truth was somewhere near. Idle doesn't output tf.Print and therefore K.print_tensor() to it's shell, so when i used cmd.exe python train.py, i saw tensor output.<\/p>\n"}]}
+{"QuestionId":57812517,"AnswerCount":1,"Tags":"<keras><tensorflow2.0>","CreationDate":"2019-09-05T20:22:30.940","AcceptedAnswerId":null,"OwnerUserId":287238.0,"Title":"How to create a keras model that depends dynamically on the input dimension (not batch size)?","Body":"<p>If you use the keras subclass api and you want to spawn a bunch of layers (n) depending on the the input dimension x = (batch_dim, n) is there a way to do this inside the build method? <\/p>\n\n<p>Or is the only way to pass the input dim into the model at <strong>init<\/strong> time so the layers can be created within the init scope? <\/p>\n\n<p>UPDATE: pseudo-code (untested) example<\/p>\n\n<pre><code>class BigModel(tf.keras.models.Model):\n    def __init__(self):\n        super.__init__()\n        self._my_submodels = list()\n\n    def build(self, input_shape):\n        for i in range(input_shape[1]):\n            self.my_submodels.append(MyModel(param=i))\n\n    def call(self, *inputs):\n        stuff = list()\n        for submodel in self.my_submodels:\n            stuff.append(submodel(*inputs))\n        # do something amazing with all the models\n        fan_in = ... # combine \n        return fan_in\n<\/code><\/pre>\n\n<p>You could <em>probably<\/em> rewrite the whole structure in a more vectorized way using one model with a lot of splits but it will be harder to read and deal with and I think the new tf 2.0 allows this kind of dynamism without any cost penalty.<\/p>\n","answers":[{"AnswerId":"57818212","CreationDate":"2019-09-06T08:17:09.687","ParentId":null,"OwnerUserId":"10371987","Title":null,"Body":"<p>Yes, instead of using <code>batch_shape = (batch_size, input_dim)<\/code>, use <code>batch_shape=(None, input_dim)<\/code> which allows arbitrary number of batch_size.<\/p>\n"}]}
+{"QuestionId":57813278,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><neural-network><deep-learning>","CreationDate":"2019-09-05T21:38:27.807","AcceptedAnswerId":57813489.0,"OwnerUserId":11186260.0,"Title":"How do you create Deep Q-Learning neural network to solve simple games like snake?","Body":"<p>I have been working for the last four days to try and create a simple working neural network(NN) that learns. I started off with the tower of Hanoi but that was quite tricky (doable with a Q-table)and no-one has really got good examples online so I decided to do it instead for snake game where there are lots of examples and tutorials. Long story short i have made a new super simple game where you have [0,0,0,0] and by picking 0, 1, 2, or 3 you change a 0 to a 1 or vice versa. So picking 1 would give an output of [0,1,0,0] and picking 1 again goes back to [0,0,0,0].  Very easy<\/p>\n\n<p>Despite the game being very easy I'm very much struggling to go from concept to practical as I have no education in coding.<\/p>\n\n<p>The final goal right now is to get the code below to be able to complete the game more than once. (it currently has run about 600 times and not once completed the 4 step problem) <\/p>\n\n<p>Current network architecture is 4 inputs 4 nodes in 1st hidden layer and 4 outputs and I would like to keep it this way even if the hidden layer is redundant just so I can learn how to do it correctly for other problems.<\/p>\n\n<p>If you cant be bothered to read the code and I don't blame you, ill put my mental psudocode here: <\/p>\n\n<ol>\n<li>setup variables, placeholders and import libraries<\/li>\n<li>run program 200 times to give it a chance to learn and each run has 20 turns<\/li>\n<li>run through the NN with \"states\" as the input and get out the output defined as \"output\" for future use<\/li>\n<li>game code<\/li>\n<li>the new reward for this specific game would just be the new set of states as (it has just occurred to me that this is the wrong way round ([0,1,0,0] for states should have rewards [1,0,1,1]) but i have already tried flipping it and it it still didn't work so that isn't the issue)<\/li>\n<li>my thinking was that i could get the next Q value by just running the new states through the NN<\/li>\n<li>this equation is taken directly from any deep q-learning tutorial on the internet and i think that maybe I have got this or one of the components for this wrong.<\/li>\n<li>run the gradient decent optimisation function<\/li>\n<\/ol>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf             ## importing libraries\nimport random\nimport numpy as np\n\nepsilon = 0.1                       ## create non tf variables\ny = 0.4\nmemory = []\nmemory1 = []\n\ninput_ = tf.placeholder(tf.float32, [None, 4], name='input_') \nW1 = tf.Variable(tf.random_normal([4, 4], stddev=0.03), name='W1') \nb1 = tf.Variable(tf.random_normal([4]), name='b1')    \nhidden_out = tf.add(tf.matmul(input_, W1), b1, name='hidden_out')   ## W for weights\nhidden_out = tf.nn.relu(hidden_out)                                 ## b for bias'\n\nW2 = tf.Variable(tf.random_normal([4, 4], stddev=0.03), name='W2')\nb2 = tf.Variable(tf.random_normal([4]), name='b2')\nQout = tf.add(tf.matmul(hidden_out, W2), b2, name='Qout')\nsig_out = tf.sigmoid(Qout, name='out')\n\n\nQ_target = tf.placeholder(shape=(None,4), dtype=tf.float32)\nloss = tf.reduce_sum(tf.square(Q_target - Qout))\noptimiser = tf.train.GradientDescentOptimizer(learning_rate=y).minimize(loss)\n\ninit_op = tf.global_variables_initializer()\n\nwith tf.compat.v1.Session() as sess:\n    sess.run(init_op)\n    for epoch in range(200):         ## run game 200 times\n        states = [0,0,0,0]\n        for _ in range(20):          ## 20 turns to do the correct 4 moves\n            if _ == 19:\n                memory1.append(states)\n            output = np.argmax(sess.run(sig_out, feed_dict={input_: [states]}))\n            ## sig_out is the output put through a sigmoid function\n            if random.random() &lt; epsilon:       ## this is the code for the game \n                output = random.randint(0,3)    ## ...\n            if states[output] == 0:             ## ...\n                states[output] = 1              ## ...\n            else:                               ## ...\n                states[output] = 0              ## ...\n            reward = states     \n            Qout1 = sess.run(sig_out, feed_dict={input_: [states]})\n            target = [reward + y*np.max(Qout1)]\n            sess.run([optimiser,loss], feed_dict={input_: [states], Q_target: target})\n<\/code><\/pre>\n\n<p>I haven't got any error messages in a while with this, the actual result would ideally be [1,1,1,1] every time.<\/p>\n\n<p>Thanks in advance for all of your help<\/p>\n\n<p>p.s. i couldn't think of an objective title for this, sorry<\/p>\n","answers":[{"AnswerId":"57813489","CreationDate":"2019-09-05T22:05:08.763","ParentId":null,"OwnerUserId":"12016524","Title":null,"Body":"<p>The <code>reward<\/code> value should be the objective value after an action has been taken. In your case, you have set <code>reward=states<\/code>. Since your function is attempting to maximize reward, the closer to [1, 1, 1, 1]  your state gets, the more reward your AI should receive. \n<br\/>\n<br\/><\/p>\n\n<p>Perhaps a reward function such as <code>reward = sum(states)<\/code> will solve your problem. <\/p>\n"}]}
+{"QuestionId":57813364,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-05T21:48:00.417","AcceptedAnswerId":null,"OwnerUserId":12027441.0,"Title":"given an array of row-numbers how to sum corresponding rows of a 2D tensor together?","Body":"<p>I have a 1D tensor that contains some integers and a 2D tensor that contains some values like this:<\/p>\n\n<pre><code>first_t = tf.const([0,2,3])\nsecond_t = tf.const([[4,1],[2,9],[1,6],[7,0]])\n<\/code><\/pre>\n\n<p>I want to sum rows of second_t which correspond first_t like this:<\/p>\n\n<pre><code>result = [12,7]\n<\/code><\/pre>\n\n<p>How can I do that?<\/p>\n","answers":[]}
+{"QuestionId":57813373,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2019-09-05T21:48:51.990","AcceptedAnswerId":57879487.0,"OwnerUserId":1056563.0,"Title":"How to parse the tensorflow events file?","Body":"<p>I would like to know how to extract the same performance results from the <code>events<\/code> file of the output of a model as does <code>Tensorboard<\/code> : specifically the Precision, Recall, and Loss numbers are most of interest. Here is a subset of them displayed on <code>Tensorboard<\/code> given the model checkpoint directory:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/rQMwU.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rQMwU.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I'm not sure if there self-documenting information or other metadata available for these models. This one in particular is the <code>Faster RNN Inception<\/code>: but are these outputs tied to a particular model or are they generic in format?<\/p>\n","answers":[{"AnswerId":"57879487","CreationDate":"2019-09-11T00:00:46.100","ParentId":null,"OwnerUserId":"1056563","Title":null,"Body":"<p>Found the approach in the <em>tensorboard<\/em> package:<\/p>\n\n<pre><code>  from tensorboard.backend.event_processing.event_accumulator import EventAccumulator\n  event_acc = EventAccumulator(evtf)\n  event_acc.Reload()\n<\/code><\/pre>\n\n<p>One of the entries is:<\/p>\n\n<pre><code>  scal_losses = event_acc.Scalars('Loss\/total_loss')\n<\/code><\/pre>\n\n<p>From that list we can extract such attributes as <code>Step<\/code> [number]  and <code>Value<\/code> (of the <em>loss<\/em>):<\/p>\n\n<pre><code> losses = sorted([[sevt.step, sevt.value] for sevt in scal_losses])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57813485,"AnswerCount":0,"Tags":"<multiprocessing><pytorch>","CreationDate":"2019-09-05T22:04:36.683","AcceptedAnswerId":null,"OwnerUserId":1408483.0,"Title":"Possible to extend PyTorch's share_memory() to other objects?","Body":"<p>In PyTorch nn.Module there's a method <code>share_memory()<\/code> which is used to share objects that implement the <code>nn.Module<\/code> interface, and this enables a model to be updated across multiple processes. This is known as 'Hogwild'. If it can be done with a PyTorch neural network model, why couldn't it be implemented for other objects as well, and negate needing to use queues to pass data between processes?<\/p>\n\n<p>For example, I'd like a global episode counter which could be increased from each process, why couldn't that object be shared using the same method as a neural net model?<\/p>\n","answers":[]}
+{"QuestionId":57813512,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-05T22:07:25.537","AcceptedAnswerId":57814159.0,"OwnerUserId":10195679.0,"Title":"Why slow learning in RNN implemented using for-loop?","Body":"<h2>Problem settings<\/h2>\n\n<p>As a beginner of RNN, I'm currently building a <strong>3-to-1<\/strong> <strong>autocompletion<\/strong> RNN model for 4-letter words, where the input is a 3-letter incomplete word and the output is a single-letter which completes the word. For example, I would desire to have the following model-prediction: <\/p>\n\n<ul>\n<li><strong>input<\/strong> : \"C\", \"A\", \"F\"<\/li>\n<li><strong>output<\/strong> : \"E\" <\/li>\n<\/ul>\n\n<hr>\n\n<h2>Codes - generate dataset<\/h2>\n\n<p>To get the desired result from an RNN model, I have made an (imbalanced) dataset as follows: <\/p>\n\n<pre><code>import string\nimport numpy as np       \nimport tensorflow as tf\nimport matplotlib.pyplot as plt\n\nalphList  = list(string.ascii_uppercase) # Define a list of alphabets\nalphToNum = {n: i for i, n in enumerate(alphList)} # dic of alphabet-numbers\n\n# Make dataset\n# define words of interest\nfourList = ['CARE', 'CODE', 'COME', 'CANE', 'COPE', 'FISH', 'JAZZ', 'GAME', 'WALK', 'QUIZ']\n\n# (len(Sequence), len(Batch), len(Observation)) following tensorflow-style\nfirst3Data = np.zeros((3, len(fourList), len(alphList)), dtype=np.int32)\nlast1Data  = np.zeros((len(fourList), len(alphList)), dtype=np.int32)\n\nfor idxObs, word in enumerate(fourList):\n    # Make an array of one-hot vectors consisting of first 3 letters\n    first3 = [alphToNum[n] for n in word[:-1]]\n    first3Data[:,idxObs,:] = np.eye(len(alphList))[first3]\n    # Make an array of one-hot vectors consisting of last 1 letter\n    last1  = alphToNum[word[3]]\n    last1Data[idxObs,:]    = np.eye(len(alphList))[last1]\n<\/code><\/pre>\n\n<p>So <code>fourList<\/code> contains the training data information, <code>first3Data<\/code> contains all the one-hot encoded first 3 letters of the training data, and <code>last1Data<\/code> contains all the one-hot encoded last 1 letter of the training data. <\/p>\n\n<hr>\n\n<h2>Codes - build model<\/h2>\n\n<p>Following the standard setting of <strong>3-to-1<\/strong> RNN model,I have made the following code. <\/p>\n\n<pre><code># Hyperparameters\nn_data        = len(fourList)\nn_input       = len(alphList)  # number of input units\nn_hidden      = 128            # number of hidden units\nn_output      = len(alphList)  # number of output units\nlearning_rate = 0.01\ntotal_epoch   = 100000\n\n# Variables (separate version)\nW_in  = tf.Variable(tf.random_normal([n_input, n_hidden]))\nW_rec = tf.Variable(tf.random_normal([n_hidden, n_hidden]))\nb_rec = tf.Variable(tf.random_normal([n_hidden]))\nW_out = tf.Variable(tf.random_normal([n_hidden, n_output]))\nb_out = tf.Variable(tf.random_normal([n_output]))\n\n# Manual calculation of RNN output\ndef RNNoutput(Xinput):\n    h_state    = tf.random_normal([1,n_hidden]) # initial hidden state\n\n    for iX in Xinput:\n        h_state = tf.nn.tanh(iX @ W_in + (h_state @ W_rec + b_rec))\n\n    rnn_output = h_state @ W_out + b_out\n    return(rnn_output)\n<\/code><\/pre>\n\n<p>Note that the <code>Manual calculation of RNN output<\/code> part basically rolls the hidden state exactly 4 times using the matrix multiplication and the <code>tanh<\/code> activation function as follows: <\/p>\n\n<pre><code>tf.nn.tanh(iX @ W_in + (h_state @ W_rec + b_rec))\n<\/code><\/pre>\n\n<p>Here, every time the whole data is passed, one epoch is completed. Thus I initialize the h_state every time I pass the data.  Additionally, note that <strong>I have not used a placeholder<\/strong>, which may be a cause of the learning instability.<\/p>\n\n<hr>\n\n<h2>Codes - train<\/h2>\n\n<p>I have used the following code to train the network. <\/p>\n\n<pre><code># Cost \/ optimizer definition\ncost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=RNNoutput(first3Data),\n                                                                 labels=last1Data))\noptimizer = tf.train.AdamOptimizer(learning_rate).minimize(cost)\n\n# Train and keep track of the loss history\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\n\nlossHistory = []\nfor epoch in range(total_epoch):\n    _, loss = sess.run([optimizer, cost])\n    lossHistory.append(loss)\n<\/code><\/pre>\n\n<hr>\n\n<h2>Question<\/h2>\n\n<p>The resulting learning curve looks as follows. Indeed, it shows an exponential decay. <\/p>\n\n<p>However, for me it looks too wiggly for this kind of simple example, showing some instabilities even in the late period of the learning. <\/p>\n\n<pre><code>plt.plot(range(total_epoch), lossHistory)\nplt.show()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/elehW.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/elehW.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<hr>\n\n<h2>Possible explanations?<\/h2>\n\n<p>I think the learning curve should show a square-like stable decay pattern as expected using <code>tensorflow<\/code> built-in functions (*). But I think this instability may be explained plausibly as follows:<\/p>\n\n<ul>\n<li>Instability in random initialization of parameters<\/li>\n<li>Numerical instability due to the successive addition when defining <code>RNNoutput<\/code><\/li>\n<li>Not using a <code>tensor for loop<\/code> but using the for loop directly in data<\/li>\n<\/ul>\n\n<p>But I don't think any of these played a crucial role. <strong>Is there any other solution to help me out<\/strong>? <\/p>\n\n<hr>\n\n<p>(*) I have seen a nearly square-patterned loss decay using <code>tensorflow<\/code> built-in functions for simple RNN. But sorry that I have not included the results to be compared, since I run out of time... I think I can edit shortly. <\/p>\n","answers":[{"AnswerId":"57814159","CreationDate":"2019-09-05T23:49:31.990","ParentId":null,"OwnerUserId":"7933388","Title":null,"Body":"<p>This modification where the initial state is set to be zero seems to solve the problem. <\/p>\n\n<pre><code># Variables (separate version)\nW_in  = tf.Variable(tf.random_normal([n_input, n_hidden]))\nW_rec = tf.Variable(tf.random_normal([n_hidden, n_hidden]))\nb_rec = tf.Variable(tf.random_normal([n_hidden]))\nW_out = tf.Variable(tf.random_normal([n_hidden, n_output]))\nb_out = tf.Variable(tf.random_normal([n_output]))\nh_init = tf.zeros([1,n_hidden])\n\n# Manual calculation of RNN output\ndef RNNoutput(Xinput):\n    h_state    =  h_init # initial hidden state\n\n    for iX in Xinput:\n        h_state = tf.nn.tanh(iX @ W_in + (h_state @ W_rec + b_rec))\n\n    rnn_output = h_state @ W_out + b_out\n    return(rnn_output)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57813531,"AnswerCount":0,"Tags":"<python><tensorflow><keras><keras-layer>","CreationDate":"2019-09-05T22:10:34.077","AcceptedAnswerId":null,"OwnerUserId":9133201.0,"Title":"How to load trained model and specify which layers if the network has multioutput?","Body":"<p>I have network like this:<\/p>\n\n<pre><code>                        +----+   +----+     \n                        &gt; C  |--&gt;| D  |--&gt;Classification\n                       \/+----+   +----+     \n                      \/                \n                     \/                 \n    +----+    +----+\/                  \n --&gt;| A  |---&gt;| B  |\\                  \n    +----+    +----+ \\                 \n                      \\                \n                       \\+----+   +----+      \n                        &gt; E  |--&gt;| F  |--&gt;Regression\n                        +----+   +----+     \n<\/code><\/pre>\n\n<p>Then, after training, I want to have layer: <code>A - B - E<\/code> for inference. How can I load my trained model for this kind of network?\nI am using Keras and Tensorflow.\nThanks!<\/p>\n","answers":[]}
+{"QuestionId":57813647,"AnswerCount":0,"Tags":"<tensorflow><random><keras><seed>","CreationDate":"2019-09-05T22:25:06.030","AcceptedAnswerId":null,"OwnerUserId":735008.0,"Title":"Random seed in tf.keras.models","Body":"<p>When we build a model using <code>tf.keras<\/code>, I want the output to have a consistent value. That is to say, I do not want the output to randomly change. For example, if I run the following code, the output is always different, and I guess this may be due to the random initialization of parameters.<\/p>\n\n<pre><code>inputs1 = tf.keras.layers.Input(shape=(3, 1)) \nlstm1 = tf.keras.layers.LSTM(1)(inputs1)\nmodel = tf.keras.models.Model(inputs = inputs1, outputs = lstm1)\n\ndata  = np.array([0.1, 0.2, 0.3]).reshape(1, 3, 1)\nprint (model.predict(data))\n<\/code><\/pre>\n\n<p>To resolve this problem, I observe that adding the following line at the beginning of the file solves the problem.<\/p>\n\n<pre><code>tf.random.set_random_seed(0)\n<\/code><\/pre>\n\n<p>I have two questions regarding this.<\/p>\n\n<ol>\n<li><p>Is the above approach a right approach when I build a model using tf.keras?<\/p><\/li>\n<li><p>If I have two models in a file, and if I want to set the random seed only for one of them, then how could I do this?<\/p><\/li>\n<\/ol>\n\n<p>Thank you!<\/p>\n","answers":[]}
+{"QuestionId":57813673,"AnswerCount":0,"Tags":"<java><android><tensorflow><canvas><view>","CreationDate":"2019-09-05T22:27:52.137","AcceptedAnswerId":null,"OwnerUserId":9137167.0,"Title":"09-06 01:23:09.141 8469-8469\/org.tensorflow.demo A\/libc: Fatal signal 11 (SIGSEGV), code 1, fault addr 0x58 in tid 8469 (tensorflow.demo)","Body":"<p>Whenever I wanted to draw something on canvas I get this error what should I do.<\/p>\n\n<p><code>09-06 01:23:09.141 8469-8469\/org.tensorflow.demo A\/libc: Fatal signal 11 (SIGSEGV), code 1, fault addr 0x58 in tid 8469 (tensorflow.demo)\n<\/code><\/p>\n","answers":[]}
+{"QuestionId":57813719,"AnswerCount":0,"Tags":"<python><tensorflow><faster-rcnn>","CreationDate":"2019-09-05T22:33:59.267","AcceptedAnswerId":null,"OwnerUserId":9316074.0,"Title":"How to produce masks instead of boxes when using object detection?","Body":"<p>I am using the TensorFlow Model Zoo Object Detection and would like to change the image output to have masks instead of boxes. <\/p>\n\n<p>I've tried to retrain the network with the masks_rcnn_inception_V2 model instead with no avail.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>vis_util.visualize_boxes_and_labels_on_image_array(\n        imageOutput,\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=2,\n        min_score_thresh=0.60,\n        skip_scores=True,\n        skip_labels=True)\ncv2.imwrite('..\/project\/test.png', imageOutput)\n    # Clean up\n    print(\"10 secs\")\n\n    time.sleep(10)\n<\/code><\/pre>\n\n<p>What argument do I need to include in this to produce masks instead of the boxes on the stored image?<\/p>\n","answers":[]}
+{"QuestionId":57813806,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2019-09-05T22:50:19.813","AcceptedAnswerId":57814782.0,"OwnerUserId":8725045.0,"Title":"Apply feature columns without tf.Estimator (Tensorflow 2.0.0-rc0)","Body":"<p>In the Tensorflow tf.Estimator and tf.feature_column docs it is well documented, how to use feature columns together with an Estimator e.g. in order to one-hot encode the categorical features in the dataset being used.<\/p>\n\n<p>However, I want to \"apply\" my feature columns directly to a tf.dataset which I create from a .csv file (with two columns: UserID, MovieID), without even defining a model or an Estimator. (Reason: I want to check what's happening exactly in my datapipeline, i.e. I'd like to be able to run a batch of samples through my the pipeline, and then see in the output how the features got encoded.)<\/p>\n\n<p>This is what I have tried so far:<\/p>\n\n<pre><code>column_names = ['UserID', 'MovieID']\n\nuser_col = tf.feature_column.categorical_column_with_hash_bucket(key='UserID', hash_bucket_size=1000)\nmovie_col = tf.feature_column.categorical_column_with_hash_bucket(key='MovieID', hash_bucket_size=1000)\nfeature_columns = [tf.feature_column.indicator_column(user_col), tf.feature_column.indicator_column(movie_col)]\n\nfeature_layer = tf.keras.layers.DenseFeatures(feature_columns=feature_columns)\n\ndef process_csv(line):\n  fields = tf.io.decode_csv(line, record_defaults=[tf.constant([], dtype=tf.int32)]*2, field_delim=\";\")\n  features = dict(zip(column_names, fields))\n\n  return features \n\nds = tf.data.TextLineDataset(csv_filepath)\nds = ds.map(process_csv, num_parallel_calls=4)\nds = ds.batch(10)\nds.map(lambda x: feature_layer(x))\n<\/code><\/pre>\n\n<p>However the last line with the map call raises the following error:<\/p>\n\n<blockquote>\n  <p>ValueError: Column dtype and SparseTensors dtype must be compatible.\n  key: MovieID, column dtype: , tensor dtype: <\/p>\n<\/blockquote>\n\n<p>I'm not sure what this error means...\nI also tried to define a tf.keras model with only the feature_layer I defined, and then run .predict() on my dataset - instead of using ds.map(lambda x: feature_layer(x)):<\/p>\n\n<pre><code>model = tf.keras.Sequential([feature_layer])\nmodel.compile()\nmodel.predict(ds)\n<\/code><\/pre>\n\n<p>However, this results exactly in the same error as above.\nDoes anybody have an idea what is going wrong? Is there maybe an easier way to achieve this?<\/p>\n","answers":[{"AnswerId":"57814782","CreationDate":"2019-09-06T01:49:37.367","ParentId":null,"OwnerUserId":"8725045","Title":null,"Body":"<p>Just found the issue:\ntf.feature_column.categorical_column_with_hash_bucket() takes an optional argument dtype, which is set to tf.dtypes.string by default.\nHowever, the datatype of my columns is numerical (tf.dtypes.int32).\nThis solved the issue:<\/p>\n\n<pre><code>tf.feature_column.categorical_column_with_hash_bucket(key='UserID', hash_bucket_size=1000, dtype=tf.dtypes.int32)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57814030,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-05T23:27:45.137","AcceptedAnswerId":null,"OwnerUserId":7933388.0,"Title":"How should I use tf.scan to implement this recurrence relation?","Body":"<p>I would like to implement the following function in Tensorflow.<\/p>\n\n<p>The input is a tensor <code>x<\/code> of shape <code>(n, 1)<\/code><\/p>\n\n<p>The output is a tensor <code>y<\/code> of shape <code>(n, 1)<\/code>.<\/p>\n\n<p>I fix a tensor <code>a<\/code> of shape <code>(2, 1)<\/code>.<\/p>\n\n<p>Then I want to calculate <code>y<\/code> by the following formula.<\/p>\n\n<pre><code>y[0] = x[0]\ny[1] = x[1] - a[0] * x[0]\ny[2] = x[2] - a[0] * x[1] - a[1] * x[0]\n... \ny[i] = x[i] - a[0] * x[i - 1] - a[1] * x[i - 2]\n<\/code><\/pre>\n\n<p>I would like the entries of <code>a<\/code> to be trainable.<\/p>\n\n<p>Thanks for any help you can provide!<\/p>\n\n<p>EDIT:  I am attempting a solution that might look like this: <\/p>\n\n<pre><code>elems = np.array([1, 0, 1, 0, 1, 1, 1, 0, 1, 0])\n\ninit = (np.array([0, 0]), 0)\n\n# a = initialize\n# y = elems[0]\n\n# test first step\n# print((np.array([a[0][1], a[1]]), a[0][0] + a[0][1] + y))\n\na1 = tf.Variable(-1)\na2 = tf.Variable(-1)\n\ndef f(x, y):\n    return (tf.stack([x[0][1], x[1]], 0), y - a1 * x[0][1] - a2 * x[0][1])\n\nc = tf.scan(f, elems, initializer=init)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print(sess.run(c))\n<\/code><\/pre>\n\n<p>Where <code>elems<\/code> is the input <code>x<\/code> and the last column of <code>c<\/code> should be <code>y<\/code>.  It doesn't work yet, the answer should be<\/p>\n\n<p><code>1     1     3     4     8    13    22    35    58    93<\/code><\/p>\n","answers":[{"AnswerId":"57814982","CreationDate":"2019-09-06T02:24:43.597","ParentId":null,"OwnerUserId":"7933388","Title":null,"Body":"<p>I was able to solve the problem.<\/p>\n\n<pre><code>elems = np.array([1, 0, 1, 0, 1, 1, 1, 0, 1, 0])\n\ninit = (np.array([0, 0]), 0)\n\na1 = tf.Variable(-1)\na2 = tf.Variable(-1)\n\ndef f(x, y):\n    return (tf.stack([x[0][1], x[1]], 0), y - a2 * x[0][1] - a1 * x[1])\n\nc = tf.scan(f, elems, initializer=init)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print(sess.run(c))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57814535,"AnswerCount":0,"Tags":"<python><pytorch><conda><torch>","CreationDate":"2019-09-06T01:03:31.660","AcceptedAnswerId":null,"OwnerUserId":3204706.0,"Title":"\"AssertionError: Torch not compiled with CUDA enabled\" in spite upgrading to CUDA version","Body":"<p>I figured out this is a popular question, but still I couldn't find a solution for that.<\/p>\n\n<p>I'm trying to run a simple repo <a href=\"https:\/\/github.com\/swathikirans\/violence-recognition-pytorch\" rel=\"nofollow noreferrer\">Here<\/a> which uses <code>PyTorch<\/code>. Although I just upgraded my Pytorch to the latest CUDA version from pytorch.org (<code>1.2.0<\/code>), it still throws the same error. I'm on Windows 10 and use conda with python 3.7.<\/p>\n\n<pre><code>    raise AssertionError(\"Torch not compiled with CUDA enabled\")\nAssertionError: Torch not compiled with CUDA enabled\n<\/code><\/pre>\n\n<p>How to fix the problem?<\/p>\n\n<p>Here is my <code>conda list<\/code>:<\/p>\n\n<pre><code># Name                    Version                   Build  Channel\n_ipyw_jlab_nb_ext_conf    0.1.0                    py37_0    anaconda\n_pytorch_select           1.1.0                       cpu    anaconda\n_tflow_select             2.3.0                       mkl    anaconda\nabsl-py                   0.7.1                    pypi_0    pypi\nalabaster                 0.7.12                   py37_0    anaconda\nanaconda                  2019.07                  py37_0    anaconda\nanaconda-client           1.7.2                    py37_0    anaconda\nanaconda-navigator        1.9.7                    py37_0    anaconda\nanaconda-project          0.8.3                      py_0    anaconda\nargparse                  1.4.0                    pypi_0    pypi\nasn1crypto                0.24.0                   py37_0    anaconda\nastor                     0.8.0                    pypi_0    pypi\nastroid                   2.2.5                    py37_0    anaconda\nastropy                   3.2.1            py37he774522_0    anaconda\natomicwrites              1.3.0                    py37_1    anaconda\nattrs                     19.1.0                   py37_1    anaconda\nbabel                     2.7.0                      py_0    anaconda\nbackcall                  0.1.0                    py37_0    anaconda\nbackports                 1.0                        py_2    anaconda\nbackports-csv             1.0.7                    pypi_0    pypi\nbackports-functools-lru-cache 1.5                      pypi_0    pypi\nbackports.functools_lru_cache 1.5                        py_2    anaconda\nbackports.os              0.1.1                    py37_0    anaconda\nbackports.shutil_get_terminal_size 1.0.0                    py37_2    anaconda\nbackports.tempfile        1.0                        py_1    anaconda\nbackports.weakref         1.0.post1                  py_1    anaconda\nbeautifulsoup4            4.7.1                    py37_1    anaconda\nbitarray                  0.9.3            py37he774522_0    anaconda\nbkcharts                  0.2                      py37_0    anaconda\nblas                      1.0                         mkl    anaconda\nbleach                    3.1.0                    py37_0    anaconda\nblosc                     1.16.3               h7bd577a_0    anaconda\nbokeh                     1.2.0                    py37_0    anaconda\nboto                      2.49.0                   py37_0    anaconda\nbottleneck                1.2.1            py37h452e1ab_1    anaconda\nbzip2                     1.0.8                he774522_0    anaconda\nca-certificates           2019.5.15                     0    anaconda\ncertifi                   2019.6.16                py37_0    anaconda\ncffi                      1.12.3           py37h7a1dbc1_0    anaconda\nchainer                   6.2.0                    pypi_0    pypi\nchardet                   3.0.4                    py37_1    anaconda\ncheroot                   6.5.5                    pypi_0    pypi\ncherrypy                  18.1.2                   pypi_0    pypi\nclick                     7.0                      py37_0    anaconda\ncloudpickle               1.2.1                      py_0    anaconda\nclyent                    1.2.2                    py37_1    anaconda\ncolorama                  0.4.1                    py37_0    anaconda\ncomtypes                  1.1.7                    py37_0    anaconda\nconda                     4.7.11                   py37_0    anaconda\nconda-build               3.18.9                   py37_3    anaconda\nconda-env                 2.6.0                         1    anaconda\nconda-package-handling    1.3.11                   py37_0    anaconda\nconda-verify              3.4.2                      py_1    anaconda\nconsole_shortcut          0.1.1                         3    anaconda\nconstants                 0.6.0                    pypi_0    pypi\ncontextlib2               0.5.5                    py37_0    anaconda\ncpuonly                   1.0                           0    pytorch\ncryptography              2.7              py37h7a1dbc1_0    anaconda\ncudatoolkit               10.0.130                      0    anaconda\ncurl                      7.65.2               h2a8f88b_0    anaconda\ncycler                    0.10.0                   py37_0    anaconda\ncython                    0.29.12          py37ha925a31_0    anaconda\ncytoolz                   0.10.0           py37he774522_0    anaconda\ndask                      2.1.0                      py_0    anaconda\ndask-core                 2.1.0                      py_0    anaconda\ndecorator                 4.4.0                    py37_1    anaconda\ndefusedxml                0.6.0                      py_0    anaconda\ndistributed               2.1.0                      py_0    anaconda\ndocutils                  0.14                     py37_0    anaconda\nentrypoints               0.3                      py37_0    anaconda\net_xmlfile                1.0.1                    py37_0    anaconda\nez-setup                  0.9                      pypi_0    pypi\nfastcache                 1.1.0            py37he774522_0    anaconda\nfasttext                  0.9.1                    pypi_0    pypi\nfeedparser                5.2.1                    pypi_0    pypi\nffmpeg                    4.1.3                h6538335_0    conda-forge\nfilelock                  3.0.12                     py_0    anaconda\nfirst                     2.0.2                    pypi_0    pypi\nflask                     1.1.1                      py_0    anaconda\nfreetype                  2.9.1                ha9979f8_1    anaconda\nfuture                    0.17.1                   py37_0    anaconda\ngast                      0.2.2                    py37_0    anaconda\nget                       2019.4.13                pypi_0    pypi\nget_terminal_size         1.0.0                h38e98db_0    anaconda\ngevent                    1.4.0            py37he774522_0    anaconda\nglob2                     0.7                        py_0    anaconda\ngoogle-pasta              0.1.7                    pypi_0    pypi\ngraphviz                  2.38.0                        4    anaconda\ngreenlet                  0.4.15           py37hfa6e2cd_0    anaconda\ngrpcio                    1.22.0                   pypi_0    pypi\nh5py                      2.9.0            py37h5e291fa_0    anaconda\nhdf5                      1.10.4               h7ebc959_0    anaconda\nheapdict                  1.0.0                    py37_2    anaconda\nhtml5lib                  1.0.1                    py37_0    anaconda\nhttp-client               0.1.22                   pypi_0    pypi\nhypothesis                4.34.0                   pypi_0    pypi\nicc_rt                    2019.0.0             h0cc432a_1    anaconda\nicu                       58.2                 ha66f8fd_1    anaconda\nidna                      2.8                      py37_0    anaconda\nimageio                   2.4.1                    pypi_0    pypi\nimageio-ffmpeg            0.3.0                    pypi_0    pypi\nimagesize                 1.1.0                    py37_0    anaconda\nimportlib_metadata        0.17                     py37_1    anaconda\nimutils                   0.5.2                    pypi_0    pypi\nintel-openmp              2019.0                   pypi_0    pypi\nipykernel                 5.1.1            py37h39e3cac_0    anaconda\nipython                   7.6.1            py37h39e3cac_0    anaconda\nipython_genutils          0.2.0                    py37_0    anaconda\nipywidgets                7.5.0                      py_0    anaconda\nisort                     4.3.21                   py37_0    anaconda\nitsdangerous              1.1.0                    py37_0    anaconda\njaraco-functools          2.0                      pypi_0    pypi\njdcal                     1.4.1                      py_0    anaconda\njedi                      0.13.3                   py37_0    anaconda\njinja2                    2.10.1                   py37_0    anaconda\njoblib                    0.13.2                   py37_0    anaconda\njpeg                      9b                   hb83a4c4_2    anaconda\njson5                     0.8.4                      py_0    anaconda\njsonschema                3.0.1                    py37_0    anaconda\njupyter                   1.0.0                    py37_7    anaconda\njupyter_client            5.3.1                      py_0    anaconda\njupyter_console           6.0.0                    py37_0    anaconda\njupyter_core              4.5.0                      py_0    anaconda\njupyterlab                1.0.2            py37hf63ae98_0    anaconda\njupyterlab_server         1.0.0                      py_0    anaconda\nkeras                     2.2.4                         0    anaconda\nkeras-applications        1.0.8                      py_0    anaconda\nkeras-base                2.2.4                    py37_0    anaconda\nkeras-preprocessing       1.1.0                      py_1    anaconda\nkeyring                   18.0.0                   py37_0    anaconda\nkiwisolver                1.1.0            py37ha925a31_0    anaconda\nkrb5                      1.16.1               hc04afaa_7\nlazy-object-proxy         1.4.1            py37he774522_0    anaconda\nlibarchive                3.3.3                h0643e63_5    anaconda\nlibcurl                   7.65.2               h2a8f88b_0    anaconda\nlibiconv                  1.15                 h1df5818_7    anaconda\nliblief                   0.9.0                ha925a31_2    anaconda\nlibmklml                  2019.0.5                      0    anaconda\nlibpng                    1.6.37               h2a8f88b_0    anaconda\nlibprotobuf               3.8.0                h7bd577a_0    anaconda\nlibsodium                 1.0.16               h9d3ae62_0    anaconda\nlibssh2                   1.8.2                h7a1dbc1_0    anaconda\nlibtiff                   4.0.10               hb898794_2    anaconda\nlibxml2                   2.9.9                h464c3ec_0    anaconda\nlibxslt                   1.1.33               h579f668_0    anaconda\nllvmlite                  0.29.0           py37ha925a31_0    anaconda\nlocket                    0.2.0                    py37_1    anaconda\nlxml                      4.3.4            py37h1350720_0    anaconda\nlz4-c                     1.8.1.2              h2fa13f4_0    anaconda\nlzo                       2.10                 h6df0209_2    anaconda\nm2w64-gcc-libgfortran     5.3.0                         6\nm2w64-gcc-libs            5.3.0                         7\nm2w64-gcc-libs-core       5.3.0                         7\nm2w64-gmp                 6.1.0                         2\nm2w64-libwinpthread-git   5.0.0.4634.697f757               2\nmake-dataset              1.0                      pypi_0    pypi\nmarkdown                  3.1.1                    py37_0    anaconda\nmarkupsafe                1.1.1            py37he774522_0    anaconda\nmatplotlib                3.1.0            py37hc8f65d3_0    anaconda\nmccabe                    0.6.1                    py37_1    anaconda\nmenuinst                  1.4.16           py37he774522_0    anaconda\nmistune                   0.8.4            py37he774522_0    anaconda\nmkl                       2019.0                   pypi_0    pypi\nmkl-service               2.0.2            py37he774522_0    anaconda\nmkl_fft                   1.0.12           py37h14836fe_0    anaconda\nmkl_random                1.0.2            py37h343c172_0    anaconda\nmock                      3.0.5                    py37_0    anaconda\nmore-itertools            7.0.0                    py37_0    anaconda\nmoviepy                   1.0.0                    pypi_0    pypi\nmpmath                    1.1.0                    py37_0    anaconda\nmsgpack-python            0.6.1            py37h74a9793_1    anaconda\nmsys2-conda-epoch         20160418                      1\nmultipledispatch          0.6.0                    py37_0    anaconda\nmysqlclient               1.4.2.post1              pypi_0    pypi\nnavigator-updater         0.2.1                    py37_0    anaconda\nnbconvert                 5.5.0                      py_0    anaconda\nnbformat                  4.4.0                    py37_0    anaconda\nnetworkx                  2.3                        py_0    anaconda\nninja                     1.9.0            py37h74a9793_0    anaconda\nnltk                      3.4.4                    py37_0    anaconda\nnose                      1.3.7                    py37_2    anaconda\nnotebook                  6.0.0                    py37_0    anaconda\nnumba                     0.44.1           py37hf9181ef_0    anaconda\nnumexpr                   2.6.9            py37hdce8814_0    anaconda\nnumpy                     1.16.4                   pypi_0    pypi\nnumpy-base                1.16.4           py37hc3f5095_0    anaconda\nnumpydoc                  0.9.1                      py_0    anaconda\nolefile                   0.46                     py37_0    anaconda\nopencv-contrib-python     4.1.0.25                 pypi_0    pypi\nopencv-python             4.1.0.25                 pypi_0    pypi\nopenpyxl                  2.6.2                      py_0    anaconda\nopenssl                   1.1.1c               he774522_1    anaconda\npackaging                 19.0                     py37_0    anaconda\npandas                    0.24.2           py37ha925a31_0    anaconda\npandoc                    2.2.3.2                       0    anaconda\npandocfilters             1.4.2                    py37_1    anaconda\nparso                     0.5.0                      py_0    anaconda\npartd                     1.0.0                      py_0    anaconda\npath.py                   12.0.1                     py_0    anaconda\npathlib2                  2.3.4                    py37_0    anaconda\npatsy                     0.5.1                    py37_0    anaconda\npattern                   3.6                      pypi_0    pypi\npdfminer-six              20181108                 pypi_0    pypi\npep8                      1.7.1                    py37_0    anaconda\npickleshare               0.7.5                    py37_0    anaconda\npillow                    6.1.0            py37hdc69c19_0    anaconda\npip                       19.1.1                   py37_0    anaconda\npkginfo                   1.5.0.1                  py37_0    anaconda\npluggy                    0.12.0                     py_0    anaconda\nply                       3.11                     py37_0    anaconda\nportend                   2.5                      pypi_0    pypi\npost                      2019.4.13                pypi_0    pypi\npowershell_shortcut       0.0.1                         2    anaconda\nproglog                   0.1.9                    pypi_0    pypi\nprometheus_client         0.7.1                      py_0    anaconda\nprompt_toolkit            2.0.9                    py37_0    anaconda\nprotobuf                  3.7.1                    pypi_0    pypi\npsutil                    5.6.3            py37he774522_0    anaconda\npublic                    2019.4.13                pypi_0    pypi\npy                        1.8.0                    py37_0    anaconda\npy-lief                   0.9.0            py37ha925a31_2    anaconda\npybind11                  2.3.0                    pypi_0    pypi\npycodestyle               2.5.0                    py37_0    anaconda\npycosat                   0.6.3            py37hfa6e2cd_0    anaconda\npycparser                 2.19                     py37_0    anaconda\npycrypto                  2.6.1            py37hfa6e2cd_9    anaconda\npycryptodome              3.8.2                    pypi_0    pypi\npycurl                    7.43.0.3         py37h7a1dbc1_0    anaconda\npydot                     1.4.1                    pypi_0    pypi\npyflakes                  2.1.1                    py37_0    anaconda\npygments                  2.4.2                      py_0    anaconda\npylint                    2.3.1                    py37_0    anaconda\npyodbc                    4.0.26           py37ha925a31_0    anaconda\npyopenssl                 19.0.0                   py37_0    anaconda\npyparsing                 2.4.0                      py_0    anaconda\npyqt                      5.9.2            py37h6538335_2    anaconda\npyreadline                2.1                      py37_1    anaconda\npyrsistent                0.14.11          py37he774522_0    anaconda\npysocks                   1.7.0                    py37_0    anaconda\npytables                  3.5.2            py37h1da0976_1    anaconda\npytest                    5.0.1                    py37_0    anaconda\npytest-arraydiff          0.3              py37h39e3cac_0    anaconda\npytest-astropy            0.5.0                    py37_0    anaconda\npytest-doctestplus        0.3.0                    py37_0    anaconda\npytest-openfiles          0.3.2                    py37_0    anaconda\npytest-remotedata         0.3.1                    py37_0    anaconda\npython                    3.7.3                h8c8aaf0_1    anaconda\npython-dateutil           2.8.0                    py37_0    anaconda\npython-docx               0.8.10                   pypi_0    pypi\npython-graphviz           0.11.1                   pypi_0    pypi\npython-libarchive-c       2.8                     py37_11    anaconda\npytorch                   1.2.0               py3.7_cpu_1  [cpuonly]  pytorch\npytube                    9.5.1                    pypi_0    pypi\npytz                      2019.1                     py_0    anaconda\npywavelets                1.0.3            py37h8c2d366_1    anaconda\npywin32                   223              py37hfa6e2cd_1    anaconda\npywinpty                  0.5.5                 py37_1000    anaconda\npyyaml                    5.1.1            py37he774522_0    anaconda\npyzmq                     18.0.0           py37ha925a31_0    anaconda\nqt                        5.9.7            vc14h73c81de_0  [vc14]  anaconda\nqtawesome                 0.5.7                    py37_1    anaconda\nqtconsole                 4.5.1                      py_0    anaconda\nqtpy                      1.8.0                      py_0    anaconda\nquery-string              2019.4.13                pypi_0    pypi\nrequest                   2019.4.13                pypi_0    pypi\nrequests                  2.22.0                   py37_0    anaconda\nrope                      0.14.0                     py_0    anaconda\nruamel_yaml               0.15.46          py37hfa6e2cd_0    anaconda\nscikit-image              0.15.0           py37ha925a31_0    anaconda\nscikit-learn              0.21.2           py37h6288b17_0    anaconda\nscipy                     1.3.0                    pypi_0    pypi\nscipy-stack               0.0.5                    pypi_0    pypi\nseaborn                   0.9.0                    py37_0    anaconda\nsend2trash                1.5.0                    py37_0    anaconda\nsetuptools                41.1.0                   pypi_0    pypi\nsimplegeneric             0.8.1                    py37_2    anaconda\nsingledispatch            3.4.0.3                  py37_0    anaconda\nsip                       4.19.8           py37h6538335_0    anaconda\nsix                       1.12.0                   py37_0    anaconda\nsnappy                    1.1.7                h777316e_3    anaconda\nsnowballstemmer           1.9.0                      py_0    anaconda\nsortedcollections         1.1.2                    py37_0    anaconda\nsortedcontainers          2.1.0                    py37_0    anaconda\nsoupsieve                 1.8                      py37_0    anaconda\nsphinx                    2.1.2                      py_0    anaconda\nsphinxcontrib             1.0                      py37_1    anaconda\nsphinxcontrib-applehelp   1.0.1                      py_0    anaconda\nsphinxcontrib-devhelp     1.0.1                      py_0    anaconda\nsphinxcontrib-htmlhelp    1.0.2                      py_0    anaconda\nsphinxcontrib-jsmath      1.0.1                      py_0    anaconda\nsphinxcontrib-qthelp      1.0.2                      py_0    anaconda\nsphinxcontrib-serializinghtml 1.1.3                      py_0    anaconda\nsphinxcontrib-websupport  1.1.2                      py_0    anaconda\nspyder                    3.3.6                    py37_0    anaconda\nspyder-kernels            0.5.1                    py37_0    anaconda\nsqlalchemy                1.3.5            py37he774522_0    anaconda\nsqlite                    3.29.0               he774522_0    anaconda\nstatsmodels               0.10.0           py37h8c2d366_0    anaconda\nsumma                     1.2.0                    pypi_0    pypi\nsympy                     1.4                      py37_0    anaconda\ntbb                       2019.4               h74a9793_0    anaconda\ntblib                     1.4.0                      py_0    anaconda\ntempora                   1.14.1                   pypi_0    pypi\ntensorboard               1.14.0           py37he3c9ec2_0    anaconda\ntensorboardx              1.8                      pypi_0    pypi\ntensorflow                1.14.0          mkl_py37h7908ca0_0    anaconda\ntensorflow-base           1.14.0          mkl_py37ha978198_0    anaconda\ntensorflow-estimator      1.14.0                     py_0    anaconda\ntensorflow-mkl            1.14.0               h4fcabd2_0    anaconda\ntermcolor                 1.1.0                    pypi_0    pypi\nterminado                 0.8.2                    py37_0    anaconda\ntestpath                  0.4.2                    py37_0    anaconda\ntk                        8.6.8                hfa6e2cd_0    anaconda\ntoolz                     0.10.0                     py_0    anaconda\ntorchvision               0.4.0                  py37_cpu  [cpuonly]  pytorch\ntornado                   6.0.3            py37he774522_0    anaconda\ntqdm                      4.32.1                     py_0    anaconda\ntraitlets                 4.3.2                    py37_0    anaconda\ntyping                    3.6.6                    pypi_0    pypi\ntyping-extensions         3.6.6                    pypi_0    pypi\nunicodecsv                0.14.1                   py37_0    anaconda\nurllib3                   1.24.2                   py37_0    anaconda\nvalidators                0.13.0                   pypi_0    pypi\nvc                        14.1                 h0510ff6_4    anaconda\nvs2015_runtime            14.15.26706          h3a45250_4    anaconda\nwcwidth                   0.1.7                    py37_0    anaconda\nwebencodings              0.5.1                    py37_1    anaconda\nwerkzeug                  0.15.4                     py_0    anaconda\nwheel                     0.33.4                   py37_0    anaconda\nwidgetsnbextension        3.5.0                    py37_0    anaconda\nwin_inet_pton             1.1.0                    py37_0    anaconda\nwin_unicode_console       0.5                      py37_0    anaconda\nwincertstore              0.2                      py37_0    anaconda\nwinpty                    0.4.3                         4    anaconda\nwrapt                     1.11.2           py37he774522_0    anaconda\nxlrd                      1.2.0                    py37_0    anaconda\nxlsxwriter                1.1.8                      py_0    anaconda\nxlwings                   0.15.8                   py37_0    anaconda\nxlwt                      1.3.0                    py37_0    anaconda\nxz                        5.2.4                h2fa13f4_4    anaconda\nyaml                      0.1.7                hc54c509_2    anaconda\nyoutube-dl                2019.8.2                 pypi_0    pypi\nzc-lockfile               1.4                      pypi_0    pypi\nzeromq                    4.3.1                h33f27b4_3    anaconda\nzict                      1.0.0                      py_0    anaconda\nzipp                      0.5.1                      py_0    anaconda\nzlib                      1.2.11               h62dcd97_3    anaconda\nzstd                      1.3.7                h508b16e_0    anaconda\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57814542,"AnswerCount":2,"Tags":"<pytorch><setup.py><caffe2>","CreationDate":"2019-09-06T01:05:53.903","AcceptedAnswerId":null,"OwnerUserId":1434041.0,"Title":"Installing pytorch fails - fatal error: libavutil\/motion_vector.h: No such file or directory","Body":"<p>I'm trying to install pytorch <\/p>\n\n<pre><code>python setup.py install\n<\/code><\/pre>\n\n<p>but it fails as it can't locate <code>libavutil\/motion_vector.h<\/code><\/p>\n\n<pre><code>[ 66%] Building CXX object caffe2\/CMakeFiles\/torch.dir\/video\/video_input_op.cc.o\nIn file included from \/root\/pytorch\/caffe2\/video\/video_input_op.h:15:0,\n                 from \/root\/pytorch\/caffe2\/video\/video_input_op.cc:1:\n\/root\/pytorch\/caffe2\/video\/video_decoder.h:15:37: fatal error: libavutil\/motion_vector.h: No such file or directory\n #include &lt;libavutil\/motion_vector.h&gt;\n                                     ^\ncompilation terminated.\n[ 66%] Building CXX object caffe2\/CMakeFiles\/torch.dir\/video\/video_io.cc.o\nmake[2]: *** [caffe2\/CMakeFiles\/torch.dir\/video\/video_input_op.cc.o] Error 1\nmake[2]: *** Waiting for unfinished jobs....\nIn file included from \/root\/pytorch\/caffe2\/video\/video_io.h:6:0,\n                 from \/root\/pytorch\/caffe2\/video\/video_io.cc:1:\n\/root\/pytorch\/caffe2\/video\/video_decoder.h:15:37: fatal error: libavutil\/motion_vector.h: No such file or directory\n #include &lt;libavutil\/motion_vector.h&gt;\n                                     ^\ncompilation terminated.\nmake[2]: *** [caffe2\/CMakeFiles\/torch.dir\/video\/video_io.cc.o] Error 1\nmake[1]: *** [caffe2\/CMakeFiles\/torch.dir\/all] Error 2\nmake: *** [all] Error 2\nTraceback (most recent call last):\n  File \"setup.py\", line 759, in &lt;module&gt;\n    build_deps()\n  File \"setup.py\", line 321, in build_deps\n    cmake=cmake)\n  File \"\/root\/pytorch\/tools\/build_pytorch_libs.py\", line 63, in build_caffe2\n    cmake.build(my_env)\n  File \"\/root\/pytorch\/tools\/setup_helpers\/cmake.py\", line 330, in build\n    self.run(build_args, my_env)\n  File \"\/root\/pytorch\/tools\/setup_helpers\/cmake.py\", line 143, in run\n    check_call(command, cwd=self.build_dir, env=env)\n  File \"\/usr\/lib\/python2.7\/subprocess.py\", line 540, in check_call\n    raise CalledProcessError(retcode, cmd)\nsubprocess.CalledProcessError: Command '['cmake', '--build', '.', '--target', 'install', '--config', 'Release', '--', '-j', '12']' returned non-zero exit status 2\n<\/code><\/pre>\n\n<p>What I've tried so far:<\/p>\n\n<ul>\n<li>installed <code>libavutil-dev<\/code> <\/li>\n<li>ensured <code>ffmpeg<\/code> is installed<\/li>\n<li>followed <a href=\"https:\/\/github.com\/facebookresearch\/VMZ\/blob\/master\/tutorials\/Installation_guide.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/facebookresearch\/VMZ\/blob\/master\/tutorials\/Installation_guide.md<\/a><\/li>\n<\/ul>\n\n<p>Any other ideas?<\/p>\n","answers":[{"AnswerId":"57952085","CreationDate":"2019-09-16T07:21:40.617","ParentId":null,"OwnerUserId":"920010","Title":null,"Body":"<p>Missing headers of libraries commonly indicate missing development packages. For Debian and Ubuntu the package you are looking for is most likely called <code>libavutil-dev<\/code>. Try installing this package using you package manager and re-running the pip installation. For Fedora and SUSE the package should end with <code>-devel<\/code>.<\/p>\n"},{"AnswerId":"57959353","CreationDate":"2019-09-16T14:49:22.767","ParentId":null,"OwnerUserId":"10637327","Title":null,"Body":"<p>For conda installation, firstly create an env, very important because you've used pip (pip stops conda seeing it's base env).<\/p>\n\n<pre><code>conda create -n pytorch_env python=3.7\nsource activate pytorch_env\n\nconda install -c anaconda mkl\n<\/code><\/pre>\n\n<p>Then install pytorch and torch vision<\/p>\n\n<pre><code>conda install -c pytorch pytorch torchvision\n<\/code><\/pre>\n\n<p>You might want to run these as administrator, for me not needed.\nYou will need to install Caffe via<\/p>\n\n<pre><code>conda install -c caffe2 pytorch-nightly\n<\/code><\/pre>\n\n<hr>\n\n<p>There's a bug for caffe2 installation (not nightly). Corrected below,<\/p>\n\n<pre><code>conda create -n pytorch_env python=3.6\nsource activate pytorch_env\nconda install -c caffe2 caffe2\n<\/code><\/pre>\n\n<p>Having done this I found <code>opencv<\/code> and <code>ffmpeg<\/code> were loaded .. I can only assume they were a dependency of caffe2. Thus in summary yes you can do the whole thing easily via conda once the 3.7 -> 3.6 bug is resolved<\/p>\n"}]}
+{"QuestionId":57815001,"AnswerCount":2,"Tags":"<python-3.x><pytorch><dataloader>","CreationDate":"2019-09-06T02:28:29.383","AcceptedAnswerId":null,"OwnerUserId":6423473.0,"Title":"pytorch collate_fn reject sample and yield another","Body":"<p>I have built a Dataset, where I'm doing various checks on the images I'm loading. I'm then passing this DataSet to a DataLoader.<\/p>\n\n<p>In my DataSet class I'm returning the sample as None if a picture fails my checks and i have a custom collate_fn function which removes all Nones from the retrieved batch and returns the remaining valid samples.<\/p>\n\n<p>However at this point the returned batch can be of varying size. Is there a way to tell the collate_fn to keep sourcing data until the batch size meets a certain length?<\/p>\n\n<pre><code>class DataSet():\n     def __init__(self, example):\n          # initialise dataset\n          # load csv file and image directory\n          self.example = example\n     def __getitem__(self,idx):\n          # load one sample\n          # if image is too dark return None\n          # else \n          # return one image and its equivalent label\n\ndataset = Dataset(csv_file='..\/', image_dir='..\/..\/')\n\ndataloader = DataLoader(dataset , batch_size=4,\n                        shuffle=True, num_workers=1, collate_fn = my_collate )\n\ndef my_collate(batch): # batch size 4 [{tensor image, tensor label},{},{},{}] could return something like G = [None, {},{},{}]\n    batch = list(filter (lambda x:x is not None, batch)) # this gets rid of nones in batch. For example above it would result to G = [{},{},{}]\n    # I want len(G) = 4\n    # so how to sample another dataset entry?\n    return torch.utils.data.dataloader.default_collate(batch) \n<\/code><\/pre>\n","answers":[{"AnswerId":"57882783","CreationDate":"2019-09-11T05:41:58.640","ParentId":null,"OwnerUserId":"6423473","Title":null,"Body":"<p>There are 2 hacks that can be used to sort out the problem, choose one way:<\/p>\n\n<p>By using the original batch sample <strong>Fast option<\/strong>:<\/p>\n\n<pre><code>def my_collate(batch):\n    len_batch = len(batch) # original batch length\n    batch = list(filter (lambda x:x is not None, batch)) # filter out all the Nones\n    if len_batch &gt; len(batch): # if there are samples missing just use existing members, doesn't work if you reject every sample in a batch\n        diff = len_batch - len(batch)\n        for i in range(diff):\n            batch = batch + batch[:diff]\n    return torch.utils.data.dataloader.default_collate(batch)\n<\/code><\/pre>\n\n<p>Otherwise just load another sample from dataset at random <strong>Better option<\/strong>:<\/p>\n\n<pre><code>def my_collate(batch):\n    len_batch = len(batch) # original batch length\n    batch = list(filter (lambda x:x is not None, batch)) # filter out all the Nones\n    if len_batch &gt; len(batch): # source all the required samples from the original dataset at random\n        diff = len_batch - len(batch)\n        for i in range(diff):\n            batch.append(dataset[np.random.randint(0, len(dataset))])\n\n    return torch.utils.data.dataloader.default_collate(batch)\n<\/code><\/pre>\n"},{"AnswerId":"57815062","CreationDate":"2019-09-06T02:40:17.813","ParentId":null,"OwnerUserId":"11969685","Title":null,"Body":"<p>Unless the code you've posted is incorrect, I would start by fixing your DataSet class definition. In python, you should always define a class like this:<\/p>\n\n<pre><code>class DataSet:\n\ndef __init__(self, example):\n    self.example = example\n<\/code><\/pre>\n\n<p>Where you can then add further attributes to your class as well as functions. I would suggest learning how to use classes first, then coming back at your problem.<\/p>\n\n<p>Let me know if this helps! Cheers!<\/p>\n"}]}
+{"QuestionId":57815034,"AnswerCount":1,"Tags":"<python><tensorflow><ubuntu><matplotlib><cmd>","CreationDate":"2019-09-06T02:35:38.727","AcceptedAnswerId":null,"OwnerUserId":7927642.0,"Title":"plt.show() not working graph on ubuntu cmd promt","Body":"<p>I am trying to learn machine learning using python and trying to get graphs to show using matplotlib. However when I run code showing the graph, it doesn't do anything. I am using the ubuntu terminal within windows cmd.<\/p>\n\n<p>I am running through the tutorial within the tensorflow documentation\n<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 know that using pycharm and running the code would show it, but i prefer to run and compile everything through the windows cmd, specifically the ubuntu terminal.<\/p>\n\n<pre><code>plt.figure()\nplt.imshow(train_images[0])\nplt.colorbar()\nplt.grid(False)\nplt.show()\n<\/code><\/pre>\n\n<p>When i run the code nothing happens. I tried to print() it but it prints None. I'm assuming it has something to do with the privileges set up with the ubuntu terminal but i have no clue how to fix it. Do I just have to use pycharm or is there a way to get this working?<\/p>\n","answers":[{"AnswerId":"57820744","CreationDate":"2019-09-06T11:01:23.773","ParentId":null,"OwnerUserId":"11885270","Title":null,"Body":"<p>Install anaconda or conda and use Jupyter notebook. For purpose of learning, I see no reason to do plotting in that way.<\/p>\n"}]}
+{"QuestionId":57815105,"AnswerCount":1,"Tags":"<python><tensorflow><keras><softmax><activation-function>","CreationDate":"2019-09-06T02:48:23.900","AcceptedAnswerId":null,"OwnerUserId":8690773.0,"Title":"How to print the activation output of specified layers","Body":"<p>I am quite new to Keras and deep learning and I have been wanting to print outputs of a section of my layers (named <code>output[x]<\/code>)<\/p>\n\n<p>Down below you can find a section of the architecture. Do note that I have not provided any reproducible code.<\/p>\n\n<p>The goal is to validate the <code>val_loss<\/code> function. This concerns a categorical cross-entropy function where the formula (simplified) is stated as:<\/p>\n\n<p><code>L=\u2212y\u22c5log(y\u0302)<\/code><\/p>\n\n<p>Where L is given when I run the architecture as the model outputs the <code>L<\/code>. The <code>y<\/code> is my ground truth, and <code>y\u0302<\/code> the estimation. <\/p>\n\n<p><strong>Goal: Print the values of output1, output2, output3, output4 and output5 so I will know what my <code>y\u0302<\/code> is. With these given variables I can then validate the formula used.<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>layer_1 = Conv2D(filters[0], kernel_size[0], activation='relu', strides=strides)(inputs)\nlayer_11 = MaxPooling2D(pool_size=pool_size, strides=maxp_strides, padding='valid')(layer_1)\nlayer_2 = Conv2D(filters[1], kernel_size[1], activation='relu', strides=strides, padding='same')(layer_11)\nlayer_3 = Conv2D(filters[2], kernel_size[2], strides=strides, activation='relu', padding='same')(layer_2)\nlayer_4 = Conv2D(filters[3], kernel_size[3], strides=strides, activation='relu', padding='same')(layer_3)\nlayer_5 = Conv2D(filters[4], kernel_size[4], strides=strides, activation='relu', padding='same')(layer_4)\nlayer_6 = Flatten()(layer_5)\nlayer_7 = Dense(units[0], activation='relu')(layer_6)\nlayer_7 = Dropout(dropout)(layer_7)\nlayer_8 = Dense(units[1], activation='relu')(layer_7)\nlayer_8 = Dropout(dropout)(layer_8)\noutput1 = Dense(output_class, activation='softmax')(layer_8)\noutput2 = Dense(output_class, activation='softmax')(layer_8)\noutput3 = Dense(output_class, activation='softmax')(layer_8)\noutput4 = Dense(output_class, activation='softmax')(layer_8)\noutput5 = Dense(output_class, activation='softmax')(layer_8)\n\nrms = optimizers.RMSprop(lr=lr, rho=rho, epsilon=epsilon, decay=decay)\nreduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=factor, patience=patience, min_lr=min_lr)\nearly_stopping = EarlyStopping(monitor='val_loss', patience=ES_patience)\n\nmodel = Model(inputs=inputs, outputs=[output1, output2, output3, output4, output5])\nmodel.compile(optimizer=rms, loss='categorical_crossentropy', metrics=['categorical_crossentropy'])\nhistory = model.fit(X_train, [self.y_train[:, 0, :], self.y_train[:, 1, :], self.y_train[:, 2, :],self.y_train[:, 3, :], self.y_train[:, 4, :]],batch_size=batch_size,epochs=epoch, validation_split=val_split, verbose=verbose, callbacks=[reduce_lr])\n<\/code><\/pre>\n\n<p><strong>This is what I have tried<\/strong><\/p>\n\n<pre><code>print(\"outputs: {} - output1 {} - output2 {} - output3 {} - output4 {} - output5 {}\".format(model.output,output1,output2,output3,output4,output5))\n# outputs: [&lt;tf.Tensor 'dense_38\/Softmax:0' shape=(?, 3) dtype=float32&gt;, &lt;tf.Tensor 'dense_39\/Softmax:0' shape=(?, 3) dtype=float32&gt;, &lt;tf.Tensor 'dense_40\/Softmax:0' shape=(?, 3) dtype=float32&gt;, &lt;tf.Tensor 'dense_41\/Softmax:0' shape=(?, 3) dtype=float32&gt;, &lt;tf.Tensor 'dense_42\/Softmax:0' shape=(?, 3) dtype=float32&gt;] - output1 Tensor(\"dense_38\/Softmax:0\", shape=(?, 3), dtype=float32) - output2 Tensor(\"dense_39\/Softmax:0\", shape=(?, 3), dtype=float32) - output3 Tensor(\"dense_40\/Softmax:0\", shape=(?, 3), dtype=float32) - output4 Tensor(\"dense_41\/Softmax:0\", shape=(?, 3), dtype=float32) - output5 Tensor(\"dense_42\/Softmax:0\", shape=(?, 3), dtype=float32)\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>print(model.layers[-1].output)\n# Tensor(\"dense_42\/Softmax:0\", shape=(?, 3), dtype=float32)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57818598","CreationDate":"2019-09-06T08:41:42.450","ParentId":null,"OwnerUserId":"1849975","Title":null,"Body":"<p>The thing is, when you define your graph like that, you don't have any values inside it. <code>output1<\/code> is just a placeholder. If you want to visualize\/display\/plot anything from your graph (or even inspect your computational graph), I would suggest you take a look at <a href=\"https:\/\/www.tensorflow.org\/guide\/summaries_and_tensorboard\" rel=\"nofollow noreferrer\">TensorBoard<\/a>.<\/p>\n"}]}
+{"QuestionId":57815114,"AnswerCount":0,"Tags":"<tensorflow><tensorboard>","CreationDate":"2019-09-06T02:50:03.400","AcceptedAnswerId":null,"OwnerUserId":5046896.0,"Title":"Reading TensorBoard data returns DataLossError","Body":"<p>I defined the following function to extract data from TensorBoard data<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def extract_tensorboard_data(path, tag='Performance\/EpisodeTotalReward'):\n    data = []\n\n    for e in tf.train.summary_iterator(path):\n        for v in e.summary.value:\n            if v.tag == tag:\n                data.append(v.simple_value)\n    return data\n<\/code><\/pre>\n\n<p>After running my process for a couple of hours, I killed my process and I wanted to extract the data from the TensorBoard file.<\/p>\n\n<p>But I got the following error:<\/p>\n\n<pre><code>~\/tf_utils.py in extract_tensorboard_data(path, tag)\n     29     data = []\n     30 \n---&gt; 31     for e in tf.train.summary_iterator(path):\n     32         for v in e.summary.value:\n     33             if v.tag == tag:\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/summary\/summary_iterator.py in summary_iterator(path)\n     66   \"\"\"\n     67   # pylint: enable=line-too-long\n---&gt; 68   for r in tf_record.tf_record_iterator(path):\n     69     yield event_pb2.Event.FromString(r)\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/lib\/io\/tf_record.py in tf_record_iterator(path, options)\n    179     while True:\n    180       try:\n--&gt; 181         reader.GetNext()\n    182       except errors.OutOfRangeError:\n    183         break\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py in GetNext(self)\n    487 \n    488     def GetNext(self):\n--&gt; 489         return _pywrap_tensorflow_internal.PyRecordReader_GetNext(self)\n    490 \n    491     def record(self):\n\nDataLossError: truncated record at 1129507685\n<\/code><\/pre>\n\n<p>I check the data with <\/p>\n\n<pre><code>$ tensorboard --inspect  --logdir=10   --tag=Performance\/RewOfAg_0\n\nI0906 10:58:59.480272 MainThread program.py:189] Not bringing up TensorBoard, but inspecting event files.\nI0906 10:58:59.480272 139992933570368 program.py:189] Not bringing up TensorBoard, but inspecting event files.\n======================================================================\nProcessing event files... (this can take a few minutes)\n======================================================================\n\nFound event files in:\n10\n\nEvent statistics for tag Performance\/RewOfAg_0 in 10:\nscalars\n   first_step           1\n   last_step            53267\n   max_step             53267\n   min_step             1\n   num_steps            53267\n   outoforder_steps     []\n======================================================================\n<\/code><\/pre>\n\n<p>I think the data was not corrupted.<\/p>\n\n<p>How can I extract the data? Is there anything wrong with the code?<\/p>\n","answers":[]}
+{"QuestionId":57815454,"AnswerCount":1,"Tags":"<python><keras><lstm><recurrent-neural-network>","CreationDate":"2019-09-06T03:46:09.177","AcceptedAnswerId":null,"OwnerUserId":3665135.0,"Title":"Why does the loss of the LSTM model rise up over several epochs?","Body":"<p>I set up a simple LSTM model for regression, which is a many to many model. The loss of model declined at the start while it rose up over several epochs. Where am I wrong?<\/p>\n\n<p>I'have run the model for a epoch number ranging from 200 to 600 on 8 CPU cores, but the results remain same.<\/p>\n\n<p>Here is my code.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential()\nmodel.add(LSTM(50, activation='relu', return_sequences=True, input_shape=(n_steps, n_features)))\nmodel.add(LSTM(50, activation='relu'))\nmodel.add(Dense(output_steps))\nmodel.compile(optimizer='adam', loss='mse')\n<\/code><\/pre>\n\n<p>I expect the loss would reduce gradually with little fluctuation.<\/p>\n\n<p>However I saw the result as follows<\/p>\n\n<pre><code>7143\/7143 [==============================] - 2s 281us\/step - loss: 6595.8919\nEpoch 63\/200\n7143\/7143 [==============================] - 2s 289us\/step - loss: 6557.3760\nEpoch 64\/200\n7143\/7143 [==============================] - 2s 280us\/step - loss: 6947.0848\nEpoch 65\/200\n7143\/7143 [==============================] - 2s 282us\/step - loss: 6439.9647\nEpoch 66\/200\n7143\/7143 [==============================] - 2s 277us\/step - loss: 6583.3354\nEpoch 67\/200\n7143\/7143 [==============================] - 2s 278us\/step - loss: 6724.0296\nEpoch 68\/200\n7143\/7143 [==============================] - 2s 279us\/step - loss: 6457.0547\nEpoch 69\/200\n7143\/7143 [==============================] - 2s 278us\/step - loss: 6371.6533\nEpoch 70\/200\n7143\/7143 [==============================] - 2s 279us\/step - loss: 6644.9585\nEpoch 71\/200\n7143\/7143 [==============================] - 2s 277us\/step - loss: 6340.0420\nEpoch 72\/200\n7143\/7143 [==============================] - 2s 279us\/step - loss: 9484.5966\nEpoch 73\/200\n7143\/7143 [==============================] - 2s 277us\/step - loss: 10975.8083\nEpoch 74\/200\n7143\/7143 [==============================] - 2s 275us\/step - loss: 10174.8291\nEpoch 75\/200\n7143\/7143 [==============================] - 2s 282us\/step - loss: 9863.0310\nEpoch 76\/200\n7143\/7143 [==============================] - 2s 278us\/step - loss: 9882.6081\nEpoch 77\/200\n7143\/7143 [==============================] - 2s 280us\/step - loss: 9398.1880\n<\/code><\/pre>\n","answers":[{"AnswerId":"57817290","CreationDate":"2019-09-06T07:09:11.840","ParentId":null,"OwnerUserId":"6070667","Title":null,"Body":"<p>The learning rate that you are using can be too big for the optimization after 71st epoch. Reason for this is the learning rate can be relatively larger when compared to the error at a later number of epochs than the error at the beginning.<\/p>\n\n<p><strong>You can use a smaller learning rate or you can reduce the lr monotonically as a solution for this.<\/strong><\/p>\n\n<p>Good luck :-)<\/p>\n"}]}
+{"QuestionId":57815515,"AnswerCount":0,"Tags":"<python><hook><pytorch><torch><autograd>","CreationDate":"2019-09-06T03:54:09.993","AcceptedAnswerId":null,"OwnerUserId":9610081.0,"Title":"Update some weight with torch.no_grad and type(weight.grad) is Nonetype","Body":"<p>I\u2019m trying to freeze some weights(value = 0) from updating during backward, but I failed for several times :frowning:<\/p>\n\n<p>The method one I tried is using trying to zero out the gradient with below codes:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def hook(self, model, inputs):\n    with torch.no_grad():  \n        print('*********************************forward***********************')\n        print(model)\n        print(model.weight.data )\n        model.weight.data = model.weight.data * self.sparse_mask[self.type[model]]\n\n\ndef back_hook(self, model,grad_input,grad_output):\n    print(model)\n    print(model.weight)\n    print('im back',type(model.weight.grad))\n\n    model.weight.grad = model.weight.grad * self.sparse_mask[self.type[model]]\n\n\ndef register_hook(self,module):\n    self.handle = module.register_forward_pre_hook(self.hook)\n    self.backhandle = module.register_backward_hook(self.back_hook)\n\nfor m in model.modules():\n     if isinstance(m, nn.Conv2d):\n           sp.register_hook(m)\n\npred = model(data)\nloss = criterion(pred, target)\nlosses.update(loss.data.cpu().numpy())\noptimizer.zero_grad()\nloss.backward()\n\n<\/code><\/pre>\n\n<p>But I got this error<\/p>\n\n<blockquote>\n  <p>TypeError: mul(): argument \u2018other\u2019 (position 1) must be Tensor, not NoneType<\/p>\n<\/blockquote>\n\n<p>And my output of <code>back_hook()<\/code> is<\/p>\n\n<blockquote>\n  <p>im back NoneType<\/p>\n<\/blockquote>\n\n<p>which means my weight.grad is None<\/p>\n\n<p>I don\u2019t have any clue why this happened! what I want is freeze that weight which value = 0 and update the others non-zero weight.<\/p>\n\n<p>Anyone would like to help me solve that problem?<\/p>\n\n<p>Really appreciate!<\/p>\n","answers":[]}
+{"QuestionId":57815577,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-06T04:04:51.783","AcceptedAnswerId":null,"OwnerUserId":11855471.0,"Title":"How can I use keras.texts_to_sequences in multiple columns data frame?","Body":"<p>I have the data frame 1000*24, and each value in the column is string, and I want to use keras.texts_to_sequences of every column,\nhow can I do that?<\/p>\n\n<p>I tried to use for-loop, but its output \"List\" of each cell<\/p>\n\n<p>Keras, Python3.6<\/p>\n\n<p>My data is look like:<\/p>\n\n<p><a href=\"https:\/\/i.imgur.com\/5RS1AQp.png\" rel=\"nofollow noreferrer\">https:\/\/i.imgur.com\/5RS1AQp.png<\/a><\/p>\n\n<p>(Sorry, I can't upload the image because I have not enough reputation)<\/p>\n","answers":[]}
+{"QuestionId":57815801,"AnswerCount":1,"Tags":"<python><neural-network><pytorch>","CreationDate":"2019-09-06T04:39:32.433","AcceptedAnswerId":null,"OwnerUserId":1264768.0,"Title":"What defines legitimate input dimensions of a pytorch nn?","Body":"<p>From the tutorial \/ example github, the network to be trained on the MNIST dataset is defined as:<\/p>\n\n<pre><code>class Net(nn.Module):\n    def __init__(self):\n        super(Net, self).__init__()\n        self.conv1 = nn.Conv2d(1, 20, 5, 1)\n        self.conv2 = nn.Conv2d(20, 50, 5, 1)\n        self.fc1 = nn.Linear(4*4*50, 500)\n        self.fc2 = nn.Linear(500, 10)\n\n    def forward(self, x):\n        x = F.relu(self.conv1(x))\n        x = F.max_pool2d(x, 2, 2)\n        x = F.relu(self.conv2(x))\n        x = F.max_pool2d(x, 2, 2)\n        x = x.view(-1, 4*4*50)\n        x = F.relu(self.fc1(x))\n        x = self.fc2(x)\n        return F.log_softmax(x, dim=1)\n<\/code><\/pre>\n\n<p>The image sizes this network is trained with are 28x28 pixels.<\/p>\n\n<p>Testing images of different sizes creates errors. To be precise: I tested images of size 27x27 and 32x32 for whom it does break with an error; an input size of 29x29 does create none.<\/p>\n\n<p>Where is the size of 28x28 actually defined? What formula can I use to discern parameters for different input sizes for different tasks? Can one use images of different sizes as input? How?<\/p>\n","answers":[{"AnswerId":"57816148","CreationDate":"2019-09-06T05:22:03.770","ParentId":null,"OwnerUserId":"6210807","Title":null,"Body":"<p>You can backtrack the size of the input image from the first <em>fully connected layer<\/em> i.e. <code>fc1 = Linear(4*4*50, 500)<\/code>. The input to the fc1 is 50x4x4 (CxHxW) (here 50 is the channel dimension as evident from the previous conv2 layer). Thus, the output of conv2 (before max-pooling operation) is 50x8x8 as you're performing pooling using 2x2 - <code>max_pool2d(x, 2, 2)<\/code>.<\/p>\n\n<pre><code>torch.nn.Conv2d(in_channels, out_channels, kernel_size, stride=1, padding=0, ...)\n<\/code><\/pre>\n\n<p>Now, you can get the input size of image before the convolution operation using the formula <code>(W-F+2P)\/S + 1 = output size<\/code>. Here, <code>W<\/code> is input size, <code>F<\/code> is filter\/kernel size, and <code>P<\/code> is padding used, and <code>S<\/code> is the stride. Thus, (W-5+2*0)\/1+1=8 => W=12. <\/p>\n\n<p>Hence, the input to conv2 is 20x12x12.<\/p>\n\n<p>In the same way, we can continue the process as follows:<\/p>\n\n<p>Output of conv1 (i.e. before max-pooling): 20x24x24<br>\nInput to conv1: 1x28x28. ((W-5+2*0)\/1+1=24 => W=28)<\/p>\n\n<p>Hence, the input image size is 1x28x28.<\/p>\n\n<hr>\n\n<p>The error is because the fully connected layers expect input of fixed size and that defines your network. In order to pass variable sized input, you may use need to transform the input to size that your network (fc layers) expects using <a href=\"https:\/\/pytorch.org\/docs\/stable\/torchvision\/transforms.html\" rel=\"nofollow noreferrer\">transformations<\/a> such as cropping.<\/p>\n\n<p>Also, there are networks that can take variable size input e.g. Fully Convolutional Networks (FCN), which doesn't contain fc layers, but only conv layers. You may also read about Spatial Pyramid Pooling (used in a network called DeepLab for semantic segmentation). <a href=\"https:\/\/stats.stackexchange.com\/q\/388859\/194589\">Read more here<\/a>.<\/p>\n"}]}
+{"QuestionId":57816021,"AnswerCount":2,"Tags":"<python-3.x><tensorflow><tf.keras>","CreationDate":"2019-09-06T05:04:42.877","AcceptedAnswerId":null,"OwnerUserId":11227488.0,"Title":"Error when importing 'keras' from 'tensorflow'","Body":"<p>I'm trying to import keras using TensorFlow 2.0 RC and using Python 3.6.3.<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras\n<\/code><\/pre>\n\n<p>I'm getting an error when using the SPYDER IDE: <\/p>\n\n<blockquote>\n  <p>ImportError: cannot import name 'keras' from 'tensorflow'<\/p>\n<\/blockquote>\n\n<p>I've tried searching on google for an answer but no avail. Any help would be appreciated :) !<\/p>\n","answers":[{"AnswerId":"57920104","CreationDate":"2019-09-13T08:39:42.223","ParentId":null,"OwnerUserId":"11759240","Title":null,"Body":"<p>From Keras repo.:<\/p>\n\n<p>Keras is a high-level neural networks API, written in Python and capable of running on top of TensorFlow, CNTK, or Theano.Before installing Keras, please install one of its backend engines: TensorFlow, Theano, or CNTK. We recommend the TensorFlow backend.<\/p>\n\n<p>So Keras is a skin (an API). TensorFlow has decided to include this skin inside itself as tf.keras. Since Keras provides APIs that TensorFlow has already implemented (unless CNTK and Theano overtake TensorFlow which is unlikely).<\/p>\n\n<p>So, we can use simply install keras by using pip \n<code>pip install keras<\/code><\/p>\n"},{"AnswerId":"57816106","CreationDate":"2019-09-06T05:16:23.627","ParentId":null,"OwnerUserId":"3796272","Title":null,"Body":"<p><code>keras<\/code>is an additional package which uses <code>tensorflow<\/code> (or <code>tensorflow-gpu<\/code>) as backend. You have to install <code>keras<\/code> too<\/p>\n\n<pre><code>python -m pip install keras\n<\/code><\/pre>\n\n<p>Now you can use it<\/p>\n\n<pre><code>&gt;&gt;&gt; import keras\nUsing TensorFlow backend.\n&gt;&gt;&gt; print(keras.__version__)\n2.2.4\n&gt;&gt;&gt;\n<\/code><\/pre>\n\n<p>It seems to be kinda new that keras is delivered with <code>tensorflow<\/code> (or my information are old...). Anyway I tested your way and it works for me:<\/p>\n\n<pre><code>python -m pip install tensorflow==2.0RC\n\nPython 3.6.8 (tags\/v3.6.8:3c6b436a57, Dec 24 2018, 00:16:47) [MSC v.1916 64 bit (AMD64)] on win32\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n&gt;&gt;&gt; import tensorflow\n&gt;&gt;&gt; from tensorflow import keras\n&gt;&gt;&gt; print(keras.__version__)\n2.2.4-tf\n&gt;&gt;&gt; print(tensorflow.__version__)\n2.0.0-rc0\n<\/code><\/pre>\n\n<p>Have you tried try to upgrade \/ reinstall your packages?<\/p>\n"}]}
+{"QuestionId":57816305,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2019-09-06T05:41:39.067","AcceptedAnswerId":null,"OwnerUserId":3337089.0,"Title":"How to restore tensorflow v1.1.0 saved model in v1.13.1","Body":"<p>I'm trying to restore the pretrained model provided <a href=\"https:\/\/github.com\/rubenvillegas\/iclr2017mcnet\" rel=\"nofollow noreferrer\">here<\/a> and continue training on a different dataset. The pretrained models available there are trained on <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/linux\/gpu\/tensorflow_gpu-1.1.0-cp27-none-linux_x86_64.whl\" rel=\"nofollow noreferrer\">tensorflow_gpu-1.1.0<\/a>. But I have <code>tensorflow_gpu-1.13.1<\/code>. When I try restoring the model, I get the below error.<\/p>\n\n<pre><code>NotFoundError (see above for traceback): 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.\n<\/code><\/pre>\n\n<p>Is it possible to convert the model to current tensorflow version? I tried a script provided <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/rnn\/python\/tools\/checkpoint_convert.py\" rel=\"nofollow noreferrer\">here<\/a>, but no luck!<\/p>\n\n<p>If not possible to convert, I'm okay to use older tensorflow version as well. But I'm not able to install properly the old version as well.\nThe command provided in the <a href=\"https:\/\/github.com\/rubenvillegas\/iclr2017mcnet\" rel=\"nofollow noreferrer\">github page<\/a> is below<\/p>\n\n<pre><code>pip install --ignore-installed --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/gpu\/tensorflow_gpu-1.1.0-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n\n<p>But when I install tensorflow using the above command, I get the below error<\/p>\n\n<pre><code>Python 2.7.16 |Anaconda, Inc.| (default, Aug 22 2019, 16:00:36) \n[GCC 7.3.0] on linux2\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 \"\/media\/nagabhushan\/Data02\/SoftwareFiles\/Anaconda\/anaconda3\/envs\/MCnet3\/lib\/python2.7\/site-packages\/tensorflow\/__init__.py\", line 24, in &lt;module&gt;\n    from tensorflow.python import *\n  File \"\/media\/nagabhushan\/Data02\/SoftwareFiles\/Anaconda\/anaconda3\/envs\/MCnet3\/lib\/python2.7\/site-packages\/tensorflow\/python\/__init__.py\", line 51, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"\/media\/nagabhushan\/Data02\/SoftwareFiles\/Anaconda\/anaconda3\/envs\/MCnet3\/lib\/python2.7\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 52, in &lt;module&gt;\n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"\/media\/nagabhushan\/Data02\/SoftwareFiles\/Anaconda\/anaconda3\/envs\/MCnet3\/lib\/python2.7\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 41, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"\/media\/nagabhushan\/Data02\/SoftwareFiles\/Anaconda\/anaconda3\/envs\/MCnet3\/lib\/python2.7\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"\/media\/nagabhushan\/Data02\/SoftwareFiles\/Anaconda\/anaconda3\/envs\/MCnet3\/lib\/python2.7\/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)\nImportError: libcublas.so.8.0: 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>If I install <code>tensorflow-1.1.0<\/code> using <code>conda<\/code>, import works, but restore model fails again with the same error<\/p>\n\n<pre><code>NotFoundError (see above for traceback): 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.\n<\/code><\/pre>\n\n<p>Kindly help!<\/p>\n","answers":[{"AnswerId":"57818431","CreationDate":"2019-09-06T08:31:34.227","ParentId":null,"OwnerUserId":"1849975","Title":null,"Body":"<p>I am really not sure if it is possible to port a model but I will try to share how I solved this issue. <\/p>\n\n<p>First off, you should be able to create whole graph independent of the <code>TensorFlow<\/code> version. If any error occurs there it should be a minimal one. Then, you can simply copy all variables from your old model to the new one with:<\/p>\n\n<pre><code>RESTORE_VARS_BLACKLIST = ['dont', 'load', 'this']\nckpt_vars = tf.train.list_variables(RESTORE_VARS_CKPT)\nass_ops = []\nfor dst_var in tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES):\n    for (ckpt_var, ckpt_shape) in ckpt_vars:\n        if dst_var.name.split(\":\")[0] == ckpt_var and dst_var.shape == ckpt_shape and ckpt_var not in RESTORE_VARS_BLACKLIST:\n            value = tf.train.load_variable(RESTORE_VARS_CKPT, ckpt_var)\n            ass_ops.append(tf.assign(dst_var, value))\n# Run assign in a session\nsess.run(ass_ops)\n<\/code><\/pre>\n\n<p>At the end, just save your new model.<\/p>\n"}]}
+{"QuestionId":57816915,"AnswerCount":1,"Tags":"<tensorflow><keras><neural-network><deep-learning><tensor>","CreationDate":"2019-09-06T06:39:34.830","AcceptedAnswerId":null,"OwnerUserId":9080239.0,"Title":"How to use tensor cores instead of cuda cores in my code?","Body":"<p>I have RTX2070 Nvidia graphic card which has tensor cores on it. I want to run my deep learning code utilizing tensor cores instead of Cuda cores. Is it possible in this graphic card? and isn't there any specific driver that I should install in order to do that?\nand how should I check to know whether the model is running on tensor cores or Cuda cores?\nI am using Keras framework on windows 10.<\/p>\n","answers":[{"AnswerId":"57818463","CreationDate":"2019-09-06T08:33:56.487","ParentId":null,"OwnerUserId":"1849975","Title":null,"Body":"<p>According to <a href=\"https:\/\/docs.nvidia.com\/deeplearning\/frameworks\/tensorflow-user-guide\/index.html#tensorcore\" rel=\"nofollow noreferrer\">NVIDIA<\/a>:<\/p>\n\n<blockquote>\n  <p>The container enables Tensor Core math by default<\/p>\n<\/blockquote>\n\n<p>If you want to disable it you can set <code>TF_DISABLE_CUDNN_TENSOR_OP_MATH<\/code> to <code>1<\/code>.<\/p>\n"}]}
+{"QuestionId":57817687,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><keras><prediction>","CreationDate":"2019-09-06T07:38:30.380","AcceptedAnswerId":null,"OwnerUserId":10376805.0,"Title":"Does the model.predict() function create batches in the same order as input data?","Body":"<p>When using <code>model.predict(array, batch_size=20)<\/code> in Keras, can I expect the output predictions to be in the same order as the inputs from the 'array'?<\/p>\n","answers":[]}
+{"QuestionId":57817884,"AnswerCount":0,"Tags":"<python><flutter><keras><output><conv-neural-network>","CreationDate":"2019-09-06T07:54:21.263","AcceptedAnswerId":null,"OwnerUserId":11919769.0,"Title":"How do I change the prediction format of the NN into a dict inside a list","Body":"<p>I am building an application which uses ML image classification, whenever I add another model which is custom made by me the app crashes at the time of output. But whenever I run it using the existing models like SSDmobilenet it works accurately. How do i format my output from <code>[output]<\/code> to <code>[{output}]<\/code> so that it can be used. <\/p>\n\n<p>I have tried building the model and saving it into model.json format, it didn't work. This is the format of the result I have while running the test from python\n<code>[[1.]]<\/code> when classified as true and <code>[[.0]]<\/code> when classified as false<\/p>\n\n<p>Image when using SSDMobilenet\n<a href=\"https:\/\/i.stack.imgur.com\/DvEeH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DvEeH.png\" alt=\"Image while predicting in the application using SSDMobilenet\"><\/a>\nImage while using the custom tflite model in python\n<a href=\"https:\/\/i.stack.imgur.com\/atUxx.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/atUxx.png\" alt=\"Image when predicted with python\"><\/a><\/p>\n\n<p>Tflite plugin uses shape like this. Is there any fix?\n<a href=\"https:\/\/i.stack.imgur.com\/muM3w.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/muM3w.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57818023,"AnswerCount":2,"Tags":"<python><opencv><tensorflow><deep-learning><training-data>","CreationDate":"2019-09-06T08:04:48.510","AcceptedAnswerId":57818250.0,"OwnerUserId":930122.0,"Title":"Tensorflow trained model speed","Body":"<p>I'm a Tensorflow newby and I'm trying to train a 1 class model for object detection. In particular I'm trying to recognize an arrow like the following:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ZWpum.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ZWpum.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I need a very fast recognition so I started wondering if a pre-trained model can contain such kind of shape.\nUnfortunately didn't find anything similar and therefor I started with my own training of the arrow using as model the faster_rcnn_inception_v2_coco_2018_01_28.<\/p>\n\n<p>I'm using his pipeline config, and I'm using his fine_tune_checkpoint as well, is this right considering that I have to train a completely different object?<\/p>\n\n<p>The result is a training with a very good accuracy but very low speed. I need to increase the framerate and I didn't understand yet if the less is the \"training loss\" the more is the \"object recognition speed\", or not.<\/p>\n\n<p>Any suggestion on how could I speedup the detection?<\/p>\n","answers":[{"AnswerId":"57818250","CreationDate":"2019-09-06T08:19:37.357","ParentId":null,"OwnerUserId":"1849975","Title":null,"Body":"<blockquote>\n  <p>I'm using his pipeline config, and I'm using his fine_tune_checkpoint\n  as well, is this right considering that I have to train a completely\n  different object?<\/p>\n<\/blockquote>\n\n<p>Yes! Every time you want to change the output of a deep NN, you should take a pretrained model. Training a model from scratch can take several weeks and you will never be able to generate enough data on your own. Taking a pretrained model and fine-tuning it is a way to go.<\/p>\n\n<blockquote>\n  <p>I didn't understand yet if the\n  less is the \"training loss\" the more is the \"object recognition\n  speed\", or not.<\/p>\n<\/blockquote>\n\n<p>No. Training loss just tells you how good your model performs with respect to the training set.<\/p>\n\n<p>The issue you are having is a classic speed vs. accuracy trade-off. I encourage you to take a look at <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/detection_model_zoo.md\" rel=\"nofollow noreferrer\">this table<\/a> and find a model which is fast enough for you (i.e. lowest run-time) but have decent accuracy. I would first check <code>SSD<\/code> here.<\/p>\n"},{"AnswerId":"57818886","CreationDate":"2019-09-06T09:01:19.343","ParentId":null,"OwnerUserId":"3125902","Title":null,"Body":"<blockquote>\n  <p>The result is a training with a very good accuracy but very low speed.<\/p>\n<\/blockquote>\n\n<p>How much FPS does your algorithm perform?\nSince you already have prepared dataset, I would suggest using Tiny-Yolo which performs 244 FPS on COCO dataset <a href=\"https:\/\/pjreddie.com\/darknet\/yolo\/\" rel=\"nofollow noreferrer\">https:\/\/pjreddie.com\/darknet\/yolo\/<\/a><\/p>\n\n<p>Preparing training dataset for Tiny-Yolo is very easy if you use this <a href=\"https:\/\/github.com\/AlexeyAB\/darknet#how-to-train-tiny-yolo-to-detect-your-custom-objects\" rel=\"nofollow noreferrer\">repository<\/a>  <\/p>\n\n<p>And<\/p>\n\n<blockquote>\n  <p>I didn't understand yet if the less is the \"training loss\" the more is the \"object recognition speed\"<\/p>\n<\/blockquote>\n\n<p>Training lost has nothing to do with speed. <\/p>\n"}]}
+{"QuestionId":57819034,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-09-06T09:11:21.873","AcceptedAnswerId":null,"OwnerUserId":8186207.0,"Title":"Why do Keras evaluate_generator and predict_generator yield different accuracy scores?","Body":"<p>I'm trying to train a conv net in Keras using generator. However, when I use the evaluate_generator and predict_generator, they yield different accuracy scores (simple accuracy, (TP + TN) \/ (TP + FP + TN + FN)).<\/p>\n\n<p>I saw some posts saying that it was about multi-processing and set workers to 1 and pickle_safe to True would do, but it didn't. Here is my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>batch_size = 16\nclasses_list = [str(i) for i in range(num_classes)]\n\ndata_flow_train_param = {\n    \"dataframe\": data_train,\n    \"x_col\": \"path\",\n    \"y_col\": \"diagnosis\",\n    \"classes\": classes_list,\n    \"batch_size\": batch_size,\n    \"target_size\": IMAGE_SIZE,\n    \"shuffle\": True\n}\ndata_flow_train = image_generator.flow_from_dataframe(**data_flow_train_param)\n\ndata_flow_val_param = {\n    \"dataframe\": data_val,\n    \"x_col\": \"path\",\n    \"y_col\": \"diagnosis\",\n    \"classes\": classes_list,\n    \"batch_size\": 1,\n    \"target_size\": IMAGE_SIZE,\n    \"shuffle\": False\n}\ndata_flow_val = image_generator.flow_from_dataframe(**data_flow_val_param)\n\nfit_params = {\n    \"generator\": data_flow_train,\n    \"steps_per_epoch\": data_train.shape[0] \/\/ batch_size,\n    \"epochs\": 10,\n    \"verbose\": 1,\n    \"validation_data\": data_flow_val,\n    \"validation_steps\": data_val.shape[0],\n    \"callbacks\": [checkpoint, lr_decay, early_stopping]\n}\n\nprint(\"Training the model...\")\nhistory = model.fit_generator(**fit_params)\nprint(\"Done!\")\n\nmodel.evaluate_generator(data_flow_val, steps=data_val.shape[0], verbose=1, pickle_safe=True)\n\ny_val_true = np.array(data_val[\"diagnosis\"]).astype(\"int\")\ny_val_pred = np.argmax(model.predict_generator(data_flow_val, steps=data_val.shape[0], verbose=1, pickle_safe=True), axis=-1)\nprint(accuracy_score(y_val_true, y_val_pred))\n\n<\/code><\/pre>\n\n<p>The evaluate_generator function gives me accuracy around 80%, while the predict_generator only has 30%.<\/p>\n","answers":[]}
+{"QuestionId":57819299,"AnswerCount":2,"Tags":"<python><keras>","CreationDate":"2019-09-06T09:28:21.767","AcceptedAnswerId":null,"OwnerUserId":11973552.0,"Title":"relation between batch_size and running time","Body":"<p>previously I thought that smaller batch_size would lead to faster training, but in practice in keras, I am receiving the opposite results which is that bigger batch_size make training faster.<\/p>\n\n<p>I am implementing a sample code, and by increasing the amount of batch_size, the training become faster. that is the opposite of my previously common believe (that smaller batch_size results in faster training),\nhere's the sample code:<\/p>\n\n<pre><code># fit model\nimport time\nstart = time.time()\nhistory = model.fit(trainX, trainy, validation_data=(testX, testy), epochs=1000,\nbatch_size= 500 , verbose=0)\nend = time.time()\n\nelapsed = end - start\nprint(elapsed)\n<\/code><\/pre>\n\n<p>I put 500, 250, 50 and 10 as batch_size respectively, i expect the lower batch_size would have faster training, but batch_size 500 results in 6.3 sec,\n250 results in 6.7sec, 50 results in 28.0sec, and 10 results in 140.2secc !!!<\/p>\n","answers":[{"AnswerId":"57819501","CreationDate":"2019-09-06T09:41:14.523","ParentId":null,"OwnerUserId":"10371987","Title":null,"Body":"<p>For hardware, GPUs are very good at parallelising the calculations, specifically, the matrix operations happens in forward-and-backward-propagation. The same thing happens in software side, tensorflow and other DL libraies optimise matrix operations.   <\/p>\n\n<p>Therefore, lager batch size makes the GPU and DL libraries able to \"optimise more matrix computation\", which leads to faster training time.<\/p>\n"},{"AnswerId":"57819471","CreationDate":"2019-09-06T09:39:05.820","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>This makes sense. I do not know what model you are using but Keras is highly optimized making use of vectorization for fast matrix operations. So if you split your data of 5000 samples into batch sizes of 500, with 1000 epochs, essentially there are (5000\/500) x 1000 iterations through the model. That's 10 000\nNow if you do that for a batch size of 10 there are (5000\/10) x 1000 iterations. That's 500 000. Many more iterations through the model, both forward and backwards.<\/p>\n\n<p>Hope that helps.<\/p>\n"}]}
+{"QuestionId":57819411,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-06T09:35:30.460","AcceptedAnswerId":null,"OwnerUserId":9454845.0,"Title":"The tensorflow diagram model loads","Body":"<p>Model loading failure\nI downloaded the <strong>vgg16_weights.npz<\/strong> pretraining model, but it didn't load<\/p>\n\n<pre><code>    args = parser.parse_args()\n\n    vgg_file = open(args.model_path, 'rb')\n\n    vgg_model = vgg_file.read()\n    vgg_file.close()\n\n    graph_def = tf.GraphDef()\n    graph_def.ParseFromString(vgg_model)\n<\/code><\/pre>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/Users\/weiyangbin\/ChineseVQA\/extract_fc7.py\", line 114, in &lt;module&gt;\n    main()\n  File \"\/Users\/weiyangbin\/ChineseVQA\/extract_fc7.py\", line 39, in main\n    graph_def.ParseFromString(vgg_model)\ngoogle.protobuf.message.DecodeError: Error parsing message\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57819459,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-06T09:38:17.117","AcceptedAnswerId":null,"OwnerUserId":3584765.0,"Title":"Difference between tf.keras.backend.eval and tf.keras.backend.get_value","Body":"<p>\u0391t some point I needed to get the value of a Tensor in keras\/tensorflow. So, I run into 2 possible solutions here:<\/p>\n\n<ul>\n<li><code>tf.eval()<\/code><\/li>\n<li><code>tf.get_value()<\/code><\/li>\n<\/ul>\n\n<p>The documentation didn't disappoint me this time also:<br>\nFor <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/backend\/eval\" rel=\"nofollow noreferrer\">eval<\/a>: <\/p>\n\n<blockquote>\n  <p>Evaluates the value of a variable.<\/p>\n<\/blockquote>\n\n<p>For <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/backend\/get_value\" rel=\"nofollow noreferrer\">get_value<\/a>:<\/p>\n\n<blockquote>\n  <p>Returns the value of a variable.<\/p>\n<\/blockquote>\n\n<p>We are playing with English definition here I think.  <\/p>\n\n<p>The code also is not really helpful:<br>\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.14\/tensorflow\/python\/keras\/backend.py#L1189-L1208\" rel=\"nofollow noreferrer\">eval<\/a> vs <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.14\/tensorflow\/python\/keras\/backend.py#L2965-L2989\" rel=\"nofollow noreferrer\">get_value<\/a>.<\/p>\n\n<p>Basically one seems to call the other at some point (!).<\/p>\n\n<pre><code>def eval(x):\n  return get_value(to_dense(x))\n<\/code><\/pre>\n\n<p>The other is heavier:<\/p>\n\n<pre><code>def get_value(x):\n  if not tensor_util.is_tensor(x):\n    return x\n  if context.executing_eagerly():\n    return x.numpy()\n  if not getattr(x, '_in_graph_mode', True):\n    with context.eager_mode():\n      return x.numpy()\n\n  if ops.executing_eagerly_outside_functions():\n    return function([], x)(x)\n\n  return x.eval(session=get_session((x,)))\n<\/code><\/pre>\n\n<p>Anyway, does anyone knows what's the difference between them and when to use either of them?<\/p>\n","answers":[]}
+{"QuestionId":57819633,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-06T09:49:09.607","AcceptedAnswerId":57819701.0,"OwnerUserId":2241766.0,"Title":"Extract rows and columns of a tensor using indices","Body":"<p>I have a 2-D tensor <code>x<\/code>, which is a placeholder <code>x = tf.placeholder(tf.int32, shape=[None, 100])<\/code>.\nNow in the graph, I want to extract randomly <code>50%<\/code> rows of x.\nHere is what I tried:<\/p>\n\n<pre><code>idxs = tf.range(tf.shape(x)[0])\nridxs = tf.random_shuffle(idxs)[: int(tf.shape(x)[0] * 0.5)]\nx = tf.gather(x, ridxs, axis=0)\n<\/code><\/pre>\n\n<p>It throws the following error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"main_gcn_random_sampling.py\", line 76, in &lt;module&gt;\n    model = GCN_RANDOM_SAMPLING(FLAGS.learning_rate, num_input, num_classes, hidden_dimensions=hidden_sizes, act=tf.nn.relu)\n  File \"\/home\/tiendh\/Projects\/EPooling\/gcn_random_sampling.py\", line 63, in __init__\n    ridxs = tf.random_shuffle(idxs)[: int(tf.shape(x)[0] * 0.5)]\nTypeError: unsupported operand type(s) for *: 'Tensor' and 'float'\n<\/code><\/pre>\n\n<p>I could declare another placeholder for predefined indices, but it would look ugly. Any other suggestion?<\/p>\n","answers":[{"AnswerId":"57819701","CreationDate":"2019-09-06T09:53:00.500","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>That's almost right, but:<\/p>\n\n<ol>\n<li>You have to cast to <code>tf.shape(x)[0]<\/code> to float before multiplying by <code>0.5<\/code>.<\/li>\n<li>You cannot use <code>int<\/code> to cast a tensor data type, you need to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/cast\" rel=\"nofollow noreferrer\"><code>tf.cast<\/code><\/a>.<\/li>\n<\/ol>\n\n<pre class=\"lang-py prettyprint-override\"><code>idxs = tf.range(tf.shape(x)[0])\nnum_half_idx = tf.cast(tf.cast(tf.shape(x)[0], tf.float32) * 0.5, tf.int32)\nridxs = tf.random_shuffle(idxs)[: num_half_idx]\nx = tf.gather(x, ridxs, axis=0)\n<\/code><\/pre>\n\n<p>Note that, if you are always getting a 50% of the rows, you can also simply do:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>idxs = tf.range(tf.shape(x)[0])\nridxs = tf.random_shuffle(idxs)[: tf.shape(x)[0] \/\/ 2]\nx = tf.gather(x, ridxs, axis=0)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57820459,"AnswerCount":0,"Tags":"<django><tensorflow><nginx><uwsgi>","CreationDate":"2019-09-06T10:39:46.463","AcceptedAnswerId":null,"OwnerUserId":4022109.0,"Title":"Django How to preload Fat Class","Body":"<p>How can I use Django preload this fat class?<\/p>\n\n<p>because when requests come in My code will do like:<\/p>\n\n<pre><code>MyFatClass()\n<\/code><\/pre>\n\n<p>then it's very slow to loading this class<\/p>\n\n<p>have any idea to preload this class in global and very django requests comes will take already preload's fat class?    <\/p>\n\n<pre><code>class Singleton(type):\n    _instances = {}\n    def __call__(cls, *args, **kwargs):\n         if cls not in cls._instances:cls._instances[cls] = super(Singleton, cls).__call__(*args, **kwargs)\n         return cls._instances[cls]\n\nclass MyFatClass(metaclass=Singleton):\n      # something about tensorflow\n<\/code><\/pre>\n\n<hr>\n\n<p>Now,  I put it under the \"manage.py\"<\/p>\n\n<pre><code>if __name__ == '__main__':\n    main()\n    MyFatClass()\n<\/code><\/pre>\n\n<p>then I works.<\/p>\n\n<p>but I think that is a little bit hackish<\/p>\n","answers":[]}
+{"QuestionId":57820829,"AnswerCount":0,"Tags":"<python><keras><embedding>","CreationDate":"2019-09-06T11:06:56.397","AcceptedAnswerId":null,"OwnerUserId":12030144.0,"Title":"How to handle a changing input shape on an embedding layer","Body":"<p>I am trying out a recommendation problem with keras\nmy model is as follows<\/p>\n\n<pre><code>def build_model(n_users, n_movies, n_factors, min_rating, max_rating):\n   user = Input(shape=(1,))\n   u = EmbeddingLayer(n_users, n_factors)(user)\n   movie = Input(shape=(1,))\n   m = EmbeddingLayer(n_movies, n_factors)(movie)\n   x = Concatenate()([u, m])\n   x = Dropout(0.05)(x)\n   x = Dense(10, kernel_initializer='he_normal')(x)\n   x = Activation('relu')(x)\n   x = Dropout(0.5)(x)\n   x = Dense(1, kernel_initializer='he_normal')(x)\n   x = Activation('sigmoid')(x)\n   x = Lambda(lambda x: x * (max_rating - min_rating) + min_rating)(x)\n   model = Model(inputs=[user, movie], outputs=x)\n   opt = Adam(lr=0.001)\n   model.compile(loss='mean_squared_error', optimizer=opt)\n   return model\n<\/code><\/pre>\n\n<p>Is there anyway to retrain the model when the number of users change over time with out starting from scratch as the embedding layer would not work if my understanding of embedding is right. I would have to retrain the model from scratch when the number of users or movies would change, is there any way to preserve the weights learned on the dense layers.<\/p>\n","answers":[]}
+{"QuestionId":57821315,"AnswerCount":0,"Tags":"<python><keras><scikit-learn>","CreationDate":"2019-09-06T11:39:00.567","AcceptedAnswerId":null,"OwnerUserId":142605.0,"Title":"KerasClassifier with multiple outputs","Body":"<p>I have biLSTM neural network with multiple outputs. Each output is a vector of a different size.<\/p>\n\n<p>I have wrapped the network with <a href=\"https:\/\/keras.io\/scikit-learn-api\/\" rel=\"nofollow noreferrer\"><code>KerasClassifier<\/code><\/a> to be able to use it with scikit-learn (e.g. for grid search). Unfortunately, it fails when trying to fit it with:<\/p>\n\n<pre><code>ValueError: could not broadcast input array from shape (100,10,3) into shape (100,10)\n<\/code><\/pre>\n\n<p>Here's minimal example to reproduce it:<\/p>\n\n<pre><code>from keras.wrappers.scikit_learn import KerasClassifier\nimport numpy as np\n\ndef create_model():\n    return None\n\nclf = KerasClassifier(build_fn=create_model)\n\nsamples_count = 100\nsample_size = 10\noutputs_sizes = [3,4,1]\n\ny = [np.zeros((samples_count, sample_size, output_size)) \n     for output_size \n     in outputs_sizes]\n\nX = np.zeros((samples_count, 100))\n\nclf.fit(X, y)\n<\/code><\/pre>\n\n<p>How can use <code>KerasClassifier<\/code> with multiple outputs?<\/p>\n","answers":[]}
+{"QuestionId":57821443,"AnswerCount":0,"Tags":"<python><keras><neural-network><deep-learning><conv-neural-network>","CreationDate":"2019-09-06T11:48:19.700","AcceptedAnswerId":null,"OwnerUserId":11795744.0,"Title":"Applying triplet loss for recognition problem","Body":"<p>I want to apply triplet loss for some recognition problem.I am doing it by making triplets and storing them in 5D array <strong>X = [total_samples, triplet_marking, channels, height, width]<\/strong>. Although there is no need of ground truth labels here since triplet_markings (0 : anchor, 1 : positive and 2 : negative) are sufficient but to fit into keras I need them. Ground truth labels are defined in 2D array <strong>Y = [total_samples, triplet_marking]<\/strong>.<\/p>\n\n<p>Firstly I have defined my basic neural network with 16 elements FC layer at last using a function. Then I give the triplets of anchor-pos-neg as inputs. Then defined the triplet loss as defined in Andrew Ng's course on coursera. Then subsequently defined, compiled and trained the model.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def network(input_shape):\n    seq = models.Sequential()\n    seq.add(layers.Conv2D(8, (3,3), (1,1), data_format=\"channels_first\", activation=\"relu\", kernel_initializer=\"glorot_uniform\"))\n    seq.add(layers.Conv2D(8, (3,3), (1,1), data_format=\"channels_first\", activation=\"relu\", kernel_initializer=\"glorot_uniform\"))\n    seq.add(layers.MaxPooling2D(pool_size=(2, 2), data_format=\"channels_first\"))\n    seq.add(Flatten())\n    seq.add(Dense(16, activation='relu'))\n    seq.add(BatchNormalization())\n    return seq\n\nimg_anc = layers.Input(shape=(3,384,384))\nimg_pos = layers.Input(shape=(3,384,384))\nimg_neg = layers.Input(shape=(3,384,384))\n\nnet = network((3,384,384))\nfeature_anc = base_network(img_anc)\nfeature_pos = base_network(img_pos)\nfeature_neg = base_network(img_neg)\n\ndef triplet_loss(y_true, y_pred):\n    a = y_pred[:,0]\n    p = y_pred[:,1]\n    n = y_pred[:,2]\n\n    margin = 1\n\n    pos_dist = np.sum(np.square(np.subtract(a,p)))\n    neg_dist = np.sum(np.square(np.subtract(a,n)))\n    basic_loss = np.subtract(pos_dist,neg_dist) + margin\n    loss = np.max(basic_loss,0)\n\n    return loss  \n\n\nmodel_train = models.Model(input=[img_anc, img_pos, img_neg], output=[feature_anc, feature_pos, feature_neg])\n\nmodel_train.compile(loss=triplet_loss, optimizer='adam')\n\nimg_a = x_train[:, 0] #using triplet_marking\nimg_p = x_train[:, 1]\nimg_n = x_train[:, 2]\n\n# [img_a, img_p, img_c] vector is my input\n# [y_train[:,0], y_train[:,1], y_train[:,2]] vector is my ground truth labels\n# loss-func is not taking labels in account, I'm giving it just for the keras to work \n\nhistory = model_train.fit([img_a, img_p, img_n], [y_train[:,0], y_train[:,1], y_train[:,2]], batch_size=16, epochs= 100, verbose=2, validation_split=.25, shuffle=True)\n<\/code><\/pre>\n\n<p>As I am giving 3 images as inputs, I am hoping y_pred to be of shape = <strong>(batch_size,3,16)<\/strong> where y_pred[:,0] will be anchor features, y_pred[:,1] will be pos features and y_pred[:,2] will be neg features, but it has shape of <strong>(batch_size,16)<\/strong>. Maybe that's why I am not getting correct results.\nCorrect me if I have done something wrong anywhere.<\/p>\n","answers":[]}
+{"QuestionId":57821681,"AnswerCount":1,"Tags":"<java><concurrency><guava><caffeine>","CreationDate":"2019-09-06T12:03:25.303","AcceptedAnswerId":57825130.0,"OwnerUserId":3231055.0,"Title":"ConcurrentHashMap compute in guava\/caffeine\/? cache","Body":"<p>Since Java 8 we can use .compute* methods on ConcurrentHashMap to synchronize processing by key, so that if two threads execute .compute* method on the same key at the same time, callbacks still will be executed one after another and not simultaneously. But ConcurrentHashMap doesn't provide ability to delete data in timely fashion as caches usually allow.<\/p>\n\n<p>Guava\/Caffeine caches provide ability to automatically delete values based on time, but you don't have that nasty feature of synchronized processing based on key, as in ConcurrentHashMap (you can get ConcurrentMap using asMap method, but .compute* implementations doesn't provide synchronization based on key)<\/p>\n\n<p>My goal is to have both processing synchronization by key as in ConcurrentHashMap AND removals by time as in Guava\/Caffeine. <\/p>\n\n<p>What is the best way to achieve it in Java?<\/p>\n","answers":[{"AnswerId":"57825130","CreationDate":"2019-09-06T15:49:05.113","ParentId":null,"OwnerUserId":"3231055","Title":null,"Body":"<p>I was wrong about Caffeine - it supports atomic operations with compute.\nGuava support was added since 21 version.<\/p>\n"}]}
+{"QuestionId":57821826,"AnswerCount":0,"Tags":"<python><keras><deep-learning><classification><multilabel-classification>","CreationDate":"2019-09-06T12:12:37.680","AcceptedAnswerId":null,"OwnerUserId":11900725.0,"Title":"How do you train a neural network to predict a set number of labels for every input?","Body":"<p>I am working on a project where we are trying to classify gene expression using a neural network. We are using Keras. We have the sequences of 35000 genes. For each of these genes, we know how much they are expressed in 28 different tissues. Therefore, each gene has 28 different labels. Each label can be one of the following: 'none', 'low', 'med', or 'high'. What I would like to see is that for every gene we feed to the network, it returns 28 labels. Each of the label corresponds to the predicted rate of expression in the corresponding tissue for that particular gene. <\/p>\n\n<p>I have tried googling for the answer and came across multi-label-classifiers. These classifiers return a binary output vector. For example: an output of [0, 1, 0, 1, 1] means that the input is classified as having label 2, 4, and 5. This would not be suitable solution for my problem as we need to assign one of the 4 labels to each tissue. Would anyone have a clue how I could solve this?<\/p>\n","answers":[]}
+{"QuestionId":57821942,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-06T12:21:31.073","AcceptedAnswerId":57826321.0,"OwnerUserId":5177587.0,"Title":"Keras: how to do weighted addition of outputs of encoders with weights given by another classifier?","Body":"<p><strong>I need to merge multidimensional outputs of 3 (or more) encoders by weighted addition, and the weights are coming from another classifier<\/strong>. How do I do this?<\/p>\n\n<p>basically I need to do this (in a vectorized form):<\/p>\n\n<p>output = dotProduct([output1, output2, output3], w) = w1*output1 + w2*output2 + w3*output3<\/p>\n\n<p>Dimensions:<\/p>\n\n<p>output = output1 = output2 = output3 = [None, 16, 16, 512]<\/p>\n\n<p>w = [None, 3]<\/p>\n\n<p>None is the unknown dimension allocated for batch.<\/p>\n","answers":[{"AnswerId":"57826321","CreationDate":"2019-09-06T17:26:02.743","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>For a generic number of encoders, you could do something along the lines of:<\/p>\n\n<pre><code>def f(x):\n    w = x[-1]\n    outputs = x[:-1]\n    outputs_ = K.concatenate([o[:, None, ...] for o in outputs],\n                             axis=1)  # Shape=(None, nb_outputs, 16, 16, 512)\n\n    out = K.sum(w[..., None, None, None] * outputs_,\n                axis=1)  # Shape=(None, 16, 16, 512)\n    return out\n\noutputs = [output1, output2, output3]  # Works with any number of outputs\noutput = Lambda(f)(outputs + [w])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57822131,"AnswerCount":0,"Tags":"<python><machine-learning><keras><deep-learning>","CreationDate":"2019-09-06T12:33:21.083","AcceptedAnswerId":null,"OwnerUserId":12030400.0,"Title":"Is it possible we have a great validation accuracy and both low sensitivity and specificity in balanced validation set?","Body":"<p>I have a problem. I'm using Keras with TensorFlow backend to image recognition. There are two classes: positive and negative tasks.<\/p>\n\n<p>I'm using a pretrained model (with freeze weights) with a dense layers (that is not freeze).<\/p>\n\n<p>I have reported a problem: training after 100 epochs, I got that results bellow:<\/p>\n\n<blockquote>\n  <p>Epoch 98\/100 488\/488 [==============================] - 222s\n  454ms\/step - loss: 0.0774 - acc: 0.9720 - sensitivity: 0.9746 -\n  specificity: 0.9683 - val_loss: 0.0660 - val_acc: 0.9697 -\n  val_sensitivity: 0.4412 - val_specificity: 0.5368<br> Epoch 99\/100\n  488\/488 [==============================] - 221s 453ms\/step - loss:\n  0.0791 - acc: 0.9687 - sensitivity: 0.9729 - specificity: 0.9647 - val_loss: 0.0930 - val_acc: 0.9570 - val_sensitivity: 0.5428 -\n  val_specificity: 0.4245<br> Epoch 100\/100 488\/488\n  [==============================] - 221s 453ms\/step - loss: 0.0775 -\n  acc: 0.9693 - sensitivity: 0.9715 - specificity: 0.9655 - val_loss:\n  0.0751 - val_acc: 0.9684 - val_sensitivity: 0.4482 - val_specificity: 0.5285<\/p>\n<\/blockquote>\n\n<p>I got an accuracy of 96%, loss of 0.0751 (binary cross entropy) and low both sensitivity and specificity. It could be a problem with the class imbalance, but I have 2600 test instances, which positive class has 1302 and negative class has 1298. <\/p>\n\n<p>If we have an accuracy of +90% it was expected that one of the two another metrics could provide a good result classification. If these metrics are for evaluate the hit of each class, and we have a high accuracy, how can we get low values for <strong>both<\/strong> metrics?<\/p>\n\n<p>I have no idea what is happening here, someone can help me?<\/p>\n\n<p>There are my sensitivity and specificity functions bellow:<\/p>\n\n<pre><code>def sensitivity(y_true, y_pred):\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    return true_positives \/ (possible_positives + K.epsilon())\ndef specificity(y_true, y_pred):\n    true_negatives = K.sum(K.round(K.clip((1-y_true) * (1-y_pred), 0, 1)))\n    possible_negatives = K.sum(K.round(K.clip(1-y_true, 0, 1)))\n    return true_negatives \/ (possible_negatives + K.epsilon())\n<\/code><\/pre>\n\n<p>I apologize for my bad English. Thank you for your help!<\/p>\n\n<p>I tried to make a binary classification with images. I used transfer learning, and added two metrics for evaluation of the model: sensitivity and specificity.<\/p>\n\n<p>If we have a great accuracy in a binary classification task with <strong>balanced<\/strong> validation data, we expect get great results in the validation sensitivity OR specificity. There is some inconsistency in the model or metric evaluation.<\/p>\n\n<p><strong>EDIT 9\/9<\/strong>\nI have tried to use another transfer learning OR simple models and I've got the same results. I really don't understand what is this.<\/p>\n\n<blockquote>\n  <p>Epoch 100\/100 488\/488 [==============================] - 217s 445ms\/step - loss: 0.0575 - acc: 0.9767 - sensitivity: 0.9793 - specificity: 0.9737 - val_loss: 0.0731 - val_acc: 0.9653 - val_sensitivity: 0.4904 - val_specificity: 0.4852<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":57822568,"AnswerCount":1,"Tags":"<python><neural-network><pytorch>","CreationDate":"2019-09-06T13:01:42.793","AcceptedAnswerId":57833425.0,"OwnerUserId":9464073.0,"Title":"Deep learning with Pytorch: understanding the neural network example","Body":"<p>I'm reading the <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/blitz\/neural_networks_tutorial.html\" rel=\"nofollow noreferrer\">Pytorch documentation<\/a> and I have a couple of questions about the neural network that is introduced. The documentation defines the following network:<\/p>\n\n<pre><code>import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\n    class Net(nn.Module):\n\n        def __init__(self):\n            super(Net, self).__init__()\n            # 1 input image channel, 6 output channels, 3x3 square convolution\n            # kernel\n            self.conv1 = nn.Conv2d(1, 6, 3)\n            self.conv2 = nn.Conv2d(6, 16, 3)\n            # an affine operation: y = Wx + b\n            self.fc1 = nn.Linear(16 * 6 * 6, 120)  # 6*6 from image dimension\n            self.fc2 = nn.Linear(120, 84)\n            self.fc3 = nn.Linear(84, 10)\n\n        def forward(self, x):\n            # Max pooling over a (2, 2) window\n            x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))\n            # If the size is a square you can only specify a single number\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        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<\/code><\/pre>\n\n<p>Later on, the following statement is made:<\/p>\n\n<p><em>Let try a random 32x32 input. Note: expected input size of this net (LeNet) is 32x32. To use this net on MNIST dataset, please resize the images from the dataset to 32x32.<\/em><\/p>\n\n<p><strong>Question 1:<\/strong> Why do the images need to be 32x32 (where I assume that this means 32 pixels by 32)? <\/p>\n\n<p>The first convolution applies six kernel to an image, with every kernel being 3x3. This means that if the input channel is 32x32, the six output channels all have dimensions 30x30 (the 3x3 kernel grid makes you lose 2 pixels in width and height). The second convolution applies more kernels so that there now are sixteen output channels of dimensions 28x28 (again the 3x3 kernel grid makes you lose 2 pixels in width and height). Now I would expect 16x28x28 nodes in the next layer, since every one of the sixteen output channels has 28x28 pixels. Somehow, this is incorrect, and the next layer contains 16x6x6 nodes. Why is this true?<\/p>\n\n<p><strong>Question 2:<\/strong> The second convolution layer goes from six input channels to sixteen output channels. How is this done? <\/p>\n\n<p>In the first convolution layer we go from one input channel to six input channels, which makes sense to me. You can just apply six kernels to the single input channel to arrive at six output channels. Going from six input channels to sixteen output channels does not make as much sense to me. How are the different kernels applied? Do you apply two kernels to the first five input channels to arrive at ten output channels, and apply six kernels to the last input channel, so that the total comes to sixteen output channels? Or does the neural network learn itself to use x kernels and apply them to the input channels that it finds most suitable?<\/p>\n","answers":[{"AnswerId":"57833425","CreationDate":"2019-09-07T11:52:34.133","ParentId":null,"OwnerUserId":"9464073","Title":null,"Body":"<p>I can now answer these questions my self. <\/p>\n\n<p><strong>Question 1:<\/strong> To see why you need a 32x32 image for this neural network to work consider the following:<\/p>\n\n<p><strong>Layer 1:<\/strong>\nFirst, convolution is applied with a 3x3 kernel. Since the image has dimensions 32x32, this will result in a grid of 30x30. Next, max pooling is applied to the grid, with a 2x2 kernel and stride of 2 resulting in a grid that has dimensions 15x15.<\/p>\n\n<p><strong>Layer 2:<\/strong>\nFirst, convolution is applied with a 3x3 kernel to the 15x15 grid, resulting in a 13x13 grid. Next, max pooling is applied with a 2x2 kernel and stride of 2 resulting in a grid that has dimensions 6x6. We get a 6x6 grid and not a 7x7 grid because by default the floor function is used and not the ceil function. <\/p>\n\n<p>Since the convolution in layer 2 has sixteen output channels, the first linear layer needs 16x6x6 nodes! We see that the required input is indeed a 32x32 image.<\/p>\n\n<p><strong>Question 2:<\/strong> Every output channel is created by applying six different kernels to each input channel and summing the results. This is explained in the <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#conv2d\" rel=\"nofollow noreferrer\">documentation<\/a>.<\/p>\n"}]}
+{"QuestionId":57823210,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><tensorflow-datasets><tensorflow-estimator>","CreationDate":"2019-09-06T13:42:53.543","AcceptedAnswerId":57919403.0,"OwnerUserId":12030773.0,"Title":"How to combine tf.data.Dataset and tf.estimator.DNNRegressor properly","Body":"<p>I am currently learning to use tensorflow and have troubles getting started.\nI would like to use the newest API, namely estimator and dataset. But if I run the code presented below I get an Error.<\/p>\n\n<p>On the tensorflow page <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/DNNRegressor\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/DNNRegressor<\/a> I found, that \"The function should construct and return one of the following: * A tf.data.Dataset object: Outputs of Dataset object must be a tuple (features, labels) with same constraints as below.\"<\/p>\n\n<p>I thought my code would provide that, but there seems to be a problem and I am out of ideas.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\ndef input_evaluation_set():\n    data = [0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1]\n    labels = []\n    for d in data:\n        labels.append(1)\n    return tf.data.Dataset.from_tensor_slices((tf.constant(data), tf.constant(labels)))\n\npoint = tf.feature_column.numeric_column('points')\nestimator = tf.estimator.DNNRegressor(feature_columns = [point],hidden_units = [100,100,100])\n\nestimator.train(input_fn = input_evaluation_set)\n<\/code><\/pre>\n\n<p>I expect to run a training session on a deep neural network with 3 hidden layers a' 100 neurons in order to approximate the 'constant 1' function;\ninstead I get the Error \"ValueError: features should be a dictionary of 'Tensor's. Given type: class, 'tensorflow.python.framework.ops.Tensor'<\/p>\n","answers":[{"AnswerId":"57919403","CreationDate":"2019-09-13T07:50:00.727","ParentId":null,"OwnerUserId":"12030773","Title":null,"Body":"<p>You need to use .batch on your database in order to have the right format.<\/p>\n\n<p>The following is working on my computer:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ndef basic_dataset(numPoints):\n    data = np.linspace(0,1,numPoints)\n    dataset = dict({'points': data})\n    labels = []\n    for d in data:\n        labels.append(1)\n    return tf.data.Dataset.from_tensor_slices((dataset, np.array(labels)))\n\ndef input_train_set():\n    dataset = basic_dataset(11)\n    return dataset.repeat(100).shuffle(1000).batch(1)\n\npoint = tf.feature_column.numeric_column('points')\nestimator = tf.estimator.DNNRegressor(feature_columns = [point],hidden_units = [100,100,100], label_dimension = 1)\n\nestimator.train(input_fn = input_train_set)\n<\/code><\/pre>\n"},{"AnswerId":"57824543","CreationDate":"2019-09-06T15:07:16.827","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>Main problem in your code was you were sending the data in dataset as simple tensors. But it input data in dataset should be dictionary with key name same as that with what was used in feature column. Apart, from that I have added extra dimension in input data. The following code will work.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n### DEFINE NEW MAP FUNCTION\ndef map_fn(d, l):\n    return {'points': d}, l\n\ndef input_evaluation_set():\n    data = [0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1]\n    labels = []\n    for d in data:\n        labels.append(1)\n\n    ### CHANGE STARTS HERE\n    data = np.array(data)\n    data = np.expand_dims(data, axis=-1)\n    labels = np.array(labels)\n    labels = np.expand_dims(labels, axis=-1)\n    ### CHANGE ENDS HERE\n\n    dataset = tf.data.Dataset.from_tensor_slices((tf.constant(data), tf.constant(labels)))\n\n    ### CREATE DICTIONARY PAIR IN INPUT DATA\n    dataset = dataset.map(map_fn)\n    return dataset\n\npoint = tf.feature_column.numeric_column('points')\nestimator = tf.estimator.DNNRegressor(feature_columns = [point],hidden_units = [100,100,100])\n\nestimator.train(input_fn = input_evaluation_set)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57823443,"AnswerCount":0,"Tags":"<numpy><torch>","CreationDate":"2019-09-06T13:58:17.067","AcceptedAnswerId":null,"OwnerUserId":4619822.0,"Title":"Equivalent of torch.addmm in Numpy","Body":"<p>Hi I'm rewriting a module writted in Pytorch, I need to tr all the code calculation in Numpy and I'm struggling to translate in Pytorch this line <code>feature.addmm_(1, -2, feat, feat.t())<\/code> \nwhat is the equivalent of Pytorch <code>addmm_<\/code> or <code>addmm<\/code> in Numpy<\/p>\n\n<p>for example:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>a = torch.tensor([[0.9619, 0.0384, 0.7012],\n        [0.5561, 0.3637, 0.9272]])\n\nb = 2\n\nc = a.sum(dim=1, keepdim=True).expand(b, b)\ny = c + c.t()\ny = y.addmm_(1, -2, a, a.t())\n\nprint(y)\n# tensor([[0.5662, 1.1504], [1.1504, 1.0916]])\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57823554,"AnswerCount":1,"Tags":"<python><tensorflow><tf.keras><nsl>","CreationDate":"2019-09-06T14:04:18.950","AcceptedAnswerId":57824837.0,"OwnerUserId":8608022.0,"Title":"InvalidArgumentError: Cannot update variable with shape [] using a Tensor with shape [32]","Body":"<p>I am trying to get started with Neural Structured Learning but when I run the example given on the page to test, I get the following error<\/p>\n\n<p>I have tried to squeeze the dimensions, I have tried a different version of tensorflow --- I am still quite new to tensorflow so at this point I would really be guessing.<\/p>\n\n<pre><code># Create a base model -- sequential, functional, or subclass.\nmodel = tf.keras.Sequential([\n    tf.keras.Input((28, 28), name='feature'),\n    tf.keras.layers.Flatten(),\n    tf.keras.layers.Dense(128, activation=tf.nn.relu),\n    tf.keras.layers.Dense(10, activation=tf.nn.softmax)\n])\n\n# Wrap the model with adversarial regularization.\nadv_config = nsl.configs.make_adv_reg_config(multiplier=0.2, adv_step_size=0.05)\nadv_model = nsl.keras.AdversarialRegularization(model, adv_config=adv_config)\n\n\n# Compile, train, and evaluate.\nadv_model.compile(optimizer='adam',loss='sparse_categorical_crossentropy',metrics=['accuracy'])\n\n\n#let us now fit the model\nadv_model.fit({'feature': x_train, 'label': y_train}, batch_size=32, epochs=5)\n\n\n\nW0906 13:48:30.427690 140388427564928 training_utils.py:1101] Output output_1 missing from loss dictionary. We assume this was done on purpose. The fit and evaluate APIs will not be expecting any data to be passed to output_1.\nEpoch 1\/5\n---------------------------------------------------------------------------\nInvalidArgumentError                      Traceback (most recent call last)\n&lt;ipython-input-21-a5b951c24c49&gt; in &lt;module&gt;()\n----&gt; 1 adv_model.fit({'feature': x_train, 'label': y_train}, batch_size=32, epochs=5)\n\n3 frames\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py in __call__(self, *args, **kwargs)\n   1456         ret = tf_session.TF_SessionRunCallable(self._session._session,\n   1457                                                self._handle, args,\n-&gt; 1458                                                run_metadata_ptr)\n   1459         if run_metadata:\n   1460           proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\nInvalidArgumentError: Cannot update variable with shape [] using a Tensor with shape [32], shapes must be equal.\n     [[{{node AdversarialRegularization_1\/AssignAddVariableOp_2}}]]\n\n<\/code><\/pre>\n\n<p>The model is supposed to train and I get some accuracy from it. I don't understand where the problem is from in my code.<\/p>\n","answers":[{"AnswerId":"57824837","CreationDate":"2019-09-06T15:28:23.917","ParentId":null,"OwnerUserId":"4685471","Title":null,"Body":"<p>Although not mentioned explicitly in any requirements, in <a href=\"https:\/\/www.tensorflow.org\/neural_structured_learning\/tutorials\/adversarial_keras_cnn_mnist\" rel=\"nofollow noreferrer\">all<\/a> <a href=\"https:\/\/www.tensorflow.org\/neural_structured_learning\/tutorials\/graph_keras_lstm_imdb\" rel=\"nofollow noreferrer\">three<\/a> (so far) <a href=\"https:\/\/www.tensorflow.org\/neural_structured_learning\/tutorials\/graph_keras_mlp_cora\" rel=\"nofollow noreferrer\">tutorials<\/a>, the first step is to install Tensorflow 2.0:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/yexTf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/yexTf.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>There is also a report of an error identical to yours in the relevant <a href=\"https:\/\/medium.com\/@amalvnair95\/i-tried-out-example-code-for-mnist-and-i-get-the-following-error-9f7ff01f52e0\" rel=\"nofollow noreferrer\">blog post<\/a>, which was <a href=\"https:\/\/medium.com\/p\/9f7ff01f52e0\/responses\/show\" rel=\"nofollow noreferrer\">resolved<\/a> by upgrading to Tensorflow 2.0<\/p>\n\n<p>So, make an environment where you install TF 2.0 and the package:<\/p>\n\n<pre><code>pip install --quiet tensorflow==2.0.0-rc0\npip install --quiet neural-structured-learning\n<\/code><\/pre>\n\n<p>and you should be fine.<\/p>\n"}]}
+{"QuestionId":57823601,"AnswerCount":0,"Tags":"<python><opencv><keras><image-thresholding>","CreationDate":"2019-09-06T14:06:27.043","AcceptedAnswerId":null,"OwnerUserId":12030299.0,"Title":"I am getting an error when trying to apply otsu threshholding in the preprocessing_function of imagedatagenerator in keras","Body":"<p>I am getting an error when trying to apply otsu threshholding in the preprocessing_function of imagedatagenerator in keras:<\/p>\n\n<p>the exact error massage is:<\/p>\n\n<pre><code>error: OpenCV(3.4.3) \/io\/opencv\/modules\/imgproc\/src\/thresh.cpp:1406: \nerror: (-215:Assertion failed) src.type() == CV_8UC1 in function \n'threshold'\n<\/code><\/pre>\n\n<p>I tried already to change all dtypes of images which come in through flow from directory to uint8. I also tried to change the dimensions of the incoming image from (150,150,1) to (150,150) and before returning (after otsu) back to (150,150,1).\nThe images i put in are already gray since i use color_mode = \"grayscale\" in my generator. I also tried color_mode = \"rgb\" and changing it manually with cv2 to grayscale before passing it into the otsu thresholding.<\/p>\n\n<p>here is my code for the preprocessing function:<\/p>\n\n<pre><code>import cv2 as cv\nimport numpy as np\n\ndef otsu(img):\n  img_blur = cv.GaussianBlur(img,(5,5),0)\n\n  ret, img_otsu = cv.threshold(img_blur,0,255,cv.THRESH_BINARY + \n  cv.THRESH_OTSU)\n\n  return img_otsu\n<\/code><\/pre>\n\n<p>here are my generators i pass it to:<\/p>\n\n<pre><code>from keras.preprocessing.image import ImageDataGenerator\n\ntrain_datagen = ImageDataGenerator(\n    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    fill_mode = 'nearest',\n    preprocessing_function = otsu)\n\nvalidation_datagen = ImageDataGenerator(rescale = 1.0\/255.,  \n    preprocessing_function = otsu)\n\ntrain_generator = train_datagen.flow_from_directory(\n    \"drive\/Drive\/images CNN binary\/TRAINING\",\n    batch_size = 20,\n    color_mode = \"grayscale\",\n    class_mode = 'binary',\n    target_size = (150, 150))\n\nvalidation_generator =  validation_datagen.flow_from_directory(\n    \"drive\/My Drive\/images CNN binary\/VALIDATION\",\n    batch_size  = 20,\n    color_mode = \"grayscale\",\n    class_mode  = 'binary', \n    target_size = (150, 150))\n<\/code><\/pre>\n\n<p>And here i use the generator<\/p>\n\n<pre><code>history = model.fit_generator(\n    train_generator,\n    validation_data = validation_generator,\n    steps_per_epoch = 100,\n    epochs = 20,\n    validation_steps = 50,\n    callbacks = callbacks_list,\n    verbose = 2)\n<\/code><\/pre>\n\n<p>Thank you in advance<\/p>\n","answers":[]}
+{"QuestionId":57823708,"AnswerCount":0,"Tags":"<tensorflow><regression><conv-neural-network>","CreationDate":"2019-09-06T14:12:18.913","AcceptedAnswerId":null,"OwnerUserId":415885.0,"Title":"How to minimize the loss?","Body":"<p>This should be a regression problem.<\/p>\n\n<p>I would like the Neural Network to be able to estimate the length of a line, in pixel, from an image, like this 3 images, each image is 200 x 200 pcs:<\/p>\n\n<p>a)<a href=\"https:\/\/i.stack.imgur.com\/5hAEG.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5hAEG.jpg\" alt=\"Train_1\"><\/a>b)<a href=\"https:\/\/i.stack.imgur.com\/uODWY.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/uODWY.jpg\" alt=\"Train_2\"><\/a>c)<a href=\"https:\/\/i.stack.imgur.com\/zbVK1.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/zbVK1.jpg\" alt=\"Train_3\"><\/a><\/p>\n\n<p>Training image of 6000 pcs, and validation image of 1000 pcs.<\/p>\n\n<p>The labels are the distance in pixel:<\/p>\n\n<p>a) 1.205404496424333018e+02<\/p>\n\n<p>b) 1.188780888137086436e+02<\/p>\n\n<p>c) 1.110180165558725918e+02<\/p>\n\n<p>Here is my training code:<\/p>\n\n<pre><code>img_size = 200\n\ndef preprocess_image(image):\n  image = tf.image.decode_jpeg(image, channels=3)\n  image = tf.image.resize(image, [img_size, img_size])\n  image \/= 255.0  # normalize to [0,1] range\n\n  return image\n\ndef load_and_preprocess_image(path):\n  image = tf.read_file(path)\n  return preprocess_image(image)\n\nAUTOTUNE = tf.data.experimental.AUTOTUNE\nBATCH_SIZE = 16\n\ntrain_labels = np.loadtxt(\"train_labels.txt\")\nval_labels = np.loadtxt(\"test_labels.txt\")\n\ntrain_images = sorted(glob.glob(\"train_img\/img_*.jpg\"))\nval_images = sorted(glob.glob(\"test_img\/img_*.jpg\"))\n\nsteps_per_epoch_count=tf.ceil(len(train_images)\/BATCH_SIZE)\n\ntrain_path_ds = tf.data.Dataset.from_tensor_slices(train_images)\nval_path_ds = tf.data.Dataset.from_tensor_slices(val_images)\n\ntrain_image_ds = train_path_ds.map(load_and_preprocess_image, \nnum_parallel_calls = AUTOTUNE)\ntrain_label_ds = \ntf.data.Dataset.from_tensor_slices(tf.cast(train_labels, tf.float32))\ntrain_image_label_ds = tf.data.Dataset.zip((train_image_ds, \ntrain_label_ds))\n\nval_image_ds = val_path_ds.map(load_and_preprocess_image, \nnum_parallel_calls = AUTOTUNE)\nval_label_ds = tf.data.Dataset.from_tensor_slices(tf.cast(val_labels, tf.float32))\nval_image_label_ds = tf.data.Dataset.zip((val_image_ds, val_label_ds))\n\n\nmodel = tf.keras.models.Sequential([\n    tf.keras.layers.Convolution2D(16,3,3, input_shape=(img_size, \nimg_size, 3), activation = 'relu'),\n\ntf.keras.layers.MaxPooling2D(pool_size=(2,2)),\ntf.keras.layers.Convolution2D(32,3,3, activation = 'relu'),\ntf.keras.layers.MaxPooling2D(pool_size=(2,2)),\n# tf.keras.layers.Convolution2D(64,3,3, activation = 'relu'),\n# tf.keras.layers.MaxPooling2D(pool_size=(2,2)),\ntf.keras.layers.Flatten(),\ntf.keras.layers.Dense(400, activation=tf.nn.relu),\ntf.keras.layers.Dropout(0.2),\ntf.keras.layers.Dense(200, activation=tf.nn.relu),\ntf.keras.layers.Dropout(0.1),\ntf.keras.layers.Dense(100, activation=tf.nn.relu),\ntf.keras.layers.Dropout(0.05),\ntf.keras.layers.Dense(1, activation=tf.nn.relu)\n])\n\nmodel.compile(optimizer=tf.keras.optimizers.RMSprop(0.01),\n    loss = \"mean_squared_error\",\n    metrics = [\"mean_absolute_error\", \"mean_squared_error\"]\n    )\n\ntrain_ds = train_image_label_ds.apply(tf.data.experimental.shuffle_and_repeat(buffer_size=len(train_images)))\ntrain_ds = train_ds.batch(BATCH_SIZE)\ntrain_ds = train_ds.prefetch(buffer_size=AUTOTUNE)\n\n\nval_ds = val_image_label_ds.apply(\n  tf.data.experimental.shuffle_and_repeat(buffer_size=len(val_images)))\nval_ds = val_ds.batch(BATCH_SIZE)\nval_ds = val_ds.prefetch(buffer_size=AUTOTUNE)\n\nhistory = model.fit(\n    train_ds, \n    epochs = 80, \n    validation_data = val_ds,\n    steps_per_epoch = 374,\n    validation_steps = 62\n    )\n<\/code><\/pre>\n\n<p>However, this is the train vs eval mean_squared_error plot:\n<a href=\"https:\/\/i.stack.imgur.com\/NmGQw.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/NmGQw.png\" alt=\"Plot\"><\/a><\/p>\n\n<p>Question:<\/p>\n\n<ol>\n<li>Why is the validation loss not stable?<\/li>\n<li>The average Mean Squared Error is about 400 in training, which seems too high. What modification I can do to improve the estimation?<\/li>\n<\/ol>\n\n<p>EDIT:<\/p>\n\n<p>This is my latest model:<\/p>\n\n<p>Learning rate = 0.01<\/p>\n\n<p>Batch size = 16<\/p>\n\n<pre><code>model = tf.keras.models.Sequential([\n    tf.keras.layers.Convolution2D(16,3,3, input_shape=(img_size, img_size, 3), activation = 'relu'),\n    tf.keras.layers.MaxPooling2D(pool_size=(2,2)),\n    tf.keras.layers.Convolution2D(32,3,3, activation = 'relu'),\n    tf.keras.layers.MaxPooling2D(pool_size=(2,2)),\n    tf.keras.layers.Flatten(),\n    tf.keras.layers.Dense(2, activation=tf.nn.relu),\n    tf.keras.layers.Dropout(0.5),\n    tf.keras.layers.Dense(2, activation=tf.nn.relu), #, kernel_regularizer = tf.keras.regularizers.l2(0.001)\n    tf.keras.layers.Dropout(0.5),\n    tf.keras.layers.Dense(2, activation=tf.nn.relu), #, kernel_regularizer = tf.keras.regularizers.l2(0.001)\n    tf.keras.layers.Dropout(0.5),\n    tf.keras.layers.Dense(2, activation=tf.nn.relu), #, kernel_regularizer = tf.keras.regularizers.l2(0.001)\n    tf.keras.layers.Dense(1, activation=\"linear\")\n])\n<\/code><\/pre>\n\n<p>The output looks like this:\n<a href=\"https:\/\/i.stack.imgur.com\/mlWUR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/mlWUR.png\" alt=\"Very closed train and validation loss\"><\/a><\/p>\n\n<p>As you can see, the train and validation loss is almost identical. The mse loss are both stabilized around 2393, which square root to 48.91 pixel error, quite high. <\/p>\n\n<p>What advice to lower it further? Is it normal?<\/p>\n","answers":[]}
+{"QuestionId":57824511,"AnswerCount":1,"Tags":"<tensorflow><pytorch><tensor><mxnet>","CreationDate":"2019-09-06T15:05:48.547","AcceptedAnswerId":null,"OwnerUserId":5576325.0,"Title":"How to convert data(tensor) across different deep learning frameworks directly in gpu without copying to cpu?","Body":"<p>For example, I have a cuda tensor in pytorch, how to directly convert it to mxnet\/tensorflow in gpu without copying it to cpu? Frequently copying data between cpu and gpu is very time-consuming.<\/p>\n","answers":[{"AnswerId":"57826866","CreationDate":"2019-09-06T18:18:59.760","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>The fastest way would be to use <a href=\"https:\/\/docs.nvidia.com\/cuda\/cuda-runtime-api\/index.html\" rel=\"nofollow noreferrer\">CUDA API<\/a> for that. <\/p>\n\n<p>Sharing tensors via PyTorch is only supported on CPU. <\/p>\n\n<p>And if you would try to do some Python exhibition you will be slow because of GIL.<\/p>\n"}]}
+{"QuestionId":57824600,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-06T15:11:03.413","AcceptedAnswerId":57825200.0,"OwnerUserId":1179925.0,"Title":"Tensorflow: why tf.get_default_graph() is not current graph?","Body":"<p>In this example <code>tf.get_default_graph()<\/code> points to new graph? when this new graph is created? why it not point to already existing <code>graph<\/code>?<\/p>\n\n<pre><code>import tensorflow as tf\n\nprint('tf.__version__', tf.__version__)\n\ngraph = tf.Graph()\nwith graph.as_default():\n    x = tf.constant([[-2.25 + 4.75j], [-3.25 + 5.75j]])\n    y = tf.abs(x)\n\nprint(tf.get_default_graph())\nprint(graph)\ngraph == tf.get_default_graph()\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>tf.__version__ 1.14.0\n\n&lt;tensorflow.python.framework.ops.Graph object at 0x13ca5f128&gt;\n&lt;tensorflow.python.framework.ops.Graph object at 0x13ca5f4e0&gt;\n\nFalse\n<\/code><\/pre>\n","answers":[{"AnswerId":"57825200","CreationDate":"2019-09-06T15:53:34.817","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>By default, TensorFlow creates a \"root\" default graph, which will be the graph to use when no other graph has been designed as default with <code>.as_default()<\/code>. You cannot \"set\" this base default graph, as it is created internally by TensorFlow, although you can drop it and replace it with a new graph with <code>tf.reset_default_graph()<\/code>.<\/p>\n\n<p>In your example, you create a new graph <code>graph<\/code>, and then make a context manager with that graph as default.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>graph = tf.Graph()\nwith graph.as_default():\n    x = tf.constant([[-2.25 + 4.75j], [-3.25 + 5.75j]])\n    y = tf.abs(x)\n<\/code><\/pre>\n\n<p>The <code>tf.constant<\/code> and <code>tf.abs<\/code> operations in there will be created within <code>graph<\/code>, because that is the default graph inside that block.<\/p>\n\n<p>However, once the block finishes, the context manager that makes <code>graph<\/code> the default graph is finished too, so you are left with no explicitly set default graph. This means that the default graph will now be the one that TensorFlow created internally. If you call <code>tf.reset_default_graph()<\/code> and then <code>tf.get_default_graph()<\/code> again you will see that you get now a different graph.<\/p>\n\n<p>So, if you want to use some specific graph as default, you always need to use the <code>.as_default()<\/code> context to make it so, and the graph will stop being the default when you are out of that.<\/p>\n"}]}
+{"QuestionId":57824757,"AnswerCount":0,"Tags":"<tensorflow><keras><time-series><conv-neural-network><lstm>","CreationDate":"2019-09-06T15:22:34.493","AcceptedAnswerId":null,"OwnerUserId":2754279.0,"Title":"How to use TimeDistributed for CNN+LSTM?","Body":"<p>I am trying to classify 6 classes time-frequency domain signal (STFT spectrogram) with a size of 3601x217 pixels. Assume that for each classes have 70 training samples, 20 validation samples, and 10 test samples. Each sample is a PNG image of size 3601x217x1 (grayscale), while in the future I am thinking to expand this into 3601x217x4 (3 color channels + alpha). But right now, I am focusing on grayscale first.<\/p>\n\n<p>After trying pure LSTM, I found out that the model overfits quickly (training accuracy > 90%, but val_acc stucks at 20%). ow I'd like to try out Time Distributed CNN+LSTM.<\/p>\n\n<p>Following is the code:<\/p>\n\n<pre><code>img_width, img_height = 3601,217\ntrain_data_dir = 'sensor1\/training'\nvalidation_data_dir = 'sensor1\/validation'\n\nnum_classes = 6\nnb_train_samples = 70\nnb_validation_samples = 20\nepochs = 50\nbatch_size = 5\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\nmodel = Sequential()\nmodel.add(TimeDistributed(Conv2D(16, (3,3), padding='same', strides=(2,2),             activation='relu', input_shape = input_shape)))\nmodel.add(TimeDistributed(MaxPooling2D(pool_size=(2, 2))))\nmodel.add(Dropout(0.5))\nmodel.add(TimeDistributed(Conv2D(32, (3,3), padding='same', strides=(2,2),     activation='relu')))\nmodel.add(TimeDistributed(MaxPooling2D(pool_size=(2, 2))))\nmodel.add(Dropout(0.5))\nmodel.add(TimeDistributed(Conv2D(64, (3,3), padding='same', strides=(2,2), activation='relu')))\nmodel.add(TimeDistributed(MaxPooling2D(pool_size=(2, 2))))\nmodel.add(Dropout(0.5))\nmodel.add(TimeDistributed(Flatten()))\nmodel.add(LSTM(units=128, return_sequences=False))\nmodel.add(LSTM(units=64, return_sequences=False))\nmodel.add(Dense(32))\nmodel.add(Dense(num_classes, activation='softmax'))\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n\ntrain_datagen = ImageDataGenerator(rescale = 1. \/ 255)\ntest_datagen = ImageDataGenerator(rescale = 1. \/ 255)\n\ntrain_generator = train_datagen.flow_from_directory(\n    train_data_dir, target_size=(img_width, img_height),\n    batch_size=batch_size, color_mode='grayscale')\n\nvalidation_generator = test_datagen.flow_from_directory(\n    validation_data_dir, target_size=(img_width, img_height),\n    batch_size=batch_size, color_mode='grayscale')\n\nmodel.fit_generator(\n    train_generator, steps_per_epoch=nb_train_samples \/\/ batch_size,\n    epochs=epochs, callbacks=[plot_losses],\n    validation_data=validation_generator,\n    validation_steps=nb_validation_samples \/\/ batch_size)\n<\/code><\/pre>\n\n<p>The code gives following error:<\/p>\n\n<p><strong>File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\engine\\training.py\", line 496, in _make_train_function raise RuntimeError('You must compile your model before using it.')\nRuntimeError: You must compile your model before using it.<\/strong><\/p>\n\n<p>What's wrong with the code? I actually compiled the model before calling model.fit_generator:<\/p>\n\n<pre><code>model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57824804,"AnswerCount":2,"Tags":"<tensorflow><pytorch><adam>","CreationDate":"2019-09-06T15:26:12.470","AcceptedAnswerId":57826893.0,"OwnerUserId":11635654.0,"Title":"epsilon parameter in Adam opitmizer","Body":"<p>Using pyTorch and tensorflow (TF), I was wandering how the Adam optimizer is implemented for curiosity. And I do not know if I am wrong or not but it seems to me that the two implementations differ and the pyTorch one is the original one from <a href=\"https:\/\/arxiv.org\/pdf\/1412.6980.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1412.6980.pdf<\/a>. <\/p>\n\n<p>My problem comes from the eps-parameter. Using the TF implentation seems to lead to a time-and-b2 dependance of this parameter, namely<\/p>\n\n<pre><code>q(t+1) = q(t) - \\gamma * sqrt[(1-b2^t)]\/(1-b1^t) * m(t)\/[sqrt[v(t)]+eps]\n<\/code><\/pre>\n\n<p>which in the original algorithm notation can be reformulate as<\/p>\n\n<pre><code> q(t+1) = q(t) - \\gamma * mhat(t)\/[sqrt[vhat(t)]+ eps\/sqrt[(1-b2^t)]]\n<\/code><\/pre>\n\n<p>and this point out the variation of the eps-parameter which is not the case neither in the original algorithm neither in the pyTorch implementation.<\/p>\n\n<p>Am I wrong? or it is well known? \nThanks for your help.<\/p>\n","answers":[{"AnswerId":"57826893","CreationDate":"2019-09-06T18:21:44.473","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>Indeed, you can check this <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/AdamOptimizer\" rel=\"nofollow noreferrer\">in the docs<\/a> for the TF Adam optimizer. To quote the relevant part: <\/p>\n\n<blockquote>\n  <p>The default value of 1e-8 for epsilon might not be a good default in general. For example, when training an Inception network on ImageNet a current good choice is 1.0 or 0.1. Note that since AdamOptimizer uses the formulation just before Section 2.1 of the Kingma and Ba paper rather than the formulation in Algorithm 1, the \"epsilon\" referred to here is \"epsilon hat\" in the paper.<\/p>\n<\/blockquote>\n\n<p>If you check the \"the formulation just before section 2.1\" in the paper, they actually include the time dependence in <code>alpha<\/code>, resulting in a time-dependent \"step size\" <code>alpha_t<\/code> but a fixed <code>epsilon<\/code>. Note that at the end of the day, this is just rewriting\/interpreting parameters in a slightly different fashion and doesn't change the actual workings of the algorithm. But you <em>will<\/em> need to be aware that choosing the same <code>epsilon<\/code> in the PyTorch and TF implementations will apparently not lead to the same results...<\/p>\n"},{"AnswerId":"57877205","CreationDate":"2019-09-10T19:35:01.463","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>You can derive the formula from the second to the first as following. So in the tensorflow implementation, the epsilon is actually the epsilon' here. Also, in the tensorflow implementation, the learning rate is adjusted to alpha' in the following formula. Hope this helps. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/yMhzD.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/yMhzD.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":57825554,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-06T16:20:00.703","AcceptedAnswerId":57826020.0,"OwnerUserId":1036017.0,"Title":"tf.train.Saver not loading in same process","Body":"<p>I am observing a strange behavior where Saver can't restore if the checkpoint was saved earlier in the same Python process. It loads fine if done from a different process. Here's some simple code that will show the problem.<\/p>\n\n<pre><code>import tensorflow.compat.v1 as tf\n\ndef train():\n    W = tf.Variable(tf.zeros([1, 1]))\n    saver = tf.train.Saver()\n\n    with tf.Session() as sess:\n        sess.run(tf.global_variables_initializer())\n        saver.save(sess, \".\/model.ckpt\")\n\ndef predict():\n    W = tf.Variable(tf.zeros([1, 1]))\n    saver = tf.train.Saver()\n\n    with tf.Session() as sess:\n        sess.run(tf.global_variables_initializer())\n        saver.restore(sess, \".\/model.ckpt\")\n\ntrain()\npredict()\n<\/code><\/pre>\n\n<p>Here we save and restore immediately after that in the same process. Restoration fails with errors like:<\/p>\n\n<pre><code>Key Variable_1 not found in checkpoint\n<\/code><\/pre>\n\n<p>But if I run just the predict() code again from a new Python process it works just fine.<\/p>\n\n<pre><code>#train()\npredict()\n<\/code><\/pre>\n\n<p>Am I doing something wrong here?<\/p>\n","answers":[{"AnswerId":"57826020","CreationDate":"2019-09-06T16:57:47.530","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>After <code>predict<\/code>, if you run:<\/p>\n\n<pre><code>print([v for v in tf.trainable_variables()])\n<\/code><\/pre>\n\n<p>you will see that two different variables are being created. That's why TF is not able to restore the value of the second one.<\/p>\n\n<hr>\n\n<p>In order to link both variables into a single one, you can either:<\/p>\n\n<ul>\n<li><p>Pass a dictionary to the argument <code>var_list<\/code> of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Saver\" rel=\"nofollow noreferrer\">tf.train.Saver<\/a>. For example:<\/p>\n\n<pre><code>saver = tf.train.Saver({'W': W})\n<\/code><\/pre><\/li>\n<li><p>Use <a href=\"https:\/\/stackoverflow.com\/questions\/38545362\/tensorflow-variable-scope-reuse-if-variable-exists\/38895052#38895052\">auto-reusing<\/a> when creating the <a href=\"https:\/\/www.tensorflow.org\/guide\/variables\" rel=\"nofollow noreferrer\">variable<\/a>. For example:<\/p>\n\n<pre><code>with tf.variable_scope('', reuse=tf.AUTO_REUSE):\n    W = tf.get_variable(initializer=lambda: tf.zeros([1, 1]),\n                        name='W')\n<\/code><\/pre><\/li>\n<\/ul>\n"}]}
+{"QuestionId":57825630,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><tensorflow-datasets>","CreationDate":"2019-09-06T16:27:02.973","AcceptedAnswerId":null,"OwnerUserId":10063472.0,"Title":"Tensorflow train.py throws Windows fatal exception","Body":"<p>I've been working with Tensorflow for quite a while now, had some issues, but they never remained unresolved. Today i wanted to train a new model, when things got interesting. At first, the training stopped after one step without any reason. It happend before, opening a new cmd window solved it. Not this time tough. After i tried again, train.py started to throw this:<\/p>\n\n<blockquote>\n  <p>Windows fatal exception: access violation<\/p>\n  \n  <p>Current thread 0x000018d4 (most recent call first):<\/p>\n  \n  <p>File\n  \"C:\\windows\\system32\\venv\\lib\\site-packages\\tensorflow\\python\\lib\\io\\file_io.py\",\n  line 84 in _preread_check   File\n  \"C:\\windows\\system32\\venv\\lib\\site-packages\\tensorflow\\python\\lib\\io\\file_io.py\",\n  line 122 in read   File\n  \"C:\\Users\\xx\\source\\TensorFlow\\models\\research\\object_detection\\utils\\label_map_util.py\",\n  line 133 in load_labelmap   File\n  \"C:\\Users\\xx\\source\\TensorFlow\\models\\research\\object_detection\\utils\\label_map_util.py\",\n  line 164 in get_label_map_dict   File\n  \"C:\\Users\\xx\\source\\TensorFlow\\models\\research\\object_detection\\data_decoders\\tf_example_decoder.py\",\n  line 59 in <strong>init<\/strong>   File\n  \"C:\\Users\\xx\\source\\TensorFlow\\models\\research\\object_detection\\data_decoders\\tf_example_decoder.py\",\n  line 314 in <strong>init<\/strong>   File\n  \"C:\\Users\\xx\\source\\TensorFlow\\models\\research\\object_detection\\builders\\dataset_builder.py\",\n  line 130 in build   File \"train.py\", line 121 in get_next   File\n  \"C:\\Users\\xx\\source\\TensorFlow\\models\\research\\object_detection\\legacy\\trainer.py\",\n  line 59 in create_input_queue   File\n  \"C:\\Users\\xx\\source\\TensorFlow\\models\\research\\object_detection\\legacy\\trainer.py\",\n  line 280 in train   File \"train.py\", line 180 in main   File\n  \"C:\\windows\\system32\\venv\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py\",\n  line 324 in new_func   File \"C:\\Program Files (x86)\\Microsoft Visual\n  Studio\\Shared\\Python37_64\\lib\\site-packages\\absl\\app.py\", line 251 in\n  _run_main File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Python37_64\\lib\\site-packages\\absl\\app.py\", line 300 in\n  run File\n  \"C:\\windows\\system32\\venv\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\",\n  line 40 in run   File \"train.py\", line 184 in <\/p>\n<\/blockquote>\n\n<p>Last time i saw this issue, it was because i was using data downloaded from the internet, and there was one particular picture that TF did not like, but removing that one from the dataset solved the issue. I was wondering if this was the case, but no. I couldnt start it with previously tried datasets either... i decided to reinstall tensorflow, set up a new virtual environment, but still nothing. Been looking for hours what the problem could be, both on the internet, and on my own trying different things, but nothing worked, same exception each time. Did anybody encounter anything similar?<\/p>\n","answers":[{"AnswerId":"58536847","CreationDate":"2019-10-24T08:04:45.197","ParentId":null,"OwnerUserId":"10063472","Title":null,"Body":"<p>I decided to share what solved my problem, might help the others. I reinstalled Tensorflow itself in a virtual environment, and upgraded it to version 1.8 (Requires Python 3.6, it is not compatible with higher versions (mine is 3.6.5 in particular)), make sure your PYTHONPATH variable is pointing to the right folder. Also, on Windows, this error message on can occur when you use generate_tfrecord.py, i ran into it many times, it usually happend, because i had image(s), that Tensorflow did not like (i'm not completly sure about the cause), at first, try removing .webp, .gif, etc.(non .png\/.jpg) files. I even had the exception for renaming an image downloaded from the internet, and TF did not stand it anymore.<\/p>\n"}]}
+{"QuestionId":57825915,"AnswerCount":1,"Tags":"<c++><tensorflow><visual-studio-code><googletest><bazel>","CreationDate":"2019-09-06T16:49:20.613","AcceptedAnswerId":57866675.0,"OwnerUserId":2196255.0,"Title":"how to debug bazel c++ google test in visual studio code","Body":"<p>I have Visual Studio Code, C++ bazel tests that I built using command like this<\/p>\n\n<pre><code>bazel test \/\/tensorflow\/lite\/kernels:xxx_test --test_arg=gtest_filter=XXXTest -c dbg\n<\/code><\/pre>\n\n<p>Then I can debug it using gdb like this<\/p>\n\n<pre><code>gdb .\/bazel-bin\/tensorflow\/lite\/kernels\/xxx_test.runfiles\/org_tensorflow\/tensorflow\/lite\/kernels\/xxx_test\n<\/code><\/pre>\n\n<p>But is it another way to debug it properly in VS Code ?\nI installed plugin for Google Tests, but it does not see them and there is no gtest.exe.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"57866675","CreationDate":"2019-09-10T08:18:49.443","ParentId":null,"OwnerUserId":"7699037","Title":null,"Body":"<p>The following steps work for me when using the Google Test Adapter plugin. Go to <code>Debug -&gt; Add Configuration...<\/code> to add a new launch target to you <code>launch.json<\/code> file and fill in the program path under <code>\"program\"<\/code> (note that a relative path does not work for me here, so maybe you need to fill in the complete path):<\/p>\n\n<pre><code>{\n    \/\/ Use IntelliSense to learn about possible attributes.\n    \/\/ Hover to view descriptions of existing attributes.\n    \/\/ For more information, visit: https:\/\/go.microsoft.com\/fwlink\/?linkid=830387\n    \"version\": \"0.2.0\",\n    \"configurations\": [\n        {\n            \"name\": \"gtest\",\n            \"type\": \"cppdbg\",\n            \"request\": \"launch\",\n            \"program\": \"(maybe insert complete path)\/bazel-bin\/tensorflow\/lite\/kernels\/xxx_test.runfiles\/org_tensorflow\/tensorflow\/lite\/kernels\/xxx_test\",\n            \"args\": [],\n            \"stopAtEntry\": false,\n            \"cwd\": \"${workspaceFolder}\",\n            \"environment\": [],\n            \"externalConsole\": false,\n            \"MIMode\": \"gdb\",\n            \"setupCommands\": [\n                {\n                    \"description\": \"Enable pretty-printing for gdb\",\n                    \"text\": \"-enable-pretty-printing\",\n                    \"ignoreFailures\": true\n                }\n            ],\n            \"preLaunchTask\": \"\",\n            \"miDebuggerPath\": \"\/usr\/bin\/gdb\"\n        }\n    ]\n}\n<\/code><\/pre>\n\n<p>Check if you can correctly debug your app, by setting a break point in your code and then pressing <code>F5<\/code>.<\/p>\n\n<p>If it works go to the Google Test Adapter Plugin and hover the mouse on the left side (where it normally shows the available tests) and in the bar where it says \"GOOGLE TESTS\" (with all the run and stop buttons) press on the double arrow <code>&lt;-&gt;<\/code> and set the hook on the above debug configuration. Now pressing on <code>Run All<\/code> should correctly work.<\/p>\n\n<p>Let me know if it works and if not, which steps don't work.<\/p>\n"}]}
+{"QuestionId":57826075,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning><signal-processing>","CreationDate":"2019-09-06T17:03:12.350","AcceptedAnswerId":null,"OwnerUserId":5800182.0,"Title":"Deep learning, signal processing and feature engineering","Body":"<ol>\n<li><p>I have a signals represented in python in dense matrices (the values are y-coordinates from a chart - eg. weather temp etc. in different locations around the world).<\/p><\/li>\n<li><p>I'm currently trying to process\/recognize these matrices via different kinds of deep neural networks.<\/p><\/li>\n<li><p>I'm trying to enhance the important parts of the signals in many different ways as well...I know when we talk about sound \/ human voice...Log mel, MFCC etc. are useful tool to do such and enhance the important parts and vise-versa...<\/p><\/li>\n<\/ol>\n\n<p>What are my best options to enhance the most important features (in a similar way) in my dense matrices (in other words to feature engineer the most important features)? I'm of course aware about the fact, that the model can't know what the most important features are and not (without some kind of help \/ feature engineering - but what more can I do as a human to improve the model). Like set the importance of each feature?<\/p>\n","answers":[]}
+{"QuestionId":57826156,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-06T17:10:10.400","AcceptedAnswerId":null,"OwnerUserId":6150318.0,"Title":"How to add additional learning parameters to Keras.Sequential","Body":"<p>I would like to add an additional layer to my network in which there are only two parameters are training parameters. How can I do that<\/p>\n\n<p>Consider a simple model like <\/p>\n\n<pre><code>modelC1 = tf.keras.models.Sequential([\n  tf.keras.layers.Flatten(input_shape=(14,)),\n  tf.keras.layers.Dense(256, activation='relu'), \n  tf.keras.layers.Dense(128, activation='relu'),\n  tf.keras.layers.Dense(1, activation='sigmoid')\n],name='test')\n<\/code><\/pre>\n\n<p>What I want to do is to an additional layer of the form f=g(Ax) where g is my point-wise activation function and x is in the input. A is the weight matrix and has a form A = a1*I + b1*J where I and J are constant matrices and a1 and b1 are parameters to be learned. <\/p>\n\n<p>How can I do this?<\/p>\n","answers":[]}
+{"QuestionId":57826612,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><lstm>","CreationDate":"2019-09-06T17:55:05.917","AcceptedAnswerId":57826771.0,"OwnerUserId":2991243.0,"Title":"Reset all weights and biases of the model in Keras (restore the model after training)","Body":"<p>Suppose that I have something like this.<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(units = 10 input_shape = (x1, x2)))\nmodel.add(Activation('tanh'))\nmodel.compile(optimizer = 'adam', loss = 'mse')\n\n## Step 1.\nmodel.fit(X_train, Y_train, epochs = 10)\n<\/code><\/pre>\n\n<p>After training the model, I want to reset everything (weights and biases) in the model. So I want to restore the model after <code>compile<\/code> function (Step 1). What is the fastest way to that in Keras?<\/p>\n","answers":[{"AnswerId":"57826771","CreationDate":"2019-09-06T18:09:41.630","ParentId":null,"OwnerUserId":"5085211","Title":null,"Body":"<p>Whether it's the fastest is probably up in the air, but it's certainly straightforward and might be good enough for your case: Serialize the initial weights, then deserialize when necessary, and use something like <code>io.BytesIO<\/code> to avoid the disk I\/O hit (and having to clean up afterwards):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from io import BytesIO\n\nmodel = Sequential()\nmodel.add(LSTM(units = 10, input_shape = (x1, x2)))\nmodel.add(Activation('tanh'))\nmodel.compile(optimizer = 'adam', loss = 'mse')\nf = BytesIO()\nmodel.save_weights(f)  # Stores the weights\nmodel.fit(X_train, Y_train, epochs = 10)\n# [Do whatever you want with your trained model here]\nmodel.load_weights(f)  # Resets the weights\n<\/code><\/pre>\n"}]}
+{"QuestionId":57827274,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning><autoencoder><tf.keras>","CreationDate":"2019-09-06T18:56:52.593","AcceptedAnswerId":57833267.0,"OwnerUserId":11731187.0,"Title":"Keras Autoencoder: Tying Weights from Encoder To Decoder not working","Body":"<p>Im creating an Autoencoder as part of my full model for a Kaggle competition. Im trying to tie the weighs of the Encoder, transposed to the Decoder. Before the first Epoch the weights are correctly sync, after that, the Decoder weights just freeze, and dont keep up with the Encoder weights that are being updated by the Gradient Descent.<\/p>\n\n<p>I look for 12 hours in almost every post about this problem i reach on google, no one seems to have the answer for my case. The closest one is this <a href=\"https:\/\/stackoverflow.com\/questions\/53751024\/tying-autoencoder-weights-in-a-dense-keras-layer\">Tying Autoencoder Weights in a Dense Keras Layer<\/a> but the problem was solved by not using a variable tensor as kernel, but im already not using that type of tensor as my Decoder kernel, so was no usefull.<\/p>\n\n<p>Im using a DenseTied Keras custom Layer class defined in this article <a href=\"https:\/\/towardsdatascience.com\/build-the-right-autoencoder-tune-and-optimize-using-pca-principles-part-ii-24b9cca69bd6\" rel=\"noreferrer\">https:\/\/towardsdatascience.com\/build-the-right-autoencoder-tune-and-optimize-using-pca-principles-part-ii-24b9cca69bd6<\/a>, is exactly the same, just change the way i reference the Keras backed for suit my imports style.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport seaborn as sns\nimport os\n<\/code><\/pre>\n\n<p>This is the custom layer definition<\/p>\n\n<pre><code>class DenseTied(tf.keras.layers.Layer):\n\n    def __init__(self, units,\n                 activation=None,\n                 use_bias=True,\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                 tied_to=None,\n                 **kwargs):\n        self.tied_to = tied_to\n        if 'input_shape' not in kwargs and 'input_dim' in kwargs:\n            kwargs['input_shape'] = (kwargs.pop('input_dim'),)\n        super().__init__(**kwargs)\n        self.units = units\n        self.activation = tf.keras.activations.get(activation)\n        self.use_bias = use_bias\n        self.kernel_initializer = tf.keras.initializers.get(kernel_initializer)\n        self.bias_initializer = tf.keras.initializers.get(bias_initializer)\n        self.kernel_regularizer = tf.keras.regularizers.get(kernel_regularizer)\n        self.bias_regularizer = tf.keras.regularizers.get(bias_regularizer)\n        self.activity_regularizer = tf.keras.regularizers.get(activity_regularizer)\n        self.kernel_constraint = tf.keras.constraints.get(kernel_constraint)\n        self.bias_constraint = tf.keras.constraints.get(bias_constraint)\n        self.input_spec = tf.keras.layers.InputSpec(min_ndim=2)\n        self.supports_masking = True\n\n    def build(self, input_shape):\n        assert len(input_shape) &gt;= 2\n        input_dim = input_shape[-1]\n\n        if self.tied_to is not None:\n            self.kernel = tf.keras.backend.transpose(self.tied_to.kernel)\n            self.non_trainable_weights.append(self.kernel)\n        else:\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        if self.use_bias:\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        else:\n            self.bias = None\n        self.input_spec = tf.keras.layers.InputSpec(min_ndim=2, axes={-1: input_dim})\n        self.built = True\n\n    def compute_output_shape(self, input_shape):\n        assert input_shape and len(input_shape) &gt;= 2\n        output_shape = list(input_shape)\n        output_shape[-1] = self.units\n        return tuple(output_shape)\n\n    def call(self, inputs):\n        output = tf.keras.backend.dot(inputs, self.kernel)\n        if self.use_bias:\n            output = tf.keras.backend.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>And this the model training and testing with a dummy data set<\/p>\n\n<pre><code>rand_samples = np.random.rand(16, 51)\ndummy_ds = tf.data.Dataset.from_tensor_slices((rand_samples, rand_samples)).shuffle(16).batch(16)\n\nencoder = tf.keras.layers.Dense(1, activation=\"linear\", input_shape=(51,), use_bias=True)\ndecoder = DenseTied(51, activation=\"linear\", tied_to=encoder, use_bias=True)\n\nautoencoder = tf.keras.Sequential()\nautoencoder.add(encoder)\nautoencoder.add(decoder)\n\nautoencoder.compile(metrics=['accuracy'],\n                    loss='mean_squared_error',\n                    optimizer='sgd')\n\nautoencoder.summary()\n\nprint(\"Encoder Kernel Before 1 Epoch\", encoder.kernel[0])\nprint(\"Decoder Kernel Before 1 Epoch\", decoder.kernel[0][0])\n\nautoencoder.fit(dummy_ds, epochs=1)\n\nprint(\"Encoder Kernel After 1 Epoch\", encoder.kernel[0])\nprint(\"Decoder Kernel After 1 Epoch\", decoder.kernel[0][0])\n<\/code><\/pre>\n\n<p>The Expected output is have the two kernels exactly the same in the first element (Print just one weight for simplicity)<\/p>\n\n<p>The current output shows that Decoder Kernel is not updated as the same as the Transposed Encoder Kernel<\/p>\n\n<pre><code>2019-09-06 14:55:42.070003: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library nvcuda.dll\n2019-09-06 14:55:42.984580: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1640] Found device 0 with properties:\nname: GeForce GTX 1060 major: 6 minor: 1 memoryClockRate(GHz): 1.733\npciBusID: 0000:01:00.0\n2019-09-06 14:55:43.088109: I tensorflow\/stream_executor\/platform\/default\/dlopen_checker_stub.cc:25] GPU libraries are statically linked, skip dlopen check.        \n2019-09-06 14:55:43.166145: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1763] Adding visible gpu devices: 0\n2019-09-06 14:55:43.203865: I tensorflow\/core\/platform\/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2\n2019-09-06 14:55:43.277988: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1640] Found device 0 with properties:\nname: GeForce GTX 1060 major: 6 minor: 1 memoryClockRate(GHz): 1.733\npciBusID: 0000:01:00.0\n2019-09-06 14:55:43.300888: I tensorflow\/stream_executor\/platform\/default\/dlopen_checker_stub.cc:25] GPU libraries are statically linked, skip dlopen check.        \n2019-09-06 14:55:43.309040: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1763] Adding visible gpu devices: 0\n2019-09-06 14:55:44.077814: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1181] Device interconnect StreamExecutor with strength 1 edge matrix:\n2019-09-06 14:55:44.094542: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1187]      0\n2019-09-06 14:55:44.099411: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1200] 0:   N\n2019-09-06 14:55:44.103424: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1326] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 4712 MB memory) -&gt; physical GPU (device: 0, name: GeForce GTX 1060, pci bus id: 0000:01:00.0, compute capability: 6.1)\nModel: \"sequential\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #\n=================================================================\ndense (Dense)                (None, 1)                 52\n_________________________________________________________________\ndense_tied (DenseTied)       (None, 51)                103\n=================================================================\nTotal params: 103\nTrainable params: 103\nNon-trainable params: 0\n_________________________________________________________________\nEncoder Kernel Before 1 Epoch tf.Tensor([0.20486075], shape=(1,), dtype=float32)\nDecoder Kernel Before 1 Epoch tf.Tensor(0.20486075, shape=(), dtype=float32)\n1\/1 [==============================] - 1s 657ms\/step - loss: 0.3396 - accuracy: 0.0000e+00\nEncoder Kernel After 1 Epoch tf.Tensor([0.20530733], shape=(1,), dtype=float32)\nDecoder Kernel After 1 Epoch tf.Tensor(0.20486075, shape=(), dtype=float32)\nPS C:\\Users\\whitm\\Desktop\\CodeProjects\\ForestClassifier-DEC&gt;\n<\/code><\/pre>\n\n<p>i dont see what im doing wrong.<\/p>\n","answers":[{"AnswerId":"57828731","CreationDate":"2019-09-06T21:29:53.447","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>The weights are not tied. You are just initializing the weights of the tied layer with the transposed weights of the first layer and then never train them. <code>transpose<\/code> returns a new tensor\/different object and <code>add_weight<\/code> creates a new variable, thus any relation between the two layers is lost after <code>build<\/code>. I think it would be better do so something like this:<\/p>\n\n<pre><code>def call(self, inputs):\n    output = tf.keras.backend.dot(inputs, tf.keras.backend.transpose(self.tied_to.kernel))\n    if self.use_bias:\n        output = tf.keras.backend.bias_add(output, self.tied_to.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>Here, the tied layer always uses the first layer's weights explicitly and would not have any weights itself (i.e. remove the <code>add_weight<\/code> parts from <code>build<\/code>).<\/p>\n"},{"AnswerId":"57833267","CreationDate":"2019-09-07T11:31:40.470","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>To tie the weights, I would suggest using the <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/#shared-layers\" rel=\"nofollow noreferrer\">Keras functional API<\/a> which enables to share layers. That said, here is an alternative implementation that ties the weights between the encoder and decoder:<\/p>\n\n<pre><code>class TransposableDense(tf.keras.layers.Dense):\n\n    def __init__(self, units, **kwargs):\n        super().__init__(units, **kwargs)\n\n    def build(self, input_shape):\n        assert len(input_shape) &gt;= 2\n        input_dim = input_shape[-1]\n        self.t_output_dim = input_dim\n\n        self.kernel = self.add_weight(shape=(int(input_dim), self.units),\n                                      initializer=self.kernel_initializer,\n                                      name='kernel',\n                                      regularizer=self.kernel_regularizer,\n                                      constraint=self.kernel_constraint)\n        if self.use_bias:\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.bias_t = self.add_weight(shape=(input_dim,),\n                                          initializer=self.bias_initializer,\n                                          name='bias_t',\n                                          regularizer=self.bias_regularizer,\n                                          constraint=self.bias_constraint)\n        else:\n            self.bias = None\n            self.bias_t = None\n        # self.input_spec = tf.keras.layers.InputSpec(min_ndim=2, axes={-1: input_dim})\n        self.built = True\n\n    def call(self, inputs, transpose=False):\n        bs, input_dim = inputs.get_shape()\n\n        kernel = self.kernel\n        bias = self.bias\n        if transpose:\n            assert input_dim == self.units\n            kernel = tf.keras.backend.transpose(kernel)\n            bias = self.bias_t\n\n        output = tf.keras.backend.dot(inputs, kernel)\n        if self.use_bias:\n            output = tf.keras.backend.bias_add(output, bias, data_format='channels_last')\n        if self.activation is not None:\n            output = self.activation(output)\n        return output\n\n    def compute_output_shape(self, input_shape):\n        bs, input_dim = input_shape\n        output_dim = self.units\n        if input_dim == self.units:\n            output_dim = self.t_output_dim\n        return bs, output_dim\n<\/code><\/pre>\n\n<p>The kernel of this dense layer can be transposed by calling the layer with <code>transpose=True<\/code>. Note that this might <em>break<\/em> some basic Keras principles (e.g. the layer has multiple output shapes), but it should work for your case.<\/p>\n\n<hr>\n\n<p>Here is an example showing how you can use it to define your model:<\/p>\n\n<pre><code>a = tf.keras.layers.Input((51,))\ndense = TransposableDense(1, activation='linear', use_bias=True)\nencoder_out = dense(a)\ndecoder_out = dense(encoder_out, transpose=True)\nencoder = tf.keras.Model(a, encoder_out)\nautoencoder = tf.keras.Model(a, decoder_out)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57827332,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-06T19:03:17.507","AcceptedAnswerId":null,"OwnerUserId":11427310.0,"Title":"tensorflow Sampled softmax loss","Body":"<p>I am using <code>tf.sampled_softmax_loss<\/code> in a RNN for translation from english to french. This is usefull to speed up the training but, is it possible to get the prediction of the samples used by <code>tf.sampled_softmax_loss<\/code> ?<\/p>\n\n<p>I'd like to be able to compute the accuracy.<\/p>\n\n<p>Indeed, what I understand is that <code>tf.sampled_softmax_loss<\/code> acts like a final layer that output all the logits for all the required class, and then apply the sampled algorithm to proccess a \"smaller\" softmax. But, I can't directly access, at least, the logits.<\/p>\n","answers":[]}
+{"QuestionId":57828198,"AnswerCount":0,"Tags":"<tensorflow><matrix-multiplication><ptx><tensorflow-xla><xla>","CreationDate":"2019-09-06T20:26:34.743","AcceptedAnswerId":null,"OwnerUserId":9637085.0,"Title":"Why PTX is not generated for a matrix multiplication tensorflow operation while I use XLA_FLAGS to dump ptx and llvm?","Body":"<p>I'm trying to dump <code>HLO<\/code>, <code>LLVM IR<\/code>, and <code>PTX<\/code> for the following <code>tensorflow 1.13.1<\/code> code:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\na = tf.placeholder(shape=(10,20), dtype=tf.float32)\nb = tf.placeholder(shape=(20,10), dtype=tf.float32)\nc = tf.placeholder(shape=(10,10), dtype=tf.float32)\n\njit_scope = tf.contrib.compiler.jit.experimental_jit_scope #using JIT compilation\nwith jit_scope():\n    d = tf.matmul(a, b) + c\n\nwith tf.Session() as sess:\n    print(sess.run(d, feed_dict={a:np.random.random((10,20)), b:np.random.random((20,10)), c:np.random.random((10,10))}))\n<\/code><\/pre>\n\n<p>I run this program with the following <code>XLA_FLAGS<\/code>:<\/p>\n\n<pre><code>XLA_FLAGS=\"--xla_generate_hlo_text_to=.\/path1 --xla_dump_ir_to=.\/path2\" python source.py\n<\/code><\/pre>\n\n<p>The <code>.\/path1<\/code> directory includes all the <code>HLO<\/code> passes outputs. But the <code>.ll<\/code> and <code>.ptx<\/code> files in <code>.\/path2<\/code> directory does not include the <code>ptx<\/code> or <code>llvm IR<\/code> code. The <code>.ptx<\/code>file is empty and the <code>ll<\/code> files have just the following lines:<\/p>\n\n<pre><code>; ModuleID = 'cluster_0__XlaCompiledKernel_true__XlaNumConstantArgs_0__XlaNumResourceArgs_0_.12'\nsource_filename = \"cluster_0__XlaCompiledKernel_true__XlaNumConstantArgs_0__XlaNumResourceArgs_0_.12\"\ntarget datalayout = \"e-i64:64-i128:128-v16:16-v32:32-n16:32:64\"\ntarget triple = \"nvptx64-nvidia-cuda\"\n<\/code><\/pre>\n\n<p>My questions is why I cannot see the <code>ptx<\/code> code for the mentioned tensorflow code? <\/p>\n\n<p>I should mention that, If I add the <code>d = d + d<\/code> line to the python code above after the <code>a * b + c<\/code>, <code>XLA<\/code> generates some <code>PTX<\/code> code which is just the adding operation and it does not include the matrix multiplication code.<\/p>\n","answers":[]}
+{"QuestionId":57828416,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2019-09-06T20:49:12.120","AcceptedAnswerId":57828786.0,"OwnerUserId":12032109.0,"Title":"How to use GradientTape with AutoGraph in Tensorflow 2?","Body":"<p>I can not figure out how to run GradientTape code on AutoGraph in Tensorflow 2.<\/p>\n\n<p>I want to run GradientTape code on TPU. I wanted to start by testing it on CPU. TPU code would run much faster using AutoGraph. I tried watching the input variable and I tried passing in the argument into the function that contained GradientTape but both failed.<\/p>\n\n<p>I made a reproducible example here:\n<a href=\"https:\/\/colab.research.google.com\/drive\/1luCk7t5SOcDHQC6YTzJzpSixIqHy_76b#scrollTo=9OQUYQtTTYIt\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/1luCk7t5SOcDHQC6YTzJzpSixIqHy_76b#scrollTo=9OQUYQtTTYIt<\/a><\/p>\n\n<p>The code and the corresponding output is as follows:\nThey all start with <code>import tensorflow as tf<\/code><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = tf.constant(3.0)\nwith tf.GradientTape() as g:\n  g.watch(x)\n  y = x * x\ndy_dx = g.gradient(y, x)\nprint(dy_dx)\n<\/code><\/pre>\n\n<p>Output: <code>tf.Tensor(6.0, shape=(), dtype=float32)<\/code>\nExplanation: Using Eager Execution, the GradientTape produces the gradient.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>@tf.function\ndef compute_me():\n    x = tf.constant(3.0)\n    with tf.GradientTape() as g:\n      g.watch(x)\n      y = x * x\n    dy_dx = g.gradient(y, x) # Will compute to 6.0\n    print(dy_dx)\ncompute_me()\n<\/code><\/pre>\n\n<p>Output: <code>Tensor(\"AddN:0\", shape=(), dtype=float32)<\/code>\nExplanation: Using AutoGraph on GradientTape in TF2 results in empty gradient<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>@tf.function\ndef compute_me_args(x):\n    with tf.GradientTape() as g:\n      g.watch(x)\n      y = x * x\n    dy_dx = g.gradient(y, x) # Will compute to 6.0\n    print(dy_dx)    \nx = tf.constant(3.0)\ncompute_me_args(x)\n<\/code><\/pre>\n\n<p>Output: <code>Tensor(\"AddN:0\", shape=(), dtype=float32)<\/code>\nExplanation: Passing in arguments also fails<\/p>\n\n<p>I was expecting all of the cells to output <code>tf.Tensor(6.0, shape=(), dtype=float32)<\/code> but instead the cells using AutoGraph output <code>Tensor(\"AddN:0\", shape=(), dtype=float32)<\/code>.<\/p>\n","answers":[{"AnswerId":"57828786","CreationDate":"2019-09-06T21:36:22.790","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>It doesn't \"fail\", it's just that <code>print<\/code>, if used in the context of a <code>tf.function<\/code> (i.e. in graph mode) will print the symbolic tensors, and these do not have a value. Try this instead:<\/p>\n\n<pre><code>@tf.function\ndef compute_me():\n    x = tf.constant(3.0)\n    with tf.GradientTape() as g:\n        g.watch(x)\n        y = x * x\n    dy_dx = g.gradient(y, x) # Will compute to 6.0\n    tf.print(dy_dx)\ncompute_me()\n<\/code><\/pre>\n\n<p>This should print <code>6<\/code>. All you need to do is use <code>tf.print<\/code> instead, which is \"smart\" enough to print the actual values if available. Or, using return values:<\/p>\n\n<pre><code>@tf.function\ndef compute_me():\n    x = tf.constant(3.0)\n    with tf.GradientTape() as g:\n        g.watch(x)\n        y = x * x\n    dy_dx = g.gradient(y, x) # Will compute to 6.0\n    return dy_dx\nresult = compute_me()\nprint(result)\n<\/code><\/pre>\n\n<p>outputs something like <code>&lt;tf.Tensor: id=43, shape=(), dtype=float32, numpy=6.0&gt;<\/code>. You can see that the value (6.0) is visible here as well. Use <code>result.numpy()<\/code> to just get <code>6.0<\/code>.<\/p>\n"}]}
+{"QuestionId":57829494,"AnswerCount":0,"Tags":"<python><tensorflow><keras><h5py>","CreationDate":"2019-09-06T23:22:31.117","AcceptedAnswerId":null,"OwnerUserId":4001750.0,"Title":"How to install tensorflow in 64-bit Windows 10?","Body":"<p>I have Python version 3.7 64-bit in Windows 10 (64-bit). I tried installing both tensorflow and keras but I end up with the following error all the time:<\/p>\n\n<pre><code>Collecting tensorflow\n  Downloading https:\/\/files.pythonhosted.org\/packages\/f7\/08\/25e47a53692c2e0dcd2211a493ddfe9007a5cd92e175d6dffa6169a0b392\/tensorflow-1.14.0-cp37-cp37m-win_amd64.whl (68.3MB)\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| 68.3MB 109kB\/s\nRequirement already satisfied: numpy&lt;2.0,&gt;=1.14.5 in c:\\users\\naseef\\appdata\\local\\programs\\python\\python37\\lib\\site-packages (from tensorflow) (1.16.4)\nCollecting keras-preprocessing&gt;=1.0.5 (from tensorflow)\n  Using cached https:\/\/files.pythonhosted.org\/packages\/28\/6a\/8c1f62c37212d9fc441a7e26736df51ce6f0e38455816445471f10da4f0a\/Keras_Preprocessing-1.1.0-py2.py3-none-any.whl\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\nCollecting tensorflow-estimator&lt;1.15.0rc0,&gt;=1.14.0rc0 (from tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/3c\/d5\/21860a5b11caf0678fbc8319341b0ae21a07156911132e0e71bffed0510d\/tensorflow_estimator-1.14.0-py2.py3-none-any.whl (488kB)\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| 491kB 133kB\/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 keras-applications&gt;=1.0.6 (from tensorflow)\n  Using cached https:\/\/files.pythonhosted.org\/packages\/71\/e3\/19762fdfc62877ae9102edf6342d71b28fbfd9dea3d2f96a882ce099b03f\/Keras_Applications-1.0.8-py3-none-any.whl\nCollecting protobuf&gt;=3.6.1 (from tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/46\/8b\/5e77963dac4a944a0c6b198c004fac4c85d7adc54221c288fc6ca9078072\/protobuf-3.9.1-cp37-cp37m-win_amd64.whl (1.0MB)\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| 1.0MB 123kB\/s\nRequirement already satisfied: six&gt;=1.10.0 in c:\\users\\naseef\\appdata\\local\\programs\\python\\python37\\lib\\site-packages (from tensorflow) (1.12.0)\nCollecting absl-py&gt;=0.7.0 (from tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/3c\/0d\/7cbf64cac3f93617a2b6b079c0182e4a83a3e7a8964d3b0cc3d9758ba002\/absl-py-0.8.0.tar.gz (102kB)\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| 112kB 2.2MB\/s\nCollecting wheel&gt;=0.26 (from tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/00\/83\/b4a77d044e78ad1a45610eb88f745be2fd2c6d658f9798a15e384b7d57c9\/wheel-0.33.6-py2.py3-none-any.whl\nRequirement already satisfied: wrapt&gt;=1.11.1 in c:\\users\\naseef\\appdata\\roaming\\python\\python37\\site-packages (from tensorflow) (1.11.1)\nCollecting termcolor&gt;=1.1.0 (from tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/8a\/48\/a76be51647d0eb9f10e2a4511bf3ffb8cc1e6b14e9e4fab46173aa79f981\/termcolor-1.1.0.tar.gz\nCollecting grpcio&gt;=1.8.6 (from tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/32\/e7\/478737fd426798caad32a2abb7cc63ddb4c12908d9e03471dd3c41992b05\/grpcio-1.23.0-cp37-cp37m-win_amd64.whl (1.6MB)\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| 1.6MB 114kB\/s\nCollecting tensorboard&lt;1.15.0,&gt;=1.14.0 (from tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/91\/2d\/2ed263449a078cd9c8a9ba50ebd50123adf1f8cfbea1492f9084169b89d9\/tensorboard-1.14.0-py3-none-any.whl (3.1MB)\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 117kB\/s\nCollecting google-pasta&gt;=0.1.6 (from tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/d0\/33\/376510eb8d6246f3c30545f416b2263eee461e40940c2a4413c711bdf62d\/google_pasta-0.1.7-py3-none-any.whl (52kB)\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| 61kB ...\nCollecting h5py (from keras-applications&gt;=1.0.6-&gt;tensorflow)\n  Using cached https:\/\/files.pythonhosted.org\/packages\/5f\/97\/a58afbcf40e8abecededd9512978b4e4915374e5b80049af082f49cebe9a\/h5py-2.10.0.tar.gz\nRequirement already satisfied: setuptools in c:\\users\\naseef\\appdata\\local\\programs\\python\\python37\\lib\\site-packages (from protobuf&gt;=3.6.1-&gt;tensorflow) (41.2.0)\nCollecting werkzeug&gt;=0.11.15 (from tensorboard&lt;1.15.0,&gt;=1.14.0-&gt;tensorflow)\n  Downloading https:\/\/files.pythonhosted.org\/packages\/b7\/61\/c0a1adf9ad80db012ed7191af98fa05faa95fa09eceb71bb6fa8b66e6a43\/Werkzeug-0.15.6-py2.py3-none-any.whl (328kB)\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| 337kB 59kB\/s\nCollecting markdown&gt;=2.6.8 (from tensorboard&lt;1.15.0,&gt;=1.14.0-&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 137kB\/s\nInstalling collected packages: keras-preprocessing, astor, tensorflow-estimator, gast, h5py, keras-applications, protobuf, absl-py, wheel, termcolor, grpcio, werkzeug, markdown, tensorboard, google-pasta, tensorflow\n  Running setup.py install for gast ... done\n  Running setup.py install for h5py ... error\n    ERROR: Complete output from command 'C:\\Users\\Naseef\\AppData\\Local\\Programs\\Python\\Python37\\python.exe' -u -c 'import setuptools, tokenize;__file__='\"'\"'C:\\\\Users\\\\Naseef\\\\AppData\\\\Local\\\\Temp\\\\pip-install-b3tjfyss\\\\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 'C:\\Users\\Naseef\\AppData\\Local\\Temp\\pip-record-lnidw34r\\install-record.txt' --single-version-externally-managed --compile:\n    ERROR: running install\n    running build\n    running build_py\n    creating build\n    creating build\\lib.win-amd64-3.7\n    creating build\\lib.win-amd64-3.7\\h5py\n    copying h5py\\h5py_warnings.py -&gt; build\\lib.win-amd64-3.7\\h5py\n    copying h5py\\highlevel.py -&gt; build\\lib.win-amd64-3.7\\h5py\n    copying h5py\\ipy_completer.py -&gt; build\\lib.win-amd64-3.7\\h5py\n    copying h5py\\version.py -&gt; build\\lib.win-amd64-3.7\\h5py\n    copying h5py\\__init__.py -&gt; build\\lib.win-amd64-3.7\\h5py\n    creating build\\lib.win-amd64-3.7\\h5py\\_hl\n    copying h5py\\_hl\\attrs.py -&gt; build\\lib.win-amd64-3.7\\h5py\\_hl\n    copying h5py\\_hl\\base.py -&gt; build\\lib.win-amd64-3.7\\h5py\\_hl\n    copying h5py\\_hl\\compat.py -&gt; build\\lib.win-amd64-3.7\\h5py\\_hl\n    copying h5py\\_hl\\dataset.py -&gt; build\\lib.win-amd64-3.7\\h5py\\_hl\n    copying h5py\\_hl\\datatype.py -&gt; build\\lib.win-amd64-3.7\\h5py\\_hl\n    copying h5py\\_hl\\dims.py -&gt; build\\lib.win-amd64-3.7\\h5py\\_hl\n    copying h5py\\_hl\\files.py -&gt; build\\lib.win-amd64-3.7\\h5py\\_hl\n    copying h5py\\_hl\\filters.py -&gt; build\\lib.win-amd64-3.7\\h5py\\_hl\n    copying h5py\\_hl\\group.py -&gt; build\\lib.win-amd64-3.7\\h5py\\_hl\n    copying h5py\\_hl\\selections.py -&gt; build\\lib.win-amd64-3.7\\h5py\\_hl\n    copying h5py\\_hl\\selections2.py -&gt; build\\lib.win-amd64-3.7\\h5py\\_hl\n    copying h5py\\_hl\\vds.py -&gt; build\\lib.win-amd64-3.7\\h5py\\_hl\n    copying h5py\\_hl\\__init__.py -&gt; build\\lib.win-amd64-3.7\\h5py\\_hl\n    creating build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\common.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_attribute_create.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_attrs.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_attrs_data.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_base.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_completions.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_dataset.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_dataset_getitem.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_dataset_swmr.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_datatype.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_deprecation.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_dimension_scales.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_dims_dimensionproxy.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_dtype.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_file.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_file2.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_file_image.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_filters.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_group.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_h5.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_h5d_direct_chunk.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_h5f.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_h5p.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_h5pl.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_h5t.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_objects.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_selections.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_slicing.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\test_threads.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    copying h5py\\tests\\__init__.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\n    creating build\\lib.win-amd64-3.7\\h5py\\tests\\test_vds\n    copying h5py\\tests\\test_vds\\test_highlevel_vds.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\\test_vds\n    copying h5py\\tests\\test_vds\\test_lowlevel_vds.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\\test_vds\n    copying h5py\\tests\\test_vds\\test_virtual_source.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\\test_vds\n    copying h5py\\tests\\test_vds\\__init__.py -&gt; build\\lib.win-amd64-3.7\\h5py\\tests\\test_vds\n    running build_ext\n    error: [WinError 126] The specified module could not be found\n    Loading library to get version: hdf5.dll\n    ----------------------------------------\nERROR: Command \"'C:\\Users\\Naseef\\AppData\\Local\\Programs\\Python\\Python37\\python.exe' -u -c 'import setuptools, tokenize;__file__='\"'\"'C:\\\\Users\\\\Naseef\\\\AppData\\\\Local\\\\Temp\\\\pip-install-b3tjfyss\\\\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 'C:\\Users\\Naseef\\AppData\\Local\\Temp\\pip-record-lnidw34r\\install-record.txt' --single-version-externally-managed --compile\" failed with error code 1 in C:\\Users\\Naseef\\AppData\\Local\\Temp\\pip-install-b3tjfyss\\h5py\\\n<\/code><\/pre>\n\n<p>Even installation of h5py using the following command > pu -m pip install h5py did not succeed. Is it a version conflict or should I be installing them from their source?<\/p>\n","answers":[]}
+{"QuestionId":57829658,"AnswerCount":1,"Tags":"<python><opencv><tensorflow><gstreamer><nvidia-jetson>","CreationDate":"2019-09-06T23:57:42.027","AcceptedAnswerId":null,"OwnerUserId":12032787.0,"Title":"How to fix \"TypeError: int() argument must be a string, a bytes-like object or a number, not 'NoneType'\"?","Body":"<p>My aim is streaming object detection video from Raspberry Pi camera to UDP. I am doing that on Nvidia Jetson Nano(Ubuntu 18.04). I am using gstreamer to capture video. I have to encode video with h264. Pixel Format of camera is 'RG10'. I tried different pipeline to capture video but nothing is changed. I think, problem is capturing video. I could not figure out. I am newbies. So I need help. <\/p>\n\n<p>Here is code:<\/p>\n\n<pre><code>\nimport os\nimport numpy as np\nimport tensorflow as tf\nimport sys\n\nsys.path.append('\/usr\/local\/lib\/python3.6\/dist-packages')\nimport cv2 as cv2\n\nfrom utils import label_map_util\nfrom utils import visualization_utils as vis_util\n\nMODEL_NAME = 'inference_graph'\nCWD_PATH = os.getcwd()\nPATH_TO_CKPT = os.path.join(CWD_PATH,MODEL_NAME,'frozen_inference_graph.pb')\nPATH_TO_LABELS = os.path.join(CWD_PATH,'training','labelmap.pbtxt')\nNUM_CLASSES = 1\n\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\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    sess = tf.Session(graph=detection_graph)\n\nimage_tensor = detection_graph.get_tensor_by_name('image_tensor:0')\n\ndetection_boxes = detection_graph.get_tensor_by_name('detection_boxes:0')\ndetection_scores = detection_graph.get_tensor_by_name('detection_scores:0')\ndetection_classes = detection_graph.get_tensor_by_name('detection_classes:0')\nnum_detections = detection_graph.get_tensor_by_name('num_detections:0')\n\nwriter = cv2.VideoWriter(\" appsrc name=appsrc_element block=true ! video\/x-raw,format=RGB,width=320,height=240,framerate=30\/1 ! identity check-imperfect-timestamp=true ! videoconvert ! x264enc ! video\/x-h264,profile=\\\"high-4:4:4\\ ! rtph264pay  ! udpsink host= 192.168.1.49 port=5000 \",cv2.CAP_GSTREAMER,0, 20, (320,240), True) # Write frames here after processing\n\n\nvideo = cv2.VideoCapture('nvarguscamerasrc ! image\/jpg,width=1280,height=720,type=video,framerate=30\/1 ! videoscale ! videoconvert ! x264enc tune=zerolatency ! rtph264pay ! appsink ', cv2.CAP_GSTREAMER)\nret = video.set(3,1280)\nret = video.set(4,720)\n\nwhile(True):\n\n\n    ret, frame = video.read()   \n    frame_expanded = np.expand_dims(frame, axis=0)\n\n\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    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    w_size = cv2.getWindowImageRect('frame')\n\n    print(w_size)\n    cv2.rectangle(frame,(int(int(w_size[2])*0.25) ,int(int(w_size[3])*0.10)),(int(int(w_size[2])*0.75),int(int(w_size[3])*0.90)),(0,255,0),3)\n\n    writer.write(frame) \n\n\n    if cv2.waitKey(1) == ord('q'):\n        break\n\nvideo.release()\nwriter.release()\ncv2.destroyAllWindows()\n<\/code><\/pre>\n\n<p>This is output error :<\/p>\n\n<pre><code>  _np_qint8 = np.dtype([(\"qint8\", np.int8, 1)])\n\/home\/berfu\/.virtualenvs\/deep_learning\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/dtypes.py:527: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint8 = np.dtype([(\"quint8\", np.uint8, 1)])\n\/home\/berfu\/.virtualenvs\/deep_learning\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/dtypes.py:528: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint16 = np.dtype([(\"qint16\", np.int16, 1)])\n\/home\/berfu\/.virtualenvs\/deep_learning\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/dtypes.py:529: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint16 = np.dtype([(\"quint16\", np.uint16, 1)])\n\/home\/berfu\/.virtualenvs\/deep_learning\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/dtypes.py:530: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint32 = np.dtype([(\"qint32\", np.int32, 1)])\n\/home\/berfu\/.virtualenvs\/deep_learning\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/dtypes.py:535: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  np_resource = np.dtype([(\"resource\", np.ubyte, 1)])\n2019-09-07 02:26:33.800624: W tensorflow\/core\/platform\/profile_utils\/cpu_utils.cc:98] Failed to find bogomips in \/proc\/cpuinfo; cannot determine CPU frequency\n2019-09-07 02:26:33.804717: I tensorflow\/compiler\/xla\/service\/service.cc:161] XLA service 0x2816fdf0 executing computations on platform Host. Devices:\n2019-09-07 02:26:33.805320: I tensorflow\/compiler\/xla\/service\/service.cc:168]   StreamExecutor device (0): &lt;undefined&gt;, &lt;undefined&gt;\n2019-09-07 02:26:33.895575: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:965] ARM64 does not support NUMA - returning NUMA node zero\n2019-09-07 02:26:33.895923: I tensorflow\/compiler\/xla\/service\/service.cc:161] XLA service 0x24aca990 executing computations on platform CUDA. Devices:\n2019-09-07 02:26:33.895984: I tensorflow\/compiler\/xla\/service\/service.cc:168]   StreamExecutor device (0): NVIDIA Tegra X1, Compute Capability 5.3\n2019-09-07 02:26:33.896452: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1433] Found device 0 with properties: \nname: NVIDIA Tegra X1 major: 5 minor: 3 memoryClockRate(GHz): 0.9216\npciBusID: 0000:00:00.0\ntotalMemory: 3.86GiB freeMemory: 272.39MiB\n2019-09-07 02:26:33.896515: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1512] Adding visible gpu devices: 0\n2019-09-07 02:26:35.073590: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:984] Device interconnect StreamExecutor with strength 1 edge matrix:\n2019-09-07 02:26:35.073769: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:990]      0 \n2019-09-07 02:26:35.073809: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1003] 0:   N \n2019-09-07 02:26:35.074121: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1115] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 42 MB memory) -&gt; physical GPU (device: 0, name: NVIDIA Tegra X1, pci bus id: 0000:00:00.0, compute capability: 5.3)\nGST_ARGUS: Creating output stream\nCONSUMER: Waiting until producer is connected...\nGST_ARGUS: Available Sensor modes :\nGST_ARGUS: 3280 x 2464 FR = 21.000000 fps Duration = 47619048 ; Analog Gain range min 1.000000, max 10.625000; Exposure Range min 13000, max 683709000;\n\nGST_ARGUS: 3280 x 1848 FR = 28.000001 fps Duration = 35714284 ; Analog Gain range min 1.000000, max 10.625000; Exposure Range min 13000, max 683709000;\n\nGST_ARGUS: 1920 x 1080 FR = 29.999999 fps Duration = 33333334 ; Analog Gain range min 1.000000, max 10.625000; Exposure Range min 13000, max 683709000;\n\nGST_ARGUS: 1280 x 720 FR = 59.999999 fps Duration = 16666667 ; Analog Gain range min 1.000000, max 10.625000; Exposure Range min 13000, max 683709000;\n\nGST_ARGUS: 1280 x 720 FR = 120.000005 fps Duration = 8333333 ; Analog Gain range min 1.000000, max 10.625000; Exposure Range min 13000, max 683709000;\n\nGST_ARGUS: Running with following settings:\n   Camera index = 0 \n   Camera mode  = 2 \n   Output Stream W = 1920 H = 1080 \n   seconds to Run    = 0 \n   Frame Rate = 29.999999 \nGST_ARGUS: PowerService: requested_clock_Hz=13608000\nGST_ARGUS: Setup Complete, Starting captures for 0 seconds\nGST_ARGUS: Starting repeat capture requests.\nCONSUMER: Producer has connected; continuing.\nGST_ARGUS: Cleaning up\nGST_ARGUS: \nPowerServiceHwVic::cleanupResources\nCONSUMER: Done Success\nGST_ARGUS: Done Success\n&lt;class 'NoneType'&gt;\nTraceback (most recent call last):\n  File \"webcam.py\", line 68, in &lt;module&gt;\n    feed_dict={image_tensor: frame_expanded})\n  File \"\/home\/berfu\/.virtualenvs\/deep_learning\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 929, in run\n    run_metadata_ptr)\n  File \"\/home\/berfu\/.virtualenvs\/deep_learning\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1121, in _run\n    np_val = np.asarray(subfeed_val, dtype=subfeed_dtype)\n  File \"\/home\/berfu\/.virtualenvs\/deep_learning\/lib\/python3.6\/site-packages\/numpy\/core\/_asarray.py\", line 85, in asarray\n    return array(a, dtype, copy=False, order=order)\nTypeError: int() argument must be a string, a bytes-like object or a number, not 'NoneType'\n<\/code><\/pre>\n\n<p>Thank you, for your help<\/p>\n","answers":[{"AnswerId":"57829687","CreationDate":"2019-09-07T00:04:55.457","ParentId":null,"OwnerUserId":"11746212","Title":null,"Body":"<p>The values in <code>feed_dict<\/code> have to be convertible to a numpy array - it looks like <code>frame_expanded<\/code> may contain <code>None<\/code>. Try printing the <code>feed_dict<\/code> values to see if it contains valid data.<\/p>\n"}]}
+{"QuestionId":57829922,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-07T01:04:02.257","AcceptedAnswerId":null,"OwnerUserId":11765076.0,"Title":"TF Lite GPU Delegator: OpenGL Vs OpenCL","Body":"<p>Has anyone done any comparison analysis of Tensorflow lite GPU Delegator for OpenGL &amp; OpenCL? I would like to talk about findings.\nRef: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/f1b58a9c2cfbfb661440ef21d625361746aa0e10\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/commit\/f1b58a9c2cfbfb661440ef21d625361746aa0e10<\/a><\/p>\n","answers":[]}
+{"QuestionId":57830251,"AnswerCount":1,"Tags":"<python><tensorflow><keras><neural-network><keras-layer>","CreationDate":"2019-09-07T02:34:05.147","AcceptedAnswerId":null,"OwnerUserId":4570144.0,"Title":"Keras Minimum() causing AttributeError: 'NoneType' object has no attribute '_inbound_nodes'","Body":"<p>What is going on, and how do I correctly use Keras' Minimum() layer?<\/p>\n\n<pre><code>start = Input(shape=(28,28,1), dtype='float32')\nc = Conv2D(20,3,use_bias=False,data_format=\"channels_last\",input_shape=(28, 28, 1),padding='same')(start)\nc = Reshape((28,28,20,1))(c)\nm = MaxPooling3D(pool_size=(1,1,20), strides=1, padding='valid', data_format=\"channels_last\")(c)\nm = Reshape((28,28,1))(m)\nm = Minimum()([m,mask])\nprint(m.shape, mask.shape)\nf = Flatten()(m)\nprint(f.shape)\nout = Dense(10,activation='softmax')(f)\nmodel = Model(inputs=start, outputs=out)\n<\/code><\/pre>\n\n<p><strong>output:<\/strong><\/p>\n\n<pre><code>(60000, 28, 28, 1) (60000, 28, 28, 1)\n(60000, 784)\n\nAttributeError                            Traceback (most recent call last)\n&lt;ipython-input-21-e0fab99c12c2&gt; in &lt;module&gt;()\n      9 print(f.shape)\n     10 out = Dense(10,activation='softmax')(f)\n---&gt; 11 model = Model(inputs=start, outputs=out)\n\nAttributeError: 'NoneType' object has no attribute '_inbound_nodes'\n<\/code><\/pre>\n\n<p>The keras API says that Minimum() is:<\/p>\n\n<blockquote>\n  <p>Layer that computes the minimum (element-wise) a list of inputs.<\/p>\n  \n  <p>It takes as input a list of tensors, all of the same shape, \n  and returns a single tensor (also of the same shape).<\/p>\n<\/blockquote>\n\n<p>So it seems like I am using it correctly to me... but I am probably doing something very stupid.<\/p>\n","answers":[{"AnswerId":"57830403","CreationDate":"2019-09-07T03:19:22.020","ParentId":null,"OwnerUserId":"4570144","Title":null,"Body":"<p>I solved the problem.  It's my mistake.\nI was trying to pass mask to minimum(), but mask was a numpy array and minimum() needs a Keras Input()<\/p>\n"}]}
+{"QuestionId":57830439,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><machine-learning><gradient-descent>","CreationDate":"2019-09-07T03:26:42.967","AcceptedAnswerId":57832408.0,"OwnerUserId":9557970.0,"Title":"gradient descent using tensors calculating wrong values","Body":"<p>I am implementing simple gradient descent algorithm using tensors. It learns two parameters <strong>m<\/strong> and <strong>c<\/strong>.<br>\n The normal python code for it is :<\/p>\n\n<pre><code>for i in range(epochs): \n    Y_pred = m*X + c  # The current predicted value of Y\n    D_m = (-2\/n) * sum(X * (Y - Y_pred))  # Derivative wrt m\n    D_c = (-2\/n) * sum(Y - Y_pred)  # Derivative wrt c\n    m = m - L * D_m  # Update m\n    c = c - L * D_c  # Update c\n    print (m, c) \n<\/code><\/pre>\n\n<p>output for python :<\/p>\n\n<pre><code>0.7424335285442664 0.014629895049575754\n1.1126970531591416 0.021962519495058154\n1.2973530613155333 0.025655870599552183\n1.3894434413955663 0.027534253868790198\n1.4353697670010162 0.028507481513901086\n<\/code><\/pre>\n\n<p>Tensorflow equivalent code :<\/p>\n\n<pre><code>#Graph of gradient descent\ny_pred = m*x + c\nd_m = (-2\/n) * tf.reduce_sum(x*(y-y_pred)) \nd_c = (-2\/n) * tf.reduce_sum(y-y_pred)  \nupm = tf.assign(m, m - learning_rate * d_m)\nupc = tf.assign(c, c - learning_rate * d_c)\n\n#starting session\nsess = tf.Session()\n\n#Training for epochs\nfor i in range(epochs):\n    sess.run(y_pred)\n    sess.run(d_m)\n    sess.run(d_c)\n    sess.run(upm)\n    sess.run(upc)\n    w = sess.run(m)\n    b = sess.run(c)\n    print(w,b)\n<\/code><\/pre>\n\n<p>Output for tensorflow :<\/p>\n\n<pre><code>0.7424335285442664 0.007335550424492317\n1.1127687194584988 0.011031122807663662\n1.2974962163433057 0.012911024540805463\n1.3896400798226038 0.013885244876397126\n1.4356019721347115 0.014407698787092268\n<\/code><\/pre>\n\n<p>The parameter <strong>m has the same value<\/strong> for both but <strong>parameter c has different value<\/strong> for both although the implementation is same for both.<br>\nThe output contains first 5 values of parameter m and c. The output of parameter c using tensors is approximately half of the normal python.<br>\nI don't know where my mistake is.<\/p>\n\n<p>For recreating the entire output:\n<a href=\"https:\/\/github.com\/aniketbote\/learning-tensorflow\/blob\/master\/Examples%20low%20level%20api\/uni-variate_linear_regression_tensors.py\" rel=\"nofollow noreferrer\">Repo containing data along with both implementations<\/a><\/p>\n\n<p>The repo also contains image of graph obtained through tensorboard in events directory<\/p>\n","answers":[{"AnswerId":"57832408","CreationDate":"2019-09-07T09:30:50.893","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>The problem is that, in the TF implementation, the updates are not being performed atomically. In other words, the implementation of the algorithm is updating <code>m<\/code> and <code>c<\/code> in an interleaved manner (e.g. the new value of <code>m<\/code> is being used when updating <code>c<\/code>). To make the updates atomic, you should simultaneously run <code>upm<\/code> and <code>upc<\/code>:<\/p>\n\n<pre><code>sess.run([upm, upc])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57830663,"AnswerCount":0,"Tags":"<keras><google-colaboratory>","CreationDate":"2019-09-07T04:24:19.130","AcceptedAnswerId":null,"OwnerUserId":7934751.0,"Title":"Keras model.save and load_model can't save optimizer state correctly in google-colab","Body":"<p>It seems that keras can't save optimizer correctly in google colab. After save and reload a model, the loss of reloaded model is quite different to original model.<\/p>\n\n<p>Although the Documentation of keras says model.save will save the state of the optimizer, are there some special cases that it won't save the entire state of the optimizer?<\/p>\n\n<p>keras==2.2.5<\/p>\n\n<p>backend==tensorflow<\/p>\n\n<p>tensorflow==1.14.0<\/p>\n\n<p>Here is my code.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from sklearn.model_selection import train_test_split\nx_train, x_valid, y_train, y_valid = train_test_split(X, Y, test_size=0.3, shuffle= True)\n\nimport keras.optimizers as ko\noptimizer = ko.Nadam()\nmodel.compile(loss='mean_squared_error', optimizer=optimizer, metrics=['accuracy'])\nhistory = model.fit(x_train, y_train, epochs=500, validation_data=(x_valid, y_valid), batch_size=64)\n#output\n\"\"\"\nEpoch 495\/500\n1750\/1750 [==============================] - 16s 9ms\/step - loss: 0.9285 - acc: 1.0000 - val_loss: 4.0664 - val_acc: 1.0000\nEpoch 496\/500\n1750\/1750 [==============================] - 16s 9ms\/step - loss: 1.8576 - acc: 1.0000 - val_loss: 1.7988 - val_acc: 1.0000\nEpoch 497\/500\n1750\/1750 [==============================] - 16s 9ms\/step - loss: 0.8660 - acc: 1.0000 - val_loss: 1.4999 - val_acc: 1.0000\nEpoch 498\/500\n1750\/1750 [==============================] - 16s 9ms\/step - loss: 0.5575 - acc: 1.0000 - val_loss: 0.6627 - val_acc: 1.0000\nEpoch 499\/500\n1750\/1750 [==============================] - 16s 9ms\/step - loss: 0.4687 - acc: 1.0000 - val_loss: 1.3086 - val_acc: 1.0000\nEpoch 500\/500\n1750\/1750 [==============================] - 16s 9ms\/step - loss: 0.9433 - acc: 1.0000 - val_loss: 5.9769 - val_acc: 1.0000\n\"\"\"\n\nimport pickle\nwith open('\/gdrive\/My Drive\/lab\/2019_9_4\/history500.pkl', 'wb') as output:\n  pickle.dump(history, output, pickle.HIGHEST_PROTOCOL)\nmodel.save(\"\/gdrive\/My Drive\/lab\/2019_9_4\/abdp.h5\")\nwith open('\/gdrive\/My Drive\/lab\/2019_9_4\/x_train.pkl', 'wb') as output:\n  pickle.dump(x_train, output, pickle.HIGHEST_PROTOCOL)\nwith open('\/gdrive\/My Drive\/lab\/2019_9_4\/x_valid.pkl', 'wb') as output:\n  pickle.dump(x_valid, output, pickle.HIGHEST_PROTOCOL)\nwith open('\/gdrive\/My Drive\/lab\/2019_9_4\/y_train.pkl', 'wb') as output:\n  pickle.dump(y_train, output, pickle.HIGHEST_PROTOCOL)\nwith open('\/gdrive\/My Drive\/lab\/2019_9_4\/y_valid.pkl', 'wb') as output:\n  pickle.dump(y_valid, output, pickle.HIGHEST_PROTOCOL)\n\nfrom keras.models import load_model\nmodel = load_model(\"\/gdrive\/My Drive\/lab\/2019_9_4\/abdp.h5\")\nwith open('\/gdrive\/My Drive\/lab\/2019_9_4\/x_train.pkl', 'rb') as f:\n  x_train = pickle.load(f)\nwith open('\/gdrive\/My Drive\/lab\/2019_9_4\/x_valid.pkl', 'rb') as f:\n  x_valid = pickle.load(f)\nwith open('\/gdrive\/My Drive\/lab\/2019_9_4\/y_train.pkl', 'rb') as f:\n  y_train = pickle.load(f)\nwith open('\/gdrive\/My Drive\/lab\/2019_9_4\/y_valid.pkl', 'rb') as f:\n  y_valid = pickle.load(f)\n\nhistory = model.fit(x_train, y_train, epochs=5, validation_data=(x_valid, y_valid), batch_size=64)\n\"\"\"\nEpoch 1\/5\n1750\/1750 [==============================] - 32s 18ms\/step - loss: 3671.9919 - acc: 1.0000 - val_loss: 27511.0797 - val_acc: 1.0000\nEpoch 2\/5\n1750\/1750 [==============================] - 29s 16ms\/step - loss: 327.5412 - acc: 1.0000 - val_loss: 3598.9627 - val_acc: 1.0000\nEpoch 3\/5\n1750\/1750 [==============================] - 29s 16ms\/step - loss: 216.6781 - acc: 1.0000 - val_loss: 1258.5962 - val_acc: 1.0000\nEpoch 4\/5\n1750\/1750 [==============================] - 29s 16ms\/step - loss: 141.3966 - acc: 1.0000 - val_loss: 1526.8668 - val_acc: 1.0000\nEpoch 5\/5\n1750\/1750 [==============================] - 29s 16ms\/step - loss: 114.3413 - acc: 1.0000 - val_loss: 1107.3428 - val_acc: 1.0000\n\"\"\"\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57830896,"AnswerCount":1,"Tags":"<python><django><tensorflow>","CreationDate":"2019-09-07T05:16:43.047","AcceptedAnswerId":57831132.0,"OwnerUserId":11369234.0,"Title":"Models trained with TensorFlow are not available in Django","Body":"<p>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<p>Key bias_1 not found in checkpoint\n     [[node save_1\/RestoreV2 (defined at Programing\\web_programing\\django\\django-vegiter\\predict\\views.py:20) ]]<\/p>\n\n<p>Trace Back<\/p>\n\n<pre><code>saver.restore(sess, save_path) \n\u2026\nerr, \"a Variable name or other graph key that is missing\") \n<\/code><\/pre>\n\n<p>Run in Python virtual environment<\/p>\n\n<p>Django               2.2.5\nKeras-Applications   1.0.8\nKeras-Preprocessing  1.1.0\nnumpy                1.17.1\npandas               0.25.1\ntensorboard          1.14.0\ntensorflow           1.14.0\ntensorflow-estimator 1.14.0\ntermcolor            1.1.0<\/p>\n\n<h1>Create your views here.<\/h1>\n\n<p>def index(request):\n    X = tf.placeholder(tf.float32, shape=[None, 4])\n    Y = tf.placeholder(tf.float32, shape=[None, 1])<\/p>\n\n<pre><code>W = tf.Variable(tf.random_normal([4, 1]), name=\"weight\")\nb = tf.Variable(tf.random_normal([1]), name=\"bias\")\n\nhypothesis = tf.matmul(X, W) + b\n\nsaver = tf.train.Saver()\nmodel = tf.global_variables_initializer()\n\nsess = tf.Session()\nsess.run(model)\nsave_path = \".\/model\/saved.cpkt\"\nsaver.restore(sess, save_path)\n\nif request.method == \"POST\":\n    avg_temp = float(request.POST['avg_temp'])\n    min_temp = float(request.POST['min_temp'])\n    max_temp = float(request.POST['max_temp'])\n    rain_fall = float(request.POST['rain_fall'])\n\n    price = 0\n\n    data = ((avg_temp, min_temp, max_temp, rain_fall), (0, 0, 0, 0))\n    arr = np.array(data, dtype=np.float32)\n\n    x_data = arr[0:4]\n    dict = sess.run(hypothesis, feed_dict={X: x_data})\n\n    price = dict[0]\nelse:\n    price = 0\nreturn render(request, 'predict\/index.html', {'price': price})\n<\/code><\/pre>\n\n<p>POST<\/p>\n\n<p>Variable        Value\ncsrfmiddlewaretoken  'BeE48x03YbOY3tIC7eF8L0tKrZME'<br>\navg_temp         '24'\nmin_temp         '12'\nmax_temp         '31'\nrain_fall        '0.9'<\/p>\n\n<p>After receiving the post data, we want the price variable to have a forecast.<\/p>\n","answers":[{"AnswerId":"57831132","CreationDate":"2019-09-07T06:02:00.870","ParentId":null,"OwnerUserId":"2851685","Title":null,"Body":"<p><strong>Concept<\/strong><\/p>\n\n<p>At high level we can thing of graph as mathematical expression like (x+3x^2-.6x^3) which defines operation to be performed on a variable x. On the other hand session is where value of x and all the intermediate values like \"3x^2\" etc are stored. <\/p>\n\n<p>To restore session we need graph same elements for which session is saved.<\/p>\n\n<p>When you are running code below<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>W = tf.Variable(tf.random_normal([4, 1]), name=\"weight\")\nb = tf.Variable(tf.random_normal([1]), name=\"bias\")\n<\/code><\/pre>\n\n<p>Tensorflow will create the variable and add that to default graph which may not have same structure as the saved graph.<\/p>\n\n<p><strong>Steps to Restore Both Graph and Session<\/strong><\/p>\n\n<ul>\n<li>Restore graph from \".meta\" file first<\/li>\n<li>Then you can get reference of tensors and operations from the graphs by using graph.get_tensor_by_name()\/graph.get_operation_by_name()<\/li>\n<li>Then you can do some addition to the graph, that won't affect session restoration<\/li>\n<li>do variable initialization before restoring the session. If you do it latter it will overwrite the restored variables<\/li>\n<li>restore the session from checkpoint<\/li>\n<\/ul>\n\n<p><strong>Code will look like below<\/strong>   <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>## restore the graph\nimporter = tf.train.import_meta_graph(\"%s.meta\"%saved_path)\ngraph = tf.get_default_graph()\n\n## get reference of output of tensor and set it to variable\n## here \":0\" indicates the output of an operation\nW = graph.get_tensor_by_name(\"weights\")\noutput = graph.get_tensor_by_name(\"hidden5_out:0\")\n\n\n## restore the variables in session\nsession = tf.Session(graph=graph)\nimporter.restore(session, saved_path)\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":57831054,"AnswerCount":0,"Tags":"<python-3.x><keras><prediction><display>","CreationDate":"2019-09-07T05:49:07.827","AcceptedAnswerId":null,"OwnerUserId":9879809.0,"Title":"How to display predicted image in CNN network?","Body":"<p>q=model.predict(a_arr, batch_size=None, verbose=0, steps=None)<\/p>\n\n<p>print(q.shape)= (1,256,256,2) <\/p>\n\n<p>I have predicted for 1 image which is of 256*256 in size and got the results in probabilities for each pixels that tells to which class it belongs to (urban or rural)<\/p>\n\n<p>My question is how do I display this as an image ?<\/p>\n","answers":[]}
+{"QuestionId":57831334,"AnswerCount":0,"Tags":"<pytorch><torchvision>","CreationDate":"2019-09-07T06:40:43.677","AcceptedAnswerId":null,"OwnerUserId":4159376.0,"Title":"Invalid argument tensors for DataLoader PyTorch","Body":"<p>I have made a dataset using PyTorch's <code>DataLoader<\/code> and <code>Imagefolder<\/code>, my dataset class has two ImageFolder datasets. These two datasets are paired (original and ground truth image). I want to feed these to a PyTorch neural network. <\/p>\n\n<p>Dataset class:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class bsds_dataset(Dataset):\n    def __init__(self, ds_main, ds_energy):\n        self.dataset1 = ds_main\n        self.dataset2 = ds_energy\n\n    def __getitem__(self, index):\n        x1 = self.dataset1[index]\n        x2 = self.dataset2[index]\n\n        return x1, x2\n\n    def __len__(self):\n        return len(self.dataset1)\n<\/code><\/pre>\n\n<p>And I'm loading images with Imagefolder:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>original_imagefolder = '.\/images\/whole'\ntarget_imagefolder = '.\/results\/whole'\n\noriginal_ds = ImageFolder(original_imagefolder, \ntransform=transforms.ToTensor())\nenergy_ds = ImageFolder(target_imagefolder, transform=transforms.ToTensor())\n\ndataset = bsds_dataset(original_ds, energy_ds)\nloader = DataLoader(dataset, batch_size=16)\n<\/code><\/pre>\n\n<p>Then I tried to iterate in batches:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>for i, x, y in enumerate(loader):\n    print(x)\n<\/code><\/pre>\n\n<p>This error happened:<\/p>\n\n<blockquote>\n  <p>RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 0. Got 321 and 481 in dimension 2 at ..\\aten\\src\\TH\/generic\/THTensor.cpp:711<\/p>\n<\/blockquote>\n\n<p>The dataset is BSDS500:<\/p>\n\n<p><a href=\"https:\/\/www2.eecs.berkeley.edu\/Research\/Projects\/CS\/vision\/grouping\/resources.html\" rel=\"nofollow noreferrer\">https:\/\/www2.eecs.berkeley.edu\/Research\/Projects\/CS\/vision\/grouping\/resources.html<\/a><\/p>\n\n<p>All images in the dataset are in 481x321 or 321x481 pixels.\nI think some transform is needed but I don't want to demolish the images and stretch them.<\/p>\n\n<p>Full traceback:<\/p>\n\n<pre><code>C:\\Anaconda3\\envs\\torchgpu\\lib\\site-packages\\ipykernel_launcher.py:77: UserWarning: nn.init.xavier_normal is now deprecated in favor of nn.init.xavier_normal_.\nC:\\Anaconda3\\envs\\torchgpu\\lib\\site-packages\\ipykernel_launcher.py:78: UserWarning: nn.init.constant is now deprecated in favor of nn.init.constant_.\n---------------------------------------------------------------------------\nRuntimeError                              Traceback (most recent call last)\n&lt;ipython-input-42-4c4ba0a13c32&gt; in &lt;module&gt;\n      5 optimizer = optim.SGD(model.parameters(), lr=0.001)\n      6 for epoch in range(epochs):\n----&gt; 7     for i, batch in enumerate(loader):\n      8         print(batch)\n\nC:\\Anaconda3\\envs\\torchgpu\\lib\\site-packages\\torch\\utils\\data\\dataloader.py in __next__(self)\n    558         if self.num_workers == 0:  # same-process loading\n    559             indices = next(self.sample_iter)  # may raise StopIteration\n--&gt; 560             batch = self.collate_fn([self.dataset[i] for i in indices])\n    561             if self.pin_memory:\n    562                 batch = _utils.pin_memory.pin_memory_batch(batch)\n\nC:\\Anaconda3\\envs\\torchgpu\\lib\\site-packages\\torch\\utils\\data\\_utils\\collate.py in default_collate(batch)\n     66     elif isinstance(batch[0], container_abcs.Sequence):\n     67         transposed = zip(*batch)\n---&gt; 68         return [default_collate(samples) for samples in transposed]\n     69 \n     70     raise TypeError((error_msg_fmt.format(type(batch[0]))))\n\nC:\\Anaconda3\\envs\\torchgpu\\lib\\site-packages\\torch\\utils\\data\\_utils\\collate.py in &lt;listcomp&gt;(.0)\n     66     elif isinstance(batch[0], container_abcs.Sequence):\n     67         transposed = zip(*batch)\n---&gt; 68         return [default_collate(samples) for samples in transposed]\n     69 \n     70     raise TypeError((error_msg_fmt.format(type(batch[0]))))\n\nC:\\Anaconda3\\envs\\torchgpu\\lib\\site-packages\\torch\\utils\\data\\_utils\\collate.py in default_collate(batch)\n     66     elif isinstance(batch[0], container_abcs.Sequence):\n     67         transposed = zip(*batch)\n---&gt; 68         return [default_collate(samples) for samples in transposed]\n     69 \n     70     raise TypeError((error_msg_fmt.format(type(batch[0]))))\n\nC:\\Anaconda3\\envs\\torchgpu\\lib\\site-packages\\torch\\utils\\data\\_utils\\collate.py in &lt;listcomp&gt;(.0)\n     66     elif isinstance(batch[0], container_abcs.Sequence):\n     67         transposed = zip(*batch)\n---&gt; 68         return [default_collate(samples) for samples in transposed]\n     69 \n     70     raise TypeError((error_msg_fmt.format(type(batch[0]))))\n\nC:\\Anaconda3\\envs\\torchgpu\\lib\\site-packages\\torch\\utils\\data\\_utils\\collate.py in default_collate(batch)\n     41             storage = batch[0].storage()._new_shared(numel)\n     42             out = batch[0].new(storage)\n---&gt; 43         return torch.stack(batch, 0, out=out)\n     44     elif elem_type.__module__ == 'numpy' and elem_type.__name__ != 'str_' \\\n     45             and elem_type.__name__ != 'string_':\n\nRuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 0. Got 321 and 481 in dimension 2 at ..\\aten\\src\\TH\/generic\/THTensor.cpp:711\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57831548,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-07T07:15:44.183","AcceptedAnswerId":null,"OwnerUserId":11377163.0,"Title":"Image classification using tf.estimator.BoostedTreesClassifier()","Body":"<p>How to train an image classification model using tf.estimator.BoostedTreesClassifier ?<\/p>\n\n<p>To use tf.estimator.BoostedTreesClassifier first we need to create a feature_columns, I'm not able to understand how to create feature_columns using images as input to tf.boostedtreeclassifier or Is it possible to give image as an input to the model ?<\/p>\n","answers":[]}
+{"QuestionId":57831935,"AnswerCount":0,"Tags":"<tensorflow><protocol-buffers>","CreationDate":"2019-09-07T08:17:33.487","AcceptedAnswerId":null,"OwnerUserId":11372242.0,"Title":"Tensorflow checkpoint errors caused by protobuf","Body":"<p>I'm facing an error when running tensorflow model (bert) on Ubuntu.<\/p>\n\n<p>Error message is as follows:<\/p>\n\n<blockquote>\n  <p>[libprotobuf FATAL external\/protobuf_archive\/src\/google\/protobuf\/map.h:1064] CHECK failed: it != end(): key not found: architecture\n  terminate called after throwing an instance of 'google::protobuf::FatalException'\n    what():  CHECK failed: it != end(): key not found: architecture<br>\n  Fatal Python error: Aborted<\/p>\n<\/blockquote>\n\n<p>I'm using python 3.6.8, tensorflow 1.14.0 and the result of <code>pip3 show protobuf<\/code> says<\/p>\n\n<p>Name: protobuf<br>\nVersion: 3.9.1<br>\nSummary: Protocol Buffers<br>\nHome-page: <a href=\"https:\/\/developers.google.com\/protocol-buffers\/\" rel=\"nofollow noreferrer\">https:\/\/developers.google.com\/protocol-buffers\/<\/a><br>\nAuthor: None<br>\nAuthor-email: None<br>\nLicense: 3-Clause BSD License<br>\nLocation: \/home\/titan-n0\/.local\/lib\/python3.6\/site-packages<br>\nRequires: six, setuptools<\/p>\n\n<p>The above error occurs right before making a checkpoint. I need this to make a saved model from tf serving.<\/p>\n","answers":[]}
+{"QuestionId":57832423,"AnswerCount":1,"Tags":"<python><numpy><pytorch>","CreationDate":"2019-09-07T09:33:12.910","AcceptedAnswerId":57832453.0,"OwnerUserId":4171008.0,"Title":"Convert CUDA tensor to NumPy","Body":"<p>First of all, I tried those solutions: <a href=\"https:\/\/stackoverflow.com\/questions\/53467215\/convert-pytorch-tensor-to-numpy-array-using-cuda\">1<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/41035617\/how-to-convert-tensor-to-numpy-array\">2<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/50931333\/convert-tensor-to-numpy-array\">3<\/a>, and <a href=\"https:\/\/stackoverflow.com\/questions\/52148590\/convert-tensor-to-numpy-without-a-session\">4<\/a>, but did not work for me.<\/p>\n\n<p>After training and testing the neural network, I am trying to show some examples to verify my work. I named the method <strong>predict<\/strong> which I pass the image to it to predict for which class it belongs:<\/p>\n\n<pre><code>def predict(model, image_path, topk=5):\n''' Predict the class (or classes) of an image using a trained deep learning model.\n'''\n\noutput = process_image(image_path)\noutput.unsqueeze_(0)\noutput = output.cuda().float()\n\nmodel.eval()\n\nwith torch.no_grad():\n    score = model(output)\n    prob, idxs = torch.topk(score, topk)\n\n    # Convert indices to classes\n    idxs = np.array(idxs)\n    idx_to_class = {val:key for key, val in model.class_to_idx.items()}\n    classes = [idx_to_class[idx] for idx in idxs[0]]\n\n    # Map the class name with collected topk classes\n    names = []\n    for cls in classes:\n        names.append(cat_to_name[str(cls)])\n\n    return prob, names\n<\/code><\/pre>\n\n<p>Then there is the final step which displays the final result based on the training of the neural network and done like this:<\/p>\n\n<pre><code># TODO: Display an image along with the top 5 classes\nx_pos, y_pos = predict(model, img_pil, topk=5)\n\nax_img = imshow(output)\nax_img.set_title(y_pos[0])\n\nplt.figure(figsize=(4,4))\nplt.barh(range(len(y_pos)), np.exp(x_pos[0]))\nplt.yticks(range(len(y_pos)), y_pos)\n\nplt.show()\n<\/code><\/pre>\n\n<p>The error is:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-45-e3f9951e9804&gt; in &lt;module&gt;()\n----&gt; 1 x_pos, y_pos = predict(model, img_pil, topk=5)\n      2\n      3 ax_img = imshow(output)\n      4 ax_img.set_title(y_pos[0])\n      5\n\n1 frames\n&lt;ipython-input-44-d77500f31561&gt; in predict(model, image_path, topk)\n     14\n     15         # Convert indices to classes\n---&gt; 16         idxs = np.array(idxs)\n     17         idx_to_class = {val:key for key, val in model.class_to_idx.items()}\n     18         classes = [idx_to_class[idx] for idx in idxs[0]]\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/tensor.py in __array__(self, dtype)\n    456     def __array__(self, dtype=None):\n    457         if dtype is None:\n--&gt; 458             return self.numpy()\n    459         else:\n    460             return self.numpy().astype(dtype, copy=False)\n\nTypeError: can't convert CUDA tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.\n<\/code><\/pre>\n\n<p><strong>How do I solve this?<\/strong><\/p>\n\n<p>I tried to change idx to <code>idxs = idxs.cpu().numpy()<\/code> and the error is:<\/p>\n\n<p><\/p>\n\n<pre><code>TypeError                                 Traceback (most recent call last)\n&lt;ipython-input-62-e3f9951e9804&gt; in &lt;module&gt;()\n      5\n      6 plt.figure(figsize=(4,4))\n----&gt; 7 plt.barh(range(len(y_pos)), np.exp(x_pos[0]))\n      8 plt.yticks(range(len(y_pos)), y_pos)\n      9\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/tensor.py in __array__(self, dtype)\n    456     def __array__(self, dtype=None):\n    457         if dtype is None:\n--&gt; 458             return self.numpy()\n    459         else:\n    460             return self.numpy().astype(dtype, copy=False)\n\nTypeError: can't convert CUDA tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.\n<\/code><\/pre>\n","answers":[{"AnswerId":"57832453","CreationDate":"2019-09-07T09:37:46.290","ParentId":null,"OwnerUserId":"7701908","Title":null,"Body":"<p>Try to change<\/p>\n\n<pre><code>idxs = np.array(idxs)\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>idxs = idxs.cpu().numpy()\n<\/code><\/pre>\n\n<p>And change<\/p>\n\n<pre><code>plt.barh(range(len(y_pos)), np.exp(x_pos[0]))\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>plt.barh(range(len(y_pos)), np.exp(x_pos[0].cpu().numpy()))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57832879,"AnswerCount":1,"Tags":"<tensorflow><redhat><bazel>","CreationDate":"2019-09-07T10:36:12.767","AcceptedAnswerId":null,"OwnerUserId":10436469.0,"Title":"How to uninstall bazel 0.29.0 in order to install 0.26.1 because of tensorflow","Body":"<p>I am using Redhat 7.3. I need to install tensorflow for that I already installed bazel 0.29.0 and when I wanted to configure tensorflow it requires bazel 0.26.1. Thats why i tried to uninstall bazel 0.29.0 but was not able to do it. I am new in Redhat community , could you please show me a way how to solve this problem ? Thanks in advance.<\/p>\n","answers":[{"AnswerId":"58152464","CreationDate":"2019-09-29T05:19:17.430","ParentId":null,"OwnerUserId":"6912318","Title":null,"Body":"<p>Try:<\/p>\n\n<pre><code>rm $HOME\/.cache\/bazel\nsudo rm \/usr\/local\/bin\/bazel \/etc\/bazelrc \/usr\/local\/lib\/bazel\n<\/code><\/pre>\n\n<p>Worked for me. I needed to reinstall bazel<\/p>\n"}]}
+{"QuestionId":57833071,"AnswerCount":1,"Tags":"<pytorch><tensor>","CreationDate":"2019-09-07T11:03:36.230","AcceptedAnswerId":57833770.0,"OwnerUserId":1767774.0,"Title":"Pytorch: choosing columns from a 3d tensor, according to indices tensor","Body":"<p>I have a 3D tensor <code>M<\/code> of dimensions <code>[BxLxD]<\/code> and a 1D tensor <code>idx<\/code> of dimensions <code>[B,1]<\/code> that contains column indices in the range <code>(0, L-1)<\/code>. I want to create a 2D tensor <code>N<\/code> of dimensions <code>[BxD]<\/code> such that <code>N[i,j] = M[i, idx[i], j]<\/code>. How can this be done efficiently?<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>B,L,D = 2,4,2\n\nM = torch.rand(B,L,D)\n\n&gt;\n\ntensor([[[0.0612, 0.7385],\n         [0.7675, 0.3444],\n         [0.9129, 0.7601],\n         [0.0567, 0.5602]],\n\n        [[0.5450, 0.3749],\n         [0.4212, 0.9243],\n         [0.1965, 0.9654],\n         [0.7230, 0.6295]]])\n\n\nidx = torch.randint(0, L, size = (B,))\n\n&gt;\n\ntensor([3, 0])\n\nN = get_N(M, idx)\n\nExpected output:\n\n&gt;\n\ntensor([[0.0567, 0.5602], \n       [0.5450, 0.3749]])\n<\/code><\/pre>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"57833770","CreationDate":"2019-09-07T12:39:05.627","ParentId":null,"OwnerUserId":"7701908","Title":null,"Body":"<pre><code>import torch\n\nB,L,D = 2,4,2\n\ndef get_N(M, idx):\n    return M[torch.arange(B), idx, :].squeeze()\n\nM = torch.tensor([[[0.0612, 0.7385],\n                   [0.7675, 0.3444],\n                   [0.9129, 0.7601],\n                   [0.0567, 0.5602]],\n\n                   [[0.5450, 0.3749],\n                   [0.4212, 0.9243],\n                   [0.1965, 0.9654],\n                   [0.7230, 0.6295]]])\nidx = torch.tensor([3,0])\nN = get_N(M, idx)\nprint(N)\n<\/code><\/pre>\n\n<p>result:<\/p>\n\n<pre><code>tensor([[0.0567, 0.5602],\n        [0.5450, 0.3749]])\n<\/code><\/pre>\n\n<p>slice along two dimensions.<\/p>\n"}]}
+{"QuestionId":57833245,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-07T11:28:10.270","AcceptedAnswerId":null,"OwnerUserId":2337125.0,"Title":"How to use TF 2.0 tf.recompute_grad?","Body":"<p>I wanted to use memory saving gradients (openai\/gradient-checkpointing) to reduce GPU memory cost of my neural network, but I found that it's impossible in TF 2.0, but also I have found that I can use tf.recompute_grad for this purpose. I didn't found any examples or tutorials on Google, so I ask here. Also, is it possible to use this with tf.keras?<\/p>\n","answers":[]}
+{"QuestionId":57833280,"AnswerCount":0,"Tags":"<keras><lstm>","CreationDate":"2019-09-07T11:33:22.473","AcceptedAnswerId":null,"OwnerUserId":7912541.0,"Title":"Functional API connection problem (CNN output to LSTM input)","Body":"<p>I am trying to CNN's output to feed LSTM in encoder of seq2seq model in Keras. But LSTM does not receive CNN's output..<\/p>\n\n<p>Anybody who has an idea?\nMany thanks.<\/p>\n\n<pre><code>from keras.layers import Input, LSTM\nfrom keras.models import Model\n\nimport keras\nregulariser = None\n\nnum_encoder_input_features=128\nhidden_neurons=64\n\ndef cnn_encode(input_tensor):\n    c1 = keras.layers.Conv2D(32, (3, 3), activation='relu', input_shape=(201, 201, 1),data_format='channels_last')(input_tensor)\n    c1 = keras.layers.MaxPooling2D((2, 2))(c1)\n    f1 = keras.layers.Flatten()(c1)\n    f1 = keras.layers.Dense(128)(f1)\n    return f1\n\ninput_tensor = keras.layers.Input(shape=( 201, 201, 1))\ncnn_output_tensor = cnn_encode(input_tensor)\n\nencoder = LSTM(hidden_neurons, return_state=True)\nencoder_outputs, state_h, state_c = encoder(cnn_output_tensor)\nencoder_states = [state_h, state_c]\n\ndecoder_inputs = Input(shape=(None, 201, 201, 1))\ndecoder_lstm = LSTM(hidden_neurons, return_sequences=True, return_state=True)\ndecoder_outputs, _, _ = decoder_lstm(decoder_inputs,\n                                     initial_state=encoder_states)\n\nmodel = Model([input_tensor, decoder_inputs], decoder_outputs)\n<\/code><\/pre>\n\n<p>Traceback (most \/model\/05_Seq2Seq_Convolution_Deconvolution.py\", line 22, in \n    encoder_outputs, state_h, state_c = encoder(cnn_output_tensor)\n  File \"\/anaconda3\/envs\/dev_tf1.13\/lib\/python3.6\/site-packages\/keras\/layers\/recurrent.py\", line 532, in <strong>call<\/strong>\n    return super(RNN, self).<strong>call<\/strong>(inputs, **kwargs)\n  File \"\/anaconda3\/envs\/dev_tf1.13\/lib\/python3.6\/site-packages\/keras\/engine\/base_layer.py\", line 414, in <strong>call<\/strong>\n    self.assert_input_compatibility(inputs)\n  File \"\/anaconda3\/envs\/dev_tf1.13\/lib\/python3.6\/site-packages\/keras\/engine\/base_layer.py\", line 311, in assert_input_compatibility\n    str(K.ndim(x)))\nValueError: Input 0 is incompatible with layer lstm_1: expected ndim=3, found ndim=2<\/p>\n","answers":[]}
+{"QuestionId":57833919,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-07T13:00:54.997","AcceptedAnswerId":null,"OwnerUserId":12034521.0,"Title":"When I try to use trt.create_inference_graph I got this error:tensorflow.python.framework.errors_impl.FailedPreconditionError: TensorRT is not enabled","Body":"<p>trt_graph = trt.create_inference_graph(input_graph_def = graph_def, outputs = outputs, max_batch_size=1, max_workspace_size_bytes=workspace_size, precision_mode='INT8', minimum_segment_size=3)<\/p>\n\n<p>tf_version = 1.10.0<\/p>\n","answers":[]}
+{"QuestionId":57833959,"AnswerCount":0,"Tags":"<python><tensorflow><keras><linear-regression><keras-layer>","CreationDate":"2019-09-07T13:07:09.813","AcceptedAnswerId":null,"OwnerUserId":7592063.0,"Title":"Keras training model to find the sum of two numbers","Body":"<p>Trying to implement a very simple training model to predict the sum of two numbers using tensorflow keras.\nI tried with Single Dense Layer.\nI tried changing no of epochs as well. \nI tried changing optimizer to linear as well.\nBut nothing worked out for me to give me the best accuracy.<\/p>\n\n<p>Below is the code What I tried so far and the output that I got.<\/p>\n\n<p><strong>Import<\/strong><\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nfrom random import randrange\n<\/code><\/pre>\n\n<p><strong>Generate Training Data<\/strong><\/p>\n\n<pre><code>trainingInput = [[i, i + randrange(5000)] for i in range(1, 5000)]\ntrainingOutput = [(input [0] + input [1]) for input  in trainingInput ]\n\ntestInput = [[5, 5], [1, 9], [2, 5], [6, 3], [1, 4]]\ntestOutput = [10, 10, 7, 9, 5]\n<\/code><\/pre>\n\n<p><strong>Build the model<\/strong><\/p>\n\n<pre><code>model = tf.keras.Sequential()\nmodel.add(tf.keras.layers.Flatten(input_shape=(2,)))\nmodel.add(tf.keras.layers.Dense(64, activation=tf.nn.relu))\nmodel.add(tf.keras.layers.Dense(64, activation=tf.nn.relu))\nmodel.add(tf.keras.layers.Dense(1))\n<\/code><\/pre>\n\n<p><strong>Compile the model<\/strong><\/p>\n\n<pre><code>model.compile(optimizer='adam', loss=tf.keras.losses.mae, metrics=['mae'])\n<\/code><\/pre>\n\n<p><strong>Training<\/strong><\/p>\n\n<pre><code>model.fit(trainData, trainOutput, batch_size=5, epochs=50)\n<\/code><\/pre>\n\n<p><strong>Final Evaluation and Prediction<\/strong><\/p>\n\n<pre><code>test_loss, test_acc = model.evaluate(testData, testOutput)\nprint(\"Test Accuracy : \", test_acc)\na = np.array([[3, 3000], [4, 5], [1,10], [2,10],[5,9], [4,10], [1,15]])\nprint(model.predict(a))\n<\/code><\/pre>\n\n<p><strong>Output<\/strong><\/p>\n\n<pre><code>[[2994.769   ]\n[   8.959281]\n[  10.961123]\n[  11.956481]\n[  13.944955]\n[  13.947194]\n[  15.949196]]\n<\/code><\/pre>\n\n<p>Can anyone help me improve my training model and accuracy? Thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":57834078,"AnswerCount":0,"Tags":"<python><tensorflow><keras><scikit-learn><deep-learning>","CreationDate":"2019-09-07T13:21:55.043","AcceptedAnswerId":null,"OwnerUserId":12012634.0,"Title":"Accuracy given from evaluating model not equal to sklearn classification_report accuracy","Body":"<p>I'm using sklearn classification_report for reporting test statistics. The accuracy given by this method is 42% while model evaluation gives 93% accuracy. Which one is the real accuracy and what's the reason of this difference?<\/p>\n\n<p>Model evaluation:<br><br>\n<code>results = model.evaluate(test_ds.values, test_lb.values)\nprint(results)<\/code> <br><br>\nOutput:<br>\n<code>7397\/7397 [==============================] - 0s 28us\/sample - loss: 0.2309 - acc: 0.9305<\/code><br><br>\nReport Classification:<br><br>\n<code>import numpy as np\nfrom sklearn.metrics import classification_report\npredictions = model.predict(test_ds)\nprint(classification_report(test_lb, np.argmax(predictions, axis=1)))<\/code><br><br>\nOutput:<br><\/p>\n\n<p><code>label    precision   recall   f1-score    support<\/code><\/p>\n\n<pre><code>       0       0.41      0.38      0.40      3700\n       1       0.43      0.46      0.44      3697\n\naccuracy                           0.42      7397\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57834128,"AnswerCount":2,"Tags":"<tensorflow><keras><rank><dimension>","CreationDate":"2019-09-07T13:28:26.237","AcceptedAnswerId":57837653.0,"OwnerUserId":735008.0,"Title":"The input dimension of the LSTM layer in Keras","Body":"<p>I'm trying keras.layers.LSTM.\nThe following code works.<\/p>\n\n<pre><code>#!\/usr\/bin\/python3\nimport tensorflow as tf\nimport numpy as np\nfrom tensorflow import keras\n\ndata = np.array([1, 2, 3]).reshape((1, 3, 1)) \nx = keras.layers.Input(shape=(3, 1)) \ny = keras.layers.LSTM(10)(x)\nmodel = keras.Model(inputs=x, outputs=y)\n\nprint (model.predict(data))\n<\/code><\/pre>\n\n<p>As shown above, the input data shape is <code>(1, 3, 1)<\/code>, and the actual input shape in the Input layer is <code>(3, 1)<\/code>. I'm a little bit confused about this inconsistency of the dimension.\nIf I use the following shape in the Input layer, it doesn't work:<\/p>\n\n<pre><code>x = keras.layers.Input(shape=(1, 3, 1)) \n<\/code><\/pre>\n\n<p>The error message is as follows:<\/p>\n\n<pre><code>ValueError: Input 0 of layer lstm is incompatible with the layer: expected ndim=3, found ndim=4. Full shape received: [None, 1, 3, 1]\n<\/code><\/pre>\n\n<p>It seems that the rank of the input must be 3, but why should we use a rank-2 shape in the Input layer?<\/p>\n","answers":[{"AnswerId":"57834385","CreationDate":"2019-09-07T14:00:58.857","ParentId":null,"OwnerUserId":"9061899","Title":null,"Body":"<p>Check out the documentation for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/Input\" rel=\"nofollow noreferrer\">tf.keras.Input<\/a>. The syntax is as-<\/p>\n\n<pre><code>tf.keras.Input(\n    shape=None,\n    batch_size=None,\n    name=None,\n    dtype=None,\n    sparse=False,\n    tensor=None,\n    **kwargs\n)\n<\/code><\/pre>\n\n<p><strong>shape:<\/strong> defines the shape of a single sample, with variable batch size.<\/p>\n\n<p>Notice, that it expects the first value as batch_size otherwise pass <code>batch_size<\/code> as a parameter explicitly<\/p>\n"},{"AnswerId":"57837653","CreationDate":"2019-09-07T21:39:55.010","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Keras works with \"batches\" of \"samples\". Since most models use variable batch sizes that you define only when fitting, for convenience you don't need to care about the batch dimension, but only with the sample dimension. <\/p>\n\n<p>That said, when you use <code>shape = (3,1)<\/code>, this is the same as defining <code>batch_shape = (None, 3, 1)<\/code> or <code>batch_input_shape = (None, 3, 1)<\/code>.   <\/p>\n\n<p>The three options mean:<\/p>\n\n<ul>\n<li>A variable batch size: <code>None<\/code>    <\/li>\n<li>With samples of shape <code>(3, 1)<\/code>.    <\/li>\n<\/ul>\n\n<p>It's important to know this distinction especially when you are going to create custom layers, losses or metrics. The actual tensors all have the batch dimension and you should take that into account when making operations with tensors. <\/p>\n"}]}
+{"QuestionId":57834137,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2019-09-07T13:30:23.680","AcceptedAnswerId":57879490.0,"OwnerUserId":1056563.0,"Title":"Tensorboard printed summary Precision and Accuracy does not show the Step count","Body":"<p>We want to correlate the StepCount with the trends of Precision and Accuracy. The Tensorboard GUI allows this: but we want to automate the results.  The printed results do not print the Step Count - at least not by default:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/sHgsn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/sHgsn.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Is there a way to tease the Step Count from Tensorboard's printed results?<\/p>\n","answers":[{"AnswerId":"57879490","CreationDate":"2019-09-11T00:01:56.840","ParentId":null,"OwnerUserId":"1056563","Title":null,"Body":"<p>The steps are actually available in the <em>TF Events<\/em> file but require coding to extract and correlate.  I show how to do that here:  <a href=\"https:\/\/stackoverflow.com\/questions\/57813373\/how-to-parse-the-tensorflow-events-file\">How to parse the tensorflow events file?<\/a><\/p>\n"}]}
+{"QuestionId":57834187,"AnswerCount":1,"Tags":"<keras><deep-learning><semantic-segmentation>","CreationDate":"2019-09-07T13:36:47.000","AcceptedAnswerId":57837433.0,"OwnerUserId":9652729.0,"Title":"How to fix input shape error in keras model","Body":"<p>I tried to perform semantic segmentation using this tutorial:\n<a href=\"https:\/\/github.com\/nikhilroxtomar\/UNet-Segmentation-in-Keras-TensorFlow\/blob\/master\/unet-segmentation.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/nikhilroxtomar\/UNet-Segmentation-in-Keras-TensorFlow\/blob\/master\/unet-segmentation.ipynb<\/a>\nI modified a bit his notebook but I successfully trained the model with a 50% accuracy.<\/p>\n\n<p>I already tried to reshape the input array but it does not work.\nHERE IS THE CODE:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>test =  X_train[0]\ntest.shape\n&gt;&gt;&gt; (480, 640, 4)\nparallel_model.predict(test)\n&gt;&gt;&gt; ValueError: Error when checking input: expected input_3 to have 4 dimensions, but got array with shape (480, 640, 4)\n<\/code><\/pre>\n\n<p>Here is the model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def UNet():\n    f = [16, 32, 64, 128, 256]\n    inputs = keras.layers.Input((480, 640, 4))\n\n    p0 = inputs\n    c1, p1 = down_block(p0, f[0]) #128 -&gt; 64\n    c2, p2 = down_block(p1, f[1]) #64 -&gt; 32\n    c3, p3 = down_block(p2, f[2]) #32 -&gt; 16\n    c4, p4 = down_block(p3, f[3]) #16-&gt;8\n\n    bn = bottleneck(p4, f[4])\n\n    u1 = up_block(bn, c4, f[3]) #8 -&gt; 16\n    u2 = up_block(u1, c3, f[2]) #16 -&gt; 32\n    u3 = up_block(u2, c2, f[1]) #32 -&gt; 64\n    u4 = up_block(u3, c1, f[0]) #64 -&gt; 128\n\n    outputs = keras.layers.Conv2D(4, (1, 1), padding=\"same\", activation=\"sigmoid\")(u4)\n    model = keras.models.Model(inputs, outputs)\n    return model\n<\/code><\/pre>\n\n<p>I know this is a noob error but I really want top solve it!<\/p>\n","answers":[{"AnswerId":"57837433","CreationDate":"2019-09-07T20:59:31.790","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Keras works with \"batches\", never with single images. <\/p>\n\n<p>That means it's expecting 4 dimensions <code>(batch_size, 480, 640, 4)<\/code>.   <\/p>\n\n<p>If you are going to predict with a single image, then you need your input array with shape <code>(1, 480, 640, 4)<\/code>.   <\/p>\n\n<pre><code>test =  X_train[0]\ntest.shape\n&gt;&gt;&gt; (480, 640, 4)\ntest = test.reshape((-1, 480, 640, 4))\ntest.shape\n&gt;&gt;&gt; (1, 480, 640, 4)\n<\/code><\/pre>\n\n<p>Now you can <code>parallel_model.predict(test)<\/code>   .<\/p>\n"}]}
+{"QuestionId":57834240,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-07T13:42:46.623","AcceptedAnswerId":null,"OwnerUserId":11810590.0,"Title":"How can i improve my CNN's accuracy evolution?","Body":"<p>So, i'm trying to create a CNN which can predict if there is any \"support devices\" in a x-ray thorax image, but when training my model it seems it's not learning anything.<\/p>\n\n<p>I'm using a dataset called \"CheXpert\" which has over 200.000 images to use. After doing some \"cleaning\", the final dataset ended up with 100.000 images.\nAs far as the model is concerned, i imported the convolutional base of the vgg16 pretrained model and added by my self 2 fully conected layers. Then, i freezed all the convolutional base and make only trainable the fully conected layers. Here's the code:<\/p>\n\n<pre><code>from keras.layers import GlobalAveragePooling2D\nfrom keras.models import Model\n\npretrained_model = VGG16(weights='imagenet', include_top=False)\n\npretrained_model.summary()\n\nfor layer in pretrained_model.layers:\n    layer.trainable = False\n\nx = pretrained_model.output\nx = GlobalAveragePooling2D()(x)\n\ndropout = Dropout(0.25)\n\n# let's add a fully-connected layer\nx = Dense(1024, activation='relu')(x)\nx = dropout(x)\nx = Dense(1024, activation = 'relu')(x)\npredictions = Dense(1, activation='sigmoid')(x)\n\nfinal_model = Model(inputs=pretrained_model.input, outputs=predictions)\n\nfinal_model.compile(loss='binary_crossentropy', optimizer='rmsprop', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>As far as i know, the normal behavior should be that the accuracy should start low and then grow up with the epochs. But here it only oscillates through the same values (0.93 and 0.95). I'm sorry i cannot upload images to show you the graphs.<\/p>\n\n<p>To sum up, i want to know if that little variance in the accuracy means that the model is not learning anything.<\/p>\n\n<p>I have an hypothesis: from all the 100.000 images of the dataset, 95.000 have the label \"1\" and only 5.000 have the label \"0\". I think that if diminish the images with \"1\" equate them with the images with \"0\" the results would change.<\/p>\n","answers":[{"AnswerId":"57834361","CreationDate":"2019-09-07T13:57:21.133","ParentId":null,"OwnerUserId":"12026600","Title":null,"Body":"<p>The lack of images labeled \"0\" doesn't help the CNN for sure. I also suggest to lower the learning rate and play around with the batch size to see if something changes.\nI wish it helps.<\/p>\n"}]}
+{"QuestionId":57834554,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-07T14:21:44.000","AcceptedAnswerId":57836681.0,"OwnerUserId":8176285.0,"Title":"Keras Generator keeps looping for no reason","Body":"<p>I have a keras generator <\/p>\n\n<pre><code>train_datagen = ImageDataGenerator(rotation_range=360,\n                                       horizontal_flip=True,\n                                       vertical_flip=True,\n                                       validation_split=0.15,\n                                       preprocessing_function=preprocess_image, \n                                       rescale=1 \/ 128.)\n\nleft_test_generator=train_datagen.flow_from_dataframe(dataframe=df_test_left, \n                                                    directory = test_images_path,\n                                                    x_col=\"pic_id\",\n                                                    target_size=(IMG_WIDTH, IMG_HEIGHT),\n                                                    batch_size=1,\n                                                    shuffle=False, \n                                                    class_mode=None, seed=SEED) \n<\/code><\/pre>\n\n<p>the length of left_test_generator is 500<\/p>\n\n<p>but when if i run this <\/p>\n\n<pre><code>for image_count, image in enumerate(right_test_generator):\n    print(image_count)\n<\/code><\/pre>\n\n<p>this will print image_count which is greater than 8000.\nWhat is going on here? Any ideas?<\/p>\n","answers":[{"AnswerId":"57836681","CreationDate":"2019-09-07T19:03:25.770","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>That is how generators work, internally they have an infinite loop that produces batches of data, the <code>__len__<\/code> method is just for reference to be aware of how many calls to the generator are required to reach one epoch.<\/p>\n\n<p>So if you want to get a specific number of samples from the generator, you should not iterate on it like that, use a for loop with a specific number of iterations.<\/p>\n"}]}
+{"QuestionId":57835069,"AnswerCount":1,"Tags":"<python><keras><kaggle>","CreationDate":"2019-09-07T15:23:27.157","AcceptedAnswerId":null,"OwnerUserId":2569186.0,"Title":"Kernel Threw Exception in kaggle submission","Body":"<p>We have six submissions to a kaggle <a href=\"https:\/\/www.kaggle.com\/c\/aptos2019-blindness-detection\" rel=\"nofollow noreferrer\">competition<\/a><\/p>\n\n<p>The second submission succeeded but all other failed, four failed for <code>Kernel Threw Exception<\/code><\/p>\n\n<p>Looking at the log there is nothing to point to the failure:<\/p>\n\n<pre><code>Predicting\n\n2019-09-07 07:41:52.231132: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcublas.so.10.0\n\n2019-09-07 07:41:52.940301: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudnn.so.7\n\ny_pred Shape:  (1928, 5)\n\n[[0.009 0.296 0.398 0.111 0.186]\n [0.015 0.156 0.273 0.276 0.281]\n [0.009 0.293 0.042 0.356 0.3  ]\n ...\n [0.021 0.148 0.283 0.34  0.208]\n [0.007 0.197 0.516 0.136 0.144]\n [0.995 0.002 0.001 0.    0.001]]\n\n(1928, 2)\n<\/code><\/pre>\n\n<p>The log stops here.  Any ideas how we can discover what is causing the exception and why it would work once and not again?<\/p>\n","answers":[{"AnswerId":"57837464","CreationDate":"2019-09-07T21:04:53.793","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Without looking at your kernel it's impossible to know. <\/p>\n\n<p>But a few hints are:<\/p>\n\n<ul>\n<li>Kernel must finish running before 9 hours    <\/li>\n<li>Make sure internet is \"off\" for those kernels    <\/li>\n<li>Make sure there is an output file called <code>submission.csv<\/code>    <\/li>\n<li>Make sure this file is in the correct format for the competition (compare to the output file of the kernel that worked)    <\/li>\n<\/ul>\n\n<p>Debugging hints:<\/p>\n\n<p>Before submitting, make a \"commit\" first. Wait for it to finish and take a look at what is happening.  <\/p>\n\n<ul>\n<li>Check each cell for errors in their outputs    <\/li>\n<li>Check for the existence of the submission file    <\/li>\n<li>Check what is inside the submission file    <\/li>\n<li>Take a look at the log page of the kernel (not inside the editor when you commit, but at the kernel's page: <code>https:\/\/www.kaggle.com\/yourname\/kernel_name<\/code> without the <code>edit<\/code> part at the end)    <\/li>\n<\/ul>\n\n<p>This is the best to do if the kernel is private for us. <\/p>\n"}]}
+{"QuestionId":57835098,"AnswerCount":1,"Tags":"<c++><build><cmake><caffe>","CreationDate":"2019-09-07T15:26:35.553","AcceptedAnswerId":null,"OwnerUserId":8103305.0,"Title":"Building Caffe on Windows without Python for C++ \/ OpenCV usage","Body":"<p>I'd like to build Caffe on my windows machine using VS2019. Executing CMake, I get the error:<\/p>\n\n<blockquote>\n  <p>Building without python. Prebuilt dependencies will default to Python 2.7\n  CMake Error at cmake\/WindowsDownloadPrebuiltDependencies.cmake:40 (message):\n    Could not find url for MSVC version = 1916 and Python version = ..\n  Call Stack (most recent call first):\n    CMakeLists.txt:77 (include)<\/p>\n<\/blockquote>\n\n<p>I read about the error in a google research and the answer is always \n<code>\"Set the python directory\"<\/code> <\/p>\n\n<p>I don't have python installed on my computer and I also wouldn't like to as I just want caffe to use in C++ and\/or OpenCV.<\/p>\n\n<p>I also set the cmake parameter <code>BUILD_python = false<\/code>and <code>BUILD_python_layer = false<\/code><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/3Tjd7.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3Tjd7.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Can somebody guide me, how to come arround that?<\/p>\n\n<p>Thank you.<\/p>\n\n<p>Maybe it's helpful - here is \"Line 40 from the <code>WindowsCreateLinkHeader.cmake<\/code>File:<\/p>\n\n<pre><code>get_filename_component(MSVC_VC_DIR [HKEY_LOCAL_MACHINE\\\\SOFTWARE\\\\Microsoft\\\\VisualStudio\\\\${MSVC_PRODUCT_VERSION_${MSVC_VERSION}}\\\\Setup\\\\VC;ProductDir] REALPATH CACHE)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57836364","CreationDate":"2019-09-07T18:18:53.780","ParentId":null,"OwnerUserId":"8386275","Title":null,"Body":"<p>okey, here goes nothing :)<\/p>\n\n<p>If you take a look at <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/windows\/cmake\/WindowsDownloadPrebuiltDependencies.cmake\" rel=\"nofollow noreferrer\">this<\/a> file(WindowsDownloadPrebuiltDependencies.cmake) you will find that DEPENDENCIES_NAME_* and DEPENDENCIES_URL_* do not contain MSCV version 1916, only 1900 and 1800. This is what the error is actually saying to you.<\/p>\n\n<p>One fix would be to change all 1900_27 to 1916_27 and pray to your favorite god that it is still compatible.<br>\n27 is by the way Python version 2.7 which the error states to be the default.<\/p>\n\n<p>Have not tried this myself so it is kinda unsure if it really would work but hey, maybe one step closer for you :)<\/p>\n"}]}
+{"QuestionId":57835371,"AnswerCount":2,"Tags":"<node.js><tensorflow><node-pre-gyp>","CreationDate":"2019-09-07T16:02:47.527","AcceptedAnswerId":null,"OwnerUserId":5326607.0,"Title":"Unable to install @tensorflow\/tfjs-node with npm","Body":"<p>I cannot install '@tensorflow\/tfjs-node'<\/p>\n\n<p>I installed it using 'npm install @tensorflow\/tfjs-node' but the installations fails. I tried installing 'node-pre-gyp' globally, with and without root permissions but no success. I run on node v12.9.0.<\/p>\n\n<pre><code>\n&gt; fsevents@1.2.9 install \/Users\/username\/Documents\/Development\/AI\/Tensorflow\/node_modules\/fsevents\n&gt; node install\n\nnode-pre-gyp WARN Using needle for node-pre-gyp https download \n[fsevents] Success: \"\/Users\/username\/Documents\/Development\/AI\/Tensorflow\/node_modules\/fsevents\/lib\/binding\/Release\/node-v72-darwin-x64\/fse.node\" is installed via remote\n\n&gt; @tensorflow\/tfjs-node@1.2.9 install \/Users\/username\/Documents\/Development\/AI\/Tensorflow\/node_modules\/@tensorflow\/tfjs-node\n&gt; node scripts\/install.js\n\nCPU-darwin-1.2.9.tar.gz\n* Downloading libtensorflow\n[==============================] 10422374\/bps 100% 0.0s\n* Building TensorFlow Node.js bindings\nnode-pre-gyp install failed with error: Error: Command failed: node-pre-gyp install --fallback-to-build\nnode-pre-gyp WARN Using needle for node-pre-gyp https download \nnode-pre-gyp WARN Tried to download(404): https:\/\/storage.googleapis.com\/tf-builds\/pre-built-binary\/napi-v4\/1.2.9\/CPU-darwin-1.2.9.tar.gz \nnode-pre-gyp WARN Pre-built binaries not found for @tensorflow\/tfjs-node@1.2.9 and node@12.9.0 (node-v72 ABI, unknown) (falling back to source compile with node-gyp) \nNo receipt for 'com.apple.pkg.CLTools_Executables' found at '\/'.\n\nNo receipt for 'com.apple.pkg.DeveloperToolsCLILeo' found at '\/'.\n\nNo receipt for 'com.apple.pkg.DeveloperToolsCLI' found at '\/'.\n\ngyp: No Xcode or CLT version detected!\ngyp ERR! configure error \ngyp ERR! stack Error: `gyp` failed with exit code: 1\ngyp ERR! stack     at ChildProcess.onCpExit (\/usr\/local\/lib\/node_modules\/npm\/node_modules\/node-gyp\/lib\/configure.js:344:16)\ngyp ERR! stack     at ChildProcess.emit (events.js:209:13)\ngyp ERR! stack     at Process.ChildProcess._handle.onexit (internal\/child_process.js:272:12)\ngyp ERR! System Darwin 19.0.0\ngyp ERR! command \"\/usr\/local\/bin\/node\" \"\/usr\/local\/lib\/node_modules\/npm\/node_modules\/node-gyp\/bin\/node-gyp.js\" \"configure\" \"--fallback-to-build\" \"--module=\/Users\/username\/Documents\/Development\/AI\/Tensorflow\/node_modules\/@tensorflow\/tfjs-node\/lib\/napi-v4\/tfjs_binding.node\" \"--module_name=tfjs_binding\" \"--module_path=\/Users\/username\/Documents\/Development\/AI\/Tensorflow\/node_modules\/@tensorflow\/tfjs-node\/lib\/napi-v4\" \"--napi_version=4\" \"--node_abi_napi=napi\" \"--napi_build_version=4\" \"--node_napi_label=napi-v4\"\ngyp ERR! cwd \/Users\/username\/Documents\/Development\/AI\/Tensorflow\/node_modules\/@tensorflow\/tfjs-node\ngyp ERR! node -v v12.9.0\ngyp ERR! node-gyp -v v5.0.3\ngyp ERR! not ok \nnode-pre-gyp ERR! build error \nnode-pre-gyp ERR! stack Error: Failed to execute '\/usr\/local\/bin\/node \/usr\/local\/lib\/node_modules\/npm\/node_modules\/node-gyp\/bin\/node-gyp.js configure --fallback-to-build --module=\/Users\/username\/Documents\/Development\/AI\/Tensorflow\/node_modules\/@tensorflow\/tfjs-node\/lib\/napi-v4\/tfjs_binding.node --module_name=tfjs_binding --module_path=\/Users\/username\/Documents\/Development\/AI\/Tensorflow\/node_modules\/@tensorflow\/tfjs-node\/lib\/napi-v4 --napi_version=4 --node_abi_napi=napi --napi_build_version=4 --node_napi_label=napi-v4' (1)\nnode-pre-gyp ERR! stack     at ChildProcess.&lt;anonymous&gt; (\/Users\/username\/Documents\/Development\/AI\/Tensorflow\/node_modules\/node-pre-gyp\/lib\/util\/compile.js:83:29)\nnode-pre-gyp ERR! stack     at ChildProcess.emit (events.js:209:13)\nnode-pre-gyp ERR! stack     at maybeClose (internal\/child_process.js:1021:16)\nnode-pre-gyp ERR! stack     at Process.ChildProcess._handle.onexit (internal\/child_process.js:283:5)\nnode-pre-gyp ERR! System Darwin 19.0.0\nnode-pre-gyp ERR! command \"\/usr\/local\/bin\/node\" \"\/Users\/username\/Documents\/Development\/AI\/Tensorflow\/node_modules\/.bin\/node-pre-gyp\" \"install\" \"--fallback-to-build\"\nnode-pre-gyp ERR! cwd \/Users\/username\/Documents\/Development\/AI\/Tensorflow\/node_modules\/@tensorflow\/tfjs-node\nnode-pre-gyp ERR! node -v v12.9.0\nnode-pre-gyp ERR! node-pre-gyp -v v0.13.0\nnode-pre-gyp ERR! not ok \n\n\n&gt; nodemon@1.19.2 postinstall \/Users\/username\/Documents\/Development\/AI\/Tensorflow\/node_modules\/nodemon\n&gt; node bin\/postinstall || exit 0\n\nnpm WARN tensorflow@1.0.0 No description\nnpm WARN tensorflow@1.0.0 No repository field.\n\nadded 362 packages from 202 contributors and audited 2419 packages in 11.642s\nfound 0 vulnerabilities\n\n<\/code><\/pre>\n\n<p>node-pre-gyp failes to install and 'CPU-darwin-1.2.9.tar.gz' no longer seems to exist, so I get an 404 error.<\/p>\n\n<p>Also this question if probably somewhat of a duplicate to this question:\n<a href=\"https:\/\/stackoverflow.com\/questions\/57537386\/cannot-import-tensorflow-tfjs-node-in-nodejs\">Cannot import @tensorflow\/tfjs-node in nodejs<\/a><\/p>\n\n<p>EDIT: Tried to downgrade to @tensorflow\/tfjs-node@1.2.8, but the error is still there:<\/p>\n\n<pre><code>CPU-darwin-1.2.8.tar.gz\n* Downloading libtensorflow\n[==============================] 10642013\/bps 100% 0.0s\n* Building TensorFlow Node.js bindings\nnode-pre-gyp install failed with error: Error: Command failed: node-pre-gyp install --fallback-to-build\nnode-pre-gyp WARN Using needle for node-pre-gyp https download \nnode-pre-gyp WARN Tried to download(404): https:\/\/storage.googleapis.com\/tf-builds\/pre-built-binary\/napi-v4\/1.2.8\/tfjs_binding-v1.2.8-node-v72-darwin-x64.tar.gz \nnode-pre-gyp WARN Pre-built binaries not found for @tensorflow\/tfjs-node@1.2.8 and node@12.9.0 (node-v72 ABI, unknown) (falling back to source compile with node-gyp) \n<\/code><\/pre>\n","answers":[{"AnswerId":"57835445","CreationDate":"2019-09-07T16:15:08.200","ParentId":null,"OwnerUserId":"547640","Title":null,"Body":"<p><code>node-pre-gyp<\/code> is a nice tool in that it can download precompiled binaries for you which can save a lot of time. The problem is that there still has to be binaries built for your exact version of nodejs.<\/p>\n\n<p>In your case it seems like none are available (the below line is in the logfile you posted).<\/p>\n\n<pre><code>node-pre-gyp WARN Tried to download(404): https:\/\/storage.googleapis.com\/tf-builds\/pre-built-binary\/napi-v4\/1.2.9\/CPU-darwin-1.2.9.tar.gz \nnode-pre-gyp WARN Pre-built binaries not found for @tensorflow\/tfjs-node@1.2.9 and node@12.9.0 (node-v72 ABI, unknown) (falling back to source compile with node-gyp) \n<\/code><\/pre>\n\n<p>If we look at the url it tried to download however there seems to be a file available for version <a href=\"https:\/\/storage.googleapis.com\/tf-builds\/pre-built-binary\/napi-v4\/1.2.8\/CPU-darwin-1.2.8.tar.gz\" rel=\"nofollow noreferrer\">1.2.8<\/a> (you can get a list of available files at <a href=\"https:\/\/storage.googleapis.com\/tf-builds\/\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tf-builds\/<\/a>).<\/p>\n\n<p>So either you downgrade @tensorflow\/tfjs-node to 1.2.8 and tries that or you check out the <a href=\"https:\/\/github.com\/nodejs\/node-gyp#on-macos\" rel=\"nofollow noreferrer\">instructions for node-gyp<\/a> to actually get it to build stuff on your machine (which is probably the best thing long term).<\/p>\n"},{"AnswerId":"57947647","CreationDate":"2019-09-15T19:43:43.560","ParentId":null,"OwnerUserId":"6063676","Title":null,"Body":"<p>I solved the same problem with this command<\/p>\n\n<p><strong>npm install --production windows-build-tools<\/strong><\/p>\n"}]}
+{"QuestionId":57835609,"AnswerCount":2,"Tags":"<python><tensorflow><keras><pycharm>","CreationDate":"2019-09-07T16:36:31.950","AcceptedAnswerId":57851715.0,"OwnerUserId":9560068.0,"Title":"pycharm says taht it cant find tensorflow.keras but it works","Body":"<p>it always says tat it tf.keras is an \"Unresolved References\" but it works. why and how can i get to reconize it?<\/p>\n\n<p>I tryed using <\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorflow.keras\n<\/code><\/pre>\n\n<p>but then it just wont work at all<\/p>\n","answers":[{"AnswerId":"57851715","CreationDate":"2019-09-09T09:58:20.420","ParentId":null,"OwnerUserId":"2787185","Title":null,"Body":"<p>Which <code>tensorflow<\/code> version do you use? There is a known issue with 2.0.0 <a href=\"https:\/\/youtrack.jetbrains.com\/issue\/PY-37589\" rel=\"nofollow noreferrer\">https:\/\/youtrack.jetbrains.com\/issue\/PY-37589<\/a><\/p>\n"},{"AnswerId":"57835631","CreationDate":"2019-09-07T16:39:48.043","ParentId":null,"OwnerUserId":"8302848","Title":null,"Body":"<p>You have to set your projects' python interpreter to the one that you are using.<\/p>\n\n<p>Try to change it in <code>File-&gt;settings-&gt;Project-&gt;project interpreter<\/code> and then select the python from the environment where you have tensorflow installed<\/p>\n"}]}
+{"QuestionId":57835678,"AnswerCount":2,"Tags":"<python><keras><scikit-learn><classification><cross-validation>","CreationDate":"2019-09-07T16:46:18.450","AcceptedAnswerId":57836943.0,"OwnerUserId":12026486.0,"Title":"Keras does not accept the target label due to its shape?","Body":"<p>I am trying to get Keras working with a classification problem, which has five categorical target labels (1, 2, 3, 4, 5). For some reason I am unable to get it working, while using StratifiedKFold. X and y are NumPy arrays with shapes (500, 20) and (500, ), respectively. <\/p>\n\n<p>The error message is \"ValueError: Error when checking target: expected dense_35 to have shape (1,) but got array with shape (5,)\", which leads me to think that the error definitely lies in the format of the target variable. It is also notable, that the number in \"dense_35\" seems to be varying for each attempt of trying to run the code.<\/p>\n\n<pre><code>random_state = 123\nn_splits = 10\ncv = StratifiedKFold(n_splits=n_splits, \nrandom_state=random_state, shuffle=False)\n\ndef baseline_model():\n    nn_model = Sequential()\n    nn_model.add(Dense(units=50, input_dim=X.shape[1], init='normal',\n                       activation= 'relu' ))\n    nn_model.add(Dense(30, init='normal', activation='relu'))\n    nn_model.add(Dense(10, init='normal', activation='relu'))\n    nn_model.add(Dense(1, init='normal', activation='softmax'))\n\n    nn_model.compile(optimizer='adam', loss='categorical_crossentropy',\n                     metrics = ['accuracy'])\n    return nn_model\n\nfor train, test in cv.split(X, y):   \n    X_train, X_test = X[train], X[test]\n    y_train, y_test = y[train], y[test]\n\n    np_utils.to_categorical(y_train)\n    np_utils.to_categorical(y_test)\n\n    estimator = KerasClassifier(build_fn=baseline_model,\n                                epochs=200, batch_size=5,\n                                verbose=0)\n\n    estimator.fit(X_train, y_train)\n    y_pred = estimator.predict(X_test)\n\nThe numpy array (y), that I am trying to split:\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 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 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 2 2 2 2 2 2 2 2 2 2 2\n 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 2 2 2 2 2\n 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 2 2 2 2 2\n 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3\n 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3\n 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3\n 3 3 3 3 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4\n 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4\n 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 5 5 5 5 5 5 5\n 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5\n 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5\n 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5]\n<\/code><\/pre>\n","answers":[{"AnswerId":"58137079","CreationDate":"2019-09-27T15:02:22.053","ParentId":null,"OwnerUserId":"3660313","Title":null,"Body":"<p>Since it's a classification problem, which has five categorical target labels, the last dense layer (output layer) must have 5 units:<\/p>\n\n<pre><code>def baseline_model():\n    nn_model = Sequential()\n    nn_model.add(Dense(units=50, input_dim=X.shape[1], init='normal',\n                       activation= 'relu' ))\n    nn_model.add(Dense(30, init='normal', activation='relu'))\n    nn_model.add(Dense(10, init='normal', activation='relu'))\n\n    #Output layer\n    nn_model.add(Dense(5, init='normal', activation='softmax'))\n    nn_model.compile(optimizer='adam', loss='categorical_crossentropy',\n                     metrics = ['accuracy'])\n    return nn_model\n<\/code><\/pre>\n"},{"AnswerId":"57836943","CreationDate":"2019-09-07T19:40:37.260","ParentId":null,"OwnerUserId":"11952444","Title":null,"Body":"<pre><code>random_state = 123\nn_splits = 10\ncv = StratifiedKFold(n_splits=n_splits, random_state=random_state, shuffle=False)\ndef baseline_model():\n    nn_model = Sequential(name='model_name')\n    nn_model.add(Dense(units=50, input_dim=X.shape[1], init='normal',\n                       activation= 'relu', name='dense1'))\n    nn_model.add(Dense(30, init='normal', activation='relu', name='dense2'))\n    nn_model.add(Dense(10, init='normal', activation='relu', name='dense3'))\n    # code changed here\n    nn_model.add(Dense(5, init='normal', activation='softmax', name='dense4'))\n\n    nn_model.compile(optimizer='adam', loss='categorical_crossentropy',\n                     metrics = ['accuracy'])\n    return nn_model\n\nfor train, test in cv.split(X, y):   \n    X_train, X_test = X[train], X[test]\n    y_train, y_test = y[train], y[test]\n\n    # the error is due to this step\n    # you have specified only one output in the last dense layer (dense4)\n    # but you are giving input of length 5\n    np_utils.to_categorical(y_train)\n    np_utils.to_categorical(y_test)\n\n    estimator = KerasClassifier(build_fn=baseline_model,\n                                epochs=200, batch_size=5,\n                                verbose=0)\n    estimator.fit(X_train, y_train)\n    y_pred = estimator.predict(X_test)\n<\/code><\/pre>\n\n<ul>\n<li>By specifying the <strong><em>name<\/em><\/strong> parameters in the layer you can name your layers. By doing so you will get the definite name of the layers in case of error every time.<\/li>\n<li><strong>model.summary()<\/strong> is another useful function with which you can check the output shape of each layer.<\/li>\n<\/ul>\n"}]}
+{"QuestionId":57835948,"AnswerCount":2,"Tags":"<python><python-2.7><pytorch><torch>","CreationDate":"2019-09-07T17:24:24.310","AcceptedAnswerId":null,"OwnerUserId":8601916.0,"Title":"how to install pytorch in python2.7?","Body":"<p>i am using python2.7 in virtual environment. i tried to install pytorch in python2.7 but i got error belows:<\/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>UnsatisfiableError: The following specifications were found\r\nto be incompatible with the existing python installation in your environment:\r\n\r\n  - pytorch-cpu -&gt; python[version='3.5.*|3.6.*']\r\n  - pytorch-cpu -&gt; python[version='&gt;=3.5,&lt;3.6.0a0|&gt;=3.6,&lt;3.7.0a0|&gt;=3.7,&lt;3.8.0a0']\r\n\r\nIf python is on the left-most side of the chain, that's the version you've asked for.\r\nWhen python appears to the right, that indicates that the thing on the left is somehow\r\nnot available for the python version you are constrained to.  Your current python version\r\nis (python=2.7).  Note that conda will not change your python version to a different minor version\r\nunless you explicitly specify that.\r\n\r\nThe following specifications were found to be incompatible with each other:\r\n\r\n\r\n\r\nPackage wheel conflicts for:\r\npython=2.7 -&gt; pip -&gt; wheel\r\npytorch-cpu -&gt; python[version='&gt;=3.6,&lt;3.7.0a0'] -&gt; pip -&gt; wheel\r\nPackage vc conflicts for:\r\npython=2.7 -&gt; sqlite[version='&gt;=3.27.2,&lt;4.0a0'] -&gt; vc[version='14.*|&gt;=14,&lt;15.0a0|&gt;=14.1,&lt;15.0a0']\r\npython=2.7 -&gt; vc[version='9.*|&gt;=9,&lt;10.0a0']\r\npytorch-cpu -&gt; numpy[version='&gt;=1.11'] -&gt; vc[version='14|14.*|&gt;=14,&lt;15.0a0']\r\npytorch-cpu -&gt; vc[version='&gt;=14.1,&lt;15.0a0']\r\nPackage cffi conflicts for:\r\npytorch-cpu -&gt; cffi\r\npytorch-cpu -&gt; python[version='&gt;=3.6,&lt;3.7.0a0'] -&gt; pip -&gt; requests -&gt; urllib3[version='&gt;=1.21.1,&lt;1.25'] -&gt; cryptography[version='&gt;=1.3.4'] -&gt; cffi[version='&gt;=1.7']\r\npython=2.7 -&gt; pip -&gt; requests -&gt; urllib3[version='&gt;=1.21.1,&lt;1.25'] -&gt; cryptography[version='&gt;=1.3.4'] -&gt; cffi[version='&gt;=1.7']\r\nPackage pip conflicts for:\r\npython=2.7 -&gt; pip\r\npytorch-cpu -&gt; python[version='&gt;=3.6,&lt;3.7.0a0'] -&gt; pip\r\nPackage setuptools conflicts for:\r\npython=2.7 -&gt; pip -&gt; setuptools\r\npytorch-cpu -&gt; python[version='&gt;=3.6,&lt;3.7.0a0'] -&gt; pip -&gt; setuptools\r\nPackage msgpack-python conflicts for:\r\npython=2.7 -&gt; pip -&gt; cachecontrol -&gt; msgpack-python\r\npytorch-cpu -&gt; python[version='&gt;=3.6,&lt;3.7.0a0'] -&gt; pip -&gt; cachecontrol -&gt; msgpack-python<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>i tried conda install pytorch-cpu -c pytorch and link(<a href=\"https:\/\/pytorch.org\/get-started\/locally\/\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/get-started\/locally\/<\/a>). but it is not worked. so what should i do for install torch in python version2.7? i want to install pytorch cpu version. <\/p>\n\n<p>plz help:)<\/p>\n","answers":[{"AnswerId":"57836170","CreationDate":"2019-09-07T17:55:03.347","ParentId":null,"OwnerUserId":"11952444","Title":null,"Body":"<p>Here's the link to the <a href=\"https:\/\/pytorch.org\" rel=\"nofollow noreferrer\">PyTorch official download page<\/a><\/p>\n\n<p>From here, you can choose the python version (2.7) and CUDA (None) and other relevant details based on your environment and OS.<\/p>\n\n<p>Other helpful links:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/49918479\/how-to-install-pytorch-in-anaconda-with-conda-or-pip\">windows<\/a><br><\/li>\n<li><a href=\"https:\/\/medium.com\/@bryant.kou\/how-to-install-pytorch-on-windows-step-by-step-cc4d004adb2a\" rel=\"nofollow noreferrer\">windows<\/a><br><\/li>\n<li><a href=\"https:\/\/dev.to\/berry_clione\/install-pytorch-on-mac-by-pip-2fga\" rel=\"nofollow noreferrer\">mac<\/a><br><\/li>\n<li><a href=\"https:\/\/www.learnopencv.com\/installing-deep-learning-frameworks-on-ubuntu-with-cuda-support\/\" rel=\"nofollow noreferrer\">ubuntu<\/a><br><\/li>\n<li><a href=\"https:\/\/www.javatpoint.com\/pytorch-installation\" rel=\"nofollow noreferrer\">all<\/a><\/li>\n<\/ul>\n"},{"AnswerId":"57836026","CreationDate":"2019-09-07T17:33:35.423","ParentId":null,"OwnerUserId":"11902407","Title":null,"Body":"<p>It seems that pytorch try to install some packages which requires python >= 3.6 version.<\/p>\n\n<pre><code>Package wheel conflicts for:\npython=2.7 -&gt; pip -&gt; wheel\npytorch-cpu -&gt; python[version='&gt;=3.6,&lt;3.7.0a0'] -&gt; pip -&gt; wheel\n<\/code><\/pre>\n\n<p>You have to specify in command for which python version would you like to install pytorch<\/p>\n"}]}
+{"QuestionId":57836126,"AnswerCount":1,"Tags":"<python><tensorflow><logging><keras>","CreationDate":"2019-09-07T17:50:39.423","AcceptedAnswerId":null,"OwnerUserId":4139143.0,"Title":"Keras interfering with python logging","Body":"<p>I want to log some python code which loads in a trained keras model. For some reason, (python) logging doesn't not work if the keras <code>load_model<\/code> is imported. But (python) logging works fine if I don't import keras.<\/p>\n\n<p>This works:<\/p>\n\n<pre><code>import logging\n\nLOG_FORMAT = \"%(levelname)s %(asctime)s - %(message)s\"\nlogging.basicConfig(filename=\"logs\/my_logs.log\",\n                    level=logging.DEBUG,\n                    format=LOG_FORMAT)\nlogger = logging.getLogger()\n\ndef do_stuff():\n    logging.info(\"About to do stuff\")\n    ... stuff gets done here...\n<\/code><\/pre>\n\n<p>This doesn't work<\/p>\n\n<pre><code>import logging\nfrom keras.models import load_model\n\nmy_model = load_model(\"fetch_preprocess\/lstm_model.h5\")\n\nLOG_FORMAT = \"%(levelname)s %(asctime)s - %(message)s\"\nlogging.basicConfig(filename=\"logs\/my_logs.log\",\n                    level=logging.DEBUG,\n                    format=LOG_FORMAT)\nlogger = logging.getLogger()\n\ndef do_stuff():\n    logging.info(\"About to do stuff\")\n    pred = my_model.predict(data)\n    ... stuff gets done here...\n<\/code><\/pre>\n\n<p>By \"doesn't work\", I mean the <code>logging<\/code> module doesn't<\/p>\n\n<ul>\n<li>create a new log file<\/li>\n<li>write to that file anytime <code>logging<\/code> is called<\/li>\n<\/ul>\n\n<p>but no error's are ever thrown. So I find this strange.<\/p>\n\n<p>I believe Keras \/ tensorflow is interfering with the logging (as the only change I make to the code is to exclude Keras and then logging works fine).<\/p>\n\n<p>Is there some way to suppress whatever keras is doing in the background so I can use the python logging?<\/p>\n","answers":[{"AnswerId":"57836709","CreationDate":"2019-09-07T19:06:52.230","ParentId":null,"OwnerUserId":"1849975","Title":null,"Body":"<p>As you mentioned, I also think that <code>Keras<\/code> is simply replacing your logger with its own. You can, however, do the following:<\/p>\n\n<pre><code>import logging as my_logging\nfrom keras.models import load_model\n\nmy_model = load_model(\"fetch_preprocess\/lstm_model.h5\")\n\nLOG_FORMAT = \"%(levelname)s %(asctime)s - %(message)s\"\nmy_logging.basicConfig(filename=\"logs\/my_logs.log\",\n                    level=my_logging.DEBUG,\n                    format=LOG_FORMAT)\nmy_logger = my_logging.getLogger()\n<\/code><\/pre>\n"}]}
+{"QuestionId":57836355,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><computer-vision><pytorch>","CreationDate":"2019-09-07T18:17:44.810","AcceptedAnswerId":57843010.0,"OwnerUserId":11465985.0,"Title":"How to split and load huge dataset that doesn't fit into memory into pytorch Dataloader?","Body":"<p>I'm training a deep learning model to do multi label classification of diseases in NIH's Chest Xray-14 dataset using Google's Colab. I can't load all images into Dataloader at once, given around 112k training examples and limited RAM.<\/p>\n\n<p>Is there a way to just store path of images in pytorch's DataLoader, reading only those images needed for current iteration during training, and once iteration is complete, the images are unloaded from memory, so on so forth until one epoch is complete.<\/p>\n","answers":[{"AnswerId":"57843010","CreationDate":"2019-09-08T14:50:54.243","ParentId":null,"OwnerUserId":"2790047","Title":null,"Body":"<p>Yes, the default behavior for the <a href=\"https:\/\/pytorch.org\/docs\/stable\/torchvision\/datasets.html#imagefolder\" rel=\"nofollow noreferrer\">ImageFolder<\/a> is to create a list of image paths and load the actual images only when needed. It doesn't support multiclass labels. However, you can write your own <a href=\"https:\/\/pytorch.org\/docs\/stable\/data.html#torch.utils.data.Dataset\" rel=\"nofollow noreferrer\">Dataset<\/a> to support multi-label, referencing the <a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torchvision\/datasets\/folder.html#ImageFolder\" rel=\"nofollow noreferrer\">ImageFolder<\/a> class for details.<\/p>\n\n<p>During <code>__init__<\/code> you construct a list of image paths and a corresponding list of labels. Images should only be loaded only when <code>__getitem__<\/code> is invoked. Below is a stub of such a dataset class, the details will depend on the organization of your files, image types, and label format.<\/p>\n\n<pre><code>class CustomDataset(torch.utils.data.Dataset):\n    def __init__(self, args):\n        \"\"\" Construct an indexed list of image paths and labels \"\"\"\n\n    def __getitem__(self, n):\n        \"\"\" Load image n in the list of image paths and return it along with its label.\n            In the case of multiclass the label will probably be a list of values\"\"\"\n\n    def __len__(self):\n        \"\"\" return the total number of images in this dataset \"\"\"\n<\/code><\/pre>\n\n<p>Once you've created a valid dataset instance an instance of <a href=\"https:\/\/pytorch.org\/docs\/stable\/data.html#torch.utils.data.DataLoader\" rel=\"nofollow noreferrer\">DataLoader<\/a> should be created, providing your dataset as an argument. A DataLoader is responsible for sampling its dataset, i.e. invoking the <code>__getitem__<\/code> method you wrote, and putting individual samples into mini-batches. It also handles parallelized loading and defines how the indices are sampled. The DataLoader itself doesn't store more than it needs. The maximum number of samples it should hold in memory at any time is <code>batch_size * num_workers<\/code> (or <code>batch_size<\/code> if <code>num_workers == 0<\/code>).<\/p>\n"}]}
+{"QuestionId":57836600,"AnswerCount":0,"Tags":"<r><keras><neural-network><classification>","CreationDate":"2019-09-07T18:51:20.510","AcceptedAnswerId":null,"OwnerUserId":12035471.0,"Title":"Neural networks only classifies to one class out of five","Body":"<p>I want to compare random forests with neural networks on a dataset that has only categorical inputs and a categorical output. The output has 5 classes, and the current neural net I fit only classifies to one class and gives the same probability for each class for all the data inputs. I'm not sure why I get the same classification for each datapoint. \nI used Keras through R. <\/p>\n\n<pre><code>model %&gt;% \n  layer_dense(units = 25, activation = \"relu\", input_shape = 41) %&gt;% \n  layer_dropout(rate = 0.11) %&gt;% \n  layer_dense(units =18, activation = \"relu\") %&gt;%\n  layer_dropout(rate = 0.001) %&gt;%\n  layer_dense(units = 5, activation = \"softmax\")\n\nmodel %&gt;% compile(\n  loss = 'categorical_crossentropy',\n  optimizer  =optimizer_sgd(lr=0.01),\n  metrics=c('accuracy')\n)\n\nhistory &lt;- model %&gt;% fit(x_train, y_train, \n                         epochs =100,batch_size = 50,\n                         validation_split = 0.2)\n\n model %&gt;% predict(x_test)\n<\/code><\/pre>\n\n<p>this is the output I get. It classifies each training point to the class with the highest probability, which is the third column in the output. <\/p>\n\n<pre><code>&gt; (predictions)\n            [,1]      [,2]      [,3]      [,4]      [,5]\n  [1,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n  [2,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n  [3,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n  [4,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n  [5,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n  [6,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n  [7,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n  [8,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n  [9,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [10,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [11,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [12,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [13,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [14,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [15,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [16,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [17,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [18,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [19,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [20,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [21,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [22,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [23,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [24,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [25,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [26,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [27,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [28,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [29,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [30,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [31,] 0.2001473 0.1990314 0.3208318 0.1130381 0.1669514\n [32,] 0.2001473 0.1990314 0.3208318 0.1130381 0.1669514\n [33,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [34,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [35,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [36,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [37,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [38,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [39,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [40,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [41,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [42,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n [43,] 0.2001473 0.1990315 0.3208320 0.1130381 0.1669514\n<\/code><\/pre>\n\n<p>but the actual test target looks like this:<\/p>\n\n<pre><code>1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 3 2 3 2 2 3 3 2 2 3 2\n2 2 3 2 3 3 3 3 3 3 2 2 3 2 3 3 2 3 2 2 2 2 2 3 3 3 3 3 2 3 2 3 3\n3 3 3 3 2 5 4 5 5 4 5 4 5 4 5 4 5 5 5 4 5 4 5 4 5 4 5 4 5 4 4 4 4\n5 4 4 4 5 4 4 5 4 5 5 5 5 5 5 5 5 5 5 5 5 \n<\/code><\/pre>\n\n<p>This is what my model predicts for the test parget:<\/p>\n\n<pre><code>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 2\n2 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\n2 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\n2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57836679,"AnswerCount":2,"Tags":"<python><machine-learning><deep-learning><pytorch>","CreationDate":"2019-09-07T19:03:11.747","AcceptedAnswerId":57836815.0,"OwnerUserId":8618806.0,"Title":"For each layer in neural networks (pytorch), how many biases should be there?","Body":"<p>I have a simple model in pytorch. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Network()\n<\/code><\/pre>\n\n<p>It's details are:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Network(\n  (hidden): Linear(in_features=784, out_features=256, bias=True)\n  (output): Linear(in_features=256, out_features=10, bias=True)\n  (sigmoid): Sigmoid()\n  (softmax): Softmax(dim=1)\n)\n<\/code><\/pre>\n\n<p>There are 3 neurons' layers in total. 1 input(786 neurons), 1 hidden(256 neurons) and 1 output (10 neurons). Therefore there'll be two weight layers. So there must be two biases (simply two floating point numbers) for both weight layers right? (correct me if i am wrong).  <\/p>\n\n<p>Now after initializing my network i was curious about the two bias values. So i wanted to check the bias value of my hidden layer so i wrote:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.hidden.bias\n<\/code><\/pre>\n\n<p>And what i got as the result was not i expected! I actually expected one value! And this is what i actually got: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tensor([-1.6868e-02, -3.5661e-02,  1.2489e-02, -2.7880e-02,  1.4025e-02,\n        -2.6085e-02,  1.2625e-02, -3.1748e-02,  5.0335e-03,  3.8031e-03,\n        -3.1648e-02, -3.4881e-02, -2.0026e-02,  1.9728e-02,  6.2461e-03,\n         9.3936e-04, -5.9270e-03, -2.7183e-02, -1.9850e-02, -3.5693e-02,\n        -1.9393e-02,  2.6555e-02,  2.3482e-02,  2.1230e-02, -2.2175e-02,\n        -2.4386e-02,  3.4848e-02, -2.6044e-02,  1.3575e-02,  9.4125e-03,\n         3.0012e-02, -2.6078e-02,  7.1615e-05, -1.7061e-02,  6.6355e-03,\n        -3.4966e-02,  2.9311e-02,  1.4060e-02, -2.5763e-02, -1.4020e-02,\n         2.9852e-02, -7.9176e-03, -1.8396e-02,  1.6927e-02, -1.1001e-03,\n         1.5595e-02,  1.2169e-02, -1.2275e-02, -2.9270e-03, -6.5685e-04,\n        -2.4297e-02,  3.0048e-02,  2.9692e-03, -2.5398e-02,  2.9955e-03,\n        -9.3653e-04, -1.2932e-02,  2.4232e-02, -3.5182e-02, -1.6163e-02,\n         3.0025e-02,  3.1227e-02, -8.2498e-04,  2.7102e-02, -2.3830e-02,\n        -3.4958e-02, -1.1886e-02,  1.6097e-02,  1.4579e-02, -2.6744e-02,\n         1.1900e-02, -3.4855e-02, -4.2208e-03, -5.2035e-03,  1.7055e-02,\n        -4.8580e-03,  3.4088e-03,  1.6923e-02,  3.5570e-04, -3.0478e-02,\n         8.4647e-03,  2.5704e-02, -2.3255e-02,  6.9396e-03, -1.2521e-03,\n        -9.4101e-03, -2.5798e-02, -1.4438e-03, -7.2684e-03,  3.5417e-02,\n        -3.4388e-02,  1.3706e-02, -5.1430e-03,  1.6174e-02,  1.8135e-03,\n        -2.9018e-02, -2.9083e-02,  7.4100e-03, -2.7758e-02,  2.4367e-02,\n        -3.8350e-03,  9.4390e-03, -1.0844e-02,  1.6381e-02, -2.5268e-02,\n         1.3553e-02, -1.0545e-02, -1.3782e-02,  2.8519e-02,  2.3630e-02,\n        -1.9703e-02, -2.0147e-02, -1.0485e-02,  2.4637e-02,  1.9989e-02,\n         5.6601e-03,  1.9121e-02, -1.5286e-02,  2.5996e-02, -2.9833e-02,\n        -2.9458e-02,  2.3944e-02, -3.0107e-02, -1.2307e-02, -1.8419e-02,\n         3.3551e-02,  1.2396e-02,  2.9356e-02,  3.3274e-02,  5.4677e-03,\n         3.1715e-02,  1.3361e-02,  3.3042e-02,  2.7843e-03,  2.2837e-02,\n        -3.4981e-02,  3.2355e-02, -2.7658e-03,  2.2184e-02, -2.0203e-02,\n        -3.3264e-02, -3.4858e-02,  1.0820e-03, -1.4279e-02, -2.8041e-02,\n         4.1962e-03,  2.4266e-02, -3.5704e-02, -2.6172e-02,  2.3335e-02,\n         2.0657e-02, -3.0387e-03, -5.7096e-03, -1.1062e-02,  1.3450e-02,\n        -3.3965e-02,  1.9623e-03, -2.0067e-02, -3.3858e-02, -2.1931e-02,\n        -1.5414e-02,  2.4454e-02,  2.5668e-02, -1.1932e-02,  5.7540e-04,\n         1.5130e-02,  1.3916e-02, -2.1521e-02, -3.0575e-02,  1.8841e-02,\n        -2.3240e-02, -2.7297e-02, -3.2668e-02, -1.5544e-02, -5.9408e-03,\n         3.0241e-02,  2.2039e-02, -2.4389e-02,  3.1703e-02,  3.5305e-02,\n        -2.7501e-03,  2.0154e-02, -5.3489e-03,  1.4177e-02,  1.6829e-02,\n         3.3066e-02, -1.3425e-02, -3.2565e-02,  6.5624e-03, -1.5681e-02,\n         2.3047e-02,  6.5880e-03, -3.3803e-02,  2.3790e-02, -5.5061e-03,\n         2.9413e-02,  1.2290e-02, -1.0958e-02,  1.2680e-03,  1.3343e-02,\n         6.6689e-03, -2.2975e-03, -1.2068e-02,  1.6523e-02, -3.1612e-02,\n        -1.7529e-02, -2.2220e-02, -1.4723e-02, -1.3495e-02, -5.1805e-03,\n        -2.9620e-02,  3.0571e-02, -3.0999e-02,  3.3681e-03,  1.3579e-02,\n         1.4837e-02,  1.5694e-02, -1.1178e-02,  4.6233e-03, -2.2583e-02,\n        -3.5281e-03,  3.0918e-02,  2.6407e-02,  1.5822e-04, -3.0181e-03,\n         8.6989e-03,  2.8998e-02, -1.5975e-02, -3.1574e-02, -1.5609e-02,\n         1.0472e-02,  5.8976e-03,  7.0131e-03, -3.2047e-02,  2.6045e-02,\n        -2.8882e-02, -2.2121e-02, -3.2960e-02,  1.8268e-02,  3.0984e-02,\n         1.4824e-02,  3.0010e-02, -5.7523e-03, -2.0017e-02,  4.8700e-03,\n         1.4997e-02, -1.4898e-02,  6.8572e-03,  9.7713e-03,  1.3410e-02,\n         4.9619e-03,  3.1016e-02,  3.1240e-02, -3.0203e-02,  2.1435e-02,\n         2.7331e-02], requires_grad=True)\n<\/code><\/pre>\n\n<p>Can someone explain to me this behaviour? Why did i get 256 values instead of one?<\/p>\n\n<p><strong>Edit1:<\/strong><\/p>\n\n<p>Here is my understanding of the layers: \nFor a whole layer of neurons, a bias is just a single value. Am i right? But what i am seeing as the output about are 256 values ? why ? did pytorch assume that i have a bias with each neuron ? is that okay?<br>\n<a href=\"https:\/\/i.stack.imgur.com\/JwEmK.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/JwEmK.png\" alt=\"Bias\"><\/a><\/p>\n","answers":[{"AnswerId":"57836815","CreationDate":"2019-09-07T19:21:29.897","ParentId":null,"OwnerUserId":"8112138","Title":null,"Body":"<p>So first it's important to realize what's going on inside of one of these layers. When you write:<\/p>\n\n<pre><code>Linear(in_features=784, out_features=256, bias=True)\n<\/code><\/pre>\n\n<p>You are modeling a <em>linear relationship<\/em> between the input and the output. You're probably familiar with this from basic math:<\/p>\n\n<pre><code>Y = MX + B\n<\/code><\/pre>\n\n<p>However instead of a \"slope\" and a \"y-intercept\", you have a weights matrix and a bias term. This is still a linear relationship, but with matrices as our input and output. <\/p>\n\n<p>Y is our output, M is our weights matrix, X is our input, and B is our Bias. You define that the input is a (N x 784) matrix, and our output is a (N x 256) matrix (N is the number of samples). <\/p>\n\n<p>If you're familiar with matrix multiplication this means that our weights matrix is (784 X 256). The output of MX will be a (N x 256) matrix, so our bias term must also be (N x 256) for the MX + B to work out.<\/p>\n\n<p>In general the number of values in the bias term will be the same as the number of <code>out_features<\/code>.<\/p>\n"},{"AnswerId":"57841674","CreationDate":"2019-09-08T11:50:39.523","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>Check this out:<\/p>\n\n<pre><code>from torchvision.models import resnet18\nmodel = resnet18(pretrained=False)    \n\nfor name, param in model.named_parameters():\n    if param.requires_grad:\n        print (name)\n<\/code><\/pre>\n\n<p>This will give you a huge list like this:<\/p>\n\n<pre><code>conv1.weight\nbn1.weight\nbn1.bias\nlayer1.0.conv1.weight\nlayer1.0.bn1.weight\nlayer1.0.bn1.bias\nlayer1.0.conv2.weight\nlayer1.0.bn2.weight\nlayer1.0.bn2.bias\nlayer1.1.conv1.weight\nlayer1.1.bn1.weight\nlayer1.1.bn1.bias\nlayer1.1.conv2.weight\nlayer1.1.bn2.weight\nlayer1.1.bn2.bias\nlayer2.0.conv1.weight\nlayer2.0.bn1.weight\nlayer2.0.bn1.bias\nlayer2.0.conv2.weight\nlayer2.0.bn2.weight\nlayer2.0.bn2.bias\nlayer2.0.downsample.0.weight\nlayer2.0.downsample.1.weight\nlayer2.0.downsample.1.bias\nlayer2.1.conv1.weight\nlayer2.1.bn1.weight\nlayer2.1.bn1.bias\nlayer2.1.conv2.weight\nlayer2.1.bn2.weight\nlayer2.1.bn2.bias\nlayer3.0.conv1.weight\nlayer3.0.bn1.weight\nlayer3.0.bn1.bias\nlayer3.0.conv2.weight\nlayer3.0.bn2.weight\nlayer3.0.bn2.bias\nlayer3.0.downsample.0.weight\nlayer3.0.downsample.1.weight\nlayer3.0.downsample.1.bias\nlayer3.1.conv1.weight\nlayer3.1.bn1.weight\nlayer3.1.bn1.bias\nlayer3.1.conv2.weight\nlayer3.1.bn2.weight\nlayer3.1.bn2.bias\nlayer4.0.conv1.weight\nlayer4.0.bn1.weight\nlayer4.0.bn1.bias\nlayer4.0.conv2.weight\nlayer4.0.bn2.weight\nlayer4.0.bn2.bias\nlayer4.0.downsample.0.weight\nlayer4.0.downsample.1.weight\nlayer4.0.downsample.1.bias\nlayer4.1.conv1.weight\nlayer4.1.bn1.weight\nlayer4.1.bn1.bias\nlayer4.1.conv2.weight\nlayer4.1.bn2.weight\nlayer4.1.bn2.bias\nfc.weight\nfc.bias\n<\/code><\/pre>\n\n<p>And you will know all the parameters and get the bases, <strong>however<\/strong> if you print the model you will get:<\/p>\n\n<pre><code>ResNet(\n  (conv1): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)\n  (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n  (relu): ReLU(inplace)\n  (maxpool): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)\n  (layer1): Sequential(\n    (0): BasicBlock(\n      (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n      (relu): ReLU(inplace)\n      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n    )\n    (1): BasicBlock(\n      (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n      (relu): ReLU(inplace)\n      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n    )\n  )\n  (layer2): Sequential(\n    (0): BasicBlock(\n      (conv1): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)\n      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n      (relu): ReLU(inplace)\n      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n      (downsample): Sequential(\n        (0): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False)\n        (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n      )\n    )\n    (1): BasicBlock(\n      (conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n      (relu): ReLU(inplace)\n      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n    )\n  )\n  (layer3): Sequential(\n    (0): BasicBlock(\n      (conv1): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)\n      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n      (relu): ReLU(inplace)\n      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n      (downsample): Sequential(\n        (0): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2), bias=False)\n        (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n      )\n    )\n    (1): BasicBlock(\n      (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n      (relu): ReLU(inplace)\n      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n    )\n  )\n  (layer4): Sequential(\n    (0): BasicBlock(\n      (conv1): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)\n      (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n      (relu): ReLU(inplace)\n      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n      (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n      (downsample): Sequential(\n        (0): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)\n        (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n      )\n    )\n    (1): BasicBlock(\n      (conv1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n      (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n      (relu): ReLU(inplace)\n      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n      (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n    )\n  )\n  (avgpool): AdaptiveAvgPool2d(output_size=(1, 1))\n  (fc): Linear(in_features=512, out_features=1000, bias=True)\n)\n<\/code><\/pre>\n\n<p>The information where bias is set to <code>True<\/code> or <code>False<\/code>, meaning if they will be actually used or not. You can check the last also by modifying the first piece of code, but hope this will be helpful. <\/p>\n"}]}
+{"QuestionId":57836849,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow><tf.keras>","CreationDate":"2019-09-07T19:26:45.373","AcceptedAnswerId":57853140.0,"OwnerUserId":11731187.0,"Title":"Tensorflow.Keras: Custom Constraint Not Working","Body":"<p>Im trying to implement the Weights Orthogonality Constraint showed <a href=\"https:\/\/towardsdatascience.com\/build-the-right-autoencoder-tune-and-optimize-using-pca-principles-part-ii-24b9cca69bd6\" rel=\"nofollow noreferrer\">here<\/a>, in section 2.0. when i try to use it on a Keras Dense Layer, An Value Error is raised.<\/p>\n\n<p>This is happening too when trying to implement the Custom Uncorrelated Features Constraint in the part 3.0 of the same Article.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nclass WeightsOrthogonalityConstraint(tf.keras.constraints.Constraint):\n    def __init__(self, encoding_dim, weightage = 1.0, axis = 0):\n        self.encoding_dim = encoding_dim\n        self.weightage = weightage\n        self.axis = axis\n\n    def weights_orthogonality(self, w):\n        if(self.axis==1):\n            w = tf.keras.backend.transpose(w)\n        if(self.encoding_dim &gt; 1):\n            m = tf.keras.backend.dot(tf.keras.backend.transpose(w), w) - tf.keras.backend.eye(self.encoding_dim)\n            return self.weightage * tf.keras.backend.sqrt(tf.keras.backend.sum(tf.keras.backend.square(m)))\n        else:\n            m = tf.keras.backend.sum(w ** 2) - 1.\n            return m\n\n    def __call__(self, w):\n        return self.weights_orthogonality(w)\n\nrand_samples = np.random.rand(16, 4)\ndummy_ds = tf.data.Dataset.from_tensor_slices((rand_samples, rand_samples)).shuffle(16).batch(16)\n\nencoder = tf.keras.layers.Dense(2, \"relu\", input_shape=(4,), kernel_regularizer=WeightsOrthogonalityConstraint(2))\ndecoder = tf.keras.layers.Dense(4, \"relu\")\n\nautoencoder = tf.keras.models.Sequential()\nautoencoder.add(encoder)\nautoencoder.add(decoder)\n\nautoencoder.compile(metrics=['accuracy'],\n                    loss='mean_squared_error',\n                    optimizer='sgd')\n\nautoencoder.summary()\n\nautoencoder.fit(dummy_ds, epochs=1)\n\n<\/code><\/pre>\n\n<p>If i stop using the contraint, there is no errors, but when used, the next error is raised:<\/p>\n\n<pre><code>2019-09-07 14:20:25.962610: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library nvcuda.dll\n2019-09-07 14:20:26.997957: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1640] Found device 0 with properties: \nname: GeForce GTX 1060 major: 6 minor: 1 memoryClockRate(GHz): 1.733\npciBusID: 0000:01:00.0\n2019-09-07 14:20:27.043016: I tensorflow\/stream_executor\/platform\/default\/dlopen_checker_stub.cc:25] GPU libraries are statically linked, skip dlopen check.        \n2019-09-07 14:20:27.050749: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1763] Adding visible gpu devices: 0\n2019-09-07 14:20:27.081369: I tensorflow\/core\/platform\/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2\n2019-09-07 14:20:27.113598: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1640] Found device 0 with properties: \nname: GeForce GTX 1060 major: 6 minor: 1 memoryClockRate(GHz): 1.733\npciBusID: 0000:01:00.0\n2019-09-07 14:20:27.144194: I tensorflow\/stream_executor\/platform\/default\/dlopen_checker_stub.cc:25] GPU libraries are statically linked, skip dlopen check.        \n2019-09-07 14:20:27.151802: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1763] Adding visible gpu devices: 0\n2019-09-07 14:20:27.800616: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1181] Device interconnect StreamExecutor with strength 1 edge matrix:\n2019-09-07 14:20:27.817323: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1187]      0 \n2019-09-07 14:20:27.840635: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1200] 0:   N \n2019-09-07 14:20:27.848536: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1326] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 4712 MB memory) -&gt; physical GPU (device: 0, name: GeForce GTX 1060, pci bus id: 0000:01:00.0, compute capability: 6.1)\nTraceback (most recent call last):\n  File \"c:\\Users\\whitm\\.vscode\\extensions\\ms-python.python-2019.9.34911\\pythonFiles\\ptvsd_launcher.py\", line 43, in &lt;module&gt;\n    main(ptvsdArgs)\n  File \"c:\\Users\\whitm\\.vscode\\extensions\\ms-python.python-2019.9.34911\\pythonFiles\\lib\\python\\ptvsd\\__main__.py\", line 432, in main\n    run()\n  File \"c:\\Users\\whitm\\.vscode\\extensions\\ms-python.python-2019.9.34911\\pythonFiles\\lib\\python\\ptvsd\\__main__.py\", line 316, in run_file\n    runpy.run_path(target, run_name='__main__')\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\runpy.py\", line 263, in run_path\n    pkg_name=pkg_name, script_name=fname)\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\runpy.py\", line 96, in _run_module_code\n    mod_name, mod_spec, pkg_name, script_name)\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\runpy.py\", line 85, in _run_code\n    exec(code, run_globals)\n  File \"c:\\Users\\whitm\\Desktop\\CodeProjects\\ForestClassifier-DEC\\Test.py\", line 35, in &lt;module&gt;\n    optimizer='sgd')\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\site-packages\\tensorflow\\python\\training\\tracking\\base.py\", line 458, in _method_wrapper\n    result = method(self, *args, **kwargs)\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 337, in compile\n    self._compile_weights_loss_and_weighted_metrics()\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\site-packages\\tensorflow\\python\\training\\tracking\\base.py\", line 458, in _method_wrapper\n    result = method(self, *args, **kwargs)\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 1494, in _compile_weights_loss_and_weighted_metrics\n    self.total_loss = self._prepare_total_loss(masks)\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py\", line 1601, in _prepare_total_loss\n    custom_losses = self.get_losses_for(None) + self.get_losses_for(\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer.py\", line 1209, in get_losses_for\n    return [l for l in self.losses if l._unconditional_loss]\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer.py\", line 835, in losses\n    return collected_losses + self._gather_children_attribute('losses')\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer.py\", line 2129, in _gather_children_attribute      \n    getattr(layer, attribute) for layer in nested_layers))\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer.py\", line 2129, in &lt;genexpr&gt;\n    getattr(layer, attribute) for layer in nested_layers))\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer.py\", line 832, in losses\n    loss_tensor = regularizer()\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer.py\", line 907, in _tag_unconditional\n    loss = loss()\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer.py\", line 1659, in _loss_for_variable\n    regularization = regularizer(v)\n  File \"c:\\Users\\whitm\\Desktop\\CodeProjects\\ForestClassifier-DEC\\Test.py\", line 21, in __call__\n    return self.weights_orthogonality(w)\n  File \"c:\\Users\\whitm\\Desktop\\CodeProjects\\ForestClassifier-DEC\\Test.py\", line 14, in weights_orthogonality\n    m = tf.keras.backend.dot(tf.keras.backend.transpose(w), w) - tf.keras.backend.eye(self.encoding_dim)\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py\", line 1310, in eye\n    return variable(linalg_ops.eye(size, dtype=tf_dtype), dtype, name)\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py\", line 785, in variable\n    constraint=constraint)\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\site-packages\\tensorflow\\python\\ops\\variables.py\", line 264, in __call__\n    return super(VariableMetaclass, cls).__call__(*args, **kwargs)\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\site-packages\\tensorflow\\python\\ops\\resource_variable_ops.py\", line 464, in __init__\n    shape=shape)\n  File \"C:\\ProgramData\\Anaconda3\\envs\\tensorflow-gpu\\lib\\site-packages\\tensorflow\\python\\ops\\resource_variable_ops.py\", line 550, in _init_from_args\n    raise ValueError(\"Tensor-typed variable initializers must either be \"\nValueError: Tensor-typed variable initializers must either be wrapped in an init_scope or callable (e.g., `tf.Variable(lambda : tf.truncated_normal([10, 40]))`) when building functions. Please file a feature request if this restriction inconveniences you.\n<\/code><\/pre>\n\n<p>Thanks in advance!<\/p>\n\n<p>PD: <a href=\"https:\/\/colab.research.google.com\/drive\/19jTi5jRaDKFey0QZ1FQgOUqiz3UqS3xm\" rel=\"nofollow noreferrer\">Here<\/a> is a Colab Notebook showing the error<\/p>\n\n<p>PD2: I anage to found the line causing the problem, is this one:<\/p>\n\n<pre><code>m = tf.keras.backend.dot(tf.keras.backend.transpose(w), w) - tf.keras.backend.eye(self.encoding_dim)\n<\/code><\/pre>\n\n<p>specifically the keras backend eye() function is causing the problem<\/p>\n","answers":[{"AnswerId":"57853140","CreationDate":"2019-09-09T11:30:44.377","ParentId":null,"OwnerUserId":"11731187","Title":null,"Body":"<p>I manage to solve this problem:<\/p>\n\n<p>the function causing the error was tf.keras.backed.eye() on line 14. I read out there that the implementation in the keras backend of this function use numpy array for the identity matrix, but tensorflow and other backends already have their impementation for this function using tensors. There is something with the lack of tensors causing the error on tf2.0, so just change tf.keras.backed.eye() to tf.eye() solve the problem.<\/p>\n"},{"AnswerId":"57837850","CreationDate":"2019-09-07T22:19:32.013","ParentId":null,"OwnerUserId":"11952444","Title":null,"Body":"<p>I checked the same with TensorFlow version <strong>1.14.0<\/strong> and it is working fine.<\/p>\n\n<p>I am sharing my colab notebook link below. Please check!<\/p>\n\n<p><a href=\"https:\/\/colab.research.google.com\/drive\/1_8AqpogMDgFKU2haYk-kCaRO8BhQgy0E\" rel=\"nofollow noreferrer\">Colab Notebook<\/a><\/p>\n"}]}
+{"QuestionId":57836894,"AnswerCount":0,"Tags":"<python><vectorization><pytorch>","CreationDate":"2019-09-07T19:32:54.383","AcceptedAnswerId":null,"OwnerUserId":4572113.0,"Title":"Vectorize loss function with complex indexing","Body":"<p>I have written a loss function in pytorch, however, it is too slow.\nThus, I'm trying to vectorize it.<\/p>\n\n<h1>Description<\/h1>\n\n<p>Given two meshes, it computes the distortion between patches (i.e. set of vertices) defined on them. The distortion is how much the distances between each pair of points changes.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/apPJu.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/apPJu.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<ul>\n<li>Patches may have different sizes<\/li>\n<li>This assumes a correspondence between M and N. <\/li>\n<li>The loss accept batches.<\/li>\n<\/ul>\n\n<h1>Non-vectorized code<\/h1>\n\n<pre class=\"lang-py prettyprint-override\"><code>def loss_func(y_pred, batch):\n    y_true = batch['vertices']\n    patches: Tuple(Tensor,...) = batch['patches']\n\n    loss = 0\n    for i in range(y_pred.shape[0]):\n        patch_loss = 0\n        for patch in patches:\n            dist_a = torch.pdist(y_pred[i][patch])\n            dist_b = torch.pdist(y_true[i][patch])\n\n            patch_loss += torch.sum((dist_a - dist_b) ** 2)\n\n        loss += patch_loss\n    return loss \/ y_pred.shape[0]\n<\/code><\/pre>\n\n<h1>Vectorization attempt<\/h1>\n\n<pre class=\"lang-py prettyprint-override\"><code>def loss_func(y_pred, batch):\n    y_true = batch['vertices']\n    patches: Tuple(Tensor,...) = batch['patches']\n\n    pred_dists = torch.cdist(y_pred, y_pred)\n    true_dists = torch.cdist(y_true, y_true)\n\n    diff_dists = (pred_dists - true_dists) ** 2\n\n    loss = 0\n    for patch in patches:\n        d = diff_dists[:, patch, :][:, :, patch]\n        loss += torch.sum(d, dim=(1, 2)) \/ 2\n\n    return loss.mean()\n<\/code><\/pre>\n\n<h1>Problem<\/h1>\n\n<p>It is still too slow.\nI'm encountering difficulties in the vectorization since the patches may have different dimensions, thus they can't be put in a tensor.<\/p>\n\n<p>I have the feeling that I could use directly the adjacency matrix to vectorize the code, without gathering the vertex indices in patches. <\/p>\n\n<p>The patches are computed in this way:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def get_patch(vertex_index, vertex2vertex):\n    return vertex2vertex[vertex_index, :].nonzero().squeeze(-1)\n<\/code><\/pre>\n\n<h1>Working example<\/h1>\n\n<p>I prepared a self contained example to test and experiment with the two functions, with some benchmarks.\nIn the benchmark there is the same loss function without patches (i.e. just the distortion between all the vertices) as reference time<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nimport time\n\ndevice = 'cuda'\nprecision = torch.double\nbatch_size = 8\nnum_vertices = 400\ndim = 2\n\ny_pred = torch.rand(batch_size, num_vertices, dim, device=device, dtype=precision)\ny_true = torch.rand(batch_size, num_vertices, dim, device=device, dtype=precision)\n\nadjacency = (torch.rand(num_vertices, num_vertices, device=device, dtype=precision) &gt; 0.95).to(precision)\n\ndef get_patch(vertex_index, vertex2vertex):\n    return vertex2vertex[vertex_index, :].nonzero().squeeze(-1)\n\npatches = tuple(get_patch(i, adjacency) for i in range(num_vertices))\nbatch = {\n    'vertices': y_true,\n    'patches': patches\n}\n\ndef loss_func(y_pred, batch):\n    y_true = batch['vertices']\n    patches: Tuple(Tensor,...) = batch['patches']\n\n    loss = 0\n    for i in range(y_pred.shape[0]):\n        patch_loss = 0\n        for patch in patches:\n            dist_a = torch.pdist(y_pred[i][patch])\n            dist_b = torch.pdist(y_true[i][patch])\n\n            patch_loss += torch.sum((dist_a - dist_b) ** 2)\n\n        loss += patch_loss\n    return loss \/ y_pred.shape[0]\n\ndef loss_func_vec(y_pred, batch):\n    y_true = batch['vertices']\n    patches: Tuple(Tensor,...) = batch['patches']\n\n    pred_dists = torch.cdist(y_pred, y_pred)\n    true_dists = torch.cdist(y_true, y_true)\n\n    diff_dists = (pred_dists - true_dists) ** 2\n\n    loss = 0\n    for patch in patches:\n        d = diff_dists[:, patch, :][:, :, patch]\n        loss += torch.sum(d, dim=(1, 2)) \/ 2\n\n    return loss.mean()\n\ndef reference_no_patches(y_pred, batch):\n    y_true = batch['vertices']\n    da = torch.cdist(y_pred, y_pred)\n    db = torch.cdist(y_true, y_true)\n    return torch.sum((da - db) ** 2) \/ y_true.shape[0] \/ 2\n\nprint(f'LOSS FUNCTION (ITERATION): TIME (LOSS)')\nprint()\n\nfor i in range(5):\n    start_time = time.time()\n    res = loss_func(y_pred, batch)\n    torch.cuda.synchronize()\n    print(f'Standard {i}:\\t{time.time() - start_time:.4f}s ({res:.5f})')\nprint()\n\nfor i in range(5):\n    start_time = time.time()\n    res = loss_func_vec(y_pred, batch)\n    torch.cuda.synchronize()\n    print(f'Vectorized {i}:\\t{time.time() - start_time:.4f}s ({res:.5f})')\nprint()\n\nfor i in range(5):\n    start_time = time.time()\n    res = reference_no_patches(y_pred, batch)\n    torch.cuda.synchronize()\n    print(f'Reference without patches {i}:\\t{time.time() - start_time:.4f}s ({res:.5f})')\n<\/code><\/pre>\n\n<pre><code>LOSS FUNCTION (ITERATION): TIME (LOSS)\n\nStandard 0: 0.3311s (9744.87875)\nStandard 1: 0.2853s (9744.87875)\nStandard 2: 0.2929s (9744.87875)\nStandard 3: 0.2972s (9744.87875)\nStandard 4: 0.2714s (9744.87875)\n\nVectorized 0:   0.0254s (9744.87875)\nVectorized 1:   0.0251s (9744.87875)\nVectorized 2:   0.0262s (9744.87875)\nVectorized 3:   0.0242s (9744.87875)\nVectorized 4:   0.0241s (9744.87875)\n\nReference without patches 0:    0.0063s (9695.41083)\nReference without patches 1:    0.0062s (9695.41083)\nReference without patches 2:    0.0062s (9695.41083)\nReference without patches 3:    0.0066s (9695.41083)\nReference without patches 4:    0.0067s (9695.41083)\n<\/code><\/pre>\n\n<h1>Context<\/h1>\n\n<p>The matrices involved are small. The number of vertices in each mesh is ~500, thus the adjacency matrix is <code>[500, 500]<\/code>.\nThe patches may vary in dimension, however each one is &lt;&lt; 500. There is one patch for each vertex (i.e. each vertex is the center of a patch), thus the number of patches is O(n!) (thus, I can't stack the patches into a single vector to perform the indexing).<\/p>\n\n<p>I used the double precision in the example only to avoid numerical errors in the results, I plan to use float precision.<\/p>\n\n<p>Finally, I will be happy to receive any suggestions on how to improve speed!<\/p>\n\n<h1>Update<\/h1>\n\n<p>Due to this <a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/24345\" rel=\"nofollow noreferrer\">bug<\/a> I have to be careful with the memory allocation.\nI rewrote the vectorized code as:<\/p>\n\n<pre><code>def loss_func_vec_mem(y_pred, batch):\n    y_true = batch['vertices']\n    patches: Tuple(Tensor,...) = batch['patches']\n\n    loss = 0\n    for patch in patches:\n        pred = y_pred[:, patch, :]\n        true = y_true[:, patch, :]\n\n        pred_dists = torch.cdist(pred, pred)\n        true_dists = torch.cdist(true, true)\n\n        diff_dists = (pred_dists - true_dists) ** 2\n\n        loss += torch.sum(diff_dists, dim=(1, 2)) \n    return loss.mean() \/ 2\n<\/code><\/pre>\n\n<p>Although it's even slower:<\/p>\n\n<pre><code>LOSS FUNCTION (ITERATION): TIME (LOSS)\n\nVec mem 0:  0.0484s (14685.45521)\nVec mem 1:  0.0489s (14685.45521)\nVec mem 2:  0.0471s (14685.45521)\nVec mem 3:  0.0471s (14685.45521)\nVec mem 4:  0.0459s (14685.45521)\n\nVectorized 0:   0.0271s (29370.91043)\nVectorized 1:   0.0262s (29370.91043)\nVectorized 2:   0.0320s (29370.91043)\nVectorized 3:   0.0266s (29370.91043)\nVectorized 4:   0.0263s (29370.91043)\n\nStandard 0: 0.1680s (14685.45521)\nStandard 1: 0.1614s (14685.45521)\nStandard 2: 0.1639s (14685.45521)\nStandard 3: 0.1606s (14685.45521)\nStandard 4: 0.1694s (14685.45521)\n\nReference without patches 0:    0.0049s (12874.01275)\nReference without patches 1:    0.0105s (12874.01275)\nReference without patches 2:    0.0103s (12874.01275)\nReference without patches 3:    0.0092s (12874.01275)\nReference without patches 4:    0.0100s (12874.01275)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57837215,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><dataset><faster-rcnn>","CreationDate":"2019-09-07T20:25:52.120","AcceptedAnswerId":null,"OwnerUserId":11847474.0,"Title":"Can I get Soil Images Dataset?","Body":"<p>As I am working on a soil related project, I need a data-set of Soil Images. What is the best way or best algorithm to be used for classification? How can I create my own data-set and perform classification on images which are not in the original data-set?<\/p>\n\n<p>What I have done right now is found images on internet and renamed them to make them supervised data.<\/p>\n\n<p>As I used images downloaded from the web, the model's accuracy fluctuates. It moves up and down from 30% to 90% to 50%.<\/p>\n\n<p>I want a proper data-set or pre-processing techniques and algorithm suggestion.<\/p>\n","answers":[{"AnswerId":"57838067","CreationDate":"2019-09-07T23:14:28.327","ParentId":null,"OwnerUserId":"9557970","Title":null,"Body":"<p>As stated in your problem you want to work on image classification problem specifically soil.<br>\nThere are few basic steps you need to follow to implement such tasks successfully.<br>\nSteps to be followed :-  <\/p>\n\n<p><strong>1. Collection of data :<\/strong><br>\nThere are various sources for acquiring dataset related to your particular problem statement.Some of them are <strong>Kaggle<\/strong>, <strong>Quandle<\/strong>, <strong>Google dataset search<\/strong>.<br>\nIf you don't find the correct data you can scrape the images online.   <\/p>\n\n<p><strong>2. Labelling the data :<\/strong><br>\nFor labelling you can create directory structure such that one directory contains a single type of soil type.  <\/p>\n\n<p><strong>3. Oversampling :<\/strong><br>\nIf you have less amount of data or suffering from class imbalance problem you can use image oversampling technique such rotating etc.<\/p>\n\n<p><strong>4. Train test split :<\/strong><br>\nSplit your dataset into 3 parts. Train, test and cross- validation. 70-15-15 is good ratio. For eg if you have 100 images of black soil and 100 images of brown soil<br>\nFor Black:<br>\ntrain : 70 , test : 15 , CV :15<br>\nsimilar for Brown<\/p>\n\n<p><strong>5. Algorithm :<\/strong><br>\nThere are pretty powerful image classification algorithm such as <strong>inception model, VGG model, Resnet model, Nasnet model<\/strong>. By using technique called as transfer learning you can train these model as per the requirements of your project. These model have there own preprocessing class so you don't have to worry about it. Keras is the easiest library which contains all the mentioned models along with other model which you can use for transfer learning. Alternatively you can create your own custom model consisting of convolutional layers and fully connected layers for your. However in almost every instance the transfer learning will provide better results. <\/p>\n\n<p><strong>6. Model tuning :<\/strong><br>\nUse cross-validation set for tuning your model by changing hyper-parameters like learning rate etc. Using separate cross validation set will help with fluctuating accuracy. <\/p>\n\n<p><strong>7. Testing :<\/strong><br>\nAfter the tuning is complete then test your model for accuracy with test set.<\/p>\n\n<p>You can follow this tutorial : <a href=\"https:\/\/towardsdatascience.com\/keras-transfer-learning-for-beginners-6c9b8b7143e\" rel=\"nofollow noreferrer\">Transfer learning keras<\/a><\/p>\n"}]}
+{"QuestionId":57837306,"AnswerCount":0,"Tags":"<tensorflow><object-detection-api><jupyter-lab>","CreationDate":"2019-09-07T20:39:56.883","AcceptedAnswerId":null,"OwnerUserId":10334964.0,"Title":"Tensorflow (object detection api) keeps blocking RAM after inference","Body":"<p>After I make an inference on a loaded image, tensorflow keeps blocking the RAM (~ 2GB). It doesn't free it up even after the execution is finished.<\/p>\n\n<p>I did transfer learning on the coco dataset with my own images and use jupyter lab for the final inference of the images. I oriented myself at an <a href=\"https:\/\/github.com\/RomRoc\/objdet_train_tensorflow_colab\/pull\/7\" rel=\"nofollow noreferrer\">adapted version<\/a> of <a href=\"https:\/\/hackernoon.com\/object-detection-in-google-colab-with-custom-dataset-5a7bb2b0e97e\" rel=\"nofollow noreferrer\">RomRoc's Dog Detector<\/a>. The model itself (<em>frozen_inference_graph<\/em>) has a size of around 50MB.<\/p>\n\n<p>I am using <em>JupyterLab version 0.35.5<\/em> on <em>Linux Mint 19.1 Cinnamon<\/em>. My RAM is only 4 GB which often leads to crashes for multiple inferences. SWAP storage is activated and used.<\/p>\n\n<p>I checked the size of all variables using the variable inspector, but they are just a fraction of the 2 GB used.<\/p>\n\n<p>I suspect that the session or graph are not closed correctly, but even the manual closing doesn't lead to freeing up memory.<\/p>\n\n<p>This is the code leading to a big increase in RAM usage:<\/p>\n\n<pre><code>def 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 ['num_detections', 'detection_boxes', 'detection_scores',\n                        'detection_classes', 'detection_masks']:\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(tensor_name)\n\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\n                # Reframe is required to translate mask from box coordinates to image coordinates and fit the image size.\n                real_num_detection = tf.cast(tensor_dict['num_detections'][0], tf.int32)\n                detection_boxes = tf.slice(detection_boxes, [0, 0], [real_num_detection, -1])\n                detection_masks = tf.slice(detection_masks, [0, 0, 0], [real_num_detection, -1, -1])\n                detection_masks_reframed = utils_ops.reframe_box_masks_to_image_masks(detection_masks, detection_boxes, \n                                                                                      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(detection_masks_reframed, 0)\n\n            image_tensor = tf.get_default_graph().get_tensor_by_name('image_tensor:0')\n\n            # Run inference\n            output_dict = sess.run(tensor_dict, 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['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]\n    return output_dict\n<\/code><\/pre>\n\n<p>I expect memory consumption going back to the normal usage level of ~ 400 MB, like before the inference. Thanks already for any advices.<\/p>\n\n<hr>\n\n<p>I could replicate the same problem for <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/object_detection_tutorial.ipynb\" rel=\"nofollow noreferrer\">the official tensorflow object detection tutorial<\/a>. Although this one is actually less memory hungry.<\/p>\n","answers":[]}
+{"QuestionId":57837712,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2019-09-07T21:51:05.510","AcceptedAnswerId":null,"OwnerUserId":10018602.0,"Title":"Another \"ValueError: Cannot feed value of shape (30, 5) for Tensor 'Placeholder:0', which has shape '(?, 30)'\"","Body":"<p>Be kind, I'm new to TensorFlow. I have found a project that trains a <strong>Policy Gradient agent<\/strong> to trade the stock market, trained on only the daily <strong>Close<\/strong> prices. And I'm interested in making it train on the <strong>Open, High, Low, and Volume<\/strong> features as well, so I'm attempting to add them in to the existing code. I've stripped away most of the project to leave only what's necessary for diagnosis.<\/p>\n\n<p>You can find my <a href=\"https:\/\/colab.research.google.com\/drive\/1U0UDHxOv_CjDCMa9ynZEyNk4I9Tkb-M5\" rel=\"nofollow noreferrer\">Colab notebook here<\/a>. And I have done my best to comment each section to make it easier to browse, and where I THINK the issues are, but I need someone to show me where and why.<\/p>\n\n<p>I'm currently getting the error:<\/p>\n\n<pre><code>ValueError: Cannot feed value of shape (30, 5) for Tensor 'Placeholder:0', which has shape '(?, 30)'\n<\/code><\/pre>\n\n<p>...which makes sense because the <code>self.X = tf.compat.v1.placeholder(tf.float32, (None, self.state_size))<\/code> is designed for the only feature (Close), but I'm trying to add in the other features into the state as well. The <code>state_size<\/code> is the <code>window_size<\/code> of <code>30<\/code> (to look back 30 rows during training). And I'm trying to change the state to include the added features. So I try to change the placeholder to <code>self.state_size,5<\/code>, but then I get the error:<\/p>\n\n<pre><code>ValueError: Cannot feed value of shape (35760, 5) for Tensor 'Placeholder:0', which has shape '(30, 5)'\n<\/code><\/pre>\n\n<p>...which I'm a little unclear about, but I don't think that's the issue. I know I'm trying to feed the tensor data in a shape that it's not expecting, but I don't know how to adapt this on my own. (I think) What I'm looking to do is add in those extra features into the <code>get_state<\/code> function so that each row is 1 window_size, and each column represents the features. Then the training should take place over the iterations.<\/p>\n\n<p>I've found similar questions\/answers at the below links to help someone who knows more about this than me out. Most of them talk about reshaping the data at the placeholder, which I thought I had tried, but now I've just outrun my knowledge. Thanks in advance.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/45966301\/tensorflow-cannot-feed-value-of-shape-100-784-for-tensor-placeholder0\/46876972\">Here<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/44216254\/tensorflow-valueerror-cannot-feed-value-of-shape-423-for-tensor-placeholde\">Here<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/40430186\/tensorflow-valueerror-cannot-feed-value-of-shape-64-64-3-for-tensor-uplace\">Here<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/45966301\/tensorflow-cannot-feed-value-of-shape-100-784-for-tensor-placeholder0\">Here<\/a><\/p>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>Boy I'm sure struggling to figure this out on my own, but I appreciate the guidance thus far and I think I'm getting close, I just have limited knowledge of what to change. Given the answer below, I understand that my window_size\/number of rows can change depending on what the window_size\/lookback value is, so that would be the None part of my the placeholder, and in this particular case, I WOULD know the number of features (in this case 5) ahead of time so setting that as a static number (5) would suffice. So I'm trying to not have individual placeholders for each new feature I want to add.<\/p>\n\n<p>So trying to use the new placeholder <code>self.X = tf.compat.v1.placeholder(tf.float32, (None,5))<\/code> now the error I'm getting is:<\/p>\n\n<pre><code>    InvalidArgumentError: Incompatible shapes: [3270,3] vs. [98100,3]\n     [[{{node sub}}]]\n\nDuring handling of the above exception, another exception occurred:\n\nInvalidArgumentError                      Traceback (most recent call last)\n&lt;ipython-input-4-78f4afe280e1&gt; in &lt;module&gt;()\n     22          skip = skip)\n     23 \n---&gt; 24 agent.train(iterations = 500, checkpoint = 10, initial_money = initial_money)\n\n&lt;ipython-input-3-1788840ff10e&gt; in train(self, iterations, checkpoint, initial_money)\n    135             cost, _ = self.sess.run([self.cost, self.optimizer], feed_dict={self.X:np.vstack(ep_history[:,0]),\n    136                                                     self.REWARDS:ep_history[:,2],\n--&gt; 137                                                     self.ACTIONS:ep_history[:,1]})\n    138 \n    139 \n<\/code><\/pre>\n\n<p>...so I get the printed output of the <code>ep_history<\/code> array to check its shape, which has a starting shape of :<\/p>\n\n<pre><code>array([[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.]]) # which is the current state (5x30)\n\n       0 10000 # which is the action and reward\/starting_money\n\n array([[ 0.000000e+00,  0.000000e+00,  0.000000e+00,  0.000000e+00,\n     0.000000e+00],\n   [ 0.000000e+00,  0.000000e+00,  0.000000e+00,  0.000000e+00,\n     0.000000e+00],\n                      ....\n   [ 0.000000e+00,  0.000000e+00,  0.000000e+00,  0.000000e+00,\n     0.000000e+00],\n   [ 6.000060e-01,  1.669999e+00,  4.899980e-01,  2.309998e+00,\n    -2.135710e+07]])] # which is the NEXT state\n<\/code><\/pre>\n\n<p>...which acts as the state (5 columns of features and 30 rows of window_size), action (0), starting_money (10,000), and next_state (again 5 x 30 with the next\/future row of the dataset).<\/p>\n\n<p>So what seems to be the issue now? Is it the array? I apologize for the lengthy code, but I want to be thorough for those helping me and to show that I'm actually trying to understand the logic behind the fix so I can apply it later. Any further input? Would it be something to do with <code>get_state<\/code> function? (I've added a couple of extra comments to the colab book as I work through this). Thanks a lot.<\/p>\n","answers":[{"AnswerId":"57901683","CreationDate":"2019-09-12T07:24:46.687","ParentId":null,"OwnerUserId":"2851685","Title":null,"Body":"<p>you have used \"self.X\" as input for 1st layer, for a model number of rows (data points) can vary but number of features should be same during training and predicting because that determines the number of neurons on that layer. <\/p>\n\n<p>But you can reuse the code for data with different number of features to create different model but it has to remain same for a model.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>self.X = tf.compat.v1.placeholder(tf.float32, (None,feature_numbers))\n<\/code><\/pre>\n\n<p>You must know feature size of your input when you start training and it can't be changed latter. If you want to track other things then you need to create other variables for that or you can pre-process your data to know feature numbers before creating tensorflow graph <\/p>\n"}]}
+{"QuestionId":57837766,"AnswerCount":0,"Tags":"<python><r><keras><r-markdown>","CreationDate":"2019-09-07T22:03:20.537","AcceptedAnswerId":null,"OwnerUserId":8581206.0,"Title":"Error with keras in Rstudio + reticulate + python","Body":"<p>I'm working with R studio v1.2.1335 (r markdown and reticulate package), Python 3.7.4 installed in environment Anaconda 4.7.10.<\/p>\n\n<p>Struggling with this code (<a href=\"https:\/\/www.guru99.com\/keras-tutorial.html\" rel=\"nofollow noreferrer\">https:\/\/www.guru99.com\/keras-tutorial.html<\/a>)<\/p>\n\n<pre><code>```{python script 1}    \nimport keras\n    from keras.models import Sequential\n    from keras.layers import Dense, Activation\n    import numpy as np\n    import matplotlib.pyplot as plt \n\n    x = data = np.linspace(1,2,200)\n    y = x*4 + np.random.randn(*x.shape) * 0.3\n\n\n    model = Sequential()\n    model.add(Dense(1, input_dim=1, activation='linear'))\n\n    model.compile(optimizer='sgd', loss='mse', metrics=['mse'])\n\n    weights = model.layers[0].get_weights()\n    w_init = weights[0][0][0]\n    b_init = weights[1][0]\n    print('Linear regression model is initialized with\n    weights w: %.2f, b: %.2f' % (w_init, b_init)) \n\n\n    model.fit(x,y, batch_size=1, epochs=30, shuffle=False)\n\n    weights = model.layers[0].get_weights()\n    w_final = weights[0][0][0]\n    b_final = weights[1][0]\n    print('Linear regression model is trained to have\n    weight w: %.2f, b: %.2f' % (w_final, b_final))\n\n    predict = model.predict(data)\n\n    plt.plot(data, predict, 'b', data , y, 'k.')\n    plt.show()\n```\n<\/code><\/pre>\n\n<p>error:<\/p>\n\n<blockquote>\n  <p>Evaluation error: option error has NULL value.<\/p>\n<\/blockquote>\n\n<p>the same error happens with other code that use keras and comand Dense.<\/p>\n\n<p>others info with py_config()<\/p>\n\n<pre><code>python:         C:\\Users\\Megaport\\Anaconda3\\envs\\r-reticulate\\python.exe\nlibpython:      C:\/Users\/Megaport\/Anaconda3\/envs\/r-reticulate\/python37.dll\npythonhome:     C:\\Users\\Megaport\\ANACON~1\\envs\\R-RETI~1\nversion:        3.7.4 (default, Aug  9 2019, 18:34:13) [MSC v.1915 64 bit (AMD64)]\nArchitecture:   64bit\nnumpy:          C:\\Users\\Megaport\\ANACON~1\\envs\\R-RETI~1\\lib\\site-packages\\numpy\nnumpy_version:  1.17.2\n\npython versions found: \n C:\\Users\\Megaport\\Anaconda3\\envs\\r-reticulate\\python.exe\n C:\\Users\\Megaport\\ANACON~1\\python.exe\n C:\\Users\\Megaport\\Anaconda3\\python.exe\n C:\\Users\\Megaport\\Anaconda3\\envs\\venv\\python.exe\n<\/code><\/pre>\n\n<p>-------------------- UPDATE -------------------\nI've run the script in Rstudio without r-markdown but only with python\nand it worked without errors. The problem must be in r-markdown or reticulate installation.<\/p>\n\n<p>Any help would be appreciated<\/p>\n","answers":[]}
+{"QuestionId":57838192,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-09-07T23:50:04.243","AcceptedAnswerId":null,"OwnerUserId":11907202.0,"Title":"TypeError: 'int' object is not callable in loss.backward()","Body":"<p>While trying to set up a pytorch model, I am getting the error that the loss object is not callable when trying to do Pytorch autograd. (Relevant code shown below)<\/p>\n\n<pre><code>optimizer = torch.optim.Adam(model.parameters(), lr=lr, \n  betas(0.0,0.9))\n\ndef train(epoch, shuffle, wisdom_model, optim, loss):\n    print('train')\n    accuracy = 0\n    batch_num = 0\n    wisdom_model.train()\n    for batch in data.train_dl:\n\n        optim.zero_grad()\n\n        result = model(batch[0])\n        loss = nn.CrossEntropyLoss()(result, batch[1].long())\n\n        loss.backward()\n\n        accuracy += accuracy(result, batch[1])\n        print(accuracy)\n        pdb.set_trace()\n        batch_num += 1\n\n    return accuracy \/ batch_num\n<\/code><\/pre>\n\n<pre><code>TypeError                                 Traceback (most recent call last)\n&lt;ipython-input-28-5b9c9fe3b320&gt; in &lt;module&gt;\n----&gt; 1 run(1, False)\n\n&lt;ipython-input-27-d0d67dbf6eb2&gt; in run(num_models, dropout)\n     71 \n     72     for epoch in range(10):\n---&gt; 73         train_accuracy = train(epoch, False, model, optimizer, loss)\n     74         accuracy.append(validate(epoch, model))\n     75 \n\n&lt;ipython-input-27-d0d67dbf6eb2&gt; in train(epoch, shuffle, model, optim, loss)\n     24         pdb.set_trace()\n     25 \n---&gt; 26         loss.backward()\n     27         optim.step()\n     28 \n\nTypeError: 'int' object is not callable\n<\/code><\/pre>\n","answers":[{"AnswerId":"57841571","CreationDate":"2019-09-08T11:35:32.287","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>The problem is in this line:<\/p>\n\n<pre><code>loss = nn.CrossEntropyLoss()(result, batch[1].long())\n<\/code><\/pre>\n\n<p>Check out the <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html?highlight=nn%20crossentropyloss#torch.nn.CrossEntropyLoss\" rel=\"nofollow noreferrer\">nn.CrossEntropyLoss<\/a>.<\/p>\n\n<p>Should not look like this:<\/p>\n\n<pre><code>nn.CrossEntropyLoss()()\n<\/code><\/pre>\n\n<p>Should look like this:<\/p>\n\n<pre><code>nn.CrossEntropyLoss()\n<\/code><\/pre>\n"}]}
+{"QuestionId":57838359,"AnswerCount":0,"Tags":"<tensorflow><nvidia-jetson><mobilenet>","CreationDate":"2019-09-08T00:31:21.680","AcceptedAnswerId":null,"OwnerUserId":231238.0,"Title":"Running MobileNet on Jetson TK1","Body":"<p>I am trying to run a simple object recognition on TK1 with TensorFlow and pretrained MobileNet v2 (tried v1 also with the same problem).<\/p>\n\n<p>So far only found TF-0.8 builds for TK1 (CUDA 6.5\/cuDNN 2.0). They are missing support for <code>FusedBatchNorm<\/code> operation which is present in the model.<\/p>\n\n<p>Is there a way to reimplement <code>FusedBatchNorm<\/code> using ops available in TF-0.8, and keep the weights (if it has any)?<\/p>\n\n<p>Also, my limited understanding tells me, that batch norm is only applicable during training time. So why is it present in the frozen model? Could I just bypass it\/replace with a multiplication by an identity matrix?<\/p>\n","answers":[]}
+{"QuestionId":57838412,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-08T00:48:15.920","AcceptedAnswerId":57838507.0,"OwnerUserId":10083113.0,"Title":"How do I create a layer from a function that does not accept Tensors\/NumPy arrays as arguments?","Body":"<p>I have two Python functions that take strings as inputs and return NumPy arrays. I am trying to use these functions to create Lambda layers that are then fed into another Keras model.<\/p>\n\n<p>I can vectorize the function, and then create a TensorFlow operation via tf.py_func, like so (full code is further down below):<\/p>\n\n<pre><code>def indices_tensor(tensor):\n  return tf.py_func(np.vectorize(indices),[tensor],tf.float32)\ndef segments_tensor(tensor):\n  return tf.py_func(np.vectorize(segments),[tensor],tf.float32)\n<\/code><\/pre>\n\n<pre><code>pretrained_path = 'uncased_L-12_H-768_A-12'\nconfig_path = os.path.join(pretrained_path, 'bert_config.json')\ncheckpoint_path = os.path.join(pretrained_path, 'bert_model.ckpt')\nvocab_path = os.path.join(pretrained_path, 'vocab.txt')\n<\/code><\/pre>\n\n<pre><code># TF_KERAS must be added to environment variables in order to use TPU\nos.environ['TF_KERAS'] = '1'\nimport codecs\nfrom keras_bert import load_trained_model_from_checkpoint\n<\/code><\/pre>\n\n<pre><code>token_dict = {}\nwith codecs.open(vocab_path, 'r', 'utf8') as reader:\n    for line in reader:\n        token = line.strip()\n        token_dict[token] = len(token_dict)\n<\/code><\/pre>\n\n<pre><code>model = load_trained_model_from_checkpoint(config_path, checkpoint_path)\nimport numpy as np\nfrom keras_bert import Tokenizer\ntokenizer = Tokenizer(token_dict)\ndef tokenize(text):\n  tokens = tokenizer.tokenize(text)\n  indices, segments = tokenizer.encode(first=text, max_len=512)\n  return indices,segments\n<\/code><\/pre>\n\n<pre><code>def indices(text):\n  return tokenize(text)[0]\ndef segments(text):\n  return tokenize(text)[1]\n#@title Get indices and segments of a tensor of strings\ndef indices_tensor(tensor):\n  return tf.py_func(np.vectorize(indices),[tensor],tf.float32)\ndef segments_tensor(tensor):\n  return tf.py_func(np.vectorize(segments),[tensor],tf.float32)\n<\/code><\/pre>\n\n<pre><code>input_layer = Input(shape=(1,),dtype=tf.string)\nindices_layer = Lambda(indices_tensor)(input_layer)\nsegments_layer = Lambda(segments_tensor)(input_layer)\n``\n**Logging**\nprint(type(indices_layer))\nprint(type(segments_layer))\n<\/code><\/pre>\n\n<p>I want my Lambda function calls to produce Lambda layers, and <a href=\"https:\/\/keras.io\/layers\/core\/\" rel=\"nofollow noreferrer\">Keras's documentation<\/a> appears to suggest they should. Instead, per my indicated logging statements, they are creating Tensors.<\/p>\n\n<p>What should I do to create working layers?<\/p>\n","answers":[{"AnswerId":"57838507","CreationDate":"2019-09-08T01:15:37.050","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>There is nothing wrong there. Just your variable names that show you think the outputs of the layers are layers, but in fact they are tensors.<\/p>\n\n<p>This is an input tensor, not an input layer:<\/p>\n\n<pre><code>input_tensor = Input(shape=(1,),dtype=tf.string)\n<\/code><\/pre>\n\n<p>This is a Lambda layer:<\/p>\n\n<pre><code>layer = Lambda(segments_tensor)    #not called on the input tensor\n<\/code><\/pre>\n\n<p>This is a tensor that is the result of feeding the input tensor to the lambda layer:<\/p>\n\n<pre><code>tensor = Lambda(segments_tensor)(input_tensor)\n<\/code><\/pre>\n\n<p>This logic is true for every layer. <\/p>\n\n<pre><code>output_tensor = Layer(parameters)(input_tensor)\n\nlayer_instance = Layer(parameters)    \noutput1 = layer_instance(input_tensor1)\noutput2 = layer_instance(input_tensor2)\noutput3 = layer_instance(input_tensor3)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57838465,"AnswerCount":1,"Tags":"<python><android><tensorflow><machine-learning><flutter>","CreationDate":"2019-09-08T00:59:42.700","AcceptedAnswerId":57839378.0,"OwnerUserId":1889006.0,"Title":"Is it feasible to perform feature extraction in a Flutter app?","Body":"<p>I am attempting to implement an audio classifier in my mobile app. When training the data, I used melspectrogram extracted from the raw audio. I am using Tensorflow Lite to integrate the model into the app.  <\/p>\n\n<p>The problem is that I need to perform the same feature extraction on the input audio from the mic before passing into the tflite model. Python's Librosa library implements all of the functions that I need. My initial idea was run Python in flutter (there is the starflut Flutter package but I couldn't get it to work). <\/p>\n\n<p>Am I going about this in the wrong way? If so, what should I be doing? I could potentially rewrite the Librosa functions in dart lang, but I don't particularly want to do that.<\/p>\n","answers":[{"AnswerId":"57839378","CreationDate":"2019-09-08T05:23:17.953","ParentId":null,"OwnerUserId":"306739","Title":null,"Body":"<p>If you are targeting mobile, check the integration with \u201cnative\u201d code. E.g. look for a java\/kotlin library that can do the same on android. And a swift\/objC one for iOS.\nThen, you could wrap that functionality  in a platform-specific module.<\/p>\n"}]}
+{"QuestionId":57839097,"AnswerCount":2,"Tags":"<tensorflow><tensorflow-federated>","CreationDate":"2019-09-08T04:03:14.263","AcceptedAnswerId":null,"OwnerUserId":11283804.0,"Title":"Are there any way to do federated learning with real multiple machines using tensorflow-federated API?","Body":"<p>I am studying about tensorflow-federated API to make federated learning with real multiple machines.\nBut I found the answer on this site that not support to make real multiple federated learning using multiple learning.<\/p>\n\n<p>Are there no way to make federated learning with real multiple machines?<\/p>\n\n<p>Even I make a network structure for federated learning with 2 clients PC and 1 server PC, Is it impossible to consist of that system using tensorflow federated API?<\/p>\n\n<p>Or even if I apply the code, can't I make the system I want?<\/p>\n\n<p>If you can modify the code to configure it, can you give me a tip?If not, when will there be an example to configure on a real computer?<\/p>\n","answers":[{"AnswerId":"58283359","CreationDate":"2019-10-08T09:13:23.977","ParentId":null,"OwnerUserId":"8505949","Title":null,"Body":"<p>In case you are still looking for something: If you're not bound to TensorFlow, you could have a look at <a href=\"https:\/\/github.com\/OpenMined\/PySyft\" rel=\"nofollow noreferrer\">PySyft<\/a>, which is using PyTorch. <a href=\"https:\/\/blog.openmined.org\/federated-learning-of-a-rnn-on-raspberry-pis\/\" rel=\"nofollow noreferrer\">Here<\/a> is a practical example of a FL system built with one server and two Raspberry Pis as clients.<\/p>\n"},{"AnswerId":"57939192","CreationDate":"2019-09-14T20:55:14.320","ParentId":null,"OwnerUserId":"7507932","Title":null,"Body":"<p>TFF is really about <em>expressing<\/em> the federated computations you wish to execute. In terms of physical deployments, TFF includes two distinct runtimes: one \"reference executor\" which simply interprets the syntactic artifact that TFF generates, serially, all in Python and without any fancy constructs or optimizations; another still under development, but demonstrated in the tutorials, which uses <code>asyncio<\/code> and hierarchies of executors to allow for flexible executor architectures. Both of these are really about simulation and FL research, and not about deploying to devices.<\/p>\n\n<p>In principle, this may address your question (in particular, see <code>tff.framework.RemoteExecutor<\/code>). But I assume that you are asking more about deployment to \"real\" FL systems, e.g. data coming from sources that you don't control. This is really out of scope for TFF. From the FAQ:<\/p>\n\n<blockquote>\n  <p>Although we designed TFF with deployment to real devices in mind, at this stage we do not currently provide any tools for this purpose. The current release is intended for experimentation uses, such as expressing novel federated algorithms, or trying out federated learning with your own datasets, using the included simulation runtime.<\/p>\n  \n  <p>We anticipate that over time the open source ecosystem around TFF will evolve to include runtimes targeting physical deployment platforms.<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":57839196,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-08T04:29:22.210","AcceptedAnswerId":null,"OwnerUserId":10249632.0,"Title":"Tensorflow: Input to reshape is a tensor with 1 values, but the requested shape requires a multiple of 5","Body":"<p>I'm solving an object detection &amp; recognition problem.\nThere are several objects with different classes on a picture to detect and recognize.<\/p>\n\n<p>Train and Test datasets are in jpeg format and saved in TFRecords.<\/p>\n\n<p>I use the following code to train:<\/p>\n\n<pre><code>MODEL_NAME=faster_rcnn_resnet50_coco_2018_01_28_v01\nPIPELINE_CONFIG_PATH=\/models\/${MODEL_NAME}\/pipeline.config\nMODEL_DIR=\/models\/${MODEL_NAME}\/\nNUM_TRAIN_STEPS=50000\nSAMPLE_1_OF_N_EVAL_EXAMPLES=1\n\npython3 \/opt\/develop\/virtualenv-tf\/lib\/python3.6\/site-packages\/tensorflow\/models\/research\/object_detection\/model_main.py \\\n    --pipeline_config_path=${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>I got the following error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/opt\/develop\/virtualenv-tf\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1334, in _do_call\n    return fn(*args)\n  File \"\/opt\/develop\/virtualenv-tf\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1319, in _run_fn\n    options, feed_dict, fetch_list, target_list, run_metadata)\n  File \"\/opt\/develop\/virtualenv-tf\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1407, in _call_tf_sessionrun\n    run_metadata)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: Input to reshape is a tensor with 1 values, but the requested shape requires a multiple of 5\n     [[{{node Reshape_13}} = Reshape[T=DT_INT64, Tshape=DT_INT32, _device=\"\/device:CPU:0\"](SparseToDense_2, Reshape_13\/shape)]]\n     [[{{node IteratorGetNext}} = IteratorGetNext[output_shapes=[[1], [1,600,600,3], [1,2], [1,3], [1,100], [1,100,4], [1,100,11], [1,100,11], [1,100], [1,100], [1,100], [1]], output_types=[DT_INT32, DT_FLOAT, DT_INT32, DT_INT32, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_INT32, DT_BOOL, DT_FLOAT, DT_INT32], _device=\"\/job:localhost\/replica:0\/task:0\/device:CPU:0\"](IteratorV2)]]\n     [[{{node Loss\/ToAbsoluteCoordinates\/Assert\/AssertGuard\/Assert\/_3034}} = _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_442_Loss\/ToAbsoluteCoordinates\/Assert\/AssertGuard\/Assert\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]()]]\n<\/code><\/pre>\n\n<p>If there is something wrong with the dataset, what shall I check there?<\/p>\n","answers":[]}
+{"QuestionId":57839323,"AnswerCount":0,"Tags":"<python><keras><sentiment-analysis>","CreationDate":"2019-09-08T05:06:17.897","AcceptedAnswerId":null,"OwnerUserId":12036469.0,"Title":"Keras deep learning accuracy 100% problem","Body":"<p>I'm going through a film classification with Keras.<\/p>\n\n<p>The file is tagged [film review, sentiment(emotion)].<\/p>\n\n<pre><code> # MLP for the IMDB problem\nimport numpy\nfrom keras.datasets import imdb\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.layers import Flatten\nfrom keras.layers.embeddings import Embedding\nfrom keras.preprocessing import sequence\n# load the dataset but only keep the top n words, zero the rest\ntop_words = 5000\n# (X_train, y_train), (X_test, y_test) = imdb.load_data(num_words=top_words)\nX_train = train_result\ny_train = train_label\nX_test = test_result\ny_test = test_label\n\nmax_words = 500\nX_train = sequence.pad_sequences(X_train, maxlen=max_words)\nX_test = sequence.pad_sequences(X_test, maxlen=max_words)\n# create the model\nmodel = Sequential()\nmodel.add(Embedding(top_words, 32, input_length=max_words))\nmodel.add(Flatten())\nmodel.add(Dense(250, activation='relu'))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\nmodel.summary()\n# Fit the model\nhist = model.fit(X_train, y_train, validation_data=(X_test, y_test), epochs=10, batch_size=128, verbose=1)\n# Final evaluation of the model\nscores = model.evaluate(X_test, y_test, verbose=1)\nprint(\"Accuracy: %.2f%%\" % (scores[1]*100))\n\n\n%matplotlib inline\nimport matplotlib.pyplot as plt\n\nfig, loss_ax = plt.subplots()\n\nacc_ax = loss_ax.twinx()\n\nloss_ax.plot(hist.history['loss'], 'y', label='train loss')\nloss_ax.plot(hist.history['val_loss'], 'r', label='val loss')\n\nacc_ax.plot(hist.history['acc'], 'b', label='train acc')\nacc_ax.plot(hist.history['val_acc'], 'g', label='val acc')\n\nloss_ax.set_xlabel('epoch')\nloss_ax.set_ylabel('loss')\nacc_ax.set_ylabel('accuray')\n\nloss_ax.legend(loc='upper left')\nacc_ax.legend(loc='lower left')\n\nplt.show()\n<\/code><\/pre>\n\n<blockquote>\n  <p>OUTPUT = always Accuracy 100%<\/p>\n<\/blockquote>\n\n<p>As a basic deep learning model, \nIf you enable the imdb annotation, it will output normally.<\/p>\n\n<p>I don't think the Text preprocessing process is wrong.<\/p>\n\n<p>I downloaded the imdb.csv file directly, processed it, printed it and it shows the same graph<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ieG5c.png\" rel=\"nofollow noreferrer\">IMDB_Graph<\/a>, <a href=\"https:\/\/i.stack.imgur.com\/yAZyZ.png\" rel=\"nofollow noreferrer\">MYData_Graph<\/a> (very strange MYData graph..)<\/p>\n\n<p>For reference, <a href=\"https:\/\/drive.google.com\/open?id=1BevfrSgFAayEg6AyXvFVgbRhn-uYina7\" rel=\"nofollow noreferrer\">my Csv file<\/a><\/p>\n\n<p>I've classified horror movies as scary or brutal films.<\/p>\n\n<p>[cruel, astonish] I've sorted it into tags.<\/p>\n\n<p>Can we sort it out frankly? I've been doubting it.<\/p>\n\n<p>I don't know if the data itself is wrong or the code is wrong.<\/p>\n\n<hr>\n\n<p>Do you have any idea what's wrong?<\/p>\n\n<p>Why is 100% accuracy always displayed?<\/p>\n","answers":[]}
+{"QuestionId":57840189,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-08T08:08:55.320","AcceptedAnswerId":null,"OwnerUserId":12036781.0,"Title":"How to create a custom loss function where intermediate training outputs(tensor y_pred) of an RNN, are fed to another predefined RNN?","Body":"<p>I wish to create a custom loss function which doesn't directly use the intermediate outputs of the RNN(y_pred), but instead feeds in the y_pred as inputs to another RNN(say RNN2, which has already been defined and trained), and use these predicted values as the parameters of loss function.<\/p>\n\n<p>I try to call my custom loss function from the model.compile function, and that produces an error. Is it because i can't feed in an object of tensor data type to RNN2? Am i wrong to assume that y_pred has the intermediate outputs of training? \nAlso a simple print command of the y_pred using sess throws an error too!\ni.e <\/p>\n\n<pre><code>sess=tf.Session()\nprint(sess.eval(y_pred))\n<\/code><\/pre>\n\n<p>so is the problem fundamental to y_pred?<\/p>\n\n<p>Anyway, this is code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\ndef custom_loss(y_true, y_pred):\n        predicted=rnn2.predict(y_pred)\n        return K.mean(K.abs( predicted-y_true), axis=-1)\n\ninput_tensor = Input(shape=(1,1))\nhidden = LSTM(100, activation='softmax',return_sequences=False)(input_tensor)\nout = Dense(1, activation='softmax')(hidden)\nmodel = Model(input_tensor, out)\nmodel.compile(loss=custom_loss, optimizer='adam')\n<\/code><\/pre>\n\n<p>error<\/p>\n\n<pre><code>You must feed a value for placeholder tensor 'input_19' with dtype float and shape [?,1,1]\n     [[{{node input_19}}]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"57848620","CreationDate":"2019-09-09T06:07:36.727","ParentId":null,"OwnerUserId":"9477338","Title":null,"Body":"<p>This probably something to do when you call <code>model.compile<\/code>. Did you try pass it(RNN2) as a new layer like.<\/p>\n\n<pre><code>RNN2.trainable = False #  [1]\nmodel = Sequential()\nmodel.add(RNN1)\nmodel.add(RNN2)\n\ndef custom_loss(y_true, y_pred):\n        # predicted=rnn2.predict(y_pred)\n        return K.mean(K.abs( y_pred-y_true), axis=-1)\n\n<\/code><\/pre>\n\n<h1>Edit<\/h1>\n\n<blockquote>\n  <p>Can you elaborate on how i can use a Model(RNN2) as a Layer in RNN1?<\/p>\n<\/blockquote>\n\n<p>If it's me I'll to something like this.<\/p>\n\n<pre><code>from keras import models, layers\n\n\ninp = layers.Input((None, 1))\nx = layers.LSTM(512, return_sequences=True)(inp)\nx = layers.LSTM(256)(x)\nx = layers.Dense(32, name='rnn2_output')(x)\nrnn2 = models.Model(inp, x)\nrnn2.trainable = False #  [2]\n\ninp2 = layers.Input((None, 32))\nx = layers.LSTM(256, return_sequences=True)(inp2)\nx = layers.Dense(1)(x)\nx = rnn2(x)\nrnn1 = models.Model(inp2, x)\nrnn1.summary()\n\n<\/code><\/pre>\n\n<p>Beware, both the recently added code [2] and the old one [1](which recently edited) have <code>trainable = False<\/code>, it means this model <strong>will not<\/strong> being trained at all. Assume that you put <code>RNN2.predict<\/code> in the loss function. If you want to train it also then remove these lines.<\/p>\n"}]}
+{"QuestionId":57840199,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><lstm>","CreationDate":"2019-09-08T08:10:54.627","AcceptedAnswerId":null,"OwnerUserId":10704050.0,"Title":"How to add dropout and attention in LSTM in kers in python","Body":"<p>I have about 1000 nodes dataset where each node has 4 time-series. Each time series is exactly 6 length long.The label is 0 or 1 (i.e. binary classification).<\/p>\n\n<p>More precisely my dataset looks as follows.<\/p>\n\n<pre><code>node, time-series1, time_series2, time_series_3, time_series4, Label\nn1, [1.2, 2.5, 3.7, 4.2, 5.6, 8.8], [6.2, 5.5, 4.7, 3.2, 2.6, 1.8], \u2026, 1\nn2, [5.2, 4.5, 3.7, 2.2, 1.6, 0.8], [8.2, 7.5, 6.7, 5.2, 4.6, 1.8], \u2026, 0\nand so on.\n<\/code><\/pre>\n\n<p>I <code>normalise<\/code> my timeseries before I feed it into my LSTM model for classification.<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(10, input_shape=(6,4)))\nmodel.add(Dense(32))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n\nprint(data.shape) # (1000, 6, 4)\nmodel.fit(data, target)\n<\/code><\/pre>\n\n<p>I am new to keras and that is why started with the simplest LSTM model. However, now I would like to make it into a level that I can use it at an industry level.<\/p>\n\n<p>I read that it is good to add <code>dropout<\/code> and <code>attention<\/code> layers to the LSTM models. Please let me know if you think that adding such layers is applicable to my problem and if so how to do it? :)<\/p>\n\n<p>Note: I am not limited to droupout and attention layers and happy to receive other suggestions that I can use to improve my model.<\/p>\n\n<p>I am happy to provide more details if needed.<\/p>\n","answers":[{"AnswerId":"57840352","CreationDate":"2019-09-08T08:36:55.260","ParentId":null,"OwnerUserId":"8621405","Title":null,"Body":"<p>if you want to add dropout in lstm cell, you can try this <\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(10, input_shape=(6,4), dropout=0.5))\nmodel.add(Dense(32))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n\nprint(data.shape) # (1000, 6, 4)\nmodel.fit(data, target)\n<\/code><\/pre>\n\n<p>or using dropout between lstm cell, may consider below<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(10, input_shape=(6,4)))\nmodel.add(Dropout(0.5))\nmodel.add(LSTM(10, input_shape=(6,4)))\nmodel.add(Dense(32))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n\nprint(data.shape) # (1000, 6, 4)\nmodel.fit(data, target)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57840750,"AnswerCount":2,"Tags":"<keras>","CreationDate":"2019-09-08T09:34:56.657","AcceptedAnswerId":57844639.0,"OwnerUserId":8176285.0,"Title":"How to pass weights to mean squared error in keras","Body":"<p>I am trying to approach a regression problem, which is multi label with 8 labels for which i am using mean squared error loss, but the data set is imbalanced and i want to pass weights to the loss function.Currently i am compiling the model this way.<\/p>\n\n<pre><code>model.compile(loss='mse', optimizer=Adam(lr=0.0001), metrics=['mse', 'acc'])\n<\/code><\/pre>\n\n<p>Could someone please suggest if it is possible to add weights to mean squared error,if so, how could i do it?<\/p>\n\n<p>Thanks in advance<\/p>\n\n<p>The labels look like so<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/kiuEh.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/kiuEh.png\" alt=\"enter image description here\"><\/a><\/p>\n\n#\n\n<pre><code>model = Sequential()    \nmodel.add(effnet)\nmodel.add(GlobalAveragePooling2D())\nmodel.add(Dropout(0.5))\nmodel.add(Dense(8,name = 'nelu', activation=elu))\nmodel.compile(loss=custom_mse(class_weights), \noptimizer=Adam(lr=0.0001), metrics=['mse', 'acc'])\n<\/code><\/pre>\n","answers":[{"AnswerId":"57843430","CreationDate":"2019-09-08T15:49:21.970","ParentId":null,"OwnerUserId":"3472735","Title":null,"Body":"<p>First create a dictionary of how much you want to weight each class, for example:<\/p>\n\n<pre><code>class_weights = {0: 1,\n                 1: 1,\n                 2: 1, \n                 3: 9, \n                 4: 1...} # Do this for all eight classes\n<\/code><\/pre>\n\n<p>Then pass them into model.fit:<\/p>\n\n<pre><code>model.fit(X, y, class_weight=class_weights)\n<\/code><\/pre>\n"},{"AnswerId":"57844639","CreationDate":"2019-09-08T18:21:49.910","ParentId":null,"OwnerUserId":"11952444","Title":null,"Body":"<pre><code>import keras\nfrom keras.models import Sequential\nfrom keras.layers import Conv2D, Flatten, Dense, Conv1D, LSTM, TimeDistributed\nimport keras.backend as K\n\n\n# custom loss function\ndef custom_mse(class_weights):\n    def loss_fixed(y_true, y_pred):\n        \"\"\"\n        :param y_true: A tensor of the same shape as `y_pred`\n        :param y_pred:  A tensor resulting from a sigmoid\n        :return: Output tensor.\n        \"\"\"\n        # print('y_pred:', K.int_shape(y_pred))\n        # print('y_true:', K.int_shape(y_true))\n        y_pred = K.reshape(y_pred, (8, 1))\n        y_pred = K.dot(class_weights, y_pred)\n        # calculating mean squared error\n        mse = K.mean(K.square(y_pred - y_true), axis=-1)\n        # print('mse:', K.int_shape(mse))\n        return mse\n\nmodel = Sequential()\nmodel.add(Conv1D(8, (1), input_shape=(28, 28)))\nmodel.add(Flatten())\nmodel.add(Dense(8))\n\n# custom class weights\nclass_weights = K.variable([[0.25, 1., 2., 3., 2., 0.6, 0.5, 0.15]])\n# print('class_weights:', K.int_shape(class_weights))\n\nmodel.compile(optimizer='adam', loss=custom_mse(class_weights), metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>Here is a small implementation of a custom loss function based on your problem statement<\/p>\n\n<ul>\n<li><p>You find more information about keras loss function from <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/losses.py\" rel=\"nofollow noreferrer\">losses.py<\/a> and also check out its official documentation from <a href=\"https:\/\/keras.io\/losses\/\" rel=\"nofollow noreferrer\">here<\/a><\/p><\/li>\n<li><p>Keras does not handle low-level operations such as tensor products, convolutions and so on itself. Instead, it relies on a specialized, well-optimized tensor manipulation library to do so, serving as the \"backend engine\" of Keras. More information about <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/backend\/tensorflow_backend.py\" rel=\"nofollow noreferrer\">keras backend<\/a> can be found here and also check out its official documentation from <a href=\"https:\/\/keras.io\/backend\/\" rel=\"nofollow noreferrer\">here<\/a><\/p><\/li>\n<li><p>Use K.int_shape(tensor_name) to find the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/backend\/shape\" rel=\"nofollow noreferrer\">dimensions of a tensor<\/a>.<\/p><\/li>\n<\/ul>\n"}]}
+{"QuestionId":57841364,"AnswerCount":1,"Tags":"<python><tensorflow><anaconda><spyder>","CreationDate":"2019-09-08T11:07:00.773","AcceptedAnswerId":null,"OwnerUserId":8581206.0,"Title":"Impossible to import tensorflow with spyder anaconda","Body":"<p>I am trying to import tensorflow without success. The installation of each packages is done with Anaconda Navigator or in the Anaconda prompt command with <code>conda install *package*<\/code> in the right env that I've created previously.<\/p>\n\n<p>For the test I found this script usefull: <a href=\"https:\/\/www.southampton.ac.uk\/~fangohr\/blog\/installation-of-python-spyder-numpy-sympy-scipy-pytest-matplotlib-via-anaconda.html#running-the-tests-with-spyder\" rel=\"nofollow noreferrer\">https:\/\/www.southampton.ac.uk\/~fangohr\/blog\/installation-of-python-spyder-numpy-sympy-scipy-pytest-matplotlib-via-anaconda.html#running-the-tests-with-spyder<\/a>\nand all others packages seems to work.<\/p>\n\n<p>Something is wrong in the installation of tensorflow.<\/p>\n\n<pre><code>import tensorflow as tf\n<\/code><\/pre>\n\n<p>Output in the console:<\/p>\n\n<pre><code>    import tensorflow as tf\nTraceback (most recent call last):\n\n  File \"&lt;ipython-input-17-64156d691fe5&gt;\", line 1, in &lt;module&gt;\n    import tensorflow as tf\n\n  File \"C:\\Users\\Megaport\\Anaconda3\\envs\\venv\\lib\\site-packages\\tensorflow\\__init__.py\", line 34, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n\n  File \"C:\\Users\\Megaport\\Anaconda3\\envs\\venv\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n\n  File \"C:\\Users\\Megaport\\Anaconda3\\envs\\venv\\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:\\Users\\Megaport\\Anaconda3\\envs\\venv\\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\\Megaport\\Anaconda3\\envs\\venv\\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\\Megaport\\Anaconda3\\envs\\venv\\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\\Megaport\\Anaconda3\\envs\\venv\\lib\\imp.py\", line 242, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\Users\\Megaport\\Anaconda3\\envs\\venv\\lib\\imp.py\", line 342, in load_dynamic\n    return _load(spec)\nImportError: DLL load failed: Impossibile trovare il modulo specificato.\n\n\nFailed to load the native TensorFlow runtime.\n\nSee https:\/\/www.tensorflow.org\/install\/errors\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>Thanks in advance.<\/p>\n\n<p>this is the list of what is installed in the env. OS x64\nAll the solutions in stackoverflow (avx support of the cpu, microsoft c++ 2015, python path) have already been tried <\/p>\n\n<pre><code># packages in environment at C:\\Users\\Megaport\\Anaconda3\\envs\\venv:\n#\n# Name                    Version                   Build  Channel\n_py-xgboost-mutex         2.0                       cpu_0\n_tflow_select             2.3.0                       mkl\nabsl-py                   0.8.0                    pypi_0    pypi\nastor                     0.8.0                    pypi_0    pypi\natomicwrites              1.3.0                    py37_1\nattrs                     19.1.0                   py37_1\nbackcall                  0.1.0                    py37_0\nblas                      1.0                         mkl\nca-certificates           2019.5.15                     1\ncertifi                   2019.6.16                py37_1\ncloudpickle               1.2.1                      py_0\ncolorama                  0.4.1                    py37_0\ncycler                    0.10.0                   py37_0\ndecorator                 4.4.0                    py37_1\nfastcache                 1.1.0            py37he774522_0\nfreetype                  2.9.1                ha9979f8_1\ngast                      0.2.2                    pypi_0    pypi\ngoogle-pasta              0.1.7                    pypi_0    pypi\ngrpcio                    1.23.0                   pypi_0    pypi\nh5py                      2.10.0                   pypi_0    pypi\nhdf5                      1.10.2               hac2f561_1\nicc_rt                    2019.0.0             h0cc432a_1\nicu                       58.2                 ha66f8fd_1\nimportlib_metadata        0.19                     py37_0\nintel-openmp              2019.4                      245\nipykernel                 5.1.2            py37h39e3cac_0\nipython                   7.8.0            py37h39e3cac_0\nipython_genutils          0.2.0                    py37_0\njedi                      0.15.1                   py37_0\njoblib                    0.13.2                   py37_0\njpeg                      9b                   hb83a4c4_2\njupyter_client            5.3.1                      py_0\njupyter_core              4.5.0                      py_0\nkeras-applications        1.0.8                      py_0\nkeras-base                2.2.4                    py37_0\nkeras-preprocessing       1.1.0                      py_1\nkiwisolver                1.1.0            py37ha925a31_0\nlibmklml                  2019.0.5                      0\nlibpng                    1.6.37               h2a8f88b_0\nlibprotobuf               3.8.0                h7bd577a_0\nlibsodium                 1.0.16               h9d3ae62_0\nlibxgboost                0.90                          0\nm2w64-gcc-libgfortran     5.3.0                         6\nm2w64-gcc-libs            5.3.0                         7\nm2w64-gcc-libs-core       5.3.0                         7\nm2w64-gmp                 6.1.0                         2\nm2w64-libwinpthread-git   5.0.0.4634.697f757               2\nmarkdown                  3.1.1                    py37_0\nmatplotlib                3.1.1            py37hc8f65d3_0\nmkl                       2019.4                      245\nmkl-service               2.0.2            py37he774522_0\nmkl_fft                   1.0.14           py37h14836fe_0\nmkl_random                1.0.2            py37h343c172_0\nmore-itertools            7.2.0                    py37_0\nmpmath                    1.1.0                    py37_0\nmsys2-conda-epoch         20160418                      1\nnumpy                     1.17.2                   pypi_0    pypi\nnumpy-base                1.16.4           py37hc3f5095_0\nopenssl                   1.1.1c               he774522_1\nopt-einsum                3.0.1                    pypi_0    pypi\npackaging                 19.1                     py37_0\npandas                    0.25.1           py37ha925a31_0\nparso                     0.5.1                      py_0\npickleshare               0.7.5                    py37_0\npip                       19.2.2                   py37_0\npluggy                    0.12.0                     py_0\nprompt_toolkit            2.0.9                    py37_0\nprotobuf                  3.9.1                    pypi_0    pypi\npy                        1.8.0                    py37_0\npy-xgboost                0.90                     py37_0\npy-xgboost-cpu            0.90                     py37_0\npygments                  2.4.2                      py_0\npyparsing                 2.4.2                      py_0\npyqt                      5.9.2            py37h6538335_2\npytest                    5.0.1                    py37_0\npython                    3.7.4                h5263a28_0\npython-dateutil           2.8.0                    py37_0\npytz                      2019.2                     py_0\npyyaml                    5.1.2            py37he774522_0\npyzmq                     18.1.0           py37ha925a31_0\nqt                        5.9.7            vc14h73c81de_0\nscikit-learn              0.21.2           py37h6288b17_0\nscipy                     1.3.1            py37h29ff71c_0\nsetuptools                41.2.0                   pypi_0    pypi\nsip                       4.19.8           py37h6538335_0\nsix                       1.12.0                   pypi_0    pypi\nspyder-kernels            0.5.1                    py37_0\nsqlite                    3.29.0               he774522_0\nsympy                     1.4                      py37_0\ntb-nightly                1.15.0a20190806          pypi_0    pypi\ntensorboard               1.14.0           py37he3c9ec2_0\ntensorflow                1.14.0          mkl_py37h7908ca0_0\ntensorflow-base           1.14.0          mkl_py37ha978198_0\ntensorflow-estimator      1.14.0                     py_0\ntermcolor                 1.1.0                    pypi_0    pypi\ntornado                   6.0.3            py37he774522_0\ntraitlets                 4.3.2                    py37_0\nvc                        14.1                 h0510ff6_4\nvs2015_runtime            14.16.27012          hf0eaf9b_0\nwcwidth                   0.1.7                    py37_0\nwerkzeug                  0.15.6                   pypi_0    pypi\nwheel                     0.33.6                   pypi_0    pypi\nwincertstore              0.2                      py37_0\nwrapt                     1.11.2           py37he774522_0\nyaml                      0.1.7                hc54c509_2\nzeromq                    4.3.1                h33f27b4_3\nzipp                      0.5.2                      py_0\nzlib                      1.2.11               h62dcd97_3\n<\/code><\/pre>\n","answers":[{"AnswerId":"57842490","CreationDate":"2019-09-08T13:42:48.610","ParentId":null,"OwnerUserId":"8581206","Title":null,"Body":"<p>Installing spyder again inside the environment was the solution.<\/p>\n"}]}
+{"QuestionId":57841557,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets><tensorflow-estimator>","CreationDate":"2019-09-08T11:33:40.180","AcceptedAnswerId":57903222.0,"OwnerUserId":11318754.0,"Title":"Feed tf.estimator.Estimator.predict batch by batch","Body":"<p>I have a trained estimator model and I need to get the prediction vectors of a very large dataset that does not fit in the memory, process these prediction vectors and save them. So far my code looks something like this:<\/p>\n\n<pre><code>def hist(predictions):\n    ...\n    return histograms\n\ndef input_fn(feat, batch_size=100):\n    dataset = tf.data.Dataset.from_tensor_slices((feat))\n    dataset = dataset.batch(batch_size)\n    dataset = dataset.map(lambda x:...)\n    return dataset\n\nsuper_batch = 100\nsplits = data.shape[0]\/\/super_batch\n\nfor s in range(splits):\n    pred = list(classifier.predict(lambda: input_fn(data[s*super_batch:(s+1)*super_batch])))\n    pred_cls = [p[\"classes\"] for p in pred]\n    hist_vec = hist(pred_cls)\n    save hist_vec\n<\/code><\/pre>\n\n<p>I know this is not the correct way to do it, cause it leaves the GPU idle for a long time and because of loading the model every time I call the classifier.predict it takes a very very long time to run. Is there any way to use something like a feed function with estimators and speed this process up?<\/p>\n","answers":[{"AnswerId":"57903222","CreationDate":"2019-09-12T09:00:34.190","ParentId":null,"OwnerUserId":"10025506","Title":null,"Body":"<p><strong>I'm assuming that the issue is with <code>tf.data.Dataset.from_tensor_slices()<\/code><\/strong><\/p>\n\n<hr>\n\n<p>If you use <code>tf.data.Dataset.from_tensor_slices()<\/code> with numpy arrays in Tensorflow 1.0 with eager mode disabled, it embeds the values in the graph as one or more <code>tf.constant<\/code> operations. The bigger your dataset is, the bigger the graph becomes. This is very inefficient and you can face <code>ValueError: GraphDef cannot be larger than 2GB<\/code> errors.<\/p>\n\n<p>With <code>tf.estimator<\/code>, you have 2 solutions: <\/p>\n\n<ul>\n<li><p>Use <code>tf.data.Dataset.from_generator<\/code>. Just convert <code>feat<\/code> into a Python Generator and you are good to go. The performances are a bit worse than <code>from_tensor_slices<\/code> as the <code>tf.data<\/code> graph's speed is limited by the Python runtime.<\/p><\/li>\n<li><p>Use <code>tf.data.Dataset.from_tensor_slices()<\/code> with Tensorflow placeholders. This is more complicated but the most efficient one. Please see <a href=\"https:\/\/stackoverflow.com\/a\/51633489\/10025506\">my answer here<\/a> for more informations. The gist of it is that you need to create a specific hook to initialize the placeholders after the session is created inside the estimator.<\/p><\/li>\n<\/ul>\n"}]}
+{"QuestionId":57842119,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-08T12:51:11.637","AcceptedAnswerId":null,"OwnerUserId":10003205.0,"Title":"How to name an operation which is done via defining one or more functions on a operation in tensorflow?","Body":"<p>I want to see the value of S after training in tensorflow where <\/p>\n\n<pre><code>S = function1(function2(function3(tf.clip_by_value)))\n<\/code><\/pre>\n\n<p>I know I can give name to <code>tf.clip_by_value<\/code> by using <code>name=&lt;my_tensor_name&gt;<\/code>, and recover it by <code>tf.get_tensor_by_name(&lt;my_tensor_name&gt;)<\/code>. But if I want to recover S, how to name it, so that I can call it after training and see the value?<\/p>\n","answers":[]}
+{"QuestionId":57842325,"AnswerCount":1,"Tags":"<numpy><tensorflow><keras>","CreationDate":"2019-09-08T13:21:15.617","AcceptedAnswerId":null,"OwnerUserId":12037700.0,"Title":"why is this error occurring cannot reshape?","Body":"<p>I'm trying to reshape the data-set.<\/p>\n\n<p>I tried changing the values: <br><\/p>\n\n<pre><code>X_train = X_train.reshape(X_train.shape[0], 1, 100, 100).astype('float32')\n<\/code><\/pre>\n\n<p>but got this error :<\/p>\n\n<pre><code>ValueError: cannot reshape array of size 144052500 into shape (57621,1,100,100)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57843377","CreationDate":"2019-09-08T15:42:48.030","ParentId":null,"OwnerUserId":"3472735","Title":null,"Body":"<p>In future, please provide more details such as the shape of X_train or what kind of data you're working with.<\/p>\n\n<p>That said, I suspect you have 4 channels or some other dimension, which you are not accounting for when reshaping. This should work:<\/p>\n\n<pre><code>X_train = X_train.reshape(X_train.shape[0], 4, 100, 100).astype('float32')\n<\/code><\/pre>\n\n<p>instead of <\/p>\n\n<pre><code>X_train = X_train.reshape(X_train.shape[0], 1, 100, 100).astype('float32')\n<\/code><\/pre>\n"}]}
+{"QuestionId":57842400,"AnswerCount":1,"Tags":"<python><neural-network><pytorch>","CreationDate":"2019-09-08T13:29:10.483","AcceptedAnswerId":57842425.0,"OwnerUserId":9464073.0,"Title":"Understanding the use of classes when training a neural network with Pytorch","Body":"<p>I'm currently reading about using the cross-entropy loss to train a neural network in the <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/blitz\/cifar10_tutorial.html#define-a-loss-function-and-optimizer\" rel=\"nofollow noreferrer\">pytorch documentation<\/a>. The criterion that is to be used to calculate the loss is called as follows<\/p>\n\n<pre><code>criterion = nn.CrossEntropyLoss()\n<\/code><\/pre>\n\n<p>According to the <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#crossentropyloss\" rel=\"nofollow noreferrer\"><code>torch.nn<\/code> documentation<\/a>, <code>CrossEntropyLoss<\/code> is a class. From my understanding this would mean that <code>criterion<\/code> is an object of type <code>nn.CrossEntropyLoss<\/code>. <\/p>\n\n<p>When training the neural network, <code>criterion<\/code> is used to calculate the loss in the following way<\/p>\n\n<pre><code>loss = criterion(input, target)\n<\/code><\/pre>\n\n<p>This is somewhat confusing to me. <strong>If<\/strong> <code>criterion<\/code> <strong>is an object, then how can it be used as a function?<\/strong> What I would expect is something like<\/p>\n\n<pre><code>loss = criterion.calculate_loss(input, target)\n<\/code><\/pre>\n\n<p>where <code>calculate_loss()<\/code> would be a method defined in the <code>nn.CrossEntropyLoss<\/code> class. Furthermore, the documentation also uses the following line of code<\/p>\n\n<pre><code>running_loss += loss.item()\n<\/code><\/pre>\n\n<p>Where does this <code>item()<\/code> method come from? I can't find a source mentioning <code>item()<\/code> online.<\/p>\n","answers":[{"AnswerId":"57842425","CreationDate":"2019-09-08T13:33:43.460","ParentId":null,"OwnerUserId":"3014199","Title":null,"Body":"<blockquote>\n  <p>If criterion is an object, then how can it be used as a function?<\/p>\n<\/blockquote>\n\n<p>The criterion object has a <code>forward<\/code> method that is called in that case. <code>criterion(input, target)<\/code> is shorthand for <code>criterion.forward(input, target)<\/code><\/p>\n\n<blockquote>\n  <p>Where does this item() method come from?<\/p>\n<\/blockquote>\n\n<p>This method returns a one dimensional <code>Tensor<\/code>. The single value can be accessed as a number with <code>item()<\/code>.<\/p>\n"}]}
+{"QuestionId":57842401,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-08T13:29:19.200","AcceptedAnswerId":null,"OwnerUserId":4622046.0,"Title":"How to get the tensor shape in run time for bucket_by_sequence_length with dynamic padding?","Body":"<p>I am trying to train a text to speech model in TensorFlow. To generate batches, I'm using <strong>bucket_by_sequence_length<\/strong> function. <\/p>\n\n<p>But, after I get each batch, I need to know the shape of each tensor. The problem is that the tensors come from different sized arrays. <\/p>\n\n<p>I have tried <code>get_shape().as_list()<\/code> and 'tf.shape' but I don't get the shape in runtime. For the temporal dimension, I always get None. But, as I am planning to add a custom layer which needs to know the temporal dimension too, I am getting errors for dimension mismatch.<\/p>\n\n<pre><code>_, (texts, mels, mags, fnames) = tf.contrib.training.bucket_by_sequence_length(\n                                            input_length=text_length,\n                                            tensors=[text, mel, mag, fname],\n                                            batch_size=hp.B,\n                                            bucket_boundaries=[i for i in range(minlen + 1, maxlen - 1, 20)],\n                                            num_threads=1,\n                                            capacity=hp.B,\n                                            dynamic_pad=True)\n\n<\/code><\/pre>\n\n<p>Each time I get batch size, for the tensors I get the following shape with <code>get_shape().as_list()<\/code> <strong>(10, None, 156)<\/strong> Here, mel, mag are two tensors with unknown temporal dimension (depends on the audio file length).<\/p>\n\n<p>Here, 10 is the batch size, and 156 is the feature size. I also need the exact temporal dimension where I get None.<\/p>\n\n<p><code>tf.shape<\/code> returns <strong>Tensor(\"Text2Mel\/AudioEnc\/strided_slice:0\", shape=(), dtype=int32)<\/strong> I don't get anything useful from this too.<\/p>\n\n<p>How can I infer the exact shape at runtime?<\/p>\n\n<p>Actual Source which I have modified: <a href=\"https:\/\/github.com\/Kyubyong\/dc_tts\/blob\/master\/data_load.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Kyubyong\/dc_tts\/blob\/master\/data_load.py<\/a><\/p>\n","answers":[]}
+{"QuestionId":57842734,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-08T14:17:58.200","AcceptedAnswerId":null,"OwnerUserId":1273818.0,"Title":"Tensorflow 2.0 - incrrect data shape for SparseCategoricalCrossentropy","Body":"<p>When using <code>SparseCategoricalCrossentropy<\/code> loss I don't understand the error about the shape mismatch of the dataset.<br>\nThe output seems to be in the correct form as per <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/losses\/SparseCategoricalCrossentropy\" rel=\"nofollow noreferrer\">SparseCategoricalCrossentropy<\/a> :<br>\nInput : [7 1]<br>\nOutput: [7 3]<br>\nWhere batch size is 7.<\/p>\n\n<p>My environment currently runs:<br>\npython 3.7.4<br>\n<strong>tensorflow 2.0.0-rc0<\/strong><br>\nnumpy 1.17.2<br>\npandas 0.24.2  <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>---------------------------------------------------------------------------\nInvalidArgumentError                      Traceback (most recent call last)\n&lt;ipython-input-5-fe1cc3e1dca1&gt; in &lt;module&gt;\n     47 print(model.summary())\n     48 \n---&gt; 49 model.fit(train.batch(BATCH_SIZE), epochs=EPOCHS, verbose=2)\n     50 model.evaluate(train, steps=None, verbose=1)\n\n\/usr\/local\/lib\/python3.7\/site-packages\/tensorflow_core\/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, validation_freq, max_queue_size, workers, use_multiprocessing, **kwargs)\n    732         max_queue_size=max_queue_size,\n    733         workers=workers,\n--&gt; 734         use_multiprocessing=use_multiprocessing)\n    735 \n    736   def evaluate(self,\n\n\/usr\/local\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/keras\/engine\/training_v2.py in fit(self, model, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_freq, **kwargs)\n    322                 mode=ModeKeys.TRAIN,\n    323                 training_context=training_context,\n--&gt; 324                 total_epochs=epochs)\n    325             cbks.make_logs(model, epoch_logs, training_result, ModeKeys.TRAIN)\n    326 \n\n\/usr\/local\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/keras\/engine\/training_v2.py in run_one_epoch(model, iterator, execution_function, dataset_size, batch_size, strategy, steps_per_epoch, num_samples, mode, training_context, total_epochs)\n    121         step=step, mode=mode, size=current_batch_size) as batch_logs:\n    122       try:\n--&gt; 123         batch_outs = execution_function(iterator)\n    124       except (StopIteration, errors.OutOfRangeError):\n    125         # TODO(kaftan): File bug about tf function and errors.OutOfRangeError?\n\n\/usr\/local\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/keras\/engine\/training_v2_utils.py in execution_function(input_fn)\n     84     # `numpy` translates Tensors to values in Eager mode.\n     85     return nest.map_structure(_non_none_constant_value,\n---&gt; 86                               distributed_function(input_fn))\n     87 \n     88   return execution_function\n\n\/usr\/local\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/eager\/def_function.py in __call__(self, *args, **kwds)\n    437         # Lifting succeeded, so variables are initialized and we can run the\n    438         # stateless function.\n--&gt; 439         return self._stateless_fn(*args, **kwds)\n    440     else:\n    441       canon_args, canon_kwds = \\\n\n\/usr\/local\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/eager\/function.py in __call__(self, *args, **kwargs)\n   1820     \"\"\"Calls a graph function specialized to the inputs.\"\"\"\n   1821     graph_function, args, kwargs = self._maybe_define_function(args, kwargs)\n-&gt; 1822     return graph_function._filtered_call(args, kwargs)  # pylint: disable=protected-access\n   1823 \n   1824   @property\n\n\/usr\/local\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/eager\/function.py in _filtered_call(self, args, kwargs)\n   1139          if isinstance(t, (ops.Tensor,\n   1140                            resource_variable_ops.BaseResourceVariable))),\n-&gt; 1141         self.captured_inputs)\n   1142 \n   1143   def _call_flat(self, args, captured_inputs, cancellation_manager=None):\n\n\/usr\/local\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/eager\/function.py in _call_flat(self, args, captured_inputs, cancellation_manager)\n   1222     if executing_eagerly:\n   1223       flat_outputs = forward_function.call(\n-&gt; 1224           ctx, args, cancellation_manager=cancellation_manager)\n   1225     else:\n   1226       gradient_name = self._delayed_rewrite_functions.register()\n\n\/usr\/local\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/eager\/function.py in call(self, ctx, args, cancellation_manager)\n    509               inputs=args,\n    510               attrs=(\"executor_type\", executor_type, \"config_proto\", config),\n--&gt; 511               ctx=ctx)\n    512         else:\n    513           outputs = execute.execute_with_cancellation(\n\n\/usr\/local\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/eager\/execute.py in quick_execute(op_name, num_outputs, inputs, attrs, ctx, name)\n     65     else:\n     66       message = e.message\n---&gt; 67     six.raise_from(core._status_to_exception(e.code, message), None)\n     68   except TypeError as e:\n     69     keras_symbolic_tensors = [\n\n~\/Library\/Python\/3.7\/lib\/python\/site-packages\/six.py in raise_from(value, from_value)\n\nInvalidArgumentError:  assertion failed: [] [Condition x == y did not hold element-wise:] [x (loss\/dense_3_loss\/SparseSoftmaxCrossEntropyWithLogits\/Shape_1:0) = ] [7 1] [y (loss\/dense_3_loss\/SparseSoftmaxCrossEntropyWithLogits\/strided_slice:0) = ] [7 3]\n     [[node loss\/dense_3_loss\/SparseSoftmaxCrossEntropyWithLogits\/assert_equal\/Assert\/Assert (defined at \/usr\/local\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/framework\/ops.py:1751) ]] [Op:__inference_distributed_function_2031]\n\nFunction call stack:\ndistributed_function\n<\/code><\/pre>\n\n<p>With tensorflow=2.0.0-beta0:  <\/p>\n\n<pre><code>InvalidArgumentError:  logits and labels must have the same first dimension, got logits shape [34,3] and labels shape [2]\n     [[node loss\/dense_2_loss\/SparseSoftmaxCrossEntropyWithLogits\/SparseSoftmaxCrossEntropyWithLogits (defined at &lt;ipython-input-3-4fa88a5fad5e&gt;:70) ]] [Op:__inference_keras_scratch_graph_5038]\n<\/code><\/pre>\n\n<p>Can someone explain or point me in the correct direction?<\/p>\n\n<p>Thanks<\/p>\n\n<p>The following code reproduces the error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport pandas as pd\nimport random\nimport tensorflow as tf\n\nINPUT_SHAPE=[3, 5]\nNUM_POINTS=20\nBATCH_SIZE=7\nEPOCHS=4\n\ndef data_gen(num=10, in_shape=[5, 3]):\n    for i in range(num):\n        x = np.random.rand(in_shape[0], in_shape[1])\n        y = random.randint(0,2)\n        yield x, y\n\ntrain = tf.data.Dataset.from_generator(\n    generator=data_gen,\n    output_types=(tf.float32, tf.int32),\n    args=([NUM_POINTS, INPUT_SHAPE])\n)\n\ndef create_model(input_shape):\n    model = tf.keras.models.Sequential([\n        tf.keras.layers.Dense(100, activation=\"tanh\",input_shape=input_shape),        \n        tf.keras.layers.Dense(3, activation=\"softmax\", kernel_regularizer= tf.keras.regularizers.l2(0.001))\n    ])\n    return model\n\nmodel = create_model(input_shape=INPUT_SHAPE)\n\nmodel.compile(\n    optimizer=tf.keras.optimizers.Adam(learning_rate=1e-4, clipvalue=1.0),\n    loss= tf.keras.losses.SparseCategoricalCrossentropy())\nprint(model.summary())\n\nmodel.fit(train.batch(BATCH_SIZE), epochs=EPOCHS, verbose=2)\nmodel.evaluate(train, steps=None, verbose=1)\n\n<\/code><\/pre>\n\n<p>Model summary:<\/p>\n\n<blockquote>\n  <p>Model: \"sequential_1\"\n  _________________________________________________________________ Layer (type)                 Output Shape              Param #<br>\n  ================================================================= dense_2 (Dense)              (None, 3, 100)            600<br>\n  _________________________________________________________________ dense_3 (Dense)              (None, 3, 3)              303<br>\n  ================================================================= Total params: 903 Trainable params: 903 Non-trainable params: 0<\/p>\n  \n  <hr>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":57842958,"AnswerCount":3,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-08T14:44:05.870","AcceptedAnswerId":57843251.0,"OwnerUserId":12037617.0,"Title":"Keras GaussianNoise layer no effect?","Body":"<p>I wanted to add some Gaussian noise to my Images in my CNN with the keras's functional API, but while testing some different stddev values, I noticed that the Gaussian Layer does nothing to the input data.\nI used the following code for testing:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport cv2\n\nstddev = 0.1\nimage = cv2.imread(&lt;img_path&gt;)\nimage = (image.astype('float32') - 127.5) \/ 127.5\n\ninput_layer = tf.keras.layers.Input(shape=(128,128,3))\ngaus = tf.keras.layers.GaussianNoise(stddev)(input_layer)\nmodel = tf.keras.models.Model(inputs=input_layer, outputs=gaus)\n\nnoisy_image = model(image)\n\nprint(f'Pixel value at 0,0: {image[0,0]}')\nprint(f'Pixel value at 0,0: {noisy_image.numpy()[0,0]}')\n# Output\n# Pixel value at 0,0: [ 0.09803922 -0.30980393 -0.56078434]\n# Pixel value at 0,0: [ 0.09803922 -0.30980393 -0.56078434]\n<\/code><\/pre>\n\n<p>It doesn't matter what value I put in for <em>stddev<\/em> (tried everything from 0.001 to 10000). I expected the values to differ slightly (or heavily when stddev=1000). Am I doing something wrong?<\/p>\n\n<p>Maybe I should mention that I'm using tensorflow-gpu==2.0.0-rc0 on Windows 10<\/p>\n","answers":[{"AnswerId":"57843277","CreationDate":"2019-09-08T15:27:36.217","ParentId":null,"OwnerUserId":"3472735","Title":null,"Body":"<p>These data augmentation layers are only active at train time, so will have no effect when simply calling the model on the image. You can see that it's working with:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport cv2\n\nstddev = 0.1\nimage = cv2.imread(&lt;img_path&gt;)\nimage = (image.astype('float32') - 127.5) \/ 127.5\n\ninput_layer = tf.keras.layers.Input(shape=(128,128,3))\ngaus = tf.keras.layers.GaussianNoise(stddev)(input_layer, training=True)\nmodel = tf.keras.models.Model(inputs=input_layer, outputs=gaus)\n\nnoisy_image = model(image)\n\nprint(f'Pixel value at 0,0: {image[0,0]}')\nprint(f'Pixel value at 0,0: {noisy_image.numpy()[0,0]}')\n# Output\n# Pixel value at 0,0: [ 0.09803922 -0.30980393 -0.56078434]\n# Pixel value at 0,0: [ 0.09803922 -0.30980393 -0.56078434]\n<\/code><\/pre>\n\n<p>But make sure you remove training=True at evaluation time so it doesn't negatively affect performance.<\/p>\n"},{"AnswerId":"57843251","CreationDate":"2019-09-08T15:24:25.863","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>If you check <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/GaussianNoise\" rel=\"nofollow noreferrer\">the docs<\/a>, it says the layer is only active during training as it's supposed to serve as a regularizer. Looking at the source code confirms this. So it looks like if you will need to make sure the model \"knows\" that it's in training mode. There are several ways to do this:<\/p>\n\n<ul>\n<li>If you use the <code>model.compile<\/code>\/ <code>model.fit<\/code> interface, this should be done automatically.<\/li>\n<li>If you use the model as a callable it should accept a <code>training<\/code> parameter that you can set as <code>training=True<\/code> to \"activate\" training mode on a per-call basis. I.e. <code>noisy_image = model(image, training=True)<\/code>.<\/li>\n<li>You can use <code>tf.keras.backend.set_learning_phase(1)<\/code> to \"globally\" activate training mode (call it again later with argument <code>0<\/code> to deactivate).<\/li>\n<\/ul>\n"},{"AnswerId":"57843333","CreationDate":"2019-09-08T15:35:52.300","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>Try\n<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>import matplotlib.pyplot as plt\n\ninput_layer = tf.keras.layers.Input(shape=(128,128,3))\ngaus        = tf.keras.layers.GaussianNoise(stddev,name='output')(input_layer)\nmodel       = tf.keras.models.Model(inputs=input_layer, outputs=gaus)\n<\/code><\/pre>\n\n<pre class=\"lang-python prettyprint-override\"><code>noisy_image = get_layer_outputs(model,'output',image,train_mode=True)\nplt.imshow(image[:,:,0],cmap='bwr')\nplt.show()\nplt.imshow(np.asarray(noisy_image)[0,:,:,0],cmap='bwr')\nplt.show()\n<\/code><\/pre>\n\n<pre class=\"lang-python prettyprint-override\"><code>def get_layer_outputs(model,layer_name,input_data,train_mode=False):\n    outs_tensor   = [layer.output for layer in model.layers if layer_name == layer.name]\n    outs_function = K.function([model.input, K.learning_phase()], outs_tensor)\n    return [outs_function([input_data,int(train_mode)])][0]\n<\/code><\/pre>\n\n<p><code>GaussianNoise<\/code> only works with <code>learning_phase=1<\/code> - i.e. in 'train mode'; unsure if feeding data directly via <code>model(...)<\/code> accomplishes this. W\/ above, using <code>image=np.random.randn(128,128,3)<\/code> and <code>stddev=1<\/code>, I get<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/jlmva.png\" width=\"450\" \/><\/p>\n"}]}
+{"QuestionId":57843070,"AnswerCount":0,"Tags":"<visual-studio><cmake><protocol-buffers><caffe>","CreationDate":"2019-09-08T14:59:51.500","AcceptedAnswerId":null,"OwnerUserId":8103305.0,"Title":"Unable to find header files after cmake generation","Body":"<p>I created a build project of caffe using cmake. Cmake completes the generation and after that, I open VS2017 and no header files can be found. As this is a problem, I'm facing quite often, I now decided to ask you for help.<\/p>\n\n<p>Here is one example:\nI specified the folder <code>C:\/Program Files\/Protobuf\/source\/src<\/code> in CMake as the including directory of the header files.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/DOY7x.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DOY7x.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>As you can see in the next image, this folder contains <code>google\\protobuf\\text_format.h<\/code><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/n3Rsf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/n3Rsf.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Loading the generated solution with VS2017, the header files are not found:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/fq4ij.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/fq4ij.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Sorry, but I'm really losing my religion, as I try to build caffe for two days now and there is always something wrong.<\/p>\n\n<p>As you can see, the same problem occures for other include directories too.<\/p>\n\n<p>By the way, it's not helpful to change the path in the solution as caffe is a quite large lib and it would become a never ending story.<\/p>\n\n<p>Thank you for all advices and regards,\nJan<\/p>\n","answers":[]}
+{"QuestionId":57843564,"AnswerCount":0,"Tags":"<visual-studio><build><caffe><signalr-hub>","CreationDate":"2019-09-08T16:05:59.377","AcceptedAnswerId":null,"OwnerUserId":8103305.0,"Title":"Headerfile <signal.h> is incomplete - SIGHUP and others are missing","Body":"<p>I'm currently building caffe. Part of that program is including signal.h. This header should contain <code>#define SIGHUP = 1<\/code><\/p>\n\n<p>I found that file in <code>C:\\...\\Windows Kits\\10\\Include\\10.0.10240.0\\ucrt<\/code> <\/p>\n\n<p>But my file is incomplete. Many things like <code>#define SIGUP = 1<\/code> or <code>SA_RESTART<\/code> are missing. The struct <code>sigaction<\/code> which should be defined in signal.h is incomplete.<\/p>\n\n<p>I can just download another signal.h file, but then other errors are occuring. One can find so many different signal.h`s in the web... <\/p>\n\n<p>Can anybody tell me, which file I have to use?<\/p>\n","answers":[]}
+{"QuestionId":57843656,"AnswerCount":1,"Tags":"<python><tensorflow><google-cloud-platform>","CreationDate":"2019-09-08T16:18:51.013","AcceptedAnswerId":null,"OwnerUserId":8625760.0,"Title":"GCP output directory is not being created when using --job-dir","Body":"<p>I am trying to follow the <a href=\"https:\/\/github.com\/GoogleCloudPlatform\/training-data-analyst\/tree\/master\/blogs\/rl-on-gcp\/cartpole_policy_gradients\" rel=\"nofollow noreferrer\">training-data-analyst<\/a> tutorial by GoogleCloudPlatform on Github. I recreated the file structure on my local machine and proceed to train locally. <\/p>\n\n<p>The <em>train model locally<\/em> commands in the tutorial are as follows:<\/p>\n\n<pre><code>&gt; OUTPUT_DIR=rl_model \n&gt; JOBNAME=rl_train_$(date -u +%y%m%d_%H%M%S)\n&gt; REGION=us-central1 \n&gt; PACKAGE_PATH=$PWD\/rl_model_code\/trainer \n&gt; export PYTHONPATH=${PYTHONPATH}:${PWD}\/rl_model_code \n&gt; rm $OUTPUT_DIR\n&gt; \n&gt; gcloud ml-engine local train\\\n&gt;     --package-path=$PACKAGE_PATH\\\n&gt;     --module-name=trainer.task\\\n&gt;     --\\\n&gt;     --outdir=$OUTPUT_DIR\n<\/code><\/pre>\n\n<p>The actual instructions given on the tutorial seem to be outdated, for example <code>--outdir<\/code> should be <code>--job-dir<\/code>, and <code>gcloud ml-engine<\/code> should be <code>gcloud ai-platform<\/code>.<\/p>\n\n<p>I run the following working commands which worked on another tutorial in the <a href=\"https:\/\/cloud.google.com\/ml-engine\/docs\/tensorflow\/getting-started-keras?authuser=1\" rel=\"nofollow noreferrer\">official google docs<\/a>:<\/p>\n\n<pre><code>gcloud ai-platform local train   --package-path trainer   --module-name trainer.task   --job-dir rl-model\n<\/code><\/pre>\n\n<h1>Result:<\/h1>\n\n<p>I am getting no errors however the output directory is not being created. Here is the console output:<\/p>\n\n<pre><code>\/anaconda3\/envs\/...\/...\/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\nUsage:\n    trainer.task --outdir=&lt;name&gt;\n                 [--eval]\n                 [--n_games_per_update=&lt;N&gt;]\n                 [--n_hidden=&lt;N&gt;]\n                 [--discount_rate=&lt;rate&gt;]\n                 [--learning_rate=&lt;rate&gt;]\nOptions:\n    -h, --help  Show this screen and exit.\n    --eval  If in eval, make a gif.\n    --outdir=&lt;name&gt; Location to save model (or of saved model if eval).\n    --n_games_per_update=&lt;N&gt; Number of games to play.  [default: 10]\n    --n_hidden=&lt;N&gt;  # of hidden units [default: 10]\n    --discount_rate=&lt;rate&gt;  Reward discount rate. [default: 0.95]\n    --learning_rate=&lt;rate&gt;  Learning rate. [default: 0.01]\n<\/code><\/pre>\n\n<p>Any suggestions as to why the job directory is not being created? How can I resolve this? Thanks in advance.<\/p>\n","answers":[{"AnswerId":"57853586","CreationDate":"2019-09-09T11:58:05.937","ParentId":null,"OwnerUserId":"11372593","Title":null,"Body":"<p>You mixed the gcloud params and your python trainer params.<\/p>\n\n<pre><code># GCloud command with the gcloud param\n&gt; gcloud ml-engine local train\\\n&gt;     --package-path=$PACKAGE_PATH\\\n&gt;     --module-name=trainer.task\\\n# Don't forget the -- which specifies the end of gcloud params and the beginning of the python params\n&gt;     --\\\n# Python params\n&gt;     --outdir=$OUTPUT_DIR\n<\/code><\/pre>\n\n<p>AI-Engine is the new name of ml-engine. All documents\/examples aren't up to date. Even the official documentation (look at the url name!). But it's the same thing.<\/p>\n"}]}
+{"QuestionId":57843831,"AnswerCount":0,"Tags":"<python><tensorflow><deployment><deep-learning>","CreationDate":"2019-09-08T16:42:46.003","AcceptedAnswerId":null,"OwnerUserId":12038198.0,"Title":"Train network in 16FP gives Nan","Body":"<p>When I train my network with 32FP everything works great. as soon as I try to train it with 16FP, using keras with tensorflow backend K.set_floatx('float16') I get Nans in many of the network params on the first batch. \nI try to deploy my network on GPU using tensorRT and nvidia states it works optimaly with 16FP.<\/p>\n\n<p>I tried gradient clipping, lower LR, removing batch normalization layer, retraining my 32FP weights (in the last case I get Nans on the second batch, otherwise I get Nans already on the first batch)<\/p>\n\n<p>Why is there such a big different between 32\/16FP?\nI would love to get a good advice! \nThank you!<\/p>\n","answers":[]}
+{"QuestionId":57844580,"AnswerCount":0,"Tags":"<python><tensorflow><semantic-segmentation>","CreationDate":"2019-09-08T18:12:06.847","AcceptedAnswerId":null,"OwnerUserId":11410413.0,"Title":"How to fix \"DataLossError : Unable to open table file .\\checkpoints: NewRandomAccessFile failed to Create\/Open: .\\checkpoints : Access is denied\"?","Body":"<p>I have trained a model (Here's <a href=\"https:\/\/github.com\/GeorgeSeif\/Semantic-Segmentation-Suite\" rel=\"nofollow noreferrer\">a link<\/a>!) I successfully trained the model by:<\/p>\n\n<pre><code>python train.py --model DeepLabV3_plus\n<\/code><\/pre>\n\n<p>and when I wanted to run test.py I got the error mentioned in the title. The checkpoints folder includes all the necessary weights for testing and I tried to assign the checkpoint to the model with:<\/p>\n\n<pre><code>python test.py --model DeepLabV3_plus --checkpoint_path checkpoints\n<\/code><\/pre>\n\n<p>but failed running it. Could you please help me to run the test.py? <\/p>\n\n<p>P.S: I eventually need to predict the model on a new dataset, any help for that part is also appreciated.<\/p>\n\n<p>I have read <a href=\"https:\/\/github.com\/davidsandberg\/facenet\/issues\/251\" rel=\"nofollow noreferrer\">https:\/\/github.com\/davidsandberg\/facenet\/issues\/251<\/a> and <a href=\"https:\/\/github.com\/balancap\/SSD-Tensorflow\/issues\/150\" rel=\"nofollow noreferrer\">https:\/\/github.com\/balancap\/SSD-Tensorflow\/issues\/150<\/a> but they couldn't help me.<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>Loading model checkpoint weights ...\n2019-09-08 20:05:13.667832: W tensorflow\/core\/util\/tensor_slice_reader.cc:95] Could not open .\\checkpoints: Unknown: NewRandomAccessFile failed to Create\/Open: .\\checkpoints : Access is denied.\n; Input\/output error\n2019-09-08 20:05:13.873491: W tensorflow\/core\/util\/tensor_slice_reader.cc:95] Could not open .\\checkpoints: Unknown: NewRandomAccessFile failed to Create\/Open: .\\checkpoints : Access is denied.\n; Input\/output error\n2019-09-08 20:05:13.878487: W tensorflow\/core\/framework\/op_kernel.cc:1273] OP_REQUIRES failed at save_restore_tensor.cc:175 : Data loss: Unable to open table file .\\checkpoints: Unknown: NewRandomAccessFile failed to Create\/Open: .\\checkpoints : Access is denied.\n; Input\/output error\nTraceback (most recent call last):\n  File \"C:\\Users\\xxxxx\\Anaconda3\\envs\\py36\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1334, in _do_call\n    return fn(*args)\n  File \"C:\\Users\\xxxxx\\Anaconda3\\envs\\py36\\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\\xxxxx\\Anaconda3\\envs\\py36\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1407, in _call_tf_sessionrun\n    run_metadata)\ntensorflow.python.framework.errors_impl.DataLossError: Unable to open table file .\\checkpoints: Unknown: NewRandomAccessFile failed to Create\/Open: .\\checkpoints : Access is denied.\n; Input\/output error\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         [[{{node save_1\/RestoreV2\/_307}} = _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_312_save_1\/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):\n  File \"test.py\", line 43, in &lt;module&gt;\n    saver.restore(sess, args.checkpoint_path)\n  File \"C:\\Users\\xxxxx\\Anaconda3\\envs\\py36\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 1546, in restore\n    {self.saver_def.filename_tensor_name: save_path})\n  File \"C:\\Users\\xxxxx\\Anaconda3\\envs\\py36\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 929, in run\n    run_metadata_ptr)\n  File \"C:\\Users\\xxxxx\\Anaconda3\\envs\\py36\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1152, in _run\n    feed_dict_tensor, options, run_metadata)\n  File \"C:\\Users\\xxxxx\\Anaconda3\\envs\\py36\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1328, in _do_run\n    run_metadata)\n  File \"C:\\Users\\xxxxx\\Anaconda3\\envs\\py36\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1348, in _do_call\n    raise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.DataLossError: Unable to open table file .\\checkpoints: Unknown: NewRandomAccessFile failed to Create\/Open: .\\checkpoints : Access is denied.\n; Input\/output error\n         [[node save_1\/RestoreV2 (defined at test.py:42)  = 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         [[{{node save_1\/RestoreV2\/_307}} = _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_312_save_1\/RestoreV2\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]()]]\n\nCaused by op 'save_1\/RestoreV2', defined at:\n  File \"test.py\", line 42, in &lt;module&gt;\n    saver=tf.train.Saver(max_to_keep=1000)\n  File \"C:\\Users\\xxxxx\\Anaconda3\\envs\\py36\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 1102, in __init__\n    self.build()\n  File \"C:\\Users\\xxxxx\\Anaconda3\\envs\\py36\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 1114, in build\n    self._build(self._filename, build_save=True, build_restore=True)\n  File \"C:\\Users\\xxxxx\\Anaconda3\\envs\\py36\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 1151, in _build\n    build_save=build_save, build_restore=build_restore)\n  File \"C:\\Users\\xxxxx\\Anaconda3\\envs\\py36\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 795, in _build_internal\n    restore_sequentially, reshape)\n  File \"C:\\Users\\xxxxx\\Anaconda3\\envs\\py36\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 406, in _AddRestoreOps\n    restore_sequentially)\n  File \"C:\\Users\\xxxxx\\Anaconda3\\envs\\py36\\lib\\site-packages\\tensorflow\\python\\training\\saver.py\", line 862, in bulk_restore\n    return io_ops.restore_v2(filename_tensor, names, slices, dtypes)\n  File \"C:\\Users\\xxxxx\\Anaconda3\\envs\\py36\\lib\\site-packages\\tensorflow\\python\\ops\\gen_io_ops.py\", line 1466, in restore_v2\n    shape_and_slices=shape_and_slices, dtypes=dtypes, name=name)\n  File \"C:\\Users\\xxxxx\\Anaconda3\\envs\\py36\\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\\xxxxx\\Anaconda3\\envs\\py36\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py\", line 488, in new_func\n    return func(*args, **kwargs)\n  File \"C:\\Users\\xxxxx\\Anaconda3\\envs\\py36\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3274, in create_op\n    op_def=op_def)\n  File \"C:\\Users\\xxxxx\\Anaconda3\\envs\\py36\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 1770, in __init__\n    self._traceback = tf_stack.extract_stack()\n\nDataLossError (see above for traceback): Unable to open table file .\\checkpoints: Unknown: NewRandomAccessFile failed to Create\/Open: .\\checkpoints : Access is denied.\n; Input\/output error\n         [[node save_1\/RestoreV2 (defined at test.py:42)  = 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         [[{{node save_1\/RestoreV2\/_307}} = _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_312_save_1\/RestoreV2\", tensor_type=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/device:GPU:0\"]()]]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57844947,"AnswerCount":0,"Tags":"<python><pytorch>","CreationDate":"2019-09-08T19:03:42.857","AcceptedAnswerId":null,"OwnerUserId":9344450.0,"Title":"Can a kernel get involve and updated in the gradient map by this way?","Body":"<p>This question might be naive but qusetioned me for a long while. If we wrap a convolution function by a custom function, and pass the parameters, like this<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>my_w = nn.Parameter(torch.randn(1,1,3,3))\nmy_b = nn.Parameter(torch.randn(1,1,1))\n\ndef conv1(x, my_w, my_b):\n    x = F.conv2d(x, weight=my_w, bias=my_b, stride, padding)\n    return x\n<\/code><\/pre>\n\n<p>Can my_w and my_b get updated? as they are global parameters, they do not really live in this function<\/p>\n\n<p>If can, why? If cant, how to fix it(without 'global' tag)?<\/p>\n\n<p>Thank you guys!<\/p>\n","answers":[]}
+{"QuestionId":57844987,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning>","CreationDate":"2019-09-08T19:10:30.993","AcceptedAnswerId":null,"OwnerUserId":11891626.0,"Title":"Keras - copy Conv2D layer","Body":"<p>I want to make a copy of a Conv2D layer.<\/p>\n\n<p>I tried this:<\/p>\n\n<p>Edit: I've changed the example code to a <a href=\"https:\/\/stackoverflow.com\/help\/minimal-reproducible-example\">mcve<\/a><\/p>\n\n<p>Edit2: I've changed the code based on <a href=\"https:\/\/stackoverflow.com\/users\/5085211\/fuglede\">fuglede<\/a>'s answer<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense, Conv2D, Flatten\nfrom keras.datasets import mnist\nfrom keras.utils import to_categorical\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport random\n\n(X_train, y_train), (X_test, y_test) = mnist.load_data()\n\n\nX_train = X_train.reshape(60000, 28, 28, 1)\nX_test = X_test.reshape(10000, 28, 28, 1)\n\ny_train = to_categorical(y_train)\ny_test = to_categorical(y_test)\n\nmodel = Sequential()\n\nmodel.add(Conv2D(random.randint(32, 64), kernel_size=random.randint(1, 3), activation='relu', input_shape=(28, 28, 1)))\nmodel.add(Conv2D(32, kernel_size=3, activation='relu'))\nmodel.add(Flatten())\nmodel.add(Dense(10, activation='softmax'))\n\nmodel.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])\n\nother_model = Sequential()\n\nlayer = model.layers[1]\n\nother_model.add(Conv2D(random.randint(32, 64), kernel_size=random.randint(1, 3), activation='relu', input_shape=(28, 28, 1)))\n\ncopy_layer = Conv2D(layer.filters, kernel_size=layer.kernel_size, activation='relu')\nother_model.add(copy_layer)\ncopy_layer.set_weights(layer.get_weights())\n<\/code><\/pre>\n\n<p>But I'm getting this error:<\/p>\n\n<pre><code>ValueError: Layer weight shape (3L, 3L, 61L, 32L) not compatible with provided weight shape (3L, 3L, 40L, 32L)\n<\/code><\/pre>\n\n<p>Edit: The purpose of this is, I'm using a genetic algorithm to evolve\/\"train\" a set a neural networks, and this a part of the crossover step. <\/p>\n","answers":[{"AnswerId":"57845384","CreationDate":"2019-09-08T20:06:52.027","ParentId":null,"OwnerUserId":"5085211","Title":null,"Body":"<p>This happens because the layer is only initialized once it's added to a model. If you swap the two last lines of your example, it should work as expected.<\/p>\n"}]}
+{"QuestionId":57845282,"AnswerCount":0,"Tags":"<python><machine-learning><pytorch>","CreationDate":"2019-09-08T19:52:11.847","AcceptedAnswerId":null,"OwnerUserId":10047785.0,"Title":"how to run Multi-level Scene Description Network in ubuntu?","Body":"<p>I am currently building a neural network in machine learning where our machine will be able to generate a sentence by gathering features from the image. <br>\nI want to test a model called Multi-level Scene Description Network  I am following along the code implementation for the model HERE <a href=\"https:\/\/github.com\/yikang-li\/MSDN\" rel=\"nofollow noreferrer\">https:\/\/github.com\/yikang-li\/MSDN<\/a>.<br>\n But when i try to run step 3. <\/p>\n\n<pre><code>cd MSDN\/faster_rcnn \n.\/make.sh\ncd ..\n<\/code><\/pre>\n\n<p>gives the following error<\/p>\n\n<pre><code>.\/make.sh\nTraceback (most recent call last):\n  File \"setup.py\", line 59, in &lt;module&gt;\n    CUDA = locate_cuda()\n File \"setup.py\", line 52, in locate_cuda\n   for k, v in cudaconfig.iteritems():\n   AttributeError: 'dict' object has no attribute 'iteritems'\n   Compiling roi pooling kernels by nvcc...\n   .\/make.sh: line 10: nvcc: command not found\n   Traceback (most recent call last):\n     File \"build.py\", line 3, in &lt;module&gt;\n       from torch.utils.ffi import create_extension\n     File \"\/home\/faaiz\/anaconda3\/lib\/python3.7\/site-packages\/torch                             \/utils\/ffi\/__init__.py\", line 1, in &lt;module&gt;\nraise ImportError(\"torch.utils.ffi is deprecated. Please use cpp extensions instead.\")\nImportError: torch.utils.ffi is deprecated. Please use cpp extensions instead.\n<\/code><\/pre>\n\n<p>I put some search into this and came to conclusion that this might be because of the python version I am using \"Python 3.7\" where as this code was made for \"Python 3.5\".<br>\nI changed my conda environment from 3.7 to 3.5 using \n<a href=\"https:\/\/docs.anaconda.com\/anaconda\/navigator\/tutorials\/create-python35-environment\/\" rel=\"nofollow noreferrer\">https:\/\/docs.anaconda.com\/anaconda\/navigator\/tutorials\/create-python35-environment\/<\/a> <br>\nbut still no difference.<br>\nif I can get some guide on this, it will be greatly appreciated<\/p>\n","answers":[]}
+{"QuestionId":57845439,"AnswerCount":2,"Tags":"<tensorflow><nlp><language-model><bert-language-model>","CreationDate":"2019-09-08T20:14:09.827","AcceptedAnswerId":null,"OwnerUserId":2056969.0,"Title":"Which model (GPT2, BERT, XLNet and etc) would you use for a text classification task? Why?","Body":"<p>I'm trying to train a model for a sentence classification task. The input is a sentence (a vector of integers) and the output is a label (0 or 1). I've seen some articles here and there about using Bert and GPT2 for text classification tasks. However, I'm not sure which one should I pick to start with. Which of these recent models in NLP such as original Transformer model, Bert, GPT2, XLNet would you use to start with? And why? I'd rather to implement in Tensorflow, but I'm flexible to go for PyTorch too. \nThanks!<\/p>\n","answers":[{"AnswerId":"57845877","CreationDate":"2019-09-08T21:28:37.183","ParentId":null,"OwnerUserId":"4451232","Title":null,"Body":"<p>It highly depends on your dataset and is part of the data scientist's job to find which model is more suitable for a particular task in terms of selected performance metric, training cost, model complexity etc.<\/p>\n\n<p>When you work on the problem you will probably test all of the above models and compare them. Which one of them to choose first? Andrew Ng in <a href=\"https:\/\/www.deeplearning.ai\/machine-learning-yearning\/\" rel=\"nofollow noreferrer\">\"Machine Learning Yearning\"<\/a> suggest starting with simple model so you can quickly iterate and test your idea, data preprocessing pipeline etc.<\/p>\n\n<blockquote>\n  <p>Don\u2019t start off trying to design and build the perfect system.\n  Instead, build and train a basic system quickly\u2014perhaps in just a few\n  days<\/p>\n<\/blockquote>\n\n<p>According to this suggestion, you can start with a simpler model such as <a href=\"https:\/\/arxiv.org\/pdf\/1801.06146.pdf\" rel=\"nofollow noreferrer\">ULMFiT<\/a> as a baseline, verify your ideas and then move on to more complex models and see how they can improve your results.<\/p>\n\n<p>Note that modern NLP models contain a large number of parameters and it is difficult to train them from scratch without a large dataset. That's why you may want to use <em>transfer learning<\/em>: you can download pre-trained model and use it as a basis and fine-tune it to your task-specific dataset to achieve better performance and reduce training time.<\/p>\n"},{"AnswerId":"57847458","CreationDate":"2019-09-09T03:15:56.247","ParentId":null,"OwnerUserId":"3518108","Title":null,"Body":"<p>I agree with Max's answer, but if the constraint is to use a state of the art large pretrained model, there is a really easy way to do this. The library by HuggingFace called pytorch-transformers. Whether you chose BERT, XLNet, or whatever, they're easy to swap out. <a href=\"https:\/\/towardsdatascience.com\/https-medium-com-chaturangarajapakshe-text-classification-with-transformer-models-d370944b50ca\" rel=\"nofollow noreferrer\">Here<\/a> is a detailed tutorial on using that library for text classification.<\/p>\n\n<p>EDIT: I just came across this repo, <a href=\"https:\/\/github.com\/ThilinaRajapakse\/pytorch-transformers-classification\" rel=\"nofollow noreferrer\">pytorch-transformers-classification<\/a> (Apache 2.0 license), which is a tool for doing exactly what you want.<\/p>\n"}]}
+{"QuestionId":57845542,"AnswerCount":0,"Tags":"<python-3.x><matplotlib><keras><operating-system><imagemagick>","CreationDate":"2019-09-08T20:30:24.383","AcceptedAnswerId":null,"OwnerUserId":10740549.0,"Title":"Imagemagick convert doesn't convert","Body":"<p>So my problem is in the convert from Imagemagick, this code should convert the image but it doesn't. Maybe it's because I'm on windows but idk. <\/p>\n\n<p>I've provided the entire path as well as instead of using convert used magick but it still didn't work, it worked on the cmd but for some reason not in the program.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    def display_mnist(self): # Get the png, apply color map and print to canvas\n        from PIL import ImageTk, Image\n        im = Image.fromarray(self.mnist_user_image*255).convert('RGB')\n        im.save(\"E:\\projects\\visuals\/mnist.png\")\n        os.system('C:\\Program Files\\ImageMagick-7.0.8-Q16\/magick.exe E:\\projects\\visuals\/mnist.png -resize 28x28 E:\\projects\\visuals\/mnist.png')\n        self.mnist_user_image = np.array(plt.imread('E:\\projects\\visuals\/mnist.png'),\n                                                    dtype=np.float64)\n<\/code><\/pre>\n\n<p>I want the program to work, or at least not give this error: \nOSError: [Errno 22] Invalid argument: 'E:\\projects\\x0bisuals\/mnist.png'<\/p>\n","answers":[]}
+{"QuestionId":57845942,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets><tfrecord>","CreationDate":"2019-09-08T21:38:30.167","AcceptedAnswerId":58012609.0,"OwnerUserId":6921786.0,"Title":"How to specify starting point of tf.Data.iterator (or skipping first X batches)?","Body":"<p>For a tensorflow dataset iterator (tf.data.Iterator), what is the best way of skipping over the first X batches, but ONLY in the first iteration, and not for subsequent iterations when repeat() is specified)?<\/p>\n\n<p>I tried the following but it didn't work:<\/p>\n\n<pre><code>import tensorflow as tf\nimport pandas as pd\nfrom pyspark.sql import SparkSession\n\nspark = SparkSession.builder.master('local[*]').config(\"spark.jars\",'some\/path\/spark-tensorflow-connector_2.11-1.10.0.jar').getOrCreate()\n\ndf = pd.DataFrame({'x': range(10), 'y': [i*2 for i in range(10)]})\ndf = spark.createDataFrame(df)\n\ndf.write.format('tfrecords').option('recordType', 'Example').mode(\"overwrite\").save('testdata')\n\ndef parse_function(proto):\n    feature_description = {\n    'x': tf.FixedLenFeature([], tf.int64),\n    'y': tf.FixedLenFeature([], tf.int64)\n    }\n\n    parsed_features = tf.parse_single_example(proto, feature_description)\n\n    x = parsed_features['x']\n    y = parsed_features['y']\n\n    return {'x': x, 'y': y}\n\ndef load_data(filename_pattern, parse_function, batch_size=200, skip_batches=0):\n    files = tf.data.Dataset.list_files(file_pattern=filename_pattern, shuffle=False)\n    dataset = tf.data.TFRecordDataset(files)\n\n    dataset = dataset.repeat()\n\n    dataset = dataset.map(parse_function)\n    dataset = dataset.apply(tf.contrib.data.batch_and_drop_remainder(batch_size))\n\n    dataset = dataset.prefetch(2)\n\n    # Create an iterator\n    iterator = dataset.make_one_shot_iterator()\n    data = iterator.get_next()\n\n    with tf.Session() as sess:\n        for i in range(skip_batches):\n            sess.run(data)\n\n    return data\n\n# skip first three batches\ndata = load_data('testdata\/part-*', parse_function, batch_size=2, skip_batches=3)\n\nsess = tf.Session()\n\nfor i in range(3):\n    print(sess.run(data))\n<\/code><\/pre>\n\n<p>Expected\/desired: <\/p>\n\n<pre><code>    {'y': array([12, 14]), 'x': array([6, 7])}\n    {'y': array([16, 18]), 'x': array([8, 9])}\n    {'y': array([0, 2]), 'x': array([0, 1])}\n<\/code><\/pre>\n\n<p>Actual: <\/p>\n\n<pre><code>    {'y': array([0, 2]), 'x': array([0, 1])}\n    {'y': array([4, 6]), 'x': array([2, 3])}\n    {'y': array([8, 10]), 'x': array([4, 5])}\n<\/code><\/pre>\n\n<p>Thanks in advance for any help!<\/p>\n","answers":[{"AnswerId":"58012609","CreationDate":"2019-09-19T13:51:11.383","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>Instead of going through the <code>tf.Dataset.iterator()<\/code>, why don't you skip the first X batches?<\/p>\n\n<p>Let's say you want 10 Batches and each has 32 elements, means total 320 elements. So you can skip these using <code>tf.Dataset.skip(320)<\/code> (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset#skip\" rel=\"nofollow noreferrer\">skip<\/a>) which gives you dataset with first 10 Batches skipped. <\/p>\n"}]}
+{"QuestionId":57846223,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorflow-hub>","CreationDate":"2019-09-08T22:32:46.400","AcceptedAnswerId":null,"OwnerUserId":12039027.0,"Title":"Why do I get 'Connecting to invalid output of source node' error with a Tensorflow-Hub KerasLayer?","Body":"<p><strong>Edit:<\/strong> I tried turning on eager execution to see if I could identify exactly where the issue was occuring, and eager execution stopped the errors and made it run successfully. No idea why that would be the case and unfortunately that really doesn't help me.<\/p>\n\n<p><strong>Original Post:<\/strong> I'm pretty new to Tensorflow and I'm trying to understand how to use Tensorflow-Hub modules in a tf.keras model. My goal is to create an email classification system to route emails in my organization.<\/p>\n\n<p>I have already built a model using data that was pre-processed using the universal sentence encoder module. This was an RNN and worked quite effectively, but I'm interested in whether I can improve my accuracy.<\/p>\n\n<p>Now I want to incorporate the module directly into my neural net so I can train it.<\/p>\n\n<p>I'm running this in a Jupyter Notebook.<\/p>\n\n<p>I built a simple, non-RNN model to try and get the Tensorflow-Hub module training.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>print(\"Version: \", tf.__version__)\nprint(\"Eager mode: \", tf.executing_eagerly())\nprint(\"Hub version: \", hub.__version__)\nprint(\"GPU is\", \"available\" if tf.test.is_gpu_available() else \"NOT AVAILABLE\")\n\nhub_module = \"https:\/\/tfhub.dev\/google\/tf2-preview\/nnlm-en-dim128\/1\"\nmodel = models.Sequential()\nmodel.add(hub.KerasLayer(hub_module, input_shape=[], dtype=tf.string, trainable=True))\nmodel.add(layers.Dense(128, activation='relu'))\nmodel.add(layers.Dense(64, activation='relu'))\nmodel.add(layers.Dense(3, activation='softmax'))\nmodel.build()\nmodel.summary()\nmodel.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy', 'mae'])\n\n#Fake data\ntrain_data = [[\"Hello how are you\"], [\"Goodbye my friend\"], [\"Happiness is a warm slice of toast\"]]\ntrain_labels = [[1, 0, 0],[0, 1, 0],[0, 0, 1]]\n\ntrain_dataset = tf.data.Dataset.from_tensor_slices((train_data, train_labels))\n\nmodel.fit(train_dataset, epochs=1, verbose=2)    \n<\/code><\/pre>\n\n<p>Here is my full console output:<\/p>\n\n<pre><code>Version:  1.14.0\nEager mode:  False\nHub version:  0.6.0\nGPU is available\nModel: \"sequential_5\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nkeras_layer_5 (KerasLayer)   (None, 128)               124642688 \n_________________________________________________________________\ndense_15 (Dense)             (None, 128)               16512     \n_________________________________________________________________\ndense_16 (Dense)             (None, 64)                8256      \n_________________________________________________________________\ndense_17 (Dense)             (None, 3)                 195       \n=================================================================\nTotal params: 124,667,651\nTrainable params: 124,667,651\nNon-trainable params: 0\n_________________________________________________________________\nWARNING:tensorflow:Expected a shuffled dataset but input dataset `x` is not shuffled. Please invoke `shuffle()` on input dataset.\n\nWARNING:tensorflow:Expected a shuffled dataset but input dataset `x` is not shuffled. Please invoke `shuffle()` on input dataset.\n\n---------------------------------------------------------------------------\nInvalidArgumentError                      Traceback (most recent call last)\n~\\Anaconda3\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\client\\session.py in _do_call(self, fn, *args)\n   1355     try:\n-&gt; 1356       return fn(*args)\n   1357     except errors.OpError as e:\n\n~\\Anaconda3\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\client\\session.py in _run_fn(feed_dict, fetch_list, target_list, options, run_metadata)\n   1338       # Ensure any changes to the graph are reflected in the runtime.\n-&gt; 1339       self._extend_graph()\n   1340       return self._call_tf_sessionrun(\n\n~\\Anaconda3\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\client\\session.py in _extend_graph(self)\n   1373     with self._graph._session_run_lock():  # pylint: disable=protected-access\n-&gt; 1374       tf_session.ExtendSession(self._session)\n   1375 \n\nInvalidArgumentError: Node 'Adam\/gradients\/keras_layer_1\/StatefulPartitionedCall_grad\/StatefulPartitionedCall': Connecting to invalid output 1 of source node keras_layer_1\/StatefulPartitionedCall which has 1 outputs\n\nDuring handling of the above exception, another exception occurred:\n\nInvalidArgumentError                      Traceback (most recent call last)\n&lt;ipython-input-11-492e87ad5d5d&gt; in &lt;module&gt;\n     28 train_dataset = tf.data.Dataset.from_tensor_slices((train_data, train_labels))\n     29 \n---&gt; 30 model.fit(train_dataset, epochs=1, verbose=2)\n\n~\\Anaconda3\\envs\\tensorflow\\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, validation_freq, max_queue_size, workers, use_multiprocessing, **kwargs)\n    778           validation_steps=validation_steps,\n    779           validation_freq=validation_freq,\n--&gt; 780           steps_name='steps_per_epoch')\n    781 \n    782   def evaluate(self,\n\n~\\Anaconda3\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_arrays.py in model_iteration(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, validation_freq, mode, validation_in_fit, prepared_feed_values_from_dataset, steps_name, **kwargs)\n    139       reset_dataset_after_each_epoch = True\n    140       steps_per_epoch = training_utils.infer_steps_for_dataset(\n--&gt; 141           inputs, steps_per_epoch, epochs=epochs, steps_name=steps_name)\n    142     input_iterator = _get_iterator(inputs, model._distribution_strategy)\n    143 \n\n~\\Anaconda3\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training_utils.py in infer_steps_for_dataset(dataset, steps, epochs, steps_name)\n   1391   \"\"\"\n   1392   assert isinstance(dataset, dataset_ops.DatasetV2)\n-&gt; 1393   size = K.get_value(cardinality.cardinality(dataset))\n   1394   if size == cardinality.INFINITE and steps is None:\n   1395     raise ValueError('When passing an infinitely repeating dataset, you '\n\n~\\Anaconda3\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py in get_value(x)\n   2987     return function([], x)(x)\n   2988 \n-&gt; 2989   return x.eval(session=get_session((x,)))\n   2990 \n   2991 \n\n~\\Anaconda3\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py in get_session(op_input_list)\n    460   if not _MANUAL_VAR_INIT:\n    461     with session.graph.as_default():\n--&gt; 462       _initialize_variables(session)\n    463   return session\n    464 \n\n~\\Anaconda3\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py in _initialize_variables(session)\n    877     # marked as initialized.\n    878     is_initialized = session.run(\n--&gt; 879         [variables_module.is_variable_initialized(v) for v in candidate_vars])\n    880     uninitialized_vars = []\n    881     for flag, v in zip(is_initialized, candidate_vars):\n\n~\\Anaconda3\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\client\\session.py in run(self, fetches, feed_dict, options, run_metadata)\n    948     try:\n    949       result = self._run(None, fetches, feed_dict, options_ptr,\n--&gt; 950                          run_metadata_ptr)\n    951       if run_metadata:\n    952         proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\n~\\Anaconda3\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\client\\session.py in _run(self, handle, fetches, feed_dict, options, run_metadata)\n   1171     if final_fetches or final_targets or (handle and feed_dict_tensor):\n   1172       results = self._do_run(handle, final_targets, final_fetches,\n-&gt; 1173                              feed_dict_tensor, options, run_metadata)\n   1174     else:\n   1175       results = []\n\n~\\Anaconda3\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\client\\session.py in _do_run(self, handle, target_list, fetch_list, feed_dict, options, run_metadata)\n   1348     if handle is None:\n   1349       return self._do_call(_run_fn, feeds, fetches, targets, options,\n-&gt; 1350                            run_metadata)\n   1351     else:\n   1352       return self._do_call(_prun_fn, handle, feeds, fetches)\n\n~\\Anaconda3\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\client\\session.py in _do_call(self, fn, *args)\n   1368           pass\n   1369       message = error_interpolation.interpolate(message, self._graph)\n-&gt; 1370       raise type(e)(node_def, op, message)\n   1371 \n   1372   def _extend_graph(self):\n\nInvalidArgumentError: Node 'Adam\/gradients\/keras_layer_1\/StatefulPartitionedCall_grad\/StatefulPartitionedCall': Connecting to invalid output 1 of source node keras_layer_1\/StatefulPartitionedCall which has 1 outputs\n<\/code><\/pre>\n","answers":[{"AnswerId":"57859381","CreationDate":"2019-09-09T18:26:42.233","ParentId":null,"OwnerUserId":"12039027","Title":null,"Body":"<p>This issue was resolved by upgrading to the TF 2.0 RC from TF 1.14.<\/p>\n"}]}
+{"QuestionId":57846414,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-08T23:14:20.873","AcceptedAnswerId":null,"OwnerUserId":4936825.0,"Title":"How to use feature_column v2 in Tensorflow (TF-Ranking)","Body":"<p>I'm using TF-Ranking to train a recommendation engine. I have encountered a problem that seems to be a version incompatibility issue concerning <code>tf.feature_column<\/code> API.<\/p>\n\n<p>The short version of my question is: What is a v2 feature column (TF 2.0?) (see <a href=\"https:\/\/github.com\/tensorflow\/ranking\/blob\/31fc134816cc4974a46a11e7bb2df0066d0a88f0\/tensorflow_ranking\/python\/feature.py#L92\" rel=\"nofollow noreferrer\">this<\/a> for instance) and how can I ensure that my feature columns are treated as v2, while I'm still using TF 1.14. <\/p>\n\n<p>Here is the details:<\/p>\n\n<p>I'm unable to shorten my code sufficiently to provide a reproducible example. But I will try to describe the problem in words.<\/p>\n\n<p>TF Version: 1.14\nOS: Ubuntu 18.04<\/p>\n\n<p>I initialy had two features in my model, <code>user<\/code> and <code>item<\/code>, both sparse categorical features which were wrapped in their own <code>tf.feature_column.embedding_column<\/code>. I was able to use the <code>train_and_evaluate<\/code> method of the <code>Estimator<\/code> and export the model for serving.<\/p>\n\n<p>Then I added a new feature <code>curr_item<\/code> which is only present during prediction (as a context feature). This shares the embeddings with <code>item<\/code>. So now I have a <code>tf.feature_column.shared_embedding_columns<\/code> which wraps both <code>item<\/code> and <code>current_item<\/code>.<\/p>\n\n<p>Now calling <code>train_and_evaluate<\/code> results in the  following error (shortened messages):<\/p>\n\n<blockquote>\n  <p>ValueError: Could not load all requested variables from checkpoint. Please make sure your model_fn does not expect variables that were not saved in the checkpoint. \n  Key input_layer\/user_embedding\/embedding_weights not found in checkpoint<\/p>\n<\/blockquote>\n\n<p>Note that calling <code>train<\/code> method only works fine. My understanding is that once it gets to evaluation, it tries to load the variables from the checkpoint, but that variable doesn't exist. I did a little debugging and found the reason:<\/p>\n\n<p>When <code>encode_listwise_features<\/code> is called during training (which in turn calls <code>encode_features<\/code>) all features (<code>user<\/code> and <code>item<\/code>) are \"V2\" (not sure what that means) and so the following if statement holds:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/ranking\/blob\/31fc134816cc4974a46a11e7bb2df0066d0a88f0\/tensorflow_ranking\/python\/feature.py#L92\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/ranking\/blob\/31fc134816cc4974a46a11e7bb2df0066d0a88f0\/tensorflow_ranking\/python\/feature.py#L92<\/a><\/p>\n\n<p>and both variables are named with an <code>encoding_layer<\/code> prefix (scope name?):<\/p>\n\n<blockquote>\n  <p>encoding_layer\/user_embedding\/embedding_weights\n  encoding_layer\/item_embedding\/embedding_weights<\/p>\n<\/blockquote>\n\n<p>But when I call the same function for all three features (a little confused wether this is in eval or predict mode), some of these are not \"V2\" and we end up in the else part of the above condition which calls <code>input_layer<\/code> direcetly and variables are named using <code>input_layer<\/code> prefix. Now TF is trying to restore<\/p>\n\n<blockquote>\n  <p>input_layer\/user_embedding\/embedding_weights<\/p>\n<\/blockquote>\n\n<p>from the check-point, but that name doesn't exist in the checkpoint, because it was called <\/p>\n\n<blockquote>\n  <p>encoding_layer\/user_embedding\/embedding_weights<\/p>\n<\/blockquote>\n\n<p>in training.<\/p>\n\n<p>So:<\/p>\n\n<p>1) How can I ensure that all my features are treated as v2 at all stages? I tried using <code>tf.compat.v2.feature_column<\/code> but that didn't help. There is already a ToDo note above that if statement for this.\n2) Can the <code>encode_feature<\/code> be modified to avoid this situation? e.g. raise an exception with a helpful message?<\/p>\n","answers":[]}
+{"QuestionId":57846434,"AnswerCount":2,"Tags":"<tensorflow><keras><lstm>","CreationDate":"2019-09-08T23:18:35.467","AcceptedAnswerId":null,"OwnerUserId":4621314.0,"Title":"Difference between sample step and time step in LSTM (Keras)","Body":"<p>I'm trying to understand how the state progresses in an LSTM-layer. If I have the following code<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential()\nmodel.add(LSTM(2, return_sequences=True,input_shape=(4,2),stateful=False,batch_size=4))\nyp=model.predict(np.array([    [[0,0],[0,1],[0,0],[1,1]],\n                               [[0,1],[0,0],[1,1],[0,0]],\n                               [[0,0],[1,1],[0,0],[0,1]],\n                               [[1,1],[0,0],[0,1],[0,0]],\n                              ]))\nprint(yp)\n<\/code><\/pre>\n\n<p>why do I not get yp[0,:,:] equal yp[:,0,:]?<\/p>\n","answers":[{"AnswerId":"57860488","CreationDate":"2019-09-09T19:58:40.270","ParentId":null,"OwnerUserId":"4621314","Title":null,"Body":"<p>It seems I misunderstood how the state variables are passed around. I had expected that the state variables were pass between sample steps. I have illustrated this in the image below. However that seems not to be that case. And if one reads the documentation carefully, it is stated how it works. <\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/OfIK2.png\" alt=\"My initial and wrong understanding of state flow\"><\/p>\n\n<p>I was fooled by the fact the with default initialization, an input of zeros does not change the state, so in my example y01=y10, y02=y11, y03=y12, and similar from sample 3 to 4.<\/p>\n\n<p>My current understanding is that state variables are pass as in the illustration below.<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/dHwOe.png\" alt=\"My current understanding\"><\/p>\n\n<p>This means that for each of the thee models below the output is the same.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>wts=[.5*np.ones((2,4)),.5*np.ones((1,4)),np.array([0,1,0,0])]\n\nmodel1 = Sequential()\nmodel1.add(LSTM(1, return_sequences=True,input_shape=(1,2),stateful=True,batch_size=1))\nmodel1.layers[0].set_weights(wts)\nyp1=model1.predict(np.array([[[0,0]]]))\nyp2=model1.predict(np.array([[[0,1]]]))\nyp3=model1.predict(np.array([[[0,0]]]))\nyp4=model1.predict(np.array([[[1,1]]]))\nprint(\"model1:\",np.hstack((yp1[:,0,0],yp2[:,0,0],yp3[:,0,0],yp4[:,0,0])))\n\nmodel2 = Sequential()\nmodel2.add(LSTM(1, return_sequences=True,input_shape=(2,2),stateful=True,batch_size=1))\nmodel2.layers[0].set_weights(wts)\nyp1=model2.predict(np.array([[[0,0],[0,1]]]))\nyp2=model2.predict(np.array([[[0,0],[1,1]]]))\nprint(\"model2:\",np.hstack((yp1.reshape((-1)),yp2.reshape((-1)))))\n\nmodel3 = Sequential()\nmodel3.add(LSTM(1, return_sequences=True,input_shape=(4,2),stateful=False,batch_size=1))\nmodel3.layers[0].set_weights(wts)\nyp=model3.predict(np.array([[[0,0],[0,1],[0,0],[1,1]]]))\nprint(\"model3:\",yp.reshape(-1))\n<\/code><\/pre>\n\n<blockquote>\n  <p>model1: [0.         0.16222607 0.12174767 0.46444273]<\/p>\n  \n  <p>model2: [0.         0.16222607 0.12174767 0.46444273]<\/p>\n  \n  <p>model3: [0.         0.16222607 0.12174767 0.46444273]<\/p>\n<\/blockquote>\n\n<p>I hope this makes sense and will help others.<\/p>\n\n<p>Thanks<\/p>\n"},{"AnswerId":"57848198","CreationDate":"2019-09-09T05:15:35.627","ParentId":null,"OwnerUserId":"10371987","Title":null,"Body":"<p>Your input is an array of <code>shape=(4,4,2)<\/code>, where the first 4 is batch size (4 samples), second 4 is time_step, and the last 2 is the input_dim of each time_step.  <\/p>\n\n<h2>Difference between sample step and time step in LSTM<\/h2>\n\n<p>Each sample can have multiple time steps, or even different time steps between samples is also valid.  <\/p>\n\n<h2>Why do I not get yp[0,:,:] equal yp[:,0,:]?<\/h2>\n\n<p>You need to understand that RNN block such as LSTM\/GRU can <strong>catch the dependency between time steps.<\/strong> More concretely, there are different gates in LSTM that can carry the information from previous time step to current time step.<br>\nIn your example, when the first sample <code>[[0,0],[0,1],[0,0],[1,1]]<\/code> get into the model, the LSTM block actually <strong>process the time step one after one, from [0,0]->[0,1]->...[1,1].<\/strong> This is how forward-propagation in RNN is done. Therefore, the output state of different time step in different sample should be different.  <\/p>\n\n<p>Here is a <a href=\"https:\/\/towardsdatascience.com\/illustrated-guide-to-lstms-and-gru-s-a-step-by-step-explanation-44e9eb85bf21\" rel=\"nofollow noreferrer\">blog<\/a> gives good explaination of LSTM\/GRU with visualisation.<\/p>\n"}]}
+{"QuestionId":57846553,"AnswerCount":1,"Tags":"<tensorflow><gradient-descent><eager-execution>","CreationDate":"2019-09-08T23:55:10.010","AcceptedAnswerId":57886724.0,"OwnerUserId":4028648.0,"Title":"TensorFlow: How can I inspect gradients and weights in eager execution?","Body":"<p>I am using TensorFlow 1.12 in eager execution, and I want to inspect the values of my gradients and my weights at different points during training for debugging purposes. <a href=\"https:\/\/stackoverflow.com\/a\/48928133\/4028648\">This answer<\/a> uses TensorBoard to get nice graphs of weight and gradient distribution over epochs, which is what I would like. However, when I use <a href=\"https:\/\/keras.io\/callbacks\/#tensorboard\" rel=\"nofollow noreferrer\">Keras' TensorBoard callback<\/a>, I get this:<\/p>\n\n<pre><code>WARNING:tensorflow:Weight and gradient histograms not supported for eagerexecution, setting `histogram_freq` to `0`.\n<\/code><\/pre>\n\n<p>In other words, this is not compatible with eager execution. Is there any other way to print gradients and\/or weigths? Most non-TensorBoard answers seem to rely on graph-based execution.<\/p>\n","answers":[{"AnswerId":"57886724","CreationDate":"2019-09-11T10:02:38.790","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>In eager execution, you can directly print the weights. As for the gradients, you can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/GradientTape\" rel=\"nofollow noreferrer\">tf.GradientTape<\/a> to get the gradients of the loss function with respect to some weights. Here is an example showing how to print gradients and weights:<\/p>\n\n<pre><code>import tensorflow as tf\n\ntf.enable_eager_execution()\n\nx = tf.ones(shape=(4, 3))\ny = tf.ones(shape=(4, 1))\ndense = tf.layers.Dense(1)\n\n# Print gradients\nwith tf.GradientTape() as t:\n    h = dense(x)\n    loss = tf.losses.mean_squared_error(y, h)\ngradients = t.gradient(loss, dense.kernel)\nprint('Gradients: ', gradients)\n\n# Print weights\nweights = dense.get_weights()\nprint('Weights: ', weights)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57847379,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch><yolo>","CreationDate":"2019-09-09T03:02:48.770","AcceptedAnswerId":null,"OwnerUserId":7371707.0,"Title":"What does `(pred[:, 2:4] > min_wh).all(1)` do in YOLO (Deep Learning)?","Body":"<p>I am using yoloV3 under pytorch. I encountered this code <code>(pred[:, 2:4] &gt; min_wh).all(1)<\/code> and do not know its function. Can anyone help? Thanks!<\/p>\n\n<p>My concern is the use of <code>().all(1)<\/code>. I know <code>.all()<\/code> or <code>.any()<\/code>, but do not know <code>.all(1)<\/code>. Please explain <code>.all(1)<\/code>, thanks.<\/p>\n","answers":[{"AnswerId":"57868805","CreationDate":"2019-09-10T10:25:20.900","ParentId":null,"OwnerUserId":"10478535","Title":null,"Body":"<p>According to the documentation <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.BoolTensor.all\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.BoolTensor.all<\/a><\/p>\n\n<p>there is <code>all(dim)<\/code> with first parameter <code>dim<\/code>. Which means it is the same as <code>all()<\/code>, but only on the selected dimension. It is basically used to select predictions (<strong>rows<\/strong>) on which <strong>both<\/strong> width and height are bigger than <code>min_wh<\/code>. <\/p>\n\n<p>In your case <code>pred<\/code> has shape <code>(number_of_predictions, 7)<\/code> or <\/p>\n\n<pre><code>[\n[x, y, w, h, object_conf, class_conf, class],\n[x, y, w, h, object_conf, class_conf, class],\n...\n]\n<\/code><\/pre>\n\n<p>After <code>pred[:, 2:4] &gt; min_wh<\/code>  the result will be something like<\/p>\n\n<pre><code>[\n[True, False],\n[True, True],\n[False, False],\n...\n]\n<\/code><\/pre>\n\n<p>we want to select rows which have both width and height bigger than <code>min_wh<\/code>, thus we need to use <code>all(1)<\/code>.<\/p>\n\n<p>Because<\/p>\n\n<p><code>all()<\/code> will give you <code>True<\/code> if all the elements are <code>True<\/code>, <code>False<\/code> otherwise<\/p>\n\n<p><code>all(0)<\/code> will give you tensor with shape <code>(2,)<\/code>, for example <code>[True, False]<\/code>. First element will be <code>True<\/code> if all elements in the first <strong>column<\/strong> are <code>True<\/code>, <code>False<\/code> otherwise. Second element will be <code>True<\/code> if all elements in the second <strong>column<\/strong> are <code>True<\/code>, <code>False<\/code> otherwise.<\/p>\n\n<p>and <code>all(1)<\/code> will give you tensor with shape <code>(number_of_predictions,)<\/code>,\nwhere each element is <code>True<\/code> only if all the elements in the row are <code>True<\/code>.<\/p>\n"}]}
+{"QuestionId":57847484,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-09-09T03:20:32.803","AcceptedAnswerId":57847958.0,"OwnerUserId":9066532.0,"Title":"Why does torch.gt function turn requires_grad into False?","Body":"<p>The <code>requires_grad<\/code> of tensor <code>b<\/code> and <code>c<\/code> are <code>True<\/code>. But the <code>requires_grad<\/code> of tensor <code>d<\/code> is <code>False<\/code>. I am curious why this change happens because all the <code>requires_grad<\/code> of inputs are True.<\/p>\n\n<p>However, the <code>requires_grad<\/code> of tensor <code>e<\/code> is True. I can still do <code>backward()<\/code> on <code>e<\/code>. But is there an error in this way?<\/p>\n\n<p>I am using Python3.7 and Pytorch1.1.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nimport torch.nn as nn\n\nnet = nn.Conv2d(1, 1, 3, padding=1)\na = torch.randn(1, 1, 10, 10)\nb = net(a)\nc = net(b)\n\nd = torch.gt(b, c)\n\ne = b - c\ne[e &gt; 0] = 1.0\ne[e &lt; 0] = 0.0\n<\/code><\/pre>\n","answers":[{"AnswerId":"57847958","CreationDate":"2019-09-09T04:40:15.073","ParentId":null,"OwnerUserId":"9828197","Title":null,"Body":"<p>I assume this is because you cannot take a gradient of greater than operation. The return type is boolean:<\/p>\n\n<pre><code>&gt;&gt;&gt; torch.gt(torch.tensor([[1, 2], [3, 4]]), torch.tensor([[1, 1], [4, 4]]))\ntensor([[False, True], [False, False]])\n<\/code><\/pre>\n\n<p>Whereas minus or other arithmetic operation returns another numeral.<\/p>\n"}]}
+{"QuestionId":57848153,"AnswerCount":0,"Tags":"<python><tensorflow><keras><computer-vision><image-segmentation>","CreationDate":"2019-09-09T05:09:55.667","AcceptedAnswerId":null,"OwnerUserId":8763097.0,"Title":"Keras: Multi-label classification with U-net architcture for effective background\/foreground image segmentation","Body":"<p>I have a series of cell images (+ masks) similar to 'D' in the following picture:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/43RM9.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/43RM9.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>However, I'm trying to set up a two-label classification network using the Unet architecture where I explicitly label the following two components of an image:<\/p>\n\n<p>1) The Cells, and<\/p>\n\n<p>2) The Background (or the borders of cells)<\/p>\n\n<p>My reasoning for having a dedicated label for the background is as follows: my original images are 2000x3000 in size, and I want to chunk the images into the Unet as 512x512 blocks. However, when training the original Unet in this manner, I find that because a fair amount of the 512x512 blocks don't contain any cells (they're mostly Background, so the accompanying masks are empty), the loss for every epoch hovers at 0.4 and the network has a lot of trouble training. This is in spite of the presence of good 512x512 blocks that do contain good cell + mask layers.<\/p>\n\n<p>My thinking is that, to get the Unet system to work with 2 labels and segment the images properly, I need to provide two things:\n1) Another set of masks for the background itself so the network can positively identify the background.\n2) An output modifier for the Unet Keras code itself so it gives out 2 labels instead of one.<\/p>\n\n<p>However, please correct me if I am wrong on 1 assumption: the Unet architecture will work better to segment images if provided with a series of negative controls (background) as well as positive controls (the cells).<\/p>\n\n<p>Here's the Keras Model based off <a href=\"https:\/\/www.kaggle.com\/keegil\/keras-u-net-starter-lb-0-277\" rel=\"nofollow noreferrer\">the Kaggle starter tutorial model<\/a>:<\/p>\n\n<pre><code>inputs = Input((IMG_HEIGHT, IMG_WIDTH, IMG_CHANNELS))\n    s = Lambda(lambda x: x \/ 255) (inputs)\n\n    c1 = Conv2D(16, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (s)\n    c1 = Dropout(0.1) (c1)\n    c1 = Conv2D(16, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (c1)\n    p1 = MaxPooling2D((2, 2)) (c1)\n\n    c2 = Conv2D(32, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (p1)\n    c2 = Dropout(0.1) (c2)\n    c2 = Conv2D(32, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (c2)\n    p2 = MaxPooling2D((2, 2)) (c2)\n\n    c3 = Conv2D(64, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (p2)\n    c3 = Dropout(0.2) (c3)\n    c3 = Conv2D(64, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (c3)\n    p3 = MaxPooling2D((2, 2)) (c3)\n\n    c4 = Conv2D(128, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (p3)\n    c4 = Dropout(0.2) (c4)\n    c4 = Conv2D(128, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (c4)\n    p4 = MaxPooling2D(pool_size=(2, 2)) (c4)\n\n    c5 = Conv2D(256, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (p4)\n    c5 = Dropout(0.3) (c5)\n    c5 = Conv2D(256, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (c5)\n\n    u6 = Conv2DTranspose(128, (2, 2), strides=(2, 2), padding='same') (c5)\n    u6 = concatenate([u6, c4])\n    c6 = Conv2D(128, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (u6)\n    c6 = Dropout(0.2) (c6)\n    c6 = Conv2D(128, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (c6)\n\n    u7 = Conv2DTranspose(64, (2, 2), strides=(2, 2), padding='same') (c6)\n    u7 = concatenate([u7, c3])\n    c7 = Conv2D(64, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (u7)\n    c7 = Dropout(0.2) (c7)\n    c7 = Conv2D(64, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (c7)\n\n    u8 = Conv2DTranspose(32, (2, 2), strides=(2, 2), padding='same') (c7)\n    u8 = concatenate([u8, c2])\n    c8 = Conv2D(32, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (u8)\n    c8 = Dropout(0.1) (c8)\n    c8 = Conv2D(32, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (c8)\n\n    u9 = Conv2DTranspose(16, (2, 2), strides=(2, 2), padding='same') (c8)\n    u9 = concatenate([u9, c1], axis=3)\n    c9 = Conv2D(16, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (u9)\n    c9 = Dropout(0.1) (c9)\n    c9 = Conv2D(16, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (c9)\n\n    outputs = Conv2D(1, (1, 1), activation='sigmoid') (c9)\n\n    model = Model(inputs=[inputs], outputs=[outputs])\n    model.compile(optimizer='adam', loss='binary_crossentropy', metrics=[mean_iou])\n    model.summary()\n<\/code><\/pre>\n\n<p><strong>1) Getting the Unet to output 2 labels:<\/strong><\/p>\n\n<p>Should I change the output layer from <\/p>\n\n<pre><code>outputs = Conv2D(1, (1, 1), activation='sigmoid') (c9)\n<\/code><\/pre>\n\n<p>to:<\/p>\n\n<pre><code>outputs = Conv2D(2, (1, 1), activation='sigmoid') (c9)\n<\/code><\/pre>\n\n<p><strong>2) How do I structure the input array for the masks<\/strong><\/p>\n\n<p>For the mask labels, I have a 3D numpy array in the shape of: <code>(# images, height, width)<\/code> with <code>dtype=uint8<\/code>. Every mask is binarised to 0 or 255 for a single label (eg the cell). Do I create another mask array with the same shape and pass that into the network as another training layer, or do I label the background as a separate number in the original mask image? To simplify: how do I train 2 types of labels in the Unet architecture? <\/p>\n","answers":[]}
+{"QuestionId":57848809,"AnswerCount":0,"Tags":"<keras><nlp><text-classification><word-embedding>","CreationDate":"2019-09-09T06:27:35.783","AcceptedAnswerId":null,"OwnerUserId":12039987.0,"Title":"How to write adversarial and virtual adversarial custom loss function in keras","Body":"<p>I am trying to write a keras implementation of this paper <a href=\"https:\/\/arxiv.org\/abs\/1605.07725\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1605.07725<\/a> on using adversarial and virtual adversarial loss fn to improve the performance on text classification.<\/p>\n\n<p>But my custom loss function is not working.<\/p>\n\n<pre><code>def custom_loss(y_true,y_pred,embedding_model):\n    embedding = embedding_model.output\n    cls_loss = K.categorical_crossentropy(y_true,y_pred)\n    grad, = tf.gradients(\n            cls_loss,\n            embedding)\n    grad = tf.stop_gradient(grad)\n    perturb = scale_l2(grad,norm )\n    embedding_perturbated = embedding + perturb\n    adv_logits = network(embedding_perturbated)\n    return cls_loss + K.categorical_crossentropy(y_true,adv_logits)\n<\/code><\/pre>\n\n<p>grad, = tf.gradients(cls_loss,embedding) returns None.<\/p>\n\n<p>How do i implement both the cost functions in keras<\/p>\n","answers":[]}
+{"QuestionId":57848855,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-09T06:31:38.237","AcceptedAnswerId":null,"OwnerUserId":11947897.0,"Title":"How to print tensorflow dense layer output","Body":"<p>I would like to print the output of my hidden layers that look like:<\/p>\n\n<pre><code>net = tf.layers.dense(s, 64, activation=tf.nn.leaky_relu, name='l2', trainable=trainable)  \nmid = tf.layers.dense(net, 64, activation=tf.nn.leaky_relu, name='l3', trainable=trainable)\na = tf.layers.dense(mid, 3, activation=tf.nn.softmax, name='la', trainable=trainable)  \n<\/code><\/pre>\n\n<p>I tried <code>sess.run()<\/code> but I don't know which arguments to give, nor how. I except the output lenght to be <code>input*units+output<\/code>.<\/p>\n\n<p>How can I do this? <\/p>\n","answers":[]}
+{"QuestionId":57849267,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-09T07:07:35.790","AcceptedAnswerId":null,"OwnerUserId":260826.0,"Title":"TF.data migration to dataset.interleave","Body":"<p>TensorFlow nightly: <code>1.15.0-dev20190730<\/code><\/p>\n\n<pre><code>filenames = tf.gfile.Glob(data_files_pattern)\ndataset = tf.data.Dataset.from_tensor_slices(filenames).repeat()\n\ndef _read_fn(f):\n  return tf.data.TFRecordDataset(f)\n\ndataset = dataset.apply(tf.data.experimental.parallel_interleave(\n    map_func=_read_fn,\n    cycle_length=CYCLE_LENGTH,\n    block_length=BLOCK_LENGTH,\n    sloppy=True,\n    buffer_output_elements=BUFFER_OUTPUT_ELEMENTS,\n    prefetch_input_elements=BUFFER_INPUT_ELEMENTS))\ndataset = dataset.batch(BATCH_SIZE, drop_remainder=False)\ndataset = dataset.prefetch(PREFETCH)\nreturn dataset\n<\/code><\/pre>\n\n<p>I get the following warning:<\/p>\n\n<pre><code>WARNING:tensorflow:From sample.py:35: parallel_interleave (from tensorflow.python.data.experimental.ops.interleave_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.data.Dataset.interleave(map_func, cycle_length, block_length, num_parallel_calls=tf.data.experimental.AUTOTUNE)` instead. If sloppy execution is desired, use `tf.data.Options.experimental_determinstic`.\nW0909 06:50:51.144233 140600866592512 deprecation.py:323] From sample.py:35: parallel_interleave (from tensorflow.python.data.experimental.ops.interleave_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.data.Dataset.interleave(map_func, cycle_length, block_length, num_parallel_calls=tf.data.experimental.AUTOTUNE)` instead. If sloppy execution is desired, use `tf.data.Options.experimental_determinstic`.\n<\/code><\/pre>\n\n<p>When I migrate to avoid warning, my reads are slower and CPU utilization decreases:<\/p>\n\n<pre><code>filenames = tf.gfile.Glob(data_files_pattern)\ndataset = tf.data.Dataset.from_tensor_slices(filenames).repeat()\n\ndef _read_fn(f):\n   return tf.data.TFRecordDataset(f)\n\noptions = tf.data.Options()\noptions.experimental_deterministic = True\ndataset = dataset.interleave(\n    map_func=_read_fn,\n    cycle_length=CYCLE_LENGTH,\n    block_length=BLOCK_LENGTH,      \n    num_parallel_calls=tf.data.experimental.AUTOTUNE).with_options(options)\ndataset = dataset.batch(BATCH_SIZE, drop_remainder=False)\ndataset = dataset.prefetch(PREFETCH)\nreturn dataset\n<\/code><\/pre>\n\n<p>Am I migrating correctly?<\/p>\n","answers":[{"AnswerId":"58090486","CreationDate":"2019-09-25T02:46:28.167","ParentId":null,"OwnerUserId":"401884","Title":null,"Body":"<p>The problem is that you are comparing sloppy (non-deterministic) <code>parallel_interleave<\/code> to deterministic <code>interleave<\/code>. You set <code>sloppy=True<\/code> for <code>parallel_interleave<\/code>, so for a correct migration you need to set<\/p>\n\n<pre><code>options.experimental_deterministic = False\n<\/code><\/pre>\n\n<p>for <code>interleave<\/code>.<\/p>\n"}]}
+{"QuestionId":57849289,"AnswerCount":1,"Tags":"<pytorch><torchvision>","CreationDate":"2019-09-09T07:09:13.273","AcceptedAnswerId":57855051.0,"OwnerUserId":5380809.0,"Title":"Torchvision normalize - how it operates on tuple of means\/sds?","Body":"<p>I don't understand how this transform works from torchvision. Ultimately I want to build a custom normalize class so I need to figure out how this works first.<\/p>\n\n<p>Here in the docs it describes the init like this:<\/p>\n\n<pre><code>def __init__(self, mean, std, inplace=False):\n        self.mean = mean\n        self.std = std\n        self.inplace = inplace\n<\/code><\/pre>\n\n<p>And when I pass these parameters usually (not custom class) I pass them as a list or tuple for each channel:<\/p>\n\n<p><code>transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])<\/code><\/p>\n\n<p>But if I look at the call:<\/p>\n\n<p><code>return F.normalize(tensor, self.mean, self.std, self.inplace)<\/code><\/p>\n\n<p>All this passes the tuple to is F.normalize() which only accepts a single value for the p parameter.<\/p>\n\n<p>The class must iterate through the channels somehow to allow this to be implemented but how does it do this and how can I implement it in custom class?<\/p>\n\n<p>Based on <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/data_loading_tutorial.html\" rel=\"nofollow noreferrer\">this tutorial<\/a>, I would describe it like this:<\/p>\n\n<pre><code>class Normalize(object):\n    \"\"\"Convert ndarrays in sample to Tensors.\"\"\"\n\n    def __init__(self, mean, std, inplace=False):\n        self.mean = mean\n        self.std = std\n        self.inplace = inplace\n\n    def __call__(self, sample):\n        image, landmarks = sample['image'], sample['landmarks']\n        return {'image': F.normalize(image, self.mean, self.std, self.inplace),\n                'landmarks': landmarks}\n<\/code><\/pre>\n\n<p>But this does not work because it does not go through each channel.<\/p>\n","answers":[{"AnswerId":"57855051","CreationDate":"2019-09-09T13:29:26.230","ParentId":null,"OwnerUserId":"4892874","Title":null,"Body":"<p>The <code>normalize<\/code> function called in there is this one <a href=\"https:\/\/github.com\/pytorch\/vision\/blob\/master\/torchvision\/transforms\/functional.py#L191\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/vision\/blob\/master\/torchvision\/transforms\/functional.py#L191<\/a><\/p>\n\n<p>The input is a tensor of shape <code>(C, H, W)<\/code> and <code>mean<\/code> and <code>std<\/code> can be sequences, that are internally converted to tensors. The normalization is done through broadcasting <a href=\"https:\/\/github.com\/pytorch\/vision\/blob\/master\/torchvision\/transforms\/functional.py#L217\" rel=\"nofollow noreferrer\">in this way<\/a>:<\/p>\n\n<p><code>tensor.sub_(mean[:, None, None]).div_(std[:, None, None])<\/code><\/p>\n"}]}
+{"QuestionId":57849392,"AnswerCount":0,"Tags":"<python><keras><reinforcement-learning>","CreationDate":"2019-09-09T07:18:26.143","AcceptedAnswerId":null,"OwnerUserId":12037158.0,"Title":"Reinforcement Learning - not converging","Body":"<p>I build a simple traffic light simulation in pygame and my goal is to have the agent choose between different traffic light modes to minimize the waiting time for the cars (the score is calculated by adding up the time spend with 0 velocity for each car between the frames). However the model does not learn the expected behaviour. <\/p>\n\n<p>I tried several discount values (0.1, 0.5 ,0.9) since i am not sure whether it is necessary to look far ahead into the future to solve this particular task.<\/p>\n\n<p>I trained the agent for about 200k steps and while it preferred certain actions(the two which make one whole street green, which makes sense to me) it does not seem to choose them at the right time and it looks more like it randomly switches between the two. The agent does not care about the cars on the street at all.<\/p>\n\n<p>The agent is also able to learn simple strageties in the same environment like just making all lights red if its goal is to maximize waiting time.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nfrom keras.optimizers import adam\nimport numpy as np\nimport tensorflow as tf\nimport keras\nimport matplotlib.pyplot as plt\n#from utils import plotlearning\n\nbatch_size = 10\ngamma = 0.9\n\nepsilon = 1\nepsilon_decay = 0.999\nepsilon_end = 0.001\ntarget_update = 200\nmemory_size = 100000\nLR = 0.001\n\nnum_episodes = 1000\ndecisions_per_episode = 200\nframes_per_step = 10\n\nclass ReplayBuffer(object):\n    def __init__(self, max_size, input_shape, n_actions):\n        self.mem_size = max_size\n        self.mem_cntr = 0\n        self.state_memory = np.zeros((self.mem_size, 106, 60, 4))\n        self.new_state_memory = np.zeros((self.mem_size, 106, 60, 4))\n        dtype = np.int8\n        self.action_memory = np.zeros((self.mem_size), dtype=dtype)\n        self.reward_memory = np.zeros(self.mem_size)\n\n    def store_transition(self, state, action, reward, state_):\n        index_num = int(self.mem_cntr % self.mem_size)\n        self.new_state_memory[index_num] = state_\n        self.state_memory[index_num] = state\n        self.reward_memory[index_num] = reward\n        self.action_memory[index_num] = action\n        self.mem_cntr +=1\n\n    def sample_buffer(self, batch_size):\n        max_mem = min(self.mem_cntr, self.mem_size)\n        batch = np.random.choice(max_mem, batch_size)\n\n        states = self.state_memory[batch]\n        states_ = self.new_state_memory[batch]\n        rewards = self.reward_memory[batch]\n        actions = self.action_memory[batch]\n\n        return states, actions, rewards, states_\n\n\ndef build_dqn(n_actions):\n    model = tf.keras.models.Sequential()\n\n\n    #convolution\n    model.add(tf.keras.layers.Conv2D(16, kernel_size=8, strides=(4, 4), activation=\"relu\", input_shape=(106,60,4)))    \n    model.add(tf.keras.layers.Conv2D(32, kernel_size=4, strides=(2, 2), activation=\"relu\"))\n\n    #flatten\n    model.add(tf.keras.layers.Flatten())\n\n    #dense\n    model.add(tf.keras.layers.Dense(256, activation=tf.nn.relu))\n\n    #output\n    model.add(tf.keras.layers.Dense(n_actions))\n\n\n\n    model.compile(loss='mse',\n              optimizer=\"adam\")\n\n    return model\n\n\n\n\n\n\n\n\nclass Agent(object):\n    def __init__(self, n_actions, input_dims, eps, eps_dec, eps_end, mem_size=1000, fname='dqn_model_positive_reward_0.5gamma.h5'):\n        self.action_space = [i for i in range(n_actions)]\n        self.n_actions = n_actions\n        self.model_file = fname\n\n        self.memory = ReplayBuffer(mem_size, input_dims, n_actions)\n        self.q_eval = build_dqn(n_actions)\n        self.target_model = build_dqn(n_actions)\n        self.epsilon = eps\n        self.epsilon_dec = eps_dec\n        self.epsilon_end = eps_end\n\n\n    def remember(self, state, action, reward, new_state):\n        self.memory.store_transition(state, action, reward, new_state)\n\n    def choose_action(self, state):\n        state = state\n        rand = np.random.random()\n\n        if rand &lt; self.epsilon:\n            action = np.random.choice(self.action_space)\n        else:\n            output = self.q_eval.predict(state)\n            print(output)\n            action = np.where(output == np.amax(output))\n            action = action[1][0]\n            print(action)\n        return action\n    def copy_nn(self):\n        self.target_model.set_weights(self.q_eval.get_weights()) \n\n    def learn(self):\n\n        if self.memory.mem_cntr &lt; batch_size:\n            return\n\n        state, action, reward, new_state = self.memory.sample_buffer(batch_size)\n\n        q_eval = self.q_eval.predict(state)\n        q_next = self.target_model.predict(new_state)\n\n        q_target = q_eval.copy()\n\n        batch_index = np.arange(batch_size, dtype = np.int32)\n\n        q_target[batch_index, action] = reward + gamma*np.max(q_next, axis=1)\n\n        _ = self.q_eval.fit(state, q_target, verbose = 0)\n\n        self.epsilon = self.epsilon*self.epsilon_dec if self.epsilon &gt; self.epsilon_end else self.epsilon_end\n\n    def save_model(self):\n        self.q_eval.save(self.model_file)\n\n    def load_model(self):\n        self.q_eval = load_model(self.model_file)\n\n\nenv = Environment()\n\n\n\nagent = Agent(input_dims = 57600, n_actions = 7, eps = 1, eps_dec= 0.1**(1\/(num_episodes*decisions_per_episode)), eps_end = 0.001)\n\nscores = []\neps_history = []\nlast_action = 0\n\n\n\nfor i in range(num_episodes):\n    env.reset()\n    old_score = 0\n    state = env.get_state(random.randint(0,6))\n    decisions = 0\n    env.spawn_delay = random.randint(20,200)\n\n    if i % target_update == 0:    \n        agent.copy_nn()\n\n    #refill active_spawns list\n    env.active_spawns =[]\n    for element in range(21):\n        env.active_spawns.append(random.randint(0,7))\n\n\n\n    while decisions&lt;decisions_per_episode:\n        decisions+=1\n        action = agent.choose_action(state)\n\n\n        for _ in range(frames_per_step):\n\n            env.step(action)\n\n        old_action = action\n\n        if len(env.all_cars) != 0:\n            reward = env.score - old_score\n        else:\n            reward = 0\n\n\n        state_ = env.get_state(old_action)\n        agent.remember(state, action, reward, state_)\n        state = state_\n        old_score = env.score\n        agent.learn()\n    eps_history.append(epsilon)\n    scores.append(env.score)\n\n\n\n    if i % 10 == 0 and i &gt; 0:\n        agent.save_model()\n\n\n\npygame.quit()\n\n<\/code><\/pre>\n\n<p>This is the part that trains the model. I left the environment part out since it should not be of big importance. If this part is needed just let me know.<\/p>\n","answers":[]}
+{"QuestionId":57849535,"AnswerCount":1,"Tags":"<python><gpu><pytorch>","CreationDate":"2019-09-09T07:28:27.113","AcceptedAnswerId":57849618.0,"OwnerUserId":9066532.0,"Title":"Would it possible to use all memory of GPUs with one model?","Body":"<p>There is a <code>model<\/code> and two GPUs. I put the <code>model<\/code> on GPU with <code>model.cuda()<\/code>. If I passed a big image to the <code>model<\/code>, it allocated all memory of GPU0 and then it raised <code>CUDA out of memory<\/code> error without allocating any memory of GPU1.<\/p>\n\n<p>Because there is only one image every <code>forward()<\/code>, I can not use such <code>torch.nn.DataParallel<\/code> things to split input.<\/p>\n\n<p>Is there any way to use all the memory of GPUs when passing a image to the <code>model<\/code>?<\/p>\n\n<p>I'm using Python3.7 and Pytorch1.1.<\/p>\n","answers":[{"AnswerId":"57849618","CreationDate":"2019-09-09T07:34:58.317","ParentId":null,"OwnerUserId":"7701908","Title":null,"Body":"<p>you can split your model into two submodule. \nlike this:<\/p>\n\n<pre><code>class MyModel(nn.Module):\n    def __init__(self, split_gpus):\n        self.large_submodule1 = ...\n        self.large_submodule2 = ...\n\n        self.split_gpus = split_gpus\n        if split_gpus:\n            self.large_submodule1.cuda(0)\n            self.large_submodule1.cuda(1)\n\n    def forward(self, x):\n        x = self.large_submodule1(x)\n        if split_gpus:\n            x = x.cuda(1) # P2P GPU transfer\n        return self.large_submodule2(x)\n<\/code><\/pre>\n\n<p>from the <a href=\"https:\/\/discuss.pytorch.org\/t\/split-single-model-in-multiple-gpus\/13239\/2?u=alpha\" rel=\"nofollow noreferrer\">pytorch discuss<\/a><\/p>\n"}]}
+{"QuestionId":57849831,"AnswerCount":2,"Tags":"<python><django><tensorflow><keras><tf.keras>","CreationDate":"2019-09-09T07:51:59.293","AcceptedAnswerId":58264389.0,"OwnerUserId":3286607.0,"Title":"How to predict in multiple and simultanous Keras classifier sessions in Django?","Body":"<p>I know similar questions have been asked before and I have read all of them but none solved my problem.<\/p>\n\n<p>I have a Django project in which I am using <code>CNNSequenceClassifier<\/code> from <code>sequence_classifiers<\/code> which is a <code>Keras<\/code> model. The model files have been fit before and saved to particular destinations from which I again load them and predict what I want.<\/p>\n\n<pre><code>clf = CNNSequenceClassifier(epochs=2)\n<\/code><\/pre>\n\n<p>When I load the model I do this, due to the suggestions I found in searches, which is using the global model before loading the model, and then the other two lines:<\/p>\n\n<pre><code>global clf\nclf = pickle.load(open(modelfilenameandpath, \"rb\"))\nglobal graph\ngraph = tf.get_default_graph()\n<\/code><\/pre>\n\n<p>and before predicting I use graph.as_default()<\/p>\n\n<pre><code>with graph.as_default():\n    probs = clf.predict_proba(the_new_vecs)\nK.clear_session()\n<\/code><\/pre>\n\n<p>I put <code>K.clear_session()<\/code> because I predict in a for loop and sometimes the next item of the for loop's prediction gets mixed up with the last one and raises tensorflow errors. But clearly <code>K.clear_session()<\/code> clears the session and makes it easy for the new item's prediction to work fine.<\/p>\n\n<p>The problem is in my <code>views.py<\/code> I have two functions which trigger prediction. And sometimes I need to use both simultanously. But since the probject is using Tensorflow Backend, only one session is defined and the predictions of these two functions get mixed up together. <code>K.clear_session()<\/code> does not help here because it was only for that particular session and it does not expect new stuff coming in the same session.<\/p>\n\n<p>I really do not know how to make the functions understand whenever they start loading the model and\/or want to predict it, start a new and independent session so that nothing gets mixed up.\nI have seen codes like:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n<\/code><\/pre>\n\n<p>but I don't know where to put the <code>probs = clf.predict_proba(the_new_vecs)<\/code>. It seems all of them require the action you want to be done in the session to be passed as an argument to <code>sess.run()<\/code> which does not work for me or at least I do not know how to make it work. I want the predict to be done in a new session and after that the values of <code>probs<\/code> are really important. Also <code>global_variables_initializer()<\/code> seems to change the value of all my variables.<\/p>\n\n<p>I have also tried <code>K.set_session()<\/code> or <code>K.get_session()<\/code> but did not work.<\/p>\n\n<p>To sum up, I think the reason I am stuck is that I do not know how to use sessions for my purpose. I do not know where to use what code?!<\/p>\n\n<p>Please help!<\/p>\n","answers":[{"AnswerId":"58264389","CreationDate":"2019-10-07T06:24:11.400","ParentId":null,"OwnerUserId":"8366805","Title":null,"Body":"<p>My understanding of <code>tensorflow<\/code> graphs and sessions is:<\/p>\n\n<p>A <code>tensorflow<\/code> graph hosts operations, <code>placeholder<\/code>s and <code>Variable<\/code>s. A <code>tensorflow<\/code> graph lives inside a <code>tensorflow<\/code> session (that's why to save a trained model using <code>tensorflow.train.Saver<\/code>, you need something like <code>Saver.save(sess, the graph)<\/code>).  <\/p>\n\n<p>Here is a simple example to help you make sense of the relation between a <code>keras<\/code> model and a <code>tensorflow<\/code> graph:<\/p>\n\n<pre><code>import tensorflow as tf \nfrom keras.layers import Input, Dense\nfrom keras.models import Model \n\ntf.reset_default_graph()\ngraph_1, graph_2 = tf.Graph(), tf.Graph()\nwith graph_1.as_default():\n    x_in = Input(shape=(1, ), name='in_graph1')\n    pred = Dense(5, name='dense1_graph1')(x_in)\n    pred = Dense(1, name='dense2_graph1')(pred)\n    model = Model(input=x_in, output=pred)\nwith graph_2.as_default():\n    x_in = Input(shape=(1, ), name='in_graph2')\n    pred = Dense(10, name='dense1_graph2')(x_in)\n    pred = Dense(1, name='dense2_graph2')(pred)\n    model = Model(input=x_in, output=pred)\n\n\nwith tf.Session() as sess:\n    default_ops = sess.graph.get_operations()\n\nwith graph_1.as_default():\n    with tf.Session() as sess:\n        one_ops = sess.graph.get_operations()\n\nwith graph_2.as_default():\n    with tf.Session() as sess:      \n        two_ops = sess.graph.get_operations()\n<\/code><\/pre>\n\n<p>As you can see by running the code, <code>default_ops<\/code> is an empty list which means there is no operation in the default graph. <code>one_ops<\/code> is a list of operations of the first <code>keras<\/code> model and <code>two_ops<\/code> is a list of operations of the second <code>keras<\/code> model. <\/p>\n\n<p><strong>So, by using <code>with graph.as_default()<\/code>, a <code>keras<\/code> model can be exclusively embedded in a <code>tensorflow<\/code> graph.<\/strong><\/p>\n\n<p>With this in mind, it becomes easy to load several <code>keras<\/code> models in a single script. I think the example script below will address your confusions:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf \nfrom keras.layers import Input, Dense\nfrom keras.models import Model\nfrom Keras.models import model_from_json\n\ntf.reset_default_graph()\nx = np.linspace(1, 4, 4)\ny = np.random.rand(4)\n\nmodels = {}\ngraph_1, graph_2 = tf.Graph(), tf.Graph()\n\n# graph_1 \nwith graph_1.as_default():\n    x_in = Input(shape=(1, ), name='in_graph1')\n    pred = Dense(5, name='dense1_graph1')(x_in)\n    pred = Dense(1, name='dense2_graph1')(pred)\n    model = Model(input=x_in, output=pred)\n    models['graph_1'] = model\n# graph_2\nwith graph_2.as_default():\n    x_in = Input(shape=(1, ), name='in_graph2')\n    pred = Dense(10, name='dense1_graph2')(x_in)\n    pred = Dense(1, name='dense2_graph2')(pred)\n    model = Model(input=x_in, output=pred)\n    models['graph_2'] = model\n\n# save the two models \nwith tf.Session(graph=graph_1) as sess:\n    with open(\"model_1.json\", \"w\") as source:\n        source.write(models['graph_1'].to_json())\n    models['graph_1'].save_weights(\"weights_1.h5\")\nwith tf.Session(graph=graph_2) as sess:\n    with open(\"model_2.json\", \"w\") as source:\n        source.write(models['graph_2'].to_json())\n    models['graph_2'].save_weights(\"weights_2.h5\")\n\n####################################################\n# play with the model \npred_one, pred_one_reloaded = [], []\npred_two, pred_two_reloaded = [], []\nfor _ in range(10):\n    print(_)\n    if _ % 2 == 0:\n\n        with graph_1.as_default():\n            with tf.Session() as sess:\n                sess.run(tf.global_variables_initializer())\n                pred_one.append(models['graph_1'].predict(x).ravel())\n        with tf.Session() as sess:\n            with open(\"model_1.json\", \"r\") as f:\n                model = model_from_json(f.read())\n            model.load_weights(\"weights_1.h5\")\n            pred_one_reloaded.append(model.predict(x).ravel())\n\n    else:\n        with graph_2.as_default():\n            with tf.Session() as sess:\n                sess.run(tf.global_variables_initializer())\n                pred_two.append(models['graph_2'].predict(x).ravel())\n        with tf.Session() as sess:\n            with open(\"model_2.json\", \"r\") as f:\n                model = model_from_json(f.read())\n            model.load_weights(\"weights_2.h5\")\n            pred_two_reloaded.append(model.predict(x).ravel())\n\npred_one = np.array(pred_one)\npred_one_reloaded = np.array(pred_one_reloaded)\npred_two = np.array(pred_two)\npred_two_reloaded = np.array(pred_two_reloaded)\n\nprint(pred_one)\nprint(pred_one_reloaded)\nprint(pred_two)\nprint(pred_two_reloaded)\n<\/code><\/pre>\n"},{"AnswerId":"58256669","CreationDate":"2019-10-06T11:11:15.887","ParentId":null,"OwnerUserId":"7838040","Title":null,"Body":"<p>You have 2 options at this point, either use multiprocessing library and spin-off a new process the moment you get a request. Be careful here as to using up the memory. <\/p>\n\n<p>The other one (Recommended) is export your TensorFlow model to be served in tensorflow serving, in that you have the option of loading multiple instances of the same model, that way you can serve simultaneously without having to worry about things like memory management or loading and unloading memory.  <\/p>\n"}]}
+{"QuestionId":57850324,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2019-09-09T08:27:57.267","AcceptedAnswerId":null,"OwnerUserId":6402099.0,"Title":"Keras throws: Input 0 is incompatible with layer conv2d_1: expected ndim=4, found ndim=3","Body":"<p><strong>Summary:<\/strong>\nI have data with shape like <code>(10, 300)<\/code>: an array of float numbers with length of 300, and 10 arrays consisting of these 300 numbers. And the batch size is 400.<\/p>\n\n<p>When I try to feed the following model with this data, it throws this exception:<\/p>\n\n<blockquote>\n  <p>ValueError: Input 0 is incompatible with layer conv2d_1: expected ndim=4, found ndim=3<\/p>\n<\/blockquote>\n\n<p>I've tried adding a reshape layer in the following code, between <code>inputs<\/code> and <code>conv_0<\/code>: <code>reshape = Reshape((10, EMBEDDING_DIM, 1))(inputs)<\/code>, but this time it threw exception:<\/p>\n\n<blockquote>\n  <p>Error when checking input: expected input_1 to have 3 dimensions, but got array with shape (400, 1)<\/p>\n<\/blockquote>\n\n<p>I use asserts to be sure the shape of data and it passes all the followings:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>assert len(X) == 400, \"Total number of documents is not 400\"\nassert len(X[0]) == 10, \"Not splitted well\"\nassert len(X[0][0]) == 300, \"Word vector error\"\n<\/code><\/pre>\n\n<p><strong>The model:<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># model\ninputs = Input(batch_shape=(None, 10, EMBEDDING_DIM), dtype='float32')\n\nconv_0 = Conv2D(256, kernel_size=(3, 3), activation='relu')(reshape)\nmaxpool_0 = MaxPooling2D(pool_size=(2, 2), strides=(1, 1))(conv_0)\n\nconv_1 = Conv2D(256, kernel_size=(3, 3), activation='relu')(maxpool_0)\nmaxpool_1 = MaxPooling2D(pool_size=(2, 2), strides=(1, 1))(conv_1)\n\nconv_2 = Conv2D(256, kernel_size=(3, 3), activation='relu')(maxpool_1)\nmaxpool_2 = MaxPooling2D(pool_size=(2, 2), strides=(1, 1))(conv_2)\n\nflatten = Flatten()(maxpool_2)\ndropout = Dropout(0.5)(flatten)\noutputs = Dense(units=len(le.classes_),\n                activation='softmax')(dropout)\n\nmodel = Model(inputs=inputs, outputs=outputs)\n\nadamoptimizer = Adam(lr=0.0001)\n\nmodel.compile(optimizer=adamoptimizer,\n                loss='sparse_categorical_crossentropy', metrics=['accuracy'])\nprint(model.summary())\n<\/code><\/pre>\n\n<p><strong>The summary:<\/strong><\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #\n=================================================================\ninput_1 (InputLayer)         (None, 10, 300)           0\n_________________________________________________________________\nreshape_1 (Reshape)          (None, 10, 300, 1)        0\n_________________________________________________________________\nconv2d_1 (Conv2D)            (None, 8, 298, 256)       2560\n_________________________________________________________________\nmax_pooling2d_1 (MaxPooling2 (None, 7, 297, 256)       0\n_________________________________________________________________\nconv2d_2 (Conv2D)            (None, 5, 295, 256)       590080\n_________________________________________________________________\nmax_pooling2d_2 (MaxPooling2 (None, 4, 294, 256)       0\n_________________________________________________________________\nconv2d_3 (Conv2D)            (None, 2, 292, 256)       590080\n_________________________________________________________________\nmax_pooling2d_3 (MaxPooling2 (None, 1, 291, 256)       0\n_________________________________________________________________\nflatten_1 (Flatten)          (None, 74496)             0\n_________________________________________________________________\ndropout_1 (Dropout)          (None, 74496)             0\n_________________________________________________________________\ndense_1 (Dense)              (None, 21)                1564437\n=================================================================\nTotal params: 2,747,157\nTrainable params: 2,747,157\nNon-trainable params: 0\n_________________________________________________________________\nNone\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57850662,"AnswerCount":0,"Tags":"<python><tensorflow><benchmarking>","CreationDate":"2019-09-09T08:51:42.547","AcceptedAnswerId":null,"OwnerUserId":12040557.0,"Title":"Low Performance while running Tensorflow Benchmark Test on CPU","Body":"<p>I'm running TensorFlow benchmark on a single Motherboard Pro WS C246-ACE (Processor: i9-i9900K, DRAM: 64GB), just trying out the performance without using GPUs. But the performance is quick low. 2.5~3.5 image\/sec. <\/p>\n\n<p>Even after 20000 epochs of training, the max top-1\/top-5 accuracy is about 0.031\/0.094. Learning_rate is still rising in the same speed.<\/p>\n\n<p>Is this the expected performance or did I use the wrong parameters?<\/p>\n\n<p><code>tf.__version__=1.15.0-dev20190821<\/code><\/p>\n\n<p>TensorFlow benchmarks Script\n<a href=\"https:\/\/github.com\/tensorflow\/benchmarks\/tree\/master\/scripts\/tf_cnn_benchmarks\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/benchmarks\/tree\/master\/scripts\/tf_cnn_benchmarks<\/a><\/p>\n\n<p>Below is my script:<\/p>\n\n<pre><code>python tf_cnn_benchmarks.py --num_gpus=1 \\\n  --batch_size=32 --model=resnet50_v1.5 \\\n  --variable_update=parameter_server --device=cpu \\\n  --data_format=NHWC --data_name=imagenet \\\n  --data_dir=F:\\ILSVRC2017_CLS-LOC --save_summaries_steps=10 \\\n  --train_dir=C:\\Users\\C246_0710\\Anaconda3\\envs\\tensorflow1\\benchmarks-master\\scripts\\tf_cnn_benchmarks\\training \\\n  --summary_verbosity=1 --print_training_accuracy \\\n  --benchmark_log_dir=C:\\Users\\C246_0710\\Anaconda3\\envs\\tensorflow1\\benchmarks-master\\scripts\\tf_cnn_benchmarks\\training \\\n  --save_model_steps=1000\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57850942,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-09T09:08:57.527","AcceptedAnswerId":null,"OwnerUserId":3385181.0,"Title":"How to set layout for custom_scalar plugin in tensorflow 2.0","Body":"<p>To set layout for custom_layout plugin in tensorflow 1.x you should use method add_summary in FileWriter class. So the code should look like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with tf.summary.FileWriter(LOGDIR) as writer:\n   layout_summary = summary_lib.custom_scalar_pb(...)\n   writer.add_summary(layout_summary) \n<\/code><\/pre>\n\n<p>In tensorflow 2 add_summary method was removed. How can I set custom_scalar layout now?<\/p>\n","answers":[]}
+{"QuestionId":57851136,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-09T09:21:12.713","AcceptedAnswerId":57851869.0,"OwnerUserId":11954461.0,"Title":"TypeError: Input 'y' of 'Add' Op has type float32 that does not match type int32 of argument 'x' tf.cast is useless","Body":"<p>today, I'm using TensorFlow to do some basic calculation, it suddenly has an error like below.<\/p>\n\n<pre><code>TypeError: Input 'y' of 'Add' Op has type float32 that does not match type int32 of argument 'x'.\n<\/code><\/pre>\n\n<p>I have tried to use tf.cast() to verify the type of the tensor, however, it seems to be useless which is based on that the types of  the four variables of 're', 're2', 're3' and 're4' are just float32 according to the output:<\/p>\n\n<pre><code>re1: Tensor(\"mul:0\", shape=(), dtype=float32)\nre2: Tensor(\"mul_1:0\", shape=(), dtype=float32)\nre3: Tensor(\"mul_4:0\", shape=(), dtype=float32)\nre4: Tensor(\"mul_7:0\", shape=(), dtype=float32)\n<\/code><\/pre>\n\n<p>now, I don't know how to deal with the problem. anyone can give me a hand?\nAny suggestion will be appreciated!!<\/p>\n\n<p>here are some relative codes<\/p>\n\n<pre><code>re=(fc4[i][2]-fc4[i][0])*(fc4[i][3]-fc4[i][1])\nre2=(bbox[i][2] - bbox[i][0]) * (bbox[i][3] - bbox[i][1])\nre3=pro_x * (fc4[i][2] - fc4[i][0]) * pro_y * (fc4[i][3] - fc4[i][1])\nre4=(pro_x*(fc4[i][2]-fc4[i][0])*pro_y*(fc4[i][3]-fc4[i][1]))  print(\"re1:\",re)  \nprint(\"re2:\",re2)\nprint(\"re3:\",re3)  \nprint(\"re4:\",re4)\noverate=re4\/(re+re2-re3)\n<\/code><\/pre>\n\n<p>the complete error is<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/home\/hp\/anaconda3\/envs\/mypython37\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 527, in _apply_op_helper\n    preferred_dtype=default_dtype)\n  File \"\/home\/hp\/anaconda3\/envs\/mypython37\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1224, in internal_convert_to_tensor\n    ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)\n  File \"\/home\/hp\/anaconda3\/envs\/mypython37\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1018, in _TensorTensorConversionFunction\n    (dtype.name, t.dtype.name, str(t)))\nValueError: Tensor conversion requested dtype int32 for Tensor with dtype float32: 'Tensor(\"truediv_3:0\", shape=(), dtype=float32)'\nDuring handling of the above exception, another exception occurred:\nTraceback (most recent call last):\n  File \"&lt;input&gt;\", line 1, in &lt;module&gt;\n  File \"\/snap\/pycharm-community\/147\/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 \"\/snap\/pycharm-community\/147\/helpers\/pydev\/_pydev_imps\/_pydev_execfile.py\", line 18, in execfile\n    exec(compile(contents+\"\\n\", file, 'exec'), glob, loc)\n  File \"\/home\/hp\/Downloads\/GOTURN-Tensorflow-master\/got-10k\/train\/train2dup.py\", line 215, in &lt;module&gt;\n    aot=ao(tracknet.fc4,tracknet.bbox,BATCH_SIZE)\n  File \"\/home\/hp\/Downloads\/GOTURN-Tensorflow-master\/got-10k\/train\/train2dup.py\", line 168, in ao\n    sum=tf.add(sum,overate)\n  File \"\/home\/hp\/anaconda3\/envs\/mypython37\/lib\/python3.7\/site-packages\/tensorflow\/python\/ops\/gen_math_ops.py\", line 387, in add\n    \"Add\", x=x, y=y, name=name)\n  File \"\/home\/hp\/anaconda3\/envs\/mypython37\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 563, in _apply_op_helper\n    inferred_from[input_arg.type_attr]))\nTypeError: Input 'y' of 'Add' Op has type float32 that does not match type int32 of argument 'x'.\n<\/code><\/pre>\n","answers":[{"AnswerId":"57851869","CreationDate":"2019-09-09T10:08:03.280","ParentId":null,"OwnerUserId":"7701908","Title":null,"Body":"<p>modify<\/p>\n\n<pre><code>sum=tf.add(sum,overate)\n<\/code><\/pre>\n\n<p>to:<\/p>\n\n<pre><code>sum=tf.add(tf.cast(sum, overate.dtype),overate)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57851198,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-09T09:24:22.580","AcceptedAnswerId":null,"OwnerUserId":11852801.0,"Title":"How to speed up code running in tensorflow\uff1f","Body":"<pre><code>The first is the pre-processing of data, the amount of data is particularly large, and the process of processing func is not slow, Python's multi-process can not be accelerated, but may also be slow (mainly consumed in serialization deserialization),\n<\/code><\/pre>\n\n<p>How does the big factory speed up the speed of this part of the process?<\/p>\n\n<p>What I am currently seeing is: cython or numba can be used to handle numerical calculations. Spacy can be used to process text, but spacy seems to be less compatible with regular processing such as regular expression substitution. (Well, in fact, mainly the documentation looks tired =. =)\n\"\"\nDuring the training phase, the process of generating a batch also creates a bottleneck. If you don't do real-time data enhancement, you can solve it by first processing the data.<\/p>\n\n<p>But what do you do if you need immediate processing?<\/p>\n\n<p>The methods I see are asynchronous, but it seems to be quite cumbersome to do. There are no best practices for reference. At least the example in pytorch's apex library doesn't feel right. Because the process of loading and feeding the model in my process is written in a loop, the operation of the model and the operation of the dataloader are still running serially.I tried to write a prefetch pipeline using tensorflow's tf.data but found that there is a certain speed problem, it seems that I can't read the file in parallel, maybe because of the dataset.from_generator I used, I can't parallel in the generator's yield process. I thought about using pytorch's dataloader and tf.data nested. I found that I should have completed this prefetch. My test code is in the attached html. Please correct me, is there a better solution? Or does the code have a better packaging method?\n    There is still a small problem, that is, the last generated data is actually the tensor of tf, then there are two steps that still need to run serially. The tensor of tf is converted to tensor of pytorch and tensor is put into cuda. If you can, how to use this? Two steps to write to prefetch?\nthe codes are as the followings:<\/p>\n\n<pre><code>import torch as t \nimport tensorflow as tf\ntf.compat.v1.enable_eager_execution()\n\nfrom torch.utils.data import Dataset as Tset\nfrom torch.utils.data import DataLoader as Tloader\nfrom tensorflow.data import Dataset as TFset\n\n\n!mkdir data\n\nimport os \n\nfor i in range(1000):\n    with open(f'data\/{i}.txt', 'w') as writer:\n        writer.write('sdfjoiwjeifojwoief')\nimport time\n\nclass tset(Tset):\n    def __init__(self):\n        self.files = [f'data\/{i}' for i in os.listdir('data')]\n\n    def __call__(self):\n        for i in range(self.__len__()):\n            yield self.__getitem__(i)\n\n    def __len__(self):\n        return len(self.files)\n\n    def __getitem__(self, index):\n        with open(self.files[index]) as reader:\n            data = reader.readline()\n        time.sleep(2)\n        return data\nclass tloader:\n    def __init__(self, dataset):\n        self.dataset = dataset\n        self.loader = Tloader(self.dataset, batch_size=4, num_workers=2)\n\n    def __call__(self):\n        for i in self.loader:\n            yield i\nt = tset()\ntll = Tloader(t, batch_size=4 ,num_workers=2)\ntl = tloader(t)\ntfset = TFset.from_generator(tl, output_types=(tf.string))\ndef process(x):\n    time.sleep(1)\n    print('----')\n    return x\ntfset = tfset.prefetch(10)\nfrom tqdm import tqdm\nfor i in tqdm(tfset):\n    time.sleep(1)\n    print(i)\nfor i in tqdm(tll):\n    time.sleep(1)\n    print(i)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57851309,"AnswerCount":1,"Tags":"<python><tensorflow><quantization><tpu><google-coral>","CreationDate":"2019-09-09T09:31:04.897","AcceptedAnswerId":57931622.0,"OwnerUserId":3504926.0,"Title":"Google Coral Edge TPU compiled model - Inference always almost the same","Body":"<p>I'm trying to get a Mobilenetv2 model (retrained last layers to my dataset) to run on the Google edge TPU Coral. I'm able to quantize and compile the model with the 'edgetpu_compiler' (followed this page <a href=\"https:\/\/coral.withgoogle.com\/docs\/edgetpu\/compiler\/#usage\" rel=\"nofollow noreferrer\">https:\/\/coral.withgoogle.com\/docs\/edgetpu\/compiler\/#usage<\/a>). But when I run inference in the TPU I'm getting a similar output for very different input images.<\/p>\n\n<p>I've used 'tflite_convert' tool to quantize the model like this:<\/p>\n\n<pre><code>tflite_convert --output_file=.\/model.tflite \n--keras_model_file=models\/MobileNet2_best-val-acc.h5 --output_format=TFLITE\n--inference_type=QUANTIZED_UINT8 --default_ranges_min=0 --default_ranges_max=6 \n--std_dev_values=127 --mean_values=128 --input_shapes=1,482,640,3 --input_arrays=input_2\n<\/code><\/pre>\n\n<p>Then I've used 'edgetpu_compiler' tool to compile it for the TPU:<\/p>\n\n<pre><code>sudo edgetpu_compiler  model.tflite\nEdge TPU Compiler version 2.0.258810407\nINFO: Initialized TensorFlow Lite runtime.\n\nModel compiled successfully in 557 ms.\n\nInput model: model.tflite\nInput size: 3.44MiB\nOutput model: model_edgetpu.tflite\nOutput size: 4.16MiB\nOn-chip memory available for caching model parameters: 4.25MiB\nOn-chip memory used for caching model parameters: 3.81MiB\nOff-chip memory used for streaming uncached model parameters: 0.00B\nNumber of Edge TPU subgraphs: 1\nTotal number of operations: 71\nOperation log: model_edgetpu.log\nSee the operation log file for individual operation details.\n<\/code><\/pre>\n\n<p>Then when I run inference using this code:<\/p>\n\n<pre><code>...\nlabels = [\"Class1\", \"Class2\", \"Class3\", \"Class4\"]\nresults = engine.ClassifyWithImage(img, top_k=4)\nfor result in results:\n    print('---------------------------')\n    print(labels[result[0]])\n    print('Score : ', result[1])\n<\/code><\/pre>\n\n<p>The output is like this (assuming labels [\"Class1\", \"Class2\", \"Class3\", \"Class4\"]):<\/p>\n\n<pre><code>---------------------------\nClass1\nScore :  0.2890625\n---------------------------\nClass2\nScore :  0.26953125\n---------------------------\nClass3\nScore :  0.21875\n---------------------------\nClass4\nScore :  0.21875\n<\/code><\/pre>\n\n<p>It is almost the same for any input image, and usually, the first two classes have the same (or very very similar) value (the same for the 3rd and 4th) as seen in the example shown above. It should be 0.99 for one class <code>(as it is in the .h5 model or even in the .tflite model without quantization)<\/code><\/p>\n\n<p>Can it be something with the parameters <code>-default_ranges_min=0 --default_ranges_max=6 --std_dev_values=127 --mean_values=128?<\/code> How can I calculate them?<\/p>\n\n<p>Edit 1: <\/p>\n\n<p>Using the answer from this <a href=\"https:\/\/stackoverflow.com\/questions\/54830869\/understanding-tf-contrib-lite-tfliteconverter-quantization-parameters?rq=1\">post<\/a> I've tried to quantize the model using both <code>--std_dev_values=127 --mean_values=128 and --std_dev_values=255 --mean_values=0<\/code>, but I'm still getting garbage inference. As mobilenet2 uses relu6, default ranges shoud be <code>-default_ranges_min=0 --default_ranges_max=6<\/code> right?<\/p>\n\n<p>The model is a MobileNetv2 retrained, the input is an RGB image (3 channels), the input shape is 1,482,640,3.<\/p>\n","answers":[{"AnswerId":"57931622","CreationDate":"2019-09-14T00:42:59.963","ParentId":null,"OwnerUserId":"9867252","Title":null,"Body":"<p>From your comment on mobilenetv1, it sounds like you are taking a retrained float model and converting it to TFLite. You intended to quantize it by running the command that you listed.<\/p>\n\n<p>I'd recommend that you take a closer look at the TensorFlow lite docs. In general, there are two ways of quantization (doing it during training time and doing it post-training). The approach you seem to want to take is post-training. <\/p>\n\n<p>The proper way of doing it post-training for something like Coral is to follow this\nguide (<a href=\"https:\/\/www.tensorflow.org\/lite\/performance\/post_training_integer_quant\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/lite\/performance\/post_training_integer_quant<\/a>), as recommended by the Coral team here (<a href=\"https:\/\/coral.withgoogle.com\/news\/updates-07-2019\/\" rel=\"nofollow noreferrer\">https:\/\/coral.withgoogle.com\/news\/updates-07-2019\/<\/a>). <\/p>\n\n<p>The flow you're using above is more geared towards training time quantization.<\/p>\n"}]}
+{"QuestionId":57851525,"AnswerCount":0,"Tags":"<python><tensorflow><tf.keras>","CreationDate":"2019-09-09T09:46:05.613","AcceptedAnswerId":null,"OwnerUserId":4515762.0,"Title":"Do I have to compile a TF.keras model that's part of a compiled model in order for its added losses to be taken into account during training?","Body":"<p>I have a VAE tf.keras model and it consists of an encoder and decoder model. I want to regularize the encoder and the decoder with different custom terms (using <code>add_loss()<\/code>) while training VAE (using <code>VAE.fit()<\/code>).<\/p>\n\n<p>In pseudocode:<\/p>\n\n<pre><code>encoder = model(input, encoding)\ndecoder = model(encoding, output)\nencoder.add_loss(custom_loss_1)\ndecoder.add_loss(custom_loss_2)\nvae = model(input, output)\nvae.compile(optimizer, standard_keras_loss_function)\nvae.fit(args)\n<\/code><\/pre>\n\n<p>Note that I can't add the custom losses to the wrapper model (vae) because I want them to be backpropagated only in their respective submodels (encoder's loss shouldnt affect decoder's gradients and vice versa). I also can't fit encoder and decoder separately, VAEs benefit from jointly training the encoder and decoder.<\/p>\n\n<p>I'm facing two challenges:<\/p>\n\n<ol>\n<li><p>Although I see no error during training, I have no easy way to heck if encoder and decoder losses are taken into account when fitting bar wrapper model. Do I need to compile encoder and decoder for the added custom losses to be taken into account?<\/p><\/li>\n<li><p>I don't want the 'standard_keras_loss' (binary crossentropy) to be backpropagated beyond the decoder, it should only affect the decoder's gradients. However, <code>vae.compile()<\/code> method requires me to add at least one loss to the VAE model. Thus, the only way to prevent this loss from backpropagating to encoder is to try and add a <code>stop_gradient<\/code> lambda layer after the encoder but this would disconnect my encoder from my decoder (in fact I'm getting disconnected graph error wherever I try to add the <code>stop_gradient<\/code> layer).<\/p><\/li>\n<\/ol>\n\n<p>Is there a <code>TF.keras<\/code> way to do this without having to write custom training loop? I'm trying to keep a clean code. <\/p>\n","answers":[]}
+{"QuestionId":57852043,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras><neural-network>","CreationDate":"2019-09-09T10:19:29.107","AcceptedAnswerId":null,"OwnerUserId":10794530.0,"Title":"Keras conditional passing one model output to another model","Body":"<p>I am trying to neural network in keras, which first check if its a cat or dog (base model). <\/p>\n\n<p>if it is a dog, then it is passes through another model (sub-model-1)<\/p>\n\n<p>if it is a cat, then it is passes through another model (sub-model-2)<\/p>\n\n<p>Sub-model are small model specially trained for classifying on the basis of breed.. so sub-model-1 will class dog into various dog's breed. . while sub-model-2 will classify cat into various cat's breed.<\/p>\n\n<p>The problem i am facing is :\nI dont know how to add conditional layer,so that if base model has 5 million neuron and each sub-model has 2 million -2million neuron.. if an image is passed through base model, then it should only pass through sub-model1 or sub-model2.. so in total only 7 million neuron at action in passing one image to final output.<\/p>\n\n<p>Any help, reference, everything would be appreciable.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/a5a8W.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/a5a8W.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57852096,"AnswerCount":2,"Tags":"<python><python-3.x><keras>","CreationDate":"2019-09-09T10:23:15.417","AcceptedAnswerId":57852812.0,"OwnerUserId":8173877.0,"Title":"Determining output of model.predict in relation to classess","Body":"<p>I'm trying to figure out how to determine the class from a one hot encoded vector that is returned from keras. The problem is I use the flow_from_directory with ImageDataGenerator to train my network, leading to keras automatically converting the folder structure into the class vector to my knowledge, how do I go about resolving this?<\/p>\n\n<p>Here's how my training is laid out in code:<\/p>\n\n<pre><code>    checkpoint = [ModelCheckpoint(\n        'model.checkpoint.hdf5',\n        period=1\n   )]\n\n    train_datagen = ImageDataGenerator(\n        rotation_range=30)\n\n    test_datagen = ImageDataGenerator(\n        rotation_range=30)\n\n    train_generator = train_datagen.flow_from_directory(\n        '\/train\/',\n        target_size=(x, y),\n        batch_size=batch_size,\n        class_mode='categorical')\n    test_generator = test_datagen.flow_from_directory(\n        '\/test\/',\n        target_size=(x, y),\n        batch_size=batch_size,\n        class_mode='categorical')\n\n    self.discriminator.fit_generator(\n        train_generator,\n        steps_per_epoch=10000,\n        epochs=epochs,\n        validation_data=test_generator,\n        validation_steps=10,\n        callbacks=checkpoint\n    )\n<\/code><\/pre>\n\n<p>My folder structure is like so:<\/p>\n\n<pre><code>root\/\n    train\/\n        cow\/\n            0.jpg\n            1.jpg\n        pig\/\n            0.jpg\n            1.jpg\n    test\/\n        cow\/\n            0.jpg\n            1.jpg\n        pig\/\n            0.jpg\n            1.jpg\n<\/code><\/pre>\n","answers":[{"AnswerId":"57852528","CreationDate":"2019-09-09T10:51:09.863","ParentId":null,"OwnerUserId":"5563864","Title":null,"Body":"<p>The prediction in keras refers to your own structure of the folder. if train is composed of cow and pig. Then <code>cow = 0<\/code> and <code>pig = 1<\/code>. So if the predict returns 0 is cow otherwise it's pig.<\/p>\n"},{"AnswerId":"57852812","CreationDate":"2019-09-09T11:10:16.207","ParentId":null,"OwnerUserId":"5597718","Title":null,"Body":"<pre><code>print(train_generator.class_indices)\n{'cow': 0, 'pig': 1}\n<\/code><\/pre>\n\n<p>From the documentation, \"The dictionary containing the mapping from class names to class indices can be obtained via the attribute class_indices.\"<\/p>\n\n<p><a href=\"https:\/\/keras.io\/preprocessing\/image\/#flow_from_directory\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/preprocessing\/image\/#flow_from_directory<\/a><\/p>\n"}]}
+{"QuestionId":57852236,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-09-09T10:31:27.797","AcceptedAnswerId":null,"OwnerUserId":2476373.0,"Title":"Cleaner way to use \"with torch.no_grad()\" conditioned on an expression","Body":"<p>My code looks like:<\/p>\n\n<pre><code>if no_grad_condition:\n  with torch.no_grad():\n    out=network(input)\nelse:\n  out=network(input)\n<\/code><\/pre>\n\n<p>Is there a cleaner way to do it, without duplicating the line <code>out=network(input)<\/code>?<\/p>\n\n<p>I am looking for something in the spirit of:<\/p>\n\n<pre><code>  with torch.no_grad(no_grad_condition):\n    out=network(input)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57852237","CreationDate":"2019-09-09T10:31:27.797","ParentId":null,"OwnerUserId":"2476373","Title":null,"Body":"<p>OP here: By writing down the question, I understood where to look for the answer. According to <a href=\"https:\/\/pytorch.org\/docs\/stable\/autograd.html#torch.autograd.set_grad_enabled\" rel=\"nofollow noreferrer\">pytorch docs<\/a>, we can use <code>set_grad_enabled<\/code>:<\/p>\n\n<pre><code>  with torch.set_grad_enabled(not no_grad_condition):\n    out=network(input)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57852271,"AnswerCount":0,"Tags":"<python><tensorflow><keras><recurrent-neural-network>","CreationDate":"2019-09-09T10:34:00.930","AcceptedAnswerId":null,"OwnerUserId":9857120.0,"Title":"Shape Mismatch: The shape of labels(received (128,)) should equal the shape of logits except for the last dimension (received (16, 424))","Body":"<p>Error \n    ValueError: in converted code:<\/p>\n\n<pre><code>&lt;ipython-input-63-1e3afece3370&gt;:10 train_step  *\n    loss += loss_func(targ, logits)\n&lt;ipython-input-43-44b2a8f6794e&gt;:11 loss_func  *\n    loss_ = loss_object(real, pred)\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow_core\/python\/keras\/losses.py:124 __call__\n    losses = self.call(y_true, y_pred)\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow_core\/python\/keras\/losses.py:216 call\n    return self.fn(y_true, y_pred, **self._fn_kwargs)\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow_core\/python\/keras\/losses.py:973 sparse_categorical_crossentropy\n    y_true, y_pred, from_logits=from_logits, axis=axis)\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow_core\/python\/keras\/backend.py:4431 sparse_categorical_crossentropy\n    labels=target, logits=output)\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow_core\/python\/ops\/nn_ops.py:3477 sparse_softmax_cross_entropy_with_logits_v2\n    labels=labels, logits=logits, name=name)\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow_core\/python\/ops\/nn_ops.py:3393 sparse_softmax_cross_entropy_with_logits\n    logits.get_shape()))\n\nValueError: Shape mismatch: The shape of labels (received (128,)) should equal the shape of logits except for the last dimension (received (16, 424)).\n<\/code><\/pre>\n\n<p>The code I refered to used legacyseq2seq.sequence_loss_by_example which is deprecated now. So I have tried SparseCategoricalCrossentropy loss methods throws the same error<\/p>\n\n<p>Model (Keras)<\/p>\n\n<pre><code>    def build_model(training=True):\n       input_ = tf.keras.layers.Input(shape=(unfold_max,), name='inputs')\n       embedding = tf.keras.layers.Embedding(input_dim=n_items + 1,\n                                      output_dim=128)(input_)\n       cells=[]\n       for _ in range(1):\n           cells.append(\n                tf.keras.layers.LSTMCell(128,\n                              dropout=0.25,\n                              activation=tf.keras.activations.tanh)\n       )\n       cell_output = tf.keras.layers.RNN(cells, return_state=False)(embedding)\n       softmax_W = tf.Variable(tf.ones(shape=(128,1 + n_items)), 'softmax_w')\n       softmax_b = tf.Variable(tf.zeros(shape=(n_items + 1)),'softmax_b')\n       output = tf.reshape(cell_output, [-1, 128])\n       logits = tf.matmul(output, softmax_W) + softmax_b\n       return tf.keras.Model(inputs=[input_], outputs=[logits])\n<\/code><\/pre>\n\n<p>Loss Function<\/p>\n\n<pre><code>    optimizer = tf.keras.optimizers.Adam()\n    loss_object = tf.keras.losses.SparseCategoricalCrossentropy(\n                     from_logits=True,\n                     reduction='none')\n\n    def loss_func(real, pred):\n        mask = tf.math.logical_not(tf.math.equal(real, 0))\n        loss_ = loss_object(real, pred)\n        mask = tf.cast(mask, dtype=loss_.dtype)\n        loss_ *= mask   \n        return tf.reduce_mean(loss_)\n<\/code><\/pre>\n\n<p>Train Step<\/p>\n\n<pre><code>    @tf.function\n    def train_step(inp, targ):\n        loss = 0\n        with tf.GradientTape() as tape:\n             logits = model(inp)\n             loss += loss_func(targ, logits)\n\n        variables = model.trainable_variables\n        gradients = tape.gradient(loss, variables)\n        optimizer.apply_gradients(zip(gradients, variables))\n        return loss\n<\/code><\/pre>\n\n<p>input : [135, 144, 0, 0, 0, 0, 0, 0]<\/p>\n\n<p>target : [144, 127, 0, 0, 0, 0, 0, 0]<\/p>\n","answers":[]}
+{"QuestionId":57852296,"AnswerCount":0,"Tags":"<pytorch><fast-ai>","CreationDate":"2019-09-09T10:35:57.647","AcceptedAnswerId":null,"OwnerUserId":6699013.0,"Title":"Need help in troubleshooting error thrown during creation of databunch","Body":"<p>I am trying to create a GAN based on the notebook \"lesson7-superres-gan\" of Part 1(2019). I am using the dataset here <a href=\"https:\/\/www.kaggle.com\/rounakbanik\/pokemon\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/rounakbanik\/pokemon<\/a>. \nWhile creating data for pretraining the critic I am facing the following error.<\/p>\n\n<p><code>RuntimeError: number of dims don't match in permute<\/code><\/p>\n\n<p>I tried to change the size of the images but it didn't seem to solve the error.<\/p>\n\n<p>The code which is creating the data is --><\/p>\n\n<pre><code>        data = image_source.label_empty().transform(size=image_size,tfm_y=True).databunch   (bs=batch_size).normalize(imagenet_stats, do_y=True)\n    data.c = 3\n    return data\n<\/code><\/pre>\n\n<p>Data --><\/p>\n\n<pre><code>data = get_databunch(image_source, path_lr, batch_size, image_size)\n<\/code><\/pre>\n\n<p>The line for code which throws the error--><\/p>\n\n<pre><code>data.show_batch(2)\n<\/code><\/pre>\n\n<p>The stack trace --><\/p>\n\n<pre><code>    ---------------------------------------------------------------------------\nRuntimeError                              Traceback (most recent call last)\n&lt;ipython-input-129-5efe2b85bc64&gt; in &lt;module&gt;()\n----&gt; 1 data.show_batch()\n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\fastai\\basic_data.py in show_batch(self, rows, ds_type, reverse, **kwargs)\n    192             ys = [self.train_ds.y.reconstruct(grab_idx(y, i), x=x) for i,x in enumerate(xs)]\n    193         else : ys = [self.train_ds.y.reconstruct(grab_idx(y, i)) for i in range(n_items)]\n--&gt; 194         self.train_ds.x.show_xys(xs, ys, **kwargs)\n    195 \n    196     def export(self, file:PathLikeOrBinaryStream='export.pkl'):\n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\fastai\\vision\\data.py in show_xys(self, xs, ys, imgsize, figsize, **kwargs)\n    302         rows = int(np.ceil(math.sqrt(len(xs))))\n    303         axs = subplots(rows, rows, imgsize=imgsize, figsize=figsize)\n--&gt; 304         for x,y,ax in zip(xs, ys, axs.flatten()): x.show(ax=ax, y=y, **kwargs)\n    305         for ax in axs.flatten()[len(xs):]: ax.axis('off')\n    306         plt.tight_layout()\n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\fastai\\vision\\image.py in show(self, ax, figsize, title, hide_axis, cmap, y, **kwargs)\n    219         cmap = ifnone(cmap, defaults.cmap)\n    220         ax = show_image(self, ax=ax, hide_axis=hide_axis, cmap=cmap, figsize=figsize)\n--&gt; 221         if y is not None: y.show(ax=ax, **kwargs)\n    222         if title is not None: ax.set_title(title)\n    223 \n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\fastai\\vision\\image.py in show(self, ax, figsize, title, hide_axis, cmap, y, **kwargs)\n    218         \"Show image on `ax` with `title`, using `cmap` if single-channel, overlaid with optional `y`\"\n    219         cmap = ifnone(cmap, defaults.cmap)\n--&gt; 220         ax = show_image(self, ax=ax, hide_axis=hide_axis, cmap=cmap, figsize=figsize)\n    221         if y is not None: y.show(ax=ax, **kwargs)\n    222         if title is not None: ax.set_title(title)\n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\fastai\\vision\\image.py in show_image(img, ax, figsize, hide_axis, cmap, alpha, **kwargs)\n    428     \"Display `Image` in notebook.\"\n    429     if ax is None: fig,ax = plt.subplots(figsize=figsize)\n--&gt; 430     ax.imshow(image2np(img.data), cmap=cmap, alpha=alpha, **kwargs)\n    431     if hide_axis: ax.axis('off')\n    432     return ax\n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\fastai\\vision\\image.py in image2np(image)\n     22 def image2np(image:Tensor)-&gt;np.ndarray:\n     23     \"Convert from torch style `image` to numpy\/matplotlib style.\"\n---&gt; 24     res = image.cpu().permute(1,2,0).numpy()\n     25     return res[...,0] if res.shape[2]==1 else res\n     26 \n\nRuntimeError: number of dims don't match in permute\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57852705,"AnswerCount":0,"Tags":"<python><tensorflow><distributed-tensorflow>","CreationDate":"2019-09-09T11:02:54.270","AcceptedAnswerId":null,"OwnerUserId":6067716.0,"Title":"How to create a custom distribution strategy on Tensorflow","Body":"<p>I', looking to write a custom distribution strategy for tensorflow.<\/p>\n\n<p>Currently there are several types strategies available (MirroredStrategy,   TPUStrategy, ...), but i would like to implement a new way of distributing training across several machines.<\/p>\n\n<p>I looked through the documentation and didn't find any tutorial or guide on how to implement a custom strategy<\/p>\n\n<p>Is there already a predefined way to do this?<\/p>\n","answers":[]}
+{"QuestionId":57853169,"AnswerCount":0,"Tags":"<python><pytorch><fast-ai>","CreationDate":"2019-09-09T11:32:41.740","AcceptedAnswerId":null,"OwnerUserId":5663844.0,"Title":"torch.no_grad save gradients of each batch","Body":"<p>so I am doing the FastAI Course and I wanted to try to implement Momentum myself. However, I struggle to understand how to save the gradients after each batch. The problem is in my code is that only zero tensors are saved. I understand that after each iteration the gradients are set back to zero but if I save them before I should get the actual ones.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\ndef update2(x,y,lr, grads):\n    wd = 1e-5\n    y_hat = model(x)\n    # weight decay\n    w2 = 0.\n    for p in model.parameters(): w2 += (p**2).sum()\n    # add to regular loss\n    loss = loss_func(y_hat, y) + w2*wd\n    loss.backward()\n    with torch.no_grad():\n        for idx,p in enumerate(model.parameters()):\n            #compute the gradient with momentum\n            gradient= p.grad*0.1+ grads[idx]*0.9\n            #save current gradients\n            grads[idx]=p.grad\n            p.sub_(lr * gradient)\n            p.grad.zero_()\n    return loss.item(), grads\n```\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57853326,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-09T11:41:38.407","AcceptedAnswerId":null,"OwnerUserId":4502109.0,"Title":"AttributeError: module 'tensorflow.tools.api.generator.api.nn' has no attribute 'max_pool1d'","Body":"<p>This is the first time I rise a query. Please help me and thanks a lot.\nI am working with a 1d max pool, but I got that error.\nMy tf version is 1.8.<\/p>\n\n<p>I've checked some similar queries but seems none of them could solve my problem.<\/p>\n\n<pre><code>max_pool = tf.nn.max_pool1d(relu3,ksize=2,strides=1,padding='SAME')\n<\/code><\/pre>\n\n<p>According to TF's online doc, it should be okay, unfortunately, it's not. \nCould you please enlighten me?<\/p>\n","answers":[]}
+{"QuestionId":57853348,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-09-09T11:43:00.007","AcceptedAnswerId":null,"OwnerUserId":5079088.0,"Title":"Keras network to handle two input images","Body":"<p>I want to use Keras CNN to classify images of coins. Each coin has two images (one per side), and I don't think it is a good idea to teach CNN on BOTH sides simultaneously.\nSo the question is, how to join two (identical but trained on different sides) CNNs, put a common fully connected classifier on them, and how to feed data to this construct (how to feed two images and make sure each half of a network gets its image).\nHelp appreciated.\nThank you.<\/p>\n","answers":[]}
+{"QuestionId":57853630,"AnswerCount":1,"Tags":"<python><tensorflow><chatbot><miniconda><rasa>","CreationDate":"2019-09-09T12:01:36.713","AcceptedAnswerId":null,"OwnerUserId":8242333.0,"Title":"How to fix 'Could not find a version that satisfies the requirement tensorflow==1.12.0' error?","Body":"<p>I am trying to create a chatbot using RASA framework . I am running the following command to install the requirements.txt file and its throwing error.<\/p>\n\n<pre><code>pip install -r requirements.txt\n\nERROR: Could not find a version that satisfies the requirement tensorflow==1.12.0 (from -r requirements.txt (line 11)) (from versions: none)\nERROR: No matching distribution found for tensorflow==1.12.0 (from -r requirements.txt (line 11))\n<\/code><\/pre>\n","answers":[{"AnswerId":"57853956","CreationDate":"2019-09-09T12:19:30.940","ParentId":null,"OwnerUserId":"2043395","Title":null,"Body":"<p>That version of tensorflow is not available. Please update your requirements.txt to update from tensorflow==1.12.0 to any of below:<\/p>\n\n<pre><code>1.13.0rc1, 1.13.0rc2, 1.13.1, 1.13.2, 1.14.0rc0, 1.14.0rc1, 1.14.0, 2.0.0a0, 2.0.0b0, 2.0.0b1, 2.0.0rc0\n<\/code><\/pre>\n"}]}
+{"QuestionId":57853775,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-09T12:10:10.173","AcceptedAnswerId":57854292.0,"OwnerUserId":10478535.0,"Title":"Why amount of used GPU memory doesn't increase when increasing batch size in keras?","Body":"<p>I have an image classifier in Keras. I want to achieve the fastest performance possible, but I notice that the amount of used GPU memory is constant no matter how big a batch is. <\/p>\n\n<p>I set <\/p>\n\n<pre><code>from keras.backend.tensorflow_backend import set_session\nconfig = tf.ConfigProto()\nconfig.gpu_options.allow_growth = True\nset_session(tf.Session(config=config))\n<\/code><\/pre>\n\n<p>to prevent tf allocating the whole memory.<\/p>\n\n<p>When I load the model <\/p>\n\n<pre><code>from keras.models import load_model\nmodel = load_model(model_h5_path)\n<\/code><\/pre>\n\n<p><code>nvidia-smi<\/code> shows that 1017MiB is used.<\/p>\n\n<p>Then I do <code>model.predict(batch)<\/code> using different batch sizes and measure the time here are results I got<\/p>\n\n<p>Batch 1 \u2014 4.5 ms<\/p>\n\n<p>Batch 10 \u2014 7 ms<\/p>\n\n<p>Batch 50 \u2014 22ms<\/p>\n\n<p>Batch 100 \u2014 45 ms<\/p>\n\n<p>Batch 1000 \u2014 470ms<\/p>\n\n<p>And each time <code>nvidia-smi<\/code> shows that only 1017MiB is used.<\/p>\n\n<p>If I don't set <code>allow_growth = True<\/code> I get the same result.  <\/p>\n\n<p>I expect the amount of used GPU to increase when I increase the batch size. \nI have GeForce RTX 2080 Ti with 11016MiB of memory. So it seems that I don't use the video card to the fullest, since I have 10 GB of unused memory. <\/p>\n\n<p>Please suggest how can I leverage the unused memory. A naive solution would be to create 10 models, since each of them consumes only 1 GB, but maybe there is a more elegant solution. Or maybe it is supposed to be that way, and it is impossible to use more memory.<\/p>\n","answers":[{"AnswerId":"57854292","CreationDate":"2019-09-09T12:42:25.403","ParentId":null,"OwnerUserId":"1329296","Title":null,"Body":"<p>According to the documentation (<a href=\"https:\/\/keras.io\/models\/sequential\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/models\/sequential\/<\/a>), <code>predict(...)<\/code> takes a parameter <code>batch_size<\/code> -- if you don't specify one, it defaults to 32. Hence the behaviour you see.<\/p>\n\n<p>It does not matter how many images you put into <code>batch<\/code>, you have to explicitly specify the batch size.<\/p>\n"}]}
+{"QuestionId":57853800,"AnswerCount":1,"Tags":"<python><keras><deep-learning><keras-layer>","CreationDate":"2019-09-09T12:11:25.957","AcceptedAnswerId":57854269.0,"OwnerUserId":2207729.0,"Title":"Why do we need to create a new model to see intermediate layers' outputs?","Body":"<p>In most examples I see where we want to see the intermediate layers' output values, it seems creating a new Model is the way to do:<\/p>\n\n<pre><code>from keras.models import Model\n\nmodel = ...  # include here your original model\n\nlayer_name = 'my_layer'\nintermediate_layer_model = Model(inputs=model.input,\n                                 outputs=model.get_layer(layer_name).output)\nintermediate_output = intermediate_layer_model.predict(data)\n<\/code><\/pre>\n\n<p>(source: <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-the-output-of-an-intermediate-layer\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-the-output-of-an-intermediate-layer<\/a>)<\/p>\n\n<p>However, what I don't understand about it is: what does this model do exactly? Why do we need it? Can't we just feed inputs to the original model having those layers, and pool its layer values directly?<\/p>\n\n<p>Do you have to train this new model as well?<\/p>\n\n<p>I cannot even imagine any diagram that can represent what is going on with this new model... You add the input layer of another model, then add a random intermediary layer of that other model as output, and feed inputs to it?<\/p>\n\n<p>Also, won't this new model affect the output? Won't it try to learn or require training, or the layer brings its own weights pre trained from the original model, thus providing \"altered\" output values?<\/p>\n","answers":[{"AnswerId":"57854269","CreationDate":"2019-09-09T12:40:56.367","ParentId":null,"OwnerUserId":"51685","Title":null,"Body":"<p>In the FAQ article you've linked, that's not <em>the<\/em> way, it's a simple way.<\/p>\n\n<p>A <code>Model<\/code>, after all, is just a graph of layers with a convenient API (<code>.train()<\/code>, <code>.predict()<\/code>, IO, ...).<\/p>\n\n<p>If you create a new Model from another model's input and output layer(s), you're essentially grabbing a slice of it that has the same convenient API.<\/p>\n\n<p>The layers in this slice model are as they were on the full model, so if the model was trained to begin with, so will this submodel. (With all likelihood, the layers are just references to the layers in the original model, so training the submodel would also train (likely to a bad result) the primary model.)<\/p>\n"}]}
+{"QuestionId":57854204,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><keras>","CreationDate":"2019-09-09T12:36:36.937","AcceptedAnswerId":57854361.0,"OwnerUserId":1157814.0,"Title":"Why model.predict computes different values in this sample?","Body":"<p><strong>Context<\/strong><\/p>\n\n<p>Given the following sample, I'm using Jupyter Notebook :<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import Dense\n\nimport numpy as np\nx_input = np.array([[1,2,3,4,5]])\ny_input = np.array([[10]])\n\nmodel = Sequential()\nmodel.add(Dense(units=32, activation=\"tanh\", input_dim=x_input.shape[1], kernel_initializer='random_normal'))\nmodel.add(Dense(units=1, kernel_initializer='random_normal'))\n\nmodel.compile(loss='mse', optimizer='sgd', metrics=['accuracy'])\nhistory = model.fit(x_input, y_input, epochs=10, batch_size=32)\n<\/code><\/pre>\n\n<p>When I run <code>model.predict(x_input)<\/code> I got:<\/p>\n\n<blockquote>\n  <p>array([[9.993563]], dtype=float32)<\/p>\n<\/blockquote>\n\n<p>When I run <code>model.predict(np.array([[1,2,5,4,5]]))<\/code> I got:<\/p>\n\n<blockquote>\n  <p>array([[10.180285]], dtype=float32)<\/p>\n<\/blockquote>\n\n<p><strong>Question<\/strong><\/p>\n\n<p>Should I get the very same prediction in both cases? (While using the same fit model)<\/p>\n","answers":[{"AnswerId":"57854361","CreationDate":"2019-09-09T12:47:12.777","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Well, the inputs are no the same, the first is <code>[1,2,3,4,5]<\/code> and the second is <code>[1,2,5,4,5]<\/code>. The third element of both arrays are not the same.<\/p>\n"}]}
+{"QuestionId":57854932,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><yolo>","CreationDate":"2019-09-09T13:22:26.260","AcceptedAnswerId":null,"OwnerUserId":519882.0,"Title":"function call returns dimensions 1787, 1781 and 1793 even after we using seed value","Body":"<p>with tf.Session() as test_a:<\/p>\n\n<pre><code>box_confidence = tf.random_normal([19, 19, 5, 1], mean=1, stddev=4, seed = 1)\n\nboxes = tf.random_normal([19, 19, 5, 4], mean=1, stddev=4, seed = 1)\n\nbox_class_probs = tf.random_normal([19, 19, 5, 80], mean=1, stddev=4, seed = 1)\n\nscores, boxes, classes = yolo_filter_boxes(box_confidence, boxes, box_class_probs, threshold = 0.5)\n\nprint (\"VRK scores eval \" , (scores.eval().shape))\n\nprint (\"VRK boxes eval \" , (boxes.eval().shape))\n\nprint (\"VRK classes eval \" , (classes.eval().shape))\n<\/code><\/pre>\n\n<p>Above I am getting below output:<\/p>\n\n<pre><code>box_scores shape (19, 19, 5, 80)\nbox_classes i.e, argmax output shape  (19, 19, 5)\nbox_class_scores i.e, max output shape  (19, 19, 5)\nFiltering mask shape (19, 19, 5)\nNumber of true values  Tensor(\"Greater_16:0\", shape=(19, 19, 5), dtype=bool)\nBoxes shape (19, 19, 5, 4)\nVRK scores eval  (1787,)\nVRK boxes eval  (1781, 4)\nVRK classes eval  (1793,)\n<\/code><\/pre>\n\n<p>Why I am getting different values as 1787, 1781, 1793, as I am expecting same value for all three tensors? As we are using seed i am expecting same value even we run different eval operations. For this question assume we have any function other than yolo_filter_boxes like calculate sum<\/p>\n","answers":[]}
+{"QuestionId":57855016,"AnswerCount":0,"Tags":"<python><image-processing><keras>","CreationDate":"2019-09-09T13:27:34.220","AcceptedAnswerId":null,"OwnerUserId":11008983.0,"Title":"How to fix \"AttributeError: 'str' object has no attribute '__array_interface__'\" while doing Image classification using python","Body":"<p>I am doing Image Classification using Keras and while training the model I am getting an error which says that \"AttributeError: 'str' object has no attribute '<strong>array_interface<\/strong>'\" and it is being generated from Image.py file of PIL and because of which my model is not getting started for training.<\/p>\n\n<p>I have already tried to do some changes in Image.py file and in my code file but, the error still exists. The code file was working earlier but now it is throwing me an \"AttributeError\".<\/p>\n\n<pre><code>img_width, img_height = 150, 150\ntrain_dir = r'D:\\DataSets\\Cats_Dogs\\train'\nvalidation_dir = r'D:\\DataSets\\Cats_Dogs\\test'\nnb_train_samples = 5000\nnb_validation_samples = 1000\nepochs = 50\nbatch_size = 20\n\ntrain_datagen = ImageDataGenerator(rescale=1.\/255, shear_range=0.2, zoom_range=0.2, horizontal_flip=True,\n                                   vertical_flip=True, rotation_range=0.2)\nvalidation_datagen = ImageDataGenerator(rescale=1.\/255)\n\ntrain_generator = train_datagen.flow_from_directory(train_dir, target_size=(img_width, img_height)\n                                                    , batch_size=batch_size, class_mode='binary')\n\nvalidation_generator = validation_datagen.flow_from_directory(validation_dir, target_size=(img_width, img_height)\n                                                        , batch_size=batch_size, class_mode='binary')\n\nmodel = Sequential()\nmodel.add(Conv2D(32, (3, 3), input_shape=input_shape, activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Dense(1))\nmodel.add(Activation('sigmoid'))\n\nmodel.summary()\n\nmodel.compile(optimizer='rmsprop', loss='binary_crossentropy', metrics=['accuracy'])\n\nmodel.fit_generator(train_generator, steps_per_epoch= nb_train_samples \/\/ batch_size, epochs=epochs,validation_data=validation_generator, validation_steps= nb_validation_samples \/\/ batch_size)\n\nUse tf.cast instead.\nEpoch 1\/50\nTraceback (most recent call last):\n  File \"C:\/Users\/user\/PycharmProjects\/ImageClassification\/BinaryClass.py\", line 56, in &lt;module&gt;\n    validation_data=validation_generator, validation_steps= nb_validation_samples \/\/ batch_size)\n  File \"C:\\Users\\user\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\keras\\legacy\\interfaces.py\", line 91, in wrapper\n    return func(*args, **kwargs)\n  File \"C:\\Users\\user\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\keras\\engine\\training.py\", line 1658, in fit_generator\n    initial_epoch=initial_epoch)\n  File \"C:\\Users\\user\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\keras\\engine\\training_generator.py\", line 181, in fit_generator\n    generator_output = next(output_generator)\n  File \"C:\\Users\\user\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\keras\\utils\\data_utils.py\", line 616, in get\n    six.reraise(*sys.exc_info())\n  File \"C:\\Users\\user\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\six.py\", line 693, in reraise\n    raise value\n  File \"C:\\Users\\user\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\keras\\utils\\data_utils.py\", line 603, in get\n    inputs = future.get(timeout=30)\n  File \"C:\\Users\\user\\AppData\\Local\\Continuum\\anaconda3\\lib\\multiprocessing\\pool.py\", line 657, in get\n    raise self._value\n  File \"C:\\Users\\user\\AppData\\Local\\Continuum\\anaconda3\\lib\\multiprocessing\\pool.py\", line 121, in worker\n    result = (True, func(*args, **kwds))\n  File \"C:\\Users\\user\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\keras\\utils\\data_utils.py\", line 406, in get_index\n    return _SHARED_SEQUENCES[uid][i]\n  File \"C:\\Users\\user\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\keras_preprocessing\\image\\iterator.py\", line 65, in __getitem__\n    return self._get_batches_of_transformed_samples(index_array)\n  File \"C:\\Users\\user\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\keras_preprocessing\\image\\iterator.py\", line 230, in _get_batches_of_transformed_samples\n    interpolation=self.interpolation)\n  File \"C:\\Users\\user\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\keras_preprocessing\\image\\utils.py\", line 110, in load_img\n    img = pil_image.fromarray(path)\n  File \"C:\\Users\\user\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\PIL\\Image.py\", line 2643, in fromarray\n    arr = obj.__array_interface__\nAttributeError: 'str' object has no attribute '__array_interface__'\n\nProcess finished with exit code 1\n<\/code><\/pre>\n\n<p>In the code, I am giving the path of my training and testing dataset and at the time of making a model fit on training and validation dataset, it is giving me an error.<\/p>\n","answers":[]}
+{"QuestionId":57855018,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2019-09-09T13:27:40.107","AcceptedAnswerId":null,"OwnerUserId":1179925.0,"Title":"Keras: get output before activation","Body":"<p>In pretrained VGG19 model layers are defined like<\/p>\n\n<pre><code>x = layers.Conv2D(64, (3, 3),\n                  activation='relu',\n                  padding='same',\n                  name='block1_conv1')(img_input)\n<\/code><\/pre>\n\n<p>I'm using already pretrained network, is it possible to get output before relu activation? or it's only an option to create new model with separate conv2d and activation and copy weights from pretrained model?<\/p>\n","answers":[]}
+{"QuestionId":57855158,"AnswerCount":1,"Tags":"<android><android-studio><tensorflow><face-detection>","CreationDate":"2019-09-09T13:36:24.677","AcceptedAnswerId":null,"OwnerUserId":5340621.0,"Title":"Is there any way to simply use google mediapipe framework in android studio for handtracking?","Body":"<p>I want to implement Hand Tracking using \"Google Mediapipe Framework\" but It is very tough for me to go with given google references.<\/p>\n\n<p>Mediapipe Github :- <a href=\"https:\/\/github.com\/google\/mediapipe\" rel=\"nofollow noreferrer\">https:\/\/github.com\/google\/mediapipe<\/a><\/p>\n\n<p>Installation Doc :-  <a href=\"https:\/\/mediapipe.readthedocs.io\/en\/latest\/install.html#setting-up-android-studio-with-mediapipe\" rel=\"nofollow noreferrer\">https:\/\/mediapipe.readthedocs.io\/en\/latest\/install.html#setting-up-android-studio-with-mediapipe<\/a><\/p>\n\n<p>I have followed the provided steps but unable to run the media-pipe library.<\/p>\n\n<p>It is throwing \"No Bazel found in the directory(xxx\/xxx\/mediapipe)\" while I have installed \"Bazel Plugin\" in android-studio.<\/p>\n\n<p>So, can anyone please suggest me to solve this problem or any good reference for \"HandTracking\" in android.<\/p>\n","answers":[{"AnswerId":"57879614","CreationDate":"2019-09-11T00:25:48.063","ParentId":null,"OwnerUserId":"12049406","Title":null,"Body":"<p>In addition to the \"Bazel Plugin\" for android-studio, you also need to install the Bazel build system.  You can follow the instructions for \"Install Bazel\" at <a href=\"http:\/\/github.com\/google\/mediapipe\/blob\/master\/mediapipe\/docs\/install.md\" rel=\"nofollow noreferrer\">http:\/\/github.com\/google\/mediapipe\/blob\/master\/mediapipe\/docs\/install.md<\/a><\/p>\n"}]}
+{"QuestionId":57855183,"AnswerCount":0,"Tags":"<pytorch><imagenet>","CreationDate":"2019-09-09T13:37:46.937","AcceptedAnswerId":null,"OwnerUserId":2047442.0,"Title":"Can I get list of classes from `torchvision.datasets.ImageNet` without downloading dataset?","Body":"<p>It looks like until recently there were no such list in <code>pytorch<\/code> (see <a href=\"https:\/\/github.com\/pytorch\/vision\/issues\/713\" rel=\"nofollow noreferrer\">here<\/a>). Now we have <code>ImageNet<\/code> class in <code>torchvision<\/code> but it's not quite clear how can I get list of classes.<\/p>\n","answers":[]}
+{"QuestionId":57855557,"AnswerCount":1,"Tags":"<python><docker><pytorch>","CreationDate":"2019-09-09T13:59:43.797","AcceptedAnswerId":57856437.0,"OwnerUserId":848510.0,"Title":"Optimizing docker image size","Body":"<p>I am building a docker image for my python app using the base image <code>python:3.5-slim-buster<\/code>. <\/p>\n\n<p>I am running the below command inside the <code>Dockerfile<\/code>:<\/p>\n\n<p><code>RUN pip install --no-cache-dir -r requirements.txt<\/code> which has <code>torch<\/code> library in the requirements file.<\/p>\n\n<p>After building the image it is having a size of <code>2.29 GB<\/code>. But if I build the image without torch inside the requirements file, it has only ~900 MB.<\/p>\n\n<p>when I manually run the image and check inside the container:<\/p>\n\n<p>The torch (<code>\/usr\/local\/lib\/python3.5\/site-packages\/torch<\/code>) directory is <code>1.3GB<\/code>.<\/p>\n\n<p>So even if I do a multistage build and try to copy the contents from \n<code>\/usr\/local\/lib\/python3.5\/site-packages<\/code> to a new image, I suppose it is not going to help me.<\/p>\n\n<p>Is there any other standard optimization practice which can help me to reduce the image size?<\/p>\n","answers":[{"AnswerId":"57856437","CreationDate":"2019-09-09T14:50:24.533","ParentId":null,"OwnerUserId":"10886420","Title":null,"Body":"<p>Assuming you want a trained model of some kind from <code>pytorch<\/code> (usually neural network) you should indeed use multistage Docker build as follows (IMO at least):<\/p>\n\n<h2>1. Training model and exporting<\/h2>\n\n<p>Write your script as per usual, including all dependencies you need. Train your model and save it as artifact using torchscript's <a href=\"https:\/\/pytorch.org\/docs\/stable\/jit.html\" rel=\"nofollow noreferrer\"><code>torch.jit.script<\/code><\/a>.<\/p>\n\n<h2>2. Second stage<\/h2>\n\n<p>Using C++ write inference code loading your net and compile the source (use <a href=\"https:\/\/pytorch.org\/cppdocs\/installing.html\" rel=\"nofollow noreferrer\"><code>libtorch<\/code><\/a>).<\/p>\n\n<h2>3. Final image<\/h2>\n\n<p>Copy binary from previous step and put it is as Docker's entrypoint so you can run it as disposable neural net within container.<\/p>\n"}]}
+{"QuestionId":57855574,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><keras>","CreationDate":"2019-09-09T14:00:35.623","AcceptedAnswerId":null,"OwnerUserId":7381947.0,"Title":"Is there are way to check that 2 keras models are the same without taking into account nested models","Body":"<p>I would like to safely refactor Keras models. I am thinking about creating some test functions to check that <code>new_model<\/code> is equivalent to <code>old_model<\/code>.<\/p>\n\n<p>Specifically, I would like to be able to be able to nest and unnest submodels used as layers while keeping the equality and match the config and not the instance.<\/p>\n\n<p>I've tried using <code>get_config<\/code> (and ignoring name) to match layers or models and I am now trying to recreate the layers graph. I feel like it's overkill.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.layers import Input, Add\nfrom keras.models import Model\n\ndef build_add_model():\n    x_1 = Input(shape=(10, 10, 3))\n    x_2 = Input(shape=(10, 10, 3))\n    y = Add()([x_1, x_2])\n    return Model([x_1, x_2], y)\n\nadd_model = build_add_model()\n\nx_1_2 = Input(shape=(10, 10, 3))\nx_2_2 = Input(shape=(10, 10, 3))\nwrap_add_model = Model([x_1_2, x_2_2], build_add_model()([x_1_2, x_2_2]))\nthrough_model = Model([x_1_2, x_2_2], x_1_2)\n\ndef is_equivalent_model(left, right):\n    # some keras, tf magics\n    return True\n\nassert is_equivalent_model(add_model, wrap_add_model)\nassert not is_equivalent_model(add_model, through_model)\n<\/code><\/pre>\n\n<p>I would like it to work with bigger models (at least resnet like with possibly multiple sub models with shared blocks).<\/p>\n","answers":[]}
+{"QuestionId":57855659,"AnswerCount":1,"Tags":"<python><numpy><random><pytorch>","CreationDate":"2019-09-09T14:05:53.273","AcceptedAnswerId":57856254.0,"OwnerUserId":11452683.0,"Title":"Numpy random.choice probablities don't sum up to 1","Body":"<p>I wanna randomly select sample points based on the probability distribution specified by <code>prob<\/code> for a given row. However, I get the error in <code>np.random.choice<\/code> that the probabilities don't add up to 1. This is very weird because I first normalize using the L1-norm along the rows and then I define a uniform distribution if the values are smaller than the threshold 1e-6.<\/p>\n\n<pre><code>import numpy as np\nimport torch.nn.functional as F\n\nprob = F.normalize(outputs, p=1, dim=1).clone().data.cpu().numpy() # outputs is a torch.Tensor of shape (14, 6890)\nall_zero = np.where(prob.max(1) &lt; 1e-6)[0] # find indices of rows where all values are smaller\nprob[all_zero] = np.full(prob.shape[1], 1 \/ prob.shape[1]) # fill those rows uniformly\n# ... somewhere later inside a method\nfor j in range(14):\n    sample = np.random.choice(6890, 4, replace=False, p=prob[j])\n<\/code><\/pre>\n\n<p>Do you understand, why that is?<\/p>\n","answers":[{"AnswerId":"57856254","CreationDate":"2019-09-09T14:40:35.943","ParentId":null,"OwnerUserId":"10886420","Title":null,"Body":"<p>As error suggests <code>prob[j]<\/code> doesn't sum to <code>1<\/code>. <\/p>\n\n<p>Your epsilon <code>1e-6<\/code> is way too big to be considered insignificant, there is no need for this operation at all. If you insist, you have to redistribute zero-ed out values across what's left to 1 (and it seems you did just that actually).<\/p>\n\n<p>All in all you didn't normalize the array to 1:<\/p>\n\n<pre><code>prob \/= prob.sum(axis=1) # make it prob dist\n<\/code><\/pre>\n\n<p>BTW. Broadcasting will extend your single number to the whole row, no need for <code>np.full<\/code>:<\/p>\n\n<pre><code>prob[all_zero] = 1 \/ prob.shape[1]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57855824,"AnswerCount":2,"Tags":"<tensorflow><object-detection-api><image-size><faster-rcnn>","CreationDate":"2019-09-09T14:15:26.637","AcceptedAnswerId":null,"OwnerUserId":10615173.0,"Title":"Training Image Size Faster-RCNN","Body":"<p>I will train my dataset with faster-rcnn for one class. All my images are 1920x1080 sizes. Should I resize or crop the images or I can train with this size?\nAlso my objects are really small (around 60x60).<\/p>\n\n<p>In the config file there are dimensions written as min_dimension: 600 and max_dimension: 1024 for this reason I am confused to train the model with 1920x1080 size images.<\/p>\n","answers":[{"AnswerId":"57893382","CreationDate":"2019-09-11T16:31:31.587","ParentId":null,"OwnerUserId":"9418354","Title":null,"Body":"<p>If you insist on the faster-rcnn to cope with this task, personally I recommend:<\/p>\n\n<ol>\n<li><p>Change input height and width, maximum and minimum value in the config file, which should work for your dataset in terms of successfully execution.<\/p><\/li>\n<li><p>Change the original region proposal parameters (should be in config file, too) to certain ratio and scale like 1:1 and 60.<\/p><\/li>\n<\/ol>\n\n<p>But if I were you, I would like to try:<\/p>\n\n<ol>\n<li><p>Add some shortcuts in backbone since it is a small object detection task which is in need of features of high resolution.<\/p><\/li>\n<li><p>Cut the fast-rcnn head off to enhance the performance, since I only need to detect one class to be THE class or not to be (being background or other class), and the output should be enough to encode the information at the RPN stage.<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"57857033","CreationDate":"2019-09-09T15:29:25.890","ParentId":null,"OwnerUserId":"6129428","Title":null,"Body":"<p>If your objects are small, resizing the images to a smaller size is not a good idea. You can change the max_dimension to 1920 or 2000 which might make the speed a bit lower. For cropping the images, you should first consider how the objects are placed in the images. If cropping will cut a lot of objects, then you will have many cases of truncation which might have a negative effect on the model's performance. <\/p>\n"}]}
+{"QuestionId":57856126,"AnswerCount":1,"Tags":"<python><keras><neural-network>","CreationDate":"2019-09-09T14:33:47.943","AcceptedAnswerId":57858286.0,"OwnerUserId":12042174.0,"Title":"Why am I getting a string to float error in my code?","Body":"<p>i am getting a \"could not convert string to float: '\\ufeffage'\" error and I don't know why<\/p>\n\n<pre><code>import matplotlib.pyplot as plt\nimport numpy as np\nimport pandas as pd\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.model_selection import cross_val_score\nfrom sklearn.model_selection import KFold\nfrom sklearn.pipeline import Pipeline\nfrom sklearn.preprocessing import MinMaxScaler\nfrom tensorflow.python.keras.models import Sequential\nfrom tensorflow.python.keras.layers import Dense\nfrom tensorflow.python.keras.wrappers.scikit_learn import KerasRegressor\nimport os;\n\npath = pd.read_csv('car.csv')\n\npath.replace(\",\",\".\")\n\ndataset=np.loadtxt(\"car.csv\", delimiter=\",\")\nx=dataset[:,0:5]\ny=dataset[:,5]\ny=np.reshape(y, (-1,1))\nscaler = MinMaxScaler()\nprint(scaler.fit(x))\nprint(scaler.fit(y))\nxscale=scaler.transform(x)\nyscale=scaler.transform(y)\n<\/code><\/pre>\n\n<p>this last part is where the error occurs <\/p>\n","answers":[{"AnswerId":"57858286","CreationDate":"2019-09-09T16:59:39.030","ParentId":null,"OwnerUserId":"271641","Title":null,"Body":"<p>The file you are reading contains byte order mark and column names. Pandas <code>read_csv<\/code> automatically takes care of this, but <code>np.loadtext<\/code> doesn't. Changing the first three lines of the script to this should help:<\/p>\n\n<pre><code>dataset=pd.read_csv('car.csv').values\n<\/code><\/pre>\n"}]}
+{"QuestionId":57856393,"AnswerCount":0,"Tags":"<tensorflow><nlp><gensim><spacy><tensorflow-hub>","CreationDate":"2019-09-09T14:47:55.530","AcceptedAnswerId":null,"OwnerUserId":3832330.0,"Title":"Load pretrained model on TF-Hub to calculate Word Mover's Distance (WMD) on Gensim or spaCy","Body":"<p>I'd like to calculate <a href=\"http:\/\/proceedings.mlr.press\/v37\/kusnerb15.pdf\" rel=\"nofollow noreferrer\">Word Mover's Distance<\/a> with <a href=\"https:\/\/tfhub.dev\/google\/universal-sentence-encoder\/2\" rel=\"nofollow noreferrer\">Universal Sentence Encoder on TensorFlow Hub<\/a> embedding. <\/p>\n\n<p>I have tried the example on <a href=\"https:\/\/spacy.io\/universe\/project\/wmd-relax\" rel=\"nofollow noreferrer\">spaCy for WMD-relax<\/a>, which loads 'en' model from spaCy, but I couldn't find another way to feed other embeddings. <\/p>\n\n<p>In <a href=\"https:\/\/github.com\/RaRe-Technologies\/gensim\/blob\/develop\/docs\/notebooks\/WMD_tutorial.ipynb\" rel=\"nofollow noreferrer\">gensim<\/a>, it seems that it only accepts <code>load_word2vec_format<\/code> file (<code>file.bin<\/code>) or <code>load<\/code> file (<code>file.vec<\/code>). <\/p>\n\n<p>As I know, someone has written a <a href=\"https:\/\/github.com\/Kyubyong\/bert-token-embeddings\" rel=\"nofollow noreferrer\">Bert to token embeddings based on pytorch<\/a>, but it's not generalized to other models on tf-hub.<\/p>\n\n<p>Is there any other approach to transfer pretrained models on tf-hub to spaCy format or word2vec format?<\/p>\n","answers":[]}
+{"QuestionId":57856606,"AnswerCount":1,"Tags":"<linux><pytorch><caffe><cudnn>","CreationDate":"2019-09-09T15:00:53.480","AcceptedAnswerId":57873271.0,"OwnerUserId":10359704.0,"Title":"Caffe2 doesn't seem to see cuDNN?","Body":"<p>I am trying to install SpConv (spatially sparse convolution library), however when running <code>python3 setup.py bdist_wheel<\/code>, I get an error which seems to be related to Caffe2 not being able to see cuDNN, as I infer from this message: <\/p>\n\n<pre><code>CMake Error at \/home\/ivan\/.local\/lib\/python3.6\/site-packages\/torch\/share\/cmake\/Caffe2\/Caffe2Config.cmake:96 (message):\n  Your installed Caffe2 version uses cuDNN but I cannot find the cuDNN\n  libraries.  Please set the proper cuDNN prefixes and \/ or install cuDNN.\n<\/code><\/pre>\n\n<p>I have tried reinstalling cuDNN but no luck. Can I manually point Caffe2 to cuDNN? This is because usually cuda is located in <code>\/usr\/local\/cuda\/<\/code>, however I have it in <code>\/usr\/lib\/cuda<\/code>.<\/p>\n\n<p>Full output below:<\/p>\n\n<pre><code>running bdist_wheel\nrunning build\nrunning build_py\nrunning build_ext\nRelease\n|||||CMAKE ARGS||||| ['-DCMAKE_PREFIX_PATH=\/home\/ivan\/.local\/lib\/python3.6\/site-packages\/torch', '-DPYBIND11_PYTHON_VERSION=3.6', '-DSPCONV_BuildTests=OFF', '-DCMAKE_CUDA_FLAGS=\"--expt-relaxed-constexpr\" -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__', '-DCMAKE_LIBRARY_OUTPUT_DIRECTORY=\/mnt\/home\/ivan\/second.pytorch\/second\/spconv\/build\/lib.linux-x86_64-3.6\/spconv', '-DCMAKE_BUILD_TYPE=Release']\n-- Caffe2: CUDA detected: 9.1\n-- Caffe2: CUDA nvcc is: \/usr\/bin\/nvcc\n-- Caffe2: CUDA toolkit directory: \/usr\n-- Caffe2: Header version is: 9.1\n-- Could NOT find CUDNN (missing: CUDNN_INCLUDE_DIR) \nCMake Warning at \/home\/ivan\/.local\/lib\/python3.6\/site-packages\/torch\/share\/cmake\/Caffe2\/public\/cuda.cmake:131 (message):\n  Caffe2: Cannot find cuDNN library.  Turning the option off\nCall Stack (most recent call first):\n  \/home\/ivan\/.local\/lib\/python3.6\/site-packages\/torch\/share\/cmake\/Caffe2\/Caffe2Config.cmake:88 (include)\n  \/home\/ivan\/.local\/lib\/python3.6\/site-packages\/torch\/share\/cmake\/Torch\/TorchConfig.cmake:40 (find_package)\n  CMakeLists.txt:20 (find_package)\n\n\n-- Autodetected CUDA architecture(s):  6.1\n-- Added CUDA NVCC flags for: -gencode;arch=compute_61,code=sm_61\nCMake Error at \/home\/ivan\/.local\/lib\/python3.6\/site-packages\/torch\/share\/cmake\/Caffe2\/Caffe2Config.cmake:96 (message):\n  Your installed Caffe2 version uses cuDNN but I cannot find the cuDNN\n  libraries.  Please set the proper cuDNN prefixes and \/ or install cuDNN.\nCall Stack (most recent call first):\n  \/home\/ivan\/.local\/lib\/python3.6\/site-packages\/torch\/share\/cmake\/Torch\/TorchConfig.cmake:40 (find_package)\n  CMakeLists.txt:20 (find_package)\n\n\n-- Configuring incomplete, errors occurred!\nSee also \"\/mnt\/home\/ivan\/second.pytorch\/second\/spconv\/build\/temp.linux-x86_64-3.6\/CMakeFiles\/CMakeOutput.log\".\nSee also \"\/mnt\/home\/ivan\/second.pytorch\/second\/spconv\/build\/temp.linux-x86_64-3.6\/CMakeFiles\/CMakeError.log\".\nTraceback (most recent call last):\n  File \"setup.py\", line 99, in &lt;module&gt;\n    zip_safe=False,\n  File \"\/home\/ivan\/.local\/lib\/python3.6\/site-packages\/setuptools\/__init__.py\", line 145, in setup\n    return distutils.core.setup(**attrs)\n  File \"\/usr\/lib\/python3.6\/distutils\/core.py\", line 148, in setup\n    dist.run_commands()\n  File \"\/usr\/lib\/python3.6\/distutils\/dist.py\", line 955, in run_commands\n    self.run_command(cmd)\n  File \"\/usr\/lib\/python3.6\/distutils\/dist.py\", line 974, in run_command\n    cmd_obj.run()\n  File \"\/usr\/lib\/python3\/dist-packages\/wheel\/bdist_wheel.py\", line 204, in run\n    self.run_command('build')\n  File \"\/usr\/lib\/python3.6\/distutils\/cmd.py\", line 313, in run_command\n    self.distribution.run_command(command)\n  File \"\/usr\/lib\/python3.6\/distutils\/dist.py\", line 974, in run_command\n    cmd_obj.run()\n  File \"\/usr\/lib\/python3.6\/distutils\/command\/build.py\", line 135, in run\n    self.run_command(cmd_name)\n  File \"\/usr\/lib\/python3.6\/distutils\/cmd.py\", line 313, in run_command\n    self.distribution.run_command(command)\n  File \"\/usr\/lib\/python3.6\/distutils\/dist.py\", line 974, in run_command\n    cmd_obj.run()\n  File \"setup.py\", line 40, in run\n    self.build_extension(ext)\n  File \"setup.py\", line 82, in build_extension\n    subprocess.check_call(['cmake', ext.sourcedir] + cmake_args, cwd=self.build_temp, env=env)\n  File \"\/usr\/lib\/python3.6\/subprocess.py\", line 311, in check_call\n    raise CalledProcessError(retcode, cmd)\nsubprocess.CalledProcessError: Command '['cmake', '\/mnt\/home\/ivan\/second.pytorch\/second\/spconv', '-DCMAKE_PREFIX_PATH=\/home\/ivan\/.local\/lib\/python3.6\/site-packages\/torch', '-DPYBIND11_PYTHON_VERSION=3.6', '-DSPCONV_BuildTests=OFF', '-DCMAKE_CUDA_FLAGS=\"--expt-relaxed-constexpr\" -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__', '-DCMAKE_LIBRARY_OUTPUT_DIRECTORY=\/mnt\/home\/ivan\/second.pytorch\/second\/spconv\/build\/lib.linux-x86_64-3.6\/spconv', '-DCMAKE_BUILD_TYPE=Release']' returned non-zero exit status 1.\n<\/code><\/pre>\n\n<p>System: Ubuntu 18.04, CUDA 9.1, cuDNN 7.6.3<\/p>\n","answers":[{"AnswerId":"57873271","CreationDate":"2019-09-10T14:45:48.990","ParentId":null,"OwnerUserId":"10359704","Title":null,"Body":"<p>Found the problem. I needed to set <code>CUDNN_INCLUDE_DIR<\/code> to <code>\/usr\/lib\/cuda\/include<\/code> (i.e. where the <code>cudnn.h<\/code> file is located).<\/p>\n\n<p>Lesson learnt: take time to understand the error, <code>missing: CUDNN_INCLUDE_DIR<\/code> was key. <\/p>\n"}]}
+{"QuestionId":57857129,"AnswerCount":0,"Tags":"<python><keras><keras-layer>","CreationDate":"2019-09-09T15:35:56.767","AcceptedAnswerId":null,"OwnerUserId":12042388.0,"Title":"Create multiple AE KERAS models","Body":"<blockquote>\n  <p>Create a single ENCODER and DECODER, map the mapping to several variables, and learn. As a result, if you look at the predictions, it seems that all the models predict the same values. I think it's like sharing a LAYER, right?<\/p>\n<\/blockquote>\n\n<pre><code>    encoding_dim = np.int64(round((train.shape[1] - 1) \/ set_dimsize))\n    input_dim = normal_training.shape[1]\n    input_data = Input(shape=(input_dim,))\n    encoded = Dense(encoding_dim, activation=set_set_model_activation)(input_data)\n    encoded = Dropout(drop_out_size)(encoded)\n    decoded = Dense(input_dim, activation=set_set_model_activation)(encoded)\n    def mappingmodel(input_data, decoded):\n        model_a = Model(input_data, decoded)\n        model_b = Model(input_data, decoded)\n        model_c = Model(input_data, decoded)\n        model_d = Model(input_data, decoded)\n        model_e = Model(input_data, decoded)\n        model_f = Model(input_data, decoded)\n        model_g = Model(input_data, decoded)\n        return model_a, model_b, model_c, model_d, model_e, model_f, model_g\n\n    autoencoder, label_2_autoencoder, label_3_autoencoder, label_4_autoencoder, label_5_autoencoder, label_6_autoencoder, label_7_autoencoder = mappingmodel(input_data, decoded)\n\n<\/code><\/pre>\n\n<blockquote>\n  <p>So I created several ENCODERs and DECODERs and mapped each one. Is there a more efficient way?<\/p>\n<\/blockquote>\n\n<pre><code>\n    #Improved Code\n\n    encoding_dim = np.int64(round((train.shape[1] - 1) \/ set_dimsize))\n    input_dim = normal_training.shape[1]\n    input_data = Input(shape=(input_dim,))\n    encoded = Dense(encoding_dim, activation=set_set_model_activation)(input_data)\n    encoded = Dropout(drop_out_size)(encoded)\n    decoded = Dense(input_dim, activation=set_set_model_activation)(encoded)\n\n    input_data2 = Input(shape=(input_dim,))\n    encoded2 = Dense(encoding_dim, activation=set_set_model_activation)(input_data2)\n    encoded2 = Dropout(drop_out_size)(encoded2)\n    decoded2 = Dense(input_dim, activation=set_set_model_activation)(encoded2)\n\n    input_data3 = Input(shape=(input_dim,))\n    encoded3 = Dense(encoding_dim, activation=set_set_model_activation)(input_data3)\n    encoded3 = Dropout(drop_out_size)(encoded3)\n    decoded3 = Dense(input_dim, activation=set_set_model_activation)(encoded3)\n\n    .\n    .\n    .  \n\n    input_data7 = Input(shape=(input_dim,))\n    encoded7 = Dense(encoding_dim, activation=set_set_model_activation)(input_data7)\n    encoded7 = Dropout(drop_out_size)(encoded7)\n    decoded7 = Dense(input_dim, activation=set_set_model_activation)(encoded7)\n\n\n\n    def mappingmodel(input_data, decoded):\n        model_a = Model(input_data, decoded)\n        model_b = Model(input_data2, decoded2)\n        model_c = Model(input_data3, decoded3)\n        model_d = Model(input_data4, decoded4)\n        model_e = Model(input_data5, decoded5)\n        model_f = Model(input_data6, decoded6)\n        model_g = Model(input_data7, decoded7)\n        return model_a, model_b, model_c, model_d, model_e, model_f, model_g\n\n    autoencoder, label_2_autoencoder, label_3_autoencoder, label_4_autoencoder, label_5_autoencoder, label_6_autoencoder, label_7_autoencoder = mappingmodel(input_data, decoded)\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57857169,"AnswerCount":0,"Tags":"<tensorflow><element>","CreationDate":"2019-09-09T15:38:46.280","AcceptedAnswerId":null,"OwnerUserId":11979471.0,"Title":"removing a single element from an arbitrary axis in a tensor","Body":"<p>I want to create a new tensor y from a tensor x, where the element \"idx\" at axis \"axis\" is removed.<\/p>\n\n<p>This is what I've come up with:<\/p>\n\n<pre><code>def delete_element(x, idx, axis):\n\n    mask = tf.one_hot(idx, tf.shape(x)[axis], on_value=0, off_value=1)\n    y = tf.boolean_mask(x, mask, axis=axis)\n\n    return y\n<\/code><\/pre>\n\n<p>This works, but feels unneccessary complicated. Any ideas on how to simplify this?<\/p>\n","answers":[]}
+{"QuestionId":57857325,"AnswerCount":1,"Tags":"<c++><tensorflow>","CreationDate":"2019-09-09T15:50:00.753","AcceptedAnswerId":57858229.0,"OwnerUserId":9873377.0,"Title":"How to load custom op library when using c++?","Body":"<p>I have build a very simple custom op <code>zero_out.dll<\/code> with bazel, it works when using python.<\/p>\n\n<pre><code>import tensorflow as tf\nzero_out_module = tf.load_op_library('.\/zero_out.dll')\nwith tf.Session(''):\n  zero_out_module.zero_out([[1, 2], [3, 4]]).eval()\n<\/code><\/pre>\n\n<p>However I have to run the inference using C++, is there any c++ api which has the similar function as <code>tf.load_op_library<\/code>, as it seems that a lot of registeration work has been done in <code>tf.load_op_library<\/code>, TF has not counterpart c++ API?<\/p>\n","answers":[{"AnswerId":"57858229","CreationDate":"2019-09-09T16:54:53.410","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>While there does not seem to be a public API for that in C++, the library loading functions are exposed in the <a href=\"https:\/\/www.tensorflow.org\/install\/lang_c\" rel=\"nofollow noreferrer\">TensorFlow API for C<\/a> (which is the API that <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/load_library\" rel=\"nofollow noreferrer\"><code>tf.load_library<\/code><\/a> uses). There is no \"nice\" documentation for it, but you can find them in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.14.0\/tensorflow\/c\/c_api.h#L1626-L1655\" rel=\"nofollow noreferrer\"><code>c\/c_api.h<\/code><\/a>:<\/p>\n\n<pre class=\"lang-c prettyprint-override\"><code>\/\/ --------------------------------------------------------------------------\n\/\/ Load plugins containing custom ops and kernels\n\n\/\/ TF_Library holds information about dynamically loaded TensorFlow plugins.\ntypedef struct TF_Library TF_Library;\n\n\/\/ Load the library specified by library_filename and register the ops and\n\/\/ kernels present in that library.\n\/\/\n\/\/ Pass \"library_filename\" to a platform-specific mechanism for dynamically\n\/\/ loading a library. The rules for determining the exact location of the\n\/\/ library are platform-specific and are not documented here.\n\/\/\n\/\/ On success, place OK in status and return the newly created library handle.\n\/\/ The caller owns the library handle.\n\/\/\n\/\/ On failure, place an error status in status and return NULL.\nTF_CAPI_EXPORT extern TF_Library* TF_LoadLibrary(const char* library_filename,\n                                                 TF_Status* status);\n\n\/\/ Get the OpList of OpDefs defined in the library pointed by lib_handle.\n\/\/\n\/\/ Returns a TF_Buffer. The memory pointed to by the result is owned by\n\/\/ lib_handle. The data in the buffer will be the serialized OpList proto for\n\/\/ ops defined in the library.\nTF_CAPI_EXPORT extern TF_Buffer TF_GetOpList(TF_Library* lib_handle);\n\n\/\/ Frees the memory associated with the library handle.\n\/\/ Does NOT unload the library.\nTF_CAPI_EXPORT extern void TF_DeleteLibraryHandle(TF_Library* lib_handle);\n<\/code><\/pre>\n\n<p>These functions do actually call C++ code (see source in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.14.0\/tensorflow\/c\/c_api.cc#L977-L995\" rel=\"nofollow noreferrer\"><code>c\/c_api.cc<\/code><\/a>). However, the called functions, defined in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.14.0\/tensorflow\/core\/framework\/load_library.cc\" rel=\"nofollow noreferrer\"><code>core\/framework\/load_library.cc<\/code><\/a> does not have a header to include. The workaround to use it in C++ code, which they use in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.14.0\/tensorflow\/c\/c_api.cc#L751-L753\" rel=\"nofollow noreferrer\"><code>c\/c_api.cc<\/code><\/a>, is to declare the function yourself, and link the TensorFlow library.<\/p>\n\n<pre class=\"lang-cpp prettyprint-override\"><code>namespace tensorflow {\n\/\/ Helpers for loading a TensorFlow plugin (a .so file).\nStatus LoadLibrary(const char* library_filename, void** result,\n                   const void** buf, size_t* len);\n}\n<\/code><\/pre>\n\n<p>As far as I can tell there is no API to unload the library. The C API allows you only to delete the library handle object. This done just by freeing the pointer, but if you want to avoid trouble you should probably use the freeing function given by TensorFlow, <code>tensorflow::port:free<\/code>, declared in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.14.0\/tensorflow\/core\/platform\/mem.h#L34\" rel=\"nofollow noreferrer\"><code>core\/platform\/mem.h<\/code><\/a>. Again, if you cannot not or don't want to include that, you can declare the function yourself and it should work as well.<\/p>\n\n<pre class=\"lang-cpp prettyprint-override\"><code>namespace tensorflow {\nnamespace port {\nvoid Free(void* ptr);\n}\n}\n<\/code><\/pre>\n"}]}
+{"QuestionId":57857683,"AnswerCount":0,"Tags":"<python><numpy><keras><deep-learning><spyder>","CreationDate":"2019-09-09T16:14:53.880","AcceptedAnswerId":null,"OwnerUserId":4959180.0,"Title":"AttributeError: 'ProgbarLogger' object has no attribute 'log_values'","Body":"<p>I have tried to run this <a href=\"https:\/\/github.com\/jocicmarko\/ultrasound-nerve-segmentation\" rel=\"nofollow noreferrer\">segmentation model<\/a> using spyder.\nWhen I run <code>data.py<\/code>, I obtain this message<\/p>\n\n<blockquote>\n  <p>File \"C:\/Users\/achaire\/Downloads\/Compressed\/ultrasound-nerve-segmentation-master\/ultrasound-nerve-segmentation-master\/data.py\", line 19, in create_train_data\n      imgs = np.ndarray((total, image_rows, image_cols), dtype=np.uint8)<\/p>\n  \n  <p>TypeError: 'float' object cannot be interpreted as an integer<\/p>\n<\/blockquote>\n\n<p>The code is:<\/p>\n\n<pre><code>**14** def create_train_data():\n**15** train_data_path = os.path.join(data_path, 'train')\n**16** images = os.listdir(train_data_path)\n**17** total = len(images) \/ 2\n\n**18** imgs = np.ndarray((total, image_rows, image_cols), dtype=np.uint8)\n**19** imgs_mask = np.ndarray((total, image_rows, image_cols), dtype=np.uint8)\n<\/code><\/pre>\n\n<p>I have replaced line <strong>17<\/strong> by <code>total = int(len(images) \/ 2)<\/code> and <code>np.uint8<\/code> by <code>float<\/code> in lines <strong>18<\/strong> and <strong>19<\/strong>. The problem is solved.<\/p>\n\n<p>When I run <code>train.py<\/code> I have this message <\/p>\n\n<pre><code>runfile('C:\/Users\/achaire\/Downloads\/Compressed\/ultrasound-nerve-segmentation-master\/ultrasound-nerve-segmentation-master\/train.py', wdir='C:\/Users\/achaire\/Downloads\/Compressed\/ultrasound-nerve-segmentation-master\/ultrasound-nerve-segmentation-master')\n------------------------------\nLoading and preprocessing train data...\n------------------------------\n------------------------------\nCreating and compiling model...\n------------------------------\nC:\\Users\\achaire\\Anaconda3\\lib\\site-packages\\numpy\\core\\fromnumeric.py:3118: RuntimeWarning: Mean of empty slice.\n  out=out, **kwargs)\nC:\\Users\\hamdi\\Anaconda3\\lib\\site-packages\\numpy\\core\\_methods.py:140: RuntimeWarning: Degrees of freedom &lt;= 0 for slice\n  keepdims=keepdims)\n------------------------------\nFitting model...\n------------------------------\nC:\/Users\/achaire\/Downloads\/Compressed\/ultrasound-nerve-segmentation-master\/ultrasound-nerve-segmentation-master\/train.py:119: UserWarning: The `nb_epoch` argument in `fit` has been renamed `epochs`.\n  callbacks=[model_checkpoint])\nTrain on 0 samples, validate on 0 samples\nEpoch 1\/20\nTraceback (most recent call last):\n\n  File \"&lt;ipython-input-15-f713d62eb4dc&gt;\", line 1, in &lt;module&gt;\n    runfile('C:\/Users\/achaire\/Downloads\/Compressed\/ultrasound-nerve-segmentation-master\/ultrasound-nerve-segmentation-master\/train.py', wdir='C:\/Users\/achaire\/Downloads\/Compressed\/ultrasound-nerve-segmentation-master\/ultrasound-nerve-segmentation-master')\n\n  File \"C:\\Users\\achaire\\Anaconda3\\lib\\site-packages\\spyder_kernels\\customize\\spydercustomize.py\", line 827, in runfile\n    execfile(filename, namespace)\n\n  File \"C:\\Users\\achaire\\Anaconda3\\lib\\site-packages\\spyder_kernels\\customize\\spydercustomize.py\", line 110, in execfile\n    exec(compile(f.read(), filename, 'exec'), namespace)\n\n  File \"C:\/Users\/achaire\/Downloads\/Compressed\/ultrasound-nerve-segmentation-master\/ultrasound-nerve-segmentation-master\/train.py\", line 153, in &lt;module&gt;\n    train_and_predict()\n\n  File \"C:\/Users\/achaire\/Downloads\/Compressed\/ultrasound-nerve-segmentation-master\/ultrasound-nerve-segmentation-master\/train.py\", line 119, in train_and_predict\n    callbacks=[model_checkpoint])\n\n  File \"C:\\Users\\achaire\\Anaconda3\\lib\\site-packages\\keras\\engine\\training.py\", line 1039, in fit\n    validation_steps=validation_steps)\n\n  File \"C:\\Users\\achaire\\Anaconda3\\lib\\site-packages\\keras\\engine\\training_arrays.py\", line 217, in fit_loop\n    callbacks.on_epoch_end(epoch, epoch_logs)\n\n  File \"C:\\Users\\achaire\\Anaconda3\\lib\\site-packages\\keras\\callbacks.py\", line 79, in on_epoch_end\n    callback.on_epoch_end(epoch, logs)\n\n  File \"C:\\Users\\achaire\\Anaconda3\\lib\\site-packages\\keras\\callbacks.py\", line 338, in on_epoch_end\n    self.progbar.update(self.seen, self.log_values)\n\nAttributeError: 'ProgbarLogger' object has no attribute 'log_values'\n<\/code><\/pre>\n\n<p>I have the last version of <code>anaconda<\/code>, <code>keras<\/code> and <code>python<\/code> ...<\/p>\n","answers":[]}
+{"QuestionId":57857690,"AnswerCount":0,"Tags":"<python><pytorch><autograd>","CreationDate":"2019-09-09T16:15:09.283","AcceptedAnswerId":null,"OwnerUserId":635387.0,"Title":"pytorch autograd obstructs script from terminating","Body":"<p>Whenever I call autograds <code>backward<\/code> my script never terminates. <code>backward<\/code>is not per se blocking, all lines after it are still executed, the script just does not terminate. It appears that there is some sort of working thread in the background which hangs, but I was not able to find any information about it.<\/p>\n\n<p>I originally encountered the problem while training neural networks, however I eventually found a very short example with the same behavior:<\/p>\n\n<pre><code>import torch\n\nx = torch.randn(3, requires_grad=True)\ny = x * 2\nprint(y)\n\ngradients = torch.tensor([0.1, 1.0, 0.0001], dtype=torch.float)\ny.backward(gradients)\nprint(x.grad)\n\nprint(\"all done\")\n<\/code><\/pre>\n\n<p>When I remove the <code>backward<\/code> line, the script finishes as expected. Otherwise I see a process call <code>python<\/code> in the task manager, if I terminate it by hand, the script execution also terminates.<\/p>\n\n<p>I installed pytorch on Windows 7 using conda (<code>conda create --name grad_test pytorch -c pytorch<\/code>) in the most recent, stable version (python 3.7, pytorch 1.2.0).<\/p>\n","answers":[]}
+{"QuestionId":57857871,"AnswerCount":0,"Tags":"<r><keras>","CreationDate":"2019-09-09T16:27:31.133","AcceptedAnswerId":null,"OwnerUserId":8959427.0,"Title":"Keras model - model specification and applying a trained model to test data","Body":"<p>I have some <code>text<\/code> data and a multi-class <code>class<\/code> variable which is labeled <code>1<\/code>, <code>2<\/code> and <code>3<\/code>. The <code>text<\/code> data is called <code>x_train<\/code> and the <code>class<\/code> data is called <code>y_train<\/code>. I have a very hard classification task (think of something like trying to predict the stock market hard) so I do not expect to obtain high accuracy from the model but I was hoping for something a little more than 33% with is no better than flipping a coin since I have 3 classes. The best model I have ran (using TF-IDF) obtains 0.45 accuracy. I was hoping to obtain a little better accuracy using Keras.<\/p>\n\n<p>The documents I have are long and I set the <code>maxlen = 25000<\/code> and <code>num_words = 20000<\/code> and pad the sequences.<\/p>\n\n<p>Question 1)<\/p>\n\n<p>Do you know of any \"out of the box\" parameters which usually works well on your keras models I could try?<\/p>\n\n<p>I have the following model:<\/p>\n\n<pre><code>batch_size &lt;- 32\nembedding_dims &lt;- 50\nfilters &lt;- 250\nkernel_size &lt;- 3\nhidden_dims &lt;- 250\nepochs &lt;- 2\n\n\nmodel &lt;- keras_model_sequential() %&gt;%  \n\n  layer_embedding(max_features, embedding_dims, input_length = maxlen) %&gt;%\n  layer_dropout(0.2) %&gt;%\n  layer_conv_1d(\n    filters, kernel_size, \n    padding = \"valid\", activation = \"relu\", strides = 1\n  ) %&gt;%\n  layer_global_max_pooling_1d() %&gt;%\n  layer_dense(hidden_dims) %&gt;%\n  layer_dropout(0.2) %&gt;%\n  layer_activation(\"relu\") %&gt;%\n  layer_dense(3) %&gt;%\n  layer_activation(\"softmax\")\n\n\n# Compile model\nmodel %&gt;%\n  compile(\n    loss = \"categorical_crossentropy\",\n    optimizer = \"adam\",\n    metrics = \"accuracy\"\n  )\n\nhistory &lt;- model %&gt;%\n  keras::fit(\n    x_train,\n    to_categorical(y_train -1), # I set -1 here since keras expects 0, 1, 2 class and not 1, 2, 3 as I currently have.\n    batch_size = batch_size, \n    epochs = epochs, \n    validation_split = 0.3\n  )\n<\/code><\/pre>\n\n<p>Which gives me this:<\/p>\n\n<pre><code>Train on 18895 samples, validate on 8098 samples\nEpoch 1\/2\n18895\/18895 [==============================] - 710s 38ms\/sample - loss: 1.0991 - acc: 0.3344 - val_loss: 1.0985 - val_acc: 0.3405\nEpoch 2\/2\n18895\/18895 [==============================] - 708s 37ms\/sample - loss: 1.0988 - acc: 0.3359 - val_loss: 1.0992 - val_acc: 0.3269\n<\/code><\/pre>\n\n<p>I make predictions using:<\/p>\n\n<pre><code>preds &lt;- model %&gt;%\n  predict_classes(x_test, batch_size = 32)\n<\/code><\/pre>\n\n<p>Where ALL of the predictions are <code>0<\/code>.<\/p>\n\n<p>Training and loss plot:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/kVCAK.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/kVCAK.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57857939,"AnswerCount":0,"Tags":"<pytorch><typeerror>","CreationDate":"2019-09-09T16:32:19.687","AcceptedAnswerId":null,"OwnerUserId":11790554.0,"Title":"I have a trouble after trying to parallelize data using nn.Dataparallel","Body":"<p>I didn't have any probelm without data parallelization, but after I put JUST A ONE LINE \"model = nn.DataParallel(model)\" the error message \"TypeError: 'list' object is not callable\" comes. If I push out that damn line the source works clean. plz help me<\/p>\n\n<p>I can't do anything but just mad. I search that google and there are some ways to solve that error message but I can't do anything. Because nn.Dataparallel is already used by other coders. Sorry English is not my mothertongue.<\/p>\n\n<pre><code>    if use_cuda:\n            [[[model = nn.DataParallel(model)]]]\n            model = model.cuda()\n            criterion = criterion.cuda()\n            print('cuda is used')\n<\/code><\/pre>\n\n<p>I just put <code>model = nn.DataParallel(model)<\/code> and error comes right after.<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):\n    File \"\/home\/scrcdeep2\/YBJ\/espnet_myself\/Main.py\", line 119, in \n      train(loader, model_, criterion_, optimizer_, use_cuda_, pretrained=None)<\/p>\n  \n  <p>File \"\/home\/scrcdeep2\/YBJ\/espnet_myself\/Main.py\", line 83, in train\n      outputs = model(inputs)<\/p>\n  \n  <p>File \"\/home\/scrcdeep2\/YBJ\/lib\/python3.5\/site-packages\/torch\/nn\/modules\/module.py\", line 493, in <strong>call<\/strong>\n      result = self.forward(*input, **kwargs)\n   File \"\/home\/scrcdeep2\/YBJ\/lib\/python3.5\/site-packages\/torch\/nn\/parallel\/data_parallel.py\", line 151, in forward\n      replicas = self.replicate(self.module, self.device_ids[:len(inputs)])<\/p>\n  \n  <p>File \"\/home\/scrcdeep2\/YBJ\/lib\/python3.5\/site-packages\/torch\/nn\/parallel\/data_parallel.py\", line 156, in replicate\n      return replicate(module, device_ids)<\/p>\n  \n  <p>File \"\/home\/scrcdeep2\/YBJ\/lib\/python3.5\/site-packages\/torch\/nn\/parallel\/replicate.py\", line 114, in replicate\n      modules = list(network.modules())<\/p>\n  \n  <p>TypeError: 'list' object is not callable<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":57858127,"AnswerCount":2,"Tags":"<python-3.x><tensorflow><object-detection><mask><valueerror>","CreationDate":"2019-09-09T16:46:06.240","AcceptedAnswerId":null,"OwnerUserId":10393794.0,"Title":"Mask-RCNN, ValueError: could not broadcast input array from shape (70) into shape (1)","Body":"<p>I am working on a project that detects buildings on SpaceNet dataset by using Mask-RCNN. When I run this code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.train(dataset_train, dataset_val, \n            learning_rate=config.LEARNING_RATE, \n            epochs=1, \n            layers='heads')\n<\/code><\/pre>\n\n<h1>I got this  error:<\/h1>\n\n<pre><code>    1772             batch_images[b] = mold_image(image.astype(np.float32), config)\n\n-&gt; 1773             batch_gt_class_ids[b, :gt_class_ids.shape[0]] = gt_class_ids\n\n   1774             batch_gt_boxes[b, :gt_boxes.shape[0]] = gt_boxes\n\n   1775             batch_gt_masks[b, :, :, :gt_masks.shape[-1]] = gt_masks\n<\/code><\/pre>\n\n<h1>ValueError: could not broadcast input array from shape (70) into shape (1)<\/h1>\n\n<p>And also got this warning:\n<a href=\"https:\/\/i.stack.imgur.com\/aOPpD.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/aOPpD.jpg\" alt=\"Warning Messages\"><\/a><\/p>\n\n<p>When I debug the variables: 'gt_class_ids' and 'batch_gt_class_ids',\nI got this: <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/pPsZ8.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/pPsZ8.jpg\" alt=\"Debug Image\"><\/a><\/p>\n","answers":[{"AnswerId":"57858547","CreationDate":"2019-09-09T17:20:54.683","ParentId":null,"OwnerUserId":"7941815","Title":null,"Body":"<p><code>ValueError: could not broadcast input array from shape (70) into shape (1)<\/code> means that there is an array shape mismatch in the line 1773 <code>batch_gt_boxes[b, :gt_boxes.shape[0]] = gt_boxes<\/code> to be more specific your trying to broadcast values from an array of shape 70 to an array of shape 1. <\/p>\n\n<p>Can you run a debugger on this script and share the value of <code>gt_class_ids<\/code> and <code>batch_gt_class_ids<\/code>? This can help solve the problem better. Thanks<\/p>\n"},{"AnswerId":"57868811","CreationDate":"2019-09-10T10:25:41.390","ParentId":null,"OwnerUserId":"10393794","Title":null,"Body":"<p>As @Mahesh said, I debugged the variable <code>gt_class_ids<\/code> and realized that the shape of it was (x, 1). 'x' could be 70, 15 or whatever it is, the problem was about the <code>gt_class_ids.shape[0]<\/code> part. <code>gt_class_ids.shape[0]<\/code> gives us 'x', and x again can be any number. So I went to the <\/p>\n\n<blockquote>\n  <p>model.py <\/p>\n  \n  <p>(The direction of mine was: C:\\Users\\MUSTAFAAKTAS\\Desktop\\SpaceNet_MaskRCNN\\mrcnn\\model.py)<\/p>\n<\/blockquote>\n\n<p>file and changed to <code>batch_gt_class_ids[b, :gt_class_ids.shape[1]] = gt_class_ids<\/code> <\/p>\n\n<p>instead <code>batch_gt_class_ids[b, :gt_class_ids.shape[0]] = gt_class_ids<\/code>.<\/p>\n\n<p>So, it return '1' instead of 'x'.<\/p>\n\n<p>This solution worked for me :) <\/p>\n"}]}
+{"QuestionId":57858219,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-09T16:54:02.230","AcceptedAnswerId":null,"OwnerUserId":8033051.0,"Title":"`loss` passed to Optimizer.compute_gradients should be a function when eager execution is enabled","Body":"<p>I am new to TensorFlow, I just started learning and understanding it.\nI am working on <code>neural style transfer<\/code> problem and I am using tensorflow <code>version 1.14<\/code>.<\/p>\n\n<p>I am getting an error <strong>loss passed to Optimizer.compute_gradients should be a function when eager execution is enabled<\/strong>.<\/p>\n\n<p>I tried to solve the problem by using TensorFlow graph instead of <code>eager execution<\/code>, but it's not working. I want to use <code>eager execution<\/code> because it looks like a more pythonic way.<\/p>\n\n<p>here is my code, sorry for putting whole code here, please suggest corrections in my code.<\/p>\n\n<pre><code>import scipy\nimport tensorflow as tf\nimport tensorflow.contrib.eager as tfe\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom PIL import Image\nfrom scipy import misc\nfrom skimage.transform import resize\nfrom tensorflow.keras.applications.vgg19 import VGG19, preprocess_input\nfrom tensorflow.keras import backend as K\n\ntf.enable_eager_execution()\nprint('Eager execution {}'.format(tf.executing_eagerly()))\n\ncontent_path = '800px-Green_Sea_Turtle_grazing_seagrass.jpg'\nstyle_path = '800px-The_Great_Wave_off_Kanagawa.jpg'\n\ncontent_img = plt.imread(content_path)\nplt.imshow(content_img)\nstyle_img = plt.imread(style_path)\nplt.imshow(style_img)\n\nMEANS = np.array([123.68, 116.779, 103.939]).reshape((1,1,1,3))\ncontent_img = resize(content_img, (552,800,3)) #resized content img because style img has shape (552,800,3)\n\ncontent_img = np.array(content_img)\ncontent_img = np.reshape(content_img, ((1,)+content_img.shape))\nstyle_img = np.array(style_img)\nstyle_img = np.reshape(style_img, ((1,)+style_img.shape))\n\nnoise_img= np.random.uniform(-20,20,(1,552,800,3)).astype('float32')\ngenerated_img = noise_img*0.6 + content_img*0.4\nplt.imshow(generated_img[0])\n\ncontent_img = content_img-MEANS\nstyle_img = style_img-MEANS\n\nmodel = VGG19(include_top=False, weights='imagenet')\n\ndef compute_content_cost(act_content_img, act_generated_img):\n    return tf.reduce_mean(tf.square(act_content_img-act_generated_img))\n\ndef gram_matrix(A):\n    gram = tf.matmul(A, tf.transpose(A))\n    return gram\n\ndef style_loss_one_layer(act_style_img, act_generated_img):\n    m,n_H,n_W,n_C = tf.shape(act_generated_img)               #act_generated_img.get_shape().as_list()\n    gram_act_style_img = gram_matrix(act_style_img)\n    gram_generated_img = gram_matrix(act_generated_img)\n    return tf.reduce_mean(tf.square(gram_act_style_img-gram_generated_img))*(1\/(4*n_C**2*(n_H*n_W)**2))\n\ncontent_layer = ['block5_conv2']\nstyle_layers = [('block1_conv1',0.2), \n                ('block2_conv1',0.2),\n                ('block3_conv1',0.2),\n                ('block4_conv1',0.2),\n                ('block5_conv1',0.2)]\n\ndef compute_style_cost(model, style_layers):\n    style_cost = total_style_cost = 0\n    for layer, coeff in style_layers:\n        act_style_img = model.get_layer(layer).output\n        act_generated_img = model.get_layer(layer).output\n        style_cost += style_loss_one_layer(act_style_img, act_generated_img)\n        total_style_cost += coeff*style_cost \n    return total_style_cost\n\ndef compute_total_cost(J_content, J_style, alpha=10, beta=40):\n    J = (alpha*tf.cast(J_content, tf.float64)) + (beta*J_style)\n    return J\n\nact_generated_img = model.get_layer('block5_conv2').output\nact_content_img = model.get_layer('block5_conv2').output\n\nJ_content = compute_content_cost(act_content_img=act_content_img, act_generated_img=act_generated_img)\nprint(J_content)\nJ_style = compute_style_cost(model, style_layers=style_layers)\nprint(J_style)\n\nJ_total_cost = compute_total_cost(J_content, J_style, alpha=10, beta=40)\nprint(J_total_cost)\n\noptimizer = tf.train.AdamOptimizer(2.0)\n\ntrain_step = optimizer.minimize(J_total_cost)        #**getting error here**\n<\/code><\/pre>\n","answers":[{"AnswerId":"58952358","CreationDate":"2019-11-20T10:40:28.883","ParentId":null,"OwnerUserId":"2115644","Title":null,"Body":"<p>The above error is mainly caused when you are trying to use TensorFlow 1.x but the system is running tensor 2.0. <\/p>\n\n<p>Initialise the TensorFlow using the code below to ensure you are trying to use the version 1.0<\/p>\n\n<blockquote>\n  <p>import tensorflow.compat.v1 as tf<\/p>\n<\/blockquote>\n\n<p>You can make the system disable that behaviour by the below command after the initialisers.<\/p>\n\n<blockquote>\n  <p>tf.disable_v2_behavior()<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":57858433,"AnswerCount":2,"Tags":"<python><pytorch><jupyter>","CreationDate":"2019-09-09T17:12:11.577","AcceptedAnswerId":null,"OwnerUserId":9178303.0,"Title":"How to clear GPU memory after PyTorch model training without restarting kernel","Body":"<p>I am training PyTorch deep learning models on a Jupyter-Lab notebook, using CUDA on a Tesla K80 GPU to train. While doing training iterations, the 12 GB of GPU memory are used. I finish training by saving the model checkpoint, but want to continue using the notebook for further analysis (analyze intermediate results, etc.).<\/p>\n\n<p>However, these 12 GB continue being occupied (as seen from <code>nvtop<\/code>) after finishing training. I would like to free up this memory so that I can use it for other notebooks.<\/p>\n\n<p>My solution so far is to restart this notebook's kernel, but that is not solving my issue because I can't continue using the same notebook and its respective output computed so far.<\/p>\n","answers":[{"AnswerId":"57858523","CreationDate":"2019-09-09T17:18:36.850","ParentId":null,"OwnerUserId":"2111837","Title":null,"Body":"<p>Never worked with PyTorch myself, but Google has several results which all basically say the same.. <code>torch.cuda.empty_cache()<\/code><\/p>\n\n<p><a href=\"https:\/\/forums.fast.ai\/t\/clearing-gpu-memory-pytorch\/14637\" rel=\"nofollow noreferrer\">https:\/\/forums.fast.ai\/t\/clearing-gpu-memory-pytorch\/14637<\/a><\/p>\n\n<p><a href=\"https:\/\/discuss.pytorch.org\/t\/how-can-we-release-gpu-memory-cache\/14530\" rel=\"nofollow noreferrer\">https:\/\/discuss.pytorch.org\/t\/how-can-we-release-gpu-memory-cache\/14530<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/55322434\/how-to-clear-cuda-memory-in-pytorch\">How to clear Cuda memory in PyTorch<\/a><\/p>\n"},{"AnswerId":"57860310","CreationDate":"2019-09-09T19:40:20.020","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>If you just set object that uses a lot of memory to <code>None<\/code> like this:<\/p>\n\n<pre><code>obj = None\n<\/code><\/pre>\n\n<p>And after that you call <\/p>\n\n<pre><code>gc.collect() # Python thing\n<\/code><\/pre>\n\n<p>This is how you may avoid restarting the notebook.<\/p>\n\n<hr>\n\n<p>If you still would like to see it clear from Nvidea smi or nvtop you may run:<\/p>\n\n<pre><code>torch.cuda.empty_cache() # PyTorch thing\n<\/code><\/pre>\n\n<p>to empty the PyTorch cache.<\/p>\n"}]}
+{"QuestionId":57858740,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-09T17:36:48.710","AcceptedAnswerId":null,"OwnerUserId":8624169.0,"Title":"Integration with tensorflow","Body":"<p>I want to calculate integral of Deep Neural Network with respect to input and use it in my loss function.<\/p>\n\n<p>I tried two different scheme.\nFirst, I used tf.gradients to go around the problem. (this method leads to noisy function approximation)\nSecond, I used tf.integral but it is not what I am looking for. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x_ph = tf.placeholder(tf.float32, [None,1])\ny_ph = tf.placeholder(tf.float32, [None,1])\ny_pred = mlp(x_ph)\n\ndef multilayer_perceptron(x,weights, biases):\n    # *********************** Hidden layer 1 ***********************#\n    layer_1 = tf.add(tf.matmul(x, weights['h1']), biases['b1'])\n\n\n    return tf.reshape(tf.convert_to_tensor(tf.gradients(layer_1, y)),(-1,1)), layer_1\n# I want to have my ouput as return layer_1, integral(layer_1,x)\n\ndef cost(x, y, z):\n    jx, fx = multilayer_perceptron(x,weights, biases, kp)\n    fx = tf.reshape(fx, (-1,1))\n    jx =  tf.reshape(jx, (-1,1))\n    cost_ = tf.reduce_mean(tf.abs(fx-y)) + tf.reduce_mean(tf.abs(jx-z))\n\n\n    return cost_ \n\n<\/code><\/pre>\n","answers":[{"AnswerId":"57860341","CreationDate":"2019-09-09T19:44:13.470","ParentId":null,"OwnerUserId":"8624169","Title":null,"Body":"<p>Here is my solution. I defined an extra placeholder t and used it to calculate the value of integral at x. <\/p>\n\n<p><a href=\"https:\/\/github.com\/moradza\/Practice-and-Test\/blob\/master\/Anti%20derivative%20calculator%20of%20Tensorflow%20network%201D%20input.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/moradza\/Practice-and-Test\/blob\/master\/Anti%20derivative%20calculator%20of%20Tensorflow%20network%201D%20input.ipynb<\/a> to see the code. <\/p>\n"}]}
+{"QuestionId":57858794,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-09T17:40:37.493","AcceptedAnswerId":null,"OwnerUserId":10083113.0,"Title":"Why doesn't a Lambda function that returns the result of tf.py_func\/tf.numpy_function have an output shape, and how can I correct this behavior?","Body":"<p>For example purposes, I've created two Python functions. One takes a value and just returns it; the other is a lambda function that returns a Tensor to evaluate the function:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nfrom tensorflow.keras import Sequential\nfrom tensorflow.keras.layers import Input,Lambda,Dense\ndef test_func(value):\n  return value\ntest_func = np.vectorize(test_func)\ntest_op = lambda x: tf.numpy_function(test_func,[x],tf.float32)\ninput = Input(shape=(1,))\ntest_layer = Lambda(test_op)\nmodel = Sequential()\nmodel.add(input)\nmodel.add(test_layer)\nmodel.summary()\n<\/code><\/pre>\n\n<p>When I run this code on Tensorflow 1.14.x + Keras 2.2.x + Python 3.6.x, and even if I set the output_shape parameter of the Lambda layer, I get a Lambda layer with a \"None\" output shape:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>Model: \"sequential_1\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nlambda_1 (Lambda)            None                      0         \n=================================================================\nTotal params: 0\nTrainable params: 0\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>When I test the code using a test tensor, I get the results I might expect:<\/p>\n\n<pre><code>with tf.Session() as session:\n    X = tf.constant([[1.0,2.0],[3.0,4.0]])\n    Y = tf.constant(2.0)\n    X_out = test_op(X)\n    Y_out = test_op(Y)\n    print(\"X:\",session.run(X))\n    print(\"X shape: \",session.run(tf.shape(X_out)))\n    print(\"Y:\",session.run(Y_out))\n    print(\"Y shape: \", session.run(tf.shape(Y_out)))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<blockquote>\n  <p>X: [[1. 2.]  [3. 4.]]\n  X shape:  [2 2]\n  Y: 2.0\n  Y shape:  <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/Lambda\" rel=\"nofollow noreferrer\">1<\/a>\n  <strong>EDIT:<\/strong><\/p>\n<\/blockquote>\n\n<p>For comparison purposes, I also created a neural network using the function found in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/Lambda\" rel=\"nofollow noreferrer\">this page of documentation<\/a>:<\/p>\n\n<pre><code>from tensorflow.keras import Model\nfrom tensorflow.keras.layers import Input,Lambda\ninput = Input(shape=(1,))\ndef antirectifier(x):\n    x -= K.mean(x, axis=1, keepdims=True)\n    x = K.l2_normalize(x, axis=1)\n    pos = K.relu(x)\n    neg = K.relu(-x)\n    return K.concatenate([pos, neg], axis=1)\nlambda_ = Lambda(antirectifier)\nlambda_ = Lambda(antirectifier)(input)\nmodel = Model(inputs=[input],outputs=[lambda_])\nmodel.summary()\n<\/code><\/pre>\n\n<p>As somewhat expected, this prints:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>Model: \"model\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_1 (InputLayer)         [(None, 1)]               0         \n_________________________________________________________________\nlambda_1 (Lambda)            (None, 2)                 0         \n=================================================================\nTotal params: 0\nTrainable params: 0\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>Now - if I understand correctly - both tf.py_func and K.concatenate return a TensorFlow Tensor object. So, the question is, Why does all the code I posted here work, EXCEPT my tf.numpy_function layer?<\/p>\n","answers":[{"AnswerId":"57958832","CreationDate":"2019-09-16T14:19:56.623","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>The problem lies in <code>np.vectorize<\/code>.   <\/p>\n\n<p>It says it uses numpy's broadcasting rules, and it might have something to do with that. <\/p>\n\n<p>Check it out:<\/p>\n\n<pre><code>def test_func(value):\n    return value\n\ntest_func = np.vectorize(test_func)\nprint(test_func(2))\nprint(test_func(2).shape)\n<\/code><\/pre>\n\n<p>Notice the result is <code>()<\/code>. So, Keras\/Tensorflow is working fine as they are getting the output shape given by numpy. Numpy probably doesn't know what it's getting as an input (because it doesn't work with symbolic tensors).   <\/p>\n\n<p>Notice how it works normally if you pass an actual numpy array to the function:    <\/p>\n\n<pre><code>print(test_func(np.array([2])))\nprint(test_func(np.array([2])).shape)\n\n&gt;&gt;&gt;&gt; [2]\n&gt;&gt;&gt;&gt; (1,)\n<\/code><\/pre>\n\n<p>But when Keras initializes a model, the actual numbers don't exist, and Numpy has no idea of what it will receive when you are \"using\" the model. <\/p>\n"}]}
+{"QuestionId":57859248,"AnswerCount":1,"Tags":"<tensorflow><keras><distributed-computing>","CreationDate":"2019-09-09T18:16:45.003","AcceptedAnswerId":null,"OwnerUserId":12043123.0,"Title":"Distributed Tensorflow Using fit_generator()","Body":"<p>Is it possible to use Tensorflow in a distributed manner and use the fit_generator()? In my research so far I have not seen anything on how to do this or if it is possible. If it is not possible then what are some possible solutions to use distributed Tensorflow when all the data will not fit in memory.<\/p>\n","answers":[{"AnswerId":"58607177","CreationDate":"2019-10-29T12:07:31.023","ParentId":null,"OwnerUserId":"11270789","Title":null,"Body":"<p>Using fit_generator() is not possible under a tensorflow distribution scope. \nhave a lookt at <a href=\"https:\/\/www.tensorflow.org\/guide\/data\" rel=\"nofollow noreferrer\">tf.data<\/a>. i rewrote all my Keras ImageDataGenerators to a tensorflow data pipeline. doesn't need much time, is more transparent and quite remarkably faster.<\/p>\n"}]}
+{"QuestionId":57859262,"AnswerCount":0,"Tags":"<keras><attention-model>","CreationDate":"2019-09-09T18:17:35.617","AcceptedAnswerId":null,"OwnerUserId":12043029.0,"Title":"Cannot get the expected target shape in keras","Body":"<p>I am new to sequential models. I am working on an image caption generator with attention model in Keras. <\/p>\n\n<p>I keep getting an error with the expected target shape in Keras. I have worked on basic models before. And for this kind of error, usually there's a mistake with the way I process my dataset.\nHowever, in this case, I have tried adjusting the shape of my 'y' array by unpacking it, or by trying to pack my 'outputs' list. But it doesn't change the error message.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def model(photo_shape, max_len, n_s, vocab_size):\n  outputs=list()\n  seq=Input(shape=(max_len,), name='inseq')#max_len is 33.\n  x = Embedding(vocab_size, 300,mask_zero=True)(seq)  \n\n  p=Input(shape=(photo_shape[0],photo_shape[1]),name='picture')\n  s,_,c=LSTM(n_s,return_state = True)(p)\n  for t in range(max_len):\n    context = attention(p,s)\n    word=Lambda(lambda x: x[:,t,:])(x)\n    context = concat([word,context])\n    context=reshape(context)\n    s,_,c = lstm(context)\n    out = den2(s) #return a dense layer with vocab_size units(none, 2791).\n    outputs.append(out)\n    #print(np.array(outputs).shape) =&gt; returns (33,)\n  model = Model(inputs=[seq,p],outputs=outputs)\n  return model\n\n#the following method goes to a generator function.\ndef build_sequences(tokenizer, max_length, desc_list, photo):\n  X1, X2, y = list(), list(), list()\n  desc = desc_list[0]\n  seq = tokenizer.texts_to_sequences([desc])[0]\n  l=len(seq)\n  in_seq, out_seq = seq[:l-1], seq[1:l]\n  in_seq = pad_sequences([in_seq],padding='post', maxlen=max_length)[0]\n  out_seq =[to_categorical([w], num_classes=vocab_size)[0] for w in out_seq]\n  out_seq = pad_sequences([out_seq],padding='post', maxlen=max_length)[0]\n  X1.append(in_seq)\n  X2.append(photo)\n  y.append(out_seq)\n  return np.array(X1), np.array(X2), np.array(y)\n<\/code><\/pre>\n\n<p>The error I get is this.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><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 33 array(s), but instead got the following list of 1 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, ..., 0, 0, 0],\n       [0, 0, 0, ..., 0, 0, 0],\n       [0, 0, 0, ..., 0, 0,...\n<\/code><\/pre>\n\n<p>The debug messages show the following shapes for my inputs and outputs.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>in_seq: (1, 33)\nphoto: (1, 196, 512)\nout_seq: (1, 33, 2971)\n<\/code><\/pre>\n\n<p>The first is the batch size, which should not be taken into consideration. So, I do not know why the 33 is not visible to keras. I have tried modifying this shape. But logically, should this not work?<\/p>\n\n<p>Please let me know if this is a data processing error or a problem with my model structure!<\/p>\n\n<p>Let me know if more code is required<\/p>\n","answers":[]}
+{"QuestionId":57860309,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><openal>","CreationDate":"2019-09-09T19:40:18.903","AcceptedAnswerId":null,"OwnerUserId":7665537.0,"Title":"problem while using RLLib - using using existing code","Body":"<p>I am getting an error trying to run existing code that uses RLlib but getting import error.<\/p>\n\n<p>I have installed the latest version of RlLib but that hasn't fixed the issue.<\/p>\n\n<pre><code>from __future__ import absolute_import\nfrom __future__ import division\nfrom __future__ import print_function\n\nimport ray\nfrom ray.rllib.tests.test_multi_agent_env import make_multiagent\nfrom ray.tune import run_experiments\nfrom ray.tune.registry import register_env\n\nif __name__ == \"__main__\":\n    ray.init()\n    MultiPendulum = make_multiagent(\"Pendulum-v0\")\n    register_env(\"multi_pend\", lambda _: MultiPendulum(1))\n    trials = run_experiments({\n        \"test\": {\n            \"run\": \"PPO\",\n            \"env\": \"multi_pend\",\n            \"stop\": {\n                \"timesteps_total\": 500000,\n                \"episode_reward_mean\": -200,\n            },\n            \"config\": {\n                \"train_batch_size\": 2048,\n                \"vf_clip_param\": 10.0,\n                \"num_workers\": 0,\n                \"num_envs_per_worker\": 10,\n                \"lambda\": 0.1,\n                \"gamma\": 0.95,\n                \"lr\": 0.0003,\n                \"sgd_minibatch_size\": 64,\n                \"num_sgd_iter\": 10,\n                \"model\": {\n                    \"fcnet_hiddens\": [64, 64],\n                },\n                \"batch_mode\": \"complete_episodes\",\n            },\n        }\n    })\n    if trials[0].last_result[\"episode_reward_mean\"] &lt; -200:\n        raise ValueError(\"Did not get to -200 reward\", trials[0].last_result)\n<\/code><\/pre>\n\n<p>I am getting the following output when running this<\/p>\n\n<pre><code>ImportError: cannot import name 'try_import_tf' from 'ray.rllib.utils' (\/\/anaconda3\/envs\/Causal\/lib\/python3.7\/site-packages\/ray\/rllib\/utils\/__init__.py)\n<\/code><\/pre>\n\n<p>Any help how to fix this issue would be appreciated.<\/p>\n","answers":[]}
+{"QuestionId":57860358,"AnswerCount":1,"Tags":"<python><tensorflow><keras><package><anaconda>","CreationDate":"2019-09-09T19:47:12.830","AcceptedAnswerId":null,"OwnerUserId":5098109.0,"Title":"Why can't I install packages in Anaconda?","Body":"<p>I installed Anaconda from this <a href=\"https:\/\/repo.anaconda.com\/archive\/Anaconda3-2019.07-Windows-x86_64.exe\" rel=\"nofollow noreferrer\">link<\/a> and when I try to install other packages from Anaconda Navigator, I get the following error:<\/p>\n\n<pre><code>UnsatisfiableError: The following specifications were found to be incompatible with each other:\n\n\n\nPackage _tflow_select conflicts for:\nkeras -&gt; tensorflow -&gt; _tflow_select[version='==2.1.0|==2.2.0|==2.3.0',build='mkl|gpu|eigen']\ntensorflow-eigen -&gt; _tflow_select==2.2.0=eigen\ntensorflow-hub -&gt; tensorflow[version='&gt;=1.7.0'] -&gt; _tflow_select[version='==2.1.0|==2.2.0|==2.3.0',build='mkl|gpu|eigen']\ntensorflow-metadata -&gt; tensorflow -&gt; _tflow_select[version='==2.1.0|==2.2.0|==2.3.0',build='mkl|gpu|eigen']\ntensorflow-gpu -&gt; tensorflow==1.9.0 -&gt; _tflow_select[version='==2.2.0|==2.3.0',build='mkl|eigen']\nkeras-gpu -&gt; tensorflow-gpu -&gt; tensorflow==1.9.0 -&gt; _tflow_select[version='==2.2.0|==2.3.0',build='mkl|eigen']\ntensorflow-probability -&gt; tensorflow[version='&gt;=1.14.0'] -&gt; _tflow_select[version='==2.1.0|==2.2.0|==2.3.0',build='mkl|gpu|eigen']\ntensorflow -&gt; _tflow_select[version='==2.1.0|==2.2.0|==2.3.0',build='mkl|gpu|eigen']\ntensorflow-datasets -&gt; tensorflow[version='&gt;=1.14'] -&gt; _tflow_select[version='==2.1.0|==2.2.0|==2.3.0',build='mkl|gpu|eigen']\ntensorflow-mkl -&gt; _tflow_select==2.3.0=mkl\ntensorflow-gpu -&gt; _tflow_select==2.1.0=gpu\nkeras-gpu -&gt; tensorflow-gpu -&gt; _tflow_select==2.1.0=gpu\nPackage tensorflow-base conflicts for:\ntensorflow-mkl -&gt; tensorflow==1.14.0 -&gt; tensorflow-base[version='==1.10.0|==1.11.0|==1.12.0|==1.13.1|==1.13.1|==1.14.0|==1.14.0',build='mkl_py36ha978198_0|mkl_py36hcaf7020_0|mkl_py36h81393da_0|mkl_py36h81393da_0|mkl_py36h81393da_0|mkl_py35h81393da_0|mkl_py37hcaf7020_0|mkl_py37ha978198_0']\nkeras -&gt; tensorflow -&gt; tensorflow-base[version='==1.10.0|==1.11.0|==1.11.0|==1.11.0|==1.12.0|==1.12.0|==1.12.0|==1.13.1|==1.13.1|==1.13.1|==1.13.1|==1.13.1|==1.13.1|==1.13.1|==1.13.1|==1.14.0|==1.14.0|==1.14.0|==1.14.0|==1.14.0|==1.14.0|==1.14.0|==1.14.0|==1.9.0|==1.9.0|==1.9.0|==1.9.0',build='gpu_py35h6e53903_0|eigen_py36h45df0d8_0|gpu_py37h55fc52a_0|gpu_py36h55fc52a_0|gpu_py37h871c8ca_0|gpu_py36h871c8ca_0|gpu_py36h0fff12a_0|eigen_py37hf8af7b3_0|eigen_py36hf8af7b3_0|mkl_py36h81393da_0|gpu_py36h6e53903_0|eigen_py36h45df0d8_0|mkl_py36h81393da_0|eigen_py36h45df0d8_0|gpu_py36h6e53903_0|eigen_py35h45df0d8_0|eigen_py36h45df0d8_0|gpu_py35h6e53903_0|mkl_py35h81393da_0|mkl_py36h81393da_0|gpu_py36h6e53903_0|gpu_py37h0fff12a_0|mkl_py36hcaf7020_0|mkl_py37hcaf7020_0|eigen_py36hdbc3f0e_0|eigen_py37hdbc3f0e_0|gpu_py36h9ee611f_0|gpu_py37h9ee611f_0|mkl_py36ha978198_0|mkl_py37ha978198_0|eigen_py35h45df0d8_0|gpu_py36h6e53903_0']\ntensorflow-estimator -&gt; tensorflow-base[version='&gt;=1.14.0,&lt;1.15.0a0']\ntensorflow -&gt; tensorflow-base[version='==1.10.0|==1.11.0|==1.11.0|==1.11.0|==1.12.0|==1.12.0|==1.12.0|==1.13.1|==1.13.1|==1.13.1|==1.13.1|==1.13.1|==1.13.1|==1.13.1|==1.13.1|==1.14.0|==1.14.0|==1.14.0|==1.14.0|==1.14.0|==1.14.0|==1.14.0|==1.14.0|==1.9.0|==1.9.0|==1.9.0|==1.9.0',build='gpu_py35h6e53903_0|eigen_py36h45df0d8_0|gpu_py37h55fc52a_0|gpu_py36h55fc52a_0|gpu_py37h871c8ca_0|gpu_py36h871c8ca_0|gpu_py36h0fff12a_0|eigen_py37hf8af7b3_0|eigen_py36hf8af7b3_0|mkl_py36h81393da_0|gpu_py36h6e53903_0|eigen_py36h45df0d8_0|mkl_py36h81393da_0|eigen_py36h45df0d8_0|gpu_py36h6e53903_0|eigen_py35h45df0d8_0|eigen_py36h45df0d8_0|gpu_py35h6e53903_0|mkl_py35h81393da_0|mkl_py36h81393da_0|gpu_py36h6e53903_0|gpu_py37h0fff12a_0|mkl_py36hcaf7020_0|mkl_py37hcaf7020_0|eigen_py36hdbc3f0e_0|eigen_py37hdbc3f0e_0|gpu_py36h9ee611f_0|gpu_py37h9ee611f_0|mkl_py36ha978198_0|mkl_py37ha978198_0|eigen_py35h45df0d8_0|gpu_py36h6e53903_0']\nkeras-gpu -&gt; tensorflow-gpu -&gt; tensorflow==1.9.0 -&gt; tensorflow-base==1.14.0=gpu_py37h9ee611f_0 -&gt; tensorflow-base[version='&gt;=1.14.0,&lt;1.15.0a0']\ntensorflow-eigen -&gt; tensorflow==1.9.0 -&gt; tensorflow-base==1.13.1=eigen_py37hf8af7b3_0 -&gt; tensorflow-base[version='&gt;=1.14.0,&lt;1.15.0a0']\ntensorflow-datasets -&gt; tensorflow[version='&gt;=1.14'] -&gt; tensorflow-estimator[version='&gt;=1.14.0,&lt;1.15.0'] -&gt; tensorflow-base[version='&gt;=1.14.0,&lt;1.15.0a0']\ntensorflow-mkl -&gt; tensorflow==1.14.0 -&gt; tensorflow-base==1.14.0=mkl_py37ha978198_0 -&gt; tensorflow-base[version='&gt;=1.14.0,&lt;1.15.0a0']\ntensorflow -&gt; tensorflow-estimator[version='&gt;=1.14.0,&lt;1.15.0'] -&gt; tensorflow-base[version='&gt;=1.14.0,&lt;1.15.0a0']\ntensorflow-base\n...\n<\/code><\/pre>\n\n<p>I mention that I have only the root environment. Does anybody know what these conflicts mean and how to resolve this issue?<\/p>\n","answers":[{"AnswerId":"57860392","CreationDate":"2019-09-09T19:50:18.043","ParentId":null,"OwnerUserId":"10180185","Title":null,"Body":"<p>you have to install into a new environment instead of the conda root environment.<\/p>\n"}]}
+{"QuestionId":57860614,"AnswerCount":1,"Tags":"<keras><lstm>","CreationDate":"2019-09-09T20:09:50.547","AcceptedAnswerId":null,"OwnerUserId":735008.0,"Title":"Specifying the batch size when subclassing keras.model","Body":"<p>I implement a model including a LSTM layer by subclassing the keras.Model.\nThe following is the code I used.<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras\n\nclass SequencePredictor(keras.Model):\n    def __init__(self, cell_size):\n        super(SequencePredictor, self).__init__(self)\n        self._mask = keras.layers.Masking(\n            mask_value=0.0)\n        self._lstm = keras.layers.LSTM(cell_size, return_sequences=True,\n                stateful=True)\n        self._dense = tf.keras.layers.Dense(1)\n\n    def call(self, inputs, training=False):\n        out = self._mask(inputs)\n        out = self._lstm(out)\n        return self._dense(out)\n\n\nsequence_predictor = SequencePredictor(cell_size=10)\ntrain_dataset = DataQueue(5000, 500).dataset\n\nsequence_predictor.compile(\n        loss=\"mean_squared_error\",\n        optimizer=keras.optimizers.Adam(lr=0.05),\n        metrics=['mse'])\nsequence_predictor.fit(train_dataset, epochs=50)\n<\/code><\/pre>\n\n<p>The above code doesn't work with the following error message.<\/p>\n\n<pre><code>  ValueError: If a RNN is stateful, it needs to know its batch size. Specify the batch size of your input tensors: \n    - If using a Sequential model, specify the batch size by passing a `batch_input_shape` argument to your first layer.\n    - If using the functional API, specify the batch size by passing a `batch_shape` argument to your Input layer.\n<\/code><\/pre>\n\n<p>The problem is that the above code is neither a sequential nor a functional model. I tried to specify batch_input_shape in the first Masking layer, but it doesn't work.<\/p>\n\n<p>How to resolve the above error? I am currently using tensorflow 2.0 rc0.<\/p>\n","answers":[{"AnswerId":"57978072","CreationDate":"2019-09-17T16:02:30.197","ParentId":null,"OwnerUserId":"12080653","Title":null,"Body":"<p>Build your LSTM layer with the desired input shape after instanciating it.<\/p>\n\n<pre><code>self._lstm = keras.layers.LSTM(cell_size, return_sequences=True, stateful=True)\nself._lstm.build((input_shape))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57860663,"AnswerCount":0,"Tags":"<python><keras><conv-neural-network><google-colaboratory>","CreationDate":"2019-09-09T20:15:47.233","AcceptedAnswerId":null,"OwnerUserId":9766348.0,"Title":"Building a cnn. I have weights from a github file in hdf5 format which i am Unable to load it","Body":"<p>Unable to upload the weights in colab notebook.Saved in hdf5 format\nDownloaded the hdf5 file in my drive in colab folder.Still not helping\nIt would reduce the training time for my cnn image classification model\nThe weights are from a github account.<\/p>\n\n<pre><code># define hyper parameters\nlearning_rate = 0.01\nweight_decay = 1e-6\nmomentum = 0.9\n\n# define weights and build model\nweights = 'all_cnn_weights_0.9088_0.4994.hdf5'\nmodel = allcnn(weights)\n\n# define optimizer and compile model\nsgd = SGD(lr=learning_rate, decay=weight_decay, momentum=momentum, \nnesterov=True)\nmodel.compile(loss='categorical_crossentropy', optimizer=sgd, metrics= \n['accuracy'])\n\n# print model summary\nprint (model.summary())\n\n# test the model with pretrained weights\nscores = model.evaluate(x_test, Y_test, verbose=1)\nprint(\"Accuracy: %.2f%%\" % (scores[1]*100))\n<\/code><\/pre>\n\n<p>Error message:<\/p>\n\n<pre><code>OSError                                   Traceback (most recent call \nlast)\n&lt;ipython-input-30-4b939a359e11&gt; in &lt;module&gt;()\n      5 # define weights and build model\n      6 weights = 'weight.hdf5'\n----&gt; 7 model = allcnn(weights)\n      8 \n      9 # define optimizer and compile model\n\n 4 frames\n \/usr\/local\/lib\/python3.6\/dist-packages\/h5py\/_hl\/files.py in \n make_fid(name, mode, userblock_size, fapl, fcpl, swmr)\n     140         if swmr and swmr_support:\n     141             flags |= h5f.ACC_SWMR_READ\n --&gt; 142         fid = h5f.open(name, flags, fapl=fapl)\n     143     elif mode == 'r+':\n     144         fid = h5f.open(name, h5f.ACC_RDWR, fapl=fapl)\n\n h5py\/_objects.pyx in h5py._objects.with_phil.wrapper()\n\n h5py\/_objects.pyx in h5py._objects.with_phil.wrapper()\n\n  h5py\/h5f.pyx in h5py.h5f.open()\n\nOSError: Unable to open file (unable to open file: name = 'weight.hdf5', \nerrno = 2, error message = 'No such file or directory', flags = 0, \n o_flags = 0)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57860689,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-09T20:18:12.667","AcceptedAnswerId":null,"OwnerUserId":10736997.0,"Title":"Tensorflow tf.train.get_checkpoint_state returns None despite being pointed to directory with a checkpoint file","Body":"<p>I am having a problem with retrieving the checkpoint. The function tf.train.get_checkpoint_state always returns None, despite the input being pointed to the correct directory. <\/p>\n\n<pre><code>tf.global_variables_initializer()\n  with tf.Session() as sess:\n    ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)\n    if ckpt and ckpt.model_checkpoint_path:\n      if os.path.isabs(ckpt.model_checkpoint_path):\n        # Restores from checkpoint with absolute path.\n        saver.restore(sess, ckpt.model_checkpoint_path)\n<\/code><\/pre>\n\n<p>Contents of the checkpoint file :<\/p>\n\n<pre><code>model_checkpoint_path: \"model.ckpt-29999\"\nall_model_checkpoint_paths: \"model.ckpt-0\"\nall_model_checkpoint_paths: \"model.ckpt-2000\"\nall_model_checkpoint_paths: \"model.ckpt-4000\"\nall_model_checkpoint_paths: \"model.ckpt-6000\"\nall_model_checkpoint_paths: \"model.ckpt-8000\"\nall_model_checkpoint_paths: \"model.ckpt-10000\"\nall_model_checkpoint_paths: \"model.ckpt-12000\"\nall_model_checkpoint_paths: \"model.ckpt-14000\"\nall_model_checkpoint_paths: \"model.ckpt-16000\"\nall_model_checkpoint_paths: \"model.ckpt-18000\"\nall_model_checkpoint_paths: \"model.ckpt-20000\"\nall_model_checkpoint_paths: \"model.ckpt-22000\"\nall_model_checkpoint_paths: \"model.ckpt-24000\"\nall_model_checkpoint_paths: \"model.ckpt-26000\"\nall_model_checkpoint_paths: \"model.ckpt-28000\"\nall_model_checkpoint_paths: \"model.ckpt-29999\"\n<\/code><\/pre>\n\n<p>Contents of the checkpoint directory:<\/p>\n\n<pre><code>checkpoint                            model.ckpt-18000.data-00000-of-00001  model.ckpt-26000.meta\nevents.out.tfevents.1567695112.lab0   model.ckpt-18000.index                model.ckpt-28000.data-00000-of-00001\nmodel.ckpt-0.data-00000-of-00001      model.ckpt-18000.meta                 model.ckpt-28000.index\nmodel.ckpt-0.index                    model.ckpt-20000.data-00000-of-00001  model.ckpt-28000.meta\nmodel.ckpt-0.meta                     model.ckpt-20000.index                model.ckpt-29999.data-00000-of-00001\nmodel.ckpt-10000.data-00000-of-00001  model.ckpt-20000.meta                 model.ckpt-29999.index\nmodel.ckpt-10000.index                model.ckpt-2000.data-00000-of-00001   model.ckpt-29999.meta\nmodel.ckpt-10000.meta                 model.ckpt-2000.index                 model.ckpt-4000.data-00000-of-00001\nmodel.ckpt-12000.data-00000-of-00001  model.ckpt-2000.meta                  model.ckpt-4000.index\nmodel.ckpt-12000.index                model.ckpt-22000.data-00000-of-00001  model.ckpt-4000.meta\nmodel.ckpt-12000.meta                 model.ckpt-22000.index                model.ckpt-6000.data-00000-of-00001\nmodel.ckpt-14000.data-00000-of-00001  model.ckpt-22000.meta                 model.ckpt-6000.index\nmodel.ckpt-14000.index                model.ckpt-24000.data-00000-of-00001  model.ckpt-6000.meta\nmodel.ckpt-14000.meta                 model.ckpt-24000.index                model.ckpt-8000.data-00000-of-00001\nmodel.ckpt-16000.data-00000-of-00001  model.ckpt-24000.meta                 model.ckpt-8000.index\nmodel.ckpt-16000.index                model.ckpt-26000.data-00000-of-00001  model.ckpt-8000.meta\nmodel.ckpt-16000.meta                 model.ckpt-26000.index                training_loss.txt\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57860720,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><keras>","CreationDate":"2019-09-09T20:21:22.790","AcceptedAnswerId":null,"OwnerUserId":3457177.0,"Title":"ML: How to create a simple machine learning model","Body":"<p>I'm new to machine learning. I try to make the simplest model but different from all \"hello world\" tutorials. <\/p>\n\n<p>I don't have a very big data sets so I decided to train the model that will determine if the input number is odd or even. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\n# create a model that defines if number is odd or even\ndef generate_data(length, is_random=False):\n    res = []\n    for i in range(0, length):\n        number = i if not is_random else random.randrange(0, 99999)\n\n        if number % 2 != 0:\n            res.append([number, 1])\n        else:\n            res.append([number, 0])\n    return res\n\n\ntrain_data = generate_data(5000, is_random=True)\n\ndata_set = np.array(train_data)\nrandom_number = data_set[:, 0:1]\nis_random_number_even = data_set[:, 1]\n\nprint('X: {}'.format(random_number))\nprint('Y: {}'.format(is_random_number_even))\n\nmodel = Sequential()\nmodel.add(Dense(3, input_dim=1, activation='relu'))\nmodel.add(Dense(3, activation='sigmoid'))\nmodel.add(Dense(4, activation='sigmoid'))\nmodel.add(Dense(1, activation='sigmoid'))\n\n# compile the keras model\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n\n# fit the keras model on the dataset\nmodel.fit(random_number, is_random_number_even, epochs=100, batch_size=30)\n\n_, accuracy = model.evaluate(random_number, is_random_number_even)\nprint('Accuracy: %.2f' % (accuracy*100))\n<\/code><\/pre>\n\n<p>In this code I generate an input data to train model, create a model with 2 internal layers and with different number of nodes (it's written in tutorial that there is no rules for number of layers and notes in every layer) and try to train it. <\/p>\n\n<p>The problem is that the accuracy is always about 50%, sometimes I had 52% when I change the number of layers and nodes in every layer but never more. It seems that despite of a lot of reading about this topic I'm still missing something. <\/p>\n\n<p>Can somebody tell me what's wrong here and what is a correct direction to proceed?<\/p>\n","answers":[]}
+{"QuestionId":57861016,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-09-09T20:49:48.537","AcceptedAnswerId":null,"OwnerUserId":2962001.0,"Title":"Advances MaxPooling2D Keras, count based instead of pixel value","Body":"<p>I am trying to implement a custom Pooling layers in Keras with Tensorflow (tf) backend. My goal is to implement a count-based max-pooling layers that returns the count of the number of values smaller than the current pixel value.<\/p>\n\n<p>I am trying to use the Keras custom layer definition, however, I cannot get the count function to work.<\/p>\n","answers":[]}
+{"QuestionId":57861021,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><transfer-learning><data-augmentation>","CreationDate":"2019-09-09T20:50:27.880","AcceptedAnswerId":57861509.0,"OwnerUserId":7065438.0,"Title":"How to use augumented data when using transfer learning?","Body":"<p>I have used VGG16 for transfer learning and got very low accuracy. Is it possible to use data augmentation technique to increase the accuracy when using transfer learning? <\/p>\n\n<pre><code>Following is the code for better understanding:\n\n\n# Show the image paths\ntrain_path = 'myNetDB\/train'      # Relative Path\nvalid_path = 'myNetDB\/valid'\ntest_path = 'myNetDB\/test'\n\ntrain_batches = ImageDataGenerator().flow_from_directory(train_path, target_size=(224, 224), classes=['dog', 'cat'], batch_size=10)\nvalid_batches = ImageDataGenerator().flow_from_directory(valid_path, target_size=(224, 224), classes=['dog', 'cat'], batch_size=4)\ntest_batches = ImageDataGenerator().flow_from_directory(test_path, target_size=(224, 224), classes=['dog', 'cat'], batch_size=10)\n\n\nvgg16_model= load_model('Fetched_VGG.h5') \n\n# transform the model to Sequential\nmodel= Sequential()\nfor layer in vgg16_model.layers[:-1]: \n    model.add(layer)\n\n# Freezing the layers (Oppose weights to be updated)\nfor layer in model.layers:\n    layer.trainable = False\n\n# adding the last layer\nmodel.add(Dense(2, activation='softmax'))\n\nmodel.compile(Adam(lr=.0001), loss='categorical_crossentropy', metrics=['accuracy'])\nmodel.fit_generator(train_batches, steps_per_epoch=4,\n                    validation_data=valid_batches, validation_steps=4, epochs=5, verbose=2)\n\n\npredictions = model.predict_generator(test_batches, steps=1, verbose=0)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57861509","CreationDate":"2019-09-09T21:37:25.580","ParentId":null,"OwnerUserId":"7943474","Title":null,"Body":"<p>If you got very low accuracy, it might be that your dataset is very different from the dataset VGG16 was trained on. There are two possibilities:<\/p>\n\n<ol>\n<li><p>your dataset is big enough such that you can train your model starting from the pre-trained weights.<\/p><\/li>\n<li><p>your dataset is small. In this case there are no shortcuts. You should consider a simpler model than VGG16 so that you're less likely to incur in overfitting.<\/p><\/li>\n<\/ol>\n\n<p>In both cases, to answer your question, yes, augmentation techniques, when done consciously, help increasing the accuracy. <\/p>\n"}]}
+{"QuestionId":57861344,"AnswerCount":2,"Tags":"<python><tensorflow><matrix><deep-learning>","CreationDate":"2019-09-09T21:21:03.933","AcceptedAnswerId":57863866.0,"OwnerUserId":7934786.0,"Title":"how to apply unique_with_counts over 2d array in tensorflow","Body":"<p>I have this tensor:<\/p>\n\n<pre><code>tf_a1 = [[0 3 1 22]\n        [3 5 2  2]\n        [2 6 3 13]\n        [1 7 0 3 ]\n        [4 9 11 10]]\n<\/code><\/pre>\n\n<p>What I want to do is to find the unique values being repeated more than a <code>threshold<\/code> across all columns.<\/p>\n\n<p>For example here, <code>3<\/code> repeated in <code>4 columns<\/code>. <code>0<\/code> repeated in <code>2 columns<\/code>. <code>2<\/code> repeated in <code>3 columns<\/code> and so on. <\/p>\n\n<p>I want my output be like this(suppose threshold is <code>2<\/code>, so indexes repeated more than 2 times will be masked).<\/p>\n\n<pre><code>[[F T F F]\n [T F T T]\n [T F T F]\n [F F F T]\n [F F F F]]\n<\/code><\/pre>\n\n<p>This is what I have done:<\/p>\n\n<pre><code>y, idx, count = tf.unique_with_counts(tf_a1)   \ntf.where(tf.where(count, tf_a1, tf.zeros_like(tf_a1)))\n<\/code><\/pre>\n\n<p>But it raises the error:<\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors_impl.InvalidArgumentError: unique\n  expects a 1D vector. [Op:UniqueWithCounts]<\/p>\n<\/blockquote>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"57862096","CreationDate":"2019-09-09T22:56:49.517","ParentId":null,"OwnerUserId":"7934786","Title":null,"Body":"<p>It seems <code>unique_with_count<\/code> should support multidimensional and axis, however I could not find anything in the documentation.<\/p>\n\n<p>The workaround for me was to first of al reshape my tensor to 1d array and then apply <code>unique_with_count<\/code> and then reshape it back to the original size:<\/p>\n\n<p>Thanks to @a_guest for sharing the idea<\/p>\n\n<pre><code>token_idx = tf.reshape(tf_a1, [-1])\ny, idx, count = tf.unique_with_counts(token_idx)\nmasked = tf.greater_equal(count, 2)\nbacked_same_size = tf.gather(masked, idx)\ntf.reshape(backed_same_size, shape=tf.shape(tf_a1))\n<\/code><\/pre>\n"},{"AnswerId":"57863866","CreationDate":"2019-09-10T03:59:19.440","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>Currently, the API of <code>unique_with_counts<\/code> that supports axis is not public yet. If you want to <code>unique_with_counts<\/code> with multi-dim tensor, you can call it like this in tf 1.14:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.python.ops import gen_array_ops\n\n# tensor 'x' is [[1, 0, 0],\n#                [1, 0, 0],\n#                [2, 0, 0]]\ny, idx, count = gen_array_ops.unique_with_counts_v2(x, [0])\ny ==&gt; [[1, 0, 0],\n       [2, 0, 0]]\nidx ==&gt; [0, 0, 1]\ncount ==&gt; [2, 1]\n<\/code><\/pre>\n\n<p>Since I cannot comment @sariii post. It should be <code>tf.greater(count,2)<\/code>. Also, I do not think this solution is satisfying the requirement of the problem. For example, if the first col are all 2 and the rest of col don't have 2. According to your requirement, 2 should be counted as 1 times. But in your solution, if you reshape to the 1d first, then you will lost this info. Please correct me if I'm wrong. <\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/16503#issuecomment-481831053\" rel=\"nofollow noreferrer\">issue#16503<\/a><\/p>\n"}]}
+{"QuestionId":57861467,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-09T21:33:09.430","AcceptedAnswerId":null,"OwnerUserId":10453757.0,"Title":"MNIST digit classification for digits only in the range [5,6,7,8,9]","Body":"<p>I'm trying to write a simple code for classifying MNIST dataset but only for digits 5,6,7,8,9. I wrote the code below. Upon running, I get an accuracy of 0 and loss of nan. I tried to do the same with digits 0,1,2,3,4 and it runs perfectly with accuracy of almost 99.4% on the training data (just changed the train_mask and test_mask below to [0,1,2,3,4]). Can someone help me to understand why the code doesn't do anything for digit classifying in the range 5,6,7,8,9? Thanks for the help in advance!<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport os\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Conv2D, Dropout, Flatten, MaxPooling2D\n\n(x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()\n\nnum_digits_to_classify = 5\n\ntrain_mask59 = np.isin(y_train,[5,6,7,8,9])\ntest_mask59 = np.isin(y_test,[5,6,7,8,9])\n\nx_train59, y_train59 = x_train[train_mask59], y_train[train_mask59]\nx_test59, y_test59 = x_test[test_mask59], y_test[test_mask59]\n\n# Reshaping the array to 4-dims so that it can work with the Keras API\nx_train59 = x_train59.reshape(x_train59.shape[0], 28, 28, 1)\nx_test59 = x_test59.reshape(x_test59.shape[0], 28, 28, 1)\ninput_shape = (28, 28, 1)\n# Making sure that the values are float so that we can get decimal points after division\nx_train59 = x_train59.astype('float32')\nx_test59 = x_test59.astype('float32')\n# Normalizing the RGB codes by dividing it to the max RGB value.\nx_train59 \/= 255\nx_test59 \/= 255\n\n\ncheckpoint_path = \"D:\/home\/work\/Fast_Learning\/training\/cp.ckpt59\"\ncheckpoint_dir = os.path.dirname(checkpoint_path)\n\ncp_callback = tf.keras.callbacks.ModelCheckpoint(checkpoint_path,\n                                                 save_weights_only=True,\n                                                 verbose=1)\n\ndef create_model():\n    model = Sequential()\n    model.add(Conv2D(28, kernel_size=(3,3), activation='relu', input_shape=input_shape))\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n    model.add(Flatten()) # Flattening the 2D arrays for fully connected layers\n    model.add(Dense(128, activation='relu'))\n    model.add(Dropout(0.2))\n    model.add(Dense(num_digits_to_classify,activation='softmax'))\n\n    model.compile(optimizer='adam', \n                  loss='sparse_categorical_crossentropy', \n                  metrics=['accuracy'])\n    return model\n\nmodel = create_model()\n\nmodel.fit(x=x_train59,y=y_train59, epochs=20,callbacks = [cp_callback])\n\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>Epoch 00019: saving model to D:\/home\/work\/MNIST\/training\/cp.ckpt59\nEpoch 20\/20\n29404\/29404 [==============================] - 3s 109us\/step - loss: nan - acc: 0.0000e+00\n<\/code><\/pre>\n","answers":[{"AnswerId":"57861607","CreationDate":"2019-09-09T21:49:54.567","ParentId":null,"OwnerUserId":"3767239","Title":null,"Body":"<p>If you use <code>sparse_categorical_crossentropy<\/code> the target labels are expected to be in the range <code>[0, num_digits_to_classify)<\/code>. When you select numbers 0 to 4 this is the case, but when you select 5 to 9 then they are offset by 5. Hence, before fitting, you should adjust the target labels:<\/p>\n\n<pre><code>y_train59 -= 5\ny_test59 -= 5\n<\/code><\/pre>\n\n<p>By the way, if you're using the <code>tensorflow.keras<\/code> module like so:<\/p>\n\n<pre><code>from tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import Dense, Conv2D, Dropout, Flatten, MaxPooling2D\n<\/code><\/pre>\n\n<p>it should generate a corresponding error message (not sure if <code>keras<\/code> does that as well):<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: Received a label value of 9 which is outside the valid range of [0, 5).\n<\/code><\/pre>\n\n<h2>For an arbitrary number of elements<\/h2>\n\n<p>You have multiple options to fit only a subset of the MNIST digits 0-9:<\/p>\n\n<ol>\n<li>Set the number of nodes in the last <code>Dense<\/code> layer to 10 (and just fit with fewer numbers),<\/li>\n<li>Compress the actually used digits (e.g. <code>[0, 1, 5, 6, 8]<\/code>) into the range <code>[0, N)<\/code>.<\/li>\n<\/ol>\n\n<p>For case 1)<\/p>\n\n<pre><code>model.add(Dense(10, activation='softmax'))  # last layer of model\n<\/code><\/pre>\n\n<p>For case 2)<\/p>\n\n<pre><code># Transform y_train (and similarly y_test).\nunique, index = np.unique(y_train, return_inverse=True)\ny_train = np.arange(len(unique))[index]\n# To get back the original labels, just index into the unique values.\nunique[y_train]\n<\/code><\/pre>\n\n<p>Note that <code>sparse_categorical_crossentropy<\/code> is for cases where your labels are integers and <code>categorical_crossentropy<\/code> is for cases where these labels are one-hot encoded. For example:<\/p>\n\n<pre><code>sparse_categorical_crossentropy: y = [0, 2, 1, 1, 2, 0]\n       categorical_crossentropy: y = [[1, 0, 0],\n                                      [0, 0, 1],\n                                      [0, 1, 0],\n                                      [0, 1, 0],\n                                      [0, 0, 1],\n                                      [1, 0, 0]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57861552,"AnswerCount":0,"Tags":"<keras><tensorflow2.0>","CreationDate":"2019-09-09T21:42:30.527","AcceptedAnswerId":null,"OwnerUserId":12043750.0,"Title":"can we train a model in tensorflow2.0 without using keras?","Body":"<p>I am trying to write a simple ML code to classify the mnist dataset in tensorflow2.0. I didn't use Keras for now since I just want to use lower API to help me understand how tensorflow works. However, after I defined the cross entropy, It seems impossible to continue. All the tf2.0 optimizers are moved to keras and I don't know how to train a model without keras in tf2.0. Is there a way that we bypass keras in tf2.0?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from __future__ import absolute_import, division, print_function, unicode_literals\n\nimport tensorflow as tf\nfrom tensorflow.keras import datasets, layers, models\n\n# Helper libraries\nimport numpy as np\nimport matplotlib.pyplot as plt\n\n\n(train_images, train_labels), (test_images, test_labels) = datasets.mnist.load_data()\n\nprint(train_images.shape)\nprint(len(train_labels))\nprint(train_images[1,:,:].shape)\n\n# plt.figure()\n# plt.imshow(train_images[0])\n# plt.colorbar()\n# plt.grid(False)\n# plt.show()\n\n# Normalize pixel values to be between 0 and 1\ntrain_images, test_images = train_images \/ 255.0, test_images \/ 255.0\n\nW = tf.Variable(tf.zeros([784, 10]))\nb = tf.Variable(tf.zeros([10]))\n\nfor i in range(1):\n    x = tf.constant(train_images[1,:,:].reshape(784), dtype = tf.float32) \n    x = tf.reshape(x, [1, 784])\n    print(tf.shape(x), tf.shape(W))\n    # define the model\n    y = tf.nn.softmax(tf.matmul(x, W) + b)\n    print(y)\n    # correct labels\n    y_ = np.zeros(10)\n    y_[train_labels[i]] = 1.0\n    y_ = tf.constant(y_, dtype = tf.float32)\n    y_ = tf.reshape(y_, [1, 10])\n    cross_entropy = -tf.reduce_sum(y_* tf.math.log(y))\n    print(cross_entropy)\n        I don't know how to continue from here.\n\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57861913,"AnswerCount":0,"Tags":"<python><tensorflow><keras><mnist><cleverhans>","CreationDate":"2019-09-09T22:29:37.847","AcceptedAnswerId":null,"OwnerUserId":9652636.0,"Title":"AttributeError : 'Conv2D' object has no attribute 'get_config'","Body":"<p>I'm trying to run fgsm attack on mnist dataset. I'm getting 'object has no attribute' error:<br>\nI used:\npython 2 (in google colab)\nKeras==2.1.3\ntensorflow==1.14.0<\/p>\n\n<p>I also run it with keras==1.2 and keras==2.1.5 ,\nI also chenged get_config into _get_config,\nHowever, all runs fail.<\/p>\n\n<p>Unit_keras.py is here: <a href=\"https:\/\/github.com\/tensorflow\/cleverhans\/blob\/master\/cleverhans\/utils_keras.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/cleverhans\/blob\/master\/cleverhans\/utils_keras.py<\/a><br>\nI also used : <a href=\"https:\/\/github.com\/tensorflow\/cleverhans\/blob\/6b065c4507f50e2545c009f75e82678f92353163\/cleverhans\/utils_keras.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/cleverhans\/blob\/6b065c4507f50e2545c009f75e82678f92353163\/cleverhans\/utils_keras.py<\/a><\/p>\n\n<pre><code>   x_adv_sub = fgsm.generate(images_tensor, **fgsm_par)\nFile \"\/content\/defensegan\/cleverhans\/attacks\/fast_gradient_method.py\", \n   line 50, in generate\n   labels, _nb_classes = self.get_or_guess_labels(x, kwargs)\nFile \"\/content\/defensegan\/cleverhans\/attacks\/attack.py\", line 278, in \n   get_or_guess_labels\n   preds = self.model.get_probs(x)\nFile \"\/content\/defensegan\/cleverhans\/utils_keras.py\", line 196, in \n   get_probs\n  :return: Names of all the layers kept by Keras\nFile \"\/content\/defensegan\/cleverhans\/utils_keras.py\", line 130, in \n  _get_softmax_name\n   def _get_abstract_layer_name(self):\nAttributeError: 'Conv2D' object has no attribute 'get_config'\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57862138,"AnswerCount":0,"Tags":"<numpy><tensorflow><keras>","CreationDate":"2019-09-09T23:03:00.783","AcceptedAnswerId":null,"OwnerUserId":11967117.0,"Title":"Errors in simple program to test Keras Backend","Body":"<p>Getting this error when running this file to test the Keras backend after installation to make sure all was set up to create my first Keras neural net. <\/p>\n\n<pre><code>ImportError: cannot import name '_ccallback_c' from 'scipy._lib' (C:\\Users\\AideenFay\\AppData\\Roaming\\Python\\Python37\\site-packages\\scipy\\_lib\\__init__.py)\n<\/code><\/pre>\n\n<p>Here is my progam to test the Keras backend <\/p>\n\n<pre><code>import numpy as np\nfrom keras import backend as kbe\n\ndata = kbe.variable(np.random.random((4,2)))\nzero_data= kbe.zeros_like(data)\nprint(kbe.eval(zero_data))\n<\/code><\/pre>\n\n<p>Any advice on this is recommended. Thank you. Might be a problem with overlapping Anaconda installations etc.<\/p>\n","answers":[]}
+{"QuestionId":57862207,"AnswerCount":0,"Tags":"<python><pytorch>","CreationDate":"2019-09-09T23:14:54.277","AcceptedAnswerId":null,"OwnerUserId":10855529.0,"Title":"How do I pass images from one neural network to another from top to bottom, and probabilities from bottom to top?","Body":"<p>desired result is top neural network outputs an image, which needs to be fed to two below neural networks, which feed image to below two neural networks, and then return probability back to the top, this implementation uses JoinProbs class to compute probability, but this probability needs to be done within Top, Base, Middle classes, when I attempt at calling forward method of Middle from Top, it does not give desired result.<\/p>\n\n<pre><code>class Top(nn.Module):\n  def __init__(self):\n    super().__init__()\n    self.encoder = nn.Sequential(nn.Conv2d(1, 1, 1), nn.Dropout(0.5))\n    self.lin = nn.Linear(10, 10)\n  def forward(self, x):\n    return self.encoder(x), self.encoder(x)\n\nnet = Top()\n\noutp1, outp2 = net(torch.randn(1, 1, 28, 28))\n\noutp1.shape, outp2.shape\n\nclass Middle(nn.Module):\n  def __init__(self):\n    super().__init__()\n    self.encoder = nn.Sequential(nn.Conv2d(1, 1, 1), nn.Dropout(0.5))\n    self.lin = nn.Linear(10, 10)\n  def forward(self, x):\n    return self.encoder(x), self.encoder(x)\n\nnet2 = Middle()\n\noutp11, outp12, outp21, outp22 = *net2(outp1), *net2(outp2)\n\ny = map(lambda x: x.shape, (outp11, outp12, outp21, outp22))\n\nlist(y)\n\nclass Base(nn.Module):\n  def __init__(self):\n    super().__init__()\n    self.lin = nn.Linear(784, 10)\n  def forward(self, x):\n    out = (x.view(x.size(0), -1))\n    return self.lin(out)\n\nnet3 = Base()\n\nprob11, prob12, prob21, prob22 = net3(outp11), net3(outp12), net3(outp21), net3(outp22)\n\ny = map(lambda x: x.shape, (prob11, prob12, prob21, prob22))\n\nlist(y)\n\nclass JoinProbs(nn.Module):\n  def __init__(self):\n    super().__init__()\n    self.lin = nn.Linear(20, 10)\n  def forward(self, *x):\n    out = torch.cat(x, dim=-1)\n    return self.lin(out)\n\nnet4 = JoinProbs()\n\na, b = net4(prob11, prob12), net4(prob21, prob22)\n\na, b\n\nnet4(a, b)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57862459,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><machine-learning><keras>","CreationDate":"2019-09-09T23:57:11.637","AcceptedAnswerId":null,"OwnerUserId":8843300.0,"Title":"Keras predict() returns a better accuracy than evaluate()","Body":"<p>I set up a model with Keras, then I trained it on a dataset of 3 records and finally I tested the resulting model with evaluate() and predict(), using the same test set for both functions (the test set has 100 records and it doesn't have any record of the training set, as much as it can be relevant, given the size of the two datasets).\nThe dataset is composed by 5 files, where 4 files represent each one a different temperature sensor, that each minute collects 60 measurements (each row contains 60 measurements), while the last file contains the class labels that I want to predict (in particular, 3 classes: 3, 20 or 100).<\/p>\n\n<p>This is the model I'm using:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>n_sensors, t_periods = 4, 60\n\nmodel = Sequential()\n\nmodel.add(Conv1D(100, 6, activation='relu', input_shape=(t_periods, n_sensors)))\n\nmodel.add(Conv1D(100, 6, activation='relu'))\n\nmodel.add(MaxPooling1D(3))\n\nmodel.add(Conv1D(160, 6, activation='relu'))\n\nmodel.add(Conv1D(160, 6, activation='relu'))\n\nmodel.add(GlobalAveragePooling1D())\n\nmodel.add(Dropout(0.5))\n\nmodel.add(Dense(3, activation='softmax'))\n\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>That I train:\n<code>self.model.fit(X_train, y_train, batch_size=3, epochs=5, verbose=1)<\/code><\/p>\n\n<p>Then I use evaluate:\n<code>self.model.evaluate(x_test, y_test, verbose=1)<\/code><\/p>\n\n<p>And predict:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>predictions = self.model.predict(data)\nresult = np.where(predictions[0] == np.amax(predictions[0]))\nif result[0][0] == 0:\n    return '3'\nelif result[0][0] == 1:\n    return '20'\nelse:\n    return '100'\n<\/code><\/pre>\n\n<p>For each class predicted, I confront it with the actual label, and then I calculate correct guesses \/ total examples, that should be equivalent to accuracy from the evaluate() function. Here's the code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>correct = 0\nfor profile in self.profile_file: #profile_file is an opened file\n    ts1 = self.ts1_file.readline()\n    ts2 = self.ts2_file.readline()\n    ts3 = self.ts3_file.readline()\n    ts4 = self.ts4_file.readline()\n    data = ts1, ts2, ts3, ts4\n    test_data = self.dl.transform(data) # see the last block of code I posted\n    prediction = self.model.predict(test_data)\n    if prediction == label:\n       correct += 1\nacc = correct \/ 100 # 100 is the number of total examples\n<\/code><\/pre>\n\n<p>Data feeded to evaluate() is taken from this function:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>label = pd.read_csv(os.path.join(self.testDir, 'profile.txt'), sep='\\t', header=None)\nlabel = np_utils.to_categorical(label[0].factorize()[0])\ndata = [os.path.join(self.testDir,'TS2.txt'),os.path.join(self.testDir, 'TS1.txt'),os.path.join(self.testDir,'TS3.txt'),os.path.join(self.testDir, 'TS4.txt')]\ndf = pd.DataFrame()\nfor txt in data:\n    read_df = pd.read_csv(txt, sep='\\t', header=None)\n    df = df.append(read_df)\ndf = df.apply(self.__predict_scale)\ndf = df.sort_index().values.reshape(-1,4,60).transpose(0,2,1)\nreturn df, label\n<\/code><\/pre>\n\n<p>While data feeded to predict() is taken from this other one:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>df = pd.DataFrame()\nfor txt in data: # data \n    read_df = pd.read_csv(StringIO(txt), sep='\\t', header=None)\n    df = df.append(read_df)\ndf = df.apply(self.__predict_scale)\ndf = df.sort_index().values.reshape(-1,4,60).transpose(0,2,1)\nreturn df\n<\/code><\/pre>\n\n<p>Accuracies yielded by evaluate() and predict() are always different: in particular, the maximum difference I noted was when evaluate() resulted in a 78% accuracy while predict() in a 95% accuracy. The only difference between the two functions is that I make predict() work on an example at a time, while evaluate() takes the entire dataset all at once, but it should result in no difference. How can it be?<\/p>\n\n<p><strong>UPDATE 1:<\/strong> It seems that the problem is in how I prepare my data. In the case of predict(), I transform only one line at a time from each file using the last block of code I posted, while in feeding evaluate(), I transform the entire files using the other function reported. Why should it be different? It seems to me that I'm applying the exact same transformation, the only difference is in the number of rows transformed.<\/p>\n","answers":[{"AnswerId":"57972395","CreationDate":"2019-09-17T10:32:13.093","ParentId":null,"OwnerUserId":"10273558","Title":null,"Body":"<p>This question was already answered <a href=\"https:\/\/stackoverflow.com\/questions\/45799474\/keras-model-evaluate-vs-model-predict-accuracy-difference-in-multi-class-nlp-ta?rq=1\">here<\/a><\/p>\n\n<p>what happens is when you evaluate the model, since your loss function is categorical_crossentropy, <code>metrics=['accuracy']<\/code> calculates categorical_accuracy.<\/p>\n\n<p>But predict has a default set to binary_accuracy.<\/p>\n\n<p>So essentially you are calculating categorical accuracy with evaluate and and binary accuracy with predict. this is the reason they are so widely different.<\/p>\n\n<p>the difference between categorical_accuracy and binary_accuracy is that categorical_accuracy check if all the outputs match with your y_test and binary_accuracy checks if each of you outputs matches with your y_test.<\/p>\n\n<p>Example(single row):<\/p>\n\n<pre><code>prediction = [0,0,1,1,0]\ny_test = [0,0,0,1,0]\n\ncategorical_accuracy = 0% \n<\/code><\/pre>\n\n<p><strong>since 1 output does not match the categorical_accuracy is 0<\/strong><\/p>\n\n<pre><code>binary_accuracy = 80% \n<\/code><\/pre>\n\n<p><strong>even though 1 output doesn't match the rest of 80% do match so accuracy is 80%<\/strong><\/p>\n"}]}
+{"QuestionId":57862505,"AnswerCount":1,"Tags":"<python><tensorflow><keras><keras-layer><differentiation>","CreationDate":"2019-09-10T00:06:44.837","AcceptedAnswerId":57871706.0,"OwnerUserId":12044023.0,"Title":"How do I find the non differentiable operation in my layer?","Body":"<p>I am trying to create a rather complex lambda-layer with many operations in keras. After I implemented it, I got a <code>ValueError: No gradients provided for any variable<\/code>. <\/p>\n\n<p>While I am using only keras operations to transform the data, (except for a constant I create using numpy which I later add onto a Tensor) I understand that there must be some operations which are not differentiable. Now I want to know how I can figure out which one it is, so I can find a workaround.<\/p>\n\n<p>I don't want publish any code yet as it is part of a competition and I want to figure this out on my own. If it is difficult to understand my problem because of that, please let me know. I can however give a list of all the functions I am using:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.keras import backend as K\nfrom tensorflow.python.keras.layers import Lambda\n\n...\ndef my_lambda_function(x):\n    # uses:\n    K.batch_dot\n    K.cast\n    K.clip\n    K.concatenate\n    K.one_hot\n    K.reshape\n    K.sum\n    K.tile  # only applied to a constant created in numpy\n\n...\n# using the function in a model like this:\nmy_lambda_layer = Lambda(my_lambda_function)\nresult_tensor = my_lambda_layer(some_input)\n\n<\/code><\/pre>\n\n<p>I think K.one_hot could be problematic, but I want a way to know this for sure before I try making it differentiable<\/p>\n","answers":[{"AnswerId":"57871706","CreationDate":"2019-09-10T13:18:08.087","ParentId":null,"OwnerUserId":"12044023","Title":null,"Body":"<p>After a few hours of sleep, here is my simple solution: Create a simple NN for testing and add a lambda layer in which I try out all functions seperately. This is however only an indirect way of finding the problem. Here is my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.python.keras.layers import Input, Conv2D, MaxPooling2D, UpSampling2D, Conv2DTranspose, Lambda\nfrom tensorflow.python.keras.models import Sequential\nfrom tensorflow.keras.datasets.mnist import load_data\n\nfrom tensorflow.keras import backend as K\n\nimport tensorflow as tf\nimport numpy as np\n\n(x_train, y_train), (x_test, y_test) = load_data()\nx_train, x_test = x_train \/ 255.0, x_test \/ 255.0\nx_train, x_test = np.reshape(x_train, (-1, 28, 28, 1)), np.reshape(x_test, (-1, 28, 28, 1))\n\n\ndef test_function(x):\n    x_int = K.cast(x, tf.int16)  # this was one of the gradient killers in my case\n    return K.cast(x_int, tf.float16)\n\n\nmodel = Sequential()\nmodel.add(Input(shape=(28, 28, 1)))\nmodel.add(Conv2D(10, (5, 5), padding='same', activation='relu'))\nmodel.add(MaxPooling2D())\nmodel.add(Lambda(test_function))\nmodel.add(UpSampling2D())\nmodel.add(Conv2DTranspose(4, (5, 5), padding='same', activation='relu'))\nmodel.add(Conv2DTranspose(1, (3, 3), padding='same', activation='sigmoid'))\n\nmodel.compile(optimizer='adam',\n              loss='mse',\n              metrics=['accuracy'])\n\nmodel.fit(x_train, x_train, epochs=5)\nmodel.evaluate(x_test, x_test)\n<\/code><\/pre>\n\n<p>This worked for me but I hope there are better solutions.<\/p>\n\n<p>Btw. I can approximate a floor operation (which is also killing the gradients) using these functions:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def a(x):\n    two_pi = 2 * math.pi\n    two_pi_x = x * two_pi\n    sine = K.sin(two_pi_x)\n    numerator = sine + two_pi_x\n    return numerator \/ two_pi\n\n\ndef approximated_floor(x):\n    x2 = a(a(a(x))) - 0.5\n    return x2\n<\/code><\/pre>\n"}]}
+{"QuestionId":57862702,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2019-09-10T00:48:07.737","AcceptedAnswerId":57862747.0,"OwnerUserId":1009508.0,"Title":"How to use a GlobalAveragePool layer as the output of a model","Body":"<p>I am receiving this error when I am trying to build a model using the output from another model.<\/p>\n\n<pre><code>Output tensors to a Model must be the output of a Keras `Layer` (thus holding past layer metadata). Found: &lt;keras.layers.pooling.GlobalAveragePooling2D object at 0x7Somthing or another\n<\/code><\/pre>\n\n<p>What I am trying to do is to use a finetuned model as the base model and retain the whole thing using a different method (an SCNN instead of a CNN). <\/p>\n\n<p>This is how I am declaring the model and where it goes wrong. <\/p>\n\n<pre><code>pair_base_model = Model(inputs = base_model.input, outputs=base_model.get_layer('glb_avg_pool'))\n<\/code><\/pre>\n\n<p>And this is how I load my previous model<\/p>\n\n<pre><code>base_model = load_model('..\/input\/base-model-reid\/0.ckpt')\nprint(base_model.summary())\n<\/code><\/pre>\n\n<p>Which gives me 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-html lang-html prettyprint-override\"><code>......... Whole bunch of other stuff\r\nblock14_sepconv2_act (Activatio (None, 7, 7, 2048)   0           block14_sepconv2_bn[0][0]        \r\n__________________________________________________________________________________________________\r\nglb_avg_pool (GlobalAveragePool (None, 2048)         0           block14_sepconv2_act[0][0]       \r\n__________________________________________________________________________________________________\r\ndense_1 (Dense)                 (None, 1502)         3077598     glb_avg_pool[0][0]               \r\n==================================================================================================\r\nTotal params: 23,939,078\r\nTrainable params: 10,980,862\r\nNon-trainable params: 12,958,216\r\n__________________________________________________________________________________________________\r\nNone<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n","answers":[{"AnswerId":"57862747","CreationDate":"2019-09-10T00:54:20.567","ParentId":null,"OwnerUserId":"1009508","Title":null,"Body":"<p>I'm stupoid I was working on this a little too late and found something simple that I messed up. \nI needed to use the .output for the layer in order to get the output I cannot just call .get_layer(xxxxxxx) as this won't work<\/p>\n\n<pre><code>pair_base_model = Model(inputs = base_model.input, outputs=base_model.get_layer('glb_avg_pool').output)\n<\/code><\/pre>\n\n<p>This is the working code.\nThanks for anyone who looked at this. <\/p>\n"}]}
+{"QuestionId":57863306,"AnswerCount":2,"Tags":"<python><tensorflow><machine-learning><keras><neural-network>","CreationDate":"2019-09-10T02:33:41.783","AcceptedAnswerId":null,"OwnerUserId":10761064.0,"Title":"Image - CNN clarification","Body":"<p>I want to clarify a few points in Convulutional Neural Network, I am implementing image recognition in CNN using Keras.<\/p>\n\n<p>1 -\nReducing size of images can help the model to learn faster. I read some blogs, in which they reduce the image size from  (150,150) to (32,32) . <\/p>\n\n<p>2 -\nIncreasing no. of layers and nodes can help increases the accuracy or not. I started to train the model at one CNN layer. But the accuracy is low. Then I add two more CNN layer, it gives high accuracy about 74. After that, I add one more layer it gives accuracy in the same range of 70s. <\/p>\n\n<p>3 -\nIs there a way to look images after every layer in CNN using keras.\nIt will help to study the image in CNN.<\/p>\n\n<p>Thank you<\/p>\n","answers":[{"AnswerId":"57863932","CreationDate":"2019-09-10T04:09:11.607","ParentId":null,"OwnerUserId":"9828197","Title":null,"Body":"<p>Here are my inputs regarding your questions:<\/p>\n\n<p>1) In an ideal world, you should be able to take a photo\/video of    anything with a camera as big as it fits your needs. However, this    approach is not practical, although can result in a quite good    performance. Actually, the more features you provide sans noise the    more accurate is your neural net. However, where will you use such a    network with such a big camera as its input generating machine? For    all practical purposes you use a camera with a lens of a size not    bigger than a tennis ball. At the same time, it is extremely    expensive to train images with sizes bigger than 64x64 pixels. You    can hardly afford a batch of 2 with a single GPU with any descent    neural net. At the same time, this will increase the training time    and you may end up waiting days for it to finish.<\/p>\n\n<p>2) The more layers do not produce better results and there are other    ways to achieve that. One problem with this is that the more layers    you have you face the exploding or vanishing gradient problem. Also    make sure you are regularizing your data, providing enough of    training data and that its distribution is similar to the valid\/test    set distribution. So as you can see there are many aspects to the    accuracy of a neural net and these were just some.<\/p>\n\n<p>3) Yes you can do it. It will take some time to write it thoroughly so I    am leaving you a link from medium. It is more detailed and has what    you need:<\/p>\n\n<p><a href=\"https:\/\/towardsdatascience.com\/visualizing-intermediate-activation-in-convolutional-neural-networks-with-keras-260b36d60d0\" rel=\"nofollow noreferrer\">https:\/\/towardsdatascience.com\/visualizing-intermediate-activation-in-convolutional-neural-networks-with-keras-260b36d60d0<\/a><\/p>\n"},{"AnswerId":"57863581","CreationDate":"2019-09-10T03:16:29.120","ParentId":null,"OwnerUserId":"1898734","Title":null,"Body":"<p>Let me attempt to answer your queries. <\/p>\n\n<ol>\n<li><p>Reducing size of the image size helps model learn faster and reduce the memory requirement. 150 X 150 pixels require more number of nodes in a single layer of neural network and thereby more memory. I am not sure about the accuracy stats of Squashing versus Center Cropping. <\/p><\/li>\n<li><p>Increasing number of nodes may not directly relate to improvement in accuracy. Bagging and Boosting can improve accuracy way better than increasing number of layers in many scenarios. Drop out is a problem in deep neural networks but with proper data normalization along with proper activation functions and sufficient training data should help improve accuracy with multiple layers.<\/p><\/li>\n<li><p>Short answer is yes. You can do it by reshaping the feature vectors to the original image shape at the end of each layer for which you have to define your own neural network.<\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":57863436,"AnswerCount":0,"Tags":"<r><tensorflow><logistic-regression>","CreationDate":"2019-09-10T02:56:16.657","AcceptedAnswerId":null,"OwnerUserId":7500268.0,"Title":"why the Logistic regression model trained with tensorflow performed so poor","Body":"<p>I trained a logistic regression model with tensorflow but the accuracy of the model was so poor (accuracy = 0.68). The model was trained using simulated dataset and the result should be very good. is there something wrong with the code ?<\/p>\n\n<pre class=\"lang-r prettyprint-override\"><code>#simulated dataSet\nsim_data &lt;- function(n=2000){\n  library(dummies)\n  age &lt;- round(abs(rnorm(n,mean = 60, sd = 20)))\n  lac &lt;- round(abs(rnorm(n,3,1)),1)\n  wbc &lt;- round(abs(rnorm(n,10,3)),1)\n  sex &lt;- factor(rbinom(n,size = 1,prob = 0.6),\n                   labels = c(\"Female\",\"Male\"));\n  type &lt;- as.factor(sample(c(\"Med\",\"Emerg\",\"Surg\"),\n                           size = n,replace = T,\n                           prob = c(0.4,0.4,0.2)))\n  linPred &lt;- cbind(1,age,lac,wbc,dummy(sex)[,-1],\n                   dummy(type)[,-1]) %*%\n    c(-30,0.2,4,1,-2,3,-3)\n  pi &lt;- 1\/(1+exp(-linPred))\n  mort &lt;- factor(rbinom(n,size = 1, prob = pi),\n                 labels = c(\"Alive\",\"Died\"))\n  dat &lt;- data.frame(age=age,lac=lac,wbc=wbc,\n                    sex=sex,type=type,\n                    mort = mort)\n  return(dat)\n}\nset.seed(123)\ndat &lt;- sim_data()\ndat_test &lt;- sim_data(n=1000)\n#logistic regression in conventional method\nmod &lt;- glm(mort~.,data = dat,family = \"binomial\")\nlibrary(tableone)\nShowRegTable(mod)\n#diagnotic accuracy\npred &lt;- predict.glm(mod,newdata = dat_test,\n                    type = \"response\")\nlibrary(pROC)\nroc(response = dat_test$mort,predictor = pred,ci=T)\npredBi &lt;- pred &gt;= 0.5\ncrossTab &lt;- table(predBi,dat_test$mort)\n(crossTab[1]+crossTab[4])\/sum(crossTab)\n#choose different cutoff for the accuracy\nDTaccuracy &lt;- data.frame()\nfor (cutoff in seq(0,1,by = 0.01)) {\n  predBi &lt;- pred &gt;= cutoff;\n  crossTab &lt;- table(predBi,dat_test$mort)\n  accuracy = (crossTab[1]+crossTab[4])\/sum(crossTab)\n  DTaccuracy &lt;- rbind(DTaccuracy,c(accuracy,cutoff))\n}\nnames(DTaccuracy) &lt;- c('Accuracy','Cutoff')\nqplot(x=Cutoff, y = Accuracy, data = DTaccuracy)\n\n\n#tensorflow method\nlibrary(caret)\ny = with(dat, model.matrix(~ mort + 0))\nx = model.matrix(~.,dat[,!names(dat)%in%\"mort\"])\ntrainIndex = createDataPartition(1:nrow(x), \n                                 p=0.7, list=FALSE,times=1)\n\nx_train = x[trainIndex,]\nx_test = x[-trainIndex,]\ny_train = y[trainIndex,]\ny_test = y[-trainIndex,]\n# Hyper-parameters\nepochs = 30             # Total number of training epochs\nbatch_size = 30        # Training batch size\ndisplay_freq = 10      # Frequency of displaying the training results\nlearning_rate = 0.1   # The optimization initial learning rate\n#Then we will define the placeholders for features and labels:\nlibrary(tensorflow)\nX &lt;- tf$placeholder(tf$float32, shape(NULL, ncol(x)),\n                    name = \"X\")\nY = tf$placeholder(tf$float32, shape(NULL, 2L), name = \"Y\")\n\n#we will define the parameters. We will randomly initialize the weights with mean \u201c0\u201d and a standard deviation of \u201c1.\u201d We will initialize bias to \u201c0.\u201d\n\nW = tf$Variable(tf$random_normal(shape(ncol(x),2L),\n                                 stddev = 1.0), \n                name = \"weghts\")\nb = tf$Variable(tf$zeros(shape(2L)), name = \"bias\")\n\n#Then we will compute the logit.\n\nlogits = tf$add(tf$matmul(X, W), b)\npred = tf$nn$sigmoid(logits)\n#The next step is to define the loss function. We will use sigmoid cross entropy with logits as a loss function.\n\nentropy = tf$nn$sigmoid_cross_entropy_with_logits(labels = Y, \n                                                  logits = logits)\nloss = tf$reduce_mean(entropy,name = \"loss\")\n#The last step of the model composition is to define the training op. We will use a gradient descent with a learning rate 0.1 to minimize cost.\n\noptimizer = tf$train$GradientDescentOptimizer(learning_rate = learning_rate)$minimize(loss)\ninit_op = tf$global_variables_initializer()\n#Now that we have trained the model, let\u2019s evaluate it:\n\ncorrect_prediction &lt;- tf$equal(tf$argmax(logits, 1L), tf$argmax(Y, 1L),\n                               name = \"correct_pred\")\naccuracy &lt;- tf$reduce_mean(tf$cast(correct_prediction, tf$float32),\n                           name = \"accuracy\")\n#Having structured the graph, let\u2019s execute it:\n\nwith(tf$Session() %as% sess, {\n  sess$run(init_op)\n  for (i in 1:5000) {\n    sess$run(optimizer,\n             feed_dict = dict(X=x_train, Y=y_train))\n  }\n  sess$run(accuracy, \n           feed_dict=dict(X = x_test, Y = y_test))\n})\n<\/code><\/pre>\n\n<p>The accuracy obtained by glm() method is quite good (accuracy=0.95), which is as expected; however, the accuracy was only 0.68 with TensorFlow method. How can I solve the problem?<\/p>\n","answers":[]}
+{"QuestionId":57863492,"AnswerCount":0,"Tags":"<pytorch><nvidia>","CreationDate":"2019-09-10T03:03:50.433","AcceptedAnswerId":null,"OwnerUserId":5164382.0,"Title":"System rebooting while training pytorch mask-rcnn with 4 GPUs","Body":"<p>My hardware details:<\/p>\n\n<pre><code>Motherboard: Asus WS X299 SAGE\/10G \nCPU: Intel Core i9-9900X\nGPU: Geforce RTX2080 TI - 11GB (4 of them)\nPower supply: Masterwatt Maker - 1500Watts\n<\/code><\/pre>\n\n<p>System details:<\/p>\n\n<pre><code>Bios Version: 1201.\nOS: Ubuntu 18.04\nNvidia driver: 418.56\ncuda under Conda: 10.0.130\n<\/code><\/pre>\n\n<p>Tested using <a href=\"https:\/\/github.com\/wilicc\/gpu-burn\" rel=\"nofollow noreferrer\">https:\/\/github.com\/wilicc\/gpu-burn<\/a>. All gpus are ok.<\/p>\n\n<p>Whenever I am training maskrcnn_resnet50_fpn (<a href=\"https:\/\/github.com\/pytorch\/vision\/tree\/master\/references\/detection\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/vision\/tree\/master\/references\/detection<\/a>) on coco dataset using 4 GPUs with batch size 4, the system is rebooting immediately. But, when I am using 3 GPUs with batch size 4 or 4 GPUs with batch size 2, it's training.<\/p>\n\n<p>What could be the reason? Power supply?\nI am dying to solve. I appreciate your comments.\nThanks in advance\nZulfi<\/p>\n","answers":[]}
+{"QuestionId":57863534,"AnswerCount":1,"Tags":"<python><keras><deep-learning>","CreationDate":"2019-09-10T03:10:29.670","AcceptedAnswerId":57863824.0,"OwnerUserId":12036469.0,"Title":"Can we achieve effective results with just integer encoding?","Body":"<p>I am learning deep learning,<\/p>\n\n<p>I'm using the IMDB data set.\nIt's [Integer-encoding] processed?<\/p>\n\n<p>Some examples show you're just doing deep learning, not converting to one-hot-encoding.<\/p>\n\n<p>Is this enough to get a valid result?<\/p>\n\n<p>if that be so<\/p>\n\n<p>What's the advantage of one-hot-encoding?<\/p>\n\n<p>is this my code<\/p>\n\n<pre><code>import numpy\nfrom keras.datasets import imdb\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.layers import Flatten\nfrom keras.layers.convolutional import Conv1D\nfrom keras.layers.convolutional import MaxPooling1D\nfrom keras.layers.embeddings import Embedding\nfrom keras.preprocessing import sequence\n# load the dataset but only keep the top n words, zero the rest\ntop_words = 5000\nmax_words = 500\n\nX_train = train_result\ny_train = train_label\nX_test = test_result\ny_test = test_label# pad dataset to a maximum review length in words\nX_train = sequence.pad_sequences(X_train, maxlen=max_words)\nX_test = sequence.pad_sequences(X_test, maxlen=max_words)\nprint(X_train[:1])\n# create the model\nmodel = Sequential()\nmodel.add(Embedding(top_words, 32, input_length=max_words))\nmodel.add(Conv1D(filters=32, kernel_size=3, padding='same', activation='relu'))\nmodel.add(MaxPooling1D(pool_size=2))\nmodel.add(Flatten())\nmodel.add(Dense(250, activation='relu'))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\nmodel.summary()\n# Fit the model\nhist = model.fit(X_train, y_train, validation_data=(X_test, y_test), epochs=20, batch_size=128, verbose=1)\n# Final evaluation of the model\nscores = model.evaluate(X_test, y_test, verbose=0)\nprint(\"Accuracy: %.2f%%\" % (scores[1]*100))\n<\/code><\/pre>\n\n<p>in X_train[1].<\/p>\n\n<pre><code>[[   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\n     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\n     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\n     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\n     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\n     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\n     0    0    0    0    0    0    0    0    0    0    0    0    0    0\n     0    0    0    0    0    0    0    3  284 2452  756    1 3075  194\n    54 3717   10  757  169 2216    5    1 1906  843   54   52 2732 3403\n     5 1819    3   34    4   54 1819    5 2532   42  668   23   54  709\n    52    7    9    2   80  172 3258  265   33    1 1467    4  683    4\n    11   21  988    1    3  110  631    2    4  321    3 3040  294  284\n   478   33    1   33   54 4349   33   54  213    2   86   54  516  420\n   754    1   84    2    8  526  473   63   20  184   20  184   20  184\n  1138   52    3   23    1 1468  101    3 1850    4   61    6  777   20\n   237  185   52 3846    5   54  149    7   34    4    1   18   54 4802\n   929    2    5   98    8   13   17    9    1  993  117  101    3  165\n    41  653  781    3  286  923 2882    7  210    3  181    5 1743    3\n   120  814 1630 1517    3 2317 4606 4425    9   43  686    5  744 1018\n   910  223  136 3782 1585  775 1391 3041  155    3  292    4 2975    2\n   136  135  120  864   24  869 3655  245  421    1 1803   10    1  120\n     2    1  261   78 1671   19   43 1288   16    1 1036    5  380    1\n  1744  121   10    1   84  252   55   51  670    2   24  200   51 1709\n     1 1256 1469    2    1  217    5 2453  423   79  929   36    9    3\n  1106    4 2754 4526   14   29   24 2393   74   34 4049   17   42   72\n     9  365    1   69   41 1804  572   41  559   76   92    2  153  112\n    11   15  835 1423  136    1   59   15   67    1 1320    5  441    2\n   733   17    1  688  890    5   26  421   55   23  208    2   31 2070\n    23    1 2998  136    6  413   44   33   40    7  119    9  668    4\n    22 3213   40    7  119  151  359    5   25  185]]\n<\/code><\/pre>\n\n<p>and this is output,<\/p>\n\n<pre><code>Epoch 1\/20\n10103\/10103 [==============================] - 5s 523us\/step - loss: 0.5812 - acc: 0.6589 - val_loss: 0.1229 - val_acc: 0.9551\nEpoch 2\/20\n10103\/10103 [==============================] - 5s 478us\/step - loss: 0.1299 - acc: 0.9485 - val_loss: 0.0693 - val_acc: 0.9663\nEpoch 3\/20\n10103\/10103 [==============================] - 5s 488us\/step - loss: 0.0544 - acc: 0.9824 - val_loss: 0.0589 - val_acc: 0.9775\nEpoch 4\/20\n10103\/10103 [==============================] - 5s 488us\/step - loss: 0.0258 - acc: 0.9923 - val_loss: 0.0371 - val_acc: 0.9850\nEpoch 5\/20\n10103\/10103 [==============================] - 5s 483us\/step - loss: 0.0120 - acc: 0.9976 - val_loss: 0.0528 - val_acc: 0.9813\nEpoch 6\/20\n10103\/10103 [==============================] - 5s 483us\/step - loss: 0.0058 - acc: 0.9991 - val_loss: 0.0464 - val_acc: 0.9850\nEpoch 7\/20\n10103\/10103 [==============================] - 5s 482us\/step - loss: 0.0032 - acc: 0.9994 - val_loss: 0.0707 - val_acc: 0.9738\nEpoch 8\/20\n10103\/10103 [==============================] - 5s 485us\/step - loss: 0.0022 - acc: 0.9997 - val_loss: 0.0471 - val_acc: 0.9925\nEpoch 9\/20\n10103\/10103 [==============================] - 5s 482us\/step - loss: 0.0011 - acc: 0.9998 - val_loss: 0.0698 - val_acc: 0.9775\nEpoch 10\/20\n10103\/10103 [==============================] - 5s 481us\/step - loss: 6.8280e-04 - acc: 1.0000 - val_loss: 0.0728 - val_acc: 0.9775\nEpoch 11\/20\n10103\/10103 [==============================] - 5s 483us\/step - loss: 4.8174e-04 - acc: 1.0000 - val_loss: 0.0873 - val_acc: 0.9738\nEpoch 12\/20\n10103\/10103 [==============================] - 5s 477us\/step - loss: 3.4037e-04 - acc: 1.0000 - val_loss: 0.0674 - val_acc: 0.9813\nEpoch 13\/20\n10103\/10103 [==============================] - 5s 478us\/step - loss: 2.6164e-04 - acc: 1.0000 - val_loss: 0.0847 - val_acc: 0.9775\nEpoch 14\/20\n10103\/10103 [==============================] - 5s 478us\/step - loss: 2.0453e-04 - acc: 1.0000 - val_loss: 0.0812 - val_acc: 0.9775\nEpoch 15\/20\n10103\/10103 [==============================] - 5s 473us\/step - loss: 1.6034e-04 - acc: 1.0000 - val_loss: 0.0831 - val_acc: 0.9775\nEpoch 16\/20\n10103\/10103 [==============================] - 5s 469us\/step - loss: 1.3443e-04 - acc: 1.0000 - val_loss: 0.0874 - val_acc: 0.9775\nEpoch 17\/20\n10103\/10103 [==============================] - 5s 467us\/step - loss: 1.1035e-04 - acc: 1.0000 - val_loss: 0.0891 - val_acc: 0.9775\nEpoch 18\/20\n10103\/10103 [==============================] - 5s 471us\/step - loss: 9.3257e-05 - acc: 1.0000 - val_loss: 0.0956 - val_acc: 0.9775\nEpoch 19\/20\n10103\/10103 [==============================] - 5s 465us\/step - loss: 7.9740e-05 - acc: 1.0000 - val_loss: 0.0965 - val_acc: 0.9775\nEpoch 20\/20\n10103\/10103 [==============================] - 5s 467us\/step - loss: 6.7700e-05 - acc: 1.0000 - val_loss: 0.0919 - val_acc: 0.9775\nAccuracy: 97.75%\n<\/code><\/pre>\n","answers":[{"AnswerId":"57863824","CreationDate":"2019-09-10T03:54:12.457","ParentId":null,"OwnerUserId":"6458246","Title":null,"Body":"<p>Integer-encoding implied there are some ordinal relationship in your labels, thus one-hot embedding is necessary while you constructing a classification model. Essentially, one-hot embedding is mapping discrete data into Euclidean space.<\/p>\n\n<p>For example, here is a dataset include 3 categories: apple, orange, banana. If you used integer-encoding: {apple => 0, orange => 1, banana => 2}, you could never say 'orange' is larger or greater than 'apple'.<\/p>\n\n<p>In your case, IMDB review dataset is a binary classification dataset, there are 2 kinds of label: negative and positive. You can deal them as continuous feature: if predicted value is closer to 1, the rate of review is more positive, and vice versa.<\/p>\n\n<p><a href=\"https:\/\/www.quora.com\/What-are-good-ways-to-handle-discrete-and-continuous-inputs-together\" rel=\"nofollow noreferrer\">https:\/\/www.quora.com\/What-are-good-ways-to-handle-discrete-and-continuous-inputs-together<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/17469835\/why-does-one-hot-encoding-improve-machine-learning-performance\">Why does one hot encoding improve machine learning performance?<\/a> <\/p>\n"}]}
+{"QuestionId":57863687,"AnswerCount":0,"Tags":"<machine-learning><keras><deep-learning><conv-neural-network><pytorch>","CreationDate":"2019-09-10T03:32:34.373","AcceptedAnswerId":null,"OwnerUserId":12044284.0,"Title":"CIFAR-10 test set classification accuracy different on PyTorch and Keras","Body":"<p>I\u2019ve made a custom CNN in PyTorch for classifying 10 classes in the CIFAR-10 dataset. My classification accuracy on the test dataset is 45.739%, this is very low and I thought it\u2019s because my model is not very deep but I implemented the same model in Keras and the classification accuracy come outs to be 78.92% on test dataset. No problem in Keras however I think there's something I'm missing in my PyTorch program.<\/p>\n\n<p>I have used the same model architecture, strides, padding, dropout rate, optimizer, loss function, learning rate, batch size, number of epochs on both PyTorch and Keras and despite that, the difference in the classification accuracy is still huge thus I\u2019m not able to decide how I should debug my PyTorch program further.<\/p>\n\n<p>For now I suspect 3 things: in Keras, I use the categorical cross entropy loss function (one hot vector labels) and in PyTorch I use the standard cross entropy loss function (scalar indices labels), can this be a problem?, if not then I suspect either my training loop or the code for calculating classification accuracy in PyTorch. I have attached both my programs below, will be grateful to any suggestions.<\/p>\n\n<p>My program in Keras:<\/p>\n\n<pre><code>#================Function that defines the CNN model===========\ndef CNN_model():\n    model = Sequential()\n\n    model.add(Conv2D(32,(3,3),activation='relu',padding='same', input_shape=(size,size,channels))) #SAME PADDING\n    model.add(Conv2D(32,(3,3),activation='relu')) #VALID PADDING\n    model.add(MaxPooling2D(pool_size=(2,2))) #VALID PADDING\n    model.add(Dropout(0.25))\n\n    model.add(Conv2D(64,(3,3),activation='relu', padding='same')) #SAME PADDING\n    model.add(Conv2D(64,(3,3),activation='relu')) #VALID PADDING\n    model.add(MaxPooling2D(pool_size=(2,2))) #VALID PADDING\n    model.add(Dropout(0.25))\n\n\n    model.add(Conv2D(128,(3,3),activation='relu', padding='same')) #SAME PADDING\n    model.add(Conv2D(128,(3,3),activation='relu')) #VALID PADDING\n    model.add(MaxPooling2D(pool_size=(2,2),name='feature_extractor_layer')) #VALID PADDING\n    model.add(Dropout(0.25))\n\n\n    model.add(Flatten())\n\n    model.add(Dense(512, activation='relu', name='second_last_layer'))\n    model.add(Dropout(0.25))\n    model.add(Dense(10, activation='softmax', name='softmax_layer')) #10 nodes in the softmax layer\n    model.summary()\n\n    return model\n\n\n\n#=====Main program starts here========\n#get_train_data() and get_test_data() are my own custom functions to get CIFAR-10 dataset\nimages_train, labels_train, class_train = get_train_data(0,10)\nimages_test, labels_test, class_test = get_test_data(0,10)\n\nmodel = CNN_model()\n\n\nmodel.compile(loss='categorical_crossentropy', #loss function of the CNN\n             optimizer=Adam(lr=1.0e-4), #Optimizer\n              metrics=['accuracy'])#'accuracy' metric is to be evaluated\n\n#images_train and images_test contain images and \n#class_train and class_test contains one hot vectors labels \nmodel.fit(images_train,class_train,\n          batch_size=128,\n          epochs=50,\n          validation_data=(images_test,class_test),\n          verbose=1)\n\nscores=model.evaluate(images_test,class_test,verbose=0)\nprint(\"Accuracy: \"+str(scores[1]*100)+\"% \\n\")\n<\/code><\/pre>\n\n<p>My program in PyTorch:<\/p>\n\n<pre><code>#========DEFINE THE CNN MODEL=====\nclass Net(nn.Module):\n    def __init__(self):\n        super(Net, self).__init__()\n        self.conv1 = nn.Conv2d(3, 32, 3,1,1)#SAME PADDING\n        self.conv2 = nn.Conv2d(32,32,3,1,0)#VALID PADDING\n        self.pool1 = nn.MaxPool2d(2,2) #VALID PADDING\n        self.drop1 = nn.Dropout2d(0.25) #DROPOUT OF 0.25\n\n        self.conv3 = nn.Conv2d(32,64,3,1,1)#SAME PADDING\n        self.conv4 = nn.Conv2d(64,64,3,1,0)#VALID PADDING\n        self.pool2 = nn.MaxPool2d(2,2)#VALID PADDING\n        self.drop2 = nn.Dropout2d(0.25) #DROPOUT OF 0.25\n\n        self.conv5 = nn.Conv2d(64,128,3,1,1)#SAME PADDING\n        self.conv6 = nn.Conv2d(128,128,3,1,0)#VALID PADDING\n        self.pool3 = nn.MaxPool2d(2,2)#VALID PADDING\n        self.drop3 = nn.Dropout2d(0.25) #DROPOUT OF 0.25\n\n        self.fc1 = nn.Linear(128*2*2, 512)#128*2*2 IS OUTPUT DIMENSION AFTER THE PREVIOUS LAYER \n        self.drop4 = nn.Dropout(0.25) #DROPOUT OF 0.25\n        self.fc2 = nn.Linear(512,10) #10 output nodes\n\n\n    def forward(self, x):\n        x = F.relu(self.conv1(x))\n        x = F.relu(self.conv2(x))\n        x = self.pool1(x)\n        x = self.drop1(x)\n\n\n        x = F.relu(self.conv3(x))\n        x = F.relu(self.conv4(x))\n        x = self.pool2(x)\n        x = self.drop2(x)\n\n        x = F.relu(self.conv5(x))\n        x = F.relu(self.conv6(x))\n        x = self.pool3(x)\n        x = self.drop3(x)\n\n\n        x = x.view(-1,2*2*128) #FLATTENING OPERATION 2*2*128 IS OUTPUT AFTER THE PREVIOUS LAYER\n        x = F.relu(self.fc1(x))\n        x = self.drop4(x)\n        x = self.fc2(x) #LAST LAYER DOES NOT NEED SOFTMAX BECAUSE THE LOSS FUNCTION WILL TAKE CARE OF IT\n        return x\n\n\n\n#=======FUNCTION TO CONVERT INPUT AND TARGET TO TORCH TENSORS AND LOADING INTO GPU======\ndef PrepareInputDataAndTargetData(device,images,labels,batch_size):\n\n    #GET MINI BATCH OF TRAINING IMAGES AND RESHAPE THE TORCH TENSOR FOR CNN PROCESSING\n    mini_batch_images = torch.tensor(images)\n    mini_batch_images = mini_batch_images.view(batch_size,3,32,32)\n\n    #GET MINI BATCH OF TRAINING LABELS, TARGET SHOULD BE IN LONG FORMAT SO CONVERT THAT TOO\n    mini_batch_labels = torch.tensor(labels)\n    mini_batch_labels = mini_batch_labels.long()\n\n    #FEED THE INPUT DATA AND TARGET LABELS TO GPU\n    mini_batch_images = mini_batch_images.to(device)\n    mini_batch_labels = mini_batch_labels.to(device)\n\n    return mini_batch_images,mini_batch_labels\n\n#==========MAIN PROGRAM==========\ndevice = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n#get_train_data() and get_test_data() are my own custom functions to get CIFAR-10 dataset \nImages_train, Labels_train, Class_train = get_train_data(0,10)\nImages_test, Labels_test, Class_test = get_test_data(0,10)\n\nnet = Net()\nnet = net.double() #https:\/\/discuss.pytorch.org\/t\/runtimeerror-expected-object-of-scalar-type-double-but-got-scalar-type-float-for-argument-2-weight\/38961\nprint(net)\n\n#MAP THE MODEL ONTO THE GPU\nnet = net.to(device)\n\n#CROSS ENTROPY LOSS FUNCTION AND ADAM OPTIMIZER\ncriterion = nn.CrossEntropyLoss()\noptimizer = optim.Adam(net.parameters(), lr=1e-4)\n\n#PREPARE THE DATALOADER\n#Images_train contains images and Labels_trains contains indices i.e. 0,1,...,9 \ndataset = TensorDataset( Tensor(Images_train), Tensor(Labels_train) )\ntrainloader = DataLoader(dataset, batch_size= 128, shuffle=True)\n\n#START TRAINING THE CNN MODEL FOR 50 EPOCHS\nfor epoch in range(0,50):\n    for i, data in enumerate(trainloader, 0):\n        inputs, labels = data\n        inputs = torch.tensor(inputs).double()\n        inputs = inputs.view(len(inputs),3,32,32) #RESHAPE THE IMAGES\n        labels = labels.long() #MUST CONVERT LABEL TO LONG FORMAT\n\n        #MAP THE INPUT AND LABELS TO THE GPU\n        inputs=inputs.to(device)\n        labels=labels.to(device)\n\n        #FORWARD PROP, BACKWARD PROP, PARAMETER UPDATE\n        optimizer.zero_grad()\n        outputs = net.forward(inputs)\n        loss = criterion(outputs, labels)\n        loss.backward()\n        optimizer.step()\n\n\n#CALCULATE CLASSIFICATION ACCURACY ON ALL 10 CLASSES\nwith torch.no_grad():\n    Images_class,Labels_class = PrepareInputDataAndTargetData(device,Images_test,Labels_test,len(Images_test))\n    network_outputs = net.forward(Images_class)\n    correct = (torch.argmax(network_outputs.data,1) == Labels_class.data).float().sum()\n    acc = float(100.0*(correct\/len(Images_class)))\n    print(\"Accuracy is: \"+str(acc)+\"\\n\")\n\n    del Images_class\n    del Labels_class\n    del network_outputs\n    del correct\n    del acc\n    torch.cuda.empty_cache()\n\n\nprint(\"Done\\n\")\n<\/code><\/pre>\n\n<p>I am not fully aware of how the actual core backend works in both libraries however I suppose that the classification accuracy of any model should be almost the same regardless of the library.<\/p>\n","answers":[]}
+{"QuestionId":57863907,"AnswerCount":2,"Tags":"<machine-learning><deep-learning><pytorch><torch>","CreationDate":"2019-09-10T04:04:20.763","AcceptedAnswerId":57867181.0,"OwnerUserId":3983454.0,"Title":"pytorch questions: how to add bias term and extract its value? class vs sequential model? and softmax","Body":"<p>I have a basic neural network model in pytorch like this:<\/p>\n\n<pre><code>import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nclass Net(nn.Module):\n    def __init__(self, input_dim, hidden_dim, output_dim):\n        super(Net, self).__init__()\n        self.fc1 = nn.Linear(input_dim, hidden_dim)\n        self.sigmoid = nn.Sigmoid()\n        self.fc2 = nn.Linear(hidden_dim, output_dim)\n\n    def forward(self, x):\n        out = self.fc1(x)\n        out = self.sigmoid(out)\n        out = self.fc2(out)\n        return out\nnet = Net(400, 512,10)\n<\/code><\/pre>\n\n<p>How can I extract bias\/intercept term from net.parameters()?\nAnd is this model equivalent to using sequential()?<\/p>\n\n<pre><code>net = nn.Sequential(nn.Linear(input_dim, hidden_dim[0]),\n                      nn.Sigmoid(),\n                      nn.Linear(hidden_dim[0], hidden_dim[1]),\n                      nn.Sigmoid(),\n                      nn.Linear(hidden_dim[1], output_dim))\n<\/code><\/pre>\n\n<p>Is nn.Softmax() optional at the end of either model for multi-class classification? If I understand correctly, with software it outputs probability of a certain class but without it returns predicted output?<\/p>\n\n<p>Thanks in advance for answering my newbie questions.<\/p>\n","answers":[{"AnswerId":"57867181","CreationDate":"2019-09-10T08:49:51.527","ParentId":null,"OwnerUserId":"9828197","Title":null,"Body":"<p>Let's answer questions one by one. <code>is this model equivalent to using sequential()<\/code>\nShort answer: No. You can see that you have added two Sigmoid and two linear layers. You can print your net and see the result:<\/p>\n\n<pre><code>net = Net(400, 512,10)\n\nprint(net.parameters())\nprint(net)\ninput_dim = 400\nhidden_dim = 512\noutput_dim = 10\n\nmodel = Net(400, 512,10)\n\nnet = nn.Sequential(nn.Linear(input_dim, hidden_dim),\n                      nn.Sigmoid(),\n                      nn.Linear(hidden_dim, hidden_dim),\n                      nn.Sigmoid(),\n                      nn.Linear(hidden_dim, output_dim))\n\nprint(net)\n<\/code><\/pre>\n\n<p>The output is:<\/p>\n\n<pre><code>Net(\n  (fc1): Linear(in_features=400, out_features=512, bias=True)\n  (sigmoid): Sigmoid()\n  (fc2): Linear(in_features=512, out_features=10, bias=True)\n)\n\nSequential(\n  (0): Linear(in_features=400, out_features=512, bias=True)\n  (1): Sigmoid()\n  (2): Linear(in_features=512, out_features=512, bias=True)\n  (3): Sigmoid()\n  (4): Linear(in_features=512, out_features=10, bias=True)\n)\n\n<\/code><\/pre>\n\n<p>I hope you can see where they differ.<\/p>\n\n<p>Your first question: <code>How can I extract bias\/intercept term from net.parameters()<\/code><\/p>\n\n<p>The answer:<\/p>\n\n<pre><code>model = Net(400, 512,10)\n\nbias = model.fc1.bias\n\nprint(bias)\n<\/code><\/pre>\n\n<p>the output is:<\/p>\n\n<pre><code>tensor([ 3.4078e-02,  3.1537e-02,  3.0819e-02,  2.6163e-03,  2.1002e-03,\n         4.6842e-05, -1.6454e-02, -2.9456e-02,  2.0646e-02, -3.7626e-02,\n         3.5531e-02,  4.7748e-02, -4.6566e-02, -1.3317e-02, -4.6593e-02,\n        -8.9996e-03, -2.6568e-02, -2.8191e-02, -1.9806e-02,  4.9720e-02,\n        ---------------------------------------------------------------\n        -4.6214e-02, -3.2799e-02, -3.3605e-02, -4.9720e-02, -1.0293e-02,\n         3.2559e-03, -6.6590e-03, -1.2456e-02, -4.4547e-02,  4.2101e-02,\n        -2.4981e-02, -3.6840e-03], requires_grad=True)\n<\/code><\/pre>\n"},{"AnswerId":"57866259","CreationDate":"2019-09-10T07:48:43.573","ParentId":null,"OwnerUserId":"10606534","Title":null,"Body":"<p>you can you state_dict for extract  bias of each layer or function in Model.<\/p>\n\n<p>the two present of your network is same but if you want do some extend of netwrok I will suggest use Net one instead of Sequential One<\/p>\n\n<p>WithOut softmax it will just output a tensor witch  min -1 max 1 if use sigmoid , it can't be the predict.<\/p>\n\n<p>AnyWay ,  you should answer your question seperate instead of one post for three question . good luck <\/p>\n"}]}
+{"QuestionId":57863942,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-10T04:11:04.523","AcceptedAnswerId":58443861.0,"OwnerUserId":10867166.0,"Title":"What's the difference between tensorflow.python.keras and tensorflow.keras?","Body":"<p>As the title says, Are they the same api? When I print the layers module in keras, the result are shown as follow:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.keras import layers\nprint(layers)\nfrom tensorflow.python.keras import layers\nprint(layers)\n<\/code><\/pre>\n\n<p><strong>result<\/strong><\/p>\n\n<pre><code>&lt;module 'tensorflow.python.keras.api._v1.keras.layers' from '\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/api\/_v1\/keras\/layers\/__init__.py'&gt;\n&lt;module 'tensorflow.python.keras.layers' from '\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/layers\/__init__.py'&gt;\n<\/code><\/pre>\n\n<p>We can see that two modules come from different source.<\/p>\n\n<p>And I find the api module from source code, there is only a BUILD file.<\/p>\n\n<p>Is there a relation between two modules, what is the mechanism of the api generator?<\/p>\n","answers":[{"AnswerId":"58443861","CreationDate":"2019-10-18T04:13:37.100","ParentId":null,"OwnerUserId":"11530462","Title":null,"Body":"<p>Anything under <code>tf.python.*<\/code> is private, intended for development only, rather than for public use. <\/p>\n\n<p>Importing from <code>tensorflow.python<\/code> or any other modules (including <code>import tensorflow_core<\/code>...) is not supported, and can break unannounced. <\/p>\n\n<p>So, it is suggested not to use anything with <code>tf.python.*<\/code>.<\/p>\n"}]}
+{"QuestionId":57865112,"AnswerCount":1,"Tags":"<deep-learning><pytorch>","CreationDate":"2019-09-10T06:26:04.840","AcceptedAnswerId":57865835.0,"OwnerUserId":2530674.0,"Title":"Training with BatchNorm in pytorch","Body":"<p>I'm wondering if I need to do anything special when training with BatchNorm in pytorch. From my understanding the <code>gamma<\/code> and <code>beta<\/code> parameters are updated with gradients as would normally be done by an optimizer. However, the mean and variance of the batches are updated slowly using momentum.<\/p>\n\n<ol>\n<li>So do we need to specify to the optimizer when the mean and variance parameters are updated, or does pytorch automatically take care of this?<\/li>\n<li>Is there a way to access the mean and variance of the BN layer so that I can make sure it was changing while I trained the model.<\/li>\n<\/ol>\n\n<p>If needed here is my model and training procedure:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def bn_drop_lin(n_in:int, n_out:int, bn:bool=True, p:float=0.):\n    \"Sequence of batchnorm (if `bn`), dropout (with `p`) and linear (`n_in`,`n_out`) layers followed by `actn`.\"\n    layers = [nn.BatchNorm1d(n_in)] if bn else []\n    if p != 0: layers.append(nn.Dropout(p))\n    layers.append(nn.Linear(n_in, n_out))\n\n    return nn.Sequential(*layers)\n\nclass Model(nn.Module):\n    def __init__(self, i, o, h=()):\n        super().__init__()\n\n        nodes = (i,) + h + (o,)\n        self.layers = nn.ModuleList([bn_drop_lin(i,o, p=0.5) \n                                     for i, o in zip(nodes[:-1], nodes[1:])])\n\n    def forward(self, x):\n        x = x.view(x.shape[0], -1)\n        for layer in self.layers[:-1]:\n            x = F.relu(layer(x))\n\n        return self.layers[-1](x)\n<\/code><\/pre>\n\n<p>Training:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>for i, data in enumerate(trainloader):\n    # get the inputs; data is a list of [inputs, labels]\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<\/code><\/pre>\n","answers":[{"AnswerId":"57865835","CreationDate":"2019-09-10T07:18:43.660","ParentId":null,"OwnerUserId":"2790047","Title":null,"Body":"<p>Batchnorm layers behave differently depending on if the model is in train or eval mode.<\/p>\n\n<p>When <code>net<\/code> is in train mode (i.e. after calling <code>net.train()<\/code>) the batch norm layers contained in <code>net<\/code> will use batch statistics along with gamma and beta parameters to adjust the mean and variance of each mini-batch. The running mean and variance will also be adjusted while in train mode. These updates to running mean and variance occur during the forward pass (when <code>net(inputs)<\/code> is called). The gamma and beta parameters are updated only once <code>optimizer.step()<\/code> is called.<\/p>\n\n<p>When <code>net<\/code> is in eval mode (<code>net.eval()<\/code>) batch norm uses the historical running mean and running variance stored during training to adjust the mean and running variance of samples.<\/p>\n\n<p>You can check the batch norm layers running mean and variance by displaying the layers <code>running_mean<\/code> and <code>running_var<\/code> members to ensure batch norm is updating them as expected. The learnable gamma and beta parameters can be accessed by displaying the <code>weight<\/code> and <code>bias<\/code> members of a batch norm layer respectively.<\/p>\n\n<p><strong>Edit<\/strong><\/p>\n\n<p>Below is a simple demonstration code showing that <code>running_mean<\/code> is updated during forward. Observe that it is not updated by the optimizer.<\/p>\n\n<pre><code>&gt;&gt;&gt; import torch\n&gt;&gt;&gt; import torch.nn as nn\n&gt;&gt;&gt; layer = nn.BatchNorm1d(5)\n&gt;&gt;&gt; layer.train()\n&gt;&gt;&gt; layer.running_mean\ntensor([0., 0., 0., 0., 0.])\n&gt;&gt;&gt; result = layer(torch.randn(5,5))\n&gt;&gt;&gt; layer.running_mean\ntensor([ 0.0271,  0.0152, -0.0403, -0.0703, -0.0056])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57865134,"AnswerCount":0,"Tags":"<keras><softmax>","CreationDate":"2019-09-10T06:27:35.260","AcceptedAnswerId":null,"OwnerUserId":11143966.0,"Title":"loss='mean_squared_error', Vs categorical_crossentropy or sparse_categorical_crossentropy","Body":"<p>Below is the model i created for the digit detection problem.<\/p>\n\n<p>Input is (28,28,1)<\/p>\n\n<p>output is categorical values between 0-9<\/p>\n\n<p>My Keras model summary is as below<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential([InputLayer(X_train[0].shape),\n\nConv2D(16,kernel_size=(3,3),strides=[1,1],padding='same',kernel_initializer=keras.initializers.glorot_uniform(),name='conv1'),\n\nActivation('relu'),\n\nMaxPool2D(pool_size=(2, 2)),\n\nConv2D(32,kernel_size=(3,3),strides=[2,2],padding='same',kernel_initializer=keras.initializers.glorot_uniform(),name='conv2'),\n\nActivation('relu'),\n\nConv2D(64,kernel_size=(3,3),strides=[2,2],padding='same',kernel_initializer=keras.initializers.glorot_uniform(),name='conv3'),\n\nActivation('relu'),\n\nMaxPool2D(pool_size=(2, 2)),\n\nFlatten(),\n\nDense(10),\n\nDense(1,activation='softmax')\n\n])\n\n<\/code><\/pre>\n\n<p>When i compile my model with loss as sparse_categorical_crossentropy the fit function gives an error as mentioned below. <\/p>\n\n<p>Received a label value of 9 which is outside the valid range of [0, 1).  Label values: 9 5 9 7 2 7<\/p>\n\n<p>When compiling it with loss as mean_squared_error, the fit function works but loss is very high and accuracy is very low. I can understand the reason for mean_squared_error giving such results, but unable to understand the sparse_categorical_crossentropy not working. How do i fix this issue ?<\/p>\n\n<p>Thanks\/Sachin<\/p>\n\n<p>I am trying to create a Keras model<\/p>\n","answers":[]}
+{"QuestionId":57865247,"AnswerCount":0,"Tags":"<python><python-3.x><keras>","CreationDate":"2019-09-10T06:36:29.497","AcceptedAnswerId":null,"OwnerUserId":8173877.0,"Title":"Behavior of fit_generator in combination with flow_from_directory at runtime","Body":"<p>I'm trying to internalize how the keras image augmentation and fitting work together based on the <strong>ImageDataGenerator.flow_from_directory<\/strong> and the <strong>fit_generator()<\/strong> functions.<\/p>\n\n<p>I'm trying to figure out how the generator is working behind the scenes so I know what to tweak and how, right now here is my understanding.<\/p>\n\n<p><strong><em>I call train_gen\/test_gen.flow_from_directory() with batch size 64, class mode categorical and pass it into fit_generator with steps per epoch as 100, and epochs as 10.<\/em><\/strong><\/p>\n\n<p>When the above happens during epoch 1 <strong>am I generating 100 steps * 64 batched<\/strong> images? <\/p>\n\n<p>Also <strong>are all 64 batched images from the generator based on a single source image<\/strong>, changing on the next step? So if I only had two images, one in each class and called it would I be generating 64 augmented images from 'image 1', or would I be generating 64 total augmented images randomly distributed from 'image 1' and 'image 2'? <\/p>\n\n<p>If the above questions answer is that it is not generating from a single image, <strong>is it generating from a single class?<\/strong> So if 'image 1' and 'image 2' were both cats and I introduced 'image 3', a dog, and ran the training would the first step be all cats or all dogs and not a combination of the two? <\/p>\n\n<p>My worry is that if I do image_count \/ batch_size amount of steps per epoch that I am only getting X amount of my training data, and that I will overfit certain classes, if that makes any sense.<\/p>\n","answers":[]}
+{"QuestionId":57865460,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><machine-learning><keras>","CreationDate":"2019-09-10T06:53:23.800","AcceptedAnswerId":57866964.0,"OwnerUserId":2856460.0,"Title":"Keras model passing compilation but crashing on Value Error at runtime","Body":"<p>I'm trying to create a basic RNN that counts the number of 1's it's seen and returns there sum. This is my first attempt at making an RNN and I'm quite new to ML and Keras so any help or advice is really appreciated!<\/p>\n\n<p>I've played around with the model and found that using an Embedding layer to feed the 1d data into the LSTM. This allowed the model to compile properly however it is crashing once it tries to perform the first epoch.<\/p>\n\n<pre><code>values = []\nanswers = []\nfor i in range(0,5):\n    random = randint(1,101)\n    values.append([int(x) for x in bin(random)[2:]])\n    answers.append(len([ones for ones in values[-1] if ones==1]))\nvalues = np.array(values)\nprint(\"shape:\",values.shape[0])\nprint(\"values:\",values)\nprint(\"answers:\",answers)\n\nRNN = layers.LSTM\nmodel = Sequential()\nmodel.add(layers.Embedding(1, 1, input_length = 1))\nmodel.add(RNN(1,return_sequences=True))\nmodel.add(layers.Flatten())\nmodel.add(layers.Dense(1, activation=\"relu\"))\n\nadam = optimizers.adam(lr=0.01)\nmodel.compile(optimizer=adam,\n             loss=\"mean_squared_error\")\n\nprint (model.summary())\nmodel.fit(values,np.array(answers),epochs=5)\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<pre><code>shape: 5\nvalues: [list([1]) list([1, 0, 1, 1, 1, 1, 0]) list([1, 1, 0, 0, 1])\n list([1, 1, 0, 1]) list([1, 0, 0, 0, 1, 0])]\nanswers: [1, 5, 3, 3, 2]\n\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nembedding_6 (Embedding)      (None, 1, 1)              1         \n_________________________________________________________________\nlstm_33 (LSTM)               (None, 1, 1)              12        \n_________________________________________________________________\nflatten_19 (Flatten)         (None, 1)                 0         \n_________________________________________________________________\ndense_62 (Dense)             (None, 1)                 2         \n=================================================================\nTotal params: 15\nTrainable params: 15\nNon-trainable params: 0\n_________________________________________________________________\nNone\nEpoch 1\/5\n---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n<\/code><\/pre>\n\n<p>The layers are pretty basic and connect to each other nicely with the help of Flatten(). However, upon running the code I get a Value error and I can't quite figure out why.<\/p>\n\n<pre><code>ValueError                                Traceback (most recent call last)\n&lt;ipython-input-129-f5e242474ef9&gt; in &lt;module&gt;\n     11 \n     12 print (model.summary())\n---&gt; 13 model.fit(values,np.array(answers),epochs=5)\n\n~\\Anaconda3\\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   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~\\Anaconda3\\lib\\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~\\Anaconda3\\lib\\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~\\Anaconda3\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py in _call(self, inputs)\n   2653                 array_vals.append(\n   2654                     np.asarray(value,\n-&gt; 2655                                dtype=tf.as_dtype(tensor.dtype).as_numpy_dtype))\n   2656         if self.feed_dict:\n   2657             for key in sorted(self.feed_dict.keys()):\n\n~\\Anaconda3\\lib\\site-packages\\numpy\\core\\numeric.py in asarray(a, dtype, order)\n    536 \n    537     \"\"\"\n--&gt; 538     return array(a, dtype, copy=False, order=order)\n    539 \n    540 \n\nValueError: setting an array element with a sequence.\n<\/code><\/pre>\n\n<p>If anyone has a better way that I should format the model\/data or general suggestions please comment or answer.<\/p>\n","answers":[{"AnswerId":"57866964","CreationDate":"2019-09-10T08:36:17.227","ParentId":null,"OwnerUserId":"5652313","Title":null,"Body":"<p>The problem is in <code>values = np.array(values)<\/code> line. Calling <code>np.array<\/code> on list of list which are of different lengths creates a weird structure of a numpy array of objects (not numbers) which are other numpy arrays with numbers. This eventually gives you the <code>ValueError<\/code>.<\/p>\n\n<p>Every data batch <strong>must consists of sequences of the same length<\/strong>. If you do not want the RNN to work with sequences different lengths, it is indeed possible, but you need to add masking (Ctrl+F for \"masking\" in the <a href=\"https:\/\/keras.io\/layers\/recurrent\/\" rel=\"nofollow noreferrer\">Keras documentation<\/a> and probably something little bit more complex than a <code>Sequential<\/code>.<\/p>\n"}]}
+{"QuestionId":57865662,"AnswerCount":0,"Tags":"<python><tensorflow><keras><pruning>","CreationDate":"2019-09-10T07:07:17.073","AcceptedAnswerId":null,"OwnerUserId":11165362.0,"Title":"Threshold based weight pruning in keras\/tensorflow?","Body":"<p>Is there a way to prune a Keras model based on a certain threshold? The method prune_low_magnitude() from the tensorflow-model-optimization toolkit only allows pruning based on a defined final sparsity (percentage of remaining weigths). Is there a function I am missing or do you know a workaround?<\/p>\n","answers":[]}
+{"QuestionId":57865705,"AnswerCount":0,"Tags":"<tensorflow><keras-layer><tensorflow-lite><tf.keras>","CreationDate":"2019-09-10T07:10:19.273","AcceptedAnswerId":null,"OwnerUserId":5892689.0,"Title":"tflite converter from a custom keras layer","Body":"<p>I'm getting a TypeError when trying to convert a keras .h5 file to tflite.\nThe new layer is a gaussian kernel (Radial Basis Layer).<br>\nTo be able to save and load the keras model I defined also the get_config() method in the custom layer. So I'm able to save and load the model correctly.<\/p>\n\n<pre><code>class RBFLayer(Layer):\n    def __init__(self, output_dim, centers=None, tol = 1e-6, gamma=0, **kwargs):\n        super(RBFLayer, self).__init__(**kwargs)\n        self.centers_ = centers\n        self.output_dim= output_dim\n        self.gamma_ = gamma\n        self.tol_ = tol\n\n    def build(self, input_shape):\n        self.mu = K.variable(self.centers_, name='centers')\n        self.gamma = K.variable(self.gamma_, name='gamma')\n        self.tol = K.constant(self.tol_,name='tol')            \n        super(RBFLayer, self).build(input_shape)\n    def call(self, inputs): #Kernel radial\n        a,b = self.mu.shape\n        diff = K.reshape( K.tile(inputs,(1,a))-K.reshape(self.mu,(1,-1)), (-1,a,b))\n        l2 =   K.sum(K.pow(diff, 2), axis=-1)\n        res =  K.exp(-1 * self.gamma * l2)\n        mask = K.greater( res, self.tol)\n        return K.switch(mask, res, K.zeros_like(res)) \n    def compute_output_shape(self, input_shape):\n        return (input_shape[0], self.output_dim)\n    def get_config(self):\n        config = {\n        'output_dim': self.output_dim,\n        'centers': self.centers_,\n        'gamma': self.gamma_,\n        'tol': self.tol_\n        }\n        base_config = super(RBFLayer, self).get_config()\n        return dict(list(base_config.items()) + list(config.items()))\n<\/code><\/pre>\n\n<p>Now I want to save the model to tflite. I use TFLiteConverter from keras file including also 'custom_objects'.<\/p>\n\n<pre><code>def save_tflite(self, base_name):\n        file =base_name +'.h5'\n        converter = tf.lite.TFLiteConverter.from_keras_model_file(file, custom_objects={'RBFLayer':RBFLayer})\n        tflite_model = converter.convert()\n        open(base_name+\".tflite\", \"wb\").write(tflite_model)\n<\/code><\/pre>\n\n<p>I expect to get the tflite model file including the K.variables used while training the complete model (centers, tol, gamma). <\/p>\n\n<p>When converting I get these error messages:<\/p>\n\n<pre><code>airgorbn.save_tflite( base_name)\nTraceback (most recent call last):\n\n  File \"&lt;ipython-input-7-cdaa1ec46233&gt;\", line 1, in &lt;module&gt;\n    airgorbn.save_tflite( base_name)\n\n  File \"C:\/Users\/AIRFI\/Hospital\/keras_RadialBasis.py\", line 158, in save_tflite\n    converter = tf.lite.TFLiteConverter.from_keras_model_file(file, custom_objects={'RBFLayer':RBFLayer})\n\n  File \"C:\\Users\\AIRFI\\Anaconda3\\envs\\tf\\lib\\site-packages\\tensorflow\\lite\\python\\lite.py\", line 747, in from_keras_model_file\n    keras_model = _keras.models.load_model(model_file, custom_objects)\n\n  File \"C:\\Users\\AIRFI\\Anaconda3\\envs\\tf\\lib\\site-packages\\tensorflow\\python\\keras\\saving\\save.py\", line 146, in load_model\n    return hdf5_format.load_model_from_hdf5(filepath, custom_objects, compile)\n\n  File \"C:\\Users\\AIRFI\\Anaconda3\\envs\\tf\\lib\\site-packages\\tensorflow\\python\\keras\\saving\\hdf5_format.py\", line 212, in load_model_from_hdf5\n    custom_objects=custom_objects)\n\n  File \"C:\\Users\\AIRFI\\Anaconda3\\envs\\tf\\lib\\site-packages\\tensorflow\\python\\keras\\saving\\model_config.py\", line 55, in model_from_config\n    return deserialize(config, custom_objects=custom_objects)\n\n  File \"C:\\Users\\AIRFI\\Anaconda3\\envs\\tf\\lib\\site-packages\\tensorflow\\python\\keras\\layers\\serialization.py\", line 89, in deserialize\n    printable_module_name='layer')\n\n  File \"C:\\Users\\AIRFI\\Anaconda3\\envs\\tf\\lib\\site-packages\\tensorflow\\python\\keras\\utils\\generic_utils.py\", line 192, in deserialize_keras_object\n    list(custom_objects.items())))\n\n  File \"C:\\Users\\AIRFI\\Anaconda3\\envs\\tf\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\sequential.py\", line 353, in from_config\n    model.add(layer)\n\n  File \"C:\\Users\\AIRFI\\Anaconda3\\envs\\tf\\lib\\site-packages\\tensorflow\\python\\training\\tracking\\base.py\", line 457, in _method_wrapper\n    result = method(self, *args, **kwargs)\n\n  File \"C:\\Users\\AIRFI\\Anaconda3\\envs\\tf\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\sequential.py\", line 154, in add\n    'Found: ' + str(layer))\n\nTypeError: The added layer must be an instance of class Layer. Found: &lt;__main__.RBFLayer object at 0x0000017D3A75AC50&gt;\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57865710,"AnswerCount":0,"Tags":"<python><tensorflow><object-detection-api><mobilenet>","CreationDate":"2019-09-10T07:10:53.563","AcceptedAnswerId":null,"OwnerUserId":5908470.0,"Title":"multiple bounding boxes are always generated during validation on a model trained with the Object Detection API","Body":"<p>I am using the Object Detection Api to train an object detection network with only one class. For my first test with this API I used the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/samples\/configs\/ssd_mobilenet_v2_coco.config\" rel=\"nofollow noreferrer\">SSDMobileNetV2 pipeline config<\/a>. During the validation phases I could observe on Tensorboard that only one or two bounding boxes were predicted.<\/p>\n\n<p>Then I applied the following changes to the pipeline configuration:<\/p>\n\n<pre><code>model {\n  ssd {\n    num_classes: 1\n    [...]\n}\n\nanchor_generator {\n      ssd_anchor_generator {\n        num_layers: 6\n        min_scale: 0.1\n        max_scale: 0.95\n        aspect_ratios: 1.0\n        aspect_ratios: 1.6\n        aspect_ratios: 0.7\n        aspect_ratios: 1.3\n        aspect_ratios: 0.75\n      }\n    }\n\nhard_example_miner {\n        num_hard_examples: 3000\n        iou_threshold: 0.99\n        loss_type: LOCALIZATION\n        max_negatives_per_positive: 3\n        min_negatives_per_image: 3\n      }\n\npost_processing {\n      batch_non_max_suppression {\n        score_threshold: 1e-8\n        iou_threshold: 0.6\n        max_detections_per_class: 6\n        max_total_detections: 6\n      }\n      score_converter: SIGMOID\n    }\n\neval_config: {\n  num_examples: 8000\n  num_visualizations: 70\n\n}\n<\/code><\/pre>\n\n<p>that is: I set the number of classes to 1, I changed the aspect ratio of the bounding boxes according to the ones in my dataset, I set the loss type to LOCALIZATION in the hard example miner because I am not interested in classification and I limited the number of detections per image to 6.<\/p>\n\n<p>Furthermore, I enriched my dataset by randomly inserting objects that were typically detected (false positives) in previous tests.<\/p>\n\n<p>I have been training the model for a while (the training loss is about 0.5) and I see in Tensorboard that 6 bounding boxes are always generated for each validation image, with high confidence score. Some of them are correct (more or less), while others are not. This has been happening since the beginning of the training. <\/p>\n\n<p>Why is this happening? Can I expect the number of bounding boxes to decrease later?<\/p>\n\n<p><strong>EDIT<\/strong>: it seems that this behavior is caused by the <code>loss_type<\/code> parameter: if I set it to BOTH or CLASSIFICATION instead of LOCALIZATION, I have only 0 to 2 bounding boxes predicted during validation, without false positives. I don't understand why.<\/p>\n","answers":[]}
+{"QuestionId":57865942,"AnswerCount":0,"Tags":"<tensorflow><neural-network><conv-neural-network><training-data><attention-model>","CreationDate":"2019-09-10T07:26:43.627","AcceptedAnswerId":null,"OwnerUserId":5341024.0,"Title":"Strange output of a trained neural network with attention mechanism","Body":"<p>I have implemented the following network in tensorflow:\n<a href=\"https:\/\/i.stack.imgur.com\/kbjck.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/kbjck.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>There are two streams: a motion and an appearance stream. The two stream is connected at two points with simple element-wise multiplication (with the learned attention mask). First I wanted to do a short overfitting test to see whether everything is OK in the network, so I trained it on few data. The loss plot seemed OK, it converges:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/gzOuu.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/gzOuu.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>meaning that the networks output is close to the reference data! <strong>The loss is around 0.01<\/strong>.<\/p>\n\n<p>The learned attention mask that is multiplied with one of the motion stream layer is the following:\n<a href=\"https:\/\/i.stack.imgur.com\/AXAVB.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/AXAVB.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>But the strange thing is if I try out the trained network <strong>with the same input data<\/strong> the result is the following:\n<a href=\"https:\/\/i.stack.imgur.com\/Fqjip.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Fqjip.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/0rkUT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0rkUT.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>From the result we can see that the network's output is similar to the reference but it is in other range and shifted -- <strong>the loss here is around 1<\/strong>. But it should exactly overfit the reference (with loss ~ 0.01) based on the acquired loss in the training. The network output  is not compatible with the loss! How is this possible?<\/p>\n\n<p>If I remove the binding part of the streams -- that is the multiplication with the mask:<\/p>\n\n<pre><code>    # Multiply with appreance mask\n    conv_result = tf.multiply(conv_result, mask2)\n<\/code><\/pre>\n\n<p>then I use only the motion stream. In that case I could obtain a correct overfit. \nSo I deduce something might be wrong at the binding of the streams. But it does not explain the phenomenon of small loss and large network output difference in the same time.<\/p>\n\n<p>Do you have any idea where can be the problem?<\/p>\n","answers":[]}
+{"QuestionId":57866058,"AnswerCount":2,"Tags":"<tensorflow><machine-learning><keras>","CreationDate":"2019-09-10T07:34:09.393","AcceptedAnswerId":57867573.0,"OwnerUserId":4725522.0,"Title":"Keras not detecting GPU, but tensorflow is","Body":"<p>I am trying to get my keras back-end to work, but it keeps telling that there is no GPU, but my tensorflow say otherwise.<\/p>\n\n<p>tensorflow<\/p>\n\n<pre><code>from tensorflow.python.client import device_lib\nprint(device_lib.list_local_devices())\n<\/code><\/pre>\n\n<p>Ouput<\/p>\n\n<pre><code>[name: \"\/device:CPU:0\"\ndevice_type: \"CPU\"\nmemory_limit: 268435456\nlocality {\n}\nincarnation: 7600201089948511065\n, name: \"\/device:XLA_CPU:0\"\ndevice_type: \"XLA_CPU\"\nmemory_limit: 17179869184\nlocality {\n}\nincarnation: 9640434864482537344\nphysical_device_desc: \"device: XLA_CPU device\"\n, name: \"\/device:XLA_GPU:0\"\ndevice_type: \"XLA_GPU\"\nmemory_limit: 17179869184\nlocality {\n}\nincarnation: 15836333527683385212\nphysical_device_desc: \"device: XLA_GPU device\"\n]\n<\/code><\/pre>\n\n<p>keras<\/p>\n\n<pre><code>from keras import backend as K\nK.tensorflow_backend._get_available_gpus()\n<\/code><\/pre>\n\n<p>Output<\/p>\n\n<pre><code>[]\n<\/code><\/pre>\n\n<p>Tensorflow-gpu works, but keras can't see for some reason<\/p>\n\n<p>Cuda is installed<\/p>\n\n<pre><code>tensorflow-gpu = 1.14.0\nkeras = 2.0.8\npython = 3.6\n<\/code><\/pre>\n","answers":[{"AnswerId":"57867573","CreationDate":"2019-09-10T09:13:19.140","ParentId":null,"OwnerUserId":"4755954","Title":null,"Body":"<p>Few things that you need to confirm before this,<\/p>\n\n<p>Which GPU are you using? Only NVIDIA works.\nDo you have tensorflow-gpu installed (keras uses this as its backend)?\nDo you have CUDA installed?<\/p>\n\n<p>Since tensorflow is detecting GPU, then its either an issue with on of the above points, or you may try upgrading the various packages. <\/p>\n\n<p>Lastly, if nothing else works I would recommend using the high level API in tensorflow2 which is actually the same as keras.<\/p>\n\n<p>Hope this helps.<\/p>\n"},{"AnswerId":"58318853","CreationDate":"2019-10-10T08:44:34.580","ParentId":null,"OwnerUserId":"4725522","Title":null,"Body":"<p>Add ppa to ubuntu repos, they have the newest driver. Risky, but worth it!<\/p>\n\n<pre><code>sudo add-apt-repository ppa:graphics-drivers\/ppa\nsudo apt update\n<\/code><\/pre>\n\n<p>Try to install newest cuda driver<\/p>\n\n<pre><code>sudo ubuntu-drivers autoinstall\n<\/code><\/pre>\n\n<p>If this fails (like it did in my case), do sudo apt install for all dependencies it complains about.<\/p>\n\n<p>In my case:<\/p>\n\n<pre><code>sudo apt install libnvidia-gl-435 nvidia-dkms-435 nvidia-kernel-source-435 \n<\/code><\/pre>\n\n<p>Tried again<\/p>\n\n<pre><code>sudo ubuntu-drivers autoinstall\n<\/code><\/pre>\n\n<p>reboot computer<\/p>\n\n<p>Create a new conda env<\/p>\n\n<pre><code>conda create -n tf python=3.7\n\nsource activate tf\n\n\nconda install -c anaconda keras-gpu tensorflow-gpu jupyter pytables pandas seaborn scikit-learn gensim -y\n<\/code><\/pre>\n"}]}
+{"QuestionId":57866295,"AnswerCount":1,"Tags":"<keras><lstm>","CreationDate":"2019-09-10T07:50:43.577","AcceptedAnswerId":null,"OwnerUserId":7631183.0,"Title":"ValueError: Input arrays should have the same number of samples as target arrays LSTM Keras","Body":"<p>Here is the part for data preparing - I just want the data to be in the correct shape <\/p>\n\n<pre><code>x_train , x_test, y_train, y_test = train_test_split(input_data, y , test_size = 0.2 , random_state = 33)\nprint(x_train.shape)\nprint(y_train.shape)\n<\/code><\/pre>\n\n<blockquote>\n  <p>(200, 3) <\/p>\n  \n  <p>(200, 1)<\/p>\n<\/blockquote>\n\n<pre><code>#Converting them into numpy arrays \ninput_x_train = x_train.as_matrix()\ninput_y_train = y_train.as_matrix()\nprint(input_x_train.shape)\nprint(input_y_train.shape)\n<\/code><\/pre>\n\n<blockquote>\n  <p>(200, 3) <\/p>\n  \n  <p>(200, 1)<\/p>\n<\/blockquote>\n\n<pre><code>input_x_test = x_test.as_matrix()\ninput_y_test = y_test.as_matrix()\nprint(input_x_test.shape)\nprint(input_y_test.shape)\n<\/code><\/pre>\n\n<blockquote>\n  <p>(51, 3) <\/p>\n  \n  <p>(51, 1)<\/p>\n<\/blockquote>\n\n<pre><code>#Reshaping into LSTM input format \ninput_x = input_x_train.reshape((1, input_x_train.shape[0], input_x_train.shape[1]))\nprint(input_x.shape)\n<\/code><\/pre>\n\n<blockquote>\n  <p>(1, 200, 3)<\/p>\n<\/blockquote>\n\n<p>Then I built my model like this: <\/p>\n\n<pre><code>import keras\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Dropout, Activation\nfrom keras.layers.recurrent import LSTM \nfrom keras.layers.normalization import BatchNormalization\nmodel = Sequential()\nmodel.add(LSTM(32, input_shape=(200, 3)))\nmodel.add(Dense(1))\nmodel.compile(loss='mse', optimizer='adam')\nmodel.fit(input_x, input_y_train, epochs=1, batch_size=16) \n<\/code><\/pre>\n\n<p>But I am getting this error <\/p>\n\n<blockquote>\n  <p>ValueError: Input arrays should have the same number of samples as\n  target arrays. Found 1 input samples and 200 target samples.<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"57874634","CreationDate":"2019-09-10T16:12:29.897","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>The <code>input_shape<\/code> parameter should not include your batch size. Given that each sample of yours dataset has three features, you should set <code>input_shape=(3,)<\/code>.<\/p>\n\n<p>In addition, you should not reshape your batch to <code>(1, batch_size, 3)<\/code>. I'm not sure why you're doing that, but as far as I can tell that will break things. Remove that line entirely.<\/p>\n"}]}
+{"QuestionId":57866547,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-lite>","CreationDate":"2019-09-10T08:10:12.453","AcceptedAnswerId":null,"OwnerUserId":10484743.0,"Title":"Running hand detection model tensorflow-lite through Python","Body":"<p>When I run the model for determining hands <strong>mediapipe<\/strong> <a href=\"https:\/\/github.com\/google\/mediapipe\/blob\/master\/mediapipe\/models\" rel=\"nofollow noreferrer\">hand_landmark.tflite<\/a> through Python, I encounter slow work in the process of determining hands! More precisely, it is\n<code>interpreter.invoke()<\/code>\nFrame rate drops sharply from 40 to 4!\nHow can I increase the frame rate to 20 or 25?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    # ...\n\n    # Called before the loop\n    def start(self):\n        self._interpreter = tf.lite.Interpreter(model_path = path_to_model)\n\n        self._interpreter.allocate_tensors()\n        self._input_details = self._interpreter.get_input_details()\n        self._output_details = self._interpreter.get_output_details()\n\n    # ...\n\n    # Called in a loop\n    def generate(self, img):\n        if img.size == 0:\n            return False\n\n        img_resize = cv2.resize(img, (256, 256))\n\n        img_resize_expanded = np.expand_dims(img_resize, axis = 0)\n        image_np_expanded = (np.float32(img_resize_expanded) - 0) \/ 255\n\n        self._interpreter.set_tensor(self._input_details[0]['index'], image_np_expanded)\n\n        self._interpreter.invoke()  # Slow !!!!!!!!!!!!!!!!!!!!!!\n\n        output_data = self._interpreter.get_tensor(self._output_details[0]['index'])\n\n        return output_data\n<\/code><\/pre>\n","answers":[{"AnswerId":"57931528","CreationDate":"2019-09-14T00:20:09.033","ParentId":null,"OwnerUserId":"9867252","Title":null,"Body":"<p>Additional information would make this easier to respond to:<\/p>\n\n<p>1) Are you running this on Desktop or a mobile phone? TFLite is optimized for mobile (e.g. has optimization for arm) and support to run quickly on desktops (e.g. x86) is still a work in progress<\/p>\n\n<p>2) MediaPipe is likely using the GPU delegate API when running on Android. Take a look at the tensorflow.org\/lite docs on running with the GPU. See what accelerators they take advantage of and find how to enable it in the above docs. If they actually run on the CPU and use quantization, that is something to look into.<\/p>\n"}]}
+{"QuestionId":57867006,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-09-10T08:38:44.460","AcceptedAnswerId":57870175.0,"OwnerUserId":11926492.0,"Title":"Correcting input dimensions for CNN, LSTM based classifier using Keras, Python","Body":"<p>I'm working on implementing a 2D (Perhaps 1D) CNN+ LSTM classifier for Network Traffic classification purposes. The CNN will essentially be used as a feature extractor and the LSTM would work for the classification. <\/p>\n\n<p>I have used the TimeDistributed layer to help combine the CNN and LSTM layers together (Code attached.) \nSince the input size varies dynamically, the number of data points has been indicated with None. \nno_rows=20 (Number of packets considered per flow for classification)\nno_cols=7 (Number of features considered for each packet)<\/p>\n\n<p>Despite using the TimeDistributed layer wrap, I am facing some input dimension issues. Not quite sure how to resolve this. <\/p>\n\n<p>Using Reshape as a layer to resolve this was one of the many fixes I came across but didn't work. Kindly let me know how to build this structure and how to fix my code.<\/p>\n\n<p>Thanks !<\/p>\n\n<p>(Using a Linux based AWS instance, Ubuntu 16.04 and Tensorflow backend to implement the code)<\/p>\n\n<ol>\n<li>Used Reshape layer from Keras core layers to fix the output of the CNN but did not resolve the issue. <\/li>\n<li>Had to remove the Flatten layer and replace it with GlobalMaxPooling2D layer due to the presence of dynamically changing input size.<\/li>\n<\/ol>\n\n<pre><code>X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size = 0.30, random_state = 36)\n\nmodel = Sequential()\n\n# Adding CNN Model layers\n\nmodel.add(TimeDistributed(Conv2D(32, kernel_size = (4 , 2), strides = 1, padding='valid', activation = 'relu', input_shape = (None,20,7,1))))\nmodel.add(TimeDistributed(BatchNormalization()))\nmodel.add(TimeDistributed(Conv2D(64, kernel_size = (4 , 2), strides = 1, padding='valid', activation = 'relu')))\nmodel.add(TimeDistributed(BatchNormalization()))\n#model.add(TimeDistributed(Reshape((-1,1))))\nmodel.add(TimeDistributed(GlobalMaxPooling2D()))\n#model.add(Reshape((1,1)))\n\n# Adding LSTM layers\n\nmodel.add(LSTM(128, recurrent_dropout=0.2))\nmodel.add(Dropout(rate = 0.2))\nmodel.add(Dense(100))\nmodel.add(Dropout(rate = 0.4))\nmodel.add(Dense(108))\n\nmodel.add(Dense(num_classes,activation='softmax'))\n\n\n# Compiling this model\n\nmodel.compile(loss = 'categorical_crossentropy', optimizer='adam',metrics = ['accuracy'])\n#print(model.summary())\n\n#Training the Network\n\nhistory = model.fit(X_train, Y_train, batch_size=32, epochs = 1, validation_data=(X_test, Y_test))\n<\/code><\/pre>\n\n<p>Running the code snippet mentioned above I face the following error message:<\/p>\n\n<pre><code>\"Input tensor must be of rank 3, 4 or 5 but was {}.\".format(n + 2))\nValueError: Input tensor must be of rank 3, 4 or 5 but was 2.\n\n<\/code><\/pre>\n","answers":[{"AnswerId":"57870175","CreationDate":"2019-09-10T11:49:47.760","ParentId":null,"OwnerUserId":"11952444","Title":null,"Body":"<p>One thing you can do is to make a batch of a fixed input (number of frames) from your video source and process that. The code for doing that would be:<\/p>\n\n<pre><code>def get_data(video_source, batch_size):\n    x_data = []\n    #Reading the Video from file path\n    cap = cv2.VideoCapture(video_source)    \n    for i in range(batch_size):\n        #To Store Frames\n        frames = []\n        for j in range(frame_to_process): #here we get frame_to_process\n            ret, frame = cap.read()\n            if not ret:\n                # print('No frames found!')\n                break\n            # converting to frmae gray\n            # frame = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY) \n            # resizing frame to a particular input shape\n            # frame = cv2.resize(frame,(30,30),interpolation=cv2.INTER_AREA)\n            frames.append(frame)\n        # appending each batch\n        x_data.append(frames)\n    return x_data\n\n# number of frames in each batch\nframe_to_process = 30\n# size of each batch\nbatch_size = 32\n# make batch of video inputs\nX_data = np.array(get_data(video_source, batch_size, frame_to_process))\n<\/code><\/pre>\n\n<p><strong>Tip:<\/strong> Also, instead of using <code>Conv2D<\/code> with <code>TimeDistrubuted<\/code> you can use <a href=\"https:\/\/keras.io\/layers\/recurrent\/\" rel=\"nofollow noreferrer\">ConvLSTM<\/a> which can give a little performance improvement.<\/p>\n\n<p>Anyways, if you want to process frames dynamically you can convert the code to Pytorch which has <a href=\"https:\/\/medium.com\/near-ai\/pytorch-dynamic-batching-f4df3dbe09ef\" rel=\"nofollow noreferrer\">Dynamic Graphs<\/a>, where you can give input with variable batch size.<\/p>\n\n<ul>\n<li><a href=\"https:\/\/ai.stackexchange.com\/questions\/3801\/what-is-a-dynamic-computational-graph\">Dynamic Computation Graph<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/46154189\/what-is-the-difference-of-static-computational-graphs-in-tensorflow-and-dynamic\">Difference between Static and Dynamic Graphs<\/a><\/li>\n<\/ul>\n"}]}
+{"QuestionId":57867091,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><tf.keras>","CreationDate":"2019-09-10T08:44:24.957","AcceptedAnswerId":null,"OwnerUserId":10210252.0,"Title":"tf.keras.Sequential binary classification model predicting [0.5, 0.5] or close to","Body":"<p>I am currently trying to build a model to classify whether or not the outcome of a given football match will be above or below 2.5 goals, based on the Home team, Away team &amp; game league, using a <code>tf.keras.Sequential<\/code> model in TensorFlow 2.0RC.<\/p>\n\n<p>The problem I am encountering is that my softmax results converge on <code>[0.5,0.5]<\/code> when using the <code>model.predict<\/code> method. What makes this odd is that my validation &amp; test accuracy and losses are about 0.94 &amp; 0.12 respectively after 1000 epochs of training, otherwise I would have put this down to an overfitting problem. I am aware that 1000 epochs is extremely likely to overfit, however, I want to understand why my accuracy increases until about 800 epochs in. My loss flattens at about 300 epochs.<\/p>\n\n<p>I have tried to alter the number of layers, number of units in each layer, the activation functions, optimizers and loss functions, number of epochs and learning rates, but can only seem to increase the losses. <\/p>\n\n<p>The results still seem to converge toward <code>[0.5,0.5]<\/code> regardless.<\/p>\n\n<p>The full code can be viewed at <a href=\"https:\/\/github.com\/AhmUgEk\/tensorflow_football_predictions\" rel=\"nofollow noreferrer\">https:\/\/github.com\/AhmUgEk\/tensorflow_football_predictions<\/a>, but below is an extract showing model composition.<\/p>\n\n<pre><code># Create Keras Sequential model:\nmodel = keras.Sequential()\n\nmodel.add(feature_layer)  # Input processing layer.\n\nmodel.add(Dense(units=32, activation='relu'))  # Hidden Layer 1.\nmodel.add(Dropout(rate=0.4))\nmodel.add(BatchNormalization())\n\nmodel.add(Dense(units=32, activation='relu'))  # Hidden Layer 2.\nmodel.add(Dropout(rate=0.4))\nmodel.add(BatchNormalization())\n\nmodel.add(Dense(units=2, activation='softmax'))  # Output layer.\n\n# Compile the model:\nmodel.compile(\n    optimizer=keras.optimizers.Adam(learning_rate=0.0001),\n    loss=keras.losses.MeanSquaredLogarithmicError(),\n    metrics=['accuracy']\n)\n\n# Compile the model:\nmodel.compile(\n    optimizer=keras.optimizers.Adam(learning_rate=0.0001),\n    loss=keras.losses.MeanSquaredLogarithmicError(),\n    metrics=['accuracy']\n)\n\n# Fit the model to the training dataset and validate against the \nvalidation dataset between epochs:\nmodel.fit(\n    train_dataset,\n    validation_data=val_dataset,\n    epochs=1000,\n    callbacks=[tensorboard_callback]\n)\n<\/code><\/pre>\n\n<p>I would expect to receive a result of <code>[0.282, 0.718]<\/code> for example for an input of:<\/p>\n\n<pre><code>model.predict_classes([np.array(['E0'], dtype='object'),\n    np.array(['Liverpool'], dtype='object'),\n    np.array(['Newcastle'], dtype='object')])[0]\n<\/code><\/pre>\n\n<p>but as per the above, receive a result of say <code>[0.5, 0.5]<\/code>.<\/p>\n\n<p>Am I missing something obvious here?<\/p>\n","answers":[{"AnswerId":"57878809","CreationDate":"2019-09-10T22:11:57.917","ParentId":null,"OwnerUserId":"11952444","Title":null,"Body":"<p>I had made some minor changes in the model. Now, I am not getting exactly [0.5, 0.5].<\/p>\n\n<p><strong>Result:<\/strong><\/p>\n\n<pre><code>[[0.61482537 0.3851746 ]\n [0.5121426  0.48785746]\n [0.48058605 0.51941395]\n [0.48913187 0.51086813]\n [0.45480043 0.5451996 ]\n [0.48933673 0.5106633 ]\n [0.43431875 0.5656812 ]\n [0.55314165 0.4468583 ]\n [0.5365097  0.4634903 ]\n [0.54371756 0.45628244]]\n<\/code><\/pre>\n\n<p><strong>Implementation:<\/strong><\/p>\n\n<pre><code>import datetime\nimport os\nimport numpy as np\nimport pandas as pd\nimport tensorflow as tf\n\nfrom gpu_limiter import limit_gpu\nfrom pipe_functions import csv_to_df, dataframe_to_dataset\nfrom sklearn.model_selection import train_test_split\nfrom tensorflow import keras\nfrom tensorflow.keras.layers import BatchNormalization, Dense, DenseFeatures, Dropout, Input\nfrom tensorflow.keras.callbacks import TensorBoard, ModelCheckpoint\nimport tensorflow.keras.backend as K\nfrom tensorflow.data import Dataset\n\n\n# Test GPU availability and instantiate memory growth limitation if True:\nif tf.test.is_gpu_available():\n    print('GPU Available\\n')\n    limit_gpu()\nelse:\n    print('Running on CPU')\n\ndf = csv_to_df(\".\/csv_files\")\n\n# Format &amp; organise imported data, making the \"Date\" column the new index:\ndf['Date'] = pd.to_datetime(df['Date'])\ndf = df[['Date', 'Div', 'HomeTeam', 'AwayTeam', 'FTHG', 'FTAG']].dropna().set_index('Date').sort_index()\ndf['Over_2.5'] = (df['FTHG'] + df['FTAG'] &gt; 2.5).astype(int)\ndf = df.drop(['FTHG', 'FTAG'], axis=1)\n\n\n# Split data into training, validation and testing data:\n# Note: random_state variable set to ensure reproducibility.\ntrain, test = train_test_split(df, test_size=0.05, random_state=42)\ntrain, val = train_test_split(train, test_size=0.05, random_state=42)\n\n# print(df['Over_2.5'].value_counts())  # Check that data is balanced.\n\n\n# Create datasets from train, val &amp; test dataframes:\ntarget_col = 'Over_2.5'\nbatch_size = 32\n\ndef df_to_dataset(features: np.ndarray, labels: np.ndarray, shuffle=True, batch_size=8) -&gt; Dataset:\n    ds = Dataset.from_tensor_slices(({\"feature\": features}, {\"target\": labels}))\n    if shuffle:\n        ds = ds.shuffle(buffer_size=len(features))\n    ds = ds.batch(batch_size)\n    return ds\n\ndef get_feature_transform() -&gt; DenseFeatures:\n    # Format features into feature columns to ensure data is in the correct format for feeding into the model:\n    feature_cols = []\n    for column in filter(lambda x: x != target_col, df.columns):\n        feature_cols.append(tf.feature_column.embedding_column(tf.feature_column.categorical_column_with_vocabulary_list(\n            key=column, vocabulary_list=df[column].unique()), dimension=5))\n    return DenseFeatures(feature_cols)\n\n# Transforms all features into dense tensors.\nfeature_transform = get_feature_transform()\ntrain_features = feature_transform(dict(train)).numpy()\nval_features = feature_transform(dict(val)).numpy()\ntest_features = feature_transform(dict(test)).numpy()\n\ntrain_dataset = df_to_dataset(train_features, train[target_col].values, shuffle=True, batch_size=batch_size)\nval_dataset = df_to_dataset(val_features, val[target_col].values, shuffle=True, batch_size=batch_size)  # Shuffle not required to validation data.\ntest_dataset = df_to_dataset(test_features, test[target_col].values, shuffle=True, batch_size=batch_size)  # Shuffle not required to test data.\n\n# Create Keras Functional API:\n# Create a feature layer from the feature columns, to be placed at the input layer of the model:\n\ndef build_model(input_shape: tuple) -&gt; keras.Model:\n\n    input_layer = keras.Input(shape=input_shape, name='feature')\n\n    model = Dense(units=1028, activation='relu', kernel_initializer='normal', name='dense0')(input_layer)  # Hidden Layer 1.\n    model = BatchNormalization(name='bc0')(model)\n\n    model = Dense(units=1028, activation='relu', kernel_initializer='normal', name='dense1')(model)  # Hidden Layer 2.\n    model = Dropout(rate=0.1)(model)\n    model = BatchNormalization(name='bc1')(model)\n\n    model = Dense(units=100, activation='relu', kernel_initializer='normal', name='dense2')(model)  # Hidden Layer 3.\n    model = Dropout(rate=0.25)(model)\n    model = BatchNormalization(name='bc2')(model)\n\n    model = Dense(units=50, activation='relu', kernel_initializer='normal', name='dense3')(model)  # Hidden Layer 4.\n    model = Dropout(rate=0.4)(model)\n    model = BatchNormalization(name='bc3')(model)\n\n    output_layer = Dense(units=2, activation='softmax', kernel_initializer='normal', name='target')(model)  # Output layer.\n\n    model = keras.Model(inputs=input_layer, outputs=output_layer, name='better-than-chance')\n\n    # Compile the model:\n    model.compile(\n        optimizer=keras.optimizers.Adam(learning_rate=0.001),\n        loss='mse',\n        metrics=['accuracy']\n    )\n    return model\n\n# # Create a TensorBoard log file (time appended) directory for every run of the model:\n# directory = \".\\\\logs\\\\\" + str(datetime.datetime.now().strftime(\"%Y%m%d-%H%M%S\"))\n# os.mkdir(directory)\n\n# # Create a TensorBoard callback to log a record of model performance for every 1 epoch:\n# tensorboard_callback = TensorBoard(log_dir=directory, histogram_freq=1, write_graph=True, write_images=True)\n\n# Run \"tensorboard --logdir .\\logs\" in anaconda prompt to review &amp; compare logged results.\n# Note: Make sure that the correct environment is activated before running.\n\nmodel = build_model((train_features.shape[1],))\nmodel.summary()\n\n# checkpoint = ModelCheckpoint('model-{epoch:03d}.h5', verbose=1, monitor='val_loss',save_best_only=True, mode='auto')  \n\n# Fit the model to the training dataset and validate against the validation dataset between epochs:\nmodel.fit(\n    train_dataset,\n    validation_data=val_dataset,\n    epochs=10)\n    # callbacks=[checkpoint]\n\n# Saves and reloads model.\n# model.save(\".\/model.h5\")\n# model_from_saved = keras.models.load_model(\".\/model.h5\")\n\n# Evaluate model accuracy against test dataset:\n# scores, accuracy = model.evaluate(train_dataset)\n# print('Accuracy:', accuracy)\n\n##############\n## OPTIONAL ##\n##############\n\n# DUBUGGING\n# inp = model.input                                           # input placeholder\n# outputs = [layer.output for layer in model.layers]          # all layer outputs\n# functors = [K.function([inp], [out]) for out in outputs]    # evaluation functions\n\n# # Testing\n# layer_outs = [func([test_features]) for func in functors]\n# print(layer_outs)\n\n\n# # # Form a prediction based on inputs:\nprediction = model.predict({\"feature\": test_features[:10]})\nprint(prediction)\n<\/code><\/pre>\n\n<ol>\n<li><p>One thing you can do is to try some ensemble Learning methods like\n<a href=\"https:\/\/towardsdatascience.com\/random-forest-in-python-24d0893d51c0\" rel=\"nofollow noreferrer\">RandomForest<\/a>\nand\n<a href=\"https:\/\/machinelearningmastery.com\/gentle-introduction-xgboost-applied-machine-learning\/\" rel=\"nofollow noreferrer\">XGBoost<\/a>\nand compare the results.   <\/p><\/li>\n<li><p>You should try is to add other Key Performance Indicators(KPI)s in\nyour data and then try to fit the model.<\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":57867107,"AnswerCount":2,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-09-10T08:45:21.923","AcceptedAnswerId":null,"OwnerUserId":12045291.0,"Title":"How to choose from \"tensorflow.keras\" and \"tensorflow.python.keras\" in tensorflow 2.0?","Body":"<p>Currently in tensorflow 2.0, <code>Keras<\/code> can be imported by both <code>tensorflow.keras<\/code> and <code>tensorflow.python.keras<\/code>. What's the difference and how should I choose from these two imports?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.keras import *\nfrom tensorflow.python.keras import *\n<\/code><\/pre>\n\n<p>From my understanding, <code>tensorflow.python.keras<\/code> is pretty much the same as the python package Keras, while <code>tensorflow.keras<\/code> is implemented by TensorFlow which might be more compatible with TF, but do not have all packages as in Keras such as Layer, InputSpec, etc. <\/p>\n\n<p>One other issue is that for some IDEs like PyCharm, they cannot find or auto-complete packages from <code>tensorflow.keras<\/code>, and the temporary solution is to use <code>tensorflow.python.keras<\/code> instead. (from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/26502\" rel=\"nofollow noreferrer\">tensorflow issue<\/a>)<\/p>\n","answers":[{"AnswerId":"57867315","CreationDate":"2019-09-10T08:58:13.413","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p><code>tensorflow.keras<\/code> and <code>tensorflow.python.keras<\/code> are exactly the same package, what tells you that they are different?<\/p>\n\n<p>You are not supposed to use <code>tensorflow.python<\/code> imports directly, as per this answer: <a href=\"https:\/\/stackoverflow.com\/a\/47306203\/6108843\">https:\/\/stackoverflow.com\/a\/47306203\/6108843<\/a><\/p>\n"},{"AnswerId":"57867455","CreationDate":"2019-09-10T09:06:34.173","ParentId":null,"OwnerUserId":"11977985","Title":null,"Body":"<p>As of the latest version of Tensorflow, it is recommended to use tensorflow.keras\nThis is likely due to depreciation since TF is evolving rapidly.<\/p>\n"}]}
+{"QuestionId":57867171,"AnswerCount":0,"Tags":"<keras><conv-neural-network><pre-trained-model><transfer-learning><vgg-net>","CreationDate":"2019-09-10T08:49:00.653","AcceptedAnswerId":null,"OwnerUserId":3802483.0,"Title":"keras transfer weights from vgg16 to smaller network","Body":"<p>I first trained a small network as below:<\/p>\n\n<pre><code>model = models.Sequential()\nmodel.add(layers.Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', input_shape=(256, 256, 3)))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(Dropout(0.2))\nmodel.add(layers.Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform'))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(Dropout(0.2))\nmodel.add(layers.Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_uniform'))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(Dropout(0.2))\nmodel.add(layers.Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_uniform'))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(Dropout(0.2))\nmodel.add(layers.Flatten())\nmodel.add(layers.Dropout(0.5))  #Dropout for regularization\nmodel.add(layers.Dense(512, activation='relu'))\nmodel.add(layers.Dense(1, activation='sigmoid'))  #Sigmoid function at the end because we have just two classes\n<\/code><\/pre>\n\n<p>which performed around 70-75% accuracy with no overfitting issues.<\/p>\n\n<p>Then I decided to train a VGG16 with pretrained weights on Imagenet, where I freeze all the conv layers -but one-, and I add two fc layers, then train it.<\/p>\n\n<p>I tried to optimize the model using Adam optimizer (which performs worse even when I set small learning rates?) I use ReduceLROnPlateau and EarlyStopping, and my network performs around 80-82% accuracy however it overfits greatly (val loss is increasing, and there is a huge gap between the training and validation accuracy)<\/p>\n\n<p>Network is as below:<\/p>\n\n<pre><code>epochs = 50\ncallbacks = []\ndecay = 0.0\n\nearlyStopping = EarlyStopping(monitor='val_loss', patience=10, verbose=0, mode='min')\nmcp_save = ModelCheckpoint('.mdl_wts.hdf5', save_best_only=True, monitor='val_loss', mode='min')\nreduce_lr_loss = ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=3, verbose=1, epsilon=1e-5, mode='min')\n\n\n# build the VGG16 network\nbase_model = applications.VGG16(weights='imagenet', include_top=False, input_shape=(224, 224, 3))\nprint('Model loaded.')\n\n\nfor layer in base_model.layers[:15]:\n    layer.trainable = False\n\n#for layer in base_model.layers:\n#    layer.trainable = False\n\n# add new classifier layers\nflat1 = Flatten()(base_model.layers[-1].output)\nclass1 = Dense(128, kernel_regularizer=l2(0.01), bias_regularizer=l2(0.01), activation='relu', kernel_initializer='he_uniform')(flat1)\noutput = Dense(1, kernel_regularizer=l2(0.01), bias_regularizer=l2(0.01), activation='sigmoid')(class1)\n# define new model\nbase_model = Model(inputs=base_model.inputs, outputs=output)\n\nbase_model.compile(loss='binary_crossentropy',\n                   optimizer=optimizers.Adam(lr=1e-4, decay=decay),\n              metrics=['accuracy'])\n\nimport matplotlib.pyplot as plt\nfrom keras.preprocessing.image import ImageDataGenerator\nimport os\n\nos.environ[\"CUDA_VISIBLE_DEVICES\"]=\"0\"\n\ntrain_dir = '\/home\/d\/Desktop\/s\/data\/train'\neval_dir = '\/home\/d\/Desktop\/s\/data\/eval'\ntest_dir = '\/home\/d\/Desktop\/s\/data\/test'\n\n\n# create a data generator\ntrain_datagen = ImageDataGenerator(rescale=1.\/255,   #Scale the image between 0 and 1\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\nval_datagen = ImageDataGenerator(rescale=1.\/255)  #We do not augment validation data. we only perform rescale\n\ntest_datagen = ImageDataGenerator(rescale=1.\/255)  #We do not augment validation data. we only perform rescale\n\n# load and iterate training dataset\ntrain_generator = train_datagen.flow_from_directory(train_dir,  target_size=(224,224),class_mode='binary', batch_size=16, shuffle='True', seed=42)\n# load and iterate validation dataset\nval_generator = val_datagen.flow_from_directory(eval_dir,  target_size=(224,224),class_mode='binary', batch_size=16, shuffle='True', seed=42)\n# load and iterate test dataset\ntest_generator = test_datagen.flow_from_directory(test_dir,  target_size=(224,224), class_mode=None, batch_size=1, shuffle='False', seed=42)\n\n#The training part\n#We train for 64 epochs with about 100 steps per epoch\nhistory = base_model.fit_generator(train_generator,\n                              steps_per_epoch=train_generator.n \/\/ train_generator.batch_size,\n                              epochs=epochs,\n                              validation_data=val_generator,\n                              validation_steps=val_generator.n \/\/ val_generator.batch_size,\n                              callbacks=[earlyStopping, mcp_save, reduce_lr_loss])\n<\/code><\/pre>\n\n<p>my dataset is small and very different than Imagenet, I do a lot of image augmentation.<\/p>\n\n<p>So only thing I can think of is that the VGG16 network is far too complex for my dataset. I want to use my smaller network but with pretrained weights transferred to initial layers, how can I do this? Any other recommendations are also welcome.<\/p>\n","answers":[]}
+{"QuestionId":57867854,"AnswerCount":0,"Tags":"<tensorflow><distributed-tensorflow>","CreationDate":"2019-09-10T09:28:59.553","AcceptedAnswerId":null,"OwnerUserId":10330789.0,"Title":"The working mechanism of sess.run(tf.variables_initializer(var_list) with distributed tensorflow","Body":"<p>I'm running a distributed tensorflow application with 20 PSs and 100 Workers.<\/p>\n\n<p>And there is a chief worker named worker0 takes charge of some extra work like variable initialization.<\/p>\n\n<p>When i try to initialize some very large variables via sess.run(tf.variables_initializer([large_variable_list])).\nThe initialization is very slow,\nit seems that the variables are created in worker0 and then send to ps.<\/p>\n\n<p><b>So my question is what the exact mechanism of variable initialization with distributed tensorflow.\nIs is was created in worker and then sent to PS?<\/p>\n\n<p>And how can i trace the process of variable initialization?<\/b><\/p>\n\n<pre><code>log of worker0\n\n2019-09-09 08:20:12.238889 Start to initialize all variables\n2019-09-09 08:20:20.549082: W tensorflow\/core\/framework\/allocator.cc:124] Allocation of 12271335136 exceeds 10% of system memory.\n2019-09-09 08:20:36.105050: W tensorflow\/core\/framework\/allocator.cc:124] Allocation of 12271335136 exceeds 10% of system memory.\n2019-09-09 08:21:06.638078: W tensorflow\/core\/framework\/allocator.cc:124] Allocation of 12271335136 exceeds 10% of system memory.\n2019-09-09 08:21:27.582467: W tensorflow\/core\/framework\/allocator.cc:124] Allocation of 12271335136 exceeds 10% of system memory.\n2019-09-09 08:21:53.365509: W tensorflow\/core\/framework\/allocator.cc:124] Allocation of 12271335136 exceeds 10% of system memory.\n2019-09-09 08:42:32.114437 Initialize all variables success.\n2019-09-09 08:42:33.471465 Start to initialize local variables\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code># code sippet\n\nterm_embeddings = tf.get_variable(\n                    \"term_embeddings\", [383479222, 5], dtype=tf.float32,\n                    initializer=variable_initializer, partitioner=tf.fixed_size_partitioner(ps_count))\n\nglobal_var_init_op = tf.variables_initializer([v for v in tf.global_variables() if v not in reusable_barrier.get_variables()])  \n\nsess.run([global_var_init_op])\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57867902,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><computer-vision><object-detection><transfer-learning>","CreationDate":"2019-09-10T09:32:15.833","AcceptedAnswerId":null,"OwnerUserId":8145641.0,"Title":"Inception Model retrained on custom dataset (cotton and weed plant), but labeling other objects e.g. car, person or cards as either cotton or weed","Body":"<p>I'm training a plant detection model using my custom dataset to detect between <code>cotton<\/code> and <code>weed<\/code> plant. I'm getting help from this <a href=\"https:\/\/www.youtube.com\/watch?v=Rgpfk6eYxJA\" rel=\"nofollow noreferrer\">Youtube Video<\/a> and it's respective <a href=\"https:\/\/github.com\/EdjeElectronics\/TensorFlow-Object-Detection-API-Tutorial-Train-Multiple-Objects-Windows-10\" rel=\"nofollow noreferrer\">Github repository<\/a>. I followed the way same as this man instructed. Everything went well. Model is detecting many <code>cotton<\/code> and <code>other<\/code> (weed) plants on test images. But real problem I'm facing, it is labeling all other objects images e.g. car, person, kites as either <code>cotton<\/code> or <code>other<\/code>. Even I fed it pics of cards, it labeled them as <code>cotton<\/code> and <code>weed<\/code>.<\/p>\n\n<p><strong>I've uploaded all code files, output images, config files and labels on my <a href=\"https:\/\/github.com\/Master-HM\/Retraining-Inception-labeling-my-custom-specified-labels-even-to-strange-objects\" rel=\"nofollow noreferrer\">github<\/a>. You can see them.<\/strong><\/p>\n\n<p>Instructor in this video is using <code>faster_rcnn_inception_v2<\/code> and <code>tensorflowv1.5<\/code>. Process to follow is as:<\/p>\n\n<p>1: Download <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/079d67d9a0b3407e8d074a200780f3835413ef99\" rel=\"nofollow noreferrer\">tensorflow v1.5 commit<\/a> and <a href=\"http:\/\/download.tensorflow.org\/models\/object_detection\/faster_rcnn_inception_v2_coco_2018_01_28.tar.gz\" rel=\"nofollow noreferrer\">faster_rcnn_inception_v2<\/a>.<\/p>\n\n<p>2: Using tool <code>LabelImg<\/code>, bound box and label all desired object within image, this tool creates a <code>xml<\/code> file for each image, which specify bounding box coordinates of each image. Later these <code>xml<\/code> files are converted to <code>csv<\/code> files, which contain all information for each image in single <code>csv<\/code> file. My dataset contained cotton and weed, so using <code>LabelImg<\/code>, I labeled my whole data images as <code>cotton<\/code> and <code>other<\/code> (weed).<\/p>\n\n<p>3: Then scripts are given, that you've to modify to suit your requirements and run them. It take <code>csv<\/code> files as input, containing image data information and it creates <code>tf_record<\/code> file for both <code>train<\/code> and <code>validation<\/code> datasets.<\/p>\n\n<p>4: Now I created a <code>pbtxt<\/code> extension file, which contain my only two class labels <code>cotton<\/code> and <code>other<\/code>.<\/p>\n\n<p><strong>Label file<\/strong><\/p>\n\n<pre><code>item {\n    id: 1\n     name: 'cotton'\n}\nitem {\n    id: 2\n    name: 'other'\n}\n<\/code><\/pre>\n\n<p><strong>generate_tfrecord.py<\/strong><\/p>\n\n<pre><code>...\ndef class_text_to_int(row_label):\n\n    if row_label == 'cotton':\n\n        return 1\n\n    elif row_label == 'other':\n\n        return 2\n\n    else:\n\n        None\n...\n<\/code><\/pre>\n\n<p>5: configured  <code>faster_rcnn_inception_v2_pets.congif<\/code> file and modified it for two classes <code>cotton<\/code> and <code>other<\/code>, and changed <code>num_class=2<\/code><\/p>\n\n<p>6: Finally trained my model with 12 hours of processing and 171,000 steps with loss 0.02 average. After exporting  <code>inference_graph<\/code> test image result was awesome, but when randomly I fed it pics of bike, it also labled them as my provided custom classes. e.g. labeling car as cotton, a pic contain many many cars and trucks, it labeled all of them as either <code>cotton<\/code> or <code>other<\/code>.<\/p>\n\n<p>After this happen, my supervisor instructed instead of giving label file containing only 2 classes, pick original mscoco file which contain 90 labels and add two of your in it. So it'll 92 in total, I settled things like that, changed <code>config<\/code> file  accordingly and trained model for one more complete night again, but result didn't changes.<\/p>\n\n<p>I <a href=\"https:\/\/github.com\/EdjeElectronics\/TensorFlow-Object-Detection-API-Tutorial-Train-Multiple-Objects-Windows-10\/issues\/353\" rel=\"nofollow noreferrer\">issued this on his github<\/a> and someone answered me \"to add other object e.g. car, bike, preson, cards in my training dataset and don't label it. model will learn to ignore it.\", but in my github file, in an output image, a bike was parked behind cotton plant, and it automatically detected it as <code>other<\/code> (weed). <\/p>\n\n<p>I've looked for many articles and tutorials but couldn't find my desired results. <\/p>\n\n<p>All <code>codes<\/code> , <code>outputs<\/code>, <code>.congif<\/code> and <code>label.pbtxt<\/code> files has been uploaded to <a href=\"https:\/\/github.com\/Master-HM\/Retraining-Inception-labeling-my-custom-specified-labels-even-to-strange-objects\" rel=\"nofollow noreferrer\">Github<\/a>. Have a look at it please. <\/p>\n\n<p>All I want from my model is to detect <strong>all object<\/strong> as it use to detect random objects accurately before feeding custom dataset. Yes before feeding and training on my custom data, it classified all objects accurately, for example car to car, truck to truck, and person to person. It becomes dirty when I feed my custom input to it. And if I don't feed my custom input to it, it don't detect <code>cotton<\/code> or <code>other<\/code>. <strong>I want my model to detect my custom classes <code>cotton<\/code>, <code>other<\/code> + all other classes<\/strong> e.g. <code>car_bike_person_card<\/code> on which google has trained it. I've looked for many articles, tutorial and guide, way to perform custom object detection on pre-trained model is same. Even on inception guide it is same as I did.<\/p>\n","answers":[]}
+{"QuestionId":57868337,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-estimator>","CreationDate":"2019-09-10T09:58:06.943","AcceptedAnswerId":null,"OwnerUserId":12045744.0,"Title":"tensorflow:Error reported to Coordinator: Items of feature_columns must be a _FeatureColumn. Given (type <class 'collections.NumericColumn'>)","Body":"<p>I tried to use tf estimator to build logistic regression model. I used iris dataset and it ran successfully in my computer. However, when I tried to apply this model in cluster(using train_and_evaluate instead of classfier.train), I came across this problem.<\/p>\n\n<p>python version:3.6.8\ntensorflow version:1.13.1<\/p>\n\n<p>Here is the code running locally:<\/p>\n\n<p>iris dataset only contains numeric data. So feature_columns is a list of NumericColumn.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>FUTURES = ['SepalLength', 'SepalWidth','PetalLength', 'PetalWidth', 'Species']\nfeature_columns = []\nfor key in FUTURES:\n    feature_columns.append(tf.feature_column.numeric_column(key=key))\n<\/code><\/pre>\n\n<p>define estimator. pass feature_columns into params<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>classifier = tf.estimator.Estimator(\n        model_fn=my_model_fn,\n        model_dir=models_path,\n        params={\n            'feature_columns': feature_columns,\n            'n_classes': 3,\n        })\n<\/code><\/pre>\n\n<p>define model_fn.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def my_model_fn(features,labels,mode,params):\n    net = tf.feature_column.input_layer(features, params['feature_columns'])\n    logits = tf.layers.dense(net, params['n_classes'], activation=None)\n\n    predicted_classes = tf.argmax(logits, 1)\n    if mode == tf.estimator.ModeKeys.PREDICT:\n        predictions = {'logits': logits}\n        return tf.estimator.EstimatorSpec(mode, predictions=predictions)\n\n    loss = tf.losses.sparse_softmax_cross_entropy(labels=labels, logits=logits)\n\n    if mode == tf.estimator.ModeKeys.TRAIN:\n        optimizer = tf.train.AdagradOptimizer(learning_rate=0.1)\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\n    accuracy = tf.metrics.accuracy(labels=labels,\n                                   predictions=predicted_classes,\n                                   name='acc_op') \n    metrics = {'accuracy': accuracy} \n    tf.summary.scalar('accuracy', accuracy[1])\n    if mode == tf.estimator.ModeKeys.EVAL:\n        return tf.estimator.EstimatorSpec(mode, loss=loss, eval_metric_ops=metrics)\n<\/code><\/pre>\n\n<p>This code works well and generates some result.<\/p>\n\n<p><code>-------------------------------------------------------------<\/code><\/p>\n\n<p>Then I want to train it in cluster. my_model_fn is the same as previous one and self._feature_numeric_col is still a list of NumericColumn. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class LogisticReg():\n   def __init__(self):\n        self._feature_col = x.columns.tolist()\n        self._feature_numeric_col = []\n        for key in self._feature_col:\n            self._feature_numeric_col.append(tf.feature_column.numeric_column(key=key))\n        self.estimator = tf.estimator.Estimator(model_fn=self.my_model_fn,\n                                                model_dir=self.model_path,\n                                                config=self.config,\n                                                params={'feature_columns':self._feature_numeric_col})\n\n    def my_model_fn(self, features, labels, mode, params):\n\n        net = tf.feature_column.input_layer(features, params['feature_columns'])\n        logits = tf.layers.dense(net, self.n_class, activation=None)\n\n        predicted_classes = tf.argmax(logits, 1)  \n        if mode == tf.estimator.ModeKeys.PREDICT:\n            predictions = {'logits': logits}\n            return tf.estimator.EstimatorSpec(mode, predictions=predictions)\n\n        loss = tf.losses.sparse_softmax_cross_entropy(labels=labels, logits=logits)\n\n        if mode == tf.estimator.ModeKeys.TRAIN:\n            optimizer = tf.train.AdagradOptimizer(learning_rate=0.1) \n            train_op = optimizer.minimize(loss, global_step=tf.train.get_global_step())  \uff01\n            return tf.estimator.EstimatorSpec(mode, loss=loss, train_op=train_op)\n\n        accuracy = tf.metrics.accuracy(labels=labels,predictions=predicted_classes) \n\n        metrics = {'accuracy': accuracy}  \n        tf.summary.scalar('accuracy', accuracy[1])  \n        if mode == tf.estimator.ModeKeys.EVAL:\n            return tf.estimator.EstimatorSpec(mode, loss=loss, eval_metric_ops=metrics)\n<\/code><\/pre>\n\n<p>use train_and_evaluate function instead of train\/eval\/predict<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># input_fn\ndef input_fn(self, X, y, mode, batch_size):\n    y = y.astype(np.int32)\n    X = X.astype(np.float32)\n    dataset = tf.data.Dataset.from_tensor_slices((dict(X), y)) # x,y:pandas\n    if mode == 'train':\n        dataset = dataset.shuffle(500)\n        dataset = dataset.repeat()  \n    dataset = dataset.batch(batch_size)\n    return dataset\n\n# train_spec\ntrain_spec = tf.estimator.TrainSpec(input_fn=lambda: self.input_fn(x_train,y_train,'train',batch_size),\n                                    max_steps=n_epochs)\n# eval_spec\neval_spec = tf.estimator.EvalSpec(input_fn=lambda: self.input_fn(x_valid, y_valid, 'valid', batch_size),\n                                          start_delay_secs=30, throttle_secs=30, steps=None)\n\n\ntf.estimator.train_and_evaluate(self.estimator, train_spec, eval_spec)\n<\/code><\/pre>\n\n<p>I expect the cluster version can generate similar output as the local one. However, I get this error.<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/usr\/local\/bin\/python3\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/coordinator.py\", line 297, in stop_on_exception\n    yield\n  File \"\/usr\/local\/bin\/python3\/lib\/python3.6\/site-packages\/tensorflow\/python\/distribute\/mirrored_strategy.py\", line 852, in run\n    self.main_result = self.main_fn(*self.main_args, **self.main_kwargs)\n  File \"\/usr\/local\/bin\/python3\/lib\/python3.6\/site-packages\/tensorflow_estimator\/python\/estimator\/estimator.py\", line 1112, in _call_model_fn\n    model_fn_results = self._model_fn(features=features, **kwargs)\n  File \"\/mnt\/glusterfs\/model-center\/train\/classify.py\", line 51, in my_model_fn\n    net = tf.feature_column.input_layer(features, params['feature_columns'])\n  File \"\/usr\/local\/bin\/python3\/lib\/python3.6\/site-packages\/tensorflow\/python\/feature_column\/feature_column.py\", line 302, in input_layer\n    cols_to_output_tensors=cols_to_output_tensors)\n  File \"\/usr\/local\/bin\/python3\/lib\/python3.6\/site-packages\/tensorflow\/python\/feature_column\/feature_column.py\", line 181, in _internal_input_layer\n    feature_columns = _normalize_feature_columns(feature_columns)\n  File \"\/usr\/local\/bin\/python3\/lib\/python3.6\/site-packages\/tensorflow\/python\/feature_column\/feature_column.py\", line 2263, in _normalize_feature_columns\n    'Given (type {}): {}.'.format(type(column), column))\nValueError: Items of feature_columns must be a _FeatureColumn. Given (type &lt;class 'collections.NumericColumn'&gt;): NumericColumn(key='sepal_length', shape=(1,), default_value=None, dtype=tf.float32, normalizer_fn=None).\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57868363,"AnswerCount":1,"Tags":"<machine-learning><keras><pycharm><image-recognition><resnet>","CreationDate":"2019-09-10T09:59:26.017","AcceptedAnswerId":57870072.0,"OwnerUserId":4136843.0,"Title":"AttributeError occurs while running ResNet50 using Keras on Pycharm","Body":"<p>I was using PyCharm and imported ResNet50 for image recognition of a sample image. When I run the code , the following error occured.<\/p>\n\n<p>I was learning using an online code which needed to be completed by learners. I configured PyCharm and installed required packages that were recommended. During learning image recognition using ResNet50 , while running the code I ended up with following error. Should I custom install ResNet50 on pycharm for this to work? The instructor said the IDE will auto download ResNet50 during execution of code. Attaching python code below.<\/p>\n\n<pre><code>import numpy as np\nfrom keras.preprocessing import image\nfrom keras.applications import resnet50\n\nmodel = resnet50.ResNet50\n\nimg = image.load_img(\"bay.jpg\", target_size=(224, 224))\n\nx = image.img_to_array(img)\n\nx = np.expand_dims(x, axis=0)\n\nx = resnet50.preprocess_input(x)\n\npredictions = model.predict(x)\n\npredicted_classes = resnet50.decode_predictions(predictions, top=9)\n\nprint(\"This is an image of:\")\n\nfor imagenet_id, name, likelihood in predicted_classes[0]:\n    print(\" - {}: {:2f} likelihood\".format(name, likelihood))\n<\/code><\/pre>\n\n<p>This is the resultant error that I am getting during execution. <\/p>\n\n<pre><code>File \"\/home\/warlock\/Downloads\/Ex_Files_Building_Deep_Learning_Apps\/\nExercise Files\/05\/image_recognition.py\", line 21, in &lt;module&gt;\n    predictions = model.predict(x)\nAttributeError: 'function' object has no attribute 'predict'\n<\/code><\/pre>\n","answers":[{"AnswerId":"57870072","CreationDate":"2019-09-10T11:42:57.933","ParentId":null,"OwnerUserId":"9794742","Title":null,"Body":"<p>You have this error because <code>ResNet50<\/code> is a fonction so you need to implement it like a fonction :<\/p>\n\n<pre><code>model = resnet50.ResNet50()\n<\/code><\/pre>\n\n<p>In order to have a resnet50 model with all default parameters<\/p>\n"}]}
+{"QuestionId":57868804,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-slim>","CreationDate":"2019-09-10T10:25:19.793","AcceptedAnswerId":null,"OwnerUserId":12045864.0,"Title":"Error when training a network using slim model library: TypeError: int() argument must be a string, a bytes-like object or a number, not 'Tensor'","Body":"<p>I am trying to add a simple network to tensorflow\/models\/research\/slim\/net to train it with the training code, it ran into this problem when i was tring to send a feature map in resnet to it.<\/p>\n\n<p>The feature map sent into it is like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Tensor(\"resnet_v1_152\/block1\/unit_3\/bottleneck_v1\/Relu:0\", shape=(16, 28, 28, 256), dtype=float32)\n<\/code><\/pre>\n\n<p>That is exracted from the endpoints the network returns.<\/p>\n\n<p>I tried to use placeholder to replace batchqueue, and the same error still appear, i don't know where the problem occurs.<\/p>\n\n<p>The following is the code of the network architecture.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\ndef dis_conv(inputs,\n             scope='dis_conv',\n             is_training=True,\n             reuse=tf.AUTO_REUSE):\n    with tf.variable_scope(scope, 'dis_conv', [inputs], reuse=reuse) as sc:\n        end_points_collection = sc.original_name_scope + '_end_points'\n        with slim.arg_scope([slim.conv2d],\n                            activation_fn=tf.nn.relu,\n                            outputs_collections=end_points_collection,\n                            reuse=tf.AUTO_REUSE,\n                            ):\n            with (slim.arg_scope([slim.batch_norm], is_training=is_training)):\n                depth = tf.shape(inputs)[-1]\n                width = tf.shape(inputs)[-2]\n                net = slim.conv2d(inputs, depth * 2, 3, stride=2, padding='SAME', scope='conv_1')\n                net = slim.conv2d(net, depth * 4, 3, stride=2, padding='SAME', scope='conv_2')\n                net = slim.conv2d(net, depth * 8, 3, stride=1, padding='SAME', scope='conv_3')\n                net = slim.conv2d(net, 1, width \/ 4, stride=1, padding='SAME', scope='conv_4')\n                net = tf.squeeze(net)\n                end_points = slim.utils.convert_collection_to_dict(end_points_collection)\n                end_points[sc.name + '\/conv_4'] = net\n                return net, end_points\n\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>Traceback (most recent call last):\n  File \"nonslim_act_GAN_distilling_train_eval.py\", line 869, in &lt;module&gt;\n    tf.app.run()\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/python\/platform\/app.py\", line 125, in run\n    _sys.exit(main(argv))\n  File \"nonslim_act_GAN_distilling_train_eval.py\", line 731, in main\n    resT = [dis(EPT[0])[0], dis(EPT[1])[0],\n  File \"\/media\/disk\/Backup\/02zhaofeng\/Distalling_inceptionV3\/nets\/discriminator.py\", line 59, in dis_conv\n    net = slim.conv2d(inputs, depth * 2, 3, stride=2, padding='SAME', scope='conv_1')\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/arg_scope.py\", line 183, in func_with_args\n    return func(*args, **current_args)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/layers\/python\/layers\/layers.py\", line 1154, in convolution2d\n    conv_dims=2)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/arg_scope.py\", line 183, in func_with_args\n    return func(*args, **current_args)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/layers\/python\/layers\/layers.py\", line 1057, in convolution\n    outputs = layer.apply(inputs)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py\", line 805, in apply\n    return self.__call__(inputs, *args, **kwargs)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/python\/layers\/base.py\", line 362, in __call__\n    outputs = super(Layer, self).__call__(inputs, *args, **kwargs)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py\", line 728, in __call__\n    self.build(input_shapes)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/layers\/convolutional.py\", line 161, in build\n    dtype=self.dtype)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/python\/layers\/base.py\", line 276, in add_weight\n    getter=vs.get_variable)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py\", line 565, in add_weight\n    aggregation=aggregation)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/python\/training\/checkpointable\/base.py\", line 535, in _add_variable_with_custom_getter\n    **kwargs_for_getter)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 1467, in get_variable\n    aggregation=aggregation)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 1217, in get_variable\n    aggregation=aggregation)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 510, in get_variable\n    return custom_getter(**custom_getter_kwargs)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/layers\/python\/layers\/layers.py\", line 1744, in layer_variable_getter\n    return _model_variable_getter(getter, *args, **kwargs)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/layers\/python\/layers\/layers.py\", line 1735, in _model_variable_getter\n    use_resource=use_resource)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/arg_scope.py\", line 183, in func_with_args\n    return func(*args, **current_args)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/variables.py\", line 297, in model_variable\n    use_resource=use_resource)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/arg_scope.py\", line 183, in func_with_args\n    return func(*args, **current_args)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/variables.py\", line 252, in variable\n    use_resource=use_resource)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 481, in _true_getter\n    aggregation=aggregation)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 836, in _get_single_variable\n    shape = tensor_shape.as_shape(shape)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/tensor_shape.py\", line 946, in as_shape\n    return TensorShape(shape)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/tensor_shape.py\", line 541, in __init__\n    self._dims = [as_dimension(d) for d in dims_iter]\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/tensor_shape.py\", line 541, in &lt;listcomp&gt;\n    self._dims = [as_dimension(d) for d in dims_iter]\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/tensor_shape.py\", line 482, in as_dimension\n    return Dimension(value)\n  File \"\/media\/disk\/Backup\/02zhaofeng\/anaconda3\/envs\/envPyTF\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/tensor_shape.py\", line 37, in __init__\n    self._value = int(value)\nTypeError: int() argument must be a string, a bytes-like object or a number, not 'Tensor'\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57869029,"AnswerCount":0,"Tags":"<tensorflow><go><tensorflow-serving>","CreationDate":"2019-09-10T10:38:36.207","AcceptedAnswerId":null,"OwnerUserId":4921432.0,"Title":"Tensorflow serving SharedBatchScheduler for golang","Body":"<p>I want to optimize serving tensorflow model in Golang and study tensorflow serving.<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/serving\/tree\/master\/tensorflow_serving\/batching#servers-with-multiple-models-model-versions-or-subtasks\" rel=\"nofollow noreferrer\">servers-with-multiple-models-model-versions-or-subtasks<\/a><\/p>\n\n<blockquote>\n  <p>SharedBatchScheduler presents an abstraction of queues, accepts requests to schedule a particular kind of task. Each batch contains tasks of just one type, i.e. from one queue. The scheduler ensures fairness by interleaving the different types of batches.<\/p>\n<\/blockquote>\n\n<p>Is there any Golang repo working on SharedBatchScheduler? Or I have to rewrite the tensorflow serving <a href=\"https:\/\/github.com\/tensorflow\/serving\/issues\/96\" rel=\"nofollow noreferrer\">BatchingSession<\/a> C++ source code to Golang version?<\/p>\n","answers":[]}
+{"QuestionId":57869144,"AnswerCount":0,"Tags":"<tensorflow><cross-entropy>","CreationDate":"2019-09-10T10:45:25.117","AcceptedAnswerId":null,"OwnerUserId":12012749.0,"Title":"ERROR:tensorflow:Model diverged with loss = NaN","Body":"<p>all\n   I want trying to training step, but I always got the NaN loss during training issue. below is my main code, dataset is tf_record, does someone any suggestion to debug this issue, thanks.<\/p>\n\n<p>Learning rate= 0.005<\/p>\n\n<p>If I don't squeeze label_batch, then I need chage tf.nn.sparse_softmax_cross_entropy_with_logits to tf.nn.softmax_cross_entropy_with_logits, but Loss value is very high<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def make_dataset(filenames, take_count, batch_size, image_size,\n                 training=False, shard=False, num_threads=10, nsummary=10):\n    shuffle_buffer_size = 10000\n    num_readers = 1\n    if hvd.size() &gt; len(filenames):\n        assert (hvd.size() % len(filenames)) == 0\n        filenames = filenames * (hvd.size() \/ len(filenames))\n\n    ds = tf.data.Dataset.from_tensor_slices(filenames)\n    if shard:\n       # split the dataset into parts for each GPU\n       ds = ds.shard(hvd.size(), hvd.rank())\n\n    if not training:\n        ds = ds.take(take_count)  # make sure all ranks have the same amount\n\n    if training:\n        ds = ds.shuffle(1000, seed=7 * (1 + hvd.rank()))\n\n    ds = ds.interleave(\n        tf.data.TFRecordDataset, cycle_length=num_readers, block_length=1)\n    counter = tf.data.Dataset.range(sys.maxsize)\n    ds = tf.data.Dataset.zip((ds, counter))\n    preproc_func = lambda record, counter_: parse_and_preprocess_image_record(\n                   record, counter_, image_size)\n    #preproc_func = lambda record, counter_: parse_and_preprocess_image_record(\n    #    record, counter_, height, width, brightness, contrast, saturation, hue,\n    #    distort=training, nsummary=nsummary if training else 0, increased_aug=increased_aug)\n    ds = ds.map(preproc_func, num_parallel_calls=num_threads)\n    if training:\n        ds = ds.apply(tf.data.experimental.shuffle_and_repeat(shuffle_buffer_size, seed=5*(1+hvd.rank())))\n    ds = ds.batch(batch_size)\n    return ds\n\ndef main(args):\n\n    hvd.init()\n\n    network = importlib.import_module(args.model_def)\n    image_size = (args.image_size, args.image_size)\n\n    subdir = datetime.strftime(datetime.now(), '%Y%m%d-%H%M%S')\n    log_dir = os.path.join(os.path.expanduser(args.logs_base_dir), subdir)\n    model_dir = os.path.join(os.path.expanduser(args.models_base_dir), subdir)\n    if hvd.rank() == 0:\n        if not os.path.isdir(log_dir):  # Create the log directory if it doesn't exist\n            os.makedirs(log_dir)\n        if not os.path.isdir(model_dir):  # Create the model directory if it doesn't exist\n            os.makedirs(model_dir)\n\n        stat_file_name = os.path.join(log_dir, 'stat.h5')\n\n        # Write arguments to a text file\n        facenet.write_arguments_to_file(args, os.path.join(log_dir, 'arguments.txt'))\n\n    # Store some git revision info in a text file in the log directory\n    src_path,_ = os.path.split(os.path.realpath(__file__))\n    facenet.store_revision_info(src_path, log_dir, ' '.join(sys.argv))\n\n    np.random.seed(seed=args.seed)\n    random.seed(args.seed)\n    dataset = facenet.get_dataset(args.data_dir)\n    if args.filter_filename:\n        dataset = filter_dataset(dataset, os.path.expanduser(args.filter_filename), \n            args.filter_percentile, args.filter_min_nrof_images_per_class)\n\n    if args.validation_set_split_ratio&gt;0.0:\n        train_set, val_set = facenet.split_dataset(dataset, args.validation_set_split_ratio, args.min_nrof_val_images_per_class, 'SPLIT_IMAGES')\n    else:\n        train_set, val_set = dataset, [] \n    nrof_classes = len(train_set)\n\n    if args.data_dir_record:\n       filename_pattern = os.path.join(args.data_dir_record, '%s-*')\n       train_filenames = sorted(tf.gfile.Glob(filename_pattern % 'CASIA-WEBFACE'))\n       num_training_samples = facenet.get_num_records(train_filenames)\n\n    training_samples_per_rank = num_training_samples \/\/ hvd.size()\n\n    print('Model directory: %s' % model_dir)\n    #print('Log directory: %s' % log_dir)\n    pretrained_model = None\n    if args.pretrained_model:\n        pretrained_model = os.path.expanduser(args.pretrained_model)\n        print('Pre-trained model: %s' % pretrained_model)\n    \"\"\"\n    if hvd.rank() == 0: \n        if args.lfw_dir:\n            print('LFW directory: %s' % args.lfw_dir)\n            # Read the file containing the pairs used for testing\n            pairs = lfw.read_pairs(os.path.expanduser(args.lfw_pairs))\n            # Get the paths for the corresponding images\n            lfw_paths, actual_issame = lfw.get_paths(os.path.expanduser(args.lfw_dir), pairs)\n    \"\"\"\n    with tf.Graph().as_default():\n        tf.set_random_seed(args.seed)\n        #global_step = tf.Variable(0, trainable=False)\n        global_step = tf.train.get_or_create_global_step()\n\n        # Get a list of image paths and their labels\n        with tf.device('\/cpu:0'):\n            ds=facenet.make_dataset(\n                   train_filenames,\n                   training_samples_per_rank,\n                   args.batch_size, args.image_size, training=True, shard=True)\n\n            image_batch, label_batch = ds.make_one_shot_iterator().get_next()\n\n        label_batch = tf.squeeze(label_batch)\n\n        print('Number of classes in training set: %d' % nrof_classes)\n\n        print('Building training graph')\n\n        # Build the inference graph\n        prelogits, _ = network.inference(image_batch, args.keep_probability, \n            phase_train=True, bottleneck_layer_size=args.embedding_size, \n            weight_decay=args.weight_decay)\n        logits = slim.fully_connected(prelogits, len(train_set), activation_fn=None, \n                weights_initializer=slim.initializers.xavier_initializer(), \n                weights_regularizer=slim.l2_regularizer(args.weight_decay),\n                scope='Logits', reuse=False)\n\n        embeddings = tf.nn.l2_normalize(prelogits, 1, 1e-10, name='embeddings')\n\n        # Norm for the prelogits\n        eps = 1e-4\n        prelogits_norm = tf.reduce_mean(tf.norm(tf.abs(prelogits)+eps, ord=args.prelogits_norm_p, axis=1))\n        tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES, prelogits_norm * args.prelogits_norm_loss_factor)\n\n        # Add center loss\n        prelogits_center_loss, _ = facenet.center_loss(prelogits, label_batch, args.center_loss_alfa, nrof_classes)\n        tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES, prelogits_center_loss * args.center_loss_factor)\n\n        learning_rate = tf.train.exponential_decay(args.learning_rate, global_step,\n            args.learning_rate_decay_epochs*args.epoch_size, args.learning_rate_decay_factor, staircase=True)\n        tf.summary.scalar('learning_rate', learning_rate)\n\n        # Calculate the average cross entropy loss across the batch\n        cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(\n            labels=label_batch, 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        correct_prediction = tf.cast(tf.equal(tf.argmax(logits, 1), tf.cast(label_batch, tf.int64)), tf.float32)\n        accuracy = tf.reduce_mean(correct_prediction)\n\n        # Calculate the total losses\n        regularization_losses = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)\n        total_loss = tf.add_n([cross_entropy_mean] + regularization_losses, name='total_loss')\n\n        # Build a Graph that trains the model with one batch of examples and updates the model parameters\n        train_op = facenet.train(total_loss, global_step, args.optimizer, \n            learning_rate, args.moving_average_decay, tf.global_variables(), args.log_histograms)\n\n        # Create a saver\n        saver = tf.train.Saver(tf.trainable_variables(), max_to_keep=3)\n\n        # Build the summary operation based on the TF collection of Summaries.\n        summary_op = tf.summary.merge_all()\n\n        config = tf.ConfigProto()\n        config.gpu_options.allow_growth = True\n        config.gpu_options.visible_device_list = str(hvd.local_rank())\n\n        class _LoggerHook(tf.train.SessionRunHook):\n            \"\"\"Logs loss and runtime.\"\"\"\n\n            def begin(self):\n                self._step = -1\n                self._start_time = time.time()\n\n            def before_run(self, run_context):\n                self._step += 1\n                return tf.train.SessionRunArgs(total_loss)  # Asks for loss value.\n\n            def after_run(self, run_context, run_values):\n                if self._step % args.log_frequency == 0:\n                    current_time = time.time()\n                    duration = current_time - self._start_time\n                    self._start_time = current_time\n\n                    loss_value = run_values.results\n                    examples_per_sec = args.log_frequency * args.batch_size \/ duration\n                    sec_per_batch = float(duration \/ args.log_frequency)\n\n                    format_str = ('%s: step %d, loss = %.2f (%.1f examples\/sec; %.3f '\n                                    'sec\/batch)')\n                    print (format_str % (datetime.now(), self._step, loss_value,\n                                        examples_per_sec, sec_per_batch))\n\n        with tf.train.MonitoredTrainingSession(\n            checkpoint_dir=os.path.join(model_dir, 'model-%s.ckpt' % subdir),\n            hooks=[hvd.BroadcastGlobalVariablesHook(0), \n                   tf.train.StopAtStepHook(last_step=args.epoch_size * args.max_nrof_epochs \/\/ hvd.size()),\n                   tf.train.NanTensorHook(total_loss),\n                   _LoggerHook()],\n            config=config,\n            save_checkpoint_steps=1000) as mon_sess:\n            while not mon_sess.should_stop():\n                mon_sess.run(train_op)\n\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>2019-09-10 10:36:19.204623: step 0, loss = 16.70 (2.1 examples\/sec; 43.300 sec\/batch)\nERROR:tensorflow:Model diverged with loss = NaN.\nTraceback (most recent call last):\n  File \"src\/train_softmax-horovod-v1-temp.py\", line 357, in &lt;module&gt;\n    main(parse_arguments(sys.argv[1:]))\n  File \"src\/train_softmax-horovod-v1-temp.py\", line 219, in main\n    mon_sess.run(train_op)\n  File \"\/usr\/lib\/python3.5\/contextlib.py\", line 77, in __exit__\n    self.gen.throw(type, value, traceback)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 5253, in get_controller\n    yield g\n  File \"\/usr\/lib\/python3.5\/contextlib.py\", line 77, in __exit__\n    self.gen.throw(type, value, traceback)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 5061, in get_controller\n    yield default\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 5253, in get_controller\n    yield g\n  File \"\/usr\/lib\/python3.5\/contextlib.py\", line 77, in __exit__\n    self.gen.throw(type, value, traceback)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/eager\/context.py\", line 415, in _mode\n    yield\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 5253, in get_controller\n    yield g\n  File \"src\/train_softmax-horovod-v1-temp.py\", line 219, in main\n    mon_sess.run(train_op)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/training\/monitored_session.py\", line 676, in run\n    run_metadata=run_metadata)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/training\/monitored_session.py\", line 1171, in run\n    run_metadata=run_metadata)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/training\/monitored_session.py\", line 1270, in run\n    raise six.reraise(*original_exc_info)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/six.py\", line 693, in reraise\n    raise value\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/training\/monitored_session.py\", line 1255, in run\n    return self._sess.run(*args, **kwargs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/training\/monitored_session.py\", line 1335, in run\n    run_metadata=run_metadata))\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/training\/basic_session_run_hooks.py\", line 753, in after_run\n    raise NanLossDuringTrainingError\ntensorflow.python.training.basic_session_run_hooks.NanLossDuringTrainingError: NaN loss during training.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57869149,"AnswerCount":1,"Tags":"<tensorflow><quantization><uint8t>","CreationDate":"2019-09-10T10:45:35.643","AcceptedAnswerId":58080063.0,"OwnerUserId":1284889.0,"Title":"post training quantization for mobilenet V1 not working","Body":"<p>I am trying to convert mobilenet V1 .pb file to quantized tflite file. I used the below command to do the quantization:<\/p>\n\n<pre><code>  tflite_convert \\\n  --output_file=\/home\/wc\/users\/Mostafiz\/TPU\/models\/mobilnet\/test2_4thSep\/mobilenetv1_test5.tflite \\\n  --graph_def_file=\/home\/wc\/users\/Mostafiz\/TPU\/models\/mobilnet\/mobileNet_frozen_graph.pb \\\n  --output_format=TFLITE \\\n  --inference_type=QUANTIZED_UINT8 \\\n  --inference_input_type=QUANTIZED_UINT8 \\\n  --input_shape=1,224,224,3 \\\n  --input_array=input \\\n  --output_array=MobilenetV1\/Predictions\/Reshape_1 \\\n  --inference_output_type=QUANTIZED_UINT8 \\\n  --default_ranges_min=0 \\\n  --default_ranges_max=6 \\\n  --std_dev_values=127 \\\n  --mean_value=128\n<\/code><\/pre>\n\n<p>The .tflile file is created without any error. But when I am trying to use the .tflile for inference the output classes are messed up. None of the test images are giving correct result. <\/p>\n\n<p>Not sure where I am doing wrong, can someone please help me?<\/p>\n\n<p>For inference I am using 'label_image.py' provided by tensorflow. here is the code:<\/p>\n\n<pre><code>\"\"\"label_image for tflite\"\"\"\n\nfrom __future__ import absolute_import\nfrom __future__ import division\nfrom __future__ import print_function\n\nimport argparse\nimport numpy as np\n\nfrom PIL import Image\n\nfrom tensorflow.lite.python import interpreter as interpreter_wrapper\n\ndef load_labels(filename):\n  my_labels = []\n  input_file = open(filename, 'r')\n  for l in input_file:\n    my_labels.append(l.strip())\n  return my_labels\n\nif __name__ == \"__main__\":\n  floating_model = False\n\n  parser = argparse.ArgumentParser()\n  parser.add_argument(\"-i\", \"--image\", default=\"\/tmp\/grace_hopper.bmp\", \\\n    help=\"image to be classified\")\n  parser.add_argument(\"-m\", \"--model_file\", \\\n    default=\"\/tmp\/mobilenet_v1_1.0_224_quant.tflite\", \\\n    help=\".tflite model to be executed\")\n  parser.add_argument(\"-l\", \"--label_file\", default=\"\/tmp\/labels.txt\", \\\n    help=\"name of file containing labels\")\n  parser.add_argument(\"--input_mean\", default=127.5, help=\"input_mean\")\n  parser.add_argument(\"--input_std\", default=127.5, \\\n    help=\"input standard deviation\")\n  args = parser.parse_args()\n\n  interpreter = interpreter_wrapper.Interpreter(model_path=args.model_file)\n  interpreter.allocate_tensors()\n\n  input_details = interpreter.get_input_details()\n  output_details = interpreter.get_output_details()\n\n  # check the type of the input tensor\n  if input_details[0]['dtype'] == np.float32:\n    floating_model = True\n\n  # NxHxWxC, H:1, W:2\n  height = input_details[0]['shape'][1]\n  width = input_details[0]['shape'][2]\n  img = Image.open(args.image)\n  img = img.resize((width, height))\n\n  # add N dim\n  input_data = np.expand_dims(img, axis=0)\n\n  if floating_model:\n    input_data = (np.float32(input_data) - args.input_mean) \/ args.input_std\n\n  interpreter.set_tensor(input_details[0]['index'], input_data)\n\n  interpreter.invoke()\n\n  output_data = interpreter.get_tensor(output_details[0]['index'])\n  results = np.squeeze(output_data)\n\n  top_k = results.argsort()[-5:][::-1]\n  labels = load_labels(args.label_file)\n  for i in top_k:\n    if floating_model:\n      print('{0:08.6f}'.format(float(results[i]))+\":\", labels[i])\n    else:\n      print('{0:08.6f}'.format(float(results[i]\/255.0))+\":\", labels[i])\n<\/code><\/pre>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"58080063","CreationDate":"2019-09-24T12:11:50.653","ParentId":null,"OwnerUserId":"1284889","Title":null,"Body":"<p>The dummy quantization may not work properly as we need to guess the default_max and defual_min values for activation functions. <\/p>\n\n<p>As Sudarsh mentioned in the comment, we should do a post training full integer quantization to convert .pb to INT8 tflite file.<\/p>\n\n<p>you can follow this link to start with - <a href=\"https:\/\/www.tensorflow.org\/lite\/performance\/post_training_integer_quant\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>Hope that helped.<\/p>\n\n<p>Regards.<\/p>\n"}]}
+{"QuestionId":57869293,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2019-09-10T10:54:52.820","AcceptedAnswerId":57870610.0,"OwnerUserId":12046003.0,"Title":"Predicting New ImageNet Pictures","Body":"<p>I'm having problems predicting new images based on the ImageNet database where objects are placed on different backgrounds. Predicting the classes work well on various models (ResNets and VGG) but it doesn't work at all for Densenets and Inception. I'm thinking this is related to the labels but I can't figure out what's wrong exactly. <\/p>\n\n<p>The labels are ordered from 0 to 999 (divided in maps) and are in the same order as the original dataset. <\/p>\n\n<p>I've changed to labels to the other format (e.g. n02676566). This didn't work for DenseNet (ResNets still worked fine). <\/p>\n\n<p>Code: <\/p>\n\n<pre><code>model = keras.applications.densenet.DenseNet201(weights='imagenet', classes=1000)\n\nlabels = [str(i) for i in range(0,1000)]\n\ngenerator_imagenet = ImageDataGenerator().flow_from_directory(path, target_size(224,224), batch_size=64, class_mode='categorical', shuffle=True, classes=labels)\n\nresnet18_model.compile(optimizer='SGD', loss='categorical_crossentropy', metrics=['acc'])\n\nresnet18_model.evaluate_generator(generator_imagenet)\n<\/code><\/pre>\n\n<p>Output ResNet 152: <\/p>\n\n<blockquote>\n  <p>Found 243 images belonging to 1000 classes. [1.6771895689238245,\n  0.5679012397189199]<\/p>\n  \n  <p>Output DenseNet 201:<\/p>\n  \n  <p>Found 243 images belonging to 1000 classes. [15.938421453468102, 0.0]<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"57870610","CreationDate":"2019-09-10T12:15:46.190","ParentId":null,"OwnerUserId":"9579075","Title":null,"Body":"<p>You will need to preprocess the images using the <code>preprocess_input<\/code> function provided with each <code>keras_applications<\/code> model.\nAdd the following changes.<\/p>\n\n<pre><code>keras.applications.densenet import preprocess_input\ngenerator_imagenet = ImageDataGenerator(preprocessing_function=preprocess_input).flow_from_directory(path, \n                                                          target_size(224,224),\n                                                          batch_size=64,\n                                                          class_mode='categorical', \n                                                          shuffle=True, \n                                                          classes=labels)\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":57869538,"AnswerCount":1,"Tags":"<keras><keras-layer>","CreationDate":"2019-09-10T11:08:13.517","AcceptedAnswerId":57870741.0,"OwnerUserId":8338613.0,"Title":"Convert a tensor axis into strings and concatenate the axis reshaping the tensor?","Body":"<p>If my last Keras Layer X outputs a tensor which looks, e.g. like this:<\/p>\n\n<pre><code>{\n  {4, [22, 16,  11, 33], 24},\n  {4, [12, 33,  87, 17], 14},\n  {3, [92, 407, 25, 27], 34}\n}\n<\/code><\/pre>\n\n<p>How can I add one more Layer Y that converts the output of X like this:<\/p>\n\n<pre><code>{\n  {4, \"22|16|11|33\",  24},\n  {4, \"12|33|87|17\",  14},\n  {3, \"92|407|25|27\", 34}\n}\n<\/code><\/pre>\n\n<p>In other words, I need to turn the second dimension into a string, with all values concatenated.<\/p>\n","answers":[{"AnswerId":"57870741","CreationDate":"2019-09-10T12:24:17.103","ParentId":null,"OwnerUserId":"11522398","Title":null,"Body":"<p>I'm not sure if Keras has method for this specific problem. A possible solution (if you can use list), is the following implementation:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Imagine you have a output layer like this:\na = [\n  [4, [22, 16,  11, 33], 24],\n  [4, [12, 33,  87, 17], 14],\n  [3, [92, 407, 25, 27], 34]\n]\n\n# \"transpose\" the list and get the second element\ncolumn1 = list(zip(*a))[1]\n\n# Join all elements with \"|\"\nnewColumn1 = ['|'.join(str(e) for e in row) for row in column1]\n\n# Put back the modified column\nb =  list(zip(*a))\nb[1] = newColumn1\n\n# \"Untranspose\" your list\na2 =  list(zip(*b))\na2\n<\/code><\/pre>\n\n<p>Your output will be:<\/p>\n\n<pre><code>[(4, '22|16|11|33', 24), (4, '12|33|87|17', 14), (3, '92|407|25|27', 34)]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57869557,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><keras><matrix-multiplication>","CreationDate":"2019-09-10T11:09:54.957","AcceptedAnswerId":57869965.0,"OwnerUserId":12042308.0,"Title":"How to multiply a 4-D Tensor by a 2-D matrix?","Body":"<p>I have a  4-D Tensor [batch_size, num_rows, num_cols, num_values] with shape [N,60,60,1], and I want to weight the WH Dimension with a [60,60] matrix.\nHow can I do that?<\/p>\n","answers":[{"AnswerId":"57869965","CreationDate":"2019-09-10T11:34:48.140","ParentId":null,"OwnerUserId":"4960855","Title":null,"Body":"<blockquote>\n  <p>I have a  4-D Tensor [batch_size, num_rows, num_cols, num_values] with\n  shape [N,60,60,1], and I want to weight the WH Dimension with a\n  [60,60] matrix. How can I do that?<\/p>\n<\/blockquote>\n\n<pre><code>tensor * matrix.reshape( (1, 60, 60, 1) )\n<\/code><\/pre>\n\n<p>This will <a href=\"https:\/\/www.tensorflow.org\/xla\/broadcasting\" rel=\"nofollow noreferrer\">broadcast<\/a> the matrix along axis 0 of the tensor.<\/p>\n"}]}
+{"QuestionId":57870582,"AnswerCount":1,"Tags":"<python><caffe><object-detection><face-detection><lmdb>","CreationDate":"2019-09-10T12:13:56.637","AcceptedAnswerId":null,"OwnerUserId":8409376.0,"Title":"Generating lmdb files for object detection in Caffe","Body":"<p>I am trying to train an SSD-based face detector from scratch in Caffe. I have a dataset of images where the bounding boxes for faces are stored in a csv file. In all the tutorials and code I've come across so far, convert_annoset tool is used to generate an lmdb file for object detection. However, in the latest Caffe version on Github (<a href=\"https:\/\/github.com\/BVLC\/caffe\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe<\/a>), this tool has been removed. <\/p>\n\n<p>The two options I see to deal with this issue are:<\/p>\n\n<ol>\n<li><p>Rewrite the convert_annoset tool using functions in the current Caffe library<\/p><\/li>\n<li><p>Use other python packages (such as lmdb and OpenCV) to manually create lmdb files from the images and bounding box information<\/p><\/li>\n<\/ol>\n\n<p>I've been trying to rewrite the code but I am unable to find certain classes and functions that were used in the original code such as <em>AnnotatedDatum_AnnotationType<\/em> and <em>LabelMap<\/em> in the current version.<\/p>\n\n<p>Any suggestions for how to proceed with creating lmdb for the object detection problem?<\/p>\n\n<p><strong>EDIT<\/strong>:\nI've just realized that <em>AnnotatedData<\/em> layer no longer exists in the master branch of Caffe. Does this mean that detection is not possible in this version? Do I have to use some older fork such as <a href=\"https:\/\/github.com\/weiliu89\/caffe\" rel=\"nofollow noreferrer\">https:\/\/github.com\/weiliu89\/caffe<\/a> for detection or is there any other option?<\/p>\n","answers":[{"AnswerId":"57949047","CreationDate":"2019-09-16T00:01:42.423","ParentId":null,"OwnerUserId":"11976900","Title":null,"Body":"<p>Since the original Caffe does not have a layer for object detection(mobilenet-ssd)<\/p>\n\n<p>so use this Caffe (<a href=\"https:\/\/github.com\/weiliu89\/caffe\" rel=\"nofollow noreferrer\">https:\/\/github.com\/weiliu89\/caffe<\/a>) ssd branch<\/p>\n\n<p>This Caffe has a convert_annoset that can generate an lmdb.<\/p>\n"}]}
+{"QuestionId":57871128,"AnswerCount":1,"Tags":"<tensorflow><keras><tensor>","CreationDate":"2019-09-10T12:46:21.687","AcceptedAnswerId":null,"OwnerUserId":2207729.0,"Title":"Strange tensors","Body":"<p>What does it mean when I print a tensor and it shows some sort of operation name instead of its value?<\/p>\n\n<p>Eg.:<\/p>\n\n<pre><code>Tensor(\"gradients\/block1_conv1\/convolution_grad\/Conv2DBackpropInput:0\", shape=(?, ?, ?, 3), dtype=float32)\n<\/code><\/pre>\n\n<pre><code>Tensor(\"truediv_1:0\", shape=(?, ?, ?, 3), dtype=float32)\n<\/code><\/pre>\n\n<p>The code that originates these prints is:<\/p>\n\n<pre><code>from keras.applications import VGG16\nfrom keras import backend as K\n\nmodel = VGG16(weights='imagenet',\n              include_top=False)\n\nlayer_name = 'block3_conv1'\nfilter_index = 0\n\nlayer_output = model.get_layer(layer_name).output\nloss = K.mean(layer_output[:, :, :, filter_index])\n\n# The call to `gradients` returns a list of tensors (of size 1 in this case)\n# hence we only keep the first element -- which is a tensor.\ngrads = K.gradients(loss, model.input)[0]\n\n\nprint(\"Print 1:\")\nprint(grads)\nprint(type(grads))\n\n# We add 1e-5 before dividing so as to avoid accidentally dividing by 0.\ngrads \/= (K.sqrt(K.mean(K.square(grads))) + 1e-5)\ngrads \/= (K.sqrt(K.mean(K.square(grads))) + 1e-5)\nprint(\"Print 2:\")\nprint(grads)\nprint(type(grads))\n<\/code><\/pre>\n\n<p>According to the documentation a Tensor has no values, but means for reaching a final value in a given CPU or GPU session. At least as I understood it.<\/p>\n\n<p>Is this what it really means?<\/p>\n\n<p>How can I list all operations within a tensor, like in a sequential way, that take me from its input to the final value?<\/p>\n\n<p>Eg. The <code>grads<\/code> tensor would be a gradient function, and two divisions, to calculate the final value.<\/p>\n","answers":[{"AnswerId":"57871249","CreationDate":"2019-09-10T12:52:31.027","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Keras\/Tensorflow (except when in eager mode, but I don't think Keras supports eager mode) is a \"symbolic graph\" language. <\/p>\n\n<p>When you create a model and its operations, you are just creating a \"graph\", not the operations themselves. Then you only have the tensors.<\/p>\n\n<p>The Keras models have \"input tensors\", which in Tensorflow are \"placeholders\". They are special tensors that \"will\" receive values when training or predicting. (In Keras: <code>model.fit<\/code>, <code>model.predict<\/code>, <code>model.evaluate<\/code>, etc. In Tensorflow, <code>session.run<\/code>)     <\/p>\n\n<p>Here an example of how to feed the inputs with data and get the actual values of the gradients: <a href=\"https:\/\/stackoverflow.com\/questions\/39561560\/getting-gradient-of-model-output-w-r-t-weights-using-keras\">Getting gradient of model output w.r.t weights using Keras<\/a> <\/p>\n"}]}
+{"QuestionId":57871287,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-09-10T12:55:00.923","AcceptedAnswerId":null,"OwnerUserId":11984833.0,"Title":"PyTorch equivalent of `numpy.unpackbits`?","Body":"<p>I am training a neural net on GPU. It uses a lot of binary input features.<\/p>\n\n<p>Since moving data to\/from GPU is expensive, I am looking for ways to make the initial representation more compact. Now, I encode my features as <strong>int8<\/strong>, move them over to GPU and <em>then<\/em> expand as <strong>float32<\/strong>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># create int8\nfeatures = torch.zeros(*dims, dtype=torch.int8)\n\n# fill in some data (set some features to 1.)\n\u2026\n\n# move int8 to GPU\nfeatures = features.to(device=cuda, non_blocking=True)\n\n# expand int8 as float32\nfeatures = features.to(dtype=float32)\n<\/code><\/pre>\n\n<p>Now, I am looking for ways to compress those binary features to bits instead of bytes.<\/p>\n\n<p>NumPy has functions <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.packbits.html\" rel=\"nofollow noreferrer\"><code>packbits<\/code><\/a> and <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.unpackbits.html\" rel=\"nofollow noreferrer\"><code>unpackbits<\/code><\/a><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>&gt;&gt;&gt; a = np.array([[2], [7], [23]], dtype=np.uint8)\n&gt;&gt;&gt; b = np.unpackbits(a, axis=1)\n&gt;&gt;&gt; b\narray([[0, 0, 0, 0, 0, 0, 1, 0],\n       [0, 0, 0, 0, 0, 1, 1, 1],\n       [0, 0, 0, 1, 0, 1, 1, 1]], dtype=uint8)\n<\/code><\/pre>\n\n<p>Is there any way to unpack bits in PyTorch on GPU?<\/p>\n","answers":[{"AnswerId":"57939875","CreationDate":"2019-09-14T22:54:32.687","ParentId":null,"OwnerUserId":"7347631","Title":null,"Body":"<p>There is no similar functions at the time of writing this answer. However, a workaround is using <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.from_numpy\" rel=\"nofollow noreferrer\"><code>torch.from_numpy<\/code><\/a> as in: <\/p>\n\n<pre><code>In[2]: import numpy as np\nIn[3]: a = np.array([[2], [7], [23]], dtype=np.uint8)\nIn[4]: b = np.unpackbits(a, axis=1)\nIn[5]: b\nOut[5]: \narray([[0, 0, 0, 0, 0, 0, 1, 0],\n       [0, 0, 0, 0, 0, 1, 1, 1],\n       [0, 0, 0, 1, 0, 1, 1, 1]], dtype=uint8)\nIn[6]: import torch\nIn[7]: torch.from_numpy(b)\nOut[7]: \ntensor([[0, 0, 0, 0, 0, 0, 1, 0],\n        [0, 0, 0, 0, 0, 1, 1, 1],\n        [0, 0, 0, 1, 0, 1, 1, 1]], dtype=torch.uint8)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57871463,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-10T13:04:51.757","AcceptedAnswerId":57887629.0,"OwnerUserId":2148718.0,"Title":"Keras Sequential without providing input shape","Body":"<p>I currently have a keras model that looks like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = keras.Sequential()\nmodel.add(keras.layers.Dense(100, activation=tf.nn.relu))\nmodel.add(keras.layers.Dense(100, activation=tf.nn.relu))\nmodel.add(keras.layers.Dense(len(labels), activation=tf.nn.softmax))\n<\/code><\/pre>\n\n<p>The <a href=\"https:\/\/keras.io\/getting-started\/sequential-model-guide\/#specifying-the-input-shape\" rel=\"nofollow noreferrer\">Keras documentation<\/a> tells me: <\/p>\n\n<blockquote>\n  <p>The model needs to know what input shape it should expect. For this reason, the first layer in a Sequential model (and only the first, because following layers can do automatic shape inference) needs to receive information about its input shape<\/p>\n<\/blockquote>\n\n<p><strong>However<\/strong>, the model as it is actually trains fine, without errors, even though I never specified the shape of the inputs. <\/p>\n\n<p>How does it know what shape to expect? What is the default behaviour if I don't provide an input shape? How will it affect my model?<\/p>\n\n<p><em>edit: this is using <code>tf.keras<\/code>, aka the Tensorflow backend for keras<\/em><\/p>\n","answers":[{"AnswerId":"57887732","CreationDate":"2019-09-11T11:04:33.800","ParentId":null,"OwnerUserId":"6117017","Title":null,"Body":"<p>There appears to be a miscorrelation between the TensorFlow version and the Keras documentation.<\/p>\n\n<p>As far as I can see, the documentation that you are referring to is the one from Keras, while the TensorFlow version that you use is 2.0.0 >= alpha version, whose Keras internal version is different.<\/p>\n\n<p>You should check the Keras documentation within TensorFlow 2.0, otherwise you might very well run into mismatches.<\/p>\n"},{"AnswerId":"57887629","CreationDate":"2019-09-11T10:58:23.980","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>Nice observation - I believe the Keras documentation should be updated. When the input shape is not provided, Keras infers it from the argument <code>x<\/code> of <code>Model.fit<\/code> and only then it builds the whole model. Concretely, this is what's happening:<\/p>\n\n<ol>\n<li>When adding Keras layers in the <code>Sequential<\/code> model, since the argument <code>input_shape<\/code> (and, by extension, <code>batch_input_shape<\/code>) is never set, the attribute <code>Model.inputs<\/code> remains <code>None<\/code> (see <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/engine\/sequential.py#L135-L180\" rel=\"nofollow noreferrer\"><code>Sequential.add<\/code><\/a>).<\/li>\n<li>Then, in <code>Model.fit<\/code>, they check whether <code>Model.inputs<\/code> has been set (see <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/engine\/training.py#L1150\" rel=\"nofollow noreferrer\"><code>Model.fit<\/code><\/a> and <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/88ca804d94a0f978d0522d82a2e2bda26362076a\/keras\/engine\/training.py#L500-L504\" rel=\"nofollow noreferrer\"><code>Model._standardize_user_data<\/code><\/a>) and, when it hasn't, they infer the input shape from the provided input array.<\/li>\n<li>Finally, in <code>Model._set_inputs<\/code>, they build the whole model with the inferred <code>input_shape<\/code> (see <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/88ca804d94a0f978d0522d82a2e2bda26362076a\/keras\/engine\/training.py#L409-L417\" rel=\"nofollow noreferrer\"><code>Model._set_inputs<\/code><\/a>).<\/li>\n<\/ol>\n\n<hr>\n\n<p>This can be verified by printing some weights (e.g. <code>print(model.layers[0].get_weights())<\/code>) before fitting the model. You will see that, when the argument <code>input_shape<\/code> or <code>batch_input_shape<\/code> is not provided to the first layer of the model, the weight's array is empty as the model is yet to be built.<\/p>\n"}]}
+{"QuestionId":57871601,"AnswerCount":1,"Tags":"<python><tensorflow><google-cloud-automl>","CreationDate":"2019-09-10T13:13:19.013","AcceptedAnswerId":null,"OwnerUserId":6712543.0,"Title":"How does AutoML Vision handle images with dimensions less than 224x224 pixels?","Body":"<p>I have trained AutoML Vision with 80x80x3 image samples. The training have finished successfully and I downloaded the edge tflite model. When implementing the tflite model in python, according this <a href=\"https:\/\/www.tensorflow.org\/lite\/guide\/inference#load_and_run_a_model_in_python\" rel=\"nofollow noreferrer\">tutorial by tensorflow<\/a> I realised that the tflite models input size is 224x224x3.<\/p>\n\n<p>My question is:<\/p>\n\n<ul>\n<li>How does AutoML Vision handle images with dimensions less than 224x224 during training? Does AutoML Vision resize the images to 224x224 with interpolation?<\/li>\n<\/ul>\n\n<p>For better prediction performance I would like to process the new images exactly the way AutoML Vision has processed the images during training.<\/p>\n\n<p>When feeding 80x80 images with input shape (1, 80, 80, 3) to the model, I get the exception \"Cannot set tensor: Dimension mismatch\", see code below.<\/p>\n\n<p>Feeding 224x224 images works without exceptions. However, I would like to use images with 80x80x3, like I used for training. Or preprocess the 80x80x3 images like they were during training in AutoML Vision, for example by resizing them to 224x224x3 or however AutoML Vision has handled it. <\/p>\n\n<pre><code>test_sample.shape\n\nOut: (80, 80, 3)\n\ntest_sample = test_sample.reshape(1, 80, 80, 3)\n\nOut: (1, 80, 80, 3)    \n\n# Load TFLite model and allocate tensors. \ninterpreter = tf.lite.Interpreter(model_path=model_path) \ninterpreter.allocate_tensors() \n\n# Get input and output tensors. \ninput_details = interpreter.get_input_details()\n\nprint(interpreter.get_input_details())\n\nOut: [{'name': 'image', 'index': 0, 'shape': array([  1, 224, 224,   3], dtype=int32), 'dtype': &lt;class 'numpy.uint8'&gt;, 'quantization': (0.007874015718698502, 128)}]\n\noutput_details = interpreter.get_output_details() \n\n# Test model on input data. \ninput_data = np.array(test_sample, dtype=np.uint8) \ninterpreter.set_tensor(input_details[0]['index'], input_data) \n\ninterpreter.invoke() \n\n\nOut: ValueError: Cannot set tensor: Dimension mismatch \nValueError                                Traceback (most recent call last)\nin engine\n----&gt; 1 interpreter.set_tensor(input_details[0]['index'], input_data)\n\n\/home\/cdsw\/.local\/lib\/python3.6\/site-packages\/tensorflow\/lite\/python\/interpreter.py in set_tensor(self, tensor_index, value)\n   173       ValueError: If the interpreter could not set the tensor.\n   174     \"\"\"\n--&gt; 175     self._interpreter.SetTensor(tensor_index, value)\n   176 \n   177   def resize_tensor_input(self, input_index, tensor_size):\n\n\/home\/cdsw\/.local\/lib\/python3.6\/site-packages\/tensorflow\/lite\/python\/interpreter_wrapper\/tensorflow_wrap_interpreter_wrapper.py in SetTensor(self, i, value)\n   134 \n   135     def SetTensor(self, i, value):\n--&gt; 136         return _tensorflow_wrap_interpreter_wrapper.InterpreterWrapper_SetTensor(self, i, value)\n   137 \n   138     def GetTensor(self, i):\n\nValueError: Cannot set tensor: Dimension mismatch\n<\/code><\/pre>\n","answers":[{"AnswerId":"58193308","CreationDate":"2019-10-01T22:58:13.020","ParentId":null,"OwnerUserId":"6194280","Title":null,"Body":"<p>I recommend you to use this implementation<\/p>\n\n<p>Set dtype<\/p>\n\n<pre><code># check the type of the input tensor\nfloating_model = input_details[0]['dtype'] == np.float32\n<\/code><\/pre>\n\n<p>Resize your image according with the model height and width required<\/p>\n\n<pre><code>height = input_details[0]['shape'][1]\nwidth = input_details[0]['shape'][2]\nimg = Image.open(args.image).resize((width, height))\n<\/code><\/pre>\n\n<p>Set input data<\/p>\n\n<pre><code>input_data = np.expand_dims(img, axis=0)\n<\/code><\/pre>\n\n<p>Here is the full example\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/lite\/examples\/python\/label_image.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/lite\/examples\/python\/label_image.py<\/a><\/p>\n"}]}
+{"QuestionId":57871657,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras><mask><faster-rcnn>","CreationDate":"2019-09-10T13:15:40.793","AcceptedAnswerId":null,"OwnerUserId":10393794.0,"Title":"Mask-RCNN, RasterioIOError","Body":"<p>I am working on MaskRCNN, when I try to train my dataset, I got error like this:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nCPLE_OpenFailed                           Traceback (most recent call last)\nrasterio\/_base.pyx in rasterio._base.DatasetReader.start (rasterio\/_base.c:2847)()\n\nrasterio\/_base.pyx in rasterio._base.DatasetReader.start (rasterio\/_base.c:2799)()\n\nrasterio\/_err.pyx in rasterio._err.CPLErrors.check (rasterio\/_err.c:1773)()\n\nCPLE_OpenFailed: D:\/DATASET\/SpaceNet\/Train\/AOI_2_Vegas_Train\/RGB-PanSharpen\/RGB-PanSharpen_AOI_2_Vegas_img1189.tif: No such file or directory\n\nDuring handling of the above exception, another exception occurred:\n\nRasterioIOError                           Traceback (most recent call last)\n&lt;ipython-input-11-83fb3ae74319&gt; in &lt;module&gt;\n      6             learning_rate=config.LEARNING_RATE,\n      7             epochs=1,\n----&gt; 8             layers='heads')\n\n~\\Desktop\\SpaceNet_MaskRCNN\\mrcnn\\model.py in train(self, train_dataset, val_dataset, learning_rate, epochs, layers, augmentation, custom_callbacks, no_augmentation_sources)\n   2372             max_queue_size=100,\n   2373             workers=workers,\n-&gt; 2374             use_multiprocessing=True,\n   2375         )\n   2376         self.epoch = max(self.epoch, epochs)\n\n~\\Anaconda3\\envs\\MaskRCNN\\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\\MaskRCNN\\lib\\site-packages\\keras\\engine\\training.py in fit_generator(self, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, validation_freq, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch)\n   1656             use_multiprocessing=use_multiprocessing,\n   1657             shuffle=shuffle,\n-&gt; 1658             initial_epoch=initial_epoch)\n   1659 \n   1660     @interfaces.legacy_generator_methods_support\n\n~\\Anaconda3\\envs\\MaskRCNN\\lib\\site-packages\\keras\\engine\\training_generator.py in fit_generator(model, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, validation_freq, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch)\n    179             batch_index = 0\n    180             while steps_done &lt; steps_per_epoch:\n--&gt; 181                 generator_output = next(output_generator)\n    182 \n    183                 if not hasattr(generator_output, '__len__'):\n\n~\\Desktop\\SpaceNet_MaskRCNN\\mrcnn\\model.py in data_generator(dataset, config, shuffle, augment, augmentation, random_rois, batch_size, detection_targets, no_augmentation_sources)\n   1707                     load_image_gt(dataset, config, image_id, augment=augment,\n   1708                                 augmentation=augmentation,\n-&gt; 1709                                 use_mini_mask=config.USE_MINI_MASK)\n   1710 \n   1711             # Skip images that have no instances. This can happen in cases\n\n~\\Desktop\\SpaceNet_MaskRCNN\\mrcnn\\model.py in load_image_gt(dataset, config, image_id, augment, augmentation, use_mini_mask)\n   1209     \"\"\"\n   1210     # Load image and mask\n-&gt; 1211     image = dataset.load_image(image_id)\n   1212     mask, class_ids = dataset.load_mask(image_id)\n   1213     original_shape = image.shape\n\n&lt;ipython-input-6-e65486a11b62&gt; in load_image(self, image_id)\n     14         ResimAdi = str(ResimPATH)+'RGB-PanSharpen_AOI_2_Vegas_img'+str(image_id)+'.tif'\n     15 \n---&gt; 16         with rasterio.open(ResimAdi,'r') as ds:\n     17             arr = ds.read()\n     18             image = reshape_as_image(arr)\n\n~\\Anaconda3\\envs\\MaskRCNN\\lib\\site-packages\\rasterio\\__init__.py in open(path, mode, driver, width, height, count, crs, transform, dtype, nodata, **kwargs)\n    191         raise ValueError(\n    192             \"mode string must be one of 'r', 'r+', or 'w', not %s\" % mode)\n--&gt; 193     s.start()\n    194     return s\n    195 \n\nrasterio\/_base.pyx in rasterio._base.DatasetReader.start (rasterio\/_base.c:2969)()\n\nRasterioIOError: D:\/DATASET\/SpaceNet\/Train\/AOI_2_Vegas_Train\/RGB-PanSharpen\/RGB-PanSharpen_AOI_2_Vegas_img1189.tif: No such file or directory\n<\/code><\/pre>\n\n<p>I know, it says that there were no direectory\/file with the name of <\/p>\n\n<blockquote>\n  <p>D:\/DATASET\/SpaceNet\/Train\/AOI_2_Vegas_Train\/RGB-PanSharpen\/RGB-PanSharpen_AOI_2_Vegas_img1189.tif<\/p>\n<\/blockquote>\n\n<p>However the problem is, the id of image (1189) <code>image_id<\/code> has never existed.<\/p>\n\n<p>Does the Tensorflow\/Keras create random number instead of the image_id?<\/p>\n","answers":[]}
+{"QuestionId":57871767,"AnswerCount":1,"Tags":"<python><machine-learning><keras><neural-network>","CreationDate":"2019-09-10T13:21:32.027","AcceptedAnswerId":null,"OwnerUserId":6367172.0,"Title":"Adding prior belief into a neural Network","Body":"<p>I am busy with a classification problem, with three classes. One of the classes is never predicted\/classified. I would like to know if there s anyway to inject a prior belief into my neural network, be design or not.<\/p>\n\n<p>My football prediction model predicts <code>[Draws , Home Win , Away Win]<\/code>. My classes are pretty balanced (40% , 30 % , 30%). The class <code>[Draw]<\/code> that accounts for 40% of the data is the one the my NN never predicts. My dataset contains 1900 samples.<\/p>\n\n<p>I am using a deep NN with 2 to 4 hidden layers.<\/p>\n\n<p>My code of my best model(based on training\/val loss) is as follows:<\/p>\n\n<pre><code>X_all = df.copy()\n\ntrain_cols = ['a_line0','a_line1','a_line2','a_line3','a_line4','a_line5',\n              'a_line6','a_line7','a_line8','a_line9','a_line10','h_line0',\n              'h_line1','h_line2','h_line3','h_line4','h_line5','h_line6',\n              'h_line7','h_line8','h_line9','h_line10','odds0','odds1','odds2']\n\n\nx = X_all[train_cols]\n\nx_v = x.values #returns a numpy array\nmin_max_scaler = preprocessing.MinMaxScaler()\nx_scaled = min_max_scaler.fit_transform(x_v)\nx = pd.DataFrame(x_scaled)\n\ny = X_all['result']\nohe = OneHotEncoder(n_values=3,categories='auto')\ny = ohe.fit_transform(y.reshape(-1,1))\n\nX_train, X_test, y_train, y_test = train_test_split(x, y, test_size=0.3, random_state=0)\n\nfor lr,ep in [(0.001,300)]:\n    model = Sequential()\n    model.add(Dense(25, input_dim=25, activation='relu'))\n    model.add(Dense(36, activation='relu'))\n    model.add(Dropout(0.2))\n    model.add(Dense(12, activation='relu'))\n    model.add(Dense(3, activation='sigmoid'))\n    adam = kr.optimizers.Adam(lr=lr, decay=1e-6)\n    model.compile(loss='categorical_crossentropy', optimizer=adam, metrics=['accuracy'])\n    model.fit(X_train, y_train, epochs=ep, batch_size=10,verbose = 0)\n    _, accuracy = model.evaluate(X_test, y_test)\n    _, accuracy1 = model.evaluate(X_train, y_train)\n    print('Testing Accuracy: %.2f' % (accuracy*100),'Train Accuracy: %.2f' % (accuracy1*100), 'learning rate : ', lr)\n<\/code><\/pre>\n\n<p>I apologise if the code is a bit messy.\nMy model also overfits by +- 16% (52% vs 68%) on this config of my network.<\/p>\n","answers":[{"AnswerId":"57876034","CreationDate":"2019-09-10T17:56:10.680","ParentId":null,"OwnerUserId":"4685471","Title":null,"Body":"<p>Since you are in a multi-class <strong>single-label<\/strong> setting (i.e. your labels are mutually exclusive), you should not use <code>sigmoid<\/code> as activation in your final layer; change it to<\/p>\n\n<pre><code>model.add(Dense(3, activation='softmax'))\n<\/code><\/pre>\n\n<p>Also, dropout should not be used by default; remove it for starters, and only add it if it improves the result.<\/p>\n"}]}
+{"QuestionId":57872199,"AnswerCount":0,"Tags":"<python><unit-testing><tensorflow><python-unittest>","CreationDate":"2019-09-10T13:45:40.330","AcceptedAnswerId":null,"OwnerUserId":3407000.0,"Title":"Advantages of using tf.test vs unittest","Body":"<p>I have recently started to introduce unit-testing in my TF-based project. In doing so, I have found that the package <code>tf.test<\/code> provides an alternative to the native Python <code>unittest<\/code>, but I could barely find any information about the topic on the Internet. This lack of information, added to some other facts, makes me think this technology might not be completely mature yet. An example of this is that the deprecated <code>test_session<\/code> method is still there, yielding a skipped test in every run.<\/p>\n\n<p>My question would then be, what would the advantages of using <code>tf.test<\/code> vs the native <code>unittest<\/code> be in a Tensorflow project? In which cases would it be reasonable to use the former?<\/p>\n","answers":[]}
+{"QuestionId":57872221,"AnswerCount":0,"Tags":"<tensorflow><kwargs>","CreationDate":"2019-09-10T13:46:42.093","AcceptedAnswerId":null,"OwnerUserId":7303036.0,"Title":"When attempting to build a conditional normalizing flow, running into errors with nested **kwargs","Body":"<p>I use some fancy machine learning terms, but I think someone with a better understanding of base python and especially how kwargs work could also help me, so don't let that scare you off.  I'm trying to build a conditional normalizing flow in tensorflow.  TF has some kwargs set up to make that task easier but it is not implemented in base tensorflow (haven't checked 2.0 but 1.13.1 is the version I'm using).  The problem is, there are some nested kwargs where one function has a kwargs which calls another function which has a kwargs and so on.  I have tried to built a nested diction like this <code>kwargs = {\"kwargs\" : {\"condit\" : tf.concat([params,x_i], axis = 0)}}<\/code>, but when I call something like this <code>dist.log_prob(y_i, **kwargs)<\/code> where <code>dist<\/code> is defined like this:<\/p>\n\n<pre><code>base_dist1 = tfd.MultivariateNormalDiag(loc=tf.zeros([dim]))\n\nfor k in range(4):\n    bijector.append(tfb.RealNVP(2, shift_and_log_scale_fn=tfb.real_nvp.real_nvp_conditional_template(\n                                     hidden_layers=[16])))\n    bijector.append(tfb.BatchNormalization(training=boolval))\n    bijector.append(tfb.Permute(permutation=[3,4,1,2,0]))#this needs to change as states and features change\n\nbijector.append(tfb.RealNVP(2, shift_and_log_scale_fn=tfb.real_nvp.real_nvp_conditional_template(\n                                     hidden_layers=[16])))\n\n\nchainedbij = tfb.Chain(list(reversed(bijector)))\n\ndist = tfd.ConditionalTransformedDistribution(\n    distribution=base_dist1,\n    bijector=chainedbij)\n<\/code><\/pre>\n\n<p>I get this error\/traceback: <\/p>\n\n<pre><code>  File \"&lt;ipython-input-1-0a984deabb65&gt;\", line 1, in &lt;module&gt;\n    runfile('\/home\/cameron\/AnacondaProjects\/Max_Like_NF\/VAR_NF_test.py', wdir='\/home\/cameron\/AnacondaProjects\/Max_Like_NF')\n\n  File \"\/home\/cameron\/anaconda3\/lib\/python3.6\/site-packages\/spyder\/utils\/site\/sitecustomize.py\", line 866, in runfile\n    execfile(filename, namespace)\n\n  File \"\/home\/cameron\/anaconda3\/lib\/python3.6\/site-packages\/spyder\/utils\/site\/sitecustomize.py\", line 102, in execfile\n    exec(compile(f.read(), filename, 'exec'), namespace)\n\n  File \"\/home\/cameron\/AnacondaProjects\/Max_Like_NF\/VAR_NF_test.py\", line 102, in &lt;module&gt;\n    dens = -calclik(ysim)\n\n  File \"\/home\/cameron\/AnacondaProjects\/Max_Like_NF\/VAR_NF_test.py\", line 98, in calclik\n    lprob = lprob + tf.reduce_sum(dist.log_prob(y_i, **kwargs))\n\n  File \"\/home\/cameron\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow_probability\/python\/internal\/distribution_util.py\", line 2073, in _fn\n    return fn(*args, **kwargs)\n\n  File \"\/home\/cameron\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow_probability\/python\/distributions\/conditional_distribution.py\", line 45, in log_prob\n    return self._call_log_prob(value, name, **condition_kwargs)\n\n  File \"\/home\/cameron\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow_probability\/python\/distributions\/distribution.py\", line 697, in _call_log_prob\n    return self._log_prob(value, **kwargs)\n\n  File \"\/home\/cameron\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow_probability\/python\/internal\/distribution_util.py\", line 2073, in _fn\n    return fn(*args, **kwargs)\n\nTypeError: _log_prob() got an unexpected keyword argument 'kwargs'\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57872232,"AnswerCount":1,"Tags":"<python><keras><conv-neural-network><cross-validation><k-fold>","CreationDate":"2019-09-10T13:47:15.350","AcceptedAnswerId":null,"OwnerUserId":8002024.0,"Title":"How to prevent one fold to perform a lot worse than the other 9 in 10-fold cross validation for CNN classification","Body":"<p>I'm currently working on a 2D CNN in Keras for MRI classification. The class ratio is about 60\/40, I have 155 patients, each with one MRI consisting of around 180 slices, the input of the CNN is a slice of an MRI image (256*256 px) (so input in total is ~27900 images, each 256*256 pixels).<\/p>\n\n<p>I tested different models and always evaluated them with shuffled stratified 10 fold cross validation and an EarlyStopping monitor and they all performed very well, around 95% to 98% validation accuracy. But everytime, one or two folds perform a lot worse then the other ones (70% to 80% validation accuracy). Since the folds are randomized I would expect the folds to all perform equally well.<\/p>\n\n<p>Can somebody explain how this could happen and how to prevent it? <\/p>\n\n<p>Plots for accuracy and loss:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/bN6SE.png\" rel=\"nofollow noreferrer\">Train accuracy and validation accuracy<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/0mxHU.png\" rel=\"nofollow noreferrer\">Train loss and validation loss<\/a><\/p>\n\n<p>This is part of one of the models: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>num_classes = 2\nimg_size = 256\nbatch_size = 200\n\n# Because of EarlyStopping monitor, the number of epochs doesn't really matter\nnum_epochs = 1000\n\nkfold_splits = 10\nskf = StratifiedKFold(n_splits=kfold_splits, shuffle=True)\n\n# Here the data is split \nfor index, (train_index, test_index) in enumerate(skf.split(x_data_paths, y_data_paths)):\n\n    x_train, x_test = np.array(x_data_paths)[train_index.astype(int)], np.array(x_data_paths)[test_index.astype(int)]\n    y_train, y_test = np.array(y_data_paths)[train_index.astype(int)], np.array(y_data_paths)[test_index.astype(int)]\n\n    training_batch_generator = BcMRISequence(x_train, y_train_one_hot, batch_size)\n    test_batch_generator = BcMRISequence(x_test, y_test_one_hot, batch_size)\n\n    # region Create model (using the functional API)\n    inputs = Input(shape=(img_size, img_size, 1))\n    conv1 = Conv2D(64, kernel_size=5, strides=1, activation='relu')(inputs)\n    pool1 = MaxPooling2D(pool_size=3, strides=(2, 2), padding='valid')(conv1)\n    conv2 = Conv2D(32, kernel_size=3, activation='relu')(pool1)\n    pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)\n    conv3 = Conv2D(16, kernel_size=3, activation='relu')(pool2)\n    pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)\n    flat = Flatten()(pool3)\n    hidden1 = Dense(10, activation='relu')(flat)\n    output = Dense(num_classes, activation='softmax')(hidden1)\n    model = Model(inputs=inputs, outputs=output)\n\n<\/code><\/pre>\n","answers":[{"AnswerId":"57885680","CreationDate":"2019-09-11T09:04:52.820","ParentId":null,"OwnerUserId":"8002024","Title":null,"Body":"<p>Just in case anyone else stumbles upon my question: By using stratified shuffle split with 10 iterations instead of 10 fold cross validation I got rid of the outlier fold. My guess is this bad fold appeared because of some kind of \"batch effect\" (the 10 fold cross validation doesn't shuffle the MRI slices, only the patients). The shuffle split on the other hand does shuffle the entire data set before splitting into train and test, therefore avoiding \"bad\" patients to appear together in one fold. <\/p>\n\n<p>These are the new plots, in case anyone is interested. Same model as before, just shuffle split instead of k fold cross validation.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/m8WH0.png\" rel=\"nofollow noreferrer\">Train accuracy and validation accuracy<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vo0IN.png\" rel=\"nofollow noreferrer\">Train loss and validation loss<\/a><\/p>\n"}]}
+{"QuestionId":57872334,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-10T13:53:50.893","AcceptedAnswerId":null,"OwnerUserId":3447343.0,"Title":"parallel inference in tensorflow (CPUs)","Body":"<p>this is a really basic question, but I can't get the answer anywhere.<\/p>\n\n<p>Using tensorflow, can I do inference (<code>tf.keras.Model.predict()<\/code>) on multiple CPUs in parallel? I am using <code>tf.data.Dataset<\/code> to represent my data, but from the <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/Model#predict\" rel=\"nofollow noreferrer\">documentation<\/a> it seems multiprocessing can be used only with generators or tf.keras.Sequence objects. Why is that? Am I supposed to somehow create an ordinary python generator from the <code>Dataset<\/code> or to manage the parallelism on my own with <code>multiprocessing<\/code> package? What would be the standard way of doing this?<\/p>\n\n<p>Thanks for any hints.<\/p>\n","answers":[]}
+{"QuestionId":57872336,"AnswerCount":0,"Tags":"<python><tensorflow><keras><cudnn>","CreationDate":"2019-09-10T13:54:00.993","AcceptedAnswerId":null,"OwnerUserId":880783.0,"Title":"Should I be able to use precompiled TensorFlow 2 on a Laptop Geforce RTX 2060 GPU?","Body":"<p><em>I changed the title from \"CUDNN_STATUS_ALLOC_FAILED with minimal network and data using CUDA 10.0 and CuDNN 7.6.x\" to better describe my problem.<\/em><\/p>\n\n<p>I have Laptop with an NVIDIA Geforce RTX 2060 GPU, that should have a Turing architecture and 7.5 compute capabilities: <a href=\"https:\/\/en.wikipedia.org\/wiki\/Turing_(microarchitecture)\" rel=\"nofollow noreferrer\">https:\/\/en.wikipedia.org\/wiki\/Turing_(microarchitecture)<\/a><\/p>\n\n<p>According to the CuDNN support matrix, that GPU should be supported up the the most current CuDNN 7.6.3: <a href=\"https:\/\/docs.nvidia.com\/deeplearning\/sdk\/cudnn-support-matrix\/index.html\" rel=\"nofollow noreferrer\">https:\/\/docs.nvidia.com\/deeplearning\/sdk\/cudnn-support-matrix\/index.html<\/a><\/p>\n\n<p>Having said that, this simple example code fails on one of my machines with <code>tensorflow-gpu<\/code>:<\/p>\n\n<pre><code>import numpy as np\nfrom tensorflow.keras import layers, models, optimizers\n\nmodel = models.Sequential()\nmodel.add(layers.Conv1D(1, 3, input_shape=(8, 1)))\n\n\noptimizer = optimizers.Adam(lr=1e-6)\nmodel.compile(optimizer=optimizer, loss='mse')\nx = np.zeros((1, *model.input.shape[1:]))\ny = np.zeros((1, *model.output.shape[1:]))\nmodel.fit(x, y)\n<\/code><\/pre>\n\n<p>The output, cleaned up quite a bit, is:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:44] Successfully opened \n  dynamic library cudart64_100.dll\nI tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:44] Successfully opened \n  dynamic library nvcuda.dll\nI tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1304] Created TensorFlow device \n  (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 4608 MB memory) -&gt; physical GPU \n  (device: 0, name: GeForce RTX 2060, pci bus id: 0000:01:00.0, compute capability: 7.5)\nI tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:44] Successfully opened \n  dynamic library cudnn64_7.dll\nE tensorflow\/stream_executor\/cuda\/cuda_dnn.cc:329] Could not create cudnn handle: \n  CUDNN_STATUS_ALLOC_FAILED\nW tensorflow\/core\/common_runtime\/base_collective_executor.cc:216] \n  BaseCollectiveExecutor::StartAbort Unknown: Failed to get convolution algorithm. This \n  is probably because cuDNN failed to initialize, so try looking to see if a warning log \n  message was printed above.\n     [[{{node sequential\/conv3d\/Conv3D}}]]\nTraceback (most recent call last):\n  File \"NotWorking.py\", line 12, in &lt;module&gt;\n    model.fit(x, y)\n...\ntensorflow.python.framework.errors_impl.UnknownError:  Failed to get convolution \n  algorithm. This is probably because cuDNN failed to initialize, so try looking to see \n  if a warning log message was printed above.\n     [[node sequential\/conv3d\/Conv3D (defined at C:\\Users\\bers\\AppData\\Local\\Programs\\\n  Python\\Python37\\lib\\site-packages\\tensorflow_core\\python\\framework\\ops.py:1751) ]] \n  [Op:__inference_distributed_function_505]\n\nFunction call stack:\ndistributed_function\n<\/code><\/pre>\n\n<p>The problem is reproducible in <code>tensorflow-gpu==1.14.0<\/code>, <code>2.0.0rc0<\/code>, <code>2.0.0rc1<\/code> (installed via <code>pip<\/code>), with CUDA 10.0.130 and several versions of CuDNN [7.6.3.30, 7.6.2.24, 7.6.0.64] for CUDA 10.0 on Python 3.7.4.<\/p>\n\n<p><em>Edit 1:<\/em> I have already uninstalled and reinstalled TensorFlow, Python, and everything that looks NVIDIA-related from my system, including CUDA, CuDNN, and graphics drivers until Windows could only assign a basic VGA driver to the NVIDIA GPU. The problem does not appear with only <code>tensorflow<\/code> installed, and not on a different computer with a Quadro M5000 with CuDNN 7.6.2 on CUDA 10.0.<\/p>\n\n<p><em>Edit 2:<\/em> I have run a number of experiments using the code below. These are my results:<\/p>\n\n<p>Failing:<\/p>\n\n<ol>\n<li>Python 3.7.4, tensorflow-gpu==2.0.0rc0\/2.0.0rc1 (compiled for CUDA 10.0, CuDNN 7.6.0), CUDA 10.0.130, CuDNN 7.6.3.30\/7.6.2.24\/7.6.0.64 for CUDA 10.0<\/li>\n<li>Python 3.7.4, tensorflow-gpu==1.14.0\/1.13.1 (compiled for CUDA 10.0, CuDNN 7.4.1), CUDA 10.0.130, CuDNN 7.6.0.64 for CUDA 10.0<\/li>\n<li>Python 3.6.8, tensorflow-gpu==1.14.0\/1.13.1 (compiled for CUDA 10.0, CuDNN 7.4.1), CUDA 10.0.130, CuDNN 7.6.3.30\/7.6.0.64 for CUDA 10.0<\/li>\n<\/ol>\n\n<p>Working:<\/p>\n\n<ol start=\"4\">\n<li>Python 3.7.4, tensorflow-gpu==1.14.0\/1.13.1 (compiled for CUDA 10.0, CuDNN 7.4.1), CUDA 10.0.130, CuDNN 7.5.1.10\/7.4.2.24\/7.4.1.5 for CUDA 10.0<\/li>\n<li>Python 3.6.8, tensorflow-gpu==1.14.0\/1.13.1 (compiled for CUDA 10.0, CuDNN 7.4.1), CUDA 10.0.130, CuDNN 7.5.1.10 for CUDA 10.0<\/li>\n<li>Python 3.6.8, tensorflow-gpu==1.12.3\/1.12.0 (compiled for CUDA 9.0), CUDA 9.0.176, CuDNN 7.6.3.30\/7.6.0.64\/7.5.1.10\/7.1.4\/7.0.5 for CUDA 9.0<\/li>\n<\/ol>\n\n<p>So the problem is independent of the Python version (compare 2 vs. 3), independent of the exact version of <code>tensorflow-gpu<\/code> (compare 1 vs. 2), including independent of the exact version of CuDNN that <code>tensorflow-gpu<\/code> is compiled for (compare 1 vs. 2).<\/p>\n\n<p>So far, the problem appears only for any version of CuDNN>=7.6.0 on CUDA 10.0 - meaning, the latest CuDNN 7.5 works fine on CUDA 10.0 (see 4 and 5), while the earliest CuDNN 7.6 fails on CUDA 10.0 (see 2 and 3), and both are working fine on CUDA 9.0 (see 6).<\/p>\n\n<p>Interestingly, the following variations of the code are also working fine on all system variations mentioned above:<\/p>\n\n<pre><code>model.add(layers.Conv1D(1, 3, input_shape=(3, 1)))  # changed input shape\n<\/code><\/pre>\n\n<p>and <\/p>\n\n<pre><code>model.add(layers.Conv1D(1, 3, input_shape=(8, 1)))\nmodel.add(layers.Dense(1))  # added Dense layer!\n<\/code><\/pre>\n\n<p>So in summary, some specific TensorFlow code (not even every TensorFlow code) fails only with CuDNN 7.6 on CUDA 10.0. Unfortunately, TensorFlow 2 has been compiled against CuDNN 7.6.0, so I am not able to run able TF2 code.<\/p>\n\n<p>What may be going on here?<\/p>\n","answers":[]}
+{"QuestionId":57872450,"AnswerCount":0,"Tags":"<machine-learning><keras>","CreationDate":"2019-09-10T13:59:37.477","AcceptedAnswerId":null,"OwnerUserId":8581901.0,"Title":"n-step predication with keras","Body":"<p>I want to create an keras model, which do the following:\nThe model gets as input x_t and  predicts y_t1\nand now use\ny_t1(input for model) predicts y_t2.....\ny_t3(input for model) predicts y_t4 and so on.\nI am new to keras functional api and I am not sure if the below code do what i want.<\/p>\n\n<pre><code> def model():\n    input = Input((self.data.shape[1], self.data.shape[2]))\n    print((input))\n    print('shape ({}, {})'.format(self.data.shape[1], self.data.shape[2]))\n    x = LSTM(128, activation='relu', return_sequences=True)(input)\n    x = Dropout(0.2)(x)\n    x = LSTM(128, activation='relu')(x)\n    x = Dropout(0.2)(x)\n    x = Dense(32, activation='relu')(x)\n    x = Dropout(0.2)(x)\n    x = Dense((21), activation='relu')(x)\n    self.model = Model(input, x)\n    for t in range(self.time_step):\n        x = Reshape(target_shape=(1, 21))(x)\n        x = self.model(x) \n    self.model = Model(input, x)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57872554,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><tf.keras>","CreationDate":"2019-09-10T14:06:09.810","AcceptedAnswerId":null,"OwnerUserId":6913957.0,"Title":"Converting tf.keras model to estimator makes loss worst","Body":"<p>I did an experiment to compare tf.keras's performance before and after conversion to estimator and got very different loss:\n1) tf.keras model (No estimator)   : 1706.100 RMSE (+\/- 260.064)\n2) tf.keras converted to estimator : 3912.574 RMSE (+\/- 132.833)<\/p>\n\n<p>I'm sure I did something wrong in estimator conversion, or in the Dataset API but I could not isolate the root cause. Any help is appreciated.<\/p>\n\n<p>It is related to my other post \"Converting CNN-LSTM from keras to tf.keras brings down the accuracy\".<\/p>\n\n<p>partial code with No estimator<\/p>\n\n<pre><code>import tensorflow as tf\nfrom math import sqrt\nfrom sklearn.metrics import mean_squared_error\nfrom tensorflow.keras import Sequential\nfrom tensorflow.keras.layers import Dense\nfrom tensorflow.keras.layers import Flatten\nfrom tensorflow.keras.layers import ConvLSTM2D\nfrom matplotlib import pyplot\nimport numpy as np\nimport pandas as pd\nimport warnings\n\ndef model_fit_convlstm(train_supv, config, first_round):\n    # train X shape = (48, 36) \n    train_x, train_y = train_supv[:, :-1], train_supv[:, -1]\n    # train X lstm shape = (48, 36, 1)\n    train_x_convlstm = train_x.reshape(train_x.shape[0], config['n_seq'], 1, config['n_steps'], config['n_feature'])\n    #\n    input_shape = (config['n_seq'], 1, config['n_steps'], config['n_feature'])\n    # relu = tf.nn.relu\n    relu = 'relu'\n    model = Sequential()\n    model.add(ConvLSTM2D(config['n_filters'], (1,config['n_kernel']), activation=relu, input_shape=input_shape))\n    model.add(Flatten())\n    model.add(Dense(config['n_nodes'], activation=relu))\n    model.add(Dense(1))    \n    adam = tf.train.AdamOptimizer()\n    # adam = 'adam'\n    mse = tf.keras.losses.mean_squared_error\n    # mse = 'mse'\n    model.compile(optimizer=adam, loss=mse) \n    model.fit(train_x_convlstm, train_y, epochs=config['n_epochs'], batch_size=config['n_batch'], verbose=0)\n    return model\n\ndef model_predict_convlstm(model, test_supv, config, first_round):\n    test_x, test_y = test_supv[:, :-1], test_supv[:, -1]\n    test_x_convlstm = test_x.reshape(test_x.shape[0], config['n_seq'], 1, config['n_steps'], config['n_feature'])\n    if first_round.state:\n        print('test X shape = {}'.format(test_x.shape))\n        print('test X convlstm shape = {}'.format(test_x_convlstm.shape))\n    #\n    predictions = np.array([])\n    for row in test_x_convlstm:\n        test_x_convlstm_row = row.reshape(1, config['n_seq'], 1, config['n_steps'], config['n_feature'])\n        yhat = model.predict(test_x_convlstm_row, verbose=0)\n        # print(yhat)\n        predictions = np.append(predictions, yhat)\n    #\n    return predictions\n<\/code><\/pre>\n\n<p>partial code with estimator conversion<\/p>\n\n<pre><code>import tensorflow as tf\nfrom math import sqrt\nfrom sklearn.metrics import mean_squared_error\nfrom tensorflow.keras import Sequential\nfrom tensorflow.keras.layers import Dense\nfrom tensorflow.keras.layers import Flatten\nfrom tensorflow.keras.layers import ConvLSTM2D\nfrom matplotlib import pyplot\nimport numpy as np\nimport pandas as pd\nimport warnings\n\ndef model_fit_convlstm(train_supv, config, first_round):\n    # train X shape = (48, 36) \n    train_x, train_y = train_supv[:, :-1], train_supv[:, -1]\n    # train X lstm shape = (48, 36, 1)\n    train_x_convlstm = train_x.reshape(train_x.shape[0], config['n_seq'], 1, config['n_steps'], config['n_feature'])\n    #\n    input_shape = (config['n_seq'], 1, config['n_steps'], config['n_feature'])\n    # relu = tf.nn.relu\n    relu = 'relu'\n    model = Sequential()\n    model.add(ConvLSTM2D(config['n_filters'], (1,config['n_kernel']), activation=relu , input_shape=input_shape))\n    model.add(Flatten())\n    model.add(Dense(config['n_nodes'], activation=relu))\n    model.add(Dense(1))    \n    adam = tf.train.AdamOptimizer()\n    # adam = 'adam'\n    mse = tf.keras.losses.mean_squared_error\n    # mse = 'mse'\n    model.compile(optimizer=adam, loss=mse) \n    print('model {}'.format(model))\n    #\n    with warnings.catch_warnings():\n      warnings.filterwarnings(\"ignore\")\n      estimator = tf.keras.estimator.model_to_estimator(keras_model=model)\n    #\n    # train X lstm shape = (48, 36, 1)\n    # numpy_input_fn takes x: train_x_lstm shape\n    # print('model.input_names {}'.format(model.input_names))\n    train_input_fn = tf.estimator.inputs.numpy_input_fn(\n        x = {model.input_names[0]: train_x_convlstm.astype(np.float32)},\n        y = train_y.astype(np.float32),\n        num_epochs = config['n_epochs'],\n        batch_size = config['n_batch'],\n        shuffle = True\n    )\n    estimator.train(input_fn=train_input_fn)\n    #\n    return model, estimator\n\ndef model_predict_convlstm(model, estimator, test_supv, config, first_round):\n    test_x, test_y = test_supv[:, :-1], test_supv[:, -1]\n    test_x_convlstm = test_x.reshape(test_x.shape[0], config['n_seq'], 1, config['n_steps'], config['n_feature'])\n    if first_round.state:\n        print('test X shape = {}'.format(test_x.shape))\n        print('test X convlstm shape = {}'.format(test_x_convlstm.shape))\n    #\n    predict_input_fn = tf.estimator.inputs.numpy_input_fn(\n        x = {model.input_names[0]: test_x_convlstm.astype(np.float32)},\n        y = None,\n        batch_size=1,\n        num_epochs=1,\n        shuffle=False\n    ) \n    yhat_list = estimator.predict(input_fn=predict_input_fn)\n    # print('yhat_list {}'.format(yhat_list))\n    predictions = np.array([])\n    for yhat_dict in yhat_list:\n      for key, val in yhat_dict.items():\n        # print('yhat = {} {}'.format(key, val))\n        predictions = np.append(predictions, val)\n    #\n    return predictions\n<\/code><\/pre>\n","answers":[{"AnswerId":"58200549","CreationDate":"2019-10-02T11:27:02.343","ParentId":null,"OwnerUserId":"6913957","Title":null,"Body":"<p>More observation from me. I tried another neural network method (MLP). In that case, keras and tf.keras prediction is very close. Conclusion is that the gap introduced by estimator conversion only happens to certain tf.keras libraries and layers.\nThe culprit are ConvLSTM2D and Flatten which I suspect tf.keras may have different default or slightly different algorithm.\nThe following example produces no gap in prediction performance<\/p>\n\n<p>tf.keras<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.keras import Sequential\nfrom tensorflow.keras.layers import Dense\nrelu = tf.keras.activations.relu\nmodel = Sequential()\nmodel.add(Dense(config['n_nodes'], activation=relu, input_dim=config['n_input']))\nmodel.add(Dense(1))\nadam = tf.keras.optimizers.Adam()\nmse = tf.keras.losses.mean_squared_error\nmodel.compile(loss=mse, optimizer=adam, verbose=0)\n# model.fit(train_x, train_y, epochs=config['n_epochs'], batch_size=config['n_batch'], verbose=0)\nwith warnings.catch_warnings():\n  warnings.filterwarnings(\"ignore\")\n  estimator = tf.keras.estimator.model_to_estimator(keras_model=model)\ntrain_input_fn = tf.estimator.inputs.numpy_input_fn(\n    x = {model.input_names[0]: train_x.astype(np.float32)},\n    y = train_y.astype(np.float32),\n    num_epochs = config['n_epochs'],\n    batch_size = config['n_batch'],\n    shuffle = False\n)\nestimator.train(input_fn=train_input_fn)\n<\/code><\/pre>\n\n<p>keras<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense\nmodel = Sequential()\nmodel.add(Dense(config['n_nodes'], activation='relu', input_dim=config['n_input']))\nmodel.add(Dense(1))\nmodel.compile(loss='mse', optimizer='adam')\nmodel.fit(train_x, train_y, epochs=config['n_epochs'], batch_size=config['n_batch'], verbose=0)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57872621,"AnswerCount":0,"Tags":"<python><ios><tensorflow><audio><wav>","CreationDate":"2019-09-10T14:09:52.180","AcceptedAnswerId":null,"OwnerUserId":5295185.0,"Title":"Audio WAV files from iOS are not read by Tensorflow audio_ops.decode_wav: Header mismatch: Expected fmt but found JUNK","Body":"<p>I have a iOS app that records audio as wav files and sends the files to server for audio recognition. The iOS app records audio as wav files with the following configurations<\/p>\n\n<pre><code>let settings = [\n                    AVFormatIDKey: Int(kAudioFormatLinearPCM),\n                    AVSampleRateKey: 16000,\n                    AVNumberOfChannelsKey: 1,\n                    AVLinearPCMBitDepthKey: 16,\n                    AVEncoderAudioQualityKey:AVAudioQuality.max.rawValue\n                    ] as [String : Any]\n<\/code><\/pre>\n\n<p>The files are playing with any media player. But when i try to read the files with tensorflow ops in the server i am getting the following error<\/p>\n\n<pre><code>InvalidArgumentError: Header mismatch: Expected fmt  but found JUNK\n     [[node DecodeWav_3 (defined at &lt;ipython-input-5-36afd53919f4&gt;:4) ]]\n\nErrors may have originated from an input operation.\nInput Source operations connected to node DecodeWav_3:\n ReadFile_3 (defined at &lt;ipython-input-5-36afd53919f4&gt;:3)\n\nOriginal stack trace for 'DecodeWav_3':\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/runpy.py\", line 193, in _run_module_as_main\n    \"__main__\", mod_spec)\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/runpy.py\", line 85, in _run_code\n    exec(code, run_globals)\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/ipykernel_launcher.py\", line 16, in &lt;module&gt;\n    app.launch_new_instance()\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/traitlets\/config\/application.py\", line 658, in launch_instance\n    app.start()\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/ipykernel\/kernelapp.py\", line 505, in start\n    self.io_loop.start()\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/tornado\/platform\/asyncio.py\", line 148, in start\n    self.asyncio_loop.run_forever()\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/asyncio\/base_events.py\", line 438, in run_forever\n    self._run_once()\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/asyncio\/base_events.py\", line 1451, in _run_once\n    handle._run()\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/asyncio\/events.py\", line 145, in _run\n    self._callback(*self._args)\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/tornado\/ioloop.py\", line 690, in &lt;lambda&gt;\n    lambda f: self._run_callback(functools.partial(callback, future))\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/tornado\/ioloop.py\", line 743, in _run_callback\n    ret = callback()\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/tornado\/gen.py\", line 787, in inner\n    self.run()\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/tornado\/gen.py\", line 748, in run\n    yielded = self.gen.send(value)\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/ipykernel\/kernelbase.py\", line 365, in process_one\n    yield gen.maybe_future(dispatch(*args))\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/tornado\/gen.py\", line 209, in wrapper\n    yielded = next(result)\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/ipykernel\/kernelbase.py\", line 272, in dispatch_shell\n    yield gen.maybe_future(handler(stream, idents, msg))\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/tornado\/gen.py\", line 209, in wrapper\n    yielded = next(result)\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/ipykernel\/kernelbase.py\", line 542, in execute_request\n    user_expressions, allow_stdin,\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/tornado\/gen.py\", line 209, in wrapper\n    yielded = next(result)\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/ipykernel\/ipkernel.py\", line 294, in do_execute\n    res = shell.run_cell(code, store_history=store_history, silent=silent)\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/ipykernel\/zmqshell.py\", line 536, in run_cell\n    return super(ZMQInteractiveShell, self).run_cell(*args, **kwargs)\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/IPython\/core\/interactiveshell.py\", line 2855, in run_cell\n    raw_cell, store_history, silent, shell_futures)\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/IPython\/core\/interactiveshell.py\", line 2881, in _run_cell\n    return runner(coro)\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/IPython\/core\/async_helpers.py\", line 68, in _pseudo_sync_runner\n    coro.send(None)\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/IPython\/core\/interactiveshell.py\", line 3058, in run_cell_async\n    interactivity=interactivity, compiler=compiler, result=result)\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/IPython\/core\/interactiveshell.py\", line 3249, in run_ast_nodes\n    if (await self.run_code(code, result,  async_=asy)):\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/IPython\/core\/interactiveshell.py\", line 3326, in run_code\n    exec(code_obj, self.user_global_ns, self.user_ns)\n  File \"&lt;ipython-input-5-36afd53919f4&gt;\", line 4, in &lt;module&gt;\n    data = audio_ops.decode_wav(audio_binary, desired_channels=1)\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/gen_audio_ops.py\", line 227, in decode_wav\n    desired_samples=desired_samples, name=name)\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 788, in _apply_op_helper\n    op_def=op_def)\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/tensorflow\/python\/util\/deprecation.py\", line 507, in new_func\n    return func(*args, **kwargs)\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 3616, in create_op\n    op_def=op_def)\n  File \"\/Users\/minimaci73\/anaconda3\/envs\/Samples\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 2005, in __init__\n    self._traceback = tf_stack.extract_stack()\n<\/code><\/pre>\n\n<p>I couldn't find the reason for the crash. My tensorflow code to read the files are as below<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.contrib.framework.python.ops import audio_ops\nfile = \"\/Users\/minimaci73\/Downloads\/1567690231185.wav\"  #wav file path\naudio_binary = tf.read_file(file)\ndata = audio_ops.decode_wav(audio_binary, desired_channels=1)\npcm16_data = tf.Session().run(data)\npcm16_data = pcm16_data[0] \nprint(pcm16_data)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57873047,"AnswerCount":0,"Tags":"<keras><conv-neural-network><pre-trained-model><transfer-learning><vgg-net>","CreationDate":"2019-09-10T14:31:41.797","AcceptedAnswerId":null,"OwnerUserId":3802483.0,"Title":"keras deleting layers from vgg16 to make it smaller (VGG13)","Body":"<p>I first trained a small network as below:<\/p>\n\n<pre><code>model = models.Sequential()\nmodel.add(layers.Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', input_shape=(256, 256, 3)))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(Dropout(0.2))\nmodel.add(layers.Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform'))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(Dropout(0.2))\nmodel.add(layers.Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_uniform'))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(Dropout(0.2))\nmodel.add(layers.Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_uniform'))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(Dropout(0.2))\nmodel.add(layers.Flatten())\nmodel.add(layers.Dropout(0.5))  #Dropout for regularization\nmodel.add(layers.Dense(512, activation='relu'))\nmodel.add(layers.Dense(1, activation='sigmoid'))  #Sigmoid function at the end because we have just two classes\n<\/code><\/pre>\n\n<p>which performed around 70-75% accuracy with no overfitting issues.<\/p>\n\n<p>Then I decided to train a VGG16 with pretrained weights on Imagenet, where I freeze all the conv layers -but one-, and I add two fc layers, then train it.<\/p>\n\n<p>I tried to optimize the model using Adam optimizer (which performs worse even when I set small learning rates?) I use ReduceLROnPlateau and EarlyStopping, and my network performs around 80-82% accuracy however it overfits greatly (val loss is increasing, and there is a huge gap between the training and validation accuracy)<\/p>\n\n<p>Network is as below:<\/p>\n\n<pre><code>epochs = 50\ncallbacks = []\ndecay = 0.0\n\nearlyStopping = EarlyStopping(monitor='val_loss', patience=10, verbose=0, mode='min')\nmcp_save = ModelCheckpoint('.mdl_wts.hdf5', save_best_only=True, monitor='val_loss', mode='min')\nreduce_lr_loss = ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=3, verbose=1, epsilon=1e-5, mode='min')\n\n\n# build the VGG16 network\nbase_model = applications.VGG16(weights='imagenet', include_top=False, input_shape=(224, 224, 3))\nprint('Model loaded.')\n\n\nfor layer in base_model.layers[:15]:\n    layer.trainable = False\n\n#for layer in base_model.layers:\n#    layer.trainable = False\n\n# add new classifier layers\nflat1 = Flatten()(base_model.layers[-1].output)\nclass1 = Dense(128, kernel_regularizer=l2(0.01), bias_regularizer=l2(0.01), activation='relu', kernel_initializer='he_uniform')(flat1)\noutput = Dense(1, kernel_regularizer=l2(0.01), bias_regularizer=l2(0.01), activation='sigmoid')(class1)\n# define new model\nbase_model = Model(inputs=base_model.inputs, outputs=output)\n\nbase_model.compile(loss='binary_crossentropy',\n                   optimizer=optimizers.Adam(lr=1e-4, decay=decay),\n              metrics=['accuracy'])\n\nimport matplotlib.pyplot as plt\nfrom keras.preprocessing.image import ImageDataGenerator\nimport os\n\nos.environ[\"CUDA_VISIBLE_DEVICES\"]=\"0\"\n\ntrain_dir = '\/home\/d\/Desktop\/s\/data\/train'\neval_dir = '\/home\/d\/Desktop\/s\/data\/eval'\ntest_dir = '\/home\/d\/Desktop\/s\/data\/test'\n\n\n# create a data generator\ntrain_datagen = ImageDataGenerator(rescale=1.\/255,   #Scale the image between 0 and 1\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\nval_datagen = ImageDataGenerator(rescale=1.\/255)  #We do not augment validation data. we only perform rescale\n\ntest_datagen = ImageDataGenerator(rescale=1.\/255)  #We do not augment validation data. we only perform rescale\n\n# load and iterate training dataset\ntrain_generator = train_datagen.flow_from_directory(train_dir,  target_size=(224,224),class_mode='binary', batch_size=16, shuffle='True', seed=42)\n# load and iterate validation dataset\nval_generator = val_datagen.flow_from_directory(eval_dir,  target_size=(224,224),class_mode='binary', batch_size=16, shuffle='True', seed=42)\n# load and iterate test dataset\ntest_generator = test_datagen.flow_from_directory(test_dir,  target_size=(224,224), class_mode=None, batch_size=1, shuffle='False', seed=42)\n\n#The training part\n#We train for 64 epochs with about 100 steps per epoch\nhistory = base_model.fit_generator(train_generator,\n                              steps_per_epoch=train_generator.n \/\/ train_generator.batch_size,\n                              epochs=epochs,\n                              validation_data=val_generator,\n                              validation_steps=val_generator.n \/\/ val_generator.batch_size,\n                              callbacks=[earlyStopping, mcp_save, reduce_lr_loss])\n<\/code><\/pre>\n\n<p>my dataset is small and very different than Imagenet, I do a lot of image augmentation.<\/p>\n\n<p>So only thing I can think of is that the VGG16 network is far too complex for my dataset. I want to use my smaller network but with pretrained weights transferred to initial layers, how can I delete layers off the VGG16 so that I will have a smaller network such as VGG13? (As far as I know layer.pop does not -delete- the layers so I am ruling that out)<\/p>\n","answers":[]}
+{"QuestionId":57873313,"AnswerCount":0,"Tags":"<python><tensorflow><keras><lstm><chatbot>","CreationDate":"2019-09-10T14:48:03.820","AcceptedAnswerId":null,"OwnerUserId":8563619.0,"Title":"Keras - LSTM Autoencoders: Chatbot Prediction","Body":"<h1>Hello everybody!<\/h1>\n\n<p>My name is Luis, nice to meet you all!<\/p>\n\n<p>The <strong>main idea<\/strong> of my project is to implement a system very similar to <em>chatbots<\/em>. I am using my own movies lines dataset for training. For the application part, the user enters one input sentence and then the system output one movie line from our database according to the closest answer to the user's input.<\/p>\n\n<p>To implement that, I chose to implement a <strong>Sequence to Sequence (Seq2Seq) model, LSTM autoencoders<\/strong>. I am following the post and tutorial proposed by Akira Takezawa <a href=\"https:\/\/towardsdatascience.com\/how-to-implement-seq2seq-lstm-model-in-keras-shortcutnlp-6f355f3e5639\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<h2>Here I show my Seq2Seq model:<\/h2>\n\n<pre class=\"lang-py prettyprint-override\"><code>def seq2seq_model_builder(HIDDEN_DIM):\n\n    encoder_inputs = Input(shape=(MAX_LEN, ), dtype='int32',)\n    encoder_embedding = embed_layer(encoder_inputs) \n    encoder_LSTM = LSTM(HIDDEN_DIM, return_state=True)\n    encoder_outputs, state_h, state_c = encoder_LSTM(encoder_embedding)\n\n    decoder_inputs = Input(shape=(MAX_LEN, ), dtype='int32',)\n    decoder_embedding = embed_layer(decoder_inputs) \n    decoder_LSTM = LSTM(HIDDEN_DIM, return_state=True, return_sequences=True)\n    decoder_outputs, _, _ = decoder_LSTM(decoder_embedding, initial_state=[state_h, state_c])\n\n    # NOTE: TimeDistributedDenselayer allows us to apply a layer to every temporal slice of an input\n    outputs = TimeDistributed(Dense(VOCAB_SIZE, activation='softmax'))(decoder_outputs) \n    model = Model([encoder_inputs, decoder_inputs], outputs)\n\n    return model\n<\/code><\/pre>\n\n<h2>What I've for the prediction part by now:<\/h2>\n\n<pre class=\"lang-py prettyprint-override\"><code>for i in range(10):\n    temp = ''\n    ran = random.randrange(0,len(encoder_input_data))\n    pred = model.predict([encoder_input_data[ran:ran+1], decoder_input_data[ran:ran+1]])\n    pred = np.argmax(pred[:,1:20], axis=2)\n    for l in pred:\n        for w in l:\n            if w != 0:\n                if index2word[w] == 'eos':\n                    temp += '\\n'\n                else:\n                    temp += index2word[w]\n                    temp += ' '\n        temp += '\\n'\n    print('-')\n    print(color.BLUE + \"Input: \\n\" + color.END, context[ran])\n    print(color.BLUE + \"Prediction: \\n\" + color.END, temp)\n<\/code><\/pre>\n\n<h2>Results:<\/h2>\n\n<p>Consists of an input (one sentence already from my database\/used on training) and a predicted output (using one sentence already from my database as a decoder input data).<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Input: \n So, everybody can watch.\nPrediction: \n i do not like watching the game with people who do not know the game people talk \n\n\nInput: \n Lane said she's with you and Rory.\nPrediction: \n ms slow down \n\n\nInput: \n Lorelai, this is hardly the place. \nPrediction: \n i are the order and it is not happening \n<\/code><\/pre>\n\n<h2>Problem:<\/h2>\n\n<p>I want to make this prediction with an \"unknown\" <em>free input sentence<\/em> from the user.\nI know that I should first make a function to transform the sentence into an int vector.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def sentence_to_seq(sentence):\n\n    sentence = clean_text(sentence)\n    return [word2index.get(word, word2index['eos']) for word in sentence.split()]\n<\/code><\/pre>\n\n<p><strong>This is working well!<\/strong><\/p>\n\n<h2>But,<\/h2>\n\n<p>What about the prediction function? What should I put on \"???\" ?\nOr, probably I should change the (sampling) prediction structure at all. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>pred = model.predict([sentence_to_seq(\"Hello World\"),????????)\n<\/code><\/pre>\n\n<p><strong>Could someone tell me how to do it?<\/strong><\/p>\n\n<p>Feel free to contact me anytime.\nThanks in advance.<\/p>\n\n<p>Best Regards!<\/p>\n","answers":[]}
+{"QuestionId":57873324,"AnswerCount":1,"Tags":"<python><numpy><keras>","CreationDate":"2019-09-10T14:48:24.310","AcceptedAnswerId":null,"OwnerUserId":11515599.0,"Title":"Keras fit_generator raises ValueError because multiple input not correct","Body":"<p>I have a working model that I can train regularly when using fit on a small dataset. However, due to the large total dataset size, I want to switch to a generator. <\/p>\n\n<p>My model is a variational autoencoder that has 3 input layers. First I got the model working using the regular fit method. When I run the model as follows it works:<\/p>\n\n<pre><code>history = bvae.fit([data_class, data_regress, n_particles],\n                        epochs = epochs,\n                        verbose = 1,\n                        batch_size = batch_size,\n                        validation_split = 0.2,\n                        callbacks = callback_list)\n<\/code><\/pre>\n\n<p>When I got this working I wanted to switch to a generator so that I could use more training samples. I tried to implement my generator as follows:<\/p>\n\n<pre><code>class DataGenerator(keras.utils.Sequence):\n    'Generates data for Keras'\n    def __init__(self, filename, sample_size, batch_size):\n        self.filename = filename\n        self.sample_size = sample_size\n        self.batch_size = batch_size          # batch size\n        self.channel = h5py.File(self.filename, 'r')\n        self.list_class_arg = range(3, 67, 4)\n        self.list_regress_arg = [x for x in range(1, 67) if x not in self.list_class_arg]\n\n    def __len__(self):\n        'Denotes the number of batches per epoch'\n        return int(np.floor(self.sample_size \/ self.batch_size))\n\n    def __getitem__(self, index):\n        'Generate one batch of data'\n        # selects data\n        events = self.channel['SM'][index*self.batch_size:(index+1)*self.batch_size]\n        # splits events into class and regress\n        events_class = self.class_norm_to_int(events[:,self.list_class_arg])\n        events_regress = events[:,self.list_regress_arg]\n        # count number of particles in event\n        n_particles = self.count_particles(events)\n\n        print(type(events_class))\n        print(type(events_regress))\n        print(type(n_particles))\n\n        return events_class, events_regress, n_particles\n\n    def count_particles(self, data):\n        'Counts the number of particles per events and returns that number in an array'\n        n_particles = np.zeros(len(data), dtype='int32')\n        for i in range(len(data)):\n            for j in range(3,67,4):\n                if data[i,j] != 0:\n                    n_particles[i] += 1\n                else:\n                    break\n        return n_particles\n\n    def class_norm_to_int(self, data_class):\n        'Transform the particle type to an interger number'\n        class_shape = np.shape(data_class)\n        data_class = np.ndarray.flatten(data_class)\n        data_class = [(x*2.1659782700803696)+5.2390088017131005 if x!=0 else x for x in data_class]\n        data_class = np.reshape(data_class, newshape=class_shape)\n        return data_class\n<\/code><\/pre>\n\n<p>But then I get this error:<\/p>\n\n<pre><code>Using TensorFlow backend.\nEpoch 1\/1\n&lt;class 'numpy.ndarray'&gt;\n&lt;class 'numpy.ndarray'&gt;\n&lt;class 'numpy.ndarray'&gt;\nTraceback (most recent call last):\n  File \"bvae_train_generator.py\", line 96, in &lt;module&gt;\n    callbacks = callback_list)\n  File \"\/home\/usr\/venv\/lib\/python3.6\/site-packages\/keras\/legacy\/interfaces.py\", line 91, in wrapper\n    return func(*args, **kwargs)\n  File \"\/home\/usr\/venv\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\", line 1658, in fit_generator\n    initial_epoch=initial_epoch)\n  File \"\/home\/usr\/venv\/lib\/python3.6\/site-packages\/keras\/engine\/training_generator.py\", line 215, in fit_generator\n    class_weight=class_weight)\n  File \"\/home\/usr\/venv\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\", line 1443, in train_on_batch\n    class_weight=class_weight)\n  File \"\/home\/usr\/venv\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\", line 757, in _standardize_user_data\n    exception_prefix='input')\n  File \"\/home\/usr\/venv\/lib\/python3.6\/site-packages\/keras\/engine\/training_utils.py\", line 105, 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([[5., 5., 6., ..., 0., 0., 0.],\n       [5., 0., 0., ..., 0., 0., 0.],\n       [5., 7., 7., ..., 0., 0., 0.],\n       ...,\n       [5., 5., 0., ..., 0., 0., 0.],\n       [5., 5., 0., ..., 0., 0., 0....\n<\/code><\/pre>\n\n<p>It seems to think that I am only returning a list with one array, but as you can see in the result of the print statements, all three given variables are of the type numpy.ndarray<\/p>\n\n<p>I've tried changing the return in __getitem__ to:<\/p>\n\n<pre><code>return [events_class, events_regress, n_particles]\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>return (events_class, events_regress, n_particles)\n<\/code><\/pre>\n\n<p>However, this doesn't seem to solve the problem. Any help is welcome!<\/p>\n\n<p><strong>Edit:<\/strong><\/p>\n\n<p>My model has the following form:<\/p>\n\n<pre><code>def initialize_bvae(beta, n_particles, latent, activation_func, extra_layer = False):\n\n    'Network parameters'\n    input_shape_class = (n_particles, )\n    input_shape_regress = (n_particles*3 + 2, )\n    hidden_layer_dim = 512\n    latent_dim = n_particles*latent\n\n\n    'Build encoder'\n    inputs_class = Input(shape=input_shape_class, name='input_1') # Input for classification\n    inputs_class_t0 = Lambda(lambda i: i)(inputs_class)\n    inputs_class_t1 = Lambda(lambda x: K.one_hot(K.cast(x, 'int32'), 8))(inputs_class_t0) # Onehot encode classification input\n    inputs_class_t2 = Reshape((n_particles*8,))(inputs_class_t1)\n\n    inputs_regress = Input(shape=input_shape_regress, name='input_2') # Input for regression\n\n    inputs = Concatenate()([inputs_class_t2, inputs_regress]) # Combine inputs\n\n    x = Dense(hidden_layer_dim, activation=activation_func)(inputs)\n    if extra_layer:\n        x = Dense(int(hidden_layer_dim), activation=activation_func)(x)\n        x = Dense(int(hidden_layer_dim\/2), activation=activation_func)(x)\n\n    x = Dense(int(hidden_layer_dim\/2), activation=activation_func)(x)\n    x = Dense(int(hidden_layer_dim\/4), activation=activation_func)(x)\n\n    'Latent space'\n    z_mean = Dense(latent_dim, activation='linear', name='z_mean')(x)\n    z_log_var = Dense(latent_dim, activation='linear', name='z_log_var')(x)\n    z = Lambda(sampling, output_shape=(latent_dim,), name='z')([z_mean, z_log_var])\n\n    encoder = Model([inputs_class, inputs_regress], [z_mean, z_log_var, z], name='encoder')\n\n    'Build decoder'\n    inputs_latent = Input(shape=(latent_dim,), name='z_sampling')\n    inputs_particles = Input(shape=(1,), name='input_3') # Input for particles\n\n    # Implement mask\n    mask_fn = lambda n: K.concatenate((K.ones(shape=latent*n, dtype='float32'), K.zeros(shape=latent_dim-(latent*n), dtype='float32')))\n    mask_layer = Lambda(lambda x: K.map_fn(mask_fn, K.cast(x, 'int32'), dtype='float32'))(inputs_particles)\n    masked_latent = Lambda(lambda x: x[0]*x[1])([inputs_latent, mask_layer])\n\n    x = Dense(int(hidden_layer_dim\/4), activation=activation_func)(masked_latent)\n    x = Dense(int(hidden_layer_dim\/2), activation=activation_func)(x)\n    if extra_layer:\n        x = Dense(int(hidden_layer_dim\/2), activation=activation_func)(x)\n        x = Dense(int(hidden_layer_dim), activation=activation_func)(x)\n    x = Dense(hidden_layer_dim, activation=activation_func)(x)\n\n    outputs_class_t = Dense(n_particles*8, activation='elu')(x)\n    outputs_class_t2 = Reshape((n_particles, 8))(outputs_class_t)\n    outputs_class = Conv1D(8, 1, activation='softmax')(outputs_class_t2) # Classification output\n    outputs_regress = Dense(n_particles*3+2, activation='linear')(x) #Regression output\n\n    decoder = Model([inputs_latent, inputs_particles], [outputs_class, outputs_regress], name='decoder')\n\n    'Build autoencoder'\n    outputs = decoder([encoder([inputs_class, inputs_regress])[2], inputs_particles])\n    bvae = Model([inputs_class, inputs_regress, inputs_particles], outputs, name='bvae')\n\n    'Calculate loss'\n    classification_loss = categorical_crossentropy(inputs_class_t1, outputs[0])\n    regression_loss = mse(inputs_regress, outputs[1])\n    reconstruction_loss = 20*K.mean(classification_loss) + K.mean(regression_loss)\n    kl_loss = -0.5 * K.sum((1+z_log_var-K.square(z_mean)-K.exp(z_log_var)), axis=-1)\n    bvae_loss = K.mean((1-beta) * reconstruction_loss + beta * kl_loss)\n    bvae.add_loss(bvae_loss)\n\n    bvae.compile(optimizer = keras.optimizers.Adam(lr=1e-3, beta_1=0.9, beta_2=0.999))\n    return bvae, encoder, decoder\n<\/code><\/pre>\n\n<p><strong>Edit 2:<\/strong><\/p>\n\n<p>Here is how I call the <code>fit_generator<\/code>:<\/p>\n\n<pre><code>history = bvae.fit_generator(load_data.DataGenerator(file_training, 1e6, 50000),\n                        epochs = epochs,\n                        verbose = 1,\n                        workers=0,\n                        validation_data = load_data.DataGenerator(file_validation, 1e5, 50000),\n                        callbacks = callback_list)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57953418","CreationDate":"2019-09-16T08:54:57.373","ParentId":null,"OwnerUserId":"7442908","Title":null,"Body":"<p>you are defining multiple inputs in your model input_regress, input_class, input_latent and input_particles. All of these should be fed data while fitting the model whereas you are just returning (None, n) dimensional arrays from your DataGenerator. The error due to mismatching dimensions will go away if you feed all your inputs.<\/p>\n"}]}
+{"QuestionId":57873420,"AnswerCount":1,"Tags":"<keras><loss><cross-entropy>","CreationDate":"2019-09-10T14:54:27.550","AcceptedAnswerId":57886613.0,"OwnerUserId":11304826.0,"Title":"Inconsistancies of loss function results with Keras","Body":"<p>I am implementing a CNN coupled to a multiple instance learning layer. In brief, I've got this, with C the number of categories:<\/p>\n\n<pre><code>    [1 batch of images, 1 label] &gt; CNN &gt; Custom final layer -&gt; [1 vector of size C]\n<\/code><\/pre>\n\n<p>My final layer just sums up the previous layer for the moment. To be clear, 1 batch of inputs only gives 1 single ouput. The batch corresponds therefore to multiple instances fetched in 1 single bag associated to 1 label. <\/p>\n\n<p>When I train my model and validate it with the same set:<\/p>\n\n<pre><code>history = model.fit_generator(\n    generator=training_generator,\n    steps_per_epoch=training_set.batch_count,\n    epochs=max_epoch,\n    validation_data=training_generator\n    validation_steps=training_set.batch_count)\n<\/code><\/pre>\n\n<p>I've got 2 different results betwen the training and the validation sets, in spite of being the same:<\/p>\n\n<pre><code>  35\/35 [==============================] - 30s 843ms\/step - loss: 1.9647 - acc: 0.2857 - val_loss: 1.9403 - val_acc: 0.3714\n<\/code><\/pre>\n\n<p>The loss function is just the categorical cross entropy as implemented in Keras (I've got 3 categories). I have implemented my own loss function to get some insight about what happens. Unfortunately, I obtain another inconsistency between the regular loss and my custom loss function:<\/p>\n\n<pre><code>  35\/35 [==============================] - 30s 843ms\/step - loss: 1.9647 - acc: 0.2857 - bag_loss: 1.1035 - val_loss: 1.9403 - val_acc: 0.3714 - val_bag_loss: 1.0874\n<\/code><\/pre>\n\n<p>My loss function:<\/p>\n\n<pre><code>def bag_loss(y_true, y_predicted):\n    y_true_mean = keras.backend.mean(y_true, axis=0, keepdims=False)\n    y_predicted_mean = keras.backend.mean(y_predicted, axis=0, keepdims=False)\n    loss = keras.losses.categorical_crossentropy(y_true_mean, y_predicted_mean)\n    return loss\n<\/code><\/pre>\n\n<p>The final layer of my model (I only shown the call part, for concision):<\/p>\n\n<pre><code>def call(self, x):\n    x = kb.sum(x, axis=0, keepdims=True)\n    x = kb.dot(x, self.kernel)\n    x = kb.bias_add(x, self.bias)\n    out = kb.sigmoid(x)\n\n    return out\n<\/code><\/pre>\n\n<p>After inspecting the code with TensorBoard and the TensorFlow Debugger, I have found out that, indeed, my beg loss and the regular loss return the same value at somme point. But then, Keras perform 6 supplemental additions on the regular sigmoid loss (1 for each layer in my model). Can someone help my to entangle this ball of surprising results? I expect the regular loss, the validation loss and my bag loss to be the same.<\/p>\n","answers":[{"AnswerId":"57886613","CreationDate":"2019-09-11T09:56:44.870","ParentId":null,"OwnerUserId":"11304826","Title":null,"Body":"<p>OK, I finally found the culprit: The L2 regularization, which was turned on, while I thought it was off. The regulirization term is obviously added to the cross-entropy to calculate the effective loss.\nMost of the time, a good night of sleep and a careful inspection your code are more than enough to answer your question. <\/p>\n"}]}
+{"QuestionId":57873624,"AnswerCount":0,"Tags":"<tensorflow><keras><keras-layer><tf.keras><cleverhans>","CreationDate":"2019-09-10T15:06:37.907","AcceptedAnswerId":null,"OwnerUserId":9652636.0,"Title":"AttributeError: 'Conv2D' object has no attribute 'get_config'","Body":"<p>I'm trying to run fgsm attack on mnist dataset. I'm getting 'object has no attribute' error:\nI used: python 2 (in google colab) Keras==2.1.3 tensorflow==1.14.0<\/p>\n\n<p>I also run it with keras==1.2 and keras==2.1.5 , I also chenged get_config into _get_config, However, I also chenged tf-keras to keras but all runs with above solutions failed.\nPlease tell me how to overcome this problem. Thank You very much<\/p>\n\n<pre><code>x_adv_sub = fgsm.generate(images_tensor, **fgsm_par)\nFile \"\/content\/defensegan\/cleverhans\/attacks\/fast_gradient_method.py\", \n  line 50, in generate\n  labels, _nb_classes = self.get_or_guess_labels(x, kwargs)\nFile \"\/content\/defensegan\/cleverhans\/attacks\/attack.py\", line 278, in \n  get_or_guess_labels\n  preds = self.model.get_probs(x)\nFile \"\/content\/defensegan\/cleverhans\/utils_keras.py\", line 196, in \n  get_probs\n  :return: Names of all the layers kept by Keras\nFile \"\/content\/defensegan\/cleverhans\/utils_keras.py\", line 130, in \n  _get_softmax_name\n  def _get_abstract_layer_name(self):\nAttributeError: 'Conv2D' object has no attribute 'get_config'\n<\/code><\/pre>\n\n<p>Unit_keras.py is here: \n   <a href=\"https:\/\/github.com\/tensorflow\/cleverhans\/blob\/master\/cleverhans\/utils_keras.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/cleverhans\/blob\/master\/cleverhans\/utils_keras.py<\/a><\/p>\n\n<p>I also used :\n<a href=\"https:\/\/github.com\/tensorflow\/cleverhans\/blob\/6b065c4507f50e2545c009f75e82678f92353163\/cleverhans\/utils_keras.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/cleverhans\/blob\/6b065c4507f50e2545c009f75e82678f92353163\/cleverhans\/utils_keras.py<\/a><\/p>\n\n<p>but I cant To fix the error.\nCan anybody solve that please? thanks a lot.<\/p>\n","answers":[]}
+{"QuestionId":57873839,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-10T15:20:34.383","AcceptedAnswerId":null,"OwnerUserId":12047637.0,"Title":"How can I get tensorflow to work on eclipse?","Body":"<p>I am a student in high school trying to get tensor flow to work but when asking my teacher for assistance I get yelled at and told to just try harder. I have no clue what to do.<\/p>\n\n<p>I want to use tensor flow on eclipse using python 3.7 and tensor flow 0.12.0<\/p>\n\n<p>Tensorflow is in the library and installed through pip, but when trying to import it, it throws an error.<\/p>\n\n<p>import tensorflow as tf<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File\n  \"\\cslab-server\\home\\CollinsC\\Application\n  Data\\Python\\Python37\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\",\n  line 18, in swig_import_helper\n      fp, pathname, description = imp.find_module('_pywrap_tensorflow', [dirname(<strong>file<\/strong>)])   File \"C:\\Program Files\n  (x86)\\Python3.7-32\\lib\\imp.py\", line 296, in find_module\n      raise ImportError(_ERR_MSG.format(name), name=name) ImportError: No module named '_pywrap_tensorflow'<\/p>\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 \"\\cslab-server\\home\\CollinsC\\Application Data\\Python\\Python37\\site-packages\\tensorflow\\python__init__.py\", line 54, in \n    from tensorflow.python import pywrap_tensorflow\n  File \"\\cslab-server\\home\\CollinsC\\Application Data\\Python\\Python37\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 28, in \n    _pywrap_tensorflow = swig_import_helper()\n  File \"\\cslab-server\\home\\CollinsC\\Application Data\\Python\\Python37\\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>During handling of the above exception, another exception occurred:<\/p>\n\n<p>Traceback (most recent call last):\n  File \"\\CSLAB-SERVER\\home\\CollinsC\\My Documents\\ISMWorkspace\\ISMMMM\\src\\tensorflowtesting.py\", line 6, in \n    import tensorflow as tf\n  File \"\\cslab-server\\home\\CollinsC\\Application Data\\Python\\Python37\\site-packages\\tensorflow__init__.py\", line 24, in \n    from tensorflow.python import *\n  File \"\\cslab-server\\home\\CollinsC\\Application Data\\Python\\Python37\\site-packages\\tensorflow\\python__init__.py\", line 60, in \n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"\\cslab-server\\home\\CollinsC\\Application Data\\Python\\Python37\\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:\\Program Files (x86)\\Python3.7-32\\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 \"\\cslab-server\\home\\CollinsC\\Application Data\\Python\\Python37\\site-packages\\tensorflow\\python__init__.py\", line 54, in \n    from tensorflow.python import pywrap_tensorflow\n  File \"\\cslab-server\\home\\CollinsC\\Application Data\\Python\\Python37\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 28, in \n    _pywrap_tensorflow = swig_import_helper()\n  File \"\\cslab-server\\home\\CollinsC\\Application Data\\Python\\Python37\\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":[]}
+{"QuestionId":57874015,"AnswerCount":0,"Tags":"<keras><keras-layer>","CreationDate":"2019-09-10T15:32:19.027","AcceptedAnswerId":null,"OwnerUserId":11962877.0,"Title":"Inconsistent dimensioning of input\/output data for layers","Body":"<p>I'm learning Keras, so I made some simple dummy code, and cannot understand how to dimension dense layers to match dimensions of input and output data.<\/p>\n\n<p>My input data is a length 10 array, and the output is length 5:<\/p>\n\n<pre><code>np.array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10]),\nnp.array([1, 2, 3, 4, 5]\n<\/code><\/pre>\n\n<p>When I configure the input layer as <\/p>\n\n<pre><code>input_shape=(10,)\n<\/code><\/pre>\n\n<p>with this line:<\/p>\n\n<pre><code>    model.add(Dense(units=10, input_shape=(10,),  activation='relu'))\n<\/code><\/pre>\n\n<p>I get this error:<\/p>\n\n<blockquote>\n  <p>Error when checking input: expected dense_1_input to have shape (10,) but <strong>got array with shape (1,)<\/strong><\/p>\n<\/blockquote>\n\n<p>If I want to interpret input as 2 arrays of length 5<\/p>\n\n<pre><code>input_shape=(2,1)\n<\/code><\/pre>\n\n<p>I get this error:<\/p>\n\n<blockquote>\n  <p>Error when checking input: expected dense_1_input to have 3 dimensions, but <strong>got array with shape (10, 1)<\/strong><\/p>\n<\/blockquote>\n\n<p>Note that before, the error statement it declared that the input array has shape <code>(1,)<\/code>, but now it says that the array has shape <code>(10,)<\/code><\/p>\n\n<p><strong>How come that the input shape is recognized as different shape, when I made no change to the input data?<\/strong><\/p>\n\n<p><code>(1,)<\/code> implies that input_shape should describe the dimension of input data, but the second error implies that <code>(10,)<\/code> describes the length of the data.<\/p>\n\n<p>I read the documentation here: <a href=\"https:\/\/keras.io\/layers\/core\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/core\/<\/a>, but the important parameter <code>input_shape<\/code> is undocumented <\/p>\n\n<pre><code>import sys\nimport numpy as np\n\ndef DataGenerator():\n    while True:\n        yield (np.array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10]),\n               np.array([1, 2, 3, 4, 5]) )\n\ndef main():\n    import tensorflow as tf\n    import keras as K\n    from keras.models import Sequential  \n\n\n    model = Sequential()\n\n    from keras.layers import Dense\n    #imput layer size 10\n    model.add(Dense(units=10, input_shape=(1,),  activation='relu'))\n    #Output layer size 5\n    model.add(Dense(units=5, activation='sigmoid',))\n\n    model.compile( optimizer='adam' , loss='mean_squared_error')\n\n    model.summary()\n\n    model.fit_generator(DataGenerator(),\n                        steps_per_epoch=10000, epochs=10)\n    Print (\"Ok\")\n\n\nif __name__ == \"__main__\":\n    main()\n<\/code><\/pre>\n\n<p>The code seems to work with<\/p>\n\n<pre><code>input_shape=(1,)\n<\/code><\/pre>\n\n<p>but then it raises this error at the last layer:<\/p>\n\n<blockquote>\n  <p>Error when checking target: expected dense_2 to have shape (5,) but got array with shape (1,)<\/p>\n<\/blockquote>\n\n<p>I cannot find documentation on how to describe the output layer to match the data. I thought that since the last layer produces 5 numbers, it would be automatically checked against the length 5 array feed.<\/p>\n\n<p><strong>Where is it documented or explained?<\/strong><\/p>\n","answers":[]}
+{"QuestionId":57874409,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><batching>","CreationDate":"2019-09-10T15:57:02.147","AcceptedAnswerId":null,"OwnerUserId":8349044.0,"Title":"tf.data.Dataset.batch() creates infinite batches","Body":"<p>I am quite a beginner in tensorflow. I was trying to implement the variational-autoencoder using tensorflow but ran into a strange problem. I am using MNIST handwritten digit dataset which has 60000 samples in training set.I am writing this code for loading the dataset:<\/p>\n\n<pre><code>def load_data(path=\"mnist.npz\",return_labels=True):\n    (train_data,train_labels),(test_data,test_labels) = tf.keras.datasets.mnist.load_data(path=path)\n    if return_labels:\n        return train_data,train_labels,test_data,test_labels\n    return train_data,test_data\n<\/code><\/pre>\n\n<p>And I am creating batches like this:<\/p>\n\n<pre><code>(train_data,test_data) = load_data(return_labels=False)\ntrain_data = train_data.reshape(train_data.shape[0],28,28,1).astype('float32')\ntest_data = test_data.reshape(test_data.shape[0],28,28,1).astype('float32')\n\nBATCH_SIZE = 100\n\ntrain_dataset = tf.data.Dataset.from_tensor_slices(train_data).batch(BATCH_SIZE,drop_remainder=True)\ntest_dataset = tf.data.Dataset.from_tensor_slices(test_data).batch(BATCH_SIZE,drop_remainder=True)\n<\/code><\/pre>\n\n<p>So with batch size 100, it should create 600 batches.But, to my amazement, when I am iterating like this:<\/p>\n\n<pre><code>for train_x in train_dataset:\n    pass #just for example\n<\/code><\/pre>\n\n<p>It is running into an infinite loop. Can someone point out what mistake I am doing here?<\/p>\n","answers":[]}
+{"QuestionId":57874436,"AnswerCount":2,"Tags":"<python><tensorflow><keras><lstm>","CreationDate":"2019-09-10T15:58:29.087","AcceptedAnswerId":57973957.0,"OwnerUserId":12047797.0,"Title":"Tensorflow Data Adapter Error: ValueError: Failed to find data adapter that can handle input","Body":"<p>While running a sentdex tutorial script of a cryptocurrency RNN, link here<\/p>\n\n<p><a href=\"https:\/\/www.youtube.com\/watch?v=yWkpRdpOiPY&amp;list=PLQVvvaa0QuDfhTox0AjmQ6tvTgMBZBEXN&amp;index=11\" rel=\"nofollow noreferrer\">YouTube Tutorial: Cryptocurrency-predicting RNN Model<\/a>,<\/p>\n\n<p>but have encountered an error when attempting to train the model. My tensorflow version is 2.0.0 and I'm running python 3.6. When attempting to train the model I receive the following error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>File \"C:\\python36-64\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training.py\", line 734, in fit\n    use_multiprocessing=use_multiprocessing)\n\nFile \"C:\\python36-64\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2.py\", line 224, in fit\n    distribution_strategy=strategy)\n\nFile \"C:\\python36-64\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2.py\", line 497, in _process_training_inputs\n    adapter_cls = data_adapter.select_data_adapter(x, y)\n\nFile \"C:\\python36-64\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\data_adapter.py\", line 628, in select_data_adapter\n    _type_name(x), _type_name(y)))\n\nValueError: Failed to find data adapter that can handle input: &lt;class 'numpy.ndarray'&gt;, (&lt;class 'list'&gt; containing values of types {\"&lt;class 'numpy.float64'&gt;\"})\n<\/code><\/pre>\n\n<p>Any advice would be greatly appreciated!<\/p>\n","answers":[{"AnswerId":"58738823","CreationDate":"2019-11-06T21:34:17.933","ParentId":null,"OwnerUserId":"10843785","Title":null,"Body":"<p>yes add the code before model.fit() function\n<code>train_y= np.asarray(train_y)<\/code>\n<code>validation_y= np.asarray(validation_y)<\/code><\/p>\n"},{"AnswerId":"57973957","CreationDate":"2019-09-17T12:12:30.463","ParentId":null,"OwnerUserId":"12079492","Title":null,"Body":"<p>Have you checked whether your training\/testing data and training\/testing labels are all numpy arrays? It might be that you're mixing numpy arrays with lists. <\/p>\n"}]}
+{"QuestionId":57874545,"AnswerCount":0,"Tags":"<machine-learning><keras><deep-learning><conv-neural-network>","CreationDate":"2019-09-10T16:07:38.577","AcceptedAnswerId":null,"OwnerUserId":11709128.0,"Title":"How to explain the predict failure with big CNN on small dataset","Body":"<p>I have a dataset of ~100 images of \"positive\" examples and ~2500 images of \"negative\" examples (a \"One vs. All\" classification, with <code>binary_crossentropy<\/code> loss in Keras). Images are of size ~500x700 on average, but as they are not all the same, I use <a href=\"https:\/\/github.com\/yhenon\/keras-spp\" rel=\"nofollow noreferrer\"><code>SpatialPyramidPooling<\/code><\/a> layer before the dense layer.<\/p>\n\n<p>As I don't have many examples of each class, I use Keras <a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"nofollow noreferrer\"><code>ImageDataGenerator<\/code><\/a> to generate more samples (5000 of each).<\/p>\n\n<p>I have already trained a small CNN and it worked well (~95% accuracy), but I thought that with a bigger network I could have better results.<\/p>\n\n<p>I tried a bigger CNN (inspired from VGG16 structure) but many outputs are <em>exactly<\/em> the same when predicted during the Test part of the training process (i.e. all classified examples from Test dataset are probability 0 or 0.34167683 using <code>predict_on_batch<\/code> with a batch of size 1):<\/p>\n\n<pre><code>[[0.34167683]]\n[[0.]]\n[[0.]]\n[[0.34167683]]\n[[0.34167683]]\n[[0.34167683]]\n[[0.]]\n[[0.]]\n[[0.]]\n[[0.34167683]]\n[[0.34167683]]\n[[0.34167683]]\n[[0.]]\n[[0.34167683]]\n[[0.34167683]]\n...\n<\/code><\/pre>\n\n<p>The number of epoch I train on doesn't change that behaviour: I get it right after the first epoch and it goes on for all others.<\/p>\n\n<p>Because the image size is somehow big, I use a batch of size 4 because my GPU (NVIDIA RTX 2080 8 GB RAM) cannot accept more (I get OOM for deep networks).<\/p>\n\n<p>I tried with and without <code>BatchNormalization<\/code> after each <code>Conv2D<\/code> but the problem still persists.<\/p>\n\n<p>Here is the small CNN:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def build_one_vs_all():\n    model = Sequential()\n\n    model.add(Conv2D(64, (3, 3), padding='same', input_shape=(None, None, 1)))\n    model.add(Activation('relu'))        \n    model.add(MaxPooling2D(pool_size=(2, 2)))\n\n    model.add(Conv2D(128, (3, 3), padding='same'))\n    model.add(Activation('relu'))\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n\n    model.add(Conv2D(256, (3, 3), padding='same'))\n    model.add(Activation('relu'))\n\n    model.add(SpatialPyramidPooling([1, 2, 4, 8]))       \n\n    model.add(Dense(1024))           \n    model.add(Activation('relu'))\n    model.add(Dense(1024))           \n    model.add(Activation('relu'))\n\n    model.add(Dense(1, activation = 'sigmoid'))\n    model.summary()\n\n    img = Input(shape=(None, None, 1))\n    validity = model(img)\n\n    return Model(img, validity)\n<\/code><\/pre>\n\n<p>And here is the big one: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def build_one_vs_all():\n    model = Sequential()\n\n    model.add(Conv2D(64, (3, 3), padding='same', input_shape=(None, None, 1)))\n    model.add(Activation('relu'))        \n    model.add(Conv2D(64, (3, 3), padding='same'))\n    model.add(Activation('relu'))\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n\n    model.add(Conv2D(128, (3, 3), padding='same'))\n    model.add(Activation('relu'))\n    model.add(Conv2D(128, (3, 3), padding='same'))\n    model.add(Activation('relu'))\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n\n    model.add(Conv2D(256, (3, 3), padding='same'))\n    model.add(Activation('relu'))\n    model.add(Conv2D(256, (3, 3), padding='same'))\n    model.add(Activation('relu'))\n    model.add(Conv2D(256, (3, 3), padding='same'))\n    model.add(Activation('relu'))\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n\n    model.add(Conv2D(512, (3, 3), padding='same'))\n    model.add(Activation('relu'))\n    model.add(Conv2D(512, (3, 3), padding='same'))\n    model.add(Activation('relu'))\n    model.add(Conv2D(512, (3, 3), padding='same'))\n    model.add(Activation('relu'))\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n\n    model.add(Conv2D(512, (3, 3), padding='same'))\n    model.add(Activation('relu'))\n    model.add(Conv2D(512, (3, 3), padding='same'))\n    model.add(Activation('relu'))\n    model.add(Conv2D(512, (3, 3), padding='same'))\n    model.add(Activation('relu'))\n\n    model.add(SpatialPyramidPooling([1, 2, 4, 8]))       \n    model.add(Dense(1024))           \n    model.add(Activation('relu'))\n    model.add(Dense(1024))           \n    model.add(Activation('relu'))\n\n    model.add(Dense(1, activation = 'sigmoid'))\n    model.summary()\n\n    img = Input(shape=(None, None, 1))\n    validity = model(img)\n\n    return Model(img, validity)\n<\/code><\/pre>\n\n<p>Why is that happening? How can I complexify the model to achieve stable classification with a small number of input images?<\/p>\n","answers":[]}
+{"QuestionId":57874584,"AnswerCount":1,"Tags":"<tensorflow><masking><bert-language-model>","CreationDate":"2019-09-10T16:09:53.087","AcceptedAnswerId":null,"OwnerUserId":10611614.0,"Title":"Masking in BERT","Body":"<p>In the original paper of BERT it is said:<\/p>\n\n<blockquote>\n  <p>Note that the purpose of the masking strategies\n  is to reduce the mismatch between pre-training\n  and fine-tuning, as the [MASK] symbol never appears during the fine-tuning stage.<\/p>\n<\/blockquote>\n\n<p>Let's consider a sentence \"I am a Liverpool fan\" which with 40% masking will be transformed into \"I [MASK] a [MASK] fan\". When predicting the first [MASK], will it be predicted by a phrase \"I [MASK] a fan\", excluding the second [MASK] or \"I [MASK] a [MASK] fan\", by a full sentence?<\/p>\n\n<p>And what is the purpose of replacing 10% of masked tokens with themselves? Does it mean they will not be predicted? Or we will predict them, having themselves in the context (like predicting the first [MASK] by \"I am a [MASK] fan\"?<\/p>\n\n<p>Will be very grateful for any help!<\/p>\n","answers":[{"AnswerId":"57953534","CreationDate":"2019-09-16T09:02:53.557","ParentId":null,"OwnerUserId":"8893595","Title":null,"Body":"<p>If you look at the paper clearly, and also check out the paper on XLNET which explains the drawback of BERT, it implies that, given a sentence \"I [MASK] a [MASK] fan\", it tries to predict both the [MASK] tokens using the sentence \"I [MASK] a [MASK] fan\", and not \"I [MASK] a Liverpool fan\".<\/p>\n\n<p>Due to this, it masks only 15% of the tokens. The idea behind retaining the 10% of the mask tokens as it is, implies that they will not be predicted.<\/p>\n"}]}
+{"QuestionId":57874617,"AnswerCount":1,"Tags":"<keras><h5py><pre-trained-model><transfer-learning><vgg-net>","CreationDate":"2019-09-10T16:11:37.817","AcceptedAnswerId":57875858.0,"OwnerUserId":3802483.0,"Title":"keras copying VGG16 pretrained weights layer by layer","Body":"<p>I want to copy some of the VGG16 layer weights layer by layer to another small network with alike layers, but I get an error that says:<\/p>\n\n<pre><code>  File \"\/home\/d\/Desktop\/s\/copyweights.py\", line 78, in &lt;module&gt;\n    list(f[\"model_weights\"].keys())\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 \"\/usr\/local\/lib\/python3.5\/dist-packages\/h5py\/_hl\/group.py\", line 262, in __getitem__\n    oid = h5o.open(self.id, self._e(name), lapl=self._lapl)\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\/h5o.pyx\", line 190, in h5py.h5o.open\nKeyError: \"Unable to open object (object 'model_weights' doesn't exist)\"\n<\/code><\/pre>\n\n<p>The file is definitely in the path, I downloaded the file again just to make sure it is not corrupted (and it doesn't cause an error when I use model.load_weights in general - I also have HDF5 installed<\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><code>from keras import applications\nfrom keras.preprocessing.image import ImageDataGenerator\nfrom keras import optimizers\nfrom keras.models import Sequential, Model\nfrom keras.layers import Dropout, Flatten, Dense, GlobalAveragePooling2D\nfrom keras import backend as k\nfrom keras.callbacks import ModelCheckpoint, LearningRateScheduler, TensorBoard, EarlyStopping\nfrom keras import layers\nfrom keras import models\nfrom keras import optimizers\nfrom keras.layers import Dropout\nfrom keras.regularizers import l2\nfrom keras.callbacks import EarlyStopping, ModelCheckpoint, ReduceLROnPlateau\nimport matplotlib.pyplot as plt\nfrom keras.preprocessing.image import ImageDataGenerator\nimport os\n\nepochs = 50\ncallbacks = []\n#schedule = None\ndecay = 0.0\n\nearlyStopping = EarlyStopping(monitor='val_loss', patience=10, verbose=0, mode='min')\nmcp_save = ModelCheckpoint('.mdl_wts.hdf5', save_best_only=True, monitor='val_loss', mode='min')\nreduce_lr_loss = ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=3, verbose=1, epsilon=1e-5, mode='min')\n\nbase_model = models.Sequential()\nbase_model.add(layers.Conv2D(64, (3, 3), activation='relu', name='block1_conv1',  input_shape=(256, 256, 3)))\nbase_model.add(layers.Conv2D(64, (3, 3), activation='relu', name='block1_conv2'))\nbase_model.add(layers.MaxPooling2D((2, 2)))\n#model.add(Dropout(0.2))\nbase_model.add(layers.Conv2D(128, (3, 3), activation='relu', name='block2_conv1'))\nbase_model.add(layers.Conv2D(128, (3, 3), activation='relu', name='block2_conv2'))\nbase_model.add(layers.MaxPooling2D((2, 2), name='block2_pool'))\n#model.add(Dropout(0.2))\n\nbase_model.summary()\n\"\"\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_1 (InputLayer)         (None, 256, 256, 3)       0         \n_________________________________________________________________\nblock1_conv1 (Conv2D)        (None, 256, 256, 64)      1792      \n_________________________________________________________________\nblock1_conv2 (Conv2D)        (None, 256, 256, 64)      36928     \n_________________________________________________________________\nblock1_pool (MaxPooling2D)   (None, 128, 128, 64)      0         \n_________________________________________________________________\nblock2_conv1 (Conv2D)        (None, 128, 128, 128)     73856     \n_________________________________________________________________\nblock2_conv2 (Conv2D)        (None, 128, 128, 128)     147584    \n_________________________________________________________________\nblock2_pool (MaxPooling2D)   (None, 64, 64, 128)       0         \n=================================================================\nTotal params: 260,160.0\nTrainable params: 260,160.0\nNon-trainable params: 0.0\n\"\"\"\n\nlayer_dict = dict([(layer.name, layer) for layer in base_model.layers])\n[layer.name for layer in base_model.layers]\n\"\"\"\n['input_1',\n 'block1_conv1',\n 'block1_conv2',\n 'block1_pool',\n 'block2_conv1',\n 'block2_conv2',\n 'block2_pool']\n\"\"\"\n\nimport h5py\nweights_path = '\/home\/d\/Desktop\/s\/vgg16_weights_new.h5' # ('https:\/\/github.com\/fchollet\/deep-learning-models\/releases\/download\/v0.1\/vgg19_weights_tf_dim_ordering_tf_kernels.h5)\nf = h5py.File(weights_path)\n\nlist(f[\"model_weights\"].keys())\n\"\"\"\n['block1_conv1',\n 'block1_conv2',\n 'block1_pool',\n 'block2_conv1',\n 'block2_conv2',\n 'block2_pool',\n 'block3_conv1',\n 'block3_conv2',\n 'block3_conv3',\n 'block3_conv4',\n 'block3_pool',\n 'block4_conv1',\n 'block4_conv2',\n 'block4_conv3',\n 'block4_conv4',\n 'block4_pool',\n 'block5_conv1',\n 'block5_conv2',\n 'block5_conv3',\n 'block5_conv4',\n 'block5_pool',\n 'dense_1',\n 'dense_2',\n 'dense_3',\n 'dropout_1',\n 'global_average_pooling2d_1',\n 'input_1']\n\"\"\"\n\n# list all the layer names which are in the model.\nlayer_names = [layer.name for layer in base_model.layers]\n\n\n\"\"\"\n# Here we are extracting model_weights for each and every layer from the .h5 file\n&gt;&gt;&gt; f[\"model_weights\"][\"block1_conv1\"].attrs[\"weight_names\"]\narray([b'block1_conv1\/kernel:0', b'block1_conv1\/bias:0'], \n      dtype='|S21')\n# we are assiging this array to weight_names below \n&gt;&gt;&gt; f[\"model_weights\"][\"block1_conv1\"][\"block1_conv1\/kernel:0]\n&lt;HDF5 dataset \"kernel:0\": shape (3, 3, 3, 64), type \"&lt;f4\"&gt;\n# The list comprehension (weights) stores these two weights and bias of both the layers \n&gt;&gt;&gt;layer_names.index(\"block1_conv1\")\n1\n&gt;&gt;&gt; model.layers[1].set_weights(weights)\n# This will set the weights for that particular layer.\nWith a for loop we can set_weights for the entire network.\n\"\"\"\nfor i in layer_dict.keys():\n    weight_names = f[\"model_weights\"][i].attrs[\"weight_names\"]\n    weights = [f[\"model_weights\"][i][j] for j in weight_names]\n    index = layer_names.index(i)\n    base_model.layers[index].set_weights(weights)\n\nbase_model.add(layers.Flatten())\nbase_model.add(layers.Dropout(0.5))  #Dropout for regularization\nbase_model.add(layers.Dense(256, activation='relu'))\nbase_model.add(layers.Dense(1, activation='sigmoid'))  #Sigmoid function at the end because we have just two classes\n\n\n# compile the model with a SGD\/momentum optimizer\n# and a very slow learning rate.\nbase_model.compile(loss='binary_crossentropy',\n                   optimizer=optimizers.Adam(lr=1e-4, decay=decay),\n                   metrics=['accuracy'])\n\n\n\nos.environ[\"CUDA_VISIBLE_DEVICES\"]=\"0\"\n\ntrain_dir = '\/home\/d\/Desktop\/s\/data\/train'\neval_dir = '\/home\/d\/Desktop\/s\/data\/eval'\ntest_dir = '\/home\/d\/Desktop\/s\/data\/test'\n\n\n# create a data generator\ntrain_datagen = ImageDataGenerator(rescale=1.\/255,   #Scale the image between 0 and 1\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\nval_datagen = ImageDataGenerator(rescale=1.\/255)  #We do not augment validation data. we only perform rescale\n\ntest_datagen = ImageDataGenerator(rescale=1.\/255)  #We do not augment validation data. we only perform rescale\n\n# load and iterate training dataset\ntrain_generator = train_datagen.flow_from_directory(train_dir,  target_size=(224,224),class_mode='binary', batch_size=16, shuffle='True', seed=42)\n# load and iterate validation dataset\nval_generator = val_datagen.flow_from_directory(eval_dir,  target_size=(224,224),class_mode='binary', batch_size=16, shuffle='True', seed=42)\n# load and iterate test dataset\ntest_generator = test_datagen.flow_from_directory(test_dir,  target_size=(224,224), class_mode=None, batch_size=1, shuffle='False', seed=42)\n\n#The training part\n#We train for 64 epochs with about 100 steps per epoch\nhistory = base_model.fit_generator(train_generator,\n                              steps_per_epoch=train_generator.n \/\/ train_generator.batch_size,\n                              epochs=epochs,\n                              validation_data=val_generator,\n                              validation_steps=val_generator.n \/\/ val_generator.batch_size,\n                              callbacks=[earlyStopping, mcp_save, reduce_lr_loss])\n\n#Save the model\n#base_model.save_weights('\/home\/d\/Desktop\/s\/base_model_weights.h5')\n#base_model.save('\/home\/d\/Desktop\/s\/base_model_keras.h5')\n\n#lets plot the train and val curve\n#get the details form the history object\nacc = history.history['acc']\nval_acc = history.history['val_acc']\nloss = history.history['loss']\nval_loss = history.history['val_loss']\n\nepochs = range(1, len(acc) + 1)\n\n#Train and validation accuracy\nplt.plot(epochs, acc, 'b', label='Training accuracy')\nplt.plot(epochs, val_acc, 'r', label='Validation accuracy')\nplt.title('Training and Validation accurarcy')\nplt.legend()\n\nplt.figure()\n#Train and validation loss\nplt.plot(epochs, loss, 'b', label='Training loss')\nplt.plot(epochs, val_loss, 'r', label='Validation loss')\nplt.title('Training and Validation loss')\nplt.legend()\n\nplt.show()\n<\/code><\/pre>\n","answers":[{"AnswerId":"57875858","CreationDate":"2019-09-10T17:39:06.567","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Take the <code>VGG16<\/code> model directly:<\/p>\n\n<pre><code>from keras.applications import VGG16\nvgg = VGG16(choose parameters)\n\nfor layer in vgg.layers:\n    layer_weights_list = layer.get_weights()\n<\/code><\/pre>\n"}]}
+{"QuestionId":57874954,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-10T16:32:33.853","AcceptedAnswerId":null,"OwnerUserId":10453757.0,"Title":"Training with pre-trained weights","Body":"<p>I want to use a set of pre-trained weights to train my model for MNIST classification. More specifically, I train my model on one dataset. I want to use the final weights as the starting weights to train the model on a different dataset. To do this, I use <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>intial_weights = model1.get_weights()\nmodel2 = create_model()\nmodel2.set_weights(initial_weights)\nmodel2.fit(x=x_train59,y=y_train59, epochs=20,callbacks = [cp_callback2])\n<\/code><\/pre>\n\n<p>My question is that whether model.fit() will ignore the initial weights set using model2.set_weights() or not. And if it does ignore, is there a way to make sure that model2.fit() uses the weights obtained previously. Also, is there a way to visualize the starting weights before model.fit() starts training. Thanks much in advance!<\/p>\n","answers":[{"AnswerId":"57875806","CreationDate":"2019-09-10T17:35:56.053","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>When you do <code>model2.set_weights<\/code>, you changed the weights of <code>model2<\/code>. That's all. <\/p>\n\n<p>You can see the weights the same way: <code>w2 = model2.get_weights()<\/code>. Then print <code>w2<\/code> in a convenient way. <\/p>\n"}]}
+{"QuestionId":57875066,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-datasets>","CreationDate":"2019-09-10T16:40:55.407","AcceptedAnswerId":null,"OwnerUserId":3956943.0,"Title":"How to import binary text data into tensorflow","Body":"<p>I am new to Tensorflow and I can't figure out how to properly import data. All of the tutorials include public data sets. I have files that I want to import but I'm not quite sure how. They are all binary text in the format {input}|{expected result}. For example:<\/p>\n\n<p>01010100|0110111011000101<\/p>\n\n<p>How would I import these files for training?<\/p>\n","answers":[]}
+{"QuestionId":57875101,"AnswerCount":0,"Tags":"<python><linux><tensorflow><redhat>","CreationDate":"2019-09-10T16:43:13.350","AcceptedAnswerId":null,"OwnerUserId":11861082.0,"Title":"GLIBCXX_3.4.17 and GLIBC_2.16 not found on redhat 6.7","Body":"<p>I have an account (not superuser) on Redhat 6.7 and it has Python 3.4.3 on it. The computer is offline. I installed TensorFlow 1.12 on it. When I want to import TensorFlow, I get the following error: <\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/home\/r.jack\/.local\/lib\/python3.4\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"\/home\/r.jack\/.local\/lib\/python3.4\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"\/home\/r.jack\/.local\/lib\/python3.4\/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 \"\/appl\/open_tools\/python\/3.4.3\/lib\/python3.4\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\nImportError: \/usr\/lib64\/libstdc++.so.6: version `GLIBCXX_3.4.17' not found (required by \/home\/r.jack\/.local\/lib\/python3.4\/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 \"\/home\/r.jack\/.local\/lib\/python3.4\/site-packages\/tensorflow\/__init__.py\", line 24, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"\/home\/r.jack\/.local\/lib\/python3.4\/site-packages\/tensorflow\/python\/__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"\/home\/r.jack\/.local\/lib\/python3.4\/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\/r.jack\/.local\/lib\/python3.4\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"\/home\/r.jack\/.local\/lib\/python3.4\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"\/home\/r.jack\/.local\/lib\/python3.4\/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 \"\/appl\/open_tools\/python\/3.4.3\/lib\/python3.4\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\nImportError: \/usr\/lib64\/libstdc++.so.6: version `GLIBCXX_3.4.17' not found (required by \/home\/r.jack\/.local\/lib\/python3.4\/site-packages\/tensorflow\/python\/_pywrap_tensorflow_internal.so)\n\nFailed to load the native TensorFlow runtime.\n\nSee https:\/\/www.tensorflow.org\/install\/errors\n<\/code><\/pre>\n\n<p>I realized it might be because of gcc. So, I looked at the gcc versions available:<\/p>\n\n<pre><code>gcc\/4.8.1       gcc\/4.9.0       gcc\/5.2.0       gcc\/6.1.0\ngcc\/4.8.5_rhel6 gcc\/4.9.3       gcc\/5.3.0       gcc\/7.3.0\n<\/code><\/pre>\n\n<p>I loaded gcc\/4.8.1. Now, I get a different error:<\/p>\n\n<pre><code> Traceback (most recent call last):\n  File \"\/home\/r.jack\/.local\/lib\/python3.4\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"\/home\/r.jack\/.local\/lib\/python3.4\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"\/home\/r.jack\/.local\/lib\/python3.4\/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 \"\/appl\/open_tools\/python\/3.4.3\/lib\/python3.4\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\nImportError: \/lib64\/libc.so.6: version `GLIBC_2.16' not found (required by \/home\/r.jack\/.local\/lib\/python3.4\/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 \"\/home\/r.jack\/.local\/lib\/python3.4\/site-packages\/tensorflow\/__init__.py\", line 24, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"\/home\/r.jack\/.local\/lib\/python3.4\/site-packages\/tensorflow\/python\/__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"\/home\/r.jack\/.local\/lib\/python3.4\/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\/r.jack\/.local\/lib\/python3.4\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"\/home\/r.jack\/.local\/lib\/python3.4\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"\/home\/r.jack\/.local\/lib\/python3.4\/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 \"\/appl\/open_tools\/python\/3.4.3\/lib\/python3.4\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\nImportError: \/lib64\/libc.so.6: version `GLIBC_2.16' not found (required by \/home\/r.jack\/.local\/lib\/python3.4\/site-packages\/tensorflow\/python\/_pywrap_tensorflow_internal.so)\n\n\nFailed to load the native TensorFlow runtime.\n\nSee https:\/\/www.tensorflow.org\/install\/errors\n<\/code><\/pre>\n\n<p>How can I resolve this? One solution is that the system admin upgrades the RedHat from 6.7 to 7. But, they don't do that. <\/p>\n\n<p>Do I have to try an earlier version of TensorFlow? Or, an earlier version of python? (They have python 2, but I prefer not to use Python 2)<\/p>\n\n<p><strong>What I tried after I got comments:<\/strong> I downloaded a new libc (libc6-2.16.90-3.x86_64.rpm) from <a href=\"http:\/\/rpmfind.net\/linux\/rpm2html\/search.php?query=libc.so.6\" rel=\"nofollow noreferrer\">http:\/\/rpmfind.net\/linux\/rpm2html\/search.php?query=libc.so.6<\/a> and then used following commands to set the library directory and preload to the load the new libc from my user's home directory:<\/p>\n\n<pre><code>export LD_LIBRARY_PATH=path_to_libc_so_6_dir:${LD_LIBRARY_PATH}.\nexport LD_PRELOAD=libc.so.6\n<\/code><\/pre>\n\n<p>Then, I ran python, but I got the same error. <\/p>\n","answers":[]}
+{"QuestionId":57875184,"AnswerCount":0,"Tags":"<keras><conv-neural-network><pre-trained-model><transfer-learning><vgg-net>","CreationDate":"2019-09-10T16:48:52.123","AcceptedAnswerId":null,"OwnerUserId":3802483.0,"Title":"Failed to create session while copying weights","Body":"<p>I want to copy some of the VGG16 layer weights to another small network with alike (but fewer) layers that shares the same name, but I get a peculiar error that says:<\/p>\n\n<pre><code>2019-09-10 18:43:06.201112: E tensorflow\/core\/common_runtime\/direct_session.cc:158] Internal: CUDA runtime implicit initialization on GPU:0 failed. Status: out of memory\n    Traceback (most recent call last):\n      File \"\/home\/d\/Desktop\/s\/copyweights.py\", line 189, in &lt;module&gt;\n        callbacks=[earlyStopping, mcp_save, reduce_lr_loss])\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 1418, in fit_generator\n        initial_epoch=initial_epoch)\n      File \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/engine\/training_generator.py\", line 217, in fit_generator\n        class_weight=class_weight)\n      File \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/engine\/training.py\", line 1217, in train_on_batch\n        outputs = self.train_function(ins)\n      File \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/backend\/tensorflow_backend.py\", line 2697, in __call__\n        if hasattr(get_session(), '_make_callable_from_options'):\n      File \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/backend\/tensorflow_backend.py\", line 186, in get_session\n        _SESSION = tf.Session(config=config)\n      File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py\", line 1563, in __init__\n        super(Session, self).__init__(target, graph, config=config)\n      File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py\", line 633, in __init__\n        self._session = tf_session.TF_NewSession(self._graph._c_graph, opts)\n    tensorflow.python.framework.errors_impl.InternalError: Failed to create session.\n\n    Process finished with exit code 1\n<\/code><\/pre>\n\n<p>Code:<\/p>\n\n<pre><code>from keras import applications\nfrom keras.preprocessing.image import ImageDataGenerator\nfrom keras import optimizers\nfrom keras.models import Sequential, Model\nfrom keras.layers import Dropout, Flatten, Dense, GlobalAveragePooling2D\nfrom keras import backend as k\nfrom keras.callbacks import ModelCheckpoint, LearningRateScheduler, TensorBoard, EarlyStopping\nfrom keras import layers\nfrom keras import models\nfrom keras import optimizers\nfrom keras.layers import Dropout\nfrom keras.regularizers import l2\nfrom keras.callbacks import EarlyStopping, ModelCheckpoint, ReduceLROnPlateau\nimport matplotlib.pyplot as plt\nfrom keras.preprocessing.image import ImageDataGenerator\nimport os\n\n\nepochs = 50\ncallbacks = []\n#schedule = None\ndecay = 0.0\n\nearlyStopping = EarlyStopping(monitor='val_loss', patience=10, verbose=0, mode='min')\nmcp_save = ModelCheckpoint('.mdl_wts.hdf5', save_best_only=True, monitor='val_loss', mode='min')\nreduce_lr_loss = ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=3, verbose=1, epsilon=1e-5, mode='min')\n\nbase_model = models.Sequential()\nbase_model.add(layers.Conv2D(64, (3, 3), activation='relu', name='block1_conv1',  input_shape=(224, 224, 3)))\nbase_model.add(layers.Conv2D(64, (3, 3), activation='relu', name='block1_conv2'))\nbase_model.add(layers.MaxPooling2D((2, 2)))\n#model.add(Dropout(0.2))\nbase_model.add(layers.Conv2D(128, (3, 3), activation='relu', name='block2_conv1'))\nbase_model.add(layers.Conv2D(128, (3, 3), activation='relu', name='block2_conv2'))\nbase_model.add(layers.MaxPooling2D((2, 2), name='block2_pool'))\n#model.add(Dropout(0.2))\n\nbase_model.summary()\n\nvgg_weights_path = '\/home\/d\/Desktop\/s\/vgg16_weights.h5'\nbase_model.load_weights(vgg_weights_path, by_name=True)\n\nbase_model.add(layers.Flatten())\nbase_model.add(layers.Dropout(0.5))  #Dropout for regularization\nbase_model.add(layers.Dense(256, activation='relu'))\nbase_model.add(layers.Dense(1, activation='sigmoid'))  #Sigmoid function at the end because we have just two classes\n\n\n# compile the model with a SGD\/momentum optimizer\n# and a very slow learning rate.\nbase_model.compile(loss='binary_crossentropy',\n                   optimizer=optimizers.Adam(lr=1e-4, decay=decay),\n                   metrics=['accuracy'])\n\n\n\nos.environ[\"CUDA_VISIBLE_DEVICES\"]=\"0\"\n\ntrain_dir = '\/home\/d\/Desktop\/s\/data\/train'\neval_dir = '\/home\/d\/Desktop\/s\/data\/eval'\ntest_dir = '\/home\/d\/Desktop\/s\/data\/test'\n\n\n# create a data generator\ntrain_datagen = ImageDataGenerator(rescale=1.\/255,   #Scale the image between 0 and 1\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\nval_datagen = ImageDataGenerator(rescale=1.\/255)  #We do not augment validation data. we only perform rescale\n\ntest_datagen = ImageDataGenerator(rescale=1.\/255)  #We do not augment validation data. we only perform rescale\n\n# load and iterate training dataset\ntrain_generator = train_datagen.flow_from_directory(train_dir,  target_size=(224,224),class_mode='binary', batch_size=8, shuffle='True', seed=42)\n# load and iterate validation dataset\nval_generator = val_datagen.flow_from_directory(eval_dir,  target_size=(224,224),class_mode='binary', batch_size=8, shuffle='True', seed=42)\n# load and iterate test dataset\ntest_generator = test_datagen.flow_from_directory(test_dir,  target_size=(224,224), class_mode=None, batch_size=1, shuffle='False', seed=42)\n\n#The training part\n#We train for 64 epochs with about 100 steps per epoch\nhistory = base_model.fit_generator(train_generator,\n                              steps_per_epoch=train_generator.n \/\/ train_generator.batch_size,\n                              epochs=epochs,\n                              validation_data=val_generator,\n                              validation_steps=val_generator.n \/\/ val_generator.batch_size,\n                              callbacks=[earlyStopping, mcp_save, reduce_lr_loss])\n\n#Save the model\n#base_model.save_weights('\/home\/d\/Desktop\/s\/base_model_weights.h5')\n#base_model.save('\/home\/d\/Desktop\/s\/base_model_keras.h5')\n\n#lets plot the train and val curve\n#get the details form the history object\nacc = history.history['acc']\nval_acc = history.history['val_acc']\nloss = history.history['loss']\nval_loss = history.history['val_loss']\n\nepochs = range(1, len(acc) + 1)\n\n#Train and validation accuracy\nplt.plot(epochs, acc, 'b', label='Training accuracy')\nplt.plot(epochs, val_acc, 'r', label='Validation accuracy')\nplt.title('Training and Validation accurarcy')\nplt.legend()\n\nplt.figure()\n#Train and validation loss\nplt.plot(epochs, loss, 'b', label='Training loss')\nplt.plot(epochs, val_loss, 'r', label='Validation loss')\nplt.title('Training and Validation loss')\nplt.legend()\n\nplt.show()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57875581,"AnswerCount":1,"Tags":"<tensorflow><keras><autoencoder>","CreationDate":"2019-09-10T17:18:58.217","AcceptedAnswerId":57875777.0,"OwnerUserId":2540407.0,"Title":"How to get output from a specific layer in keras.tf, the bottleneck layer in autoencoder?","Body":"<p>I am developing an autoencoder for clustering certain groups of images.<\/p>\n\n<pre><code>input_images-&gt;...-&gt;bottleneck-&gt;...-&gt;output_images\n<\/code><\/pre>\n\n<p>I have calibrated the autoencoder to my satisfaction and saved the model; everything has been developed using keras.tensorflow on python3. <\/p>\n\n<p>The next step is to apply the autoencoder to a ton of images and cluster them according to cosine distance in the bottleneck layer. Oops, I just realized that I don't know the syntax in keras.tf for running the model on a batch up to a specific layer rather than to the output layer. Thus the question:<\/p>\n\n<p>How do I run something like <code>Model.predict_on_batch<\/code> or <code>Model.predict_generator<\/code> up to the certain \"bottleneck\" layer and retrieve the values on that layer rather than the values on the output layer?<\/p>\n","answers":[{"AnswerId":"57875777","CreationDate":"2019-09-10T17:33:47.343","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>You need to define a new model (if you didn't define the encoder and decoder as separate models initially, which is usually the easiest option).    <\/p>\n\n<p>If your model was defined without reusing layers, it's just:<\/p>\n\n<pre><code>inputs = model.input   \noutputs= model.get_layer('bottleneck').output\n\nencoder = Model(inputs, outputs)\n<\/code><\/pre>\n\n<p>Use the <code>encoder<\/code> model as any other model. <\/p>\n"}]}
+{"QuestionId":57875887,"AnswerCount":2,"Tags":"<java><android><fragment><torch><flashlight>","CreationDate":"2019-09-10T17:41:37.853","AcceptedAnswerId":57876146.0,"OwnerUserId":11611503.0,"Title":"Flashlight Fragment Android Studio","Body":"<p>I have created a flashlight fragment for my app and I have implemented this code and turning the flashlight on works, but i cant turn it off.<\/p>\n\n<p>I tried to switch the light on and off from one button or like now from two but it made no difference.<\/p>\n\n<p>When i switch the fragment, the light turns off.<\/p>\n\n<p>I hope anyone knows what to do. \nThx<\/p>\n\n<pre><code>\n\n\npublic class Flashlight extends Fragment implements SurfaceHolder.Callback {\n\n    private View view;\n    Button button_On, button_Off;\n\n    private boolean lightIsOn = false;\n\n\n    Camera camera;\n    android.hardware.Camera.Parameters parameters;\n\n    @Override\n    public void onStart() {\n        super.onStart();\n        SurfaceView preview = (SurfaceView) getView().findViewById(R.id.PREVIEW);\n        SurfaceHolder mHolder = preview.getHolder();\n        mHolder.addCallback(this);\n    }\n\n    @Override\n    public void onPause() {\n        super.onPause();\n        turnOffLight();\n    }\n\n    @Override\n    public View onCreateView(final LayoutInflater inflater, final ViewGroup container,\n                             final Bundle savedInstanceState) {\n\n        view = inflater.inflate(R.layout.fragment_flashlight, container, false);\n\n        button_On = (Button) view.findViewById(R.id.button_On);\n        button_Off = (Button) view.findViewById(R.id.button_Off);\n\n\n        if (ContextCompat.checkSelfPermission(getActivity(), Manifest.permission.CAMERA)\n                != PackageManager.PERMISSION_GRANTED) {\n            requestPermissions(new String[]{Manifest.permission.CAMERA,}, PackageManager.PERMISSION_GRANTED);\n        }\n        if (!getActivity().getApplicationContext().getPackageManager().hasSystemFeature(PackageManager.FEATURE_CAMERA_FLASH)) {\n            Toast.makeText(getActivity(), \"Device has no flashlight\", Toast.LENGTH_SHORT).show();\n        }\n\n\n        button_On.setOnClickListener(new View.OnClickListener() {\n            @Override\n            public void onClick(View v) {\n\n                    turnOnLight();\n\n            }\n        });\n\n        button_Off.setOnClickListener(new View.OnClickListener() {\n            @Override\n            public void onClick(View v) {\n                turnOffLight();\n            }\n        });\n\n        return view;\n    }\n<\/code><\/pre>\n\n<p>Turn On\/Off Light:<\/p>\n\n<pre><code>    private void turnOnLight() {\n        if (!lightIsOn) {\n            if (camera == null || parameters == null) {\n                return;\n            }\n            parameters = camera.getParameters();\n            parameters.setFlashMode(Camera.Parameters.FLASH_MODE_TORCH);\n            camera.setParameters(parameters);\n            camera.startPreview();\n            lightIsOn = true;\n        }\n    }\n\n    private void turnOffLight() {\n\n        if (lightIsOn) {\n            if (camera == null || parameters == null) {\n                return;\n            }\n\n            parameters = camera.getParameters();\n            parameters.setFlashMode(Camera.Parameters.FLASH_MODE_OFF);\n            camera.setParameters(parameters);\n            camera.stopPreview();\n            lightIsOn = false;\n        }\n    }\n\n    @Override\n    public void surfaceChanged(SurfaceHolder holder, int format, int width, int height) {\n\n    }\n\n    @Override\n    public void surfaceDestroyed(SurfaceHolder holder) {\n        if (camera != null) {\n            camera.stopPreview();\n            camera.setPreviewCallback(null);\n            camera.release();\n            camera = null;\n        }\n    }\n\n    @Override\n    public void surfaceCreated(SurfaceHolder holder) {\n        if (camera == null) {\n            camera = Camera.open();\n            parameters = camera.getParameters();\n            try {\n                camera.setPreviewDisplay(holder);\n            } catch (IOException e) {\n                camera.release();\n                camera = null;\n            }\n        }\n    }\n\n}\n<\/code><\/pre>\n\n<p>Manifest:<\/p>\n\n<pre><code>&lt;uses-permission android:name=\"android.permission.CAMERA\" \/&gt;\n    &lt;uses-feature android:name=\"android.hardware.camera\" \/&gt;\n    &lt;uses-permission android:name=\"android.permission.FLASHLIGHT\" \/&gt;\n<\/code><\/pre>\n\n<p>XML:<\/p>\n\n<pre><code>&lt;?xml version=\"1.0\" encoding=\"utf-8\"?&gt;\n&lt;FrameLayout xmlns:android=\"http:\/\/schemas.android.com\/apk\/res\/android\"\n    xmlns:app=\"http:\/\/schemas.android.com\/apk\/res-auto\"\n    xmlns:tools=\"http:\/\/schemas.android.com\/tools\"\n    android:layout_width=\"match_parent\"\n    android:layout_height=\"match_parent\"\n    tools:context=\".Flashlight\"&gt;\n\n    &lt;androidx.constraintlayout.widget.ConstraintLayout\n        android:layout_width=\"match_parent\"\n        android:layout_height=\"match_parent\"&gt;\n\n        &lt;SurfaceView\n            android:id=\"@+id\/PREVIEW\"\n            android:layout_width=\"1dp\"\n            android:layout_height=\"1dp\"\n            android:layout_marginBottom=\"730dp\"\n            app:layout_constraintBottom_toBottomOf=\"parent\"\n            app:layout_constraintEnd_toEndOf=\"parent\"\n            app:layout_constraintStart_toStartOf=\"parent\"\n            app:layout_constraintTop_toTopOf=\"parent\" \/&gt;\n\n        &lt;LinearLayout\n            android:layout_width=\"match_parent\"\n            android:layout_height=\"match_parent\"\n            android:orientation=\"vertical\"&gt;\n\n            &lt;Button\n                android:id=\"@+id\/button_On\"\n                android:layout_width=\"match_parent\"\n                android:layout_height=\"wrap_content\"\n                android:text=\"on\" \/&gt;\n\n            &lt;Button\n                android:id=\"@+id\/button_Off\"\n                android:layout_width=\"match_parent\"\n                android:layout_height=\"wrap_content\"\n                android:text=\"off\" \/&gt;\n\n        &lt;\/LinearLayout&gt;\n\n\n    &lt;\/androidx.constraintlayout.widget.ConstraintLayout&gt;\n&lt;\/FrameLayout&gt;\n<\/code><\/pre>\n","answers":[{"AnswerId":"57877811","CreationDate":"2019-09-10T20:29:00.010","ParentId":null,"OwnerUserId":"11611503","Title":null,"Body":"<p>In which line?<\/p>\n\n<pre><code>private void turnOffLight() {\n\n        if (lightIsOn) {\n            if (camera == null || parameters == null) {\n                return;\n            }\n            camera = Camera.open();\n\n            parameters = camera.getParameters();\n            parameters.setFlashMode(Camera.Parameters.FLASH_MODE_OFF);\n            camera.setParameters(parameters);\n            camera.stopPreview();\n            camera.release(); \/\/ Very Important line\n            lightIsOn = false;\n            button_OnOff.setText(\"On\");\n\n        }\n    }\n<\/code><\/pre>\n"},{"AnswerId":"57876146","CreationDate":"2019-09-10T18:05:42.970","ParentId":null,"OwnerUserId":"6715226","Title":null,"Body":"<p>Okay update the permission in manifest file as below:<\/p>\n\n<pre><code>&lt;permission android:name=\"android.permission.FLASHLIGHT\"\n             android:permissionGroup=\"android.permission-group.HARDWARE_CONTROLS\"\n             android:protectionLevel=\"normal\"\n             android:label=\"@string\/permlab_flashlight\"\n             android:description=\"@string\/permdesc_flashlight\" \/&gt;\n<\/code><\/pre>\n\n<p>Try to off the light using below code:<\/p>\n\n<pre><code>cam.stopPreview();\ncam.release(); \/\/ Very Important line\n<\/code><\/pre>\n"}]}
+{"QuestionId":57875937,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2019-09-10T17:46:26.237","AcceptedAnswerId":null,"OwnerUserId":7212365.0,"Title":"Inconsistent tf.print results","Body":"<p>I was trying to understand the executing order in tensorflow graph. When I was running the following code. <code>tf.print<\/code> prints undetermined results. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\n\nfor _ in range(20):\n    tf.reset_default_graph()\n    use_resource = False\n    v = tf.Variable(0, trainable=False, use_resource=use_resource, name='v')\n    v_op1 = v.assign_add(1, name='v_op1')\n    v_op2 = v.assign_add(2, name='v_op2')\n\n    with tf.control_dependencies([v_op1]):\n        w1 = tf.print('msg1:', v, name='v--w')\n    with tf.control_dependencies([w1, v_op2]):\n        w2 = tf.print('msg2:', v, '\\n', name='w--w')\n    with tf.Session() as sess:\n        sess.run(tf.global_variables_initializer())\n        sess.run(w2)\n<\/code><\/pre>\n\n<p>Here are the output of 20 repetitions. Sometimes, it prints \"msg1: 3\", sometimes it prints \"msg1: 1\". So what's going on here?  <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>msg1: 3\nmsg2: 3 \n\nmsg1: 3\nmsg2: 3 \n\nmsg1: 3\nmsg2: 3 \n\nmsg1: 3\nmsg2: 3 \n\nmsg1: 3\nmsg2: 3 \n\nmsg1: 3\nmsg2: 3 \n\nmsg1: 3\nmsg2: 3 \n\nmsg1: 3\nmsg2: 3 \n\nmsg1: 3\nmsg2: 3 \n\nmsg1: 3\nmsg2: 3 \n\nmsg1: 1\nmsg2: 3 \n\nmsg1: 3\nmsg2: 3 \n\nmsg1: 3\nmsg2: 3 \n\nmsg1: 3\nmsg2: 3 \n\nmsg1: 3\nmsg2: 3 \n\nmsg1: 3\nmsg2: 3 \n\nmsg1: 3\nmsg2: 3 \n\nmsg1: 3\nmsg2: 3 \n\nmsg1: 3\nmsg2: 3 \n\nmsg1: 1\nmsg2: 3 \n<\/code><\/pre>\n","answers":[{"AnswerId":"57876450","CreationDate":"2019-09-10T18:30:36.873","ParentId":null,"OwnerUserId":"7416201","Title":null,"Body":"<p>I think what's happening is you're creating a race between w1 and v_op2, and they're both being executed simultaneously. (with theses lines)<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with tf.control_dependencies([w1, v_op2]):\n<\/code><\/pre>\n\n<p>for example if v_op2 is executed before w1 then 2 is added to <code>v<\/code> then when w1 is executed it adds 1 to it making msg_1 print 3. However when w1 is able to be completed before v_op2 is executed then 2 is added to <code>v<\/code> after.<\/p>\n\n<p>e.g. case 1<\/p>\n\n<pre><code>v_op2 executed:\nv = 2\nw1 executed:\nv = 3\nprint(v)\n<\/code><\/pre>\n\n<p>case 2<\/p>\n\n<pre><code>w1 executed\nv = 1\nprint(v)\nv_op2 executed\nv=3\n<\/code><\/pre>\n"}]}
+{"QuestionId":57876290,"AnswerCount":1,"Tags":"<tensorflow><3d><generative-adversarial-network>","CreationDate":"2019-09-10T18:17:52.347","AcceptedAnswerId":57914344.0,"OwnerUserId":11384699.0,"Title":"How to Train a GAN with a 3D Model Instead of a Million Images","Body":"<p>I'm attempting to train a GAN on a 3D model of a chair with TensorFlow. The purpose is for a GAN model to have full context of a chair and there after be able to generate images with the chair based on the 3D model.<\/p>\n\n<p>What I have been doing is reading a 3D model into python and rotating the model around it's (x,y,z) axis' and storing the images(to be used for training). The result is 1.3 Million images of a chair from every angle. I'm wondering if there is a better approach to this instead of generating million images for a single 3D model.<\/p>\n\n<p>It would be much more efficient to have the GAN learn a 3D model and then generate images with the learned model of the chair in a realistic scene.<\/p>\n\n<p>Python code i'm using for rotating the 3D model and saving the images<\/p>\n\n<pre><code>from stl import mesh\nfrom mpl_toolkits import mplot3d\nfrom matplotlib import pyplot\n\nstl_mesh = mesh.Mesh.from_file('.\/chair.stl')\n\ndef generate_save_figure(elev,azim,dist):         \n    figure = pyplot.figure(figsize=(1,1))\n    axes = mplot3d.Axes3D(figure)\n    axes.grid(False)\n    axes._axis3don=False    \n    axes.add_collection3d(mplot3d.art3d.Poly3DCollection(stl_mesh.vectors))\n    scale = stl_mesh.points.flatten(-1)\n    axes.auto_scale_xyz(scale, scale, scale)\n    axes.view_init(elev=elev,azim=azim)\n    axes.dist = dist\n    axes.autoscale(True)\n    figure.savefig('.\/numpy-stl-images\/elev({})-azim({})-dist({}).png'.format(elev,azim,dist))   \n    print('saved elev {}, azim {}, dist {}'.format(elev,azim,dist))\n    del figure,axes,scale\n    pyplot.close('all')\n\nfor elev in range(0,180,1):\n    for azim in range(0,360,1):\n        for dist in range(5,25,1):\n            generate_save_figure(elev,azim,dist)\n<\/code><\/pre>\n\n<p>link to the github repo i'm working on for additional context of this question(note that the chair dataset is not yet available)\n<a href=\"https:\/\/github.com\/RauxaDataScience\/GansContextDataSets\" rel=\"nofollow noreferrer\">https:\/\/github.com\/RauxaDataScience\/GansContextDataSets<\/a><\/p>\n","answers":[{"AnswerId":"57914344","CreationDate":"2019-09-12T20:52:35.037","ParentId":null,"OwnerUserId":"11384699","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/graphics\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/graphics\/<\/a> is the answer to training an ML model with 3D data<\/p>\n"}]}
+{"QuestionId":57876587,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-10T18:41:21.087","AcceptedAnswerId":null,"OwnerUserId":6802252.0,"Title":"Tensorflow hub imagenet models preprocessing functions?","Body":"<p>I am going through some Tensorflow 2.0 tutorials, and if you go through their\ncolab notebooks, they are using a simple preprocess function for the select imagenet models\nthey are representing in the tutorials. Below is one such example:<\/p>\n\n<p><a href=\"https:\/\/tfhub.dev\/google\/tf2-preview\/mobilenet_v2\/feature_vector\/4\" rel=\"nofollow noreferrer\">https:\/\/tfhub.dev\/google\/tf2-preview\/mobilenet_v2\/feature_vector\/4<\/a> (just hit Open Colab Notebook)<\/p>\n\n<p>The preprocess function used is just 1\/255 (see rescale).<\/p>\n\n<p>Have they been retrained to use this simple preprocess function?<\/p>\n\n<p>In Keras all of the imagenet models with Tensorflow backened are scaling it in between -1 and 1,\nwhile I have seen some models use standardization (get mean and std of each channel, and subtract from each image etc etc) which I thought was the official way to preprocess imagenet models.<\/p>\n\n<p>So is the preprocess function (1\/255) not the original preprocesss function and just a simplification of the preprocess function? If so what is the original preprocess function?<\/p>\n","answers":[]}
+{"QuestionId":57876757,"AnswerCount":1,"Tags":"<python><linux><tensorflow><package>","CreationDate":"2019-09-10T18:55:38.367","AcceptedAnswerId":57876835.0,"OwnerUserId":11861082.0,"Title":"How to see what packages are required to install another package?","Body":"<p>I want to install many packages on an offline computer, I download and install manually using the following code:<\/p>\n\n<pre><code>pip install XXXXX.whl --user\n<\/code><\/pre>\n\n<p>Although the code above does the offline installation, the package needs other packages so it tries to connect to the internet. I can see what package it is looking for. So, I will download manually and again install that. If there are a lot of packages required, it becomes overwhelming. <\/p>\n\n<p>Any better solution? Can I know from the beginning that what packages have to be downloaded for installing my package?<\/p>\n","answers":[{"AnswerId":"57876835","CreationDate":"2019-09-10T19:03:25.163","ParentId":null,"OwnerUserId":"11702228","Title":null,"Body":"<p>Do you use anaconda?\nMake a conda environment on your online pc and install all your packages you need.<\/p>\n\n<p>Then check with <\/p>\n\n<p>'''\nconda list\n'''<\/p>\n\n<p>What packages you need.<\/p>\n"}]}
+{"QuestionId":57876973,"AnswerCount":0,"Tags":"<c++><tensorflow><restore>","CreationDate":"2019-09-10T19:15:06.543","AcceptedAnswerId":null,"OwnerUserId":12048454.0,"Title":"Saving a tensorflow model in Python using SaveModelBuilder and Loading it in C++","Body":"<p>I want to save a model in python using the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/saved_model\/Builder\" rel=\"nofollow noreferrer\">SaveModelBuilder<\/a> API and then load it in C++. <\/p>\n\n<p>I was able to successfully implement the tensorflow <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/label_image\" rel=\"nofollow noreferrer\">label_image<\/a> example, and load the downloaded model in c++. Although when I try to do the same with my own custom model it does not work. I then implemented <a href=\"https:\/\/stackoverflow.com\/questions\/33759623\/tensorflow-how-to-save-restore-a-model\/47235448#47235448\">Tom's solution<\/a> to export my .pb file. Here I tried to load the model as shown in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/label_image\" rel=\"nofollow noreferrer\">label_image<\/a> example and also using Meta Graph, but it did not work. <\/p>\n\n<p>My code to load Graph Def is <\/p>\n\n<pre><code> tensorflow::GraphDef graph_def;\n  Status load_graph_status =  ReadBinaryProto(tensorflow::Env::Default(), graph_file_name, &amp;graph_def);\n  session-&gt;reset(tensorflow::NewSession(tensorflow::SessionOptions()));\n  Status session_create_status = (*session)-&gt;Create(graph_def);\n<\/code><\/pre>\n\n<p>and to load Meta Graph Def is<\/p>\n\n<pre><code> tensorflow::MetaGraphDef graph_def;\n  Status load_graph_status =  ReadBinaryProto(tensorflow::Env::Default(), graph_file_name, &amp;graph_def);\n  session-&gt;reset(tensorflow::NewSession(tensorflow::SessionOptions()));\n  Status session_create_status = (*session)-&gt;Create(graph_def.graph_def());\n<\/code><\/pre>\n\n<p>I saved the model in python as per <a href=\"https:\/\/stackoverflow.com\/questions\/33759623\/tensorflow-how-to-save-restore-a-model\/47235448#47235448\">Tom's article<\/a> as:<\/p>\n\n<pre><code>prediction_signature = ( tf.saved_model.signature_def_utils.build_signature_def(\n            inputs = {\"inputs\": model_input}, outputs = {\"outputs\":model_output}, method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME))            \nbuilder.add_meta_graph_and_variables(sess, [tf.saved_model.tag_constants.SERVING], signature_def_map= { tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:prediction_signature },)\nbuilder.save()\n<\/code><\/pre>\n\n<p>When I run the code. I just get the error <\/p>\n\n<pre><code> Failed to load compute graph\n<\/code><\/pre>\n\n<p>when I have defined the graph as a normal graph, although if I define it as MetaGraph, I get the following error:<\/p>\n\n<pre><code>[libprotobuf ERROR external\/protobuf_archive\/src\/google\/protobuf\/wire_format_lite.cc:611] String field 'tensorflow.NodeDef.op' contains invalid UTF-8 data when parsing a protocol buffer. Use the 'bytes' type if you intend to send raw bytes.\n<\/code><\/pre>\n\n<p>Note that the I am able to run the label_image example, so there no other issues like incorrect path to model or general syntax issues.<\/p>\n\n<p>I want helping understanding and using:<\/p>\n\n<pre><code>SavedModelBundle bundle;\n...\nLoadSavedModel(session_options, run_options, export_dir, {kSavedModelTagTrain},\n               &amp;bundle);\n<\/code><\/pre>\n\n<p>as instructed on the <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model\" rel=\"nofollow noreferrer\">Save &amp; Restore<\/a> Documentation provided by tensorflow.<\/p>\n\n<p>To conclude I want a way to save tensorflow model in python using SaveModelBuilder and load it in c++.<\/p>\n","answers":[]}
+{"QuestionId":57877064,"AnswerCount":1,"Tags":"<keras><imbalanced-data>","CreationDate":"2019-09-10T19:23:36.473","AcceptedAnswerId":null,"OwnerUserId":6450489.0,"Title":"Sample weighting didn't help in imbalanced data training","Body":"<p>I am training a two-layer LSTM network with 16 through 32 cells in each layer and had a fairly imbalanced dataset for training. Based on my seven class frequencies, the sample weights calculated through the simple formula of total_samples\/class_frequency is [3.7, 5.6, 26.4, 3.2, 191.6, 8.4, 13.2], and I add this weight for each sample to the tuple of (data, label) output of my dataset generator to run my Keras <code>model.fit()<\/code> function. The training code was:<\/p>\n\n<pre><code>model.compile(optimizer=optimizer, loss='categorical_crossentropy', metrics=['accuracy'])\nmc = ModelCheckpoint(model_file, monitor='val_acc', mode='max', verbose=1, save_best_only=True)\nes = EarlyStopping(monitor='val_acc', mode='max', verbose=1, patience=50)\nhistory = model.fit(train_data, epochs=epochs, steps_per_epoch = train_steps, validation_data=val_data,\n                            validation_steps = val_steps, verbose=verbose, callbacks=[es, mc])\n<\/code><\/pre>\n\n<p>Then I used the best saved model to evaluate it and calculate performance statistics by this code (my data is in tensorflow datasets):<\/p>\n\n<pre><code>saved_model = load_model(model_file)\niterator = test_data.make_one_shot_iterator()\nnext_element = iterator.get_next()\ny_test = y_pred = np.empty(0)\nfor i in range(test_steps):\n    batch = sess.run(next_element)\n    x_test_batch = batch[0]\n    y_test_batch = batch[1]\n    y_pred_batch = saved_model.predict_on_batch(x_test_batch)\n    y_test = np.append(y_test, np.argmax(y_test_batch, axis=1))\n    y_pred = np.append(y_pred, np.argmax(y_pred_batch, axis=1))\nprint('\\nTest data classification report:\\n{}\\n'.format(classification_report(y_test, y_pred)))\n<\/code><\/pre>\n\n<p>But what I see in the output statistics is that the weighted stats are overall worse than unweighted ones (setting all weights equally to 1), even for rare classes (highest weights). Here is the stat:<\/p>\n\n<p>For weighted run:<\/p>\n\n<pre><code>     class     prec.     recall    f1       support\n     0.0       1.00      0.97      0.98     79785\n     1.0       0.89      0.88      0.88     52614\n     2.0       0.61      0.76      0.68     11090\n     3.0       0.96      0.93      0.95     91160\n     4.0       0.59      0.92      0.72      1530\n     5.0       0.89      0.90      0.89     34746\n     6.0       0.81      0.87      0.84     22289\n\naccuracy                           0.92    293214\nmacro avg      0.82      0.89      0.85    293214\n<\/code><\/pre>\n\n<p>For unweighted run:<\/p>\n\n<pre><code>     class     prec.     recall    f1       support\n     0.0       0.99      0.98      0.99     79785\n     1.0       0.89      0.90      0.90     52614\n     2.0       0.79      0.66      0.72     11090\n     3.0       0.95      0.96      0.95     91160\n     4.0       0.85      0.82      0.83      1530\n     5.0       0.89      0.92      0.90     34746\n     6.0       0.88      0.86      0.87     22289\n\naccuracy                           0.93    293214\nmacro avg      0.89      0.87      0.88    293214\n<\/code><\/pre>\n\n<p>what is wrong here?<\/p>\n","answers":[{"AnswerId":"57883500","CreationDate":"2019-09-11T06:44:57.423","ParentId":null,"OwnerUserId":"1729802","Title":null,"Body":"<p>You should be using the <code>class_weight<\/code> in <code>fit<\/code> function or <code>fit_generator<\/code> to apply weights to your classes.<\/p>\n\n<p>First you have to create a dictionary with <code>label:weight<\/code> format:<\/p>\n\n<pre><code>class_weight = {0: 3.7,\n            1: 5.6,\n            2: 2.64,...}\n<\/code><\/pre>\n\n<p>Then apply it to your fit function:<\/p>\n\n<pre><code>history = model.fit(train_data, epochs=epochs, steps_per_epoch = train_steps, validation_data=val_data, \n                        class_weight=class_weight, validation_steps = val_steps, verbose=verbose, callbacks=[es, mc])\n<\/code><\/pre>\n\n<p>If you want to apply a weight per instance, then you need to create an array that contains the weight for the corresponding instance in the training data and set it in <code>sample_weight<\/code> in fit function.<\/p>\n"}]}
+{"QuestionId":57877842,"AnswerCount":0,"Tags":"<python><tensorflow><placeholder><serving>","CreationDate":"2019-09-10T20:31:04.057","AcceptedAnswerId":null,"OwnerUserId":12046218.0,"Title":"Feeding arrays to served Tensorflow model with JSON","Body":"<p>I trained a model in Tensorflow 1.14 and deployed it on AI platform. Now I want to get predictions, but two of my features are actually lists which are multi-hot encoded through indicator columns. How do I feed this to the model?<\/p>\n\n<p>I made a servingInputReceiver as follows: <\/p>\n\n<pre><code>def serving_input_fn() -&gt; tf.estimator.export.ServingInputReceiver:\n\"\"\"Defines the function providing the inputs at serving time.\n\nReturns:\n    A serving-time input function.\n\"\"\"\n\ndef get_placeholder(key):\n    if FEATURES[key] == int:\n        return tf.placeholder(tf.int64, [None])\n    elif FEATURES[key] == float:\n        return tf.placeholder(tf.float64, [None])\n    else:\n        return tf.placeholder(tf.int64, [None, None])\n\nplaceholders = {key: get_placeholder(key) for key in FEATURES}\nfeatures = {k: tf.expand_dims(v, -1) for k, v in placeholders.items()}\n\nreturn tf.estimator.export.ServingInputReceiver(features, placeholders)\n<\/code><\/pre>\n\n<p>Now I want to get predictions, requested with the following JSON example:<\/p>\n\n<pre><code>{\n\"instances\": [\n    {\n        \"feat_1\": 3.91,\n        \"feat_2\": 0,\n        \"feat_3\": 0,\n        \"feat_4\": 0.00,\n        \"feat_5\": [],\n        \"feat_6\": [73,101,7,9,1]\n    }\n]\n<\/code><\/pre>\n\n<p>}<\/p>\n\n<p>as feat_5 and feat_6 are the features that have <code>tf.placeholder(tf.int64, [None, None, None])<\/code> placeholder (they are also lists in my training data). \nHowever, the JSON above does not work: <\/p>\n\n<pre><code>AbortionError(code=StatusCode.INVALID_ARGUMENT, details=\\\"Input to reshape is a tensor with 5000 values, but the requested shape has 1000\\n\\t [[{{node dnn\/input_from_feature_columns\/input_layer\/feat_6_indicator\/Reshape}}]]\n<\/code><\/pre>\n\n<p>This also happens when I try to feed a list for feat_5. I can only feed single values in the JSON file. Can anybody help me with providing a list\/array feature as input into a deployed model?<\/p>\n\n<p>Edit: \nThe model is a DNNClassifier with one input layer and 2 hidden layers. <\/p>\n","answers":[]}
+{"QuestionId":57877959,"AnswerCount":1,"Tags":"<tensorflow><dataset><tensorflow-lite>","CreationDate":"2019-09-10T20:40:08.973","AcceptedAnswerId":null,"OwnerUserId":8050962.0,"Title":"What is the correct way to create representative dataset for TFliteconverter?","Body":"<p>I am trying to infer <code>tinyYOLO-V2<\/code> with <code>INT8<\/code> weights and activation. I can convert the weights to INT8 with TFliteConverter. For <code>INT8<\/code> activation, I have to give representative dataset to estimate the scaling factor. My method of creating such dataset seems wrong. <\/p>\n\n<p>What is the correct procedure ?<\/p>\n\n<pre><code>def rep_data_gen():\n    a = []\n    for i in range(160):\n        inst = anns[i]\n        file_name = inst['filename']\n        img = cv2.imread(img_dir + file_name)\n        img = cv2.resize(img, (NORM_H, NORM_W))\n        img = img \/ 255.0\n        img = img.astype('float32')\n        a.append(img)\n    a = np.array(a)\n    print(a.shape) # a is np array of 160 3D images\n    img = tf.data.Dataset.from_tensor_slices(a).batch(1)\n    for i in img.take(BATCH_SIZE):\n        print(i)\n        yield [i]\n# https:\/\/www.tensorflow.org\/lite\/performance\/post_training_quantization\nconverter = tf.lite.TFLiteConverter.from_keras_model_file(\".\/yolo.h5\")\nconverter.optimizations = [tf.lite.Optimize.DEFAULT]\nconverter.target_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]\nconverter.inference_input_type = [tf.int8]\nconverter.inference_output_type = [tf.int8]\nconverter.representative_dataset=rep_data_gen\ntflite_quant_model = converter.convert()\n<\/code><\/pre>\n\n<blockquote>\n  <p>ValueError: Cannot set tensor: Got tensor of type STRING but expected type FLOAT32 for input 27, name: input_1 <\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"59051374","CreationDate":"2019-11-26T12:50:15.333","ParentId":null,"OwnerUserId":"10295473","Title":null,"Body":"<p>I used your code for reading in a dataset and found the error:<\/p>\n\n<p>img = img.astype('float32') should be<\/p>\n\n<p>img = img.astype(np.float32)<\/p>\n\n<p>Hope this helps<\/p>\n"}]}
+{"QuestionId":57878160,"AnswerCount":1,"Tags":"<tensorflow><keras><tensorflow2.0>","CreationDate":"2019-09-10T21:01:14.267","AcceptedAnswerId":null,"OwnerUserId":11804809.0,"Title":"Reuse the weight matrix from embedding layer with @tf.function","Body":"<p>Without using @tf.function, the script work perfectly<\/p>\n\n<p>I want to use it to speed up training, but it's giving me error where I reuse the weight matrix from the embedding layers. <\/p>\n\n<p>I think the error is caused by get_weights(), because it converts tensor back to numpy <\/p>\n\n<p>I tried to use a tf.keras.layers.Dense instead of re-using the weights from embedding, and it worked perfectly.<\/p>\n\n<pre><code>class Example(tf.keras.Model):\n    def __init__(self,):\n        super(Example, self).__init__()\n        self.embed_dim = embed_dim\n        self.vocab_size = vocab_size\n        self.embed = tf.keras.layers.Embedding(self.vocab_size, self.embed_dim)\n        ...\n\n    def call(self, inputs, trianing):\n        ...\n        embed_matrix = self.embed.get_weights()\n\n        # a dense layer\n        Vhid = tf.matmul(self.kernel, tf.transpose(embed_matrix[0]))\n        pred_w = tf.matmul(pred, Vhid) + self.bias\n<\/code><\/pre>\n\n<p>In my train script.\nI did<\/p>\n\n<pre><code>@tf.function\ndef train_step(x, y, training=None):\n\n    with tf.GradientTape() as tape:\n        pred = model(x, y, training)\n        losses = compute_loss(y, pred)\n\n    grads = tape.gradient(losses, model.trainable_variables)\n\n    optimizer.apply_gradients(zip(grads, model.trainable_variables))\n\n    return losses\n\n\/home\/thomas\/projects\/tf_convsent\/models\/.py:195 call  *\n   embed_matrix = self.embed.get_weights()  # [vocab_size, 300]\n\/home\/thomas\/.conda\/envs\/tf2_p37\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py:1177 get_weights\n   return backend.batch_get_value(params)\n\/home\/thomas\/.conda\/envs\/tf2_p37\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/backend.py:3011 batch_get_value\n   raise RuntimeError('Cannot get value inside Tensorflow graph function.')\n\nRuntimeError: Cannot get value inside Tensorflow graph function.\n<\/code><\/pre>\n","answers":[{"AnswerId":"57910278","CreationDate":"2019-09-12T15:42:50.607","ParentId":null,"OwnerUserId":"11804809","Title":null,"Body":"<p>Found the easiest solution which improved 50% training speed(122 hrs to ~65 hrs)<\/p>\n\n<p>just change<\/p>\n\n<pre><code>embed_matrix = self.embed.get_weights()\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>embed_matrix = self.embed.weights\n<\/code><\/pre>\n\n<p>will do the trick.<\/p>\n"}]}
+{"QuestionId":57878623,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><entropy>","CreationDate":"2019-09-10T21:47:28.580","AcceptedAnswerId":null,"OwnerUserId":9706484.0,"Title":"Error with tf.nn.sparse_softmax_cross_entropy_with_logits","Body":"<p>I am using <code>tf.nn.sparse_softmax_cross_entropy_with_logits<\/code> and when I pass through the labels and logits I get the following error<\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors_impl.InvalidArgumentError: labels\n  must be 1-D, but got shape [50,1]<\/p>\n<\/blockquote>\n\n<p>I don't understnad how having a shape <code>[50,1]<\/code> is not the same as being 1D<\/p>\n","answers":[{"AnswerId":"57887051","CreationDate":"2019-09-11T10:22:46.277","ParentId":null,"OwnerUserId":"10025506","Title":null,"Body":"<blockquote>\n  <p>I don't understand how having a shape [50,1] is not the same as being 1D.<\/p>\n<\/blockquote>\n\n<p>While you can reshape a [50, 1] 2D matrix into a [50] 1D matrix just with a simple squeeze, Tensorflow will never do that automatically. <\/p>\n\n<p>The only heuristic the <code>tf.nn.sparse_softmax_cross_entropy_with_logits<\/code> uses to check if the input shape is correct is to check the number of dimensions it has. If it's not 1D, it fails without trying other heuristics like checking if the input could be squeezed. This is a security feature.<\/p>\n"}]}
+{"QuestionId":57878624,"AnswerCount":2,"Tags":"<keras><deep-learning><data-augmentation>","CreationDate":"2019-09-10T21:47:31.420","AcceptedAnswerId":57882787.0,"OwnerUserId":7065438.0,"Title":"Is it the right way for data augumentation for training a model?","Body":"<p>I'm new to keras and deep learning. I'have tried to use data augmentation for for training my model, but not sure if i'm doing it the right way. Can anyone assure me it my approach is correct? here is my code:<\/p>\n\n<pre><code>train_path = 'Digital_Mamo\/OPTIMAM'      # Relative Path\nvalid_path = 'Digital_Mamo\/InBreast'\ntest_path = 'Digital_Mamo\/BCDR'\n\n\nvalid_batches = ImageDataGenerator().flow_from_directory(valid_path, target_size=(224, 224), classes=['Benign', 'Malignant'], batch_size=9)\ntest_batches = ImageDataGenerator().flow_from_directory(test_path, target_size=(224, 224), classes=['Benign', 'Malignant'], batch_size=7)\n\ndatagen = ImageDataGenerator(rotation_range=10, width_shift_range=0.1,\n       height_shift_range=0.1, shear_range=0.15, zoom_range=0.1,\n       channel_shift_range=10., horizontal_flip=True)\n\ntrain_batches = datagen.flow_from_directory(\n        train_path,\n        target_size=(224, 224),\n        batch_size=10,\n        classes=['Benign','Malignant'])\n\n\nvgg16_model= load_model('Fetched_VGG.h5')\n\n# transform the model to Sequential\nmodel= Sequential()\nfor layer in vgg16_model.layers[:-1]:\n    model.add(layer)\n\nmodel.summary()\n\n# Freezing the layers (Oppose weights to be updated)\nfor layer in model.layers:\n    layer.trainable = False\n\n\nmodel.add(Dense(2, activation='softmax', name='predictions'))\n\n\n### Compile the model\nmodel.compile(Adam(lr=.0001), loss='categorical_crossentropy', metrics=['accuracy'])\n\n# train the model\nmodel.fit_generator(train_batches, steps_per_epoch=28, validation_data=valid_batches, validation_steps=3, epochs=5, verbose=2)\n\n#test\npredictions = model.predict_generator(test_batches, steps=3, verbose=0)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57881900","CreationDate":"2019-09-11T03:54:01.853","ParentId":null,"OwnerUserId":"9477338","Title":null,"Body":"<p>It's a correct way, but you can use <code>steps_per_epoch=len(train_batches)<\/code> and <code>validation_steps=len(val_batches)<\/code>  for easier life when you add more data though. Also you can just include <code>test<\/code> in <code>valid<\/code> since it won't help anything even you're using them separately.<\/p>\n\n<h1>Edit:<\/h1>\n\n<p>As @Matias point out that you shouldn't use augmentation on validate. So it's not totally wrong as I said in comment but not really correct.<\/p>\n"},{"AnswerId":"57882787","CreationDate":"2019-09-11T05:42:23.457","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Actually this is not correct, because the way you coded this, you are applying data augmentation to the validation and test sets, and you should only apply augmentation to the training set.<\/p>\n\n<p>You would need to create a second instance of <code>ImageDataGenerator<\/code> for the validation and test sets, without any augmentations set.<\/p>\n"}]}
+{"QuestionId":57879190,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2019-09-10T23:05:19.970","AcceptedAnswerId":null,"OwnerUserId":8338613.0,"Title":"Add Keras Layer to Flatten input by a given axis","Body":"<p>If my last Keras Layer X outputs a tensor which looks, e.g. like this:<\/p>\n\n<pre><code>{\n  {4, [22, 16,  11, 33], 24},\n  {4, [12, 33,  87, 17], 14},\n  {3, [92, 407, 25, 27], 34}\n}\n<\/code><\/pre>\n\n<p>How can I add one more Layer Y that converts the output of X to this:<\/p>\n\n<pre><code>{\n  {4, 22, 16,  11, 33, 24},\n  {4, 12, 33,  87, 17, 14},\n  {3, 92, 407, 25, 27, 34}\n}\n<\/code><\/pre>\n","answers":[{"AnswerId":"57881779","CreationDate":"2019-09-11T03:37:06.460","ParentId":null,"OwnerUserId":"10841164","Title":null,"Body":"<p>There is actually a type of layer called Flatten that suits your need (Keras core layer documentation <a href=\"https:\/\/keras.io\/layers\/core\/\" rel=\"nofollow noreferrer\">here<\/a>):<\/p>\n\n<pre><code>from keras.layers import Flatten\n\n# if you are using sequential model\n...\nmodel.add(Flatten())\n...\n\n# if you are using Model api\n...\nY = Flatten()(X)\n...\n<\/code><\/pre>\n"}]}
+{"QuestionId":57879217,"AnswerCount":0,"Tags":"<docker><tensorflow><ros>","CreationDate":"2019-09-10T23:09:42.243","AcceptedAnswerId":null,"OwnerUserId":1945237.0,"Title":"Use multiple docker images at once","Body":"<p>I am trying to containerize my ROS + tensorflow application. The problem is that I want to use GPU. I can either use GPU and forget about ROS, or use ROS and forget about GPU, but I don't know how I can enable both of them.<\/p>\n\n<p>I have tried starting FROM a cuda image and installing ROS as described <a href=\"http:\/\/wiki.ros.org\/kinetic\/Installation\/Ubuntu\" rel=\"nofollow noreferrer\">here<\/a>, but the ros couldn't be installed, not being able to find the package.<\/p>\n\n<p>I also tried building them one by one in the same Dockerfile and copying all the ROS-related stuff from a ROS build, but that failed too.<\/p>\n\n<p>Ideally, I want to make it work as if I \"include\" a Cuda image and a ROS image, but resource on building from multiple images like this could hardly be found. Any help or pointer would be appreciated.<\/p>\n","answers":[]}
+{"QuestionId":57879330,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2019-09-10T23:30:28.203","AcceptedAnswerId":57910498.0,"OwnerUserId":6802252.0,"Title":"Getting imagenet model pnasnet_large with to work with hub.KerasLayer","Body":"<p>I am trying out various Tensorflow models from hub but \nI can't seem to get this one to work with KerasLayer:<\/p>\n\n<p>\"<a href=\"https:\/\/tfhub.dev\/google\/imagenet\/pnasnet_large\/feature_vector\/3\" rel=\"nofollow noreferrer\">https:\/\/tfhub.dev\/google\/imagenet\/pnasnet_large\/feature_vector\/3<\/a>\"<\/p>\n\n<p>I am using the same procedure used within the examples in the documentation:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/tutorials\/images\/hub_with_keras\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/images\/hub_with_keras<\/a><\/p>\n\n<p><code>feature_extractor = hub.KerasLayer(URL,\n                            input_shape=(height, width,3))<\/code><\/p>\n\n<p>even tried a few amendments such as including:\n<code>trainable=True, tags={\"train\"})<\/code>,<\/p>\n\n<p>so it would look like this:<\/p>\n\n<p><code>feature_extractor = hub.KerasLayer(URL,\n                            input_shape=(height, width,3), trainable=True, tags={\"train\"})<\/code>)<\/p>\n\n<p>because that's what it said to do in the docs. <\/p>\n\n<p>However, I am still getting this error:<\/p>\n\n<blockquote>\n  <p>ValueError: Importing a SavedModel with tf.saved_model.load requires a 'tags=' argument if there is more than one MetaGraph. Got 'tags=None', but there are 2 MetaGraphs in the SavedModel with tag sets [[], ['train']]. Pass a 'tags=' argument to load this SavedModel<\/p>\n<\/blockquote>\n\n<pre><code>import tensorflow as tf\nimport tensorflow_hub as hub\nimport time\nimport numpy as np\nimport matplotlib.pylab as plt\n\nimport tensorflow_datasets as tfds\ntfds.disable_progress_bar()\n\nfrom tensorflow.keras import layers\nfrom tensorflow.keras.applications import imagenet_utils\nfrom tensorflow.keras.callbacks import ModelCheckpoint\n\n\ntf.compat.v1.disable_eager_execution()\n\n\nURL = \"https:\/\/tfhub.dev\/google\/imagenet\/pnasnet_large\/feature_vector\/3\"\nmodule = hub.Module(URL)\nheight, width = hub.get_expected_image_size(module)\n\n\n# here is where the error comes from\nfeature_extractor = hub.KerasLayer(URL,\n                        input_shape=(height, width,3), trainable=True, tags={\"train\"})\n<\/code><\/pre>\n","answers":[{"AnswerId":"57910498","CreationDate":"2019-09-12T15:55:15.507","ParentId":null,"OwnerUserId":"7061249","Title":null,"Body":"<p>At this time, <code>hub.KerasLayer<\/code> only works for TF2-style models like <a href=\"https:\/\/tfhub.dev\/google\/tf2-preview\/inception_v3\/feature_vector\/4\" rel=\"nofollow noreferrer\">https:\/\/tfhub.dev\/google\/tf2-preview\/inception_v3\/feature_vector\/4<\/a>.<\/p>\n\n<p>Please stay tuned for more choice of TF2-style models as TensorFlow 2.0 is getting released.<\/p>\n"}]}
+{"QuestionId":57879641,"AnswerCount":1,"Tags":"<docker><tensorflow><tensorflow-serving>","CreationDate":"2019-09-11T00:31:38.567","AcceptedAnswerId":null,"OwnerUserId":7643739.0,"Title":"When mounting using docker, bind source path does not exist","Body":"<p>When trying to mount using docker, I tried all sorts of things like restarting docker and using different directories.<\/p>\n\n<pre><code>MacBook-Pro:user$ sudo docker run -p 8501:8501 \n--mount type=bind,\nsource=\/user\/Documents\/Repo\/models\/userCategorization\/saved_model\/,target=\/models\/saved_model \n-e MODEL_NAME=saved_model \n-t tensorflow\/serving\n\ndocker: Error response from daemon: invalid mount config for type \"bind\": \nbind source path does not exist: \/user\/Documents\/Repo\/models\/userCategorization\/saved_model\/.\n\n\n<\/code><\/pre>\n\n<p>Wanted to add that \"Users\/\" directory is indeed in the file sharing preferences and the directory for the model file would be \"Users\/user\/Documents\/..etc\".<\/p>\n","answers":[{"AnswerId":"57895503","CreationDate":"2019-09-11T19:13:33.537","ParentId":null,"OwnerUserId":"7643739","Title":null,"Body":"<p>As it turns out:<\/p>\n\n<p>My first issue was that I was not targeting the correct directory, such as <code>\/model_folder<\/code><\/p>\n\n<p>My second issue was that I was that I did not have a specific version in the model folder such ash\n<code>\/model_folder\/1\/<\/code><\/p>\n"}]}
+{"QuestionId":57879708,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><object-detection>","CreationDate":"2019-09-11T00:46:34.683","AcceptedAnswerId":57888733.0,"OwnerUserId":8829431.0,"Title":"Using tensorflow object detection for either or detection","Body":"<p>I have used Tensorflow object detection for quite awhile now. I am more of a user, I dont really know how it works. I am wondering is it possible to train it to recognize an object is something and not something? For example, I want to detect cracks on the tiles. Can i use object detection to do so where i show an image of a tile and it can tell me if there is a crack (and also show the location), or it will tell me if there is no crack on the tile? <\/p>\n\n<p>I have tried to train using pictures with and without defect, using 2 classes (1 for defect and 1 for no defect). But the results keep showing both (if the picture have defect) in 1 picture. Is there a way to show only the one with defect?<\/p>\n\n<p>Basically i would like to do defect checking. This is a simplistic case of 1 defect. but the actual case will have a few defects.\nThank you.<\/p>\n","answers":[{"AnswerId":"57888733","CreationDate":"2019-09-11T12:03:07.447","ParentId":null,"OwnerUserId":"10263280","Title":null,"Body":"<p>In case you're only expecting input images of tiles, either with defects or not, you don't need a class for no defect.\nThe API adds a background class for everything which is not the other classes.\nSo you simply need to state one class - defect, and tiles which are not detected as such are not defected. \nSo in your training set - simply give bounding boxes of defects, and no bounding box in case of no defect, and then your model should learn to detect the defects as mentioned above.<\/p>\n"}]}
+{"QuestionId":57879722,"AnswerCount":0,"Tags":"<javascript><alignment><caffe><face>","CreationDate":"2019-09-11T00:49:30.853","AcceptedAnswerId":null,"OwnerUserId":12049397.0,"Title":"Check failed: axis_index < num_axes() (1 vs. 0) axis 1 out of range for 0-D Blob with shape (0)","Body":"<p>When I run the code from GitHub\uff0cwhich is <code>Face_Alignment_Two_Stage_Re-initialization<\/code>,I got the following error:<\/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>I0911 08:08:29.921020 12228 layer_factory.hpp:88] Creating layer input\r\nI0911 08:08:29.922020 12228 net.cpp:100] Creating Layer input\r\nI0911 08:08:29.922020 12228 net.cpp:418] input -&gt; data\r\nI0911 08:08:29.923020 12228 net.cpp:150] Setting up input\r\nI0911 08:08:29.924021 12228 net.cpp:157] Top shape: 1 3 448 448 (602112)\r\nI0911 08:08:29.927022 12228 net.cpp:165] Memory required for data: 2408448\r\nI0911 08:08:29.927022 12228 layer_factory.hpp:88] Creating layer data_input_0_split\r\nI0911 08:08:29.928025 12228 net.cpp:100] Creating Layer data_input_0_split\r\nI0911 08:08:29.928025 12228 net.cpp:444] data_input_0_split &lt;- data\r\nI0911 08:08:29.929024 12228 net.cpp:418] data_input_0_split -&gt; data_input_0_split_0\r\nI0911 08:08:29.930024 12228 net.cpp:418] data_input_0_split -&gt; data_input_0_split_1\r\nI0911 08:08:29.930024 12228 net.cpp:150] Setting up data_input_0_split\r\nI0911 08:08:29.930024 12228 net.cpp:157] Top shape: 1 3 448 448 (602112)\r\nI0911 08:08:29.932025 12228 net.cpp:157] Top shape: 1 3 448 448 (602112)\r\nI0911 08:08:29.932025 12228 net.cpp:165] Memory required for data: 7225344\r\nI0911 08:08:29.932025 12228 layer_factory.hpp:88] Creating layer downsample_data\r\nI0911 08:08:29.932025 12228 net.cpp:100] Creating Layer downsample_data\r\nI0911 08:08:29.932025 12228 net.cpp:444] downsample_data &lt;- data_input_0_split_0\r\nI0911 08:08:29.932025 12228 net.cpp:418] downsample_data -&gt; downsample_data\r\nI0911 08:08:29.932025 12228 net.cpp:150] Setting up downsample_data\r\nI0911 08:08:29.932025 12228 net.cpp:157] Top shape: (0)\r\nI0911 08:08:29.932025 12228 net.cpp:165] Memory required for data: 7225344\r\nI0911 08:08:29.933027 12228 layer_factory.hpp:88] Creating layer net1_conv1\r\nI0911 08:08:29.933027 12228 net.cpp:100] Creating Layer net1_conv1\r\nI0911 08:08:29.933027 12228 net.cpp:444] net1_conv1 &lt;- downsample_data\r\nI0911 08:08:29.933027 12228 net.cpp:418] net1_conv1 -&gt; net1_conv1\r\nF0911 08:08:29.933027 12228 blob.hpp:122] Check failed: axis_index &lt; num_axes() (1 vs. 0) axis 1 out of range for 0-D Blob with shape (0)\r\n*** Check failure stack trace: ***<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n\nIt seems like that the error occurred because of <code>SubsamplePooling<\/code> layer, and why the top shape is 0? I think the top shape of <code>SubsamplePooling<\/code> layer should be 1 3 60 60.\nThe part of the deploy is as follows.<\/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>name: \"facial_point_net\"\r\ninput: \"data\"\r\ninput_dim: 1\r\ninput_dim: 3\r\ninput_dim: 448\r\ninput_dim: 448\r\n\r\nlayer {\r\n  name: \"downsample_data\"\r\n  type: \"SubsamplePooling\"\r\n  bottom: \"data\"\r\n  top: \"downsample_data\"\r\n  subsample_pooling_param {\r\n    output_H: 60\r\n    output_W: 60\r\n  }\r\n}\r\n##########################################\r\nlayer {\r\n  name: \"net1_conv1\"\r\n  type: \"Convolution\"\r\n  bottom: \"downsample_data\"\r\n  top: \"net1_conv1\"\r\n  param {\r\n    lr_mult: 1\r\n    decay_mult: 1\r\n  }\r\n  param {\r\n    lr_mult: 2\r\n    decay_mult: 0\r\n  }\r\n  convolution_param {\r\n    num_output: 20\r\n    pad: 0\r\n    kernel_size: 5\r\n    stride: 1\r\n    weight_filler {\r\n      type: \"gaussian\"\r\n      std: 0.01\r\n    }\r\n    bias_filler {\r\n      type: \"constant\"\r\n      value: 0\r\n    }\r\n  }\r\n}\r\nlayer {\r\n  name: \"net1_PReLU1\"\r\n  type: \"PReLU\"\r\n  bottom: \"net1_conv1\"\r\n  top: \"net1_conv1\"\r\n}\r\n\r\nlayer {\r\n  name: \"net1_pool1\"\r\n  type: \"Pooling\"\r\n  bottom: \"net1_conv1\"\r\n  top: \"net1_pool1\"\r\n  pooling_param {\r\n    pool: MAX\r\n    kernel_size: 2\r\n    stride: 2\r\n  }\r\n}<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n\nDoes anyone know what the problem is ? I need help.<\/p>\n","answers":[]}
+{"QuestionId":57881029,"AnswerCount":1,"Tags":"<python><html><flask><keras><data-science>","CreationDate":"2019-09-11T01:55:41.063","AcceptedAnswerId":null,"OwnerUserId":11936417.0,"Title":"Pass html form input as argument for python function (recommender system)","Body":"<p>I have built a recommender system using a Keras CNN. I am deploying it with Flask as a web application. Currently it displays images in the form of a gallery. I want to use html form to click the image and then generate recommendations from the image clicked. <\/p>\n\n<p>I am using <code>PIL<\/code> to pass into the model. So far everything works fine, I'm a bit stuck on how to make it pass into the function and return the results. Thank you!  I'm new and this is my first go around with Flask. <\/p>\n\n<p>Tried several ways to pass the form, however not sure if I have been doing it the correct way. <\/p>\n\n<p>Also have tried several variations on the function.<br>\nI want to pass the image clicked into my python function, then return back the related images. So far I have this for <code>index.html<\/code>:<\/p>\n\n<pre><code>&lt;form action=\"{{ url_for('predict')}}\" method=\"POST\"&gt;\n     &lt;div class=\"gallery\"&gt;\n        {% for image in images %}\n            &lt;button type=\"submit\" class=\"btn btn-primary\"&gt; &lt;img \n             src=\"static\/img\/{{ image }}\"&gt;&lt;\/button&gt;\n        {% endfor %}\n    &lt;\/div&gt;\n&lt;\/form&gt;\n#Here is my recommender function:\n\n@app.route('\/predict', methods=['POST'])\ndef predict():\n    if request.method == 'POST':\n        product = request.form['product']\n\n        # compute cosine similarities between images\n        cosine = cosine_similarity(imgs_features)\n\n        # creates a dataframe of the cosine similarites\n        cosine_df = pd.DataFrame(cosine, columns=products, \n        index=products)\n\n        sim_images = 5\n\n        p_name = os.path.basename(str(product))\n        image = load_img(product, target_size=(img_size, img_size))\n        plt.imshow(image)\n        plt.show()\n        recommend_imgs = cosine_df[product].sort_values(ascending=False) \n        [1:sim_images + 1].index\n        for i in range(0, len(recommend_imgs)):\n            image = load_img(recommend_imgs[i], target_size=(img_size, \n            img_size))\n            # plt.imshow(image)\n            # plt.show()\n            rec_name = os.path.basename(str(recommend_imgs[i]))\n        return render_template('results.html', name=product, \n        res=rec_name, pred_imgs=plt.imshow(image))\n<\/code><\/pre>\n\n<p>I get this error mostly: <\/p>\n\n<blockquote>\n  <p>werkzeug.exceptions.HTTPException.wrap..newcls: 400 Bad Request: KeyError: 'product'<\/p>\n<\/blockquote>\n\n<p>Along with: <\/p>\n\n<pre><code>return self.wsgi_app(environ, start_response)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57946038","CreationDate":"2019-09-15T16:25:48.967","ParentId":null,"OwnerUserId":"8810723","Title":null,"Body":"<p>This is one way to start the function. I am not sure how to pass the image value that you need to the form, unless each image as a specific value name you can pass it to the function. <\/p>\n\n<p>Html: <\/p>\n\n<pre><code>{% for image in images %}\n\n&lt;form action=\"{{ url_for('predict')}}\" method=\"POST\"&gt;\n\n    &lt;button type=\"submit\" name=\"submit\" value=\"{{ image.value}}\"&gt; \n\n      &lt;img src=\"static\/img\/{{ image }}\"&gt;\n\n    &lt;\/button&gt;\n\n&lt;\/form&gt;\n\n{% endfor %}\n<\/code><\/pre>\n\n<p>python:<\/p>\n\n<pre><code>@app.route('\/predict', methods=['POST'])\ndef predict():\n    if request.method == 'POST':\n        if request.form['submit'] == '&lt;image.value_1&gt;':\n            # do something\n            ...\n\n\n        elif request.form['submit'] == '&lt;image.value_2&gt;':\n            # do something\n            ...\n        ...\n<\/code><\/pre>\n\n<p>well, that's all.             <\/p>\n"}]}
+{"QuestionId":57881350,"AnswerCount":2,"Tags":"<python><keras><deep-learning><loss-function>","CreationDate":"2019-09-11T02:42:36.983","AcceptedAnswerId":57894061.0,"OwnerUserId":12050636.0,"Title":"Is there a way to put a customized keras model inside a python class method?","Body":"<p>I'm trying to build a deep neural network in python using a class method. ((The main idea behind it is to try to customize the loss function later on))\nI'm trying to use Keras in the function that defines the neural network structure but it doesn't seem to be working. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># create a class to\nclass PGNN(keras.Sequential):\n\n    def __init__(self,x,y):\n        super().__init__()\n\n        X = np.concatenate([x,y], axis=1)\n        self.X = X\n        self.x = X[:,0:1]\n        self.y = X[:,1:2]\n\n    def build_model_u(self):\n        model_u = models.Sequential\n        model_u.add(layers.Dense(64, activation='tanh', input_shape= 1000))\n        model_u.add(layers.Dense(32, activation='tanh'))\n        model_u.add(layers.Dense(16, activation='tanh'))\n        model_u.add(layers.Dense(8, activation='tanh'))\n        model_u.add(layers.Dense(4, activation='tanh'))\n        model_u.add(layers.Dense(1))\n        model_u.compile(optimizer='Adam', loss='mse', metrics=['mae'])\n\n\n    def train(self, x_train, y_train):\n        model = build_model_u(self)\n        model.fit()\n\n    def predict(self, x_test):\n        model.predict(x_test)\n\n    def validation(self, x_test, y_test):\n        model.evaluate(x_test,y_test, verbose=2)\n<\/code><\/pre>\n\n<p>I expected the model to start training when I call model.fit(x_train,y_train) but I always get the error \"build_model_u is not defined\" <\/p>\n\n<pre><code>model = build_model_u(self)\n<\/code><\/pre>\n\n<p>NameError: name 'build_model_u' is not defined<\/p>\n","answers":[{"AnswerId":"57888317","CreationDate":"2019-09-11T11:37:29.560","ParentId":null,"OwnerUserId":"11368194","Title":null,"Body":"<p>You must be calling the method like<\/p>\n\n<pre><code>model = build_model_u\nmodel.fit(x_train,y_train)\n<\/code><\/pre>\n\n<p>Call the function like this<\/p>\n\n<pre><code>model = build_model_u()\nmodel.fit(x_train,y_train)\n<\/code><\/pre>\n\n<p>and also you can remove the output variable<\/p>\n\n<pre><code>output= model_u.add(layers.Dense(3))\n<\/code><\/pre>\n\n<p>and keep it this way<\/p>\n\n<pre><code>model_u.add(layers.Dense(3))\n<\/code><\/pre>\n"},{"AnswerId":"57894061","CreationDate":"2019-09-11T17:24:06.360","ParentId":null,"OwnerUserId":"7801852","Title":null,"Body":"<p>Yes there is way. You have to extends your custom model class with keras.Model and overtire call method then you can call fit method from your own custom class.\nYou can follow the following Keras documentations.\n<a href=\"https:\/\/keras.io\/models\/about-keras-models\/#model-subclassing\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/models\/about-keras-models\/#model-subclassing<\/a><\/p>\n"}]}
+{"QuestionId":57881499,"AnswerCount":0,"Tags":"<python><machine-learning><nlp><pytorch>","CreationDate":"2019-09-11T03:00:22.980","AcceptedAnswerId":null,"OwnerUserId":11972508.0,"Title":"Reconciling classifier input requirements with tensor dimensions","Body":"<p>New to machine learning, so please bear with me. <\/p>\n\n<p>I'm trying to train a convolutional neural network to classify Rotten Tomatoes reviews as either positive or negative. The original dataframe itself is not too complicated. I added a column indicating whether the row was part of the training, validation, or test splits and cleaned up the text, but aside from that it looks more or less like the below: <\/p>\n\n<pre><code>   freshness                                             review\n0          1   Manakamana doesn't answer any questions, yet ...\n1          1   Wilfully offensive and powered by a chest-thu...\n2          0   It would be difficult to imagine material mor...\n3          0   Despite the gusto its star brings to the role...\n4          0   If there was a good idea at the core of this ... \n<\/code><\/pre>\n\n<p>What's hard for me to understand is how to figure out what the proper dimensions are for the inputs that go into my classifier. The relevant code is below. <\/p>\n\n<p>Class definitions:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class ReviewDataset(Dataset):\n    def __init__(self, review_df, vectorizer):\n        self.review_df = review_df\n        self.vectorizer_ = vectorizer\n\n        self.train_df = self.review_df[self.review_df[\"split\"] == \"train\"]\n        self.train_size = len(self.train_df)\n\n        self.val_df = self.review_df[self.review_df[\"split\"] == \"val\"]\n        self.val_size = len(self.val_df)\n\n        self.test_df = self.review_df[self.review_df[\"split\"] == \"test\"]\n        self.test_size = len(self.test_df)\n\n        self.lookup_dict_ = {\n            \"train\": (self.train_df, self.train_size),\n            \"val\": (self.val_df, self.val_size),\n            \"test\": (self.test_df, self.test_size)\n        }\n\n        self.set_split(\"train\")\n\n    @classmethod\n    def load_and_vectorize(cls, review_csv, sample=True):\n        review_df = pd.read_csv(review_csv)\n\n        if sample != True:\n            review_df.columns = [ col.lower() for col in review_df.columns ]\n            review_df = split_into_sets(review_df)\n            review_df[\"review\"] = review_df[\"review\"].apply(clean_text)\n            return cls(review_df, Vectorizer.from_dataframe(review_df))\n        else:\n            # sample data to facilitate model tuning\n            sampleData = review_df.sample(frac=.1, random_state=1)\n            sampleData.columns = [ col.lower() for col in sampleData.columns ]\n            # remove class imbalance within sample\n            posRating = sampleData[sampleData[\"freshness\"] == 1]\n            negRating = sampleData[sampleData[\"freshness\"] == 0]\n            neg_downsampled = resample(negRating,\n                                       replace=False,\n                                       n_samples=23926,\n                                       random_state=1)\n            balancedSample = pd.concat([neg_downsampled, posRating])\n            balancedSample = split_into_sets(balancedSample)\n            balancedSample[\"review\"] = balancedSample[\"review\"].apply(clean_text)\n            return cls(balancedSample, Vectorizer.from_dataframe(balancedSample))\n\n\n    def get_vectorizer(self):\n        return self.vectorizer_\n\n    def set_split(self, split=\"train\"):\n        self.target_split_ = split\n        self.target_df_, self.target_size_ = self.lookup_dict_[split]\n\n    def __len__(self):\n        return self.target_size_\n\n    def __getitem__(self, index):\n        row = self.target_df_.iloc[index]\n\n        review_vector = \\\n            self.vectorizer_.vectorize(row[\"review\"])\n\n        rating_index = \\\n            self.vectorizer_.rating_vocab.lookup_token(row[\"freshness\"])\n\n        return {\n            \"x_data\": review_vector,\n            \"y_target\": rating_index\n        }\n\n    def get_n_batches(self, batch_size):\n        return len(self) \/\/ batch_size\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>class Vocabulary(object):\n    def __init__(self, token_to_idx=None, add_unk=True, unk_token=\"&lt;UNK&gt;\"):\n        if token_to_idx is None:\n            token_to_idx = {}\n        self.token_to_idx_ = token_to_idx\n        self.idx_to_token_ = {\n            idx: token\n            for token, idx in self.token_to_idx_.items()\n        }\n        self.add_unk_ = add_unk\n        self.unk_token_ = unk_token\n        self.unk_index = -1\n        if self.add_unk_:\n            self.unk_index_ = self.add_token(unk_token)\n\n    def to_serializable(self):\n        return {\n            \"token_to_idx\": self.token_to_idx_,\n            \"add_unk\": self.add_unk_,\n            \"unk_token\": self.unk_token_\n        }\n\n    @classmethod\n    def from_serializable(cls, contents):\n        return cls(**contents)\n\n    def add_token(self, token):\n        if token in self.token_to_idx_:\n            index = self.token_to_idx_[token]\n        else:\n            index = len(self.token_to_idx_)\n            self.token_to_idx_[token] = index\n            self.idx_to_token_[index] = token\n        return index\n\n    def lookup_token(self, token):\n        if self.add_unk_:\n            return self.token_to_idx_.get(token, self.unk_index)\n        else:\n            return self.token_to_idx_[token]\n\n    def lookup_index(self, index):\n        if index not in self.idx_to_token_:\n            raise KeyError(\"index (%d) is not in the Vocabulary\" % index)\n        return self.idx_to_token_[index]\n\n    def __str__(self):\n        return \"&lt;Vocabulary(size=%d)&gt;\" % len(self)\n\n    def __len__(self):\n        return len(self.token_to_idx_)\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>class Vectorizer(object):\n    def __init__(self, review_vocab, rating_vocab):\n        self.review_vocab = review_vocab\n        self.rating_vocab = rating_vocab\n\n    def vectorize(self, review):\n        # create collapsed one-hot vector for new observations\n        one_hot = np.zeros(len(self.review_vocab), dtype=np.float64)\n\n        for token in review.split(\" \"):\n            if token not in string.punctuation:\n                one_hot[self.review_vocab.lookup_token(token)] = 1\n\n        return one_hot\n\n    @classmethod\n    def from_dataframe(cls, review_df, cutoff=10):\n        review_vocab = Vocabulary(add_unk=True)\n        rating_vocab = Vocabulary(add_unk=False)\n\n        for rating in sorted(set(review_df[\"freshness\"])):\n            rating_vocab.add_token(rating)\n\n        word_counts = Counter()\n        for review in review_df[\"review\"]:\n            for word in review.split(\" \"):\n                if word not in string.punctuation:\n                    word_counts[word] += 1\n\n        for word, count in word_counts.items():\n            if count &gt; cutoff:\n                review_vocab.add_token(word)\n\n        return cls(review_vocab, rating_vocab)\n\n    @classmethod\n    def from_serializable(cls, contents):\n        review_vocab = Vocabulary.from_serializable(contents[\"review_vocab\"])\n        rating_vocab = Vocabulary.from_serializable(contents[\"rating_vocab\"])\n\n        return cls(review_vocab=review_vocab, rating_vocab=rating_vocab)\n\n    def to_serializable(self):\n        return {\n            \"review_vocab\": self.review_vocab.to_serializable(),\n            \"rating_vocab\": self.rating_vocab.to_serializable()\n        }\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>class Classifier(nn.Module):\n    def __init__(self, initial_n_channels, n_classes, network_n_channels):\n        super(Classifier, self).__init__()\n        self.network = nn.Sequential(\n            nn.Conv1d(in_channels=initial_n_channels,\n                      out_channels=network_n_channels,\n                      kernel_size=args[\"kernel_size\"]),\n            nn.ReLU(),\n            nn.Conv1d(in_channels=network_n_channels,\n                      out_channels=network_n_channels,\n                      kernel_size=args[\"kernel_size\"],\n                      stride=args[\"stride\"]),\n            nn.ReLU(),\n            nn.Conv1d(in_channels=network_n_channels,\n                      out_channels=network_n_channels,\n                      kernel_size=args[\"kernel_size\"],\n                      stride=args[\"stride\"]),\n            nn.ReLU(),\n            nn.Conv1d(in_channels=network_n_channels,\n                      out_channels=network_n_channels,\n                      kernel_size=args[\"kernel_size\"],\n                      stride=args[\"stride\"]),\n            nn.ReLU()\n        )\n        self.fc = nn.Linear(network_n_channels, n_classes)\n\n    def forward(self, x_in, apply_sigmoid=False):\n        features = self.network(x_in) # squeeze?\n        prediction_vector = self.linear(features)\n        if apply_sigmoid:\n            prediction_vector = F.sigmoid(prediction_vector, dim=1)\n        return prediction_vector\n<\/code><\/pre>\n\n<p>Instantiation:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># dataset and vectorizer\ndataset = ReviewDataset.load_and_vectorize(args[\"review_csv\"])\nvectorizer = dataset.get_vectorizer()\n\n# model\nclassifier = Classifier(initial_n_channels=len(vectorizer.review_vocab),\n                        n_classes=len(vectorizer.rating_vocab),\n                        network_n_channels=args[\"num_channels\"])\n\n# loss and optimizer\nloss_func = nn.BCEWithLogitsLoss()\noptimizer = optim.Adam(classifier.parameters(), lr=args[\"learning_rate\"])\n<\/code><\/pre>\n\n<p>The result of running the above code: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>RuntimeError: Expected 3-dimensional input for 3-dimensional weight 128 7882, but got 2-dimensional input of size [128, 7882] instead\n<\/code><\/pre>\n\n<p>I figured a two-dimensional tensor in the shape <code>(batch_size, len(review_vocab))<\/code> would be good, but apparently my CNN requires a third dimension to the data? What can I do to add the extra dimension, or change what the classifier itself requires?<\/p>\n","answers":[]}
+{"QuestionId":57881527,"AnswerCount":1,"Tags":"<python><keras><neural-network>","CreationDate":"2019-09-11T03:03:51.333","AcceptedAnswerId":57881607.0,"OwnerUserId":11580689.0,"Title":"Dropout entire input layer","Body":"<p>Suppose I have two inputs (each with a number of features), that I want to feed into a <code>Dropout<\/code> layer. I want each iteration to drop out a whole input, with all of its associated features, and keep the whole of the other input.<\/p>\n\n<p>After concatenating the inputs, I think I need to use the <code>noise_shape<\/code> parameter for <code>Dropout<\/code>, but the shape of the concatenated layer doesn't really let me do that. For two inputs of shape (15,), the concatenated shape is (None, 30), rather than (None, 15, 2), so one of the axes is lost and I can't drop out along it.<\/p>\n\n<p>Any suggestions for what I could do? Thanks.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.layers import Input, concatenate, Dense, Dropout\n\nx = Input((15,))  # 15 features for the 1st input\ny = Input((15,))  # 15 features for the 2nd input\nxy = concatenate([x, y])\nprint(xy._keras_shape)\n# (None, 30)\n\nlayer = Dropout(rate=0.5, noise_shape=[xy.shape[0], 1])(xy)\n...\n<\/code><\/pre>\n","answers":[{"AnswerId":"57881607","CreationDate":"2019-09-11T03:13:48.863","ParentId":null,"OwnerUserId":"10841164","Title":null,"Body":"<p>EDIT :<\/p>\n\n<p>Seems like I misunderstood your question, here is the updated answer based on your requirement.<\/p>\n\n<p>To achieve what you want, x and y effectively become the timesteps, and according to Keras documentation, <code>noise_shape=(batch_size, 1, features)<\/code> if your input shape is <code>(batch_size, timesteps, features)<\/code>:<\/p>\n\n<pre><code>x = Input((15,1))  # 15 features for the 1st input\ny = Input((15,1))  # 15 features for the 2nd input\nxy = concatenate([x, y])\n\ndropout_layer = Dropout(rate=0.5, noise_shape=[None, 1, 2])(xy)\n...\n<\/code><\/pre>\n\n<p>To test that you are getting the correct behavior, you can inspect the intermediate <code>xy<\/code> layer and <code>dropout_layer<\/code> using the following code (<a href=\"https:\/\/stackoverflow.com\/questions\/41711190\/keras-how-to-get-the-output-of-each-layer\">reference link<\/a>):<\/p>\n\n<pre><code>### Define your model ###\n\nfrom keras.layers import Input, concatenate, Dropout\nfrom keras.models import Model\nfrom keras import backend as K\n\n# Learning phase must be set to 1 for dropout to work\nK.set_learning_phase(1)\n\nx = Input((15,1))  # 15 features for the 1st input\ny = Input((15,1))  # 15 features for the 2nd input\nxy = concatenate([x, y])\n\ndropout_layer = Dropout(rate=0.5, noise_shape=[None, 1, 2])(xy)\n\nmodel = Model(inputs=[x,y], output=dropout_layer)\n\n# specify inputs and output of the model\n\nx_inp = model.input[0]                                           \ny_inp = model.input[1]\noutp = [layer.output for layer in model.layers[2:]]        \nfunctor = K.function([x_inp, y_inp], outp)\n\n### Get some random inputs ###\n\nimport numpy as np\n\ninput_1 = np.random.random((1,15,1))\ninput_2 = np.random.random((1,15,1))\n\nlayer_outs = functor([input_1,input_2])\nprint('Intermediate xy layer:\\n\\n',layer_outs[0])\nprint('Dropout layer:\\n\\n', layer_outs[1])\n\n<\/code><\/pre>\n\n<p>You should see that the entire x or y are dropped randomly (50% chance) per your requirement:<\/p>\n\n<pre><code>Intermediate xy layer:\n\n [[[0.32093528 0.70682645]\n  [0.46162075 0.74063486]\n  [0.522718   0.22318116]\n  [0.7897043  0.7849486 ]\n  [0.49387926 0.13929296]\n  [0.5754296  0.6273373 ]\n  [0.17157765 0.92996144]\n  [0.36210892 0.02305864]\n  [0.52637625 0.88259524]\n  [0.3184462  0.00197006]\n  [0.67196816 0.40147918]\n  [0.24782693 0.5766827 ]\n  [0.25653633 0.00514544]\n  [0.8130438  0.2764429 ]\n  [0.25275478 0.44348967]]]\n\nDropout layer:\n\n [[[0.         1.4136529 ]\n  [0.         1.4812697 ]\n  [0.         0.44636232]\n  [0.         1.5698972 ]\n  [0.         0.2785859 ]\n  [0.         1.2546746 ]\n  [0.         1.8599229 ]\n  [0.         0.04611728]\n  [0.         1.7651905 ]\n  [0.         0.00394012]\n  [0.         0.80295837]\n  [0.         1.1533654 ]\n  [0.         0.01029088]\n  [0.         0.5528858 ]\n  [0.         0.88697934]]]\n<\/code><\/pre>\n\n<p>If you are wondering why all the elements are multiplied by 2, take a look at how tensorflow implemented dropout <a href=\"http:\/\/laid.delanover.com\/dropout-explained-and-implementation-in-tensorflow\/\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>Hope this helps.<\/p>\n"}]}
+{"QuestionId":57881799,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-11T03:40:00.997","AcceptedAnswerId":null,"OwnerUserId":5026962.0,"Title":"How to define custom gradient for keras layer with tensorflow 1.14","Body":"<p>I am trying to implement gradient reversal layer in tf.keras (tf ver 1.14). I defined a custom gradient using tf.custom_gradient but the grad function is never called and the result is wrong.<\/p>\n\n<p>I am using the following to test the layer<\/p>\n\n<pre><code>from tensorflow.keras.models import Model\nfrom tensorflow.keras.layers import Input, Dense\nfrom tensorflow.keras import backend as k\nimport tensorflow as tf\ntf.enable_eager_execution()\nimport numpy as np\n\n@tf.custom_gradient\ndef custom_op(x):\n    result = x\n    def custom_grad(dy):\n        grad = tf.negative(dy)\n        return grad\n    return result, custom_grad\n\nclass GradientReversal(tf.keras.layers.Layer):\n  def __init__(self):\n    super(GradientReversal, self).__init__()\n\n  def call(self, inputs):\n    return custom_op(inputs)\n\n\na = Input(shape=(2,))\nb = GradientReversal()(a)\nc = Dense(1)(b)\nmodel = Model(inputs=a, outputs=c)\nmodel.compile(loss='binary_crossentropy', optimizer='sgd')\n\n@tf.function\ndef test():  \n  with tf.GradientTape() as tape:\n    out = model(np.ones((1, 2)))\n  gradients = tape.gradient(out, model.trainable_variables)\n  return gradients\n\nprint(test())\n<\/code><\/pre>\n","answers":[{"AnswerId":"57892130","CreationDate":"2019-09-11T15:11:59.650","ParentId":null,"OwnerUserId":"5026962","Title":null,"Body":"<p>Nevermind, I was supposed to use GradientReversal layer after the Dense to get the result I want.<\/p>\n"}]}
+{"QuestionId":57881840,"AnswerCount":1,"Tags":"<python><tensorflow><keras><loss-function>","CreationDate":"2019-09-11T03:44:55.093","AcceptedAnswerId":null,"OwnerUserId":12042308.0,"Title":"How to use different loss function in training and test phase","Body":"<p>I want to use a weighted loss function in the training phase and an unweighted loss function in the test phase. <\/p>\n\n<p>I've tried in_train_phase, but in the test phase, it still uses weighted loss function.<\/p>\n\n<pre><code>def custom_loss():  \n        def euclidian_distance_loss(y_true, y_pred):\n        edl_in_train = K.sqrt(K.sum(K.square((y_pred - y_true) * weight_matrix().reshape(1, 41, 41, 1))))\n        edl = K.sqrt(K.sum(K.square(y_pred - y_true)))\n        return K.in_train_phase(edl_in_train,edl)\n\n    return euclidian_distance_loss\n<\/code><\/pre>\n\n<p>Because the shape of the training data is constant and the test is not so the error came out at the start of the test phase:<\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors_impl.InvalidArgumentError: Incompatible shapes<\/p>\n<\/blockquote>\n\n<p>How can I make this code work or is there any other way to use different loss functions in these 2 phases?<\/p>\n","answers":[{"AnswerId":"57890130","CreationDate":"2019-09-11T13:17:56.363","ParentId":null,"OwnerUserId":"3472735","Title":null,"Body":"<p>At test time, as you're not minimising on the loss function, it doesn't matter exactly what goes in the loss function. What I would do is feed the unweighted loss function as a \"metric\" - this way it won't impact on training at train time, but at test time you can analyse it yourself to see how your model is dong. <\/p>\n"}]}
+{"QuestionId":57882172,"AnswerCount":0,"Tags":"<python><keras><neural-network>","CreationDate":"2019-09-11T04:29:57.310","AcceptedAnswerId":null,"OwnerUserId":11580689.0,"Title":"Check dropped nodes","Body":"<p>I have a NN model with <code>Dropout<\/code> layers. I'd like to check which nodes of the network have been dropped at each iteration. Or, which weights have been set to 0. <\/p>\n\n<p>I tried looking at the model weights after one iteration, but it didn't contain any zeros, presumably because the dropped nodes don't transfer to the final model.<\/p>\n\n<p>This question has been asked <a href=\"https:\/\/stackoverflow.com\/q\/49630178\">before<\/a>, but it didn't get an answer.<\/p>\n","answers":[]}
+{"QuestionId":57882443,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-09-11T05:06:30.940","AcceptedAnswerId":null,"OwnerUserId":12051039.0,"Title":"Keras model with high accuracy but low val_acc","Body":"<p>I am using resnet50 transfer learning for the Oxford-IIIT Pet Dataset to classify 37 breeds of cats and dogs. The idea is to follow fastai implementation closely using Keras code. However, I managed to get a training accuracy as high as 90% but can't seem to increase my val_accuracy higher than a random guess (1\/37 or ~ 3% val_acc).<\/p>\n\n<p>Any idea how do Keras compute the validation acc and how can I improve it? Or is there something wrong with my preprocessing steps? Thanks a lot.<\/p>\n\n<p>To get my validation steps, I use sklearn StratifiedShuffleSplit to get a balanced validation set.<\/p>\n\n<pre><code># Create dataframe with labels and filenames\nannotations = pd.read_csv(\"annotation\/list.txt\",header=None,delim_whitespace=True)\n\nannotations.drop([1,2,3],axis=1, inplace=True)\nannotations.columns = [\"filenames\"]\n\n# Create label columns\ntrans = str.maketrans(\"_0123456789\",\"           \")\nannotations[\"labels\"] = annotations[\"filenames\"].str.translate(trans).str.strip()\nannotations[\"filenames\"] = annotations[\"filenames\"] +\".jpg\" \n\n<\/code><\/pre>\n\n<pre><code># Creating a validation set\nfrom sklearn.model_selection import StratifiedShuffleSplit\n\ndf_array = annotations.to_numpy(copy=True)\n\nsss = StratifiedShuffleSplit(n_splits = 1, test_size=0.2)\n\nvalid_idx = [test for _,test in sss.split(df_array[:,0],df_array[:,1])]\n\nvalidation = annotations.iloc[valid_idx[0]]\nannotations.drop(valid_idx[0], inplace=True)\n<\/code><\/pre>\n\n<p>Then, constructing my generator and training my model.<\/p>\n\n<pre><code>from tensorflow.keras.preprocessing.image import ImageDataGenerator\n\nbs = 64\n\ndef normalize(x):\n  imagenet_mean = np.array([0.485, 0.456, 0.406]).reshape(1,1,3)\n  imagenet_sd = np.array([0.229, 0.224, 0.225]).reshape(1,1,3)\n  return (x- imagenet_mean)\/imagenet_sd\n\n\n\ntrain_datagen = ImageDataGenerator(rescale=1\/255.,\n                                  horizontal_flip = True,\n                                  rotation_range=10,\n                                  width_shift_range = 0.1,\n                                  height_shift_range =0.1,\n                                  brightness_range =(0.9,1.1),\n                                  shear_range =0.1,\n                                  preprocessing_function=normalize)\n\ntrain_generator = train_datagen.flow_from_dataframe(dataframe=annotations,\n                                                   directory =os.getcwd(),\n                                                   x_col=\"filenames\",\n                                                   y_col=\"labels\",\n                                                   target_size = (224,224),\n                                                   batch_size = bs,\n                                                   )\n\n\nval_datagen = ImageDataGenerator(rescale=1\/255.,\n                                preprocessing_function=normalize)\n\nvalidation_generator = val_datagen.flow_from_dataframe(dataframe=validation,\n                                                   directory =os.getcwd(),\n                                                   x_col=\"filenames\",\n                                                   y_col=\"labels\",\n                                                   target_size = (224,224),\n                                                   batch_size=bs,\n                                                   )\n\n\nfrom tensorflow.keras.applications.resnet50 import ResNet50\nfrom tensorflow.keras import optimizers\nfrom tensorflow.keras import Model\nfrom tensorflow.keras.layers import Dense, GlobalAveragePooling2D, Flatten, BatchNormalization, Dropout\n\n\nbase_model = ResNet50(include_top=False,weights=\"imagenet\")\nx = base_model.output\nx = GlobalAveragePooling2D()(x)\nx = Flatten()(x)\nx = BatchNormalization(epsilon=1e-05,momentum=0.1)(x)\nx = Dropout(0.25)(x)\nx = Dense(512,activation=\"relu\")(x)\nx = BatchNormalization(epsilon=1e-05,momentum=0.1)(x)\nx = Dropout(0.5)(x)\npredictions = Dense(37,activation=\"softmax\")(x)\n\nmodel = Model(inputs=base_model.input,outputs=predictions)\n\nfor layer in base_model.layers:\n    layer.trainable = False\n\nlr= 0.001\nopti = optimizers.Adam(lr=lr, decay=lr\/50)\nmodel.compile(optimizer=opti,\n             loss=\"categorical_crossentropy\",\n             metrics=[\"accuracy\"])\n\nmodel.fit_generator(train_generator,\n                   epochs=10,\n                   validation_data = validation_generator)\n\nfor layer in base_model.layers:\n    layer.trainable = True\n\nmodel.fit_generator(train_generator,\n                   epochs=10,\n                   validation_data = validation_generator)\n\n<\/code><\/pre>\n\n<pre><code>By the 10 epochs before unfreezing my layers\n\nloss = 0.2189\nacc = 0.9255\nval_loss = 5.5082\nval_acc = 0.0401\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57882924,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><while-loop>","CreationDate":"2019-09-11T05:54:57.217","AcceptedAnswerId":null,"OwnerUserId":12051197.0,"Title":"TypeError when using \"tf.while_loop\" : 'Tensor' object does not support item assignment","Body":"<p>I am training an unsupervised CNN, for this, I defined a loss function in which CNN input is mapped to the cnn output through a complicated expression. I am using <code>tf.while_loop<\/code> to update the variables of <code>gamma_tilde_tensor_local<\/code> and <code>D_mat_tensor_local<\/code>. I have uploaded my code &amp; the error below.\nplease suggest me how to update tensor variables in <code>while_loop<\/code>.<\/p>\n\n<p>To avoid confusion, I have avoided the complex loops. originally <code>gamma_tilde_tensor_local<\/code> size <code>(batch_size, M,M,K,K)<\/code>.<\/p>\n\n<pre><code>def function(self):\n    gamma_tilde_tensor_local = tf.Variable(tf.zeros(shape = [self.mini_batchSize, self.num_users]),dtype=tf.float32)\n    D_mat_tensor_local = tf.Variable(tf.zeros(shape = [self.mini_batchSize, , self.num_users]), dtype=tf.float32)\n    condition_g1  = lambda k_iter1, gamma_tilde_tensor_local, D_mat_tensor_local : k_iter1 &lt; self.num_users\n    def body_gm1( k_iter1, gamma_tilde_tensor_local, D_mat_tensor_local):\n        tf_x_k = tf.expand_dims(self.PHI_batch[:,:,k_iter1],axis=-1)          \n        tf_x_kd = tf.expand_dims(self.PHI_batch[:,:,k_iter2],axis=-1)        \n        phi_to_reduce =  tf.matmul( tf.transpose(tf_x_kd, perm=[0, 2, 1]), tf_x_k)\n        phi_t = tf.squeeze(phi_to_reduce,[1,2])\n        gamma_till = tf.divide(  tf.multiply( self.gammafun[:,1, 2 ], self.channel_Gain[:,1, k_iter1]), self.channel_Gain[:,1, 2])\n        gamma_tilde_tensor_local[: k_iter1] = tf.multiply(phi_t, gamma_till )    \n        D_mat_tensor_local[:,k_iter1] = tf.sqrt(  tf.multiply( self.gammafun[1, 2], self.channel_Gain[1,k_iter1] ))                 \n\n        return tf.add(k_iter1, 1), gamma_tilde_tensor_local, D_mat_tensor_local\n    k_iter1, gamma_tilde_tensor_local, D_mat_tensor_local = tf.while_loop(condition_g1, body_gm1, [0, gamma_tilde_tensor_local, D_mat_tensor_local])\n<\/code><\/pre>\n\n<p>Giving error at the line:\n<code>gamma_tilde_tensor_local[: k_iter1] = tf.multiply(phi_t, gamma_till )<\/code>.<\/p>\n","answers":[]}
+{"QuestionId":57883517,"AnswerCount":0,"Tags":"<python><keras><sequence><lstm>","CreationDate":"2019-09-11T06:46:10.363","AcceptedAnswerId":null,"OwnerUserId":7959519.0,"Title":"Is there a method for model to know about sequence length?","Body":"<p>I want to train my LSTM model with input shape (num of samples, timesteps, features) like [a, b, c] is a one input at t time.\nand then [b, c, d] is next input at t+1 time. <\/p>\n\n<p>My data like this.<\/p>\n\n<pre><code>[ \n  [ [1, 2, 3], [2, 3, 4] ], # &lt;- this is first data for training\n  [ [3, 4, 5], [3, 1, 2], [2, 4, 3] ],\n  [ [1, 1, 4], [1, 2, 6], [1, 5, 2], [1, 2, 5], [1, 3, 2] ]\n  [ [1, 2, 3] ]\n]\n<\/code><\/pre>\n\n<p>Every data has different length (different timesteps at every data)<\/p>\n\n<p>I append some padding in my data.<\/p>\n\n<pre><code>[ \n  [ [1, 2, 3], [2, 3, 4], [0, 0, 0], [0, 0, 0], [0, 0, 0] ]\n  [ [3, 4, 5], [3, 1, 2], [2, 4, 3], [0, 0, 0], [0, 0, 0] ],\n  [ [1, 1, 4], [1, 2, 6], [1, 5, 2], [1, 2, 5], [1, 3, 2] ]\n  [ [1, 2, 3], [0, 0, 0], [0, 0, 0], [0, 0, 0], [0, 0, 0] ]\n]\n<\/code><\/pre>\n\n<p>It trains well, but I Want to let the model know about sequence length without [0, 0, 0] because [0, 0, 0] is not worthy data.<\/p>\n\n<pre><code> sequence.pad_sequences(all_data, padding='post', maxlen = max_sequence_length)\n<\/code><\/pre>\n\n<p>for this code, I append some padding for my data.<\/p>\n","answers":[]}
+{"QuestionId":57883804,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-11T07:06:01.063","AcceptedAnswerId":null,"OwnerUserId":6746235.0,"Title":"How to modify the architecture of a custom Estimator after training?","Body":"<p>I build a custom Estimator and train it for several steps. Now I want to modify the architecture (e.g. add or remove layers) of the Estimator. Can somebody give an example of codes?<\/p>\n\n<p>If Keras model is used, I can easily modify the architecture of the Keras model by its API. However, I am unfamiliar with APIs of Estimator.<\/p>\n","answers":[]}
+{"QuestionId":57884632,"AnswerCount":0,"Tags":"<python><tensorflow><onnx>","CreationDate":"2019-09-11T07:59:36.793","AcceptedAnswerId":null,"OwnerUserId":8638900.0,"Title":"while transfer onnx model to tensorflow","Body":"<p>ValueError: Tensor conversion requested dtype float32 for Tensor with dtype float64<\/p>\n\n<p>TypeError: Input 'y' of 'Add' Op has type float64 that does not match type float32 of argument 'x'<\/p>\n","answers":[]}
+{"QuestionId":57885165,"AnswerCount":0,"Tags":"<python><python-3.x><keras><python-imaging-library><raster>","CreationDate":"2019-09-11T08:32:11.123","AcceptedAnswerId":null,"OwnerUserId":4986143.0,"Title":"How should I structure these raster inputs for my CNN?","Body":"<p>I am following <a href=\"https:\/\/www.pyimagesearch.com\/2019\/02\/04\/keras-multiple-inputs-and-mixed-data\/\" rel=\"nofollow noreferrer\">this guide from pyimagesearch<\/a> to make a CNN which takes images and numeric data and make a single output. On the image side, this guide offers four colored images per sample, then loads, resizes, and stacks them in a montage as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import cv2\n\ninputImages = []\noutputImage = np.zeros((64, 64, 3), dtype=\"uint8\")\n\n# loop over the input image paths\n    for inputImage in imagePaths:\n    # load the input image, resize it to be 32 32, and then\n    # update the list of input images\n    image = cv2.imread(inputImage)\n    image = cv2.resize(image, (32, 32))\n    inputImages.append(image)\n\n# tile the four input images in the output image such the first\n# image goes in the top-right corner, the second image in the\n# top-left corner, the third image in the bottom-right corner,\n# and the final image in the bottom-left corner\noutputImage[0:32, 0:32] = inputImages[0]\noutputImage[0:32, 32:64] = inputImages[1]\noutputImage[32:64, 32:64] = inputImages[2]\noutputImage[32:64, 0:32] = inputImages[3]\n<\/code><\/pre>\n\n<p>The result is a montage is 4 colored images side by side like this:\n<img src=\"https:\/\/www.pyimagesearch.com\/wp-content\/uploads\/2019\/01\/keras_regression_cnns_houses.jpg\" alt=\"image montage\"><\/p>\n\n<p>My images are rasters and as such only have a single layers each. My four images represent different indices of the same place in an aerial shot, so I would have them stack on top of each other rather than side by side.\nIn the line:<\/p>\n\n<pre><code>outputImage = np.zeros((64, 64, 3), dtype=\"uint8\")\n<\/code><\/pre>\n\n<p>I presume the dimension <code>3<\/code> refers to the RGB element of the original images.\nTo make my own montage I reduced this to just <code>np.zeros((64, 64), dtype=\"uint8)<\/code> but my resulting <code>outputImage<\/code> wasn't an image montage as suggested, but instead just a 64X64 array of numbers.<\/p>\n\n<p>To load my rasters I am using Pillow and my code currently looks as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from PIL import Image\n\ninputImages = []\noutputImage = np.zeros((64, 64), dtype = 'uint8')\n\n# relevant images\nfor inputMap in relevantMaps:\n    mapPath = os.path.sep.join([zonePath, inputMap])\n    # load input image, resize it to be 32 32, the update the input image list\n    image = Image.open(mapPath)\n    image = image.resize((32, 32))\n    inputImages.append(image)\n\n# tile the four input images in the output image such that:\n# first image -&gt; top left, second image -&gt; top right\n# third image -&gt; bottom left, fourth image -&gt; bottom right\noutputImage[0:32, 0:32] = inputImages[0]\noutputImage[32:64, 0:32] = inputImages[1]\noutputImage[0:32, 32:64] = inputImages[2]\noutputImage[32:64, 32:64] = inputImages[3]\n<\/code><\/pre>\n\n<p>My <code>outputImage<\/code> looks like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>array([[123, 125, 131, ...,   0,   0,   0],\n       [120, 125, 127, ...,   0,   0,   0],\n       [123, 127, 126, ...,   0,   0,   0],\n       ...,\n       [205, 153,  55, ..., 242, 242, 242],\n       [  4, 241,  28, ..., 242, 242, 242],\n       [243,  90, 211, ..., 242, 242, 242]], dtype=uint8)\n<\/code><\/pre>\n\n<p>Rather than an image montage.<\/p>\n\n<p>I face two problems:<\/p>\n\n<ol>\n<li><p>I'm not sure if I should stack my images next to each other like in the example, or on top of each other, as they are different layers of the same exact area.<\/p><\/li>\n<li><p>I am not sure how to stack the images and have them remain as images, my method turns them into matrices.<\/p><\/li>\n<\/ol>\n\n<p>I expect to get a usable image montage that I could then plop into a CNN as input.<\/p>\n","answers":[]}
+{"QuestionId":57885196,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><deep-learning><conv-neural-network>","CreationDate":"2019-09-11T08:33:16.010","AcceptedAnswerId":57890289.0,"OwnerUserId":2467772.0,"Title":"Why shape is changed in feeding to Tensorflow graph for Batch processing","Body":"<p>The graph needs to process in batch and input shapes are defined as <code>(None,60,80,19)<\/code> and <code>(None,60,80,38)<\/code>.<\/p>\n\n<p>The TensorFlow graph definition is as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def __init__(self, tf_config=None):\n     self.tensor_heatMat = tf.placeholder(\n         dtype=tf.float32, shape=(None, 60, 80, 19), name='heatMat_in')\n     self.tensor_pafMat = tf.placeholder(\n         dtype=tf.float32, shape=(None, 60, 80, 38), name='pafMat_in')\n     self.upsample_size = tf.placeholder(\n         dtype=tf.int32, shape=(2,), name='upsample_size')\n     self.tensor_heatMat_up = tf.image.resize_area(\n         self.tensor_heatMat, self.upsample_size, align_corners=False, name='upsample_heatmat')\n     self.tensor_pafMat_up = tf.image.resize_area(\n         self.tensor_pafMat, self.upsample_size, align_corners=False, name='upsample_pafmat')\n     smoother = Smoother({'data': self.tensor_heatMat_up}, 25, 3.0)\n     gaussian_heatMat = smoother.get_output()\n     max_pooled_in_tensor = tf.nn.pool(gaussian_heatMat, window_shape=(\n         3, 3), pooling_type='MAX', padding='SAME')\n     self.tensor_peaks = tf.where(tf.equal(\n         gaussian_heatMat, max_pooled_in_tensor), gaussian_heatMat, tf.zeros_like(gaussian_heatMat))\n     self.heatMat = self.pafMat = None\n     self.persistent_sess = tf.InteractiveSession()\n     self.persistent_sess.run(tf.variables_initializer(\n        [v for v in tf.global_variables() if\n         v.name.split(':')[0] in [x.decode('utf-8') for x in\n                                  self.persistent_sess.run(tf.report_uninitialized_variables())]\n         ])\n     )\n\n\ndef inference(self, heatmat, pafmat, upsample_size=4.0):\n    peaks, heatMat_up, pafMat_up = self.persistent_sess.run(\n        [self.tensor_peaks, self.tensor_heatMat_up, self.tensor_pafMat_up], feed_dict={\n            self.tensor_heatMat: [heatmat], self.tensor_pafMat: [pafmat], self.upsample_size: (240, 320)\n        })\n    peaks = peaks[0]\n    self.heatMat = heatMat_up[0]\n    self.pafMat = pafMat_up[0\n    humans = PoseEstimator.estimate_paf(peaks, self.heatMat, self.pafMat)\n    return humans\n<\/code><\/pre>\n\n<p>So <code>self.tensor_heatMat<\/code> and <code>self.tensor_pafMat<\/code> require batched tensors.<\/p>\n\n<p>My input data to those placeholders are:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>outputs = outputs.reshape(32, 60, 80, 57)\nheat_maps = outputs[:, :, :, : 19]\npuf_maps = outputs[:, :, :, 19:]     \nhumans = inference(heat_maps, puf_maps,4.0)\n<\/code><\/pre>\n\n<p><code>heat_maps<\/code> and <code>puf_maps<\/code> shapes are <code>(32, 60, 80, 19)<\/code> and <code>(32, 60, 80, 38)<\/code>.\nBut when I run the session with input tensors, there is an error:<\/p>\n\n<blockquote>\n  <p>ValueError: Cannot feed value of shape (1, 32, 60, 80, 19) for Tensor 'heatMat_in:0', which has shape '(?, 60, 80, 19)'<\/p>\n<\/blockquote>\n\n<p>What could be the issue?<\/p>\n","answers":[{"AnswerId":"57890289","CreationDate":"2019-09-11T13:26:16.433","ParentId":null,"OwnerUserId":"5652313","Title":null,"Body":"<p>The error is in the <code>feed_dict<\/code>. If you put the tensors in lists, TensorFlow interprets the list as the first dimension of the tensor. This is where the same <code>(1, 32, ...)<\/code> comes from. You should do instead<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>feed_dict={\n    self.tensor_heatMat: heatmat,\n    self.tensor_pafMat: pafmat,\n    self.upsample_size:(240,320)}\n<\/code><\/pre>\n\n<p>The first dimension will be 32 which is the variable batch that you leave as <code>None<\/code> in the model init method.<\/p>\n"}]}
+{"QuestionId":57885446,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-09-11T08:49:43.313","AcceptedAnswerId":57887334.0,"OwnerUserId":11635654.0,"Title":"Simple Conv Net","Body":"<p>I'm new on pytorch and I would like your light on a small net I try to setup.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nclass PzConv2d(nn.Module):\n    \"\"\" Convolution 2D Layer followed by PReLU activation\n    \"\"\"\n    def __init__(self, n_in_channels, n_out_channels, **kwargs):\n        super(PzConv2d, self).__init__()\n        self.conv = nn.Conv2d(n_in_channels, n_out_channels, bias=True,\n                            **kwargs)\n        self.activ = nn.ReLU()\n\n    def forward(self, x):\n        x = self.conv(x)\n        return self.activ(x)\n\n\nclass PzPool2d(nn.Module):\n    \"\"\" Average Pooling Layer\n    \"\"\"\n    def __init__(self, kernel_size, stride, padding=0):\n        super(PzPool2d, self).__init__()\n        self.pool = nn.AvgPool2d(kernel_size=kernel_size,\n                                 stride=stride,\n                                 padding=padding,\n                                 ceil_mode=True,\n                                 count_include_pad=False)\n\n    def forward(self, x):\n        return self.pool(x)\n\nclass PzFullyConnected(nn.Module):\n    \"\"\" Dense or Fully Connected Layer followed by ReLU\n    \"\"\"\n    def __init__(self, n_inputs, n_outputs, withrelu=True, **kwargs):\n        super(PzFullyConnected, self).__init__()\n        self.withrelu = withrelu\n        self.linear = nn.Linear(n_inputs, n_outputs, bias=True)\n        self.activ = nn.ReLU()\n\n    def forward(self, x):\n        x = self.linear(x)\n        if self.withrelu:\n            x = self.activ(x)\n        return x\n\n\nclass NetCNN(nn.Module):\n    def __init__(self,n_input_channels,debug=False):\n        super(NetCNN, self).__init__()\n\n        self.n_bins = 180\n        self.debug = debug\n        self.conv0 = PzConv2d(n_in_channels=n_input_channels,\n                              n_out_channels=64,\n                              kernel_size=5,padding=2)\n        self.pool0 = PzPool2d(kernel_size=2,stride=2,padding=0)\n\n        self.conv1 = PzConv2d(n_in_channels=64,\n                              n_out_channels=92,\n                              kernel_size=3,padding=2)\n        self.pool1 = PzPool2d(kernel_size=2,stride=2,padding=0)\n\n        self.conv2 = PzConv2d(n_in_channels=92,\n                              n_out_channels=128,\n                              kernel_size=3,padding=2)\n        self.pool2 = PzPool2d(kernel_size=2,stride=2,padding=0)\n\n\n        self.fc0 = PzFullyConnected(n_inputs=12800,n_outputs=1024)\n        self.fc1 = PzFullyConnected(n_inputs=1024,n_outputs=self.n_bins)\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\n    def forward(self, x, dummy):\n        # x:image tensor N_batch, Channels, Height, Width\n        #    size N, Channels:=5 filtres, H,W = 64 pixels\n        # dummy: is not used\n\n        # stage 0 conv 64 x 5x5\n        x = self.conv0(x)\n        x = self.pool0(x)\n\n        # stage 1 conv 92 x 3x3\n        x = self.conv1(x)\n        x = self.pool1(x)\n\n        # stage 2 conv 128 x 3x3\n        x = self.conv2(x)\n        x = self.pool2(x)\n\n        x = self.fc0(x.view(-1,self.num_flat_features(x)))\n        x = self.fc1(x)\n\n        output = x\n\n\n        return output\n<\/code><\/pre>\n\n<p>I have checked that the dimensions of the intermediate \"x\" tensor in the forward process is well shaped (at least when I send a random input image tensor). But if you see something strange let me know, please.<\/p>\n\n<p>Now, I have seen code with F.\"function\" sequence in the forward method, instead of declaring the different layers as I have done. Does it make a difference?<\/p>\n\n<p>(Notice that I'm using  F.cross_entropy as loss function, so I do not end may network by a SoftMax.)<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"57887334","CreationDate":"2019-09-11T10:40:22.237","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>As you can see from pytorch doc there are <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/nn.html#convolution-layers\" rel=\"nofollow noreferrer\">\"layers\"<\/a> and there are <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/nn.html#torch-nn-functional\" rel=\"nofollow noreferrer\">\"functionals\"<\/a>. You already noticed that they are quite similar, <strong>but<\/strong> there is a difference:<br>\nA layer is often more than just a \"functional\" it also wraps around <em>trainable parameters<\/em>.<br>\nTherefore, you can have a <code>F.conv2d(...)<\/code> in your <code>forward()<\/code> function, but you will have to manually provide (store\/update) the weights\/kernel for this convolution. On the other hand, if you are using <code>nn.Conv2d<\/code> pytorch takes care of managing\/storing\/updating the weights\/kernels for you.<br>\nSome layers have no internal parameters\/buffers (e.g., <code>nn.ReLU<\/code>, <code>nn.Softmax<\/code> etc.) thus you can choose if you want to have a \"layer\" for this operation or only calling the appropriate functional in your <code>forward()<\/code> function. It is a matter of convenience and habits, it's up to you.<\/p>\n"}]}
+{"QuestionId":57885634,"AnswerCount":1,"Tags":"<tensorflow><model><dataset>","CreationDate":"2019-09-11T09:01:55.747","AcceptedAnswerId":null,"OwnerUserId":12051957.0,"Title":"Tensorflow model fit method throws exception for dimensions","Body":"<p>I want to exercise this tensorflow tutorial about image classification <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>and have loaded my data into a dataset with this tutorial\n<a href=\"https:\/\/www.tensorflow.org\/tutorials\/load_data\/images\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/load_data\/images<\/a><\/p>\n\n<p>If i am calling the model fit method i get following exception:\n<b>Error when checking input: expected flatten_6_input to have 4 dimensions, but got array with shape (28, 28, 3)<\/b><\/p>\n\n<p>The images are in the shape of (28, 28, 3)<\/p>\n\n<p>I have already tried to pass several dimension as input parameter and also read out the data from the dataset to verify the data.<\/p>\n\n<pre><code>image_ds = path_ds.map(load_and_preprocess_image, num_parallel_calls=AUTOTUNE)\nlabel_ds = tf.data.Dataset.from_tensor_slices(tf.cast(all_image_labels, tf.int64))\nimage_label_ds = tf.data.Dataset.zip((image_ds, label_ds))\n\nmodel = keras.Sequential([\n    keras.layers.Flatten(input_shape=(28, 28, 3)),\n    keras.layers.Dense(256, activation=tf.nn.relu),\n    keras.layers.Dense(2, activation=tf.nn.softmax)\n])\n\nmodel.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\n\nmodel.fit(image_label_ds.shuffle(5), epochs=5)\n<\/code><\/pre>\n\n<pre><code>Epoch 1\/5\n---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-114-a638eaff329f&gt; in &lt;module&gt;()\n----&gt; 1 model.fit(train_dataset.shuffle(5), epochs=5)\n\n5 frames\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/training_utils.py in standardize_input_data(data, names, shapes, check_batch_axis, exception_prefix)\n    374                            ': expected ' + names[i] + ' to have ' +\n    375                            str(len(shape)) + ' dimensions, but got array '\n--&gt; 376                            'with shape ' + str(data_shape))\n    377         if not check_batch_axis:\n    378           data_shape = data_shape[1:]\n\nValueError: Error when checking input: expected flatten_6_input to have 4 dimensions, but got array with shape (28, 28, 3)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57886015","CreationDate":"2019-09-11T09:25:03.387","ParentId":null,"OwnerUserId":"12051957","Title":null,"Body":"<p>I just included the command<\/p>\n\n<p>'''\nimage_label_ds= image_label_ds.batch(2)\n'''<\/p>\n\n<p>and it worked<\/p>\n"}]}
+{"QuestionId":57885649,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-09-11T09:03:03.407","AcceptedAnswerId":null,"OwnerUserId":9181725.0,"Title":"How to quantize weight into 8 bits fixed point in Keras","Body":"<p>I need to implement MNIST into FPGA, so I should express input, weight, bias, output in 8 bits fixed point for saving memory.<\/p>\n\n<p>My first idea is just convert trained weight(floating point) into 8 bits fixed point. But that will decrease accuracy a lot.<\/p>\n\n<p>So, I think I should add 8 bits fixed point converter between each two layers. But I have no idea how to add this kind of function between two layers. Can give me some hint? Thanks! Following is my keras code.<\/p>\n\n<pre><code>from __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\n\nbatch_size = 128\nnum_classes = 10\nepochs = 12\n\n# input image dimensions\nimg_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\nif 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)\nelse:\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\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\n# convert 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(2, kernel_size=(3, 3),\n                activation='relu',\n                input_shape=input_shape))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\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<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57885776,"AnswerCount":0,"Tags":"<tensorflow><keras><lstm><tensorflow-datasets>","CreationDate":"2019-09-11T09:11:01.973","AcceptedAnswerId":null,"OwnerUserId":11545782.0,"Title":"How to train TensorFlow LSTM using TFRecordDataset from sequences of different length","Body":"<p>The TensorFlow LSTM implementation offers a stateful and stateless variant. By default, the stateless variant is used which automatically resets the LSTM state after one batch of training. Given the idea of LSTM networks, the input data is likely some kind of sequence or series that should be learned. From my understanding, the state should typically be reset after exactly one (of many) complete sequence. The input size of a LSTM is given as<\/p>\n\n<pre><code>batch_input_shape=(batch_size, num_timesteps, num_features)\n<\/code><\/pre>\n\n<p>and hence (from my understanding), the length of one series should be <code>batch_size * num_timesteps<\/code>. Alternatively, when the whole series is too big to fit into one batch, one can use the stateful variant to manually reset after all batches of one series have been trained.<\/p>\n\n<p>Now, I have a lot of series, each in their individual <code>.tfrecord<\/code> file and I am currently training the stateless variant. However, I am not sure how to handle the series inside TensorFlow. With <\/p>\n\n<pre><code>ds = tf.data.TFRecordDataset(files)\n<\/code><\/pre>\n\n<p>and some pre-processing I can create a dataset from these files and then train them using<\/p>\n\n<pre><code>model.fit(ds, epochs=num_epochs, shuffle=False)\n<\/code><\/pre>\n\n<p>However, from my understanding the dataset does not distinguish between single files any more and simply concatenates all of them together. This means that even for sequences of different lengths, the training always uses the same batch size (i.e. without padding) and potentially mixes up different sequences.<\/p>\n\n<p>Consequently, my idea would be to manually iterate the files while using a stateful LSTM, training with smaller batches and manually resetting the state after each sequence. How can I iterate the individual files of a <code>TFRecordDataset<\/code>. Is this possible? Is there any better way?<\/p>\n","answers":[]}
+{"QuestionId":57886093,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><optimization><neural-network>","CreationDate":"2019-09-11T09:30:12.603","AcceptedAnswerId":null,"OwnerUserId":10611614.0,"Title":"Tensorflow exponential decay learning rate seems not to work","Body":"<p>I have built a neural network with tensorflow and it looks as follows:<\/p>\n\n<pre><code>n_hidden = 32\nsteps = 10**5*4\ndecay_rate = 1e-4\ninitial_lr = 1e-3\n\ntf.reset_default_graph()\ng = tf.Graph()\n\ndropout_rate = tf.placeholder_with_default(0.2, (), name='dropout')\ncurr_step = tf.placeholder_with_default(1, (), name='current_step')\nlearning_rate = tf.train.exponential_decay(initial_lr, global_step=curr_step, decay_steps=steps,\n                                           decay_rate=decay_rate, name='learning_rate')\n\nX_tensor = tf.placeholder(tf.float32, shape=[None, X.shape[1]], name='X_input')\ny_tensor = tf.placeholder(tf.int64, shape=[None], name='y_input')\n\nw = tf.Variable(tf.random_normal([X.shape[1], n_hidden]), name='w_0')\nb = tf.Variable(tf.random.normal([n_hidden]), name='b_0')\nproduct = tf.nn.leaky_relu(tf.matmul(X_tensor, tf.nn.dropout(w, rate=dropout_rate, name='w_0_dropout'),\n                                     name='matmul_0') + tf.nn.dropout(b, rate=dropout_rate, name='b_0_dropout'),\n                           name='activation_0')\n\nw_1 = tf.Variable(tf.random_normal([n_hidden, n_hidden]), name='w_1')\nb_1 = tf.Variable(tf.random_normal([n_hidden]), name='b_1')\nproduct_1 = tf.nn.leaky_relu(tf.matmul(product, tf.nn.dropout(w_1, rate=dropout_rate, name='w_1_dropout'),\n                                       name='matmul_1') + tf.nn.dropout(b_1, rate=dropout_rate, name='b_1_dropout'),\n                             name='activation_1')\n\nw_2 = tf.Variable(tf.random_normal([n_hidden, 1]), name='w_2')\nb_2 = tf.Variable(tf.random_normal([1]), name='b_2')\nproduct_2 = tf.reshape(tf.nn.leaky_relu(tf.matmul(product_1, tf.nn.dropout(w_2, rate=dropout_rate,\n                                                                           name='w_2_dropout'),\n                                                  name='matmul_2') + b_2, name='activation_2'), [-1],\n                       name='reshape')\ncost = tf.losses.mean_squared_error(labels=y_tensor, predictions=product_2)\n\n#correct_predictions = tf.equal(tf.argmax(product, axis=1), y_tensor)\n#accuracy = tf.reduce_mean(tf.cast(correct_predictions, tf.float64))\nmae = tf.losses.absolute_difference(y_tensor, product_2)\ncorrect_predictions = tf.equal(tf.cast(tf.round(product_2), tf.int64), y_tensor, name='correct')\naccuracy = tf.reduce_mean(tf.cast(correct_predictions, tf.float64), name='accuracy')\noptimizer = tf.train.GradientDescentOptimizer(learning_rate, name='optimizer').minimize(cost)\n<\/code><\/pre>\n\n<p>Even when I decrease learning rate to worthless (1e-100), the loss anyway fluctuates:<\/p>\n\n<pre><code>Step 2500, Minibatch Loss= 2.8308, Training Accuracy= 0.2525, Training MAE= 1.3107, lr= 0.00000000000000\nStep 5000, Minibatch Loss= 2.7827, Training Accuracy= 0.2664, Training MAE= 1.2948, lr= 0.00000000000000\nStep 7500, Minibatch Loss= 2.6718, Training Accuracy= 0.2481, Training MAE= 1.2784, lr= 0.00000000000000\nStep 10000, Minibatch Loss= 2.6464, Training Accuracy= 0.2603, Training MAE= 1.2718, lr= 0.00000000000000\nStep 12500, Minibatch Loss= 2.8204, Training Accuracy= 0.2614, Training MAE= 1.3014, lr= 0.00000000000000\n<\/code><\/pre>\n\n<p>Maybe I have mixed up something? All the data is scaled, so lr=1e-100 cannot influence, although it does.<\/p>\n\n<p>Will be grateful for any help!<\/p>\n","answers":[{"AnswerId":"57893019","CreationDate":"2019-09-11T16:06:28.430","ParentId":null,"OwnerUserId":"3399825","Title":null,"Body":"<p>Are you sure that the parameters fluctuate though? You don't display your execution code, however, it is quite possible, that all the metrics that are displayed are only <strong>an average over all the batches seen in the current epoch<\/strong>. That would mean that the first row is average over 2500 batches, the second row would be average over 5000 batches, etc.<\/p>\n\n<p>This would explain the fluctuation. So try printing out your parameters after epochs, if they are indeed changing too, you can eliminate this hypothesis.<\/p>\n"}]}
+{"QuestionId":57886176,"AnswerCount":1,"Tags":"<python><machine-learning><keras><scikit-learn><one-hot-encoding>","CreationDate":"2019-09-11T09:34:54.157","AcceptedAnswerId":null,"OwnerUserId":12026486.0,"Title":"How to validly encode a numerical target variable in scikit-learn?","Body":"<p>I am trying to get Scikit-learn working with a classification problem for a data set with 500 observations, 20 features, and 5 categorical target labels (1, 2, 3, 4, 5). The features are all of type float64, and they have been normalized to z-scores. The fitting and validation works fine, as long as the target variable (y) is fed as a one-dimensional NumPy array.<\/p>\n\n<p>However, I recently realized the importance of one-hot encoding the target variable, since the outcome of the classificiation seems to slightly vary depending on the integer a class is assigned. <\/p>\n\n<p>For example, here's a quick overview of the mean results for a k-NN classifier by using different orders for the encoded variables:\n(1, 2, 3, 4, 5) = 52,7%,\n(2, 4, 6, 8, 10) = 52,5% and\n(4, 3, 5, 1, 2) = 52,1%. This is problematic, since the labels for y are originally strings, and the integers themselves don't tell anything about the data.<\/p>\n\n<p>One way to solve this problem would of course be to feed y as string values (e.g. \"1\", \"2\", \"3\", \"4\", \"5\"). However, I found out that scikit-learn automatically converts strings to integers, as long as the values are convertable to integers. This does not eliminate the problem itself, since (\"1\", \"2\", \"3\", \"4\", \"5\") would still result in a mean accuracy of exactly 52,7%. Using values like (\"a\", \"b\", \"c\", \"d\", \"e\") could be a solution, but it results in the error message:<\/p>\n\n<p>ValueError: invalid literal for int() with base 10: 'a'<\/p>\n\n<p>Unlike in Keras, it appears that the target variable has to be fed as a one-dimensional NumPy array, and thus, can't be given as a one-hot encoded matrix with the shape (500, 5). This results in the following error message:<\/p>\n\n<p>ValueError: bad input shape (500, 5)<\/p>\n\n<p>Here's an example of the code for a simple Gaussian Naive Bayes model:<\/p>\n\n<pre><code>random_state = 123\nn_splits = 10\ncv = StratifiedKFold(n_splits=n_splits, random_state=random_state,\nshuffle=False)\n\nmodel = GaussianNB()\n\nfor train, test in cv.split(X, y): \n    y_dummies = np_utils.to_categorical(y)\n    y_dummies = y_dummies[:, 1:]\n\n    X_train, X_test = X[train], X[test]\n    y_train, y_test = y_dummies[train], y_dummies[test]\n\n    model.fit(X_train, y_train) \n    y_pred = model.predict(X_test)\n\n    conf = confusion_matrix(y_test.argmax(axis=1), y_pred)\n    print(conf)\n<\/code><\/pre>\n\n<p>In the end, I would like to be able to output the confusion matrix (conf), that has eliminated the effect of the abstract numeric label that were assigned for each of the five classes.<\/p>\n","answers":[{"AnswerId":"57886581","CreationDate":"2019-09-11T09:54:37.847","ParentId":null,"OwnerUserId":"3653343","Title":null,"Body":"<p>When using K-NN the way in which you represent the classes\/target labels would not influence the accuracy\/precision\/recall all other metrics that you are using to evaluate the model<\/p>\n\n<p>Since you classify an instance as correctly classified if the majority of the k nearest neighboors are of the same class of the instance<\/p>\n\n<p>I think that there is some other problem in your implemantation. Maybe there is some shuffle or other random sampling on your dataset before the evaluation is done (which of course would impact, slightly or a lot, depending on the size of your dataset, your performance). <\/p>\n\n<p>Or maybe you are not consistently modifying your labels when doing the changes<\/p>\n"}]}
+{"QuestionId":57886686,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-11T10:01:08.580","AcceptedAnswerId":null,"OwnerUserId":11520685.0,"Title":"How do I make my model return probability map instead of arg_max of that?","Body":"<p>I have created a U-Net class, with which one can define a U-Net with variable depth. I am training it to segment only one class out of an image, but the problem is, that when I use a trained model to infer an image an array with only ones and zeroes is returned, I would like to see the actual probabilities of the class. How can I achieve this?<\/p>\n\n<p>Class with which I create the model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class UNet():\n    def __init__(self, depth, shape):\n          encoder_layers = [[0, 0] for x in range(depth)]\n          decoder_layers = [0 for x in range(depth)]\n          inputs = layers.Input(shape=shape, name = 'Input')                                                     # Input layer\n          encoder_layers[0] = self.encoder_block(inputs, 32)                                     # First encoder block\n          for i in range(1, depth):                                                              # Assemble the rest encoder blocks\n              encoder_layers[i] = self.encoder_block(encoder_layers[i-1][0], 32*2**i)\n          center = self.conv_block(encoder_layers[-1][0], 32*2**depth)                           # Center\n          decoder_layers[0] = self.decoder_block(center, encoder_layers[-1][1], 32*2**(depth-1)) # First decoder block\n          for i in range(1, depth):                                                              # Assemble the rest decoder blocks\n              decoder_layers[i] = self.decoder_block(decoder_layers[i -1], encoder_layers[- i -1][1], 32*2**(i-1))\n          outputs = layers.Conv2D(1, (1, 1), activation='sigmoid', name = 'Output')(decoder_layers[-1]) # Output layer\n          self.model = models.Model(inputs=[inputs], outputs=[outputs])\n\n    def encoder_block(self, input_tensor, num_filters):\n          encoder = self.conv_block(input_tensor, num_filters)\n          encoder_pool = layers.MaxPooling2D((2, 2), strides=(2, 2))(encoder)\n\n          return encoder_pool, encoder      \n\n    @staticmethod    \n    def conv_block(input_tensor, num_filters):\n          encoder = layers.Conv2D(num_filters, (3, 3), padding='same')(input_tensor)\n          encoder = layers.BatchNormalization()(encoder)\n          encoder = layers.Activation('relu')(encoder)\n          encoder = layers.Conv2D(num_filters, (3, 3), padding='same')(encoder)\n          encoder = layers.BatchNormalization()(encoder)\n          encoder = layers.Activation('relu')(encoder)\n          return encoder\n\n    @staticmethod\n    def decoder_block(input_tensor, concat_tensor, num_filters):\n          decoder = layers.Conv2DTranspose(num_filters, (2, 2), strides=(2, 2), padding='same')(input_tensor)\n          decoder = layers.concatenate([concat_tensor, decoder], axis=-1)\n          decoder = layers.BatchNormalization()(decoder)\n          decoder = layers.Activation('relu')(decoder)\n          decoder = layers.Conv2D(num_filters, (3, 3), padding='same')(decoder)\n          decoder = layers.BatchNormalization()(decoder)\n          decoder = layers.Activation('relu')(decoder)\n          decoder = layers.Conv2D(num_filters, (3, 3), padding='same')(decoder)\n          decoder = layers.BatchNormalization()(decoder)\n          decoder = layers.Activation('relu')(decoder)\n          return decoder\n<\/code><\/pre>\n\n<p>Class which I use for predictions:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class predictor():\n    def __init__(self, model_path):\n        self.model = models.load_model(model_path, custom_objects={'wce_dice_loss': wce_dice_loss,\n                                                                   'dice_loss': dice_loss,\n                                                                   'IoU': IoU})\n    def predict(self, image_path, batch_size):\n        self.img_arr = plt.imread(image_path)[:,:,0]\n        img_slicer = lip(self.img_arr)\n        img_slices =img_slicer.streched_img_map\n        msk_slices = np.zeros_like(img_slices)\n        index = 0\n        while True:\n            msk_slices[index : index + batch_size] = self.model.predict(img_slices[index : index + batch_size])\n            index += batch_size\n            if index &gt;= len(img_slices): break\n        self.mask_arr = img_slicer.mask_gluer(msk_slices)\n\n    def save(self, save_path):\n        self.mask_arr.imsave(save_path, cmap = 'gray')\n\n    def show(self):\n        plt.imshow(self.mask_arr, cmap = 'gray')\n<\/code><\/pre>\n\n<p>model.summarry(), for a model with a depth of 2:<\/p>\n\n<pre><code>Model: \"model\"\n__________________________________________________________________________________________________\nLayer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\nInput (InputLayer)              [(None, 512, 512, 1) 0                                            \n__________________________________________________________________________________________________\nconv2d (Conv2D)                 (None, 512, 512, 32) 320         Input[0][0]                      \n__________________________________________________________________________________________________\nbatch_normalization (BatchNorma (None, 512, 512, 32) 128         conv2d[0][0]                     \n__________________________________________________________________________________________________\nactivation (Activation)         (None, 512, 512, 32) 0           batch_normalization[0][0]        \n__________________________________________________________________________________________________\nconv2d_1 (Conv2D)               (None, 512, 512, 32) 9248        activation[0][0]                 \n__________________________________________________________________________________________________\nbatch_normalization_1 (BatchNor (None, 512, 512, 32) 128         conv2d_1[0][0]                   \n__________________________________________________________________________________________________\nactivation_1 (Activation)       (None, 512, 512, 32) 0           batch_normalization_1[0][0]      \n__________________________________________________________________________________________________\nmax_pooling2d (MaxPooling2D)    (None, 256, 256, 32) 0           activation_1[0][0]               \n__________________________________________________________________________________________________\nconv2d_2 (Conv2D)               (None, 256, 256, 64) 18496       max_pooling2d[0][0]              \n__________________________________________________________________________________________________\nbatch_normalization_2 (BatchNor (None, 256, 256, 64) 256         conv2d_2[0][0]                   \n__________________________________________________________________________________________________\nactivation_2 (Activation)       (None, 256, 256, 64) 0           batch_normalization_2[0][0]      \n__________________________________________________________________________________________________\nconv2d_3 (Conv2D)               (None, 256, 256, 64) 36928       activation_2[0][0]               \n__________________________________________________________________________________________________\nbatch_normalization_3 (BatchNor (None, 256, 256, 64) 256         conv2d_3[0][0]                   \n__________________________________________________________________________________________________\nactivation_3 (Activation)       (None, 256, 256, 64) 0           batch_normalization_3[0][0]      \n__________________________________________________________________________________________________\nmax_pooling2d_1 (MaxPooling2D)  (None, 128, 128, 64) 0           activation_3[0][0]               \n__________________________________________________________________________________________________\nconv2d_4 (Conv2D)               (None, 128, 128, 128 73856       max_pooling2d_1[0][0]            \n__________________________________________________________________________________________________\nbatch_normalization_4 (BatchNor (None, 128, 128, 128 512         conv2d_4[0][0]                   \n__________________________________________________________________________________________________\nactivation_4 (Activation)       (None, 128, 128, 128 0           batch_normalization_4[0][0]      \n__________________________________________________________________________________________________\nconv2d_5 (Conv2D)               (None, 128, 128, 128 147584      activation_4[0][0]               \n__________________________________________________________________________________________________\nbatch_normalization_5 (BatchNor (None, 128, 128, 128 512         conv2d_5[0][0]                   \n__________________________________________________________________________________________________\nactivation_5 (Activation)       (None, 128, 128, 128 0           batch_normalization_5[0][0]      \n__________________________________________________________________________________________________\nconv2d_transpose (Conv2DTranspo (None, 256, 256, 64) 32832       activation_5[0][0]               \n__________________________________________________________________________________________________\nconcatenate (Concatenate)       (None, 256, 256, 128 0           activation_3[0][0]               \n                                                                 conv2d_transpose[0][0]           \n__________________________________________________________________________________________________\nbatch_normalization_6 (BatchNor (None, 256, 256, 128 512         concatenate[0][0]                \n__________________________________________________________________________________________________\nactivation_6 (Activation)       (None, 256, 256, 128 0           batch_normalization_6[0][0]      \n__________________________________________________________________________________________________\nconv2d_6 (Conv2D)               (None, 256, 256, 64) 73792       activation_6[0][0]               \n__________________________________________________________________________________________________\nbatch_normalization_7 (BatchNor (None, 256, 256, 64) 256         conv2d_6[0][0]                   \n__________________________________________________________________________________________________\nactivation_7 (Activation)       (None, 256, 256, 64) 0           batch_normalization_7[0][0]      \n__________________________________________________________________________________________________\nconv2d_7 (Conv2D)               (None, 256, 256, 64) 36928       activation_7[0][0]               \n__________________________________________________________________________________________________\nbatch_normalization_8 (BatchNor (None, 256, 256, 64) 256         conv2d_7[0][0]                   \n__________________________________________________________________________________________________\nactivation_8 (Activation)       (None, 256, 256, 64) 0           batch_normalization_8[0][0]      \n__________________________________________________________________________________________________\nconv2d_transpose_1 (Conv2DTrans (None, 512, 512, 32) 8224        activation_8[0][0]               \n__________________________________________________________________________________________________\nconcatenate_1 (Concatenate)     (None, 512, 512, 64) 0           activation_1[0][0]               \n                                                                 conv2d_transpose_1[0][0]         \n__________________________________________________________________________________________________\nbatch_normalization_9 (BatchNor (None, 512, 512, 64) 256         concatenate_1[0][0]              \n__________________________________________________________________________________________________\nactivation_9 (Activation)       (None, 512, 512, 64) 0           batch_normalization_9[0][0]      \n__________________________________________________________________________________________________\nconv2d_8 (Conv2D)               (None, 512, 512, 32) 18464       activation_9[0][0]               \n__________________________________________________________________________________________________\nbatch_normalization_10 (BatchNo (None, 512, 512, 32) 128         conv2d_8[0][0]                   \n__________________________________________________________________________________________________\nactivation_10 (Activation)      (None, 512, 512, 32) 0           batch_normalization_10[0][0]     \n__________________________________________________________________________________________________\nconv2d_9 (Conv2D)               (None, 512, 512, 32) 9248        activation_10[0][0]              \n__________________________________________________________________________________________________\nbatch_normalization_11 (BatchNo (None, 512, 512, 32) 128         conv2d_9[0][0]                   \n__________________________________________________________________________________________________\nactivation_11 (Activation)      (None, 512, 512, 32) 0           batch_normalization_11[0][0]     \n__________________________________________________________________________________________________\nOutput (Conv2D)                 (None, 512, 512, 1)  33          activation_11[0][0]              \n==================================================================================================\nTotal params: 469,281\nTrainable params: 467,617\nNon-trainable params: 1,664\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57886871,"AnswerCount":1,"Tags":"<python><tensorflow><keras><neural-network><conv-neural-network>","CreationDate":"2019-09-11T10:10:45.050","AcceptedAnswerId":null,"OwnerUserId":12052274.0,"Title":"Variables are same for all Epochs","Body":"<p>I am experimenting the image classifier using CNN with keras<\/p>\n\n<p>My code -<\/p>\n\n<pre><code>model = Sequential()\n\nmodel.add(Conv2D(32, (3, 3), padding='same', input_shape=(224, 224, 3), activation=\"relu\"))\nmodel.add(Conv2D(32, (3, 3), padding='same', activation=\"relu\"))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Flatten())\nmodel.add(Dense(5, activation=\"softmax\"))\n\nmodel.compile(\n    loss='categorical_crossentropy',\n    optimizer='rmsprop',\n    metrics=['accuracy']\n)\nmodel.summary()\n\ncheck=ModelCheckpoint(filepath=r'C:\\Users\\softloft\\AppData\\Local\\Programs\\Python\\Python37\\Scripts\\Untitled Folder\\i_models.hdf5', verbose=1, save_best_only = True)\n\nhistory=model.fit(\n    x_train,\n    y_train,\n    batch_size=100,\n    epochs=30,\n    validation_data=(x_test, y_test),\n    shuffle=True,\n    callbacks=[check],\n)\n<\/code><\/pre>\n\n<p>I run this code with one CNN layer and it gives val_acc of about 71.<\/p>\n\n<p>Then, I add another CNN layer(which is the above code) and it gives the result -<\/p>\n\n<pre><code>Train on 3242 samples, validate on 1081 samples\nEpoch 1\/30\n3242\/3242 [==============================] - 235s 73ms\/step - loss: 13.0773 - acc: 0.1771 - val_loss: 13.8219 - val_acc: 0.1425\nEpoch 00001: val_loss improved from inf to 13.82190, saving model to i_models.hdf5\nEpoch 2\/30\n3242\/3242 [==============================] - 235s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00002: val_loss did not improve from 13.82190\nEpoch 3\/30\n3242\/3242 [==============================] - 235s 72ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00003: val_loss did not improve from 13.82190\nEpoch 4\/30\n3242\/3242 [==============================] - 235s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00004: val_loss did not improve from 13.82190\nEpoch 5\/30\n3242\/3242 [==============================] - 235s 72ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00005: val_loss did not improve from 13.82190\nEpoch 6\/30\n3242\/3242 [==============================] - 234s 72ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00006: val_loss did not improve from 13.82190\nEpoch 7\/30\n3242\/3242 [==============================] - 235s 72ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00007: val_loss did not improve from 13.82190\nEpoch 8\/30\n3242\/3242 [==============================] - 235s 72ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00008: val_loss did not improve from 13.82190\nEpoch 9\/30\n3242\/3242 [==============================] - 235s 72ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00009: val_loss did not improve from 13.82190\nEpoch 10\/30\n3242\/3242 [==============================] - 236s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00010: val_loss did not improve from 13.82190\nEpoch 11\/30\n3242\/3242 [==============================] - 236s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00011: val_loss did not improve from 13.82190\nEpoch 12\/30\n3242\/3242 [==============================] - 236s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00012: val_loss did not improve from 13.82190\nEpoch 13\/30\n3242\/3242 [==============================] - 235s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00013: val_loss did not improve from 13.82190\nEpoch 14\/30\n3242\/3242 [==============================] - 236s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00014: val_loss did not improve from 13.82190\nEpoch 15\/30\n3242\/3242 [==============================] - 236s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00015: val_loss did not improve from 13.82190\nEpoch 16\/30\n3242\/3242 [==============================] - 235s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00016: val_loss did not improve from 13.82190\nEpoch 17\/30\n3242\/3242 [==============================] - 236s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00017: val_loss did not improve from 13.82190\nEpoch 18\/30\n3242\/3242 [==============================] - 235s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00018: val_loss did not improve from 13.82190\nEpoch 19\/30\n3242\/3242 [==============================] - 235s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00019: val_loss did not improve from 13.82190\nEpoch 20\/30\n3242\/3242 [==============================] - 236s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00020: val_loss did not improve from 13.82190\nEpoch 21\/30\n3242\/3242 [==============================] - 236s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00021: val_loss did not improve from 13.82190\nEpoch 22\/30\n3242\/3242 [==============================] - 235s 72ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00022: val_loss did not improve from 13.82190\nEpoch 23\/30\n3242\/3242 [==============================] - 236s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00023: val_loss did not improve from 13.82190\nEpoch 24\/30\n3242\/3242 [==============================] - 236s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00024: val_loss did not improve from 13.82190\nEpoch 25\/30\n3242\/3242 [==============================] - 236s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00025: val_loss did not improve from 13.82190\nEpoch 26\/30\n3242\/3242 [==============================] - 235s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00026: val_loss did not improve from 13.82190\nEpoch 27\/30\n3242\/3242 [==============================] - 235s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00027: val_loss did not improve from 13.82190\nEpoch 28\/30\n3242\/3242 [==============================] - 236s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00028: val_loss did not improve from 13.82190\nEpoch 29\/30\n3242\/3242 [==============================] - 236s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00029: val_loss did not improve from 13.82190\nEpoch 30\/30\n3242\/3242 [==============================] - 236s 73ms\/step - loss: 13.2345 - acc: 0.1789 - val_loss: 13.8219 - val_acc: 0.1425\n\nEpoch 00030: val_loss did not improve from 13.82190\n<\/code><\/pre>\n\n<p>What happens and what to do..<\/p>\n\n<p>How all the variables are same for all epochs<\/p>\n\n<p>thank you<\/p>\n","answers":[{"AnswerId":"57888015","CreationDate":"2019-09-11T11:21:13.827","ParentId":null,"OwnerUserId":"11368194","Title":null,"Body":"<p>Try changing this<\/p>\n\n<pre><code>model.compile(loss = \"categorical_crossentropy\", optimizer = \"rmsprop\")\n<\/code><\/pre>\n\n<p>to <\/p>\n\n<pre><code>model.compile(loss = \"categorical_crossentropy\", optimizer = 'adam')\n<\/code><\/pre>\n"}]}
+{"QuestionId":57886932,"AnswerCount":0,"Tags":"<keras><conv-neural-network><pre-trained-model><transfer-learning><vgg-net>","CreationDate":"2019-09-11T10:15:00.690","AcceptedAnswerId":null,"OwnerUserId":3802483.0,"Title":"keras model does not learn anything after setting pretrained weights by layer","Body":"<p>I am training a smaller VGG like model, and I set the pretrained weights of VGG16 to the first conv layers which are identical. My model acts pretty strange though, and does not learn anything at all - the loss stays the same, the accuracy stays the same. What is wrong and how can I fix it?<\/p>\n\n<pre><code>from keras import applications\nfrom keras.preprocessing.image import ImageDataGenerator\nfrom keras import optimizers\nfrom keras.models import Sequential, Model\nfrom keras.layers import Dropout, Flatten, Dense, GlobalAveragePooling2D\nfrom keras import backend as k\nfrom keras.callbacks import ModelCheckpoint, LearningRateScheduler, TensorBoard, EarlyStopping\nfrom keras import layers\nfrom keras import models\nfrom keras import optimizers\nfrom keras.layers import Dropout\nfrom keras.regularizers import l2\nfrom keras.callbacks import EarlyStopping, ModelCheckpoint, ReduceLROnPlateau\nimport matplotlib.pyplot as plt\nfrom keras.preprocessing.image import ImageDataGenerator\nimport os\n\n\n\nbase_model = models.Sequential()\nbase_model.add(layers.Conv2D(64, (3, 3), activation='relu', name='block1_conv1',  input_shape=(224, 224, 3)))\nbase_model.add(layers.Conv2D(64, (3, 3), activation='relu', name='block1_conv2'))\nbase_model.add(layers.MaxPooling2D((2, 2)))\n#model.add(Dropout(0.2))\nbase_model.add(layers.Conv2D(128, (3, 3), activation='relu', name='block2_conv1'))\nbase_model.add(layers.Conv2D(128, (3, 3), activation='relu', name='block2_conv2'))\nbase_model.add(layers.MaxPooling2D((2, 2), name='block2_pool'))\n#model.add(Dropout(0.2))\n\nbase_model.summary()\n\"\"\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================   \n_________________________________________________________________\nblock1_conv1 (Conv2D)        (None, 256, 256, 64)      1792      \n_________________________________________________________________\nblock1_conv2 (Conv2D)        (None, 256, 256, 64)      36928     \n_________________________________________________________________\nblock1_pool (MaxPooling2D)   (None, 128, 128, 64)      0         \n_________________________________________________________________\nblock2_conv1 (Conv2D)        (None, 128, 128, 128)     73856     \n_________________________________________________________________\nblock2_conv2 (Conv2D)        (None, 128, 128, 128)     147584    \n_________________________________________________________________\nblock2_pool (MaxPooling2D)   (None, 64, 64, 128)       0         \n=================================================================\nTotal params: 260,160.0\nTrainable params: 260,160.0\nNon-trainable params: 0.0\n\"\"\"\n\n\n\n\nbase_model.add(layers.Flatten())\n#base_model.add(layers.Dropout(0.5))  #Dropout for regularization\nbase_model.add(layers.Dense(256, activation='relu'))\nbase_model.add(layers.Dense(1, activation='sigmoid'))  #Sigmoid function at the end because we have just two classes\n\n\nepochs = 50\ncallbacks = []\n#schedule = None\ndecay = 0.0\n\n#earlyStopping = EarlyStopping(monitor='val_loss', patience=10, verbose=0, mode='min')\n#mcp_save = ModelCheckpoint('.mdl_wts.hdf5', save_best_only=True, monitor='val_loss', mode='min')\n#reduce_lr_loss = ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=3, verbose=1, epsilon=1e-5, mode='min')\n\n\n# compile the model with a SGD\/momentum optimizer\n# and a very slow learning rate.\nbase_model.compile(loss='binary_crossentropy',\n                   optimizer=optimizers.SGD(lr=1e-4,decay=1e-6, momentum=0.9, nesterov=True),\n                   metrics=['accuracy'])\n\n\nvgg = applications.VGG16(weights='imagenet', include_top=False, input_shape=(224, 224, 3))\nprint('Model loaded.')\n\n\"\"\"\n['block1_conv1',\n 'block1_conv2',\n 'block1_pool',\n 'block2_conv1',\n 'block2_conv2',\n 'block2_pool',\n 'block3_conv1',\n 'block3_conv2',\n 'block3_conv3',\n 'block3_conv4',\n 'block3_pool',\n 'block4_conv1',\n 'block4_conv2',\n 'block4_conv3',\n 'block4_conv4',\n 'block4_pool',\n 'block5_conv1',\n 'block5_conv2',\n 'block5_conv3',\n 'block5_conv4',\n 'block5_pool',\n 'dense_1',\n 'dense_2',\n 'dense_3',\n 'dropout_1',\n 'global_average_pooling2d_1',\n 'input_1']\n\"\"\"\n\n\nfor layer in vgg.layers:\n    if layer.name == 'block1_conv1':\n        base_model.layers[0].set_weights(layer.get_weights())\n    elif layer.name == 'block1_conv2':\n        base_model.layers[1].set_weights(layer.get_weights())\n    elif layer.name == 'block2_conv1':\n        base_model.layers[3].set_weights(layer.get_weights())\n    elif layer.name == 'block2_conv2':\n        base_model.layers[4].set_weights(layer.get_weights())\n\nos.environ[\"CUDA_VISIBLE_DEVICES\"]=\"0\"\n\ntrain_dir = '\/home\/d\/Desktop\/s\/data\/train'\neval_dir = '\/home\/d\/Desktop\/s\/data\/eval'\ntest_dir = '\/home\/d\/Desktop\/s\/data\/test'\n\n\n# create a data generator\ntrain_datagen = ImageDataGenerator(rescale=1.\/255,   #Scale the image between 0 and 1\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\nval_datagen = ImageDataGenerator(rescale=1.\/255)  #We do not augment validation data. we only perform rescale\n\ntest_datagen = ImageDataGenerator(rescale=1.\/255)  #We do not augment validation data. we only perform rescale\n\n# load and iterate training dataset\ntrain_generator = train_datagen.flow_from_directory(train_dir,  target_size=(224,224),class_mode='binary', batch_size=16, shuffle='True', seed=42)\n# load and iterate validation dataset\nval_generator = val_datagen.flow_from_directory(eval_dir,  target_size=(224,224),class_mode='binary', batch_size=16, shuffle='True', seed=42)\n# load and iterate test dataset\ntest_generator = test_datagen.flow_from_directory(test_dir,  target_size=(224,224), class_mode=None, batch_size=1, shuffle='False', seed=42)\n\n#The training part\n#We train for 64 epochs with about 100 steps per epoch\nhistory = base_model.fit_generator(train_generator,\n                              steps_per_epoch=train_generator.n \/\/ train_generator.batch_size,\n                              epochs=epochs,\n                              validation_data=val_generator,\n                              validation_steps=val_generator.n \/\/ val_generator.batch_size) #,\n                              #callbacks=[earlyStopping, mcp_save, reduce_lr_loss])\n\n#Save the model\n#base_model.save_weights('\/home\/d\/Desktop\/s\/base_model_weights.h5')\n#base_model.save('\/home\/d\/Desktop\/s\/base_model_keras.h5')\n\n#lets plot the train and val curve\n#get the details form the history object\nacc = history.history['acc']\nval_acc = history.history['val_acc']\nloss = history.history['loss']\nval_loss = history.history['val_loss']\n\nepochs = range(1, len(acc) + 1)\n\n#Train and validation accuracy\nplt.plot(epochs, acc, 'b', label='Training accuracy')\nplt.plot(epochs, val_acc, 'r', label='Validation accuracy')\nplt.title('Training and Validation accurarcy')\nplt.legend()\n\nplt.figure()\n#Train and validation loss\nplt.plot(epochs, loss, 'b', label='Training loss')\nplt.plot(epochs, val_loss, 'r', label='Validation loss')\nplt.title('Training and Validation loss')\nplt.legend()\n\nplt.show()\n<\/code><\/pre>\n\n<p>and the training just goes on forever like this (no changes from epoch to epoch either):<\/p>\n\n<pre><code>2625\/4002 [==================&gt;...........] - ETA: 3:49 - loss: 7.9723 - acc: 0.5053\n2626\/4002 [==================&gt;...........] - ETA: 3:49 - loss: 7.9720 - acc: 0.5053\n2627\/4002 [==================&gt;...........] - ETA: 3:49 - loss: 7.9735 - acc: 0.5052\n2628\/4002 [==================&gt;...........] - ETA: 3:48 - loss: 7.9732 - acc: 0.5052\n2629\/4002 [==================&gt;...........] - ETA: 3:48 - loss: 7.9732 - acc: 0.5052\n2630\/4002 [==================&gt;...........] - ETA: 3:48 - loss: 7.9729 - acc: 0.5052\n2631\/4002 [==================&gt;...........] - ETA: 3:48 - loss: 7.9725 - acc: 0.5052\n2632\/4002 [==================&gt;...........] - ETA: 3:48 - loss: 7.9729 - acc: 0.5052\n2633\/4002 [==================&gt;...........] - ETA: 3:48 - loss: 7.9733 - acc: 0.5052\n2634\/4002 [==================&gt;...........] - ETA: 3:47 - loss: 7.9722 - acc: 0.5053\n2635\/4002 [==================&gt;...........] - ETA: 3:47 - loss: 7.9730 - acc: 0.5052\n2636\/4002 [==================&gt;...........] - ETA: 3:47 - loss: 7.9719 - acc: 0.5053\n2637\/4002 [==================&gt;...........] - ETA: 3:47 - loss: 7.9727 - acc: 0.5052\n2638\/4002 [==================&gt;...........] - ETA: 3:47 - loss: 7.9731 - acc: 0.5052\n2639\/4002 [==================&gt;...........] - ETA: 3:47 - loss: 7.9732 - acc: 0.5052\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57887035,"AnswerCount":1,"Tags":"<python><tensorflow><gpu><handwriting-recognition>","CreationDate":"2019-09-11T10:21:57.793","AcceptedAnswerId":null,"OwnerUserId":11118507.0,"Title":"What statements and exactly where I have to add to train the given Handwriting recognition on GPU (Specifically on google colab's GPU)?","Body":"<p>This code is training perfectly on Google Colab's CPU . But when I am trying to use GPU , I get the same training time.<\/p>\n\n<p>Code Link : <a href=\"https:\/\/github.com\/githubharald\/SimpleHTR?source=post_page---------------------------\" rel=\"nofollow noreferrer\">https:\/\/github.com\/githubharald\/SimpleHTR?source=post_page---------------------------<\/a><\/p>\n\n<p>I have tried using with tf.device(\"\/device:GPU:0\") context manager above setupCNN() and setupRNN() and also above see.run() . But nothing worked . \nI am trying to run it on GPU from last 48hrs .<\/p>\n\n<p>Since Colab's GPU should be approx 20 times faster. I want to run it on GPU.\nAny help will be appreciated . \nThankYou.<\/p>\n","answers":[{"AnswerId":"58516013","CreationDate":"2019-10-23T05:27:37.720","ParentId":null,"OwnerUserId":"12198627","Title":null,"Body":"<p>Check for the GPU memory in colab,<\/p>\n\n<pre><code>!ln -sf \/opt\/bin\/nvidia-smi \/usr\/bin\/nvidia-smi\n!pip install gputil\n!pip install psutil\n!pip install humanize\nimport psutil\nimport humanize\nimport os\nimport GPUtil as GPU\nGPUs = GPU.getGPUs()\ngpu = GPUs[0]\ndef checkMemory():\n process = psutil.Process(os.getpid())\n print(\"Gen RAM Free: \" + humanize.naturalsize( psutil.virtual_memory().available ), \" | Proc size: \" + humanize.naturalsize( process.memory_info().rss))\n print(\"GPU RAM Free: {0:.0f}MB | Used: {1:.0f}MB | Util {2:3.0f}% | Total {3:.0f}MB\".format(gpu.memoryFree, gpu.memoryUsed, gpu.memoryUtil*100, gpu.memoryTotal))\ncheckMemory()\n<\/code><\/pre>\n"}]}
+{"QuestionId":57887040,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><glove>","CreationDate":"2019-09-11T10:22:15.323","AcceptedAnswerId":null,"OwnerUserId":11368194.0,"Title":"Do we need a GPU system to train an deep learning model?","Body":"<p>I have created an encoder-decoder model with pre-trained 100D glove embedding, to create an abstractive text summarizer. The data set has <code>4300<\/code> article, its summary data. Vocabulary size is <code>48549<\/code> for articles and <code>19130<\/code> for summary. <code>Total memory size of input, output variables = 7.5Gb<\/code><\/p>\n\n<p>Following is the basic encoder-decoder model:<\/p>\n\n<pre><code>latent_dim = 1024\nencoder_inputs = Input(shape=(max_x_len,))\nemb1 = Embedding(len(x_voc), 100, weights=[x_voc], trainable = False)(encoder_inputs)\n\nencoder = LSTM(latent_dim, return_state=True)\nencoder_outputs, state_h, state_c = encoder(emb1)\n\ndecoder_inputs = Input(shape=(None,))\nemb2 = Embedding(len(y_voc), 100, weights=[y_voc], trainable = False)(decoder_inputs)\n\ndecoder_lstm = LSTM(latent_dim, return_sequences=True, return_state=True)\ndecoder_outputs0, _, _ = decoder_lstm(emb2, initial_state=[state_h, state_c])\n\ndecoder_dense = Dense(len(y_voc), activation='softmax')\ndecoder_outputs = decoder_dense(decoder_outputs0)\n\nmodel = Model([encoder_inputs, decoder_inputs], decoder_outputs)\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>When I train on whole data the spyder consumes 99% of the memory and system stops.<\/p>\n\n<p>My System configuration are as follows:<\/p>\n\n<pre><code>OS - windows 10 (64-bit)\nRam - 8Gb\nProcessor - Intel(R) Core(TM) i5-3470\nROM - 300Gb\n<\/code><\/pre>\n\n<p>Further I want to -<\/p>\n\n<ul>\n<li>Add more data and layers to the model<\/li>\n<li>Add attention layer<\/li>\n<li>Implement Bert<\/li>\n<\/ul>\n\n<p>Kindly suggest a solution or an suitable system configuration.<\/p>\n","answers":[{"AnswerId":"57893975","CreationDate":"2019-09-11T17:17:45.777","ParentId":null,"OwnerUserId":"7801852","Title":null,"Body":"<p>There is difference between executing Deep learning program and simple ML program.\nIn deep lering we are actually working on tensors(means smallest vector) so for processing deep model we need some processing unit which are efficient to work on tensors. There may be some dedicated system which works on neural network program.\nSo for executing neural network model we need GPU or TPU for processing data which passed from one layer of neurons to others.\nCPU may work but CPU is not dedicated to work only for Neural network model. CPU was assigned to work on whole system basically computation program execute faster on CPU.\nHope This will help you.<\/p>\n"}]}
+{"QuestionId":57887217,"AnswerCount":0,"Tags":"<python><python-2.7><tensorflow><python-3.6><flags>","CreationDate":"2019-09-11T10:33:24.360","AcceptedAnswerId":null,"OwnerUserId":9652636.0,"Title":"absl.flags._exceptions.UnrecognizedFlagError: Unknown command line flag 'cfg'","Body":"<p>I want run this code in python 3.6 (on google colab). \nNote that It's correctly runs in python 2.7. What's the problem?<\/p>\n\n<pre><code>     tf.app.flags.DEFINE_string('cfg_path', cfg_path, 'config path.')\n     for (k, v) in cfg.items():\n         if set_flag:\n             setattr(flags, k.lower(), v)\n         else:\n             if hasattr(flags, k.lower()):\n                 setattr(flags, k.lower(), v)\n             else:\n                 def_func = type_to_define_fn[type(v)]\n                 .\n                 .\n     cfg['cfg_path'] = cfg_path\n     cfg['model_type']  = FLAGS.model\n     cfg['model_folder'] = FLAGS.model_folder\n     cfg['defense_type'] = FLAGS.defense_type\n     cfg['batch_size'] = FLAGS.batch_size\n     cfg['rec_rr'] = FLAGS.rec_rr\n     cfg['rec_iters'] = FLAGS.rec_iters\n     return cfg\n<\/code><\/pre>\n\n<p>complete code is <a href=\"https:\/\/github.com\/anvinhnguyendinh\/DefenseGANwithBPDA\/blob\/master\/utils\/config.py\" rel=\"nofollow noreferrer\">here<\/a> . \nHow can I run it in python3? Thank you so much.<\/p>\n","answers":[]}
+{"QuestionId":57887503,"AnswerCount":0,"Tags":"<numpy><tensorflow><image-processing><riff>","CreationDate":"2019-09-11T10:50:33.363","AcceptedAnswerId":null,"OwnerUserId":12052478.0,"Title":"how to handle an image encoding starting with RIFF","Body":"<p>please i need assistance with this encoding, it turns out the RIFF at the beginning doesn't let me decode the image. i'm using tensorflow , i'll place a copy of the code below. i need help in handling this numpy encoding some other way.<\/p>\n\n<pre><code>&gt;&gt;&gt;img_raw = tf.io.read_file(img_path)\nprint(repr(img_raw)[:100]+\"...\")\n\n&gt;&gt;&gt;img_tensor = tf.image.decode_jpeg(img_raw, channels=3)\nimg_tensor = tf.convert_to_tensor((img_tensor))\n# print(img_tensor)\nprint(img_tensor.shape)\nprint(img_tensor.dtype)\n\nthis is the beginning of the numpy encoding\n\n&gt;&gt;&gt;b'RIFF`\\xd4\\x00\\x00WEBPVP8 T\\xd4\\x00\\x00p~\\x02\\x9d\\x01*\\x89\\x02m\\x01&gt;A\\x16\\x88D\"\\xa1!\\x1e\\x7fn&lt;(\\x04\\x04\\xb2\\xb6\\x99\\xb1\\x8dS\\x1a\\xe1Cl\\x0\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57887524,"AnswerCount":1,"Tags":"<tensorflow><intel-mkl>","CreationDate":"2019-09-11T10:51:51.260","AcceptedAnswerId":58074452.0,"OwnerUserId":4921432.0,"Title":"Serving multiple Tensorflow model in same process with intel mkl","Body":"<p>I serve multiple models in one process, and each model create a Tensorflow session. Let say there are 8 model, so 8 tf.session been created.<\/p>\n\n<p>And I follow <a href=\"https:\/\/www.tensorflow.org\/guide\/performance\/overview#tuning_mkl_for_the_best_performance\" rel=\"nofollow noreferrer\">Optimizing for CPU, Tuning_mkl_for_the_best_performance<\/a> to open MKL.\nMy machine have 8 core and 2 threads. I set each tf.session as below.<\/p>\n\n<pre><code>config = tf.ConfigProto()\nconfig.intra_op_parallelism_threads = 8\nconfig.inter_op_parallelism_threads = 1\ntf.Session(config=config)\n<\/code><\/pre>\n\n<p>Also set <\/p>\n\n<pre><code>OMP_NUM_THREADS=8, \nKMP_BLOCKTIME=1; \nKMP_AFFINITY='granularity=fine,verbose,compact,1,0'; \nKMP_SETTINGS=1\n<\/code><\/pre>\n\n<p>However, it may cause Cpu over-subscript, the Golang process create 87 threads.\nDoes something wrong in my setting?<\/p>\n\n<p>Here is the log from OMP.<\/p>\n\n<pre><code>OMP: Info #212: KMP_AFFINITY: decoding x2APIC ids.\nOMP: Info #210: KMP_AFFINITY: Affinity capable, using global cpuid leaf 11 info\nOMP: Info #154: KMP_AFFINITY: Initial OS proc set respected: 2,4,5,12,17,27,32,47\nOMP: Info #156: KMP_AFFINITY: 8 available OS procs\nOMP: Info #157: KMP_AFFINITY: Uniform topology\nOMP: Info #179: KMP_AFFINITY: 2 packages x 4 cores\/pkg x 1 threads\/core (8 total cores)\nOMP: Info #214: KMP_AFFINITY: OS proc to physical thread map:\nOMP: Info #171: KMP_AFFINITY: OS proc 4 maps to package 0 core 1\nOMP: Info #171: KMP_AFFINITY: OS proc 32 maps to package 0 core 2\nOMP: Info #171: KMP_AFFINITY: OS proc 2 maps to package 0 core 5\nOMP: Info #171: KMP_AFFINITY: OS proc 12 maps to package 0 core 8\nOMP: Info #171: KMP_AFFINITY: OS proc 5 maps to package 1 core 1\nOMP: Info #171: KMP_AFFINITY: OS proc 27 maps to package 1 core 5\nOMP: Info #171: KMP_AFFINITY: OS proc 17 maps to package 1 core 9\nOMP: Info #171: KMP_AFFINITY: OS proc 47 maps to package 1 core 11\nOMP: Info #250: KMP_AFFINITY: pid 832 tid 946 thread 0 bound to OS proc set 4\nOMP: Info #250: KMP_AFFINITY: pid 832 tid 946 thread 1 bound to OS proc set 5\nOMP: Info #250: KMP_AFFINITY: pid 832 tid 945 thread 2 bound to OS proc set 32\nOMP: Info #250: KMP_AFFINITY: pid 832 tid 1583 thread 3 bound to OS proc set 27\nOMP: Info #250: KMP_AFFINITY: pid 832 tid 1584 thread 4 bound to OS proc set 2\nOMP: Info #250: KMP_AFFINITY: pid 832 tid 1585 thread 5 bound to OS proc set 17\nOMP: Info #250: KMP_AFFINITY: pid 832 tid 1586 thread 6 bound to OS proc set 12\nOMP: Info #250: KMP_AFFINITY: pid 832 tid 1587 thread 7 bound to OS proc set 47\nOMP: Info #250: KMP_AFFINITY: pid 832 tid 1590 thread 10 bound to OS proc set 32\nOMP: Info #250: KMP_AFFINITY: pid 832 tid 1589 thread 9 bound to OS proc set 5\nOMP: Info #250: KMP_AFFINITY: pid 832 tid 1591 thread 11 bound to OS proc set 27\nOMP: Info #250: KMP_AFFINITY: pid 832 tid 1588 thread 8 bound to OS proc set 4\nOMP: Info #250: KMP_AFFINITY: pid 832 tid 2120 thread 13 bound to OS proc set 17\nOMP: Info #250: KMP_AFFINITY: pid 832 tid 2122 thread 15 bound to OS proc set 47\nOMP: Info #250: KMP_AFFINITY: pid 832 tid 2123 thread 16 bound to OS proc set 4\nOMP: Info #250: KMP_AFFINITY: pid 832 tid 2121 thread 14 bound to OS proc set 12\nOMP: Info #250: KMP_AFFINITY: pid 832 tid 2119 thread 12 bound to OS proc set 2\n<\/code><\/pre>\n","answers":[{"AnswerId":"58074452","CreationDate":"2019-09-24T06:42:03.200","ParentId":null,"OwnerUserId":"9465768","Title":null,"Body":"<p>From the available information, looks like you have 16 (8x2) physical cores and 32 (8x2x2) logical cores. Recommended settings of 'intra_op_parallelism_threads' is equal to physical cores and 'inter_op_parallelism_threads' is equal to the number of sockets.<\/p>\n\n<p>In your case, assuming 8 models at a time, I would suggest you to try out with the following configurations.<\/p>\n\n<pre><code>config = tf.ConfigProto()\nconfig.intra_op_parallelism_threads = 2\nconfig.inter_op_parallelism_threads = 2\ntf.Session(config=config)\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>OMP_NUM_THREADS=2, \nKMP_BLOCKTIME=1; \nKMP_AFFINITY='granularity=fine,verbose,compact,1,0'; \nKMP_SETTINGS=1\n<\/code><\/pre>\n\n<p>Also try 'config.intra_op_parallelism_threads=1' and 'OMP_NUM_THREADS=1'.<\/p>\n\n<p>For more details, you can refer <a href=\"https:\/\/software.intel.com\/en-us\/articles\/maximize-tensorflow-performance-on-cpu-considerations-and-recommendations-for-inference\" rel=\"nofollow noreferrer\">https:\/\/software.intel.com\/en-us\/articles\/maximize-tensorflow-performance-on-cpu-considerations-and-recommendations-for-inference<\/a><\/p>\n\n<p>Hope this helps.<\/p>\n"}]}
+{"QuestionId":57887597,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-11T10:56:32.623","AcceptedAnswerId":null,"OwnerUserId":8273896.0,"Title":"WARNING:tensorflow:Entity","Body":"<p>i tried to run the code from this repo <a href=\"https:\/\/github.com\/zzh8829\/yolov3-tf2\/\" rel=\"nofollow noreferrer\">https:\/\/github.com\/zzh8829\/yolov3-tf2\/<\/a><\/p>\n\n<p>then I ran in to this WARNING , it does find with the CPU with the WARNING below<\/p>\n\n<pre><code>WARNING:tensorflow:Entity &lt;bound method BatchNormalization.call of &lt;yolov3_tf2.batch_norm.BatchNormalization object at 0x00000297EAA0D828&gt;&gt; could not be transformed and will be executed as-is. Please report this to the AutoGraph team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output. Cause: Bad argument number for Name: 3, expecting 4 \n<\/code><\/pre>\n\n<p>after I installed tensorflow2.0rc\nit does not work<\/p>\n\n<p>my PC specifications:\nCPU: i5-9400F<\/p>\n\n<p>GPU: MSI RTX 2060<\/p>\n\n<p>OS: Win 10<\/p>\n\n<p>CUDA: 10<\/p>\n\n<p>thank you guys<\/p>\n","answers":[{"AnswerId":"58469698","CreationDate":"2019-10-20T02:59:31.937","ParentId":null,"OwnerUserId":"746754","Title":null,"Body":"<p>I have the same problem, best I can find is that it is caused by latest version of gast (0.3.0). Need to install 0.2.2: <code>pip install -U gast==0.2.2<\/code><\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/32448\" rel=\"nofollow noreferrer\">Source1<\/a><br>\n<a href=\"https:\/\/eehoeskrap.tistory.com\/363\" rel=\"nofollow noreferrer\">Source2<\/a><\/p>\n"}]}
+{"QuestionId":57888089,"AnswerCount":1,"Tags":"<python><windows><tensorflow><gpu>","CreationDate":"2019-09-11T11:25:30.420","AcceptedAnswerId":null,"OwnerUserId":11065351.0,"Title":"gpu utilization question( nvidia-smi prints that I am using gpu but task manager shows I am not using gpu )","Body":"<p>I am running tensorflow program and just curious why nvidia-smi shows that I am using gpu but\nwhy task manager shows I am not using a gpu.<\/p>\n\n<p>I was running same tensorflow code in my computer(windows 10) and in server(ubuntu 16.04).\nI figured out the program runs fast in server and I was curious what makes it faster.\ngpu memories were more utilized in my computer than in server when I saw the result of nvidia-smi.\nand I found that nvidia-smi prints that prgram is using a gpu but task manager prints the program is not using a gpu!<\/p>\n\n<pre><code>(base) PS C:\\Users\\82102&gt; nvidia-smi\nWed Sep 11 19:35:47 2019\n+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 430.86       Driver Version: 430.86       CUDA Version: 10.2     |\n|-------------------------------+----------------------+----------------------+\n| GPU  Name            TCC\/WDDM | 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   WDDM  | 00000000:23:00.0  On |                  N\/A |\n| 55%   69C    P2   163W \/ 175W |   7745MiB \/  8192MiB |     38%      Default |\n+-------------------------------+----------------------+----------------------+\n\n<\/code><\/pre>\n\n<p>task manager shows the python program does not use gpu! ( I cannot upload images yet ) <\/p>\n\n<p>In server, the program is using gpu too.<\/p>\n\n<pre><code>[0] TITAN Xp         | 84'C,  66 % |  4516 \/ 12192 MB | yoonah(4505M)\n[1] TITAN Xp         | 31'C,   0 % |   158 \/ 12196 MB | yoonah(147M)\n<\/code><\/pre>\n\n<p>why task manager shows the program is not using gpu? really curious<\/p>\n","answers":[{"AnswerId":"58516200","CreationDate":"2019-10-23T05:46:43.557","ParentId":null,"OwnerUserId":"12198627","Title":null,"Body":"<p>Please try enabling the GPU configurations and WDDM for windows, below links can help you.<\/p>\n\n<pre><code>https:\/\/www.thewindowsclub.com\/monitor-gpu-usage-iusing-task-manager\nhttps:\/\/www.howtogeek.com\/351073\/how-to-monitor-gpu-usage-in-the-windows-task-manager\n<\/code><\/pre>\n"}]}
+{"QuestionId":57888191,"AnswerCount":1,"Tags":"<out-of-memory><pytorch><dataloader>","CreationDate":"2019-09-11T11:30:54.893","AcceptedAnswerId":57890506.0,"OwnerUserId":11090634.0,"Title":"Memory error when iterate over two dataloaders simultaneously in pytorch","Body":"<p>I am trying to train my model using 2 dataloaders from 2 different datasets. <\/p>\n\n<p>I found how to set up this by using <code>cycle() and zip()<\/code> because my datasets are not the same length from here: <a href=\"https:\/\/stackoverflow.com\/questions\/51444059\/how-to-iterate-over-two-dataloaders-simultaneously-using-pytorch\">How to iterate over two dataloaders simultaneously using pytorch?<\/a><\/p>\n\n<pre><code>  File \"\/home\/Desktop\/example\/train.py\", line 229, in train_2\n    for i, (x1, x2) in enumerate(zip(cycle(train_loader_1), train_loader_2)):\n  File \"\/home\/.conda\/envs\/3dcnn\/lib\/python3.7\/site-packages\/torch\/utils\/data\/dataloader.py\", line 346, in __next__\n    data = self.dataset_fetcher.fetch(index)  # may raise StopIteration\n  File \"\/home\/.conda\/envs\/3dcnn\/lib\/python3.7\/site-packages\/torch\/utils\/data\/_utils\/fetch.py\", line 47, in fetch\n    return self.collate_fn(data)\n  File \"\/home\/.conda\/envs\/3dcnn\/lib\/python3.7\/site-packages\/torch\/utils\/data\/_utils\/collate.py\", line 80, in default_collate\n    return [default_collate(samples) for samples in transposed]\n  File \"\/home\/.conda\/envs\/3dcnn\/lib\/python3.7\/site-packages\/torch\/utils\/data\/_utils\/collate.py\", line 80, in &lt;listcomp&gt;\n    return [default_collate(samples) for samples in transposed]\n  File \"\/home\/.conda\/envs\/3dcnn\/lib\/python3.7\/site-packages\/torch\/utils\/data\/_utils\/collate.py\", line 56, in default_collate\n    return torch.stack(batch, 0, out=out)\nRuntimeError: [enforce fail at CPUAllocator.cpp:64] . DefaultCPUAllocator: can't allocate memory: you tried to allocate 154140672 bytes. Error code 12 (Cannot allocate memory)\n\n<\/code><\/pre>\n\n<p>I tried to solve that by setting <code>num_workers=0<\/code>, decreasing the batch size, using <code>pinned_memory=False<\/code> and <code>shuffle=False<\/code>...\nBut none of it worked... I am having 256GB of RAM and 4 NVIDIA TESLA V100 GPUs. <\/p>\n\n<p>I tried to run it just by not training in 2 dataloaders simultaneously but individually and it worked. However for my project I need this parallel training with 2 datasets...<\/p>\n","answers":[{"AnswerId":"57890506","CreationDate":"2019-09-11T13:39:57.050","ParentId":null,"OwnerUserId":"11090634","Title":null,"Body":"<p>Based on <a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/1917#issuecomment-433698337\" rel=\"nofollow noreferrer\">this<\/a> discussion, instead of <code>cycle()<\/code> and <code>zip()<\/code> I avoid any errors by using:<\/p>\n\n<pre><code>  try:\n     data, target = next(dataloader_iterator)\n  except StopIteration:\n     dataloader_iterator = iter(dataloader)\n     data, target = next(dataloader_iterator)\n<\/code><\/pre>\n\n<p>kudos to @srossi93 from this pytorch post!<\/p>\n"}]}
+{"QuestionId":57888418,"AnswerCount":1,"Tags":"<c++><macos><tensorflow><bazel>","CreationDate":"2019-09-11T11:44:25.523","AcceptedAnswerId":null,"OwnerUserId":10003721.0,"Title":"Add LC_VERSION_MIN_MACOSX to libtensorflow_cc.so and libtensorflow_framework.so","Body":"<p>I want to set <code>LC_VERSION_MIN_MACOSX<\/code> to libtensorflow dynamic libraries, which I build using bazel.<\/p>\n\n<p>I've already tried bazel flag: <code>--copt=-mmacosx-version-min=10.12<\/code><br>\nand set environment: <code>export MACOSX_DEPLOYMENT_TARGET=10.12<\/code>. <br>\nHowever libtensorflow_framework.so and libtensorflow_cc.so still without <code>LC_VERSION_MIN_MACOSX<\/code> flag.<\/p>\n","answers":[{"AnswerId":"57902656","CreationDate":"2019-09-12T08:25:21.453","ParentId":null,"OwnerUserId":"10003721","Title":null,"Body":"<p>I managed to do it:\nadding <code>--action_env MACOSX_DEPLOYMENT_TARGET=10.12<\/code> resolved my problem<\/p>\n"}]}
+{"QuestionId":57888656,"AnswerCount":1,"Tags":"<python><tensorflow><batch-normalization><chainer>","CreationDate":"2019-09-11T11:58:30.620","AcceptedAnswerId":57889857.0,"OwnerUserId":12051692.0,"Title":"Why does Chainer Batchnormalization not work well with mnist_nn","Body":"<p>Chainer batchnormalization does not work well with my code although batch normalization of tensorflow works. I use the dataset mnist the code below shows. <\/p>\n\n<p>Using chainer(version=6.1.0), without batchnormalization, the validation accuracy is within 0.97 and 0.98 after 100 epochs whereas with batchnormalization, it is less than 0.80 after 100 epochs.<\/p>\n\n<p>When I use the same way with tensorflow(version=1.14.0), the validation accuracy is around 0.98 in a both way, with batchnormalization or without batchnormalization.<\/p>\n\n<p>This is the part of my code.\nThe number of epoch is 100, and its batchsize is 1000. I use the Adam as the optimizer with learning_rate 0.01.<\/p>\n\n<p>dataset, train data, validation data<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>(train_images, train_labels), (test_images, test_labels) = mnist.load_data()\ntrain_images=train_images.reshape(60000, 28*28)\ntest_images = test_images.reshape(10000, 28*28)\n\nx_train = train_images.astype('float32')\/255\ny_train = train_labels.astype('int32')\nx_val = test_images.astype('float32')\/255\ny_val = test_labels.astype('int32')\n<\/code><\/pre>\n\n<p>model and condition (chainer)<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Define model\nclass MyModel(Chain):\n    def __init__(self,n_in=784,n_hidden=100,n_out=10):\n        initializer = chainer.initializers.HeNormal()\n        super().__init__()\n        with self.init_scope():\n            self.l1=L.Linear(n_in, n_hidden, initialW=initializer)\n            self.l2=L.Linear(n_hidden, n_hidden, initialW=initializer)\n            self.l3=L.Linear(n_hidden, n_out, initialW=initializer)\n            self.bn=L.BatchNormalization(n_hidden, decay=0.99, eps=0.001)\n\n    def forward(self,x):\n        h = F.relu(self.bn(self.l1(x)))\n        h = F.relu(self.bn(self.l2(h)))\n        return self.l3(h)\n\nmodel = MyModel()\noptimizer = optimizers.Adam()\noptimizer.setup(model)       \n\nn_epoch = 100\nn_batchsize = 1000 \n<\/code><\/pre>\n\n<p>model and condition (tensorflow)<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>n_inputs = 28 * 28\nn_hidden1 = 100\nn_hidden2 = 100\nn_outputs = 10\n\n\nbatch_norm_momentum = 0.9\n\nX = tf.placeholder(tf.float32, shape=(None, n_inputs), name=\"X\")\ny = tf.placeholder(tf.int32, shape=(None), name=\"y\")\ntraining = tf.placeholder_with_default(False, shape=(), name='training')\n\nwith tf.name_scope(\"dnn\"):\n    he_init = tf.variance_scaling_initializer()\n    my_batch_norm_layer = partial(tf.layers.batch_normalization, \n                                  training=training,\n                                  momentum=batch_norm_momentum)\n\n    my_dense_layer = partial(tf.layers.dense,\n                             kernel_initializer=he_init)\n\n    hidden1 = my_dense_layer(X, n_hidden1, name=\"hidden1\")\n    bn1 = tf.nn.relu(my_batch_norm_layer(hidden1))\n    hidden2 = my_dense_layer(bn1, n_hidden2, name=\"hidden2\")\n    bn2 = tf.nn.relu(my_batch_norm_layer(hidden2))\n    logits_before_bn = my_dense_layer(bn2, n_outputs, name=\"outputs\")\n    logits = my_batch_norm_layer(logits_before_bn)\n\nwith tf.name_scope(\"loss\"):\n    xentropy = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y, logits=logits)\n    loss = tf.reduce_mean(xentropy, name=\"loss\")\n\nwith tf.name_scope(\"train\"):\n    optimizer = tf.train.AdamOptimizer(learning_rate=0.001)\n    training_op = optimizer.minimize(loss)\n\nwith tf.name_scope(\"eval\"):\n    correct = tf.nn.in_top_k(logits, y, 1)\n    accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))\n<\/code><\/pre>\n\n<p>All code I use with chainer<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport chainer\nfrom chainer import cuda, Function, gradient_check, report, training, utils, Variable\nfrom chainer import datasets, iterators, optimizers, serializers\nfrom chainer import Link, Chain, ChainList\nimport chainer.functions as F\nimport chainer.links as L\nfrom chainer.training import extensions\nfrom keras.datasets import mnist\nimport cupy as cp\n\n\n\n(train_images, train_labels), (test_images, test_labels) = mnist.load_data()\ntrain_images=train_images.reshape(60000, 28*28)\ntest_images = test_images.reshape(10000, 28*28)\n\nx_train = train_images.astype('float32')\/255\ny_train = train_labels.astype('int32')\nx_val = test_images.astype('float32')\/255\ny_val = test_labels.astype('int32')\n\n\n# Define model\nclass MyModel(Chain):\n    def __init__(self,n_in=784,n_hidden=100,n_out=10):\n        initializer = chainer.initializers.HeNormal()\n        super().__init__()\n        with self.init_scope():\n            self.l1=L.Linear(n_in, n_hidden, initialW=initializer)\n            self.l2=L.Linear(n_hidden, n_hidden, initialW=initializer)\n            self.l3=L.Linear(n_hidden, n_out, initialW=initializer)\n            self.bn=L.BatchNormalization(n_hidden, decay=0.9, eps=0.001)\n\n    def forward(self,x):\n        h = F.relu(self.bn(self.l1(x)))\n        h = F.relu(self.bn(self.l2(h)))\n        return self.l3(h)\n\n# define optimizer\nmodel = MyModel()\noptimizer = optimizers.Adam(alpha=0.01)\noptimizer.setup(model)       \n\n## learn network\nn_epoch = 100\nn_batchsize = 1000 \niteration = 0\n\ngpu_id = 0\ncuda.get_device(gpu_id).use()\n\n# send the network to gpu memory\nmodel.to_gpu(gpu_id)\n\nprint(\"epoch  train\/loss  val\/loss  train\/acc   val\/acc\")\n\nfor epoch in range(n_epoch):\n    # order dataset randomly\n    order = np.random.permutation(range(len(x_train)))\n\n    loss_list = []\n    accuracy_list = []\n\n    for i in range(0, len(order), n_batchsize):\n        index = order[i:i+n_batchsize]\n        x_train_batch = x_train[index,:]\n        y_train_batch = y_train[index]\n\n        x_train_batch = cp.asarray(x_train_batch)\n        y_train_batch = cp.asarray(y_train_batch)\n\n        output_train = model(x_train_batch)\n\n        loss_train_batch = F.softmax_cross_entropy(output_train, y_train_batch)\n        accuracy_train_batch = F.accuracy(output_train, y_train_batch)\n\n        loss_list.append(cuda.to_cpu(loss_train_batch.array))\n        accuracy_list.append(cuda.to_cpu(accuracy_train_batch.array))\n\n        model.cleargrads()\n        loss_train_batch.backward()\n\n        optimizer.update()\n\n        iteration += 1\n\n    loss_train = np.mean(loss_list)\n    accuracy_train = np.mean(accuracy_list)\n\n    # after one epoch, evaluate with validation data\n    x_val = cp.asarray(x_val)\n    y_val = cp.asarray(y_val)\n    with chainer.using_config('train', False), chainer.using_config('enable_backprop', False):\n        output_val = model(x_val)\n\n    loss_val = F.softmax_cross_entropy(output_val, y_val)\n    loss_val = cuda.to_cpu(loss_val.array)\n    accuracy_val = F.accuracy(output_val, y_val)\n    accuracy_val = cuda.to_cpu(accuracy_val.array)\n\n    print('{0:&gt;4d} {1:&gt;10.4f} {2:&gt;10.4f} {3:&gt;10.4f} {4:&gt;10.4f}'.format(epoch,loss_train,loss_val,accuracy_train,accuracy_val))\n<\/code><\/pre>\n\n<p>All code I use with tensorflow<\/p>\n\n<pre><code>python\nimport tensorflow as tf\nfrom keras.datasets import mnist\nfrom functools import partial\nimport numpy as np\n\ndef shuffle_batch(X, y, batch_size):\n    rnd_idx = np.random.permutation(len(X))\n    n_batches = len(X) \/\/ batch_size\n    for batch_idx in np.array_split(rnd_idx, n_batches):\n        X_batch, y_batch = X[batch_idx], y[batch_idx]\n        yield X_batch, y_batch\n\n(train_images, train_labels), (test_images, test_labels) = mnist.load_data()\ntrain_images=train_images.reshape(60000, 28*28)\ntest_images = test_images.reshape(10000, 28*28)\n\nX_train = train_images.astype('float32')\/255\ny_train = train_labels.astype('int32')\nX_valid = test_images.astype('float32')\/255\ny_valid = test_labels.astype('int32')\n\nn_inputs = 28 * 28\nn_hidden1 = 100\nn_hidden2 = 100\nn_outputs = 10\n\nbatch_norm_momentum = 0.9\n\nX = tf.placeholder(tf.float32, shape=(None, n_inputs), name=\"X\")\ny = tf.placeholder(tf.int32, shape=(None), name=\"y\")\ntraining = tf.placeholder_with_default(False, shape=(), name='training')\n\nwith tf.name_scope(\"dnn\"):\n    he_init = tf.variance_scaling_initializer()\n    my_batch_norm_layer = partial(tf.layers.batch_normalization, \n                                  training=training,\n                                  momentum=batch_norm_momentum)\n\n    my_dense_layer = partial(tf.layers.dense)\n\n    hidden1 = my_dense_layer(X, n_hidden1, name=\"hidden1\")\n    bn1 = tf.nn.relu(my_batch_norm_layer(hidden1))\n    hidden2 = my_dense_layer(bn1, n_hidden2, name=\"hidden2\")\n    bn2 = tf.nn.relu(my_batch_norm_layer(hidden2))\n    logits_before_bn = my_dense_layer(bn2, n_outputs, name=\"outputs\")\n    logits = my_batch_norm_layer(logits_before_bn)\n\nwith tf.name_scope(\"loss\"):\n    xentropy = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y, logits=logits)\n    loss = tf.reduce_mean(xentropy, name=\"loss\")\n\nlearning_rate = 0.01\n\nwith tf.name_scope(\"train\"):\n    optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)\n    training_op = optimizer.minimize(loss)\n\nwith tf.name_scope(\"eval\"):\n    correct = tf.nn.in_top_k(logits, y, 1)\n    accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))\n\ninit = tf.global_variables_initializer()\nsaver = tf.train.Saver()\n\nn_epochs = 100\nbatch_size = 1000\n\nextra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\n\nprint(\"epoch  train\/loss  val\/loss  train\/acc   val\/acc\")\n\nwith tf.Session() as sess:\n    init.run()\n    for epoch in range(n_epochs):\n        loss_list = []\n        accuracy_list = []\n\n        for X_batch, y_batch in shuffle_batch(X_train, y_train, batch_size):\n            sess.run([training_op, extra_update_ops], \n                     feed_dict={training: True, X: X_batch, y: y_batch})\n\n            loss_batch = loss.eval(feed_dict={X: X_batch, y: y_batch})\n            accuracy_batch = accuracy.eval(feed_dict={X: X_batch, y: y_batch})\n\n            loss_list.append(loss_batch)\n            accuracy_list.append(accuracy_batch)\n\n        loss_val = loss.eval(feed_dict={X: X_valid, y: y_valid}) \n        accuracy_val = accuracy.eval(feed_dict={X: X_valid, y: y_valid})\n\n        print('{0:&gt;4d} {1:&gt;10.4f} {2:&gt;10.4f} {3:&gt;10.4f} {4:&gt;10.4f}'\n              .format(epoch,np.mean(loss_list),loss_val,np.mean(accuracy_list),accuracy_val))\n<\/code><\/pre>\n\n<p>I expect batch normalization with chainer would be around 98% but it got less than 80%.\nAm I using the batchnormalization with chainer in a wrong way, or the structure of batchnormalization differs a lot between chainer and tensorflow??<\/p>\n","answers":[{"AnswerId":"57889857","CreationDate":"2019-09-11T13:03:27.307","ParentId":null,"OwnerUserId":"10688898","Title":null,"Body":"<p>In order to use different batch statistics between the layers, the model definition has to be like the following code, which achieves 98% validation accuracy after 100 epochs in my environment.<\/p>\n\n<pre><code>class MyModel(Chain):\n    def __init__(self,n_in=784,n_hidden=100,n_out=10):\n        initializer = chainer.initializers.HeNormal()\n        super().__init__()\n        with self.init_scope():\n            self.l1=L.Linear(n_in, n_hidden, initialW=initializer)\n            self.l2=L.Linear(n_hidden, n_hidden, initialW=initializer)\n            self.l3=L.Linear(n_hidden, n_out, initialW=initializer)\n            self.bn1=L.BatchNormalization(n_hidden, decay=0.9, eps=0.001)\n            self.bn2=L.BatchNormalization(n_hidden, decay=0.9, eps=0.001)\n\n    def forward(self,x):\n        h = F.relu(self.bn1(self.l1(x)))\n        h = F.relu(self.bn2(self.l2(h)))\n        return self.l3(h)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57889082,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><predict><bert-language-model>","CreationDate":"2019-09-11T12:22:49.433","AcceptedAnswerId":null,"OwnerUserId":7131957.0,"Title":"How to predict masked word in a sentence in BERT-base from Tensorflow checkpoint (ckpt) files?","Body":"<p>I have BERT-base model checkpoints which I trained from scratch in Tensorflow. How can I use those checkpoints to predict masked word in a given sentence?<\/p>\n\n<p>Like, let say sentence is,\n\"[CLS] abc pqr [MASK] xyz [SEP]\"\nAnd I want to predict word at [MASK] position.<\/p>\n\n<p>How can I do it?\nI searched a lot online but everyone is using BERT for their task specific classification tasks.\nNot using BERT to predict masked word.<\/p>\n\n<p>Please help me in solving this prediction problem. <\/p>\n\n<p>I created data using <code>create_pretraining_data.py<\/code> &amp; trained model from scratch using <code>run_pretraining.py<\/code> from official BERT repo (<a href=\"https:\/\/github.com\/google-research\/bert\" rel=\"nofollow noreferrer\">https:\/\/github.com\/google-research\/bert<\/a>)<\/p>\n\n<p>I have searched in issues in official bert repo. But didn't found any solution.<\/p>\n\n<p>Also looked at code in that repo. They're using Estimator which they are training not using from checkpoints weights.<\/p>\n\n<p>Didn't found any way to use way to use Tensorflow checkpoints of BERT-base model (trained from scratch) to predict word masked token (i.e. [MASK]).<\/p>\n","answers":[{"AnswerId":"57891376","CreationDate":"2019-09-11T14:27:16.760","ParentId":null,"OwnerUserId":"3518108","Title":null,"Body":"<p>Do you definitely need to start from a TF checkpoint? If you can use one of the pretrained models used in the <a href=\"https:\/\/huggingface.co\/pytorch-transformers\/\" rel=\"nofollow noreferrer\">pytorch-transformers<\/a> library, I wrote a library for doing exactly this: <a href=\"https:\/\/pypi.org\/project\/fitbert\/\" rel=\"nofollow noreferrer\">FitBERT<\/a>. <\/p>\n\n<p>If you have to start with a TF checkpoint, there are scripts for converting from a TF checkpoint to something pytorch-transformers can use, <a href=\"https:\/\/huggingface.co\/pytorch-transformers\/converting_tensorflow_models.html#bert\" rel=\"nofollow noreferrer\">link<\/a>, and after converting you should be able to use FitBERT, or you can just see what we're doing in the code. <\/p>\n"}]}
+{"QuestionId":57889159,"AnswerCount":0,"Tags":"<python><tensorflow><lstm><tensorflow2.0>","CreationDate":"2019-09-11T12:27:22.873","AcceptedAnswerId":null,"OwnerUserId":11698102.0,"Title":"TF2.0 Custom LSTMCell with RNN: always outputs as a single class","Body":"<p>Implementation on TF 2.0 :<\/p>\n\n<p>I have implemented a Custom LSTMCell and fed that to an <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/layers\/RNN\" rel=\"nofollow noreferrer\">RNN-Encapsulation layer<\/a><\/p>\n\n<p>My Target is to have an identical behaviour like tf.keras.layers.LSTM.\nAnd I am not looking to optimize the same for GPU, at this point.<\/p>\n\n<p>The model is build to predict 7 output classes and the model always predicts a single class.<\/p>\n\n<p><strong>Everything is correct with the dataset, works well with the standard tf.keras.layers.LSTM<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/e2tph.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/e2tph.png\" alt=\"Here is how the confusion matrix looks like\"><\/a><\/p>\n\n<p>May be there is an obvious error which I am unable to figure out.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class LSTMCell_Layer(Layer):\n\n    def __init__(self, units=100, n_input=6, batch_size=16, n_classes=7):\n        self.units = units\n        self.state_size = units\n        super(LSTMCell_Layer, self).__init__()\n\n        w_init = tf.random_normal_initializer(stddev=random_stddev)\n\n        self.Wxh  = tf.Variable(initial_value=w_init(shape=(n_input, 4*n_hidden),dtype='float32'),trainable=True, name='Wxh')\n\n        self.Whh = tf.Variable(initial_value=w_init(shape=(n_hidden, 4 *n_hidden),dtype='float32'),trainable=True, name='Whh')\n\n        b_init = tf.zeros_initializer()\n        self.bias = tf.Variable(initial_value=b_init(shape=(4*n_hidden), dtype='float32'),trainable=True,name='bias')\n\n        self.h = tf.Variable(initial_value=w_init(shape=(batch_size, n_hidden),dtype='float32'),trainable=True, name='h')\n\n        self.forget_bias = 1.0\n        self.c = tf.Variable(initial_value=w_init(shape=(batch_size, n_hidden),dtype='float32'),trainable=True, name='c')\n\n    def call(self, X, state):\n\n        self.h = state[0]\n        W_full = tf.concat([self.Wxh, self.Whh], 0)\n        concat = tf.concat([X, self.h], 1) # concat for speed.\n        concat = tf.matmul(concat, W_full) + self.bias\n        i, j, f, o = tf.split(concat, 4, 1)\n        g = tf.tanh(j)\n        new_c = self.c*tf.sigmoid(f+self.forget_bias) + tf.sigmoid(i)*g\n        self.c = new_c\n        new_h = tf.tanh(new_c) * tf.sigmoid(o)\n    return new_c, [new_h]\n<\/code><\/pre>\n\n<p>Followed by the model definition :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class MyModel(tf.keras.Model):\n    def __init__(self):\n        super(MyModel, self).__init__()\n        self.b1 = RNN(LSTMCell_Layer(100))\n        self.b2 = Dense(100, activation='relu')\n        self.b3 = Dense(n_classes, activation='softmax')\n\n    def call(self,inputs):\n        x = self.b1(inputs)\n        x = self.b2(x)\n        x = self.b3(x)\nreturn x\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57889504,"AnswerCount":0,"Tags":"<tensorflow><dataset><distributed>","CreationDate":"2019-09-11T12:45:13.580","AcceptedAnswerId":null,"OwnerUserId":6494418.0,"Title":"boosting training speed with tf.estimator and tf.data.dataset","Body":"<p>I'm training a net on a distributed cluster,\nI reference the [<a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/data_performance\" rel=\"nofollow noreferrer\">official API: tf.data Performance<\/a>]<\/p>\n\n<p>and code my <code>input_fn<\/code><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def input_fn(\n        file_pattern,\n        feature_spec,\n        label_name,\n        batch_size=32,\n        num_epochs=1,\n        shuffle_buffer_size=2048,\n        num_processor=8):\n    \"\"\"base input_fn.\"\"\"\n    dataset = tf.data.Dataset.list_files(file_pattern)  # shuffle default\n    # num_parallel_calls must less than or equal to cycle_length.\n    dataset = dataset.interleave(tf.data.TFRecordDataset, cycle_length=15, num_parallel_calls=15)\n    dataset = dataset.map(lambda x: _parse_fn(x, feature_spec, label_name), num_parallel_calls=15)\n    if shuffle_buffer_size &gt; 1: dataset = dataset.shuffle(buffer_size=shuffle_buffer_size)\n    if num_epochs &gt; 1: dataset = dataset.repeat(num_epochs)\n    dataset = dataset.batch(batch_size)\n    dataset = dataset.prefetch(prefetch_buffer_size=15 * batch_size)\n\n    return dataset\n<\/code><\/pre>\n\n<p>I try different args of <code>cycle_length<\/code>, <code>num_parallel_calls<\/code>, <code>prefetch_buffer_size<\/code>, but my training also not full my CPU or GPU or MEM or NET_BRANCH.<\/p>\n\n<p>I can't figure out the bottleneck of my code, and I would appreciate for your reply!<\/p>\n\n<p>here is my log of compute resource,<\/p>\n\n<p>AND: my data storage with HDFS format with TFRecord, and classifier is a custom estimator!<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/fqq1P.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/fqq1P.png\" alt=\"CPU and MEM log\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/3wMct.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3wMct.png\" alt=\"NETWORK log\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/nMriG.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/nMriG.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57890162,"AnswerCount":0,"Tags":"<tensorflow><unity3d><tensorflow-lite><ml-agent>","CreationDate":"2019-09-11T13:19:37.907","AcceptedAnswerId":null,"OwnerUserId":4334452.0,"Title":"How to use tensorflow lite mode in unity3d's ml-agent inference engine","Body":"<p>I tried a lot but there isn't a straight answer to this .My inquiry ,is it possible to use tensorflow lite model in unity3d's inference engine of ML-Agents ,if yes ,then how ? if not ,what is the alternative (if any) ?<\/p>\n\n<p>Really appreciate any help here\nAyad<\/p>\n","answers":[]}
+{"QuestionId":57890219,"AnswerCount":2,"Tags":"<python-3.x><tensorflow>","CreationDate":"2019-09-11T13:22:29.003","AcceptedAnswerId":null,"OwnerUserId":8270077.0,"Title":"ImportError: cannot import name 'extract_zipped_paths'","Body":"<p>I am trying to access the listed datasets in Tensorflow 2.0.  I use Windows 10.<\/p>\n\n<p>I have run the following commands:<\/p>\n\n<pre><code>pip install tfds-nightly\n\npip install tensorflow-datasets\n<\/code><\/pre>\n\n<p>The version of Tensorflow in my PC is:<\/p>\n\n<pre><code>2.0.0-alpha0\n<\/code><\/pre>\n\n<p>(An attempt to upgrade it to the more recent version b run into an exception:<\/p>\n\n<pre><code>  Found existing installation: wrapt 1.10.11\nERROR: Cannot uninstall 'wrapt'. It is a distutils installed project and thus we cannot accurately determine which files belong to it which would lead to only a partial uninstall.\n<\/code><\/pre>\n\n<p>)<\/p>\n\n<p>When I try to import the module:<\/p>\n\n<pre><code>import tensorflow_datasets as tfds\n<\/code><\/pre>\n\n<p>I get the following error message:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nImportError                               Traceback (most recent call last)\n&lt;ipython-input-30-46a8a2031c9c&gt; in &lt;module&gt;()\n----&gt; 1 import tensorflow_datasets as tfds\n\nC:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Anaconda3_64\\lib\\site-packages\\tensorflow_datasets\\__init__.py in &lt;module&gt;()\n     44 # needs to happen before anything else, since the imports below will try to\n     45 # import tensorflow, too.\n---&gt; 46 from tensorflow_datasets.core import tf_compat\n     47 tf_compat.ensure_tf_install()\n     48 \n\nC:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Anaconda3_64\\lib\\site-packages\\tensorflow_datasets\\core\\__init__.py in &lt;module&gt;()\n     21 tf_compat.ensure_tf_install()\n     22 \n---&gt; 23 from tensorflow_datasets.core.dataset_builder import BeamBasedBuilder  # pylint:disable=g-import-not-at-top\n     24 from tensorflow_datasets.core.dataset_builder import BuilderConfig\n     25 from tensorflow_datasets.core.dataset_builder import DatasetBuilder\n\nC:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Anaconda3_64\\lib\\site-packages\\tensorflow_datasets\\core\\dataset_builder.py in &lt;module&gt;()\n     33 from tensorflow_datasets.core import constants\n     34 from tensorflow_datasets.core import dataset_utils\n---&gt; 35 from tensorflow_datasets.core import download\n     36 from tensorflow_datasets.core import file_format_adapter\n     37 from tensorflow_datasets.core import lazy_imports_lib\n\nC:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Anaconda3_64\\lib\\site-packages\\tensorflow_datasets\\core\\download\\__init__.py in &lt;module&gt;()\n     17 \n     18 from tensorflow_datasets.core.download.checksums import add_checksums_dir\n---&gt; 19 from tensorflow_datasets.core.download.download_manager import DownloadConfig\n     20 from tensorflow_datasets.core.download.download_manager import DownloadManager\n     21 from tensorflow_datasets.core.download.extractor import iter_archive\n\nC:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Anaconda3_64\\lib\\site-packages\\tensorflow_datasets\\core\\download\\download_manager.py in &lt;module&gt;()\n     32 from tensorflow_datasets.core import utils\n     33 from tensorflow_datasets.core.download import checksums\n---&gt; 34 from tensorflow_datasets.core.download import downloader\n     35 from tensorflow_datasets.core.download import extractor\n     36 from tensorflow_datasets.core.download import resource as resource_lib\n\nC:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Anaconda3_64\\lib\\site-packages\\tensorflow_datasets\\core\\download\\downloader.py in &lt;module&gt;()\n     31 import promise\n     32 import requests\n---&gt; 33 from requests.utils import extract_zipped_paths\n     34 from six.moves import urllib\n     35 import tensorflow as tf\n\nImportError: cannot import name 'extract_zipped_paths'\n<\/code><\/pre>\n","answers":[{"AnswerId":"57901225","CreationDate":"2019-09-12T06:55:01.560","ParentId":null,"OwnerUserId":"10335017","Title":null,"Body":"<p>You need to downgrade the version of requests library to 2.21.0<\/p>\n\n<pre><code>pip install requests==2.21.0\n<\/code><\/pre>\n\n<p>I had the exact same problem, with <strong>'2.0.0-beta1'<\/strong> version of tensorflow, and had I pip installed tensorflow_datasets, which resulted in the error you have mentioned. <\/p>\n\n<p>I noticed that while installing tensorflow_datasets, the requests module gets updated to latest version 2.22.0, which seems to be unable to import 'extract_zipped_paths', even though this function is defined in the requests.utils submodule, which you can find here:<\/p>\n\n<p><a href=\"https:\/\/2.python-requests.org\/en\/master\/_modules\/requests\/utils\/\" rel=\"nofollow noreferrer\">https:\/\/2.python-requests.org\/en\/master\/_modules\/requests\/utils\/<\/a><\/p>\n\n<p>Also, the requests module seemed to be working just fine on its own, without TF. I dont know where the problem is originating from, but, downgrading the requests module to 2.21.0 works without any issue.<\/p>\n\n<p>Summary:<\/p>\n\n<pre><code>tf.__version__ == '2.0.0-beta1'\n\npip install tensorflow_datasets \n\npip install requests==2.21.0\n<\/code><\/pre>\n"},{"AnswerId":"57898810","CreationDate":"2019-09-12T01:59:27.400","ParentId":null,"OwnerUserId":"2886527","Title":null,"Body":"<p>Remove all existing version of requests. And install specifically <code>pip install requests==2.22.0<\/code><\/p>\n"}]}
+{"QuestionId":57890249,"AnswerCount":0,"Tags":"<tensorflow><kaldi>","CreationDate":"2019-09-11T13:24:21.660","AcceptedAnswerId":null,"OwnerUserId":10271919.0,"Title":"I can't use once saved LSTM model","Body":"<p>I cannot reproduce Trained model in LSTM.<\/p>\n\n<p>I would like to change the following code:<\/p>\n\n<p><a href=\"https:\/\/github.com\/kaldi-asr\/kaldi\/blob\/master\/egs\/wsj\/s5\/steps\/tfrnnlm\/lstm_fast.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/kaldi-asr\/kaldi\/blob\/master\/egs\/wsj\/s5\/steps\/tfrnnlm\/lstm_fast.py<\/a><\/p>\n\n<p>Specifically speaking, I would like to load Train model and\nwant to use it in another code after saved it , since \nit is useful that we separate code to save , and code to load.<\/p>\n\n<p>I saved Train (and Valid)model, and in another code ,\nI challenged loading it,\nbut I encountered the next Error;<\/p>\n\n<pre><code>initializer = tf.random_uniform_initializer(-config.init_scale,\n                                                config.init_scale)\nwith tf.Session() as sess:\n    saver = tf.train.import_meta_graph(\"~~(mypath)~~\/rnnlm\/model.ckpt-0.meta\")\n    saver.restore(sess,tf.train.latest_checkpoint(\"~~(my path)~~\/rnnlm\/.\"))\n    valid_input = RnnlmInput(config=config, data=valid_data,name=\"ValidInput\")\n    mvalid = RnnlmModel(is_training=False, config=config, input_=valid_input)\n    valid_perplexity = run_epoch(sess,mvalid)\n\nFailedPreconditionError :\nAttempting to use unitialized value embedding.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57890259,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2019-09-11T13:25:02.570","AcceptedAnswerId":null,"OwnerUserId":5252187.0,"Title":"How do I add mulitipile preprocessing functions to a DataGenerator in keras?","Body":"<p>I have a a training DataGenerator with a preprocessing function (could be InceptionV3's,ResNet50's etc. preprocess input) as follows:<\/p>\n\n<pre><code>train_datagen = ImageDataGenerator(\n        preprocessing_function = preprocess_input,\n        rotation_range=30,\n        width_shift_range=0.2,\n        height_shift_range=0.2,\n        shear_range=0,\n        brightness_range= [0.5,1.0],\n        zoom_range=0.1,\n        channel_shift_range=10,\n        vertical_flip=True,\n        horizontal_flip=True,\n        fill_mode='nearest')\n<\/code><\/pre>\n\n<p>And now I want to add another preprocessing function:<\/p>\n\n<pre><code>get_random_eraser(v_l=0, v_h=1, pixel_level=False)\n<\/code><\/pre>\n\n<p>So all in all it would chain the model's preprocessing functions (as it was pretrained in keras) and implement another preprocessing function. How do I do that? (preprocessing_function can't get a list [f1, f2], and I cannot repeat preprocessing_function declaration in a new line(makes sense, but tried that anyway))<\/p>\n","answers":[{"AnswerId":"57890807","CreationDate":"2019-09-11T13:57:47.367","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>You need to create a new generator with your custom function:<\/p>\n\n<pre><code>def custom_function(input_image):\n    input_image = preprocess_input(input_image)\n    return get_random_eraser(v_l=0, v_h=1, pixel_level=False)(input_image)\n\nnew_train_gen = train_datagen = ImageDataGenerator(\n    preprocessing_function = custom_function,\n    rotation_range=30,\n    width_shift_range=0.2,\n    height_shift_range=0.2,\n    shear_range=0,\n    brightness_range= [0.5,1.0],\n    zoom_range=0.1,\n    channel_shift_range=10,\n    vertical_flip=True,\n    horizontal_flip=True,\n    fill_mode='nearest')\n<\/code><\/pre>\n\n<p>It may be interesting to add a <code>print(input_image.shape)<\/code> inside the <code>custom_function<\/code> in the first attempt just to make sure whether <code>input_image<\/code> is a single image or a batch. You may need to adjust your inner functions accordingly. <\/p>\n"}]}
+{"QuestionId":57890875,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-11T14:00:55.700","AcceptedAnswerId":null,"OwnerUserId":8270077.0,"Title":"AttributeError: type object 'DatasetV2' has no attribute 'from_tensor'","Body":"<p>I am learning Tensorflow 2.0 and I am getting an error as I run the following code:<\/p>\n\n<pre><code>data = np.random.randint(0,10, (3,4))\ndataset1 = tf.data.Dataset.from_tensor(data)\n\n---------------------------------------------------------------------------\nAttributeError                            Traceback (most recent call last)\n&lt;ipython-input-38-9873d7e2b8bc&gt; in &lt;module&gt;()\n----&gt; 1 dataset1 = tf.data.Dataset.from_tensor(data)\nAttributeError: type object 'DatasetV2' has no attribute 'from_tensor'\n<\/code><\/pre>\n\n<p>The same here:<\/p>\n\n<pre><code>e= tf.data.Dataset.from_element(10)\nAttributeError: type object 'DatasetV2' has no attribute 'from_element'\n<\/code><\/pre>\n\n<p>But I can run the code:<\/p>\n\n<pre><code>dataset = tf.data.Dataset.range(10)\n<\/code><\/pre>\n\n<p>without a problem.<\/p>\n","answers":[]}
+{"QuestionId":57890885,"AnswerCount":1,"Tags":"<pytorch><image-recognition>","CreationDate":"2019-09-11T14:01:36.937","AcceptedAnswerId":null,"OwnerUserId":2878388.0,"Title":"Applying a pytorch CNN to video?","Body":"<p>I am looking for some advice on how to apply a pytorch CNN to a video as opposed to an image.  <\/p>\n\n<p>Picture a drone flying over an area and using video to capture some objects below. I have a CNN trained on images of objects, and want to count the objects in the video.  <\/p>\n\n<p>Currently my strategy has been to convert the video to frames as PNGs and running the CNN on those PNGs. this seems inefficient, and I am struggling with how to count the objects without duplicating (frame 1 and frame 1+n will overlap).  <\/p>\n\n<p>It would be appreciated if someone had some advice, or a suggested tutorial\/code set that did this.  Thanks in advance.  <\/p>\n","answers":[{"AnswerId":"57911745","CreationDate":"2019-09-12T17:22:43.830","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>PyTorch at the moment doesn't have <a href=\"https:\/\/pytorch.org\/docs\/stable\/torchvision\/io.html\" rel=\"nofollow noreferrer\">support to detect and track objects<\/a> in a video. <\/p>\n\n<p>You would need to create your own logic for that.<\/p>\n\n<p>The support is limited to read the video and audio from a file, read frames and timestamps, and write the video read more in <a href=\"https:\/\/pytorch.org\/blog\/pytorch-1.2-and-domain-api-release\/#torchvision-04-with-support-for-video\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>What you will basically need to do is to create an object tracking, frame by frame together by keeping their with their square positions and based on that decide if the same object or not.<\/p>\n\n<p>If you have a drone flying and inspecting people you may check <a href=\"https:\/\/deepmind.com\/research\/open-source\/kinetics\" rel=\"nofollow noreferrer\">Kinetics<\/a> to detect human actions:<\/p>\n\n<ul>\n<li>ResNet 3D 18<\/li>\n<li>ResNet MC 18<\/li>\n<li>ResNet (2+1)D<\/li>\n<\/ul>\n\n<p>All based on <a href=\"https:\/\/arxiv.org\/abs\/1711.11248\" rel=\"nofollow noreferrer\">Kinetics-400<\/a>\nBut the newer one is <a href=\"https:\/\/arxiv.org\/abs\/1907.06987\" rel=\"nofollow noreferrer\">Kinetics-700<\/a>.<\/p>\n"}]}
+{"QuestionId":57891174,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-11T14:16:45.783","AcceptedAnswerId":57892583.0,"OwnerUserId":4166865.0,"Title":"Error incompatible shapes in function model.fit()","Body":"<p>I am new in Keras. I want to try U-net. I used this tutorial from tensorflow: <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/samples\/outreach\/blogs\/segmentation_blogpost\/image_segmentation.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/samples\/outreach\/blogs\/segmentation_blogpost\/image_segmentation.ipynb<\/a>.\nI used the code for U-net creation with my own dataset. They have got images 256x256x3 and I made my images with same shape.\nNow, I got error:<\/p>\n\n<pre><code> InvalidArgumentError: Incompatible shapes: [1376256] vs. [458752]\n     [[{{node training\/Adam\/gradients\/loss\/conv2d_23_loss\/mul_grad\/BroadcastGradientArgs}}] ]\n<\/code><\/pre>\n\n<p>It is in function model.fit(). I have got 130 training examples and batch size is 5 (I know, that those number are small...).\nPlease, Can anybody know, what can cause this error in function model.fit()?<\/p>\n\n<p>Thank you for your help.<\/p>\n","answers":[{"AnswerId":"57892583","CreationDate":"2019-09-11T15:40:49.430","ParentId":null,"OwnerUserId":"3472735","Title":null,"Body":"<p>1376256 is exactly 3 x 458752. I suspect you are not correctly accounting for your channels somewhere. As this appears to be on your output layer, it may be that you're trying to predict 3 classes when there are only 1. <\/p>\n\n<p>In future, or if this doesn't help, please provide more information including the code for your model and number of classes you're trying to predict, so people can better help. <\/p>\n"}]}
+{"QuestionId":57891175,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-11T14:16:49.560","AcceptedAnswerId":57891473.0,"OwnerUserId":2886575.0,"Title":"tensorflow find minimum distance to real points","Body":"<p>I have a Bx3 tensor, <code>foo<\/code> of B= batch size 3D points.  Through some fanciness or another, I get another tensor, <code>bar<\/code> that is of shape Bx6x3, and where each of the B 6x3 matrices corresponds to a point in <code>foo<\/code>.  That 6x3 matrix consists of 6 complex-valued 3D points.  What I would like to do is, for each of my B points, find the closest <em>real-valued<\/em> point out of the 6 in <code>bar<\/code> to the corresponding point in <code>foo<\/code>, ending up with a new Bx3 <code>min_bar<\/code> consisting of the closest points in <code>bar<\/code> to points in <code>foo<\/code>.<\/p>\n\n<p>In <code>numpy<\/code>, I can accomplish this feat with masked arrays:<\/p>\n\n<pre><code>foo = np.array([\n    [1,2,3],\n    [4,5,6],\n    [7,8,9]])\n# here bar is only Bx2x3 for simplicity, but the solution generalizes\nbar = np.array([\n    [[2,3,4],[1+0.1j,2+0.1j,3+0.1j]],\n    [[6,5,4],[4,5,7]],\n    [[1j,1j,1j],[0,0,0]],\n])\n\n#mask complex elements of bar\nbar_with_masked_imag = np.ma.array(bar)\ncandidates = bar_with_masked_imag.imag == 0\nbar_with_masked_imag.mask = ~candidates\n\ndists = np.sum(bar_with_masked_imag**2, axis=1)\nmindists = np.argmin(dists, axis=1)\nfoo_indices = np.arange(foo.shape[0])\nmin_bar = np.array(\n    bar_with_masked_imag[foo_indices,mindists,:], \n    dtype=float\n)\n\nprint(min_bar)\n#[[2. 3. 4.]\n# [4. 5. 7.]\n# [0. 0. 0.]]\n<\/code><\/pre>\n\n<p>However, tensorflow doesn't have masked arrays and such.  How do I translate this into tensorflow?<\/p>\n","answers":[{"AnswerId":"57891473","CreationDate":"2019-09-11T14:33:00.117","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Here is a way to do that:<\/p>\n\n<pre><code>import tensorflow as tf\nimport math\n\ndef solution_tf(foo, bar):\n    foo = tf.convert_to_tensor(foo)\n    bar = tf.convert_to_tensor(bar)\n    # Get real and imaginary parts\n    bar_r = tf.cast(tf.real(bar), foo.dtype)\n    bar_i = tf.imag(bar)\n    # Mask of all real-valued points\n    m = tf.reduce_all(tf.equal(bar_i, 0), axis=-1)\n    # Distance to every corresponding point\n    d = tf.reduce_sum(tf.squared_difference(tf.expand_dims(foo, 1), bar_r), axis=-1)\n    # Replace distances of complex points with infinity\n    d2 = tf.where(m, d, tf.fill(tf.shape(d), tf.constant(math.inf, d.dtype)))\n    # Find smallest distances\n    idx = tf.argmin(d2, axis=1)\n    # Get points with smallest distances\n    b = tf.range(tf.shape(foo, out_type=idx.dtype)[0])\n    return tf.gather_nd(bar_r, tf.stack([b, idx], axis=1))\n\n# Test\nwith tf.Graph().as_default(), tf.Session() as sess:\n    foo = tf.constant([\n        [1,2,3],\n        [4,5,6],\n        [7,8,9]], dtype=tf.float32)\n    bar = tf.constant([\n        [[2,3,4],[1+0.1j,2+0.1j,3+0.1j]],\n        [[6,5,4],[4,5,7]],\n        [[1j,1j,1j],[0,0,0]]], dtype=tf.complex64)\n    sol_tf = solution_tf(foo, bar)\n    print(sess.run(sol_tf))\n    # [[2. 3. 4.]\n    #  [4. 5. 7.]\n    #  [0. 0. 0.]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57891504,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-11T14:34:54.347","AcceptedAnswerId":57891832.0,"OwnerUserId":7327257.0,"Title":"Tensorflow: fill tensor inside a function","Body":"<p>What I'm trying to define is the following idea:<\/p>\n\n<p>Consider whe have these tensors<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>a = tf.constant([1., 1.5, 1.2]) # tensor with shape [3,]\nb = tf.constant([1., 2., 3.])   # \"\"\nc = tf.constant([3., 0., 6.])   # \"\"\n\nt = tf.constant([0.5, 0.6, 0.7, 2., 4., 5., 6.]) # tensor with shape [7,]\n\n<\/code><\/pre>\n\n<p>Now let's consider I want to compute a new tensor, working with each element of previous tensors, for example:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def new_tensor(a, b, c, t):\n    X = tf.constant([[tf.sin(a*t[1]), b*t[3], c+t[4]], \n        [tf.cos(b*t[5]), tf.atan2(t[5], c), a+t[2]+b],  \n        [a+t[4], a+b, c*t[0]]])\n    return X\n<\/code><\/pre>\n\n<p><code>X<\/code> should be a tensor with shape <code>[3, 3, 3]<\/code>. That is, I want to define a function that takes four tensors as input: three of them of same shape and the fourth with a different one. I want the function to compute a tensor (<code>X<\/code>) for each value of first three inputs (<code>a, b, c<\/code>).<\/p>\n\n<p>With this code TensorFlow gives this error:<\/p>\n\n<pre><code>TypeError: List of Tensors when single Tensor expected\n<\/code><\/pre>\n\n<p>According to this <a href=\"https:\/\/stackoverflow.com\/questions\/35661677\/tensorflow-generating-a-random-constant\/35662013#35662013\">post<\/a>, this is beacuse <code>tf.constant<\/code> cannot take a tensor as input, and they recommend to use <code>tf.Variable<\/code> instead. But I don't think that fits for me because I have to work later with <code>X<\/code> and don't want to initializate it, etc. I have also read this <a href=\"https:\/\/stackoverflow.com\/questions\/49148375\/updating-a-tensor-in-tensorflow\">other post<\/a>, but couldn't find any answer to my problem.<\/p>\n\n<p>Is there any way to do what I want? Does my code make sense for my purpose? Thank you in advance.<\/p>\n\n<p><strong>UPDATE: with jdehesa answer<\/strong><\/p>\n\n<p>Taking @jdehesa answer and making resulting tensor more simple:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def new_tensor(a, b, c, t):\n    # Could also use tf.convert_to_tensor\n    X = tf.stack([[a+t[1],   b*t[1],    c+t[1]],\n                  [b*t[0],  t[5]+ c,  a+t[2]+b],\n                  [a+t[4],      a+b,   c*t[0]]])\n    return X\n<\/code><\/pre>\n\n<p>And with tensors:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>a = tf.constant([1., 1., 1.]) # tensor with shape [3,]\nb = tf.constant([2., 2., 2.])   # \"\"\nc = tf.constant([3., 3., 3.])   # \"\"\n\nt = tf.constant([1., 1., 1., 1., 1., 1., 1.]) # tensor with shape [7,]\n<\/code><\/pre>\n\n<p>What I get is the following tensor:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># When evaluating x = new_tensor(a,b,c,t)\n[[[2. 2. 2.]\n  [2. 2. 2.]\n  [4. 4. 4.]]\n\n [[2. 2. 2.]\n  [4. 4. 4.]\n  [4. 4. 4.]]\n\n [[2. 2. 2.]\n  [3. 3. 3.]\n  [3. 3. 3.]]]\n\n<\/code><\/pre>\n\n<p>But what I would expected is the following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>[[[2. 2. 4.]\n  [2. 4. 4.]\n  [2. 3. 3.]]\n\n [[2. 2. 4.]\n  [2. 4. 4.]\n  [2. 3. 3.]]\n\n [[2. 2. 4.]\n  [2. 4. 4.]\n  [2. 3. 3.]]]\n<\/code><\/pre>\n\n<p>As what I want it to evaluate to each element of input tensors.<\/p>\n","answers":[{"AnswerId":"57891832","CreationDate":"2019-09-11T14:55:10.573","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>That's correct, you can only pass Python or NumPy values to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/constant\" rel=\"nofollow noreferrer\"><code>tf.constant<\/code><\/a>, but you can build your tensor with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/stack\" rel=\"nofollow noreferrer\"><code>tf.stack<\/code><\/a> or, if you prefer, generally with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/convert_to_tensor\" rel=\"nofollow noreferrer\"><code>tf.convert_to_tensor<\/code><\/a>:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef new_tensor(a, b, c, t):\n    # Could also use tf.convert_to_tensor\n    X = tf.stack([[tf.sin(a*t[1]),            b*t[3],   c+t[4]],\n                  [tf.cos(b*t[5]), tf.atan2(t[5], c), a+t[2]+b],\n                  [        a+t[4],               a+b,   c*t[0]]])\n    return X\n\nwith tf.Graph().as_default(), tf.Session() as sess:\n    a = tf.constant([1., 1.5, 1.2]) # tensor with shape [3,]\n    b = tf.constant([1., 2., 3.])   # \"\"\n    c = tf.constant([3., 0., 6.])   # \"\"\n    t = tf.constant([0.5, 0.6, 0.7, 2., 4., 5., 6.]) # tensor with shape [7,]\n    x = new_tensor(a, b, c, t)\n    print(sess.run(x))\n    # [[[ 0.5646425   0.7833269   0.65938467]\n    #   [ 2.          4.          6.        ]\n    #   [ 7.          4.         10.        ]]\n    # \n    #  [[ 0.2836622  -0.8390715  -0.7596879 ]\n    #   [ 1.0303768   1.5707964   0.69473827]\n    #   [ 2.7         4.2         4.9       ]]\n    # \n    #  [[ 5.          5.5         5.2       ]\n    #   [ 2.          3.5         4.2       ]\n    #   [ 1.5         0.          3.        ]]]\n<\/code><\/pre>\n\n<hr>\n\n<p>EDIT: For your second example, to get the result that you want you need to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/transpose\" rel=\"nofollow noreferrer\"><code>tf.transpose<\/code><\/a> to change the order of the dimensions of the tensor:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef new_tensor(a, b, c, t):\n    # Could also use tf.convert_to_tensor\n    X = tf.stack([[a+t[1],   b*t[1],    c+t[1]],\n                  [b*t[0],  t[5]+ c,  a+t[2]+b],\n                  [a+t[4],      a+b,   c*t[0]]])\n    X = tf.transpose(X, (2, 0, 1))\n    return X\n\nwith tf.Graph().as_default(), tf.Session() as sess:\n    a = tf.constant([1., 1., 1.]) # tensor with shape [3,]\n    b = tf.constant([2., 2., 2.])   # \"\"\n    c = tf.constant([3., 3., 3.])   # \"\"\n    t = tf.constant([1., 1., 1., 1., 1., 1., 1.]) # tensor with shape [7,]\n    x = new_tensor(a, b, c, t)\n    print(sess.run(x))\n    # [[[2. 2. 4.]\n    #   [2. 4. 4.]\n    #   [2. 3. 3.]]\n    # \n    #  [[2. 2. 4.]\n    #   [2. 4. 4.]\n    #   [2. 3. 3.]]\n    # \n    #  [[2. 2. 4.]\n    #   [2. 4. 4.]\n    #   [2. 3. 3.]]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57891620,"AnswerCount":0,"Tags":"<python><python-3.x><signal-processing><pytorch><max-pooling>","CreationDate":"2019-09-11T14:42:29.183","AcceptedAnswerId":null,"OwnerUserId":9793316.0,"Title":"Downsample audios with PyTorch AvgPool1d","Body":"<p>Since MaxPool2d or AvgPool2d can down-sample 2D images. I am wondering if AvgPool1d can also be used to do down-sample audios.<\/p>\n\n<p>Here's my little experiment<\/p>\n\n<h2>Generating dummy signal<\/h2>\n\n<pre><code>import numpy as np\nimport torch\nfrom torch.nn import AvgPool1d\nimport matplotlib.pyplot as plt\nimport resampy\nimport scipy\nfs = 500\ns = np.linspace(0,1,fs)\nx = np.sin(2*np.pi*5*s) + np.cos(2*np.pi*10*s)\\\n    + np.sin(2*np.pi*15*s) + np.sin(2*np.pi*30*s)\nplt.plot(x)\nplt.title('Original signal')\nplt.xlabel('samples')\nplt.ylabel('amplitude')\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/icLxd.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/icLxd.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<h2>Comparing Kaiser resampling and AvgPool1D resampling<\/h2>\n\n<pre><code>re_x1 = resampy.resample(x, 2,1) # resampling using kaiser\ntorch_resample = AvgPool1d(2,2) # resampling using Pooling\nx_torch = torch_resample(torch.tensor(x[None,None,:]))\n\nplt.plot(re_x1)\nplt.plot(x_torch.view(-1).numpy(), 'r--')\nplt.legend(['kaiser_best', 'torch pooling'])\nplt.title('Down-sampled signal')\nplt.xlabel('samples')\nplt.ylabel('amplitude')\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/hjPXE.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/hjPXE.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Checking if they are close to each other<\/p>\n\n<pre><code>np.allclose(x_torch.view(-1).numpy()[10:-10], re_x1[10:-10], 1)\n&gt;&gt;&gt; False\n<\/code><\/pre>\n\n<h2>My Question<\/h2>\n\n<p>Is there any problem to down-sample audios by using AvgPool1d rather than traditional methods by using fft or kaiser?<\/p>\n\n<p>Also, is there anyway to approximate traditional down-sampling methods by using AvgPool1d as close as possible?<\/p>\n","answers":[]}
+{"QuestionId":57891759,"AnswerCount":0,"Tags":"<python><nlp><pytorch>","CreationDate":"2019-09-11T14:51:08.630","AcceptedAnswerId":null,"OwnerUserId":7198441.0,"Title":"Code example in Hugging Face Pytorch-Transformers quickstart documentation","Body":"<p>I am following the quick start guide <a href=\"https:\/\/huggingface.co\/pytorch-transformers\/quickstart.html\" rel=\"nofollow noreferrer\">here<\/a>. The problem is they have provided code for GPU machine and I am running the code on CPU based Ubuntu machine. I have commented the lines that put everything in CUDA. The code is now showing error and I don't know how to resolve it.\nThe question is \"How can I make this work?\"<\/p>\n\n<p>I have checked <a href=\"https:\/\/stackoverflow.com\/questions\/50495053\/if-im-not-specifying-to-use-cpu-gpu-which-one-is-my-script-using\">this<\/a> answer and this is not what I'm looking for.<\/p>\n\n<p>The full code is <a href=\"https:\/\/huggingface.co\/pytorch-transformers\/quickstart.html#bert-example\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n1. Using BertModel to encode inputs in hidden-states:\n\n<pre><code>#Load pre-trained model (weights)\nmodel = BertModel.from_pretrained('bert-base-uncased')\n\n#Set the model in evaluation mode to desactivate the DropOut modules\n# This is IMPORTANT to have reproducible results during evaluation!\nmodel.eval()\n\n#***I have commented these 3 lines*** \n\n# If you have a GPU, put everything on cuda\n#tokens_tensor = tokens_tensor.to('cuda')\n#segments_tensors = segments_tensors.to('cuda')\n#model.to('cuda')\n\n#Rest all is untouched\n# *** -----------------***---------------***\n\n# Predict hidden states features for each layer\nwith torch.no_grad():\n    # See the models docstrings for the detail of the inputs\n    outputs = model(tokens_tensor, token_type_ids=segments_tensors)\n # PyTorch-Transformers models always output tuples.\n # See the models docstrings for the detail of all the outputs\n # In our case, the first element is the hidden state of the last layer of the Bert model\n    encoded_layers = outputs[0]\n# We have encoded our input sequence in a FloatTensor of shape (batch size, sequence length, model hidden dimension)\nassert tuple(encoded_layers.shape) == (1, len(indexed_tokens), model.config.hidden_size)\n<\/code><\/pre>\n\n<p>Error for 1:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nAttributeError                            Traceback (most recent call last)\n&lt;ipython-input-40-a86e9643e7f3&gt; in &lt;module&gt;\n     11 \n     12 # We have encoded our input sequence in a FloatTensor of shape (batch size, sequence length, model hidden dimension)\n---&gt; 13 assert tuple(encoded_layers).shape == (1, len(indexed_tokens), model.config.hidden_size)\n\nAttributeError: 'tuple' object has no attribute 'shape'\n<\/code><\/pre>\n\n<h1>2. Using BertForMaskedLM to predict a masked token:<\/h1>\n\n<pre><code># Load pre-trained model (weights)\nmodel = BertForMaskedLM.from_pretrained('bert-base-uncased')\nmodel.eval()\n\n#***---------------Commented--------------------------\n# If you have a GPU, put everything on cuda\n#tokens_tensor = tokens_tensor.to('cuda')\n#segments_tensors = segments_tensors.to('cuda')\n#model.to('cuda')\n\n#***---------------------------------------------\n\n# Predict all tokens\nwith torch.no_grad():\n    outputs = model(tokens_tensor, token_type_ids=segments_tensors)\n    predictions = outputs[0]\n\n# confirm we were able to predict 'henson'\npredicted_index = torch.argmax(predictions[0, masked_index]).item()\npredicted_token = tokenizer.convert_ids_to_tokens([predicted_index])[0]\nassert predicted_token == 'henson'\n<\/code><\/pre>\n\n<p>Error for 2:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nAssertionError                            Traceback (most recent call last)\n&lt;ipython-input-42-9b965490d278&gt; in &lt;module&gt;\n     17 predicted_index = torch.argmax(predictions[0, masked_index]).item()\n     18 predicted_token = tokenizer.convert_ids_to_tokens([predicted_index])[0]\n---&gt; 19 assert predicted_token == 'henson'\n\nAssertionError:\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57891845,"AnswerCount":1,"Tags":"<tensorflow><dataset><pipeline>","CreationDate":"2019-09-11T14:56:08.180","AcceptedAnswerId":null,"OwnerUserId":10399430.0,"Title":"Speeding up data pipeline with tf.data.dataset API","Body":"<p>I am having trouble speeding up the data pipeline for training with tf.data.dataset and I think I am missing something here. With the different options within dataset to pre-load data, dataset speed is still slow. <\/p>\n\n<p>I have a complex data pipeline but I simplified to a small example below. I tried fine-tuning the num_parallel_calls, cycle_length, prefetch etc but I cannot seem to get to smooth dataset generation. What am I missing? Any suggestions? enter code here<\/p>\n\n<pre><code>import tensorflow as tf\ntf.enable_eager_execution()\nfrom timeit import default_timer as timer\n\nfeature_count = 400\nbatch_size = 1024\nlook_back = 100\ntarget_groups = 21\n\ndef random_data_generator(x=0):\n    while True:\n        x_data = tf.random.uniform(\n            shape=(batch_size,  look_back, feature_count),\n            minval=-1.0,\n            maxval=5,\n            dtype=tf.dtypes.float32)\n\n        Y_data = tf.random.uniform(\n            shape=(batch_size, target_groups),\n            minval=1,\n            maxval=21,\n            dtype=tf.dtypes.int32)\n\n        yield x_data, Y_data\n\n\ndef get_simple_Dataset_generator(): \n\n    dataset = tf.data.Dataset.from_tensor_slices([0,1,2])\n    dataset = dataset.interleave(lambda x: tf.data.Dataset.from_generator(random_data_generator, \n                            output_types=(tf.float32, tf.float32), args=(x,)),\n                            cycle_length=3,\n                            block_length=3,\n                            num_parallel_calls= tf.data.experimental.AUTOTUNE)\n\n    #dataset = dataset.prefetch(2)\n    while True: \n        for x, Y in dataset:\n            yield x, Y\n\n\ndef test_speed():\n    generator = get_simple_Dataset_generator()\n    print(\"Testing generator speed \")\n\n    for i in range(1,100): \n        start_time = timer()\n        next(generator)\n        lap_time = timer()-start_time\n        print(\"%s Time - %fsec \"%(i, lap_time))\n\n\nif __name__ == '__main__': \n    test_speed()```\n\n\n\n\nI was hoping to see consistent generator speed but it still very erratic.\nOutput\n1 Time - 3.417578sec\n2 Time - 1.257846sec\n3 Time - 1.286210sec\n4 Time - 0.000456sec\n5 Time - 0.027772sec\n6 Time - 0.058985sec\n7 Time - 0.000416sec\n8 Time - 0.026721sec\n9 Time - 0.027316sec\n10 Time - 0.777332sec\n11 Time - 1.379266sec\n12 Time - 1.172304sec\n13 Time - 0.000365sec\n14 Time - 0.026909sec\n15 Time - 0.045409sec\n16 Time - 0.000708sec\n17 Time - 0.025682sec\n18 Time - 0.027223sec\n19 Time - 0.577131sec\n20 Time - 1.220682sec\n21 Time - 1.189601sec\n22 Time - 0.000573sec\n23 Time - 0.079531sec\n24 Time - 0.624080sec\n25 Time - 0.038932sec\n<\/code><\/pre>\n","answers":[{"AnswerId":"57983222","CreationDate":"2019-09-17T23:36:51.960","ParentId":null,"OwnerUserId":"1097517","Title":null,"Body":"<p>I believe that eager exicution has some kinks that are still being worked out. Changing your code to graph based execution I get each iteration is about <code>0.06<\/code> seconds in stead of the <code>1<\/code> second times from the eager version.<\/p>\n\n<p>Here is the code snippet:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom timeit import default_timer as timer\n\nfeature_count = 400\nbatch_size = 1024\nlook_back = 100\ntarget_groups = 21\n\ndef random_data(x=0):\n    x_data = tf.random.uniform(\n            shape=(batch_size,  look_back, feature_count),\n            minval=-1.0,\n            maxval=5,\n            dtype=tf.dtypes.float32)\n\n    Y_data = tf.random.uniform(\n            shape=(batch_size, target_groups),\n            minval=1,\n            maxval=21,\n            dtype=tf.dtypes.int32)\n\n    return x_data, Y_data\n\n\ndef get_simple_Dataset(): \n\n    dataset = tf.data.Dataset.from_tensor_slices(tf.zeros(100))\n    dataset = dataset.map(random_data)\n\n    dataset = dataset.prefetch(10)\n\n    return dataset\n\ndef test_speed():\n    dataset = get_simple_Dataset()\n    iterator = dataset.make_one_shot_iterator()\n    fetch = iterator.get_next()\n\n    with tf.Session() as sess:     \n        for i in range(1,100): \n            start_time = timer()\n            sess.run(fetch)\n            lap_time = timer()-start_time\n            print(\"%s Time - %fsec \"%(i, lap_time))\n\n\nif __name__ == '__main__': \n    test_speed()\n\n1 Time - 0.108968sec \n2 Time - 0.071986sec \n3 Time - 0.068198sec \n4 Time - 0.065433sec \n5 Time - 0.066582sec \n6 Time - 0.064175sec \n7 Time - 0.067372sec \n8 Time - 0.064265sec \n9 Time - 0.065510sec \n10 Time - 0.068043sec\n<\/code><\/pre>\n"}]}
+{"QuestionId":57891854,"AnswerCount":1,"Tags":"<python><tensorflow><keras><pyspark><apache-spark-sql>","CreationDate":"2019-09-11T14:56:39.483","AcceptedAnswerId":null,"OwnerUserId":1439962.0,"Title":"Best practice for feeding spark dataframes for training Tensorflow network","Body":"<p>I want to feed data coming from spark clusters, to train a deep network. I do not have GPUs in the nodes, so distributed TensorFlow or packages like <code>elephas<\/code> is not an option. <\/p>\n\n<p>I have come up with the following generator which does the job. It just retrieves the next batch from Spark. In order to handle batches I am adding an extra column <code>index<\/code> (which is simply and incremental id column), and filter on that on each call for the next batch.  <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nclass SparkBatchGenerator(tfk.utils.Sequence):\n    def __init__(self, spark_df, batch_size, sample_count=None, feature_col='features', label_col='labels'):\n        w = Window().partitionBy(sf.lit('a')).orderBy(sf.lit('a'))\n        df = spark_df.withColumn('index', sf.row_number().over(w)).sort('index')\n        self.X = df.select([feature_col, 'index'])\n        self.y = df.select([label_col, 'index'])\n\n        self.data_count = sample_count if sample_count else spark_df.count()\n        self.feature_col = feature_col\n        self.label_col = label_col\n        self.batch_size = batch_size\n\n    def __len__(self):\n        return np.ceil(self.data_count \/self.batch_size).astype(int)\n\n\n    def __getitem__(self, idx):\n        start, end = idx * self.batch_size, (idx + 1) * self.batch_size\n        batch_x = (\n            self.X.filter(f'index &gt;= {start} and index &lt; {end}')\n                  .toPandas()[self.feature_col]\n                  .apply(lambda x: x.toArray()).tolist()\n        )\n        batch_y = (\n            self.y.filter(f'index &gt;= {start} and index &lt; {end}')\n                  .toPandas()[self.label_col].tolist()\n        )\n\n\n        return np.array(batch_x), np.array(batch_y)\n\n<\/code><\/pre>\n\n<p>This works, but of course is slow, especial when <code>batch_size<\/code> is small. I was just wondering if anyone has a better solution.<\/p>\n","answers":[{"AnswerId":"57964059","CreationDate":"2019-09-16T20:34:03.057","ParentId":null,"OwnerUserId":"1439962","Title":null,"Body":"<p>I used <code>tf.data.Dataset<\/code> to handel this. I can buffer the data coming from spark, and then leave the job of batch creation to tensorflow dataset api. It is now much faster:<\/p>\n\n<pre><code>class MyGenerator(object):\n    def __init__(self, spark_df, buffer_size, feature_col='features', label_col='labels'):\n        w = Window().partitionBy(sf.lit('a')).orderBy(sf.lit('a'))\n        self.df = (\n            spark_df.withColumn('index', sf.row_number().over(w) - 1)\n                    .sort('index')\n                    .select([feature_col, 'index', label_col])\n        )\n\n        self.feature_col = feature_col\n        self.label_col = label_col\n        self.buffer_size = buffer_size\n\n\n    def generate_data(self):\n        idx = 0\n        buffer_counter = 0\n        buffer = self.df.filter(f'index &gt;= {idx} and index &lt; {buffer_size}').toPandas()\n        while len(buffer) &gt; 0:\n            if idx &lt; len(buffer):\n                X = (\n                    buffer.iloc[idx][self.feature_col].toArray()\/255.0\n                )\n                y = (\n                   buffer.iloc[idx][self.label_col]\n                )\n\n                idx += 1\n\n                yield X.reshape((28, 28)), y\n            else:\n                buffer = self.df.filter(\n                    f'index &gt;= {buffer_counter * buffer_size} and index &lt; {(buffer_counter + 1)*buffer_size}'\n                ).toPandas()\n                idx = 0\n                buffer_counter += 1\n<\/code><\/pre>\n\n<pre><code>batch_size = 128\nbuffer_size = 4*1024\n\nmy_gen = MyGenerator(feature_df, buffer_size)\ndataset = tf.data.Dataset.from_generator(my_gen.generate_data, output_types=(tf.float32, tf.int32))\ndataset = dataset.prefetch(tf.contrib.data.AUTOTUNE).batch(batch_size, drop_remainder=True)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57891940,"AnswerCount":0,"Tags":"<python><windows><tensorflow><tensorboard>","CreationDate":"2019-09-11T15:01:24.197","AcceptedAnswerId":null,"OwnerUserId":11902119.0,"Title":"Python Tensorflow Tensorboard 'NotFoundError' error message in Windows","Body":"<p>I'm using Tensorflow with Python 3.7.4 (64bit) in Windows 10. <\/p>\n\n<p>I've built a convolutional neural network model and it runs fine in Jupyter. Now I'd like to visualise it's performance using Tensorboard. But trying to set this up I get an error message. <\/p>\n\n<pre><code># Setting up Tensorboard to view model performance \nNAME = \"Trains_vs_Cars_16by2_CNN_{}\".format(int(time.time()))\ntensorboard = TensorBoard(log_dir=\"logs\/{}\".format(NAME))\n\nmodel.fit(X, y,\n      batch_size=25,\n      epochs=5,\n      validation_split=0.2,\n      callbacks=[tensorboard])\n\n# ERROR MESSAGE \n     NotFoundError                             Traceback (most recent call last)\n     &lt;ipython-input-6-c627053c0717&gt; in &lt;module&gt;\n     67           epochs=5,\n     68           validation_split=0.2,\n---&gt; 69           callbacks=[tensorboard])\n<\/code><\/pre>\n\n<p>A poster on this page (<a href=\"https:\/\/github.com\/tensorflow\/tensorboard\/issues\/2023#\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorboard\/issues\/2023#<\/a>) mentions there's a windows specific Tensorflow bug. Is that what I've run into? I'm new to Tensorflow (and Python). <\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
+{"QuestionId":57891997,"AnswerCount":0,"Tags":"<python-3.x><keras><deep-learning><amd-processor>","CreationDate":"2019-09-11T15:04:10.823","AcceptedAnswerId":null,"OwnerUserId":11452219.0,"Title":"Plaidml doesn't use the GPU card","Body":"<p>I installed PlaidML on Anaconda enviroment to use my GPU (Radeon RX Vega 56). The installing process was fine, but when I tried to run a LSTM algorithm only run using my CPU.<\/p>\n\n<p>Another problem was an error with the constraints in the network, if I tried to run with Plaidml and the constranints an exception occurs:<\/p>\n\n<blockquote>\n  <p>PlaidMLKerasException: Unsupported variable constraint.<\/p>\n<\/blockquote>\n\n<p>If I delete the constraints, the script work, but with the issue before mentioned.<\/p>\n\n<p>For the GPU issue I tried both approches that I saw:<\/p>\n\n<pre><code>import plaidml.keras\nplaidml.keras.install_backend()\n\nimport os\nos.environ[\"KERAS_BACKEND\"] = \"plaidml.keras.backend\"\n<\/code><\/pre>\n\n<p>With both approches the script runs but on my CPU, so what I need to change\/configure to be able to use the GPU??<\/p>\n","answers":[]}
+{"QuestionId":57892128,"AnswerCount":0,"Tags":"<tensorflow><bazel>","CreationDate":"2019-09-11T15:11:44.943","AcceptedAnswerId":null,"OwnerUserId":4334452.0,"Title":"Cannot install bazel on mac","Body":"<p><a href=\"https:\/\/i.stack.imgur.com\/lX0eX.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lX0eX.png\" alt=\"enter image description here\"><\/a>Im trying to install speciifc bazel (0.26.1) version on my Mac machine and doesnt seem to be installed ,when I run bazel --version ,I get command cannot be found .\nI following the exact steps on the bazel site :\n<a href=\"https:\/\/docs.bazel.build\/versions\/master\/install-os-x.html#install-with-installer-mac-os-x\" rel=\"nofollow noreferrer\">https:\/\/docs.bazel.build\/versions\/master\/install-os-x.html#install-with-installer-mac-os-x<\/a> .\nWhen I checked the bin folder as per the guidelines ,I find a shortcut and that's all <\/p>\n\n<p>I include snapshots showing the issue .<\/p>\n\n<p>Appreciate any help \nAyad<a href=\"https:\/\/i.stack.imgur.com\/p12TV.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/p12TV.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57892492,"AnswerCount":1,"Tags":"<tensorflow><gradient>","CreationDate":"2019-09-11T15:34:53.153","AcceptedAnswerId":null,"OwnerUserId":2927842.0,"Title":"Why the gradient of categorical crossentropy loss with respect to logits is 0 with gradient tape in TF2.0?","Body":"<p>I am learning Tensorflow 2.0 and I am trying to figure out how Gradient Tapes work. I have this simple example, in which, I evaluate the cross entropy loss between logits and labels. I am wondering why the gradients with respect to logits is being zero. (Please look at the code below).\nThe version of TF is tensorflow-gpu==2.0.0-rc0.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>logits = tf.Variable([[1, 0, 0], [1, 0, 0], [1, 0, 0]], type=tf.float32)\nlabels = tf.constant([[1, 0, 0], [0, 1, 0], [0, 0, 1]],dtype=tf.float32)\nwith tf.GradientTape(persistent=True) as tape:\n    loss = tf.reduce_sum(tf.losses.categorical_crossentropy(labels, logits))\n\ngrads = tape.gradient(loss, logits)\nprint(grads)\n<\/code><\/pre>\n\n<p>I am getting<\/p>\n\n<pre><code> tf.Tensor(\n[[0. 0. 0.]\n [0. 0. 0.]\n [0. 0. 0.]], shape=(3, 3), dtype=float32)\n\n<\/code><\/pre>\n\n<p>as a result, but should not it tell me how much should I change logits in order to minimize the loss?<\/p>\n","answers":[{"AnswerId":"57894760","CreationDate":"2019-09-11T18:19:11.307","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>When calculate the cross entropy loss, set <code>from_logits=True<\/code> in the <code>tf.losses.categorical_crossentropy()<\/code>. In default, it's false, which means you are directly calculate the cross entropy loss using <code>-p*log(q)<\/code>. By setting the <code>from_logits=True<\/code>, you are using <code>-p*log(softmax(q))<\/code> to calculate the loss. <\/p>\n\n<p>Update:<\/p>\n\n<p>Just find one interesting results. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>logits = tf.Variable([[0.8, 0.1, 0.1]], dtype=tf.float32)\nlabels = tf.constant([[1, 0, 0]],dtype=tf.float32)\n\nwith tf.GradientTape(persistent=True) as tape:\n    loss = tf.reduce_sum(tf.keras.losses.categorical_crossentropy(labels, logits, from_logits=False))\n\ngrads = tape.gradient(loss, logits)\nprint(grads)\n<\/code><\/pre>\n\n<p>The grads will be <code>tf.Tensor([[-0.25  1.    1.  ]], shape=(1, 3), dtype=float32)<\/code><\/p>\n\n<p>Previously, I thought tensorflow will use <code>loss=-\\Sigma_i(p_i)\\log(q_i)<\/code> to calculate the loss, and if we derive on <code>q_i<\/code>, we will have the derivative be <code>-p_i\/q_i<\/code>. So, the expected grads should be <code>[-1.25, 0, 0]<\/code>. But the output grads looks like all increased by 1. But it won't affect the optimization process. <\/p>\n\n<p>For now, I'm still trying to figure out why the grads will be increased by one. After reading the source code of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/f03fe1bf795061408fafda9704c485aea7bddde2\/tensorflow\/python\/keras\/backend.py#L4351\" rel=\"nofollow noreferrer\">tf.categorical_crossentropy<\/a>, I found that even though we set <code>from_logits=False<\/code>, it still normalize the probabilities. That will change the final gradient expression. Specifically, the gradient will be <code>-p_i\/q_i+p_i\/sum_j(q_j)<\/code>. If <code>p_i=1<\/code> and <code>sum_j(q_j)=1<\/code>, the final gradient will plus one. That's why the gradient will be <code>-0.25<\/code>, however, I haven't figured out why the last two gradients would be <code>1.<\/code>.<\/p>\n\n<p>To prove that all gradients are increased by <code>1\/sum_j(q_j)<\/code>,<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>logits = tf.Variable([[0.5, 0.1, 0.1]], dtype=tf.float32)\nlabels = tf.constant([[1, 0, 0]],dtype=tf.float32)\n\nwith tf.GradientTape(persistent=True) as tape:\n    loss = tf.reduce_sum(tf.keras.losses.categorical_crossentropy(labels, logits, from_logits=False))\n\ngrads = tape.gradient(loss, logits)\nprint(grads)\n<\/code><\/pre>\n\n<p>The grads are <code>tf.Tensor([[-0.57142866  1.4285713   1.4285713 ]]<\/code>, which should be <code>[-2,0,0]<\/code>. <\/p>\n\n<p>It shows that all gradients are increased by <code>1\/(0.5+0.1+0.1)<\/code>. For the <code>p_i==1<\/code>, the gradient increased by <code>1\/(0.5+0.1+0.1)<\/code> makes sense to me. But I don't understand why <code>p_i==0<\/code>, the gradient is still increased by <code>1\/(0.5+0.1+0.1)<\/code>. <\/p>\n"}]}
+{"QuestionId":57892892,"AnswerCount":2,"Tags":"<java><tensorflow><tensorflow-serving>","CreationDate":"2019-09-11T15:58:58.443","AcceptedAnswerId":57917473.0,"OwnerUserId":6265370.0,"Title":"Java gRPC client predict call to half_plus_two example model","Body":"<p>I'm trying to make a call from a Java client to Tensorflow Serving. The running model is the half_plus_two example model. I can make a REST call successfully. But cannot make the gRPC equivalent call.<\/p>\n\n<p>I have tried passing a string as model input and also an array of floats into tensor proto builder. The tensor proto seems to contain correct data when I print it out:\n[1.0, 2.0, 5.0]<\/p>\n\n<pre><code>String host = \"localhost\";\n        int port = 8500;\n        \/\/ the model's name.\n        String modelName = \"half_plus_two\";\n        \/\/ model's version\n        long modelVersion = 123;\n        \/\/ assume this model takes input of free text, and make some sentiment prediction.\n\/\/        String modelInput = \"some text input to make prediction with\";\n        String modelInput = \"{\\\"instances\\\": [1.0, 2.0, 5.0]\";\n\n        \/\/ create a channel\n        ManagedChannel channel = ManagedChannelBuilder.forAddress(host, port).usePlaintext().build();\n        tensorflow.serving.PredictionServiceGrpc.PredictionServiceBlockingStub stub = tensorflow.serving.PredictionServiceGrpc.newBlockingStub(channel);\n\n        \/\/ create a modelspec\n        tensorflow.serving.Model.ModelSpec.Builder modelSpecBuilder = tensorflow.serving.Model.ModelSpec.newBuilder();\n        modelSpecBuilder.setName(modelName);\n        modelSpecBuilder.setVersion(Int64Value.of(modelVersion));\n        modelSpecBuilder.setSignatureName(\"serving_default\");\n\n        Predict.PredictRequest.Builder builder = Predict.PredictRequest.newBuilder();\n        builder.setModelSpec(modelSpecBuilder);\n\n        \/\/ create the TensorProto and request\n\n        float[] floatData = new float[3];\n        floatData[0] = 1.0f;\n        floatData[1] = 2.0f;\n        floatData[2] = 5.0f;\n\n\n        org.tensorflow.framework.TensorProto.Builder tensorProtoBuilder = org.tensorflow.framework.TensorProto.newBuilder();\n        tensorProtoBuilder.setDtype(DataType.DT_FLOAT);\n        org.tensorflow.framework.TensorShapeProto.Builder tensorShapeBuilder = org.tensorflow.framework.TensorShapeProto.newBuilder();\n        tensorShapeBuilder.addDim(org.tensorflow.framework.TensorShapeProto.Dim.newBuilder().setSize(3));\n        tensorProtoBuilder.setTensorShape(tensorShapeBuilder.build());\n\n        \/\/ Set the float_val field.\n        for (int i = 0; i &lt; floatData.length; i++) {\n            tensorProtoBuilder.addFloatVal(floatData[i]);\n        }\n\n\n        org.tensorflow.framework.TensorProto tp = tensorProtoBuilder.build();\n\n        System.out.println(tp.getFloatValList());\n\n        builder.putInputs(\"inputs\", tp);\n\n        Predict.PredictRequest request = builder.build();\n        Predict.PredictResponse response = stub.predict(request);\n<\/code><\/pre>\n\n<p>When I print request the shape is:<\/p>\n\n<pre><code>model_spec {\n  name: \"half_plus_two\"\n  version {\n    value: 123\n  }\n  signature_name: \"serving_default\"\n}\ninputs {\n  key: \"inputs\"\n  value {\n    dtype: DT_FLOAT\n    tensor_shape {\n      dim {\n        size: -1\n      }\n      dim {\n        size: 1\n      }\n    }\n    float_val: 1.0\n    float_val: 2.0\n    float_val: 5.0\n  }\n}\n<\/code><\/pre>\n\n<p>Get this exception:<\/p>\n\n<pre><code>Exception in thread \"main\" io.grpc.StatusRuntimeException: INVALID_ARGUMENT: input tensor alias not found in signature: inputs. Inputs expected to be in the set {x}.\n    at io.grpc.stub.ClientCalls.toStatusRuntimeException(ClientCalls.java:233)\n    at io.grpc.stub.ClientCalls.getUnchecked(ClientCalls.java:214)\n    at io.grpc.stub.ClientCalls.blockingUnaryCall(ClientCalls.java:139)\n    at tensorflow.serving.PredictionServiceGrpc$PredictionServiceBlockingStub.predict(PredictionServiceGrpc.java:446)\n    at com.avaya.ccml.grpc.GrpcClient.main(GrpcClient.java:72)`\n<\/code><\/pre>\n\n<p>Edit:\nStill working on this.<\/p>\n\n<p>It looks like the tensor proto I'm supplying is not correct.<\/p>\n\n<p>Did an inspect with saved_model_cli and it shows the correct shape:<\/p>\n\n<pre><code>The given SavedModel SignatureDef contains the following input(s):\n  inputs['x'] tensor_info:\n      dtype: DT_FLOAT\n      shape: (-1, 1)\n      name: x:0\nThe given SavedModel SignatureDef contains the following output(s):\n  outputs['y'] tensor_info:\n      dtype: DT_FLOAT\n      shape: (-1, 1)\n      name: y:0\nMethod name is: tensorflow\/serving\/predict\n<\/code><\/pre>\n\n<p>So next need to figure out how to create tensor proto of this structure<\/p>\n\n<p>Current <\/p>\n","answers":[{"AnswerId":"57921492","CreationDate":"2019-09-13T10:02:03.333","ParentId":null,"OwnerUserId":"6265370","Title":null,"Body":"<p>I figured this out.<\/p>\n\n<p>The answer was staring me in the face the whole time.<\/p>\n\n<p>The exception states that input signature must be 'x'<\/p>\n\n<pre><code>Exception in thread \"main\" io.grpc.StatusRuntimeException: INVALID_ARGUMENT: input tensor alias not found in signature: inputs. Inputs expected to be in the set {x}.\n<\/code><\/pre>\n\n<p>And the output of the CLI also looks for 'x' as input name<\/p>\n\n<pre><code>The given SavedModel SignatureDef contains the following input(s):\n  inputs['x'] tensor_info:\n<\/code><\/pre>\n\n<p>So I changed line<\/p>\n\n<pre><code>requestBuilder.putInputs(\"inputs\", proto);\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>requestBuilder.putInputs(\"x\", proto);\n<\/code><\/pre>\n\n<p>Full working code<\/p>\n\n<pre><code>import com.google.protobuf.Int64Value;\nimport io.grpc.ManagedChannel;\nimport io.grpc.ManagedChannelBuilder;\nimport org.tensorflow.framework.DataType;\nimport tensorflow.serving.Predict;\n\npublic class GrpcClient {\n    public static void main(String[] args) {\n        String host = \"localhost\";\n        int port = 8500;\n        \/\/ the model's name.\n        String modelName = \"half_plus_two\";\n        \/\/ model's version\n        long modelVersion = 123;\n\n        \/\/ create a channel\n        ManagedChannel channel = ManagedChannelBuilder.forAddress(host, port).usePlaintext().build();\n        tensorflow.serving.PredictionServiceGrpc.PredictionServiceBlockingStub stub = tensorflow.serving.PredictionServiceGrpc.newBlockingStub(channel);\n\n        \/\/ create PredictRequest\n        Predict.PredictRequest.Builder requestBuilder = Predict.PredictRequest.newBuilder();\n\n        \/\/ create ModelSpec\n        tensorflow.serving.Model.ModelSpec.Builder modelSpecBuilder = tensorflow.serving.Model.ModelSpec.newBuilder();\n        modelSpecBuilder.setName(modelName);\n        modelSpecBuilder.setVersion(Int64Value.of(modelVersion));\n        modelSpecBuilder.setSignatureName(\"serving_default\");\n\n        \/\/ set model for request\n        requestBuilder.setModelSpec(modelSpecBuilder);\n\n        \/\/ create TensorProto with 3 floats\n        org.tensorflow.framework.TensorProto.Builder tensorProtoBuilder = org.tensorflow.framework.TensorProto.newBuilder();\n        tensorProtoBuilder.setDtype(DataType.DT_FLOAT);\n        tensorProtoBuilder.addFloatVal(1.0f);\n        tensorProtoBuilder.addFloatVal(2.0f);\n        tensorProtoBuilder.addFloatVal(5.0f);\n\n        \/\/ create TensorShapeProto\n        org.tensorflow.framework.TensorShapeProto.Builder tensorShapeBuilder = org.tensorflow.framework.TensorShapeProto.newBuilder();\n        tensorShapeBuilder.addDim(org.tensorflow.framework.TensorShapeProto.Dim.newBuilder().setSize(3));\n\n        \/\/ set shape for proto\n        tensorProtoBuilder.setTensorShape(tensorShapeBuilder.build());\n\n        \/\/ build proto\n        org.tensorflow.framework.TensorProto proto = tensorProtoBuilder.build();\n\n        \/\/ set proto for request\n        requestBuilder.putInputs(\"x\", proto);\n\n        \/\/ build request\n        Predict.PredictRequest request = requestBuilder.build();\n        System.out.println(\"Printing request \\n\" + request.toString());\n\n        \/\/ run predict\n        Predict.PredictResponse response = stub.predict(request);\n        System.out.println(response.toString());\n    }\n}\n<\/code><\/pre>\n"},{"AnswerId":"57917473","CreationDate":"2019-09-13T04:54:15.107","ParentId":null,"OwnerUserId":"10528984","Title":null,"Body":"<p>in the example for half_plus_two here they use instances label for input values; <a href=\"https:\/\/www.tensorflow.org\/tfx\/serving\/docker#serving_example\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tfx\/serving\/docker#serving_example<\/a><\/p>\n\n<p>could you try to set it to instances like this?<\/p>\n\n<pre><code>    builder.putInputs(\"instances\", tp);\n<\/code><\/pre>\n\n<p>I also believe that the DType can be problematic. instead of DT_STRING, i think you should use DT_FLOAT as the inspection result shows<\/p>\n\n<pre><code>    tensorProtoBuilder.setDtype(DataType.DT_FLOAT);\n<\/code><\/pre>\n\n<p><strong>Edit<\/strong><\/p>\n\n<p>I am working with Python, couldnt spot the mistake on yours but, this is how we send a predict request (with a <code>PredictRequest<\/code> proto). Maybe you can try out the Predict proto or there is something that I am missing out and you may spot the difference yourself<\/p>\n\n<pre><code>request = predict_pb2.PredictRequest()\nrequest.model_spec.name = model_name\nrequest.model_spec.signature_name = signature_name\nrequest.inputs['x'].dtype = types_pb2.DT_FLOAT\nrequest.inputs['x'].float_val.append(2.0)\n\nchannel = grpc.insecure_channel(model_server_address)\nstub = prediction_service_pb2_grpc.PredictionServiceStub(channel)\nresult = stub.Predict(request, RPC_TIMEOUT)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57893179,"AnswerCount":1,"Tags":"<python-3.x><keras><loss-function>","CreationDate":"2019-09-11T16:16:33.507","AcceptedAnswerId":null,"OwnerUserId":11863622.0,"Title":"How to define sample weights within custom loss function?","Body":"<p>I require to classify binary sequences, where I want to have assign higher weight to later parts of the sequence.<\/p>\n\n<p>For example, I want to assign: <\/p>\n\n<pre><code>sequence      -- weights for the samples in this sequence\n[0,0,0,1,1,1] -- [1,1,1,2,3,4]\n<\/code><\/pre>\n\n<p>Input is in shape m x n, thus m samples of length n. <\/p>\n\n<p>I want to assign weights within a custom loss function, as I want to learn how this works (and by using sample_weight I get all kinds of dimensionality problems).<\/p>\n\n<p>Now I use the following loss function:<\/p>\n\n<pre><code>def weightedLoss(weight):\n    def binaryPart(yTrue,yPred):\n        return K.mean(loss.binary_crossentropy(yTrue,yPred)*weight)  \n    return binaryPart\n<\/code><\/pre>\n\n<p>With weight being a m x n matrix containing the sample weights. The error message I get is: (m=20000,n=63)<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError: Incompatible shapes: [64] vs. [20000,63]\n     [[{{node loss_39\/dense_120_loss\/mul}}]]<\/p>\n<\/blockquote>\n\n<p>I do not understand what this error means. dense_120 is here my output layer, giving as output shape (None,63).<\/p>\n\n<p>I feel like there is an error in the definition of my loss function, what am I missing?<\/p>\n","answers":[{"AnswerId":"57893661","CreationDate":"2019-09-11T16:51:50.813","ParentId":null,"OwnerUserId":"3472735","Title":null,"Body":"<p>As Daniel said, your output layer should contain 64 neurons, but it looks like it currently contains 63. This should fix your error. <\/p>\n\n<p>To answer your question regarding sample weights, the Keras .fit method already takes an argument for sample_weights, so you can use this without creating a custom loss function. <\/p>\n\n<pre><code>fit(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, validation_freq=1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57893415,"AnswerCount":1,"Tags":"<python><pandas><pytorch><torch>","CreationDate":"2019-09-11T16:33:59.893","AcceptedAnswerId":null,"OwnerUserId":1164799.0,"Title":"PyTorch: Dataloader for time series task","Body":"<p>I have a Pandas dataframe with <em>n<\/em> rows and <em>k<\/em> columns loaded into memory. I would like to get batches for a forecasting task where the first training example of a batch should have shape <code>(*q*,*k*)<\/code> with <code>q<\/code> referring to the number of rows from the original dataframe (e.g. 0:128). The next example should be <code>(128:256, *k*)<\/code> and so on. So, ultimately, one batch should have the shape <code>(32,*q*,*k*)<\/code> with 32 corresponding to the batch size. <\/p>\n\n<p>Since <code>TensorDataset<\/code> from <code>data_utils<\/code> does not work here, I am wondering what the best way would be. I tried to use <code>np.array_split()<\/code> to get as first dimension the number of possible splits of <em>q<\/em> values in order to write a custom DataLoader but then reshaping is not guaranteed to work since not all arrays have the same shape.<\/p>\n\n<p>Here is a minimal example to make it more clear. In this case, batch size is 3 and <em>q<\/em> is 2:<\/p>\n\n<pre><code>import pandas as pd\nimport numpy as np\ndf = pd.DataFrame(data=np.arange(0,30).reshape(10,3),columns=['A','B','C'])\n<\/code><\/pre>\n\n<p>The dataset:<\/p>\n\n<pre><code>    A   B   C\n0   0   1   2\n1   3   4   5\n2   6   7   8\n3   9   10  11\n4   12  13  14\n5   15  16  17\n6   18  19  20\n7   21  22  23\n8   24  25  26\n9   27  28  29\n<\/code><\/pre>\n\n<p>The first batch in this case should have the shape (3,2,3) and look like:<\/p>\n\n<pre><code>array([[[ 0.,  1.,  2.],\n        [ 3.,  4.,  5.]],\n\n       [[ 3.,  4.,  5.],\n        [ 6.,  7.,  8.]],\n\n       [[ 6.,  7.,  8.],\n        [ 9., 10., 11.]]])\n<\/code><\/pre>\n","answers":[{"AnswerId":"57893665","CreationDate":"2019-09-11T16:52:02.367","ParentId":null,"OwnerUserId":"12054165","Title":null,"Body":"<p>You can write your analog of the TensorDataset. To do this you need to inherit from the Dataset class.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from torch.utils.data import Dataset, DataLoader\n\nclass MyDataset(Dataset):\n    def __init__(self, data_frame, q):\n        self.data = data_frame.values\n        self.q = q\n\n    def __len__(self):\n        return self.data.shape[0] \/\/ self.q\n\n    def __getitem__(self, index):\n        return self.data[index * self.q: (index+1) * self.q]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57893542,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-11T16:43:20.827","AcceptedAnswerId":57894509.0,"OwnerUserId":10496951.0,"Title":"Convolution2D \"Depth of input is not a multiple of input depth of filter\"","Body":"<p>I want to use fft2d transform in tensorflow and then analyse the magnitude and phase result with convolutional layers... I have made a system with Lambda layers to use tensorflow functions and get the magnitude and the phase. But when I add the Conv2d layer, I get the error <\/p>\n\n<p><code>Depth of input (3) is not a multiple of input depth of filter (199) for '1_Magn_Conv_f500_k2_2\/convolution' (op: 'Conv2D') with input shapes: [?,199,199,3], [2,2,199,500]<\/code> <\/p>\n\n<p>I don't understand what is the <code>shape [2,2,199,500]<\/code> and what is causing this error.<\/p>\n\n<p>I have tried to reduce the number of layers in my networks to detect which one creates the problem. I have checked that magn_angle outputs two tensors with <code>shapes [None,199,199,3]<\/code>.\nI am working with google colab.<\/p>\n\n<p>Here is the minimal code to reproduce the error<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>inpt = Input(shape = (199, 199, 3),name=str(0)+'_'+'Image')\ndef magn_angle(x):\n    x = Lambda(lambda x:K.cast(x,dtype=tf.complex64))(x)\n    x_list_magn = []\n    x_list_angle = []\n    for i in range(3):\n        fft = Lambda(lambda x: tf.fft2d(x[:,:,:,i]), output_shape=(None,199,199))(x)# 2-dimensional discrete Fourier transform over the inner-most 2 dimensions\n        x_list_magn.append(Lambda(lambda fft:K.expand_dims(tf.math.abs(fft),axis=-1), output_shape=(None,199,199))(fft))\n        x_list_angle.append(Lambda(lambda fft: K.expand_dims(tf.math.angle(fft),axis=-1), output_shape=(None,199,199))(fft))\n    magn = Concatenate()(x_list_magn)\n    angle = Concatenate()(x_list_angle)\n    magn = Lambda(lambda magn: K.cast(magn,dtype=tf.float32), output_shape=(None,199,199))(magn)\n    angle = Lambda(lambda angle: K.cast(angle,dtype=tf.float32), output_shape=(None,199,199))(angle)\n    return magn,angle\nmagn, angle = magn_angle(inpt)\nmagn = Conv2D(filters=500,kernel_size=(2,2),activation=None,strides=(1,1),padding='SAME',name=str(1)+'_'+'Magn_Conv_f500_k2',data_format=\"channels_last\")(magn)\n...\n<\/code><\/pre>\n\n<p>Which ouputs<\/p>\n\n<p><code>InvalidArgumentError: Depth of input (3) is not a multiple of input depth of filter (199) for '1_Magn_Conv_f500_k2_3\/convolution' (op: 'Conv2D') with input shapes: [?,199,199,3], [2,2,199,500]<\/code>.<\/p>\n","answers":[{"AnswerId":"57894509","CreationDate":"2019-09-11T18:01:02.087","ParentId":null,"OwnerUserId":"10496951","Title":null,"Body":"<blockquote>\n  <p>I ran your code without errors in a colab notebook using tf.keras - could be a version mismatch \u2013 Colin Torney<\/p>\n<\/blockquote>\n\n<p>Changing <code>keras. ...<\/code> imports to <code>tensorflow.keras. ...<\/code> has solved the problem.<\/p>\n"}]}
+{"QuestionId":57894056,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><deep-learning>","CreationDate":"2019-09-11T17:23:41.567","AcceptedAnswerId":null,"OwnerUserId":9517778.0,"Title":"You must feed a value for placeholder tensor while using tf.get_collection(tf.GraphKeys.UPDATE_OPS)","Body":"<p>I am defining an architecture which consists of Batch Normalization implemented in tensorflow. <\/p>\n\n<p>I am not getting into the entire architecture just showing the relevant part of the issue.<\/p>\n\n<pre><code>self.w3 = tf.layers.Dense(self.hidden_dimension)\n\n\/\/ code\n\nh = tf.layers.dropout(tf.layers.batch_normalization(tf.nn.sigmoid(self.w3(x)), \ntraining = self.is_training, renorm = True), rate = 0.15, training = self.is_training)\n\n\/\/ code\n<\/code><\/pre>\n\n<p>The official tensorflow implementation for Batch Normalization says 'update ops are placed in tf.GraphKeys.UPDATE_OPS, so they need to be executed alongside the train_op.'<\/p>\n\n<p>Now, when I DON'T use <\/p>\n\n<pre><code>update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\nself.optimize = tf.group([self.opt, update_ops])\n<\/code><\/pre>\n\n<p>The training continues normally WITH NO ERROR. However, I know batch normalization is not being performed here.<\/p>\n\n<p>But as soon as these two lines are being introduced it gives an unexpected error <\/p>\n\n<pre><code>You must feed a value for placeholder tensor\n<\/code><\/pre>\n\n<p>Interestingly, each time I try to execute this function above error keeps changing for different placeholders and does not fix to a particular placeholder. However, I am sure that I have feeding values to all the placeholders (As I already mentioned, training works fine when those two lines are not there).<\/p>\n\n<p>Just in case you want to see the loss and optimize function<\/p>\n\n<pre><code>with tf.variable_scope(name_or_scope = 'op', reuse = tf.AUTO_REUSE) as scope:\n        self.cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits = self.p_y,\n                                                                            labels = self.ans))\n\n        self.optimizer = tf.train.AdadeltaOptimizer(learning_rate=self.learning_rate)\n\n        gradients, variables = zip(*self.optimizer.compute_gradients(self.cost))\n        gradients, _ = tf.clip_by_global_norm(gradients, self.clip_rate)\n        self.opt = self.optimizer.apply_gradients(zip(gradients, variables))\n\n        update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\n        self.optimize = tf.group([self.opt, update_ops])\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57894221,"AnswerCount":1,"Tags":"<tensorflow><anaconda><gpu>","CreationDate":"2019-09-11T17:36:42.767","AcceptedAnswerId":null,"OwnerUserId":6911037.0,"Title":"I saw 2 GPU indicators, Could it be the cause that I keep failing to install tensorflow-GPU?","Body":"<p>I'd tried dozen times to install tensorflow-GPU, but all failed, thougth I followed the instructions indicated by those who successfully installed.\nI just noticed that I have 2 GPU indicators, meaning I have 2 GUP set? One Intel and one Nvidia. <\/p>\n\n<p>Could it be the reason? If so, how can I kick in the proper one (Nvidia)?\nThanks for advice.\n<a href=\"https:\/\/i.stack.imgur.com\/JFNZH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/JFNZH.png\" alt=\"2\"><\/a><\/p>\n","answers":[{"AnswerId":"58538727","CreationDate":"2019-10-24T09:51:41.437","ParentId":null,"OwnerUserId":"12198627","Title":null,"Body":"<p>Verify proper installation of tensorflow-GPU by using following steps,<\/p>\n\n<ul>\n<li><p>Can you check if tensorflow is using GPU,<\/p>\n\n<p><code>import tensorflow as tf\nprint(tf.test.is_gpu_available())<\/code><\/p><\/li>\n<li><p>To find out which device is used, you can enable log device placement\nlike this:<\/p>\n\n<p><code>sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))<\/code><\/p><\/li>\n<li><p>you can try to assign a computation to the gpu and see whether you have an error<\/p>\n\n<p><code>import tensorflow as tf\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')\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)\nwith tf.Session() as sess:\n    print (sess.run(c))<\/code><\/p><\/li>\n<li><p>Check the GPU usage using\n<code>nvidia-smi<\/code><\/p><\/li>\n<\/ul>\n"}]}
+{"QuestionId":57894349,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-11T17:47:29.840","AcceptedAnswerId":null,"OwnerUserId":11159873.0,"Title":"Combine output of multiple ImageDataGenerators to one tensor","Body":"<p>Let's say I want to apply two different <code>preprocessing_functions<\/code> in two<code>ImageDataGenerator<\/code>s to an Image and then concatenate those outputs to a one tensor to train the model.<\/p>\n\n<p>i.e., If I have RGB images, my model would expect six channels in the input tensor because I want to concatenate the output of two generators (each will have 3 channels).<\/p>\n\n<p>Concretely:<\/p>\n\n<pre><code># first preproc function\ndef preproc_function_1(image):\n    image = np.array(image)\n    return image * 0.1\n\n# second preproc function\ndef preproc_function_2(image):\n    image = np.array(image)\n    return image * 0.2\n\n# datagen_1 - that applies preproc_function_1 \ntrain_datagen_1 = ImageDataGenerator(\n     preprocessing_function=preproc_function_1)\n\n# datagen_2 - that applies preproc_function_2\ntrain_datagen_2 = ImageDataGenerator(\n     preprocessing_function=preproc_function_2)\n<\/code><\/pre>\n\n<p>If we have such parallel <code>ImageDataGenerator<\/code>s I assume, we want to use two validation generators as well<\/p>\n\n<pre><code> validation_datagen_1 = ImageDataGenerator()\n\n validation_datagen_2 = ImageDataGenerator()\n<\/code><\/pre>\n\n<p><em>Question:<\/em><\/p>\n\n<p>How can I proceed to use <code>model.fit_generator()<\/code>?<\/p>\n\n<p>I think I have to use <code>&lt;generator&gt;.flow_from_directory()<\/code> (my data can be read from <code>flow_from_directory<\/code>) to supply <code>generator<\/code> in   <a href=\"https:\/\/keras.io\/models\/model\/#fit_generator\" rel=\"nofollow noreferrer\"><code>model.fit_generator()<\/code><\/a> to train the model.<\/p>\n\n<p>Do I need to create two <code>&lt;generator&gt;.flow_from_directory()<\/code> objects? I'm lost here.<\/p>\n\n<p>Any help is much appreciated<\/p>\n","answers":[{"AnswerId":"57894634","CreationDate":"2019-09-11T18:09:54.330","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>It's easier if you just create a single preprocessing function:<\/p>\n\n<pre><code>def preproc(image):\n     return np.concatenate([\n                            preproc_function_1(image), \n                            preproc_function_2(image)\n                           ], axis=-1) #or axis=1 if channels_first\n<\/code><\/pre>\n\n<h2>Option 2 - Using two generators<\/h2>\n\n<p>Since you have the 6 channel problem, this custom generator should work. But you will not have 'preprocessing', you will have 'postprocessing' instead:<\/p>\n\n<pre><code>class Processor(keras.utils.Sequence):\n    def __init__(self, keras_generator)\n        self.keras_generator = keras_generator\n\n    def __len__(self):\n        return len(self.keras_generator(\n\n    def __getitem__(self, i):\n        batch_x, batch_y = self.keras_generator[i]\n\n        #do batch processing here\n\n        return processed_batch_x, batch_y\n\n    def on_epoch_end(self):\n        self.keras_generator.on_epoch_end()\n\ncustom_generator = Processor(validation_datagen.flow_from_directory(...))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57894439,"AnswerCount":3,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-11T17:55:00.487","AcceptedAnswerId":57894589.0,"OwnerUserId":2584754.0,"Title":"Keras Conv3D Layer with Discrete Values","Body":"<p>I'm trying to build a model that will learn features of a 3D space. Unlike image processing, the values of the 3D matrix are not continuous; they represent some discrete value of what \"material\" can be found at that specific coordinate (grass with value 1 or stairs with value 2 for example).<\/p>\n\n<p>Is it possible to train a model to learn the features of the space without interpolating in-between values? For example, I don't want the neural net to deduce 1.5 to be some kind of grass stairs. <\/p>\n","answers":[{"AnswerId":"57894703","CreationDate":"2019-09-11T18:14:09.153","ParentId":null,"OwnerUserId":"10376805","Title":null,"Body":"<p>Although I am not sure if this is what you are asking for, I would imagine that t after the bottleneck of the convolutional network, one would typically use a flatten layer and then the output goes to a dense layer. The output layer, if using sigmoid activation will give you probabilities for each of the classes which have to be one-hot encoded, as others have suggested.<\/p>\n\n<p>If you want the output of the network itself to be in discreet values, I suppose you can use some sort of step-wise activation function in the output layer. However you have to take care that your loss remains differentiable throughout the network (which is why such activation functions are not available in keras). This might be of interest: <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/7370\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/7370<\/a><\/p>\n"},{"AnswerId":"57894592","CreationDate":"2019-09-11T18:06:39.203","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>As any categorical model, this should be a \"one-hot\" data.    <\/p>\n\n<p>The \"channels\" dimension of your data should have a size of <code>n-materials<\/code>.   <\/p>\n\n<ul>\n<li>Values = 0 mean there is no presence of that material    <\/li>\n<li>Values = 1 mean there is presence of that material    <\/li>\n<\/ul>\n\n<p>So, your input shape will be something like <code>(samples, spatial1, spatial2, spatial3, materials)<\/code>. If your data is currently shaped as <code>(samples, s1, s2, s3)<\/code> and has the materias as integers as you described, you can use <code>to_categorical<\/code> to transform the integers to \"one-hot\".   <\/p>\n"},{"AnswerId":"57894589","CreationDate":"2019-09-11T18:06:28.060","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>You'll want to use one-hot encoding, which represents categorical values as arrays of zeroes with a single value set to one. This means that grass (id = 1) would be <code>[0, 1, 0, 0, ...]<\/code> and stairs (id = 2) would be <code>[0, 0, 1, 0, ...]<\/code>. To perform one-hot encoding, look into keras' <code>to_categorical<\/code> function.<\/p>\n\n<p>Further reading:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/machinelearningmastery.com\/how-to-one-hot-encode-sequence-data-in-python\/\" rel=\"nofollow noreferrer\">one-hot encoding tutorial<\/a><\/li>\n<li><a href=\"https:\/\/keras.io\/utils\/#to_categorical\" rel=\"nofollow noreferrer\">one-hot preprocessing using <code>to_categorical<\/code><\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/a\/50446946\/2846923\">one-hot on the fly using an embedding layer<\/a><\/li>\n<\/ul>\n"}]}
+{"QuestionId":57894819,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><tensorflow-datasets>","CreationDate":"2019-09-11T18:23:17.677","AcceptedAnswerId":null,"OwnerUserId":265683.0,"Title":"TensorFlow sound recognition","Body":"<p>I am making my own version of the example for <a href=\"https:\/\/data-flair.training\/blogs\/tensorflow-audio-recognition\/\" rel=\"nofollow noreferrer\">TensorFlow audio recognition<\/a> to recognise some sound effects instead of speech.\nI am getting the following error during training of my model for sound recognition:<\/p>\n\n<blockquote>\n  <p>2019-09-11 19:16:38.221677: E\n  tensorflow\/core\/kernels\/mfcc_mel_filterbank.cc:153] Missing 5 bands \n  starting at 0 in mel-frequency design. Perhaps too many channels or\n  not enough frequency resolution in spectrum. (input_length: 257\n  input_sample_rate: 44100 output_channel_count: 40\n  lower_frequency_limit: 20 upper_frequency_limit: 4000<\/p>\n<\/blockquote>\n\n<p>Can you please explain what this means, and how I can fix the problem? My audio clips are all around 1 second long, 44.1khz and are in stereo.<\/p>\n\n<p>Many thanks!<\/p>\n","answers":[{"AnswerId":"57894908","CreationDate":"2019-09-11T18:29:00.450","ParentId":null,"OwnerUserId":"265683","Title":null,"Body":"<p>The problem was the example expects a sample rate of audio files of 16000khz however I was supplying 44100khz files.<\/p>\n\n<p>I fixed the problem by adding the following flag:<\/p>\n\n<pre><code>--sample_rate=44100\n<\/code><\/pre>\n"}]}
+{"QuestionId":57894869,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras><neural-network><tensorflow-datasets>","CreationDate":"2019-09-11T18:26:35.950","AcceptedAnswerId":57964173.0,"OwnerUserId":1166332.0,"Title":"tf.dataset, multiple path inputs, and mapping per batch to load images","Body":"<p>I'm loading a dataset with multiple input images. The input image paths should only be decoded at batch time, in order to handle a large dataset.<\/p>\n\n<p>The data set is <strong>N<\/strong> image path inputs and <strong>M<\/strong> float outputs.  The images for each input have different resolutions.<\/p>\n\n<pre><code>Data is ([img_input_1.png, img_input_2.png, ...], [0.65, 0.7, 0.8])\n<\/code><\/pre>\n\n<p>The model is using the Keras functional api in symbolic mode.<\/p>\n\n<p>Here is the most recently EDITED code<\/p>\n\n<pre><code>from itertools import zip_longest\n\ndef read_image(path, shape):\n    try:\n        image = tf.io.read_file(path)\n        image = tf.image.decode_png(image)\n        image = tf.image.resize(image, [shape[1],shape[2]])\n        image \/= 255.0\n        return image\n    except:\n        print('ERROR: preprocess_image: bad path', path)    \n\ndef load_image(x, y, shp):\n    pout = [(k, x[k]) for k in x.keys()]\n    l1   = tf.convert_to_tensor(list(x))\n    l2   = tf.convert_to_tensor(list(x.values()))\n\n    pl = tf.map_fn(\n        lambda args: (read_image(args[0], shp), args[1]), [l1, l2], dtype=(tf.float32, tf.float32)\n    )\n    pl = {path: (pl[0][i], pl[1][i]) for i, path in enumerate(x)}\n    return (pl,y)\n\ndef dataset_prep(json_data, seq, batch_size):\n    # LOAD DATA FROM JSON\n    x,y = json_parse_x_y(json_data[seq])\n    xx  = [*zip_longest(*x)] # NOTE: goes from variable sized input to {'input_N':...}\n    yy  = [*zip_longest(*y)]\n\n    # GET SHAPES (hard coded atm)\n    lns = [[len(xxx)] for xxx in xx]\n    rzs = [[24,512,1],[96,512,1]] # TEMP TODO! grab grom [(v['h'],v['w'],v['c']) for v in xx]\n    shp = [*zip_longest(*[lns,rzs])]\n    shp = [list(s) for s in shp]\n    shp = [[*itertools.chain.from_iterable(s)] for s in shp]\n\n    xd  = dict([[ \"input_{}\".format(i+1),np.array(y)] for i,y in [*enumerate(xx)]])\n    yd  = dict([[\"output_{}\".format(i+1),np.array(y)] for i,y in [*enumerate(yy)]])\n\n    ds  = tf.data.Dataset.from_tensor_slices((xd, yd))\n\n    ds  = ds.shuffle(10000)\n    ds  = ds.repeat()\n    ds  = ds.map(map_func=lambda x,y: load_image(x, y, shp), num_parallel_calls=AUTOTUNE)\n    ds  = ds.batch(batch_size) if batch_size else ds \n    ds  = ds.prefetch(AUTOTUNE) \n    return ds\n<\/code><\/pre>\n\n<p>This is the error I'm getting:<\/p>\n\n<pre><code>Traceback (most recent call last):\n    File \"\/home\/me\/.local\/bin\/wavfeat\", line 11, in &lt;module&gt;\n        load_entry_point('wavfeat==0.1.0', 'console_scripts', 'wavfeat')()\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/wavfeat\/__main__.py\", line 91, in main\n        analysis_batch_sql(obj)\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/wavfeat\/analysis_run_csv.py\", line 50, in analysis_batch_sql\n        qy = [*map(lambda c: run_elm(c[0], c[1]), ch)]\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/wavfeat\/analysis_run_csv.py\", line 50, in &lt;lambda&gt;\n        qy = [*map(lambda c: run_elm(c[0], c[1]), ch)]\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/wavfeat\/analysis_run_csv.py\", line 23, in run_elm\n        out = fn(input, elm)\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/wavfeat\/one_sec_onset.py\", line 196, in one_sec_onset_train\n        return train(input, elm)\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/wavfeat\/one_sec_onset.py\", line 182, in train\n        ts = dataset_prep(jd, 'train', bc)\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/wavfeat\/one_sec_onset.py\", line 123, in dataset_prep\n        ds  = ds.map(map_func=lambda x,y: load_image(x, y, shp), num_parallel_calls=AUTOTUNE)\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/data\/ops\/dataset_ops.py\", line 1146, in map\n        self, map_func, num_parallel_calls, preserve_cardinality=True)\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/data\/ops\/dataset_ops.py\", line 3264, in __init__\n        use_legacy_function=use_legacy_function)\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/data\/ops\/dataset_ops.py\", line 2591, in __init__\n        self._function = wrapper_fn._get_concrete_function_internal()\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/eager\/function.py\", line 1366, in _get_concrete_function_internal\n        *args, **kwargs)\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/eager\/function.py\", line 1360, in _get_concrete_function_internal_garbage_collected\n        graph_function, _, _ = self._maybe_define_function(args, kwargs)\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/eager\/function.py\", line 1648, in _maybe_define_function\n        graph_function = self._create_graph_function(args, kwargs)\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/eager\/function.py\", line 1541, in _create_graph_function\n        capture_by_value=self._capture_by_value),\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/func_graph.py\", line 716, in func_graph_from_py_func\n        func_outputs = python_func(*func_args, **func_kwargs)\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/data\/ops\/dataset_ops.py\", line 2585, in wrapper_fn\n        ret = _wrapper_helper(*args)\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/data\/ops\/dataset_ops.py\", line 2530, in _wrapper_helper\n        ret = func(*nested_args)\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/wavfeat\/one_sec_onset.py\", line 123, in &lt;lambda&gt;\n        ds  = ds.map(map_func=lambda x,y: load_image(x, y, shp), num_parallel_calls=AUTOTUNE)\n    File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/wavfeat\/one_sec_data_loader.py\", line 91, in load_image\n        print(\"x['input_1'].values(): \", x['input_1'].values())\nAttributeError: 'Tensor' object has no attribute 'values'\n<\/code><\/pre>\n\n<p>What am I doing that is preventing the paths from being loaded?<\/p>\n\n<p>EDIT:<\/p>\n\n<p>Attempting pandrey's fix, I'm getting input errors. Here is the data before from_tensor_slices and ds.map and then after:<\/p>\n\n<pre><code>pre_from_tensor_slices x:  {'input_1': array(['\/media\/me\/sp_data\/sp_data\/datasets\/chr_01\/one_sec_onset_11_oac-leg\/7388_39216_30--id=7388__sql_table=oac_1__sql_idx=405167__pitch=30__onset=39216.png',\n       '\/media\/me\/sp_data\/sp_data\/datasets\/chr_01\/one_sec_onset_11_oac-leg\/2447_864_27--id=2447__sql_table=oac_1__sql_idx=415458__pitch=27__onset=864.png',\n       '\/media\/me\/sp_data\/sp_data\/datasets\/chr_01\/one_sec_onset_11_oac-leg\/2386_20208_38--id=2386__sql_table=oac_1__sql_idx=433248__pitch=38__onset=20208.png',\n       ...,\n       '\/media\/me\/sp_data\/sp_data\/datasets\/chr_01\/one_sec_onset_11_oac-leg\/6261_24528_57--id=6261__sql_table=oac_1__sql_idx=449753__pitch=57__onset=24528.png',\n       '\/media\/me\/sp_data\/sp_data\/datasets\/chr_01\/one_sec_onset_11_oac-leg\/3727_22944_31--id=3727__sql_table=oac_1__sql_idx=407620__pitch=31__onset=22944.png',\n       '\/media\/me\/sp_data\/sp_data\/datasets\/chr_01\/one_sec_onset_11_oac-leg\/1668_7056_60--id=1668__sql_table=oac_1__sql_idx=381152__pitch=60__onset=7056.png'],\n      dtype='&lt;U162'), 'input_2': array(['\/media\/me\/sp_data\/sp_data\/datasets\/mel_01\/one_sec_onset_11_oac-leg\/7388_39216_30--id=7388__sql_table=oac_1__sql_idx=405167__pitch=30__onset=39216.png',\n       '\/media\/me\/sp_data\/sp_data\/datasets\/mel_01\/one_sec_onset_11_oac-leg\/2447_864_27--id=2447__sql_table=oac_1__sql_idx=415458__pitch=27__onset=864.png',\n       '\/media\/me\/sp_data\/sp_data\/datasets\/mel_01\/one_sec_onset_11_oac-leg\/2386_20208_38--id=2386__sql_table=oac_1__sql_idx=433248__pitch=38__onset=20208.png',\n       ...,\n       '\/media\/me\/sp_data\/sp_data\/datasets\/mel_01\/one_sec_onset_11_oac-leg\/6261_24528_57--id=6261__sql_table=oac_1__sql_idx=449753__pitch=57__onset=24528.png',\n       '\/media\/me\/sp_data\/sp_data\/datasets\/mel_01\/one_sec_onset_11_oac-leg\/3727_22944_31--id=3727__sql_table=oac_1__sql_idx=407620__pitch=31__onset=22944.png',\n       '\/media\/me\/sp_data\/sp_data\/datasets\/mel_01\/one_sec_onset_11_oac-leg\/1668_7056_60--id=1668__sql_table=oac_1__sql_idx=381152__pitch=60__onset=7056.png'],\n      dtype='&lt;U162')}\npre_from_tensor_slices y:  {'output_1': array([0.817, 0.018, 0.421, ..., 0.511, 0.478, 0.147])}\n_________________________\ny:  {'output_1': &lt;tf.Tensor 'args_2:0' shape=() dtype=float64&gt;}\nx:  {'input_1': &lt;tf.Tensor 'args_0:0' shape=() dtype=string&gt;, 'input_2': &lt;tf.Tensor 'args_1:0' shape=() dtype=string&gt;}\nx.values():  dict_values([&lt;tf.Tensor 'args_0:0' shape=() dtype=string&gt;, &lt;tf.Tensor 'args_1:0' shape=() dtype=string&gt;])\nx['input_1']:  Tensor(\"args_0:0\", shape=(), dtype=string)\n\n<\/code><\/pre>\n\n<p>Running x['input_1'].values() throws an error: 'Tensor' object has no attribute 'values'<\/p>\n\n<p>I get an error situated around map_fn <\/p>\n\n<pre><code>  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/constant_op.py\", line 284, in _constant_impl\n    allow_broadcast=allow_broadcast))\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/tensor_util.py\", line 455, in make_tensor_proto\n    raise ValueError(\"None values not supported.\")\nValueError: None values not supported.\n\n<\/code><\/pre>\n\n<p>EDIT 2<\/p>\n\n<p>Attempting the latest I get the following error<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/home\/me\/.local\/bin\/wavfeat\", line 11, in &lt;module&gt;\n    load_entry_point('wavfeat==0.1.0', 'console_scripts', 'wavfeat')()\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/wavfeat\/__main__.py\", line 91, in main\n    analysis_batch_sql(obj)\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/wavfeat\/analysis_run_csv.py\", line 50, in analysis_batch_sql\n    qy = [*map(lambda c: run_elm(c[0], c[1]), ch)]\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/wavfeat\/analysis_run_csv.py\", line 50, in &lt;lambda&gt;\n    qy = [*map(lambda c: run_elm(c[0], c[1]), ch)]\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/wavfeat\/analysis_run_csv.py\", line 23, in run_elm\n    out = fn(input, elm)\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/wavfeat\/one_sec_onset.py\", line 216, in one_sec_onset_train\n    return train(input, elm)\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/wavfeat\/one_sec_onset.py\", line 203, in train\n    vs = validation_prep(jd, 'validation', bc)\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/wavfeat\/one_sec_onset.py\", line 176, in validation_prep\n    ds  = ds.map(map_func=load_and_preprocess_from_path_label, num_parallel_calls=AUTOTUNE)\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/data\/ops\/dataset_ops.py\", line 1146, in map\n    self, map_func, num_parallel_calls, preserve_cardinality=True)\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/data\/ops\/dataset_ops.py\", line 3264, in __init__\n    use_legacy_function=use_legacy_function)\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/data\/ops\/dataset_ops.py\", line 2591, in __init__\n    self._function = wrapper_fn._get_concrete_function_internal()\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/eager\/function.py\", line 1366, in _get_concrete_function_internal\n    *args, **kwargs)\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/eager\/function.py\", line 1360, in _get_concrete_function_internal_garbage_collected\n    graph_function, _, _ = self._maybe_define_function(args, kwargs)\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/eager\/function.py\", line 1648, in _maybe_define_function\n    graph_function = self._create_graph_function(args, kwargs)\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/eager\/function.py\", line 1541, in _create_graph_function\n    capture_by_value=self._capture_by_value),\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/func_graph.py\", line 716, in func_graph_from_py_func\n    func_outputs = python_func(*func_args, **func_kwargs)\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/data\/ops\/dataset_ops.py\", line 2585, in wrapper_fn\n    ret = _wrapper_helper(*args)\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/data\/ops\/dataset_ops.py\", line 2530, in _wrapper_helper\n    ret = func(*nested_args)\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/wavfeat\/one_sec_data_loader.py\", line 47, in load_and_preprocess_from_path_label\n    pl   = dict([(pk, tf.map_fn(load_and_preprocess_image, po, dtype=tf.float32)) for pk,po in pout])\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/wavfeat\/one_sec_data_loader.py\", line 47, in &lt;listcomp&gt;\n    pl   = dict([(pk, tf.map_fn(load_and_preprocess_image, po, dtype=tf.float32)) for pk,po in pout])\n  File \"\/home\/me\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/map_fn.py\", line 214, in map_fn\n    raise ValueError(\"elems must be a 1+ dimensional Tensor, not a scalar\")\nValueError: elems must be a 1+ dimensional Tensor, not a scalar\n<\/code><\/pre>\n","answers":[{"AnswerId":"57964173","CreationDate":"2019-09-16T20:44:15.807","ParentId":null,"OwnerUserId":"12076271","Title":null,"Body":"<h3>Add-on: not using dict structures<\/h3>\n\n<p>This is a full code (save for defining <code>json_parse_x_y<\/code> and declaring <code>AUTOTUNE<\/code>) to achieve what you are attempting without using dict structures.<\/p>\n\n<p>I tested that <code>make_dataset<\/code> works (see example below), so if you encounter an issue it should be due to a specification error regarding <code>load_tensors<\/code>.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from itertools import zip_longest\n\nimport tensorflow as tf\n\n# additionnally, `json_parse_x_y` must be defined\n# and `AUTOTUNE` must be declared (in my example, I set it to 2)\n\n\ndef read_image(path, shape):\n    \"\"\"Read an image of givent filepath and tensor shape.\n\n    Return a float tensor of given shape.\n    \"\"\"\n    try:\n        image = tf.io.read_file(path)\n        image = tf.image.decode_png(image)\n        image = tf.image.resize(image, [shape[1], shape[2]])\n        image \/= 255.0\n        return image\n    except:\n        raise FileNotFoundError(\"preprocess_image: bad path '%s'\" % path)\n\n\ndef load_images(paths, shapes):\n    \"\"\"Load an ensemble of images (associated with a single sample).\n\n    paths  : rank-1 string Tensor\n    shapes : list of images' shapes (same length as `paths`)\n\n    Return a tuple of float tensors containing the loaded images.\n    \"\"\"\n    return tuple((\n        read_image(paths[i], shapes[i])\n        for i in range(len(shapes))\n    ))\n\n\ndef load_tensors(json_data, seq):\n    \"\"\"Load images descriptors from a json dump.\n\n    Return a tuple containing:\n        * a rank-2 tensor containing lists of image paths (str)\n        * a rank-2 tensor containing resolution values (float)\n        * a list of image shapes, of same length as the rank-2\n          tensor's second axis\n    \"\"\"\n    x,y = json_parse_x_y(json_data[seq])\n    xx  = [*zip_longest(*x)] # NOTE: goes from variable sized input to {'input_N':...}\n    yy  = [*zip_longest(*y)]\n\n    # GET SHAPES (hard coded atm)\n    lns = [[len(xxx)] for xxx in xx]\n    rzs = [[24,512,1],[96,512,1]] # TEMP TODO! grab grom [(v['h'],v['w'],v['c']) for v in xx]\n    shp = [*zip_longest(*[lns,rzs])]\n    shp = [list(s) for s in shp]\n    shp = [[*itertools.chain.from_iterable(s)] for s in shp]\n    return (xx, yy, shp)\n\n\ndef make_dataset(xx, yy, shp, batch_size):\n    \"\"\"Build a Dataset instance containing loaded images.\n\n    xx, yy, shp : see the specification of `load_tensors`'s outputs\n    batch_size  : batch size to set on the Dataset\n\n    Return a Dataset instance where each batched sample is a tuple\n    containing two elements: first, a tuple containing N loaded images'\n    rank-3 tensors; second, a rank-1 tensor containing M float values.\n    (to be clear: batching adds a dimension to all those tensors)\n    \"\"\"\n    data = tf.data.Dataset.from_tensor_slices((xx, yy))\n    data = data.shuffle(10000)\n    data = data.map(lambda x, y: (load_images(x, shapes), y))\n    data = data.repeat()\n    data = data.batch(batch_size) if batch_size else data\n    data = data.prefetch(AUTOTUNE) \n    return data\n\n\ndef dataset_prep(json_data, seq, batch_size):\n    \"\"\"Full pipeline to making a Dataset from json.\"\"\"\n    xx, yy, shapes = load_tensors(json_data, seq)\n    return make_dataset(xx, yy, shapes)\n<\/code><\/pre>\n\n<p>Example, using \"hand-made' values ; all images are actually\n<a href=\"https:\/\/upload.wikimedia.org\/wikipedia\/en\/7\/7d\/Lenna_%28test_image%29.png\" rel=\"nofollow noreferrer\">this classic image<\/a>, of shape [512, 512, 3].<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nimport numpy as np\nimport tensorflow as tf\n\n# import previous code\n\n# Here, N = 2, and I make 2 samples.\nx = tf.convert_to_tensor(np.array([\n    ['image_1a.png', 'image_1b.png'],\n    ['image_2a.png', 'image_2b.png']\n]))\nshapes = [[1, 512, 512], [1, 512, 512]]  # images are initially [512, 512, 3]\n# Here, M = 3, and I make 2 samples. Values are purely random.\ny = tf.convert_to_tensor(np.array([\n    [.087, .92, .276],\n    [.242, .37, .205]\n]))\n\n# This should work.\ndata = make_dataset(x, y, shapes, batch_size=1)\n# Output signature is &lt;PrefetchDataset shapes:\n#     (((None, 512, 512, None), (None, 512, 512, None)), (None, 3)),\n#     types: ((tf.float32, tf.float32), tf.float64)\n# &gt;\n# Where the first None is actually `batch_size`\n# and the second is, in this case, 3.\n<\/code><\/pre>\n\n<hr>\n\n<h3>Answer to the current question:<\/h3>\n\n<p>Okay, the problem you are now encountering is that the revised <code>load_image<\/code> function does not fit the specifications of the Dataset, hence the exception raising. Please find below a full edited code that seems to work (I ran a test using custom images on my computer, with <code>xd<\/code> \/ <code>yd<\/code> dict initialized to look like your reported x and y in-dataset tensors). It is not pretty, and I would personally advise to drop the dict structures, but it works:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from itertools import zip_longest\n\ndef read_image(path, shape):\n    try:\n        image = tf.io.read_file(path)\n        image = tf.image.decode_png(image)\n        image = tf.image.resize(image, [shape[1],shape[2]])\n        image \/= 255.0\n        return image\n    except:\n        raise FileNotFoundError(\"preprocess_image: bad path '%s'\" % path)\n\n# CHANGED: load_image is actually useless\n\ndef dataset_prep(json_data, seq, batch_size):\n    # LOAD DATA FROM JSON\n    x,y = json_parse_x_y(json_data[seq])\n    xx  = [*zip_longest(*x)] # NOTE: goes from variable sized input to {'input_N':...}\n    yy  = [*zip_longest(*y)]\n\n    # GET SHAPES (hard coded atm)\n    lns = [[len(xxx)] for xxx in xx]\n    rzs = [[24,512,1],[96,512,1]] # TEMP TODO! grab grom [(v['h'],v['w'],v['c']) for v in xx]\n    shp = [*zip_longest(*[lns,rzs])]\n    shp = [list(s) for s in shp]\n    shp = [[*itertools.chain.from_iterable(s)] for s in shp]\n\n    xd  = dict([[ \"input_{}\".format(i+1),np.array(y)] for i,y in [*enumerate(xx)]])\n    yd  = dict([[\"output_{}\".format(i+1),np.array(y)] for i,y in [*enumerate(yy)]])\n\n    ds  = tf.data.Dataset.from_tensor_slices((xd, yd))\n    ds  = ds.shuffle(10000)\n\n    # CHANGED: the following line, to run images import (also moved epeat instruction later)\n    ds  = ds.map(\n        lambda x, y: (\n            {key: read_image(path, shp[i]) for i, (key, path) in enumerate(x.items())},\n            y\n        ),\n        num_parallel_calls=AUTOTUNE\n    )\n    ds  = ds.repeat()\n    ds  = ds.batch(batch_size) if batch_size else ds \n    ds  = ds.prefetch(AUTOTUNE) \n    return ds\n<\/code><\/pre>\n\n<hr>\n\n<h3>Initial answer (before question edit):<\/h3>\n\n<p>I will only deal with the exception raised by <code>load_image<\/code> in this answer, but there might be additional work to perform on the rest - I did not test for that, not having a convenient dataset at hand.<\/p>\n\n<p>The exception message is actually quite explicit: you are passing a scalar element (<em>e.g.<\/em> <code>n<\/code> in <code>[(k, tf.map_fn(lambda x: read_image(x, shp), n, dtype=tf.float32)) for k,n in pout]<\/code>) as <code>elems<\/code> argument to <code>tf.map_fn<\/code>, when it expects a tensor (or (possibly nested) list or tuple of tensors), as clearly specified in its documentation.<\/p>\n\n<p>You are also using <code>tf.map_fn<\/code> the wrong way in the quoted line of code, because basically you are mixing it up with a python intention list, when you should use either one or the other.<\/p>\n\n<p>With intention list (also replacing the useless previous lines of the <code>load_image<\/code> function):\n<code>pl = {path: (load_image(path, shp), res) for path, res in x.items()}<\/code><\/p>\n\n<p>With <code>tf.map_fn<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Read all images, return two tensors, one with images, the other with resolutions.\n# (so, resolutions inclusion in this is actually useless and should be redesigned)\npl = tf.map_fn(\n    lambda args: (read_image(args[0], shp), args[1]),\n    [tf.convert_to_tensor(list(x)), tf.convert_to_tensor(list(x.values()))],\n    dtype=(tf.float32, tf.float32)\n)\n# If you really, really want to return a dict, but is it an optimal design?\npl = {path: (pl[0][i], pl[1][i]) for i, path in enumerate(x)}\n<\/code><\/pre>\n\n<p>I do not know whether returning a dict specified in this way is optimal (or even compatible) with Dataset instantiation, however if the rest of your code is working, this should do the trick.<\/p>\n\n<p>At any rate, if you want to iterate over a dict, go ahead and use either the first version or a modified version of the second one (which may have the advantage of parallelizing images reading).<\/p>\n\n<p>I hope this helps :-)<\/p>\n"}]}
+{"QuestionId":57894989,"AnswerCount":1,"Tags":"<javascript><html><tensorflow><tensorflow.js><tensorflowjs-converter>","CreationDate":"2019-09-11T18:35:14.113","AcceptedAnswerId":57896367.0,"OwnerUserId":11335442.0,"Title":"Keep getting: \"Attempted import error: 'loadFrozenModel' is not exported from '@tensorflow\/tfjs-converter'\"","Body":"<p>I'm trying to convert a Tensorflow-for-Poets model to a Tensorflow.js model, so I can use the in a front-end environment, like a website. I was trying to follow this tutorial:\n<a href=\"https:\/\/gist.github.com\/woudsma\/d01eeda8998c9ab972d05ec9e9843886\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/woudsma\/d01eeda8998c9ab972d05ec9e9843886<\/a><\/p>\n\n<p>I've followed all the directions but when I try to launch localhost, I keep getting the titular error:<\/p>\n\n<pre><code>src\/index.js\nAttempted import error: 'loadFrozenModel' is not exported from \n'@tensorflow\/tfjs-converter'\n<\/code><\/pre>\n\n<p>I've trained a Tensorflow model using:<\/p>\n\n<ul>\n<li><p>Tensorflow v. 1.7.0<\/p><\/li>\n<li><p>TensorflowJS v. 1.2.9<\/p><\/li>\n<li><p>Numpy v. 1.16.5<\/p><\/li>\n<\/ul>\n\n<p>I also looked at these previously asked questions:<\/p>\n\n<p><a href=\"http:\/\/www.github.com\/tensorflow\/tfjs\/issues\/149\" rel=\"nofollow noreferrer\">http:\/\/www.github.com\/tensorflow\/tfjs\/issues\/149<\/a><\/p>\n\n<p><a href=\"http:\/\/www.stackoverflow.com\/questions\/49718162\/tfjs-converter-html-javascript-trouble-importing-class\">http:\/\/www.stackoverflow.com\/questions\/49718162\/tfjs-converter-html-javascript-trouble-importing-class<\/a><\/p>\n\n<p>But this hasn't fixed my issue. <\/p>\n\n<p>This is the example-project I've found in the tutorial. It contains the that I also used in my project. \n<a href=\"https:\/\/github.com\/woudsma\/retrain-mobilenet-for-the-web\" rel=\"nofollow noreferrer\">https:\/\/github.com\/woudsma\/retrain-mobilenet-for-the-web<\/a><\/p>\n\n<p>I can't find anything about this specific error, does anyone know what's going wrong?<\/p>\n\n<p>PS:This is also my first question posted to Stack Overflow, so let me know if something is missing from \/ wrong about this post.<\/p>\n\n<p>EDIT: Added my index.js:<\/p>\n\n<pre><code>import { loadFrozenModel } from '@tensorflow\/tfjs-converter'\nimport labels from '.\/labels.json'\n\nconst ASSETS_URL = `${window.location.origin}\/assets`\nconst MODEL_URL = `${ASSETS_URL}\/mobilenet-v2\/tensorflowjs_model.pb`\nconst WEIGHTS_URL = `${ASSETS_URL}\/mobilenet-v2\/weights_manifest.json`\nconst IMAGE_SIZE = 224 \/\/ Model input size\n\nconst loadModel = async () =&gt; {\n  const model = await loadFrozenModel(MODEL_URL, WEIGHTS_URL)\n  const input = tf.zeros([1, IMAGE_SIZE, IMAGE_SIZE, 3])\n   \/\/ Warm up GPU\n  \/\/ model.predict({ input }) \/\/ MobileNet V1\n  model.predict({ Placeholder: input }) \/\/ MobileNet V2\n  return model\n}\n\nconst predict = async (img, model) =&gt; {\n  const t0 = performance.now()\n  const image = tf.fromPixels(img).toFloat()\n  const resized = tf.image.resizeBilinear(image, [IMAGE_SIZE, IMAGE_SIZE])\n  const offset = tf.scalar(255 \/ 2)\n  const normalized = resized.sub(offset).div(offset)\n  const input = normalized.expandDims(0)\n  \/\/ const output = await tf.tidy(() =&gt; model.predict({ input })).data() \n\n\/\/ MobileNet V2\n  const predictions = labels\n    .map((label, index) =&gt; ({ label, accuracy: output[index] }))\n    .sort((a, b) =&gt; b.accuracy - a.accuracy)\n  const time = `${(performance.now() - t0).toFixed(1)} ms`\n  return { predictions, time }\n}\n\nconst start = async () =&gt; {\n  const input = document.getElementById('input')\n  const output = document.getElementById('output')\n  const model = await loadModel()\n  const predictions = await predict(input, model)\n  output.append(JSON.stringify(predictions, null, 2))\n}\n\nstart()\n<\/code><\/pre>\n\n<p>EDIT: I also added the HTML-file, just to be sure. <\/p>\n\n<pre><code>&lt;!DOCTYPE html&gt;\n&lt;html lang=\"en\"&gt;\n  &lt;head&gt;\n    &lt;title&gt;Image classifier&lt;\/title&gt;\n  &lt;\/head&gt;\n  &lt;body&gt;\n    &lt;img id=\"input\" src=\"assets\/images\/some-flower.jpg\" \/&gt;\n    &lt;pre id=\"output\"&gt;&lt;\/pre&gt;\n  &lt;\/body&gt;\n&lt;script src=\"https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/tfjs@latest\"&gt;&lt;\/script&gt;\n&lt;\/html&gt;\n<\/code><\/pre>\n","answers":[{"AnswerId":"57896367","CreationDate":"2019-09-11T20:27:08.513","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<blockquote>\n  <p>import { loadFrozenModel } from '@tensorflow\/tfjs-converter'<\/p>\n<\/blockquote>\n\n<p><code>loadFrozenModel<\/code> is not exported from <code>@tensorflow\/tfjs-converter<\/code>. It is rather in the namespace of <code>@tensorflow\/tfjs<\/code>. Since you've already imported the CDN scripts, you only need to load the model using <code>tf.loadFrozenModel<\/code><\/p>\n\n<pre><code>const model = await loadFrozenModel(MODEL_URL, WEIGHTS_URL)\n<\/code><\/pre>\n\n<p>Also <code>tf.fromPixels<\/code> has been changed to <code>tf.browser.fromPixels<\/code><\/p>\n"}]}
+{"QuestionId":57895021,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><audio><audio-streaming>","CreationDate":"2019-09-11T18:37:46.757","AcceptedAnswerId":null,"OwnerUserId":265683.0,"Title":"In Python how can I make this model check against the audio output stream for a PC","Body":"<p>I have been following the tutorial on <a href=\"https:\/\/www.tensorflow.org\/tutorials\/sequences\/audio_recognition\" rel=\"nofollow noreferrer\">Audio Recognition for TensorFlow<\/a>.<\/p>\n\n<p>I have trained a model, and I want to run that model against a constant stream which is the audio output device on a computer. I.e whatever is coming out of the speakers. The model is super simple, and not complicated testing for a very simple tone, the model is almost certainly overtrained but at this point I am not concerned about this. The output from the training is:<\/p>\n\n<pre><code>I0911 19:29:05.357377 4477539776 fishem.py:268] Step 250: Validation accuracy = 100.0% (N=4)\nI0911 19:29:05.357587 4477539776 fishem.py:276] Saving to \"logs\/commands_train\/conv.ckpt-250\"\nI0911 19:29:05.419842 4477539776 fishem.py:280] set_size=4\nW0911 19:29:05.487061 4477539776 fishem.py:299] Confusion Matrix:\n [[1 0 0 0 0]\n [0 0 0 0 0]\n [0 0 2 0 0]\n [0 0 0 0 0]\n [0 0 0 0 1]]\nW0911 19:29:05.487216 4477539776 fishem.py:301] Final test accuracy = 100.0% (N=4)\n<\/code><\/pre>\n\n<p>The article doesn't explain how to set a model to check against an audio stream such as this, instead focussing on a <a href=\"https:\/\/www.tensorflow.org\/tutorials\/sequences\/audio_recognition#streaming_accuracy\" rel=\"nofollow noreferrer\">generated stream of audio<\/a> and mobile device audio streams.<\/p>\n\n<p>This is a question both for Python and TensorFlow communities, given a model generated using the initial article, how can I make it test against the audio output on a computer.<\/p>\n\n<p>Many thanks!<\/p>\n","answers":[{"AnswerId":"57895676","CreationDate":"2019-09-11T19:28:25.983","ParentId":null,"OwnerUserId":"5134333","Title":null,"Body":"<p>The simplest way I can think of is a while loop without any breaking condition. Within this while loop, you'll have to capture the audio stream (with <a href=\"https:\/\/people.csail.mit.edu\/hubert\/pyaudio\/\" rel=\"nofollow noreferrer\">pyaudio<\/a> for example) and vectorize it to the same vector input structure of your model. Next, you will have to feed this vector to your model and capture the event that you are interested in. <\/p>\n"}]}
+{"QuestionId":57895924,"AnswerCount":0,"Tags":"<tensorflow><keras><neural-network>","CreationDate":"2019-09-11T19:50:57.360","AcceptedAnswerId":null,"OwnerUserId":10295417.0,"Title":"CuDNNGRU model overfitting, acc and val_acc not changing","Body":"<p>so i am fairly new at this. working on a model to predict a time series problem, and it overfits quickly after around epoch 25, i did some dropout and reduces the network complexity, but still the model can't resist overfitting after a while. also the acc and val_acc are frozen from the beginning to the end. here is the network architecture:<\/p>\n\n<pre><code>regressor = Sequential()\n\nregressor.add(CuDNNGRU(units = 50, return_sequences = True, input_shape = (X_train.shape[1], 1)))\nregressor.add(Dropout(0.3))\n\nregressor.add(CuDNNGRU(units = 50, return_sequences = True))\nregressor.add(Dropout(0.3))\n\nregressor.add(CuDNNGRU(units = 50, return_sequences =False))\nregressor.add(Dropout(0.3))\n\n# i've intentionally commented out this layer to reduce complexity\n# regressor.add(CuDNNGRU(units = 50))\n# regressor.add(Dropout(0.2))\n\nregressor.add(Dense(units = 1))\n\nregressor.compile(optimizer = 'adam', loss = 'mean_squared_error', metrics=['accuracy'])\n\nhistory = regressor.fit(X_train, y_train, epochs = 60, batch_size = 64, validation_split=0.2)\n\n<\/code><\/pre>\n\n<p>and this is basically how the epochs unravel:<\/p>\n\n<pre><code>Train on 127936 samples, validate on 31984 samples\nEpoch 1\/60\n127936\/127936 [==============================] - 99s 772us\/step - loss: 0.0015 - acc: 7.8164e-06 - val_loss: 0.0012 - val_acc: 3.1266e-05\nEpoch 2\/60\n127936\/127936 [==============================] - 98s 763us\/step - loss: 3.1164e-04 - acc: 7.8164e-06 - val_loss: 8.9437e-04 - val_acc: 3.1266e-05\nEpoch 3\/60\n127936\/127936 [==============================] - 97s 762us\/step - loss: 2.3211e-04 - acc: 7.8164e-06 - val_loss: 0.0010 - val_acc: 3.1266e-05\nEpoch 4\/60\n127936\/127936 [==============================] - 98s 765us\/step - loss: 2.2029e-04 - acc: 7.8164e-06 - val_loss: 0.0016 - val_acc: 3.1266e-05\nEpoch 5\/60\n127936\/127936 [==============================] - 97s 760us\/step - loss: 2.1758e-04 - acc: 7.8164e-06 - val_loss: 0.0028 - val_acc: 3.1266e-05\nEpoch 6\/60\n127936\/127936 [==============================] - 98s 765us\/step - loss: 2.1232e-04 - acc: 7.8164e-06 - val_loss: 0.0016 - val_acc: 3.1266e-05\nEpoch 7\/60\n127936\/127936 [==============================] - 97s 761us\/step - loss: 2.1088e-04 - acc: 7.8164e-06 - val_loss: 0.0012 - val_acc: 3.1266e-05\nEpoch 8\/60\n127936\/127936 [==============================] - 97s 760us\/step - loss: 2.0391e-04 - acc: 7.8164e-06 - val_loss: 0.0012 - val_acc: 3.1266e-05\nEpoch 9\/60\n127936\/127936 [==============================] - 97s 761us\/step - loss: 2.0780e-04 - acc: 7.8164e-06 - val_loss: 6.3458e-04 - val_acc: 3.1266e-05\nEpoch 10\/60\n127936\/127936 [==============================] - 97s 762us\/step - loss: 2.0697e-04 - acc: 7.8164e-06 - val_loss: 5.3828e-04 - val_acc: 3.1266e-05\nEpoch 11\/60\n127936\/127936 [==============================] - 98s 763us\/step - loss: 2.0546e-04 - acc: 7.8164e-06 - val_loss: 4.7978e-04 - val_acc: 3.1266e-05\nEpoch 12\/60\n127936\/127936 [==============================] - 98s 763us\/step - loss: 2.0410e-04 - acc: 7.8164e-06 - val_loss: 3.6436e-04 - val_acc: 3.1266e-05\nEpoch 13\/60\n127936\/127936 [==============================] - 98s 763us\/step - loss: 2.0250e-04 - acc: 7.8164e-06 - val_loss: 3.5326e-04 - val_acc: 3.1266e-05\nEpoch 14\/60\n127936\/127936 [==============================] - 97s 758us\/step - loss: 2.0288e-04 - acc: 7.8164e-06 - val_loss: 4.8453e-04 - val_acc: 3.1266e-05\nEpoch 15\/60\n127936\/127936 [==============================] - 98s 763us\/step - loss: 2.0580e-04 - acc: 7.8164e-06 - val_loss: 0.0014 - val_acc: 3.1266e-05\nEpoch 16\/60\n127936\/127936 [==============================] - 97s 760us\/step - loss: 2.0156e-04 - acc: 7.8164e-06 - val_loss: 0.0011 - val_acc: 3.1266e-05\nEpoch 17\/60\n127936\/127936 [==============================] - 97s 762us\/step - loss: 2.0055e-04 - acc: 7.8164e-06 - val_loss: 0.0012 - val_acc: 3.1266e-05\nEpoch 18\/60\n127936\/127936 [==============================] - 97s 759us\/step - loss: 2.0162e-04 - acc: 7.8164e-06 - val_loss: 0.0013 - val_acc: 3.1266e-05\nEpoch 19\/60\n127936\/127936 [==============================] - 97s 761us\/step - loss: 1.9856e-04 - acc: 7.8164e-06 - val_loss: 7.1617e-04 - val_acc: 3.1266e-05\nEpoch 20\/60\n127936\/127936 [==============================] - 97s 758us\/step - loss: 2.0146e-04 - acc: 7.8164e-06 - val_loss: 8.7160e-04 - val_acc: 3.1266e-05\nEpoch 21\/60\n127936\/127936 [==============================] - 97s 761us\/step - loss: 2.0139e-04 - acc: 7.8164e-06 - val_loss: 0.0017 - val_acc: 3.1266e-05\nEpoch 22\/60\n127936\/127936 [==============================] - 97s 760us\/step - loss: 2.0001e-04 - acc: 7.8164e-06 - val_loss: 0.0013 - val_acc: 3.1266e-05\nEpoch 23\/60\n127936\/127936 [==============================] - 97s 761us\/step - loss: 2.0003e-04 - acc: 7.8164e-06 - val_loss: 7.9431e-04 - val_acc: 3.1266e-05\nEpoch 24\/60\n127936\/127936 [==============================] - 98s 763us\/step - loss: 1.9823e-04 - acc: 7.8164e-06 - val_loss: 0.0011 - val_acc: 3.1266e-05\nEpoch 25\/60\n127936\/127936 [==============================] - 98s 762us\/step - loss: 1.9902e-04 - acc: 7.8164e-06 - val_loss: 0.0012 - val_acc: 3.1266e-05\nEpoch 26\/60\n127936\/127936 [==============================] - 97s 762us\/step - loss: 1.9857e-04 - acc: 7.8164e-06 - val_loss: 0.0012 - val_acc: 3.1266e-05\nEpoch 27\/60\n127936\/127936 [==============================] - 97s 761us\/step - loss: 1.9652e-04 - acc: 7.8164e-06 - val_loss: 0.0020 - val_acc: 3.1266e-05\nEpoch 28\/60\n127936\/127936 [==============================] - 97s 760us\/step - loss: 1.9940e-04 - acc: 7.8164e-06 - val_loss: 0.0017 - val_acc: 3.1266e-05\nEpoch 29\/60\n127936\/127936 [==============================] - 97s 758us\/step - loss: 1.9802e-04 - acc: 7.8164e-06 - val_loss: 0.0032 - val_acc: 3.1266e-05\nEpoch 30\/60\n127936\/127936 [==============================] - 97s 761us\/step - loss: 1.9759e-04 - acc: 7.8164e-06 - val_loss: 0.0018 - val_acc: 3.1266e-05\nEpoch 31\/60\n127936\/127936 [==============================] - 97s 757us\/step - loss: 1.9716e-04 - acc: 7.8164e-06 - val_loss: 0.0033 - val_acc: 3.1266e-05\nEpoch 32\/60\n127936\/127936 [==============================] - 97s 760us\/step - loss: 1.9611e-04 - acc: 7.8164e-06 - val_loss: 0.0038 - val_acc: 3.1266e-05\nEpoch 33\/60\n127936\/127936 [==============================] - 97s 762us\/step - loss: 1.9661e-04 - acc: 7.8164e-06 - val_loss: 0.0066 - val_acc: 3.1266e-05\nEpoch 34\/60\n127936\/127936 [==============================] - 98s 762us\/step - loss: 1.9186e-04 - acc: 7.8164e-06 - val_loss: 0.0068 - val_acc: 3.1266e-05\nEpoch 35\/60\n127936\/127936 [==============================] - 97s 762us\/step - loss: 1.9449e-04 - acc: 7.8164e-06 - val_loss: 0.0077 - val_acc: 3.1266e-05\n<\/code><\/pre>\n\n<p>as you can see, after epoch 23 the val_loss started soaring. also why is there ever no chance in acc and val_acc in this model?<\/p>\n","answers":[]}
+{"QuestionId":57895958,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-11T19:53:42.807","AcceptedAnswerId":null,"OwnerUserId":2886575.0,"Title":"tensorflow intermediate numpy calculations","Body":"<p>In <code>tensorflow<\/code>, we define a computation graph and execute the computation graph as separate exercises.  Valid components of the computation graph include <code>tensorflow<\/code> c++ OPS.  However, I am in need of doing some intermediate calculations that are relatively easy to do in <code>numpy<\/code>, but have no <code>tensorflow<\/code> OP.  Specifically, I need to compute some eigenvalues for a non-hermitian matrix.  I have also implemented the derivatives for all of these calculations.  However, this is just one step in a long pipeline of computations, 95% of which can be done in <code>tensorflow<\/code>.<\/p>\n\n<p>Ideally:<\/p>\n\n<pre><code>class myNumpyTensorflowOP(object):\n    def forward(self, x):\n        pass\n    def derivatives(self, x):\n        pass\n\nwith tf.Graph().as_default(), tf.Session() as sess:\n    foo = np.array([1,2,3])\n    tf_foo = tf.placeholder(foo.dtype, foo.shape)\n    some_tf_ops = tf.reshape(tf_foo, (foo.shape[0], 1))\n    my_ops = myNumpyTensorflowOP.forward(some_tf_ops)\n    more_tf_ops = tf.reshape(my_ops, -1)\n    sess.run(more_tf_ops, feed_dict={tf_foo:foo})\n<\/code><\/pre>\n\n<p>However, I'm pretty confident that such a thing does not exist.  Can I emulate this behavior somehow in a not totally hacky sort of way?  What options do I have for merging built-in <code>tensorflow<\/code> ops and my own hand-rolled operations?<\/p>\n","answers":[]}
+{"QuestionId":57896055,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-09-11T20:03:07.313","AcceptedAnswerId":57907026.0,"OwnerUserId":6763224.0,"Title":"Keras returns wrong value for sum","Body":"<p>I have this code to understand how Jaccard Index works and I am trying to implement it as a Keras custom metric, but the return of this function is wrong and I have no idea why (I am new to Keras btw).<\/p>\n\n<p>This function creates a rectangular area to test intersections:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def block(start, width=20, heigth=10, length=60):\n    block = np.zeros((1, heigth, length))\n    block[:, :, start:(start + width)] = 1\n    return block.astype('uint8')\n<\/code><\/pre>\n\n<p>If I get a <code>block(10)<\/code> and a <code>block(20)<\/code> there should be two 10x20 blocks with a 10x10 area of intersection so the union must be <code>2*(10*20) - 10*10<\/code> which is confirmed by this Numpy line:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>np.sum(((block(10) + block(20)) &gt; 0).astype('uint8'))\n<\/code><\/pre>\n\n<p>But when I try to use this Keras function:<\/p>\n\n<pre><code>def test(a, b):\n    return K.sum(K.cast((a + b) &gt; 0, dtype='uint8'))\n<\/code><\/pre>\n\n<p>The result is 44 when I call it with <code>K.eval(test(block(10), block(20)))<\/code><\/p>\n\n<p>EDIT (new test):<\/p>\n\n<pre><code>def test(a, b):\n    return K.sum(a), K.sum(b), K.sum(a) + K.sum(b)\n\n[K.eval(elem) for elem in test(block(10), block(20))]\n<\/code><\/pre>\n\n<p>This is the result: <code>[200, 200, 144]<\/code><\/p>\n\n<p>What am I doing wrong to get this result?<\/p>\n","answers":[{"AnswerId":"57907026","CreationDate":"2019-09-12T12:44:14.023","ParentId":null,"OwnerUserId":"2653663","Title":null,"Body":"<p>The dtype of the Keras sum is inferred from the object that is summed over, so in this case the sum will be of type <code>uint8<\/code>. Your result is 300. <code>uint8<\/code> can take 256 different values, and <code>300 - 44 = 256<\/code>. If you take a closer look on your python code:<\/p>\n\n<pre><code>blocks = block(10) + block(20) # dtype='uint8'\nmask = (blocks &gt; 0) # dtype=bool\nmask_as_uint = mask.astype('uint8') # dtype='uint8', which is an unnescessary conversion\nnp.sum(mask_as_uint) # dtype='uint64\n<\/code><\/pre>\n\n<p>So your result is of type <code>uint64<\/code>, while you ask Keras to give you a result in <code>uint8<\/code>. In fact the following gives 44 as an answer as well:<\/p>\n\n<pre><code>np.int8(300)\n<\/code><\/pre>\n\n<p>In short, switch to a bigger dtype for the sum or do the operations in an order manually where <code>uint8<\/code> doesn't have shortcomings.<\/p>\n"}]}
+{"QuestionId":57896088,"AnswerCount":0,"Tags":"<tensorflow><profiling><tensorboard>","CreationDate":"2019-09-11T20:05:09.060","AcceptedAnswerId":null,"OwnerUserId":2971299.0,"Title":"How to read the \"local.trace\" file generated by tensorflow summary callback?","Body":"<p>I could not find a way to read the <code>local.trace<\/code> file that gets generated after I fit my model with a summary callback enabled:<\/p>\n\n<pre><code>LOG_DIR = \"\/tmp\/tfds-keras\/tl\"+str(counter)\nmodel.fit(\n    dataset,\n    validation_data=dataset_test,\n    epochs=EPOCHS,\n    steps_per_epoch=DATASET_SIZE \/\/ BATCH_SIZE\/\/10,\n    validation_steps=20,\n    callbacks=[tf.keras.callbacks.TensorBoard(log_dir=LOG_DIR, update_freq=100\n                                            )],\n)\ncounter=counter+1\n<\/code><\/pre>\n\n<p>The file gets saved to:<\/p>\n\n<pre><code>\u2514\u2500\u2500 tl1\n    \u251c\u2500\u2500 events.out.tfevents.1568229131.aurora01\n    \u251c\u2500\u2500 events.out.tfevents.1568229134.aurora01.profile-empty\n    \u2514\u2500\u2500 plugins\n        \u2514\u2500\u2500 profile\n            \u2514\u2500\u2500 2019-09-11_15-12-14\n                \u2514\u2500\u2500 local.trace\n<\/code><\/pre>\n\n<p>And visual inspection shows that it displays information of operations executed by the gpu.<\/p>\n\n<pre><code>...\n'tf_data_parallel_interleave_worker_pool^P&lt;80&gt;&lt;AE&gt;&lt;90&gt;&lt;E5&gt;^M\"J^Z*edge_2139_metrics\/acc\/Identity::MemcpyDtoHH&lt;C0&gt;&lt;80&gt;&lt;D8&gt;&lt;F1&gt;&lt;DB&gt;^DP&lt;C0&gt;&lt;96&gt;&lt;B1&gt;^BZ^P\n^Elabel^R^G4 bytes\"&gt;^Z^^edge_2137_loss\/mul::MemcpyDtoHH&lt;80&gt;&lt;A0&gt;&lt;83&gt;&lt;F5&gt;&lt;DB&gt;^DP&lt;C0&gt;&lt;9F&gt;&lt;AB&gt;^CZ^P\n^Elabel^R^G4 bytes\"^A^P^A^Z%edge_2130_IteratorGetNext::MemcpyHtoDH&lt;C0&gt;&lt;9D&gt;&lt;EE&gt;&lt;D7&gt;&lt;C9&gt;^DP&lt;C0&gt;&lt;FC&gt;&lt;F0&gt;5Z^U\n^Elabel^R^L393216 bytes\"^A^Z$edge_287_IteratorGetNext::MemcpyHtoDH&lt;80&gt;&lt;BD&gt;&lt;99&gt;&lt;DB&gt;&lt;C9&gt;^DP&lt;80&gt;&lt;D4&gt;&lt;CB&gt;1Z^R\n^Elabel^R       256 bytes\"^A^Z2edge_1494_loss\/dense_loss\/num_elements::MemcpyHtoDH&lt;80&gt;&lt;D9&gt;\u05e9&lt;DB&gt;^DP&lt;80&gt;&lt;92&gt;&lt;F4&gt;^AZ^P\n^Elabel^R^G4 bytes\"^A^Z&amp;edge_1497_metrics\/acc\/Size::MemcpyHtoDH&lt;80&gt;&lt;F4&gt;\u016c&lt;DB&gt;^DP&lt;C0&gt;&lt;96&gt;&lt;B1&gt;^BZ^P\n...\n<\/code><\/pre>\n\n<p>I have tried to open the <code>tl1<\/code> folder on tensorboard and although it displays a profile tab, nothing can be visualized.<\/p>\n","answers":[]}
+{"QuestionId":57896235,"AnswerCount":0,"Tags":"<python><tensorflow><keras><neural-network>","CreationDate":"2019-09-11T20:17:34.683","AcceptedAnswerId":null,"OwnerUserId":8658456.0,"Title":"TypeError: object of type 'InputLayer' has no len()","Body":"<p>I want to build a CNN from another CNN to extract the featured vector of an image. The idea is only to take the first 13 layers of the first CNN and build the second with these layers.<\/p>\n\n<p>I'm using a Google Colab Notebook with GPU's<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.models import Model\n\nlayer_input_f_nmist = model_f_mnist.input\nlayer_outputs = [layer.output for layer in model_f_mnist.layers[:13]] \n\nmodel_mnist_featured = Model(inputs = layer_input_f_nmist, outputs = layer_outputs)\n\nfeatured_f_mnist_train = model_mnist_featured.predict(X_f_mnist_train)\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>TypeError                                 Traceback (most recent call last)\n\n&lt;ipython-input-74-ceb627fe9c83&gt; in &lt;module&gt;()\n      4 layer_outputs = [layer.output for layer in model_f_mnist.layers[:13]]\n      5 \n----&gt; 6 model_mnist_featured = Model(inputs = layer_input_f_nmist, outputs = layer_outputs)\n      7 \n      8 featured_f_mnist_train = model_mnist_featured.predict(X_f_mnist_train)\n\n4 frames\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/engine\/network.py in build_map(tensor, finished_nodes, nodes_in_progress, layer, node_index, tensor_index)\n   1400 \n   1401         # Propagate to all previous tensors connected to this node.\n-&gt; 1402         for i in range(len(node.inbound_layers)):\n   1403             x = node.input_tensors[i]\n   1404             layer = node.inbound_layers[i]\n\nTypeError: object of type 'InputLayer' has no len()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57896249,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><multilabel-classification>","CreationDate":"2019-09-11T20:18:27.927","AcceptedAnswerId":null,"OwnerUserId":7718241.0,"Title":"Tensorflow - Multilabel classification loss decreasing slowly\/not changing at all","Body":"<p>I am working on a neural network in Tensorflow. My network performs multilabel classification to predict which users bid on which cars. I have a CSV with three columns; User IDs, Highest bid, and Make names (car models), however I only use User IDs and Make names. User IDs are the input and Make names are the labels, where 86 unique Make names exist. My problem is that when I train my model with the dataset, my loss is decreasing very slowly and accuracy is almost not changing at all.<\/p>\n\n<p>I have tried changing the amount of layers I am using, as well as trying different amounts of neurons in each hidden layer, however I did not have any luck with that.<\/p>\n\n<pre><code>from __future__ import absolute_import, division, print_function, unicode_literals\n\nimport os\nimport matplotlib.pyplot as plt\nimport tensorflow as tf\nimport pandas as pd\nimport numpy as np\n\n#\n# READ FILE AND SET FEATURE &amp; LABEL\n#\nraw_train_data = pd.read_csv(\"data2016NEW.csv\", usecols=[\"USERID\"])\n#raw_test_data = pd.read_csv(\"data2017.csv\", usecols=[\"USERID\"])\n\ntrain_target = pd.read_csv(\"data2016NEW.csv\", usecols=[\"MAKENAME\"])\n#test_target = pd.read_csv(\"data2017.csv\", usecols=[\"MAKENAME\"])\n\ntrain_dataset = tf.data.Dataset.from_tensor_slices((raw_train_data.values, train_target.values))\n#test_dataset = tf.data.Dataset.from_tensor_slices((raw_test_data.values, test_target.values))\n\n\ntrain_dataset_shuffled_batched = train_dataset.shuffle(len(raw_train_data)).batch(512)\n#test_dataset_shuffled_batched = test_dataset.shuffle(len(raw_test_data)).batch(512)\n\ndef get_compiled_model():\n    model = tf.keras.Sequential([\n        #input layer\n        tf.keras.layers.Dropout(0.2, input_shape=raw_train_data.shape[1:]),\n        #hidden layers\n        tf.keras.layers.Dense(360, activation='relu'),\n        tf.keras.layers.Dense(360, activation='relu'),\n        tf.keras.layers.Dense(360, activation='relu'),\n        tf.keras.layers.Dense(360, activation='relu'),\n        tf.keras.layers.Dense(360, activation='relu'),\n        tf.keras.layers.Dense(360, activation='relu'),\n        #output layer\n        tf.keras.layers.Dense(90, activation='softmax')\n    ])\n\n    model.compile(\n        optimizer=tf.keras.optimizers.Adam(learning_rate=0.01),\n        loss=tf.keras.losses.sparse_categorical_crossentropy,\n        metrics=['accuracy']\n    )\n\n    return model\n\nmodel = get_compiled_model()\nmodel.summary()\n\nhistory = model.fit(train_dataset_shuffled_batched, epochs=1000, verbose=2)\n<\/code><\/pre>\n\n<p>Here is an image that shows a small sample of my data (my dataset is around 160000 rows):\n<img src=\"https:\/\/i.imgur.com\/Cj02zhJ.png\" alt=\"Dataset example\"><\/p>\n\n<p>As you can see, User IDs are normalized between 0 and 1, while the label Make name are integers.<\/p>\n\n<p>Appreciate all the help, cheers!<\/p>\n","answers":[{"AnswerId":"57896459","CreationDate":"2019-09-11T20:33:58.720","ParentId":null,"OwnerUserId":"11244336","Title":null,"Body":"<p>Before you train, you should see whether you can see a correlation between your features.<\/p>\n\n<p>For example : <em>can you predict if what car you will buy based on your birthday ? Is this sole information enough ?<\/em><\/p>\n\n<p>The fact that the learning rate changes slowly can be linked to <strong>the absence of correlation between x and y<\/strong><\/p>\n\n<p>To see if it is the case, simply replace your values with random values, then with value that have a clear pattern like (ID&lt;0.5 => target >20). If your accuracy improves with the pattern but not the random values, it may indicate you need more features.<\/p>\n\n<p>If you are sure about the correlation Did you try changing the learning rate ? A very high learning rate on a well-defined task would create a \"vibrating\" pattern, where learning rate goes up and down. This could imply you need to fine-tune it.<\/p>\n\n<p>Hope this helps.<\/p>\n"}]}
+{"QuestionId":57896357,"AnswerCount":1,"Tags":"<pytorch><repeat>","CreationDate":"2019-09-11T20:26:48.177","AcceptedAnswerId":null,"OwnerUserId":7037749.0,"Title":"How to repeat tensor in a specific new dimension in PyTorch","Body":"<p>If I have a tensor <code>A<\/code> which has shape <code>[M, N]<\/code>,\nI want to repeat the tensor K times so that the result <code>B<\/code> has shape <code>[M, K, N]<\/code>\nand each slice <code>B[:, k, :]<\/code> should has the same data as <code>A<\/code>.\nWhich is the best practice without a for loop.\n<code>K<\/code> might be in other dimension.<\/p>\n\n<p><code>torch.repeat_interleave()<\/code> and <code>tensor.repeat()<\/code> does not seem to work. Or I am using it in a wrong way.<\/p>\n","answers":[{"AnswerId":"57896754","CreationDate":"2019-09-11T21:01:48.523","ParentId":null,"OwnerUserId":"2790047","Title":null,"Body":"<p><a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.Tensor.repeat\" rel=\"nofollow noreferrer\">tensor.repeat<\/a> should suit your needs but you need to insert a unitary dimension first. For this we could use either <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.Tensor.reshape\" rel=\"nofollow noreferrer\"><code>tensor.reshape<\/code><\/a> or <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.Tensor.unsqueeze\" rel=\"nofollow noreferrer\"><code>tensor.unsqueeze<\/code><\/a>. Since <code>unsqueeze<\/code> is specifically defined to insert a unitary dimension we will use that.<\/p>\n\n<pre><code>B = A.unsqueeze(1).repeat(1, K, 1)\n<\/code><\/pre>\n\n<p><em>Code Description<\/em> <code>A.unsqueeze(1)<\/code> turns <code>A<\/code> from an <code>[M, N]<\/code> to <code>[M, 1, N]<\/code> and <code>.repeat(1, K, 1)<\/code> repeats the tensor <code>K<\/code> times along the second dimension.<\/p>\n"}]}
+{"QuestionId":57896578,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-11T20:45:07.157","AcceptedAnswerId":null,"OwnerUserId":12055148.0,"Title":"Converting session to Tensorflow 2.0 on a prebuilt model","Body":"<p>I am trying to convert code from a project supported by Tensorflow 1.15 to Tensorflow 2.0 that loads a model supported in Tensorflow 1.15 and runs it against test data to generate predictions. Is this model still supported in Tensorflow 2.0? If so, can anybody help me convert it to Tensorflow 2.0? I am not very familiar with Tensorflow in general and because I am loading a prebuilt model I do not have a lot of underlying knowledge about that model. <\/p>\n\n<p>I have already run tf_upgrade_v2 on the project and everything else converted over as far as I can tell. My main remaining issue rests in removing sess.run from the equation. <\/p>\n\n<p>I have attempted to combine all the relevant code together. This consists of nearly all that is related to tensorflow in the file and the file does not reference any other files in the project. Let me know if there is anything missing I need to add.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>session = tf.Session()\nwith tf.io.gfile.GFile(fname, \"rb\") as handle:\n    graph_def = tf.compat.v1.GraphDef()\n    graph_def.ParseFromString(handle.read())\n\ntf.import_graph_def(graph_def, name=\"net\")\ninput_var = tf.compat.v1.get_default_graph().get_tensor_by_name(\"net\/images:0\")\noutput_var = tf.compat.v1.get_default_graph().get_tensor_by_name(\"net\/features:0\")\nfeature_dim = self.output_var.get_shape().as_list()[-1]\n# data_x is an ndarray of shape (9, 128, 64, 3)\nout = session.run(self.output_var, feed_dict={input_var: data_x})\n<\/code><\/pre>\n\n<p>Any help would be greatly appreciated. Thank you!<\/p>\n","answers":[]}
+{"QuestionId":57896583,"AnswerCount":1,"Tags":"<machine-learning><keras><neural-network><deep-learning>","CreationDate":"2019-09-11T20:45:30.733","AcceptedAnswerId":null,"OwnerUserId":11760847.0,"Title":"Keras accuracy plots flat while loss plots not flat","Body":"<p>I am doing deep learning using a multi-layer perceptron for regression. The loss curve turns flat in the third epoch however accuracy curve remains flat at the beginning. I wonder whether this makes sense.<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/jx1q4.png\" alt=\"enter image description here\"><\/p>\n","answers":[{"AnswerId":"57911910","CreationDate":"2019-09-12T17:34:38.067","ParentId":null,"OwnerUserId":"1729802","Title":null,"Body":"<p>Since you didn't provide the code, it would be harder to narrow down what is the problem. Being said, here are some pointers that might help you see what is the problem:<\/p>\n\n<ul>\n<li>Validation set is either small or it is a bad representation of your training set. (bear in mind, if you are using <code>validation_split<\/code> in <code>fit<\/code> function, then keras will only take the last percentage of your training set and will keep it the same for all epochs. <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-is-the-validation-split-computed\" rel=\"nofollow noreferrer\">link<\/a>]).<\/li>\n<li>You are not using any regularization (Dropout, Regularization, Constraints).<\/li>\n<li>The model could be small (layers- and neurons-wise), so it is underfitting.<\/li>\n<\/ul>\n\n<p>Hope these pointers help you with your problem.<\/p>\n"}]}
+{"QuestionId":57896633,"AnswerCount":1,"Tags":"<python><python-3.x><pycharm><pytorch>","CreationDate":"2019-09-11T20:49:47.027","AcceptedAnswerId":null,"OwnerUserId":11657863.0,"Title":"Unable to Install Torch or torch vision in pycharm I am running python 3.6","Body":"<p>Installing torch from PyCharm interpreter but error occurs. Python 3.6 <\/p>\n\n<pre><code>Collecting torch==0.4.1.post2\n\n  Could not find a version that satisfies the requirement torch==0.4.1.post2 (from versions: 0.1.2, 0.1.2.post1, 0.1.2.post2)\nNo matching distribution found for torch==0.4.1.post2\n<\/code><\/pre>\n","answers":[{"AnswerId":"57896922","CreationDate":"2019-09-11T21:18:00.590","ParentId":null,"OwnerUserId":"9494334","Title":null,"Body":"<p>Try installing lower version of torch, go to:<\/p>\n\n<blockquote>\n  <p>File->Setting->Project[project_name]-> Project Interpreter -> + ->\n  search for torch, in right lower corner is check box 'Specify version'<\/p>\n<\/blockquote>\n\n<p>check it and try sevral starting from top. From what i see current is 1.2.0<\/p>\n"}]}
+{"QuestionId":57897080,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-11T21:33:42.607","AcceptedAnswerId":null,"OwnerUserId":2985049.0,"Title":"Load custom loss with extra input in keras","Body":"<p>I have a custom loss function that takes the input to the model as one of the arguments. If I load in the same session in which I train, I can load it no problem using <a href=\"https:\/\/stackoverflow.com\/questions\/48373845\/loading-model-with-custom-loss-keras\">this<\/a> technique.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\ndef custom_loss(inputs):\n    def loss(y_true, y_pred):\n        return ...\n    return loss\n\ninputs = keras.layers.Input(shape=(...))\ny = keras.layers.Activation('tanh')(inputs)\n\nmodel = keras.models.Model(inputs=inputs, outputs=y)\n\nmodel.compile(loss=custom_loss(inputs), optimizer='Adam')\nmodel.fit(...)\nmodel.save('mymodel.h5')\nload_model('mymodel.h5', custom_objects={'custom_loss': custom_loss(inputs})\n<\/code><\/pre>\n\n<p>However, I run into problems when I try to load the model in a later session, because this time I don't have access to the original input tensor. If I make a new inputs placeholder, then the model expects two different sets of inputs and I error out.<\/p>\n\n<pre><code>inputs = keras.layers.Input(shape=(...))\nload_model('mymodel.h5', custom_objects={'custom_loss': custom_loss(inputs})\n<\/code><\/pre>\n\n<p>Is there a good way to solve this problem? The issue at the end of the day is that the inputs haven't been deserialized yet so they can't be passed in to the custom objects.<\/p>\n\n<p>I don't want to just save the weights and create a new model with the same weights because I lose the optimizer state.<\/p>\n","answers":[{"AnswerId":"57919819","CreationDate":"2019-09-13T08:19:46.297","ParentId":null,"OwnerUserId":"9579075","Title":null,"Body":"<p>An alternate way is to compute the loss inside a Keras layer and pass a dummy loss function that just returns the model's output as the loss in the compile method.\nThere are other ways of doing this. But this is the one that I prefer.<\/p>\n\n<pre><code>import tensorflow as tf\nprint('Tensorflow', tf.__version__)\n\ndef custom_loss(tensor):\n    y_true, y_pred, inputs = tensor[0], tensor[1], tensor[1]\n    loss = ...\n    return tf.constant([0], dtype=tf.float32)\n\ndef dummy_loss(y_true, y_pred):\n    return y_pred\n\ndef get_model(training=False):\n    inputs = tf.keras.layers.Input(shape=(10,))\n    y = tf.keras.layers.Activation('tanh')(inputs)\n    if training:\n        targets = tf.keras.layers.Input(shape=(10,)) \n        loss_layer = tf.keras.layers.Lambda(custom_loss)([targets, y, inputs])\n        model = tf.keras.models.Model(inputs=[inputs, targets], outputs=loss_layer)\n    else:\n        model = tf.keras.models.Model(inputs=inputs, outputs=y)\n    return model\n\n\nmodel = get_model(training=True)\nmodel.compile(optimizer='sgd', loss=dummy_loss)\nmodel.save('model.h5')\n\nnew_model = tf.keras.models.load_model('model.h5', custom_objects={'dummy_loss':dummy_loss})\n<\/code><\/pre>\n"}]}
+{"QuestionId":57897237,"AnswerCount":1,"Tags":"<python><docker><tensorflow><tensorflow-serving>","CreationDate":"2019-09-11T21:49:43.993","AcceptedAnswerId":57931593.0,"OwnerUserId":7643739.0,"Title":"How to preprocess data for tensorflow\/serving model?","Body":"<p>I'm relatively new to Tensorflow. I have this model deployed locally, using Docker but the data I'm inputting is not preprocessed so it gives me an error. This might be due to the fact that I preprocessed the data outside the model when creating the model?<\/p>\n\n<p>When I input the data, I get this error from the deployed tensorflow\/serving through Docker:<\/p>\n\n<pre><code>{ \"error\": \"Failed to process element: 0 of \\'instances\\' list. Error: Invalid argument: JSON Value: \\\"Hello\\\" Type: String is not of expected type: float\" }\n<\/code><\/pre>\n\n<p>Meaning it expects an array of tokenized words so, [[32],[1]]. Something like that.<\/p>\n\n<p><strong>This is what my model looks like:<\/strong><\/p>\n\n<pre><code>    model = tf.keras.Sequential([\n            tf.keras.layers.Embedding(\n                vocab_size, embedding_dim, input_length=max_length),\n            tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(32)),\n            tf.keras.layers.Dense(128, activation=\"relu\"),\n            tf.keras.layers.Dense(3, activation=\"softmax\")\n        ])\n<\/code><\/pre>\n\n<p>How can I tokenize and pad the data the same way I did with the initial training data or how can I add that step to the model?<\/p>\n","answers":[{"AnswerId":"57931593","CreationDate":"2019-09-14T00:37:29.860","ParentId":null,"OwnerUserId":"7643739","Title":null,"Body":"<p><strong>UPDATE:<\/strong><\/p>\n\n<p>After a long time searching for the answer, I don't know why it took so long for me to come across pickling, the answer to tokenizing the input request with the same tokenizer would be to pickle, so<\/p>\n\n<pre><code># saving\nwith open('tokenizer.pickle', 'wb') as handle:\n    pickle.dump(tokenizer, handle, protocol=pickle.HIGHEST_PROTOCOL)\n\n# loading\nwith open('tokenizer.pickle', 'rb') as handle:\n    tokenizer = pickle.load(handle)\n<\/code><\/pre>\n\n<p>Same can be done with the padding if needed.<\/p>\n"}]}
+{"QuestionId":57897315,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><keras><deep-learning><conv-neural-network>","CreationDate":"2019-09-11T21:56:52.150","AcceptedAnswerId":null,"OwnerUserId":9059089.0,"Title":"Why we have target_size for DeepLab while CNN can accept any sizes?","Body":"<p>I still have not understood a concept. One reason that we use fully convolutional layer at the end in a CNN network is to handle different images sizes during training. My question is that if this is the case why we always crop or squeeze images into squared sizes in the input section. Please do not say the question is repeated, we use squared images to make it easier, check pyramid pooling, and so on. <\/p>\n\n<p>For example, Here's <a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/models\/blob\/master\/research\/deeplab\/deeplab_demo.ipynb#scrollTo=edGukUHXyymr\" rel=\"nofollow noreferrer\">a link<\/a><\/p>\n\n<p>DeepLab can accept any images with different sizes. But in its code, there is a target_size as (513). Now, if CNN can accept images with different sizes, why we need to use target_size. If this is for converting images into a standard format, why 513?\nDuring training, we should specify batch size. What is our batch_size in this case: (5, None, None, None). Is it possible to have images with different sizes in a batch? \nI read many posts and still, I am confused with these questions:\n- How can we train a model on images with different sizes (imagine that sizes are standard)? I see some codes use a batch size of one. I think it is not a solution. \n- Is there any snipped code that shows how can we define batches for a model like FCN to accept dataset with different sizes?\n- In this paper: Here's <a href=\"https:\/\/arxiv.org\/abs\/1406.4729\" rel=\"nofollow noreferrer\">a link<\/a> my problem was explained but authors again resized images into squared format, if we can use batches comprises of images with different sizes why they proposed that idea of using squared images between 180 by 180 and 224 by 224. <\/p>\n\n<p>Has DeepLab used this part: <a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/models\/blob\/master\/research\/deeplab\/deeplab_demo.ipynb#scrollTo=edGukUHXyymr\" rel=\"nofollow noreferrer\">link<\/a> to make images into a standard format? or for other reason?<\/p>\n\n<pre><code>width, height = image.size\nresize_ratio = 1.0 * 513 \/ max(width, height)\ntarget_size = (int(resize_ratio * width), int(resize_ratio * height))\n<\/code><\/pre>\n\n<p>I could not find the place of their code when they training the model on PASCAL dataset. <\/p>\n\n<p>I expected to find a simple code for Keras or Tensorflow whereas it shows easily that we can apply a CNN model such as FCN or DeepLab for a dataset such as PASCAL VOC2012 (for Segmentation) with images of different sizes without any resizing or cropping. Still, I am looking.<br>\nThank you for detail answers in advance. Please do not repeat answers like you can use batch size one, squared images are common and better, you can add black margins to the images, fully connected layer is the problem, you can use global max pooling, and so on. I am looking to find a code that works on images with different sizes. \nI could not find the place of DeepLab model in TensorFlow GitHub where it accepts batches with different sizes?? <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/deeplab\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>Also in here <a href=\"https:\/\/github.com\/aurora95\/Keras-FCN\" rel=\"nofollow noreferrer\">FCN<\/a> it is trained on COCO dataset with target_size of 320 by 320. Why? it should be any size for FCN.<\/p>\n\n<p>Also, could one explain to me how can we have a batch of images with different sizes? Could we have an np array of different sized images? Batch = [5, none, none, 3]  each of 5 with different sizes.<\/p>\n\n<p>I also found another confusing part in semantic segmentation. Using Keras Augmentation we can not augment image with more than 4 channels. It means that using Keras augmentation, we can not train PASCAL dataset with 21 channels. ?? <\/p>\n","answers":[]}
+{"QuestionId":57897320,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><loss-function>","CreationDate":"2019-09-11T21:57:17.600","AcceptedAnswerId":null,"OwnerUserId":5372627.0,"Title":"how to interpret this TensorBoard graph of loss value?","Body":"<p>After the training i got this graph of loss value . When the loss become approximately smaller than 25,  at some point this value become instantaneously very big (150, 140). <\/p>\n\n<p>Somebody suggested that the problem can be over fitting , but even after i tried different percentages of dropout (30%, 40% 60%) the problem remains.\n<a href=\"https:\/\/i.stack.imgur.com\/LYvxw.png\" rel=\"nofollow noreferrer\">a graph of loss function during training<\/a><\/p>\n","answers":[]}
+{"QuestionId":57897850,"AnswerCount":0,"Tags":"<tensorflow><distributed><tensorflow-estimator>","CreationDate":"2019-09-11T23:11:18.300","AcceptedAnswerId":null,"OwnerUserId":1582366.0,"Title":"Issues with Tensor-flow distributed training (Early stopping)","Body":"<p>During TensorFlow distributed MultiWorkerMirroredStrategy training with I am facing errors during early_stopping.<\/p>\n\n<p>I am using early stopping criterion for TrainSpec for distributed training with 2 nodes.With early stopping hook, I receive the following error immediately on all the worker nodes simultaneously. <\/p>\n\n<p>If the early stopping hook is removed the code finishes to completion. <\/p>\n\n<pre><code>strategy = tf.distribute.experimental.MultiWorkerMirroredStrategy()\n\nrun_config = tf.estimator.RunConfig(\n                    model_dir=model_output_dir,\n                    save_checkpoints_steps=5000,\n                    keep_checkpoint_max=1,\n                    train_distribute=strategy\n\nearly_stopping_mae = tf.estimator.experimental.stop_if_no_decrease_hook(\n                estimator, metric_name='mae', \n                max_steps_without_decrease=20000, min_steps=100)\n\ntrain_spec = tf.estimator.TrainSpec(\n                input_fn=lambda: csv_input_fn(\n                    train_filepath,\n                    hparams['batch_size'],\n                    hparams['num_epochs']),\n                hooks=[early_stopping_mae]\n                )\n<\/code><\/pre>\n\n<p>With early stopping, I receive the following error. Its unclear if early stopping is even supported with multi worker strategy. <\/p>\n\n<pre><code>  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/collective_ops.py\", line 133, in broadcast_send\n    instance_key=instance_key)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/gen_collective_ops.py\", line 159, in collective_bcast_send\n    shape=shape, name=name)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 626, in _apply_op_helper\n    param_name=input_name)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 60, in _SatisfiesTypeConstraint\n    \", \".join(dtypes.as_dtype(x).name for x in allowed_list)))\nTypeError: Value passed to parameter 'input' has DataType bool not in list of allowed values: float32, float16, float64, int32, int64\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57898004,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><conv-neural-network>","CreationDate":"2019-09-11T23:38:23.640","AcceptedAnswerId":57943332.0,"OwnerUserId":4173867.0,"Title":"Error: expected conv3d_1_input to have 5 dimensions, but got array with shape (10, 224, 224, 3)","Body":"<p>I'm trying to train a Neural Network on a dataset for liveness anti-spoofing. I have some videos in two folders named <code>genuine<\/code> and <code>fake<\/code>. I have extracted 10 frames of each video and saved them in two folders with aforementioned names under a new directory <code>tarining<\/code>.<\/p>\n\n<pre><code>--\/training\/\n----\/genuine\/   #containes 10frame*300videos=3000images\n----\/fake\/   #containes 10frame*800videos=8000images\n<\/code><\/pre>\n\n<p>I designed the following 3D Convent using Keras as my first try, but after running it, it throws the following exception:<\/p>\n\n<pre><code>from keras.preprocessing.image import ImageDataGenerator\nfrom keras import Model, optimizers, activations, losses, regularizers, backend, Sequential\nfrom keras.layers import Dense, MaxPooling3D, AveragePooling3D, Conv3D, Input, Flatten, BatchNormalization\n\nBATCH_SIZE = 10\nTARGET_SIZE = (224, 224)\n\ntrain_datagen = ImageDataGenerator(rescale=1.0\/255,\n                                   data_format='channels_last',\n                                   validation_split=0.2,\n                                   shear_range=0.0,\n                                   zoom_range=0,\n                                   horizontal_flip=False,\n                                   featurewise_center=False,\n                                   featurewise_std_normalization=False,\n                                   width_shift_range=False,\n                                   height_shift_range=False)\n\ntrain_generator = train_datagen.flow_from_directory(\".\/training\/\",\n                                                    target_size=TARGET_SIZE,\n                                                    batch_size=BATCH_SIZE,\n                                                    class_mode='binary',\n                                                    shuffle=False,\n                                                    subset='training')\n\nvalidation_generator = train_datagen.flow_from_directory(\".\/training\/\",\n                                                    target_size=TARGET_SIZE,\n                                                    batch_size=BATCH_SIZE,\n                                                    class_mode='binary',\n                                                    shuffle=False,\n                                                    subset='validation')\n\nSHAPE = (10, 224, 224, 3)\nmodel = Sequential()\nmodel.add(Conv3D(filters=128, kernel_size=(1, 3, 3), data_format='channels_last', activation='relu', input_shape=(10, 224, 224, 3)))\nmodel.add(MaxPooling3D(data_format='channels_last', pool_size=(1, 2, 2)))\nmodel.add(Conv3D(filters=64, kernel_size=(2, 3, 3), activation='relu'))\nmodel.add(MaxPooling3D(pool_size=(1, 2, 2)))\nmodel.add(Conv3D(filters=32, kernel_size=(2, 3, 3), activation='relu'))\nmodel.add(Conv3D(filters=32, kernel_size=(2, 3, 3), activation='relu'))\nmodel.add(MaxPooling3D(pool_size=(1, 2, 2)))\nmodel.add(Conv3D(filters=16, kernel_size=(2, 3, 3), activation='relu'))\nmodel.add(Conv3D(filters=16, kernel_size=(2, 3, 3), activation='relu'))\nmodel.add(AveragePooling3D())\nmodel.add(BatchNormalization())\nmodel.add(Dense(32, activation='relu'))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.summary()\nmodel.compile(optimizer=optimizers.adam(), loss=losses.binary_crossentropy, metrics=['accuracy'])\nmodel.fit_generator(train_generator, steps_per_epoch=train_generator.samples\/train_generator.batch_size, epochs=5, validation_data=validation_generator, validation_steps=validation_generator.samples\/validation_generator.batch_size)\nmodel.save('3d.h5')\n<\/code><\/pre>\n\n<p>Here is the Error:<\/p>\n\n<pre><code>ValueError: Error when checking input: expected conv3d_1_input to have 5 dimensions, but got array with shape (10, 224, 224, 3)\n<\/code><\/pre>\n\n<p>And this is the output of <code>model.summary()<\/code><\/p>\n\n<pre><code>Model: \"sequential_1\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nconv3d_1 (Conv3D)            (None, 10, 222, 222, 128) 3584      \n_________________________________________________________________\nmax_pooling3d_1 (MaxPooling3 (None, 10, 111, 111, 128) 0         \n_________________________________________________________________\nconv3d_2 (Conv3D)            (None, 9, 109, 109, 64)   147520    \n_________________________________________________________________\nmax_pooling3d_2 (MaxPooling3 (None, 9, 54, 54, 64)     0         \n_________________________________________________________________\nconv3d_3 (Conv3D)            (None, 8, 52, 52, 32)     36896     \n_________________________________________________________________\nconv3d_4 (Conv3D)            (None, 7, 50, 50, 32)     18464     \n_________________________________________________________________\nmax_pooling3d_3 (MaxPooling3 (None, 7, 25, 25, 32)     0         \n_________________________________________________________________\nconv3d_5 (Conv3D)            (None, 6, 23, 23, 16)     9232      \n_________________________________________________________________\nconv3d_6 (Conv3D)            (None, 5, 21, 21, 16)     4624      \n_________________________________________________________________\naverage_pooling3d_1 (Average (None, 2, 10, 10, 16)     0         \n_________________________________________________________________\nbatch_normalization_1 (Batch (None, 2, 10, 10, 16)     64        \n_________________________________________________________________\ndense_1 (Dense)              (None, 2, 10, 10, 32)     544       \n_________________________________________________________________\ndense_2 (Dense)              (None, 2, 10, 10, 1)      33        \n=================================================================\nTotal params: 220,961\nTrainable params: 220,929\nNon-trainable params: 32\n__________________________________________________________\n<\/code><\/pre>\n\n<p>I'd appreciate any help to fix the exception. By the way, I'm using TensorFlow as backend if it helps to solve the problem.<\/p>\n","answers":[{"AnswerId":"57943332","CreationDate":"2019-09-15T10:40:36.410","ParentId":null,"OwnerUserId":"4173867","Title":null,"Body":"<p>As @thushv89 mentioned in the comments Keras has no build-in video generator which causes a lot of problems for those who will work with big video datasets. Therefore, I wrote a simple VideoDataGenerator which works almost as simple as ImageDataGenerator. The script could be found <a href=\"https:\/\/github.com\/keras-team\/keras\" rel=\"nofollow noreferrer\">here on my github<\/a> in case someone needs it in the future.<\/p>\n"}]}
+{"QuestionId":57898063,"AnswerCount":0,"Tags":"<tensorflow><tensor>","CreationDate":"2019-09-11T23:49:50.487","AcceptedAnswerId":null,"OwnerUserId":395573.0,"Title":"Squeeze all but one axis in TensorFlow","Body":"<p>How can I squeeze all but specific axis(es) of a <code>Tensor<\/code>?<\/p>\n\n<p>The input size (and how many <code>1<\/code> dimension there is) is unknown in advance but I know that I wanna keep the first and second axis only (i.e. the rest can be only 1).<\/p>\n\n<p>e.g. If I have a <code>Tensor<\/code> of shape <code>(1, 4, 1, 1, 1)<\/code> and I want it to be squeezed into <code>(1, 4)<\/code><\/p>\n","answers":[]}
+{"QuestionId":57898331,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><conv-neural-network>","CreationDate":"2019-09-12T00:39:51.413","AcceptedAnswerId":57915112.0,"OwnerUserId":9813733.0,"Title":"Applying convolution on individual layers using keras","Body":"<p>I have a pretrained VGG16 network and I modified the network for PASCAL VOC 2012 dataset using transfer learning. Now, I want to take the output of every layer from the modified VGG16 network and apply convolution on each layer, then upsample them to the same size and add them. This is to identify the important regions in the image. I have taken the output from every layer <\/p>\n\n<pre><code>output = [layer.output for layer in model.layers]\n<\/code><\/pre>\n\n<p>Now I want something like<\/p>\n\n<pre><code>hypercolumns = []\nfor op in output:\n            #apply convolution on this output layer \n            #upsample it to the size of the input image\n            #store this in hypercolumns list\n<\/code><\/pre>\n\n<p>Finally after all the layers are upsampled, I will add them from the list to get a single matrix. Now, I am confused on how to apply convolution without creating a model and do upsampling. Is there a way in keras.<\/p>\n","answers":[{"AnswerId":"57915112","CreationDate":"2019-09-12T22:13:38.067","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>There are several ways to achieve this, some of which are listed <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-the-output-of-an-intermediate-layer\" rel=\"nofollow noreferrer\">here in the Keras FAQ<\/a>. I don't know of any significant difference between the two main ways listed there, so try both and see what works. Here is an example of what you can do:<\/p>\n\n<pre><code>input = model.layers[0].input\nfor op in output:\n    intermediate_function = K.function([input], [output])\n    layer_output = intermediate_function([YOUR_INPUT_DATA_HERE])[0]\n    # do the upsampling and stuff here\n<\/code><\/pre>\n"}]}
+{"QuestionId":57898405,"AnswerCount":1,"Tags":"<python><tensorflow><keras><cross-platform>","CreationDate":"2019-09-12T00:55:04.377","AcceptedAnswerId":null,"OwnerUserId":1801968.0,"Title":"How do I save a model in Tensorflow 2.0 that was trained on a GPU to be loaded on a CPU","Body":"<p>I have a bag of models that I trained and saved on a machine with a GPU. \nI used the following model that was trained and run on a GPU<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    model = Sequential()\n    model.add(CuDNNLSTM(units=30,input_shape=(None, 11), return_sequences=True, name='LAYER1'))\n    model.add(Dropout(.9, name='LAYER2'))\n    model.add(Dense(units=10, activation=\"relu\",name='LAYER3'))\n    model.add(Dropout(.1, name='LAYER4'))\n    model.add(CuDNNLSTM(units=20,return_sequences=False,name='LAYER5'))\n    model.add(Dropout(.1, name='LAYER6'))\n    model.add(Dense(units=3,activation=\"linear\",name='LEVEL7'))\n    rmsprop_opt = RMSprop(lr=learning_rate)\n    model.compile(loss=\"mse\", optimizer=rmsprop_opt)\n<\/code><\/pre>\n\n<p>I save the graph of the model using:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    model_json_dict = json.loads(model.to_json())\n    json.dump(model_json_dict, open(\"my_model_graph.json\", \"w\"))\n<\/code><\/pre>\n\n<p>I then saved the weights using a checkpoint method:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    callback_checkpoint = ModelCheckpoint(filepath=\"model_checkpoint.h5\",\n                                          monitor='val_loss',\n                                          verbose=1,\n                                          save_weights_only=True,\n                                          save_best_only=True)\n\n    callbacks = [callback_checkpoint]\n<\/code><\/pre>\n\n<p>And I fit the model using:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    history = model.fit(feature_train,\n                        label_train,\n                        batch_size=batch_size,\n                        epochs=epochs,\n                        validation_split=validation_split,\n                        callbacks=callbacks)\n<\/code><\/pre>\n\n<p>I would like to read the model back into a machine for prediction that only has a CPU.  I've loaded the model and loaded the weights as follows on a second machine and TF complains about the CPU\/GPU issue.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    model = model_from_json(json.dumps(json.load(open(\"my_model_graph.json\",\"r\"))))\n    model.load_weights(\"model_checkpoint.h5\")\n<\/code><\/pre>\n\n<p>So the question is how do I convert these saved models and their weights into a form that can be reloaded into the second machine with only a CPU?<\/p>\n\n<p>Its confusing about the proper method to do this. There is a SO that shows using a Saver() class. <a href=\"https:\/\/stackoverflow.com\/questions\/33759623\/tensorflow-how-to-save-restore-a-model\">Tensorflow: how to save\/restore a model?<\/a>. And another post that says it cant be done, and another that says its transparent. Whats the recommended method of converting these existing models?  (Retraining them is not an option!)<\/p>\n","answers":[{"AnswerId":"58149058","CreationDate":"2019-09-28T18:09:12.900","ParentId":null,"OwnerUserId":"1801968","Title":null,"Body":"<p>This is how I solved it. My model looks something like this:<\/p>\n\n<pre><code>def build_model(layers, machine, learning_rate, dropout_rate):\n    model = Sequential() #   tf.keras.models.Sequential()\n    if machine == \"GPU\":\n        model.add(\n            CuDNNLSTM(\n                units=layers[1],\n                input_shape=(None, layers[0]),\n                return_sequences=True,\n                name='FIRST_LAYER')\n        )\n    else:\n        model.add(\n            LSTM(\n                units=layers[1],\n                input_shape=(None, layers[0]),\n                return_sequences=True,\n                name='FIRST_LAYER')\n        )\n...\n<\/code><\/pre>\n\n<p>I create the initial model like this:<\/p>\n\n<pre><code>my_model = build_model(layers=layer_sizes, machine='GPU', learning_rate=0.003,dropout_rate=0.05)\n<\/code><\/pre>\n\n<p>... then train the model and save the weights using:<\/p>\n\n<pre><code>my_model.save_weights(\"my_model_weights.h5\")\n<\/code><\/pre>\n\n<p>Now I switch to the CPU instance as follows:<\/p>\n\n<pre><code>my_model = build_model(layers=layer_sizes, machine='CPU', learning_rate=0.003,dropout_rate=0.05)\n<\/code><\/pre>\n\n<p>... then I can load the saved weights as follows:<\/p>\n\n<pre><code>my_model.load_weights(\"my_model_weights.h5\")\n<\/code><\/pre>\n\n<p>Its a shame that the model_from_json api doesn't work across machines!<\/p>\n"}]}
+{"QuestionId":57898521,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><lstm>","CreationDate":"2019-09-12T01:14:16.920","AcceptedAnswerId":null,"OwnerUserId":7215963.0,"Title":"Get Keras LSTM output inside Tensorflow code","Body":"<p>I'm working with time-variant graph embedding, where at each time step, the adjacency matrix of the graph changes. The main idea is to perform the node embedding of each timestep of the graph by looking to a set of node features and the adjacency matrix. The node embedding step is long and complicated, and is not part of the core of the problem, so I will skip this part. Suffice it to say that I use Graph Convolutional Network to embed the nodes.<\/p>\n\n<p>Consider that I have a stack of B adjacency matrices A with sizes NxN, where B = batch size and N = number of nodes in the graph. Also, the matrices are stacked according to a time series, where matrix in index i comes before matrix in index i+1. I have already embedded the nodes of the graph, which results in a matrix of dimensions B x N x E, where E = size of the embedding (parameter). Note that the model has to deal with any graph, therefore, N is not a parameter. Another important comment is that each batch contains adjacency matrices from the same graph, and therefore all matrices of a batch have the same number of node, but the matrices of other batches may have different number of nodes.<\/p>\n\n<p>I now need to pass these embedding through an LSTM cell. I never used Keras before, so I'm having a hard time making the Keras LSTM blend in my Tensorflow code. What I want to do is: pass each node embedding through an LSTM such that the number of timesteps = B and the LSTM batch size = N, that is, the input to my LSTM has the shape [N, B, E], where N and B are only known through execution time. I want the output of my LSTM to have the shape of [B, E*E]. The embedding matrix is called here self.embed_mat. Here is my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def _LSTM_layer(self):\n    with tf.variable_scope(self.scope, reuse=tf.AUTO_REUSE), tf.device(self.device): \n            in_shape = tf.shape(self.embed_mat)\n            lstm_input = tf.reshape(self.embed_mat, [in_shape[1], in_shape[0], EMBED_SIZE]) #lstm = [N, B, E]\n            input_plh = K.placeholder(name=\"lstm_input\", shape=(None, None, EMBED_SIZE))\n            lstm = LSTM(EMBED_SIZE*EMBED_SIZE, input_shape=(None, None, EMBED_SIZE))\n            get_output = K.function(inputs=[input_plh], outputs=[lstm(input_plh)])\n            h = get_output([lstm_input])\n<\/code><\/pre>\n\n<p>I am a bit lost with the K.function part. All I want is the output tensor of the LSTM cell. I've seen that in order to get that with Keras, we need to use K.function, but I don't quite get it what it does. When I call get_output([lstm_input]), I get the following error:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'worker_global\/A\/shape' with dtype int64 and shape [?]\n<\/code><\/pre>\n\n<p>Here, A is the stacked adjacency matrices with dimension BxNxN. What is going on here? Does the value of N needs to be known during graph building step? I think I made some dumb mistake with the LSTM cell, but I can't get what it is.<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"58122825","CreationDate":"2019-09-26T18:32:22.360","ParentId":null,"OwnerUserId":"6826845","Title":null,"Body":"<p>If you want to get the output of your LSTM layer \"out\" given input of \"inp\" in a keras Sequential() model called \"model,\" where \"inp\" is your first \/ input layer and \"out\" is an LSTM layer that happens to be, for the sake of this example, in the 4th position in your sequential model, you would obtain the output of that LSTM layer from the  data you call \"lstm_input\" above with the following code:<\/p>\n\n<pre><code>inp = model.layers[0].input\n\nout = model.layers[3].output\n\ninp_to_out = K.function([inp], [out])\n\noutput = inp_to_out([lstm_input])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57898998,"AnswerCount":4,"Tags":"<python><google-cloud-storage><pytorch><google-cloud-dataflow>","CreationDate":"2019-09-12T02:36:30.803","AcceptedAnswerId":58399675.0,"OwnerUserId":5555483.0,"Title":"Is it possible to load a pretrained Pytorch model from a GCS bucket URL without first persisting locally?","Body":"<p>I'm asking this in the context of Google Dataflow, but also generally. <\/p>\n\n<p>Using PyTorch, I can reference a local directory containing multiple files that comprise a pretrained model. I happen to be working with a Roberta model, but the interface is the same for others.<\/p>\n\n<pre><code>ls some-directory\/\n      added_tokens.json\n      config.json             \n      merges.txt              \n      pytorch_model.bin       \n      special_tokens_map.json vocab.json\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>from pytorch_transformers import RobertaModel\n\n# this works\nmodel = RobertaModel.from_pretrained('\/path\/to\/some-directory\/')\n<\/code><\/pre>\n\n<p>However, my pretrained model is stored in a GCS bucket. Let's call it <code>gs:\/\/my-bucket\/roberta\/<\/code>. <\/p>\n\n<p>In the context of loading this model in Google Dataflow, I'm trying to remain stateless and avoid persisting to disk, so my preference would be to get this model straight from GCS. As I understand it, the PyTorch general interface method <code>from_pretrained()<\/code> can take the string representation of a local dir OR a URL. However, I can't seem to load the model from a GCS URL.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># this fails\nmodel = RobertaModel.from_pretrained('gs:\/\/my-bucket\/roberta\/')\n# ValueError: unable to parse gs:\/\/mahmed_bucket\/roberta-base as a URL or as a local path\n<\/code><\/pre>\n\n<p>If I try to use the public https URL of the directory blob, it will also fail, although that is likely due to <a href=\"https:\/\/cloud.google.com\/storage\/docs\/request-endpoints#cookieauth\" rel=\"nofollow noreferrer\">lack of authentication<\/a> since the credentials referenced in the python environment that can create clients don't translate to public requests to <code>https:\/\/storage.googleapis<\/code><\/p>\n\n<pre><code># this fails, probably due to auth\nbucket = gcs_client.get_bucket('my-bucket')\ndirectory_blob = bucket.blob(prefix='roberta')\nmodel = RobertaModel.from_pretrained(directory_blob.public_url)\n# ValueError: No JSON object could be decoded\n\n# and for good measure, it also fails if I append a trailing \/\nmodel = RobertaModel.from_pretrained(directory_blob.public_url + '\/')\n# ValueError: No JSON object could be decoded\n<\/code><\/pre>\n\n<p>I understand that <a href=\"https:\/\/cloud.google.com\/storage\/docs\/gsutil\/addlhelp\/HowSubdirectoriesWork\" rel=\"nofollow noreferrer\">GCS doesn't actually have subdirectories<\/a> and it's actually just being a flat namespace under the bucket name. However, it seems like I'm blocked by the necessity of authentication and a PyTorch not speaking <code>gs:\/\/<\/code>.<\/p>\n\n<p>I can get around this by persisting the files locally first.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from pytorch_transformers import RobertaModel\nfrom google.cloud import storage\nimport tempfile\n\nlocal_dir = tempfile.mkdtemp()\ngcs = storage.Client()\nbucket = gcs.get_bucket(bucket_name)\nblobs = bucket.list_blobs(prefix=blob_prefix)\nfor blob in blobs:\n    blob.download_to_filename(local_dir + '\/' + os.path.basename(blob.name))\nmodel = RobertaModel.from_pretrained(local_dir)\n<\/code><\/pre>\n\n<p>But this seems like such a hack, and I keep thinking I must be missing something. Surely there's a way to stay stateless and not have to rely on disk persistence!<\/p>\n\n<ul>\n<li>So is there a way to load a pretrained model stored in GCS? <\/li>\n<li>Is there a way to authenticate when doing the public URL request in this context? <\/li>\n<li>Even if there is a way to authenticate, will the non-existence of subdirectories still be an issue?<\/li>\n<\/ul>\n\n<p>Thanks for the help! I'm also happy to be pointed to any duplicate questions 'cause I sure couldn't find any.<\/p>\n\n<hr>\n\n<p>Edits and Clarifications<\/p>\n\n<ul>\n<li><p><em>My Python session is already authenticated to GCS<\/em>, which is why I'm able to download the blob files locally and then point to that local directory with <code>load_frompretrained()<\/code><\/p><\/li>\n<li><p><code>load_frompretrained()<\/code> requires a directory reference because it needs all the files listed at the top of the question, not just <code>pytorch-model.bin<\/code><\/p><\/li>\n<li><p>To clarify question #2, I was wondering if there's some way of giving the PyTorch method a request URL that had encrypted credentials embedded or something like that. Kind of a longshot, but I wanted to make sure I hadn't missed anything.<\/p><\/li>\n<li><p>To clarify question #3 (in addition to the comment on one answer below), even if there's a way to embed credentials in the URL that I don't know about, I still need to reference a <em>directory<\/em> rather than a single blob, and I don't know if the GCS subdirectory would be recognized as such because (as the Google docs state) subdirectories in GCS are an illusion and they don't represent a real directory structure. So I think this question is irrelevant or at least blocked by question #2, but it's a thread I chased for a bit so I'm still curious.<\/p><\/li>\n<\/ul>\n","answers":[{"AnswerId":"58657444","CreationDate":"2019-11-01T10:13:32.993","ParentId":null,"OwnerUserId":"1270539","Title":null,"Body":"<p>currently im not playing with Roberta, but with Bert for Token classification for NER, but i think it has the same mechanism..<\/p>\n\n<p>below its my code:<\/p>\n\n<pre><code>os.environ['GOOGLE_APPLICATION_CREDENTIALS'] = 'your_gcs_auth.json'\n\n# initiate storage\nclient = storage.Client()\nen_bucket = client.get_bucket('your-gcs-bucketname')\n\n# get blob\nen_model_blob = en_bucket.get_blob('your-modelname-in-gcsbucket.bin')\nen_model = en_model_blob.download_as_string()\n\n# because model downloaded into string, need to convert it back\nbuffer = io.BytesIO(en_model)\n\n# prepare loading model\nstate_dict = torch.load(buffer, map_location=torch.device('cpu'))\nmodel = BertForTokenClassification.from_pretrained(pretrained_model_name_or_path=None, state_dict=state_dict, config=main_config)\nmodel.load_state_dict(state_dict)\n<\/code><\/pre>\n\n<p>im not to sure whether <code>download_as_string()<\/code> method save the data into local disk or not, but from what i experience if i execute <code>download_to_filename()<\/code> that function will download the model into my local.<\/p>\n\n<p>also if you modified the config for your transformers network(and you put this in GCS and need to load also), you need to modify class <code>PretrainedConfig<\/code> as well, since it can handle file produced by <code>download_as_string()<\/code> function.<\/p>\n\n<p>cheers, hope it helps<\/p>\n"},{"AnswerId":"57902701","CreationDate":"2019-09-12T08:27:40.887","ParentId":null,"OwnerUserId":"3607203","Title":null,"Body":"<p>As you correctly stated, it seems that out of the box <code>pytorch-transformers<\/code> does not support this, but mainly just because it does not recognize the file link as an URL.<\/p>\n\n<p>After some searching, I found the corresponding error message in <a href=\"https:\/\/github.com\/huggingface\/pytorch-transformers\/blob\/ed717635ff5c2bd5dfa8fd0266f309e314a3e44f\/pytorch_transformers\/file_utils.py\" rel=\"nofollow noreferrer\">this source file<\/a>, around line 144-155.<\/p>\n\n<p>Of course, you could try adding your <code>'gs'<\/code> tag to line 144, and then interpret your connection to GCS as a <code>http<\/code> request (lines 269-272). If GCS accepts this, that should be the only thing required to change in order to work.<br\/>\nIf this does not work, the only immediate fix would be to implement something analogous to the  Amazon S3 bucket functions, but I don't know enough about S3 and GCS buckets to claim any meaningful judgement here.<\/p>\n"},{"AnswerId":"57905776","CreationDate":"2019-09-12T11:30:57.183","ParentId":null,"OwnerUserId":"7806223","Title":null,"Body":"<p>I don't know much about Pytorch or Roberta model, but I'll try to answer your inquiries refering to GCS :<\/p>\n\n<p>1.- \"So is there a way to load a pretrained model stored in GCS?\"<\/p>\n\n<p>In case your model can load the Blob directly from binary: <\/p>\n\n<pre><code>from google.cloud import storage\n\nclient = storage.Client()\nbucket = client.get_bucket(\"bucket name\")\nblob = bucket.blob(\"path_to_blob\/blob_name.ext\")\ndata = blob.download_as_string() # you will have your binary data transformed into string here.\n<\/code><\/pre>\n\n<p>2.- \"Is there a way to authenticate when doing the public URL request in this context?\" <\/p>\n\n<p>Here's the tricky part, because depending on which context you are running the script, it will be authenticated with a default service account. So when you are using the official GCP libs you can:<\/p>\n\n<p>A.- Give permissions to that default service account to access to your bucket\/objects.<\/p>\n\n<p>B.- Create a new service account and authenticate with it inside the script (you will need to generate the authentication token for that service account as well):<\/p>\n\n<pre><code>from google.cloud import storage\nfrom google.oauth2 import service_account\n\nVISION_SCOPES = ['https:\/\/www.googleapis.com\/auth\/devstorage']\nSERVICE_ACCOUNT_FILE = 'key.json'\n\ncred = service_account.Credentials.from_service_account_file(SERVICE_ACCOUNT_FILE, scopes=VISION_SCOPES)\n\nclient = storage.Client(credentials=cred)\nbucket = client.get_bucket(\"bucket_name\")\nblob = bucket.blob(\"path\/object.ext\")\ndata = blob.download_as_string()\n<\/code><\/pre>\n\n<p>However that works because the official libs handle the authentication to the API calls on the background, so in the case of from_pretrained() function not work.<\/p>\n\n<p>So an alternative to that is making the object public, so you can access to it when using the public url.<\/p>\n\n<p>3.- \"Even if there is a way to authenticate, will the non-existence of subdirectories still be an issue?\"<\/p>\n\n<p>Not sure that you mean here, you can have folders inside your bucket.<\/p>\n"},{"AnswerId":"58399675","CreationDate":"2019-10-15T17:13:48.857","ParentId":null,"OwnerUserId":"5555483","Title":null,"Body":"<p>MAJOR EDIT:<\/p>\n\n<p>You can install wheel files on Dataflow workers, and you can also use worker temp storage to persist binary files locally!<\/p>\n\n<p>It's true that (currently as of Nov 2019) you can't do this by supplying a <code>--requirements<\/code> argument. Instead you have to use <code>setup.py<\/code> like this. Assume any constants IN CAPS are defined elsewhere.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>REQUIRED_PACKAGES = [\n    'torch==1.3.0',\n    'pytorch-transformers==1.2.0',\n]\n\nsetup(\n    name='project_dir',\n    version=VERSION,\n    packages=find_packages(),\n    install_requires=REQUIRED_PACKAGES)\n<\/code><\/pre>\n\n<p>Run script<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>python setup.py sdist\n\npython project_dir\/my_dataflow_job.py \\\n--runner DataflowRunner \\\n--project ${GCP_PROJECT} \\\n--extra_package dist\/project_dir-0.1.0.tar.gz \\\n# SNIP custom args for your job and required Dataflow Temp and Staging buckets #\n<\/code><\/pre>\n\n<p>And within the job, here's downloading and using the model from GCS in the context of a custom Dataflow operator. For convenience we wrapped a few utility methods in a SEPARATE MODULE (important to get around Dataflow dependency uploads) and imported them at the LOCAL SCOPE of the custom operator, not global.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class AddColumn(beam.DoFn):\n    PRETRAINED_MODEL = 'gs:\/\/my-bucket\/blah\/roberta-model-files'\n\n    def get_model_tokenizer_wrapper(self):\n        import shutil\n        import tempfile\n        import dataflow_util as util\n        try:\n            return self.model_tokenizer_wrapper\n        except AttributeError:\n            tmp_dir = tempfile.mkdtemp() + '\/'\n            util.download_tree(self.PRETRAINED_MODEL, tmp_dir)\n            model, tokenizer = util.create_model_and_tokenizer(tmp_dir)\n            model_tokenizer_wrapper = util.PretrainedPyTorchModelWrapper(\n                model, tokenizer)\n            shutil.rmtree(tmp_dir)\n            self.model_tokenizer_wrapper = model_tokenizer_wrapper\n            logging.info(\n                'Successfully created PretrainedPyTorchModelWrapper')\n            return self.model_tokenizer_wrapper\n\n    def process(self, elem):\n        model_tokenizer_wrapper = self.get_model_tokenizer_wrapper()\n\n        # And now use that wrapper to process your elem however you need.\n        # Note that when you read from BQ your elements are dictionaries\n        # of the column names and values for each BQ row.\n\n<\/code><\/pre>\n\n<p>Utility functions in SEPARATE MODULE within the codebase. In our case in the project root this was in dataflow_util\/init.py but you don't have to do it that way.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from contextlib import closing\nimport logging\n\nimport apache_beam as beam\nimport numpy as np\nfrom pytorch_transformers import RobertaModel, RobertaTokenizer\nimport torch\n\nclass PretrainedPyTorchModelWrapper():\n    def __init__(self, model, tokenizer):\n        self.model = model\n        self.tokenizer = tokenizer\n\ndef download_tree(gcs_dir, local_dir):\n    gcs = beam.io.gcp.gcsio.GcsIO()\n    assert gcs_dir.endswith('\/')\n    assert local_dir.endswith('\/')\n    for entry in gcs.list_prefix(gcs_dir):\n        download_file(gcs, gcs_dir, local_dir, entry)\n\n\ndef download_file(gcs, gcs_dir, local_dir, entry):\n    rel_path = entry[len(gcs_dir):]\n    dest_path = local_dir + rel_path\n    logging.info('Downloading %s', dest_path)\n    with closing(gcs.open(entry)) as f_read:\n        with open(dest_path, 'wb') as f_write:\n            # Download the file in chunks to avoid requiring large amounts of\n            # RAM when downloading large files.\n            while True:\n                file_data_chunk = f_read.read(\n                    beam.io.gcp.gcsio.DEFAULT_READ_BUFFER_SIZE)\n                if len(file_data_chunk):\n                    f_write.write(file_data_chunk)\n                else:\n                    break\n\n\ndef create_model_and_tokenizer(local_model_path_str):\n    \"\"\"\n    Instantiate transformer model and tokenizer\n\n      :param local_model_path_str: string representation of the local path \n             to the directory containing the pretrained model\n      :return: model, tokenizer\n    \"\"\"\n    model_class, tokenizer_class = (RobertaModel, RobertaTokenizer)\n\n    # Load the pretrained tokenizer and model\n    tokenizer = tokenizer_class.from_pretrained(local_model_path_str)\n    model = model_class.from_pretrained(local_model_path_str)\n\n    return model, tokenizer\n<\/code><\/pre>\n\n<p>And there you have it folks! More details can be found here: <a href=\"https:\/\/beam.apache.org\/documentation\/sdks\/python-pipeline-dependencies\/\" rel=\"nofollow noreferrer\">https:\/\/beam.apache.org\/documentation\/sdks\/python-pipeline-dependencies\/<\/a><\/p>\n\n<hr>\n\n<p><s>What I've discovered is that this whole chain of questioning is irrelevant because Dataflow only allows you to install source distribution packages on workers which means you can't actually install PyTorch.<\/p>\n\n<p>When you supply a <code>requirements.txt<\/code> file, Dataflow will install with the <code>--no-binary<\/code> flag which prevents installation of Wheel (.whl) packages and only allows source distributions (.tar.gz). I decided trying to roll my own source distribution for PyTorch on Google Dataflow where it's half C++ and part Cuda and part who knows what was a fool's errand.<\/p>\n\n<p>Thanks for the input along the way y'all.<\/s><\/p>\n"}]}
+{"QuestionId":57899426,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-lite>","CreationDate":"2019-09-12T03:42:18.240","AcceptedAnswerId":null,"OwnerUserId":8819056.0,"Title":"What value should I set for \"setNumThreads\" for best performance in tflite interpreter?","Body":"<p>I don't know what value to \"setNumThreads\" for each device to get best performance. <\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/lite\/api_docs\/cc\/class\/tflite\/interpreter#setnumthreads\" rel=\"nofollow noreferrer\">Tensorflow-doc tflite interpreter setnumthreads<\/a><\/p>\n\n<p>And it is not written what value to set in this link.<\/p>\n\n<blockquote>\n  <p>Multi-threaded execution, however, comes at the cost of increased performance variability depending on what else is executed concurrently. This is particularly the case for mobile apps. For example, isolated tests may show 2x speed-up vs single-threaded, but, if another app is executing at the same time, it may result in worse performance than single-threaded.<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/lite\/performance\/best_practices\" rel=\"nofollow noreferrer\">Tensorflow-doc tflite performance best_practices<\/a><\/p>\n","answers":[]}
+{"QuestionId":57899583,"AnswerCount":0,"Tags":"<ios><tensorflow>","CreationDate":"2019-09-12T04:05:45.930","AcceptedAnswerId":null,"OwnerUserId":4556344.0,"Title":"Tensorflow face recognition for iOS","Body":"<p>For face recognition for iOS app, I am using tensorflow.<br>\nI trained facenet model with face images but the pb file is not modified.<br>\nOnly generated a pkl file.<br>\nHow can I use it on iOS app?<\/p>\n","answers":[]}
+{"QuestionId":57899789,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-12T04:34:01.857","AcceptedAnswerId":null,"OwnerUserId":744616.0,"Title":"Why Tokenizer is keeping track of more words than num_words?","Body":"<p>I have following code<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras\nfrom tensorflow.keras.preprocessing.text import Tokenizer\n\nsentences = [\n             'I love my dog',\n             'I love my cat',\n             'You love my dog',\n]\n\ntokenizer = Tokenizer(num_words=3)\ntokenizer.fit_on_texts(sentences)\nword_index = tokenizer.word_index\n\nprint(word_index)\n<\/code><\/pre>\n\n<blockquote>\n  <p>Output: {'love': 1, 'my': 2, 'i': 3, 'dog': 4, 'cat': 5, 'you': 6}<\/p>\n<\/blockquote>\n\n<p>Now how my code is keeping track of more than 3 unique frequent words? <\/p>\n\n<p>Am I missing anything here?<\/p>\n","answers":[{"AnswerId":"57902504","CreationDate":"2019-09-12T08:16:17.523","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>Looking at the <a href=\"https:\/\/github.com\/keras-team\/keras-preprocessing\/blob\/master\/keras_preprocessing\/text.py\" rel=\"nofollow noreferrer\">source code<\/a>, it seems as though indices are assigned to all words encountered. However, once you actually use the tokenizer to convert text to index sequences (e.g. using <code>texts_to_sequences<\/code>), all \"infrequent words\" will be replaced by the OOV token. Note that this will only be done if you actually specify an OOV token (which you haven't). Examples:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras\nfrom tensorflow.keras.preprocessing.text import Tokenizer\n\nsentences = [\n             'I love my dog',\n             'I love my cat',\n             'You love my dog dog',\n]\n\ntokenizer = Tokenizer(num_words=4, oov_token=None)\ntokenizer.fit_on_texts(sentences)\nword_index = tokenizer.word_index\n\nprint(word_index)\ntokenizer.texts_to_sequences([\"I love my cat\"])\n<\/code><\/pre>\n\n<blockquote>\n  <p>{'love': 1, 'you': 6, 'i': 4, 'dog': 3, 'my': 2, 'cat': 5}<br>\n  [[1, 2]]<\/p>\n<\/blockquote>\n\n<p>I slightly modified the text to break the tie between \"dog\" and \"I\" and increased the number of stored words by one (for whatever reason, specifying 4 actually only uses the three most common words...). You can see that OOV words (\"I\" and \"cat\") are simply left out of the text even though they have an index assigned.  <\/p>\n\n<p>Here's what happens if we specify an OOV token:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras\nfrom tensorflow.keras.preprocessing.text import Tokenizer\n\nsentences = [\n             'I love my dog',\n             'I love my cat',\n             'You love my dog dog',\n]\n\ntokenizer = Tokenizer(num_words=4, oov_token=\"oov\")\ntokenizer.fit_on_texts(sentences)\nword_index = tokenizer.word_index\n\nprint(word_index)\ntokenizer.texts_to_sequences([\"I love my cat\"])\n<\/code><\/pre>\n\n<blockquote>\n  <p>{'love': 2, 'you': 7, 'i': 5, 'dog': 4, 'my': 3, 'cat': 6, 'oov': 1}<br>\n  [[1, 2, 3, 1]]<\/p>\n<\/blockquote>\n\n<p>As you can see, index 1 is now reserved for the OOV token and infrequent words are assigned to this index when converted.<\/p>\n"}]}
+{"QuestionId":57900136,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-estimator>","CreationDate":"2019-09-12T05:18:21.340","AcceptedAnswerId":null,"OwnerUserId":1582366.0,"Title":"Issues with Tensor-flow distributed training (batch normalization)","Body":"<p>During TensorFlow distributed MultiWorkerMirroredStrategy training with I am facing errors during batch normalization. <\/p>\n\n<p>For model building, I am using batch normalization.<br>\nWith batch normalization, I receive the following error only on the chief node. <\/p>\n\n<p>If batch normalization is removed the code finishes to completion.<\/p>\n\n<pre><code>def _model_fn(features, labels, mode, params):\n\n\n    with tf.variable_scope('model', reuse=tf.AUTO_REUSE):\n\n        layer_1 = tf.keras.layers.LSTM(64,return_sequences=True)\n        LSTM_output = layer_1(features)\n        d = tf.keras.layers.Dense(32)(LSTM_output)\n        d = tf.keras.layers.Activation('relu')(d)   \n        d = tf.keras.layers.BatchNormalization()(\n          d, training= mode == tf.estimator.ModeKeys.TRAIN)  \n        d = tf.keras.layers.Dense(params['label_size'])(d)\n<\/code><\/pre>\n\n<pre><code>File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow\/python\/distribute\/estimator_training.py\", line 290, in train_and_evaluate\n    session_config=run_config.session_config)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow\/python\/distribute\/distribute_coordinator.py\", line 853, in run_distribute_coordinator\n    task_id, session_config, rpc_layer)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow\/python\/distribute\/distribute_coordinator.py\", line 360, in _run_single_worker\n    return worker_fn(strategy)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow\/python\/distribute\/estimator_training.py\", line 252, in _worker_fn\n    hooks=hooks)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow_estimator\/python\/estimator\/estimator.py\", line 367, in train\n    loss = self._train_model(input_fn, hooks, saving_listeners)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow_estimator\/python\/estimator\/estimator.py\", line 1156, in _train_model\n    return self._train_model_distributed(input_fn, hooks, saving_listeners)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow_estimator\/python\/estimator\/estimator.py\", line 1219, in _train_model_distributed\n    self._config._train_distribute, input_fn, hooks, saving_listeners)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow_estimator\/python\/estimator\/estimator.py\", line 1329, in _actual_train_model_distributed\n    saving_listeners)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow_estimator\/python\/estimator\/estimator.py\", line 1406, in _train_with_estimator_spec\n    estimator_spec, worker_hooks, saving_listeners)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow_estimator\/python\/estimator\/estimator.py\", line 1352, in _train_with_estimator_spec_distributed\n    _, loss = mon_sess.run([estimator_spec.train_op, estimator_spec.loss])\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 754, in run\n    run_metadata=run_metadata)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 1252, in run\n    run_metadata=run_metadata)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 1353, in run\n    raise six.reraise(*original_exc_info)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 1338, in run\n    return self._sess.run(*args, **kwargs)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 1411, in run\n    run_metadata=run_metadata)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/monitored_session.py\", line 1169, in run\n    return self._sess.run(*args, **kwargs)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 950, in run\n    run_metadata_ptr)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1173, in _run\n    feed_dict_tensor, options, run_metadata)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1350, in _do_run\n    run_metadata)\n  File \"\/home\/ab981s\/anaconda3\/envs\/py2tensorflow_nightly\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1370, in _do_call\n    raise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: From \/job:worker\/replica:0\/task:0:\nUpper bound check fail for input 10 from node geolocation_model\/gradients\/geolocation_model\/batch_normalization\/batchnorm\/add_1_grad\/Reshape_1 to node scoped_allocator_concat_1_15 input bounds = [0x7f035f94a200, 0x7f035f94a280] backing_tensor bounds = [0x7f035ee22900, 0x7f035eebe7c8]\n         [[{{node scoped_allocator_concat_1_15}}]]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57900190,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-09-12T05:24:34.193","AcceptedAnswerId":null,"OwnerUserId":4663377.0,"Title":"How to persist a process in python","Body":"<p>I have the following Pytorch code, what the code does is:<\/p>\n\n<ol>\n<li>Loads Pytorch object detection model into RAM using <code>net=CRAFT()<\/code><\/li>\n<li>using <code>net.eval()<\/code> it runs inference<\/li>\n<li>Using the for loop I am supplying images to run inference on one by one<\/li>\n<\/ol>\n\n<p>code:<\/p>\n\n<pre><code>def copyStateDict(state_dict):\n    if list(state_dict.keys())[0].startswith(\"module\"):\n        start_idx = 1\n    else:\n        start_idx = 0\n    new_state_dict = OrderedDict()\n    for k, v in state_dict.items():\n        name = \".\".join(k.split(\".\")[start_idx:])\n        new_state_dict[name] = v\n    return new_state_dict\n\n\nif __name__ == '__main__':\n    # load net\n    net = CRAFT()     # initialize\n\n    print('Loading weights from checkpoint (' + args.trained_model + ')')\n    if args.cuda:\n        net.load_state_dict(copyStateDict(torch.load(args.trained_model)))\n    else:\n        net.load_state_dict(copyStateDict(torch.load(args.trained_model, map_location='cpu')))\n\n    if args.cuda:\n        net = net.cuda()\n        net = torch.nn.DataParallel(net)\n        cudnn.benchmark = False\n\n    net.eval()\n\n    t = time.time()\n\n    # load data\n    for k, image_path in enumerate(image_list):\n        print(\"Test image {:d}\/{:d}: {:s}\".format(k+1, len(image_list), image_path), end='\\r')\n        image = imgproc.loadImage(image_path)\n\n        bboxes, polys, score_text = test_net(net, image, args.text_threshold, args.link_threshold, args.low_text, args.cuda, args.poly)\n\n        # save score text\n        #filename, file_ext = os.path.splitext(os.path.basename(image_path))\n        #mask_file = result_folder + \"\/res_\" + filename + '_mask.jpg'\n        #cv2.imwrite(mask_file, score_text)\n\n        file_utils.saveResult(image_path, image[:,:,::-1], polys, dirname=result_folder)\n\n    print(\"elapsed time : {}s\".format(time.time() - t))\n<\/code><\/pre>\n\n<p>There is a caveat in this code, especially the fact that it occupies <code>1.8 GB<\/code> of RAM each time I do step <strong>1<\/strong> and <strong>2<\/strong>. The actual inference on images takes much lesser RAM (40mb per image), but the majority of my memory consumption comes from loading the <code>ML<\/code> model into RAM. Is there a way in which I can keep loaded this model into RAM on a server and then just perform inferences on the images?<\/p>\n","answers":[{"AnswerId":"57900316","CreationDate":"2019-09-12T05:35:52.787","ParentId":null,"OwnerUserId":"31667","Title":null,"Body":"<p>You pretty much described a web server. You can create a trivial API around this with something like <a href=\"https:\/\/palletsprojects.com\/p\/flask\/\" rel=\"nofollow noreferrer\">flask<\/a> which would load the model at startup and receive either images or paths in the web request, then return a response. Run it with <code>uwsgi<\/code> or something similar to have more than one worker available at a time (just make sure the model is initialised once, not per-worker).<\/p>\n\n<p>Or you could go more fancy and use one of the many RPC libraries to create a similar kind of local server.<\/p>\n"}]}
+{"QuestionId":57900254,"AnswerCount":1,"Tags":"<keras><deep-learning>","CreationDate":"2019-09-12T05:30:12.303","AcceptedAnswerId":null,"OwnerUserId":8176285.0,"Title":"Pass text features along with image data to a CNN model","Body":"<p>I am trying to approach a multi label image classification problem,for which i have image data but i also have some other features like gender etc, but the issue is that i will get this information during testing, in other words during testing only the image information will be provided.<\/p>\n\n<p>My question is how can i use these extra features to help my image model which is a convolution Neural Network even though i wont have this info during testing?<\/p>\n\n<p>Any advice will be helpful.Thanks in advance.<\/p>\n","answers":[{"AnswerId":"57900576","CreationDate":"2019-09-12T06:02:55.803","ParentId":null,"OwnerUserId":"8112138","Title":null,"Body":"<p>This is a really open ended question. I can give you some general guidelines on how this can work. <\/p>\n\n<p><code>keras<\/code> model API supports multiple inputs as well as <a href=\"https:\/\/keras.io\/layers\/merge\/\" rel=\"nofollow noreferrer\"><code>merge<\/code><\/a> layers. For example you can have something like this:<\/p>\n\n<pre><code>from keras.layers import Input\nfrom keras.models import Model\n\nimage = Input(...)\ntext = Input(...)\n\n... # apply layers onto image and text\n\nfrom keras.layers.merge import Concatenate\ncombined = Concatenate()([image, text])\n\n... # apply layers onto combined\n\nmodel = Model([image, text], [combined])\n<\/code><\/pre>\n\n<p>This way you can have a <code>model<\/code> that takes multiple inputs that can make use of all of your data sources. <code>keras<\/code> has tools to combine your different inputs to produce one output. The part where this becomes open ended is the architecture. <\/p>\n\n<p>Right now you should probably pass <code>image<\/code> through a <code>CNN<\/code>, and then <code>merge<\/code> the output with <code>text<\/code>. You have to tweak the exact specifications, such as how you handle each input, your <code>merge<\/code> method, and how you handle the combined output. <\/p>\n\n<p>A good example of <code>merge<\/code> being used is <a href=\"https:\/\/github.com\/eriklindernoren\/Keras-GAN\/blob\/master\/cgan\/cgan.py\" rel=\"nofollow noreferrer\">here<\/a>, where a GAN is given latent noise in the form of an image but also a label to determine what kind of image it should generate. Both the <code>discriminator<\/code> and the <code>generator<\/code> make use of the <code>multiply<\/code> merge layer to combine their inputs.<\/p>\n"}]}
+{"QuestionId":57900546,"AnswerCount":1,"Tags":"<java><tensorflow>","CreationDate":"2019-09-12T06:00:44.267","AcceptedAnswerId":57902454.0,"OwnerUserId":10784798.0,"Title":"How can I load a Tensorflow SavedModel in Java with 'predict' Sgnature Def?","Body":"<p>I'm training a Tensorflow Estimator and use <code>export_saved_model<\/code> to save the model in SavedModel format. Now I want to load this model with Tensorflow Java API (I don't want to use an model-server, I need to directly load it in Java). \nNow the problem is that <code>Estimator.export_saved_model<\/code> only exports the 'predict' signature_def, while the <code>SavedModelBundle<\/code> in Java seems to only support models with a 'serving_default' signature def.\nSo the question is: is there a way to make <code>Estimator.export_saved_model<\/code> to include the 'serving_default' signature def? Or is it possible to load the model with the 'predict' signature def in java? or are there any other options I could try?<\/p>\n\n<p>Here is the code for exporting the model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>feature_cols = [\n        tf.feature_column.numeric_column('numeric_feature'),\n        tf.feature_column.indicator_column( tf.feature_column.categorical_column_with_vocabulary_list('categorial_text_feature', vocabulary_list=['WORD1', 'WORD1']))\n]\n\nestimator = tf.estimator.LinearRegressor(\n    feature_columns=feature_cols,\n    model_dir=model_dir,\n    label_dimension=1)\n\n    estimator.train(input_fn=input_fn)\n\nserving_input_receiver_fn = tf.estimator.export.build_raw_serving_input_receiver_fn({\n        'numeric_feature': tf.placeholder(tf.float32, shape=(None,)),\n        'categorial_text_feature': tf.placeholder(tf.string, shape=(None,))\n})\nestimator.export_saved_model(\n    export_dir_base=model_dir,\n    serving_input_receiver_fn=serving_input_receiver_fn)\n<\/code><\/pre>\n\n<p>If I inspect the model with <code>saved_model_cli show --tag_set serve<\/code> I get:<\/p>\n\n<pre><code>The given SavedModel MetaGraphDef contains SignatureDefs with the following keys:\nSignatureDef key: \"predict\"\n<\/code><\/pre>\n\n<p>And with <code>saved_model_cli show --tag_set serve --signature_def predict<\/code>:<\/p>\n\n<pre><code>The given SavedModel SignatureDef contains the following input(s):\n  inputs['numeric_feature'] tensor_info:\n      dtype: DT_FLOAT\n      shape: (-1)\n      name: Placeholder:0\n  inputs['categorial_text_feature'] tensor_info:\n      dtype: DT_STRING\n      shape: (-1)\n      name: Placeholder_1:0\nThe given SavedModel SignatureDef contains the following output(s):\n  outputs['predictions'] tensor_info:\n      dtype: DT_FLOAT\n      shape: (-1)\n      name: linear\/linear_model\/linear_model\/linear_model\/weighted_sum:0\nMethod name is: tensorflow\/serving\/predict\n\n<\/code><\/pre>\n\n<p>And with <code>saved_model_cli show --tag_set serve --signature_def serving_default<\/code>:<\/p>\n\n<pre><code>The given SavedModel SignatureDef contains the following input(s):\nThe given SavedModel SignatureDef contains the following output(s):\nMethod name is: \n\n<\/code><\/pre>\n\n<p>Now when I load the model in Java and feed some values:<\/p>\n\n<pre><code>SavedModelBundle model = SavedModelBundle.load(modelPath, \"serve\");\n\nTensor&lt;?&gt; numericTensor = Tensors.create(new float[] { 10.3 });\nTensor&lt;?&gt; stringTensor = Tensors.create(new byte[][] { \"WORD1\".getBytes() });\nTensor&lt;?&gt; output = model\n                .session()\n                .runner()\n                .feed(\"numeric_feature\", numericTensor)\n                .feed(\"categorial_text_feature\", stringTensor)\n                .fetch(\"predictions\")\n                .run()\n                .get(0);\n<\/code><\/pre>\n\n<p>This results to following error:<\/p>\n\n<pre><code>java.lang.IllegalArgumentException: No Operation named [numeric_feature] in the Graph\n    at org.tensorflow.Session$Runner.operationByName(Session.java:372)\n    at org.tensorflow.Session$Runner.parseOutput(Session.java:381)\n    at org.tensorflow.Session$Runner.feed(Session.java:131)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57902454","CreationDate":"2019-09-12T08:12:58.490","ParentId":null,"OwnerUserId":"10784798","Title":null,"Body":"<p>Found a (not perfect, but simple) workaround:<\/p>\n\n<p>I just exported the model with <code>as_text=True<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>estimator.export_saved_model(\n        export_dir_base=model_dir,\n        serving_input_receiver_fn=serving_input_receiver_fn,\n        as_text=True)\n<\/code><\/pre>\n\n<p>and then just change the .pbtxt file by hand so that the signature def is called 'serving_default'<\/p>\n"}]}
+{"QuestionId":57901482,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><gpu>","CreationDate":"2019-09-12T07:12:46.563","AcceptedAnswerId":57902663.0,"OwnerUserId":11805922.0,"Title":"How to specify which GPU to use when running tensorflow?","Body":"<p>We have a DGX-1 in Lab.<\/p>\n\n<p>I see many tasks are running on different GPU.<\/p>\n\n<p>For MLperf docker application, I can use NV_GPU=x to assign which GPU to use.<\/p>\n\n<p>However, I have a python Keras\/TensorFlow code, I used this same way, the loading doesn't go to the specified GPU.<\/p>\n","answers":[{"AnswerId":"57902663","CreationDate":"2019-09-12T08:25:41.103","ParentId":null,"OwnerUserId":"11651301","Title":null,"Body":"<p>You could use <code>CUDA_VISIBLE_DEVICES<\/code> to specify the GPU to be used by your model:<\/p>\n\n<pre><code>import os \nos.environ[\"CUDA_VISIBLE_DEVICES\"] = 0,1  #Will assign GPUs 0 and 1 to the model\n<\/code><\/pre>\n"}]}
+{"QuestionId":57901517,"AnswerCount":0,"Tags":"<r><tensorflow><keras>","CreationDate":"2019-09-12T07:15:25.420","AcceptedAnswerId":null,"OwnerUserId":12056842.0,"Title":"How to obtain the weights from a keras model?","Body":"<p>I want to get the weights for each variable of my keras model in R.<\/p>\n\n<p>For example we have the following model:<\/p>\n\n<p><code>y= a + bx1 +cx2 + dx3<\/code><\/p>\n\n<p>I would like to retrieve <code>b<\/code>, <code>c<\/code> and <code>d<\/code> but when using the get_weights() command in R I get a matrix that is difficult to interpret.<\/p>\n\n<p>Any help on the matter would be appreciated.<\/p>\n","answers":[]}
+{"QuestionId":57901522,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras><deep-learning><boolean>","CreationDate":"2019-09-12T07:15:49.603","AcceptedAnswerId":null,"OwnerUserId":10393794.0,"Title":"Mask RCNN: IndexError: boolean index did not match indexed array","Body":"<p>I am working on a project, and I need to train Mask RCNN regarding to SpaceNet Dataset. <\/p>\n\n<p>So, When I try to train my model, lots of warnings and errors show up. <\/p>\n\n<p>The Error message is:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nIndexError                                Traceback (most recent call last)\n&lt;ipython-input-11-a73fb1f7a961&gt; in &lt;module&gt;\n      8             learning_rate=config.LEARNING_RATE,\n      9             epochs=10,\n---&gt; 10             layers='heads')\n     11 \n     12 # Training - Stage 2\n\n~\\Desktop\\SpaceNet_MaskRCNN\\mrcnn\\model.py in train(self, train_dataset, val_dataset, learning_rate, epochs, layers, augmentation, custom_callbacks, no_augmentation_sources)\n   2372             max_queue_size=100,\n   2373             workers=workers,\n-&gt; 2374             use_multiprocessing=True,\n   2375         )\n   2376         self.epoch = max(self.epoch, epochs)\n\n~\\Anaconda3\\envs\\MaskRCNN\\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\\MaskRCNN\\lib\\site-packages\\keras\\engine\\training.py in fit_generator(self, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, validation_freq, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch)\n   1656             use_multiprocessing=use_multiprocessing,\n   1657             shuffle=shuffle,\n-&gt; 1658             initial_epoch=initial_epoch)\n   1659 \n   1660     @interfaces.legacy_generator_methods_support\n\n~\\Anaconda3\\envs\\MaskRCNN\\lib\\site-packages\\keras\\engine\\training_generator.py in fit_generator(model, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, validation_freq, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch)\n    179             batch_index = 0\n    180             while steps_done &lt; steps_per_epoch:\n--&gt; 181                 generator_output = next(output_generator)\n    182 \n    183                 if not hasattr(generator_output, '__len__'):\n\n~\\Desktop\\SpaceNet_MaskRCNN\\mrcnn\\model.py in data_generator(dataset, config, shuffle, augment, augmentation, random_rois, batch_size, detection_targets, no_augmentation_sources)\n   1707                     load_image_gt(dataset, config, image_id, augment=augment,\n   1708                                 augmentation=augmentation,\n-&gt; 1709                                 use_mini_mask=config.USE_MINI_MASK)\n   1710 \n   1711             # Skip images that have no instances. This can happen in cases\n\n~\\Desktop\\SpaceNet_MaskRCNN\\mrcnn\\model.py in load_image_gt(dataset, config, image_id, augment, augmentation, use_mini_mask)\n   1263     _idx = np.sum(mask, axis=(0, 1)) &gt; 0\n   1264     mask = mask[:, :, _idx]\n-&gt; 1265     class_ids = class_ids[_idx]\n   1266     # Bounding boxes. Note that some boxes might be all zeros\n   1267     # if the corresponding mask got cropped out.\n\nIndexError: boolean index did not match indexed array along dimension 0; dimension is 1 but corresponding boolean dimension is 650\n\n<\/code><\/pre>\n\n<p>The warnings are:<\/p>\n\n<blockquote>\n  <p>ERROR:root:Error processing image {'id': 219, 'source': 'yapi', 'path': None, 'width': 650, 'height': 650}\n  Traceback (most recent call last):\n    File \"C:\\Users\\MUSTAFAAKTAS\\Desktop\\SpaceNet_MaskRCNN\\mrcnn\\model.py\", line 1710, in data_generator\n      use_mini_mask=config.USE_MINI_MASK)\n    File \"C:\\Users\\MUSTAFAAKTAS\\Desktop\\SpaceNet_MaskRCNN\\mrcnn\\model.py\", line 1266, in load_image_gt\n      class_ids = class_ids[_idx]\n  IndexError: boolean index did not match indexed array along dimension 0; dimension is 1 but corresponding boolean dimension is 650<\/p>\n<\/blockquote>\n\n<p>--<\/p>\n\n<blockquote>\n  <p>ERROR:root:Error processing image {'id': 448, 'source': 'yapi', 'path': None, 'width': 650, 'height': 650}\n  Traceback (most recent call last):\n    File \"C:\\Users\\MUSTAFAAKTAS\\Desktop\\SpaceNet_MaskRCNN\\mrcnn\\model.py\", line 1710, in data_generator\n      use_mini_mask=config.USE_MINI_MASK)\n    File \"C:\\Users\\MUSTAFAAKTAS\\Desktop\\SpaceNet_MaskRCNN\\mrcnn\\model.py\", line 1266, in load_image_gt\n      class_ids = class_ids[_idx]\n  IndexError: boolean index did not match indexed array along dimension 0; dimension is 1 but corresponding boolean dimension is 650<\/p>\n<\/blockquote>\n\n<p>--<\/p>\n\n<blockquote>\n  <p>Also there are warnings for image_id: 219-348-444-448-3986-3023<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":57901526,"AnswerCount":2,"Tags":"<tensorflow><scikit-learn><nlp><nltk><text-parsing>","CreationDate":"2019-09-12T07:15:58.060","AcceptedAnswerId":null,"OwnerUserId":10954152.0,"Title":"Classify intent of random utterance of chat bot from training data and give different graphical visualization using random forest?","Body":"<p>I am creating a nlp model to detect the intent from the provided utterance from a excel file which I am using for training having 2 columns like shown below:<\/p>\n\n<pre><code>Utterence                                       Intent\nhi can I have an Apple Watch                   service\nhow much I will be paying monthly              service\nyou still around                              YOU_THERE\nare you still there                           YOU_THERE\nyou there                                     YOU_THERE\nSpeak to me if you are there.                 YOU_THERE\nyou around                                    YOU_THERE\n<\/code><\/pre>\n\n<p>There are like around 3000 utterances in the training files and many intents.<\/p>\n\n<p>I trained my model using scikit learn module and my code looks like this.<\/p>\n\n<pre><code>from sklearn.feature_extraction.text import TfidfVectorizer\nimport pandas as pd\nimport numpy as np\nimport re\ndef preprocessing(userQuery):\n    letters_only = re.sub(\"[^a-zA-Z\\\\d]\", \" \", userQuery)\n    words = letters_only.lower().split()\n    return( \" \".join(words ))\n#read utterance data from a xlsx file\ntrain = pd.read_excel('training.xlsx')\nquery_features = train['Utterence']\n#create tfidf\ntfidf_vectorizer = TfidfVectorizer(ngram_range=(1, 1))\nnew_query = [preprocessing(query) for query in query_features]\nfeatures = tfidf_vectorizer.fit_transform(new_query).toarray()\n#create random forest classification model\nmodel = RandomForestClassifier()\nmodel.fit(features, train['Intent'])\n#intent prediction on user query\nuserQuery = \"I want apple watch\"\nuserQueryList=[]\nuserQueryList.append(preprocessing(userQuery))\nutfidf = tfidf_vectorizer.transform(userQueryList)\nprint(\" prediction: \", model.predict(utfidf))\n<\/code><\/pre>\n\n<p>The one of problem for me here is for example: when i run for utterance <code>I want apple watch<\/code> it gives predicted intent as <code>you_there<\/code> instead of <code>service<\/code> as shown below(confirmation on training snapshot above):<\/p>\n\n<pre><code>C:\\U\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\sklearn\\ensemble\\forest.py:246: FutureWarning: The default value of n_estimators will change from 10 in version 0.20 to 100 in 0.22.\n  \"10 in version 0.20 to 100 in 0.22.\", FutureWarning)\n prediction:  ['YOU_THERE']\n<\/code><\/pre>\n\n<p>Please help me how should i train my  model and what changes should I make to fix such issues and how i can check accuracy? Also I want to see graphical visualization and ROC curve how it can achieved using random forest. I am not very verse in NLP any help would be appreciated.<\/p>\n","answers":[{"AnswerId":"57987670","CreationDate":"2019-09-18T07:52:15.033","ParentId":null,"OwnerUserId":"9607338","Title":null,"Body":"<p>Let's address your first issue:<\/p>\n\n<blockquote>\n  <p>how should i train my model and what changes should I make to fix such issues<\/p>\n<\/blockquote>\n\n<p>Below I'm using word2vec approach which rather than just converting the utterances to vectors using TFIDF approach (losing the semantic information contained in that particular sentence), it maintains the semantic info.<\/p>\n\n<p>To understand more about word2vec, refer this blog : <\/p>\n\n<p>[1]<a href=\"https:\/\/www.analyticsvidhya.com\/blog\/2017\/06\/word-embeddings-count-word2veec\/\" rel=\"nofollow noreferrer\">https:\/\/www.analyticsvidhya.com\/blog\/2017\/06\/word-embeddings-count-word2veec\/<\/a><\/p>\n\n<p>Below is the code for predicted the intent using word2vec approach (Note - It's same as your code, just instead of using TFIDFVectorizer, I'm using word2vec to obtain the vectors. Also the code is divided into different functions to get a good overview of logics that will be evident by there names).<\/p>\n\n<pre><code>import pandas as pd\nimport numpy as np\nfrom gensim.models import Word2Vec\nfrom sklearn import preprocessing\nfrom sklearn.ensemble import RandomForestClassifier\nfrom sklearn.svm import SVC\n\ndef preprocess_lower(token):\n    #utility for preprocessing\n    return token.lower()\n\ndef load_data(file_name):\n    # load a csv file in memory\n    return pd.read_csv(file_name)\n\n\ndef process_training_data(training_data):\n    # process the training data and split it between independent and dependent variables\n    training_sentences = [list(map(preprocess_lower,sentence.split(\" \"))) for sentence in list(training_data.Utterence.values)]\n    target_class = training_data.Intent.values\n    label_encoded_Y = preprocessing.LabelEncoder().fit_transform(list(target_class))\n    return target_class, training_sentences, label_encoded_Y\n\n\ndef process_user_query(training_data):\n    # process the training data and split it between independent and dependent variables\n    training_sentences = [list(map(preprocess_lower,sentence.split(\" \"))) for sentence in training_data]\n    return training_sentences\n\ndef train_word2vec_model(train_sentences_list):\n    # training word2vec on sentences list (inputted by user)\n    model = Word2Vec(train_sentences_list, size=100, window=4, min_count=1, workers=4)\n    return model\n\ndef convert_training_data_vectors(model, train_sentences_list):\n    #get the sentences average vector\n    training_sectences_vector = list()\n    for sentence in train_sentences_list:\n        sentence_vetor = [list(model.wv[token]) for token in sentence if token in model.wv.vocab ]\n        training_sectences_vector.append(list(np.mean(sentence_vetor, axis=0)))\n    return training_sectences_vector\n\ndef training_rf_prediction_model(training_data_vectors, label_encoded_Y):\nfrom sklearn.metrics import accuracy_score\nfrom sklearn.model_selection import train_test_split\n# training model on user inputted data \nrf_model = RandomForestClassifier()\n# here use the split function and divide the data into training and testing \nx_train,x_test,y_train,y_test=train_test_split(training_data_vectors,label_encoded_Y,\ntrain_size=0.8,test_size=0.2)\nrf_model.fit(x_train, y_train)\ny_pred = rf_model.predict(x_test)\nprint(accuracy_score(y_test, y_pred))\nreturn rf_model\n\ndef training_svm_prediction_model(training_data_vectors, label_encoded_Y):\n    svm_model = SVC(gamma='auto')\n    svm_model.fit(training_data_vectors, label_encoded_Y)\n    return svm_model\n\ndef process_data_flow(file_name):\n    training_data = load_data(file_name)\n    target_class, training_sentences, label_encoded_Y = process_training_data(training_data)\n    word2vec_model = train_word2vec_model(train_sentences_list=training_sentences)\n    training_data_vectors = convert_training_data_vectors(word2vec_model, train_sentences_list=training_sentences)\n    prediction_model = training_rf_prediction_model(training_data_vectors, label_encoded_Y)\n    #intent prediction on user query\n    userQuery = [\"i want apple watch\"]\n    user_query_vectors = convert_training_data_vectors(word2vec_model, process_user_query(userQuery))\n    predicted_class = prediction_model.predict(user_query_vectors)[0]\n    predicted_intent = target_class[list(label_encoded_Y).index(predicted_class)]\n    return predicted_intent\n\nprint(\"Predicted class: \", process_data_flow(\"sample_intent_data.csv\"))\n<\/code><\/pre>\n\n<p>sample data file is in csv format, you just need to format and paste the data in below format :<\/p>\n\n<pre><code>#sample_input_data.csv\nUtterence,Intent\nhi can I have an Apple Watch,service\nhow much I will be paying monthly,service\nyou still around,YOU_THERE\nare you still there,YOU_THERE\nyou there,YOU_THERE\nSpeak to me if you are there,YOU_THERE\nyou around,YOU_THERE\n<\/code><\/pre>\n\n<p>Also note, your training data should contain good amount of training utterances for each intents for the approach to work.<\/p>\n\n<p>For accuracy, you can use below approach:<\/p>\n\n<p>Divide the data into training and testing (mention the split ratio) :<\/p>\n\n<pre><code>x_train,x_test,y_train,y_test=train_test_split(training_vectors,label_encoded_Y,\ntrain_size=0.8,\ntest_size=0.2)\n<\/code><\/pre>\n\n<p>And after training the model, use predict function on x_test to get the predictions. Now just match the prediction for testing data from the model and actual from the data set and you will be able to easily determine the accuracy.<\/p>\n\n<p>Edit: Added the accuracy score calculation while predicting.<\/p>\n"},{"AnswerId":"57901817","CreationDate":"2019-09-12T07:33:40.610","ParentId":null,"OwnerUserId":"10026409","Title":null,"Body":"<p>You are using word bags approach which does not perform well on sequence data.\nFor your problem, sequential is material to classification.\nI would suggest to you that use LSTM (performs better on sequence data)<\/p>\n"}]}
+{"QuestionId":57901603,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2019-09-12T07:20:32.090","AcceptedAnswerId":58220795.0,"OwnerUserId":3120842.0,"Title":"Keras issue with top k categorical accuracy with 3d tensors","Body":"<p>I'm having trouble with the top k categorical accuracy metric in keras. I'm training on genomic data, with about 10,000 individuals, each individual has a stretch of 4,000 nucleotides that can be one of four bases, so they are one hot encoded. The output is similar, but instead of four bases, there are three categories.<\/p>\n\n<p>input shape (10000, 4000, 4)\noutput shape (10000, 4000, 3)<\/p>\n\n<p>I try to use top k categorical accuracy, but I get:\nValueError: Shape must be rank 2 but is rank 3 for 'metrics\/top_k_categorical_accuracy\/in_top_k\/InTopKV2' (op: 'InTopKV2') with input shapes: [?,?,3], [?,?], [].<\/p>\n\n<p>Evidently keras doesn't like the 3d tensor. How do I get around this?<\/p>\n","answers":[{"AnswerId":"58220795","CreationDate":"2019-10-03T14:01:11.597","ParentId":null,"OwnerUserId":"9084663","Title":null,"Body":"<p><a href=\"https:\/\/stackoverflow.com\/questions\/58100322\/top-k-categorical-accuracy-for-time-distributed-lstm-results\/58102380#58102380\">This<\/a> might help you.<\/p>\n\n<p>Simply put, top k categorical accuracy was not defined to be dimensionally flexible.<\/p>\n\n<p>You need to transform the data to fit.<\/p>\n\n<pre><code>import keras.backend as K \ndef my_top_k(true, pred, num):\n    true = K.reshape(true, (-1, features_num))   \n    pred = K.reshape(pred, (-1, features_num))\n    return top_k_categorical_accuracy(true, pred, k=num)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57901697,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><multiclass-classification>","CreationDate":"2019-09-12T07:25:42.050","AcceptedAnswerId":null,"OwnerUserId":4352170.0,"Title":"Keras model evaluation accuracy unchanged, and designing model","Body":"<p>I'm trying to design a CNN in Keras to classify small images of emojis in other images. Below is an example of one of the 13 classes. All images are the same size and all the emojis are of the same size as well. I would think that one should rather easily be able to achieve VERY high accuracy when classifying, as emojis from one class is exactly the same! My intuition told me that if an emoji is 50x50 I could create a convolutional layer of the same size to match one type of emoji. My supervisor did not think that was feasible however. Anyway, my problem is that, no matter how I design my model, I always get the same validation accuracy for each epoch, which corresponds to 1\/13 (or simply guessing that each emoji belongs to the same class).<\/p>\n\n<p>My model looks like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential()\nmodel.add(Conv2D(16, kernel_size=3, activation=\"relu\", input_shape=IMG_SIZE))\nmodel.add(Dropout(0.5))\nmodel.add(Conv2D(32, kernel_size=3, activation=\"relu\"))\nmodel.add(Conv2D(64, kernel_size=3, activation=\"relu\"))\nmodel.add(Conv2D(128, kernel_size=3, activation=\"relu\"))\n#model.add(Conv2D(256, kernel_size=3, activation=\"relu\"))\nmodel.add(Dropout(0.5))\nmodel.add(Flatten()) \n#model.add(Dense(256, activation=\"relu\"))\nmodel.add(Dense(128, activation=\"relu\"))\nmodel.add(Dense(64, activation=\"relu\"))\nmodel.add(Dense(NUM_CLASSES, activation='softmax', name=\"Output\")) \n<\/code><\/pre>\n\n<p>And I train it like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># ------------------ Compile and train ---------------\nsgd = optimizers.SGD(lr=0.001, decay=1e-6, momentum=0.9, nesterov=True)\nrms = optimizers.RMSprop(lr=0.004, rho=0.9, epsilon=None, decay=0.0)\n\nmodel.compile(optimizer=sgd, loss='categorical_crossentropy', metrics=[\"accuracy\"])  # TODO Read more about this\ntrain_hist = model.fit_generator(\n    train_generator,\n    steps_per_epoch=train_generator.n \/\/ BATCH_SIZE,\n    validation_steps=validation_generator.n \/\/ BATCH_SIZE,  # TODO que?\n    epochs=EPOCHS,\n    validation_data=validation_generator,\n    #callbacks=[EarlyStopping(patience=3, restore_best_weights=True)]\n)\n<\/code><\/pre>\n\n<p>Even with this model, that has over 200 million parameters, I get exactly 0.0773 in validation accuracy for each epoch:<\/p>\n\n<pre><code>Epoch 1\/10\n56\/56 [==============================] - 21s 379ms\/step - loss: 14.9091 - acc: 0.0737 - val_loss: 14.8719 - val_acc: 0.0773\nEpoch 2\/10\n56\/56 [==============================] - 6s 108ms\/step - loss: 14.9308 - acc: 0.0737 - val_loss: 14.8719 - val_acc: 0.0773\nEpoch 3\/10\n56\/56 [==============================] - 6s 108ms\/step - loss: 14.7869 - acc: 0.0826 - val_loss: 14.8719 - val_acc: 0.0773\nEpoch 4\/10\n56\/56 [==============================] - 6s 108ms\/step - loss: 14.8948 - acc: 0.0759 - val_loss: 14.8719 - val_acc: 0.0773\nEpoch 5\/10\n56\/56 [==============================] - 6s 109ms\/step - loss: 14.8897 - acc: 0.0762 - val_loss: 14.8719 - val_acc: 0.0773\nEpoch 6\/10\n56\/56 [==============================] - 6s 109ms\/step - loss: 14.8178 - acc: 0.0807 - val_loss: 14.8719 - val_acc: 0.0773\nEpoch 7\/10\n56\/56 [==============================] - 6s 108ms\/step - loss: 15.0747 - acc: 0.0647 - val_loss: 14.8719 - val_acc: 0.0773\nEpoch 8\/10\n56\/56 [==============================] - 6s 108ms\/step - loss: 14.7509 - acc: 0.0848 - val_loss: 14.8719 - val_acc: 0.0773\nEpoch 9\/10\n56\/56 [==============================] - 6s 108ms\/step - loss: 14.8948 - acc: 0.0759 - val_loss: 14.8719 - val_acc: 0.0773\nEpoch 10\/10\n56\/56 [==============================] - 6s 108ms\/step - loss: 14.8228 - acc: 0.0804 - val_loss: 14.8719 - val_acc: 0.0773\n<\/code><\/pre>\n\n<p>Because its not learning anything, I'm starting to think that it's not my models fault, but maybe the dataset or how I train it. I have tried training with \"adam\" as well but get the same result. I try to change the input size of the images but still, same result. Below is a sample from my dataset. Do you guys have any ideas what could be wrong?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ZaV16.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ZaV16.jpg\" alt=\"Same from dataset\"><\/a><\/p>\n","answers":[{"AnswerId":"57910675","CreationDate":"2019-09-12T16:06:18.640","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>I think the main issue currently is that your model has way too many parameters relative to how few samples you have for training. For image classification nowadays, you generally want to just have conv layers, a Global<em>Something<\/em>Pooling layer, and then a single Dense layer for your outputs. You just need to make sure that your conv section ends up with a large enough receptive field to be able to find all the features you need.<\/p>\n\n<p>The first thing to think about is making sure that you have a large enough receptive field (<a href=\"https:\/\/medium.com\/mlreview\/a-guide-to-receptive-field-arithmetic-for-convolutional-neural-networks-e0f514068807\" rel=\"nofollow noreferrer\">further reading about that here<\/a>). There are three main ways to achieve that: pooling, stride >= 2, and\/or dilation >= 2. Because you have only 13 \"features\" that you want to identify, and all of them will always be pixel-perfect, I'm thinking dilation will be the way to go so that the model can easily \"overfit\" on those 13 \"features\". If we use 4 conv layers with dilations of 1, 2, 4, and 8, respectively, then we'll end up with a receptive field of 31. This should be enough to easily recognize 50-pixel emoji.<\/p>\n\n<p>Next, how many filters should each layer have? Normally, you start with a few filters, and increase as you go through the model, as you are doing here. However, we want to \"overfit\" on specific features, so we should probably increase the amounts in the earlier layers. Just to make it easy, let's give all layers 64 filters.<\/p>\n\n<p>Last, how do we convert this all to a single prediction? Rather than using a dense layer, which would use a <em>ton<\/em> of parameters and would not be translation invariant, people nowadays use GlobalAveragePooling or GlobalMaxPooling. GlobalAveragePooling is more common, because it's good for helping to find combinations of many features. However, we just want to find 13 or so exact features here, so GlobalMaxPooling might work even better. Then a single dense layer after that will be enough to get a prediction. Because we're using <em>Global<\/em>MaxPooling, it doesn't need to be flattened first\u2014global pooling already does that for us.<\/p>\n\n<p>Resulting model:<\/p>\n\n<pre><code>Conv2D(64, kernel_size=3, activation=\"relu\", input_shape=IMG_SIZE)\nConv2D(64, kernel_size=3, dilation_rate=2, activation=\"relu\")\nConv2D(64, kernel_size=3, dilation_rate=4, activation=\"relu\")\nConv2D(64, kernel_size=3, dilation_rate=8, activation=\"relu\")\nGlobalMaxPooling()\nDense(NUM_CLASSES, activation='softmax', name=\"Output\")\n<\/code><\/pre>\n\n<p>Try that. You'll probably also want to add BatchNormalization layers in there after every layers except the last. Once you get decent training accuracy, check if it's overfitting. If it is (which is likely), try these steps:<\/p>\n\n<ul>\n<li>batch norm if you haven't already<\/li>\n<li>weight decay on all conv and dense layers<\/li>\n<li>SpatialDropout2D after the conv layers and regular Dropout after the pooling layer<\/li>\n<li>augment your data with an ImageDataGenerator<\/li>\n<\/ul>\n\n<p>Designing nets like this is almost more of an art than a science, so change stuff around if you think you should. Eventually, you will get an intuition for what is more or less likely to work in any given situation.<\/p>\n"}]}
+{"QuestionId":57901930,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><distributed-computing><parameter-server>","CreationDate":"2019-09-12T07:40:30.073","AcceptedAnswerId":null,"OwnerUserId":2570677.0,"Title":"Tenseflow Momentum Optimizer on a distributed cluster: where is the momentum saved?","Body":"<p>I am working on training a model on a distributed cluster using Tensorflow.<\/p>\n\n<p>I'm using <code>tf.train.MomentumOptimizer<\/code> as the optimizer.\nAnd I'm using a parameter server to maintain a global state on a PySpark cluster.<\/p>\n\n<p>I want to know if the accumulated momentum is saved in the parameter server (it is global for all workers) or is the momentum local for each worker?\nI couldn't find a reference to it in the documentation.<\/p>\n\n<p>Any lead on where to find the answer is welcome.<\/p>\n","answers":[]}
+{"QuestionId":57901933,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-12T07:40:53.903","AcceptedAnswerId":null,"OwnerUserId":6485370.0,"Title":"Want to check Intermediate Operations inside Keras Layer","Body":"<p>I am facing floating point resolution loss during convolution operation while porting the code on my embedded processor which supports only half precision, so I want to test the intermediate operations that are performed layer by layer in my Keras based model which is performing good while on Full precision on my desktop.<\/p>\n\n<p>In the following snippet of code I want to compute the <strong>1DConv<\/strong> on the 1500x3 shaped input data. The kernel size is 10 and Kernel shape is (10x3x16).<\/p>\n\n<p>To compute the 1D-Convolution, Keras does the Expand Dimensions on input shape and add one more dimension to it, which becomes suitable for 2D Convolution operation.<\/p>\n\n<p>Then series of operations are called e.g. <strong>Conv2D<\/strong> followed by <strong>Squeeze<\/strong> and finally BiasAdd. \nFinally the output of the Conv1D layer is pushed in\n<strong>conv1d_20\/Elu<\/strong> layer.<\/p>\n\n<p>Please find the picture attached for full description of operations involved.<\/p>\n\n<p>Now, I want to test the output much before the actual output of a Layer is produced.<\/p>\n\n<p>Please see the below code:<\/p>\n\n<pre><code>Input_sequence = keras.layers.Input(shape=(1500,3))\nencoder_conv1 = keras.layers.Conv1D(filters=16, kernel_size=10, padding='same', activation=tf.nn.elu)(Input_sequence)\n<\/code><\/pre>\n\n<p><strong>The Model summary shows:<\/strong><\/p>\n\n<pre><code>Model: \"model_5\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_5 (InputLayer)         [(None, 1500, 3)]         0         \n_________________________________________________________________\nconv1d_20 (Conv1D)           (None, 1500, 16)          496  \n<\/code><\/pre>\n\n<p>I want to define the model output at conv1d_20\/Conv2D but it gives me error. But the below is accepted at compilation.<\/p>\n\n<pre><code>encoder = keras.Model(inputs=autoencoder.input, outputs=autoencoder.get_layer('conv1d_20').output)\nencoder.get_output_at(0)\n<\/code><\/pre>\n\n<p><strong>It outputs<\/strong><\/p>\n\n<pre><code>&lt;tf.Tensor 'conv1d_20\/Elu:0' shape=(?, 1500, 16) dtype=float32&gt;\n<\/code><\/pre>\n\n<p>I want to test the output of Conv2D operation but it produces the output of conv1d_20\/Elu. <\/p>\n\n<p>How can I do this test. Please help me.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/VsIqP.png\" rel=\"nofollow noreferrer\">Conv1D operation<\/a><\/p>\n","answers":[{"AnswerId":"57906444","CreationDate":"2019-09-12T12:11:05.430","ParentId":null,"OwnerUserId":"5597718","Title":null,"Body":"<p>You can disable the bias(<code>use_bias=False<\/code>) and activation functions(<code>activation=None<\/code>) when defining the <code>Conv1D<\/code> operation.<\/p>\n\n<pre><code>Input_sequence = keras.layers.Input(shape=(1500,3))\nencoder_conv1 = keras.layers.Conv1D(filters=16, kernel_size=10, \n                                    padding='same', use_bias=False, \n                                    activation=None)(Input_sequence)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57901999,"AnswerCount":0,"Tags":"<r><keras>","CreationDate":"2019-09-12T07:45:03.017","AcceptedAnswerId":null,"OwnerUserId":283538.0,"Title":"Keras model with several output and custom loss function","Body":"<p>I came across <a href=\"https:\/\/github.com\/rstudio\/keras\/issues\/451\" rel=\"nofollow noreferrer\">this post<\/a> and also <a href=\"https:\/\/towardsdatascience.com\/deep-quantile-regression-c85481548b5a\" rel=\"nofollow noreferrer\">this<\/a> in the context of fitting a Keras ANN to perform quantile regression. Unfortunately, it only appears possible to create one model per quantile. This may cause quantile crossing. I am just wondering, could one fit a multi output ANN, using Keras and R, where different outputs correspond to different quantiles?<\/p>\n\n<p>This is the code from the cited first link for reproducible:<\/p>\n\n<pre><code>quantile &lt;- 0.5\ntilted_loss &lt;- function(q, y, f) {\n  e &lt;- y - f\n  k_mean(k_maximum(q * e, (q - 1) * e), axis = 2)\n}\n\nsess &lt;- k_get_session()\nys &lt;- k_constant(c(1,2,3,4), shape = c(2,2))\nyhats &lt;- k_constant(c(1,3,3,4), shape = c(2,2))\nsess$run(tilted_loss(quantile, ys, yhats))\n\nx_train &lt;-\n  iris[1:120, c(\"Petal.Length\", \"Sepal.Length\")] %&gt;% as.matrix()\ny_train &lt;-\n  iris[1:120, c(\"Petal.Width\", \"Sepal.Width\")] %&gt;% as.matrix()\nx_test &lt;-\n  iris[121:150, c(\"Petal.Length\", \"Sepal.Length\")] %&gt;% as.matrix()\ny_test &lt;-\n  iris[121:150, c(\"Petal.Width\", \"Sepal.Width\")] %&gt;% as.matrix()\n\nmodel &lt;- keras_model_sequential()\nmodel %&gt;%\n  layer_dense(units = 32, input_shape = 2) %&gt;%\n  layer_activation_leaky_relu() %&gt;%\n  layer_dropout(rate = 0.2) %&gt;%\n  layer_dense(units = 2)\n\nmodel %&gt;% compile(\n  optimizer = \"adam\",\n  loss = function(y_true, y_pred)\n    tilted_loss(quantile, y_true, y_pred),\n  metrics = \"mae\"\n)\n\nhistory &lt;-\n  model %&gt;% fit(x_train,\n                y_train,\n                batch_size = 10,\n                epochs = 120)\n\nplot(history) \n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57902084,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2019-09-12T07:50:15.810","AcceptedAnswerId":null,"OwnerUserId":12057006.0,"Title":"Retraining on COCO from scratch using given code, got stuck when saving the checkpoint 0","Body":"<p>I want to retrain faster-rcnn on MSCOCO dataset from scratch with model_main.py. First I generate tfrecord file using create_coco_tf_record.py with COCO2017 Detection, and I got train\/val file like this: coco_train.record-00000-of-00100.\nAfter that, I ran model_main.py , and the commang window outputs many warning logs. Then I got stuck at Saving checkpoints for 0 into \/data\/code\/vision_ori\/my_checkpoints\/model.ckpt. <\/p>\n\n<p>I checked carefully and found the process stuck while creating a new MonitoredSession object.<\/p>\n\n<h3>Source code \/ logs<\/h3>\n\n<h2>logs:<\/h2>\n\n<pre><code>2019-09-12 14:36:27.183614: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1\n640] Found device 0 with properties: \nname: GeForce RTX 2080 Ti major: 7 minor: 5 memoryClockRate(GHz): 1.545\npciBusID: 0000:1b:00.0\n2019-09-12 14:36:27.185145: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1\n640] Found device 1 with properties: \nname: GeForce RTX 2080 Ti major: 7 minor: 5 memoryClockRate(GHz): 1.545\npciBusID: 0000:1c:00.0\n2019-09-12 14:36:27.186642: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1\n640] Found device 2 with properties: \nname: GeForce RTX 2080 Ti major: 7 minor: 5 memoryClockRate(GHz): 1.545\npciBusID: 0000:1d:00.0\n2019-09-12 14:36:27.188176: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1\n640] Found device 3 with properties: \nname: GeForce RTX 2080 Ti major: 7 minor: 5 memoryClockRate(GHz): 1.545\npciBusID: 0000:1e:00.0\n2019-09-12 14:36:27.190511: I tensorflow\/stream_executor\/platform\/default\/dso_lo\nader.cc:42] Successfully opened dynamic library libcudart.so.10.0\n2019-09-12 14:36:27.242760: I tensorflow\/stream_executor\/platform\/default\/dso_lo\nader.cc:42] Successfully opened dynamic library libcufft.so.10.0\n2019-09-12 14:36:27.293652: I tensorflow\/stream_executor\/platform\/default\/dso_lo\nader.cc:42] Successfully opened dynamic library libcurand.so.10.0\n2019-09-12 14:36:27.324475: I tensorflow\/stream_executor\/platform\/default\/dso_lo\nader.cc:42] Successfully opened dynamic library libcusolver.so.10.0\n2019-09-12 14:36:27.345063: I tensorflow\/stream_executor\/platform\/default\/dso_lo\nader.cc:42] Successfully opened dynamic library libcusparse.so.10.0\n2019-09-12 14:36:27.355112: I tensorflow\/stream_executor\/platform\/default\/dso_lo\nader.cc:53] Could not dlopen library 'libcudnn.so.7'; dlerror: libcudnn.so.7: ca\nnnot open shared object file: No such file or directory; LD_LIBRARY_PATH: \/usr\/l\nocal\/nvidia\/lib64:\/data\/cuda\/cuda-10.0\/cuda\/lib64:\/data\/cuda\/cuda-10.0\/cuda\/lib6\n4\n2019-09-12 14:36:27.355137: W tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1\n663] Cannot dlopen some GPU libraries. Skipping registering GPU devices...\n2019-09-12 14:36:27.355137: W tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1\n663] Cannot dlopen some GPU libraries. Skipping registering GPU devices...\n2019-09-12 14:36:27.355518: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1\n181] Device interconnect StreamExecutor with strength 1 edge matrix:\n2019-09-12 14:36:27.355536: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1\n187]      0 1 2 3 \n2019-09-12 14:36:27.355553: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1\n200] 0:   N N N N \n2019-09-12 14:36:27.355569: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1\n200] 1:   N N N N \n2019-09-12 14:36:27.355584: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1\n200] 2:   N N N N \n2019-09-12 14:36:27.355600: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1\n200] 3:   N N N N \nWARNING:tensorflow:From \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\n\/training\/saver.py:1276: checkpoint_exists (from tensorflow.python.training.chec\nkpoint_management) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse standard file APIs to check for files with this prefix.\nW0912 14:36:27.357658 140603477600000 deprecation.py:323] From \/usr\/local\/lib\/py\nthon3.5\/dist-packages\/tensorflow\/python\/training\/saver.py:1276: checkpoint_exist\ns (from tensorflow.python.training.checkpoint_management) is deprecated and will\n be removed in a future version.\nInstructions for updating:\nUse standard file APIs to check for files with this prefix.\nINFO:tensorflow:Restoring parameters from \/data\/code\/vision_ori\/my_checkpoints\/m\nodel.ckpt-0\nI0912 14:36:27.359247 140603477600000 saver.py:1280] Restoring parameters from \/\ndata\/code\/vision_ori\/my_checkpoints\/model.ckpt-0\n2019-09-12 14:36:28.141141: W tensorflow\/compiler\/jit\/mark_for_compilation_pass.\ncc:1412] (One-time warning): Not using XLA:CPU for cluster because envvar TF_XLA\n_FLAGS=--tf_xla_cpu_global_jit was not set.  If you want XLA:CPU, either set tha\nt envvar, or use experimental_jit_scope to enable XLA:CPU.  To confirm that XLA \nis active, pass --vmodule=xla_compilation_cache=1 (as a proper command-line flag\n, not via TF_XLA_FLAGS) or set the envvar XLA_FLAGS=--xla_hlo_profile.\nWARNING:tensorflow:From \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\n\/training\/saver.py:1066: get_checkpoint_mtimes (from tensorflow.python.training.\ncheckpoint_management) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse standard file utilities to get mtimes.\nW0912 14:36:28.481194 140603477600000 deprecation.py:323] From \/usr\/local\/lib\/py\nthon3.5\/dist-packages\/tensorflow\/python\/training\/saver.py:1066: get_checkpoint_m\ntimes (from tensorflow.python.training.checkpoint_management) is deprecated and \nwill be removed in a future version.\nInstructions for updating:\nUse standard file utilities to get mtimes.\nINFO:tensorflow:Running local_init_op.\nI0912 14:36:29.053177 140603477600000 session_manager.py:500] Running local_init\n_op.\nINFO:tensorflow:Done running local_init_op.\nI0912 14:36:29.196373 140603477600000 session_manager.py:502] Done running local\n_init_op.\nINFO:tensorflow:Saving checkpoints for 0 into \/data\/code\/vision_ori\/my_checkpoin\nts\/model.ckpt.\nI0912 14:36:33.405228 140603477600000 basic_session_run_hooks.py:606] Saving che\nckpoints for 0 into \/data\/code\/vision_ori\/my_checkpoints\/model.ckpt.\n<\/code><\/pre>\n\n<p>It gets stuck here for days and just won't go on.<\/p>\n\n<h2>building tf-record:<\/h2>\n\n<pre><code>python3 create_coco_tf_record.py --logtostderr \\\n--train_image_dir=\"\/data\/code\/vision_ori\/dataset\/train2017\" \\\n--val_image_dir=\"\/data\/code\/vision_ori\/dataset\/val2017\" \\\n--test_image_dir=\"\/data\/code\/vision_ori\/dataset\/test2017\" \\\n--train_annotations_file=\"\/data\/code\/vision_ori\/dataset\/anno\/instances_train2017.json\" \\\n--val_annotations_file=\"\/data\/code\/vision_ori\/dataset\/anno\/annotations\/instances_val2017.json\" \\\n--testdev_annotations_file=\"\/data\/code\/vision_ori\/dataset\/anno\/annotations\/image_info_test-dev2017.json\" \\\n--output_dir=\"cocodata\"\n\n<\/code><\/pre>\n\n<h2>trying to train:<\/h2>\n\n<pre><code>python3 object_detection\/model_main.py \\\n    --pipeline_config_path=\"\/data\/code\/vision_ori\/models\/research\/object_detection\/samples\/configs\/faster_rcnn_inception_resnet_v2_atrous_coco.config\" \\\n    --model_dir=\"\/data\/code\/vision_ori\/my_checkpoints\" \\\n    --num_train_steps=200000 \\\n    --sample_1_of_n_eval_examples=1 \\\n    --alsologtostderr\n\n<\/code><\/pre>\n\n<h2>config file:<\/h2>\n\n<p>I would like to train from scratch so I removed 2 lines of code:<\/p>\n\n<pre><code>fine_tune_checkpoint: \"PATH_TO_BE_CONFIGURED\/model.ckpt\"\n  from_detection_checkpoint: true\n<\/code><\/pre>\n\n<pre><code>model {\n  faster_rcnn {\n    num_classes: 90\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_inception_resnet_v2'\n      first_stage_features_stride: 8\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: 8\n        width_stride: 8\n      }\n    }\n    first_stage_atrous_rate: 2\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: 17\n    maxpool_kernel_size: 1\n    maxpool_stride: 1\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: 100\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  optimizer {\n    momentum_optimizer: {\n      learning_rate: {\n        manual_step_learning_rate {\n          initial_learning_rate: 0.0003\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  # 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}\n\ntrain_input_reader: {\n  tf_record_input_reader {\n    input_path: \"\/data\/code\/vision_ori\/models\/research\/object_detection\/dataset_tools\/cocodata\/coco_train.record-00000-of-00100\"\n  }\n  label_map_path: \"\/data\/code\/vision_ori\/models\/research\/object_detection\/data\/mscoco_label_map.pbtxt\"\n}\n\neval_config: {\n  num_examples: 5000\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: \"\/data\/code\/vision_ori\/models\/research\/object_detection\/dataset_tools\/cocodata\/coco_val.record-00000-of-00010\"\n  }\n  label_map_path: \"\/data\/code\/vision_ori\/models\/research\/object_detection\/data\/mscoco_label_map.pbtxt\"\n  shuffle: false\n  num_readers: 1\n}\n\n<\/code><\/pre>\n\n<p>No error information. Just got stuck here.<\/p>\n","answers":[]}
+{"QuestionId":57902381,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-12T08:09:13.010","AcceptedAnswerId":57928137.0,"OwnerUserId":11305248.0,"Title":"Keras loss seems to be wrong - weighted sum is wrong","Body":"<p>I'm using Keras with tensorflow backend to train a model with multiple outputs but only one loss (MSE on output layer \"parts\"). This is defined as follows:<\/p>\n\n<pre><code>model = Model(inputs=[image, year], outputs=outputs) # outputs is a list of 8 output layers\nmodel_parallel = keras.utils.multi_gpu_model(model, gpus=2, cpu_merge=True, cpu_relocation=False)\nlosses = {'parts': keras.losses.mean_squared_error}\nloss_weights = {'parts': 1.}\nmodel_parallel.compile(loss=losses,\n                       loss_weights=loss_weights,\n                       optimizer=keras.optimizers.RMSprop(lr=args.lr)\n                       )\nmodel_parallel.fit_generator(generator=generator_training,\n                             epochs=args.epochs,\n                             verbose=1\n                             )\n<\/code><\/pre>\n\n<p>fit_generator prints the total loss and my parts_loss:<\/p>\n\n<pre><code>33\/5730 [..............................] - ETA: 14:28:59 - loss: 117.6318 - parts_loss: 97.1686\n<\/code><\/pre>\n\n<p>Why is the total loss (= weighted sum of all losses) different to parts_loss?<\/p>\n","answers":[{"AnswerId":"57928137","CreationDate":"2019-09-13T17:38:16.643","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>The total loss is different from <code>parts_loss<\/code> because it includes the regularization term. In other words, <code>loss<\/code> is computed as <code>loss = parts_loss + k*R<\/code>, where <code>R<\/code> is the regularization term (typically the L1 or L2 norm of the model's weights) and <code>k<\/code> a hyperparameter that controls the contribution of the regularization loss in the total loss.<\/p>\n"}]}
+{"QuestionId":57902387,"AnswerCount":3,"Tags":"<python><tensorflow><tensorflow2.0><matrix-factorization>","CreationDate":"2019-09-12T08:09:34.017","AcceptedAnswerId":null,"OwnerUserId":6244166.0,"Title":"Matrix Factorization in tensorflow 2.0 using WALS Method","Body":"<p>I am using WALS method in order to perform matrix factorization. Initially in tensorflow 1.13 I can import factorization_ops using<\/p>\n\n<pre><code>from tensorflow.contrib.factorization.python.ops import factorization_ops \n<\/code><\/pre>\n\n<p>As described in the <a href=\"https:\/\/github.com\/GoogleCloudPlatform\/tensorflow-recommendation-wals\/blob\/master\/wals_ml_engine\/trainer\/wals.py\" rel=\"nofollow noreferrer\">documentation<\/a> <\/p>\n\n<p>Wals model can be called from factorization_ops by using<\/p>\n\n<pre><code>factorization_ops.WALSModel\n<\/code><\/pre>\n\n<p>Using same command in tensorflow 2.0 giving me following error<\/p>\n\n<p><strong>ModuleNotFoundError: No module named 'tensorflow.contrib.factorization<\/strong><\/p>\n\n<p>Going through the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/31350\" rel=\"nofollow noreferrer\">issue<\/a> there appears to be no way out to use WALSModel in tensorflow 2.0+. <\/p>\n\n<p>Also it has been mentioned <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/releases\" rel=\"nofollow noreferrer\">here<\/a> in tensorflow release updates that tf.contrib has been deprecated, and functionality has been either migrated to the core TensorFlow API, to an ecosystem project such as tensorflow\/addons or tensorflow\/io, or removed entirely.<\/p>\n\n<p>How can I use WALS model in tensorflow 2.0 (Currently I am using 2.0.0-rc0 on windows machine) ? Is WALSModel has been removed or I am missing out some information ?<\/p>\n","answers":[{"AnswerId":"58057252","CreationDate":"2019-09-23T07:07:21.710","ParentId":null,"OwnerUserId":"4373898","Title":null,"Body":"<p>I think that WALS has been removed. As part of <code>tf.contrib<\/code> it is not supported by TF2 and I\u00a0do not think it fits in any of core or sub-projects.\nYour best bet is probably to make it available as a third-party library.<\/p>\n\n<p>I expect to use it for my project, but the need to re-write it (mainly copy what was in TF1 and make it work as a separate library compatible with TF2) reduce the priority of this task...<\/p>\n\n<p>Let us know if you start to code something. Thanks.<\/p>\n\n<p>Alexis.<\/p>\n"},{"AnswerId":"58079028","CreationDate":"2019-09-24T11:13:09.620","ParentId":null,"OwnerUserId":"11374349","Title":null,"Body":"<p>I have the same issue, but I don't really have time to write a library myself unfortunately. There are several potential options that I am considering: <\/p>\n\n<ol>\n<li><p>Stick with TF1.X until someone creates a library <\/p><\/li>\n<li><p>Switch to using lightfm to continue using WALS <\/p><\/li>\n<li><p>Switch to neural collaborative filtering using embedding layers with keras and a dot product layer. See this paper <a href=\"https:\/\/arxiv.org\/abs\/1708.05031\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1708.05031<\/a>, and this code implementation:<\/p><\/li>\n<\/ol>\n\n<pre><code>from tensorflow.keras.layers import Input, Embedding, Flatten, Dot, Dense\nfrom tensorflow.keras.models import Model\n#import tensorflow.distribute\n\ndef get_compiled_model(n_users, n_items, embedding_dims=20):\n    # Product embedding\n    prod_input = Input(shape=[1], name=\"Item-Input\")\n    prod_embedding = Embedding(n_items+1, embedding_dims, name=\"Item-Embedding\")(prod_input)\n    prod_vec = Flatten(name=\"Flatten-Product\")(prod_embedding)\n\n    # User embedding\n    user_input = Input(shape=[1], name=\"User-Input\")\n    user_embedding = Embedding(n_users+1, embedding_dims, name=\"User-Embedding\")(user_input)\n    user_vec = Flatten(name=\"Flatten-Users\")(user_embedding)\n\n    # The output is the dot product of the two, i.e. a one-hot vector\n    dot_product = Dot(name=\"Dot-Product\", axes=1)([prod_vec, user_vec])\n\n    # compile - uncomment these two lines to make training distributed\n    # dist_strat = distribute.Strategy()\n    # with dist_strat.scope():\n    model = Model(inputs = [user_input, prod_input], outputs = dot_product)\n    model.compile(\n        optimizer='adam',\n        loss='mean_squared_error'\n    )\n    return model\n\n\n<\/code><\/pre>\n"},{"AnswerId":"58081290","CreationDate":"2019-09-24T13:21:01.010","ParentId":null,"OwnerUserId":"3134944","Title":null,"Body":"<p>I believe WALS is not supported in TF 2.0 ...The official recommendation model is <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/08bb9eb5ad79e6bceffc71aeea6af809cc78694b\/official\/recommendation\" rel=\"nofollow noreferrer\">Neural Collaborative Filter (NCF)<\/a>  <\/p>\n\n<p>I hope this helps.<\/p>\n\n<p>M<\/p>\n"}]}
+{"QuestionId":57902596,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2019-09-12T08:21:57.483","AcceptedAnswerId":null,"OwnerUserId":11805922.0,"Title":"How to merge .h5 weights from different training result?","Body":"<p>I have NB, PC and Server ( plus Colab), I want all of them running the same one my Keras RL code.<\/p>\n\n<p>However, the problem is how to merge their different training result (.h5) files?<\/p>\n\n<p>When GPU is doing parallel training, is it doing: <\/p>\n\n<pre><code>New_weight=  ( Trained_Result_A + Trained_Result_B ) \/ 2 ?\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57902687,"AnswerCount":1,"Tags":"<python><tensorflow><keras><loss-function><loss>","CreationDate":"2019-09-12T08:27:04.703","AcceptedAnswerId":null,"OwnerUserId":12042308.0,"Title":"Can I disable val_loss caculation in keras?","Body":"<p>Actually I want to use different loss functions in training and validation phase. I tried in_tarin_phase but it doesn't work. <\/p>\n\n<p>So I just wonder can I disable the val_loss calculation?<\/p>\n","answers":[{"AnswerId":"57902999","CreationDate":"2019-09-12T08:45:35.190","ParentId":null,"OwnerUserId":"10579274","Title":null,"Body":"<p>Below has a custom loss function:<\/p>\n\n<pre><code># Build a model\ninputs = Input(shape=(128,))\nlayer1 = Dense(64, activation='relu')(inputs)\nlayer2 = Dense(64, activation='relu')(layer1)\npredictions = Dense(10, activation='softmax')(layer2)\nmodel = Model(inputs=inputs, outputs=predictions)\n\n# Define custom loss\ndef custom_loss(layer):\n\n    # Create a loss function that adds the MSE loss to the mean of all squared activations of a specific layer\n    def loss(y_true,y_pred):\n        return K.mean(K.square(y_pred - y_true) + K.square(layer), axis=-1)\n\n    # Return a function\n    return loss\n\n# Compile the model\nmodel.compile(optimizer='adam',\n              loss=custom_loss(layer), # Call the loss function with the selected layer\n              metrics=['accuracy'])\n\n# train\nmodel.fit(data, labels) \n<\/code><\/pre>\n"}]}
+{"QuestionId":57902980,"AnswerCount":4,"Tags":"<python-3.x><tensorflow><object-detection>","CreationDate":"2019-09-12T08:44:32.900","AcceptedAnswerId":57903964.0,"OwnerUserId":8829431.0,"Title":"Tensorflow Object Detection: Using new detected images to retrain model","Body":"<p>I have searched through this forum for similar questions but was unanswered (<a href=\"https:\/\/stackoverflow.com\/q\/51884713\/8829431\">Updating Tensorflow Object detection model with new images<\/a>). I have managed to create my custom train model (lets name it model1). Was wondering if can i use new images that are processed by model1 to further train model1? will it improve the accuracy of the model?<\/p>\n","answers":[{"AnswerId":"57903286","CreationDate":"2019-09-12T09:04:24.507","ParentId":null,"OwnerUserId":"2269826","Title":null,"Body":"<p>Accuracy will depend on the number of correctly classified images and not only on the total number of training images. <a href=\"https:\/\/developers.google.com\/machine-learning\/crash-course\/classification\/accuracy\" rel=\"nofollow noreferrer\">https:\/\/developers.google.com\/machine-learning\/crash-course\/classification\/accuracy<\/a>. If you consider that the new images are to be used for training (have correct labels), then you should consider re-training the model. Take a look at this post <a href=\"https:\/\/datascience.stackexchange.com\/questions\/12761\/should-a-model-be-re-trained-if-new-observations-are-available\">https:\/\/datascience.stackexchange.com\/questions\/12761\/should-a-model-be-re-trained-if-new-observations-are-available<\/a><\/p>\n"},{"AnswerId":"57903154","CreationDate":"2019-09-12T08:56:05.430","ParentId":null,"OwnerUserId":"9033534","Title":null,"Body":"<p>It might improve it but it is tricky. It would lead to overfitting. Improving the data set would actually help, but not with images detected by its own model. This kind of images are detected cause the model already performs well on them, so not much help.<\/p>\n\n<p>What you need actually is quite the opposite. You need to teach the model to recognize the images that it didn't recognize before<\/p>\n"},{"AnswerId":"58064108","CreationDate":"2019-09-23T13:56:38.497","ParentId":null,"OwnerUserId":"5576984","Title":null,"Body":"<p>You can use your current model (model1) in a number of ways:<\/p>\n\n<ul>\n<li>on new images to detect bad results (hard examples) for new training<\/li>\n<li>on new images to detect good results for evaluation<\/li>\n<li>on the images in the existing dataset to detect bad images (wrong label etc.)<\/li>\n<\/ul>\n\n<p>Some of the bad results from new images will be non-objects (adversarial) and not directly usable for training (but see this: <a href=\"https:\/\/github.com\/tensorflow\/models\/issues\/3578#issuecomment-375267920\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/issues\/3578#issuecomment-375267920<\/a>).<\/p>\n\n<p>Removal of bad images from the existing dataset requires retraining from scratch unless there is some funky way of \"untraining\" images from a model.<\/p>\n\n<p>Eventually one would end up approaching a perfect dataset that makes best use of the capacity of the chosen model architecture, although the domain may evolve over time.<\/p>\n\n<p>I think the reason this is not much discussed is because most researchers have to work with common datasets so they can compare their approaches (brilliant read: <a href=\"https:\/\/www.ncbi.nlm.nih.gov\/pmc\/articles\/PMC5697567\/\" rel=\"nofollow noreferrer\">https:\/\/www.ncbi.nlm.nih.gov\/pmc\/articles\/PMC5697567\/<\/a>). <\/p>\n"},{"AnswerId":"57903964","CreationDate":"2019-09-12T09:43:05.940","ParentId":null,"OwnerUserId":"3871961","Title":null,"Body":"<p>The main problem of machine learning (that is the approach you are using for object detection here) is that of <strong>generalization<\/strong>. In your case, it is the ability to recognize objects of the same type as image you used for training, in images that were not used during training.<\/p>\n\n<p>Obviously, if you were able to use all the possible images during training, your system would be perfect (actually, it would be a simple exact image matching problem). In a more realistic setup, the more training image you are using, the higher chance you have to obtain a better object detector.<\/p>\n\n<p>Usually, it is however more valuable to add <strong>hard examples<\/strong> to your training set. Hence, if your application allows it (in terms of computation time in particular) you can indeed add all the images that are wrongly detected in your dataset (with the correct label) and it will probably help to get a better model, able to detect the object in harder condition on new images.<\/p>\n\n<p>However, it really depends on what you are doing. If you want to compare your system to another one, you need to use the same (training and) test images to be fair. <strong>For benchmarking, you are not allowed to include test images in the training dataset<\/strong>! When you compute the accuracy (on a validation\/test dataset) to compare several settings, be sure you are fair in this comparison.<\/p>\n"}]}
+{"QuestionId":57903518,"AnswerCount":2,"Tags":"<python><tensorflow><sigmoid>","CreationDate":"2019-09-12T09:18:42.263","AcceptedAnswerId":null,"OwnerUserId":102720.0,"Title":"Interpreting a sigmoid result as probability in neural networks","Body":"<p>I've created a neural network with a sigmoid activation function in the last layer, so I get results between 0 and 1. I want to classify things in 2 classes, so I check \"is the number > 0.5, then class 1 else class 0\". All basic.\nHowever, I would like to say \"the probability of it being in class 0 is <code>x<\/code> and in class 1 is <code>y<\/code>\".<\/p>\n\n<p>How can I do this?<\/p>\n\n<ul>\n<li>Does a number like 0.73 tell me it's 73% sure to be in class 1? And then 1-0.73 = 0.27 so 27% in class 0?<\/li>\n<li>When it's 0.27, does that mean it's 27% sure in class 0, 73% in class 1? Makes no sense.<\/li>\n<\/ul>\n\n<p>Should I work with the 0.5 and look \"how far away from the center is the number, and then that's the percentage\"?<\/p>\n\n<p>Or am I misunderstanding the result of the NN?<\/p>\n","answers":[{"AnswerId":"57907269","CreationDate":"2019-09-12T12:57:16.360","ParentId":null,"OwnerUserId":"6490241","Title":null,"Body":"<pre><code>Does a number like 0.73 tell me it's 73% sure to be in class 1? And then 1-0.73 = 0.27 so 27% in class 0?\n<\/code><\/pre>\n\n<p>The Answer is No. When we are using Sigmoid Function the sum of the results will not sum to 1.There are chances that sum of results of the classes will be less than 1 or in some cases it will be greater than 1.<\/p>\n\n<p>In the same case,when we use the softmax function. The sum of all the outputs will be added to 1.<\/p>\n"},{"AnswerId":"57907511","CreationDate":"2019-09-12T13:10:40.880","ParentId":null,"OwnerUserId":"8244247","Title":null,"Body":"<p>As pointed out by Teja, the short answer is no, however, depending on the loss you use, it may be closer to truth than you may think.<\/p>\n\n<p>Imagine you try to train your network to differentiate numbers into two arbitrary categories that are <code>beautiful<\/code> and <code>ugly<\/code>. Say your input number are either <code>0<\/code> or <code>1<\/code> and <code>0<\/code>s have a <code>0.2<\/code> probability of being labelled <code>ugly<\/code> whereas <code>1<\/code>s have o <code>0.6<\/code>probability of being <code>ugly<\/code>.<\/p>\n\n<p>Imagine that your neural network takes as inputs <code>0<\/code>s and <code>1<\/code>s, passes them into some layers, and ends in a softmax function. If your loss is <em>binary cross-entropy<\/em>, then the optimal solution for your network is to output <code>0.2<\/code> when it sees a <code>0<\/code> in input and <code>0.6<\/code> when it sees a <code>1<\/code> in input (this is a property of the cross-entropy which is minimized when you output the true probabilities of each label). Therefore, you can interpret these numbers as probabilities.<\/p>\n\n<p>Of course, real world examples are not that easy and are generally deterministic so the interpretation is a little bit tricky. However, I believe that it is not entirely false to think of your results as probabilities as long as you use the cross-entropy as a loss.<\/p>\n\n<p>I'm sorry, this answer is not black or white, but reality is sometimes complex ;)<\/p>\n"}]}
+{"QuestionId":57904166,"AnswerCount":0,"Tags":"<python><keras><neural-network>","CreationDate":"2019-09-12T09:55:02.423","AcceptedAnswerId":null,"OwnerUserId":11580689.0,"Title":"Feature weights","Body":"<p>Is there a way to give weights to features before training a neural network model? The meaning of <em>weight<\/em> here would be similar to that in linear regression - a representation of how much to trust each feature. This is unrelated to node weights or feature importance estimation after training. <\/p>\n\n<p>Something similar to this <a href=\"https:\/\/arxiv.org\/pdf\/1101.4918\" rel=\"nofollow noreferrer\">publication<\/a>.<\/p>\n","answers":[]}
+{"QuestionId":57904201,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-12T09:57:21.253","AcceptedAnswerId":null,"OwnerUserId":11208496.0,"Title":"What could be the reason that a fully connected layer not connected to the computation graph, but affects the accuracy?","Body":"<p>I'm working on the CitationNetwork datasets. After some point, I found that if I insert a fully connected layer to my code (not connect it to the other part of the graph), it affects the accuracy. I tried 100 runs with different seeds: the one where the FC was there, but not connected to the computational graph, reached better accuracy on Cora dataset than the other one, where the code didn't include that line. (Instead of 80.4, 78.2). \nOn Pubmed dataset, the effect was reversed, the code without FC overperformed the code with FC. (77, 78.8).<\/p>\n\n<pre><code>self.fc = tf.contrib.layers.fully_connected(\n    features,\n    num_output_layers=0,\n    activation_fn=tf.nn.sigmoid)\n<\/code><\/pre>\n\n<p>Changing the num_output_layers parameter didn't affect this question at all.\nAt the beginning, I thought this is some kind of initialization-problem, but since I ran it with 100 different seed, I have no more idea. <\/p>\n\n<p>Edit: Adding\n    <code>weights_initializer=tf.initializers.truncated_normal<\/code>\nas a parameter to the FC, the results will be the same as without FC.<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[]}
+{"QuestionId":57904536,"AnswerCount":0,"Tags":"<tensorflow><deeplab>","CreationDate":"2019-09-12T10:14:30.077","AcceptedAnswerId":null,"OwnerUserId":1591509.0,"Title":"Get more accuracy DeepLab then the pretrainned model","Body":"<p>I have a question in deeplab v3 accuracy right now the best accuracy(mIOU) achieved by the model xception65_coco_voc_trainval 87.80. I was wondering that can we further increase it if i train the model for less number of class lets say for person only using the same dataset or it will not make any difference.\nThe problem is along the edges model has problem identifying it correctly. The orignal image had all green background and after applying the mask some greenish remains. So 2 questions in short\n1. If i train for single class then will accuracy improve ?\n2.  If not then is there any other method i can apply to achive better results\nNote:\nBackground doesn't have to be green in all cases.\ni saw the remove.bg website and they have done good job, does any one knows what approach have they taken to achieve good results\nThanks<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/7zJSH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/7zJSH.png\" alt=\"enter image description here\"><\/a> <\/p>\n","answers":[]}
+{"QuestionId":57904704,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><audio><pyaudio>","CreationDate":"2019-09-12T10:23:51.873","AcceptedAnswerId":null,"OwnerUserId":265683.0,"Title":"How do you use a TensorFlow model to classify a sound using this code","Body":"<p>I have been following this tutorial for creating a <a href=\"http:\/\/ttps:\/\/www.tensorflow.org\/tutorials\/sequences\/audio_recognition\" rel=\"nofollow noreferrer\">TensorFlow Audio Recognition model<\/a>.<\/p>\n\n<p>I have completed training of my model (which is based upon a smaller set of simple data from the content in the example).<\/p>\n\n<p>Now I have built the model, what is the lines of code I need in order to actually classify against the created model?<\/p>\n\n<p>Please see the following code which has built my model which is a slightly altered version of the article's code:<\/p>\n\n<p>train.py:<\/p>\n\n<pre><code>from __future__ import absolute_import\nfrom __future__ import division\nfrom __future__ import print_function\n\nimport argparse\nimport os.path\nimport sys\n\nimport numpy as np\nfrom six.moves import xrange  # pylint: disable=redefined-builtin\nimport tensorflow as tf\n\nimport input_data\nimport models\nfrom tensorflow.python.platform import gfile\n\nFLAGS = None\n\n\ndef main(_):\n  # Set the verbosity based on flags (default is INFO, so we see all messages)\n  tf.compat.v1.logging.set_verbosity(FLAGS.verbosity)\n\n  # Start a new TensorFlow session.\n  sess = tf.compat.v1.InteractiveSession()\n\n  # Begin by making sure we have the training data we need. If you already have\n  # training data of your own, use `--data_url= ` on the command line to avoid\n  # downloading.c\n  model_settings = models.prepare_model_settings(\n      len(input_data.prepare_words_list(FLAGS.wanted_words.split(','))),\n      FLAGS.sample_rate, FLAGS.clip_duration_ms, FLAGS.window_size_ms,\n      FLAGS.window_stride_ms, FLAGS.feature_bin_count, FLAGS.preprocess)\n  audio_processor = input_data.AudioProcessor(\n      FLAGS.data_url, FLAGS.data_dir,\n      FLAGS.silence_percentage, FLAGS.unknown_percentage,\n      FLAGS.wanted_words.split(','), FLAGS.validation_percentage,\n      FLAGS.testing_percentage, model_settings, FLAGS.summaries_dir)\n  fingerprint_size = model_settings['fingerprint_size']\n  label_count = model_settings['label_count']\n  time_shift_samples = int((FLAGS.time_shift_ms * FLAGS.sample_rate) \/ 1000)\n  # Figure out the learning rates for each training phase. Since it's often\n  # effective to have high learning rates at the start of training, followed by\n  # lower levels towards the end, the number of steps and learning rates can be\n  # specified as comma-separated lists to define the rate at each stage. For\n  # example --how_many_training_steps=10000,3000 --learning_rate=0.001,0.0001\n  # will run 13,000 training loops in total, with a rate of 0.001 for the first\n  # 10,000, and 0.0001 for the final 3,000.\n  training_steps_list = list(map(int, FLAGS.how_many_training_steps.split(',')))\n  learning_rates_list = list(map(float, FLAGS.learning_rate.split(',')))\n  if len(training_steps_list) != len(learning_rates_list):\n    raise Exception(\n        '--how_many_training_steps and --learning_rate must be equal length '\n        'lists, but are %d and %d long instead' % (len(training_steps_list),\n                                                   len(learning_rates_list)))\n\n  input_placeholder = tf.compat.v1.placeholder(\n      tf.float32, [None, fingerprint_size], name='fingerprint_input')\n  if FLAGS.quantize:\n    fingerprint_min, fingerprint_max = input_data.get_features_range(\n        model_settings)\n    fingerprint_input = tf.quantization.fake_quant_with_min_max_args(\n        input_placeholder, fingerprint_min, fingerprint_max)\n  else:\n    fingerprint_input = input_placeholder\n\n  logits, dropout_prob = models.create_model(\n      fingerprint_input,\n      model_settings,\n      FLAGS.model_architecture,\n      is_training=True)\n\n  # Define loss and optimizer\n  ground_truth_input = tf.compat.v1.placeholder(\n      tf.int64, [None], name='groundtruth_input')\n\n  # Optionally we can add runtime checks to spot when NaNs or other symptoms of\n  # numerical errors start occurring during training.\n  control_dependencies = []\n  if FLAGS.check_nans:\n    checks = tf.compat.v1.add_check_numerics_ops()\n    control_dependencies = [checks]\n\n  # Create the back propagation and training evaluation machinery in the graph.\n  with tf.compat.v1.name_scope('cross_entropy'):\n    cross_entropy_mean = tf.compat.v1.losses.sparse_softmax_cross_entropy(\n        labels=ground_truth_input, logits=logits)\n  if FLAGS.quantize:\n    tf.contrib.quantize.create_training_graph(quant_delay=0)\n  with tf.compat.v1.name_scope('train'), tf.control_dependencies(\n      control_dependencies):\n    learning_rate_input = tf.compat.v1.placeholder(\n        tf.float32, [], name='learning_rate_input')\n    train_step = tf.compat.v1.train.GradientDescentOptimizer(\n        learning_rate_input).minimize(cross_entropy_mean)\n  predicted_indices = tf.argmax(input=logits, axis=1)\n  correct_prediction = tf.equal(predicted_indices, ground_truth_input)\n  confusion_matrix = tf.math.confusion_matrix(labels=ground_truth_input,\n                                              predictions=predicted_indices,\n                                              num_classes=label_count)\n  evaluation_step = tf.reduce_mean(input_tensor=tf.cast(correct_prediction,\n                                                        tf.float32))\n  with tf.compat.v1.get_default_graph().name_scope('eval'):\n    tf.compat.v1.summary.scalar('cross_entropy', cross_entropy_mean)\n    tf.compat.v1.summary.scalar('accuracy', evaluation_step)\n\n  global_step = tf.compat.v1.train.get_or_create_global_step()\n  increment_global_step = tf.compat.v1.assign(global_step, global_step + 1)\n\n  saver = tf.compat.v1.train.Saver(tf.compat.v1.global_variables())\n\n  # Merge all the summaries and write them out to \/tmp\/retrain_logs (by default)\n  merged_summaries = tf.compat.v1.summary.merge_all(scope='eval')\n  train_writer = tf.compat.v1.summary.FileWriter(FLAGS.summaries_dir + '\/train',\n                                                 sess.graph)\n  validation_writer = tf.compat.v1.summary.FileWriter(\n      FLAGS.summaries_dir + '\/validation')\n\n  tf.compat.v1.global_variables_initializer().run()\n\n  start_step = 1\n\n  if FLAGS.start_checkpoint:\n    models.load_variables_from_checkpoint(sess, FLAGS.start_checkpoint)\n    start_step = global_step.eval(session=sess)\n\n  tf.compat.v1.logging.info('Training from step: %d ', start_step)\n\n  # Save graph.pbtxt.\n  tf.io.write_graph(sess.graph_def, FLAGS.train_dir,\n                    FLAGS.model_architecture + '.pbtxt')\n\n  # Save list of words.\n  with gfile.GFile(\n      os.path.join(FLAGS.train_dir, FLAGS.model_architecture + '_labels.txt'),\n      'w') as f:\n    f.write('\\n'.join(audio_processor.words_list))\n\n  # Training loop.\n  training_steps_max = np.sum(training_steps_list)\n  for training_step in xrange(start_step, training_steps_max + 1):\n    # Figure out what the current learning rate is.\n    training_steps_sum = 0\n    for i in range(len(training_steps_list)):\n      training_steps_sum += training_steps_list[i]\n      if training_step &lt;= training_steps_sum:\n        learning_rate_value = learning_rates_list[i]\n        break\n    # Pull the audio samples we'll use for training.\n    train_fingerprints, train_ground_truth = audio_processor.get_data(\n        FLAGS.batch_size, 0, model_settings, FLAGS.background_frequency,\n        FLAGS.background_volume, time_shift_samples, 'training', sess)\n    # Run the graph with this batch of training data.\n    train_summary, train_accuracy, cross_entropy_value, _, _ = sess.run(\n        [\n            merged_summaries,\n            evaluation_step,\n            cross_entropy_mean,\n            train_step,\n            increment_global_step,\n        ],\n        feed_dict={\n            fingerprint_input: train_fingerprints,\n            ground_truth_input: train_ground_truth,\n            learning_rate_input: learning_rate_value,\n            dropout_prob: 0.5\n        })\n    train_writer.add_summary(train_summary, training_step)\n    tf.compat.v1.logging.info(\n        'Step #%d: rate %f, accuracy %.1f%%, cross entropy %f' %\n        (training_step, learning_rate_value, train_accuracy * 100,\n         cross_entropy_value))\n    is_last_step = (training_step == training_steps_max)\n    if (training_step % FLAGS.eval_step_interval) == 0 or is_last_step:\n      set_size = audio_processor.set_size('validation')\n      total_accuracy = 0\n      total_conf_matrix = None\n      for i in xrange(0, set_size, FLAGS.batch_size):\n        validation_fingerprints, validation_ground_truth = (\n            audio_processor.get_data(FLAGS.batch_size, i, model_settings, 0.0,\n                                     0.0, 0, 'validation', sess))\n        # Run a validation step and capture training summaries for TensorBoard\n        # with the `merged` op.\n        validation_summary, validation_accuracy, conf_matrix = sess.run(\n            [merged_summaries, evaluation_step, confusion_matrix],\n            feed_dict={\n                fingerprint_input: validation_fingerprints,\n                ground_truth_input: validation_ground_truth,\n                dropout_prob: 1.0\n            })\n        validation_writer.add_summary(validation_summary, training_step)\n        batch_size = min(FLAGS.batch_size, set_size - i)\n        total_accuracy += (validation_accuracy * batch_size) \/ set_size\n        if total_conf_matrix is None:\n          total_conf_matrix = conf_matrix\n        else:\n          total_conf_matrix += conf_matrix\n      tf.compat.v1.logging.info('Confusion Matrix:\\n %s' % (total_conf_matrix))\n      tf.compat.v1.logging.info('Step %d: Validation accuracy = %.1f%% (N=%d)' %\n                                (training_step, total_accuracy * 100, set_size))\n\n    # Save the model checkpoint periodically.\n    if (training_step % FLAGS.save_step_interval == 0 or\n        training_step == training_steps_max):\n      checkpoint_path = os.path.join(FLAGS.train_dir,\n                                     FLAGS.model_architecture + '.ckpt')\n      tf.compat.v1.logging.info('Saving to \"%s-%d\"', checkpoint_path,\n                                training_step)\n      saver.save(sess, checkpoint_path, global_step=training_step)\n\n  set_size = audio_processor.set_size('testing')\n  tf.compat.v1.logging.info('set_size=%d', set_size)\n  total_accuracy = 0\n  total_conf_matrix = None\n  for i in xrange(0, set_size, FLAGS.batch_size):\n    test_fingerprints, test_ground_truth = audio_processor.get_data(\n        FLAGS.batch_size, i, model_settings, 0.0, 0.0, 0, 'testing', sess)\n    test_accuracy, conf_matrix = sess.run(\n        [evaluation_step, confusion_matrix],\n        feed_dict={\n            fingerprint_input: test_fingerprints,\n            ground_truth_input: test_ground_truth,\n            dropout_prob: 1.0\n        })\n    batch_size = min(FLAGS.batch_size, set_size - i)\n    total_accuracy += (test_accuracy * batch_size) \/ set_size\n    if total_conf_matrix is None:\n      total_conf_matrix = conf_matrix\n    else:\n      total_conf_matrix += conf_matrix\n  tf.compat.v1.logging.warn('Confusion Matrix:\\n %s' % (total_conf_matrix))\n  tf.compat.v1.logging.warn('Final test accuracy = %.1f%% (N=%d)' %\n                            (total_accuracy * 100, set_size))\n\n\nif __name__ == '__main__':\n  parser = argparse.ArgumentParser()\n  parser.add_argument(\n      '--data_dir',\n      type=str,\n      default='Audio\/training_data',\n      help=\"\"\"\\\n      Where to download the speech training data to.\n      \"\"\")\n  parser.add_argument(\n      '--background_volume',\n      type=float,\n      default=0.1,\n      help=\"\"\"\\\n      How loud the background noise should be, between 0 and 1.\n      \"\"\")\n  parser.add_argument(\n      '--background_frequency',\n      type=float,\n      default=0.0,\n      help=\"\"\"\\\n      How many of the training samples have background noise mixed in.\n      \"\"\")\n  parser.add_argument(\n      '--silence_percentage',\n      type=float,\n      default=10.0,\n      help=\"\"\"\\\n      How much of the training data should be silence.\n      \"\"\")\n  parser.add_argument(\n      '--unknown_percentage',\n      type=float,\n      default=10.0,\n      help=\"\"\"\\\n      How much of the training data should be unknown words.\n      \"\"\")\n  parser.add_argument(\n      '--time_shift_ms',\n      type=float,\n      default=100.0,\n      help=\"\"\"\\\n      Range to randomly shift the training audio by in time.\n      \"\"\")\n  parser.add_argument(\n      '--testing_percentage',\n      type=int,\n      default=10,\n      help='What percentage of wavs to use as a test set.')\n  parser.add_argument(\n      '--validation_percentage',\n      type=int,\n      default=10,\n      help='What percentage of wavs to use as a validation set.')\n  parser.add_argument(\n      '--sample_rate',\n      type=int,\n      default=16000,\n      help='Expected sample rate of the wavs',)\n  parser.add_argument(\n      '--clip_duration_ms',\n      type=int,\n      default=1000,\n      help='Expected duration in milliseconds of the wavs',)\n  parser.add_argument(\n      '--window_size_ms',\n      type=float,\n      default=30.0,\n      help='How long each spectrogram timeslice is.',)\n  parser.add_argument(\n      '--window_stride_ms',\n      type=float,\n      default=10.0,\n      help='How far to move in time between spectogram timeslices.',)\n  parser.add_argument(\n      '--feature_bin_count',\n      type=int,\n      default=40,\n      help='How many bins to use for the MFCC fingerprint',\n  )\n  parser.add_argument(\n      '--how_many_training_steps',\n      type=str,\n      default='200,50',\n      help='How many training loops to run',)\n  parser.add_argument(\n      '--eval_step_interval',\n      type=int,\n      default=50,\n      help='How often to evaluate the training results.')\n  parser.add_argument(\n      '--learning_rate',\n      type=str,\n      default='0.001,0.0001',\n      help='How large a learning rate to use when training.')\n  parser.add_argument(\n      '--batch_size',\n      type=int,\n      default=10,\n      help='How many items to train with at once',)\n  parser.add_argument(\n      '--summaries_dir',\n      type=str,\n      default='logs\/retrain_logs',\n      help='Where to save summary logs for TensorBoard.')\n  parser.add_argument(\n      '--wanted_words',\n      type=str,\n      default='splash,footsteps,enemy',\n      help='Words to use (others will be added to an unknown label)',)\n  parser.add_argument(\n      '--train_dir',\n      type=str,\n      default='logs\/commands_train',\n      help='Directory to write event logs and checkpoint.')\n  parser.add_argument(\n      '--save_step_interval',\n      type=int,\n      default=10,\n      help='Save model checkpoint every save_steps.')\n  parser.add_argument(\n      '--start_checkpoint',\n      type=str,\n      default='',\n      help='If specified, restore this pretrained model before any training.')\n  parser.add_argument(\n      '--model_architecture',\n      type=str,\n      default='conv',\n      help='What model architecture to use')\n  parser.add_argument(\n      '--check_nans',\n      type=bool,\n      default=False,\n      help='Whether to check for invalid numbers during processing')\n  parser.add_argument(\n      '--quantize',\n      type=bool,\n      default=False,\n      help='Whether to train the model for eight-bit deployment')\n  parser.add_argument(\n      '--preprocess',\n      type=str,\n      default='mfcc',\n      help='Spectrogram processing mode. Can be \"mfcc\", \"average\", or \"micro\"')\n  parser.add_argument(\n      '--data_url',\n      type=str,\n      default='Audio\/training_data',\n      help='Directory where training data resides')\n\n  # Function used to parse --verbosity argument\n  def verbosity_arg(value):\n    \"\"\"Parses verbosity argument.\n\n    Args:\n      value: A member of tf.logging.\n    Raises:\n      ArgumentTypeError: Not an expected value.\n    \"\"\"\n    value = value.upper()\n    if value == 'INFO':\n      return tf.compat.v1.logging.INFO\n    elif value == 'DEBUG':\n      return tf.compat.v1.logging.DEBUG\n    elif value == 'ERROR':\n      return tf.compat.v1.logging.ERROR\n    elif value == 'FATAL':\n      return tf.compat.v1.logging.FATAL\n    elif value == 'WARN':\n      return tf.compat.v1.logging.WARN\n    else:\n      raise argparse.ArgumentTypeError('Not an expected value')\n  parser.add_argument(\n      '--verbosity',\n      type=verbosity_arg,\n      default=tf.compat.v1.logging.INFO,\n      help='Log verbosity. Can be \"INFO\", \"DEBUG\", \"ERROR\", \"FATAL\", or \"WARN\"')\n\n  FLAGS, unparsed = parser.parse_known_args()\n  tf.compat.v1.app.run(main=main, argv=[sys.argv[0]] + unparsed)\n<\/code><\/pre>\n\n<p>My question is, how do I actually classify a wav file against this model?<\/p>\n","answers":[{"AnswerId":"57921462","CreationDate":"2019-09-13T10:00:17.567","ParentId":null,"OwnerUserId":"10606534","Title":null,"Body":"<p>it's a simple argmax classfy  you just need to get the softmax predict output by add it to sess run params<\/p>\n\n<pre><code> pred , train_summary, train_accuracy, cross_entropy_value, _, _ = sess.run(\n        [\n            predicted_indices ,\n            merged_summaries,\n            evaluation_step,\n            cross_entropy_mean,\n            train_step,\n            increment_global_step,\n        ],\n<\/code><\/pre>\n\n<p>then compare the pred array with train_ground_truth which should be the dict or array container audio label.<\/p>\n"}]}
+{"QuestionId":57905600,"AnswerCount":1,"Tags":"<python><parsing><pytorch>","CreationDate":"2019-09-12T11:20:20.507","AcceptedAnswerId":null,"OwnerUserId":11739632.0,"Title":"ImportError: cannot import name 'parameter_parser' from 'parser' (unknown location)","Body":"<p>I am trying to Import Parameter_parser from Parser. but it is showing the error below:<\/p>\n\n<pre><code>ImportError: cannot import name 'parameter_parser' from 'parser'\n<\/code><\/pre>\n\n<p>In the line below that I also get:<\/p>\n\n<pre><code>ModuleNotFoundError: No module named 'load_data'\n<\/code><\/pre>\n\n<p>This is my code:<\/p>\n\n<pre><code>    import matplotlib \n    matplotlib.use('agg') \n    import numpy as np \n    import time \n    import os \n    import torch.utils.data \n    import torch.nn.functional as F \n    import torch.optim as optim \n    import torch.optim.lr_scheduler as lr_scheduler \n    from torch.utils.data import DataLoader \n    from os.path import join as pjoin \n    from parser import parameter_parser \n    from load_data import split_ids, GraphData, collate_batch \n    from models.gcn_modify import GCN_MODIFY \n    from models.gcn_origin import GCN_ORIGIN \n    from models.gat import GAT \n    from models.mgcn import MGCN \n    from sklearn import metrics`\n<\/code><\/pre>\n","answers":[{"AnswerId":"57910734","CreationDate":"2019-09-12T16:10:03.443","ParentId":null,"OwnerUserId":"1807569","Title":null,"Body":"<p>When I try the same thing in my python console I get this:<\/p>\n\n<pre><code>&gt;&gt;&gt;  from parser import parameter_parser\n\n\nFile \"&lt;stdin&gt;\", line 1\n    from parser import parameter_parser\n    ^\nIndentationError: unexpected indent\n&gt;&gt;&gt; from parser import parameter_parser\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\nImportError: cannot import name parameter_parser\n<\/code><\/pre>\n\n<p>Is this the same problem for you? This is because you don't have the module installed via pip (<code>pip install PACKAGE_NAME<\/code>) or whatever you use to install your packages. Another idea is that you have set up a virtual env, installed it there and did not activate it.<\/p>\n\n<p>At any case although I didn't downvote your answer (I think there are no wrong questions!) I assume the person that did was not able to find the additional information to help you solve your problem. for next time try adding what OS you are using, what package is causing the problem and what solutions have you already tried (did you find other answers on stackoverflow? did you google the problem? did you try importing the package by itself in the console?).<\/p>\n"}]}
+{"QuestionId":57905851,"AnswerCount":0,"Tags":"<python><tensorflow><inference>","CreationDate":"2019-09-12T11:34:54.703","AcceptedAnswerId":null,"OwnerUserId":5380012.0,"Title":"Slower inference speed after freezing graph in Tensorflow","Body":"<p>I'm trying to optimize for inference <a href=\"https:\/\/github.com\/Rayhane-mamah\/Tacotron-2\" rel=\"nofollow noreferrer\">this<\/a> implementation of Tacotron. After freezing and\/or optimizing it takes about 30 seconds for evaluating one sentence instead of 2-3 seconds before \"optimization\".  <\/p>\n\n<p>I'm following <a href=\"https:\/\/leimao.github.io\/blog\/Save-Load-Inference-From-TF-Frozen-Graph\/\" rel=\"nofollow noreferrer\">this<\/a> (or any similar) guide. Just using <code>convert_variables_to_constants<\/code> function to freeze graph and then <code>optimize_for_inference_lib.optimize_for_inference<\/code> for optimizing.<\/p>\n\n<p>To run evaluation I'm importing the graph and call <code>sess.run()<\/code>. This call is dramatically slower after \"optimization\"<\/p>\n\n<p>What am I doing wrong?<\/p>\n","answers":[]}
+{"QuestionId":57906490,"AnswerCount":1,"Tags":"<docker><tensorflow><keras><docker-swarm><nvidia-docker>","CreationDate":"2019-09-12T12:14:35.583","AcceptedAnswerId":58803991.0,"OwnerUserId":1514114.0,"Title":"Docker service with GPU configured in compose file; no GPU recognized by Keras","Body":"<p>I've got a multi-service application configured in a v3.5 docker compose file.<\/p>\n\n<p>One of the services is to have access to the (one) GPU on the (one) node in the swarm.\nHowever, if I start the service using the docker compose file, I don't seem to have access to the GPU, as reported by keras:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>   import keras\n   from tensorflow.python.client import device_lib\n   print(device_lib.list_local_devices())\n<\/code><\/pre>\n\n<p>prints<\/p>\n\n<pre><code>Using TensorFlow backend.\n\n[name: \"\/device:CPU:0\"\n device_type: \"CPU\"\n memory_limit: 268435456\n locality {  }\n incarnation: 10790773049987428954,\n name: \"\/device:XLA_CPU:0\"\n device_type: \"XLA_CPU\"\n memory_limit: 17179869184\n locality {  }\n incarnation: 239154712796449863\n physical_device_desc: \"device: XLA_CPU device\"]\n<\/code><\/pre>\n\n<p>If I run the same image from the command line like this:<\/p>\n\n<p><code>docker run -it --rm $(ls \/dev\/nvidia* | xargs -I{} echo '--device={}') $(ls \/usr\/lib\/*-linux-gnu\/{libcuda,libnvidia}* | xargs -I{} echo '-v {}:{}:ro') -v $(pwd):\/srv --entrypoint \/bin\/bash ${MY_IMG}<\/code><\/p>\n\n<p>The output is<\/p>\n\n<pre><code>[name: \"\/device:CPU:0\"\n device_type: \"CPU\"\n memory_limit: 268435456\n locality {\n }\n incarnation: 3178082198631681841, name: \"\/device:XLA_CPU:0\"\n device_type: \"XLA_CPU\"\n memory_limit: 17179869184\n locality {\n }\n incarnation: 15685155444461741733\n physical_device_desc: \"device: XLA_CPU device\", name: \"\/device:XLA_GPU:0\"\n device_type: \"XLA_GPU\"\n memory_limit: 17179869184\n locality {\n }\n incarnation: 4056441191345727860\n physical_device_desc: \"device: XLA_GPU device\"]\n<\/code><\/pre>\n\n<p><strong>Config:<\/strong><\/p>\n\n<p>I've installed nvidia-docker and configured the node according to <a href=\"https:\/\/medium.com\/@kyeongwook.ma\/deploy-nvidia-docker-to-docker-swarm-service-in-10-mins-4ba603048e85\" rel=\"nofollow noreferrer\">this guide<\/a>:<\/p>\n\n<pre><code>\/etc\/systemd\/system\/docker.service.d\/override.conf:\n\n[Service]\nExecStart=\nExecStart=\/usr\/bin\/dockerd -H fd:\/\/ --containerd=\/run\/containerd\/containerd.sock --default-runtime=nvidia --node-generic-resource gpu=GPU-b7ad85d5\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>\/etc\/nvidia-container-runtime\/config.toml:\n\ndisable-require = false\nswarm-resource = \"DOCKER_RESOURCE_GPU\"\n\n[nvidia-container-cli]\n#root = \"\/run\/nvidia\/driver\"\n#path = \"\/usr\/bin\/nvidia-container-cli\"\nenvironment = []\n#debug = \"\/var\/log\/nvidia-container-toolkit.log\"\n#ldcache = \"\/etc\/ld.so.cache\"\nload-kmods = true\n#no-cgroups = false\n#user = \"root:video\"\nldconfig = \"@\/sbin\/ldconfig.real\"\n\n[nvidia-container-runtime]\n#debug = \"\/var\/log\/nvidia-container-runtime.log\"\n<\/code><\/pre>\n\n<p>The relevant part of the docker compose file:<\/p>\n\n<pre><code>docker-compose.yaml:\n\nversion: '3.5'\n   ...\n   services:\n        ...\n        my-service:\n            ...\n            deploy:\n              resources:\n                reservations:\n                  generic_resources:\n                    - discrete_resource_spec:\n                        kind: 'gpu'\n                        value: 1\n<\/code><\/pre>\n\n<p><strong>Question:<\/strong> \nWhat else is needed to get access to the GPU in that docker service?<\/p>\n","answers":[{"AnswerId":"58803991","CreationDate":"2019-11-11T15:11:41.120","ParentId":null,"OwnerUserId":"9177493","Title":null,"Body":"<p>NVIDIA-Docker is only working on Docker Compose 2.3<\/p>\n\n<p>Change version to <code>version: '2.3'<\/code><\/p>\n\n<p><a href=\"https:\/\/github.com\/NVIDIA\/nvidia-docker\/wiki#do-you-support-docker-compose\" rel=\"nofollow noreferrer\">https:\/\/github.com\/NVIDIA\/nvidia-docker\/wiki#do-you-support-docker-compose<\/a>.<\/p>\n"}]}
+{"QuestionId":57906768,"AnswerCount":0,"Tags":"<python><tensorflow><gradient>","CreationDate":"2019-09-12T12:30:13.710","AcceptedAnswerId":null,"OwnerUserId":12055533.0,"Title":"The output of tape.gradient cannot be applied in optimizer.apply_gradients","Body":"<p>The output of tape.gradient cannot be applied in optimizer.apply_gradients<\/p>\n\n<p>I am using a strange cost function and need to compute the gradients and apply them explicitly.<\/p>\n\n<p>The format of the function calls is strange. I was expecting to be able to apply directly the output of <\/p>\n\n<pre><code>dw = t.gradient( cost, wl, unconnected_gradients=tf.UnconnectedGradients.ZERO)\n<\/code><\/pre>\n\n<p>into<\/p>\n\n<pre><code>optimizer.apply_gradients(dw, wl)  # optimizer = Adam\n<\/code><\/pre>\n\n<p>If wl is not a vector but a list of variables it works. But as I want to make a bigger system I don't want to define the variables independently.<\/p>\n\n<p>dw = t.gradient( cost, wl, unconnected_gradients=tf.UnconnectedGradients.ZERO)\noptimizer.apply_gradients(dw, wl)  # optimizer = Adam<\/p>\n\n<p>The content of the variables is:`<\/p>\n\n<pre><code>print(dw)\ntf.Tensor(\n[[  12.000002]\n [  10.000002]\n [  98.00002 ]\n [ 250.00003 ]\n [1594.      ]], shape=(5, 1), dtype=float32)\n\nprint(wl)\n&lt;tf.Variable 'Shape\/Variable:0' shape=(5, 1) dtype=float32, numpy=\narray([[1.],\n       [1.],\n       [1.],\n       [1.],\n       [1.]], dtype=float32)&gt;\n<\/code><\/pre>\n\n<pre><code>---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-144-a875badce41c&gt; in &lt;module&gt;()\n     41 #     print(d4y_dx4)\n     42     print(dw)\n---&gt; 43     optimizer.apply_gradients(dw, wl)  # optimizer = Adam\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/training\/optimizer.py in apply_gradients(self, grads_and_vars, global_step, name)\n    574       raise ValueError(\"No variables provided.\")\n    575     converted_grads_and_vars = []\n--&gt; 576     for g, v in grads_and_vars:\n    577       if g is not None:\n    578         try:\n\nValueError: not enough values to unpack (expected 2, got 1)\n<\/code><\/pre>\n\n<p>Using the zip solution does not work either<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-146-d722d545815d&gt; in &lt;module&gt;()\n     41 #     print(d4y_dx4)\n     42     print(dw)\n---&gt; 43     optimizer.apply_gradients(zip(dw, wl))  # optimizer = Adam\n     44     #tf.print(wl)\n     45     #tf.print(tf.math.reduce_mean((2*wl[2] + 6 * wl[3] * x + 12*wl[4] * x**2-1)**2))\n\nTypeError: zip argument #2 must support iteration\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57907321,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><ssl-certificate><urllib><ubuntu-18.04>","CreationDate":"2019-09-12T12:59:47.967","AcceptedAnswerId":null,"OwnerUserId":10258933.0,"Title":"ubuntu, python, ssl.SSLError: [SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed (_ssl.c:852)","Body":"<p>I am running some sample code to test out the installation of Tensorflow for running on a GPU. Python works fine andto  I am able to import Tensorflow without any errors.<\/p>\n\n<p>However, when I try to load a dataset from the Tesorflow website, I run into problems and I receive the error messages below.<\/p>\n\n<p>The problem is related to some SSL certificates, something I have little knowledge of and therefor I am reaching out for help.<\/p>\n\n<p><em>(Relevant)<\/em> <strong>CODE<\/strong>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\n(x_train, y_train),(x_test, y_test) = tf.keras.datasets.mnist.load_data()\n<\/code><\/pre>\n\n<p><strong>ERROR MESSAGES<\/strong>:<\/p>\n\n<pre><code>urllib.error.URLError: &lt;urlopen error [SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed (_ssl.c:852)&gt;\n\n...\n\nurllib.error.URLError: &lt;urlopen error [SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed (_ssl.c:852)&gt;\n\n...\n\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:852)\n<\/code><\/pre>\n\n<p><strong>FULL TRACEBACK<\/strong>:<\/p>\n\n<pre><code>Downloading data from https:\/\/storage.googleapis.com\/tensorflow\/tf-keras-datasets\/mnist.npz\nTraceback (most recent call last):\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/urllib\/request.py\", line 1318, in do_open\n    encode_chunked=req.has_header('Transfer-encoding'))\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/http\/client.py\", line 1254, in request\n    self._send_request(method, url, body, headers, encode_chunked)\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/http\/client.py\", line 1300, in _send_request\n    self.endheaders(body, encode_chunked=encode_chunked)\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/http\/client.py\", line 1249, in endheaders\n    self._send_output(message_body, encode_chunked=encode_chunked)\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/http\/client.py\", line 1036, in _send_output\n    self.send(msg)\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/http\/client.py\", line 974, in send\n    self.connect()\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/http\/client.py\", line 1415, in connect\n    server_hostname=server_hostname)\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/ssl.py\", line 407, in wrap_socket\n    _context=self, _session=session)\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/ssl.py\", line 817, in __init__\n    self.do_handshake()\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/ssl.py\", line 1077, in do_handshake\n    self._sslobj.do_handshake()\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/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:852)\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/utils\/data_utils.py\", line 251, in get_file\n    urlretrieve(origin, fpath, dl_progress)\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/urllib\/request.py\", line 248, in urlretrieve\n    with contextlib.closing(urlopen(url, data)) as fp:\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/urllib\/request.py\", line 223, in urlopen\n    return opener.open(url, data, timeout)\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/urllib\/request.py\", line 526, in open\n    response = self._open(req, data)\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/urllib\/request.py\", line 544, in _open\n    '_open', req)\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/urllib\/request.py\", line 504, in _call_chain\n    result = func(*args)\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/urllib\/request.py\", line 1361, in https_open\n    context=self._context, check_hostname=self._check_hostname)\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/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:852)&gt;\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"\/home\/filip\/PycharmProjects\/mist_test\/mist_test_file.py\", line 2, in &lt;module&gt;\n    (x_train, y_train),(x_test, y_test) = tf.keras.datasets.mnist.load_data()\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/datasets\/mnist.py\", line 49, in load_data\n    file_hash='8a61469f7ea1b51cbae51d4f78837e45')\n  File \"\/home\/filip\/anaconda3\/envs\/tf\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/utils\/data_utils.py\", line 255, 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:852)\n<\/code><\/pre>\n\n<p><strong>SYSTEM INFO<\/strong>:<\/p>\n\n<pre><code># packages in environment at \/home\/filip\/anaconda3\/envs\/tf:\n#\n# Name                    Version                   Build  Channel\n_libgcc_mutex             0.1                        main  \nabsl-py                   0.8.0                    pypi_0    pypi\nastor                     0.8.0                    pypi_0    pypi\nca-certificates           2019.5.15                     1  \ncertifi                   2019.6.16                py36_1  \ngast                      0.3.1                    pypi_0    pypi\ngoogle-pasta              0.1.7                    pypi_0    pypi\ngrpcio                    1.23.0                   pypi_0    pypi\nh5py                      2.10.0                   pypi_0    pypi\nkeras-applications        1.0.8                    pypi_0    pypi\nkeras-preprocessing       1.1.0                    pypi_0    pypi\nlibedit                   3.1.20181209         hc058e9b_0  \nlibffi                    3.2.1                hd88cf55_4  \nlibgcc-ng                 9.1.0                hdf63c60_0  \nlibstdcxx-ng              9.1.0                hdf63c60_0  \nmarkdown                  3.1.1                    pypi_0    pypi\nncurses                   6.1                  he6710b0_1  \nnumpy                     1.17.2                   pypi_0    pypi\nopenssl                   1.1.1d               h7b6447c_0  \npip                       19.2.2                   py36_0  \nprotobuf                  3.9.1                    pypi_0    pypi\npython                    3.6.9                h265db76_0  \nreadline                  7.0                  h7b6447c_5  \nsetuptools                41.0.1                   py36_0  \nsix                       1.12.0                   pypi_0    pypi\nsqlite                    3.29.0               h7b6447c_0  \ntensorboard               1.14.0                   pypi_0    pypi\ntensorflow-estimator      1.14.0                   pypi_0    pypi\ntensorflow-gpu            1.14.0                   pypi_0    pypi\ntermcolor                 1.1.0                    pypi_0    pypi\ntk                        8.6.8                hbc83047_0  \nwerkzeug                  0.15.6                   pypi_0    pypi\nwheel                     0.33.4                   py36_0  \nwrapt                     1.11.2                   pypi_0    pypi\nxz                        5.2.4                h14c3975_4  \nzlib                      1.2.11               h7b6447c_3\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57907411,"AnswerCount":0,"Tags":"<pytorch><yolo>","CreationDate":"2019-09-12T13:04:01.537","AcceptedAnswerId":null,"OwnerUserId":1684315.0,"Title":"Yolov3 - Concatenate output layers to speed up training","Body":"<p>I am using Yolov3 in Pytorch and have got 5 classes. Each class consists of approx. 500 images and after 2.500 iterations I get an mAP of 0.80%. However, the training takes approx. 2 days and I am strongly wondering how to decrease the training time.<\/p>\n\n<p>One idea is to take a pretrained yolov3 network weights file, concatenate the three output layers and use these layers for training.<\/p>\n\n<p>What do you think of this idea? Or will this not work?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
+{"QuestionId":57907556,"AnswerCount":0,"Tags":"<python><tensorflow><keras><cudnn>","CreationDate":"2019-09-12T13:13:01.217","AcceptedAnswerId":null,"OwnerUserId":12058424.0,"Title":"CUDA_ERROR_OUT_OF_MEMORY with batch_size 1 and 256x256x3 images on pix2pix gan","Body":"<p>When running the Jason Brownlee example code of the pix2pix gan I run into an OOM error, but it still starts training and the results are actually good.<\/p>\n\n<p>Why is it? And is there a way to solve this knowing that I can't reduce the batch_size anymore and also the image resolution?<\/p>\n\n<p>(Already tried using FP16 but error still persisted)<\/p>\n\n<p>This is the error that is displayed:<\/p>\n\n<pre><code>019-09-12 14:00:37.770651: E tensorflow\/stream_executor\/cuda\/cuda_driver.cc:828] failed to allocate 3.78G (4056085248 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY: out of memory\n<\/code><\/pre>\n\n<p>btw I'm using an RTX2070 with 8GB VRAM<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># -- COMMANDS NEEDED FOR GPU TO WORK -- #\nimport tensorflow as tf\nimport keras.backend as K\nconfig = tf.ConfigProto()\nconfig.gpu_options.allow_growth = True\nK.set_session(tf.Session(config=config))\n# ------------------------------------- #\n\n# example of pix2pix gan for satellite to map image-to-image translation\nfrom numpy import load\nfrom numpy import zeros\nfrom numpy import ones\nfrom numpy.random import randint\nfrom keras.optimizers import Adam\nfrom keras.initializers import RandomNormal\nfrom keras.models import Model\nfrom keras.models import Input\nfrom keras.layers import Conv2D\nfrom keras.layers import Conv2DTranspose\nfrom keras.layers import LeakyReLU\nfrom keras.layers import Activation\nfrom keras.layers import Concatenate\nfrom keras.layers import Dropout\nfrom keras.layers import BatchNormalization\nfrom keras.layers import LeakyReLU\nfrom matplotlib import pyplot\nimport os\n\n\n# define the discriminator model\ndef define_discriminator(image_shape):\n    # weight initialization\n    init = RandomNormal(stddev=0.02)\n    # source image input\n    in_src_image = Input(shape=image_shape)\n    # target image input\n    in_target_image = Input(shape=image_shape)\n    # concatenate images channel-wise\n    merged = Concatenate()([in_src_image, in_target_image])\n    # C64\n    d = Conv2D(64, (4,4), strides=(2,2), padding='same', kernel_initializer=init)(merged)\n    d = LeakyReLU(alpha=0.2)(d)\n    # C128\n    d = Conv2D(128, (4,4), strides=(2,2), padding='same', kernel_initializer=init)(d)\n    d = BatchNormalization()(d)\n    d = LeakyReLU(alpha=0.2)(d)\n    # C256\n    d = Conv2D(256, (4,4), strides=(2,2), padding='same', kernel_initializer=init)(d)\n    d = BatchNormalization()(d)\n    d = LeakyReLU(alpha=0.2)(d)\n    # C512\n    d = Conv2D(512, (4,4), strides=(2,2), padding='same', kernel_initializer=init)(d)\n    d = BatchNormalization()(d)\n    d = LeakyReLU(alpha=0.2)(d)\n    # second last output layer\n    d = Conv2D(512, (4,4), padding='same', kernel_initializer=init)(d)\n    d = BatchNormalization()(d)\n    d = LeakyReLU(alpha=0.2)(d)\n    # patch output\n    d = Conv2D(1, (4,4), padding='same', kernel_initializer=init)(d)\n    patch_out = Activation('sigmoid')(d)\n    # define model\n    model = Model([in_src_image, in_target_image], patch_out)\n    # compile model\n    opt = Adam(lr=0.0002, beta_1=0.5)\n    model.compile(loss='binary_crossentropy', optimizer=opt, loss_weights=[0.5])\n    return model\n\n# define an encoder block\ndef define_encoder_block(layer_in, n_filters, batchnorm=True):\n    # weight initialization\n    init = RandomNormal(stddev=0.02)\n    # add downsampling layer\n    g = Conv2D(n_filters, (4,4), strides=(2,2), padding='same', kernel_initializer=init)(layer_in)\n    # conditionally add batch normalization\n    if batchnorm:\n        g = BatchNormalization()(g, training=True)\n    # leaky relu activation\n    g = LeakyReLU(alpha=0.2)(g)\n    return g\n\n# define a decoder block\ndef decoder_block(layer_in, skip_in, n_filters, dropout=True):\n    # weight initialization\n    init = RandomNormal(stddev=0.02)\n    # add upsampling layer\n    g = Conv2DTranspose(n_filters, (4,4), strides=(2,2), padding='same', kernel_initializer=init)(layer_in)\n    # add batch normalization\n    g = BatchNormalization()(g, training=True)\n    # conditionally add dropout\n    if dropout:\n        g = Dropout(0.5)(g, training=True)\n    # merge with skip connection\n    g = Concatenate()([g, skip_in])\n    # relu activation\n    g = Activation('relu')(g)\n    return g\n\n# define the standalone generator model\ndef define_generator(image_shape=(256,256,3)):\n    # weight initialization\n    init = RandomNormal(stddev=0.02)\n    # image input\n    in_image = Input(shape=image_shape)\n    # encoder model\n    e1 = define_encoder_block(in_image, 64, batchnorm=False)\n    e2 = define_encoder_block(e1, 128)\n    e3 = define_encoder_block(e2, 256)\n    e4 = define_encoder_block(e3, 512)\n    e5 = define_encoder_block(e4, 512)\n    e6 = define_encoder_block(e5, 512)\n    e7 = define_encoder_block(e6, 512)\n    # bottleneck, no batch norm and relu\n    b = Conv2D(512, (4,4), strides=(2,2), padding='same', kernel_initializer=init)(e7)\n    b = Activation('relu')(b)\n    # decoder model\n    d1 = decoder_block(b, e7, 512)\n    d2 = decoder_block(d1, e6, 512)\n    d3 = decoder_block(d2, e5, 512)\n    d4 = decoder_block(d3, e4, 512, dropout=False)\n    d5 = decoder_block(d4, e3, 256, dropout=False)\n    d6 = decoder_block(d5, e2, 128, dropout=False)\n    d7 = decoder_block(d6, e1, 64, dropout=False)\n    # output\n    g = Conv2DTranspose(3, (4,4), strides=(2,2), padding='same', kernel_initializer=init)(d7)\n    out_image = Activation('tanh')(g)\n    # define model\n    model = Model(in_image, out_image)\n    return model\n\n# define the combined generator and discriminator model, for updating the generator\ndef define_gan(g_model, d_model, image_shape):\n    # make weights in the discriminator not trainable\n    d_model.trainable = False\n    # define the source image\n    in_src = Input(shape=image_shape)\n    # connect the source image to the generator input\n    gen_out = g_model(in_src)\n    # connect the source input and generator output to the discriminator input\n    dis_out = d_model([in_src, gen_out])\n    # src image as input, generated image and classification output\n    model = Model(in_src, [dis_out, gen_out])\n    # compile model\n    opt = Adam(lr=0.0002, beta_1=0.5)\n    model.compile(loss=['binary_crossentropy', 'mae'], optimizer=opt, loss_weights=[1,100])\n    return model\n\n# load and prepare training images\ndef load_real_samples(filename):\n    # load compressed arrays\n    data = load(filename)\n    # unpack arrays\n    X1, X2 = data['arr_0'], data['arr_1']\n    # scale from [0,255] to [-1,1]\n    X1 = (X1 - 127.5) \/ 127.5\n    X2 = (X2 - 127.5) \/ 127.5\n    return [X1, X2]\n\n# select a batch of random samples, returns images and target\ndef generate_real_samples(dataset, n_samples, patch_shape):\n    # unpack dataset\n    trainA, trainB = dataset\n    # choose random instances\n    ix = randint(0, trainA.shape[0], n_samples)\n    # retrieve selected images\n    X1, X2 = trainA[ix], trainB[ix]\n    # generate 'real' class labels (1)\n    y = ones((n_samples, patch_shape, patch_shape, 1))\n    return [X1, X2], y\n\n# generate a batch of images, returns images and targets\ndef generate_fake_samples(g_model, samples, patch_shape):\n    # generate fake instance\n    X = g_model.predict(samples)\n    # create 'fake' class labels (0)\n    y = zeros((len(X), patch_shape, patch_shape, 1))\n    return X, y\n\n# generate samples and save as a plot and save the model\ndef summarize_performance(step, g_model, dataset, n_samples=3):\n    # select a sample of input images\n    [X_realA, X_realB], _ = generate_real_samples(dataset, n_samples, 1)\n    # generate a batch of fake samples\n    X_fakeB, _ = generate_fake_samples(g_model, X_realA, 1)\n    # scale all pixels from [-1,1] to [0,1]\n    X_realA = (X_realA + 1) \/ 2.0\n    X_realB = (X_realB + 1) \/ 2.0\n    X_fakeB = (X_fakeB + 1) \/ 2.0\n    # plot real source images\n    for i in range(n_samples):\n        pyplot.subplot(3, n_samples, 1 + i)\n        pyplot.axis('off')\n        pyplot.imshow(X_realA[i])\n    # plot generated target image\n    for i in range(n_samples):\n        pyplot.subplot(3, n_samples, 1 + n_samples + i)\n        pyplot.axis('off')\n        pyplot.imshow(X_fakeB[i])\n    # plot real target image\n    for i in range(n_samples):\n        pyplot.subplot(3, n_samples, 1 + n_samples*2 + i)\n        pyplot.axis('off')\n        pyplot.imshow(X_realB[i])\n    # save plot to file\n    filename1 = 'results\/plot_%06d.png' % (step+1)\n    pyplot.savefig(filename1)\n    pyplot.close()\n    # save the generator model\n    filename2 = 'results\/model_%06d.h5' % (step+1)\n    g_model.save(filename2)\n    print('&gt;Saved: %s and %s' % (filename1, filename2))\n\n# train pix2pix models\ndef train(d_model, g_model, gan_model, dataset, n_epochs=100, n_batch=1):\n    # determine the output square shape of the discriminator\n    n_patch = d_model.output_shape[1]\n    # unpack dataset\n    trainA, trainB = dataset\n    # calculate the number of batches per training epoch\n    bat_per_epo = int(len(trainA) \/ n_batch)\n    # calculate the number of training iterations\n    n_steps = bat_per_epo * n_epochs\n    print('batch per epoch ', bat_per_epo)\n    # manually enumerate epochs\n    for i in range(n_steps):\n        # select a batch of real samples\n        [X_realA, X_realB], y_real = generate_real_samples(dataset, n_batch, n_patch)\n        # generate a batch of fake samples\n        X_fakeB, y_fake = generate_fake_samples(g_model, X_realA, n_patch)\n        # update discriminator for real samples\n        d_loss1 = d_model.train_on_batch([X_realA, X_realB], y_real)\n        # update discriminator for generated samples\n        d_loss2 = d_model.train_on_batch([X_realA, X_fakeB], y_fake)\n        # update the generator\n        g_loss, _, _ = gan_model.train_on_batch(X_realA, [y_real, X_realB])\n        # summarize performance\n        print('&gt;%d, d1[%.3f] d2[%.3f] g[%.3f]' % (i+1, d_loss1, d_loss2, g_loss), end='\\r')\n        # summarize model performance\n        if (i+1) % (bat_per_epo * 10) == 0:\n            print('&gt;%d, d1[%.3f] d2[%.3f] g[%.3f]' % (i+1, d_loss1, d_loss2, g_loss))\n            summarize_performance(i, g_model, dataset)\n\n# load image data\ndataset = load_real_samples('maps_256.npz')\nprint('Loaded', dataset[0].shape, dataset[1].shape)\n# define input shape based on the loaded dataset\nimage_shape = dataset[0].shape[1:]\n# define the models\nd_model = define_discriminator(image_shape)\ng_model = define_generator(image_shape)\n# define the composite model\ngan_model = define_gan(g_model, d_model, image_shape)\ngan_model.summary()\n# train model\ntrain(d_model, g_model, gan_model, dataset)\n<\/code><\/pre>\n\n<p>To download the dataset<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># load, split and scale the maps dataset ready for training\nfrom os import listdir\nfrom numpy import asarray\nfrom numpy import vstack\nfrom keras.preprocessing.image import img_to_array\nfrom keras.preprocessing.image import load_img\nfrom numpy import savez_compressed\n\n# load all images in a directory into memory\ndef load_images(path, size=(256,512)):\n    src_list, tar_list = list(), list()\n    # enumerate filenames in directory, assume all are images\n    for filename in listdir(path):\n        # load and resize the image\n        pixels = load_img(path + filename, target_size=size)\n        # convert to numpy array\n        pixels = img_to_array(pixels)\n        # split into satellite and map\n        sat_img, map_img = pixels[:, :256], pixels[:, 256:]\n        src_list.append(sat_img)\n        tar_list.append(map_img)\n    return [asarray(src_list), asarray(tar_list)]\n\n# dataset path\npath = '.\/maps\/train\/'\n# load dataset\n[src_images, tar_images] = load_images(path)\nprint('Loaded: ', src_images.shape, tar_images.shape)\n# save as compressed numpy array\nfilename = 'maps_256.npz'\nsavez_compressed(filename, src_images, tar_images)\nprint('Saved dataset: ', filename) \n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57907702,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras><face-recognition><keras-vggface>","CreationDate":"2019-09-12T13:22:16.657","AcceptedAnswerId":null,"OwnerUserId":347484.0,"Title":"Keras VGGFace celebrities names encoding","Body":"<p>I run TF 1.13 with keras and keras-vggface 0.6<\/p>\n\n<pre><code>results = decode_predictions(y)\n\nfor result in results[0]:\n    confidence = result[1] * 100\n    #log_str += '%s: %.3f%%' % (result[0], confidence)\n    name = result[0].replace(\"'\", \"\").replace(\"b \", \"\").replace(\"_\", \" \")\n\n# the name is a str object \"b' Emir_Had\\xc5\\xbeihafizbegovi\\xc4\\x87'\" \n# (which looks like byte literal but its not)\n\n# I strip leftovers and make it:\ns = str(\"Emir Had\\xc5\\xbeihafizbegovi\\xc4\\x87\")\nb = s.encode()\nprint(b.decode('utf-8'))\n<\/code><\/pre>\n\n<p>and I get <code>Emir Had\u00c5\u00beihafizbegovi\u00c4<\/code> while Wikipedia shows <code>Emir Had\u017eihafizbegovi\u0107<\/code><\/p>\n\n<p>So I guess it wasn't utf-8.. If I change to utf-16 it gives my hieroglyphs.. <\/p>\n","answers":[]}
+{"QuestionId":57907926,"AnswerCount":0,"Tags":"<tensorflow><tensorflow2.0>","CreationDate":"2019-09-12T13:35:07.910","AcceptedAnswerId":null,"OwnerUserId":287238.0,"Title":"How to manually clear the tf.function caches (or manage the max size) in tensorflow 2.0?","Body":"<p>The example below shows a simple way of clearing the caches manually. Is there a more standard\/stable way of managing the cache going forward? Or perhaps a pattern to avoid this in the first place? <\/p>\n\n<p>We have some case where the batch-size varies a lot and we hit memory issues as I the def_fun is not going out of scope and the cache is likely not clearing.<\/p>\n\n<pre><code>In [164]: @tf.function\n     ...: def f(x):\n     ...:     return dict(something=x ** 2)\n     ...:\n     ...:\n     ...:\n\nIn [165]: f._list_all_concrete_functions_for_serialization()\nOut[165]: []\n\nIn [166]: _ = f(tf.convert_to_tensor(np.random.randn(109, 3).astype(np.float32)))\n\nIn [167]: _ = f(tf.convert_to_tensor(np.random.randn(111, 3).astype(np.float32)))\n\nIn [168]: f._list_all_concrete_functions_for_serialization()\nOut[168]:\n[&lt;tensorflow.python.eager.function.ConcreteFunction at 0x7fac73e0d358&gt;,\n &lt;tensorflow.python.eager.function.ConcreteFunction at 0x7fac71d41a58&gt;]\n\nIn [169]: f._stateful_fn._function_cache._garbage_collectors\nOut[169]:\n[&lt;tensorflow.python.eager.function._FunctionGarbageCollector at 0x7fac94252390&gt;,\n &lt;tensorflow.python.eager.function._FunctionGarbageCollector at 0x7fac7b0c6048&gt;,\n &lt;tensorflow.python.eager.function._FunctionGarbageCollector at 0x7fac7b0c6d68&gt;]\n\nIn [170]: f._stateful_fn._function_cache._garbage_collectors[0]\nOut[170]: &lt;tensorflow.python.eager.function._FunctionGarbageCollector at 0x7fac94252390&gt;\n\nIn [171]: f._stateful_fn._function_cache._garbage_collectors[0]._cache\nOut[171]:\nOrderedDict([(CacheKey(input_signature=('UTd1s109-3-u', None), parent_graph=None, device_functions=(), colocation_stack=(), in_cross_replica_context=False),\n              &lt;tensorflow.python.eager.function.ConcreteFunction at 0x7fac7371def0&gt;),\n             (CacheKey(input_signature=('UTd1s111-3-u', None), parent_graph=None, device_functions=(), colocation_stack=(), in_cross_replica_context=False),\n              &lt;tensorflow.python.eager.function.ConcreteFunction at 0x7fac77a514a8&gt;),\n             (CacheKey(input_signature=('URu', (TensorSpec(shape=(111, 3), dtype=tf.float32, name='x'),)), parent_graph=None, device_functions=(), colocation_stack=(), in_cross_replica_context=False),\n              &lt;tensorflow.python.eager.function.ConcreteFunction at 0x7fac73e0d358&gt;),\n             (CacheKey(input_signature=('URu', (TensorSpec(shape=(109, 3), dtype=tf.float32, name='x'),)), parent_graph=None, device_functions=(), colocation_stack=(), in_cross_replica_context=False),\n              &lt;tensorflow.python.eager.function.ConcreteFunction at 0x7fac71d41a58&gt;)])\n\nIn [172]: f._stateful_fn._function_cache._garbage_collectors[0]._cache.popitem()\nOut[172]:\n(CacheKey(input_signature=('URu', (TensorSpec(shape=(109, 3), dtype=tf.float32, name='x'),)), parent_graph=None, device_functions=(), colocation_stack=(), in_cross_replica_context=False),\n &lt;tensorflow.python.eager.function.ConcreteFunction at 0x7fac71d41a58&gt;)\n\nIn [173]: f._stateful_fn._function_cache._garbage_collectors[0]._cache.popitem()\nOut[173]:\n(CacheKey(input_signature=('URu', (TensorSpec(shape=(111, 3), dtype=tf.float32, name='x'),)), parent_graph=None, device_functions=(), colocation_stack=(), in_cross_replica_context=False),\n &lt;tensorflow.python.eager.function.ConcreteFunction at 0x7fac73e0d358&gt;)\n\nIn [174]: f._stateful_fn._function_cache._garbage_collectors[0]._cache.popitem()\nOut[174]:\n(CacheKey(input_signature=('UTd1s111-3-u', None), parent_graph=None, device_functions=(), colocation_stack=(), in_cross_replica_context=False),\n &lt;tensorflow.python.eager.function.ConcreteFunction at 0x7fac77a514a8&gt;)\n\nIn [175]: f._stateful_fn._function_cache._garbage_collectors[0]._cache.popitem()\nOut[175]:\n(CacheKey(input_signature=('UTd1s109-3-u', None), parent_graph=None, device_functions=(), colocation_stack=(), in_cross_replica_context=False),\n &lt;tensorflow.python.eager.function.ConcreteFunction at 0x7fac7371def0&gt;)\n\nIn [176]: f._stateful_fn._function_cache._garbage_collectors[0]._cache.popitem()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57908199,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><machine-learning><object-detection-api>","CreationDate":"2019-09-12T13:49:18.413","AcceptedAnswerId":null,"OwnerUserId":12058610.0,"Title":"In tensor-flow how can I object detect 100s of images rather than one?","Body":"<p>I am using TensorFlow on windows 10 with this '<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>I want to object detect 100s of images at one run time, currently, the code does this;<\/p>\n\n<p>its hardcoded to only take in one image and the image name has to be specified in the code. <\/p>\n\n<p>import os\nimport cv2\nimport numpy as np\nimport tensorflow as tf\nimport sys<\/p>\n\n<pre><code>sys.path.append(\"..\")\n\nfrom utils import label_map_util\nfrom utils import visualization_utils as vis_util\n\nMODEL_NAME = 'inference_graph'\nIMAGE_NAME = 'test11.jpg'\n\nCWD_PATH = os.getcwd()\n\nPATH_TO_CKPT = os.path.join(CWD_PATH,MODEL_NAME,'frozen_inference_graph.pb')\n\nPATH_TO_LABELS = os.path.join(CWD_PATH,'training','labelmap.pbtxt')\n\nPATH_TO_IMAGE = os.path.join(CWD_PATH,IMAGE_NAME)\n\nNUM_CLASSES = 2\n\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\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\nimage_tensor = detection_graph.get_tensor_by_name('image_tensor:0')\n\ndetection_boxes = detection_graph.get_tensor_by_name('detection_boxes:0')\n\ndetection_scores = detection_graph.get_tensor_by_name('detection_scores:0')\ndetection_classes = detection_graph.get_tensor_by_name('detection_classes:0')\n\nnum_detections = detection_graph.get_tensor_by_name('num_detections:0')\n\nimage = cv2.imread(PATH_TO_IMAGE)\nimage_expanded = np.expand_dims(image, axis=0)\n\n(boxes, scores, classes, num) = sess.run(\n    [detection_boxes, detection_scores, detection_classes, num_detections],\n    feed_dict={image_tensor: image_expanded})\n\n\nvis_util.visualize_boxes_and_labels_on_image_array(\n    image,\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=30,\n    min_score_thresh=0.60)\n\ncv2.imshow('Object detector', image)\n\ncv2.waitKey(0)\n\ncv2.destroyAllWindows()\n\n\n\nI want the code to detect a whole folder of images in TensorFlow and export that data, for example, the name of the image and the label name to be exported in a spreadsheet.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57908214,"AnswerCount":0,"Tags":"<python><tensorflow><recurrent-neural-network>","CreationDate":"2019-09-12T13:49:55.490","AcceptedAnswerId":null,"OwnerUserId":10449967.0,"Title":"Tensorflow: tape.gradient() returns None for GRU layer","Body":"<p>I build my model with the following code(tensorflow==1.14):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class Model(tf.keras.Model):\n  def __init__(self):\n    super(Model, self).__init__()\n\n    self.embedding = tf.keras.layers.Embedding(10, 5)\n    self.rnn = tf.keras.layers.GRU(100)  # neither GRU nor LSTM works\n    self.final_layer = tf.keras.layers.Dense(10)\n    self.loss_obj = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True, reduction='none')\n\n  def call(self, inp):\n    inp_em = self.embedding(inp)  # (batch_size, seq_len, embedding_size)\n    inp_enc = self.rnn(inp_em)  # (batch_size, hidden_size)\n    logits = self.final_layer(inp_enc)  # (batch_size, class_num)\n\n    return logits\n\nmodel = Model()\n\ninp = np.random.randint(0, 10, [5, 50], dtype=np.int32)\nout = np.random.randint(0, 10, [5], dtype=np.int32)\n\nwith tf.GradientTape() as tape:\n  logits = model(inp)\n  loss = model.loss_obj(out, logits)\n  print(loss)\n\ngradients = tape.gradient(tf.reduce_mean(loss), model.trainable_variables)\n\nprint('==========  Trainable Variables  ==========')\nfor v in model.trainable_variables:\n  print(v)\n\nprint('==========  Gradients  ==========')\nfor g in gradients:\n    print(g)\n<\/code><\/pre>\n\n<p>but when i print the gridents, the output is :<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>Tensor(\"categorical_crossentropy\/weighted_loss\/Mul:0\", shape=(5,), dtype=float32)\n==========  Trainable Variables  ==========\n&lt;tf.Variable 'model\/embedding\/embeddings:0' shape=(10, 5) dtype=float32&gt;\n&lt;tf.Variable 'model\/gru\/kernel:0' shape=(5, 300) dtype=float32&gt;\n&lt;tf.Variable 'model\/gru\/recurrent_kernel:0' shape=(100, 300) dtype=float32&gt;\n&lt;tf.Variable 'model\/gru\/bias:0' shape=(300,) dtype=float32&gt;\n&lt;tf.Variable 'model\/dense\/kernel:0' shape=(100, 10) dtype=float32&gt;\n&lt;tf.Variable 'model\/dense\/bias:0' shape=(10,) dtype=float32&gt;\n==========  Gradients  ==========\nNone\nNone\nNone\nNone\nTensor(\"MatMul:0\", shape=(100, 10), dtype=float32)\nTensor(\"BiasAddGrad:0\", shape=(10,), dtype=float32)\n<\/code><\/pre>\n\n<p>the gridents for final layer works well, but None for GRU layers and so on.<\/p>\n\n<p>I have tried both <code>tf.keras.layers.LSTM<\/code> and <code>tf.keras.layers.GRU<\/code>, the same problem exists.<\/p>\n\n<h2>Update<\/h2>\n\n<p>finally, I replace <code>tf.GradientTape().gradient()<\/code> with <code>tf.graidents()<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>logits = model(inp)\nloss = model.loss_obj(out, logits)\n\ngradients = tf.gradients(tf.reduce_mean(loss), model.trainable_variables)\n<\/code><\/pre>\n\n<p>the gradients works. But I still don't know what's the difference between these two implements.<\/p>\n","answers":[]}
+{"QuestionId":57908226,"AnswerCount":0,"Tags":"<tensorflow><cudnn>","CreationDate":"2019-09-12T13:50:36.073","AcceptedAnswerId":null,"OwnerUserId":10295417.0,"Title":"tensorflow, tensorflow-gpu not recognizing the GPU","Body":"<p>this might be a repeated post but none of the other posts worked for me. \ni have tensorflow 1.14.0 and tensorflow-gpu 1.14.0 i have already installed the compatible Cuda and Cudnn versions required, and specified their path in environment variables. but still the tensorflow and\/or tensorflow-gpu, don't recognize my GPU.\nso the of the validating code:<\/p>\n\n<pre><code>from tensorflow.python.client import device_lib\ndef get_available_devices():\n    local_device_protos = device_lib.list_local_devices()\n    return [x.name for x in local_device_protos]\n\nprint(get_available_devices()) \n\n<\/code><\/pre>\n\n<p>is: <\/p>\n\n<pre><code>['\/device:CPU:0']\n<\/code><\/pre>\n\n<p>OP: windows 10<\/p>\n\n<p>IDE: Jupyter (anaconda)<\/p>\n\n<p>GPU version: GTX 960M <\/p>\n\n<p>note: some posts mention that you have to uninstall tensorflow at the end, but when i do, the code above doesn't even work...<\/p>\n","answers":[]}
+{"QuestionId":57908550,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-12T14:08:30.533","AcceptedAnswerId":null,"OwnerUserId":12058659.0,"Title":"Fidelity custom loss function in tensorflow - input shape problem","Body":"<p>I am having trouble with inputs to a custom loss function as part of a multi-objective elementary neural network (GRU so far). <\/p>\n\n<p>I have a multi-objective function that takes as input a flattened (,16) shape (4x4) matrix (\"input matrix\") and generates a sequence of 10 4x4 matrices (represented via a flattened output shape (,160)) as a first objective using 'mse' loss. This part of the model, which takes <code>data.lab<\/code> as training data and <code>inputreshape<\/code> as label data works fine. I need to, as a second objective, however convert this (,160) tensor to a (10,4,4) tensor (reshaping to 10 4x4 matrices), then matmul them together to obtain a (4,4) matrix \"output matrix\" (done via the productlayer custom layer). This output matrix then needs to be input into a function to calculate the 'fidelity' via the custom loss function 'fidelity2' by comparing with the original input from <code>data.lab<\/code>. I did this by inputting <code>data.lab<\/code> a second time, but now as a label. However (1) model.summary() seems to show the output of productlayer is (4,4) when I think it should be (None,4,4) and (2) I receive the error:<\/p>\n\n<p><code>InvalidArgumentError: In[0] is not a matrix. Instead it has shape [10,4,4]\n     [[{{node loss_7\/fidout_loss\/ArithmeticOptimizer\/FoldTransposeIntoMatMul_matmul}}]]<\/code><\/p>\n\n<p>The y_true in the custom loss function seems to be taking the entire label set from model.fit (see code below), rather than batching, so this means it dumps the (10,4,4) label set in but is trying to compare against the (4,4) y_pred. Unsure how to fix this problem. Code is below. Ultimately I need to be able to calculate fidelity - actually infidelity eventually - in this way for arbitrary batch sizes (I set batchsize in model.fit to 1 just to try to get it working). Suggestions appreciated.<\/p>\n\n<pre><code>def fidelity2(y_true, y_pred):\n    y_truetp = tf.transpose(y_true)\n    t1 = (y_truetp @ y_pred)\n    tr = tf.trace(t1)\n    mxdim = tf.cast(tf.shape(y_pred)[0], tf.float32)\n    fidelity = (tf.abs((tr)** 2) \/ mxdim ** 2)\n    return fidelity    \n\n\n\nx = layers.Input(shape=(data.realdim,data.realdim), name='input1', batch_size=None)\nx1 = layers.GRU(data.Uj_dim, return_sequences=True)(x)\nx1 = layers.Dropout(rate=0.2)(x1)\nx1 = layers.GRU(data.Uj_dim, return_sequences=True)(x1)\nx1 = layers.Dropout(rate=0.2)(x1)\nx1 = layers.GRU(data.Uj_dim, return_sequences=True)(x1)\nx1 = layers.Flatten()(x1)\ny = layers.Dense(160, activation='relu', name='output_data')(x1)\nxreshape = layers.Reshape((4,4))(x)\ny2 = productlayer(trainable = True, name=\"fidout\")(x1)\nmodel = tf.keras.models.Model(inputs=x, outputs=[y,y2])\n\n\n\nbatchsz = 1\n#===============\nmodel.compile(optimizer='adam', loss=['mse',fidelity2], metrics=['mse',fidelity2])\nmodel.summary()\n\nmodel.fit(data.lab, [inputreshape,data.lab], epochs=2, batch_size=batchsz, validation_split = 0, shuffle=False, steps_per_epoch=1)\n\n#==============```\n\nModel: \"model_13\"\n__________________________________________________________________________________________________\nLayer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput1 (InputLayer)             [(None, 4, 4)]       0                                            \n__________________________________________________________________________________________________\ngru_39 (GRU)                    (None, 4, 40)        5400        input1[0][0]                     \n__________________________________________________________________________________________________\ndropout_26 (Dropout)            (None, 4, 40)        0           gru_39[0][0]                     \n__________________________________________________________________________________________________\ngru_40 (GRU)                    (None, 4, 40)        9720        dropout_26[0][0]                 \n__________________________________________________________________________________________________\ndropout_27 (Dropout)            (None, 4, 40)        0           gru_40[0][0]                     \n__________________________________________________________________________________________________\ngru_41 (GRU)                    (None, 4, 40)        9720        dropout_27[0][0]                 \n__________________________________________________________________________________________________\nflatten_13 (Flatten)            (None, 160)          0           gru_41[0][0]                     \n__________________________________________________________________________________________________\noutput_data (Dense)             (None, 160)          25760       flatten_13[0][0]                 \n__________________________________________________________________________________________________\nfidout (productlayer)           (4, 4)               0           flatten_13[0][0]                 \n==================================================================================================\nTotal params: 50,600\nTrainable params: 50,600\nNon-trainable params: 0\n__________________________________________________________________________________________________\nEpoch 1\/2\n---------------------------------------------------------------------------\nInvalidArgumentError...InvalidArgumentError: In[0] is not a matrix. Instead it has shape [10,4,4]\n     [[{{node loss_7\/fidout_loss\/ArithmeticOptimizer\/FoldTransposeIntoMatMul_matmul}}]]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57908752,"AnswerCount":0,"Tags":"<tensorflow><recurrent-neural-network>","CreationDate":"2019-09-12T14:18:22.973","AcceptedAnswerId":null,"OwnerUserId":11206288.0,"Title":"In tensorflow, how to compute rnn output's gradient to rnn input(when output and input is from different timestep)?","Body":"<p>In tensorflow, how to compute rnn output's gradient to rnn input(when output and input is from different timestep)?<\/p>\n\n<p>I am using bidirectional rnn:<\/p>\n\n<pre><code>((fw_outputs, bw_outputs), (fw_state, bw_state)) = tf.nn.bidirectional_dynamic_rnn(fw_cell, bw_cell, inputs, dtype=tf.float32)\n\n# Concat foward and backward outputs\noutputs = tf.concat((fw_outputs, bw_outputs), 2)\n<\/code><\/pre>\n\n<p>I want to compute gradient from rnn output in time step A to rnn input in timestep B. For example, my rnn inputs has shape (batch_size, time_steps, hidden_units), rnn outputs has shape (batch_size, time_steps, 2*hidden_units), I try to compute the gradient using the code below:<\/p>\n\n<pre><code>### Set input time step to be 1, and output time step to be 2.\n\n### this line return None\ngrad = tf.gradient(outputs[:,2,:], inputs[:,1,:])\n\n### And I have tried grad_ys(in which only gradient of timestep 2 is 1), it returns zero grad the input of time step 1.\n\ngrad = sess.run(tf.gradient(outputs, inputs, grad_ys), feed_dict)\n# grad[:,1,:] is zeros, only grad[:,2,:] has some non-zero values.\n<\/code><\/pre>\n\n<p>So my question is how can I achieve my purpose?<\/p>\n","answers":[]}
+{"QuestionId":57908966,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2019-09-12T14:30:23.837","AcceptedAnswerId":null,"OwnerUserId":12010475.0,"Title":"Speed up gradient calculation in custom training with loss function in eager mode in Tensorflow","Body":"<p>I am using Tensorflow v2-rc0 to build and train a neural network. My loss function is rather complicated and cannot be expressed in terms of tensor operations. For this reason, I want to run the loss function in eager mode but construct a graph for all model related computations. In order to calculate the gradient for training, I need to calculate the Jacobians of the network and the loss function separately and take their dot product.<\/p>\n\n<p>Here is how I implemented the gradient calculation.\nThis is just a mock up of what the final code would look like. I am using numpy here to simulate the need for eager execution in the loss function.<\/p>\n\n<pre><code>def mse(yTrue, yPred):\n    return np.mean((yTrue - yPred)**2)\n\ndef msejac(yTrue, yPred):\n    yt = np.array(yTrue)\n    yp = np.array(yPred)\n    return -2.0\/len(yt) * (yt - yp)\n\n@tf.function\ndef jacobian(model, inputs):\n    with tf.GradientTape() as gtape:\n        outputs = model(inputs)\n    jacs = gtape.jacobian(outputs, model.trainable_variables)\n    return outputs, jacs\n\ndef grad(model, inputs, targets):\n    outputs, outjacs = jacobian(model, inputs)\n    loss_value = mse(outputs, targets)\n\n    grads = [\n        np.mean([tf.tensordot(msejac(targets[b, :], outputs[b, :]),\n                              outjac[b, :],\n                              [-1, 0]).numpy()\n                 for b in range(inputs.shape[0])], axis=0)  # iter over batches\n        for outjac in outjacs]   # iter over model variables\n\n    return loss_value, grads\n<\/code><\/pre>\n\n<p>The calculation of the model Jacobian and the dot products at the end of <code>grad<\/code> take by far the longest in my training loop. And are orders of magnitude slower than the same calculation but computing the gradient directly on the loss, i.e.<\/p>\n\n<pre><code>def tf_mse(yTrue, yPred):\n    return tf.math.reduce_mean((yTrue - yPred)**2)\n\n@tf.function\ndef tf_grad(model, inputs, targets):\n    with tf.GradientTape() as gtape:\n        outputs = model(inputs)\n        lossValue = tf_mse(outputs, targets)\n    return lossValue, gtape.gradient(lossValue, model.trainable_variables)\n<\/code><\/pre>\n\n<p>My code becomes more efficient compared to the rest of the training loop when reducing the batch size because the Jacobians become much smaller. But even for a batch size of 1, this does not reach the performance of the pure TF-based implementation. In addition, I need to have a large batch size in order for my loss function to be meaningful.<\/p>\n\n<p>Is there a way I can speed things up? Is it even a sane thing to compute the Jacobian or is there a better way of getting the gradient in this situation?<\/p>\n","answers":[]}
+{"QuestionId":57909201,"AnswerCount":0,"Tags":"<python><android><tensorflow><machine-learning><keras>","CreationDate":"2019-09-12T14:43:38.173","AcceptedAnswerId":null,"OwnerUserId":7158019.0,"Title":"Trained Tensorflow model performs poorly on inference","Body":"<p>I trained an image classification model using Keras with Tensorflow backend. The model got good accuracy on validation dataset as well as on the testing data, I save the entire model to <code>.h5<\/code> format, here is my checkpoint callback.<\/p>\n\n<p><code>checkpoint = ModelCheckpoint(model_name+\".h5\", monitor='val_acc', verbose=1, save_best_only=True, save_weights_only=False, mode='auto', period=1)<\/code>. <\/p>\n\n<p>As I was hoping to use this model on Android so I refreeze the model to binary protobuf <code>(.pb)<\/code> using <a href=\"https:\/\/github.com\/amir-abdi\/keras_to_tensorflow\" rel=\"nofollow noreferrer\">keras_to_tensorflow.<\/a><\/p>\n\n<p>When performing inference using the model on mobile I noticed the model gives very wrong and random predictions. I have tried exploring other reasons why this could be the case like I found <a href=\"https:\/\/stackoverflow.com\/questions\/49474467\/export-keras-model-to-pb-file-and-optimize-for-inference-gives-random-guess-on\/49582776#49582776\">here<\/a> still, it seems clear the issue is not with loading the images.<\/p>\n\n<p>Also, performing inference with the converted model on Tensorflow Python still gives the same wrong\/random predictions. Here is my code for performing inference in Python.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def model_predict( model_path, image_path, model_input, model_output, class_names ):\n\n    with tf.Graph().as_default() as graph: # Set default graph as graph\n\n        with tf.Session() as sess:\n            # Load the graph in graph_def\n            print(\"load graph\")\n\n            # We load the protobuf file from the disk and parse it to retrive the unserialized graph_drf\n            with gfile.FastGFile(model_path,'rb') as f:\n\n                print(\"Load Image...\")\n                # Read the image &amp; get statstics\n                np_image = Image.open(image_path)\n                np_image = np.array(np_image).astype('float32')\/255\n                np_image = np.resize(np_image, (224, 224, 3))\n                np_image = np.expand_dims(np_image, axis=0)\n\n\n                # Set FCN graph to the default graph\n                graph_def = tf.GraphDef()\n                graph_def.ParseFromString(f.read())\n                sess.graph.as_default()\n\n                # Import a graph_def into the current default Graph (In this case, the weights are (typically) embedded in the graph)\n\n                tf.import_graph_def(\n                graph_def,\n                input_map=None,\n                return_elements=None,\n                name=\"\",\n                op_dict=None,\n                producer_op_list=None\n                )\n\n                # INFERENCE Here\n                m_input = graph.get_tensor_by_name(model_input) # Input Tensor\n                m_output = graph.get_tensor_by_name(model_output) # Output Tensor\n\n                print (\"Shape of input : \", tf.shape(m_input))\n                #initialize_all_variables\n                tf.global_variables_initializer()\n\n                # Run model on single image\n                Session_out = sess.run( m_output, feed_dict = {m_input : np_image} )\n\n                print(\"Predicted class:\", class_names[Session_out[0].argmax()] )\n<\/code><\/pre>\n\n<p>How do I perform inference using Tensorflow Python\/Android with a save <code>.pb<\/code> model?<\/p>\n\n<p>Others have suggested I save the session used for training and load them to Tensorflow when performing inference. If this is the case how do I load the saved session in Tensorflow android?<\/p>\n\n<p>I am sure the model did not overfit, it performs very well when using Keras.<\/p>\n","answers":[]}
+{"QuestionId":57909397,"AnswerCount":0,"Tags":"<python><keras><deep-learning><classification>","CreationDate":"2019-09-12T14:53:42.500","AcceptedAnswerId":null,"OwnerUserId":8855704.0,"Title":"Correct activations and loss functions for binary classification","Body":"<p>I am doing a binary image classification (two mutually exclusive classes) with keras and tensorflow and have been looking at the following options for the loss function and output layer activation:<\/p>\n\n<ol>\n<li><p>Categorical cross-entropy and softmax function with 2 neurons output<\/p><\/li>\n<li><p>Binary cross-entropy and softmax function with 2 neurons output<\/p><\/li>\n<li><p>Categorical cross-entropy and sigmoid function with 2 neurons output<\/p><\/li>\n<li><p>Binary cross-entropy and sigmoid function with 1 neuron output<\/p><\/li>\n<\/ol>\n\n<p>Before all of them there are several convolutional layers and Flatten().<\/p>\n\n<p>Do they all make sense? I expect 1 and 4 do, but I want to know more about the other 2 as well, because 2 seems to work fine but I don't know if it is correct.<\/p>\n","answers":[]}
+{"QuestionId":57909524,"AnswerCount":0,"Tags":"<tensorflow><keras><multiple-gpu>","CreationDate":"2019-09-12T14:58:41.753","AcceptedAnswerId":null,"OwnerUserId":10510693.0,"Title":"tensor-flow does not recognise 2nd GPU (\/gpu:1)","Body":"<p>I am trying to use 2 GPUs, tensorflow does not recognise the 2nd one. the 2nd GPU is working fine (in widows environment)<\/p>\n\n<pre><code>When I set CUDA_VISIBLE_DEVICES=0 and run the program I see RTX2070 as GPU0\nWhen I set CUDA_VISIBLE_DEVICES=1 and run the program I see GTX1050as GPU0\nWhen I set CUDA_VISIBLE_DEVICES=0,1 and run the program I see RTX2070 as GPU0\n<\/code><\/pre>\n\n<p>so basically, TF does not recognise GPU1, it one GPU at the same time (GPU 0)\nIs there any command to manually define GPU1?<\/p>\n\n<p>uninstalled and re-installed, Cudann, python 3.7, tensorflow and keras (GPU versions). I am using anaconda on windows 10. tried to change the CUDA_VISIBLE_DEVICES to 0, 1. I dont see any error, but the 2nd GPU does not appear anywhere in python.<\/p>\n\n<p>the main GPU is RTX2070 (8GB) and 2nd GPU is GTX1050 (2GB). Before i submit i spent sometime searching for solution and did whatever I could find on the internet. drivers are up to date, 64bit version anf latest versions of the software are installed. I dont see any issue, beside not appearing the 2nd GPU.<\/p>\n\n<p>The codes are working fine on first GPU, both have > 3.5 computational capacity.<\/p>\n","answers":[]}
+{"QuestionId":57909596,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2019-09-12T15:02:44.533","AcceptedAnswerId":57959123.0,"OwnerUserId":10584990.0,"Title":"Custom loss in keras produces misleading outputs during training of an autoencoder","Body":"<p>I have rank-3 tensors of size (100,100,4) that I try to compress and reconstruct with an autoencoder. I use a physically motivated loss function. Mathematically it is<\/p>\n\n<p>L = - overlap(y_true,y_pred) + |1 - norm(y_pred)^2|<\/p>\n\n<p>In code it reads:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def physical_loss(y_true,y_pred,norm=None):\n    return - tf.tensordot(y_true,y_pred,axes=([0,1,2],[0,1,2])) + tf.math.abs(1 - tf.tensordot(y_pred,y_pred,axes=([0,1,2],[0,1,2])))\n<\/code><\/pre>\n\n<p>The first term is the overlap between the input and output tensor (that I want to maximize), the second term keeps the output tensor from exploding and most desirably at norm 1 (overlap between tensor and itself = norm). So ideally the loss goes down to -1 (i.e. y_pred = y_true and norm(y_pred)=1)<\/p>\n\n<p>My model is a convolutional AutoEncoder that reduces the size to (50,50,4),(25,25,4) and builds it up again.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>cnn.compile(loss=physical_loss,optimizer=\"adam\")\n\ncnn.fit(x_train, x_train, epochs=num_epochs, batch_size=16,validation_data=(x_train,x_train))\n<\/code><\/pre>\n\n<p>Now when I train my cnn I the loss gradually goes down in 16 epochs eventually reaching val_loss = -0.67. But when I manually check the outcome (i.e. computing the loss between x_train and cnn.predict(x_train) directly myself), on average I obtain 0.75 (positive!).<\/p>\n\n<p><strong>So my question is what is keras\/tensorflow outputting when evaluating?<\/strong><\/p>\n\n<p>This might be related or another question, but when I cnn.evaluate(x_train,x_train,batch_size=X) the result varies from +0.88 to -0.88 and values in between depending on how I choose the batch size X (gets the same -0.67 when chossing None, so 32). I know keras does everything in batches but this shouldnt alter the overal output no? Also this makes it hard to interpret the output \/ training success.<\/p>\n","answers":[{"AnswerId":"57959123","CreationDate":"2019-09-16T14:36:33.167","ParentId":null,"OwnerUserId":"10584990","Title":null,"Body":"<p>I think I figured it out, the problem is that keras\/tensorflow is expecting an already vectorized function. The above definition is intended for rank-3 tensors, but keras\/tensorflow is always dealing with batches, so \"lists\" of rank-3 tensors that are not actually lists but rank-4 tensors (first dimension is the \"list\"). <\/p>\n\n<p>I found vectorizing this loss function not so trivial but found a working solution with tf.map_fn(function,listed_tensor,dtype) where listed_tensor is the mentioned rank-4 batch.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def single_eval(duo):\n    return - tf.tensordot(duo[0],duo[1],axes=([0,1,2],[0,1,2])) + tf.keras.backend.abs(1 - tf.tensordot(duo[1],duo[1],axes=([0,1,2],[0,1,2])))\n\ndef physical_loss(y_true,y_pred,norm=None):\n    return tf.map_fn(single_eval,(y_true,y_pred),dtype=(tf.float32))\n<\/code><\/pre>\n\n<p>The specification of the dtype is necessary in order to be able to use a tuple as an argument (i.e. use two arguments).<\/p>\n"}]}
+{"QuestionId":57909920,"AnswerCount":0,"Tags":"<r><keras><time-series><recurrent-neural-network>","CreationDate":"2019-09-12T15:23:06.813","AcceptedAnswerId":null,"OwnerUserId":7327058.0,"Title":"Fitting a keras time-series generator in R to time-series and exogenous data","Body":"<p>I'm using the R interface to Keras (for the 1st time) with a <code>timeseries_generator<\/code> per the reprex below. How do I get <code>fit_generator<\/code> to complete all epochs?<\/p>\n\n<p>I'm looking at custom generators as a possible alternative, but would ideally like to use the <code>timeseries_generator<\/code> assuming it is appropriate to my intended usage.<\/p>\n\n<p>The intent is to predict the actual values several steps ahead by using the \"feature\" and three lagged copies of the actuals as predictors.<\/p>\n\n<p>I'm assuming that differencing is not required since the feature is expected to indicate trend.<\/p>\n\n<pre class=\"lang-r prettyprint-override\"><code>library(tidyverse)\nlibrary(keras)\n\n# Function to generate multiple lags\n# See: https:\/\/gist.github.com\/romainfrancois\/469ed8224ba4be48befec28cb1e1ff80\nmany_lags &lt;- function(var, n = 10) {\n  var &lt;- enquo(var)\n\n  indices &lt;- seq_len(n)\n  map(indices, ~ quo(lag(!!var, !!.x))) %&gt;%\n    set_names(sprintf(\"lag_%s_%02d\", rlang::quo_text(var), indices))\n}\n\n# --- Simulate time-series\nset.seed(42)\nlags &lt;- 3\nsamples &lt;- 1000\n\nts &lt;- tibble(time = 1:(1000 + lags)) %&gt;%\n  mutate(\n    actual = seq(0.5, (1000 + lags) \/ 2, by = 0.5) + arima.sim(model = list(\n      ar = c(0.2, 0.3),\n      ma = 0.2,\n      order = c(2, 1, 1)\n    ), 999 + lags) %&gt;% as.double(),\n    feature = seq(0.5, (1000 + lags) \/ 2, by = 0.5)\n  ) %&gt;%\n  mutate(!!!many_lags(actual, lags)) %&gt;% \n  tail(samples)\n\nts %&gt;% ggplot(aes(time, actual)) +\n  geom_line()\n<\/code><\/pre>\n\n<p><img src=\"https:\/\/i.imgur.com\/HU2q7hK.png\" alt=\"\"><\/p>\n\n<pre class=\"lang-r prettyprint-override\"><code>\n# --- Centre &amp; scale predictors\ntest_window &lt;- 200\n\ntemp &lt;- ts %&gt;%\n  head(-test_window)\n\ncol_mean &lt;- map_dbl(temp, mean) %&gt;% replace(c(1:2), 0)\ncol_sd &lt;- map_dbl(temp, sd) %&gt;% replace(c(1:2), 1)\n\ntrain_data &lt;- ts %&gt;%\n  head(-test_window) %&gt;%\n  scale(col_mean, col_sd)\n\ntest_data &lt;- ts %&gt;%\n  tail(test_window) %&gt;%\n  scale(col_mean, col_sd)\n\n# --- Create generator objects\nbatch_size = 5\nlength &lt;- 5 # forecast timesteps\n\ntrain_gen &lt;- timeseries_generator(\n  train_data[, 3:6],\n  targets = train_data[, 2],\n  length = length,\n  sampling_rate = 1,\n  stride = 1,\n  start_index = 1,\n  end_index = 600,\n  shuffle = FALSE,\n  reverse = FALSE,\n  batch_size = batch_size\n)\n\nval_gen &lt;- timeseries_generator(\n  train_data[, 3:6],\n  targets = train_data[, 2],\n  length = length,\n  sampling_rate = 1,\n  stride = 1,\n  start_index = 601,\n  end_index = 800,\n  shuffle = FALSE,\n  reverse = FALSE,\n  batch_size = batch_size\n)\n\ntest_gen &lt;- timeseries_generator(\n  test_data[, 3:6],\n  targets = test_data[, 2],\n  length = length,\n  sampling_rate = 1,\n  stride = 1,\n  start_index = 1,\n  end_index = 200,\n  shuffle = FALSE,\n  reverse = FALSE,\n  batch_size = batch_size\n)\n\n# --- Train the model\nmodel &lt;- keras_model_sequential() %&gt;%\n  layer_gru(\n    units = ncol(train_data[, 3:6]),\n    batch_size = batch_size,\n    input_shape = list(NULL, ncol(train_data[, 3:6]))\n  ) %&gt;%\n  layer_dense(units = 1)\n\nsummary(model)\n#&gt; Model: \"sequential\"\n#&gt; ___________________________________________________________________________\n#&gt; Layer (type)                     Output Shape                  Param #     \n#&gt; ===========================================================================\n#&gt; gru (GRU)                        (5, 4)                        108         \n#&gt; ___________________________________________________________________________\n#&gt; dense (Dense)                    (5, 1)                        5           \n#&gt; ===========================================================================\n#&gt; Total params: 113\n#&gt; Trainable params: 113\n#&gt; Non-trainable params: 0\n#&gt; ___________________________________________________________________________\n\nmodel %&gt;% compile(loss = \"mse\",\n                  optimizer = optimizer_rmsprop(),\n                  metrics = list(\"mae\"))\n\nhistory &lt;- model %&gt;% \n  fit_generator(\n    train_gen,\n    steps_per_epoch = floor((train_gen$end_index - train_gen$start_index + 1) \/ batch_size),\n    epochs = 100,\n    validation_data = val_gen,\n    validation_steps = floor((val_gen$end_index - val_gen$start_index + 1) \/ batch_size)\n  )\n<\/code><\/pre>\n\n<p><sup>Created on 2019-09-12 by the <a href=\"https:\/\/reprex.tidyverse.org\" rel=\"nofollow noreferrer\">reprex package<\/a> (v0.3.0)<\/sup><\/p>\n\n<p>I get the following messages:<\/p>\n\n<pre><code>2019-09-12 15:42:44.613462: I **tensorflow\/core\/platform\/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA**\nEpoch 1\/100\n119\/119 [==============================] - 1s 10ms\/step - loss: 20351.4386 - mean_absolute_error: 120.7280 - val_loss: 77980.3295 - val_mean_absolute_error: 279.1719\nEpoch 2\/100\n**WARNING:tensorflow:Your dataset iterator ran out of data; interrupting training. Make sure that your iterator can generate at least `steps_per_epoch * epochs` batches** (in this case, 11900 batches). You may need touse the repeat() function when building your dataset.\n  0\/119 [..............................] - ETA: 0s - loss: 0.0000e+00 - mean_absolute_error: 120.7280 - val_loss: 0.0000e+00 - val_mean_absolute_error: 0.0000e+00\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57909951,"AnswerCount":1,"Tags":"<python><memory-management><keras><deep-learning><python-3.5>","CreationDate":"2019-09-12T15:25:00.740","AcceptedAnswerId":null,"OwnerUserId":3204706.0,"Title":"MemoryError: Unable to allocate array with shape (200, 20, 244, 244, 3) and data type float64","Body":"<p>I was trying to run a training module from a Github repo located <a href=\"https:\/\/github.com\/liorsidi\/ViolenceDetection_CNNLSTM\" rel=\"nofollow noreferrer\">HERE<\/a>. I'm running it on a cloud server, but I don't know why after some rounds, I get this error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"run.py\", line 229, in &lt;module&gt;\n    classes=classes, use_augs=use_augs, fix_lens=fix_lens)\n  File \"run.py\", line 165, in hyper_tune_network\n    classes=classes)\n  File \"run.py\", line 111, in get_generators\n    classes=classes)\n  File \"\/home\/code\/DatasetBuilder.py\", line 221, in get_sequences\n    X = pad_sequences(X, maxlen=seq_length, padding='pre', truncating='pre')\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/keras\/preprocessing\/sequence.py\", line 61, in pad_sequences\n    x = (np.ones((num_samples, maxlen) + sample_shape) * value).astype(dtype)\n  File \"\/root\/.local\/lib\/python3.5\/site-packages\/numpy\/core\/numeric.py\", line 214, in ones\n    a = empty(shape, dtype, order)\n\n    MemoryError: Unable to allocate array with shape (200, 20, 244, 244, 3) and data type float64\n<\/code><\/pre>\n\n<p>Frist, what does it mean? Is there any ways to fix it? What's the solution?<\/p>\n","answers":[{"AnswerId":"58666686","CreationDate":"2019-11-01T22:43:49.417","ParentId":null,"OwnerUserId":"8173793","Title":null,"Body":"<p>It has something to do with how linux handles RAM overcommits (I assume you're running linux). If you're requesting too much, your process will be killed or you get an error as you describe. As <strong>root<\/strong>, you can try:<\/p>\n\n<pre><code>echo 1 &gt; \/proc\/sys\/vm\/overcommit_memory\n<\/code><\/pre>\n\n<p>This will let you request as much as you want. If you request too much, you'll still get an out of memory error, but this will let you use larger arrays. <\/p>\n"}]}
+{"QuestionId":57910680,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-09-12T16:06:45.507","AcceptedAnswerId":57911642.0,"OwnerUserId":7157352.0,"Title":"How to name custom metrics in Keras fit output","Body":"<p>I am using custom metrics when training my Keras model. It works fine, except that the metrics names in the output of <code>model.fit_generator(...)<\/code> are not interpretable (NB: Tensorboard also uses these wrong names).<\/p>\n\n<p>Here is a reproducible example of what I am doing: the metrics are using one parameter (in addition to the prediction and ground truth), so I defined a factory to generate a parameterless metric function, similar to this:<\/p>\n\n<pre><code>def my_dummy_metric(y_true, y_pred, the_param=1.0):\n    return the_param * keras.backend.ones((1))\n\ndef my_metric_factory(the_param=1.0):\n    def fn(y_true, y_pred):\n        return my_dummy_metric(y_true, y_pred, the_param=the_param)\n\n    return fn\n\nmy_second_metric = my_metric_factory(2.0)\nmy_other_metric = my_metric_factory(3.14)\n<\/code><\/pre>\n\n<p>Then I compile and train my model:<\/p>\n\n<pre><code>model.compile(my_optim, my_loss, [my_second_metric, my_other_metric])\nhistory = model.fit_generator(...)\nprint(history.params['metrics'])\n<\/code><\/pre>\n\n<p>My trouble is that the metric names in <code>history<\/code> are <code>fn<\/code>, <code>fn_1<\/code>, <code>val_fn<\/code> and <code>val_fn_1<\/code>. These names are also used by Tensorboard and you need to know the implem details to understand them.<\/p>\n\n<p>On the contrary, I don't have this problem when using a simple custom function, without factory:<\/p>\n\n<pre><code>model.compile(my_optim, my_loss, [my_dummy_metric])\nhistory = model.fit_generator(...)\nprint(history.params['metrics'])\n<\/code><\/pre>\n\n<p>Would it be possible to obtain <code>my_XXX_metric<\/code> as output names also in the factory-based use case?<\/p>\n\n<p><strong>Environment:<\/strong> Using Keras 2.2.4, TF 1.14.0, Python 3.7<\/p>\n","answers":[{"AnswerId":"57911642","CreationDate":"2019-09-12T17:15:29.927","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>Yes, this is possible. In the metric factory, just set an appropriate <code>__name__<\/code> of metric function. For example:<\/p>\n\n<pre><code>def my_metric_factory(the_param=1.0):\n    def fn(y_true, y_pred):\n        return my_dummy_metric(y_true, y_pred, the_param=the_param)\n\n    fn.__name__ = 'metricname_{}'.format(the_param)\n    return fn\n<\/code><\/pre>\n"}]}
+{"QuestionId":57910710,"AnswerCount":1,"Tags":"<python-3.x><keras>","CreationDate":"2019-09-12T16:08:46.540","AcceptedAnswerId":57911180.0,"OwnerUserId":11938775.0,"Title":"Keras ImageDataGenerator wrong mapping","Body":"<p>I would like use an ImageDatagenerator from Keras.<\/p>\n\n<p>My Code: <\/p>\n\n<pre><code>image_datagen = image.ImageDataGenerator()\nmask_datagen  = image.ImageDataGenerator()\nimage_datagen.fit(X_train[:int(X_train.shape[0]*validation_split)],seed=42)\nmask_datagen.fit(Y_train[:int(Y_train.shape[0]*validation_split)],seed=42)\nx=image_datagen.flow(X_train[:int(X_train.shape[0]*validation_split)], batch_size=BATCH_SIZE,shuffle=True)\ny=mask_datagen.flow(Y_train[:int(Y_train.shape[0]*validation_split)], batch_size=BATCH_SIZE,shuffle=True)\ntrain_generator = zip(x, y)\nreturn train_generator\n<\/code><\/pre>\n\n<p>So, my images are grayscale and my label are binary images. But if I plot the x and y data, it becomes a different label for the original image. There is a mapping error. But where? If I plot my <code>X_train[0]<\/code> and <code>Y_train[0]<\/code> everything is ok.<\/p>\n","answers":[{"AnswerId":"57911180","CreationDate":"2019-09-12T16:40:01.690","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>You are shuffling the data with <code>shuffle=True<\/code> in the flow function calls. Because you are using two separate generators, <code>x<\/code> and <code>y<\/code> are getting shuffled independently. To fix this, I believe you need to add <code>seed=something<\/code> to your <code>.flow<\/code> calls.<\/p>\n\n<p>Also, you shouldn't need to call <code>.fit<\/code> at all, since you aren't using any transformations that are data-dependent.<\/p>\n"}]}
+{"QuestionId":57911414,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow2.0>","CreationDate":"2019-09-12T16:58:30.117","AcceptedAnswerId":null,"OwnerUserId":2985049.0,"Title":"Update external state using tensorflow autograph and tf.function","Body":"<p>I'm creating a layer that is a variation on batch normalization and using the <code>tf.function<\/code> decorator to speed it up. However, I'm getting an error that says autograph doesn't know whether or not I want to reuse the variable that I'm trying to update.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class MyClass(tf.keras.layers.Layer):\n\n    def build():\n        self.foo = self.add_weight(...)\n\n    @tf.function\n    def call(inputs, training=None):\n        lst = [K.moving_average_update(self.foo, .5, .999)]\n        self.add_updates(lst)\n\n<\/code><\/pre>\n\n<p>This throws an error that is something like:<\/p>\n\n<pre><code>ValueError: Variable my_class\/foo\/biased already exists, disallowed. Did you mean to set reuse=True or reuse=tf.AUTO_REUSE in VarScope?\n<\/code><\/pre>\n\n<p>What is the idiomatic way to tell autograph to reuse state outside the function?<\/p>\n","answers":[{"AnswerId":"58130767","CreationDate":"2019-09-27T08:34:03.227","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>In order to reuse the variable <code>self.foo<\/code> you could define within a tf.variable_scope block, setting the argument <code>reuse = tf.AUTO_REUSE<\/code>. For example:<\/p>\n\n<pre><code>def build(self, input_shape):\n\n    with tf.variable_scope('scope', reuse=tf.AUTO_REUSE):\n        self.foo = tf.Variable(initial_value=YOUR_INITIAL_VALUE, name='var')\n\n    trainable = None  # True or False\n    if trainable:\n        self._trainable_weights.append(self.foo)\n    else:\n        self._non_trainable_weights.append(self.foo)\n<\/code><\/pre>\n\n<p><strong>NOTE:<\/strong> Not tested.<\/p>\n"}]}
+{"QuestionId":57911524,"AnswerCount":2,"Tags":"<python><tensorflow><keras><tf.keras>","CreationDate":"2019-09-12T17:06:00.887","AcceptedAnswerId":null,"OwnerUserId":8908465.0,"Title":"Why can't I use a Keras model like an tf op in my flow graph","Body":"<p>I am trying to use my pre-trained keras model somewhere in my computational flow-graph in a tf session, so I've tried this simplified graph and I still get the same errors.<\/p>\n\n<pre><code>model = load_model('models\/vtcnn3.h5')\n\ninput = tf.placeholder(tf.float32,shape=(1,*x_test.shape[1:]))\noutput = model(input)\n\nsess = tf.compat.v1.Session()\n\nguess_0 = sess.run(output, {input:x_test[0:1]} )\n<\/code><\/pre>\n\n<p>when the session runs I get a large traceback that eventually says<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.FailedPreconditionError: Error while reading resource variable dense2\/kernel from Container: localhost. This could mean that the variable was uninitialized. Not found: Container localhost does not exist. (Could not find resource: localhost\/dense2\/kernel)\n     [[node sequential\/dense2\/MatMul\/ReadVariableOp (defined at code\/soq.py:114) ]]\n\nErrors may have originated from an input operation.\nInput Source operations connected to node sequential\/dense2\/MatMul\/ReadVariableOp:\n dense2\/kernel (defined at code\/soq.py:111)\n\nOriginal stack trace for 'sequential\/dense2\/MatMul\/ReadVariableOp':\n  File \"code\/soq.py\", line 114, in &lt;module&gt;\n    output = model(input)\n  File \"\/Users\/yaba\/miniconda3\/envs\/cs1\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/engine\/base_layer.py\", line 634, in __call__\n    outputs = call_fn(inputs, *args, **kwargs)\n.\n.\n.\n.\n<\/code><\/pre>\n\n<p><code>dense2<\/code> is the name of a layer in my model. When I make other changes to the flow-graph, the traceback puts the error into a different layer of my model.<\/p>\n","answers":[{"AnswerId":"57919336","CreationDate":"2019-09-13T07:44:21.543","ParentId":null,"OwnerUserId":"9579075","Title":null,"Body":"<p>Before you try to run your inputs through your Keras model, you need to initialize the variables. The following code works fine with <code>tensorflow==1.14.0<\/code><\/p>\n\n<pre><code>import tensorflow as tf\n\nmodel = tf.keras.applications.ResNet50()\n\ninit_op = tf.global_variables_initializer()\nwith tf.compat.v1.Session() as sess:\n    random_image, _ = sess.run([tf.random.normal(shape=[1, 224, 224, 3]), init_op])\n    outputs = sess.run(model.output, feed_dict={model.input:random_image})\n<\/code><\/pre>\n"},{"AnswerId":"57928077","CreationDate":"2019-09-13T17:34:17.197","ParentId":null,"OwnerUserId":"8908465","Title":null,"Body":"<p>Thank you to Srihari Humbardwadi for the initialization answer, that fixed the error I was getting but now the output of the network was different than if I used <code>model.predict(...)<\/code>!<\/p>\n\n<p>If you want to load a Keras model fmor a file and use it in your Tensorflow graph, you must set your <code>tf.Session()<\/code> into the Keras backend, THEN initialize variables, THEN load the Keras model as shown below<\/p>\n\n<pre><code>sess = tf.Session()\n\n# make sure keras has the same session as this code BEFORE initializing\ntf.keras.backend.set_session(sess)\n\n# Do this BEFORE loading a keras model\ninit_op = tf.global_variables_initializer()\nsess.run(init_op)\n\n\nmodel = models.load_model('models\/vtcnn3.h5')\n\n\ninput = tf.placeholder(tf.float32,shape=(1,*x_test.shape[1:]))\noutput = model(input)\n\n\nguess_0 = sess.run(output, feed_dict={input:x_test[0:1]} )\n<\/code><\/pre>\n\n<p>This <code>guess_0<\/code> will now be the same as if you were to do <\/p>\n\n<pre><code>model = models.load_model('models\/vtcnn3.h5')\nguess_0 = model.predict(x_test[0:1])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57911734,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-12T17:22:07.387","AcceptedAnswerId":null,"OwnerUserId":3667142.0,"Title":"Tensorflow Float16 for VGG19 model parameters","Body":"<p>I'm doing <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/generative\/style_transfer\" rel=\"nofollow noreferrer\">neural style transfer<\/a> using TensorFlow(with the 19.01 Nvidia TensorFlow Docker image) in Python. I have an Nvidia 2070 graphics card, and I've been getting Out of Memory errors when I try to run a newer version of the TensorFlow docker image(19.08 for example). So I decided that perhaps it is time to consider using 16 bit precision instead of 32 bit for storing the parameters of the VGG19 CNN.<\/p>\n\n<p>My initial research when I built my machine had led me to believe that switching from 32 to 16 was a cakewalk, but that hasn't been my experience now that I'm actively trying to make the transition.<\/p>\n\n<p>This is what I have done:<\/p>\n\n<ol>\n<li>I set tf.keras.backend.set_floatx('float16')<\/li>\n<li>Set tf.keras.backend.set_epsilon(1e-4)<\/li>\n<li>Change my image input to the VGG19 network to a float16, and any other miscellaneous parts of my code that use the float32 datatype in conjunction with the float16.<\/li>\n<\/ol>\n\n<p>When I run the code, nvidia-smi still reports that essentially 100% of my GPU is being used. Has anyone had any success with reducing their model memory footprint by switching to float16 in TensorFlow?<\/p>\n","answers":[]}
+{"QuestionId":57911771,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><logistic-regression>","CreationDate":"2019-09-12T17:23:46.043","AcceptedAnswerId":null,"OwnerUserId":2779330.0,"Title":"Multiclass logistic regression - implementation question","Body":"<p>This is my try to implement multi-class logistic regression in python using softmax as activation function and mnist digit data set as training and test set.<\/p>\n\n<pre><code>import numpy as np\n\ndef softmax(z):\n   return np.array([(np.exp(el)\/np.sum(np.exp(el))) for el in z])\n\ndef cost(W,F,L):\n   m = F.shape[0] #get number of rows\n   mul = np.dot(F, W)\n   sm_mul_T = softmax(mul)\n   return -(1\/m) * np.sum(L * np.log(sm_mul_T))\n\ndef gradient(W,F,L):\n   m = F.shape[0]  # get number of rows\n   mul = np.dot(F, W)\n   sm_mul_T = softmax(mul)\n   return -(1 \/ m) * np.dot(F.T , (L - sm_mul_T))\n\nfrom tensorflow.examples.tutorials.mnist import input_data\nmnist = input_data.read_data_sets(\".\/datasets\/MNIST_data\/\", one_hot=True)\n\nW = np.zeros((785, 10)) #784 features + 1 bias\n\nfor _ in range(10000):\n   F, L = mnist.train.next_batch(100)\n\n   F = np.insert(F,0, values=1, axis=1)\n\n   total_cost = cost(W,F,L)\n   print(\"Total cost is {}\".format(total_cost))\n\n   gradients = gradient(W,F,L)\n\n   W = W - (0.1 * gradients)\n\nFU = mnist.test.images\n\nFU = np.insert(FU,0, values=1, axis=1)\n\nLU = mnist.test.labels\n\nmulU = np.dot(FU, W)\nsm_mulU = softmax(mulU)\n\nOK=0\nNOK=0\n\nfor i in range(10000):\n   a1 = np.argmax(sm_mulU[i])\n   a2 = np.argmax(LU[i])\n   if a1 == a2:\n      OK = OK + 1\n   else:\n      NOK = NOK + 1\n\nprint(\"{} OK vs {} NOK\".format(OK, NOK))\nprint(\"accur {}%\".format((OK\/(NOK+OK))*100))\n<\/code><\/pre>\n\n<p>What I wanted to do is basically to try to implement it myself and try to get similar results as using Tensor flow. But the problem is that tensor flow implementation gets in the end 91% accuracy, while I can get only around 70%. Also, it seems like my model diverges and cost starts to increase pretty fast.<\/p>\n\n<p>Is my implementation wrong, or is it due to more advanced algorithm inside TensorFlow implementation?<\/p>\n","answers":[]}
+{"QuestionId":57912754,"AnswerCount":1,"Tags":"<python><tensorflow><compilation><static-linking>","CreationDate":"2019-09-12T18:38:02.027","AcceptedAnswerId":null,"OwnerUserId":2886575.0,"Title":"linking into the tensorflow libraries","Body":"<p>I am writing a custom OP for <code>tensorflow<\/code>.  My operation is a linear algebra type thing, so I have been using the boilerplate from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/00fad90125b18b80fe054de1055770cfb8fe4ba3\/tensorflow\/core\/kernels\/self_adjoint_eig_op.cc#L32\" rel=\"nofollow noreferrer\">here<\/a> to implement a similar thing:<\/p>\n\n<pre><code>template &lt;class Scalar&gt;\nclass FooOp : public ::tensorflow::LinearAlgebraOp&lt;Scalar&gt; {\n public:\n  typedef LinearAlgebraOp&lt;Scalar&gt; Base;\n  explicit FooOp(OpKernelConstruction* context) : Base(context) {}\n\n  using TensorShapes = typename Base::TensorShapes;\n  using Matrix = typename Base::Matrix;\n  using MatrixMaps = typename Base::MatrixMaps;\n  using ConstMatrixMap = typename Base::ConstMatrixMap;\n  using ConstMatrixMaps = typename Base::ConstMatrixMaps;\n\n  TensorShapes GetOutputMatrixShapes(\n      const TensorShapes&amp; input_matrix_shapes) const final {}\n\n  void ComputeMatrix(OpKernelContext* context, const ConstMatrixMaps&amp; inputs,\n                     MatrixMaps* outputs) final {};\n<\/code><\/pre>\n\n<p>Some of these items, such as the <code>LinearAlgebraOp<\/code> class, come from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/512a3c3a6c0b36f6750394103b7eb316fe4ce853\/tensorflow\/core\/kernels\/linalg_ops_common.cc#L32\" rel=\"nofollow noreferrer\"><code>tensorflow\/core\/kernels\/linalg_ops_common.cc<\/code><\/a>.  However when I try to build my new op using <a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/op#build_the_op_library\" rel=\"nofollow noreferrer\">the instructions<\/a>, I get an error about that class.  How do I use this in my user ops?<\/p>\n","answers":[{"AnswerId":"57966486","CreationDate":"2019-09-17T02:11:47.360","ParentId":null,"OwnerUserId":"2886575","Title":null,"Body":"<p>In order to access the class <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/512a3c3a6c0b36f6750394103b7eb316fe4ce853\/tensorflow\/core\/kernels\/linalg_ops_common.h#L40\" rel=\"nofollow noreferrer\"><code>LinearAlgebraOp<\/code><\/a> from <a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/op#compile_the_op_using_your_system_compiler_tensorflow_binary_installation\" rel=\"nofollow noreferrer\">user ops<\/a>, you'll need to export some new symbols in the shared libraries.  This means you'll need to <a href=\"https:\/\/www.tensorflow.org\/install\/source\" rel=\"nofollow noreferrer\">compile tensorflow from source<\/a>.<\/p>\n\n<p>The compile-script framework that tensorflow uses is <a href=\"https:\/\/bazel.build\/\" rel=\"nofollow noreferrer\">bazel<\/a> and the BUILD scripts list the various files and whatnot to be compiled and which functions to \"advertise\".  We need to \"advertise\" the files that the desired class is in.<\/p>\n\n<p>First, we're going to change: <code>\/tensorflow\/core\/BUILD<\/code> to tell bazel that our user ops will depend on these files.  Then, we'll change <code>tensorflow\/core\/kernels\/BUILD<\/code> to make them available.  Here's the git diff:<\/p>\n\n<hr>\n\n\n\n<pre><code>--- a\/tensorflow\/core\/BUILD\n+++ b\/tensorflow\/core\/BUILD\n@@ -1266,7 +1266,7 @@ cc_library(\n     copts = tf_copts(),\n     linkstatic = 1,\n     visibility = [\"\/\/visibility:public\"],\n-    deps = [\":framework\"],\n+    deps = [\":framework\", \"\/\/tensorflow\/core\/kernels:linalg_ops_common\"],\n     alwayslink = 1,\n )\n\n--- a\/tensorflow\/core\/kernels\/BUILD\n+++ b\/tensorflow\/core\/kernels\/BUILD\n@@ -3317,7 +3317,7 @@ cc_library(\n     name = \"linalg_ops_common\",\n     srcs = [\"linalg_ops_common.cc\"],\n     hdrs = [\"linalg_ops_common.h\"],\n-    visibility = [\"\/\/visibility:private\"],\n+    visibility = [\"\/\/visibility:public\"],\n     deps = [\n         \"\/\/tensorflow\/core:framework\",\n         \"\/\/tensorflow\/core:lib\",\n<\/code><\/pre>\n\n<p>Great.  Now make sure that you list these dependencies in your <a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/op#compile_the_op_using_bazel_tensorflow_source_installation\" rel=\"nofollow noreferrer\">user op bazel BUILD file<\/a><\/p>\n\n<pre><code>+    deps = [\"\/\/tensorflow\/core\/kernels:linalg_ops_common\",],\n<\/code><\/pre>\n\n<p>Now, <a href=\"https:\/\/www.tensorflow.org\/install\/source\" rel=\"nofollow noreferrer\">compile tensorflow<\/a>, and then <a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/op#compile_the_op_using_bazel_tensorflow_source_installation\" rel=\"nofollow noreferrer\">compile your module<\/a>.  Now install your compiled custom version of tensorflow.  The <code>.so<\/code> file should now work on this install using the <a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/op#use_the_op_in_python\" rel=\"nofollow noreferrer\">usual user op import strategy<\/a>.<\/p>\n"}]}
+{"QuestionId":57913519,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><anaconda>","CreationDate":"2019-09-12T19:41:25.877","AcceptedAnswerId":null,"OwnerUserId":11674981.0,"Title":"import tensorflow as tf error occur in Anaconda Navigator","Body":"<p>while importing tenserflow module in anaconda notebook I'm facing this error numpy.core.multiarray failed to import<\/p>\n","answers":[{"AnswerId":"57913894","CreationDate":"2019-09-12T20:12:04.030","ParentId":null,"OwnerUserId":"3254869","Title":null,"Body":"<p>Welcome to StackOverflow. \nPlease consider <a href=\"https:\/\/stackoverflow.com\/help\/how-to-ask\">reviewing \"How do i ask a good question\" guidelines<\/a> to provide more context. <\/p>\n\n<p>With your question, it is likely you have an older version of numpy installed in Anaconda through pip. Try using anaconda <a href=\"https:\/\/docs.conda.io\/projects\/conda\/en\/latest\/user-guide\/tasks\/manage-environments.html\" rel=\"nofollow noreferrer\">environments<\/a> and install tensorflow directly <a href=\"https:\/\/docs.anaconda.com\/anaconda\/user-guide\/tasks\/tensorflow\/\" rel=\"nofollow noreferrer\">via this guide.<\/a><\/p>\n\n<p>The commands you should run are to install the current release of CPU-only TensorFlow, recommended for beginners are:<\/p>\n\n<pre><code>conda create -n tf tensorflow\nconda activate tf\n<\/code><\/pre>\n"}]}
+{"QuestionId":57913757,"AnswerCount":0,"Tags":"<python-3.x><torch>","CreationDate":"2019-09-12T20:00:45.263","AcceptedAnswerId":null,"OwnerUserId":4762630.0,"Title":"How do I install torch version 1 above in my windows machine?","Body":"<p>I'm trying to install torch version 1>, unfortunately I'm having this error.<\/p>\n\n<p>ERROR: Could not find a version that satisfies the requirement torch==1.1.0 (from versions: 0.1.2, 0.1.2.post1, 0.1.2.post2)\nERROR: No matching distribution found for torch==1.1.0<\/p>\n\n<p>This is what I execute.<\/p>\n\n<p>pip install torch==1.1.0<\/p>\n","answers":[]}
+{"QuestionId":57913825,"AnswerCount":2,"Tags":"<python><pytorch>","CreationDate":"2019-09-12T20:05:57.027","AcceptedAnswerId":57915262.0,"OwnerUserId":12060255.0,"Title":"How to select specific labels in pytorch MNIST dataset","Body":"<p>I'm trying to create dataloaders using only a specific digit from PyTorch Mnist dataset <\/p>\n\n<p>I already tried to create my own Sampler but it doesn't work and I'm not sure I'm using correctly the mask.<\/p>\n\n<pre><code>class YourSampler(torch.utils.data.sampler.Sampler):\n\n    def __init__(self, mask):\n\n        self.mask = mask\n\n\n    def __iter__(self):\n\n        return (self.indices[i] for i in torch.nonzero(self.mask))\n\n\n    def __len__(self):\n\n        return len(self.mask)\n\n\nmnist = datasets.MNIST(root=dataroot, train=True, download=True, transform = transform)   \n\nmask = [True if mnist[i][1] == 5 else False for i in range(len(mnist))]\n\nmask = torch.tensor(mask)   \n\nsampler = YourSampler(mask)\n\ntrainloader = torch.utils.data.DataLoader(mnist, batch_size=4, sampler = sampler, shuffle=False, num_workers=2)\n\n<\/code><\/pre>\n\n<p>So far I had many different types of errors. For this implementation, it's \"Stop Iteration\".\nI feel like this is very easy\/stupid but I can't find a simple way to do it.\nThank you for your help!<\/p>\n","answers":[{"AnswerId":"57914468","CreationDate":"2019-09-12T21:03:20.430","ParentId":null,"OwnerUserId":"1818327","Title":null,"Body":"<p><code>StopIteration<\/code> is raised when your iterator is exhausted. Are you sure your mask is working correctly? it seems like you pass list of boolean values, yet torch.nonzero would expect floats or ints.<\/p>\n\n<p>You should write:<\/p>\n\n<p><code>mask = [1 if mnist[i][1] == 5 else 0 for i in range(len(mnist))]<\/code><\/p>\n\n<p>You should also need to pass the dataset to your sampler such as:<\/p>\n\n<p><code>sampler = YourSampler(dataset, mask=mask)<\/code><\/p>\n\n<p>with this class definition<\/p>\n\n<pre><code>class YourSampler(torch.utils.data.sampler.Sampler):\n\n    def __init__(self, dataset, mask):\n\n        self.mask = mask\n        self.dataset = dataset\n...\n<\/code><\/pre>\n\n<p>For more details, you can refer to pytorch documentation(which shows the source code) to see how they implemented more advanced samplers:<a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/utils\/data\/sampler.html#SequentialSampler\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/utils\/data\/sampler.html#SequentialSampler<\/a><\/p>\n"},{"AnswerId":"57915262","CreationDate":"2019-09-12T22:31:50.307","ParentId":null,"OwnerUserId":"12060255","Title":null,"Body":"<p>Thank you for your help.\nAfter a while I figured out a solution (but might not be the best at all):<\/p>\n\n<pre><code>class YourSampler(torch.utils.data.sampler.Sampler):\n    def __init__(self, mask, data_source):\n        self.mask = mask\n        self.data_source = data_source\n\n    def __iter__(self):\n        return iter([i.item() for i in torch.nonzero(mask)])\n\n    def __len__(self):\n        return len(self.data_source)\n\nmnist = datasets.MNIST(root=dataroot, train=True, download=True, transform = transform)    \nmask = [1 if mnist[i][1] == 5 else 0 for i in range(len(mnist))]\nmask = torch.tensor(mask)   \nsampler = YourSampler(mask, mnist)\ntrainloader = torch.utils.data.DataLoader(mnist, batch_size=batch_size,sampler = sampler, shuffle=False, num_workers=workers)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57914148,"AnswerCount":0,"Tags":"<tensorflow><keras><prediction>","CreationDate":"2019-09-12T20:34:30.107","AcceptedAnswerId":null,"OwnerUserId":8894647.0,"Title":"I have question regarding image classification code","Body":"<p>I am doing Image classification model living and nonliving and I have 500 train images of human and 600 images of car and I am following the same code that is in tensorflow official documentation Here is the link:<\/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 problem I am having is during prediction part where I plot my image against barchart. I am always having living label despite it is nonliving in actual.<\/p>\n\n<pre><code>class_name = [\"living\", \"nonliving\"]\npickle_in = open(\"x_pickle\", \"rb\")\nX_train_img = pickle.load(pickle_in)\nX_train_img = np.array(X_train_img)\nX_train_img = X_train_img\/255\n\npickle_in = open(\"y_pickle\", \"rb\")\nX_train_label = pickle.load(pickle_in)\n\npickle_in = open(\"a_pickle\", \"rb\")\nX_test_img = pickle.load(pickle_in)\nX_test_img = np.array(X_test_img)\nX_test_img = X_test_img\/255\n\npickle_in = open(\"b_pickle\", \"rb\")\nX_test_label = pickle.load(pickle_in)\ndef plot_image(i, predictions_array, true_label, img):\n    predictions_array, true_label, img = predictions_array[i], \ntrue_label[i], img[i]\n    plt.grid(False)\n    plt.xticks([])\n    plt.yticks([])\n\n    plt.imshow(img, cmap=plt.cm.binary)\n\n    predicted_label = np.argmax(predictions_array)\n    if predicted_label == true_label:\n        color = 'blue'\n\n    else:\n        color = 'red'\n\n    plt.xlabel(\"{} {:2.0f}% ({})\".format(class_name[predicted_label],\n                                     100 * np.max(predictions_array),\n                                     class_name[true_label]), color=color)\ndef plot_value_array(i, predictions_array, true_label):\n    predictions_array, true_label = predictions_array[i], true_label[i]\n    plt.grid(False)\n    plt.xticks([])\n    plt.yticks([])\n    thisplot = plt.bar(range(10), predictions_array, color=\"#777777\")\n    plt.ylim([0, 1])\n    predicted_label = np.argmax(predictions_array)\n\n    thisplot[predicted_label].set_color('red')\n    thisplot[true_label].set_color('blue')\n    model = create_model()\n    model_details = model.fit(X_train_img, X_train_label, batch_size=64, \n    epochs=50, validation_data=(X_test_img,X_test_label),\n                       callbacks=[cp_callback])\n\nmodel.load_weights(checkpoint_path)\n\nevaluation = model.evaluate(X_test_img,X_test_label)\nprint(\"Accuracy: %.2f%%\" % (evaluation[1]*100))\n\npredictions = model.predict(X_test_img)\nprint(predictions[24])\n\nnp.argmax(predictions[24])\nprint(np.argmax(predictions[24]))\nprint(X_test_label[24])\n\ni = 24\nplt.figure(figsize=(6,3))\nplt.subplot(1,2,1)\nplot_image(i, predictions, X_test_label, X_test_img)\nplt.subplot(1,2,2)\nplot_value_array(i, predictions, X_test_label)\nplt.show()\n<\/code><\/pre>\n\n<p>just like in tutorial documentation where ankle is being classified correctly with correct label my car should be classified as nonliving but it is always classified  as living<\/p>\n\n<p>I think I have hard coded it somewhere or something is not getting it right because it always shows living whether it is image of car or human<\/p>\n","answers":[]}
+{"QuestionId":57914176,"AnswerCount":1,"Tags":"<python><keras><generator><lstm>","CreationDate":"2019-09-12T20:36:42.753","AcceptedAnswerId":null,"OwnerUserId":12060390.0,"Title":"How to correctly specify input shape for a Keras LSTM model","Body":"<p>I have a generator that yields my x and y which is what I'm using to train my LSMT model in Keras.<\/p>\n\n<p>The X is of the format [0.0, 1.0, 0.0004].<\/p>\n\n<p>I set the shape of this data earlier:<\/p>\n\n<pre><code>x_out = np.array(list(x_drop.values())).reshape(len(x_drop), 1)\ny_out = np.array([y])\n<\/code><\/pre>\n\n<p>Which I get in an async generator, so then I pass into a syncing function and then use model.fit_generator():<\/p>\n\n<pre><code>train_gen = to_sync_generator(replay(traintime_0, traintime_1, test=0))\n    model = lstm_model()\n\nmodel.fit_generator(\n    train_gen,\n    epochs = 5,\n    verbose = 1,\n    steps_per_epoch=500,\n    initial_epoch=0,\n    validation_data=valid_gen,\n    validation_steps=500)\n<\/code><\/pre>\n\n<p>My model code is below.<\/p>\n\n<pre><code>   def lstm_model():\n       model = Sequential()\n       model.add(LSTM(100, input_shape=(3,1)))\n       model.add(Dense(1))\n       model.compile(loss='mean_squared_error', optimizer='adam')\n       return model\n<\/code><\/pre>\n\n<p>I get the following error: <\/p>\n\n<pre><code>ValueError: Error when checking input: expected lstm_1_input to have 3 dimensions, but got array with shape (3, 1)\n<\/code><\/pre>\n\n<p>Note, I can not preprocess this data. <\/p>\n\n<p>I haven't really seen examples of using lstm with non-preprocessed data, any advise is greatly appreciated! thank you.<\/p>\n","answers":[{"AnswerId":"57915653","CreationDate":"2019-09-12T23:36:44.767","ParentId":null,"OwnerUserId":"11518214","Title":null,"Body":"<p>Your data must be a tuple using the format (x, y, z), where => x is your batch number (you can initially start this with 1), y is your data array and z is your data shape.\nBy your case, I can deduce your data have one feature and has 3 steps. So, you can prepare it using the following code:<\/p>\n\n<pre><code>data = data.reshape((1, 3, 1))\n<\/code><\/pre>\n\n<p>If you cannot change the data directly, you can merge it into a new array, then fit the array into the model<\/p>\n"}]}
+{"QuestionId":57914193,"AnswerCount":3,"Tags":"<machine-learning><keras>","CreationDate":"2019-09-12T20:38:01.070","AcceptedAnswerId":null,"OwnerUserId":7317985.0,"Title":"input_dim, output_dim not working in updated keras","Body":"<p>i am doing this online course on udemy and everything works fine but when i try to intialize the first hidden layer it gave the following error <\/p>\n\n<pre><code>TypeError: __init__() missing 1 required positional argument: 'units'.\n<\/code><\/pre>\n\n<p>then I did the ctrl+I on the spyder and chnged the output_dim and init arguments but i dont know what to replace others with ..<\/p>\n\n<pre><code>import keras\nfrom keras.models import Sequential\nfrom keras.layers import Dense\n\n#initializing the ANN\nclassifier = Sequential()\n\n#adding the input layer and the first hidden layer\nclassifier.add(Dense(units =6, kernel_initializer = 'uniform' , activation = 'relu', input_dim =11 ))\n\n#adding the second layer\n\nclassifier.add(Dense(Output_dim = 6 , kernel_initializer = 'uniform' , activation = 'relu'))\n<\/code><\/pre>\n\n<p>should work fine with no error<\/p>\n","answers":[{"AnswerId":"57914991","CreationDate":"2019-09-12T21:57:36.647","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>In a <code>Dense<\/code> layer, the number of units is equivalent to the output dimensionality. However, the argument <code>Output_dim<\/code> does not exist. So, replace <code>Dense(Output_dim=6, ...)<\/code> with <code>Dense(units=6, ...)<\/code> (or even just <code>Dense(6, ...)<\/code>).<\/p>\n"},{"AnswerId":"57942636","CreationDate":"2019-09-15T09:03:41.480","ParentId":null,"OwnerUserId":"12069506","Title":null,"Body":"<p>In the new documentation of Dense function output_dim is replaced by units, and the input_dim is replaced by input_shape. However in the input_shape argument you have to specify a tuple.<\/p>\n\n<p>Eg:<\/p>\n\n<p>Adding the input layer and the first hidden layer<\/p>\n\n<p><code>classifier.add(Dense(units=6, activation = 'relu', kernel_initializer = 'uniform', input_shape = (11, )))<\/code><\/p>\n\n<p>Adding the second layer<\/p>\n\n<p><code>classifier.add(Dense(units = 6 , kernel_initializer = 'uniform' , activation = 'relu'))<\/code><\/p>\n"},{"AnswerId":"58254693","CreationDate":"2019-10-06T06:11:52.923","ParentId":null,"OwnerUserId":"7021733","Title":null,"Body":"<h1>Add the first ANN layers (Input and Hidden Layers)<\/h1>\n\n<p><code>classifier.add(Dense(units=6, activation='relu', kernel_initializer='uniform', input_dim = 11))<\/code><\/p>\n\n<h1>Adding the second hidden layer<\/h1>\n\n<p><code>classifier.add(Dense(units = 6, kernel_initializer = 'uniform', activation = 'relu'))<\/code><\/p>\n"}]}
+{"QuestionId":57914354,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras><deep-learning><vgg-net>","CreationDate":"2019-09-12T20:53:53.300","AcceptedAnswerId":57915450.0,"OwnerUserId":7065438.0,"Title":"How to replace the input channel shape from (224, 224, 3) to (224, 224, 1) in VGG16?","Body":"<p>I am using VGG16 for transfer learning. My images are grayscale. So, I need to change the input channel shape of Vgg16 from (224, 224, 3) to (224, 224, 1). I tried the following code and got error: <\/p>\n\n<pre><code>TypeError: build() takes from 1 to 2 positional arguments but 4 were given\n<\/code><\/pre>\n\n<p>Can anyone help me where Am I doing it wrong?<\/p>\n\n<pre><code>vgg16_model= load_model('Fetched_VGG.h5')\nvgg16_model.summary()\n\n# transform the model to Sequential\nmodel= Sequential()\nfor layer in vgg16_model.layers[1:-1]:\n    model.add(layer)\n\n# Freezing the layers (Oppose weights to be updated)\nfor layer in model.layers:\n    layer.trainable = False\n\nmodel.build(224,224,1)\nmodel.add(Dense(2, activation='softmax', name='predictions'))\n<\/code><\/pre>\n","answers":[{"AnswerId":"57915450","CreationDate":"2019-09-12T23:02:27.097","ParentId":null,"OwnerUserId":"10528078","Title":null,"Body":"<p>you can't, even if you get rid of the input layer, this model has a graph that has already been compiled and your first <strong>conv<\/strong> layer expects an input with 3 channels. I don't think there is really an easy work around to make it accept 1 channel if there is any at all.<\/p>\n\n<p>you need to repeat your data in third dimension and have the same grayscale image in all 3 bands instead of RGB, that works just fine.<\/p>\n\n<p>if your image has the shape of : (224,224,1):<\/p>\n\n<pre><code>import numpy as np\ngray_image_3band = np.repeat(gray_img, repeats = 3, axis = -1)\n<\/code><\/pre>\n\n<p>if your image has the shape of : (224,224)<\/p>\n\n<pre><code>gray_image_3band = np.repeat(gray_img[..., np.newaxis], repeats = 3, axis = -1)\n<\/code><\/pre>\n\n<p>you don't need to call the model.build() anymore this way, keep the input layer. but if you ever wanted to call it you need to pass the shape as a tuple like this:<\/p>\n\n<pre><code>model.build( (224, 224, 1) ) # this is correct, notice the parentheses\n<\/code><\/pre>\n"}]}
+{"QuestionId":57914611,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-estimator>","CreationDate":"2019-09-12T21:16:03.557","AcceptedAnswerId":null,"OwnerUserId":8725045.0,"Title":"Has tf.Estimator become obsolete in Tensorflow 2.0?","Body":"<p>Today I've set up a custom model using its tf.Estimator high-level API in Tensorflow 2.0.\nIt was a pain in the *** to get it running, and there are very few complete examples online that implement custom Estimators in Tensorflow 2, which made me questioning the reasons for using this API.<\/p>\n\n<p>According to the docs, the main advantages of using the tf.Estimator API are:<\/p>\n\n<ol>\n<li><p>You can run Estimator-based models on a local host or on a distributed multi-server environment without changing your model. Furthermore, you can run Estimator-based models on CPUs, GPUs, or TPUs without recoding your model.<\/p><\/li>\n<li><p>You no longer have to worry about creating the computational graph or sessions since Estimators handle all the \"plumbing\" for you<\/p><\/li>\n<\/ol>\n\n<p>Advantage 2. clearly doesn't apply to Tensorflow 2.0 anymore, as it runs in eager mode by default, so you don't have to worry about sessions anyways.<\/p>\n\n<p>Advantage 1. also seems quite irrelevant in Tensorflow 2.0 - using tf.distribute.Strategy, you can now easily run even high level tf.Keras models in a distributed fashion and on CPUs\/GPUs\/TPUs.\ntf.Keras models are so much easier and faster to set up, so why did they even bother to keep the tf.Estimator API in Tensorflow 2.0? Are there other advantages of using this API?<\/p>\n","answers":[]}
+{"QuestionId":57915007,"AnswerCount":1,"Tags":"<tensorflow><out-of-memory><gpu>","CreationDate":"2019-09-12T22:00:05.193","AcceptedAnswerId":null,"OwnerUserId":866082.0,"Title":"How to measure (manually) how much of my GPU memory is used \/ available","Body":"<p>I'm experimenting with Tensorflow by creating different models and testing them. The problem I'm having right now is that I don't have a clear sense of how big my model could be before I face the OOM (out of memory) error. For sure, I can keep adding layers or making them bigger and see when I'm hitting the limit. But it would be nice if I could measure how much memory my models are occupying.<\/p>\n\n<p>Using <code>nvidia-smi<\/code> command I can see this:<\/p>\n\n<pre><code>+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 430.40       Driver Version: 430.40       CUDA Version: 10.1     |\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 Off |                  N\/A |\n| 14%   56C    P2    87W \/ 280W |  10767MiB \/ 11178MiB |     37%      Default |\n+-------------------------------+----------------------+----------------------+\n\n+-----------------------------------------------------------------------------+\n| Processes:                                                       GPU Memory |\n|  GPU       PID   Type   Process name                             Usage      |\n|=============================================================================|\n|    0     24329      C   python                                     10747MiB |\n+-----------------------------------------------------------------------------+\n<\/code><\/pre>\n\n<p>But the problem I have with the memory usage here is that no matter how small or big my model is, it always shows the same number! As if Tensorflow will take all no matter what.<\/p>\n\n<p>I'm looking for a formula that gives me the memory needed by providing it the number of model parameters, batch size, and sample size.<\/p>\n","answers":[{"AnswerId":"57915486","CreationDate":"2019-09-12T23:09:43.300","ParentId":null,"OwnerUserId":"867889","Title":null,"Body":"<p>You can use next function:<\/p>\n\n<pre><code>def get_max_memory_usage(sess):\n    \"\"\"Might be unprecise. Run after training\"\"\"\n    if sess is None: sess = tf.get_default_session()\n    max_mem = int(sess.run(tf.contrib.memory_stats.MaxBytesInUse()))\n    print(\"Contrib max memory: %f G\" % (max_mem \/ 1024. \/ 1024. \/ 1024.))\n    return max_mem\n<\/code><\/pre>\n\n<p>Or look into <a href=\"https:\/\/github.com\/yaroslavvb\/memory_util\" rel=\"nofollow noreferrer\">mem_utils<\/a> by Vladislav Bulatov<\/p>\n"}]}
+{"QuestionId":57915055,"AnswerCount":0,"Tags":"<python><windows><powershell><tensorflow><command-line>","CreationDate":"2019-09-12T22:06:24.850","AcceptedAnswerId":null,"OwnerUserId":9179887.0,"Title":"How to Fix \"Failed to load the native TensorFlow runtime\" Error","Body":"<p>I'm trying to run I program I've written for Tensorflow from a file in Notepad using Powershell. Whenever I run <code>python main.py<\/code> to open and execute the code, I get the error below.<\/p>\n\n<p>I haven't really found anything to try, and the only thing I felt I could do is uninstall and re-install Tensorflow, but that isn't working.<\/p>\n\n<pre><code>C:\\Users\\stxrsstillfade&gt; cd tensorflow-demo\nC:\\Users\\stxrsstillfade\\tensorflow-demo&gt; python main.py\n<\/code><\/pre>\n\n<p>The code error isn't even with the code. The traceback is saying that the problem is with the first line, <code>import tensorflow<\/code>.<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"main.py\", line 1, in &lt;module&gt;\n    import tensorflow as tf\n  File \"C:\\Python27\\lib\\site-packages\\tensorflow\\__init__.py\", line 24, in &lt;module&gt;\n    from tensorflow.python import *\n  File \"C:\\Python27\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 72, in &lt;module&gt;\n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"C:\\Python27\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 66, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"C:\\Python27\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow = swig_import_helper()\n  File \"C:\\Python27\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 20, in swig_import_helper\n    import _pywrap_tensorflow\nImportError: No module named _pywrap_tensorflow\n\n\nFailed to load the native TensorFlow runtime.\n\nSee https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/get_started\/os_setup.md#import_error\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've tried getting to the GitHub link, but it seems to have been deleted.<\/p>\n\n<p>Thank you for your answer in advance.<\/p>\n","answers":[]}
+{"QuestionId":57915086,"AnswerCount":0,"Tags":"<c++><caffe2><vcpkg><dumpbin>","CreationDate":"2019-09-12T22:10:33.340","AcceptedAnswerId":null,"OwnerUserId":8844604.0,"Title":"Problem with dumpbin.exe using vcpkg for compile caffe2 win10","Body":"<p>When try to use <code>vcpkg<\/code> with <code>vs2015<\/code> to install <code>caffe2:x64-windows<\/code> in windows 10 I obtain an error with <code>dumpbin.exe<\/code>: <\/p>\n\n<blockquote>\n  <p>Dump of file E:\\vcpkg\\packages\\caffe2_x64-windows\\debug\\lib\\Caffe2_CPU.lib\n      LINK : fatal error LNK1104: cannot open file 'E:\\vcpkg\\packages\\caffe2_x64-windows\\debug\\lib\\Caffe2_CPU.lib'<\/p>\n<\/blockquote>\n\n<p>I think that the problem is about the version of <code>dumpbin.exe<\/code> used in the process, it can't open big lib files, this one is bigger than 2 GB, but in the directory <code>C:\\Program Files (x86)\\Microsoft Visual Studio 14.0\\VC\\bin<\/code> I found directories like amd64_x86 with other version of <code>dumpbin<\/code> but I dont know how to configure <code>vcpkg<\/code> to use this version of <code>dumpbin<\/code> program.<\/p>\n\n<p><strong>Resolved<\/strong><\/p>\n\n<p>I resolve my problem checking the code in the solution of vcpkg, I changed all the paths of the type \"VC\" \/ \"bin\" for \"VC\" \/ \"amd64_x86\", in this way I use the chaintool used by vcpkg, after compiling caffe2 I undo my changes.<\/p>\n","answers":[]}
+{"QuestionId":57915233,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><transfer-learning><vgg-net>","CreationDate":"2019-09-12T22:28:13.493","AcceptedAnswerId":null,"OwnerUserId":7065438.0,"Title":"How to add customm layers inside vgg16 when doing transfer learning?","Body":"<p>I am trying to use transfer learning using vgg16. My main concept is to train the first few layers of vgg16, and add my own layer, afterwords add the rest of the layers from vgg16, and add my own output layer to the end. To do this I follow this sequence: (1) load layers and freez layers, (2) add my layers, (3) load the rest of layers (except the output layer) [THIS IS WHERE I ENCOUNTER THE FOLLOWING ERROR] and freez the layer, (4) add output layer. Is my approach ok? If not, then where I am doing wrong? Here's the error:<\/p>\n\n<blockquote>\n  <p>ValueError: Input 0 is incompatible with layer block3_conv1: expected axis -1 of input shape to have value 128 but got shape (None, 64, 56, 64)<\/p>\n<\/blockquote>\n\n<p>The full code is here for better understanding:<\/p>\n\n<pre><code>    vgg16_model= load_model('Fetched_VGG.h5')\n    vgg16_model.summary()\n\n    model= Sequential()\n\n    #add vgg layer (inputLayer, block1, block2)\n    for layer in vgg16_model.layers[0:6]:\n        model.add(layer)\n\n    #frees\n    # Freezing the layers (Oppose weights to be updated)\n    for layer in model.layers:\n        layer.trainable = False\n\n    #add custom\n    model.add(Conv2D(64, (3, 3), activation='relu', padding='same', name='block66_conv1_m') )\n    model.add( Conv2D(64, (3, 3), activation='relu', padding='same', name='block66_conv2_m') )\n    model.add( Conv2D(64, (3, 3), activation='relu', padding='same', name='block66_conv3_m') )\n    model.add( MaxPooling2D((2, 2), strides=(2, 2), name='block66_pool_m'))\n\n    # add vgg layer (block 3 to last layer (except the output dense layer))\n    for layer in vgg16_model.layers[7:-1]:\n        model.add(layer)\n\n    # Freezing the layers (Oppose weights to be updated)\n    for layer in model.layers:\n        layer.trainable = False\n\n    # add out out layer\n    model.add(Dense(2, activation='softmax', name='predictions'))\n    model.summary()\n<\/code><\/pre>\n","answers":[{"AnswerId":"57925551","CreationDate":"2019-09-13T14:28:56.187","ParentId":null,"OwnerUserId":"10675934","Title":null,"Body":"<p>As VGG16 layer 7 is expecting 128 filters you'll need to match this with your final Conv2D<\/p>\n\n<pre><code>model.add( Conv2D(128, (3, 3), activation='relu', padding='same', name='block66_conv3_m') )\n<\/code><\/pre>\n\n<p>If the dimensions match you should be able to build your model but it's not clear what you're trying to achieve. Your approach of adding to the middle of the VGG16 model will mean that all the downstream layers will need to be retrained<\/p>\n"}]}
+{"QuestionId":57915333,"AnswerCount":1,"Tags":"<python><tensorflow><memory-leaks>","CreationDate":"2019-09-12T22:45:45.963","AcceptedAnswerId":null,"OwnerUserId":10289143.0,"Title":"Why do I have memory leaks when loading batches for training using Tensorflow?","Body":"<p>I am training a neural network using batches of images. I want to import the images in the for training loop to avoid importing all the images at once. Once a loop iteration is done, I want to use a new batch and forget about previous images to free the CPU memory. I am reallocating the variables every time but the memory (and thus the running time) keeps on increasing. Do you know how to free memory from the previous batches?<\/p>\n\n<p>I am using Python 3.6.8 and Tensorflow 1.14.0 on a GPU Tesla K80 (memory_limit:11.3 GB).<\/p>\n\n<p>I have tried gc.collect() but it does not work.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport tensorflow as tf\nimport matplotlib.pyplot as plt\nimport os\nimport time\nimport gc\nimport psutil\n\ndir_img = \"..\/dir_img_png\/data\/\"\ndata = [os.path.join(dir_img, f) for f in os.listdir(dir_img) if os.path.isfile(os.path.join(dir_img, f))]\nnp.random.shuffle(data)\ndata_train = data[10:]\ndata_test = data[:10]\n\n# Hyperparameters\ninput_size=256\nbatch_size=64\nepochs = 20\n\ndef memory():\n    pid = os.getpid()\n    py = psutil.Process(pid)\n    memoryUse = py.memory_info()[0]\/2.**30  # memory use in GB\n    print('memory use:', memoryUse)\n\ndef preprocess_image(path):\n    raw_img = tf.read_file(path)\n    img = tf.io.decode_png(raw_img, channels=1)\n    img = tf.cast(img, tf.float32)\n    img -= 127.5\n    img *= 1.\/127.5\n    return img\n\n# Networks IO\nreal_images = tf.placeholder(tf.float32, [None, input_size, input_size, 1])\n\ninit = tf.global_variables_initializer()\n\nsess = tf.Session()\nsess.run(init)\n\nfor epoch in range(epochs):\n    for batch in range(len(data_train)\/\/batch_size):\n        tps = time.time()\n        batch_images = np.array([sess.run(preprocess_image(d)) for d in data_train[batch*batch_size:batch*batch_size+batch_size]])\n        print(\"tps1: {t}\".format(t=time.time()-tps))\n\n        tps2 = time.time()\n        gc.collect()\n        print(\"tps2: {t}\".format(t=time.time()-tps2))\n\n        memory()\n\nsess.close()\n<\/code><\/pre>\n\n<p>Here is the output I get:<\/p>\n\n<pre><code>tps1: 10.445663928985596\ntps2: 0.0995786190032959\nmemory use: 0.871917724609375\ntps1: 9.142687320709229\ntps2: 0.10767912864685059\nmemory use: 0.9062271118164062\ntps1: 12.030094146728516\ntps2: 0.10630679130554199\nmemory use: 0.9415740966796875\ntps1: 13.415296077728271\ntps2: 0.11185669898986816\nmemory use: 0.9608650207519531\ntps1: 12.053950548171997\ntps2: 0.11706066131591797\nmemory use: 0.9794692993164062\ntps1: 14.279714584350586\ntps2: 0.11610865592956543\nmemory use: 0.9980583190917969\ntps1: 11.772900342941284\ntps2: 0.12384176254272461\nmemory use: 1.0166587829589844\ntps1: 15.43606686592102\ntps2: 0.12571096420288086\n<\/code><\/pre>\n\n<p>The memory and running time keep on increasing.<\/p>\n","answers":[{"AnswerId":"57918903","CreationDate":"2019-09-13T07:11:38.580","ParentId":null,"OwnerUserId":"10289143","Title":null,"Body":"<p>Every method tf.* (like tf.read_file, tf.io.decode_png or tf.cast) must be outside of the loop.<\/p>\n\n<p>We need to define the graph before the loop and then run only the tensor we want :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport tensorflow as tf\nimport os\nimport time\nimport psutil\n\ndir_img = \"..\/dir_img_png\/data\/\"\ndata = [os.path.join(dir_img, f) for f in os.listdir(dir_img) if os.path.isfile(os.path.join(dir_img, f))]\nnp.random.shuffle(data)\ndata_train = data[10:]\ndata_test = data[:10]\n\n# Hyperparameters\ninput_size=256\nbatch_size=64\nepochs = 20\n\ndef memory():\n    pid = os.getpid()\n    py = psutil.Process(pid)\n    memoryUse = py.memory_info()[0]\/2.**30  # memory use in GB\n    print('memory use:', memoryUse)\n\n# Networks IO\nimg_path = tf.placeholder(tf.string)\nreal_images = tf.placeholder(tf.float32, [None, input_size, input_size, 1])\n\n# Preprocess image\nraw_img = tf.read_file(img_path)\npng_img = tf.io.decode_png(raw_img, channels=1)\npng_img = tf.cast(png_img, tf.float32)\npng_img = tf.subtract(png_img, 127.5)\npng_img = tf.divide(png_img, 127.5)\n\ninit = tf.global_variables_initializer()\n\nsess = tf.Session()\nsess.run(init)\n\nfor epoch in range(epochs):\n    for batch in range(len(data_train)\/\/batch_size):\n        tps = time.time()\n        batch_images = np.array([sess.run(png_img, feed_dict={img_path: d}) for d in data_train[batch*batch_size:batch*batch_size+batch_size]])\n        print(\"tps1: {t}\".format(t=time.time()-tps))\n\n        memory()\n\nsess.close()\n<\/code><\/pre>\n\n<p>The output shows then that the memory is constant:<\/p>\n\n<pre><code>tps1: 11.155524492263794\nmemory use: 0.8702735900878906\ntps1: 9.429716110229492\nmemory use: 0.8859291076660156\ntps1: 9.62732195854187\nmemory use: 0.8859291076660156\ntps1: 11.327840089797974\nmemory use: 0.8859291076660156\ntps1: 8.580215215682983\nmemory use: 0.8859291076660156\ntps1: 11.039035081863403\nmemory use: 0.8859291076660156\ntps1: 12.866767168045044\nmemory use: 0.8859291076660156\ntps1: 12.425249576568604\nmemory use: 0.8859291076660156\n<\/code><\/pre>\n"}]}
+{"QuestionId":57915767,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-12T23:53:40.073","AcceptedAnswerId":null,"OwnerUserId":2457019.0,"Title":"Keras permute_dimension causes weird tensor shapes","Body":"<p>I'm running into an issue with a model I'm trying to build. I've been trying to debug it and ran into an oddity that I think may be the cause, but I'm not sure what I'm doing wrong. I've reduced what I think the problem is into a small snippet you can run on colab.<\/p>\n\n<p>Here's a colab where you can try running this:\n<a href=\"https:\/\/colab.research.google.com\/drive\/1pSTwCwMFGlWgJOP3gn9WF6pZq2CiP4XJ\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/1pSTwCwMFGlWgJOP3gn9WF6pZq2CiP4XJ<\/a><\/p>\n\n<pre><code>import keras\nfrom keras.layers import Layer, Dense, Input, Reshape\nimport keras.backend as K\n\nclass SimplePermute(Layer):\n    def __init__(self, **kwargs):\n        super(SimplePermute, self).__init__(**kwargs)\n\n    def call(self, inputs, **kwargs):\n        return K.permute_dimensions(inputs, [0,2,1])\ntest_i = Input(shape=(10, 256))\ntest = SimplePermute()(test_i)\nprint(test.get_shape())\nprint(K.int_shape(test))\ntest = Dense(units=100, activation=\"softmax\", name=\"sft2\")(test)\nprint(test.get_shape())\nprint(K.int_shape(test))\n<\/code><\/pre>\n\n<p>I'd expect the second series of prints to print the permuted tensor shape - that is [?, 256, 10]. However, the K.int_shape() returns [?, 10, 256], while TF's get_shape() returns the properly permuted shape. <\/p>\n\n<p>I believe this internal mismatch is causing the errors I'm seeing downstream in the model.<\/p>\n","answers":[{"AnswerId":"57916820","CreationDate":"2019-09-13T03:06:40.410","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Your custom layer doesn't have the <code>compute_output_shape<\/code> method implemented. This is what Keras uses to determine the <code>_keras_shape<\/code> property of the tensors, which is returned by <code>K.int_shape<\/code>.   <\/p>\n\n<ul>\n<li>You can use the standard <code>Permute((2,1))<\/code> layer.    <\/li>\n<li>Or you can use a <code>Lambda(lambda x: K.permute_dimensions(x, [0,2,1]))<\/code> layer.     <\/li>\n<li>Or you can implement the <code>compute_output_shape<\/code> method:  <\/li>\n<\/ul>\n\n<p>.<\/p>\n\n<pre><code>def compute_output_shape(self, input_shape): \n    return (input_shape[0], input_shape[2], input_shape[1])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57915922,"AnswerCount":1,"Tags":"<tensorflow><rasa>","CreationDate":"2019-09-13T00:23:33.190","AcceptedAnswerId":null,"OwnerUserId":11498751.0,"Title":"How to use lib rasa without the tensorflow pipeline","Body":"<p>I am trying to use the rasa lib to create a chatbot and am getting the following error while trying to create the project:<\/p>\n\n<p><code>The TensorFlow library was compiled to use AVX instructions, but these aren't available on your machine.<\/code><\/p>\n\n<p>So I would like to know how I can install the rasa lib without using the tensorflow pipeline.<\/p>\n\n<p>Can anyone help?<\/p>\n","answers":[{"AnswerId":"57950810","CreationDate":"2019-09-16T05:26:04.693","ParentId":null,"OwnerUserId":"3265215","Title":null,"Body":"<p>You can choose the <a href=\"https:\/\/rasa.com\/docs\/rasa\/nlu\/choosing-a-pipeline\/\" rel=\"nofollow noreferrer\">Pretrained Embeddings<\/a> pipeline, that is based on spaCy and sklearn. To do this, change the pipeline line in your <code>config.yml<\/code> file to <\/p>\n\n<pre><code>pipeline: \"pretrained_embeddings_spacy\"\n<\/code><\/pre>\n\n<p>Although, it's worth figuring out why Tensorflow is not working on your machine, as there are use cases for which Tensorflow supervised embeddings <a href=\"https:\/\/blog.rasa.com\/rasa-nlu-in-depth-part-1-intent-classification\/\" rel=\"nofollow noreferrer\">make more sense<\/a>. People in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/24548\" rel=\"nofollow noreferrer\">this thread<\/a> recommend installing Anaconda as a solution. <\/p>\n"}]}
+{"QuestionId":57915955,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensor>","CreationDate":"2019-09-13T00:28:30.890","AcceptedAnswerId":null,"OwnerUserId":3120842.0,"Title":"3D tensor subsetting operation","Body":"<p>I'm trying to perform a specific operation on a 3D tensor. I'll just show it below using a 3D numpy array instead. Basically I want to do the same operation as below but on an equivalent tensor.<\/p>\n\n<pre><code>a3 = np.array([[[1,0,0],[1,0,0],[1,0,0],[1,0,0],[0,1,0],[0,1,0],[0,0,1],[0,0,1]], [[1,0,0],[1,0,0],[1,0,0],[1,0,0],[0,0,1],[0,0,1],[0,0,1],[0,0,1]], [[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\nexpression = (a3.sum(axis=(1,2)) &gt;= 1)\n\na3[expression, :, 1].flatten()\n<\/code><\/pre>\n\n<p>It should return a list of the same length as the first dimension of the initial 3D tensor.<\/p>\n\n<p>I've tried to write some code based on tf.constant, but it has not worked so far. Here's my attempt:<\/p>\n\n<pre><code>a3 = tf.convert_to_tensor(a3)\none = tf.constant(1, dtype=tf.int32)\nmask = tf.keras.backend.greater_equal(tf.keras.backend.sum(a3,axis=(1,2)),one)\ntf.boolean_mask(a3,mask,axis=0)\nsess = tf.compat.v1.Session()\n\nwith sess.as_default():\n    assert tf.compat.v1.get_default_session() is sess\n    print(a3.eval())\n<\/code><\/pre>\n","answers":[{"AnswerId":"57916097","CreationDate":"2019-09-13T00:56:34.753","ParentId":null,"OwnerUserId":"3120842","Title":null,"Body":"<p>Ok figured it out:<\/p>\n\n<pre><code>a3 = tf.convert_to_tensor(a3)\none = tf.constant(1, dtype=tf.int32)\nmask = tf.keras.backend.greater_equal(tf.keras.backend.sum(a3,axis=(1,2)),one)\na4masked = tf.boolean_mask(a3,mask,axis=0)\na4flat = tf.reshape(a4masked, [tf.size(a4masked)])\n\nsess = tf.compat.v1.Session()\n\nwith sess.as_default():\n  assert tf.compat.v1.get_default_session() is sess\n  print(a4flat.eval())\n<\/code><\/pre>\n"}]}
+{"QuestionId":57915956,"AnswerCount":2,"Tags":"<opencv><tensorflow><caffe><tensorflow-lite><caffe2>","CreationDate":"2019-09-13T00:28:31.280","AcceptedAnswerId":null,"OwnerUserId":11981194.0,"Title":"Which is faster when deploying cnn models by TensorFlow Lite, Caffe2 or OpenCV?","Body":"<p>We can deploy MobileNet on Smartphone by TensorFlow Lite, Caffe2 or OpenCV, and I think Caffe2 will provide the best performance with higher fps. But why? Is the performance gap between them so large? Thanks.<\/p>\n","answers":[{"AnswerId":"57947914","CreationDate":"2019-09-15T20:20:49.763","ParentId":null,"OwnerUserId":"12071381","Title":null,"Body":"<p>You should probably go for TensorFlow Lite. Last I looked, Caffe2 had almost zero smartphone GPU support, while TFLite now supports both iOS and many Android devices (all that have OpenGLES >=3.1). Using the GPU generally makes things several times faster, and you can reduce the inference precision to half-float (FP16) with TFLite for even more speed and not too much of a performance hit.<\/p>\n\n<p>When you can't use the mobile GPU, you'll probably want to quantize your network to int8, which is easily doable with TensorFlow and TensorFlow Lite, whether during or after training. Caffe2 seems to need QNNPACK for quantization, which is claimed to be as much as 2 times faster. The catch is that it only works with two pre-trained models that they released (<a href=\"https:\/\/github.com\/pytorch\/QNNPACK\/issues\/12\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/QNNPACK\/issues\/12<\/a>), so you can't convert your own model.<\/p>\n\n<p>So I can't really think of a reason to use Caffe2 over TFLite.<\/p>\n\n<p>I'm not sure about OpenCV's DNN module, but I seriously doubt it has mobile GPU support. There's a slight chance it has quantization.<\/p>\n"},{"AnswerId":"57917534","CreationDate":"2019-09-13T05:02:16.987","ParentId":null,"OwnerUserId":"192373","Title":null,"Body":"<p>Each framework introduces their own optimizations, the result may be significantly different for different devices.<\/p>\n"}]}
+{"QuestionId":57916295,"AnswerCount":0,"Tags":"<python><machine-learning><keras><scikit-learn><neural-network>","CreationDate":"2019-09-13T01:33:15.557","AcceptedAnswerId":null,"OwnerUserId":10844173.0,"Title":"Grid Search in ANN keras model - TypeError: can't pickle _thread.RLock objects","Body":"<p>I am trying to perform grid search on an ANN model. The model alone worked, but when I try to tune it with Grid Search I get the error:<\/p>\n\n<p>TypeError: can't pickle _thread.RLock objects<\/p>\n\n<p>I saw in other threads that Keras doesn't allow pickling and one of the parameters people get wrong is the \"n_jobs = -1\", but I specifically set it as 1, just to be sure.<\/p>\n\n<pre><code>import pandas as pd\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.wrappers.scikit_learn import KerasClassifier\nfrom sklearn.model_selection import GridSearchCV\nfrom sklearn.model_selection import train_test_split\n\n# Reading data\nX_dum = pd.read_csv(\"X_dum.csv\")\nY = pd. read_csv(\"Y.csv\")\n\n# Splitting the dataset into the Training set and Test set\nX_train, X_test, y_train, y_test = train_test_split(X_dum, Y, test_size = 0.2)\n\ndef createANN():\n\n    # Initialising the ANN\n    model = Sequential()\n    model.add(Dense(units = 200, kernel_initializer = 'uniform', activation = 'relu', input_dim = 224))\n    model.add(Dense(units = 150, kernel_initializer = 'uniform', activation = 'relu'))\n    model.add(Dense(units = 100, kernel_initializer = 'uniform', activation = 'relu'))\n    model.add(Dense(units = 50, kernel_initializer = 'uniform', activation = 'relu'))\n    model.add(Dense(units = 1, kernel_initializer = 'uniform', activation = 'sigmoid'))\n    # Compiling the ANN\n    model.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])\n\n    return model\n\nprint(\"Grid Search - ANN\")\n# Wrap Keras model so it can be used by scikit-learn\nann_sk = KerasClassifier(build_fn = createANN(), epochs = 5, batch_size = 10)\n\n# Applying Grid Search to find the best model and the best parameters\n\nparameters = [{'batch_size': [10, 100]}]\ngrid_search = GridSearchCV(estimator = ann_sk, param_grid = parameters, cv = 3, n_jobs = 1)\ngrid_search = grid_search.fit(X_train, y_train)\nann_best_accuracy = grid_search.best_score_\nann_best_parameters = grid_search.best_params_\n\n# Results\nprint(\"Accuracy: {0}\".format(ann_best_accuracy))\nprint(\"Parameters: {0}\".format(ann_best_parameters))\n<\/code><\/pre>\n\n<p>I get the following error stream:<\/p>\n\n<pre><code>\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\spyder_kernels\\customize\\spydercustomize.py\", line 827, in runfile\n    execfile(filename, namespace)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\spyder_kernels\\customize\\spydercustomize.py\", line 110, in execfile\n    exec(compile(f.read(), filename, 'exec'), namespace)\n\n  File \"C:\/XXXXXXXXXXXXX\/GridTest.py\", line 37, in &lt;module&gt;\n    grid_search = grid_search.fit(X_train, y_train)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\sklearn\\model_selection\\_search.py\", line 632, in fit\n    base_estimator = clone(self.estimator)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\sklearn\\base.py\", line 64, in clone\n    new_object_params[name] = clone(param, safe=False)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\sklearn\\base.py\", line 55, in clone\n    return copy.deepcopy(estimator)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 180, in deepcopy\n    y = _reconstruct(x, memo, *rv)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 280, in _reconstruct\n    state = deepcopy(state, memo)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 150, in deepcopy\n    y = copier(x, memo)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 240, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 180, in deepcopy\n    y = _reconstruct(x, memo, *rv)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 280, in _reconstruct\n    state = deepcopy(state, memo)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 150, in deepcopy\n    y = copier(x, memo)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 240, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 180, in deepcopy\n    y = _reconstruct(x, memo, *rv)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 280, in _reconstruct\n    state = deepcopy(state, memo)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 150, in deepcopy\n    y = copier(x, memo)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 240, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 180, in deepcopy\n    y = _reconstruct(x, memo, *rv)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 280, in _reconstruct\n    state = deepcopy(state, memo)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 150, in deepcopy\n    y = copier(x, memo)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 240, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 180, in deepcopy\n    y = _reconstruct(x, memo, *rv)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 280, in _reconstruct\n    state = deepcopy(state, memo)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 150, in deepcopy\n    y = copier(x, memo)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 240, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 180, in deepcopy\n    y = _reconstruct(x, memo, *rv)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 280, in _reconstruct\n    state = deepcopy(state, memo)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 150, in deepcopy\n    y = copier(x, memo)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 240, in _deepcopy_dict\n    y[deepcopy(key, memo)] = deepcopy(value, memo)\n\n  File \"C:\\ProgramData\\Anaconda3\\lib\\copy.py\", line 169, in deepcopy\n    rv = reductor(4)\n\nTypeError: can't pickle _thread.RLock objects\n<\/code><\/pre>\n\n<p>I really cannot see the mistake here.<\/p>\n","answers":[]}
+{"QuestionId":57916527,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-13T02:14:33.950","AcceptedAnswerId":null,"OwnerUserId":12061039.0,"Title":"Superposition of conv2d(s) to function as a conv3d","Body":"<p>Working on Tensorflow Lite for a mobile project. I am trying to convert an existing model based on Tensorflow 1.x to Tensorflow Lite. It's encountering an issue on the compatibility of conv3d with Tensorflow Lite.<\/p>\n\n<p>Since higher dimension convolution is a superposition of lower dimension convolutions. Would it be possible to achieve it by a set of conv2d(s) to function exactly like a conv3d in Tensorflow Lite? Or Tensorflow 2.x, since it is a superset of Tensorflow Lite.<\/p>\n\n<p><em>For example:<\/em><\/p>\n\n<pre><code>        update_dense = tf.compat.v1.layers.dense(fc7,8192,activation=tf.nn.leaky_relu,use_bias=True)\n        update_dense = tf.reshape(update_dense, [4, 4, 4, -1, 128]) \n\n        reset_dense = tf.compat.v1.layers.dense(fc7,8192,activation=tf.nn.leaky_relu,use_bias=True)\n        reset_dense = tf.reshape(reset_dense, [4, 4, 4, -1, 128]) \n\n        hidden_dense = tf.compat.v1.layers.dense(fc7,8192,activation=tf.nn.leaky_relu,use_bias=True)\n        hidden_dense = tf.reshape(hidden_dense, [4, 4, 4, -1, 128])  \n\n        t_x_s_update = tf.nn.conv3d(prev_hidden, weights['update_gate'], strides=[1, 1, 1, 1, 1], padding=\"SAME\") + update_dense\n        t_x_s_update = tf.add(t_x_s_update, biases['update_gate']) \n        update_gate = tf.sigmoid(t_x_s_update)\n\n        t_x_s_reset = tf.nn.conv3d(prev_hidden, weights['reset_gate'], strides=[1, 1, 1, 1, 1], padding=\"SAME\") + reset_dense\n        t_x_s_reset = tf.add(t_x_s_reset, biases['reset_gate'])\n        reset_gate = tf.sigmoid(t_x_s_reset)\n\n\n        hidden_gate = tf.nn.conv3d(reset_gate * prev_hidden, weights['hidden_gate'], strides=[1, 1, 1, 1, 1], padding=\"SAME\") + hidden_dense\n        hidden_gate = tf.add(hidden_gate,biases['hidden_gate'])\n<\/code><\/pre>\n\n<p>Does anyone can provide a guide or better code form of solution on making conv3d by conv2d in Tensorflow?<\/p>\n","answers":[]}
+{"QuestionId":57916648,"AnswerCount":0,"Tags":"<pytorch>","CreationDate":"2019-09-13T02:36:12.417","AcceptedAnswerId":null,"OwnerUserId":12061114.0,"Title":"Pytorch - For loop LSTM and Trying to backward through the graph a second time, but the buffers have already been freed","Body":"<p>I'm a pytorch beginner. I have a for loop LSTM in my model for some reason. When I train the model in the second epoch I got the RuntimeError like this:\nTrying to backward through the graph a second time, but the buffers have already been freed. Specify retain_graph=True when calling backward the first time.<\/p>\n\n<p>I tried to use retain_graph=True in loss, but this causes the training process really slow.<\/p>\n\n<p>My network architecture as follows: LSTM and full connect layer<\/p>\n\n<pre><code>def __init__(self,embedding_dim,hidden_size,batch_size,compare_num):\n    super(MultiLayerClassifer,self).__init__()\n    self.studentPath_layers=compare_num\n    self.embedding_dim=embedding_dim\n    self.hidden_size=hidden_size\n    self.batch_size=batch_size\n    self.lstm=nn.LSTM(self.embedding_dim*3,hidden_size,batch_first=True)\n    self.l1=nn.Linear(hidden_size,int(hidden_size\/2))\n    self.l2=nn.Linear(int(hidden_size\/2),1)\n    self.allLinear1=nn.Linear(self.studentPath_layers,class_num)\n<\/code><\/pre>\n\n<p>And forward:<\/p>\n\n<pre><code>def forward(self,inputs):\n    sList=[]\n    for i in range(self.studentPath_layers):\n        if i &gt;= inputs.size()[1]:\n            break\n        sentence=inputs[:,i]\n\n        sentence_embedding=sentence.view\n        (sentence.shape[0],sentence.shape[1],-1)\n\n        self.hidden=self.init_hidden(sentence.shape[0])\n        output,self.hidden=self.lstm(sentence_embedding,self.hidden)\n\n        s=self.l1(output[:,-1,:])\n        s=self.l2(s)\n        s=torch.tanh(s)\n        sList.append(s)\n    sVariable=torch.stack(sList,-1)\n    y=self.allLinear1(sVariable)\n    return y\n<\/code><\/pre>\n\n<p>Train process in an epoch as follows:<\/p>\n\n<pre><code>for studentIdList,courseIdList,scoreList,sentenceList,embeddingList,typeList in t:\n\n    model.train()\n\n    embeddings_in=embeddingList.to(device)\n    target=torch.tensor([tag_to_ix[s] for s in     scoreList],dtype=torch.long).to(device)\n\n\n    output=model(embeddings_in)\n    output=output.reshape((output.shape[0],3)).to(device)\n    train_acc+=(torch.max(output, 1)[1]==target).sum().item()\n\n    loss=loss_function(output,target)\n    if weight_decay &gt; 0:\n        loss = loss + reg_loss(model)\n    train_loss+=loss.item()\n\n    optimizer.zero_grad()\n    loss.backward()\n\n    optimizer.step()\n    train_num+=embeddings_in.shape[0]\n    step+=1\n<\/code><\/pre>\n\n<p>Can anyone help me? I think the for loop cause the problem, but I'm confused after tried several methods.<\/p>\n","answers":[]}
+{"QuestionId":57916801,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-13T03:03:24.950","AcceptedAnswerId":null,"OwnerUserId":10811770.0,"Title":"Multiplying tensor with tensordot, keeping batch dimension","Body":"<p>I want to do a tensor multiplikation were all images in the batch are handled separately.<\/p>\n\n<p>What i want to do is equivalent to<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>dot = tf.stack([ tf.tensordot(f1[i], f2[i], [2, 2]) for i in range(batchsize) ], axis=0)\n<\/code><\/pre>\n\n<p>where f1 and f2 are tensor with 3 dimensions. I was wondering if there is a more direct way to get this tensor, especially since batchsize is not always known and concatenating could cost speed?\nIs the variation with summing by hand faster?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>f1 = tf.reshape(f1, [-1, hw, 1, c])\nf2 = tf.reshape(f2, [-1, 1, hw, c])\ndot = tf.reduce_sum(f1*f2, axis=-1)\n<\/code><\/pre>\n\n<p>I'm not sure if tensorflow will create that f1*f2 tensor before summing it which would cost vram and speed?\nIs there a most elegant way?<\/p>\n","answers":[]}
+{"QuestionId":57916843,"AnswerCount":0,"Tags":"<tensorflow><tpu>","CreationDate":"2019-09-13T03:11:22.423","AcceptedAnswerId":null,"OwnerUserId":4071437.0,"Title":"tf.where - Is there any work around for using tf.where with TPUEstimator","Body":"<p>I run the LSTM model for NLP task and have additional function for getting particular elements of output_layer (before concat them and softmax) whose locations\/indexs are found by matching word_ids in input_tensor (words are already embedded into ids). Search function is tf.where().<\/p>\n\n<p>The problem is \"Error recorded from training_loop: Compilation failure: Detected unsupported operations when trying to compile graph... Where (No registered 'Where' OpKernel for XLA_TPU_JIT devices compatible\"<\/p>\n\n<p>I also found similar issue here: <a href=\"https:\/\/github.com\/tensorflow\/tpu\/issues\/236\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tpu\/issues\/236<\/a><\/p>\n\n<p>My question is \"Is there any ways to work around that issue in my case?\"<\/p>\n\n<p>My additional function is below. Thanks for your help.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def get_entity_tensor(input_ten, output_ten, sign=999):\n  sign = tf.constant([sign], dtype=tf.int32) # input_ids has tf.int32 type\n  p = tf.where(tf.equal(input_ten, sign)) # tf.where return tf.int64 type\n  ent = tf.gather_nd(output_ten, p)\n  return ent\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57916950,"AnswerCount":1,"Tags":"<c><tensorflow><dyld>","CreationDate":"2019-09-13T03:28:59.247","AcceptedAnswerId":null,"OwnerUserId":3985177.0,"Title":"Faulty TensorFlow C install on MacOS?: dyld errors \"Library not loaded: @rpath\/libtensorflow.1.dylib\" and \"Symbol not found: __cg_DGifCloseFile\"","Body":"<h1>Question<\/h1>\n\n<p>How should I install the TensorFlow C library on MacOS or otherwise address the errors:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>dyld: Library not loaded: @rpath\/libtensorflow.1.dylib\n  Referenced from: \/Users\/Me\/.\/hello_tf\n  Reason: image not found\nAbort trap: 6\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>dyld: Symbol not found: __cg_DGifCloseFile\n  Referenced from: \/System\/Library\/Frameworks\/ImageIO.framework\/Versions\/A\/ImageIO\n  Expected in: \/usr\/local\/lib\/libGIF.dylib\n in \/System\/Library\/Frameworks\/ImageIO.framework\/Versions\/A\/ImageIO\nAbort trap: 6\n<\/code><\/pre>\n\n<h1>Context<\/h1>\n\n<p>I followed the <a href=\"https:\/\/www.tensorflow.org\/install\/lang_c\" rel=\"nofollow noreferrer\">installation instructions<\/a> for MacOS using the \u201cmacOS CPU only\u201d library verbatim until the <code>sudo ldconfig<\/code> bit.  This being Mac, I substituted <code>update_dyld_shared_cache<\/code> for <code>ldconfig<\/code> and restarted, then built the example program provided.<\/p>\n\n<pre class=\"lang-c prettyprint-override\"><code>#include &lt;stdio.h&gt;\n#include &lt;tensorflow\/c\/c_api.h&gt;\nint main() {\n  printf(\"Hello from TensorFlow C library version %s\\n\", TF_Version());\n  return 0;\n}\n<\/code><\/pre>\n\n<p>I got:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>~$ gcc hello_tf.c -ltensorflow -o hello_tf\n~$ .\/hello_tf\ndyld: Library not loaded: @rpath\/libtensorflow.1.dylib\n  Referenced from: \/Users\/Me\/.\/hello_tf\n  Reason: image not found\nAbort trap: 6\n<\/code><\/pre>\n\n<p>Following the instructions\u2019 advice, I rebuild with <code>gcc -I\/usr\/local\/include -L\/usr\/local\/lib hello_tf.c -ltensorflow -o hello_tf<\/code> and got the same result.<\/p>\n\n<p>The installation instructions say to set <code>LIBRARY_PATH<\/code> and <code>DYLD_LIBRARY_PATH<\/code> for installs in non-system directories.  I decided to give it a shot:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>export LIBRARY_PATH=$LIBRARY_PATH:\/usr\/local\/lib\nexport DYLD_LIBRARY_PATH=$DYLD_LIBRARY_PATH:\/usr\/local\/lib\n<\/code><\/pre>\n\n<p>Compiling with both methods (that is, without and with <code>-I\/usr\/local\/include -L\/usr\/local\/lib<\/code>) now give me:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>dyld: Symbol not found: __cg_DGifCloseFile\n  Referenced from: \/System\/Library\/Frameworks\/ImageIO.framework\/Versions\/A\/ImageIO\n  Expected in: \/usr\/local\/lib\/libGIF.dylib\n in \/System\/Library\/Frameworks\/ImageIO.framework\/Versions\/A\/ImageIO\nAbort trap: 6\n<\/code><\/pre>\n\n<p>(Well, at least it\u2019s a different error.)<\/p>\n\n<p>My <code>\/usr\/local\/lib<\/code> contains the TensorFlow files:<\/p>\n\n<pre><code>libtensorflow.1.14.0.dylib\nlibtensorflow.1.dylib\nlibtensorflow.dylib\nlibtensorflow_framework.1.14.0.dylib\nlibtensorflow_framework.1.dylib\nlibtensorflow_framework.dylib\n<\/code><\/pre>\n\n<p><code>libtensorflow.dylib<\/code> and <code>libtensorflow.1.dylib<\/code> are links that point to <code>libtensorflow.1.14.0.dylib<\/code>, and <code>libtensorflow_framework.dylib<\/code> and <code>libtensorflow_framework.1.dylib<\/code> likewise point to <code>libtensorflow.1.14.0.dylib<\/code>, as in the provided <code>.tar.gz<\/code> archive.<\/p>\n\n<h1>Desired Outcome<\/h1>\n\n<p>The primary desired effect is that the following occurs:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>~$ gcc hello_tf.c -ltensorflow -o hello_tf\n~$ .\/hello_tf\nHello from TensorFlow C library version 1.14.0\n<\/code><\/pre>\n\n<p>(That\u2019s the version I\u2019m using.)<\/p>\n\n<p>A secondary desired effect is understanding of why the mentioned errors occurred.<\/p>\n","answers":[{"AnswerId":"57937424","CreationDate":"2019-09-14T16:37:58.150","ParentId":null,"OwnerUserId":"3985177","Title":null,"Body":"<p>I fixed it by installing with Homebrew, adding <code>\/usr\/local\/lib<\/code> to <code>LD_LIBRARY_PATH<\/code>, and compiling with the <code>-ltensorflow<\/code> flag as the last of the GCC arguments.  It seems dyld is rather picky; if possible, I\u2019d still like to have a better explanation than that.<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>~$ brew install libtensorflow\n # Homebrew output omitted\n~$ export LD_LIBRARY_PATH=\"${LD_LIBRARY_PATH}:\/usr\/local\/lib\"\n~$ gcc hello_tf.c -o hello_tf -ltensorflow\n~$ .\/hello_tf\nTensorFlow C lib version 1.14.0\n<\/code><\/pre>\n\n<p>This also didn\u2019t work with the modifications to <code>DYLD_LIBRARY_PATH<\/code> and <code>LIBRARY_PATH<\/code> mentioned in the question.<\/p>\n\n<p>It might be wise to add <code>\/usr\/local\/lib<\/code> to <code>LD_LIBRARY_PATH<\/code> in <code>~\/.bash_profile<\/code> or some build script.<\/p>\n"}]}
+{"QuestionId":57917039,"AnswerCount":0,"Tags":"<python-3.x><image-processing><pytorch>","CreationDate":"2019-09-13T03:45:56.330","AcceptedAnswerId":null,"OwnerUserId":1812810.0,"Title":"Facing issue in understanding and porting pytorch code to keras","Body":"<p><a href=\"https:\/\/github.com\/uoguelph-mlrg\/Cutout\/blob\/master\/util\/cutout.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/uoguelph-mlrg\/Cutout\/blob\/master\/util\/cutout.py<\/a>\nThis is the code for patching of images, called cutout which leads to regularisation. The code uses torch which I need to migrate to keras.\nCode snippet :<\/p>\n\n<pre><code>mask = torch.from_numpy(mask)\nmask = mask.expand_as(img)\nimg = img * mask\n<\/code><\/pre>\n\n<p>Now mask is numpy array(32,32).\nAnd img is array(32,32,3). \nAnd how do I implement this in keras?\nThis may be a simple code for those with torch knowledge, which is new to me.<\/p>\n","answers":[]}
+{"QuestionId":57917398,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-13T04:44:46.180","AcceptedAnswerId":57917847.0,"OwnerUserId":3120842.0,"Title":"Error for tf.sets.intersection: ValueError: Shape must be at least rank 2 but is rank 1 for 'DenseToDenseSetOperation_12'","Body":"<p>I've reproduced and isolated my error below. I'm super new to Tensorflow. Is this error occurring because I didn't initialise the idx1 and idx2 variables correctly? If so, what do I have to do?<\/p>\n\n<p>The error I get is \"ValueError: Shape must be at least rank 2 but is rank 1 for 'DenseToDenseSetOperation_12' (op: 'DenseToDenseSetOperation') with input shapes: [3], [3].\"<\/p>\n\n<pre><code>import tensorflow as tf\n\nidx1 = [1,2,3]\nidx2 = [2,4,5]\n\nintersection = tf.sets.intersection(tf.convert_to_tensor(idx1), tf.convert_to_tensor(idx2))\nsess = tf.compat.v1.Session()\nwith sess.as_default():\n  assert tf.compat.v1.get_default_session() is sess\n  print(intersection.eval())\n<\/code><\/pre>\n","answers":[{"AnswerId":"57917847","CreationDate":"2019-09-13T05:40:22.440","ParentId":null,"OwnerUserId":"7217806","Title":null,"Body":"<p>The error states that your <code>idx<\/code> variables to not meet the requirements of the <code>intersection<\/code> method (see the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/rank\" rel=\"nofollow noreferrer\">docs<\/a> for details about the <code>rank<\/code>). As an example, you should not get this error using a two dimensional array:<\/p>\n\n<pre><code>idx1 = [[1, 2, 3]]\nidx2 = [[2, 4, 5]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57917450,"AnswerCount":1,"Tags":"<tensorflow><tensorflow.js>","CreationDate":"2019-09-13T04:51:21.940","AcceptedAnswerId":57925803.0,"OwnerUserId":1490302.0,"Title":"How to get\/set weights for a supervised model in tensorflow.js?","Body":"<p>I'd like to change the weights of a supervised model but I get the same exact result after changing the weights.  What am I doing wrong?<\/p>\n\n<pre><code>const model = tf.sequential();\nmodel.add(tf.layers.dense({...}));\nmodel.add(tf.layers.dense({...}));\nmodel.add(tf.layers.dense({...}));\nmodel.compile({...});\nmodel.fit({});\n\nconst result1 = model.predict(tf.tensor2d(...)).dataSync();\n\nconst newWeights = [];\nmodel.layers.map((layer, i) =&gt; {\n  newWeights[i] = []\n  const weights = layer.getWeights();\n  newWeights[i][0] = weights[0].arraySync()\n  newWeights[i][1] = weights[1].arraySync()\n\n  newWeights[i][0].map(tensor =&gt; tensor.map(x =&gt; {\n    if (random(1) &lt; 0.5) {\n      return x + offset();\n    }\n\n    return x;\n  })\n\n  layer.setWeights([tf.tensor2d(newWeights[i][0], [newWeights[i][0].length, newWeights[i][0][0].length]), tf.tensor(newWeights[i][1])])\n})\n\nconst result2 = model.predict(tf.tensor2d(...)).dataSync();\n<\/code><\/pre>\n\n<p>Code snippets:<\/p>\n\n<pre><code>const random = (max) =&gt; {\n  return floor(Math.random() * Math.floor(max), 2);\n} \n\nconst floor = (num, toDecimal) =&gt; {\n  let  dec = Math.pow(10, toDecimal);\n  return Number(Math.floor(num * dec) \/ dec);\n}\n\nconst offset = () =&gt; {\n  randomGaussian() * 0.5\n}\n\nlet previous = false;\nlet y2 = 0;\nconst randomGaussian = (mean, sd) =&gt; {\n  let y1, x1, x2, w;\n  if (previous) {\n    y1 = y2;\n    previous = false;\n  } else {\n    do {\n      x1 = random(2) - 1;\n      x2 = random(2) - 1;\n      w = x1 * x1 + x2 * x2;\n    } while (w &gt;= 1);\n    w = Math.sqrt(-2 * Math.log(w) \/ w);\n    y1 = x1 * w;\n    y2 = x2 * w;\n    previous = true;\n  }\n\n  let m = mean || 0;\n  let s = sd || 1;\n  return y1 * s + m;\n};\n<\/code><\/pre>\n\n<p>result1 === result2 but why?<\/p>\n","answers":[{"AnswerId":"57925803","CreationDate":"2019-09-13T14:45:11.530","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>Most likely that the new weights are identical to that of the first model.<\/p>\n\n<p><strong>Example: Simple example to change weights of a model<\/strong>\n<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() =&gt; {\r\nconst model = tf.sequential({\r\n        layers: [tf.layers.dense({units: 1, inputShape: [10]})]\r\n    });\r\n    model.compile({optimizer: 'sgd', loss: 'meanSquaredError'});\r\n    for (let i = 1; i &lt; 5 ; ++i) {\r\n      const h = await model.fit(tf.ones([8, 10]), tf.ones([8, 1]), {\r\n          batchSize: 4,\r\n          epochs: 3\r\n      });\r\n      console.log(\"Loss after Epoch \" + i + \" : \" + h.history.loss[0]);\r\n    }\r\n    \r\n    const p = await model.predict(tf.zeros([1, 10]))\r\n    p.print()\r\n    const layers = model.layers\r\n\r\n    layers[0].setWeights([tf.zeros([10, 1]), tf.zeros([1])])\r\n    \r\n    const q = await model.predict(tf.zeros([1, 10]))\r\n    q.print()\r\n\r\n\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@latest\"&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><strong>Issue of the code<\/strong><\/p>\n\n<p>The <code>newWeights<\/code> created is not assigned to <code>newWeights<\/code>. <code>map<\/code> is not an in-place operator. The array returned by <code>map<\/code> should be assigned back to <code>newWeights<\/code>.<\/p>\n\n<pre><code>newWeights[i][0] = newWeights[i][0].map(tensor =&gt; tensor.map(x =&gt; {\n    if (random(1) &lt; 0.5) {\n      return x + offset();\n    }\n\n    return x;\n  })\n<\/code><\/pre>\n"}]}
+{"QuestionId":57917458,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-13T04:52:34.787","AcceptedAnswerId":null,"OwnerUserId":9014519.0,"Title":"hyperparameter by looking at the .meta file in tensorflow? and Can I change the meta file to be human readable?","Body":"<p>I'm a pytorch user. However, TensorFlow offers pretrained weights. We need hyperparameters to reproduce. Can I read meta files and know hyperparameters?<\/p>\n","answers":[]}
+{"QuestionId":57917795,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2019-09-13T05:34:51.513","AcceptedAnswerId":57926597.0,"OwnerUserId":10018602.0,"Title":"How to re-load a saved model (with graph?) to create the same results on future testing data?","Body":"<p>I have trained a model and saved all the files (meta, index, checkpoint, etc.) using the <code>saver = tf.compat.v1.train.Saver()<\/code> function, and now I want to re-load that model in order to test on new data. It works fine, but my question is, every time I run the restored model on the same dataset (i.e. run it once on a testing dataset, and then start it over and run it again on that same dataset) I get very different results. I'm hoping to be able to run it over and over again on the same dataset, but get the same results.<\/p>\n\n<p>I have two separate .py files, one for training and one for testing\/loading the model to test on the dataset. My training variables\/placeholders look something like this in the training.py file (in case it's relevant):<\/p>\n\n<pre><code>    # set some tensorflow variables and placeholders, etc.\n    self.X = tf.compat.v1.placeholder(tf.float32, (None, self.state_size))                                                                  \n    self.REWARDS = tf.compat.v1.placeholder(tf.float32, (None))\n    self.ACTIONS = tf.compat.v1.placeholder(tf.int32, (None))\n    feed_forward = tf.layers.dense(self.X, self.LAYER_SIZE, activation = tf.nn.relu)\n    self.logits = tf.layers.dense(feed_forward, self.OUTPUT_SIZE, activation = tf.nn.softmax)\n    input_y = tf.one_hot(self.ACTIONS, self.OUTPUT_SIZE)\n    loglike = tf.math.log((input_y * (input_y - self.logits) + (1 - input_y) * (input_y + self.logits)) + 1) # tf.log\n    rewards = tf.tile(tf.reshape(self.REWARDS, (-1,1)), [1, self.OUTPUT_SIZE])\n    self.cost = -tf.reduce_mean(loglike * (rewards + 1)) # leave this as a negative, so that the minimize function of the Adam optimizer will keep improving \n    # Adam Optimizer\n    self.optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate = self.LEARNING_RATE).minimize(self.cost) # minimize(self.cost)\n    # Start the Tensorflow session\n    self.sess = tf.compat.v1.InteractiveSession()\n    self.sess.run(tf.compat.v1.global_variables_initializer())\n...\n            saver = tf.compat.v1.train.Saver()\n            save_path = saver.save(self.sess, \".\/agent_output\/\" + name + \"_model\")\n<\/code><\/pre>\n\n<p>And in the testing.py file, it looks something like this:<\/p>\n\n<pre><code>    ...\n    # Start the Tensorflow session\n    self.sess = tf.compat.v1.InteractiveSession()\n\n    new_saver = tf.train.import_meta_graph('.\/agent_output\/' + name + '_model.meta')\n    new_saver.restore(self.sess, tf.train.latest_checkpoint('.\/agent_output\/'))\n    print('Model loaded step 1')\n\n    #saver = tf.compat.v1.train.Saver()\n    #saver.restore(self.sess, \".\/agent_output\/\" + name + \"_model\")\n    #print('Model Restored!')\n\n    self.sess.run(tf.compat.v1.global_variables_initializer())\n<\/code><\/pre>\n\n<p>Just to give you an idea of what I'm working with. As you can see, I've tried the import_meta_graph and the commented out <code>saver.restore<\/code> method, but I think I'm missing something, or if it's even possible in my case?<\/p>\n\n<p>I'm just hoping someone can point me in the right direction. What I've discovered on my own is that there should be a way to not only load the variables, but also the graph? Or maybe I need to implement that during the training? I'm running Python 3.6 and Tensorflow 1.14 (I believe? Not 2.0).<\/p>\n","answers":[{"AnswerId":"57926597","CreationDate":"2019-09-13T15:39:27.040","ParentId":null,"OwnerUserId":"2160809","Title":null,"Body":"<p>Your problem is probably running <code>self.sess.run(tf.compat.v1.global_variables_initializer())<\/code> after restoring the model. You only need to run this for a fresh model not a restored one. Try it without this line.<\/p>\n"}]}
+{"QuestionId":57917981,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-09-13T05:53:27.800","AcceptedAnswerId":57920676.0,"OwnerUserId":5939036.0,"Title":"Pytorch how to stack tensor like for loop","Body":"<p>I want to concat tensor generated in for loop, and get 2dTensor.<br>\nstandard python, like below.<\/p>\n\n<pre><code>li = []\nfor i in range(0, len(items)):\n    # calc something\n    li.append(calc_result)\n<\/code><\/pre>\n\n<p>In my case, in for loop, generate <code>torch.Size([768])<\/code> Tensor, and I want to get <code>torch.Size([len(item),768])<\/code> Tensor.<br>\nHow to do this?<\/p>\n","answers":[{"AnswerId":"57920676","CreationDate":"2019-09-13T09:15:26.073","ParentId":null,"OwnerUserId":"4892874","Title":null,"Body":"<p>You can use <code>torch.stack<\/code>:<\/p>\n\n<pre><code>torch.stack(li, dim=0)\n<\/code><\/pre>\n\n<p>after the for loop will give you a <code>torch.Tensor<\/code> of that size.<\/p>\n\n<p>Note that if you know in advance the size of the final tensor, you can allocate an empty tensor beforehand and fill it in the for loop:<\/p>\n\n<pre><code>x = torch.empty(size=(len(items), 768))\nfor i in range(len(items)):\n    x[i] = calc_result\n<\/code><\/pre>\n\n<p>This is usually faster than doing the stack.<\/p>\n"}]}
+{"QuestionId":57918070,"AnswerCount":0,"Tags":"<python><tensorflow><ubuntu><python-3.6><tensorboard>","CreationDate":"2019-09-13T06:01:17.997","AcceptedAnswerId":null,"OwnerUserId":10575386.0,"Title":"ImportError: TensorBoard","Body":"<p>I'm using Python in Anaconda environment (jupyter notebook)<\/p>\n\n<p>OS: Ubuntu\nTensorflow version: 1.14.0\nPython version: 3.6<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorboard\/issues\/1425\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorboard\/issues\/1425<\/a>\nThis is  the same issue,but the answer didn't help. \nIs there a workaround for this issue other than trying to reinstall  Tensorflow? Or should I upgrade my tensorflow version? <\/p>\n\n<pre><code>    from keras.callbacks import  TensorBoard\n\n    tensorboard = TensorBoard(log_dir='graph_1', histogram_freq=0, \n    batch_size=512, write_graph=True, write_grads=False, \n    write_images=False, embeddings_freq=0, embeddings_layer_names=None, \n    embeddings_metadata=None, embeddings_data=None, update_freq='epoch')\n<\/code><\/pre>\n\n<p>throws an error like:<\/p>\n\n<pre><code>ImportError                               Traceback (most recent call last)\n~\/anaconda3\/lib\/python3.6\/site-packages\/keras\/callbacks.py in __init__(self, log_dir, histogram_freq, batch_size, write_graph, write_grads, write_images, embeddings_freq, embeddings_layer_names, embeddings_metadata, embeddings_data, update_freq)\n    744             import tensorflow as tf\n--&gt; 745             from tensorflow.contrib.tensorboard.plugins import projector\n    746         except ImportError:\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/__init__.py in &lt;module&gt;\n     40 from tensorflow.contrib import distribute\n---&gt; 41 from tensorflow.contrib import distributions\n     42 from tensorflow.contrib import estimator\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/distributions\/__init__.py in &lt;module&gt;\n     43   from tensorflow.contrib.distributions.python.ops.distribution_util import tridiag\n---&gt; 44   from tensorflow.contrib.distributions.python.ops.estimator import *\n     45   from tensorflow.contrib.distributions.python.ops.geometric import *\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/distributions\/python\/ops\/estimator.py in &lt;module&gt;\n     20 \n---&gt; 21 from tensorflow.contrib.learn.python.learn.estimators.head import _compute_weighted_loss\n     22 from tensorflow.contrib.learn.python.learn.estimators.head import _RegressionHead\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/learn\/__init__.py in &lt;module&gt;\n     92 # pylint: disable=wildcard-import\n---&gt; 93 from tensorflow.contrib.learn.python.learn import *\n     94 # pylint: enable=wildcard-import\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/learn\/python\/__init__.py in &lt;module&gt;\n     27 # pylint: disable=wildcard-import\n---&gt; 28 from tensorflow.contrib.learn.python.learn import *\n     29 # pylint: enable=wildcard-import\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/__init__.py in &lt;module&gt;\n     29 from tensorflow.contrib.learn.python.learn import datasets\n---&gt; 30 from tensorflow.contrib.learn.python.learn import estimators\n     31 from tensorflow.contrib.learn.python.learn import graph_actions\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/__init__.py in &lt;module&gt;\n    301 from tensorflow.contrib.learn.python.learn.estimators.constants import ProblemType\n--&gt; 302 from tensorflow.contrib.learn.python.learn.estimators.dnn import DNNClassifier\n    303 from tensorflow.contrib.learn.python.learn.estimators.dnn import DNNEstimator\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/dnn.py in &lt;module&gt;\n     33 from tensorflow.contrib.learn.python.learn import metric_spec\n---&gt; 34 from tensorflow.contrib.learn.python.learn.estimators import dnn_linear_combined\n     35 from tensorflow.contrib.learn.python.learn.estimators import estimator\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/dnn_linear_combined.py in &lt;module&gt;\n     35 from tensorflow.contrib.learn.python.learn import metric_spec\n---&gt; 36 from tensorflow.contrib.learn.python.learn.estimators import estimator\n     37 from tensorflow.contrib.learn.python.learn.estimators import head as head_lib\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/estimator.py in &lt;module&gt;\n     51 from tensorflow.contrib.learn.python.learn.estimators._sklearn import NotFittedError\n---&gt; 52 from tensorflow.contrib.learn.python.learn.learn_io import data_feeder\n     53 from tensorflow.contrib.learn.python.learn.utils import export\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/learn_io\/__init__.py in &lt;module&gt;\n     36 from tensorflow.contrib.learn.python.learn.learn_io.graph_io import read_keyed_batch_features_shared_queue\n---&gt; 37 from tensorflow.contrib.learn.python.learn.learn_io.numpy_io import numpy_input_fn\n     38 from tensorflow.contrib.learn.python.learn.learn_io.pandas_io import extract_pandas_data\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/learn_io\/numpy_io.py in &lt;module&gt;\n     25 \n---&gt; 26 from tensorflow.python.estimator.inputs.numpy_io import numpy_input_fn as core_numpy_input_fn\n     27 from tensorflow.python.util.deprecation import deprecated\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/estimator\/inputs\/numpy_io.py in &lt;module&gt;\n     25 \n---&gt; 26 from tensorflow_estimator.python.estimator.inputs import numpy_io\n     27 \n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow_estimator\/python\/estimator\/inputs\/numpy_io.py in &lt;module&gt;\n     25 \n---&gt; 26 from tensorflow_estimator.python.estimator.inputs.queues import feeding_functions\n     27 from tensorflow.python.util.tf_export import estimator_export\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow_estimator\/__init__.py in &lt;module&gt;\n      9 \n---&gt; 10 from tensorflow_estimator._api.v2 import estimator\n     11 _names_with_underscore = []\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow_estimator\/_api\/v2\/estimator\/__init__.py in &lt;module&gt;\n      9 \n---&gt; 10 from tensorflow_estimator._api.v2.estimator import experimental\n     11 from tensorflow_estimator._api.v2.estimator import export\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow_estimator\/_api\/v2\/estimator\/experimental\/__init__.py in &lt;module&gt;\n     10 from tensorflow_estimator.python.estimator.canned.linear import LinearSDCA\n---&gt; 11 from tensorflow_estimator.python.estimator.canned.rnn import RNNClassifier\n     12 from tensorflow_estimator.python.estimator.canned.rnn import RNNEstimator\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow_estimator\/python\/estimator\/canned\/rnn.py in &lt;module&gt;\n     22 \n---&gt; 23 from tensorflow.python.feature_column import dense_features\n     24 from tensorflow.python.feature_column import feature_column_lib as fc\n\nImportError: cannot import name 'dense_features'\n\nDuring handling of the above exception, another exception occurred:\n\nImportError                               Traceback (most recent call last)\n&lt;ipython-input-83-0d8f7236c1d1&gt; in &lt;module&gt;\n     16 reduce_lr_1 = ReduceLROnPlateau(monitor = 'val_loss', factor = 0.2, patience = 1, verbose = 1, min_delta = 0.0001)\n     17 \n---&gt; 18 tensorboard_1 = TensorBoard(log_dir='graph_1', histogram_freq=0, batch_size=512, write_graph=True, write_grads=False, write_images=False, embeddings_freq=0, embeddings_layer_names=None, embeddings_metadata=None, embeddings_data=None, update_freq='epoch')\n     19 \n     20 callbacks_1 = [checkpoint_1,earlystop_1,tensorboard_1,reduce_lr_1]\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/keras\/callbacks.py in __init__(self, log_dir, histogram_freq, batch_size, write_graph, write_grads, write_images, embeddings_freq, embeddings_layer_names, embeddings_metadata, embeddings_data, update_freq)\n    745             from tensorflow.contrib.tensorboard.plugins import projector\n    746         except ImportError:\n--&gt; 747             raise ImportError('You need the TensorFlow module installed to '\n    748                               'use TensorBoard.')\n    749 \n\nImportError: You need the TensorFlow module installed to use TensorBoard.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57918128,"AnswerCount":1,"Tags":"<keras><nlp>","CreationDate":"2019-09-13T06:08:12.670","AcceptedAnswerId":null,"OwnerUserId":11452211.0,"Title":"How does Keras tokenizer handle unseen data?","Body":"<p>Say I have trained a tokenizer with some vocabulary, now what exactly does keras tokenizer() do when it encounters a word which was not present in vocabulary(training data)  does it simply ignore it?<\/p>\n","answers":[{"AnswerId":"57919151","CreationDate":"2019-09-13T07:30:38.367","ParentId":null,"OwnerUserId":"9579075","Title":null,"Body":"<p>According to the documentation if you set the oov_token you should be able to handle unseen words.<\/p>\n\n<pre><code>oov_token`: if given, it will be added to word_index and used\n            to replace out-of-vocabulary words during text_to_sequence calls\n<\/code><\/pre>\n"}]}
+{"QuestionId":57918572,"AnswerCount":1,"Tags":"<keras><scikit-learn>","CreationDate":"2019-09-13T06:47:46.697","AcceptedAnswerId":null,"OwnerUserId":3120842.0,"Title":"Keras metric equivalent to scikit learn's average precision score metric","Body":"<p>I've had a look at the Keras metrics documentation and couldn't find an equivalent to scikit learn's <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.metrics.average_precision_score.html\" rel=\"nofollow noreferrer\">average precision score<\/a> metric (which I think is the same as the area under the precision-recall curve, AUPRC). It isn't the same as the average_precision_at_k, I believe unless someone can correct me on that.<\/p>\n","answers":[{"AnswerId":"57958580","CreationDate":"2019-09-16T14:05:21.747","ParentId":null,"OwnerUserId":"3660313","Title":null,"Body":"<p>You can implement custom metrics for keras to be passed at the compilation step. (<a href=\"https:\/\/keras.io\/metrics\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/metrics\/<\/a>) The function would need to take (y_true, y_pred) as arguments and return a single tensor value.<\/p>\n\n<p>Here is an implementation of average_precision for keras:<\/p>\n\n<pre><code>import keras.backend as K\n\ndef average_precision(y_true, y_pred):\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<\/code><\/pre>\n"}]}
+{"QuestionId":57918716,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-13T06:59:38.293","AcceptedAnswerId":null,"OwnerUserId":3282532.0,"Title":"Keras: is KMP_AFFINITY a critical output?","Body":"<p>My main concern is if the message \"KMP_AFFINITY\" is something I should react about. The problem doesnt occur on my laptop with CPU, but only on my PC with GPU. First I tried to run my own Keras implementation but received some warnings. Then I found some simple cnn code from <a href=\"https:\/\/raw.githubusercontent.com\/adventuresinML\/adventures-in-ml-code\/master\/keras_cnn.py\" rel=\"nofollow noreferrer\">github<\/a> to see if the same warnings occured.<\/p>\n\n<p>I tried to search through the internet, but what I did not understand from those replies, was if its a critical warning and if not, how to get rid of it. Am I missing some packages or something else? The output from anaconda prompt is shown below. After a while the script continues and after some time finishes. <\/p>\n\n<p>I found one solution to remove the warnings (<a href=\"https:\/\/stackoverflow.com\/questions\/56224689\/reduce-console-verbosity\">stackoverflow)<\/a> by including:<\/p>\n\n<pre><code>import os\nos.environ['KMP_WARNINGS'] = 'off'\n<\/code><\/pre>\n\n<p>But does the warnings mean something important or?<\/p>\n\n<pre><code>Train on 60000 samples, validate on 10000 samples\nEpoch 1\/10\n2019-09-13 08:36:03.426583: I tensorflow\/core\/platform\/cpu_feature_guard.cc:145]\n This TensorFlow binary is optimized with Intel(R) MKL-DNN to use the following\nCPU instructions in performance critical operations:  AVX AVX2\nTo enable them in non-MKL-DNN operations, rebuild TensorFlow with the appropriat\ne compiler flags.\n2019-09-13 08:36:03.525592: I tensorflow\/core\/common_runtime\/process_util.cc:115\n] Creating new thread pool with default inter op setting: 12. Tune using inter_o\np_parallelism_threads for best performance.\nOMP: Info #212: KMP_AFFINITY: decoding x2APIC ids.\nOMP: Info #210: KMP_AFFINITY: Affinity capable, using global cpuid leaf 11 info\nOMP: Info #154: KMP_AFFINITY: Initial OS proc set respected: 0-11\nOMP: Info #156: KMP_AFFINITY: 12 available OS procs\nOMP: Info #157: KMP_AFFINITY: Uniform topology\nOMP: Info #179: KMP_AFFINITY: 1 packages x 6 cores\/pkg x 2 threads\/core (6 total\n cores)\nOMP: Info #214: KMP_AFFINITY: OS proc to physical thread map:\nOMP: Info #171: KMP_AFFINITY: OS proc 0 maps to package 0 core 0 thread 0\nOMP: Info #171: KMP_AFFINITY: OS proc 1 maps to package 0 core 0 thread 1\nOMP: Info #171: KMP_AFFINITY: OS proc 2 maps to package 0 core 1 thread 0\nOMP: Info #171: KMP_AFFINITY: OS proc 3 maps to package 0 core 1 thread 1\nOMP: Info #171: KMP_AFFINITY: OS proc 4 maps to package 0 core 2 thread 0\nOMP: Info #171: KMP_AFFINITY: OS proc 5 maps to package 0 core 2 thread 1\nOMP: Info #171: KMP_AFFINITY: OS proc 6 maps to package 0 core 3 thread 0\nOMP: Info #171: KMP_AFFINITY: OS proc 7 maps to package 0 core 3 thread 1\nOMP: Info #171: KMP_AFFINITY: OS proc 8 maps to package 0 core 4 thread 0\nOMP: Info #171: KMP_AFFINITY: OS proc 9 maps to package 0 core 4 thread 1\nOMP: Info #171: KMP_AFFINITY: OS proc 10 maps to package 0 core 5 thread 0\nOMP: Info #171: KMP_AFFINITY: OS proc 11 maps to package 0 core 5 thread 1\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 11860 thread 0 bound to OS proc set\n0\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 8300 thread 1 bound to OS proc set 2\n\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 13076 thread 2 bound to OS proc set\n4\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 11736 thread 3 bound to OS proc set\n6\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 12028 thread 4 bound to OS proc set\n8\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 12700 thread 5 bound to OS proc set\n10\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 12608 thread 6 bound to OS proc set\n1\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 12232 thread 7 bound to OS proc set\n3\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 3580 thread 8 bound to OS proc set 5\n\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 11228 thread 9 bound to OS proc set\n7\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 13292 thread 10 bound to OS proc set\n 9\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 11672 thread 11 bound to OS proc set\n 11\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 12708 thread 12 bound to OS proc set\n 0\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 12464 thread 13 bound to OS proc set\n 2\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 11448 thread 14 bound to OS proc set\n 4\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 11800 thread 15 bound to OS proc set\n 6\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 3800 thread 16 bound to OS proc set\n8\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 10096 thread 17 bound to OS proc set\n 10\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 9532 thread 18 bound to OS proc set\n1\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 5248 thread 19 bound to OS proc set\n3\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 2600 thread 20 bound to OS proc set\n5\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 7976 thread 21 bound to OS proc set\n7\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 12804 thread 22 bound to OS proc set\n 9\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 1700 thread 23 bound to OS proc set\n11\n  128\/60000 [..............................] - ETA: 7:16 - loss: 2.3163 - acc: 0\n.0547OMP: Info #250: KMP_AFFINITY: pid 12848 tid 11852 thread 24 bound to OS pro\nc set 0\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 12864 thread 25 bound to OS proc set\n 2\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 12724 thread 26 bound to OS proc set 4\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 11544 thread 27 bound to OS proc set 6\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 13024 thread 28 bound to OS proc set 8\nOMP: Info #250: KMP_AFFINITY: pid 12848 tid 3372 thread 29 bound to OS proc set\n10\n...\n42624\/60000 [====================&gt;.........] - ETA: 11s - loss: 0.0207 - acc: 0.\n42752\/60000 [====================&gt;.........] - ETA: 11s - loss: 0.0208 - acc: 0.\n42880\/60000 [====================&gt;.........] - ETA: 11s - loss: 0.0208 - acc: 0.\n43008\/60000 [====================&gt;.........] - ETA: 11s - loss: 0.0208 - acc: 0.\n43136\/60000 [====================&gt;.........] - ETA: 11s - loss: 0.0208 - acc: 0.\n43264\/60000 [====================&gt;.........] - ETA: 11s - loss: 0.0208 - acc: 0.\n43392\/60000 [====================&gt;.........] - ETA: 10s - loss: 0.0209 - acc: 0.\n43520\/60000 [====================&gt;.........] - ETA: 10s - loss: 0.0209 - acc: 0.\n43648\/60000 [====================&gt;.........] - ETA: 10s - loss: 0.0208 - acc: 0.\n...\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57919276,"AnswerCount":0,"Tags":"<tensorflow><keras><conv-neural-network><training-data>","CreationDate":"2019-09-13T07:40:27.223","AcceptedAnswerId":null,"OwnerUserId":11505921.0,"Title":"Can i indicate how much images are needed for a Keras model?","Body":"<p>Lets say I have 1000 pictures and want to train a Model on them.\nI want to train it 10 times with limit = [100, 200, ... , 1000] to indicate how much data is truly needed for 90% accuracy.<\/p>\n\n<p>Can I pass these limits into a Keras model?<\/p>\n\n<p>Idealy it would look like this:<\/p>\n\n<pre><code>limit = [100, 200, ... , 1000]\nhistory = []\n\nfor x in limit\n   var = model.fit(limit, ...)\n   history.append( var )\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57919429,"AnswerCount":0,"Tags":"<tensorflow><object-detection-api>","CreationDate":"2019-09-13T07:51:47.307","AcceptedAnswerId":null,"OwnerUserId":5356625.0,"Title":"TensorFlow Object Detection API After training model the file pipeline.config is empty","Body":"<p>I try retraining model whith TensorFlow Object Detection API. In file pipeline.config I change:<\/p>\n\n<ul>\n<li>path<\/li>\n<li>num_classes:1<\/li>\n<li>batch_size: 1 (for speed of training)<\/li>\n<li>num_steps: 1000 (for speed of training)<\/li>\n<\/ul>\n\n<p>I am doing<\/p>\n\n<pre><code>C:\\Distr\\models-master\\models-master\\research\\object_detection&gt;python legacy\/train.py --logtostderr --train_dir=training\/ --pipeline_config_path=training\/pipeline.config\n<\/code><\/pre>\n\n<p>Training is successful, but pipeline.config is empty after training. I have a warning:<\/p>\n\n<pre><code>I0913 12:14:47.846972 28680 learning.py:507] global step 995: loss = 6.8707 (0.798 sec\/step)\nINFO:tensorflow:global step 996: loss = 2.7868 (0.820 sec\/step)\nI0913 12:14:48.669283 28680 learning.py:507] global step 996: loss = 2.7868 (0.820 sec\/step)\nINFO:tensorflow:global step 997: loss = 3.4453 (0.797 sec\/step)\nI0913 12:14:49.466856 28680 learning.py:507] global step 997: loss = 3.4453 (0.797 sec\/step)\nINFO:tensorflow:global step 998: loss = 5.1915 (0.804 sec\/step)\nI0913 12:14:50.271611 28680 learning.py:507] global step 998: loss = 5.1915 (0.804 sec\/step)\nINFO:tensorflow:global step 999: loss = 7.8720 (0.950 sec\/step)\nI0913 12:14:51.224293 28680 learning.py:507] global step 999: loss = 7.8720 (0.950 sec\/step)\nINFO:tensorflow:Recording summary at step 999.\nI0913 12:14:51.677052 21604 supervisor.py:1050] Recording summary at step 999.\nINFO:tensorflow:global_step\/sec: 1.20015\nI0913 12:14:52.015100 25292 supervisor.py:1099] global_step\/sec: 1.20015\nINFO:tensorflow:global step 1000: loss = 3.9892 (0.837 sec\/step)\nI0913 12:14:52.063099 28680 learning.py:507] global step 1000: loss = 3.9892 (0.837 sec\/step)\nINFO:tensorflow:Stopping Training.\nI0913 12:14:52.064100 28680 learning.py:777] Stopping Training.\nINFO:tensorflow:Finished training! Saving model to disk.\nI0913 12:14:52.064100 28680 learning.py:785] Finished training! Saving model to disk.\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\summary\\writer\\writer.py:386: UserWarning: Attempting to use a closed FileWriter. The operation will be a noop unless the FileWriter is explicitly reopened.\n  warnings.warn(\"Attempting to use a closed FileWriter. \"\nenter code here\n<\/code><\/pre>\n\n<p>When I try export model I have error. I use command:<\/p>\n\n<pre><code>python export_inference_graph.py --input_type image_tensor --pipeline_config_path training\/pipeline.config --trained_checkpoint_prefix training\/model.ckpt-1000 --output_directory inference_graph\n<\/code><\/pre>\n\n<p>Error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"export_inference_graph.py\", line 162, in &lt;module&gt;\n    tf.app.run()\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\platform\\app.py\", line 40, in run\n    _run(main=main, argv=argv, flags_parser=_parse_flags_tolerate_undef)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\absl\\app.py\", line 299, in run\n    _run_main(main, args)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\absl\\app.py\", line 250, in _run_main\n    sys.exit(main(argv))\n  File \"export_inference_graph.py\", line 158, in main\n    write_inference_graph=FLAGS.write_inference_graph)\n  File \"C:\\Distr\\models-master\\models-master\\research\\object_detection\\exporter.py\", line 473, in export_inference_graph\n    is_training=False)\n  File \"C:\\Distr\\models-master\\models-master\\research\\object_detection\\builders\\model_builder.py\", line 136, in build\n    raise ValueError('Unknown meta architecture: {}'.format(meta_architecture))\nValueError: Unknown meta architecture: None\n<\/code><\/pre>\n\n<p>If I change empty pipeline.config on initial, then export work succesfull.<\/p>\n\n<p>What needs to be fixed to get the correct pipeline.config file after training?<\/p>\n","answers":[]}
+{"QuestionId":57919530,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-13T07:59:32.620","AcceptedAnswerId":null,"OwnerUserId":826983.0,"Title":"Getting interrupted by signal 11: SIGSEGV","Body":"<p>All I know is that the error occurs when this branch gets executed and the <code>weights<\/code> from it get passed down to <code>tf.data.experimental.sample_from_datasets<\/code>:<\/p>\n\n<pre><code># ...\nelif pretrain_cfg.schedule == PretrainSchedule.CONVERGE_LINEARLY:\n    logger.info('[%s] - Pretrain: Using CONVERGE_LINEARLY schedule' % self.name)\n    a = tf.minimum(tf.constant(1.0, dtype=tf.float64, shape=(1,)), global_step \/ max_pretrain_steps)\n    b = tf.maximum(tf.constant(0.0, dtype=tf.float64, shape=(1,)), 1 - global_step \/ max_pretrain_steps)\n    weights = a * const_task_weights + b * pretrain_task_weights\n\nreturn tf.data.experimental.sample_from_datasets(datasets, weights=weights)\n<\/code><\/pre>\n\n<p>The following works:<\/p>\n\n<pre><code>weights = tf.cond(\n    tf.greater(global_step, max_pretrain_steps),\n    true_fn=lambda: const_task_weights,\n    false_fn=lambda: pretrain_task_weights\n)\n<\/code><\/pre>\n\n<p>but for some reason this here causes the <code>SIGSEGV<\/code>:<\/p>\n\n<pre><code>a = tf.minimum(tf.constant(1.0, dtype=tf.float64, shape=(1,)), global_step \/ max_pretrain_steps)\nb = tf.maximum(tf.constant(0.0, dtype=tf.float64, shape=(1,)), 1 - global_step \/ max_pretrain_steps)\nweights = a * const_task_weights + b * pretrain_task_weights\n<\/code><\/pre>\n\n<p>I don't really see what the problem is but the problem comes definitely from this line: <\/p>\n\n<pre><code>weights = a * const_task_weights + b * pretrain_task_weights\n<\/code><\/pre>\n\n<p>The question is why. It might not be valid to have a dependency to the <code>global_step<\/code> in this context as since the <code>weights<\/code> parameter of <code>sample_from_datasets<\/code>.<\/p>\n\n<p>However, in <code>sample_from_datasets<\/code> I don't see anything suspicious since inside <code>sample_from_datasets<\/code> the first thing that happens is<\/p>\n\n<pre><code>weights = ops.convert_to_tensor(weights, name=\"weights\")\n<\/code><\/pre>\n\n<p>So passing a tensor to it should be fine. <\/p>\n\n<p>Any ideas?<\/p>\n\n<hr>\n\n<p>Error output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>INFO:tensorflow:Running local_init_op.\nINFO:tensorflow:Done running local_init_op.\nINFO:tensorflow:Saving checkpoints for 0 into \/data\/translation\/multi-problem\/hi2en\/model\/512-3-1-1024\/de2en.hi2en\/c19cfad259cad911\/model.ckpt.\nbash: line 1:  4153 Segmentation fault      (core dumped) env \"CUDA_VISIBLE_DEVICES\"=\"0\" \"LIBRARY_ROOTS\"=\"\/Users\/username\/Library\/Caches\/PyCharm2018.2\/remote_sou...\n\nProcess finished with exit code 139 (interrupted by signal 11: SIGSEGV)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57923756","CreationDate":"2019-09-13T12:34:16.660","ParentId":null,"OwnerUserId":"826983","Title":null,"Body":"<p>Okay, it turns out the problem is <strong>not<\/strong> with tensorflow directly - actually not at all.<\/p>\n\n<p>The problem was because <code>const_task_weights<\/code> and <code>pretrain_task_weights<\/code> did not have the same shape. I did not validate the input and had a bug somewhere else.<\/p>\n\n<p>Just be aware that you might get this kind of error if the shapes do not match.<\/p>\n\n<p>I guess this cannot be checked or determined by tensorflow so that will be something the user has to take care of (citation needed).<\/p>\n"}]}
+{"QuestionId":57919790,"AnswerCount":1,"Tags":"<python><caffe><object-detection><mobilenet>","CreationDate":"2019-09-13T08:17:35.677","AcceptedAnswerId":null,"OwnerUserId":11571383.0,"Title":"Count number of persons in a video","Body":"<p>I have code from pyimagesearch \"<a href=\"https:\/\/www.pyimagesearch.com\/2017\/09\/18\/real-time-object-detection-with-deep-learning-and-opencv\/#comment-550946\" rel=\"nofollow noreferrer\">https:\/\/www.pyimagesearch.com\/2017\/09\/18\/real-time-object-detection-with-deep-learning-and-opencv\/#comment-550946<\/a>\" for object detection using caffe model. What I want to do is whenever a person is detected it should count and finally I want to count the number of person passed .<\/p>\n\n<p>I have tried something but it incrementing the count for every frame, even if the person is same and appeared in multiple frames , it is counting in all the frames.<\/p>\n\n<pre><code>import numpy as np\nimport argparse\nimport cv2\nimport imutils\n\n\n\nap = argparse.ArgumentParser()\nap.add_argument(\"-v\", \"--video\", default=\"1.avi\",\n    help=\"path to input video\")\nap.add_argument(\"-p\", \"--prototxt\", \ndefault=\"MobileNetSSD_deploy.prototxt.txt\",\n    help=\"path to Caffe 'deploy' prototxt file\")\nap.add_argument(\"-m\", \"--model\", default=\"MobileNetSSD_deploy.caffemodel\",\n    help=\"path to Caffe pre-trained model\")\nap.add_argument(\"-c\", \"--confidence\", type=float, default=0.6,\n    help=\"minimum probability to filter weak detections\")\nargs = vars(ap.parse_args())\n\nCLASSES = [\"background\", \"aeroplane\", \"bicycle\", \"bird\", \"boat\",\n    \"bottle\", \"bus\", \"car\", \"cat\", \"chair\", \"cow\", \"diningtable\",\n    \"dog\", \"horse\", \"motorbike\", \"person\", \"pottedplant\", \"sheep\",\n    \"sofa\", \"train\", \"tvmonitor\"]\n\nCOLORS = np.random.uniform(0, 255, size=(len(CLASSES), 3))\n\n\nprint(\"[INFO] loading model...\")\nnet = cv2.dnn.readNetFromCaffe(args[\"prototxt\"], args[\"model\"])\n\ncap = cv2.VideoCapture(\"1.mp4\")\n\n\ncount= 0\nwhile True:\n\n    ret, image = cap.read()\n    image = imutils.resize(image, width=500)\n\n    image = np.rot90(image)\n    image = np.rot90(image)\n    image = np.rot90(image)\n\n    (h, w) = image.shape[:2]\n    blob = cv2.dnn.blobFromImage(cv2.resize(image, (300, 300)), \n0.007843, (300, 300), 127.5)\n    net.setInput(blob)\n    detections = net.forward()\n    for i in np.arange(0, detections.shape[2]):\n        confidence = detections[0, 0, i, 2]\n        if confidence &gt; args[\"confidence\"]:\n\n            idx = int(detections[0, 0, i, 1])\n            box = detections[0, 0, i, 3:7] * np.array([w, h, \nw, h])\n            (startX, startY, endX, endY) = box.astype(\"int\")\n\n            label = \"{}: {:.2f}%\".format(CLASSES[idx], \nconfidence * 100)\n            print(\"[INFO] {}\".format(label))\n            if CLASSES[idx]==\"person\":\n                            count += 1 \n                            print(\"person\",count)\n\ncap.release()\ncv2.destroyAllWindows()\n<\/code><\/pre>\n\n<p>Actually what I want is whenever a person comes it should count one and when the next person comes it should count 2 like that the counting process should continue.<\/p>\n","answers":[{"AnswerId":"57949002","CreationDate":"2019-09-15T23:51:35.663","ParentId":null,"OwnerUserId":"11976900","Title":null,"Body":"<p>Even if the person disappears from the frame, the value \"person\" will remain in CLASSED[idx].<\/p>\n\n<p>So you Initialize CLASSED[idx] or need to take other steps<\/p>\n"}]}
+{"QuestionId":57919799,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><tensorflow2.0>","CreationDate":"2019-09-13T08:18:08.620","AcceptedAnswerId":null,"OwnerUserId":4373898.0,"Title":"Why does TF BaselineRegressor and a hand made Zero Rule Regressor do not predict the same thing?","Body":"<p>I know that TF\u00a0provides premade estimators, like the <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/estimator\/BaselineRegressor\" rel=\"nofollow noreferrer\"><code>BaselineRegressor<\/code><\/a>. From what I understand, this is a <a href=\"http:\/\/weka.sourceforge.net\/doc.dev\/weka\/classifiers\/rules\/ZeroR.html\" rel=\"nofollow noreferrer\">ZeroR<\/a> regressor, predicting the average value of the training labels.<\/p>\n\n<p>After training this estimator, I evaluated it and got these metrics:<\/p>\n\n<pre><code>'loss': 4037909.0,\n'average_loss': 4037909.0,\n\n'label\/mean': 1039.9038,\n'prediction\/mean': 4.178101\n<\/code><\/pre>\n\n<p>Here, the <code>prediction\/mean<\/code> is the mean computed during training (which is the value predicted by the estimator, as can be shown below:<\/p>\n\n<pre><code>&gt; estimator.predict(...) \n[{'predictions': array([4.1781044], dtype=float32)},\n {'predictions': array([4.1781044], dtype=float32)},\n ...,\n {'predictions': array([4.1781044], dtype=float32)}]\n<\/code><\/pre>\n\n<p>However, this is far from the expected mean which should be around 1000 (as for the <code>label\/mean<\/code>).<\/p>\n\n<p>Thus, I implement my own ZeroR algorithm:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def zero_rule_estimation(dataset):\n    fit_mean, num_batches = 0, 0\n    for sample in dataset.skip(1):\n        num_batches += 1\n        batch_mean = tf.math.reduce_mean(sample[1])\n        fit_mean = fit_mean + ((batch_mean - fit_mean) \/ num_batches) # Running average\n\n    label_mean, error, num_batches = 0, 0, 0\n    for sample in dataset.take(1):\n        num_batches += 1\n        batch_mean = tf.math.reduce_mean(sample[1])\n        label_mean = label_mean + ((batch_mean - label_mean) \/ num_batches)\n        error = tf.math.reduce_mean(tf.math.squared_difference(sample[1], fit_mean))\n\n    print(\"Label mean:\", label_mean.numpy())\n    print(\"Prediction mean:\", fit_mean.numpy())\n    print(\"ZR RMSE:\", tf.math.sqrt(error).numpy())\n<\/code><\/pre>\n\n<p>Running it gives these metrics:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>Label mean: 1039.9038\nPrediction mean: 1062.3342\nZR RMSE: 1722.1163\n<\/code><\/pre>\n\n<p>Here, the prediction mean is more like what I expected.<\/p>\n\n<p>So why does the BaselineRegressor and my ZeroR do not behave the same way?<\/p>\n\n<p>Thanks for your answer.\nAlexis.<\/p>\n","answers":[]}
+{"QuestionId":57919973,"AnswerCount":1,"Tags":"<deep-learning><pytorch><gradient-descent><generative-adversarial-network>","CreationDate":"2019-09-13T08:30:23.363","AcceptedAnswerId":null,"OwnerUserId":9151048.0,"Title":"Interpret GAN loss","Body":"<p>I am currently training the standard DCGAN network on my dataset. After 40 epochs, the loss of both generator and discriminator is 45-50. Can someone please explain the reason and possible solution for this?<\/p>\n","answers":[{"AnswerId":"57920316","CreationDate":"2019-09-13T08:53:24.500","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>This interpretation may be added to <a href=\"https:\/\/en.wikipedia.org\/wiki\/List_of_unsolved_problems_in_mathematics\" rel=\"nofollow noreferrer\">unsolved problems<\/a>.<\/p>\n\n<p>You cannot interpret the loss of generator and discriminator. Since when one improves it will be harder for the other. When generator improves it will be harder for the critic. When critic improves it will be harder for the generator.<\/p>\n\n<p>The values totally depend on your loss function. You may expect that numbers should be \"about the same\" over time.<\/p>\n"}]}
+{"QuestionId":57920089,"AnswerCount":0,"Tags":"<python><pytorch><gpytorch>","CreationDate":"2019-09-13T08:38:53.390","AcceptedAnswerId":null,"OwnerUserId":9875707.0,"Title":"What is `gpytorch.settings.max_preconditioner_size`? And `gpytorch.beta_features.checkpoint_kernel(checkpoint_size)`?","Body":"<p>I am following the tutorial <a href=\"https:\/\/gpytorch.readthedocs.io\/en\/latest\/examples\/01_Simple_GP_Regression\/Simple_MultiGPU_GP_Regression.html\" rel=\"nofollow noreferrer\">Simple_MultiGPU_GP_Regression<\/a> and I noticed that during the training two options were established:<\/p>\n\n<pre><code>    with gpytorch.beta_features.checkpoint_kernel(checkpoint_size), \\\n         gpytorch.settings.max_preconditioner_size(preconditioner_size):\n<\/code><\/pre>\n\n<p>What are or refer to the variables <code>checkpoint_kernel<\/code> and <code>preconditioner_size<\/code>?<\/p>\n\n<p>I have check the documentation <a href=\"https:\/\/gpytorch.readthedocs.io\/en\/latest\/settings.html#gpytorch.settings.max_preconditioner_size\" rel=\"nofollow noreferrer\">preconditioner_size<\/a> but it is not quite clear to me what it refers to. <\/p>\n\n<p>I intuit that <code>checkpoint_size<\/code> refers to something related to the number of training points associated with each GPU. But it is just an intuition.<\/p>\n\n<p>Help is appreciated. <\/p>\n\n<p>Thanks<\/p>\n","answers":[]}
+{"QuestionId":57920584,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-13T09:09:13.520","AcceptedAnswerId":57920765.0,"OwnerUserId":1849975.0,"Title":"Gather a slice given an idex array","Body":"<p>I have an input tensor <code>params<\/code> of size <code>(BxHxWx200)<\/code> and an index array <code>idx<\/code> of size <code>(Bx1000x2)<\/code>. <code>idx<\/code> holds 1000 different locations of <code>param<\/code> tensor, i.e. the last axis (the one of size 2) is <code>HxW<\/code> location which corresponds to the <code>params<\/code> input tensor. I want to get the slice of <code>params<\/code> of size <code>(Bx1000x200)<\/code> that corresponds to the locations in <code>idx<\/code>. <\/p>\n\n<p>In NumPy, it would be as simple as <code>params[idx]<\/code>. I am not sure how to do it in TensorFlow.<\/p>\n\n<p>If I try <code>output = tf.gather_nd(params, idx)<\/code> I am getting a tensor of size <code>(Bx1000xWx200)<\/code>. Why is that and how can I fix it?<\/p>\n","answers":[{"AnswerId":"57920765","CreationDate":"2019-09-13T09:20:25.510","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You need another index in the last dimension of <code>idx<\/code> to index the first axis of <code>params<\/code>:<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Input data (can be dynamic shapes)\nB, H, W = 10, 20, 30\nparams = tf.placeholder(tf.float32, [B, H, W, 200])\nidx = tf.placeholder(tf.int32, [B, 1000, 2])\n# Dimensions of idx\nidx_s = tf.shape(idx, out_type=idx.dtype)\n# Index for first dimension of params\nr = tf.range(idx_s[0])\n# Tile r to match the shape of idx\nidx_b = tf.tile(r[:, tf.newaxis, tf.newaxis], (1, idx_s[1], 1))\n# Make the complete index\nidx_full = tf.concat([idx_b, idx], axis=-1)\n# Gather the result\nout = tf.gather_nd(params, idx_full)\nprint(out.shape)\n# (10, 1000, 200)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57920604,"AnswerCount":0,"Tags":"<python><numpy><tensorflow><keras><python-imaging-library>","CreationDate":"2019-09-13T09:10:08.133","AcceptedAnswerId":null,"OwnerUserId":10078816.0,"Title":"PIL conversion difference between RGB and L modes","Body":"<p>I have several dozen images I'm using to test my deep learning application however when I converted my code to use grayscale for both testing and training I get a problem with the dimensions of the return value for grayscale mode.<\/p>\n\n<p>Here's what I have:<\/p>\n\n<pre><code>def test_batch(self, path):\n    \"\"\"Evaluates a group of images on the network\n\n           Keyword arguments:\n           path -- a directory containing images to be evaluated\n    \"\"\"\n    for folders, subfolders, files in os.walk(path):\n        for file in files:\n            if '.jpg' in file:\n\n                if self.channels == 3:\n                    img = PIL.Image.open(os.path.join(path, file)).convert('RGB')\n                else:\n                    img = PIL.Image.open(os.path.join(path, file)).convert('L')\n                img = img.resize((128, 128), resample=PIL.Image.LANCZOS)\n                print(file)\n                estimations = self.discriminator.predict(\n                    np.expand_dims(np.array(img).astype('float32') \/ 255, axis=0))\n                data = []\n                for i in range(len(estimations[0])):\n                    data.append('{0:2f}'.format(estimations[0][i] * 100))\n<\/code><\/pre>\n\n<p>I'm getting this error: <\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking input: expected input_1 to have 4 dimensions, but got array with shape (1, 128, 128)<\/p>\n<\/blockquote>\n\n<p>Normally when run with 3 channel mode (RGB mode) on my network for both training\/testing the array that is input is <\/p>\n\n<blockquote>\n  <p>(image_count, image_width, image_height, 3)<\/p>\n<\/blockquote>\n\n<p>It's my assumption that the final element is the channels and as such I don't see why converting it to grayscale would eliminate this element, does anyone have any idea as to what might be going on?<\/p>\n\n<p><strong>Edit:<\/strong>\nA quick explanation of what is going on behind the scenes is that I'm training the data with a generator with fit_generator in keras, here is the train generator definition:<\/p>\n\n<pre><code> train_datagen = ImageDataGenerator(\n        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        fill_mode='nearest')\n  train_generator = train_datagen.flow_from_directory(\n        train_path,\n        target_size=(128, 128),\n        shuffle=True,\n        batch_size=batch_size,\n        class_mode='categorical',\n        color_mode='grayscale' if self.channels == 1 else 'rgb'\n    )\n<\/code><\/pre>\n\n<p>As you can see in the above I squash pixel values between 0-1 and apply color_mode grayscale when the channel size is 1.<\/p>\n","answers":[]}
+{"QuestionId":57920748,"AnswerCount":1,"Tags":"<tensorflow><neural-network><classification><conv-neural-network>","CreationDate":"2019-09-13T09:19:48.943","AcceptedAnswerId":null,"OwnerUserId":10653982.0,"Title":"Several FC layers in a row","Body":"<p>I have a question about the role of the Fully Connected layer in the last layers of a CNN. <\/p>\n\n<p>1- Is FC layer acts as a learner classifier? <\/p>\n\n<p>2- Why we first use a linear activation function followed by a non-linear (e.g. softmax)?<\/p>\n\n<p>3- What is the reason for adding several FC layers in a row on top of the network?<\/p>\n\n<pre><code>M_L = KL.Dense(512, activation='relu')(M_L)\nM_L = KL.Dropout(DROPOUT_PROB)(M_L)\nM_L = KL.Dense(256, activation='relu')(M_L)\nM_L = KL.Dropout(DROPOUT_PROB)(M_L)\nM_L = KL.Dense(128, activation='relu')(M_L)\nM_L = KL.Dropout(DROPOUT_PROB)(M_L)\nM_L = KL.Dense(64, activation='relu')(M_L)\nM_L = KL.Dropout(DROPOUT_PROB)(M_L)\nM_L = KL.Dense(1, activation='Sigmoid')(M_L)\n<\/code><\/pre>\n\n<p>4- What would be the difference if we only do like this:<\/p>\n\n<pre><code>M_L = KL.Dense(512, activation='relu')(M_L)\nM_L = KL.Dropout(DROPOUT_PROB)(M_L)\nM_L = KL.Dense(1, activation='Sigmoid')(M_L)\n<\/code><\/pre>\n\n<p>Or even:<\/p>\n\n<pre><code>M_L = KL.Dense(1, activation='Sigmoid')(M_L)\n<\/code><\/pre>\n\n<p>My intuition is that by adding more FC layers we have more trainable parameters. So, it will help to a multi-task network to have some specific parameters for a specific task. Am I right?<\/p>\n\n<p>5- Do we have any other reason for adding several consecutive FC  layers? is decreasing the features smoothly helps for training a classifier?<\/p>\n","answers":[{"AnswerId":"57921124","CreationDate":"2019-09-13T09:42:31.160","ParentId":null,"OwnerUserId":"8198272","Title":null,"Body":"<p>The <a href=\"https:\/\/en.wikipedia.org\/wiki\/Universal_approximation_theorem\" rel=\"nofollow noreferrer\">Universal Approximation Theorem<\/a> states that a neural network needs only a single hidden layer with non-linear activation functions to model any function.  That single layer might need to have an infinite number of filters to model the function perfectly -- but we can get an approximation of arbitrary accuracy by choosing a sufficiently large number of filters.<\/p>\n\n<p>So you are right, in principle, your second architecture will be able to approximate your function.  It won't do it as well as the first, but it will do something.<\/p>\n\n<p>The third architecture is extremely weak, because you don't have a hidden layer.  You only have a single filter with a sigmoid activation function.  Presumably the function you want to model is constrained to the range 0 to 1.  That's why there's a sigmoid output layer in all your architectures.  Presumably, you have many inputs.  All that will happen in your third architecture is that you will take a weighted, linear sum of your inputs, add one scalar weight (the bias), and then take the sigmoid of the result.  That's not very expressive.  You can't get arbitrarily close to an arbitrary function with this architecture.<\/p>\n\n<p>Now, what's special about the first, \"deep\" architecture?  The Universal Approximation Theorem says we only need one hidden layer, and the second architecture has that.  So we could just make that single hidden layer wider, right?  Well, the Universal Approximation Theorem doesn't say that a single hidden layer is the BEST way to model a function.  Frequently, we find that multiple layers with progressively smaller numbers of filters produce better results.  To achieve comparable results with the second architecture as you do with the first, you might need 10,000 filters in your hidden layer.<\/p>\n\n<p>Before the introduction of ReLU, deep architectures trained very slowly or got stuck.  That's not much of an issue now.<\/p>\n"}]}
+{"QuestionId":57920804,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><tensorflow-probability>","CreationDate":"2019-09-13T09:22:52.477","AcceptedAnswerId":null,"OwnerUserId":10385423.0,"Title":"How to enforce a probability distribution on intermediate layer features?","Body":"<p>I have a CNN-LSTM architecture. Now, I want to enforce a probability distribution, say, mixture normal distribution, on the features of an intermediate layer. How can I do this in Keras?<\/p>\n\n<p>I saw MDN implementation for keras (<a href=\"https:\/\/github.com\/cpmpercussion\/keras-mdn-layer\" rel=\"nofollow noreferrer\">https:\/\/github.com\/cpmpercussion\/keras-mdn-layer<\/a>). But this requires that it should be used as last layer of the network. Also, I took a look at tensorflow_probability, but I am not sure if I can use a tensorflow_probability layer for this. <\/p>\n\n<pre><code>enc.add(LSTM(units=64,activation='tanh',return_sequences=False))\n\n# probability layer\nenc.add(tfpl.MixtureNormal(num_classes,[64]))\n\nenc.add(Dense(units=num_classes,activation='softmax'))\n<\/code><\/pre>\n\n<p>When I tried the above code, it gave me an error as follows:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: 2 root error(s) found.\n\n  (0) Invalid argument: Number of ways to split should evenly divide the split dimension, but got split_dim 2 (size = 15) and num_split 2\n     [[{{node mixture_normal\/MixtureSameFamily\/independent_normal\/IndependentNormal\/split}}]]\n     [[metrics\/categorical_accuracy\/Identity\/_251]]\n\n  (1) Invalid argument: Number of ways to split should evenly divide the split dimension, but got split_dim 2 (size = 15) and num_split 2\n     [[{{node mixture_normal\/MixtureSameFamily\/independent_normal\/IndependentNormal\/split}}]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"57931532","CreationDate":"2019-09-14T00:21:22.877","ParentId":null,"OwnerUserId":"9946929","Title":null,"Body":"<p>You might consider using <a href=\"https:\/\/www.tensorflow.org\/probability\/api_docs\/python\/tfp\/layers\/KLDivergenceRegularizer\" rel=\"nofollow noreferrer\"><code>KLDivergenceRegularizer<\/code><\/a> with <code>use_exact=False<\/code>, if you don't need to strictly enforce distribution but rather penalize divergence.<\/p>\n\n<p>Note that most Keras layers accept an <code>activity_regularizer<\/code> arg, e.g.\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/Dense\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/Dense<\/a><\/p>\n"}]}
+{"QuestionId":57921044,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-13T09:37:45.760","AcceptedAnswerId":null,"OwnerUserId":7327257.0,"Title":"Tensorflow: how to keep batch dimension when using tf.where()?","Body":"<p>I'm trying to select elements differents from zero and work with them later. My input tensor has batch dimension, so I want to keep it and don't mix data over batches. I think <code>tf.gather_nd()<\/code> would work for me, but first I have to get the indexes of the desired data and I found <code>tf.where()<\/code>. I have tried the following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>img = tf.constant([[[1., 0., 0.], \n                    [0., 0., 2.],\n                    [0., 3, 0.]], \n                   [[1., 2., 3.], \n                    [0., 0., 1.], \n                    [0., 0., 0.]]], dtype='float32') # shape [2, 3, 3]\n\nindexes = tf.where(tf.not_equal(img, 0.))\n<\/code><\/pre>\n\n<p>I would expect <code>indexes<\/code> to keep batch dimension, however it has shape <code>[7, 2]<\/code>. I suspect the problem comes from having different number of points that satisfies the condition in different batches. <\/p>\n\n<p>Is there a way to get the indexes keeping batch dimension? Thanks in advance.<\/p>\n\n<p><strong>EDIT:<\/strong> <code>indexes<\/code> has shape <code>[7, 3]<\/code> where first dim refers to number of points and the second dim refers to the position of the point (incluiding which batch it belongs to). But I need <code>indexes<\/code> to have the specific batch dimension, because later I want to use it to ghater data from <code>img<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Y = tf.gather_nd(img, indexes)\n<\/code><\/pre>\n\n<p>I want <code>Y<\/code> to have batch dimension, but as <code>indexes<\/code> hasn't, I get a flat tensor with data from different bateches mixed.<\/p>\n","answers":[{"AnswerId":"57921375","CreationDate":"2019-09-13T09:56:03.937","ParentId":null,"OwnerUserId":"8244247","Title":null,"Body":"<p>Actually, you may have done something wrong : when I run your code, <code>indexes<\/code> is of dimension <code>(7,3)<\/code> and not <code>(7,2)<\/code>. The <code>3<\/code> correspond to your 3 dimensions, whereas the <code>7<\/code> corresponds to the number of non-zero elements in <code>img<\/code>.<\/p>\n\n<p>Full result of <code>sess.run(indexes)<\/code> : <\/p>\n\n<pre><code>array([[0, 0, 0],\n      [0, 1, 2],\n      [0, 2, 1],\n      [1, 0, 0],\n      [1, 0, 1],\n      [1, 0, 2],\n      [1, 1, 2]])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57921463,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow2.0>","CreationDate":"2019-09-13T10:00:20.407","AcceptedAnswerId":57939445.0,"OwnerUserId":8270077.0,"Title":"How tf.data.experimental.group_by_window() operates in Tensorflow 2.0","Body":"<p>I am trying to understand the tf.data.experimental.group_by_window() method in Tensorflow 2 but I have some difficulties.<\/p>\n\n<p>For a reproducible example I use the one presented in the documentation:<\/p>\n\n<pre><code>components = np.arange(100).astype(np.int64)\ndataset20 = tf.data.Dataset.from_tensor_slices(components)\ndataset20 = dataset.apply(tf.data.experimental.group_by_window(key_func=lambda x: x%2, reduce_func=lambda _,\\\n                                                          els: els.batch(10), window_size=100))\n\ni = 0\n\nfor elem in dataset20:\n\n    print('i is {0}\\n'.format(i))\n\n    print('elem is {0}'.format(elem.numpy()))\n\n    i += 1\n\n    print('\\n--------------------------------\\n')\n\ni is 0\n\nelem is [0 2 4 6 8]\n\n--------------------------------\n\ni is 1\n\nelem is [1 3 5 7 9]\n\n--------------------------------\n<\/code><\/pre>\n","answers":[{"AnswerId":"57939445","CreationDate":"2019-09-14T21:37:15.793","ParentId":null,"OwnerUserId":"5587428","Title":null,"Body":"<p>Part of the confusion may be that the output doesn't correspond to the example code. The actual output from this:<\/p>\n\n<pre><code>components = np.arange(100).astype(np.int64)\ndataset20 = tf.data.Dataset.from_tensor_slices(components)\ndataset20 = dataset20.apply(tf.data.experimental.group_by_window(key_func=lambda x: x%2, reduce_func=lambda _,els: els.batch(10), window_size=100))\nfor i, d in enumerate(dataset20): \n    print(i, d.numpy())\n<\/code><\/pre>\n\n<p>is<\/p>\n\n<pre><code>0 [ 0  2  4  6  8 10 12 14 16 18]\n1 [20 22 24 26 28 30 32 34 36 38]\n2 [40 42 44 46 48 50 52 54 56 58]\n3 [60 62 64 66 68 70 72 74 76 78]\n4 [80 82 84 86 88 90 92 94 96 98]\n5 [ 1  3  5  7  9 11 13 15 17 19]\n6 [21 23 25 27 29 31 33 35 37 39]\n7 [41 43 45 47 49 51 53 55 57 59]\n8 [61 63 65 67 69 71 73 75 77 79]\n9 [81 83 85 87 89 91 93 95 97 99]\n<\/code><\/pre>\n\n<p>As described in the documentation <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/data\/experimental\/group_by_window\" rel=\"nofollow noreferrer\">here<\/a>, the <code>key func<\/code> separates the data into groups with associated key values. In the example the <code>key func<\/code> separates the data [0, 99] into even and odd groups. The <code>reduce_func<\/code> then operates on the key, group pairs to produce another dataset. Note though that <code>reduce_func<\/code> only operates on groups of data no greater than <code>window_size<\/code>. In the example, the window size is greater than the two group sizes (100 vs 50 elements), so has no effect and all evens are given in batches of 10 followed by all odds. If <code>window size<\/code> is changed to a value less than 50 then it does have an effect. For example, if the window size is changed to 5 and also the batching is moved to outside the <code>group_by_window<\/code> function:<\/p>\n\n<pre><code>dataset20 = dataset20.apply(tf.data.experimental.group_by_window(key_func=lambda x: x%2, reduce_func=lambda _, els: els, window_size=5)).batch(10)\n<\/code><\/pre>\n\n<p>then the following output is produced:<\/p>\n\n<pre><code>0 [0 2 4 6 8 1 3 5 7 9]\n1 [10 12 14 16 18 11 13 15 17 19]\n2 [20 22 24 26 28 21 23 25 27 29]\n3 [30 32 34 36 38 31 33 35 37 39]\n4 [40 42 44 46 48 41 43 45 47 49]\n5 [50 52 54 56 58 51 53 55 57 59]\n6 [60 62 64 66 68 61 63 65 67 69]\n7 [70 72 74 76 78 71 73 75 77 79]\n8 [80 82 84 86 88 81 83 85 87 89]\n9 [90 92 94 96 98 91 93 95 97 99]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57921773,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><neural-network><recommendation-engine>","CreationDate":"2019-09-13T10:20:34.353","AcceptedAnswerId":null,"OwnerUserId":2739654.0,"Title":"How to implement the cross-product transformation in Google's wide-and-deep model","Body":"<p>Having read the paper on wide and deep models for recommender systems, I'm interested in trialing something similar. I'm struggling however to conceptualize how the cross-product transformation they used in the paper was done. The tutorial for the DNNLinearCombinedClassifier demonstrates the use of tensorflow.feature_columns.crossed_columns to perform that transformation, however that implementation only makes sense in a one-to-one relationship between the two columns to be crossed (I.E. when creating the crossed column \"workclass_x_education\",  for each sample in the dataset there is only one value for workclass and one value for education). In Google's paper they say that:<\/p>\n\n<blockquote>\n  <p>The wide component consists of  the  cross-product transformation  of  user  installed  apps and impression apps<\/p>\n<\/blockquote>\n\n<p>and additionally that:<\/p>\n\n<blockquote>\n  <p>Each example corresponds to one impression.  The label is app acquisition: 1 if the impressed app was installed, and 0 otherwise.<\/p>\n<\/blockquote>\n\n<p>From that I gather that each sample is a single impression and that there's a feature for \"impressed app\" which details the app that this impression was marketing. In that context, there can't then be a single feature for \"installed app\", because a user might have hundreds of installed apps at the time of the impression. This is confusing, because the figures in the paper do imply that it is a single feature:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9n8l4.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9n8l4.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Should I interpret that \"User Installed App\" Categorical Feature not as a single feature, but actually as their dataset having a feature-per-app with a binary flag denoting whether the user has that app installed or not, and that to replicate this functionality in tensorflow I'd need to do something like...<\/p>\n\n<pre><code>\napp_list = ['app1', 'app2', 'app3', 'app4', ...]\ninstalled_app_features = ['app1_installed', 'app2_installed', 'app3_installed', 'app4_installed', ...]\n\nimpressed_app = tf.feature_column.categorical_column_with_vocabulary_list(key='impressed_app', vocabulary_list=app_list)\n\ninstalled_apps = []\n\ncrossed_features = []\n\nfor app in installed_app_features:\n    feature = tf.feature_column.categorical_column_with_vocabulary_list(key='app', vocabulary_list=[0,1])\n    installed_apps.append(feature)\n    crossed_features.append(\n        tf.feature_column.crossed_column(\n            [impressed_app, feature],\n            100\n        )    \n    )\n<\/code><\/pre>\n\n<p>...or am I just completely missing something and need to think about this in a different way?<\/p>\n\n<p>Many thanks in advance for any help offered.<\/p>\n\n<p>Dan<\/p>\n","answers":[]}
+{"QuestionId":57922111,"AnswerCount":0,"Tags":"<python><tensorflow><keras><gradient><loss-function>","CreationDate":"2019-09-13T10:42:08.553","AcceptedAnswerId":null,"OwnerUserId":5741822.0,"Title":"How to make a custom loss function+Gradient with Python in Tensorflow?","Body":"<p>This is generally about using tf.numpy_function in writing custom loss function. Where gradient of custom loss function must be provided manually as well.<\/p>\n\n<p>Assume that a custom loss function with extra input arguments is written.<\/p>\n\n<pre><code>def custom_loss(self, targetPosition):\n    def loss(y_true, y_pred):\n        loss_ = tf.numpy_function(self.forAllWrapped, [y_pred, targetPosition], tf.float32)\n        return loss_;\n    return loss\n<\/code><\/pre>\n\n<p>where loss function is written in a tf.numpy_function wrapper which gets the tensor values as Numpy arrays and sends it through GRPC to some C# code to calculate the loss. <\/p>\n\n<pre><code>def forAllWrapped(self, y_pred, targetPosition):\n    result= np.empty((0,targetPosition.shape[1]), np.float)\n    i=0\n    numFKs = y_pred.shape[0]\n    while i &lt; numFKs :\n        res_ =sres_=GET-INDIVIDUAL-LOSS-FROM-GRPC-METHOD(y_pred[i,],targetPosition[i,])\n        result = np.append(result, self.trimResultFK(res_) , axis=0)\n        i=i+1\n    loss__ = K.mean(K.square(targetPosition - result), axis=1)\n    return loss__.astype(tf.float32)\n<\/code><\/pre>\n\n<p>The loss calculation part works perfectly. But when I am gonna use my custom loss function, I am getting an error:<\/p>\n\n<blockquote>\n  <p>Variable xxxxxx  has <code>None<\/code> 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.<\/p>\n<\/blockquote>\n\n<p>Due to usage of non-Tensorflow functions to calculate the loss value, then gradient is unknown to the optimizer and it cannot automatically derive it using <a href=\"https:\/\/www.tensorflow.org\/tutorials\/eager\/automatic_differentiation\" rel=\"nofollow noreferrer\">Automatic differentiation and gradient<\/a>.\nTo solve this problem, I tried to implement custom gradient and attach it to my custom loss function based on <a href=\"https:\/\/stackoverflow.com\/a\/39921608\">this nice answer<\/a>. \nHere is the corresponding parts which attach the gradient using my version of self.numpy_function method:<\/p>\n\n<pre><code>def numpy_function(self, func, inp, dtype=tf.float32, grad=None):\n    # Need to generate a unique name to avoid duplicates:\n    rnd_name = 'PyFuncGrad' + str(np.random.randint(0, 1E+8))\n    tf.RegisterGradient(rnd_name)(grad)\n    g = tf.get_default_graph()\n    with g.gradient_override_map({\"PyFunc\": rnd_name}):\n        return tf.numpy_function(func, inp, dtype)\n<\/code><\/pre>\n\n<p>I have now 3 methods to override the gradient. First gradient is calculated in numpy domain (dummy outputs 1 and 0 for y_pred, targetPosition with the same shape as their numpy array shape):<\/p>\n\n<pre><code>def np_d_customGrad(self, y_pred, targetPosition):\n    return np.ones((self.batch_size,2)).astype(np.float32),np.zeros((self.batch_size ,3 )).astype(np.float32)\n<\/code><\/pre>\n\n<p>Then it is wrapped for Tensorflow:<\/p>\n\n<pre><code>def tf_customGrad(self, y_pred, targetPosition):\n    d_y_pred_, d_targetPosition_ = tf.numpy_function(self.np_d_customGrad,[y_pred, targetPosition],[tf.float32,tf.float32])\n    return d_y_pred_, d_targetPosition_\n<\/code><\/pre>\n\n<p>and finally used to calculate gradient for two inputs of our custom loss function (y_pred,and targetPosition):<\/p>\n\n<pre><code>def customGrad(self, op, grad):\n    y_pred = op.inputs[0]\n    targetPosition = op.inputs[1]\n    d_y_pred, d_targetPosition  = self.tf_customGrad(y_pred,targetPosition)\n    return grad * d_y_pred , grad * d_targetPosition\n<\/code><\/pre>\n\n<p>This results in dimension error:<\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors_impl.InvalidArgumentError: Incompatible shapes: [1000,2] vs. [1000]<\/p>\n<\/blockquote>\n\n<p>I have tried to return different shapes of zeros and ones in <strong>np_d_customGrad<\/strong> function. but it does not work. I assume that my function must return two gradient one per input parameters of custom loss function. (Please correct me if I am wrong)<\/p>\n\n<p>I have also tried to return the numpy array <strong>result<\/strong> instead of <strong>loss__<\/strong> values and use<\/p>\n\n<pre><code>return K.mean(K.square(targetPosition - actualPosition__), axis=1)\n<\/code><\/pre>\n\n<p>in <em>custom_loss<\/em> hoping that Tensorflow derives the gradient itself, but it could not.<\/p>\n\n<p>Please note that I cannot implement the <strong>GET-INDIVIDUAL-LOSS-FROM-GRPC-METHOD<\/strong> method in python using TF as it takes months of work due to complexity.<\/p>\n","answers":[]}
+{"QuestionId":57922433,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2019-09-13T11:03:50.947","AcceptedAnswerId":null,"OwnerUserId":7955314.0,"Title":"Multi class classification Metric Finding argmax along each row of a 2d tensor, and change them to 1 in keras backed K for TP,FP,FN calculation","Body":"<p>I am having a multi class classification problem. My keras model giving continuous vectors at the end of prediction. <\/p>\n\n<p>Here is the code i have in numpy based that i use for evaluation from model.predict.\nSo, I am trying for TP, FP, FN calculation in keras metric itself which i can get from <code>keras.evaluate<\/code>.<\/p>\n\n<p>Numpy Snippet Used:<\/p>\n\n<pre><code>#   y_true is one hot of classes.\n\n# Continuos values\ny_pred_continuos = model.predict(xfeature)    \n# Discrete Values\ny_pred = np.zeros_like(y_pred_continuos)\ny_pred[np.arange(len(y_pred_continuos)), y_pred_continuos.argmax(1)] = 1   \n\ncur_TP = np.multiply(y_true,y_pred)\ncur_FP = y_pred - cur_TP\ncur_FN = y_true - cur_TP\ncur_TP = cur_TP.sum(axis=0)\ncur_FP = cur_FP.sum(axis=0)\ncur_FN = cur_FN.sum(axis=0)\n<\/code><\/pre>\n\n<p>y_true, and y_pred will be avialble in keras <a href=\"https:\/\/keras.io\/metrics\/\" rel=\"nofollow noreferrer\">Custom Metrics<\/a>.<br>\nIn Keras, I have tried like:<\/p>\n\n<pre><code>\nfrom keras import backend as K\nimport numpy as np\ncy_true = np.array([[0,0,1,0],[0,0,0,1]])\ncy_pred = np.random.normal(size=8).reshape(2,4)\n\ny_true = K.variable(cy_true,)\ny_pred = K.variable(cy_pred)\ny_pred_discrete = K.zeros_like(y_pred)\n\n<\/code><\/pre>\n\n<p>I am unaware on how to loop over and find one hot of continuos vector.<\/p>\n\n<p>Also Using scikit-learn's classification report, I can get the same. To do that, I have to run it out of Keras Metrics by saving predictions and then analysis. <\/p>\n\n<p>But I want them as keras metric to find TP,FP,FN,TN for multi class classfication . <\/p>\n\n<p>I have written them for multi label classification where I have just sorting of predictions, followed by selecting top k.<\/p>\n\n<p>But in case of multi class, I dont know how to loop over a keras tensor. <\/p>\n\n<p>If I find them in Keras Metric itself, then I can write a callback in keras using BaseAccumulator to cumulate the values over batches and at the epoch end, i can get metric F1 too,  Also I cann use to monitor my training process.<\/p>\n","answers":[]}
+{"QuestionId":57923280,"AnswerCount":0,"Tags":"<pytorch><hyperopt><python-sacred>","CreationDate":"2019-09-13T12:01:04.280","AcceptedAnswerId":null,"OwnerUserId":10222761.0,"Title":"How to merge Hyperopt and Sacred together?","Body":"<p>I am trying to use omniboard to keep a track of all the experiments hyperopt library is doing. The link below is from sacred library documentation and I am not able to use it to merge hyperopt and sacred together. Right now when I am defining an objective function and putting it in fmin(), I am getting a TypeError saying 'NoneType object is not callable'. Can somebody please help me figure out a way around this.<\/p>\n\n<p><a href=\"https:\/\/github.com\/yuvalatzmon\/SACRED_HyperOpt_v2\" rel=\"nofollow noreferrer\">https:\/\/github.com\/yuvalatzmon\/SACRED_HyperOpt_v2<\/a><\/p>\n","answers":[]}
+{"QuestionId":57923352,"AnswerCount":0,"Tags":"<python><tensorflow><python-3.6><gradient><loss>","CreationDate":"2019-09-13T12:06:24.993","AcceptedAnswerId":null,"OwnerUserId":9652636.0,"Title":"TypeError: Tensor objects are only iterable when eager execution is enabled. To iterate over this tensor use tf.map_fn","Body":"<p>I want to run this <a href=\"https:\/\/github.com\/kabkabm\/defensegan\/blob\/master\/blackbox.py\" rel=\"nofollow noreferrer\">code<\/a>.\nThe part of code that has error is:<\/p>\n\n<pre><code>   if defense_type:\n         if 'gan' in defense_type:\n              # Load cached dataset reconstructions.\n              ...\n\n         if FLAGS.debug:\n             train_images_bb = train_images_bb[:20 * batch_size]\n             train_labels_bb = train_labels_bb[:20 * batch_size]\n\n         #Prepare the black_box model.\n         prep_bbox_out = prep_bbox(\n         sess, images_tensor, labels_tensor, train_images_bb,\n         train_labels_bb, test_images_bb, test_labels_bb, nb_epochs,\n         batch_size, learning_rate, rng=rng, gan=cur_gan,\n         adv_training=adv_training,\n         cnn_arch=bb_model)\n   else:\n         prep_bbox_out = prep_bbox(sess, images_tensor, labels_tensor,\n                              train_images_bb, train_labels_bb,\n                              test_images_bb, test_labels_bb,\n                              nb_epochs, batch_size, learning_rate,\n                              rng=rng, gan=cur_gan,\n                              adv_training=adv_training,\n                              cnn_arch=bb_model)\n\n   model, bbox_preds, accuracies['bbox'] = prep_bbox_out\n\n   # Train substitute method\n   print(\"Training the substitute model.\")\n   reconstructed_tensors = tf.stop_gradient(\n           gan.reconstruct(images_tensor, batch_size=batch_size,\n                    reconstructor_id=1))\n   model_sub, preds_sub = train_sub(\n        sess, images_tensor, labels_tensor,\n        model(reconstructed_tensors), images_sub,\n        labels_sub,\n        nb_classes, nb_epochs_s, batch_size,\n        learning_rate, data_aug, lmbda, rng=rng,\n        substitute_model=sub_model,\n      )\n<\/code><\/pre>\n\n<p>But I'm getting this error: <\/p>\n\n<pre><code>      File \"blackbox.py\", line 770, in &lt;lambda&gt;\n          main_cfg = lambda x: main(cfg, x)\n      File \"blackbox.py\", line 686, in main\n         defense_type=FLAGS.defense_type)\n      File \"blackbox.py\", line 512, in blackbox\n         reconstructor_id=1)))\n      File \"\/usr\/local\/lib\/python3.6\/dist- \n         packages\/tensorflow\/python\/framework\/ops.py\", line 477, in \n        __iter__\n     \"Tensor objects are only iterable when eager execution is \"\n TypeError: Tensor objects are only iterable when eager execution is \n  enabled. To iterate over this tensor use tf.map_fn.  \n<\/code><\/pre>\n\n<p>The original code in <a href=\"https:\/\/github.com\/kabkabm\/defensegan\/blob\/master\/blackbox.py\" rel=\"nofollow noreferrer\">here<\/a>, implemented with python 2.7, but I turned it into python 3.6. \nI use python 3.6 (google colab)\nTensorslow 1.14.0 , I also checked tf 1.12.0 and tf 1.8.0.<\/p>\n\n<p>Kindly, help me get through this issue? How I can rerwite the code to do iterate the tensor?<\/p>\n","answers":[]}
+{"QuestionId":57924014,"AnswerCount":0,"Tags":"<tensorflow><keras><google-colaboratory><tensorboard>","CreationDate":"2019-09-13T12:53:00.377","AcceptedAnswerId":null,"OwnerUserId":519882.0,"Title":"tensor flow core estimator not found while running tensor board in goolge colab","Body":"<p>I am trying to learn tensor board on google colab at below link<\/p>\n\n<p><a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/tensorboard\/blob\/master\/docs\/r2\/get_started.ipynb\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/github\/tensorflow\/tensorboard\/blob\/master\/docs\/r2\/get_started.ipynb<\/a><\/p>\n\n<p>While running I am getting below error when i run below command<\/p>\n\n<pre><code>%tensorboard --logdir logs\/fit\n<\/code><\/pre>\n\n<p>Error is <\/p>\n\n<pre><code>contents of stderr:\nTraceback (most recent call last):\n  File \"\/usr\/local\/bin\/tensorboard\", line 6, in &lt;module&gt;\n    from tensorboard.main import run_main\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorboard\/main.py\", line 40, in &lt;module&gt;\n    from tensorboard import default\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorboard\/default.py\", line 39, in &lt;module&gt;\n    from tensorboard.plugins.beholder import beholder_plugin_loader\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorboard\/plugins\/beholder\/__init__.py\", line 22, in &lt;module&gt;\n    from tensorboard.plugins.beholder.beholder import Beholder\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorboard\/plugins\/beholder\/beholder.py\", line 199, in &lt;module&gt;\n    class BeholderHook(tf.estimator.SessionRunHook):\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/__init__.py\", line 50, in __getattr__\n    module = self._load()\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/__init__.py\", line 44, in _load\n    module = _importlib.import_module(self.__name__)\n  File \"\/usr\/lib\/python3.6\/importlib\/__init__.py\", line 126, in import_module\n    return _bootstrap._gcd_import(name[level:], package, level)\nModuleNotFoundError: No module named 'tensorflow_core.estimator'\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57924284,"AnswerCount":1,"Tags":"<python><tensorflow><anaconda>","CreationDate":"2019-09-13T13:11:01.503","AcceptedAnswerId":null,"OwnerUserId":12059180.0,"Title":"How to run anaconda?","Body":"<p>I just started using python, and I want to have some fun with TensorFlow. But I have stuck in the installation of Anaconda. When I installed it this is the error what I'm getting:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/D8Ixa.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/D8Ixa.png\" alt=\"When I installed it this is the error what I&#39;m getting\"><\/a><\/p>\n\n<p>What should I do? How should I add TensorFlow to my project. I'm an absolute beginner.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/dMxXH.png\" rel=\"nofollow noreferrer\">These ate my variables, sorry for the Hungarian language<\/a><\/p>\n","answers":[{"AnswerId":"57924609","CreationDate":"2019-09-13T13:32:16.210","ParentId":null,"OwnerUserId":"12024442","Title":null,"Body":"<p>To troubleshoot this issue :<\/p>\n\n<ol>\n<li><p>Access the Environmental Variables dialog:<\/p>\n\n<p>a. Right-click the Windows Start button and then select System.\nb. Click the 'Advanced system settings' link.\nc. Click Environment Variables.<\/p><\/li>\n<li><p>Under user and system variables, look for the QT_PLUGIN_PATH variable. If this variable exists, take note of the current value in case you need to reverse the change. Then select the variable and click 'Delete'.<\/p><\/li>\n<li><p>Close all remaining windows by clicking OK.<\/p><\/li>\n<li><p>Reinstall Anaconda<\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":57924724,"AnswerCount":2,"Tags":"<pytorch><dataloader>","CreationDate":"2019-09-13T13:39:22.797","AcceptedAnswerId":null,"OwnerUserId":6780589.0,"Title":"DataLoader messing up transformed data","Body":"<p>I am testing the MNIST dataset in Pytorch, and after I apply a transformation to the X data, it seems the DataLoader puts all values out of the original order, potentially messing up the training step.<\/p>\n\n<p>My transformation is to divide all values by 255. One should notice that the transformation itself does not change positions, as shown by the first scatterplots. But after the data is passed to the DataLoader and I retrieve it back, they are out of order. If I make no transformation, everything is fine (not shown). The distribution of the values is the same among <em>before<\/em>, <em>after1<\/em> (divided by 255\/before DataLoader) and <em>after2<\/em> (divided by 255\/after DataLoader) (also not shown), only the order seems to be affected.<\/p>\n\n<pre><code>import torch\nfrom torchvision import datasets\nimport torchvision.transforms as transforms\nimport matplotlib.pyplot as plt\n\ntransform = transforms.ToTensor()\n\ntrain = datasets.MNIST(root = '.', train = True, download = True, transform = transform)\ntest = datasets.MNIST(root = '.', train = False, download = True, transform = transform)\n\nbefore = train.data[0]\n\ntrain.data = train.data.float()\/255\nafter1 = train.data[0]\n\ntrain_loader = torch.utils.data.DataLoader(train, batch_size = 128)\ntest_loader = torch.utils.data.DataLoader(test, batch_size = 128)\n\nfig, ax = plt.subplots(1, 2)\nax[0].scatter(range(len(before.view(-1))), before.view(-1))\nax[0].set_title('Before')\nax[1].scatter(range(len(after1.view(-1))), after1.view(-1))\nax[1].set_title('After1')\n\nafter2 = next(iter(train_loader))[0][0]\n\nfig, ax = plt.subplots(1, 2)\nax[0].scatter(range(len(before.view(-1))), before.view(-1))\nax[0].set_title('Before')\nax[1].scatter(range(len(after2.view(-1))), after2.view(-1))\nax[1].set_title('After2')\n\nfig, ax = plt.subplots(1, 3)\nax[0].imshow(before, cmap = 'gray')\nax[1].imshow(after1, cmap = 'gray')\nax[2].imshow(after2.view(28, 28), cmap = 'gray')\n<\/code><\/pre>\n\n<p>I know that this might not be the best way to deal with this data (<code>transforms.Normalize<\/code> should solve it), but I would really like to understand what is happening.<\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"57939180","CreationDate":"2019-09-14T20:52:55.667","ParentId":null,"OwnerUserId":"6780589","Title":null,"Body":"<p>So... I <a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/26235\" rel=\"nofollow noreferrer\">posted this same question at Pytorch's GitHub page<\/a>, and they answered the following:<\/p>\n\n<blockquote>\n  <p>It's unrelated to data loader. You are messing with an attribute of\n  the particular dataset object, however, the actual <code>__getitem__<\/code> of that\n  object does much more:\n  <a href=\"https:\/\/github.com\/pytorch\/vision\/blob\/6de158c473b83cf43344a0651d7c01128c7850e6\/torchvision\/datasets\/mnist.py#L92\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/vision\/blob\/6de158c473b83cf43344a0651d7c01128c7850e6\/torchvision\/datasets\/mnist.py#L92<\/a><\/p>\n  \n  <p>In particular this line (<code>mode='L'<\/code>) assumes <code>uint8<\/code> input. Since you\n  replaced it with float, it is wrong.<\/p>\n<\/blockquote>\n\n<p>Then I guess the preferred approach would be to apply a transform when preparing the dataset at the very beginning of my code.<\/p>\n"},{"AnswerId":"57926266","CreationDate":"2019-09-13T15:16:05.977","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>Originally I haven't tested the code you wrote.\nRewrote the original:<\/p>\n\n<pre><code>import torch\nfrom torchvision import datasets\nimport torchvision.transforms as transforms\nfrom torch.utils.data import DataLoader, Dataset, TensorDataset\nimport matplotlib.pyplot as plt\n\ntransform = transforms.ToTensor()\n\ntrain = datasets.MNIST(root = '.', train = True, download = True, transform = transform)\ntest = datasets.MNIST(root = '.', train = False, download = True, transform = transform)\n\ndl = DataLoader(train)\n\nimages = dl.dataset.data.float()\/255\nlabels = dl.dataset.targets\n\ntrain_ds = TensorDataset(images, labels)\ntrain_loader = DataLoader(train_ds, batch_size=128)\n# img, target = next(iter(train_loader))\n\nbefore = train.data[0]\ntrain.data = train.data.float()\/255\nafter1 = train.data[0]\n\n# train_loader = torch.utils.data.DataLoader(train, batch_size = 128)\ntest_loader = torch.utils.data.DataLoader(test, batch_size = 128)\n\nfig, ax = plt.subplots(1, 2)\nax[0].scatter(range(len(before.view(-1))), before.view(-1))\nax[0].set_title('Before')\nax[1].scatter(range(len(after1.view(-1))), after1.view(-1))\nax[1].set_title('After1')\n\nafter2 = next(iter(train_loader))[0][0]\n\nfig, ax = plt.subplots(1, 2)\nax[0].scatter(range(len(before.view(-1))), before.view(-1))\nax[0].set_title('Before')\nax[1].scatter(range(len(after2.view(-1))), after2.view(-1))\nax[1].set_title('After2')\n\nfig, ax = plt.subplots(1, 3)\nax[0].imshow(before, cmap = 'gray')\nax[1].imshow(after1, cmap = 'gray')\nax[2].imshow(after2.view(28, 28), cmap = 'gray')\n<\/code><\/pre>\n"}]}
+{"QuestionId":57924759,"AnswerCount":0,"Tags":"<python><tensorflow><dataset><object-detection-api><tfrecord>","CreationDate":"2019-09-13T13:41:30.710","AcceptedAnswerId":null,"OwnerUserId":8432716.0,"Title":"Is it possible to use tif files for tfrecord? If so, how?","Body":"<p>I want to try out a new dataset I have built. The image files are all 16-bit files in tif format. For a quick check I want to use the Tensorflow Object Detection API. For this I need to produce tfrecord files. Is it possible with tif files (already did my research, didn't found an easy way yet, nor any way to be honest). I don't want to convert the files in png\/jpg, because I don't want to lose any information within the files. Any help?<\/p>\n","answers":[]}
+{"QuestionId":57925061,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-13T13:59:35.943","AcceptedAnswerId":57933374.0,"OwnerUserId":1519955.0,"Title":"How can I reduce the number of CPUs used by Tensorlfow\/Keras?","Body":"<p>I am using the Keras api of Tensorflow 2.0.<\/p>\n\n<p>When calling <code>fit<\/code> on my Keras model, it uses all availabel CPUs.<\/p>\n\n<p>I would like to limit the number of used CPUs. However the way it used to work in former versions of Tensorflow cannot be used anymore:<\/p>\n\n<pre><code>tf.keras.backend.set_session(tf.compat.v1.Session(config=tf.compat.v1.ConfigProto(\n       intra_op_parallelism_threads=2, inter_op_parallelism_threads=2)))\n<\/code><\/pre>\n\n<blockquote>\n  <p>AttributeError: module 'tensorflow.python.keras.api._v2.keras.backend' has no attribute 'set_session'<\/p>\n<\/blockquote>\n\n<p>How could I do that?<\/p>\n","answers":[{"AnswerId":"57926618","CreationDate":"2019-09-13T15:41:14.617","ParentId":null,"OwnerUserId":"8234169","Title":null,"Body":"<p>The second answer in <a href=\"https:\/\/stackoverflow.com\/questions\/40690598\/can-keras-with-tensorflow-backend-be-forced-to-use-cpu-or-gpu-at-will\">this post<\/a> could be a solution to limit the number of CPUs used. You may change the code as <\/p>\n\n<pre><code>import tensorflow as tf\nfrom keras import backend as K\n\nnum_cores = 4\n\nnum_CPU = 1\nnum_GPU = 0\n\nconfig = tf.ConfigProto(intra_op_parallelism_threads=num_cores,\n                        inter_op_parallelism_threads=num_cores, \n                        allow_soft_placement=True,\n                        device_count = {'CPU' : num_CPU,\n                                        'GPU' : num_GPU}\n                       )\n\nsession = tf.Session(config=config)\nK.set_session(session)\n<\/code><\/pre>\n"},{"AnswerId":"57933374","CreationDate":"2019-09-14T07:24:10.590","ParentId":null,"OwnerUserId":"5597718","Title":null,"Body":"<p>In Tensorflow 2.0, there is no session anymore. In eager execution, directly use the config API to set the parallelism at the start of the program like this. <\/p>\n\n<pre><code>import tensorflow as tf\n\ntf.config.threading.set_intra_op_parallelism_threads(2)\ntf.config.threading.set_inter_op_parallelism_threads(2)\nwith tf.device('\/CPU:0'):\n    model = tf.keras.models.Sequential([...\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/config\/threading\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/config\/threading<\/a><\/p>\n"}]}
+{"QuestionId":57925256,"AnswerCount":1,"Tags":"<pytorch><semantic-analysis><allennlp>","CreationDate":"2019-09-13T14:10:26.697","AcceptedAnswerId":57925788.0,"OwnerUserId":7263895.0,"Title":"Replicating Semantic Analysis Model in Demo","Body":"<p>Good day, I am a student that is interested in NLP. I have come across the demo on  <a href=\"https:\/\/demo.allennlp.org\/sentiment-analysis\" rel=\"nofollow noreferrer\">AllenNLP's homepage<\/a>, which stated that:<\/p>\n\n<blockquote>\n  <p>The model is a simple LSTM using GloVe embeddings that is trained on the binary classification setting of the Stanford Sentiment Treebank. It achieves about 87% accuracy on the test set.<\/p>\n<\/blockquote>\n\n<p>Is there any reference to the sample code or any tutorial that I can follow to replicate this result, so that I can learn more about this subject? I am trying to obtain a Regression Output (Instead of classification). <\/p>\n\n<p>I hope that someone can point me in the right direction.. Any help is much appreciated. Thank you!<\/p>\n","answers":[{"AnswerId":"57925788","CreationDate":"2019-09-13T14:44:21.097","ParentId":null,"OwnerUserId":"3288004","Title":null,"Body":"<p>AllenAI provides all code for examples and lib opensource on Git, including AllenNLP.<\/p>\n\n<p>I found exactly how the example was run here: <a href=\"https:\/\/github.com\/allenai\/allennlp\/blob\/master\/allennlp\/tests\/data\/dataset_readers\/stanford_sentiment_tree_bank_test.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/allenai\/allennlp\/blob\/master\/allennlp\/tests\/data\/dataset_readers\/stanford_sentiment_tree_bank_test.py<\/a><\/p>\n\n<p>However, to make it a Regression task, you'll have to tweak directly on Pytorch, which is the underlying technology for AllenNLP.<\/p>\n"}]}
+{"QuestionId":57925467,"AnswerCount":0,"Tags":"<tensorflow><keras><google-colaboratory><tpu>","CreationDate":"2019-09-13T14:23:43.333","AcceptedAnswerId":null,"OwnerUserId":2491350.0,"Title":"AttributeError: KerasTPUModel' object has no attribute _ckpt_saved_epoch","Body":"<p>I am trying to train a model on Google Colab, in order to play around with training on TPU. However, I am running into the following error:<\/p>\n\n<pre><code>AttributeError                            Traceback (most recent call last)\n\n&lt;ipython-input-82-e74efc36d872&gt; in &lt;module&gt;()\n----&gt; 1 tpu_model.fit_generator(training_set, steps_per_epoch = 8000, epochs = 25)\n\n2 frames\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/callbacks.py in configure_callbacks(callbacks, model, do_validation, batch_size, epochs, steps_per_epoch, samples, verbose, count_mode, mode)\n    118   callback_list.model.stop_training = False\n    119   # pylint: disable=protected-access\n--&gt; 120   if callback_list.model._ckpt_saved_epoch is not None:\n    121     # The attribute `_ckpt_saved_epoch` is supposed to be None at the start of\n    122     # training (it should be made None at the end of successful multi-worker\n\nAttributeError: 'KerasTPUModel' object has no attribute '_ckpt_saved_epoch'\n<\/code><\/pre>\n\n<p>While trying to run the following code. <\/p>\n\n<pre><code>    import tensorflow as tf\n    from tensorflow.keras import layers\n    from tensorflow.keras.preprocessing.image import ImageDataGenerator\n    import os\n    import zipfile\n    print(tf.VERSION)\n\n    local_zip = '\/home\/cats_and_dogs_filtered.zip'\n    zip_ref = zipfile.ZipFile(local_zip, 'r')\n    zip_ref.extractall('\/home')\n    zip_ref.close()\n\n    def create_model():\n        classifier = tf.keras.models.Sequential()\n\n        classifier.add(layers.Conv2D(32, (3, 3), input_shape=(64, 64, 3), activation='relu'))\n        classifier.add(layers.MaxPooling2D(pool_size=(2, 2)))\n\n        classifier.add(layers.Conv2D(32, (3, 3), activation= 'relu'))\n        classifier.add(layers.MaxPooling2D(pool_size=(2, 2)))\n\n        classifier.add(layers.Flatten())\n\n        classifier.add(layers.Dense(units=128, activation= 'relu'))\n        classifier.add(layers.Dense(units=1, activation= 'sigmoid'))\n\n        return classifier\n\n    train_datagen = ImageDataGenerator(rescale = 1.\/255, shear_range = 0.2, zoom_range = 0.2, horizontal_flip = True)\n    training_set = train_datagen.flow_from_directory('\/home\/cats_and_dogs_filtered\/train', target_size = (64, 64), batch_size = 32, class_mode = 'binary')\n\n    model = create_model()\n\n    TPU_WORKER = 'grpc:\/\/' + os.environ['COLAB_TPU_ADDR']\n    tpu_model = tf.contrib.tpu.keras_to_tpu_model(    model,    strategy=tf.contrib.tpu.TPUDistributionStrategy(tf.contrib.cluster_resolver.TPUClusterResolver(TPU_WORKER)))\n    tpu_model.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])\n    tpu_model.save_weights('.\/tpu_model.h5', overwrite=True)      \n\n\n    tpu_model.fit_generator(training_set, steps_per_epoch = 8000, epochs = 25)\n<\/code><\/pre>\n\n<p>I am not sure what is going on. I used similar code to train it on CPU (takes a long time to train).<\/p>\n","answers":[]}
+{"QuestionId":57925817,"AnswerCount":1,"Tags":"<tensorflow><tensorflow2.0>","CreationDate":"2019-09-13T14:46:39.780","AcceptedAnswerId":57927515.0,"OwnerUserId":3115923.0,"Title":"TensorFlow 2.0 scatter add","Body":"<p>I would like to implement the following design in TensorFlow 2.0.<\/p>\n\n<p>Given a <code>memory<\/code> tensor of shape <code>[a, b, c]<\/code>,<br>\nan <code>indices<\/code> tensor of shape <code>[a, 1]<\/code>,<br>\nand an <code>updates<\/code> tensor of shape <code>[a, c]<\/code>,  <\/p>\n\n<p>I want to increment <code>memory<\/code> at the positions indicated by <code>indices<\/code> with the values in <code>updates<\/code>.<\/p>\n\n<p><code>tf.tensor_scatter_nd_add<\/code> does not seem to work:<\/p>\n\n<p><code>tf.tensor_scatter_nd_add(memory, indices, updates)<\/code> returns <code>{InvalidArgumentError}Inner dimensions of output shape must match inner dimensions of updates shape. Output: [a,b,c] updates: [a,c] [Op:TensorScatterAdd]<\/code>.<\/p>\n\n<p>Is it really necessary for <code>updates<\/code> to have as many inner dimensions as <code>memory<\/code> ? In my logic, <code>memory[indices]<\/code> (as a pseudocode) should already be a tensor of shape <code>[a, c]<\/code>. Furthermore, the shape of <code>tf.gather_nd(params=memory, indices=indices, batch_dims=1)<\/code> is already <code>[a, c]<\/code>.<\/p>\n\n<p>Could you please recommend an alternative ?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"57927515","CreationDate":"2019-09-13T16:47:10.890","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>I think what you want is this:<\/p>\n\n<pre><code>import tensorflow as tf\n\na, b, c = 3, 4, 5\nmemory = tf.ones([a, b, c])\nindices = tf.constant([[2], [0], [3]])\nupdates = 10 * tf.reshape(tf.range(a * c, dtype=memory.dtype), [a, c])\nprint(updates.numpy())\n# [[  0.  10.  20.  30.  40.]\n#  [ 50.  60.  70.  80.  90.]\n#  [100. 110. 120. 130. 140.]]\n\n# Make indices for first dimension\nind_a = tf.range(tf.shape(indices, out_type=indices.dtype)[0])\n# Make full indices\nindices_2 = tf.concat([tf.expand_dims(ind_a, 1), indices], axis=1)\n# Scatter add\nout = tf.tensor_scatter_nd_add(memory, indices_2, updates)\nprint(out.numpy())\n# [[[  1.   1.   1.   1.   1.]\n#   [  1.   1.   1.   1.   1.]\n#   [  1.  11.  21.  31.  41.]\n#   [  1.   1.   1.   1.   1.]]\n# \n#  [[ 51.  61.  71.  81.  91.]\n#   [  1.   1.   1.   1.   1.]\n#   [  1.   1.   1.   1.   1.]\n#   [  1.   1.   1.   1.   1.]]\n# \n#  [[  1.   1.   1.   1.   1.]\n#   [  1.   1.   1.   1.   1.]\n#   [  1.   1.   1.   1.   1.]\n#   [101. 111. 121. 131. 141.]]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57926860,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorflow-serving><tf.keras>","CreationDate":"2019-09-13T15:55:41.207","AcceptedAnswerId":null,"OwnerUserId":9891949.0,"Title":"tensorflow\/serving - Input to reshape is a tensor with 100 values, but the requested shape has 10000","Body":"<p>When I train my tf.keras Functional API model and serve it with the tensorflow\/serving docker image I'm getting a shape error when I call the API.<\/p>\n\n<p><strong>How I build my model:<\/strong><\/p>\n\n<pre><code>from tensorflow.keras.layers import Dense, DenseFeatures, Input\nfrom tensorflow.keras.models import Model\nfrom tensorflow import feature_column, string as tf_string\nimport os\n\nfeature_layer_inputs = {}\nfeature_columns = []\nfor header in ['categorical_one', 'categorical_two', 'categorical_three']:\n    feature_columns.append(feature_column.indicator_column(\n        feature_column.categorical_column_with_hash_bucket(header, hash_bucket_size=100)))\n    feature_layer_inputs[header] = Input(shape=(1,), name=header, dtype=tf_string)\n\n\nfc_layer = DenseFeatures(feature_columns)\nfc = fc_layer(feature_layer_inputs)\n\n''' Feature Columns to Dense and merge with attention output '''\n\nfc = Dense(300, activation='relu')(fc)\nfc = Dense(324, activation='relu')(fc)\n\npred = Dense(num_classes, activation='softmax')(fc)\n\ninputs = [v for v in feature_layer_inputs.values()]\n\nmodel = Model(inputs=inputs, outputs=pred)\n\nmodel.compile(...)\n\nmodel.fit(...)\n\nsaved_model_path = \"C:\/Temp\/saved_models\/{}\".format(int(time.time()))\nos.mkdir(saved_model_path)\n\nmodel.save(saved_model_path)\n<\/code><\/pre>\n\n<p><strong>How my Serving Signature Definition looks like:<\/strong><\/p>\n\n<pre><code>The given SavedModel SignatureDef contains the following input(s):\ninputs['categorical_one'] tensor_info:\ndtype: DT_STRING\nshape: (-1, 1)\nname: serving_default_categorical_one:0\ninputs['categorical_two'] tensor_info:\ndtype: DT_STRING\nshape: (-1, 1)\nname: serving_default_categorical_two:0\ninputs['categorical_three'] tensor_info:\ndtype: DT_STRING\nshape: (-1, 1)\nname: serving_default_categorical_three:0\nThe given SavedModel SignatureDef contains the following output(s):\noutputs['dense'] tensor_info:\ndtype: DT_FLOAT\nshape: (-1, num_classes)\nname: StatefulPartitionedCall:0\nMethod name is: tensorflow\/serving\/predict\n<\/code><\/pre>\n\n<p><strong>How I call the Serving API:<\/strong><\/p>\n\n<pre><code>curl -d '{\"instances\": [ {\"categorical_one\": \"ABC\", \"categorical_two\": \"DEF\", \"categorical_three\": \"GHI\"} ] }' -X POST http:\/\/localhost:8501\/v1\/models\/my-model-name\/versions\/1:predict\n<\/code><\/pre>\n\n<p><strong>The error message I get:<\/strong><\/p>\n\n<pre><code>{ \"error\": \"Input to reshape is a tensor with 100 values, but the requested shape has 10000\\n\\t [[{{node StatefulPartitionedCall\/StatefulPartitionedCall\/model\/dense_features\/categorical_one_indicator\/Reshape}}]]\" }\n<\/code><\/pre>\n\n<p><em>Please notice that I just posted the crucial parts of my code and not every line.<\/em><\/p>\n\n<p>Any idea is welcome!<\/p>\n","answers":[{"AnswerId":"57959278","CreationDate":"2019-09-16T14:44:57.690","ParentId":null,"OwnerUserId":"9891949","Title":null,"Body":"<p>It seems like, tensorflow doesn't like the explicitly defined shapes in the input layers for the DenseFeatures.\nI changed the following line of code and it worked.<\/p>\n\n<pre><code>feature_layer_inputs[header] = Input(shape=(), name=header, dtype=tf_string)\n<\/code><\/pre>\n\n<p>I will further check the results with a real trained model ..<\/p>\n"}]}
+{"QuestionId":57926993,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-09-13T16:05:23.710","AcceptedAnswerId":null,"OwnerUserId":12064059.0,"Title":"Prepare input array for keras LSTM with multiple features and multiple timesteps","Body":"<p>I'm trying to prepare the input array for an LSTM, with the goal of predicting the next timestep.<\/p>\n\n<p>My data is made of 2 features and 3 observations, and it's represented by a 3x2 matrix (input).\nMy target is a variable with a single value, and I have two data points. I want to use 2 timesteps to predict the next value.<\/p>\n\n<pre><code>import numpy as np\ninput = np.array([[1,2],[3,4],[5,6]])\ntarget = np.array([10,11])\n<\/code><\/pre>\n\n<p>What should be my input matrix for model.fit, and my input_shape?<\/p>\n","answers":[]}
+{"QuestionId":57927187,"AnswerCount":0,"Tags":"<keras><neural-network><deep-learning><time-series><prediction>","CreationDate":"2019-09-13T16:19:55.547","AcceptedAnswerId":null,"OwnerUserId":11545433.0,"Title":"LSTM Time-Series produces shifted forecast?","Body":"<p>I am doing a time-series forecast with a LSTM NN and Keras. As input features there are two variables (precipitation and temperature) and the one target to be predicted is groundwater-level.<\/p>\n\n<p>It seems to be working quite all right, though there is a serious offset between the actual data and the output (see image). <\/p>\n\n<p>Now I've read that this is can be a classic sign of the network not working, as it seems to be mimicing the output and <\/p>\n\n<blockquote>\n  <p>what the model is actually doing is that when predicting the value at\n  time \u201ct+1\u201d, it simply uses the value at time \u201ct\u201d as its prediction <a href=\"https:\/\/towardsdatascience.com\/how-not-to-use-machine-learning-for-time-series-forecasting-avoiding-the-pitfalls-19f9d7adf424\" rel=\"nofollow noreferrer\">https:\/\/towardsdatascience.com\/how-not-to-use-machine-learning-for-time-series-forecasting-avoiding-the-pitfalls-19f9d7adf424<\/a><\/p>\n<\/blockquote>\n\n<p>However, this is not actually possible in my case, as the target-values are not used as input variable. I am using a multi variate time-series with two features, independent of the output feature. \nAlso, the predicted values are not offset in future (t+1) but rather seem to lag behind (t-1).<\/p>\n\n<p>Does anyone know what could cause this problem? <a href=\"https:\/\/i.stack.imgur.com\/96MLV.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/96MLV.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>This is the complete code of my network: <\/p>\n\n<pre><code># Split in Input and Output Data \nx_1 = data[['MeanT']].values\nx_2 = data[['Precip']].values\ny = data[['Z_424A_6857']].values\n\n# Scale Data\nx = np.hstack([x_1, x_2])\nscaler = MinMaxScaler(feature_range=(0, 1))\nx = scaler.fit_transform(x)\n\nscaler_out = MinMaxScaler(feature_range=(0, 1))\ny = scaler_out.fit_transform(y)\n\n# Reshape Data\nx_1, x_2, y = H.create2feature_data(x_1, x_2, y, window)\ntrain_size = int(len(x_1) * .8)\ntest_size = int(len(x_1)) #  * .5\n\nx_1 = np.expand_dims(x_1, 2) # 3D tensor with shape (batch_size, timesteps, input_dim) \/\/ (nr. of samples, nr. of timesteps, nr. of features)\nx_2 = np.expand_dims(x_2, 2)\ny = np.expand_dims(y, 1)\n\n# Split Training Data\nx_1_train = x_1[:train_size]\nx_2_train = x_2[:train_size]\ny_train = y[:train_size]\n\n# Split Test Data\nx_1_test = x_1[train_size:test_size]\nx_2_test = x_2[train_size:test_size]\ny_test = y[train_size:test_size]\n\n# Define Model Input Sets\ninputA = Input(shape=(window, 1))\ninputB = Input(shape=(window, 1))\n\n# Build Model Branch 1\nbranch_1 = layers.GRU(16, activation=act, dropout=0, return_sequences=False, stateful=False, batch_input_shape=(batch, 30, 1))(inputA)\nbranch_1 = layers.Dense(8, activation=act)(branch_1)\n#branch_1 = layers.Dropout(0.2)(branch_1)\nbranch_1 = Model(inputs=inputA, outputs=branch_1) \n\n# Build Model Branch 2\nbranch_2 = layers.GRU(16, activation=act, dropout=0, return_sequences=False, stateful=False, batch_input_shape=(batch, 30, 1))(inputB)\nbranch_2 = layers.Dense(8, activation=act)(branch_2)\n#branch_2 = layers.Dropout(0.2)(branch_2)\nbranch_2 = Model(inputs=inputB, outputs=branch_2) \n\n# Combine Model Branches\ncombined = layers.concatenate([branch_1.output, branch_2.output])\n\n# apply a FC layer and then a regression prediction on the combined outputs\ncomb = layers.Dense(6, activation=act)(combined)\ncomb = layers.Dense(1, activation=\"linear\")(comb)\n\n# Accept the inputs of the two branches and then output a single value\nmodel = Model(inputs=[branch_1.input, branch_2.input], outputs=comb)\nmodel.compile(loss='mse', optimizer='adam', metrics=['mse', H.r2_score])\n\nmodel.summary()\n\n# Training\nmodel.fit([x_1_train, x_2_train], y_train, epochs=epoch, batch_size=batch, validation_split=0.2, callbacks=[tensorboard])\nmodel.reset_states()\n\n# Evaluation\nprint('Train evaluation')\nprint(model.evaluate([x_1_train, x_2_train], y_train))\n\nprint('Test evaluation')\nprint(model.evaluate([x_1_test, x_2_test], y_test))\n\n# Predictions\npredictions_train = model.predict([x_1_train, x_2_train])\npredictions_test = model.predict([x_1_test, x_2_test])\n\npredictions_train = np.reshape(predictions_train, (-1,1))\npredictions_test = np.reshape(predictions_test, (-1,1))\n\n# Reverse Scaling\npredictions_train = scaler_out.inverse_transform(predictions_train)\npredictions_test = scaler_out.inverse_transform(predictions_test)\n\n# Plot results\nplt.figure(figsize=(15, 6))\nplt.plot(orig_data, color='blue', label='True GWL')  \nplt.plot(range(train_size), predictions_train, color='red', label='Predicted GWL (Training)')\nplt.plot(range(train_size, test_size), predictions_test, color='green', label='Predicted GWL (Test)')\nplt.title('GWL Prediction')  \nplt.xlabel('Day')  \nplt.ylabel('GWL')  \nplt.legend()  \nplt.show()   \n<\/code><\/pre>\n\n<p>I am using a batch size of 30 timesteps, a lookback of 90 timesteps, with a total data size of around 7500 time steps. <\/p>\n\n<p>Any help would be greatly appreciated :-) Thank you!<\/p>\n","answers":[]}
+{"QuestionId":57927688,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-13T17:01:40.890","AcceptedAnswerId":57958211.0,"OwnerUserId":5050423.0,"Title":"Tensorflow - conversion from frozen .pb to .tflite","Body":"<p>I am trying to convert a custom frozen .pb file (saved_model.pb &amp; saved_model.pbtxt) to .tflite format so I can load it onto a Coral Dev Board. Trying to convert with the following python code from the <a href=\"https:\/\/www.tensorflow.org\/lite\/convert\/python_api\" rel=\"nofollow noreferrer\">docs<\/a> (Exporting a quantized GraphDef). The model used was MobileNetV2.<\/p>\n\n<p>File name: convert.py<\/p>\n\n<pre><code>import tensorflow as tf\n\nimg = tf.placeholder(name=\"normalized_input_image_tensor\", dtype=tf.float32, shape=(1, 300, 300, 3))\nconst = tf.constant([1., 2., 3.]) + tf.constant([1., 4., 4.])\nval = img + const\nout = tf.fake_quant_with_min_max_args(val, min=0., max=1., name=\"TFLite_Detection_PostProcess\")\n\nwith tf.Session() as sess:\n  converter = tf.lite.TFLiteConverter.from_session(sess, [img], [out])\n  converter.inference_type = tf.lite.constants.QUANTIZED_UINT8\n  input_arrays = converter.get_input_arrays()\n  converter.quantized_input_stats = {input_arrays[0] : (0., 1.)}  # mean, std_dev\n  tflite_model = converter.convert()\n  open(\"converted_model.tflite\", \"wb\").write(tflite_model)\n<\/code><\/pre>\n\n<p>Being new and naive to tensorflow, I <em>think<\/em> my input and outputs are as follows<\/p>\n\n<p><strong>Input<\/strong>\n<a href=\"https:\/\/i.stack.imgur.com\/OsbNg.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/OsbNg.jpg\" alt=\"Input\"><\/a><\/p>\n\n<p><strong>Output<\/strong>\n<a href=\"https:\/\/i.stack.imgur.com\/n96Zw.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/n96Zw.jpg\" alt=\"Output\"><\/a><\/p>\n\n<p>After running the python code above, the converted .tflite model that outputs is only 1kb in size and looks too simplistic. \n<a href=\"https:\/\/i.stack.imgur.com\/rQplk.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rQplk.jpg\" alt=\"tflite output\"><\/a><\/p>\n\n<p>There are several questions really in this. <\/p>\n\n<ol>\n<li>Is the input in the python code correct? Particularly, name=\"normalized_input_image_tensor\"<\/li>\n<li>Is the output in the python code correct? Particularly, name=\"TFLite_Detection_PostProcess\"<\/li>\n<li>What shape should the .tflite output take - should it be more complex?<\/li>\n<\/ol>\n\n<p>Thanks! Appreciate your patience with newbs as I stumble through.<\/p>\n\n<p>Visualizations with - <a href=\"https:\/\/github.com\/lutzroeder\/Netron\" rel=\"nofollow noreferrer\">https:\/\/github.com\/lutzroeder\/Netron<\/a> <\/p>\n\n<p>EDIT, found this which gave me the input name<\/p>\n\n<pre><code>&gt;&gt;&gt; import tensorflow as tf\n&gt;&gt;&gt; gf = tf.GraphDef()\n&gt;&gt;&gt; gf.ParseFromString(open('saved_model.pb','rb').read())\n&gt;&gt;&gt; [n.name + '=&gt;' +  n.op for n in gf.node if n.op in ( 'Softmax','Placeholder')]\n<\/code><\/pre>\n\n<p>OUTPUT: ['normalized_input_impage_tesnor=>Placeholder']<\/p>\n\n<p>changed code to <\/p>\n\n<pre><code>img = tf.placeholder(name=\"Placeholder\", dtype=tf.float32, shape=(1, 300, 300, 3))\nconst = tf.constant([1., 2., 3.]) + tf.constant([1., 4., 4.])\nval = img + const\nout = tf.fake_quant_with_min_max_args(val, min=0., max=1., name=\"MobilenetV2\/Predictions\/Softmax\")\n<\/code><\/pre>\n\n<p>with same result<\/p>\n","answers":[{"AnswerId":"57958211","CreationDate":"2019-09-16T13:45:25.240","ParentId":null,"OwnerUserId":"5050423","Title":null,"Body":"<pre><code>tflite_convert --output_file=converted_model.tflite --graph_def_file=saved_model.pb --input_arrays=input --output_arrays=TFLite_Detection_PostProcess,TFLite_Detection_PostProcess:1,TFLite_Detection_PostProcess:2,TFLite_Detection_PostProcess:3  --input_shape=1,300,300,3 --allow_custom_ops\n<\/code><\/pre>\n\n<p>And now it works. <\/p>\n"}]}
+{"QuestionId":57927764,"AnswerCount":0,"Tags":"<tensorflow><object-detection-api><data-augmentation>","CreationDate":"2019-09-13T17:09:17.100","AcceptedAnswerId":null,"OwnerUserId":1886357.0,"Title":"Tensorflow object detection api: how to use imgaug for augmentation?","Body":"<p>I've been hand-rolling augmenters using <a href=\"https:\/\/github.com\/aleju\/imgaug\" rel=\"noreferrer\">imgaug<\/a>, as I really like some of the <a href=\"https:\/\/stackoverflow.com\/questions\/44906317\/what-are-possible-values-for-data-augmentation-options-in-the-tensorflow-object\">options that are not available in the tf object detection api<\/a>. For instance, I use motion blur because so much of my data has fast-moving, blurry objects.<\/p>\n\n<p>How can I best integrate my augmentation sequence with the api for on-the-fly training? <\/p>\n\n<p>E.g., say I have an augmenter:<\/p>\n\n<pre><code>aug = iaa.SomeOf((0, 2), \n                 [iaa.Fliplr(0.5), iaa.Flipud(0.5), iaa.Affine(rotate=(-10, 10))])\n<\/code><\/pre>\n\n<p>Is there some way to configure the object detection api to work with this?<\/p>\n\n<p>What I am currently doing is using <code>imgaug<\/code> to generate (augmented) training data, and then creating <code>tfrecord<\/code> files from each iteration of this augmentation pipeline. This is very inefficient as I am saving large amounts of data to disk rather than running augmentation on the fly, during training.<\/p>\n","answers":[]}
+{"QuestionId":57928097,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-13T17:35:24.807","AcceptedAnswerId":null,"OwnerUserId":8270077.0,"Title":"How to deal with this exception? 'ERROR: Cannot uninstall 'wrapt'. It is a distutils installed project etc.'","Body":"<p>I am trying to update wrapt in order to update tensorflow 2.0 to its latest version but I am running into the following exception:<\/p>\n\n<blockquote>\n  <p>ERROR: Cannot uninstall 'wrapt'. It is a distutils installed project and thus we cannot accurately determine which files belong to it which would lead to only a partial uninstall.<\/p>\n<\/blockquote>\n\n<p>How can I deal with this?<\/p>\n\n<p>For the full error message:<\/p>\n\n<pre><code>(base) D:\\DIGITAL_LIBRARY\\Jupyter_Notebook_Files&gt;pip install wrapt --upgrade\nCollecting wrapt\n  Using cached https:\/\/files.pythonhosted.org\/packages\/23\/84\/323c2415280bc4fc880ac5050dddfb3c8062c2552b34c2e512eb4aa68f79\/wrapt-1.11.2.tar.gz\nBuilding wheels for collected packages: wrapt\n  Building wheel for wrapt (setup.py) ... error\n  ERROR: Command errored out with exit status 1:\n   command: 'c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\python.exe' -u -c 'import sys, setuptools, tokenize; sys.argv[0] = '\"'\"'C:\\\\Users\\\\Alienware\\\\AppData\\\\Local\\\\Temp\\\\pip-install-4116x4uj\\\\wrapt\\\\setup.py'\"'\"'; __file__='\"'\"'C:\\\\Users\\\\Alienware\\\\AppData\\\\Local\\\\Temp\\\\pip-install-4116x4uj\\\\wrapt\\\\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\\Alienware\\AppData\\Local\\Temp\\pip-wheel-yzv7oa9k' --python-tag cp36\n       cwd: C:\\Users\\Alienware\\AppData\\Local\\Temp\\pip-install-4116x4uj\\wrapt\\\n  Complete output (46 lines):\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\\wrapt\n  copying src\\wrapt\\decorators.py -&gt; build\\lib.win-amd64-3.6\\wrapt\n  copying src\\wrapt\\importer.py -&gt; build\\lib.win-amd64-3.6\\wrapt\n  copying src\\wrapt\\wrappers.py -&gt; build\\lib.win-amd64-3.6\\wrapt\n  copying src\\wrapt\\__init__.py -&gt; build\\lib.win-amd64-3.6\\wrapt\n  running build_ext\n  Traceback (most recent call last):\n    File \"&lt;string&gt;\", line 1, in &lt;module&gt;\n    File \"C:\\Users\\Alienware\\AppData\\Local\\Temp\\pip-install-4116x4uj\\wrapt\\setup.py\", line 101, in &lt;module&gt;\n      run_setup(with_extensions=True)\n    File \"C:\\Users\\Alienware\\AppData\\Local\\Temp\\pip-install-4116x4uj\\wrapt\\setup.py\", line 71, in run_setup\n      setup(**setup_kwargs_tmp)\n    File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\distutils\\core.py\", line 148, in setup\n      dist.run_commands()\n    File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\distutils\\dist.py\", line 955, in run_commands\n      self.run_command(cmd)\n    File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\distutils\\dist.py\", line 974, in run_command\n      cmd_obj.run()\n    File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\site-packages\\wheel\\bdist_wheel.py\", line 202, in run\n      self.run_command('build')\n    File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\distutils\\cmd.py\", line 313, in run_command\n      self.distribution.run_command(command)\n    File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\distutils\\dist.py\", line 974, in run_command\n      cmd_obj.run()\n    File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\distutils\\command\\build.py\", line 135, in run\n      self.run_command(cmd_name)\n    File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\distutils\\cmd.py\", line 313, in run_command\n      self.distribution.run_command(command)\n    File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\distutils\\dist.py\", line 974, in run_command\n      cmd_obj.run()\n    File \"C:\\Users\\Alienware\\AppData\\Local\\Temp\\pip-install-4116x4uj\\wrapt\\setup.py\", line 25, in run\n      build_ext.run(self)\n    File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\distutils\\command\\build_ext.py\", line 308, in run\n      force=self.force)\n    File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\distutils\\ccompiler.py\", line 1031, in new_compiler\n      return klass(None, dry_run, force)\n    File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\distutils\\cygwinccompiler.py\", line 285, in __init__\n      CygwinCCompiler.__init__ (self, verbose, dry_run, force)\n    File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\distutils\\cygwinccompiler.py\", line 129, in __init__\n      if self.ld_version &gt;= \"2.10.90\":\n  TypeError: '&gt;=' not supported between instances of 'NoneType' and 'str'\n  ----------------------------------------\n  ERROR: Failed building wheel for wrapt\n  Running setup.py clean for wrapt\nFailed to build wrapt\nInstalling collected packages: wrapt\n  Found existing installation: wrapt 1.10.11\nERROR: Cannot uninstall 'wrapt'. It is a distutils installed project and thus we cannot accurately determine which files belong to it which would lead to only a partial uninstall.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57928270,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-13T17:50:00.143","AcceptedAnswerId":null,"OwnerUserId":12064494.0,"Title":"ValueError: Error when checking target: expected dense_1 to have 2 dimensions, but got array with shape (68, 50, 50, 50, 1)","Body":"<p>I'm trying to build a convolutional network that will work on a 3D voxel grid.  I try to add a fully connected layer but get an error:<\/p>\n\n<p>ValueError: Error when checking target: expected dense_1 to have 2 dimensions, but got array with shape (68, 50, 50, 50, 1)<\/p>\n\n<p>How can this be happening when I have a flatten layer first?  Shouldn't my input to the dense layer at that point be, well, flat?<\/p>\n\n<pre><code>x, y = load_data(directory)\nx_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.25, random_state=42)\nmodel = Sequential()\nmodel.add(Convolution3D(1, kernel_size=(3, 3, 3), activation='relu',\n                        border_mode='same', name='conv1',\n                        input_shape=(50, 50, 50, 1)))\nmodel.add(MaxPooling3D(pool_size=(2, 2, 2)))\nmodel.add(Flatten())\nmodel.add(Dense(32))\nmodel.compile(\n    loss='mean_squared_error',\n    optimizer='adam',\n    metrics=['accuracy']\n    )\nmodel.fit(\n    x_train,\n    y_train,\n    epochs=10,\n    batch_size=32,\n    )\nmodel.evaluate(x_test, y_test)\n<\/code><\/pre>\n\n\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nconv1 (Conv3D)               (None, 50, 50, 50, 1)     28        \n_________________________________________________________________\nmax_pooling3d_1 (MaxPooling3 (None, 25, 25, 25, 1)     0         \n_________________________________________________________________\nflatten_1 (Flatten)          (None, 15625)             0         \n_________________________________________________________________\ndense_1 (Dense)              (None, 32)                500032    \n=================================================================\n<\/code><\/pre>\n","answers":[{"AnswerId":"57933175","CreationDate":"2019-09-14T06:53:57.027","ParentId":null,"OwnerUserId":"5825953","Title":null,"Body":"<p><code>train_test_split<\/code> method split arrays into train and test set. If input to the method are list of arrays, the methods returns train and test tuples.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>train_set, test_set = train_test_split(x, y, test_size=0.25, random_state=42)\nx_train, y_train = train_set\nx_test, y_test = test_set\n<\/code><\/pre>\n\n<p>or since python support left side assignment to tuples, <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>(x_train, y_train), (x_test, y_test) = train_test_split(x, y, test_size=0.25, random_state=42)\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":57928562,"AnswerCount":0,"Tags":"<tensorflow><prediction><tensorflow-serving><tensorflow-estimator>","CreationDate":"2019-09-13T18:17:59.260","AcceptedAnswerId":null,"OwnerUserId":2666311.0,"Title":"REST api predict to Tensorflow serving for BoostedTreesClassifier returns error","Body":"<p>I am attempting to make predictions by serving the saved BoostedTreesClassifier model through Tensorflow's Restful API as directed in the following link: <a href=\"https:\/\/www.tensorflow.org\/tfx\/serving\/api_rest#model_metadata_api\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tfx\/serving\/api_rest#model_metadata_api<\/a>. <\/p>\n\n<p>In the following code, I train my boostedTreesClassifier:<\/p>\n\n<pre><code>def trainAndEvalBoostedTrees(train_x, train_y, test_x, test_y, selected_col_objs):\n    def make_inmemory_train_input_fn(train_x, train_y):\n        def input_fn():\n            return dict(train_x), train_y\n\n        return input_fn\n\n    def make_input_fn(test_x, train_y, n_epochs=None, shuffle=True):\n        def input_fn():\n            dataset = tf.data.Dataset.from_tensor_slices((dict(test_x), train_y))\n            if shuffle:\n                dataset = dataset.shuffle(len(train_x))\n            # For training, cycle thru dataset as many times as need (n_epochs=None).\n            dataset = dataset.repeat(n_epochs)\n            # In memory training doesn't use batching.\n            dataset = dataset.batch(len(train_x))\n            return dataset\n\n        return input_fn\n\n    train_input_fn = make_inmemory_train_input_fn(train_x, train_y)\n    eval_input_fn = make_input_fn(test_x, test_y, shuffle=False, n_epochs=1)\n    est = tf.contrib.estimator.boosted_trees_classifier_train_in_memory(\n        train_input_fn,\n        selected_col_objs)\n    return est.evaluate(eval_input_fn)['accuracy'], est\n<\/code><\/pre>\n\n<p>This is how I save the model:<\/p>\n\n<pre><code>def serving_input_receiver_fn():\n    features = {}\n    for key in train_x.keys():\n        dt = train_x[key].dtype\n        features[key] = tf.placeholder(dtype=dt, name=key, shape=[1, 1])\n    return tf.estimator.export.ServingInputReceiver(features, features)\n    boostedTreesLoc = boostedTreesClassifier.export_savedmodel(boosted_trees_save_path, serving_input_receiver_fn=serving_input_receiver_fn)\n<\/code><\/pre>\n\n<p>The post request, requests.post(\"<a href=\"http:\/\/localhost:8501\/v1\/models\/boostedTrees:predict\" rel=\"nofollow noreferrer\">http:\/\/localhost:8501\/v1\/models\/boostedTrees:predict<\/a>\", json=data) reached the saved BoostedTreesClassifier, which was being served on a Docker container, but returned the following error: <\/p>\n\n<p>b'{ \"error\": \"axis = 1 not in [-1, 1)\\n\\t [[{{node boosted_trees\/unstack_1}}]]\" }'\n{'error': 'axis = 1 not in [-1, 1)\\n\\t [[{{node boosted_trees\/unstack_1}}]]'}. <\/p>\n\n<p>I confirmed that the input data was formatted properly, but I am not sure how to debug or fix this problem. I would appreciate any suggestions.<\/p>\n","answers":[]}
+{"QuestionId":57928619,"AnswerCount":1,"Tags":"<tensorflow><keras-layer><tf.keras><tensorflow-hub>","CreationDate":"2019-09-13T18:22:55.797","AcceptedAnswerId":57952186.0,"OwnerUserId":2605956.0,"Title":"Memory leak when running universal-sentence-encoder-large itterating on dataframe","Body":"<p>I have 140K sentences I want to get embeddings for. I am using TF_HUB Universal Sentence Encoder and am iterating over the sentences(I know it's not the best way but when I try to feed over 500 sentences into the model it crashes). \nMy Environment is:\nUbuntu 18.04\nPython 3.7.4\nTF 1.14\nRam: 16gb\nprocessor: i-5<\/p>\n\n<p>my code is:<\/p>\n\n<p><strong>version 1<\/strong>\nI iterate inside the tf.session context manager<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>embed = hub.Module(\"https:\/\/tfhub.dev\/google\/universal-sentence-encoder-large\/3\")\n    df = pandas_repository.get_dataframe_from_table('sentences')\n    with tf.compat.v1.Session() as session:\n        session.run(tf.global_variables_initializer())\n        session.run(tf.tables_initializer())\n        sentence_embedding = None\n        for i, row in df.iterrows():\n            sentence = row['content']\n            embeddings = embed([sentence])\n            sentence_embedding = session.run(embeddings)\n            df.at[i, 'embedding'] = sentence_embedding\n            print('processed index:', i)\n<\/code><\/pre>\n\n<p><strong>version 2<\/strong>\nI open and close a session within each iteration<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>embed = hub.Module(\"https:\/\/tfhub.dev\/google\/universal-sentence-encoder-large\/3\")\n    df = pandas_repository.get_dataframe_from_table('sentences')\n    for i, row in df.iterrows():\n        sentence = row['content']\n        embeddings = embed([sentence])\n        sentence_embedding = None\n        with tf.compat.v1.Session() as session:\n            session.run(tf.global_variables_initializer())\n            session.run(tf.tables_initializer())\n            sentence_embedding = session.run(embeddings)\n            df.at[i, 'embedding'] = sentence_embedding\n            print('processed index:', i)\n<\/code><\/pre>\n\n<p>While <strong>version 2<\/strong> does seem to have some sort of GC and memory is cleared a bit. It still goes over 50 items and explodes.<\/p>\n\n<p><strong>version 1<\/strong> just goes on gobbling memory. <\/p>\n\n<p><strong>The correct solution as given by <a href=\"https:\/\/stackoverflow.com\/users\/7061249\/arnoegw\">arnoegw<\/a><\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def calculate_embeddings(dataframe, table_name):\n    sql_get_sentences = \"SELECT * FROM semantic_similarity.sentences WHERE embedding IS NULL LIMIT 1500\"\n    sql_update = 'UPDATE {} SET embedding = data.embedding FROM (VALUES %s) AS data(id, embedding) WHERE {}.id = data.id'.format(table_name, table_name)\n    df = pandas_repository.get_dataframe_from_sql(sql_get_sentences) \n    with hub.eval_function_for_module(\"https:\/\/tfhub.dev\/google\/universal-sentence-encoder-large\/3\") as embed:    \n        while len(df) &gt;= 0:\n            sentence_array = df['content'].values\n            sentence_embeddings = embed(sentence_array)\n            df['embedding'] = sentence_embeddings.tolist()\n            values = [tuple(x) for x in df[['id', 'embedding']].values]\n            pandas_repository.update_db_from_df('semantic_similarity.sentences', sql_update, values)       \n            df = pandas_repository.get_dataframe_from_sql(sql_get_sentences)\n<\/code><\/pre>\n\n<p>I am a newbee to TF and can use any help I can get.<\/p>\n","answers":[{"AnswerId":"57952186","CreationDate":"2019-09-16T07:28:48.730","ParentId":null,"OwnerUserId":"7061249","Title":null,"Body":"<p>Your code uses tf.Session, so it falls under the TF1.x programming model of first building a dataflow graph and then running it repeatedly with inputs being fed and outputs being fetched from the graph.<\/p>\n\n<p>But your code does not align well with that programming model. Both versions keep adding new applications of (calls to) the hub.Module to the default TensorFlow graph instead of applying it once and running the same graph repeatedly for the various inputs. Version 2 keeps going into and out of tf.Sessions, which frees some memory but is very inefficient.<\/p>\n\n<p>Please see my answer to \"<a href=\"https:\/\/stackoverflow.com\/questions\/56488857\/strongly-increasing-memory-consumption-when-using-elmo-from-tensorflow-hub\/56493291#56493291\">Strongly increasing memory consumption when using ELMo from Tensorflow-Hub<\/a>\" for guidance how to do it right in the graph-based programming model of TensorFlow 1.x.<\/p>\n\n<p>TensorFlow 2.0, which is going to be released soon, defaults to the programming model of \"eager execution\", which does away with graphs and sessions and would have avoided this confusion. TensorFlow Hub will be updated in due course for TF2.0. For a preview close to your use-case, see <a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/hub\/blob\/master\/examples\/colab\/tf2_text_classification.ipynb\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/github\/tensorflow\/hub\/blob\/master\/examples\/colab\/tf2_text_classification.ipynb<\/a><\/p>\n"}]}
+{"QuestionId":57929276,"AnswerCount":0,"Tags":"<deep-learning><pytorch><lstm><recurrent-neural-network><seq2seq>","CreationDate":"2019-09-13T19:19:14.763","AcceptedAnswerId":null,"OwnerUserId":11871369.0,"Title":"LSTM many-to-many synced sequence input and output Pytorch loss not decreasing?","Body":"<p>I am working on a a project where I built an LSTM model for seq2seq, where I have a synced sequence input and output. My audio time series is 32000 in length and my labels are also 32000 in length. And we wish to make a classification (fake or real audio) decision on each sample of the audio. So my tensors look like this for 1 audio example<\/p>\n\n<p>time series<\/p>\n\n<pre><code>tensor([-0.1635, -0.1510, -0.1455,  ..., -0.3338, -0.3163, -0.2944])\n<\/code><\/pre>\n\n<p>labels 0s and 1s corresponding to each sample<\/p>\n\n<pre><code>tensor([0, 0, 0,  ..., 0, 0, 0])\n<\/code><\/pre>\n\n<p>In all the audios, there is about 45-50% of the audio samples that is labeled 1 for fake, and basically the other half labeled 0 for real. So class imbalance isn't an issue here. Also, the fake audio will always be in a certain region of the audio. So for example in a audio time series of length 32000, the fake region could be from 22000 to 30000, or 5000 to 24000 just to give two examples. <\/p>\n\n<p>I initially tried this problem with an LSTM that takes a fixed input size of 32000 and trained it. What I noticed is that my model was only predicting one class and the loss was never decreasing, after like 5 epochs even. I then read that LSTMs with long input sequences like the audio that I have, is prone to vanishing gradients and possibly an unlearnable model. So I changed my training method so that for each batch in the Pytorch Dataloader, we split the audio time series of 32000 in length to smaller chunks (I tried 100, 200, 500, 1000...etc) and we do backpropogation on these smaller chunks. With this method my model was able to have some recall during training so it was predicting both classes, but still the LOSS was not decreasing at all, even after like 5 epochs.<\/p>\n\n<pre><code>class LSTM(nn.Module):\n    def __init__(self, input_size, hidden_size, num_layers, output_dim = 2, batch_size = batch_size):\n        super(LSTM , self).__init__()\n        self.num_layers = num_layers\n        self.input_size = input_size\n        self.hidden_size = hidden_size\n        self.batch_size = batch_size\n        self.rnn_state = None\n        self.lstm = nn.LSTM(self.input_size , self.hidden_size , self.num_layers, batch_first = True)\n        self.to_linear = nn.Linear(self.hidden_size, output_dim)\n    def repackage_rnn_state(self):\n\n        self.rnn_state = self._detach_rnn_state(self.rnn_state)\n\n    def _detach_rnn_state(self, h):\n\n        if isinstance(h, torch.Tensor):\n            return h.detach()\n        else:\n            return tuple(self._detach_rnn_state(v) for v in h)\n\n\n    def forward(self, x):\n        lstm_out, self.rnn_state = self.lstm(x , self.rnn_state)\n        logits = self.to_linear(lstm_out)\n        scores = F.softmax(logits, dim = 2)\n        return scores\n\n\n\nfor epoch in range(num_epochs):\n\n    running_loss_train , accuracy_train = 0 , 0\n\n    network.hidden = network.init_hidden()\n\n    for batch_idx, batch in enumerate(train_loader):\n\n\n        new_time_series , new_labels = batch\n\n        new_time_series, new_labels = new_time_series.to(device), new_labels.to(device)\n\n        optimizer.zero_grad()\n\n        # forward pass feeding the padded sequence\n\n        preds = network(new_time_series.unsqueeze(2))\n\n        preds = preds.view(preds.shape[0]*preds.shape[1], preds.shape[2])\n\n        train_loss = criterion(preds, new_labels.flatten().long())\n\n        train_loss.backward()\n\n        # gradient clipping here just before the parameter updates, clips inplace, clip value = 1\n\n        torch.nn.utils.clip_grad_norm_(network.parameters(),clip)\n\n        optimizer.step()\n\n        running_loss_train+=train_loss.detach().item()\/new_time_series.shape[0]\n\n        _ ,predictions = torch.max(preds,1)\n\n        metrics = get_evaluations_ignoring_padding(new_labels, predictions, 2)\n\n        accuracy_train+=metrics['Accuracy']\n\n\n        if batch_idx%100==0:\n            print('Epoch : %d | BatchID : %d | Train Loss : %.3f | Train Acc : %.3f ' % \n                  (epoch , batch_idx, train_loss.item() \/ new_time_series.shape[0] ,\n                   metrics['Accuracy']))\n<\/code><\/pre>\n\n<p>The above code is for when I did the \"truncated\" backpropogation splitting my time series into multiple chunks. So before I didn't have the repackage_rnn_state or _detach_rnn_state when I trained on the 32000 length. Doing it the original way and with the chunk method again both did not decrease loss.<\/p>\n\n<p>I previously did a many-to-one classification on real vs fake audio, where half of the audio didn't have any fake regions and the other Half has some fake regions. I got 99% accuracy on that CNN model and only after 3 epochs. So I'm thinking my data is pretty easy to train on and just because it's now a many-to-many problem there shouldn't be learning issues.<\/p>\n\n<p>Is there something wrong with my code??<\/p>\n\n<p>Is it not appropriate to train an LSTM on my kind of problem?<\/p>\n\n<p>I'd like some advice about how you would go about training this kind of task.<\/p>\n\n<p>Please let me know if you need more information and I will update my question and thanks<\/p>\n","answers":[]}
+{"QuestionId":57929299,"AnswerCount":2,"Tags":"<pytorch>","CreationDate":"2019-09-13T19:21:17.407","AcceptedAnswerId":57937613.0,"OwnerUserId":2736559.0,"Title":"How to share weights between modules in Pytorch?","Body":"<p>What is the correct way of sharing weights between two layers(modules) in Pytorch?<br>\nBased on my findings in the Pytorch discussion forum, there are several ways for doing this.<br>\nAs an example, based on <a href=\"https:\/\/discuss.pytorch.org\/t\/how-to-share-part-of-weights-in-different-models\/26696\" rel=\"nofollow noreferrer\">this discussion<\/a>, I thought simply assigning the transposed weights would do it. That is doing : <\/p>\n\n<pre><code> self.decoder[0].weight = self.encoder[0].weight.t()\n<\/code><\/pre>\n\n<p>This however, proved to be wrong and causes an error. \nI then tried wrapping the above line in a <code>nn.Parameter()<\/code>: <\/p>\n\n<pre><code>self.decoder[0].weight = nn.Parameter(self.encoder[0].weight.t())\n<\/code><\/pre>\n\n<p>This eliminates the error, but then again, there is no sharing happening here. by this I just initialized a <strong>new<\/strong> tensor with the same values as the <code>encoder[0].weight.t()<\/code>. <\/p>\n\n<p>I then found this <a href=\"https:\/\/gist.github.com\/InnovArul\/500e0c57e88300651f8005f9bd0d12bc\" rel=\"nofollow noreferrer\">link<\/a> which provides different ways for sharing weights. however, I'm skeptical if all methods given there are actually correct.<br>\nFor example, one way is demonstrated like this : <\/p>\n\n<pre><code># tied autoencoder using off the shelf nn modules\nclass TiedAutoEncoderOffTheShelf(nn.Module):\n    def __init__(self, inp, out, weight):\n        super().__init__()\n        self.encoder = nn.Linear(inp, out, bias=False)\n        self.decoder = nn.Linear(out, inp, bias=False)\n\n        # tie the weights\n        self.encoder.weight.data = weight.clone()\n        self.decoder.weight.data = self.encoder.weight.data.transpose(0,1)\n\n    def forward(self, input):\n        encoded_feats = self.encoder(input)\n        reconstructed_output = self.decoder(encoded_feats)\n        return encoded_feats, reconstructed_output\n<\/code><\/pre>\n\n<p>Basically it creates a new weight tensor using <code>nn.Parameter()<\/code> and assigns it to each layer\/module like this : <\/p>\n\n<pre><code>weights = nn.Parameter(torch.randn_like(self.encoder[0].weight))\nself.encoder[0].weight.data = weights.clone()\nself.decoder[0].weight.data = self.encoder[0].weight.data.transpose(0, 1)\n<\/code><\/pre>\n\n<p>This really confuses me, how is this sharing the same variable between these two layers? \nIs it not just cloning the 'raw' <strong>data<\/strong>?<br>\nWhen I used this approach, and visualized the weights, I noticed the visualizations were different and that make me even more certain something is not right.<br>\nI'm not sure if the different visualizations were solely due to one being the transpose of the other one, or as I just already suspected, they are optimized independently (i.e. the weights are not shared between layers) <\/p>\n\n<p>example weight initialization : \n<a href=\"https:\/\/i.stack.imgur.com\/o0n0X.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/o0n0X.jpg\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/h9lve.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/h9lve.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"57929915","CreationDate":"2019-09-13T20:20:30.290","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>Ai questions in general have the tendency to be wrongly understood, this one in particular. \nI will rephrase your question, can layer A from module M1 and layer B from module M2 share the weights WA = WB, and possible WA = WB_transposed.<\/p>\n\n<p>This is possible via PyTorch hooks where you would update forward hook of A to alter the WB and possible you would freeze WB in M2 autograd.<\/p>\n\n<p>So just use hooks.<\/p>\n\n<hr>\n\n<pre><code>from time import sleep\n\nimport torch\nimport torch.nn as nn\nclass M(nn.Module):\n    def __init__(self):        \n        super().__init__()        \n        self.l1 = nn.Linear(1,2)\n\n    def forward(self, x):                      \n        x = self.l1(x)\n        return x\n\nmodel = M()\nmodel.train()\n\ndef printh(module, inp, outp):\n    sleep(1)    \n    print(\"update other model parameter in here...\")   \n\n\nh = model.register_forward_hook(printh)\nfor i in range(1,4):\n\n    x = torch.randn(1)\n    output = model(x)\n\nh.remove()\n<\/code><\/pre>\n"},{"AnswerId":"57937613","CreationDate":"2019-09-14T17:05:05.740","ParentId":null,"OwnerUserId":"2736559","Title":null,"Body":"<p>As it turns out, after further investigation, which was simply retransposing the decoder's weight and visualized it, they were indeed shared.<br>\nBelow is the visualization for encoder and decoders weights : \n<a href=\"https:\/\/i.stack.imgur.com\/wGUkN.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/wGUkN.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/T3t8x.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/T3t8x.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":57929529,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-13T19:43:27.763","AcceptedAnswerId":null,"OwnerUserId":11126450.0,"Title":"Background threads and absl::Notifications in TensorFlow custom-ops","Body":"<p>I am adding a custom op to TensorFlow that essentially needs to launch and wait on a background thread, but seem to have run into a deadlock issue.<\/p>\n\n<p>This is warned against in the TensorFlow documentation as it might prevent other ops from being executed (see tensorflow\/core\/framework\/op_kernel.h), and <code>AsyncOpKernel<\/code> is likely a better match for the task, yet I would like to understand why only option B and C below seem to work, with option A apparently resulting in a deadlock.<\/p>\n\n<p>Running the same code outside TensorFlow gives the behaviour I expected, leading me to believe this is somehow related to the mix between absl and TensorFlow.<\/p>\n\n<pre class=\"lang-cpp prettyprint-override\"><code>class AsyncTestOp : public OpKernel {\n public:\n  explicit AsyncTestOp(OpKernelConstruction* ctx) : OpKernel(ctx) {}\n\n  void Compute(OpKernelContext* ctx) override {\n\n    absl::Notification notification;\n\n    std::thread thread(\n      [](absl::Notification* notification) {\n        std::this_thread::sleep_for(std::chrono::seconds(5));\n        notification-&gt;Notify();\n      },\n      &amp;notification);\n\n    \/\/ Option A: doesn't work\n    notification.WaitForNotification();\n\n    \/\/ Option B: works\n    \/\/ while (!notification.HasBeenNotified()) {\n    \/\/   notification.WaitForNotificationWithTimeout(absl::Milliseconds(100));\n    \/\/ }\n\n    \/\/ Option C: works\n    \/\/ while (!notification.HasBeenNotified()) {\n    \/\/   std::this_thread::sleep_for(std::chrono::milliseconds(100));\n    \/\/ }\n\n    thread.join();\n  }\n};\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57929803,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow2.0>","CreationDate":"2019-09-13T20:09:22.337","AcceptedAnswerId":null,"OwnerUserId":314864.0,"Title":"What is the proper way to convert Tracing Code using RunOptions to Tensorflow 2.0?","Body":"<p>I'm having difficulty finding any documentation on how to migrate tracing code from 1.x to 2.0.<\/p>\n\n<p>In tensorflow 1.x you could do the following:<\/p>\n\n<pre><code>run_options = tf.compat.v1.RunOptions(trace_level=tf.compat.v1.RunOptions.FULL_TRACE)\nrun_metadata = tf.compat.v1.RunMetadata()\nfinal_result = sess.run(result, feed_dict={...},\n                        options=run_options,\n                        run_metadata=run_metadata)\n\ntrace = fetched_timeline = timeline.Timeline(run_metadata.step_stats)\nchrome_trace = fetched_timeline.generate_chrome_trace_format()\nwith open('timeline.json', 'w') as f:\n    f.write(chrome_trace)\n<\/code><\/pre>\n\n<p>How can you do a similar thing with a @tf.function call?<\/p>\n\n<pre><code>@tf.function\ndef predict(x1, x2):\n    ...\n#=============\n# Set run options and RunMetadata variables\n#=============\n#=============\nfinal_result = predict(x1_val, x2_val)\n\n#=============\n# Dump Trace (assuming run_metadata is the RunMetaData object we configured previously)\n#=============\ntrace = fetched_timeline = timeline.Timeline(run_metadata.step_stats)\nchrome_trace = fetched_timeline.generate_chrome_trace_format()\nwith open('timeline.json', 'w') as f:\n    f.write(chrome_trace)\n#=============\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57930073,"AnswerCount":3,"Tags":"<python><tensorflow><keras><batch-normalization>","CreationDate":"2019-09-13T20:38:14.453","AcceptedAnswerId":58051800.0,"OwnerUserId":3954958.0,"Title":"Keras Batchnormalization and sample weights","Body":"<p>I am trying the the training and evaluation example on the tensorflow <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/keras\/training_and_evaluation#training_evaluation_from_tfdata_datasets\" rel=\"noreferrer\">website<\/a>.\nSpecifically, this part:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nfrom tensorflow import keras\nfrom tensorflow.keras import layers\n\n(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()\n\nx_train = x_train.reshape(60000, 784).astype('float32') \/ 255\nx_test = x_test.reshape(10000, 784).astype('float32') \/ 255\n\ny_train = y_train.astype('float32')\ny_test = y_test.astype('float32')\n\ndef get_uncompiled_model():\n  inputs = keras.Input(shape=(784,), name='digits')\n  x = layers.Dense(64, activation='relu', name='dense_1')(inputs)\n  x = layers.BatchNormalization()(x)\n  x = layers.Dense(64, activation='relu', name='dense_2')(x)\n  outputs = layers.Dense(10, activation='softmax', name='predictions')(x)\n  model = keras.Model(inputs=inputs, outputs=outputs)\n  return model\n\ndef get_compiled_model():\n  model = get_uncompiled_model()\n  model.compile(optimizer=keras.optimizers.RMSprop(learning_rate=1e-3),\n              loss='sparse_categorical_crossentropy',\n              metrics=['sparse_categorical_accuracy'])\n  return model\n\nsample_weight = np.ones(shape=(len(y_train),))\nsample_weight[y_train == 5] = 2.\n\n# Create a Dataset that includes sample weights\n# (3rd element in the return tuple).\ntrain_dataset = tf.data.Dataset.from_tensor_slices(\n    (x_train, y_train, sample_weight))\n\n# Shuffle and slice the dataset.\ntrain_dataset = train_dataset.shuffle(buffer_size=1024).batch(64)\n\nmodel = get_compiled_model()\nmodel.fit(train_dataset, epochs=3)\n<\/code><\/pre>\n\n<p>It appears that if I add the batch normalization layer (this line: <code>x = layers.BatchNormalization()(x)<\/code>) I get the following error: <\/p>\n\n<p><code>InvalidArgumentError: The second input must be a scalar, but it has shape [64]\n     [[{{node batch_normalization_2\/cond\/ReadVariableOp\/Switch}}]]<\/code><\/p>\n\n<p>Any ideas?<\/p>\n","answers":[{"AnswerId":"57983037","CreationDate":"2019-09-17T23:05:36.577","ParentId":null,"OwnerUserId":"8788960","Title":null,"Body":"<p>If you just want to use sample weights, you don't have to use <code>tf.data.Dataset<\/code>, you can simply run:<\/p>\n\n<pre><code>model.fit(x=x_train, y=y_train, sample_weight=sample_weight, batch_size=64, epochs=3)\n<\/code><\/pre>\n\n<p>and it works for me (when I change <code>learning_rate<\/code> to <code>lr<\/code> as @ASHu2 mentioned).<\/p>\n\n<p>It gets 97% accuracy after 3 epochs:<\/p>\n\n<pre><code>...\n57408\/60000 [===========================&gt;..] - ETA: 0s - loss: 0.1010 - sparse_categorical_accuracy: 0.9709\n58816\/60000 [============================&gt;.] - ETA: 0s - loss: 0.1011 - sparse_categorical_accuracy: 0.9708\n60000\/60000 [==============================] - 2s 37us\/sample - loss: 0.1007 - sparse_categorical_accuracy: 0.9709\n<\/code><\/pre>\n\n<p>I used TF 1.14.0 on windows.<\/p>\n"},{"AnswerId":"57967151","CreationDate":"2019-09-17T03:59:41.137","ParentId":null,"OwnerUserId":"7454706","Title":null,"Body":"<p>The same code works for me.<\/p>\n\n<p>The only lines I changed are : <\/p>\n\n<p><code>model.compile(optimizer=keras.optimizers.RMSprop(learning_rate=1e-3)<\/code>\nto <code>model.compile(optimizer=keras.optimizers.RMSprop(lr=1e-3)<\/code>\n(which is version specific)<\/p>\n\n<p>Then \n<code>model.fit(train_dataset, epochs=3)<\/code> to <code>model.fit(train_dataset, epochs=3, steps_per_epoch=30)<\/code><br>\nReason :  When using iterators as input to a model, you should specify the <code>steps_per_epoch<\/code> argument<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/sVZqS.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/sVZqS.jpg\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"58051800","CreationDate":"2019-09-22T17:38:39.087","ParentId":null,"OwnerUserId":"3954958","Title":null,"Body":"<p>The problem was solved when I updated tensorflow from version 1.14.1 to 2.0.0-rc1.<\/p>\n"}]}
+{"QuestionId":57930476,"AnswerCount":0,"Tags":"<python><keras><neural-network><generator><conv-neural-network>","CreationDate":"2019-09-13T21:20:39.497","AcceptedAnswerId":null,"OwnerUserId":9119473.0,"Title":"Keras: Create a custom generator for two input model using flow_from _directory() function","Body":"<p>I was trying to train my siamese network with <code>fit_generator()<\/code>,I learned from this answer: <a href=\"https:\/\/stackoverflow.com\/questions\/49404993\/keras-how-to-use-fit-generator-with-multiple-inputs\">Keras: How to use fit_generator with multiple inputs<\/a> that the best way to do this was to create your own generator that yield the multiple data points, my problem was that I retrieve my data with <code>flow_from_directory()<\/code> function and I didn't know if that was possible.<p>\nThis is my attempt to readapt a generator for my problem:<p><\/p>\n\n<pre><code>from keras.models import load_model\nfrom keras import optimizers\nfrom keras.preprocessing.image import ImageDataGenerator\nimport numpy as np\n\nmodel = load_model(\"siamese_model.h5\")\n\ntrain_datagen = ImageDataGenerator(rescale = 1.\/255)\n\ndef generator():\n    t1 = train_datagen.flow_from_directory(base_dir,target_size = (150, 150), batch_size = 20, class_mode = 'categorical',shuffle = True)\n    t2 = train_datagen.flow_from_directory(base_dir,target_size = (150, 150), batch_size = 20, class_mode = 'categorical', shuffle = True)\n    while True:\n        d1,y = t1.next()\n        d2 = t2.next()\n        yield ([d1[0], d2[0]],y)\n\nmodel.compile(loss = 'categorical_crossentropy',optimizer= optimizers.RMSprop(lr=2e-5),metrics=['acc'])\n\nhistory = model.fit_generator(generator(),\n                              steps_per_epoch = 10,\n                              epochs = 5)\n\n<\/code><\/pre>\n\n<p><p><\/p>\n\n<p>My code give me the exact same error as when I tried to fit my model without the custom generator:<p><\/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([[[[0.14509805, 0.15686275, 0.16862746],\n         [0.14509805, 0.15686275, 0.16862746],\n         [0.14509805, 0.15686275, 0.16862746],\n         ...,\n         [0.14117648, 0.15294118, 0.16862746...\n<\/code><\/pre>\n\n<p>What am I doing wrong?<\/p>\n","answers":[]}
+{"QuestionId":57930495,"AnswerCount":1,"Tags":"<audio><deep-learning><conv-neural-network><pytorch><recurrent-neural-network>","CreationDate":"2019-09-13T21:22:11.870","AcceptedAnswerId":null,"OwnerUserId":11871369.0,"Title":"LSTM after CNN how to feed in and what dimensions (input size)?","Body":"<p>I am trying to build a Convolutional Recurrent Network that takes a fixed input size of 32000 x 1 raw audio time series. It's just a numpy array of length 32000. So for example if we have a batch size of 1, my dimensions would be<\/p>\n\n<pre><code>torch.size([1,32000,1])\n<\/code><\/pre>\n\n<p>1 audio time series, 32000 in length, 1 channel<\/p>\n\n<p>I had to reshape that to be<\/p>\n\n<pre><code>torch.size([1,1,32000])\n<\/code><\/pre>\n\n<p>so that it could be fed into the convolutional layer<\/p>\n\n<p>As this is fed into Convolutional \/ Pooling layers my dimensions change as follows<\/p>\n\n<pre><code>torch.Size([1, 1, 32000])\ntorch.Size([1, 64, 7981])\ntorch.Size([1, 64, 1995])\ntorch.Size([1, 64, 1993])\ntorch.Size([1, 64, 498])\ntorch.Size([1, 128, 496])\ntorch.Size([1, 128, 494])\ntorch.Size([1, 128, 123])\ntorch.Size([1, 256, 121])\ntorch.Size([1, 256, 119])\ntorch.Size([1, 256, 29])\ntorch.Size([1, 512, 27])\ntorch.Size([1, 512, 25])\ntorch.Size([1, 512, 6])\n<\/code><\/pre>\n\n<p>So after the final max pooling layer, the size of the data is <\/p>\n\n<pre><code>torch.Size([1, 512, 6])\n<\/code><\/pre>\n\n<p>How do I go about this output and take it into the LSTM?<\/p>\n\n<p>How would we feed this result into the LSTM? would it involve flattening our output from the last max pooling layer? Or is the input size for the LSTM going to be 512 or 6? <\/p>\n\n<p>Please advise and thank you.<\/p>\n\n<p>LSTM expects a 3D input shape of #Samples (audio time series), # Time steps (32000) , # features<\/p>\n","answers":[{"AnswerId":"57936761","CreationDate":"2019-09-14T15:10:25.503","ParentId":null,"OwnerUserId":"2624532","Title":null,"Body":"<p>Adding an LSTM on top of the CNN here doesn\u2019t make sense. You\u2019ve already used the CNN to learn features based on the time dimension, so using an LSTM on top to try and learn the relationship over time of the CNN features that represent the relationship over time doesn\u2019t make sense. <\/p>\n\n<p>You use an LSTM on top of a CNN when you\u2019re using the CNN to learn the spatial dimensions, then the LSTM to learn the time dimension. For example video classification where you can use a CNN to learn features from each image, then an LSTM that takes those features as input and learns features over time.<\/p>\n"}]}
+{"QuestionId":57930566,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-13T21:30:09.453","AcceptedAnswerId":null,"OwnerUserId":7197896.0,"Title":"Trying to add an additional loss to my tensorflow model","Body":"<p>I would like to add a regularization loss to my tf model whereby i penalize the SSIM between in the input and an intermediate represenation.  My code worked fine with tensorflow 1.13.1 but I upgraded today to 1.15 and it no longer works.  <\/p>\n\n<p>Tensorflow says that i must use a placeholder.  But where???<\/p>\n\n<p>I build my model and then add my loss:<\/p>\n\n<pre><code>class MS_SSIM:\n    # Input must be in range [0, max_value]\n\n    def __init__(self, max_value=1.0):\n        self.__name__ = 'MS_SSIM_Objective'\n        self.max_value = max_value\n\n    def __int_shape(self, x):\n        return tf.keras.backend.int_shape(x) if self.backend == 'tensorflow' else tf.keras.backend.shape(x)\n\n    def __call__(self, y_true, y_pred):\n        ms_ssim_val = tf.image.ssim_multiscale(y_true, y_pred, self.max_value)\n        return tf.keras.backend.mean(1.0 - ms_ssim_val)\n\nloss = losses.MS_SSIM()  \nreconstructionLoss = 1e-6*loss(reconstruction.inputs[0], reconstruction.outputs[0])\nmodel.add_loss(reconstructionLoss)  # This is the line where it errors when using tf 1.14 and 1.5 but not 1.13.\n<\/code><\/pre>\n\n<p>I am not sure what to do here. Any thoughts on a fix?<\/p>\n","answers":[]}
+{"QuestionId":57931067,"AnswerCount":1,"Tags":"<keras><tensor><batch-normalization>","CreationDate":"2019-09-13T22:40:53.113","AcceptedAnswerId":57932195.0,"OwnerUserId":12065268.0,"Title":"Unexpectedly found an instance of type of BatchNormalization. Expected a symbolic tensor instance","Body":"<p>I got an error when implementing Residual Network in Keras. Below is the code that gives me error (the error comes from the first line of the final step in the function definition):<\/p>\n\n<p>Load packages:<\/p>\n\n<pre><code>import numpy as np\nfrom keras import layers\nfrom keras.layers import Input, Add, Concatenate, Dense, Activation, ZeroPadding2D, BatchNormalization, Flatten, Conv2D, AveragePooling2D, MaxPooling2D, GlobalMaxPooling2D\nfrom keras.models import Model, load_model\nfrom keras.preprocessing import image\nfrom keras.utils import layer_utils\nfrom keras.utils.data_utils import get_file\nfrom keras.applications.imagenet_utils import preprocess_input\nimport pydot\nfrom IPython.display import SVG\nfrom keras.utils.vis_utils import model_to_dot\nfrom keras.utils import plot_model\nfrom resnets_utils import *\nfrom keras.initializers import glorot_uniform\nimport scipy.misc\nfrom matplotlib.pyplot import imshow\n%matplotlib inline\n\nimport keras.backend as K\nK.set_image_data_format('channels_last')\nK.set_learning_phase(1)\n<\/code><\/pre>\n\n<p>Define the function: (it's the first line of the \"final step\" that gives me the error)<\/p>\n\n<pre><code>def identity_block(X, f, filters, stage, block):\n    \"\"\"\n    Implementation of the identity block as defined in Figure 4\n\n    Arguments:\n    X -- input tensor of shape (m, n_H_prev, n_W_prev, n_C_prev)\n    f -- integer, specifying the shape of the middle CONV's window for the main path\n    filters -- python list of integers, defining the number of filters in the CONV layers of the main path\n    stage -- integer, used to name the layers, depending on their position in the network\n    block -- string\/character, used to name the layers, depending on their position in the network\n\n    Returns:\n    X -- output of the identity block, tensor of shape (n_H, n_W, n_C)\n    \"\"\"\n\n    # defining name basis\n    conv_name_base = 'res' + str(stage) + block + '_branch'\n    bn_name_base = 'bn' + str(stage) + block + '_branch'\n    # Save the input value. You'll need this later to add back to the main path. \n    X_shortcut = X\n\n    # First component of main path\n    X = Conv2D(filters = F1, kernel_size = (1, 1), strides = (1,1), padding = 'valid', name = conv_name_base + '2a', kernel_initializer = glorot_uniform(seed=0))(X)\n    X = BatchNormalization(axis = 3, name = bn_name_base + '2a')(X)\n    X = Activation('relu')(X)\n\n    # Second component of main path \n    X = Conv2D(filters=F2, kernel_size=(f,f),strides=(1,1),padding='same',name=conv_name_base+'2b',kernel_initializer=glorot_uniform(seed=0))(X)\n    X = BatchNormalization(axis=3,name=bn_name_base+'2b')(X)\n    X = Activation('relu')(X)\n\n    # Third component of main path\n    X = Conv2D(filters=F3,kernel_size=(1,1),strides=(1,1),padding='valid',name=conv_name_base+'2c',kernel_initializer=glorot_uniform(seed=0))(X)\n    print(f'before BatchNormalization: X={X}');X = BatchNormalization(axis=3,name=bn_name_base+'2c');print(f'after  BatchNormalization: X={X}');\n\n    # Final step: Add shortcut value to main path, and pass it through a RELU activation\n    X = Add()([X_shortcut,X])\n    X = Activation('relu')(X)\n\n    ### END CODE HERE ###\n\nreturn X\n<\/code><\/pre>\n\n<p>Call\/test the above function:<\/p>\n\n<pre><code>tf.reset_default_graph()\n\nwith tf.Session() as test:\n    np.random.seed(1)\n    A_prev = tf.placeholder(\"float\", [3, 4, 4, 6])\n    X = np.random.randn(3, 4, 4, 6)\n    A = identity_block(A_prev, f = 2, filters = [2, 4, 6], stage = 1, block = 'a')\n    test.run(tf.global_variables_initializer())\n    out = test.run([A], feed_dict={A_prev: X, K.learning_phase(): 0})\n    print(\"out = \" + str(out[0][1][1][0]))\n<\/code><\/pre>\n\n<p>Below is the print message and error message:<\/p>\n\n<p>before BatchNormalization: X=Tensor(\"res1a_branch2c\/BiasAdd:0\", shape=(3, 4, 4, 6), dtype=float32)\nafter  BatchNormalization: X= <code>&lt;keras.layers.normalization.BatchNormalization object at 0x7f169c6d9668&gt;<\/code><\/p>\n\n<pre><code>ValueError: Unexpectedly found an instance of type `&lt;class 'keras.layers.normalization.BatchNormalization'&gt;`. Expected a symbolic tensor instance.\n<\/code><\/pre>\n\n<p>Below is the complete log (in case you need it)<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n\/opt\/conda\/lib\/python3.6\/site-packages\/keras\/engine\/topology.py in assert_input_compatibility(self, inputs)\n    424             try:\n--&gt; 425                 K.is_keras_tensor(x)\n    426             except ValueError:\n\n\/opt\/conda\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py in is_keras_tensor(x)\n    399                           tf.SparseTensor)):\n--&gt; 400         raise ValueError('Unexpectedly found an instance of type `' + str(type(x)) + '`. '\n    401                          'Expected a symbolic tensor instance.')\n\nValueError: Unexpectedly found an instance of type `&lt;class 'keras.layers.normalization.BatchNormalization'&gt;`. Expected a symbolic tensor instance.\n\nDuring handling of the above exception, another exception occurred:\n\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-6-b3d1050f50dc&gt; in &lt;module&gt;()\n      5     A_prev = tf.placeholder(\"float\", [3, 4, 4, 6])\n      6     X = np.random.randn(3, 4, 4, 6)\n----&gt; 7     A = identity_block(A_prev, f = 2, filters = [2, 4, 6], stage = 1, block = 'a')\n      8     test.run(tf.global_variables_initializer())\n      9     out = test.run([A], feed_dict={A_prev: X, K.learning_phase(): 0})\n\n&lt;ipython-input-5-013941ce79d6&gt; in identity_block(X, f, filters, stage, block)\n     43 \n     44     # Final step: Add shortcut value to main path, and pass it through a RELU activation (\u22482 lines)\n---&gt; 45     X = Add()([X_shortcut,X])\n     46     X = Activation('relu')(X)\n     47 \n\n\/opt\/conda\/lib\/python3.6\/site-packages\/keras\/engine\/topology.py in __call__(self, inputs, **kwargs)\n    556                 # Raise exceptions in case the input is not compatible\n    557                 # with the input_spec specified in the layer constructor.\n--&gt; 558                 self.assert_input_compatibility(inputs)\n    559 \n    560                 # Collect input shapes to build layer.\n\n\/opt\/conda\/lib\/python3.6\/site-packages\/keras\/engine\/topology.py in assert_input_compatibility(self, inputs)\n    429                                  'Received type: ' +\n    430                                  str(type(x)) + '. Full input: ' +\n--&gt; 431                                  str(inputs) + '. All inputs to the layer '\n    432                                  'should be tensors.')\n    433 \n\nValueError: Layer add_1 was called with an input that isn't a symbolic tensor. Received type: &lt;class 'keras.layers.normalization.BatchNormalization'&gt;. Full input: [&lt;tf.Tensor 'Placeholder:0' shape=(3, 4, 4, 6) dtype=float32&gt;, &lt;keras.layers.normalization.BatchNormalization object at 0x7f169c6d9668&gt;]. All inputs to the layer should be tensors.\n<\/code><\/pre>\n\n<p>I guess that I missed something in the final step of the function definition, but I have no idea why I got the error. Could any Keras expert here help me out?<\/p>\n","answers":[{"AnswerId":"57932195","CreationDate":"2019-09-14T03:18:12.633","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>Always remember to pass tensors into layers:<\/p>\n\n<pre><code>print(f'before BatchNormalization: X={X}');\n#X = BatchNormalization(axis=3,name=bn_name_base+'2c')    # &lt;--- INCORRECT\nX = BatchNormalization(axis=3,name=bn_name_base+'2c')(X)  # &lt;--- CORRECT\nprint(f'after  BatchNormalization: X={X}');\n<\/code><\/pre>\n\n<p>The difference between 'CORRECT' and 'INCORRECT' is, latter yields a <em>layer<\/em> - whereas former <em>evaluates that layer into a tensor<\/em> when fed with <code>X<\/code>.<\/p>\n\n<p>Furthermore, your <code>identity_block()<\/code> lacks a return, which will throw another error; add: \n<code>return X<\/code>. Lastly, <code>F1, F2, F3<\/code> are neither defined within the function nor passed as arguments - which you may not see as an error since they were defined <em>outside<\/em> the function - e.g. in your local namespace.<\/p>\n"}]}
+{"QuestionId":57931255,"AnswerCount":0,"Tags":"<keras><tensorboard>","CreationDate":"2019-09-13T23:12:14.293","AcceptedAnswerId":null,"OwnerUserId":866082.0,"Title":"How to get metrics (scalar) samples more frequent than per epoch in tensorboard","Body":"<p>I'm trying to show the metrics of my training phase in Tensorboard. This is how I've setup my tensorboard at the moment:<\/p>\n\n<pre><code>model.compile(optimizer='rmsprop', loss='categorical_crossentropy',\n              metrics=['accuracy'])\n\ntensorboard = callbacks.TensorBoard(log_dir='.\/logs', write_graph=False,\n                                    batch_size=28)\n\nmodel.fit_generator(training_generator, epochs=80, steps_per_epoch=25, validation_steps=10,\n                    callbacks=[tensorboard],\n                    validation_data=validation_generator,\n                    workers=3, use_multiprocessing=True,  shuffle=True)\n<\/code><\/pre>\n\n<p>And each epoch I can see a new <code>loss<\/code> and <code>acc<\/code> are added to the chart in tensorboard. But that's not frequent enough for me. I need to increase the frequency of the metrics calculated. Based on what the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/callbacks\/TensorBoard\" rel=\"nofollow noreferrer\">documentation says<\/a>, I should be able to pass an <code>update_freq<\/code> parameter to control this behaviour. The default value for this parameter is <code>epoch<\/code> and other possible values are <code>batch<\/code>, or an integer that indicates the number of samples trained before metrics are recalculated by tensorboard.<\/p>\n\n<p>As long as the <code>update_freq<\/code> is <code>epoch<\/code> (explicitly or by its default value), everything works fine. I see four charts by the names <code>epoch_acc<\/code>, <code>epoch_loss<\/code>, <code>epoch_val_acc<\/code>, and <code>epoch_val_loss<\/code>. And they are updated by each epoch. But the other values for <code>update_freq<\/code> seem not to work. No matter what other value I set for it (<code>batch<\/code> or any integer), I see two charts by the names <code>batch_acc<\/code> and <code>batch_loss<\/code> (no problem so far) but they hold only one data point. And they are not updated (I've waited like 10 epochs before giving up).<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/1QmNY.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1QmNY.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Does anyone know what's going on? Could it have anything to do with my model's structure?<\/p>\n","answers":[]}
+{"QuestionId":57931380,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><pytorch>","CreationDate":"2019-09-13T23:40:06.490","AcceptedAnswerId":null,"OwnerUserId":2603684.0,"Title":"How to convert the following CNN written in tensor-flow to the model in Pytorch?","Body":"<p>I have a existing CNN network model written in tensorflow. I am changing my project to pytorch. I am a newbie in Pytorch. I believe that I have done mostly right in the conversion process of the model. However I am getting an error saying: invalid argument 0: Sizes of tensors must match except in dimension 3. Got 144 and 72 in dimension 1<\/p>\n\n<pre><code>#My tensorflow network that works perfectly\ndef FluidNet( nClasses, nClasses1 ,  input_height=128, input_width=128):\n    ## input_height and width must be devisible by 32 because maxpooling with filter size = (2,2) is operated 5 times,\n    ## which makes the input_height and width 2^5 = 32 times smaller\n    assert input_height%32 == 0\n    assert input_width%32 == 0\n    IMAGE_ORDERING =  \"channels_last\" \n\n    img_input = Input(shape=(input_height,input_width, 6), name='combined_input') ## Assume 128,128,6\n\n    ## Block 1 128x128\n    x = Conv2D(18, (2, 2), activation='relu', padding='same', name='block1_conv1', data_format=IMAGE_ORDERING )(img_input)\n    x = Conv2D(18, (2, 2), activation='relu', padding='same', name='block1_conv2', data_format=IMAGE_ORDERING )(x)\n    x = MaxPooling2D((2, 2), strides=(2, 2), name='block1_pool', data_format=IMAGE_ORDERING )(x)\n    f1 = x\n\n    # Block 2 64x64\n    x = Conv2D(36, (2, 2), activation='relu', padding='same', name='block2_conv1', data_format=IMAGE_ORDERING )(x)\n    x = Conv2D(36, (2, 2), activation='relu', padding='same', name='block2_conv2', data_format=IMAGE_ORDERING )(x)\n    x = MaxPooling2D((2, 2), strides=(2, 2), name='block2_pool', data_format=IMAGE_ORDERING )(x)\n    f2 = x\n\n    # Block 3 32x32\n    x = Conv2D(72, (2, 2), activation='relu', padding='same', name='block3_conv1', data_format=IMAGE_ORDERING )(x)\n    x = Conv2D(72, (2, 3), activation='relu', padding='same', name='block3_conv2', data_format=IMAGE_ORDERING )(x)\n    x = Conv2D(72, (2, 2), activation='relu', padding='same', name='block3_conv3', data_format=IMAGE_ORDERING )(x)\n    x = MaxPooling2D((2, 2), strides=(2, 2), name='block3_pool', data_format=IMAGE_ORDERING )(x)\n    pool3 = x\n\n    # Block 4 16x16\n    x = Conv2D(144, (2, 2), activation='relu', padding='same', name='block4_conv1', data_format=IMAGE_ORDERING )(x)\n    x = Conv2D(144, (2, 2), activation='relu', padding='same', name='block4_conv2', data_format=IMAGE_ORDERING )(x)\n    x = Conv2D(144, (2, 2), activation='relu', padding='same', name='block4_conv3', data_format=IMAGE_ORDERING )(x)\n    pool4 = MaxPooling2D((2, 2), strides=(2, 2), name='block4_pool', data_format=IMAGE_ORDERING )(x)\n\n    # Block 5 8x8\n    x = Conv2D(144, (2, 2), activation='relu', padding='same', name='block5_conv1', data_format=IMAGE_ORDERING )(pool4)\n    x = Conv2D(144, (2, 2), activation='relu', padding='same', name='block5_conv2', data_format=IMAGE_ORDERING )(x)\n    x = Conv2D(144, (2, 2), activation='relu', padding='same', name='block5_conv3', data_format=IMAGE_ORDERING )(x)\n    pool5 = MaxPooling2D((2, 2), strides=(2, 2), name='block5_pool', data_format=IMAGE_ORDERING )(x)\n\n    # Block Transpose &lt;DECODER&gt; : Depth\n    #1st deconv layer 4x4\n    x = (Conv2DTranspose( 72, kernel_size=(4,4) ,  strides=(2,2) , padding='same', dilation_rate = (1,1), use_bias=False, data_format=IMAGE_ORDERING, name=\"Transpose_pool5\" ) (pool5))\n\n    #concatinate x and pool4 for 2nd Deconv layer 8X8\n    x = concatenate ([x, pool4],axis = 3)\n    x = (Conv2DTranspose( 36 , kernel_size=(6,6) ,  strides=(2,2) ,padding='same', dilation_rate = (1,1), use_bias=False, data_format=IMAGE_ORDERING, name=\"Transpose_pool4\")(x))\n\n    #concatinate x and pool3 for 3rd Deconv layer 28x28\n    x = concatenate ([x, pool3],axis = 3)    \n    x= (Conv2DTranspose( 18 , kernel_size=(4,4) ,  strides=(2,2) , padding='same',dilation_rate = (1,1), use_bias=False, data_format=IMAGE_ORDERING, name=\"Transpose_pool3\" )(x))\n\n    #concatinate x and f2 for 4th Deconv layer\n    x = concatenate ([x, f2],axis = 3)    \n    x = (Conv2DTranspose( 9 , kernel_size=(4,4) ,  strides=(2,2) ,  padding='same',dilation_rate = (1,1), use_bias=False, data_format=IMAGE_ORDERING, name=\"Transpose_pool2\" )(x))\n\n    #concatinate x and f1 for 5th Deconv layer\n\n    x = concatenate ([x, f1],axis = 3)    \n    x = (Conv2DTranspose( nClasses + nClasses1 , kernel_size=(3,3) ,  strides=(2,2) , padding='same',dilation_rate = (1,1), use_bias=False, data_format=IMAGE_ORDERING, name=\"Transpose_pool1\" )(x))\n\n    o = x\n    o = (Activation('sigmoid', name=\"depth_out\"))(o)\n\n    x2 = (Conv2DTranspose( 72, kernel_size=(4,4) ,  strides=(2,2) , padding='same', dilation_rate = (1,1), use_bias=False, data_format=IMAGE_ORDERING, name=\"Transpose_pool5_2\" ) (pool5))\n\n    x2 = concatenate ([x2, pool4],axis = 3)\n    x2 = (Conv2DTranspose( 36 , kernel_size=(6,6) ,  strides=(2,2) ,padding='same', dilation_rate = (1,1), use_bias=False, data_format=IMAGE_ORDERING, name=\"Transpose_pool4_2\")(x2))\n\n    x2 = concatenate ([x2, pool3],axis = 3)    \n    x2= (Conv2DTranspose( 18 , kernel_size=(4,4) ,  strides=(2,2) , padding='same',dilation_rate = (1,1), use_bias=False, data_format=IMAGE_ORDERING, name=\"Transpose_pool3_2\" )(x2))\n\n    x2 = concatenate ([x2, f2],axis = 3)    \n    x2 = (Conv2DTranspose( 9 , kernel_size=(4,4) ,  strides=(2,2) ,  padding='same',dilation_rate = (1,1), use_bias=False, data_format=IMAGE_ORDERING, name=\"Transpose_pool2_2\" )(x2))\n\n    x2 = concatenate ([x2, f1],axis = 3)    \n    x2 = (Conv2DTranspose( 7 , kernel_size=(3,3) ,  strides=(2,2) , padding='same',dilation_rate = (1,1), use_bias=False, data_format=IMAGE_ORDERING, name=\"Transpose_pool1_2\" )(x2))\n\n    o2 = x2\n    o2 = (Activation('sigmoid', name=\"scale_out\"))(o2)\n\n    singleOut = concatenate([o,o2],axis = 3, name=\"single_out\")\n\n    #model creation\n    model = Model(img_input, singleOut)\n\n    return model\n\n\n# My pytorch network that gives an error\nclass FluidNet (nn.Module):\ndef __init__(self,nClasses=4,nClasses1=7):\n    super(FluidNet,self).__init__()\n\n    # conv layers: (in_channel size, out_channels size, kernel_size, stride, padding)\n    self.conv1_1 = nn.Conv2d(6, 18, kernel_size=2, padding=1)\n    self.conv1_2 = nn.Conv2d(18, 18, kernel_size=2, padding=1)\n\n    self.conv2_1 = nn.Conv2d(18, 36, kernel_size=2, padding=1)\n    self.conv2_2 = nn.Conv2d(36, 36, kernel_size=2, padding=1)\n\n    self.conv3_1 = nn.Conv2d(36, 72, kernel_size=2, padding=1)\n    self.conv3_2 = nn.Conv2d(72, 72, kernel_size=2, padding=1)\n    self.conv3_3 = nn.Conv2d(72, 72, kernel_size=2, padding=1)\n\n    self.conv4_1 = nn.Conv2d(72, 144, kernel_size=2, padding=1)\n    self.conv4_2 = nn.Conv2d(144, 144, kernel_size=2, padding=1)\n    self.conv4_3 = nn.Conv2d(144, 144, kernel_size=2, padding=1)\n\n    self.conv5_1 = nn.Conv2d(144, 144, kernel_size=2, padding=1)\n    self.conv5_2 = nn.Conv2d(144, 144, kernel_size=2, padding=1)\n    self.conv5_3 = nn.Conv2d(144, 144, kernel_size=2, padding=1)\n\n    # max pooling (kernel_size, stride)\n    self.pool = nn.MaxPool2d(kernel_size=2, stride=2,padding=0)\n\n    # Block Transpose &lt;DECODER&gt; : Depth and Normal : nClasses = 4\n    self.up5_a = nn.ConvTranspose2d(144,72,4, stride=2,  padding=2, bias=False)\n    self.up4_a = nn.ConvTranspose2d(72,36,6, stride=2,  padding=2, bias=False)\n    self.up3_a = nn.ConvTranspose2d(36,18,4, stride=2,  padding=2, bias=False)\n    self.up2_a = nn.ConvTranspose2d(18,9,4, stride=2, padding=2, bias=False)\n    self.up1_a = nn.ConvTranspose2d(9,nClasses,3,stride=2, padding=2, bias=False)\n\n    # Block Transpose &lt;DECODER&gt; : Image+Ref+Scale: nClasses1 = 7\n    self.up5_b = nn.ConvTranspose2d(144,72,4, stride=2, padding=2, bias=False)\n    self.up4_b = nn.ConvTranspose2d(72,36,6, stride=2,  padding=2, bias=False)\n    self.up3_b = nn.ConvTranspose2d(36,18,4, stride=2,  padding=2, bias=False)\n    self.up2_b = nn.ConvTranspose2d(18,9,4, stride=2,  padding=2, bias=False)\n    self.up1_b = nn.ConvTranspose2d(9,nClasses1,3, stride=2,  padding=2, bias=False)\n\ndef forward(self, x, training=True):\n    x = F.relu(self.conv1_1(x))\n    x = F.relu(self.conv1_2(x))\n    x = self.pool(x)\n    pool1 = x\n\n    x = F.relu(self.conv2_1(x))\n    x = F.relu(self.conv2_2(x))\n    x = self.pool(x)\n    pool2 = x\n\n    x = F.relu(self.conv3_1(x))\n    x = F.relu(self.conv3_2(x))\n    x = F.relu(self.conv3_3(x))\n    x = self.pool(x)\n    pool3 = x\n\n    x = F.relu(self.conv4_1(x))\n    x = F.relu(self.conv4_2(x))\n    x = F.relu(self.conv4_3(x))\n    x = self.pool(x)\n    pool4 = x\n\n    x = F.relu(self.conv5_1(x))\n    x = F.relu(self.conv5_2(x))\n    x = F.relu(self.conv5_3(x))\n    x = self.pool(x)\n    pool5 = x\n\n    x = self.up5_a(x)\n    print(x.size())\n    print(pool4.size())\n    x = torch.cat((x,pool4),3)\n    x = self.up4_a(x)\n    x = torch.cat((x,pool3),3)\n    x = self.up3_a(x)\n    x = torch.cat((x,pool2),3)\n    x = self.up2_a(x)\n    x = torch.cat((x,pool1),3)\n    x = self.up1_a(x)\n    o = x\n    o = torch.sigmoid(o)\n\n    x1 = self.up5_a(pool5)\n    x1 = torch.cat((x1,pool4),3)\n    x1 = self.up4_a(x1)\n    x1 = torch.cat((x1,pool3),3)\n    x1 = self.up3_a(x1)\n    x1 = torch.cat((x1,pool2),3)\n    x1 = self.up2_a(x1)\n    x1 = torch.cat((x1,pool1),3)\n    x1 = self.up1_a(x1)\n    o1 = x1\n    o1 = torch.sigmoid(o1)\n\n    return o,o1\n<\/code><\/pre>\n\n<p>These print statements in my pytorch code (print(x.size()), print(pool4.size())) prints out to be: torch.Size([1, 72, 10, 10])\ntorch.Size([1, 144, 10, 10]). Please help me make this transfer. Thank you.<\/p>\n","answers":[{"AnswerId":"57936939","CreationDate":"2019-09-14T15:35:26.050","ParentId":null,"OwnerUserId":"2624532","Title":null,"Body":"<p>It looks like you\u2019re trying to concatenate your output channels\/feature maps which are dimension 1 in Pytorch, not 3. This can be seen in your print outputs.<\/p>\n"}]}
+{"QuestionId":57931586,"AnswerCount":1,"Tags":"<python><neural-network><pytorch><seq2seq>","CreationDate":"2019-09-14T00:35:20.980","AcceptedAnswerId":57938618.0,"OwnerUserId":2612484.0,"Title":"Masking: Mask everything after a specified token (eos)","Body":"<p>My <code>tgt<\/code> tensor is in shape of <code>[12, 32, 1]<\/code> which is <code>sequence_length, batch_size, token_idx<\/code>.<\/p>\n\n<p>What is the best way to create a mask which has ones for entries with <code>&lt;eos&gt;<\/code> and before in sequence, and zeros afterwards?<\/p>\n\n<p>Currently I'm calculating my mask like this, which simply puts zeros where <code>&lt;blank&gt;<\/code> is, ones otherwise.<\/p>\n\n<p><code>mask = torch.zeros_like(tgt).masked_scatter_((tgt != tgt_padding), torch.ones_like(tgt))<\/code><\/p>\n\n<p>But the problem is, that my <code>tgt<\/code> can contain <code>&lt;blank&gt;<\/code> as well (before <code>&lt;eos&gt;<\/code>), in which cases I don't want to mask it out.<\/p>\n\n<p>My temporary solution:<\/p>\n\n<pre><code>mask = torch.ones_like(tgt)\nfor eos_token in (tgt == tgt_eos).nonzero():\n    mask[eos_token[0]+1:,eos_token[1]] = 0\n<\/code><\/pre>\n","answers":[{"AnswerId":"57938618","CreationDate":"2019-09-14T19:28:33.353","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p>I guess you are trying to create a mask for the PAD tokens. There are several ways. One of them is as follows.<\/p>\n\n<pre><code># tensor is of shape [seq_len, batch_size, 1]\ntensor = tensor.mul(tensor.ne(PAD).float())\n<\/code><\/pre>\n\n<p>Here, <code>PAD<\/code> stands for the index of the <code>PAD_TOKEN<\/code>. <code>tensor.ne(PAD)<\/code> will create a byte tensor where at <code>PAD_TOKEN<\/code> positions, 0 will be assigned and 1 elsewhere.<\/p>\n\n<hr>\n\n<p>If you have examples like, <code>\"&lt;s&gt; I think &lt;pad&gt; so &lt;\/s&gt; &lt;pad&gt; &lt;pad&gt;\"<\/code>. Then, I would suggest using different PAD tokens, for before and after <code>&lt;\/s&gt;<\/code>. <\/p>\n\n<p>OR, if you have the length information for each sentence (in the above example, the sentence length is 6), then you can create the mask using the following function.<\/p>\n\n<pre><code>def sequence_mask(lengths, max_len=None):\n    \"\"\"\n    Creates a boolean mask from sequence lengths.\n    :param lengths: 1d tensor [batch_size]\n    :param max_len: int\n    \"\"\"\n    batch_size = lengths.numel()\n    max_len = max_len or lengths.max()\n    return (torch.arange(0, max_len, device=lengths.device)  # (0 for pad positions)\n            .type_as(lengths)\n            .repeat(batch_size, 1)\n            .lt(lengths.unsqueeze(1)))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57931967,"AnswerCount":2,"Tags":"<pytorch><normalization>","CreationDate":"2019-09-14T02:13:18.723","AcceptedAnswerId":null,"OwnerUserId":5380809.0,"Title":"PyTorch normalize two sets of gradients during training","Body":"<p>In this <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/dcgan_faces_tutorial.html\" rel=\"nofollow noreferrer\">GAN tutorial<\/a>, if you scroll down to the training loop you can see they combine the gradients \n <code>errD = errD_real + errD_fake<\/code> like this. Where <code>errD_real = criterion(output, label)<\/code> and <code>errD_fake = criterion(output, label)<\/code> and <code>criterion = nn.BCELoss()<\/code>. I want to do the same thing but before doing a backward pass I want to normalize both gradients to the lower Euclidean norm of the two. How would I do that?<\/p>\n\n<p>I know I can access the gradients of each weight individually on netD by printing out <code>netD.weight.grad<\/code>, but is there some way to batchnorm them to the lower Euclidean norm of the two?<\/p>\n\n<p>Here's the part of the training loop I'm talking about:<\/p>\n\n<pre><code>for epoch in range(num_epochs):\n    # For each batch in the dataloader\n    for i, data in enumerate(dataloader, 0):\n\n        ############################\n        # (1) Update D network: maximize log(D(x)) + log(1 - D(G(z)))\n        ###########################\n        ## Train with all-real batch\n        netD.zero_grad()\n        # Format batch\n        real_cpu = data[0].to(device)\n        b_size = real_cpu.size(0)\n        label = torch.full((b_size,), real_label, device=device)\n        # Forward pass real batch through D\n        output = netD(real_cpu).view(-1)\n        # Calculate loss on all-real batch\n        errD_real = criterion(output, label)\n        # Calculate gradients for D in backward pass\n        errD_real.backward()\n        D_x = output.mean().item()\n\n        ## Train with all-fake batch\n        # Generate batch of latent vectors\n        noise = torch.randn(b_size, nz, 1, 1, device=device)\n        # Generate fake image batch with G\n        fake = netG(noise)\n        label.fill_(fake_label)\n        # Classify all fake batch with D\n        output = netD(fake.detach()).view(-1)\n        # Calculate D's loss on the all-fake batch\n        errD_fake = criterion(output, label)\n        # Calculate the gradients for this batch\n        errD_fake.backward()\n        D_G_z1 = output.mean().item()\n        # Add the gradients from the all-real and all-fake batches\n        errD = errD_real + errD_fake\n        # Update D\n        optimizerD.step()\n        ...\n<\/code><\/pre>\n","answers":[{"AnswerId":"57939007","CreationDate":"2019-09-14T20:22:48.870","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p>You mentioned two things which are incorrect. <\/p>\n\n<ol>\n<li><code>errD = errD_real + errD_fake<\/code> basically combines two objective function, not gradients. <\/li>\n<li>You need to change the gradients after the backward pass, not before. With the backward pass, the gradients will be computed and then you can modify the gradients before calling <code>optimizer.step()<\/code> which updates the model parameters.<\/li>\n<\/ol>\n\n<p>Answer to your question:<\/p>\n\n<p>I believe there is no direct API support from PyTorch to achieve your goal. However, you can easily modify the gradients by yourself after the backward pass. To compute the Euclidean norm, you can use the <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html?highlight=norm#torch.norm\" rel=\"nofollow noreferrer\">norm<\/a> function. Then you can simply <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html?highlight=div_#torch.Tensor.div_\" rel=\"nofollow noreferrer\">divide<\/a> the gradients (inplace).<\/p>\n"},{"AnswerId":"57943594","CreationDate":"2019-09-15T11:17:10.840","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>You better leave the gradients intact and make your optimizer so that it will count the effects you need.<\/p>\n\n<p>Gradients will in most cases be deleted before a new forward anyway.<\/p>\n\n<p>Some newer algorithms such as <a href=\"https:\/\/arxiv.org\/abs\/1904.00962\" rel=\"nofollow noreferrer\">Lamb<\/a> do the trick of parameter-wise (layer-wise) gradient normalization (MSE in this case) like you plan.<\/p>\n\n<p>Also, check the <a href=\"https:\/\/arxiv.org\/abs\/1805.08318\" rel=\"nofollow noreferrer\">SAGAN<\/a> paper and why they used <code>nn.utils.spectral_norm<\/code> since you mantioned GANs.<\/p>\n"}]}
+{"QuestionId":57932156,"AnswerCount":0,"Tags":"<python><numpy><tensorflow><scipy>","CreationDate":"2019-09-14T03:09:40.820","AcceptedAnswerId":null,"OwnerUserId":363028.0,"Title":"Numerical package for uint256?","Body":"<p>I have some code that is beautifully expressed using vectors of 256-bit unsigned integers.<\/p>\n\n<p>Currently I use normal Python to loop over the vector, but I would like to speed this up.<\/p>\n\n<p>TensorFlow, NumPy and SciPy does not seem to support uint256.<\/p>\n\n<p>Is there a numerical package in which I can use uint256 in vectors? Does this package support arbitrarily long uints?<\/p>\n","answers":[]}
+{"QuestionId":57932318,"AnswerCount":0,"Tags":"<pytorch>","CreationDate":"2019-09-14T03:54:29.453","AcceptedAnswerId":null,"OwnerUserId":12065771.0,"Title":"AttributeError: Can't get attribute '_default_unk_index' on <module 'torchtext.vocab'","Body":"<p>I run my code in google colab, but I meet some error.<\/p>\n\n<p>torch:1.1.0, torchtext:0.4.0, I update them, but it's useless<\/p>\n\n<pre><code>[2019-09-14 02:57:15,574 INFO]  * src vocab size = 49766\n[2019-09-14 02:57:15,574 INFO]  * history vocab size = 49766\n[2019-09-14 02:57:15,574 INFO]  * tgt vocab size = 49766\n[2019-09-14 02:57:15,574 INFO] Building model...\nTraceback (most recent call last):\n  File \"\/content\/drive\/My Drive\/ReDR\/train.py\", line 109, in &lt;module&gt;\n    main(opt)\n  File \"\/content\/drive\/My Drive\/ReDR\/train.py\", line 41, in main\n    single_main(opt, -1)\n  File \"\/content\/drive\/My Drive\/ReDR\/onmt\/train_single.py\", line 86, in main\n    model = build_model(model_opt, opt, fields, checkpoint)\n  File \"\/content\/drive\/My Drive\/ReDR\/onmt\/model_builder.py\", line 235, in build_model\n    vocab = torch.load(opt.drqa_vocab_path)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/serialization.py\", line 387, in load\n    return _load(f, map_location, pickle_module, **pickle_load_args)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/serialization.py\", line 574, in _load\n    result = unpickler.load()\nAttributeError: Can't get attribute '_default_unk_index' on &lt;module 'torchtext.vocab' from '\/usr\/local\/lib\/python3.6\/dist-packages\/torchtext\/vocab.py'&gt;\n<\/code><\/pre>\n\n<p>I will appreciate any help from you!<\/p>\n","answers":[]}
+{"QuestionId":57932584,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-14T04:57:13.643","AcceptedAnswerId":null,"OwnerUserId":3496060.0,"Title":"Equivalent of from tensorflow.contrib.framework.python.ops import arg_scope in tf2.0?","Body":"<p>I'm looking for the equivalent of from tensorflow.contrib.framework.python.ops import arg_scope and add_arg_scope since the contrib module has been eliminated in tf 2.0.<\/p>\n","answers":[]}
+{"QuestionId":57932647,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><generative-adversarial-network>","CreationDate":"2019-09-14T05:11:19.787","AcceptedAnswerId":null,"OwnerUserId":10435959.0,"Title":"Fetch argument None has invalid type <type 'NoneType'>","Body":"<p>I am trying to train a 'Generator' for GAN. When I wanna feed the placeholder i get the following error:<\/p>\n\n<pre><code>_grads = mySess.run(params[\"gradient\"],\n         feed_dict={params[\"data\"][\"x\"]: batch_Z})\n<\/code><\/pre>\n\n<p>The <code>params[\"gradient\"]<\/code> is:<\/p>\n\n<pre><code>params[\"gradient\"] = optimizer.compute_gradients(loss, trainable_vars)\n<\/code><\/pre>\n\n<p>params[\"data\"][\"x\"] = tf.placeholder(tf.float32, [None, z_shape])<\/p>\n\n<p>loss and trainable_vars are:<\/p>\n\n<pre><code>loss = cost(tf.ones_like(disc_logits_fake), disc_logits_fake)\ntrainable_vars = [generator.W_conv1, generator.b_conv1, generator.W_conv2, generator.b_conv2, generator.W_conv3, generator.b_conv3, generator.W_conv4, generator.b_conv4]\n<\/code><\/pre>\n\n<p>The error I get is:<\/p>\n\n<blockquote>\n  <p>TypeError: Fetch argument None has invalid type <\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":57933011,"AnswerCount":1,"Tags":"<pytorch><torchtext>","CreationDate":"2019-09-14T06:23:30.713","AcceptedAnswerId":57936614.0,"OwnerUserId":11473132.0,"Title":"torchtext data build_vocab \/ data_field","Body":"<p>I want to ask you some about torchtext.<\/p>\n\n<p>I have a task about abstractive text summarization, and I build a seq2seq model with pytorch.<\/p>\n\n<p>I just wonder about data_field constructed by build_vocab function in torchtext.<\/p>\n\n<p>In machine translation, i accept that two data_fields(input, output) are needed.<\/p>\n\n<p>But, in summarization, input data and output data are same language.<\/p>\n\n<p>Here, should I make two data_field(full_sentence, abstract_sentence) in here?<\/p>\n\n<p>Or is it okay to use only one data_field?<\/p>\n\n<p>I'm afraid that my wrong choice make model's performance down.<\/p>\n\n<p>Please, give me a hint.<\/p>\n","answers":[{"AnswerId":"57936614","CreationDate":"2019-09-14T14:52:04.193","ParentId":null,"OwnerUserId":"2160809","Title":null,"Body":"<p>You are right in the case of summarization and other tasks, it makes sense to build and use the same vocab for input and output<\/p>\n"}]}
+{"QuestionId":57933053,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><object-detection>","CreationDate":"2019-09-14T06:31:42.073","AcceptedAnswerId":null,"OwnerUserId":12065950.0,"Title":"How to add additional classes to a pre-trained object detection model and train it to detect all of the classes (pre-trained + new)?","Body":"<p>I had followed this blog --> <a href=\"https:\/\/medium.com\/@teyou21\/training-your-object-detection-model-on-tensorflow-part-2-e9e12714bdf\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@teyou21\/training-your-object-detection-model-on-tensorflow-part-2-e9e12714bdf<\/a> , and built a SSD Mobilenet model which is pre-trained on the COCO Dataset called \"ssd_mobilenet_v2_quantized_coco\". <\/p>\n\n<p>What happens here is that it perfectly detects my new classes, but I want to include the pre-trained classes as well. <\/p>\n\n<p>I tried changing the number of classes to 96 ( 90 pre-trained + 6 new ) and edited the \"labelmap.pbtxt\" with the name and corresponding id of all labels from the COCO Dataset with the new classes being added at the last from ids 91 - 96.<\/p>\n\n<p>It still detects only the new classes only.<\/p>\n\n<p>What should I do to detect both the pre-trained and new classes?<\/p>\n","answers":[]}
+{"QuestionId":57933393,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-09-14T07:27:33.843","AcceptedAnswerId":null,"OwnerUserId":8262057.0,"Title":"Implementing cosine_decay with LearningRateScheduler","Body":"<p>I want to use <code>LearningRateScheduler<\/code> to implment the cosine decay learning rate, and whenever I implement the following code it gave the error below:<\/p>\n\n<blockquote>\n  <p>ValueError: The output of the \"schedule\" function should be float.<\/p>\n<\/blockquote>\n\n<pre><code>class CosineDecreaseX(object):\n    def __init__(self,  decay_steps , baseLR):\n      self.decay_steps = decay_steps\n      self.baseLR = baseLR\n\n   def lr_func(self, epoch):\n      LR = tf.compat.v1.train.cosine_decay(learning_rate=self.baseLR,global_step= epoch,decay_steps= self.decay_steps)\n      return LR\n\n decay = CosineDecrease( decay_steps=1000 ,  baseLR= 0.0001)\n reduce_lr = tf.compat.v1.keras.callbacks.LearningRateScheduler(decay.lr_func, verbose=1)\n<\/code><\/pre>\n\n<p>Can anyone tell me what is wrong with my code?<\/p>\n","answers":[]}
+{"QuestionId":57933781,"AnswerCount":1,"Tags":"<python><pytorch><torch><tensor>","CreationDate":"2019-09-14T08:27:32.307","AcceptedAnswerId":null,"OwnerUserId":10606134.0,"Title":"How Pytorch Tensor get the index of elements?","Body":"<p>I have 2 Tensors named <strong>x<\/strong> and <strong>list<\/strong> and their definitions are below: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = torch.tensor(3)\nlist = torch.tensor([1,2,3,4,5])\n<\/code><\/pre>\n\n<p>Now I want to get the index of element <strong>x<\/strong> from <strong>list<\/strong>. The expected output is an Integer: <\/p>\n\n<pre><code>2\n<\/code><\/pre>\n\n<p>How can I do in an easy way?<\/p>\n","answers":[{"AnswerId":"57938564","CreationDate":"2019-09-14T19:21:21.213","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<pre><code>import torch\n\nx = torch.tensor(3)\n\nlist = torch.tensor([1,2,3,4,5])\nidx = (list == x).nonzero().flatten()\nprint (idx.tolist()) # [2]\n\nlist = torch.tensor([1,2,3,3,5])\nidx = (list == x).nonzero().flatten()\nprint (idx.tolist()) # [2, 3]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57934357,"AnswerCount":0,"Tags":"<python><tensorflow><python-unicode>","CreationDate":"2019-09-14T09:53:11.717","AcceptedAnswerId":null,"OwnerUserId":11473190.0,"Title":"How to solve UnicodeDecodeError when importing tensorflow.contrib.slim","Body":"<p>I have this problem when I import tensorflow.contrib.slim<\/p>\n\n<pre><code>import tensorflow.contrib.slim as slim\n<\/code><\/pre>\n\n<p>the traceback is:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"finetune-sintel.py\", line 22, in &lt;module&gt;\n    from model_pwcnet import ModelPWCNet, _DEFAULT_PWCNET_FINETUNE_OPTIONS\n  File \"\/home\/sxl\/ext4-2T-3\/code\/tfoptflow-master\/tfoptflow\/model_pwcnet.py\", line 21, in &lt;module&gt;\n    from model_base import ModelBase\n  File \"\/home\/sxl\/ext4-2T-3\/code\/tfoptflow-master\/tfoptflow\/model_base.py\", line 15, in &lt;module&gt;\n    import tensorflow.contrib.slim as slim\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/__init__.py\", line 22, in &lt;module&gt;\n    from tensorflow.contrib import bayesflow\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/bayesflow\/__init__.py\", line 24, in &lt;module&gt;\n    from tensorflow.contrib.bayesflow.python.ops import entropy\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/bayesflow\/python\/ops\/entropy.py\", line 23, in &lt;module&gt;\n    from tensorflow.contrib.bayesflow.python.ops.entropy_impl import *\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/bayesflow\/python\/ops\/entropy_impl.py\", line 30, in &lt;module&gt;\n    from tensorflow.contrib.bayesflow.python.ops import variational_inference\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/bayesflow\/python\/ops\/variational_inference.py\", line 26, in &lt;module&gt;\n    from tensorflow.contrib.bayesflow.python.ops.variational_inference_impl import *\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/bayesflow\/python\/ops\/variational_inference_impl.py\", line 29, in &lt;module&gt;\n    from tensorflow.contrib.bayesflow.python.ops import stochastic_graph_impl as sg\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/bayesflow\/python\/ops\/stochastic_graph_impl.py\", line 28, in &lt;module&gt;\n    from tensorflow.contrib.bayesflow.python.ops import stochastic_tensor_impl\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/bayesflow\/python\/ops\/stochastic_tensor_impl.py\", line 51, in &lt;module&gt;\n    from tensorflow.contrib.distributions.python.ops import distribution\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/distributions\/__init__.py\", line 91, in &lt;module&gt;\n    from tensorflow.contrib.distributions.python.ops.bernoulli import *\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/distributions\/python\/ops\/bernoulli.py\", line 21, in &lt;module&gt;\n    from tensorflow.contrib.distributions.python.ops import distribution\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/distributions\/python\/ops\/distribution.py\", line 29, in &lt;module&gt;\n    from tensorflow.contrib import framework as contrib_framework\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/__init__.py\", line 81, in &lt;module&gt;\n    from tensorflow.contrib.framework.python.ops import *\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/__init__.py\", line 26, in &lt;module&gt;\n    from tensorflow.contrib.framework.python.ops.variables import *\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/variables.py\", line 26, in &lt;module&gt;\n    from tensorflow.contrib.framework.python.ops import gen_variable_ops\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/gen_variable_ops.py\", line 40, in &lt;module&gt;\n    _ops.RegisterShape(\"ZeroInitializer\")(None)\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1697, in __call__\n    self._op_type)\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/registry.py\", line 67, in register\n    stack = traceback.extract_stack()\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/traceback.py\", line 211, in extract_stack\n    stack = StackSummary.extract(walk_stack(f), limit=limit)\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/traceback.py\", line 364, in extract\n    f.line\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/traceback.py\", line 286, in line\n    self._line = linecache.getline(self.filename, self.lineno).strip()\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/linecache.py\", line 16, in getline\n    lines = getlines(filename, module_globals)\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/linecache.py\", line 47, in getlines\n    return updatecache(filename, module_globals)\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/linecache.py\", line 137, in updatecache\n    lines = fp.readlines()\n  File \"\/home\/sxl\/anaconda3\/envs\/tfpy36\/lib\/python3.6\/codecs.py\", line 321, in decode\n    (result, consumed) = self._buffer_decode(data, self.errors, final)\nUnicodeDecodeError: 'utf-8' codec can't decode byte 0xcf in position 4307: invalid continuation byte\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57934577,"AnswerCount":0,"Tags":"<python><tensorflow><tensorboard><tensorflow-datasets>","CreationDate":"2019-09-14T10:23:24.550","AcceptedAnswerId":null,"OwnerUserId":7377486.0,"Title":"Understanding tensorboard with tensorflow model training","Body":"<p>I am trying to create a accident detector (from pictures or videos) for that I collected about 150 images (downloaded from google) of normal and accidental cars. I then used labelimg to create xmls and then converted those xmls to csv to tfrecords (used someone elses script). For model I downloaded ssd mobilenet with ssd pets v1 config. My batch size is 120 and I am using Azure GPU VM with K80. I am using tensorflow\/models\/research\/object_detection\/model_train.py to train my data and It is taking too long (its been 30+ hours) but since I am a newbie with no ML background I am not sure if is working or not? I dont understand all these tensorboard graphs etc, so I need help to understand what these graphs and images are trying to tell me? And how much time will it take?\nMy tensorboard: <a href=\"http:\/\/danial.eastus.cloudapp.azure.com:9001\/\" rel=\"nofollow noreferrer\">http:\/\/danial.eastus.cloudapp.azure.com:9001\/<\/a><\/p>\n\n<p>Screenshot from my VPN:\n<a href=\"https:\/\/i.stack.imgur.com\/lWqL0.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lWqL0.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/13A0d.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/13A0d.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57934638,"AnswerCount":2,"Tags":"<tensorflow><keras><loss-function>","CreationDate":"2019-09-14T10:31:28.957","AcceptedAnswerId":57944795.0,"OwnerUserId":4028648.0,"Title":"How does TensorFlow\/Keras's class_weight parameter of the fit() function work?","Body":"<p>I do semantic segmentation with TensorFlow 1.12 and Keras. I supply a vector of weights (size equal to the number of classes) to <code>tf.keras.Model.fit()<\/code> using its <code>class_weight<\/code> parameter. I was wondering how this works internally. I use a custom loss function(s) (dice loss and focal loss amongst others), and the weights cannot be premultiplied with the predictions or the one-hot ground truth before being fed to the loss function, since that wouldn't make any sense. My loss function outputs one scalar value, so it also cannot be multiplied with the function output. So where and how exactly as the class weights taken into account?<\/p>\n\n<p>My custom loss function is:<\/p>\n\n<pre><code>def cross_entropy_loss(onehots_true, logits): # Inputs are [BATCH_SIZE, height, width, num_classes]\n    logits, onehots_true = mask_pixels(onehots_true, logits) # Removes pixels for which no ground truth exists, and returns shape [num_gt_pixels, num_classes]\n    return tf.losses.softmax_cross_entropy(onehots_true, logits)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57944795","CreationDate":"2019-09-15T13:57:24.053","ParentId":null,"OwnerUserId":"9828197","Title":null,"Body":"<p>You can refer to the below code from keras source code in github:<\/p>\n\n<pre><code>    class_sample_weight = np.asarray(\n        [class_weight[cls] for cls in y_classes if cls in class_weight])\n\n    if len(class_sample_weight) != len(y_classes):\n      # subtract the sets to pick all missing classes\n      existing_classes = set(y_classes)\n      existing_class_weight = set(class_weight.keys())\n      raise ValueError(\n          '`class_weight` must contain all classes in the data.'\n          ' The classes %s exist in the data but not in '\n          '`class_weight`.' % (existing_classes - existing_class_weight))\n\n  if class_sample_weight is not None and sample_weight is not None:\n    # Multiply weights if both are provided.\n    return class_sample_weight * sample_weight\n<\/code><\/pre>\n\n<p>so as you can see, first <code>class_weight<\/code> is transformed into a numpy array <code>class_sample_weight<\/code> and then it is multiplied with the sample_weight.<\/p>\n\n<p>source: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/keras\/engine\/training_utils.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/keras\/engine\/training_utils.py<\/a> <\/p>\n"},{"AnswerId":"57936450","CreationDate":"2019-09-14T14:29:55.253","ParentId":null,"OwnerUserId":"10878733","Title":null,"Body":"<p>As mentioned in the <a href=\"https:\/\/keras.io\/models\/sequential\/\" rel=\"nofollow noreferrer\">Keras Official Docs<\/a>,<\/p>\n\n<blockquote>\n  <p><code>class_weight<\/code>: Optional dictionary mapping class indices (integers)\n  to a weight (float) value, used for weighting the loss function\n  (during training only). This can be useful to tell the model to \"pay\n  more attention\" to samples from an under-represented class.<\/p>\n<\/blockquote>\n\n<p>Basically, we provide class weights where we have a <strong>class imbalanc<\/strong>e. Meaning, the training samples are not uniformly distributed among all the classes. Some classes have fewer samples whereas some classes have higher samples.<\/p>\n\n<p>We need the classifier to make more <em>attention<\/em> to the classes which are less in number. One way could be to increase the loss value for classes with low samples. A higher loss means higher optimization which results in efficient classification.<\/p>\n\n<p>In terms of Keras, we pass a <code>dict<\/code> mapping class indices to their weights ( <strong>factors by which the loss value will be multiplied<\/strong> ). Let's take an example,<\/p>\n\n<pre><code>class_weights = { 0 : 1.2 , 1 : 0.9 }\n<\/code><\/pre>\n\n<p>Internally, the loss values for classes 0 and 1 will be multiplied by their corresponding weight values.<\/p>\n\n<pre><code>weighed_loss_class0 = loss0 * class_weights[0]\nweighed_loss_class1 = loss1 * class_weights[1]\n<\/code><\/pre>\n\n<p>Now, <code>the weighed_loss_class0<\/code> and <code>weighed_loss_class1<\/code> will be used for backpropagation.<\/p>\n\n<p>See <a href=\"https:\/\/towardsdatascience.com\/handling-imbalanced-datasets-in-deep-learning-f48407a0e758\" rel=\"nofollow noreferrer\">this<\/a> and <a href=\"https:\/\/www.researchgate.net\/post\/In_classification_how_do_i_handle_an_unbalanced_training_set\" rel=\"nofollow noreferrer\">this<\/a>.<\/p>\n"}]}
+{"QuestionId":57934683,"AnswerCount":1,"Tags":"<python><tensorflow><reinforcement-learning><tensorflow.js><tensorflowjs-converter>","CreationDate":"2019-09-14T10:36:48.063","AcceptedAnswerId":57948296.0,"OwnerUserId":4288795.0,"Title":"convert stable-baselines tensorflow model to tensorflowjs","Body":"<p><strong>I am trying to convert the underlying tensorflow model from <a href=\"https:\/\/github.com\/hill-a\/stable-baselines\" rel=\"nofollow noreferrer\">stable-baselines<\/a> to <a href=\"https:\/\/github.com\/tensorflow\/tfjs\" rel=\"nofollow noreferrer\">tensorflowjs<\/a> to be able to use the model on the browser. But I could not make the conversion work<\/strong><\/p>\n\n<p>I followed <a href=\"https:\/\/github.com\/hill-a\/stable-baselines\/issues\/329\" rel=\"nofollow noreferrer\">this github issue<\/a> to create the necessary tensorflow files using the code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def generate_checkpoint_from_model(model, checkpoint_name):  \n        tf.saved_model.simple_save(model.sess, checkpoint_name, inputs={\"obs\": model.act_model.obs_ph}, outputs={\"action\": model.action_ph})\n<\/code><\/pre>\n\n<p>Then I try to transform the model using <a href=\"https:\/\/www.tensorflow.org\/js\/tutorials\/conversion\/import_saved_model\" rel=\"nofollow noreferrer\">tensorflowjs_converter<\/a><\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>tensorflowjs_converter --input_format=tf_saved_model test\/ web_test\n<\/code><\/pre>\n\n<p>However, it gives me the following error:<\/p>\n\n<pre><code>Unable to lift tensor &lt;tf.Tensor 'loss\/action_ph:0' shape=(?,) dtype=int32&gt; because it depends transitively on placeholder &lt;tf.Operation 'loss\/action_ph' type=Placeholder&gt; via at least one path, e.g.: loss\/action_ph (Placeholder)\n<\/code><\/pre>\n\n<p>I created the following <a href=\"https:\/\/colab.research.google.com\/drive\/17g9sPh9I1NDFeHlp3T1VG_tKV6HUFL61\" rel=\"nofollow noreferrer\">colab notebook<\/a> with the error so you can try it.<\/p>\n\n<p><strong>Does anyone knows how to make this conversion work?<\/strong><\/p>\n\n<p>Thank you for the help<\/p>\n","answers":[{"AnswerId":"57948296","CreationDate":"2019-09-15T21:26:44.317","ParentId":null,"OwnerUserId":"4288795","Title":null,"Body":"<p>I posted the question as an <a href=\"https:\/\/github.com\/hill-a\/stable-baselines\/issues\/474\" rel=\"nofollow noreferrer\">issue in stable-baselines<\/a> and they answered. I will copy here as a reference to others:<\/p>\n\n<blockquote>\n  <p>You are trying to save the action placeholder used in PPO training\n  (part of PPO agent), but for inference you only need the trained\n  policy and its placeholders (<code>model.act_model<\/code>). The code on colab runs\n  without errors by changing call to <code>simple_save<\/code> to this:<\/p>\n  \n  <p><code>tf.saved_model.simple_save(model.sess, checkpoint_name,\n  inputs={\"obs\": model.act_model.obs_ph}, outputs={\"action\": model.act_model._policy_proba})<\/code><\/p>\n  \n  <p>The value of _policy_proba <a href=\"https:\/\/github.com\/hill-a\/stable-baselines\/blob\/master\/stable_baselines\/common\/policies.py#L233\" rel=\"nofollow noreferrer\">depends on the environment\/algorithm.<\/a><\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":57934994,"AnswerCount":0,"Tags":"<java><python><keras>","CreationDate":"2019-09-14T11:21:57.800","AcceptedAnswerId":null,"OwnerUserId":5937741.0,"Title":"Keeping keras loaded for evaluation","Body":"<p>I'm trying to get my python application which utilizes keras\/tfgpu to make predictions based on images and return the predictions to a JavaFX application however the load time for the tensorflow backend is quite high compared to the networks evaluation methods so the app feels very slow. <\/p>\n\n<p>I'd do something with networking however it seems a bit overkill seeing as how the apps will always be on the same machine and I don't necessarily want to deal with that whole process. What are the common methods for dealing with this kind of situation?<\/p>\n","answers":[]}
+{"QuestionId":57935189,"AnswerCount":1,"Tags":"<tensorflow><keras><metrics>","CreationDate":"2019-09-14T11:45:30.190","AcceptedAnswerId":57935678.0,"OwnerUserId":5536513.0,"Title":"Custom metric access X input data","Body":"<p>I'd like to write a custom metric for a spelling correction model that counts correctly substituted letters that were previously incorrect. And it should be counted incorrectly substituted letters that were previously correct.<\/p>\n\n<p>That's why I need access to the x_input data. Unfortunately, only y_true and y_pred are accessible by default. Is there a workaround to get to the matching x_input?<\/p>\n\n<p>Is:<\/p>\n\n<pre><code>def custom_metric(y_true, y_pred):\n<\/code><\/pre>\n\n<p>Wanted:<\/p>\n\n<pre><code>def custom_metric(x_input, y_true, y_pred):\n<\/code><\/pre>\n","answers":[{"AnswerId":"57935678","CreationDate":"2019-09-14T12:52:54.783","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<pre><code>def custom_loss(x_input):\n    def loss_fn(y_true, y_pred):\n        # Use your x_input here directly\n        return #Your loss value\n    return loss_fn\n\nmodel = # Define your model\nmodel.compile(loss=custom_loss(x_input))   \n# Values of y_true and y_pred will be passed implicitly by Keras\n<\/code><\/pre>\n\n<p>Remember that <code>x_input<\/code> will be having same values across all batches of input while model is getting trained.<\/p>\n\n<p><strong>EDIT<\/strong>:<\/p>\n\n<p>Since you need <code>x_input<\/code> data <strong>only<\/strong> of every batch for estimating during the loss function and you are having your own custom loss function, why don't you pass the <code>x_input<\/code> as labels. Something like this:<\/p>\n\n<pre><code>model.fit(x=x_input, y=x_input)\nmodel.compile(loss=custom_loss())\n\ndef custom_loss(y_true, y_pred):\n  # y_true corresponds to x_input data\n<\/code><\/pre>\n\n<p>If you need x_input and you need to pass some other data, you can do like this:<\/p>\n\n<pre><code>model.fit(x=x_input, y=[x_input, other_data])\nmodel.compile(loss=custom_loss())\n<\/code><\/pre>\n\n<p>You just need to decouple the data in <code>y_true<\/code> now.<\/p>\n"}]}
+{"QuestionId":57935203,"AnswerCount":0,"Tags":"<python-3.x><numpy><keras>","CreationDate":"2019-09-14T11:47:23.147","AcceptedAnswerId":null,"OwnerUserId":11880411.0,"Title":"How to fix this python error? QA memory network system","Body":"<p>I have QA memory network model after training I have given story and question to the trained model in order to get answer prediction. But I having this error.<\/p>\n\n<p>bAbI task Story: Mary went to the bathroom . John moved to the hallway . Mary travelled to the office . Question: Where is Mary ?<\/p>\n\n<p>Error: x = [word_idx[w] for w in story]<\/p>\n\n<p>KeyError: 'Mary<\/p>\n\n<p><pre><code>\n    def vectorize_stories(data, word_idx, story_maxlen, query_maxlen):\n    X = []\n    Xq = []\n    Y = []\n          for story, query, answer in data:\n     line 56 x = [word_idx[w] for w in story]\n           xq = [word_idx[w] for w in query]\n        # let's not forget that index 0 is reserved\n        y = np.zeros(len(word_idx) + 1)\n        y[word_idx[answer]] = 1\n        X.append(x)\n        Xq.append(xq)\n        Y.append(y)\n        return (pad_sequences(X, maxlen=story_maxlen),\n            pad_sequences(Xq, maxlen=query_maxlen), np.array(Y))<\/p>\n\n<pre><code>   vocab_size = len(vocab) + 1\n   story_maxlen = max(map(len, (x for x, _, _ in train_stories + \n  test_stories)))\n  query_maxlen = max(map(len, (x for _, x, _ in train_stories + \n  test_stories)))\n<\/code><\/pre>\n\n<p>word_idx = dict((c, i + 1) for i, c in enumerate(vocab))\n     idx_word = dict((i+1, c) for i,c in enumerate(vocab))\n     inputs_train, queries_train, answers_train = \n     vectorize_stories(train_stories,\n                                                               word_idx,\n                                                               story_maxlen,\n                                                               query_maxlen)\n     inputs_test, queries_test, answers_test = \n     vectorize_stories(test_stories,\n                                                            word_idx,\n                                                            story_maxlen,\n                                                            query_maxlen)<\/p>\n\n print('Please input a story')\n user_story_inp = input().split(' ')\n print('Please input a query')\n user_query_inp = input().split(' ')\n user_story, user_query, user_ans = vectorize_stories([[user_story_inp, \n user_query_inp, '.']], word_idx, story_maxlen, query_maxlen)\n user_prediction = model.predict([user_story, user_query])\n user_prediction = idx_word[np.argmax(user_prediction)]\n print('Result')\n print(' '.join(user_story_inp), ' '.join(user_query_inp), '| \n Prediction:', user_prediction)\n&lt;\/pre&gt;&lt;\/code&gt;\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57935439,"AnswerCount":1,"Tags":"<tensorflow><keras><keras-layer><dropout>","CreationDate":"2019-09-14T12:22:14.870","AcceptedAnswerId":57935502.0,"OwnerUserId":11713857.0,"Title":"Why does non-zero values change in Keras Dropout?","Body":"<p>Suppose I have a tensor:<\/p>\n\n<pre><code>x = tf.reshape(tf.constant(tf.range(1, 21, dtype=tf.float32)), (5,4))\n\n&lt;tf.Tensor: id=1080557, shape=(5, 4), dtype=float32, numpy=\narray([[ 1.,  2.,  3.,  4.],\n       [ 5.,  6.,  7.,  8.],\n       [ 9., 10., 11., 12.],\n       [13., 14., 15., 16.],\n       [17., 18., 19., 20.]], dtype=float32)&gt;\n<\/code><\/pre>\n\n<p>And I apply dropout to it:<\/p>\n\n<pre><code>dropout = tf.keras.layers.Dropout(0.1)\ndropout(x, training=True)\n\n&lt;tf.Tensor: id=1080704, shape=(5, 4), dtype=float32, numpy=\narray([[ 1.1111112,  2.2222223,  3.3333335,  0.       ],\n       [ 5.555556 ,  6.666667 ,  7.777778 ,  8.888889 ],\n       [10.       , 11.111112 , 12.222223 ,  0.       ],\n       [14.444445 , 15.555556 , 16.666668 , 17.777779 ],\n       [18.88889  ,  0.       , 21.111113 , 22.222223 ]], dtype=float32)&gt;\n<\/code><\/pre>\n\n<p>Each time I run it I have from 1 to 3 zeroed values which is not exactly <code>rate=0.1<\/code>. What range of rates does it actually apply and why did non-zeroed values change?<\/p>\n\n<p>To visualize Celius Stingher answer:<\/p>\n\n<pre><code>l = 10000; r = range(l)\nf = np.zeros((5,4))\nfor i in r:\n  d = dropout(x, training=True)\n  f += d\nf = f\/l\nf\n\n&lt;tf.Tensor: id=1234967, shape=(5, 4), dtype=float32, numpy=\narray([[ 1.0006623,  1.999991 ,  2.988533 ,  4.017763 ],\n       [ 5.000613 ,  6.0477467,  7.0076656,  8.0248575],\n       [ 9.048    , 10.06455  , 10.980609 , 12.010143 ],\n       [12.918334 , 14.100925 , 15.039784 , 16.014153 ],\n       [17.0579   , 18.112    , 19.064175 , 20.024672 ]], dtype=float32)&gt;\n<\/code><\/pre>\n","answers":[{"AnswerId":"57935502","CreationDate":"2019-09-14T12:30:39.587","ParentId":null,"OwnerUserId":"11897007","Title":null,"Body":"<p>Because dropout works in a way that every neuron is set to 0 with a probability equal to the value you are passing. You can think of it as a binomial distribution[*] with <code>p = 0.1<\/code> and <code>n = 20<\/code> the <strong>expected value equals 2<\/strong> and the <strong>Standard deviation equals ~1.34<\/strong> so that explains why most of the time you'll see between 1 and 3 neuron (values) being forced to 0. That is why you can set a random seed within the dropout function to ensure reproducibility<\/p>\n\n<p>[*] In this paper<a href=\"http:\/\/www.jmlr.org\/papers\/volume15\/srivastava14a\/srivastava14a.pdf\" rel=\"nofollow noreferrer\">1<\/a> you find further detail where they assume r(j) follows a Bernoulli distribution (hence the repetition of multiple bernoulli distributions follow a binomial one). <\/p>\n\n<p>OP asked: Thank you. I got you, but what about other values? Why did non-zeroed values change? <\/p>\n\n<p>EDIT: Given how the function works, your model, and given your set of values, the expected value in the long run after applying dropout has to equal the values before applying them. So if you run this code for let's say 1000 iterations, I expect the average of the average value for each run to tend to 10.5 (or 210 total). The only to achieve this is to increase each value by the same rate as the dropout. If you get the <code>worse<\/code> case you'll drop the last 3 numbers, but in the <code>best<\/code> case you'll drop the first 3 the average of both averages is 10.5, the initial average values. The interpretation is from the paper I linked. (P.1933 )<\/p>\n"}]}
+{"QuestionId":57935576,"AnswerCount":2,"Tags":"<python><keras><deep-learning><image-recognition><transfer-learning>","CreationDate":"2019-09-14T12:39:22.980","AcceptedAnswerId":57936560.0,"OwnerUserId":9149062.0,"Title":"Is it normal that transfer learning (VGG16) performs worse on CIFAR-10?","Body":"<p><strong>Note:<\/strong> I am not sure this is the right website to ask these kind of questions. Please tell me where I should ask them before downvoting this \"because this isn't the right place to ask\". Thanks!<\/p>\n\n<p>I am currently experimenting with deep learning using <a href=\"https:\/\/keras.io\" rel=\"nofollow noreferrer\">Keras<\/a>. I tried already a model similar to the one to be found on <a href=\"https:\/\/keras.io\/examples\/cifar10_cnn\/\" rel=\"nofollow noreferrer\">the Keras example<\/a>. This yields expecting results:<\/p>\n\n<ul>\n<li>80% after 10-15 epochs without data augmentation before overfitting around the 15th epoch and<\/li>\n<li>80% after 50 epochs with data augmentation without any signs of overfitting.<\/li>\n<\/ul>\n\n<p>After this I wanted to try transfer learning. I did this by using the <a href=\"https:\/\/keras.io\/applications\/#vgg16\" rel=\"nofollow noreferrer\">VGG16<\/a> network without retraining its weights (see code below). This gave very poor results: 63% accuracy after 10 epochs with a very shallow curve (see picture below) which seems to be indicating that it will achieve acceptable results only (if ever) after a very long training time (I would expect 200-300 epochs before it reaches 80%).<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/MWmsT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/MWmsT.png\" alt=\"Transfer learning results\"><\/a><\/p>\n\n<p>Is this normal behavior for this kind of application? Here are a few things I could imagine to be the cause of these bad results:<\/p>\n\n<ul>\n<li>the CIFAR-10 dataset has images of <code>32x32<\/code> pixels, which might be too few for the VGG16 net<\/li>\n<li>The filters of VGG16 are not good for CIFAR-10, which would be solved by setting the weights to <code>trainable<\/code> or by starting with random weights (only copying the model and not the weights)<\/li>\n<\/ul>\n\n<p>Thanks in advance!<\/p>\n\n<hr>\n\n<p><strong>My code:<\/strong><\/p>\n\n<p><em>Note that the inputs are 2 datasets (50000 training images and 10000 testing images) which are labeled images with shape <code>32x32x3<\/code>. Each pixel value is a float in the range <code>[0.0, 1.0]<\/code>.<\/em><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import keras\n\n# load and preprocess data...\n\n# get VGG16 base model and define new input shape\nvgg16 = keras.applications.vgg16.VGG16(input_shape=(32, 32, 3),\n                                       weights='imagenet',\n                                       include_top=False)\n\n# add new dense layers at the top\nx = keras.layers.Flatten()(vgg16.output)\nx = keras.layers.Dense(1024, activation='relu')(x)\nx = keras.layers.Dropout(0.5)(x)\nx = keras.layers.Dense(128, activation='relu')(x)\npredictions = keras.layers.Dense(10, activation='softmax')(x)\n\n# define and compile model\nmodel = keras.Model(inputs=vgg16.inputs, outputs=predictions)\nfor layer in vgg16.layers:\n    layer.trainable = False\nmodel.compile(optimizer='adam',\n              loss='categorical_crossentropy',\n              metrics=['accuracy'])\n\n# training and validation\nmodel.fit(x_train, y_train,\n          batch_size=256,\n          epochs=10,\n          validation_data=(x_test, y_test))\n\nmodel.evaluate(x_test, y_test)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57936560","CreationDate":"2019-09-14T14:44:26.003","ParentId":null,"OwnerUserId":"2624532","Title":null,"Body":"<p>It\u2019s not that the VGG16 model doesn\u2019t work on that input size it\u2019s that the weights you\u2019re using have been pre-trained on a different input size (imagenet). You need your source and target dataset to have the same input size so the pre-trained weights can transfer. So you could either do pre-training with rescaled imagenet images to 32x32x3 or pick a target dataset that is roughly the same as the pre-training was done on (often 224x224x3 or similar for imagenet) and scale to match. I have seen a paper recently where they transferred from imagenet to cifar10 and 100, by up scaling the latter, which worked reasonably well, but that wouldn\u2019t be ideal. <\/p>\n\n<p><a href=\"http:\/\/openaccess.thecvf.com\/content_CVPR_2019\/papers\/Kornblith_Do_Better_ImageNet_Models_Transfer_Better_CVPR_2019_paper.pdf\" rel=\"nofollow noreferrer\">http:\/\/openaccess.thecvf.com\/content_CVPR_2019\/papers\/Kornblith_Do_Better_ImageNet_Models_Transfer_Better_CVPR_2019_paper.pdf<\/a><\/p>\n\n<p>Second with a target dataset with that many training examples freezing all the layers transferred is unlikely to be a good solution in fact setting all the layers to trainable will probably work best.<\/p>\n"},{"AnswerId":"57935822","CreationDate":"2019-09-14T13:11:22.167","ParentId":null,"OwnerUserId":"12067060","Title":null,"Body":"<p>i think that the CIFAR-10 dataset has images of 32x32 pixels, which might be too few for the VGG16 net\nthe program is not compatible for CIFAR-10 to use<\/p>\n"}]}
+{"QuestionId":57935766,"AnswerCount":0,"Tags":"<python><machine-learning><keras>","CreationDate":"2019-09-14T13:04:06.043","AcceptedAnswerId":null,"OwnerUserId":9007012.0,"Title":"Keras unable to create dataset from array","Body":"<p>I am trying to fit a dataset through my network. My network has 2 inputs and 3 outputs. \nI create an array to hold these generated data points <\/p>\n\n<pre><code>xs.append(current_state) # this having a shape of [2]\nys.append(action_pred) # this having a shape of [3]\n<\/code><\/pre>\n\n<p>Then I try to fit those arrays through the network using the fit function :<\/p>\n\n<pre><code>xs = np.array(xs, dtype='float')\nys = np.array(ys, dtype='float')\nmodel.fit(x=xs, y=ys, batch_size=200)\n<\/code><\/pre>\n\n<p>But I get the error <\/p>\n\n<pre><code>expected dense_1_input to have 2 dimensions, but got array with shape (200, 1, 2)\n<\/code><\/pre>\n\n<p>Help me understand what concept I am not understanding. Why doesn't it slice it up into 200 single data points and train on it?\nHere is the network:<\/p>\n\n<pre><code>model.add(Dense(4, input_shape=(2,)))\nmodel.add(Dense(3))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57935919,"AnswerCount":0,"Tags":"<deep-learning><pytorch><recurrent-neural-network>","CreationDate":"2019-09-14T13:22:43.573","AcceptedAnswerId":null,"OwnerUserId":10782229.0,"Title":"How to change a CRNN model based on input size in pytorch?","Body":"<p>I am fairly new to deep learning and pytorch. I am trying to develop a word based OCR for bangla words following <a href=\"https:\/\/github.com\/meijieru\/crnn.pytorch\" rel=\"nofollow noreferrer\">this<\/a> code. Here they resize all images to 32,100 size. As my images are of different size and quite larger ( 90,752 ) I can't resize them to 32,100. Rather I am padding all my images to have the maximum height and width of my data set in the data loader. In that case the model won't work as it requires the height to come down to 1 after CNN layers. <\/p>\n\n<p>So my question is , how do I change the model?<\/p>\n","answers":[]}
+{"QuestionId":57935988,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-14T13:30:42.690","AcceptedAnswerId":57936161.0,"OwnerUserId":10895862.0,"Title":"UnboundLocalError: local variable 'batch_index' referenced before assignment","Body":"<p>This is NOT MY code by here is the line, where it shows a problem:<\/p>\n\n<p>model.fit(trainX, trainY, batch_size=2, epochs=200, verbose=2)<\/p>\n\n<p>(As I am thinking now, it is very possible this code uses an older version of TF, because 'epochs' was written as 'nb_epoch').<\/p>\n\n<p>The last update of the code is from: Jan 11, 2017!<\/p>\n\n<p>I have tried everything from the internet (which is not so much), including looking inside the source code of tensorflow\/keras for some hints. Just to make it clear that I don't have a variable, called 'batch_index' inside the code.<\/p>\n\n<p>So far I have looked inside different versions of TF (tensorflow\/tensorflow\/python\/keras\/engine\/training_arrays.py). It appears that all are from 2018 copyright, but some start with the function fit_loop, and other with model_iteration (which is probably an update of fit_loop).<\/p>\n\n<p>So, this 'batch_index' variable can be seen only in the first function. <\/p>\n\n<p>I wonder if I am going in the right direction at all??!<\/p>\n\n<p>There is no point in showing the code, because, as I explained, there is no such variable in the first place inside the code.<\/p>\n\n<p>but, here is some code of the function 'stock_prediction', that gives the error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\ndef stock_prediction():\n\n    # Collect data points from csv\n    dataset = []\n\n    with open(FILE_NAME) as f:\n        for n, line in enumerate(f):\n            if n != 0:\n                dataset.append(float(line.split(',')[1]))\n\n    dataset = np.array(dataset)\n\n    # Create dataset matrix (X=t and Y=t+1)\n    def create_dataset(dataset):\n        dataX = [dataset[n+1] for n in range(len(dataset)-2)]\n        return np.array(dataX), dataset[2:]\n\n    trainX, trainY = create_dataset(dataset)\n\n    # Create and fit Multilinear Perceptron model\n    model = Sequential()\n    model.add(Dense(8, input_dim=1, activation='relu'))\n    model.add(Dense(1))\n    model.compile(loss='mean_squared_error', optimizer='adam')\n    model.fit(trainX, trainY, nb_epoch=200, batch_size=2, verbose=2)\n\n    # Our prediction for tomorrow\n    prediction = model.predict(np.array([dataset[0]]))\n    result = 'The price will move from %s to %s' % (dataset[0], prediction[0][0])\n\n    return result\n\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>\n---------------------------------------------------------------------------\nUnboundLocalError                         Traceback (most recent call last)\n&lt;ipython-input-19-3dde95909d6e&gt; in &lt;module&gt;\n     14 \n     15 # We have our file so we create the neural net and get the prediction\n---&gt; 16 print(stock_prediction())\n     17 \n     18 # We are done so we delete the csv file\n\n&lt;ipython-input-18-8bbf4f61c738&gt; in stock_prediction()\n     23     model.add(Dense(1))\n     24     model.compile(loss='mean_squared_error', optimizer='adam')\n---&gt; 25     model.fit(trainX, trainY, batch_size=1, epochs=200, verbose=2)\n     26 \n     27     # Our prediction for tomorrow\n\n~\\Anaconda3\\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, validation_freq, max_queue_size, workers, use_multiprocessing, **kwargs)\n   1176                                         steps_per_epoch=steps_per_epoch,\n   1177                                         validation_steps=validation_steps,\n-&gt; 1178                                         validation_freq=validation_freq)\n   1179 \n   1180     def evaluate(self,\n\n~\\Anaconda3\\lib\\site-packages\\keras\\engine\\training_arrays.py in fit_loop(model, fit_function, fit_inputs, out_labels, batch_size, epochs, verbose, callbacks, val_function, val_inputs, shuffle, callback_metrics, initial_epoch, steps_per_epoch, validation_steps, validation_freq)\n    211                     break\n    212 \n--&gt; 213             if batch_index == len(batches) - 1:  # Last batch.\n    214                 if do_validation and should_run_validation(validation_freq, epoch):\n    215                     val_outs = test_loop(model, val_function, val_inputs,\n\nUnboundLocalError: local variable 'batch_index' referenced before assignment\n<\/code><\/pre>\n\n<p>A little clarification:<\/p>\n\n<p>I tried to see my version of tf\/keras and here is it:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.python import keras\nprint(keras.__version__)\nimport keras\nprint(keras.__version__)\nimport tensorflow\nprint(tensorflow.__version__)\n<\/code><\/pre>\n\n<p>2.2.4-tf<\/p>\n\n<p>2.2.5<\/p>\n\n<p>1.14.0<\/p>\n\n<p>Why keras shows different versions??<\/p>\n","answers":[{"AnswerId":"57997864","CreationDate":"2019-09-18T17:16:51.027","ParentId":null,"OwnerUserId":"10895862","Title":null,"Body":"<h3>The problem was resolved!<\/h3>\n\n<p>I had to import the correct libraries (Tensorflow and not Keras directly):<\/p>\n\n<h3>instead of importing directly Keras:<\/h3>\n\n<pre><code>from tensorflow.python import keras.models.Sequential\nfrom tensorflow.python import keras.layers.Dense\n<\/code><\/pre>\n\n<h3>importing of only tensorflow works:<\/h3>\n\n<pre><code>from tensorflow.python.keras.layers import Input, Dense\nfrom tensorflow.python.keras.models import Sequential\n<\/code><\/pre>\n\n<p>Apparently this is related to the different version issue of Keras.<\/p>\n"},{"AnswerId":"57936161","CreationDate":"2019-09-14T13:53:08.750","ParentId":null,"OwnerUserId":"3653343","Title":null,"Body":"<p>I checked in the <code>training_arrays.py<\/code> (<a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/engine\/training_arrays.py\" rel=\"nofollow noreferrer\">here<\/a>) the function in which you got the error and, as I can see, I think the problem could be in these statements (from lines 177 - 205):<\/p>\n\n<pre><code>batches = make_batches(num_train_samples, batch_size)\nfor batch_index, (batch_start, batch_end) in enumerate(batches): # the problem is here\n    # do stuff\n    ...\nif batch_index == len(batches) - 1:\n    # do stuff\n    ...\n<\/code><\/pre>\n\n<p>If batches is an <em>empty list,<\/em> you could get this error. Could be that your training set has some problem?<\/p>\n"}]}
+{"QuestionId":57936004,"AnswerCount":1,"Tags":"<python-3.x><keras><reinforcement-learning><openai-gym>","CreationDate":"2019-09-14T13:32:55.820","AcceptedAnswerId":null,"OwnerUserId":12067106.0,"Title":"Why my network returns an integer larger than the output space length","Body":"<p>I am learning Deep-Q-Network and have tried changing the code at <a href=\"https:\/\/github.com\/keon\/deep-q-learning\/blob\/master\/dqn.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keon\/deep-q-learning\/blob\/master\/dqn.py<\/a> to try out atari breakout.\nBut the output of my network is sometimes larger than the output space.\nThere was no problem with the original author's code, but after changing the code, there was a problem.<\/p>\n\n<p>I'm running python 3.7 on Windows 10 using Numpy, keras-gpu, and running on AMD Ryzen 3700x and rtx 2070 super.\nI thought it was a neural network model problem, but I can't do anything because I didn't know how to solve it.<\/p>\n\n<pre><code>class DQNAgent:\n    def __init__(self, state_size, action_size):\n        self.state_size = state_size\n        self.action_size = action_size\n        self.memory = deque(maxlen=2000)\n        self.gamma = 0.95    # discount rate\n        self.epsilon = 1.0  # exploration rate\n        self.epsilon_min = 0.01\n        self.epsilon_decay = 0.995\n        self.learning_rate = 0.001\n        self.model = self._build_model()\n\n    def _build_model(self):\n        # Neural Net for Deep-Q learning Model\n        model = Sequential()\n        model.add(Conv2D(32, kernel_size=(5, 5), input_shape=(105, 80, 1), \nactivation='relu'))\n        model.add(Dense(24, input_dim=self.state_size, activation='relu'))\n        model.add(Dense(24, activation='relu'))\n        model.add(Dense(self.action_size, activation='linear'))\n        model.compile(loss='mse',\n                      optimizer=Adam(lr=self.learning_rate))\n        return model\n\n    def remember(self, state, action, reward, next_state, done):\n        self.memory.append((state, action, reward, next_state, done))\n\n    def act(self, state):\n        if np.random.rand() &lt;= self.epsilon:\n            return random.randrange(self.action_size)\n        state = np.expand_dims(state, 3)\n        state = np.expand_dims(state, 0)\n        act_values = self.model.predict(state)\n        return np.argmax(act_values[0])\n\n    def replay(self, batch_size):\n        minibatch = random.sample(self.memory, batch_size)\n        for state, action, reward, next_state, done in minibatch:\n            target = reward\n            if not done:\n                next_state = np.expand_dims(next_state, 3)\n                next_state = np.expand_dims(next_state, 0)\n                state = np.expand_dims(state, 3)\n                state = np.expand_dims(state, 0)\n                target = (reward + self.gamma *\n                      np.amax(self.model.predict(next_state)[0]))\n            target_f = self.model.predict(state)\n            target_f[0][action] = target\n            self.model.fit(state, target_f, epochs=1, verbose=0)\n        if self.epsilon &gt; self.epsilon_min:\n            self.epsilon *= self.epsilon_decay\n<\/code><\/pre>\n\n<p>Expect output 0 ~ 3 of act() but sometimes act() returns more than 2000.<\/p>\n","answers":[{"AnswerId":"57936900","CreationDate":"2019-09-14T15:30:26.377","ParentId":null,"OwnerUserId":"9437380","Title":null,"Body":"<p>The output size of your neural network is dependent from the...<\/p>\n\n<p>self.action_space = action_space <\/p>\n\n<p>...line in <strong>init<\/strong>(...). What are you passing into the second argument when initializing the Agent? This is what you are setting as the output size of the neural network.<\/p>\n"}]}
+{"QuestionId":57936175,"AnswerCount":2,"Tags":"<python><tensorflow><scikit-learn><neural-network><deep-learning>","CreationDate":"2019-09-14T13:54:43.580","AcceptedAnswerId":null,"OwnerUserId":9706484.0,"Title":"Neural network doesn't appear to learn","Body":"<p>I am trying to implement a neural network for the heart dataset from kaggle<\/p>\n\n<p><a href=\"https:\/\/www.kaggle.com\/ronitf\/heart-disease-uci\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/ronitf\/heart-disease-uci<\/a><\/p>\n\n<p>Even though I am training the network on the entire dataset it doesn't seem to learn at all with the accuracy staying around 50% , even if it is overfitting a Neural network should be able to get higher in sample accuracy. Can anyone see any issues with this code (i've tried many learning rates , batches and epochs and nothing seems to work.<\/p>\n\n<pre><code>#need to switch from the binary output to the singular value output\n\nimport pandas\nimport numpy as np\nimport tensorflow as tf\nimport random\nfrom sklearn.preprocessing import normalize\n\n\n\n\n#pulling in data from excel\nsheet = pandas.read_excel(\".\/heart.xlsx\")\nsheet2=sheet.values\n\n#getting my data arrays\nX_data,y_data = np.split(sheet2, [13], axis=1)\ny_data=np.reshape(y_data, [-1])   #need to learn more about why this works\n\nX_data =  normalize(X_data, axis=0, norm='max')\n\n\n\n#random shuffle\n\nind_list = [i for i in range(len(X_data))]\nrandom.shuffle(ind_list)\n\n\nX_data=X_data[ind_list]\ny_data=y_data[ind_list]\n\n\n\n#initialising values\nepochs = 1000\nlearning_rate=0.001\nbatch_size=50\n\n\nX = tf.placeholder(tf.float32,shape=(None,n_inputs),name=\"X\")\ny = tf.placeholder(tf.int64,shape=(None) ,name=\"y\")\n\n\n\n#creating the structure of the neural network\n\n\nwith tf.name_scope(\"dnn\"):\n    input = tf.layers.dense(X,13,name=\"hidden1\" ,activation=tf.nn.relu)\n    input = tf.layers.dense(input,7,name=\"hidden2\" ,activation=tf.nn.relu)\n    logits = tf.layers.dense(input,1,name=\"outputs\")\n\n\n\nwith tf.name_scope(\"loss\"):\n    entropy = tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.cast(y,tf.float32),logits=logits)\n    loss=tf.reduce_mean(entropy,name=\"loss\")\n\nwith tf.name_scope(\"train\"):\n    optimizer=tf.train.AdamOptimizer(learning_rate)\n    training_operation=optimizer.minimize(loss)\n\nwith tf.name_scope(\"accuracy\"):\n    predicted = tf.nn.sigmoid(logits)\n    correct_pred = tf.equal(tf.round(predicted), tf.cast(y,tf.float32))\n    accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))\n\n\n#defining functions\n\ndef shuffle_batch(X, y, batch_size):\n    rnd_idx = np.random.permutation(len(X))\n    n_batches = len(X) \/\/ batch_size\n    for batch_idx in np.array_split(rnd_idx, n_batches):\n        X_batch, y_batch = X[batch_idx], y[batch_idx]\n        yield X_batch, y_batch\n\n\n\n#initialising the graph and running the optimisation\n\ninit= tf.global_variables_initializer()\nsave = tf.train.Saver()\n\n\n\nwith tf.Session() as sess:\n    init.run()\n\n    for epoch in range(epochs):\n\n        for X_batch,y_batch in shuffle_batch(X_data,y_data,batch_size):\n            sess.run(training_operation,feed_dict={X:X_batch,y:y_batch})\n\n        acc_train=accuracy.eval(feed_dict={X:X_data,y:y_data})\n\n        print(\"Epoch\",epoch,\"training acuuracy\",acc_train)\n<\/code><\/pre>\n\n<p>output<\/p>\n\n<pre><code>2019-09-15 07:52:04.662488: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2\nEpoch 0 training acuuracy 0.5089697\nEpoch 1 training acuuracy 0.5089697\nEpoch 2 training acuuracy 0.5089697\nEpoch 3 training acuuracy 0.50838155\nEpoch 4 training acuuracy 0.5089697\nEpoch 5 training acuuracy 0.5092638\nEpoch 6 training acuuracy 0.5101461\nEpoch 7 training acuuracy 0.5101461\nEpoch 8 training acuuracy 0.5101461\nEpoch 9 training acuuracy 0.51073426\nEpoch 10 training acuuracy 0.51102835\nEpoch 11 training acuuracy 0.51102835\nEpoch 12 training acuuracy 0.51073426\nEpoch 13 training acuuracy 0.51132244\nEpoch 14 training acuuracy 0.5122047\nEpoch 15 training acuuracy 0.5124988\nEpoch 16 training acuuracy 0.5127929\nEpoch 17 training acuuracy 0.513087\nEpoch 18 training acuuracy 0.51338106\nEpoch 19 training acuuracy 0.51338106\nEpoch 20 training acuuracy 0.51367515\nEpoch 21 training acuuracy 0.51396924\nEpoch 22 training acuuracy 0.5151456\nEpoch 23 training acuuracy 0.51426333\nEpoch 24 training acuuracy 0.5154397\nEpoch 25 training acuuracy 0.5151456\nEpoch 26 training acuuracy 0.51632196\nEpoch 27 training acuuracy 0.5172042\nEpoch 28 training acuuracy 0.51808643\nEpoch 29 training acuuracy 0.5183805\nEpoch 30 training acuuracy 0.5192628\nEpoch 31 training acuuracy 0.5189687\nEpoch 32 training acuuracy 0.51985097\nEpoch 33 training acuuracy 0.52043915\nEpoch 34 training acuuracy 0.52043915\nEpoch 35 training acuuracy 0.52043915\nEpoch 36 training acuuracy 0.5213214\nEpoch 37 training acuuracy 0.5219096\nEpoch 38 training acuuracy 0.5222037\nEpoch 39 training acuuracy 0.5222037\nEpoch 40 training acuuracy 0.5224978\nEpoch 41 training acuuracy 0.5224978\nEpoch 42 training acuuracy 0.5224978\nEpoch 43 training acuuracy 0.5222037\nEpoch 44 training acuuracy 0.5219096\nEpoch 45 training acuuracy 0.5222037\nEpoch 46 training acuuracy 0.5222037\nEpoch 47 training acuuracy 0.5219096\nEpoch 48 training acuuracy 0.5222037\nEpoch 49 training acuuracy 0.52279186\nEpoch 50 training acuuracy 0.52308595\nEpoch 51 training acuuracy 0.52338004\nEpoch 52 training acuuracy 0.52308595\nEpoch 53 training acuuracy 0.52308595\nEpoch 54 training acuuracy 0.52308595\nEpoch 55 training acuuracy 0.52367413\nEpoch 56 training acuuracy 0.5245564\nEpoch 57 training acuuracy 0.52602684\nEpoch 58 training acuuracy 0.52543867\nEpoch 59 training acuuracy 0.52573276\nEpoch 60 training acuuracy 0.526615\nEpoch 61 training acuuracy 0.52573276\nEpoch 62 training acuuracy 0.5269091\nEpoch 63 training acuuracy 0.5269091\nEpoch 64 training acuuracy 0.52602684\nEpoch 65 training acuuracy 0.526615\nEpoch 66 training acuuracy 0.5277914\nEpoch 67 training acuuracy 0.5272032\nEpoch 68 training acuuracy 0.526615\nEpoch 69 training acuuracy 0.526615\nEpoch 70 training acuuracy 0.526615\nEpoch 71 training acuuracy 0.5269091\nEpoch 72 training acuuracy 0.526615\nEpoch 73 training acuuracy 0.526615\nEpoch 74 training acuuracy 0.5277914\nEpoch 75 training acuuracy 0.52867365\nEpoch 76 training acuuracy 0.5277914\nEpoch 77 training acuuracy 0.5277914\nEpoch 78 training acuuracy 0.52837956\nEpoch 79 training acuuracy 0.52837956\nEpoch 80 training acuuracy 0.52837956\nEpoch 81 training acuuracy 0.5295559\nEpoch 82 training acuuracy 0.52985\nEpoch 83 training acuuracy 0.5304382\nEpoch 84 training acuuracy 0.5301441\nEpoch 85 training acuuracy 0.5301441\nEpoch 86 training acuuracy 0.5307323\nEpoch 87 training acuuracy 0.5304382\nEpoch 88 training acuuracy 0.5304382\nEpoch 89 training acuuracy 0.53132045\nEpoch 90 training acuuracy 0.5322027\nEpoch 91 training acuuracy 0.5322027\nEpoch 92 training acuuracy 0.5322027\nEpoch 93 training acuuracy 0.533085\nEpoch 94 training acuuracy 0.5348495\nEpoch 95 training acuuracy 0.53455544\nEpoch 96 training acuuracy 0.5333791\nEpoch 97 training acuuracy 0.5333791\nEpoch 98 training acuuracy 0.53426135\nEpoch 99 training acuuracy 0.53455544\nEpoch 100 training acuuracy 0.53455544\nEpoch 101 training acuuracy 0.5354377\nEpoch 102 training acuuracy 0.5351436\nEpoch 103 training acuuracy 0.53455544\nEpoch 104 training acuuracy 0.5348495\nEpoch 105 training acuuracy 0.53396726\nEpoch 106 training acuuracy 0.53455544\nEpoch 107 training acuuracy 0.53426135\nEpoch 108 training acuuracy 0.5351436\nEpoch 109 training acuuracy 0.5354377\nEpoch 110 training acuuracy 0.5354377\nEpoch 111 training acuuracy 0.53690815\nEpoch 112 training acuuracy 0.53720224\nEpoch 113 training acuuracy 0.5377904\nEpoch 114 training acuuracy 0.5380845\nEpoch 115 training acuuracy 0.5380845\nEpoch 116 training acuuracy 0.5377904\nEpoch 117 training acuuracy 0.5377904\nEpoch 118 training acuuracy 0.53720224\nEpoch 119 training acuuracy 0.5377904\nEpoch 120 training acuuracy 0.5377904\nEpoch 121 training acuuracy 0.5377904\nEpoch 122 training acuuracy 0.5383786\nEpoch 123 training acuuracy 0.5383786\nEpoch 124 training acuuracy 0.5383786\nEpoch 125 training acuuracy 0.5386727\nEpoch 126 training acuuracy 0.5383786\nEpoch 127 training acuuracy 0.5383786\nEpoch 128 training acuuracy 0.5383786\nEpoch 129 training acuuracy 0.5383786\nEpoch 130 training acuuracy 0.5389668\nEpoch 131 training acuuracy 0.5386727\nEpoch 132 training acuuracy 0.5389668\nEpoch 133 training acuuracy 0.53926086\nEpoch 134 training acuuracy 0.53926086\nEpoch 135 training acuuracy 0.53926086\nEpoch 136 training acuuracy 0.53955495\nEpoch 137 training acuuracy 0.53955495\nEpoch 138 training acuuracy 0.53984904\nEpoch 139 training acuuracy 0.53984904\nEpoch 140 training acuuracy 0.53984904\nEpoch 141 training acuuracy 0.54014313\nEpoch 142 training acuuracy 0.5404372\nEpoch 143 training acuuracy 0.54014313\nEpoch 144 training acuuracy 0.53955495\nEpoch 145 training acuuracy 0.5404372\nEpoch 146 training acuuracy 0.54014313\nEpoch 147 training acuuracy 0.54014313\nEpoch 148 training acuuracy 0.53984904\nEpoch 149 training acuuracy 0.53955495\nEpoch 150 training acuuracy 0.54014313\nEpoch 151 training acuuracy 0.5404372\nEpoch 152 training acuuracy 0.5407313\nEpoch 153 training acuuracy 0.5410254\nEpoch 154 training acuuracy 0.5410254\nEpoch 155 training acuuracy 0.5413195\nEpoch 156 training acuuracy 0.5413195\nEpoch 157 training acuuracy 0.5410254\nEpoch 158 training acuuracy 0.5410254\nEpoch 159 training acuuracy 0.5410254\nEpoch 160 training acuuracy 0.5413195\nEpoch 161 training acuuracy 0.5413195\nEpoch 162 training acuuracy 0.5413195\nEpoch 163 training acuuracy 0.5413195\nEpoch 164 training acuuracy 0.5410254\nEpoch 165 training acuuracy 0.5413195\nEpoch 166 training acuuracy 0.54190767\nEpoch 167 training acuuracy 0.54220176\nEpoch 168 training acuuracy 0.54190767\nEpoch 169 training acuuracy 0.54220176\nEpoch 170 training acuuracy 0.54220176\nEpoch 171 training acuuracy 0.54220176\nEpoch 172 training acuuracy 0.54220176\nEpoch 173 training acuuracy 0.54220176\nEpoch 174 training acuuracy 0.54220176\nEpoch 175 training acuuracy 0.54220176\nEpoch 176 training acuuracy 0.54190767\nEpoch 177 training acuuracy 0.5416136\nEpoch 178 training acuuracy 0.5416136\nEpoch 179 training acuuracy 0.54190767\nEpoch 180 training acuuracy 0.54220176\nEpoch 181 training acuuracy 0.54220176\nEpoch 182 training acuuracy 0.54220176\nEpoch 183 training acuuracy 0.54220176\nEpoch 184 training acuuracy 0.54220176\nEpoch 185 training acuuracy 0.54220176\nEpoch 186 training acuuracy 0.54220176\nEpoch 187 training acuuracy 0.54220176\nEpoch 188 training acuuracy 0.54220176\nEpoch 189 training acuuracy 0.54220176\nEpoch 190 training acuuracy 0.54220176\nEpoch 191 training acuuracy 0.54220176\nEpoch 192 training acuuracy 0.54220176\nEpoch 193 training acuuracy 0.54220176\nEpoch 194 training acuuracy 0.54278994\nEpoch 195 training acuuracy 0.543084\nEpoch 196 training acuuracy 0.543084\nEpoch 197 training acuuracy 0.543084\nEpoch 198 training acuuracy 0.543084\nEpoch 199 training acuuracy 0.543084\nEpoch 200 training acuuracy 0.54278994\nEpoch 201 training acuuracy 0.54249585\nEpoch 202 training acuuracy 0.54220176\nEpoch 203 training acuuracy 0.54249585\nEpoch 204 training acuuracy 0.54220176\nEpoch 205 training acuuracy 0.54220176\nEpoch 206 training acuuracy 0.54249585\nEpoch 207 training acuuracy 0.54249585\nEpoch 208 training acuuracy 0.54249585\nEpoch 209 training acuuracy 0.54249585\nEpoch 210 training acuuracy 0.54249585\nEpoch 211 training acuuracy 0.54249585\nEpoch 212 training acuuracy 0.543084\nEpoch 213 training acuuracy 0.543084\nEpoch 214 training acuuracy 0.54278994\nEpoch 215 training acuuracy 0.54249585\nEpoch 216 training acuuracy 0.54278994\nEpoch 217 training acuuracy 0.543084\nEpoch 218 training acuuracy 0.54278994\nEpoch 219 training acuuracy 0.54278994\nEpoch 220 training acuuracy 0.54278994\nEpoch 221 training acuuracy 0.54278994\nEpoch 222 training acuuracy 0.54278994\nEpoch 223 training acuuracy 0.54249585\nEpoch 224 training acuuracy 0.54249585\nEpoch 225 training acuuracy 0.54249585\nEpoch 226 training acuuracy 0.54278994\nEpoch 227 training acuuracy 0.54278994\nEpoch 228 training acuuracy 0.54278994\nEpoch 229 training acuuracy 0.54278994\nEpoch 230 training acuuracy 0.54278994\nEpoch 231 training acuuracy 0.54278994\nEpoch 232 training acuuracy 0.54249585\nEpoch 233 training acuuracy 0.54249585\nEpoch 234 training acuuracy 0.54278994\nEpoch 235 training acuuracy 0.543084\nEpoch 236 training acuuracy 0.543084\nEpoch 237 training acuuracy 0.5433781\nEpoch 238 training acuuracy 0.543084\nEpoch 239 training acuuracy 0.5433781\nEpoch 240 training acuuracy 0.5433781\nEpoch 241 training acuuracy 0.5436722\nEpoch 242 training acuuracy 0.5433781\nEpoch 243 training acuuracy 0.543084\nEpoch 244 training acuuracy 0.543084\nEpoch 245 training acuuracy 0.543084\nEpoch 246 training acuuracy 0.543084\nEpoch 247 training acuuracy 0.543084\nEpoch 248 training acuuracy 0.543084\n<\/code><\/pre>\n","answers":[{"AnswerId":"57938058","CreationDate":"2019-09-14T18:08:59.137","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>I am posting only the parts of the code I have changed.<\/p>\n\n<pre><code>X = tf.placeholder(tf.float32,shape=(None,n_inputs),name=\"X\")\ny = tf.placeholder(tf.int32,shape=(None) ,name=\"y\")\n\n#creating the structure of the neural network\nwith tf.name_scope(\"dnn\"):\n    input = tf.layers.dense(X,13,name=\"hidden1\" ,activation=tf.nn.relu)\n    input = tf.layers.dense(input,7,name=\"hidden2\" ,activation=tf.nn.relu)\n    logits = tf.layers.dense(input,2,name=\"outputs\", activation=tf.nn.softmax)\n\nwith tf.name_scope(\"loss\"):\n    entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y,logits=logits)\n    loss=tf.reduce_mean(entropy,name=\"loss\")\n\nwith tf.name_scope(\"train\"):\n    optimizer=tf.train.AdamOptimizer(learning_rate)\n    training_operation=optimizer.minimize(loss)\n\nwith tf.name_scope(\"accuracy\"):\n    predicted = tf.argmax(logits, axis=-1, output_type=tf.int32)\n    correct_pred = tf.equal(predicted, y)\n    accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))\n<\/code><\/pre>\n\n<p>You will get 0.953 accuracy now.<\/p>\n\n<p>Convert your sigmoid loss to softmax loss as they are mutually exclusive class output values.<\/p>\n"},{"AnswerId":"58018290","CreationDate":"2019-09-19T20:06:29.257","ParentId":null,"OwnerUserId":"10675934","Title":null,"Body":"<p>The code below will learn with a single output. The sigmoid loss doesn't impose mutual exclusivity but it's straightforward to write a binary ce loss as below if you want to stick with the single output structure. I don't know if that was the reason your code didn't learn as I'd expect some learning even with the sigmoid loss - possibly the cast operator is the culprit.<\/p>\n\n<pre><code>X = tf.placeholder(tf.float32,shape=(None,n_inputs),name=\"X\")\ny = tf.placeholder(tf.float32,shape=(None) ,name=\"y\")\n\n#creating the structure of the neural network\nwith tf.name_scope(\"dnn\"):\n    input = tf.layers.dense(X,13,name=\"hidden1\" ,activation=tf.nn.relu)\n    input = tf.layers.dense(input,7,name=\"hidden2\" ,activation=tf.nn.relu)\n    logits = tf.layers.dense(input,1,name=\"outputs\", activation=tf.nn.sigmoid)\n\n\nwith tf.name_scope(\"loss\"):\n    entropy = -tf.expand_dims(y,-1)*tf.log(logits + 1e-10) - (1 - tf.expand_dims(y,-1))*tf.log(1 - logits + 1e-10)\n    loss=tf.reduce_mean(entropy, name=\"loss\")\n\nwith tf.name_scope(\"train\"):\n    optimizer=tf.train.AdamOptimizer(learning_rate)\n    training_operation=optimizer.minimize(loss)\n\nwith tf.name_scope(\"accuracy\"):\n    predicted = tf.round(logits)\n    correct_pred = tf.equal(predicted, tf.expand_dims(y,-1))\n    accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))\n\n\nepochs = 1000\ninit= tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n    init.run()\n\n    for epoch in range(epochs):\n\n        for X_batch,y_batch in shuffle_batch(X_data,y_data,batch_size):\n            sess.run(training_operation,feed_dict={X:X_batch,y:y_batch})\n\n        acc_train=accuracy.eval(feed_dict={X:X_data,y:y_data})\n\n        print(\"Epoch\",epoch,\"training acuuracy\",acc_train)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57936515,"AnswerCount":0,"Tags":"<c++><pytorch><caffe2>","CreationDate":"2019-09-14T14:38:24.927","AcceptedAnswerId":null,"OwnerUserId":5553929.0,"Title":"Cant find caffe2 library with find_library(CAFFE2_LIB caffe2) in CMAKE","Body":"<p>I'm trying to compile @leonardvandriel <a href=\"https:\/\/github.com\/leonardvandriel\/caffe2_cpp_tutorial\" rel=\"nofollow noreferrer\">https:\/\/github.com\/leonardvandriel\/caffe2_cpp_tutorial<\/a>\nc++ caffe2 torurials and i have trouble with finding the caffe2 library. <\/p>\n\n<p>I followed the instructions on: \n<a href=\"https:\/\/caffe2.ai\/docs\/getting-started.html?platform=ubuntu&amp;configuration=compile\" rel=\"nofollow noreferrer\">https:\/\/caffe2.ai\/docs\/getting-started.html?platform=ubuntu&amp;configuration=compile<\/a>\nto install caffe2.<\/p>\n\n<p>I also installed caffe2 through anconda and it is working through python, however, through c++ i can't get to compile it successfully since the caffe2 library is not found in the compilation process.<\/p>\n\n<p>Thanks for any answer regarding this issue. <\/p>\n","answers":[]}
+{"QuestionId":57936544,"AnswerCount":0,"Tags":"<python><multiprocessing><pytorch><dataloader>","CreationDate":"2019-09-14T14:42:16.053","AcceptedAnswerId":null,"OwnerUserId":12066927.0,"Title":"OSError: [Errno 24] Too many open files in Windows 10 when using pytorch dataloader","Body":"<p>I'm doing a model of a CNN network which takes images that are located in a directory \"images\", and each image is located inside that directory in a directory named as the ID of the image (ex. image 1 is in: images\/1\/1.png).<\/p>\n\n<pre><code>class SSMDataset(Dataset):\n    songs_list = []\n    def __init__(self, root_dir, transform=None):\n        \"\"\"lista de nombre de canciones\"\"\"\n        self.root_dir = root_dir\n        self.songs_list = []\n        self.labels_list = []\n        for (dirpath, dirnames, filenames) in os.walk(self.root_dir):\n            for f in filenames:\n                if f.endswith('.png'):\n                    o = {}\n                    o['img_path'] = dirpath + '\/' + f\n                    self.songs_list.append(o)\n                elif f.endswith('.csv'):\n                    o['label_path'] = dirpath + '\/' + f\n                    self.labels_list.append(o)\n        self.transform = transform\n\n    def __len__(self):\n        return len(self.songs_list)\n\n    def __getitem__(self, index):\n        img_path = self.songs_list[index]['img_path']\n        image = Image.open(img_path)\n        fp = image.fp\n        image.load()\n        fp.closed\n        image = np.array(image)\n        image = image[np.newaxis, :, :]\n        label_path = self.labels_list[index]['label_path']\n        label = np.genfromtxt(label_path)\n        if self.transform is not None:\n            for t in self.transform:\n                image, labels = t(image, label)\n        return image, label\n<\/code><\/pre>\n\n<p>When I train the model, at the epoch 23 and always in taht epoch the following error raises: <\/p>\n\n<p>Epoch: 23\n39% 136\/350 [00:32&lt;00:37,  5.67it\/s, accuracy=0.221, loss=1.97]<\/p>\n\n<p>Traceback (most recent call last):\nFile dataloader.py line 560, in <strong>next<\/strong>\nFile dataloader.py line 560, in \nFile  line 55, in getitem\nFile Image.py line 2543, in open<\/p>\n\n<p>OSError: [Errno 24] Too many open files: C:\\Users\\...\\images\\1296\n\/1296.png<\/p>\n\n<p>Process finished with exit code 1<\/p>\n\n<p>I have made proof with different libraries as cv2, matplotlib...to open the image, I have use the command \"with Image.open(img_path) as im:\" to make sure the image is closed after it is taken but it always raises the same error and in the same epoch. I also included: <\/p>\n\n<pre><code>import torch.multiprocessing\ntorch.multiprocessing.set_sharing_strategy('file_system')\n<\/code><\/pre>\n\n<p>in the code but the error continues to appear at the same epoch. Any idea of how could it be solved? Maybe is it a Windows problem...?<\/p>\n","answers":[]}
+{"QuestionId":57936588,"AnswerCount":0,"Tags":"<r><tensorflow>","CreationDate":"2019-09-14T14:49:07.523","AcceptedAnswerId":null,"OwnerUserId":12067376.0,"Title":"Error installing tensorflow in R: \/Users\/emmeran\/.virtualenvs\/r-reticulate\/bin\/pip: No such file or directory","Body":"<p>After installing the <code>tensorflow<\/code> package for R from Github, and loading with <code>library(tensorflow)<\/code>, when doing <code>install_tensorflow()<\/code>, I get the error message below. I don't know how to get it to find the file or where I can even know where the file is. Thanks<\/p>\n\n<pre><code>install_tensorflow()\n\nUsing virtual environment '~\/.virtualenvs\/r-reticulate' ...\n\nsh: \/Users\/emmeran\/.virtualenvs\/r-reticulate\/bin\/pip: No such file or directory\n<\/code><\/pre>\n\n<blockquote>\n  <p>Error: Error installing package(s): 'pip'<\/p>\n  \n  <p>In addition: Warning message:<\/p>\n  \n  <p>In system2(pip, args) : error in running command<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":57936595,"AnswerCount":0,"Tags":"<python><keras><feature-extraction>","CreationDate":"2019-09-14T14:49:40.110","AcceptedAnswerId":null,"OwnerUserId":9916923.0,"Title":"CNN Feature Extraction Time","Body":"<p>I have a dataset consist of 260 thousands images that are extracted from several videos. I want to extract features of these images and use them for frame retrieval. I used VGG16 (pretrained on imagenet) that implemented in Keras library with 'avg' pooling in the last convolutional layer. VGG16 gives me a vector consist of 512 number (feature) for each image. The only reason that bothers me is that this scenario is too time-consuming. For my dataset, it took about a day and 6 hours which is too much.<br>\nIs this elapsed time normal?<br>\nBecause of low performance, I switch from VGG16 to DenseNet121 that already implemented in Keras. For this model (now) it took a day and 18 hours to extract features from 33% of my images (about 86000).<br>\nI ask again: <strong>Is this elapsed time normal?<\/strong><br>\nIs there any way to extract feature faster? Even without using of implemented algorithms?\nIf you need more clarification, just ask for it. Thank You!<\/p>\n","answers":[]}
+{"QuestionId":57936622,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2019-09-14T14:53:00.863","AcceptedAnswerId":null,"OwnerUserId":10068973.0,"Title":"Can someone tell me how to fix this error , here I'm working with tensorflow fashion_mnish dataset","Body":"<p>Here is the code<\/p>\n\n<pre><code>plt.figure(figsize=(10,10))\n\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(train_image[i], cmap=plt.cm.binary)\n    plt.xlabel(class_names[train_label[i]])\nplt.show()\n<\/code><\/pre>\n\n<p>Here is an error that occur when I run the above code<\/p>\n\n<pre><code>TypeError                                 Traceback (most recent call last)\n 7     plt.grid(False)\n  8     plt.imshow(train_image[i], cmap=plt.cm.binary)\n  9     plt.xlabel(class_names[train_label[i]])\n 10 plt.show()\n\n  TypeError: only integer scalar arrays can be converted to a scalar index\n<\/code><\/pre>\n","answers":[{"AnswerId":"57936688","CreationDate":"2019-09-14T15:02:45.970","ParentId":null,"OwnerUserId":"9437380","Title":null,"Body":"<p>It is hard to tell, as I don't know which form your variables have. My guess would be that:<\/p>\n\n<p>plt.xlabel(class_names[train_label[i]])<\/p>\n\n<p>train_label[i] is not returning an integer. Can you check if train_label is an int and not float?<\/p>\n\n<p>My other guess would be that the resulting shape of train_label[i] is not correct.<\/p>\n"}]}
+{"QuestionId":57937040,"AnswerCount":2,"Tags":"<tensorflow><keras><gpu><resnet><vgg-net>","CreationDate":"2019-09-14T15:50:11.643","AcceptedAnswerId":null,"OwnerUserId":347484.0,"Title":"Can I train Keras\/TF model layer by layer?","Body":"<p>I am looking to train a large face identification network. Resnet or VGG-16\/19. TensorFlow 1.14<\/p>\n\n<p>My question is - if I run out of GPU memory - <strong>is it valid strategy to train sets of layers one by one?<\/strong><\/p>\n\n<p>For example train 2 cnn and maxpooling layer as one set, then \"freeze the weights\" somehow and train next set etc..<\/p>\n\n<p><em>I know I can train on multi-gpu in tensorflow but what if I want to stick to just one GPU..<\/em><\/p>\n","answers":[{"AnswerId":"57938143","CreationDate":"2019-09-14T18:21:35.483","ParentId":null,"OwnerUserId":"1849975","Title":null,"Body":"<p>I may be wrong <em>but<\/em>, even if you freeze your weights, they still need to be loaded into the memory (you need to do whole forward pass in order to compute the loss).<\/p>\n\n<p>Comments on this are appreciated.<\/p>\n"},{"AnswerId":"57937286","CreationDate":"2019-09-14T16:19:38.247","ParentId":null,"OwnerUserId":"2160809","Title":null,"Body":"<p>The usual approach is to use transfer learning: use a pretrained model and fine-tune it for the task.<\/p>\n\n<p>For fine-tuning in computer vision, a known approach is re-training only the last couple of layers. See for example:<br>\n<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"}]}
+{"QuestionId":57937097,"AnswerCount":2,"Tags":"<python><tensorflow><keras><resnet><vgg-net>","CreationDate":"2019-09-14T15:55:19.207","AcceptedAnswerId":57937219.0,"OwnerUserId":347484.0,"Title":"Is it possible to validate a deep learning model by training small data subset?","Body":"<p>I am looking to train a large model (resnet or vgg) for face identification.<\/p>\n\n<p><strong>Is it valid strategy to train on few faces (1..3) to validate a model?<\/strong><\/p>\n\n<p>In other words - if a model learns one face well - is it evidence that the model is good for the task?<\/p>\n\n<p><em>point here is that I don't want to spend a week of GPU expensive time only to find out that my model is no good or data has errors or my TF coding has a bug<\/em><\/p>\n","answers":[{"AnswerId":"57937143","CreationDate":"2019-09-14T16:01:49.397","ParentId":null,"OwnerUserId":"3653343","Title":null,"Body":"<p><strong>Short answer:<\/strong> No, because Deep Learning works well on huge amount of data.<\/p>\n\n<p><strong>Long answer:<\/strong> No. The problem is that learning only one face could overfit your model on that specific face, without learning features not present in your examples. Because for example, the model has learn to detect your face thanks to a specific, very simple, pattern in that face (that's called <em>overfitting<\/em>). <\/p>\n\n<p>Making a stupid simple example, your model has learn to detect that face because there is a mole on your right cheek, and it has learn to identify it<\/p>\n\n<p>To make your model perform well on the general case, you need an huge amount of data, making your model capable to learn different kind of patterns<\/p>\n\n<p><strong>Suggestion:<\/strong>\nBecause the training of a deep neural network is a time consuming task, usually one does not train one single neural network at time, but many neural network are trained in parallel, with different hyperparameters (layers, nodes, activation functions, learning rate, etc).<\/p>\n\n<p>Edit because of the discussion below:<\/p>\n\n<p>If your dataset is small is quite impossible to have a good performance on the general case, because the neural network will learn the <em>easiest pattern<\/em>, which is <em>usually<\/em> not the general\/better one. <\/p>\n\n<p>Adding data <em>you force<\/em> the neural network to extract <em>good<\/em> patterns, that work on the general case. <\/p>\n\n<p>It's a tradeoff, but <em>usually<\/em> a training on a small dataset would not lead to a good classifier on the general case<\/p>\n\n<p>edit2: refrasing everything to make it more clear. A good performance on a small dataset don't tell you if your model when trained on all the dataset is a good model. That's why you train to \nthe majority of your dataset and test\/validate on a smaller dataset<\/p>\n"},{"AnswerId":"57937219","CreationDate":"2019-09-14T16:11:32.467","ParentId":null,"OwnerUserId":"2160809","Title":null,"Body":"<p>For face recognition, usually a siamese net or triplet loss are used. This is an approach for one-shot learning. Which means it could perform really well given only few examples per class (person face here), but you still need to train it on many examples (different person faces). See for example:<br>\n<a href=\"https:\/\/towardsdatascience.com\/one-shot-learning-with-siamese-networks-using-keras-17f34e75bb3d\" rel=\"nofollow noreferrer\">https:\/\/towardsdatascience.com\/one-shot-learning-with-siamese-networks-using-keras-17f34e75bb3d<\/a><\/p>\n\n<p>You wouldn't train your model from scratch but use a pretrained model anyways and fine-tune it for your task<\/p>\n\n<p>You could also have a look at pretrained face recognition models for better results like facenet<br>\n<a href=\"https:\/\/github.com\/davidsandberg\/facenet\" rel=\"nofollow noreferrer\">https:\/\/github.com\/davidsandberg\/facenet<\/a><\/p>\n"}]}
+{"QuestionId":57937696,"AnswerCount":1,"Tags":"<python><tensorflow><shapes><inference><seq2seq>","CreationDate":"2019-09-14T17:16:14.587","AcceptedAnswerId":null,"OwnerUserId":3894673.0,"Title":"In tensorflow is there a way to find out a shape (rank) of elements' outputs while building a graph?","Body":"<p>I'm trying to build a graph in tensorflow. But it gives me an error that I have wrong rank of shape. So, I'm trying to locate at which step something went wrong. Is there a chance to find out shapes of elements's outputs while building a graph?<\/p>\n\n<p>For example, my code is:<\/p>\n\n<pre><code>def inference_decoding_layer(start_token, end_token, embeddings, dec_cell, initial_state, output_layer,\n                         max_summary_length, batch_size):\n'''Create the inference logits'''\n\n\nstart_tokens = tf.tile(tf.constant([start_token], dtype=tf.int32), [batch_size], name='start_tokens')\n\ninference_helper = tf.contrib.seq2seq.GreedyEmbeddingHelper(embeddings, #shape (2000,25,768)\n                                                            start_tokens,\n                                                            end_token)\n\n\ninference_decoder = tf.contrib.seq2seq.BasicDecoder(dec_cell,\n                                                    inference_helper,\n                                                    initial_state,\n                                                    output_layer)\n\ninference_logits, _ , _ = tf.contrib.seq2seq.dynamic_decode(inference_decoder,\n                                                        output_time_major=False,\n                                                        impute_finished=True,\n                                                        maximum_iterations=max_summary_length) \n\nreturn inference_decoder\n<\/code><\/pre>\n\n<p>The problem appears at dynamic_decoder. Here is the error:<\/p>\n\n<blockquote>\n  <p>ValueError: Shape must be rank 3 but is rank 2 for 'decode\/decoder\/while\/BasicDecoderStep\/decoder\/attention_wrapper\/concat_6' (op: 'ConcatV2') with input shapes: [32,25,768], [32,256], [].<\/p>\n<\/blockquote>\n\n<p>So, I'm wondering is there a way to find out, for example, what shape of the value we get from GreedyEmbeddingHelper and then from BasicDecoder... Or maybe of other thing in my whole code. So, I would locate where the problem lays.<\/p>\n\n<p>P.S. If there are any other ways\/suggestions of how to locate the problem in this case I would be very grateful!<\/p>\n","answers":[{"AnswerId":"57938258","CreationDate":"2019-09-14T18:37:50.050","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>For the sake of easy debugging, <code>eager<\/code> mode has been introduced. With eager mode, you can keep printing the output shape after each line of code is executed.<\/p>\n\n<p>In TF 1.x, to enable it, you have to run the following code:<\/p>\n\n<pre><code>tf.enable_eager_execution()\n<\/code><\/pre>\n\n<p>In TF 2.0, by default eager mode will be enabled. Also, the packages you are working on has been moved to <a href=\"https:\/\/github.com\/tensorflow\/addons\/tree\/master\/tensorflow_addons\/seq2seq\" rel=\"nofollow noreferrer\">TensorFlow Addons<\/a> in TF 2.0. <\/p>\n"}]}
+{"QuestionId":57937753,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-14T17:24:43.050","AcceptedAnswerId":null,"OwnerUserId":12067790.0,"Title":"Keras validation with fit_generator and dropout layer","Body":"<p>I hope someone can help me answer a question about Keras dropout layers in combination with a fit_generator.<\/p>\n\n<p>Currently I am training a model containing multiple dropout layers. I understood however that when predicting with a model containing dropout layers, Keras will automatically disable them by utilizing a learning phase variable. <\/p>\n\n<p>I wonder if this also applied to training and validating with a fit_generator. Where upon validating with a generator (validation_data) the dropout layers will be deactivated?<\/p>\n","answers":[]}
+{"QuestionId":57937938,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-09-14T17:53:18.707","AcceptedAnswerId":58208735.0,"OwnerUserId":11579387.0,"Title":"Failed to get convolution algorithm","Body":"<p>im trying to train a convolutional network, i will omit the part of the code that import and create folders for the pictures and clases, now this is the main code of the net, this example was taken from the book \"deep learning in python\" from Fran\u00e7ois Collet<\/p>\n\n<pre><code>from keras import layers\nfrom keras import models\nmodel = 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.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.Dense(512, activation='relu'))\nmodel.add(layers.Dense(1177, activation='softmax'))\n\nfrom keras import optimizers\nmodel.compile(loss='categorical_crossentropy',\noptimizer=optimizers.RMSprop(lr=1e-4),\nmetrics=['acc'])\n\nfrom keras.preprocessing.image import ImageDataGenerator\ntrain_datagen = ImageDataGenerator(rescale=1.\/255)\ntest_datagen = ImageDataGenerator(rescale=1.\/255)\ntrain_generator = train_datagen.flow_from_directory(NEW_DATA_DIR, target_size=(150, 150), batch_size=10,class_mode='categorical')\n\nhistory = model.fit_generator(\ntrain_generator,\nsteps_per_epoch=100,\nepochs=30)\n<\/code><\/pre>\n\n<p>and then i get this error<\/p>\n\n<pre><code>WARNING:tensorflow:From \/home\/mimus\/anaconda3\/envs\/tf_gpu\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/math_ops.py:3066: to_int32 (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse tf.cast instead.\nEpoch 1\/30\n\n---------------------------------------------------------------------------\nUnknownError                              Traceback (most recent call last)\n&lt;ipython-input-7-380fa4b50355&gt; in &lt;module&gt;\n      2 train_generator,\n      3 steps_per_epoch=100,\n----&gt; 4 epochs=30)\n\n~\/anaconda3\/envs\/tf_gpu\/lib\/python3.6\/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\/tf_gpu\/lib\/python3.6\/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\/envs\/tf_gpu\/lib\/python3.6\/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\/envs\/tf_gpu\/lib\/python3.6\/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\/envs\/tf_gpu\/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~\/anaconda3\/envs\/tf_gpu\/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~\/anaconda3\/envs\/tf_gpu\/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~\/anaconda3\/envs\/tf_gpu\/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 conv2d_1\/convolution}}]]\n     [[{{node loss\/mul}}]]\n<\/code><\/pre>\n\n<p>im using python 3.6, tensorflowgpu 1.13.1, kerasgpu 2.2.4, cudatoolkit 10.0.130, cudnn 7.3.1, numpy 1.16.5. i really hope you can helpme to solve this, thank you in advance<\/p>\n\n<p>edit:<\/p>\n\n<p>using nvidia-smi i get this<\/p>\n\n<pre><code>NVIDIA-SMI 430.26       Driver Version: 430.26       CUDA Version: 10.2     |\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 2060    Off  | 00000000:01:00.0  On |                  N\/A |\n|  0%   52C    P8    19W \/ 170W |   5895MiB \/  5931MiB |     24%      Default |\n+-------------------------------+----------------------+----------------------+\n\n+-----------------------------------------------------------------------------+\n| Processes:                                                       GPU Memory |\n|  GPU       PID   Type   Process name                             Usage      |\n|=============================================================================|\n|    0      1305      G   \/usr\/lib\/xorg\/Xorg                            20MiB |\n|    0      1356      G   \/usr\/bin\/gnome-shell                          51MiB |\n|    0      1556      G   \/usr\/lib\/xorg\/Xorg                           123MiB |\n|    0      1703      G   \/usr\/bin\/gnome-shell                          74MiB |\n|    0      2816      G   \/usr\/lib\/firefox\/firefox                       3MiB |\n|    0     10133      G   \/usr\/lib\/firefox\/firefox                       3MiB |\n|    0     18578      C   ...\/mimus\/anaconda3\/envs\/tf_gpu\/bin\/python  5605MiB |\n+-----------------------------------------------------------------------------+\n<\/code><\/pre>\n","answers":[{"AnswerId":"58208735","CreationDate":"2019-10-02T20:12:12.080","ParentId":null,"OwnerUserId":"11579387","Title":null,"Body":"<p>the problem was that my gpu have not enough ram for the complexity of the net, this can be solved by trying a smaller model, thank you all<\/p>\n"}]}
+{"QuestionId":57938283,"AnswerCount":1,"Tags":"<python><tensorflow><keras><gpu><cpu-usage>","CreationDate":"2019-09-14T18:41:54.427","AcceptedAnswerId":null,"OwnerUserId":9895768.0,"Title":"Understanding the Hardware usage when training a Classifier on a GPU","Body":"<p>I am training a Cat-Dog Classifier with the help transfer learning on a <strong>GPU<\/strong> using <strong>TF 2.0<\/strong>. I used Keras ImageDataGenerator for performing data augmentation. While training the model, I monitored the usage of GPU, Disk(HDD) and CPU, and noted the following:-<\/p>\n\n<ol>\n<li>For the first epoch, the Disk usage is at a constant <strong>20-25%<\/strong>.\n    However, near the end of the epoch, during the validation phase, it spikes to <strong>40-50%<\/strong>. For the subsequent epochs, the Disk usage is\n    strangely <strong>zero<\/strong>.  <\/li>\n<li>The CPU utilization is at <strong>~60%<\/strong> throughout an epoch, except at\n        the end when it spikes to 100%.<\/li>\n<li>When I loaded <strong>any<\/strong> Keras Model (like VGG16,Xception,InceptionResNetV2 etc), all of them took about <strong>6.5 gigs<\/strong> of VRAM.  <\/li>\n<li>The GPU usage is at a constant <strong>29-30%<\/strong> for almost the entirety\n        of an epoch, expect at the end when it spikes to <strong>70%<\/strong>.<\/li>\n<\/ol>\n\n<p>From these observations, I made the following inferences:-<\/p>\n\n<ol>\n<li>During the First epoch, the images are loaded from the HDD to RAM in\nbatches of 8(batch_size=8), and as soon as a batch is loaded, the\nCPU performs data augmentation on the batch and sends the augmented batch to the GPU\nfor training. Now, when the validation stage comes, since there is\nno data augmentation to be performed on the Validation data, the CPU\ncan directly pass the batch of image from RAM to GPU without any\npreprocessing. Thus, the overhead is reduced, and all 3 namely,\nDisk, CPU and GPU work at a higher speed, hence the higher usage in\nall of them during the end of an epoch.<\/li>\n<li>For the subsequent epochs, since all the images(training+validation)\nhave already been loaded onto RAM, there is no need to fetch it from\nthe Disk, and so Disk usage remains zero.<\/li>\n<\/ol>\n\n<p>However, there were a few things I could not wrap my head around:-<\/p>\n\n<ol>\n<li>In the <a href=\"https:\/\/keras.io\/applications\/\" rel=\"nofollow noreferrer\">Keras Appplications Page<\/a>, VGG19 has the largest size in\nterms of memory footprint (549 MB), and I can't understand how can\nthis go on to consume about 6.5 GB of VRAM?<\/li>\n<li>Why is it that all the models took the same amount of VRAM when\n    loaded, even though they are vastly different in terms of size(both\n    total number of layers and memory footprint) ?<\/li>\n<\/ol>\n\n<p>Here are a few code snippets:-<\/p>\n\n<pre><code>train_datagen = ImageDataGenerator(rotation_range = 30,\n                                   width_shift_range = 0.4,\n                                   height_shift_range = 0.4,\n                                   shear_range = 0.4,\n                                   zoom_range = 0.25,\n                                   horizontal_flip = True,\n                                   brightness_range = [0.5, 1.5],\n                                   preprocessing_function = preprocess_input) \n\nvalid_datagen = ImageDataGenerator(preprocessing_function = preprocess_input)\n\ntrain_generator = train_datagen.flow_from_dataframe(train_data,\n                                                    directory = 'train\/',\n                                                    x_col = 'Photo',\n                                                    y_col = 'Class',\n                                                    target_size = (299,299),\n                                                    class_mode = 'binary',\n                                                    seed = 42,\n                                                    batch_size = 8)\n\nvalidation_generator = valid_datagen.flow_from_dataframe(valid_data,\n                                                    directory = 'train\/',\n                                                    x_col = 'Photo',\n                                                    y_col = 'Class',\n                                                    target_size = (299,299),\n                                                    class_mode = 'binary',\n                                                    seed = 42,\n                                                    batch_size = 8)\n\ninception_resnet_v2 = InceptionResNetV2(include_top = False,\n                                    weights = 'imagenet',\n                                    input_shape = (299, 299, 3),\n                                    pooling = 'avg',\n                                    classes = 2)\ninception_resnet_v2.trainable = False\n\nout = Dense(1, activation = 'sigmoid')(inception_resnet_v2.output)\n\nmodel = Model(inputs = inception_resnet_v2.inputs, outputs = out)\n\ncheckpoint = ModelCheckpoint('model.h5',\n                             monitor = 'val_accuracy',\n                             verbose = 0,\n                             save_best_only = True,\n                             save_weights_only = False, \n                             mode = 'max',\n                             period = 1)\n\noptim = tf.keras.optimizers.Adam(lr = 0.0001)\nmodel.compile(optimizer = optim, loss = 'binary_crossentropy', metrics = ['accuracy'])\n\nhist = model.fit_generator(train_generator,\n                           steps_per_epoch = len(train_generator),\n                           epochs = 10,\n                           callbacks = [checkpoint],\n                           validation_data = validation_generator,\n                           verbose = 1,\n                           validation_steps = len(validation_generator),\n                           validation_freq = 1)\n<\/code><\/pre>\n\n<p>I would be grateful if someone could answer my questions and also point out, if my inferences were correct or wrong.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"58539635","CreationDate":"2019-10-24T10:44:45.233","ParentId":null,"OwnerUserId":"12198627","Title":null,"Body":"<p>By default, TensorFlow maps nearly all of the GPU memory of all GPUs (subject to CUDA_VISIBLE_DEVICES) visible to the process. <\/p>\n\n<p>You can limit the same using,<\/p>\n\n<pre><code>import tensorflow as tf\nfrom keras import backend as k\n\nconfig = tf.ConfigProto() # TensorFlow wizardry\nconfig.gpu_options.allow_growth = True # Don't pre-allocate memory; allocate as-needed\nconfig.gpu_options.per_process_gpu_memory_fraction = 0.95 # Only allow a total fraction the GPU memory to be allocated\nk.tensorflow_backend.set_session(tf.Session(config=config)) # Create a session with the above options specified.\n<\/code><\/pre>\n\n<p>Also you can check this for more information : <a href=\"https:\/\/www.tensorflow.org\/guide\/gpu\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/gpu<\/a><\/p>\n"}]}
+{"QuestionId":57938405,"AnswerCount":0,"Tags":"<python><macos><tensorflow>","CreationDate":"2019-09-14T18:58:51.797","AcceptedAnswerId":null,"OwnerUserId":11474049.0,"Title":"No module named tensorflow while i installed it with pip","Body":"<p>I'm new to Tensorflow and machine learning. I'm following some tutorials on my PC, and everything worked just fine so far, importation and test with python are OK. I tried to test my code on my Macbook. So I installed Tensorflow, Matplotlib and Pandas using Pip3 on the command prompt. But when I ran the code on my Mac I got this :<\/p>\n\n<pre><code>Traceback (most recent call last):\n   File \"test.py\", line 3, in &lt;module&gt;\nimport tensorflow as ts\nModuleNotFoundError: No module named 'tensorflow'\n<\/code><\/pre>\n\n<p>There are the output and the Pip3 list :<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9gWWY.jpg\" rel=\"nofollow noreferrer\">Result of the pip install tensorflow<\/a><\/p>\n\n<pre><code>pip3 show tensorflow\n<\/code><\/pre>\n\n<p>gave me :<\/p>\n\n<pre><code>Name: tensorflow\nVersion: 2.0.0a0\nSummary: TensorFlow is an open source machine learning framework for everyone.\nHome-page: https:\/\/www.tensorflow.org\/\nAuthor: Google Inc.\nAuthor-email: packages@tensorflow.org\nLicense: Apache 2.0\nLocation: \/usr\/local\/lib\/python3.7\/site-packages\nRequires: google-pasta, gast, absl-py, astor, keras-preprocessing, keras-applications, six,           numpy, wheel, grpcio, protobuf, tf-estimator-nightly, termcolor, tb-nightly\nRequired-by:\n<\/code><\/pre>\n\n<p>And :<\/p>\n\n<pre><code>pip3 list\n<\/code><\/pre>\n\n<p>gave me :<\/p>\n\n<pre><code>Package              Version\n-------------------- --------------------\nabsl-py              0.8.0\nastor                0.8.0\ncycler               0.10.0\ngast                 0.3.2\ngoogle-pasta         0.1.7\ngrpcio               1.23.0\nh5py                 2.10.0\nKeras-Applications   1.0.8\nKeras-Preprocessing  1.1.0\nkiwisolver           1.1.0\nMarkdown             3.1.1\nmatplotlib           3.1.1\nnumpy                1.17.2\npandas               0.25.1\npip                  19.1.1\nprotobuf             3.9.1\npyparsing            2.4.2\npython-dateutil      2.8.0\npytz                 2019.2\nsetuptools           41.0.1\nsix                  1.12.0\ntb-nightly           1.14.0a20190301\ntensorflow           2.0.0a0\ntermcolor            1.1.0\ntf-estimator-nightly 1.14.0.dev2019030115\nvirtualenv           16.7.5\nWerkzeug             0.15.6\nwheel                0.33.4\n<\/code><\/pre>\n\n<p>I did the exact same things on my Mac that I did on my PC, but I have more than one version of Python on my Mac, maybe this is the problem, but I just don't know how to solve it.<\/p>\n\n<p>Sorry for my English, I'm not quit used to post on forum like these, and thanks for your help, I can give more information on my problem if you want.<\/p>\n","answers":[]}
+{"QuestionId":57938506,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-14T19:12:40.927","AcceptedAnswerId":58509928.0,"OwnerUserId":560449.0,"Title":"Tensorflow Graph - check if a node depends on a placeholder","Body":"<p>In a <code>Tensorflow<\/code> graph, is there a way to find out if a <code>node<\/code> depends on a <code>placeholder<\/code>, like <code>node.depends(placeholder) -&gt; Bool<\/code><\/p>\n\n<pre><code>import tensorflow as tf\n\nx = tf.placeholder(name='X', dtype=tf.int64, shape=[])\ny = tf.placeholder(name='Y', dtype=tf.int64, shape=[])\n\np = tf.add(x, y)\nq = tf.add(x, x)\nsess = tf.Session()\n\nresult = sess.run([p,q], feed_dict={x: 1, y: 2})\nprint(result)\n\nresult = sess.run([p,q], feed_dict={x: 1, y: 3})\nprint(result)\n<\/code><\/pre>\n\n<p>In the code example above, <code>q<\/code> does not depend on <code>y<\/code>. In the second call of <code>session.run<\/code>, we modify only <code>y<\/code>. Thus <code>q<\/code> does not need to be evaluated again. Does session automatically reuse existing values  in these cases? If so, is there any way to find out which are the nodes that were evaluated during <code>.run<\/code>?<\/p>\n\n<p>Otherwise, if I can quickly find out which nodes are dependent on the placeholders I modify, I can send only those as input to run, and reuse existing values (in a dictionary as cache).<\/p>\n\n<p>The idea is to avoid costly evaluations and more importantly, in my application, to minimize costly operations (outside <code>tensorflow<\/code>) that need to be triggered whenever the output nodes change - a necessity in my application.<\/p>\n","answers":[{"AnswerId":"58509928","CreationDate":"2019-10-22T17:58:06.857","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Checking dependency between tensors in a graph can be done with a function like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Checks if tensor a depends on tensor b\ndef tensor_depends(a, b):\n    if a.graph is not b.graph:\n        return False\n    gd = a.graph.as_graph_def()\n    gd_sub = tf.graph_util.extract_sub_graph(gd, [a.op.name])\n    return b.op.name in {n.name for n in gd_sub.node}\n<\/code><\/pre>\n\n<p>For example:<\/p>\n\n<pre><code>import tensorflow as tf\n\nx = tf.placeholder(name='X', dtype=tf.int64, shape=[])\ny = tf.placeholder(name='Y', dtype=tf.int64, shape=[])\np = tf.add(x, y)\nq = tf.add(x, x)\nprint(tensor_depends(q, x))\n# True\nprint(tensor_depends(q, y))\n# False\n<\/code><\/pre>\n\n<p>About your questions, generally TensorFlow does recompute everything on each run, even if the inputs do not change. If you want to cache results, you can do it yourself, at a higher level - for TensorFlow, it would not be clear what results it should keep (only the latest ones, a few recent ones, ...). And in any case, even if the inputs do not change, the output may change, as is the case with recurrent models, or more generally due to any change in <a href=\"https:\/\/stackoverflow.com\/q\/52636943\">stateful objects<\/a> such as variables or datasets. There are some optimization opportunities that are lost, but it would probably not be reasonable to expect TensorFlow to address them (analyze the model to determine whether it can cache results or not, what results can be cached, how much, how to configure it, etc.). If you know that some intermediate values in the graph will not change, you can also compute that part and then feed it as input (you can actually feed values for any tensor in the <code>feed_dict<\/code>). Otherwise, as you suggest, you can just check what depends on what and recompute only as necessary.<\/p>\n"}]}
+{"QuestionId":57938514,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-14T19:13:47.707","AcceptedAnswerId":null,"OwnerUserId":8847560.0,"Title":"How to modify weights of a layer in another layer in Keras?","Body":"<p>I am trying to modify the weights of a keras layer , such as Conv2D, in another custom layer.<\/p>\n\n<p>I have tried to pass the layer (which's weights will be modified) to custom layer so that I can use get_weights, set_weights methods in the custom layer. However, it turned out I cannot pass a layer or a model to a custom layer as arguments.<\/p>\n\n<p>For example, I want to use the model or a specific layer in the custom layer below. To do this, I need to pass the model or the specific layer to ExampleFunc. Note that I cannot pass the model or layer in _ <em>init<\/em> _.<\/p>\n\n<pre><code>    class exampleLayer(tf.keras.layers.Layer):\n        def __init__(self, indexofOtherLayer):\n            super(exampleLayer, self).__init__()\n            self.indL = indexofOtherLayer\n\n        def call(self, inputs, **kwargs):\n            return ExampleFunc(MODEL,inputs,self.indL)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57939110,"AnswerCount":1,"Tags":"<keras><tokenize>","CreationDate":"2019-09-14T20:40:51.870","AcceptedAnswerId":57940292.0,"OwnerUserId":5203203.0,"Title":"Keras Tokenizer Character Level Not Working","Body":"<p>I am sending a list of lists through the Keras Tokenizer with char_level = True, yet the result is word tokenization, not character tokenization. <\/p>\n\n<pre><code>    from tensorflow import keras\n    from keras.preprocessing.text import Tokenizer\n\n\n    # List of lists\n    train_data = [['SMITH', 'JOHN', '', 'CHESTERTOWN', 'MD', '21620', '555555555', 'F'], ['CROW', 'JOE', '', 'FREDERICK', 'MD', '217011313', '9999999999', 'F']]\n\n    t = Tokenizer(filters='!\"#$%&amp;()*+,-.\/:;&lt;=&gt;?@[\\\\]^_`{|}~\\t\\n', split=',', char_level=True, oov_token=True) \n    t.fit_on_texts(train_data)\n    train_token = np.array(t.texts_to_sequences(train_data)) \n\n    print(train_token)\n    array([[ 5,  6,  2,  7,  3,  8,  9,  4], [10, 11,  2, 12,  3, 13, 14,  4]])\n<\/code><\/pre>\n","answers":[{"AnswerId":"57940292","CreationDate":"2019-09-15T00:39:50.150","ParentId":null,"OwnerUserId":"10956606","Title":null,"Body":"<p>This is happening because you your data should be a string, not a list. If you concatenate all words into one string, it will work as expected.<\/p>\n\n<p>Just add the following to your code:<\/p>\n\n<pre><code>def concat_list(l):\n    concat = ''\n    for word in l:\n        concat += word + ' '\n    return concat\n\ntrain_data = [concat_list(data) for data in train_data]\n<\/code><\/pre>\n\n<p>You will then get:<\/p>\n\n<pre><code>&gt;&gt;&gt; [list([16, 9, 17, 10, 11, 2, 18, 7, 11, 19, 2, 2, 12, 11, 5, 16, 10, 5, 8, 10, 7, 20, 19, 2, 9, 13, 2, 14, 6, 23, 14, 21, 2, 4, 4, 4, 4, 4, 4, 4, 4, 4, 2, 15, 2])\n     list([12, 8, 7, 20, 2, 18, 7, 5, 2, 2, 15, 8, 5, 13, 5, 8, 17, 12, 24, 2, 9, 13, 2, 14, 6, 25, 21, 6, 6, 22, 6, 22, 2, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 2, 15, 2])]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57939122,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-14T20:43:37.027","AcceptedAnswerId":57941751.0,"OwnerUserId":12068262.0,"Title":"Learning about keras library, improving weights and bias","Body":"<p>I am a new user trying to learn more about neural network. \nMy question is. \nWhen keras is training. and is changing the accurasy (weights and bias improving), Is keras saving the last accuray with more percent?<\/p>\n\n<p>for example. (the last NET that I did)<\/p>\n\n<pre><code>1570\/1570 [==============================] - 1s 352us\/sample - loss: 0.1273 - acc: 0.9605\nEpoch 13573\/20000\n1570\/1570 [==============================] - 1s 380us\/sample - loss: 0.1106 - acc: 0.9580\nEpoch 13574\/20000\n1570\/1570 [==============================] - 1s 397us\/sample - loss: 0.0660 - acc: 0.9764\nEpoch 13575\/20000\n1570\/1570 [==============================] - 0s 308us\/sample - loss: 0.0849 - acc: 0.9707\nEpoch 13576\/20000\n1570\/1570 [==============================] - 1s 359us\/sample - loss: 0.0549 - acc: 0.9815\nEpoch 13577\/20000\n1570\/1570 [==============================] - 1s 359us\/sample - loss: 0.0502 - acc: 0.9828\nEpoch 13578\/20000\n1570\/1570 [==============================] - 1s 331us\/sample - loss: 0.0492 - acc: 0.9834\nEpoch 13579\/20000\n1570\/1570 [==============================] - 1s 375us\/sample - loss: 0.0531 - acc: 0.9841\nEpoch 13580\/20000\n1570\/1570 [==============================] - 0s 312us\/sample - loss: 0.0445 - acc: 0.9866\nEpoch 13581\/20000\n1570\/1570 [==============================] - 1s 375us\/sample - loss: 0.0438 - acc: 0.9860\nEpoch 13582\/20000\n1570\/1570 [==============================] - 1s 373us\/sample - loss: 0.0601 - acc: 0.9796\nEpoch 13583\/20000\n1570\/1570 [==============================] - 1s 406us\/sample - loss: 0.0905 - acc: 0.9669\nEpoch 13584\/20000\n1570\/1570 [==============================] - 1s 420us\/sample - loss: 0.1169 - acc: 0.9580\n<\/code><\/pre>\n\n<p>If I decide to end the current training. Keras take the weights and bias with the acc: 0.9866 or Keras will take the last? (in this case 0.9580) with more loss. <\/p>\n\n<p>In case Keras take the last. I would like to know if there are any posibility for be able to add some lines and save weights and bias (when the acc(last) > acc(past)).<\/p>\n\n<p>just in case. A simple code example. <\/p>\n\n<pre><code>X = # Data imput\nY = # Data output\n\nmodel = keras.Sequential([ keras.layers.Flatten(input_shape=(1, 14)),\n                           keras.layers.Dense(56, activation='relu'),\n                           keras.layers.Dense(28, activation='relu'),\n                           keras.layers.Dense(14, activation='relu'),\n                           keras.layers.Dense(7, activation='relu'),\n                           keras.layers.Dense(2, activation='softmax') ])\nmodel.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\nmodel.fit(X, Y, epochs=20000)\n\ntest_loss,test_acc = model.evaluate(X,Y)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57941751","CreationDate":"2019-09-15T06:41:08.797","ParentId":null,"OwnerUserId":"5597718","Title":null,"Body":"<blockquote>\n  <p>If I decide to end the current training. Keras take the weights and\n  bias with the acc: 0.9866 or Keras will take the last? (in this case\n  0.9580) with more loss.<\/p>\n<\/blockquote>\n\n<p>Keras will take the last. To save the best weights, you can pass the <code>ModelCheckpoint<\/code> callback via the <code>callbacks<\/code> argument to <code>model.fit<\/code>. For writing custom logic to save the weights you can use <code>LambdaCallback<\/code>.<\/p>\n\n<p><a href=\"https:\/\/keras.io\/callbacks\/#modelcheckpoint\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/callbacks\/#modelcheckpoint<\/a><\/p>\n\n<p><a href=\"https:\/\/keras.io\/callbacks\/#lambdacallback\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/callbacks\/#lambdacallback<\/a><\/p>\n"}]}
+{"QuestionId":57939204,"AnswerCount":1,"Tags":"<machine-learning><keras>","CreationDate":"2019-09-14T20:56:46.727","AcceptedAnswerId":null,"OwnerUserId":7317985.0,"Title":"ValueError: Error when checking target: expected dense_3 to have shape (1,) but got array with shape (11,)","Body":"<p>I'm trying to train my model with Keras and I'm taking this online course from udemy. Now everything works fine but when I try to fit the ANN to the training set it gives the following error. everything works fine but when I EXECUTE this last line it gives the error.\nIt should work fine without this error or is there any other way to fit the ANN to the training set?<\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\n\ndataset = pd.read_csv('Churn_Modelling.csv')\nX = dataset.iloc[:, 3:13].values\ny = dataset.iloc[:, 13].values\n\nfrom sklearn.preprocessing import OneHotEncoder, LabelEncoder\nlabelencoder_X_1 = LabelEncoder()\nX[:, 1] = labelencoder_X_1.fit_transform(X[:, 1])\nlabelencoder_X_2 = LabelEncoder()\nX[:, 2] = labelencoder_X_2.fit_transform(X[:, 2])\n\nonehotencoder = OneHotEncoder(categorical_features = [1])\nX = onehotencoder.fit_transform(X).toarray()\nX = X[:, 1:]\n\nfrom sklearn.model_selection import train_test_split\nX_train , y_train , X_test, y_test = train_test_split(X,y, test_size = 0.2 , random_state = 0)\n\n\n#convert X_test into a 'numpy' array to acoid valur error for 1D array \nX_test = np.reshape(y, (-1,1))\n\nfrom sklearn.preprocessing import StandardScaler\nsc = StandardScaler()\nX_train = sc.fit_transform(X_train)\nX_test = sc.fit_transform(X_test)\n\n\nimport keras\nfrom keras.models import Sequential\nfrom keras.layers import Dense\n\n#initializing the ANN\nclassifier = Sequential()\n\n#adding the input layer and the first hidden layer\nclassifier.add(Dense(units =6, kernel_initializer = 'uniform' , activation = 'relu', input_dim =11 ))\n\n#adding the second layer\nclassifier.add(Dense(units = 6 , kernel_initializer = 'uniform' , activation = 'relu'))\n\n#adding the output layer\nclassifier.add(Dense(units = 1 , kernel_initializer = 'uniform' , activation = 'sigmoid'))\n\n#compiling the ANN\nclassifier.compile(optimizer = 'adam' , loss = 'binary_crossentropy', metrics = ['accuracy'])\n# 'optimizer' is the algorithm that u wanna use for the wights adjustments\n\n#fitting the ann to the trainging set \nclassifier.fit(X_train , y_train , batch_size =10 , epochs = 100 )\n<\/code><\/pre>\n","answers":[{"AnswerId":"57975398","CreationDate":"2019-09-17T13:32:13.063","ParentId":null,"OwnerUserId":"3660313","Title":null,"Body":"<p>It seems that <code>input_shape<\/code> is not set correctly.<\/p>\n\n<p><strong>from <a href=\"https:\/\/keras.io\/layers\/core\/\" rel=\"nofollow noreferrer\">docs<\/a>:<\/strong><\/p>\n\n<blockquote>\n  <p>Input shape<\/p>\n  \n  <p>nD tensor with shape: (batch_size, ..., input_dim). The most common\n  situation would be a 2D input with shape (batch_size, input_dim).<\/p>\n<\/blockquote>\n\n<p>In your case <code>input_shape=(X_train.shape[1],)<\/code><\/p>\n\n<p>Try this:<\/p>\n\n<pre><code>#initializing the ANN\nclassifier = Sequential()\n\n#adding the input layer and the first hidden layer\nclassifier.add(Dense(units=6,\n                     kernel_initializer='uniform',\n                     activation='relu',\n                     input_shape=(X_train.shape[1],))\n\n...\n<\/code><\/pre>\n"}]}
+{"QuestionId":57939268,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-datasets><tensorflow-estimator>","CreationDate":"2019-09-14T21:07:58.237","AcceptedAnswerId":null,"OwnerUserId":8315741.0,"Title":"How to use Tensorflow's two different install types together?","Body":"<p>I have run into a situation wherein I need to use a Colab notebook (GPU) for<\/p>\n\n<pre><code>!pip install tf-nightly-gpu-2.0-preview\n<\/code><\/pre>\n\n<p>and then, in the same notebook, I also need to use <\/p>\n\n<pre><code>!pip install -q gpt-2-simple\n<\/code><\/pre>\n\n<p>But when I import GPT-2 using,<\/p>\n\n<pre><code>import gpt_2_simple as gpt2\n<\/code><\/pre>\n\n<p>I get the error\n<code>AttributeError: module<\/code> <code>'tensorflow_estimator.python.estimator.api._v1.estimator' has no attribute '__file__'<\/code><\/p>\n\n<p>Clearly, GPT-2 depends on Tensorflow. I am very confused and have spent hours trying to work this out in terms of Tensorflow Versions, but nothing seems to sort this. I basically need to use the <code>nightly-gpu<\/code> version as this has a major role in my program's operation, but with this, GPT-2 doesn't work. Any help will be appreciated!<\/p>\n","answers":[]}
+{"QuestionId":57939906,"AnswerCount":0,"Tags":"<tensorflow><tensor><tensorflow-datasets><resnet>","CreationDate":"2019-09-14T23:03:35.840","AcceptedAnswerId":null,"OwnerUserId":3268219.0,"Title":"How are H, W, C, N tensor attributes form image data related to tensor attributes of Batch (B), M, and DModel that are used in models like BERT?","Body":"<p>For 4d tensor attributes are H, W, C, N. But there are tensor parameters B, M, DModel in BERT benchmark topology. How are they related?<\/p>\n","answers":[]}
+{"QuestionId":57939910,"AnswerCount":0,"Tags":"<c++><opencv><deep-learning><pytorch><caffe2>","CreationDate":"2019-09-14T23:04:22.757","AcceptedAnswerId":null,"OwnerUserId":5553929.0,"Title":"Running Caffe2 Model on android - my model running very slow on the android device(dosen't finish the init phase)","Body":"<p>I'm trying to run a Caffe2 model on an android device, i ran the AICameraDemo App successfully but my model init net is 7 times larger than sqeeznet model, and the load of the init net to the workspace with the predictor or with RunNetOnce takes forever.<\/p>\n\n<p>I saw posts about this, most of them from 2018 so i'm wondering if anything change since then.<\/p>\n\n<p>If it is not possible, is there another framework which i could run a model this big on an android device ? <\/p>\n\n<p>Thanks for any answer. <\/p>\n","answers":[]}
+{"QuestionId":57940151,"AnswerCount":1,"Tags":"<python><python-3.x><pytorch>","CreationDate":"2019-09-15T00:01:22.943","AcceptedAnswerId":null,"OwnerUserId":2111778.0,"Title":"PyTorch: Sharing a model between threads, do I need to lock it myself?","Body":"<p>I'm sharing a PyTorch neural network model between a main thread which <code>trains<\/code> the model and a number of worker threads which <code>eval<\/code> the model to generate training samples (\u00e0 la AlphaGo).<\/p>\n\n<p>My question is, do I need to create a separate mutex to lock and unlock when accessing the model in different threads?<\/p>\n\n<hr>\n\n<p>I did not see any mutex in the PyTorch <a href=\"https:\/\/github.com\/pytorch\/examples\/blob\/master\/mnist_hogwild\/main.py\" rel=\"nofollow noreferrer\">Hogwild<\/a> example, however the difference there is that all threads are in <code>train<\/code> mode.<\/p>\n\n<p>Apart from the official <a href=\"https:\/\/pytorch.org\/docs\/stable\/notes\/multiprocessing.html\" rel=\"nofollow noreferrer\">tutorial<\/a>, I'm unable to find even basic information (like this question) online.<\/p>\n\n<p>I'm using the <code>torch.multiprocessing<\/code> module for this and I'm using <code>model.share_memory()<\/code> and <code>torch.multiprocessing.set_start_method('spawn', force=True)<\/code>.<\/p>\n","answers":[{"AnswerId":"57943357","CreationDate":"2019-09-15T10:44:10.693","ParentId":null,"OwnerUserId":"5688831","Title":null,"Body":"<p>You need to <code>.join()<\/code> all worker threads if you wish to continue using main thread.<\/p>\n"}]}
+{"QuestionId":57940947,"AnswerCount":1,"Tags":"<keras><scikit-learn><hyperparameters><talos>","CreationDate":"2019-09-15T03:35:26.653","AcceptedAnswerId":null,"OwnerUserId":7088205.0,"Title":"Functional API\/Multi Input model hyperparameter optimization","Body":"<p>The only hyperparameter optimization library I've found that works with Keras functional API is Talos.<\/p>\n\n<p><a href=\"https:\/\/github.com\/autonomio\/talos\/blob\/9890d71d31451af3d7e8d91a75841bc7904db958\/docs\/Examples_Multiple_Inputs.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/autonomio\/talos\/blob\/9890d71d31451af3d7e8d91a75841bc7904db958\/docs\/Examples_Multiple_Inputs.md<\/a><\/p>\n\n<p>Does anyone know any others that would work?<\/p>\n","answers":[{"AnswerId":"57978732","CreationDate":"2019-09-17T16:48:21.843","ParentId":null,"OwnerUserId":"3660313","Title":null,"Body":"<p>You can perform Keras Hyperparameter Tuning using Sklearn Grid Search with Cross Validation.<\/p>\n\n<p>To perform Grid Search with Sequential Keras models (single-input only), you must turn these models into sklearn-compatible estimators by using <a href=\"http:\/\/keras.io\/scikit-learn-api\" rel=\"nofollow noreferrer\">Keras Wrappers for the Scikit-Learn API<\/a>.<\/p>\n\n<blockquote>\n  <p>A sklearn estimator is a class object with fit(X,y) , predict(x)and\n  scoremethods. (and optionnaly predict_proba method)<\/p>\n<\/blockquote>\n\n<p>No need to do that from scratch, you can use Sequential Keras models as part of your Scikit-Learn workflow by implementing one of two wrappers from keras.wrappers.scikit_learnpackage:<\/p>\n\n<pre><code>KerasClassifier(build_fn=None, **sk_params): which implements the Scikit-Learn classifier interface.\nKerasRegressor(build_fn=None, **sk_params): which implements the Scikit-Learn regressor interface.\n<\/code><\/pre>\n\n<p><strong>Arguments<\/strong><\/p>\n\n<pre><code>build_fn: callable function or class instance the should construct, compile and return a Keras model, which will then be used to fit\/predict.\nsk_params: model parameters &amp; fitting parameters.\n<\/code><\/pre>\n\n<blockquote>\n  <p>Note that like all other estimators in scikit-learn, build_fn should\n  provide default values for its arguments, so that you could create the\n  estimator without passing any values to sk_params.<\/p>\n<\/blockquote>\n\n<h2>Example: Simple binary classification with Keras<\/h2>\n\n<h2><strong>Implement Keras Model creator function<\/strong><\/h2>\n\n<p>We want to fine-tune these hyperparameters: <strong>optimizer, dropout_rate, kernel_init method and dense_layer_sizes.<\/strong><\/p>\n\n<p>These parameters must be defined in the signature of <strong>create_model()<\/strong> function with <strong>default parameters<\/strong>. You can add other hyperparameters if you want such as learning_rate, ...<\/p>\n\n<p><strong>binary_crossentropy<\/strong> is perfect for Two-class classification problem.<\/p>\n\n<pre><code>def create_model(dense_layer_sizes, optimizer=\"adam\", dropout=0.1, init='uniform', nbr_features=2500, dense_nparams=256):\n    model = Sequential()\n    model.add(Dense(dense_nparams, activation='relu', input_shape=(nbr_features,), kernel_initializer=init,)) \n    model.add(Dropout(dropout), )\n    for layer_size in dense_layer_sizes:\n        model.add(Dense(layer_size, activation='relu'))\n        model.add(Dropout(dropout), )\n    model.add(Dense(1, activation='sigmoid'))\n    model.compile(loss='binary_crossentropy', optimizer=optimizer,metrics=[\"accuracy\"])\n    return model\n<\/code><\/pre>\n\n<h2><strong>Create sklearn-like estimator<\/strong><\/h2>\n\n<p>It\u2019s a classification problem so we are using KerasClassifier wrapper.<\/p>\n\n<pre><code>kears_estimator = KerasClassifier(build_fn=create_model, verbose=1)\n<\/code><\/pre>\n\n<h2><strong>Defining Hyperparamers Space<\/strong><\/h2>\n\n<p>We define here our <strong>hyperparameters space<\/strong> including keras <strong>fit hyperparameters<\/strong>: <strong>epochs<\/strong> and <strong>batch_size<\/strong>:<\/p>\n\n<pre><code># define the grid search parameters\nparam_grid = {\n    epochs': [10, 100, ],\n    dense_nparams': [32, 256, 512],\n    init': [ 'uniform', 'zeros', 'normal', ], \n    batch_size':[2, 16, 32],\n    optimizer':['RMSprop', 'Adam', 'Adamax', 'sgd'],\n    dropout': [0.5, 0.4, 0.3, 0.2, 0.1, 0]\n}\n<\/code><\/pre>\n\n<h2><strong>And Finally Performing Grid Search with KFold Cross Validation<\/strong><\/h2>\n\n<pre><code>kfold_splits = 5\ngrid = GridSearchCV(estimator=kears_estimator,  \n                    n_jobs=-1, \n                    verbose=1,\n                    return_train_score=True,\n                    cv=kfold_splits,  #StratifiedKFold(n_splits=kfold_splits, shuffle=True)\n                    param_grid=param_grid,)\n\ngrid_result = grid.fit(X, y, )\n\n# summarize results\nprint(\"Best: %f using %s\" % (grid_result.best_score_, grid_result.best_params_))\nmeans = grid_result.cv_results_['mean_test_score']\nstds = grid_result.cv_results_['std_test_score']\nparams = grid_result.cv_results_['params']\nfor mean, stdev, param in zip(means, stds, params):\n    print(\"%f (%f) with: %r\" % (mean, stdev, param))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57941665,"AnswerCount":1,"Tags":"<python><tensorflow><keras><jupyter>","CreationDate":"2019-09-15T06:25:19.493","AcceptedAnswerId":57943082.0,"OwnerUserId":12069030.0,"Title":"How to combine the weights of 5 trained models to predict test dataset with better accuracy?","Body":"<p>Am using MNIST Fashion data set\n1 - i have divided train data into 5 different location, which each contain only 2 class images.\n2 - i have trained each location data set in different model and test data content all (0 to 9) class images for prediction.<\/p>\n\n<p>Question->how to combine all 5 trained models to one model which can predict my test data with better accuracy ?<\/p>\n","answers":[{"AnswerId":"57943082","CreationDate":"2019-09-15T10:08:27.030","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>There are many ways you can combine the output from this 5 separate models. I am listing some of them.<\/p>\n\n<ol>\n<li>Simply add the probability obtained for each class in those 5 models and then whichever is having maximum, give the verdict as that class.<\/li>\n<li>Go with voting pattern. Which class is being predicted as output by most number of classifiers. In case of tie, check whose probability is highest.<\/li>\n<li>Ensemble them. Pass the final hidden layer (or layers) features of all classifiers to another network. This new network can be as simple as a single dense layer or some complicated <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/BoostedTreesClassifier\" rel=\"nofollow noreferrer\">boosted tree models<\/a>.<\/li>\n<\/ol>\n"}]}
+{"QuestionId":57941674,"AnswerCount":2,"Tags":"<python><tensorflow><keras><conv-neural-network>","CreationDate":"2019-09-15T06:26:58.700","AcceptedAnswerId":null,"OwnerUserId":10849918.0,"Title":"Getting Validation Accuracy of 99% with MNIST with less than 10000 parameters CNN","Body":"<p>Given MNIST dataset in keras,the challenge is to develop a CNN neural net model with less than 10k parameters with 99% validation accuracy.  <\/p>\n\n<p>I tried making the model for the same but am getting accuracy as 98.71.  <\/p>\n\n<pre><code>    def create_model():\n        lay1=Conv2D(2,kernel_size=(1,1),activation='relu',padding='same')(inputs)\n        lay1=Conv2D(2,kernel_size=(7,7),strides=(2,2),activation='relu',padding='same')(lay1)\n        lay1=Conv2D(2,kernel_size=(1,1),activation='relu',padding='same')(lay1)\n        lay1=MaxPooling2D(pool_size=(7,7),strides=(2,2),padding='same')(lay1)\n        lay2=Conv2D(4,kernel_size=(1,1),activation='relu',padding='same')(inputs)\n        lay2=Conv2D(4,kernel_size=(7,7),strides=(2,2),activation='relu',padding='same')(lay2)\n        lay2=Conv2D(4,kernel_size=(1,1),activation='relu',padding='same')(lay2)\n        lay2=MaxPooling2D(pool_size=(7,7),strides=(2,2),padding='same')(lay2)\n        lay3=Conv2D(6,kernel_size=(1,1),activation='relu',padding='same')(inputs)\n        lay3=Conv2D(6,kernel_size=(7,7),strides=(2,2),activation='relu',padding='same')(lay3)\n        lay3=Conv2D(6,kernel_size=(1,1),activation='relu',padding='same')(lay3)\n        lay3=MaxPooling2D(pool_size=(7,7),strides=(2,2),padding='same')(lay3)\n        fc=concatenate([lay1,lay2,lay3])\n        fc=Flatten()(fc)\n        fc=Dense(10,activation='relu')(fc)\n        outputs=Dense(10,activation='softmax')(fc)\n        model=Model(input=inputs,output=outputs)\n        model.compile(loss=keras.losses.categorical_crossentropy,\n          optimizer=keras.optimizers.Adadelta(),\n          metrics=['accuracy'])\n        return model\n<\/code><\/pre>\n\n<p>The total parameters coming are 8,862 and the batch size used for the above is 32 and the number of epochs are 10.  <\/p>\n\n<p>Can you please suggest ways to improve the model with the constraints on the number of parameters so that the validation accuracy is 99% or above?  <\/p>\n","answers":[{"AnswerId":"58187206","CreationDate":"2019-10-01T14:39:54.640","ParentId":null,"OwnerUserId":"6624101","Title":null,"Body":"<p>Here is a <a href=\"https:\/\/myrtle.ai\/how-to-train-your-resnet-8-bag-of-tricks\/\" rel=\"nofollow noreferrer\">blog post<\/a> which aims to train a CIFAR10 ResNet Classifier to reach 94% in few seconds. Although the goal is a liitle bit different, some tricks might be useful for you, such as CELU activation or label smoothing.<\/p>\n"},{"AnswerId":"57944351","CreationDate":"2019-09-15T13:01:48.547","ParentId":null,"OwnerUserId":"7534984","Title":null,"Body":"<p>You should be able to add another Conv\/Pool block when reducing the kernel_size in the previous ones to 3 or 5 instead of 7. If research told us one thing recently is that deeper is (mostly) better.<\/p>\n\n<p>Otherwise play with the hyperparameters (learning-rate, batch-size, ...) or augment the input data randomly. Batch- or Layer-Normalization have proven themselves very util lately (and they only introduce a handful of parameters).<\/p>\n"}]}
+{"QuestionId":57941970,"AnswerCount":1,"Tags":"<tensorflow><computer-vision><conv-neural-network>","CreationDate":"2019-09-15T07:18:07.877","AcceptedAnswerId":null,"OwnerUserId":12069286.0,"Title":"Reducing overfitting in CNN by increasing training data size versus augmenting images (preprocessing data) using DataImageGenerator","Body":"<p>Does increasing the size of the training data helps in reducing overfitting ? Or is it suggested to go for image augmentation (data preprocessing) using the ImageDataGenerator in Tensorflow to skew or rotate the image to decrease overfitting ? \nWhich method is better to reduce overfitting ??<\/p>\n","answers":[{"AnswerId":"57942999","CreationDate":"2019-09-15T09:56:20.050","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>By means of image augmentation, basically you are increasing the size of your training data. If you have data which is different from your existing training data, then adding that data into your training data is good.<\/p>\n\n<p>So in short, both the methods are good to overcome over-fitting.<\/p>\n"}]}
+{"QuestionId":57942487,"AnswerCount":1,"Tags":"<python><pandas><pytorch>","CreationDate":"2019-09-15T08:43:53.587","AcceptedAnswerId":null,"OwnerUserId":5740200.0,"Title":"How to convert torch tensor to pandas dataframe?","Body":"<p>I'd like to convert a torch tensor to pandas dataframe but by using <code>pd.DataFrame<\/code> I'm getting a dataframe filled with tensors instead of numeric values.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nimport pandas as  pd\nx = torch.rand(4,4)\npx = pd.DataFrame(x)\n<\/code><\/pre>\n\n<p>Here's what I get when clicking on <code>px<\/code> in the variable explorer:<\/p>\n\n<pre><code>0   1   2   3\ntensor(0.3880)  tensor(0.4598)  tensor(0.4239)  tensor(0.7376)\ntensor(0.4174)  tensor(0.9581)  tensor(0.0987)  tensor(0.6359)\ntensor(0.6199)  tensor(0.8235)  tensor(0.9947)  tensor(0.9679)\ntensor(0.7164)  tensor(0.9270)  tensor(0.7853)  tensor(0.6921)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57942576","CreationDate":"2019-09-15T08:56:22.773","ParentId":null,"OwnerUserId":"5740200","Title":null,"Body":"<p>I found one possible way by converting torch first to numpy:<\/p>\n\n<pre><code>import torch\nimport pandas as  pd\n x = torch.rand(4,4)\n px = x.numpy()\n px = pd.DataFrame(px)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57942507,"AnswerCount":2,"Tags":"<out-of-memory><pytorch>","CreationDate":"2019-09-15T08:46:29.880","AcceptedAnswerId":null,"OwnerUserId":3429266.0,"Title":"How to save GPU memory usage in PyTorch","Body":"<p>In PyTorch I wrote a very simple CNN discriminator and trained it. Now I need to deploy it to make predictions. But the target machine has a small GPU memory and got out of memory error. So I think that I can set <code>requires_grad = False<\/code> to prevent PyTorch from storing the gradient values. However I didn't find it making any difference.<\/p>\n\n<p>There are about 5 millions of parameters in my model. But when predicting a single batch of input, it consumes about 1.2GB of memory. I think there should be no need for such large memory.<\/p>\n\n<p>The question is how to save GPU memory usage when I just want to use my model to make predictions?<\/p>\n\n<hr>\n\n<p>Here is a demo, I use <code>discriminator.requires_grad_<\/code> to disable\/enable autograd of all parameters. But it seems to be no use.<\/p>\n\n<pre><code>import torch\nimport numpy as np\nimport torch.nn as nn\nimport torch.nn.functional as functional\n\nfrom pynvml.smi import nvidia_smi\nnvsmi = nvidia_smi.getInstance()\n\ndef getMemoryUsage():\n    usage = nvsmi.DeviceQuery(\"memory.used\")[\"gpu\"][0][\"fb_memory_usage\"]\n    return \"%d %s\" % (usage[\"used\"], usage[\"unit\"])\n\nprint(\"Before GPU Memory: %s\" % getMemoryUsage())\n\nclass Discriminator(nn.Module):\n    def __init__(self):\n        super().__init__()\n        # trainable layers\n        # input: 2x256x256\n        self.conv1 = nn.Conv2d(2, 8, 5, padding=2) # 8x256x256\n        self.pool1 = nn.MaxPool2d(2) # 8x128x128\n        self.conv2 = nn.Conv2d(8, 32, 5, padding=2) # 32x128x128\n        self.pool2 = nn.MaxPool2d(2) # 32x64x64\n        self.conv3 = nn.Conv2d(32, 96, 5, padding=2) # 96x64x64\n        self.pool3 = nn.MaxPool2d(4) # 96x16x16\n        self.conv4 = nn.Conv2d(96, 256, 5, padding=2) # 256x16x16\n        self.pool4 = nn.MaxPool2d(4) # 256x4x4\n        self.num_flat_features = 4096\n        self.fc1 = nn.Linear(4096, 1024)\n        self.fc2 = nn.Linear(1024, 256)\n        self.fc3 = nn.Linear(256, 1)\n        # loss function\n        self.loss = nn.MSELoss()\n        # other properties\n        self.requires_grad = True\n    def forward(self, x):\n        y = x\n        y = self.conv1(y)\n        y = self.pool1(y)\n        y = functional.relu(y)\n        y = self.conv2(y)\n        y = self.pool2(y)\n        y = functional.relu(y)\n        y = self.conv3(y)\n        y = self.pool3(y)\n        y = functional.relu(y)\n        y = self.conv4(y)\n        y = self.pool4(y)\n        y = functional.relu(y)\n        y = y.view((-1,self.num_flat_features))\n        y = self.fc1(y)\n        y = functional.relu(y)\n        y = self.fc2(y)\n        y = functional.relu(y)\n        y = self.fc3(y)\n        y = torch.sigmoid(y)\n        return y\n    def predict(self, x, score_th=0.5):\n        if len(x.shape) == 3:\n            singlebatch = True\n            x = x.view([1]+list(x.shape))\n        else:\n            singlebatch = False\n        y = self.forward(x)\n        label = (y &gt; float(score_th))\n        if singlebatch:\n            y = y.view(list(y.shape)[1:])\n        return label, y\n    def requires_grad_(self, requires_grad=True):\n        for parameter in self.parameters():\n            parameter.requires_grad_(requires_grad)\n        self.requires_grad = requires_grad\n\n\nx = torch.cuda.FloatTensor(np.zeros([2, 256, 256]))\ndiscriminator = Discriminator()\ndiscriminator.to(\"cuda:0\")\n\n# comment\/uncomment this line to make difference\ndiscriminator.requires_grad_(False)\n\ndiscriminator.predict(x)\n\nprint(\"Requires grad\", discriminator.requires_grad)\nprint(\"After GPU Memory: %s\" % getMemoryUsage())\n<\/code><\/pre>\n\n<p>By comment out the line <code>discriminator.requires_grad_(False)<\/code>, I got output:<\/p>\n\n<pre><code>Before GPU Memory: 6350MiB\nRequires grad True\nAfter GPU Memory: 7547MiB\n<\/code><\/pre>\n\n<p>While by uncomment the line, I got:<\/p>\n\n<pre><code>Before GPU Memory: 6350MiB\nRequires grad False\nAfter GPU Memory: 7543MiB\n<\/code><\/pre>\n","answers":[{"AnswerId":"57943399","CreationDate":"2019-09-15T10:49:49.103","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>You can use <a href=\"https:\/\/pypi.org\/project\/pynvml\/\" rel=\"nofollow noreferrer\">pynvml<\/a>.<\/p>\n\n<p>This python tool made Nvidia so you can Python query like this:<\/p>\n\n<pre><code>from pynvml.smi import nvidia_smi\nnvsmi = nvidia_smi.getInstance()\nnvsmi.DeviceQuery('memory.free, memory.total')\n<\/code><\/pre>\n\n<p>You can always also execute:<\/p>\n\n<pre><code>torch.cuda.empty_cache()\n<\/code><\/pre>\n\n<p>To empty the cache and you will find even more free memory that way.<\/p>\n\n<p>Before calling <code>torch.cuda.empty_cache()<\/code> if you have objects you don't use anymore you can call this:<\/p>\n\n<pre><code>obj = None\n<\/code><\/pre>\n\n<p>And after that you call<\/p>\n\n<pre><code>gc.collect()\n<\/code><\/pre>\n"},{"AnswerId":"57942942","CreationDate":"2019-09-15T09:49:03.267","ParentId":null,"OwnerUserId":"10749432","Title":null,"Body":"<p>Try to use <code>model.eval()<\/code> with <code>torch.no_grad()<\/code> on your target machine when making predictions. <code>model.eval()<\/code> will switch model layers to eval mode. <code>torch.no_grad()<\/code> will deactivate autograd engine and as a result memory usage will be reduced.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = torch.cuda.FloatTensor(np.zeros([2, 256, 256]))\ndiscriminator = Discriminator()\ndiscriminator.to(\"cuda:0\")\n\ndiscriminator.eval()\nwith torch.no_grad():\n    discriminator.predict(x)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57942575,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><reinforcement-learning>","CreationDate":"2019-09-15T08:56:20.987","AcceptedAnswerId":null,"OwnerUserId":4184024.0,"Title":"Replacing conv2d layers with a pre-made convolutional architecture produces an InvalidArgumentError with GatherV2 operation in A3C network definition","Body":"<p>Thanks for giving my problem a look. I'm a newbie in tensorflow. \nI am trying to train an A3C implementation for arcade fighting games. The network definition of an A3C implementation is quite simple with two convolutional layers followed by a fully connected layer with 256 units and terminating with two separate output layers sharing the same body, one with 63 outputs of action-probabilities with softmax, predicting the policy, and another with a single output, predicting the state-value V(s). This implementation works fine with all worker-agents terminating gracefully after their assigned combined number of steps ends.<\/p>\n\n<p>I also modified the architecture to allow for the training of more visually complex agents by replacing the convolutional layers (tf.contrib.layers.conv2d) with pre-made architectures included in tensorflow (tensorflow.contrib.slim.nets), including AlexNet and ResNet50, making it output a flattened vector of 512 classes. The frames were scaled to 224 x 224 grayscale, the native resolution of AlexNet or ResNet, and the latest 8 or 16 grayscale frames were stacked together, resulting in a tensor of 224 x 224 x 8 or 224 x 224 x 16, which is not the native number of channels for 3 for these networks, but 3 is a very small amount of latest frames to be kept.<\/p>\n\n<p>Whenever I try to run this implementation, the agent runs fine, for some time, but there is eventually an out-of-bounds exception with tensorflow's GatherV2 operation, which might somehow originate from the input operations. The bug takes its time to appear (I had to let the agents train for 2 hours to reproduce it) and I'm not sure what causes it. All tensors seem to be of their expected dimensions at all times.<\/p>\n\n<p>I've pasted both the code for network definitions of the regular and modified implementations.<\/p>\n\n<p>Thanks!<\/p>\n\n<p>I've tried both flattening the output tensor without providing number of classes argument and relying on the built-in flattening operation when providing a number of classes. <\/p>\n\n<p>Simple conv layer version:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nimport tensorflow as tf\nimport numpy as np\n\n\ndef build_feature_extractor(input_, frames_per_step):\n  # We only want to create the weights once\n  # In all future calls we should set reuse = True\n\n  # Normalize: scale the inputs from 0..255 to 0..1\n  input_ = tf.to_float(input_) \/ 255.0\n\n  # Extract normalized handcrafted inputs from frame data\n  # (game's memory values for P1 health, P2 health, Remaining time\n  # in addition to one hot encoding of character selected)\n  X_feat = tf.slice(input_[0, :, :, frames_per_step - 1], [0, 0], [1, 22])\n  X_feat = tf.squeeze(X_feat)\n\n\n  # conv layers\n  conv1 = tf.contrib.layers.conv2d(\n    input_,\n    16, # num output feature maps\n    8,  # kernel size\n    4,  # stride\n    activation_fn=tf.nn.relu,\n    scope=\"conv1\")\n  conv2 = tf.contrib.layers.conv2d(\n    conv1,\n    32, # num output feature maps\n    4,  # kernel size\n    2,  # stride\n    activation_fn=tf.nn.relu,\n    scope=\"conv2\")\n\n  # Flatten data\n  # image -&gt; feature vector\n  flat = tf.contrib.layers.flatten(conv2)\n  X_feat = X_feat[None]  # Add an extra dimension to feature vector to make it 2D\n  X_feat = tf.tile(X_feat, [tf.shape(flat)[0], 1])  # Repeat feature vector minibatch_size (num_workers) amount of times\n  flat = tf.concat([X_feat, flat], axis=1)  # Concatenate feature vector to each flattened vector of the minibatch\n  # dense layer\n  fc1 = tf.contrib.layers.fully_connected(\n    inputs=flat,\n    num_outputs=256,\n    scope=\"fc1\")\n\n  return fc1\n\nclass PolicyNetwork:\n  def __init__(self, num_outputs, im_height, im_width, frames_per_step, reg=0.01):\n    self.num_outputs = num_outputs\n\n    # Graph inputs\n    # After resizing we have 8 consecutive frames of size 200 x 127\n    self.states = tf.placeholder(shape=[None, im_height, im_width, frames_per_step], dtype=tf.uint8, name=\"X\")\n    # Advantage = G - V(s)\n    self.advantage = tf.placeholder(shape=[None], dtype=tf.float32, name=\"y\")\n    # Selected actions\n    self.actions = tf.placeholder(shape=[None], dtype=tf.int32, name=\"actions\")\n\n    # Since we set reuse=False here, that means we MUST\n    # create the PolicyNetwork before creating the ValueNetwork\n    # ValueNetwork will use reuse=True\n    with tf.variable_scope(\"shared\", reuse=False):\n      fc1 = build_feature_extractor(self.states, frames_per_step)\n\n    # Use a separate scope for output and loss\n    with tf.variable_scope(\"policy_network\"):\n      # Final dense layer that maps to the action space\n      # Gets logits for each action\n      self.logits = tf.contrib.layers.fully_connected(fc1, num_outputs, activation_fn=None)\n      # Apply softmax to get probabilities of each action\n      self.probs = tf.nn.softmax(self.logits)\n\n      # Sample an action\n      # Categorical distribution object which accepts logits\n      cdist = tf.distributions.Categorical(logits=self.logits)\n      # Use this to sample actions\n      self.sample_action = cdist.sample()\n\n      # Add regularization to increase exploration\n      self.entropy = -tf.reduce_sum(self.probs * tf.log(self.probs), axis=1)\n\n      # Get the predictions for the chosen actions only\n      batch_size = tf.shape(self.states)[0]\n      gather_indices = tf.range(batch_size) * tf.shape(self.probs)[1] + self.actions\n      self.selected_action_probs = tf.gather(tf.reshape(self.probs, [-1]), gather_indices)\n\n      #Calculate and minimize cost\n      self.loss = tf.log(self.selected_action_probs) * self.advantage + reg * self.entropy\n      self.loss = -tf.reduce_sum(self.loss, name=\"loss\")\n\n      # Training\n      # Create RMSProp optimizer\n      self.optimizer = tf.train.RMSPropOptimizer(0.0002, 0.99, 0.0, 5e-7)\n\n      # Use optimizer to compute the gradients of each variable in the network and store them in a list\n      # we'll need these later for running gradient descent steps\n      self.grads_and_vars = self.optimizer.compute_gradients(self.loss)\n      self.grads_and_vars = [[grad, var] for grad, var in self.grads_and_vars if grad is not None]\n\n\nclass ValueNetwork:\n  def __init__(self, im_height, im_width, frames_per_step):\n    # Placeholders for our input\n    # After resizing we have 4 consecutive frames of size 84 x 84\n    self.states = tf.placeholder(shape=[None, im_height, im_width, frames_per_step], dtype=tf.uint8, name=\"X\")\n    # The TD target value\n    self.targets = tf.placeholder(shape=[None], dtype=tf.float32, name=\"y\")\n\n    # Since we set reuse=True here, that means we MUST\n    # create the PolicyNetwork before creating the ValueNetwork\n    # PolicyNetwork will use reuse=False.\n    # When build_feature_extractor() is called with reuse = True,\n    # it is going to use the same variables created previously.\n    # [Share weights between two networks when reuse = True]\n    # This allows for the main body of the two neural networks to be shared\n    with tf.variable_scope(\"shared\", reuse=True):\n      fc1 = build_feature_extractor(self.states, frames_per_step)\n\n    # Use a separate scope for output and loss\n    # Fully connected final layer with 1 output (the predicted value)\n    with tf.variable_scope(\"value_network\"):\n      self.vhat = tf.contrib.layers.fully_connected(\n        inputs=fc1,\n        num_outputs=1,\n        activation_fn=None)\n      #Squeeze to convert into 1-D tensor\n      self.vhat = tf.squeeze(self.vhat, squeeze_dims=[1], name=\"vhat\")\n\n      self.loss = tf.squared_difference(self.vhat, self.targets)\n      self.loss = tf.reduce_sum(self.loss, name=\"loss\")\n\n      # training (e: 1e-6)\n      self.optimizer = tf.train.RMSPropOptimizer(0.0002, 0.99, 0.0, 5e-7)\n\n      # Obtain list of each variable and its corresponding gradient\n      # we'll need these later for running gradient descent steps\n      self.grads_and_vars = self.optimizer.compute_gradients(self.loss)\n      self.grads_and_vars = [[grad, var] for grad, var in self.grads_and_vars if grad is not None]\n\n\n# Should use this to create networks\n# to ensure they're created in the correct order\ndef create_networks(num_outputs, im_height, im_width, frames_per_step):\n  policy_network = PolicyNetwork(num_outputs=num_outputs, im_height=im_height, im_width=im_width, frames_per_step=frames_per_step)\n  value_network = ValueNetwork(im_height=im_height, im_width=im_width, frames_per_step=frames_per_step)\n  return policy_network, value_network\n<\/code><\/pre>\n\n<p>In the modifications, the architecture of the shared network body (feature extractor) was changed as such (here using AlexNet):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nimport tensorflow as tf\nimport numpy as np\nfrom Image_Transformer_AlexNet import IM_HEIGHT_AFTER_PAD, IM_WIDTH_AFTER_PAD\n\nfrom tensorflow.contrib.slim.nets import alexnet\n\ndef build_feature_extractor(input_, frames_per_step):\n  # We only want to create the weights once\n  # In all future calls we should set reuse = True\n\n  # Scale the inputs from 0..255 to 0..1\n  input_ = tf.to_float(input_) \/ 255.0\n\n  # Extract normalized handcrafted inputs from frame data\n  # (game's memory values for P1 health, P2 health, Remaining time\n  # in addition to one hot encoding of character selected)\n  X_feat = tf.slice(input_[0, :, :, frames_per_step - 1], [0, 0], [1, 22])\n  X_feat = tf.squeeze(X_feat)\n\n  # Pass data through the Alexnet Convolutional architecture\n  # Output: A 2D tensor, Flattens with max pooling the output to (minibatch_size, classes)\n  Z_flat, end_points = alexnet.alexnet_v2(input_, 512)\n\n  #Z_flat = tf.contrib.layers.flatten(Z)\n\n  X_feat = X_feat[None]  # Add an extra dimension to feature vector to make it 2D\n  X_feat = tf.tile(X_feat, [tf.shape(Z_flat)[0], 1])  # Repeat feature vector minibatch_size (num_workers) amount of times\n  Z_flat = tf.concat([X_feat, Z_flat], axis=1)  # Concatenate feature vector to each flattened vector of the minibatch\n\n  # dense layer\n  fc1 = tf.contrib.layers.fully_connected(\n    inputs=Z_flat,\n    num_outputs=256,\n    scope=\"fc1\")\n\n  return fc1\n<\/code><\/pre>\n\n<p>Exception stack:<\/p>\n\n<pre><code>Exception in thread Thread-17:\nTraceback (most recent call last):\n  File \"\/home\/rielfox4\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1356, in _do_call\n    return fn(*args)\n  File \"\/home\/rielfox4\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1341, in _run_fn\n    options, feed_dict, fetch_list, target_list, run_metadata)\n  File \"\/home\/rielfox4\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1429, in _call_tf_sessionrun\n    run_metadata)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: indices[9] = 630 is not in [0, 630)\n     [[{{node worker_0\/policy_network\/GatherV2}}]]\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"\/home\/rielfox4\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/threading.py\", line 916, in _bootstrap_inner\n    self.run()\n  File \"\/home\/rielfox4\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/threading.py\", line 864, in run\n    self._target(*self._args, **self._kwargs)\n  File \"\/home\/rielfox4\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/MAMEToolkit\/umk3_agents\/A3C_AlexNet_main.py\", line 119, in &lt;lambda&gt;\n    worker_fn = lambda: worker.run(sess, coord, STEPS_PER_UPDATE)\n  File \"\/home\/rielfox4\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/MAMEToolkit\/umk3_agents\/A3C_AlexNet_worker.py\", line 188, in run\n    self.current_pnet_loss, self.current_vnet_loss = self.update(steps, sess)\n  File \"\/home\/rielfox4\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/MAMEToolkit\/umk3_agents\/A3C_AlexNet_worker.py\", line 372, in update\n    ], feed_dict)\n  File \"\/home\/rielfox4\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 950, in run\n    run_metadata_ptr)\n  File \"\/home\/rielfox4\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1173, in _run\n    feed_dict_tensor, options, run_metadata)\n  File \"\/home\/rielfox4\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1350, in _do_run\n    run_metadata)\n  File \"\/home\/rielfox4\/anaconda3\/envs\/tensorflow\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1370, in _do_call\n    raise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: indices[9] = 630 is not in [0, 630)\n     [[node worker_0\/policy_network\/GatherV2 (defined at \/MAMEToolkit\/umk3_agents\/A3C_AlexNet_networks.py:76) ]]\n\nErrors may have originated from an input operation.\nInput Source operations connected to node worker_0\/policy_network\/GatherV2:\n worker_0\/policy_network\/add (defined at \/MAMEToolkit\/umk3_agents\/A3C_AlexNet_networks.py:75)\n\nOriginal stack trace for 'worker_0\/policy_network\/GatherV2':\n  File \"\/MAMEToolkit\/umk3_agents\/A3C_AlexNet_main.py\", line 96, in &lt;module&gt;\n    max_global_steps=MAX_GLOBAL_STEPS)\n  File \"\/MAMEToolkit\/umk3_agents\/A3C_AlexNet_worker.py\", line 130, in __init__\n    self.policy_net, self.value_net = create_networks(policy_net.num_outputs, im_height, im_width, frames_per_step)\n  File \"\/MAMEToolkit\/umk3_agents\/A3C_AlexNet_networks.py\", line 135, in create_networks\n    policy_network = PolicyNetwork(num_outputs=num_outputs, im_height=im_height, im_width=im_width, frames_per_step=frames_per_step)\n  File \"\/MAMEToolkit\/umk3_agents\/A3C_AlexNet_networks.py\", line 76, in __init__\n    self.selected_action_probs = tf.gather(tf.reshape(self.probs, [-1]), gather_indices)\n  File \"\/tensorflow\/python\/util\/dispatch.py\", line 180, in wrapper\n    return target(*args, **kwargs)\n  File \"\/tensorflow\/python\/ops\/array_ops.py\", line 3475, in gather\n    return gen_array_ops.gather_v2(params, indices, axis, name=name)\n  File \"\/tensorflow\/python\/ops\/gen_array_ops.py\", line 4097, in gather_v2\n    batch_dims=batch_dims, name=name)\n  File \"\/tensorflow\/python\/framework\/op_def_library.py\", line 788, in _apply_op_helper\n    op_def=op_def)\n  File \"\/tensorflow\/python\/util\/deprecation.py\", line 507, in new_func\n    return func(*args, **kwargs)\n  File \"\/tensorflow\/python\/framework\/ops.py\", line 3616, in create_op\n    op_def=op_def)\n  File \"\/tensorflow\/python\/framework\/ops.py\", line 2005, in __init__\n    self._traceback = tf_stack.extract_stack()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57942768,"AnswerCount":0,"Tags":"<dataframe><keras><tf.keras>","CreationDate":"2019-09-15T09:23:00.463","AcceptedAnswerId":null,"OwnerUserId":11879598.0,"Title":"How to fix\"DataFrameIterator' object has no attribute 'num_classes'\"?","Body":"<p>I'm loading the training images with tf.keras.preprocessing.image.ImageDataGenerator class. I choose to use flow_from_dataframe function. However, when I try to try to create a layer with num_classes, I got the error: AttributeError: 'DataFrameIterator' object has no attribute 'num_classes' <\/p>\n\n<pre><code>    datagen=tf.keras.preprocessing.image.ImageDataGenerator(rescale=1.\/255.)\n\nimage_data=datagen.flow_from_dataframe(\ndataframe=df[:1700],\ndirectory='\/tf\/notebooks\/2222\/Face Images',\nx_col=\"Filename\",\ny_col=\"Race\",\nbatch_size=32,\n#seed=42,\nshuffle=True,\nclass_mode=\"categorical\",\ntarget_size=(224,224))\n\nmodel = tf.keras.Sequential([\n  feature_extractor_layer,\n  layers.Dense(image_data.num_classes, activation='softmax')\n])\n\nmodel.summary()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57942911,"AnswerCount":2,"Tags":"<python-3.x><tensorflow><keras><mnist>","CreationDate":"2019-09-15T09:45:12.463","AcceptedAnswerId":57942928.0,"OwnerUserId":5114585.0,"Title":"Printing image from fashion_MNIST data gives colored image?","Body":"<p>The well-known fashion_MNIST data set for deep leanring\/computer vision is a collection of <strong>greyscale images<\/strong>[<a href=\"https:\/\/github.com\/zalandoresearch\/fashion-mnist]\" rel=\"nofollow noreferrer\">https:\/\/github.com\/zalandoresearch\/fashion-mnist]<\/a>.<\/p>\n\n<p>This dataset is already available with keras [<a href=\"https:\/\/keras.io\/datasets]\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/datasets]<\/a>.<\/p>\n\n<p>However, when I loaded the data set in Tensorflow Keras API and tried to print some image. I got <strong>colored images<\/strong>. I wonder why?\nCan anyone please explain. Here's my code:<\/p>\n\n<pre><code>import tensorflow as tf\nmnist = tf.keras.datasets.fashion_mnist\n(training_images, training_labels), (test_images, test_labels) = mnist.load_data()\n\nimport matplotlib.pyplot as plt\nplt.imshow(training_images[0])\n#print(training_labels[0])\n#print(training_images[0])\n<\/code><\/pre>\n","answers":[{"AnswerId":"57942928","CreationDate":"2019-09-15T09:47:32.507","ParentId":null,"OwnerUserId":"1428102","Title":null,"Body":"<p>You must set the color map.<\/p>\n\n<pre><code>plt.imshow(IMG, cmap='gray')\n<\/code><\/pre>\n"},{"AnswerId":"57943508","CreationDate":"2019-09-15T11:04:44.367","ParentId":null,"OwnerUserId":"8209005","Title":null,"Body":"<p>It happens because of <strong>imshow()<\/strong> method of matplotlib.\nBy default, it uses <strong>rcParams[\"image.cmap\"] to 'viridis'<\/strong> due to which we got a colored image. You can see here <a href=\"https:\/\/matplotlib.org\/3.1.0\/api\/_as_gen\/matplotlib.pyplot.imshow.html\" rel=\"nofollow noreferrer\">https:\/\/matplotlib.org\/3.1.0\/api\/_as_gen\/matplotlib.pyplot.imshow.html<\/a>. \nYou can also check these parameters in your terminal by running this code:<\/p>\n\n<pre><code> import matplotlib.pyplot as plt\n print(plt.rcParams)\n<\/code><\/pre>\n\n<p>If we set it to \"Greys\" then we will get grey image. You can set it by running this code:<\/p>\n\n<pre><code>plt.rcParams['image.cmap'] = 'Greys'\n<\/code><\/pre>\n\n<p>You can read more about 'viridis' here: <a href=\"https:\/\/matplotlib.org\/3.1.1\/tutorials\/colors\/colormaps.html\" rel=\"nofollow noreferrer\">https:\/\/matplotlib.org\/3.1.1\/tutorials\/colors\/colormaps.html<\/a><\/p>\n"}]}
+{"QuestionId":57943032,"AnswerCount":1,"Tags":"<python><image><matplotlib><pytorch>","CreationDate":"2019-09-15T10:02:06.793","AcceptedAnswerId":null,"OwnerUserId":11760140.0,"Title":"plt.imshow() gives TypeError: Image data of dtype object cannot be converted to float using PyTorch","Body":"<p>I am trying to create a custom Dataset Processor for a set of images. However, when I try to view the images in my dataset, I get hit with the TypeError: Image data of dtype object cannot be converted to float. <\/p>\n\n<p>I tried to check if I am passing in a PIL Image into the plt.imshow() function and I am.<\/p>\n\n<pre><code>class DatasetProcessing(Dataset):\n\n    def __init__(self, input_data, output_data, transform=None): \n        self.transform = transform\n        self.input_data = \n        input_data.reshape((-1,64,64)).astype(np.float32)[:,:,:,None]\n        self.output_data = output_data \n\n    def __getitem__(self, index): \n        return self.transform(self.input_data[index]), self.output_data[index]\n\n    def __len__(self): \n        return len(list(self.input_data))\n\ntransform = transforms.Compose([transforms.ToPILImage()])\n\n\ndset_train = DatasetProcessing(X_slices_train, Y_train, transform)\n\n\ntrain_loader = torch.utils.data.DataLoader(dset_train, batch_size=4, \n                                      shuffle=True, num_workers=4) \n\nplt.figure(figsize = (16, 4))\nfor num, x in enumerate(dset_train):\n    plt.subplot(1,6,num+1)\n    plt.axis('off')\n    print(x)\n    plt.imshow(np.asarray(x))\n    plt.title(y_train[num])\n<\/code><\/pre>\n\n<p>I expected to get pictures of my dataset, but instead I get the following error message:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-10-8b8caac49d97&gt; in &lt;module&gt;\n      4     plt.axis('off')\n      5     print(x)\n----&gt; 6     plt.imshow(np.asarray(x))\n      7     plt.title(y_train[num])\n\n~\/anaconda3\/lib\/python3.7\/site-packages\/matplotlib\/pyplot.py in imshow(X, cmap, norm, aspect, interpolation, alpha, vmin, vmax, origin, extent, shape, filternorm, filterrad, imlim, resample, url, data, **kwargs)\n   2675         filternorm=filternorm, filterrad=filterrad, imlim=imlim,\n   2676         resample=resample, url=url, **({\"data\": data} if data is not\n-&gt; 2677         None else {}), **kwargs)\n   2678     sci(__ret)\n   2679     return __ret\n\n~\/anaconda3\/lib\/python3.7\/site-packages\/matplotlib\/__init__.py in inner(ax, data, *args, **kwargs)\n   1587     def inner(ax, *args, data=None, **kwargs):\n   1588         if data is None:\n-&gt; 1589             return func(ax, *map(sanitize_sequence, args), **kwargs)\n   1590 \n   1591         bound = new_sig.bind(ax, *args, **kwargs)\n\n~\/anaconda3\/lib\/python3.7\/site-packages\/matplotlib\/cbook\/deprecation.py in wrapper(*args, **kwargs)\n    367                 f\"%(removal)s.  If any parameter follows {name!r}, they \"\n    368                 f\"should be pass as keyword, not positionally.\")\n--&gt; 369         return func(*args, **kwargs)\n    370 \n    371     return wrapper\n\n~\/anaconda3\/lib\/python3.7\/site-packages\/matplotlib\/cbook\/deprecation.py in wrapper(*args, **kwargs)\n    367                 f\"%(removal)s.  If any parameter follows {name!r}, they \"\n    368                 f\"should be pass as keyword, not positionally.\")\n--&gt; 369         return func(*args, **kwargs)\n    370 \n    371     return wrapper\n\n~\/anaconda3\/lib\/python3.7\/site-packages\/matplotlib\/axes\/_axes.py in imshow(self, X, cmap, norm, aspect, interpolation, alpha, vmin, vmax, origin, extent, shape, filternorm, filterrad, imlim, resample, url, **kwargs)\n   5658                               resample=resample, **kwargs)\n   5659 \n-&gt; 5660         im.set_data(X)\n   5661         im.set_alpha(alpha)\n   5662         if im.get_clip_path() is None:\n\n~\/anaconda3\/lib\/python3.7\/site-packages\/matplotlib\/image.py in set_data(self, A)\n    676                 not np.can_cast(self._A.dtype, float, \"same_kind\")):\n    677             raise TypeError(\"Image data of dtype {} cannot be converted to \"\n--&gt; 678                             \"float\".format(self._A.dtype))\n    679 \n    680         if not (self._A.ndim == 2\n\nTypeError: Image data of dtype object cannot be converted to float\n<\/code><\/pre>\n","answers":[{"AnswerId":"57943752","CreationDate":"2019-09-15T11:41:11.037","ParentId":null,"OwnerUserId":"10749432","Title":null,"Body":"<p>Your <code>dset_train<\/code> yields <code>self.transform(self.input_data[index]), self.output_data[index]<\/code> if understood correctly <code>self.transform(self.input_data[index])<\/code> is an image tensor (data) and <code>self.output_data[index]<\/code> is a label, but here: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>plt.imshow(np.asarray(x))\n<\/code><\/pre>\n\n<p>you are passing unpacked <code>x<\/code> which is actually (data, label)<\/p>\n\n<p>So, you need to unpack it first:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>plt.figure(figsize = (16, 4))\nfor num, x in enumerate(dset_train):\n    data, label = x\n    plt.subplot(1,6,num+1)\n    plt.axis('off')\n    print(x)\n    plt.imshow(np.asarray(data))\n    plt.title(y_train[num])\n<\/code><\/pre>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<blockquote>\n  <p>Why I have to unpack x?<\/p>\n<\/blockquote>\n\n<p>You're inheriting from PyTorch's <code>Dataset<\/code>, and according to <a href=\"https:\/\/pytorch.org\/docs\/stable\/data.html#torch.utils.data.Dataset\" rel=\"nofollow noreferrer\">docs<\/a>:<\/p>\n\n<blockquote>\n  <p>All datasets that represent a map from keys to data samples should subclass it. All subclasses should overrite <code>__getitem__()<\/code>, supporting fetching a data sample for a given key.<\/p>\n<\/blockquote>\n\n<p>In your defined <code>DatasetProcessing<\/code> class <code>__getitem__()<\/code> returns a tuple of 2 items: <code>self.transform(self.input_data[index])<\/code> and <code>self.output_data[index]<\/code>, the first one is data, the second one is appropriate label. And that's why you need to unpack it like <code>data, y = x<\/code>, because your <code>DatasetProcessing<\/code> dataset yields data and a label.<\/p>\n\n<blockquote>\n  <p>Is there any documentation\/tutorials you can link me to?<\/p>\n<\/blockquote>\n\n<p>I can recommend you this links:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/data_loading_tutorial.html\" rel=\"nofollow noreferrer\">Data Loading and Processing Tutorial<\/a><\/li>\n<li><a href=\"https:\/\/pytorch.org\/docs\/stable\/data.html#torch.utils.data.Dataset\" rel=\"nofollow noreferrer\">Dataset docs<\/a> <\/li>\n<li><a href=\"https:\/\/pytorch.org\/docs\/stable\/data.html\" rel=\"nofollow noreferrer\"><code>torch.utils.data<\/code> docs<\/a><\/li>\n<\/ul>\n"}]}
+{"QuestionId":57943425,"AnswerCount":2,"Tags":"<python><machine-learning><keras><deep-learning><conv-neural-network>","CreationDate":"2019-09-15T10:53:49.690","AcceptedAnswerId":57943566.0,"OwnerUserId":8211505.0,"Title":"CNN architecture: classifying \"good\" and \"bad\" images","Body":"<p>I'm researching the possibility of implementing a CNN in order to classify images as \"good\" or \"bad\" but am having no luck with my current architecture.<\/p>\n\n<p>Characteristics that denote a \"bad\" image:<\/p>\n\n<ul>\n<li>Overexposure<\/li>\n<li>Oversaturation<\/li>\n<li>Incorrect white balance<\/li>\n<li>Blurriness<\/li>\n<\/ul>\n\n<p>Would it be feasible to implement a neural network to classify images based on these characteristics or is it best left to a traditional algorithm that simply looks at the variance in brightness\/contrast throughout an image and classifies it that way?<\/p>\n\n<p>I have attempted training a CNN using the VGGNet architecture but I always seem to get a biased and unreliable model, regardless of the number of epochs or number of steps.<\/p>\n\n<p>Examples:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/BXdtL.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/BXdtL.jpg\" alt=\"&quot;Good&quot; Photo\"><\/a><a href=\"https:\/\/i.stack.imgur.com\/Syt0b.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Syt0b.jpg\" alt=\"&quot;Bad&quot; Photo\"><\/a><\/p>\n\n<p>My current model's architecture is very simple (as I am new to the whole machine learning world) but seemed to work fine with other classification problems, and I have modified it slightly to work better with this binary classification problem:<\/p>\n\n<pre><code>    # CONV =&gt; RELU =&gt; POOL layer set\n    # define convolutional layers, use \"ReLU\" activation function\n    # and reduce the spatial size (width and height) with pool layers\n    model.add(Conv2D(32, (3, 3), padding=\"same\", input_shape=input_shape)) # 32 3x3 filters (height, width, depth)\n    model.add(Activation(\"relu\"))\n    model.add(BatchNormalization(axis=channel_dimension))\n    model.add(MaxPooling2D(pool_size=(2,2)))\n    model.add(Dropout(0.25)) # helps prevent overfitting (25% of neurons disconnected randomly)\n\n    # (CONV =&gt; RELU) * 2 =&gt; POOL layer set (increasing number of layers as you go deeper into CNN)\n    model.add(Conv2D(64, (3, 3), padding=\"same\", input_shape=input_shape)) # 64 3x3 filters\n    model.add(Activation(\"relu\"))\n    model.add(BatchNormalization(axis=channel_dimension))\n    model.add(Conv2D(64, (3, 3), padding=\"same\", input_shape=input_shape)) # 64 3x3 filters\n    model.add(Activation(\"relu\"))\n    model.add(BatchNormalization(axis=channel_dimension))\n    model.add(MaxPooling2D(pool_size=(2,2)))\n    model.add(Dropout(0.25)) # helps prevent overfitting (25% of neurons disconnected randomly)\n\n    # (CONV =&gt; RELU) * 3 =&gt; POOL layer set (input volume size becoming smaller and smaller)\n    model.add(Conv2D(128, (3, 3), padding=\"same\", input_shape=input_shape)) # 128 3x3 filters\n    model.add(Activation(\"relu\"))\n    model.add(BatchNormalization(axis=channel_dimension))\n    model.add(Conv2D(128, (3, 3), padding=\"same\", input_shape=input_shape)) # 128 3x3 filters\n    model.add(Activation(\"relu\"))\n    model.add(BatchNormalization(axis=channel_dimension))\n    model.add(Conv2D(128, (3, 3), padding=\"same\", input_shape=input_shape)) # 128 3x3 filters\n    model.add(Activation(\"relu\"))\n    model.add(BatchNormalization(axis=channel_dimension))\n    model.add(MaxPooling2D(pool_size=(2,2)))\n    model.add(Dropout(0.25)) # helps prevent overfitting (25% of neurons disconnected randomly)\n\n    # only set of FC =&gt; RELU layers\n    model.add(Flatten())\n    model.add(Dense(512))\n    model.add(Activation(\"relu\"))\n    model.add(BatchNormalization())\n    model.add(Dropout(0.5))\n\n    # sigmoid classifier (output layer)\n    model.add(Dense(classes))\n    model.add(Activation(\"sigmoid\"))\n<\/code><\/pre>\n\n<p>Is there any glaring omissions or mistakes with this model or can I simply not solve this problem using deep learning (with my current GPU, a GTX 970)?<\/p>\n\n<p>Thanks for your time and experience,<\/p>\n\n<p>Josh<\/p>\n\n<p>EDIT:\nHere is my code for compiling\/training the model:<\/p>\n\n<pre><code># initialise the model and optimiser\nprint(\"[INFO] Training network...\")\nopt = SGD(lr=initial_lr, decay=initial_lr \/ epochs)\nmodel.compile(loss=\"sparse_categorical_crossentropy\", optimizer=opt, metrics=[\"accuracy\"])\n\n# set up checkpoints\nmodel_name = \"output\/50_epochs_{epoch:02d}_{val_acc:.2f}.model\"\ncheckpoint = ModelCheckpoint(model_name, monitor='val_acc', verbose=1, \nsave_best_only=True, mode='max')\nreduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.2,\n                          patience=5, min_lr=0.001)\ntensorboard = TensorBoard(log_dir=\"logs\/{}\".format(time()))\ncallbacks_list = [checkpoint, reduce_lr, tensorboard]\n\n# train the network\nH = model.fit_generator(training_set, steps_per_epoch=500, epochs=50, validation_data=test_set, validation_steps=150, callbacks=callbacks_list)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57943566","CreationDate":"2019-09-15T11:13:56.607","ParentId":null,"OwnerUserId":"7331631","Title":null,"Body":"<p>I suggest you go for transfer learning instead of training the whole network.\nuse the weights trained on a huge Dataset like <a href=\"http:\/\/www.image-net.org\/\" rel=\"nofollow noreferrer\">ImageNet<\/a>  <\/p>\n\n<p>you can easily do this using Keras you just need to import model with weights like <strong>xception<\/strong> and remove last layer which represents 1000 classes of imagenet dataset to 2 node dense layer cause you have only 2 classes and set <code>trainable=False<\/code> for the base layer and <code>trainable=True<\/code> for custom added layers like dense layer having node = 2.<\/p>\n\n<p>and you can train the model as usual way.<\/p>\n\n<p>Demo code - <\/p>\n\n<pre><code>from keras.applications import *\nfrom keras.models import Model\n\nbase_model = Xception(input_shape=(img_width, img_height, 3), weights='imagenet', include_top=False\nx = base_model.output\nx = GlobalAveragePooling2D()(x)\npredictions = Dense(2, activation='softmax')(x)\nmodel = Model(base_model.input, predictions)\n# freezing the base layer weights\nfor layer in base_model.layers:\n    layer.trainable = False\n<\/code><\/pre>\n"},{"AnswerId":"57943674","CreationDate":"2019-09-15T11:28:23.277","ParentId":null,"OwnerUserId":"4685471","Title":null,"Body":"<p>Independently of any other advice (including the answer already provided), and assuming <code>classes=2<\/code> (which you don't clarify - there is a reason we ask for a <a href=\"https:\/\/stackoverflow.com\/help\/minimal-reproducible-example\">MCVE<\/a> here), you seem to perform a fundamental mistake in your final layer, i.e.:<\/p>\n\n<pre><code># sigmoid classifier (output layer)\nmodel.add(Dense(classes))\nmodel.add(Activation(\"sigmoid\"))\n<\/code><\/pre>\n\n<p>A sigmoid activation is suitable only if your final layer consists of a <strong>single<\/strong> node; if <code>classes=2<\/code>, as I suspect, based also on your puzzling statement in the comments that <\/p>\n\n<blockquote>\n  <p>with three different images, my results are 0.987 bad and 0.999 good <\/p>\n<\/blockquote>\n\n<p>and <\/p>\n\n<blockquote>\n  <p>I was giving you the predictions from the model previously<\/p>\n<\/blockquote>\n\n<p>you should use a <strong>softmax<\/strong> activation, i.e.<\/p>\n\n<pre><code>model.add(Dense(classes))\nmodel.add(Activation(\"softmax\"))\n<\/code><\/pre>\n\n<p>Alternatively, you could use sigmoid, but your final layer should consist of a single node, i.e.<\/p>\n\n<pre><code>model.add(Dense(1))\nmodel.add(Activation(\"sigmoid\"))\n<\/code><\/pre>\n\n<p>The latter is usually preferred in binary classification settings, but the results should be the same in principle.<\/p>\n\n<p><strong>UPDATE<\/strong> (after updating the question):<\/p>\n\n<p><code>sparse_categorical_crossentropy<\/code> is not the correct loss here, either.<\/p>\n\n<p>All in all, try the following changes:<\/p>\n\n<pre><code>model.compile(loss=\"binary_crossentropy\", optimizer=Adam(), metrics=[\"accuracy\"])\n\n# final layer:\nmodel.add(Dense(1))\nmodel.add(Activation(\"sigmoid\"))\n<\/code><\/pre>\n\n<p>with Adam optimizer (needs import). Also, dropout should not be used <em>by default<\/em> - see <a href=\"https:\/\/stackoverflow.com\/questions\/57894274\/reducing-versus-delaying-overfitting-in-neural-network\">this thread<\/a>; start without it and only add if necessary (i.e. if you see signs of <a href=\"https:\/\/stackoverflow.com\/questions\/56780450\/is-this-overfitting\">overfitting<\/a>).<\/p>\n"}]}
+{"QuestionId":57943785,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-15T11:45:41.440","AcceptedAnswerId":null,"OwnerUserId":7205164.0,"Title":"tensorflow vs keras execution","Body":"<p>I use this code to build a regression model<\/p>\n\n<pre><code>training_input_func = tf.estimator.inputs.pandas_input_fn( x=x_train,\n                                                  y=y_train['Price'],\n                                                  batch_size=256,\n                                                  num_epochs=500,\n                                                  shuffle=True )\nregressor = tf.estimator.DNNRegressor(feature_columns = feature_cols, \n                                          activation_fn = tf.nn.relu,\n                                          hidden_units=[100, 50, 100],\n                                          model_dir = 'model',\n                                          optimizer = tf.train.GradientDescentOptimizer( learning_rate= 0.01 ))\nregressor.train(input_fn = training_input_func, steps=2000)\n<\/code><\/pre>\n\n<p>It takes 2-3 min to execute, But when i try this code using keras<\/p>\n\n<pre><code>epochs = 500\nbatch_size = 256\nmodel_1 = keras.Sequential()\nmodel_1.add(Dense(100, activation =\"tanh\"))\nmodel_1.add(Dense(50, activation =\"relu\"))\nmodel_1.add(Dense(y_train_array.shape[0]))\nmodel_1.compile(loss='mean_squared_error', optimizer=Adam(), metrics=[metrics.mae])\nmodel_1.fit(x_train_array, y_train_array,\n        batch_size=batch_size,\n        epochs=epochs,\n        shuffle=True,\n        verbose=2, # Change it to 2, if wished to observe execution.  o if not\n        validation_data=(x_validation_array, y_validation_array),\n        callbacks=keras_callbacks,\n        use_multiprocessing=True,\n        workers=50)\n<\/code><\/pre>\n\n<p>It takes almost 3-4 hour for first epoch. The size of training example is around 3M and validation is 30 k. Is there any problem in my code? I know keras take more time time compare to tensorflow. <\/p>\n","answers":[]}
+{"QuestionId":57943845,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><computer-vision><pytorch>","CreationDate":"2019-09-15T11:53:12.193","AcceptedAnswerId":null,"OwnerUserId":1424739.0,"Title":"How a robust background removal is implemented?","Body":"<p>I found that a deep-learning-based method (e.g., 1) is much more robust than a non-deep-learning-based method (e.g., 2, using OpenCV).<\/p>\n\n<ol>\n<li><a href=\"https:\/\/www.remove.bg\" rel=\"nofollow noreferrer\">https:\/\/www.remove.bg<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/29313667\/how-do-i-remove-the-background-from-this-kind-of-image\">How do I remove the background from this kind of image?<\/a><\/li>\n<\/ol>\n\n<p>In the OpenCV example, Canny is used to detect the edges. But this step can be very sensitive to the image. The contour detection may end up with wrong contours. It is also difficult to determine wich contours should be kept.<\/p>\n\n<p>How a robust deep-learning method is implemented? Is any good example code? Thanks.<\/p>\n","answers":[{"AnswerId":"57945285","CreationDate":"2019-09-15T14:57:12.907","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>For that to work you need to use Unet. You can search for that on github.\nUnet transofrm is: I->I.\nSpace of the image will become image (of same or similar size). <\/p>\n\n<p>You need to have say 10.000 images with bg removed. People, (long hair people), cats, cars, shoes, T-shirts, etc. <\/p>\n\n<p>So you set different backgrounds on all these images as source and prediction should be images with removed background.<\/p>\n\n<p>You can also do a <a href=\"https:\/\/pytorch.org\/hub\/pytorch_vision_deeplabv3_resnet101\/\" rel=\"nofollow noreferrer\">segmentation model<\/a> and when you find the foreground you can remove the bg.<\/p>\n"}]}
+{"QuestionId":57943944,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning>","CreationDate":"2019-09-15T12:05:43.067","AcceptedAnswerId":null,"OwnerUserId":6716438.0,"Title":"Is there any function in TensorFlow equivalent to python reduce?","Body":"<p>I want a function in TensorFlow which has the same effect as Pythion <code>reduce()<\/code><\/p>\n\n<p>For example, if I have a tensor <code>a<\/code> with value <code>[a1, a2, a3]<\/code> and a function <code>func()<\/code>, I want <code>[func(a1, a2), func(func(a1, a2), a3)]<\/code>. If <code>a<\/code> is a Python list I can simply do <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from functools import reduce\n# a = [a1, a2, a3]\nA = reduce(func, a)\n<\/code><\/pre>\n\n<p>What if <code>a<\/code> is a Tensor (not a python list!) of TF object and I want the same A? If no alternative TF function, how can I do it effectively?<\/p>\n","answers":[{"AnswerId":"57944316","CreationDate":"2019-09-15T12:56:36.727","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>You have not mentioned which type of <code>func<\/code> you would like to achieve, i.e whether you just wanted to simply add all tensors, or take mean etc. For those type of operations there is <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/math\/reduce_sum\" rel=\"nofollow noreferrer\">tf.reduce_sum<\/a>, and set of other reduce operations. But if you want to have your own custom operation to be implemented by means of reduce, you can achieve it with <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/py_function\" rel=\"nofollow noreferrer\">tf.py_function<\/a>. Check the code below: <\/p>\n\n<pre><code>import tensorflow as tf\nfrom functools import reduce\n\ntf.enable_eager_execution()\n\nx = tf.convert_to_tensor(tf.random.normal([1, 4]))\ny = tf.convert_to_tensor(tf.random.normal([1, 4]))\nz = tf.convert_to_tensor(tf.random.normal([1, 4]))\n\nprint('X = ', x.numpy())\nprint('Y = ', y.numpy())\nprint('Z = ', z.numpy())\n\ndef custom_function(*elements):\n    data = reduce(lambda x, y: x + y, elements)\n    return data\n\n\nresult = tf.py_function(custom_function, [x, y, z], Tout=tf.float32)\nprint('Result = ', result.numpy())\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>X =  [[ 0.9403639  -0.41230875  0.08281742  0.24365784]]\nY =  [[-1.2166423  -0.18330204 -1.2322044  -1.1254508 ]]\nZ =  [[-0.42474216  0.997245   -2.319666    1.4623599 ]]\nResult =  [[-0.70102054  0.40163422 -3.4690528   0.5805669 ]]\n<\/code><\/pre>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>The output of any <code>reduce<\/code> function in Python has to be a single value. In your case, you wish to accumulate the values of subsequent elements in a tensor and in the end produce another 1 dimensional Tensor. <\/p>\n\n<p>I can't think of any built-in TensorFlow equivalent operation which can be used in this way. But, we can do the processing by indexing of tensor elements. However, I am not sure, how much this solution can integrate with your overall model architecture. Check the code below:<\/p>\n\n<pre><code>import tensorflow as tf\ntf.enable_eager_execution()\n\nx = tf.convert_to_tensor(tf.random.normal([5]))\nprint('X = ', x.numpy())\n\ny = []\nsum = x[0]\nfor index in range(1, len(x)):\n    sum += x[index]\n    y.append(sum)\ny = tf.convert_to_tensor(y)\nprint('Sum = ', y.numpy())\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>X =  [ 1.0281714  -1.5870513  -1.7896062   0.11207424  0.50049555]\nSum =  [-0.55887985 -2.348486   -2.2364118  -1.7359163 ]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57944382,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2019-09-15T13:06:29.603","AcceptedAnswerId":null,"OwnerUserId":12070086.0,"Title":"I am getting UserWarning: Update your `Model` call to the Keras 2 API: `Model(inputs=Tensor(\"in..., outputs=Tensor(\"de...)`","Body":"<p>I am working on a dog cat classifier and I found these code on GitHub <a href=\"https:\/\/github.com\/bradleypallen\/keras-dogs-vs-cats\/blob\/master\/keras-dogs-vs-cats-vgg16-transfer.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/bradleypallen\/keras-dogs-vs-cats\/blob\/master\/keras-dogs-vs-cats-vgg16-transfer.ipynb<\/a>. But I am getting an error in the code mentioned below.<\/p>\n\n<p>UserWarning: Update your <code>Model<\/code> call to the Keras 2 API: <code>Model(inputs=Tensor(\"in..., outputs=Tensor(\"de...)<\/code><\/p>\n\n<p>So, in the end, I am getting an error\n'ProgbarLogger' object has no attribute 'log_values'<\/p>\n\n<p>Actually I am new to keras programming so I could not try anything else.<\/p>\n\n<pre><code>input=Input(shape=(IMG_WIDTH,IMG_HEIGHT,3))\noutput_vgg16_conv=model_vgg16_conv(input)\nx=Flatten()(output_vgg16_conv)\nx=Dense(64,activation=\"relu\")(x)\nx=Dense(2,activation=\"softmax\")(x)\nmodel=Model(input=input, output=x) \n<\/code><\/pre>\n\n<p>I am getting a warning in the last line<\/p>\n\n<p>'ProgbarLogger' object has no attribute 'log_values' I am getting this error and as I googled this, people were talking about degenerate input and I have no idea about it. So anyone please help me with it<\/p>\n","answers":[]}
+{"QuestionId":57944774,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-09-15T13:55:26.240","AcceptedAnswerId":null,"OwnerUserId":12070181.0,"Title":"How to slice a torch tensor without using python indexing","Body":"<p>My pytorch code below keeps getting jit tracer warning (in pytorch 1.1.0 environment) complaining that \"Pytorch 1.0 Tracer Warning: Converting a tensor to a Python index might ...\"<\/p>\n\n<p>Is there a way to implement the code line marked (A) below without using python indexing?<\/p>\n\n<pre><code>N,C,H,W = input.size()\nCout=4*C\nHout=H\/\/2\nWout=W\/\/2\ndownsampled=torch.zeros([N,Cout,Hout,Wout], dtype= torch.FloatTensor)\ndownsampled[:,1:Cout:4,:,:]=input[:,:,0::2,1::2] ---- (A)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57950995","CreationDate":"2019-09-16T05:47:39.930","ParentId":null,"OwnerUserId":"12070181","Title":null,"Body":"<p>I confirmed that jit tracer no longer complains the python indexing in Pytorch 1.2 (as Umang Gupta commented).<\/p>\n\n<p>BTW, I came up with an implementation with no slicing (but still using indexing) as follow:<\/p>\n\n<pre><code>import torch\n\ninput=torch.arange(100)\ninput=input.view(10,10)\ninput=input[None, None, ...].expand(2,3,10,10) #torch.Size([2,3,10,10])\n\nN,C,H,W=input.size()\nCout=4*C\nHout=H\/\/2\nWout=W\/\/2\n\ndownsampled=torch.zeros([N,Cout,Hout,Wout],dtype=torch.int8) #torch.Size([2,12,5,5])\n\ndim2_idx=torch.tensor([k for k in range(0,H,2)])\ndim3_idx=torch.tensor([k for k in range(1,W,2)])\nsliced_input=input.index_select(2,dim2_idx).index_select(3,dim3_idx) #torch.Size([2,3,5,5])\n\n#downsampled.index_select(1,torch.tensor([k for k in range(1,Cout,4)]))=temp &lt;---Error: Can't assign to function call\n\nfor idx in range(1,Cout,4):\n    downsampled[:,idx,:,:]=sliced_input[:,idx\/\/4,:,:]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57944786,"AnswerCount":0,"Tags":"<python><c><tensorflow><keras>","CreationDate":"2019-09-15T13:56:23.817","AcceptedAnswerId":null,"OwnerUserId":4045358.0,"Title":"keras.models.load_model fails with 'tags' = train","Body":"<p>I am exploring the c API for tensorflow 2.0. <\/p>\n\n<p>Problem:\nWhen loading model into python,weights are not restored and thus model seems to be untrained.<\/p>\n\n<p>Workflow:\nI am using TF 2.0 C api to handle the training of my model.\nThe general setup I follow is:<\/p>\n\n<p><strong>1. Define model in python using TF keras api.<\/strong> <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nfrom tensorflow import keras\n\nmodel = keras.Sequential([keras.layers.Dense(128,\n                                             input_shape=(784,),\n                                             activation='relu'),\n                          keras.layers.Dense(10, activation='softmax')\n                        ])\nmodel.compile(optimizer='adam',\n              loss=\"categorical_crossentropy\",\n              metrics=['accuracy'])\n\nkeras.experimental.export_saved_model(model,\"keras_model\")\n<\/code><\/pre>\n\n<p>I am using keras.experimental.export_saved_model() because I need the \"signature_def['train']\" which is not saved when using keras.Model.save().<\/p>\n\n<p><strong>2. Train model in C using TF 2.0 C api<\/strong>\nSaved model is then loaded into my C program via:<\/p>\n\n<pre class=\"lang-c prettyprint-override\"><code>TF_LoadSessionFromSavedModel()\n<\/code><\/pre>\n\n<p>Subsequently it is trained and checkpoints are saved with:<\/p>\n\n<pre class=\"lang-c prettyprint-override\"><code>TF_SessionRun()\n<\/code><\/pre>\n\n<p>Saving the model creates new checkpoints files (\"checkpoint.index\" and \"checkpoint.data-00000-of-00001\") in the \"variables\" folder of where my model is stored.<\/p>\n\n<p><strong>3. PROBLEM Reload model in python<\/strong>\nAfter training I reload my model in python. This is where I find out that the loaded model has wights corresponding to the untrained model. I know this because predictions are giberish when my trained model in C was predicting accurately.\nI load my model via:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nfrom tensorflow import keras\n\nmodel = keras.experimental.load_from_saved_model(\"keras_model\")\n<\/code><\/pre>\n\n<p>Again, I am using keras.experimental.load_from_saved_model() because when I use keras.models.load_model() I get the following ValueError:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>ValueError: Importing a SavedModel with tf.saved_model.load requires a 'tags=' argument if there is more than one MetaGraph. Got 'tags=None', but there are 3 MetaGraphs in the SavedModel with tag sets [['train'], ['eval'], ['serve']]. Pass a 'tags=' argument to load this SavedModel.\n<\/code><\/pre>\n\n<p>If I pass a \"tags=serve\" argument to keras.model.load_model() I get the following TypeError:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>TypeError: load_model() got an unexpected keyword argument 'tags'\n<\/code><\/pre>\n\n<p>Trying to load the saved checkpoints file into my model via keras.Model.load_weights() results in an OSErro due to the file format:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>h5py\/_objects.pyx in h5py._objects.with_phil.wrapper()\nh5py\/_objects.pyx in h5py._objects.with_phil.wrapper()\nh5py\/h5f.pyx in h5py.h5f.open()\nOSError: Unable to open file (file signature not found)\n<\/code><\/pre>\n\n<p><strong>Questions<\/strong>\nWhat is the correct way to load a trained model via it's checkpoints file?<\/p>\n\n<p>How do I save a model <strong>not<\/strong> using keras.experimental.export_saved_model and still be able to access the signature_def['train']?<\/p>\n","answers":[]}
+{"QuestionId":57944938,"AnswerCount":0,"Tags":"<nlp><pytorch><bert-language-model>","CreationDate":"2019-09-15T14:14:39.330","AcceptedAnswerId":null,"OwnerUserId":6151940.0,"Title":"Accuracy not increasing with BERT Large model","Body":"<p>I used both <code>BERT_base_cased<\/code> and <code>BERT_large_Cased<\/code> model for multi class text classification. With BERT_base_cased, I got satisfactory results. When I tried with BERT_large_cased model, the accuracy is same for all the epochs<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/T31Yg.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/T31Yg.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/CTiiO.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/CTiiO.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>With <code>BERT_base_cased<\/code>, there is no such problem. But with <code>BERT_large_cased<\/code>, why accuracy is same in all the epochs? Any help is really appreciated.............<\/p>\n","answers":[]}
+{"QuestionId":57945356,"AnswerCount":1,"Tags":"<python><neural-network><pytorch><reinforcement-learning>","CreationDate":"2019-09-15T15:05:33.547","AcceptedAnswerId":null,"OwnerUserId":2683985.0,"Title":"How can I disable gradient updates for some modules in autograd backpropagation?","Body":"<p>I'm building a multi-model neural network for reinforcement learning to include an action network, a world model network, and a critic. The idea is train the world model to emulate whatever simulation you are trying to master based on input from the action network and the previous state, to train the critic to maximize the Bellman equation (total reinforcement over time) based on the world model output, and then backpropagate the critic value through the world model to provide gradient targets for training the actions. So - from some state, the action network outputs an action which is fed into the model to generate the next state, and that state feeds into the critic network for evaluation against some goal state.<\/p>\n\n<p>For all this to work, I must use 3 separate loss functions, one for each network, and they all add something to the gradients in one or more networks but they can be in conflict. For example - to train the world model I use a target from an environmental simulation and for the critic I use a target of the current state reward + discount * next state forecast value. However, to train the a actor I just use the negative critic value as a loss and backpropagate all the way through all three models to calibrate the best action. <\/p>\n\n<p>I can make this work without any batching by zeroing out gradients incrementally, but that is inefficient and doesn't let me accumulate gradients for any kind of \"time-series batching\" optimizer update step. Each model has it's own trainable parameters, but the execution graph flows through all three networks. So inside the calibration loop after firing the networks in sequence:\n...<\/p>\n\n<pre><code>        if self.actor.calibrating:\n            self.actor.optimizer.zero_grad()\n            #Pick loss For maximizing the value of all actions\n            loss = -self.critic.value\n            #Backpropagate through all three networks to train actor output\n            #How do I stop the critic and model networks from incrementing their gradient values?\n            loss.backward(retain_graph=True)\n            self.actor.optimizer.step()\n        if self.model.calibrating:\n            self.model.optimizer.zero_grad()\n            #Reduce loss for ambiguous actions\n            loss = self.model.get_loss() * self.actor.get_confidence()**2\n            #How can I block this from backpropagating through action network?\n            loss.backward(retain_graph=True)\n            self.model.optimizer.step()\n        if self.critic.calibrating:\n            self.critic.optimizer.zero_grad()\n            #Reduce loss for ambiguous actions\n            loss = self.critic.get_loss(self.goal) * self.actor.get_confidence()**2\n            #How do I stop this from backpropagating through the model and action networks?\n            loss.backward(retain_graph=True)\n            self.critic.optimizer.step()\n<\/code><\/pre>\n\n<p>...<\/p>\n\n<p>Finally - my question is in two parts:<\/p>\n\n<ol>\n<li>How can I temporarily stop loss.backward() at a given layer without detaching it forever?<\/li>\n<li>How can I block loss.backward() from updating some gradients where I'm just flowing through a model to get gradients for another model?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"57958534","CreationDate":"2019-09-16T14:02:26.383","ParentId":null,"OwnerUserId":"2683985","Title":null,"Body":"<p>Got this figured out thanks to a suggestion from a colleague to try the requires_grad setting. (I had assumed that would break the execution graph, but it doesn't)<\/p>\n\n<p>So - to answer my own two questions:<\/p>\n\n<ol>\n<li>If you calibrate the chained models in the correct order, you can detach them one at a time so that loss.backward() doesn't run over models that aren't needed. I was thinking that this would break the graph but... this is Pytorch, not Tensorflow 1.x and the graph is regenerated on every forward pass anyway. Silly me for missing this yesterday.<\/li>\n<li>If you set requires_grad to False for a model (or a layer or an individual weight) then loss.backward() will STILL traverse the entire connected graph but it will leave those individual gradients as they were while still setting any gradients earlier in the graph. Exactly what I wanted.<\/li>\n<\/ol>\n\n<p>This code works to minimize the execution of unnecessary graph traversals and gradient updates. I still need to refactor it for staggered updates over time so that it can accumulate gradients for several cycles before stepping the optimizers, but this definitely works as intended.<\/p>\n\n<pre><code>#Step through all models in a chain to create gradient paths from critic back through the world model, to the actor.\n    def step(self):\n        #Get the current state from the simulation\n        state = self.world.state\n        #Fire the actor to select a softmax action.\n        self.actor(state)\n        #run the world simulation on that action.\n        self.world.step(self.actor.action)\n        #Combine the action and starting state as input to the world model.\n        if self.actor.calibrating:\n            action_state = torch.cat([self.actor.value, state], dim=0)\n        else:\n            #Push softmax action closer to 1.0\n            action_state = torch.cat([self.actor.hard_value, state], dim=0)\n        #Run the model and then the critic on the action_state\n        self.critic(self.model(action_state))\n        if self.actor.calibrating:\n            self.actor.optimizer.zero_grad()\n            self.model.requires_grad = False\n            self.critic.requires_grad = False\n            #Pick loss For maximizing the value of the action choice\n            loss = -self.critic.value * self.actor.get_confidence()\n            loss.backward(retain_graph=True)\n            self.actor.optimizer.step()\n        if self.model.calibrating:\n            #Don't need to backpropagate through actor again\n            self.actor.value.detach_()\n            self.model.optimizer.zero_grad()\n            self.model.requires_grad = True\n            #Reduce loss for ambiguous actions\n            loss = self.model.get_loss() * self.actor.get_confidence()**2\n            loss.backward(retain_graph=True)\n            self.model.optimizer.step()\n        if self.critic.calibrating:\n            #Don't need to backpropagate through the model or actor again\n            self.model.value.detach_()\n            self.critic.optimizer.zero_grad()\n            self.critic.requires_grad = True\n            #Reduce loss for ambiguous actions\n            loss = self.critic.get_loss(self.goal) * self.actor.get_confidence()**2\n            loss.backward(retain_graph=True)\n            self.critic.optimizer.step()\n<\/code><\/pre>\n"}]}
+{"QuestionId":57945364,"AnswerCount":1,"Tags":"<python><keras><deep-learning>","CreationDate":"2019-09-15T15:06:51.453","AcceptedAnswerId":57947146.0,"OwnerUserId":397192.0,"Title":"Keras fit_generator : Error when checking target: expected activation_32 to have 4 dimensions, but got array with shape (4, 4)","Body":"<p>I'm getting the error below:<\/p>\n\n<pre><code>  **File \"C:\\Users\\sergi\\Anaconda3\\envs\\tensorpy36gpu\\lib\\site-packages\\keras\\engine\\training_utils.py\", line 131, in standardize_input_data\n    'with shape ' + str(data_shape))\n\nValueError: Error when checking target: expected activation_32 to have 4 dimensions, but got array with shape (4, 4)**\n<\/code><\/pre>\n\n<p>when I run fit_generator:<\/p>\n\n<pre><code>fit = model.fit_generator(trainX, steps_per_epoch=len(trainX) \/\/ batch_size, validation_data=trainY,\n                          validation_steps=len(trainY) \/\/ batch_size, epochs=10, verbose=2)\n<\/code><\/pre>\n\n<p>The code complete:<\/p>\n\n<pre><code> def model(rows, cols, channels): #channels_last\n\n   model = Sequential()        \n   model.add(Conv2D(32, (3, 3), padding='same', input_shape=(rows, cols, channels)))\n   model.add(Activation('relu'))\n   model.add(MaxPooling2D(pool_size=(2,2)))\n   model.add(Conv2D(64, (5, 5), padding='same'))\n   model.add(Activation('relu'))\n   model.add(MaxPooling2D(pool_size=(2,2)))\n   model.add(Conv2D(128, (5, 5), padding='same'))\n   model.add(Activation('relu'))\n   model.add(MaxPooling2D(pool_size=(2,2)))\n   model.add(Dense(4))\n   model.add(Activation('softmax'))\n\n   return model\n\ndef summarize_diagnostics(history):\n    # plot loss\n    plt.subplot(211)\n    plt.title('Cross Entropy Loss')\n    plt.plot(history.history['loss'], color='blue', label='train')\n    plt.plot(history.history['val_loss'], color='orange', label='test')\n    # plot accuracy\n    plt.subplot(212)\n    plt.title('Classification Accuracy')\n    plt.plot(history.history['acc'], color='blue', label='train')\n    plt.plot(history.history['val_acc'], color='orange', label='test')\n    plt.show()\n\nopt = SGD(lr=0.001)\n\nmodel = model(196, 196, 3)\nmodel.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy'])\n\ndatagen = ImageDataGenerator(validation_split=0.30)\n\nbatch_size = 16\n\ntrainX = datagen.flow_from_directory(\".\/Train\/\", batch_size=16, target_size=(196, 196), subset='training', \n                                     class_mode = 'categorical')\ntrainY = datagen.flow_from_directory(\".\/Train\/\", batch_size=16, target_size=(196, 196), subset='validation', \n                                     class_mode = 'categorical')\n\nfit = model.fit_generator(trainX, steps_per_epoch=len(trainX) \/\/ batch_size, validation_data=trainY,\n                          validation_steps=len(trainY) \/\/ batch_size, epochs=10, verbose=2)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57947146","CreationDate":"2019-09-15T18:36:17.860","ParentId":null,"OwnerUserId":"10801861","Title":null,"Body":"<p>You have to Flatten data between the Conv layers and Dense layers\n<a href=\"https:\/\/i.stack.imgur.com\/8PmM7.png\" rel=\"nofollow noreferrer\">image<\/a><\/p>\n\n<pre><code>   model.add(Conv2D(128, (5, 5), padding='same'))\n   model.add(Activation('relu'))\n   model.add(MaxPooling2D(pool_size=(2,2)))\n\n   model.add(Flatten())\n\n   model.add(Dense(4))\n   model.add(Activation('softmax'))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57945706,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><save>","CreationDate":"2019-09-15T15:47:11.260","AcceptedAnswerId":null,"OwnerUserId":11335094.0,"Title":"Python Tensorflow- Saving my own neural network","Body":"<p>I am new to machine learning and have tried to implement a fully connectod neural network class in python. Here is my code: <\/p>\n\n<pre><code>class Network():\n\n    def __init__(self, n_x, hiddenLayerDimensions, n_y):\n        self.n_x = n_x\n        self.n_y = n_y\n        self.hidden_layer_dimensions = hiddenLayerDimensions\n        self.num_layers = len(hiddenLayerDimensions) + 1\n        layerDimensions = [n_x] + hiddenLayerDimensions + [n_y]\n        self.W, self.b = {}, {}\n        self.sess = tf.Session()\n        for i in range(1, len(layerDimensions)):\n            self.W[i] = tf.Variable(tf.random_normal([layerDimensions[i], layerDimensions[i-1]], stddev=0.03), name='W'+str(i))\n            self.b[i] = tf.Variable(tf.random_normal([layerDimensions[i],1]), name='b'+str(i))\n\n        self.saver = tf.train.Saver()\n\n    def train(self, trainSetX, trainSetY, batchSize = 100, learning_rate = 0.5, epochs = 10):\n        L = self.num_layers\n        m = trainSetX.shape[1]\n        while m % batchSize != 0:\n            batchSize += 1\n\n        X = tf.placeholder(dtype= tf.float32, shape=[self.n_x, None])\n        Y = tf.placeholder(dtype = tf.float32, shape = [self.n_y, None])\n        A = X\n        for i in range(1, L):\n            A = tf.nn.relu(self.W[i] @ A + self.b[i])\n\n        Z = self.W[L] @ A + self.b[L]\n        cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=tf.transpose(Z), labels=tf.transpose(Y)))\n        optimiser = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(cost)\n\n        init = tf.global_variables_initializer()\n        correct_prediction = tf.equal(tf.argmax(Y, axis=0), tf.argmax(Z, axis=0))\n\n        accuracy = tf.reduce_mean(tf.cast(correct_prediction, \"float\"))\n\n        self.sess.run(init)\n\n        totalBatches = m \/\/ batchSize\n        for epoch in range(epochs):\n            averageCost = 0\n            for batchX, batchY in splitIntoBatches(trainSetX, trainSetY, batchSize):\n                _, c = self.sess.run([optimiser, cost], feed_dict={X:batchX, Y:batchY})\n                averageCost += c\n\n            averageCost \/= totalBatches\n\n            if epoch % 10 == 9:\n                print(\"Cost after epoch {}: {}\".format(epoch + 1, c))\n\n        acc = self.sess.run(accuracy, feed_dict={X:trainSetX, Y:trainSetY})\n        print(\"Finished training. Training accuracy: \" + str(acc))\n\n\n    def test(self, testsetX, testsetY):\n        L = self.num_layers\n        X = tf.placeholder(dtype= tf.float32, shape = [self.n_x, None])\n        Y = tf.placeholder(dtype= tf.float32, shape = [self.n_y, None])\n\n        A = X\n        for i in range(1, L):\n            Z = self.W[i] @ A + self.b[i]\n            A = tf.nn.relu(Z)\n\n        Z = self.W[L] @ A + self.b[L]\n        correct_prediction = tf.equal(tf.argmax(Y, axis=0), tf.argmax(Z, axis=0))\n        accuracy = tf.reduce_mean(tf.cast(correct_prediction, \"float\"))\n        return self.sess.run(accuracy, feed_dict= {X: testsetX, Y:testsetY})\n\n    def classify(self, v):\n        v = v.reshape(self.n_x, 1)\n\n        L = self.num_layers\n        X = tf.placeholder(dtype=tf.float32, shape=[self.n_x, None])\n\n        A = X\n        for i in range(1, L):\n            Z = self.W[i] @ A + self.b[i]\n            A = tf.nn.relu(Z)\n\n        Z = self.W[L] @ A + self.b[L]\n        Z = self.sess.run(Z, feed_dict= {X : v})\n        return np.argmax(Z, axis = 0)[0]\n<\/code><\/pre>\n\n<p>I am looking for some way of saving and restoring trained networks, e.g so I can write code like this:<\/p>\n\n<pre><code>net = Network(...)\n\nnet.train(...)\n\nsave_network(net, path)\n<\/code><\/pre>\n\n<p>then next time I open my computer I could write:<\/p>\n\n<pre><code>restored_net = restore_network(path)\n<\/code><\/pre>\n\n<p>Can you help me write <code>save_network<\/code> and <code>restore_network<\/code> functions? I tried using pickle but I get\n<code>cannot pickle _thread.Rlock objects<\/code> error, and have also tried using <code>tf.train.Saver<\/code>, but not got anything working.<\/p>\n","answers":[]}
+{"QuestionId":57946286,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><conv-neural-network>","CreationDate":"2019-09-15T16:57:01.053","AcceptedAnswerId":null,"OwnerUserId":8301238.0,"Title":"Keras: expected dense_5 to have shape (None, 1) but got array with shape (60000, 10)","Body":"<p>I'm running the following code on python 3.6:<\/p>\n\n<pre><code>(train_images, train_labels), (test_images, test_labels) = mnist.load_data()\ntrain_images = train_images.reshape((60000, 28, 28, 1)).astype('float')\/255\ntest_images = test_images.reshape((10000, 28, 28, 1)).astype('float')\/255\ntrain_labels = ku.to_categorical(train_labels)\ntest_labels = ku.to_categorical(test_labels)\nfilepath = 'my_model_file.hdf5' # define where the model is saved\ncallbacks_list = [\n        keras.callbacks.EarlyStopping(\n            monitor = 'val_loss', # Use accuracy to monitor the model\n            patience = 1 # Stop after one step with lower accuracy\n        ),\n        keras.callbacks.ModelCheckpoint(\n            filepath = filepath, # file where the checkpoint is saved\n            monitor = 'val_loss', # Don't overwrite the saved model unless val_loss is worse\n            save_best_only = True # Only save model if it is the best\n        )\n]\nmodel = models.Sequential()\nmodel.add(layers.Conv2D(32, (3, 3), activation='relu', input_shape=(28, 28, 1)))\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(64, (3, 3), activation='relu'))\nmodel.add(layers.Flatten())\nmodel.add(layers.Dense(64, activation='relu', kernel_regularizer=l2(0.001)))\nmodel.add(Dropout(0.5))\nmodel.add(layers.Dense(10, activation='softmax'))\nmodel.summary()\n## Compile the model\nmodel.compile(optimizer='adam',loss='sparse_categorical_crossentropy',metrics=['accuracy'])\n\n## Now fit the model\nnr.seed(2356)\nset_random_seed(2333)\nhistory = model.fit(train_images, train_labels, epochs=40, batch_size = 128, callbacks = callbacks_list, validation_data = (test_images, test_labels))  # Call backs argument here)\n<\/code><\/pre>\n\n<p>and get the following error: <\/p>\n\n<pre><code>ValueError: Error when checking target: expected dense_5 to have shape (None, 1) but got array with shape (60000, 10)\n<\/code><\/pre>\n\n<p>The previous posts I look talk about datalabel should be convert to categorical data but this is done. Actually most of the code is copy and past from tensor flow tutorial so it's little bit hard to see where things where wrong.<\/p>\n","answers":[{"AnswerId":"57947023","CreationDate":"2019-09-15T18:23:53.303","ParentId":null,"OwnerUserId":"10801861","Title":null,"Body":"<p>Use loss functions categorical_crossentropy<\/p>\n\n<blockquote>\n  <p>loss='categorical_crossentropy'<\/p>\n<\/blockquote>\n\n<p>beacause your label was encoded at <\/p>\n\n<blockquote>\n  <p>train_labels = ku.to_categorical(train_labels)<\/p>\n<\/blockquote>\n\n<p>PS.\nIf your label are endcoded(one-hot encoded) use categorical_crossentropy\nex:<\/p>\n\n<blockquote>\n  <p>[1,0,0]\n  [0,1,0]\n  [0,0,1]<\/p>\n<\/blockquote>\n\n<p>But if your targets are integers, use sparse_categorical_crossentropy\nex:<\/p>\n\n<blockquote>\n  <p>[1,2,3]<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":57946626,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-15T17:37:25.357","AcceptedAnswerId":57958415.0,"OwnerUserId":11013499.0,"Title":"why does this error \" The following are legacy tf.layers.Layers\" happen when I have my custom layer?","Body":"<p>I am using Keras and I have a custom layer but when I use it, the following error happens and I do not know what is the problem. could you please help me with this issue? it is weird when I used the same code on another system, this error does not appear!<\/p>\n\n<pre><code>import os\nimport time\nfrom os import listdir\nfrom os.path import isfile, join\nimport numpy as np\nimport matplotlib\nimport matplotlib.pyplot as plt\nfrom mpl_toolkits.mplot3d import Axes3D\nfrom matplotlib import cm\nimport scipy.io as sio\nfrom scipy.misc import imread\nimport cv2\nimport skimage.transform as imgTrans\nfrom skimage.measure import compare_ssim, compare_psnr\nimport PIL\nfrom PIL import Image\nfrom PIL import ImageFont\nfrom PIL import ImageDraw \nfrom tqdm import tqdm\nimport tensorflow as tf\nkeras = tf.keras\nlayers = keras.layers\nfrom my_circular_layer_new import Conv2D_circular\nimport Various_Functions as vf\nfrom scipy.ndimage.filters import convolve, median_filter\nfrom scipy.ndimage.filters import gaussian_filter\ndef buildModel(model_path, patch_rows=32, patch_cols=32, channels=1, block_size=8 ,num_bitplane=1, use_circular=True):\n\n    conv2d_layer = layers.Conv2D if use_circular == False else Conv2D_circular\n\n    w_rows = int((patch_rows) \/ block_size)\n    w_cols = int((patch_cols) \/ block_size)\n\n    input_img = layers.Input(shape=(patch_rows, patch_cols, 1), name='input_img')\n    input_strenght_alpha = layers.Input(shape=(1,), name='strenght_factor_alpha')\n    input_watermark = layers.Input(shape=(w_rows, w_cols, num_bitplane), name='input_watermark')\n\n    # Rearrange input \n    rearranged_img = l1 = layers.Lambda(tf.space_to_depth, arguments={'block_size':block_size}, name='rearrange_img')(input_img)\n\n\n    dct_layer = layers.Conv2D(64, (1, 1), activation='linear', padding='same', use_bias=False, trainable=False, name='dct1')\n    dct_layer2 = layers.Conv2D(64, (1, 1), activation='linear', padding='same', use_bias=False, trainable=False, name='dct2')\n    idct_layer = layers.Conv2D(64, (1, 1), activation='linear', padding='same', use_bias=False, trainable=False, name='idct')\n    dct_layer_img = dct_layer(rearranged_img)\n\n    # Concatenating The Image's dct coefs and watermark\n    encoder_input = layers.Concatenate(axis=-1, name='encoder_input')([dct_layer_img, input_watermark])\n\n    # Encoder\n    encoder_model = layers.Conv2D(64, (1, 1), dilation_rate=1, activation='elu', padding='same', name='enc_conv1')(encoder_input)\n    encoder_model = conv2d_layer(64, (2, 2), dilation_rate=1, activation='elu', padding='same', name='enc_conv2')(encoder_model)\n    encoder_model = conv2d_layer(64, (2, 2), dilation_rate=1, activation='elu', padding='same', name='enc_conv3')(encoder_model)\n    encoder_model = conv2d_layer(64, (2, 2), dilation_rate=1, activation='elu', padding='same', name='enc_conv4')(encoder_model)\n    encoder_model = conv2d_layer(64, (2, 2), dilation_rate=1, activation='elu', padding='same', name='enc_conv5')(encoder_model)\n    encoder_model = idct_layer(encoder_model)\n\n    # Strength\n    encoder_model = layers.Lambda(multiply_scalar, arguments={'scalar':input_strenght_alpha}, name='strenght_factor')(encoder_model)\n\n    encoder_model = layers.Add(name='residual_add')([encoder_model, l1])\n    encoder_model = x = layers.Lambda(tf.depth_to_space, arguments={'block_size':block_size}, name='enc_output_depth2space')(encoder_model)\n\n    # Attack (The attacks occure in test phase)\n\n    # Watermark decoder\n    input_attacked_img = layers.Input(shape=(patch_rows, patch_cols, 1), name='input_attacked_img')\n    decoder_model = layers.Lambda(tf.space_to_depth, arguments={'block_size':block_size}, name='dec_input_space2depth')(input_attacked_img)\n    decoder_model = dct_layer2(decoder_model)\n    decoder_model = layers.Conv2D(64, (1, 1), dilation_rate=1, activation='elu', padding='same', name='dec_conv1')(decoder_model)\n    decoder_model = conv2d_layer(64, (2, 2), dilation_rate=1, activation='elu', padding='same', name='dec_conv2')(decoder_model)\n    decoder_model = conv2d_layer(64, (2, 2), dilation_rate=1, activation='elu', padding='same', name='dec_conv3')(decoder_model)\n    decoder_model = conv2d_layer(64, (2, 2), dilation_rate=1, activation='elu', padding='same', name='dec_conv4')(decoder_model)\n    decoder_model = layers.Conv2D(num_bitplane, (1, 1), dilation_rate=1, activation='sigmoid', padding='same', name='dec_output_depth2space')(decoder_model)\n\n    # Whole model\n    embedding_net = tf.keras.models.Model(inputs=[input_img, input_watermark, input_strenght_alpha], outputs=[x])\n    extractor_net = tf.keras.models.Model(inputs=[input_attacked_img], outputs=[decoder_model])\n\n    # Set weights\n    DCT_MTX = sio.loadmat('.\/Weights\/Transforms\/DCT_coef.mat')['DCT_coef']\n    dct_mtx = np.reshape(DCT_MTX, [1,1,64,64])\n    embedding_net.get_layer('dct1').set_weights(np.array([dct_mtx]))\n    extractor_net.get_layer('dct2').set_weights(np.array([dct_mtx]))\n\n    IDCT_MTX = sio.loadmat('.\/Weights\/Transforms\/IDCT_coef.mat')['IDCT_coef']\n    idct_mtx = np.reshape(IDCT_MTX, [1,1,64,64])\n    embedding_net.get_layer('idct').set_weights(np.array([idct_mtx]))\n\n    embedding_net.load_weights(model_path,by_name = True)\n    extractor_net.load_weights(model_path,by_name = True)\n    return embedding_net, extractor_net\n\n\n&gt; TypeError: The following are legacy tf.layers.Layers:  \n&gt; &lt;my_circular_layer_new.Conv2D_circular object at 0x000001F773AC2160&gt;  \n&gt; &lt;my_circular_layer_new.Conv2D_circular object at 0x000001F773B1F978&gt;  \n&gt; &lt;my_circular_layer_new.Conv2D_circular object at 0x000001F773B1F9E8&gt;  \n&gt; &lt;my_circular_layer_new.Conv2D_circular object at 0x000001F773B1F630&gt;\n&gt; To use keras as a framework (for instance using the Network, Model, or\n&gt; Sequential classes), please use the tf.keras.layers implementation\n&gt; instead. (Or, if writing custom layers, subclass from tf.keras.layers\n&gt; rather than tf.layers)\n<\/code><\/pre>\n\n<blockquote>\n  <blockquote>\n    <p>TypeError: The following are legacy tf.layers.Layers:<br>\n    <br>\n    <br>\n    <br>\n    \n    To use keras as a framework (for instance using the Network, Model, or\n    Sequential classes), please use the tf.keras.layers implementation\n    instead. (Or, if writing custom layers, subclass from tf.keras.layers\n    rather than tf.layers)<\/p>\n  <\/blockquote>\n<\/blockquote>\n","answers":[{"AnswerId":"57958415","CreationDate":"2019-09-16T13:56:48.253","ParentId":null,"OwnerUserId":"5432214","Title":null,"Body":"<p>I would guess that you have two different versions of the same libraries, and the system that gives you the error has a newer version than the other one. Either downgrade one system or upgrade the other one<\/p>\n"}]}
+{"QuestionId":57946665,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><scikit-learn><neural-network><deep-learning>","CreationDate":"2019-09-15T17:43:00.400","AcceptedAnswerId":null,"OwnerUserId":9706484.0,"Title":"Binary classification using neural network in tensorflow failing to learn","Body":"<p>I am trying to implement a neural network for the heart dataset from kaggle<\/p>\n\n<p><a href=\"https:\/\/www.kaggle.com\/ronitf\/heart-disease-uci\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/ronitf\/heart-disease-uci<\/a><\/p>\n\n<p>Even though I am training the network on the entire dataset it doesn't seem to learn at all with the accuracy staying around 50% , even if it is overfitting a Neural network should be able to get higher in sample accuracy. Can anyone see any issues with this code (i've tried many learning rates , batches and epochs and nothing seems to work.<\/p>\n\n<p>As I understand for binary classification it is better to have 1 output to classify. I can already make this work using two outputs so please don't give answers suggesting this.<\/p>\n\n<pre><code>#need to switch from the binary output to the singular value output\n\nimport pandas\nimport numpy as np\nimport tensorflow as tf\nimport random\nfrom sklearn.preprocessing import normalize\n\n\n\n\n#pulling in data from excel\nsheet = pandas.read_excel(\".\/heart.xlsx\")\nsheet2=sheet.values\n\n#getting my data arrays\nX_data,y_data = np.split(sheet2, [13], axis=1)\ny_data=np.reshape(y_data, [-1])   #need to learn more about why this works\n\nX_data =  normalize(X_data, axis=0, norm='max')\n\n\n\n#random shuffle\n\nind_list = [i for i in range(len(X_data))]\nrandom.shuffle(ind_list)\n\n\nX_data=X_data[ind_list]\ny_data=y_data[ind_list]\n\n\n\n#initialising values\nepochs = 1000\nlearning_rate=0.001\nbatch_size=50\n\n\nX = tf.placeholder(tf.float32,shape=(None,n_inputs),name=\"X\")\ny = tf.placeholder(tf.int64,shape=(None) ,name=\"y\")\n\n\n\n#creating the structure of the neural network\n\n\nwith tf.name_scope(\"dnn\"):\n    input = tf.layers.dense(X,13,name=\"hidden1\" ,activation=tf.nn.relu)\n    input = tf.layers.dense(input,7,name=\"hidden2\" ,activation=tf.nn.relu)\n    logits = tf.layers.dense(input,1,name=\"outputs\")\n\n\n\nwith tf.name_scope(\"loss\"):\n    entropy = tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.cast(y,tf.float32),logits=logits)\n    loss=tf.reduce_mean(entropy,name=\"loss\")\n\nwith tf.name_scope(\"train\"):\n    optimizer=tf.train.AdamOptimizer(learning_rate)\n    training_operation=optimizer.minimize(loss)\n\nwith tf.name_scope(\"accuracy\"):\n    predicted = tf.nn.sigmoid(logits)\n    correct_pred = tf.equal(tf.round(predicted), tf.cast(y,tf.float32))\n    accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))\n\n\n#defining functions\n\ndef shuffle_batch(X, y, batch_size):\n    rnd_idx = np.random.permutation(len(X))\n    n_batches = len(X) \/\/ batch_size\n    for batch_idx in np.array_split(rnd_idx, n_batches):\n        X_batch, y_batch = X[batch_idx], y[batch_idx]\n        yield X_batch, y_batch\n\n\n\n#initialising the graph and running the optimisation\n\ninit= tf.global_variables_initializer()\nsave = tf.train.Saver()\n\n\n\nwith tf.Session() as sess:\n    init.run()\n\n    for epoch in range(epochs):\n\n        for X_batch,y_batch in shuffle_batch(X_data,y_data,batch_size):\n            sess.run(training_operation,feed_dict={X:X_batch,y:y_batch})\n\n        acc_train=accuracy.eval(feed_dict={X:X_data,y:y_data})\n\n        print(\"Epoch\",epoch,\"training acuuracy\",acc_train)\n<\/code><\/pre>\n\n<p>output<\/p>\n\n<pre><code>2019-09-15 07:52:04.662488: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2\nEpoch 0 training acuuracy 0.5089697\nEpoch 1 training acuuracy 0.5089697\nEpoch 2 training acuuracy 0.5089697\nEpoch 3 training acuuracy 0.50838155\nEpoch 4 training acuuracy 0.5089697\nEpoch 5 training acuuracy 0.5092638\nEpoch 6 training acuuracy 0.5101461\nEpoch 7 training acuuracy 0.5101461\nEpoch 8 training acuuracy 0.5101461\nEpoch 9 training acuuracy 0.51073426\nEpoch 10 training acuuracy 0.51102835\nEpoch 11 training acuuracy 0.51102835\nEpoch 12 training acuuracy 0.51073426\nEpoch 13 training acuuracy 0.51132244\nEpoch 14 training acuuracy 0.5122047\nEpoch 15 training acuuracy 0.5124988\nEpoch 16 training acuuracy 0.5127929\nEpoch 17 training acuuracy 0.513087\nEpoch 18 training acuuracy 0.51338106\nEpoch 19 training acuuracy 0.51338106\nEpoch 20 training acuuracy 0.51367515\nEpoch 21 training acuuracy 0.51396924\nEpoch 22 training acuuracy 0.5151456\nEpoch 23 training acuuracy 0.51426333\nEpoch 24 training acuuracy 0.5154397\nEpoch 25 training acuuracy 0.5151456\nEpoch 26 training acuuracy 0.51632196\nEpoch 27 training acuuracy 0.5172042\nEpoch 28 training acuuracy 0.51808643\nEpoch 29 training acuuracy 0.5183805\nEpoch 30 training acuuracy 0.5192628\nEpoch 31 training acuuracy 0.5189687\nEpoch 32 training acuuracy 0.51985097\nEpoch 33 training acuuracy 0.52043915\nEpoch 34 training acuuracy 0.52043915\nEpoch 35 training acuuracy 0.52043915\nEpoch 36 training acuuracy 0.5213214\nEpoch 37 training acuuracy 0.5219096\nEpoch 38 training acuuracy 0.5222037\nEpoch 39 training acuuracy 0.5222037\nEpoch 40 training acuuracy 0.5224978\nEpoch 41 training acuuracy 0.5224978\nEpoch 42 training acuuracy 0.5224978\nEpoch 43 training acuuracy 0.5222037\nEpoch 44 training acuuracy 0.5219096\nEpoch 45 training acuuracy 0.5222037\nEpoch 46 training acuuracy 0.5222037\nEpoch 47 training acuuracy 0.5219096\nEpoch 48 training acuuracy 0.5222037\nEpoch 49 training acuuracy 0.52279186\nEpoch 50 training acuuracy 0.52308595\nEpoch 51 training acuuracy 0.52338004\nEpoch 52 training acuuracy 0.52308595\nEpoch 53 training acuuracy 0.52308595\nEpoch 54 training acuuracy 0.52308595\nEpoch 55 training acuuracy 0.52367413\nEpoch 56 training acuuracy 0.5245564\nEpoch 57 training acuuracy 0.52602684\nEpoch 58 training acuuracy 0.52543867\nEpoch 59 training acuuracy 0.52573276\nEpoch 60 training acuuracy 0.526615\nEpoch 61 training acuuracy 0.52573276\nEpoch 62 training acuuracy 0.5269091\nEpoch 63 training acuuracy 0.5269091\nEpoch 64 training acuuracy 0.52602684\nEpoch 65 training acuuracy 0.526615\nEpoch 66 training acuuracy 0.5277914\nEpoch 67 training acuuracy 0.5272032\nEpoch 68 training acuuracy 0.526615\nEpoch 69 training acuuracy 0.526615\nEpoch 70 training acuuracy 0.526615\nEpoch 71 training acuuracy 0.5269091\nEpoch 72 training acuuracy 0.526615\nEpoch 73 training acuuracy 0.526615\nEpoch 74 training acuuracy 0.5277914\nEpoch 75 training acuuracy 0.52867365\nEpoch 76 training acuuracy 0.5277914\nEpoch 77 training acuuracy 0.5277914\nEpoch 78 training acuuracy 0.52837956\nEpoch 79 training acuuracy 0.52837956\nEpoch 80 training acuuracy 0.52837956\nEpoch 81 training acuuracy 0.5295559\nEpoch 82 training acuuracy 0.52985\nEpoch 83 training acuuracy 0.5304382\nEpoch 84 training acuuracy 0.5301441\nEpoch 85 training acuuracy 0.5301441\nEpoch 86 training acuuracy 0.5307323\nEpoch 87 training acuuracy 0.5304382\nEpoch 88 training acuuracy 0.5304382\nEpoch 89 training acuuracy 0.53132045\nEpoch 90 training acuuracy 0.5322027\nEpoch 91 training acuuracy 0.5322027\nEpoch 92 training acuuracy 0.5322027\nEpoch 93 training acuuracy 0.533085\nEpoch 94 training acuuracy 0.5348495\nEpoch 95 training acuuracy 0.53455544\nEpoch 96 training acuuracy 0.5333791\nEpoch 97 training acuuracy 0.5333791\nEpoch 98 training acuuracy 0.53426135\nEpoch 99 training acuuracy 0.53455544\nEpoch 100 training acuuracy 0.53455544\nEpoch 101 training acuuracy 0.5354377\nEpoch 102 training acuuracy 0.5351436\nEpoch 103 training acuuracy 0.53455544\nEpoch 104 training acuuracy 0.5348495\nEpoch 105 training acuuracy 0.53396726\nEpoch 106 training acuuracy 0.53455544\nEpoch 107 training acuuracy 0.53426135\nEpoch 108 training acuuracy 0.5351436\nEpoch 109 training acuuracy 0.5354377\nEpoch 110 training acuuracy 0.5354377\nEpoch 111 training acuuracy 0.53690815\nEpoch 112 training acuuracy 0.53720224\nEpoch 113 training acuuracy 0.5377904\nEpoch 114 training acuuracy 0.5380845\nEpoch 115 training acuuracy 0.5380845\nEpoch 116 training acuuracy 0.5377904\nEpoch 117 training acuuracy 0.5377904\nEpoch 118 training acuuracy 0.53720224\nEpoch 119 training acuuracy 0.5377904\nEpoch 120 training acuuracy 0.5377904\nEpoch 121 training acuuracy 0.5377904\nEpoch 122 training acuuracy 0.5383786\nEpoch 123 training acuuracy 0.5383786\nEpoch 124 training acuuracy 0.5383786\nEpoch 125 training acuuracy 0.5386727\nEpoch 126 training acuuracy 0.5383786\nEpoch 127 training acuuracy 0.5383786\nEpoch 128 training acuuracy 0.5383786\nEpoch 129 training acuuracy 0.5383786\nEpoch 130 training acuuracy 0.5389668\nEpoch 131 training acuuracy 0.5386727\nEpoch 132 training acuuracy 0.5389668\nEpoch 133 training acuuracy 0.53926086\nEpoch 134 training acuuracy 0.53926086\nEpoch 135 training acuuracy 0.53926086\nEpoch 136 training acuuracy 0.53955495\nEpoch 137 training acuuracy 0.53955495\nEpoch 138 training acuuracy 0.53984904\nEpoch 139 training acuuracy 0.53984904\nEpoch 140 training acuuracy 0.53984904\nEpoch 141 training acuuracy 0.54014313\nEpoch 142 training acuuracy 0.5404372\nEpoch 143 training acuuracy 0.54014313\nEpoch 144 training acuuracy 0.53955495\nEpoch 145 training acuuracy 0.5404372\nEpoch 146 training acuuracy 0.54014313\nEpoch 147 training acuuracy 0.54014313\nEpoch 148 training acuuracy 0.53984904\nEpoch 149 training acuuracy 0.53955495\nEpoch 150 training acuuracy 0.54014313\nEpoch 151 training acuuracy 0.5404372\nEpoch 152 training acuuracy 0.5407313\nEpoch 153 training acuuracy 0.5410254\nEpoch 154 training acuuracy 0.5410254\nEpoch 155 training acuuracy 0.5413195\nEpoch 156 training acuuracy 0.5413195\nEpoch 157 training acuuracy 0.5410254\nEpoch 158 training acuuracy 0.5410254\nEpoch 159 training acuuracy 0.5410254\nEpoch 160 training acuuracy 0.5413195\nEpoch 161 training acuuracy 0.5413195\nEpoch 162 training acuuracy 0.5413195\nEpoch 163 training acuuracy 0.5413195\nEpoch 164 training acuuracy 0.5410254\nEpoch 165 training acuuracy 0.5413195\nEpoch 166 training acuuracy 0.54190767\nEpoch 167 training acuuracy 0.54220176\nEpoch 168 training acuuracy 0.54190767\nEpoch 169 training acuuracy 0.54220176\nEpoch 170 training acuuracy 0.54220176\nEpoch 171 training acuuracy 0.54220176\nEpoch 172 training acuuracy 0.54220176\nEpoch 173 training acuuracy 0.54220176\nEpoch 174 training acuuracy 0.54220176\nEpoch 175 training acuuracy 0.54220176\nEpoch 176 training acuuracy 0.54190767\nEpoch 177 training acuuracy 0.5416136\nEpoch 178 training acuuracy 0.5416136\nEpoch 179 training acuuracy 0.54190767\nEpoch 180 training acuuracy 0.54220176\nEpoch 181 training acuuracy 0.54220176\nEpoch 182 training acuuracy 0.54220176\nEpoch 183 training acuuracy 0.54220176\nEpoch 184 training acuuracy 0.54220176\nEpoch 185 training acuuracy 0.54220176\nEpoch 186 training acuuracy 0.54220176\nEpoch 187 training acuuracy 0.54220176\nEpoch 188 training acuuracy 0.54220176\nEpoch 189 training acuuracy 0.54220176\nEpoch 190 training acuuracy 0.54220176\nEpoch 191 training acuuracy 0.54220176\nEpoch 192 training acuuracy 0.54220176\nEpoch 193 training acuuracy 0.54220176\nEpoch 194 training acuuracy 0.54278994\nEpoch 195 training acuuracy 0.543084\nEpoch 196 training acuuracy 0.543084\nEpoch 197 training acuuracy 0.543084\nEpoch 198 training acuuracy 0.543084\nEpoch 199 training acuuracy 0.543084\nEpoch 200 training acuuracy 0.54278994\nEpoch 201 training acuuracy 0.54249585\nEpoch 202 training acuuracy 0.54220176\nEpoch 203 training acuuracy 0.54249585\nEpoch 204 training acuuracy 0.54220176\nEpoch 205 training acuuracy 0.54220176\nEpoch 206 training acuuracy 0.54249585\nEpoch 207 training acuuracy 0.54249585\nEpoch 208 training acuuracy 0.54249585\nEpoch 209 training acuuracy 0.54249585\nEpoch 210 training acuuracy 0.54249585\nEpoch 211 training acuuracy 0.54249585\nEpoch 212 training acuuracy 0.543084\nEpoch 213 training acuuracy 0.543084\nEpoch 214 training acuuracy 0.54278994\nEpoch 215 training acuuracy 0.54249585\nEpoch 216 training acuuracy 0.54278994\nEpoch 217 training acuuracy 0.543084\nEpoch 218 training acuuracy 0.54278994\nEpoch 219 training acuuracy 0.54278994\nEpoch 220 training acuuracy 0.54278994\nEpoch 221 training acuuracy 0.54278994\nEpoch 222 training acuuracy 0.54278994\nEpoch 223 training acuuracy 0.54249585\nEpoch 224 training acuuracy 0.54249585\nEpoch 225 training acuuracy 0.54249585\nEpoch 226 training acuuracy 0.54278994\nEpoch 227 training acuuracy 0.54278994\nEpoch 228 training acuuracy 0.54278994\nEpoch 229 training acuuracy 0.54278994\nEpoch 230 training acuuracy 0.54278994\nEpoch 231 training acuuracy 0.54278994\nEpoch 232 training acuuracy 0.54249585\nEpoch 233 training acuuracy 0.54249585\nEpoch 234 training acuuracy 0.54278994\nEpoch 235 training acuuracy 0.543084\nEpoch 236 training acuuracy 0.543084\nEpoch 237 training acuuracy 0.5433781\nEpoch 238 training acuuracy 0.543084\nEpoch 239 training acuuracy 0.5433781\nEpoch 240 training acuuracy 0.5433781\nEpoch 241 training acuuracy 0.5436722\nEpoch 242 training acuuracy 0.5433781\nEpoch 243 training acuuracy 0.543084\nEpoch 244 training acuuracy 0.543084\nEpoch 245 training acuuracy 0.543084\nEpoch 246 training acuuracy 0.543084\nEpoch 247 training acuuracy 0.543084\nEpoch 248 training acuuracy 0.543084\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57947094,"AnswerCount":1,"Tags":"<python><regression><pytorch><normalization>","CreationDate":"2019-09-15T18:31:02.697","AcceptedAnswerId":null,"OwnerUserId":11725460.0,"Title":"How to \"denormalize\" a pytorch tensor?","Body":"<p>I would like to normalize the labels for my neural network but also be able to reverse the normalization process, so that I can scale my prediction outputs back to the original size.<\/p>\n\n<p>My labels are of the form:<\/p>\n\n<pre><code>[[ 2.25, 2345123.23],\n [ 1.13, 234565.11],\n ...\n [0.12, 990232.98]]\n<\/code><\/pre>\n\n<p>Now I would like to normalize each column such that the values range from 0 to 1.<\/p>\n\n<p>I found that pytorch has the torch.nn.functional.normalize function which allows me to normalize along a specific dimension using whichever p-norm I desire.<\/p>\n\n<p>Is there an existing inverse function that allows me to scale my normalized values?<\/p>\n\n<p>Currently, I am using a custom function that performs feature scaling as follows which scales both columns together, however, I wanted to make sure I am not creating a function that already exists in the torch library before rewriting it.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code> A = lbl_min\n B = lbl_max\n a = lower_bound\n b = upper_bound\n self.labels = a*np.ones(shape) + (self.labels - A*np.ones(shape))*(b-a)\/(B-A)\n<\/code><\/pre>\n\n<p>Additional information:\nI am using a regression model to output continuous real world values. I need to \"denormalize\" my predictions in order to validate them against the true values. Also, if I were to use my model in a practical application, I would need my output back in its original units.<\/p>\n\n<ul>\n<li>My neural network predicts multiple values at once that are in very different ranges.<\/li>\n<li>The two values have different distributions but can be predicted from the same input.<\/li>\n<li>My input is an image into two CNN layers that then are fed into two fully connected layers<\/li>\n<\/ul>\n","answers":[{"AnswerId":"57948219","CreationDate":"2019-09-15T21:14:08.370","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>This is a trick question right.\nIt has <strong>no sense to denormalize<\/strong> your values.<\/p>\n\n<p>Why? Because once you normalize them, the neural net should learn to create the predictions based on your normalized values.<\/p>\n\n<p>So you are safe.<\/p>\n\n<p>BTW, you will typically use mean and std of your data to normalize.<\/p>\n\n<hr>\n\n<p>In case of the regression problems it is the same. At least the architectures I created so far. For instance: Create the model M1 to learn sum <code>+<\/code> operation.\nYou provide pair of values <code>a,b<\/code> and you set the sum <code>c<\/code> as prediction.<\/p>\n\n<p>In case you train another model M2 with <code>a\/10, b\/10<\/code> and you again provide the <code>c<\/code> as prediction, this means you normalize the input.<\/p>\n\n<p>All you have to do in the inference phase is to <strong>normalize<\/strong> your inputs (divide by 10) and you get the correct output.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ZIPgZ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ZIPgZ.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":57947300,"AnswerCount":0,"Tags":"<python-3.x><tensorflow>","CreationDate":"2019-09-15T18:57:18.160","AcceptedAnswerId":null,"OwnerUserId":8270077.0,"Title":"How to get information about a dataset in Tensorflow 2.0","Body":"<p>I would like to get information describing the contents of a dataset object in Tensorflow 2.0 but I don't know how this can be done.<\/p>\n\n<p>In the previous version of Tensorflow there were the methods:<\/p>\n\n<pre><code>tf.data.get_output_classes\ntf.data.get_output_shapes\ntf.data.get_output_types\n<\/code><\/pre>\n\n<p>But in Tensorflow 2.0 these are not included.<\/p>\n\n<p>There is a property called element_spec, but when I tried to invoke it it resulted in exception:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/62Qa8.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/62Qa8.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices(np.random.randn(3,4))\nfor elem in dataset:\n\n        print(elem.element_spec)\n\n---------------------------------------------------------------------------\nAttributeError                            Traceback (most recent call last)\n&lt;ipython-input-33-d43716bf830c&gt; in &lt;module&gt;()\n      1 for elem in dataset:\n      2 \n----&gt; 3         print(elem.element_spec)\n\nAttributeError: 'tensorflow.python.framework.ops.EagerTensor' object has no attribute 'element_spec'\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57947705,"AnswerCount":0,"Tags":"<c++><tensorflow><g++>","CreationDate":"2019-09-15T19:51:55.350","AcceptedAnswerId":null,"OwnerUserId":12044023.0,"Title":"Tensorflow compile custom op undefined reference","Body":"<p>I am trying to compile a custom tensorflow op using the pip installed binarys but I get many <code>undefined reference<\/code>-errors. I am new to c++ and therefore have no idea why I get these errors, as I <code>#include<\/code> everything and even added the <code>using ...<\/code> lines where necessary. As far as I understand, the errors are related to linking and I am probably missing a <code>-L<\/code> folder but <code>tf.sysconfig.get_link_flags()<\/code> as suggested by <a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/op\" rel=\"nofollow noreferrer\">the TF guide for adding a new op<\/a> just returns an empty list.<\/p>\n\n<p>this is the command I use to compile:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>\"D:\\Program Files (x86)\\mingw-w64\\i686-8.1.0-posix-dwarf-rt_v6-rev0\\mingw32\\bin\\g++.exe\" \n-std=c++11 -shared soft_hot_encode.cc -o soft_hot_encode.so -fPIC \n-I \"D:\\Python\\Python36\\lib\\site-packages\\tensorflow\\include\" \n-I \"D:\\Python\\Python36\\Lib\\site-packages\\tensorflow\\include\\google\\protobuf_archive\\src\" \n-I \"D:\\Python\\Python36\\Lib\\site-packages\\tensorflow\\include\\tensorflow\\include\\external\\com_google_absl\" \n-I \"D:\\Python\\Python36\\Lib\\site-packages\\tensorflow\\include\\tensorflow\\include\\external\\eigen_archive\" \n-D_GLIBCXX_USE_CXX11_ABI=0 -O2 -w \n<\/code><\/pre>\n\n<p>errors look like this:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>D:\\AppData\\Local\\Temp\\ccZAntyd.o:soft_hot_encode.cc:(.text+0x2f): undefined reference to `tensorflow::OpKernel::OpKernel(tensorflow::OpKernelConstruction*)'\nD:\\AppData\\Local\\Temp\\ccZAntyd.o:soft_hot_encode.cc:(.text+0x85): undefined reference to `tensorflow::shape_inference::InferenceContext::MakeDimForScalarInput(int, tensorflow::shape_inference::DimensionHandle*)'\nD:\\AppData\\Local\\Temp\\ccZAntyd.o:soft_hot_encode.cc:(.text+0xf5): undefined reference to `tensorflow::shape_inference::InferenceContext::Subshape(tensorflow::shape_inference::ShapeHandle, long long, long long, tensorflow::shape_inference::ShapeHandle*)'\nD:\\AppData\\Local\\Temp\\ccZAntyd.o:soft_hot_encode.cc:(.text+0x138): undefined reference to `tensorflow::shape_inference::InferenceContext::Subshape(tensorflow::shape_inference::ShapeHandle, long long, tensorflow::shape_inference::ShapeHandle*)'\nD:\\AppData\\Local\\Temp\\ccZAntyd.o:soft_hot_encode.cc:(.text+0x187): undefined reference to `tensorflow::shape_inference::InferenceContext::Vector(tensorflow::shape_inference::DimensionOrConstant)'\n    ...\n<\/code><\/pre>\n\n<p>Edit:<\/p>\n\n<p><strong>I solved this problem<\/strong> by compiling using bazel and the tensorflow source. To get it to compile properly, I needed to use the source of a proper release (in my case v2.0.0-beta1, commit: 8e423e3) by invoking <code>git checkout 8e423e3d56390671f0d954c90f4fd163ab02a9c1<\/code> in the previously cloned repository folder. After that I followed the <a href=\"https:\/\/www.tensorflow.org\/install\/source_windows\" rel=\"nofollow noreferrer\">tensorflow guide to compiling from source<\/a> to setup bazel.<\/p>\n","answers":[]}
+{"QuestionId":57947725,"AnswerCount":1,"Tags":"<python><pytorch><vector-multiplication>","CreationDate":"2019-09-15T19:55:02.773","AcceptedAnswerId":null,"OwnerUserId":11725460.0,"Title":"Multiply rows of matrix by vector elementwise in pytorch?","Body":"<p>I would like to do the below but using <strong>PyTorch<\/strong>.<\/p>\n\n<p>The below example and description is from this <a href=\"https:\/\/stackoverflow.com\/questions\/3643555\/multiply-rows-of-matrix-by-vector\">post<\/a>.<\/p>\n\n<blockquote>\n  <p>I have a numeric matrix with 25 columns and 23 rows, and a vector of\n  length 25. How can I multiply each row of the matrix by the vector\n  without using a for loop?<\/p>\n  \n  <p>The result should be a 25x23 matrix (the same size as the input), but\n  each row has been multiplied by the vector.<\/p>\n<\/blockquote>\n\n<p>Example Code in R (source: <a href=\"https:\/\/stackoverflow.com\/questions\/3643555\/multiply-rows-of-matrix-by-vector\">reproducible example from @hatmatrix's answer<\/a>):<\/p>\n\n<pre><code>matrix &lt;- matrix(rep(1:3,each=5),nrow=3,ncol=5,byrow=TRUE)\n\n     [,1] [,2] [,3] [,4] [,5]\n[1,]    1    1    1    1    1\n[2,]    2    2    2    2    2\n[3,]    3    3    3    3    3\n\nvector &lt;- 1:5\n<\/code><\/pre>\n\n<p>Desired output:<\/p>\n\n<pre><code>     [,1] [,2] [,3] [,4] [,5]\n[1,]    1    2    3    4    5\n[2,]    2    4    6    8   10\n[3,]    3    6    9   12   15\n<\/code><\/pre>\n\n<p>What is the best way of doing this using Pytorch?<\/p>\n","answers":[{"AnswerId":"57947795","CreationDate":"2019-09-15T20:03:40.580","ParentId":null,"OwnerUserId":"11725460","Title":null,"Body":"<p>The answer was so trivial that I overlooked it.<\/p>\n\n<p>For simplicity I used a smaller vector and matrix in this answer.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>X = torch.tensor([[3, 5],[5, 5],[1, 0]])                                                                                                                                                                          \ny = torch.tensor([7,4])                                                                                                                                                                                   \nX*y\n# or alternatively\ny*X\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<pre><code>tensor([[21, 20],\n        [35, 20],\n        [ 7,  0]])\n\ntensor([[21, 20],\n        [35, 20],\n        [ 7,  0]])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57947752,"AnswerCount":0,"Tags":"<python><tensorflow><jupyter-notebook>","CreationDate":"2019-09-15T19:58:27.037","AcceptedAnswerId":null,"OwnerUserId":9819895.0,"Title":"jupyter book points wrong module location. Couldnt find 'cudart64_100.dll'","Body":"<p>TensorFlow: 12.0.0\nCuda: 9.0<\/p>\n\n<p>I am trying to compile object_detection_tutorial.ipynb from object_detection in one of the TensorFlow's sections in jupyter notebook. Jupyter has been installed via pip install. When I'm compiling script I get error code telling me to install Cuda 10.0 due to lack of 'cudart64_100.dll' in bin folder. Jupyter points the wrong path to TensorFlow module, which is located in python's pip folder, this can cause the whole issue. How can I specify a path to python interpreter, not to one chosen by jupyter by default?<\/p>\n\n<pre><code> c:\\users\\\u0142ukasz\\pycharmprojects\\untitled2\\venv\\lib\\site-packages\\tensorflow\\python\\platform\\self_check.py in preload_check()\n 80               \"environment variable. Download and install CUDA %s from \"\n 81               \"this URL: https:\/\/developer.nvidia.com\/cuda-90-download-archive\"\n ---&gt; 82               % (build_info.cudart_dll_name, build_info.cuda_version_number))\n 83 \n 84       if hasattr(build_info, \"cudnn_dll_name\") and hasattr(\n\nImportError: Could not find 'cudart64_100.dll'. TensorFlow requires that this DLL be installed    in a directory that is named in your %PATH% environment variable. Download and install CUDA 10.0 from this URL: https:\/\/developer.nvidia.com\/cuda-90-download-archive\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57947967,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><linear-regression><tensorboard><tensorflow-estimator>","CreationDate":"2019-09-15T20:28:48.803","AcceptedAnswerId":null,"OwnerUserId":8756425.0,"Title":"Calculating Log Loss and Displaying it on Tensorboard","Body":"<p>I am new to TensorFlow\/TensorBoard and trying to run a sample Linear Classifier with Logistic Regression. <\/p>\n\n<p>Right now, the code runs OK and TensorBoard displays a graph called \"Loss\" and I'm guessing that is not the Log Loss I need. <\/p>\n\n<p>How can I add Log Loss to the code below without complicating it (changing it too much) and how can I display it on the TensorBoard the way Loss, Accuracy, etc are displayed?<\/p>\n\n<p>Thank you very much!<\/p>\n\n<pre><code>import tensorflow as tf\nimport pandas as pd\nimport os\nos.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'\n\nCOLUMNS = ['numPackets', 'NeighborSize', 'Entropy', 'SampledFlowSize', 'AverageMacDelay',\n          'ReceivedBroadcasts', 'SentPackets', 'label']\n\nFEATURES = ['numPackets', 'NeighborSize', 'Entropy', 'SampledFlowSize', 'AverageMacDelay',\n          'ReceivedBroadcasts', 'SentPackets']\n\ncontinuous_features = [tf.feature_column.numeric_column(k) for k in FEATURES]\n\nLABEL= 'label'\n\ndef get_input_fn(data_set, num_epochs=None, n_batch = 256, shuffle=True):\n    return tf.estimator.inputs.pandas_input_fn(\n       x=pd.DataFrame({k: data_set[k].values for k in FEATURES}),\n       y = pd.Series(data_set[LABEL].values),\n       batch_size=n_batch,\n       num_epochs=num_epochs,\n       shuffle=shuffle)\n\nPATH = \"\/tensor\/final_dataset\/dataset.small\"\nPATH_test = \"\/tensor\/final_dataset\/dataset.small.test\"\n\ndf_train = pd.read_csv(PATH, skipinitialspace=True, names = COLUMNS, index_col=False)\ndf_test = pd.read_csv(PATH_test, skipinitialspace=True, names = COLUMNS, index_col=False)\n\nmodel = tf.estimator.LinearClassifier(\n    n_classes = 2,\n    model_dir=\"ongoing\/train\",\n    feature_columns=continuous_features, optimizer=tf.train.FtrlOptimizer(\n        learning_rate=0.4,\n        l1_regularization_strength=0.9,\n        l2_regularization_strength=5))\n\n\nmodel.train(input_fn=get_input_fn(df_train,\n                                      num_epochs=None,\n                                      n_batch = 32,\n                                      shuffle=False),\n                                      steps=1000)\n\nresult = model.evaluate(input_fn=get_input_fn(df_test,\n                                      num_epochs=None,\n                                      n_batch = 32,\n                                      shuffle=False),\n                                      steps=1000)\n\n#predictions = model.predict(df_test)\n\nfor key,value in sorted(result.items()):\n  print('%s: %s' % (key, value))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57948024,"AnswerCount":0,"Tags":"<python><nlp><pytorch>","CreationDate":"2019-09-15T20:38:41.977","AcceptedAnswerId":null,"OwnerUserId":11599315.0,"Title":"RuntimeError: Expected object of backend CPU but got backend CUDA for argument #3 'index'","Body":"<p>I'm experimenting with argumentation mining with BERT and try to identify argument components (BIO-classification task). I closely followed the code from this article, <a href=\"https:\/\/www.depends-on-the-definition.com\/named-entity-recognition-with-bert\" rel=\"nofollow noreferrer\">Named Entity Recognition With Bert<\/a>, but adapted it to fit my data. I run the code on colab (hardware accelerator is set to GPU). <\/p>\n\n<p>Does anybody know a solution to my problem? Thanks in advance!<\/p>\n\n<pre><code>epochs = 5\nmax_grad_norm = 1.0\n\nfor _ in trange(epochs, desc=\"Epoch\"):\n    # TRAIN loop\n    model.train()\n    tr_loss = 0\n    nb_tr_examples, nb_tr_steps = 0, 0\n    for step, batch in enumerate(train_dataloader):\n        # add batch to gpu\n        batch = tuple(t.to(device) for t in batch)\n        b_input_ids, b_input_mask, b_labels = batch\n        # forward pass\n        loss = model(b_input_ids, token_type_ids=None, attention_mask=b_input_mask, labels=b_labels)\n        # backward pass\n        loss.backward()\n        # track train loss\n        tr_loss += loss.item()\n        nb_tr_examples += b_input_ids.size(0)\n        nb_tr_steps += 1\n        # gradient clipping\n        torch.nn.utils.clip_grad_norm_(parameters=model.parameters(), max_norm=max_grad_norm)\n        # update parameters\n        optimizer.step()\n        model.zero_grad()\n    # print train loss per epoch\n    print(\"Train loss: {}\".format(tr_loss\/nb_tr_steps))\n    # VALIDATION on validation set\n    model.eval()\n    eval_loss, eval_accuracy = 0, 0\n    nb_eval_steps, nb_eval_examples = 0, 0\n    predictions , true_labels = [], []\n    for batch in valid_dataloader:\n        batch = tuple(t.to(device) for t in batch)\n        b_input_ids, b_input_mask, b_labels = batch\n\n        with torch.no_grad():\n            tmp_eval_loss = model(b_input_ids, token_type_ids=None,\n                                  attention_mask=b_input_mask, labels=b_labels)\n            logits = model(b_input_ids, token_type_ids=None,\n                           attention_mask=b_input_mask)\n        logits = logits.detach().cpu().numpy()\n        label_ids = b_labels.to('cpu').numpy()\n        predictions.extend([list(p) for p in np.argmax(logits, axis=2)])\n        true_labels.append(label_ids)\n\n        tmp_eval_accuracy = flat_accuracy(logits, label_ids)\n\n        eval_loss += tmp_eval_loss.mean().item()\n        eval_accuracy += tmp_eval_accuracy\n\n        nb_eval_examples += b_input_ids.size(0)\n        nb_eval_steps += 1\n    eval_loss = eval_loss\/nb_eval_steps\n    print(\"Validation loss: {}\".format(eval_loss))\n    print(\"Validation Accuracy: {}\".format(eval_accuracy\/nb_eval_steps))\n    pred_tags = [tags_vals[p_i] for p in predictions for p_i in p]\n    valid_tags = [tags_vals[l_ii] for l in true_labels for l_i in l for l_ii in l_i]\n    print(\"F1-Score: {}\".format(f1_score(pred_tags, valid_tags)))\n<\/code><\/pre>\n\n<p>Although I closely followed the original code, it does not seem to work for me and I keep on getting this mistake...<\/p>\n\n<pre><code>RuntimeError                              Traceback (most recent call last)\n&lt;ipython-input-45-aa659bbf8fac&gt; in &lt;module&gt;()\n     12 \n     13         # forward pass\n---&gt; 14         loss = model(b_input_ids, token_type_ids=None, attention_mask=b_input_mask, labels=b_labels)\n     15         # backward pass\n     16         loss.backward()\n\n8 frames\n\/usr\/local\/lib\/python3.6\/dist-packages\/torch\/nn\/functional.py in embedding(input, weight, padding_idx, max_norm, norm_type, scale_grad_by_freq, sparse)\n   1465         # remove once script supports set_grad_enabled\n   1466         _no_grad_embedding_renorm_(weight, input, max_norm, norm_type)\n-&gt; 1467     return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)\n   1468 \n   1469 \n\nRuntimeError: Expected object of backend CPU but got backend CUDA for argument #3 'index'\n<\/code><\/pre>\n\n<p><strong>SOLUTION<\/strong>:\nsetting model = model.to(device) plus reducing the batch size from 32 to 4 allowed the program to run without any runtimeerrors. <\/p>\n","answers":[]}
+{"QuestionId":57948111,"AnswerCount":0,"Tags":"<tensorflow><time-series><convolution><torch><autoencoder>","CreationDate":"2019-09-15T20:54:19.447","AcceptedAnswerId":null,"OwnerUserId":3641187.0,"Title":"1D Causal De-Convolution in autoencoders - does it make sense?","Body":"<p>I am currently trying to adapted the Temporal Convolutional Net (TCN) framework to an autoencoder model for creating latent representations of multivariate time series. A nice explanation for TCN can be found here:<\/p>\n\n<p><a href=\"https:\/\/arxiv.org\/pdf\/1803.01271.pdf\" rel=\"nofollow noreferrer\">Paper on TCN sequence modeling<\/a><\/p>\n\n<p>Almost all TCN variants employ a form of 1D convolution on their inputs which can be seen as time series of sorts (multivariate should be supported). The 1D convolution is understood to be \"causal\", meaning that zero-padding is generally only applied to the \"left\" side of the input to avoid leakage from future events in the downstream network flow during the encoding procedure.<\/p>\n\n<p>I like to build my Encoder \/ Decoder architectures in a  mirrored fashion - pooling cells in the encoder become upsampling cells in the decoder. Following this rule, 1D convolutions should be come 1D de-convolutions (or transposed convolutions as they sometimes are called).<\/p>\n\n<p>My question: does it make sense to think about \"causality\" during 1D de-convolution steps in the decoder? Not that I am not generating loss via a predictive reconstruction (reconstruct the time step ahead is often being used for predictive autoencoders) - i train on reconstructing the original time series.<\/p>\n\n<p>Intuitively, it seems that 1D deconvolution operations with \"valid\" or \"same\" padding does not create \"future leakage\", since by construction the decoder input contains only causal temporal correlations. Am I correct in this assumption?<\/p>\n","answers":[]}
+{"QuestionId":57948195,"AnswerCount":1,"Tags":"<python><tensorflow><anaconda><sublimetext3>","CreationDate":"2019-09-15T21:10:17.727","AcceptedAnswerId":57982533.0,"OwnerUserId":12071453.0,"Title":"Sublime text errors. (import?)","Body":"<p>Tried to run this code in Sublime:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.keras.datasets import cifar10\nfrom tensorflow.keras.preprocessing.image import ImageDataGenerator\nfrom tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import Dense, Dropout, Activation, Flatten\nfrom tensorflow.keras.layers import Conv2D, MaxPooling2D\nimport pickle\n\npickle_in = open(\"X.pickle\",\"rb\")\nX = pickle.load(pickle_in)\n\npickle_in = open(\"y.pickle\",\"rb\")\ny = pickle.load(pickle_in)\n\nX = X\/255.0\n\nmodel = Sequential()\n\nmodel.add(Conv2D(256, (3, 3), input_shape=X.shape[1:]))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Conv2D(256, (3, 3)))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Flatten())  # this converts our 3D feature maps to 1D feature vectors\n\nmodel.add(Dense(64))\nmodel.add(Activation('relu'))\n\nmodel.add(Dense(1))\nmodel.add(Activation('sigmoid'))\n\nmodel.compile(loss='binary_crossentropy',\n              optimizer='adam',\n              metrics=['accuracy'])\n\nmodel.fit(X, y, batch_size=32, epochs=10, validation_split=0.3)\n<\/code><\/pre>\n\n<p>but I keep getting those errors:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"D:\\catsdogsai\\catsdogsai\\catsdogsai.py\", line 1, in &lt;module&gt;\n    import tensorflow as tf\n  File \"C:\\Users\\Edward\\Anaconda3\\envs\\aitest\\lib\\site-packages\\tensorflow\\__init__.py\", line 28, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"C:\\Users\\Edward\\Anaconda3\\envs\\aitest\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 47, in &lt;module&gt;\n    import numpy as np\n  File \"C:\\Users\\Edward\\Anaconda3\\envs\\aitest\\lib\\site-packages\\numpy\\__init__.py\", line 140, in &lt;module&gt;\n    from . import _distributor_init\n  File \"C:\\Users\\Edward\\Anaconda3\\envs\\aitest\\lib\\site-packages\\numpy\\_distributor_init.py\", line 34, in &lt;module&gt;\n    from . import _mklinit\nImportError: DLL load failed: The specified module could not be found.\n<\/code><\/pre>\n\n<p>I have installed all the modules, activated the env on Sublime too but still nothing. <\/p>\n\n<p>Triend reinstalling everything but it's still nothing good, same errors.<\/p>\n","answers":[{"AnswerId":"57982533","CreationDate":"2019-09-17T21:53:38.440","ParentId":null,"OwnerUserId":"12071453","Title":null,"Body":"<p>Solved!<\/p>\n\n<p>I had to downgrade the CUDA from 10.1 to 10.0 since apparently tensorflow only supports 10.0.<\/p>\n"}]}
+{"QuestionId":57948362,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><jupyter-notebook><jupyter>","CreationDate":"2019-09-15T21:41:14.363","AcceptedAnswerId":null,"OwnerUserId":1362485.0,"Title":"TensorFlow deactivates Python when imported in Jupyter notebook","Body":"<p>I have installed TensorFlow with Miniconda 3 in the base environment and it works fine in the PyCharm editor. But when I try to use TensorFlow in a Jupyter Notebook, it crashes Python. This is what I see, even with a simple <code>import tensorflow as tf<\/code>, what could be the problem and how to fix this?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/wAqN5.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/wAqN5.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"57948631","CreationDate":"2019-09-15T22:28:03.337","ParentId":null,"OwnerUserId":"7876832","Title":null,"Body":"<p>Could be a memory issue with your <code>Jupyter<\/code> editor.  I'd suggest, if you can, try importing only the appropriate modules required for your tensorflow application and see if that solves the issue. <\/p>\n"}]}
+{"QuestionId":57948726,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-15T22:46:21.690","AcceptedAnswerId":57949580.0,"OwnerUserId":4496896.0,"Title":"TensorFlow: \"FailedPreconditionError: Table not initialized\" in Keras model layer","Body":"<p>I'm trying to add a LookupTable with a trained Keras classification model as final layer. My model basically predicts a label and score for which I would like to convert the label into another value.<\/p>\n\n<ul>\n<li>A high level code:<\/li>\n<\/ul>\n\n<pre class=\"lang-py prettyprint-override\"><code>def my_func(x):\n    import tensorflow as tf\n\n    keys = tf.constant([0, 1, 2, 3, 4], dtype=tf.int32)\n    values = tf.constant([676754, 425362, 918376, 152678], dtype=tf.float32)\n\n    init = tf.lookup.KeyValueTensorInitializer(keys, values)\n    table = tf.lookup.StaticHashTable(init, -1)\n    init.initialize(table)\n\n    label_tensor = tf.cast(x[:, :, 1], tf.int32)\n    score_tensor = x[:, :, 2]\n\n    result = table.lookup(label_tensor)\n    return tf.concat((tf.reshape(result, (-1, 2, 1)), tf.reshape(score_tensor, (-1, 2, 1))), axis=1)\n\nnew_layer = layers.Lambda(my_func)(model.output)\nnew_model = Model(inputs=model.input, outputs=new_layer)\n<\/code><\/pre>\n\n<p><code>FailedPreconditionError: Table not initialized<\/code> Error I'm getting:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nFailedPreconditionError                   Traceback (most recent call last)\n&lt;ipython-input-111-7598eb4fb2fb&gt; in &lt;module&gt;\n      1 \n      2 \n----&gt; 3 predictions = new_model.predict(pad_sequences(encoded_data, maxlen=15, padding='post'))\n\n~\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/training.py 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~\/anaconda3\/lib\/python3.6\/site-packages\/keras\/engine\/training_arrays.py 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~\/anaconda3\/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~\/anaconda3\/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~\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py in __call__(self, *args, **kwargs)\n   1456         ret = tf_session.TF_SessionRunCallable(self._session._session,\n   1457                                                self._handle, args,\n-&gt; 1458                                                run_metadata_ptr)\n   1459         if run_metadata:\n   1460           proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\nFailedPreconditionError: Table not initialized.\n     [[{{node new_model.lambda.LookupTableFindV2}}]]\n<\/code><\/pre>\n\n<p>Things that I have tried:<\/p>\n\n<ul>\n<li>Instead of using <code>init.initialize(table)<\/code> I've also tried using <code>tf.tables_initializer()<\/code><\/li>\n<\/ul>\n","answers":[{"AnswerId":"57949580","CreationDate":"2019-09-16T02:02:57.200","ParentId":null,"OwnerUserId":"4496896","Title":null,"Body":"<p><strong>SOLVED:<\/strong><\/p>\n\n<ul>\n<li>You have to initialize all tables outside the lambda function, before the <code>model.predict()<\/code><\/li>\n<li>You can also remove <strike><strong>init.initialize(table)<\/strong><\/strike> from the lambda function.<\/li>\n<\/ul>\n\n<pre class=\"lang-py prettyprint-override\"><code>K.get_session().run(tf.tables_initializer(name='init_all_tables'))\nnew_model.predict(input)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57948851,"AnswerCount":0,"Tags":"<python><tensorflow><garbage-collection>","CreationDate":"2019-09-15T23:14:54.353","AcceptedAnswerId":null,"OwnerUserId":6917400.0,"Title":"How to reset or delete a tensorflow graph?","Body":"<p>Some questions exist discussing resetting a tf Graph object. The solution is, that in Tensorflow one can not actually empty or reset a graph. There just exist graph objects, and even tf.reset_default_graph() is of no use, because it only sets the default graph to a new graph object. <\/p>\n\n<p>Sometimes it seems useful to get rid of the content of graphs though, as their contents can be quite RAM consuming. How can this then be achieved, i.e. how can one get rid of graphs or variables? Are graphs (together with their contents) simply garbage collected, when there is no more reference to them, or does Tensorflow actually keep a reference to them somewhere internally? \nWhat is the clean way, to free RAM from large tensorflow graphs or variables?<\/p>\n","answers":[]}
+{"QuestionId":57949520,"AnswerCount":0,"Tags":"<python><tensorflow><google-cloud-platform><object-detection-api>","CreationDate":"2019-09-16T01:50:26.393","AcceptedAnswerId":null,"OwnerUserId":10834282.0,"Title":"Tensorflow object detection wont start to train on gcp . Error after installing pycocotools","Body":"<p>I have been trying to run the google object detection module on gcp and after packing slim , objectdetection and pycocotools just like the examples, I am always getting the same error <\/p>\n\n<pre><code>please install the python-tk package' ImportError: No module named _tkinter, please install the python-tk package\n<\/code><\/pre>\n\n<p>I have tried multiple models and the suggestions on this site to change backend of matplotlib to agg as well but it didn't help my case.<\/p>\n\n<p>please help <\/p>\n","answers":[]}
+{"QuestionId":57949625,"AnswerCount":2,"Tags":"<pytorch>","CreationDate":"2019-09-16T02:12:27.730","AcceptedAnswerId":57962290.0,"OwnerUserId":3983454.0,"Title":"With pytorch DataLoader how to take in two ndarray (data & label)?","Body":"<p>I have a training data features in ndarray of shape (100, 400, 3) as it's 100 images of 20x20 with RGB channel and label in shape (100, ). Do I need to combine them into one dataset or how can I pass it to Pytorch dataLoader in order to iterate over image and label later?  <\/p>\n\n<p>What I've tried so far<\/p>\n\n<pre><code>#turn ndarray of features and labels into tensors\ntransform = transforms.Compose([transforms.ToPILImage(),\n                                transforms.ToTensor()])\n<\/code><\/pre>\n","answers":[{"AnswerId":"57950625","CreationDate":"2019-09-16T05:00:21.137","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.from_numpy\" rel=\"nofollow noreferrer\">convert<\/a> your data\/label ndarrays to <code>torch.tensor<\/code> and use <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/data.html#torch.utils.data.TensorDataset\" rel=\"nofollow noreferrer\"><code>torch.utils.data.TensorDataset<\/code><\/a> to create a dataset that iterates over your examples.<br>\nOnce you have a dataset, you can wrap a <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/data.html#torch.utils.data.DataLoader\" rel=\"nofollow noreferrer\"><code>DataLoader<\/code><\/a> around it to be used for training.<\/p>\n"},{"AnswerId":"57962290","CreationDate":"2019-09-16T18:10:51.473","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>As @Shai mentioned, <code>DataLoader<\/code> requires the input to be the <code>Dataset<\/code> class or its subclass. One of the simplest subclasses is <code>TensorDataset<\/code> and you can convert it from ndarray. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nimport numpy as np\nimport torch.utils as utils\n\ntrain_x = torch.Tensor(np.random.randn(100,400,3))\ntrain_y = torch.Tensor(np.random.randint(0,2,100))\n\ndataset = utils.data.TensorDataset(train_x, train_y)\ndataloader = utils.data.DataLoader(dataset)\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":57949866,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-16T02:54:06.413","AcceptedAnswerId":57959306.0,"OwnerUserId":11607378.0,"Title":"What types of model files does Tensorflow support and how to load .ckpt + .pb","Body":"<p>I found TensorFlow supports various of model files, <code>.ckpt, .meta, .pb<\/code><\/p>\n\n<p>1) now I have a resnet50 model, with <code>.ckpt<\/code> 102MB and <code>.pb<\/code> 459KB, how to load them into memory by using TensorFlow?<\/p>\n\n<p>2) And how many types of model files does Tensorflow support? Could anyone provide a list or material like (such as 1. ckpt+pb 2. frozen.pb 3. meta + ??) <\/p>\n","answers":[{"AnswerId":"57959306","CreationDate":"2019-09-16T14:46:43.960","ParentId":null,"OwnerUserId":"10111013","Title":null,"Body":"<p>These files are not different file types to store complete models. Each contains different information and collectively they store a single model. <\/p>\n\n<p>The <code>.meta<\/code> file stores the bulk of the structure of the model, while the <code>.ckpt<\/code> file is a checkpoint file storing information about the values of the Variables within the network (weights and biases for a neural network) at a stage of training where the network was saved. The <code>.pb<\/code> files are protocol buffer files. <\/p>\n\n<p>To see a working example of how to save and restore a model using these files see here:<\/p>\n\n<p><a href=\"https:\/\/bitbucket.org\/FinleyGibson\/save-and-load-tf-model\/\" rel=\"nofollow noreferrer\">https:\/\/bitbucket.org\/FinleyGibson\/save-and-load-tf-model\/<\/a><\/p>\n\n<p>and open the .ipynb files with the jupyter notebook file viewer<\/p>\n"}]}
+{"QuestionId":57950679,"AnswerCount":1,"Tags":"<python><keras><deep-learning><text-classification>","CreationDate":"2019-09-16T05:08:14.400","AcceptedAnswerId":57950818.0,"OwnerUserId":8754011.0,"Title":"How to classify a resume segments?","Body":"<p>I need to apply deep learning in resume analysis, by taking the text and split it into paragraphs, then pass each paragraph into deep learning model that predict if this paragraph is skills or education etc...<\/p>\n\n<p>By the way I have a dataset that contains paragraphs and its labels!<\/p>\n\n<p>Can I perform that with text classification in keras?<\/p>\n","answers":[{"AnswerId":"57950818","CreationDate":"2019-09-16T05:26:49.877","ParentId":null,"OwnerUserId":"11952444","Title":null,"Body":"<p>Yes, you can apply a simple text using CNN and recurrent neural network to automate the task of parsing.<\/p>\n\n<ol>\n<li><p><strong>CNN<\/strong><\/p>\n\n<ul>\n<li>1-D CNN with Word Embedding<\/li>\n<li>Multi-Channel CNN with categorical cross-entropy loss function<\/li>\n<\/ul><\/li>\n<li><p><strong>CNN + LSTM<\/strong><\/p>\n\n<ul>\n<li>1-D CNN + LSTM with Word Embedding<\/li>\n<\/ul><\/li>\n<li><p><strong>LSTM<\/strong><\/p>\n\n<ul>\n<li>LSTM with category cross-entropy loss function<\/li>\n<li>Bi-directional LSTM\/GRU with categorical cross-entropy loss function<\/li>\n<\/ul><\/li>\n<\/ol>\n\n<p>You can check the link down below:<\/p>\n\n<p><a href=\"https:\/\/github.com\/chen0040\/keras-english-resume-parser-and-analyzer\" rel=\"nofollow noreferrer\">Github<\/a><\/p>\n"}]}
+{"QuestionId":57950746,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><machine-learning><deep-learning>","CreationDate":"2019-09-16T05:16:28.900","AcceptedAnswerId":57952836.0,"OwnerUserId":6997665.0,"Title":"Restoring saved model in Tensorflow v1.14","Body":"<p>I am using tensorflow v1.14. I have a saved model and I'm trying to restore the model using the following code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>loader = tf.train.import_meta_graph(\"models\/fcnn0\/model.ckpt.meta\")\ngraph = tf.get_default_graph()\nsess = tf.Session()\nloader.restore(sess, \"models\/fcnn0\/model.ckpt\")\n\n<\/code><\/pre>\n\n<p>I used to use the same piece of code in Tensorflow v1.13 and it used to work without errors. But now I'm getting the error<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py\", line 1356, in _do_call\n    return fn(*args)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py\", line 1341, 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 1429, in _call_tf_sessionrun\n    run_metadata)\ntensorflow.python.framework.errors_impl.DataLossError: file is too short to be an sstable\n     [[{{node save\/RestoreV2}}]]\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"\/home\/sandesh\/PycharmProjects\/fading\/finding_code\/src\/load_32_64.py\", line 8, in &lt;module&gt;\n    loader.restore(sess, \"models\/fcnn_32_64_aenc_1331_747_3870000\/model.ckpt\")\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/training\/saver.py\", line 1286, in restore\n    {self.saver_def.filename_tensor_name: save_path})\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py\", line 950, in run\n    run_metadata_ptr)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py\", line 1173, in _run\n    feed_dict_tensor, options, run_metadata)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py\", line 1350, in _do_run\n    run_metadata)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/client\/session.py\", line 1370, in _do_call\n    raise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.DataLossError: file is too short to be an sstable\n     [[node save\/RestoreV2 (defined at home\/sandesh\/PycharmProjects\/fading\/finding_code\/src\/load_32_64.py:5) ]]\n\nOriginal stack trace for 'save\/RestoreV2':\n  File \"home\/sandesh\/PycharmProjects\/fading\/finding_code\/src\/load_32_64.py\", line 5, in &lt;module&gt;\n    loader = tf.train.import_meta_graph(\"models\/fcnn_32_64_aenc0\/model.ckpt.meta\")\n  File \"usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/training\/saver.py\", line 1449, in import_meta_graph\n    **kwargs)[0]\n  File \"usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/training\/saver.py\", line 1473, in _import_meta_graph_with_return_elements\n    **kwargs))\n  File \"usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/meta_graph.py\", line 857, in import_scoped_meta_graph_with_return_elements\n    return_elements=return_elements)\n  File \"usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/util\/deprecation.py\", line 507, in new_func\n    return func(*args, **kwargs)\n  File \"usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/importer.py\", line 443, in import_graph_def\n    _ProcessNewOps(graph)\n  File \"usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/importer.py\", line 236, in _ProcessNewOps\n    for new_op in graph._add_new_tf_operations(compute_devices=False):  # pylint: disable=protected-access\n  File \"usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 3751, in _add_new_tf_operations\n    for c_op in c_api_util.new_tf_operations(self)\n  File \"usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 3751, in &lt;listcomp&gt;\n    for c_op in c_api_util.new_tf_operations(self)\n  File \"usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 3641, in _create_op_from_tf_operation\n    ret = Operation(c_op, self)\n  File \"usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 2005, in __init__\n    self._traceback = tf_stack.extract_stack()\n\n<\/code><\/pre>\n\n<p>Can someone point me as to what I'm doing wrong? Thanks in advance.<\/p>\n","answers":[{"AnswerId":"57952836","CreationDate":"2019-09-16T08:15:10.090","ParentId":null,"OwnerUserId":"6997665","Title":null,"Body":"<p>I was looking into the folder where the model files were saved and found that the <code>model.ckpt.meta<\/code> file had not been written to disk properly. I reran the training and saved the model and then it worked perfectly.<\/p>\n"}]}
+{"QuestionId":57950830,"AnswerCount":1,"Tags":"<tensorflow><graph>","CreationDate":"2019-09-16T05:28:17.280","AcceptedAnswerId":57958200.0,"OwnerUserId":9237371.0,"Title":"\"the operation's graph is different from the session's graph\" Error","Body":"<p>I am new to tensorflow and got this error message when I tried to create a graph and then execute some operations:<\/p>\n\n<pre><code>ValueError                                Traceback (most recent call last)\n&lt;ipython-input-136-9e5ed7cede4c&gt; in &lt;module&gt;()\n----&gt; 1 get_ipython().run_cell_magic('time', '', '\\nn_epochs = 20\\nbatch_size = 5\\n\\ninit = tf.global_variables_initializer()\\n\\nwith tf.Session(graph=graph) as sess:\\n    print(0)\\n    init.run()\\n    print(1)\\n    # init = tf.global_variables_initializer()\\n    # saver = tf.train.Saver()\\n    print(2)\\n    for epoch in range(n_epochs):\\n        for X_batch, y_batch in shuffle_batch(X_train, y_train, batch_size):\\n            sess.run(training_op, feed_dict={X: X_batch, y: y_batch})\\n        mse_batch = loss.eval(feed_dict={X: X_batch, y: y_batch})\\n        mse_valid = loss.eval(feed_dict={X: X_valid, y: y_valid})\\n        print(epoch, \"Batch mse:\", mse_batch, \"Validation mse:\", mse_valid)\\n\\n    # save_path = saver.save(sess, \".\/my_model_final.ckpt\")')\n\n4 frames\n&lt;\/usr\/local\/lib\/python3.6\/dist-packages\/decorator.py:decorator-gen-60&gt; in time(self, line, cell, local_ns)\n\n&lt;timed exec&gt; in &lt;module&gt;()\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/ops.py in _run_using_default_session(operation, feed_dict, graph, session)\n   5603                        \"`run(session=sess)`\")\n   5604     if session.graph is not graph:\n-&gt; 5605       raise ValueError(\"Cannot use the default session to execute operation: \"\n   5606                        \"the operation's graph is different from the \"\n   5607                        \"session's graph. Pass an explicit session to \"\n\nValueError: Cannot use the default session to execute operation: the operation's graph is different from the session's graph. Pass an explicit session to run(session=sess)\n<\/code><\/pre>\n\n<p>I'm trying to have a method to build multiple graphs for hyperparameter tuning (no of neurons per layer, no of hidden layers, etc.). The code is mainly from <a href=\"https:\/\/github.com\/ageron\/handson-ml\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ageron\/handson-ml<\/a> as an example. I don't think I'm constructing and using the graph correctly.  <\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport pandas as pd\n\n\ndef reset_graph(seed=42):\n    tf.reset_default_graph()\n    tf.set_random_seed(seed)\n    np.random.seed(seed)\n\n\ndef create_graph(n_inputs, n_outputs, n_hidden_layers=2, n_neurons_per_layer=100, activation_function=tf.nn.relu, learning_rate=0.01, \n                 optimize_method='nesterov'):\n\n  g = tf.Graph()\n\n\n  with g.as_default():\n    X = tf.placeholder(tf.float32, shape=(None, n_inputs), name=\"X\")\n    y = tf.placeholder(tf.int32, shape=(None), name=\"y\")\n\n    # create layers\n    with tf.name_scope(\"dnn\"):\n      hidden_layers = {}\n      for i in range(n_hidden_layers):\n        if i == 0:\n          hidden_layers['hidden_1'] = tf.layers.dense(X, n_neurons_per_layer, activation=activation_function, name=\"hidden_1\")\n        else:\n          name_last_layer = 'hidden_' + str(i)\n          name_this_layer = 'hidden_' + str(i+1)\n          hidden_layers[name_this_layer] = tf.layers.dense(hidden_layers[name_last_layer], n_neurons_per_layer, \n                                                           activation=activation_function, name=\"hidden_{}\".format(i+1))\n      name_last_hidden = 'hidden_' + str(n_hidden_layers)\n      logits = tf.layers.dense(hidden_layers[name_last_hidden], n_outputs, name=\"logits\")\n\n    with tf.name_scope(\"loss\"):\n        xentropy = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y, logits=logits)\n        loss = tf.reduce_mean(xentropy, name=\"loss\")\n        y_proba = tf.nn.softmax(logits)\n\n    with tf.name_scope(\"eval\"):\n        correct = tf.nn.in_top_k(logits, y, 1)\n        accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))\n\n    with tf.name_scope(\"train\"):\n      if optimize_method == 'nesterov': # nesterov optimizer\n        optimizer = tf.train.MomentumOptimizer(learning_rate=learning_rate, momentum=0.9, use_nesterov=True)\n      elif optimize_method == 'adam': # adam optimizer\n        optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)\n      else: # momentum optimizer\n        optimizer = tf.train.MomentumOptimizer(learning_rate=learning_rate, momentum=0.9)\n      training_op = optimizer.minimize(loss)\n\n  return g, loss, training_op, X, y\n\n\ndef shuffle_batch(X, y, batch_size):\n    rnd_idx = np.random.permutation(len(X))\n    n_batches = len(X) \/\/ batch_size\n    for batch_idx in np.array_split(rnd_idx, n_batches):\n        X_batch, y_batch = X[batch_idx], y[batch_idx]\n        yield X_batch, y_batch\n\ngraph, loss, training_op, X, y = create_graph(100, 10)\n\nn_epochs = 20\nbatch_size = 5\n\ninit = tf.global_variables_initializer()\n# saver = tf.train.Saver()\n\nwith tf.Session(graph=graph) as sess:\n    init.run()\n    print(2)\n    for epoch in range(n_epochs):\n        for X_batch, y_batch in shuffle_batch(X_train, y_train, batch_size):\n            sess.run(training_op, feed_dict={X: X_batch, y: y_batch})\n        mse_batch = loss.eval(feed_dict={X: X_batch, y: y_batch})\n        mse_valid = loss.eval(feed_dict={X: X_valid, y: y_valid})\n        print(epoch, \"Batch mse:\", mse_batch, \"Validation mse:\", mse_valid)\n\n    # save_path = saver.save(sess, \".\/my_model_final.ckpt\")\n<\/code><\/pre>\n","answers":[{"AnswerId":"57958200","CreationDate":"2019-09-16T13:44:38.737","ParentId":null,"OwnerUserId":"10111013","Title":null,"Body":"<p>so the problem here is that you are specifying a graph object other than the default graph for the bulk of your operations (defined in the function as <code>g<\/code>, and outside the function as <code>graph<\/code>), but the operation <code>tf.global_variables_initializer()<\/code> is added to the default graph not <code>graph<\/code>.<\/p>\n\n<p>Operations are added to the graph object when they are declared, not when called by session, so although you have specified<\/p>\n\n<pre><code>with tf.Session(graph=graph) as sess:\n<\/code><\/pre>\n\n<p>the <code>init<\/code> operation has been defined on a different graph to the one specified, and so cant be called in this session. <\/p>\n\n<p>changing the <code>init<\/code> operation declaration to <\/p>\n\n<pre><code>with graph.as_default():\n    init = tf.global_variables_initializer()\n<\/code><\/pre>\n\n<p>will solve. <\/p>\n\n<p><strong>Alternatively<\/strong> you can remove the declaration of <code>init<\/code> all together and just call <code>tf.global_variables_initializer()<\/code> directly within the session like so:<\/p>\n\n<pre><code>with tf.Session(graph=graph) as sess:\n    sess.run(tf.global_variables_initializer())\n    print(2)\n    for epoch in range(n_epochs):\n    ...\n<\/code><\/pre>\n\n<p>which will automatically initialize variables on the graph associated with <code>sess<\/code>. <\/p>\n"}]}
+{"QuestionId":57950901,"AnswerCount":1,"Tags":"<python><pandas><keras><one-hot-encoding><multilabel-classification>","CreationDate":"2019-09-16T05:37:04.067","AcceptedAnswerId":57951017.0,"OwnerUserId":9813733.0,"Title":"One hot encoding of multi label images in keras","Body":"<p>I am using PASCAL VOC 2012 dataset for image classification. A few images have multiple labels where as a few of them have single labels as shown below. <\/p>\n\n<pre><code>    0  2007_000027.jpg               {'person'}\n    1  2007_000032.jpg  {'aeroplane', 'person'}\n    2  2007_000033.jpg            {'aeroplane'}\n    3  2007_000039.jpg            {'tvmonitor'}\n    4  2007_000042.jpg                {'train'}\n<\/code><\/pre>\n\n<p>I want to do one-hot encoding of these labels to train the model. However, I couldn't use keras.utils.to_categorical, as these labels are not integers and pandas.get_dummies is not giving me the results as expected. get_dummies is giving different categories as below, i.e. it is taking each unique combination of labels as one category.<\/p>\n\n<pre><code> {'aeroplane', 'bus', 'car'}  {'aeroplane', 'bus'}  {'tvmonitor', 'sofa'}  {'tvmonitor'} ...\n<\/code><\/pre>\n\n<p>What is the best way to one-hot encode these labels as we don't have specific number of labels for each image. <\/p>\n","answers":[{"AnswerId":"57951017","CreationDate":"2019-09-16T05:49:12.947","ParentId":null,"OwnerUserId":"2901002","Title":null,"Body":"<p>If there are <code>set<\/code>s in second column is possible use <code>MultiLabelBinarizer<\/code>:<\/p>\n\n<pre><code>print (df)\n                 a                        b\n0  2007_000027.jpg               {'person'}\n1  2007_000032.jpg  {'aeroplane', 'person'}\n2  2007_000033.jpg            {'aeroplane'}\n3  2007_000039.jpg            {'tvmonitor'}\n4  2007_000042.jpg                {'train'}\n<\/code><\/pre>\n\n<hr>\n\n<pre><code>from sklearn.preprocessing import MultiLabelBinarizer\n\nmlb = MultiLabelBinarizer()\ndf = pd.DataFrame(mlb.fit_transform(df['b']),columns=mlb.classes_)\nprint (df)\n   aeroplane  person  train  tvmonitor\n0          0       1      0          0\n1          1       1      0          0\n2          1       0      0          0\n3          0       0      0          1\n4          0       0      1          0\n<\/code><\/pre>\n\n<p>Or <a href=\"http:\/\/pandas.pydata.org\/pandas-docs\/stable\/reference\/api\/pandas.Series.str.join.html\" rel=\"nofollow noreferrer\"><code>Series.str.join<\/code><\/a> with <a href=\"http:\/\/pandas.pydata.org\/pandas-docs\/stable\/reference\/api\/pandas.Series.str.get_dummies.html\" rel=\"nofollow noreferrer\"><code>Series.str.get_dummies<\/code><\/a>, but it should be slowier in large DataFrame:<\/p>\n\n<pre><code>df = df['b'].str.join('|').str.get_dummies()\nprint (df)\n\n   aeroplane  person  train  tvmonitor\n0          0       1      0          0\n1          1       1      0          0\n2          1       0      0          0\n3          0       0      0          1\n4          0       0      1          0\n<\/code><\/pre>\n"}]}
+{"QuestionId":57951499,"AnswerCount":2,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-16T06:35:15.380","AcceptedAnswerId":57952272.0,"OwnerUserId":2361465.0,"Title":"Depreciation warnings in Tensorflow Keras official tutorial","Body":"<p>The example code<\/p>\n\n<pre><code>model = tf.keras.Sequential()\n# Adds a densely-connected layer with 64 units to the model:\nmodel.add(layers.Dense(64, activation='relu'))\n# Add another:\nmodel.add(layers.Dense(64, activation='relu'))\n# Add a softmax layer with 10 output units:\nmodel.add(layers.Dense(10, activation='softmax'))\n<\/code><\/pre>\n\n<p>in the official <a href=\"https:\/\/www.tensorflow.org\/guide\/keras\" rel=\"nofollow noreferrer\">Tensorflow website<\/a> results in a warning in the output as seen in that page itself.<\/p>\n\n<pre><code>calling VarianceScaling.__init__ (from tensorflow.python.ops.init_ops) with dtype is deprecated and will be removed in a future version.\n<\/code><\/pre>\n\n<p>What is the reason of the warning? How should the code be changed to avoid such warnings?<\/p>\n","answers":[{"AnswerId":"57953648","CreationDate":"2019-09-16T09:10:32.910","ParentId":null,"OwnerUserId":"5597718","Title":null,"Body":"<blockquote>\n  <p>What is the reason of the warning? How should the code be changed to\n  avoid such warnings?<\/p>\n<\/blockquote>\n\n<p>This warning is applicable only for pre-Tensorflow 2.0. The warning is there because starting from Tensorflow 2.0, <code>dtype<\/code> argument should be passed when calling the initializer instance and not to the constructor. <\/p>\n\n<p>Code snippet to reproduce the warning in pre-tensorflow2.0:<\/p>\n\n<pre><code>from tensorflow.keras.layers import Dense\ndense = Dense(64, activation='relu')\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>calling VarianceScaling.__init__ (from tensorflow.python.ops.init_ops) with dtype is deprecated and will be removed in a future version.\nInstructions for updating:\nCall initializer instance with the dtype argument instead of passing it to the constructor\n<\/code><\/pre>\n\n<p>It is not immediately apparent why the code snippet would cause this warning because no initializer is created. What happens is that internally Tensorflow creates the initializer when a <code>Dense<\/code> layer is initiated, since the default kernel initializer is <code>glorot_uniform<\/code>. Something like this:<\/p>\n\n<pre><code>from tensorflow.keras import initializers\ninitializer = initializers.get('glorot_uniform') \n<\/code><\/pre>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/024675edc62dddab17bc439b69eb9bd71f1a1a9a\/tensorflow\/python\/keras\/layers\/core.py#L997\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/024675edc62dddab17bc439b69eb9bd71f1a1a9a\/tensorflow\/python\/keras\/layers\/core.py#L997<\/a>\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/024675edc62dddab17bc439b69eb9bd71f1a1a9a\/tensorflow\/python\/keras\/layers\/core.py#L1014\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/024675edc62dddab17bc439b69eb9bd71f1a1a9a\/tensorflow\/python\/keras\/layers\/core.py#L1014<\/a><\/p>\n"},{"AnswerId":"57952272","CreationDate":"2019-09-16T07:36:13.203","ParentId":null,"OwnerUserId":"9828197","Title":null,"Body":"<p>Welcome to the club :)  I mean tensorflow is notorious for not being backward compatible and worse yet there are warnings like this now for a long time. <\/p>\n\n<p>See here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/25996\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/25996<\/a><\/p>\n\n<p>So your best bet is just ignore those warnings and, if you wish to be proactive, then contribute with a fix. Otherwise, make sure they do not distract you from your main task. Warnings now are normal in tensorflow.<\/p>\n"}]}
+{"QuestionId":57952045,"AnswerCount":1,"Tags":"<tensorflow><optimization><keras><neural-network>","CreationDate":"2019-09-16T07:18:37.597","AcceptedAnswerId":null,"OwnerUserId":5071038.0,"Title":"Constrained Optimization Tensorflow","Body":"<p>I have a trained classifier neural network in Keras. Let the neural network be f(x). I want to find the vectors x such that when ||x||^2 = 1, f(x) is maximized.  I currently have trained my neural network with Keras<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(500, activation='sigmoid'))\nmodel.add(Dense(500, activation='sigmoid'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(1, activation='sigmoid'))\n\nmodel.compile(loss='binary_crossentropy', optimizer='adam',  metrics=['accuracy', auc])\nmodel.fit(x_train, y_train, validation_data=(x_test, y_test), epochs=2, verbose = 1, callbacks=[earlyStopping])\n<\/code><\/pre>\n\n<p>I want to know if there is a way to solve this constrained optimization problem once my Neural network has already been trained. There is a scipy optimize which can do this for general functions. Is there a way to do this for a neural network. Please include a code sample.<\/p>\n","answers":[{"AnswerId":"57958687","CreationDate":"2019-09-16T14:11:25.313","ParentId":null,"OwnerUserId":"10111013","Title":null,"Body":"<p>If I understand you correctly you have finished training your neural network and would like to find the input <code>x<\/code> which maximizes the probability of it being in a certain class (where the output is close to 1.0)<\/p>\n\n<p>You could write a small function to assess the performance of your network using the <code>predict_proba()<\/code> method to get classification probabilities on test data, and then optimise this function using scipy:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(500, activation='sigmoid'))\nmodel.add(Dense(500, activation='sigmoid'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(1, activation='sigmoid'))\n\nmodel.compile(loss='binary_crossentropy', optimizer='adam',  metrics=['accuracy', auc])\nmodel.fit(x_train, y_train, validation_data=(x_test, y_test), epochs=2, verbose = 1, callbacks=[earlyStopping])\n\ndef f(x):\n    prediction = model.predict_proba(x)\n    return -prediction\n\na = scipy.optimize.minimize(f, x0=np.random.randn(500))\noptimal_x = a.x\n<\/code><\/pre>\n\n<p><code>optimal_x<\/code> will be the input <code>x<\/code> which maximises the certainty with which your classifier puts it in one specific class. <\/p>\n"}]}
+{"QuestionId":57952297,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><deserialization><tfrecord><eager-execution>","CreationDate":"2019-09-16T07:38:11.997","AcceptedAnswerId":null,"OwnerUserId":6677037.0,"Title":"npz decompression in tfrecord parsing","Body":"<p>I want to decompress the numpy array data compressed using savez_compressed within a TFrecord parsing and I'm getting an error: <\/p>\n\n<pre><code>TypeError: integer argument expected, got 'Tensor'. \n<\/code><\/pre>\n\n<p><strong>The writing of the TFRecord:<\/strong> <\/p>\n\n<pre><code>tf.enable_eager_execution()\n\ndef _bytes_feature(value):\n    return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))\n\ndef record_example(record):\n    feature_key_value_pair = {'record': _bytes_feature(record)}\n    features = tf.train.Features(feature=feature_key_value_pair)\n    example = tf.train.Example(features=features)\n    return example\n\n\ndef write_tfrecord(data,path):\n    with tf.python_io.TFRecordWriter(path) as writer:\n        compressed_array = io.BytesIO()\n        np.savez_compressed(compressed_array, data)\n        compressed_array.seek(0)\n        example = record_example(compressed_array.read())\n        writer.write(example.SerializeToString())\n<\/code><\/pre>\n\n<p><strong>The reading of the TFRecord:<\/strong><\/p>\n\n<pre><code>def parse_tfrecord(serialized_example):\n    f_dict={'record': tf.FixedLenFeature([], tf.string)}\n    features = tf.parse_single_example(serialized_example, f_dict)\n    compressed_array = io.BytesIO()\n    compressed_array.read(features['record'])\n    compressed_array.seek(0)\n    feat = np.load(compressed_array)['arr_0']\n    return {'record':feat}\n\nmy_record=np.random.random((1,500)\nwrite_tfrecord(my_record),'temp.tfr')\nds=tf.data.TFRecordDataset(['temp.tfr'])\nall_data=ds.map(parse_tfrecord)\n\nfor d in all_data:\n    rec=d['record']\n    print(rec)\n<\/code><\/pre>\n\n<p>The writing of the tfrecords working well, please help me with the reading.<\/p>\n\n<p>thanks!<\/p>\n","answers":[]}
+{"QuestionId":57953004,"AnswerCount":1,"Tags":"<macos><installation><pytorch>","CreationDate":"2019-09-16T08:26:25.527","AcceptedAnswerId":null,"OwnerUserId":12073062.0,"Title":"GPU driver and CUDA is not enabled and accessible by PyTorch on MacOS","Body":"<p>GPU driver and CUDA is not enabled and accessible by PyTorch.\n<code>torch.cuda.is_available()<\/code> returns <code>false<\/code><\/p>\n\n<p>I am using macOS Mojave 10.14.6\nI have installed Cuda 10.0 version of pytorch. \nI tried verfication on <a href=\"https:\/\/pytorch.org\/get-started\/locally\/\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/get-started\/locally\/<\/a> and constructing a randomly initialized tensor works just fine. <\/p>\n\n<p>But when I tried <\/p>\n\n<pre><code>import torch\ntorch.cuda.is_availalbe()\n<\/code><\/pre>\n\n<p>it returns false.<\/p>\n\n<p>Therefore, I followed instructions on Pytorch and installed Anaconda and Cuda. <\/p>\n\n<p>Then tried this: <\/p>\n\n<p><code>git clone --recursive https:\/\/github.com\/pytorch\/pytorch\ncd pytorch\nMACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py install<\/code><\/p>\n\n<p>In terminal, I got <\/p>\n\n<pre><code>fatal error: 'string.h' file not found\n#include_next &lt;string.h&gt;\n<\/code><\/pre>\n\n<p>I searched on stackoverflow and came up with this. <a href=\"https:\/\/stackoverflow.com\/questions\/55107466\/build-pytorch-from-source\">Build Pytorch from source<\/a> . So I tried:<\/p>\n\n<pre><code>$ find \/Library\/Developer\/CommandLineTools\/usr -type f -name string.h\n<\/code><\/pre>\n\n<p>which returned <code>\/Library\/Developer\/CommandLineTools\/usr\/include\/c++\/v1\/string.h<\/code><\/p>\n\n<p>Doesn't this mean I already have <code>string.h<\/code>?<\/p>\n\n<p>How can I solve this problem?<\/p>\n","answers":[{"AnswerId":"57953283","CreationDate":"2019-09-16T08:46:55.530","ParentId":null,"OwnerUserId":"3411080","Title":null,"Body":"<p>Are you installing from a conda env? According to the github this should work:\n - Create a conda env\n - <code>conda install numpy ninja pyyaml mkl mkl-include setuptools cmake cffi typing<\/code>  (that installs some requirements)\n Then this (which I assume you've already done):<\/p>\n\n<pre><code>git clone --recursive https:\/\/github.com\/pytorch\/pytorch\ncd pytorch\n# if you are updating an existing checkout\ngit submodule sync\ngit submodule update --init --recursive\n<\/code><\/pre>\n\n<p>And finally set up the conda variable and install:<\/p>\n\n<pre><code>export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-\"$(dirname $(which conda))\/..\/\"}\nMACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py install\n<\/code><\/pre>\n\n<p>There are issues on the git reporting that behavior <a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/25669\" rel=\"nofollow noreferrer\">here<\/a> which suggest to add something like:\n<code>MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ NO_CUDA=1 python setup.py install<\/code><\/p>\n\n<p>Check that<code>NO_CUDA<\/code>, this issue as also been mentioned <a href=\"https:\/\/discuss.pytorch.org\/t\/fatal-error-during-mac-source-build\/38436\/3\" rel=\"nofollow noreferrer\">HERE<\/a> in the forums and it seems to be that it could be an issue caused by the OS and driver versions. If that should be the case I recommend to use <a href=\"https:\/\/github.com\/NVIDIA\/nvidia-docker\" rel=\"nofollow noreferrer\">Nvidia Docker<\/a> (hopefully it has mac support) with a pytorch container from <a href=\"https:\/\/ngc.nvidia.com\/catalog\/landing\" rel=\"nofollow noreferrer\">https:\/\/ngc.nvidia.com\/catalog\/landing<\/a>\nIt that should also fail, your best bet is to install without CUDA support.<\/p>\n"}]}
+{"QuestionId":57953100,"AnswerCount":1,"Tags":"<python><lambda><pytorch>","CreationDate":"2019-09-16T08:33:07.160","AcceptedAnswerId":null,"OwnerUserId":8847950.0,"Title":"Python 3 unable to pickle lambda","Body":"<p>I was following a tutorial online about <a href=\"https:\/\/towardsdatascience.com\/deep-learning-for-self-driving-cars-7f198ef4cfa2\" rel=\"nofollow noreferrer\">Deep Learning for Self-Driving Cars<\/a> with pytorch. But then I encounter some problem when python trying to pickle lambda.<\/p>\n\n<p>I've tried to define my own function but then the result is still the same<\/p>\n\n<p>Below is where I define my dataset for training.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>params = {'batch_size': 32, 'shuffle': True, 'num_workers':4}\ntraining_set = Dataset(train_samples, transform=transdorms.Compose([transforms.Lambda(lambda x: (x\/255)-0.5)])\ntraining_generator = DataLoader(training_set, **params)\n<\/code><\/pre>\n\n<p>Then the problem is at the following code.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>for local_batch, (centers, lefts, rights) in enumerate(training generator):\n        ... Some code\n<\/code><\/pre>\n\n<p>The following error is shown<\/p>\n\n<blockquote>\n  <p>_pickle.PicklingError: Can't pickle <code>&lt;function &lt;lambda&gt; at 0x000001B915587948&gt;<\/code>: attribute lookup  on <strong>main<\/strong> failed<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"57953311","CreationDate":"2019-09-16T08:48:14.700","ParentId":null,"OwnerUserId":"9609843","Title":null,"Body":"<p>Python documentation says that <code>lambda<\/code> functions cannot be pickled.<\/p>\n\n<p><a href=\"https:\/\/docs.python.org\/3\/library\/pickle.html#id2\" rel=\"nofollow noreferrer\">https:\/\/docs.python.org\/3\/library\/pickle.html#id2<\/a>:<\/p>\n\n<blockquote>\n  <p>Note that functions (built-in and user-defined) are pickled by \u201cfully\n  qualified\u201d name reference, not by value. 2 This means that only the\n  function name is pickled, along with the name of the module the\n  function is defined in. Neither the function\u2019s code, nor any of its\n  function attributes are pickled. Thus the defining module must be\n  importable in the unpickling environment, and the module must contain\n  the named object, otherwise an exception will be raised.<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/docs.python.org\/3\/library\/pickle.html#id7\" rel=\"nofollow noreferrer\">https:\/\/docs.python.org\/3\/library\/pickle.html#id7<\/a>:<\/p>\n\n<blockquote>\n  <p>This is why lambda functions cannot be pickled: all lambda functions\n  share the same name: .<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":57953585,"AnswerCount":2,"Tags":"<python><matlab><caffe>","CreationDate":"2019-09-16T09:06:04.970","AcceptedAnswerId":57965593.0,"OwnerUserId":11801228.0,"Title":"Problem converting net forward on caffe from MATLAB to Python","Body":"<p>I am trying to convert some MATLAB code to Python because I was unsuccessful in running the code on MATLAB (it crashed whenever I run it). So far I have finished converting the code and run it but the result is very wrong. So I tried to debug it and some part of the codes still gives me uncertainty since I can't check the result on MATLAB<\/p>\n\n<p>This is the MATLAB code that I want to convert<\/p>\n\n<pre class=\"lang-matlab prettyprint-override\"><code>%Prepare images\nim=reshape(im,[size(im)]); im=single(im)\/255;\nim_data = im(:, :, [3, 2, 1]);  % permute channels from RGB to BGR\nim_data = permute(im_data, [2, 1, 3]); \n\n%pass images  \nout_im = net.forward({im_data});\nn_out=out_im{2}; al_out=out_im{1}; light_out=out_im{3};\n<\/code><\/pre>\n\n<p>This is the Python code that I created <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#Prepare images (rotate,flip,change color,reshape)\nim=np.reshape(im,(im.shape))\nim=np.float32(im)\/255\n#already convert to BGR on top\n#im_data = cv2.cvtColor(im, cv2.COLOR_RGB2BGR)\n#Dibalik lalu di rotate 90 ke kiri\nim_data = np.transpose(im, (1, 0, 2)) \nim_data=np.moveaxis(im_data, -1, 0) \nim_input = im_data[np.newaxis, : , :, :]\nnet.blobs['data'].reshape(*im_input.shape)\nnet.blobs['data'].data[...] = im_input\n#pass images  \nout_im = net.forward()\nn_out=out_im['Nconv0'][0] \nal_out=out_im['Aconv0'][0] \nlight_out=out_im['fc_light'][0]\n<\/code><\/pre>\n\n<p>I tried to follow the tutorial for caffe user which is the input should be (N,channel,width,height) so I convert the input into (1,3,128,128) but on the MATLAB code I believe the shape is (width,height,channel). Last time I tried to input with just 3 dimension I got error.<\/p>\n","answers":[{"AnswerId":"57965593","CreationDate":"2019-09-16T23:39:22.593","ParentId":null,"OwnerUserId":"11976900","Title":null,"Body":"<p>When transpose, it is (2,0,1), not (1,0,2).<\/p>\n\n<pre><code>inputs = cv.resize(inputs,(128,128))\ninputs = inputs.transpose((2,0,1))\ninputs = inputs[None,:]\nnet.forward_all( data = inputs )\n<\/code><\/pre>\n\n<p>try this one<\/p>\n"},{"AnswerId":"57968108","CreationDate":"2019-09-17T05:57:18.867","ParentId":null,"OwnerUserId":"11976900","Title":null,"Body":"<pre><code>inputs = cv.resize(inputs,(128,128))\ninputs = inputs * 0.00390625\ninputs = inputs.transpose((2,0,1))\ninputs = inputs[None,:]\nnet.forward_all( data = inputs )\n<\/code><\/pre>\n\n<p>try this one again<\/p>\n"}]}
+{"QuestionId":57953884,"AnswerCount":2,"Tags":"<python-3.x><keras><neural-network><deep-learning>","CreationDate":"2019-09-16T09:24:12.820","AcceptedAnswerId":57954784.0,"OwnerUserId":11435927.0,"Title":"What is difference between metrics=accuracy and precision from confusion matrix?","Body":"<p>What is difference between metrics=accuracy and precision from each class that obtained from confusion matrix?<\/p>\n\n<p>Is that any relationship between accuracy and precision?<\/p>\n\n<p>How can i calculate loss from precision of each class? for example if precision from class 1 was 0.95, my loss is 0.05? <\/p>\n","answers":[{"AnswerId":"57954956","CreationDate":"2019-09-16T10:29:49.190","ParentId":null,"OwnerUserId":"2851685","Title":null,"Body":"<p><strong>Accuracy<\/strong>: Is fraction of correct predictions in all prediction made by model. i.e. \n<code>no. of correct predictions \/ no. of total predictions<\/code><\/p>\n\n<p><strong>Precision<\/strong>: Is fraction of correct positive predictions in all positive predictions made by the model i.e<\/p>\n\n<p><code>no. of correct positives\/ no. of total positive predictions<\/code><\/p>\n\n<p>which is also equal to <\/p>\n\n<p><code>no. of correct positives\/ (no. of correct positives + no. of wrong positives)<\/code><\/p>\n\n<p>There is also a third metric you want to consider<\/p>\n\n<p><strong>Recall<\/strong>: Is fraction of correct positive predictions made in actual positive data i.e<\/p>\n\n<p><code>correct positive predictions\/actual positives<\/code><\/p>\n\n<p>If your data are balanced, i.e. have nearly same number of instance in each class then you can use accuracy, else you can use precision or accuracy based on either classifying instance in positive class or negative class is more important <\/p>\n\n<p>There is not direct relation between loss and precision or any of above parameters because these parameters are calculated based on predictions but loss are calculated with probability of prediction for each class. When we make prediction based on probability we choose class with highest probability in doing so we lost information about it so we can't go back from prediction to probability.<\/p>\n"},{"AnswerId":"57954784","CreationDate":"2019-09-16T10:18:56.637","ParentId":null,"OwnerUserId":"5597718","Title":null,"Body":"<blockquote>\n  <p>What is difference between metrics=accuracy and precision from each\n  class that obtained from confusion matrix?<\/p>\n<\/blockquote>\n\n<p>Accuracy = (Correct predictions)\/(Total number of cases)<\/p>\n\n<p>Precision = (True positive)\/(Total number of positive predictions)<\/p>\n\n<blockquote>\n  <p>How can i calculate loss from precision of each class?<\/p>\n<\/blockquote>\n\n<p>Loss has to be calculated from the loss function defined for your network like <code>binary_crossentropy<\/code>.<\/p>\n"}]}
+{"QuestionId":57953992,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2019-09-16T09:30:58.083","AcceptedAnswerId":null,"OwnerUserId":11027413.0,"Title":"Tensorflow 1.14 can't detect GPU with python 3.6 but see it in Python 2.7","Body":"<p>I am trying to work with Mask R-CNN samples to practice detection and segmentation models. I am using python 3.6.8 with Tensorflow and Keras but program model starts training only with CPU and can't detect any GPU. At the same time Python 2.7 can see my GPU as it shown in code below.<\/p>\n\n<p>I have installed NVidia driver 410, CUDA 10.0, cudnn 7.4.2, tensorflow-gpu==1.14.0 installed with pip and pip3.<\/p>\n\n<p>To check is GPU available i used small python script:<\/p>\n\n<pre><code>from tensorflow.python.client import device_lib\nprint(device_lib.list_local_devices())\n\nimport tensorflow as tf\ntf.test.is_gpu_available()\n<\/code><\/pre>\n\n<p><strong>Output for python:<\/strong><\/p>\n\n<pre><code>2019-09-16 12:17:21.621640: I tensorflow\/core\/platform\/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\n2019-09-16 12:17:21.635692: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcuda.so.1\n2019-09-16 12:17:21.720910: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-09-16 12:17:21.721429: I tensorflow\/compiler\/xla\/service\/service.cc:168] XLA service 0x558027aed070 executing computations on platform CUDA. Devices:\n2019-09-16 12:17:21.721444: I tensorflow\/compiler\/xla\/service\/service.cc:175]   StreamExecutor device (0): GeForce GTX 1060 3GB, Compute Capability 6.1\n2019-09-16 12:17:21.741624: I tensorflow\/core\/platform\/profile_utils\/cpu_utils.cc:94] CPU Frequency: 3408000000 Hz\n2019-09-16 12:17:21.742628: I tensorflow\/compiler\/xla\/service\/service.cc:168] XLA service 0x558027ba37a0 executing computations on platform Host. Devices:\n2019-09-16 12:17:21.742663: I tensorflow\/compiler\/xla\/service\/service.cc:175]   StreamExecutor device (0): &lt;undefined&gt;, &lt;undefined&gt;\n2019-09-16 12:17:21.742937: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-09-16 12:17:21.743542: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1640] Found device 0 with properties: \nname: GeForce GTX 1060 3GB major: 6 minor: 1 memoryClockRate(GHz): 1.7845\npciBusID: 0000:01:00.0\n2019-09-16 12:17:21.743884: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudart.so.10.0\n2019-09-16 12:17:21.745776: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcublas.so.10.0\n2019-09-16 12:17:21.747601: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcufft.so.10.0\n2019-09-16 12:17:21.748031: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcurand.so.10.0\n2019-09-16 12:17:21.750885: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcusolver.so.10.0\n2019-09-16 12:17:21.752588: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcusparse.so.10.0\n2019-09-16 12:17:21.757589: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42]     Successfully opened dynamic library libcudnn.so.7\n2019-09-16 12:17:21.757765: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-09-16 12:17:21.758440: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-09-16 12:17:21.759102: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1763] Adding visible gpu devices: 0\n2019-09-16 12:17:21.759173: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudart.so.10.0\n2019-09-16 12:17:21.760528: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1181] Device interconnect StreamExecutor with strength 1 edge matrix:\n2019-09-16 12:17:21.760553: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1187]      0 \n2019-09-16 12:17:21.760564: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1200] 0:   N \n2019-09-16 12:17:21.760807: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-09-16 12:17:21.761483: I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2019-09-16 12:17:21.762083: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1326] Created TensorFlow device (\/device:GPU:0 with 2287 MB memory) -&gt; physical GPU (device: 0, name: GeForce GTX 1060 3GB, pci bus id: 0000:01:00.0, compute capability: 6.1)\n[name: \"\/device:CPU:0\"\ndevice_type: \"CPU\"\nmemory_limit: 268435456\nlocality {\n}\nincarnation: 13231116977719504797\n, name: \"\/device:XLA_GPU:0\"\ndevice_type: \"XLA_GPU\"\nmemory_limit: 17179869184\nlocality {\n}\nincarnation: 2220022223072533235\nphysical_device_desc: \"device: XLA_GPU device\"\n, name: \"\/device:XLA_CPU:0\"\ndevice_type: \"XLA_CPU\"\nmemory_limit: 17179869184\nlocality {\n}\nincarnation: 579036656168733531\nphysical_device_desc: \"device: XLA_CPU device\"\n, name: \"\/device:GPU:0\"\ndevice_type: \"GPU\"\nmemory_limit: 2398158848\nlocality {\n  bus_id: 1\n  links {\n  }\n}\nincarnation: 14902085726432239020\nphysical_device_desc: \"device: 0, name: GeForce GTX 1060 3GB, pci bus id: 0000:01:00.0, compute capability: 6.1\"\n]\n<\/code><\/pre>\n\n<p><strong>Output for python3:<\/strong><\/p>\n\n<pre><code>WARNING: Logging before flag parsing goes to stderr.\nW0916 12:10:46.067314 140410557671232 __init__.py:308] Limited tf.compat.v2.summary API due to missing TensorBoard installation.\n2019-09-16 12:10:46.084425: I tensorflow\/core\/platform\/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\n2019-09-16 12:10:46.113540: I tensorflow\/core\/platform\/profile_utils\/cpu_utils.cc:94] CPU Frequency: 3408000000 Hz\n2019-09-16 12:10:46.114038: I tensorflow\/compiler\/xla\/service\/service.cc:168] XLA service 0x495fd00 executing computations on platform Host. Devices:\n2019-09-16 12:10:46.114054: I tensorflow\/compiler\/xla\/service\/service.cc:175]   StreamExecutor device (0): &lt;undefined&gt;, &lt;undefined&gt;\n[name: \"\/device:CPU:0\"\ndevice_type: \"CPU\"\nmemory_limit: 268435456\nlocality {\n}\nincarnation: 5355211451834126024\n, name: \"\/device:XLA_CPU:0\"\ndevice_type: \"XLA_CPU\"\nmemory_limit: 17179869184\nlocality {\n}\nincarnation: 8307771055363446310\nphysical_device_desc: \"device: XLA_CPU device\"\n]\n<\/code><\/pre>\n\n<p>Where did I make mistake and how to fix this issue and detect GPU for Python3?<\/p>\n","answers":[]}
+{"QuestionId":57953999,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-16T09:31:35.613","AcceptedAnswerId":57954082.0,"OwnerUserId":8047378.0,"Title":"TypeError: __call__() takes 2 positional arguments but 3 were given","Body":"<p>I get this error when I use my own layer:<\/p>\n\n<pre><code>Traceback (most recent call last):\nFile \"E:\/fffan\/try.py\", line 40, in &lt;module&gt;\n run(input, 5000)\nFile \"E:\/fffan\/try.py\", line 36, in run\n out = SelfAttention(nclass)(output, state)\nTypeError: __call__() takes 2 positional arguments but 3 were given\n<\/code><\/pre>\n\n<p>Here is my code, and holp someone can tell me how to fix it.<\/p>\n\n<pre><code>from keras.engine.topology import Layer\nfrom keras.layers.core import Dense\nfrom keras import backend as K\nfrom keras.layers import Input,CuDNNGRU\nfrom keras.activations import softmax\n\nclass SelfAttention(Layer):   #### \u5bf9\u4e8e\u957f\u5e8f\u5217\u6548\u679c\u8f83\u5dee\n    def __init__(self,units,**kwargs):\n        self.W1 = Dense(units)\n        self.W2 = Dense(units)\n        self.V = Dense(1)\n        super(SelfAttention, self).__init__(**kwargs)\n\n    def call(self,features, hidden):\n        hidden_with_time_axis = K.expand_dims(hidden, 1)\n        score = self.V(K.tanh(self.W1(features) + self.W2(hidden_with_time_axis)))\n        attention_weights = softmax(score, axis=1)\n        context_vector = attention_weights * features\n        return context_vector\n\ndef GRU(units):\n    return CuDNNGRU(units, return_sequences=True,\n                return_state=True,\n                recurrent_initializer='glorot_uniform')\n\ndef run(input,nclass):\n    output, state = GRU(nclass)(input)\n    out = SelfAttention(nclass)(output, state)\n\nif __name__ == '__main__':\n    input = Input(shape=(35, 512), name='the_input')\n    run(input, 5000)\n<\/code><\/pre>\n\n<p>My tensorflow version is 1.14.0\uff0c and my keras is 2.1.5<\/p>\n\n<p>Does somebody know anything adout this issue ?<\/p>\n","answers":[{"AnswerId":"57954082","CreationDate":"2019-09-16T09:37:10.677","ParentId":null,"OwnerUserId":"5099803","Title":null,"Body":"<p>It should be:<\/p>\n\n<pre><code>def __call__(self, features, hidden):\n<\/code><\/pre>\n\n<p>instead of:<\/p>\n\n<pre><code>def call(self, features, hidden):\n<\/code><\/pre>\n"}]}
+{"QuestionId":57954717,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-lite><tensorflow.js><coreml>","CreationDate":"2019-09-16T10:15:33.450","AcceptedAnswerId":null,"OwnerUserId":4511899.0,"Title":"Converting TensorflowJS's BodyPix model to TensorFlow Lite","Body":"<p>I have come across googles new BodyPix TensorflowJS model today, and I want to get it running on Android and iOS, but using TensorFlow Lite and CoreML. I was wondering if someone could point me towards the best way to convert this model into TensorFlow Lite. I have done conversions between TensorFlow Lite and CoreML so that's no problem.<\/p>\n\n<p>I've read a few documents on how to do this, but im a little confused as those documents mention a model.json file or something similar, which the BodyPix src directory contains a multitude of files, which im unsure on what any of them do.<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tfjs-models\/tree\/master\/body-pix\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tfjs-models\/tree\/master\/body-pix<\/a><\/p>\n\n<p>Any help or pointers would be appreciated.<\/p>\n","answers":[]}
+{"QuestionId":57955005,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><lstm>","CreationDate":"2019-09-16T10:33:00.863","AcceptedAnswerId":null,"OwnerUserId":8597324.0,"Title":"time and volume prediction","Body":"<p>I am learning ML and did some basic model training on colab. Right now I wanted to train a model which can predict time and volume. Here's the scenario:<\/p>\n\n<p>I have a database with N number of users (consider user_id), and then each user has some job to do (consider this job_no). Every job has a time_taken field (which is in minutes) and a quantity field (consider this volume).<\/p>\n\n<p>Now I wanted to train a model, which can predict time_taken and quantity for a given job of a given user. Let's say I wanted to predict time_taken and quantity for user-1, job-2.<\/p>\n\n<p>Can anyone point me in the right direction with the help of some examples or reference links<\/p>\n","answers":[{"AnswerId":"57955369","CreationDate":"2019-09-16T10:56:07.420","ParentId":null,"OwnerUserId":"2851685","Title":null,"Body":"<p>Your question is quite broad, based on my understanding of your problem I would direct you to following two options<\/p>\n\n<ol>\n<li><p>A simple solution, calculate median value of time and volume for each combination of user_id and job_id and use it as prediction to make prediction for existing user_id and job_id combination<\/p><\/li>\n<li><p>Use regression, in my opinion your data is not suitable for this case as userid is not related in anyway related to time or volume so it is hard for regression model to find any meaningful relation, (in other word any user_id could have be assigned to any user and new user's will also be assigned any user_id) so you may want to find some more data about the user like, age, weight etc. which has some relation with the output.<br>\nFor job_id you can convert it to on hot encoded category and use as features for regression model like linear regression.<br>\nNote: in case of user_id if you want to use the model with to predict only for users in the training dataset then model may find some complex relation there too<\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":57955471,"AnswerCount":0,"Tags":"<tensorflow><computer-vision><dataset><label><classification>","CreationDate":"2019-09-16T11:03:04.747","AcceptedAnswerId":null,"OwnerUserId":11621702.0,"Title":"Creating an .csv Image Dataset for Classification - with label & pixel values","Body":"<p>I have 2 folders of images and I need to prepare an image dataset for image classification using Tensorflow 2.0. I only have two classes, each folder contains one class. <\/p>\n\n<p>Ideally I would love to know the following: \n- How to create the .csv file with \n1st column: label \nfollowing columns: 255 pixel values<\/p>\n\n<p>and finally how do I split the dataset to 2 .csv files for train &amp; test set randomly? <\/p>\n","answers":[]}
+{"QuestionId":57955969,"AnswerCount":0,"Tags":"<python><keras><layer><categorical-data><tensor>","CreationDate":"2019-09-16T11:34:15.900","AcceptedAnswerId":null,"OwnerUserId":11903541.0,"Title":"ValueError: Output tensors to a Model must be the output of a Keras `Layer` (thus holding past layer metadata)","Body":"<p>I have read the similar question but the answers did not work for me even If I try it in several ways. This are some of the questions I search and try here:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/50825248\/keras-backend-reshape-typeerror-failed-to-convert-object-of-type-class-list\">Keras.backend.reshape: TypeError: Failed to convert object of type &lt;class &#39;list&#39;&gt; to Tensor. Consider casting elements to a supported type<\/a><\/p>\n\n<p>or<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/55908329\/tensorflow-valueerror-output-tensors-to-a-model-must-be-the-output-of-a-tenso\">Tensorflow - ValueError: Output tensors to a Model must be the output of a TensorFlow `Layer`<\/a><\/p>\n\n<p>I will explain my code.\nThe dataset is 15 million rows with 4 dimensions and represent one day of device_id location. \nI have two type of data: numerical data and categorical date. The categorical data is extremely huge (more than 4 millions different device_id) therefore I can't use one hot encoder because it create an incredible huge sparse matrix. I did use for the categorical data Tensorflow embedding.\nI did use Functional Models and merge the categorical layers and the numerical layer into one. The goal of the Neural Network is to predict future device_id location.\nI will put a sniped of the data and then code and the code<\/p>\n\n<pre><code>device_id                    seconds      latitude     longitude\n\njg4M\/taYRc2cBJPGa8c7vw==       752        53.392060    -2.069796\njg4M\/taYRc2cBJPGa8c8vw==       753        53.392060    -2.069796\nPdhyfpkIT8Weslb54thwuQ==       754        51.496483     0.116620   \n\n<\/code><\/pre>\n\n<p>I put the device ID in one panda series and transformed into tensors<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>data1000 = data1['device_id']\n\ndata_tf = tf.convert_to_tensor(data1000)\n\nsess = tf.InteractiveSession()  \nprint(data_tf.eval())\n\nsess.close()\n<\/code><\/pre>\n\n<p>I did the same with each one of the numerical columns: Long, Lat and Seconds.\nSince now the have dimension (1,) but the embedding layer will add another dimension, the output dimension, then the categorical layer will become (none,1,1) I did reshape the numerical data one into (1,1)\nI did this so they can have the same shape when numerical and categorical layers merge. I know this because it id raise an error.\nHere is the code for this. I will show the code only for Latitude assuming is self explanatory for Longitude and Seconds since the process is the same.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>data1_lat = data1['latitude']\n\ntype(data1_sec)\npandas.core.series.Series\ndata1_lat.values.reshape(15904370,1)\n\ndata_tf_lat = tf.convert_to_tensor(data1_lat)\n\nsess = tf.InteractiveSession()  \nprint(data_tf_lat.eval())\n\nsess.close()\n<\/code><\/pre>\n\n<p>Now we go to the Neural Network<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>data_tf_lat = Input(shape=(1,1))\ndata_tf_long = Input(shape=(1,1))\ndata_tf_sec = Input(shape=(1,1))\ncat_in       = Input(shape=(1,))\nembed_layer  = Embedding(input_dim=1, output_dim=1, input_length=(len_data1000))(cat_in)\n#embed_layer  = Flatten(embed_layer)\nmerged_layer = concatenate([embed_layer,data_tf_lat,data_tf_long,data_tf_sec])\noutput       = LSTM(128)(merged_layer)\noutput       = Dense(10)(output)\noutput       = Dense(1)\nmodel        = Model(inputs=[data_tf_lat,data_tf_long,data_tf_sec, cat_in], outputs=[output])\n\nValueError: Output tensors to a Model must be the output of a Keras `Layer` (thus holding past layer metadata). Found: &lt;keras.layers.core.Dense object at 0x7ff04499db38&gt;\n<\/code><\/pre>\n\n<p>Any help\nI try several methodologies I find here and in other places but I can't find a solution?`\nThank <\/p>\n","answers":[]}
+{"QuestionId":57956630,"AnswerCount":1,"Tags":"<python><keras><keras-layer><tf.keras><regularized>","CreationDate":"2019-09-16T12:10:29.923","AcceptedAnswerId":57957591.0,"OwnerUserId":2011067.0,"Title":"Regularization of activations once the model is initialized","Body":"<p>I would like to add regularization of activations in Tensorflow.keras on a pretrained network, using a loop over layers.<\/p>\n\n<p>If I want to regularize <strong>weights<\/strong> or <strong>biases<\/strong>, I can do:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>l1=0.001\nl2=0.001\nfor layer in model.layers:\n    if isinstance(layer, DepthwiseConv2D):\n        layer.add_loss(regularizers.l1_l2(l1,l2)(layer.depthwise_kernel))\n    elif isinstance(layer, layers.Conv2D) or isinstance(layer, layers.Dense):\n        layer.add_loss(regularizers.l1_l2(l1,l2)(layer.kernel))\n    if hasattr(layer, 'bias_regularizer') and layer.use_bias: \n        layer.add_loss(regularizers.l1_l2(l1,l2)(layer.bias))\n<\/code><\/pre>\n\n<p>As far as I understood and tested: this is working.<\/p>\n\n<p>However I do not see clearly how to do this for <strong>activations<\/strong> regularization. Specifically, I want to add to the loss the OUTPUT of the Activation layer.<\/p>\n\n<p>I guess I should do something like:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>for layer in model.layers:\n    if isinstance(layer, Activation):\n        layer.add_loss(regularizers.l1_l2(l1,l2)(layer.XXX))\n<\/code><\/pre>\n\n<p>But it is not clear to me what should be replacing XXX in the above.<\/p>\n\n<p>Thanks in advance for your help!<\/p>\n","answers":[{"AnswerId":"57957591","CreationDate":"2019-09-16T13:10:32.000","ParentId":null,"OwnerUserId":"5597718","Title":null,"Body":"<blockquote>\n  <p>But it is not clear to me what should be replacing XXX in the above.<\/p>\n<\/blockquote>\n\n<pre><code>layer.output\n<\/code><\/pre>\n"}]}
+{"QuestionId":57957101,"AnswerCount":1,"Tags":"<python><pip><pytorch>","CreationDate":"2019-09-16T12:40:37.417","AcceptedAnswerId":57999435.0,"OwnerUserId":11383441.0,"Title":"pytorch cannot install using anaconda prompt","Body":"<p><code>C:\\Users\\User&gt;pip install pytorch<\/code><\/p>\n\n<p>When I run this command anaconda prompt display this error.<\/p>\n\n<pre><code>Collecting pytorch\n  Downloading https:\/\/files.pythonhosted.org\/packages\/ee\/67\/f403d4ae6e9cd74b546ee88cccdb29b8415a9c1b3d80aebeb20c9ea91d96\/pytorch-1.0.2.tar.gz\nBuilding wheels for collected packages: pytorch\n  Building wheel for pytorch (setup.py) ... error\n  ERROR: Complete output from command 'C:\\Users\\User\\Anaconda3\\python.exe' -u -c 'import setuptools, tokenize;__file__='\"'\"'C:\\\\Users\\\\User\\\\AppData\\\\Local\\\\Temp\\\\pip-install-yto7o19p\\\\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\\User\\AppData\\Local\\Temp\\pip-wheel-lqbohuk2' --python-tag cp37:\n  ERROR: Traceback (most recent call last):\n    File \"&lt;string&gt;\", line 1, in &lt;module&gt;\n    File \"C:\\Users\\User\\AppData\\Local\\Temp\\pip-install-yto7o19p\\pytorch\\setup.py\", line 15, in &lt;module&gt;\n      raise Exception(message)\n  Exception: You tried to install \"pytorch\". The package named for PyTorch is \"torch\"\n  ----------------------------------------\n  ERROR: Failed building wheel for pytorch\n  Running setup.py clean for pytorch\nFailed to build pytorch\nInstalling collected packages: pytorch\n  Running setup.py install for pytorch ... error\n    ERROR: Complete output from command 'C:\\Users\\User\\Anaconda3\\python.exe' -u -c 'import setuptools, tokenize;__file__='\"'\"'C:\\\\Users\\\\User\\\\AppData\\\\Local\\\\Temp\\\\pip-install-yto7o19p\\\\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\\User\\AppData\\Local\\Temp\\pip-record-oah_3fo2\\install-record.txt' --single-version-externally-managed --compile:\n    ERROR: Traceback (most recent call last):\n      File \"&lt;string&gt;\", line 1, in &lt;module&gt;\n      File \"C:\\Users\\User\\AppData\\Local\\Temp\\pip-install-yto7o19p\\pytorch\\setup.py\", line 11, in &lt;module&gt;\n        raise Exception(message)\n    Exception: You tried to install \"pytorch\". The package named for PyTorch is \"torch\"\n    ----------------------------------------\nERROR: Command \"'C:\\Users\\User\\Anaconda3\\python.exe' -u -c 'import setuptools, tokenize;__file__='\"'\"'C:\\\\Users\\\\User\\\\AppData\\\\Local\\\\Temp\\\\pip-install-yto7o19p\\\\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\\User\\AppData\\Local\\Temp\\pip-record-oah_3fo2\\install-record.txt' --single-version-externally-managed --compile\" failed with error code 1 in C:\\Users\\User\\AppData\\Local\\Temp\\pip-install-yto7o19p\\pytorch\\\n<\/code><\/pre>\n\n<p><code>C:\\Users\\User&gt;pip install torch<\/code><\/p>\n\n<p>When I tried to install <code>torch<\/code> it gives me this error message<\/p>\n\n<pre><code>Collecting torch\n  Downloading https:\/\/files.pythonhosted.org\/packages\/f8\/02\/880b468bd382dc79896eaecbeb8ce95e9c4b99a24902874a2cef0b562cea\/torch-0.1.2.post2.tar.gz (128kB)\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| 133kB 544kB\/s\nRequirement already satisfied: pyyaml in c:\\users\\user\\anaconda3\\lib\\site-packages (from torch) (5.1.1)\nBuilding wheels for collected packages: torch\n  Building wheel for torch (setup.py) ... error\n  ERROR: Complete output from command 'C:\\Users\\User\\Anaconda3\\python.exe' -u -c 'import setuptools, tokenize;__file__='\"'\"'C:\\\\Users\\\\User\\\\AppData\\\\Local\\\\Temp\\\\pip-install-5hkp2cz9\\\\torch\\\\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\\User\\AppData\\Local\\Temp\\pip-wheel-64iazehi' --python-tag cp37:\n  ERROR: running bdist_wheel\n  running build\n  running build_deps\n  Traceback (most recent call last):\n    File \"&lt;string&gt;\", line 1, in &lt;module&gt;\n    File \"C:\\Users\\User\\AppData\\Local\\Temp\\pip-install-5hkp2cz9\\torch\\setup.py\", line 265, in &lt;module&gt;\n      description=\"Tensors and Dynamic neural networks in Python with strong GPU acceleration\",\n    File \"C:\\Users\\User\\Anaconda3\\lib\\site-packages\\setuptools\\__init__.py\", line 145, in setup\n      return distutils.core.setup(**attrs)\n    File \"C:\\Users\\User\\Anaconda3\\lib\\distutils\\core.py\", line 148, in setup\n      dist.run_commands()\n    File \"C:\\Users\\User\\Anaconda3\\lib\\distutils\\dist.py\", line 966, in run_commands\n      self.run_command(cmd)\n    File \"C:\\Users\\User\\Anaconda3\\lib\\distutils\\dist.py\", line 985, in run_command\n      cmd_obj.run()\n    File \"C:\\Users\\User\\Anaconda3\\lib\\site-packages\\wheel\\bdist_wheel.py\", line 192, in run\n      self.run_command('build')\n    File \"C:\\Users\\User\\Anaconda3\\lib\\distutils\\cmd.py\", line 313, in run_command\n      self.distribution.run_command(command)\n    File \"C:\\Users\\User\\Anaconda3\\lib\\distutils\\dist.py\", line 985, in run_command\n      cmd_obj.run()\n    File \"C:\\Users\\User\\Anaconda3\\lib\\distutils\\command\\build.py\", line 135, in run\n      self.run_command(cmd_name)\n    File \"C:\\Users\\User\\Anaconda3\\lib\\distutils\\cmd.py\", line 313, in run_command\n      self.distribution.run_command(command)\n    File \"C:\\Users\\User\\Anaconda3\\lib\\distutils\\dist.py\", line 985, in run_command\n      cmd_obj.run()\n    File \"C:\\Users\\User\\AppData\\Local\\Temp\\pip-install-5hkp2cz9\\torch\\setup.py\", line 51, in run\n      from tools.nnwrap import generate_wrappers as generate_nn_wrappers\n  ModuleNotFoundError: No module named 'tools.nnwrap'\n  ----------------------------------------\n  ERROR: Failed building wheel for torch\n  Running setup.py clean for torch\n  ERROR: Complete output from command 'C:\\Users\\User\\Anaconda3\\python.exe' -u -c 'import setuptools, tokenize;__file__='\"'\"'C:\\\\Users\\\\User\\\\AppData\\\\Local\\\\Temp\\\\pip-install-5hkp2cz9\\\\torch\\\\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  ERROR: running clean\n  error: [Errno 2] No such file or directory: '.gitignore'\n  ----------------------------------------\n  ERROR: Failed cleaning build dir for torch\nFailed to build torch\nInstalling collected packages: torch\n  Running setup.py install for torch ... error\n    ERROR: Complete output from command 'C:\\Users\\User\\Anaconda3\\python.exe' -u -c 'import setuptools, tokenize;__file__='\"'\"'C:\\\\Users\\\\User\\\\AppData\\\\Local\\\\Temp\\\\pip-install-5hkp2cz9\\\\torch\\\\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\\User\\AppData\\Local\\Temp\\pip-record-en00_2m3\\install-record.txt' --single-version-externally-managed --compile:\n    ERROR: running install\n    running build_deps\n    Traceback (most recent call last):\n      File \"&lt;string&gt;\", line 1, in &lt;module&gt;\n      File \"C:\\Users\\User\\AppData\\Local\\Temp\\pip-install-5hkp2cz9\\torch\\setup.py\", line 265, in &lt;module&gt;\n        description=\"Tensors and Dynamic neural networks in Python with strong GPU acceleration\",\n      File \"C:\\Users\\User\\Anaconda3\\lib\\site-packages\\setuptools\\__init__.py\", line 145, in setup\n        return distutils.core.setup(**attrs)\n      File \"C:\\Users\\User\\Anaconda3\\lib\\distutils\\core.py\", line 148, in setup\n        dist.run_commands()\n      File \"C:\\Users\\User\\Anaconda3\\lib\\distutils\\dist.py\", line 966, in run_commands\n        self.run_command(cmd)\n      File \"C:\\Users\\User\\Anaconda3\\lib\\distutils\\dist.py\", line 985, in run_command\n        cmd_obj.run()\n      File \"C:\\Users\\User\\AppData\\Local\\Temp\\pip-install-5hkp2cz9\\torch\\setup.py\", line 99, in run\n        self.run_command('build_deps')\n      File \"C:\\Users\\User\\Anaconda3\\lib\\distutils\\cmd.py\", line 313, in run_command\n        self.distribution.run_command(command)\n      File \"C:\\Users\\User\\Anaconda3\\lib\\distutils\\dist.py\", line 985, in run_command\n        cmd_obj.run()\n      File \"C:\\Users\\User\\AppData\\Local\\Temp\\pip-install-5hkp2cz9\\torch\\setup.py\", line 51, in run\n        from tools.nnwrap import generate_wrappers as generate_nn_wrappers\n    ModuleNotFoundError: No module named 'tools.nnwrap'\n    ----------------------------------------\nERROR: Command \"'C:\\Users\\User\\Anaconda3\\python.exe' -u -c 'import setuptools, tokenize;__file__='\"'\"'C:\\\\Users\\\\User\\\\AppData\\\\Local\\\\Temp\\\\pip-install-5hkp2cz9\\\\torch\\\\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\\User\\AppData\\Local\\Temp\\pip-record-en00_2m3\\install-record.txt' --single-version-externally-managed --compile\" failed with error code 1 in C:\\Users\\User\\AppData\\Local\\Temp\\pip-install-5hkp2cz9\\torch\\\n<\/code><\/pre>\n\n<p>Other packsges can install by using <code>pip install<\/code> command but <code>pytorch<\/code> and <code>torch<\/code> didn't work like that<\/p>\n","answers":[{"AnswerId":"57999435","CreationDate":"2019-09-18T19:13:48.163","ParentId":null,"OwnerUserId":"11383441","Title":null,"Body":"<p>I found solution for my problem. I have use this command<\/p>\n\n<pre><code>C:\\Users\\User&gt;conda install PyTorch -c PyTorch\n<\/code><\/pre>\n\n<p>then I have run<\/p>\n\n<pre><code>C:\\Users\\User&gt;pip3 install torchvision\n<\/code><\/pre>\n"}]}
+{"QuestionId":57957346,"AnswerCount":0,"Tags":"<keras><conv-neural-network><lstm>","CreationDate":"2019-09-16T12:54:06.100","AcceptedAnswerId":null,"OwnerUserId":12074156.0,"Title":"Binary Classification of Vibration Measurements (1D) via CNN-LSTM Network. Input shape problem","Body":"<p>I'd like to make a binary classification of vibration measurements. In a paper I found the following network, which I wanted to copy simplified. (<a href=\"https:\/\/i.stack.imgur.com\/g55ur.png\" rel=\"nofollow noreferrer\">CNN-LSTM structure<\/a>, <a href=\"https:\/\/www.google.com\/url?sa=t&amp;rct=j&amp;q=&amp;esrc=s&amp;source=web&amp;cd=1&amp;cad=rja&amp;uact=8&amp;ved=2ahUKEwiY-eXZrtXkAhXPDOwKHRTmD8QQFjAAegQIAhAC&amp;url=https%3A%2F%2Fwww.phmpapers.org%2Findex.php%2Fphmconf%2Farticle%2Fdownload%2F545%2Fphmc_18_545&amp;usg=AOvVaw20q8oCReaXESO8zuCRV2Ie\" rel=\"nofollow noreferrer\">LINK to the paper - download<\/a>)<\/p>\n\n<p>I now have a problem with the input for the TimeDistributed(Conv1D()) and can't find the error. The error massage is copied below the code. It says I am missing one input dimenstion.<\/p>\n\n<p>My data has the following structure:<\/p>\n\n<p>2951 Measurements (Label = good) + 3043 Measurements (Label = bad) = 5994 Samples in total.<\/p>\n\n<p>16000 measuring points per measurement<\/p>\n\n<p>1D measurements<\/p>\n\n<p>I hope you can help me, many thanks for your time!<\/p>\n\n<p>My Code:<\/p>\n\n<pre><code>xGood = dfGood.values\nyGood = pd.DataFrame(index=range(len(dfGood)), columns=range(1))\nyGood.iloc[:len(yGood)] = 0\nyGood = yGood.values\n\nxBad = dfBad.values\nyBad = pd.DataFrame(index=range(len(dfBad)), columns=range(1))\nyBad.iloc[:len(yBad)] = 1\nyBad = yBad.values\n\nx = np.concatenate((xBad, xGood), axis=0)\ny = np.concatenate((yBad, yGood), axis=0)\n\n\n\ntrainX, testX, trainy, testy = train_test_split(x, y, test_size=0.33, random_state=42)\ntrainy = to_categorical(trainy)\ntesty = to_categorical(testy)\nn_timesteps, n_outputs = trainX.shape[1], trainy.shape[1]\ntrainX = np.expand_dims(trainX, axis=2)\ntestX = np.expand_dims(testX, axis=2)\n\n#Model\ndef evaluate_model(trainX, trainy, testX, testy):\n    verbose, epochs, batch_size = 1, 50, 400\n    n_timesteps, n_outputs = trainX.shape[1], trainy.shape[1]\n\n    model = Sequential()\n    #CNN Model\n    model.add(TimeDistributed(Conv1D(filters=16, kernel_size=12001, activation='relu'), input_shape=(5994,n_timesteps, 1)))\n    model.add(TimeDistributed(Conv1D(filters=250, kernel_size=3985, activation='relu')))\n    model.add(TimeDistributed(MaxPooling1D()))\n    model.add(TimeDistributed(Flatten()))\n    #LSTM Model\n    model.add(LSTM(16))\n    model.add(Dense(n_outputs, activation='softmax'))\n\n    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n    # fit network\n    model.fit(trainX, trainy, epochs=epochs, batch_size=batch_size, verbose=verbose)\n    # evaluate model\n    _, accuracy = model.evaluate(testX, testy, batch_size=batch_size, verbose=0)\n    print(model.summary())\n    print(accuracy)\n\n    model.save(pathModel)\n    return accuracy\n\nevaluate_model(trainX, trainy, testX, testy)\n<\/code><\/pre>\n\n<p>The Error I get: <\/p>\n\n<p>ValueError: Error when checking input: expected time_distributed_28_input to have 4 dimensions, but got array with shape (4017, 16000, 1)<\/p>\n","answers":[]}
+{"QuestionId":57957742,"AnswerCount":0,"Tags":"<tensorflow><keras><lstm><tensorflow-lite>","CreationDate":"2019-09-16T13:17:38.170","AcceptedAnswerId":null,"OwnerUserId":11947073.0,"Title":"Converting keras model to tensorflow lite gives \"FailedPreconditionError\"","Body":"<p>I have a model in <code>keras<\/code> using 1 layer of LSTM with bidirectional wrapper, which I want to convert to <code>tensorflow<\/code> lite. <\/p>\n\n<p>I'm using the callback <code>ModelCheckpoint<\/code> while training the model to save the model and the best weights.<\/p>\n\n<p>Then I'm using this code to reload the best trained model from the checkpoint:<\/p>\n\n<pre><code>predictor = None\npath_Load = os.path.join(os.getcwd(),'LSTMB_CheckPoints.hdf5')\npredictor = load_model(path_Load)\npredictor.load_weights(path_Load)\n<\/code><\/pre>\n\n<p>Upon checking with validation data, the model loads succesfully and works as intended. Now I want to convert it to Tensorflow Lite, with a bit of code I found on stackoverflow - <\/p>\n\n<pre><code>keras_file = path_Load\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>I thought maybe the checkpoint file was causing problems, so I resaved the model, and called converter again using - <\/p>\n\n<pre><code>keras_file = \"keras_model.h5\"\ntf.keras.models.save_model(predictor, keras_file)\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>upon which I am getting this error - <\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.FailedPreconditionError: Attempting to use uninitialized value bidirectional_1\/backward_lstm_1\/kernel\n     [[{{node _retval_bidirectional_1\/backward_lstm_1\/kernel_0_0}}]]\n<\/code><\/pre>\n\n<p>I tried to run the global variable initializer function of tensorflow <\/p>\n\n<pre><code>predictor = None\npath_Load = os.path.join(os.getcwd(),'LSTMB_CheckPoints.hdf5')\npredictor = load_model(path_Load)\npredictor.load_weights(path_Load)\nwith tf.Session() as sess:\n     sess.run(tf.global_variables_initializer())\nkeras_file = \"keras_model.h5\"\ntf.keras.models.save_model(predictor, keras_file)\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>before resaving the model, to check whether the model variables were getting uninitialized, but the error still remains. <\/p>\n\n<p>has anyone faced a similar issue and found a solution? Is there a way to convert the sequential model to tflite directly without saving and reloading the file?<\/p>\n","answers":[]}
+{"QuestionId":57957761,"AnswerCount":0,"Tags":"<python><tensorflow><keras><spectrogram>","CreationDate":"2019-09-16T13:18:53.467","AcceptedAnswerId":null,"OwnerUserId":11911827.0,"Title":"feeding variable length spectrograms to tensorflow keras","Body":"<p>My inputs are N mp3 files which i've turned into spectrograms of shape <code>(128,X)<\/code>. Each one has a different <code>X<\/code> depending on the duration of the mp3 example. <\/p>\n\n<p>How do I feed these spectrograms to keras. If they were 1D data of a fixed length, I would make a single feature <code>ndarray<\/code> where each example would correspond to a single row in the <code>ndarray<\/code> 'matrix'. But since the length isn't fixed and each example isn't a vector but a 2d <code>ndarray<\/code>, i'm not sure how to create a single feature collection to pass to Keras. Is there a trick to it? <\/p>\n","answers":[]}
+{"QuestionId":57958230,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-16T13:46:23.737","AcceptedAnswerId":null,"OwnerUserId":9582951.0,"Title":"cuda memory allocation failure in tensorflow","Body":"<p>I'm trying to solve CUDA MEMORY ALLOCATION failure problem in tensorflow, please help me out.\nusing:\ncuda=10 \ntensorflow==1.14<\/p>\n\n<blockquote>\n  <p>E tensorflow\/stream_executor\/cuda\/cuda_driver.cc:828] failed to\n  allocate 5.40G (5798205696 bytes) from device:\n  CUDA_ERROR_OUT_OF_MEMORY: out of memory E\n  tensorflow\/stream_executor\/cuda\/cuda_driver.cc:828] failed to allocate\n  4.86G (5218384896 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY: out of memory E tensorflow\/stream_executor\/cuda\/cuda_driver.cc:828] failed to\n  allocate 4.37G (4696546304 bytes) from device:\n  CUDA_ERROR_OUT_OF_MEMORY: out of memory LLVM ERROR: out of memory<\/p>\n<\/blockquote>\n\n<p>i have tried some soln which is already there but still it's not working.\ni have tried:<\/p>\n\n<p>1.<\/p>\n\n<pre><code>gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.9)\n  config = tf.ConfigProto(gpu_options=gpu_options)\n<\/code><\/pre>\n\n<p>2.<\/p>\n\n<pre><code>config = tf.configConfigProto()\n  config.gpu_options.allow_growth = True\/False\n<\/code><\/pre>\n\n<p>i'm working a project, which is already on github.\nif someone wants more details please check <a href=\"https:\/\/github.com\/hcmlab\/vadnet\/blob\/master\/train\/code\/model\/model.py\" rel=\"nofollow noreferrer\">repo<\/a>.<\/p>\n","answers":[]}
+{"QuestionId":57959004,"AnswerCount":0,"Tags":"<pytorch><torchvision>","CreationDate":"2019-09-16T14:29:51.450","AcceptedAnswerId":null,"OwnerUserId":4159376.0,"Title":"Pytorch feeding dataloader batch with custom dataset and collate_fn() to model is not working","Body":"<p>Hi I have created a custom dataset using dataset class that original and target are images, like semantic segmentation and pix2pix,I am loading dataset using Imagefolder and a custom collate function and I am trying to load my custom dataset using dataloader for training the neural network, but an error occurred saying that input should be tensor not list:<\/p>\n\n<p>my collate function<\/p>\n\n<pre><code>def my_collate(batch):\ndata = [item[0] for item in batch]\ntarget = [item[1] for item in batch]\nreturn [data, target]\n<\/code><\/pre>\n\n<p>Dataset class:<\/p>\n\n<pre><code>class bsds_dataset(Dataset):\ndef __init__(self, ds_main, ds_energy):\n    self.dataset1 = ds_main\n    self.dataset2 = ds_energy\n\ndef __getitem__(self, index):\n    x1 = self.dataset1[index]\n    x2 = self.dataset2[index]\n\n    return x1, x2\n\ndef __len__(self):\n    return len(self.dataset1)\n<\/code><\/pre>\n\n<p>Loading dataset:<\/p>\n\n<pre><code>generic_transform = transforms.Compose([\ntransforms.ToTensor(),\ntransforms.ToPILImage(),\n#transforms.CenterCrop(size=128),\n#transforms.Lambda(lambda x: myimresize(x, (128, 128))),\ntransforms.ToTensor(),\n#transforms.Normalize((0., 0., 0.), (6, 6, 6))\n])\noriginal_imagefolder = '.\/images\/whole'\ntarget_imagefolder = '.\/results\/whole'\n\noriginal_ds = ImageFolder(original_imagefolder, \ntransform=generic_transform)\nenergy_ds = ImageFolder(target_imagefolder, transform=generic_transform)\n\ndataset = bsds_dataset(original_ds, energy_ds)\nloader = DataLoader(dataset, batch_size=16, collate_fn=my_collate)\n\nepochs = 2\nmodel = UNet(1, depth=5, merge_mode='concat')\nmodel.cuda()\nloss = torch.nn.MSELoss()\ncriterion_pixelwise = torch.nn.L1Loss()\n\nloss.cuda()\ncriterion_pixelwise.cuda()\n\noptimizer = optim.SGD(model.parameters(), lr=0.001)\n\nTensor = torch.cuda.FloatTensor\n#main loop\nfor epoch in range(epochs):\n   for i, batch in enumerate(loader):\n       original, target = batch\n       out = model(original)\n<\/code><\/pre>\n\n<p>And this error occurred:<\/p>\n\n<pre><code>  TypeError: conv2d(): argument 'input' (position 1) must be Tensor, not \n  list\n<\/code><\/pre>\n\n<p>I am trying to forward each instance of batch into the model. The batch in iteration is list but it should be tensor I don't know how to convert it to tensor or load each instance of batch. Please help. Thanks a lot.<\/p>\n\n<p>Full Traceback:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-147-d1dea1bc00f8&gt; in &lt;module&gt;\n     15     for i, batch in enumerate(loader):\n     16         original, target = batch\n---&gt; 17         out = model(original)\n\nC:\\Anaconda3\\envs\\torchgpu\\lib\\site-packages\\torch\\nn\\modules\\module.py in __call__(self, *input, **kwargs)\n    491             result = self._slow_forward(*input, **kwargs)\n    492         else:\n--&gt; 493             result = self.forward(*input, **kwargs)\n    494         for hook in self._forward_hooks.values():\n    495             hook_result = hook(self, input, result)\n\n&lt;ipython-input-7-5f743c3455c4&gt; in forward(self, x)\n     89         # encoder pathway, save outputs for merging\n     90         for i, module in enumerate(self.down_convs):\n---&gt; 91             x, before_pool = module(x)\n     92             encoder_outs.append(before_pool)\n     93 \n\nC:\\Anaconda3\\envs\\torchgpu\\lib\\site-packages\\torch\\nn\\modules\\module.py in __call__(self, *input, **kwargs)\n    491             result = self._slow_forward(*input, **kwargs)\n    492         else:\n--&gt; 493             result = self.forward(*input, **kwargs)\n    494         for hook in self._forward_hooks.values():\n    495             hook_result = hook(self, input, result)\n\n&lt;ipython-input-5-26a0f7e21ea6&gt; in forward(self, x)\n     14 \n     15     def forward(self, x):\n---&gt; 16         x = F.relu(self.conv1(x))\n     17         x = F.relu(self.conv2(x))\n     18         before_pool = x\n\nC:\\Anaconda3\\envs\\torchgpu\\lib\\site-packages\\torch\\nn\\modules\\module.py in __call__(self, *input, **kwargs)\n    491             result = self._slow_forward(*input, **kwargs)\n    492         else:\n--&gt; 493             result = self.forward(*input, **kwargs)\n    494         for hook in self._forward_hooks.values():\n    495             hook_result = hook(self, input, result)\n\nC:\\Anaconda3\\envs\\torchgpu\\lib\\site-packages\\torch\\nn\\modules\\conv.py in forward(self, input)\n    336                             _pair(0), self.dilation, self.groups)\n    337         return F.conv2d(input, self.weight, self.bias, self.stride,\n--&gt; 338                         self.padding, self.dilation, self.groups)\n    339 \n    340 \n\nTypeError: conv2d(): argument 'input' (position 1) must be Tensor, not list\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57959010,"AnswerCount":0,"Tags":"<python><keras><svm><lstm><pca>","CreationDate":"2019-09-16T14:30:09.257","AcceptedAnswerId":null,"OwnerUserId":10385423.0,"Title":"How to classify features at the output of an LSTM?","Body":"<p>In my DNN architecture in Keras, the output of an LSTM layer (with return sequences=True) is (batch_size, dim_1, dim_2). I want to take these intermediate layer features and classify it using an off-the-shelf classifier like SVM. In sklearn SVM (<a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.svm.SVC.html\" rel=\"nofollow noreferrer\">https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.svm.SVC.html<\/a>), the fit method takes X of shape  (n_samples, n_features). What to do when I have a matrix as n_features? Can I do a PCA like technique before feeding to SVM?<\/p>\n","answers":[]}
+{"QuestionId":57959097,"AnswerCount":0,"Tags":"<python><fft><torch>","CreationDate":"2019-09-16T14:34:50.597","AcceptedAnswerId":null,"OwnerUserId":12074834.0,"Title":"Input and Output of torch.fft?","Body":"<p>I am trying to transform my RGB images from spatial domain to frequency domain using <code>torch.rfft<\/code>, but I really do not understand what should be the signal dimension of my input as I have 3 color channels and <code>torch.rfft<\/code> works when I insert <code>signal_ndim=1<\/code>, \" and 3!<\/p>\n\n<p>I also do not understand what is the output exactly, as it does not contain any complex numbers and why I have the output of <code>(100,100,2,2)<\/code> for the input image size of <code>(100,100,3)<\/code>?<\/p>\n\n<p>What is this <code>2*2<\/code> at the end? Why not <code>(100,100,3,2)<\/code>? What is the redundancy here and what exactly are the output matrices?<\/p>\n","answers":[]}
+{"QuestionId":57959305,"AnswerCount":2,"Tags":"<python><numpy><tensorflow>","CreationDate":"2019-09-16T14:46:38.943","AcceptedAnswerId":57976743.0,"OwnerUserId":4496896.0,"Title":"TensorFlow: How to combine rows of tensor with summing the 2nd element of tensor which has the same 1st element?","Body":"<p>For example, I want to <code>add<\/code> the 2nd element of this tensor where the 1st element is same. Any Numpy based solution is also welcomed! <\/p>\n\n<ul>\n<li>From :<\/li>\n<\/ul>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = tf.constant([\n    [1., 0.9],\n    [2., 0.7],\n    [1., 0.7],\n    [3., 0.4],\n    [4., 0.8]\n], dtype=tf.float32)\n<\/code><\/pre>\n\n<ul>\n<li>To:<\/li>\n<\/ul>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = tf.constant([\n    [1., 1.6],\n    [2., 0.7],\n    [3., 0.4],\n    [4., 0.8]\n], dtype=tf.float32)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57976743","CreationDate":"2019-09-17T14:42:38.943","ParentId":null,"OwnerUserId":"4496896","Title":null,"Body":"<p>Thanks <a href=\"https:\/\/stackoverflow.com\/users\/10111013\/finleygibson\">FinleyGibson<\/a> for the Numpy solution and some useful TensorFlow pointers! This is my solution in TF with using <code>tf.unique_with_counts()<\/code> and <code>tf.segment_sum()<\/code>:<\/p>\n\n<p><strike><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = tf.constant([\n    [1., 0.9],\n    [2., 0.7],\n    [1., 0.7],\n    [3., 0.4],\n    [4., 0.8]\n], dtype=tf.float32)\n\nwith tf.Session() as sess:\n    y, idx, y_counts = tf.unique_with_counts(x[:, 0])\n    idx_sorted = tf.sort(idx, axis=-1, direction='ASCENDING')\n    score_sum = tf.segment_sum(x[:, 1], idx_sorted)\n    result = tf.stack((y, score_sum), axis=1)\n    print(sess.run(result))\n\n[[1.       1.5999999]\n[2.        0.7      ]\n[3.        0.4      ]\n[4.        0.8      ]]\n<\/code><\/pre>\n\n<p><\/strike>\n<strong>EDIT:<\/strong><\/p>\n\n<ul>\n<li>Seems like the above solution was not properly ordering rows\/columns pairs. Here is the fixed version.<\/li>\n<\/ul>\n\n<pre class=\"lang-py prettyprint-override\"><code>with tf.Session() as sess:\n    x = tf.constant([\n        [2., 0.7],\n        [1., 0.1],\n        [3., 0.4],\n        [4., 0.8],\n        [1., 0.9]], dtype=tf.float32)\n\n    def matrix_sort(a, col):\n        return tf.gather(a, tf.nn.top_k(-a[:, col], k=a.get_shape()[0].value).indices)\n\n    sorted_x = matrix_sort(matrix_sort(x, 1), 0)\n    labels = sorted_x[:, 0]\n    scores = sorted_x[:, 1]\n    y, idx, y_counts = tf.unique_with_counts(labels)\n    score_sum = tf.segment_sum(scores, idx)    \n    result = tf.stack((y, score_sum), axis=1) \n    print(sess.run(result))\n\n[[1.  1. ]\n [2.  0.7]\n [3.  0.4]\n [4.  0.8]]\n\n<\/code><\/pre>\n"},{"AnswerId":"57959532","CreationDate":"2019-09-16T14:58:50.390","ParentId":null,"OwnerUserId":"10111013","Title":null,"Body":"<p>numpy solution:<\/p>\n\n<pre><code>x = np.array([\n    [1., 0.9],\n    [2., 0.7],\n    [1., 0.7],\n    [3., 0.4],\n    [4., 0.8]])\n\nans = np.array([[i,np.sum(x[np.where(x[:,0]==i), 1])] for i in set(x[:,0])])\n<\/code><\/pre>\n\n<p>gives <\/p>\n\n<pre><code>array([[1. , 1.6],\n       [2. , 0.7],\n       [3. , 0.4],\n       [4. , 0.8]])\n<\/code><\/pre>\n\n<p>you will not be able to do this for a 'tf.constant()', as it is a <em>constant<\/em> variable and does not support having its values changed. If you want to change values within tensorflow data structures it is best to either pass values to a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/placeholder\" rel=\"nofollow noreferrer\">tf.placeholder<\/a> or use a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Variable\" rel=\"nofollow noreferrer\">tf.Variable<\/a>. However these require predefined dimensions, and cannot have their sizes changed as desired in your question.<\/p>\n"}]}
+{"QuestionId":57959445,"AnswerCount":0,"Tags":"<pytorch><tensor><information-theory>","CreationDate":"2019-09-16T14:54:33.293","AcceptedAnswerId":null,"OwnerUserId":1767774.0,"Title":"Memory efficient batch-pairwise KL divergence in pytorch","Body":"<p>Given a Batch tensor <code>S<\/code> of dimensions <code>BxD<\/code> where each row is a probability distribution over <code>D<\/code> dimensions, I want to calculate the batch-pairwise KL divergence matrix <code>KL<\/code> (of dimensions <code>BxB<\/code>) such that <code>KL[i,j] = KL-DIVERGENCE(S[i],S[j])<\/code>.<\/p>\n\n<p>This works:<\/p>\n\n<pre><code>softmax = torch.nn.Softmax(dim = 1)\nB,D = 100, 1000\nS = softmax(torch.rand(B,D))\nKL = torch.mean(S[:,None,:] * (S[:,None,:]\/S).log(),dim = 2)\n<\/code><\/pre>\n\n<p>but is very memory inefficient. Is there a way to calculate this matrix more efficiently?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[]}
+{"QuestionId":57959462,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-datasets>","CreationDate":"2019-09-16T14:55:21.663","AcceptedAnswerId":57973310.0,"OwnerUserId":1268426.0,"Title":"Build a dataset for TensorFlow","Body":"<p>I have a large number of JPGs representing vehicles. I want to create a dataset for TensorFlow with a categorization such that every vehicle image describes the side, the angle or the roof, i.e. I want to create nine subsets of images (front, back, driver side, driver front angle, driver back angle, passenger side, passenger front angle, passenge back angle, roof). At the moment the filename of each JPG describes the desired point.<\/p>\n\n<p>How can I turn this set to be a dataset that TensorFlow can easily manipulate? Also, should I run a procedure which crop the JPG to extract only the vehicle portion? How could I do that using TensorFlow?<\/p>\n\n<p>I apologize in advance for not providing details and examples to this question, but I don't really know how can I achieve an entry point for this problem. The tutorials I'm following all assume an already created dataset ready to use.<\/p>\n","answers":[{"AnswerId":"57973310","CreationDate":"2019-09-17T11:29:21.133","ParentId":null,"OwnerUserId":"10111013","Title":null,"Body":"<p>Okay, I'm going to try to answer this as well as I can, but producing and pre-processing data for use in ML algorithms is laborious and often expensive (hence the repeated use of well known data sets for testing algorithm designs).<\/p>\n\n<p><strong>To address a few straight-forward questions first:<\/strong><\/p>\n\n<p><code>should I run a procedure which crop the JPG to extract only the vehicle portion?<\/code><\/p>\n\n<p>No. This isn't necessary. The neural network will sort the relevant information in the images from the irrelevant itself and having a diverse set of images will help to build a robust classifier. Also you would likely make life a lot more difficult for yourself later on by resizing images (see point <strong>1.<\/strong> below for more).<\/p>\n\n<p><code>How could I do that using TensorFlow?<\/code> <\/p>\n\n<p>You wouldn't. Tensorflow is designed to build and test ML models, and does not have tool for pre-processing data. (well perhaps TensorFlow Extended does, but this shouldn't be necessary)<\/p>\n\n<p><strong>Now a rough guideline for how you would go about creating a data set from the files described:<\/strong><\/p>\n\n<p><strong>1)<\/strong> The first thing you will <em>need<\/em> to do is to load your .jpg images into python and resize them all to be identical. A neural network will need the same number of inputs (pixels in this case) in every training example, so having different sized images will not work.<\/p>\n\n<ul>\n<li>There is a good answer detailing how to load images using python image library (PIL) on stack overflow <a href=\"https:\/\/stackoverflow.com\/a\/36777043\/10111013\">here<\/a>. <\/li>\n<li>The PIL image instances (elements of the list <code>loadedImages<\/code> in the example above) can then be converted to numpy arrays using <code>data = np.asarray(image)<\/code>, which tensorflow can work with. <\/li>\n<li><p>In addition to building a set of numpy arrays of your data, you will also need a second numpy array of labels for this data. A typical way to encode this will be as a numpy array the same length as your number of images with an integer value for each point representing the class to which that image belongs (0-8 for your 9 classes). You could input these by hand, but this will be labour intensive, and I would suggest using python strings inbuilt <a href=\"https:\/\/www.programiz.com\/python-programming\/methods\/string\/find\" rel=\"nofollow noreferrer\">find<\/a> method to locate key words within the filenames to automate determining their class. This could be done within the <\/p>\n\n<pre><code>for image in imagesList:\n<\/code><\/pre>\n\n<p>loop in the above link, as <code>image<\/code> should be a string containing the image filename.<\/p>\n\n<ul>\n<li>As I mentioned above, resizing the images is necessary to make sure they are all identical. You could do this with numpy, using <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.indexing.html\" rel=\"nofollow noreferrer\">indexing<\/a> to choose a subsection of each image array, or using PIL's <a href=\"https:\/\/www.geeksforgeeks.org\/python-pil-image-resize-method\/\" rel=\"nofollow noreferrer\">resize<\/a> function before converting to numpy. There is no right answer here, and many methods have been used to resize images for this purpose, from padding, to stretching to cropping. <\/li>\n<\/ul><\/li>\n<\/ul>\n\n<p>Then end result here should be 2 numpy arrays. One of image data which has shape <code>[w,h,3,n]<\/code> where <code>w<\/code>=image width, <code>h<\/code>=image height, 3 = the three RGB layers (provided images are in colour) and <code>n<\/code>= the number of images you have. The second of labels associated with these images, of shape <code>[n,]<\/code> where every element of the length <code>n<\/code> array is an integer from 0-8 specifying its class. <\/p>\n\n<p>At this point it would be a good idea to save the dataset in this format using <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.save.html\" rel=\"nofollow noreferrer\">numpy.save()<\/a> so that you don't have to go through this process again. <\/p>\n\n<p><strong>2)<\/strong> Once you have your images in this format, tensorflow has a class called <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset\" rel=\"nofollow noreferrer\">tf.Dataset<\/a> into which you can load the image and label data described above and will allow you to shuffle and sample data from it. <\/p>\n\n<p>I hope that was helpful, and I am sorry that there is no quick-fix solution to this (at least not one I am aware of). Good luck.<\/p>\n"}]}
+{"QuestionId":57959755,"AnswerCount":0,"Tags":"<python><tensorflow><out-of-memory>","CreationDate":"2019-09-16T15:10:54.710","AcceptedAnswerId":null,"OwnerUserId":11208496.0,"Title":"How to solve ResourceExhaustedError? Why is there such a big tensor?","Body":"<p>My loss function would be tf.losses.mean_squared_error, but with this, ResourceExhaustedError occurs.\nBecause of this, I tried to implement the loss function on my own:<\/p>\n\n<pre><code>diff = tf.subtract(ground_truth_ph, self.pred, name='sub') \nloss_op = tf.math.square(diff, name='squared')\n<\/code><\/pre>\n\n<p>I still get the error:<\/p>\n\n<pre><code>ResourceExhaustedError: OOM when allocating tensor with shape[1138314,1138314] and type float on \/job:localhost\/replica:0\/task:0\/device:CPU:0 by allocator cpu\n     [[{{node sub}}]]\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.\n<\/code><\/pre>\n\n<p>My self.pred have a shape of (1138314, 1), and ground_truth_ph also (initialized with (None, 1))<\/p>\n\n<p>I don't understand why is it needed to allocate this amount of memory for an elementwise subtraction. <\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[]}
+{"QuestionId":57959915,"AnswerCount":1,"Tags":"<python><tensorflow><autoencoder>","CreationDate":"2019-09-16T15:19:10.630","AcceptedAnswerId":null,"OwnerUserId":3029270.0,"Title":"combining two graph i tensorflow and optimze the middle model only","Body":"<p>I have auto-encoder (E1 - D1) trained already \nNow I want another auto-encoder lets call it (E2 - D2)<\/p>\n\n<p>so the model its like that E1 -> E2 -> D2 -> D1 where the model optimize E2 D2 only by using output of D1 and input to E1<\/p>\n\n<p>E1 -> 5 CNN layers + 1 FC (already trained)<\/p>\n\n<p>E2 -> 5 CNN layers + 1 FC <\/p>\n\n<p>D2 -> 5 DeCNN layers + 1 FC<\/p>\n\n<p>D1 -> 5 DeCNN layers + 1 FC (already trained)<\/p>\n\n<p>Note E1 D1 already trained (blue in the image)<\/p>\n\n<p>please see the image<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/nASVq.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/nASVq.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I want to use X (input) Y (output) to optimize only E2 D2<\/p>\n","answers":[{"AnswerId":"57960900","CreationDate":"2019-09-16T16:21:35.593","ParentId":null,"OwnerUserId":"10111013","Title":null,"Body":"<p>The easiest way to do this in tensorflow would be to first train E1 and D1 (already done). And then extract and store the values of the <code>tf.Variables<\/code> (weights and biases) for these networks, then create an entirely new network with structure that encompasses all of the auto-encoders described above, and inject the extracted weight values back into the new network.<\/p>\n\n<p>You can do this by traning the E1 and D1 networks and then calling <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Graph#get_tensor_by_name\" rel=\"nofollow noreferrer\">tf.get_tensor_by_name<\/a> within the session to extract the weight values as numpy arrays. These can then be saved for later using <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.save.html\" rel=\"nofollow noreferrer\">np.save<\/a>.<\/p>\n\n<p>Create your new full network with all the desired layers:<\/p>\n\n<p>5 CNN layers + 1 FC + 5 CNN layers + 1 FC + 5 DeCNN layers + 1 FC 5 DeCNN layers + 1 FC <\/p>\n\n<p>And then you can initialise the values of the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Variable\" rel=\"nofollow noreferrer\">tf.Variables<\/a> in the E1 and D1 layers to the values saved earlier, by loading the previously saved values using <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.load.html\" rel=\"nofollow noreferrer\">np.load<\/a> and then passing these to your tf.Variables as the <code>&lt;initial-value&gt;<\/code> argument described <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Variable\" rel=\"nofollow noreferrer\">here<\/a>. <\/p>\n\n<p>You can set these restored tf.Variables to untrainable so that they remain fixed during the rest of the training.<\/p>\n"}]}
+{"QuestionId":57960407,"AnswerCount":0,"Tags":"<tensorflow><distributed-computing>","CreationDate":"2019-09-16T15:49:41.987","AcceptedAnswerId":null,"OwnerUserId":6450489.0,"Title":"Feeding tensorflow estimator with dataset in distributed mode","Body":"<p>I have below piece of code for running a keras LSTM classifier network on a dataset containing sequence of features and it runs well without problem (details omitted to keep the text readable). My Tensorflow version is 1.13:<\/p>\n\n<pre><code>def parse_tfrecord(example):\n    features = tf.parse_single_example(example, featuresDict)\n    label = features['label']\n    data = tf.decode_raw(features['data'], tf.int64)\n    return data, label\n\ndef read_datasets(pattern, numFiles, numEpochs=None, batchSize=None):\n    files = tf.data.Dataset.list_files(pattern)\n    def _parse(x):\n        x = tf.data.TFRecordDataset(x, compression_type='GZIP')\n        return x\n    dataset = files.interleave(_parse, cycle_length=numFiles, block_length=1).map(parse_tfrecord)\n    dataset = dataset.batch(batchSize)\n    dataset = dataset.repeat(numEpochs)\n    return dataset\n\ndef keras_model(...)\n\ndef main(args):\n    ...   \n    train_data = read_datasets(...)\n    val_data = read_datasets(...)\n    test_data = read_datasets(...)\n\n    model = keras_model(...)\n    model.compile(...)\n    model.fit(train_data, epochs=epochs, steps_per_epoch = train_steps, validation_data=val_data, validation_steps = val_steps)\n\nif __name__ == '__main__':\n    ...\n    main(args)\n<\/code><\/pre>\n\n<p>I want to run it in distributed mode, so I encapsulate the keras classifier with an estimator object. Distributed learning documentation is unclear on the net, but what I understand from it (and expect it to be) is that I should just modify the main() function like this:<\/p>\n\n<pre><code>def main(args):\n    ...\n    train_data = read_datasets(...)\n    val_data = read_datasets(...)\n    test_data = read_datasets(...)\n\n    model = keras_model(...)\n    model.compile(...)\n    #model.fit(train_data, epochs=epochs, steps_per_epoch = train_steps, validation_data=val_data, validation_steps = val_steps)\n    runConfig = tf.estimator.RunConfig(\n            session_config=config,\n            model_dir=log_dir,\n            save_summary_steps=1,\n            save_checkpoints_steps=train_steps\n            )\n    estimator = tf.keras.estimator.model_to_estimator(model, model_dir=log_dir, config=runConfig)\n    train_input_fn = train_data.make_one_shot_iterator().get_next\n    eval_input_fn = val_data.make_one_shot_iterator().get_next\n    train_spec = tf.estimator.TrainSpec(input_fn=train_input_fn, max_steps=epochs*train_steps)\n    eval_spec = tf.estimator.EvalSpec(input_fn=eval_input_fn, start_delay_secs=1, throttle_secs=1, steps=None)\n    tf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)\n\nif __name__ == '__main__':\n    ...\n    TF_CONFIG = {\n        'task': ...\n        'cluster': ...\n    }\n    os.environ['TF_CONFIG'] = json.dumps(TF_CONFIG)\n    main(args)\n<\/code><\/pre>\n\n<p>However I get below error traceback:<\/p>\n\n<pre><code>INFO:tensorflow:Calling model_fn.\nTraceback (most recent call last):\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\ops\\array_ops.py\", line 1811, in zeros\n    tensor_shape.TensorShape(shape))\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\tensor_shape.py\", line 690, in __init__\n    self._dims = [as_dimension(d) for d in dims_iter]\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\tensor_shape.py\", line 690, in &lt;listcomp&gt;\n    self._dims = [as_dimension(d) for d in dims_iter]\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\tensor_shape.py\", line 632, in as_dimension\n    return Dimension(value)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\tensor_shape.py\", line 185, in __init__\n    self._value = int(value)\nTypeError: int() argument must be a string, a bytes-like object or a number, not 'Tensor'\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"LSTM_TFdataset-distributed.py\", line 323, in &lt;module&gt;\n    main(a)\n  File \"LSTM_TFdataset-distributed.py\", line 233, in main\n    tf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow_estimator\\python\\estimator\\training.py\", line 471, in train_and_evaluate\n    return executor.run()\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow_estimator\\python\\estimator\\training.py\", line 638, in run\n    getattr(self, task_to_run)()\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow_estimator\\python\\estimator\\training.py\", line 643, in run_chief\n    return self._start_distributed_training()\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow_estimator\\python\\estimator\\training.py\", line 789, in _start_distributed_training\n    saving_listeners=saving_listeners)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow_estimator\\python\\estimator\\estimator.py\", line 358, in train\n    loss = self._train_model(input_fn, hooks, saving_listeners)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow_estimator\\python\\estimator\\estimator.py\", line 1124, in _train_model\n    return self._train_model_default(input_fn, hooks, saving_listeners)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow_estimator\\python\\estimator\\estimator.py\", line 1154, in _train_model_default\n    features, labels, model_fn_lib.ModeKeys.TRAIN, self.config)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow_estimator\\python\\estimator\\estimator.py\", line 1112, in _call_model_fn\n    model_fn_results = self._model_fn(features=features, **kwargs)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow_estimator\\python\\estimator\\keras.py\", line 278, in model_fn\n    labels)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow_estimator\\python\\estimator\\keras.py\", line 201, in _clone_and_build_model\n    optimizer_iterations=global_step)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\models.py\", line 466, in clone_and_build_model\n    clone = clone_model(model, input_tensors=input_tensors)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\models.py\", line 271, in clone_model\n    return _clone_functional_model(model, input_tensors=input_tensors)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\models.py\", line 161, in _clone_functional_model\n    **kwargs))\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\layers\\recurrent.py\", line 701, in __call__\n    return super(RNN, self).__call__(inputs, **kwargs)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\base_layer.py\", line 554, in __call__\n    outputs = self.call(inputs, *args, **kwargs)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\layers\\recurrent.py\", line 2416, in call\n    inputs, mask=mask, training=training, initial_state=initial_state)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\layers\\recurrent.py\", line 759, in call\n    inputs, initial_state, constants)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\layers\\recurrent.py\", line 863, in _process_inputs\n    initial_state = self.get_initial_state(inputs)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\layers\\recurrent.py\", line 679, in get_initial_state\n    inputs=None, batch_size=batch_size, dtype=dtype)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\layers\\recurrent.py\", line 2194, in get_initial_state\n    self, inputs, batch_size, dtype))\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\layers\\recurrent.py\", line 3001, in _generate_zero_filled_state_for_cell\n    return _generate_zero_filled_state(batch_size, cell.state_size, dtype)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\layers\\recurrent.py\", line 3017, in _generate_zero_filled_state\n    return nest.map_structure(create_zeros, state_size)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\util\\nest.py\", line 381, in map_structure\n    structure[0], [func(*x) for x in entries])\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\util\\nest.py\", line 381, in &lt;listcomp&gt;\n    structure[0], [func(*x) for x in entries])\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\keras\\layers\\recurrent.py\", line 3014, in create_zeros\n    return array_ops.zeros(init_state_size, dtype=dtype)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\ops\\array_ops.py\", line 1814, in zeros\n    shape = ops.convert_to_tensor(shape, dtype=dtypes.int32)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 1039, in convert_to_tensor\n    return convert_to_tensor_v2(value, dtype, preferred_dtype, name)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 1097, in convert_to_tensor_v2\n    as_ref=False)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 1175, in internal_convert_to_tensor\n    ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\ops\\array_ops.py\", line 1102, in _autopacking_conversion_function\n    return _autopacking_helper(v, dtype, name or \"packed\")\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\ops\\array_ops.py\", line 1054, in _autopacking_helper\n    return gen_array_ops.pack(elems_as_tensors, name=scope)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\ops\\gen_array_ops.py\", line 6426, in pack\n    \"Pack\", values=values, axis=axis, name=name)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\op_def_library.py\", line 350, in _apply_op_helper\n    g = ops._get_graph_from_inputs(_Flatten(keywords.values()))\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 5713, in _get_graph_from_inputs\n    _assert_same_graph(original_graph_element, graph_element)\n  File \"D:\\Shahriar\\Python36\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 5649, in _assert_same_graph\n    original_item))\nValueError: Tensor(\"lstm\/zeros\/packed\/1:0\", shape=(), dtype=int32, device=\/job:chief\/task:0) must be from the same graph as Tensor(\"strided_slice:0\", shape=(), dtype=int32).\n<\/code><\/pre>\n\n<p>Can anyone help me to solve this problem?<\/p>\n","answers":[]}
+{"QuestionId":57960524,"AnswerCount":0,"Tags":"<tensorflow><google-cloud-platform>","CreationDate":"2019-09-16T15:56:14.823","AcceptedAnswerId":null,"OwnerUserId":6244166.0,"Title":"The replica master 0 exited with a non-zero status of 1(GCP Tensorflow error)","Body":"<p>I am using the recommendation engine tutorial available <a href=\"https:\/\/cloud.google.com\/solutions\/machine-learning\/recommendation-system-tensorflow-overview\" rel=\"nofollow noreferrer\">here<\/a>. I am exactly following the code but I am encountering an error(error log details given below) while running <\/p>\n\n<pre><code>.\/mltrain.sh train ${BUCKET} data\/ratings.dat --delimiter ::\n<\/code><\/pre>\n\n<p>The error log details are<\/p>\n\n<pre><code>The replica master 0 exited with a non-zero status of 1. Traceback (most recent call last): File \"\/usr\/lib\/python2.7\/runpy.py\", line 162, in _run_module_as_main \"__main__\", fname, loader, pkg_name) File \"\/usr\/lib\/python2.7\/runpy.py\", line 72, in _run_code exec code in run_globals File \"\/root\/.local\/lib\/python2.7\/site-packages\/trainer\/task.py\", line 206, in &lt;module&gt; main(job_args) File \"\/root\/.local\/lib\/python2.7\/site-packages\/trainer\/task.py\", line 31, in main args, input_file, args['data_type']) File \"\/root\/.local\/lib\/python2.7\/site-packages\/trainer\/model.py\", line 80, in create_test_and_train_sets input_file) File \"\/root\/.local\/lib\/python2.7\/site-packages\/trainer\/model.py\", line 116, in _ratings_train_and_test 'timestamp': np.int32, File \"\/usr\/local\/lib\/python2.7\/dist-packages\/pandas\/io\/parsers.py\", line 498, in parser_f return _read(filepath_or_buffer, kwds) File \"\/usr\/local\/lib\/python2.7\/dist-packages\/pandas\/io\/parsers.py\", line 275, in _read parser = TextFileReader(filepath_or_buffer, **kwds) File \"\/usr\/local\/lib\/python2.7\/dist-packages\/pandas\/io\/parsers.py\", line 586, in __init__ self.options, self.engine = self._clean_options(options, engine) File \"\/usr\/local\/lib\/python2.7\/dist-packages\/pandas\/io\/parsers.py\", line 665, in _clean_options raise ValueError(msg) ValueError: Falling back to the 'python' engine because the 'c' engine does not support regex separators, but this causes 'dtype' to be ignored as it is not supported by the 'python' engine. (Note the 'converters' option provides similar functionality.) To find out more about why your job exited please check the logs: https:\/\/console.cloud.google.com\/logs\/viewer?project=570527225445&amp;resource=ml_job%2Fjob_id%2Fwals_ml_train_20190916_150714&amp;advancedFilter=resource.type%3D%\n<\/code><\/pre>\n\n<p>I am following the below code<\/p>\n\n<pre><code>sudo apt-get update\nsudo apt-get install -y git bzip2 unzip\ngit clone https:\/\/github.com\/GoogleCloudPlatform\/tensorflow-recommendation-wals\nwget https:\/\/repo.continuum.io\/miniconda\/Miniconda2-latest-Linux-x86_64.sh\nbash Miniconda2-latest-Linux-x86_64.sh\ncd tensorflow-recommendation-wals\nconda create -n tfrec\nconda install -n tfrec --file conda.txt\nsource activate tfrec\npip install -r requirements.txt\npip install tensorflow\ncurl -O 'http:\/\/files.grouplens.org\/datasets\/movielens\/ml-1m.zip'\nunzip ml-1m.zip\nmkdir -p data\ncp ml-1m\/ratings.dat data\/\n<\/code><\/pre>\n\n<p>my bucket name is re_bucket<\/p>\n\n<pre><code>BUCKET=gs:\/\/re_bucket\ngsutil cp -r data\/ratings.dat $BUCKET\/data\/ratings.dat\ncd wals_ml_engine\n.\/mltrain.sh train ${BUCKET} data\/ratings.dat --delimiter ::\n<\/code><\/pre>\n\n<p>I have tried several solutions posted <a href=\"https:\/\/stackoverflow.com\/questions\/49434874\/tensorflow-on-ml-engine-the-replica-master-0-exited-with-a-non-zero-status-of-1\">here<\/a> but still getting the same error. I have also created several complete new instances many times to see whether that can resolve the error or not but there is no success in that. Please let me know what are the possible steps that I can follow to resolve issue.<\/p>\n\n<p>Note: When I am trying the same thing in my local storage using<\/p>\n\n<pre><code>.\/mltrain.sh local ..\/data re_data.dat --delimiter ::\n<\/code><\/pre>\n\n<p>things are working fine for me. It might be the case that error is related with GCP bucket.<\/p>\n\n<p>Also attaching the screenshot for stack driver information<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/rw07I.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rw07I.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57960757,"AnswerCount":2,"Tags":"<keras><neural-network><deep-learning><conv-neural-network><training-data>","CreationDate":"2019-09-16T16:12:12.217","AcceptedAnswerId":null,"OwnerUserId":3802483.0,"Title":"Why does my model predict the same label?","Body":"<p>I am training a small network and the training seems to go fine, the val loss decreases, I reach validation accuracy around 80, and it actually stops training once there is no more improvement (patience=10). It trained for 40 epochs. However, it keeps predicting only one class for every test image! I tried to initialize the conv layers randomly, I added regularizers, I switched from Adam to SGD, I added clipvalue, I added dropouts. I also switched to softmax (I have only two labels but I saw some recommendation on using softmax and Dense layer with 2 neurons). Some or one of these helped with the overfitting, but nothing worked for the prediction problem. <strong>The data is balanced<\/strong>, though it is a small dataset, so it doesn't make sense that it reaches 80% if it predicts the same labels for evaluation set as well.<\/p>\n\n<p>What is wrong with my model and how can I fix it? Any comments are welcome.<\/p>\n\n<pre><code>#Import some packages to use\nimport cv2\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nfrom keras.preprocessing.image import ImageDataGenerator\nimport os\nfrom keras.regularizers import l2\nfrom keras.callbacks import EarlyStopping, ModelCheckpoint, ReduceLROnPlateau\nfrom keras.layers.core import Dense, Dropout, Flatten\nfrom keras.layers.convolutional import Conv2D, MaxPooling2D\nfrom keras.initializers import RandomNormal\n\nos.environ[\"CUDA_VISIBLE_DEVICES\"]=\"0\"\n\nepochs = 200\ncallbacks = []\n#schedule = None\ndecay = 0.0\n\nearlyStopping = EarlyStopping(monitor='val_loss', patience=10, verbose=0, mode='min')\nmcp_save = ModelCheckpoint('.mdl_wts.hdf5', save_best_only=True, monitor='val_loss', mode='min')\nreduce_lr_loss = ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=3, verbose=1, epsilon=1e-5, mode='min')\n\ntrain_dir = '\/home\/d\/Desktop\/s\/data\/train'\neval_dir = '\/home\/d\/Desktop\/s\/data\/eval'\ntest_dir = '\/home\/d\/Desktop\/s\/data\/test'\n\n\n\n# create a data generator\ntrain_datagen = ImageDataGenerator(rescale=1.\/255,   #Scale the image between 0 and 1\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\nval_datagen = ImageDataGenerator(rescale=1.\/255)  #We do not augment validation data. we only perform rescale\n\ntest_datagen = ImageDataGenerator(rescale=1.\/255)  #We do not augment validation data. we only perform rescale\n\n# load and iterate training dataset\ntrain_generator = train_datagen.flow_from_directory(train_dir,  target_size=(224,224),class_mode='categorical', batch_size=16, shuffle='True', seed=42)\n# load and iterate validation dataset\nval_generator = val_datagen.flow_from_directory(eval_dir,  target_size=(224,224),class_mode='categorical', batch_size=16, shuffle='True', seed=42)\n# load and iterate test dataset\ntest_generator = test_datagen.flow_from_directory(test_dir,  target_size=(224,224), class_mode=None, batch_size=1, shuffle='False', seed=42)\n#We will use a batch size of 32. Note: batch size should be a factor of 2.***4,8,16,32,64...***\n#batch_size = 4\n\n\n\n#from keras import layers\nfrom keras import models\nfrom keras import optimizers\n#from keras.layers import Dropout\n#from keras.preprocessing.image import ImageDataGenerator\nfrom keras.preprocessing.image import img_to_array, load_img\n\nmodel = models.Sequential()\nmodel.add(Conv2D(64, (3, 3), activation='relu', name='block1_conv1', kernel_initializer=RandomNormal(\n        mean=0.0, stddev=0.05), bias_initializer=RandomNormal(mean=0.0, stddev=0.05), input_shape=(224, 224, 3)))\nmodel.add(Conv2D(64, (3, 3), activation='relu', name='block1_conv2', kernel_initializer=RandomNormal(\n        mean=0.0, stddev=0.05), bias_initializer=RandomNormal(mean=0.0, stddev=0.05)))\nmodel.add(MaxPooling2D((2, 2)))\nmodel.add(Dropout(0.2))\nmodel.add(Conv2D(128, (3, 3), activation='relu', name='block2_conv1', kernel_initializer=RandomNormal(\n        mean=0.0, stddev=0.05), bias_initializer=RandomNormal(mean=0.0, stddev=0.05)))\nmodel.add(Conv2D(128, (3, 3), activation='relu', name='block2_conv2',kernel_initializer=RandomNormal(\n        mean=0.0, stddev=0.05), bias_initializer=RandomNormal(mean=0.0, stddev=0.05)))\nmodel.add(MaxPooling2D((2, 2), name='block2_pool'))\nmodel.add(Dropout(0.2))\nmodel.add(Conv2D(256, (3, 3), activation='relu', name='block3_conv1', kernel_initializer=RandomNormal(\n        mean=0.0, stddev=0.05), bias_initializer=RandomNormal(mean=0.0, stddev=0.05)))\nmodel.add(Conv2D(256, (3, 3), activation='relu', name='block3_conv2', kernel_initializer=RandomNormal(\n        mean=0.0, stddev=0.05), bias_initializer=RandomNormal(mean=0.0, stddev=0.05)))\nmodel.add(Conv2D(256, (3, 3), activation='relu', name='block3_conv3', kernel_initializer=RandomNormal(\n        mean=0.0, stddev=0.05), bias_initializer=RandomNormal(mean=0.0, stddev=0.05)))\nmodel.add(MaxPooling2D((2, 2), name='block3_pool'))\nmodel.add(Dropout(0.2))\n#model.add(layers.Conv2D(512, (3, 3), activation='relu', name='block4_conv1'))\n#model.add(layers.Conv2D(512, (3, 3), activation='relu', name='block4_conv2'))\n#model.add(layers.Conv2D(512, (3, 3), activation='relu', name='block4_conv3'))\n#model.add(layers.MaxPooling2D((2, 2), name='block4_pool'))\nmodel.add(Flatten())\nmodel.add(Dense(256, kernel_regularizer=l2(0.01), bias_regularizer=l2(0.01), activation='relu', kernel_initializer='he_uniform'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(2, kernel_regularizer=l2(0.01), bias_regularizer=l2(0.01), activation='softmax'))\n\n#Lets see our model\nmodel.summary()\n\n#We'll use the RMSprop optimizer with a learning rate of 0.0001\n#We'll use binary_crossentropy loss because its a binary classification\n#model.compile(loss='binary_crossentropy', optimizer=optimizers.SGD(lr=1e-5, momentum=0.9), metrics=['acc'])\nmodel.compile(loss='categorical_crossentropy',\n                   #optimizer=optimizers.Adadelta(lr=1.0, rho=0.95, epsilon=1e-08, decay=decay),\n                    optimizer=optimizers.SGD(lr= 0.0001, clipvalue = 0.5, decay=1e-6, momentum=0.9, nesterov=True),\n              metrics=['accuracy'])\n\n\n\n#The training part\n#We train for 64 epochs with about 100 steps per epoch\nhistory = model.fit_generator(train_generator,\n                              steps_per_epoch=train_generator.n \/\/ train_generator.batch_size,\n                              epochs=epochs,\n                              validation_data=val_generator,\n                              validation_steps=val_generator.n \/\/ val_generator.batch_size,\n                              callbacks=[earlyStopping, mcp_save]) #, reduce_lr_loss])\n\n\n#Save the model\nmodel.save_weights('\/home\/d\/Desktop\/s\/categorical_weights.h5')\nmodel.save('\/home\/d\/Desktop\/s\/categorical_model_keras.h5')\n\n#lets plot the train and val curve\n#get the details form the history object\nacc = history.history['acc']\nval_acc = history.history['val_acc']\nloss = history.history['loss']\nval_loss = history.history['val_loss']\n\nepochs = range(1, len(acc) + 1)\n\n#Train and validation accuracy\nplt.plot(epochs, acc, 'b', label='Training accuracy')\nplt.plot(epochs, val_acc, 'r', label='Validation accuracy')\nplt.title('Training and Validation accuracy')\nplt.legend()\n\nplt.figure()\n#Train and validation loss\nplt.plot(epochs, loss, 'b', label='Training loss')\nplt.plot(epochs, val_loss, 'r', label='Validation loss')\nplt.title('Training and Validation loss')\nplt.legend()\n\nplt.show()\n\nmodel.evaluate_generator(generator=val_generator, steps=val_generator.n \/\/ val_generator.batch_size)\n\nSTEP_SIZE_TEST=test_generator.n\/\/test_generator.batch_size\ntest_generator.reset()\npred=model.predict_generator(test_generator,\nsteps=STEP_SIZE_TEST,\nverbose=1)\n\npredicted_class_indices=np.argmax(pred,axis=1)\n\nlabels = (train_generator.class_indices)\nnp.save('\/home\/d\/Desktop\/s\/classes', labels)\n\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(\"categorical_results.csv\",index=False)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57962111","CreationDate":"2019-09-16T17:57:05.927","ParentId":null,"OwnerUserId":"1729802","Title":null,"Body":"<p>One of the problems that could lead to such behavior is imbalanced dataset. Your model found out that if it predicts the dominant class each time, it would get a good results.<\/p>\n\n<p>There are many ways to tackle an imbalance dataset. Here is a good <a href=\"https:\/\/towardsdatascience.com\/methods-for-dealing-with-imbalanced-data-5b761be45a18\" rel=\"nofollow noreferrer\">tutorial<\/a>.\nOne of the easiest yet powerful solution is to apply higher penalty to your loss if it wrongly predicted the smaller class. This can be implemented in keras by setting the parameter <code>class_weight<\/code> in the <code>fit<\/code>or <code>fit_generator<\/code> function.<\/p>\n\n<p>It can be a dictionary of example:<\/p>\n\n<pre><code>class_weight = {0: 0.75, 1: 0.25}  # does not necessarily add to up 1.\nhistory = model.fit_generator(train_generator,\n                              steps_per_epoch=train_generator.n \/\/ train_generator.batch_size,\n                              epochs=epochs,\n                              class_weight= class_weight,  # this is the important part\n                              validation_data=val_generator,\n                              validation_steps=val_generator.n \/\/ val_generator.batch_size,\n                              callbacks=[earlyStopping, mcp_save]) #, reduce_lr_loss])\n<\/code><\/pre>\n"},{"AnswerId":"57962575","CreationDate":"2019-09-16T18:31:23.740","ParentId":null,"OwnerUserId":"10484131","Title":null,"Body":"<p>Adding to Coderji's answer, it might also prove advantageous to counter class imbalance using <a href=\"https:\/\/stats.stackexchange.com\/questions\/49540\/understanding-stratified-cross-validation\">stratified <code>k<\/code>-fold cross-validation<\/a>, with <code>k = 5<\/code> being common practice. This basically splits your data set up into <code>k<\/code> splits like regular cross-validation, but also stratifies these splits. In the case of class imbalance, each of these splits contain over-\/undersampled classes compensating for their lower\/higher occurence within the data set.<\/p>\n\n<p>As of yet <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/796\" rel=\"nofollow noreferrer\">Keras does not have it's own way<\/a> to use stratified <code>k<\/code>-fold cross-validation. Instead it's advised to use <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.model_selection.StratifiedKFold.html\" rel=\"nofollow noreferrer\">sklearn's <code>StratifiedKFold<\/code><\/a>. <a href=\"https:\/\/medium.com\/@literallywords\/stratified-k-fold-with-keras-e57c487b1416\" rel=\"nofollow noreferrer\">This article<\/a> gives a detailed overview how to achieve this in Keras, \nwith the gist of it being:<\/p>\n\n<pre><code>from sklearn.model_selection import StratifiedKFold# Instantiate the cross validator\nskf = StratifiedKFold(n_splits=kfold_splits, shuffle=True)# Loop through the indices the split() method returns\nfor index, (train_indices, val_indices) in enumerate(skf.split(X, y)):\n    print \"Training on fold \" + str(index+1) + \"\/10...\"    # Generate batches from indices\n    xtrain, xval = X[train_indices], X[val_indices]\n    ytrain, yval = y[train_indices], y[val_indices]    # Clear model, and create it\n    model = None\n    model = create_model()\n\n    # Debug message I guess\n    # print \"Training new iteration on \" + str(xtrain.shape[0]) + \" training samples, \" + str(xval.shape[0]) + \" validation samples, this may be a while...\"\n\n    history = train_model(model, xtrain, ytrain, xval, yval)\n    accuracy_history = history.history['acc']\n    val_accuracy_history = history.history['val_acc']\n    print \"Last training accuracy: \" + str(accuracy_history[-1]) + \", last validation accuracy: \" + str(val_accuracy_history[-1])\n<\/code><\/pre>\n\n<ul>\n<li><code>create_model()<\/code> returns a compiled Keras model<\/li>\n<li><code>train_model()<\/code> returns last <a href=\"https:\/\/keras.io\/models\/model\/#fit\" rel=\"nofollow noreferrer\"><code>history<\/code> object<\/a> of its last <code>model.fit()<\/code> operation<\/li>\n<\/ul>\n"}]}
+{"QuestionId":57961113,"AnswerCount":2,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2019-09-16T16:38:52.360","AcceptedAnswerId":null,"OwnerUserId":3711266.0,"Title":"Randomness in TensorFlow Dataset map function","Body":"<p>I have a Dataset of raw timeseries data that I have stored in TFRecords on disk:<\/p>\n\n<pre><code>dataset = TFRecordDataset(tfrecords)  # tfrecords is a list of filenames\ndataset = dataset.map(lambda x: do_something(x))\ndataset = dataset.shuffle(1024)\ndataset = dataset.repeat()\ndataset = dataset.batch(128)        \n<\/code><\/pre>\n\n<p>What I would like my <code>do_something<\/code> function to do is, for each raw instance, take a random slice of the data so that I have a small window of data from the instance. But on the next epoch, I would like to ensure that I get a <em>different<\/em> random slice from each instance. My main question is, if introduce randomness into the <code>map<\/code> function (i.e. my <code>do_something<\/code> function), will it: <\/p>\n\n<ol>\n<li>Just take random slices once from each raw instance and then continue to iterate over those same slices on each epoch. <\/li>\n<li>Give me different random slices from each of the raw instances on each epoch.<\/li>\n<\/ol>\n\n<p>I desire (2), so if that is not happening, is there an alternative way to achieve it?<\/p>\n\n<p>For example, say I have 100 initial samples, each a timeseries of 50 data points. I want to generate 2000 samples of smaller slices, say 5-data-point slices. If I randomly select slices in my <code>map<\/code> function, will I just get the same 100 5-data-point slices on every <code>repeat<\/code>, or is there a way that I can get 100 <em>different<\/em> 5-data-point slices everytime I cycle through the 100 (50-data-point) initial samples?<\/p>\n","answers":[{"AnswerId":"57963467","CreationDate":"2019-09-16T19:44:08.100","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>Without considering your <code>do_something<\/code> function, in <code>tf.data<\/code>, as long as you are having <code>shuffle<\/code> operation before <code>repeat<\/code>, you are guaranteed to have no repetitions of any sample till all the samples are iterated through once. So, by default you are bound to get the second characteristic which is <em>Give me different random slices from each of the raw instances on each epoch.<\/em><\/p>\n\n<p>Next, <code>map<\/code> function is meant for pre-processing stuff. It may be something like standardizing your data input, or reading images from filenames etc. You shouldn't be using <code>map<\/code> to shuffle your data. You should allow <code>shuffle<\/code> operation to perform random shuffling of your data.<\/p>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>Your question is still little unclear to me, but I believe you are having totally 100 samples each of 50 datapoints (let us call them <em>main samples<\/em>) and inside map function, you are going to randomly slice out 5 points from it (let us call them <em>subsamples<\/em>). Whatever I had written in my original answer, holds true for <em>main samples<\/em>. You will have no problems in repetitions of 100 <em>main samples<\/em> you generate in each <code>batch<\/code>. But at <em>subsamples<\/em> level, repeatibility depends upon how you have written your <code>do_something<\/code> function locally. <\/p>\n"},{"AnswerId":"58103265","CreationDate":"2019-09-25T16:57:42.120","ParentId":null,"OwnerUserId":"401884","Title":null,"Body":"<p>You will get different random slices each epoch. Each epoch will call your <code>map<\/code> function again, so as long as your <code>map<\/code> function generates different slices each time it's called, you will get different slices.<\/p>\n"}]}
+{"QuestionId":57961215,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras><neural-network>","CreationDate":"2019-09-16T16:46:36.157","AcceptedAnswerId":null,"OwnerUserId":11124740.0,"Title":"neural network doesnt change output for 1 out of 2 predicted variables after training","Body":"<p><strong>tldr<\/strong>; if I train the network on 1 training example, the outcome sometimes makes no sense at all, sometimes is as expected. If I train it on more examples and higher iterations, the network, which produces two outcomes (<code>p<\/code> and <code>v<\/code>) always predicts exactly <code>0<\/code> for <code>v<\/code> and I would like to change that.<\/p>\n\n<p>I think that the main problem is the behaviour of my NN on regular data, I included the overfitting part thought since I read it as one approach to debugging, and I wanted to provide as much info as possible<\/p>\n\n<p>In the following post I will provide all code necessary to reproduce the problem.<br>\nThe first part is on its overfitting behaviour (training on one input results to exploding loss), the second part is on its behaviour on regular data (one of the prediction always stays at 0)  <\/p>\n\n<p>I am training a neural network on the same input. The wanted outcome for a value \"<code>v<\/code>\" is <code>1<\/code>. If I create the network and train it, sometimes the predicted outcome will be <code>1<\/code>, sometimes it will be <code>-<\/code>1.<br>\n Also, the <code>loss<\/code> seems to flip between <code>0<\/code> and <code>4<\/code> during training epochs.<br>\n Additionally, the loss blows up immensly, even though both losses for the outcome layers are close to zero.<br>\n I do not understand where this behaviour comes from. I used <code>Leaky-ReLU<\/code> to make sure it can handle negative input, I used a high <code>learning rate<\/code> to make sure the data in this example is sufficient on the training, and the input is the same all the time.<\/p>\n\n<p>My Neural network looks like this:<\/p>\n\n<pre><code>input_layer = keras.Input(shape=(6,7),)    \nformatted_input_layer = keras.layers.Reshape((6,7, 1))(input_layer)       \nconv_layer1 = self.create_conv_layer(formatted_input_layer)\nres_layer1 = self.create_res_layer(conv_layer1)\nres_layer2 = self.create_res_layer(res_layer1)\nres_layer3 = self.create_res_layer(res_layer2)\nres_layer4 = self.create_res_layer(res_layer3)\npolicy_head = self.create_policy_head(res_layer4)\nvalue_head = self.create_value_head(res_layer4)\nmodel = keras.Model(inputs=input_layer, outputs=[policy_head, value_head])\noptimizer = keras.optimizers.SGD(lr=args['lr'],momentum=args['momentum'])\nmodel.compile(loss = {'policy_head' : 'categorical_crossentropy', 'value_head' : 'mean_squared_error'}, optimizer=optimizer, loss_weights={'policy_head':0.5, 'value_head':0.5})\n<\/code><\/pre>\n\n<p>Methods for the different layers:<br>\nconv_layer:<\/p>\n\n<pre><code>def create_conv_layer(self, input_layer):\n    conv_layer = keras.layers.Conv2D(filters=256,\n                                     kernel_size=3,\n                                     strides=1,\n                                     padding='same',\n                                     use_bias=False,\n                                     data_format=\"channels_last\",\n                                     activation = \"linear\",\n                                     kernel_regularizer = keras.regularizers.l2(0.0001))(input_layer)\n    conv_layer= keras.layers.BatchNormalization(axis=-1)(conv_layer)\n    conv_layer = keras.layers.LeakyReLU()(conv_layer)\n    return conv_layer\n<\/code><\/pre>\n\n<p>res_layer:  <\/p>\n\n<pre><code>def create_res_layer(self, input_layer):\n        conv_layer = self.create_conv_layer(input_layer)\n        res_layer = keras.layers.Conv2D(filters=256,\n                                         kernel_size=3,\n                                         strides=1,\n                                         padding='same',\n                                         use_bias=False,\n                                         data_format=\"channels_last\",\n                                         activation = \"linear\",\n                                         kernel_regularizer = keras.regularizers.l2(0.0001))(conv_layer)\n        res_layer= keras.layers.BatchNormalization(axis=-1)(res_layer)\n        res_layer = keras.layers.add([input_layer, res_layer])\n        res_layer = keras.layers.LeakyReLU()(res_layer)\n        return res_layer\n<\/code><\/pre>\n\n<p>policy head:  <\/p>\n\n<pre><code>def create_policy_head(self, input_layer):\n        policy_head = keras.layers.Conv2D(filters=2,\n                                          kernel_size=1,\n                                          strides=1,\n                                          padding='same',\n                                          use_bias = False,\n                                          data_format='channels_last',\n                                          activation='linear',\n                                          kernel_regularizer = keras.regularizers.l2(0.0001))(input_layer)\n        policy_head = keras.layers.BatchNormalization(axis=-1)(policy_head)\n        policy_head = keras.layers.LeakyReLU()(policy_head)\n        policy_head = keras.layers.Flatten()(policy_head)\n        policy_head = keras.layers.Dense(units = 7,\n                                         use_bias = False,\n                                         activation = 'softmax',\n                                         kernel_regularizer = keras.regularizers.l2(0.0001),\n                                         name = \"policy_head\"\n                                         )(policy_head)\n        return policy_head\n<\/code><\/pre>\n\n<p>value head:  <\/p>\n\n<pre><code>def create_value_head(self, input_layer):\n        value_head = keras.layers.Conv2D(filters=1,\n                                          kernel_size=1,\n                                          strides=1,\n                                          padding='same',\n                                          use_bias = False,\n                                          data_format='channels_last',\n                                          activation='linear',\n                                          kernel_regularizer = keras.regularizers.l2(0.0001))(input_layer)\n        value_head = keras.layers.BatchNormalization(axis=-1)(value_head)\n        value_head = keras.layers.LeakyReLU()(value_head)  \n        value_head = keras.layers.Flatten()(value_head)        \n        value_head = keras.layers.Dense(units = 21,\n                                         use_bias = False,\n                                         activation = 'linear',\n                                         kernel_regularizer = keras.regularizers.l2(0.0001)\n                                         )(value_head)\n        value_head = keras.layers.LeakyReLU()(value_head)      \n        value_head = keras.layers.Dense(units = 1,\n                                         use_bias = False,\n                                         activation = 'tanh',\n                                         kernel_regularizer = keras.regularizers.l2(0.0001),\n                                         name = \"value_head\"                                     \n                                         )(value_head)\n        return value_head\n\n                                )(value_head)\n<\/code><\/pre>\n\n<p>The way I am testing my NN:<\/p>\n\n<pre><code>    canonicalBoard = np.zeros(shape = (6,7), dtype=int) \n\n\n    Pi = [0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0] \n\n    trainExamples = [[canonicalBoard, Pi, 1]]*50        \n\n    nnetwrapper.train(trainExamples)\n\n    board = canonicalBoard[np.newaxis, :, :]\n\n    p, v = nnetwrapper.nnet.model.predict(board)\n<\/code><\/pre>\n\n<p>which results in the training looking like this:<\/p>\n\n<pre><code>Epoch 1\/10\n50\/50 [==============================] - 4s 71ms\/step - loss: 1.6829 - policy_head_loss: 1.9459 - value_head_loss: 1.0000\nEpoch 2\/10\n50\/50 [==============================] - 1s 15ms\/step - loss: 2.3218 - policy_head_loss: 3.8470 - value_head_loss: 0.3768\nEpoch 3\/10\n50\/50 [==============================] - 1s 13ms\/step - loss: 4456112.5000 - policy_head_loss: 0.9027 - value_head_loss: 0.8510\nEpoch 4\/10\n50\/50 [==============================] - 1s 14ms\/step - loss: 16085884.0000 - policy_head_loss: 0.0945 - value_head_loss: 3.9925\nEpoch 5\/10\n50\/50 [==============================] - 1s 14ms\/step - loss: 32722448.0000 - policy_head_loss: 2.6572 - value_head_loss: 4.0000\nEpoch 6\/10\n50\/50 [==============================] - 1s 14ms\/step - loss: 52690084.0000 - policy_head_loss: 9.6345 - value_head_loss: 3.1810e-12\nEpoch 7\/10\n50\/50 [==============================] - 1s 14ms\/step - loss: 74703120.0000 - policy_head_loss: 1.1921e-07 - value_head_loss: 4.0000\nEpoch 8\/10\n50\/50 [==============================] - 1s 14ms\/step - loss: 97784832.0000 - policy_head_loss: 1.1921e-07 - value_head_loss: 4.0000\nEpoch 9\/10\n50\/50 [==============================] - 1s 14ms\/step - loss: 121202520.0000 - policy_head_loss: 2.0802e-05 - value_head_loss: 4.0000\nEpoch 10\/10\n50\/50 [==============================] - 1s 14ms\/step - loss: 144415040.0000 - policy_head_loss: 1.1921e-07 - value_head_loss: 4.0000\n<\/code><\/pre>\n\n<p>and my prediction looking like this:<\/p>\n\n<pre><code>p: [[0. 0. 0. 0. 0. 1. 0.]]  v: [[-1.]]\n<\/code><\/pre>\n\n<p>another outcome could be:<\/p>\n\n<pre><code>Epoch 1\/10\n50\/50 [==============================] - 4s 82ms\/step - loss: 1.6829 - policy_head_loss: 1.9459 - value_head_loss: 1.0000\nEpoch 2\/10\n50\/50 [==============================] - 1s 17ms\/step - loss: 2.2826 - policy_head_loss: 2.0001 - value_head_loss: 2.1454\nEpoch 3\/10\n50\/50 [==============================] - 1s 16ms\/step - loss: 1718694.1250 - policy_head_loss: 0.5434 - value_head_loss: 3.9772\nEpoch 4\/10\n50\/50 [==============================] - 1s 14ms\/step - loss: 6204218.0000 - policy_head_loss: 9.4180e-05 - value_head_loss: 0.0000e+00\nEpoch 5\/10\n50\/50 [==============================] - 1s 14ms\/step - loss: 12620835.0000 - policy_head_loss: 1.1921e-07 - value_head_loss: 0.0000e+00\nEpoch 6\/10\n50\/50 [==============================] - 1s 14ms\/step - loss: 20322232.0000 - policy_head_loss: 7.7489 - value_head_loss: 0.0000e+00\nEpoch 7\/10\n50\/50 [==============================] - 1s 14ms\/step - loss: 28812526.0000 - policy_head_loss: 1.1921e-07 - value_head_loss: 1.5966\nEpoch 8\/10\n50\/50 [==============================] - 1s 14ms\/step - loss: 37715012.0000 - policy_head_loss: 1.1921e-07 - value_head_loss: 0.0000e+00\nEpoch 9\/10\n50\/50 [==============================] - 1s 15ms\/step - loss: 46747064.0000 - policy_head_loss: 1.1921e-07 - value_head_loss: 0.0000e+00\nEpoch 10\/10\n50\/50 [==============================] - 1s 14ms\/step - loss: 55699992.0000 - policy_head_loss: 1.1921e-07 - value_head_loss: 0.0000e+00\n<\/code><\/pre>\n\n<p>with the predictions:<\/p>\n\n<pre><code>p: [[0. 0. 0. 0. 0. 1. 0.]]  v: [[1.]]\n<\/code><\/pre>\n\n<p>which are the correct ones as I would have expected.<\/p>\n\n<p>How come on some trainings, my NN doesnt fit the data at all?   I wanted to start the training process for a whole week now, but before I do so I want to make sure there are no errors in the way I layed out my NN.   And this looks like I am missing something here.<\/p>\n\n<p>And here at the <code>predict<\/code> \/ <code>train<\/code> methods of my neuralnet, (the NN is part of an alpha-zero replica and the involved game is connect4, I omitted these in the example to make it easier to actually replicate the problem. This is why you see some transform operations in <code>predict<\/code> and <code>train<\/code> methods)<\/p>\n\n<pre><code>def train(self, examples):\n        input_boards, target_pis, target_vs = list(zip(*examples))       \n        input_boards = np.asarray(input_boards)\n        target_pis = np.asarray(target_pis)\n        target_vs = np.asarray(target_vs)        \n        logger.debug(\"Passing to nn: x: {}, y: {}, batch_size: {}, epochs: {}\".format(input_boards, [target_pis, target_vs], self.args[\"batch_size\"], self.args[\"epochs\"]))\n        self.nnet.model.fit(x = input_boards, y = [target_pis, target_vs], batch_size = self.args[\"batch_size\"], epochs = self.args[\"epochs\"])\n\n\ndef predict(self, board):\n        board = board.nn_board_2d\n        # preparing input\n        board = board[np.newaxis, :, :] # this has to be done for the conv2d to work\n\n        # run\n        pi, v = self.nnet.model.predict(board)\n        return pi[0], v[0]\n<\/code><\/pre>\n\n<p>The parameters I used for this example:<\/p>\n\n<pre><code>'lr': 0.2\n'dropout': 0.1\n'epochs': 10\n'num_channels': 512,\n'filters': 256\n'momentum':0.9\n<\/code><\/pre>\n\n<p>As soon as I use a lower learning rate and more iterations, my <code>p<\/code> changes, but my <code>v<\/code> stays exactly at <code>0<\/code>. This is what is bothering me the most:<\/p>\n\n<pre><code>Epoch 1\/10\n118\/118 [==============================] - 5s 44ms\/step - loss: 1.8037 - policy_head_loss: 2.4305 - value_head_loss: 0.7578\nEpoch 2\/10\n118\/118 [==============================] - 2s 14ms\/step - loss: 1.1666 - policy_head_loss: 1.7723 - value_head_loss: 0.1416\nEpoch 3\/10\n118\/118 [==============================] - 2s 13ms\/step - loss: 1.0987 - policy_head_loss: 1.6832 - value_head_loss: 0.0948\nEpoch 4\/10\n118\/118 [==============================] - 2s 13ms\/step - loss: 1.0430 - policy_head_loss: 1.5787 - value_head_loss: 0.0876\nEpoch 5\/10\n118\/118 [==============================] - 2s 13ms\/step - loss: 0.9943 - policy_head_loss: 1.4859 - value_head_loss: 0.0826\nEpoch 6\/10\n118\/118 [==============================] - 2s 13ms\/step - loss: 0.9469 - policy_head_loss: 1.3959 - value_head_loss: 0.0777\nEpoch 7\/10\n118\/118 [==============================] - 2s 13ms\/step - loss: 0.9046 - policy_head_loss: 1.3180 - value_head_loss: 0.0707\nEpoch 8\/10\n118\/118 [==============================] - 2s 13ms\/step - loss: 0.8629 - policy_head_loss: 1.2403 - value_head_loss: 0.0647\nEpoch 9\/10\n118\/118 [==============================] - 2s 14ms\/step - loss: 0.8068 - policy_head_loss: 1.1344 - value_head_loss: 0.0583\nEpoch 10\/10\n118\/118 [==============================] - 2s 13ms\/step - loss: 0.7335 - policy_head_loss: 0.9922 - value_head_loss: 0.0538\nI0916 14:24:38.595780  7116 trainingonly1NN.py:232] Values for empty board: new network: P : [0.20057818 0.11068489 0.15129891 0.20042823 0.13117987 0.04705378\n 0.15877616]  v: 0\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57961355,"AnswerCount":0,"Tags":"<tensorflow><tensorflow2.0>","CreationDate":"2019-09-16T16:57:52.083","AcceptedAnswerId":null,"OwnerUserId":8029329.0,"Title":"AttributeError: 'TensorSliceDataset' object has no attribute \"'element_spec'\"","Body":"<p>I am very new to tensorflow. I ran following code - <\/p>\n\n<pre><code>import tensorflow as tf\nd = tf.data.Dataset.from_tensor_slices(tf.random.uniform([3,6]))\nd.element_spec\n<\/code><\/pre>\n\n<p>And got following error -<\/p>\n\n<pre><code>---------------------------------------------------------------------------\n\nAttributeError                            Traceback (most recent call last)\n\n&lt;ipython-input-4-a00819363729&gt; in &lt;module&gt;()\n      1 import tensorflow as tf\n      2 d = tf.data.Dataset.from_tensor_slices(tf.random.uniform([3,6]))\n----&gt; 3 d.element_spec\n\nAttributeError: 'TensorSliceDataset' object has no attribute 'element_spec'\n<\/code><\/pre>\n\n<p>I cannot find solution for this error. I just want to print the type of each element component in variable d. What is going wrong here??<\/p>\n\n<p>I am using tensorflow 2.0. <\/p>\n","answers":[]}
+{"QuestionId":57962268,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><deep-learning><conv-neural-network>","CreationDate":"2019-09-16T18:09:33.447","AcceptedAnswerId":57962959.0,"OwnerUserId":7539208.0,"Title":"Printing tensor value output of model.output in keras","Body":"<p>So I'm trying to print the value of the model.output (which is a tensor) of pretrained VGG16 model, but unfortunately all the below listed methods didn't work in my case, which I found from the previously answered similar questions.<\/p>\n\n<p>1) Firstly on the basis of <a href=\"https:\/\/stackoverflow.com\/questions\/54882511\/actually-printing-values-from-tensor-object\">this<\/a> post I tried using <\/p>\n\n<blockquote>\n  <p>print(K.eval(model.output()))<\/p>\n<\/blockquote>\n\n<p>but it threw an error as<\/p>\n\n<blockquote>\n  <p>TypeError: 'Tensor' object is not callable<\/p>\n<\/blockquote>\n\n<p>2) Then upon going through <a href=\"https:\/\/stackoverflow.com\/questions\/43448029\/debugging-keras-tensor-values\">this<\/a> post I tried using <\/p>\n\n<blockquote>\n  <p>K.print_tensor(model.output, message='model.output = ')<\/p>\n<\/blockquote>\n\n<p>approach but this time there was no output printed.<\/p>\n\n<p>Here is my code:<\/p>\n\n<pre><code>from keras.applications.vgg16 import VGG16\nfrom keras.applications.vgg16 import preprocess_input\nfrom keras.preprocessing.image import load_img\nfrom keras.preprocessing.image import img_to_array\nfrom keras.models import Model\nimport keras.backend as K\nfrom matplotlib import pyplot\nfrom numpy import expand_dims\nimport numpy as np\n\n\n# load the model\nmodel = VGG16()\n\nimg = load_img('..\/input\/treebird\/bird.jpg', target_size=(224, 224))\nimg = img_to_array(img)           \nimg = expand_dims(img, axis=0)      \nimg = preprocess_input(img)       \n\nprediction = model.predict(img)\n\nprint(model.output)                                 #Tensor(\"predictions_19\/Softmax:0\",       shape=(?, 1000), dtype=float32)\nprint(K.eval(model.output()))                          # throws TypeError: 'Tensor' object is not callable\nK.print_tensor(model.output, message='model.output = ')\n<\/code><\/pre>\n\n<p>Am I lacking somewhere in the implementation of the above methods or are there other methods that I am supposed to use to print the tensor in this case?<\/p>\n","answers":[{"AnswerId":"57962959","CreationDate":"2019-09-16T19:03:15.910","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>Posting only the relevant parts of the code.<\/p>\n\n<p>First you will have to clear the session so that your input node in model will have same name everytime. Use <code>model.summary()<\/code> to know what is the input node name.<\/p>\n\n<p>Then, you will have to make use of <code>feed_dict<\/code> to pass the tensor as shown in the code.<\/p>\n\n<pre><code>K.clear_session()\n\n# load the model\nmodel = VGG16()\nmodel.summary()\n\nprint(model.output.eval(session=K.get_session(), feed_dict={'input_1:0': img}))   \n<\/code><\/pre>\n"}]}
+{"QuestionId":57962430,"AnswerCount":0,"Tags":"<keras><deep-learning><classification><vgg-net><bias-neuron>","CreationDate":"2019-09-16T18:21:23.303","AcceptedAnswerId":null,"OwnerUserId":7065438.0,"Title":"Why CNN model tends to classify most instances in a specific class?","Body":"<p>I used the Vgg16 model with my data set for a binary classification problem (Benign or malignant). I trained the model using transfer learning, and it turned out that the model is predicting most of the images as benign. The interesting thing here is that, in my training data, the number of training sample for benign is 136, and the number of malignant sample is 144. So, why the model is predicting most of the test image as benign? To investigate more, I test the model with some of the malignant images that were used for training, and surprisingly find out that the model is predicting the same malignant image(that were used for training the model) as benign when testing. My question is why the model is predicting most of the testing images as benign, when the no of images in the training sample for benign cases are less then the malignant samples.<\/p>\n\n<pre><code>train_path = 'Digital_Mamo\/OPTIMAM'     \nvalid_path = 'Digital_Mamo\/InBreast'\ntest_path = 'Digital_Mamo\/OPTIMAM\/Test_malignat_big'\n\nvalid_batches = ImageDataGenerator().flow_from_directory(valid_path, target_size=(224, 224), classes=['Benign', 'Malignant'], batch_size=9) \ntest_batches = ImageDataGenerator().flow_from_directory(test_path, target_size=(224, 224), classes=['Benign', 'Malignant'], batch_size=14)\n\ntrain_batches = datagen.flow_from_directory(\n        train_path,\n        target_size=(224, 224),\n        batch_size=10,\n        classes=['Benign','Malignant'])\n\nvgg16_model= load_model('Fetched_VGG.h5')\nmodel= Sequential()\nfor layer in vgg16_model.layers[:-1]:\n    model.add(layer)\n\nfor layer in model.layers:\n    layer.trainable = False\n\nmodel.add(Dense(2, activation='softmax', name='predictions'))\n\nmodel.compile(Adam(lr=.0001), loss='categorical_crossentropy', metrics=['accuracy'])\n\nhist = model.fit_generator(train_batches, steps_per_epoch=10, validation_data=valid_batches, validation_steps=10, epochs=5, verbose=2)\n\nscoreSeg = model.evaluate_generator(test_batches)\n\npredictions = model.predict_generator(test_batches, steps=10, verbose=0)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57962520,"AnswerCount":1,"Tags":"<java><spring-boot><caffeine><caffeine-cache>","CreationDate":"2019-09-16T18:27:21.343","AcceptedAnswerId":null,"OwnerUserId":105408.0,"Title":"Caffeine cache \"cannot be cast to org.springframework.cache.Cache\"","Body":"<p>I am attempting to use Spring Boot caching.  I have a cached class:<\/p>\n\n<pre><code>@Component\npublic class GlobalAssetIdCache {\n\n  private static final Logger LOG = LogManager.getLogger(GlobalAssetIdCache.class);\n\n  @Cacheable(value = \"GLOBAL_ASSET_ID\", cacheManager = \"cafeine\")\n  public String fetchGlobalAssetId(String make, String serialNumber) {\n    LOG.info(\"Getting cache value for:{}-{}\", make, serialNumber);\n    return \"MARKISCOOL\";\n  }\n}\n<\/code><\/pre>\n\n<p>I have a config:<\/p>\n\n<pre><code>@EnableCaching\npublic class LambdaAppTestConfig {\n\n  @Bean(name = \"caffeine\")\n  public CacheManager cacheManager() {\n    return StewardCache.getCacheManager();\n  }\n<\/code><\/pre>\n\n<p>I have a cache:<\/p>\n\n<pre><code>public enum StewardCache {\n\n  GLOBAL_ASSET_ID(500000, 0);\n\n  private Cache cache;\n\n  \/**\n   * Build cache.\n   *\/\n  StewardCache(int size, int expire) {\n    Caffeine cache = Caffeine.newBuilder()\n        .maximumSize(size)\n        .recordStats();\n    if (expire &gt; 0) {\n      cache.expireAfterAccess(expire, TimeUnit.MINUTES);\n    }\n    this.cache = cache.build();\n  }\n\n  public static CacheManager getCacheManager() {\n    SimpleCacheManager simpleCacheManager = new SimpleCacheManager();\n    List caches = new ArrayList();\n    for (int i = 0; i &lt; StewardCache.values().length; i++) {\n      StewardCache aCache = StewardCache.values()[i];\n      caches.add(aCache.cache);\n    }\n    simpleCacheManager.setCaches(caches);\n    return simpleCacheManager;\n  }\n\n}\n<\/code><\/pre>\n\n<p>I am getting this error:<\/p>\n\n<pre><code> java.lang.ClassCastException: com.github.benmanes.caffeine.cache.BoundedLocalCache$BoundedLocalManualCache cannot be cast to org.springframework.cache.Cache\n<\/code><\/pre>\n\n<p>I have tried using Guava cache as well, but I get the same error.  What am I missing here?  How do I force Spring to use Caffeine?<\/p>\n","answers":[{"AnswerId":"57977964","CreationDate":"2019-09-17T15:55:59.863","ParentId":null,"OwnerUserId":"105408","Title":null,"Body":"<p>@GvSharma gave me the answer.<\/p>\n\n<p>I changed to this:<\/p>\n\n<pre><code> public static CacheManager getCacheManager() {\n    SimpleCacheManager simpleCacheManager = new SimpleCacheManager();\n    List caches = new ArrayList();\n    for (int i = 0; i &lt; StewardCache.values().length; i++) {\n      StewardCache aCache = StewardCache.values()[i];\n      caches.add(new CaffeineCache(aCache.name(), aCache.cache));\n    }\n    simpleCacheManager.setCaches(caches);\n    return simpleCacheManager;\n  }\n<\/code><\/pre>\n\n<p>So Spring has a fist full of wrappers for various caches.  I did not find this documented anywhere.  Once I did this, it all worked great.<\/p>\n"}]}
+{"QuestionId":57962694,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning><google-colaboratory>","CreationDate":"2019-09-16T18:42:10.083","AcceptedAnswerId":null,"OwnerUserId":8803677.0,"Title":"Why does my Generative Adverserial Network keep running out of memory when training?","Body":"<p>I am having an issues with training my GAN on Google Colab. My training dataset is a folder of MIDI files, which are converted to numpy arrays using the pypianoroll package. It keeps running out of the 25 gigabytes of memory which is available, I assume that amount should definitely be enough. <\/p>\n\n<p>I tried lowering the batch size to 1 but it didn't help.<\/p>\n\n<pre><code>class GANDrum(object):\n    def __init__(self,buffer_size=60000,batch_size=1,epochs=20,noise_dim=100,num_examples_to_generate=1):\n    self.train_files=[]\n    self.buffer_size=buffer_size\n    self.batch_size=batch_size\n    self.load_data()\n\n    self.generator_optimizer=tf.keras.optimizers.Adam(1e-4)\n    self.discriminator_optimizer=tf.keras.optimizers.Adam(1e-4)\n\n    self.generator=self.make_generator_model()\n    self.discriminator=self.make_discriminator_model()\n\n    self.checkpoint_prefix=os.path.join('.\/training_checkpoints',\"ckpt\")\n    self.checkpoint=tf.train.Checkpoint(generator_optimizer=self.generator_optimizer,discriminator_optimizer=self.discriminator_optimizer,generator=self.generator,discriminator=self.discriminator)\n\n    self.epochs=epochs\n    self.noise_dim=noise_dim\n    self.seed=tf.random.normal([num_examples_to_generate, noise_dim])\n\n\ndef load_data(self):\n    i=0\n    entries=os.listdir(r'\/content\/drive\/My Drive\/midi_d_processed\/')\n    for entry in entries:\n        try:\n            data=(pp.parse(r'\/content\/drive\/My Drive\/midi_d_processed\/'+entry))\n            mtx=tf.reshape(tensor=tf.convert_to_tensor(data.tracks[0].pianoroll),shape=(8,12,128))\n\n            mtx=tf.cast(mtx, tf.float32)\n            self.train_files.append(mtx)\n        except:\n            i=i+1\n    print(i,'files were not in the correct format')\n    self.train_dataset=tf.data.Dataset.from_tensor_slices(self.train_files).shuffle(self.buffer_size).batch(self.batch_size)\n\ndef make_generator_model(self):\n    try:\n        model=tf.keras.Sequential()\n\n\n        model.add(layers.Dense(units=1536,input_shape=(100,)))\n        model.add(layers.BatchNormalization())\n        model.add(layers.LeakyReLU())\n\n        model.add(layers.Dense(units=768))\n        model.add(layers.BatchNormalization())\n        model.add(layers.LeakyReLU())\n\n\n        model.add(layers.Reshape((2,3,128)))\n\n\n        model.add(layers.Conv2DTranspose(filters=1,strides=2,kernel_size=2))\n        model.add(layers.BatchNormalization())\n        model.add(layers.LeakyReLU())\n\n        model.add(layers.Conv2DTranspose(filters=1,strides=2,kernel_size=2))\n        model.add(layers.BatchNormalization())\n        model.add(layers.LeakyReLU())\n\n        model.add(layers.Conv2DTranspose(filters=128,strides=1,kernel_size=1,activation='tanh'))\n        model.add(layers.BatchNormalization())\n        model.add(layers.Softmax())\n\n        return model\n    except:\n        print('make_generator_model')\n\n\n\ndef make_discriminator_model(self):\n        model=tf.keras.Sequential()\n\n        model.add(layers.Conv2D(filters=128,strides=1,kernel_size=1,input_shape=[8,12,128]))\n        model.add(layers.LeakyReLU())\n        model.add(layers.Conv2D(filters=1,strides=2,kernel_size=2,input_shape=[8,12,128]))\n        model.add(layers.LeakyReLU())\n\n        model.add(layers.Conv2D(filters=1,strides=2,kernel_size=2))\n        model.add(layers.LeakyReLU())\n\n        model.add(layers.Conv2D(filters=1,kernel_size=2))\n        model.add(layers.LeakyReLU())\n        model.add(layers.Flatten())\n        model.add(layers.Dense(1))\n\n        return model\n\n\ndef discriminator_loss(self,real_output,fake_output):\n    cross_entropy=tf.keras.losses.BinaryCrossentropy(from_logits=True)\n\n    real_loss=cross_entropy(tf.ones_like(real_output),real_output)\n    fake_loss=cross_entropy(tf.zeros_like(fake_output),fake_output)\n    total_loss=real_loss+fake_loss\n    return total_loss\n\n\ndef generator_loss(self,fake_output):\n    cross_entropy=tf.keras.losses.BinaryCrossentropy(from_logits=True)\n\n    return cross_entropy(tf.ones_like(fake_output),fake_output)\n\n\ndef train_step(self,files):\n    noise = tf.random.normal([self.batch_size,self.noise_dim])\n\n    with tf.GradientTape() as gen_tape,tf.GradientTape() as disc_tape:\n        generated_midi=self.generator(noise,training=True)\n\n        real_output=self.discriminator(files,training=True)\n        fake_output=self.discriminator(generated_midi,training=True)\n\n        gen_loss=self.generator_loss(fake_output)\n        disc_loss=self.discriminator_loss(real_output,fake_output)\n\n    gradients_of_generator=gen_tape.gradient(gen_loss,self.generator.trainable_variables)\n    gradients_of_discriminator=disc_tape.gradient(disc_loss,self.discriminator.trainable_variables)\n\n    self.generator_optimizer.apply_gradients(zip(gradients_of_generator,self.generator.trainable_variables))\n    self.discriminator_optimizer.apply_gradients(zip(gradients_of_discriminator,self.discriminator.trainable_variables))\n\n\ndef train(self,dataset,epochs):\n    for self.epoch in range(epochs):\n        start=time.time()\n\n    for midi_batch in self.train_dataset:\n        self.train_step(midi_batch)\n\n    self.generate_and_save_midi(self.generator,self.epoch+1,self.seed)\n\n    if(self.epoch+1)%15==0:\n        checkpoint.save(file_prefix=self.checkpoint_prefix)\n\n    print ('Time for epoch {} is {} sec'.format(self.epoch+1,time.time()-start))\n\n\ndef generate_and_save_midi(self,model,epoch,test_input):\n        predictions=model(test_input,training=False)\n\n        shaped_predictions=np.reshape(a=predictions.numpy(),newshape=(96,128))\n\n        track=pp.Track(pianoroll=shaped_predictions,is_drum=True)\n\n        multitrack=pp.Multitrack(tracks=[track])\n\n        multitrack.write('.\/generated_midi')\n\n\ndef run(self):\n    self.train(self.train_dataset,self.epochs)\n<\/code><\/pre>\n\n<p>I usually get an error message that says that I'm out of memory, but sometimes I get an unknown error. What could be the problem?<\/p>\n","answers":[]}
+{"QuestionId":57963062,"AnswerCount":0,"Tags":"<android><tensorflow><mobile><object-detection>","CreationDate":"2019-09-16T19:09:54.737","AcceptedAnswerId":null,"OwnerUserId":3965296.0,"Title":"Is the current object detection technology advanced enough to detect small objects like football players on the field on mobile phone?","Body":"<p>I've been looking at implementations of object detection models like YOLO to frameworks like tensorflowLite, I was wondering if the mobile phone's RAM\/computing power is able to process recorded videos like this on the cell phone itself. On tensorFlowLite website, it also says that max 10 objects can be detected.<\/p>\n","answers":[]}
+{"QuestionId":57963254,"AnswerCount":1,"Tags":"<matlab><tensorflow><dataset><image-segmentation><deeplab>","CreationDate":"2019-09-16T19:25:16.920","AcceptedAnswerId":58734226.0,"OwnerUserId":10389501.0,"Title":"How to generate proper labelled image from MATLAB (Image Labeler) for image segmentation","Body":"<p>I get black image whenever exporting labels to file in the Image Labeler App in MATLAB R2019a<\/p>\n\n<p>Here is what i do:<\/p>\n\n<pre><code>Export Labels &gt; To File\n<\/code><\/pre>\n\n<p>Also, I know that PNG file's image value is composed as 0 1 2 and this is the reason image apperas black.\n<a href=\"https:\/\/i.stack.imgur.com\/SmIRF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/SmIRF.png\" alt=\"b\"><\/a><\/p>\n\n<p>I tried this piece of code and I get the result with distingished but the dimensions of the image are changed.<\/p>\n\n<pre><code>    figure;\n    [i, m] = imread('PixelLabelData\/Label_1.png');\n    imshow(i,m);\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/poEOz.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/poEOz.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I need is Labelled png image with proper dimensions. How can I do this ?<\/p>\n","answers":[{"AnswerId":"58734226","CreationDate":"2019-11-06T16:05:31.270","ParentId":null,"OwnerUserId":"10389501","Title":null,"Body":"<p>I am using <a href=\"https:\/\/supervise.ly\" rel=\"nofollow noreferrer\">https:\/\/supervise.ly<\/a> instead of MATLAB. Development is much easier and faster.<\/p>\n"}]}
+{"QuestionId":57963649,"AnswerCount":0,"Tags":"<python><python-3.x><keras><theano>","CreationDate":"2019-09-16T19:58:48.447","AcceptedAnswerId":null,"OwnerUserId":11415906.0,"Title":"error when using the keras api on convelutional layers","Body":"<p>I'm trying to build a model that looks like this:<\/p>\n\n<pre><code>input\n  \/\nconvlayers\/flatten\n   \/         \\\nfirst_output  \\\n     \\        \/\n   second_output\n<\/code><\/pre>\n\n<p>but it fails at the first conv layers with the error:<\/p>\n\n<pre><code>ValueError: Layer conv2d_4 was called with an input that isn't a symbolic tensor.\nReceived type: &lt;class 'keras.layers.convolutional.Conv2D'&gt;.\nFull input: [&lt;keras.layers.convolutional.Conv2D object at 0x7f450d7b8630&gt;].\nAll inputs to the layer should be tensors.\n<\/code><\/pre>\n\n<p>and the error points to the layer after the first conv with the inputshape call.<\/p>\n\n<p>Help would be appreciated.<\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><code>conv1 = Conv2D(8, 4, padding = \"same\", strides = 2)(inputs)\nconv2 = Conv2D(16 ,4, padding = \"same\", strides = 2)(conv1)\nflat = Flatten()(conv2)\ndense1 = Dense(32)(flat)\n\ndense2 = Dense(32)(dense1)\nfirst_output = Dense(64)(dense2)\nmerged = concatenate([flat,first_output])\nsecond_output_dense1 = Dense(32)(merged)\nsecond_output_dense2 = Dense(32)(second_output_dense1)\nsecond_output = Dense(64)(second_output_dense2)\n\nmodel = Model(inputs=conv1, outputs=[first_output,second_output])\nmodel.compile(loss = \"mse\", optimizer = \"adam\" )\n<\/code><\/pre>\n\n<p>Answer:<\/p>\n\n<p>i was under the impression that you could call the model without an input layer and just define the input in the first layer : conv1 = Conv2D(8, 4, padding = \"same\", strides = 2, <strong><em>input_shape = (6,8,8,)<\/em><\/strong>)<\/p>\n\n<p>but that didnt work so instead you have to delete the input shape thing and create an input layer here is the fixed code<\/p>\n\n<pre><code>inputs = Input(shape=(6,8,8,))\nconv1 = Conv2D(8, 4, padding = \"same\", strides = 2, input_shape = (6,8,8,))\nconv2 = Conv2D(16 ,4, padding = \"same\", strides = 2)(conv1)\nflat = Flatten()(conv2)\ndense1 = Dense(32)(flat)\n\ndense2 = Dense(32)(dense1)\nfirst_output = Dense(64)(dense2)\nmerged = concatenate([flat,first_output])\nsecond_output_dense1 = Dense(32)(merged)\nsecond_output_dense2 = Dense(32)(second_output_dense1)\nsecond_output = Dense(64)(second_output_dense2)\n\nmodel = Model(inputs=inputs, outputs=[first_output,second_output])\nmodel.compile(loss = \"mse\", optimizer = \"adam\" )\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57964165,"AnswerCount":1,"Tags":"<python><tensorflow><julia>","CreationDate":"2019-09-16T20:43:18.173","AcceptedAnswerId":null,"OwnerUserId":10281232.0,"Title":"How to convert tf.layer_norm to julia?","Body":"<p>I'm trying to convert tensorflow code to julia but I didn't understand how <code>layer_norm<\/code> function works.<\/p>\n\n<p>I'm trying to convert that part:<\/p>\n\n<pre><code>output = tf.contrib.layers.layer_norm(attn_out + h, begin_norm_axis=-1,\n                                              scope='LayerNorm')\n<\/code><\/pre>\n\n<p>shape of the tensor <code>attn_out + h<\/code> is <code>( qlen,bsz,d_model )<\/code> and when I downloaded the trained model, variables gamma and beta have shape  <code>(d_model,)<\/code>. Also, output of <code>nn.moments<\/code> should have shape <code>[qlen,bsz,1]<\/code>. <code>layer_norm<\/code> function calls <code>batch_norm<\/code> function with these parameters:<\/p>\n\n<pre><code>add_arg_scope\ndef layer_norm(inputs,\n               center=True,\n               scale=True,\n               activation_fn=None,\n               reuse=None,\n               variables_collections=None,\n               outputs_collections=None,\n               trainable=True,\n               begin_norm_axis=1,\n               begin_params_axis=-1,\n               scope=None):\n  \"\"\"Adds a Layer Normalization layer.\n  Based on the paper:\n    \"Layer Normalization\"\n    Jimmy Lei Ba, Jamie Ryan Kiros, Geoffrey E. Hinton\n    https:\/\/arxiv.org\/abs\/1607.06450.\n  Can be used as a normalizer function for conv2d and fully_connected.\n  Given a tensor `inputs` of rank `R`, moments are calculated and normalization\n  is performed over axes `begin_norm_axis ... R - 1`.  Scaling and centering,\n  if requested, is performed over axes `begin_params_axis .. R - 1`.\n  By default, `begin_norm_axis = 1` and `begin_params_axis = -1`,\n  meaning that normalization is performed over all but the first axis\n  (the `HWC` if `inputs` is `NHWC`), while the `beta` and `gamma` trainable\n  parameters are calculated for the rightmost axis (the `C` if `inputs` is\n  `NHWC`).  Scaling and recentering is performed via broadcast of the\n  `beta` and `gamma` parameters with the normalized tensor.\n  The shapes of `beta` and `gamma` are `inputs.shape[begin_params_axis:]`,\n  and this part of the inputs' shape must be fully defined.\n  Args:\n    inputs: A tensor having rank `R`. The normalization is performed over axes\n      `begin_norm_axis ... R - 1` and centering and scaling parameters are\n      calculated over `begin_params_axis ... R - 1`.\n    center: If True, add offset of `beta` to normalized tensor. If False, `beta`\n      is ignored.\n    scale: If True, multiply by `gamma`. If False, `gamma` is not used. When the\n      next layer is linear (also e.g. `nn.relu`), this can be disabled since the\n      scaling can be done by the next layer.\n    activation_fn: Activation function, default set to None to skip it and\n      maintain a linear activation.\n    reuse: Whether or not the layer and its variables should be reused. To be\n      able to reuse the layer scope must be given.\n    variables_collections: Optional collections for the variables.\n    outputs_collections: Collections to add the outputs.\n    trainable: If `True` also add variables to the graph collection\n      `GraphKeys.TRAINABLE_VARIABLES` (see tf.Variable).\n    begin_norm_axis: The first normalization dimension: normalization will be\n      performed along dimensions `begin_norm_axis : rank(inputs)`\n    begin_params_axis: The first parameter (beta, gamma) dimension: scale and\n      centering parameters will have dimensions\n      `begin_params_axis : rank(inputs)` and will be broadcast with the\n        normalized inputs accordingly.\n    scope: Optional scope for `variable_scope`.\n  Returns:\n    A `Tensor` representing the output of the operation, having the same\n    shape and dtype as `inputs`.\n  Raises:\n    ValueError: If the rank of `inputs` is not known at graph build time,\n      or if `inputs.shape[begin_params_axis:]` is not fully defined at\n      graph build time.\n  \"\"\"\n  with variable_scope.variable_scope(\n      scope, 'LayerNorm', [inputs], reuse=reuse) as sc:\n    inputs = ops.convert_to_tensor(inputs)\n    inputs_shape = inputs.shape\n    inputs_rank = inputs_shape.ndims\n    if inputs_rank is None:\n      raise ValueError('Inputs %s has undefined rank.' % inputs.name)\n    dtype = inputs.dtype.base_dtype\n    if begin_norm_axis &lt; 0:\n      begin_norm_axis = inputs_rank + begin_norm_axis\n    if begin_params_axis &gt;= inputs_rank or begin_norm_axis &gt;= inputs_rank:\n      raise ValueError('begin_params_axis (%d) and begin_norm_axis (%d) '\n                       'must be &lt; rank(inputs) (%d)' %\n                       (begin_params_axis, begin_norm_axis, inputs_rank))\n    params_shape = inputs_shape[begin_params_axis:]\n    if not params_shape.is_fully_defined():\n      raise ValueError(\n          'Inputs %s: shape(inputs)[%s:] is not fully defined: %s' %\n          (inputs.name, begin_params_axis, inputs_shape))\n    # Allocate parameters for the beta and gamma of the normalization.\n    beta, gamma = None, None\n    if center:\n      beta_collections = utils.get_variable_collections(variables_collections,\n                                                        'beta')\n      beta = variables.model_variable(\n          'beta',\n          shape=params_shape,\n          dtype=dtype,\n          initializer=init_ops.zeros_initializer(),\n          collections=beta_collections,\n          trainable=trainable)\n    if scale:\n      gamma_collections = utils.get_variable_collections(\n          variables_collections, 'gamma')\n      gamma = variables.model_variable(\n          'gamma',\n          shape=params_shape,\n          dtype=dtype,\n          initializer=init_ops.ones_initializer(),\n          collections=gamma_collections,\n          trainable=trainable)\n    # By default, compute the moments across all the dimensions except the one with index 0.\n    norm_axes = list(range(begin_norm_axis, inputs_rank))\n    mean, variance = nn.moments(inputs, norm_axes, keep_dims=True)\n    # Compute layer normalization using the batch_normalization function.\n    # Note that epsilon must be increased for float16 due to the limited\n    # representable range.\n    variance_epsilon = 1e-12 if dtype != dtypes.float16 else 1e-3\n    outputs = nn.batch_normalization(\n        inputs,\n        mean,\n        variance,\n        offset=beta,\n        scale=gamma,\n        variance_epsilon=variance_epsilon)\n    outputs.set_shape(inputs_shape)\n    if activation_fn is not None:\n      outputs = activation_fn(outputs)\n    return utils.collect_named_outputs(outputs_collections, sc.name, outputs)\n\n<\/code><\/pre>\n\n<p>However none of these tensors <code>(gamma,beta,mean,variance)<\/code> satisfy the input restrictions of <code>batch_norm<\/code>:<\/p>\n\n<pre><code>@tf_export(\"nn.batch_normalization\")\ndef batch_normalization(x,\n                        mean,\n                        variance,\n                        offset,\n                        scale,\n                        variance_epsilon,\n                        name=None):\n  r\"\"\"Batch normalization.\n  Normalizes a tensor by `mean` and `variance`, and applies (optionally) a\n  `scale` \\\\(\\gamma\\\\) to it, as well as an `offset` \\\\(\\beta\\\\):\n  \\\\(\\frac{\\gamma(x-\\mu)}{\\sigma}+\\beta\\\\)\n  `mean`, `variance`, `offset` and `scale` are all expected to be of one of two\n  shapes:\n    * In all generality, they can have the same number of dimensions as the\n      input `x`, with identical sizes as `x` for the dimensions that are not\n      normalized over (the 'depth' dimension(s)), and dimension 1 for the\n      others which are being normalized over.\n      `mean` and `variance` in this case would typically be the outputs of\n      `tf.nn.moments(..., keep_dims=True)` during training, or running averages\n      thereof during inference.\n    * In the common case where the 'depth' dimension is the last dimension in\n      the input tensor `x`, they may be one dimensional tensors of the same\n      size as the 'depth' dimension.\n      This is the case for example for the common `[batch, depth]` layout of\n      fully-connected layers, and `[batch, height, width, depth]` for\n      convolutions.\n      `mean` and `variance` in this case would typically be the outputs of\n      `tf.nn.moments(..., keep_dims=False)` during training, or running averages\n      thereof during inference.\n  See Source: [Batch Normalization: Accelerating Deep Network Training by\n  Reducing Internal Covariate Shift; S. Ioffe, C. Szegedy]\n  (http:\/\/arxiv.org\/abs\/1502.03167).\n  Args:\n    x: Input `Tensor` of arbitrary dimensionality.\n    mean: A mean `Tensor`.\n    variance: A variance `Tensor`.\n    offset: An offset `Tensor`, often denoted \\\\(\\beta\\\\) in equations, or\n      None. If present, will be added to the normalized tensor.\n    scale: A scale `Tensor`, often denoted \\\\(\\gamma\\\\) in equations, or\n      `None`. If present, the scale is applied to the normalized tensor.\n    variance_epsilon: A small float number to avoid dividing by 0.\n    name: A name for this operation (optional).\n  Returns:\n    the normalized, scaled, offset tensor.\n  \"\"\"\n  with ops.name_scope(name, \"batchnorm\", [x, mean, variance, scale, offset]):\n    inv = math_ops.rsqrt(variance + variance_epsilon)\n    if scale is not None:\n      inv *= scale\n    # Note: tensorflow\/contrib\/quantize\/python\/fold_batch_norms.py depends on\n    # the precise order of ops that are generated by the expression below.\n    return x * math_ops.cast(inv, x.dtype) + math_ops.cast(\n        offset - mean * inv if offset is not None else -mean * inv, x.dtype)\n<\/code><\/pre>\n\n<p>I didn't understand how it works and could't convert to julia code. I'm using the same order for tensor shapes. If x has shape <code>[a,b,c]<\/code> in tf, it has <code>[a,b,c]<\/code> shape in my julia code aswell.<\/p>\n","answers":[{"AnswerId":"57994294","CreationDate":"2019-09-18T13:48:13.997","ParentId":null,"OwnerUserId":"4717097","Title":null,"Body":"<p><code>Flux<\/code> provides batch normalization and layer normalization (and some others) in Julia (called <code>LayerNorm<\/code> and <code>BatchNorm<\/code>. In your question you don't specify what library you use in Julia, but if you use something else than <code>Flux<\/code>, then the code might be helpful for your own implementation. See: <a href=\"https:\/\/github.com\/FluxML\/Flux.jl\/blob\/master\/src\/layers\/normalise.jl\" rel=\"nofollow noreferrer\">https:\/\/github.com\/FluxML\/Flux.jl\/blob\/master\/src\/layers\/normalise.jl<\/a> <\/p>\n\n<p>If you are using <code>Knet<\/code> (which I never did) there is at least batch normalization already implemented, I am not sure about other normalization layers. They might be there as well. <\/p>\n\n<p>Please also post what you already tried so for in Julia. <\/p>\n"}]}
+{"QuestionId":57964208,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-16T20:48:03.657","AcceptedAnswerId":57972368.0,"OwnerUserId":2082026.0,"Title":"Padded batches does not handle the last batch of the data with smaller size than batch size","Body":"<p>I have a dataset where I have images and per image I have 10 questions and thus 10 answers. I have successfully trained and checkpointed the model. The model consists of two inputs, one input as image to a CNN and the other input as question to the LSTM. So for every image I would feed in 10 questions. The results of both would then be concatenated and input to the FC layers.<\/p>\n\n<p>Considering I have a batch size of 64, I would feed in 64 images and 640 questions. At the time of concatenation I would need to have the dimension at <code>axis:0<\/code> to be equal, to avoid concatenation error due to different dimensions. Thus, I flatten my CNN network output and repeat it 10 times and then concatenate it to the LSTM output.<\/p>\n\n<p>Inside the <strong>model_rn.py<\/strong>, I do following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class Model:\n    def __init__(self):\n        self.img = tf.placeholder(\n            name='img',\n            dtype=tf.float32,\n            shape=[self.batch_size, self.img_size, self.img_size, 3]\n        )\n        self.q = tf.placeholder(\n            name='ques',\n            dtype=tf.float32,\n            shape=[self.batch_size * 10, self.ques_dim]\n        )\n        self.ans = tf.placeholder(\n            name='ans',\n            dtype=tf.float32,\n            shape=[self.batch_size * 10, self.ans_dim]\n        )\n        # and some more class variables\n        self.build()\n\n    def build(self):\n        def cnn(img, q, scope):\n            # some Conv2D and BatchNormalization\n            flat = Flatten(name='flatten')(bn_4)  # layer where data is flattened before concatenate\n            flat = tf.keras.backend.repeat_elements(flat, 10, axis=0)  # repeat 10 times\n        # some statements to feed data into LSTM and CNN\n<\/code><\/pre>\n\n<p>Then I load my model and try to run it on the test dataset which contains 20 images, 200 questions and thus 200 answers. But then I get the error:<\/p>\n\n<blockquote>\n  <p>ValueError: Cannot feed value of shape (20, 128, 128, 3) for Tensor 'img_1:0', which has shape '(640, 128, 128, 3)'<\/p>\n<\/blockquote>\n\n<p>To feed the batches from the test dataset I have used <code>padded_batch<\/code>.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>dataset_img = Dataset.from_tensor_slices((images)).padded_batch(\n    64, padded_shapes=(128, 128, 3)\n)\ndataset_ques = Dataset.from_tensor_slices((questions)).padded_batch(\n    64 * 10, padded_shapes=(14)\n)\ndataset_ans = Dataset.from_tensor_slices((answers)).padded_batch(\n    64 * 10, padded_shapes=(22)\n)\n<\/code><\/pre>\n\n<p>Can someone help me with this?<\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"57972368","CreationDate":"2019-09-17T10:30:14.113","ParentId":null,"OwnerUserId":"2851685","Title":null,"Body":"<p>you have constrained your placeholder to take exactly \"batch_size\" number of rows. To take any rows you can create placeholder like this<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>self.img = tf.placeholder(\n            name='img',\n            dtype=tf.float32,\n            shape=[None, self.img_size, self.img_size, 3]\n        )\n<\/code><\/pre>\n\n<p>similarly for self.q and self.ans<\/p>\n"}]}
+{"QuestionId":57965025,"AnswerCount":0,"Tags":"<tensorflow><keras><neural-network><conv-neural-network>","CreationDate":"2019-09-16T22:14:20.823","AcceptedAnswerId":null,"OwnerUserId":4173867.0,"Title":"Validation loss is not decreasing less than 0.4 in all architectures","Body":"<p>For research, I'm working on face anti-spoofing. I have the rose-youtu dataset which spec could be found <a href=\"http:\/\/rose1.ntu.edu.sg\/datasets\/faceLivenessDetection.asp\" rel=\"nofollow noreferrer\">here<\/a> in details. The problem is no matter what architecture I use for the Model, the validation loss would not be less than ~0.5, Although I tried different architectures (3D CNN, 2D CNN, Fine-tuning, Inception) with different combinations of regularization (Dropout, L1, L2, L1_L2) but the thing is the val_loss will end up at ~0.5 with and when I do evaluation on the test set I get ~0.7 loss and ~0.85% accuracy. Note that there is no overfitting happening, both loss and val_loss are close and converging until they reach ~0.3 which causes the loss to decrease and val_loss fluctuating in the range of ~0.4 0.6.<\/p>\n\n<p>What is that I'm not considering? Does it mean that the dataset could not be improved more than this?<\/p>\n","answers":[]}
+{"QuestionId":57965554,"AnswerCount":1,"Tags":"<python><tensorflow><object-detection-api>","CreationDate":"2019-09-16T23:32:57.150","AcceptedAnswerId":57982884.0,"OwnerUserId":12076599.0,"Title":"Invalid argument: Input to reshape is a tensor with x values, but requested shape requires a multiple of y. {node Reshape_13}","Body":"<p>I am using tensorflow's object detection api with faster_rcnn_resnet101 and get the following error when trying to train: <\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors_impl.InvalidArgumentError: 2 root error(s) found.\n   (0) Invalid argument: Input to reshape is a tensor with 36 values, but the requested shape requires a multiple of 16<\/p>\n  \n  <p>[[{{node Reshape_13}}]]<\/p>\n  \n  <p>[[IteratorGetNext]]<\/p>\n  \n  <p>[[IteratorGetNext\/_7243]]<\/p>\n  \n  <p>(1) Invalid argument: Input to reshape is a tensor with 36 values, but the requested shape requires a multiple of 16<\/p>\n  \n  <p>[[{{node Reshape_13}}]]<\/p>\n  \n  <p>[[IteratorGetNext]]<\/p>\n  \n  <p>0 successful operations.\n  0 derived errors ignored.<\/p>\n<\/blockquote>\n\n<p>I am using a slightly modified version of the pets-train.sh file to run the training (only paths have been altered).  I am trying to train on tf.record files containing jpg images of size (1280, 720) and have made no changes to the network architecture (I have confirmed that all images in the record are of this size).  <\/p>\n\n<p>Curiously, I can successfully run inference on these images when I do something equivalent to what's in the tutorial file detect_pets.py.  This makes me think something is wrong with the way that I've created the tf.record files (code below) rather than anything to do with the shape of the images, despite the error having to do with reshape.  However,I've successfully trained on tf.records created in the same way before (from images of size (600, 600), (1024, 1024), and (720, 480), all with the same network).  Moreover, I've previously encountered a similar error (only the numbers were different but the error was still with node Reshape_13) on a different data set of images with size (600, 600).  <\/p>\n\n<p>I am using python 3.7, tf version 1.14.0, cuda 10.2, Ubuntu 18.04<\/p>\n\n<p>I've looked extensively at various other posts (<a href=\"https:\/\/stackoverflow.com\/questions\/46465925\/input-to-reshape-is-a-tensor-with-37632-values-but-the-requested-shape-has-1505\/52240509#52240509\">here<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/50552806\/tensorflow-image-segmentation-convnet-invalidargumenterror-input-to-reshape-is?rq=1\">here<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/40955223\/tensorflow-python-framework-errors-invalidargumenterror-input-to-reshape-is-a-t?rq=1\">here<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/38397258\/error-tensorflow-cnn-dimension\">here<\/a>, and <a href=\"https:\/\/stackoverflow.com\/questions\/43010339\/python-tensorflow-input-to-reshape-is-a-tensor-with-92416-values-but-the-re\">here<\/a>) but I wasn't able to make any progress.  <\/p>\n\n<p>I've tried adjusting the keep_aspect_ratio_resizer parameters (originally min_dimension=600, max_dimension=1024 but I've also tried min, max = (720, 1280) and have tried pad_to_max_dimension: true with both of these min\/max choices as well).  <\/p>\n\n<p>This is the code I'm using to create the tf.record file (apologies or indentations being off here):<\/p>\n\n<pre><code>def make_example(imfile, boxes):\nwith tf.gfile.GFile(imfile, \"rb\") as fid:\n    encoded_jpg = fid.read()\n\nencoded_jpg_io = io.BytesIO(encoded_jpg)\nimage = PIL.Image.open(encoded_jpg_io)\nif image.format != \"JPEG\":\n    raise Exception(\"Images need to be in JPG format\")\n\nheight = image.height\nwidth = image.width\n\nxmins = []\nxmaxs = []\nymins = []\nymaxs = []\nfor box in boxes:\n    xc, yc, w, h = box\n    xmin = xc - w \/ 2\n    xmax = xc + w \/ 2\n    ymin = yc - h \/ 2\n    ymax = yc + h \/ 2\n\n    new_xmin = np.clip(xmin, 0, width - 1)\n    new_xmax = np.clip(xmax, 0, width - 1)\n    new_ymin = np.clip(ymin, 0, height - 1)\n    new_ymax = np.clip(ymax, 0, height - 1)\n\n    area = (ymax - ymin) * (xmax - xmin)\n    new_area = (new_ymax - new_ymin) * (new_xmax - new_xmin)\n    if new_area &gt; 0.3 * area:\n        xmins.append(new_xmin \/ width)\n        xmaxs.append(new_xmax \/ width)\n        ymins.append(new_ymin \/ height)\n        ymaxs.append(new_ymax \/ height)\n\nclasses_text = [\"vehicle\".encode(\"utf8\")] * len(boxes)\nclasses = [1] * len(boxes)\nabs_imfile = os.path.abspath(imfile)\ndifficult = [0] * len(boxes)\n\nexample = tf.train.Example(\n    features=tf.train.Features(\n        feature={\n            \"image\/height\": int64_feature(height),\n            \"image\/width\": int64_feature(width),\n            \"image\/filename\": bytes_feature(imfile.encode(\"utf8\")),\n            \"image\/source_id\": bytes_feature(abs_imfile.encode(\"utf8\")),\n            \"image\/encoded\": bytes_feature(encoded_jpg),\n            \"image\/format\": bytes_feature(\"jpeg\".encode(\"utf8\")),\n            \"image\/object\/bbox\/xmin\": float_list_feature(xmins),\n            \"image\/object\/bbox\/xmax\": float_list_feature(xmaxs),\n            \"image\/object\/bbox\/ymin\": float_list_feature(ymins),\n            \"image\/object\/bbox\/ymax\": float_list_feature(ymaxs),\n            \"image\/object\/class\/text\": bytes_list_feature(classes_text),\n            \"image\/object\/class\/label\": int64_list_feature(classes),\n            \"image\/object\/difficult\": int64_list_feature(difficult),\n        }\n    )\n)\nreturn example\n\n\ndef make_tfrecord(outfile, imfiles, truthfiles):\nwriter = tf.python_io.TFRecordWriter(outfile)\n\nfor imfile, truthfile in zip(imfiles, truthfiles):\n    print(imfile)\n    boxes = pd.read_csv(truthfile)\n    if boxes.empty:\n        boxes = []\n    else:\n        boxes = [\n            (box.Xc, box.Yc, box.Width, box.Height) for box in boxes.itertuples()\n        ]\n\n    example = make_example(imfile, boxes)\n    writer.write(example.SerializeToString())\nwriter.close()\n\ndef make_combined_train_dset(names):\nimfiles = []\ntruthfiles = []\ntraindir = os.path.join(tf_datadir, \"train\")\nvaldir = os.path.join(tf_datadir, \"val\")\n\nfor name in names:\n    imdir = os.path.join(processed_datadir, name, \"images\")\n    truthdir = os.path.join(processed_datadir, name, \"truth\")\n\n    imfiles.extend(sorted(glob.glob(os.path.join(imdir, \"*.jpg\"))))\n    truthfiles.extend(sorted(glob.glob(os.path.join(truthdir, \"*.csv\"))))\n\ninds = list(range(len(imfiles)))\nnp.random.shuffle(inds)\nimfiles = [imfiles[i] for i in inds]\ntruthfiles = [truthfiles[i] for i in inds]\n\nntrain = round(0.9 * len(imfiles))\ntrain_imfiles = imfiles[:ntrain]\ntrain_truthfiles = truthfiles[:ntrain]\nval_imfiles = imfiles[ntrain:]\nval_truthfiles = truthfiles[ntrain:]\n\nchunksize = 1500\n\nfor d in [traindir, valdir]:\n    if not os.path.exists(d):\n        os.mkdir(d)\n\nfor i in range(0, len(train_imfiles), chunksize):\n    print(f\"{i} \/ {len(train_imfiles)}\", end=\"\\r\")\n    cur_imfiles = train_imfiles[i : i + chunksize]\n    cur_truthfiles = train_truthfiles[i : i + chunksize]\n    testfile = os.path.join(traindir, f\"{i}.tfrecord\")\n    make_tfrecord(testfile, cur_imfiles, cur_truthfiles)\n\nfor i in range(0, len(val_imfiles), chunksize):\n    print(f\"{i} \/ {len(val_imfiles)}\", end=\"\\r\")\n    cur_imfiles = val_imfiles[i : i + chunksize]\n    cur_truthfiles = val_truthfiles[i : i + chunksize]\n    testfile = os.path.join(valdir, f\"{i}.tfrecord\")\n    make_tfrecord(testfile, cur_imfiles, cur_truthfiles)\n\ndef make_train_dset(name, train_inc=1, val_inc=1, test_inc=1):\ntrainfile = os.path.join(tf_datadir, name + \"-train.tfrecord\")\nvalfile = os.path.join(tf_datadir, name + \"-val.tfrecord\")\n\nimdir = os.path.join(processed_datadir, name, \"images\")\ntruthdir = os.path.join(processed_datadir, name, \"truth\")\n\nimfiles = sorted(glob.glob(os.path.join(imdir, \"*.jpg\")))\ntruthfiles = sorted(glob.glob(os.path.join(truthdir, \"*.csv\")))\n\nn = len(imfiles)\nntrain = round(0.9 * n)\n\nprint(trainfile)\nmake_tfrecord(trainfile, imfiles[:ntrain:train_inc], truthfiles[:ntrain:train_inc])\nprint(valfile)\nmake_tfrecord(valfile, imfiles[ntrain::val_inc], truthfiles[ntrain::val_inc])\n<\/code><\/pre>\n\n<p>With other datasets, I've been able to just create a tf.record using the function make_combined_train_dset (or make_train_dset) defined above and then provide the path to these datasets in the faster_rcnn_resnet101.config file and then training proceeded normally (just as in the tutorial example).  With this new dataset (and at least one other dataset), I'm getting the aforementioned reshape error.  Nonetheless, I can still run inference on the images in this dataset so it makes me think that the problem is with the tf records or with how they are being read in rather than an intrinsic problem with the images or their size.  <\/p>\n\n<p>Any assistance anyone can offer would be greatly appreciated as I've struggled with this for several days now.<\/p>\n","answers":[{"AnswerId":"57982884","CreationDate":"2019-09-17T22:41:43.703","ParentId":null,"OwnerUserId":"12076599","Title":null,"Body":"<p>I'm an idiot: confirmed.<\/p>\n\n<p>The problem was that classes_text, classes, and difficult were the wrong length.<\/p>\n\n<p>Replaced<\/p>\n\n<pre><code>classes_text = [\"vehicle\".encode(\"utf8\")] * len(boxes)\nclasses = [1] * len(boxes)\ndifficult = [0] * len(boxes)\n<\/code><\/pre>\n\n<p>with <\/p>\n\n<pre><code>classes_text = [\"vehicle\".encode(\"utf8\")] * len(xmins)\nclasses = [1] * len(xmins\ndifficult = [0] * len(xmins)\n<\/code><\/pre>\n\n<p>and it runs fine.  Posting this in case anyone else struggles with a similar issue.<\/p>\n\n<p>Thank you to anyone who put time or thought into my question.  Hope this helps someone not waste time.<\/p>\n"}]}
+{"QuestionId":57965609,"AnswerCount":0,"Tags":"<tensorflow><keras><deep-learning>","CreationDate":"2019-09-16T23:41:36.433","AcceptedAnswerId":null,"OwnerUserId":12076679.0,"Title":"Different results from Tensorflow and Keras","Body":"<p>I get different results from Tensorflow and Keras with the same network structure. <\/p>\n\n<p>The loss function looks like<\/p>\n\n<pre><code>class MaskedMultiCrossEntropy(object):\n    def loss(self, y_true, y_pred):\n        vec = tf.nn.softmax_cross_entropy_with_logits(logits=y_pred, labels=y_true, dim=1)\n        mask = tf.equal(y_true[:,0,:], -1)\n        zer = tf.zeros_like(vec)\n        loss = tf.where(mask, x=zer, y=vec)\n        return loss\n<\/code><\/pre>\n\n<p>The network layer I used is called CrowdsClassification, which is implemented by Keras. Then I build the network by<\/p>\n\n<pre><code>x = Dense(128, input_shape=(input_dim,), activation='relu')(inputs)\nx = Dropout(0.5)(x)\nx = Dense(N_CLASSES)(x)\nx = Activation(\"softmax\")(x)\ncrowd = CrowdsClassification(num_classes, num_oracles, conn_type=\"MW\")\nx = crowd(x)\n<\/code><\/pre>\n\n<p>Train the model with Keras<\/p>\n\n<pre><code>model = Model(inputs=inputs, outputs=x)\nmodel.compile(optimizer='adam', loss=loss)\nmodel.fit(inputs,\n           true_class, epochs=100, shuffle=False, verbose=2, validation_split=0.1))\n<\/code><\/pre>\n\n<p>Train the model with tensorflow<\/p>\n\n<pre><code>optimizer = tf.train.AdamOptimizer(learning_rate=0.01, beta1=0.9, beta2=0.999)\nopt_op = optimizer.minimize(loss, global_step=global_step)\nsess.run(tf.global_variables_initializer())\nfor epoch in range(100):\n    sess.run([loss, opt_op], feed_dict=train_feed_dict)\n<\/code><\/pre>\n\n<p>The Tensorflow will get a wrong prediction. It seems that the issue comes from the loss function, that Tensorflow cannot backproporgate the masked loss. Anyone can give some advices? Thx a lot.<\/p>\n","answers":[]}
+{"QuestionId":57965732,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2019-09-17T00:03:10.520","AcceptedAnswerId":58403997.0,"OwnerUserId":7212365.0,"Title":"Why are gradients incorrect for categorical crossentropy?","Body":"<p>After answering <a href=\"https:\/\/stackoverflow.com\/questions\/57892492\/why-the-gradient-of-categorical-crossentropy-loss-with-respect-to-logits-is-0-wi\/57894760#57894760\">this question<\/a>, there are some interesting but confused findings I met in tensorflow 2.0. The gradients of <code>logits<\/code> looks incorrect to me. Let's say we have <code>logits<\/code> and <code>labels<\/code> here. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>logits = tf.Variable([[0.8, 0.1, 0.1]], dtype=tf.float32)\nlabels = tf.constant([[1, 0, 0]],dtype=tf.float32)\n\nwith tf.GradientTape(persistent=True) as tape:\n    loss = tf.reduce_sum(tf.keras.losses.categorical_crossentropy(labels, logits, \n                                                                  from_logits=False))\ngrads = tape.gradient(loss, logits)\nprint(grads)\n<\/code><\/pre>\n\n<p>Since <code>logits<\/code> is already a prob distribution, so I set <code>from_logits=False<\/code> in the loss function. <\/p>\n\n<p>I thought tensorflow will use <code>loss=-\\Sigma_i(p_i)\\log(q_i)<\/code> to calculate the loss, and if we derive on <code>q_i<\/code>, we will have the derivative be <code>-p_i\/q_i<\/code>. So, the expected grads should be [-1.25,0,0]. However, tensorflow will return [-0.25,1,1]. <\/p>\n\n<p>After reading the source code of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/f03fe1bf795061408fafda9704c485aea7bddde2\/tensorflow\/python\/keras\/backend.py#L4351\" rel=\"nofollow noreferrer\"><code>tf.categorical_crossentropy<\/code><\/a>, I found that even though we set <code>from_logits=False<\/code>, it still normalize the probabilities. That will change the final gradient expression. Specifically, the gradient will be <code>-p_i\/q_i+p_i\/sum_j(q_j)<\/code>. If <code>p_i=1<\/code> and <code>sum_j(q_j)=1<\/code>, the final gradient will plus one. That's why the gradient will be -0.25, however, I haven't figured out why the last two gradients would be 1.<\/p>\n\n<p>To prove that all gradients are increased by <code>1\/sum_j(q_j)<\/code>, I made up a logits, which is not prob distribution, and set <code>from_logits=False<\/code> still. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>logits = tf.Variable([[0.5, 0.1, 0.1]], dtype=tf.float32)\nlabels = tf.constant([[1, 0, 0]],dtype=tf.float32)\n\nwith tf.GradientTape(persistent=True) as tape:\n    loss = tf.reduce_sum(tf.keras.losses.categorical_crossentropy(labels, logits,\n                                                                  from_logits=False))\ngrads = tape.gradient(loss, logits)\nprint(grads)\n<\/code><\/pre>\n\n<p>The grads returned by tensorflow is <code>[-0.57142866,1.4285713,1.4285713 ]<\/code>, which I thought should be <code>[-2,0,0]<\/code>.<\/p>\n\n<p>It shows that all gradients are increased by <code>1\/(0.5+0.1+0.1)<\/code>. For the <code>p_i==1<\/code>, the gradient increased by <code>1\/(0.5+0.1+0.1)<\/code> makes sense to me. But I don't understand why <code>p_i==0<\/code>, the gradient is still increased by <code>1\/(0.5+0.1+0.1)<\/code>.<\/p>\n\n<p><strong>Update<\/strong><\/p>\n\n<p>Thanks for @OverLordGoldDragon's kind reminder. After normalizing the probs, the correct gradients formula should be <code>-p_i\/q_i+1\/sum_j(q_j)<\/code>. So the behaviors in the question are expected. <\/p>\n","answers":[{"AnswerId":"58403997","CreationDate":"2019-10-15T23:56:50.807","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>Categorical crossentropy is tricky, particularly w.r.t. one-hot encodings; the problem arises out of presuming that some predictions are \"tossed out\" in computing loss or gradient, when looking at how loss is computed: <\/p>\n\n<p><code>loss = f(labels * preds) = f([1, 0, 0] * preds)<\/code><\/p>\n\n<p><strong>Why are the gradients incorrect?<\/strong> Above may suggest that <code>preds[1:]<\/code> don't matter, but note that this isn't actually <em>preds<\/em> - it's <code>preds_normalized<\/code>, which involves single element of <code>preds<\/code>. To get a better idea of what's happening, the <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/985521ee7050df39f9c06f53b54e17927bd1e6ea\/keras\/backend\/numpy_backend.py#L333\" rel=\"nofollow noreferrer\">Numpy backend<\/a> is helpful; assuming <code>from_logits=False<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>losses = []\nfor label, pred in zip(labels, preds):\n    pred_norm = pred \/ pred.sum(axis=-1, keepdims=True)\n    losses.append(np.sum(label * -np.log(pred_norm), axis=-1, keepdims=False))\n<\/code><\/pre>\n\n<p>A more complete explanation of above - <a href=\"https:\/\/stackoverflow.com\/questions\/58159154\/how-to-calculate-categorical-cross-entropy-by-hand\/58160633#58160633\">here<\/a>. Below is my derivation of the gradients formula, with examples comparing its Numpy implementation with <code>tf.GradientTape<\/code> results. To skip the meaty details, scroll to \"Main idea\".<\/p>\n\n<hr>\n\n<p><strong>Formula + Derivation<\/strong>: proof of correctness at the bottom.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\"\"\"\ngrad = -y * sum(p_zeros) \/ (p_one * sum(pred)) + p_mask \/ sum(pred)\n\np_mask  = abs(y - 1)\np_zeros = p_mask * pred\n\ny = label:      1D array of length N, one-hot \np = prediction: 1D array of length N, float32 from 0 to 1\np_norm = normalized predictions\np_mask = prediction masks (see below)\n\"\"\"\n<\/code><\/pre>\n\n<p><em>What's happening?<\/em> Begin with a simple example to understand what <code>tf.GradientTape<\/code> is doing:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>w = tf.Variable([0.5, 0.1, 0.1])\n\nwith tf.GradientTape(persistent=True) as tape:\n    f1 = w[0] + w[1]  # f = function\n    f2 = w[0] \/ w[1]\n    f3 = w[0] \/ (w[0] + w[1] + w[2])\nprint(tape.gradient(f1, w))  # [1.    1.  0.]\nprint(tape.gradient(f2, w))  # [10. -50.  0.]\nprint(tape.gradient(f3, w))  # [0.40816 -1.02040 -1.02040]\n<\/code><\/pre>\n\n<p>Let <code>w = [w1, w2, w3]<\/code>. Then:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\"\"\"\ngrad = [df1\/dw1, df1\/dw2, df1\/dw3]\n\ngrad1 = [d(w1 + w2)\/w1, d(w1 + w2)\/w2, d(w1 + w2)\/w3] = [1, 1, 0]\ngrad2 = [d(w1 \/ w2)\/w1, d(w1 \/ w2)\/w2, d(w1 + w2)\/w3] = [1\/w2, -w1\/w2^2, 0] = [10, -50, 0]\ngrad3 = [(w1 + w2)\/K, - w2\/K, -w3\/K] = [0.40816 -1.02040 -1.02040] -- K = (w1 + w2 + w3)^2\n\"\"\"\n<\/code><\/pre>\n\n<p>In other words, <code>tf.GradientTape<\/code> treats <em>each element<\/em> of the input tensor its differentiating against as a <em>variable<\/em>. This in mind, it suffices to implement categorical crossentropy via elementary <code>tf<\/code>functions then derive its derivative by hand and see if they agree. It's what I've done at the bottom code, with loss better explained in answer linked above.<\/p>\n\n<hr>\n\n<p><strong>Formula explanation<\/strong>: <\/p>\n\n<p><code>f3<\/code> above is the most insightful, as it's actually <code>pred_norm<\/code>; all we need now is to add a natural log, and handle two separate cases: grads for <code>y==1<\/code>, and for <code>y==0<\/code>; with a handy Wolf, derivatives can be computed in a flash. Adding more variables to denominator, we can see the following pattern:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.wolframalpha.com\/input\/?i=d%2Fdx+ln%28x+%2F+%28x+%2B+y+%2B+z%29%29\" rel=\"nofollow noreferrer\"><code>d(loss)\/d(p_one)<\/code><\/a> <code>= p_zeros \/ (p_one * sum(pred))<\/code><\/li>\n<li><a href=\"https:\/\/www.wolframalpha.com\/input\/?i=d%2Fdx+ln%28y+%2F+%28x+%2B+y+%2B+z%29%29\" rel=\"nofollow noreferrer\"><code>d(loss)\/d(p_non_one)<\/code><\/a> <code>= -1 \/ sum(pred)<\/code><\/li>\n<\/ul>\n\n<p>where <code>p_one<\/code> is <code>pred<\/code> where <code>label == 1<\/code>, <code>p_non_one<\/code> is any other <code>pred<\/code> element, and <code>p_zeros<\/code> is all <code>pred<\/code> elements except <code>p_one<\/code>. The code at bottom is simply an implementation of exactly this, using compact syntax.<\/p>\n\n<hr>\n\n<p><strong>Explanation example<\/strong>:<\/p>\n\n<p>Suppose <code>label = [1, 0, 0]; pred = [.5, .1, .1]<\/code>. Below is <code>numpy_gradient<\/code>, step-by-step:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>p_mask == [0, 1, 1]  # effectively `label` \"inverted\", to exclude `p_one`\np_one  == .5         # pred where `label` == 1\n\n## grad_zeros\np_mask \/ np.sum(pred) == [0, 1, 1] \/ (.5 + .1 + .1) = [0, 1\/.7, 1\/.7]\n\n## grad_one\np_one  * np.sum(pred) == .5 * (.5 + .1 + .1) = .5 * .7 = .35\np_mask * pred         == [0, 1, 1] * [.5, .1, .1] = [0, .1, .1]\nnp.sum(p_mask * pred) == .2\nlabel * np.sum(p_mask * pred) == .2 * [1, 0, 0] = [.2, 0, 0]\n\nlabel * np.sum(p_mask * pred) \/ (p_one * np.sum(pred)) \n== [.2, 0, 0] \/ .35 = 0.57142854\n<\/code><\/pre>\n\n<p>Per above, we can see that the gradient is effectively divided into two computations: <code>grad_one<\/code>, and <code>grad_zeros<\/code>.<\/p>\n\n<hr>\n\n<p><strong>Main idea<\/strong>: understandably, that's a lot of detail, so here's the main idea: <em>every element<\/em> of <code>label<\/code> and <code>pred<\/code> affects <code>grad<\/code>, and loss is computed using <code>pred_norm<\/code>, not <code>pred<\/code>, and the <strong>normalization step is backpropagated<\/strong>. We can run a little visual to confirm this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>labels = tf.constant([[1, 0, 0]],dtype=tf.float32)\ngrads = []\nfor i in np.linspace(0, 1, 100):\n    logits = tf.Variable([[0.5, 0.1, i]], dtype=tf.float32)\n    with tf.GradientTape(persistent=True) as tape:\n        loss = tf.keras.losses.categorical_crossentropy(\n              labels, logits, from_logits=False)  \n    grads.append(tape.gradient(loss, logits))\ngrads = np.vstack(grads)\nplt.plot(grads)\n<\/code><\/pre>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/PPTrW.png\" width=\"450\"><\/p>\n\n<p><em>Even though<\/em> only <code>logits[2]<\/code> is varied, <code>grads[1]<\/code> varies <em>exactly the same<\/em>. The explanation's clear from <code>grad_zeros<\/code> above, but more intuitively, categorical crossentropy <em>doesn't care<\/em> \"how wrong\" the zero-label predictions are individually, only <em>collectively<\/em> - because it only semi-directly computes loss from <code>pred[0]<\/code> (i.e. <code>pred[0] \/ sum(pred)<\/code>), which is normalized by all other <code>pred<\/code>. So whether <code>pred[1] == .9<\/code> and <code>pred[2] == .2<\/code> or vice versa, <code>p_norm<\/code> is exactly the same.<\/p>\n\n<hr>\n\n<p><strong>Closing note<\/strong>: derived formulas are intended for a 1D case for simplicity, and may not work for N-dimensional <code>labels<\/code> and <code>preds<\/code> tensors, but can be easily generalized.<\/p>\n\n<hr>\n\n<p><strong>Numpy vs. tf.GradientTape<\/strong>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def numpy_gradient(label, pred):\n    p_mask = np.abs(label - 1)\n    p_one  = pred[np.where(label==1)[0][0]]\n    return p_mask \/ np.sum(pred) \\\n           - label * np.sum(p_mask * pred) \/ (p_one * np.sum(pred))\n\ndef gtape_gradient(label, pred):\n    pred  = tf.Variable(pred)\n    label = tf.Variable(label)\n\n    with tf.GradientTape() as tape:\n        loss = - tf.math.log(tf.reduce_sum(label * pred) \/ tf.reduce_sum(pred))\n    return tape.gradient(loss, pred).numpy()\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>label = np.array([1.,   0., 0. ])\npred  = np.array([0.5, 0.1, 0.1])\nprint(numpy_gradient(label, pred))\nprint(gtape_gradient(label, pred))\n\n# [-0.57142854  1.4285713   1.4285713 ]  &lt;-- 100% agreement\n# [-0.57142866  1.4285713   1.4285713 ]  &lt;-- 100% agreement\n<\/code><\/pre>\n"}]}
+{"QuestionId":57966054,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-09-17T00:58:10.867","AcceptedAnswerId":57969204.0,"OwnerUserId":2893024.0,"Title":"Do I need to match the representation of my output layer and the target labels in PyTorch Nllloss?","Body":"<p>I'm trying to solve a classification problem with 3 possible outputs: <code>0, 1, or 2<\/code>. \nMy output layer finally outputs a vector of probabilities for each label, say <code>[0.3,0.4,0.3]<\/code><\/p>\n\n<p>My loss function is defined such:<\/p>\n\n<pre><code>loss = criterion(output_batch, label_batch) #criterion = nn.NLLLoss()\n<\/code><\/pre>\n\n<p>Now my question has to do with the outputs and labels not matching in the way the store data. The output is in the form of a size=3 probability vector (adding to 1 using soft max), and my target labels are simple scalars. <\/p>\n\n<p>I can convert my labels to vectors when the loss function is calculated but I'm not sure if this is necessary<\/p>\n\n<pre><code>0 ==&gt; [1,0,0]\n1 ==&gt; [0,1,0]\n2 ==&gt; [0,0,1]\n<\/code><\/pre>\n\n<p>Can someone please shed light on this issue? Thanks! <\/p>\n","answers":[{"AnswerId":"57969204","CreationDate":"2019-09-17T07:20:03.383","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>Let's say your classes are: cat, dog and capibara.<\/p>\n\n<p>You have so called <code>softmax<\/code> predictions.<\/p>\n\n<p>[0.3,0.4,0.3]<\/p>\n\n<p>The <code>softmax<\/code> function is pumping one result at a top. In this case if dog is under 0.4 our output is predicting the dog.<\/p>\n\n<p>Note how the predictions sum to 1 = 0.3+0.4+0.3.<\/p>\n\n<p>Now you need to calculate the log of that which is log softmax, and then NLL is just negative of that.\n<a href=\"https:\/\/i.stack.imgur.com\/OeDen.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/OeDen.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<blockquote>\n  <p>I can convert my labels to vectors when the loss function is calculated but I'm not sure if this is necessary?<\/p>\n<\/blockquote>\n\n<pre><code>0 ==&gt; [1,0,0]\n1 ==&gt; [0,1,0]\n2 ==&gt; [0,0,1]\n<\/code><\/pre>\n\n<p>This is not necessary in your case. This means we had three different estimations (bs=3) while you showed just one.<\/p>\n\n<hr>\n\n<p>Here is a little exercise:<\/p>\n\n<pre><code>batch_size, n_classes = 10, 5\nx = torch.randn(batch_size, n_classes)\nprint(\"x:\",x)\n\ntarget = torch.randint(n_classes, size=(batch_size,), dtype=torch.long)\nprint(\"target:\",target)\n\n\ndef log_softmax(x): \n    return x - x.exp().sum(-1).log().unsqueeze(-1)\n\ndef nll_loss(p, target):\n    return -p[range(target.shape[0]), target].mean()\n\npred = log_softmax(x)\nprint (\"pred:\", pred)\nohe = torch.zeros(batch_size, n_classes)\nohe[range(ohe.shape[0]), target]=1\nprint(\"ohe:\",ohe)\n\npe = pred[range(target.shape[0]), target]\nprint(\"pe:\",pe)\n\nmean = pred[range(target.shape[0]), target].mean()\nprint(\"mean:\",mean)\n\nnegmean = -mean\nprint(\"negmean:\", negmean)\n\nloss = nll_loss(pred, target)\nprint(\"loss:\",loss)\n<\/code><\/pre>\n\n<p>Out:<\/p>\n\n<pre><code>x: tensor([[ 1.5837, -1.3132,  1.5513,  1.4422,  0.8072],\n        [ 1.1740,  1.9250,  0.4258, -1.0320, -0.4650],\n        [-1.2447, -0.5360, -1.4950,  1.2020,  1.2724],\n        [ 0.2300,  0.2587, -0.4463, -0.1397, -0.3617],\n        [-0.7983,  0.7742,  0.0035,  0.9963, -0.7926],\n        [ 0.7575, -0.8008,  0.7995,  0.0448,  0.6621],\n        [-1.7153,  0.7672, -0.6841, -0.4826, -0.8614],\n        [ 0.0263,  0.7244,  0.8751, -1.0226, -1.3762],\n        [ 0.0192, -0.4368, -0.4010, -1.0660,  0.0364],\n        [-0.5120, -1.4871,  0.6758,  1.2975,  0.2879]])\ntarget: tensor([0, 4, 3, 0, 0, 4, 1, 2, 4, 2])\npred: tensor([[-1.2094, -4.1063, -1.2418, -1.3509, -1.9859],\n        [-1.3601, -0.6091, -2.1083, -3.5661, -2.9991],\n        [-3.3233, -2.6146, -3.5736, -0.8766, -0.8063],\n        [-1.3302, -1.3015, -2.0065, -1.7000, -1.9220],\n        [-2.7128, -1.1403, -1.9109, -0.9181, -2.7070],\n        [-1.2955, -2.8538, -1.2535, -2.0081, -1.3909],\n        [-3.0705, -0.5881, -2.0394, -1.8379, -2.2167],\n        [-1.7823, -1.0841, -0.9334, -2.8311, -3.1847],\n        [-1.2936, -1.7496, -1.7138, -2.3788, -1.2764],\n        [-2.5641, -3.5393, -1.3764, -0.7546, -1.7643]])\nohe: tensor([[1., 0., 0., 0., 0.],\n        [0., 0., 0., 0., 1.],\n        [0., 0., 0., 1., 0.],\n        [1., 0., 0., 0., 0.],\n        [1., 0., 0., 0., 0.],\n        [0., 0., 0., 0., 1.],\n        [0., 1., 0., 0., 0.],\n        [0., 0., 1., 0., 0.],\n        [0., 0., 0., 0., 1.],\n        [0., 0., 1., 0., 0.]])\npe: tensor([-1.2094, -2.9991, -0.8766, -1.3302, -2.7128, -1.3909, -0.5881, -0.9334,\n        -1.2764, -1.3764])\nmean: tensor(-1.4693)\nnegmean: tensor(1.4693)\nloss: tensor(1.4693)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57966420,"AnswerCount":1,"Tags":"<r><tensorflow><split>","CreationDate":"2019-09-17T02:02:29.373","AcceptedAnswerId":null,"OwnerUserId":7500268.0,"Title":"why the tensorflow split function report error when X is a tensor","Body":"<p>The tensorflow function split report error and I cannot figure it out<\/p>\n\n<pre class=\"lang-r prettyprint-override\"><code>X = tf$random_uniform(minval=0,\n                      maxval=10,\n                      shape(256, 32),\n                      name = \"X\");\nY = tf$split(X, num_or_size_splits = 2, axis = 0)\n<\/code><\/pre>\n\n<p>the output are expected to be correct but it reports error message:<\/p>\n\n<blockquote>\n  <p>Error in py_call_impl(callable, dots$args, dots$keywords): <\/p>\n  \n  <p>ValueError: \n  Rank-0 tensors are not supported as the num_or_size_splits argument to split. \n  Argument provided: 2.0<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"58884455","CreationDate":"2019-11-15T20:46:52.310","ParentId":null,"OwnerUserId":"5099715","Title":null,"Body":"<p>it works fine in tensorflow 1.14 with eager enabled:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf \ntf.enable_eager_execution()\nX = tf.random_uniform(shape=(256, 32),minval=0,maxval=10,name = \"X\");\nY = tf.split(X, num_or_size_splits = 2, axis = 0)\n<\/code><\/pre>\n\n<p>if you are trying to use a tensor as num_or_size_splits make sure it's a list with elements indicating the size of each split<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf \ntf.enable_eager_execution()\nX = tf.random_uniform(shape=(256, 32),minval=0,maxval=10,name = \"X\");\nn_split = tf.constant([128, 128])\nY= tf.split(X, num_or_size_splits = n_split, axis = 0)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57966951,"AnswerCount":2,"Tags":"<python><image><pytorch>","CreationDate":"2019-09-17T03:27:34.087","AcceptedAnswerId":57967098.0,"OwnerUserId":3983454.0,"Title":"how to use RGB values in feedforward neural network?","Body":"<p>I have data set of colored images in the form of ndarray (100, 20, 20, 3) and 100 corresponding labels. When passing them as input to a fully connected neural network (not CNN), what should I do with the 3 values of RGB? Average them perhaps lose some information, but if not manipulating them, my main issue is batch size, as demo-ed below in pytorch. <\/p>\n\n<pre><code>for epoch in range(n_epochs):\n    for i, (images, labels) in enumerate(train_loader):\n        # because of rgb values, now images is 3 times the length of labels\n        images = Variable(images.view(-1, 400))\n        labels = Variable(labels)\n        optimizer.zero_grad()\n        outputs = net(images)\n        loss = criterion(outputs, labels)\n        loss.backward()\n        optimizer.step()\n<\/code><\/pre>\n\n<p>This returns 'ValueError: Expected input batch_size (300) to match target batch_size (100).' Should I have reshaped images into (1, 1200) dimension tensors? Thanks in advance for answers.<\/p>\n","answers":[{"AnswerId":"57967087","CreationDate":"2019-09-17T03:49:00.997","ParentId":null,"OwnerUserId":"12013459","Title":null,"Body":"<p>Use Grey scaled image to reduce the batch size<\/p>\n"},{"AnswerId":"57967098","CreationDate":"2019-09-17T03:50:35.150","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>Since size of labels is <code>(100,)<\/code>, so your batch data should be with shape of <code>(100, H, W, C)<\/code>. I'm assuming your data loader is returning a tensor with shape of <code>(100,20,20,3)<\/code>. The error happens because you reshape the tensor to <code>(300,400)<\/code>. <\/p>\n\n<ol>\n<li><p>Check your network architecture whether the input tensor shape is <code>(20,20,3)<\/code>. <\/p><\/li>\n<li><p>If your network can only accept single channel images, you can first convert your RGB to grayscale images. <\/p><\/li>\n<li><p>Or, modify your network architecture to make it accept 3 channels images. One convenient way is adding an extra layer reducing 3 channels to 1 channel, and you do not need to change the other parts of the network.  <\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":57967466,"AnswerCount":1,"Tags":"<android><tensorflow><tensorflow-lite>","CreationDate":"2019-09-17T04:42:09.667","AcceptedAnswerId":57986942.0,"OwnerUserId":11955242.0,"Title":"Fail to use custom model in tensorflow lite object detection android app","Body":"<p>I trained model using Google AutoMl then produce tensorflow lite model to detect plastic bottle etc. I would like to use in the tensorflowlite object detection android example but failed <\/p>\n\n<p>this the github that i refer : <a href=\"https:\/\/github.com\/tensorflow\/examples\/tree\/master\/lite\/examples\/object_detection\/android\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/examples\/tree\/master\/lite\/examples\/object_detection\/android<\/a><\/p>\n\n<p>i replaced the tflite file and the txt file with my own project, the installation on android studio its work well but the apps its crashed and didnt work  <\/p>\n\n<pre><code>public class DetectorActivity extends CameraActivity implements OnImageAvailableListener {\n  private static final Logger LOGGER = new Logger();\n\n  \/\/ Configuration values for the prepackaged SSD model.\n  private static final int TF_OD_API_INPUT_SIZE = 300;\n  private static final boolean TF_OD_API_IS_QUANTIZED = true;\n  private static final String TF_OD_API_MODEL_FILE = \"swai.tflite\";\n  private static final String TF_OD_API_LABELS_FILE = \"file:\/\/\/android_asset\/swai.txt\";\n  private static final DetectorMode MODE = DetectorMode.TF_OD_API;\n  \/\/ Minimum detection confidence to track a detection.\n  private static final float MINIMUM_CONFIDENCE_TF_OD_API = 0.5f;\n  private static final boolean MAINTAIN_ASPECT = false;\n  private static final Size DESIRED_PREVIEW_SIZE = new Size(640, 480);\n  private static final boolean SAVE_PREVIEW_BITMAP = false;\n  private static final float TEXT_SIZE_DIP = 10;\n  OverlayView trackingOverlay;\n\n<\/code><\/pre>\n\n<p>this the eror that i found when deploying in the virtual devices <\/p>\n\n<pre><code>09-17 13:32:09.283 1599-1856\/? D\/gralloc_ranchu: gralloc_alloc: Creating ashmem region of size 462848\n09-17 13:32:09.325 9980-10000\/org.tensorflow.lite.examples.detection E\/AndroidRuntime: FATAL EXCEPTION: inference\n    Process: org.tensorflow.lite.examples.detection, PID: 9980\n    java.lang.IllegalArgumentException: Cannot convert between a TensorFlowLite buffer with 786432 bytes and a ByteBuffer with 270000 bytes.\n        at org.tensorflow.lite.Tensor.throwIfShapeIsIncompatible(Tensor.java:272)\n        at org.tensorflow.lite.Tensor.throwIfDataIsIncompatible(Tensor.java:249)\n        at org.tensorflow.lite.Tensor.setTo(Tensor.java:110)\n        at org.tensorflow.lite.NativeInterpreterWrapper.run(NativeInterpreterWrapper.java:151)\n        at org.tensorflow.lite.Interpreter.runForMultipleInputsOutputs(Interpreter.java:275)\n        at org.tensorflow.lite.examples.detection.tflite.TFLiteObjectDetectionAPIModel.recognizeImage(TFLiteObjectDetectionAPIModel.java:193)\n        at org.tensorflow.lite.examples.detection.DetectorActivity$2.run(DetectorActivity.java:181)\n        at android.os.Handler.handleCallback(Handler.java:873)\n        at android.os.Handler.dispatchMessage(Handler.java:99)\n        at android.os.Looper.loop(Looper.java:193)\n        at android.os.HandlerThread.run(HandlerThread.java:65)\n\n\n<\/code><\/pre>\n\n<p>I think there is no settings to implement this model work, Any Sugesstion please? Thanks!<\/p>\n","answers":[{"AnswerId":"57986942","CreationDate":"2019-09-18T07:08:20.737","ParentId":null,"OwnerUserId":"11955242","Title":null,"Body":"<p>case close, solved with changing the TF_OD_API_INPUT_SIZE from 300 to 512. 786432 is represent 512x512x3 but my input is only 270000 (300x3003). <\/p>\n"}]}
+{"QuestionId":57967475,"AnswerCount":3,"Tags":"<tensorflow><keras><deep-learning><data-generation>","CreationDate":"2019-09-17T04:44:01.833","AcceptedAnswerId":null,"OwnerUserId":2995747.0,"Title":"How to repeat data with flow_from_directory in Keras","Body":"<p>I am trying to use <strong><code>keras<\/code><\/strong> <strong><em>flow_from_directory<\/em><\/strong> to train a model. But it does not repeat\ndata after the epoch(i.e. when all the data has been iterated). I could not find any \noption to do so either. Below is my code for data generation while training.\nFor example if total images = 70\n               batch_size   = 32\nthen in 1st and 2nd iteration is gives 32 images, but in third it gives 6 images.<\/p>\n\n<pre><code># data generation from directory without labels  \ntrn = datagen.flow_from_directory(os.path.join(BASE, 'train_gen'),\n                                         batch_size=batch_size,\n                                         target_size=(inp_shape[:2]),\n                                         class_mode=None)\nX = trn.next() # getting a batch of data.\n<\/code><\/pre>\n\n<p>I want the data generator to start repeating data after it's exhausted.<\/p>\n\n<p>Actually I am trying to train a GAN, where a batch images are generated from Generator-Model and then it is concatenated with a batch of real images and then passed to Discriminator-Model and GAN-Model to train. I can't figure out how can I use <strong><em>fit_generator<\/em><\/strong> in this, Code is as below:<\/p>\n\n<pre><code>def train(self, inp_shape, batch_size=1, n_epochs=1000):\n    BASE = '\/content\/gdrive\/My Drive\/Dataset\/GAN'\n\n    datagen = ImageDataGenerator(rescale=1.\/255)\n    trn_dist = datagen.flow_from_directory(os.path.join(BASE, 'train_gen'),\n                                                      batch_size=batch_size,\n                                                      target_size=(inp_shape[:2]),\n                                                      seed = 1360000,\n                                                      class_mode=None)\n\n    val_dist = datagen.flow_from_directory(os.path.join(BASE, 'test_gen'),\n                                                      batch_size=batch_size,\n                                                      target_size=(inp_shape[:2]),\n                                                      class_mode=None)\n\n    trn_real = datagen.flow_from_directory(os.path.join(BASE, 'train_real'),\n                                                      batch_size=batch_size,\n                                                      target_size=(inp_shape[:2]),\n                                                      seed = 1360000,\n                                                      class_mode=None)\n\n    for e in range(n_epochs):\n\n      real_images = trn_real.next()\n\n      dist_images = trn_dist.next()\n\n      gen_images = self.generator.predict(dist_images)\n\n      factor = inp_shape[0]\/250\n      gen_res = ndi.zoom(gen_images, (1, factor, factor, 1), order=2)      \n\n      X = np.concatenate([real_images, gen_res])\n\n      y = np.zeros(2*batch_size)\n      y[:batch_size] = 1.\n\n      self.discriminator.trainable = True\n      self.discriminator.fit(X, y, batch, n_epochs)\n\n      self.discriminator.trainable = False\n\n      self.model.fit(gen_res, y[:batch_size])\n      print ('&gt; training --- epoch=%d\/%d' % (e, n_epochs))\n      if e &gt; 0 and e % 2000 == 0:\n        self.model.save('%s\/models\/gan_model_%d_.h5'%(BASE, e))\n<\/code><\/pre>\n\n<p>PS: I am new to Gans please correct me if I am doing something wrong.<\/p>\n","answers":[{"AnswerId":"57967796","CreationDate":"2019-09-17T05:23:50.473","ParentId":null,"OwnerUserId":"7378795","Title":null,"Body":"<p>The <code>flow_from_directory<\/code> method is made to be used with the <code>fit_generator<\/code> function. The <code>fit_generator<\/code> function allows you to specify the number of epochs. <\/p>\n\n<pre><code>model.fit_generator(trn, epochs=epochs)\n<\/code><\/pre>\n\n<p>Where <code>model<\/code> refers to the model object you want to train. Should solve your problem. These functions are well explained in the Keras documentation<\/p>\n"},{"AnswerId":"57969474","CreationDate":"2019-09-17T07:37:38.477","ParentId":null,"OwnerUserId":"9579075","Title":null,"Body":"<p>You can always specify <code>steps_per_epoch<\/code> argument in <code>fit_generator<\/code> method. This will enable you to repeat data when <code>steps_per_epoch<\/code> > total_samples \/\/ batch_size.<\/p>\n"},{"AnswerId":"57970229","CreationDate":"2019-09-17T08:24:58.853","ParentId":null,"OwnerUserId":"4173867","Title":null,"Body":"<p>To shed some light on the problem, First, you need to know the parameters of <code>flow_from_directory<\/code>. <code>batch_size<\/code> determines the number of samples to be loaded for computation and the <code>epoch<\/code> determines the number of times that you what Keras to pass through all your data. In essence, if you set your <code>epoch=2<\/code> and <code>batch_size=32<\/code> it means that Keras will go through all your data twice with splitting your data in <code>mini-batches<\/code> with 32 samples of your data. then what's missing in your code is essentially the epoch parameter.\nI recommend setting the steps_per_epoch and validation_data as well. the <code>steps_per_epoch<\/code> determined the number of batches in each epoch than to visit all your samples in each epoch set the <code>steps_per_epoch<\/code> as follows.<\/p>\n\n<pre><code>model.fit_generator(train_generator, steps_per_epoch=train_generator.samples\/train_generator.batch_size, epochs=10, validation_data=validation_generator, validation_steps=validation_generator.samples\/validation_generator.batch_size)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57967668,"AnswerCount":0,"Tags":"<python><numpy><pytorch>","CreationDate":"2019-09-17T05:08:23.043","AcceptedAnswerId":null,"OwnerUserId":12077478.0,"Title":"Converting matrix inconsistent shape of np.array to torch.tensor?","Body":"<p>For example, I have an np array containing elements with np. array<\/p>\n\n<pre><code> (5, 10)\n(10, 10)\n(32, 10)\n(2, 10)\n(48, 10)\n(72, 10)\n(42, 10)\n<\/code><\/pre>\n\n<p>how can I convert this kind of array into torch.tensor?<\/p>\n\n<p>TypeError: can't convert np.ndarray of type numpy.object_. The only supported types are: <code>float64, float32, float16, int64, int32, int16, int8, and uint8.<\/code><\/p>\n","answers":[]}
+{"QuestionId":57967696,"AnswerCount":0,"Tags":"<deep-learning><pytorch><gradient-descent>","CreationDate":"2019-09-17T05:12:02.793","AcceptedAnswerId":null,"OwnerUserId":3983454.0,"Title":"How to customize learning rate decay till a certain iteration with pytorch using Goodfellow's algorithm?","Body":"<p>I need to adjust SGD learning rate using GoodFellow's algorithm (details at the end) <\/p>\n\n<pre><code>eps_k = (1 \u2212 \u03b1)eps_0 + \u03b1*eps_\u03c4\n<\/code><\/pre>\n\n<p>where eps_0 is the initial learning rate, tau is the number of iteration after which lr no longer changes, alpha = iteration\/tau and eps_\u03c4 is supposed to be 1% of eps_0<\/p>\n\n<p>I used this function to \ntau = 200<\/p>\n\n<pre><code>def adjust_learning_rate(optimizer, initial_lr, tau, i):\n    \"\"\"Sets the learning rate to reduce until tau-th iteration\"\"\"\n    if i &lt; tau:\n      alpha = i\/tau\n      lr = (1 - alpha) * initial_lr + alpha * 0.01 * initial_lr\n      for param_group in optimizer.param_groups:\n          param_group['lr'] = lr\n<\/code><\/pre>\n\n<p>During the training stage, should I call this decay function after running zero_grad() on optimizer like this?<\/p>\n\n<pre><code>#loop over epochs:\noptimizer.zero_grad()\n\n#add it here?\nadjust_learning_rate(optimizer, initial_lr, tau, i)\noutputs = net(images)\nloss = criterion(outputs, labels)\nloss.backward()\n<\/code><\/pre>\n\n<p>I looked up official documentation on how to use scheduler to adjust lr, but nothing quite like this method. Is there an existing scheduler function I should use instead or create my own custom lr decay function?<\/p>\n\n<p>Attached the formula from GoodFellow book<a href=\"https:\/\/i.stack.imgur.com\/7QEPc.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/7QEPc.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57968030,"AnswerCount":0,"Tags":"<python><keras><path-finding>","CreationDate":"2019-09-17T05:50:15.430","AcceptedAnswerId":null,"OwnerUserId":500564.0,"Title":"Iterative Keras model with both constant inputs and some temporal inputs","Body":"<p>I am having a play with Keras and trying to wrap my head around an iterative model for path finding.<\/p>\n\n<p>My data set includes:<\/p>\n\n<ul>\n<li>Maps - a 2d image with walls in black, areas that can be move across in white, these are 7x7 pixels<\/li>\n<li>Paths - each map has at least 1 path, which is is set of 2d coordinates that lead from the start point, which is always (0, 0) to the endpoint (6, 6). These have variable lengths<\/li>\n<\/ul>\n\n<p>For example:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/5Jh6D.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5Jh6D.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Here is a sample map, with walls in black, and the path that I want expect to be generated in red.<\/p>\n\n<p>I want to pass into the model the map (which I convert to a 1D array)\nThe set of points of the path that have been generated to date, after 3 iterations I would expect: [(0,0), (1, 0), (2, 0)].\nIf not the whole path at last the last n-points.<\/p>\n\n<p>I then expect as the output either a direction to the next cell such as:<\/p>\n\n<pre><code>class PathAction(Enum):\n  NONE = 0\n  UP = 1\n  RIGHT = 2\n  DOWN = 3\n  LEFT = 4\n  MAX = 5\n<\/code><\/pre>\n\n<p>but would also be happy with the coordinates of the next cell.<\/p>\n\n<p>So I am stuck trying to work out how to set this model up in Keras.\nShould I simply set up a Dense layer that takes the map as 1D array + n-points from the path and trained on what I expect as the action to get to the next cell, repeating this until the endpoint is reached.<\/p>\n\n<p>OR should I use some sort of LTSM layer for the path + map as 1d in a Dense layer? If I go down this road, I'm not quite sure how the iterative training part is combined with the LTSM.<\/p>\n\n<p>Any help appreciated.<\/p>\n","answers":[]}
+{"QuestionId":57968087,"AnswerCount":0,"Tags":"<python><deep-learning><pytorch><yolo>","CreationDate":"2019-09-17T05:55:53.793","AcceptedAnswerId":null,"OwnerUserId":12069056.0,"Title":"compute yolo v1 loss when output w,h is negative","Body":"<p><img src=\"https:\/\/i.stack.imgur.com\/fe9kH.png\" alt=\"yolo loss function\"><\/p>\n\n<p>Then I start to train my model, the output w and h is a negative value, i don't know how to compute the loss in this situation. I can't find any description in the paper.<\/p>\n","answers":[]}
+{"QuestionId":57968110,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-17T05:57:23.743","AcceptedAnswerId":null,"OwnerUserId":8186207.0,"Title":"Keras custom loss function yields weird result","Body":"<p>I'm trying to write a custom loss function of weighted binary cross-entropy in Keras. However, when I compiled my model with the custom loss function, both of the loss and the accuracy went down. Normally the accuracy is around 90% when I train the model with plain BCE, but it came down to 3-10% when I used my custom loss function. Here is my custom loss function:<\/p>\n\n<pre><code>def weighted_crossentropy_core(y_true, y_pred, pos_weight, neg_weight):\n\n    wcel = K.binary_crossentropy(y_true, y_pred)\n    weight = y_true * pos_weight + (1.0 - y_true) * neg_weight\n    wcel = K.mean(weight * cel)\n\n    return wcel\n\ndef weighted_crossentropy_wrapper(pos_weight=1, neg_weight=1):\n    def weighted_crossentropy_return(y_true, y_pred):\n        return weighted_crossentropy_core(y_true, y_pred, pos_weight, neg_weight)\n    return weighted_crossentropy_return\n\nwcel = weighted_crossentropy_wrapper()\nmodel.compile(Adam(init_lr), loss=wcel, metrics=[\"accuracy\"])\n<\/code><\/pre>\n\n<p>The most weird thing is that when I wrapped the tf.keras.losses.binary_crossentropy loss function and then sent the wrapper in (which is essentially the same as keras' bce loss), the resulting losses were totally different from those I got when passing keras bce without the wrapper! Here is how I wrap the bce:<\/p>\n\n<pre><code>def wrapped_bce(y_true, y_pred):\n\n    bce = tf.keras.losses.binary_crossentropy(y_true, y_pred)\n\n    return bce\n\nwcel = wrapped_bce()\nmodel.compile(Adam(init_lr), loss=wcel, metrics=[\"accuracy\"])\n<\/code><\/pre>\n\n<p>Is there anything wrong with the way I wrap my loss function? Thanks in advance!<\/p>\n","answers":[{"AnswerId":"57968176","CreationDate":"2019-09-17T06:02:31.353","ParentId":null,"OwnerUserId":"2042297","Title":null,"Body":"<p>To wrap your loss in the second example, you should use <code>wcel = wrapped_bce<\/code> (without parenthesis). <\/p>\n\n<p>Same thing in the first example, but your function <code>weighted_crossentropy_wrapper<\/code> is a wrapper around another wrapper and omits <code>y_true<\/code> and <code>y_pred<\/code>, this looks a bit strange to me.<\/p>\n"}]}
+{"QuestionId":57968421,"AnswerCount":2,"Tags":"<python><keras><lstm><mean-square-error>","CreationDate":"2019-09-17T06:21:48.130","AcceptedAnswerId":57971032.0,"OwnerUserId":7215349.0,"Title":"Mean squred error interpretation in LSTM model (bidirectional or multiparallel)","Body":"<p>I'm playing with time series and Keras LSTM 1) bidirectional and 2) multiparallel model. I'm saving\nthe best model according to the \"mean_squared_error\" metrics. My dataset is normalized with MinMaxScaler (default range from 0 to 1). Mean squared error is 0.02 on the test part of the dataset. Does it mean that my model's mean error is 14% - that is 0.02^0.5. Is it a good practical interpretation of the model's precision? <\/p>\n\n<p>Suppose I want to predict fourth value in this sequence:<\/p>\n\n<pre><code>[10, 20, 30, 40, 50, 60, 70, 80, 90]\n<\/code><\/pre>\n\n<p>So my x_test and y_test looks like this:<\/p>\n\n<pre><code>[10 20 30] 40\n[20 30 40] 50\n[30 40 50] 60\n[40 50 60] 70\n[50 60 70] 80\n[60 70 80] 90\n<\/code><\/pre>\n\n<p>And the code:<\/p>\n\n<pre><code>cp = [ModelCheckpoint(filepath=path+\"\/epochBi.h5\", monitor='mean_squared_error', verbose=1, save_best_only=True)]\nmodel.compile(optimizer='adam', loss='mse', metrics= ['mean_squared_error'])\nhistory_callback = model.fit(X_train, y_train, epochs=200, verbose=1, callbacks=cp)\nmodel.load_weights(path+\"\/epochBi.h5\")\nscore = model.evaluate(X_test, y_test, verbose=1)\n<\/code><\/pre>\n\n<p>Suppose that I evaluate on original dataset, how would I interpret the MSE=0.02?<\/p>\n","answers":[{"AnswerId":"57968766","CreationDate":"2019-09-17T06:49:08.997","ParentId":null,"OwnerUserId":"10017933","Title":null,"Body":"<p>You are using MSE that measure the distance between your model answer and actual ones. Maybe your model never exactly guesses your target but its guess is enough near to target. So, you can not compute a percent or accuracy here.\nFor more information, you may read about the differences between regression and classification models.<\/p>\n"},{"AnswerId":"57971032","CreationDate":"2019-09-17T09:12:45.337","ParentId":null,"OwnerUserId":"9828197","Title":null,"Body":"<p><strong>UPDATE<\/strong> The question was updated and hence a new comment<\/p>\n\n<p>Again, your MSE of 0.2 DOES NOT MEAN 14% of anything, it is just a scalar and you cannot take it as a percentage of your ground truth. <\/p>\n\n<blockquote>\n  <p>Is it a good practical interpretation of the model's precision? <\/p>\n<\/blockquote>\n\n<p>I would ask this question differently, <code>what does it mean to me?<\/code><\/p>\n\n<p>Well, looking at your target values that range from 40 to 90, you can say that on average, your error was 0.14. Now if you need more details, it means that when you were predicting 40, you guessed 39.86, when 50 --> 50.14, 60--> 59.86, etc. But this is on average.<\/p>\n\n<p>Hope you got the idea by now<\/p>\n\n<hr>\n\n<p>Let's have a look at the formula of mean squared error (MSE):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/0HY5i.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0HY5i.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>So it is just the average of the squared difference between the predicted and the actual data points. In your case, sqaure root of MSE (RMSE) is  0.02^0.5 or around 0.14 (rounded).<\/p>\n\n<blockquote>\n  <p>Does it mean that my model mean error is 14%<\/p>\n<\/blockquote>\n\n<p>No, you cannot say that your the error is 14% when you get the loss of 0.14. 0.14 is just the value of the error, and your each individual errors could have been negative or positive, making their sum zero but not their squared sum.<\/p>\n\n<blockquote>\n  <p>Is it good practical interpretation of the model precision? <\/p>\n<\/blockquote>\n\n<p>Depends on your goals. You usually measure accuracy depending on some other metrics too like MAE, R sqaured and others. Assuming you did not rescale your data to its original magnitude, then there is no way to know how good or bad this is on absolute terms. So I suggest you measure MSE on your rescaled data and then decide, if not done so already.<\/p>\n"}]}
+{"QuestionId":57968503,"AnswerCount":0,"Tags":"<python><python-3.x><pandas><image-processing><keras>","CreationDate":"2019-09-17T06:28:01.330","AcceptedAnswerId":null,"OwnerUserId":7810784.0,"Title":"Keras ImageGenerator : IndexError: list index out of range","Body":"<p>I loaded the data from kaggle kernel to my machine for reprodicing, now the code is not working, but works on the keras on same python environment.<\/p>\n\n<p>Here is the code and the bug.<\/p>\n\n<pre><code>def flow_from_dataframe(img_data_gen, in_df, path_col, y_col, **dflow_args):\n    base_dir = os.path.dirname(in_df[[path_col]].values[0])\n    print('## Ignore next message from keras, values are replaced anyways')\n    df_gen = img_data_gen.flow_from_directory(base_dir, \n                                     class_mode = 'sparse',\n                                    **dflow_args)\n    df_gen.filenames = in_df[path_col].values\n    df_gen.classes = np.stack(in_df[y_col].values)\n    df_gen.samples = in_df.shape[0]\n    df_gen.n = in_df.shape[0]\n    df_gen._set_index_array()\n    df_gen.directory = '' # since we have the full path\n    print('Reinserting dataframe: {} images'.format(in_df.shape[0]))\n    return df_gen\n\n\ntrain_gen = flow_from_dataframe(core_idg, train_df, \n                             path_col = 'path',\n                            y_col = 'disease_vec', \n                            target_size = IMG_SIZE,\n                             color_mode = 'rgb',\n                            batch_size = 32)\n\nvalid_gen = flow_from_dataframe(core_idg, valid_df, \n                             path_col = 'path',\n                            y_col = 'disease_vec', \n                            target_size = IMG_SIZE,\n                             color_mode = 'rgb',\n                            batch_size = 256) # we can use much larger batches for evaluation\n# used a fixed dataset for evaluating the algorithm\ntest_X, test_Y = next(flow_from_dataframe(core_idg, \n                               valid_df, \n                             path_col = 'path',\n                            y_col = 'disease_vec', \n                            target_size = IMG_SIZE,\n                             color_mode = 'rgb',\n                            batch_size = 1024)) # one big batch\n\nt_x, t_y = next(train_gen)\nfig, m_axs = plt.subplots(4, 4, figsize = (16, 16))\n<\/code><\/pre>\n\n<p>I am able to see the paths and path lists, but not sure, where it is coming from.<\/p>\n\n<pre><code>Found 0 images belonging to 0 classes.\nReinserting dataframe: 10000 images\nTraceback (most recent call last):\n  File \"main.py\", line 181, in &lt;module&gt;\n    batch_size = 32)) # one big batch\n  File \"\/home\/user\/.conda\/envs\/test_env\/lib\/python3.6\/site-packages\/keras_prepro         cessing\/image\/iterator.py\", line 104, in __next__\n    return self.next(*args, **kwargs)\n  File \"\/home\/user\/.conda\/envs\/test_env\/lib\/python3.6\/site-packages\/keras_prepro         cessing\/image\/iterator.py\", line 116, in next\n    return self._get_batches_of_transformed_samples(index_array)\n  File \"\/home\/user\/.conda\/envs\/test_env\/lib\/python3.6\/site-packages\/keras_prepro         cessing\/image\/iterator.py\", line 227, in _get_batches_of_transformed_samples\n    img = load_img(filepaths[j],\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57968555,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-09-17T06:32:51.543","AcceptedAnswerId":57978487.0,"OwnerUserId":2908528.0,"Title":"Extracting Classification Module From the CapsNet","Body":"<p>With Reference to the concept of <a href=\"https:\/\/papers.nips.cc\/paper\/6975-dynamic-routing-between-capsules.pdf\" rel=\"nofollow noreferrer\">Capsule Network<\/a>, I am trying to extract just the classification module from the <a href=\"https:\/\/software.intel.com\/en-us\/articles\/understanding-capsule-network-architecture\" rel=\"nofollow noreferrer\">Intel's Implementation of Capsules in Keras<\/a>, As I don't need the decoder or reconstruction part that is provided in the mentioned link.<\/p>\n\n<p><strong>My Try:<\/strong><\/p>\n\n<p>So I just commented out the decoder part of the network<\/p>\n\n<pre><code>#dec = Dense(512, activation='relu')(mask)\n#dec = Dense(1024, activation='relu')(dec)\n#dec = Dense(784, activation='sigmoid')(dec)\n#dec = Reshape(input_shape)(dec)\n<\/code><\/pre>\n\n<p>and the decoder argument in the following line<\/p>\n\n<pre><code>#model = Model([x, mask_input], [output_capsule, dec])\nmodel = Model([x, mask_input], [output_capsule])\nmodel.compile(optimizer='adam', loss=[ margin_loss, 'mae' ], metrics=[ margin_loss, 'mae', 'accuracy'])\nmodel.fit([X, Y], [Y, X], batch_size=128, epochs=3, validation_split=0.2)\n<\/code><\/pre>\n\n<p><strong>Error<\/strong><\/p>\n\n<p>I am getting the following error.<\/p>\n\n<pre><code>ValueError: When passing a list as loss, it should have one entry per model outputs. The model has 1 outputs, but you passed loss=[&lt;function margin_loss at 0x0000020C3E7A30D0&gt;, 'mae']\n<\/code><\/pre>\n\n<p><strong>Help Required:<\/strong> <\/p>\n\n<p>Can somebody guide me how can I use that only the classification part of the Module because I have images that have dimension 90 x 90 and I want to use the classification part to check the accuracy and later on I would analyze each capsules?<\/p>\n","answers":[{"AnswerId":"57978487","CreationDate":"2019-09-17T16:32:06.783","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>To answer why you are getting that error:<\/p>\n\n<p>The original model has a two outputs. Each output has its own loss function. In this case, the outputs are <code>[output_capsule, dec]<\/code> and the corresponding loss functions are <code>[margin_loss, 'mae']<\/code>. Because you removed the <code>dec<\/code> output, you need to remove its loss function from the compilation of the model.<\/p>\n\n<p>In addition, you will need to make sure that you are passing the correct inputs and outputs everywhere else. In the code that you have here, that matters for <code>model.fit<\/code>. The first argument is the inputs. You still have two inputs to your model (not sure if you <em>want<\/em> two inputs, but that's a different problem), so passing in a list of two inputs to <code>fit<\/code> is good. However, the second argument is the desired outputs. You're currently passing in two output arrays, but you only have one output for your model. Since you removed your model's second output, you should remove the second output that you pass in to <code>fit<\/code>, i.e. <code>X<\/code>.<\/p>\n\n<p>Modified code:<\/p>\n\n<pre><code>model = Model([x, mask_input], [output_capsule])\nmodel.compile(optimizer='adam', loss=[margin_loss], metrics=[margin_loss, 'mae', 'accuracy'])\nmodel.fit([X, Y], [Y], batch_size=128, epochs=3, validation_split=0.2)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57968883,"AnswerCount":1,"Tags":"<python><pip><pytorch><setuptools>","CreationDate":"2019-09-17T06:57:53.967","AcceptedAnswerId":57970570.0,"OwnerUserId":482601.0,"Title":"Unexpected import behavior: non-existing module imported","Body":"<p>I have a torch extension (but I think the error I get is independent of torch), let's call it <code>foo<\/code>, that is built with <code>setuptools<\/code>. However, it depends on whether there is a GPU available in the system, in which case some CUDA code is compiled:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from torch.utils.cpp_extension import CppExtension, BuildExtension, CUDAExtension\n\nif cuda_available:\n    ext = CUDAExtension(\n            'foo_gpu',\n            prefixed(prefix, ['foo.cpp', 'foo_kernel.cu']),\n            define_macros=[('COMPILE_CUDA', '1')])\nelse:\n    ext = CppExtension(\n            'foo_cpu',\n            prefixed(prefix, ['foo.cpp']))\n\nsetup(name=ext.name,\n      version='1.0.0',\n      ext_modules=[ext_module],\n      extra_compile_args=['-mmacosx-version-min=10.9'],\n      cmdclass={'build_ext': BuildExtension})\n<\/code><\/pre>\n\n<p>Notice how the module is compiled to be either <code>foo_gpu<\/code> or <code>foo_cpu<\/code>.<\/p>\n\n<p>Later, I import it as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>try:\n    import foo_gpu as foo_backend\n    CUDA_SUPPORTED = True\nexcept ImportError:\n    CUDA_SUPPORTED = False\n    # Try importing the cpu version\n    try:\n        import foo_cpu as foo_backend\n    except ImportError:\n        ...\n<\/code><\/pre>\n\n<p>Let's say I compile it with CUDA support, so <code>foo_gpu<\/code> goes into PIP (actually goes into PIP as <code>foo-gpu<\/code>, not sure why?)<\/p>\n\n<p>Now, I uninstalled it, <code>pip uninstall foo-gpu<\/code>, and compile it with CPU support only, and pip shows <code>foo-cpu 1.0.0<\/code>.<\/p>\n\n<p><em>BUT<\/em>, now, when I run the import code above, it still finds <code>foo-gpu<\/code>, i.e., the first import statement succeeds! Even though it does not show up in <code>pip<\/code>.<\/p>\n\n<hr>\n\n<p><strong>EDIT<\/strong>. I checked <code>sys.path<\/code> and found there is one folder that contains something with <code>gpu<\/code> in my conda env:<\/p>\n\n<pre><code>$ ls ~\/miniconda3\/envs\/cpu_env\/lib\/python3.7\/site-packages\/foo_cpu-1.0.0-py3.7-linux-x86_64.egg\/ \n EGG-INFO\n __pycache__\n foo_cpu.cpython-37m-x86_64-linux-gnu.so\n foo_cpu.py\n foo_gpu.cpython-37m-x86_64-linux-gnu.so\n<\/code><\/pre>\n\n<p>But how should it get there? In this env (<code>cpu_env<\/code>), I never compiled with GPU support. Are there some caches that get invoked?<\/p>\n","answers":[{"AnswerId":"57970570","CreationDate":"2019-09-17T08:47:03.587","ParentId":null,"OwnerUserId":"482601","Title":null,"Body":"<p>Turns out the GPU build in was cached by setuptools (<code>build\/<\/code>, <code>dist\/<\/code>, etc.) and also installed!<\/p>\n"}]}
+{"QuestionId":57969322,"AnswerCount":0,"Tags":"<python><tensorflow><bazel><tensorflow-model-analysis>","CreationDate":"2019-09-17T07:28:17.900","AcceptedAnswerId":null,"OwnerUserId":11840960.0,"Title":"Unable to find pip package","Body":"<p>I have been using bazel build for my im2txt model from TensorFlow and it shows me <\/p>\n\n<p><code>ModuleNotFoundError: No module named 'nltk'<\/code> <\/p>\n\n<p>I have installed the package nltk and tried to even create an environment and run the bazel script<\/p>\n\n<p>Is there anyway i need to link my python to bazel for the python modules to be installed separately on an environment ?<\/p>\n\n<p>Repo i am trying to run : <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/im2txt\" rel=\"nofollow noreferrer\">im2txt<\/a><\/p>\n\n<p>Code sample which iam running : <\/p>\n\n<pre><code># Location to save the MSCOCO data.\nMSCOCO_DIR=\"${HOME}\/im2txt\/data\/mscoco\"\n\n# Build the preprocessing script.\ncd research\/im2txt\nbazel build \/\/im2txt:download_and_preprocess_mscoco\n\n# Run the preprocessing script.\nbazel-bin\/im2txt\/download_and_preprocess_mscoco \"${MSCOCO_DIR}\"   ##Error here &lt;----\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57969421,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning><recurrent-neural-network>","CreationDate":"2019-09-17T07:34:12.263","AcceptedAnswerId":57970366.0,"OwnerUserId":6997665.0,"Title":"Encoder Decoder model for RNN in tensorflow","Body":"<p>I am implementing an encoder decoder model using bidirectional RNN for both encoder and decoder. Since I initialize the bidirectional RNN on the encoder side and the weights and vectors associated with the bidirectional RNN is already initialized, I get the following error when I try to initialize another instance on the decoder side:<\/p>\n\n<pre><code>ValueError: Variable bidirectional_rnn\/fw\/gru_cell\/w_ru already exists, disallowed. Did you mean to set reuse=True or reuse=tf.AUTO_REUSE in VarScope?\n<\/code><\/pre>\n\n<p>I tried defining each within it's own <code>name_scope<\/code> like below but to no avail:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def enc(message, weights, biases):\n    message = tf.unstack(message, timesteps_enc, 1)\n    fw_cell = rnn.GRUBlockCell(num_hidden_enc)\n    bw_cell = rnn.GRUBlockCell(num_hidden_enc)\n    with tf.name_scope(\"encoder\"):\n        outputs, _, _ = rnn.static_bidirectional_rnn(fw_cell, bw_cell, message, dtype=tf.float32)\n    return tf.matmul(outputs[-1], weights) + biases\n\n\ndef dec(codeword, weights, biases):\n    codeword = tf.expand_dims(codeword, axis=2)\n    codeword = tf.unstack(codeword, timesteps_dec, 1)\n    fw_cell = rnn.GRUBlockCell(num_hidden_dec)\n    bw_cell = rnn.GRUBlockCell(num_hidden_dec)\n    with tf.name_scope(\"decoder\"):\n        outputs, _, _ = rnn.static_bidirectional_rnn(fw_cell, bw_cell, codeword, dtype=tf.float32)\n    return tf.matmul(outputs[-1], weights) + biases\n<\/code><\/pre>\n\n<p>Can someone please hint at what I am doing wrong?<\/p>\n","answers":[{"AnswerId":"57970366","CreationDate":"2019-09-17T08:33:32.110","ParentId":null,"OwnerUserId":"1849975","Title":null,"Body":"<p>Just putting it as an answer:<\/p>\n\n<p>Just try to exchange <code>name_scope<\/code> for <code>variable_scope<\/code>. I'm not sure if it is still valid, but for older versions of TF, usage of <code>name_scope<\/code> was not encouraged. From your variable name <code>bidirectional_rnn\/fw\/gru_cell\/w_ru<\/code> you can see that the scope is not applied. <\/p>\n"},{"AnswerId":"57970922","CreationDate":"2019-09-17T09:06:39.617","ParentId":null,"OwnerUserId":"5652313","Title":null,"Body":"<p>One thing is that you cannot create variables with the same name in the same scope, so changing <code>name_scope<\/code> for <code>variable_scope<\/code> will fix the training.<\/p>\n\n<p>The other thing is that such a model cannot work as an encoder-decoder model because the decoder RNN cannot be bidirectional. You indeed have the entire target sequences at the training time, but at the inference time, you generate the target left-to-right. This means you only have the left context for the forward RNN, but you don't have the right context for the backward RNN.<\/p>\n"}]}
+{"QuestionId":57969874,"AnswerCount":0,"Tags":"<python><tensorflow><nvidia><tensorrt>","CreationDate":"2019-09-17T08:03:01.627","AcceptedAnswerId":null,"OwnerUserId":7747161.0,"Title":"can the TF_TRTwork with faster Rcnn inception v2 for object detection on nvidia geforce","Body":"<p>I am working on a object detection model to detect the car plates\ni used faster rcnn inception v2 (COCO) to train and i got impressive result for accuracy, but the speed isn't that good i am trying to implement the system for real time human\nso dose the tf_trt work with this model , and if dose can it work with Nvidia Geforce graphic card<\/p>\n\n<p>please help me<\/p>\n","answers":[]}
+{"QuestionId":57970097,"AnswerCount":0,"Tags":"<tensorflow><mnist><dataloader>","CreationDate":"2019-09-17T08:17:09.757","AcceptedAnswerId":null,"OwnerUserId":12078221.0,"Title":"\"cannot unpack non-iterable function object\" : trying to run a code through tensorflow","Body":"<p>while running this code I am getting error<\/p>\n\n<pre><code>mnist = tf.keras.datasets.mnist\n(x_train, y_train), (x_test, y_test) = mnist.load_data\n<\/code><\/pre>\n\n<p>this should run fine .<\/p>\n\n<p>Anybody know how to resolve this?\nPlease help me<\/p>\n","answers":[]}
+{"QuestionId":57970504,"AnswerCount":0,"Tags":"<tensorflow><object-detection>","CreationDate":"2019-09-17T08:43:15.420","AcceptedAnswerId":null,"OwnerUserId":3284573.0,"Title":"TensorFlow: identify all rectangles in binary image","Body":"<p>I have a binary image <code>(H, W)<\/code> in TensorFlow that looks like this. I want to extract all rectangles in it. Right now, I'm using <code>tf.contrib.image.connected_components<\/code>, but it identifies the corners too.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/gRpv0.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/gRpv0.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>For example, given a corner that looks like the following, two rectangles should be extracted. Is there any TF utility that can achieve this?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/3FT2K.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3FT2K.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":57970556,"AnswerCount":1,"Tags":"<machine-learning><keras><neural-network>","CreationDate":"2019-09-17T08:46:23.667","AcceptedAnswerId":null,"OwnerUserId":8427065.0,"Title":"Why does output changes after I perform cross validation?","Body":"<p>I have built a neural network for performing regression. However, if I'm performing cross-validation before making prediction, the output changes. Below are the graphs with and without cross validation.<\/p>\n\n<p><strong>With Cross Validation<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/dckzk.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/dckzk.png\" alt=\"enter image description here\"><\/a>  <\/p>\n\n<p><strong>Without Cross Validation<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/yp3HY.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/yp3HY.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The code that I use for cross validation<\/p>\n\n<pre><code>from keras.wrappers.scikit_learn import KerasRegressor\nfrom sklearn.model_selection import KFold\n\nepoch = 5000\n\nn_cols = X_train.shape[1]\n\ndef baseline_model():\n\n    model = Sequential()\n    model.add(Dense(3, activation='sigmoid', input_shape=(n_cols,)))\n    model.add(Dense(1, activation = 'linear'))\n    model.compile(optimizer='adam', loss='mse')\n\n    return model\n\nestimator = KerasRegressor(build_fn=baseline_model, epochs=epoch, batch_size=16, verbose = 0)\n\nkfold = KFold(n_splits=5)\nresults = cross_val_score(estimator, X_train, y_train, cv=kfold)\nprint(\"Results: %.10f (%.10f) MSE\" % (results.mean(), results.std()))\nprint(\"RMSE:\", np.sqrt(abs(results.mean())))\n\nprint(results)\n<\/code><\/pre>\n\n<p>for prediction<\/p>\n\n<pre><code>epoch = 5000\nn_cols = X_train.shape[1]\n\ndef modelling():\n\n    model = Sequential()\n    model.add(Dense(4, activation='tanh', input_shape=(n_cols,)))\n    model.add(Dense(1, activation = 'linear'))\n    model.compile(optimizer='adam', loss='mse')\n\n    return model\nmodel = modelling()\nhistory = model.fit(X_train, y_train, epochs= epoch, validation_split = 0.3, batch_size= 16, verbose = 0)\n<\/code><\/pre>\n\n<p>Using keras with tensorflow backend<\/p>\n","answers":[{"AnswerId":"57970775","CreationDate":"2019-09-17T08:58:28.900","ParentId":null,"OwnerUserId":"7350293","Title":null,"Body":"<p>That's the essence of cross-validation. Instead of one evaluation, it yields the mean and std of many evaluations. For your example, you are using a 5 split Kfold, which means you will be learning on 4\/5 of train data and testing on the remaining 1\/5 for 5 times. \nCross validation is used to be sure that your model is not overfitting. <\/p>\n"}]}
+{"QuestionId":57970717,"AnswerCount":1,"Tags":"<python><tensorflow><keras><gan>","CreationDate":"2019-09-17T08:54:48.017","AcceptedAnswerId":57971850.0,"OwnerUserId":11825110.0,"Title":"Using pretrained convolutional network as a GAN discriminator","Body":"<p>I've pulled some code from TF2.0 documentation to generate images from a custom dataset. The code is <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/generative\/dcgan\" rel=\"nofollow noreferrer\">here<\/a> <\/p>\n\n<p>Since the documentation uses Keras i figured i might change the discriminator network to a pretrained network e.g InceptionV3, and only train the top layers. I've found <a href=\"https:\/\/keras.io\/applications\/\" rel=\"nofollow noreferrer\">this<\/a> code (Fine-tune InceptionV3 on a new set of classes). I cant seem to figure out how to replace the the one with the other. I understand that im trying to replace Sequential mode with the Functional API. But i guess they are somehow interconnected. However, im not a frequent Keras user.<\/p>\n\n<p>My questions is: How do i replace a custom CNN in Sequential mode with a pretrained one from the Functional API to use as a discriminator?<\/p>\n\n<p>EDIT: I would be happy if anyone has examples of doing it with the GANEstimator instead as im more used to TF.<\/p>\n\n<p><strong>Use the generator to generate a random image<\/strong><\/p>\n\n<pre><code>def make_generator_model():\n    model = tf.keras.Sequential()\n    model.add(layers.Dense(7*7*256, use_bias=False, input_shape=(100,)))\n    model.add(layers.BatchNormalization())\n    model.add(layers.LeakyReLU())\n\n    model.add(layers.Reshape((7, 7, 256)))\n    assert model.output_shape == (None, 7, 7, 256) # Note: None is the batch size\n\n    model.add(layers.Conv2DTranspose(128, (5, 5), strides=(1, 1), padding='same', use_bias=False))\n    assert model.output_shape == (None, 7, 7, 128)\n    model.add(layers.BatchNormalization())\n    model.add(layers.LeakyReLU())\n\n    model.add(layers.Conv2DTranspose(64, (5, 5), strides=(2, 2), padding='same', use_bias=False))\n    assert model.output_shape == (None, 14, 14, 64)\n    model.add(layers.BatchNormalization())\n    model.add(layers.LeakyReLU())\n\n    model.add(layers.Conv2DTranspose(3, (5, 5), strides=(2, 2), padding='same', use_bias=False, activation='tanh'))\n    assert model.output_shape == (None, 28, 28, 3)\n\n    return model\n\ngenerator = make_generator_model()\nnoise = tf.random.normal([1, 100])\ngenerated_image = generator(noise, training=False)\n<\/code><\/pre>\n\n<p><strong>The current discriminator and helpers (Outputs tf.Tensor([[-0.0003378]], shape=(1, 1), dtype=float32))<\/strong><\/p>\n\n<pre><code>def make_discriminator_model():\n    model = tf.keras.Sequential()\n    model.add(layers.Conv2D(64, (5, 5), strides=(2, 2), padding='same',\n                                     input_shape=[28, 28, 1]))\n    model.add(layers.LeakyReLU())\n    model.add(layers.Dropout(0.3))\n\n    model.add(layers.Conv2D(128, (5, 5), strides=(2, 2), padding='same'))\n    model.add(layers.LeakyReLU())\n    model.add(layers.Dropout(0.3))\n\n    model.add(layers.Flatten())\n    model.add(layers.Dense(1))\n\n    return model\n\ncross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=True)\n\ndef discriminator_loss(real_output, fake_output):\n    real_loss = cross_entropy(tf.ones_like(real_output), real_output)\n    fake_loss = cross_entropy(tf.zeros_like(fake_output), fake_output)\n    total_loss = real_loss + fake_loss\n    return total_loss\n\ndiscriminator = make_discriminator_model()\ndecision = discriminator(generated_image)\nprint (decision)\n<\/code><\/pre>\n\n<p><strong>The desired discriminator<\/strong><\/p>\n\n<pre><code>def make_discriminator_model():\n    # create the base pre-trained model\n    model = InceptionV3(weights='imagenet', include_top=False)\n\n    # ADD TOP LAYERS\n\n    # FREEZE ALL LAYERS EXCEPT TOP LAYERS\n\n    return model\n\n# COMPILE\n\ndef discriminator_loss(real_output, fake_output):\n    real_loss = ??? # Real Loss\n    fake_loss = ??? # Fake loss\n    total_loss = real_loss + fake_loss\n    return total_loss\n\nnoise = tf.random.normal([1, 100])\ngenerated_image = generator(noise, training=False)\n\ndiscriminator = make_discriminator_model()\ndecision = discriminator(generated_image)\nprint (decision)\n<\/code><\/pre>\n\n<p><strong>All imports<\/strong><\/p>\n\n<pre><code>  from __future__ import absolute_import, division, print_function, unicode_literals\n\ntry:\n  # %tensorflow_version only exists in Colab.\n  %tensorflow_version 2.x\nexcept Exception:\n  pass\nimport tensorflow as tf\nprint('TF version: {}'.format(tf.__version__))\n\nimport glob\nimport imageio\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport os\nimport PIL\nfrom PIL import Image\nfrom tensorflow.keras import layers\nimport time\n\nfrom IPython import display\nfrom tensorflow.keras.preprocessing import image\nfrom tensorflow.keras.applications import vgg16\nimport os.path\nfrom tensorflow.keras.applications.inception_v3 import InceptionV3\nfrom tensorflow.keras.preprocessing import image\nfrom tensorflow.keras.models import Model\nfrom tensorflow.keras.layers import Dense, GlobalAveragePooling2D\nfrom tensorflow.keras import backend as K\n<\/code><\/pre>\n\n<p><strong>EDIT:\nThis was the discriminator i ended up with! Thanks to @pandrey<\/strong><\/p>\n\n<pre><code>def make_discriminator_model():\n    pre_trained = tf.keras.applications.InceptionV3(\n        weights='imagenet', include_top=False, input_shape=IMG_SHAPE\n    )\n    pre_trained.trainable = False  # mark all weights as non-trainable\n    model = tf.keras.Sequential([pre_trained])\n    model.add(layers.GlobalAveragePooling2D())\n    model.add(layers.Dense(1))   \n    return model\n<\/code><\/pre>\n","answers":[{"AnswerId":"57971850","CreationDate":"2019-09-17T10:00:08.170","ParentId":null,"OwnerUserId":"12076271","Title":null,"Body":"<p>Here is a way to achieve the building of a partly-pretrained-and-frozen model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Load the pre-trained model and freeze it.\npre_trained = tf.keras.applications.InceptionV3(\n    weights='imagenet', include_top=False\n)\npre_trained.trainable = False  # mark all weights as non-trainable\n# Define a Sequential model, adding trainable layers on top of the previous.\nmodel = tf.keras.Sequential([pre_trained])\n# ADD TOP LAYERS HERE, just as you would in your example code\n# You can check that the proper parameters are frozen or learnable with:\nmodel.summary()\n<\/code><\/pre>\n\n<p>I do not think you need to alter the loss function compared to the one you gave as an example.<\/p>\n"}]}
+{"QuestionId":57971007,"AnswerCount":1,"Tags":"<python><tensorflow><progress-bar>","CreationDate":"2019-09-17T09:11:28.870","AcceptedAnswerId":57971496.0,"OwnerUserId":12078222.0,"Title":"TensorFlow 2.0: display progress bar in custom training loop","Body":"<p>I am training a CNN for an audio classification task, and I am using TensorFlow 2.0 RC with a custom training loop (as explained in <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/keras\/training_and_evaluation#part_ii_writing_your_own_training_evaluation_loops_from_scratch\" rel=\"nofollow noreferrer\">this guide<\/a> from their official site). I would find it really handy to have a nice progress bar, similarly to the usual Keras <code>model.fit<\/code>.<\/p>\n\n<p>This is an outline of my training code (I am using 4 GPU's, with a mirrored distribution strategy):<\/p>\n\n<pre><code>strategy = distribute.MirroredStrategy()\n\ndistr_train_dataset = strategy.experimental_distribute_dataset(train_dataset)\n\nif valid_dataset:\n    distr_valid_dataset = strategy.experimental_distribute_dataset(valid_dataset)\n\nwith strategy.scope():\n\n    model = build_model() # build the model\n\n    optimizer = # define optimizer\n    train_loss = # define training loss\n    train_metrics_1 = # AUC-ROC\n    train_metrics_2 = # AUC-PR\n    valid_metrics_1 = # AUC-ROC for validation\n    valid_metrics_2 = # AUC-PR for validation\n\n    # rescale loss\n    def compute_loss(labels, predictions):\n        per_example_loss = train_loss(labels, predictions)\n        return per_example_loss\/config.batch_size\n\n    def train_step(batch):\n        audio_batch, label_batch = batch\n        with tf.GradientTape() as tape:\n            logits = model(audio_batch)\n            loss = compute_loss(label_batch, logits)\n        variables = model.trainable_variables\n        grads = tape.gradient(loss, variables)\n        optimizer.apply_gradients(zip(grads, variables))\n\n        train_metrics_1.update_state(label_batch, logits)\n        train_metrics_2.update_state(label_batch, logits)\n        train_mean_loss.update_state(loss)\n        return loss\n\n    def valid_step(batch):\n        audio_batch, label_batch = batch\n        logits = model(audio_batch, training=False)\n        loss = compute_loss(label_batch, logits)\n\n        val_metrics_1.update_state(label_batch, logits)\n        val_metrics_2.update_state(label_batch, logits)\n        val_loss.update_state(loss)\n        return loss\n\n    @tf.function \n    def distributed_train(batch):\n        num_batches = 0\n        for batch in distr_train_dataset:\n            num_batches += 1\n            strategy.experimental_run_v2(train_step, args=(batch, ))\n            # print progress here\n            tf.print('Step', num_batches, '; Loss', train_mean_loss.result(), '; ROC_AUC', train_metrics_1.result(), '; PR_AUC', train_metrics_2.result())\n            gc.collect()\n\n    @tf.function\n    def distributed_valid(batch):\n        for batch in distr_valid_dataset:\n            strategy.experimental_run_v2(valid_step, args=(batch, ))\n            gc.collect()\n\nfor epoch in range(epochs):\n    distributed_train(distr_train_dataset)\n    gc.collect()\n    train_metrics_1.reset_states()\n    train_metrics_2.reset_states()\n    train_mean_loss.reset_states()\n\n    if valid_dataset:\n        distributed_valid(distr_valid_dataset)\n        gc.collect()\n        val_metrics_1.reset_states()\n        val_metrics_2.reset_states()\n        val_loss.reset_states()\n<\/code><\/pre>\n\n<p>Here <code>train_dataset<\/code> and <code>valid_dataset<\/code> are two tf.data.TFRecordDataset generated with the usual tf.data input pipeline.<\/p>\n\n<p>TensorFlow provides a really nice tf.keras.utils.Progbar (which is indeed what you see when you train using <code>model.fit<\/code>). I have taken a look at its <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/r1.14\/tensorflow\/python\/keras\/utils\/generic_utils.py#L313-L480\" rel=\"nofollow noreferrer\">source code<\/a>, and it relies on numpy, so I can't use it in place of the <code>tf.print()<\/code> statement (which is executed in graph mode).<\/p>\n\n<p>How can I implement a similar progress bar in my custom training loop (with my training function running in graph mode)?<\/p>\n\n<p>How does <code>model.fit<\/code> display a progress bar in the first place?<\/p>\n","answers":[{"AnswerId":"57971496","CreationDate":"2019-09-17T09:38:35.717","ParentId":null,"OwnerUserId":"12076271","Title":null,"Body":"<blockquote>\n  <p>How can I implement a similar progress bar in my custom training loop (with my training function running in graph mode)?<\/p>\n<\/blockquote>\n\n<p>Why not change a bit the structure of your code, so that you encapsulate single <code>strategy.experimental_run_v2<\/code> calls in <code>tf.function<\/code>-decorated functions and have them return the metrics you want to display, then run those within a non-decorated <code>for<\/code> loop and use a <code>tf.keras.utils.Progbar<\/code>?<\/p>\n\n<blockquote>\n  <p>How does <code>model.fit<\/code> display a progress bar in the first place?<\/p>\n<\/blockquote>\n\n<p>In v2, <code>model.fit<\/code> displays a Progress bar through the use of a <code>TrainingContext<\/code> object, which encapsulates a <code>Progbar<\/code> as well as other specified callbacks, which are doted with <code>on_epoch_end<\/code>, <code>on_batch_begin<\/code>, etc. methods which process logs. To be honest, I am not quite sure how to implement similar mechanisms in a custom training loop, but it might be worth investigating the default one, whose source is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/keras\/engine\/training_v2.py\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":57971128,"AnswerCount":0,"Tags":"<python><keras><neural-network><reinforcement-learning><tensor>","CreationDate":"2019-09-17T09:17:50.043","AcceptedAnswerId":null,"OwnerUserId":9126289.0,"Title":"how to divide a problem into states and choose matrix size","Body":"<p>I'm new to deep learning, so I'm sorry if my question is out of place. I'm currently working on a reinforcement deep learning code to apply on a game similar to bubble shooter. I'm a bit confused about how to divide this problem into different states, and what is the recommended tensor size for the states (for example, if I have 4 states, how should I choose the size of the matrix)?<\/p>\n\n<p>I was thinking about different state for every ball color, but I'm not sure how to continue. How can I notify my agent about what's happening in the environment so he'll choose an action? should I scan the environment and tell him in which coordinates there is a matching color for the one in the cannon? is it a logical solution? I tried this approach but got stuck.<\/p>\n\n<p>I assume there are varied considerations for choosing the states, but if anyone could give me some points to think about, or a link to read about (I was searching online, but haven't found anything useful), I would be very grateful.<\/p>\n\n<p>Thank you in advance!<\/p>\n","answers":[]}
+{"QuestionId":57971261,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2019-09-17T09:25:57.343","AcceptedAnswerId":null,"OwnerUserId":9140187.0,"Title":"Protocol message RunOptions has no \"report_tensor_allocations_upon_oom\" field","Body":"<p>I'm training ctpn: <a href=\"https:\/\/github.com\/eragonruan\/text-detection-ctpn\" rel=\"nofollow noreferrer\">GitHub link<\/a> <br>\nusing <code>tensorflow-gpu==1.3<\/code> <br><\/p>\n\n<blockquote>\n  <p>Exception:<\/p>\n  \n  <p>Traceback (most recent call last):\n    File \"\/home\/xxx\/text-detection-ctpn\/ctpn\/train_net.py\", line 54, in \n      restore=True)<br>\n    File \"\/home\/xxx\/text-detection-ctpn\/lib\/fast_rcnn\/train.py\", line 245, in train_net\n      sw.train_model(sess, max_iters, restore=restore)<br>\n    File \"\/home\/xxx\/text-detection-ctpn\/lib\/fast_rcnn\/train.py\", line 178, in train_model\n      run_options = tf.RunOptions(report_tensor_allocations_upon_oom=True)  <br>\n  ValueError: Protocol message RunOptions has no \"report_tensor_allocations_upon_oom\" field.<\/p>\n<\/blockquote>\n\n<p>this is code:<\/p>\n\n<pre><code>def train_model(self, sess, max_iters, restore=False):\n    \"\"\"Network training loop.\"\"\"\n    with tf.device('\/device:GPU:0'):\n        data_layer = get_data_layer(self.roidb, self.imdb.num_classes)\n        total_loss, model_loss, rpn_cross_entropy, rpn_loss_box = self.net.build_loss(ohem=cfg.TRAIN.OHEM)\n        # scalar summary\n        tf.summary.scalar('rpn_reg_loss', rpn_loss_box)\n        tf.summary.scalar('rpn_cls_loss', rpn_cross_entropy)\n        tf.summary.scalar('model_loss', model_loss)\n        tf.summary.scalar('total_loss', total_loss)\n        summary_op = tf.summary.merge_all()\n\n        # log_image, log_image_data, log_image_name =\\\n        #     self.build_image_summary()\n\n        # optimizer\n        lr = tf.Variable(cfg.TRAIN.LEARNING_RATE, trainable=False)\n        if cfg.TRAIN.SOLVER == 'Adam':\n            opt = tf.train.AdamOptimizer(cfg.TRAIN.LEARNING_RATE)\n        elif cfg.TRAIN.SOLVER == 'RMS':\n            opt = tf.train.RMSPropOptimizer(cfg.TRAIN.LEARNING_RATE)\n        else:\n            # lr = tf.Variable(0.0, trainable=False)\n            momentum = cfg.TRAIN.MOMENTUM\n            opt = tf.train.MomentumOptimizer(lr, momentum)\n\n        global_step = tf.Variable(0, trainable=False)\n        with_clip = True\n        if with_clip:\n            tvars = tf.trainable_variables()\n            grads, norm = tf.clip_by_global_norm(tf.gradients(total_loss, tvars), 10.0)\n            train_op = opt.apply_gradients(list(zip(grads, tvars)), global_step=global_step)\n        else:\n            train_op = opt.minimize(total_loss, global_step=global_step)\n\n        # intialize variables\n        sess.run(tf.global_variables_initializer())\n        restore_iter = 0\n\n        # load vgg16\n        if self.pretrained_model is not None and not restore:\n            try:\n                print(('Loading pretrained model '\n                   'weights from {:s}').format(self.pretrained_model))\n                self.net.load(self.pretrained_model, sess, True)  \n            except:\n                raise 'Check your pretrained model {:s}'.format(self.pretrained_model)\n\n        if restore:\n            dir = '\/home\/xxx\/deploy\/dev\/checkpoints'\n\n            ckpt = tf.train.get_checkpoint_state(dir)\n            print('Restoring from {}...'.format(ckpt.model_checkpoint_path), end=' ')\n            self.saver.restore(sess, ckpt.model_checkpoint_path)\n            stem = os.path.splitext(os.path.basename(ckpt.model_checkpoint_path))[0]\n            #restore_iter = int(stem.split('_')[-1])\n            sess.run(global_step.assign(restore_iter))\n            print('done')\n\n        last_snapshot_iter = -1\n        timer = Timer()\n        #print(\"kkkk\")\n        for iter in range(restore_iter, max_iters):\n            timer.tic()\n            # learning rate\n            if iter != 0 and iter % cfg.TRAIN.STEPSIZE == 0:\n                sess.run(tf.assign(lr, lr.eval() * cfg.TRAIN.GAMMA))\n                print(lr)\n\n            # get one batch\n            blobs = data_layer.forward()  ###########################\n            #print(\"qqqq\")\n            feed_dict = {\n                self.net.data: blobs['data'],\n                self.net.im_info: blobs['im_info'],\n                self.net.keep_prob: 0.5,\n                self.net.gt_boxes: blobs['gt_boxes'],    \n                self.net.gt_ishard: blobs['gt_ishard'],\n                self.net.dontcare_areas: blobs['dontcare_areas']\n            }\n            res_fetches = []\n            fetch_list = [total_loss, model_loss, rpn_cross_entropy, rpn_loss_box,\n                          summary_op,\n                          train_op] + res_fetches\n            run_options = tf.RunOptions(report_tensor_allocations_upon_oom=True)    \n\n            total_loss_val, model_loss_val, rpn_loss_cls_val, rpn_loss_box_val, \\\n                summary_str, _ = sess.run(fetches=fetch_list, feed_dict=feed_dict, options=run_options)\n\n            self.writer.add_summary(summary=summary_str, global_step=global_step.eval())\n\n            _diff_time = timer.toc(average=False)\n\n\n            if (iter) % (cfg.TRAIN.DISPLAY) == 0:\n                print('iter: %d \/ %d, total loss: %.4f, model loss: %.4f, rpn_loss_cls: %.4f, rpn_loss_box: %.4f, lr: %f'%\\\n                        (iter, max_iters, total_loss_val, model_loss_val, rpn_loss_cls_val, rpn_loss_box_val, lr.eval()))\n                print('speed: {:.3f}s \/ iter'.format(_diff_time))\n\n            if (iter+1) % cfg.TRAIN.SNAPSHOT_ITERS == 0:\n                last_snapshot_iter = iter\n                self.snapshot(sess, iter)\n<\/code><\/pre>\n\n<p>how to fix?<\/p>\n","answers":[]}
+{"QuestionId":57971362,"AnswerCount":0,"Tags":"<java><python><tensorflow><export><sonnet>","CreationDate":"2019-09-17T09:31:56.353","AcceptedAnswerId":null,"OwnerUserId":12078446.0,"Title":"Unable to load tensorflow model in Java that was originally created in python (file not found exception)","Body":"<p>I train a neural network in python 3.7 using tensorflow in version 1.14 and sonnet in version 1.34. After training, I want to export my neural network in order to use it in Java for predictions. Unfortunately, it always displays an error while trying to load the model in Java, saying the file cannot be found.<\/p>\n\n<ul>\n<li>I double checked the model path a hundred times and it is correct (including transformation of slashes to backslashes because it runs on windows)<\/li>\n<li>I tried several other export methods in python (including SavedModelBuilder and ModelSaver)<\/li>\n<\/ul>\n\n<p>Below this you can see the Python code to export the model.<\/p>\n\n<pre><code>tf.saved_model.simple_save(sess,\n                           path_graph_model_export_dir,\n                           inputs=model_inputs, outputs={'sim': similarity})\n<\/code><\/pre>\n\n<p>This is the structure of the exported directory.<\/p>\n\n<pre><code>saved_model\n---- saved_model.pb\n---- variables\n-------- variables.data-00000-of-00001\n-------- variables.index\n<\/code><\/pre>\n\n<p>And the code in Java to load the model..<\/p>\n\n<pre><code>SavedModelBundle model = SavedModelBundle.load(modelPath, \"serve\")\n<\/code><\/pre>\n\n<p>results in this exception.<\/p>\n\n<pre><code>2019-09-17 10:59:03.905538: I tensorflow\/cc\/saved_model\/reader.cc:31] Reading SavedModel from: saved_model\n2019-09-17 10:59:03.905997: I tensorflow\/cc\/saved_model\/loader.cc:311] SavedModel load for tags { serve }; Status: fail. Took 457 microseconds.\norg.tensorflow.TensorFlowException: Could not find SavedModel .pb or .pbtxt at supplied export directory path: saved_model\n<\/code><\/pre>\n\n<p>I cannot figure out why this exception appears when the file is clearly located at this directory. I hope someone can help me with this. Thanks in advance!<\/p>\n","answers":[]}
+{"QuestionId":57971680,"AnswerCount":1,"Tags":"<python><multiprocessing><pytorch><python-multiprocessing>","CreationDate":"2019-09-17T09:49:32.287","AcceptedAnswerId":null,"OwnerUserId":7697546.0,"Title":"How to get rid of zombie processes using torch.multiprocessing.Pool (Python)","Body":"<p>I am using <code>torch.multiprocessing.Pool<\/code> to speed up my NN in inference, like this:<\/p>\n\n<pre><code>import torch.multiprocessing as mp\nmp = torch.multiprocessing.get_context('forkserver')\n\ndef parallel_predict(predict_func, sequences, args):\n    predicted_cluster_ids = []\n    pool = mp.Pool(args.num_workers, maxtasksperchild=1)\n    out = pool.imap(\n        func=functools.partial(predict_func, args=args),\n        iterable=sequences,\n        chunksize=1)\n    for item in tqdm(out, total=len(sequences), ncols=85):\n        predicted_cluster_ids.append(item)\n    pool.close()\n    pool.terminate()\n    pool.join()\n    return predicted_cluster_ids\n<\/code><\/pre>\n\n<p>Note 1) I am using <code>imap<\/code> because I want to be able to show a progress bar with <code>tqdm<\/code>.<br>\nNote 2) I tried with both <code>forkserver<\/code> and <code>spawn<\/code> but no luck. I cannot use other methods because of how they interact (poorly) with CUDA.<br>\nNote 3) I am using <code>maxtasksperchild=1<\/code> and <code>chunksize=1<\/code> so for each sequence in <code>sequences<\/code> it spawns a new process.<br>\nNote 4) Adding or removing <code>pool.terminate()<\/code> and <code>pool.join()<\/code> makes no difference.<br>\nNote 5) <code>predict_func<\/code> is a method of a class I created. I could also pass the whole model to <code>parallel_predict<\/code> but it does not change anything.<\/p>\n\n<p>Everything works fine except the fact that after a while I run out of memory on the CPU (while on the GPU everything works as expected). Using <code>htop<\/code> to monitor memory usage I notice that, for every process I spawn with pool I get a zombie that uses 0.4% of the memory. They don't get cleared, so they keep using space. Still, <code>parallel_predict<\/code> does return the correct result and the computation goes on. My script is structured in a way that id does validation multiple times so next time <code>parallel_predict<\/code> is called the zombies add up. <\/p>\n\n<p>This is what I get in <code>htop<\/code>:\n<a href=\"https:\/\/i.stack.imgur.com\/11u1i.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/11u1i.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Usually, these zombies get cleared after <code>ctrl-c<\/code> but in some rare cases I need to <code>killall<\/code>.<\/p>\n\n<p>Is there some way I can force the <code>Pool<\/code> to close them?<\/p>\n\n<p>UPDATE:\nI tried to kill the zombie processes using this:<\/p>\n\n<pre><code>def kill(pool):\n    import multiprocessing\n    import signal\n    # stop repopulating new child\n    pool._state = multiprocessing.pool.TERMINATE\n    pool._worker_handler._state = multiprocessing.pool.TERMINATE\n    for p in pool._pool:\n        os.kill(p.pid, signal.SIGKILL)\n    # .is_alive() will reap dead process\n    while any(p.is_alive() for p in pool._pool):\n        pass\n    pool.terminate()\n<\/code><\/pre>\n\n<p>But it does not work. It gets stuck at <code>pool.terminate()<\/code><\/p>\n\n<p>UPDATE2:\nI tried to use the <code>initializer<\/code> arg in <code>imap<\/code> to catch signals like this:<\/p>\n\n<pre><code>def process_initializer():\n    def handler(_signal, frame):\n        print('exiting')\n        exit(0)\n    signal.signal(signal.SIGTERM, handler)\n\n\ndef parallel_predict(predict_func, sequences, args):\n    predicted_cluster_ids = []\n    with mp.Pool(args.num_workers, initializer=process_initializer, maxtasksperchild=1) as pool:\n        out = pool.imap(\n            func=functools.partial(predict_func, args=args),\n            iterable=sequences,\n            chunksize=1)\n        for item in tqdm(out, total=len(sequences), ncols=85):\n            predicted_cluster_ids.append(item)\n        for p in pool._pool:\n            os.kill(p.pid, signal.SIGTERM)\n        pool.close()\n        pool.terminate()\n        pool.join()\n    return predicted_cluster_ids\n<\/code><\/pre>\n\n<p>but again it does not free memory.<\/p>\n","answers":[{"AnswerId":"57997262","CreationDate":"2019-09-18T16:35:22.887","ParentId":null,"OwnerUserId":"7697546","Title":null,"Body":"<p>Ok, I have more insights to share with you. <del>Indeed this is not a bug, it is actually the \"supposed\" behavior for the multiprocessing module in Python (torch.multiprocessing wraps it).<\/del> What happens is that, although the <code>Pool<\/code> terminates all the processes, the memory is not released (given back to the OS). <del>This is also stated in the <a href=\"https:\/\/docs.python.org\/3.4\/library\/multiprocessing.html?highlight=process#module-multiprocessing.pool\" rel=\"nofollow noreferrer\">documentation<\/a>, though in a very confusing way.<\/del>\nIn the <a href=\"https:\/\/docs.python.org\/3\/library\/multiprocessing.html#module-multiprocessing.pool\" rel=\"nofollow noreferrer\">documentation<\/a> it says that <\/p>\n\n<blockquote>\n  <p>Worker processes within a Pool typically live for the complete duration of the Pool\u2019s work queue<\/p>\n<\/blockquote>\n\n<p>but also:<\/p>\n\n<blockquote>\n  <p>A frequent pattern found in other systems (such as Apache, mod_wsgi, etc) to free resources held by workers is to allow a worker within a pool to complete only a set amount of work before being exiting, being cleaned up and a new process spawned to replace the old one. The maxtasksperchild argument to the Pool exposes this ability to the end user<\/p>\n<\/blockquote>\n\n<p>but the \"clean up\" does NOT happen.<\/p>\n\n<p>To make things worse I found this <a href=\"https:\/\/discuss.pytorch.org\/t\/a-call-to-torch-cuda-is-available-makes-an-unrelated-multi-processing-computation-crash\/4075\/8\" rel=\"nofollow noreferrer\">post<\/a> in which they recommend to use <code>maxtasksperchild=1<\/code>. This increases the memory leak, because this way the number of zombies goes with the number of data points to be predicted, and since <code>pool.close()<\/code> does not free memory they add up. <\/p>\n\n<p>This is very bad if you are using multiprocessing for example in validation. For every validation step I was reinitializing the pool but the memory didn't get freed from the previous iteration.<\/p>\n\n<p>The SOLUTION here is to move <code>pool = mp.Pool(args.num_workers)<\/code> outside the training loop, so the pool does not get closed and reopened, and therefore it always reuses the same processes. NOTE: again remember to remove <code>maxtasksperchild=1<\/code> and <code>chunksize=1<\/code>.<\/p>\n\n<p>I think this should be included in the <a href=\"https:\/\/pytorch.org\/docs\/stable\/notes\/multiprocessing.html\" rel=\"nofollow noreferrer\">best practices<\/a> page.<\/p>\n\n<p>BTW in my opinion this behavior of the multiprocessing library should be considered as a bug and should be fixed Python side (not Pytorch side)<\/p>\n"}]}
+{"QuestionId":57971689,"AnswerCount":1,"Tags":"<tensorflow><azure-machine-learning-service>","CreationDate":"2019-09-17T09:50:01.943","AcceptedAnswerId":57977973.0,"OwnerUserId":3946083.0,"Title":"Disk I\/O extremely slow on P100-NC6s-V2","Body":"<p>I am training an image segmentation model on azure ML pipeline. During the testing step, I'm saving the output of the model to the associated blob storage. Then I want to find the IOU (Intersection over Union) between the calculated output and the ground truth. Both of these set of images lie on the blob storage. However, IOU calculation is extremely slow, and I think it's disk bound. In my IOU calculation code, I'm just loading the two images (commented out other code), still, it's taking close to 6 seconds per iteration, while training and testing were fast enough. <\/p>\n\n<p>Is this behavior normal? How do I debug this step?<\/p>\n","answers":[{"AnswerId":"57977973","CreationDate":"2019-09-17T15:56:27.297","ParentId":null,"OwnerUserId":"8821969","Title":null,"Body":"<p>A few notes on the drives that an AzureML remote run has available:<\/p>\n\n<p>Here is what I see when I run <code>df<\/code> on a remote run (in this one, I am using a blob <code>Datastore<\/code> via <code>as_mount()<\/code>):<\/p>\n\n<pre><code>Filesystem                             1K-blocks     Used  Available Use% Mounted on\noverlay                                103080160 11530364   86290588  12% \/\ntmpfs                                      65536        0      65536   0% \/dev\ntmpfs                                    3568556        0    3568556   0% \/sys\/fs\/cgroup\n\/dev\/sdb1                              103080160 11530364   86290588  12% \/etc\/hosts\nshm                                      2097152        0    2097152   0% \/dev\/shm\n\/\/danielscstorageezoh...-620830f140ab 5368709120  3702848 5365006272   1% \/mnt\/batch\/tasks\/...\/workspacefilestore\nblobfuse                               103080160 11530364   86290588  12% \/mnt\/batch\/tasks\/...\/workspaceblobstore\n<\/code><\/pre>\n\n<p>The interesting items are <code>overlay<\/code>, <code>\/dev\/sdb1<\/code>, <code>\/\/danielscstorageezoh...-620830f140ab<\/code> and <code>blobfuse<\/code>:<\/p>\n\n<ol>\n<li><code>overlay<\/code> and <code>\/dev\/sdb1<\/code> are both the mount of the <strong>local SSD<\/strong> on the machine (I am using a STANDARD_D2_V2 which has a 100GB SSD).<\/li>\n<li><code>\/\/danielscstorageezoh...-620830f140ab<\/code> is the mount of the <strong>Azure File Share<\/strong> that contains the project files (your script, etc.). It is also the <em>current working directory<\/em> for your run.<\/li>\n<li><strong><code>blobfuse<\/code><\/strong> is the blob store that I had requested to mount in the <code>Estimator<\/code> as I executed the run.<\/li>\n<\/ol>\n\n<p>I was curious about the performance differences between these 3 types of drives. My mini benchmark was to download and extract this file: <a href=\"http:\/\/download.tensorflow.org\/example_images\/flower_photos.tgz\" rel=\"nofollow noreferrer\">http:\/\/download.tensorflow.org\/example_images\/flower_photos.tgz<\/a> (it is a 220 MB tar file that contains about 3600 jpeg images of flowers).<\/p>\n\n<p>Here the results:<\/p>\n\n<pre><code>Filesystem\/Drive         Download_and_save       Extract\nLocal_SSD                               2s            2s  \nAzure File Share                        9s          386s\nPremium File Share                     10s          120s\nBlobfuse                               10s          133s\nBlobfuse w\/ Premium Blob                8s          121s\n<\/code><\/pre>\n\n<p>In summary, writing small files is much, much slower on the network drives, so it is highly recommended to use \/tmp or Python <code>tempfile<\/code> if you are writing smaller files. <\/p>\n\n<p>For reference, here the script I ran to measure: <a href=\"https:\/\/gist.github.com\/danielsc\/9f062da5e66421d48ac5ed84aabf8535\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/danielsc\/9f062da5e66421d48ac5ed84aabf8535<\/a><\/p>\n\n<p>And this is how I ran it: <a href=\"https:\/\/gist.github.com\/danielsc\/6273a43c9b1790d82216bdaea6e10e5c\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/danielsc\/6273a43c9b1790d82216bdaea6e10e5c<\/a><\/p>\n"}]}
+{"QuestionId":57972019,"AnswerCount":2,"Tags":"<python><django><python-2.7><tensorflow>","CreationDate":"2019-09-17T10:10:52.183","AcceptedAnswerId":null,"OwnerUserId":11889190.0,"Title":"django.core.exceptions.ImproperlyConfigured: Error loading psycopg2 module: DLL load failed: The specified module could not be found","Body":"<p>After running any command on my project directory i got this:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"C:\\Python27\\Lib\\site-packages\\django\\db\\backends\\postgresql\\base.py\", line 21, in &lt;module&gt;\n    import psycopg2 as Database\n  File \"C:\\Python27\\Lib\\site-packages\\psycopg2\\__init__.py\", line 50, in &lt;module&gt;\n    from psycopg2._psycopg import (                     # noqa\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 \"manage.py\", line 15, in &lt;module&gt;\n    execute_from_command_line(sys.argv)\n  File \"C:\\Python27\\Lib\\site-packages\\django\\core\\management\\__init__.py\", line 364, in execute_from_command_line\n    utility.execute()\n  File \"C:\\Python27\\Lib\\site-packages\\django\\core\\management\\__init__.py\", line 338, in execute\n    django.setup()\n  File \"C:\\Python27\\Lib\\site-packages\\django\\__init__.py\", line 27, in setup\n    apps.populate(settings.INSTALLED_APPS)\n  File \"C:\\Python27\\Lib\\site-packages\\django\\apps\\registry.py\", line 108, in populate\n    app_config.import_models()\n  File \"C:\\Python27\\Lib\\site-packages\\django\\apps\\config.py\", line 202, in import_models\n    self.models_module = import_module(models_module_name)\n  File \"C:\\Users\\Ga\\AppData\\Local\\Programs\\Python\\Python37\\lib\\importlib\\__init__.py\", line 127, in import_module\n    return _bootstrap._gcd_import(name[level:], package, level)\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 1006, in _gcd_import\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 983, in _find_and_load\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 967, in _find_and_load_unlocked\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 677, in _load_unlocked\n  File \"&lt;frozen importlib._bootstrap_external&gt;\", line 728, in exec_module\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 219, in _call_with_frames_removed\n  File \"C:\\Python27\\Lib\\site-packages\\django\\contrib\\auth\\models.py\", line 4, in &lt;module&gt;\n    from django.contrib.auth.base_user import AbstractBaseUser, BaseUserManager\n  File \"C:\\Python27\\Lib\\site-packages\\django\\contrib\\auth\\base_user.py\", line 52, in &lt;module&gt;\n    class AbstractBaseUser(models.Model):\n  File \"C:\\Python27\\Lib\\site-packages\\django\\db\\models\\base.py\", line 124, in __new__\n    new_class.add_to_class('_meta', Options(meta, app_label))\n  File \"C:\\Python27\\Lib\\site-packages\\django\\db\\models\\base.py\", line 325, in add_to_class\n    value.contribute_to_class(cls, name)\n  File \"C:\\Python27\\Lib\\site-packages\\django\\db\\models\\options.py\", line 214, in contribute_to_class\n    self.db_table = truncate_name(self.db_table, connection.ops.max_name_length())\n  File \"C:\\Python27\\Lib\\site-packages\\django\\db\\__init__.py\", line 33, in __getattr__\n    return getattr(connections[DEFAULT_DB_ALIAS], item)\n  File \"C:\\Python27\\Lib\\site-packages\\django\\db\\utils.py\", line 211, in __getitem__\n    backend = load_backend(db['ENGINE'])\n  File \"C:\\Python27\\Lib\\site-packages\\django\\db\\utils.py\", line 115, in load_backend\n    return import_module('%s.base' % backend_name)\n  File \"C:\\Users\\Ga\\AppData\\Local\\Programs\\Python\\Python37\\lib\\importlib\\__init__.py\", line 127, in import_module\n    return _bootstrap._gcd_import(name[level:], package, level)\n  File \"C:\\Python27\\Lib\\site-packages\\django\\db\\backends\\postgresql\\base.py\", line 25, in &lt;module&gt;\n    raise ImproperlyConfigured(\"Error loading psycopg2 module: %s\" % e)\ndjango.core.exceptions.ImproperlyConfigured: Error loading psycopg2 module: DLL load failed: The specified module could not be found.\n<\/code><\/pre>\n\n<p>Is I have installed this \"pip install psycopg2-binary\" on my virtualenv and \"pip install psycopg2==2.7.4\" this one also but none of the command fixed my issue? I have checked all the postgres setting but till on same issue. Is that problem with my postgres setting or installation issue?? <\/p>\n","answers":[{"AnswerId":"57972289","CreationDate":"2019-09-17T10:25:39.890","ParentId":null,"OwnerUserId":"5099803","Title":null,"Body":"<p>Make sure that you have <code>psycopg2<\/code> installed. Reading <a href=\"http:\/\/initd.org\/psycopg\/docs\/install.html\" rel=\"nofollow noreferrer\">this<\/a> might be helpful.<\/p>\n"},{"AnswerId":"57972235","CreationDate":"2019-09-17T10:22:57.773","ParentId":null,"OwnerUserId":"12078894","Title":null,"Body":"<p>Configure django database settings<\/p>\n\n<p>Check settings.py and change Engine to django.db.backends.postgresql_psycopg2<\/p>\n"}]}
+{"QuestionId":57972053,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2019-09-17T10:13:04.040","AcceptedAnswerId":null,"OwnerUserId":12078826.0,"Title":"Few questions about Keras documentation","Body":"<p>In Keras documentation named activations.md, it says \"Activations can either be used through an Activation layer, or through the activation argument supported by <strong>all forward layers<\/strong>.\". Then what is the meaning of <strong>forward layers<\/strong>? I think some layers don't have an activation parameter.(ex. Dropout layer)<\/p>\n\n<p>And \"Activations that are more complex than a simple TensorFlow\/Theano\/CNTK function (eg. learnable activations, <strong>which maintain a state<\/strong>) are available as Advanced Activation layers, and can be found in the module keras.layers.advanced_activations. These include PReLU and LeakyReLU.\". Then what is the meaning of <strong>state<\/strong> in this case?<\/p>\n","answers":[{"AnswerId":"57974433","CreationDate":"2019-09-17T12:40:56.127","ParentId":null,"OwnerUserId":"12076271","Title":null,"Body":"<p>I am not sure there is a strict definition of \"forward layers\" in this context, but basically what it means is that the \"classic\", keras-built-in types of layers comprising one or more sets of weights used to transform an input matrix into an output one have a <code>activation<\/code> argument. Typically, <code>Dense<\/code> layers have one, as well as the various kinds of RNN and CNN layers.<\/p>\n\n<p>It would not make sense for Dropout layers to have an activation function : they simply add a mechanism triggered at training to (hopefully) improve convergence rate and decrease overfitting chances.<\/p>\n\n<p>As for the idea of \"maintaining a state\", it refers to activation functions that would not behave independently on each and every fed-in sample, but would instead retain some learnable information (the so-called <em>state<\/em>). Typically, for a LeakyReLU activation, you could adjust the leak parameter through training (and it would, in the documentation's terminology, be referred to as a <em>state<\/em> of this activation function).<\/p>\n"}]}
+{"QuestionId":57972482,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2019-09-17T10:36:57.413","AcceptedAnswerId":null,"OwnerUserId":11824234.0,"Title":"Implementing MSE loss","Body":"<p>I'm new to deep learning and I want to implement an autoencoder. I'm using <strong>keras<\/strong> and <strong>mse loss function<\/strong>. But when I use MSE function that I implemented althogh the output of my function and keras.losses.mse are approximately the same but the result is significantly worse.<\/p>\n\n<p>Duo to some needs in my project I wanted to modify MSE. But when implemented the quality of the results became much worse.<\/p>\n\n<p>The input of the function is two 23*32*3 ndarrays.\nAnd I had to do slicing because of the modifications I wanted to do.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nimport tensorflow as tf\nimport numpy as np\nimport keras as k\n\ndef custom_loss(im1, im2):\n    im11 = im1[:, :, 0]\n    im12 = im1[:, :, 1]\n    im13 = im1[:, :, 2]\n    im21 = im2[:,:,0]\n    im22 = im2[:,:,1]\n    im23 = im2[:,:,2]\n    diff1 = tf.square(im11 - im21)\n    diff2 = tf.square(im12 - im22)\n    diff3 = tf.square(im13 - im23)\n    res = tf.divide((diff1 + diff2 + diff3), 3)\n    return res\n\n\n<\/code><\/pre>\n\n<p>but when I test it the results are almost the same. But when I use it in model the results are much worse.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\na = np.random.rand(32, 32, 3)\nb = np.random.rand(32, 32, 3)\nv1 = tf.placeholder(tf.float32, (32, 32, 3), 'v1')\nv2 = tf.placeholder(tf.float32, (32, 32, 3), 'v2')\n\nwith tf.Session() as sess:\n    me = sess.run(custom_loss(v1, v2), feed_dict={v1: a, v2: b})\n    true = sess.run(k.losses.mse(v1, v2), feed_dict={v1: a, v2: b})\n\nfor i in range(32):\n    for j in range(32):\n        print(i, j, true[i][j], me[i][j])\n\n<\/code><\/pre>\n","answers":[{"AnswerId":"57972897","CreationDate":"2019-09-17T11:02:21.763","ParentId":null,"OwnerUserId":"5536259","Title":null,"Body":"<p>I believe you are missing summing of the diffs.<\/p>\n\n<p>try:<\/p>\n\n<pre><code>def custom_loss(im1, im2):\n    im11 = im1[:, :, 0]\n    im12 = im1[:, :, 1]\n    im13 = im1[:, :, 2]\n    im21 = im2[:,:,0]\n    im22 = im2[:,:,1]\n    im23 = im2[:,:,2]\n    diff1 = (im11 - im21)**2\n    diff2 = (im12 - im22)**2\n    diff3 = (im13 - im23)**2\n    res = tf.divide((diff1 + diff2 + diff3).sum(), 3)\n    return res\n<\/code><\/pre>\n"}]}
+{"QuestionId":57972486,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-17T10:37:22.130","AcceptedAnswerId":null,"OwnerUserId":12078322.0,"Title":"U-net: TypeError: __call__() missing 1 required positional argument: 'inputs'","Body":"<p>i am working on segmentation of brain tumor dataset provided by BraTS challenge based on u-net. After defining the model at the stage of training an error occurred. <\/p>\n\n<p>The error is:<\/p>\n\n<pre><code>TypeError                                 Traceback (most recent call last)\n&lt;ipython-input-56-d745053f879c&gt; in &lt;module&gt;()\n      1 Folders = ReadDataset(PATH)\n      2 TrainGenerator = MyGenerator(PATH)\n----&gt; 3 Model=unet()\n      4 Model.fit_generator(TrainGenerator, epochs=1, steps_per_epoch=25)\n      5 #for Index, Folder in enumerate(Folders):\n\n&lt;ipython-input-54-baecfef1ddd3&gt; in unet(pretrained_weights, input_size)\n     41     conv10 = Conv2D(1, 1, activation = 'sigmoid')(conv9)\n     42 \n---&gt; 43     model = Model(input = inputs, output = conv10)\n     44 \n     45     model.compile(optimizer = Adam(lr = 1e-4), loss = 'binary_crossentropy', metrics = ['accuracy'])\n\nTypeError: __call__() missing 1 required positional argument: 'inputs'\n<\/code><\/pre>\n\n<pre><code>def unet(pretrained_weights = None,input_size = (240, 240,4)):\n    inputs = Input(input_size)\n    conv1 = Conv2D(64, 3, activation = 'relu', padding = 'same', \n            kernel_initializer = 'he_normal')(inputs)\n    conv1 = Conv2D(64, 3, activation = 'relu', padding = 'same', \n            kernel_initializer = 'he_normal')(conv1)\n\n    .\n    .\n    .\n\n\n    conv10 = Conv2D(1, 1, activation = 'sigmoid')(conv9)\n\n    model = Model(input = inputs, output = conv10)\n\n    model.compile(optimizer = Adam(lr = 1e-4), loss = 'binary_crossentropy', \n                              metrics = ['accuracy'])\n\n\n    return model\n<\/code><\/pre>\n\n<p>And the training code is:<\/p>\n\n<pre><code>Folders = ReadDataset(PATH)\nTrainGenerator = MyGenerator(PATH)\nModel=unet()\nModel.fit_generator(TrainGenerator, epochs=1, steps_per_epoch=25)\nTrain, Test, Valid = Split(Folders, 0.8)\n<\/code><\/pre>\n\n<p>I am just trying to get some result as i am beginner in deep learning. Any help would be highly appreciated and i would be very grateful.<\/p>\n","answers":[{"AnswerId":"57980706","CreationDate":"2019-09-17T19:17:01.003","ParentId":null,"OwnerUserId":"12067496","Title":null,"Body":"<p>Older versions of Keras require you to use 'inputs' instead of 'input' as the name of the first input parameter for Model. Check if you can update your Keras version (current stable is 2.2.5) or try this line:<\/p>\n\n<pre><code>model = Model(inputs = inputs, outputs = conv10)\n<\/code><\/pre>\n\n<p>possibly:<\/p>\n\n<pre><code>model = Model(inputs = [inputs], outputs = [conv10])\n<\/code><\/pre>\n\n<p>Instead of:<\/p>\n\n<pre><code>model = Model(input = inputs, output = conv10)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57972913,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-17T11:03:27.860","AcceptedAnswerId":57974089.0,"OwnerUserId":6503551.0,"Title":"How to map sequence to sequence in tensorflow?","Body":"<p>I have a 3 dimensional matrix of shape (height, width, 4). In fact it is a bitmap with RGBA values per pixel. I would like to reduce each RGBA set to a set with two values, say [x,y]. <\/p>\n\n<p>see image at imgur com \/ Blr2EQC<\/p>\n\n<p>I have tried using map_fn<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import cv2\nimport tensorflow as tf\n\ndef map_pixel_to_vector(elt):\n    b = elt[0] - 127\n    g = elt[1] - 127\n    r = elt[2] - 127\n    a = elt[3] - 127\n\n    dx = (g * 127) + r\n    dy = (a * 127) + b\n    return [dx,dy]\n\nfile = \"example.png\"\nframe = cv2.imread(file, cv2.IMREAD_UNCHANGED\ns = tf.shape(frame)\n\n# reshape to list of pixels\nelts = tf.reshape(frame, (s[0]*s[1],4))\n\n# cast from uint8 to int32 to support negative output\nelts = tf.dtypes.cast(elts, tf.int32)\n\n# map each pixel to output\nelts = tf.map_fn(map_pixel_to_vector, elts)\n\n# reshape back to image resolution\nelts = tf.reshape(elts, (s[0], s[1], 2)\n<\/code><\/pre>\n\n<p>Now I would expect this to work, each [rgba] pixel would be reduced to [xy] pixel, but instead I get <\/p>\n\n<pre><code>ValueError: The two structures don't have the same nested structure.\n\nFirst structure: type=DType str=&lt;dtype: 'int32'&gt;\n\nSecond structure: type=list str=[&lt;tf.Tensor: id=262537, shape=(), dtype=int32, numpy=98&gt;, &lt;tf.Tensor: id=262540, shape=(), dtype=int32, numpy=210&gt;]\n\nMore specifically: Substructure \"type=list str=[&lt;tf.Tensor: id=262537, shape=(), dtype=int32, numpy=98&gt;, &lt;tf.Tensor: id=262540, shape=(), dtype=int32, numpy=210&gt;]\" is a sequence, while substructure \"type=DType str=&lt;dtype: 'int32'&gt;\" is not\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"main.py\", line 97, in &lt;module&gt;\n    loss = loss_fn(exc, [outputs[-1]], [inputs[-1]])\n  File \"main.py\", line 36, in loss_fn\n    elts = tf.map_fn(reduce_pixel_to_vector, elts)\n  File \"\/usr\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/ops\/map_fn.py\", line 268, in map_fn\n    maximum_iterations=n)\n  File \"\/usr\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/ops\/control_flow_ops.py\", line 2714, in while_loop\n    loop_vars = body(*loop_vars)\n  File \"\/usr\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/ops\/control_flow_ops.py\", line 2705, in &lt;lambda&gt;\n    body = lambda i, lv: (i + 1, orig_body(*lv))\n  File \"\/usr\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/ops\/map_fn.py\", line 258, in compute\n    nest.assert_same_structure(dtype or elems, packed_fn_values)\n  File \"\/usr\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/util\/nest.py\", line 313, in assert_same_structure\n    % (str(e), str1, str2))\n\n<\/code><\/pre>\n\n<p>Any help would be greatly appreciated.<\/p>\n","answers":[{"AnswerId":"57974089","CreationDate":"2019-09-17T12:20:10.153","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Your function <code>map_pixel_to_vector<\/code> is returning a list, not a tensor. You can make it into a tensor for example with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/stack\" rel=\"nofollow noreferrer\"><code>tf.stack<\/code><\/a> or <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/convert_to_tensor\" rel=\"nofollow noreferrer\"><code>tf.convert_to_tensor<\/code><\/a>:<\/p>\n\n<pre><code>def map_pixel_to_vector(elt):\n    b = elt[0] - 127\n    g = elt[1] - 127\n    r = elt[2] - 127\n    a = elt[3] - 127\n\n    dx = (g * 127) + r\n    dy = (a * 127) + b\n    return tf.stack([dx, dy])\n<\/code><\/pre>\n\n<p>However, you can do the same operation without <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/map_fn\" rel=\"nofollow noreferrer\"><code>tf.map_fn<\/code><\/a> more simply and efficiently like this:<\/p>\n\n<pre><code>import tensorflow as tf\nimport cv2\n\nfile = \"example.png\"\nframe = tf.constant(cv2.imread(file, cv2.IMREAD_UNCHANGED))\nelts = tf.dtypes.cast(frame, tf.int32)\nr, g, b, a = tf.unstack(elts - 127, num=4, axis=-1)\nelts = tf.stack([(g * 127) + r, (a * 127) + b], axis=-1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57973223,"AnswerCount":1,"Tags":"<java><tensorflow><tensorflow-serving>","CreationDate":"2019-09-17T11:24:16.663","AcceptedAnswerId":null,"OwnerUserId":6265370.0,"Title":"How to create a TensorProto from a Java Map","Body":"<p>I can create a Tensorproto with single values like floats.<\/p>\n\n<pre><code>\/\/ create TensorProto with 3 floats\n        org.tensorflow.framework.TensorProto.Builder tensorProtoBuilder = org.tensorflow.framework.TensorProto.newBuilder();\n        tensorProtoBuilder.setDtype(DataType.DT_FLOAT);\n        tensorProtoBuilder.addFloatVal(1.0f);\n        tensorProtoBuilder.addFloatVal(2.0f);\n        tensorProtoBuilder.addFloatVal(5.0f);\n\n        \/\/ create TensorShapeProto\n        org.tensorflow.framework.TensorShapeProto.Builder tensorShapeBuilder = org.tensorflow.framework.TensorShapeProto.newBuilder();\n        tensorShapeBuilder.addDim(org.tensorflow.framework.TensorShapeProto.Dim.newBuilder().setSize(3));\n\n        \/\/ set shape for proto\n        tensorProtoBuilder.setTensorShape(tensorShapeBuilder.build());\n\n        \/\/ build proto\n        org.tensorflow.framework.TensorProto proto = tensorProtoBuilder.build();\n\n        System.out.println(proto.toString());\n<\/code><\/pre>\n\n<p>But how do I create a Tensorproto with key-value pairs ? The key-value pairs represent features that I want to feed into a classification model.<\/p>\n\n<pre><code>\"country\": \"ireland\", \"currency\": \"euro\"\n<\/code><\/pre>\n\n<p>Is this a case of 2 arrays, one with keys (feature names) and one with values ?<\/p>\n\n<p>Edit:\nIs there a way of converting a collection of org.tensorflow.example.Feature to tensorproto ? <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/54091670\/how-to-provide-input-for-a-tensorflow-dnnregressor-in-java\">How to provide input for a TensorFlow DNNRegressor in Java?<\/a><\/p>\n","answers":[{"AnswerId":"58008633","CreationDate":"2019-09-19T10:09:12.550","ParentId":null,"OwnerUserId":"6265370","Title":null,"Body":"<p>I didn't get any feedback on this. <\/p>\n\n<p>But managed to figure it out so hope my answer and code will help someone in the future.<\/p>\n\n<p>Note: It is not necessary to create a TensorProto for this problem. <\/p>\n\n<p>In a nutshell I had to:\n- create Feature objs\n- add them to Features List obj\n- add Features List obj to Example obj\n- add Example obj to Examples List obj\n- create Input obj\n- add Examples List obj to Input obj<\/p>\n\n<p>Here is the working code.<\/p>\n\n<pre><code>import com.google.protobuf.ByteString;\nimport com.google.protobuf.Int64Value;\nimport io.grpc.ManagedChannel;\nimport io.grpc.ManagedChannelBuilder;\nimport tensorflow.serving.Classification;\nimport tensorflow.serving.Classification.ClassificationRequest;\nimport tensorflow.serving.PredictionServiceGrpc;\nimport tensorflow.serving.PredictionServiceGrpc.PredictionServiceBlockingStub;\nimport tensorflow.serving.Model.ModelSpec;\nimport org.tensorflow.example.Example;\nimport tensorflow.serving.InputOuterClass.Input;\nimport tensorflow.serving.InputOuterClass.ExampleList;\nimport org.tensorflow.example.Features;\n\nimport java.util.HashMap;\nimport java.util.Map;\n\n\npublic class GrpcClient {\n\n    public static void main(String[] args) throws Exception {\n\n        String host = \"localhost\";\n        int port = 8500;\n        String modelName = \"Model_1\";\n        long modelVersion = 1568645807;\n\n        \/\/ create a channel\n        ManagedChannel channel = ManagedChannelBuilder.forAddress(host, port).usePlaintext().build();\n        PredictionServiceBlockingStub stub = PredictionServiceGrpc.newBlockingStub(channel);\n\n        \/\/ create ClassificationRequest\n        ClassificationRequest.Builder classificationRequestBuilder = ClassificationRequest.newBuilder();\n\n        \/\/ create ModelSpec\n        ModelSpec.Builder modelSpecBuilder = ModelSpec.newBuilder();\n        modelSpecBuilder.setName(modelName);\n        modelSpecBuilder.setVersion(Int64Value.of(modelVersion));\n        modelSpecBuilder.setSignatureName(\"serving_default\");\n\n        \/\/ set model for request\n        classificationRequestBuilder.setModelSpec(modelSpecBuilder);\n\n        \/\/ map of input features\n        Map&lt;String, Float&gt; myMap = new HashMap&lt;&gt;();\n        myMap.put(\"feature1\", 1.0f);\n        myMap.put(\"feature2\", 0.0f);\n\n        \/\/ convert map to example list\n        ExampleList.Builder exampleListBuilder = ExampleList.newBuilder();\n        exampleListBuilder.addExamples(buildExample(myMap));\n        ExampleList exampleList = exampleListBuilder.build();\n\n        \/\/ create input\n        Input.Builder inputBuilder = Input.newBuilder();\n        inputBuilder.setExampleList(exampleList);\n\n        \/\/ set input for request\n        Input input = inputBuilder.build();\n        classificationRequestBuilder.setInput(input);\n\n        \/\/ build request\n        Classification.ClassificationRequest request = classificationRequestBuilder.build();\n\n        \/\/ run classification\n        Classification.ClassificationResponse response = stub.classify(request);\n\n        System.out.println(response.toString());\n    }\n\n    public static Example buildExample(Map&lt;String, ?&gt; featureMap) {\n        Features.Builder featuresBuilder = Features.newBuilder();\n        for (String attr : featureMap.keySet()) {\n            Object value = featureMap.get(attr);\n            if (value instanceof Float) {\n                featuresBuilder.putFeature(attr, feature((Float) value));\n            } else if (value instanceof float[]) {\n                featuresBuilder.putFeature(attr, feature((float[]) value));\n            } else if (value instanceof String) {\n                featuresBuilder.putFeature(attr, feature((String) value));\n            } else if (value instanceof String[]) {\n                featuresBuilder.putFeature(attr, feature((String[]) value));\n            } else if (value instanceof Long) {\n                featuresBuilder.putFeature(attr, feature((Long) value));\n            } else if (value instanceof long[]) {\n                featuresBuilder.putFeature(attr, feature((long[]) value));\n            } else {\n                throw new UnsupportedOperationException(\"Not supported attribute value data type!\");\n            }\n        }\n        Features features = featuresBuilder.build();\n        return Example.newBuilder().setFeatures(features).build();\n    }\n\n    private static org.tensorflow.example.Feature feature(String... strings) {\n        org.tensorflow.example.BytesList.Builder b = org.tensorflow.example.BytesList.newBuilder();\n        for (String s : strings) {\n            b.addValue(ByteString.copyFromUtf8(s));\n        }\n        return org.tensorflow.example.Feature.newBuilder().setBytesList(b).build();\n    }\n\n    private static org.tensorflow.example.Feature feature(float... values) {\n        org.tensorflow.example.FloatList.Builder b = org.tensorflow.example.FloatList.newBuilder();\n        for (float v : values) {\n            b.addValue(v);\n        }\n        return org.tensorflow.example.Feature.newBuilder().setFloatList(b).build();\n    }\n\n    private static org.tensorflow.example.Feature feature(long... values) {\n        org.tensorflow.example.Int64List.Builder b = org.tensorflow.example.Int64List.newBuilder();\n        for (long v : values) {\n            b.addValue(v);\n        }\n        return org.tensorflow.example.Feature.newBuilder().setInt64List(b).build();\n    }\n}\n<\/code><\/pre>\n"}]}
+{"QuestionId":57973374,"AnswerCount":0,"Tags":"<tensorflow><keras><tensorflow-estimator>","CreationDate":"2019-09-17T11:33:50.283","AcceptedAnswerId":null,"OwnerUserId":12079217.0,"Title":"MNIST get 10 times worse accuracy rate when use tensorflow.keras.estimator.model_to_estimator","Body":"<p>Same model, model_to_estimator get 10 times worse accuracy rate than keras model fit<\/p>\n\n<p>model_to_estimator: accuracy = <strong>0.10129999<\/strong>\nkeras model fit : accuracy = <strong>0.9694<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nfrom tensorflow import keras\nfrom tensorflow.keras import layers\n\ndef dense_model() :\n    inputs = keras.Input(shape=(784,), name='input_1')\n    x = layers.Dense(64, activation='relu')(inputs)\n    x = layers.Dense(32, activation='relu')(x)\n    outputs = layers.Dense(10, activation='softmax')(x)\n\n    return keras.Model(inputs = inputs, outputs = outputs, name='mnist_model')\n\ndef get_estimator(config) :\n    model = dense_model()\n    model.summary()\n    model.compile(loss='sparse_categorical_crossentropy',\n             optimizer=keras.optimizers.RMSprop(),\n             metrics=['accuracy'])\n\n    return model, keras.estimator.model_to_estimator(keras_model=model, model_dir='.\/model\/', config=config)\n\ndef train_input_fn(features, labels, epochs, batch_size):\n    dataset = tf.data.Dataset.from_tensor_slices((dict(features), labels))\n    return dataset.repeat(epochs).batch(batch_size)\n\ndef main() :\n    (x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()\n\n    x_train = x_train.reshape(-1, 784).astype('float32')\/255\n    x_test= x_test.reshape(-1, 784).astype('float32')\/255\n\n    config = tf.estimator.RunConfig()\n    config.mode_dir = '.\/model\/'\n\n    model, estimator = get_estimator(config)\n    tf.logging.set_verbosity(tf.logging.INFO)\n\n    estimator.train(input_fn=lambda:train_input_fn({'input_1' : x_train}, y_train, 5, 64)\n        , steps=5000)\n\n    estimator.evaluate(input_fn=lambda:train_input_fn({'input_1' : x_test}, y_test, 1, 100))\n\n    ##########\n\n    history = model.fit(x_train, y_train, batch_size=64, epochs=5, validation_split=0.2)\n    test_scores = model.evaluate(x_test, y_test, verbose=0)\n    print('Test loss:', test_scores[0])\n    print('Test accuracy:', test_scores[1])\n\nif __name__ == '__main__':\n\n    main()\n<\/code><\/pre>\n\n<p>INFO:tensorflow:Saving dict for global step 4689: accuracy = <strong>0.10129999<\/strong>, global_step = 4689, loss = 0.106062405<\/p>\n\n<p>INFO:tensorflow:Saving 'checkpoint_path' summary for global step 4689: .\/model\/model.ckpt-4689<\/p>\n\n<p>Train on 48000 samples, validate on 12000 samples<\/p>\n\n<p>Test loss: 0.10144322782349773\nTest accuracy: <strong>0.9694<\/strong><\/p>\n","answers":[]}
+{"QuestionId":57973791,"AnswerCount":0,"Tags":"<pandas><python-3.6><pytorch><lstm>","CreationDate":"2019-09-17T12:01:17.443","AcceptedAnswerId":null,"OwnerUserId":12079288.0,"Title":"How to encode my data to use it with an Pytorch-LSTM","Body":"<p>I am trying to set up a Pytorch-LSTM for an affect-analyis-problem. I am currently stuck on how I should correctly encode my data and what kind of data the LSTM expects. <\/p>\n\n<p>I have a dataset of roughly 12000 sentences, all tagges with a specific emotion. My data is encoded as follows: <\/p>\n\n<p>I created a seperate dataset that consists of the Part-of-Speech-Tags for every word in the sentences. The POS-Tags are one-hot-encoded:<\/p>\n\n<pre><code>0           1               83\n[1,0,0..0], [0,1,0..0], ... [0,0,0..0] \n[0,0,0..1], [0,0,1..0], ... [1,0,0..0]\n...\n<\/code><\/pre>\n\n<p>Every row is one sentence. All rows have a length of 84 (shorter sentences are padded with a zero-vector). The lists in each cell correspond to a specific pos-tags, they have a length of 17. <\/p>\n\n<p>The target-dataset has the same number of rows as the dataset above, but every row only contains one list, encoding one of four emotions: <\/p>\n\n<pre><code>0           \n[1,0,0,0] \n[0,0,0,1]\n...\n<\/code><\/pre>\n\n<p>I am now stuck on how to instert this data (especially the pos-dataset) into the lstm. The docs say, the LSTM expects an input-dimension and output-dimension parameter. What is exactly is my input dimension here? Every sequence has a length of 84, but every item in this sequence is a list of length 17.<\/p>\n\n<p>I've read the docs and some tutorials, however none of these seem to use a similar data structure as I do. Any help would be greatly appreciated. <\/p>\n","answers":[]}
+{"QuestionId":57973921,"AnswerCount":0,"Tags":"<tensorflow><keras><conv-neural-network><lstm>","CreationDate":"2019-09-17T12:09:29.843","AcceptedAnswerId":null,"OwnerUserId":11655562.0,"Title":"How TimeDistributed layer works?","Body":"<p>I looking for details of work of TimeDistributed to understand how backpropagation through time affect on CNN weights (in other words, can I reuse previous CNN prediction's embedding to classify new sequence of images (something like a queue where full CNN prediction calculation just for first image)).<\/p>\n\n<p>Docs:\n<a href=\"https:\/\/keras.io\/layers\/wrappers\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/wrappers\/<\/a>\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/TimeDistributed\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/TimeDistributed<\/a><\/p>\n\n<p>Now I try to dive in into source code but it is difficult)<\/p>\n","answers":[]}
+{"QuestionId":57973934,"AnswerCount":0,"Tags":"<nlp><pytorch>","CreationDate":"2019-09-17T12:10:26.617","AcceptedAnswerId":null,"OwnerUserId":12047252.0,"Title":"Pytorch loss function with ignore_index cause that model doesn't converge","Body":"<p>I was going through Pytorch official tutorial regarding the implementation of seq2eq Chatbot with Attention. To gain better understanding I used different methods to achieve similar results. One of as I thought improvements was the implementation of simple nn.NLLLoss function with attribute ignore_index = padding_token. Previous solution that used binary mask tensor describing the padding of the target tensor seemed to be too complicated, I wanted a simpler solution, but as it tuned out adding few lines of code made everything worse. Now model doesn't converge at all. I tried to change learning rate, but it didn\u2019t help.<\/p>\n\n<p>Pytorch chatbot tutirial: <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/chatbot_tutorial.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/tutorials\/beginner\/chatbot_tutorial.html<\/a><\/p>\n\n<pre><code>criterion = nn.NLLLoss(ignore_index=PAD_token, reduction='mean')\n<\/code><\/pre>\n\n<p>In the definition of train function I only made small change in the loop when we forward propagate through decoder one step at a time:<\/p>\n\n<pre><code>if use_teacher_forcing:\n    for t in range(max_target_len):\n        decoder_output, decoder_hidden = decoder(\n            decoder_input, decoder_hidden, encoder_outputs\n        )\n        # Teacher forcing: next input is current target\n        decoder_input = target_variable[t].view(1, -1)\n\n        # Calculate and accumulate loss\n\n        loss = criterion(decoder_output, target_variable[t])\n        total_loss += loss\n        print_losses.append(loss.item())\n\n\n# Perform backpropatation\ntotal_loss.backward()\n\n# Clip gradients: gradients are modified in place\n_ = nn.utils.clip_grad_norm_(encoder.parameters(), clip)\n_ = nn.utils.clip_grad_norm_(decoder.parameters(), clip)\n\n# Adjust model weights\nencoder_optimizer.step()\ndecoder_optimizer.step()\n\nreturn sum(print_losses)\/max_target_len\n\nOutput:\nIteration: 1; Percent complete: 0.0%; Average loss: 8.5246\nIteration: 2; Percent complete: 0.1%; Average loss: 8.5403\nIteration: 3; Percent complete: 0.1%; Average loss: 8.5663\nIteration: 4; Percent complete: 0.1%; Average loss: 8.5691\nIteration: 5; Percent complete: 0.1%; Average loss: 8.4637\nIteration: 6; Percent complete: 0.1%; Average loss: 8.5378\nIteration: 7; Percent complete: 0.2%; Average loss: 8.5575\nIteration: 8; Percent complete: 0.2%; Average loss: 8.5145\nIteration: 9; Percent complete: 0.2%; Average loss: 8.5717\nIteration: 10; Percent complete: 0.2%; Average loss: 8.5122\nIteration: 11; Percent complete: 0.3%; Average loss: 8.5587\nIteration: 12; Percent complete: 0.3%; Average loss: 8.5491\nIteration: 13; Percent complete: 0.3%; Average loss: 8.5933\nIteration: 14; Percent complete: 0.4%; Average loss: 8.5231\nIteration: 15; Percent complete: 0.4%; Average loss: 8.5239\n<\/code><\/pre>\n\n<p>\u2026 and so on<\/p>\n\n<p>Total loss increase or fluctuate ate the same value, but it doesn\u2019t want to decrease. Did anybody try to use any kind of loss functions with ignore_index in seq2seq model with variate length sequences and has some positive results? nn.CrossEntropyLoss behaves similarly. I am starting to think that the loss with ignore_index parameter isn't calculated properly...<\/p>\n\n<p>I would like to emphasize that tensors passed to loss function have the following dimensions:<\/p>\n\n<pre><code>loss = criterion(decoder_output, target_variable[t])\n# decoder_output: (batch_size, vocab_size)\n# target_variable[t]: (batch_size)\n<\/code><\/pre>\n\n<p>I think that these are proper dimensions, but maybe I should reshape these tensors? In original code they calculate negative log manually passing to it \u201cconcatenated\u201d tensors with final shape: <\/p>\n\n<pre><code>    crossEntropy = -torch.log(torch.gather(inp, 1, target.view(-1, 1)).squeeze(1)) # shape (batch_size)\n<\/code><\/pre>\n\n<p>One additional question. In my version I used:<\/p>\n\n<pre><code>packed = nn.utils.rnn.pack_padded_sequence(embedded, input_lengths, batch_first=True, enforce_sorted=False)\n<\/code><\/pre>\n\n<p>thus I didn\u2019t sort batch in descending order according to sequence lengths. I noticed that my code process slower than original, could be that cause by using unsorted batch, or it should work well anyway?<\/p>\n\n<p>Any suggestion would be very appreciated.<\/p>\n\n<p>Best regards.<\/p>\n","answers":[]}
+{"QuestionId":57974040,"AnswerCount":1,"Tags":"<tensorflow><lambda><keras><layer>","CreationDate":"2019-09-17T12:17:04.470","AcceptedAnswerId":null,"OwnerUserId":11979471.0,"Title":"Error in Keras Lambda layer when wrapped function expects non-float argument","Body":"<p>I want to wrap a tensorflow function in a Keras Lambda layer as per the docs. However, my inputs are complex64. Here is a more complete example of the code i am using to replicate this behavior:<\/p>\n\n<pre><code>import numpy as np\nfrom keras.models import Model\nfrom keras.layers import Input, Lambda\nimport tensorflow as tf\nnp.set_printoptions(precision=3, threshold=3, edgeitems=3)\n\n\ndef layer0(inp):\n    z = inp[0] + inp[1]\n    num = tf.cast(tf.real(z), tf.complex64)\n    return z\/num\n\n\nif __name__ == \"__main__\":\n\n    shape = (1,10,5)\n    z1 = Input(shape=shape[1:], dtype=np.complex64)\n    z2 = Input(shape=shape[1:], dtype=np.complex64)\n\n    #s = Lambda(layer0, output_shape=shape)([z1, z2])\n    s = Lambda(layer0)([z1, z2])\n\n    model = Model(inputs=[z1,z2], outputs=s)\n\n    z1_in = np.asarray(np.random.normal(size=shape) + np.random.normal(size=shape)*1j, 'complex64')\n    z2_in = np.asarray(np.random.normal(size=shape) + np.random.normal(size=shape)*1j, 'complex64')\n\n    s_out = model.predict([z1_in, z2_in])\n    print(s_out)\n<\/code><\/pre>\n\n<p>which gives the following error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"complex_lambda.py\", line 32, in &lt;module&gt;\n    s = Lambda(layer0)([z1, z2])\n  File \"complex_lambda.py\", line 18, in layer0\n    return z\/num\nTypeError: x and y must have the same dtype, got tf.float32 != tf.complex64\n<\/code><\/pre>\n\n<p>However, if I use the commented line instead: \n<code>s = Lambda(layer0, output_shape=shape)([z1, z2])<\/code><\/p>\n\n<p>The code runs just fine. It seems that \"output_shape=(...)\" is necessary to make the division in the lambda function work. While this solution solves the problem for a single output variable, it doesn't work when having multiple outputs.<\/p>\n","answers":[{"AnswerId":"57974226","CreationDate":"2019-09-17T12:28:28.273","ParentId":null,"OwnerUserId":"12076271","Title":null,"Body":"<p>I cannot replicate your issue. Which version of <code>tensorflow<\/code> are you using? Are you using the <code>keras<\/code> package, or the <code>tensorflow.keras<\/code> submodule ?<\/p>\n\n<p>At any rate, I think you can fix your issue by specifying the <code>dtype<\/code> of the <code>Lambda<\/code> layer : <code>s = Lambda(lambda x: tf.math.real(x[0] + x[1]), dtype='complex64')([z1, s2])<\/code><\/p>\n"}]}
+{"QuestionId":57974098,"AnswerCount":0,"Tags":"<tensorflow><quantization>","CreationDate":"2019-09-17T12:20:28.503","AcceptedAnswerId":null,"OwnerUserId":7830286.0,"Title":"Errors occurred while converting 32-bit float TensorFlow model into 8-bit fixed TensorFlow model","Body":"<p>I am following the procedures listed in the github, Quantization-aware training, <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>.\nTo quantize my own TF model, <strong>landing_retrained_graph.pb<\/strong>, I fed it into the instructions of quantizations.<\/p>\n\n<pre><code>freeze_graph \\\n  --input_graph=landing_retrained_graph.pb \\\n  --input_checkpoint=checkpoint \\\n  --output_graph=landing_frozen_eval_graph.pb --output_node_names=outputs\n<\/code><\/pre>\n\n<p>Then the error pops up like below.<\/p>\n\n<pre><code>(ztdl) Jisoos-MacBook-Pro:tf_files jisooyu$ freeze_graph \\\n&gt;   --input_graph=landing_retrained_graph.pb \\\n&gt;   --input_checkpoint=checkpoint \\\n&gt;   --output_graph=landing_frozen_eval_graph.pb --output_node_names=outputs\n\n\nTraceback (most recent call last):\n  File \"\/Users\/jisooyu\/anaconda3\/envs\/ztdl\/bin\/freeze_graph\", line 11, in &lt;module&gt;\n    sys.exit(run_main())\n  File \"\/Users\/jisooyu\/anaconda3\/envs\/ztdl\/lib\/python3.7\/site-packages\/tensorflow\/python\/tools\/freeze_graph.py\", line 487, in run_main\n    app.run(main=my_main, argv=[sys.argv[0]] + unparsed)\n  File \"\/Users\/jisooyu\/anaconda3\/envs\/ztdl\/lib\/python3.7\/site-packages\/tensorflow\/python\/platform\/app.py\", line 40, in run\n    _run(main=main, argv=argv, flags_parser=_parse_flags_tolerate_undef)\n  File \"\/Users\/jisooyu\/anaconda3\/envs\/ztdl\/lib\/python3.7\/site-packages\/absl\/app.py\", line 300, in run\n    _run_main(main, args)\n  File \"\/Users\/jisooyu\/anaconda3\/envs\/ztdl\/lib\/python3.7\/site-packages\/absl\/app.py\", line 251, in _run_main\n    sys.exit(main(argv))\n  File \"\/Users\/jisooyu\/anaconda3\/envs\/ztdl\/lib\/python3.7\/site-packages\/tensorflow\/python\/tools\/freeze_graph.py\", line 486, in &lt;lambda&gt;\n    my_main = lambda unused_args: main(unused_args, flags)\n  File \"\/Users\/jisooyu\/anaconda3\/envs\/ztdl\/lib\/python3.7\/site-packages\/tensorflow\/python\/tools\/freeze_graph.py\", line 378, in main\n    flags.saved_model_tags, checkpoint_version)\n  File \"\/Users\/jisooyu\/anaconda3\/envs\/ztdl\/lib\/python3.7\/site-packages\/tensorflow\/python\/tools\/freeze_graph.py\", line 338, in freeze_graph\n    input_graph_def = _parse_input_graph_proto(input_graph, input_binary)\n  File \"\/Users\/jisooyu\/anaconda3\/envs\/ztdl\/lib\/python3.7\/site-packages\/tensorflow\/python\/tools\/freeze_graph.py\", line 253, in _parse_input_graph_proto\n    text_format.Merge(f.read(), input_graph_def)\n  File \"\/Users\/jisooyu\/anaconda3\/envs\/ztdl\/lib\/python3.7\/site-packages\/tensorflow\/python\/lib\/io\/file_io.py\", line 128, in read\n    pywrap_tensorflow.ReadFromStream(self._read_buf, length))\n  File \"\/Users\/jisooyu\/anaconda3\/envs\/ztdl\/lib\/python3.7\/site-packages\/tensorflow\/python\/lib\/io\/file_io.py\", line 98, in _prepare_value\n    return compat.as_str_any(val)\n  File \"\/Users\/jisooyu\/anaconda3\/envs\/ztdl\/lib\/python3.7\/site-packages\/tensorflow\/python\/util\/compat.py\", line 117, in as_str_any\n    return as_str(value)\n  File \"\/Users\/jisooyu\/anaconda3\/envs\/ztdl\/lib\/python3.7\/site-packages\/tensorflow\/python\/util\/compat.py\", line 87, in as_text\n    return bytes_or_text.decode(encoding)\n\nUnicodeDecodeError: 'utf-8' codec can't decode byte 0xe0 in position 55: invalid continuation byte\n<\/code><\/pre>\n\n<p>Any suggestions to fix the error are appreciated. <\/p>\n","answers":[]}
+{"QuestionId":57974213,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><deep-learning><tensorflow2.0>","CreationDate":"2019-09-17T12:27:53.533","AcceptedAnswerId":57974493.0,"OwnerUserId":10392393.0,"Title":"How can I combine two gradient tapes in TensorFlow 2.0","Body":"<p>How can I combine the two following gradient tape, into one:<\/p>\n\n<pre><code>x = tf.Variable(x, dtype=tf.float32)\nwith tf.GradientTape() as t:\n    m, v = DGP.predict(x)\n    dm_dx = t.gradient(m, x)\nwith tf.GradientTape() as t:\n    m, v = DGP.predict(x)\n    dv_dx = t.gradient(v, x)\n<\/code><\/pre>\n\n<p>Here's what I prefer but does not work the way I have written it:<\/p>\n\n<pre><code>with tf.GradientTape() as t:\n    m, v = DGP.predict(x)\n    dm_dx, dv_dx = t.gradient([m,v], x)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57974493","CreationDate":"2019-09-17T12:43:58.673","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You should be able to do this:<\/p>\n\n<pre><code>x = tf.Variable(x, dtype=tf.float32)\nwith tf.GradientTape(persistent=True) as t:\n    m, v = DGP.predict(x)\n    dm_dx = t.gradient(m, x)\n    dv_dx = t.gradient(v, x)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57974580,"AnswerCount":2,"Tags":"<python><machine-learning><keras><deep-learning>","CreationDate":"2019-09-17T12:49:29.757","AcceptedAnswerId":57982157.0,"OwnerUserId":8334684.0,"Title":"How should I approach a 300 classes classification machine learning problem?","Body":"<p>I am trying to make a Multi-Class classification application, but my dataset has 300 classes, is it possible to train my model with all these classes with a normal PC?<\/p>\n","answers":[{"AnswerId":"57974896","CreationDate":"2019-09-17T13:05:46.073","ParentId":null,"OwnerUserId":"10311011","Title":null,"Body":"<p>It is difficult to answer this question. The training time of your model depends on a number of factors. It might be best to train your model for a certain amount of hours and evaluate the performance. You could make use of fitting a learning curve, which could provide an esitmation of how many data points your require for training to achieve a certain performance. After that you could link the required amount of data points to computation time. <\/p>\n\n<p>Here is an article provides an algorithm for fitting a learning curve: <a href=\"https:\/\/www.ncbi.nlm.nih.gov\/pmc\/articles\/PMC3307431\/\" rel=\"nofollow noreferrer\">https:\/\/www.ncbi.nlm.nih.gov\/pmc\/articles\/PMC3307431\/<\/a>.<\/p>\n"},{"AnswerId":"57982157","CreationDate":"2019-09-17T21:15:11.390","ParentId":null,"OwnerUserId":"10542451","Title":null,"Body":"<p>Sure it is. You can even train imagenet with 1000 categories or more, if you have enough time! ;)<\/p>\n\n<p>You just have to think about which loss function you want (categorical crossentropy, sparse categorical crossentropy or even binary if you want to penalize each output node independently), apart from that there's not really much difference between 10, 100 or a 1000 classes.<\/p>\n\n<p>Of course you have to increase your model size to account for more classes, so RAM may be an issue, but then you can always decrease batch size. If you are using images and convnets and your model is still too large, try to downsample the images, use pooling layers or larger strides.<\/p>\n\n<p>If your computer is too old and slow, you can also try Google Colab which offers free GPU and even TPU online!<\/p>\n"}]}
+{"QuestionId":57974605,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><tf.keras><batch-normalization>","CreationDate":"2019-09-17T12:50:36.100","AcceptedAnswerId":57975860.0,"OwnerUserId":5048010.0,"Title":"How to implement Batch Normalization on tensorflow with Keras as a high-level API","Body":"<p>BatchNormalization (BN) operates slightly differently when in training and in inference. In training, it uses the average and variance of the current mini-batch to scale its inputs; this means that the exact result of the application of batch normalization depends not only on the current input, but also on all other elements of the mini-batch. This is clearly not desirable when in inference mode, where we want a deterministic result. Therefore, in that case, a fixed statistic of the global average and variance over the entire training set is used.<\/p>\n\n<p>In Tensorflow, this behavior is controlled by a boolean switch <code>training<\/code> that needs to be specified when calling the layer, see <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/BatchNormalization\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/BatchNormalization<\/a>. How do I deal with this switch when using Keras high-level API? Am I correct in assuming that it is dealt with automatically, depending whether we are using <code>model.fit(x, ...)<\/code> or <code>model.predict(x, ...)<\/code>?<\/p>\n\n<hr>\n\n<p>To test this, I have written this example. We start with a random distribution and we want to classify whether the input is positive or negative. However, we also have a test dataset coming from a different distribution where the inputs are displaced by 2 (and consequently the labels check whether x>2).<\/p>\n\n<pre><code>import numpy as np\nfrom math import ceil\nfrom tensorflow.python.data import Dataset\nfrom tensorflow.python.keras import Input, Model\nfrom tensorflow.python.keras.layers import Dense, BatchNormalization\n\nnp.random.seed(18)\nxt = np.random.randn(10_000, 1)\nyt = np.array([[int(x &gt; 0)] for x in xt])\ntrain_data = Dataset.from_tensor_slices((xt, yt)).shuffle(10_000).repeat().batch(32).prefetch(2)\n\nxv = np.random.randn(100, 1)\nyv = np.array([[int(x &gt; 0)] for x in xv])\nvalid_data = Dataset.from_tensor_slices((xv, yv)).repeat().batch(32).prefetch(2)\n\nxs = np.random.randn(100, 1) + 2\nys = np.array([[int(x &gt; 2)] for x in xs])\ntest_data = Dataset.from_tensor_slices((xs, ys)).repeat().batch(32).prefetch(2)\n\nx = Input(shape=(1,))\na = BatchNormalization()(x)\na = Dense(8, activation='sigmoid')(a)\na = BatchNormalization()(a)\ny = Dense(1, activation='sigmoid')(a)\nmodel = Model(inputs=x, outputs=y, )\nmodel.summary()\n\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\nmodel.fit(train_data, epochs=10, steps_per_epoch=ceil(10_000 \/ 32), validation_data=valid_data,\n          validation_steps=ceil(100 \/ 32))\nzs = model.predict(test_data, steps=ceil(100 \/ 32))\nprint(sum([ys[i] == int(zs[i] &gt; 0.5) for i in range(100)]))\n<\/code><\/pre>\n\n<p>Running the code prints the value 0.5, meaning that half the examples are labeled properly. This is what I would expect if the system was using the global statistics on the training set to implement BN.<\/p>\n\n<p>If we change the BN layers to read<\/p>\n\n<pre><code>x = Input(shape=(1,))\na = BatchNormalization()(x, training=True)\na = Dense(8, activation='sigmoid')(a)\na = BatchNormalization()(a, training=True)\ny = Dense(1, activation='sigmoid')(a)\n<\/code><\/pre>\n\n<p>and run the code again we find 0.87. Forcing always the training state, the percentage of correct prediction has changed. This is consistent with the idea that <code>model.predict(x, ...)<\/code> is now using the statistic of the mini-batch to implement BN, and is therefore able to slightly \"correct\" the mismatch in the source distributions between training and test data.<\/p>\n\n<p>Is that correct?<\/p>\n","answers":[{"AnswerId":"57975860","CreationDate":"2019-09-17T13:57:02.680","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>If I'm understanding your question correctly, then yes, keras does automatically manage training vs inference behavior based on <code>fit<\/code> vs <code>predict<\/code>\/<code>evaluate<\/code>. The flag is called <code>learning_phase<\/code>, and it determines the behavior of batch norm, dropout, and potentially other things. The current learning phase can be seen with <code>keras.backend.learning_phase()<\/code>, and set with <code>keras.backend.set_learning_phase()<\/code>.<\/p>\n\n<p><a href=\"https:\/\/keras.io\/backend\/#learning_phase\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/backend\/#learning_phase<\/a><\/p>\n"}]}
+{"QuestionId":57975375,"AnswerCount":0,"Tags":"<python-3.x><keras><lstm>","CreationDate":"2019-09-17T13:31:07.477","AcceptedAnswerId":null,"OwnerUserId":5105541.0,"Title":"input 3d data to lstm","Body":"<p>I want to fit model on 3D data.<\/p>\n\n<pre><code>Y = pd.get_dummies(tag_data['tag']).values\nX_train, X_test, Y_train, Y_test = train_test_split(X,Y, test_size = 0.20, random_state = 42)\nprint(X_train.shape,Y_train.shape)\nprint(X_test.shape,Y_test.shape)\nbatch_size = 32\nmodel.fit(X_train, Y_train, epochs = 10, batch_size=batch_size, verbose = 2)\n<\/code><\/pre>\n\n\n\n<pre><code>(7, 273, 10) (7, 2)\n(2, 273, 10) (2, 2)\n\nValueError: Error when checking input: expected embedding_2_input to have 2 dimensions, but got array with shape (7, 273, 10)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57975576,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><deep-learning>","CreationDate":"2019-09-17T13:41:16.337","AcceptedAnswerId":null,"OwnerUserId":4565802.0,"Title":"`Checkpoint` was expecting a trackable object (an object derived from `TrackableBase`), got name: \"Adam_1\"","Body":"<p>I am using TensorFlow 1.14 to create a GAN model, but I get the following error <\/p>\n\n<pre><code>ValueError: `Checkpoint` was expecting a trackable object (an object derived from `TrackableBase`), got name: \"Adam_1\"\nop: \"NoOp\"\ninput: \"^generator\/batch_trans_conv1\/AssignMovingAvg\"\ninput: \"^generator\/batch_trans_conv1\/AssignMovingAvg_1\"\ninput: \"^generator\/batch_trans_conv2\/AssignMovingAvg\"\ninput: \"^generator\/batch_trans_conv2\/AssignMovingAvg_1\"\ninput: \"^generator\/batch_trans_conv3\/AssignMovingAvg\"\ninput: \"^generator\/batch_trans_conv3\/AssignMovingAvg_1\"\ninput: \"^generator\/batch_trans_conv4\/AssignMovingAvg\"\ninput: \"^generator\/batch_trans_conv4\/AssignMovingAvg_1\"\ninput: \"^Adam_1\/update_generator\/dense\/kernel\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/dense\/bias\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/trans_conv1\/kernel\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/trans_conv1\/bias\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/batch_trans_conv1\/gamma\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/batch_trans_conv1\/beta\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/trans_conv2\/kernel\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/trans_conv2\/bias\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/batch_trans_conv2\/gamma\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/batch_trans_conv2\/beta\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/trans_conv3\/kernel\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/trans_conv3\/bias\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/batch_trans_conv3\/gamma\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/batch_trans_conv3\/beta\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/trans_conv4\/kernel\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/trans_conv4\/bias\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/batch_trans_conv4\/gamma\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/batch_trans_conv4\/beta\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/logits\/kernel\/ApplyAdam\"\ninput: \"^Adam_1\/update_generator\/logits\/bias\/ApplyAdam\"\ninput: \"^Adam_1\/Assign\"\ninput: \"^Adam_1\/Assign_1\"\n<\/code><\/pre>\n\n<p>I am using the AdamOptimizer and checkpointing the generator optimizer and discriminator optimizer using the following code<\/p>\n\n<pre><code>LR_D = 0.00004\nLR_G = 0.0004\nBETA1 = 0.5\n\nd_train_opt = tf.train.AdamOptimizer(learning_rate=LR_D, beta1=BETA1)\ng_train_opt = tf.train.AdamOptimizer(learning_rate=LR_G, beta1=BETA1)  \ncheckpoint_dir = 'training_checkpoints'\ncheckpoint_prefix = os.path.join(checkpoint_dir, \"ckpt\") \n\ncheckpoint =  tf.train.Checkpoint(generator_optimizer=g_train_opt, discriminator_optimizer=d_train_opt)  \n<\/code><\/pre>\n\n<p><strong>How to save the AdamOptimizer values using checkpoints?<\/strong><\/p>\n","answers":[]}
+{"QuestionId":57975591,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-17T13:42:11.697","AcceptedAnswerId":57975782.0,"OwnerUserId":3353601.0,"Title":"How to assign to pixels in an image in eager tensorflow?","Body":"<p>I have a need to rasterize points into image form inside my model, and I can't figure out the right way to do this. The issues are 1) tensorflow doesn't handle \"assignment\" to tensors very intuitively and 2) eager mode doesn't handle creation of tensors very intuitively<\/p>\n\n<p>Here's what I'm trying to do<\/p>\n\n<pre><code>rope_images = tfe.Variable(\n            initial_value=lambda: tf.zeros([batch_size, self.sdf_shape[0], self.sdf_shape[1], self.n_points]),\n            name='rope_images',\n            trainable=False)\nrow_y_indeces = tf.cast(points[:, :, 1] \/ resolution[:, 0:1] + origin[:, 0:1], tf.int64)\ncol_x_indeces = tf.cast(points[:, :, 0] \/ resolution[:, 1:2] + origin[:, 1:2], tf.int64)\nbatch_indeces = tf.reshape(tf.tile(tf.reshape(tf.range(batch_size), [-1, 1]), [1, self.n_points]), [-1])\nrow_indeces = tf.squeeze(row_y_indeces)\ncol_indeces = tf.squeeze(col_x_indeces)\npoint_channel_indeces = tf.tile(tf.range(self.n_points, dtype=tf.int64), [batch_size])\nindeces = tf.transpose(tf.stack([batch_indeces, row_indeces, col_indeces, point_channel_indeces]))\non_pixels = tf.gather_nd(rope_images, indeces)\ntf.assign(on_pixels, 1)\n\n<\/code><\/pre>\n\n<p>But I get the error:<\/p>\n\n<pre><code>AttributeError: 'Tensor' object has no attribute 'assign'\n<\/code><\/pre>\n\n<p>What's the right way to do this?<\/p>\n","answers":[{"AnswerId":"57975782","CreationDate":"2019-09-17T13:52:17.227","ParentId":null,"OwnerUserId":"12076271","Title":null,"Body":"<p>You should be able to turn your Tensor into a ResourceVariable using <code>tf.Variable<\/code> (you will possibly want to specify that <code>trainable=False<\/code>), which supports assignement.<\/p>\n\n<p>If you want to replace some values within a Tensor, you can also use <code>tf.where<\/code>, <em>e.g.<\/em> if you want to replace the <code>indeces<\/code> values of <code>rope_images<\/code> with <code>1<\/code>: <code>rope_images = tf.where(indeces, 1, rope_images)<\/code>, which effectively creates a new Tensor with either the retained initial values or the substitution ones. Note that you can replace <code>1<\/code> with a full-shape Tensor of substitute values, in contexts when you need more flexibility.<\/p>\n\n<p>I hope this helps :-)<\/p>\n"}]}
+{"QuestionId":57975738,"AnswerCount":0,"Tags":"<tensorflow><jupyter-notebook><google-colaboratory>","CreationDate":"2019-09-17T13:49:27.443","AcceptedAnswerId":null,"OwnerUserId":12078003.0,"Title":"How to run a specific code from Google Colaboratory in Jupyter Notebook?","Body":"<p>I am currently doing a Tensorflow course on Udacity. The code is executed there in Google Colaboratory.<\/p>\n\n<p>Now I tryed to run the code locally on Jupyter Notebook. By doing this I get an error message. <\/p>\n\n<p>The code is:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    from __future__ import absolute_import, division, print_function, unicode_literals\n\n    import os\n    import matplotlib.pyplot as plt\n    import numpy as np\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>    import tensorflow as tf\n    from tensorflow.keras.preprocessing.image import ImageDataGenerator\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>    import logging\n    logger = tf.get_logger()\n    logger.setLevel(logging.ERROR)\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>    _URL = 'https:\/\/storage.googleapis.com\/mledu-datasets\/cats_and_dogs_filtered.zip'\n    zip_dir = tf.keras.utils.get_file('cats_and_dogs_filterted.zip', origin=_URL, extract=True)\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>    zip_dir_base = os.path.dirname(zip_dir)\n    !find $zip_dir_base -type d -print\n<\/code><\/pre>\n\n<p>In Google Colaboratory it runs fine. On Jupyter Notebook I get after the last part the error message that it is the false date type.<\/p>\n\n<p>What does the following line means?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    !find $zip_dir_base -type d -print\n<\/code><\/pre>\n\n<p>What do I have to change to run it in Jupyter Notebook?<\/p>\n","answers":[]}
+{"QuestionId":57975763,"AnswerCount":1,"Tags":"<keras><loss-function>","CreationDate":"2019-09-17T13:51:37.140","AcceptedAnswerId":57976403.0,"OwnerUserId":6296435.0,"Title":"Calculate loss in Keras without running the model","Body":"<p>Is there a way to get the loss of the model, with it's current weights, without running evaluate, or fit, on it?<\/p>\n\n<pre><code>model = keras.Sequential([\n    keras.layers.Input(400),\n    keras.layers.Dense(25, activation=tf.nn.sigmoid, kernel_regularizer=regularizers.l2(lambd)),\n    keras.layers.Dense(10, activation=tf.nn.softmax)\n])\nmodel.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy',  # does the 1-hot encoding for us\n              metrics=['accuracy'])\nmodel.set_weights([Theta1.T, np.zeros(25), Theta2.T, np.zeros(10)])\nprob = model.predict(X)\npred = np.argmax(prob, axis=1).reshape(-1, 1)\npred_y = pred == y\nprint(f'Training Set Accuracy: {np.mean(pred_y)*100:.2f}%') \n# How do I get the loss now?\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/stackoverflow.com\/a\/46863768\/6296435\">This<\/a> didn't work for me.<\/p>\n","answers":[{"AnswerId":"57976403","CreationDate":"2019-09-17T14:23:50.967","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>You can, via passing the outputs of <code>model.predict(x)<\/code> to an implementation of the loss function. In addition, you'll need a function to compute model regularization losses - <code>reg_loss(model)<\/code>. Below is an implementation of <code>binary_crossentropy<\/code>, and <code>l1<\/code>, <code>l2<\/code>, and <code>l1_l2<\/code> losses from all layers, including recurrent - but does <em>not<\/em> include <code>activity_regularizer<\/code> losses, which aren't weight losses. You can replace <code>binary_crossentropy<\/code> with your own function - e.g. <code>sparse_softmax_crossentropy_with_logits<\/code>:<\/p>\n\n\n\n<p><hr>\n<strong>WORKING IMPLEMENTATION:<\/strong> <em>(numerically stable version)<\/em><\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>def binary_crossentropy(y_true, y_pred, sample_weight=1):\n    if len(y_pred.shape)==1:\n        y_pred = np.atleast_2d(y_pred).T\n    y_pred = [max(min(pred[0], 1-K.epsilon()), K.epsilon()) for pred in y_pred]\n    y_true,y_pred,sample_weight = force_2d_shape([y_true,y_pred,sample_weight])\n\n    logits = np.log(y_pred) - np.log(1-y_pred) # sigmoid inverse\n    neg_abs_logits = -np.abs(logits)\n    relu_logits    = (logits &gt; 0)*logits\n\n    loss_vec = relu_logits - logits*y_true + np.log(1 + np.exp(neg_abs_logits))\n    return np.mean(sample_weight*loss_vec)\n\ndef force_2d_shape(arr_list):\n    for arr_idx, arr in enumerate(arr_list):\n        if len(np.array(arr).shape) != 2:\n            arr_list[arr_idx] = np.atleast_2d(arr).T\n    return arr_list\n<\/code><\/pre>\n\n<pre class=\"lang-python prettyprint-override\"><code>def l1l2_weight_loss(model):\n    l1l2_loss = 0\n    for layer in model.layers:\n        if 'layer' in layer.__dict__ or 'cell' in layer.__dict__:\n            l1l2_loss += _l1l2_rnn_loss(layer)\n            continue\n\n        if 'kernel_regularizer' in layer.__dict__ or \\\n           'bias_regularizer'   in layer.__dict__:\n            l1l2_lambda_k, l1l2_lambda_b = [0,0], [0,0] # defaults\n            if layer.__dict__['kernel_regularizer'] is not None:\n                l1l2_lambda_k = list(layer.kernel_regularizer.__dict__.values())\n            if layer.__dict__['bias_regularizer']   is not None:\n                l1l2_lambda_b = list(layer.bias_regularizer.__dict__.values())\n\n            if any([(_lambda != 0) for _lambda in (l1l2_lambda_k + l1l2_lambda_b)]):\n                W = layer.get_weights()\n\n                for idx,_lambda in enumerate(l1l2_lambda_k + l1l2_lambda_b):\n                    if _lambda != 0:\n                        _pow = 2**(idx % 2) # 1 if idx is even (l1), 2 if odd (l2)\n                        l1l2_loss += _lambda*np.sum(np.abs(W[idx\/\/2])**_pow)\n    return l1l2_loss\n<\/code><\/pre>\n\n<pre class=\"lang-python prettyprint-override\"><code>def _l1l2_rnn_loss(layer):\n    l1l2_loss = 0\n    if 'backward_layer' in layer.__dict__:\n        bidirectional = True\n        _layer = layer.layer\n    else:\n        _layer = layer\n        bidirectional = False\n    ldict = _layer.cell.__dict__\n\n    if 'kernel_regularizer'    in ldict or \\\n       'recurrent_regularizer' in ldict or \\\n       'bias_regularizer'      in ldict:\n        l1l2_lambda_k, l1l2_lambda_r, l1l2_lambda_b = [0,0], [0,0], [0,0]\n        if ldict['kernel_regularizer']    is not None:\n            l1l2_lambda_k = list(_layer.kernel_regularizer.__dict__.values())\n        if ldict['recurrent_regularizer'] is not None:\n            l1l2_lambda_r = list(_layer.recurrent_regularizer.__dict__.values())\n        if ldict['bias_regularizer']      is not None:\n            l1l2_lambda_b = list(_layer.bias_regularizer.__dict__.values())\n\n        all_lambda = l1l2_lambda_k + l1l2_lambda_r + l1l2_lambda_b\n        if any([(_lambda != 0) for _lambda in all_lambda]):\n            W = layer.get_weights()\n            idx_incr = len(W)\/\/2 # accounts for 'use_bias'\n\n            for idx,_lambda in enumerate(all_lambda):\n                if _lambda != 0:\n                    _pow = 2**(idx % 2) # 1 if idx is even (l1), 2 if odd (l2)\n                    l1l2_loss += _lambda*np.sum(np.abs(W[idx\/\/2])**_pow)\n                    if bidirectional:\n                        l1l2_loss += _lambda*np.sum(\n                                    np.abs(W[idx\/\/2 + idx_incr])**_pow)\n        return l1l2_loss  \n<\/code><\/pre>\n\n<p><hr>\n<b>TESTING IMPLEMENTATION:<\/b><\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>from keras.layers import Input, Dense, LSTM, GRU, Bidirectional\nfrom keras.models import Model\nfrom keras.regularizers import l1, l2, l1_l2\nimport numpy as np \n\nipt   = Input(shape=(1200,16))\nx     = LSTM(60, activation='relu', return_sequences=True,\n                                                 recurrent_regularizer=l2(1e-3),)(ipt)\nx     = Bidirectional(GRU(60, activation='relu', bias_regularizer     =l1(1e-4)))(x)\nout   = Dense(1,  activation='sigmoid',          kernel_regularizer   =l1_l2(2e-4))(x)\nmodel = Model(ipt,out)\n\nmodel.compile(loss='binary_crossentropy', optimizer='adam')\n<\/code><\/pre>\n\n<pre class=\"lang-python prettyprint-override\"><code>X = np.random.rand(10,1200,16) # (batch_size, timesteps, input_dim)\nY = np.random.randint(0,2,(10,1))\nclass_weights = {'0':1, '1': 6}\nsample_weights = np.array([class_weights[str(label[0])] for label in Y])\n<\/code><\/pre>\n\n<pre class=\"lang-python prettyprint-override\"><code>keras_loss   = model.evaluate(X,Y,sample_weight=sample_weights)\ncustom_loss  = binary_crossentropy(Y, model.predict(X))\ncustom_loss += l1l2_weight_loss(model)\n\nprint('%.6f'%keras_loss  + ' -- keras_loss')\nprint('%.6f'%custom_loss + ' -- custom_loss') \n<\/code><\/pre>\n\n<p><b>\n0.763822 -- keras_loss<br>\n0.763822 -- custom_loss\n<\/b><\/p>\n"}]}
+{"QuestionId":57975958,"AnswerCount":0,"Tags":"<python><neural-network><conv-neural-network><pytorch><recurrent-neural-network>","CreationDate":"2019-09-17T14:01:07.127","AcceptedAnswerId":null,"OwnerUserId":10782229.0,"Title":"CRNN model for word data not learning anything after 10 epochs on 932 train images","Body":"<p>I am trying to predict words from images of Bangla words using CRNN approach. I am training it in a local environment with 932 images and 10 epochs. The loss is decreasing but accuracy is 0% . The model is kinda doing random guessing , and not learning anything . Following is my model :<\/p>\n\n<pre><code>import torch.nn as nn\n\n\nclass BidirectionalLSTM(nn.Module):\n\n    def __init__(self, number_of_input, number_of_hidden, nunmer_of_out):\n        super(BidirectionalLSTM, self).__init__()\n\n        self.rnn = nn.LSTM(number_of_input, number_of_hidden, bidirectional=True)\n        self.embedding = nn.Linear(number_of_hidden * 2, nunmer_of_out)\n\n    def forward(self, input):\n        recurrent, _ = self.rnn(input)\n        T, b, h = recurrent.size()\n        # print(\"size from recurrent forward :{} {} {}\".format(T,b,h))\n        t_rec = recurrent.view(T * b, h)\n        # print(\"size after recurrent view : {}\".format(t_rec.size()))\n        output = self.embedding(t_rec)  # [T * b, nOut]\n\n        output = output.view(T, b, -1)\n        return output\n\n\nclass CRNN(nn.Module):\n\n    def __init__(self, imgH, nc, nclass, nh, n_rnn=2, leakyRelu=False):\n        super(CRNN, self).__init__()\n        # assert imgH % 16 == 0, 'imgH has to be a multiple of 16'\n\n        ks = [3, 3, 3, 3, 3, 3, 2]  # original [3, 3, 3, 3, 3, 3, 2]\n        ps = [1, 1, 1, 1, 1, 1, 0]\n        ss = [1, 1, 1, 1, 1, 1, 1]\n        nm = [64, 128, 256, 256, 512, 512, 512]\n\n        cnn = nn.Sequential()\n\n        def conv_relu(i, batchNormalization=False):\n            nIn = nc if i == 0 else nm[i - 1]\n            nOut = nm[i]\n            cnn.add_module('conv{0}'.format(i),\n                           nn.Conv2d(nIn, nOut, ks[i], ss[i], ps[i]))\n            if batchNormalization:\n                cnn.add_module('batchnorm{0}'.format(i), nn.BatchNorm2d(nOut))\n            if leakyRelu:\n                cnn.add_module('relu{0}'.format(i),\n                               nn.LeakyReLU(0.2, inplace=True))\n            else:\n                cnn.add_module('relu{0}'.format(i), nn.ReLU(True))\n\n        conv_relu(0)\n        cnn.add_module('pooling{0}'.format(0), nn.MaxPool2d(2, 2))  # 64x16x64 original\n        conv_relu(1)\n        cnn.add_module('pooling{0}'.format(1), nn.MaxPool2d(2, 2))  # 128x8x32 original\n        conv_relu(2, True)\n        conv_relu(3)\n        cnn.add_module('pooling{0}'.format(2),\n                       nn.MaxPool2d((2, 2), (2, 1), (0, 1)))  # 256x4x16 original\n        conv_relu(4, True)\n        conv_relu(5)\n        cnn.add_module('pooling{0}'.format(3),\n                       nn.MaxPool2d((2, 2), (2, 1), (0, 1)))  # 512x2x16 original\n        conv_relu(6, True)  # 512x1x16\n        # Extra max pull to bring down height to 1\n        cnn.add_module('pooling{0}'.format(4),\n                       nn.MaxPool2d((3, 3), (2, 4), (0, 1))) # final size 512x2x32\n        self.cnn = cnn\n        self.rnn = nn.Sequential(\n            BidirectionalLSTM(512, nh, nh),\n            BidirectionalLSTM(nh, nh, nclass))\n\n    def forward(self, input):\n        # conv features\n        conv = self.cnn(input)\n        b, c, h, w = conv.size()\n        # assert h == 1, \"the height of conv must be 1\"\n        # print('size before squeeze')\n        # print(conv.size())\n        conv = conv.squeeze(2)\n        # print('size after squeeze')\n        # print(conv.size())\n        conv = conv.permute(2, 0, 1)  # [w, b, c]\n\n        # rnn features\n        output = self.rnn(conv)\n        # print(\"from model {}\".format(output))\n        return output\n\n<\/code><\/pre>\n\n<p>And following is my main :<\/p>\n\n<pre><code>from data_preparation.dataloader import DataSetOCR\nfrom utills.dataloader_services import *\nfrom torch.utils.data import DataLoader\nimport parameters\nfrom models import crnn\nimport torch.optim as optim\nfrom torch.autograd import Variable\nfrom torch.nn import CTCLoss\nfrom utills.string_label_converter import averager, StrLabelConverter\nimport torch.nn.functional as F\n\ntrain_dataset = DataSetOCR(\n    csv_file_path= parameters.train_csv_path,\n    text_file_path= parameters.text_file_path,\n    root_directory= parameters.train_root)\nassert train_dataset\n\ntest_dataset = DataSetOCR(\n    csv_file_path= parameters.test_csv_path,\n    text_file_path= parameters.text_file_path,\n    root_directory= parameters.test_root)\nassert test_dataset\n\ndataloader_params = {\n    'batch_size': 2,\n    'shuffle': True,\n    'collate_fn': my_collate\n\n}\n\ntrain_loader = DataLoader(train_dataset, **dataloader_params)\ntrain_iter = iter(train_loader)\n\n# custom weights initialization called on crnn\ndef weights_init(m):\n    classname = m.__class__.__name__\n    if classname.find('Conv') != -1:\n        m.weight.data.normal_(0.0, 0.02)\n    elif classname.find('BatchNorm') != -1:\n        m.weight.data.normal_(1.0, 0.02)\n        m.bias.data.fill_(0)\n\n\ncrnn = crnn.CRNN(parameters.max_image_height, 3, parameters.number_of_classes, 256)\ncrnn.apply(weights_init)\ncriterion = CTCLoss()\noptimizer = optim.Adam(crnn.parameters(), lr=0.001)\nloss_avg = averager()\n\nimage = torch.FloatTensor(2, 3, parameters.max_image_width, parameters.max_image_height)\ntext = torch.LongTensor(2 * 5)\nlength = torch.LongTensor(2)\n\n\nstring_converter = StrLabelConverter()\n# string_converter.convert_integer_to_string()\n\ndef val(net, dataset, criterion, max_iter=100):\n    print('Start val')\n\n    for p in crnn.parameters():\n        p.requires_grad = False\n\n    net.eval()\n    data_loader = torch.utils.data.DataLoader(\n        dataset, shuffle=True, batch_size=2, collate_fn=my_collate)\n    val_iter = iter(data_loader)\n\n    i = 0\n    n_correct = 0\n    loss_avg = averager()\n\n    max_iter = min(max_iter, len(data_loader))\n    # print('length of dataloader')\n    # print(len(data_loader))\n    for i in range(max_iter):\n        data = val_iter.next()\n        i += 1\n        images, texts = data\n        batch_size = images.size(0)\n        # print(images)\n        # print(image)\n        loadData(image, images)\n        t, l = string_converter.convert_string_to_integer(texts, [])\n        loadData(text, t)\n        loadData(length, l)\n        # print('actual label')\n        # print(text)\n        # print(text.size())\n\n\n        # print('input from val')\n        # print(i)\n        print(image)\n        preds = crnn(image)\n        preds = F.log_softmax(preds, 2)\n        print(preds)\n        preds_size = Variable(torch.IntTensor([preds.size(0)] * batch_size))\n        # print('predictions')\n        # print(preds.size())\n        cost = criterion(preds, text, preds_size, length) \/ batch_size\n        print('validation {}'.format(cost))\n        loss_avg.add(cost)\n\n        # print('predictions before max')\n        # print(preds.size())\n        _, preds = preds.max(2)\n        # print('after max function')\n        # print(preds)\n        # print(preds.size())\n        preds = preds.squeeze(1) # originial 2\n        # print('after squeeze')\n        # print(preds.size())\n        preds = preds.transpose(1, 0).contiguous().view(-1)\n        # print('after transpose')\n        # print(preds.size())\n        sim_preds = string_converter.convert_integer_to_string(preds.data, preds_size.data)\n\n        cpu_texts = string_converter.convert_integer_to_string(text, length)\n\n        # for pred, target in zip(sim_preds, cpu_texts):\n        #     # if pred == target:\n        #     #     n_correct += 1\n        #     print((pred, target))\n    #\n    # raw_preds = string_converter.convert_integer_to_string(preds.data, preds_size.data)[:2]\n    # for raw_pred, pred, gt in zip(raw_preds, sim_preds, cpu_texts):\n    #     print('%-20s =&gt; %-20s, gt: %-20s' % (raw_pred, pred, gt))\n    #\n    # accuracy = n_correct \/ float(max_iter * 100)\n    # print('Test loss: %f, accuray: %f' % (loss_avg.val(), accuracy))\n\ndef trainBatch(train_iter, criterion, optimizer):\n    try:\n        data = train_iter.next()\n    except StopIteration:\n        train_iter = iter(train_loader)\n        data = train_iter.next()\n\n    images,texts = data\n    loadData(image,images)\n    t,l =string_converter.convert_string_to_integer(texts, [])\n    loadData(text, t)\n    loadData(length,l)\n    batch_size = dataloader_params['batch_size']\n    optimizer.zero_grad()\n    # # test_util.loadData(image, data['image'])\n    # text = Variable(data['integer_sequence_label'])\n    # print(\"from train {}\".format(image))\n    preds = crnn(image)\n    preds = F.log_softmax(preds, 2)\n    preds_size = Variable(torch.IntTensor([preds.size(0)] * batch_size))\n\n\n\n    cost = criterion(preds, text, preds_size, length) \/ batch_size\n    # print('cost from train {}'.format(cost))\n    crnn.zero_grad()\n    cost.backward()\n    optimizer.step()\n    return cost\n\nfor epoch in range(25):\n    i = 0\n    while i &lt; len(train_loader):\n        for p in crnn.parameters():\n            p.requires_grad = True\n        crnn.train()\n\n        cost = trainBatch(train_iter, criterion, optimizer)\n        loss_avg.add(cost)\n        i += 1\n\n        if i % 50 == 0:\n            print('[%d\/%d][%d\/%d] Loss: %f' % (epoch, 25, i, len(train_loader), loss_avg.val()))\n            loss_avg.reset()\n\n        if i % 50 == 0:\n            val(crnn, test_dataset, criterion)\n\n        # #do checkpointing\n        # if i % 1 == 0:\n        #     torch.save(crnn.state_dict(), '{0}\/netCRNN_{1}_{2}.pth'.format('\/home\/bjit-531\/PycharmProjects\/python\/bangla-ocr-version-2\/bangla-ocr-version-2\/weights\/', epoch, i))\n        #\n<\/code><\/pre>\n\n<p>I am generating the images in the data set with various fonts so they have different height and width. I am padding them during data loading , so the model receives each image with size 90 x 752 . \nWhat am i doing wrong here? Why the model accuracy isn't improving ?<\/p>\n","answers":[]}
+{"QuestionId":57975995,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-17T14:03:14.187","AcceptedAnswerId":58024997.0,"OwnerUserId":11135519.0,"Title":"Train two consecutive models in tensorflow","Body":"<p>I am trying to build a model in tensorflow, while I use two consecutive models. Unfortunately I can't include them within one model. The first Model is basically an encoder, the second returns my needed value.<\/p>\n\n<pre><code>out = Model_a(image_input)\nvalue = Model_b(out)\nloss = f(value)\n<\/code><\/pre>\n\n<p>I can train Model_b using the given loss, but would then need the gradients of the first layer (of Model_b) regarding the loss to proceed for the gradient calculation in Model_a. Furthermore I would need somehow a function that calculates the gradients based on these gradients, instead of a loss function. Does anyone have an idea if tensorflow already has such functionality or had to tackle similar problems?<\/p>\n\n<p>Cheers<\/p>\n","answers":[{"AnswerId":"58024997","CreationDate":"2019-09-20T08:51:36.487","ParentId":null,"OwnerUserId":"11135519","Title":null,"Body":"<p>I found a working solution, for any who has similar problems. Using Tensorflow 2.0 and the keras eager-mode (using GradientTape) one can construct any loss function as desired, even including consecutive models. Important is that the predict function will not work, one needs to use the direct call.\nNow the gradients can be calculated for each model regarding that loss function, which seems to work so far. Important is that the models itself are included in the loss function and not a copy of the output or at least the copy is generated within the Tape environment. An example is found below:<\/p>\n\n<pre><code>optimizer = tf.keras.optimizers.Adam(lr=0.1)\nwith tf.GradientTape(persistent=True) as tape:\n    error = (model2(model1(x)) - y)\n    loss_value = tf.reduce_mean(tf.square(error))\ngradients1 = tape.gradient(loss_value, model1.variables)\ngradients2 = tape.gradient(loss_value, model2.variables)\noptimizer.apply_gradients(zip(gradients1, model1.variables))\noptimizer.apply_gradients(zip(gradients2, model2.variables))\n<\/code><\/pre>\n\n<p>If anyone finds a more efficient or \"prettier\" solution I would be happy if he\/she shares it.<\/p>\n\n<p>Cheers<\/p>\n"}]}
+{"QuestionId":57976012,"AnswerCount":1,"Tags":"<python><tensorflow><keras><reinforcement-learning><reward>","CreationDate":"2019-09-17T14:04:04.537","AcceptedAnswerId":null,"OwnerUserId":10515022.0,"Title":"Why is my reward function returning None in Python?","Body":"<p>OK, so, I'm trying to make a intrinsic-curiosity agent using keras and tensorflow. This agent's reward function is the difference of an autoencoder's loss between the previous and current state, and the autoencoder's loss between the current state and an imagined next state. However this reward function always returns None instead of the actual difference. I've tried printing the loss out but it always gives the correct values.<\/p>\n\n<p>Reward function\/replay code:<\/p>\n\n<pre><code>    def replay(self, batch):\n        minibatch = R.sample(self.memory, batch)\n\n        for prev_state, actions, state, reward, imagined_next_state in minibatch:\n            target = []\n\n            imagined_next_state = np.add(np.random.random(self.state_size), imagined_next_state)\n            target_m = self.model.predict(state)\n            for i in range(len(target_m)):\n                target_m[i][0][actions[i]]=reward\n\n            history_m = self.model.fit(state, target_m, epochs=1, verbose=0)\n            history_ae_ps = self.autoencoder.fit(prev_state, state, epochs=1, verbose=0)\n            history_ae_ns = self.autoencoder.fit(state, imagined_next_state, epochs=1, verbose=0)\n\n            loss_m = history_m.history['loss'][-1]\n            loss_ae_ps = history_ae_ps.history['loss'][-1]\n            loss_ae_ns = history_ae_ns.history['loss'][-1]\n            print(\"LOSS AE PS:\", loss_ae_ps)\n            print(\"LOSS AE NS:\", loss_ae_ns)\n\n            loss_ae = loss_ae_ns - loss_ae_ps\n            print(reward, loss_ae)\n            return loss_ae\n<\/code><\/pre>\n\n<p>Agent environment loop code:<\/p>\n\n<pre><code>    def loop(self, times='inf'):\n        if times is 'inf':\n            times = 2**31\n\n        reward = 0.0001\n        prev_shot = self.get_shot()\n\n        for i in range(times):\n            acts, ins, act_probs, shot = self.get_act()\n\n            act_0 = acts[0]\n            act_1 = acts[1]\n            act_2 = acts[2]\n            act_3 = acts[3]\n\n            self.act_to_mouse(act_0, act_1)\n            self.act_to_click(act_2)\n            self.act_to_keys(act_3)\n\n            reward = self.remember_and_replay(prev_shot, acts, shot, reward, ins)\n            if reward is None:\n                raise(RewardError(\"Rewards are none.\"))\n            prev_shot = shot\n<\/code><\/pre>\n","answers":[{"AnswerId":"57976013","CreationDate":"2019-09-17T14:04:04.537","ParentId":null,"OwnerUserId":"10515022","Title":null,"Body":"<p>I just solved it while typing the question. I simply wasn't returning the reward in the remember_and_replay method... <\/p>\n\n<p>The remember_and_replay method looked like this:<\/p>\n\n<pre><code>def remember_and_replay(self, prev_shot, action, shot, reward, ins):\n        self.dqn.remember(prev_shot, action, shot, reward, ins)\n        self.dqn.replay(1)\n\n<\/code><\/pre>\n\n<p>when it should have been like this:<\/p>\n\n<pre><code>def remember_and_replay(self, prev_shot, action, shot, reward, ins):\n        self.dqn.remember(prev_shot, action, shot, reward, ins)\n        rew = self.dqn.replay(1)\n        return rew\n<\/code><\/pre>\n\n<p>Hope I helped anyone else. :)<\/p>\n"}]}
+{"QuestionId":57976140,"AnswerCount":2,"Tags":"<keras><neural-network>","CreationDate":"2019-09-17T14:09:27.507","AcceptedAnswerId":58050687.0,"OwnerUserId":2919560.0,"Title":"Keras model learning to give wrong answers training accuracy raising to 0.8 to then deacrease dramatically down to 0.1","Body":"<p>I'm going through the book \"Deep Learning with Python\" from Fran\u00e7ois Chollet.<\/p>\n\n<p>In section 7.9, there is an example neural network using Conv1D layers to process the IMDB dataset. What surprises me is that it starts to learn and both training and validation accuracy improve, but after a couple epochs, both the training as well as the validation accuracy drop.<\/p>\n\n<p>Seeing the validation accuracy drop does not surprises me, it's quite typical of overfitting, but <strong>what I don't understand is how the training accuracy can drop down to as low as 12%. It's almost as if the network was learning the opposite of what it should.<\/strong><\/p>\n\n<p>The code is:<\/p>\n\n<pre><code>import keras\nfrom keras import layers\nfrom keras.datasets import imdb\nfrom keras.preprocessing import sequence\nfrom keras.optimizers import Adam, RMSprop\n\nmax_features = 2000\nmax_len = 500\n\n(x_train, y_train), (x_test, y_test) = imdb.load_data(num_words=max_features)\nx_train = sequence.pad_sequences(x_train, maxlen=max_len)\nx_test = sequence.pad_sequences(x_test, maxlen=max_len)\n\nmodel = keras.models.Sequential()\nmodel.add(layers.Embedding(max_features, 128,\n                           input_length=max_len,\n                           name='embed'))\nmodel.add(layers.Conv1D(32, 7, activation='relu'))\nmodel.add(layers.MaxPooling1D(5))\nmodel.add(layers.Conv1D(32, 7, activation='relu'))\nmodel.add(layers.GlobalMaxPooling1D())\nmodel.add(layers.Dense(1))\n\nmodel.summary()\n\nmodel.compile(optimizer=RMSprop(),\n              loss='binary_crossentropy',\n              metrics=['acc'])\n\nhistory = model.fit(x_train, y_train,\n                    epochs=20,\n                    batch_size=128,\n                    validation_split=0.2,\n                    # callbacks=callbacks,\n                    )\n<\/code><\/pre>\n\n<p>And the training results are as follows:<\/p>\n\n<pre><code>Train on 20000 samples, validate on 5000 samples\nEpoch 1\/20\n20000\/20000 [==============================] - 4s 212us\/step - loss: 0.7043 - acc: 0.6076 - val_loss: 0.4488 - val_acc: 0.8166\nEpoch 2\/20\n20000\/20000 [==============================] - 3s 151us\/step - loss: 0.4509 - acc: 0.8179 - val_loss: 0.6575 - val_acc: 0.7594\nEpoch 3\/20\n20000\/20000 [==============================] - 3s 151us\/step - loss: 0.4082 - acc: 0.7923 - val_loss: 0.4759 - val_acc: 0.7874\nEpoch 4\/20\n20000\/20000 [==============================] - 3s 152us\/step - loss: 0.3633 - acc: 0.7526 - val_loss: 0.5385 - val_acc: 0.7356\nEpoch 5\/20\n20000\/20000 [==============================] - 3s 154us\/step - loss: 0.3333 - acc: 0.7235 - val_loss: 0.5658 - val_acc: 0.7056\nEpoch 6\/20\n20000\/20000 [==============================] - 3s 152us\/step - loss: 0.2793 - acc: 0.6868 - val_loss: 0.5790 - val_acc: 0.6494\nEpoch 7\/20\n20000\/20000 [==============================] - 3s 151us\/step - loss: 0.2433 - acc: 0.6408 - val_loss: 0.6710 - val_acc: 0.5726\nEpoch 8\/20\n20000\/20000 [==============================] - 3s 149us\/step - loss: 0.2061 - acc: 0.5789 - val_loss: 1.7192 - val_acc: 0.3538\nEpoch 9\/20\n20000\/20000 [==============================] - 3s 151us\/step - loss: 0.1769 - acc: 0.5144 - val_loss: 0.8144 - val_acc: 0.4416\nEpoch 10\/20\n20000\/20000 [==============================] - 3s 151us\/step - loss: 0.1507 - acc: 0.4365 - val_loss: 1.1555 - val_acc: 0.3682\nEpoch 11\/20\n20000\/20000 [==============================] - 3s 152us\/step - loss: 0.1395 - acc: 0.3675 - val_loss: 1.1440 - val_acc: 0.3412\nEpoch 12\/20\n20000\/20000 [==============================] - 3s 156us\/step - loss: 0.1241 - acc: 0.3159 - val_loss: 1.8202 - val_acc: 0.2686\nEpoch 13\/20\n20000\/20000 [==============================] - 3s 155us\/step - loss: 0.1225 - acc: 0.2756 - val_loss: 1.0667 - val_acc: 0.2944\nEpoch 14\/20\n20000\/20000 [==============================] - 3s 152us\/step - loss: 0.1183 - acc: 0.2422 - val_loss: 1.1143 - val_acc: 0.2794\nEpoch 15\/20\n20000\/20000 [==============================] - 3s 151us\/step - loss: 0.1153 - acc: 0.2142 - val_loss: 1.1599 - val_acc: 0.2686\nEpoch 16\/20\n20000\/20000 [==============================] - 3s 153us\/step - loss: 0.1150 - acc: 0.1930 - val_loss: 1.2467 - val_acc: 0.2544\nEpoch 17\/20\n20000\/20000 [==============================] - 3s 151us\/step - loss: 0.1145 - acc: 0.1766 - val_loss: 1.1953 - val_acc: 0.2492\nEpoch 18\/20\n20000\/20000 [==============================] - 3s 153us\/step - loss: 0.1115 - acc: 0.1508 - val_loss: 1.4812 - val_acc: 0.2226\nEpoch 19\/20\n20000\/20000 [==============================] - 3s 156us\/step - loss: 0.1119 - acc: 0.1355 - val_loss: 1.2690 - val_acc: 0.2288\nEpoch 20\/20\n20000\/20000 [==============================] - 3s 155us\/step - loss: 0.1127 - acc: 0.1248 - val_loss: 1.2903 - val_acc: 0.2148\n<\/code><\/pre>\n\n<p>I could of course stop it early while the validation accuracy is at the top, but what I would is to understand how the training accuracy can drop so dramatically. And more surprisingly, how it can drop below 0.5, being a softmax 0\/1 type output layer, I would expect the accuracy to be around 0.5. It really looks like it learns to give wrong answers. <\/p>\n","answers":[{"AnswerId":"57982921","CreationDate":"2019-09-17T22:47:28.937","ParentId":null,"OwnerUserId":"10542451","Title":null,"Body":"<p>Sounds like exploding gradients or high variance. Try batch normalization before activations (or some other kind of regularization).<\/p>\n\n<p>Use some easier model like the one below and work your way up:<\/p>\n\n<pre><code>model = keras.models.Sequential()\nmodel.add(layers.Embedding(max_features, 128,\n                           input_length=max_len,\n                           name='embed'))\nmodel.add(layers.Flatten())\nmodel.add(layers.Dense(512))\nmodel.add(layers.BatchNormalization())\nmodel.add(layers.ReLU())\nmodel.add(layers.Dense(1))\nmodel.add(layers.ReLU())\n<\/code><\/pre>\n\n<p>Got to 70% after 20 episodes with that one.<\/p>\n\n<p>Another problem could be too few examples in your data, try getting more samples, maybe through data augmentation, or reduce features.<\/p>\n"},{"AnswerId":"58050687","CreationDate":"2019-09-22T15:30:27.933","ParentId":null,"OwnerUserId":"2919560","Title":null,"Body":"<p>Michael, thanks for the pointers on gradient explosion, I had not heard of this problem before.<\/p>\n\n<p>What was really unusual with that network was that I was getting 12.5% accuracy for a classification problem with two answers with balanced samples. In the case the network is not learning I would have expected something around 50%.<\/p>\n\n<p>Comming back to it after a break, I realized the problem is that the last Dense layer is missing an activation function. This would be OK for a regression problem, but not for a classification problem like this one. After adding activation='sigmoid' the network trains properly, with training accuracy raising continuously. It shows overfitting but that's expected.<\/p>\n\n<p>So in short the problem was the use of the loss function <code>binary_crossentropy<\/code> with no activation function.<\/p>\n"}]}
+{"QuestionId":57976730,"AnswerCount":1,"Tags":"<keras><neural-network><deep-learning><multiclass-classification>","CreationDate":"2019-09-17T14:41:43.920","AcceptedAnswerId":null,"OwnerUserId":11359194.0,"Title":"Keras Multiclass Classification (Dense model) - Confusion Matrix Incorrect","Body":"<p><a href=\"https:\/\/i.stack.imgur.com\/5dtoF.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5dtoF.jpg\" alt=\"enter image description here\"><\/a><a href=\"https:\/\/i.stack.imgur.com\/57W4k.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/57W4k.jpg\" alt=\"enter image description here\"><\/a>I have a labeled dataset. last column (78) contains 4 types of attack. following codes confusion matrix is correct for two types of attack. can any one help to modify the code for keras multiclass attack detection and correction for get correct confusion matrix? and for correct code for precision, FPR,TPR for multiclass. Thanks.<\/p>\n\n<pre><code>import pandas as pd\nfrom sklearn.preprocessing import LabelEncoder, StandardScaler\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.model_selection import GridSearchCV\nfrom tensorflow.keras.wrappers.scikit_learn import KerasClassifier\nfrom tensorflow.keras.models import Sequential, load_model\nfrom tensorflow.keras.layers import Dense\nfrom sklearn.metrics import confusion_matrix\nimport matplotlib.pyplot as plt\nimport seaborn as sns\nfrom keras.utils.np_utils import to_categorical\n\ndataset_original = pd.read_csv('.\/XYZ.csv')\n\n# Dron NaN value from Data Frame\ndataset = dataset_original.dropna()\n\n\n# data cleansing\nX = dataset.iloc[:, 0:78]\nprint(X.info())\nprint(type(X))\ny = dataset.iloc[:, 78] #78 is labeled column contains 4 anomaly type\nprint(y)\n\n\n# encode the labels to 0, 1 respectively\nprint(y[100:110])\nencoder = LabelEncoder()\ny = encoder.fit_transform(y)\nprint([y[100:110]])\n\n# Split the dataset now\nXTrain, XTest, yTrain, yTest = train_test_split(X, y, test_size=0.2, random_state=0)\n\n\n# feature scaling\nscalar = StandardScaler()\nXTrain = scalar.fit_transform(XTrain)\nXTest = scalar.transform(XTest)\n\n\n# modeling\nmodel = Sequential()\nmodel.add(Dense(units=16, kernel_initializer='uniform', activation='relu', input_dim=78))\nmodel.add(Dense(units=8, kernel_initializer='uniform', activation='relu'))\nmodel.add(Dense(units=6, kernel_initializer='uniform', activation='relu'))\nmodel.add(Dense(units=1, kernel_initializer='uniform', activation='sigmoid'))\nmodel.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])\nmodel.fit(XTrain, yTrain, batch_size=1000, epochs=10)\n\n\nhistory = model.fit(XTrain, yTrain, batch_size=1000, epochs=10, verbose=1, validation_data=(XTest, \nyTest))\n\n\nyPred = model.predict(XTest)\nyPred = [1 if y &gt; 0.5 else 0 for y in yPred]\nmatrix = confusion_matrix(yTest, yPred)`enter code here`\nprint(matrix)\naccuracy = (matrix[0][0] + matrix[1][1]) \/ (matrix[0][0] + matrix[0][1] + matrix[1][0] + matrix[1][1])\nprint(\"Accuracy: \" + str(accuracy * 100) + \"%\")\n<\/code><\/pre>\n","answers":[{"AnswerId":"57977615","CreationDate":"2019-09-17T15:33:11.840","ParentId":null,"OwnerUserId":"5136550","Title":null,"Body":"<p>If i understand correctly, you are trying to solve a multiclass classification problem where your target label belongs to 4 different attacks. Therefore, you should use the output Dense layer having 4 units instead of 1 with a 'softmax' activation function (not 'sigmoid' activation). Additionally, you should use 'categorical_crossentropy' loss in place of 'binary_crossentropy' while compiling your model.\nFurthermore, with this setting, applying argmax on prediction result (that has 4 class probability values for each test sample) you will get the final label\/class.<\/p>\n\n<hr>\n\n<p>[Edit]\nYour confusion matrix and high accuracy indicates that you are working with an imbalanced dataset. May be very high number of samples are from class 0 and few samples are from the remaining 3 classes. To handle this you may want to apply weighting samples or over-sampling\/under-sampling approaches.<\/p>\n"}]}
+{"QuestionId":57976816,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><vectorization><convolution>","CreationDate":"2019-09-17T14:46:07.547","AcceptedAnswerId":57977904.0,"OwnerUserId":1682673.0,"Title":"Numpy way to generate linear operation matrix from a convolution kernel","Body":"<p>A 2D convolution kernel, <code>K<\/code>, of shape <code>(k1, k2, n_channel, n_filter)<\/code> applies on a 2D vector, <code>A<\/code>, of shape <code>(m1, m2, n_channel)<\/code> and generates another 2D vector, <code>B<\/code>, of shape <code>(m1 - k1 + 1, m2 - k2 + 1, n_filter)<\/code> (with <em>valid<\/em> padding).<\/p>\n\n<p>It is also true that for each <code>K<\/code>, there exists a <code>W_K<\/code> of shape <code>(m1 - k1 + 1, m2 - k2 + 1, n_filter, m1, m2, n_channel)<\/code>, such that tensor dot of <code>W_K<\/code> and <code>A<\/code> is equal to <code>B<\/code>. i.e. <code>B = np.tensordot(W_K, A, 3)<\/code>.<\/p>\n\n<p>I am trying to find a pure NumPy solution to generate this <code>W_K<\/code> from <code>K<\/code> without using any python loops.<\/p>\n\n<p>I can see <code>W_K[i,j,f] == np.pad(K[...,f], ((i,m1-i-k1), (j,m2-j-k2)), 'constant', constant_values=0)<\/code> or simply <code>W_K[i, j, f, i:i+k1, j:j+k2, ...] == K[..., f]<\/code>.<\/p>\n\n<p>What I'm looking for is almost similar to a Toeplitz matrix. But I need it in multi-dimensions.<\/p>\n\n<p>Example in loopy code:<\/p>\n\n<pre><code>import numpy as np\n\n# 5x5 image with 3-channels\nA = np.random.random((5,5,3))\n# 2x2 Conv2D kernel with 2 filters for A  \nK = np.random.random((2,2,3,2))\n\n# It should be of (4,4,2,5,5,3), but I create this way for convenience. I move the axis at the end.\nW_K = np.empty((4,4,5,5,3,2))\nfor i, j in np.ndindex(4, 4):\n  W_K[i, j] = np.pad(K, ((i, 5-i-2),(j, 5-j-2), (0, 0), (0, 0)), 'constant', constant_values=0)\n\n# above lines can also be rewritten as\nW_K = np.zeros((4,4,5,5,3,2))\nfor i, j in np.ndindex(4, 4):\n  W_K[i, j, i:i+2, j:j+2, ...] = K[...]\n\nW_K = np.moveaxis(W_K, -1, 2)\n\n# now I can do\nB = np.tensordot(W_K, A, 3)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57977904","CreationDate":"2019-09-17T15:51:30.273","ParentId":null,"OwnerUserId":"5067311","Title":null,"Body":"<p>What you want needs a bit of <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.indexing.html#advanced-indexing\" rel=\"nofollow noreferrer\">fancy indexing<\/a> gymnastics but it's not very cumbersome to code. The idea is to create 4-dimensional index arrays that apply the <code>W_K[i, j, i:i+2, j:j+2, ...]<\/code> part of your second loopy example.<\/p>\n\n<p>Here's a slightly modified version of your example, just to make sure that some relevant dimensions differ (because this makes bugs easier to find: they would be proper errors rather than mangled values):<\/p>\n\n<pre><code>import numpy as np\n\n# parameter setup\nk1, k2, nch, nf = 2, 4, 3, 2 \nm1, m2 = 5, 6 \nw1, w2 = m1 - k1 + 1, m2 - k2 + 1 \nK = np.random.random((k1, k2, nch, nf)) \nA = np.random.random((m1, m2, nch)) \n\n# your loopy version for comparison\nW_K = np.zeros((w1, w2, nf, m1, m2, nch)) \nfor i, j in np.ndindex(w1, w2): \n    W_K[i, j, :, i:i+k1, j:j+k2, ...] = K.transpose(-1, 0, 1, 2) \n\nW_K2 = np.zeros((w1, w2, m1, m2, nch, nf))  # to be transposed back\ni,j = np.mgrid[:w1, :w2][..., None, None]  # shape (w1, w2, 1, 1) \nk,l = np.mgrid[:k1, :k2]  # shape (k1, k2) ~ (1, 1, k1, k2)  \n\nW_K2[i, j, i+k, j+l, ...] = K \nW_K2 = np.moveaxis(W_K2, -1, 2) \n\nprint(np.array_equal(W_K, W_K2))  # True\n<\/code><\/pre>\n\n<p>We first create an index mesh <code>i,j<\/code> that span the first two dimensions of <code>W_K<\/code>, then create two similar meshes that span its (pre-<code>moveaxis<\/code>) second and third dimensions. By injecting two trailing singleton dimensions into the former we end up with 4d index arrays that together span the first four dimensions of <code>W_K<\/code>.<\/p>\n\n<p>All that's left is to assign to this slice using the original <code>K<\/code>, and move back the dimension. Due to how advanced indexing changes behaviour when the sliced (non-advanced) indices in an expression are not all next to one another, this is much easier to do with your <code>moveaxis<\/code> approach. I first tried to create <code>W_K2<\/code> with its final dimensions, but then we'd have <code>W_K[i, j, :, i+k, j+l, ...]<\/code> that has subtly different behaviour (in particular, different shape).<\/p>\n"}]}
+{"QuestionId":57976938,"AnswerCount":0,"Tags":"<python><python-3.x><machine-learning><neural-network><pytorch>","CreationDate":"2019-09-17T14:52:37.123","AcceptedAnswerId":null,"OwnerUserId":2111778.0,"Title":"PyTorch: How to do inference on multiple GPUs?","Body":"<p>PyTorch has <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/former_torchies\/parallelism_tutorial.html\" rel=\"nofollow noreferrer\">DataParallel<\/a>, which splits a batch among multiple GPUs. However, this is not the kind of parallelism I want, as this does not offer any speed benefits unless the batch size is too large to fit into the GPU's memory. <\/p>\n\n<p>I want real-time low-latency evaluation in multiple CPU threads. <\/p>\n\n<p>So for 2 GPUs, I would want to alternatively use GPU 0 and GPU 1 for inference (both controlled by a corresponding CPU thread). So when new data arrives, even if we just started a job for one GPU so it is occupied, we can prepare a batch for the free GPU.<\/p>\n\n<p>The reason I'm interested in low latency is because I'm implementing a variant of the AlphaGo algorithm.<\/p>\n","answers":[]}
+{"QuestionId":57977217,"AnswerCount":0,"Tags":"<python><tensorflow><object-detection-api>","CreationDate":"2019-09-17T15:08:17.470","AcceptedAnswerId":null,"OwnerUserId":12062750.0,"Title":"Reading multiple files in a folder and saving bounding box predictions in a CSV file","Body":"<p>I have just started programming and started using python and the Tensorflow Object Detection API. I have successfully tested my network using the code below:<\/p>\n\n<pre><code>    ######## Image Object Detection Using Tensorflow-trained Classifier #########\n#\n# Author: Evan Juras\n# Date: 1\/15\/18\n# Description: \n# This program uses a TensorFlow-trained classifier to perform object detection.\n# It loads the classifier uses it to perform object detection on an image.\n# It draws boxes and scores around the objects of interest in the image.\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# Import packages\nimport os\nimport cv2\nimport numpy as np\nimport tensorflow as tf\nimport sys\n\n# This is needed since the notebook is stored in 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 = 'inference_graph'\nIMAGE_NAME = 'testImage3.jpg'\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,'training','labelmap.pbtxt')\n\n# Path to image\nPATH_TO_IMAGE = os.path.join(CWD_PATH,IMAGE_NAME)\n\n# Number of classes the object detector can identify\nNUM_CLASSES = 6\n\n# Load the label map.\n# Label maps map indices to category names, so that when our convolution\n# network predicts `5`, we know that this corresponds to `king`.\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# 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# Load image using OpenCV and\n# expand image 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\nimage = cv2.imread(PATH_TO_IMAGE)\nimage_expanded = np.expand_dims(image, 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: image_expanded})\n\n# Draw the results of the detection (aka 'visulaize the results')\n\nvis_util.visualize_boxes_and_labels_on_image_array(\n    image,\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=0.60)\n\n# All the results have been drawn on image. Now display the image.\ncv2.imshow('Object detector', image)\n\n# Press any key to close the image\ncv2.waitKey(0)\n\n# Clean up\ncv2.destroyAllWindows()\n<\/code><\/pre>\n\n<p>However, I would now like to use this same code so that I am able to read multiple images as well as save their predicted co-ordinates into a CSV file.<\/p>\n\n<p>All  my efforts to solve this has been futile.<\/p>\n\n<p>Any help would be appreciated!<\/p>\n","answers":[]}
+{"QuestionId":57977352,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-17T15:17:10.870","AcceptedAnswerId":57984086.0,"OwnerUserId":5777338.0,"Title":"How to plot the image tensor returned by tf.image.sobel_edges","Body":"<p>I want to use sobel edge in tensorflow using <code>tf.image.sobel_edges<\/code>.<\/p>\n\n<p>The following is the code <\/p>\n\n<pre><code>import tensorflow as tf\nimport skimage.io\nimport numpy as np\nfrom tensorflow import keras\nimport matplotlib.image as mpimg\nimport matplotlib.pyplot as plt\n\nimage = skimage.io.imread('table1.jpg')\n\nimage=np.array(image)\nimage = tf.cast(image, tf.float32)\nimage=tf.compat.v1.expand_dims(image, 0)\n\nsobel= tf.image.sobel_edges(image)\n\nsess = tf.Session()\nSobelImage=sess.run(sobel)\nplt.imshow(SobelImage)\nsess.close() \n<\/code><\/pre>\n\n<p>When I run the program I get<\/p>\n\n<blockquote>\n  <p>runfile('E:\/projects\/Github programs\/image recognition\/sobleEdge.py', wdir='E:\/projects\/Github programs\/image recognition')\n  Traceback (most recent call last):<\/p>\n  \n  <p>File \"\", line 1, in \n      runfile('E:\/projects\/Github programs\/image recognition\/sobleEdge.py', wdir='E:\/projects\/Github programs\/image recognition')<\/p>\n  \n  <p>File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\spyder_kernels\\customize\\spydercustomize.py\", line 827, in runfile\n      execfile(filename, namespace)<\/p>\n  \n  <p>File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\spyder_kernels\\customize\\spydercustomize.py\", line 110, in execfile\n      exec(compile(f.read(), filename, 'exec'), namespace)<\/p>\n  \n  <p>File \"E:\/projects\/Github programs\/image recognition\/sobleEdge.py\", line 29, in \n      plt.imshow(SobelImage)<\/p>\n  \n  <p>File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\matplotlib\\pyplot.py\", line 2677, in imshow\n      None else {}), **kwargs)<\/p>\n  \n  <p>File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\matplotlib__init__.py\", line 1589, in inner\n      return func(ax, *map(sanitize_sequence, args), **kwargs)<\/p>\n  \n  <p>File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\matplotlib\\cbook\\deprecation.py\", line 369, in wrapper\n      return func(*args, **kwargs)<\/p>\n  \n  <p>File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\matplotlib\\cbook\\deprecation.py\", line 369, in wrapper\n      return func(*args, **kwargs)<\/p>\n  \n  <p>File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\matplotlib\\axes_axes.py\", line 5660, in imshow\n      im.set_data(X)<\/p>\n  \n  <p>File \"C:\\ProgramData\\Anaconda3\\lib\\site-packages\\matplotlib\\image.py\", line 683, in set_data\n      .format(self._A.shape))<\/p>\n  \n  <p>TypeError: Invalid shape (1, 565, 750, 3, 2) for image data<\/p>\n<\/blockquote>\n\n<p>Can Someone help me how to plot the tensor returned by <code>tf.image.sobel_edges<\/code><\/p>\n","answers":[{"AnswerId":"57984086","CreationDate":"2019-09-18T02:05:33.313","ParentId":null,"OwnerUserId":"5136550","Title":null,"Body":"<pre><code>sobel= tf.image.sobel_edges(image)\nprint(sobel.shape)\ncv2.imshow(\"image\", image)\nsobel_x = np.asarray(sobel[0, :, :, :, 0])  #  Sobel_X\nsobel_y = np.asarray(sobel[0, :, :, :, 1])  #  Sobel_Y\ncv2.imshow(\"Sobel_x\",sobel_x)\ncv2.imshow(\"Sobel_y\",sobel_y)\ncv2.waitKey(0)\n<\/code><\/pre>\n\n<p>Since the output of the sobel_edge function is a 5D tensor with shape [batch_size, h, w, d, 2] where the last two dimensions hold Sobel edge responses in X and Y direction. Additionally, to display you need to convert tensor to np array first as cv2 cant handle tensor directly.\nOutput: <a href=\"https:\/\/i.stack.imgur.com\/L4obH.jpg\" rel=\"nofollow noreferrer\">click to see the output image here<\/a><\/p>\n"}]}
+{"QuestionId":57977393,"AnswerCount":0,"Tags":"<docker><tensorflow><tensorflow-serving>","CreationDate":"2019-09-17T15:19:35.990","AcceptedAnswerId":null,"OwnerUserId":12080392.0,"Title":"What is the difference between \"direct_session_runs\" and \"graph_runs\" in tf_serving docker","Body":"<p>I'm trying to create some dash boards in splunk using tensorflow-serving docker logs. It returns two metrics for each event that I'm not sure what their difference or meaning is, \"direct_session_runs\" and \"graph_runs\"<\/p>\n\n<p>I've tried looking in the tf documentation but I'm not seeing an explanation of what the values are. I originally thought they would be the same but I'm seeing them occasionally diverge.<\/p>\n\n<pre><code># TYPE :tensorflow:core:direct_session_runs counter\n:tensorflow:core:direct_session_runs{} 82\n\n# TYPE :tensorflow:core:graph_runs counter\n:tensorflow:core:graph_runs{} 82\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57977408,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow><tensorflow-datasets><tensorflow2.0>","CreationDate":"2019-09-17T15:20:20.760","AcceptedAnswerId":58220646.0,"OwnerUserId":4373898.0,"Title":"How to cache and iterate through a Dataset of unknown size?","Body":"<p>While adding the <code>.cache()<\/code> step to my dataset pipeline, successives training epochs still download the data from the network storage.<\/p>\n\n<p>I have a dataset on a network storage. I want to cache it, but not to repeat it: a training epoch must run through the whole dataset.\nHere is my dataset building pipeline:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>return tf.data.Dataset.list_files(\n        file_pattern\n    ).interleave(\n        tf.data.TFRecordDataset,\n        num_parallel_calls=tf.data.experimental.AUTOTUNE\n    ).shuffle(\n        buffer_size=2048\n    ).batch(\n        batch_size=2048,\n        drop_remainder=True,\n    ).cache(\n    ).map(\n        map_func=_parse_example_batch,\n        num_parallel_calls=tf.data.experimental.AUTOTUNE\n    ).prefetch(\n        buffer_size=32\n    )\n<\/code><\/pre>\n\n<p>If I use it as is, the dataset is downloaded at each epoch. To avoid this, I have to add the <code>.repeat()<\/code> step to the pipeline and use the <code>steps_per_epoch<\/code> keyword of the <code>model.fit<\/code> function. However, I do not know the size of the complete dataset and thus I\u00a0cannot pass the right <code>steps_per_epoch<\/code> value.<\/p>\n\n<p>What is the right way to cache and use dataset of unknown size?<\/p>\n\n<p>Thanks.<\/p>\n\n<hr>\n\n<h2>Edit<\/h2>\n\n<p>While reading some TF\u00a0code, I (re)discovered the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset#make_initializable_iterator\" rel=\"nofollow noreferrer\"><code>make_initializable_iterator<\/code><\/a>. It seems that it is what I am looking for, that is to say iterate multiple times through the same dataset (taking advantage of the cache after the first iteration). However, this is deprecated and no longer part of the main API in TF2.<\/p>\n\n<p>Updating instruction is to manually iterate over the Dataset with <code>for ... in dataset<\/code>. Is it not what is done by the <code>keras.Model.fit<\/code> function? Have I to write the training loop manually to get cache advantages?<\/p>\n\n<p>Kind.<\/p>\n","answers":[{"AnswerId":"58220646","CreationDate":"2019-10-03T13:52:24.137","ParentId":null,"OwnerUserId":"4373898","Title":null,"Body":"<p>Good news! Final v2.0.0 release fix this behavior.<\/p>\n\n<p>Here is a code snippet to highlight the different behaviors.<\/p>\n\n<pre><code>import time\n\nimport tensorflow as tf\nimport tensorflow.keras as keras\n\n# Simple layer that just print its inputs\nclass Print(keras.layers.Layer):\n\n       def compute_output_signature(self, input_signature):\n              return input_signature\n\n       def call(self, inputs, **kwargs):\n              tf.print(inputs)\n              return inputs\n\n# Generator returning incremented values each time it is re-initialized\ngenerator_list = [0]\ndef generator():\n       v = generator_list[-1]\n       generator_list.append(v+1)\n       tf.print(\"Generating samples with value {}\".format(v))\n       time.sleep(2)\n       for i in range(2):\n              yield (tf.constant([v]), tf.constant(v))\n\n\ndef main():\n       model_input = keras.layers.Input(shape=(1,))\n       model_output = Print()(model_input)\n       model = keras.Model(inputs=model_input, outputs=model_output)\n       model.compile(\"adam\", loss=\"mae\")\n\n       ds = tf.data.Dataset.from_generator(\n              generator, (tf.int64, tf.int64), ([1], [])\n       )\n       cached_ds = ds.cache()\n\n       tf.print(\"Fit\")\n       model.fit(\n              cached_ds,\n              epochs=3,\n              verbose=2\n       )\n\n       tf.print(\"For ... in ...\")\n       for i in range(3):\n              for x, y in cached_ds:\n                     model(x)\n\nif __name__ == '__main__':\n    main()\n<\/code><\/pre>\n\n<p>With tensorflow 2.0.0-b1 (used on Google AI Platform), here is the output:<\/p>\n\n<pre><code>Fit\nEpoch 1\/3\nGenerating samples with value 0\n# sleep 2s\n2019-10-03 15:45:32.718522: W tensorflow\/compiler\/jit\/mark_for_compilation_pass.cc:1483] (One-time warning): Not using XLA:CPU for cluster because envvar TF_XLA_FLAGS=--tf_xla_cpu_global_jit was not set.  If you want XLA:CPU, either set that envvar, or use experimental_jit_scope to enable XLA:CPU.  To confirm that XLA is active, pass --vmodule=xla_compilation_cache=1 (as a proper command-line flag, not via TF_XLA_FLAGS) or set the envvar XLA_FLAGS=--xla_hlo_profile.\n[[0]]\n[[0]]\n2\/2 - 2s - loss: 0.0000e+00\nGenerating samples with value 1\n# sleep 2s\nEpoch 2\/3\n[[1]]\n[[1]]\n2\/2 - 2s - loss: 0.0000e+00\nEpoch 3\/3\n2019-10-03 15:45:34.774195: W tensorflow\/core\/kernels\/data\/cache_dataset_ops.cc:815] The calling iterator did not fully read the dataset being cached. In order to avoid unexpected truncation of the dataset, the partially cached contents of the dataset will be discarded. This can happen if you have an input pipeline similar to `dataset.cache().take(k).repeat()`. You should use `dataset.take(k).cache().repeat()` instead.\nGenerating samples with value 2\n# sleep 2s\n[[2]]\n[[2]]\n2019-10-03 15:45:36.782046: W tensorflow\/core\/kernels\/data\/cache_dataset_ops.cc:815] The calling iterator did not fully read the dataset being cached. In order to avoid unexpected truncation of the dataset, the partially cached contents of the dataset will be discarded. This can happen if you have an input pipeline similar to `dataset.cache().take(k).repeat()`. You should use `dataset.take(k).cache().repeat()` instead.\n2\/2 - 2s - loss: 0.0000e+00\nFor ... in ...\nGenerating samples with value 3\n# sleep 2s\n[3]\n[3]\nGenerating samples with value 4\n# sleep 2s\n[4]\n[4]\nGenerating samples with value 5\n# sleep 2s\n[5]\n[5]\n<\/code><\/pre>\n\n<p>You can see, that the value of the tensor is incremented for each epoch, and the sleep instruction is executed each time. Moreover, we get the warning about truncated iterator...<\/p>\n\n<p>Now, with tensorflow 2.0.0:<\/p>\n\n<pre><code>Fit\nEpoch 1\/3\nWARNING:tensorflow:The list of trainable weights is empty. Make sure that you are not setting model.trainable to False before compiling the model.\nGenerating samples with value 0\n# sleep 2s\n[[0]]\n[[0]]\n2019-10-03 15:49:59.587796: W tensorflow\/core\/common_runtime\/base_collective_executor.cc:216] BaseCollectiveExecutor::StartAbort Out of range: End of sequence\n     [[{{node IteratorGetNext}}]]\n2\/2 - 2s - loss: 0.0000e+00\nEpoch 2\/3\n[[0]]\n[[0]]\n2019-10-03 15:49:59.598144: W tensorflow\/core\/common_runtime\/base_collective_executor.cc:216] BaseCollectiveExecutor::StartAbort Out of range: End of sequence\n     [[{{node IteratorGetNext}}]]\n2\/2 - 0s - loss: 0.0000e+00\nEpoch 3\/3\n[[0]]\n[[0]]\n2019-10-03 15:49:59.605260: W tensorflow\/core\/common_runtime\/base_collective_executor.cc:216] BaseCollectiveExecutor::StartAbort Out of range: End of sequence\n     [[{{node IteratorGetNext}}]]\nFor ... in ...\n2\/2 - 0s - loss: 0.0000e+00\n[0]\n[0]\n[0]\n[0]\n[0]\n[0]\n<\/code><\/pre>\n\n<p>And 'Voila'! The generator function is executed only once, with no more sleeps and always the same value of the tensor. I just have some warnings about end of sequence, but I can support it!<\/p>\n\n<p>Kind.<\/p>\n"},{"AnswerId":"57994034","CreationDate":"2019-09-18T13:34:13.947","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>In TF2.0, you do not need <code>.repeat()<\/code>. By <\/p>\n\n<blockquote>\n  <p>successives training epochs still download the data from the network storage.<\/p>\n<\/blockquote>\n\n<p>I think you got confused with the message <code>filling up shuffle buffer<\/code>. This happens before every epoch if you are using <code>shuffle()<\/code> function. Maybe try without <code>shuffle()<\/code>, just to see the difference. \nAlso, I would suggest you to use <code>cache()<\/code> after <code>map()<\/code> and before <code>batch()<\/code>.<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<blockquote>\n  <p>filling up shuffle buffer<\/p>\n<\/blockquote>\n\n<p>is a message you get when using <code>shuffle<\/code> function. You can still <code>shuffle()<\/code> the dataset after using <code>cache()<\/code>. Look <a href=\"https:\/\/www.tensorflow.org\/tutorials\/load_data\/images#cache\" rel=\"nofollow noreferrer\">here<\/a>\nAlso, if I understood it correctly you are feeding the resulted dataset from <code>map()<\/code> to your model for training, then you should <code>cache()<\/code> this dataset not the other one because training will be done on this.\nFor counting the number of elements in your dataset you can use following code <\/p>\n\n<pre><code>num_elements = 0\nfor element in dataset: # tf.dataset type\n  num_elements += 1\nprint ('Total number of elements in the file: ',num_elements)\n<\/code><\/pre>\n\n<p>Now, by diving this <code>num_elements<\/code> with your <code>batch_size<\/code> you would get <code>steps_per_epoch<\/code><\/p>\n"}]}
+{"QuestionId":57977713,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2019-09-17T15:39:35.303","AcceptedAnswerId":null,"OwnerUserId":12080428.0,"Title":"How to calculate backpropagation through tf.while_loop to use as loss function","Body":"<p>I want to implement an Fourier Ring Correlation Loss for two images to train a GAN. Therefore I'd like to loop over a specific amount of times and calculate the loss. This works fine for a normal Python loop. To speed up the process I want to use the <code>tf.while_loop<\/code> but unfortunately I am not able to track the gradients through my while loop. I constructed a dummy example just to calculate gradients during a while loop but it doesn't work. First, the working python loop : <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = tf.constant(3.0)\ny = tf.constant(2.0)\n\nfor i in range(3):    \n    y = y * x\n\ngrad = tf.gradients(y, x) \nwith tf.Session() as ses:\n    print(\"output : \", ses.run(grad))\n<\/code><\/pre>\n\n<p>This works and gives the output<\/p>\n\n<p>[54]<\/p>\n\n<p>If i do the same with a tf.while_loop it doesn't work:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>a = tf.constant(0, dtype = tf.int64)\nb = tf.constant(3, dtype = tf.int64)\n\nx = tf.constant(3.0)\ny = tf.constant(2.0)\n\ndef cond(a,b,x,y):\n    return tf.less(a,b)\n\ndef body(a,b,x,y):\n\n    y = y * x \n    with tf.control_dependencies([y]):\n        a = a + 1\n    return [a,b,x,y]\n\nresults = tf.while_loop(cond, body, [a,b,x,y], back_prop = True)\n\ngrad = tf.gradients(y, results[2])\n\nwith tf.Session() as ses:\n    print(\"grad : \", ses.run(grad))\n\n<\/code><\/pre>\n\n<p>The output is :<\/p>\n\n<blockquote>\n  <p>TypeError: Fetch argument None has invalid type '&lt;'class 'NoneType'><\/p>\n<\/blockquote>\n\n<p>So I guess somehow tensorflow is not able to do the backpropagation.<\/p>\n\n<p>The problem still accours if you use tf.GradientTape() instead of tf.gradients().<\/p>\n","answers":[{"AnswerId":"57988654","CreationDate":"2019-09-18T08:47:12.157","ParentId":null,"OwnerUserId":"7146757","Title":null,"Body":"<p>I changed the code so that it now outputs the gradients:<\/p>\n\n<pre><code>import tensorflow as tf\n\na = tf.constant(0, dtype = tf.int64)\nb = tf.constant(3, dtype = tf.int64)\n\nx = tf.Variable(3.0, tf.float32)\ny = tf.Variable(2.0, tf.float32)\ndy = tf.Variable(0.0, tf.float32)\n\ndef cond(a,b,x,y,dy):\n    return tf.less(a,b)\n\ndef body(a,b,x,y,dy):\n\n    y = y * x \n    dy = tf.gradients(y, x)[0]\n    with tf.control_dependencies([y]):\n        a = a + 1\n    return [a,b,x,y,dy]\n\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as ses:\n    ses.run(init)\n    results = ses.run(tf.while_loop(cond, body, [a,b,x,y,dy], back_prop = True))\n    print(\"grad : \", results[-1])\n<\/code><\/pre>\n\n<p>The things I modified:<\/p>\n\n<ol>\n<li>I made <code>x<\/code> and <code>y<\/code> into variables and added their initialisation <code>init<\/code>.<\/li>\n<li>I added a variable called <code>dy<\/code> which will contain the gradient of y.<\/li>\n<li>I moved the <code>tf.while_loop<\/code> inside the session.<\/li>\n<li>Put the evaluation of the gradient inside the body function<\/li>\n<\/ol>\n\n<p>I think the problem before was that when you define <code>grad = tf.gradients(y, results[2])<\/code> the loop has not run yet, so y is not a function of x. Therefore, there is no gradient.<\/p>\n\n<p>Hope this helps.<\/p>\n"}]}
+{"QuestionId":57977880,"AnswerCount":2,"Tags":"<deep-learning><pytorch>","CreationDate":"2019-09-17T15:50:06.200","AcceptedAnswerId":null,"OwnerUserId":288609.0,"Title":"any script to test the installation of Pytorch","Body":"<p>I have installed the pytorch, and would like to check are there any script to test whether the installation is correct, e.g., whether it can enable CUDA or not, etc?<\/p>\n","answers":[{"AnswerId":"57978762","CreationDate":"2019-09-17T16:50:40.413","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>If you installed it from <a href=\"https:\/\/pytorch.org\/get-started\/locally\/#installing-on-linux\" rel=\"nofollow noreferrer\">here<\/a> you are doing fine.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/l7mRT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/l7mRT.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Check <a href=\"https:\/\/stackoverflow.com\/questions\/48152674\/how-to-check-if-pytorch-is-using-the-gpu\">this<\/a>:<\/p>\n\n<pre><code>import torch\ndev = torch.device(\"cuda\") if torch.cuda.is_available() else torch.device(\"cpu\")\nprint(dev)\n<\/code><\/pre>\n\n<p>If you have your GPU installed correctly you should have <code>nvidia-smi<\/code>.\n(On Windows it should be inside <code>C:\\Program Files\\NVIDIA Corporation\\NVSMI<\/code>)<\/p>\n"},{"AnswerId":"57986484","CreationDate":"2019-09-18T06:38:02.947","ParentId":null,"OwnerUserId":"6210807","Title":null,"Body":"<p>You can use the <a href=\"https:\/\/raw.githubusercontent.com\/pytorch\/pytorch\/master\/torch\/utils\/collect_env.py\" rel=\"nofollow noreferrer\"><code>collect_env.py<\/code><\/a> script provided in the PyTorch <a href=\"https:\/\/github.com\/pytorch\/pytorch\/tree\/master\/torch\/utils\" rel=\"nofollow noreferrer\">utils<\/a> folder.<\/p>\n\n<p>Its output is as follows:<\/p>\n\n<pre><code>Collecting environment information...\nPyTorch version: 1.2.0\nIs debug build: No\nCUDA used to build PyTorch: 10.0.130\n\nOS: Ubuntu 16.04.6 LTS\nGCC version: (Ubuntu 5.4.0-6ubuntu1~16.04.11) 5.4.0 20160609\nCMake version: version 3.14.6\n\nPython version: 3.7\nIs CUDA available: Yes\nCUDA runtime version: Could not collect\nGPU models and configuration: GPU 0: GeForce RTX 2080\nNvidia driver version: 410.48\ncuDNN version: \/usr\/lib\/x86_64-linux-gnu\/libcudnn.so.7.4.1\n\nVersions of relevant libraries:\n[pip] numpy==1.16.4\n[pip] torch==1.2.0\n[pip] torchsample==0.1.3\n[pip] torchsummary==1.5.1\n[pip] torchvision==0.4.0a0+6b959ee\n[conda] blas                      1.0                         mkl  \n[conda] mkl                       2019.4                      243  \n[conda] mkl-service               2.0.2            py37h7b6447c_0  \n[conda] mkl_fft                   1.0.14           py37ha843d7b_0  \n[conda] mkl_random                1.0.2            py37hd81dba3_0  \n[conda] pytorch                   1.2.0           py3.7_cuda10.0.130_cudnn7.6.2_0    pytorch\n[conda] torchsample               0.1.3                    pypi_0    pypi\n[conda] torchsummary              1.5.1                    pypi_0    pypi\n[conda] torchvision               0.4.0                py37_cu100    pytorch\n<\/code><\/pre>\n"}]}
+{"QuestionId":57978105,"AnswerCount":0,"Tags":"<keras><mlp>","CreationDate":"2019-09-17T16:04:09.907","AcceptedAnswerId":null,"OwnerUserId":262653.0,"Title":"multiple output in keras logisitc code snippet","Body":"<p>I have the following code snippet. how can I adapt this for 2 output variables instead of one in this logistic regression example?<\/p>\n\n<pre><code># split into input (X) and output (Y) variables\nX = dataset[:,0:13]\nY = dataset[:,13]\n# define base model\ndef baseline_model():\n    # create model\n    model = Sequential()\n    model.add(Dense(13, input_dim=13, kernel_initializer='normal',  activation='relu'))\n    model.add(Dense(1, kernel_initializer='normal'))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57978458,"AnswerCount":2,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-17T16:29:51.847","AcceptedAnswerId":57978624.0,"OwnerUserId":1090446.0,"Title":"simple element wise with Keras over TF","Body":"<p>i am trying to implement the following in TensorFlow:<\/p>\n\n<pre><code>Input * const\n<\/code><\/pre>\n\n<p>matrix multiplication of 640x800x6\n Here is the code \n    ssValues = np.zeros(shape=(6,640,800),dtype=np.float16)<\/p>\n\n<pre><code>inputPlaceHolder = tf.compat.v1.placeholder(shape=(6,640,800), name='InputTensor', dtype=tf.dtypes.float16)\ninputLayer = tf.keras.Input(shape=(6,640,800,),\n                            batch_size=1,\n                            name='inputLayer',\n                            dtype=tf.dtypes.float16,\n                            tensor=inputPlaceHolder)\n\nssConstant = tf.constant(ssValues, dtype=tf.dtypes.float16, shape=(6,640,800), name='ss')\nssm = tf.keras.layers.Multiply()([inputPlaceHolder,inputPlaceHolder])\nmodel = tf.keras.models.Model(inputs=inputLayer, outputs=ssm)\n\ninput = np.zeros(shape=(6,640,800),dtype=np.float16)\n\noutput = model.predict(input)\n<\/code><\/pre>\n\n<p>i get the following error:\nValueError: ('Error when checking model input: expected no data, but got:', array([[[1., 1., 1., ..., 1., 1., 1.],<\/p>\n\n<ul>\n<li><p>how to overcome this error and run the predict function ?<\/p><\/li>\n<li><p>why tf.keras.layers.multiply doesn't return a Layer object ?<\/p><\/li>\n<\/ul>\n","answers":[{"AnswerId":"57978583","CreationDate":"2019-09-17T16:39:14.997","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>When you use <code>Input(shape)<\/code> you have a placeholder already. It doesn't make sense to create a placeholder to the pass it to <code>Input(tensor=placeholder)<\/code> because this is not how Keras works. <\/p>\n\n<p>You must:<\/p>\n\n<pre><code>inputs = Input(shape=(6,640,800))\nssm_tensor = Multiply()([inputs, inputs])\nmodel = Model(inputs, ssm)\n<\/code><\/pre>\n\n<p>Since you always have a batch size with Keras:<\/p>\n\n<pre><code>input = np.zeros(shape=(1,6,640,800))\n<\/code><\/pre>\n"},{"AnswerId":"57978624","CreationDate":"2019-09-17T16:41:45.643","ParentId":null,"OwnerUserId":"12076271","Title":null,"Body":"<p>Your issue comes from the fact that you declared your operation on a v1 placeholder, when it should simply use the <code>inputLayer<\/code> (which already acts as a placeholder for inputs following the provided specification).<\/p>\n\n<p>Additionnally, you wrote a multiplication that returns $x \\times x$, when I think you wanted $x \\times constant$ ; so here would be the code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>inputLayer = tf.keras.Input(shape=(6,640,800,),\n                            batch_size=1,\n                            name='inputLayer',\n                            dtype=tf.dtypes.float16)\nssConstant = tf.constant(  # also fixed a shape issue here\n    ssValues, dtype=tf.dtypes.float16, shape=(1, 6,640,800), name='ss'\n)\nssm = tf.keras.layers.Multiply(dtype=tf.dtypes.float16)([inputLayer, ssConstant])\nmodel = tf.keras.models.Model(inputs=inputLayer, outputs=ssm)\n\ninputs = np.zeros(shape=(1,6,640,800), dtype=np.float16)\noutput = model.predict(inputs)\n<\/code><\/pre>\n\n<p>Furthermore, since this is not an actual model, in the sense that it uses a constant and not learnable weights, you might want to use <code>tf.keras.backend.function<\/code> instead of <code>tf.keras.Model<\/code> (but that is really up to you).<\/p>\n\n<p>Note that the shapes are probably not suited to what you actually want, with the batch-size of 1... Please consider using a batch-size of 6 to remove the useless dimensions.<\/p>\n"}]}
+{"QuestionId":57978510,"AnswerCount":2,"Tags":"<python><pytorch>","CreationDate":"2019-09-17T16:33:19.223","AcceptedAnswerId":57979420.0,"OwnerUserId":9464073.0,"Title":"How do I reshape this tensor?","Body":"<p>I have a <code>torch.tensor<\/code> that looks like this:<\/p>\n\n<pre><code>tensor([[[A,B,C],\n         [D,E,F],\n         [G,H,I]],\n\n        [[J,K,L],\n         [M,N,O],\n         [P,Q,R]]]\n<\/code><\/pre>\n\n<p>I want to reshape this tensor so that its dimensions are <code>(18, 1)<\/code>. I want the new tensor to look like this:<\/p>\n\n<pre><code>tensor([A,\n        J,\n        B,\n        K,\n        C,\n        L,\n        D,\n        M,\n       ...\n        I,\n        R])\n<\/code><\/pre>\n\n<p>I've tried <code>tensor.view(-1,1)<\/code> but this doesn't work..<\/p>\n","answers":[{"AnswerId":"57978712","CreationDate":"2019-09-17T16:47:36.380","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>Both <code>view<\/code> or <code>reshape<\/code> work in here:<\/p>\n\n<pre><code>t = torch.tensor([[[1,2,3],\n         [4,5,6],\n         [7,8,9]],\n\n        [[1,2,3],\n         [1,2,3],\n         [1,2,3]]])\nprint(t.size())\nt = t.permute(1,2,0).reshape(-1,1)\nprint(t)\nprint(t.size())\n<\/code><\/pre>\n"},{"AnswerId":"57979420","CreationDate":"2019-09-17T17:39:04.660","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<pre class=\"lang-py prettyprint-override\"><code>a = torch.arange(18).view(2,3,3)\n\nprint(a)\n#tensor([[[ 0,  1,  2],\n#         [ 3,  4,  5],\n#         [ 6,  7,  8]],\n#\n#        [[ 9, 10, 11],\n#         [12, 13, 14],\n#         [15, 16, 17]]])\n\naa = a.permute(1,2,0).flatten()\n\nprint(aa)\n#tensor([ 0,  9,  1, 10,  2, 11,  3, 12,  4, 13,  5, 14,  6, 15,  7, 16,  8, 17])\n<\/code><\/pre>\n"}]}
+{"QuestionId":57979255,"AnswerCount":2,"Tags":"<python><python-3.x><keras><multilabel-classification><imbalanced-data>","CreationDate":"2019-09-17T17:26:23.573","AcceptedAnswerId":57979861.0,"OwnerUserId":4764604.0,"Title":"Passing a list as loss_weights, it should have one entry per model output. Keras tells me that the model has 1 output but I thought having more","Body":"<p>I have a dataset <code>df<\/code> for a multiclass classification problem. I have a huge class imbalance. Namely, <code>grade_F<\/code> and <code>grade_G<\/code>.<\/p>\n\n<pre><code>&gt;&gt;&gt; percentage = 1. \/ df['grade'].value_counts(normalize=True)\n&gt;&gt;&gt; print(percentage )\n\nB    0.295436\nC    0.295362\nA    0.204064\nD    0.136386\nE    0.048788\nF    0.014684\nG    0.005279\n<\/code><\/pre>\n\n<p>At the same time I have I have very innacurrate predictions for less represented classes, as one can see <a href=\"https:\/\/datascience.stackexchange.com\/questions\/60194\/improving-classifcation-when-some-are-less-represented\">here<\/a>.<\/p>\n\n<p>I have a neural network with an output dimension of 7. I mean the array I want to predict is :<\/p>\n\n<pre><code>&gt;&gt;&gt; print(y_train.head())\n        grade_A  grade_B  grade_C  grade_D  grade_E  grade_F  grade_G\n689526        0        1        0        0        0        0        0\n523913        1        0        0        0        0        0        0\n266122        0        0        1        0        0        0        0\n362552        0        0        0        1        0        0        0\n484987        1        0        0        0        0        0        0\n...\n<\/code><\/pre>\n\n<p>So I tried the following neural network:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense, Dropout\nfrom keras.constraints import maxnorm\n\ndef create_model(input_dim, output_dim):\n    print(output_dim)\n    # create model\n    model = Sequential()\n    # input layer\n    model.add(Dense(100, input_dim=input_dim, activation='relu', kernel_constraint=maxnorm(3)))\n    model.add(Dropout(0.2))\n\n    # hidden layer\n    model.add(Dense(60, activation='relu', kernel_constraint=maxnorm(3)))\n    model.add(Dropout(0.2))\n\n    # output layer\n    model.add(Dense(output_dim, activation='softmax'))\n\n    # Compile model\n    model.compile(loss='categorical_crossentropy', loss_weights=lossWeights, optimizer='adam', metrics=['accuracy'])\n    return model\n\nfrom keras.callbacks import ModelCheckpoint\nfrom keras.models import load_model\n\nmodel = create_model(x_train.shape[1], y_train.shape[1])\n\nepochs =  35\nbatch_sz = 64\n\nprint(\"Beginning model training with batch size {} and {} epochs\".format(batch_sz, epochs))\n\ncheckpoint = ModelCheckpoint(\"lc_model.h5\", monitor='val_acc', verbose=0, save_best_only=True, mode='auto', period=1)\n# train the model\nhistory = model.fit(x_train.as_matrix(),\n                y_train.as_matrix(),\n                validation_split=0.2,\n                epochs=epochs,  \n                batch_size=batch_sz, # Can I tweak the batch here to get evenly distributed data ?\n                verbose=2,\n                callbacks=[checkpoint])\n\n# revert to the best model encountered during training\nmodel = load_model(\"lc_model.h5\")\n<\/code><\/pre>\n\n<p>So I fed a vector of weights inversely proportional to class frequency:<\/p>\n\n<pre><code>lossWeights = df['grade'].value_counts(normalize=True)\nlossWeights = lossWeights.sort_index().tolist()\n<\/code><\/pre>\n\n<p>However it told me that the output was of size 1 :<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-66-bf262c22c9dc&gt; in &lt;module&gt;\n      2 from keras.models import load_model\n      3 \n----&gt; 4 model = create_model(x_train.shape[1], y_train.shape[1])\n      5 \n      6 epochs =  35\n\n&lt;ipython-input-65-9290b177bace&gt; in create_model(input_dim, output_dim)\n     19 \n     20     # Compile model\n---&gt; 21     model.compile(loss='categorical_crossentropy', loss_weights=lossWeights, optimizer='adam', metrics=['accuracy'])\n     22     return model\n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\engine\\training.py in compile(self, optimizer, loss, metrics, loss_weights, sample_weight_mode, weighted_metrics, target_tensors, **kwargs)\n    178                                  'The model has ' + str(len(self.outputs)) +\n    179                                  ' outputs, but you passed loss_weights=' +\n--&gt; 180                                  str(loss_weights))\n    181             loss_weights_list = loss_weights\n    182         else:\n\nValueError: 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=[4.9004224502112255, 3.3848266392035704, 3.385677583130476, 7.33212052000478, 20.49667767920116, 68.10064134188455, 189.42024013722127]\n<\/code><\/pre>\n","answers":[{"AnswerId":"57979861","CreationDate":"2019-09-17T18:11:48.230","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>What you are looking for is <code>class_weight<\/code> in the <code>fit<\/code> function. <\/p>\n\n<pre><code>weights = {0: 1 \/ 0.204064,\n           1: 1 \/ 0.295436, \n           2: 1 \/ 0.295362,\n           3: 1 \/ 0.136386, \n           4: 1 \/ 0.048788,\n           5: 1 \/ 0.014684,\n           6: 1 \/ 0.005279}\n<\/code><\/pre>\n\n<p>You may want to reduce their sizes, because you have weights from about 3 to 200 there, but the relation is more important. <\/p>\n\n<p>Then:<\/p>\n\n<pre><code>model.fit(....\n          class_weight = weights, \n         )\n<\/code><\/pre>\n"},{"AnswerId":"57979604","CreationDate":"2019-09-17T17:51:01.530","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p><code>loss_weights<\/code> does not weight different classes, it weights different outputs. Your model has only one output. Yes, that output is a list, but it is still treated as a single entity by keras.<\/p>\n\n<p>A model made with the functional API can have multiple outputs, each with its own loss function. While training the model, the loss is then defined to be the sum of all the loss functions applied to their respective outputs. In that case, <code>loss_weights<\/code> can be used to weight the different outputs.<\/p>\n\n<p>However, I do not believe it is useful for what <em>you<\/em> want to do.<\/p>\n"}]}
+{"QuestionId":57979365,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-17T17:34:08.860","AcceptedAnswerId":null,"OwnerUserId":1081069.0,"Title":"How to retrieve the value of eval_metrics in TensorFlow","Body":"<p>I have a working pipeline that train a model and evaluate metrics (defined in <code>EstimatorSpec<\/code>). I can also see the aggregated value of metrics (for instance, <code>accuracy<\/code>) in Tensorboard. <\/p>\n\n<p>Now I would like to calculate some business quantity as a function of aggregated eval_metrics for the entire evaluation dataset. (as a made-up example, consider <code>f(accuracy)=sqrt(accuracy)<\/code>). <\/p>\n\n<p>But I don't know how to retrieve the tensor of accuracy in the first place. I tried \n<code>tf.get_default_graph().get_tensor_by_name(\"accuracy:0\")<\/code> and I get<\/p>\n\n<pre><code>\"The name 'accuracy:0' refers to a Tensor which does not exist. The operation, 'accuracy', does not exist in the graph.\"\n<\/code><\/pre>\n\n<p>Do I miss anything?<\/p>\n\n<p>Thanks<\/p>\n","answers":[]}
+{"QuestionId":57979546,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-17T17:47:13.357","AcceptedAnswerId":null,"OwnerUserId":1090446.0,"Title":"exporting TensorFlow model with no variables","Body":"<p>i am exporting Keras TF model without a luck:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nssValues = np.zeros(shape=(640,800,6),dtype=np.float16)\nssValues += 3.\n\nssKerasConstant = tf.keras.backend.constant(value=ssValues, dtype=tf.dtypes.float16, shape=(1,640,800,6));\ninputLayer = tf.keras.Input(shape=(640,800,6),\n                            name='inputLayer',\n                            batch_size=None,\n                            dtype=tf.dtypes.float16)\n\nssConstant = tf.constant(ssValues, dtype=tf.dtypes.float16, shape=(1,640,800,6), name='ss')\nssm = tf.keras.layers.Multiply()([inputLayer,ssKerasConstant])\nmodel = tf.keras.models.Model(inputs=inputLayer, outputs=ssm)\ntf.keras.experimental.export_saved_model(model, '~\/models\/model7.pb')\n<\/code><\/pre>\n\n<p>and i get the following error:\n graph = inputs[0].graph\nIndexError: list index out of range<\/p>\n\n<p>even though i am able to predict the model.<\/p>\n","answers":[]}
+{"QuestionId":57979712,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2019-09-17T17:59:55.443","AcceptedAnswerId":null,"OwnerUserId":12081115.0,"Title":"AttributeError: module 'tensorflow' has no attribute 'get_default_graph'. I am getting this error when adding a new layer to a model","Body":"<pre><code>ml = Sequential()\nml.add(LSTM(64,dropout=0.5, recurrent_dropout=0.5,return_sequences=True))\n<\/code><\/pre>\n\n<p>This is giving me a error: <code>AttributeError: module 'tensorflow' has no attribute 'get_default_graph'<\/code>. <\/p>\n\n<p><code>ml<\/code> is the name of the model. I am unable to add any layers to a model. Please help.<\/p>\n\n<p>The dependencies that I am using are:<\/p>\n\n<ul>\n<li>Tensorflow: 2.0.0-beta1<\/li>\n<li>Keras: 2.2.4<\/li>\n<li>Python: 3.7.3<\/li>\n<\/ul>\n","answers":[]}
+{"QuestionId":57979815,"AnswerCount":0,"Tags":"<tensorflow><distributed-tensorflow>","CreationDate":"2019-09-17T18:08:35.977","AcceptedAnswerId":null,"OwnerUserId":1304861.0,"Title":"Is TLS supported in Distributed Tensorflow gRPC communication","Body":"<p>I was wondering is TLS supported in current distributed tensorflow with gRPC?<\/p>\n\n<p>I am reading through the code, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.14\/tensorflow\/core\/distributed_runtime\/rpc\/grpc_server_lib.h#L105\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.14\/tensorflow\/core\/distributed_runtime\/rpc\/grpc_server_lib.h#L105<\/a> <\/p>\n\n<p>the implementation of GetServerCredentials\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.14.0\/tensorflow\/core\/distributed_runtime\/rpc\/grpc_server_lib.cc#L417\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.14.0\/tensorflow\/core\/distributed_runtime\/rpc\/grpc_server_lib.cc#L417<\/a>  <\/p>\n\n<p>Looks like TLS is not supported (<a href=\"https:\/\/grpc.io\/docs\/guides\/auth\/\" rel=\"nofollow noreferrer\">https:\/\/grpc.io\/docs\/guides\/auth\/<\/a>)<\/p>\n\n<p>So my questions is: \n1) double check TLS is not supported in case i missed something.\n2) if currently not support, what's the plan?<\/p>\n","answers":[]}
+{"QuestionId":57979968,"AnswerCount":0,"Tags":"<python-2.7><tensorflow><matplotlib><typing><pep8>","CreationDate":"2019-09-17T18:20:40.063","AcceptedAnswerId":null,"OwnerUserId":851699.0,"Title":"pep8 gives strange failures when typing module is combined with matplotlib or tensorflow","Body":"<p>I'm getting a strange errors when running <code>pep8<\/code> in Python2.7 on files using the <code>typing<\/code> module.<\/p>\n\n<p>If I make a file <code>my_file.py<\/code> containing <\/p>\n\n<pre><code>import typing\nimport tensorflow\n<\/code><\/pre>\n\n<p>Then <code>py.test -k pep8 my_file.py<\/code> gives <code>Segmentation fault (core dumped)<\/code>.  Removing either of the two imports, or reversing their order, stops the error.  This only seems to happen in <code>pep8<\/code>, not when running normally in Python.<\/p>\n\n<p>It causes a different error with <code>matplotlib<\/code>:<\/p>\n\n<pre><code>import typing\nimport matplotlib\n<\/code><\/pre>\n\n<p>Here, <code>py.test -k pep8  pepcheck_file_a.py<\/code> gives <code>E   RuntimeError: maximum recursion depth exceeded while calling a Python object<\/code>.  <\/p>\n\n<p>As I mentioned, in either case, reversing the import order to: <\/p>\n\n<pre><code>import tensorflow\nimport typing\n<\/code><\/pre>\n\n<p>Stops the error.<\/p>\n\n<p>It's annoying in a large codebase to have to make sure that typing is always imported <em>after<\/em> various other packages (one has to manually rearrange the imports in each file).  <\/p>\n\n<p>Is there some easier solution to this?<\/p>\n","answers":[]}
+{"QuestionId":57980047,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-17T18:25:50.753","AcceptedAnswerId":null,"OwnerUserId":11952444.0,"Title":"No change in loss while training Deep Convolutional GAN on CIFAR10 dataset","Body":"<p>The model is not training at all. I think the gradients are not flowing properly or maybe there is some logical error.<\/p>\n\n<p>There is no change in loss even after few epochs.\nI tried with many variations of the code but could not debug the problem.<\/p>\n\n<p><strong>No change in discriminator and generator loss.<\/strong><\/p>\n\n<p>Below is my code:<\/p>\n\n<pre><code>from keras.layers import Conv2D, Dense, LeakyReLU, BatchNormalization, Reshape \nfrom keras.layers import Flatten, Input, MaxPooling2D, Dropout, AlphaDropout, Conv2DTranspose\nfrom keras.optimizers import Adam\nfrom keras.models import Model, Sequential\nfrom keras.datasets import cifar10\nimport keras.backend as K\nimport cv2\nimport pickle\nimport numpy as np\nimport os\nimport matplotlib.pyplot as plt\n\nclass CifarGAN:\n    def __init__(self, img_rows, img_cols, channels, epochs, batch_size, latent_dim):\n        self.img_rows = img_rows\n        self.img_cols = img_cols\n        self.channels = channels\n        self.img_shape = (self.img_rows, self.img_cols, self.channels)\n        self.epochs = epochs\n        self.batch_size = batch_size\n        self.latent_dim = latent_dim\n\n        self.optimizer = Adam(0.0002, beta_1=0.5)\n\n        # build and compile the discriminator\n        self.discriminator = self.build_discriminator()\n        self.discriminator.compile(optimizer=self.optimizer, loss='binary_crossentropy', metrics=['accuracy'])\n\n        # build and compile generator\n        self.generator = self.build_generator()\n        self.generator.compile(optimizer=self.optimizer, loss='binary_crossentropy')        \n\n        # input noise\n        noise = Input((self.latent_dim, ))\n        img = self.generator(noise)\n\n        # for the combined model we will only train the generator\n        self.discriminator.trainable = False\n\n        valid = self.discriminator(img)\n\n        self.combined = Model(inputs=noise, outputs=valid, name='combined')\n        self.combined.compile(optimizer=self.optimizer, loss='binary_crossentropy')\n        # self.combined.summary()\n\n    def build_generator(self):\n\n        model = Sequential(name='generator')\n\n        model.add(Dense(256*4*4, name='gen_dense_2'))\n        model.add(LeakyReLU(alpha=0.2, name='gen_lr_2'))\n\n        model.add(Reshape((4, 4, 256), name='gen_reshape_1'))\n\n        model.add(Conv2DTranspose(128, kernel_size=(4, 4), strides=(2, 2), padding='same', kernel_initializer='he_normal'))\n        # model.add(BatchNormalization(name='gen_bc_1'))\n        model.add(LeakyReLU(alpha=0.2, name='gen_lr_3'))\n        # model.add(AlphaDropout(0.1))\n\n        model.add(Conv2DTranspose(128, kernel_size=(4, 4), strides=(2, 2), padding='same', kernel_initializer='he_normal'))\n        # model.add(BatchNormalization(name='gen_bc_2'))\n        model.add(LeakyReLU(alpha=0.2, name='gen_lr_4'))\n        # model.add(AlphaDropout(0.1))\n\n        model.add(Conv2DTranspose(256, kernel_size=(4, 4), strides=(2, 2), padding='same', kernel_initializer='he_normal'))\n        # model.add(BatchNormalization(name='gen_bc_3'))\n        model.add(LeakyReLU(alpha=0.2, name='gen_lr_5'))\n        # model.add(AlphaDropout(0.1))\n\n        model.add(Conv2D(3, kernel_size=(3, 3), padding='same', activation='tanh'))\n\n        noise = Input((self.latent_dim, ), name='gen_input')\n        img = model(noise)\n        # model.summary()\n        return Model(noise, img)\n\n\n    def build_discriminator(self):\n        model = Sequential(name='discriminator')\n\n        model.add(Conv2D(128, kernel_size=(3, 3), strides=(1, 1), padding='same', kernel_initializer='he_normal', name='dis_conv_1'))\n        model.add(BatchNormalization(name='dis_bc_1'))\n        model.add(LeakyReLU(alpha=0.2, name='dis_lr_1'))\n\n        model.add(Conv2D(128, kernel_size=(3, 3), strides=(1, 1), padding='same', kernel_initializer='he_normal', name='dis_conv_2'))\n        model.add(BatchNormalization(name='dis_bc_2'))\n        model.add(LeakyReLU(alpha=0.2, name='dis_lr_2'))\n\n        model.add(Flatten(name='dis_flatten'))\n\n        model.add(Dense(256, name='dis_dense_1'))\n        model.add(LeakyReLU(alpha=0.2, name='dis_lr_4'))\n\n        model.add(Dense(1, activation='sigmoid'))\n\n        image = Input(self.img_shape, name='dis_input')\n        valid = model(image)\n        # model.summary()\n\n        return Model(image, valid)\n\n    def generate_real_samples(self):\n        gen_imgs = np.random.normal(0, 1, (self.batch_size, self.latent_dim))\n        labels = np.ones((self.batch_size, 1))\n        return gen_imgs, labels\n\n    def generate_fake_samples(self):\n        noise = np.random.normal(0, 1, (self.batch_size, self.latent_dim))\n        imgs = self.generator.predict(noise)\n        labels = np.zeros((self.batch_size, 1))\n        return imgs, labels\n\n    def sample_original_image(self):\n        idx = np.random.randint(0, self.samples, self.batch_size)\n        imgs = self.X_train[idx]\n        labels = np.ones((self.batch_size, 1))\n        return imgs, labels\n\n    def get_data(self):\n        (X_train, _), (_, _) = cifar10.load_data()\n        self.samples = X_train.shape[0]\n        # Rescale the image to -1 to 1\n        X_train = (X_train.astype('float32') - 127.5) \/ 127.5\n        return X_train\n\n    def train(self):\n        self.X_train = self.get_data()\n\n        mini_batch = int(self.batch_size \/ 2)\n        for epoch in range(1, self.epochs+1):\n            for batch in range(mini_batch):\n\n                imgs, labels = self.sample_original_image()\n                gen_imgs, gen_labels = self.generate_fake_samples()\n\n                # Train the discriminator\n                d_loss_real, _ = self.discriminator.train_on_batch(imgs, labels)\n                d_loss_fake, _ = self.discriminator.train_on_batch(gen_imgs, gen_labels)\n\n                # The generator wants the discriminator to label the generated samples\n                # as valid (ones)\n                noise, valid_y = self.generate_real_samples()\n\n                # Train the generator\n                g_loss = self.combined.train_on_batch(noise, valid_y)\n\n            # Plot the progress every 100 iterations:\n            if epoch % 10 == 0:\n                print (\"epoch: %d \/ %d [D loss real: %f, D loss fake: %f] [G loss: %f]\" % (epoch, self.epochs, d_loss_real, d_loss_fake, g_loss))\n\n\nimg_rows, img_cols, channels, epochs, batch_size, latent_dim = 32, 32, 3, 100, 64, 100\ncifar_gan = CifarGAN(img_rows, img_cols, channels, epochs, batch_size, latent_dim)\ncifar_gan.train()\n<\/code><\/pre>\n\n<p><strong>Epochs<\/strong><\/p>\n\n<pre><code>epoch: 0 \/ 100 [D loss real: 0.000000, D loss fake: 15.942385] [G loss: 0.000000]\nepoch: 10 \/ 100 [D loss real: 0.000000, D loss fake: 15.942385] [G loss: 0.000000]\nepoch: 20 \/ 100 [D loss real: 0.000000, D loss fake: 15.942385] [G loss: 0.000000]\n<\/code><\/pre>\n\n<p>Thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":57980088,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><computer-vision><google-coral>","CreationDate":"2019-09-17T18:29:12.783","AcceptedAnswerId":null,"OwnerUserId":5050423.0,"Title":".tflite model initializing the same detection boxes everytime","Body":"<p>I'm trying to implement a trained-from-scratch model onto a USB webcam attached to a google dev board. I cut the training at 15k. The model had not converged fully at 15k but I wanted to workout how to freeze the graph, tflite and move it onto a google coral dev board. I realise this may be an issue in terms of accuracy. MobileNetv2.<\/p>\n\n<p>Below are that of a ceiling and a desk. The detection boxes appear in the exact same place each and everytime, no matter what the camera is looking at. The boxes do not move, nor does the percent change when the camera moves. <\/p>\n\n<p>Train model<\/p>\n\n<pre><code>python train.py --logtostderr --train_dir=training --pipeline_config_path=training\/ssd_mobilenet_v1_coco.config\n<\/code><\/pre>\n\n<p>Freeze model<\/p>\n\n<pre><code>python export_tflite_ssd_graph.py --pipeline_config_path=training\/ssd_mobilenet_v2_coco.config --trained_checkpoint_prefix=training\/model.ckpt-9070 --output_directory=inference_graph --add_postprocessing_op=true \n<\/code><\/pre>\n\n<p>Quantize<\/p>\n\n<pre><code>tflite_convert --graph_def_file=inference_graph\/tflite_graph.pb --output_file=inference_graph\/detect.tflite --inference_type=QUANTIZED_UINT8 --input_shapes=1,300,300,3 --input_arrays=normalized_input_image_tensor --output_arrays=TFLite_Detection_PostProcess,TFLite_Detection_PostProcess:1,TFLite_Detection_PostProcess:2,TFLite_Detection_PostProcess:3 --mean_values=128 --std_dev_values=127 --allow_custom_ops --default_ranges_min=0 --default_ranges_max=6\n<\/code><\/pre>\n\n<p>The model is ran as follows on the google coral dev board;<\/p>\n\n<pre><code>export DISPLAY=:0 &amp;&amp; edgetpu_detect \\\n--source \/dev\/video1:YUY2:1280x720:20\/1  \\\n--model ${DEMO_FILES}\/converted_tflite_quant_model.tflite\n<\/code><\/pre>\n\n<p>The google coral came prepacked with a demo facial recognition model (below). The only difference is the model. The below command works perfectly and tracks my face across the screen. The above command results in the photos below.<\/p>\n\n<pre><code>export DISPLAY=:0 &amp;&amp; edgetpu_detect \\\n--source \/dev\/video1:YUY2:1280x720:20\/1  \\\n--model ${DEMO_FILES}\/mobilenet_ssd_v2_face_quant_postprocess_edgetpu.tflite\n<\/code><\/pre>\n\n<p>What could possibly be going wrong here? <\/p>\n\n<p>Edit, a few months later, I've retrained it with mobilenet_v1, to the exact same result. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/TK9WK.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/TK9WK.jpg\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/PuZ3n.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/PuZ3n.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"58384064","CreationDate":"2019-10-14T21:06:13.360","ParentId":null,"OwnerUserId":"6262757","Title":null,"Body":"<p>Humm,I wonder if the source camera could be a problem also?\nCan you double check if the pixel format, size, and FPS values are correct?\nAlso, on both images that you posted, the process seems to be loading your <code>converted_tflite_quant_model.tflite<\/code> model, are you sure you tried other model?\nWere you able to run the demo with the correct result?<\/p>\n"}]}
+{"QuestionId":57980170,"AnswerCount":1,"Tags":"<python><keras><conv-neural-network>","CreationDate":"2019-09-17T18:35:40.070","AcceptedAnswerId":57980274.0,"OwnerUserId":8855704.0,"Title":"Why are there a set of parameters per input filter in Keras CNN?","Body":"<p>I am trying to understand the reason behind the answer to <a href=\"https:\/\/stackoverflow.com\/questions\/44608552\/keras-cnn-model-parameters-calculation\">this<\/a> question. I was expecting the number of parameters to be:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>total_params = (filter_height * filter_width  + 1) * number_of_filters\n<\/code><\/pre>\n\n<p>BUT you have to multiply the height and width by the number of input channels. Why is this? Isn't there parameter sharing for this dimension? If this is the case, how does this help with feature recognition?<\/p>\n\n<p>I would expect a CNN to be able to infer relationships between channels, but I haven't seen how this is explicitly done.<\/p>\n","answers":[{"AnswerId":"57980274","CreationDate":"2019-09-17T18:43:48.000","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Imagine you have an RGB image and want to pass a single filter: <code>number_of_filters = 1<\/code>. <\/p>\n\n<blockquote>\n  <p>How would this filter treat each of the input channels: R, G and B?    <\/p>\n<\/blockquote>\n\n<p>Should the filter treat all input channels equally? Does the green channel bring the same information as the red? <\/p>\n\n<p>Well, no, each channel has its own information and the filter must consider all input channels, otherwise it will not be looking at the whole image. <\/p>\n\n<p>This is exactly the same as with dense\/fully connected networks, where you have:   <\/p>\n\n<pre><code>total_params =( input_dim + 1 ) * units    \n<\/code><\/pre>\n\n<p>The only difference is that a convolutional filter has height and width. <\/p>\n"}]}
+{"QuestionId":57980214,"AnswerCount":1,"Tags":"<keras><lstm><keras-layer>","CreationDate":"2019-09-17T18:38:18.203","AcceptedAnswerId":null,"OwnerUserId":6499614.0,"Title":"Keras: change LSTM argument return_sequences=True after compile","Body":"<p>I want to use a (2 layered) pretrained LSTM model and I want to add a new LSTM layer into that before the last Dense layer. So the layer I will add will be third. Since the second LSTM for the pretrained model has set return_sequences=True. I am unable to add third LSTM layer. How can I change the configuration of any layer of pretrained model with LSTM to add another LSTM layer. I am not keen on making another copy of model and copying the weights. I want to change in the existing model itself.<\/p>\n\n<p>I am trying it as:<\/p>\n\n<pre><code>model.layers[-1].return_sequences=True\n<\/code><\/pre>\n\n<p>This statement does not generate any error. But the layer configuration still shows return_sequences=False<\/p>\n\n<p>I also tried changing the configuration of layer explicitly as:<\/p>\n\n<pre><code>config=model.layers[-1].get_config()\nconfig['layers']['config']['return_sequences']=True\n<\/code><\/pre>\n\n<p>This changes the value of return_sequences in dictionary config. But I do not know how to change the layer. Something like, is not working.<\/p>\n\n<p>model.layers[-1]=LSTM.from_config(config)<\/p>\n\n<p>It gives init takes at least two arguments.<\/p>\n\n<p>Layer[-1] of model is actually a bidirectional wrapped LSTM.<\/p>\n","answers":[{"AnswerId":"57980306","CreationDate":"2019-09-17T18:45:53.163","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>I think making another copy of the model and copying the weights is your best bet. If you study the source code you'd probably be able to figure out a way to hack another layer on, but that would take effort, would potentially not work, and might break in the future.<\/p>\n"}]}
+{"QuestionId":57980344,"AnswerCount":2,"Tags":"<python><keras><conv-neural-network><recurrent-neural-network>","CreationDate":"2019-09-17T18:49:05.507","AcceptedAnswerId":57980583.0,"OwnerUserId":7003616.0,"Title":"How to \"standardize\" a dataset that has a variable length?","Body":"<p>I'm not sure if I'm wording my question correctly, but here's the gist:\nThe data set, SVC 2004, I am working with has x files, each with y 7-tuples, hence the shape of the dataset becomes (x, y, 7). I have normalized the data and am inserting it into a 1D CNN for feature extraction and an RNN as the classifier. But here is the problem: <strong>y is never the same for each file<\/strong>. This causes a problem when creating the Sequential model as it needs a constant shape. Here is some of my code:<\/p>\n\n<pre><code>\/\/DataPreprocessing\ndef load_dataset_normalized(path):\nfile_names = os.listdir(path)\n\nnum_of_persons = len(file_names)\n\ninitial_starting_point = np.zeros(np.shape([7]))\n\nhighest_num_of_points = find_largest_num_of_points(path)\n\nx_dataset = []\ny_dataset = []\n\ncurrent_file = 0\n\nfor infile in file_names:\n    full_file_name = os.path.join(path, infile)\n    file = open(full_file_name, \"r\")\n    file_lines = file.readlines()\n\n    num_of_points = int(file_lines[0])\n\n    x = []\n    y = []\n    time_stamp = []\n    button_status = []\n    azimuth_angles = []\n    altitude = []\n    pressure = []\n\n    for idx, line in enumerate(file_lines[1:]):\n        idx+=1\n        nums = line.split(' ')\n\n        if idx == 1:\n            nums[2] = 0\n            initial_starting_point = nums\n\n            x.append(float(nums[0]))\n            y.append(float(nums[1]))\n            time_stamp.append(0.0)\n            button_status.append(float(nums[3]))\n            azimuth_angles.append(float(nums[4]))\n            altitude.append(float(nums[5]))\n            pressure.append(float(nums[6]))\n\n        else:\n            x.append(float(nums[0]))\n            y.append(float(nums[1]))\n            time_stamp.append(10)\n            button_status.append(float(nums[3]))\n            azimuth_angles.append(float(nums[4]))\n            altitude.append(float(nums[5]))\n            pressure.append(float(nums[6]))\n\n    max_x = max(x)\n    max_y = max(y)\n    max_azimuth_angle = max(azimuth_angles)\n    max_altitude = max(altitude)\n    max_pressure = max(pressure)\n\n    min_x = min(x)\n    min_y = min(y)\n    min_azimuth_angle = min(azimuth_angles)\n    min_altitude = min(altitude)\n    min_pressure = min(pressure)\n\n    #Alignment normalization:\n    for i in range(num_of_points):\n        x[i] -= float(initial_starting_point[0])\n        y[i] -= float(initial_starting_point[1])\n        azimuth_angles[i] -= float(initial_starting_point[4])\n        altitude[i] -= float(initial_starting_point[5])\n        pressure[i] -= float(initial_starting_point[6])\n\n    #Size normalization\n    for i in range(num_of_points):\n        x[i] = ((x[i] - max_x) \/ (min_x - max_x))\n        y[i] = ((y[i] - max_y) \/ (min_y - max_y))\n        azimuth_angles[i] = ((azimuth_angles[i] - max_azimuth_angle) \/ (min_azimuth_angle - max_azimuth_angle))\n        altitude[i] = ((altitude[i] - max_altitude) \/ (min_altitude - max_altitude))\n        pressure[i] = ((pressure[i] - max_pressure) \/ (min_pressure - max_pressure))\n\n    #data points to dataset\n    x_line = []\n    for i in range (num_of_points):\n        x_line.append(([x[i], y[i], time_stamp[i], button_status[i], azimuth_angles[i], altitude[i], pressure[i]]))\n        if (num_of_points &lt; 713) and (i == num_of_points-1):\n           for idx in range(713 - num_of_points):\n               x_line.append([0, 0, 0, 0, 0, 0, 0])\n\n        if i == num_of_points-1:\n            x_dataset.append(x_line)\n\n    current_file += 1\n\n    infile_without_extension = infile.replace('.TXT','')\n    index_of_s = infile_without_extension.find(\"S\")\n    index_of_num = index_of_s + 1\n    sig_ID = int(infile_without_extension[index_of_num:])\n    if sig_ID &lt; 21:\n        y_dataset.append([1,0])\n    else:\n        y_dataset.append([0,1])\n\nx_dataset = np.array([np.array(xi) for xi in x_dataset])\ny_dataset = np.asarray(y_dataset)\nreturn x_dataset, y_dataset, highest_num_of_points\n\n    \/\/Class that creates my model (creation of model works perfectly)\n    class crnn_model:\n    def build_model(self, input_shape_num, x_train, y_train, x_test, y_test):\n    model = Sequential()\n    model.add(Conv1D(filters=50, kernel_size=3, activation='sigmoid', input_shape = (713, 7)))\n    model.add(MaxPooling1D(pool_size=3))\n    model.add(LSTM(2))\n    model.compile(optimizer='adam', loss='mse', metrics = ['accuracy'])\n    model.summary()\n\n    print(model.fit(x_train, y_train, epochs=50, verbose=0))\n\n    yhat = model.predict(x_test, verbose=0)\n    print(yhat)\n<\/code><\/pre>\n\n<p>I have thought of using the file with the most number of 7-tuples as the shape, as I have hardcoded, for now, with the above code (713). Would this be a good alternative? If not, how do I go about \"standardizing\" or \"normalizing\" the number of points (y) for the input shape of the CNN? <\/p>\n","answers":[{"AnswerId":"57980497","CreationDate":"2019-09-17T18:59:59.430","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>In my experience, people tend to handpick an arbitrary length. Any shorter sequences get padded (generally with zeroes), and any longer sequences get truncated.<\/p>\n\n<p>What length you pick depends on the dataset; generally I make a histogram of the lengths, and then eyeball a number that keeps most of the data without too much unnecessary padding. Often that's right around the 90th-95th percentile. However, if the longest length isn't too much longer than the average length (e.g. it's less than a factor of two), it wouldn't be unreasonable to make that the number.<\/p>\n\n<hr>\n\n<p>Another <em>possible<\/em> option\u2014though I have no experience with this and thus cannot recommend it\u2014is to merely normalize each batch. Any given batch needs to have the same length for all samples, but independent batches can have different lengths. Theoretically, you could partition your data based on sequence lengths, and then make batches with as little padding\/truncating as possible. However, this could take a lot of effort, and might only work if sequence length is not correlated with anything else. I would recommend not attempting this approach.<\/p>\n"},{"AnswerId":"57980583","CreationDate":"2019-09-17T19:06:24.613","ParentId":null,"OwnerUserId":"9758922","Title":null,"Body":"<p>Here you are referring to variable length sequences which you can achieve by letting the <code>input_shape=(None, 7)<\/code>. This says that each batch the <code>y<\/code> could vary <em>but not within a batch<\/em>. So we solve this problem by:<\/p>\n\n<ul>\n<li>Implementing a <a href=\"https:\/\/keras.io\/preprocessing\/sequence\/\" rel=\"nofollow noreferrer\">Sequence<\/a> that takes the data and chops into batches. For each batch it returns <code>(batch_size, y_max_within_batch, 7)<\/code> by <strong>padding<\/strong> using <a href=\"https:\/\/keras.io\/preprocessing\/sequence\/\" rel=\"nofollow noreferrer\">pad_sequences<\/a>.<\/li>\n<li>Setting the <code>input_shape=(None, 7)<\/code> such that at runtime your model can feed on different length inputs. This says I'll however many 7-tuples at each batch.<\/li>\n<li>Optionally using a <a href=\"https:\/\/keras.io\/layers\/core\/\" rel=\"nofollow noreferrer\">Masking Layer<\/a> to mask out the padded sequences so your model doesn't use the padding as features and only predicts on the correctly length of sample. You can also predict on padding which will allow predictions to take into the length of the sequence easier.<\/li>\n<\/ul>\n\n<p>Those are mostly the tools \/ layers you need to handle varying length sequences in Keras. How you use them and effects of padding etc depend on the task you are trying to solve.<\/p>\n"}]}
+{"QuestionId":57980392,"AnswerCount":2,"Tags":"<python><tensorflow><tf.keras><tfrecord>","CreationDate":"2019-09-17T18:52:31.097","AcceptedAnswerId":null,"OwnerUserId":5155356.0,"Title":"TFrecord - How to print features correctly?","Body":"<p>I made a tfrecords file that contains images and segmentation-images. I tried printing one record, but the image and segmentation-image do not match.<\/p>\n\n<p>my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def extract_features(data_record):\n    features = {'label': tf.FixedLenFeature([], tf.string),\n                'image': tf.FixedLenFeature([], tf.string)}\n\n    sample = tf.parse_single_example(data_record, features)\n\n    image = tf.image.decode_image(sample['image'], dtype = tf.float32)        \n    image = tf.reshape(image, [height, width, 3])\n\n    label = tf.image.decode_image(sample['label'], dtype = tf.float32)        \n    label = tf.reshape(label, [height, width, 3])\n\n    return image, label\n\ndef create_dataset(ds_path):\n\n    dataset = tf.data.TFRecordDataset(ds_path)\n    dataset = dataset.shuffle(69863)\n    dataset = dataset.map(extract_features, num_parallel_calls=8)\n    dataset = dataset.repeat()\n    dataset = dataset.batch(1)\n\n    iterator = dataset.make_one_shot_iterator()\n\n    return iterator\n\niterator = create_dataset('train.tfrecords')\nimages, labels = iterator.get_next()\n\nprint(images.shape)\nplt.imshow(tf.keras.backend.eval(images[0]))\nplt.show()\nplt.imshow(tf.keras.backend.eval(labels[0]))\nplt.show()\n<\/code><\/pre>\n\n<p>When I try to show the data like this, the label and image are not matching. I think my mistake lies in my usage of tf.keras.backend.eval(), so I tried to use it only once like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>batch = iterator.get_next()\nbatch = tf.keras.backend.eval(batch)\nimages, labels = batch\n\nprint(images)\n\nplt.imshow(images[0])\nplt.show()\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Tensor(\"IteratorGetNext_6:0\", shape=(?, 540, 676, 3), dtype=float32)\n\nTypeError: Image data of dtype object cannot be converted to float\n<\/code><\/pre>\n\n<p>But then the images and labels are still tensors and I can't print them. Does anyone have an idea, how I could get this code to work?<\/p>\n","answers":[{"AnswerId":"57984268","CreationDate":"2019-09-18T02:31:31.647","ParentId":null,"OwnerUserId":"5136550","Title":null,"Body":"<p>You need to convert the Tensor to np array first before showing.<\/p>\n\n<pre><code>images, labels = iterator.get_next()  # Get next Batch\nimg = np.asarray(images[0])  # First image of the batch\nprint(img)  # print images value\ncv2.imshow(\"Image\",img)  # Display image using cv2; not tested with pyplot.imshow(img)\ncv2.waitKey(0)\n<\/code><\/pre>\n\n<p>[Works perfectly without any issue with tensorflow_gpu 1.14.0 in eager mode]<\/p>\n"},{"AnswerId":"57999271","CreationDate":"2019-09-18T19:00:59.937","ParentId":null,"OwnerUserId":"5155356","Title":null,"Body":"<p><strong>Solution:<\/strong><\/p>\n\n<p>If the whole batch is converted with tf.keras.backend.eval([batch]), both the image and the label are from the same batch. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>images, labels = iterator.get_next()\nimages, labels = tf.keras.backend.eval([image_batch, label_batch])\n\nprint(images.shape)\nplt.imshow(images[0])\nplt.show()\n\nprint(labels.shape)\nplt.imshow(labels[0])\nplt.show()\n<\/code><\/pre>\n"}]}
+{"QuestionId":57980625,"AnswerCount":0,"Tags":"<c++><linux><tensorflow><deep-learning><inference>","CreationDate":"2019-09-17T19:09:09.257","AcceptedAnswerId":null,"OwnerUserId":8250471.0,"Title":"Inference in C++ Tensorflow: Session->Run hangs when ran with input and output parameters supplied","Body":"<p>I trained a FasterRCNN for point detection using TensorFlow and its python API. I was able to do the inference successfully in the same without any issues.\nNow for production requirements, I need to build a standalone application or library which can perform the inference in C++.<\/p>\n\n<p>In order to achieve that I successfully managed to build the TensorFlow library from scratch for C++ API and ran a few test programs.<\/p>\n\n<p>Steps followed for inference on my trained model:<\/p>\n\n<ol>\n<li>Method #1: freeze the graph to *.pb file using modified <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/freeze_graph.py\" rel=\"nofollow noreferrer\">freeze_graph.py<\/a> and restore it.<\/li>\n<li>Method #2: Directly restore from checkpoint using MetaGraphDef and ReadBinaryProto.<\/li>\n<\/ol>\n\n<p>I tried both the methods and the loading of the model seems to happen without any error. I printed the node names as well.<\/p>\n\n<p>After loading the graph I created a handler for image input, previously the input during training was getting handled by QueueRunner, but because it's just inference I created the handler myself with appropriate modifications to it as can be seen below.<\/p>\n\n<pre><code>    Status fasterRcnn::CreateGraphForImage(bool unstack)\n{\n    fileNameVar=Placeholder(iRoot.WithOpName(\"input\"),DT_STRING);\n    auto fileReader=ReadFile(iRoot.WithOpName(\"fileReader\"),fileNameVar);\n    auto imageReader=DecodeJpeg(iRoot.WithOpName(\"jpegReader\"),fileReader,DecodeJpeg::Channels(imageChannels));\n    auto floatCaster=Cast(iRoot.WithOpName(\"floatCaster\"),imageReader,DT_FLOAT);\n    auto dimsExpander=ExpandDims(iRoot.WithOpName(\"dim\"),floatCaster,0);\n    \/\/auto resized=ResizeBilinear(iRoot.WithOpName(\"size\"),dimsExpander,Const(iRoot,{imageSide,imageSide}));\n    \/\/auto div=Div(iRoot.WithOpName(\"normalized\"),resized,{255.f});\n    imageTensorVar=dimsExpander;\/\/div;\n    return iRoot.status();\n\n}\n<\/code><\/pre>\n\n<p>NOTE: iRoot is the private variable of fasterRcnn class-> Scope used by the graph for loading images into tensors<\/p>\n\n<pre><code>   class fasterRcnn\n{\nprivate:\n    Scope iRoot;\/\/graph for loading images into tensors\n    const int imageSide;\n    const int imageChannels;\/\/rgb\n    \/\/load image vars\n    Output fileNameVar;\n    Output imageTensorVar;\n    \/\/\n    std::unique_ptr&lt;Session&gt; fSession;\/\/file session\n    std::unique_ptr&lt;GraphDef&gt; graphDef;\npublic:\n    fasterRcnn(int side,int channels):iRoot(Scope::NewRootScope()),imageSide(side),imageChannels(channels){}\n    Status CreateGraphForImage(bool unstack);\n    Status LoadSavedModel(string &amp;fileName);\/\/via frozen graph\n    Status LoadSavedModel(std::string graph_fn,std::string checkpoint_fn);\/\/via checkpoints\n    Status PredictFromFrozen(Tensor &amp;image,int&amp;results);\/\/\n    Status ReadTensorFromImageFile(string&amp; file_name, Tensor&amp; outTensor);\n    Status ReadFileTensors(string&amp; folder_name,vector&lt;Tensor&gt;&amp; file_tensors);\n\n};\n<\/code><\/pre>\n\n<p><strong>Now the when the time comes for Prediction when I run Session->Run with particular input node name and output node name the program hangs<\/strong>:<\/p>\n\n<pre><code>Status fasterRcnn::PredictFromFrozen(Tensor&amp; image, int&amp; result)\n{\n    vector&lt;Tensor&gt; outTensors;\n    cout&lt;&lt;\"Prediction about to start\"&lt;&lt;endl;\n    TF_CHECK_OK(fSession-&gt;Run({{\"fasterrcnn\/truncated_base_network\/sub\",image}},{\"fasterrcnn\/rcnn\/rcnn_proposal\/GatherV2_1\"},{},&amp;outTensors));\n    cout&lt;&lt;\"Prediction done\"&lt;&lt;endl;\n    auto output_c = outTensors[0].scalar&lt;float&gt;();\n    for (int i=0;i&lt;outTensors.size();i++)\n    {\n        cout &lt;&lt; \"Output dimension of the image\" &lt;&lt; outTensors[i].DebugString()&lt;&lt;\"\\n\";\n    }\n    cout &lt;&lt; \"Output dimension of the image\" &lt;&lt; outTensors[0].DebugString()&lt;&lt;\"\\n\";\n    return Status::OK();\n}\n<\/code><\/pre>\n\n<p><strong>TF_CHECK_OK(fSession->Run({{\"fasterrcnn\/truncated_base_network\/sub\",image}},{\"fasterrcnn\/rcnn\/rcnn_proposal\/GatherV2_1\"},{},&amp;outTensors));<\/strong><\/p>\n\n<p><strong><em>It hangs at the above statement.<\/em><\/strong><\/p>\n\n<p><strong>Up until a few Debug thoughts:<\/strong><\/p>\n\n<ol>\n<li>I thought maybe the QueueRunner might be creating the issue, but according to <a href=\"https:\/\/stackoverflow.com\/a\/35346656\/8250471\">this answer<\/a> I handled that as I created the image handler graph.<\/li>\n<li>I thought maybe I didn't freeze the graph well, that's I loaded directly from the checkpoints, but still, it hangs.<\/li>\n<li>The input to the VGG_16 truncated network in graph expects a 4D float tensor, which I am providing, double-checked that.<\/li>\n<\/ol>\n\n<p><strong>Links to Code and Graph:<\/strong><\/p>\n\n<ol>\n<li><a href=\"https:\/\/drive.google.com\/drive\/folders\/1iO7JIitD_BfHnnLqgzvUygFwAAW6v_5G?usp=sharing\" rel=\"nofollow noreferrer\">CodeBase and GraphPng and GIF hosted on GDrive<\/a><\/li>\n<\/ol>\n\n<p><strong>Question:<\/strong><\/p>\n\n<ol>\n<li>Is there any step that I missing, for successful inferencing in C++.<\/li>\n<li>How to verify if my input and output node names are correct? I referred to the python code and they seem right.<\/li>\n<li>Any steps for debugging that I should follow to provide more info on this?<\/li>\n<\/ol>\n","answers":[]}
+{"QuestionId":57981126,"AnswerCount":1,"Tags":"<python><tensorflow><derivative>","CreationDate":"2019-09-17T19:47:45.927","AcceptedAnswerId":null,"OwnerUserId":2886575.0,"Title":"Shapes of arguments in tensorflow RegisterGradient?","Body":"<p>When composing a tensorflow op, often one wants to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/RegisterGradient\" rel=\"nofollow noreferrer\">register a gradient for that op<\/a>.  The signature for registering a gradient is:<\/p>\n\n<pre><code>@tf.RegisterGradient(\"FooOp\")\ndef _foo_grad(op, grad):\n    return bar\n<\/code><\/pre>\n\n<p>Suppose that <code>op<\/code> has only one input and one output.<br>\nLet <code>op.inputs[0].shape = [...,N,M]<\/code><br>\nLet <code>op.outputs[0].shape = [...,K,L]<\/code><br>\nWhat shape will <code>grad<\/code> be?  What shape should <code>bar<\/code> be?  The number of partial derivatives of <code>FooOp<\/code> is one per input per output, so bar would be shaped something like <code>[...,N,M,K,L]<\/code>.  Is this intuition correct?<\/p>\n","answers":[{"AnswerId":"57981307","CreationDate":"2019-09-17T20:02:07.270","ParentId":null,"OwnerUserId":"29238","Title":null,"Body":"<p>Depends on axes length.<\/p>\n\n<p>I can't answer for sure this particular question, but take a look at Numpy Broadcast rules to see if help you: <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/user\/basics.broadcasting.html\" rel=\"nofollow noreferrer\">https:\/\/docs.scipy.org\/doc\/numpy\/user\/basics.broadcasting.html<\/a><\/p>\n"}]}
+{"QuestionId":57981959,"AnswerCount":0,"Tags":"<python-2.7><tensorflow>","CreationDate":"2019-09-17T20:55:37.667","AcceptedAnswerId":null,"OwnerUserId":12081765.0,"Title":"Cannot assign a device for operation 'Variable_4\/Adam_1'","Body":"<p>I am trying to execute the script 'train.py' which i cloned from the following github repository:<\/p>\n\n<p><a href=\"https:\/\/github.com\/xiaojunxu\/dnn-binary-code-similarity\" rel=\"nofollow noreferrer\">https:\/\/github.com\/xiaojunxu\/dnn-binary-code-similarity<\/a><\/p>\n\n<p>After installing all the requirements of the above repository (requirements.txt), I run 'train.py' and get the following error which I could not find a solution to fix them:<\/p>\n\n<pre><code>2019-09-17 20:43:51.186970: I tensorflow\/core\/platform\/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA\n  Traceback (most recent call last):\n    File \"train.py\", line 124, in &lt;module&gt;\n      gnn.init(LOAD_PATH, LOG_PATH)\n    File \"\/ws\/Gemini\/graphnnSiamese.py\", line 120, in init\n      sess.run(tf.global_variables_initializer())\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 889, in run\n      run_metadata_ptr)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1120, in _run\n      feed_dict_tensor, options, run_metadata)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1317, in _do_run\n      options, run_metadata)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1336, in _do_call\n      raise type(e)(node_def, op, message)\n  tensorflow.python.framework.errors_impl.InvalidArgumentError: Cannot assign a device for operation 'Variable_4\/Adam_1': Operation was explicitly assigned to \/device:GPU:0 but available devices are [ \/job:localhost\/replica:0\/task:0\/device:CPU:0 ]. Make sure the device specification refers to a valid device.\n     [[Node: Variable_4\/Adam_1 = VariableV2[_class=[\"loc:@Variable_4\"], container=\"\", dtype=DT_FLOAT, shape=[64], shared_name=\"\", _device=\"\/device:GPU:0\"]()]]\n\n  Caused by op u'Variable_4\/Adam_1', defined at:\n    File \"train.py\", line 122, in &lt;module&gt;\n      lr = LEARNING_RATE\n    File \"\/ws\/Gemini\/graphnnSiamese.py\", line 93, in __init__\n      optimizer = tf.train.AdamOptimizer(learning_rate=lr).minimize(loss)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/optimizer.py\", line 353, in minimize\n      name=name)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/optimizer.py\", line 474, in apply_gradients\n      self._create_slots([_get_variable_for(v) for v in var_list])\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/adam.py\", line 137, in _create_slots\n      self._zeros_slot(v, \"v\", self._name)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/optimizer.py\", line 796, in _zeros_slot\n      named_slots[_var_key(var)] = slot_creator.create_zeros_slot(var, op_name)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/slot_creator.py\", line 174, in create_zeros_slot\n      colocate_with_primary=colocate_with_primary)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/slot_creator.py\", line 148, in create_slot_with_initializer\n      dtype)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/slot_creator.py\", line 67, in _create_slot_var\n      validate_shape=validate_shape)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 1203, in get_variable\n      constraint=constraint)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 1092, in get_variable\n      constraint=constraint)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 425, in get_variable\n      constraint=constraint)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 394, in _true_getter\n      use_resource=use_resource, constraint=constraint)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 805, in _get_single_variable\n      constraint=constraint)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/variables.py\", line 213, in __init__\n      constraint=constraint)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/variables.py\", line 309, in _init_from_args\n      name=name)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/state_ops.py\", line 133, in variable_op_v2\n      shared_name=shared_name)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/gen_state_ops.py\", line 927, in _variable_v2\n      shared_name=shared_name, name=name)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 787, in _apply_op_helper\n      op_def=op_def)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 2956, in create_op\n      op_def=op_def)\n    File \"\/root\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 1470, in __init__\n      self._traceback = self._graph._extract_stack()  # pylint: disable=protected-access\n\n  InvalidArgumentError (see above for traceback): Cannot assign a device for operation 'Variable_4\/Adam_1': Operation was explicitly assigned to \/device:GPU:0 but available devices are [ \/job:localhost\/replica:0\/task:0\/device:CPU:0 ]. Make sure the device specification refers to a valid device.\n     [[Node: Variable_4\/Adam_1 = VariableV2[_class=[\"loc:@Variable_4\"], container=\"\", dtype=DT_FLOAT, shape=[64], shared_name=\"\", _device=\"\/device:GPU:0\"]()]]\n<\/code><\/pre>\n\n<p>I have found in the suggestions that one should try to change the following to \"0\":<\/p>\n\n<pre><code>os.environ[\"CUDA_VISIBLE_DEVICES\"]= \"0\"\n<\/code><\/pre>\n\n<p>But it did not work for me.<\/p>\n\n<p>I would really appreciate if someone could help me fix the issue. Thanks.<\/p>\n","answers":[]}
+{"QuestionId":57982158,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><keras><loss-function>","CreationDate":"2019-09-17T21:15:15.617","AcceptedAnswerId":57985788.0,"OwnerUserId":4764604.0,"Title":"ValueError: Unknown loss function:focal_loss_fixed when loading model with my custom loss function","Body":"<p>I designed my own loss function. However when trying to revert to the best model encountered during training with <\/p>\n\n<pre><code>model = load_model(\"lc_model.h5\")\n<\/code><\/pre>\n\n<p>I got the following error:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-105-9d09ef163b0a&gt; in &lt;module&gt;\n     23 \n     24 # revert to the best model encountered during training\n---&gt; 25 model = load_model(\"lc_model.h5\")\n\nC:\\ProgramData\\Anaconda3\\lib\\site-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\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\engine\\saving.py in _deserialize_model(f, custom_objects, compile)\n    310                       metrics=metrics,\n    311                       loss_weights=loss_weights,\n--&gt; 312                       sample_weight_mode=sample_weight_mode)\n    313 \n    314         # Set optimizer weights.\n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\engine\\training.py in compile(self, optimizer, loss, metrics, loss_weights, sample_weight_mode, weighted_metrics, target_tensors, **kwargs)\n    137             loss_functions = [losses.get(l) for l in loss]\n    138         else:\n--&gt; 139             loss_function = losses.get(loss)\n    140             loss_functions = [loss_function for _ in range(len(self.outputs))]\n    141         self.loss_functions = loss_functions\n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\losses.py in get(identifier)\n    131     if isinstance(identifier, six.string_types):\n    132         identifier = str(identifier)\n--&gt; 133         return deserialize(identifier)\n    134     if isinstance(identifier, dict):\n    135         return deserialize(identifier)\n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\losses.py in deserialize(name, custom_objects)\n    112                                     module_objects=globals(),\n    113                                     custom_objects=custom_objects,\n--&gt; 114                                     printable_module_name='loss function')\n    115 \n    116 \n\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\keras\\utils\\generic_utils.py in deserialize_keras_object(identifier, module_objects, custom_objects, printable_module_name)\n    163             if fn is None:\n    164                 raise ValueError('Unknown ' + printable_module_name +\n--&gt; 165                                  ':' + function_name)\n    166         return fn\n    167     else:\n\nValueError: Unknown loss function:focal_loss_fixed\n<\/code><\/pre>\n\n<p>Here is the neural network :<\/p>\n\n<pre><code>from keras.callbacks import ModelCheckpoint\nfrom keras.models import load_model\n\nmodel = create_model(x_train.shape[1], y_train.shape[1])\n\nepochs =  35\nbatch_sz = 64\n\nprint(\"Beginning model training with batch size {} and {} epochs\".format(batch_sz, epochs))\n\ncheckpoint = ModelCheckpoint(\"lc_model.h5\", monitor='val_acc', verbose=0, save_best_only=True, mode='auto', period=1)\n\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Dropout\nfrom keras.constraints import maxnorm\n\ndef create_model(input_dim, output_dim):\n    print(output_dim)\n    # create model\n    model = Sequential()\n    # input layer\n    model.add(Dense(100, input_dim=input_dim, activation='relu', kernel_constraint=maxnorm(3)))\n    model.add(Dropout(0.2))\n\n    # hidden layer\n    model.add(Dense(60, activation='relu', kernel_constraint=maxnorm(3)))\n    model.add(Dropout(0.2))\n\n    # output layer\n    model.add(Dense(output_dim, activation='softmax'))\n\n    # Compile model\n    # model.compile(loss='categorical_crossentropy', loss_weights=None, optimizer='adam', metrics=['accuracy'])\n    model.compile(loss=focal_loss(alpha=1), loss_weights=None, optimizer='adam', metrics=['accuracy'])\n\n    return model\n\n# train the model\nhistory = model.fit(x_train.as_matrix(),\n                y_train.as_matrix(),\n                validation_split=0.2,\n                epochs=epochs,  \n                batch_size=batch_sz, # Can I tweak the batch here to get evenly distributed data ?\n                verbose=2,\n                class_weight = weights, # class_weight tells the model to \"pay more attention\" to samples from an under-represented fraud class.\n                callbacks=[checkpoint])\n\n# revert to the best model encountered during training\nmodel = load_model(\"lc_model.h5\")\n<\/code><\/pre>\n\n<p>And here is my loss function:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef focal_loss(gamma=2., alpha=4.):\n\n    gamma = float(gamma)\n    alpha = float(alpha)\n\n    def focal_loss_fixed(y_true, y_pred):\n        \"\"\"Focal loss for multi-classification\n        FL(p_t)=-alpha(1-p_t)^{gamma}ln(p_t)\n        Notice: y_pred is probability after softmax\n        gradient is d(Fl)\/d(p_t) not d(Fl)\/d(x) as described in paper\n        d(Fl)\/d(p_t) * [p_t(1-p_t)] = d(Fl)\/d(x)\n        Focal Loss for Dense Object Detection\n        https:\/\/arxiv.org\/abs\/1708.02002\n\n        Arguments:\n            y_true {tensor} -- ground truth labels, shape of [batch_size, num_cls]\n            y_pred {tensor} -- model's output, shape of [batch_size, num_cls]\n\n        Keyword Arguments:\n            gamma {float} -- (default: {2.0})\n            alpha {float} -- (default: {4.0})\n\n        Returns:\n            [tensor] -- loss.\n        \"\"\"\n        epsilon = 1.e-9\n        y_true = tf.convert_to_tensor(y_true, tf.float32)\n        y_pred = tf.convert_to_tensor(y_pred, tf.float32)\n\n        model_out = tf.add(y_pred, epsilon)\n        ce = tf.multiply(y_true, -tf.log(model_out))\n        weight = tf.multiply(y_true, tf.pow(tf.subtract(1., model_out), gamma))\n        fl = tf.multiply(alpha, tf.multiply(weight, ce))\n        reduced_fl = tf.reduce_max(fl, axis=1)\n        return tf.reduce_mean(reduced_fl)\n    return focal_loss_fixed\n\n# model.compile(loss=focal_loss(alpha=1), optimizer='nadam', metrics=['accuracy'])\n# model.fit(X_train, y_train, epochs=3, batch_size=1000)\n<\/code><\/pre>\n","answers":[{"AnswerId":"57985788","CreationDate":"2019-09-18T05:41:37.607","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>You have to load the <code>custom_objects<\/code> of focal_loss_fixed as shown below:<\/p>\n\n<pre><code>model = load_model(\"lc_model.h5\", custom_objects={'focal_loss_fixed': focal_loss()})\n<\/code><\/pre>\n\n<p>However, if you wish to just perform inference with your model and not further optimization or training your model, you can simply wish to ignore the loss function like this:<\/p>\n\n<pre><code>model = load_model(\"lc_model.h5\", compile=False)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57982210,"AnswerCount":2,"Tags":"<tensorflow><anaconda>","CreationDate":"2019-09-17T21:20:22.493","AcceptedAnswerId":null,"OwnerUserId":9894578.0,"Title":"Can I update TensorFlow in Anaconda?","Body":"<p>I have Anaconda under Windows 8.1, Python 3.7 e TensorFlow 1.14. I tried some pip commands but the 1.14 is the only version installed of TensorFlow.\nThere are other ways to update, for example, version 1.5 ?<\/p>\n\n<p>Thank a lot for any help!<\/p>\n","answers":[{"AnswerId":"57984909","CreationDate":"2019-09-18T03:59:03.583","ParentId":null,"OwnerUserId":"5101148","Title":null,"Body":"<p>Tensorflow 1.14 is the latest release for the time being. If you wanna install a specific version<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>conda install tensorflow=1.5.0\n<\/code><\/pre>\n\n<p>The problem is that tensorflow 1.5 is not compatible with Python 3.7 before 1.13.0rc1.<\/p>\n\n<p>If you need version <code>1.5.0<\/code>, you need to create a virtual environment with Python 3.6 using <code>conda<\/code>.<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>conda create -n py36 python=3.6\nconda activate py36\nconda install tensorflow=1.5.0\n\n# you can also install tensorflow using pip\n# choose the package manager you want\nconda install tensorflow==1.5.0\n<\/code><\/pre>\n\n<p><strong>Note<\/strong>: Don't use <code>pip<\/code> and <code>conda<\/code> to install pkg at the same time in one environment. Check <a href=\"https:\/\/www.anaconda.com\/using-pip-in-a-conda-environment\/\" rel=\"nofollow noreferrer\">Using Pip in a Conda Environment<\/a> for more info.<\/p>\n"},{"AnswerId":"57982307","CreationDate":"2019-09-17T21:30:44.637","ParentId":null,"OwnerUserId":"10637327","Title":null,"Body":"<p>Anaconda and pip don't get on that is why. I believe pip  stops Anaconda seeing it's base env. You need to update anaconda only via conda and once pip is run the Anaconda base env, in my experience, is damaged.<\/p>\n\n<p>To recover the situation you will need to write a new env via conda and load everything again. I can forward the codes if this is helpful. Creating and loading new envs is good practice for a major.project because it ensures code doesn't break from updates.<\/p>\n"}]}
+{"QuestionId":57982223,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-09-17T21:22:03.783","AcceptedAnswerId":null,"OwnerUserId":11674132.0,"Title":"How to selecting images for DataGenerator in Keras for prediction?","Body":"<p>I load the images in the following format for bulk predictions. This method works fine if all images are in a parent folder as follows : <\/p>\n\n<blockquote>\n  <p>..Parent Folder > Sub folder > All image files.<\/p>\n<\/blockquote>\n\n<p>So I need to set my 'directory' = 'Parent Folder'. It reads all the sub folders inside 'Parent folder' and their images.<\/p>\n\n<pre><code>predict_datagen = ImageDataGenerator(rescale = 1.\/255)\n            predict_generator = predict_datagen.flow_from_directory(directory = parent_image_folder,\n                                                target_size = (size, size),\n                                                batch_size = batch_size,\n                                                class_mode = None, \n                                                shuffle = False)\n\npredict_generator.reset()\n            predict_filenames = predict_generator.filenames\n            predictions = model.predict_generator(predict_generator, verbose=1, steps = int(len(predict_generator.filenames) \/ batch_size)+1)\n<\/code><\/pre>\n\n<p>But in my project, I need to load only the images in selected subfolders folders in 'Parent folder'. I could achieve it by copying the selected folders into a separate folder(new parent folder). But is there any other better method?<\/p>\n","answers":[{"AnswerId":"57982730","CreationDate":"2019-09-17T22:20:16.990","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>There is no way to do this without modifying ImageDataGenerator. If you're on linux, the best way I can think of is to create a separate parent directory, and then create symlinks for the selected subfolders. Make sure to set <code>follow_links=True<\/code> in <code>flow_from_directory<\/code>.<\/p>\n"}]}
+{"QuestionId":57982230,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><floating-point>","CreationDate":"2019-09-17T21:23:05.360","AcceptedAnswerId":null,"OwnerUserId":498873.0,"Title":"How to avoid Numpy type conversions?","Body":"<p>Is it possible to avoid or emit warnings for automatic Numpy type conversions from integer and <code>32 bit float arrays<\/code> to <code>64 bit float arrays<\/code>?<\/p>\n\n<p>My use case for this is that I'm developing a large analysis package (20k lines of Python &amp; Numpy), and currently have a mix of float 32 and 64 and some int dtypes, very likely resulting in sub-optimal performance and wasted memory, and basically I'd like to use float32 consistently everywhere.<\/p>\n\n<p>I know that in <em>Tensorflow<\/em> combining two arrays of different dtype gives an error - precisely because implicit conversions to float64 would lead to bad performance and is \"infectious\" on all computed tensors and hard to find where it is introduced if done implicitly.<\/p>\n\n<p>Looking for an option in Numpy or a way to monkey-patch Numpy so that it behaves in this regard like Tensorflow, i.e. emits an error on implicit type conversion on operations like <code>np.add<\/code>, <code>np.mul<\/code> etc., or even better, emits a warning with printed traceback, so that execution continues but I see where it happened. Possible?<\/p>\n","answers":[{"AnswerId":"57983423","CreationDate":"2019-09-18T00:13:35.413","ParentId":null,"OwnerUserId":"7207392","Title":null,"Body":"<p>Disclaimer: I didn't test this in any serious way, but it seems to be a promising route.<\/p>\n\n<p>A comparatively painless way of manipulating ufunc behavior seems to be <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.classes.html\" rel=\"nofollow noreferrer\">subclassing <code>ndarray<\/code><\/a> and overriding <code>__array_ufunc__<\/code>. For example, if you are content with catching anything that produces <code>float64<\/code><\/p>\n\n<pre><code>class no64(np.ndarray):\n    def __array_ufunc__(self, ufunc, method, *inputs, **kwds):\n        ret = getattr(ufunc, method)(*map(np.asarray,inputs), **kwds)\n        # some ufuncs return multiple arrays:\n        if isinstance(ret,tuple):\n            if any(x.dtype == np.float64 for x in ret):\n                raise ValueError\n            return (*(x.view(no64) for x in ret),)\n        if ret.dtype == np.float64:\n             raise ValueError\n        return ret.view(no64)\n\nx = np.arange(6,dtype=np.float32).view(no64)\n<\/code><\/pre>\n\n<p>Now let's see what our class can do:<\/p>\n\n<pre><code>x*x\n# no64([ 0.,  1.,  4.,  9., 16., 25.], dtype=float32)\nnp.sin(x)\n# no64([ 0.        ,  0.84147096,  0.9092974 ,  0.14112   , -0.7568025 ,\n#       -0.9589243 ], dtype=float32)\nnp.frexp(x)\n# (no64([0.   , 0.5  , 0.5  , 0.75 , 0.5  , 0.625], dtype=float32), no64([0, 1, 2, 2, 3, 3], dtype=int32))\n<\/code><\/pre>\n\n<p>Now let's pair it with a 64 bit argument:<\/p>\n\n<pre><code>x + np.arange(6)\n# Traceback (most recent call last):\n#   File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n#   File \"&lt;stdin&gt;\", line 9, in __array_ufunc__\n# ValueError\n\nnp.multiply.outer(x, np.arange(2.0))\n# Traceback (most recent call last):\n#   File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n#   File \"&lt;stdin&gt;\", line 9, in __array_ufunc__\n# ValueError\n<\/code><\/pre>\n\n<p>What doesn't work (I'm sure there is much more)<\/p>\n\n<pre><code>np.outer(x, np.arange(2.0))  # not a ufunc, so slips through\n# array([[0., 0.],\n#        [0., 1.],\n#        [0., 2.],\n#        [0., 3.],\n#        [0., 4.],\n#        [0., 5.]])\n<\/code><\/pre>\n\n<p><code>__array_function__<\/code> seems to be what catches those.<\/p>\n"}]}
+{"QuestionId":57982376,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-09-17T21:36:06.103","AcceptedAnswerId":null,"OwnerUserId":8840416.0,"Title":"Pytorch tensor multiplication with Float tensor giving wrong answer","Body":"<p>I am seeing some strange behavior when i multiply two pytorch tensors.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = torch.tensor([99397544.0])\ny = torch.tensor([0.1])\nx * y\n<\/code><\/pre>\n\n<p>This outputs <\/p>\n\n<pre><code>tensor([9939755.])\n<\/code><\/pre>\n\n<p>However, the answer should be <code>9939754.4<\/code><\/p>\n","answers":[{"AnswerId":"57982766","CreationDate":"2019-09-17T22:25:01.163","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>In default, the tensor dtype is <code>torch.float32<\/code> in pytorch. Change it to <code>torch.float64<\/code> will give the right result. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = torch.tensor([99397544.0], dtype=torch.float64)\ny = torch.tensor([0.1], dtype=torch.float64)\nx * y\n# tensor([9939754.4000])\n<\/code><\/pre>\n\n<p>The mismatched result for <code>torch.float32<\/code> caused by rounding error if you do not have enough precision to calculate (represent) it. <\/p>\n\n<p><a href=\"https:\/\/docs.oracle.com\/cd\/E19957-01\/806-3568\/ncg_goldberg.html\" rel=\"nofollow noreferrer\">What Every Computer Scientist Should Know About Floating-Point Arithmetic<\/a><\/p>\n"}]}
+{"QuestionId":57983098,"AnswerCount":0,"Tags":"<tensorflow><distributed-computing>","CreationDate":"2019-09-17T23:17:11.107","AcceptedAnswerId":null,"OwnerUserId":6450489.0,"Title":"Noisy loss in distributed training in tensorflow","Body":"<p>I am working on distributed learning in tensorflow through estimators API, and I always get below training loss graphs in tensorboard when comparing one-machine configuration (one ps and one chief on the same machine) and two-machine configuration (one ps and one chief on machine#1, one worker on machine#2).<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/8jIwA.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/8jIwA.jpg\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/9UBWy.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9UBWy.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Also when I download the graph data, I see that some steps are lost in two-machine distributed training case.\nMy model is defined in keras and I use the standard estimator functions for distributed training using dataset API in the following format (read_dataset is a custom function with generates training or evaluation dataset):<\/p>\n\n<pre><code>runConfig = tf.estimator.RunConfig(\n        session_config=config,\n        model_dir=log_dir,\n        save_summary_steps=1,\n        save_checkpoints_steps=train_steps\n        )\nestimator = tf.keras.estimator.model_to_estimator(model, model_dir=log_dir, config=runConfig)\ntrain_spec = tf.estimator.TrainSpec(input_fn=lambda: read_dataset(...), max_steps=...)\neval_spec = tf.estimator.EvalSpec(input_fn=lambda: read_dataset(...), start_delay_secs=1, throttle_secs=1, steps=None)\ntf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)\n<\/code><\/pre>\n\n<p>What can be the cause for this noisy behavior and missing logs?<\/p>\n\n<p>And another question: I assume that the above code implements data parallelism, in which input dataset is splitted between different machines and so each sample is used once during training but possibly by a different machine. Is it true? I am using Tensorflow 1.12 and 1.13.<\/p>\n","answers":[]}
+{"QuestionId":57983533,"AnswerCount":1,"Tags":"<python><tensorflow><google-cloud-platform><google-cloud-functions><tensorflow2.0>","CreationDate":"2019-09-18T00:33:11.560","AcceptedAnswerId":null,"OwnerUserId":7679261.0,"Title":"Google Cloud Function: \"Error: memory limit exceeded. Function invocation was interrupted.\" when deploying NLP model","Body":"<p>I am trying to deploy an NLP model which has been trained in Tesorflow 2.0 beta.  I use <code>model.save('pb_model')<\/code> to save model in .pb format (this gives me pb_model.pb and variables folders as usual).  <\/p>\n\n<p>I placed pb files in Google Cloud storage bucket which all together are about 370Mb.  I am using Google Cloud Functions since this is currently the only serverless platform which supports TF 2.0.  For my GCP Function, I selected the highest memory allocation available which is currently 2Gb.  I trigger my Function with an HTTP request.<\/p>\n\n<p>My <code>main.py<\/code> code looks like this:<\/p>\n\n<pre><code>import sys\nimport os\nimport psutil\nimport shutil\nimport tensorflow as tf\n\nimport googleapiclient.discovery\nfrom google.cloud import storage\n\nmodel = None\n\nclass CustomModel():\n    def __init__(self):\n        print('Starting CustomModel build...')\n        print('1:', psutil.virtual_memory())\n        dirName = '\/tmp\/model_dir\/variables\/'\n\n        # Create target directory &amp; all intermediate directories if don't exists\n        os.makedirs(dirName, exist_ok=True)\n\n        pb_model_dir = 'gs:\/\/&lt;cloud_bucket_name&gt;\/model_dir'  # storage bucket file path of pb files\n        model = tf.saved_model.load(pb_model_dir)\n        shutil.rmtree('\/tmp\/model_dir', ignore_errors=True) # remove tmp files to maintain low memory\n        print('loaded model')\n        return model\n\ndef handler(request):\n    global model\n\n    # Model load which only happens during cold starts\n    if model is None:\n        model = CustomModel()\n\n    model_inputs = &lt;some input text&gt;\n    predictions = model(model_inputs)\n    print(predictions)\n\n    return predictions\n<\/code><\/pre>\n\n<p>When looking at the logs, the output of print statement <code>print('1:', psutil.virtual_memory())<\/code> shows that I have 1.9Gb free out of the allocated 2Gb, before the model is loaded.  <\/p>\n\n<p>When I run this exact code on Google Colab, the maximum amount of memory which is occupied by the loaded model is about 0.9Gb.  Therefore, in principle, that means that when I deploy model in GCP Function, 2Gb is more than enough for my model however I keep getting the memory limit exceeded error: <\/p>\n\n<blockquote>\n  <p>Error: memory limit exceeded. Function invocation was interrupted.<\/p>\n<\/blockquote>\n\n<p>Does anyone know why I am getting this error?  Am I constructing GCP Function correctly?  I'm still kinda of new to GCP.  Btw, I even tried to import pb model files from bucket into my Function \/tmp\/ folder instead of pointing to bucket path (<code>pb_model_dir = 'gs:\/\/&lt;cloud_bucket_name&gt;\/model_dir'<\/code>) but that also gives me memory limit exceeded error. \n Not sure what else to try... Thank you in advance for your help!<\/p>\n","answers":[{"AnswerId":"57987654","CreationDate":"2019-09-18T07:51:39.667","ParentId":null,"OwnerUserId":"11928088","Title":null,"Body":"<p>There could be more reasons for this situation to be encountered. As I can see that you are already aware, temporary files can be a cause of a memory leak. Temporary files can also be created from errors, or warnings. First of all, I would check my function logs for errors and warnings and I would try to fix them.<\/p>\n\n<p>As I see your Cloud Function is interacting a lot with the storage bucket. This causes high queries per second which might be a possible reason for the Cloud Function instance to fail, because of the GCS errors. This issue was encountered before and as you can see <a href=\"https:\/\/github.com\/googleapis\/google-cloud-go\/issues\/1253\" rel=\"nofollow noreferrer\">here<\/a>, it was mitigated by treating the GCS errors. So, I would suggest to investigate more following this path, go where your function interacts with the bucket and try to treat the possible errors coming from the bucket.<\/p>\n\n<p>Another thing that you can do is investigate a bit the way of the memory is used in your function. For this you have some nice tools you can use like <a href=\"http:\/\/guppy-pe.sourceforge.net\/#Heapy\" rel=\"nofollow noreferrer\">Heap Checker<\/a> and  <a href=\"https:\/\/pythonhosted.org\/Pympler\/\" rel=\"nofollow noreferrer\">Pympler<\/a>.<\/p>\n\n<p>If these is not the issue, I would ask you to show some logs(from CF and Storage Bucket) and some tests for the function, because on a first look, your code looks fine to me.<\/p>\n\n<p>You may also take into consideration using another method to achieve your technical purpose. You have <a href=\"https:\/\/cloud.google.com\/blog\/products\/ai-machine-learning\/how-to-serve-deep-learning-models-using-tensorflow-2-0-with-cloud-functions\" rel=\"nofollow noreferrer\">here<\/a> more information than can present you all the options you have.<\/p>\n"}]}
+{"QuestionId":57984056,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-serving>","CreationDate":"2019-09-18T02:01:48.213","AcceptedAnswerId":null,"OwnerUserId":10981539.0,"Title":"How to have retrained Inception model running on Tensorflow-Serving accept images via REST API","Body":"<p><strong>Summary:<\/strong><\/p>\n\n<p>I have a working retrained version of the Inception model and I am trying to deploy it with Tensorflow-Serving and use Tensorflow-Serving's REST API. However, when using the REST API, I can only send the input image as a base64-encoded string. The problem is that this appears not to be a valid input to the model and I cannot figure out how to correctly configure the model to accept the base64-encoded string and do the preprocessing necessary to transform it to a valid input.<\/p>\n\n<p><strong>Background:<\/strong><\/p>\n\n<p>I have a retrained Inception model produced by following the instructions here:\n<a href=\"https:\/\/www.tensorflow.org\/hub\/tutorials\/image_retraining\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/hub\/tutorials\/image_retraining<\/a><\/p>\n\n<p>More specifically, I create the saved_model.py by running this script with appropriate arguments:\n<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 am able to successfully run inferences locally using the example label_image script (which appears to do the necessary image preprocessing):\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/label_image\/label_image.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/label_image\/label_image.py<\/a><\/p>\n\n<p><strong>Problem:<\/strong><\/p>\n\n<p>I send the image in a json document like so:<\/p>\n\n<pre><code>{\n  \"instances\": [\n      {\"image\": [{\"b64\": base64Thumbnail}]}\n  ]\n}\n<\/code><\/pre>\n\n<p>to this address:<\/p>\n\n<pre><code>localhost:8501\/v1\/models\/inception_retrain\/:predict\n<\/code><\/pre>\n\n<p>Tensorflow-Serving sends back a response with this error in the body (I've truncated the base64 string here to save space):<\/p>\n\n<pre><code>\"error\": \"Failed to process element: 0 key: image of \\'instances\\' list.\n\nError: Invalid argument: JSON Value: {\\n    \\\"b64\\\": \\\"\/9j\/4AAQSk...\\\"\\n} \n\nType: Object is not of expected type: float\"\n<\/code><\/pre>\n\n<p>Which makes sense, as when I check the metadata of my model through Tensorflow-Serving, I get this description of the model's input, indicating it expects a float:<\/p>\n\n<pre><code>\"inputs\": {\n    \"image\": {\n        \"dtype\": \"DT_FLOAT\",\n        \"tensor_shape\": {\n            \"dim\": [\n                {\n                    \"size\": \"-1\",\n                    \"name\": \"\"\n                 },\n                 {\n                    \"size\": \"299\",\n                    \"name\": \"\"\n                 },\n                 {\n                    \"size\": \"299\",\n                    \"name\": \"\"\n                 },\n                 {\n                    \"size\": \"3\",\n                    \"name\": \"\"\n                 }\n              ],\n            \"unknown_rank\": false\n        },\n        \"name\": \"Placeholder:0\"\n    }\n}\n<\/code><\/pre>\n\n<p>What do I need to do so that the model will accept and analyze the base64-encoded image? <\/p>\n\n<p>In most examples, they just some gRPC-related client code (which seems simpler), but that option is not available to me, as my client is written in javascript\/node and I can only send the image via Tensorflow-Serving's REST API.<\/p>\n","answers":[]}
+{"QuestionId":57985220,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2019-09-18T04:43:25.387","AcceptedAnswerId":58002941.0,"OwnerUserId":10520937.0,"Title":"Converting a singular variable regression to multi-variable regression in Tensorflow","Body":"<p>Note: I didn't write the code, here is the code I'm trying to modify <a href=\"https:\/\/gist.github.com\/balzer82\/b5cac8fb214e0b52efa0ea7b8a058681\" rel=\"nofollow noreferrer\">this<\/a> and \nmy repository is <a href=\"https:\/\/github.com\/itisyeetimetoday\/reggression\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>I'm currently trying to convert a singular variable regression to a multi-variable DNN regression. Line 76 in <code>regressor_full.py<\/code> always returns an error about incompatible shape. I've changed the input to the shape of 11, however, I don't know to change the output layer to 11 too.<\/p>\n\n<pre><code>Line 76 : feature_columns = [tf.feature_column.numeric_column('X', shape=(11,))]\n<\/code><\/pre>\n\n<p>However, I'm sure that the input and output tensors aren't the only 2 things that have to be changed. Could you guys help me with adapting the repo? Thank you<\/p>\n","answers":[{"AnswerId":"58002941","CreationDate":"2019-09-19T02:34:31.740","ParentId":null,"OwnerUserId":"5136550","Title":null,"Body":"<p>To adapt the regressor for 11 targets\/outputs, change the value of label_dimension to 11 as follows [Line #97 in your code]:<\/p>\n\n<pre><code>regressor = skflow.DNNRegressor(feature_columns=feature_columns,\n                label_dimension=11,\n                hidden_units=hidden_layers,\n                model_dir=MODEL_PATH,\n                dropout=dropout,\n                config=test_config)\n<\/code><\/pre>\n\n<p>Successful training log:\n<a href=\"https:\/\/i.stack.imgur.com\/AOE8Q.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/AOE8Q.jpg\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":57985297,"AnswerCount":0,"Tags":"<python><tensorflow><nvidia>","CreationDate":"2019-09-18T04:52:42.583","AcceptedAnswerId":null,"OwnerUserId":268645.0,"Title":"Two python application if run simultaneously on windows 7 crashes after sometime","Body":"<p>We have created Two python (3.6) application one for object detection in video using tensorflow and another one GUI based on pyside2. <\/p>\n\n<p>Both applications are running fine with out crashing if run one application at a time. <\/p>\n\n<p>But if run simultaneously on windows 7, after some time (some time after one hour or two) crashes with Ntdll.dll error code c0000374.<\/p>\n\n<p>We are using NVidia graphics card, cuda, numpy etc<\/p>\n\n<p>New to python &amp; tensorflow<\/p>\n","answers":[]}
+{"QuestionId":57985406,"AnswerCount":2,"Tags":"<python><tensorflow><keras><oraclelinux>","CreationDate":"2019-09-18T05:05:16.563","AcceptedAnswerId":57985905.0,"OwnerUserId":11332292.0,"Title":"Cannot import name 'tf_utils' when using importing keras","Body":"<p>I'm using Oracle Linux 7.7, and I installed python3.6 using yum (epel repos). Then I install tensorflow 1.5(since if it goes newer ver I got core dumped) and keras. If I'm importing tensorflow, I got nothing.\nBut when I import keras, I got <\/p>\n\n<pre><code>ImportError: cannot import name 'tf_utils'\n<\/code><\/pre>\n\n<p>Here's the full output:<\/p>\n\n<pre><code>$ python\nPython 3.6.8 (default, Aug  7 2019, 08:02:28) \n[GCC 4.8.5 20150623 (Red Hat 4.8.5-39.0.1)] on linux\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n&gt;&gt;&gt; import tensorflow\n&gt;&gt;&gt; import keras\nUsing TensorFlow backend.\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/home\/reyhan\/project\/.virtualenvs\/keras_tf\/lib\/python3.6\/site-packages\/keras\/__init__.py\", line 3, in &lt;module&gt;\nfrom . import utils\n  File \"\/home\/reyhan\/project\/.virtualenvs\/keras_tf\/lib\/python3.6\/site-packages\/keras\/utils\/__init__.py\", line 6, in &lt;module&gt;\nfrom . import conv_utils\n  File \"\/home\/reyhan\/project\/.virtualenvs\/keras_tf\/lib\/python3.6\/site-packages\/keras\/utils\/conv_utils.py\", line 9, in &lt;module&gt;\nfrom .. import backend as K\n  File \"\/home\/reyhan\/project\/.virtualenvs\/keras_tf\/lib\/python3.6\/site-packages\/keras\/backend\/__init__.py\", line 1, in &lt;module&gt;\nfrom .load_backend import epsilon\n  File \"\/home\/reyhan\/project\/.virtualenvs\/keras_tf\/lib\/python3.6\/site-packages\/keras\/backend\/load_backend.py\", line 90, in &lt;module&gt;\nfrom .tensorflow_backend import *\n  File \"\/home\/reyhan\/project\/.virtualenvs\/keras_tf\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 13, in &lt;module&gt;\nfrom tensorflow.python.keras.utils import tf_utils\nImportError: cannot import name 'tf_utils'\n<\/code><\/pre>\n\n<p>I was using python 3.6 by building it from source before and keras worked fine but since I can't install tkinter for pyplot I uninstall it and using the one from yum instead.<\/p>\n","answers":[{"AnswerId":"58521094","CreationDate":"2019-10-23T10:45:03.743","ParentId":null,"OwnerUserId":"12262626","Title":null,"Body":"<p>I had the same problem, but upgraded Tensorflow via pip rather than downgrading Keras and this solved the problem :)<\/p>\n\n<p>Using Python 3.6.4, Keras 2.3.1, Tensorflow 2.0<\/p>\n"},{"AnswerId":"57985905","CreationDate":"2019-09-18T05:54:12.983","ParentId":null,"OwnerUserId":"11332292","Title":null,"Body":"<p>Seems like it was a problem with keras 2.3.0, I installed keras 2.1.5 using pip and it works fine. <\/p>\n"}]}
+{"QuestionId":57985690,"AnswerCount":1,"Tags":"<tensorflow><keras><machine-learning-model>","CreationDate":"2019-09-18T05:32:23.670","AcceptedAnswerId":null,"OwnerUserId":12083017.0,"Title":"How to convert tensorflow model to keras model? .pb file to .hdf5?","Body":"<p>I have searched for hours on google but have not seen any code\/package to convert.<br>\nAny guidance on this would be highly appreciated.<\/p>\n","answers":[{"AnswerId":"57985896","CreationDate":"2019-09-18T05:53:16.350","ParentId":null,"OwnerUserId":"5136550","Title":null,"Body":"<p>First load .pb model and then save as keras .h5 model.<\/p>\n\n<pre><code>best_model_file_pb = MODEL_DIR + \"best_model.pb\"\nbest_model_file_h5 = MODEL_DIR + \"best_model_.h5\"\nloaded_model = tf.keras.models.load_model(best_model_file_pb)\ntf.keras.models.save_keras_model(loaded_model, best_model_file_h5)\nloaded_model_from_h5 = tf.keras.models.load_model(best_model_file_h5)\n<\/code><\/pre>\n\n<p>This converted model will work in the tensorflow ecosystem, using tf.keras. However, if you want tf.keras model to work in keras (with tensorflow backend), then this model won't work.<\/p>\n"}]}
+{"QuestionId":57986005,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><matplotlib><undefined-variable>","CreationDate":"2019-09-18T06:03:38.497","AcceptedAnswerId":57995737.0,"OwnerUserId":554517.0,"Title":"matplotlib.cm.coolwarm (used in Colaboratory TensorFlow tutorial from Google) is an undefined variable in Eclipse","Body":"<p>I was trying to take code I found on a Google Colaboratory tensorflow tutorial (in a Jupyter workbook) and implement it in an Eclipse enviroment on my own computer: \n(<a href=\"https:\/\/colab.research.google.com\/notebooks\/mlcc\/first_steps_with_tensor_flow.ipynb?utm_source=mlcc&amp;utm_campaign=colab-external&amp;utm_medium=referral&amp;utm_content=firststeps-colab&amp;hl=en#scrollTo=wgSMeD5UU81N\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/notebooks\/mlcc\/first_steps_with_tensor_flow.ipynb?utm_source=mlcc&amp;utm_campaign=colab-external&amp;utm_medium=referral&amp;utm_content=firststeps-colab&amp;hl=en#scrollTo=wgSMeD5UU81N<\/a>)\nThis code executes fine on the Colaboratory jupyter workbook.\nHowever Eclipse gives me get an \"undefined variable\" error from a matlablib import.  <\/p>\n\n<p>my system:<br>\nMac OS: 10.13.6<br>\neclipse: 019-03 (4.11.0)<br>\nAnaconda installation including: python3.7.3, matplotlib 3.1.0. \n(Also the shasum check for the Anaconda download was correct before installing Anaconda packages)<br>\nI also installed tensorflow per googles website: tensorflow 1.14.0  <\/p>\n\n<p>Using Eclipse and python3.7<br>\nThis line generates no errors:  <\/p>\n\n<pre><code>from matplotlib import cm\"   #(&lt;--no error generated here)\n<\/code><\/pre>\n\n<p>However when I go to use \"cm.coolwarm\" later, Eclipse gives me an \"undefined variable from import\" error related to \"coolwarm\":  <\/p>\n\n<pre><code>colors = [cm.coolwarm(x) for x in np.linspace(-1, 1, periods)] \n<\/code><\/pre>\n\n<p>When I look at the cm.py file in my downloaded Matplotlib package, there is in fact no \"coolwarm\" variable\/method\/function defined. Going to the Matplotlib website, there also doesn't appear to be any coolwarm defined in version 3.1.1 (the latest version), or older stable 3.0.1.  <\/p>\n\n<p>My conclusion is that the Google Colabaratory jupyter workbook must be using a very old version of matplotlib and that cm.coolwarm is deprecated. But I am a relative newbie, and just thought I would check to see if there was something I am missing...  <\/p>\n\n<p>Thanks for your thoughts.<\/p>\n","answers":[{"AnswerId":"57995737","CreationDate":"2019-09-18T14:58:37.990","ParentId":null,"OwnerUserId":"4124317","Title":null,"Body":"<p>The coolwarm colormap is not deprecated in any newer version of matplotlib. As far as I can tell there is also no plan to do so in the future.<\/p>\n\n<p>The following works fine with any version of matplotlib<\/p>\n\n<pre><code>from matplotlib import cm\nprint(cm.coolwarm(0.5))\n<\/code><\/pre>\n\n<p>The reason you did not find <code>cm.coolwarm<\/code> in the source code is that those names are generated on the fly. All internal colormaps, which you can get via <code>plt.get_cmap(\"name_of_colormap\")<\/code>, are also made available as objects in the <code>cm<\/code> module as <code>cm.name_of_colormap<\/code> via <a href=\"https:\/\/github.com\/matplotlib\/matplotlib\/blob\/4b27b46d14d18ec00c7acfd46c397e53a14374d6\/lib\/matplotlib\/cm.py#L73\" rel=\"nofollow noreferrer\">this line<\/a><\/p>\n\n<pre><code>locals().update(cmap_d)\n<\/code><\/pre>\n\n<hr>\n\n<p>That being said I have no idea as to why you get the error in Eclipse.  <\/p>\n"}]}
+{"QuestionId":57986019,"AnswerCount":0,"Tags":"<keras><onnx>","CreationDate":"2019-09-18T06:05:11.520","AcceptedAnswerId":null,"OwnerUserId":12083128.0,"Title":"ONNX to Keras Conversion","Body":"<p>I am using the following link for my implementation. <a href=\"https:\/\/github.com\/nerox8664\/onnx2keras\/\" rel=\"nofollow noreferrer\">https:\/\/github.com\/nerox8664\/onnx2keras\/<\/a><\/p>\n\n<p>when trying to execute the code i am getting following error.<\/p>\n\n<p>k_model = onnx_to_keras(onnx_model, ['input']) <\/p>\n\n<hr>\n\n<p>Traceback (most recent call last):\n  File \"import_convt2keras.py\", line 8, in \n    k_model = onnx_to_keras(onnx_model, ['input'])\n  File \"\/home\/ubuntu\/anaconda3\/envs\/tensorflow_p36\/lib\/python3.6\/site-packages\/onnx2keras\/converter.py\", line 80, in onnx_to_keras\n    weights[onnx_extracted_weights_name] = numpy_helper.to_array(onnx_w)<\/p>\n\n<h2>TypeError: unhashable type: 'google.protobuf.pyext._message.RepeatedScalarContainer'<\/h2>\n","answers":[]}
+{"QuestionId":57986086,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-18T06:10:38.753","AcceptedAnswerId":57994784.0,"OwnerUserId":9839954.0,"Title":"tensorflow 1.13 how to use tf.searchsort safe?","Body":"<p>With code <\/p>\n\n<pre><code>tf.searchsorted(input, input2)\n<\/code><\/pre>\n\n<p>I got <strong>first error<\/strong><\/p>\n\n<blockquote>\n  <p>InvalidArgumentError (see above for traceback): Reshape cannot infer\n  the missing input size for an empty tensor unless all specified input\n  sizes are non-zero<\/p>\n<\/blockquote>\n\n<p>It also reminded me in line 3459 of <\/p>\n\n<blockquote>\n  <p>tensorflow\/python\/ops\/array_ops.py<\/p>\n  \n  <p>searchsorted <\/p>\n<\/blockquote>\n\n<p>got<\/p>\n\n<pre><code>sorted_sequence_2d = reshape(sorted_sequence, [-1, sequence_size])\n<\/code><\/pre>\n\n<p>But when tensor shape contain 0 dimension,  <code>input.shape=(0,)<\/code>, it will return an error. see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/10609\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>I want to check the tensor shape <strong>before<\/strong> using <code>tf.searchsorted<\/code>, and I know <strong>the dimension is<\/strong> <code>None<\/code><\/p>\n\n<p>So I use <\/p>\n\n<pre><code>if not tf.equal(input.shape[0], 0):\n  tf.searchsorted(input, input2)\n<\/code><\/pre>\n\n<p>Then I got my <strong>second error<\/strong>, and I know that <code>tf.equal<\/code> will return bool tensor, which cannot be used like bool. But I don't know how to solve my first error.<\/p>\n\n<blockquote>\n  <p>ValueError: Tried to convert 'x' to a tensor and failed. Error: Cannot\n  convert an unknown Dimension to a Tensor: ?<\/p>\n<\/blockquote>\n\n<p>My question is how to use <code>tf.searchsort<\/code> safe if the first error trigered by <code>0<\/code> dimension<\/p>\n","answers":[{"AnswerId":"57994784","CreationDate":"2019-09-18T14:11:05.697","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Although it is not explicitly stated, it is implied that <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/searchsorted\" rel=\"nofollow noreferrer\"><code>tf.searchsorted<\/code><\/a> does not work with an empty sequence in the first parameter.<\/p>\n\n<p>You can however use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/cond\" rel=\"nofollow noreferrer\"><code>tf.cond<\/code><\/a> to express something like \"use searchsorted if it is not an empty sequence, otherwise return all zeros\" (or whatever you prefer to return instead):<\/p>\n\n<pre><code>tf.cond(tf.not_equal(tf.size(input), 0),\n        lambda: tf.searchsorted(input, input2),\n        lambda: tf.zeros_like(input2, dtype=tf.int32))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57986783,"AnswerCount":1,"Tags":"<deep-learning><pytorch><transformer><seq2seq>","CreationDate":"2019-09-18T06:59:52.050","AcceptedAnswerId":null,"OwnerUserId":11390236.0,"Title":"embedding layer outputs nan","Body":"<p>I am trying to learn a seq2seq model.\nAn embedding layer is located in the encoder and it sometimes outputs nan value after some iterations.\nI cannot identify the reason.\nHow can I solve this??\nThe problem is the first emb_layer in the forward function in the code below.<\/p>\n\n<pre><code>\nclass TransformerEncoder(nn.Module):\n    def __init__(self, vocab_size, hidden_size=1024, num_layers=6, dropout=0.2, input_pad=1, batch_first=False, embedder=None, init_weight=0.1):\n        super(TransformerEncoder, self).__init__()\n        self.input_pad = input_pad\n        self.vocab_size = vocab_size\n        self.num_layers = num_layers\n        self.embedder = embedder\n\n        if embedder is not None:\n            self.emb_layer = embedder\n        else:\n            self.emb_layer = nn.Embedding(vocab_size, hidden_size, padding_idx=1)\n\n        self.positional_encoder = PositionalEncoder()\n        self.transformer_layers = nn.ModuleList()\n        for _ in range(num_layers):\n            self.transformer_layers.append(\n                    TransformerEncoderBlock(num_heads=8, embedding_dim=1024, dropout=dropout))\n\n    def set_mask(self, inputs):\n        self.input_mask = (inputs == self.input_pad).unsqueeze(1)\n\n    def forward(self, inputs):\n        x = self.emb_layer(inputs)\n        x = self.positional_encoder(x)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58018804","CreationDate":"2019-09-19T20:52:22.600","ParentId":null,"OwnerUserId":"12054165","Title":null,"Body":"<p>It looks like some weights become nan. The one of the possible reasons is that on some iteration a layer output is +-inf. If it output is +-inf on forward, on backward it will have a +-inf and as inf - inf = none, the weights will become none, and at all following iterations will output none. <\/p>\n\n<p>You may check this just by tracking inf outputs in emb_layer. <\/p>\n\n<p>If this is the reason, just try to avoid functions that may return inf values.<\/p>\n"}]}
+{"QuestionId":57987351,"AnswerCount":1,"Tags":"<python-2.7><tensorflow><debugging><machine-learning>","CreationDate":"2019-09-18T07:32:23.980","AcceptedAnswerId":57987810.0,"OwnerUserId":3293726.0,"Title":"TensorFlow code reading: what does `sess.run()` do except evaluate?","Body":"<p>I am reading the code of <a href=\"https:\/\/github.com\/kohpangwei\/influence-release\" rel=\"nofollow noreferrer\">influence function<\/a> (ICML 2017 best paper). There is a confusing <a href=\"https:\/\/github.com\/kohpangwei\/influence-release\/blob\/master\/influence\/logisticRegressionWithLBFGS.py#L190\" rel=\"nofollow noreferrer\">line<\/a> (190) which returns nothing <code>self.sess.run(self.set_params_op, feed_dict=params_feed_dict)<\/code>. Relevant snippets as follows:<\/p>\n\n<pre><code>params_feed_dict = {}\nparams_feed_dict[self.W_placeholder] = W\n# params_feed_dict[self.b_placeholder] = b\nself.sess.run(self.set_params_op, feed_dict=params_feed_dict)\nif save_checkpoints: self.saver.save(self.sess, self.checkpoint_file, global_step=0)\n\nif verbose:\n    print('LBFGS training took %s iter.' % model.n_iter_)\n    print('After training with LBFGS: ')\n    self.print_model_eval()\n<\/code><\/pre>\n\n<p>The <code>self.set_params_op<\/code> is assigned by the following function, which is done in init function (whether I did it or not, it will be executed):<\/p>\n\n<pre><code>def set_params(self):\n    self.W_placeholder = tf.placeholder(\n        tf.float32,\n        shape=[self.input_dim * self.num_classes],\n        name='W_placeholder')\n\n    set_weights = tf.assign(self.weights, self.W_placeholder, validate_shape=True)\n    return [set_weights]\n<\/code><\/pre>\n\n<p>Then I tried to comment that line, but the print message of <code>self.print_model_eval()<\/code> changed.<\/p>\n\n<p>The useful <a href=\"https:\/\/github.com\/kohpangwei\/influence-release\/blob\/578bc458b4d7cc39ed7343b9b271a04b60c782b1\/influence\/genericNeuralNet.py#L305\" rel=\"nofollow noreferrer\">snippets<\/a> about <code>print_model_eval()<\/code> defines as follows:<\/p>\n\n<pre><code>def print_model_eval():\n    ...\n    grad_loss_val, loss_no_reg_val, loss_val, train_acc_val = self.sess.run(\n        [self.grad_total_loss_op, self.loss_no_reg, self.total_loss, self.accuracy_op],\n        feed_dict=self.all_train_feed_dict)\n\n    test_loss_val, test_acc_val = self.sess.run(\n        [self.loss_no_reg, self.accuracy_op],\n        feed_dict=self.all_test_feed_dict)\n<\/code><\/pre>\n\n<p>The tensors are obtained by:<\/p>\n\n<pre><code>self.total_loss, self.loss_no_reg, self.indiv_loss_no_reg = self.loss(\n            self.logits, \n            self.labels_placeholder)\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>def loss(self, logits, labels):\n    labels = tf.one_hot(labels, depth=self.num_classes)\n    # correct_prob = tf.reduce_sum(tf.multiply(labels, tf.nn.softmax(logits)), reduction_indices=1)\n    cross_entropy = - tf.reduce_sum(tf.multiply(labels, tf.nn.log_softmax(logits)), reduction_indices=1)\n\n    indiv_loss_no_reg = cross_entropy\n    loss_no_reg = tf.reduce_mean(cross_entropy, name='xentropy_mean')\n    tf.add_to_collection('losses', loss_no_reg)\n\n    total_loss = tf.add_n(tf.get_collection('losses'), name='total_loss')\n\n    return total_loss, loss_no_reg, indiv_loss_no_reg\n<\/code><\/pre>\n\n<p>My questions are:<\/p>\n\n<ol>\n<li>How does <code>sess.run()<\/code> without return value affect the computation of other tensors?<\/li>\n<li>What is the best practice to debug TensorFlow code?<\/li>\n<\/ol>\n\n<p>Any suggestion would be appreciated. Thank you in advance.<\/p>\n","answers":[{"AnswerId":"57987810","CreationDate":"2019-09-18T08:00:46.433","ParentId":null,"OwnerUserId":"7146757","Title":null,"Body":"<p>With TensorFlow in non-eager execution mode, first you create a graph with operations as the nodes. \nThen, within a session, you run the operations. Certain operations, like the initialisation of variables, don't return anything. They just assign values to the variables in the graph.<\/p>\n\n<p>Here, <code>self.set_params_op<\/code> sounds like it could be an operation of this type. It may be setting values for some parameters. If you comment it out, my guess is that these parameters are not being set to the 'post training' values. So when you evaluate the model, you get different results.<\/p>\n"}]}
+{"QuestionId":57987766,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2019-09-18T07:58:04.677","AcceptedAnswerId":57988017.0,"OwnerUserId":11170350.0,"Title":"Keras evaluate give different result than history. Which one is accurate","Body":"<p>I am wokring on the classification of EEG data (time series) and i am getting different accuracy using evaluate function in keras and from history of iterations in keras.\nI want to know which one is more accurate to consider.\n Here is my keras model<\/p>\n\n<pre><code>model=Sequential()\nmodel.add(Conv1D(filters=60,kernel_size=5,strides=1,padding='same',input_shape=X_train.shape[1::]))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Conv1D(filters=40,kernel_size=5,strides=1,padding='same'))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Conv1D(filters=20,kernel_size=3,strides=1,padding='same'))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Flatten())\nmodel.add(Dense(50,activation='tanh'))\nmodel.add(Dense(20,activation='tanh'))\nmodel.add(Dense(2,activation='sigmoid'))\noptimizer=optimizers.Adam(lr=0.0001)\nmodel.compile(optimizer=optimizer,loss='binary_crossentropy',metrics=['accuracy'])\nhistory=model.fit(X_train,y_train,batch_size=20,epochs=100,\n          verbose=2,validation_split=0.3)\nprint('accuracy',np.average(np.array(history.history['val_acc'])))\nmodel.evaluate(X_test,y_test)\n<\/code><\/pre>\n\n<p>The accuracy I got by using evaluate function is 0.448. And accuracy I got by using history by averaging the validation accuracy is 0.709.<\/p>\n\n<p>Here is accuracy log of last 20 epochs<\/p>\n\n<pre><code>Epoch 81\/100\n - 2s - loss: 0.3923 - acc: 0.8217 - val_loss: 0.5482 - val_acc: 0.7327\nEpoch 82\/100\n - 2s - loss: 0.3843 - acc: 0.8227 - val_loss: 0.5522 - val_acc: 0.7305\nEpoch 83\/100\n - 2s - loss: 0.3885 - acc: 0.8235 - val_loss: 0.5589 - val_acc: 0.7268\nEpoch 84\/100\n - 2s - loss: 0.4001 - acc: 0.8142 - val_loss: 0.5781 - val_acc: 0.7177\nEpoch 85\/100\n - 2s - loss: 0.3938 - acc: 0.8165 - val_loss: 0.5474 - val_acc: 0.7396\nEpoch 86\/100\n - 2s - loss: 0.3965 - acc: 0.8187 - val_loss: 0.5587 - val_acc: 0.7230\nEpoch 87\/100\n - 2s - loss: 0.3768 - acc: 0.8279 - val_loss: 0.5947 - val_acc: 0.7107\nEpoch 88\/100\n - 2s - loss: 0.3789 - acc: 0.8288 - val_loss: 0.5583 - val_acc: 0.7334\nEpoch 89\/100\n - 2s - loss: 0.3805 - acc: 0.8262 - val_loss: 0.5690 - val_acc: 0.7227\nEpoch 90\/100\n - 2s - loss: 0.3712 - acc: 0.8302 - val_loss: 0.5661 - val_acc: 0.7271\nEpoch 91\/100\n - 2s - loss: 0.3678 - acc: 0.8361 - val_loss: 0.5798 - val_acc: 0.7180\nEpoch 92\/100\n - 2s - loss: 0.3727 - acc: 0.8352 - val_loss: 0.5927 - val_acc: 0.7085\nEpoch 93\/100\n - 2s - loss: 0.3720 - acc: 0.8271 - val_loss: 0.5622 - val_acc: 0.7321\nEpoch 94\/100\n - 2s - loss: 0.3829 - acc: 0.8200 - val_loss: 0.5799 - val_acc: 0.7155\nEpoch 95\/100\n - 2s - loss: 0.3630 - acc: 0.8384 - val_loss: 0.5623 - val_acc: 0.7337\nEpoch 96\/100\n - 2s - loss: 0.3655 - acc: 0.8330 - val_loss: 0.5817 - val_acc: 0.7164\nEpoch 97\/100\n - 2s - loss: 0.3722 - acc: 0.8314 - val_loss: 0.5892 - val_acc: 0.7164\nEpoch 98\/100\n - 2s - loss: 0.3591 - acc: 0.8373 - val_loss: 0.5750 - val_acc: 0.7155\nEpoch 99\/100\n - 2s - loss: 0.3710 - acc: 0.8327 - val_loss: 0.5674 - val_acc: 0.7258\nEpoch 100\/100\n - 2s - loss: 0.3669 - acc: 0.8319 - val_loss: 0.5900 - val_acc: 0.7104\n<\/code><\/pre>\n\n<p>I am bit confused which one is more accurate. Figure shows validation accuracy and train accuracy\n<a href=\"https:\/\/i.stack.imgur.com\/vYrNM.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/vYrNM.png\" alt=\"Validation accuracy and train accuracy\"><\/a><\/p>\n","answers":[{"AnswerId":"57988017","CreationDate":"2019-09-18T08:11:44.230","ParentId":null,"OwnerUserId":"5661194","Title":null,"Body":"<p>Both are correct - you are simply using a different data set each time.<\/p>\n\n<p>You specify<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.fit(X_train,y_train,batch_size=20,epochs=100,\n          verbose=2,validation_split=0.3)\n<\/code><\/pre>\n\n<p>which means that 30% of the training data set is not used for training and set aside for validation (<code>validation_split=0.3<\/code>). This is the validation loss you see in the history.<\/p>\n\n<p>However, during testing, you do this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.evaluate(X_test,y_test)\n<\/code><\/pre>\n\n<p>which uses an entirely new data set <code>X_test<\/code>. Since you are getting significantly lower accuracies on this data set, it must be fundamentally different from your training set in some way, and your model fails to capture this variability.<\/p>\n\n<p>Now, which accuracy reflects the \"true\" performance? Neither. Both. It depends on what <code>X_test<\/code> contains, in what way it is different from <code>X_train<\/code>, and what you want your model to be good at. But usually, you would report the model performance on <em>new<\/em> and <em>unseen<\/em> data, i.e., <code>X_test<\/code>.<\/p>\n"}]}
+{"QuestionId":57988410,"AnswerCount":0,"Tags":"<python><keras><conv-neural-network><backtrace>","CreationDate":"2019-09-18T08:34:19.840","AcceptedAnswerId":null,"OwnerUserId":4012737.0,"Title":"FCNN: Is there a keras utility which can give me the input region of an image from output pixel coordinates","Body":"<p>Lets say input image is of dimension 128*128 and the final output layer is of dimension 10*10.\nI want to know which region of input controls the value on (4,5)th co-ordinate of the output. I'm able to compute it manually. An answer for some network can be a square region of length 20 pixels with (12,15) being the co-ordinate of top left vertex.<\/p>\n\n<p><strong>Manual traceback<\/strong>: We can start tracking the region backwards. For example <code>MaxPooling2D(2)<\/code> layer will double your region. <code>Conv2D(padding ='same')<\/code> output size will shift your region. <code>Conv2D(padding ='valid')<\/code>  will not change your region.<\/p>\n\n<p>Since this is a deterministic and <em>easy<\/em> problem, there should be a function somewhere which takes in a [keras] model and output co-ordinate and returns back the input region bounding box. \nThanks !<\/p>\n","answers":[]}
+{"QuestionId":57988620,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><deep-learning><data-science>","CreationDate":"2019-09-18T08:45:12.020","AcceptedAnswerId":58158765.0,"OwnerUserId":5210052.0,"Title":"How to better preprocess images for a better deep learning result?","Body":"<p>We are experimenting applying convolutional neural network to classify good surfaces and surfaces with defects.<\/p>\n\n<p>The Good and Bad images are mostly like the following:<\/p>\n\n<p>Good ones:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/50fkw.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/50fkw.jpg\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/P591n.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/P591n.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Bad ones:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/S9ZyH.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/S9ZyH.jpg\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/4vCIK.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/4vCIK.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<ol>\n<li>The image is relatively big (Height:800 pixels, width: 500 pixels)<\/li>\n<li>The defect very <strong>local<\/strong> and <strong>small<\/strong> relative to image<\/li>\n<li>The background is very <strong>noisy<\/strong><\/li>\n<li>The deep learning (6 x conv+pooling -> flatten -> dense64-> dense32) result is very bad \n(perhaps due to limited Bad samples and very small defect pattern)<\/li>\n<li>There are other defect patterns like very <strong>subtle<\/strong> scratches, residuals and stains, etc., which is one of the main reasons that we want to use deep learning instead of specific feature engineering.<\/li>\n<li>We can and are willing to accumulate more images of defects.<\/li>\n<\/ol>\n\n<p>So the question are:<\/p>\n\n<ol>\n<li>Is deep learning even an appropriate tool for defect detection like this in practice.<\/li>\n<li>If yes, how can we adapt or pre-process the images to the formats that the deep learning models can really work with. (Could we apply some known filters to make the background much less noisy?)<\/li>\n<li>If no, what are other practical techniques that can be used other than deep models.\nWill things like template matching or anything else actually be a fit for this type of problems?<\/li>\n<\/ol>\n\n<hr>\n\n<p>Update:\nVery good idea to come up with an explicit circular stripes checker.\nIt might be directly used to check where the pattern is disturbed or be used as per-processing step for deep learning.<\/p>\n\n<hr>\n\n<p>Update:\nA more subtle pattern 'scratch'.\nThere is a scratch starting from the bottom of the fan area going up and a little to the right.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Aje2N.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Aje2N.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"58158765","CreationDate":"2019-09-29T20:02:55.277","ParentId":null,"OwnerUserId":"5082048","Title":null,"Body":"<blockquote>\n  <p>Is deep learning even an appropriate tool for defect detection like\n  this in practice.<\/p>\n<\/blockquote>\n\n<p>Deep learning certainly is a possibility that promises to be universal. In general, it should rather be the last resort than the first approach. Downsides include:<\/p>\n\n<ul>\n<li>It is difficult to include prior knowledge.<\/li>\n<li>You therefore need an extreme amount of data to train the classifier for the general case.<\/li>\n<li>If you succeed, the model is opaque. It might depend on subtle properties, which cause it to fail if the manufacturing process is changed in the slightest way and there is no easy way to fix it.<\/li>\n<\/ul>\n\n<blockquote>\n  <p>If yes, how can we adapt or pre-process the images to the formats that\n  the deep learning models can really work with. (Could we apply some\n  known filters to make the background much less noisy?)<\/p>\n<\/blockquote>\n\n<p>Independent of the classifier you eventually decide to use, preprocessing should be optimal. <\/p>\n\n<p><strong>Illumination<\/strong>: The illumination is uneven. I'd suggest to define a region of interested, in which the illumination is bright enough to see something. I'd suggest to calculate the average intensity over many images and use this to normalize the brightness. The result would be an image cropped to the region of interest, where the illumination is homogenous.<\/p>\n\n<p><strong>Circular stripes<\/strong>: In the images you show, as the stripes are circular, their orientation depends on the position in the image. I would suggest to use a transformation, which transforms the region of interest (fraction of a circle) into a trapezoid, where each stripe is horizontal and the length of each stripe is retained. <\/p>\n\n<blockquote>\n  <p>If no, what are other practical techniques that can be used other than\n  deep models. Will things like template matching or anything else\n  actually be a fit for this type of problems?<\/p>\n<\/blockquote>\n\n<p>Rather than identifying defects, you could try identifying the intact structure, which has relatively constant properties. (This would be the circular stripes checker that I have suggested in the comment). Here, one obvious thing to test would be a 2D fourier transformation at each pixel within an image preprocessed as described above. If the stripes are intact, you should see that the frequency of intensity change is much lower in horizontal than in vertical direction. I would just plot these two quantities for many \"good\" and \"bad\" pixels and check, whether that might already allow some classification.<\/p>\n\n<p>If you can preselect possible defects with that method, you could then crop out a small image and subject it to deep learning or whatever other method you want to use. <\/p>\n"}]}
+{"QuestionId":57988817,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-18T08:55:56.877","AcceptedAnswerId":null,"OwnerUserId":7205164.0,"Title":"Keras Multiple Checkpoint","Body":"<p>I am trying to implement different checkpoint in model training using keras. When I use single checkpoint it is working well but when i give multiple checkpoints as a list, it is giving me  an error message. <\/p>\n\n<pre><code>model_1_checkpoint = ModelCheckpoint('model\/model1\/model.{epoch:02d}-{val_loss:.2f}.hdf5', monitor='val_loss', \\\n                                         save_best_only=True, verbose=1)\n    model_1_tensorboard = TensorBoard(log_dir=str('logs\/scalars\/'+datetime.datetime.now().strftime(\"%Y%m%d-%H%M%S\")),\\\n                                      histogram_freq=0, write_graph=True, write_images=True, embeddings_freq=0, \\\n                                      embeddings_layer_names=None, embeddings_metadata=None),\n\n    keras_callbacks_1 = EarlyStopping(monitor='val_mean_absolute_error', patience=20, verbose=1, restore_best_weights=True)\n\n    callback = [model_1_checkpoint, model_1_tensorboard, keras_callbacks_1]\n\n    model_1.fit(x_train_array, y_train_array,\n        batch_size=batch_size,\n        epochs=epochs,\n        shuffle=True,\n        verbose=1, # Change it to 2, if wished to observe execution one line per epoch.  o if not. 1 for progress bar\n        validation_data=(x_validation_array, y_validation_array),\n        callbacks=callback,\n        use_multiprocessing=True,\n        workers=30)\n<\/code><\/pre>\n\n<p>Error Message:<\/p>\n\n<pre><code>    AttributeError                            Traceback (most recent call last)\n    &lt;ipython-input-203-3b2d923f4ffd&gt; in &lt;module&gt;\n          7     callbacks=callback,\n          8     use_multiprocessing=True,\n    ----&gt; 9     workers=30)\n\n    \/app\/anaconda3\/lib\/python3.7\/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, validation_freq, max_queue_size, workers, use_multiprocessing, **kwargs)\n        778           validation_steps=validation_steps,\n        779           validation_freq=validation_freq,\n    --&gt; 780           steps_name='steps_per_epoch')\n        781 \n        782   def evaluate(self,\n\n    \/app\/anaconda3\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/engine\/training_arrays.py in model_iteration(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, validation_freq, mode, validation_in_fit, prepared_feed_values_from_dataset, steps_name, **kwargs)\n        211       samples=num_samples_or_steps,\n        212       verbose=0,  # Handle ProgBarLogger separately in this loop.\n    --&gt; 213       mode=mode)\n        214   # TODO(omalleyt): Handle ProgBar as part of Callbacks once hooks are ready.\n        215   progbar = training_utils.get_progbar(model, count_mode)\n\n    \/app\/anaconda3\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/callbacks.py in configure_callbacks(callbacks, model, do_validation, batch_size, epochs, steps_per_epoch, samples, verbose, count_mode, mode)\n        103   # Set callback model\n        104   callback_model = model._get_callback_model()  # pylint: disable=protected-access\n    --&gt; 105   callback_list.set_model(callback_model)\n        106 \n        107   set_callback_parameters(\n\n    \/app\/anaconda3\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/callbacks.py in set_model(self, model)\n        229     self.model = model\n        230     for callback in self.callbacks:\n    --&gt; 231       callback.set_model(model)\n        232 \n        233   def _call_batch_hook(self, mode, hook, batch, logs=None):\n\n    AttributeError: 'tuple' object has no attribute 'set_model'\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57988897,"AnswerCount":1,"Tags":"<python><windows><pandas><pytorch><torchtext>","CreationDate":"2019-09-18T09:00:02.233","AcceptedAnswerId":58071155.0,"OwnerUserId":5443758.0,"Title":"OverflowError: Python int too large to convert to C long torchtext.datasets.text_classification.DATASETS['AG_NEWS']()","Body":"<p>I have 64 bit windows 10 OS\nI have installed python 3.6.8\nI have installed torch and torchtext using pip.\ntorch version is 1.2.0<\/p>\n\n<p>I am trying to load AG_NEWS dataset using below code:<\/p>\n\n<pre><code>import torch\nimport torchtext\nfrom torchtext.datasets import text_classification\nNGRAMS = 2\nimport os\nif not os.path.isdir('.\/.data'):\n    os.mkdir('.\/.data')\ntrain_dataset, test_dataset = text_classification.DATASETS['AG_NEWS'](root='.\/.data', ngrams=NGRAMS, vocab=None)\n<\/code><\/pre>\n\n<p>On the last statement of above code, I am getting below error:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nOverflowError                             Traceback (most recent call last)\n&lt;ipython-input-1-7e8544fdaaf6&gt; in &lt;module&gt;\n      6 if not os.path.isdir('.\/.data'):\n      7     os.mkdir('.\/.data')\n----&gt; 8 train_dataset, test_dataset = text_classification.DATASETS['AG_NEWS'](root='.\/.data', ngrams=NGRAMS, vocab=None)\n      9 # BATCH_SIZE = 16\n     10 # device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n\nc:\\users\\pramodp\\appdata\\local\\programs\\python\\python36\\lib\\site-packages\\torchtext\\datasets\\text_classification.py in AG_NEWS(*args, **kwargs)\n    168     \"\"\"\n    169 \n--&gt; 170     return _setup_datasets(*((\"AG_NEWS\",) + args), **kwargs)\n    171 \n    172 \n\nc:\\users\\pramodp\\appdata\\local\\programs\\python\\python36\\lib\\site-packages\\torchtext\\datasets\\text_classification.py in _setup_datasets(dataset_name, root, ngrams, vocab, include_unk)\n    126     if vocab is None:\n    127         logging.info('Building Vocab based on {}'.format(train_csv_path))\n--&gt; 128         vocab = build_vocab_from_iterator(_csv_iterator(train_csv_path, ngrams))\n    129     else:\n    130         if not isinstance(vocab, Vocab):\n\nc:\\users\\pramodp\\appdata\\local\\programs\\python\\python36\\lib\\site-packages\\torchtext\\vocab.py in build_vocab_from_iterator(iterator)\n    555     counter = Counter()\n    556     with tqdm(unit_scale=0, unit='lines') as t:\n--&gt; 557         for tokens in iterator:\n    558             counter.update(tokens)\n    559             t.update(1)\n\nc:\\users\\pramodp\\appdata\\local\\programs\\python\\python36\\lib\\site-packages\\torchtext\\datasets\\text_classification.py in _csv_iterator(data_path, ngrams, yield_cls)\n     33     with io.open(data_path, encoding=\"utf8\") as f:\n     34         reader = unicode_csv_reader(f)\n---&gt; 35         for row in reader:\n     36             tokens = ' '.join(row[1:])\n     37             tokens = tokenizer(tokens)\n\nc:\\users\\pramodp\\appdata\\local\\programs\\python\\python36\\lib\\site-packages\\torchtext\\utils.py in unicode_csv_reader(unicode_csv_data, **kwargs)\n    128             maxInt = int(maxInt \/ 10)\n    129 \n--&gt; 130     csv.field_size_limit(sys.maxsize)\n    131 \n    132     if six.PY2:\n\nOverflowError: Python int too large to convert to C long\n<\/code><\/pre>\n\n<p>I think the issue is with either windows os or torchtext because I am getting same error for below code as well.<\/p>\n\n<pre><code>pos = data.TabularDataset( path='data\/pos\/pos_wsj_train.tsv', format='tsv', fields=[('text', data.Field()),\n        ('labels', data.Field())])\n<\/code><\/pre>\n\n<p>Can somebody please help? and mainly I don't have any large numerical values in the file.<\/p>\n","answers":[{"AnswerId":"58071155","CreationDate":"2019-09-23T23:04:07.050","ParentId":null,"OwnerUserId":"12110116","Title":null,"Body":"<p>I also encountered a similar problem. I changed a line of code in my torchtext\\utils.py file and my error disappeared.<\/p>\n\n<p>csv.field_size_limit(sys.maxsize)  -- Changed this<br>\ncsv.field_size_limit(maxInt)       -- To this<\/p>\n\n<p>Hope this helps.<\/p>\n"}]}
+{"QuestionId":57989716,"AnswerCount":1,"Tags":"<python><numpy><serialization><pytorch>","CreationDate":"2019-09-18T09:41:12.223","AcceptedAnswerId":null,"OwnerUserId":9687599.0,"Title":"Loading .npy files as dataset for pytorch","Body":"<p>I have preprocessed data in .npy files, let's call it X.npy for raw data and Y.npy for labels. They're organized to match every element from both files (first element from X has first label from Y etc.). How can I load it as dataset using <code>torch.utils.data.DataLoader<\/code>? I'm very new to pytorch, and any help will be useful.<\/p>\n","answers":[{"AnswerId":"57989934","CreationDate":"2019-09-18T09:52:39.110","ParentId":null,"OwnerUserId":"8289159","Title":null,"Body":"<p>You could also use DatasetFolder, which basically is the underlying class of ImageFolder. Using this class you can provide your own file extensions and loader to load the samples.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def npy_loader(path):\n    return torch.from_numpy(np.load(path))\n<\/code><\/pre>\n"}]}
+{"QuestionId":57989945,"AnswerCount":0,"Tags":"<tensorflow><tf.keras>","CreationDate":"2019-09-18T09:53:24.217","AcceptedAnswerId":null,"OwnerUserId":5048010.0,"Title":"Behavior of Dense layers in tf.keras","Body":"<p>I wonder about the behavior of a tensorflow.python.keras.layers.Dense layer (simply Dense from now on). In the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/Dense\" rel=\"nofollow noreferrer\">documentation<\/a> one can read <\/p>\n\n<blockquote>\n  <p>Note: If the input to the layer has a rank greater than 2, then it is\n  flattened prior to the initial dot product with kernel.<\/p>\n<\/blockquote>\n\n<p>Therefore, I expect the output of a Dense layer to always have two dimensions, one given by the batch size and the other by the number of units in the layer. However, I'm not able to replicate this: I always obtain the same number of dimensions in input and in output. In practice, Dense behaves exactly as I would expect TimeDistributed(Dense) to behave (the only difference being that TimeDistributed(Dense) raises an error when applied to 2D inputs).<\/p>\n\n<p>Am I not understanding the doc correctly? <\/p>\n\n<hr>\n\n<p>A minimal working example to show this:<\/p>\n\n<pre><code>from tensorflow.python.keras import Input, Model\nfrom tensorflow.python.keras.layers import Dense, TimeDistributed\n\nshape = (80, 32, 14, 2)  # this doesn't include the batch size, change at will but keep len(shape) &gt;= 2\n\nx = Input(shape=shape)\n\ny1 = TimeDistributed(Dense(3))(x)\nmodel1 = Model(inputs=x, outputs=y1)\nmodel1.summary()\n\ny2 = Dense(3)(x)\nmodel2 = Model(inputs=x, outputs=y2)\nmodel2.summary()\n<\/code><\/pre>\n\n<p>The summaries show that the models have the same output shape and the same number of trainable parameters.<\/p>\n","answers":[]}
+{"QuestionId":57990046,"AnswerCount":0,"Tags":"<python><tensorflow><object-detection><object-detection-api>","CreationDate":"2019-09-18T09:59:01.693","AcceptedAnswerId":null,"OwnerUserId":12084288.0,"Title":"Tensorflow object detection: Box data error while running eval.py","Body":"<p><strong>ValueError: Invalid box data. data must be a numpy array of N[y_min, x_min, y_max, x_max]<\/strong><\/p>\n\n<p>I am genetaring tfrecord from csv file by using <a href=\"https:\/\/github.com\/datitran\/raccoon_dataset\/blob\/master\/generate_tfrecord.py\" rel=\"nofollow noreferrer\">extraneous code<\/a> ( google says it is the only way to do this), here is part with correct order of coordinates:<\/p>\n\n<pre><code>for index, row in group.object.iterrows():\n    ymins.append(row['ymin'] \/ width)\n    xmins.append(row['xmin'] \/ width)\n    ymaxs.append(row['ymax'] \/ height)\n    xmaxs.append(row['xmax'] \/ height)\n    classes_text.append(row['class'].encode('utf8'))\n    classes.append(class_text_to_int(row['class']))\n\ntf_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\/ymin': dataset_util.float_list_feature(ymins),\n    'image\/object\/bbox\/xmin': dataset_util.float_list_feature(xmins),\n    'image\/object\/bbox\/ymax': dataset_util.float_list_feature(ymaxs),\n    'image\/object\/bbox\/xmax': dataset_util.float_list_feature(xmaxs),\n    'image\/object\/class\/text': dataset_util.bytes_list_feature(classes_text),\n    'image\/object\/class\/label': dataset_util.int64_list_feature(classes)\n<\/code><\/pre>\n\n<p>What am I doing wrong? All of tf object detection tutorials are using this generate_tfrecord.py from raccoon detector. I have found an avice at github to put coordinates in the right order, but it didn`t help me at all (code is given above).<\/p>\n","answers":[]}
+{"QuestionId":57990372,"AnswerCount":2,"Tags":"<python><numpy><tensorflow><tensorflow2.0>","CreationDate":"2019-09-18T10:14:39.863","AcceptedAnswerId":null,"OwnerUserId":10202729.0,"Title":"How to shuffle two numpy datasets using TensorFlow 2.0?","Body":"<p>I wish to write a function in <strong>TensorFlow 2.0<\/strong> than shuffles data and their target labels before each training iteration.<\/p>\n\n<p>Let's say I have two numpy datasets, X and y, representing data and labels for classification. How can I shuffle them <em>at the same time<\/em>?<\/p>\n\n<p>Using <code>sklearn<\/code> it's pretty easy:<\/p>\n\n<pre><code>from sklearn.utils import shuffle\nX, y = shuffle(X, y)\n<\/code><\/pre>\n\n<p>How can I do the same in <strong>TensorFlow 2.0<\/strong> ? The only tool I found in the documentation is <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/random\/shuffle\" rel=\"nofollow noreferrer\">tf.random.shuffle<\/a>, but it takes only one object at a time, I need to feed two.<\/p>\n","answers":[{"AnswerId":"57990622","CreationDate":"2019-09-18T10:27:32.370","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>If you just want to shuffle two arrays in the same way, you can do:<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Assuming X and y are initially NumPy arrays\nX = tf.convert_to_tensor(X)\ny = tf.convert_to_tensor(y)\n# Make random permutation\nperm = tf.random.shuffle(tf.range(tf.shape(X)[0]))\n# Reorder according to permutation\nX = tf.gather(X, perm, axis=0)\ny = tf.gather(y, perm, axis=0)\n<\/code><\/pre>\n\n<p>However, you may consider using a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset\" rel=\"nofollow noreferrer\"><code>tf.data.Dataset<\/code><\/a>, which already provides a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset#shuffle\" rel=\"nofollow noreferrer\"><code>shuffle<\/code><\/a> method.<\/p>\n\n<pre><code>import tensorflow as tf\n\n# You may use a placeholder if in graph mode\n# (see https:\/\/www.tensorflow.org\/guide\/datasets#consuming_numpy_arrays)\nds = tf.data.Dataset.from_tensor_slices((X, y))\n# Shuffle with some buffer size (len(X) will use a buffer as big as X)\nds = ds.shuffle(buffer_size=len(X))\n<\/code><\/pre>\n"},{"AnswerId":"57990643","CreationDate":"2019-09-18T10:28:14.337","ParentId":null,"OwnerUserId":"10288309","Title":null,"Body":"<p>Instead of shuffling x and y , its much easier to shuffle their indices, so first generate a list of indices<\/p>\n\n<p><code>indices = tf.range(start=0, limit=tf.shape(x_data)[0], dtype=tf.int32)<\/code><\/p>\n\n<p>then shuffle these indices<\/p>\n\n<p><code>idx = tf.random.shuffle(indices)<\/code><\/p>\n\n<p>and use these indices to shuffle the data<\/p>\n\n<pre><code>x_data = tf.gather(x_data, idx)\ny_data = tf.gather(y_data, idx)\n<\/code><\/pre>\n\n<p>and youll have shuffled data<\/p>\n"}]}
+{"QuestionId":57990468,"AnswerCount":0,"Tags":"<python><pandas><pytorch>","CreationDate":"2019-09-18T10:19:23.593","AcceptedAnswerId":null,"OwnerUserId":6557410.0,"Title":"How to load\/covert pandas dataframe with string values into pytorch dataset","Body":"<p>I am working on windows machine.\nI am trying to load pandas dataframe which has one columns as text(string) and another column as label (int). I want to load this dataframe into pytorch dataset.\nLet's say above data is in dataframe df.<\/p>\n\n<p>I am using below <\/p>\n\n<pre><code>train_target = torch.DoubleTensor(df['Text'].values)\n<\/code><\/pre>\n\n<p>I also used below to load directly from csv file.<\/p>\n\n<pre><code>TEXT1 = data.Field(tokenize = 'spacy', include_lengths = True)\nLABEL1 = data.LabelField(dtype = torch.float)\nfields = [('text',TEXT1),(label,LABEL1)]\ntrain_data, valid_data, test_data = data.TabularDataset.splits(\n                                        path = '.',\n                                        train = '1.csv',\n                                        validation = '1.csv',\n                                        test = '1.csv',\n                                        format = 'csv',\n                                        fields = fields,\n                                        skip_header = True\n)\n<\/code><\/pre>\n\n<p>But it is giving error as <\/p>\n\n<pre><code>OverflowError: Python int too large to convert to C long\n<\/code><\/pre>\n\n<p>I am searching for such text conversion into pytorch,tensor object . <\/p>\n","answers":[]}
+{"QuestionId":57991573,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-18T11:20:01.060","AcceptedAnswerId":null,"OwnerUserId":1179925.0,"Title":"Tensorflow: frozen graph and graph.get_collection","Body":"<p>Trying to print <code>tf.GraphKeys.GLOBAL_VARIABLES<\/code>, etc. after loading frozen graph:<\/p>\n\n<pre><code># Load frozen graph\nmodel_filepath = '.\/inception_v3_2016_08_28_frozen.pb'\n\ngraph_def = tf.GraphDef()\nwith tf.gfile.GFile(model_filepath, \"rb\") as f:\n    graph_def.ParseFromString(f.read())\n    print('type(graph_def)', type(graph_def))\n\ngraph = tf.Graph()\nwith graph.as_default():\n    tf.import_graph_def(graph_def, name=\"\")\n    print('type(graph)', type(graph))\n\nprint('len(graph.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)))',\n      len(graph.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)))\n\nprint('len(graph.get_collection(tf.GraphKeys.MODEL_VARIABLES)))',\n      len(graph.get_collection(tf.GraphKeys.MODEL_VARIABLES)))\n\nprint('len(graph.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)))',\n      len(graph.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>type(graph_def) &lt;class 'tensorflow.core.framework.graph_pb2.GraphDef'&gt;\ntype(graph) &lt;class 'tensorflow.python.framework.ops.Graph'&gt;\nlen(graph.get_collection(tf.GraphKeys.GLOBAL_VARIABLES))) 0\nlen(graph.get_collection(tf.GraphKeys.MODEL_VARIABLES))) 0\nlen(graph.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES))) 0\n<\/code><\/pre>\n\n<p>Why it return empty list of variables?<\/p>\n\n<p><strong>Update:<\/strong><\/p>\n\n<p>Looks related:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/39007396\/how-do-you-list-the-variables-in-the-graph-in-tensorflow\">How do you list the variables in the graph in tensorflow?<\/a><\/p>\n\n<blockquote>\n  <p>That shows contents of VARIABLES collection which does not get\n  restored with import_graph_def. However, it does get restored when you\n  import meta graph<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":57991733,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><artificial-intelligence><nsl>","CreationDate":"2019-09-18T11:28:45.520","AcceptedAnswerId":null,"OwnerUserId":112050.0,"Title":"Can Tensorflow-NSL solve the shortest path problem?","Body":"<p>I'm looking at the general class of graph algorithms that are solvable using neural networks. <\/p>\n\n<p>For example - <a href=\"https:\/\/medium.com\/octavian-ai\/finding-shortest-paths-with-graph-networks-807c5bbfc9c8\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/octavian-ai\/finding-shortest-paths-with-graph-networks-807c5bbfc9c8<\/a> - refers to an interesting solution to the shortest problem using graph neural networks. <\/p>\n\n<p>All the examples in NSL illustrate examples where the graph adds to information in an existing model. But can NSL solve graph problems itself ?<\/p>\n","answers":[{"AnswerId":"58139650","CreationDate":"2019-09-27T18:16:24.383","ParentId":null,"OwnerUserId":"12132173","Title":null,"Body":"<p>Although NSL could potentially be used to improve the accuracy or robustness of a neural network specifically trained to solve the shortest-path problem (like the RNN described in the Medium article you linked to), NSL was not developed to solve graph problems directly. That said, we'd certainly be interested in hearing about any progress in using our libraries to do so.<\/p>\n"}]}
+{"QuestionId":57992585,"AnswerCount":0,"Tags":"<python><performance><keras><neural-network><deep-learning>","CreationDate":"2019-09-18T12:15:36.217","AcceptedAnswerId":null,"OwnerUserId":12084489.0,"Title":"Why is training so slow on my neural network?","Body":"<p>I have two feed forward neural networks (built with keras in python) performing a similar task, yet the training time is completely different, I can't figure out the cause of this.<\/p>\n\n<p>I have a neural network taking inputs and outputs of shape:\ninput shape: (100000, 21) output shape: (100000, 9),\nthe first axis being the number of training samples, and the second the number of nodes in the input and output layer respectively.\nThen I have another neural network taking a subset of the above inputs and yielding a subset of the above outputs as such:\ninput shape: (100000, 13) output shape: (100000, 1)\nOther than this difference, the settings are exactly the same. Yet the first network is significantly faster to train, what could be causing this?<\/p>\n\n<p>These settings are:\na regression problem with 5 hidden layers of 128 nodes with relu activation functions, adam optimizer, linear activation for the output layer. I suppose the number of epochs is also different, but that shouldn't make a difference for the training time per epoch as far as I'm concerned. Batch size is 64, glorot-normal initialization.<\/p>\n\n<p>Here is some of the training output that keras provides of each network respectively, note that the network with less input and output nodes is significantly slower:<\/p>\n\n<pre><code>input shape: (100000, 21) output shape: (100000, 9),\nTrain on 90000 samples, validate on 10000 samples\nEpoch 1\/500\n90000\/90000 [==============================] - 3s 31us\/step - loss: 0.3145 - val_loss: 0.1749\nEpoch 2\/500\n90000\/90000 [==============================] - 2s 23us\/step - loss: 0.1426 - val_loss: 0.1298\nEpoch 3\/500\n90000\/90000 [==============================] - 2s 23us\/step - loss: 0.1161 - val_loss: 0.0999\nEpoch 4\/500\n90000\/90000 [==============================] - 2s \n<\/code><\/pre>\n\n<pre><code>input shape: (100000, 13) output shape: (100000, 1)\nTrain on 90000 samples, validate on 10000 samples\nEpoch 1\/100\n90000\/90000 [==============================] - 36s 398us\/step - loss: 0.0922 - val_loss: 0.0525\nEpoch 2\/100\n44608\/90000 [=============&gt;................] - ETA: 12s - loss: 0.0503\n<\/code><\/pre>\n\n<p>The training time per epoch does not decrease beyond 24 seconds in the second case, although I expected it to be faster given any form of stochastic gradient descent requires fewer gradient calculations given less input.<\/p>\n\n<p>Here is the code for the first neural network:<\/p>\n\n<pre><code># initiate:\nmodel = Sequential()\n\n# add hidden layers\nmodel.add(Dense(nodes, input_dim = n, activation='relu', kernel_initializer = 'glorot_normal'))\nfor i in range(layers):\n    model.add(Dense(nodes, activation = 'relu', kernel_initializer = 'glorot_normal'))\n\n# add output\nmodel.add(Dense(out, activation = 'linear', kernel_initializer = 'glorot_normal')) \n\n# compile the keras model\nmodel.compile(loss='mean_squared_error', optimizer='adam')\n\n# fit the keras model on the dataset\nprint('input shape:',np.shape(X_norm), 'output shape:', np.shape(L_norm))\nhistory = model.fit(X_norm, L_norm, epochs=epochs, batch_size=batch_size, validation_split = (1\/10) \n                   , callbacks = [earlystop]) \n<\/code><\/pre>\n\n<p>The second network is exactly the same, apart from the number of input in model.fit:<\/p>\n\n<pre><code># fit the keras model on the dataset\n    history = model.fit(inputdt, outputdt, epochs=epochs, batch_size=batch_size, validation_split = (1\/10) \n                       , verbose = 1, callbacks = [earlystop]) \n<\/code><\/pre>\n\n<p>where inputdt and outputdt are slices of X_norm and L_norm respectively.<\/p>\n","answers":[]}
+{"QuestionId":57993201,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-18T12:47:34.947","AcceptedAnswerId":null,"OwnerUserId":1946300.0,"Title":"How to assign a result of an operation to Layer.kernel?","Body":"<p>I have this simple model:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python.framework.graph_util import convert_variables_to_constants\n\ninput_shape = (28, 28, 1)\n\nl = tf.keras.layers\n\nmodel = tf.keras.Sequential([\n  l.Conv2D(32, 5, padding='same', activation='relu', input_shape=input_shape),\n])\n<\/code><\/pre>\n\n<p>Literally, a single convolution. Now I want to assign a result of different op to <code>model.layers[0].kernel<\/code>, i.e. I want the <code>kernel<\/code> to be formed in a special way during training and depend on another trainable parameter.<\/p>\n\n<p><em>Why would I want that? I'm experimenting with weight clustering and weights are supposed to be looked up from a trainable array. Lookup would happen with gather op.<\/em><\/p>\n\n<p>For now, I do a simple thing:<\/p>\n\n<pre><code>multiplier = tf.placeholder(shape=(1,), dtype=tf.float32)\nmodel.layers[0].kernel = tf.assign(model.layers[0].kernel, multiplier * model.layers[0].kernel)\n<\/code><\/pre>\n\n<p>I would hope that this will create another op in the graph that the convolutional kernel would depend on. However, after running this:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python.framework.graph_util import convert_variables_to_constants\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.all_variables()).difference(keep_var_names or []))\n    output_names = output_names or []\n    output_names += [v.op.name for v in tf.all_variables()]\n    # Graph -&gt; GraphDef ProtoBuf\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 = convert_variables_to_constants(session, input_graph_def,\n                                                  output_names, freeze_var_names)\n    return frozen_graph\nfrozen_graph = freeze_session(tf.keras.backend.get_session(),\n                              output_names=[out.op.name for out in model.outputs])\ntf.train.write_graph(frozen_graph, \"model\", \"tf_model.pb\", as_text=False)\n<\/code><\/pre>\n\n<p>I get this picture:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/IDTfH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/IDTfH.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The kernel does not appear to be dependant on the placeholder I've created. \nHowever, if I stop in a debugger and do:<\/p>\n\n<pre><code>&gt;&gt;&gt;model.layers[0].kernel\n<\/code><\/pre>\n\n<p>I get this:<\/p>\n\n<pre><code>&lt;tf.Variable 'AssignVariableOp' shape=(5, 5, 1, 32) dtype=float32&gt;\n<\/code><\/pre>\n\n<p>So the assignment clearly happens in the computational graph. \nWhat am I doing wrong?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"57994148","CreationDate":"2019-09-18T13:40:33.433","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Keras is not meant to use \"placeholders\" like that. It should use <code>Input<\/code> to create placeholders automatically and use them correctly in the model. The placeholder you created externally is certainly not considered a part of the \"Keras\" model (it's not performed in a Keras layer, it's not a Keras input).    <\/p>\n\n<p>So, you should have two Keras inputs:<\/p>\n\n<pre><code>data_input = Input((28,28,1))\nkernel_input = Input(tensor = your_placeholder)    \n<\/code><\/pre>\n\n<p>I used <code>tensor=<\/code> in the <code>Input<\/code> because it has a different batch size, and Keras usually complains about inputs that don't follow the same batch size of the others. <\/p>\n\n<p>Now this tensor should be part of the Keras model and appear in the graph. Notice that because it's a placeholder, it will not work with <code>model.fit<\/code>, you will need Tensorflow training where you feed this placeholder yourself. (You may use a variable that you update instead of a placeholder too).    <\/p>\n\n<p>Now about the Kernel. You will probably need to create a custom layer. That because <code>kernel<\/code> will only exist after you <code>build<\/code> the layer, and the layer will only be built when you call it on a tensor. That means, all the layer's ops will probably be already created by the point you are trying to change the Kernel, so it will probably not take effect, or take effect only on the next batch.   <\/p>\n\n<p>So, the suggestion is:<\/p>\n\n<pre><code>class MultConv2D(Conv2D):\n    def build(self, input_shape):\n        super(MultConv2D, self).build(input_shape)\n        self.kernel = kernel * multiplier \n<\/code><\/pre>\n\n<p>The troublesome part is where to input the multiplier, maybe in <code>__init__<\/code>, maybe in <code>call<\/code> (but rewriting the call method is way more complicated and will only support that the layer be called once)    <\/p>\n"}]}
+{"QuestionId":57993767,"AnswerCount":0,"Tags":"<c++><tensorflow><bazel><bin>","CreationDate":"2019-09-18T13:19:48.210","AcceptedAnswerId":null,"OwnerUserId":9855322.0,"Title":"How to deploy a release build of a Tensorflow project with Bazel","Body":"<p>I have a collegue that works on a project who needs a CNN model I built. I prepared the C++ code that will take from a pipe the data she wants to make prediction on and my process will send back the prediction. The problem is I don't know how to make a standalone release using Bazel. I would like to give her my bin file and avoid her to have to deal with tensorflow and bazel.\nDoes anyone know how to get the bin file with all the needs to run it standalone ?<\/p>\n\n<p>I know I could use a server approach, but my code is already here and I just want to give it to her.<\/p>\n\n<p>Thanks !<\/p>\n","answers":[]}
+{"QuestionId":57993832,"AnswerCount":0,"Tags":"<python><c><tensorflow><keras>","CreationDate":"2019-09-18T13:23:21.390","AcceptedAnswerId":null,"OwnerUserId":4045358.0,"Title":"Load weights from checkpoint not working in keras model","Body":"<p>I am going insane over this.<\/p>\n\n<p>I define a sequential model using tensorflow keras:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = tf.keras.Sequential([tf.keras.layer.Dense(128,input_shape(784,),activation=\"relu\"),\n                             tf.keras.layer.Dense(10,activation=\"softmax\"])\nmodel.compile(optimizer=\"adam\",loss=\"mse\")\nkeras.experimental.export_saved_model(model,\"keras_model\")\n<\/code><\/pre>\n\n<p>I train said model in a <a href=\"https:\/\/gist.github.com\/7PintsOfCherryGarcia\/d6d3df8e7282f2e635bf9362c81f9292\" rel=\"nofollow noreferrer\">C program<\/a> using c_api.h<\/p>\n\n<p>C program saves weights in a checkpoint file.<\/p>\n\n<p>When trying to restore weights in python from checkpoint file with:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>keras.experimental.load_from_saved_model(\"keras_model\/\")\n#OR\nmodel = tf.keras.Sequential([tf.keras.layer.Dense(128,input_shape(784,),activation=\"relu\"),\n                             tf.keras.layer.Dense(10,activation=\"softmax\"])\nmodel.load_weights(\"keras_model\/variables\/variables\")\n#OR\ncheckpoint = tf.train.Checkpoint(model=model)\nstatus = checkpoint.restore(\"keras_model\/variables\/variables\")\n<\/code><\/pre>\n\n<p>I end up getting an error and no weights are restored. <\/p>\n\n<p>I am able to restore weights and continue training in my C program<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>keras.experimental.load_from_saved_model(\"keras_model\/\")\nWARNING: Logging before flag parsing goes to stderr.\nW0918 15:18:04.350199 140418474760000 deprecation.py:323] From &lt;ipython-input-2-06ea110fdc8e&gt;:1: load_from_saved_model (from tensorflow.python.keras.saving.saved_model_experimental) is deprecated and will be removed in a future version.\nInstructions for updating:\nThe experimental save and load functions have been  deprecated. Please switch to `tf.keras.models.load_model`.\n2019-09-18 15:18:04.390271: I tensorflow\/core\/platform\/profile_utils\/cpu_utils.cc:94] CPU Frequency: 1696040000 Hz\n2019-09-18 15:18:04.390913: I tensorflow\/compiler\/xla\/service\/service.cc:168] XLA service 0x4bf4790 executing computations on platform Host. Devices:\n2019-09-18 15:18:04.390961: I tensorflow\/compiler\/xla\/service\/service.cc:175]   StreamExecutor device (0): Host, Default Version\nW0918 15:18:04.436281 140418474760000 deprecation.py:323] From \/home\/jregalado\/Projects\/tensorflow\/env\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/training\/tracking\/util.py:1249: NameBasedSaverStatus.__init__ (from tensorflow.python.training.tracking.util) is deprecated and will be removed in a future version.\nInstructions for updating:\nRestoring 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---------------------------------------------------------------------------\nAssertionError                            Traceback (most recent call last)\n&lt;ipython-input-2-06ea110fdc8e&gt; in &lt;module&gt;\n----&gt; 1 keras.experimental.load_from_saved_model(\"keras_model\/\")\n\n~\/Projects\/tensorflow\/env\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/util\/deprecation.py in new_func(*args, **kwargs)\n322               'in a future version' if date is None else ('after %s' % date),\n323               instructions)\n--&gt; 324       return func(*args, **kwargs)\n325     return tf_decorator.make_decorator(\n326         func, new_func, 'deprecated',\n\n~\/Projects\/tensorflow\/env\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/keras\/saving\/saved_model_experimental.py in load_from_saved_model(saved_model_path, custom_objects)\n425       compat.as_text(constants.VARIABLES_DIRECTORY),\n426       compat.as_text(constants.VARIABLES_FILENAME))\n--&gt; 427   model.load_weights(checkpoint_prefix)\n428   return model\n\n~\/Projects\/tensorflow\/env\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/keras\/engine\/training.py in load_weights(self, filepath, by_name)\n179         raise ValueError('Load weights is not yet supported with TPUStrategy '\n180                          'with steps_per_run greater than 1.')\n--&gt; 181     return super(Model, self).load_weights(filepath, by_name)\n182\n183   @trackable.no_automatic_dependency_tracking\n\n~\/Projects\/tensorflow\/env\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/keras\/engine\/network.py in load_weights(self, filepath, by_name)\n1372         # streaming restore for any variables created in the future.\n1373         trackable_utils.streaming_restore(status=status, session=session)\n-&gt; 1374       status.assert_nontrivial_match()\n1375       return status\n1376     if h5py is None:\n\n~\/Projects\/tensorflow\/env\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/training\/tracking\/util.py in assert_nontrivial_match(self)\n964     # assert_nontrivial_match and assert_consumed (and both are less\n965     # useful since we don't touch Python objects or Python state).\n--&gt; 966     return self.assert_consumed()\n967\n968   def _gather_saveable_objects(self):\n\n~\/Projects\/tensorflow\/env\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/training\/tracking\/util.py in assert_consumed(self)\n941       raise AssertionError(\n942           \"Some objects had attributes which were not restored:{}\".format(\n--&gt; 943               \"\".join(unused_attribute_strings)))\n944     for trackable in self._graph_view.list_objects():\n945       # pylint: disable=protected-access\n\nAssertionError: Some objects had attributes which were not restored:\n&lt;tf.Variable 'a\/kernel:0' shape=(784, 128) dtype=float32, numpy=\narray([[-0.03716458, -0.04911711, -0.01023878, ...,  0.0636776 ,\n0.02892563, -0.05542086],\n[-0.02324755, -0.07362694, -0.0399951 , ...,  0.0680329 ,\n0.05201877, -0.05149256],\n[ 0.00954343,  0.05673491,  0.05108347, ...,  0.01994208,\n-0.01107961,  0.06192174],\n...,\n[ 0.07091486, -0.07734856, -0.04417738, ...,  0.01921409,\n-0.01908814, -0.05070668],\n[ 0.01353646, -0.05189713, -0.01391671, ..., -0.05795977,\n0.04801518,  0.00801209],\n[-0.05304915,  0.01870193,  0.05657425, ..., -0.06819408,\n-0.00760372, -0.0106293 ]], dtype=float32)&gt;: ['a\/kernel']\n&lt;tf.Variable 'a\/bias:0' shape=(128,) dtype=float32, numpy=\narray([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,\n0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,\n0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,\n0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,\n0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,\n0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,\n0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,\n0., 0., 0., 0., 0., 0., 0., 0., 0.], dtype=float32)&gt;: ['a\/bias']\n&lt;tf.Variable 'b\/kernel:0' shape=(128, 10) dtype=float32, numpy=\narray([[-0.1759212 , -0.09282549, -0.11045764, ..., -0.13727605,\n-0.02849793,  0.14510198],\n[ 0.06857841, -0.01459177,  0.08369003, ...,  0.05089156,\n-0.05319159, -0.08594933],\n[-0.180914  , -0.18932283,  0.20551099, ..., -0.17210156,\n-0.10069884,  0.06433241],\n...,\n[ 0.09097584, -0.03930017, -0.15125516, ...,  0.02359283,\n-0.16158347, -0.13176063],\n[-0.04145582, -0.03205152,  0.20097663, ..., -0.15124482,\n0.16874255, -0.15434337],\n[-0.13188484,  0.04145408,  0.05036192, ..., -0.10489662,\n0.12316228,  0.08794598]], dtype=float32)&gt;: ['b\/kernel']\n&lt;tf.Variable 'b\/bias:0' shape=(10,) dtype=float32, numpy=array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0.], dtype=float32)&gt;: ['b\/bias']\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57993899,"AnswerCount":1,"Tags":"<python><pytorch><google-colaboratory><tensor><index-error>","CreationDate":"2019-09-18T13:27:33.710","AcceptedAnswerId":58002399.0,"OwnerUserId":12079890.0,"Title":"How to use tensor.item() ? IndexError: invalid index of a 0-dim tensor. Use tensor.item() to convert a 0-dim tensor to a Python number","Body":"<p>Im pretty new to Siamese Neural Networks and recently found <a href=\"https:\/\/innovationincubator.com\/siamese-neural-network-with-pytorch-code-example\/\" rel=\"nofollow noreferrer\">this example<\/a> and <a href=\"https:\/\/colab.research.google.com\/drive\/1FsixLon5Zz3_rFA0xIHzc8Tvnnw8FLr8#scrollTo=EDMURKIHPI7E\" rel=\"nofollow noreferrer\">Colab notebook<\/a>.<\/p>\n\n<p>When running the code i get the following error:    <\/p>\n\n<blockquote>\n  <p>IndexError: invalid index of a 0-dim tensor. Use tensor.item() to\n  convert a 0-dim tensor to a Python number<\/p>\n<\/blockquote>\n\n<p>on the line:<\/p>\n\n<pre><code>result=torch.max(res,1)[1][0][0][0].data[0].tolist()\n<\/code><\/pre>\n\n<p>I found something about the <code>tensor.item()<\/code> but i really just don't know how to use it here.<\/p>\n\n<p>EDIT:<\/p>\n\n<pre><code>test_dataloader = DataLoader(test_dataset,num_workers=6,batch_size=1,shuffle=True)\naccuracy=0\ncounter=0\ncorrect=0\nfor i, data in enumerate(test_dataloader,0): \nx0, x1 , label = data\n# onehsot applies in the output of 128 dense vectors which is then  converted to 2 dense vectors\noutput1,output2 = model(x0.to(device),x1.to(device))\nres=torch.abs(output1.cuda() - output2.cuda())\nlabel=label[0].tolist()\nlabel=int(label[0])\nresult=torch.max(res,1)[1][0][0][0].data.item().tolist()\nif label == result:\ncorrect=correct+1\ncounter=counter+1\n#   if counter ==20:\n#      break\n\naccuracy=(correct\/len(test_dataloader))*100\nprint(\"Accuracy:{}%\".format(accuracy))\n<\/code><\/pre>\n\n<p>Thats the code i get the error on. <\/p>\n","answers":[{"AnswerId":"58002399","CreationDate":"2019-09-19T01:03:56.273","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<p>What this error message says is that you're trying to index into an array which has just one item in it. For example,<\/p>\n\n<pre><code>In [10]: aten = torch.tensor(2)   \n\nIn [11]: aten  \nOut[11]: tensor(2)\n\nIn [12]: aten[0]\n---------------------------------------------------------------------------\nIndexError Traceback (most recent call last)\n&lt;ipython-input-12-5c40f6ab046a&gt; in &lt;module&gt;\n----&gt; 1 aten[0]\n\nIndexError: invalid index of a 0-dim tensor.  Use tensor.item() to convert a 0-dim \ntensor to a Python number\n<\/code><\/pre>\n\n<p>In the above case, <code>aten<\/code> is a tensor with just a single number in it. So, using an index (or more) to retrieve that number throws the <code>IndexError<\/code>.<\/p>\n\n<p>The correct way to extract the number(item) out of the tensor is to use <code>tensor.item()<\/code>, here <code>aten.item()<\/code> as in:<\/p>\n\n<pre><code>In [14]: aten.item()\nOut[14]: 2\n<\/code><\/pre>\n"}]}
+{"QuestionId":57993939,"AnswerCount":0,"Tags":"<python><pytorch><torch><tensor><numpy-ndarray>","CreationDate":"2019-09-18T13:29:37.020","AcceptedAnswerId":null,"OwnerUserId":9687599.0,"Title":"Convert n-dimensional numpy array into torch tensor","Body":"<p>I have 4-dimensional numpy array (<code>type(x)<\/code> is <code>ndarray<\/code>, <code>type(x[i])<\/code> is <code>ndarray<\/code>, <code>type(x[i][j])<\/code> is <code>ndarray<\/code>, and <code>type(x[i][j][k])<\/code> is <code>float32<\/code>). I need to convert it to torch tensor to use it as part of dataset. I tried to stack <code>torch.Tensor<\/code> like that:<\/p>\n\n<pre><code>def convert_to_tensor(ndarr):\n    for i in ndarr:\n        for j in ndarr[i]:\n            tensor_ndarr = torch.stack([torch.Tensor(k) for k in ndarr[i][j]])\n    return tensor_ndarr\n<\/code><\/pre>\n\n<p>but not sure if that gonna work.\nIs there more simple or convenient way to do that?<\/p>\n\n<p>P.S. X looks like that: <\/p>\n\n<pre><code>[array([[0.0000000e+00, 0.0000000e+00, 0.0000000e+00, ..., 9.1042261e-07,\n        4.2489510e-06, 1.0829879e-07],\n       [0.0000000e+00, 0.0000000e+00, 0.0000000e+00, ..., 4.5446509e-06,\n        1.0727225e-05, 7.5697239e-06],\n       [0.0000000e+00, 0.0000000e+00, 0.0000000e+00, ..., 1.4139889e-06,\n        8.1363060e-06, 5.8842788e-06],\n       ...,\n       [0.0000000e+00, 0.0000000e+00, 0.0000000e+00, ..., 1.2733157e-07,\n        3.6470925e-07, 2.5473477e-07],\n       [0.0000000e+00, 0.0000000e+00, 0.0000000e+00, ..., 7.2166131e-08,\n        2.0409759e-07, 2.3035474e-07],\n       [0.0000000e+00, 0.0000000e+00, 0.0000000e+00, ..., 5.7966993e-08,\n        1.7371592e-07, 1.5426794e-07]], dtype=float32)\n...\n array([[0.0000000e+00, 0.0000000e+00, 0.0000000e+00, ..., 9.1042261e-07,\n        4.2489510e-06, 1.0829879e-07],\n       [0.0000000e+00, 0.0000000e+00, 0.0000000e+00, ..., 4.5446509e-06,\n        1.0727225e-05, 7.5697239e-06],\n       [0.0000000e+00, 0.0000000e+00, 0.0000000e+00, ..., 1.4139889e-06,\n        8.1363060e-06, 5.8842788e-06],\n       ...,\n       [0.0000000e+00, 0.0000000e+00, 0.0000000e+00, ..., 1.2733157e-07,\n        3.6470925e-07, 2.5473477e-07],\n       [0.0000000e+00, 0.0000000e+00, 0.0000000e+00, ..., 7.2166131e-08,\n        2.0409759e-07, 2.3035474e-07],\n       [0.0000000e+00, 0.0000000e+00, 0.0000000e+00, ..., 5.7966993e-08,\n        1.7371592e-07, 1.5426794e-07]], dtype=float32)]\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57993992,"AnswerCount":0,"Tags":"<tensorflow><keras><lstm><tf.keras>","CreationDate":"2019-09-18T13:32:22.537","AcceptedAnswerId":null,"OwnerUserId":11541912.0,"Title":"How to convert keras CuDNNLSTM layer to LSTM (and CuDNNGRU to GRU)?","Body":"<p>I am trying to convert <code>tf.keras.layers.CuDNNLSTM<\/code> to <code>tf.keras.layers.LSTM<\/code> (and <code>tf.keras.layers.CuDNNGRU<\/code> to <code>tf.keras.layers.GRU<\/code>). <\/p>\n\n<p>I see from <a href=\"https:\/\/stackoverflow.com\/questions\/47003418\/using-keras-how-can-i-load-weights-generated-from-cudnnlstm-into-lstm-model\">this StackOverflow post<\/a> and <a href=\"https:\/\/github.com\/keras-team\/keras\/pull\/8307\" rel=\"nofollow noreferrer\">this GitHub PR<\/a> that the CuDNNLSTM --> LSTM conversion is possible, and I see from <a href=\"https:\/\/github.com\/keras-team\/keras\/pull\/9112\" rel=\"nofollow noreferrer\">this GitHub PR<\/a> that the conversion is possible, but I'm seeing some unexpected behavior during inference, so I'm not sure if I'm doing it correctly.<\/p>\n\n<p>The way I am doing this now is through creating (and training) a model which has a <code>CuDNNLSTM<\/code> layer...<\/p>\n\n<pre><code>    inputs = tf.keras.Input(shape=(120,))\n    x = layers.CuDNNLSTM(100)(inputs)   # or x = layers.CuDNNGRU(100)(inputs)\n    outputs = layers.Dense(5, activation='softmax')(x)\n\n    model = tf.keras.Model(inputs=inputs, outputs=outputs)\n<\/code><\/pre>\n\n<p>... compiling and training the model ...<\/p>\n\n<pre><code>    model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])\n    model.fit(...)\n<\/code><\/pre>\n\n<p>... saving the model as both a json for the structure and an hdf5 file for weights, and editing the file to say it's an <code>LSTM<\/code> layer instead of a <code>CuDNNLSTM<\/code> layer.<\/p>\n\n<pre><code>    model.save_weights('\/path\/to\/weights.h5')\n    # get json, edit layer name, dump into file\n    json_config = model.to_json()\n    with open('\/path\/to\/structure.json', 'w') as json_file:\n        json_config_ = json.loads(json_config)\n        layers = json_config_['config']['layers']\n        for layer in layers:\n            if layer['class_name'].lower() == 'cudnnlstm':\n                layer['class_name'] = 'LSTM'\n            elif layer['class_name'].lower() == 'cudnngru':\n                layer['class_name'] = 'GRU'\n\n        json.dump(json_config_, json_file)\n<\/code><\/pre>\n\n<p>I am taking <code>'\/path\/to\/weights.h5'<\/code> and <code>'\/path\/to\/structure.json'<\/code> and converting them to coreml to do inference on an iphone, so it might be failing during the coreml conversion. I want to check if I am doing the conversion from CuDNN layers to LSTM\/GRU layers correctly first, before investigating the coreml converter. <b>Is this the correct way to convert from CuDNNLSTM to LSTM and from CuDNNGRU to GRU? Is there another way to do the conversion that I should be doing instead?<\/b><\/p>\n","answers":[]}
+{"QuestionId":57995171,"AnswerCount":0,"Tags":"<tensorflow><keras><distributed>","CreationDate":"2019-09-18T14:28:52.087","AcceptedAnswerId":null,"OwnerUserId":6450489.0,"Title":"Need a clear simple approach to distributed learning in Tensorflow\/Keras","Body":"<p>I have multiple machines, some with GPUs and some others not. I also have a keras model that works fine on a single machine but I want to train it in a distributed mode because I want to test it with a huge dataset and with a bigger number of layers.\nThere is quite a lot of pages discussing the distribution strategy in tf.distribute, but at the same time there are a lot other pages showing how to do it with encapsulating keras model with an estimator, setting up the TF_CONFIG parameter and then call <code>tf.estimator.train_and_evaluate<\/code>.\nI used the second approach personally as it was more straightforward and am struggling to tune and debug it. It works anyway, but I am very confused what is the point in all of strategy-related stuff as I don't see any use them in the second approach, and the documentation is not helping to clear it.<\/p>\n\n<p>I also have some doubt if my file setting environment is correct: My understanding is that the PS server is going to hold the model parameters and the chief server is going to administer the whole training process, distributing data, and saving summaries and checkpoints. So I assume that:<\/p>\n\n<p>0- I need only one chief server, at least one PS server, possibly some workers, and one evaluator. The data and parameter sharing and communication between all of these servers is done by system and I am not engaged in it.<\/p>\n\n<p>1- The main python code for all machines should be exactly the same, except in TF_CONFIG definition that defines the task and index for that specific machine.<\/p>\n\n<p>2- I should have one shared copy of data in a folder available to all chief and workers.<\/p>\n\n<p>3- I should have one shared log directory accessible by all machines as defined in tf.estimator.RunConfig.<\/p>\n\n<p>4- Having this setting then a piece of code such as below will do the job (assuming the <code>model<\/code> has been defined elsewhere and the <code>read_datasets<\/code> function returns features and labels for running the model):<\/p>\n\n<pre><code>runConfig = tf.estimator.RunConfig(\n        session_config=config,\n        model_dir=log_dir,\n        save_summary_steps=1,\n        save_checkpoints_steps=train_steps\n        )\nestimator = tf.keras.estimator.model_to_estimator(model, model_dir=log_dir, config=runConfig)\ntrain_spec = tf.estimator.TrainSpec(input_fn=lambda: read_datasets(...), max_steps=epochs*train_steps)\neval_spec = tf.estimator.EvalSpec(input_fn=lambda: read_datasets(...), start_delay_secs=1, throttle_secs=1)\ntf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)\n<\/code><\/pre>\n\n<p>Although the above approach seems to work fine, I still have some difficulties understanding how the chief is partitioning the dataset among the workers and how to set the train_step and batch_size in this approach. Also I don't know how can I report accuracy and other metrics such as precision\/recall\/F1 in addition to loss when running the <code>tf.estimator.evaluate<\/code> without writing a custom model_fn for my encapsulated keras estimator.<\/p>\n","answers":[]}
+{"QuestionId":57995821,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><lstm><recurrent-neural-network>","CreationDate":"2019-09-18T15:03:17.467","AcceptedAnswerId":null,"OwnerUserId":11931220.0,"Title":"Variable names at inference time don't match checkpoint vars when using TimeDistributed and RNN layers for training","Body":"<p>I'm training an LSTM Encoder-Decoder to take in an input sequence and predict the next N steps. At training time I train on sequences, using tf.keras.layers.TimeDistributed() to wrap my convolutional layers and tf.keras.layers.RNN() to unroll my LSTM cell for the sequence length.\nAt inference time however, I am no longer running on sequences but passing in single examples. I therefore do not use the TimeDistributed and RNN layers anymore. The problem is that the variables I am restoring from my checkpoint now have different names from my inference graph's variables, as the training variables are named by the TimeDistributed and RNN layers. e.g. 'Network\/encoder\/time_distributed\/kernel' at training vs. 'Network\/encoder\/conv2d\/kernel' at inference. The shapes match, but the names do not. Are there any simple solutions to this which avoid me having to manually rename all of the variables in my checkpoint file?<\/p>\n","answers":[]}
+{"QuestionId":57995859,"AnswerCount":1,"Tags":"<python><neural-network><pytorch>","CreationDate":"2019-09-18T15:05:25.340","AcceptedAnswerId":57999454.0,"OwnerUserId":11452683.0,"Title":"PyTorch: How to write a neural network that only returns the weights?","Body":"<p>I'm training a neural network that learns some weights and based on those weights, I compute transformations that produce the predicted model in combination with the weights. My network doesn't learn properly and therefore I'm writing a different network that does nothing but returning the weights independent from the input <code>x<\/code> (after normalization with softmax and transpose). This way, I want to find out whether the problem lies in the network or in the transformation estimation outside the network. But this doesn't work. This is what I've got.<\/p>\n\n<pre><code>class DoNothingNet(torch.nn.Module):\n    def __init__(self, n_vertices=6890, n_joints=14):\n        super(DoNothingNet, self).__init__()\n        self.weights = nn.parameter.Parameter(torch.randn(n_vertices, n_joints))\n\n    def forward(self, x, indices):\n        self.weights = F.softmax(self.weights, dim=1)\n        return self.weights.transpose(0,1)\n<\/code><\/pre>\n\n<p>But the line <code>self.weights = F.softmax(self.weights, dim=1)<\/code> doesn't work and produces the error <code>TypeError: cannot assign 'torch.cuda.FloatTensor' as parameter 'weights' (torch.nn.Parameter or None expected)<\/code>. How do I fix this? And does the code even make sense?<\/p>\n","answers":[{"AnswerId":"57999454","CreationDate":"2019-09-18T19:15:19.397","ParentId":null,"OwnerUserId":"12054165","Title":null,"Body":"<p>nn.Module tracks all fields of type nn.Parameter for training. In your code every forward call you try to change parameters weights by assigning it to Tensor type, so the error occurs.<\/p>\n\n<p>The following code outputs normalised weights without changing the stored ones. Hope this will help.<\/p>\n\n<pre><code>import torch\nfrom torch import nn\nfrom torch.nn import functional as F\n\nclass DoNothingNet(torch.nn.Module):\n    def __init__(self, n_vertices=6890, n_joints=14):\n        super(DoNothingNet, self).__init__()\n        self.weights = nn.parameter.Parameter(torch.randn(n_vertices, n_joints))\n\n    def forward(self, x, indices):\n        output = F.softmax(self.weights, dim=1)\n        return output.transpose(0,1)\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":57995975,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2019-09-18T15:12:06.447","AcceptedAnswerId":null,"OwnerUserId":3407000.0,"Title":"Tensorflow: what role does the scope in which a tensor is created play on a tf.data.Dataset?","Body":"<p>I have been experimenting with <code>tf.data.Dataset<\/code>, and I have found a situation in which I am struggling to understand Tensorflow's behavior. A short example is as follows:<\/p>\n\n<pre><code>import tensorflow as tf\n\nrd = tf.random_uniform((), 0, 1.0)\n\n\ndef _process_example(example):\n    rand = tf.random_uniform((), 0, 1.0)\n\n    rd2 = rand\n    # rd2 = rd\n    example['val'] = rd2\n\n    return example\n\n\ndef _create_dataset():\n    def generator():\n        for n in [i for i in range(10)]:\n            yield {'n': n}\n\n    dataset = tf.data.Dataset.from_generator(generator, {'n': tf.int32})\n    dataset = dataset.map(_process_example)\n\n    return dataset\n\n\ndef _create_iterator():\n    dataset = _create_dataset()\n    data_iter = dataset.make_one_shot_iterator()\n    elem = data_iter.get_next()\n\n    return elem\n\n\nif __name__ == '__main__':\n    elem = _create_iterator()\n\n    with tf.Session() as sess:\n        for i in range(10):\n            print(sess.run(elem))\n<\/code><\/pre>\n\n<p>In the code above, Tensorflow allows me to create the dataset and print the elements without errors. However, if I comment out <code>rd2 = rand<\/code> and uncomment <code>rd2 = rd<\/code> effectively changing the variable <code>rd2<\/code> to point to the tensor created in the global scope instead of the one created in the scope of the function (apart from this, both are created identically), Tensorflow crashes with the following error:<\/p>\n\n<pre><code>ValueError: Failed to create a one-shot iterator for a dataset. `Dataset.make_one_shot_iterator()` does not support datasets that capture stateful objects, such as a `Variable` or `LookupTable`. In these cases, use `Dataset.make_initializable_iterator()`. (Original error: Cannot capture a stateful node (name:random_uniform\/RandomUniform, type:RandomUniform) by value.)\n<\/code><\/pre>\n\n<p>I am aware that <code>tf.data.Iterator<\/code> can be used to create an initiable iterator instead of the one-shot iterator that I am using. However, after trying that I have found that the dataset takes the random value generated at the time the <code>init_op<\/code> of the iterator is run for every element of the dataset if I use the tensor created in the global scope, but not so if I use the tensor created within the function, in which case I get a different random number for each element of the dataset.<\/p>\n\n<p>What is the explanation behind this behavior?<\/p>\n","answers":[]}
+{"QuestionId":57996019,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-18T15:14:03.657","AcceptedAnswerId":null,"OwnerUserId":3413239.0,"Title":"How to use layer normalization in tensorflow 1.12?","Body":"<p>I am stuck with tensorflow 1.12, and I need to use layer normalization. I can't find some examples of this, and as I am new to tensorflow I am unable to figure out where I am going wrong. <\/p>\n\n<p><code>tf.contrib.layers.layer_norm<\/code> is the function that I want to include in my <code>tf.keras.Sequential()<\/code> like this - <\/p>\n\n<pre><code>        self.module = K.Sequential([\n            tf.contrib.layers.layer_norm(trainable=True),\n            K.layers.Activation(self.activation),\n            K.layers.Dense(units=self.output_size, activation=None, kernel_initializer=self.initializer)\n        ])\n<\/code><\/pre>\n\n<p>I also tried using <\/p>\n\n<pre><code>self.ln = tf.contrib.layers.layer_norm(trainable=True)\n### and in call()\nself.ln(self.module)\n<\/code><\/pre>\n\n<p>In all the cases, it throws the error at the line defining <code>tf.contrib.layers.layer_norm(trainable=True)<\/code>-<\/p>\n\n<pre><code>TypeError: layer_norm() missing 1 required positional argument: 'inputs'\n<\/code><\/pre>\n\n<p>I understand that the inputs need to be given as the argument to layernorm, but if I want it to trainable, it can only be defined in <code>__init__()<\/code>. Where am I going wrong?<\/p>\n\n<p>I use mainly PyTorch, so it is quite obvious that I am not able to grasp the ideology of <code>tf<\/code>. Any suggestions will be very helpful!<\/p>\n","answers":[{"AnswerId":"57999613","CreationDate":"2019-09-18T19:28:56.270","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p><code>Sequential<\/code> needs to be initialized by a list of <code>Layer<\/code> instances, such as <code>tf.keras.layers.Activation<\/code>, <code>tf.keras.layers.Dense<\/code>. <code>tf.contrib.layers.layer_norm<\/code> is functional instead of <code>Layer<\/code> instance. <\/p>\n\n<p>There is a third party implementation of layer normalization in keras style - <a href=\"https:\/\/github.com\/CyberZHG\/keras-layer-normalization\" rel=\"nofollow noreferrer\">keras-layer-normalization<\/a>. But I haven't tested in tensorflow. <\/p>\n"}]}
+{"QuestionId":57996557,"AnswerCount":0,"Tags":"<tensorflow><go>","CreationDate":"2019-09-18T15:47:30.753","AcceptedAnswerId":null,"OwnerUserId":6228832.0,"Title":"Create a tf graph of shape [1,224,224,3] without using any resizing function in Go","Body":"<p>I am trying to use a TensorFlow model in Go. Frankly, I know nothing about Machine learning. I am just trying to run the model. I've been successful so far by doing as below<\/p>\n\n<pre><code>    func (m *Engine) Run(ctx context.Context, r io.Reader, resizType bool) (probability float64, err error) {\n        \/\/ Create a session for inference over graph.\n        session, err := tf.NewSession(m.model, nil)\n        if err != nil {\n            log.Fatal(err)\n        }\n        defer session.Close()\n\n        tensor, err := makeTensorFromImage(ctx, r, resizType)\n        if err != nil {\n            return 0, err\n        }\n\n        output, err := session.Run(\n            map[tf.Output]*tf.Tensor{\n                m.model.Operation(\"model_2_input\").Output(0): tensor,\n            },\n            []tf.Output{\n                m.model.Operation(\"dense_4\/Sigmoid\").Output(0),\n            },\n            nil)\n        if err != nil {\n            fmt.Println(err)\n        }\n\n        probabilities := output[0].Value().([][]float32)[0]\n\n        return float64(probabilities[0]), nil\n    }\n\nfunc makeTensorFromImage(ctx context.Context, image io.Reader, resizType bool) (*tf.Tensor, error) {\n    bytes := StreamToByte(image)\n\n    tensor, err := tf.NewTensor(string(bytes))\n    if err != nil {\n        return nil, err\n    }\n\n    \/\/ Construct a graph to normalize the image\n    graph, input, output, err := constructGraphToNormalizeImage(resizType)\n    if err != nil {\n        return nil, err\n    }\n\n    \/\/ Execute that graph to normalize this one image\n    session, err := tf.NewSession(graph, nil)\n    if err != nil {\n        return nil, err\n    }\n    defer session.Close()\n    normalized, err := session.Run(\n        map[tf.Output]*tf.Tensor{input: tensor},\n        []tf.Output{output},\n        nil)\n    if err != nil {\n        return nil, err\n    }\n\n    return normalized[0], nil\n}\n<\/code><\/pre>\n\n<p>The below func give me a graph of shape [1,224,224,3] .<\/p>\n\n<pre><code>func constructGraphToNormalizeImage(resizType bool) (graph *tf.Graph, input, output tf.Output, err error) {\n    const (\n        H, W  = 224, 224\n        Mean  = float32(0)\n        Scale = float32(255)\n    )\n    s := op.NewScope()\n    input = op.Placeholder(s, tf.String)\n    if resizType {\n        output = op.Div(s,\n            op.Sub(s,\n                op.ResizeArea(s,\n                    op.ExpandDims(s,\n                        op.Cast(s,\n                            op.DecodeJpeg(s, input, op.DecodeJpegChannels(3)), tf.Float),\n                        op.Const(s.SubScope(\"make_batch\"), int32(0))),\n                    op.Const(s.SubScope(\"size\"), []int32{H, W})),\n                op.Const(s.SubScope(\"mean\"), Mean)),\n            op.Const(s.SubScope(\"scale\"), Scale))\n        graph, err = s.Finalize()\n    } else {\n        output = op.Div(s,\n            op.Sub(s,\n                op.ResizeNearestNeighbor(s,\n                    op.ExpandDims(s,\n                        op.Cast(s,\n                            op.DecodeJpeg(s, input, op.DecodeJpegChannels(3)), tf.Float),\n                        op.Const(s.SubScope(\"make_batch\"), int32(0))),\n                    op.Const(s.SubScope(\"size\"), []int32{H, W})),\n                op.Const(s.SubScope(\"mean\"), Mean)),\n            op.Const(s.SubScope(\"scale\"), Scale))\n        graph, err = s.Finalize()\n    }\n    return graph, input, output, err\n}\n<\/code><\/pre>\n\n<p>Now I don't want to use any resize function of any kind. I tried removing the resize func bit it gives me a shape of [1,?,?,3] obviously. So my question is how do I achieve what I need without resizing.<\/p>\n","answers":[]}
+{"QuestionId":57996658,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><computer-vision>","CreationDate":"2019-09-18T15:53:43.130","AcceptedAnswerId":null,"OwnerUserId":3029270.0,"Title":"understanding optimizer in tensorflow","Body":"<p>in tensorflow <\/p>\n\n<p>if I have optimizer like below <\/p>\n\n<pre><code> with tf.device('\/gpu:'+GPU0):\n        aeu_var = [var for var in tf.trainable_variables() if var.name.startswith('aeu_2')]\n\n        self.aeu_g_optim = tf.train.AdamOptimizer(learning_rate=0.0001, beta1=0.9, beta2=0.999, epsilon=1e-8).\\\n                        minimize(self.aeu_loss, var_list=aeu_var)\n<\/code><\/pre>\n\n<p>in my main loop \nI calculate the loss <code>self.aeu_loss<\/code> \nthen I run session as follow<\/p>\n\n<pre><code> self.sess2.run(self.aeu_g_optim, feed_dict={self.aeu_loss:loss})\n<\/code><\/pre>\n\n<p>so I already did the work of computing the loss and feed it when I want to run  the optimizer \nhowever an error still shown asking for feeding PLACEHOLDER which is an input to the model <\/p>\n\n<p>my Qs:<\/p>\n\n<p>1- this optimizer only will try to optimize only variables under scope 'aeu_2' , correct?<\/p>\n\n<p>2- if I provide loss to optimizer why ask for feeding PLACEHOLDER?<\/p>\n\n<p>3- how to optimize part of graph  ?<\/p>\n","answers":[]}
+{"QuestionId":57996947,"AnswerCount":0,"Tags":"<python-3.x><opencv><keras>","CreationDate":"2019-09-18T16:14:01.893","AcceptedAnswerId":null,"OwnerUserId":12017671.0,"Title":"Condition to check if an image is not in correct orientation and then correct it's form before processing?","Body":"<p>I am using keras-retinanet model to detect an object from an image. Problem I am facing with this is if image is not orientated correct then model is unable to detect the object. For example if image is like this:<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/9976K.jpg\" alt=\"sample input image\"><\/p>\n\n<p>If I manually flip this image by 180 then model is able to detect the object.<\/p>\n\n<p>I am doing following image processing before sending to model-<\/p>\n\n<pre><code>def _preprocess_image(self, image):\n    draw = image.copy()\n    draw = cv2.cvtColor(draw, cv2.COLOR_BGR2RGB)\n    image = preprocess_image(image)\n    image, scale = resize_image(image)\n    return draw, image, scale\n<\/code><\/pre>\n\n<p>In above method how can I add a check if orientation is not correct and then can correct it?<\/p>\n","answers":[]}
+{"QuestionId":57997201,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><device-driver>","CreationDate":"2019-09-18T16:30:51.820","AcceptedAnswerId":null,"OwnerUserId":11833149.0,"Title":"pixel-cnn (tensorflow-gpu) not recognising GPU","Body":"<p>I'm trying to run the pixel-cnn neural network available on <a href=\"https:\/\/github.com\/openai\/pixel-cnn\" rel=\"nofollow noreferrer\">github<\/a>. Following the instructions in README.md I run the following code in cmd:<\/p>\n\n<pre><code>train.py -i .\/data_dir\/ -o .\/save_dir -g 1\n<\/code><\/pre>\n\n<p>I'm using one gpu and created the two folders .\/data_dir and .\/save_dir within the same directory as train.py for loading &amp; saving the data. When doing so I get the following error message:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: Cannot assign a device for operation model_1\/ones: node model_1\/ones (defined at \\OneDrive - MNG\\Matura Arbeit\\Projects\\pixel-cnn-master\\pixel_cnn_pp\\model.py:36) was explicitly assigned to \/device:GPU:0 but available devices are [ \/job:localhost\/replica:0\/task:0\/device:CPU:0 ]. Make sure the device specification refers to a valid device. The requested device appears to be a GPU, but CUDA is not enabled.\n<\/code><\/pre>\n\n<p>It seems that the tensorflow doesn't recognise the GPU but when checking the devices available to tensorflow (as described <a href=\"https:\/\/stackoverflow.com\/questions\/38559755\/how-to-get-current-available-gpus-in-tensorflow\">here<\/a>) both my CPU and GPU show up as \"\/device:CPU:0\" and \/device:GPU:0\". Also, when running other programs with tensorflow-gpu it work perfectly fine.<\/p>\n\n<p>I have installed <em>tensorflow-gpu==1.14.0<\/em>. As for the CUDA I'm pretty sure I have installed version 10.0, as shown by <code>nvcc --version<\/code>. Although when running <code>nvidia-smi<\/code> it shows that <em>CUDA version 10.1<\/em> is installed.<\/p>\n\n<p>(edited:)I am using an Anaconda evironment (Windows 10) with <em>tensorflow-gpu==1.14.0<\/em>. The GPU I'm using is a <em>GTX 1050Ti<\/em> with Max-Q Design and <em>driver version 436.30<\/em>. As for CUDA I'm pretty sure I have installed <em>version 10.0<\/em>, as shown by <code>nvcc --version<\/code>. Although when running <code>nvidia-smi<\/code> it shows that <em>CUDA version 10.1<\/em> is installed.<\/p>\n","answers":[]}
+{"QuestionId":57997369,"AnswerCount":1,"Tags":"<tensorflow><eigen>","CreationDate":"2019-09-18T16:43:04.567","AcceptedAnswerId":null,"OwnerUserId":5200092.0,"Title":"How to convert a uint8 to Eigen:half?","Body":"<p>I am building a c++ op in tensorflow. It requires to convert a <code>uint8<\/code> value to <code>Eigen::half<\/code> ? However, I did not find a function that does that. But I do notice there is a <code>half_to_float()<\/code>. Is there a similar function <code>uint8_to_float()<\/code> in Eigen library ?  <\/p>\n","answers":[{"AnswerId":"58005638","CreationDate":"2019-09-19T07:15:44.603","ParentId":null,"OwnerUserId":"6870253","Title":null,"Body":"<p>You can directly cast from <code>uint8_t<\/code> to <code>Eigen::half<\/code>:<\/p>\n\n<pre><code>Eigen::half foo(uint8_t x)\n{\n    return Eigen::half(x);\n}\n<\/code><\/pre>\n\n<p>The conversion is likely not very efficient, but it should work: <a href=\"https:\/\/godbolt.org\/z\/Wc_j3X\" rel=\"nofollow noreferrer\">https:\/\/godbolt.org\/z\/Wc_j3X<\/a><\/p>\n\n<p>If you need this a lot, you could think about implementing a small look-up table.<\/p>\n"}]}
+{"QuestionId":57998233,"AnswerCount":1,"Tags":"<python-3.x><numpy><anaconda><conda><torch>","CreationDate":"2019-09-18T17:44:42.040","AcceptedAnswerId":null,"OwnerUserId":12086551.0,"Title":"Error with packages that cause OSError: [WinError 193] %1 is not a valid Win32 application","Body":"<p>I'm trying to run the program main file however when installing torch and cuda there seems to be an error that pops up: \"<code>OSError: [WinError 193] %1 is not a valid Win32 application<\/code>\" <\/p>\n\n<p>I have already downgraded torch, torchvision and numpy<\/p>\n\n<pre><code>runfile('\/\/nsq025vs\/u10\/aczc973\/Desktop\/MScProjectConvert\/main.py', wdir='\/\/nsq025vs\/u10\/aczc973\/Desktop\/MScProjectConvert')\nTraceback (most recent call last):\n\n  File \"&lt;ipython-input-1-e4181834ff1d&gt;\", line 1, in &lt;module&gt;\n    runfile('\/\/nsq025vs\/u10\/aczc973\/Desktop\/MScProjectConvert\/main.py', wdir='\/\/nsq025vs\/u10\/aczc973\/Desktop\/MScProjectConvert')\n\n  File \"C:\\Anaconda3\\lib\\site-packages\\spyder_kernels\\customize\\spydercustomize.py\", line 827, in runfile\n    execfile(filename, namespace)\n\n  File \"C:\\Anaconda3\\lib\\site-packages\\spyder_kernels\\customize\\spydercustomize.py\", line 110, in execfile\n    exec(compile(f.read(), filename, 'exec'), namespace)\n\n  File \"\/\/nsq025vs\/u10\/aczc973\/Desktop\/MScProjectConvert\/main.py\", line 4, in &lt;module&gt;\n    import torch\n\n  File \"C:\\Anaconda3\\lib\\site-packages\\torch\\__init__.py\", line 42, in &lt;module&gt;\n    import numpy as _np\n\n  File \"\\\\nsq025ps\\p10\\aczc973\\AppData\\Python\\Python37\\site-packages\\numpy\\__init__.py\", line 140, in &lt;module&gt;\n    from . import _distributor_init\n\n  File \"\\\\nsq025ps\\p10\\aczc973\\AppData\\Python\\Python37\\site-packages\\numpy\\_distributor_init.py\", line 26, in &lt;module&gt;\n    WinDLL(os.path.abspath(filename))\n\n  File \"C:\\Anaconda3\\lib\\ctypes\\__init__.py\", line 356, in __init__\n    self._handle = _dlopen(self._name, mode)\n\nOSError: [WinError 193] %1 is not a valid Win32 application\n<\/code><\/pre>\n","answers":[{"AnswerId":"58011460","CreationDate":"2019-09-19T12:52:37.383","ParentId":null,"OwnerUserId":"5012099","Title":null,"Body":"<p>There is a mix between differnt python installations: <code>C:\\Anaconda3\\lib\\site-packages<\/code> and <code>\\\\nsq025ps\\p10\\aczc973\\AppData\\Python\\Python37\\site-packages<\/code> which probably have a different 32\/64bit setup. Make sure that your PYTHONPATH does not include the reference to <code>\\\\nsq025ps\\p10\\aczc973\\AppData\\Python\\Python37\\site-packages<\/code> when you are running stuff with your <code>anaconda<\/code> distribution and vice versa.<\/p>\n\n<p>Even safer: Completely uninstall <code>\\\\nsq025ps\\p10\\aczc973\\AppData\\Python\\Python37<\/code>, you should not need a python installation besides your <code>anaconda<\/code><\/p>\n"}]}
+{"QuestionId":57998816,"AnswerCount":2,"Tags":"<python><numpy><pytorch>","CreationDate":"2019-09-18T18:25:23.707","AcceptedAnswerId":58004525.0,"OwnerUserId":2612484.0,"Title":"How to exponentially decay a value along an axis?","Body":"<p>I have a tensor of shape <code>(sequence_length,batch_size,1)<\/code> (which is <code>(5,5,1)<\/code> in the following example). Every sequence has a single entry that I want to exponentially decay along the sequence (for example with a decay factor of <code>0.99<\/code>) for <code>n<\/code> steps (here: <code>n = 2<\/code>). <\/p>\n\n<p><strong>Example<\/strong>:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Rm6P1.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Rm6P1.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Is there an efficient way to realize something like that (either in <code>PyTorch<\/code> or <code>numpy<\/code>).<\/p>\n\n<p>Edit (visualizing the additive nature):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/CAHfs.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/CAHfs.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"58000959","CreationDate":"2019-09-18T21:30:03.290","ParentId":null,"OwnerUserId":"3666197","Title":null,"Body":"<h2><sup>Given the last comments,<\/sup> the <strong><code>numpy<\/code><\/strong>-based problem solution is doable:<\/h2>\n\n<p>While there may be smarter <code>numpy<\/code>-vectorised trics doable, the non-interfering adding of many 1D-\"diffusion\"-kernels over a <strong><code>.view<\/code><\/strong>-based strided-tricks data seems uncertain ATM not to introduce unwanted side-effects of skewed-superposition(s) of both original plus kernel-added values from re-propagation of superposed values instead of 1D-\"diffusion\"-kernel propagations of just the original tensor values ( before any kernel-added products get added as these may reach the same tensor-cell location )<\/p>\n\n<hr>\n\n<p><strong>- step 0:<\/strong> create a constant, Q-learning gamma-based \"kernel\"-vector <strong><code>kernel<\/code><\/strong> and a zeroed-copy of the <code>ORIGINAL_TENSOR<\/code> in <code>RESULT_TENSOR<\/code><\/p>\n\n<hr>\n\n<pre><code>&gt;&gt;&gt; N_STEPS  = 2\n&gt;&gt;&gt; Q_FACTOR = 0.99\n&gt;&gt;&gt; kernel = np.ones( N_STEPS )\n&gt;&gt;&gt; for i in range( 1, kernel.shape[0] ):\n...     kernel[:i] *= Q_FACTOR\n... \n&gt;&gt;&gt; kernel\narray([0.9801, 0.99  , 1.    ])\n&gt;&gt;&gt; RESULT_TENSOR = np.zeros( ORIGINAL_TENSOR.shape )\n<\/code><\/pre>\n\n<p>In case of <code>[SPACE]<\/code>-domain related problem, the interim adder storage may keep way smaller, using just a current-column-wise mapping, so as to minimise the overall RAM-footprint.<\/p>\n\n<hr>\n\n<p><strong>- step 1:<\/strong> for each tensor-column, use the <code>numpy<\/code> striding-tricks for forming a 2D-meta-view of each of the such tensor-column, that has depth equal to n-steps ( for reversed-propagated q-rewards ) - like was used in this <code>rolling_window()<\/code> example for 2D-matrix<sub> (needs adapting for using over a vector-column or if going a bit wilder, going here straight across the while original tensor)<\/sub>:<\/p>\n\n<hr>\n\n<pre><code>def rolling_window( aMatrix, aRollingWindowLENGTH ):                    #\n    \"\"\"                                                                 __doc__\n    USAGE:   rolling_window( aMatrix, aRollingWindowLENGTH )\n\n    PARAMS:  aMatrix                a numpy array\n             aRollingWindowLENGTH   a LENGTH of a Rolling Window\n\n    RETURNS: a stride_trick'ed numpy array with rolling windows\n\n    THROWS:  n\/a\n\n    EXAMPLE: &gt;&gt;&gt; x = np.arange( 10 ).reshape( ( 2, 5 ) )\n\n             &gt;&gt;&gt; rolling_window( x, 3 )\n             array([[[0, 1, 2], [1, 2, 3], [2, 3, 4]],\n                    [[5, 6, 7], [6, 7, 8], [7, 8, 9]]])\n\n             &gt;&gt;&gt; np.mean( rolling_window( x, 3 ), -1 )\n             array([[ 1.,  2.,  3.],\n                    [ 6.,  7.,  8.]])\n    \"\"\"\n    new_shape   = aMatrix.shape[:-1] + ( aMatrix.shape[-1]\n                                       - aRollingWindowLENGTH + 1,\n                                         aRollingWindowLENGTH\n                                         )\n    new_strides = aMatrix.strides    + ( aMatrix.strides[-1], )\n    return np.lib.stride_tricks.as_strided( aMatrix,\n                                            shape   = new_shape,\n                                            strides = new_strides\n                                            )\n\nmeta2D = rolling_window( aTensorCOLUMN_to_PROCESS, N_STEPS )\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>- step 2:<\/strong> for the case of column-wise tensor processing, multiply the <code>kernel<\/code> by a scalar - by each current column's 2D-meta-view matrix \"rows'\"[-1] value and <strong><code>+=<\/code> store<\/strong> such produced \"scaled\"-kernel into a separate, <code>RESULT_TENSOR<\/code> (or the above explained <code>[SPACE]<\/code>-domain motivated re-useable\/re-erased interim column-adder ) accumulating additive results ( IMHO this cannot be in-place, as the kernel-\"back-propagated\" rewards must not \"increase\" the other (though additive in result) Q-rewards, not yet processed by the <code>kernel<\/code>, as the strided-tricks are just a view and storing in-place would interfere with the not yet processed sections of the <code>ORIGINAL_TENSOR<\/code>.<\/p>\n\n<hr>\n\n<p><strong>- step 3:<\/strong> you are done, use the <code>RESULT_TENSOR<\/code> accumulated result, as it now holds all the Q-learning n-steps deep propagations of all non-zero values summed up with any cross-interfering side-effects avoided.<\/p>\n"},{"AnswerId":"58004525","CreationDate":"2019-09-19T05:56:04.430","ParentId":null,"OwnerUserId":"7207392","Title":null,"Body":"<p>This looks like convolution to me:<\/p>\n\n<pre><code>import numpy as np\nfrom scipy.ndimage import convolve1d\n\n# just for printing:\nimport numpy.lib.recfunctions as nlr\ndef show(x): print(nlr.unstructured_to_structured(x))\n\na = \"00000\",\"00100\",\"10000\",\"03502\",\"00010\"\nA = np.array(a).view('U1').reshape(5,5,1).astype(float)\n\nshow(A)\n# [[(0.,) (0.,) (0.,) (0.,) (0.,)]\n#  [(0.,) (0.,) (1.,) (0.,) (0.,)]\n#  [(1.,) (0.,) (0.,) (0.,) (0.,)]\n#  [(0.,) (3.,) (5.,) (0.,) (2.,)]\n#  [(0.,) (0.,) (0.,) (1.,) (0.,)]]\n\nshow(convolve1d(A,np.r_[np.logspace(2,0,3,base=0.99),0,0],axis=0,mode=\"constant\"))\n# [[(0.9801,) (0.    ,) (0.99  ,) (0.    ,) (0.    ,)]\n#  [(0.99  ,) (2.9403,) (5.9005,) (0.    ,) (1.9602,)]\n#  [(1.    ,) (2.97  ,) (4.95  ,) (0.9801,) (1.98  ,)]\n#  [(0.    ,) (3.    ,) (5.    ,) (0.99  ,) (2.    ,)]\n#  [(0.    ,) (0.    ,) (0.    ,) (1.    ,) (0.    ,)]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":57998859,"AnswerCount":1,"Tags":"<python><tensorflow><sparse-matrix><tensorflow2.0>","CreationDate":"2019-09-18T18:28:51.140","AcceptedAnswerId":57999172.0,"OwnerUserId":7519842.0,"Title":"Efficient boolean masking with Tensorflow SparseTensors","Body":"<p>So, I want to mask out entire rows of a <code>SparseTensor<\/code>. This would be easy to do with <code>tf.boolean_mask<\/code>, but there isn't an equivalent for <code>SparseTensor<\/code>s. Currently, something that is possible is for me to just go through all of the indices in <code>SparseTensor.indices<\/code> and filter out all of the ones that aren't a masked row, e.g.:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>masked_indices = list(filter(lambda index: masked_rows[index[0]], indices))\n<\/code><\/pre>\n\n<p>where masked_rows is a 1D array of whether or not the row at that index is masked.<\/p>\n\n<p>However, this is really slow, since my SparseTensor is fairly large (it has 90k indices, but will be growing to be significantly larger). It takes quite a few seconds on a single data point, before I even apply <code>SparseTensor.mask<\/code> on the filtered indices. Another flaw of the approach is that it doesn't actually remove the rows, either (although, in my case, a row of all zeros is just as fine).<\/p>\n\n<p>Is there a better way to mask a SparseTensor by row, or is this the best approach? <\/p>\n","answers":[{"AnswerId":"57999172","CreationDate":"2019-09-18T18:52:12.960","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You can do that like this:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef boolean_mask_sparse_1d(sparse_tensor, mask, axis=0):  # mask is assumed to be 1D\n    mask = tf.convert_to_tensor(mask)\n    ind = sparse_tensor.indices[:, axis]\n    mask_sp = tf.gather(mask, ind)\n    new_size = tf.math.count_nonzero(mask)\n    new_shape = tf.concat([sparse_tensor.shape[:axis], [new_size],\n                           sparse_tensor.shape[axis + 1:]], axis=0)\n    new_shape = tf.dtypes.cast(new_shape, tf.int64)\n    mask_count = tf.cumsum(tf.dtypes.cast(mask, tf.int64), exclusive=True)\n    masked_idx = tf.boolean_mask(sparse_tensor.indices, mask_sp)\n    new_idx_axis = tf.gather(mask_count, masked_idx[:, axis])\n    new_idx = tf.concat([masked_idx[:, :axis],\n                         tf.expand_dims(new_idx_axis, 1),\n                         masked_idx[:, axis + 1:]], axis=1)\n    new_values = tf.boolean_mask(sparse_tensor.values, mask_sp)\n    return tf.SparseTensor(new_idx, new_values, new_shape)\n\n# Test\nsp = tf.SparseTensor([[1], [3], [4], [6]], [1, 2, 3, 4], [7])\nmask = tf.constant([True, False, True, True, False, False, True])\nout = boolean_mask_sparse_1d(sp, mask)\nprint(out.indices.numpy())\n# [[2]\n#  [3]]\nprint(out.values.numpy())\n# [2 4]\nprint(out.shape)\n# (4,)\n<\/code><\/pre>\n"}]}
+{"QuestionId":57999090,"AnswerCount":0,"Tags":"<python><tensorflow><memory>","CreationDate":"2019-09-18T18:45:19.713","AcceptedAnswerId":null,"OwnerUserId":12086720.0,"Title":"How to release part of the GPU memory which is no more in used in a Tensorflow session?","Body":"<p>In some of my application, I evaluate the gradient with a large amount of data. After getting a summed average of the gradient, data used to evaluate that part of gradient is no longer in used and the corresponding GPU memory should be freed and can be reused within the same session. However, I didn't find any methods that can manage memory in this way. I Will highly appreciate it if somebody can help with this.  <\/p>\n\n<p>I know we can free the memory if a session is closed. However, since my variables and network parameters should be maintained for further iterations, I still cannot release that part of memory by closing a session.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\n\n\nparam = tf.trainable_variables()\ngrad = tf.gradients(loss, param)\n\nwith tf.session() as sess:\n    sess.run(grad, feed_dict=feed_dict)\n    '''\n    After evaluating the grad, those data in used can be freed since grad is obtained.\n    And I am able to reused that part of memory to do other operations.\n    So I am wondering how can we release the memory within the same session?\n    '''\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57999225,"AnswerCount":2,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-09-18T18:58:00.683","AcceptedAnswerId":58002329.0,"OwnerUserId":567620.0,"Title":"How to access sample weights in a Keras custom loss function supplied by a generator?","Body":"<p>I have a generator function that infinitely cycles over some directories of images and outputs 3-tuples of batches the form<\/p>\n\n<pre><code>[img1, img2], label, weight\n<\/code><\/pre>\n\n<p>where <code>img1<\/code> and <code>img2<\/code> are <code>batch_size x M x N x 3<\/code> tensors, and <code>label<\/code> and <code>weight<\/code> are each <code>batch_size<\/code> x 1 tensors.<\/p>\n\n<p>I provide this generator to the <code>fit_generator<\/code> function when training a model with Keras.<\/p>\n\n<p>For this model I have a custom cosine contrastive loss function,<\/p>\n\n<pre><code>def cosine_constrastive_loss(y_true, y_pred):\n    cosine_distance = 1 - y_pred\n    margin = 0.9\n    cdist = y_true * y_pred + (1 - y_true) * keras.backend.maximum(margin - y_pred, 0.0)\n    return keras.backend.mean(cdist)\n<\/code><\/pre>\n\n<p>Structurally everything runs OK with my model. There are no errors and it is consuming the inputs and labels from the generator as expected.<\/p>\n\n<p>But now I am seeking to directly use the weights parameter per each batch and perform some customized logic inside of <code>cosine_contrastive_loss<\/code> based on the sample-specific weight.<\/p>\n\n<p>How can I access this parameter from the structure of a batch of samples at the moment of the loss function being executed?<\/p>\n\n<p>Note that since it is an infinitely cycling generator, it is not possible to precompute weights or compute them on the fly to either curry the weights into the loss function or generate them. <\/p>\n\n<p>They have to be generated in unison with the samples being generated, and indeed there is custom logic in my data generator that determines the weights dynamically from properties of <code>img1<\/code>, <code>img2<\/code> and <code>label<\/code> at the moment they are generated for a batch. <\/p>\n","answers":[{"AnswerId":"58599783","CreationDate":"2019-10-28T23:46:42.347","ParentId":null,"OwnerUserId":"4867484","Title":null,"Body":"<p>The loss function in Keras Tensorflow v2 is called with the sample weighs <\/p>\n\n<pre><code>output_loss = loss_fn(y_true, y_pred, sample_weight=sample_weight)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/tf-2\/keras\/engine\/training.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/blob\/tf-2\/keras\/engine\/training.py<\/a><\/p>\n\n<hr>\n\n<p>You can use GradientTape for custom training, see <a href=\"https:\/\/www.tensorflow.org\/guide\/keras\/train_and_evaluate#part_ii_writing_your_own_training_evaluation_loops_from_scratch\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/keras\/train_and_evaluate#part_ii_writing_your_own_training_evaluation_loops_from_scratch<\/a><\/p>\n"},{"AnswerId":"58002329","CreationDate":"2019-09-19T00:51:00.117","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<h2>Manual training loop alternative<\/h2>\n\n<p>The only thing I can think of is a manual training loop where you get the weights yourself.<\/p>\n\n<p>Have a weights tensor and a non variable batch size:<\/p>\n\n<pre><code>weights = K.variable(np.zeros((batch_size,)))\n<\/code><\/pre>\n\n<p>Use them in your custom loss:<\/p>\n\n<pre><code>def custom_loss(true, pred):\n    return someCalculation(true, pred, weights)\n<\/code><\/pre>\n\n<p>For a \"generator\":<\/p>\n\n<pre><code>for e in range(epochs):\n    for s in range(steps_per_epoch):\n        x, y, w = next(generator) #or generator.next(), not sure\n        K.set_value(weights, w)\n\n        model.train_on_batch(x, y)\n<\/code><\/pre>\n\n<p>For a <code>keras.utils.Sequence<\/code>:<\/p>\n\n<pre><code>for e in range(epochs):\n    for s in range(len(generator)):\n        x,y,w = generator[s]\n\n        K.set_value(weights, w)\n        model.train_on_batch(x,y)\n<\/code><\/pre>\n\n<p>I know this answer is not optimal because it does not parallelize getting data from the generator as it happens with <code>fit_generator<\/code>. But it's the best easy solution I can think of. Keras didn't expose the weights, they are applied automatically in some hidden source code. <\/p>\n\n<hr>\n\n<h1>Let the model calculate the weights alternative<\/h1>\n\n<p>If calculating the weights can be done from <code>x<\/code> and <code>y<\/code>, you can delegate this task to the loss function itself.<\/p>\n\n<p>This is sort of hacky, but may work:<\/p>\n\n<pre><code>input1 = Input(shape1)\ninput2 = Input(shape2)\n\n# .... model creation .... #\n\nmodel = Model([input1, input2], outputs)\n<\/code><\/pre>\n\n<p>Let the loss have access to <code>input1<\/code> and <code>input2<\/code>:<\/p>\n\n<pre><code>def custom_loss(y_true, y_pred):\n    w = calculate_weights(input1, input2, y_pred)\n    # .... rest of the loss .... #\n<\/code><\/pre>\n\n<p>The issue here is whether you can or not calculate the weigths as a tensor from the inputs. <\/p>\n"}]}
+{"QuestionId":57999582,"AnswerCount":1,"Tags":"<c++><linux><ubuntu><caffe><ld>","CreationDate":"2019-09-18T19:26:34.843","AcceptedAnswerId":57999628.0,"OwnerUserId":3204706.0,"Title":"\/usr\/bin\/ld: cannot find -lcaffe","Body":"<p>I installed Caffe-cpu on my Ubuntu 18.04 via the apt-get command, as it instructs on their <a href=\"https:\/\/caffe.berkeleyvision.org\/install_apt.html\" rel=\"nofollow noreferrer\">official website<\/a>:<\/p>\n\n<pre><code>sudo apt install caffe-cpu\n<\/code><\/pre>\n\n<p>But when trying to run a sample from a <a href=\"https:\/\/github.com\/Pratik08\/Vis-DSS\" rel=\"nofollow noreferrer\">git repo<\/a>, it needs the header files (<code>caffe.hpp<\/code>). For that, I downloaded the latest Caffe Git repo and pointed to the sources. But it also complains at the make time:<\/p>\n\n<pre><code>\/usr\/bin\/ld: cannot find -lcaffe\n<\/code><\/pre>\n\n<p>My understanding of similar problems is that apparently it needs to link to <code>caffe.so<\/code>, but I can't find it anywhere on my machine. What's the problem? How to fix it?<\/p>\n","answers":[{"AnswerId":"57999628","CreationDate":"2019-09-18T19:29:31.773","ParentId":null,"OwnerUserId":"6277384","Title":null,"Body":"<pre><code>sudo apt install libcaffe-cpu-dev\n<\/code><\/pre>\n\n<p>It's widely known that to use some library on Ubuntu, we need to install <code>libxxx-dev<\/code> package.<\/p>\n"}]}
+{"QuestionId":57999776,"AnswerCount":1,"Tags":"<tensorflow><gpu><distributed-computing><google-colaboratory>","CreationDate":"2019-09-18T19:42:51.083","AcceptedAnswerId":null,"OwnerUserId":12086875.0,"Title":"Distributed training in Tensorflow using multiple GPUs in Google Colab","Body":"<p>I have recently become interested in incorporating distributed training into my Tensorflow projects. I am using Google Colab and Python 3 to implement a Neural Network with customized, distributed, training loops, as described in this guide: \n<a href=\"https:\/\/www.tensorflow.org\/tutorials\/distribute\/training_loops\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/distribute\/training_loops<\/a><\/p>\n\n<p>In that guide under section 'Create a strategy to distribute the variables and the graph', there is a picture of some code that basically sets up a 'MirroredStrategy' and then prints the number of generated replicas of the model, see below.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/hrIxb.png\" rel=\"nofollow noreferrer\">Console output<\/a><\/p>\n\n<p>From what I can understand, the output indicates that the MirroredStrategy has only created one replica of the model, and thereofore, only one GPU will be used to train the model. My question: is Google Colab limited to training on a single GPU? <\/p>\n\n<p>I have tried to call MirroredStrategy() both with, and without, GPU acceleration, but I only get one model replica every time. This is a bit surprising because when I use the multiprocessing package in Python, I get four threads. I therefore expected that it would be possible to train four models in parallel in Google Colab. Are there issues with Tensorflows implementation of distributed training?<\/p>\n","answers":[{"AnswerId":"58032694","CreationDate":"2019-09-20T17:16:43.253","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>On google colab, you can only use one GPU, that is the limit from Google. However, you can run different programs on different gpu instances so by creating different colab files and connect them with gpus but you can not place the same model on many gpu instances in parallel. \nThere are no problems with mirrored startegy, talking from personal experience it works fine if you have more than one GPU. <\/p>\n"}]}
+{"QuestionId":57999830,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning>","CreationDate":"2019-09-18T19:47:13.060","AcceptedAnswerId":57999901.0,"OwnerUserId":7789594.0,"Title":"Non-uniform penalty in loss function calculation using Keras","Body":"<p>Is there a simple approach to use have non-uniform penalty while calculating the loss in Keras (or Tensorflow)? \nI mean, assume you have 5000 training samples and they are not of the same importance and we prefer to get more accurate prediction in some cases even at the expense of inaccurate estimation for several other samples. I already have some coefficients for samples representing the importance of them. <\/p>\n\n<p>Is there any way to implement such models on Keras or Tensorflow? <\/p>\n","answers":[{"AnswerId":"57999901","CreationDate":"2019-09-18T19:53:46.900","ParentId":null,"OwnerUserId":"1729802","Title":null,"Body":"<p>Yes there is a way is to use <code>sample_weight<\/code> in the fit or <code>fit_generator<\/code> function of keras. This will apply the weight per sample. Therefore, the sample_weight vector should have same length as your training set.<\/p>\n\n<p>Here is sample code:<\/p>\n\n<pre><code>history = model.fit(X_train, y_train, epochs=1000, sample_weight=sample_coefficients)\n<\/code><\/pre>\n\n<p>From keras <a href=\"https:\/\/keras.io\/models\/model\/\" rel=\"nofollow noreferrer\">documentation<\/a>:<\/p>\n\n<blockquote>\n  <p>sample_weight: 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"}]}
+{"QuestionId":57999861,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-18T19:50:35.787","AcceptedAnswerId":null,"OwnerUserId":3614179.0,"Title":"How to print tensorflow tensor value in this example?","Body":"<p>I'm new to tensorflow and basically I copied the example somewhere but cannot compile it.<\/p>\n\n<pre><code>import  tensorflow as tf\nfrom    tensorflow.keras import datasets, layers, optimizers, Sequential, metrics\n\n(xs, ys),_ = datasets.mnist.load_data()\nprint('datasets:', xs.shape, ys.shape, xs.min(), xs.max())\n\nxs = tf.convert_to_tensor(xs, dtype=tf.float32) \/ 255.\ndb = tf.data.Dataset.from_tensor_slices((xs,ys))\ndb = db.batch(32).repeat(10)\n\nnetwork = Sequential([layers.Dense(256, activation='relu'),\n    layers.Dense(256, activation='relu'),\n    layers.Dense(256, activation='relu'),\n    layers.Dense(10)])\nnetwork.build(input_shape=(None, 28*28))\nnetwork.summary()\n\noptimizer = optimizers.SGD(lr=0.01)\nacc_meter = metrics.Accuracy()\n\nfor step, (x,y) in enumerate(db):\n\n    with tf.GradientTape() as tape:\n        x = tf.reshape(x, (-1, 28*28))\n        out = network(x)\n\n        y_onehot = tf.one_hot(y, depth=10)\n        loss = tf.square(out-y_onehot)\n        loss = tf.reduce_sum(loss) \/ 32\n        acc_meter.update_state(tf.argmax(out, axis=1), y)\n        grads = tape.gradient(loss, network.trainable_variables)\n        optimizer.apply_gradients(zip(grads, network.trainable_variables))\n\n        if step % 200==0:\n            print(float(loss))\n            exit()\n<\/code><\/pre>\n\n<p>This gives the error :<\/p>\n\n<pre><code>TypeError: float() argument must be a string or a number, not 'Tensor'\n<\/code><\/pre>\n\n<p>on the second-last line.<\/p>\n\n<p>However I've tried <code>loss.eval()<\/code>, which says <code>No default session is registered.<\/code> But if I write <\/p>\n\n<pre><code>tf.Session() as sess:\n    print(sess.run(loss))\n<\/code><\/pre>\n\n<p>it leads to some really complicated error. \nIf I write <code>print(loss.numpy())<\/code>, it says <code>AttributeError: 'Tensor' object has no attribute 'numpy'<\/code><\/p>\n\n<p>Everything solution I searched on the internet required the code to have a <code>tf.Session()<\/code> running, which in this example there isn't. How do I print out the value of the <code>loss<\/code> variable?<\/p>\n","answers":[{"AnswerId":"58000015","CreationDate":"2019-09-18T20:04:34.793","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>If you're using tf-1.x, you should put <code>tf.enable_eager_execution()<\/code> at first. I only add this line and the code works.<\/p>\n"}]}
+{"QuestionId":57999882,"AnswerCount":0,"Tags":"<keras><keras-layer>","CreationDate":"2019-09-18T19:52:29.120","AcceptedAnswerId":null,"OwnerUserId":6499614.0,"Title":"Keras: unable to extract layer output while pushing layers back to model","Body":"<p>I have a Keras classification network with last two layers as Dense and activation respectively. I popped the layers from the model as :<\/p>\n\n<pre><code>activationlayer=classifier.layers.pop()\ndenselayer=classifier.layers.pop()\n<\/code><\/pre>\n\n<p>Then I pushed the layers back into the model:<\/p>\n\n<pre><code>outt1=denselayer(classifier.layers[-1].output)\noutt=activationlayer(outt1)\nclassifier=Model(classifier.input,outt)\n<\/code><\/pre>\n\n<p>Now while indexing the output of dense layer I get error:<\/p>\n\n<pre><code>classifier.layers[-2].output\n\nError: Layer out2 has multiple inbound nodes, hence the notion of \"layer output\" is ill-defined.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":57999985,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-18T20:01:31.060","AcceptedAnswerId":null,"OwnerUserId":564653.0,"Title":"Keras when I try to train my model I get a \" <no registered kernels>\" error","Body":"<p>when I try to create a model, it crashes when I try to train it \ncode to create model, then train it<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(CuDNNLSTM(128, input_shape=(train_x.shape[1:]), return_sequences=True))\nmodel.add(Dropout(0.2))\nmodel.add(BatchNormalization())\n\nmodel.add(CuDNNLSTM(128, return_sequences=True))\nmodel.add(Dropout(0.1))\nmodel.add(BatchNormalization())\n\nmodel.add(CuDNNLSTM(128))\nmodel.add(Dropout(0.2))\nmodel.add(BatchNormalization())\n\nmodel.add(Dense(32, activation='relu'))\nmodel.add(Dropout(0.2))\n\nmodel.add(Dense(2, activation='softmax'))\n\n\nopt = tf.keras.optimizers.Adam(lr=0.001, decay=1e-6)\n\n# Compile model\nmodel.compile(\n    loss='sparse_categorical_crossentropy',\n    optimizer=opt,\n    metrics=['accuracy'])\n\n# program crashes hear\nhistory = model.fit(\n    train_x, train_y,\n    batch_size=BATCH_SIZE,\n    epochs=EPOCHS,\n    validation_data=(validation_x, validation_y),\n    callbacks=[tensorboard, checkpoint] )\n<\/code><\/pre>\n\n<p>I get the following error<\/p>\n\n<blockquote>\n  <p>[seed=0, dropout=0, T=DT_FLOAT, input_mode=\"linear_input\",\n  direction=\"unidirectional\", rnn_mode=\"lstm\", is_training=true,\n  seed2=0] Registered devices: [CPU] Registered kernels:   <\/p>\n<\/blockquote>\n\n<p>complete stack trace<\/p>\n\n<pre><code>    K.batch_set_value([(v, 0) for v in self.variables])\n  File \"C:\\python37-32\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py\", line 3071, in batch_set_value\n    get_session().run(assign_ops, feed_dict=feed_dict)\n  File \"C:\\python37-32\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py\", line 462, in get_session\n    _initialize_variables(session)\n  File \"C:\\python37-32\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py\", line 879, in _initialize_variables\n    [variables_module.is_variable_initialized(v) for v in candidate_vars])\n  File \"C:\\python37-32\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 950, in run\n    run_metadata_ptr)\n  File \"C:\\python37-32\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1173, in _run\n    feed_dict_tensor, options, run_metadata)\n  File \"C:\\python37-32\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1350, in _do_run\n    run_metadata)\n  File \"C:\\python37-32\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1370, in _do_call\n    raise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: No OpKernel was registered to support Op 'CudnnRNN' used by node cu_dnnlstm\/CudnnRNN (defined at \/Users\/tpottel\/programming\/java\/eclipse projects\/BBitBotCoinUsing2\/pythonproject\/deeplearning\/lstm-bitcoin.py:128) with these attrs: [seed=0, dropout=0, T=DT_FLOAT, input_mode=\"linear_input\", direction=\"unidirectional\", rnn_mode=\"lstm\", is_training=true, seed2=0]\nRegistered devices: [CPU]\nRegistered kernels:\n  &lt;no registered kernels&gt;\n\\\n<\/code><\/pre>\n","answers":[{"AnswerId":"58013995","CreationDate":"2019-09-19T15:05:19.050","ParentId":null,"OwnerUserId":"2464597","Title":null,"Body":"<p>Taken from my comment, since that seemed to help: it looks like you don't have <a href=\"https:\/\/en.wikipedia.org\/wiki\/CUDA\" rel=\"nofollow noreferrer\">CUDA<\/a>\/<a href=\"https:\/\/developer.nvidia.com\/cudnn\" rel=\"nofollow noreferrer\">CuDNN<\/a> set up on your system, or possibly no GPU. You could run with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/LSTM\" rel=\"nofollow noreferrer\"><code>LSTM<\/code><\/a> layers instead of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/CuDNNLSTM\" rel=\"nofollow noreferrer\"><code>CuDNNLSTM<\/code><\/a>, but if the GPU acceleration is important you are going to need to configure your GPU. That's on the OS side, not Keras or Python.<\/p>\n\n<p>To use <code>LSTM<\/code> layers:<\/p>\n\n<pre><code>from keras.layers import LSTM\n\nmodel = Sequential()\nmodel.add(LSTM(128, input_shape=(train_x.shape[1:]), return_sequences=True))\nmodel.add(Dropout(0.2))\n# ...\n<\/code><\/pre>\n\n<p>If you have a GPU and you want to use it, try installing <code>tensorflow-gpu<\/code> with pip and seeing if you can't get GPU acceleration working. That's kind of a whole thing, so I encourage you to do some research online and then come back to SO when you have specific questions.<\/p>\n"}]}
+{"QuestionId":58000160,"AnswerCount":1,"Tags":"<numpy><random><pytorch><random-seed>","CreationDate":"2019-09-18T20:15:58.620","AcceptedAnswerId":58001583.0,"OwnerUserId":781938.0,"Title":"Does PyTorch have a RandomState-like object for random number generation?","Body":"<p>in numpy i can<\/p>\n\n<pre><code>import numpy as np\n\nrs = np.random.RandomState(seed=0)\n<\/code><\/pre>\n\n<p>and then pass that object around, eg for dependency injection.<\/p>\n\n<p>Does PyTorch have a similar interface? I can't find anything in the docs, but maybe i'm missing something.<\/p>\n","answers":[{"AnswerId":"58001583","CreationDate":"2019-09-18T22:40:43.680","ParentId":null,"OwnerUserId":"8112138","Title":null,"Body":"<p>The closest thing would be <a href=\"https:\/\/pytorch.org\/docs\/master\/torch.html?highlight=manual_seed#torch.manual_seed\" rel=\"nofollow noreferrer\"><code>torch.manual_seed<\/code><\/a>, which sets the seed for generating random numbers and returns a <code>torch.Generator<\/code>. This thread <a href=\"https:\/\/discuss.pytorch.org\/t\/random-seed-initialization\/7854\" rel=\"nofollow noreferrer\">here<\/a> has more information, apparently there may be some inconsistencies depending on whether you are using GPU or a CPU. <\/p>\n"}]}
+{"QuestionId":58000377,"AnswerCount":0,"Tags":"<tensorflow><deprecation-warning><audacity>","CreationDate":"2019-09-18T20:33:16.797","AcceptedAnswerId":null,"OwnerUserId":12087127.0,"Title":"colab Udasity tensor flow warning","Body":"<p>colab is using the code below to split the \"fashion_mnist\" data set in training and testing data, but warning appears and it is not supposed to happen.<\/p>\n\n<p>This is the colab code: <\/p>\n\n<pre><code>dataset, metadata = tfds.load('fashion_mnist', as_supervised=True, with_info=True)\ntrain_dataset, test_dataset = dataset['train'], dataset['test']\n<\/code><\/pre>\n\n<p>Below is the warning message, I already tried by changing the values in True to False for each one, but it did not help.<\/p>\n\n<pre><code>WARNING: absl:Warning: Setting shuffle_files=True because split=TRAIN and shuffle_files=None. This behavior will be deprecated on 2019-08-06, at which point shuffle_files=False will be the default for all splits.\n<\/code><\/pre>\n\n<p>This is the link for the explanation from the tensor flow (tf) course of udacity, which says to just run the colab code and that's it. Apparently they didn't consider that it was going to be ending on 2019-08-06.<\/p>\n","answers":[]}
+{"QuestionId":58000760,"AnswerCount":0,"Tags":"<c++><deep-learning><blob><caffe><caffe2>","CreationDate":"2019-09-18T21:10:15.800","AcceptedAnswerId":null,"OwnerUserId":3204706.0,"Title":"Error running Caffe example: blob.cpp: Check failed: count_ == proto.data_size() (1 vs. 0)","Body":"<p>I was trying to run a sample Caffe code from a <a href=\"https:\/\/github.com\/Pratik08\/Vis-DSS\" rel=\"nofollow noreferrer\">git repo<\/a>.<\/p>\n\n<p>I installed Caffe on my Ubuntu 18.04 with <code>apt-get install caffe-cpu<\/code> and <code>libcaffe-cpu-dev<\/code>, and compiled the code and generated <code>DeepVideoSummExample<\/code> successfully. <\/p>\n\n<p>For running, this executable asks for <code>googlenet_places205<\/code> model files as well as <code>hybridCNN<\/code> mean file, which I downloaded them online. But when running <code>DeepVideoSummExample<\/code> with the following example command as stated in the repo, I get an error right away:<\/p>\n\n<p><code>.\/build\/DeepVideoSummExample -videoFile Sample_NEW.mp4 -videoSaveFile outdeep.avi -summaryModelSim 0 -simcover 0 -segmentType 0 -summaryAlgo 0 -featureLayer loss3\/classifier -network_file Models\/googlenet_places205\/deploy_places205.protxt -trained_file Models\/googlenet_places205\/googlelet_places205_train_iter_2400000.caffemodel -mean_file Models\/hybridCNN\/hybridCNN_mean.binaryproto -label_file Models\/googlenet_places205\/categoryIndex_places205.csv -budget 30<\/code><\/p>\n\n<p>And error (a lot of outputs....but here is the end):<\/p>\n\n<pre><code>F0918 20:11:34.079298 17067 blob.cpp:496] Check failed: count_ == proto.data_size() (1 vs. 0)\n*** Check failure stack trace: ***\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.dropbox.com\/s\/4xbkuyzcoqzc10w\/log.txt?dl=0\" rel=\"nofollow noreferrer\">Here<\/a> are the extended errors. <\/p>\n\n<p>What's the problem? How to fix it?<\/p>\n","answers":[]}
+{"QuestionId":58000830,"AnswerCount":1,"Tags":"<tensorflow><google-colaboratory>","CreationDate":"2019-09-18T21:16:18.120","AcceptedAnswerId":null,"OwnerUserId":12086875.0,"Title":"Does anyone get the Tensorflow distributed training notebook to work?","Body":"<p>When I run this notebook, which is linked from Tensorflows documentation site, <\/p>\n\n<p><a href=\"https:\/\/colab.research.google.com\/github\/tensorflow\/docs\/blob\/r2.0rc\/site\/en\/tutorials\/distribute\/training_loops.ipynb#scrollTo=7x7s5iYAYSGD\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/github\/tensorflow\/docs\/blob\/r2.0rc\/site\/en\/tutorials\/distribute\/training_loops.ipynb#scrollTo=7x7s5iYAYSGD<\/a><\/p>\n\n<p>I get the following error:<\/p>\n\n<blockquote>\n  <p>module 'tensorflow._api.v1.nn' has no attribute 'compute_average_loss'<\/p>\n<\/blockquote>\n\n<p>If anyone can get it to work, what settings are you using in Google Colab? Python 3\/2, GPU, Tensorflow version, etc.<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"58005898","CreationDate":"2019-09-19T07:30:43.907","ParentId":null,"OwnerUserId":"9828197","Title":null,"Body":"<p>The github code is from may this year: <\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/docs\/commits\/r2.0rc\/site\/en\/tutorials\/distribute\/training_loops.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/docs\/commits\/r2.0rc\/site\/en\/tutorials\/distribute\/training_loops.ipynb<\/a><\/p>\n\n<p>while the colab is loading tf 1.14:<\/p>\n\n<pre><code>from __future__ import absolute_import, division, print_function, unicode_literals\n\n# Import TensorFlow\nimport tensorflow as tf\n\n# Helper libraries\nimport numpy as np\nimport os\n\nprint(tf.__version__)\n\n1.14.0\n<\/code><\/pre>\n\n<p>so your best bet is to run the program with tf 2.0.<\/p>\n\n<p>hope it helps :)<\/p>\n"}]}
+{"QuestionId":58001215,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-18T21:54:38.920","AcceptedAnswerId":null,"OwnerUserId":11223907.0,"Title":"Pass non-symbolic tensor to Keras Lambda layer","Body":"<p>I am trying to pass a <code>RNNCell<\/code> object to a Keras lambda layer so that I can use the Tensorflow layer within a Keras model, as follows.<\/p>\n\n<pre><code>conv_cell = ConvGRUCell(shape = [14, 14],\n                       filters = 32,\n                       kernel = [3,3],\n                       padding = 'SAME')\n\ndef convGRU(inputs, cell, length):\n    output, final = tf.nn.bidirectional_dynamic_rnn(\n            cell, cell, x, length, dtype=tf.float32)\n    output = tf.concat(output, -1)\n    final = tf.concat(final, -1)\n    return [output, final]\n\nlm = Lambda(lambda x: convGRU(x[0], x[1], x[2])([input, conv_cell, length])\n<\/code><\/pre>\n\n<p>However, I get an error that <code>conv_cell<\/code> is not a symbolic tensor (it is a custom layer based on Tensorflow's GRUCell).<\/p>\n\n<p>Is there any way to pass the cell to the lambda layer? I got it to work with functools.partial but it fails to save\/load the model because it cannot access the function inside the model.<\/p>\n","answers":[{"AnswerId":"58002258","CreationDate":"2019-09-19T00:36:37.843","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<pre><code>def convGRU(cell, length): # if length is produced by the model, use it with the inputs    \n    def inner_func(inputs):\n        code...\n    return inner_func\n\nlm = Lambda(convGRU(cell, length))(input)\n<\/code><\/pre>\n\n<p>For save\/load you need to use <code>custom_objects = {'convGRU': convGRU, 'cell':cell, 'length': length}<\/code>, etc. Whatever Keras doesn't know automatically needs to be in <code>custom_objects<\/code> for loading a saved model. <\/p>\n"}]}
+{"QuestionId":58001318,"AnswerCount":1,"Tags":"<python><tensorflow><nlp><tensor2tensor>","CreationDate":"2019-09-18T22:06:45.723","AcceptedAnswerId":null,"OwnerUserId":11157630.0,"Title":"no attribute 'contrib' for problems()","Body":"<p>I am trying to see the available problems() but it is giving Error.\nCan you please let me know if I am missing anything<\/p>\n\n<pre><code>&gt;&gt;&gt; from tensor2tensor import problems\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"C:\\Users\\\\Anaconda3\\lib\\site-packages\\tensor2tensor\\problems.py\", line 22, in &lt;module&gt;\n    from tensor2tensor.utils import registry\n  File \"C:\\Users\\\\Anaconda3\\lib\\site-packages\\tensor2tensor\\utils\\registry.py\", line 551, in &lt;module&gt;\n    attacks = tf.contrib.framework.deprecated(None, \"Use registry.attack\")(attack)\nAttributeError: module 'tensorflow' has no attribute 'contrib'\n&gt;&gt;&gt; tf.__version__\n'2.0.0-beta1'\n&gt;&gt;&gt;\n<\/code><\/pre>\n\n<p>I am working on windows<\/p>\n","answers":[{"AnswerId":"58926639","CreationDate":"2019-11-19T04:26:24.933","ParentId":null,"OwnerUserId":"5008968","Title":null,"Body":"<p>Tensor2Tensor library <a href=\"https:\/\/github.com\/tensorflow\/tensor2tensor\/issues\/1524#issuecomment-476321189\" rel=\"nofollow noreferrer\"><strong>is not yet compatible with Tensorflow 2.0<\/strong><\/a> because it still uses a lot of deprecated APIs. Your only option currently is to downgrade to an older version, such as Tensorflow 1.15.<\/p>\n\n<pre><code>pip3 install tensorflow==1.15.0\n<\/code><\/pre>\n"}]}
+{"QuestionId":58001321,"AnswerCount":1,"Tags":"<keras><lstm>","CreationDate":"2019-09-18T22:07:15.417","AcceptedAnswerId":58001358.0,"OwnerUserId":4335178.0,"Title":"convLSTM2D dilation_rate seems not to work","Body":"<p>I'm trying to do an upsampling using the dilation_rate from the convLSTM2D (Keras with Tenosrflow as backend)<\/p>\n\n<pre><code>input = Input(shape=(10, 64, 64, 1), name='encoder_input')\nlayer1 = ConvLSTM2D(filters=33, kernel_size=(5,5), dilation_rate=(2, 2))\nmodel = Model(input, layer1(input))\nplot_model(model, show_shapes=True, show_layer_names=True)\n<\/code><\/pre>\n\n<p>I would expect the output shape to be (None,128,128,33) but I got (None,64,64,33).\nWouldn't this dilation_rate=(2, 2) be the opposite to strides=(2, 2)?<\/p>\n","answers":[{"AnswerId":"58001358","CreationDate":"2019-09-18T22:11:54.600","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>Dilation, unlike stride, does not change the shape of the data. It simply increases the \"spread\" of the kernels. In this gif, you can see how it works:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/PrlC3.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/PrlC3.gif\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The only change in the shape of the data comes from cutting off 2 from each side, because no padding is used.<\/p>\n"}]}
+{"QuestionId":58001456,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><deep-learning><tf.keras>","CreationDate":"2019-09-18T22:24:18.977","AcceptedAnswerId":58001684.0,"OwnerUserId":1917160.0,"Title":"TensorFlow Keras 'accuracy' metric under the hood implementation","Body":"<p>When building a classifier using TensorFlow Keras, one often monitors model accuracy by specifying <code>metrics=['accuracy']<\/code> during the compilation step:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = tf.keras.Model(...)\nmodel.compile(optimizer=..., loss=..., metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>This behaves correctly whether or not the model outputs logits or class probabilities, and whether or not the model expects ground truth labels to be one-hot-encoded vectors or integer indices (i.e., integers in the interval <code>[0, n_classes)<\/code>).<\/p>\n\n<p>This is not the case if one wants to use cross-entropy loss: each of the four combinations of the cases mentioned above requires a different <code>loss<\/code> value to be passed during the compilation step:<\/p>\n\n<ol>\n<li><p>If the model outputs probabilities and the ground truth labels are one-hot-encoded, then <code>loss='categorical_crossentropy'<\/code> works.<\/p><\/li>\n<li><p>If the model outputs probabilities and the ground truth labels are integer indices, then <code>loss='sparse_categorical_crossentropy'<\/code> works.<\/p><\/li>\n<li><p>If the model outputs logits and the ground truth labels are one-hot-encoded, then <code>loss=tf.keras.losses.CategoricalCrossentropy(from_logits=True)<\/code> works.<\/p><\/li>\n<li><p>If the model outputs logits and the ground truth labels are integer indices, then <code>loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)<\/code> works.<\/p><\/li>\n<\/ol>\n\n<p>It seems that just specifying <code>loss='categorical_crossentropy'<\/code> is not robust enough to handle these four cases, whereas specifying <code>metrics=['accuracy']<\/code> <em>is<\/em> robust enough.<\/p>\n\n<hr>\n\n<p><strong>Question<\/strong>\nWhat is happening behind the scenes when the user specifies <code>metrics=['accuracy']<\/code> in the model compilation step that allows the accuracy computation to be performed correctly regardless of whether the model outputs logits or probabilities and whether the ground truth labels are one-hot-encoded vectors or integer indices?<\/p>\n\n<hr>\n\n<p>I suspect that the logits versus probabilities case is simple since the predicted class can be obtained as an argmax either way, but I would ideally like to be pointed to where in the TensorFlow 2 source code the computation is actually done.<\/p>\n\n<p>Please note that I am currently using <a href=\"https:\/\/pypi.org\/project\/tensorflow\/2.0.0rc1\/\" rel=\"nofollow noreferrer\">TensorFlow 2.0.0-rc1<\/a>.<\/p>\n\n<hr>\n\n<p><strong>Edit<\/strong>\nIn pure Keras, <code>metrics=['accuracy']<\/code> is handled explicitly in the <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/d8b226f26b35348d934edb1213061993e7e5a1fa\/keras\/engine\/training.py#L651-L665\" rel=\"nofollow noreferrer\"><code>Model.compile<\/code> method<\/a>.<\/p>\n","answers":[{"AnswerId":"58001684","CreationDate":"2019-09-18T22:55:23.997","ParentId":null,"OwnerUserId":"1917160","Title":null,"Body":"<p>Found it: this is handled in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/41474120daa68791e09eae99c8bdd8df093eda31\/tensorflow\/python\/keras\/engine\/training_utils.py#L1007-L1046\" rel=\"nofollow noreferrer\"><code>tensorflow.python.keras.engine.training_utils.get_metric_function<\/code><\/a>. In particular, the output shape is inspected to determine which accuracy function to use.<\/p>\n\n<p>To elaborate, in the current implementation <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/41474120daa68791e09eae99c8bdd8df093eda31\/tensorflow\/python\/keras\/engine\/training.py#L184-L400\" rel=\"nofollow noreferrer\"><code>Model.compile<\/code><\/a> either delegates metric processing to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/41474120daa68791e09eae99c8bdd8df093eda31\/tensorflow\/python\/keras\/engine\/training.py#L1563-L1575\" rel=\"nofollow noreferrer\"><code>Model._compile_eagerly<\/code><\/a> (if executing eagerly) or does it directly. In either case, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/41474120daa68791e09eae99c8bdd8df093eda31\/tensorflow\/python\/keras\/engine\/training.py#L1892-L1908\" rel=\"nofollow noreferrer\"><code>Model._cache_output_metric_attributes<\/code><\/a> is called, which calls <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/41474120daa68791e09eae99c8bdd8df093eda31\/tensorflow\/python\/keras\/engine\/training_utils.py#L748-L824\" rel=\"nofollow noreferrer\"><code>collect_per_output_metric_info<\/code><\/a> for both the unweighted and weighted metrics. This function loops over the provided metrics, calling <code>get_metric_function<\/code> on each one.<\/p>\n"}]}
+{"QuestionId":58001703,"AnswerCount":0,"Tags":"<python><pytorch>","CreationDate":"2019-09-18T22:58:06.003","AcceptedAnswerId":null,"OwnerUserId":7432133.0,"Title":"Images change color after unfolding using pytorch","Body":"<p>I am new to pytorch, for one of my projects I am dividing a large image into smaller tiles\/patches. I am using unfold to make this happen. My code is as follows<\/p>\n\n<pre><code>data = training_set[1][0].data.unfold(1, 64, 64).unfold(2, 64, 64).unfold(3, 64, 64)\n<\/code><\/pre>\n\n<p>After doing this I transpose the resultant matrix since the images are flipped, like this sample code<\/p>\n\n<pre><code>torch.t(data [0][0][0][0])\n<\/code><\/pre>\n\n<p>but the resultant images lose color, or get discolored for some reason, and I am worried that this might affect any calculations I do based on these patches.<\/p>\n\n<p>The following is a screenshot of the problem<a href=\"https:\/\/i.stack.imgur.com\/JKept.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/JKept.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The top is the patch and the bottom one is the complete picture<\/p>\n\n<p>Any help is appreciated, thanks<\/p>\n","answers":[]}
+{"QuestionId":58001774,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2019-09-18T23:09:29.877","AcceptedAnswerId":58108256.0,"OwnerUserId":4790871.0,"Title":"Tensorflow get the default device name","Body":"<p>In Tensorflow 1.14 I'm trying to use <code>tf.data.experimental.prefetch_to_device(device=...)<\/code> to prefetch my data to the GPU. But I'm not always training on a GPU, I often times train on a CPU (especially during development). <\/p>\n\n<p>Is there a way to get the current default device in use? Tensorflow either picks the CPU (when I set <code>CUDA_VISIBLE_DEVICES=-1<\/code>) otherwise it'll pick the GPU, the default usually works. <\/p>\n\n<p>So far I can only find a way to list visible devices with <code>sess.list_devices()<\/code>, but there must be a way to query the current default device so I don't have to manually change it in <code>prefetch_to_device<\/code> every time, right?<\/p>\n","answers":[{"AnswerId":"58157571","CreationDate":"2019-09-29T17:27:57.927","ParentId":null,"OwnerUserId":"4790871","Title":null,"Body":"<p>The best solution I've found so far is to use <code>tf.test.is_gpu_available<\/code>:<\/p>\n\n<pre><code>device = 'gpu:0' if tf.test.is_gpu_available() else 'cpu'\n<\/code><\/pre>\n"},{"AnswerId":"58108256","CreationDate":"2019-09-26T01:01:39.403","ParentId":null,"OwnerUserId":"9828197","Title":null,"Body":"<p>There is no API way of doing what you said currently. The closest <\/p>\n\n<blockquote>\n  <p><code>device = 'gpu:0' if tf.test.is_gpu_available() else 'cpu'<\/code><\/p>\n<\/blockquote>\n\n<p>is what you have already said. <\/p>\n\n<p>The reasons I think so is that the allocation is done at a low level: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/cf4dbb45ffb4d6ea0dc9c2ecfb514e874092cd16\/tensorflow\/core\/common_runtime\/colocation_graph.cc\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/cf4dbb45ffb4d6ea0dc9c2ecfb514e874092cd16\/tensorflow\/core\/common_runtime\/colocation_graph.cc<\/a><\/p>\n\n<p>Maybe you can also try with <a href=\"https:\/\/www.tensorflow.org\/guide\/using_gpu\" rel=\"nofollow noreferrer\"><code>soft placement<\/code><\/a><\/p>\n\n<p>Hope it helps.<\/p>\n"}]}
+{"QuestionId":58001987,"AnswerCount":1,"Tags":"<python><validation><keras>","CreationDate":"2019-09-18T23:45:48.797","AcceptedAnswerId":null,"OwnerUserId":11736147.0,"Title":"Bayesian Inference with dropout, validation set","Body":"<p>I have implemented the Bayesian SegNet using Keras (<a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/9412\" rel=\"nofollow noreferrer\">1<\/a>). To do so, I used the following custom loss function, which performs Bayesian inference for validation loss:<\/p>\n\n<pre><code>def custom_loss_Bayesian(y_true, y_pred):\n\n    train_loss = K.categorical_crossentropy(y_true, y_pred)                    \n\n    output_list = []\n    for i in range(n_MoteCarlo_Samples):                                       \n        output_list.append(K.categorical_crossentropy(y_true, y_pred))\n\n    Monty_sample_bin = K.stack(output_list,axis=0)  \n    val_loss_Bayesian=K.mean(Monty_sample_bin,axis=0)\n\n    return K.in_train_phase(train_loss, val_loss_Bayesian)\n<\/code><\/pre>\n\n<p>It works fine, but this approach has a big problem. By increasing the number of Monte Carlo samples, the training process takes a lot longer. This is probably due to the fact that the loop for validation loss is calculated for each training batch, but it is not used anywhere. I only need <code>val_loss_Bayesian<\/code> after the end of each epoch. Is there a smarter way to do this? Please note, that I want to use <code>ModelCheckpoint<\/code> to save the set of weights with the lowest <code>val_loss_Bayesian<\/code>. I tried to implement the solution by lovecambi as in <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/3155\" rel=\"nofollow noreferrer\">2<\/a>, such that <code>n_MonteCarlo_Samples<\/code> is selected based on batch number, but it did not work. It appears that when you compile a model, the properties inside the loss function cannot be changed. Any suggestions are much appreciated.<\/p>\n","answers":[{"AnswerId":"58034140","CreationDate":"2019-09-20T19:25:03.390","ParentId":null,"OwnerUserId":"11736147","Title":null,"Body":"<p>Based on the discussion in <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/10426\" rel=\"nofollow noreferrer\">1<\/a>, it seems that Keras calculates the validation loss for each batch at the same time as the training loss. However, during the training process, only a running average is constantly being reported in the log for training. Therefore, the code snippet given above is doing its job as it is supposed to. It is taking log because validation loss is being calculated for each mini-batch. If I need Bayesian inference monitored in each epoch, I have to pay the computational price.  <\/p>\n"}]}
+{"QuestionId":58002079,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><computer-vision><unsupervised-learning><gan>","CreationDate":"2019-09-19T00:04:24.657","AcceptedAnswerId":null,"OwnerUserId":10094585.0,"Title":"Creating DCGANs for non power of 2 dimensional images","Body":"<p>I am attempting to make a DCGAN for images that have an unusual shape of (75,95,3).  The general formula I have seen in most DCGAN implementations is the generator has each set of layers double the the width and height while halving the channel size.  The opposite being done for the discriminator.  I would like to produce images with the above dimensions, is there a best way to do this?<\/p>\n\n<p>My first thought for the generator is that often the first layer is a dense layer which I could somehow set the number of nodes to make it convenient to blow up to (75,95,3).<\/p>\n\n<p>Another thought I had was using the general formula from above and then cropping the last layer to my shape, but my intuition tells me this would cause many issues with gradients.<\/p>\n\n<p>Any thoughts or suggestions would be greatly appreciated.<\/p>\n","answers":[]}
+{"QuestionId":58002524,"AnswerCount":0,"Tags":"<pytorch>","CreationDate":"2019-09-19T01:25:08.187","AcceptedAnswerId":null,"OwnerUserId":10364071.0,"Title":"GroupNorm is considerably slower and consumes higher GPU memory than BatchNorm in Pytorch","Body":"<p>I use GroupNorm in pytorch instead of BatchNorm and keep all the others (network architecture) unchanged. It shows that in Imagenet dataset, using resnet50 architecture, GroupNorm is 40% slower than BatchNorm, and consumes 33% more GPU memory than BatchNorm. I am really confused because GroupNorm shouldn\u2019t need more calculation than BatchNorm. The details are listed below.<\/p>\n\n<p>For details of Group Normalization, one can see this paper: <a href=\"https:\/\/arxiv.org\/pdf\/1803.08494.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1803.08494.pdf<\/a><\/p>\n\n<p>For BatchNorm, one minibatch consumes 12.8 seconds with GPU memory 7.51GB;<\/p>\n\n<p>For GroupNorm, one minibatch consumes 17.9 seconds with GPU memory 10.02GB.<\/p>\n\n<p>I use the following code to convert all the BatchNorm layers to GroupNorm layers.<\/p>\n\n<pre><code>def convert_bn_model_to_gn(module, num_groups=16):\n\"\"\"\nRecursively traverse module and its children to replace all instances of\n``torch.nn.modules.batchnorm._BatchNorm`` with :class:`torch.nn.GroupNorm`.\nArgs:\n    module: your network module\n    num_groups: num_groups of GN\n\"\"\"\nmod = module\nif isinstance(module, nn.modules.batchnorm._BatchNorm):\n    mod = nn.GroupNorm(num_groups, module.num_features,\n                       eps=module.eps, affine=module.affine)\n    # mod = nn.modules.linear.Identity()\n    if module.affine:\n        mod.weight.data = module.weight.data.clone().detach()\n        mod.bias.data = module.bias.data.clone().detach()\nfor name, child in module.named_children():\n    mod.add_module(name, convert_bn_model_to_gn(\n        child, num_groups=num_groups))\ndel module\nreturn mod\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58002531,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2019-09-19T01:27:17.593","AcceptedAnswerId":null,"OwnerUserId":741099.0,"Title":"Disable Tensorflow\/Numpy Deprecation Warning Messages","Body":"<p>Running my Python 3.7.4 app that uses <code>tensorflow<\/code> v1.14.0 causes a large list of deprecation warnings to appear. The following code clears up most of it.<\/p>\n\n<pre><code>try:\n    from tensorflow.python.util import module_wrapper as deprecation\nexcept ImportError:\n    from tensorflow.python.util import deprecation_wrapper as deprecation\ndeprecation._PER_MODULE_WARNING_LIMIT = 0\n<\/code><\/pre>\n\n<p>However, none of the warnings are removed. Updating <code>tensorflow<\/code> to v2.x is not an option right now.<\/p>\n\n<p>How can these messages be removed?<\/p>\n\n<p><strong>Warning messages:<\/strong><\/p>\n\n<pre><code>\/anaconda3\/envs\/ml\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/dtypes.py:516: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint8 = np.dtype([(\"qint8\", np.int8, 1)])\n\/anaconda3\/envs\/ml\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/dtypes.py:517: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint8 = np.dtype([(\"quint8\", np.uint8, 1)])\n\/anaconda3\/envs\/ml\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/dtypes.py:518: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint16 = np.dtype([(\"qint16\", np.int16, 1)])\n\/anaconda3\/envs\/ml\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/dtypes.py:519: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint16 = np.dtype([(\"quint16\", np.uint16, 1)])\n\/anaconda3\/envs\/ml\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/dtypes.py:520: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint32 = np.dtype([(\"qint32\", np.int32, 1)])\n\/anaconda3\/envs\/ml\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/dtypes.py:525: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  np_resource = np.dtype([(\"resource\", np.ubyte, 1)])\nWARNING:tensorflow:From \/Users\/x\/foo\/dnnlib\/tflib\/tfutil.py:34: The name tf.Dimension is deprecated. Please use tf.compat.v1.Dimension instead.\n\nWARNING:tensorflow:From \/Users\/x\/foo\/dnnlib\/tflib\/tfutil.py:74: The name tf.variable_scope is deprecated. Please use tf.compat.v1.variable_scope instead.\n\nWARNING:tensorflow:From \/Users\/x\/foo\/dnnlib\/tflib\/tfutil.py:128: The name tf.Session is deprecated. Please use tf.compat.v1.Session instead.\n\n<\/code><\/pre>\n","answers":[{"AnswerId":"58267619","CreationDate":"2019-10-07T10:20:35.967","ParentId":null,"OwnerUserId":"9805238","Title":null,"Body":"<p>You can ignore all \"FutureWarning\" warnings by putting the following code at the beginning of your scripts (<a href=\"https:\/\/machinelearningmastery.com\/how-to-fix-futurewarning-messages-in-scikit-learn\/\" rel=\"nofollow noreferrer\">source<\/a>):<\/p>\n\n<pre><code>from warnings import simplefilter \nsimplefilter(action='ignore', category=FutureWarning)\n<\/code><\/pre>\n\n<p>Please note that you have to put those lines before any other imports as some of those might already import dependencies on their own.<\/p>\n"}]}
+{"QuestionId":58002537,"AnswerCount":0,"Tags":"<python><anaconda><conda><caffe><pycaffe>","CreationDate":"2019-09-19T01:28:01.680","AcceptedAnswerId":null,"OwnerUserId":12065774.0,"Title":"Can't Install Caffe Using Conda","Body":"<p>I installed cafe on my GPU machine, but importing it gives 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\/tafa\/anaconda2\/envs\/caffe\/lib\/python2.7\/site-packages\/caffe\/__init__.py\", line 1, in &lt;module&gt;\n    from .pycaffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, RMSPropSolver, AdaDeltaSolver, AdamSolver, NCCL, Timer\n  File \"\/home\/tafa\/anaconda2\/envs\/caffe\/lib\/python2.7\/site-packages\/caffe\/pycaffe.py\", line 13, in &lt;module&gt;\n    from ._caffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, \\\nImportError: \/home\/tafa\/anaconda2\/envs\/caffe\/lib\/python2.7\/site-packages\/caffe\/..\/..\/..\/libcaffe.so.1.0.0: undefined symbol: _ZNK6google8protobuf7Message11GetTypeNameB5cxx11Ev\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58002600,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow2.0><tf.keras>","CreationDate":"2019-09-19T01:38:59.713","AcceptedAnswerId":null,"OwnerUserId":12087822.0,"Title":"How do sessions and parallelism work in TF2.0?","Body":"<p>I am trying to run two tensorflow models in parallel in the same process.<\/p>\n\n<p>In Tensorflow 1.x we could do e.g. <a href=\"https:\/\/stackoverflow.com\/questions\/46725323\/keras-tensorflow-exception-while-predicting-from-multiple-threads\">Keras Tensorflow - Exception while predicting from multiple threads<\/a><\/p>\n\n<pre><code>graph = tf.Graph()\nwith graph.as_default():\n    session = tf.Session()\n    with session.as_default():\n        # Create and use the model ...\n<\/code><\/pre>\n\n<p>In Tensorflow 2.0, the session has been removed.<\/p>\n\n<p>When I try to run two models in parallel within the same process, Tensorflow crashes. Is there a way around this problem is TF 2.0?<\/p>\n\n<p>Error message:<\/p>\n\n<pre><code>2019-09-18 16:26:56.978352: E tensorflow\/stream_executor\/cuda\/cuda_blas.cc:238] failed to create cublas handle: CUBLAS_STATUS_NOT_INITIALIZED\n2019-09-18 16:26:56.981418: E tensorflow\/stream_executor\/cuda\/cuda_blas.cc:238] failed to create cublas handle: CUBLAS_STATUS_NOT_INITIALIZED\n2019-09-18 16:26:56.983202: E tensorflow\/stream_executor\/cuda\/cuda_blas.cc:238] failed to create cublas handle: CUBLAS_STATUS_NOT_INITIALIZED\n2019-09-18 16:26:56.985757: E tensorflow\/stream_executor\/cuda\/cuda_blas.cc:238] failed to create cublas handle: CUBLAS_STATUS_NOT_INITIALIZED\n2019-09-18 16:26:56.987541: E tensorflow\/stream_executor\/cuda\/cuda_blas.cc:238] failed to create cublas handle: CUBLAS_STATUS_NOT_INITIALIZED\n2019-09-18 16:26:56.990140: E tensorflow\/stream_executor\/cuda\/cuda_blas.cc:238] failed to create cublas handle: CUBLAS_STATUS_NOT_INITIALIZED\n2019-09-18 16:26:56.992499: E tensorflow\/stream_executor\/cuda\/cuda_blas.cc:238] failed to create cublas handle: CUBLAS_STATUS_NOT_INITIALIZED\n2019-09-18 16:26:56.994886: E tensorflow\/stream_executor\/cuda\/cuda_blas.cc:238] failed to create cublas handle: CUBLAS_STATUS_NOT_INITIALIZED\n2019-09-18 16:26:56.997778: E tensorflow\/stream_executor\/cuda\/cuda_blas.cc:238] failed to create cublas handle: CUBLAS_STATUS_NOT_INITIALIZED\n2019-09-18 16:26:56.997802: W tensorflow\/stream_executor\/stream.cc:1919] attempting to perform BLAS operation using StreamExecutor without BLAS support\n2019-09-18 16:26:56.997829: W tensorflow\/core\/common_runtime\/base_collective_executor.cc:216] BaseCollectiveExecutor::StartAbort Internal: Blas GEMM launch failed : a.shape=(1, 2520), b.shape=(2520, 2520), m=1, n=2520, k=2520\n     [[{{node dense\/MatMul}}]]\n2019-09-18 16:26:57.000709: E tensorflow\/stream_executor\/cuda\/cuda_blas.cc:238] failed to create cublas handle: CUBLAS_STATUS_NOT_INITIALIZED\n2019-09-18 16:26:57.000744: W tensorflow\/stream_executor\/stream.cc:1919] attempting to perform BLAS operation using StreamExecutor without BLAS support\nTraceback (most recent call last):\n  File \n    result = model.predict([numpy.array([inV,]), numpy.array([inS,])])\n  File \"\/usr\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/keras\/engine\/training.py\", line 915, in predict\n    use_multiprocessing=use_multiprocessing)\n  File \"\/usr\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/keras\/engine\/training_arrays.py\", line 722, in predict\n    callbacks=callbacks)\n  File \"\/usr\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/keras\/engine\/training_arrays.py\", line 393, in model_iteration\n    batch_outs = f(ins_batch)\n  File \"\/usr\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/keras\/backend.py\", line 3625, in __call__\n    outputs = self._graph_fn(*converted_inputs)\n  File \"\/usr\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/eager\/function.py\", line 1081, in __call__\n    return self._call_impl(args, kwargs)\n  File \"\/usr\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/eager\/function.py\", line 1121, in _call_impl\n    return self._call_flat(args, self.captured_inputs, cancellation_manager)\n  File \"\/usr\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/eager\/function.py\", line 1224, in _call_flat\n    ctx, args, cancellation_manager=cancellation_manager)\n  File \"\/usr\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/eager\/function.py\", line 511, in call\n    ctx=ctx)\n  File \"\/usr\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/eager\/execute.py\", line 67, in quick_execute\n    six.raise_from(core._status_to_exception(e.code, message), None)\n  File \"&lt;string&gt;\", line 3, in raise_from\ntensorflow.python.framework.errors_impl.InternalError:  Blas GEMM launch failed : a.shape=(1, 2520), b.shape=(2520, 2520), m=1, n=2520, k=2520\n     [[node dense\/MatMul (defined at usr\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/framework\/ops.py:1751) ]] [Op:__inference_keras_scratch_graph_378]\n\nFunction call stack:\nkeras_scratch_graph\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58002728,"AnswerCount":0,"Tags":"<python><tensorflow><keras><lstm><seq2seq>","CreationDate":"2019-09-19T02:00:58.617","AcceptedAnswerId":null,"OwnerUserId":8240510.0,"Title":"Seq2seq encoder, decoder model stack LSTM","Body":"<p>I have tried <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/lstm_seq2seq.py\" rel=\"nofollow noreferrer\">this<\/a> keras example of seq2seq, encoder-decoder model for language translation. Model definition:<\/p>\n\n<pre><code># Define an input sequence and process it.\nencoder_inputs = Input(shape=(None, num_encoder_tokens))\nencoder = CuDNNLSTM(latent_dim, return_state=True)\nencoder_outputs, state_h, state_c = encoder(encoder_inputs)\n# We discard `encoder_outputs` and only keep the states.\nencoder_states = [state_h, state_c]\n\n# Set up the decoder, using `encoder_states` as initial state.\ndecoder_inputs = Input(shape=(None, num_decoder_tokens))\n# We set up our decoder to return full output sequences,\n# and to return internal states as well. We don't use the\n# return states in the training model, but we will use them in inference.\ndecoder_lstm = CuDNNLSTM(latent_dim, return_sequences=True, return_state=True)\ndecoder_outputs, _, _ = decoder_lstm(decoder_inputs,\n                                     initial_state=encoder_states)\ndecoder_dense = Dense(num_decoder_tokens, activation='softmax')\ndecoder_outputs = decoder_dense(decoder_outputs)\n\n# Define the model that will turn\n# `encoder_input_data` &amp; `decoder_input_data` into `decoder_target_data`\nmodel = Model([encoder_inputs, decoder_inputs], decoder_outputs)\n<\/code><\/pre>\n\n<p>Model successfully learned english to russian character level translation to some extend. But as one can see encoder and decoder consists of only on <code>LSTM layer<\/code>. From common sense stacking many <code>LSTM<\/code> layers together would give better results. <\/p>\n\n<blockquote>\n  <p>Is it good idea to stack LSTM layers, if yes how in this case?<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":58002754,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-09-19T02:05:16.280","AcceptedAnswerId":58003697.0,"OwnerUserId":10547514.0,"Title":"I've installed PyTorch, but CUDA is not recognized","Body":"<p>I installed <a href=\"https:\/\/en.wikipedia.org\/wiki\/PyTorch\" rel=\"nofollow noreferrer\">PyTorch<\/a> using this command:<\/p>\n\n<pre><code>pip3 install torch===1.2.0 torchvision===0.4.0 -f https:\/\/download.pytorch.org\/whl\/torch_stable.html\n<\/code><\/pre>\n\n<p>And it seemed to work. When importing torch, there isn't any error. \nMy laptop has a <a href=\"https:\/\/en.wikipedia.org\/wiki\/GeForce_10_series#GeForce_10_(10xx)_series\" rel=\"nofollow noreferrer\">GeForce 1060 TI<\/a>, which I assumed would work with CUDA.<\/p>\n\n<p>Here is the error in the IDE (Eclipse):<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/ey7wU.png\" alt=\"Screenshot with code\"><\/p>\n","answers":[{"AnswerId":"58003697","CreationDate":"2019-09-19T04:23:54.133","ParentId":null,"OwnerUserId":"4171008","Title":null,"Body":"<p><strong>You should include code, not a graph!<\/strong><\/p>\n\n<p>Remove PyTorch and install it again with this command:<\/p>\n\n<pre><code>conda install pytorch -c pytorch\npip3 install torchvision\n<\/code><\/pre>\n\n<p>after this try checking it with this command:<\/p>\n\n<pre><code>import torch\nimport torchvision\ntrain_on_gpu = torch.cuda.is_available()\n\nif train_on_gpu:\n    print('CUDA is available, Training on GPU ...')\nelse:\n    print('CUDA is not available!  Training on CPU ...')\n<\/code><\/pre>\n"}]}
+{"QuestionId":58002836,"AnswerCount":2,"Tags":"<python><pytorch><torchvision>","CreationDate":"2019-09-19T02:18:36.300","AcceptedAnswerId":58003869.0,"OwnerUserId":6423473.0,"Title":"PyTorch [1 if x > 0.5 else 0 for x in outputs ] with tensors","Body":"<p>I have a list outputs from a sigmoid function as a tensor in PyTorch<\/p>\n\n<p>E.g<\/p>\n\n<pre><code>output (type) = torch.Size([4]) tensor([0.4481, 0.4014, 0.5820, 0.2877], device='cuda:0',\n<\/code><\/pre>\n\n<p>As I'm doing binary classification I want to turn all values bellow 0.5 to 0 and above 0.5 to 1.<\/p>\n\n<p>Traditionally with a NumPy array you can use list iterators:<\/p>\n\n<pre><code>output_prediction = [1 if x &gt; 0.5 else 0 for x in outputs ]\n<\/code><\/pre>\n\n<p>This would work, however I have to later convert output_prediction back to a tensor to use<\/p>\n\n<pre><code>torch.sum(ouput_prediction == labels.data)\n<\/code><\/pre>\n\n<p>Where labels.data is a binary tensor of labels.<\/p>\n\n<p>Is there a way to use list iterators with tensors?<\/p>\n","answers":[{"AnswerId":"58003869","CreationDate":"2019-09-19T04:45:54.820","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<pre class=\"lang-py prettyprint-override\"><code>prob = torch.tensor([0.3,0.4,0.6,0.7])\n\nout = (prob&gt;0.5).float()\n# tensor([0.,0.,1.,1.])\n<\/code><\/pre>\n\n<p>Explanation: In pytorch, you can directly use <code>prob&gt;0.5<\/code> to get a <code>torch.bool<\/code> type tensor. Then you can convert to float type via <code>.float()<\/code>.<\/p>\n"},{"AnswerId":"58004502","CreationDate":"2019-09-19T05:53:19.920","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<p>Why not consider using a <em>loop<\/em>less solution? Maybe something like below would suffice:<\/p>\n\n<pre><code>In [34]: output = torch.tensor([0.4481, 0.4014, 0.5820, 0.2877]) \n\n# subtract off the threshold value (0.5), create a boolean mask, \n# and then cast the resultant tensor to an `int` type\nIn [35]: result = torch.as_tensor((output - 0.5) &gt; 0, dtype=torch.int32) \n\nIn [36]: result        \nOut[36]: tensor([0, 0, 1, 0], dtype=torch.int32)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58003010,"AnswerCount":2,"Tags":"<python><tensorflow><tfrecord>","CreationDate":"2019-09-19T02:46:02.880","AcceptedAnswerId":null,"OwnerUserId":10284833.0,"Title":"tensorflow TFRecords cannot parse serialized example","Body":"<p>I am getting the following error when trying to decode a TFRecord example : <\/p>\n\n<blockquote>\n  <p>InvalidArgumentError: Name: , Feature: relevance (data type:\n  float) is required but could not be found. [Op:ParseExample]<\/p>\n<\/blockquote>\n\n<p>To decode the example, I am using <code>tf.io.parse_example<\/code> as follows<\/p>\n\n<pre><code>example_features = tf.compat.v1.io.parse_example(\n     tf.reshape(serialized_list, [-1]), peritem_feature_spec)\n<\/code><\/pre>\n\n<p>where <code>serialized_list<\/code> looks like <\/p>\n\n<pre><code>[ &lt;example_object&gt;, ...\nb'\\n\\xcc\\x01\\n\\x15\\n\\trelevance\\x12\\x08\\x12\\x06\\n\\x04\\x9a\\xe9`D\\n\\xb2\n\\x01\\n\\x13encoded_clust_index\\x12\\x9a\\00\\ &lt;more...&gt;]\n<\/code><\/pre>\n\n<p>and <code>peritem_feature_spec<\/code> looks like<\/p>\n\n<pre><code>peritem_feature_spec = {\n    'relevance':tf.FixedLenFeature([], tf.float32),\n    'encoded_clust_index':tf.VarLenFeature(tf.float32)\n    }\n<\/code><\/pre>\n\n<p>I'm very confused why the feature 'relevance' can't be found.  I think I encoded and created my TFRecord objects correctly.  Am I creating the feature_spec incorrectly?  My thought is <code>tf.VarLenFeature<\/code> is the incorrect feature type to use, but I cannot figure out what is correct. <\/p>\n\n<p>Using <code>tensorflow_ranking.python.data.parse_from_example_in_example<\/code> is able to correctly decode the TFRecord into its features, but I don't know why <code>tf.io.parse_example<\/code> fails<\/p>\n","answers":[{"AnswerId":"58020973","CreationDate":"2019-09-20T02:22:12.807","ParentId":null,"OwnerUserId":"10284833","Title":null,"Body":"<p>EDIT : My old answer is below this answer<\/p>\n\n<p>The right answer was to provide a <code>default_value<\/code> to the feature spec<\/p>\n\n<pre><code>peritem_feature_spec = {\n    'relevance':tf.FixedLenFeature([], tf.float32, default_value=0.0),\n    'encoded_clust_index':tf.VarLenFeature(tf.float32, default_value=0.0)\n    }\n<\/code><\/pre>\n\n<p>My <strong>old<\/strong> (works, but incorrect) answer is below<\/p>\n\n<p>So the issue came down to how my features were being padded in the tensorflow_ranking library. It was padding a list feature like : <\/p>\n\n<pre><code>def pad_fn():\n    return tf.pad(\n        tensor=serialized_list,\n        paddings=[[0, 0], [0, list_size -cur_list_size]],\n            constant_values=\"\")\n<\/code><\/pre>\n\n<p>This method appends empty bytes to the end of the tensor.  The parser looks for a <code>feature_name<\/code> in the empty tensors, and responds that it cant find it.  My workaround was to append serialized TFRecord example protos instead of an empty byte string.  I accomplished this like so : <\/p>\n\n<pre><code>def pad_fn():\n    # Create feature spec for tf.train.Example to append\n    pad_spec = {}\n    # Default values are 0 or an empty byte string depending on \n    # original serialized data type\n    dtype_map = {tf.float32:tf.train.Feature(\n            float_list=tf.train.FloatList(value=[0.0])), \n                 tf.int32:tf.train.Feature(\n                         int64_list=tf.train.Int64List(value=[0])), \n                 tf.string:tf.train.Feature(\n                         bytes_list=tf.train.BytesList(\n                                 value=[bytes('', encoding='UTF-8')]))}\n     # Create the feature spec\n    for key, item in peritem_feature_spec.items():\n        dtype = item.dtype\n        pad_spec[key] = dtype_map[dtype]\n    # Make and serialize example to append\n    constant_values = tf.train.Example(\n            features=tf.train.Features(feature=pad_spec))\n    constant_val_str = constant_values.SerializeToString()\n\n    # Add serialized padding to end of list\n    return tf.pad(\n        tensor=serialized_list,\n        paddings=[[0, 0], [0, list_size - cur_list_size]],\n        constant_values=constant_val_str)\n<\/code><\/pre>\n"},{"AnswerId":"58012472","CreationDate":"2019-09-19T13:44:30.077","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>Instead of <code>VarLenFeature<\/code> can you try <code>tf.FixedLenSequenceFeature([], tf.float32, allow_missing = True ,default_value=0)<\/code> explained <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/io\/FixedLenSequenceFeature\" rel=\"nofollow noreferrer\">here<\/a> and see if that works?<\/p>\n"}]}
+{"QuestionId":58003156,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-19T03:05:25.703","AcceptedAnswerId":null,"OwnerUserId":5440895.0,"Title":"tensorflow restored model naming and ways to restore variables","Body":"<p>I have a problem with loading the saved model in tensorflow.<\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><code>import tensorflow as tf\ntrain = False\nif train:\n    with tf.variable_scope('Composition',  reuse=tf.AUTO_REUSE):\n        myw = tf.get_variable(\"L\",shape=[2,3], initializer=tf.random_uniform_initializer(-0.05,0.05))      \n    with tf.Session() as sess:\n        init = tf.initializers.global_variables()                                                                                                                          \n        sess.run([init])\n        saver = tf.train.Saver()\n        saver.save(sess, '.\/example')                                                                                                                                              \n\nif not train:\n    with tf.Session() as sess:\n        saver = tf.train.import_meta_graph('example.meta')                                                                                                                                 \n        saver.restore(sess,tf.train.latest_checkpoint('.\/'))\n        print('-----------variable names:',  [var.name for var in tf.global_variables()])\n        with tf.variable_scope('Composition',  reuse=True):\n            myl = tf.get_variable(\"L\",shape=[2,3]) \n<\/code><\/pre>\n\n<p>print result:\n'-----------variable names: ['Composition\/L:0']'<\/p>\n\n<p>error:<\/p>\n\n<pre><code>traceback (most recent call last):\n  File \"code.py\", line 22, in &lt;module&gt;\n    myl = tf.get_variable(\"L\",shape=[2,3])\n  File \"\/usr\/local\/lib\/python3.7\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 1479, in get_variable\n    aggregation=aggregation)\n  File \"\/usr\/local\/lib\/python3.7\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 1220, in get_variable\n    aggregation=aggregation)\n  File \"\/usr\/local\/lib\/python3.7\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 547, in get_variable\n    aggregation=aggregation)\n  File \"\/usr\/local\/lib\/python3.7\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 499, in _true_getter\n    aggregation=aggregation)\n  File \"\/usr\/local\/lib\/python3.7\/site-packages\/tensorflow\/python\/ops\/variable_scope.py\", line 866, in _get_single_variable\n    \"reuse=tf.AUTO_REUSE in VarScope?\" % name)\nValueError: Variable Composition\/L does not exist, or was not created with tf.get_variable(). Did you mean to set reuse=tf.AUTO_REUSE in VarScope?\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58003288,"AnswerCount":0,"Tags":"<python><tensorflow><keras><generative-adversarial-network><dcgan>","CreationDate":"2019-09-19T03:22:28.483","AcceptedAnswerId":null,"OwnerUserId":11794827.0,"Title":"3D DCGAN doesn't learn as expected","Body":"<p>I'm trying to create a 3D DCGan in Keras, following this paper: <a href=\"http:\/\/3dgan.csail.mit.edu\/papers\/3dgan_nips.pdf\" rel=\"nofollow noreferrer\">http:\/\/3dgan.csail.mit.edu\/papers\/3dgan_nips.pdf<\/a><\/p>\n\n<p>In my use case, I wish to output a (16,16,16) array of integers.\nThe inputs are scaled between [-1,1] when they are fed into the model, then scaled up and rounded before visualization.<\/p>\n\n<p>To visualize the learning, I displayed the first \"slice\" of the array as an image.<\/p>\n\n<p>The generator is made like so:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def build_generator():\n    z_size = 200\n    gen_filters = [64, 32, 16, 1]\n    gen_kernel_sizes = [2, 4, 4, 4]\n    gen_strides = [1, 2, 2, 2]\n    gen_input_shape = (1, 1, 1, z_size)\n    gen_activations = ['relu', 'relu', 'relu', 'tanh']\n    gen_convolutional_blocks = 4\n\n    input_layer = layers.Input(shape=gen_input_shape)\n\n    a = layers.Conv3DTranspose(filters=gen_filters[0], \n                        kernel_size=gen_kernel_sizes[0],\n                        strides=gen_strides[0])(input_layer)\n    a = layers.BatchNormalization()(a, training=True)\n    a = layers.Activation(activation='relu')(a)\n\n    for i in range(gen_convolutional_blocks - 1):\n        a = layers.Conv3DTranspose(filters=gen_filters[i + 1], \n                            kernel_size=gen_kernel_sizes[i + 1],\n                            strides=gen_strides[i + 1], padding='same')(a)\n        a = layers.BatchNormalization()(a, training=True)\n        a = layers.Activation(activation=gen_activations[i + 1])(a)\n    gen_model = Model(inputs=input_layer, outputs=a)\n    gen_model.summary()\n    return gen_model\n<\/code><\/pre>\n\n<pre><code>Model: \"model_2\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_3 (InputLayer)         [(None, 1, 1, 1, 200)]    0         \n_________________________________________________________________\nconv3d_transpose_4 (Conv3DTr (None, 2, 2, 2, 64)       102464    \n_________________________________________________________________\nbatch_normalization_8 (Batch (None, 2, 2, 2, 64)       256       \n_________________________________________________________________\nactivation_5 (Activation)    (None, 2, 2, 2, 64)       0         \n_________________________________________________________________\nconv3d_transpose_5 (Conv3DTr (None, 4, 4, 4, 32)       131104    \n_________________________________________________________________\nbatch_normalization_9 (Batch (None, 4, 4, 4, 32)       128       \n_________________________________________________________________\nactivation_6 (Activation)    (None, 4, 4, 4, 32)       0         \n_________________________________________________________________\nconv3d_transpose_6 (Conv3DTr (None, 8, 8, 8, 16)       32784     \n_________________________________________________________________\nbatch_normalization_10 (Batc (None, 8, 8, 8, 16)       64        \n_________________________________________________________________\nactivation_7 (Activation)    (None, 8, 8, 8, 16)       0         \n_________________________________________________________________\nconv3d_transpose_7 (Conv3DTr (None, 16, 16, 16, 1)     1025      \n_________________________________________________________________\nbatch_normalization_11 (Batc (None, 16, 16, 16, 1)     4         \n_________________________________________________________________\nactivation_8 (Activation)    (None, 16, 16, 16, 1)     0         \n=================================================================\nTotal params: 267,829\nTrainable params: 267,603\nNon-trainable params: 226\n\n<\/code><\/pre>\n\n<p>The discriminator:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def build_discriminator():\n    dis_input_shape = (16, 16, 16, 1)\n    dis_filters = [16, 32, 64, 1]\n    dis_kernel_sizes = [4, 4, 4, 4]\n    dis_strides = [2, 2, 2, 2]\n    dis_paddings = ['same', 'same', 'same', 'valid']\n    dis_alphas = [0.2, 0.2, 0.2, 0.2, 0.2]\n    dis_activations = ['leaky_relu', 'leaky_relu', 'leaky_relu', \n                       'sigmoid']\n    dis_convolutional_blocks = 4\n\n    dis_input_layer = layers.Input(shape=dis_input_shape)\n\n    a = layers.Conv3D(filters=dis_filters[0],\n               kernel_size=dis_kernel_sizes[0],\n               strides=dis_strides[0],\n               padding=dis_paddings[0])(dis_input_layer)\n    a = layers.BatchNormalization()(a, training=True)\n    a = layers.LeakyReLU(dis_alphas[0])(a)\n\n    for i in range(dis_convolutional_blocks - 1):\n        a = layers.Conv3D(filters=dis_filters[i + 1],\n                   kernel_size=dis_kernel_sizes[i + 1],\n                   strides=dis_strides[i + 1],\n                   padding=dis_paddings[i + 1])(a)\n        a = layers.BatchNormalization()(a, training=True)\n        if dis_activations[i + 1] == 'leaky_relu':\n            a = layers.LeakyReLU(dis_alphas[i + 1])(a)\n        elif dis_activations[i + 1] == 'sigmoid':\n            a = layers.Activation(activation='sigmoid')(a)\n\n    dis_model = Model(inputs=dis_input_layer, outputs=a)\n    print(dis_model.summary())\n    return dis_model\n\n<\/code><\/pre>\n\n<pre><code>_________________________________________________________________\nModel: \"model_3\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_4 (InputLayer)         [(None, 16, 16, 16, 1)]   0         \n_________________________________________________________________\nconv3d_4 (Conv3D)            (None, 8, 8, 8, 16)       1040      \n_________________________________________________________________\nbatch_normalization_12 (Batc (None, 8, 8, 8, 16)       64        \n_________________________________________________________________\nleaky_re_lu_3 (LeakyReLU)    (None, 8, 8, 8, 16)       0         \n_________________________________________________________________\nconv3d_5 (Conv3D)            (None, 4, 4, 4, 32)       32800     \n_________________________________________________________________\nbatch_normalization_13 (Batc (None, 4, 4, 4, 32)       128       \n_________________________________________________________________\nleaky_re_lu_4 (LeakyReLU)    (None, 4, 4, 4, 32)       0         \n_________________________________________________________________\nconv3d_6 (Conv3D)            (None, 2, 2, 2, 64)       131136    \n_________________________________________________________________\nbatch_normalization_14 (Batc (None, 2, 2, 2, 64)       256       \n_________________________________________________________________\nleaky_re_lu_5 (LeakyReLU)    (None, 2, 2, 2, 64)       0         \n_________________________________________________________________\nconv3d_7 (Conv3D)            (None, 1, 1, 1, 1)        4097      \n_________________________________________________________________\nbatch_normalization_15 (Batc (None, 1, 1, 1, 1)        4         \n_________________________________________________________________\nactivation_9 (Activation)    (None, 1, 1, 1, 1)        0         \n=================================================================\nTotal params: 169,525\nTrainable params: 169,299\nNon-trainable params: 226\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>An adversarial model is then created:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>adversarial_model = tf.keras.models.Sequential()\nadversarial_model.add(generator)\nadversarial_model.add(discriminator)\nadversarial_model.compile(loss=\"binary_crossentropy\", optimizer=tf.keras.optimizers.Adam(lr=gen_learning_rate, beta_1=beta))\n<\/code><\/pre>\n\n<p>The data is preprocessed like so:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>processed_inputs = np.subtract(np.multiply(np.divide(dataset_list, scalar),2),1)\n<\/code><\/pre>\n\n<p>and converted back like so:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def generate_save_heatmaps(model, epoch, test_input):\n  predictions = model(test_input, training=False)\n  processed_predictions = np.divide(np.multiply(np.add(predictions, 1), scalar),2)\n  processed_predictions = np.around(processed_predictions).astype(int)\n  processed_predictions = np.squeeze(processed_predictions)\n  generate_heatmap(processed_predictions[0][0], epoch, True)\n<\/code><\/pre>\n\n<p>The model is trained via this function:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>for epoch in range(epochs):\n    gen_losses = []\n    dis_losses = []\n    number_of_batches = int(processed_inputs.shape[0] \/ batch_size)\n    for index in range(number_of_batches):\n        z_sample = np.random.normal(0, 0.33, size=[batch_size, 1, 1, 1, \n                            z_size]).astype(np.float32)\n        struct_batch = processed_inputs[index * batch_size:(index + 1) * batch_size, :, :, :]\n        gen_structs = generator.predict(z_sample)\n        # Make the discriminator network trainable\n        discriminator.trainable = True\n\n        # Train the discriminator network\n        loss_real = discriminator.train_on_batch(struct_batch, \n                                                 labels_real)\n        loss_fake = discriminator.train_on_batch(gen_structs, \n                                                 labels_fake)        \n        d_loss = 0.5 * np.add(loss_real, loss_fake)\n        discriminator.trainable = False\n\n        z = np.random.normal(0, 0.33, size=[batch_size, 1, 1, 1, z_size]).astype(np.float32)\n        # Train the adversarial model\n        g_loss = adversarial_model.train_on_batch(z, labels_real)\n\n        gen_losses.append(g_loss)\n        dis_losses.append(d_loss)\n\n        write_log('g_loss', np.mean(gen_losses), counter)\n        write_log('d_loss', np.mean(dis_losses), counter)\n<\/code><\/pre>\n\n<p>In all of the training data, the first slice looks like this:\n<img src=\"https:\/\/i.imgur.com\/UHR79Xj.png\" alt=\"\"><\/p>\n\n<p>Most of the data looks like that, so I hoped the net would quickly pick up an affinity for the number 12.<\/p>\n\n<p>However, an evolution over a few hundred epochs shows otherwise:\n<img src=\"https:\/\/i.imgur.com\/0ZPF8TM.gif\" alt=\"\"><\/p>\n\n<p>The tensorboard loss graph looks like this:\n<img src=\"https:\/\/i.imgur.com\/lccoy6a.png\" alt=\"\"><\/p>\n\n<p>I have tried different learning rates, changing the amount of convolution blocks, etc. I don't know at this point what to do.<\/p>\n","answers":[]}
+{"QuestionId":58003448,"AnswerCount":1,"Tags":"<python><tensorflow><linear-regression><valueerror><sequencenumericcolumn>","CreationDate":"2019-09-19T03:44:50.097","AcceptedAnswerId":null,"OwnerUserId":1508724.0,"Title":"ValueError: Items of feature_columns must be either a DenseColumn or CategoricalColumn","Body":"<p>The code is simple:<\/p>\n\n<pre><code>    x_data = np.linspace(0, 10.0, 1000000)\n    y_true = (0.5 * x_data) + 5 \n    x_train, x_eval, y_train, y_eval = train_test_split(x_data, y_true, test_size = 0.25, random_state=101)\n    input_func = tf.estimator.inputs.numpy_input_fn({'x':x_train}, y_train, \n                                                batch_size=8, num_epochs=None, shuffle= True)\n    estimator = tf.estimator.LinearRegressor(feature_columns=feat_cols)\n    estimator.train(input_fn=input_func, steps=1000)\n<\/code><\/pre>\n\n<p>The error:<\/p>\n\n<blockquote>\n  <p>INFO:tensorflow:Calling model_fn.\n  --------------------------------------------------------------------------- ValueError                                Traceback (most recent call\n  last)  in ()\n  ----> 1 estimator.train(input_fn=input_func, steps=1000)\n        2 #eval_metrics = estimator.evaluate(input_fn=eval_input_func, steps=1000)<\/p>\n  \n  <p>8 frames\n  \/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/feature_column\/feature_column_v2.py\n  in <strong>init<\/strong>(self, feature_columns, units, sparse_combiner, trainable,\n  name, **kwargs)\n      498         raise ValueError(\n      499             'Items of feature_columns must be either a '\n  --> 500             'DenseColumn or CategoricalColumn. Given: {}'.format(column))\n      501 \n      502     self._units = units<\/p>\n  \n  <p>ValueError: Items of feature_columns must be either a DenseColumn or\n  CategoricalColumn. Given: SequenceNumericColumn(key='x', shape=(1,),\n  default_value=0.0, dtype=tf.float32, normalizer_fn=None)<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"58004067","CreationDate":"2019-09-19T05:08:21.000","ParentId":null,"OwnerUserId":"5136550","Title":null,"Body":"<p>The following code works fine for training and prediction.<\/p>\n\n<pre><code>x_data = np.linspace(0, 10.0, 1000)\nprint(x_data.shape)\ny_true = (0.5 * x_data) + 5\nprint(y_true.shape)\nx_train, x_eval, y_train, y_eval = train_test_split(x_data, y_true, test_size=0.25, random_state=101)\ntrain_func = tf.estimator.inputs.numpy_input_fn({'x': x_train}, y_train, batch_size=8, num_epochs=10, shuffle=True)\n\nfeatures = [tf.contrib.layers.real_valued_column(\"x\", dimension=1)]\nestimator = tf.estimator.LinearRegressor(feature_columns=features)\n\nestimator.train(input_fn=train_func, steps=100) # Fit the model to training data.\n\neval_func = tf.estimator.inputs.numpy_input_fn({'x': x_eval}, batch_size=1, num_epochs=1, shuffle=False)\n\nresult = estimator.predict(eval_func) # Predict scores\n\nprint(\"predict_scores\", list(result))\n<\/code><\/pre>\n"}]}
+{"QuestionId":58003475,"AnswerCount":0,"Tags":"<python><tensorflow><flask><keras>","CreationDate":"2019-09-19T03:50:27.160","AcceptedAnswerId":null,"OwnerUserId":741099.0,"Title":"Tensorflow.Session is None if Tensorflow Object is Saved to Global variable and Retrieved","Body":"<p><strong>Scenario:<\/strong> <\/p>\n\n<p>I am encountering problems with a Flask app that tries to save a Tensorflow model from disk into memory (a global variable <code>generator<\/code>) inside <code>load_model()<\/code> when the app first starts. <\/p>\n\n<p>When a user visits the <code>\/test<\/code> endpoint, the preloaded model <code>generator<\/code> will be used to generate some data.<\/p>\n\n<p><strong>Problem:<\/strong> <\/p>\n\n<p><code>generator.run()<\/code> works properly when it is being called in <code>load_model()<\/code> when the app first starts. However the exact same code for <code>generator.run()<\/code> throws an error when it is called inside <code>test()<\/code>.<\/p>\n\n<p>I managed to narrow down the problem occurring at the <code>tf.compat.v1.Session()<\/code> variable <code>sess<\/code>. <\/p>\n\n<p>When called from <code>load_model()<\/code>, <code>sess<\/code> is <code>&lt;tensorflow.python.client.session.Session object at 0x12d317fd0&gt;<\/code>. But when called from <code>test()<\/code>, <code>sess<\/code> is <code>None<\/code>.<\/p>\n\n<p>Anyone knows how to solve this problem? It is highly preferred to load the huge model to memory just once (~10s load time), then load it everytime the endpoint is queried.<\/p>\n\n<p>Thank you!<\/p>\n\n<pre><code>from flask import Flask\n\napp = Flask(__name__)\ngenerator = None\n\n# EVERYTHING WORKS WELL HERE\ndef load_model():\n    tflib.init_tf()\n\n    # Load model from disk\n    with open(model_path, \"rb\") as f:\n        _G, _D, Gs = pickle.load(f, encoding='latin1')\n\n    # Update global variable generator\n    global generator\n    generator = Gs\n\n    # Run model\n    latent = np.random.randn(1, generator.input_shape[1])\n    img = generator.run(latent)[0]              # `sess` is &lt;tensorflow.python.client.session.Session object at 0x12d317fd0&gt;\n    print(img.shape)                            # prints: (512, 512, 3)\n\n\n# PROBLEM OCCURS HERE\n@app.route('\/test')\ndef test():\n    print('generator: ', generator)             # prints: &lt;dnnlib.tflib.network.Network object at 0x13a80e5d0&gt;\n    print('generator.run: ', generator.run)     # prints: &lt;bound method Network.run of &lt;dnnlib.tflib.network.Network object at 0x13a80e5d0&gt;&gt;\n\n    # Run model\n    latent = np.random.randn(1, generator.input_shape[1])\n    img = generator.run(latent)[0]              # `sess` is None\n\n\nif __name__ == '__main__':\n    load_model()                                # Loads the model from disk\n    app.run()\n<\/code><\/pre>\n\n<p><strong>Error after visiting <code>\/test<\/code><\/strong><\/p>\n\n<pre><code>[2019-09-18 23:26:33,083] ERROR in app: Exception on \/test [GET]\nTraceback (most recent call last):\n  File \"\/anaconda3\/envs\/ml\/lib\/python3.7\/site-packages\/flask\/app.py\", line 2446, in wsgi_app\n    response = self.full_dispatch_request()\n  File \"\/anaconda3\/envs\/ml\/lib\/python3.7\/site-packages\/flask\/app.py\", line 1951, in full_dispatch_request\n    rv = self.handle_user_exception(e)\n  File \"\/anaconda3\/envs\/ml\/lib\/python3.7\/site-packages\/flask\/app.py\", line 1820, in handle_user_exception\n    reraise(exc_type, exc_value, tb)\n  File \"\/anaconda3\/envs\/ml\/lib\/python3.7\/site-packages\/flask\/_compat.py\", line 39, in reraise\n    raise value\n  File \"\/anaconda3\/envs\/ml\/lib\/python3.7\/site-packages\/flask\/app.py\", line 1949, in full_dispatch_request\n    rv = self.dispatch_request()\n  File \"\/anaconda3\/envs\/ml\/lib\/python3.7\/site-packages\/flask\/app.py\", line 1935, in dispatch_request\n    return self.view_functions[rule.endpoint](**req.view_args)\n  File \"app.py\", line 71, in test\n    img = generator.run(latent)[0]\n  File \"\/Users\/x\/foo\/dnnlib\/tflib\/network.py\", line 460, in run\n    mb_out = sess.run(out_expr, dict(zip(in_expr, mb_in)))\nAttributeError: 'NoneType' object has no attribute 'run'\n<\/code><\/pre>\n\n<p><strong>Code around <code>sess.run()<\/code><\/strong> <a href=\"https:\/\/gist.github.com\/nyxynyx\/a8ef6069d03fc757ccf09d2f9357c981#file-dnnlib-tflib-network-py-L453-L455\" rel=\"nofollow noreferrer\">[Full file <code>network.py<\/code>]<\/a><\/p>\n\n<pre><code>...\n        for mb_begin in range(0, num_items, minibatch_size):\n            if print_progress:\n                print(\"\\r%d \/ %d\" % (mb_begin, num_items), end=\"\")\n\n            mb_end = min(mb_begin + minibatch_size, num_items)\n            mb_num = mb_end - mb_begin\n            mb_in = [src[mb_begin : mb_end] if src is not None else np.zeros([mb_num] + shape[1:]) for src, shape in zip(in_arrays, self.input_shapes)]\n            #\u8dd1\n            # [&lt;tf.Tensor 'Gs\/_Run\/concat:0' shape=(?, 3, 512, 512) dtype=float32&gt;]\n\n            # &lt;tf.Tensor 'Gs\/_Run\/labels_in:0' shape=&lt;unknown&gt; dtype=float32&gt;: array([], shape=(1, 0), dtype=float64)\n            # &lt;tf.Tensor 'Gs\/_Run\/latents_in:0' shape=&lt;unknown&gt; dtype=float32&gt;: latents.shape (1, 512)\n            #mb_out = tf.get_default_session().run(out_expr, dict(zip(in_expr, mb_in)))\n            sess = tf.get_default_session()\n            #writer = tf.summary.FileWriter(\"logs\/\", sess.graph)\n            mb_out = sess.run(out_expr, dict(zip(in_expr, mb_in)))\n\n            for dst, src in zip(out_arrays, mb_out):\n                dst[mb_begin: mb_end] = src\n...\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58003739,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm>","CreationDate":"2019-09-19T04:29:18.927","AcceptedAnswerId":null,"OwnerUserId":10834644.0,"Title":"Tensorflow 2.0 Combine CNN + LSTM","Body":"<p>How can you add an LSTM Layer after (flattened) conv2d Layer in Tensorflow 2.0 \/ Keras? My Training input data has the following shape (size, sequence_length, height, width, channels). For a convolutional layer, I can only process one image a a time, for the LSTM Layer I need a sequence of features. Is there a way to reshape your data before the LSTM Layer, so you can combine both?<\/p>\n","answers":[{"AnswerId":"58004719","CreationDate":"2019-09-19T06:12:58.490","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>From an overview of shape you have provided which is <code>(size, sequence_length, height, width, channels)<\/code>, it appears that you have sequences of images for each label. For this purpose, we usually make use of <code>Conv3D<\/code>. I am enclosing a sample code below:<\/p>\n\n<pre><code>import tensorflow as tf\n\nSIZE = 64\nSEQUENCE_LENGTH = 50\nHEIGHT = 128\nWIDTH = 128\nCHANNELS = 3\n\ndata = tf.random.normal((SIZE, SEQUENCE_LENGTH, HEIGHT, WIDTH, CHANNELS))\n\ninput = tf.keras.layers.Input((SEQUENCE_LENGTH, HEIGHT, WIDTH, CHANNELS))\nhidden = tf.keras.layers.Conv3D(32, (3, 3, 3))(input)\nhidden = tf.keras.layers.Reshape((-1, 32))(hidden)\nhidden = tf.keras.layers.LSTM(200)(hidden)\n\nmodel = tf.keras.models.Model(inputs=input, outputs=hidden)\nmodel.summary()\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>Model: \"model\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_1 (InputLayer)         [(None, 50, 128, 128, 3)] 0         \n_________________________________________________________________\nconv3d (Conv3D)              (None, 48, 126, 126, 32)  2624      \n_________________________________________________________________\nreshape (Reshape)            (None, None, 32)          0         \n_________________________________________________________________\nlstm (LSTM)                  (None, 200)               186400    \n=================================================================\nTotal params: 189,024\nTrainable params: 189,024\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>If you still wanted to make use of <code>Conv2D<\/code> which is not recommended in your case, you will have to do something like shown below. Basically, you are appending the sequence of images across the height dimension, which will make you to loose temporal dimensions.<\/p>\n\n<pre><code>import tensorflow as tf\n\nSIZE = 64\nSEQUENCE_LENGTH = 50\nHEIGHT = 128\nWIDTH = 128\nCHANNELS = 3\n\ndata = tf.random.normal((SIZE, SEQUENCE_LENGTH, HEIGHT, WIDTH, CHANNELS))\n\ninput = tf.keras.layers.Input((SEQUENCE_LENGTH, HEIGHT, WIDTH, CHANNELS))\nhidden = tf.keras.layers.Reshape((SEQUENCE_LENGTH * HEIGHT, WIDTH, CHANNELS))(input)\nhidden = tf.keras.layers.Conv2D(32, (3, 3))(hidden)\nhidden = tf.keras.layers.Reshape((-1, 32))(hidden)\nhidden = tf.keras.layers.LSTM(200)(hidden)\n\nmodel = tf.keras.models.Model(inputs=input, outputs=hidden)\nmodel.summary()\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>Model: \"model\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_1 (InputLayer)         [(None, 50, 128, 128, 3)] 0         \n_________________________________________________________________\nreshape (Reshape)            (None, 6400, 128, 3)      0         \n_________________________________________________________________\nconv2d (Conv2D)              (None, 6398, 126, 32)     896       \n_________________________________________________________________\nreshape_1 (Reshape)          (None, None, 32)          0         \n_________________________________________________________________\nlstm (LSTM)                  (None, 200)               186400    \n=================================================================\nTotal params: 187,296\nTrainable params: 187,296\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n"}]}
+{"QuestionId":58004256,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-estimator>","CreationDate":"2019-09-19T05:30:32.827","AcceptedAnswerId":null,"OwnerUserId":4859841.0,"Title":"How does tensorflow Estimator API reuse variables across batches?","Body":"<p>so for each Estimator we need to define:<\/p>\n\n<pre><code>def model_fn(\n   features, # This is batch_features from input_fn\n   labels,   # This is batch_labels from input_fn\n   mode,     # An instance of tf.estimator.ModeKeys\n   params):  # Additional configuration\n<\/code><\/pre>\n\n<p>This <code>model_fn<\/code> directly takes the features and use it:<\/p>\n\n<pre><code>...\nwith tf.variable_scope(\"model\") as scope:\n  weight = tf.get_variable(\"var1\", [output_size, batch_size])\n  output = features[\"input\"] * weight\n...\n<\/code><\/pre>\n\n<p>So my question is the variables like the <code>weight<\/code> variable seems to be created every time for each new batch, and how does the tensorflow Estimator reuse the same variable across batches?\nAnd do we have to use the <code>tf.get_variable()<\/code> not the <code>tf.Variable()<\/code> inside the Estimator in order to make the training reuse the variable across batches?<\/p>\n","answers":[]}
+{"QuestionId":58004563,"AnswerCount":2,"Tags":"<python><machine-learning><keras><neural-network><lstm>","CreationDate":"2019-09-19T05:59:14.647","AcceptedAnswerId":null,"OwnerUserId":10977971.0,"Title":"How to set up LSTM network for predict multi-sequence?","Body":"<p>I am learning how to set up the RNN-LSTM network for prediction. I have created the dataset with one input variable.<\/p>\n\n<pre><code>x  y\n1  2.5\n2  6\n3  8.6\n4  11.2\n5  13.8\n6  16.4\n...\n<\/code><\/pre>\n\n<p>By the following python code, I have created the window data, like [x(t-2), x(t-1), x(t)] to predict [y(t)]:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>df= pd.read_excel('dataset.xlsx')\n\n# split a univariate dataset into train\/test sets\ndef split_dataset(data):\n    train, test = data[:-328], data[-328:-6]\n    return train, test\n\ntrain, test  = split_dataset(df.values)\n\n# scale train and test data\ndef scale(train, test):\n    # fit scaler\n    scaler = MinMaxScaler(feature_range=(0,1))\n    scaler = scaler.fit(train)\n    # transform train\n    #train = train.reshape(train.shape[0], train.shape[1])\n    train_scaled = scaler.transform(train)\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\nscaler, train_scaled, test_scaled = scale(train, test)\n\ndef to_supervised(train, n_input, n_out=7):\n    # flatten data\n    data = train\n    X, y = list(), list()\n    in_start = 0\n    # step over the entire history one time step at a time\n    for _ in range(len(data)):\n        # define the end of the input sequence\n        in_end = in_start + n_input\n        out_end = in_end + n_out\n        # ensure we have enough data for this instance\n        if out_end &lt;= len(data):\n            x_input = data[in_start:in_end, 0]\n            x_input = x_input.reshape((len(x_input), 1))\n            X.append(x_input)\n            y.append(data[in_end:out_end, 0])\n        # move along one time step\n        in_start += 1\n    return np.array(X), np.array(y)\ntrain_x, train_y = to_supervised(train_scaled, n_input = 3, n_out = 1)\ntest_x, test_y =  to_supervised(test_scaled, n_input = 3, n_out = 1)\n\nverbose, epochs, batch_size = 0, 20, 16\nn_timesteps, n_features, n_outputs = train_x.shape[1], train_x.shape[2], train_y.shape[1]\n\n\nmodel = Sequential()\nmodel.add(LSTM(200, return_sequences= False, input_shape = (train_x.shape[1],train_x.shape[2])))\nmodel.add(Dense(1))\nmodel.compile(loss = 'mse', optimizer = 'adam')\nhistory = model.fit(train_x, train_y, epochs=epochs, verbose=verbose, validation_data = (test_x, test_y))\n<\/code><\/pre>\n\n<p>However, I have other questions about this:<\/p>\n\n<p>Q1: What is the meaning of units in LSTM? [model.add(LSTM(units, ...))]<\/p>\n\n<p>(I have tried different units for the model, it would be more accurate as units increased.)<\/p>\n\n<p>Q2: How many layers should I set?<\/p>\n\n<p>Q3: How can I predict multi-steps ? e.g base on (x(t),x(t-1)) to predict y(t), y(t+1) I have tried to set the n_out = 2 in the to_supervised function, but when I applied the same method, it returned the error<\/p>\n\n<pre><code>train_x, train_y = to_supervised(train_scaled, n_input = 3, n_out = 2)\ntest_x, test_y =  to_supervised(test_scaled, n_input = 3, n_out = 2)\n\nverbose, epochs, batch_size = 0, 20, 16\nn_timesteps, n_features, n_outputs = train_x.shape[1], train_x.shape[2], train_y.shape[1]\n\nmodel = Sequential()\nmodel.add(LSTM(200, return_sequences= False, input_shape = (train_x.shape[1],train_x.shape[2])))\nmodel.add(Dense(1))\nmodel.compile(loss = 'mse', optimizer = 'adam')\nhistory = model.fit(train_x, train_y, epochs=epochs, verbose=verbose, validation_data = (test_x, test_y))\n<\/code><\/pre>\n\n<pre><code>ValueError: Error when checking target: expected dense_27 to have shape (1,) but got array with shape (2,)\n<\/code><\/pre>\n\n<p>Q3(cont): What should I add or change in the model setting?<\/p>\n\n<p>Q3(cont): What is the return_sequences ? When should I set True?<\/p>\n","answers":[{"AnswerId":"58009720","CreationDate":"2019-09-19T11:13:14.073","ParentId":null,"OwnerUserId":"9742432","Title":null,"Body":"<p>Re Q1: It is the number of LSTM cells (=LSTM units), which consist of several neurons themselves but have (in the standard case as given) only one output each. Thus, the number of units corresponds directly to the dimensionality of your output.<\/p>\n"},{"AnswerId":"58006741","CreationDate":"2019-09-19T08:23:56.620","ParentId":null,"OwnerUserId":"5143695","Title":null,"Body":"<p>Q1. Units in LSTM is the number of neurons in your LSTM layer.<\/p>\n\n<p>Q2. That depends on your model \/ data. Try changing them around to see the effect.<\/p>\n\n<p>Q3. That depends which apporach you take.<\/p>\n\n<p>Q4. Ideally you'll want to predict a single time step every time.\nIt is possible to predict several at a time, but in my experience you will get better results like as i have described below<\/p>\n\n<p>e.g<\/p>\n\n<p>use y(t-1), y(t) to predict y_hat(t+1)<\/p>\n\n<p>THEN<\/p>\n\n<p>use y(t),  y_hat(t+1) to predict y_hat(t+2)<\/p>\n\n<p>Are you sure you're actually using X to predict Y in this case?\nhow does train x\/y and test x\/y look like?<\/p>\n"}]}
+{"QuestionId":58004850,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-19T06:23:33.280","AcceptedAnswerId":null,"OwnerUserId":12088605.0,"Title":"TF2.0RC1\/RC0 the output shape of custom layer is None. The same code works in TF2.0 Beta1","Body":"<p>In this layer, we use a custom c++ op writen by ourself. Is there any tutorial for defining a custom c++ op for tf2.0 rc1\uff08our c++ op works in tf2.0 beta1\uff09?<\/p>\n","answers":[]}
+{"QuestionId":58005117,"AnswerCount":1,"Tags":"<python><tensorflow><image-processing><boost><tree>","CreationDate":"2019-09-19T06:42:33.533","AcceptedAnswerId":58008493.0,"OwnerUserId":11377163.0,"Title":"How to us Tensorflow boosted tree for image clasification task?","Body":"<p>I was doing image classification using CNN models. I would like to try image classification using decision tree or boosted tree models. I found <strong>Tensorflow<\/strong> as Boosted tree model, but I'm not able to understand how to give image as an input to the model. If you know how to train image classification using <em>tf.boosted<\/em> tree please guide me.<\/p>\n","answers":[{"AnswerId":"58008493","CreationDate":"2019-09-19T10:01:42.853","ParentId":null,"OwnerUserId":"5136550","Title":null,"Body":"<p>The simplest way is to use the flatten image feature vector as input. You can use the following code sample that i have tested with mnist Dataset (modified for 2 class). Note that current implementation of tensorflow BoostedTreesClassifier doesn't support pruning for multi class classifier.<\/p>\n\n<pre><code>import tensorflow as tf\n\ntf.enable_eager_execution()\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\nfor i in range(len(y_train)):\n    if y_train[i] is not 0:\n        y_train[i] = 1\n\nfor i in range(len(y_test)):\n    if y_test[i] is not 0:\n        y_test[i] = 1\n\nNUM_FEATURES = 28*28\n\ntrain_input_fn = tf.estimator.inputs.numpy_input_fn({'x': x_train.reshape(-1,NUM_FEATURES)}, y_train, batch_size=128, num_epochs=5, shuffle=True)\neval_input_fn = tf.estimator.inputs.numpy_input_fn({'x': x_test.reshape(-1,NUM_FEATURES)}, y_test, batch_size=128, num_epochs=1, shuffle=False)\n\nfeatures = [tf.feature_column.numeric_column(\"x\", shape=(NUM_FEATURES,))]\n\nest = tf.estimator.BoostedTreesClassifier(features, n_batches_per_layer=1)\nest.train(train_input_fn)\nresults = est.evaluate(eval_input_fn)\nprint('Accuracy : ', results['accuracy'])\n<\/code><\/pre>\n"}]}
+{"QuestionId":58005297,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-09-19T06:53:14.060","AcceptedAnswerId":null,"OwnerUserId":8186207.0,"Title":"PyTorch having trouble detecting CUDA","Body":"<p>I am running CNN on PyTorch. The torch.cuda.is_available() function returned false and no GPU is detected. However, I can run Keras model with GPU. Here is my system information:<\/p>\n\n<ul>\n<li>OS: Ubuntu 18.04.3<\/li>\n<li>Python 3.7.3 (Conda)<\/li>\n<li>GPU: GTX1080Ti<\/li>\n<li>Nvidia driver: 430.50<\/li>\n<\/ul>\n\n<p>When I check nvidia-smi, the output said that the CUDA version is 10.1. However, the nvcc -V command tells me that it is CUDA 9.1.<\/p>\n\n<p>I downloaded NVIDIA-Linux-x86_64-430.50.run from the official site and install it with command line. I installed CUDA 10.1 using these following command line recommended by the official site:<\/p>\n\n<pre><code>wget http:\/\/developer.download.nvidia.com\/compute\/cuda\/10.1\/Prod\/local_installers\/cuda_10.1.243_418.87.00_linux.run\nsudo sh cuda_10.1.243_418.87.00_linux.run\n<\/code><\/pre>\n\n<p>I installed PyTorch through pip install. What is wrong? Thanks in advance!<\/p>\n","answers":[{"AnswerId":"58005585","CreationDate":"2019-09-19T07:12:10.503","ParentId":null,"OwnerUserId":"2650622","Title":null,"Body":"<p>The default Pytorch 1.2 package depends on CUDA 10.0, but you have CUDA 9.1. The output of <code>nvidia-smi<\/code> just tells you the maximum CUDA version your GPU supports, <code>nvcc<\/code> gives the CUDA installed on your system. It seems that your installation of CUDA 10.1 was unsuccessful.<\/p>\n\n<p>In addition to CUDA 10.0, Pytorch also supports CUDA 9.2 and I've found that the Pytorch package compiled for CUDA 10.0 also works with CUDA 10.1. So you can either upgrade your CUDA installation to 9.2 and install the Pytorch CUDA 9.2 package with<\/p>\n\n<pre><code>pip3 install torch==1.2.0+cu92 torchvision==0.4.0+cu92 -f https:\/\/download.pytorch.org\/whl\/torch_stable.html\n<\/code><\/pre>\n\n<p>Or get a working installation of CUDA 10.1. There are detailed Linux instructions <a href=\"https:\/\/docs.nvidia.com\/cuda\/cuda-installation-guide-linux\/index.html\" rel=\"nofollow noreferrer\">here.<\/a> (Note that you may have to remove previous installations of CUDA before installing a new one.)<\/p>\n"}]}
+{"QuestionId":58005443,"AnswerCount":1,"Tags":"<tensorflow><object-detection><object-detection-api>","CreationDate":"2019-09-19T07:02:01.403","AcceptedAnswerId":58010514.0,"OwnerUserId":2585410.0,"Title":"Does one step in Object Detection API mean processing one picture or one bounding box?","Body":"<p>In pipeline.config file in <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/object_detection\" rel=\"nofollow noreferrer\">Tensorflow Object Detection API<\/a> we have parameter <code>NUM_STEPS<\/code>. <\/p>\n\n<p>Does <strong>one step<\/strong> mean processing one whole picture, or one bounding box? <\/p>\n\n<p>In the config file, we have:<\/p>\n\n<pre><code>model {\n  faster_rcnn {\n    # (...)\n  }\n\n  train_config: {\n    batch_size: 1\n    optimizer {\n      # (...)\n    }\n    gradient_clipping_by_norm: 10.0\n    # (...)\n    num_steps: 200000  # &lt;-- HERE IT IS\n    # (...)\n  }\n}\n<\/code><\/pre>\n\n<p>E.g. We've got a training TFRecord with 2 pictures, 10 bboxes each. If I have NUM_STEPS set to 10, does this mean, that I would process first 10 bboxes, or each photo 5 times? <\/p>\n\n<p>Full config file can be found here:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/32dadfc2def4f05faeedacce98e4c4099be4c433\/research\/object_detection\/samples\/configs\/faster_rcnn_inception_v2_coco.config#L113\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/32dadfc2def4f05faeedacce98e4c4099be4c433\/research\/object_detection\/samples\/configs\/faster_rcnn_inception_v2_coco.config#L113<\/a><\/p>\n","answers":[{"AnswerId":"58010514","CreationDate":"2019-09-19T11:57:37.323","ParentId":null,"OwnerUserId":"3871961","Title":null,"Body":"<p>One 'step' corresponds to one batch processing.<\/p>\n\n<p>The input of faster-RCNN is a full image and your  batch size is 1, thus it means that you are using one image each time. In your case, the first step will process the five box of the first image and the second step the five of the second one.<\/p>\n"}]}
+{"QuestionId":58005484,"AnswerCount":0,"Tags":"<python><pandas><tensorflow><optimization>","CreationDate":"2019-09-19T07:04:37.687","AcceptedAnswerId":null,"OwnerUserId":9997633.0,"Title":"I would like to Manipulate data under a column by changing the values that are both strings and integers","Body":"<p>I'm trying to change the data of under a column in pandas. The data is both strings and integers (i.e. IT-1). It's large data structure so changing each item individually would not be ideal. Please Help me if you know of any loop or structure i could use. I'm really stuck. Thank you.<\/p>\n\n<h1>This is what i tried but clearly doing this for 10,000 items would not be effective<\/h1>\n\n<pre><code>s = df.item\n\n\ndf.at[0:2, 'item'] = 'IT-0','IT-1','IT-2'\n\n\ndisplay(s)\n<\/code><\/pre>\n\n<p>In the results below you can see that i can iterate over the first 3 items using the mechanical method. Some kind of loop structure would really be helpful<\/p>\n\n<pre><code>0         IT-0\n\n1         IT-1\n\n2         IT-2\n\n3         IT-8\n\n4         IT-56\n         ...\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58006304,"AnswerCount":0,"Tags":"<java><android><tensorflow><neural-network><speech-recognition>","CreationDate":"2019-09-19T07:55:14.507","AcceptedAnswerId":null,"OwnerUserId":12088791.0,"Title":"How to implement three first layers of network from Simple Audio Recognition Tensorflow tutorial in java","Body":"<p>I'm trying to use model trained using Simple Audio Recognition Tensorlow tutorial in Android application, but I want to implement three first layers not in network, but on CPU side. Basicly I used (without changing anything) training from Tensorflow tutorial, but freezed model without decoding sample data, calculating spectrogram and mfcc:<\/p>\n\n<p>wav_data_placeholder<\/p>\n\n<pre><code>Tensor(\"wav_data:0, shape=(), dtype=string)\n<\/code><\/pre>\n\n<p>decoded_sample_data<\/p>\n\n<pre><code>DecodeWav(audio=(tf.Tensor 'decoded_sample_data:0' shape=(16000, 1) dtype=flat32), sample_rate(tf.Tensor 'decoded_sample_data:1' shape=() dtype=int32))\n<\/code><\/pre>\n\n<p>spectrogram<\/p>\n\n<pre><code>Tensor(\"AudioSpectrogram:0, shape=(1, 98, 257), dtype=float32)\n<\/code><\/pre>\n\n<p>fingerprint_input<\/p>\n\n<pre><code>Tensor(\"Mfcc:0, shape=(1, 98, 40), dtype=float32)\n<\/code><\/pre>\n\n<hr>\n\n<p>So my freezed model has input shape (1, 98, 40) and takes values that are calculated by the mfcc method which has input values from spectrogram output which has input from decoded sample data.<\/p>\n\n<p>I tried librosa mfcc implementation in java, but unfortunately, it does not provide the same results. <\/p>\n\n<p>How can I compute my .wav file in Android application to successfully satisfy my model input to get the right predictions? <\/p>\n","answers":[]}
+{"QuestionId":58007127,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-09-19T08:46:20.977","AcceptedAnswerId":58008361.0,"OwnerUserId":1295035.0,"Title":"Pytorch differentiable conditional (index-based) sum","Body":"<p>I have an <code>idx<\/code> array like <code>[0, 1, 0, 2, 3, 1]<\/code> and another 2d array <code>data<\/code> like the following:<\/p>\n\n<pre><code>[[0,  1,  2],\n [3,  4,  5],\n [6,  7,  8],\n [9,  10, 11],\n [12, 13, 14],\n [15, 16, 17]]\n<\/code><\/pre>\n\n<p>I want my output to be <code>4x3<\/code> in which 4 is the max of <code>idx<\/code> and 3 is the feature size (<code>data.shape[1]<\/code>) and in the output each element is the sum of features with corresponding index in the <code>idx<\/code>. Then output in this example would be:<\/p>\n\n<pre><code>[[6,  8,  10],\n [18, 20, 22],\n [9,  10, 11],\n [12, 13, 14]]\n<\/code><\/pre>\n\n<p>I can do it with iterating over <code>range(3)<\/code> and creating a mask on data and summing up them but it's not differentiable (I suppose). Is there any function(s) in Pytorch for this purpose? something like <code>scatter()<\/code>.<\/p>\n\n<p>Update: It seems I am looking for something named <em>scatter sum<\/em> which is implemented in <a href=\"https:\/\/github.com\/rusty1s\/pytorch_scatter\" rel=\"nofollow noreferrer\">this<\/a> repository.<\/p>\n","answers":[{"AnswerId":"58008361","CreationDate":"2019-09-19T09:54:06.557","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You are looking for <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.Tensor.index_add_\" rel=\"nofollow noreferrer\"><code>index_add_<\/code><\/a>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\n\nx = torch.tensor([[ 0.,  1.,  2.],\n        [ 3.,  4.,  5.],\n        [ 6.,  7.,  8.],\n        [ 9., 10., 11.],\n        [12., 13., 14.],\n        [15., 16., 17.]], dtype=torch.float)\nidx = torch.tensor([0, 1, 0, 2, 3, 1], dtype=torch.long)  # note the dtype here, must be \"long\"\n# init the sums to zero\ny = torch.zeros((idx.max()+1, x.shape[1]), dtype=x.dtype)\n\n# do the magic\ny.index_add_(0, idx, x)\n<\/code><\/pre>\n\n<p>Gives the desired output<\/p>\n\n<blockquote>\n<pre><code>tensor([[ 6.,  8., 10.],\n        [18., 20., 22.],\n        [ 9., 10., 11.],\n        [12., 13., 14.]])\n<\/code><\/pre>\n<\/blockquote>\n"}]}
+{"QuestionId":58007140,"AnswerCount":0,"Tags":"<tensorflow><keras><tensorflow-estimator>","CreationDate":"2019-09-19T08:47:13.073","AcceptedAnswerId":null,"OwnerUserId":10867166.0,"Title":"How to print metric value when training a custom kera model using estimator?","Body":"<p>I've defined a custom keras model using <code>Model subclass<\/code>. And I want to convert this model to estimator for distributed training. Simple codes are as follow:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    tf.logging.set_verbosity(tf.logging.INFO)\n\n    model = EstimatorModel()\n\n    model.compile(optimizer=tf.keras.optimizers.Adam(),\n                  loss='categorical_crossentropy',\n                  metrics=['accuracy'])\n\n    estimator_model = tf.keras.estimator.model_to_estimator(\n        keras_model=model, model_dir=FLAGS.model_path)\n\n    estimator_model.train(input_fn=input_fn)\n<\/code><\/pre>\n\n<p>Code can run normally. Now, the question is the estimator only print the <code>loss<\/code>, and didn't print the <code>metric<\/code>. I've tried to add metric to <code>estimator_model<\/code> using <code>tf.estimator.add_metrics(estimator_model, my_auc)<\/code>, but it still not work. What should I do to solve this? <\/p>\n","answers":[]}
+{"QuestionId":58007391,"AnswerCount":1,"Tags":"<neural-network><nlp><pytorch><transformer><attention-model>","CreationDate":"2019-09-19T09:01:32.667","AcceptedAnswerId":58007592.0,"OwnerUserId":11237667.0,"Title":"Attention Text Generation in Character-by-Character fashion","Body":"<p>I am searching the web for a couple of days for any <strong>text generation<\/strong> model that would use only attention mechanisms.<\/p>\n\n<p>The <strong>Transformer<\/strong> architecture that made waves in the context of <strong>Seq-to-Seq<\/strong> models is actually based solely on <strong>Attention<\/strong> mechanisms but is mainly designed and used for translation or chat bot tasks so it doesn't fit to the purpose, but the principle does.<\/p>\n\n<p>My question is:<\/p>\n\n<p>Does anyone knows or heard of a text generation model <strong>based solely on Attention without any recurrence<\/strong>?<\/p>\n\n<p>Thanks a lot!<\/p>\n\n<p>P.S. I'm familiar with <strong>PyTorch<\/strong>.<\/p>\n","answers":[{"AnswerId":"58007592","CreationDate":"2019-09-19T09:12:21.517","ParentId":null,"OwnerUserId":"5652313","Title":null,"Body":"<p>Building a character-level self-attentive model is a challenging task. Character-level models are usually based on RNNs. Whereas in a word\/subword model, it is clear from the beginning what are the units carrying meaning (and therefore the units the attention mechanism can attend to), a character-level model needs to learn word meaning in the following layers. This makes it quite difficult for the model to learn.<\/p>\n\n<p>Text generation models are nothing more than conditional languages model. Google AI recently published a paper on <a href=\"https:\/\/arxiv.org\/abs\/1908.10322\" rel=\"nofollow noreferrer\">Transformer character language model<\/a>, but it is the only work I know of.<\/p>\n\n<p>Anyway, you should consider either using subwords units (as BPE, SentencePiece) or if you really need to go for character level, use RNNs instead.<\/p>\n"}]}
+{"QuestionId":58007478,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-19T09:05:44.783","AcceptedAnswerId":null,"OwnerUserId":1889418.0,"Title":"CNN with Keras - accurary incredibly low and loss negative - clearly an error on my end","Body":"<p>I am attempting to build a CNN to classify skin-cancer related images into seven categories.\nI'm relatively new to the concept of CNN and have been adapting the dog\/cat classification use-case to a known skin-cancer database challenge.\nThe issue is that the loss and accuracy are extremely low as well as constant through the epochs.\nI'm not sure where the problem is, though - my first assumption is that the number of images used is too low: 436 samples for training and 109 validation. I reduce the number of images used from 10000+ because i'm using my macbook pro.<\/p>\n\n<p>script:<\/p>\n\n<pre><code>    import tensorflow as tf\n    from tensorflow.keras.models import Sequential\n    from tensorflow.keras.layers import Dense, Dropout, Activation, Flatten, Conv2D, MaxPooling2D\n\n    import numpy as np\n    import pandas as pd\n\n    import matplotlib.pyplot as plt\n\n    import sys\n    import os\n    import cv2\n\n    DATA_DIR = \"\/Users\/namefolder\/PycharmProjects\/skin-cancer\/HAM10000_images_part_1\"\n\n    metadata = pd.read_csv(os.path.join(DATA_DIR, 'HAM10000_metadata.csv'))\n\n    lesion_type_dict = {'nv': 'Melanocytic nevi',\n        'mel': 'Melanoma',\n        'bkl': 'Benign keratosis-like lesions ',\n        'bcc': 'Basal cell carcinoma',\n        'akiec': 'Actinic keratoses',\n        'vasc': 'Vascular lesions',\n        'df': 'Dermatofibroma'}\n\n    metadata['cell_type'] = metadata['dx'].map(lesion_type_dict.get)\n    metadata['dx_code'] = pd.Categorical(metadata['dx']).codes\n\n    # save array of image-id and diagnosis-type (categorical)\n    metadata = metadata[['image_id', 'dx', 'dx_type', 'dx_code']]\n\n    training_data = []\n\n    IMG_SIZE=40\n\n    # preparing training data\n\n    def creating_training_data(path):\n        for img in os.listdir(path):\n            try:\n                img_array = cv2.imread(os.path.join(path, img), cv2.IMREAD_GRAYSCALE)\n                new_array = cv2.resize(img_array, (IMG_SIZE, IMG_SIZE))\n                for index, row in metadata.iterrows():\n                    if img == row['image_id']+'.jpg':\n                        try:\n                            training_data.append([new_array, row['dx_code']])\n                        except Exception as ee:\n                            pass\n            except Exception as e:\n                pass\n\n        return training_data\n\n    training_data = creating_training_data(DATA_DIR)\n\n    import random\n\n    random.shuffle(training_data)\n\n    # Splitting data into X features and Y label\n    X_train = []\n    y_train = []\n    for features, label in training_data:\n        X_train.append(features)\n        y_train.append(label)\n\n    # Reshaping of the data - required by Tensorflow and Keras (*necessary step of deep-learning using these repos)\n    X_train = np.array(X_train).reshape(-1, IMG_SIZE, IMG_SIZE, 1)\n\n    # Normalize data - to reduce processing requirements\n    X_train = X_train\/255.0\n\n    # model configuration\n    model = Sequential()\n    model.add(Conv2D(64, (3,3), input_shape = X_train.shape[1:]))\n    model.add(Activation(\"relu\"))\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n\n    model.add(Conv2D(64, (3,3)))\n    model.add(Activation(\"relu\"))\n    model.add(MaxPooling2D(pool_size=(2, 2)))\n\n    model.add(Flatten())\n    model.add(Dense(64))\n\n    model.add(Dense(1))\n    model.add(Activation(\"softmax\"))\n\n    model.compile(loss=\"mean_squared_error\",\n                 optimizer=\"adam\",\n                 metrics=[\"accuracy\"])\n<\/code><\/pre>\n\n<p>Training model:<\/p>\n\n<pre><code> Model fitting output:\n\n            Train on 436 samples, validate on 109 samples\n            Epoch 1\/20\n            436\/436 [==============================] - 1s 2ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 2\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 3\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 4\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 5\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 6\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 7\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 8\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 9\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 10\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 11\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 12\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 13\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 14\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 15\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 16\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 17\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 18\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 19\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n            Epoch 20\/20\n            436\/436 [==============================] - 1s 1ms\/sample - loss: 11.7890 - acc: 0.0688 - val_loss: 13.6697 - val_acc: 0.0642\n<\/code><\/pre>\n\n<p>Model summary:<\/p>\n\n<pre><code>Model: \"sequential_16\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nconv2d_30 (Conv2D)           (None, 38, 38, 64)        640       \n_________________________________________________________________\nactivation_44 (Activation)   (None, 38, 38, 64)        0         \n_________________________________________________________________\nmax_pooling2d_30 (MaxPooling (None, 19, 19, 64)        0         \n_________________________________________________________________\nconv2d_31 (Conv2D)           (None, 17, 17, 64)        36928     \n_________________________________________________________________\nactivation_45 (Activation)   (None, 17, 17, 64)        0         \n_________________________________________________________________\nmax_pooling2d_31 (MaxPooling (None, 8, 8, 64)          0         \n_________________________________________________________________\nflatten_14 (Flatten)         (None, 4096)              0         \n_________________________________________________________________\ndense_28 (Dense)             (None, 64)                262208    \n_________________________________________________________________\ndense_29 (Dense)             (None, 1)                 65        \n_________________________________________________________________\nactivation_46 (Activation)   (None, 1)                 0         \n=================================================================\nTotal params: 299,841\nTrainable params: 299,841\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>Can someone please advise me if this could be the case or do . you see other areas I need to change\/fix?<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"58010051","CreationDate":"2019-09-19T11:31:20.093","ParentId":null,"OwnerUserId":"9742432","Title":null,"Body":"<p>You are using <\/p>\n\n<pre><code>model.add(Dense(1))\nmodel.add(Activation(\"softmax\"))\n<\/code><\/pre>\n\n<p>i.e., a Dense layer of only one neuron with softmax? This does not work, you need at least a dimensionality of the output of 2 in order to use softmax.<\/p>\n\n<p>How are your labels represented?<\/p>\n"},{"AnswerId":"58010351","CreationDate":"2019-09-19T11:49:07.570","ParentId":null,"OwnerUserId":"10411744","Title":null,"Body":"<p>Hi there writing my suggestions here because I've not earned the right to comment yet.<\/p>\n\n<p>First of all you might be completly right with your assumption that you need more data. Also you might to consider that the data might be skewed so that one class can appeare more often in the data. I don't really know how you chose your samples ...but you might want to look out for a realisitc distribution of the classes in your small sample.<\/p>\n\n<p>As for suggestions, I'm not sure what you try to predict, but i guess you want to find out if a image is cancerous or not. If this is the case you have a binary classification problem like with cats and dogs. So you should use a \"sigmoid\" activation function in your output layer instead of a \"softmax\". Softmax is mostlybe used for multi classification.<\/p>\n\n<p>Futhermore I don't see any deeper issues with your code. So try to change the activation function and use more samples with the right distribution if possible.<\/p>\n\n<p>Hope this helps :)<\/p>\n"}]}
+{"QuestionId":58007490,"AnswerCount":0,"Tags":"<python><tensorflow><eager-execution>","CreationDate":"2019-09-19T09:06:36.687","AcceptedAnswerId":null,"OwnerUserId":9266202.0,"Title":"Tensorflow - performance issue between eager and graph modes","Body":"<p>I'm facing an issue and I need your help.\nI'm trying to benchmark execution time between eager and graph modes.\nI'm using a very simple model (only one dense layer).<\/p>\n\n<p>Here is the graph version: <\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport timeit\n\ngraph = tf.Graph()\nwith graph.as_default():\n  x = tf.placeholder(tf.float32, [10, 200, 200, 3])\n  f = tf.keras.layers.Flatten()(x)\n  c = tf.keras.layers.Dense(512, activation='relu')(f)\n\nwith tf.Session(graph=graph) as sess:\n  sess.run(tf.global_variables_initializer())\n  val = np.ones((10, 200, 200, 3)) * 3.5\n  # warmup\n  sess.run(c, {x: val})\n  print(timeit.timeit(lambda: sess.run(c, {x: val}), number=1000)\/1000)\n<\/code><\/pre>\n\n<p>Here is the eager version<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport timeit\ntf.enable_eager_execution()\n\nflatten_layer = tf.keras.layers.Flatten()\ndense_layer = tf.keras.layers.Dense(512, activation='relu')\n\ndef c(x):\n  f = flatten_layer(x)\n  return dense_layer(f)\n\nval = tf.math.scalar_mul(3.5, tf.ones((10, 200, 200, 3)))\n# warmup\nc(val)\nprint(timeit.timeit(lambda: c(val), number=1000)\/1000)\n<\/code><\/pre>\n\n<p>I'm using tf  1.14 and the issue i'm facing is (if it is one) that the graph version is slower than the eager version.\nHere are the output of the graph and eager versions (respectively):<\/p>\n\n<pre><code>0.003849512729793787    \n\n0.0008663320459891111\n<\/code><\/pre>\n\n<p>I was expected that the result would be inverted, so I really don't understand why the graph version is slower than the eager one.\nIf I use the <em>tf.function<\/em> decorator in the eager version, the difference is even greater.\nHas someone an explanation ?<\/p>\n\n<p>Thank you<\/p>\n","answers":[]}
+{"QuestionId":58007525,"AnswerCount":1,"Tags":"<python><pytorch><tensor><zero-padding>","CreationDate":"2019-09-19T09:08:35.437","AcceptedAnswerId":58007819.0,"OwnerUserId":11658424.0,"Title":"padding zeroes to a tensor on both dimensions","Body":"<p>I have a tensor t<\/p>\n\n<pre><code>1 2\n3 4\n5 6\n7 8\n<\/code><\/pre>\n\n<p>And I would like to make it <\/p>\n\n<pre><code>0 0 0 0\n0 1 2 0\n0 3 4 0\n0 5 6 0\n0 7 8 0\n0 0 0 0 \n<\/code><\/pre>\n\n<p>I tried stacking with new=torch.tensor([0. 0. 0. 0.]) tensor four times but that did not work. <\/p>\n\n<pre><code>t = torch.arange(8).reshape(1,4,2).float()\nprint(t)\nnew=torch.tensor([[0., 0., 0.,0.]])\nprint(new)\nr = torch.stack([t,new])  # invalid argument 0: Tensors must have same number of dimensions: got 4 and 3\nnew=torch.tensor([[[0., 0., 0.,0.]]])\nprint(new)\nr = torch.stack([t,new])  # invalid argument 0: Sizes of tensors must match except in dimension 0.\n<\/code><\/pre>\n\n<p>I also tried cat, that did not work either.<\/p>\n","answers":[{"AnswerId":"58007819","CreationDate":"2019-09-19T09:25:50.863","ParentId":null,"OwnerUserId":"11734904","Title":null,"Body":"<p>It is probably better to initialize an array of the desired shape first, and then add the data at the appropriate indices.<\/p>\n\n<pre><code>import torch\n\nt = torch.arange(8).reshape(1,4,2).float()\nx = torch.zeros((1, t.shape[1]+2, t.shape[2]+2))\nx[:, 1:-1, 1:-1] = t\n\nprint(x)\n<\/code><\/pre>\n\n<p>On the other hand, if you just want to pad your tensor with zeroes (and not just add extra zeroes somewhere), you can use <code>torch.nn.functional.pad<\/code>:<\/p>\n\n<pre><code>import torch\n\nt = torch.arange(8).reshape(1, 4, 2).float()\nx = torch.nn.functional.pad(t, (1, 1, 1, 1))\n\nprint(x)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58007782,"AnswerCount":1,"Tags":"<python><pandas><validation><keras><deep-learning>","CreationDate":"2019-09-19T09:23:51.293","AcceptedAnswerId":null,"OwnerUserId":9118733.0,"Title":"Measuring performance for keras model being trained in batches","Body":"<p>I have a keras model which i am trying to train with a large dataset in chunks<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>for chunk in pd.read_csv(input_file, chunksize=chunk_size, usecols = FEATURE_COLUMNS, low_memory = False):\n\n       (X_train, y_train, X_val, y_val,full_pipeline) = dataPrep.get_data(data=chunk, mean=mean, variance=variance)\n\n\n\n       print(\"Clicks in Training Set =&gt;  {} , in CV =&gt;  {}\".format(np.sum(y_train == 1),np.sum(y_val == 1)))\n       print(\"SMS in Training Set =&gt;  {} , in CV =&gt;  {}\".format(np.sum(y_train == 0),np.sum(y_val == 0)))\n\n       preprocessor = full_pipeline\n       model.fit(X_train, y_train, batch_size=batch_size, validation_data=(X_val, y_val), epochs=epochs, verbose=1)\n       score = model.evaluate(X_val, y_val, verbose=0)\n\n       print('Validation Metrics    :', score)\n<\/code><\/pre>\n\n<p>Is there any way to measure the f1 score of the model after training on the entire dataset? Instead of performance of each chunk<\/p>\n","answers":[{"AnswerId":"58009099","CreationDate":"2019-09-19T10:35:18.603","ParentId":null,"OwnerUserId":"10411744","Title":null,"Body":"<p>Hi there writing my suggestion here because I'm not able to comment...<\/p>\n\n<p>correct me if I'm wrong but you could just get the f1-score after every chunk and then after the training is done you could get the mean of the gathered f1-scores.    <\/p>\n\n<p>Alernatively you also could get the recall and precision for every chunk and then use them in the final formular for the f1-score(<a href=\"https:\/\/en.wikipedia.org\/wiki\/F1_score\" rel=\"nofollow noreferrer\">Formular look here<\/a>)<\/p>\n\n<p>To see how to implement the f1-score, recall and precision you could take a look <a href=\"https:\/\/datascience.stackexchange.com\/questions\/45165\/how-to-get-accuracy-f1-precision-and-recall-for-a-keras-model\">here<\/a><\/p>\n\n<p>I hope this helps :)<\/p>\n"}]}
+{"QuestionId":58007918,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-19T09:31:57.543","AcceptedAnswerId":58015079.0,"OwnerUserId":6296435.0,"Title":"tf.metrics.accuracy doesn't match real accuracy","Body":"<p>I'm trying to reproduce the NN exercises of Coursera ML course using TensorFlow (and not Keras). <\/p>\n\n<p>I find that calculating the accuracy using <code>tf.metrics.accuracy<\/code> gives results that are lower than the accuracy when I calculate it.<\/p>\n\n<p>Relevant code is:<\/p>\n\n<pre><code>accuracy, update_op = tf.metrics.accuracy(labels=y, predictions=tf.argmax(tf.sigmoid(output), axis=1))\n...\n# in session:\nacc = sess.run(accuracy, feed_dict={tf_x: X, tf_y: y})\nsess.run(update_op, feed_dict={tf_x: X, tf_y: y})\nprint(f'step {step} - accuracy: {acc}')\n...\n# real accuracy\npredictions = sess.run(tf.argmax(tf.sigmoid(output), axis=1), feed_dict={tf_x: X})\npred_y = predictions == y\nprint(f'Training Set Accuracy after training: {np.mean(pred_y) * 100}%')\n<\/code><\/pre>\n\n<p>It can be even 30% difference (i.e. acc is 0.5 and real accuracy is 0.8)<\/p>\n\n<p>Am I doing something wrong?<\/p>\n\n<p>Note that if I do this:<\/p>\n\n<pre><code>equal = tf.equal(tf.cast(tf.argmax(tf.sigmoid(output), 1), tf.int32), y)\nacc_op = tf.reduce_mean(tf.cast(equal, tf.float32))\nacc = sess.run(acc_op, feed_dict={tf_x: X, tf_y: y})\n<\/code><\/pre>\n\n<p>I get the same results... so is tf.metrics.accuracy calculated in some other way?<\/p>\n","answers":[{"AnswerId":"58015079","CreationDate":"2019-09-19T16:09:57.583","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p><strong>SOLUTION<\/strong>: First call <code>sess.run(update_op, feed_dict)<\/code>, then <code>sess.run(accuracy)<\/code>. If feeding a new batch, and accuracy <em>on that batch<\/em> is desired, some hidden vars must first be reset - the workflow's as follows:<\/p>\n\n<pre><code>accuracy, update_op = tf.metrics.accuracy(tf_labels, tf_predictions, scope=\"my_metrics\")\nrunning_vars = tf.get_collection(tf.GraphKeys.LOCAL_VARIABLES, scope=\"my_metrics\")\nrunning_vars_initializer = tf.variables_initializer(var_list=running_vars)\n<\/code><\/pre>\n\n<pre><code>for i in range(num_batches):\n    # explicitly initialize\/reset 'total' and 'count' to 0\n    sess.run(running_vars_initializer) \n\n    # feed labels and predictions at i-th batch to update_ops\n    feed_dict={tf_labels: y[i], tf_predictions: tf.argmax(tf.sigmoid(output[i]), axis=1)}\n    session.run(update_op, feed_dict=feed_dict)\n\n    # compute and print accuracy from current 'total' and 'count'\n    print('Batch {} accuracy: {}'.format(i, session.run(accuracy)))\n<\/code><\/pre>\n\n<p><hr>\n<strong>DETAILS<\/strong>: <code>tf.metrics.accuracy<\/code> utilizes two runtime variables, <code>total<\/code> (# of correct predictions) and <code>count<\/code> (# of labels fed), initialized locally behind the scenes. <code>accuracy<\/code> gets updated <em>only<\/em> once <code>update_op<\/code> is called - in steps:<\/p>\n\n<ul>\n<li><code>total<\/code> and <code>count<\/code> are initialized to zero<\/li>\n<li><code>sess.run(update_op, feed_dict)<\/code> --> <code>total<\/code> and <code>count<\/code> are updated per <code>feed_dict<\/code><\/li>\n<li><code>sess.run(accuracy)<\/code> --> <code>accuracy<\/code> uses the <em>current<\/em> <code>total<\/code> and <code>count<\/code> to compute the metric<\/li>\n<li><code>sess.run(accuracy, feed_dict)<\/code> --> <code>accuracy<\/code> uses the <em>current<\/em> <code>total<\/code> and <code>count<\/code> to compute the metric<\/li>\n<\/ul>\n\n<p>What the last two say is, <code>feed_dict<\/code> actually does nothing to change <code>accuracy<\/code>; <code>accuracy<\/code> operates on <code>total<\/code> and <code>count<\/code>, which are only updated via <code>update_op<\/code>. Lastly,<\/p>\n\n<ul>\n<li><code>sess.run(accuracy, ...)<\/code> does <em>not<\/em> reset <code>total<\/code> and <code>count<\/code> to 0<\/li>\n<\/ul>\n\n<p>This is largely why <code>total<\/code> and <code>count<\/code> are used at all - for <em>scalability<\/em>; it enables computing metrics of data too big to fit into memory in one go, by keeping a running history. <\/p>\n\n<p>Lastly, your placeholder logic looks off - you feed data into <code>tf_x<\/code> and <code>tf_y<\/code>, yet neither are to be found anywhere within <code>tf.metrics.accuracy(...)<\/code> - but that's an easy fix.\n<hr>\n<strong>References\/further reading<\/strong>: <a href=\"https:\/\/stackoverflow.com\/questions\/46409626\/how-to-properly-use-tf-metrics-accuracy\">StackOverflow<\/a>, a good <a href=\"http:\/\/ronny.rest\/blog\/post_2017_09_11_tf_metrics\/\" rel=\"nofollow noreferrer\">blog entry<\/a><\/p>\n"}]}
+{"QuestionId":58008541,"AnswerCount":0,"Tags":"<python><numpy><tensorflow>","CreationDate":"2019-09-19T10:04:26.970","AcceptedAnswerId":null,"OwnerUserId":7121726.0,"Title":"Why tf.linspace() will return a long decimal number?","Body":"<p>Today, I tried to equally split a number range into a number list with TensorFlow's <code>linspace<\/code> function, I found it return a very annoying result:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nprint(tf.linspace(-3.0, 3.0, 11))\n<\/code><\/pre>\n\n<p>The output:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tf.Tensor(\n[-3.         -2.4        -1.8        -1.1999999  -0.5999999   0.\n  0.60000014  1.2000003   1.8000002   2.4         3.        ], shape=(11,), dtype=float32)\n<\/code><\/pre>\n\n<p>But if I use the same function in Numpy, it will show a more reasonable result:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nprint(np.linspace(-3.0, 3.0, 11))\n<\/code><\/pre>\n\n<p>The output:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>[-3.  -2.4 -1.8 -1.2 -0.6  0.   0.6  1.2  1.8  2.4  3. ]\n<\/code><\/pre>\n\n<p>Why TensorFlow will return a number that has a long decimal, like <code>-0.5999999<\/code> not just <code>-0.6<\/code>?<\/p>\n","answers":[]}
+{"QuestionId":58008573,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-layers>","CreationDate":"2019-09-19T10:06:09.897","AcceptedAnswerId":null,"OwnerUserId":6296435.0,"Title":"Easy way to set pre-trained weights in tensorflow?","Body":"<p>Is there an easy way to set pretrained weights to a tensorflow layer? (w\/o keras)<\/p>\n\n<p>In keras it's easy as <code>model.set_weights(...)<\/code> or <code>layer.set_weights(...)<\/code> - but I'm breaking my head around this for more than a day in tf, and can't seem to find anything that works. <\/p>\n","answers":[]}
+{"QuestionId":58009029,"AnswerCount":1,"Tags":"<tensorflow><google-colaboratory><tpu><google-cloud-tpu>","CreationDate":"2019-09-19T10:31:02.893","AcceptedAnswerId":null,"OwnerUserId":8842694.0,"Title":"TF1.14][TPU]Can not use custom TFrecord dataset on Colab using TPU","Body":"<p>I have created a TFRecord dataset file consisting elements and their corresponding labels. I want to use it for training model on Colab using free TPU. I can load the TFRecord file and even run an iterator just to see the contents however, before the beginning of the epoch it throws following error-<\/p>\n\n<pre><code>UnimplementedError: From \/job:worker\/replica:0\/task:0:\nFile system scheme '[local]' not implemented (file: '\/content\/gdrive\/My Drive\/data\/encodeddata_inGZIP.tfrecord')\n     [[{{node MultiDeviceIteratorGetNextFromShard}}]]\n     [[RemoteCall]]\n     [[IteratorGetNextAsOptional_1]]\n<\/code><\/pre>\n\n<p>In my understanding, it wants the TFRecord file on the TPU bucket, I don't know how to do that on Colab. How can one use a TFRecord file directly on Colab TPU?<\/p>\n","answers":[{"AnswerId":"58854212","CreationDate":"2019-11-14T10:11:32.813","ParentId":null,"OwnerUserId":"12371629","Title":null,"Body":"<p>You need to host it on Google Cloud Storage:<\/p>\n\n<blockquote>\n  <p>All input files and the model directory must use a cloud storage bucket path (gs:\/\/bucket-name\/...), and this bucket must be accessible from the TPU server. Note that all data processing and model checkpointing is performed on the TPU server, not the local machine.<\/p>\n<\/blockquote>\n\n<p>As mentioned on Google's troubleshooting page: <a href=\"https:\/\/cloud.google.com\/tpu\/docs\/troubleshooting#cannot_use_local_filesystem\" rel=\"nofollow noreferrer\">https:\/\/cloud.google.com\/tpu\/docs\/troubleshooting#cannot_use_local_filesystem<\/a><\/p>\n\n<p>Hope this helps!<\/p>\n"}]}
+{"QuestionId":58009538,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><loss-function>","CreationDate":"2019-09-19T11:00:26.790","AcceptedAnswerId":null,"OwnerUserId":3029270.0,"Title":"tensorflow inject loss value before running optimizer","Body":"<p>I want to fetch loss value from graph then change it and return it back before optimization step.<\/p>\n\n<p>how to do that in tensorflow <\/p>\n","answers":[]}
+{"QuestionId":58009788,"AnswerCount":1,"Tags":"<tensorflow><multiprocessing>","CreationDate":"2019-09-19T11:16:52.910","AcceptedAnswerId":null,"OwnerUserId":10261865.0,"Title":"Problem when predicting via multiprocess with Tensorflow","Body":"<p>I have 4 (or more) models (same structure but different training data). Now I want to ensemble them to make a prediction. I want to pre-load the models and then predict one input message (one message at a time) in parallel via multiprocess. However, the program always stops at \"session.run\" step. I could not figure it out why.<\/p>\n\n<p>I tried passing all arguments to the function in each process, as shown in the code below. I also tried using a Queue object and put all the data (except the model object) in the queue. I also tried to set the number of process to 1. It made no difference.<\/p>\n\n<pre><code>with Manager() as manager:\n    first_level_test_features=manager.list()\n    procs =[]\n    for id in range(4):\n        p = Process(target=predict, args=(id, (message, models, configs, vocabs, emoji_dict,first_level_test_features)))\n        procs.append(p)\n        p.start()\n\n    for p in procs:\n        p.join()\n<\/code><\/pre>\n\n<p>I did not get any error message since it is just stuck there. I would expect the program can start multiple processes and each process uses the model pass to it to make the prediction.<\/p>\n","answers":[{"AnswerId":"58010228","CreationDate":"2019-09-19T11:42:16.687","ParentId":null,"OwnerUserId":"12076271","Title":null,"Body":"<p>I am unsure how session sharing along different Processes would work, and this is probably where your issue comes from. Given the way TensorFlow works, I would advise implementing the ensemble call as a graph operation, so that it can be run through a single <code>session.run<\/code> call, with TF handling the parallelization of computations wherever possible.<\/p>\n\n<p>In practice, if you have symbolic tensors representing the models' predictions, you could use a TF operation to aggregate them (<code>tf.concat<\/code>, <code>tf.reduce_mean<\/code>, <code>tf.add_n<\/code>... whichever suits your design) and end up with a single symbolic tensor representing the ensemble prediction.<\/p>\n\n<p>I hope this helps; if not, please provide some more details as to what your setting is, notably which form your models have.<\/p>\n"}]}
+{"QuestionId":58010467,"AnswerCount":1,"Tags":"<python><tensorflow><model><coreml>","CreationDate":"2019-09-19T11:55:00.630","AcceptedAnswerId":null,"OwnerUserId":11856751.0,"Title":"ValueError: output name: logits\/BiasAdd, was provided, but the Tensorflow graph does not contain a tensor with this name","Body":"<p>I working on a project that detects eye region and I have a frozen inference model.\nI'm trying to convert this model to .mlmodel ( using tfcoreml ) but I have got this error :<\/p>\n\n<blockquote>\n  <p>ValueError: output name: logits\/BiasAdd was provided, but the Tensorflow graph does not contain a tensor with this name.<\/p>\n<\/blockquote>\n\n<p>But the tensor is existed on the graph, this :\n<a href=\"https:\/\/i.stack.imgur.com\/R3ttA.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/R3ttA.png\" alt=\"graph of the model\"><\/a><\/p>\n\n<p><strong>+ This is the input tensor(input_to_float) :<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Dzywa.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Dzywa.png\" alt=\"Input tensor\"><\/a><\/p>\n\n<p><strong>+ This is the output tensor(logits\/BiasAdd):<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/DZ6CF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DZ6CF.png\" alt=\"output tensor\"><\/a><\/p>\n\n<p>and this the code that I use for converting:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nimport tfcoreml as tf_converter\ntf_converter.convert(tf_model_path='C:\\\\Users\\\\User\\\\cnn-facial-landmark\\\\irismodel\\\\freeze2\\\\frozen_inf_model_iris.pb',\n                     mlmodel_path='irismodelios.mlmodel',\n                     output_feature_names=['logits\/BiasAdd'],\n                     input_name_shape_dict={'input_to_float': [1, 112, 112, 3]})\n<\/code><\/pre>\n\n<p>How can I solve this error?<\/p>\n","answers":[{"AnswerId":"58018299","CreationDate":"2019-09-19T20:07:12.907","ParentId":null,"OwnerUserId":"11856751","Title":null,"Body":"<p>I can say that I solved my problem by following this: <a href=\"https:\/\/stackoverflow.com\/questions\/48117854\/how-to-use-a-retrained-tensorflow-for-poets-graph-on-ios\">How to use a retrained &quot;tensorflow for poets&quot; graph on iOS?<\/a><\/p>\n\n<p>After optimizing the frozen graph, I get placeholder instead of OneShotterator, IteratorGetNext and cast.\nBefore converting: ( optimized graph )<\/p>\n\n<p>, but I'm not sure because of I doubt that the input_to_float should have [1,112,112,3] instead of [3,112,112]:<\/p>\n\n<p>Logs of CMD:<\/p>\n\n<pre><code>Core ML input(s):\n[name: \"input_to_float__0\"\ntype {\nmultiArrayType {\nshape: 3\nshape: 112\nshape: 112\ndataType: DOUBLE\n}\n}\n]\nCore ML output(s):\n[name: \"logits__BiasAdd__0\"\ntype {\nmultiArrayType {\nshape: 80\ndataType: DOUBLE\n}\n}\n]\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/dKLUN.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/dKLUN.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>After converting ( .mlmodel file )<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/1o8Fq.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1o8Fq.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":58010924,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-19T12:21:46.940","AcceptedAnswerId":null,"OwnerUserId":11622712.0,"Title":"Keras: 'Sequential' object has no attribute '_feed_input_names'","Body":"<p>I try to create a multi-label classificator using Keras. The training data has the following format:<\/p>\n\n<p><strong>X_train<\/strong><\/p>\n\n<pre><code>f1  f2  f3\n1   1   5\n0   2   4\n1   0   4\n<\/code><\/pre>\n\n<p><strong>y_train:<\/strong><\/p>\n\n<pre><code>c1  c2\n0   1\n1   1\n0   0\n<\/code><\/pre>\n\n<p>This is the code that I use to build the model:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense\nimport math\n\ndef softmax(z):\n    z_exp = [math.exp(i) for i in z]\n    sum_z_exp = sum(z_exp)\n    return [i \/ sum_z_exp for i in z_exp]\n\nnn = Sequential()\nnn.add(Dense(10, activation=\"relu\", input_shape=(10,)))\nnn.add(Dense(2, activation=\"sigmoid\"))\n\nnn.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])\n\nn = 10000\nmodel.fit(X_train.values, y_train.values, batch_size=16, epochs=5, verbose=1, validation_split=0.1)\n<\/code><\/pre>\n\n<p>When I run this code, it fails with the message:<\/p>\n\n<blockquote>\n  <p>AttributeError: 'Sequential' object has no attribute '_feed_input_names'<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"58012832","CreationDate":"2019-09-19T14:04:25.080","ParentId":null,"OwnerUserId":"10411744","Title":null,"Body":"<p>I' m writing my suggestions here because i can't comment yet.\nI think your input_shape might be off. You can find <a href=\"https:\/\/stackoverflow.com\/questions\/53758374\/attributeerror-sequential-object-has-no-attribute-feed-input-names?rq=1\">here<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/55062447\/attributeerror-sequential-object-has-no-attribute-feed-input-names-in-kera\">here<\/a> that someone other had the same problem. Maybe try something like:<\/p>\n\n<pre><code>nn.add(Dense(.... input_dim=X_train.shape[1]))\n<\/code><\/pre>\n\n<p>I hope this helps :)<\/p>\n"}]}
+{"QuestionId":58011262,"AnswerCount":0,"Tags":"<python><keras><conda>","CreationDate":"2019-09-19T12:41:26.563","AcceptedAnswerId":null,"OwnerUserId":11345941.0,"Title":"keras\"s error on different conda environment","Body":"<p>I am working with python 3.7, win 10. I have defined a new environment using conda, unfortunately when  I activate my environment using<\/p>\n\n<pre><code>conda activate the \"name of my environment\"\n\n<\/code><\/pre>\n\n<p>I want to use my last notebook when I run in a cell Keras import commands <\/p>\n\n<pre><code>from keras.preprocessing.text import Tokenizer\nfrom keras.preprocessing.sequence import pad_sequences\nfrom keras import metrics\nimport keras.backend as K\nfrom keras.preprocessing import sequence\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Dropout, Activation\nfrom keras.layers import Embedding\nfrom keras.layers import Conv1D, MaxPooling1D\n<\/code><\/pre>\n\n<p>I faced with the error regarding Keras or TensorFlow<\/p>\n\n<p>the error is so:<\/p>\n\n<pre><code>Using TensorFlow backend.\n---------------------------------------------------------------------------\nImportError                               Traceback (most recent call last)\nc:\\users\\moha\\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\nc:\\users\\moha\\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\nc:\\users\\moha\\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\\envs\\envc\\lib\\imp.py in load_module(name, file, filename, details)\n    241         else:\n--&gt; 242             return load_dynamic(name, filename, file)\n    243     elif type_ == PKG_DIRECTORY:\n\n~\\Anaconda3\\envs\\envc\\lib\\imp.py in load_dynamic(name, path, file)\n    341             name=name, loader=loader, origin=path)\n--&gt; 342         return _load(spec)\n    343 \n\nImportError: DLL load failed: The specified module could not be found.\n\nDuring handling of the above exception, another exception occurred:\n\nImportError                               Traceback (most recent call last)\n&lt;ipython-input-15-d5531b57a346&gt; in &lt;module&gt;\n----&gt; 1 from keras.preprocessing.text import Tokenizer\n      2 from keras.preprocessing.sequence import pad_sequences\n      3 from keras import metrics\n      4 import keras.backend as K\n      5 from keras.preprocessing import sequence\n\nc:\\users\\moha\\anaconda3\\lib\\site-packages\\keras\\__init__.py in &lt;module&gt;\n      1 from __future__ import absolute_import\n      2 \n----&gt; 3 from . import utils\n      4 from . import activations\n      5 from . import applications\n\nc:\\users\\moha\\anaconda3\\lib\\site-packages\\keras\\utils\\__init__.py in &lt;module&gt;\n      4 from . import data_utils\n      5 from . import io_utils\n----&gt; 6 from . import conv_utils\n      7 \n      8 # Globally-importable utils.\n\nc:\\users\\moha\\anaconda3\\lib\\site-packages\\keras\\utils\\conv_utils.py in &lt;module&gt;\n      7 from six.moves import range\n      8 import numpy as np\n----&gt; 9 from .. import backend as K\n     10 \n     11 \n\nc:\\users\\moha\\anaconda3\\lib\\site-packages\\keras\\backend\\__init__.py in &lt;module&gt;\n     87 elif _BACKEND == 'tensorflow':\n     88     sys.stderr.write('Using TensorFlow backend.\\n')\n---&gt; 89     from .tensorflow_backend import *\n     90 else:\n     91     # Try and load external backend.\n\nc:\\users\\moha\\anaconda3\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py in &lt;module&gt;\n      3 from __future__ import print_function\n      4 \n----&gt; 5 import tensorflow as tf\n      6 from tensorflow.python.framework import ops as tf_ops\n      7 from tensorflow.python.training import moving_averages\n\nc:\\users\\moha\\anaconda3\\lib\\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 from tensorflow._api.v1 import app\n\nc:\\users\\moha\\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\nc:\\users\\moha\\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\\moha\\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\\moha\\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\\moha\\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\\moha\\Anaconda3\\envs\\envc\\lib\\imp.py\", line 242, in load_module\n    return load_dynamic(name, filename, file)\n  File \"C:\\Users\\moha\\Anaconda3\\envs\\envc\\lib\\imp.py\", line 342, 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\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 have tried a different solution here<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/22872\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/22872<\/a><\/p>\n\n<p>does not work or faced with other erros?<\/p>\n\n<p>do you have any easy solution to solve the issue?<\/p>\n\n<p>I need to assert that it works on my main environment, I think it is related mostly to Keras with backend TensorFlow   ...<\/p>\n","answers":[]}
+{"QuestionId":58011264,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-19T12:41:28.740","AcceptedAnswerId":null,"OwnerUserId":9839954.0,"Title":"Transfer python code to tensorflow, while the AutoGraph don't work","Body":"<p>A simple example to illustrate <strong>function of my code:<\/strong><\/p>\n\n<pre><code>map = [[0,0,1,1,4],\n       [0,1,2,1,4],\n       [2,2,0,0,3]\n       ]\ncore = [2,2]\nmap_new = most_down(map,core)\nprint(map_new)\n[[0 1 4]\n [2 0 3]]\n<\/code><\/pre>\n\n<p>map and core are <code>int<\/code>. I upsample int feature map with <strong>most<\/strong> count value. For example\n<code>map_new[0][0] = 0<\/code>came from \n<code>map[0:2,0:2]<\/code>, because there are three <code>0<\/code> and one <code>1<\/code>. <strong>working code in python:<\/strong><\/p>\n\n<pre><code>import numpy as np\nimport statistics\ndef most_down(map, core=(16,16)):\n    map = np.array(map)\n    width = core[0]\n    core0 = width\n\n    height = core[1]\n    core1=height\n    shape = np.shape(map)\n\n    pro1 = int(shape[0]\/width)\n    pro2 = int(shape[1]\/height)\n    if not shape[0]%width==0:\n        pro1+=1\n    if not shape[1]%height==0:\n        pro2+=1\n    map_new = np.zeros([pro1,pro2],dtype=np.int)\n\n    for i in range(pro1):\n        for j in range(pro2):\n            if i&lt;pro1-1:\n                w = np.array([i*core0,(i+1)*core0])\n            else:\n                w = np.array([i * core0, shape[0]])\n            if j&lt;pro2-1:\n                h = np.array([j*core1,(j+1)*core1])\n            else:\n                h = np.array([j*core1,shape[1]])\n\n            block = map[w[0]:w[1], h[0]:h[1]]\n            block = np.reshape(block,[-1,])\n\n            most_hit =np.argmax(np.bincount(block))\n            map_new[i,j]=most_hit\n    return map_new\n\nmap = [[0,0,1,1,4],\n       [0,1,2,1,4],\n       [2,2,0,0,3]\n       ]\ncore = [2,2]\nmap_new = most_down(map,core)\nprint(map_new)\n<\/code><\/pre>\n\n<p>Then I try <code>@autograph.convert()<\/code> and then I got <strong>relentless error on <a href=\"https:\/\/www.tensorflow.org\/guide\/autograph?hl=zh-cn\" rel=\"nofollow noreferrer\">AutoGraph<\/a><\/strong>. I can't post all of them.<\/p>\n\n<p><strong>my question is<\/strong> :<\/p>\n\n<p><strong>1.<\/strong> How to use tf core code? I can only finish some of it:<\/p>\n\n<pre><code>def get_proporttion(map, core=(16,16)):\n\n    width = core[0]\n    height = core[1]\n    shape = map.shape\n    pro1 = tf.to_int32(tf.to_float(shape[0])\/tf.to_float(width))\n    pro2 = tf.to_int32(tf.to_float(shape[1])\/tf.to_float(height))\n    pro1=tf.cond(tf.not_equal(shape[0]%width,0),\n            lambda :tf.add(pro1,1),\n            lambda :pro1)\n    pro2=tf.cond(tf.not_equal(shape[1]%height,0),\n            lambda :tf.add(pro2,1),\n            lambda :pro2)\n<\/code><\/pre>\n\n<p><strong>2<\/strong>.how can I make <code>@autograph.convert()<\/code> work?<\/p>\n\n<p>Explicit description of the function:<\/p>\n\n<pre><code>def most_down:\n#Input1: map tensor tf.int32 shape=[width,height,channels]\n#Input2:core tensor tf.int32 shape=[2,2]\n#Output:new_map tensor tf.int32 shape will depend on the input.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58011410,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-09-19T12:49:31.537","AcceptedAnswerId":null,"OwnerUserId":1451579.0,"Title":"How to get current batch prediction inside keras training cycle?","Body":"<p>I am using <code>_loss, _acc = model.train_on_batch(x, y)<\/code> training approach. Now I want to get current batch model predictions(outputs) to find incorrect predictions and save same to use later in hard negative mining. How to get current outputs?<\/p>\n","answers":[]}
+{"QuestionId":58011563,"AnswerCount":1,"Tags":"<nlp><anaconda><pytorch>","CreationDate":"2019-09-19T12:58:36.003","AcceptedAnswerId":null,"OwnerUserId":7698202.0,"Title":"Module Not Found Error when importing Pytorch_Transformers","Body":"<p>After downloading pytorch_transformers through Anaconda and executing the import command through the Jupyter Notebook, I am facing several errors related to missing modules. <\/p>\n\n<p>I tried searching sacremoses to import the package via Anaconda, but it is only available for Linux machines. Has anyone else faced similar issues? Thanks in advance!<\/p>\n\n<p><code>from pytorch_transformers import BertTokenizer, BertModel, BertForMaskedLM<\/code><\/p>\n\n<p>This is the error:<\/p>\n\n<hr>\n\n<pre><code>&lt;ipython-input-5-218d0858d00f&gt; in &lt;module&gt;\n----&gt; 1 from pytorch_transformers import BertTokenizer, BertModel, BertForMaskedLM\n\n~\\Anaconda3\\lib\\site-packages\\pytorch_transformers\\__init__.py in &lt;module&gt;\n      1 __version__ = \"1.2.0\"\n----&gt; 2 from .tokenization_auto import AutoTokenizer\n      3 from .tokenization_bert import BertTokenizer, BasicTokenizer, WordpieceTokenizer\n      4 from .tokenization_openai import OpenAIGPTTokenizer\n      5 from .tokenization_transfo_xl import (TransfoXLTokenizer, TransfoXLCorpus)\n\n~\\Anaconda3\\lib\\site-packages\\pytorch_transformers\\tokenization_auto.py in &lt;module&gt;\n     24 from .tokenization_transfo_xl import TransfoXLTokenizer\n     25 from .tokenization_xlnet import XLNetTokenizer\n---&gt; 26 from .tokenization_xlm import XLMTokenizer\n     27 from .tokenization_roberta import RobertaTokenizer\n     28 from .tokenization_distilbert import DistilBertTokenizer\n\n~\\Anaconda3\\lib\\site-packages\\pytorch_transformers\\tokenization_xlm.py in &lt;module&gt;\n     25 from io import open\n     26 \n---&gt; 27 import sacremoses as sm\n     28 \n     29 from .tokenization_utils import PreTrainedTokenizer\n\nModuleNotFoundError: No module named 'sacremoses'```\n<\/code><\/pre>\n","answers":[{"AnswerId":"58012582","CreationDate":"2019-09-19T13:49:39.443","ParentId":null,"OwnerUserId":"10389198","Title":null,"Body":"<p>Please try to create a conda environment and install the packages in the created environment using the below steps:<\/p>\n\n<ol>\n<li>conda create -n env_pytorch -c intel python=3.5<\/li>\n<li>source activate env_pytorch<\/li>\n<li>pip install pytorch-transformers<\/li>\n<\/ol>\n\n<p>Hope this helps.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/0lYxL.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0lYxL.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":58012094,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><dropout>","CreationDate":"2019-09-19T13:24:57.923","AcceptedAnswerId":58012524.0,"OwnerUserId":2097240.0,"Title":"Tensorflow Custom Training With Phases","Body":"<p>I need to create a custom training loop with Tensorflow \/ Keras (because I want to have more than one optimizer and tell which weights each optimizer should act upon).<\/p>\n\n<p>Although <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/keras\/training_and_evaluation\" rel=\"nofollow noreferrer\">this tutorial<\/a> and <a href=\"https:\/\/www.tensorflow.org\/tutorials\/eager\/custom_training_walkthrough\" rel=\"nofollow noreferrer\">that one too<\/a> are quite clear regarding this matter, they miss a <strong>very important<\/strong> point: how do I predict for training phase and how do I predict for validation phase?   <\/p>\n\n<p>Suppose my model has <code>Dropout<\/code> layers, or <code>BatchNormalization<\/code> layers. They certainly work in a completely different way whether they are in training or validation. <\/p>\n\n<p>How do I adapt these tutorials? This is a dummy example (may contain one or two pieces of pseudocode):<\/p>\n\n<pre><code># Iterate over epochs.\nfor epoch in range(3):\n\n\n    # Iterate over the batches of the dataset.\n    for step, (x_batch_train, y_batch_train) in enumerate(train_dataset):\n        with tf.GradientTape() as tape:\n\n            #model with two outputs\n            #IMPORTANT: must be in training phase (use dropouts, calculate batch statistics)\n            logits1, logits2 = model(x_batch_train) #must be \"training\"\n\n            loss_value1 = loss_fn1(y_batch_train[0], logits1)\n            loss_value2 = loss_fn2(y_batch_train[1], logits2)\n\n            grads1 = tape.gradient(loss_value1, model.trainable_weights[selection1])    \n            grads2 = tape.gradient(loss_value2, model.trainable_weights[selection2])\n\n            optimizer1.apply_gradients(zip(grads1, model.trainable_weights[selection1]))\n            optimizer2.apply_gradients(zip(grads2, model.trainable_weights[selection2]))\n\n\n\n    # Run a validation loop at the end of each epoch.\n    for x_batch_val, y_batch_val in val_dataset:\n\n        ##Important: must be validation phase\n            #dropouts are off: calculate all neurons and divide value    \n            #batch norms use previously calculated statistics    \n        val_logits1, val_logits2 = model(x_batch_val)\n\n        #.... do the evaluations\n<\/code><\/pre>\n","answers":[{"AnswerId":"58012524","CreationDate":"2019-09-19T13:47:10.873","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>I think you can just pass a <code>training<\/code> parameter when you call a <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/Model\" rel=\"nofollow noreferrer\"><code>tf.keras.Model<\/code><\/a>, and it will be passed down to the layers:<\/p>\n\n<pre><code># On training\nlogits1, logits2 = model(x_batch_train, training=True)\n# On evaluation\nval_logits1, val_logits2 = model(x_batch_val, training=False)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58012178,"AnswerCount":0,"Tags":"<multiprocessing><pytorch><distributed-computing><distributed>","CreationDate":"2019-09-19T13:28:16.727","AcceptedAnswerId":null,"OwnerUserId":6129790.0,"Title":"Resolving pytorch distributed execution printing multiple log statements for each process spawned?","Body":"<p>I am running pytorch distributed environment to train some models and in the same script I am also using logging to print status of the program. The problem is that with pytorch distributed since its spawning multiple processes I see my log statements being printed <code>n<\/code> times where <code>n<\/code> is the number of processes being spawned. Here's an example of it:<\/p>\n\n<pre><code>1.0, 0.05, 2.1823, 0.1703, 1.9799, 0.2352\n1.0, 0.05, 2.1804, 0.1674, 1.9767, 0.2406\n1.0, 0.05, 2.1814, 0.1697, 2.0053, 0.2154\n2.0, 0.05, 2.1593, 0.1741, 2.0935, 0.192\n2.0, 0.05, 2.1526, 0.1779, 2.1166, 0.1908\n2.0, 0.05, 2.1562, 0.1812, 2.0868, 0.2076\n3.0, 0.05, 1.9319, 0.2473, 1.8041, 0.2903\n3.0, 0.05, 1.9386, 0.2413, 1.8037, 0.3017\n3.0, 0.05, 1.9286, 0.2443, 1.815, 0.2939\n4.0, 0.05, 1.7522, 0.3153, 1.828, 0.3131\n4.0, 0.05, 1.7504, 0.3207, 1.7613, 0.3245\n4.0, 0.05, 1.7522, 0.3223, 1.7841, 0.3209\n5.0, 0.05, 1.5815, 0.3951, 1.5559, 0.4307\n5.0, 0.05, 1.5767, 0.3939, 1.5326, 0.4205\n5.0, 0.05, 1.588, 0.3909, 1.5882, 0.3995\n<\/code><\/pre>\n\n<p>Any ideas on how to avoid or resolve this issue? Thanks!<\/p>\n","answers":[]}
+{"QuestionId":58012182,"AnswerCount":1,"Tags":"<keras><deep-learning><lstm><lstm-stateful>","CreationDate":"2019-09-19T13:28:23.713","AcceptedAnswerId":null,"OwnerUserId":3937300.0,"Title":"How to implement multi state LSTM RNN in keras","Body":"<p>I have 1000 distinct users and the dataset consists activities of these users over the past 1 year. Total records are over 300K. The inputs for the LSTM RNN has the feature vectors corresponding to these users. The user is also included because behavior of each user may vary from person to person. The network should learn behavior of each user and should be able to predict the next behavior based on the past information of the same user. \nHow to maintain separate hidden states for each user within an LSTM RNN.<\/p>\n\n<p>Following blog post is similar to my problem:<\/p>\n\n<p><a href=\"https:\/\/towardsdatascience.com\/multi-state-lstms-for-categorical-features-66cc974df1dc\" rel=\"nofollow noreferrer\">https:\/\/towardsdatascience.com\/multi-state-lstms-for-categorical-features-66cc974df1dc<\/a><\/p>\n\n<h2><strong>Update<\/strong><\/h2>\n\n<p>My dataset looks like:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/poL31.jpg\" rel=\"nofollow noreferrer\">DATASET<\/a><\/p>\n\n<p>I transformed my dataset into a 3D the numpy array and reshaped it as (No of records, timesteps, n_features).<\/p>\n\n<p>The questions are:<\/p>\n\n<p>1) Is it necessary to encode the \"user\" attribute?<\/p>\n\n<p>2) what is the correct batch size for this problem? Is it batch = 1000 (no. of distinct users)?<\/p>\n\n<p>3) Do I need to include each user's data in each batch input to the model?<\/p>\n\n<p>OR<\/p>\n\n<p>Please suggest the correct implementation of this problem.<\/p>\n","answers":[{"AnswerId":"58012477","CreationDate":"2019-09-19T13:44:50.637","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>This is just automatic. You don't need to do anything. <\/p>\n\n<p>The <code>LSTM<\/code> layer will certainly have a state matrix the size of your batch of users. (Otherwise it wouldn't be useful)   <\/p>\n"}]}
+{"QuestionId":58012185,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-19T13:28:33.747","AcceptedAnswerId":null,"OwnerUserId":6450489.0,"Title":"Combining map and padded_batch in tensorflow","Body":"<p>To increase the performance on input pipeline tensorflow documentation recommends using <code>tf.contrib.data.map_and_batch<\/code> instead of separate <code>.map<\/code> and <code>.batch<\/code> functions (<a href=\"https:\/\/www.tensorflow.org\/guide\/performance\/datasets\" rel=\"nofollow noreferrer\">Data input pipeline performance<\/a>). But how to combine <code>.map<\/code> and <code>.padded_batch<\/code> functions in input pipeline?<\/p>\n","answers":[]}
+{"QuestionId":58012199,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><pytorch>","CreationDate":"2019-09-19T13:29:32.017","AcceptedAnswerId":null,"OwnerUserId":10227103.0,"Title":"Model weights are not being updatesd, but loss is decreasing","Body":"<p>The following code is to train an MLP with images of size 64*64, while using the loss ||output - input||^2.<\/p>\n\n<p>For some reason, my weights per epoch are not being updated as shown at the end.<\/p>\n\n<pre><code>class MLP(nn.Module):\n    def __init__(self, size_list):\n        super(MLP, self).__init__()\n        layers = []\n        self.size_list = size_list\n        for i in range(len(size_list) - 2):\n            layers.append(nn.Linear(size_list[i],size_list[i+1]))\n            layers.append(nn.ReLU())\n        layers.append(nn.Linear(size_list[-2], size_list[-1]))\n        self.net = nn.Sequential(*layers)\n\n    def forward(self, x):\n        return self.net(x)\n\nmodel_1 = MLP([4096, 64, 4096])\n<\/code><\/pre>\n\n<p>And for training each epoch:<\/p>\n\n<pre><code>def train_epoch(model, train_loader, criterion, optimizer):\n    model.train()\n    model.to(device)\n\nrunning_loss = 0.0\n\n    start_time = time.time()\n    # train batch\n    for batch_idx, (data) in enumerate(train_loader):   \n        optimizer.zero_grad() \n\n        data = data.to(device)\n\n        outputs = model(data)\n        loss = criterion(outputs, data)\n        running_loss += loss.item()\n\n        loss.backward()\n        optimizer.step()\n\n    end_time = time.time()\n\n    weight_ll = model.net[0].weight\n    running_loss \/= len(train_loader)\n\n    print('Training Loss: ', running_loss, 'Time: ',end_time - start_time, 's')\n    return running_loss, outputs, weight_ll\n<\/code><\/pre>\n\n<p>for training the data:<\/p>\n\n<pre><code>n_epochs = 20\nTrain_loss = []\nweights=[]\n\ncriterion = nn.MSELoss()\n\noptimizer = optim.SGD(model_1.parameters(), lr = 0.1)\n\n\nfor i in range(n_epochs):\n    train_loss, output, weights_ll = train_epoch(model_1, trainloader, criterion, optimizer)\n    Train_loss.append(train_loss)\n    weights.append(weights_ll)\n    print('='*20)\n<\/code><\/pre>\n\n<p>Now, when I print the weights of the first fully connected layer per epoch they aren't being updated.<\/p>\n\n<pre><code>print(weights[0][0])\nprint(weights[19][0])\n<\/code><\/pre>\n\n<p>The output for the above is (showing the weight in epoch 0 and in epoch 19): <\/p>\n\n<pre><code>tensor([ 0.0086,  0.0069, -0.0048,  ..., -0.0082, -0.0115, -0.0133],\n       grad_fn=&lt;SelectBackward&gt;)\ntensor([ 0.0086,  0.0069, -0.0048,  ..., -0.0082, -0.0115, -0.0133],\n       grad_fn=&lt;SelectBackward&gt;)\n<\/code><\/pre>\n\n<p>What may be going wrong? Looking at my loss, it's decreasing at a steady rate but there is no change in the weights.<\/p>\n","answers":[{"AnswerId":"58020405","CreationDate":"2019-09-20T00:46:39.510","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>Try to change it <code>weight_ll = model.net[0].weight.clone().detach()<\/code> or just <code>weight_ll = model.net[0].weight.clone()<\/code> in your <code>train_epoch()<\/code> function. And you will see the weights differ.<\/p>\n\n<p>Explanation: <code>weights_ll<\/code> are always the last epoch values if you do not clone it. It will be regarded as the same tensor in the graph. That's why your <code>weights[0][0]<\/code> equals to <code>weights[19][0]<\/code>, they are actually the same tensor. <\/p>\n"}]}
+{"QuestionId":58012229,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tf.keras>","CreationDate":"2019-09-19T13:30:43.557","AcceptedAnswerId":null,"OwnerUserId":1109811.0,"Title":"`fit_generator` is not yet enabled for unbuilt Model subclasses","Body":"<p>I am trying to utilize <code>Model.fit_generator<\/code> of Keras in <strong>tensorflow 1.10<\/strong>.<\/p>\n\n<p>Simplified reproducible code:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nclass TestNet(tf.keras.Model):\n    def __init__(self, class_count, name='TestNet', **kwargs):\n        super(TestNet, self).__init__(name=name, **kwargs)\n        self.convolution = tf.keras.layers.Conv1D(class_count, kernel_size=1, input_shape=(None, 3))\n\n    def call(self, points):\n        return self.convolution(points)\n\ndef segmentation_loss(labels, logits):\n    cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(\n        labels=labels, logits=logits)\n    return tf.reduce_mean(cross_entropy)\n\ndef generate():\n    while True:\n        yield (np.zeros(shape=(100,3)), np.zeros(shape=(100)))\n\nif __name__ == \"__main__\":\n    test_net = TestNet(class_count=5)\n    optimizer = tf.keras.optimizers.Adam()\n    test_net.compile(optimizer, loss=segmentation_loss)\n    history = test_net.fit_generator(generate, steps_per_epoch=1000, epochs=10)\n<\/code><\/pre>\n\n<p>While this works in tensorflow 1.14, executing this in 1.10 yields the <code>NotImplementedError<\/code> in the title:<\/p>\n\n<blockquote>\n  <p>NotImplementedError: <code>fit_generator<\/code> is not yet enabled for unbuilt Model subclasses<\/p>\n<\/blockquote>\n\n<p>Anybody knows how to work around this?<\/p>\n","answers":[{"AnswerId":"58012555","CreationDate":"2019-09-19T13:48:20.390","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>I see a mistake in <code>fit_generator<\/code>, update to this:    <\/p>\n\n<pre><code>history = test_net.fit_generator(generate(), steps_per_epoch=1000, epochs=10)\n<\/code><\/pre>\n\n<p>Now, I don't know how Tensorflow 1.10 works, but this kind of modeling is rather new, usually a Keras model is built like this:<\/p>\n\n<pre><code>#model's inputs\ninputs = Input((None,3))\n\n#model's layers\nconvolution = Conv1D(class_count, kernel_size=1)\n\n#model's call\noutputs = convolution(inputs)\n\n#model finish\ntest_net = Model(inputs, outputs)\n<\/code><\/pre>\n\n<p>A workaround for you:<\/p>\n\n<pre><code>inputs = Input(shape)\ntest_net = TestNet(...)\noutputs = test_net.call(inputs)\ntest_net_model = Model(inputs, outputs)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58012270,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2019-09-19T13:32:57.680","AcceptedAnswerId":58064122.0,"OwnerUserId":2611712.0,"Title":"tensorflow: two models initialized in two separate graphs but they seen only in last one graph","Body":"<pre><code>import tensorflow as tf\nimport numpy as np\n\nclass SimpleModel():\n    pass\n\ndef declare_placeholders(self):\n\n    self.input_batch = tf.placeholder(dtype=tf.int32, shape=[None, None], name='input_batch') \n\nSimpleModel.__declare_placeholders = classmethod(declare_placeholders) \n\ndef init_model(self):\n    self.__declare_placeholders()\n\nSimpleModel.__init__ = classmethod(init_model)\n\ng_1 = tf.Graph()\nwith g_1.as_default():\n    model1 = SimpleModel()\n\ng_2 = tf.Graph()\nwith g_2.as_default(): \n    model2 = SimpleModel()\n<\/code><\/pre>\n\n<p>I'dont expect AssertionError in next situation:<\/p>\n\n<pre><code>assert model1.input_batch.graph is g_1\n\n---------------------------------------------------------------------------\nAssertionError                            Traceback (most recent call last)\n&lt;ipython-input-5-1ffc6a2e45a2&gt; in &lt;module&gt;\n----&gt; 1 assert model1.input_batch.graph is g_1\n\nAssertionError: \n<\/code><\/pre>\n\n<p>But I see next:<\/p>\n\n<pre><code>assert model1.input_batch.graph is g_2 (????)\n<\/code><\/pre>\n\n<p>But I've initialized model1 inside g_1!<\/p>\n\n<p>How can I modify SimpleModel to solve this problem?<\/p>\n","answers":[{"AnswerId":"58031196","CreationDate":"2019-09-20T15:23:29.430","ParentId":null,"OwnerUserId":"11952444","Title":null,"Body":"<p>I did something like this to load 2 different models in 2 separate graphs:<\/p>\n\n<p>Here PATH_TO_MODEL_CKPT is the path of the saved model file or you can construct the model inside the graph.<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python.platform import gfile\nimport os\n\nclass Graph1(object):\n    def __init__(self, PATH_TO_MODEL_CKPT):\n        graph1 = tf.Graph()\n        with self.graph1.as_default():\n            model_exp = os.path.expanduser(PATH_TO_CKPT)\n            print('Model filename: %s' % model_exp)\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                sess1 = tf.Session(graph=graph1)\n\nclass Graph2(object):\n    def __init__(self, PATH_TO_MODEL_CKPT):\n        graph2 = tf.Graph()\n        with self.graph2.as_default():\n            model_exp = os.path.expanduser(PATH_TO_CKPT)\n            print('Model filename: %s' % model_exp)\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                sess2 = tf.Session(graph=graph2)\n<\/code><\/pre>\n"},{"AnswerId":"58064122","CreationDate":"2019-09-23T13:57:15.650","ParentId":null,"OwnerUserId":"2611712","Title":null,"Body":"<p>I've have solve problem: use class definition without pass and classmethod<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nclass SimpleModel():\n\n    def declare_placeholders(self):    \n        self.input_batch = tf.placeholder(dtype=tf.int32, shape=[None, None], name='input_batch') \n\n    def __init__(self):\n        self.declare_placeholders()\n\ng_1 = tf.Graph()\nwith g_1.as_default():\n    model1 = SimpleModel()\n\ng_2 = tf.Graph()\nwith g_2.as_default(): \n    model2 = SimpleModel()\n<\/code><\/pre>\n"}]}
+{"QuestionId":58012584,"AnswerCount":1,"Tags":"<python-3.x><matplotlib><plot><keras>","CreationDate":"2019-09-19T13:49:46.963","AcceptedAnswerId":58012684.0,"OwnerUserId":11819525.0,"Title":"Keras : KeyError: 'acc' , during plotting","Body":"<p>i'm trying to visualize the Training plot of my Model , and i get this Error :<\/p>\n\n<pre><code>    Traceback (most recent call last):\n    File \"train.py\", line 120, in &lt;module&gt;\n    plt.plot(H.history['acc'])\n    KeyError: 'acc'\n<\/code><\/pre>\n\n<p>&amp; here's the Full Code :<\/p>\n\n<pre><code>     # set the matplotlib backend so figures can be saved in the background\n     import matplotlib\n     matplotlib.use(\"Agg\")\n\n     # import the necessary packages\n     from pyimagesearch.resnet import ResNet\n     from sklearn.preprocessing import LabelEncoder\n     from sklearn.model_selection import train_test_split\n     from sklearn.metrics import classification_report\n     from keras.preprocessing.image import ImageDataGenerator\n     from keras.optimizers import SGD\n     from keras.utils import np_utils\n     from imutils import paths\n     import matplotlib.pyplot as plt\n     import numpy as np\n     import argparse\n     import cv2\n     import os\n\n     # construct the argument parser and parse the arguments\n     ap = argparse.ArgumentParser()\n     ap.add_argument(\"-d\", \"--dataset\", required=True,\n         help=\"path to input dataset\")\n     ap.add_argument(\"-a\", \"--augment\", type=int, default=-1,\n         help=\"whether or not 'on the fly' data augmentation should be used\")\n     ap.add_argument(\"-p\", \"--plot\", type=str, default=\"plot.png\",\n         help=\"path to output loss\/accuracy plot\")\n     args = vars(ap.parse_args())\n\n     # initialize the initial learning rate, batch size, and number of\n     # epochs to train for\n     INIT_LR = 1e-1\n     BS = 8\n     EPOCHS = 50\n\n     # grab the list of images in our dataset directory, then initialize\n     # the list of data (i.e., images) and class images\n     print(\"[INFO] loading images...\")\n     imagePaths = list(paths.list_images(args[\"dataset\"]))\n     data = []\n     labels = []\n\n     # loop over the image paths\n     for imagePath in imagePaths:\n             # extract the class label from the filename, load the image, and\n             # resize it to be a fixed 64x64 pixels, ignoring aspect ratio\n             label = imagePath.split(os.path.sep)[-2]\n             image = cv2.imread(imagePath)\n             image = cv2.resize(image, (64, 64))\n\n             # update the data and labels lists, respectively\n             data.append(image)\n             labels.append(label)\n\n     # convert the data into a NumPy array, then preprocess it by scaling\n     # all pixel intensities to the range [0, 1]\n     data = np.array(data, dtype=\"float\") \/ 255.0\n\n     # encode the labels (which are currently strings) as integers and then\n     # one-hot encode them\n     le = LabelEncoder()\n     labels = le.fit_transform(labels)\n     labels = np_utils.to_categorical(labels, 3)\n\n     # partition the data into training and testing splits using 75% of\n     # the data for training and the remaining 25% for testing\n     (trainX, testX, trainY, testY) = train_test_split(data, labels,\n             test_size=0.25, random_state=42)\n\n     # initialize an our data augmenter as an \"empty\" image data generator\n     aug = ImageDataGenerator()\n\n     # check to see if we are applying \"on the fly\" data augmentation, and\n     # if so, re-instantiate the object\n     if args[\"augment\"] &gt; 0:\n            print(\"[INFO] performing 'on the fly' data augmentation\")\n            aug = ImageDataGenerator(\n                    rotation_range=20,\n                    zoom_range=0.15,\n                    width_shift_range=0.2,\n                    height_shift_range=0.2,\n                    shear_range=0.15,\n                    horizontal_flip=True,\n                    fill_mode=\"nearest\")\n\n    # initialize the optimizer and model\n    print(\"[INFO] compiling model...\")\n    opt = SGD(lr=INIT_LR, momentum=0.9, decay=INIT_LR \/ EPOCHS)\n    model = ResNet.build(64, 64, 3, 2, (2, 3, 4),\n            (32, 64, 128, 256), reg=0.0001)\n    model.compile(loss=\"categorical_crossentropy\", optimizer=opt,\n            metrics=[\"accuracy\"])\n\n    # train the network\n    print(\"[INFO] training network for {} epochs...\".format(EPOCHS))\n    H = model.fit_generator(\n            aug.flow(trainX, trainY, batch_size=BS),\n            validation_data=(testX, testY),\n            steps_per_epoch=len(trainX) \/\/ BS,\n            epochs=EPOCHS)\n\n    # evaluate the network\n    print(\"[INFO] evaluating network...\")\n    predictions = model.predict(testX, batch_size=BS)\n    print(classification_report(testY.argmax(axis=1),\n            predictions.argmax(axis=1), target_names=le.classes_))\n\n    # plot the training loss and accuracy\n    N = np.arange(0, EPOCHS)\n    plt.style.use(\"ggplot\")\n    plt.figure()\n    plt.plot(N, H.history[\"loss\"], label=\"train_loss\")\n    plt.plot(N, H.history['val_loss'], label=\"val_loss\")\n    plt.plot(N, H.history['acc'], label=\"accuracy\")\n    plt.plot(N, H.history['val_acc'], label=\"val_acc\")\n    plt.title(\"Training Loss and Accuracy on Dataset\")\n    plt.xlabel(\"Epoch #\")\n    plt.ylabel(\"Loss\/Accuracy\")\n    plt.legend(loc=\"lower left\")\n    plt.savefig(args[\"plot\"])\n<\/code><\/pre>\n\n<p>So, this code , is originally from the blog Pyimagesearch , on data augmentation , so , i decided to give it a try , the original \"pipeline\" was meant for a binary classification , so , since i was intrested in a mutli-class \"task\" , i brought some changes.<\/p>\n\n<p>but the plotting part of the code , seems correct to me , i cheked on multiple posts here on Stack , checked the keras documentation , and nothing , i do not understand , why it dose not work !!!!\nany suggestion would be much appreciated .\nthank you.<\/p>\n","answers":[{"AnswerId":"58012684","CreationDate":"2019-09-19T13:55:51.787","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Have you tried <code>H.history['accuracy']<\/code>? Since you compiled using <code>'accuracy'<\/code> it will probably have the same string.    <\/p>\n\n<p>Now you can always inspect what you've got:<\/p>\n\n<pre><code>for key in H.history.keys():\n    print(key)\n<\/code><\/pre>\n\n<p>You will see what is logged there <\/p>\n"}]}
+{"QuestionId":58012586,"AnswerCount":2,"Tags":"<pytorch>","CreationDate":"2019-09-19T13:49:49.510","AcceptedAnswerId":58017357.0,"OwnerUserId":169658.0,"Title":"How to add to pytorch tensor at indices?","Body":"<p>I have to admit, I'm a bit confused by the scatter* and index* operations - I'm not sure any of them do exactly what I'm looking for, which is very simple:<\/p>\n\n<p>Given some 2-D tensor<\/p>\n\n<pre><code>z = tensor([[1., 1., 1., 1.],\n            [1., 1., 1., 1.],\n            [1., 1., 1., 1.]])\n<\/code><\/pre>\n\n<p>And a list (or tensor?) of 2-d indexes:<\/p>\n\n<pre><code>inds = tensor([[0, 0],\n               [1, 1],\n               [1, 2]])\n<\/code><\/pre>\n\n<p>I want to add a scalar to z at those indexes (and do it efficiently):<\/p>\n\n<pre><code>znew = z.something_add(inds, 3)\n-&gt;\nznew = tensor([[4., 1., 1., 1.],\n               [1., 4., 4., 1.],\n               [1., 1., 1., 1.]])\n<\/code><\/pre>\n\n<p>If I have to I can make that scalar a tensor of whatever shape (where all elements = 3), but I'd rather not...<\/p>\n","answers":[{"AnswerId":"58017489","CreationDate":"2019-09-19T19:01:45.973","ParentId":null,"OwnerUserId":"3090085","Title":null,"Body":"<p>This code achieves what you want:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>z_new = z.clone() # copy the tensor\nz_new[inds[:, 0], inds[:, 1]] += 3 # modify selected indices of new tensor\n<\/code><\/pre>\n\n<p>In PyTorch, you can index each axis of a tensor with another tensor.<\/p>\n"},{"AnswerId":"58017357","CreationDate":"2019-09-19T18:52:33.930","ParentId":null,"OwnerUserId":"604734","Title":null,"Body":"<p>You must provide two lists to your indexing. The first having the row positions and the second the column positions. In your example, it would be:<\/p>\n\n<pre><code>z[[0, 1, 1], [0, 1, 2]] += 3\n<\/code><\/pre>\n\n<p>torch.Tensor indexing follows Numpy. See <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.indexing.html#integer-array-indexing\" rel=\"nofollow noreferrer\">https:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.indexing.html#integer-array-indexing<\/a> for more details.<\/p>\n"}]}
+{"QuestionId":58012741,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-19T13:59:03.623","AcceptedAnswerId":58064287.0,"OwnerUserId":8682939.0,"Title":"Error importing tensorflow \"AlreadyExistsError: Another metric with the same name already exists.\"","Body":"<p>I am running this simple code on Spyder 3.3 with Python 3.7 and Tensorlow 2.0:<\/p>\n\n<pre><code>import tensorflow as tf\nprint(tf.__version__)\n<\/code><\/pre>\n\n<p>When I try to run it again in the same IPython console, I get the following error:<\/p>\n\n<blockquote>\n  <p>File \"\/home\/rodrigo\/.local\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/eager\/monitoring.py\", line 121, in <strong>init<\/strong>\n      self._metric = self._metric_methods[self._label_length].create(*args)<\/p>\n  \n  <p>AlreadyExistsError: Another metric with the same name already exists.<\/p>\n<\/blockquote>\n\n<p>If I close the IPython console, and then open it again, it works fine. I am getting this error in every code that imports Tensorflow. <strong>Does anyone know how to solve this?<\/strong><\/p>\n\n<p>System configuration:<\/p>\n\n<ul>\n<li>Ubuntu 19.04 <\/li>\n<li>Spyder 3.3.2<\/li>\n<li>Python 3.7.3<\/li>\n<li>IPython 5.8.0<\/li>\n<li>TensorFlow 2.0.0-rc2<\/li>\n<\/ul>\n","answers":[{"AnswerId":"58064287","CreationDate":"2019-09-23T14:05:51.727","ParentId":null,"OwnerUserId":"371834","Title":null,"Body":"<p>Tensorflow constructs singletons as side effects during import. Importing twice results in the singletons being created again, which is not supported. Please never import twice.<\/p>\n"}]}
+{"QuestionId":58013114,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-datasets>","CreationDate":"2019-09-19T14:18:45.813","AcceptedAnswerId":null,"OwnerUserId":1500986.0,"Title":"Flipping labels in Tensorflow using map function of tf.data","Body":"<p>I am training a <code>tensorflow<\/code> model that outputs 3 labels: straight, left, right based on a image input.<\/p>\n\n<p>I am using a tf.data dataset and want to augment the data by flipping the image. The problem is, that if I flip the image, I also need to flip the label (as left becomes right when flipping the image).<\/p>\n\n<p>I have implemented the map function and tried to flip the labels using tf.gather\n(Was suggested here <a href=\"https:\/\/stackoverflow.com\/questions\/50439661\/how-to-horizontally-flip-labelled-directions-in-tensorflow\">How to horizontally flip labelled directions in tensorflow<\/a>)<\/p>\n\n<pre><code>augmented2 = trainDataset.map(\n    lambda x, y: flip(x, y), num_parallel_calls=10)\n\n\ndef flip(x: tf.Tensor, y) -&gt; tf.Tensor:\n    y = tf.gather( y,[0, 2, 1])\n    x = flip_left_right(x)\n    return x, y\n<\/code><\/pre>\n\n<p>The problem is that I get an error during training:<\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors_impl.InvalidArgumentError:  logits\n  and labels must be broadcastable: logits_size=[32,3] labels_size=[3,3]<\/p>\n<\/blockquote>\n\n<p>What am I doing wrong? Is the a better\/easier way to flip the labels?<\/p>\n","answers":[]}
+{"QuestionId":58013229,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-19T14:24:57.440","AcceptedAnswerId":58014425.0,"OwnerUserId":2908528.0,"Title":"Got unexpected keyword argument shape","Body":"<p>In my following Code <\/p>\n\n<pre><code>class cnnUtils:\n\n    def get_weight(shape):\n        init=tf.truncated_normal(shape,stddev=0.1)\n        return tf.Variable(init)\n    def get_bias(shape):\n        init=tf.constant(0.1,shape=shape)\n        return tf.Variable(init)\n    def conv2d(x,w):\n        return tf.nn.conv2d(x,w,strides=[1,1,1,1],padding=\"SAME\")\n    def maxpool_2d(x):\n        return tf.nn.max_pool(x,ksize=[1,2,2,1],strides=[1,2,2,1],padding=\"SAME\")\n    def conv_layer(input,shape):\n        b=get_bias([shape[3]])\n        w=get_weight(shape)\n        return tf.nn.relu(conv2d(input,w)+b)\n    def full_layer(input,size):\n        in_size=int(input.get_shape()[1])\n        w=get_weight([in_size,size])\n        b=get_bias([size])\n        return tf.matmul(input,w)+b\nutils=CnnUtils()\nx=tf.placeholder(tf.float32,shape=[None,32,32,3])\ny=tf.placeholder(tf.float32,shape=[None,10])\nconv1=utils.conv_layer(x,shape=[5,5,3,32])\n<\/code><\/pre>\n\n<p>I am getting following error<\/p>\n\n<hr>\n\n<p>TypeError                                 Traceback (most recent call last)\n in \n----> 1 conv1=utils.conv_layer(x,shape=[5,5,3,32])<\/p>\n\n<p>TypeError: conv_layer() got an unexpected keyword argument 'shape'<\/p>\n\n<p>But when I move the class keyword and use the code as simple function call like <\/p>\n\n<p>conv1=conv_layer(x,shape=[5,5,3,32])<\/p>\n\n<p>Erors got finished. Can somebody explain me what is happening here? My understanding is that the keyword \"shape\" is in a mess here.<\/p>\n","answers":[{"AnswerId":"58014425","CreationDate":"2019-09-19T15:28:44.287","ParentId":null,"OwnerUserId":"5136550","Title":null,"Body":"<p>In case of conv_layer as a method of CnnUtils class, 1st argument of conv_layer method, input, refers to the instance of class CnnUtils. Therefore, when you call utils.conv_layer(x,shape=[5,5,3,32]), x is assigned as the value of shape. [just print the value of input and shape in conv_layer method]. So the working implementation is as follows:<\/p>\n\n<pre><code>import tensorflow as tf\n\n\nclass CnnUtils:\n\n    def get_weight(self, shape):\n        init=tf.truncated_normal(shape,stddev=0.1)\n        return tf.Variable(init)\n\n    def get_bias(self, shape):\n        init=tf.constant(0.1,shape=shape)\n        return tf.Variable(init)\n\n    def conv2d(self, x, w):\n        return tf.nn.conv2d(x,w,strides=[1,1,1,1],padding=\"SAME\")\n\n    def maxpool_2d(self, x):\n        return tf.nn.max_pool(x,ksize=[1,2,2,1],strides=[1,2,2,1],padding=\"SAME\")\n\n    def conv_layer(self, input, shape):\n        b=self.get_bias([shape[3]])\n        w=self.get_weight(shape)\n        return tf.nn.relu(self.conv2d(input,w)+b)\n\n    def full_layer(self, input, size):\n        in_size=int(input.get_shape()[1])\n        w=self.get_weight([in_size,size])\n        b=self.get_bias([size])\n        return tf.matmul(input,w)+b\n\n\nutils=CnnUtils()\nx=tf.placeholder(tf.float32,shape=[None,32,32,3])\ny=tf.placeholder(tf.float32,shape=[None,10])\nconv1=utils.conv_layer(x, shape=[5,5,3,32])\n<\/code><\/pre>\n"}]}
+{"QuestionId":58013244,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2019-09-19T14:25:30.967","AcceptedAnswerId":58014072.0,"OwnerUserId":11713857.0,"Title":"Why doesn't `tf.matmul` work with transposed tensor?","Body":"<p>Why doesn't <code>tf.matmul<\/code> work with transposed tensor?<\/p>\n\n<p><code>transpose_b=True<\/code> is ok, but not <code>tf.transpose(inp)<\/code>.<\/p>\n\n<p>This screenshot was made in Colab with <code>tensorflow-gpu==2.0.0-rc1<\/code>:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/X7SOA.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/X7SOA.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"58014641","CreationDate":"2019-09-19T15:42:15.540","ParentId":null,"OwnerUserId":"2036766","Title":null,"Body":"<p>What Tensorflow is telling you is that the dimensions do not match up when multiplying the two tensors together. Think of it in basic linear algebra terms. When multiplying matrices, you can only multiply together matrices, where the last dimension of the first matrix is the same as the first dimension of the second. E.g.  you can multiply a 2x4 Matrix with a 4x2 matrix (which is what <code>transpose<\/code> does for you. From <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/transpose\" rel=\"nofollow noreferrer\">the docs<\/a>:<\/p>\n\n<blockquote>\n  <p>If <code>perm<\/code> is not given, it is set to (n-1...0), where n is the rank of the input tensor. Hence by default, this operation performs a regular matrix transpose on 2-D input Tensors.<\/p>\n<\/blockquote>\n\n<p>so if you omit perm in higher dimensions, <code>tf.transform()<\/code> switches dimensions just like it would for 2d tensors (matrices):<\/p>\n\n<pre><code>inp_t_without_perm = tf.transpose(inp)\ninp_t_without_perm\n# Output: &lt;tf.Tensor 'transpose_8:0' shape=(1, 4, 2) dtype=float32&gt;\n<\/code><\/pre>\n\n<p>so it just switches the last dimension for the first and leaves the second one unaltered. This is equivalent to:<\/p>\n\n<pre><code>inp_t_with_wrong_perm = tf.transpose(inp, perm=[2,1,0])\ninp_t_with_wrong_perm\n# Output: &lt;tf.Tensor 'transpose_8:0' shape=(1, 4, 2) dtype=float32&gt;\n<\/code><\/pre>\n\n<p>if you then do:<\/p>\n\n<pre><code>mul = tf.matmul(inp, inp_t_without_perm) # or with inp_t_with_wrong_perm\n<\/code><\/pre>\n\n<p>you get this error, because either your first two or last two dimensions do not match up.<\/p>\n\n<p>Now, when multiplying higher-order tensors together, you have to align the dimension which differ in the same way you would do in 2d (think about it as dividing up your tensor into matrices and vectors. In your case, you have a vector and a matrix... Sorry, I did not come up yet with a better metaphor, and when I find a quiet half-hour with pen and paper, I could make it more formal by using Einstein notation, but this is basically how it works...).<\/p>\n\n<p>For your case, what works is:<\/p>\n\n<pre><code>inp = tf.reshape(tf.linspace(-1.0, 1.0, 8), (2,4,1))\n# switch the last two dimensions so you can multiply 4x1 by 1x4\n# and leave first dimension as it is.\ninp_t = tf.transpose(inp, perm=[0,2,1])\nmul = tf.matmul(inp, inp_t)\nmul\n# Output: &lt;tf.Tensor 'MatMul_8:0' shape=(2, 4, 4) dtype=float32&gt;\n<\/code><\/pre>\n\n<p>Note that in your case, this is the only permutation which works, since this kind of multiplication is non-commutative. So you will have to match up dimensions from left to right (again, sorry for the hand-waving, but a formal mathematical proof would require me to do some higher-order tensor algebra, but I think this is precisely what you want to achieve...). I did not go too deep into the documentation, but I think the <code>transform_b<\/code> parameter is doing precicely this permutation for you. Hope that helps. Please comment for further questions.<\/p>\n"},{"AnswerId":"58014072","CreationDate":"2019-09-19T15:08:57.847","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p><code>transpose_b=True<\/code> in <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/linalg\/matmul\" rel=\"nofollow noreferrer\"><code>tf.linalg.matmul<\/code><\/a> transposes only the two last axes of the second given tensor, while <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/transpose\" rel=\"nofollow noreferrer\"><code>tf.transpose<\/code><\/a>, without more arguments, reverses the dimensions completely. The equivalent would be:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>inp_t = tf.transpose(inp, (0, 2, 1))\ntf.matmul(inp, inp_t)\n<\/code><\/pre>\n"},{"AnswerId":"58014154","CreationDate":"2019-09-19T15:14:03.870","ParentId":null,"OwnerUserId":"10418086","Title":null,"Body":"<p><code>tf.transpose()<\/code> performs regular 2-D matrix transpose by default (it sets the perm parameter to input_tensor_rank-1) if you don't explicitly specify the <code>perm<\/code>(permutation) parameter.\nSo set the perm parameter appropriately<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>inp_t = tf.transpose(inp, perm=[0,2,1])\ny = tf.matmul(inp, x)\nprint(y)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58013427,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><tensorflow-datasets>","CreationDate":"2019-09-19T14:33:40.850","AcceptedAnswerId":null,"OwnerUserId":11926707.0,"Title":"After I use dataset,concatenate, dataset.map only act on original data","Body":"<p>As the title says, I concatenate two dataset and use map function to change value's position and rescale value. Before I use map, all Tensor's shape are\nmatched, but after using map function, and use for loop iterate dataset to print index, and iteration's break point is on two dataset joint. <\/p>\n\n<p>I run into this problem in Colab using GPU,\nand use Python 3.6, tensorflow-gpu 2.0.0b1<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>dataset_crop = tf.data.Dataset.from_generator(img_resize_and_crop_genr, (tf.float32, tf.float32),((7,), (48,48,1)))\ndataset = dataset.concatenate(dataset_crop)\ndataset = dataset.map(lambda label, img_raw: (tf.cast(img_raw, tf.float32)\/float(255), label))\nfor i,(label, img) in enumerate(dataset):\n  print(i)\n<\/code><\/pre>\n\n<p>By the way, dataset's total row has 19984 before concatenating \n<br\/>\nWhat the concatenate hell..<\/p>\n\n<pre><code>...\n19982\n19983\n---------------------------------------------------------------------------\nInvalidArgumentError                      Traceback (most recent call last)\n&lt;ipython-input-26-36305ee0e8ef&gt; in &lt;module&gt;()\n----&gt; 1 for i,(label, img) in enumerate(dataset):\n      2   print(i)\n\n4 frames\n\/usr\/local\/lib\/python3.6\/dist-packages\/six.py in raise_from(value, from_value)\n\nInvalidArgumentError: ValueError: Tensor's shape (7,) is not compatible with supplied shape [48, 48, 1]\nTraceback (most recent call last):\n\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/script_ops.py\", line 209, in __call__\n    ret = func(*args)\n\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/data\/ops\/dataset_ops.py\", line 525, in generator_py_func\n    values = next(generator_state.get_iterator(iterator_id))\n\n  File \"&lt;ipython-input-25-196a9ac04fc0&gt;\", line 5, in img_resize_and_crop_genr\n    img.set_shape([side_len, side_len,1])\n\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 981, in set_shape\n    (self.shape, shape))\n\nValueError: Tensor's shape (7,) is not compatible with supplied shape [48, 48, 1]\n\n\n     [[{{node PyFunc}}]] [Op:IteratorGetNextSync]\n<\/code><\/pre>\n","answers":[{"AnswerId":"58048367","CreationDate":"2019-09-22T10:40:25.183","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>The problem is in your <code>from_generator<\/code> function. When you are passing the <code>output_shapes<\/code> parameter, a strict checking is done to see whether the output shape is exactly the same shape being generated or not. In your case, you are getting a <code>ValueError<\/code> stating that it is expecting (48, 48, 1) but (7,) shape got generated. <\/p>\n\n<p>A similar error can be generated using the following code:<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices((np.zeros(19984, dtype=np.float32), np.ones(19984, dtype=np.float32)))\n\ndef img_resize_and_crop_genr():\n    yield np.zeros((7,)), np.ones((48, 48, 1))\n\ndataset_crop = tf.data.Dataset.from_generator(img_resize_and_crop_genr, (tf.float32, tf.float32),((48,48,1), (7,)))\ndataset = dataset.concatenate(dataset_crop)\ndataset = dataset.map(lambda label, img_raw: (tf.cast(img_raw, tf.float32)\/float(255), label))\nfor i,(label, img) in enumerate(dataset):\n  print(i)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>ValueError: `generator` yielded an element of shape (7,) where an element of shape (48, 48, 1) was expected.\n<\/code><\/pre>\n\n<p>I believe you have interchanged your <code>output_shapes<\/code>. If that is the case, you can make the correction as:<\/p>\n\n<pre><code>dataset_crop = tf.data.Dataset.from_generator(img_resize_and_crop_genr, \n                                              (tf.float32, tf.float32),((7,), (48,48,1)))\n<\/code><\/pre>\n\n<p>Also, <code>output_shapes<\/code> is an optional parameter. You can avoid the problem whole together by not passing the parameter as shown below:<\/p>\n\n<pre><code>dataset_crop = tf.data.Dataset.from_generator(img_resize_and_crop_genr, \n                                              (tf.float32, tf.float32))\n<\/code><\/pre>\n"}]}
+{"QuestionId":58013655,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-lite>","CreationDate":"2019-09-19T14:47:08.143","AcceptedAnswerId":null,"OwnerUserId":11426718.0,"Title":"Failed to allocate tensors in tflite","Body":"<p>I am trying to load a tflite version of a tensorflow implementation of <a href=\"http:\/\/openaccess.thecvf.com\/content_cvpr_2017\/papers\/Zhou_EAST_An_Efficient_CVPR_2017_paper.pdf\" rel=\"nofollow noreferrer\">EAST<\/a>\ntext detector with tflite's python interpreter. When I try to allocate the tensors, I get the error:<\/p>\n\n<pre><code>tensorflow\/contrib\/lite\/kernels\/concatenation.cc:75 t-&gt;dims-&gt;data[d] != t0-&gt;dims-&gt;data[d] (45 != 46)\nNode number 78 (CONCATENATION) failed to prepare.\n\nFailed to allocate tensors!\n<\/code><\/pre>\n\n<p>I have fixed the size of the model during conversion and I get the error even when keeping the size choosen. Here is a sample of the code that fails:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.lite.python.interpreter import Interpreter\n\nmodel = Interpreter(model_path='frozen_east_text_detection.tflite')\nmodel.allocate_tensors()\n<\/code><\/pre>\n\n<p>Even though I found similar issues on tensorflow's <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\" rel=\"nofollow noreferrer\">github<\/a>, I can't find any solution, does anyone have a solution? Or at least know what is the cause of the problem?<\/p>\n\n<p>Thank you in advance.<\/p>\n","answers":[]}
+{"QuestionId":58013710,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-19T14:49:38.397","AcceptedAnswerId":58013955.0,"OwnerUserId":8704463.0,"Title":"tf.identity doesn't copy value before assignment","Body":"<p>I expect the program to<\/p>\n\n<ol>\n<li>copy the old value of <code>x<\/code> with <code>tf.identity<\/code>,<\/li>\n<li>assign new value <code>y<\/code> to <code>x<\/code><\/li>\n<\/ol>\n\n<p>But <code>tf.identity<\/code> doesn't keep the old value of <code>x<\/code>.<\/p>\n\n<p>The <code>tf.control_dependencies<\/code> should require doing step 1 before step 2.<\/p>\n\n<p>Result:<\/p>\n\n<pre><code>version: 1.14.1-dev20190531\nx_before_: 0.0\nx_ident_: 1.0\nx_after_: 1.0\n<\/code><\/pre>\n\n<p>Test:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\ndtype = np.float32\nx = tf.get_variable('x', shape=(), dtype=dtype,\n                    initializer=tf.zeros_initializer)\ny = tf.constant(1, dtype=dtype)\n\n# Copy value of x before assigning new value y to x\nx_ident = tf.identity(x)\nwith tf.control_dependencies([x_ident]):\n    assign_op = tf.assign(x, y)\n\n# Run\ninit_op = x.initializer\nwith tf.Session() as sess:\n    sess.run(init_op)\n    x_before_ = sess.run(x)\n    x_ident_ = sess.run([x_ident, assign_op])[0]\n    x_after_ = sess.run(x)\n\n# Check\nprint(\"version:\", tf.__version__)\nprint(\"x_before_:\", x_before_)\nprint(\"x_ident_:\", x_ident_)\nprint(\"x_after_:\", x_after_)\nassert np.allclose(x_ident_, 0)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58013955","CreationDate":"2019-09-19T15:02:24.800","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>I have found this problem in the past, and I think you are right, it should work as you expect. Maybe it warrants <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\" rel=\"nofollow noreferrer\">filing an issue<\/a>. I don't have a solution as such, but my workaround is to have a useless operation like adding zero:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\ndtype = np.float32\nx = tf.get_variable('x', shape=(), dtype=dtype,\n                    initializer=tf.zeros_initializer)\ny = tf.constant(1, dtype=dtype)\n\n# Force snapshot of x with zero addition\nx_ident = x + 0\nwith tf.control_dependencies([x_ident]):\n    assign_op = tf.assign(x, y)\n\n# Run\ninit_op = x.initializer\nwith tf.Session() as sess:\n    sess.run(init_op)\n    x_before_ = sess.run(x)\n    x_ident_ = sess.run([x_ident, assign_op])[0]\n    x_after_ = sess.run(x)\n\n# Check\nprint(\"version:\", tf.__version__)\nprint(\"x_before_:\", x_before_)\nprint(\"x_ident_:\", x_ident_)\nprint(\"x_after_:\", x_after_)\nassert np.allclose(x_ident_, 0)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code>version: 1.14.0\nx_before_: 0.0\nx_ident_: 0.0\nx_after_: 1.0\n<\/code><\/pre>\n"}]}
+{"QuestionId":58013754,"AnswerCount":1,"Tags":"<python><tensorflow><pickle><tensorflow-datasets>","CreationDate":"2019-09-19T14:51:45.383","AcceptedAnswerId":58042925.0,"OwnerUserId":9065637.0,"Title":"How to read multiple .mat files (which are too large to fit in memory) using tensorflow dataset","Body":"<p>I have around 550K samples, each sample being 200x50x1. The size of this dataset is around 57GB.<\/p>\n\n<p>I want to train a network on this set but I am having trouble reading it. <\/p>\n\n<pre><code>batch_size=8\n\ndef _read_py_function(filename,labels_slice):\n    with h5py.File(filename, 'r') as f:\n        data_slice = np.asarray(f['feats'])\n        print(data_slice.shape)\n    return data_slice, labels_slice\n\nplaceholder_files = tf.placeholder(tf.string, [None])\nplaceholder_labels = tf.placeholder(tf.int32, [None])\n\ndataset = tf.data.Dataset.from_tensor_slices((placeholder_files,placeholder_labels))\ndataset = dataset.map(\n    lambda filename, label: tuple(tf.py_func(\n        _read_py_function, [filename,label], [tf.uint8, tf.int32])))\n\ndataset = dataset.shuffle(buffer_size=50000)\ndataset = dataset.batch(batch_size)\n\niterator = tf.data.Iterator.from_structure(dataset.output_types, dataset.output_shapes)\ndata_X, data_y = iterator.get_next()\ndata_y = tf.cast(data_y, tf.int32)\n\nnet = conv_layer(inputs=data_X,num_outputs=8, kernel_size=3, stride=2, scope='rcl_0')\nnet = pool_layer(inputs=net,kernel_size=2,scope='pl_0')\n\nnet = dropout_layer(inputs=net,scope='dl_0')\n\nnet = flatten_layer(inputs=net,scope='flatten_0')\nnet = dense_layer(inputs=net,num_outputs=256,scope='dense_0')\nnet = dense_layer(inputs=net,num_outputs=64,scope='dense_1')\nout = dense_layer(inputs=net,num_outputs=10,scope='dense_2')\n\n<\/code><\/pre>\n\n<p>And I run the session using :<\/p>\n\n<pre><code>sess.run(train_iterator, feed_dict = {placeholder_files: filenames, placeholder_labels: ytrain})\n        try:\n            while True:\n                _, loss, acc = sess.run([train_op, loss_op, accuracy_op])\n                train_loss += loss \n                train_accuracy += acc\n        except tf.errors.OutOfRangeError:\n            pass\n\n<\/code><\/pre>\n\n<p>But I am getting the error even before running the session :<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"SFCC-trial-134.py\", line 297, in &lt;module&gt;\n    net = rcnn_layer(inputs=data_X,num_outputs=8, kernel_size=3, stride=2, scope='rcl_0')\n  File \"SFCC-trial-134.py\", line 123, in rcnn_layer\n    reuse=False)\n  File \"SFCC-trial-134.py\", line 109, in conv_layer\n    reuse         = reuse\n  File \"\/home\/priyam.jain\/tensorflow-gpu-python3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/arg_scope.py\", line 183, in func_with_args\n    return func(*args, **current_args)\n  File \"\/home\/priyam.jain\/tensorflow-gpu-python3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/layers\/python\/layers\/layers.py\", line 1154, in convolution2d\n    conv_dims=2)\n  File \"\/home\/priyam.jain\/tensorflow-gpu-python3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/framework\/python\/ops\/arg_scope.py\", line 183, in func_with_args\n    return func(*args, **current_args)\n  File \"\/home\/priyam.jain\/tensorflow-gpu-python3\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/layers\/python\/layers\/layers.py\", line 1025, in convolution\n    (conv_dims + 2, input_rank))\nTypeError: %d format: a number is required, not NoneType\n<\/code><\/pre>\n\n<p>I though about using TFRecords but had a hard time creating those. Couldn't find a good post where I learn to create them for my kind of dataset.<\/p>\n\n<p>conv_layer is defined as follows :<\/p>\n\n<pre><code>def conv_layer(inputs, num_outputs, kernel_size, stride, normalizer_fn=None, activation_fn=nn.relu, trainable=True, scope='noname', reuse=False):\n\n    net = slim.conv2d(inputs = inputs,\n        num_outputs   = num_outputs,\n        kernel_size   = kernel_size,\n        stride        = stride,\n        normalizer_fn = normalizer_fn,\n        activation_fn = activation_fn,\n        trainable     = trainable,\n        scope         = scope,\n        reuse         = reuse\n        )\n    return net\n<\/code><\/pre>\n","answers":[{"AnswerId":"58042925","CreationDate":"2019-09-21T18:14:17.847","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>Do not pass <code>tf.py_func<\/code> inside your map function. You can read the file image by passing the function name directly inside your <code>map<\/code> function. I am posing only the relevant parts of the code.<\/p>\n\n<pre><code>def _read_py_function(filename, label):\n    return tf.zeros((224, 224, 3), dtype=tf.float32), tf.ones((1,), dtype=tf.int32)\n\ndataset = dataset.map(lambda filename, label: _read_py_function(filename, label))\n<\/code><\/pre>\n\n<p>Another change is your <code>iterator<\/code> will expect only floating point of input. So you will have to change your <code>tf.uint8<\/code> type of output to <code>float<\/code>.<\/p>\n"}]}
+{"QuestionId":58013898,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning><lstm>","CreationDate":"2019-09-19T14:59:25.920","AcceptedAnswerId":null,"OwnerUserId":10589938.0,"Title":"Problem with Keras LSTM input_shape: expected lstm_1_input to have shape (500, 2) but got array with shape (500, 5)","Body":"<p><code>x_train<\/code> and <code>y_train<\/code> are input and output of my model with shapes of <code>(6508, 500, 5), (6508, 5)<\/code> respectively.<\/p>\n\n<p>And the model is like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential()\nmodel.add(LSTM(units=96, return_sequences=True, input_shape=x_train.shape[1:]))\nmodel.add(Dropout(0.2))\nmodel.add(LSTM(units=96, return_sequences=True))\nmodel.add(Dropout(0.2))\nmodel.add(LSTM(units=96))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(units=5))\n\nmodel.compile(loss='mean_squared_error', optimizer='adam', metrics=['mse'])\n\nmodel.fit(x_train, y_train, epochs=epochs, batch_size=batch_size)\n<\/code><\/pre>\n\n<p>Model Summary:<\/p>\n\n<pre><code>Model: \"sequential_1\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nlstm_1 (LSTM)                (None, 500, 96)           39168     \n_________________________________________________________________\ndropout_1 (Dropout)          (None, 500, 96)           0         \n_________________________________________________________________\nlstm_2 (LSTM)                (None, 500, 96)           74112     \n_________________________________________________________________\ndropout_2 (Dropout)          (None, 500, 96)           0         \n_________________________________________________________________\nlstm_3 (LSTM)                (None, 96)                74112     \n_________________________________________________________________\ndropout_3 (Dropout)          (None, 96)                0         \n_________________________________________________________________\ndense_1 (Dense)              (None, 5)                 485       \n=================================================================\nTotal params: 187,877\nTrainable params: 187,877\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>The problem is <code>lstm_1<\/code> requires input_shape (500, 2) and my data shape is (500, 5):<\/p>\n\n<pre><code>ValueError: Error when checking input: expected lstm_1_input to have shape (500, 2) but got array with shape (500, 5)\n<\/code><\/pre>\n\n<p>And I print layers' shape:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>for layer in model.layers:\n    print(layer.input_shape, end='\\t')\n\n# (None, 500, 5)  (None, 500, 96) (None, 500, 96) (None, 500, 96) (None, 500, 96) (None, 96)      (None, 96)\n<\/code><\/pre>\n\n<p>It prints <code>(None, 500, 5)<\/code> for <code>lstm_1<\/code> so I can't figure out the problem.<\/p>\n\n<pre><code>Keras==2.3.0\ntf==1.14.0\n<\/code><\/pre>\n\n<p><strong>UPDATE:<\/strong><\/p>\n\n<p>Using <code>keras==2.2.5<\/code> or <code>tf.keras<\/code> solves the problem.<\/p>\n","answers":[]}
+{"QuestionId":58014123,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><tensorflow-datasets><tensorflow2.0>","CreationDate":"2019-09-19T15:11:58.703","AcceptedAnswerId":59063925.0,"OwnerUserId":4373898.0,"Title":"How to improve data input pipeline performance?","Body":"<p>I try to optimize my data input pipeline.\nThe dataset is a set of 450 TFRecord files of size ~70MB each, hosted on GCS.\nThe job is executed with GCP\u00a0ML Engine. There is no GPU.<\/p>\n\n<p>Here is the pipeline:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def build_dataset(file_pattern):\n    return tf.data.Dataset.list_files(\n        file_pattern\n    ).interleave(\n        tf.data.TFRecordDataset,\n        num_parallel_calls=tf.data.experimental.AUTOTUNE\n    ).shuffle(\n        buffer_size=2048\n    ).batch(\n        batch_size=2048,\n        drop_remainder=True,\n    ).cache(\n    ).repeat(\n    ).map(\n        map_func=_parse_example_batch,\n        num_parallel_calls=tf.data.experimental.AUTOTUNE\n    ).prefetch(\n        buffer_size=1\n    )\n<\/code><\/pre>\n\n<p>With the mapped function:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def _bit_to_float(string_batch: tf.Tensor):\n    return tf.reshape(tf.math.floormod(tf.dtypes.cast(tf.bitwise.right_shift(\n        tf.expand_dims(tf.io.decode_raw(string_batch, tf.uint8), 2),\n        tf.reshape(tf.dtypes.cast(tf.range(7, -1, -1), tf.uint8), (1, 1, 8))\n    ), tf.float32), 2), (tf.shape(string_batch)[0], -1))\n\n\ndef _parse_example_batch(example_batch):\n    preprocessed_sample_columns = {\n        \"features\": tf.io.VarLenFeature(tf.float32),\n        \"booleanFeatures\": tf.io.FixedLenFeature((), tf.string, \"\"),\n        \"label\": tf.io.FixedLenFeature((), tf.float32, -1)\n    }\n    samples = tf.io.parse_example(example_batch, preprocessed_sample_columns)\n    dense_float = tf.sparse.to_dense(samples[\"features\"])\n    bits_to_float = _bit_to_float(samples[\"booleanFeatures\"])\n    return (\n        tf.concat([dense_float, bits_to_float], 1),\n        tf.reshape(samples[\"label\"], (-1, 1))\n    )\n<\/code><\/pre>\n\n<p>I tried to follow the best practices of the <a href=\"https:\/\/www.tensorflow.org\/guide\/performance\/datasets\" rel=\"nofollow noreferrer\">data pipeline tutorial<\/a>, and vectorize my mapped function (as advised by <a href=\"https:\/\/stackoverflow.com\/a\/53585477\/11543584\">mrry<\/a>).<\/p>\n\n<p>With this settings, while data are downloaded at high-speed (bandwidth is around 200MB\/s) the CPU is under-used (14%) and the training is very slow (more than 1hour for a epoch).<\/p>\n\n<p>I tried some parameters configuration, changing the <code>interleave()<\/code> arguments like <code>num_parallel_calls<\/code> or <code>cycle_length<\/code> or the <code>TFRecordDataset<\/code> arguments like <code>num_parallel_calls<\/code>.<\/p>\n\n<p>The fastest configuration uses this set of parameters:<\/p>\n\n<ul>\n<li><code>interleave.num_parallel_calls<\/code>: 1<\/li>\n<li><code>interleave.cycle_length<\/code>: 8<\/li>\n<li><code>TFRecordDataset.num_parallel_calls<\/code>: 8<\/li>\n<\/ul>\n\n<p>With this one, one epoch only take ~20 minutes to run. <strong>However, CPU usage is only at 50% while bandwidth consumption is around 55MB\/s<\/strong><\/p>\n\n<h2>Questions:<\/h2>\n\n<ol>\n<li>How to optimize the pipeline to reach 100% CPU usage (and something like 100MB\/s of bandwidth consumption)?<\/li>\n<li>Why does <code>tf.data.experimental.AUTOTUNE<\/code> not find best value to speed up the training?<\/li>\n<\/ol>\n\n<p>Kind,\nAlexis.<\/p>\n\n<hr>\n\n<h2>Edit<\/h2>\n\n<p>After some more experimentations, I came to the following solution.<\/p>\n\n<ol>\n<li>Remove the <code>interleave<\/code> step which is already handled by <code>TFRecordDataset<\/code> if <code>num_parallel_calls<\/code> is greater than 0.<\/li>\n<li>Update the mapped function to only do <code>parse_example<\/code> and <code>decode_raw<\/code>, returning a tuple `((, ), ())<\/li>\n<li><code>cache<\/code> after the <code>map<\/code><\/li>\n<li>Move the <code>_bit_to_float<\/code> function as a component of the model<\/li>\n<\/ol>\n\n<p>Finally, here is the data pipeline code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def build_dataset(file_pattern):\n    return tf.data.TFRecordDataset(\n        tf.data.Dataset.list_files(file_pattern),\n        num_parallel_reads=multiprocessing.cpu_count(),\n        buffer_size=70*1000*1000\n    ).shuffle(\n        buffer_size=2048\n    ).map(\n        map_func=split,\n        num_parallel_calls=tf.data.experimental.AUTOTUNE\n    ).batch(\n        batch_size=2048,\n        drop_remainder=True,\n    ).cache(\n    ).repeat(\n    ).prefetch(\n        buffer_size=32\n    )\n\n\ndef split(example):\n    preprocessed_sample_columns = {\n        \"features\": tf.io.VarLenFeature(tf.float32),\n        \"booleanFeatures\": tf.io.FixedLenFeature((), tf.string, \"\"),\n        \"label\": tf.io.FixedLenFeature((), tf.float32, -1)\n    }\n    samples = tf.io.parse_single_example(example, preprocessed_sample_columns)\n    dense_float = tf.sparse.to_dense(samples[\"features\"])\n    bits_to_float = tf.io.decode_raw(samples[\"booleanFeatures\"], tf.uint8)\n    return (\n        (dense_float, bits_to_float),\n        tf.reshape(samples[\"label\"], (1,))\n    )\n\n\ndef build_model(input_shape):\n    feature = keras.Input(shape=(N,))\n    bool_feature = keras.Input(shape=(M,), dtype=\"uint8\")\n    one_hot = dataset._bit_to_float(bool_feature)\n    dense_input = tf.reshape(\n        keras.backend.concatenate([feature, one_hot], 1),\n        input_shape)\n    output = actual_model(dense_input)\n\n    model = keras.Model([feature, bool_feature], output)\n    return model\n\ndef _bit_to_float(string_batch: tf.Tensor):\n    return tf.dtypes.cast(tf.reshape(\n        tf.bitwise.bitwise_and(\n            tf.bitwise.right_shift(\n                tf.expand_dims(string_batch, 2),\n                tf.reshape(\n                    tf.dtypes.cast(tf.range(7, -1, -1), tf.uint8),\n                    (1, 1, 8)\n                ),\n            ),\n            tf.constant(0x01, dtype=tf.uint8)\n        ),\n        (tf.shape(string_batch)[0], -1)\n    ), tf.float32)\n<\/code><\/pre>\n\n<p>Thanks to all these optimizations:<\/p>\n\n<ul>\n<li>Bandwidth consumption is around 90MB\/s<\/li>\n<li>CPU usage is around 20%<\/li>\n<li>First epoch spends 20 minutes<\/li>\n<li>Successives epochs spend 5 minutes each<\/li>\n<\/ul>\n\n<p>So this seems to be a good first setup. But CPU and BW are still not overused, so any advice is still welcomed!<\/p>\n\n<hr>\n\n<h2>Edit Bis<\/h2>\n\n<p>So, after some benchmarking I came accross what I think is our best input pipeline:<\/p>\n\n<pre><code>def build_dataset(file_pattern):\n    tf.data.Dataset.list_files(\n        file_pattern\n    ).interleave(\n        TFRecordDataset,\n        cycle_length=tf.data.experimental.AUTOTUNE,\n        num_parallel_calls=tf.data.experimental.AUTOTUNE\n    ).shuffle(\n        2048\n    ).batch(\n        batch_size=64,\n        drop_remainder=True,\n    ).map(\n        map_func=parse_examples_batch,\n        num_parallel_calls=tf.data.experimental.AUTOTUNE\n    ).cache(\n    ).prefetch(\n        tf.data.experimental.AUTOTUNE\n    )\n\ndef parse_examples_batch(examples):\n    preprocessed_sample_columns = {\n        \"features\": tf.io.FixedLenSequenceFeature((), tf.float32, allow_missing=True),\n        \"booleanFeatures\": tf.io.FixedLenFeature((), tf.string, \"\"),\n        \"label\": tf.io.FixedLenFeature((), tf.float32, -1)\n    }\n    samples = tf.io.parse_example(examples, preprocessed_sample_columns)\n    bits_to_float = tf.io.decode_raw(samples[\"booleanFeatures\"], tf.uint8)\n    return (\n        (samples['features'], bits_to_float),\n        tf.expand_dims(samples[\"label\"], 1)\n    )\n<\/code><\/pre>\n\n<p>So, what's new:<\/p>\n\n<ul>\n<li>According to this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/31019#issuecomment-517044334\" rel=\"nofollow noreferrer\">GitHub issue<\/a>, the <code>TFRecordDataset<\/code> interleaving is a legacy one, so <code>interleave<\/code> function is better.<\/li>\n<li><code>batch<\/code> before <code>map<\/code> is a good habit (<a href=\"https:\/\/www.tensorflow.org\/guide\/function#batching\" rel=\"nofollow noreferrer\">vectorizing your function<\/a>) and reduce the number of times the mapped function is called.<\/li>\n<li>No need of <code>repeat<\/code> anymore. Since TF2.0, the Keras model API supports the dataset API and can use cache (see the <a href=\"https:\/\/stackoverflow.com\/a\/58220646\/11543584\">SO post<\/a>)<\/li>\n<li>Switch from a <code>VarLenFeature<\/code> to a <code>FixedLenSequenceFeature<\/code>, removing a useless call to <code>tf.sparse.to_dense<\/code>.<\/li>\n<\/ul>\n\n<p>Hope this can help. Advices are still welcomed.<\/p>\n","answers":[{"AnswerId":"59063925","CreationDate":"2019-11-27T05:59:30.087","ParentId":null,"OwnerUserId":"11530462","Title":null,"Body":"<p>Mentioning the Solution and the Important observations of @AlexisBRENON in the Answer Section, for the benefit of the Community.<\/p>\n\n<p>Below mentioned are the Important Observations:<\/p>\n\n<ol>\n<li>According to this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/31019#issuecomment-517044334\" rel=\"nofollow noreferrer\">GitHub issue<\/a>, the <code>TFRecordDataset<\/code> <code>interleaving<\/code> is a legacy one, so <code>interleave<\/code> function is better.<\/li>\n<li><code>batch<\/code> before <code>map<\/code> is a good habit (<a href=\"https:\/\/www.tensorflow.org\/guide\/function#batching\" rel=\"nofollow noreferrer\">vectorizing your function<\/a>) and reduce the number of times the mapped function is called.<\/li>\n<li>No need of <code>repeat<\/code> anymore. Since TF2.0, the Keras model API supports the dataset API and can use cache (see the <a href=\"https:\/\/stackoverflow.com\/questions\/57977408\/how-to-cache-and-iterate-through-a-dataset-of-unknown-size\/58220646#58220646\">SO post<\/a>)<\/li>\n<li>Switch from a <code>VarLenFeature<\/code> to a <code>FixedLenSequenceFeature<\/code>, removing a useless call to <code>tf.sparse.to_dense<\/code>.<\/li>\n<\/ol>\n\n<p>Code for the Pipeline, with improved performance, in line with above observations is mentioned below:<\/p>\n\n<pre><code>def build_dataset(file_pattern):\n    tf.data.Dataset.list_files(\n        file_pattern\n    ).interleave(\n        TFRecordDataset,\n        cycle_length=tf.data.experimental.AUTOTUNE,\n        num_parallel_calls=tf.data.experimental.AUTOTUNE\n    ).shuffle(\n        2048\n    ).batch(\n        batch_size=64,\n        drop_remainder=True,\n    ).map(\n        map_func=parse_examples_batch,\n        num_parallel_calls=tf.data.experimental.AUTOTUNE\n    ).cache(\n    ).prefetch(\n        tf.data.experimental.AUTOTUNE\n    )\n\ndef parse_examples_batch(examples):\n    preprocessed_sample_columns = {\n        \"features\": tf.io.FixedLenSequenceFeature((), tf.float32, allow_missing=True),\n        \"booleanFeatures\": tf.io.FixedLenFeature((), tf.string, \"\"),\n        \"label\": tf.io.FixedLenFeature((), tf.float32, -1)\n    }\n    samples = tf.io.parse_example(examples, preprocessed_sample_columns)\n    bits_to_float = tf.io.decode_raw(samples[\"booleanFeatures\"], tf.uint8)\n    return (\n        (samples['features'], bits_to_float),\n        tf.expand_dims(samples[\"label\"], 1)\n    )\n<\/code><\/pre>\n"}]}
+{"QuestionId":58014671,"AnswerCount":1,"Tags":"<python><keras><deep-learning>","CreationDate":"2019-09-19T15:44:12.203","AcceptedAnswerId":58034149.0,"OwnerUserId":11973552.0,"Title":"How to get shape of previous layer and pass it to next layer?","Body":"<p>I want to take the shape of Input data which is passed to Input layer with (None,) shape, and use it in a for loop for some purpose.<\/p>\n\n<p>Here's part of my code implementation:<\/p>\n\n<pre><code>lst_rfrm = []\nInpt_lyr = keras.Input(shape = (None,))\nfor k in range(tm_stp):\n  F = keras.layers.Lambda(lambda x, i, j: x[:, None, j : j + i])\n  F.arguments = {'i' : sub_len, 'j' : k}\n  tmp_rfrm = F(Inpt_lyr)\n  lst_rfrm.append(tmp_rfrm)\ncnctnt_lyr = keras.layers.merge.Concatenate(axis = 1)(lst_rfrm)\n#defining other layers ...\n<\/code><\/pre>\n\n<p>because the Input shape is (None,), I don't know what to give to for loop as range( at the code i describe it with 'tm_stp'). how can i get the shape of the input layer (the data that is passed to input layer) in this situation? \nany help is deeply appreciated<\/p>\n","answers":[{"AnswerId":"58034149","CreationDate":"2019-09-20T19:26:03.660","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>You can try a different type of loop. It seems you are trying sliding windows, right?\nYou don't know the \"length\" to run, but you know the window size and how much of the borders to remove... so....<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/FXpR4.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/FXpR4.png\" alt=\"Slicing differently\"><\/a><\/p>\n\n<p>This function gets the slices following that principle:<\/p>\n\n<pre><code>windowSize = sub_len\ndef getWindows(x):\n    borderCut = windowSize - 1 #lost length in the length dimension\n\n    leftCut = range(windowSize) #start of sequence\n    rightCut = [i - borderCut for i in leftCut] #end of sequence - negative\n    rightCut[-1] = None #because it can't be zero for slicing\n\n    croppedSequences = K.stack([x[:, l: r] for l,r in zip(leftCut, rightCut)], axis=-1)\n    return croppedSequences\n<\/code><\/pre>\n\n<p>Running test:<\/p>\n\n<pre><code>from keras.layers import *\nfrom keras.models import Model\nimport keras.backend as K\nimport numpy as np\n\nwindowSize = 3\nbatchSize = 5\n\nrandomLength = np.random.randint(5,10)\ninputData = np.arange(randomLength * batchSize).reshape((batchSize, randomLength))\n\ndef getWindows(x):\n    borderCut = windowSize - 1\n\n    leftCut = range(windowSize)\n    rightCut = [i - borderCut for i in leftCut]\n    rightCut[-1] = None\n\n    croppedSequences = K.stack([x[:, l: r] for l,r in zip(leftCut, rightCut)], axis=-1)\n    return croppedSequences\n\ninputs = Input((None,))\noutputs = Lambda(getWindows)(inputs)\nmodel = Model(inputs, outputs)\n\npreds = model.predict(inputData)\n\nfor i, (inData, pred) in enumerate(zip(inputData, preds)):\n    print('sample: ', i)\n    print('input sequence: ', inData)\n    print('output sequence: \\n', pred, '\\n\\n')\n<\/code><\/pre>\n"}]}
+{"QuestionId":58015151,"AnswerCount":0,"Tags":"<tensorflow><keras><tf.keras>","CreationDate":"2019-09-19T16:14:57.650","AcceptedAnswerId":null,"OwnerUserId":11541912.0,"Title":"zero_output_for_mask RNN backend parameter not present in Keras","Body":"<p>I am using LSTM layers in <code>tf.keras<\/code> and I am trying to build an identical LSTM layer in <code>keras<\/code>.<\/p>\n\n<p>I noticed that <code>tf.keras.backend.rnn<\/code> has a parameter <code>zero_output_for_mask<\/code> <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.13\/api_docs\/python\/tf\/keras\/backend\/rnn\" rel=\"nofollow noreferrer\">link<\/a>. I understand that this parameter (if True) says to use an array of 0s when a timestep is missed\/skipped, and (if False) says to use the previous output of the LSTM when a timestep is missed\/skipped. I do not see this parameter in Keras RNN layers <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/95fab0e01d0451b70ee7b45469242438fefdb1cb\/keras\/layers\/recurrent.py#L239\" rel=\"nofollow noreferrer\">link<\/a>. <\/p>\n\n<p>I am wondering if the default behavior of <code>tf.keras.layers.LSTM<\/code> with <code>zero_output_for_mask=False<\/code> (the default behavior) is the same as the default behavior of <code>keras.layers.LSTM<\/code>? If not, how do I replicate the behavior of <code>tf.keras.layers.LSTM<\/code> in <code>keras.layers.LSTM<\/code>?<\/p>\n","answers":[]}
+{"QuestionId":58015227,"AnswerCount":0,"Tags":"<tensorflow><keras><tf.keras>","CreationDate":"2019-09-19T16:21:03.950","AcceptedAnswerId":null,"OwnerUserId":2130515.0,"Title":"Customer MultiSGD optimizer for tensorflow.keras","Body":"<p>I tried to implement the MultiSGD optimizer using the tf.keras. The code is modified from <a href=\"https:\/\/gist.github.com\/mkocabas\/99658da8186145f6f1e2fc70e882dac0\" rel=\"nofollow noreferrer\">MultiSGD<\/a> wich written for keras.\nhere is my version:<\/p>\n\n<pre><code>@tf_export('keras.optimizers.MultiSGD')\nclass MultiSGD(Optimizer):\n    \"\"\"Stochastic gradient descent optimizer.\n\n    Includes support for momentum,\n    learning rate decay, and Nesterov momentum.\n\n    Arguments:\n        lr: float &gt;= 0. Learning rate.\n        momentum: float &gt;= 0. Parameter that accelerates SGD\n            in the relevant direction and dampens oscillations.\n        decay: float &gt;= 0. Learning rate decay over each update.\n        nesterov: boolean. Whether to apply Nesterov momentum.\n    \"\"\"\n\n    def __init__(self, lr=0.01, momentum=0., decay=0., nesterov=False, lr_mult=None, **kwargs):\n        super(MultiSGD, self).__init__(**kwargs)\n        with K.name_scope(self.__class__.__name__):\n            self.iterations = K.variable(0, dtype='int64', name='iterations')\n            self.lr = K.variable(lr, name='lr')\n            self.momentum = K.variable(momentum, name='momentum')\n            self.decay = K.variable(decay, name='decay')\n        self.initial_decay = decay\n        self.nesterov = nesterov\n        self.lr_mult = lr_mult\n\n    def get_updates(self, loss, params):\n        grads = self.get_gradients(loss, params)\n        self.updates = [state_ops.assign_add(self.iterations, 1)]\n\n        lr = self.lr\n        if self.initial_decay &gt; 0:\n            lr = lr * (  # pylint: disable=g-no-augmented-assignment\n                    1. \/ (1. + self.decay * math_ops.cast(self.iterations,\n                                                          K.dtype(self.decay))))\n        # momentum\n        shapes = [K.int_shape(p) for p in params]\n        moments = [K.zeros(shape) for shape in shapes]\n        self.weights = [self.iterations] + moments\n        for p, g, m in zip(params, grads, moments):\n            if p.name in self.lr_mult:\n                multiplied_lr = lr * self.lr_mult[p.name]\n            else:\n                multiplied_lr = lr\n            # v = self.momentum * m - lr * g  # velocity\n            v = self.momentum * m - multiplied_lr * g  # velocity\n            self.updates.append(state_ops.assign(m, v))\n\n            if self.nesterov:\n                # new_p = p + self.momentum * v - lr * g\n                new_p = p + self.momentum * v - multiplied_lr * g\n            else:\n                new_p = p + v\n\n            # Apply constraints.\n            if getattr(p, 'constraint', None) is not None:\n                new_p = p.constraint(new_p)\n\n            self.updates.append(state_ops.assign(p, new_p))\n        return self.updates\n\n    def get_config(self):\n        config = {\n            'lr': float(K.get_value(self.lr)),\n            'momentum': float(K.get_value(self.momentum)),\n            'decay': float(K.get_value(self.decay)),\n            'nesterov': self.nesterov\n        }\n        base_config = super(MultiSGD, self).get_config()\n        return dict(list(base_config.items()) + list(config.items()))\n<\/code><\/pre>\n\n<p>I got an error that I couldnot fix it:\n<strong>tensorflow.python.framework.errors_impl.FailedPreconditionError: Error while reading resource variable training\/MultiSGD\/Variable_180 from Container: localhost. This could mean that the variable was uninitialized<\/strong><\/p>\n","answers":[]}
+{"QuestionId":58015489,"AnswerCount":4,"Tags":"<python><tensorflow><flask><keras>","CreationDate":"2019-09-19T16:38:49.263","AcceptedAnswerId":null,"OwnerUserId":11306005.0,"Title":"Flask and Keras model Error ''_thread._local' object has no attribute 'value''?","Body":"<p>I am using the following:\npython 3.6.4<\/p>\n\n<p>Flask = 1.1.1,<\/p>\n\n<p>Keras = 2.3.0,<\/p>\n\n<p>TensorFlow = 1.14.0,\nI have a Flask server that gets pictures from the clients. using Keras model with a TensorFlow back-end I try to get a prediction from a pre-trained model.<\/p>\n\n<p>I am using the following function to upload the model( as part of a class)<\/p>\n\n<pre><code>\n model_path = self.conf[\"model_path\"] \/\/ path in conf to model\n self.model = load_model(model_path)  \/\/ uploading the model\n self.model._make_predict_function()\n p_log.info(\"model had been upload successfully \")\n<\/code><\/pre>\n\n<p>and I use the following line for prediction:<\/p>\n\n<pre><code>cm_prediction = self.model.predict([face, reye, leye, fg])[0]\n<\/code><\/pre>\n\n<p>Until today I didn't have any problem, always got a prediction.\n<strong>now I get the following error<\/strong>: <\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"D:\\code_project\\path to project\", line 75, in predict\n    cm_prediction = self.model.predict([face, reye, leye, fg])[0]\n  File \"D:\\code_project\\path to project\", line 1462, in predict\n    callbacks=callbacks)\n  File \"D:\\code_project\\predictserver\\venv\\lib\\site-packages\\keras\\engine\\training_arrays.py\", line 276, in predict_loop\n    callbacks.model.stop_training = False\n  File \"D:\\code_project\\predictserver\\venv\\lib\\site-packages\\keras\\engine\\network.py\", line 323, in __setattr__\n    super(Network, self).__setattr__(name, value)\n  File \"D:\\code_project\\predictserver\\venv\\lib\\site-packages\\keras\\engine\\base_layer.py\", line 1215, in __setattr__\n    if not _DISABLE_TRACKING.value:\nAttributeError: '_thread._local' object has no attribute 'value'\n<\/code><\/pre>\n\n<p>I have a simple Flask server running:<\/p>\n\n<pre><code>if __name__ == '__main__':\n    pre = predictor()\n    # app.run(debug=True)\n    app.run(host='0.0.0.0', port=12345)\n<\/code><\/pre>\n\n<p>The model is always being uploaded.<\/p>\n\n<p>If I am running the program <strong>without the Flask server<\/strong>, hence giving manually input, <strong>I get a prediction<\/strong>, but as soon as the <strong>server is on<\/strong> the error appears and I <strong>stop getting a predictions<\/strong><\/p>\n\n<p>I tried to look on the web for some similar problem but didnt found any, if someone knows what the problem and how to solve it, I will appreciate sharing it.<\/p>\n","answers":[{"AnswerId":"58023136","CreationDate":"2019-09-20T06:49:24.153","ParentId":null,"OwnerUserId":"11425739","Title":null,"Body":"<p>I had the same problem when building a docker container today, which had worked perfectly before. Fixed it by downgrading Keras version to 2.2.4.<\/p>\n"},{"AnswerId":"58204527","CreationDate":"2019-10-02T15:19:58.277","ParentId":null,"OwnerUserId":"12154623","Title":null,"Body":"<p>I had the same problem with <strong>Keras 2.3.0<\/strong>.<\/p>\n\n<p>Another fix for those that don't want to downgrade is to set  <code>threaded=False<\/code> in <code>app.run()<\/code>.<\/p>\n"},{"AnswerId":"58023399","CreationDate":"2019-09-20T07:08:39.197","ParentId":null,"OwnerUserId":"11306005","Title":null,"Body":"<p>So after a long night,\nKeras had released a new version 2.3.0 in Sep 17,19. \nAs part of revision update I did, I updated all libraries, Keras among them. Since I did it the message appeared. <\/p>\n\n<p><strong>After I downgraded back to Keras 2.2.5 The problem disappeared.<\/strong><\/p>\n"},{"AnswerId":"58666381","CreationDate":"2019-11-01T21:58:45.313","ParentId":null,"OwnerUserId":"545698","Title":null,"Body":"<p>If you're having issues and are a little slow like myself, set debug=False as well<\/p>\n"}]}
+{"QuestionId":58015528,"AnswerCount":0,"Tags":"<python-3.x><python-2.7><numpy><tensorflow><ros>","CreationDate":"2019-09-19T16:41:44.857","AcceptedAnswerId":null,"OwnerUserId":9954063.0,"Title":"Does runtimewarning: overflow encountered in add, cause a ValueError: math domain error?","Body":"<p>I have a small piece of code where I am taking a set of 24 angles and then using a NN to predict next set of 24 angles. before passing the input to NN, I am processing the input using a complex math function involving cos, sin, multiplications and addition. While trying to run the simulation after certain time like 40 minutes, I am facing this <code>RuntimeWarning: overflow encountered in add<\/code> which then leads to an <code>ValueError: math domain error<\/code><\/p>\n\n<p>I have initially though the math domain error might be occurring because the angles that might be fed t cos might be out of bound and so remapped those angles in the cos range and the error still continues. I am using numpy and tensorflow for processing. <\/p>\n\n<pre><code>term_1_1 = (self.m_l*9.8*self.velocity*(self.r_1*math.sin(self.j_angles[0])+2*self.l_1*math.sin(self.j_angles[0])+self.r_1*math.sin(self.j_angles[0]+self.j_angles[1])+self.r_1*math.sin(self.j_angles[0]+self.j_angles[1]+self.j_angles[2]))+self.m_l*9.8*self.velocity*(self.r_1*math.sin(self.j_angles[0]+self.j_angles[1])+self.l_1*math.sin(self.j_angles[1])+self.r_1*math.sin(self.j_angles[0]+self.j_angles[1]+self.j_angles[2]))+self.m_l*9.8*self.velocity*(self.r_1*math.sin(self.j_angles[0]+self.j_angles[1]+self.j_angles[2])))\n<\/code><\/pre>\n\n<p>Can anyone help me by pointing what exactly is causing this overflow and error. All the numpy arrays of dtype float64, same as tensorflow variables and other placeholders. j_angles is a numpy initialized array of 24 elements followed by the ROS subscriber output.<\/p>\n","answers":[]}
+{"QuestionId":58015589,"AnswerCount":1,"Tags":"<keras><keras-layer>","CreationDate":"2019-09-19T16:45:37.737","AcceptedAnswerId":58016516.0,"OwnerUserId":6499614.0,"Title":"Keras: layer trainable set false not working","Body":"<p>I have a pretrained keras based Bidirectional LSTM based model. I want to set trainable false for the first few layers. But it is not working.<\/p>\n\n<pre><code> classifier.layers[1].trainable=False\n<\/code><\/pre>\n\n<p>I also tried creating a dummy copy of the model using config with trainable key changed to false. But even the weights of that dummy model change after training.<\/p>\n","answers":[{"AnswerId":"58016516","CreationDate":"2019-09-19T17:51:45.543","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Since something is strange, you can try to deeply find all the layers involved and set <code>trainable=False<\/code> in all of them: <\/p>\n\n<pre><code>classifier.layers[i].forward_layer.trainable = False\nclassifier.layers[i].backward_layer.trainable = False \nclassifier.layers[i].trainable = False\n<\/code><\/pre>\n\n<p>And use <code>classifier.compile<\/code> after that. <\/p>\n\n<p>In case your model involves submodels, you might need recurrency:<\/p>\n\n<pre><code>classifier.layers[i].layers[j].trainable = False\n<\/code><\/pre>\n"}]}
+{"QuestionId":58015667,"AnswerCount":2,"Tags":"<python><tensorflow><tensorflow-estimator>","CreationDate":"2019-09-19T16:51:03.057","AcceptedAnswerId":null,"OwnerUserId":4462149.0,"Title":"TensorFlow graph error in Estimator (ValueError: Tensor (...) must be from the same graph as Tensor(...))","Body":"<p><strong>UPDATE:<\/strong> Testing the same code with tensorflow-gpu 1.13.1 works both on my PC and on Google Cloud. <\/p>\n\n<hr>\n\n<p>Using TensorFlow Estimator and running train_and_evaluate gives me the following error message:<\/p>\n\n<p><strong><em>\"ValueError: Tensor(\"Const:0\", shape=(3,), dtype=float32) must be from the same graph as Tensor(\"ParallelMapDataset:0\", shape=(), dtype=variant, device=\/device:CPU:0).\"<\/em><\/strong> (see the full error output near bottom)<\/p>\n\n<p>This happens when training the CNN on <strong>my PC<\/strong> with a GPU (GeForge RTX 2070). I am using Python 3.7 with tensorflow-gpu\/tensorflow 1.14.0, Keras 2.2.4, running in a Conda environment.<\/p>\n\n<p>It happens after the following log message \"... Saving checkpoints for 2716 into C:\/EstimatorOutput\/10\/model.ckpt.\" and appear to be when the input function for the evaluation step is being processed.<\/p>\n\n<p>The code, as it is now, has run previously with no issue, but this has suddenly changed for reasons that are unclear to me. <\/p>\n\n<p>I ran similar code on <strong>Google Cloud<\/strong> (which also previously ran fine), and the same problem occur (see error output near bottom; Run on GPU (BASIC_GPU); TensorFlow 1.14; Keras 2.2.4)<\/p>\n\n<p>The error seems to be related to the evaluation step when the graph is created for some reason the new graph is not compatible.<\/p>\n\n<p>Here is my code - ><\/p>\n\n<p>My task module:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nfrom train_model import model #(\"train_model\" is local folder)\nfrom train_model.model import create_estimator \n\nif __name__ == '__main__':\n\n    model_num = 10\n\n    # Throw properties into params dict to pass to other functions\n    params = {}\n    params['train csv'] = \"train_set_local.csv\"\n    params['eval csv'] = \"eval_set_local.csv\"\n    params['output path'] = \"C:\/EstimatorOutput\/\" + str(model_num) + \"\/\"\n    params['data path'] = \"C:\/Databases\/Birds_dB\/Images\"\n    params['image size'] = [244, 224]\n    params[\"batch size\"] = 16*2\n    params['use random flip'] = True\n    params['learning rate'] = 0.000001  \n    params['dropout rate'] = 0.50\n    params['num classes'] = 123\n    params['train steps'] = 65000\n    params['eval steps'] = 20\n    params['eval_throttle_secs'] = 600\n    params['num parallel calls'] = 4\n\n    # Run the training job\n    model.go_train(params) # (See \"go_train\" below in model script -&gt;)\n<\/code><\/pre>\n\n<p>My model module<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nfrom tensorflow.python.keras import estimator as kes\nfrom tensorflow.python.keras.applications.vgg16 import VGG16\nfrom tensorflow.python.keras.models import Model\nfrom tensorflow.python.keras.layers import Dropout, Flatten, Dense\nfrom train_model.input_fn import make_input_fn\n\n\ndef create_estimator(params):\n    # Import VGG16 model for transfer learning\n    base_model = VGG16(weights='imagenet')\n    base_model.summary()\n\n    x = base_model.get_layer('fc2').output\n\n    x = Dropout(params['dropout rate'])(x)\n\n    predictions = Dense(params['num classes'], activation=\"sigmoid\", name=\"sm_out\")(x)\n\n    model = Model(inputs=base_model.input, outputs=predictions)\n\n    for layer in model.layers:\n        layer.trainable = True\n\n    model.compile(\n        loss=\"binary_crossentropy\",\n        optimizer=tf.train.AdamOptimizer(params['learning rate'],\n                                         beta1=0.9,\n                                         beta2=0.999),\n        metrics=[\"categorical_accuracy\"]\n    )\n\n\n\n    config = tf.ConfigProto()\n    config.gpu_options.allow_growth = True\n    config.gpu_options.per_process_gpu_memory_fraction = 0.95\n    run_config = tf.estimator.RunConfig(\n            session_config=config,\n            model_dir=params['output path']\n    )\n\n    # Convert to Estimator\n    estimator_model = kes.model_to_estimator(\n        keras_model=model,\n        config=run_config\n    )\n\n    return estimator_model\n\n\ndef go_train(params):\n    # Create the estimator\n    Est = create_estimator(params)\n\n    # Set up Estimator train and evaluation specifications\n    train_spec = tf.estimator.TrainSpec(\n        input_fn=make_input_fn(params['train csv'], tf.estimator.ModeKeys.TRAIN, params, augment=True),\n        max_steps=params['train steps']\n    )\n    eval_spec = tf.estimator.EvalSpec(\n        input_fn=make_input_fn(params['eval csv'], tf.estimator.ModeKeys.EVAL, params, augment=True),\n        steps=params['eval steps'],  # Evaluates on \"eval steps\" batches\n        throttle_secs=params['eval_throttle_secs']\n    )\n\n\n    # Run training and evaluation\n    tf.estimator.train_and_evaluate(Est, train_spec, eval_spec)\n\n<\/code><\/pre>\n\n<p>My input module:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nfrom keras.applications.vgg16 import preprocess_input\n\ntf.logging.set_verbosity(v=tf.logging.INFO)\n\nHEIGHT = 224\nWIDTH = 224\nNUM_CHANNELS = 3\nNCLASSES = 123\n\n\ndef read_and_preprocess_with_augment(image_bytes, label=None):\n    return read_and_preprocess(image_bytes, label, augment=True)\n\n\ndef read_and_preprocess(image_bytes, label=None, augment=False):\n\n    image = tf.image.decode_jpeg(contents=image_bytes, channels=NUM_CHANNELS)\n    image = tf.image.convert_image_dtype(image=image, dtype=tf.float32)  # 0-1\n    image = tf.expand_dims(input=image, axis=0)  # resize_bilinear needs batches\n\n    if augment:\n\n        # Resize to slightly larger than target size\n        image = tf.image.resize_bilinear(images=image, size=[HEIGHT + 50, WIDTH + 50], align_corners=False)\n\n        # Image random rotation\n        degree_angle = tf.random.uniform((), minval=-25, maxval=25, dtype=tf.dtypes.float32)\n        radian = degree_angle * 3.14 \/ 180\n        image = tf.contrib.image.rotate(image, radian, interpolation='NEAREST')\n\n        # remove batch dimension\n        image = tf.squeeze(input=image, axis=0)\n\n        # Random Crop\n        image = tf.random_crop(value=image, size=[HEIGHT, WIDTH, NUM_CHANNELS])\n        # Random L-R flip\n        image = tf.image.random_flip_left_right(image=image)\n        # Random brightness\n        image = tf.image.random_brightness(image=image, max_delta=63.0 \/ 255.0)\n        # Random contrast\n        image = tf.image.random_contrast(image=image, lower=0.2, upper=1.8)\n\n    else:\n        image = tf.image.resize_bilinear(images=image, size=[HEIGHT, WIDTH], align_corners=False)\n        image = tf.squeeze(input=image, axis=0)  # remove batch dimension\n\n    image = tf.cast(tf.round(image * 255), tf.int32)\n    image = preprocess_input(image)\n\n    label = tf.one_hot(tf.strings.to_number(label, out_type=tf.int32), depth=NCLASSES)\n\n    return {\"input_1\": image}, label\n\n\ndef make_input_fn(csv_of_filenames, mode, params, augment=False):\n    def _input_fn():\n        def decode_csv(csv_row):\n            filename, label = tf.decode_csv(records=csv_row, record_defaults=[[\"\"], [\"\"]])\n            image_bytes = tf.read_file(filename=filename)\n            return image_bytes, label\n\n        # Create tf.data.dataset from filename\n        dataset = tf.data.TextLineDataset(filenames=csv_of_filenames).map(map_func=decode_csv, num_parallel_calls=params['num parallel calls'])\n\n        if augment:\n            dataset = dataset.map(map_func=read_and_preprocess_with_augment, num_parallel_calls=params['num parallel calls'])\n        else:\n            dataset = dataset.map(map_func=read_and_preprocess, num_parallel_calls=params['num parallel calls'])\n\n        if mode == tf.estimator.ModeKeys.TRAIN:\n            num_epochs = None  \n            dataset = dataset.shuffle(buffer_size=10*params[\"batch size\"])\n        else:\n            num_epochs = 1  \n\n        dataset = dataset.repeat(count=num_epochs).batch(batch_size=params[\"batch size\"]).prefetch(4)\n        images, labels = dataset.make_one_shot_iterator().get_next()\n\n        return images, labels\n    return _input_fn\n<\/code><\/pre>\n\n<p><strong>Error output on PC<\/strong><\/p>\n\n<p>As mentioned, the above code when running locally on my GPU results is this series of error messages(abbreviated):<\/p>\n\n<p>Saving checkpoints for 2716 into ....\n...\n...\n  File \"C:...\\estimator.py\", line 501, in _evaluate\n    self._evaluate_build_graph(input_fn, hooks, checkpoint_path))<\/p>\n\n<p>File \"C:...\\estimator.py\", line 1501, in _evaluate_build_graph\n    self._call_model_fn_eval(input_fn, self.config))<\/p>\n\n<p>File \"C:...\\estimator.py\", line 1534, in _call_model_fn_eval\n    input_fn, ModeKeys.EVAL)<\/p>\n\n<p>File \"C:...\\estimator.py\", line 1022, in _get_features_and_labels_from_input_fn\n    self._call_input_fn(input_fn, mode))<\/p>\n\n<p>File \"C:...\\estimator.py\", line 1113, in _call_input_fn\n    return input_fn(**kwargs)<\/p>\n\n<p>File \"C:...\\input_fn.py\", line 71, in _input_fn\n    dataset = dataset.map(map_func=read_and_preprocess_with_augment, num_parallel_calls=params['num parallel calls'])<\/p>\n\n<p>File \"C:...dataset_ops.py\", line 1776, in map\n    self, map_func, num_parallel_calls, preserve_cardinality=False))<\/p>\n\n<p>File \"C:...\\dataset_ops.py\", line 3239, in <strong>init<\/strong>\n    **flat_structure(self))<\/p>\n\n<p>File \"C:...\\gen_dataset_ops.py\", line 4179, in parallel_map_dataset\n    name=name)<\/p>\n\n<p>File \"C:...\\op_def_library.py\", line 366, in _apply_op_helper\n    g = ops._get_graph_from_inputs(_Flatten(keywords.values()))<\/p>\n\n<p>File \"C:...\\ops.py\", line 6135, in _get_graph_from_inputs\n    _assert_same_graph(original_graph_element, graph_element)<\/p>\n\n<p>File \"C:...ops.py\", line 6071, in _assert_same_graph\n    (item, original_item))<\/p>\n\n<p><em>ValueError: Tensor(\"Const:0\", shape=(3,), dtype=float32) must be from the same graph as Tensor(\"ParallelMapDataset:0\", shape=(), dtype=variant, device=\/device:CPU:0).<\/em><\/p>\n\n<p><strong>Error output on Google Cloud<\/strong><\/p>\n\n<p>service\nThe replica master 0 exited with a non-zero status of 1. \nTraceback (most recent call last): [...] <\/p>\n\n<p>File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow_estimator\/python\/estimator\/estimator.py\", line 1534, in _call_model_fn_eval input_fn, ModeKeys.EVAL) <\/p>\n\n<p>File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow_estimator\/python\/estimator\/estimator.py\", line 1022, in _get_features_and_labels_from_input_fn self._call_input_fn(input_fn, mode)) <\/p>\n\n<p>File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow_estimator\/python\/estimator\/estimator.py\", line 1113, in _call_input_fn return input_fn(**kwargs) <\/p>\n\n<p>File \"\/root\/.local\/lib\/python3.5\/site-packages\/train_model\/input_fn.py\", line 87, in _input_fn dataset = dataset.map(map_func=read_and_preprocess_with_augment, num_parallel_calls=params['num parallel calls']) <\/p>\n\n<p>File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/data\/ops\/dataset_ops.py\", line 1776, in map self, map_func, num_parallel_calls, preserve_cardinality=False)) <\/p>\n\n<p>File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/data\/ops\/dataset_ops.py\", line 3239, in <strong>init<\/strong> **flat_structure(self)) File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/ops\/gen_dataset_ops.py\", line 4179, in parallel_map_dataset name=name) File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/framework\/op_def_library.py\", line 366, in _apply_op_helper g = ops._get_graph_from_inputs(_Flatten(keywords.values())) <\/p>\n\n<p>File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 6135, in _get_graph_from_inputs _assert_same_graph(original_graph_element, graph_element) <\/p>\n\n<p>File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/framework\/ops.py\", line 6071, in _assert_same_graph (item, original_item)) <\/p>\n\n<p><em>ValueError: Tensor(\"Const_1:0\", shape=(3,), dtype=float32, device=\/device:CPU:0) must be from the same graph as Tensor(\"ParallelMapDataset:0\", shape=(), dtype=variant, device=\/device:CPU:0).<\/em> <\/p>\n\n<p><strong><em>Any help\/hint is much appreciated. I am stuck at this point and do not know how to debug this one!<\/em><\/strong><\/p>\n","answers":[{"AnswerId":"58184402","CreationDate":"2019-10-01T12:05:01.497","ParentId":null,"OwnerUserId":"9973234","Title":null,"Body":"<p>I was facing the same issue but somehow it got solved by using the preprocess input function from mobilenet for all the models like VGG, MobileNet, ResNet and Xception.<\/p>\n\n<p>use this line:<\/p>\n\n<p><code>from tensorflow.keras.applications.mobilenet import preprocess_input<\/code><\/p>\n\n<p>instead of:<\/p>\n\n<p><code>from keras.applications.vgg16 import preprocess_input<\/code><\/p>\n"},{"AnswerId":"58184400","CreationDate":"2019-10-01T12:04:52.997","ParentId":null,"OwnerUserId":"11652623","Title":null,"Body":"<p>use this preprocess function:<\/p>\n\n<pre><code>from tensorflow.keras.applications.mobilenet import preprocess_input\n<\/code><\/pre>\n\n<p>It has same functionality to that of VGGs preprocess input.<\/p>\n"}]}
+{"QuestionId":58015766,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><tensorflow2.0><tf.keras>","CreationDate":"2019-09-19T16:57:31.823","AcceptedAnswerId":null,"OwnerUserId":8382950.0,"Title":"Setting adaptable learning rate for Adam without using callbacks","Body":"<p>I am exploring the translation model with attention from Tensorflow docs - <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/text\/nmt_with_attention\" rel=\"nofollow noreferrer\">NMT with Attention<\/a> . Here in TF 2.0, optimizer is defined as <\/p>\n\n<pre><code>optimizer = tf.keras.optimizers.Adam()\n<\/code><\/pre>\n\n<p>How do I set a learning rate in this case? Is it just initializing the argument like below? How do I set an adaptable learning rate?<\/p>\n\n<pre><code>tf.keras.optimizers.Adam(learning_rate=0.001)\n<\/code><\/pre>\n\n<p>In the <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/text\/nmt_with_attention\" rel=\"nofollow noreferrer\">NMT with Attention<\/a> model, they dont use Keras to define the model and I could not use Callbacks or 'model.fit' like below:<\/p>\n\n<pre><code>model.fit(x_train, y_train, callbacks=[LearningRateReducerCb()], epochs=5)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58016567,"AnswerCount":0,"Tags":"<python><tensorflow><keras><face-recognition><dlib>","CreationDate":"2019-09-19T17:54:44.433","AcceptedAnswerId":null,"OwnerUserId":12091488.0,"Title":"Keras TypeError: incompatible function arguments (dlib model)","Body":"<p>I am generating some face images using <code>Conv2DTranspose<\/code> followed by a 1x1 <code>Conv2D<\/code>. I need to have a face detection Keras custom layer that later will be used to generate facial landmarks for my loss function.  <\/p>\n\n<p>I used <code>dlib.cnn_face_detection_model_v1<\/code> to detect faces but the problem is it only accepts list or arrays and not tensors. Plus in graph definition stage the tensors have not yet been fed with any values. So, it keeps throwing me this error:<\/p>\n\n<pre><code>TypeError: __call__(): incompatible function arguments. The following argument types are supported:\n    1. (self: dlib.cnn_face_detection_model_v1, imgs: list, upsample_num_times: int=0, batch_size: int=128) -&gt; std::vector&lt;std::vector&lt;dlib::mmod_rect,std::allocator&lt;dlib::mmod_rect&gt; &gt;,std::allocator&lt;std::vector&lt;dlib::mmod_rect,std::allocator&lt;dlib::mmod_rect&gt; &gt; &gt; &gt;\n    2. (self: dlib.cnn_face_detection_model_v1, img: array, upsample_num_times: int=0) -&gt; std::vector&lt;dlib::mmod_rect,std::allocator&lt;dlib::mmod_rect&gt; &gt;\n\nInvoked with: &lt;dlib.cnn_face_detection_model_v1 object at 0x0000013474884D50&gt;, &lt;tf.Tensor 'conv2d\/BiasAdd:0' shape=(?, 160, 160, 3) dtype=float32&gt;, 1\n\nDid you forget to `#include &lt;pybind11\/stl.h&gt;`? Or &lt;pybind11\/complex.h&gt;,\n&lt;pybind11\/functional.h&gt;, &lt;pybind11\/chrono.h&gt;, etc. Some automatic\nconversions are optional and require extra headers to be included\nwhen compiling your pybind11 module.\n<\/code><\/pre>\n\n<p>Here's how I defined my landmarks detector layer:<\/p>\n\n<pre><code>class landmarks_layer(Layer):\n  def __init__(self, num_outputs):\n    super(landmarks_layer, self).__init__()\n    self.num_outputs = num_outputs\n  def call(self, input_):\n    landmarks = find_landmarks68(input_)\n    return landmarks\n<\/code><\/pre>\n\n<p>and here's my function to detect landmarks:<\/p>\n\n<pre><code>def find_landmarks68(img):\n    cnn_face_detector =   \n    dlib.cnn_face_detection_model_v1('mmod_human_face_detector.dat')\n    dets = cnn_face_detector(img, 1)\n    rects = dlib.rectangles()\n    rects.extend([d.rect for d in dets])\n    predictor = dlib.shape_predictor(\"shape_predictor_68_face_landmarks.dat\") \n    shape = predictor(img, rects[0])\n    (img_y,img_x,img_z) = img.shape\n    landmarks = []\n    for i in shape.parts():\n        landmarks.append([i.y, i.x])    \n    return landmarks\n<\/code><\/pre>\n\n<p>I expected that the layer would return a (None, 68, 2) tensor in graph definition stage but it executes the code and when it faces a tensor with no value, it throws me the error. So, I am stuck in the face detection stage right now and it doesn't proceed to landmarks detection. I thought I should've used <code>tf.cond<\/code> for checking on whether the batch number is None or not and if it is I don't know it just return a (None, 68, 2) tensor but if it is not, then it should've performed the landmarks detection in graph execution stage but it runs the code even graph definition stage.<\/p>\n","answers":[]}
+{"QuestionId":58017274,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-19T18:45:41.467","AcceptedAnswerId":null,"OwnerUserId":12092080.0,"Title":"Simple Custom Layer fails to build in simple model","Body":"<p>When using a custom class to replace a lambda function, building a model fails.<\/p>\n\n<p>I previously had this code in a lambda function and it worked fine but I was unable to save a model.  I need to save the model that I'm building which has dependencies on this code snippet.<\/p>\n\n<pre><code>import keras\nimport tensorflow as tf \n\nclass ShapePositionLayer(keras.layers.Layer): \n\n    def call(self, x):\n         assert isinstance(x, list)\n         a, b = x\n         return keras.backend.gather(keras.backend.shape(a), b)\n\n    def compute_output_shape(self, input_shape):\n        return (1)\n\ncaptions = keras.layers.Input(shape=[5,1024], name='captions')\nbatch_size  = ShapePositionLayer()([captions,tf.constant(0, \ndtype=tf.int32)])\nmodel = keras.models.Model(inputs=[captions], outputs=[batch_size])\n<\/code><\/pre>\n\n<p>I expected to be able to build a model.<\/p>\n\n<p>Instead receive error:\nAttributeError: 'NoneType' object has no attribute '_inbound_nodes'<\/p>\n","answers":[]}
+{"QuestionId":58017567,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-19T19:07:47.147","AcceptedAnswerId":58029774.0,"OwnerUserId":5985593.0,"Title":"Keras (tensorflow) finds GPU, but only runs on cpu w\/ Cuda 10.1","Body":"<p>Already many questions are posted about this, but none of them really answers mine or there is a small difference with what I came across. <\/p>\n\n<p>I'm on ubuntu 18.04 and installed keras following the default instructions with CUDA 10.1 AND tensorflow-gpu.<\/p>\n\n<p>When running something tensorflow detects I have a GPU, but when I'm checking cpu vs gpu usage, he still only seem to run on cpu. I came across <a href=\"https:\/\/datascience.stackexchange.com\/a\/41958\">this<\/a> thread and run that script. It confirms what I was guessing, that he can't use my gpu for some reason:<\/p>\n\n<pre><code>2019-09-19 21:05:57.730197: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudnn.so.7\n2019-09-19 21:05:57.730247: W tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1663] Cannot dlopen some GPU libraries. Skipping registering GPU devices...\n2019-09-19 21:05:57.730281: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1181] Device interconnect StreamExecutor with strength 1 edge matrix:\n2019-09-19 21:05:57.730303: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1187]      0 \n2019-09-19 21:05:57.730317: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1200] 0:   N \n2019-09-19 21:05:57.922335: W tensorflow\/compiler\/jit\/mark_for_compilation_pass.cc:1412] (One-time warning): Not using XLA:CPU for cluster because envvar TF_XLA_FLAGS=--tf_xla_cpu_global_jit was not set.  If you want XLA:CPU, either set that envvar, or use experimental_jit_scope to enable XLA:CPU.  To confirm that XLA is active, pass --vmodule=xla_compilation_cache=1 (as a proper command-line flag, not via TF_XLA_FLAGS) or set the envvar XLA_FLAGS=--xla_hlo_profile.\n<\/code><\/pre>\n\n<p>When listing the devices it says:<\/p>\n\n<pre><code>[name: \"\/device:CPU:0\"\ndevice_type: \"CPU\"\nmemory_limit: 268435456\nlocality {\n}\nincarnation: 57580461479478464\n, name: \"\/device:XLA_GPU:0\"\ndevice_type: \"XLA_GPU\"\nmemory_limit: 17179869184\nlocality {\n}\nincarnation: 6376288845656491190\nphysical_device_desc: \"device: XLA_GPU device\"\n, name: \"\/device:XLA_CPU:0\"\ndevice_type: \"XLA_CPU\"\nmemory_limit: 17179869184\nlocality {\n}\nincarnation: 17409275481256463364\nphysical_device_desc: \"device: XLA_CPU device\"\n]\n<\/code><\/pre>\n\n<p>But halfway the logs, tensorflow outputs this:<\/p>\n\n<pre><code>2019-09-19 20:44:32.676537: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1640] Found device 0 with properties: \nname: GeForce GTX 860M major: 5 minor: 0 memoryClockRate(GHz): 1.0195\npciBusID: 0000:01:00.0\n<\/code><\/pre>\n\n<p>.\/deviceQuery outputs this:<\/p>\n\n<pre><code>.\/deviceQuery Starting...\n\n CUDA Device Query (Runtime API) version (CUDART static linking)\n\nDetected 1 CUDA Capable device(s)\n\nDevice 0: \"GeForce GTX 860M\"\n  CUDA Driver Version \/ Runtime Version          10.1 \/ 10.1\n  CUDA Capability Major\/Minor version number:    5.0\n  Total amount of global memory:                 2004 MBytes (2101870592 bytes)\n  ( 5) Multiprocessors, (128) CUDA Cores\/MP:     640 CUDA Cores\n  GPU Max Clock rate:                            1020 MHz (1.02 GHz)\n  Memory Clock rate:                             2505 Mhz\n  Memory Bus Width:                              128-bit\n  L2 Cache Size:                                 2097152 bytes\n  Maximum Texture Dimension Size (x,y,z)         1D=(65536), 2D=(65536, 65536), 3D=(4096, 4096, 4096)\n  Maximum Layered 1D Texture Size, (num) layers  1D=(16384), 2048 layers\n  Maximum Layered 2D Texture Size, (num) layers  2D=(16384, 16384), 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 1 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  Device supports Compute Preemption:            No\n  Supports Cooperative Kernel Launch:            No\n  Supports MultiDevice Co-op Kernel Launch:      No\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) &gt;\n\ndeviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 10.1, CUDA Runtime Version = 10.1, NumDevs = 1\nResult = PASS\n<\/code><\/pre>\n\n<p>Anyone knows why tensorflow can't find my GPU or how to make it available?<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"58029774","CreationDate":"2019-09-20T13:55:08.203","ParentId":null,"OwnerUserId":"9637085","Title":null,"Body":"<p>It's because of cuda 10.1. You need to downgrade to cuda 10.0.<\/p>\n\n<p>Here's a similar <a href=\"https:\/\/stackoverflow.com\/a\/56803006\/9637085\">solution<\/a>.<\/p>\n"}]}
+{"QuestionId":58018009,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-09-19T19:43:43.970","AcceptedAnswerId":null,"OwnerUserId":9188181.0,"Title":"AttributeError: 'gevent._local.local' object has no attribute 'value'","Body":"<p>I have a keras Sequential model which works perfectly fine during model training, But when I am trying to run model.predict command on trained model, <\/p>\n\n<p>I am getting following error:-<\/p>\n\n<pre><code>Traceback (most recent call last):\n File \"\/opt\/program\/predictor.py\", line 69, in predict\n    a=model.predict(x_input, verbose=0)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/engine\/training.py\", line 1456, in predict\n    self._make_predict_function()\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/engine\/training.py\", line 378, in _make_predict_function\n    **kwargs)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/engine\/network.py\", line 323, in __setattr__\n    super(Network, self).__setattr__(name, value)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/engine\/base_layer.py\", line 1215, in __setattr__\n    if not _DISABLE_TRACKING.value:\n  File \"src\/gevent\/local.py\", line 408, in gevent._local.local.__getattribute__\n\nAttributeError: 'gevent._local.local' object has no attribute 'value'\n<\/code><\/pre>\n\n<p>I have searched for this error but couldn't found any solution. \nDoes anyone know any fix for this?<\/p>\n","answers":[{"AnswerId":"58018010","CreationDate":"2019-09-19T19:43:43.970","ParentId":null,"OwnerUserId":"9188181","Title":null,"Body":"<p>I got the reason for this error:<\/p>\n\n<p>My docker container automatically installs latest keras version which got changed to 2.3.0(New stable version) which was not available till last week so my container was working perfectly fine previously. Keras new stable version : 2.3.0 might have some issue with Gevent which was causing this conflict.<\/p>\n\n<p><strong>I uninstalled keras and installed 2.2.5 version which resolved the error.<\/strong><\/p>\n\n<pre><code>pip uninstall keras\npip install keras==2.2.5\n<\/code><\/pre>\n"}]}
+{"QuestionId":58018623,"AnswerCount":1,"Tags":"<python><tensorflow><object-detection-api>","CreationDate":"2019-09-19T20:36:01.403","AcceptedAnswerId":null,"OwnerUserId":1886357.0,"Title":"Input image size for tensorflow faster-rcnn in prediction mode?","Body":"<p>I have just done some transfer learning with a faster-rcnn using tensorflow object detection api. I am in tensorflow 1.14, the backbone network is faster_rcnn_resnet101_coco. Do frozen networks resize images fed to them when making predictions? <\/p>\n\n<p>I ask because when I feed the model an image that is much larger than those I trained on, it doesn't recognize any of the objects. When I crop the image down to 1200x1200, the objects are all identical, but it works great. <\/p>\n\n<p>Does the model include image size constraints? Should I be making predictions using similar dimensions to those in the config file, even though the objects are the same size in the 3000x3000 image?<\/p>\n\n<p>In the config file for training, I constrain the input images:<\/p>\n\n<pre><code>image_resizer {\n  keep_aspect_ratio_resizer {\n    min_dimension: 200\n    max_dimension: 1200\n  }\n}\n<\/code><\/pre>\n\n<p>Does this mean that in the trained model, that I am now using, if I feed it an image larger than 1200x1200, it will scale it down? Here is how I do the prediction in the loaded frozen model:<\/p>\n\n<pre><code>with model.as_default():\n        with tf.Session(graph=model) as sess:\n            imageTensor = model.get_tensor_by_name(\"image_tensor:0\")\n            boxesTensor = model.get_tensor_by_name(\"detection_boxes:0\")\n            scoresTensor = model.get_tensor_by_name(\"detection_scores:0\")\n            classesTensor = model.get_tensor_by_name(\"detection_classes:0\")\n            numDetections = model.get_tensor_by_name(\"num_detections:0\")\n\n            # Make prediction\n            (boxes, scores, labels, N) = sess.run([boxesTensor, scoresTensor, classesTensor, numDetections],\n                                                   feed_dict = {imageTensor: image})\n<\/code><\/pre>\n\n<p>Related: <a href=\"https:\/\/stackoverflow.com\/questions\/57855824\/training-image-size-faster-rcnn\">Training Image Size Faster-RCNN<\/a><\/p>\n\n<p>Also, this post makes me think it should handle any input size, but it clearly doesn't handle them the same, so I'm confused: <a href=\"https:\/\/stackoverflow.com\/questions\/53387059\/faster-rcnn-inception-v2-input-size?rq=1\">Faster RCNN + inception v2 input size<\/a><\/p>\n","answers":[{"AnswerId":"58600604","CreationDate":"2019-10-29T02:06:38.777","ParentId":null,"OwnerUserId":"12288987","Title":null,"Body":"<p>In <a href=\"https:\/\/arxiv.org\/pdf\/1612.03144.pdf\" rel=\"nofollow noreferrer\">Feature Pyramid Networks for Object Detection<\/a>, Faster RCNN shows different mAP on object of different size. The model has higher mAP on large objects than on small objects. In <a href=\"http:\/\/papers.nips.cc\/paper\/5638-faster-r-cnn-towards-real-time-object-detection-with-region-proposal-networks.pdf\" rel=\"nofollow noreferrer\">Faster R-CNN: Towards Real-Time Object Detection\nwith Region Proposal Networks<\/a>, faster RCNN resizes input images such that their shorter side is 600 pixels. Thus, in my option, relative size of objects in images does matter in detection. Cropping a large image and use the smaller image as input may facilitate the detection of small objects in the raw image for small objects become relatively large objects in the new image. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/4xReZ.png\" rel=\"nofollow noreferrer\">FPN in a basic Faster R-CNN system has different performance on small, middle and large objects.<\/a><\/p>\n\n<ol>\n<li><a href=\"https:\/\/github.com\/rbgirshick\/py-faster-rcnn\/issues\/304\" rel=\"nofollow noreferrer\">Discussion on GitHub<\/a>   <\/li>\n<li><a href=\"https:\/\/github.com\/rbgirshick\/py-faster-rcnn\/issues\/371\" rel=\"nofollow noreferrer\">Another discussion on GitHub<\/a><\/li>\n<\/ol>\n"}]}
+{"QuestionId":58019166,"AnswerCount":0,"Tags":"<python><tensorflow><tf.keras>","CreationDate":"2019-09-19T21:31:04.667","AcceptedAnswerId":null,"OwnerUserId":8089695.0,"Title":"Tensorflow He initializer for layers with with skip connections","Body":"<p>In tensorflow <a href=\"https:\/\/arxiv.org\/pdf\/1502.01852.pdf\" rel=\"nofollow noreferrer\">he initializer<\/a> is implemented <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/initializers\/he_uniform\" rel=\"nofollow noreferrer\">here<\/a>. My network has additive skip connections (like residual networks) in it. So the <code>fan_in<\/code> should change accordingly if unit variance is to be maintained by the network. Does the tensorflow initializer take care of that or do i need to write my own initializer for that?<\/p>\n","answers":[]}
+{"QuestionId":58019431,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><deep-learning>","CreationDate":"2019-09-19T22:04:26.493","AcceptedAnswerId":null,"OwnerUserId":12071911.0,"Title":"Tensorflow: Loss function using scatter_nd: ValueError: No gradients provided for any variable","Body":"<p>I am trying to process the output of my network before I compute the loss. The network outputs 2 values, let's say x,y. I want to generate a histogram matrix for the output. Basically, I want to count the number of times a particular (x,y) occurs and store that value in a matrix at (x,y) in it. I am able to successfully do it for my ground truth an predictions using <strong>tf.scatter_nd<\/strong> but when I calculate the l2 loss between the ground truth count matrix and predicted count matrix, I get <strong>ValueError: No gradients provided for any variable, check your graph for ops that do not support gradients<\/strong>. Below I have shared a code, please let me know the mistake or the way around. <\/p>\n\n<pre><code>l2_loss = tf.constant([0], dtype=tf.float32)\n for j in range(batch_size):\n    count_matrix_estimated_frame1 = tf.scatter_nd(int_estimated_frame1[j,:,:], value_estimated_frame1, shape, name='scatter_nd1')\n    count_matrix_frame1_gt_events = tf.scatter_nd(int_frame1_gt_events[j,:,:], value_frame1_gt_events, shape, name='scatter_nd2')\n    count_matrix_frame2_gt_events = tf.scatter_nd(int_frame2_gt_events[j,:,:], value_frame1_gt_events, shape, name='scatter_nd3')\n\n    l2_loss = tf.add(tf.reduce_mean(\\\n                tf.reduce_sum((count_matrix_estimated_frame1 - count_matrix_frame1_gt_events)\\\n                    *(count_matrix_estimated_frame1 - count_matrix_frame1_gt_events), axis=1, name='reduce_sum')\/2.0, name='reduce_mean'), l2_loss)\n\nreturn l2_loss\n<\/code><\/pre>\n\n<p>Above mentioned is the excerpt of the loss function. The function just contains declarations of the variables other than this. When I run the code using this loss function I get the following error: <\/p>\n\n<pre><code>File \"train.py\", line 300, in &lt;module&gt;\n    train()\n  File \"train.py\", line 128, in train\n    train_op = optimizer.minimize(loss, global_step=batch)\n  File \"\/home\/&lt;name_of_user&gt;\/.local\/lib\/python3.5\/site-packages\/tensorflow\/python\/training\/optimizer.py\", line 406, in minimize\n    ([str(v) for _, v in grads_and_vars], loss))\nValueError: No gradients provided for any variable, check your graph for ops that do not support gradients, \nbetween variables [\"&lt;tf.Variable 'Variable:0' shape=() dtype=int32_ref&gt;\", \"&lt;tf.Variable 'sa1\/layer1\/conv0\/weights:0' \nshape=(1, 1, 6, 32) dtype=float32_ref&gt;\" ........\n\n<\/code><\/pre>\n\n<p>The error continues<\/p>\n\n<p>I first thought that scatter doesn't have a gradient but using get_gradient_function() I found that scatter_nd has gradient. I don't know what the problem. Any help would be great.<\/p>\n","answers":[]}
+{"QuestionId":58019435,"AnswerCount":1,"Tags":"<pytorch><imshow>","CreationDate":"2019-09-19T22:05:31.513","AcceptedAnswerId":null,"OwnerUserId":3983454.0,"Title":"how to identify wrong classification with batches in pytorch","Body":"<p>I have a script like this where batches of images were used <\/p>\n\n<pre><code>correct = 0\ntotal = 0\nincorrect_classification=[]\nfor (i, [images, labels]) in enumerate(test_loader):\n  images = Variable(images.view(-1, n_pixel*n_pixel))\n  outputs = net(images)\n  _, predicted = torch.min(outputs.data, 1)\n  total += labels.size(0)                    \n  correct += (predicted == labels).sum() \nprint('Accuracy: %d %%' %\n      (100 * correct \/ total))\n<\/code><\/pre>\n\n<p>With a batch size of 10, each enumerate returns 10 x image size tensor. How can I save all the wrong classifications into the array incorrect_classification or wrong img and their probability into an dictionary so I use can plt.imshow to inspect them later? <\/p>\n\n<p>If batch size is 1 I could use this:<\/p>\n\n<pre><code>if (predicted==labels).item()==0:\n    incorrect_examples.append(images.numpy())\n<\/code><\/pre>\n\n<p>But with a batch size specified (like 100 images per batch) how should I save the wrong classifications?<\/p>\n\n<p>Thanks in advance for any answers.<\/p>\n","answers":[{"AnswerId":"58044708","CreationDate":"2019-09-21T22:54:40.377","ParentId":null,"OwnerUserId":"3058060","Title":null,"Body":"<p>As already said in the comment of @zihaozhihao, <code>images[predicted==labels]<\/code> should do the work.<\/p>\n\n<p>In other words, you will get a mask of indexes and then access the images you want with this mask:<\/p>\n\n<pre><code>correct = 0\ntotal = 0\nincorrect_examples=[]\nfor (i, [images, labels]) in enumerate(test_loader):\n    images = Variable(images.view(-1, n_pixel*n_pixel))\n    outputs = net(images)\n    _, predicted = torch.min(outputs.data, 1)\n    total += labels.size(0)                    \n    correct += (predicted == labels).sum() \n    print('Accuracy: %d %%' % (100 * correct \/ total))\n\n    # if (predicted==labels).item()==0:\n    #     incorrect_examples.append(images.numpy())\n\n    idxs_mask = (predicted == labels).view(-1)\n    incorrect_examples.append(images[idxs_mask].numpy()) \n<\/code><\/pre>\n\n<p>The <code>view(-1)<\/code> will flatten the mask which will be used to mask the batch channel of the images tensor. <\/p>\n\n<p>At the end of the loop (out of it), the itens in the list <code>incorrect_examples<\/code> will have shape <code>[batch_size, n_pixel, n_pixel]<\/code>, and for convenience, you can group all of them in one tensor by concatenating them:<\/p>\n\n<pre><code>incorrect_images = torch.cat(incorrect_examples)\n# incorrect_images.size() -&gt; (n_incorrect_images, n_pixel, n_pixel)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58019950,"AnswerCount":1,"Tags":"<python><docker><keras><theano><h5py>","CreationDate":"2019-09-19T23:24:51.007","AcceptedAnswerId":null,"OwnerUserId":11724570.0,"Title":"Keras model.load_weights() error: ValueError: Invalid high library version bound (invalid high library version bound)","Body":"<p>I use <a href=\"https:\/\/hub.docker.com\/r\/gadgetron\/ubuntu_1804_no_cuda\" rel=\"nofollow noreferrer\">Docker from Gadgetron<\/a> to realize a Gadget with Keras. Theano as backend. Here is the <a href=\"https:\/\/github.com\/thomaskuestner\/CNNArt\/blob\/yang_liu\/Gadgetron\/Dockerfile\" rel=\"nofollow noreferrer\">Dockerfile<\/a> and the result of <code>pip list<\/code> is <\/p>\n\n<pre><code>Package              Version    \n-------------------- -----------\nabsl-py              0.7.1      \nastor                0.8.0      \ncertifi              2019.6.16  \nchardet              3.0.4      \ncycler               0.10.0     \nCython               0.26.1     \ndecorator            4.4.0      \ndicom                0.9.9.post1\ndicom-numpy          0.1.6      \nengineering-notation 0.6.0      \ngast                 0.2.2      \ngoogle-pasta         0.1.7      \ngrpcio               1.21.1     \nh5py                 2.9.0      \nidna                 2.8        \nimageio              2.5.0      \nismrmrd              1.6.1      \nismrmrd-python-tools 0.3        \njoblib               0.13.2     \nKeras                2.2.4      \nKeras-Applications   1.0.8      \nKeras-Preprocessing  1.1.0      \nkiwisolver           1.1.0      \nlxml                 4.2.1      \nMarkdown             3.1.1      \nmatplotlib           3.1.0      \nnetworkx             2.3        \nnumpy                1.16.4     \nopencv-python        4.1.0.25   \nPillow               6.0.0      \npip                  19.2.3     \nprotobuf             3.8.0      \npsutil               5.4.2      \npydicom              1.2.2      \npygobject            3.26.1     \npynrrd               0.4.0      \npyparsing            2.4.0      \nPyQt5                5.12.2     \nPyQt5-sip            4.19.17    \npython-apt           1.6.4      \npython-dateutil      2.8.0      \npytz                 2019.2     \nPyWavelets           1.0.3      \nPyXB                 1.2.6      \nPyYAML               5.1.1      \npyzmq                18.0.2     \nrequests             2.22.0     \nscikit-image         0.15.0     \nscikit-learn         0.21.2     \nscipy                1.3.0      \nsetuptools           41.0.1     \nsix                  1.12.0     \nsklearn              0.0        \ntensorboard          1.14.0     \ntensorboardX         1.8        \ntensorflow           1.14.0     \ntensorflow-estimator 1.14.0     \ntensorflow-gpu       1.14.0     \ntermcolor            1.1.0      \nTheano               1.0.4      \ntk-tools             0.11.0     \ntorch                1.1.0      \ntorchfile            0.1.0      \ntorchvision          0.3.0      \ntornado              6.0.3      \nurllib3              1.25.3     \nvisdom               0.1.8.8    \nwebsocket-client     0.56.0     \nWerkzeug             0.15.4     \nwheel                0.33.4     \nwrapt                1.11.2     \n<\/code><\/pre>\n\n<p>Mycode is like this:<\/p>\n\n<pre><code>from keras.models import model_from_json\n\nwith open('\/opt\/data\/model.json', 'r') as f:\n    model = model_from_json(f.read())\nmodel.load_weights('\/opt\/data\/weight.h5')\n<\/code><\/pre>\n\n<p>When loading the weights, an error appeared:<\/p>\n\n<pre><code>File \"\/opt\/data\/CNNArt\/Gadgetron\/gadget_cnnart.py\", line 29, in process\n    model.load_weights('\/opt\/data\/weight.h5')\n\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/keras\/engine\/network.py\", line 1157, in load_weights\n    with h5py.File(filepath, mode='r') as f:\n\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/h5py\/_hl\/files.py\", line 391, in __init__\n    fapl = make_fapl(driver, libver, rdcc_nslots, rdcc_nbytes, rdcc_w0, **kwds)\n\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/h5py\/_hl\/files.py\", line 108, in make_fapl\n    plist.set_libver_bounds(low, high)\n\n  File \"h5py\/_objects.pyx\", line 54, in h5py._objects.with_phil.wrapper\n\n  File \"h5py\/_objects.pyx\", line 55, in h5py._objects.with_phil.wrapper\n\n  File \"h5py\/h5p.pyx\", line 1140, in h5py.h5p.PropFAID.set_libver_bounds\n\nValueError: Invalid high library version bound (invalid high library version bound)\n<\/code><\/pre>\n\n<p>However, in Docker, when I opened python <strong>in the terminal<\/strong>, I can load the model and its parameters. <\/p>\n\n<pre><code>&gt;&gt;&gt; from keras.models import model_from_json\n&gt;&gt;&gt; with open('\/opt\/data\/model.json', 'r') as f:\n...     model = model_from_json(f.read())\n...\n&gt;&gt;&gt; model.load_weights('\/opt\/data\/weight.h5')\n&gt;&gt;&gt; \n<\/code><\/pre>\n\n<p>I have tried to reinstall <code>h5py<\/code> with <code>pip<\/code>, but it doesn't work. <\/p>\n\n<p>The problem is it works directly in python interaction terminal but doesn't work in Gadget script.<\/p>\n","answers":[{"AnswerId":"58044535","CreationDate":"2019-09-21T22:16:23.780","ParentId":null,"OwnerUserId":"11724570","Title":null,"Body":"<p>Uninstall h5py and reinstall with this command:<\/p>\n\n<pre><code>pip install --no-binary=h5py h5py \n<\/code><\/pre>\n\n<p><a href=\"https:\/\/fenicsproject.discourse.group\/t\/gmsh-4-4-1-in-fenics-meshio\/1415\" rel=\"nofollow noreferrer\">Source<\/a><\/p>\n"}]}
+{"QuestionId":58020289,"AnswerCount":0,"Tags":"<tensorflow><keras><checkpointing>","CreationDate":"2019-09-20T00:26:47.040","AcceptedAnswerId":null,"OwnerUserId":11876924.0,"Title":"TF Keras ModelCheckpoint filepath batch number","Body":"<p>I am using <code>ModelCheckpoint<\/code> to save checkpoints every 500 batches in every epoch. It is documented here <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/callbacks\/ModelCheckpoint\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/callbacks\/ModelCheckpoint<\/a>.<\/p>\n\n<p>How would I set <code>filepath<\/code> to include the batch number? I know I can use <code>{epoch}<\/code> and parameters in <code>logs<\/code>. <\/p>\n","answers":[]}
+{"QuestionId":58020333,"AnswerCount":0,"Tags":"<python><machine-learning><deep-learning><computer-vision><pytorch>","CreationDate":"2019-09-20T00:34:40.410","AcceptedAnswerId":null,"OwnerUserId":11960163.0,"Title":"3D Conv on non-overlapping blocks of a 3D tensor then remapping them (pytorch)","Body":"<p>Hi I have a 3D tensor of size (128,128,128) as input to my model. When it enters the model it has the shape (8, 4, 128, 128, 128) which is (Batch, Channels, H, W, D). <\/p>\n\n<p>I would like to separate the channels and perform a convolution on blocks of (32,32,32) for this (128,128,128) input. Then I wish to take the conv weights and multiply it with the input values to the conv and remap them to a (128,128,128) block. <\/p>\n\n<p>My current inefficient solution (using many for loops, converting between numpy&lt;->tensor, and scikit-image) is below however it takes too long and requires too much memory. What's the best way to do this on tensors?<\/p>\n\n<pre><code>from skimage.util.shape import view_as_blocks\n\nclass LFBlock(nn.Module):\n    def __init__(self, input_shape=(128,128,128), kernel_size=(1,1,1), blk_div=4):\n        super(LFBlock,self).__init__()      \n\n        # Divides the (128,128,128)\/\/4 -&gt; (32,32,32)\n        self.block_shape = (input_shape[0]\/\/blk_div, input_shape[1]\/\/blk_div, input_shape[2]\/\/blk_div)\n\n        self.num_blocks = (input_shape[0]\/\/self.block_shape[0])*(input_shape[0]\/\/self.block_shape[0])*\\\n        (input_shape[0]\/\/self.block_shape[0])\n\n        conv_list = []\n        for n in range(self.num_blocks):\n            conv_list.append(nn.Conv3d(1,1, kernel_size=kernel_size, stride=1, padding=0, bias=True))\n\n        self.conv1x1s = nn.ModuleList(conv_list)        \n\n    def forward(self, lf_in):\n        # Batch\n        for i in range(lf_in.shape[0]):\n            # Modality\n            for ch in range(lf_in.shape[1]):\n                x_lf = lf_in[i,ch,:]\n                lf_blocks = view_as_blocks(x_lf.cpu().numpy(), block_shape=self.block_shape)\n\n                # Do Conv3d on each block \n                for x in range(len(lf_blocks)):\n                    for y in range(len(lf_blocks)):\n                        for z in range(len(lf_blocks)):\n\n                            conv_idx = x*len(lf_blocks) + y*len(lf_blocks) + z\n\n                            # Convolve the block, then multiply with the weight of the block. \n                            tensor_img = torch.from_numpy(lf_blocks[x,y,z])[None, None,:]\n                            conv = self.conv1x1s[conv_idx](tensor_img.cuda())\n\n                            # w * x. \n                            # view_as_blocks returns a view so modifications are done in-place\n\n                            lf_blocks[x,y,z] = tensor_img.cpu()*self.conv1x1s[conv_idx].weight.data.cpu()\n\n        # Linearly sum the modalities together\n        # out = w0*x0 + w1*x1 + w2*x2 + w3*x3 \n        out = (lf_in[:,0]+lf_in[:,1]+lf_in[:,2]+lf_in[:,3])[:,None]\n\n        return out\n<\/code><\/pre>\n\n<p>Any help is appreciated. Thank you!<\/p>\n","answers":[]}
+{"QuestionId":58020635,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-20T01:26:13.527","AcceptedAnswerId":null,"OwnerUserId":1118236.0,"Title":"Set the shape of a Variable to the batch size in TensorFlow","Body":"<p>I am trying to initialize a Tensorflow variable, whose shape should be the same as a placeholder:<\/p>\n\n<pre><code>import tensorflow as tf\nevents = tf.placeholder(tf.float32, shape=[None], name='events')\nopt_res = tf.get_variable('opt_res', initializer=tf.truncated_normal(shape=tf.shape(events), mean=0.0, stddev=1e-1))\n\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n    init.run()\n    sess.run([opt_res], feed_dict={'events:0': [0.1, 0.2, 0.3]})\n<\/code><\/pre>\n\n<p>I get below error. The error is in the <code>tf.shape(events)<\/code>.<\/p>\n\n<pre><code>ValueError: initial_value must have a shape specified: Tensor(\"truncated_normal:0\", shape=(?,), dtype=float32)\n<\/code><\/pre>\n\n<p>The <code>events<\/code> variable is a vector whose length is equal to the batch size. The batch size is set to 2048. However, the last batch is usually less than 2048. So I have to set the shape to <code>[None]<\/code>.<\/p>\n\n<p>In this case, how should I copy the shaple of <code>events<\/code> to <code>opt_res<\/code>?<\/p>\n","answers":[{"AnswerId":"58020891","CreationDate":"2019-09-20T02:09:38.983","ParentId":null,"OwnerUserId":"2851685","Title":null,"Body":"<p>you don't need to create variable. you can do it like this<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nevents = tf.placeholder(tf.float32, shape=[None], name='events')\nopt_res = tf.truncated_normal(shape=tf.shape(events), mean=0.0, stddev=1e-1)\n\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n    init.run()\n    val = sess.run([opt_res], feed_dict={events: [0.1, 0.2, 0.3]})\n    print(val)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58020775,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-20T01:48:15.603","AcceptedAnswerId":null,"OwnerUserId":2106753.0,"Title":"Batch tf.data.Dataset.filter()?","Body":"<p>I know one can filter a dataset with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/data\/Dataset#filter\" rel=\"nofollow noreferrer\"><code>tf.data.Dataset.filter()<\/code><\/a>:<\/p>\n\n<pre><code>d = tf.data.Dataset.from_tensor_slices([1, 2, 3])\n\nd = d.filter(lambda x: x &lt; 3)  # ==&gt; [1, 2]\n\n# `tf.math.equal(x, y)` is required for equality comparison\ndef filter_fn(x):\n  return tf.math.equal(x, 1)\n\nd = d.filter(filter_fn)  # ==&gt; [1]\n<\/code><\/pre>\n\n<p>What if I want to do a \"batch filter\"? By this, I mean given a batch of strings <code>['str1', 'str2', 'str3', 'str4']<\/code>, how do I make a dataset that is able to return me a filtered dataset that spits out a batch of values that correspond to those strings: <code>[val1_respects_str1, val2_respects_str2, val3_respects_str3, val4_respects_str4]<\/code>?<\/p>\n","answers":[{"AnswerId":"58342400","CreationDate":"2019-10-11T13:32:49.300","ParentId":null,"OwnerUserId":"1259561","Title":null,"Body":"<p>What you want is not a filter but a map. The map function will compute the mapped values, such as:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>d = tf.data.Dataset.from_tensor_slices(list(range(100)))\n\ndef map_fn(x):\n  return x*2  # if you need arbitrary python logic, use tf.py_function to wrap it\n\nd = d.shuffle(100).batch(10).map(map_fn)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58021471,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2019-09-20T03:37:57.277","AcceptedAnswerId":null,"OwnerUserId":6486211.0,"Title":"How to write a prediction function for LSTM model","Body":"<p><a href=\"https:\/\/www.kaggle.com\/anunnikrishnan\/keras-classifier-based-spam-classification\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/anunnikrishnan\/keras-classifier-based-spam-classification<\/a><\/p>\n\n<p>Based on this notebook how would a prediction function look like that takes in a single sms message as test input?<\/p>\n\n<pre><code>sms = \" You won a free giftcard \"\n\nmax_words = 1000\nmax_len = 150\ntok = Tokenizer(num_words=max_words)\ntok.fit_on_texts(sms)\nsequences = tok.texts_to_sequences(text)\nsms_matrix = sequence.pad_sequences(sequences,maxlen=max_len)\nprediction = model.predict(sms_matrix)\n<\/code><\/pre>\n\n<p>the results look like this<\/p>\n\n<pre><code>array([[0.00395709],\n       [0.00090399],\n       [0.00528631],\n       [0.00549346],\n       [0.00347543],\n       [0.00090399],\n       [0.00751105],\n       [0.00549346],\n       [0.00736764],\n       [0.00549346],\n       [0.00568002],\n       [0.00696304],\n       [0.00551164],\n       [0.00551164],\n       [0.00549346],\n       [0.0014942 ],\n       [0.00137451],\n       [0.00568   ],\n       [0.01121344],\n       [0.01121931],\n       [0.00736762],\n       [0.00696306],\n       [0.00619166]], dtype=float32)\n<\/code><\/pre>\n\n<p>I'm having trouble how to interpret these results in a binary way.<\/p>\n","answers":[]}
+{"QuestionId":58021826,"AnswerCount":0,"Tags":"<python><scipy><pyinstaller><theano><pymc3>","CreationDate":"2019-09-20T04:35:17.083","AcceptedAnswerId":null,"OwnerUserId":5057891.0,"Title":"How to create a standalone python executable with pymc3 package?","Body":"<p>I have created a mathematical model in python. I want to make a standalone python executable for my model. The python script uses several python libraries like numpy, pandas, pymc3. I tried to make a python executable using pyinstaller and also auto-py-to-exe. In both cases, it failed to create the executable. The error occurs while importing the pymc3 module (theano library). My primary model is dependent on pymc3 module. I cannot bypass this module. Could anyone please help me to sort out this issue?<\/p>\n\n<p>Here is the error message I get when I execute the .exe file,<\/p>\n\n<pre><code>WARNING (theano.tensor.blas): Using NumPy C-API based implementation for BLAS functions.\nc:\\python\\lib\\site-packages\\PyInstaller\\loader\\pyimod03_importers.py:621: MatplotlibDeprecationWarning:\nThe MATPLOTLIBDATA environment variable was deprecated in Matplotlib 3.1 and will be removed in 3.3.\n  exec(bytecode, module.__dict__)\nTraceback (most recent call last):\n  File \"sample.py\", line 46, in &lt;module&gt;\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 983, in _find_and_load\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 967, in _find_and_load_unlocked\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 677, in _load_unlocked\n  File \"c:\\python\\lib\\site-packages\\PyInstaller\\loader\\pyimod03_importers.py\", line 621, in exec_module\n    exec(bytecode, module.__dict__)\n  File \"site-packages\\pymc3\\__init__.py\", line 11, in &lt;module&gt;\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 983, in _find_and_load\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 967, in _find_and_load_unlocked\n  File \"&lt;frozen importlib._bootstrap&gt;\", line 677, in _load_unlocked\n  File \"c:\\python\\lib\\site-packages\\PyInstaller\\loader\\pyimod03_importers.py\", line 621, in exec_module\n    exec(bytecode, module.__dict__)\n  File \"site-packages\\pymc3\\stats.py\", line 20, in &lt;module&gt;\n  File \"site-packages\\pkg_resources\\__init__.py\", line 481, in get_distribution\n  File \"site-packages\\pkg_resources\\__init__.py\", line 357, in get_provider\n  File \"site-packages\\pkg_resources\\__init__.py\", line 900, in require\n  File \"site-packages\\pkg_resources\\__init__.py\", line 786, in resolve\npkg_resources.DistributionNotFound: The 'scipy' distribution was not found and is required by the application\n[13796] Failed to execute script sample\n<\/code><\/pre>\n\n<p>Here is the python script,<\/p>\n\n<pre><code>class modelFit:\n    # Initialises the attributes\n    def __init__(self, x, y):\n        self.x = x\n        self.y = y\n\n    # Performs the Bayes estimation\n    def coeffEstimation(self):\n        X = np.array(self.x)\n        Y = np.array(self.y)\n\n        with pm.Model() as linearRegModel:\n            # Define priors\n            #regCoeff = pm.HalfNormal('regCoeff', sd = 20, shape=X.shape[1])\n            regCoeff = pm.Uniform('regCoeff', lower = 0, upper = 100, shape=X.shape[1])\n\n            # Standard deviation\n            sigma = pm.HalfNormal('sigma', sd = 5)\n\n            # Estimate of mean\n            mean = pm.math.dot(X, regCoeff)\n\n            # Likelihood estimate\n            likelihood = pm.Normal('Y_estimated', mu = mean, sd = sigma, observed = Y)\n\n            # Sampler\n            step = pm.NUTS()\n\n            # Posterior distribution\n            linearTrace = pm.sample(draws = 500, chains = 1,\n                                    tune = 500, nuts_kwargs=dict(target_accept=0.95))\n\n        with open('Summary.txt', 'w') as f:\n            print(pm.summary(linearTrace).round(3), file=f)\n            diverging = linearTrace['diverging']\n            print('Number of Divergent Chains: {}'.format(diverging.nonzero()[0].size), file=f)\n            divergingPer = diverging.nonzero()[0].size \/ len(linearTrace) * 100\n            print('Percentage of Divergent Chains: {:.1f}'.format(divergingPer), file=f)\n\n        print('Bayesian Inference Done! Exporting the results, Please hold on!')\n        return(linearTrace)\n\n\nimport theano.tensor.shared_randomstreams\nimport scipy\nimport pymc3 as pm\nimport numpy as np\nimport pandas as pd\ndataLocation = '.\/input\/'\ninputData = pd.read_excel(dataLocation + 'input.xlsx')\nsizeInputData = inputData.shape\nx = inputData.iloc[:,1:sizeInputData[1]]\ny = inputData.iloc[:,0]\n\n\n# Executes the Bayes method\nlinReg = modelFit(x, y)\nregTrace = linReg.coeffEstimation()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58021925,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-20T04:47:27.790","AcceptedAnswerId":null,"OwnerUserId":9839954.0,"Title":"tensorflow: how to use flags.DEFINE_multi_float()","Body":"<p>I use bash code to run a python file with lots of parameters. like:<\/p>\n\n<pre><code>python \"${WORK_DIR}\"\/eval.py \\\n  --logtostderr \\\n  --eval_split=\"val\" \\\n  --model_variant=\"xception_65\" \\\n  --atrous_rates=6 \\\n  --atrous_rates=12 \\\n  --atrous_rates=18 \\\n  --output_stride=16 \\\n  --decoder_output_stride=4 \\\n  --eval_crop_size=\"513,513\" \\\n  --checkpoint_dir=\"${TRAIN_LOGDIR}\" \\\n  --eval_logdir=\"${EVAL_LOGDIR}\" \\\n  --dataset_dir=\"${PASCAL_DATASET}\" \\\n  --max_number_of_evaluations=1 \\\n  --eval_scales=[0.5,0.25,1.75]\n<\/code><\/pre>\n\n<p>But then I got error:<\/p>\n\n<blockquote>\n  <p>absl.flags._exceptions.IllegalFlagValueError: flag\n  --eval_scales=[0.5,0.25,1.75]: could not convert string to float: '[0.5,0.25,1.75]'<\/p>\n<\/blockquote>\n\n<p><strong>So what is the right format<\/strong> to pass parameter to variable defined by <code>flags.DEFINE_multi_float()<\/code><\/p>\n\n<pre><code># Change to [0.5, 0.75, 1.0, 1.25, 1.5, 1.75] for multi-scale test.\nflags.DEFINE_multi_float('eval_scales', [1.0],\n                         'The scales to resize images for evaluation.')\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58022629,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-20T06:07:32.993","AcceptedAnswerId":58023734.0,"OwnerUserId":7140567.0,"Title":"How to divide slices of a keras 3D tensor with elements of a vector?","Body":"<p>I have to perform a tensor operation where each slice of the tensor is divided by the corresponding element from a vector. For example, a tensor K of shape (4,80,50) have 4 slices of shape(80,50) along the axes-0. Each of the 4 slices have to be divided by elements of the vector P of shape (4,1) ie. K[0,:,:] \/p[0] , K[1,:,:]\/p[1] etc. Is there a keras function which can perform this ?<\/p>\n","answers":[{"AnswerId":"58023734","CreationDate":"2019-09-20T07:28:53.717","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>Since <code>K<\/code> and <code>P<\/code> have the same size of 0-axis, you need to expand dimension of <code>P<\/code> at the end. In tf2.0, here is an example<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>K = tf.reshape(tf.range(24),(4,2,3))\n#array([[[ 0,  1,  2],\n#        [ 3,  4,  5]],\n#\n#       [[ 6,  7,  8],\n#        [ 9, 10, 11]],\n#\n#       [[12, 13, 14],\n#        [15, 16, 17]],\n#\n#       [[18, 19, 20],\n#        [21, 22, 23]]])\n\nP = tf.reshape(tf.constant([1,2,4,8]),(4,1))\n\nK \/ tf.expand_dims(P,axis=-1)\n#array([[[0.   , 1.   , 2.   ],\n#        [3.   , 4.   , 5.   ]],\n#\n#       [[3.   , 3.5  , 4.   ],\n#        [4.5  , 5.   , 5.5  ]],\n#\n#       [[3.   , 3.25 , 3.5  ],\n#        [3.75 , 4.   , 4.25 ]],\n#\n#       [[2.25 , 2.375, 2.5  ],\n#        [2.625, 2.75 , 2.875]]])\n<\/code><\/pre>\n"}]}
+{"QuestionId":58022713,"AnswerCount":2,"Tags":"<tensorflow><keras><tensorflow2.0>","CreationDate":"2019-09-20T06:14:36.580","AcceptedAnswerId":58980972.0,"OwnerUserId":6263317.0,"Title":"Tensorflow 2.0 Custom loss function with multiple inputs","Body":"<p>I am trying to optimize a model with the following two loss functions<\/p>\n\n<pre><code>def loss_1(pred, weights, logits):\n    weighted_sparse_ce = kls.SparseCategoricalCrossentropy(from_logits=True)\n    policy_loss = weighted_sparse_ce(pred, logits, sample_weight=advantages)\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>def loss_2(y_pred, y):\n    return kls.mean_squared_error(y_pred, y)\n<\/code><\/pre>\n\n<p>however, because TensorFlow 2 expects loss function to be of the form<\/p>\n\n<pre><code>def fn(y_pred, y_true):\n    ...\n<\/code><\/pre>\n\n<p>I am using a work-around for <code>loss_1<\/code> where I pack <code>pred<\/code> and <code>weights<\/code> into a single tensor before passing to <code>loss_1<\/code> in the call to <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/Model#fit\" rel=\"nofollow noreferrer\"><code>model.fit<\/code><\/a> and then unpack them in <code>loss_1<\/code>. This is inelegant and nasty because <code>pred<\/code> and <code>weights<\/code> are of different data types and so this requires an additional cast, pack, un-pack and un-cast each time I call <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/Model#fit\" rel=\"nofollow noreferrer\"><code>model.fit<\/code><\/a>.<\/p>\n\n<p>Furthermore, I am aware of the <code>sample_weight<\/code> argument to <code>fit<\/code>, which is kind of like the solution to <a href=\"https:\/\/stackoverflow.com\/questions\/57414114\/implement-custom-loss-function-in-tensorflow-2-0\">this question<\/a>. This might be a workable solution were it not for the fact that I am using two loss functions and I only want the <code>sample_weight<\/code> applied to one of them. Also, even if this were a solution, would it not be generalizable to other types of custom loss functions.<\/p>\n\n<hr>\n\n<p>All that being said, my question, said concisely, is:<\/p>\n\n<p>What is the best way to create a loss function with an arbitrary number of\narguments in TensorFlow 2?<\/p>\n\n<p>Another thing I have tried is passing a <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/tuple\" rel=\"nofollow noreferrer\"><code>tf.tuple<\/code><\/a> but that also seems to violate TensorFlow's desires for a loss function input.<\/p>\n","answers":[{"AnswerId":"58033185","CreationDate":"2019-09-20T17:59:02.813","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>In tf 1.x we have <code>tf.nn.weighted_cross_entropy_with_logits<\/code> function which allows us trade off recall and precision by adding extra positive weights for each class. In multi-label classification, it should be a (N,) tensor or numpy array. However, in tf 2.0, I haven't found similar loss functions yet, so I wrote my own loss function with extra arguments <code>pos_w_arr<\/code>. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.keras.backend import epsilon\n\ndef pos_w_loss(pos_w_arr):\n    \"\"\"\n    Define positive weighted loss function\n    \"\"\"\n    def fn(y_true, y_pred):\n        _epsilon = tf.convert_to_tensor(epsilon(), dtype=y_pred.dtype.base_dtype)\n        _y_pred = tf.clip_by_value(y_pred, _epsilon, 1. - _epsilon)\n        cost = tf.multiply(tf.multiply(y_true, tf.math.log(\n            _y_pred)), pos_w_arr)+tf.multiply((1-y_true), tf.math.log(1-_y_pred))\n        return -tf.reduce_mean(cost)\n    return fn\n<\/code><\/pre>\n\n<p>Not sure what do you mean it wouldn't work when using eager tensors or numpy array as inputs though. Please correct me if I'm wrong. <\/p>\n"},{"AnswerId":"58980972","CreationDate":"2019-11-21T17:54:29.293","ParentId":null,"OwnerUserId":"6263317","Title":null,"Body":"<p>This problem can be easily solved using <a href=\"https:\/\/www.tensorflow.org\/tutorials\/customization\/custom_training_walkthrough\" rel=\"nofollow noreferrer\">custom training<\/a> in TF2. You need only compute your two-component loss function within a <code>GradientTape<\/code> context and then call an optimizer with the produced gradients. For example, you could create a function <code>custom_loss<\/code> which computes both losses given the arguments to each:<\/p>\n\n<pre><code>def custom_loss(model, loss1_args, loss2_args):\n  # model: tf.model.Keras\n  # loss1_args: arguments to loss_1, as tuple.\n  # loss2_args: arguments to loss_2, as tuple.\n  with tf.GradientTape() as tape:\n    l1_value = loss_1(*loss1_args)\n    l2_value = loss_2(*loss2_args)\n    loss_value = [l1_value, l2_value]\n  return loss_value, tape.gradient(loss_value, model.trainable_variables)\n\n# In training loop:\nloss_values, grads = custom_loss(model, loss1_args, loss2_args)\noptimizer.apply_gradients(zip(grads, model.trainable_variables))\n<\/code><\/pre>\n\n<p>In this way, each loss function can take an arbitrary number of eager tensors, regardless of whether they are inputs or outputs to the model. The sets of arguments to each loss function need not be disjoint as shown in this example.<\/p>\n"}]}
+{"QuestionId":58022817,"AnswerCount":0,"Tags":"<gpu><pytorch><nvidia>","CreationDate":"2019-09-20T06:25:20.277","AcceptedAnswerId":null,"OwnerUserId":4777141.0,"Title":"Pytorch Cuda Runtime Error (with Properly installed Drivers)","Body":"<p>When I run<\/p>\n\n<pre><code>import torch\ntorch.cuda.FloatTensor([1.])\n<\/code><\/pre>\n\n<p>I seem to be getting the error: <code>RuntimeError: cuda runtime error (38) : no CUDA-capable device is detected at \/pytorch\/aten\/src\/THC\/THCGeneral.cpp:50<\/code>.<\/p>\n\n<p>I'm using Debian (Stretch). Checking the nvidia driveron the console:<\/p>\n\n<pre><code>\/sbin\/modinfo nvidia_current\n<\/code><\/pre>\n\n<p>returns:<\/p>\n\n<pre><code>filename:       \/lib\/modules\/4.9.0-11-amd64\/updates\/dkms\/nvidia-current.ko\nalias:          char-major-195-*\nversion:        418.74\nsupported:      external\nlicense:        NVIDIA\nsrcversion:     AB4044DE27C9CA55579A110\n<\/code><\/pre>\n\n<p>when I do <code>nvcc --version<\/code>, I get:<\/p>\n\n<pre><code>nvcc: NVIDIA (R) Cuda compiler driver\nCopyright (c) 2005-2018 NVIDIA Corporation\nBuilt on Tue_Jun_12_23:07:04_CDT_2018\nCuda compilation tools, release 9.2, V9.2.148\n<\/code><\/pre>\n\n<p>On python, when I do<\/p>\n\n<pre><code>import torch\nprint(torch.version.cuda)  \nprint(torch.cuda.device_count())\nprint(torch.cuda.is_available())\n<\/code><\/pre>\n\n<p>I get:<\/p>\n\n<pre><code>9.2.148\n0\nFalse\n<\/code><\/pre>\n\n<p>I've tried doing <code>os.environ[\"CUDA_VISIBLE_DEVICES\"] = '0<\/code>, but it still doesn't work. <\/p>\n\n<p>Any help would be much appreciated!<\/p>\n","answers":[]}
+{"QuestionId":58022846,"AnswerCount":2,"Tags":"<arrays><python-3.x><numpy><pytorch><torch>","CreationDate":"2019-09-20T06:27:21.253","AcceptedAnswerId":null,"OwnerUserId":6894995.0,"Title":"numpy argmax in array with multiple brackets","Body":"<p>I have an issue in apply argmax to an array which has multiple brackets.\nIn real life I am getting this as a result of a pytorch tensor.\nHere I can put an example:<\/p>\n\n<pre><code>a = np.array([[1.0, 1.1],[2.1,2.0]])\nnp.argmax(a,axis=1)\n\narray([1, 0])\n<\/code><\/pre>\n\n<p>It is correct. But:<\/p>\n\n<pre><code>a = np.array([[[1.0, 1.1]],[[2.1,2.0]]])\nnp.argmax(a,axis=1)\n\narray([[0, 0],\n       [0, 0]])\n<\/code><\/pre>\n\n<p>It does not give me what I expect.\nConsider that in reality I have this level of inner brackets:<\/p>\n\n<pre><code>a = np.array([[[[1.0, 1.1]]],[[[2.1,2.0]]]])\n<\/code><\/pre>\n","answers":[{"AnswerId":"58023039","CreationDate":"2019-09-20T06:41:46.727","ParentId":null,"OwnerUserId":"4427777","Title":null,"Body":"<p>Use <code>.squeeze()<\/code> and a negative index.<\/p>\n\n<pre><code>a = np.array([[[[1.0, 1.1]]], [[[2.1, 2.0]]]])\nnp.argmax(a, axis = -1).squeeze()\n\narray([1, 0], dtype=int32)\n<\/code><\/pre>\n"},{"AnswerId":"58022847","CreationDate":"2019-09-20T06:27:21.253","ParentId":null,"OwnerUserId":"6894995","Title":null,"Body":"<p>A possible solution is to increment axis value:<\/p>\n\n<pre><code>a = np.array([[[[1.0, 1.1]]],[[[2.1,2.0]]]])\nnp.argmax(a,axis=3)\n\narray([[[1]],\n       [[0]]])\n<\/code><\/pre>\n\n<p>But I still have inner brackets.<\/p>\n"}]}
+{"QuestionId":58023424,"AnswerCount":0,"Tags":"<python><deep-learning><artificial-intelligence><caffe>","CreationDate":"2019-09-20T07:10:30.250","AcceptedAnswerId":null,"OwnerUserId":11976900.0,"Title":"How can I set Caffe output feature to int?","Body":"<p>I'm doing inference with caffe<\/p>\n\n<p>Some of the results after going through the conv1 layer.<\/p>\n\n<pre><code>'''\n [[ 0.         0.         0.        ...  0.         0.         0.       ]\n  [ 2.4921875  8.1796875  8.125     ...  8.15625    8.15625    8.15625  ]\n  [ 2.4921875  8.15625    8.125     ...  8.15625    8.15625    8.15625  ]\n  ...\n  [ 7.96875   14.4296875  0.        ...  0.        12.953125   0.6171875]\n  [ 0.         0.         1.9140625 ... 19.6875     0.         0.       ]\n  [ 1.2109375  0.         0.        ...  0.         0.        19.71875  ]]\n...\n<\/code><\/pre>\n\n<p>I want to use these as inputs to the next layer without the decimal point.<\/p>\n\n<p>like this <\/p>\n\n<pre><code>'''\n [[ 0  0  0     ...  0    0    0  ]\n  [ 2  8  8     ...  8    8    8  ]\n  [ 2  8  8     ...  8    8    8  ]\n  ...\n  [ 7   14  0   ...  0    12   0  ]\n  [ 0   0   1   ... 19    0    0  ]\n  [ 1   0   0   ...  0    0   19  ]]\n...\n<\/code><\/pre>\n\n<p>Is this possible?<\/p>\n","answers":[]}
+{"QuestionId":58023585,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><conv-neural-network>","CreationDate":"2019-09-20T07:20:13.500","AcceptedAnswerId":null,"OwnerUserId":7849245.0,"Title":"TensorFlow : Fine tuning only the Fully Connected Layer for Different learning rates with a single python file","Body":"<p>I am doing some experiments with transfer learning :<br>\nI have a script file where in 1st Part i train the model on a subset of mnist data set and then successfully save it . \nThe model architecture consists of 2 C.N.N layers and 1 fully connected layer.<\/p>\n\n<pre><code>for epoch in range(1, params['epochs'] + 1):\n                 shuffle = np.random.permutation(len(y_train))\n                 x_train, y_train = x_train[shuffle], y_train[shuffle]\n\n\n                 for i in range(0, len(y_train), params['batch_size']):\n                     x_train_mb, y_train_mb = x_train[i:i + params['batch_size']], y_train[i:i + params['batch_size']]\n\n                     sess.run(model.optimize, feed_dict={model.input: x_train_mb, model.target: y_train_mb, model.is_task1: True,\n model.is_train: True, model.learning_rate:\n temp_learning_rate_source_training})\n\n\n\n\n                 valid_acc = classification_batch_evaluation(sess, model, model.metrics, params['batch_size'], True, x_valid, y=y_valid,\n stream=True)\n\n                 print('valid [{} \/ {}] valid accuracy: {} learning Rate :{}'.format(epoch, params['epochs'] + 1,\n valid_acc,temp_learning_rate_source_training))\n                 if valid_acc &gt; initial_best_epoch['valid_acc']:\n                     initial_best_epoch['epoch'] = epoch\n                     initial_best_epoch['valid_acc'] = valid_acc\n                     model.save_model(sess, epoch) \n\n                 if epoch - initial_best_epoch['epoch'] &gt;= params['patience']:\n                     print('Early Stopping Epoch: {}\\n'.format(epoch))\n                     logging.info('Early Stopping Epoch: {}\\n'.format(epoch))\n                     break\n\n\n         print('Initial training done \\n',file=f)\n         logging.info('Initial training done \\n')\n         sess.close()      \n\n     model.restore_model(sess) ##Restores the model after creating it .\n<\/code><\/pre>\n\n<p>Now i  want to do transfer learning by keeping the architecture same and  transferring the parameters for the C.N.N layers and re-initializing the fully connected layer . And then training the  3 layers again for the limited new data set  by using different learning rates and \"decay_after_epoch\" to analyze the result . Now because of high number of combinations i have written 2 for - loops to automate the process as follows :     <\/p>\n\n<pre><code>for temp_learning_rate_target_training in (0.001,0.005,0.01):\n\n        for decay_after_epoch in (3,5,10): \n            learning_rate = temp_learning_rate_target_training\n            model.restore_model(sess) ##Restores the model after creating it .\n            with open(\"\/home\/abhishek\/Desktop\/{}_{}_{}.txt\".format(params[\"dataset\"],params[\"k\"],params[\"n\"])) as f1:\n                with open(\"\/home\/abhishek\/Desktop\/{}_{}_{}_{}_{}.txt\".format(params[\"dataset\"],params[\"k\"],params[\"n\"],temp_learning_rate_target_training,decay_after_epoch), \"w\") as f:\n                    for x in f1.readlines():\n                        f.write(x)\n                    print(\"Target Training Begins\",file=f)\n                    for epoch in range(1, params['epochs'] + 1):\n                        shuffle = np.random.permutation(len(y_train2))\n                        x_train2, y_train2 = x_train2[shuffle], y_train2[shuffle]\n\n\n                        if epoch%decay_after_epoch==0 and epoch &lt;=decay_after_epoch:\n                            learning_rate = learning_rate *0.1\n                        elif (epoch-decay_after_epoch)%30==0:\n                            learning_rate = learning_rate *0.1\n\n\n\n                        for i in range(0, len(y_train2), params['batch_size']):\n                            x_train_mb, y_train_mb = x_train2[i:i + params['batch_size']], y_train2[i:i + params['batch_size']]\n                            sess.run(model.optimize, feed_dict={model.input: x_train_mb, model.target: y_train_mb, model.is_task1: False, model.is_train: True, model.learning_rate: params['learning_rate']})\n\n                        train_acc = classification_batch_evaluation(sess, model, model.metrics, params['batch_size'], False, x_train2, y=y_train2, stream=True)\n                        sess.close()\n\n                        print('train [{} \/ {}] train accuracy: {} learning Rate:{} '.format(epoch, params['epochs'] + 1, train_acc,learning_rate),file=f)\n                        print('train [{} \/ {}] train accuracy: {} learning Rate :{}'.format(epoch, params['epochs'] + 1, train_acc,learning_rate))\n                        logging.info('train [{} \/ {}] train accuracy: {}'.format(epoch, params['epochs'] + 1, train_acc))\n\n                        if train_acc &gt; transfer_best_epoch['train_acc']:\n                            transfer_best_epoch['epoch'] = epoch\n                            transfer_best_epoch['train_acc'] = train_acc\n                            test_acc = classification_batch_evaluation(sess, model, model.metrics, params['batch_size'], False, x_test2, y=y_test2, stream=True)\n                            transfer_best_epoch['test_acc'] = test_acc\n\n                        if epoch % params['patience'] == 0:\n                            acc_diff = transfer_best_epoch['train_acc'] - es_acc\n                            if acc_diff &lt; params['percentage_es'] * es_acc:\n                                print('Early Stopping Epoch: {}\\n'.format(epoch))\n                                logging.info('Early Stopping Epoch: {}\\n'.format(epoch))\n                                break\n                            es_acc = transfer_best_epoch['train_acc']\n\n                    print('Transfer training done \\n',file=f)\n                    print('TARGET test accuracy: {}'.format(transfer_best_epoch['test_acc']),file=f)\n<\/code><\/pre>\n\n<p>Now after the first loop is  run with temp_learning_rate_target_training =0.0001 and decay_after_epoch = 3 the model is trained , and i have the test accuracy , and let the weights and biases for different(3) layers are given by a set S2 .\nNow when the loop is run again , the parameter  model.is_task1: False makes sure that the fully connected layer is re-initialized  , but the parameters   of the C.N.N layers are copied over from the set S2 . (Why i am saying this is because i am getting exact same accuracy logs for all the combination of learning Rates and decay_after_epoch ).  However , i want to train different loops with same initial parameters for the C.N.N layers which are given by S1<br>\nI have tried closing the session with  <code>sess.close()<\/code> , after each loop and then restoring the saved model(which was trained in part1 of the code) with  <code>model.restore_model(sess)<\/code> but it still does not give the expected result.  How should i be proceeding ?<\/p>\n","answers":[{"AnswerId":"58024838","CreationDate":"2019-09-20T08:39:38.163","ParentId":null,"OwnerUserId":"5652313","Title":null,"Body":"<p>If you need to tweak your models like this, you need to be aware of what variables you want to train.<\/p>\n\n<p>All layers should have their variables in their variable scope. You can get the variables:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='my_scope')\n<\/code><\/pre>\n\n<p>For variables you want to re-initialize, you can do just:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>sess.run([v.initializer for v in variables_to_reset])\n<\/code><\/pre>\n\n<p>When you initializer the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Optimizer#minimize\" rel=\"nofollow noreferrer\">minimize method<\/a> (it is hidden in your <code>model<\/code> object, this what gets you the <code>model.optimize<\/code> op that you call in your training loop), you can specify <code>var_list<\/code> which is the list of variables you want to train, the rest will remain intact.<\/p>\n"}]}
+{"QuestionId":58023677,"AnswerCount":0,"Tags":"<python><keras><stride>","CreationDate":"2019-09-20T07:25:34.443","AcceptedAnswerId":null,"OwnerUserId":12012993.0,"Title":"i want to control stride in convolutional neural network but i dont know how to write stride","Body":"<p>I am beginner of medical deep learning. I am practicing medical deep learning. I don't know how to control stride in ConVNN. I want to know if I control stride in ConVNN,that I think image is different. so I want to see that.<\/p>\n\n<p>already I saw many information about stride in ConVNN in Keras. but I didn't understand.<\/p>\n\n<pre><code>def build_SRCNN():\n    input_img = Input(shape=(512, 512, 1), name='image_input')\n\n\n    x = Conv2D(64, (9, 9), strides=(1,1), activation='relu',kernel_initializer='he_normal', padding='same')(input_img)\n    x = Conv2D(32, (3, 3), strides=(1,1), activation='relu', kernel_initializer='he_normal', padding='same')(x)\n    x = Conv2D(1, (5, 5), strides=(1,1), activation='relu',kernel_initializer='he_normal', padding='same')(x)\n\n    # model\n    SRCNN = Model(inputs=input_img, outputs=x)\n\n    return SRCNN\n<\/code><\/pre>\n\n<p>I wrote <code>stride = (1,1)<\/code> to <code>(3,3)<\/code> or <code>(5,5)<\/code> or <code>(7,7)<\/code> ... but error is same.<\/p>\n","answers":[]}
+{"QuestionId":58023853,"AnswerCount":0,"Tags":"<tensorflow><keras><precision-recall>","CreationDate":"2019-09-20T07:36:09.920","AcceptedAnswerId":null,"OwnerUserId":2080475.0,"Title":"Use thesholds parameter in tf.keras.metrics.Recall","Body":"<p>I would like to know how recall gets calculated when multiple thresholds are specified. \nBelow is the snippet of description from <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/metrics\/Recall\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/metrics\/Recall<\/a><\/p>\n\n<blockquote>\n  <p>thresholds: (Optional) A float value or a python list\/tuple of float threshold values in [0, 1]. A threshold is compared with prediction values to determine the truth value of predictions (i.e., above the threshold is true, below is false). One metric value is generated for each threshold value. If neither thresholds nor top_k are set, the default is to calculate recall with thresholds=0.5.<\/p>\n<\/blockquote>\n\n<p>I am trying to passing a list of 3 threshold values , and as per the description I expected 3 recall values would be generated (i.e one recall per threshold) , but it isnt working in this way, only 1 recall metric is generated.<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import Embedding,Flatten,Dense\nfrom tensorflow.keras.layers import LSTM\nfrom tensorflow.keras.layers import Bidirectional\nfrom tensorflow.keras.layers import Dropout\nfrom tensorflow.keras import layers\nmodel=Sequential()\nmodel.add(Embedding(len(tokens)+1,embedding_dim,input_length=MAX_TEXT_LEN,weights=[embedding_matrix]))\nmodel.add(LSTM(128))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(9,activation='sigmoid'))\nopt=tf.keras.optimizers.Adam(lr=0.0001)\nmodel.compile(loss='binary_crossentropy', optimizer=opt, metrics=[tf.keras.metrics.Recall(thresholds=[0.2,0.4,0.8]))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58024109,"AnswerCount":0,"Tags":"<python><tensorflow><keras><neural-network><deep-learning>","CreationDate":"2019-09-20T07:51:50.440","AcceptedAnswerId":null,"OwnerUserId":5108626.0,"Title":"Keras - pass inputs through layers only if condition passes","Body":"<p>I'm trying to develop a model which picks the best candidate out of a list of candidates:<\/p>\n\n<ul>\n<li>Input: A list of <code>K<\/code> candidates<\/li>\n<li>Target output: A one-hot vector <code>x<\/code> with <code>x[t] = p_t<\/code> where <code>p_t<\/code> is the probability of <code>t<\/code> being the best candidate<\/li>\n<\/ul>\n\n<p>This is how I'm trying to do it in Keras:<\/p>\n\n<ol>\n<li>For a given training row, represent each candidate <code>i<\/code> as a feature vector <code>f_i<\/code> (all features are positive). The row may have less than <code>K<\/code> candidates; in that case, I mask\n\/ pad up the list of candidates with dummy <code>f_i<\/code> vectors (of all\n-1's)<\/li>\n<li>Pass each of the actual <code>f_i<\/code> vectors through a set of layers to get some score <code>s_i<\/code>; the dummy feature vectors shouldn't be passed\nthrough these layers and should directly get a score of 0<\/li>\n<li>Do\nsoftmax over all the non-zero <code>s_i<\/code> and get a final probability<\/li>\n<\/ol>\n\n<p>Step 1 is trivial, and step 3 can be done with a custom activation as shown <a href=\"https:\/\/stackoverflow.com\/questions\/56078515\/keras-masking-zero-before-softmax\">here<\/a>. Step 2 is what I don't know how to do; I tried using a <code>Lambda<\/code> layer from Keras with <code>K.switch<\/code> like:<\/p>\n\n<pre><code>score_layer = Lambda(lambda x: K.switch(K.sum(x) &lt; 0, 0.0, score_using_layers(x))\n<\/code><\/pre>\n\n<p>but this didn't work. When I tried to compile the model and print the summary, it wouldn't even show the layers being used to score the real feature vectors. How do I do this?<\/p>\n","answers":[]}
+{"QuestionId":58024561,"AnswerCount":0,"Tags":"<lua><torch>","CreationDate":"2019-09-20T08:21:42.047","AcceptedAnswerId":null,"OwnerUserId":11081280.0,"Title":"Segmentation fault thrown when importing cudnn","Body":"<p>I'm new to Lua and Torch7 but after successfully (it seems) installing the environment I keep getting segmentation error on line: <\/p>\n\n<pre><code>require 'cudnn'\n<\/code><\/pre>\n\n<p>I should also note that when I do<\/p>\n\n<pre><code>require 'cunn' or require 'torch'\n<\/code><\/pre>\n\n<p>I get no errors...<\/p>\n\n<p>My system is \nCUDA 10.0\nGPU: Tesla K80<\/p>\n","answers":[]}
+{"QuestionId":58024781,"AnswerCount":1,"Tags":"<tensorflow><tensorflow.js>","CreationDate":"2019-09-20T08:36:21.800","AcceptedAnswerId":58032402.0,"OwnerUserId":11335442.0,"Title":"Error: Tensorflow Op is not supported: AddN","Body":"<p>I've retrained a Mobilenet V1 model using my own dataset. Right now, I'm trying to get my model to load into this example project:\n<a href=\"https:\/\/github.com\/shivangidas\/image-classifier\" rel=\"nofollow noreferrer\">https:\/\/github.com\/shivangidas\/image-classifier<\/a><\/p>\n\n<p>It keeps throwing the following error <\/p>\n\n<pre><code> Error: Tensorflow Op is not supported: AddN\n<\/code><\/pre>\n\n<p>I don't understand what this means or why this error is being shown. I also can't seem to find anything about it online. <\/p>\n\n<p>This is the code I'm using:<\/p>\n\n<pre><code>const MODEL_URL =\n  \"https:\/\/localhost\/path\/to\/model\";\nconst WEIGHTS_URL =\n  \"https:\/\/localhost\/path\/to\/weights_manifest\";\nlet model;\nlet IMAGENET_CLASSES = [];\nlet offset = tf.scalar(128);\nasync function loadModelAndClasses() {\n  $.getJSON(\n   \"https:\/\/localhost\/path\/to\/labels.json\",\n    function(data) {\n      $.each(data, function(key, val) {\n        IMAGENET_CLASSES.push(val);\n      });\n    }\n  );\n  model = await tf.loadFrozenModel(MODEL_URL, WEIGHTS_URL);\n  \/\/console.log(\"After model is loaded: \" + tf.memory().numTensors);\n  $(\".loadingDiv\").hide();\n  $(\"#inputImage\").attr(\"disabled\", false);\n}\nloadModelAndClasses();\nfunction readURL(input) {\n  if (input.files &amp;&amp; input.files[0]) {\n    var reader = new FileReader();\n\nreader.onload = function(e) {\n  $(\"#imageSrc\")\n    .attr(\"src\", e.target.result)\n    .width(224)\n    .height(224);\n};\n\nreader.readAsDataURL(input.files[0]);\n\/\/console.log(\"After image is loaded: \" + tf.memory().numTensors);\n\nreader.onloadend = async function() {\n  console.log(\"Before predictions: \" + tf.memory().numTensors);\n\n  let imageData = document.getElementById(\"imageSrc\");\n\n  \/\/console.log(\"After offset: \" + tf.memory().numTensors);\n  let pixels1 = tf.fromPixels(imageData);\n  let pixel2 = pixels1.resizeNearestNeighbor([224, 224]);\n  let pixel3 = pixel2.toFloat();\n  console.log(\"After pixels are formed: \" + tf.memory().numTensors);\n\n  let pixels = pixel3.sub(offset);\n  let pixels4 = pixels.div(offset);\n  let pixels5 = pixels4.expandDims();\n  console.log(\"After pre-processing: \" + tf.memory().numTensors);\n\n  const output = await model.predict(pixels5);\n  console.log(\"After output: \" + tf.memory().numTensors);\n  const predictions = Array.from(output.dataSync())\n    .map(function(p, i) {\n      return {\n        probabilty: p,\n        classname: IMAGENET_CLASSES[i]\n      };\n    })\n    .sort((a, b) =&gt; b.probabilty - a.probabilty)\n    .slice(0, 10);\n\n  \/\/console.log(predictions);\n  var html = \"\";\n  for (let i = 0; i &lt; 10; i++) {\n    html += \"&lt;li&gt;\" + predictions[i].classname + \"&lt;\/li&gt;\";\n  }\n  $(\".predictionList\").html(html);\n  console.log(\"After predictions: \" + tf.memory().numTensors);\n\n  pixels.dispose();\n  pixels1.dispose();\n  pixel2.dispose();\n  pixel3.dispose();\n  pixels4.dispose();\n  pixels5.dispose();\n  output.dispose();\n  console.log(\"After dispose: \" + tf.memory().numTensors);\n     };\n    }\n   }\n<\/code><\/pre>\n","answers":[{"AnswerId":"58032402","CreationDate":"2019-09-20T16:51:05.533","ParentId":null,"OwnerUserId":"5627599","Title":null,"Body":"<p>The <code>addN<\/code> operation is supported by the converter according to <a href=\"https:\/\/github.com\/tensorflow\/tfjs\/blob\/master\/tfjs-converter\/docs\/supported_ops.md\" rel=\"nofollow noreferrer\">this list<\/a>. It looks like you are using a rather old version of Tensorflow.js. I noticed the function <code>loadFrozenModel<\/code> which has been renamed to <code>loadGraphModel<\/code> since version 1.0 (released in March 2019).<\/p>\n\n<p>Converting <code>addN<\/code> is supported <a href=\"https:\/\/github.com\/tensorflow\/tfjs-converter\/blob\/v0.5.6\/docs\/supported_ops.md\" rel=\"nofollow noreferrer\">since version 0.5.6<\/a> (see <a href=\"https:\/\/github.com\/tensorflow\/tfjs-converter\/commit\/d7427eb84df7bc05c9254010719d7fd416ef8b4c#diff-86b322812e9000253eed437a8b196c3f\" rel=\"nofollow noreferrer\">this commit<\/a>). If your Tensorflow.js version is older than <code>0.5.6<\/code> you can simply upgrade to a more recent version and it should work.<\/p>\n"}]}
+{"QuestionId":58024879,"AnswerCount":1,"Tags":"<python><tensorflow><keras><conv-neural-network>","CreationDate":"2019-09-20T08:42:29.870","AcceptedAnswerId":null,"OwnerUserId":11088375.0,"Title":"Should I use only \"exactly same\" input shape for transfer learning?","Body":"<p>I trained a CNN model with (5x128x128x3) size of input shape\nand I got trained weight of (5x128x128x3)<\/p>\n\n<p>by the way, I wanna use this weight for training (7x128x128x3) size of input data<\/p>\n\n<p>So, this is my question\nshould I use only same shape of input?\nI wonder if I can use another size (in this case, 7x128x128x3) of input for transfer learning<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>ValueError: Error when checking input: expected input_1 to have shape (5, 128, 128, 3) but got arry with shape (7, 128, 128, 3)```\n<\/code><\/pre>\n","answers":[{"AnswerId":"58025265","CreationDate":"2019-09-20T09:08:37.067","ParentId":null,"OwnerUserId":"2650622","Title":null,"Body":"<p>Let's break down the dimensions <strong>(5x128x128x3)<\/strong>:<\/p>\n\n<p>The <strong>first<\/strong> dimension is the batch size (which was 5 when the original model was trained). This is irrelevant and you can set it to <code>None<\/code> as pointed out in the comments to feed arbitrary sized batches to the model.<\/p>\n\n<p>The <strong>second to third<\/strong> dimensions (128x128) are the width and height of the image and you may be able to change these, but it's hard to say for sure without knowing the model architecture and which layer output you're using for transfer learning. The reason you can change these is that 2d convolutional filters are repeated across the 2d dimensions (width and height) of the image, so they will remain valid for different widths and heights (assuming compatible padding). But if you change the 2d dimensions too much, it is possible that the receptive fields of the layers are changed in a way that hurts transfer learning performance. Eg. if the 7th conv layer in the network for 128x128 input can see the entire input image in each activation (a receptive field of 128x128), then if you double the width and height, it won't anymore and the layer may not recognize certain global features.<\/p>\n\n<p>The <strong>fourth<\/strong> dimension is the number of channels in the input images and you can't change this, as the filters in the first layer will have 3 weights across the depth dimension.<\/p>\n"}]}
+{"QuestionId":58024931,"AnswerCount":2,"Tags":"<tensorflow><tensorflow-datasets>","CreationDate":"2019-09-20T08:45:47.963","AcceptedAnswerId":null,"OwnerUserId":11998352.0,"Title":"How to make tf.data.Dataset only shuffle a subset of data column?","Body":"<p>Let's see we have a dataset:<\/p>\n\n<pre><code>   ds = tf.data.Dataset.from_tensor_slices((x_train,y_train ))\n<\/code><\/pre>\n\n<p>if I apply ds = ds.shuffle(), then both x_train and y_train will be shuffled. How to do if I only want shuffle x_train while keeping y_train unchanged?<\/p>\n","answers":[{"AnswerId":"58025484","CreationDate":"2019-09-20T09:21:47.033","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>Not sure why you want to do this since wrong labels will be mapped to the features. \nBut still if you want to do this anyways, you can shuffle <code>x_train<\/code>, before feeding it to <code>from_tensor_slices()<\/code>.<\/p>\n"},{"AnswerId":"58026304","CreationDate":"2019-09-20T10:11:53.330","ParentId":null,"OwnerUserId":"11478478","Title":null,"Body":"<p>You can achieve what u want with a map function that returns only 1 of the outputs. And then you can zip them.<\/p>\n\n<pre><code>ds = tf.data.Dataset.from_tensor_slices((x_train, y_train))\nds_x = ds.map(lambda x, y: x)\nds_y = ds.map(lambda x, y: y)\nds_x = ds_x.shuffle(buffer_size=50) #change buffer_size as u like\nds = tf.data.Dataset.zip((ds_x, ds_y))\n<\/code><\/pre>\n"}]}
+{"QuestionId":58025069,"AnswerCount":2,"Tags":"<python><tensorflow><tensorflow2.0>","CreationDate":"2019-09-20T08:56:05.957","AcceptedAnswerId":null,"OwnerUserId":12094550.0,"Title":"How to enable cuda unified memory in tensorflow v2","Body":"<p>In <code>tensorflow<\/code> 1.x, there is an option like <code>use_unified_memory<\/code> and <code>per_process_gpu_memory_fraction<\/code> which is potential to trigger CUDA UVM used. But how can this be done in <code>tensorflow<\/code> 2.0?<\/p>\n\n<p><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\n<pre><code>\/\/ If true, uses CUDA unified memory for memory allocations. If\n\/\/ per_process_gpu_memory_fraction option is greater than 1.0, then unified\n\/\/ memory is used regardless of the value for this field. See comments for\n\/\/ per_process_gpu_memory_fraction field for more details and requirements\n\/\/ of the unified memory. This option is useful to oversubscribe memory if\n\/\/ multiple processes are sharing a single GPU while individually using less\n\/\/ than 1.0 per process memory fraction.\nbool use_unified_memory = 2;\n<\/code><\/pre>\n","answers":[{"AnswerId":"58965019","CreationDate":"2019-11-20T23:33:45.197","ParentId":null,"OwnerUserId":"3723199","Title":null,"Body":"<p>If someone's looking to enable UVM in 1.x, just set the <code>per_process_gpu_memory_fraction<\/code> over 1 (to any number you want).<\/p>\n\n<p><code>use_unified_memory<\/code> doesn't do anything.<\/p>\n\n<p>Another potential bug of TensorFlow: you may want to move your model definition to after you establish the session. Like<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with tf.Session(GPUOptions...) as s:\n    model = xxx\n    s.run(model)\n<\/code><\/pre>\n"},{"AnswerId":"58071462","CreationDate":"2019-09-23T23:57:40.123","ParentId":null,"OwnerUserId":"11530462","Title":null,"Body":"<p>For <strong>TensorFlow 2.0<\/strong> you may use <code>tf.config.experimental<\/code> api.<br>\nSimilarly to TF 1.X there are two methods to limit gpu usage as listed below:  <\/p>\n\n<p><strong>(1) Allow GPU memory growth<\/strong><br>\nThe first option is to turn on memory growth by calling <code>tf.config.experimental.set_memory_growth<\/code><br>\nFor instance;<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>gpus = tf.config.experimental.list_physical_devices('GPU')\ntf.config.experimental.set_memory_growth(gpus[0], True)\n<\/code><\/pre>\n\n<p>See <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/config\/experimental\/set_memory_growth\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/config\/experimental\/set_memory_growth<\/a><\/p>\n\n<p><strong>(2) Set hard limit to GPU memory growth<\/strong><br>\nThe second method is to configure a virtual GPU device with <code>tf.config.experimental.set_virtual_device_configuration<\/code> and set a hard limit on the total memory to allocate on the GPU.<br>\nFor instance;<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>gpus = tf.config.experimental.list_physical_devices('GPU')\ntf.config.experimental.set_virtual_device_configuration(\n          gpus[0],\n            [tf.config.experimental.VirtualDeviceConfiguration(memory_limit=1024)])\n<\/code><\/pre>\n\n<p>See following link to know more <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/config\/experimental\/set_virtual_device_configuration\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/config\/experimental\/set_virtual_device_configuration<\/a><\/p>\n\n<p>Hope this helps.<\/p>\n"}]}
+{"QuestionId":58025115,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm><autoencoder>","CreationDate":"2019-09-20T08:58:41.337","AcceptedAnswerId":58025761.0,"OwnerUserId":11726162.0,"Title":"Variable length input for LSTM autoencoder- Keras","Body":"<p>I am trying an autoencoder model with LSTM layers in Keras for text outlier detection. I have encoded every sentence into a sequence of numbers, with each number representing a letter.<\/p>\n\n<p>So far I have already trained a model with a fixed-length input, by padding zeros to each of the 4000 sequences up to maxlength = 40 thus training the model with a [4000,40,1] shaped array ([batch_size, timesteps, features]).<\/p>\n\n<p>Now I am wondering how I can use such an autoencoder model without padding zeros to each sequence (sentence) thus training and predicting with the actual size of each sentence (sequence).<\/p>\n\n<p>At the moment I have standardized every sequence so my train data (x_train) is a list of arrays and every array in list represents a standardized sequence of numbers of different lengths.<\/p>\n\n<p>To input this data to the LSTM model I am trying to reshape into 3d array with:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x_train=np.reshape(x_train, (len(x_train), 1, 1))\n<\/code><\/pre>\n\n<p>not sure if this is correct though.<\/p>\n\n<p>My model looks like this (I've removed the input_shape parameter so the model can accept variable-length input): <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nmodel = Sequential()\nmodel.add(LSTM(20, activation='tanh',return_sequences=True))\nmodel.add(LSTM(15, activation='tanh', return_sequences=True))\nmodel.add(LSTM(5, activation='tanh', return_sequences=True))\nmodel.add(LSTM(15, activation='tanh', return_sequences=True))\nmodel.add(LSTM(20, activation='tanh', return_sequences=True))\nmodel.add((Dense(1,activation='tanh')))\n\n<\/code><\/pre>\n\n<p>Then when trying to compile and train the model<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>nb_epoch = 10\nmodel.compile(optimizer='rmsprop', loss='mse')\ncheckpointer = ModelCheckpoint(filepath=\"text_model.h5\",\n                               verbose=0,\n                               save_best_only=True)\n\nes_callback = keras.callbacks.EarlyStopping(monitor='val_loss')\n\nhistory = model.fit(x_train, x_train,\n                    epochs=nb_epoch,\n                    shuffle=True,\n                    validation_data=(x_test, x_test),\n                    verbose=0,\n                    callbacks=[checkpointer,es_callback])\n<\/code><\/pre>\n\n<p>I get error : \"ValueError: setting an array element with a sequence.\"<\/p>\n\n<p>My model summary is the following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nlstm_6 (LSTM)                (None, 1, 20)             1760      \n_________________________________________________________________\nlstm_7 (LSTM)                (None, 1, 15)             2160      \n_________________________________________________________________\nlstm_8 (LSTM)                (None, 1, 5)              420       \n_________________________________________________________________\nlstm_9 (LSTM)                (None, 1, 15)             1260      \n_________________________________________________________________\nlstm_10 (LSTM)               (None, 1, 20)             2880      \n_________________________________________________________________\ndense_2 (Dense)              (None, 1, 1)              21        \n=================================================================\nTotal params: 8,501\nTrainable params: 8,501\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>So my question is if it's possible to train and predict with variable-length input sequence in an LSTM autoencoder model.<\/p>\n\n<p>And if my thinking process on text outlier detection using such a model architecture is correct.<\/p>\n","answers":[{"AnswerId":"58025761","CreationDate":"2019-09-20T09:38:24.123","ParentId":null,"OwnerUserId":"10371987","Title":null,"Body":"<p>Padding still has to be done such that the input can be 3d array(tensor), but Keras actually provides <a href=\"https:\/\/keras.io\/layers\/core\/#masking\" rel=\"nofollow noreferrer\">masking<\/a> layer for you to ignore the padded 0s in the input tensor. So that the model will not be affected by the paddings.<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import LSTM, Dense, Masking\n\nmodel = Sequential()\nmodel.add(Masking(mask_value=0.0, input_shape=(timesteps, features)))\nmodel.add(LSTM(20, activation='tanh',return_sequences=True))\nmodel.add(LSTM(15, activation='tanh', return_sequences=True))\nmodel.add(LSTM(5, activation='tanh', return_sequences=True))\nmodel.add(LSTM(15, activation='tanh', return_sequences=True))\nmodel.add(LSTM(20, activation='tanh', return_sequences=True))\nmodel.add((Dense(1,activation='tanh')))\n<\/code><\/pre>\n"}]}
+{"QuestionId":58025837,"AnswerCount":0,"Tags":"<python><keras><tf.keras>","CreationDate":"2019-09-20T09:42:46.647","AcceptedAnswerId":null,"OwnerUserId":12094772.0,"Title":"ELU activation stops improving loss after second epoch","Body":"<p>I built the following neural net using only ReLU activation and it converged some acceptable level. I would like to try another activation since I read it should improve the training in deeper layers. Here is my model:<\/p>\n\n<pre><code>model = tfk.Sequential([\n          tfk.layers.Dense(h1, input_shape=(4,), activation='elu',kernel_initializer='he_uniform'),\n          tfk.layers.GaussianDropout(drop1),    \n          tfk.layers.Dense(h2, activation= 'relu',kernel_initializer='he_uniform'),\n          tfk.layers.Dense(h3, activation= 'elu',kernel_initializer='he_uniform'),\n          tfk.layers.Dense(h4, activation= 'elu',kernel_initializer='he_uniform'),\n          tfk.layers.Dense(h4, activation= 'elu',kernel_initializer='he_uniform'),\n          tfk.layers.Dense(1,activation='relu',kernel_initializer='he_uniform'),\n       ])\n<\/code><\/pre>\n\n<p>The reason why I use a ReLu activation also in the output layer is that I expect only positive output. <\/p>\n\n<p>The problem is that it stops training after the first two epochs.<\/p>\n\n<pre><code>Epoch 1\/200\n270625\/270625 [==============================] - 41s 153us\/sample - loss: 0.3676 - val_loss: 0.3654\nEpoch 2\/200\n270625\/270625 [==============================] - 39s 146us\/sample - loss: 0.3674 - val_loss: 0.3654\nEpoch 3\/200\n270625\/270625 [==============================] - 40s 147us\/sample - loss: 0.3674 - val_loss: 0.3654\n<\/code><\/pre>\n\n<p>No matter how many epoch, the loss won't move. I tried using glorot_uniform initialization with similar results. <\/p>\n\n<p>I was wondering if the problem is in the way I pass the ELU activation in tfk.layers.Dense. Because it should receive the argument <em>alpha<\/em> but I cannot figure out how to do so in tfk.keras<\/p>\n","answers":[]}
+{"QuestionId":58026237,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><keras><resnet>","CreationDate":"2019-09-20T10:07:43.190","AcceptedAnswerId":null,"OwnerUserId":3995211.0,"Title":"Keras: Skip connections betwenn non-consecutive layers","Body":"<p>I want to implement a Keras Wrapper that modifies a layer such that the usual ResNet progression   <\/p>\n\n<p><code>x_{k+1} = x_k + f(x_k)<\/code><\/p>\n\n<p>where f is a layer will become<\/p>\n\n<p><code>x_{k+1} = g(x_{k-1}) + x_k + f(x_k)<\/code><\/p>\n\n<p>For the usual ResNet progression I know how to do this. It will look similar to this:<\/p>\n\n<pre><code>class Residual(Wrapper):\n\n    def call(self, x):\n        layer_output = self.layer.call(x)\n        output = tf.math.add(x, layer_output)\n        return output\n<\/code><\/pre>\n\n<p>However, is <code>x_{k-1}<\/code> is needed not only the output of the last but also of the forelast layer is need. I'm unsure what's the correct way to do this. The Residual Wrapper will be called like this:<\/p>\n\n<pre><code>layer_2 = Residual(Dense(28**2,input_shape=input_shape))(layer_1)\n<\/code><\/pre>\n\n<p>For the other wrapper that will be called Residual2 I tried:<\/p>\n\n<pre><code>layer_2 = Residual2(Dense(28**2,input_shape=input_shape))([layer_1,layer_0])\n<\/code><\/pre>\n\n<p>I get the following Exception;<\/p>\n\n<blockquote>\n  <p>ValueError: Layer laplace expects 1 inputs, but it received 2 input tensors. Inputs received: <code>[&lt;tf.Tensor 'dense\/BiasAdd:0' shape=(?, 784) dtype=float32&gt;, &lt;tf.Tensor 'dense\/BiasAdd:0' shape=(?, 784) dtype=float32&gt;]<\/code><\/p>\n<\/blockquote>\n\n<p>I tired to debug looking into the original keras code, but I'm not such an experienced programmer that this is really leading somewhere. Seemingly, I don't know how to properly set the input dimensions of my Wrapper. Can someone explain how to properly do this ?<\/p>\n\n<p>Thanks for your time<\/p>\n","answers":[]}
+{"QuestionId":58026250,"AnswerCount":0,"Tags":"<machine-learning><neural-network><deep-learning><conv-neural-network><pytorch>","CreationDate":"2019-09-20T10:08:27.537","AcceptedAnswerId":null,"OwnerUserId":7147097.0,"Title":"PyTorch: Dynamically freeze layers in my network?","Body":"<p>I have a network and I want to only update the first few layers' weights. The thing is, I want to change the amount of layers I update dynamically through the training process.<\/p>\n\n<p>Is there a way to do this without having to change each weight to trainable = False? (or requires_grad = False)<\/p>\n","answers":[]}
+{"QuestionId":58026858,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-20T10:44:55.393","AcceptedAnswerId":null,"OwnerUserId":12095015.0,"Title":"Difference in results based on the way keras is imported","Body":"<p>using keras API with Tensorflow, I get different results based on how I import the keras module<\/p>\n\n<p>I am testing how the keras high level API works with different frameworks. When running a simple mnist_cnn.py example from the git repo (<a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/mnist_cnn.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/mnist_cnn.py<\/a>) I get very different results based on how I import keras.<\/p>\n\n<p>Just using \"import keras\"\nresult in \nTest loss: 0.058285388482175765\nTest accuracy: 0.9814000129699707<\/p>\n\n<p>whereas if I use the tensorflow imports \"from tensorflow import keras as keras\", I get \nTest loss: 2.2431459686279296\nTest accuracy: 0.3012000024318695<\/p>\n\n<p>Why is there such a huge difference?<\/p>\n\n<p>With <code>import keras<\/code> only <\/p>\n\n<pre><code>Using TensorFlow backend.\nx_train shape: (60000, 28, 28, 1)\n60000 train samples\n10000 test samples\n2019-09-20 11:26:17.534247: I tensorflow\/core\/platform\/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\nModel: \"sequential_1\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nconv2d_1 (Conv2D)            (None, 26, 26, 32)        320       \n_________________________________________________________________\nconv2d_2 (Conv2D)            (None, 24, 24, 64)        18496     \n_________________________________________________________________\nmax_pooling2d_1 (MaxPooling2 (None, 12, 12, 64)        0         \n_________________________________________________________________\ndropout_1 (Dropout)          (None, 12, 12, 64)        0         \n_________________________________________________________________\nflatten_1 (Flatten)          (None, 9216)              0         \n_________________________________________________________________\ndense_1 (Dense)              (None, 128)               1179776   \n_________________________________________________________________\ndropout_2 (Dropout)          (None, 128)               0         \n_________________________________________________________________\ndense_2 (Dense)              (None, 10)                1290      \n=================================================================\nTotal params: 1,199,882\nTrainable params: 1,199,882\nNon-trainable params: 0\n_________________________________________________________________\nTrain on 60000 samples, validate on 10000 samples\nEpoch 1\/1\n60000\/60000 [==============================] - 83s 1ms\/step - loss: 0.2724 - accuracy: 0.9155 - val_loss: 0.0594 - val_accuracy: 0.9824\nTest loss: 0.059407291333144534\nTest accuracy: 0.9824000000953674 \n<\/code><\/pre>\n\n<p>With below statements instead of just \"import keras\"<\/p>\n\n<pre><code>import tensorflow as tf\nprint(tf.__version__)\nfrom tensorflow import keras as keras\n<\/code><\/pre>\n\n<p>I get the below results<\/p>\n\n<pre><code>2.0.0-beta1\nUsing TensorFlow backend.\nx_train shape: (60000, 28, 28, 1)\n60000 train samples\n10000 test samples\n2019-09-20 11:28:37.691293: I tensorflow\/core\/platform\/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\nModel: \"sequential_1\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nconv2d_1 (Conv2D)            (None, 26, 26, 32)        320       \n_________________________________________________________________\nconv2d_2 (Conv2D)            (None, 24, 24, 64)        18496     \n_________________________________________________________________\nmax_pooling2d_1 (MaxPooling2 (None, 12, 12, 64)        0         \n_________________________________________________________________\ndropout_1 (Dropout)          (None, 12, 12, 64)        0         \n_________________________________________________________________\nflatten_1 (Flatten)          (None, 9216)              0         \n_________________________________________________________________\ndense_1 (Dense)              (None, 128)               1179776   \n_________________________________________________________________\ndropout_2 (Dropout)          (None, 128)               0         \n_________________________________________________________________\ndense_2 (Dense)              (None, 10)                1290      \n=================================================================\nTotal params: 1,199,882\nTrainable params: 1,199,882\nNon-trainable params: 0\n_________________________________________________________________\nTrain on 60000 samples, validate on 10000 samples\nEpoch 1\/1\n60000\/60000 [==============================] - 75s 1ms\/step - loss: 2.2834 - accuracy: 0.1298 - val_loss: 2.2431 - val_accuracy: 0.3012\nTest loss: 2.2431459686279296\nTest accuracy: 0.3012000024318695 \n<\/code><\/pre>\n\n<p>Included above.<\/p>\n","answers":[]}
+{"QuestionId":58026949,"AnswerCount":1,"Tags":"<python><scikit-learn><pytorch><svd>","CreationDate":"2019-09-20T10:49:40.650","AcceptedAnswerId":58237693.0,"OwnerUserId":5516760.0,"Title":"Truncate SVD decomposition of Pytorch tensor without transfering to cpu","Body":"<p>I'm training a model in Pytorch and I want to use truncated SVD decomposition of input. For calculating SVD I transfer input witch is a Pytorch Cuda Tensor to CPU and using <code>TruncatedSVD<\/code> from <code>scikit-learn<\/code> perform truncate, after that, I transfer the result back to GPU. The following is code for my model:<\/p>\n\n<pre><code> class ImgEmb(nn.Module):\n    def __init__(self, input_size, hidden_size):\n        super(ImgEmb, self).__init__()\n        self.input_size = input_size\n        self.hidden_size = hidden_size\n        self.drop = nn.Dropout(0.2)\n        self.mlp = nn.Linear(input_size\/2, hidden_size)\n        self.relu = nn.Tanh()\n        self.svd = TruncatedSVD(n_components=input_size\/2)\n\n    def forward(self, input):\n        svd=self.svd.fit_transform(input.cpu())\n        svd_tensor=torch.from_numpy(svd)\n        svd_tensor=svd_tensor.cuda()\n        mlp=self.mlp(svd_tensor)\n        res = self.relu(mlp)\n        return res\n<\/code><\/pre>\n\n<p>I wonder is a way to implement truncated SVD without transferring back and forth to GPU? (Because it's very time consuming and is not efficient at all)<\/p>\n","answers":[{"AnswerId":"58237693","CreationDate":"2019-10-04T13:47:03.050","ParentId":null,"OwnerUserId":"4802196","Title":null,"Body":"<p>You could directly use PyTorch's SVD and truncate it manually, or you can use the truncated SVD from <a href=\"http:\/\/tensorly.org\/dev\/index.html\" rel=\"nofollow noreferrer\">TensorLy<\/a>, with the PyTorch backend:<\/p>\n\n<pre><code>import tensorly as tl\ntl.set_backend('pytorch')\n\nU, S, V = tl.truncated_svd(matrix, n_eigenvecs=10)\n<\/code><\/pre>\n\n<p>However, the GPU SVD does not scale very well on large matrices. You can also use TensorLy's partial svd which will still copy your input to CPU but will be much faster if you keep only a few eigenvalues as it will use a sparse eigendecomposition. In Scikit-learn's truncated SVD, you can also use 'algorithm = arpack' to use Scipy's sparse SVD which again might be faster if you only need a few components. <\/p>\n"}]}
+{"QuestionId":58026964,"AnswerCount":2,"Tags":"<tensorflow><deep-learning><conv-neural-network><object-detection>","CreationDate":"2019-09-20T10:50:59.533","AcceptedAnswerId":null,"OwnerUserId":1207608.0,"Title":"What is the purpose of the scales factor in Faster Rcnn Box Coder?","Body":"<p>I'm using the object detection api and tuning the parameters for a SSD task. My question refers to the box coder at <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/box_coders\/faster_rcnn_box_coder.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/box_coders\/faster_rcnn_box_coder.py<\/a>. \nWhy setting these scales factors to [10,10,5,5]? The original paper does not explain it. I suspect that it has to do either assigning a different weight to the 4 components of location error (tx, ty, tw, th) or  with some numerical stability issue, but I would like to have a confirmation. Thanks<\/p>\n","answers":[{"AnswerId":"58908960","CreationDate":"2019-11-18T06:21:50.250","ParentId":null,"OwnerUserId":"6319495","Title":null,"Body":"<p>The network predicts changes for each anchor box. That is, for each anchor block, it predicts an offset for x, y position and width, height.<\/p>\n\n<p>a short description of these parameters can be found, for example to link:\n<a href=\"https:\/\/medium.com\/@smallfishbigsea\/understand-ssd-and-implement-your-own-caa3232cd6ad\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@smallfishbigsea\/understand-ssd-and-implement-your-own-caa3232cd6ad<\/a>\n<a href=\"https:\/\/lambdalabs.com\/blog\/how-to-implement-ssd-object-detection-in-tensorflow\/\" rel=\"nofollow noreferrer\">https:\/\/lambdalabs.com\/blog\/how-to-implement-ssd-object-detection-in-tensorflow\/<\/a><\/p>\n"},{"AnswerId":"58919370","CreationDate":"2019-11-18T16:49:38.867","ParentId":null,"OwnerUserId":"1207608","Title":null,"Body":"<p>I find the answer here <a href=\"https:\/\/leimao.github.io\/blog\/Bounding-Box-Encoding-Decoding\/\" rel=\"nofollow noreferrer\">https:\/\/leimao.github.io\/blog\/Bounding-Box-Encoding-Decoding\/<\/a>, where the variables are used as some sort of Representation Encoding With Variance. The question was also the subject of this issue <a href=\"https:\/\/github.com\/rykov8\/ssd_keras\/issues\/53\" rel=\"nofollow noreferrer\">https:\/\/github.com\/rykov8\/ssd_keras\/issues\/53<\/a><\/p>\n"}]}
+{"QuestionId":58027224,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2019-09-20T11:08:07.530","AcceptedAnswerId":null,"OwnerUserId":9032335.0,"Title":"load Tensor Flow dataset \"stanford_dogs\"","Body":"<p>I'd like to use the 'stanford_dogs' dataset of the tensor_flow_datasets collection.\nIt is listed here:\n<a href=\"https:\/\/www.tensorflow.org\/datasets\/catalog\/stanford_dogs\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/datasets\/catalog\/stanford_dogs<\/a><\/p>\n\n<p>but when trying to load, it's not listed among the available datasets.\nHow to load 'Stanford_Dogs'?<\/p>\n\n<pre><code>import tensorflow_datasets as tfds\n\nSPLIT_WEIGHTS = (8, 1, 1)\nsplits = tfds.Split.TRAIN.subsplit(weighted=SPLIT_WEIGHTS)\n(raw_train, raw_validation, raw_test), metadata = tfds.load(name=\"stanford_dogs\", \n                                                            with_info=True,\n                                                            split=list(splits),\n                                                             batch_size=-1,\n                                                            as_supervised=True)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58027364,"AnswerCount":0,"Tags":"<python-3.x><keras><deep-learning><anaconda>","CreationDate":"2019-09-20T11:18:01.207","AcceptedAnswerId":null,"OwnerUserId":7380691.0,"Title":"TerminatedWorkerError: A worker process managed by the executor was unexpectedly terminated","Body":"<p>I am trying to find optimal hyperparameters for my Deep learning model. When I used 'GridSearchCV' approach and tried fitting my model, I ended up with error <\/p>\n\n<blockquote>\n  <p>\"TerminatedWorkerError: A worker process managed by the executor was unexpectedly terminated. This could be caused by a segmentation fault while calling the function or by an excessive memory usage causing the Operating System to kill the worker. The exit codes of the workers are {EXIT(1)}\"<\/p>\n<\/blockquote>\n\n<p>I have updated and tested scipy libraries, and that works fine.\nI can't use n_jobs=-1. I must use 1 upto 16 CPUs (n_jobs=16).<\/p>\n\n<p>My platform:\nNAME=\"Ubuntu\"\nVERSION=\"16.04.6 LTS (Xenial Xerus)\".\nI am accessing sklearn using conda 4.7.12 in python 3.6.9.\nWe can't update Ubuntu at this time!<\/p>\n\n<pre><code>import pandas as pd\nimport numpy\nimport keras\nimport scipy\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.wrappers.scikit_learn import KerasClassifier\nfrom sklearn.model_selection import GridSearchCV\nfrom sklearn.model_selection import train_test_split\nfrom keras.optimizers import SGD\nfrom sklearn.externals import joblib\nfrom math import sqrt\nimport tensorflow as tf\nfrom datetime import datetime\n\n\n# Make grid input\nbatch_size  = [1644, 822, 548, 411, 328, 274, 234, 205, 182, 164]\nepochs = [10, 50, 100, 150, 200, 250, 300]\nlearn_rate = [0.00001, 0.0001, 0.001, 0.01, 0.1, 0.2, 0.3]\nmomentum = [0.0, 0.2, 0.4, 0.6, 0.8, 0.9]\nparam_grid = dict(epochs = epochs, batch_size = batch_size, learn_rate=learn_rate, momentum=momentum)\n\n# Define model\ndef get_compiled_model_advanced(learn_rate=learn_rate, momentum=momentum):\n    model = tf.keras.Sequential([\n        tf.keras.layers.Dense(200, activation='relu'),\n        tf.keras.layers.Dense(50, activation='relu'),\n        tf.keras.layers.Dense(10, activation='relu'),\n        tf.keras.layers.Dense(1, activation='sigmoid')])\n    # Compile model\n    optimizer = SGD(lr=learn_rate, momentum=momentum, decay=0.5, nesterov=False)\n    model.compile(optimizer=optimizer, loss='binary_crossentropy', metrics=['accuracy'])\n    return model\n\nmodel_sgd = KerasClassifier(build_fn=get_compiled_model_advanced, verbose=0)\n\ngrid_sgd = GridSearchCV(estimator = model_sgd, param_grid = param_grid, n_jobs = 16, verbose=0)\nsgd_fit = grid_sgd.fit(X, Y)\n<\/code><\/pre>\n\n<p>I would like to get my model running!\nCould someone tell me what am I doing wrong?<\/p>\n","answers":[]}
+{"QuestionId":58027458,"AnswerCount":1,"Tags":"<php><python><python-3.x><macos><tensorflow>","CreationDate":"2019-09-20T11:24:37.243","AcceptedAnswerId":null,"OwnerUserId":9786502.0,"Title":"PHP returns NULL running Python Script with Tensorflow","Body":"<p>First of all, sorry if this seems like a duplicate question. But honestly, nothing has worked for me. Searched for hours here.<\/p>\n\n<p>Okay, so I have a python file called <code>label_image.py<\/code>, the code is the following:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nprint(\"TEST\")\n<\/code><\/pre>\n\n<p>When I tried to call the script via a PHP file:<\/p>\n\n<pre><code>$outpt = shell_exec('python label_image.py');\n\nvar_dump($outpt);\n<\/code><\/pre>\n\n<p>this returns NULL. <\/p>\n\n<p>But if I remove the import of tensorflow in the python file, so it's the following:<\/p>\n\n<pre><code>import numpy as np\nprint(\"TEST\")\n<\/code><\/pre>\n\n<p>and runs the PHP file, it outputs the following:<\/p>\n\n<pre><code>string(5) \"TEST \" \n<\/code><\/pre>\n\n<p>With this test, it can be deduced that the line of importing Tensorflow causes the return of NULL. <\/p>\n\n<p>I tried putting this on top of my PHP file, as I've read on another stackoverflow thread:<\/p>\n\n<pre><code>PUTENV(\"\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/\");\n<\/code><\/pre>\n\n<p>but it doesn't work, still. I am using Mac OS by the way.<\/p>\n\n<p>I would appreciate any help.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"58027859","CreationDate":"2019-09-20T11:51:50.957","ParentId":null,"OwnerUserId":"372172","Title":null,"Body":"<p>If there is a problem locating the tensor flow library, I think you may try setting <code>PYTHONPATH<\/code> like this:<\/p>\n\n<pre class=\"lang-php prettyprint-override\"><code>&lt;?php\n\n\/\/ PYTHONPATH setup\n$pythonpath = getenv('PYTHONPATH'); \/\/ see if there is current setting\n$additional = '\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/'; \/\/ the location where the Tensorflow library is installed.\nputenv(($pythonpath === false) ? \"PYTHONPATH=$additional\" : \"PYTHONPATH=$pythonpath:$additional\");\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":58027705,"AnswerCount":0,"Tags":"<python><tensorflow><keras><neural-network>","CreationDate":"2019-09-20T11:41:48.757","AcceptedAnswerId":null,"OwnerUserId":12095275.0,"Title":"What is the maximum value the categorical cross-entropy loss can achieve?","Body":"<p>I'm developing a Neural Network using Keras and Tensorflow. The aim of the network is to segment a fragment os a spectrogram in three possible classes. I'm using the softmax function as the activation and the categorical cross-entropy loss. My tags are categorical so [1 0 0], [0 1 0] and [0 0 1]. Every pixel should have one. My doubt comes when I'm doing the training and validation and I have the output of the loss for each epoch. Do I have a maximum value that it can get?<\/p>\n","answers":[]}
+{"QuestionId":58027839,"AnswerCount":2,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-09-20T11:50:30.050","AcceptedAnswerId":null,"OwnerUserId":5380380.0,"Title":"How to fine-tune a keras model with existing plus newer classes?","Body":"<p>Good day!<\/p>\n\n<p>I have a celebrity dataset on which I want to fine-tune a keras built-in model. SO far what I have explored and done, we remove the top layers of the original model (or preferably, pass the include_top=False) and add our own layers, and then train our newly added layers while keeping the previous layers frozen. This whole thing is pretty much like intuitive.<\/p>\n\n<p>Now what I require is, that my model learns to identify the celebrity faces, while also being able to detect all the other objects it has been trained on before. Originally, the models trained on imagenet come with an output layer of 1000 neurons, each representing a separate class. I'm confused about how it should be able to detect the new classes? All the transfer learning and fine-tuning articles and blogs tell us to replace the original 1000-neuron output layer with a different N-neuron layer (N=number of new classes). In my case, I have two celebrities, so if I have a new layer with 2 neurons, I don't know how the model is going to classify the original 1000 imagenet objects.<\/p>\n\n<p>I need a pointer on this whole thing, that how exactly can I have a pre-trained model taught two new celebrity faces while also maintaining its ability to recognize all the 1000 imagenet objects as well.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"58028414","CreationDate":"2019-09-20T12:28:21.070","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>With transfer learning, you can make the trained model classify among the new classes on which you just trained using the features learned from the new dataset and the features learned by the model from the dataset on which it was trained in the first place. Unfortunately, you can not make the model to classify between all the classes (original dataset classes + second time used dataset classes), because when you add the new classes, it keeps their weights only for classification.\nBut, let's say for experimentation you change the number of output neurons (equal to the number of old + new classes) in the last layer, then it will now give random weights to these neurons which on prediction will not give you meaningful result. <\/p>\n\n<p>This whole thing of making the model to classify among old + new classes experimentation is still in research area. \nHowever, one way you can achieve it is to train your model from scratch on the whole data (old + new).<\/p>\n"},{"AnswerId":"58029620","CreationDate":"2019-09-20T13:45:02.903","ParentId":null,"OwnerUserId":"5136550","Title":null,"Body":"<p>CNN is prone to forgetting the previously learned knowledge when retrained for a new task on novel domain and this phenomenon is often called as catastrophic forgetting which is an active and challenging research domain.<\/p>\n\n<p>Coming to the point, One obvious way to enabling model to classify new class along with old class is to train from scratch on the accumulated (old+new) dataset that is time consuming. \nOn the other hand, using transfer learning approach, you freeze the weight of backbone model for feature extraction and learn the output layer's weight (from new dataset) with your target number of new class\/node. So, we lost previously learned knowledge.\nFinally, one work around to solve your problem (i.e. enabling model to work for both old+new classes) is to concat old and new output layer keeping the fixed backbone, as depicted in the figure:\n<a href=\"https:\/\/i.stack.imgur.com\/sGqup.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/sGqup.jpg\" alt=\"enter image description here\"><\/a> <\/p>\n\n<p>More precisely, you can build new model with freezed backbone and old+new number of class in output layer followed by loading weight of output layer from  previous two models (pre-trained and newly trained model) in a specialized way.<\/p>\n"}]}
+{"QuestionId":58028064,"AnswerCount":1,"Tags":"<scikit-learn><pytorch>","CreationDate":"2019-09-20T12:04:37.627","AcceptedAnswerId":58028716.0,"OwnerUserId":11948285.0,"Title":"sklearn pipeline with multiple inputs\/outputs","Body":"<p>How can I build a sklearn pipeline to do the following?<\/p>\n\n<p>What I have:<\/p>\n\n<pre><code>A, B = getAB(X_train)\nX_train = transform(X_train)\nmodel(A, B, X_train)\n<\/code><\/pre>\n\n<p>What I want:<\/p>\n\n<pre><code>pipe = Pipeline([\n(\u2018ab\u2019, getAB),\n(\u2018tranf\u2019, transform),\n(\u2018net\u2019, net)\n]\npipe.fit(X_train, y_train)\n<\/code><\/pre>\n\n<p>Please help!<\/p>\n","answers":[{"AnswerId":"58028716","CreationDate":"2019-09-20T12:48:12.323","ParentId":null,"OwnerUserId":"1729802","Title":null,"Body":"<p>Yes it is doable by writing a custom transformer that has a fit\/transform function. This can be your class:<\/p>\n\n<pre><code>from sklearn.base import BaseEstimator, TransformerMixin\n\ndef getABTransformer(BaseEstimator, TransformerMixin):\n    def __init__(self): # no *args or **kargs\n        pass\n\n    def fit(self, X, y=None):\n        return self # nothing else to do\n\n    def transform(self, X, y=None):\n        return getAB(X)\n<\/code><\/pre>\n\n<p>Then you can create your <code>ColumnTransformer<\/code> as following:<\/p>\n\n<pre><code>from sklearn.compose import ColumnTransformer\n\nclm_pipe = ColumnTransformer([\n(\u2018ab\u2019, getABTransformer, np.arange(0, len(X_train)),  # list of columns indices\n(\u2018tranf\u2019, transform, np.arange(0, len(X_train))),  # list of columns indices\n]\n<\/code><\/pre>\n\n<p>and a final pipeline with the model:<\/p>\n\n<pre><code>pipe = Pipeline([\n(\u2018clm_pipe\u2019, clm_pipe),\n(\u2018net\u2019, net)\n]\n<\/code><\/pre>\n\n<p>You can read more about <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.compose.ColumnTransformer.html\" rel=\"nofollow noreferrer\">ColumnTransformer<\/a><\/p>\n"}]}
+{"QuestionId":58028376,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><linear-regression>","CreationDate":"2019-09-20T12:25:47.560","AcceptedAnswerId":58028767.0,"OwnerUserId":9625777.0,"Title":"Unable to understand format of test data while evaluating training model","Body":"<p>I am training a regression model that takes approximates the weights for the equation :\n<strong>Y = R+B+G<\/strong>\nFor this, I provide pre-determined values of R, B and G and Y, as training data.<\/p>\n\n<pre><code>    R    = np.array([-4, -10,  -2,  8, 5, 22,  3],  dtype=float)\n    B    = np.array([4, -10,  0,  0, 15, 5,  1],  dtype=float)\n    G    = np.array([0, 10,  5,  8, 1, 2,  38],  dtype=float)\n\n    Y    = np.array([0, -10, 3, 16, 21, 29, 42],  dtype=float)\n<\/code><\/pre>\n\n<p>The training batch consisted of 1x3 array corresponding to Ith value of R, B and G.<\/p>\n\n<pre><code>    RBG = np.array([R,B,G]).transpose()\n    print(RBG)\n\n    [[ -4.   4.   0.]\n    [-10. -10.  10.]\n    [ -2.   0.   5.]\n    [  8.   0.   8.]\n    [  5.  15.   1.]\n    [ 22.   5.   2.]\n    [  3.   1.  38.]]\n<\/code><\/pre>\n\n<p>I used a neural network with 3 inputs, 1 dense layer (hidden layer) with 2 neurons and the output layer (output) with a single neuron.<\/p>\n\n<pre><code>    hidden = tf.keras.layers.Dense(units=2, input_shape=[3])\n    output = tf.keras.layers.Dense(units=1)\n<\/code><\/pre>\n\n<p>Further, I trained the model<\/p>\n\n<pre><code>    model = tf.keras.Sequential([hidden, output])\n    model.compile(loss='mean_squared_error', \n    optimizer=tf.keras.optimizers.Adam(0.1))\n    history = model.fit(RBG,Y, epochs=500, verbose=False)\n    print(\"Finished training the model\")\n<\/code><\/pre>\n\n<p>The loss vs epoch plot was as normal, decreasing and then flat.<\/p>\n\n<p>But when I tested the model, using random values of R, B and G as<\/p>\n\n<pre><code>    print(model.predict([[1],[1],[1]]))\n<\/code><\/pre>\n\n<p>expecting the output to be 1+1+1 = 3, but got the Value Error:<\/p>\n\n<pre><code>    ValueError: Error when checking input: expected dense_2_input to have shape (3,) but got array with shape (1,)\n<\/code><\/pre>\n\n<p>Any idea where I might be getting wrong?<\/p>\n\n<p>Surprisingly, the only input it responds to, is the training data itself. i.e, <\/p>\n\n<pre><code>    print(model.predict(RBG))\n\n    [[ 2.1606684e-07]\n    [-3.0000000e+01]\n    [-3.2782555e-07]\n    [ 2.4000002e+01]\n    [ 4.4999996e+01]\n    [ 2.9000000e+01]\n    [ 4.2000000e+01]]\n<\/code><\/pre>\n","answers":[{"AnswerId":"58028767","CreationDate":"2019-09-20T12:51:30.510","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>As the error says, the problem is in your shape of the input. You need to transpose <code>[[1],[1],[1]]<\/code> this input then you will have the shape that is expected by the model.<\/p>\n\n<p>so <code>npq = np.array([[1],[1],[1]]).transpose()<\/code> and now feed this to <code>model.predict(npq)<\/code><\/p>\n"}]}
+{"QuestionId":58028909,"AnswerCount":1,"Tags":"<python><numpy><matplotlib><pytorch><networkx>","CreationDate":"2019-09-20T13:01:15.670","AcceptedAnswerId":null,"OwnerUserId":1066031.0,"Title":"Categorical Classification probabilites as a graph","Body":"<p>What is the best way to take the categorical classification probabilities typically output from a neural network and plot them as a graph using something like <code>networkx<\/code>.<\/p>\n\n<p>Given 8 categories, say <code>['A', 'B', 'C', 'D', 'E', 'F', 'G', 'H']<\/code> and the predictions across 5 samples with shape <code>(8, 5)<\/code> I'd like to plot a graph where the node label is the category and the probability for all indices where the probability is greater than 0.0.  <\/p>\n\n<pre><code>array([[0.        , 0.        , 0.        , 0.        , 0.        ],\n       [0.        , 0.        , 0.        , 0.        , 0.        ],\n       [0.17817138, 0.11233618, 0.12554741, 0.16154018, 0.16248149],\n       [0.        , 0.        , 0.        , 0.        , 0.        ],\n       [0.70687366, 0.86215913, 0.85997397, 0.8285762 , 0.828603  ],\n       [0.        , 0.        , 0.        , 0.        , 0.        ],\n       [0.        , 0.        , 0.        , 0.        , 0.        ],\n       [0.        , 0.        , 0.        , 0.        , 0.        ]],\n      dtype=float32)\n<\/code><\/pre>\n\n<p>Valid edges are between all the non zero values of consecutive columns in the array.<\/p>\n","answers":[{"AnswerId":"58080288","CreationDate":"2019-09-24T12:24:16.230","ParentId":null,"OwnerUserId":"3381305","Title":null,"Body":"<p>If I understand correctly, you want a graph in which vertices represent classes and edges represent two classes having non-zero probability for the same data instance.<\/p>\n\n<p>Starting with data similar to yours (since yours has only 2 classes ever having non-zero proability)...<\/p>\n\n<pre><code>import numpy as np\n\nA = np.array([[0.        , 0.        , 0.        , 0.8285762 , 0.        ],\n              [0.        , 0.        , 0.        , 0.        , 0.        ],\n              [0.17817138, 0.        , 0.12554741, 0.        , 0.16248149],\n              [0.        , 0.        , 0.        , 0.10000000, 0.        ],\n              [0.70687366, 0.        , 0.85997397, 0.06154018, 0.828603  ],\n              [0.        , 0.11233618, 0.        , 0.        , 0.        ],\n              [0.11495425, 0.        , 0.        , 0.        , 0.        ],\n              [0.        , 0.86215913, 0.        , 0.        , 0.        ]\n             ])\n<\/code><\/pre>\n\n<p>There's probably a cunning NumPy move here, but with a loop over classes we can build an edge list:<\/p>\n\n<pre><code>from itertools import combinations\n\nedge_list = []\nidx = np.argwhere(A.T)\nfor i, _ in enumerate(A):\n    this = idx[idx[:, 0]==i]\n    combs = combinations(this[:, 1], r=2)\n    edge_list.extend(list(combs))\nedge_list\n<\/code><\/pre>\n\n<p>You can instantiate a <code>networkx.Graph<\/code> and add the edges like so:<\/p>\n\n<pre><code>import networkx as nx\n\nG = nx.Graph()\nG.add_edges_from(edge_list)\n<\/code><\/pre>\n\n<p>Then you can draw the graph with <code>nx.draw(G)<\/code> but th default drawing is not great. We can prepare the node size and labels...<\/p>\n\n<pre><code>nx.set_node_attributes(G, {i:chr(65+i) for i in range(len(A))}, 'name')\nnx.set_node_attributes(G, {i:sum(a) for i, a in enumerate(A)}, 'prob')\n\nprobs = nx.get_node_attributes(G, 'prob')\nsizes = [1000*x for x in probs.values()]\nlabels = nx.get_node_attributes(G, 'name')\n<\/code><\/pre>\n\n<p>Then make a better plot:<\/p>\n\n<pre><code>%matplotlib inline\nimport matplotlib.pyplot as plt\n\nfig, ax = plt.subplots(figsize=(10, 6))\n\npos = nx.spring_layout(G)\nnx.draw_networkx_nodes(G, pos, ax=ax, node_size=sizes, node_color='orange')\nnx.draw_networkx_edges(G, pos, ax=ax, width=4, splines='curved')\nnx.draw_networkx_labels(G, pos,\n                        labels=labels,\n                        font_size=20,\n                        font_family='sans-serif',\n                        font_color='blue')\n\nplt.axis('off')\nplt.show()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vynoV.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/vynoV.png\" alt=\"A plot of the resulting graph.\"><\/a><\/p>\n"}]}
+{"QuestionId":58028976,"AnswerCount":6,"Tags":"<python><tensorflow><keras><deep-learning><conv-neural-network>","CreationDate":"2019-09-20T13:05:45.660","AcceptedAnswerId":null,"OwnerUserId":8191929.0,"Title":"TypeError: Unexpected keyword argument passed to optimizer: learning_rate","Body":"<p>I am trying to load a Keras model which was trained on an Azure VM (NC promo). But I am getting the following error.<\/p>\n\n<blockquote>\n  <p>TypeError: Unexpected keyword argument passed to optimizer:learning_rate<\/p>\n<\/blockquote>\n\n<p>EDIT:<\/p>\n\n<p>Here is the code snippet that I am using to load my model:<\/p>\n\n<pre><code>from keras.models import load_model\nmodel = load_model('my_model_name.h5')\n<\/code><\/pre>\n","answers":[{"AnswerId":"58029409","CreationDate":"2019-09-20T13:32:28.883","ParentId":null,"OwnerUserId":"9794932","Title":null,"Body":"<p>Did you use a custom optimizer?<\/p>\n\n<p>If so, you can load like this:<\/p>\n\n<pre><code>model = load_model('my_model_name.h5', custom_objects={\n    'Adam': lambda **kwargs: hvd.DistributedOptimizer(keras.optimizers.Adam(**kwargs))\n})\n<\/code><\/pre>\n\n<p>Alternatively you can load your model with <code>model = load_model('my_model_name.h5', compile=False)<\/code> and then add an optimizer and recompile, but that will lose your saved weights.<\/p>\n"},{"AnswerId":"58569974","CreationDate":"2019-10-26T10:32:48.497","ParentId":null,"OwnerUserId":"9158477","Title":null,"Body":"<p>I had the same problem. Using Keras version 2.3.1 and TensorFlow-GPU version 1.13, I had to upgrade Tensorflow-GPU to version 1.15<\/p>\n\n<pre><code>pip uninstall tensorflow-gpu\npip install tensorflow-gpu==1.15\n<\/code><\/pre>\n"},{"AnswerId":"58550566","CreationDate":"2019-10-25T00:00:06.073","ParentId":null,"OwnerUserId":"12271466","Title":null,"Body":"<p>This happened to me too. Most likely because the <code>learning_rate<\/code> was renamed from version 2.2.* to 2.3.0 in September 2018. \n(see release notes: <a href=\"https:\/\/github.com\/keras-team\/keras\/releases\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/releases<\/a> : \nRename lr to learning_rate for all optimizers. )<\/p>\n\n<p>This worked for me:\n    sudo pip install keras --upgrade <\/p>\n"},{"AnswerId":"58582766","CreationDate":"2019-10-27T19:46:13.437","ParentId":null,"OwnerUserId":"12282702","Title":null,"Body":"<p>i got the same error while i was working in two different PC. \nin some versions of tensorflow is tf.keras.optimizers.SGD(lr = x) while in other verions istf.keras.optimizers.SGD(learning rate = x). <\/p>\n"},{"AnswerId":"58829361","CreationDate":"2019-11-13T02:14:40.590","ParentId":null,"OwnerUserId":"9254522","Title":null,"Body":"<p>I've had a similar problem.<\/p>\n\n<p>You you have this issue, try to use <code>lr<\/code> instead of <code>learning_rate<\/code> when defining the learning rate in your optimizer. <\/p>\n"},{"AnswerId":"58368962","CreationDate":"2019-10-14T01:05:55.170","ParentId":null,"OwnerUserId":"9603973","Title":null,"Body":"<p>I am also experiencing this when I try to load my model on another machine. Also trained the initial modal on an azure VM. I have tried the suggestions above and can't figure out what is causing it. Any other thoughts? Below is my code to train the model.<\/p>\n\n<p>Models were trained and are being used in my api projects using the following versions:\nkeras 2.3.0\ntensorflow 1.14.0<\/p>\n\n<pre><code>history = model.fit(X, y,validation_split=0.1, \\\n                epochs=20, \\\n                batch_size=1000, \\\n                class_weight = cw)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58029586,"AnswerCount":0,"Tags":"<python><tensorflow><computer-vision><quantization>","CreationDate":"2019-09-20T13:43:13.513","AcceptedAnswerId":null,"OwnerUserId":10715858.0,"Title":"How to quantize model to deploy in tensorflow?","Body":"<p>I am trying to quantize my model with tensorflow and TransformGraph quantize_weights <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/tools\/graph_transforms\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/tools\/graph_transforms<\/a> this is good working when model not quantize during training. <\/p>\n\n<p>But I am try train with tf.contrib.quantize.create_training_graph, and when I freeze my model with TransformGraph optimization i get violation of accuracy. <\/p>\n\n<p>What difference between tf.contrib.quantize.create_training_graph and TransformGraph quantize_weights?<\/p>\n\n<p>How I can train quantize model, freeze it and deploy to protobuf file to inference in my pc (not tflite!)?<\/p>\n","answers":[]}
+{"QuestionId":58030543,"AnswerCount":2,"Tags":"<python><keras><iterable-unpacking>","CreationDate":"2019-09-20T14:40:35.197","AcceptedAnswerId":58032704.0,"OwnerUserId":5017267.0,"Title":"list of keras callbacks generates error: 'tuple' object has no attribute 'set_model'","Body":"<p>I am writing a keras model where I want to use a few built-in keras callbacks, however I am probably making a grammar mistake somewhere that I cannot spot. The piece of code giving me troubles is the following:<\/p>\n\n<pre><code>from keras.callbacks import EarlyStopping, ModelCheckpoint, TensorBoard\n...\n...\nes = EarlyStopping(monitor='val_loss', min_delta=0.01, verbose=1, patience=5)\ntb = TensorBoard(log_dir=logdir, write_graph=True, write_images=True, histogram_freq=0)\nmc = ModelCheckpoint(filepath=filepath, save_best_only=True, monitor='val_loss', mode='min')\n\nhistory = model.fit(X_train, y_train,\n                    batch_size=batch_size,\n                    epochs=n_epochs,\n                    verbose=1,\n                    validation_split=0.3,\n                    callbacks=[es, tb, mc])\n<\/code><\/pre>\n\n<p>however in doing so I get the error <code>'tuple' object has no attribute 'set_model'<\/code>. Referring to <a href=\"https:\/\/stackoverflow.com\/questions\/53063339\/attributeerror-list-object-has-no-attribute-set-model\">this<\/a> other question it seems that the problem is generated by the fact that <code>es, tb<\/code> are already tuples per s\u00e9 and therefore positioning them into a list (in the call <code>callbacks=[es, tb, mc]<\/code>) raises the error. In fact <\/p>\n\n<pre><code>print(type(es))\nprint(type(tb))\nprint(type(mc))\n\n&lt;class 'tuple'&gt;\n&lt;class 'tuple'&gt;\n&lt;class 'keras.callbacks.ModelCheckpoint'&gt;\n<\/code><\/pre>\n\n<p>This said, I do not understand how to work around it. <code>EarlyStopping<\/code> and <code>TensorBoard<\/code> return tuples, how are they supposed to be called in the keras callbacks list?<\/p>\n","answers":[{"AnswerId":"58032704","CreationDate":"2019-09-20T17:17:32.073","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>Unpack your tuples - in this case, it's simple: <code>(object,)[0] == object<\/code> - but in general, you may have <code>(object1, object2)<\/code>, etc, which you can handle via <code>callbacks=[*es, *tb, *mc]<\/code>.<\/p>\n\n<p>The <code>*<\/code> <em>unpacks iterables<\/em> - as a demo:<\/p>\n\n<pre><code>def print_unpacked(*positional_args):\n    print(positional_args)\n    print(*positional_args)\na = 1\nb = ('dog',5)\n<\/code><\/pre>\n\n<pre><code>print_unpacked(a,b)\n# &gt;&gt; (1, ('dog',5))\n# &gt;&gt; 1 ('dog',5)\nprint(a,b)\n# &gt;&gt; 1 ('dog',5)\nprint(a,*b)\n# &gt;&gt; 1 'dog' 5\n<\/code><\/pre>\n"},{"AnswerId":"58033158","CreationDate":"2019-09-20T17:56:57.557","ParentId":null,"OwnerUserId":"2409563","Title":null,"Body":"<p>You should remove comma at the end of following lines in your code posted above<\/p>\n\n<pre><code>es = EarlyStopping(monitor='val_loss', min_delta=0.01, verbose=1, patience=5),\ntb = TensorBoard(log_dir=logdir, write_graph=True, write_images=True, histogram_freq=0),\n<\/code><\/pre>\n"}]}
+{"QuestionId":58031018,"AnswerCount":1,"Tags":"<pandas><csv><keras><save><prediction>","CreationDate":"2019-09-20T15:09:21.097","AcceptedAnswerId":null,"OwnerUserId":11540430.0,"Title":"How can i save Keras predicted values?","Body":"<p>i want to save the predicted steering angle values from keras model in csv or txt file.<\/p>\n\n<p>I want to use it for drawing line graph. I tried to draw without saving values but it looked complicated.<\/p>\n\n<p>I used pandas to saving values but i got an error message.<\/p>\n\n<p>I wrote:<\/p>\n\n<pre><code>degrees = float(model.predict(image, batch_size=1))\ndegrees = pd.DataFrame(degrees, columns=['pred1']).to_csv('data1.csv')\n<\/code><\/pre>\n\n<p>and i got this message.<\/p>\n\n<pre><code>ValueError: DataFrame constructor not properly called!\n<\/code><\/pre>\n\n<p>How can i save the predicted values properly?<\/p>\n","answers":[{"AnswerId":"58032711","CreationDate":"2019-09-20T17:18:06.727","ParentId":null,"OwnerUserId":"5679285","Title":null,"Body":"<p>From what I can understand from your question, your model is returning a single value which you are converting into <code>float<\/code>. Pandas dataframe are basically tables having more than one column. Since <code>degrees<\/code> variable is a single value, you should use Pandas Series here. This should work:<\/p>\n\n<pre><code>degrees = float(model.predict(image, batch_size=1))\ndegrees = pd.Series(degrees).to_csv('data1.csv', header='pred1')\n<\/code><\/pre>\n"}]}
+{"QuestionId":58031343,"AnswerCount":2,"Tags":"<tensorflow><keras><tf.keras>","CreationDate":"2019-09-20T15:32:03.690","AcceptedAnswerId":58172123.0,"OwnerUserId":12092080.0,"Title":"Strange dimension behaviour: needs both dimension 2 and 3 unsure why","Body":"<p>I am trying to convert a simple model to TFLite and run into the following issue with dimensions.<\/p>\n\n<p>I've already tried using perm=[1,0] and perm=[0,2,1] the first one will generate an error requiring 3 dimensions and the 2nd one will generate an error requiring 2 dimensions.<\/p>\n\n<pre><code>import tensorflow as tf\n\ncaptions = tf.keras.layers.Input(shape=[5,1024], name='captions')\ncap_i = tf.keras.layers.Lambda(lambda q: q[0][:5,:])([captions])\ncap_iT = tf.keras.layers.Lambda(lambda query:tf.transpose(query, \nperm=[0,2,1]))(cap_i)\nmodel = tf.keras.models.Model(inputs=[captions], outputs=[cap_iT])\nmodel.save('my_model.hd5')\nconverter = \ntf.lite.TFLiteConverter.from_keras_model_file('my_model.hd5')\ntflite_model = converter.convert()\nopen(\"converted_modelfile.tflite\", \"wb\").write(tflite_model)\n<\/code><\/pre>\n\n<blockquote>\n  <p>ValueError: Dimension must be 2 but is 3 for 'lambda_1\/transpose' (op: 'Transpose') with input shapes: [5,1024], [3].<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"58034634","CreationDate":"2019-09-20T20:14:55.727","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>You are probably getting the error in two different places.<\/p>\n\n<p>You are throwing away the batch size dimension in the first <code>Lambda<\/code> with  <code>q[0]<\/code>. You should not do this, you will need the batch dimension at the end of the Keras model (probably the location of the other error). Although you are passing <code>[captions]<\/code> inside a list, it is probably automatically getting the element inside the list because it's a single tensor. <\/p>\n\n<p>The message in your question is certainly in the second <code>Lambda<\/code>, where you have a tensor with two dimensions <code>[5,1024]<\/code> (because you threw away the batch size in the first <code>Lambda<\/code>) and you are trying to permute 3 dimensions with <code>[0,2,1]<\/code>.   <\/p>\n"},{"AnswerId":"58172123","CreationDate":"2019-09-30T16:49:34.217","ParentId":null,"OwnerUserId":"12092080","Title":null,"Body":"<p>Found a nice way to fix the inputs using a compatible operation in TFLite.<\/p>\n\n<pre><code>import tensorflow.compat.v1 as tf\nimport numpy as np\n\ntf.disable_v2_behavior()\ninitial_input = tf.placeholder(dtype=tf.float32, shape=(None,5,1024))\ncap_i = tf.strided_slice(initial_input, [0,0,0], [0,5,1024], [1,1,1], shrink_axis_mask=1)\ncap_i_reshaped =tf.reshape(cap_i,[1,5,1024])\ncap_iT = tf.transpose(cap_i_reshaped, perm=[0,2,1])\n\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\ntf.io.write_graph(sess.graph_def, '', 'train.pbtxt')\nconverter = tf.lite.TFLiteConverter.from_session(sess, [initial_input], [cap_iT])\nconverter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS, \ntf.lite.OpsSet.SELECT_TF_OPS]\ntflite_model = converter.convert()\nopen('converted_model.tflite', \"wb\").write(tflite_model)\nsess.close()\n<\/code><\/pre>\n"}]}
+{"QuestionId":58031787,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-20T16:00:41.800","AcceptedAnswerId":null,"OwnerUserId":5892689.0,"Title":"InvalidArgumentError while trying to make a Regression and Classification with Keras functional API and SparseCategoricalCrossentropy Loss","Body":"<p>I want to make a Regression and a Classification in the same model. The reason is I want to port-it to TFLite. The classification I need to use SparseCategoricalcrossEntropy so I can get predicted directly the class index.\nTo build it in the same model I use Functional API from Keras.<\/p>\n\n<p>So I would like to get a prediction of the class_index and the regression in the sam model.<\/p>\n\n<p>A simplified version of the code is this:<\/p>\n\n<pre><code>import numpy as np\nfrom keras.layers import Input, Dense\nfrom keras.models import Model \n\nX=np.random.random(size=(100,5)) \ny=np.random.randint(0,100,size=(100,2)).astype(float)   #Regression\nclass_index=np.random.randint(0,2,size=(100,1))         #Classification\n\ninput1 = Input(shape=(5,))\nt = Dense(10, activation='relu')(input1)\nout1 = Dense(2, activation='softmax')(t)\n\nt2 = Dense(50, activation='relu')(t) \nt3 = Dense(50, activation='relu')(t2) \nout2 = Dense(2)(t3)\n\nmodel = Model(inputs=input1, outputs=[out1, out2])\nmodel.compile(\n    optimizer='adam',\n    loss=['sparse_categorical_crossentropy', 'mean_squared_error'],\n    metrics=['sparse_categorical_accuracy','accuracy']\n    )\n\nhistory = model.fit(X, [ class_index, y ], epochs=10, batch_size=64)\n<\/code><\/pre>\n\n<p>The error that I get is:<\/p>\n\n<pre><code>  File \"C:\/Users\/...\/Sin t\u00edtulo5.py\", line 31, in &lt;module&gt;\n    history = model.fit(X, [ class_index, y ], epochs=10, batch_size=64)\n\n  File \"C:\\Users\\AIRFI\\Anaconda3\\envs\\tf\\lib\\site-packages\\keras\\engine\\training.py\", line 1178, in fit\n    validation_freq=validation_freq)\n\n  File \"C:\\Users\\AIRFI\\Anaconda3\\envs\\tf\\lib\\site-packages\\keras\\engine\\training_arrays.py\", line 204, in fit_loop\n    outs = fit_function(ins_batch)\n\n  File \"C:\\Users\\AIRFI\\Anaconda3\\envs\\tf\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py\", line 2979, in __call__\n    return self._call(inputs)\n\n  File \"C:\\Users\\AIRFI\\Anaconda3\\envs\\tf\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py\", line 2937, in _call\n    fetched = self._callable_fn(*array_vals)\n\n  File \"C:\\Users\\AIRFI\\Anaconda3\\envs\\tf\\lib\\site-packages\\tensorflow_core\\python\\client\\session.py\", line 1472, in __call__\n    run_metadata_ptr)\n\nInvalidArgumentError: 2 root error(s) found.\n  (0) Invalid argument: Can not squeeze dim[1], expected a dimension of 1, got 2\n     [[{{node metrics_35\/sparse_categorical_accuracy_1\/Squeeze}}]]\n     [[metrics_35\/acc_1\/Mean\/_567]]\n  (1) Invalid argument: Can not squeeze dim[1], expected a dimension of 1, got 2\n     [[{{node metrics_35\/sparse_categorical_accuracy_1\/Squeeze}}]]\n0 successful operations.\n0 derived errors ignored.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58032114,"AnswerCount":1,"Tags":"<python><tensorflow><jupyter-notebook>","CreationDate":"2019-09-20T16:27:47.417","AcceptedAnswerId":null,"OwnerUserId":11927380.0,"Title":"How to resolve ValueError: Cannot feed value of shape (230,) for Tensor 'Placeholder_1:0', which has shape '(?, 4) in python using Jupyter notebook","Body":"<p>I can't seem to resolve the issue as I am new to tensorflow I think the issue is the graph mismatching\n   but cant resolve it please help. I want to resolve this as i will be using this for android app.<\/p>\n\n<p><strong>Defining placeholders:<\/strong><\/p>\n\n<pre><code>x = tf.placeholder(tf.float32,shape=[None,80,80,3])\ny_true = tf.placeholder(tf.float32,shape=[None,4])\n<\/code><\/pre>\n\n<p><strong>CNN Layer:<\/strong><\/p>\n\n<pre><code>%%time\n#Changing settings for GPU running.\nconfig = tf.ConfigProto(allow_soft_placement=True)\nconfig.gpu_options.allow_growth = True\nconfig.gpu_options.allocator_type = 'BFC'\n#Training and saving the result\nwith tf.Session(config=config) as sess:\n    sess.run(init)\n    summary_writer = tf.summary.FileWriter(TRAIN_DIR, graph=tf.get_default_graph())\n    for i in range(epochs):\n        for j in range(0,steps,step_size):            \n            _ , c , summary,d = sess.run([train,cross_entropy,merged_summary_op,acc],feed_dict={x:X[j:j+step_size] , y_true:Y[j:j+step_size] ,hold_prob1:0.5,hold_prob2:0.5,hold_prob3:0.5,hold_prob4:0.5})\n            summary_writer.add_summary(summary, i * total_batch + j)\n            acc_train.append(d)\n            mean_of_cross_entropy = sess.run(cross_entropy,feed_dict={x:cv_x,y_true:cv_y ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0})\n            mean_of_acc = sess.run(acc,feed_dict={x:cv_x ,y_true:cv_y ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0})\n            cross_entropy_list.append(mean_of_cross_entropy)\n            acc_list.append(mean_of_acc)\n        print(i,mean_of_cross_entropy,mean_of_acc)\n    saver.save(sess, \"C:\\\\Users\\\\blessie\\\\Desktop\\DATASETS - TOMATO LEAF DISEASE\\\\Models\\\\CNN_MC.ckpt\")\n    print(\"test accuracy = \",np.mean([sess.run(acc,feed_dict={x:test_x[:230],y_true:test_y[:230] ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0}),sess.run(acc,feed_dict={x:test_x[230:460],y_true:test_y[230:460] ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0}),sess.run(acc,feed_dict={x:test_x[460:],y_true:test_y[460:] ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0})]))   \n    print(\"cross_entropy loss = \",np.mean([sess.run(cross_entropy,feed_dict={x:test_x[:230],y_true:test_y[:230] ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0}),sess.run(cross_entropy,feed_dict={x:test_x[230:460],y_true:test_y[230:460] ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0}),sess.run(cross_entropy,feed_dict={x:test_x[460:],y_true:test_y[460:] ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0})]))\n<\/code><\/pre>\n\n<p>After running the code and pre-processing the data, I got a value error stating: <\/p>\n\n<p><strong>ValueError: Cannot feed value of shape (230,) for Tensor 'Placeholder_1:0', which has shape '(?, 4)'<\/strong><\/p>\n\n<p>How can I resolve this problem\/error?<\/p>\n","answers":[{"AnswerId":"58048772","CreationDate":"2019-09-22T11:30:10.220","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>You haven't given the full code, but from the error message, it appears that your <code>y_true<\/code> placeholder is being supplied with data of shape <code>[230]<\/code> which is basically of type <code>[None]<\/code>. So changing your <code>y_true<\/code> to should solve this particular problem:<\/p>\n\n<pre><code>y_true = tf.placeholder(tf.float32,shape=[None])\n<\/code><\/pre>\n\n<p>But you have used <code>y_true<\/code> previously in your code already. So, is the shape of <code>cv_y<\/code> correctly being fed as <code>[None, 4]<\/code>. <\/p>\n\n<p>Other particular thing what I can guess is you are supplying one hot encoded values to <code>cv_y<\/code> already but in the subsequent lines of code you are not. So, you can change your code as below:<\/p>\n\n<pre><code>y_true = tf.placeholder(tf.float32,shape=[None])\ny_true = tf.one_hot(y_true, 4)\n<\/code><\/pre>\n\n<p>And now, you will have to change your <code>cv_y<\/code> just before you are passing it in <code>feed_dict<\/code> as <\/p>\n\n<pre><code>cv_y = np.argmax(cv_y, axis=-1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58032270,"AnswerCount":0,"Tags":"<python><tensorflow><google-colaboratory>","CreationDate":"2019-09-20T16:40:33.223","AcceptedAnswerId":null,"OwnerUserId":12096615.0,"Title":"How to fix tf.flags ERROR on google colab","Body":"<p>My code runs without error on my system, but gives this error on google colab...<\/p>\n\n<h1>How do fix this problem???<\/h1>\n\n<p>code:<\/p>\n\n<pre><code>flags = tf.flags\n\nflags.DEFINE_string(\"mode\", \"train\", \"train\/debug\/test\")\n\nflags.DEFINE_string(\"target_dir\", target_dir, \"\")\nflags.DEFINE_string(\"log_dir\", log_dir, \"\")\nflags.DEFINE_string(\"save_dir\", save_dir, \"\")\nflags.DEFINE_string(\"train_file\", train_file, \"\")\nflags.DEFINE_string(\"dev_file\", dev_file, \"\")\n.\n.\n.\n<\/code><\/pre>\n\n<p>==========\nERROR:<\/p>\n\n<pre><code>DuplicateFlagError                        Traceback (most recent call last)\n\n&lt;ipython-input-5-52773f56b9d0&gt; in &lt;module&gt;()\n     48     os.makedirs(answer_dir)\n     49 \n---&gt; 50 flags.DEFINE_string(\"mode\", 'train', \"prepro\/train\/debug\/test\")\n     51 \n     52 flags.DEFINE_string(\"target_dir\", target_dir, \"\")\n\n4 frames\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/absl\/flags\/_flagvalues.py in __setitem__(self, name, flag)\n    428         # module is simply being imported a subsequent time.\n    429         return\n--&gt; 430       raise _exceptions.DuplicateFlagError.from_flag(name, self)\n    431     short_name = flag.short_name\n    432     # If a new flag overrides an old one, we need to cleanup the old flag's\n\nDuplicateFlagError: The flag 'mode' is defined twice. First from \/usr\/local\/lib\/python3.6\/dist-packages\/ipykernel_launcher.py, Second from \/usr\/local\/lib\/python3.6\/dist-packages\/ipykernel_launcher.py.  Description from first occurrence: prepro\/train\/debug\/test\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58032882,"AnswerCount":3,"Tags":"<python><machine-learning><image-processing><keras><conv-neural-network>","CreationDate":"2019-09-20T17:34:19.247","AcceptedAnswerId":58033078.0,"OwnerUserId":12096782.0,"Title":"Keras CNN high accuracy but bad predictions. How to improve it","Body":"<p>I have to identify a retinal disease with CNN. I have 1400 images, 700 each class. My classes are (0 - no PDR) and (1 - PDR). I'm trying to make a model to identify if an input retina have the disease in level 4 or not. <\/p>\n\n<p>I'm making a follow manipulation with my images and rezise all to 256x256:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>ImageCV[index] = cv2.addWeighted(ImageCV[index],4, cv2.GaussianBlur(ImageCV[index],(0,0), 256\/30), -4, 128)\n<\/code><\/pre>\n\n<p>And it made the follow with my imgs: \n<a href=\"https:\/\/imgur.com\/X1p9G1c\" rel=\"nofollow noreferrer\">https:\/\/imgur.com\/X1p9G1c<\/a><\/p>\n\n<p>Then, when I train my model I got a very high accuracy (like 99....) but when I try to predict some test images, it fails.. for example, I've putted 10 PDR examples in test folder and tries to predict them (all must be 1).. this is the result:<\/p>\n\n<pre><code>[[0.]]\n[[0.]]\n[[1.]]\n[[0.]]\n[[0.]]\n[[0.]]\n[[1.]]\n[[0.]]\n[[0.]]\n[[0.]]\n<\/code><\/pre>\n\n<p>This is my model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>visible = Input(shape=(256,256,3))\nconv1 = Conv2D(16, kernel_size=(3,3), activation='relu', strides=(1, 1))(visible)\nconv2 = Conv2D(16, kernel_size=(3,3), activation='relu', strides=(1, 1))(conv1)\nbat1 = BatchNormalization()(conv2)\nconv3 = ZeroPadding2D(padding=(1, 1))(bat1)\npool1 = MaxPooling2D(pool_size=(2, 2))(conv3)\n\nconv4 = Conv2D(32, kernel_size=(3,3), activation='relu', padding='valid', kernel_regularizer=regularizers.l2(0.01))(pool1)\nconv5 = Conv2D(32, kernel_size=(3,3), activation='relu', padding='valid', kernel_regularizer=regularizers.l2(0.01))(conv4)\nbat2 = BatchNormalization()(conv5)\npool2 = MaxPooling2D(pool_size=(1, 1))(bat2)\n\nconv6 = Conv2D(64, kernel_size=(3,3), activation='relu',strides=(1, 1), padding='valid')(pool2)\nconv7 = Conv2D(64, kernel_size=(3,3), activation='relu',strides=(1, 1), padding='valid')(conv6)\nbat3 = BatchNormalization()(conv7)\nconv7 = ZeroPadding2D(padding=(1, 1))(bat3)\npool3 = MaxPooling2D(pool_size=(1, 1))(conv7)\n\nconv8 = Conv2D(128, kernel_size=(3,3), activation='relu', padding='valid', kernel_regularizer=regularizers.l2(0.01))(pool3)\nconv9 = Conv2D(128, kernel_size=(2,2), activation='relu', strides=(1, 1), padding='valid')(conv8)\nbat4 = BatchNormalization()(conv9)\npool4 = MaxPooling2D(pool_size=(1, 1))(bat4)\n\nflat = Flatten()(pool4)\n\noutput = Dense(1, activation='sigmoid')(flat)\nmodel = Model(inputs=visible, outputs=output)\n\nopt = optimizers.adam(lr=0.001, decay=0.0)\n\nmodel.compile(optimizer= opt, loss='binary_crossentropy', metrics=['accuracy'])\n\n\ndata, labels = ReadImages(TRAIN_DIR)\n\ntest, lt = ReadImages(TEST_DIR)\n\ndata = np.array(data)\nlabels = np.array(labels)\ntest = np.array(test)\nlt = np.array(lt)\n\nnp.random.permutation(len(data))\nnp.random.permutation(len(labels))\nnp.random.permutation(len(test))\nnp.random.permutation(len(lt))\n\nmodel.fit(data, labels, epochs=7, validation_data = (test,lt))\n\nmodel.save('model.h5')\n<\/code><\/pre>\n\n<p>And this is predict.py<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = load_model('model.h5')\n\nfor filename in os.listdir(r'v\/'):\n    if filename.endswith(\".jpg\") or filename.endswith(\".ppm\") or filename.endswith(\".jpeg\"):\n        ImageCV = cv2.resize(cv2.imread(os.path.join(TEST_DIR) + filename), (256,256))\n        ImageCV = cv2.addWeighted(ImageCV,4, cv2.GaussianBlur(ImageCV,(0,0), 256\/30), -4, 128)\n        cv2.imshow('image', ImageCV)\n        cv2.waitKey(0)\n        cv2.destroyAllWindows()\n        ImageCV = ImageCV.reshape(-1,256,256,3)\n        print(model.predict(ImageCV))\n<\/code><\/pre>\n\n<p>What do I could to do to improve my predictions at all? \nI strongly appreciate your help<\/p>\n\n<p><strong>UPDATE<\/strong> \nWell, I tried to do all of were said in answers but still'nt working...\nthis is my code now:<\/p>\n\n<pre><code>visible = Input(shape=(256,256,3))\nconv1 = Conv2D(16, kernel_size=(3,3), activation='relu', strides=(1, 1))(visible)\nconv2 = Conv2D(32, kernel_size=(3,3), activation='relu', strides=(1, 1))(conv1)\nbat1 = BatchNormalization()(conv2)\nconv3 = ZeroPadding2D(padding=(1, 1))(bat1)\npool1 = MaxPooling2D(pool_size=(2, 2))(conv3)\ndrop1 = Dropout(0.30)(pool1)\n\nconv4 = Conv2D(32, kernel_size=(3,3), activation='relu', padding='valid', kernel_regularizer=regularizers.l2(0.01))(drop1)\nconv5 = Conv2D(64, kernel_size=(3,3), activation='relu', padding='valid', kernel_regularizer=regularizers.l2(0.01))(conv4)\nbat2 = BatchNormalization()(conv5)\npool2 = MaxPooling2D(pool_size=(1, 1))(bat2)\ndrop1 = Dropout(0.30)(pool2)\n\nconv6 = Conv2D(128, kernel_size=(3,3), activation='relu', padding='valid', kernel_regularizer=regularizers.l2(0.01))(pool2)\nconv7 = Conv2D(128, kernel_size=(2,2), activation='relu', strides=(1, 1), padding='valid')(conv6)\nbat3 = BatchNormalization()(conv7)\npool3 = MaxPooling2D(pool_size=(1, 1))(bat3)\ndrop1 = Dropout(0.30)(pool3)\n\nflat = Flatten()(pool3)\ndrop4 = Dropout(0.50)(flat)\n\noutput = Dense(1, activation='sigmoid')(drop4)\nmodel = Model(inputs=visible, outputs=output)\n\nopt = optimizers.adam(lr=0.001, decay=0.0)\n\nmodel.compile(optimizer= opt, loss='binary_crossentropy', metrics=['accuracy'])\n\ndata, labels = ReadImages(TRAIN_DIR)\ntest, lt = ReadImages(TEST_DIR)\n\ndata = np.array(data)\nlabels = np.array(labels)\n\nperm = np.random.permutation(len(data))\ndata = data[perm]\nlabels = labels[perm]\n#model.fit(data, labels, epochs=8, validation_data = (np.array(test), np.array(lt)))\n\naug = ImageDataGenerator(rotation_range=20, width_shift_range=0.2, height_shift_range=0.2, shear_range=0.15,\n    horizontal_flip=True)\n\n# train the network\nmodel.fit_generator(aug.flow(data, labels, batch_size=32),\n    validation_data=(np.array(test), np.array(lt)), steps_per_epoch=len(data) \/\/ 32,\n    epochs=7)\n<\/code><\/pre>\n\n<p>And this is the return: <\/p>\n\n<pre><code>Epoch 1\/7\n43\/43 [==============================] - 1004s 23s\/step - loss: 1.8090 - acc: 0.9724 - val_loss: 1.7871 - val_acc: 0.9861\nEpoch 2\/7\n43\/43 [==============================] - 1003s 23s\/step - loss: 1.8449 - acc: 0.9801 - val_loss: 1.4828 - val_acc: 1.0000\nEpoch 3\/7\n43\/43 [==============================] - 1092s 25s\/step - loss: 1.5704 - acc: 0.9920 - val_loss: 1.3985 - val_acc: 1.0000\nEpoch 4\/7\n43\/43 [==============================] - 1062s 25s\/step - loss: 1.5219 - acc: 0.9898 - val_loss: 1.3167 - val_acc: 1.0000\nEpoch 5\/7\n43\/43 [==============================] - 990s 23s\/step - loss: 2.5744 - acc: 0.9222 - val_loss: 2.9347 - val_acc: 0.9028\nEpoch 6\/7\n43\/43 [==============================] - 983s 23s\/step - loss: 1.6053 - acc: 0.9840 - val_loss: 1.3299 - val_acc: 1.0000\nEpoch 7\/7\n43\/43 [==============================] - 974s 23s\/step - loss: 1.6180 - acc: 0.9801 - val_loss: 1.5181 - val_acc: 0.9861\n<\/code><\/pre>\n\n<p>I'd have added dropouts, reduce the model layers, put data augmentation, and doesn't work at all (all the predictions returns 0)...<\/p>\n\n<p>Please anyone can help in it.<\/p>\n","answers":[{"AnswerId":"58033078","CreationDate":"2019-09-20T17:50:54.113","ParentId":null,"OwnerUserId":"11897007","Title":null,"Body":"<p>It seems you having issues with overfitting. I have kind of a personal dilemma in here whether this is off-topic or not because the approach one can give it's somewhat opinion-based, but here I go:\nFirst off, if you need to regularize an overfitting network, you want to use <code>dropout<\/code> starting at <code>0.25<\/code> and checking if that improves the model. <code>Data augmentation<\/code> is a must when dealing with overfitting, together with <code>batch normalization<\/code> (which you are applying).\nIf this still not solves your overfitting problems, then you should try working on your network architecture to tweak it in order to make it generalize better. Did you check a simple of the inputs being used for train and test? <\/p>\n\n<p>TLDR: Try dropout and data augmentation, if it doesn't work and your data is correct, you might have to work on improving the architecture to make a better generalizing model.<\/p>\n\n<p>EDIT: The consensus on approaching this kind of models is to have it overfit at first, with a decent accuracy and then work towards generalizing it without losing accuracy if possible.<\/p>\n"},{"AnswerId":"58033201","CreationDate":"2019-09-20T18:00:11.643","ParentId":null,"OwnerUserId":"2409563","Title":null,"Body":"<p>Your model is very big for your training set size, leading to overfitting. Use less layers<\/p>\n"},{"AnswerId":"58045344","CreationDate":"2019-09-22T01:21:24.350","ParentId":null,"OwnerUserId":"6417245","Title":null,"Body":"<p>You're overfitting i.e. have a high variance problem. Try adding some soft dropout (0.2-0.3) at the end of your convolution blocks. You can also add a couple of Dense layers with decreasing unit numbers before the output layer and put more dropout layers in between those (0.5+).<\/p>\n\n<p>You should also implement more data augmentation such as rotations, flips, random noise, random brightness, etc. Check out Keras' documentation for the ImageDataGenerator class.<\/p>\n"}]}
+{"QuestionId":58033171,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras><reproducible-research><hyperopt>","CreationDate":"2019-09-20T17:58:07.773","AcceptedAnswerId":null,"OwnerUserId":3198674.0,"Title":"How to reproduce the MLP results with the best set selected via hyperopt","Body":"<p>I am trying to select the best hyperparameter set for my MLP models in a 4 fold CV using hyperopt. As seen in the code below, the averaged rsq is trying to be maximized. The data set contains time series data related to stock market. Moreover, rsq is not the only metric that I am interested, that is, I am also trying to get a positive correlation with a possible smallest rsq (-0.02).<\/p>\n\n<p>With hyperopt, I obtained the best parameters leading to the smallest rsq with a correlation of 0.26 which was quite acceptable for me. However, when I repeat the model make with the selected parameters I got quite different results. Please see the code below:<\/p>\n\n<pre><code>space = {\n\n    'act': hp.choice('act', ['tanh','relu']),\n    'batch' : hp.choice('batch', [32,64,128]),\n    'dropout': hp.choice('dropout', [0, 0.25, 0.5, 0.75]),\n    'epochs' :  hp.choice('epochs', [10, 20, 50, 100]),\n    'hidden': hp.choice('hidden', [1, 2, 3, 4]),\n    'l2': hp.choice('l2', [0.1, 0.01, 0.001, 0.0001]),\n    'lr': hp.choice('lr', [0.1,0.01,0.001,0.0001]),\n    'neurons': hp.choice('neurons', [1, 2, 4, 6]),\n    'opt': hp.choice('opt', ['adam','rmsprop'])\n\n}\n\ndef objective(params):\n    from keras.regularizers import l1,l2\n\n    import keras\n    from keras import backend as KB\n    import tensorflow as tf\n    from sklearn.metrics import r2_score,mean_squared_error\n    from keras.callbacks import EarlyStopping\n    from keras.layers.normalization import BatchNormalization\n    from keras.layers import GaussianNoise\n\n    dfResultsCV = pd.DataFrame()\n\n    mseList = []\n    corrList = []\n    rsqList = []\n\n    print(mseList)\n\n    for i in range(0,4):\n\n        if i == 0:\n            trainCV = pd.concat([train[train['fold'] == 0], train[train['fold'] == 1], train[train['fold'] == 2]], axis=0, ignore_index=True)\n            testCV = train[train['fold'] == 3]\n            fold = 4\n            sel_colsCV = CVcols[0]\n        elif i == 1:\n            trainCV = pd.concat([train[train['fold'] == 0], train[train['fold'] == 1], train[train['fold'] == 3]], axis=0, ignore_index=True)\n            testCV = train[train['fold'] == 2]\n            fold = 3\n            sel_colsCV = CVcols[1]\n        elif i == 2:\n            trainCV = pd.concat([train[train['fold'] == 0], train[train['fold'] == 2], train[train['fold'] == 3]], axis=0, ignore_index=True)\n            testCV = train[train['fold'] == 1]\n            fold = 2\n            sel_colsCV = CVcols[2]\n        else:\n            trainCV = pd.concat([train[train['fold'] == 1], train[train['fold'] == 2], train[train['fold'] == 3]], axis=0, ignore_index=True)\n            testCV = train[train['fold'] == 0]\n            fold = 1\n            sel_colsCV = CVcols[3]\n\n\n        ### Scale the trainCV data\n        ###\n        trainCVsca = trainCV.copy()\n        testCVsca = testCV.copy()\n\n        trainCVsca[sel_colsCV] = scaler.fit_transform(trainCV[sel_colsCV])\n        testCVsca[sel_colsCV] = scaler.transform(testCV[sel_colsCV])\n\n        ###\n        ### Scale the trainCV data\n\n        np.random.seed(0)\n        tf.set_random_seed(0)\n        sess = tf.Session(graph=tf.get_default_graph())\n        KB.set_session(sess)\n\n        early_stop = EarlyStopping(monitor='val_loss', min_delta=0.00001, patience=5, verbose=2, mode='auto')\n        MLPmodel = Sequential()\n        MLPmodel.add(GaussianNoise(0.001, input_shape=(len(sel_colsCV),)))\n        MLPmodel.add(Dense(params['neurons'], input_dim=trainCVsca[sel_colsCV].shape[1], activation=params['act'], \n                           activity_regularizer=l2(params['l2'])))\n\n        for j in range(0,params['hidden']):\n\n            MLPmodel.add(Dense(params['neurons'], activation=params['act'], activity_regularizer=l2(params['l2'])))\n            MLPmodel.add(BatchNormalization())\n            MLPmodel.add(Dropout(params['dropout']))\n\n        MLPmodel.add(Dense(1, activation='linear'))\n\n        if params['opt'] == 'adam':\n            optim = keras.optimizers.Adam(lr=params['lr'])\n            MLPmodel.compile(loss='mse', optimizer=params['opt'])\n        else:\n            optim = keras.optimizers.RMSprop(lr=params['lr'])\n            MLPmodel.compile(loss='mse', optimizer=params['opt'])\n\n        MLPmodel.fit(trainCVsca[sel_colsCV], trainCVsca[target], batch_size=params['batch'], epochs=params['epochs'], \n                                validation_data=(testCVsca[sel_colsCV],testCVsca[target]),callbacks=[early_stop],verbose=2)\n\n        mse = mean_squared_error(testCVsca[target], MLPmodel.predict(testCVsca[sel_colsCV]))\n        rsq = r2_score(testCVsca[target], MLPmodel.predict(testCVsca[sel_colsCV]))\n        corr = np.corrcoef(pd.DataFrame(testCVsca[target])['IT_3M_REL_RETURN'],pd.DataFrame((MLPmodel.predict(testCVsca[sel_colsCV]).reshape(-1,1)))[0])[0][1]\n\n        mseList.append(mse)\n        rsqList.append(rsq)\n        corrList.append(corr)\n\n        del MLPmodel\n        KB.clear_session()\n\n    AvgMse = np.average(mseList)\n    AvgRsq = np.average(rsqList)\n    AvgCorr = np.average(corrList)\n\n    print(\"Average Mse is: \", AvgMse)\n    print(\"Average Rsq is: \", AvgRsq)\n    print(\"Average Corr is: \", AvgCorr)\n\n    score = -AvgRsq\n\n    return {'loss': score, 'status': STATUS_OK}\n\ntrials = Trials()\n\nbest = fmin(objective, space, algo=tpe.suggest, trials=trials, max_evals=500, rstate=np.random.RandomState(23))\n<\/code><\/pre>\n\n<p>When I run the model with the parameters:<\/p>\n\n<pre><code>act = space_eval(space, best)['act']\nbatch = space_eval(space, best)['batch']\ndropout= space_eval(space, best)['dropout']\nepoch = space_eval(space, best)['epochs']\nhidden = space_eval(space, best)['hidden']\nl2_reg = space_eval(space, best)['l2']\nlr = space_eval(space, best)['lr']\nneurons = space_eval(space, best)['neurons']\nopt = space_eval(space, best)['opt']\n<\/code><\/pre>\n\n<p>I got an rsq of -2.02 with a correlation value of 0.155 which are very different from the results hyperopt reached. What might be the reason?<\/p>\n","answers":[]}
+{"QuestionId":58033286,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><tensorflow2.0><attention-model>","CreationDate":"2019-09-20T18:07:02.813","AcceptedAnswerId":null,"OwnerUserId":8382950.0,"Title":"Tensorflow @tf.function: AttributeError: in converted code","Body":"<p>I created a class and defined the train_step function inside it: TF tutorial: <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/text\/nmt_with_attention\" rel=\"nofollow noreferrer\">NMT_attention<\/a> \nWithout using the @tf.function significantly increases the training time. On defining it, I get a conversion error for the private variables declared inside the class.<\/p>\n\n<pre><code>@tf.function\n    def train_step(self, input, target, encoderHidden):\n      loss = 0\n\n      with tf.GradientTape() as tape:\n        encoderOutput, encoderHidden = self.__encoder(input, encoderHidden) #throws error\n<\/code><\/pre>\n\n<p>Below is the traceback:<\/p>\n\n<pre><code>    Using TensorFlow backend.\n        2019-09-20 12:54:32.676302: I tensorflow\/core\/platform\/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\n        Traceback (most recent call last):\n          File \"\/Users\/Users\/Library\/Python\/lib\/python\/site-packages\/proj\/Models\/attention_model.py\", line 499, in &lt;module&gt;\n            model.fit(path, epochs)\n          File \"\/Users\/Users\/Library\/Python\/lib\/python\/site-packages\/proj\/Models\/attention_model.py\", line 383, in fit\n            loss = self.train_step(input, target, encoderHidden)\n          File \"\/Users\/Users\/work_env\/lib\/python3.7\/site-packages\/tensorflow\/python\/eager\/def_function.py\", line 416, in __call__\n            self._initialize(args, kwds, add_initializers_to=initializer_map)\n          File \"\/Users\/Users\/work_env\/lib\/python3.7\/site-packages\/tensorflow\/python\/eager\/def_function.py\", line 359, in _initialize\n            *args, **kwds))\n          File \"\/Users\/Users\/work_env\/lib\/python3.7\/site-packages\/tensorflow\/python\/eager\/function.py\", line 1360, in _get_concrete_function_internal_garbage_collected\n            graph_function, _, _ = self._maybe_define_function(args, kwargs)\n          File \"\/Users\/Users\/work_env\/lib\/python3.7\/site-packages\/tensorflow\/python\/eager\/function.py\", line 1648, in _maybe_define_function\n            graph_function = self._create_graph_function(args, kwargs)\n          File \"\/Users\/Users\/work_env\/lib\/python3.7\/site-packages\/tensorflow\/python\/eager\/function.py\", line 1541, in _create_graph_function\n            capture_by_value=self._capture_by_value),\n          File \"\/Users\/Users\/work_env\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/func_graph.py\", line 716, in func_graph_from_py_func\n            func_outputs = python_func(*func_args, **func_kwargs)\n          File \"\/Users\/Users\/work_env\/lib\/python3.7\/site-packages\/tensorflow\/python\/eager\/def_function.py\", line 309, in wrapped_fn\n            return weak_wrapped_fn().__wrapped__(*args, **kwds)\n          File \"\/Users\/Users\/work_env\/lib\/python3.7\/site-packages\/tensorflow\/python\/eager\/function.py\", line 2155, in bound_method_wrapper\n            return wrapped_fn(*args, **kwargs)\n          File \"\/Users\/Users\/work_env\/lib\/python3.7\/site-packages\/tensorflow\/python\/framework\/func_graph.py\", line 706, in wrapper\n            raise e.ag_error_metadata.to_exception(type(e))\n        AttributeError: in converted code:\n            relative to \/Users\/Users:\n\n            Library\/Python\/lib\/python\/site-packages\/proj\/Models\/attention_model.py:262 train_step  *\n                encoderOutput, encoderHidden = self.__encoder(input, encoderHidden)\n            work_env\/lib\/python3.7\/site-packages\/tensorflow\/python\/autograph\/impl\/api.py:329 converted_call\n                f = getattr(owner, f)\n\n            AttributeError: 'Model' object has no attribute '__encoder'\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58033895,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-20T19:01:32.100","AcceptedAnswerId":null,"OwnerUserId":12097102.0,"Title":"Extracting the dropout mask from a keras dropout layer?","Body":"<p>I would like to extract and store the dropout mask [array of 1\/0s] from a dropout layer in a Sequential Keras model at each batch while training. I was wondering if there was a straight forward way way to do this within Keras or if I would need to switch over to tensorflow (<a href=\"https:\/\/stackoverflow.com\/questions\/37463863\/how-to-get-the-dropout-mask-in-tensorflow\">How to get the dropout mask in Tensorflow<\/a>). <\/p>\n\n<p>Would appreciate any help! I'm quite new to TensorFlow and Keras. <\/p>\n\n<p>There are a couple of functions (dropout_layer.get_output_mask(), dropout_layer.get_input_mask()) for the dropout layer that I tried using but got <code>None<\/code> after calling on the previous layer.  <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = tf.keras.Sequential()\nmodel.add(tf.keras.layers.Flatten(name=\"flat\", input_shape=(28, 28, 1)))\nmodel.add(tf.keras.layers.Dense(\n    512,\n    activation='relu',\n    name = 'dense_1',\n    kernel_initializer=tf.keras.initializers.GlorotUniform(seed=123),\n    bias_initializer='zeros'))\ndropout = tf.keras.layers.Dropout(0.2, name = 'dropout') #want this layer's mask\n\nmodel.add(dropout)\nx = dropout.output_mask\ny = dropout.input_mask\nmodel.add(tf.keras.layers.Dense(\n    10,\n    activation='softmax',\n    name='dense_2',\n    kernel_initializer=tf.keras.initializers.GlorotUniform(seed=123),\n    bias_initializer='zeros'))\n\nmodel.compile(...)\nmodel.fit(...)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58649330","CreationDate":"2019-10-31T18:27:42.343","ParentId":null,"OwnerUserId":"12097102","Title":null,"Body":"<p>I found a very hacky way to do this by trivially extending the provided dropout layer. (Almost all code from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/cc47b982ab00bff3080ad3ad5e8c7e3a365391b4\/tensorflow\/python\/ops\/nn_ops.py\" rel=\"nofollow noreferrer\">TF<\/a>.)<\/p>\n\n<pre><code>class MyDR(tf.keras.layers.Layer):\ndef __init__(self,rate,**kwargs):\n    super(MyDR, self).__init__(**kwargs)\n\n    self.noise_shape = None\n    self.rate = rate\n\n\ndef _get_noise_shape(self,x, noise_shape=None):\n    # If noise_shape is none return immediately.\n    if noise_shape is None:\n        return array_ops.shape(x)\n    try:\n        # Best effort to figure out the intended shape.\n        # If not possible, let the op to handle it.\n        # In eager mode exception will show up.\n        noise_shape_ = tensor_shape.as_shape(noise_shape)\n    except (TypeError, ValueError):\n        return noise_shape\n\n    if x.shape.dims is not None and len(x.shape.dims) == len(noise_shape_.dims):\n        new_dims = []\n        for i, dim in enumerate(x.shape.dims):\n            if noise_shape_.dims[i].value is None and dim.value is not None:\n                new_dims.append(dim.value)\n            else:\n                new_dims.append(noise_shape_.dims[i].value)\n        return tensor_shape.TensorShape(new_dims)\n\n    return noise_shape\n\ndef build(self, input_shape):\n    self.noise_shape = input_shape\n    print(self.noise_shape)\n    super(MyDR,self).build(input_shape)\n\n@tf.function\ndef call(self,input):\n    self.noise_shape = self._get_noise_shape(input)\n    random_tensor = tf.random.uniform(self.noise_shape, seed=1235, dtype=input.dtype)\n    keep_prob = 1 - self.rate\n    scale = 1 \/ keep_prob\n    # NOTE: if (1.0 + rate) - 1 is equal to rate, then we want to consider that\n    # float to be selected, hence we use a &gt;= comparison.\n    self.keep_mask = random_tensor &gt;= self.rate\n    #NOTE: here is where I save the binary masks. \n    #the file grows quite big!\n    tf.print(self.keep_mask,output_stream=\"file:\/\/temp\/droput_mask.txt\")\n\n    ret = input * scale * math_ops.cast(self.keep_mask, input.dtype)\n    return ret\n<\/code><\/pre>\n"},{"AnswerId":"58034527","CreationDate":"2019-09-20T20:03:28.327","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>It's not easily exposed in Keras. It goes deep until it calls the Tensorflow dropout. <\/p>\n\n<p>So, although you're using Keras, it's will also be a tensor in the graph that can be gotten by name (finding it's name: <a href=\"https:\/\/stackoverflow.com\/questions\/36883949\/in-tensorflow-get-the-names-of-all-the-tensors-in-a-graph\">In Tensorflow, get the names of all the Tensors in a graph<\/a>). <\/p>\n\n<p>This option, of course will lack some keras information, you should probably have to do that inside a Lambda layer so Keras adds certain information to the tensor. And you must take extra care because the tensor will exist even when not training (where the mask is skipped)   <\/p>\n\n<p>Now, you can also use a less hacky way, that may consume a little processing:<\/p>\n\n<pre><code>def getMask(x):\n    boolMask = tf.not_equal(x, 0)\n    floatMask = tf.cast(boolMask, tf.float32) #or tf.float64\n    return floatMask\n<\/code><\/pre>\n\n<p>Use a <code>Lambda(getMasc)(output_of_dropout_layer)<\/code>    <\/p>\n\n<p>But instead of using a <code>Sequential<\/code> model, you will need a functional API <code>Model<\/code>.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>inputs = tf.keras.layers.Input((28, 28, 1))\noutputs = tf.keras.layers.Flatten(name=\"flat\")(inputs)\noutputs = tf.keras.layers.Dense(\n    512,\n    #    activation='relu', #relu will be a problem here\n    name = 'dense_1',\n    kernel_initializer=tf.keras.initializers.GlorotUniform(seed=123),\n    bias_initializer='zeros')(outputs)\n\noutputs = tf.keras.layers.Dropout(0.2, name = 'dropout')(outputs)\nmask = Lambda(getMask)(outputs)\n#there isn't \"input_mask\"\n\n\n#add the missing relu: \noutputs = tf.keras.layers.Activation('relu')(outputs)\noutputs = tf.keras.layers.Dense(\n    10,\n    activation='softmax',\n    name='dense_2',\n    kernel_initializer=tf.keras.initializers.GlorotUniform(seed=123),\n    bias_initializer='zeros')(outputs)\n\nmodel = Model(inputs, outputs)\nmodel.compile(...)\nmodel.fit(...)\n<\/code><\/pre>\n\n<h2>Training and predicting<\/h2>\n\n<p>Since you can't train the masks (it doesn't make any sense), it should not be an output of the model for training.   <\/p>\n\n<p>Now, we could try this:<\/p>\n\n<pre><code>trainingModel = Model(inputs, outputs)    \npredictingModel = Model(inputs, [output, mask])    \n<\/code><\/pre>\n\n<p>But masks don't exist in prediction, because dropout is only applied in training. So this doesn't bring us anything good in the end.    <\/p>\n\n<p>The only way for training is then using a dummy loss and dummy targets:    <\/p>\n\n<pre><code>def dummyLoss(y_true, y_pred):\n    return y_true #but this might evoke a \"None\" gradient problem since it's not trainable, there is no connection to any weights, etc.    \n\nmodel.compile(loss=[loss_for_main_output, dummyLoss], ....)\n\nmodel.fit(x_train, [y_train, np.zeros((len(y_Train),) + mask_shape), ...)\n<\/code><\/pre>\n\n<p>It's not guaranteed that these will work. <\/p>\n"}]}
+{"QuestionId":58034824,"AnswerCount":1,"Tags":"<python><keras><conv-neural-network>","CreationDate":"2019-09-20T20:31:21.117","AcceptedAnswerId":58039767.0,"OwnerUserId":12097353.0,"Title":"How to correctly define loss function in keras network that takes an image and outputs an image?","Body":"<p>I'm trying to build a model that takes grayscale image and generates another image as output by residual CNN using keras. And the key idea here is that the amount of non-white pixels is way smaller than the amount of white ones. So the error in each non-white pixel should have more weight in the loss function than in the white one. And the darker the pixel the bigger error should be. \nThat's how I've done it presently.\nHere, total_pred_score means all the correctly guessed pixels with their weights defined by some int coefficient and total_true_score means the total expected score for all images and pixels in the train set.\nBut I'm not sure it's right.Can you help me please?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def custom_loss(y_true, y_pred):\n    y_true1, y_pred1 = (255 - y_true) \/ 255, (255 - y_pred) \/ 255\n    dif = (1 - K.abs(y_true1 - y_pred1))\n    weight = (coeff - 1) * y_true1 + 1 \n    total_true_score = K.sum(weight, axis = [0,1,2,3])\n    total_pred_score = K.sum(multiply([dif,weight]), axis = [0,1,2,3])\n    return K.abs(total_true_score - total_pred_score) \/ total_true_score\n<\/code><\/pre>\n","answers":[{"AnswerId":"58039767","CreationDate":"2019-09-21T11:20:21.823","ParentId":null,"OwnerUserId":"11651301","Title":null,"Body":"<p>The loss function that you are trying to build should have two properties:<\/p>\n\n<ol>\n<li>Error in each non-white pixel should have more weight than in the white ones<\/li>\n<li>Darker the pixel the bigger error should be<\/li>\n<\/ol>\n\n<p>So to make it simpler you want your error to be more if you classify a dark pixel as a white one, but really not too much worried about the cases where you classify the white pixel as dark ones but you still want them to contribute to your loss function. Also, you want to make sure that your loss is proportional to the intensity of the pixel (darker the pixel more the error).<\/p>\n\n<p>I have come up with the following loss function which obeys these two properties:<\/p>\n\n<pre><code>def custom_loss(y_true, y_pred, coeff):\n    y_true1, y_pred1 = (255 - y_true) \/ 255, (255 - y_pred) \/ 255\n    dif = y_true1 - y_pred1\n    temp1 = K.l2_normalize(K.cast(K.greater(dif, 0),\"float32\")*y_true1,axis=-1) * coeff\n    temp2 = K.cast(K.less(dif, 0),\"float32\")\n    weight = temp1 + temp2  \n    loss = K.abs(weight*dif)\n    average_loss = K.mean(loss)   ##you need to return this when you use in your code\n    return K.eval(loss)\n<\/code><\/pre>\n\n<p>So basically we are finding the difference between true value and the predicted value after doing the normalization and multiplying weights proportional only to positive values in <code>dif<\/code> (which corresponds to darker pixels) -<code>temp1<\/code>. And then we just add a weight of <code>1<\/code> to pixels corresponding to negative values (which corresponds to white). <\/p>\n\n<blockquote>\n  <p>SAMPLE OUTPUT<\/p>\n<\/blockquote>\n\n<pre><code>coeff = 5\ny_t = np.array([0,250,30,0,255,160,210,0,2,4])\ny_p = np.array([50,0,80,10,255,160,210,250,2,80])\ncustom_loss(y_t,y_p,coeff ) \n\narray([0.449957, 0.98039216, 0.39702088, 0.08999141, 0., 0., 0., 2.249785, 0., 0.67320627],dtype=float32)\n<\/code><\/pre>\n\n<p>Here if you look at pixels at bit <code>2<\/code> and <code>8<\/code> carefully, pixel <code>8<\/code> corresponds to the case where we predict a dark pixel as a while pixel which is terribly bad in our case so the loss value is really high <code>2.249785<\/code>, likewise pixel <code>2<\/code> corresponds to the case where we predict a white pixel as a dark pixel which is kind of okay for us, so the loss value is <code>0.98039216<\/code>  which is not as high as the previous case. This corresponds to property <code>1<\/code>.<\/p>\n\n<p>Likewise, if you look as pixels <code>1<\/code> and <code>3<\/code> the magnitude of error in both the cases are the same (they are off by <code>50<\/code>) but pixel <code>1<\/code> is much darker than pixel <code>3<\/code> so we want the error on pixel <code>1<\/code> to be higher than pixel <code>3<\/code>, which is what we observe in the error vector. <\/p>\n\n<p>And in cases where our true and predicted values match, we have <code>0<\/code> error. <\/p>\n\n<p>Hope this helps!<\/p>\n"}]}
+{"QuestionId":58035125,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-20T20:59:46.727","AcceptedAnswerId":58035736.0,"OwnerUserId":6575669.0,"Title":"How to provide input layer to tf.layers.Conv3D?","Body":"<p>tf.layers have two classes for 3d convolution. 1) tf.layers.conv3d and 2) tf.layers.Conv3D<\/p>\n\n<p>According to the documentation, tf.layers.conv3d is going to be deprecated. So I started using tf.layers.Conv3D.<\/p>\n\n<p>But tf.layers.Conv3D doesn't have a parameter to specify the input layer where as tf.layers.conv3d has. <\/p>\n\n<p>Can someone help me figure out how to give inputs to tf.layers.Conv3D layer ?<\/p>\n","answers":[{"AnswerId":"58035736","CreationDate":"2019-09-20T22:22:02.573","ParentId":null,"OwnerUserId":"1540563","Title":null,"Body":"<p>You could rewrite your code in the following manner:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    # Old\n    output = tf.conv3d(input, *other_args)\n\n    # New\n    output = tf.keras.layers.Conv3D(*other_args)(input)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58036681,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow><keras>","CreationDate":"2019-09-21T01:48:04.093","AcceptedAnswerId":58037507.0,"OwnerUserId":6179333.0,"Title":"Integrating the loss of a keras model into a tensorflow graph","Body":"<p>If I define this simple Keras model<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras\nimport numpy as np\n\nl1 = keras.layers.Input(shape=(32))\nl2 = keras.layers.Dense(10)(l1)\nmodel = keras.Model(inputs=l1, outputs=l2)\n\nmodel.compile(loss='mse', optimizer='adam')\n<\/code><\/pre>\n\n<p>Let's say I have the input and labels values stored in <code>train_examples<\/code> and <code>train_labels<\/code> respectively<\/p>\n\n<p>If I also define a variable <code>some_var<\/code> that depends on that model's loss (I just use <code>model_loss<\/code> here for the sake of this example)<\/p>\n\n<pre><code>some_var = model.total_loss\n<\/code><\/pre>\n\n<p>How do I evaluate the value for <code>some_var<\/code>? I know it should be something like:<\/p>\n\n<pre><code>with keras.backend.get_session() as sess:\n    sess.run(some_var, feed_dict={ ?: train_examples, ?: train_labels })\n<\/code><\/pre>\n\n<p>what should go in place of the question marks?\nI don't want to modify the model's loss function, just use whatever has been defined in the definition of another variable<\/p>\n\n<p>thank you in advance<\/p>\n","answers":[{"AnswerId":"58037189","CreationDate":"2019-09-21T04:03:55.443","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Since you want to \"evaluate\" and not use it in some further graph calculations, you can simply use a callback:<\/p>\n\n<pre><code>from keras.callbacks import LambdaCallback\n\ndef getLoss(epoch, logs):\n    print(logs['loss']) #or val_loss (print the keys of logs if in doubt)\n\ncallback = LambdaCallback(on_epoch_end = getLoss)\n<\/code><\/pre>\n\n<p>Use this callback when fitting:<\/p>\n\n<pre><code>model.fit(x, y, callbacks = [callback])\n<\/code><\/pre>\n\n<p>If you can wait until training finishes, you can simply get the history at the end:<\/p>\n\n<pre><code>historyCallback = model.fit(x, y)\n\nprint(historyCallback.history['loss'])\n<\/code><\/pre>\n"},{"AnswerId":"58037507","CreationDate":"2019-09-21T05:24:15.713","ParentId":null,"OwnerUserId":"5597718","Title":null,"Body":"<blockquote>\n  <p>How do I evaluate the value for some_var? <\/p>\n<\/blockquote>\n\n<pre><code>some_var = model.total_loss\ntrain_examples = np.ones((1,32))\ntrain_labels=np.ones((1,10))\nwith keras.backend.get_session() as sess:\n    loss = sess.run(some_var, feed_dict={ 'input_1:0': train_examples,\n                                          'dense_target:0' : train_labels })\n    print(f'loss is {loss}')\n\n\nloss is 3.1767444610595703\n<\/code><\/pre>\n\n<p>Regarding <code>feed_dict<\/code> keys, <\/p>\n\n<ol>\n<li><p>for Keras input layer it's layer_name:0, in the above example, input_1:0 (layer name assigned by keras can be known from <code>model.summary()<\/code>)<\/p><\/li>\n<li><p>When a particular key:value pair is not provided in <code>feed_dict<\/code>, tensorflow throws an error with the key that is missing. <\/p>\n\n<p><em>tensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor '<strong>dense_target<\/strong>' with dtype float and shape [?,?]<\/em><\/p><\/li>\n<li><p>Long answer on why ':0' is needed for the tensor name. <a href=\"http:\/\/stackoverflow.com\/a\/37870634\/419116\">http:\/\/stackoverflow.com\/a\/37870634\/419116<\/a><\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":58037171,"AnswerCount":1,"Tags":"<python-3.x><keras><gpu><conv-neural-network>","CreationDate":"2019-09-21T03:59:58.290","AcceptedAnswerId":null,"OwnerUserId":11435927.0,"Title":"What is difference between the result of using GPU or not?","Body":"<p>I have a CNN with 2 hidden layers. When i use keras on cpu with 8GB RAM, sometimes i have \"Memory Error\" or sometimes precision class was 0 but some classes at the same time were 1.00. If i use keras on GPU,will it solve my problem?<\/p>\n","answers":[{"AnswerId":"58045304","CreationDate":"2019-09-22T01:11:06.093","ParentId":null,"OwnerUserId":"6417245","Title":null,"Body":"<p>You probably don't have enough memory to fit all the images in the CPU during training. Using a GPU will only help if it has more memory. If this is happening because you have too many images or they're resolution is too high, you can try using keras' ImageDataGenerator and any of the flow methods to feed your data in batches. <\/p>\n"}]}
+{"QuestionId":58037276,"AnswerCount":0,"Tags":"<tensorflow><neural-network><theano><lasagne>","CreationDate":"2019-09-21T04:27:10.157","AcceptedAnswerId":null,"OwnerUserId":12098218.0,"Title":"TypeError: Failed to instantiate <class 'lasagne.layers.pool.MaxPool2DLayer'>. Maybe parameter names have changed?","Body":"<pre><code>from lasagne import layers\nlayers=[('input', layers.InputLayer),('conv1', layers.Conv2DLayer),('pool1', layers.MaxPool2DLayer),\n('conv2', layers.Conv2DLayer),\n('pool2', layers.MaxPool2DLayer),\n('conv3', layers.Conv2DLayer),\n('pool3', layers.MaxPool2DLayer),\n('hidden4', layers.DenseLayer),\n('hidden5', layers.DenseLayer),\n('output', layers.DenseLayer),\n]**strong text**\n\nfrom nolearn.lasagne import NeuralNet\nnnet = NeuralNet(layers=layers,input_shape=(None, 3, 32, 32),conv1_num_filters=32,\nconv1_filter_size=(3, 3),\nconv2_num_filters=64,\nconv2_filter_size=(2, 2),\nconv3_num_filters=128,\nconv3_filter_size=(2, 2),\npool1_ds=(2,2),\npool2_ds=(2,2),\npool3_ds=(2,2),\nhidden4_num_units=500,\nhidden5_num_units=500,\noutput_num_units=10,\noutput_nonlinearity=softmax,\nupdate_learning_rate=0.01,\nupdate_momentum=0.9,\nregression=True,\nmax_epochs=3,\nverbose=1)\n\nnnet.fit(X_train, y_train)\n<\/code><\/pre>\n\n<blockquote>\n  <p>nnet.fit(X_train, y_train)                                                                              >File \"\/home\/pandu\/anaconda3\/lib\/python3.7\/site-packages\/nolearn-0.6.1.dev0->py3.7.egg\/nolearn\/lasagne\/base.py\", line 697, in fit                                                                                       >self.initialize()                                                                                       >File \"\/home\/pandu\/anaconda3\/lib\/python3.7\/site-packages\/nolearn-0.6.1.dev0->py3.7.egg\/nolearn\/lasagne\/base.py\", line 485, in initialize                                                                                >self.initialize_layers()                                                                                >File \"\/home\/pandu\/anaconda3\/lib\/python3.7\/site-packages\/nolearn-0.6.1.dev0->py3.7.egg\/nolearn\/lasagne\/base.py\", line 609, in initialize_layers                                                                         >chain_exception(TypeError(msg), e)                                                                      >File \"\/home\/pandu\/anaconda3\/lib\/python3.7\/site-packages\/nolearn-0.6.1.dev0->py3.7.egg\/nolearn\/_compat.py\", line 25, in chain_exception                                                                                 >exec(\"raise exc1 from exc2\")                                                                            >File \"\", line 1, in                                                                     >TypeError: Failed to instantiate  with args {'name': >'pool1', 'ds': (2, 2), 'incoming': }.                    >Maybe parameter names have changed?<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":58037682,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><linear-regression>","CreationDate":"2019-09-21T06:00:06.237","AcceptedAnswerId":58038421.0,"OwnerUserId":9625777.0,"Title":"How to train a Regression model for single input and multiple output?","Body":"<p>I have trained a regression model that approximates the weights for the equation :\n<strong>Y = R+B+G<\/strong>\nFor this, I provide pre-determined values of R, B and G and Y, as training data and after training the model, the model is successfully able to predict the value of Y for given values of R, B and G. I used a neural network with 3 inputs, 1 dense layer (hidden layer) with 2 neurons and the output layer (output) with a single neuron.<\/p>\n\n<pre><code>    hidden = tf.keras.layers.Dense(units=2, input_shape=[3])\n    output = tf.keras.layers.Dense(units=1)\n<\/code><\/pre>\n\n<p>But, I need to implement the inverse of this. i.e., I need to train a model that takes in value of Y and predicts values of R, B and G that corrspond to that value of Y. \nI have just learnt that regression is all about a <strong>single<\/strong> output. So, I am unable to think of a solution and the path to it. \nKindly Help.<br>\n(P.S Is it possible to use the model that I have already trained, to do this? I mean, once, the weights have been determined for R, B and G, is it possible to manipulate the model to use these weights to map Y to R, B and G?)<\/p>\n","answers":[{"AnswerId":"58038421","CreationDate":"2019-09-21T08:08:25.440","ParentId":null,"OwnerUserId":"5136550","Title":null,"Body":"<p>Here is an example to start solving your problem using neural network in tensorflow.<\/p>\n\n<pre><code>import numpy as np\nfrom tensorflow.python.keras.layers import Input, Dense\nfrom tensorflow.python.keras.models import Model\n\nX=np.random.random(size=(100,1))\ny=np.random.randint(0,100,size=(100,3)).astype(float)   #Regression\n\ninput1 = Input(shape=(1,))\nl1 = Dense(10, activation='relu')(input1)\nl2 = Dense(50, activation='relu')(l1)\nl3 = Dense(50, activation='relu')(l2)\nout = Dense(3)(l3)\n\nmodel = Model(inputs=input1, outputs=[out])\nmodel.compile(\n    optimizer='adam',\n    loss=['mean_squared_error']\n    )\n\nhistory = model.fit(X, [y], epochs=10, batch_size=64)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58038162,"AnswerCount":0,"Tags":"<keras><deep-learning><conv-neural-network>","CreationDate":"2019-09-21T07:25:05.623","AcceptedAnswerId":null,"OwnerUserId":1841219.0,"Title":"CNN model gives accuracy 0f 90% on Validation Images but when predicted on the same validation set, accuracy is hardly 5%","Body":"<p>I am new to CNN. I was implementing CNN to classify a dataset with 18000 training images and 10,000 test images into 102 output classes.\nI am using Transfer learning and i can see that training accuracy is around 98% and Validation accuracy is around 97%. But when i predict on the same validation dataset, i hardly get 5% accuracy.\nI tried L1\/L2 regularizarion in Dense layers as looks like overfitting case.\nI tried dropouts But didn't able to get more accuracy than 5-6% on Validation dataset itself. Needed some suggestions and input here.\nValidation Dataset i had created out of train dataset by randomly selecting 50 images of each of the 102 classes , so validation dataset contains around 5000 images. Below is the code<\/p>\n\n<pre><code>from keras.applications.mobilenet import preprocess_input\n\nbase_model=MobileNet(weights='imagenet',include_top=False)\nx=base_model.output\nx=GlobalAveragePooling2D()(x)\nx=Dense(1024,kernel_regularizer=regularizers.l2(0.3),activation='relu')(x)\nx=Dense(1024,kernel_regularizer=regularizers.l2(0.3), activation='relu')(x)\npreds=Dense(102,activation='softmax')(x)\nmodel=Model(inputs=base_model.input,outputs=preds)\n\nfor layer in model.layers[:44]:\n    layer.trainable=False\nfor layer in model.layers[44:]:\n    layer.trainable=True\n\ntrain_datagen=ImageDataGenerator(preprocessing_function=preprocess_input) \n\ntrain_generator=train_datagen.flow_from_directory('.\/train', \n                                                 target_size=(224,224),\n                                                 color_mode='rgb',\n                                                 batch_size=32,\n                                                 class_mode='categorical',\n                                                 shuffle=True)\nvalidation_datagen=ImageDataGenerator(preprocessing_function=preprocess_input)\n\nvalidation_generator=validation_datagen.flow_from_directory('.\/Validation', \n                                                 target_size=(224,224),\n                                                 color_mode='rgb',\n                                                 batch_size=32,\n                                                 class_mode='categorical',\n                                                 shuffle=True)\nmodel.compile(optimizer='adam',loss='categorical_crossentropy',metrics=['accuracy'])\nstep_size_train=train_generator.n\/\/train_generator.batch_size\nMobileNet=model.fit_generator(generator=train_generator,\n                   steps_per_epoch=step_size_train,\n                    validation_data=validation_generator,\n                              validation_steps=validation_generator.n \/\/ validation_generator.batch_size,\n                   epochs=30)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58038394,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><jupyter-notebook><anaconda><virtualenv>","CreationDate":"2019-09-21T08:04:01.997","AcceptedAnswerId":null,"OwnerUserId":9519666.0,"Title":"Unable to activate an virtualenv in Jupyter notebook","Body":"<p>I'm not able to activate the virtual environment created in cmd on Jupyter notebook.<\/p>\n\n<p>I have created a virtual environment for installing tensorflow using anaconda. I can activate the virtualenv in cmd and run the 'sess.run(hello)' test with no problem. However, it fails when I run the same code in Jupyter notebook. I tried to activate the virutalenv on notebook and gained the feedback that said restarting kernel may need. But it remains at the default environment after I restarted kernel.  <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>conda activate tensorflow_cpu\n<\/code><\/pre>\n\n<p>Error: \"Note: you may need to restart the kernel to use updated packages.\"<\/p>\n","answers":[]}
+{"QuestionId":58038826,"AnswerCount":0,"Tags":"<tensorflow><object-detection-api><transfer-learning>","CreationDate":"2019-09-21T09:06:35.080","AcceptedAnswerId":null,"OwnerUserId":2277994.0,"Title":"How can I customize how many weights to use from imagenet pre-trained model?","Body":"<p>I am trying to build a detection model for a few class problem, using Tensorflow Object Detection API.<\/p>\n\n<p>I use an Inception V2 architecture and I try to transfer learning from a model pre-trained on Imagenet classification (<a href=\"http:\/\/download.tensorflow.org\/models\/inception_v2_2016_08_28.tar.gz\" rel=\"nofollow noreferrer\">http:\/\/download.tensorflow.org\/models\/inception_v2_2016_08_28.tar.gz<\/a>). I think I am doing something wrong, because the model does not seem to learn enough. I guess the tasks and dataset are pretty different, so I would need to tell the API how many pre-trained weights to keep. <\/p>\n\n<p>This is my train config.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>train_config: {\n  batch_size: 8\n  optimizer {\n    momentum_optimizer: {\n      learning_rate: {\n        manual_step_learning_rate {\n          initial_learning_rate: 0.0001\n          schedule {\n            step: 900000\n            learning_rate: .00001\n          }\n          schedule {\n            step: 1200000\n            learning_rate: .000001\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_type: \"classification\"\n  fine_tune_checkpoint: \"\/inception_v2.ckpt\"\n  load_all_detection_checkpoint_vars: true\n}\n<\/code><\/pre>\n\n<p>I have only changed the checkpoint type and the checkpoint.<\/p>\n\n<p>Any suggestion on how to improve the performance is welcomed.<\/p>\n\n<p>Thank you!<\/p>\n","answers":[]}
+{"QuestionId":58039031,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-21T09:38:24.503","AcceptedAnswerId":null,"OwnerUserId":6244166.0,"Title":"tf.pow gives negative results and zero with large value of exponent (power)","Body":"<p>I am using tensorflow 1.14.0 version. While performing tf.pow operation I have observed that tf.pow returns negative or zero value for large value of exponent.<\/p>\n\n<pre><code>a = tf.constant(50)\nb = tf.constant(9)\nc = tf.constant(10)\nd = tf.constant(15)\ne = tf.constant(100)\n\nwith tf.Session() as sess:\n  a_b, a_c, a_d, a_e = sess.run([tf.pow(a,b),\n                                 tf.pow(a, c),\n                                 tf.pow(a, d),\n                                 tf.pow(a, e)])\n  print('a_b is: ', a_b)\n  print('a_c is: ', a_c)\n  print('a_d is: ',a_d)\n  print('a_e is:', a_e)\n<\/code><\/pre>\n\n<p>Printed values are:<\/p>\n\n<pre><code> a_b is:  1507045888\n a_c is:  -1957116928\n a_d is:  -606830592\n a_e is: 0\n<\/code><\/pre>\n\n<p>Can anyone provides an explanation of why this is happening ? Is this a bug in tensorflow or something else which I might not be aware of  ?<\/p>\n","answers":[]}
+{"QuestionId":58039174,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-21T09:56:57.340","AcceptedAnswerId":null,"OwnerUserId":2461399.0,"Title":"Intialize tensorflow one time on server","Body":"<p>I'm working on models using tensorflow and keras.\nThe system calls tensoflorw and instialize it every time I run the code.\nI want to store the initialization in RAM from the run in the first time only. So I don't have to wait all this long every time. \nIs there any solution for such idea?<\/p>\n","answers":[]}
+{"QuestionId":58039221,"AnswerCount":0,"Tags":"<python><numpy><matrix><pytorch><cupy>","CreationDate":"2019-09-21T10:04:29.223","AcceptedAnswerId":null,"OwnerUserId":10987015.0,"Title":"Basic matrix operation in PyTorch\/CuPy using GPU","Body":"<p>I have a numpy script where I do the following operation with big matrices (can go over <code>10000x10000<\/code> with <code>float<\/code> values):<\/p>\n\n<pre><code>F = (I - Q)^-1 * R\n<\/code><\/pre>\n\n<p>I first used pytorch tensors on CPU (i7-8750H) and it runs 2 times faster:<\/p>\n\n<pre><code>tensorQ = torch.from_numpy(Q)\ntensorR = torch.from_numpy(R)\nsub= torch.eye(a * d, dtype=float) - tensorQ\ninv= torch.inverse(sub)\ntensorF = torch.mm(inv, tensorR)\nF = tensorF.numpy()\n<\/code><\/pre>\n\n<p>Now I'm trying to execute it on GPU (1050Ti Max-Q) to see if I can get another speedup but the code run slower than numpy version (I've already installed CUDA and cuDNN). Maybe pytorch it's not even the best library to do this kind of things, but I'm learning it now and I thought it could help me:<\/p>\n\n<pre><code>dev = torch.device('cuda')\n\ntensorQ = torch.from_numpy(Q).to(dev)\ntensorR = torch.from_numpy(R).to(dev)\nsub= torch.eye(a * d, dtype=float).to(dev) - tensorQ\ninv= torch.inverse(sub).to(dev)\ntensorF = torch.mm(inv, tensorR).cpu()\nF = tensorF.numpy()\n<\/code><\/pre>\n\n<p>Am I missing something?<\/p>\n\n<p><strong>Edit:<\/strong>\nI also tried using CuPy but it's still slow:<\/p>\n\n<pre><code>Q = cp.array(matrixQ)\nR = cp.array(matrixR)\nsub = cp.identity(attacker * defender) - matrixQ\ninv = cp.linalg.inv(sub)\nF = cp.matmul(inv, matrixR)\nF = cp.asnumpy(matrixF)\n<\/code><\/pre>\n\n<p>Probably the overhead of memory allocation is just too big compared to the computation of few operations.<\/p>\n","answers":[]}
+{"QuestionId":58039243,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-21T10:07:07.827","AcceptedAnswerId":58040943.0,"OwnerUserId":10419891.0,"Title":"Keras for find sentences similarities from pre-trained word2vec","Body":"<p>I have pre-trained <code>word2vec<\/code> from <code>gensim<\/code>. And Using <code>gensim<\/code> for finding the similarities between <code>words<\/code> works as expected. But I am having problem in finding the similarities between two different sentences. Using of <code>cosine<\/code> <code>similarities<\/code> is not a good option for sentences and Its not giving good result. <code>Soft Cosine similarities<\/code> in <code>gensim<\/code> gives a little better results but still, it is also not looking good. <\/p>\n\n<p>I found <code>WMDsimilarities<\/code> in <code>gensim<\/code>. This is a bit better than <code>softcosine<\/code> and <code>cosine<\/code>.<\/p>\n\n<p>I am thinking if there is more option like using deep learning like <code>keras<\/code> and <code>tensorflow<\/code> to find the sentences similarities from pre-trained word2vec. I know the classification can be done using <code>word embbeding<\/code>  and this seems somewhat better options but then I need to find a training data and labeled it from the scratch. <\/p>\n\n<p>So, I am wondering if there is any other option which can be used pre-trained <code>word2vec<\/code> in <code>keras<\/code> and get the sentences similarities. Is there way. I am open to any suggestions and advice.<\/p>\n","answers":[{"AnswerId":"58040943","CreationDate":"2019-09-21T13:52:16.453","ParentId":null,"OwnerUserId":"7383560","Title":null,"Body":"<p>Before reimplementing the wheel I'd suggest to try <a href=\"https:\/\/radimrehurek.com\/gensim\/models\/doc2vec.html\" rel=\"nofollow noreferrer\"><code>doc2vec<\/code> method from gensim<\/a>, it works quite well and it's easy to use.<\/p>\n\n<p>To implement it in Keras reusing the embeddings you have computed with <code>gensim<\/code>:<\/p>\n\n<ol>\n<li>Store the word embeddings in a file, one word per line with the corresponding embedding. Alternatively you can do as @Paul suggested and skip the step 2 and reuse the layer in step 3.<\/li>\n<li>Load word embeddings into a Keras Embedding layer. You can checkout this <a href=\"https:\/\/keras.io\/examples\/pretrained_word_embeddings\/\" rel=\"nofollow noreferrer\">Keras tutorial<\/a> for more details (check how <code>embedding_layer<\/code> variable is initialized).<\/li>\n<li>Then a sequence to sequence model can be used to compute the embedding of the text. In which you have an encoder that embeds the string and the decoder that converts the embedding back to a string. Here is a <a href=\"https:\/\/keras.io\/examples\/lstm_seq2seq\/\" rel=\"nofollow noreferrer\">Keras tutorial that translates from English to French<\/a>. You can use a similar process to transform your text into your text and pick the internal embedding for your similarity metric.<\/li>\n<\/ol>\n\n<p>You can also have a look how the <a href=\"https:\/\/cs.stanford.edu\/~quocle\/paragraph_vector.pdf\" rel=\"nofollow noreferrer\">paragraph to vector model<\/a> works, you can also implement it using Keras and loading the word embedding weights that you have computed.<\/p>\n"}]}
+{"QuestionId":58039344,"AnswerCount":0,"Tags":"<python><python-3.x><pytorch><onnx>","CreationDate":"2019-09-21T10:22:51.250","AcceptedAnswerId":null,"OwnerUserId":12099010.0,"Title":"ERFNet (pytorch) to ONNX","Body":"<p>When I convert a ERFNet(pytorch) model to ONNX. <\/p>\n\n<p>(This model is from <a href=\"https:\/\/github.com\/wvangansbeke\/LaneDetection_End2End\" rel=\"nofollow noreferrer\">https:\/\/github.com\/wvangansbeke\/LaneDetection_End2End<\/a>)<br \/>\nMy code is as follows.<\/p>\n\n<pre><code>from torch.autograd import Variable\nimport torch.onnx\nimport torchvision\nfrom Networks.LSQ_layer import Net\nfrom Networks.utils import define_args, save_weightmap, first_run, \\\n                           mkdir_if_missing, Logger, define_init_weights,\\\n                           define_scheduler, define_optim, AverageMeter\n\ndummy_input = Variable(torch.randn(1,3,256,512)).cuda()\nparser = define_args()\nargs = parser.parse_known_args()[0]  \n\nmodel = Net(args)\ndefine_init_weights(model, args.weight_init)\ncheckpoint = torch.load(\"model_best_epoch_204.pth.tar\")\nmodel.load_state_dict(checkpoint['state_dict'])\nmodel = model.cuda()\n\ntorch.onnx.export(model, dummy_input, \"LaneDetection.onnx\", verbose=True)\n<\/code><\/pre>\n\n<p>This error appears.<br \/>\nValueError: Auto nesting doesn't know how to process an input object of type int. Accepted types: Tensors, or lists\/tuples of them<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/tmp\/pycharm_project_633\/venv\/Scripts\/darknet2onnx.py\", line 22, in &lt;module&gt;\n    torch.onnx.export(model, dummy_input, \"LaneDetection.onnx\", verbose=True)\n  File \"\/work\/dependence\/anaconda3\/lib\/python3.6\/site-packages\/torch\/onnx\/__init__.py\", line 27, in export\n    return utils.export(*args, **kwargs)\n  File \"\/work\/dependence\/anaconda3\/lib\/python3.6\/site-packages\/torch\/onnx\/utils.py\", line 104, in export\n    operator_export_type=operator_export_type)\n  File \"\/work\/dependence\/anaconda3\/lib\/python3.6\/site-packages\/torch\/onnx\/utils.py\", line 281, in _export\n    example_outputs, propagate)\n  File \"\/work\/dependence\/anaconda3\/lib\/python3.6\/site-packages\/torch\/onnx\/utils.py\", line 224, in _model_to_graph\n    graph, torch_out = _trace_and_get_graph_from_model(model, args, training)\n  File \"\/work\/dependence\/anaconda3\/lib\/python3.6\/site-packages\/torch\/onnx\/utils.py\", line 192, in _trace_and_get_graph_from_model\n    trace, torch_out = torch.jit.get_trace_graph(model, args, _force_outplace=True)\n  File \"\/work\/dependence\/anaconda3\/lib\/python3.6\/site-packages\/torch\/jit\/__init__.py\", line 197, in get_trace_graph\n    return LegacyTracedModule(f, _force_outplace)(*args, **kwargs)\n  File \"\/work\/dependence\/anaconda3\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/module.py\", line 489, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/work\/dependence\/anaconda3\/lib\/python3.6\/site-packages\/torch\/jit\/__init__.py\", line 252, in forward\n    out = self.inner(*trace_inputs)\n  File \"\/work\/dependence\/anaconda3\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/module.py\", line 487, in __call__\n    result = self._slow_forward(*input, **kwargs)\n  File \"\/work\/dependence\/anaconda3\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/module.py\", line 477, in _slow_forward\n    result = self.forward(*input, **kwargs)\n  File \"\/tmp\/pycharm_project_633\/venv\/Scripts\/Networks\/LSQ_layer.py\", line 295, in forward\n    shared_encoder, output = self.net(input, end_to_end*self.pretrained)\n  File \"\/work\/dependence\/anaconda3\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/module.py\", line 487, in __call__\n    result = self._slow_forward(*input, **kwargs)\n  File \"\/work\/dependence\/anaconda3\/lib\/python3.6\/site-packages\/torch\/nn\/modules\/module.py\", line 464, in _slow_forward\n    input_vars = tuple(torch.autograd.function._iter_tensors(input)) #lulu \n  File \"\/work\/dependence\/anaconda3\/lib\/python3.6\/site-packages\/torch\/autograd\/function.py\", line 284, in _iter\n    for var in _iter(o):\n  File \"\/work\/dependence\/anaconda3\/lib\/python3.6\/site-packages\/torch\/autograd\/function.py\", line 293, in _iter\n    if condition_msg else \"\"))\nValueError: Auto nesting doesn't know how to process an input object of type int. Accepted types: Tensors, or lists\/tuples of them\n<\/code><\/pre>\n\n<p>After debug, I find the problem is from \nanaconda3\/lib\/python3.6\/site-packages\/torch\/autograd\/function.py, line 293, in _iter<\/p>\n\n<pre><code>def _iter_filter(condition, allow_unknown=False, condition_msg=None,\n                 conversion=None):#lulu change allow_unknown=False to True\n    def _iter(obj):\n        if conversion is not None:\n            obj = conversion(obj)\n        if condition(obj):\n            yield obj\n        elif obj is None:\n            return\n        elif isinstance(obj, (list, tuple)):\n            for o in obj:\n                for var in _iter(o):\n                    yield var\n        elif allow_unknown:\n            yield obj\n        else:\n            raise ValueError(\"Auto nesting doesn't know how to process \"\n                             \"an input object of type \" + torch.typename(obj) +\n                             (\". Accepted types: \" + condition_msg +\n                              \", or lists\/tuples of them\"\n                              if condition_msg else \"\"))\n\n    return _iter\n<\/code><\/pre>\n\n<p>In this function,<\/p>\n\n<pre><code>for var in _iter(o):\n    yield var\n<\/code><\/pre>\n\n<p>It let o = (int) 0 in a cycle and report ValueError.<\/p>\n\n<p>Why does this zero be generated? How can I fix it?\nIf you want more details, please comment. I will reply you as soon as I can.<\/p>\n","answers":[]}
+{"QuestionId":58039747,"AnswerCount":1,"Tags":"<artificial-intelligence><conv-neural-network><pytorch><vgg-net>","CreationDate":"2019-09-21T11:17:36.153","AcceptedAnswerId":null,"OwnerUserId":12099199.0,"Title":"Why is my VGG-19 implementation not performing well?","Body":"<p>I have to implement VGG-19 manually in order to have the best accuracy rate on CIFAR-10 data test.<\/p>\n\n<p>I read the paper about VGGNet and implemented the VGG-19 network but I have an accuracy rate of 10% on the test dataset... I tried changing Batch size, learning rate with no improvement.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class MyClassifier(nn.Module):\n  def __init__(self):\n        super(MyClassifier, self).__init__()\n        self.conv1 = nn.Conv2d(in_channels=3, out_channels=64, kernel_size=3, stride=1, padding=1)\n        self.conv12 = nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=1, padding=1)\n\n        self.conv2 = nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, stride=1, padding=1)\n        self.conv22 = nn.Conv2d(in_channels=128, out_channels=128, kernel_size=3, stride=1, padding=1)\n\n        self.conv3 = nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, stride=1, padding=1)\n        self.conv32 = nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, stride=1, padding=1)\n\n        self.conv4 = nn.Conv2d(in_channels=256, out_channels=512, kernel_size=3, stride=1, padding=1)\n        self.conv42 = nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1)\n\n\n        self.relu = nn.ReLU()\n        self.pool = nn.MaxPool2d(kernel_size=2, stride=2)\n\n        self.Dropout = nn.Dropout(p=0.5)\n        self.fc1 = nn.Linear(in_features=512*1*1, out_features=4096)\n        self.fc2 = nn.Linear(in_features=4096, out_features=4096)\n        self.fc3 = nn.Linear(in_features=4096, out_features=output_dim)\n\n  def forward(self, x):\n      x = self.relu(self.conv1(x))\n      x = self.pool(self.relu(self.conv12(x)))\n\n      x = self.relu(self.conv2(x))\n      x = self.pool(self.relu(self.conv22(x)))\n\n      x = self.relu(self.conv3(x))\n      x = self.pool(self.relu(self.conv32(x)))\n\n      x = self.relu(self.conv4(x))\n      x = self.relu(self.conv42(x))\n      x = self.relu(self.conv42(x))\n      x = self.pool(self.relu(self.conv42(x)))\n\n      x = self.relu(self.conv42(x))\n      x = self.relu(self.conv42(x))\n      x = self.relu(self.conv42(x))\n      x = self.pool(self.relu(self.conv42(x)))\n\n      x = x.view(-1, 512*1*1)\n      x = self.Dropout(self.relu(self.fc1(x)))\n      x = self.Dropout(self.relu(self.fc2(x)))\n      outputs = self.fc3(x)\n      return outputs\n<\/code><\/pre>\n\n<p>Here are my results:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>[epoch:0, iteration:2000] train loss : 2.5708 train accuracy : 0.0000\n[epoch:0, iteration:4000] train loss : 2.4595 train accuracy : 0.0000\n[epoch:0, iteration:6000] train loss : 2.2051 train accuracy : 0.0000\n[epoch:0, iteration:8000] train loss : 2.4449 train accuracy : 0.0000\n[epoch:0, iteration:10000] train loss : 2.3113 train accuracy : 0.2500\n[epoch:0, iteration:12000] train loss : 2.3602 train accuracy : 0.0000\n[epoch:0, iteration:12500] test_loss : 2.3092 test accuracy : 0.1000\ncheckpoint is saved !\n<\/code><\/pre>\n\n<p>My test accuracy is 0.1 (10%) at epoch:0 but stays 0.1 even after 10 epochs. I believed I could achieve at least 0.8 (80%) with VGG-19's code. \nDo you see any problem with the code or is it something else ?<\/p>\n","answers":[{"AnswerId":"58048261","CreationDate":"2019-09-22T10:27:14.527","ParentId":null,"OwnerUserId":"12102363","Title":null,"Body":"<p>If I had to guess, I would think it has something to do with your loss function during training. Are you using Cross-Entropy loss? If you're using NLLLoss, you need a log softmax layer at the end of your network. <\/p>\n"}]}
+{"QuestionId":58039953,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><ml.net>","CreationDate":"2019-09-21T11:44:09.640","AcceptedAnswerId":null,"OwnerUserId":960567.0,"Title":"Person recolonization using ML.NET\/TensorFlow","Body":"<p>I am noob in ML. I have a Person table that have,<\/p>\n\n<pre><code>-----------------------------------\n               User\n-----------------------------------\nUserId | UserName | UserPicturePath\n1      |   MyName | MyName.jpeg\n<\/code><\/pre>\n\n<p>Now I have tens of millions of persons in my database. I wanna train my model to predict the <code>UserId<\/code> by giving images(png\/jpeg\/tiff) in bytes. So, input will be images and the output I am looking is <code>UserId<\/code>. Right now I am looking for a solution in ML.NET but I am open to switch to TensorFlow.<\/p>\n","answers":[{"AnswerId":"58263413","CreationDate":"2019-10-07T04:07:15.323","ParentId":null,"OwnerUserId":"9828197","Title":null,"Body":"<p>Well, this is nothing but a mapping problem, particularly an id-to-face mapping problem, and neural nets excell at this more than on anything else.<\/p>\n\n<p>As you have understood by now, you can do this using tensorflow, pytorch or any library of the same purpose.<\/p>\n\n<p>But if you want to use tensorflow, read on for a ready code at the end. It is easiest to achieve your task by transfer learning, i.e. by loading some pretrained model, freezing all but last layer and then training the network to produce a latent one-dimensional vector for a given face image. Then you can save this vector into a database and map it into an id. <\/p>\n\n<p>Then, whenever there is a new image and you want to predict an id for the image, you run your image through the network, get your vector and compute cosine similarity with vectors in your database. If the similarity is above some threshold and it is the highest among other similarities, you have found your id.<\/p>\n\n<p>There are many ways to go about this. Sure you have to preprocess your data, and augment it at the same time, but if you want some ready code to play with then have a look at this famous happy house tutorial from Andrew NG and his team:<\/p>\n\n<p><a href=\"https:\/\/github.com\/gemaatienza\/Deep-Learning-Coursera\/blob\/master\/4.%20Convolutional%20Neural%20Networks\/Keras%20-%20Tutorial%20-%20Happy%20House%20v2.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/gemaatienza\/Deep-Learning-Coursera\/blob\/master\/4.%20Convolutional%20Neural%20Networks\/Keras%20-%20Tutorial%20-%20Happy%20House%20v2.ipynb<\/a><\/p>\n\n<p>This should suffice your needs.<\/p>\n\n<p>Hope it helps!<\/p>\n"}]}
+{"QuestionId":58040010,"AnswerCount":1,"Tags":"<tensorflow><tensorflow2.0><tf.keras>","CreationDate":"2019-09-21T11:51:32.547","AcceptedAnswerId":58045484.0,"OwnerUserId":7657215.0,"Title":"How to implement matmul-based nn written in TF1 to TF2","Body":"<p>I want to implement simple, matmul-based neural network written in TF1 to TF2.<\/p>\n\n<p><a href=\"https:\/\/github.com\/golbin\/TensorFlow-Tutorials\/blob\/master\/04%20-%20Neural%20Network%20Basic\/02%20-%20Deep%20NN.py\" rel=\"nofollow noreferrer\">Here is source.<\/a> (don't mind Korean comments, it's tutorial written in Korean)<\/p>\n\n<p>So I found 'how to migrate TF1 into TF2', and I know I have to remove placeholders.<\/p>\n\n<p>Here is my code overall:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\nx_data = np.array(\n    [[0, 0], [1, 0], [1, 1], [0, 0], [0, 0], [0, 1]])\n\ny_data = np.array([\n    [1, 0, 0],\n    [0, 1, 0],\n    [0, 0, 1],\n    [1, 0, 0],\n    [1, 0, 0],\n    [0, 0, 1]\n])\n\nx_data = tf.cast(x_data, tf.float32)\ny_data = tf.cast(y_data, tf.float32)\n\nW1 = tf.Variable(tf.random.uniform([2, 10], -1., 1.))\nW2 = tf.Variable(tf.random.uniform([10, 3], -1., 1.))\nb1 = tf.Variable(tf.zeros([10]))\nb2 = tf.Variable(tf.zeros([3]))\n\nLayer1 = tf.matmul(x_data, W1) + b1\nLayer1 = tf.nn.relu(Layer1)\n\nmodel = tf.matmul(Layer1, W2) + b2\n\ncost = tf.reduce_mean(\n    tf.nn.softmax_cross_entropy_with_logits(labels=y_data, logits=model)\n)\noptimizer = tf.keras.optimizers.Adam(learning_rate=0.01)\n\noptimizer.minimize(loss=lambda: cost , var_list=[W1, W2, b1, b2])\n<\/code><\/pre>\n\n<p>I found <a href=\"https:\/\/stackoverflow.com\/questions\/55953344\/issue-with-adamoptimizer\">Issue with AdamOptimizer<\/a> , and I changed loss to lambda.<\/p>\n\n<p>And, got this Error:<\/p>\n\n<pre><code>ValueError: No gradients provided for any variable: ['Variable:0', 'Variable:0', 'Variable:0', 'Variable:0'].\n<\/code><\/pre>\n\n<p>And I have no idea how to fix it.<\/p>\n\n<p>So I'd like to know proper way to implement something like this in TF2.<\/p>\n","answers":[{"AnswerId":"58045484","CreationDate":"2019-09-22T02:06:12.263","ParentId":null,"OwnerUserId":"7657215","Title":null,"Body":"<p>All right. I went through <a href=\"https:\/\/www.tensorflow.org\/guide\/eager\" rel=\"nofollow noreferrer\">official guide for eager execution<\/a>, and finally done it.<\/p>\n\n<p>Here's the code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\nx_data = np.array(\n    [[0, 0], [1, 0], [1, 1], [0, 0], [0, 0], [0, 1]])\n\ny_data = np.array([\n    [1, 0, 0],\n    [0, 1, 0],\n    [0, 0, 1],\n    [1, 0, 0],\n    [1, 0, 0],\n    [0, 0, 1]\n])\n\nx_data = tf.cast(x_data, tf.float32)\ny_data = tf.cast(y_data, tf.float32)\n\nclass Model(tf.keras.Model):\n    def __init__(self):\n        super(Model, self).__init__()\n        self.W1 = tf.Variable(tf.random.uniform([2, 10], -1., 1.))\n        self.W2 = tf.Variable(tf.random.uniform([10, 3], -1., 1.))\n        self.b1 = tf.Variable(tf.zeros([10]))\n        self.b2 = tf.Variable(tf.zeros([3]))\n    def _calc_layer(x, w, b):\n        return tf.matmul(x, w) + b\n    def __call__(self, x):\n        layer1 = tf.nn.relu(Model._calc_layer(x_data, self.W1, self.b1))\n        return Model._calc_layer(layer1, self.W2, self.b2)\n\ndef cost(model, inputs, targets):\n    return tf.reduce_mean(\n        tf.nn.softmax_cross_entropy_with_logits(labels=targets,\n                                                logits=model(inputs))\n    )\n\nmodel = Model()\n\ndef cost_tominimize():\n    return cost(model, x_data, y_data)\n\noptimizer = tf.keras.optimizers.Adam(learning_rate=0.01)\n\nfor i in range(100):\n    optimizer.minimize(loss=cost_tominimize,\n                        var_list=model.trainable_variables)\n    #print(cost_tominimize().numpy())\n\n#test\nprediction = tf.argmax(model(x_data), 1)\ntarget = tf.argmax(y_data, 1)\nprint(\"prediction : \", prediction.numpy())\nprint(\"real : \", target.numpy())\n\nis_correct = tf.equal(prediction, target)\naccuracy = tf.reduce_mean(tf.cast(is_correct, tf.float32))\nprint('accuracy: %.2f%%' % (accuracy * 100))\n<\/code><\/pre>\n\n<p>Also there is version which doesn't use class:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\nx_data = np.array(\n    [[0, 0], [1, 0], [1, 1], [0, 0], [0, 0], [0, 1]])\n\ny_data = np.array([\n    [1, 0, 0],\n    [0, 1, 0],\n    [0, 0, 1],\n    [1, 0, 0],\n    [1, 0, 0],\n    [0, 0, 1]\n])\n\nx_data = tf.cast(x_data, tf.float32)\ny_data = tf.cast(y_data, tf.float32)\n\nW1 = tf.Variable(tf.random.uniform([2, 10], -1., 1.))\nW2 = tf.Variable(tf.random.uniform([10, 3], -1., 1.))\nb1 = tf.Variable(tf.zeros([10]))\nb2 = tf.Variable(tf.zeros([3]))\n\ndef calc_layer(x, w, b):\n    return tf.matmul(x, w) + b\n\ndef model(x):\n    layer1 = tf.nn.relu(calc_layer(x, W1, b1))\n    return calc_layer(layer1, W2, b2)\n\ndef cost(model, inputs, targets):\n    return tf.reduce_mean(\n        tf.nn.softmax_cross_entropy_with_logits(labels=targets,\n                                                logits=model(inputs))\n    )\n\ndef cost_tominimize():\n    return cost(model, x_data, y_data)\n\noptimizer = tf.keras.optimizers.Adam(learning_rate = 0.01)\n\nfor i in range(100):\n    optimizer.minimize(loss=cost_tominimize,\n                       var_list = [W1, W2, b1, b2])\n    print(cost_tominimize().numpy())\n\n#...and test part here...\n<\/code><\/pre>\n"}]}
+{"QuestionId":58040012,"AnswerCount":0,"Tags":"<tensorflow><computer-vision>","CreationDate":"2019-09-21T11:51:40.287","AcceptedAnswerId":null,"OwnerUserId":3284573.0,"Title":"Visibility on map with a range of angles","Body":"<p>I have a map for a robot that looks like the left figure below. I am writing code to simulate visibility in map when the robot is facing a particular angle. For example, the right figure states that, if the robot stands in the yellow region and faces the top-left corner, it will look outside the map.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Xc8gf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Xc8gf.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The problem right now is that a given image, like the following image, is taken for a range of angles (e.g. [-30, 30] degrees), according to the field-of-view of the robot. The yellow regions in this image correspond to regions outside the map. However, my code is for 1 angle only.<\/p>\n\n<p>Are there any established algorithms in computer graphics for computing the visibility region, corresponding to a given image and a range of angles? The implementation is preferably in TensorFlow.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/15ORK.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/15ORK.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":58040020,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><tflearn>","CreationDate":"2019-09-21T11:53:49.200","AcceptedAnswerId":58040081.0,"OwnerUserId":9052730.0,"Title":"How to run python code without the tensorflow warning","Body":"<p>I am using python 3.6.9 and tensorflow 1.14.0 on windows 10 and have installed and updated all libraries on python 3.6. When I run my code I expect it to just print out the json data but instead I get a big tensorflow\/tflearn error and I don't know why. <\/p>\n\n<p>My code :<\/p>\n\n<pre><code>import nltk\nimport numpy\nimport tflearn\nimport tensorflow\nimport random\nimport json\n\nfrom nltk.stem.lancaster import LancasterStemmer\nstemmer = LancasterStemmer()\n\n\nwith open(\"intents.json\") as file:\n    data = json.load(file)\n\nprint(data)\n<\/code><\/pre>\n\n<p>Error I get :<\/p>\n\n<pre><code>C:\\Users\\PC\\PycharmProjects\\Machine Learning\\venv\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:516: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint8 = np.dtype([(\"qint8\", np.int8, 1)])\nC:\\Users\\PC\\PycharmProjects\\Machine Learning\\venv\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:517: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint8 = np.dtype([(\"quint8\", np.uint8, 1)])\nC:\\Users\\PC\\PycharmProjects\\Machine Learning\\venv\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:518: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint16 = np.dtype([(\"qint16\", np.int16, 1)])\nC:\\Users\\PC\\PycharmProjects\\Machine Learning\\venv\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:519: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint16 = np.dtype([(\"quint16\", np.uint16, 1)])\nC:\\Users\\PC\\PycharmProjects\\Machine Learning\\venv\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:520: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint32 = np.dtype([(\"qint32\", np.int32, 1)])\nC:\\Users\\PC\\PycharmProjects\\Machine Learning\\venv\\lib\\site-packages\\tensorflow\\python\\framework\\dtypes.py:525: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  np_resource = np.dtype([(\"resource\", np.ubyte, 1)])\nC:\\Users\\PC\\PycharmProjects\\Machine Learning\\venv\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:541: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint8 = np.dtype([(\"qint8\", np.int8, 1)])\nC:\\Users\\PC\\PycharmProjects\\Machine Learning\\venv\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:542: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint8 = np.dtype([(\"quint8\", np.uint8, 1)])\nC:\\Users\\PC\\PycharmProjects\\Machine Learning\\venv\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:543: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint16 = np.dtype([(\"qint16\", np.int16, 1)])\nC:\\Users\\PC\\PycharmProjects\\Machine Learning\\venv\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:544: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_quint16 = np.dtype([(\"quint16\", np.uint16, 1)])\nC:\\Users\\PC\\PycharmProjects\\Machine Learning\\venv\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:545: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  _np_qint32 = np.dtype([(\"qint32\", np.int32, 1)])\nC:\\Users\\PC\\PycharmProjects\\Machine Learning\\venv\\lib\\site-packages\\tensorboard\\compat\\tensorflow_stub\\dtypes.py:550: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) \/ '(1,)type'.\n  np_resource = np.dtype([(\"resource\", np.ubyte, 1)])\nWARNING:tensorflow:From C:\\Users\\PC\\PycharmProjects\\Machine Learning\\venv\\lib\\site-packages\\tflearn\\helpers\\summarizer.py:9: The name tf.summary.merge is deprecated. Please use tf.compat.v1.summary.merge instead.\n\nWARNING:tensorflow:From C:\\Users\\PC\\PycharmProjects\\Machine Learning\\venv\\lib\\site-packages\\tflearn\\helpers\\trainer.py:25: The name tf.summary.FileWriter is deprecated. Please use tf.compat.v1.summary.FileWriter instead.\n\nWARNING:tensorflow:From C:\\Users\\PC\\PycharmProjects\\Machine Learning\\venv\\lib\\site-packages\\tflearn\\collections.py:13: The name tf.GraphKeys is deprecated. Please use tf.compat.v1.GraphKeys instead.\n\nWARNING:tensorflow:From C:\\Users\\PC\\PycharmProjects\\Machine Learning\\venv\\lib\\site-packages\\tflearn\\config.py:123: The name tf.get_collection is deprecated. Please use tf.compat.v1.get_collection instead.\n\nWARNING:tensorflow:From C:\\Users\\PC\\PycharmProjects\\Machine Learning\\venv\\lib\\site-packages\\tflearn\\config.py:129: The name tf.add_to_collection is deprecated. Please use tf.compat.v1.add_to_collection instead.\n\nWARNING:tensorflow:From C:\\Users\\PC\\PycharmProjects\\Machine Learning\\venv\\lib\\site-packages\\tflearn\\config.py:131: The name tf.assign is deprecated. Please use tf.compat.v1.assign instead.\n<\/code><\/pre>\n\n<p>What do I do to get rid of that error and run my code properly?<\/p>\n","answers":[{"AnswerId":"58040081","CreationDate":"2019-09-21T12:02:05.933","ParentId":null,"OwnerUserId":"11841571","Title":null,"Body":"<p><code>Numpy<\/code> will gonna deprecate some API. See the <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/release.html#shape-1-fields-in-dtypes-won-t-be-collapsed-to-scalars-in-a-future-version\" rel=\"nofollow noreferrer\">release note<\/a>.<\/p>\n\n<blockquote>\n  <p>Shape-1 fields in dtypes won\u2019t be collapsed to scalars in a future version<\/p>\n  \n  <p>Currently, a field specified as [(name, dtype, 1)] or \"1type\" is interpreted as a scalar field (i.e., the same as [(name, dtype)] or [(name, dtype, ()]). This now raises a FutureWarning; in a future version, it will be interpreted as a shape-(1,) field, i.e. the same as [(name, dtype, (1,))] or \"(1,)type\" (consistently with [(name, dtype, n)] \/ \"ntype\" with n>1, which is already equivalent to [(name, dtype, (n,)] \/ \"(n,)type\").<\/p>\n<\/blockquote>\n\n<p>But your <code>tensorflow<\/code> version uses [(name, dtype, 1)] and it cause the deprecation warning. Since it is warning, you can just ignore and later <code>tensorflow<\/code> version or latest will take care about this warning.<\/p>\n\n<p>This is already implemented into <code>tensorflow<\/code> package, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/3402d7118460857cf484f57338d14d9113597d15\" rel=\"nofollow noreferrer\">see this<\/a>. Please update package to the latest, it will resolve the issue I think.<\/p>\n\n<p>One more thing is that the <code>tflearn<\/code> package also have deprecated commands. It is again used in the <code>tensorflow<\/code> package but since <code>tflearn<\/code> package is not active recently, I think you have to ignore this warning, that will not be resolved. See <a href=\"https:\/\/github.com\/tflearn\/tflearn\" rel=\"nofollow noreferrer\">the repository<\/a>.<\/p>\n"}]}
+{"QuestionId":58040324,"AnswerCount":1,"Tags":"<python><machine-learning><deep-learning><pytorch><matrix-multiplication>","CreationDate":"2019-09-21T12:35:33.573","AcceptedAnswerId":null,"OwnerUserId":3342270.0,"Title":"Difference of torch.matmul and python built-in @ operator to do matrix multiplication","Body":"<p>Can I always replace torch.matmul with python's built-in @ operator to do the matrix multiplication? Please assume that I know the difference between torch.matmul, torch.mm and many others. I just want to make sure how many of them can be safely replaced by @ operator without sacrificing speed or some native support from torch. <\/p>\n\n<p>If it does no harm, I would like to extensively use them in the future. <\/p>\n","answers":[{"AnswerId":"58042440","CreationDate":"2019-09-21T17:11:44.793","ParentId":null,"OwnerUserId":"3236925","Title":null,"Body":"<p>yes you can always use <code>@<\/code> in place of <code>.matmul<\/code>. You can also do the same in <code>numpy<\/code>. Also see <a href=\"https:\/\/stackoverflow.com\/questions\/53370003\/pytorch-mapping-operators-to-functions\">PyTorch mapping operators to functions<\/a><\/p>\n"}]}
+{"QuestionId":58040773,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-21T13:30:22.113","AcceptedAnswerId":null,"OwnerUserId":9278776.0,"Title":"How can I reset the metrics every 1000 steps in Tensorflow Estimator?","Body":"<p>I created a classifier with tensorflow(1.14.1) estimator and calculated the metrics every step. I used tensorboard to show the metrics and it agregrated every step. I want to reset it each epoch. <\/p>\n\n<p>This is for HPC, running Python(3.7) and tensorflow(1.14.1).<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def fn_builder(...):\n    def model_fn(features, labels, mode, params):\n        # create bert-mlp net\n        ids = features['ids']\n        words = features['words']\n        masks = features['mask']\n        label_mask = features['label_mask']\n\n        # model\n        is_training = (mode == tf.estimator.ModeKeys.TRAIN)\n        loss, prd_label = create_model(...)\n\n        pre = tf.metrics.precision(labels, prd_label, weights=label_mask)\n\n        metrics = {'Precision': pre}\n        tf.summary.scalar('p', pre[1])\n\n\n        if mode == tf.estimator.ModeKeys.TRAIN:\n            optimizer = Adam(learning_rate=learning_rate)\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)\n\n\n\n    return pi_model_fn\n\nif __name__ == \"__main__\":\n    batch_size = 5\n    epochs = 30\n    num_eval_steps = 20 \/ batch_size + 1\n    num_train_steps = 20 * epochs \/ batch_size\n\n    pi_checkpointing_config = tf.estimator.RunConfig(\n        model_dir='data\/pi',\n        save_checkpoints_secs=10 * 60,  # Save checkpoints every 10min.\n        keep_checkpoint_max=10,  # Retain the 10 most recent checkpoints.\n\n    )\n\n    pi = tf.estimator.Estimator(\n        model_fn=fn_builder(...),\n        config=pi_checkpointing_config)\n    pi.train(\n        input_fn=input_fn_creator(...),\n        max_steps=num_train_steps\n    )\n\n<\/code><\/pre>\n\n<p>This is mine: <a href=\"https:\/\/tva1.sinaimg.cn\/large\/006y8mN6ly1g77gj8vie2j30m808ut8y.jpg\" rel=\"nofollow noreferrer\">https:\/\/tva1.sinaimg.cn\/large\/006y8mN6ly1g77gj8vie2j30m808ut8y.jpg<\/a><\/p>\n\n<p>And this what I want: <a href=\"https:\/\/i.stack.imgur.com\/LraT7.png\" rel=\"nofollow noreferrer\">https:\/\/i.stack.imgur.com\/LraT7.png<\/a><\/p>\n","answers":[]}
+{"QuestionId":58040789,"AnswerCount":1,"Tags":"<python><tensorflow><keras><casting>","CreationDate":"2019-09-21T13:31:59.723","AcceptedAnswerId":58041364.0,"OwnerUserId":12099555.0,"Title":"Custom binary crossentropy loss in keras that ignores columns with no non-zero values","Body":"<p>I'm trying to segment data where the label can be quite sparse. Therefore I want to only calculate gradients in columns that have at least one nonzero value.<\/p>\n\n<p>I've tried some methods where I apply an extra input which is the mask of these nonzero columns, but given that all the necessary information already is contained in <code>y_true<\/code>, a method which only looks at <code>y_true<\/code> to find the mask would definitely be preferable.<\/p>\n\n<p>If I would implement it with numpy, it would probably look something like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def loss(y_true, y_pred):\n    indices = np.where(np.sum(y_true, axis=1) &gt; 0)\n    return binary_crossentropy(y_true[indices], y_pred[indices])\n<\/code><\/pre>\n\n<p><code>y_true<\/code> and <code>y_pred<\/code> are in this example vectorized 2D images.<\/p>\n\n<p>How could this be \"translated\" to a differentiable Keras loss function?<\/p>\n","answers":[{"AnswerId":"58041364","CreationDate":"2019-09-21T14:46:22.633","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>Use <code>tf<\/code>-compatible operations, via <code>tf<\/code> and <code>keras.backend<\/code>:<\/p>\n\n<pre><code>import tensorflow as tf\nimport keras.backend as K\nfrom keras.losses import binary_crossentropy\n\ndef custom_loss(y_true, y_pred):\n    indices = K.squeeze(tf.where(K.sum(y_true, axis=1) &gt; 0))\n    y_true_sparse = K.cast(K.gather(y_true, indices), dtype='float32')\n    y_pred_sparse = K.cast(K.gather(y_pred, indices), dtype='float32')\n    return binary_crossentropy(y_true_sparse, y_pred_sparse) # returns a tensor\n<\/code><\/pre>\n\n<p>I'm unsure about the exact dimensionality specs of your problem, but loss must evaluate to a single value - which above doesn't, since you're passing multi-dimensional predictions and labels. To reduce dims, wrap the return above with e.g. <code>K.mean<\/code>. Example:<\/p>\n\n<pre><code>y_true = np.random.randint(0,2,(10,2))\ny_pred = np.abs(np.random.randn(10,2))\ny_pred \/= np.max(y_pred) # scale between 0 and 1\n\nprint(K.get_value(custom_loss(y_true, y_pred))) # get_value evaluates returned tensor\nprint(K.get_value(K.mean(custom_loss(y_true, y_pred))\n&gt;&gt; [1.1489482  1.2705883  0.76229745  5.101402  3.1309896] # sparse; 5 \/ 10 results\n&gt;&gt; 2.28284 # single value, as required\n<\/code><\/pre>\n\n<p>(Lastly, note that this sparsity will bias the loss by excluding all-zero columns from the total label\/pred count; if undesired, you can average via <code>K.sum<\/code> and <code>K.shape<\/code> or <code>K.size<\/code>)<\/p>\n"}]}
+{"QuestionId":58040996,"AnswerCount":1,"Tags":"<keras><neural-network><time-series><lstm>","CreationDate":"2019-09-21T14:00:03.550","AcceptedAnswerId":null,"OwnerUserId":11212687.0,"Title":"input_shape in LSTM","Body":"<p>I have the following code snippet:<\/p>\n\n<pre><code>    X_train = np.reshape(X_train, (X_train.shape[0], 1, X_train.shape[1]))\n    X_test = np.reshape(X_test, (X_test.shape[0], 1, X_test.shape[1]))\n\n\n    model = Sequential()\n    model.add(LSTM(200, activation='relu', input_shape=(X_train.shape[0], 1, X_train.shape[2]), return_sequences=False))\n    model.compile(optimizer='adam', loss='mean_squared_error')\n    model.fit(X_train, y_train, epochs=100, batch_size=32)\n    y_pred = model.predict(X_test)\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, found ndim=4\n<\/code><\/pre>\n\n<p>The original shapes of X_train and X_test:<\/p>\n\n<pre><code>X_train: 1483, 13\nX_test: 360, 13\n<\/code><\/pre>\n\n<p>and after reshaping they become:<\/p>\n\n<pre><code>X_train: 1483, 1, 13\nX_test: 360, 1, 13\n<\/code><\/pre>\n\n<p>I know this might be a duplicate, but none of the answers online seem to work for me.<\/p>\n","answers":[{"AnswerId":"58043390","CreationDate":"2019-09-21T19:18:31.067","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p><code>input_shape=(X_train.shape[0], 1, X_train.shape[2])<\/code> is wrong. <\/p>\n\n<p>An LSTM should have 2D input shapes (which means 3D internal tensors).<br>\nl\n- The input shape must contain <code>(sequence_length, features_per_step)<\/code>.<br>\n- This means the internal shape will be <code>(free_batch_size, sequence_length, features_per_step)<\/code>    <\/p>\n\n<p>Your data then must be 3D, ok, but the <code>input_shape<\/code> should be 2D.<br>\nNow, <code>sequence_length<\/code> is absolutely necessary for a recurrent layer to work, if you have <code>sequence_length = 1<\/code> it's totaly useless, unless you are going for <code>stateful=True<\/code> which involves much more complicated code.<\/p>\n"}]}
+{"QuestionId":58041017,"AnswerCount":1,"Tags":"<python><keras><model>","CreationDate":"2019-09-21T14:01:52.133","AcceptedAnswerId":null,"OwnerUserId":6896705.0,"Title":"How to deal with the model output to compute the loss","Body":"<p>I am struggling to train the model in Keras, by minimizing the loss between the correct data and \"input*output\", but do not know how to deal with it.<\/p>\n\n<p>Given that<\/p>\n\n<p>X: model input (training data)<\/p>\n\n<p>Y: model output<\/p>\n\n<p>T: correct data<\/p>\n\n<p>model = Model(inputs=X, outputs=Y)<\/p>\n\n<p>Then, in my understanding,\nmodel.fit(X,T) trains the model to minimize the distance between Y(=model(X)) and T, according to the user-defined loss function.<\/p>\n\n<p>My question is:\nWhat if I want to minimize the distance between Y*X and T?<\/p>\n\n<p>I thought writing such as \"model.fit(X * model.predict(X), T)\" would work well? (It did not, actually)\nI wonder how to write the code to do that.<\/p>\n\n<p>Thank you for the advice in advance.<\/p>\n","answers":[{"AnswerId":"58043411","CreationDate":"2019-09-21T19:21:26.917","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Make a functional API model:<\/p>\n\n<pre><code>inputs = Input(input_shape)\noutputs = SomeLayer(...)(inputs)\noutputs = SomeLayer(...)(outputs)\noutputs = SomeLayer(...)(outputs)\n....\n\noutputs = Multiply()([inputs, outputs])\n\nmodel = Model(inputs, outputs)\nmodel.compile(optimizer=optimizer, loss=loss, metrics=metrics)\n\nmodel.fit(X, T, ...)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58041183,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-21T14:23:24.913","AcceptedAnswerId":null,"OwnerUserId":12099719.0,"Title":"Are weights and trainable variables the same in Keras?","Body":"<p>That's all, I need to know if they are the same or correspond to different concepts.<\/p>\n\n<p>When I use the <code>model.summary()<\/code> method, it gives me the amount of trainable variables, I need to know if they are the same as the weights<\/p>\n","answers":[{"AnswerId":"58043446","CreationDate":"2019-09-21T19:26:38.950","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>\"Almost\".<\/p>\n\n<p>In most cases, yes, they are. But there are layers that use non-trainable weights for other purposes.   <\/p>\n\n<p>For instance, a <code>BatchNormalization<\/code> layers has four weight variables:<\/p>\n\n<ul>\n<li>mean: not trainable with backpropagation, but learnable from taking statistics from the data    <\/li>\n<li>variance: not trainable with backpropagation, but learnable from taking statistics from  the data   <\/li>\n<li>scale: trainable with backpropagation    <\/li>\n<li>offset: trainable with backpropagation    <\/li>\n<\/ul>\n"}]}
+{"QuestionId":58041187,"AnswerCount":1,"Tags":"<python><keras><deep-learning>","CreationDate":"2019-09-21T14:23:53.213","AcceptedAnswerId":58047808.0,"OwnerUserId":11973552.0,"Title":"How to repeat Input layer until it exceeds a certain amount of neurons?","Body":"<p>the shape of my Input layer is (None,), and I want to repeat the Input until it exceeds a certain amount. (for example until exceeds (7,)) <\/p>\n\n<p>here is my code:<\/p>\n\n<pre><code>import keras\n\nleast_accepted = 7 #the least accepted length\nInpt_lyr = keras.Input(shape = (None,))\nnxt_lyr = keras.layers.RepeatVector(least_accepted)(Inpt_lyr)\nnxt_lyr = keras.layers.Reshape((-1,))\n# and so on...\n<\/code><\/pre>\n\n<p>but i'm not sure if it is the optimized way. for example, if the minimum  accepted length  is 7, i want for inputs with length 1, 2, 3, 4, 5, 6, results be like these:\n1 -> 7, 2 -> 8, 3 -> 9, 4 -> 8, 5 -> 10, 6 -> 12. that is their first multiples which is greater than 7.\nany help is really appreciated...<\/p>\n","answers":[{"AnswerId":"58047808","CreationDate":"2019-09-22T09:18:08.367","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>The function:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.layers import *\nfrom keras.models import Model\nimport keras.backend as K\nfrom keras.backend import tf\nimport numpy as np\n\nminLength = 7\n\ndef expandData(x):\n    one = K.ones((1,), dtype='int32')\n    repetitions = K.cast(tf.ceil(minLength \/ K.shape(x)[-1:]), 'int32')\n    repetitions = K.concatenate([one, repetitions])\n    return K.tile(x, repetitions)\n<\/code><\/pre>\n\n<p>A working test: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>inputs = Input((None,))\noutputs = Lambda(expandData)(inputs)\nmodel = Model(inputs, outputs)\n\ninputData = [np.arange(i+1).reshape((1,-1)) for i in range(10)]\nfor inputs in inputData:\n    preds = model.predict(inputs)\n\n    print('\\ninput sequence - len=', len(inputs[0]), ' - ', inputs)\n    print('output sequence - len=', len(preds[0]), ' - ',  preds)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58041403,"AnswerCount":1,"Tags":"<python><arrays><numpy><pytorch>","CreationDate":"2019-09-21T14:52:25.637","AcceptedAnswerId":58041648.0,"OwnerUserId":1975277.0,"Title":"Pytorch's stack() adds dimension?","Body":"<p>I am looking to stack one <code>shape=(1,2)<\/code> tensor array on top of another <code>shape=(1,2)<\/code> array, across <code>dim=1<\/code>, using pytorch's <code>stack()<\/code> method.<\/p>\n\n<pre><code>&gt;&gt;&gt; import numpy as np\n&gt;&gt;&gt; import torch\n&gt;&gt;&gt; np_a = np.array([[1,2]])\n&gt;&gt;&gt; np_b = np.array([[3,4]])\n&gt;&gt;&gt; print(np_a)\n[[1 2]]\n&gt;&gt;&gt; print(np_b)\n[[3 4]]\n&gt;&gt;&gt; t_a = torch.from_numpy(np_a)\n&gt;&gt;&gt; t_b = torch.from_numpy(np_b)\n&gt;&gt;&gt; print(t_a)\ntensor([[1, 2]])\n&gt;&gt;&gt; print(t_b)\ntensor([[3, 4]])\n&gt;&gt;&gt; t_stacked = torch.stack((t_a, t_b), dim=1)\n&gt;&gt;&gt; print(t_stacked)\ntensor([[[1, 2],\n         [3, 4]]])\n<\/code><\/pre>\n\n<p>The resulting tensor has an added dimension and now has a <code>shape=(1,2,2)<\/code>.  Why doesn't pytorch's <code>stack()<\/code> behave like numpy's <code>vstack()<\/code>?  See below:<\/p>\n\n<pre><code>&gt;&gt;&gt; import numpy as np\n&gt;&gt;&gt; np_a = np.array([[1,2]])\n&gt;&gt;&gt; np_b = np.array([[3,4]])\n&gt;&gt;&gt; stacked = np.vstack((np_a, np_b))\n&gt;&gt;&gt; print(stacked)\n[[1 2]\n [3 4]]\n<\/code><\/pre>\n\n<p>How do I make pytorch not add a dimension?<\/p>\n","answers":[{"AnswerId":"58041648","CreationDate":"2019-09-21T15:26:40.693","ParentId":null,"OwnerUserId":"3014199","Title":null,"Body":"<p>I can't tell you why it was decided to have <code>pytorch<\/code>'s <code>stack<\/code> behave differently from numpy (maybe compatibility with luatorch?). Anyway, to get the desired outcome you can use <code>torch.cat<\/code>:<\/p>\n\n<pre><code>&gt;&gt;&gt; torch.cat((t_a, t_b), dim=0)     \ntensor([[1, 2],\n        [3, 4]])\n<\/code><\/pre>\n"}]}
+{"QuestionId":58041611,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-21T15:21:24.413","AcceptedAnswerId":null,"OwnerUserId":11013499.0,"Title":"how can I cropout some part of a tensor?","Body":"<p>I want to crop some parts of a tensor and assign these parts with zero. for this aim, I used the following code, but it produces the following error. how can solve this problem? I am using Keras and I have a CNN network. x is the output of the convolution layer and a tensor.\nThanks.<\/p>\n\n<pre><code>def diff_cropout(x, rate=0.6):\n    global block_size\n    cover_img =x[0]\n    Iw = x[1]\n    diff = Iw - cover_img\n    diff_shape = diff.get_shape().as_list()\n    h, w = diff_shape[1:3]\n\n    blacked_pixels = int(rate * (h*w))\n    height = int(np.floor(np.sqrt(blacked_pixels)))\n    width = int(np.ceil(np.sqrt(blacked_pixels)))\n\n    x = np.random.randint(0, w - width)\n    y = np.random.randint(0, h - height)\n\n    new_diff = diff\n    new_diff = new_diff[y:y+height, x:x+width].assign(tf.zeros([height, width]))\n    new_Iw = cover_img + new_diff\n\n    return new_I\n<\/code><\/pre>\n\n<p>in my network:<\/p>\n\n<pre><code>diff_cropout_attacked = layers.Lambda(diff_cropout, arguments={'rate':diff_cropRate}, name='diff_cropout_attack')([image, actimg])\n<\/code><\/pre>\n\n<p>the error:<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):<\/p>\n  \n  <p>File \"\", line 1, in \n      runfile('E:\/New folder\/testattack.py',\n  wdir='E:\/New folder')<\/p>\n  \n  <p>File\n  \"D:\\software\\Anaconda3\\envs\\py36\\lib\\site-packages\\spyder_kernels\\customize\\spydercustomize.py\",\n  line 704, in runfile\n      execfile(filename, namespace)<\/p>\n  \n  <p>File\n  \"D:\\software\\Anaconda3\\envs\\py36\\lib\\site-packages\\spyder_kernels\\customize\\spydercustomize.py\",\n  line 108, in execfile\n      exec(compile(f.read(), filename, 'exec'), namespace)<\/p>\n  \n  <p>File \"E:\/Newfolder\/testattack.py\", line 319, in \n      diff_cropout_attacked = layers.Lambda(diff_cropout, arguments={'rate':diff_cropRate},\n  name='residual_cropout_attack')([image, actimg])<\/p>\n  \n  <p>File\n  \"D:\\software\\Anaconda3\\envs\\py36\\lib\\site-packages\\keras\\engine\\base_layer.py\",\n  line 457, in <strong>call<\/strong>\n      output = self.call(inputs, **kwargs)<\/p>\n  \n  <p>File\n  \"D:\\software\\Anaconda3\\envs\\py36\\lib\\site-packages\\keras\\layers\\core.py\",\n  line 687, in call\n      return self.function(inputs, **arguments)<\/p>\n  \n  <p>File \"E:\/Newfolder\/testattack.py\", line 199, in diff_cropout\n      new_diff = new_diff[y:y+height, x:x+width].assign(tf.zeros([height, width]))<\/p>\n  \n  <p>File\n  \"D:\\software\\Anaconda3\\envs\\py36\\lib\\site-packages\\tensorflow\\python\\ops\\array_ops.py\",\n  line 700, in assign\n      raise ValueError(\"Sliced assignment is only supported for variables\")<\/p>\n  \n  <p>ValueError: Sliced assignment is only supported for variables<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":58041868,"AnswerCount":0,"Tags":"<python><tensorflow><keras><neural-network>","CreationDate":"2019-09-21T15:57:29.790","AcceptedAnswerId":null,"OwnerUserId":6296435.0,"Title":"NN: Why does the manual implementation converges better than Keras\/TF?","Body":"<p>I've <a href=\"https:\/\/github.com\/MaverickMeerkat\/MachineLearningPython\/blob\/master\/ex4\/NN.py\" rel=\"nofollow noreferrer\">implemented<\/a> a basic 1-hidden layer NN to identify digits from part of the MNIST dataset (This is adapted from the Coursera ML course). It is implemented using matrices and conjugated-gradient (fmincg) optimization. <\/p>\n\n<p>In addition I've implemented both a <a href=\"https:\/\/github.com\/MaverickMeerkat\/MachineLearningPython\/blob\/master\/ex4\/KerasImplementation.py\" rel=\"nofollow noreferrer\">Keras implementation<\/a> and a vanilla <a href=\"https:\/\/github.com\/MaverickMeerkat\/MachineLearningPython\/blob\/master\/ex4\/TFImplementation.py\" rel=\"nofollow noreferrer\">TF implementation<\/a>. <\/p>\n\n<p>In the manual implementation, there's 50 iterations (or epochs, but since I'm using the whole batch, each epoch is of size 1).<\/p>\n\n<p>Now, what I've noticed, is that the manual implementation takes a while to compute, but after 50 iterations gives me really good results.<\/p>\n\n<p>Keras and TF on the other hand, takes less time to compute, but need about 500-1000 iterations if I'm using the whole batch to give me the same results. (Though using 50 epochs, with each processing mini-batches of size 32, seems to work good as well) <\/p>\n\n<p>So I wonder, how come the manual impl. converges better? And why does it take longer to compute even though it's the same number of iterations (50)? <\/p>\n","answers":[]}
+{"QuestionId":58042219,"AnswerCount":0,"Tags":"<python><tensorflow><anaconda>","CreationDate":"2019-09-21T16:39:58.187","AcceptedAnswerId":null,"OwnerUserId":9914673.0,"Title":"Import module error in anaconda with tensorflow","Body":"<p>I have installed tensorflow in anaconda.  However, When I tried to build a python file I have a build error:<\/p>\n\n<pre><code>ImportError:No module named 'tensorflow.python.eager'.  \n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/CV2Er.png\" rel=\"nofollow noreferrer\">example<\/a><\/p>\n\n<p>Which library do I need to install to solve this problem.  Thanks.<\/p>\n","answers":[]}
+{"QuestionId":58042592,"AnswerCount":1,"Tags":"<python><keras><deep-learning>","CreationDate":"2019-09-21T17:31:33.407","AcceptedAnswerId":58042918.0,"OwnerUserId":12036469.0,"Title":"keras .h5 file + other data use","Body":"<p>Thanks to all Stackoverflow users.<\/p>\n\n<p>I have a question about a .h5 file.\nFor example, data [A, B, C, D, E, Train] data.<\/p>\n\n<p>Train with [Train + A] data.\nBinary classifications with tags Good and Bad respectively.<\/p>\n\n<p>Save the verified .h5 file.<\/p>\n\n<p>And can you derive meaningful predictions later from the .h5 file? (Model.predict method)<\/p>\n\n<p>In other words, does the prediction value of [.h5 (model) + B] data make sense?<\/p>\n\n<p>B differs only in data, but the classification table is the same.[Good,Bad]<\/p>\n\n<p>If this is not possible, do we have to retrain to determine the B data set?<\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"58042918","CreationDate":"2019-09-21T18:13:43.647","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>You can save a trained model, and load it to use for predictions:<\/p>\n\n<pre><code>from keras.models import load_model\n\nmodel.save(\".\/model1.h5\") # SAVE\nmodel = load_model(\".\/model1.h5\") # LOAD (already compiled)\n<\/code><\/pre>\n\n<p>This will load the model architecture, optimizer state, and weights. You can then use the model to run inference: <code>model.predict(input_data)<\/code><\/p>\n\n<p>To save only the <em>weights<\/em> of the model, see below; when loading, be sure to <em>compile<\/em> the model after, and that the model has the same layer weights. <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-save-a-keras-model\" rel=\"nofollow noreferrer\">Further reading<\/a><\/p>\n\n<pre><code>model.save_weights(\".\/model1_weights.h5\") # SAVE\nmodel.load_weights(\".\/model1_weights.h5\") # LOAD\nmodel.compile(...) # COMPILE\n<\/code><\/pre>\n"}]}
+{"QuestionId":58042674,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-21T17:42:06.740","AcceptedAnswerId":58043152.0,"OwnerUserId":1042017.0,"Title":"How to get output from randomly sampled k entries from a tensor","Body":"<p>I have a keras\/tf problem using sub-sampling of values from a tensor. My model is given below:<\/p>\n\n<pre><code>x_input = Input((input_size,))\nenc1 = Dense(encoder_size[0], activation='relu')(x_input)\ndrop = Dropout(keep_prob)(enc1)\nenc2 = Dense(encoder_size[1], activation='relu')(drop)\ndrop = Dropout(keep_prob)(enc2)\nmu = Dense(latent_dim, activation='linear', name='encoder_mean')(drop)\nencoder = Model(x_input,mu)\n<\/code><\/pre>\n\n<p>I want to sample from the input randomly and then get the encoded values of the input. The error I am getting is <\/p>\n\n<p><code>ValueError: When feeding symbolic tensors to a model, we expect the tensors to have a static batch size. Got tensor with shape: (None, 13)<\/code> <\/p>\n\n<p>which I can understand is because \"predict\" does not work on placeholder but I am not sure what to pass to get the output for a placeholder. <\/p>\n\n<pre><code># sample input randomly\nsample_num = 500\nidxs = tf.range(tf.shape(x_input)[0])\nridxs = tf.random_shuffle(idxs)[:sample_num]\nsample_input = tf.gather(x_input, ridxs)\n# get sample shape\nsample_shape = K.shape(sample_input)\n# sample from encoded value\nsample_encoded = encoder.predict(sample_input) &lt;----- Error\n<\/code><\/pre>\n","answers":[{"AnswerId":"58043152","CreationDate":"2019-09-21T18:45:46.547","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>If you see the <a href=\"https:\/\/keras.io\/models\/model\/#predict\" rel=\"nofollow noreferrer\"><code>predict<\/code><\/a> function documentation, it doesn't expect a placeholder or a tensor node as an expected set of input. You have to pass directly the Numpy array (in your case).<\/p>\n\n<p>If you wish to perform some special data preprocessing which is not part of your regular model, you have to do it in Numpy and avoid Tensor computations for it.<\/p>\n"}]}
+{"QuestionId":58042801,"AnswerCount":0,"Tags":"<python><tensorflow><stochastic-gradient>","CreationDate":"2019-09-21T18:00:18.800","AcceptedAnswerId":null,"OwnerUserId":10516607.0,"Title":"How many times does tf.train.GradientDescentOptimizer update weights","Body":"<p>As illustrated in the title, does op = tf.train.GradientDescentOptimizer.minimize only update the network weights once if I call sess.run(op) for one iteration? If the input batch_size is not equal to one, does it still update once by all the samples in each batch?<\/p>\n","answers":[]}
+{"QuestionId":58042955,"AnswerCount":0,"Tags":"<keras><lstm>","CreationDate":"2019-09-21T18:18:21.020","AcceptedAnswerId":null,"OwnerUserId":5638494.0,"Title":"A ConvLSTM2DCell's error: 'tuple' object has no attribute '_keras_shape'","Body":"<p>I'm setting up a new Model which need to use Keras' ConvLSTM2DCells.\nbut I've encountered a error that says <code>'tuple' object has no attribute '_keras_shape'<\/code><\/p>\n\n<p>This is the snippet of my code:<\/p>\n\n<pre><code>inputs = keras.Input(shape=(64, 32, 4))\nh0 = keras.Input(shape=(64, 32, 64))\nCL2DCell = keras.layers.ConvLSTM2DCell(filters=64, kernel_size=(3, 3), padding='same')\nx, [h, c] = CL2DCell(inputs, states=[h0, h0])\n<\/code><\/pre>\n\n<p>and Here're the additional detail.<\/p>\n\n<pre><code>    563         # Update tensor history, _keras_shape and _uses_learning_phase.\n    564         for i in range(len(output_tensors)):\n--&gt; 565             output_tensors[i]._keras_shape = output_shapes[i]\n    566             uses_lp = any(\n    567                 [getattr(x, '_uses_learning_phase', False)\n\nAttributeError: 'tuple' object has no attribute '_keras_shape'\n<\/code><\/pre>\n\n<p>Thank you.<\/p>\n","answers":[]}
+{"QuestionId":58043002,"AnswerCount":1,"Tags":"<python><keras><conv-neural-network><recurrent-neural-network>","CreationDate":"2019-09-21T18:25:34.187","AcceptedAnswerId":58044089.0,"OwnerUserId":7003616.0,"Title":"Training and Testing accuracy not increasing for a CNN followed by a RNN for signature verification","Body":"<p>I'm currently working on online signature verification. The dataset has a variable shape of (x, 7) where x is the number of points a person used to sign their signature. I have the following model:<\/p>\n\n<pre><code>    model = Sequential()\n    #CNN\n    model.add(Conv1D(filters=64, kernel_size=3, activation='sigmoid', input_shape=(None, 7)))\n    model.add(MaxPooling1D(pool_size=3))\n    model.add(Conv1D(filters=64, kernel_size=2, activation='sigmoid'))\n\n    #RNN\n    model.add(Masking(mask_value=0.0))\n    model.add(LSTM(8))\n    model.add(Dense(2, activation='softmax'))\n\n    opt = Adam(lr=0.0001)\n    model.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['accuracy'])\n    model.summary()\n\n    print(model.fit(x_train, y_train, epochs=100, verbose=2, batch_size=50))\n\n    score, accuracy = model.evaluate(x_test,y_test, verbose=2)\n    print(score, accuracy)\n<\/code><\/pre>\n\n<p>I know it may not be the best model but this is the first time I'm building a neural network. I have to use a CNN and RNN as it is required for my honours project. At the moment, I achieve 0.5142 as the highest training accuracy and 0.54 testing accuracy. I have tried increasing the number of epochs, changing the activation function, add more layers, moving the layers around, changing the learning rate and changing the optimizer. <\/p>\n\n<p>Please share some advice on changing my model or dataset. Any help is much appreciated.  <\/p>\n","answers":[{"AnswerId":"58044089","CreationDate":"2019-09-21T21:05:51.713","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>For CNN-RNN, some promising things to try:<\/p>\n\n<ul>\n<li><strong>Conv1D layers<\/strong>: <code>activation='relu'<\/code>, <code>kernel_initializer='he_normal'<\/code><\/li>\n<li><strong>LSTM layer<\/strong>: <code>activation='tanh'<\/code>, and <code>recurrent_dropout=.1, .2, .3<\/code><\/li>\n<li><strong>Optimizer<\/strong>: <code>Nadam<\/code>, <code>lr=2e-4<\/code> (Nadam may significantly outperform all other optimizers for RNNs)<\/li>\n<li><strong>batch_size<\/strong>: lower it. Unless you have 200+ batches in total, set <code>batch_size=32<\/code>; lower batch size better exploits the Stochastic mechanism of the optimizer and can improve generalization<\/li>\n<li><strong>Dropout<\/strong>: right after second <code>Conv1D<\/code>, with a rate <code>.1, .2<\/code> - or, after first <code>Conv1D<\/code>, with a rate <code>.25, .3<\/code>, but <em>only<\/em> if you use SqueezeExcite (see below), else <code>MaxPooling<\/code> won't work as well<\/li>\n<li><a href=\"https:\/\/arxiv.org\/abs\/1709.01507\" rel=\"nofollow noreferrer\">SqueezeExcite<\/a>: shown to enhance all CNN performance across a large variety of tasks; Keras implementation you can use below<\/li>\n<li><strong>BatchNormalization<\/strong>: while your model isn't large, it's still deep, and may benefit from one BN layer right after second <code>Conv1D<\/code><\/li>\n<li><strong>L2 weight decay<\/strong>: on <em>first<\/em> <code>Conv1D<\/code>, to prevent it from memorizing the input; try <code>1e-5, 1e-4<\/code>, e.g. <code>kernel_regularizer=l2(1e-4) # from keras.regularizers import l2<\/code><\/li>\n<li><strong>Preprocessing<\/strong>: make sure all data is normalized (or standardized if time-series), and batches are shuffled each epoch<\/li>\n<\/ul>\n\n<pre><code>def SqueezeExcite(_input):\n    filters = _input._keras_shape[-1]\n\n    se = GlobalAveragePooling1D()(_input)\n    se = Reshape((1, filters))(se)\n    se = Dense(filters\/\/16,activation='relu',   \n               kernel_initializer='he_normal', use_bias=False)(se)\n    se = Dense(filters,    activation='sigmoid',\n               kernel_initializer='he_normal', use_bias=False)(se)\n\n    return multiply([_input, se])\n<\/code><\/pre>\n\n<pre><code># Example usage\nx = Conv1D(filters=64, kernel_size=4, activation='relu', kernel_initializer='he_normal')(x)\nx = SqueezeExcite(x) # place after EACH Conv1D\n<\/code><\/pre>\n"}]}
+{"QuestionId":58043606,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-21T19:50:00.553","AcceptedAnswerId":58043830.0,"OwnerUserId":2607447.0,"Title":"Keras predicts same result for any input","Body":"<p>I'm learning machine learning and my dataset consists from 6 columns:<\/p>\n\n<pre><code>home_team, away_team, home_odds, away_odds, home_score, away_score, 1_if_home_wins_else_0\n<\/code><\/pre>\n\n<p>To be able to feed the Tensorflow with teams, I converted every team to integer so the first two columns are integers (like database ids)<\/p>\n\n<p>There is a 10k rows in the csv.<\/p>\n\n<p><strong>example<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/jMUjp.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/jMUjp.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Now I'm trying to modify the code for <a href=\"https:\/\/machinelearningmastery.com\/tutorial-first-neural-network-python-keras\/\" rel=\"nofollow noreferrer\">pima indians diabetes<\/a> to predict <strong>winnings<\/strong> of home team.<\/p>\n\n<p>But it returns the same prediction (0) to any input. When I tried it only on odds, the predictions were more accurate and not the same at least.<\/p>\n\n<p><strong>code<\/strong><\/p>\n\n<pre><code># load the dataset\ndataset = loadtxt('football_data.csv', delimiter=',')\n# split into input (X) and output (y) variables\nX = dataset[:, 0:4]\ny = dataset[:, 6]\n# define the keras model\nmodel = Sequential()\nmodel.add(Dense(12, input_dim=4, activation='relu'))\nmodel.add(Dense(8, activation='relu'))\nmodel.add(Dense(1, activation='sigmoid'))\n# compile the keras model\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n# fit the keras model on the dataset\nmodel.fit(X, y, epochs=150, batch_size=10)\n# evaluate the keras model\n_, accuracy = model.evaluate(X, y)\nprint('Accuracy: %.2f' % (accuracy * 100))\n\n# make class predictions with the model\npredictions = model.predict_classes(X)\n# summarize the first 5 cases\nfor i in range(50):\n    print('%s =&gt; %d (expected %d)' % (X[i].tolist(), predictions[i], y[i]))\n<\/code><\/pre>\n\n<p>Do you know where is the problem?<\/p>\n","answers":[{"AnswerId":"58043830","CreationDate":"2019-09-21T20:23:09.473","ParentId":null,"OwnerUserId":"9149062","Title":null,"Body":"<p>The fact that you convert the first two columns (team names) to integers does not make any sense. This way you would be implying that teams that have similar IDs, such as <code>1146<\/code> and <code>1179<\/code>, will perform similar and that teams with completely different  IDs, such as <code>4<\/code> and <code>6542<\/code>, would perform very differently. Usually this kind of data would be presented in a different manner or even excluded from the dataset.<\/p>\n\n<p>I would exclude those columns in this case since the odds seem to contain all necessary data, I wouldn't even use neural networks for this but just compare the odds. However I understand that you want to use a simple dataset for learning purposes in which case only using the odds would be fine.<\/p>\n\n<p>Mind though that the neural network will probably learn to assign the win to the team with the biggest odds of winning, like the following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>if home_odds &gt; away_odds:\n    one_if_home_wins_else_zero = 1\nelse:\n    one_if_home_wins_else_zero = 0\n<\/code><\/pre>\n"}]}
+{"QuestionId":58043637,"AnswerCount":0,"Tags":"<tensorflow><keras><deep-learning><binary-reproducibility>","CreationDate":"2019-09-21T19:54:56.817","AcceptedAnswerId":null,"OwnerUserId":7065438.0,"Title":"Reproducible results with Keras with Tensorflow background","Body":"<p>I have my own network. But it is giving me different outputs each time I run the code.  I'm using keras (with Tensorflow backend), write the following code for reproducibility. My training sample: 280, validation sample # 27, test sample # 21. <\/p>\n\n<pre><code># The following  lines are for reproducibility\nimport os\nimport random as rn\nos.environ['PYTHONHASHSEED'] = '0'\n# random seed for NP genreator of ranodm numbers\nnp.random.seed(37) \nrn.seed(1254)      # specifying the seed for python-generated random     numbers:\n\nimport tensorflow as tf\ntf.compat.v1.set_random_seed(89)    #tf.set_random_seed(89)\n\nimport keras.backend.tensorflow_backend as K\nsession_conf =  tf.compat.v1.ConfigProto(intra_op_parallelism_threads=1,    inter_op_parallelism_threads=1)\nsess=tf.compat.v1.Session(graph=tf.compat.v1.get_default_graph(),    config= session_conf)\n K.set_session(sess)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58043843,"AnswerCount":1,"Tags":"<r><tensorflow><keras>","CreationDate":"2019-09-21T20:24:41.763","AcceptedAnswerId":58044175.0,"OwnerUserId":2155179.0,"Title":"Installing Keras\/Tensor Flow on R","Body":"<p>I am trying to install tensorflow on R. As far as I understand I have already installed successlly many times but it won't load<\/p>\n\n<p>Everytime I run a function from Tensorflow and Keras it does not work. <\/p>\n\n<pre><code>&gt; testlabels=to_categorical(trainy)\nError: Python module tensorflow.python.keras was not found.\n\nDetected Python configuration:\n\npython:         C:\\Users\\AUSTER~1\\ANACON~1\\python.exe\nlibpython:      C:\/Users\/AUSTER~1\/ANACON~1\/python37.dll\npythonhome:     C:\\Users\\AUSTER~1\\ANACON~1\nversion:        3.7.3 (default, Apr 24 2019, 15:29:51) [MSC v.1915 64 bit (AMD64)]\nArchitecture:   64bit\nnumpy:          C:\\Users\\AUSTER~1\\ANACON~1\\lib\\site-packages\\numpy\nnumpy_version:  1.16.4\ntensorflow:     [NOT FOUND]\n<\/code><\/pre>\n\n<p>When I load the libraries I have no issues. They all load up sucessfully.<\/p>\n\n<pre><code>library(EBImage)\nlibrary(keras)\nlibrary(tensorflow)\n\n\nI have ran all these following commands.\ninstall.packages(\"keras\")\nsource(\"http:\/\/bioconductor.org\/biocLite.R\")\nbiocLite(\"EBImage\")\n\ninstall.packages(\"reticulate\")\ninstall.packages(\"remotes\")\nremotes::install_github(\"rstudio\/tensorflow\")\n\nreticulate::py_discover_config(\"keras\")\nreticulate::py_discover_config(\"tensorflow\")\n\ninstall.packages(\"devtools\")\ndevtools::install_github(\"rstudio\/keras\")\ntensorflow::install_tensorflow()\ndevtools::install_github(\"rstudio\/keras\")\n<\/code><\/pre>\n\n<p>What could be causing the issue. After I installed tensforflow it said \"<code>Installation Complete<\/code>\" and then restarted R. Is R simply unable to find the installation?<\/p>\n\n<p>I have also ran the following commands still with no luck. <\/p>\n\n<pre><code>library(tensorflow)\nSys.setenv(TENSORFLOW_PYTHON=\"\/usr\/local\/bin\/python\")\n<\/code><\/pre>\n\n<p><strong>Update<\/strong><\/p>\n\n<p>I was able to get Tensorflow\/Keras to work by opening an Anaconda command prompt and running the following commands. Then my R code using Keras\/Tensorflow started to work. <\/p>\n\n<pre><code>conda install -c conda-forge keras\n<\/code><\/pre>\n","answers":[{"AnswerId":"58044175","CreationDate":"2019-09-21T21:15:23.877","ParentId":null,"OwnerUserId":"4752675","Title":null,"Body":"<p>The keras package makes use of keras through python. When you install the package, it installs the R interface, but <em>not<\/em> the python package. You must separately install the python package. There is a nice guide for doing that on <a href=\"https:\/\/www.r-bloggers.com\/a-newbies-install-of-keras-tensorflow-on-windows-10-with-r\/\" rel=\"nofollow noreferrer\">RBloggers<\/a><\/p>\n"}]}
+{"QuestionId":58044133,"AnswerCount":1,"Tags":"<python-3.x><neural-network><pytorch>","CreationDate":"2019-09-21T21:11:01.617","AcceptedAnswerId":58046379.0,"OwnerUserId":1450312.0,"Title":"Training a model with multiple learning rate in PyTorch","Body":"<p>I am new to PyTorch and getting used to some concepts. <\/p>\n\n<p>I need to train a Neural Network. For optimization, I need to use Adam optimizer with 4 different <code>learning rates = [2e-5, 3e-5, 4e-5, 5e-5]<\/code><\/p>\n\n<p>The optimizer function is defined as below<\/p>\n\n<pre><code>def optimizer(no_decay = ['bias', 'gamma', 'beta'], lr=2e-5):\n    param_optimizer = list(model.named_parameters())\n    optimizer_grouped_parameters = [\n    {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],\n     'weight_decay_rate': 0.01},\n    {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)],\n     'weight_decay_rate': 0.0}\n     ]\n     # This variable contains all of the hyperparemeter information our training loop needs\n     optimizer = BertAdam(optimizer_grouped_parameters, lr, warmup=.1)\n     return optimizer\n<\/code><\/pre>\n\n<p>How do I make sure the optimizer uses my specified set of learning rate and returns the best model?<\/p>\n\n<p>During training, we use the optimizer like below where I don't see a way to tell it to try different learning rate<\/p>\n\n<pre><code>def model_train():\n    #other code\n    # clear out the gradient\n    optimizer.zero_grad()\n    # Forward pass\n    loss = model(b_input_ids, token_type_ids=None, attention_mask=b_input_mask, labels=b_labels)\n    train_loss_set.append(loss.item())\n    # Backward pass\n    loss.backward()\n    # Update parameters and take a step using the computed gradient\n    optimizer.step()\n<\/code><\/pre>\n\n<p>I know that the <code>optimizer.step()<\/code> internally steps through to optimize the gradient. But how do I make sure the optimizer tries my specified set of learning rate and returns the best model to me?<\/p>\n\n<p>Please suggest.<\/p>\n","answers":[{"AnswerId":"58046379","CreationDate":"2019-09-22T05:44:06.810","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you want to train four times with four different learning rates and then compare you need not only four <code>optimizer<\/code>s but also four <code>model<\/code>s: Using different learning rate (or any other meta-parameter for this matter) yields a different trajectory of the weights in the high-dimensional \"parameter space\". That is, after a few steps its not only the learning rate that differentiate between the models, but the trained weights themselves - this is what yield the actual difference between the models.  <\/p>\n\n<p>therefore, you need to train 4 times using 4 separate instances of <code>model<\/code> using 4 instances of <code>optimizer<\/code> with different learning rates.  <\/p>\n"}]}
+{"QuestionId":58044416,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-09-21T21:59:10.160","AcceptedAnswerId":58044625.0,"OwnerUserId":10521726.0,"Title":"How to correctly shape my CNN-LSTM input layer","Body":"<p>I have a data set with the shape (3340, 6). I want to use a CNN-LSTM to read a sequence of 30 rows and predict the next row's (6) elements. From what I have read, this is considered a multi-parallel time series. I have been primarily following this <a href=\"https:\/\/machinelearningmastery.com\/how-to-develop-lstm-models-for-time-series-forecasting\/\" rel=\"nofollow noreferrer\">machine learning mastery tutorial<\/a> and am having trouble implementing the CNN-LSTM architecture for a multi-parallel time series.<\/p>\n\n<p>I have used this function to split the data into 30 day time step frames<\/p>\n\n<pre><code># split a multivariate sequence into samples\ndef split_sequences(sequences, n_steps):\n    X, y = list(), list()\n    for i in range(len(sequences)):\n        # find the end of this pattern\n        end_ix = i + n_steps\n        # check if we are beyond the dataset\n        if end_ix &gt; len(sequences)-1:\n            break\n        # gather input and output parts of the pattern\n        seq_x, seq_y = sequences[i:end_ix, :], sequences[end_ix, :]\n        X.append(seq_x)\n        y.append(seq_y)\n    return array(X), array(y)\n<\/code><\/pre>\n\n<p>Here is a sample of the data frames produced by the function above.<\/p>\n\n<pre><code>   # 30 Time Step Input Frame X[0], X.shape = (3310, 30, 6)\n   [4.951e-02, 8.585e-02, 5.941e-02, 8.584e-02, 8.584e-02, 5.000e+00],\n   [8.584e-02, 9.307e-02, 7.723e-02, 8.080e-02, 8.080e-02, 4.900e+01],\n   [8.080e-02, 8.181e-02, 7.426e-02, 7.474e-02, 7.474e-02, 2.000e+01],\n   [7.474e-02, 7.921e-02, 6.634e-02, 7.921e-02, 7.921e-02, 4.200e+01],\n   ...\n\n   # 1 Time Step Output Array y[0], y.shape = (3310, 6)\n   [6.550e-02, 7.690e-02, 6.243e-02, 7.000e-02, 7.000e-02, 9.150e+02]\n<\/code><\/pre>\n\n<p>Here is the following model that I am using:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(TimeDistributed(Conv1D(64, 1, activation='relu'), input_shape=(None, 30, 6)))\nmodel.add(TimeDistributed(MaxPooling1D(pool_size=2)))\nmodel.add(TimeDistributed(Flatten()))\nmodel.add(LSTM(50, activation='relu', return_sequences=True))\nmodel.add(Dense(6))\nmodel.compile(optimizer='adam', loss='mse')\n<\/code><\/pre>\n\n<p>When I run model.fit, I receive the following error:<\/p>\n\n<pre><code>ValueError: Error when checking input: expected time_distributed_59_input to have \n4 dimensions, but got array with shape (3310, 30, 6)\n<\/code><\/pre>\n\n<p>I am at a loss at how to properly shape my input layer so that I can get this model learning. I have done several <code>Conv2D<\/code> nets in the past but this is my first time series model so I apologize if there's an obvious answer here that I am missing.<\/p>\n","answers":[{"AnswerId":"58044625","CreationDate":"2019-09-21T22:36:56.390","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<ul>\n<li>Remove <code>TimeDistributed<\/code> from <code>Conv1D<\/code> and <code>MaxPooling1D<\/code>; 3D inputs are supported<\/li>\n<li>Remove <code>Flatten()<\/code>, as it destroys <code>timesteps<\/code>-<code>channels<\/code> relationships<\/li>\n<li>Add <code>TimeDistributed<\/code> to the last <code>Dense<\/code> layer, as <code>Dense<\/code> does not support <code>3D<\/code> inputs (returned by <code>LSTM(return_sequences=True)<\/code>; alternatively, use <code>return_sequences=False<\/code>)<\/li>\n<\/ul>\n"}]}
+{"QuestionId":58044421,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-21T21:59:39.980","AcceptedAnswerId":58046396.0,"OwnerUserId":1042017.0,"Title":"Keras custom loss function with samples from complete input dataset","Body":"<p>I am trying to devise a custom loss function for Variational auto-encoder in Keras with two parts: <code>reconstruction loss<\/code> and <code>divergence loss<\/code>. However, instead of using the gaussian distribution for divergence loss, I want to sample randomly from the input and then perform the divergence loss based on the sampled inputs. However, I do not know how to sample inputs which are from the complete datastet and then perform a loss with respect to it. The encoder model is:<\/p>\n\n<pre><code>x_input = Input((input_size,))\nenc1 = Dense(encoder_size[0], activation='relu')(x_input)\ndrop = Dropout(keep_prob)(enc1)\nenc2 = Dense(encoder_size[1], activation='relu')(drop)\ndrop = Dropout(keep_prob)(enc2)\nmu = Dense(latent_dim, activation='linear', name='encoder_mean')(drop)\nencoder = Model(x_input,mu)\n<\/code><\/pre>\n\n<p>The structure of loss should be:<\/p>\n\n<pre><code># the input is the placeholder for the complete input\ndef loss(x, y, input):\n    reconstruction_loss = mean_squared_error(x, y)\n    sample_num = 100\n    sample_input = sample_from_input(input, sample_num)\n    sample_encoded = encoder.predict(sample_input) &lt;-- this would not work with placeholder\n    sample_prior = gaussian(mean=0, std=1)\n    # perform KL divergence between sample_encoded and sample_prior\n<\/code><\/pre>\n\n<p>I have not found anything similar given. It would be great if somebody can point me in the right direction. <\/p>\n","answers":[{"AnswerId":"58046396","CreationDate":"2019-09-22T05:48:45.343","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>There are couple of problems in your code. First, when you create your custom loss function, it expects only two (equivalent) parameters of <code>y_true<\/code> and <code>y_pred<\/code>. So you will not be able to pass explicitly the parameter of <code>input<\/code> in your case. If you wish to pass additional parameters, you have to use the concept of nested function.<\/p>\n\n<p>Next thing is inside <code>predict<\/code> function you will not be able to pass TensorFlow placeholders. You will have to pass Numpy array equivalents in it. So I would recommend you to rewrite your <code>sample_from_input<\/code> which samples from a set of file path inputs, reads it and sends a Numpy array of file data. Also, in the parameter of <code>input_data<\/code>, pass it the file paths where your data is present. <\/p>\n\n<p>I have enclosed only the relevant parts of code.<\/p>\n\n<pre><code>def custom_loss(input_data):\n  def loss(y_true, y_pred):\n    reconstruction_loss = mean_squared_error(x, y)\n    sample_num = 100\n    sample_input = sample_from_input(input_data)\n    # sample_input is a Numpy array\n    sample_encoded = encoder.predict(sample_input) \n    sample_prior = gaussian(mean=0, std=1)\n    # perform KL divergence between sample_encoded and sample_prior\n    divergence_loss = # Your logic returning a numeric value\n    return reconstruction_loss + divergence_loss\n  return loss\n\nencoder.compile(optimizer='adam', loss=custom_loss('&lt;&lt;input_data_path&gt;&gt;'))\n<\/code><\/pre>\n"}]}
+{"QuestionId":58044728,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><pytorch>","CreationDate":"2019-09-21T22:57:54.203","AcceptedAnswerId":58045560.0,"OwnerUserId":9856528.0,"Title":"Pytorch simple model not improving","Body":"<p>I am making a simple PyTorch neural net to approximate the sine function on x = [0, 2pi]. This is a simple architecture I use with different deep learning libraries to test whether I understand how to use it or not. The neural net, when untrained, always produces a straight horizontal line, and when trained, produces a straight line at y = 0. In general, it always produces a straight line at y = (The mean of the function). This leads me to believe something is wrong with the forward prop portion of it, as the boundary should not just be a straight line when untrained. Here is the code for the net:<\/p>\n\n<pre><code>class Net(nn.Module):\n    def __init__(self):\n      super(Net, self).__init__()\n      self.model = nn.Sequential(\n      nn.Linear(1, 20),\n      nn.Sigmoid(),\n      nn.Linear(20, 50),\n      nn.Sigmoid(),\n      nn.Linear(50, 50),\n      nn.Sigmoid(),\n      nn.Linear(50, 1)\n      )\n\n    def forward(self, x):\n        x = self.model(x)\n        return x\n<\/code><\/pre>\n\n<p>Here is the training loop<\/p>\n\n<pre><code>def train(net, trainloader, valloader, learningrate, n_epochs):\n    net = net.train()\n    loss = nn.MSELoss()\n    optimizer = torch.optim.SGD(net.parameters(), lr = learningrate)\n\n    for epoch in range(n_epochs):\n\n        for X, y in trainloader:\n            X = X.reshape(-1, 1)\n            y = y.view(-1, 1)\n            optimizer.zero_grad()\n\n            outputs = net(X)\n\n            error   = loss(outputs, y)\n            error.backward()\n            #net.parameters()  net.parameters() * learningrate\n            optimizer.step()\n\n        total_loss = 0\n        for X, y in valloader:\n            X = X.reshape(-1, 1).float()\n            y = y.view(-1, 1)\n            outputs = net(X)\n            error   = loss(outputs, y)\n            total_loss += error.data\n\n        print('Val loss for epoch', epoch, 'is', total_loss \/ len(valloader) )\n\n<\/code><\/pre>\n\n<p>it is called as:<\/p>\n\n<pre><code>net = Net()\nlosslist = train(net, trainloader, valloader, .0001, n_epochs = 4)\n<\/code><\/pre>\n\n<p>Where trainloader and valloader are the training and validation loaders. Can anyone help me see what's wrong with this? I know its not the learning rate since its the one I use in other frameworks, and I know its not the fact im using SGD or sigmoid activation functions, although I have a suspicion the error is in the activation functions somewhere.<\/p>\n\n<p>Does anyone know how to fix this? Thanks.<\/p>\n","answers":[{"AnswerId":"58045593","CreationDate":"2019-09-22T02:37:14.967","ParentId":null,"OwnerUserId":"4587151","Title":null,"Body":"<blockquote>\n  <p>In general, it always produces a straight line at y = (The mean of the function).<\/p>\n<\/blockquote>\n\n<p>Usually, this means that the NN has only successfully trained the final layer so far. You need to train it for longer or with better optimizations, as ViniciusArruda shows here.<\/p>\n\n<p>Edit: To explain further.. When only the final layer has been trained, the NN is effectively trying to guess the output y with no knowledge of the input X. In this case, the best guess it can make is the mean value. That way, it can minimize its MSE loss.<\/p>\n"},{"AnswerId":"58045560","CreationDate":"2019-09-22T02:29:57.663","ParentId":null,"OwnerUserId":"3058060","Title":null,"Body":"<p>After a while playing with some hyperparameters, modifying the net and changing the optimizer (following <a href=\"http:\/\/karpathy.github.io\/2019\/04\/25\/recipe\/\" rel=\"nofollow noreferrer\">this<\/a> excellent recipe) I ended up with changing the line <code>optimizer = torch.optim.SGD(net.parameters(), lr = learningrate)<\/code> to <code>optimizer = torch.optim.Adam(net.parameters())<\/code> (the <a href=\"https:\/\/pytorch.org\/docs\/stable\/optim.html#torch.optim.Adam\" rel=\"nofollow noreferrer\">default<\/a> optimizer parameters was used), running for 100 epochs and batch size equal to 1.<\/p>\n\n<p>The following code was used (tested on CPU only):<\/p>\n\n<pre><code>import torch\nimport torch.nn as nn\nfrom torch.utils import data\nimport numpy as np\nimport matplotlib.pyplot as plt\n\n# for reproducibility\ntorch.manual_seed(0)\nnp.random.seed(0)\n\nclass Dataset(data.Dataset):\n\n    def __init__(self, init, end, n):\n\n        self.n = n\n        self.x = np.random.rand(self.n, 1) * (end - init) + init\n        self.y = np.sin(self.x)\n\n    def __len__(self):\n\n        return self.n\n\n    def __getitem__(self, idx):\n\n        x = self.x[idx, np.newaxis]\n        y = self.y[idx, np.newaxis]\n\n        return torch.Tensor(x), torch.Tensor(y)\n\n\nclass Net(nn.Module):\n    def __init__(self):\n        super(Net, self).__init__()\n        self.model = nn.Sequential(\n        nn.Linear(1, 20),\n        nn.Sigmoid(),\n        nn.Linear(20, 50),\n        nn.Sigmoid(),\n        nn.Linear(50, 50),\n        nn.Sigmoid(),\n        nn.Linear(50, 1)\n        )\n\n    def forward(self, x):\n        x = self.model(x)\n        return x\n\ndef train(net, trainloader, valloader, n_epochs):\n\n    loss = nn.MSELoss()\n    # Switch the two following lines and run the code\n    # optimizer = torch.optim.SGD(net.parameters(), lr = 0.0001)\n    optimizer = torch.optim.Adam(net.parameters())\n\n    for epoch in range(n_epochs):\n\n        net.train()\n        for x, y in trainloader:\n            optimizer.zero_grad()\n            outputs = net(x).view(-1)\n            error   = loss(outputs, y)\n            error.backward()\n            optimizer.step()\n\n        net.eval()\n        total_loss = 0\n        for x, y in valloader:\n            outputs = net(x)\n            error   = loss(outputs, y)\n            total_loss += error.data\n\n        print('Val loss for epoch', epoch, 'is', total_loss \/ len(valloader) )    \n\n    net.eval()\n\n    f, (ax1, ax2) = plt.subplots(1, 2, sharey=True)\n\n    def plot_result(ax, dataloader):\n        out, xx, yy = [], [], []\n        for x, y in dataloader:\n            out.append(net(x))\n            xx.append(x)\n            yy.append(y)\n        out = torch.cat(out, dim=0).detach().numpy().reshape(-1)\n        xx = torch.cat(xx, dim=0).numpy().reshape(-1)\n        yy = torch.cat(yy, dim=0).numpy().reshape(-1)\n        ax.scatter(xx, yy, facecolor='green')\n        ax.scatter(xx, out, facecolor='red')\n        xx = np.linspace(0.0, 3.14159*2, 1000)\n        ax.plot(xx, np.sin(xx), color='green')\n\n    plot_result(ax1, trainloader)\n    plot_result(ax2, valloader)\n    plt.show()\n\n\ntrain_dataset = Dataset(0.0, 3.14159*2, 100)\nval_dataset = Dataset(0.0, 3.14159*2, 30)\n\nparams = {'batch_size': 1,\n          'shuffle': True,\n          'num_workers': 4}\n\ntrainloader = data.DataLoader(train_dataset, **params)\nvalloader = data.DataLoader(val_dataset, **params)\n\nnet = Net()\nlosslist = train(net, trainloader, valloader, n_epochs = 100)        \n<\/code><\/pre>\n\n<p>Result with Adam optimizer:\n<a href=\"https:\/\/i.stack.imgur.com\/5i3Qy.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5i3Qy.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Result with SGD optimizer:\n<a href=\"https:\/\/i.stack.imgur.com\/PfSLr.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/PfSLr.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":58044796,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2019-09-21T23:09:34.413","AcceptedAnswerId":null,"OwnerUserId":1150108.0,"Title":"How to shuffle the training data in Keras for an Autoencoder","Body":"<p>I'm using an autoencoder in Keras. I'm looking to shuffle the training data <code>x_train<\/code> so that the autoencoder will reconstruct the data to a different sample from the same class. Is this possible? <\/p>\n\n<pre><code>model_train = autoencoder.fit(x_train, x_train,\n          batch_size=32,\n          epochs=1000,\n          shuffle=True,\n          callbacks=[checkpoint, early_stopping],\n          validation_data=(x_test, x_test))\n<\/code><\/pre>\n\n<p>I assume that <code>shuffle=True<\/code> is shuffling <code>x_train<\/code> and calculating the loss based on the <em>same<\/em> pairs, and that's not what I want it to do.<\/p>\n","answers":[{"AnswerId":"58045173","CreationDate":"2019-09-22T00:39:53.943","ParentId":null,"OwnerUserId":"10807234","Title":null,"Body":"<p>It is possible, but Keras won't do it for you as it shuffles the data and labels together. Assuming you've got the labels, I found this function quite useful for your purposes:<\/p>\n\n<pre><code>import numpy as np\n\ndef create_pairs(data, labels):\n    # Exclude batch dimension\n    pairs = np.empty(0, 2, *data.shape[1:])\n\n    for label in np.unique(labels):\n        idxs = np.where(labels == label)[0]\n        # Indexes must be even in order to create pairs\n        idxs = idxs if len(idxs) % 2 == 0 else idxs[:-1]\n        np.random.shuffle(idxs)\n\n        samples = data[idxs].reshape((-1, 2, *data.shape[1:]))\n        pairs = np.vstack((pairs, samples))\n    return pairs[:, 0], pairs[:, 1]\n<\/code><\/pre>\n\n<p>Now that the data is shuffled and divided into pairs you can train your model:<\/p>\n\n<pre><code>x_train, y_train = create_pairs(data, labels)\nhistory = model.fit(\n    x_train, y_train,\n    batch_size=32,\n    epochs=1000,\n    shuffle=True,\n    callbacks=[checkpoint, early_stopping],\n    validation_split=0.2)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58045085,"AnswerCount":1,"Tags":"<machine-learning><keras><neural-network>","CreationDate":"2019-09-22T00:14:04.440","AcceptedAnswerId":58045175.0,"OwnerUserId":3444294.0,"Title":"Magnitude of Sample Weights in Keras","Body":"<p>Keras model.fit supports per-sample weights. What is the range of acceptable values for these weights? Must they sum to 1 across all training samples? Or does keras accept any weight values and then perform some sort of normalization? The keras source includes, e.g. <code>training_utils.standardize_weights<\/code> but that does not appear to be doing statistical standardization.<\/p>\n","answers":[{"AnswerId":"58045175","CreationDate":"2019-09-22T00:40:31.417","ParentId":null,"OwnerUserId":"6697120","Title":null,"Body":"<p>After looking at the source <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/engine\/training_utils.py\" rel=\"nofollow noreferrer\">here<\/a>, I've found that you should be able to pass any acceptable numerical values (within overflow bounds) for both sample weights and class weights.  They do not need to sum to 1 across all training samples and each weight may be greater than one.  The only sort of normalization that appears to be happening is taking the max of 2D class weight inputs.<\/p>\n\n<p>If both class weights and samples weights are provided, it provides the product of the two.<\/p>\n\n<p>I think the unspoken component here is that the activation function should be dealing with normalization.<\/p>\n"}]}
+{"QuestionId":58045108,"AnswerCount":0,"Tags":"<python><keras><neural-network><conv-neural-network>","CreationDate":"2019-09-22T00:20:28.847","AcceptedAnswerId":null,"OwnerUserId":6417245.0,"Title":"Using flow_from_directory with double input model","Body":"<p><a href=\"https:\/\/i.stack.imgur.com\/0Tw7U.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0Tw7U.png\" alt=\"General network architecture\"><\/a><\/p>\n\n<p>So I have some images that I need to classify. I also have extracted some metrics from these images, which I plan on feeding to a multilayer perceptron and then merging with the CNN's output. I am using keras' flow_from_directory() to feed the images to the network, but now that I am trying to make the double input network shown above, I am unsure as to how to feed the data. <\/p>\n\n<p>Is there any way to train a model like this using flow_from_directory() when one of the inputs takes images and the other an array of floats? Would I need to create a separate generator for the metrics and have the data in a different directory, or should I dump the images (paths) and metrics into a dataframe and use flow_from_dataframe() instead?<\/p>\n","answers":[]}
+{"QuestionId":58045596,"AnswerCount":1,"Tags":"<python><keras><conv-neural-network><max-pooling>","CreationDate":"2019-09-22T02:38:14.060","AcceptedAnswerId":58045694.0,"OwnerUserId":9644683.0,"Title":"AttributeError: module 'keras.backend' has no attribute 'set_image_dim_ordering'","Body":"<pre><code>model = Sequential()\nK.set_image_dim_ordering('th')\nmodel.add(Convolution2D(30, 5, 5, border_mode= 'valid' , input_shape=(1, 10, 10),activation= 'relu' ))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Convolution2D(15, 3, 3, activation= 'relu' ))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.2))\nmodel.add(Flatten())\nmodel.add(Dense(128, activation= 'relu' ))\nmodel.add(Dense(50, activation= 'relu' ))\nmodel.add(Dense(10, activation= 'softmax' ))\n# Compile model\nmodel.compile(loss= 'categorical_crossentropy' , optimizer= 'adam' , metrics=[ 'accuracy' ])\n<\/code><\/pre>\n\n<p>I get an error when I use <code>set_image_dim_ordering()<\/code> from <code>keras.backend<\/code><\/p>\n\n<p>This the error report : <code>AttributeError: module 'keras.backend' has no attribute 'set_image_dim_ordering'<\/code><\/p>\n\n<p>My Import statement <\/p>\n\n<pre><code>import numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I\/O (e.g. pd.read_csv)\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.layers import Dropout\nfrom keras.layers import Flatten\nfrom keras.layers.convolutional import Convolution2D\nfrom keras.layers.convolutional import MaxPooling2D\nfrom sklearn.preprocessing import LabelEncoder,OneHotEncoder\nfrom keras import backend as K\nfrom subprocess import check_output\n<\/code><\/pre>\n","answers":[{"AnswerId":"58045694","CreationDate":"2019-09-22T02:59:55.800","ParentId":null,"OwnerUserId":"10366273","Title":null,"Body":"<p>A couple things to look at. This is a <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/12649\" rel=\"nofollow noreferrer\">known issue<\/a> and reported for version 2.2.5 but should work with 2.2.4 and before.<\/p>\n\n<p>However, you should stop using this method anyway because it is legacy now and has been replaced by <a href=\"https:\/\/keras.io\/backend\/#image_data_format\" rel=\"nofollow noreferrer\">image_data_format<\/a>: [<a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/1cf5218edb23e575a827ca4d849f1d52d21b4bb0\/keras\/backend\/common.py#L194\" rel=\"nofollow noreferrer\">code<\/a>]<\/p>\n\n<pre><code>keras.backend.image_data_format()\nkeras.backend.set_image_data_format(data_format)\n<\/code><\/pre>\n\n<p>It should continue to work but there is a bug right now, at least.<\/p>\n\n<p><strong>EDIT:<\/strong> <em>Poster reported below does not work or returns an error with their code, though the method does seem to exist.<\/em><\/p>\n\n<blockquote>\n  <p>Some have reported that you may be able to access the method using <code>K.common<\/code> but I haven't tested:<\/p>\n\n<pre><code>K.common.image_dim_ordering()\nK.common.set_image_dim_ordering(dim_ordering)\n<\/code><\/pre>\n<\/blockquote>\n"}]}
+{"QuestionId":58045715,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-09-22T03:05:30.650","AcceptedAnswerId":58045827.0,"OwnerUserId":4910495.0,"Title":"Why does my keras model require a 3D list to predict?","Body":"<p>I've read a few answers on here saying that models require a 2D array to predict, even though we may just want one prediction. However, for as long as I can remember my models require 3D arrays to predict on. My input shape is 2D (shape 47145, 120), though when predicting I need to place my array inside two other lists in order for it to work correctly. Output shape is a 1D array.<\/p>\n\n<pre><code>model.predict([[1, 2, 3...]])[0]\n<\/code><\/pre>\n\n<p>doesn't work, but<\/p>\n\n<pre><code>model.predict([[[1, 2, 3...]]])[0][0]\n<\/code><\/pre>\n\n<p>does. I don't mind it, and it predicts perfectly fine for me, but having to add an extra <code>[0][0]<\/code> after each line of prediction has led me to question if this is the expected behavior.<\/p>\n\n<p>Here's the error message I get when I try to predict on a 2D array:<\/p>\n\n<pre><code>In [154] a = [0] * 120\n\nIn [155]: model.predict([a])\nValueError: Error when checking input: expected dense_1_input to have shape (120,) but got array with shape (1,) \n<\/code><\/pre>\n\n<p>Here is the code I use to create my model:<\/p>\n\n<pre><code>opt = keras.optimizers.Adam(lr=0.00055, decay=1e-6)\n\nlayer_num = 5\nnodes = 80\n\nmodel = Sequential()\nmodel.add(Dense(x_train.shape[1], activation=\"relu\", input_shape=(x_train.shape[1:])))\nfor i in range(layer_num):\n    model.add(Dense(nodes, activation=\"relu\"))\nmodel.add(Dense(1, activation='linear'))\n\ntensorboard = TensorBoard(log_dir=\"logs\\{}-layers-{}-nodes-{}-opt-adam-decay2\".format(\n                          layer_num, nodes, 'relu'), histogram_freq=0, write_graph=True)\ncallbacks = [tensorboard]\n\nmodel.compile(loss='mean_squared_error', optimizer=opt, metrics=['mean_absolute_error'])\nmodel.fit(x_train, y_train, shuffle=True, batch_size=64, validation_data=(x_test, y_test),\n                                                         callbacks=callbacks, epochs=80)\n<\/code><\/pre>\n\n<p>My shapes:<\/p>\n\n<pre><code>In [146]: x_train.shape\nOut[146]: (47145, 120)\n\nIn [147]: y_train.shape\nOut[147]: (47145,)\n\nIn [148]: x_train[0]\nOut[148]: \narray([0.51050081, 0.48233333, 0.12769333, 0.12920803, 0.13162067,\n       0.12359916, 0.51050081, 0.488     , 0.12325733, 0.14299455,\n       0.13225505, 0.12455976, 0.51050081, 0.49366667, 0.12325733,\n       0.14299455, 0.13225505, 0.12455976, 0.51050081, 0.499     ,\n       0.13022811, 0.13016997, 0.13352359, 0.12584059, 0.51050081,\n       0.49566667, 0.12896078, 0.11991025, 0.13225505, 0.13096383,\n       0.51050081, 0.49233333, 0.12769333, 0.11638345, 0.12940056,\n       0.13032347, 0.51050081, 0.48933333, 0.12325733, 0.12407822,\n       0.12718045, 0.12808203, 0.51050081, 0.49266667, 0.12547533,\n       0.12696372, 0.12718045, 0.12231832, 0.51050081, 0.48933333,\n       0.12547533, 0.12696372, 0.12718045, 0.12231832, 0.51050081,\n       0.486     , 0.12737644, 0.12568135, 0.12654618, 0.12231832,\n       0.51050081, 0.48266667, 0.12769333, 0.12119272, 0.12496034,\n       0.12584059, 0.51050081, 0.47933333, 0.13181244, 0.12151326,\n       0.12718045, 0.12936276, 0.51050081, 0.476     , 0.12896078,\n       0.12696372, 0.1284491 , 0.12391928, 0.51050081, 0.47266667,\n       0.12896078, 0.12696372, 0.1284491 , 0.12391928, 0.51050081,\n       0.46933333, 0.126109  , 0.12728437, 0.12654618, 0.11783545,\n       0.51050081, 0.466     , 0.12547533, 0.12343704, 0.12305742,\n       0.11143137, 0.51050081, 0.46266667, 0.13022811, 0.11926895,\n       0.11988585, 0.11879605, 0.51050081, 0.45933333, 0.12737644,\n       0.12407822, 0.12147158, 0.12968299, 0.51050081, 0.456     ,\n       0.126109  , 0.12568135, 0.12718045, 0.14761449, 0.51050081,\n       0.45266667, 0.126109  , 0.12568135, 0.12718045, 0.14761449])\n<\/code><\/pre>\n","answers":[{"AnswerId":"58045827","CreationDate":"2019-09-22T03:38:02.613","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>Not expected behavior; the problem arises from passing in a <em>list<\/em>, not an array - where behind the scenes, the former's handled as an iterable (<code>len([[[1,2,3]]][0].shape)==2)<\/code>), but a Numpy array is processed directly. Try <code>model.predict(np.array([[1,2,3,...]]))<\/code><\/p>\n\n<p>To your edit, <code>a = [0] * 120<\/code> is <em>not<\/em> a 2D array, but still a list - even as <code>[a]<\/code>.<\/p>\n"}]}
+{"QuestionId":58045718,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><tensorboard>","CreationDate":"2019-09-22T03:06:05.457","AcceptedAnswerId":58049841.0,"OwnerUserId":8238158.0,"Title":"See all correctly and incorrectly identified images when training on the mnist dataset","Body":"<p>I'm trying to find a way to visualize which numbers in the mnist dataset a model was able to correctly identify and which ones it wasn't. \nWhat I can't seem to find is if such a visualization is possible in tensorboard or if I would need to use\/create something else to achieve it.<\/p>\n\n<p>I'm currently working from the <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/quickstart\/beginner\" rel=\"nofollow noreferrer\">basic tutorial<\/a> provided for tensorflow 2.0 with tensorboard added.<\/p>\n\n<pre><code>import datetime\nimport tensorflow as tf\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(input_shape=(28, 28)),\n  tf.keras.layers.Dense(128, activation='relu'),\n  tf.keras.layers.Dropout(0.2),\n  tf.keras.layers.Dense(10, activation='softmax')\n])\n\nmodel.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\n\nlog_dir=\"logs\/fit\/\" + datetime.datetime.now().strftime(\"%Y%m%d-%H%M%S\")\ntensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=log_dir, histogram_freq=1)\n\nmodel.fit(x_train,\n          y_train,\n          epochs=5,\n          validation_data=(x_test, y_test),\n          callbacks=[tensorboard_callback])\n\nmodel.evaluate(x_test, y_test)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58049841","CreationDate":"2019-09-22T13:50:55.100","ParentId":null,"OwnerUserId":"8238158","Title":null,"Body":"<p>It appears the what-if tool is what I was looking for, it allows you to visually sort testing data depending on whether it was <a href=\"https:\/\/i.stack.imgur.com\/xrxiW.jpg\" rel=\"nofollow noreferrer\">correctly or incorrectly identified by the model.<\/a><\/p>\n\n<p>If you want to test it out <a href=\"https:\/\/pair-code.github.io\/what-if-tool\/age.html\" rel=\"nofollow noreferrer\">here<\/a> is their demo that I used to get the above image and they have multiple other demos on the tools <a href=\"https:\/\/pair-code.github.io\/what-if-tool\/index.html#demos\" rel=\"nofollow noreferrer\">site<\/a>.<\/p>\n"}]}
+{"QuestionId":58045909,"AnswerCount":1,"Tags":"<matrix><pytorch><matrix-multiplication><tensor>","CreationDate":"2019-09-22T03:57:53.580","AcceptedAnswerId":null,"OwnerUserId":12001135.0,"Title":"How do I have to multiply the following 4darrays to obtain a 2d array in Pytorch?","Body":"<p>I am using Pytorch. I have a 4d array with shape (C, H, W, K) and I want to multiply it to another 4d-array with shape (N, C, H, W).\nI want a final 2-d matrix (N, K).\nHow should I implement this in Pytorch?\nAnd, more generally, how does higher order matrix multiplication works? Are there any rules to keep in mind in general?<\/p>\n","answers":[{"AnswerId":"58046202","CreationDate":"2019-09-22T05:05:33.953","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can use <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.Tensor.view\" rel=\"nofollow noreferrer\"><code>view<\/code><\/a> to reshape the 4D tensors into simple 2D matrices and then multiply them<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x1 = x1.view(-1, x1.shape[-1])  # from C-H-W-k -&gt; CHW-K\nx2 = x2.view(x2.shape[0], -1)   # from N-C-H-W -&gt; N-CHW\nout = x2 @ x1  # matrix multiplication\n<\/code><\/pre>\n\n<p>Note that the complexity of <code>view<\/code>, in generalm is minimal as it usually does not incur copying and moving of tensor values in memory - only changes to the header of the tensor representation.  <\/p>\n\n<p>For more advanced and more flexible tensor multiplication, you might want to look at <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.einsum\" rel=\"nofollow noreferrer\"><code>einsum<\/code><\/a>.<\/p>\n"}]}
+{"QuestionId":58046140,"AnswerCount":0,"Tags":"<python-3.x><theano>","CreationDate":"2019-09-22T04:52:46.040","AcceptedAnswerId":null,"OwnerUserId":10583531.0,"Title":"ValueError: invalid literal for int() with base 10: 'Active code page: 65001\\r\\n4'","Body":"<p>I met this error just as the title when I tried to run some <code>Theano<\/code> test code in Pycharm. What made me really confused is that if I run the code in CLI(I mean, using <code>python XXX.py<\/code> directly), it should work well!<\/p>\n\n<p>Here is the error statement:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"D:\/test_gpu.py\", line 1, in &lt;module&gt;\n    from theano import function, config, shared, tensor\n  File \"D:\\Anaconda3\\envs\\theano\\lib\\site-packages\\theano\\__init__.py\", line 156, in &lt;module&gt;\n    import theano.gpuarray\n  File \"D:\\Anaconda3\\envs\\theano\\lib\\site-packages\\theano\\gpuarray\\__init__.py\", line 33, in &lt;module&gt;\n    from . import fft, dnn, opt, extra_ops, multinomial, reduction, sort, rng_mrg, ctc\n  File \"D:\\Anaconda3\\envs\\theano\\lib\\site-packages\\theano\\gpuarray\\fft.py\", line 14, in &lt;module&gt;\n    from .opt import register_opt, op_lifter, register_opt2\n  File \"D:\\Anaconda3\\envs\\theano\\lib\\site-packages\\theano\\gpuarray\\opt.py\", line 2801, in &lt;module&gt;\n    from .dnn import (local_abstractconv_cudnn,\n  File \"D:\\Anaconda3\\envs\\theano\\lib\\site-packages\\theano\\gpuarray\\dnn.py\", line 339, in &lt;module&gt;\n    handle_type = CUDNNDataType('cudnnHandle_t', 'cudnnDestroy')\n  File \"D:\\Anaconda3\\envs\\theano\\lib\\site-packages\\theano\\gpuarray\\dnn.py\", line 259, in CUDNNDataType\n    version=version(raises=False))\n  File \"D:\\Anaconda3\\envs\\theano\\lib\\site-packages\\theano\\gpuarray\\dnn.py\", line 319, in version\n    if not dnn_present():\n  File \"D:\\Anaconda3\\envs\\theano\\lib\\site-packages\\theano\\gpuarray\\dnn.py\", line 207, in dnn_present\n    dnn_present.avail, dnn_present.msg = _dnn_check_compile()\n  File \"D:\\Anaconda3\\envs\\theano\\lib\\site-packages\\theano\\gpuarray\\dnn.py\", line 169, in _dnn_check_compile\n    try_run=False, output=True)\n  File \"D:\\Anaconda3\\envs\\theano\\lib\\site-packages\\theano\\gof\\cmodule.py\", line 2230, in try_flags\n    theano.config.cxx, comp_args)\n  File \"D:\\Anaconda3\\envs\\theano\\lib\\site-packages\\theano\\gof\\cmodule.py\", line 1879, in _try_flags\n    comp_args=comp_args)\n  File \"D:\\Anaconda3\\envs\\theano\\lib\\site-packages\\theano\\gof\\cmodule.py\", line 1796, in _try_compile_tmp\n    args = cls.compile_args()\n  File \"D:\\Anaconda3\\envs\\theano\\lib\\site-packages\\theano\\gof\\cmodule.py\", line 2109, in compile_args\n    mj, mn, patch = [int(vp) for vp in version]\n  File \"D:\\Anaconda3\\envs\\theano\\lib\\site-packages\\theano\\gof\\cmodule.py\", line 2109, in &lt;listcomp&gt;\n    mj, mn, patch = [int(vp) for vp in version]\nValueError: invalid literal for int() with base 10: 'Active code page: 65001\\r\\n4'\n<\/code><\/pre>\n\n<p>My platform is Windows 10 and my python version is 3.6.5, and my Theano version is 1.0.2. I think there is no problem with my installation. Please notice the <code>Active code page: 65001\\r\\n4<\/code> in the statement. I know it is the CLI window I just opened, but I have no idea what it means.<\/p>\n\n<p>I tried to google some solutions about the <code>invalid literal for int() with base 10<\/code>, which is about transformation error between and str and int. But it may not be in my case.<\/p>\n\n<p>Here is my code, copied from the documentation:<\/p>\n\n<pre><code>from theano import function, config, shared, tensor\nimport numpy\nimport time\n\nvlen = 10 * 30 * 768  # 10 x #cores x # threads per core\niters = 1000\n\nrng = numpy.random.RandomState(22)\nx = shared(numpy.asarray(rng.rand(vlen), config.floatX))\nf = function([], tensor.exp(x))\nprint(f.maker.fgraph.toposort())\nt0 = time.time()\nfor i in range(iters):\n    r = f()\nt1 = time.time()\nprint(\"Looping %d times took %f seconds\" % (iters, t1 - t0))\nprint(\"Result is %s\" % (r,))\nif numpy.any([isinstance(x.op, tensor.Elemwise) and\n              ('Gpu' not in type(x.op).__name__)\n              for x in f.maker.fgraph.toposort()]):\n    print('Used the cpu')\nelse:\n    print('Used the gpu')\n\n<\/code><\/pre>\n\n<p>And my running results in CIL is as follows:<\/p>\n\n<pre><code>D:\\Anaconda3\\envs\\theano\\lib\\site-packages\\theano\\gpuarray\\dnn.py:184: UserWarning: Your cuDNN version is more recent than Theano. If you encounter problems, try updating Theano or downgrading cuDNN to a version &gt;= v5 and &lt;= v7.\n  warnings.warn(\"Your cuDNN version is more recent than \"\nUsing cuDNN version 7600 on context None\nMapped name None to device cuda: GeForce GTX 965M (0000:01:00.0)\n[GpuElemwise{exp,no_inplace}(&lt;GpuArrayType&lt;None&gt;(float32, vector)&gt;), HostFromGpu(gpuarray)(GpuElemwise{exp,no_inplace}.0)]\nLooping 1000 times took 0.627322 seconds\nResult is [1.2317803 1.6187935 1.5227807 ... 2.2077181 2.2996776 1.623233 ]\nUsed the gpu\n<\/code><\/pre>\n\n<p>Does anyone meet a similar error? What can I do to fix this problem?<\/p>\n\n<p>Thx!<\/p>\n","answers":[]}
+{"QuestionId":58046178,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-09-22T05:00:39.703","AcceptedAnswerId":58046263.0,"OwnerUserId":3983454.0,"Title":"How to have incrementing batch size in pytorch","Body":"<p>In pytorch, DataLoader will split a dataset into batches of set size with additional options of shuffling etc, which one can then loop over.<\/p>\n\n<p>But if I need the batch size to increment, such as first 10 batch of size 50, next 5 batch of size 100 and so on, what's the best way of doing so?<\/p>\n\n<p>I tried splitting the tensor then concat them:<\/p>\n\n<pre><code>#10x50 + 5*100\noriginalTensor = torch.randn(1000, 80)\nsplit1=torch.split(originalTensor, 500, dim=0)\nsplit2=torch.split(list(split1)[0], 100, dim=0)\n<\/code><\/pre>\n\n<p>Thereafter is there a way to pass the concatenated tensor into dataLoader or any other way to directly turn the concat tensor into a generator (which might lose shuffling and other functionalities)?<\/p>\n","answers":[{"AnswerId":"58046263","CreationDate":"2019-09-22T05:18:44.093","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>I think you can do that by simply providing a non-default <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/data.html#torch.utils.data.BatchSampler\" rel=\"nofollow noreferrer\"><code>batch_sampler<\/code><\/a> to your <a href=\"https:\/\/pytorch.org\/docs\/1.1.0\/data.html#torch.utils.data.DataLoader\" rel=\"nofollow noreferrer\"><code>DataLoader<\/code><\/a>.<br>\nFor instance:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class VaryingSizeBatchSampler(Sampler):\n    r\"\"\"Wraps another sampler to yield a varying-size mini-batch of indices.\n\n    Args:\n        sampler (Sampler): Base sampler.\n        batch_size_fn (function): Size of current mini-batch.\n        drop_last (bool): If ``True``, the sampler will drop the last batch if\n            its size would be less than ``batch_size``\n    \"\"\"\n\n    def __init__(self, sampler, batch_size_fn, drop_last):\n        if not isinstance(sampler, Sampler):\n            raise ValueError(\"sampler should be an instance of \"\n                             \"torch.utils.data.Sampler, but got sampler={}\"\n                             .format(sampler))\n        self.sampler = sampler\n        self.batch_size_fn = batch_size_fn\n        self.drop_last = drop_last\n        self.batch_counter = 0\n\n    def __iter__(self):\n        batch = []\n        cur_batch_size = self.batch_size_fn(self.batch_counter)  # get current batch size\n        for idx in self.sampler:\n            batch.append(idx)\n            if len(batch) == cur_batch_size:\n                yield batch\n                self.batch_counter += 1\n                cur_batch_size = self.batch_size_fn(self.batch_counter)  # get current batch size                \n                batch = []\n        if len(batch) &gt; 0 and not self.drop_last:\n            yield batch            \n\n    def __len__(self):\n        raise NotImplementedError('You need to implement it yourself!')\n<\/code><\/pre>\n"}]}
+{"QuestionId":58046402,"AnswerCount":1,"Tags":"<android><tensorflow><speech-recognition>","CreationDate":"2019-09-22T05:49:26.687","AcceptedAnswerId":null,"OwnerUserId":2088427.0,"Title":"What's the most effective way to implement an offline continuous speech recognizer on Android using Tensorflow?","Body":"<p>I'm looking for a way to implement an offline continuous speech recognizer running on Android and preferably using the Tensorflow engine.<\/p>\n\n<p>The app needs to be able to recognize any word out of a predefined set of words in a continuous stream of speech and the assumption is that the Android device may be in an area that does not support an internet connection, so Google services and such are out of the question.<\/p>\n\n<p>Are there perhaps any RNN or CNN models that can achieve this?<\/p>\n","answers":[{"AnswerId":"58047357","CreationDate":"2019-09-22T08:21:30.893","ParentId":null,"OwnerUserId":"10878733","Title":null,"Body":"<p>You can use the <a href=\"https:\/\/www.tensorflow.org\/lite\/examples\" rel=\"nofollow noreferrer\">Speech Recognition<\/a> ( scroll down and search for \"<em>Speech Recognition<\/em>\" ) model available on <a href=\"https:\/\/www.tensorflow.org\/lite\" rel=\"nofollow noreferrer\">TensorFlow Lite<\/a>. It can generate probability scores for words spoken in an audio stream. The best thing is that they have provided an Android sample <a href=\"https:\/\/github.com\/tensorflow\/examples\/tree\/master\/lite\/examples\/speech_commands\/android\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>Using a TensorFlow Lite model, you will not require an active internet connection to make predictions. This can compromise the app's size ( the <code>.tflite<\/code> model file could have a large size though ).<\/p>\n\n<p>You can build a simple method which extracts necessary words from the model's predictions.<\/p>\n"}]}
+{"QuestionId":58046516,"AnswerCount":1,"Tags":"<python><tensorflow><keras><training-data>","CreationDate":"2019-09-22T06:10:35.540","AcceptedAnswerId":58046619.0,"OwnerUserId":9644683.0,"Title":"ValueError: Error when checking input: expected conv2d_36_input to have shape (3, 32, 32) but got array with shape (1, 10, 10)","Body":"<p>When I try to run my training model the error comes what should I do to pass the error report?<\/p>\n\n<p>Is there any problem with my model or my reshape part?<\/p>\n\n<p>This is my reshape part<\/p>\n\n<pre><code># Reshape and normalize training data\ntrainX = train.reshape(train.shape[0], 1, 10, 10).astype( 'float32' )\nx_train = trainX \/ 255.0\ny_train = train[:,99]\n# print(y_train)\n# # # Reshape and normalize test data\ntestX = test.reshape(test.shape[0], 1, 10, 10).astype(     'float32' )\nx_test = testX \/ 255.0\ny_test = test[:,99]\n# print(y_test)\n<\/code><\/pre>\n\n<p>This is my model :<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Conv2D(32, (3, 3), input_shape=(3, 32, 32), \nactivation='relu', padding='same'))\nmodel.add(Dropout(0.2))\nmodel.add(BatchNormalization())\nmodel.add(Conv2D(64, (3, 3), padding='same'))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.2))\nmodel.add(BatchNormalization())\n<\/code><\/pre>\n\n<p>Then If I reshape my data like into 3*32*32 Then the value error report appears  : <\/p>\n\n<pre><code>---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-31-36bf3e556ae8&gt; in &lt;module&gt;()\n----&gt; 1 trainX = train.reshape(train.shape[0], 3, 32, 32).astype( 'float32' )\n  2 x_train = trainX \/ 255.0\n  3 y_train = train[:,10]\n  4 # print(y_train)\n  5 # # # Reshape and normalize test data\n<\/code><\/pre>\n","answers":[{"AnswerId":"58046619","CreationDate":"2019-09-22T06:27:19.600","ParentId":null,"OwnerUserId":"12100866","Title":null,"Body":"<p>Input shape of your Train and Test dataset differ from your model. When using <strong>input_shape()<\/strong> you need to keep in mind what is the size of the desired input. in you case your shape looks as (1,10,10) which is a 10x10 image with one level channel\/depth (i.e. black and white image). However, the model you are using requires a (3,32,32) shape which is translated into a color images (i.e. 3 channel represents RGB colors) with 32x32 dimensions.<\/p>\n\n<p>Thus if you change the model as followed code, it might work, but there other parameters you need to optimize as well, such as Conv2D desired feature map:<\/p>\n\n<pre><code>model.add(Conv2D(32, (3, 3), **input_shape=(1, 10, 10)**,activation='relu', padding='same'))\n<\/code><\/pre>\n"}]}
+{"QuestionId":58046994,"AnswerCount":1,"Tags":"<python><tensorflow><keras><import>","CreationDate":"2019-09-22T07:24:45.420","AcceptedAnswerId":58047024.0,"OwnerUserId":735008.0,"Title":"How to use \"keras\" directly rather than to use \"tensorflow.keras\" in importing lines","Body":"<p>I am trying to use a third party library that is using <code>tensorflow<\/code> and <code>keras<\/code> internally.\nI encountered one issue in using this library.<\/p>\n\n<p>In that library, it is assumed that <code>keras<\/code> is available as a top-level module name.<\/p>\n\n<p>For example, the following lines are frequently used inside this library.<\/p>\n\n<pre><code>from keras.layers import Dense\n<\/code><\/pre>\n\n<p>In my configuration, it is somewhat different, so I need to import lines as follows:<\/p>\n\n<pre><code>from tensorflow.keras.layers import Dense\n<\/code><\/pre>\n\n<p>Since that library is not that small, it is not efficient for me to manually edit all such lines. What would be the best way to resolve this problem?<\/p>\n","answers":[{"AnswerId":"58047024","CreationDate":"2019-09-22T07:30:38.127","ParentId":null,"OwnerUserId":"12100853","Title":null,"Body":"<p>You will need to install keras:<\/p>\n\n<ul>\n<li><p>If you are using Windows or virtual environment: \n<code>pip install keras<\/code><\/p><\/li>\n<li><p>For Linux or Mac environment:\n<code>sudo pip install keras<\/code><\/p><\/li>\n<\/ul>\n\n<p>Then, afterwards you can use <code>from keras.layers import Dense<\/code> instead.<\/p>\n"}]}
+{"QuestionId":58047284,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2019-09-22T08:08:44.463","AcceptedAnswerId":58048052.0,"OwnerUserId":8684109.0,"Title":"What's the effect of monitor in keras?","Body":"<p>In Keras, we use ModelCheckpoint to save our trained models. In a document of Keras, it explains that \"monitor: quantity to monitor.\", but I still can't understand it. What's the effect of monitor in our machine learning process?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>keras.callbacks.callbacks.ModelCheckpoint(filepath, monitor='val_loss', verbose=0, save_best_only=False, save_weights_only=False, mode='auto', period=1)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58048052","CreationDate":"2019-09-22T09:54:57.857","ParentId":null,"OwnerUserId":"4622046","Title":null,"Body":"<p><a href=\"https:\/\/keras.io\/callbacks\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/callbacks\/<\/a><\/p>\n\n<p>From the keras documentation, I'm explaining the parameters of the ModelCheckpoint. It's used to save your best model while training, the reason is maybe after training for few epochs the model may start to diverge or show poor performance\/ may get overfitted. Many epochs do not always mean best performance, so it's better to keep saving the weights while training.<\/p>\n\n<p><strong>save_best_only: if save_best_only=True, the latest best model according to the quantity monitored will not be overwritten.<\/strong> Here, it clearly says, the model will be saved based on the value of the quantity to be monitored. <\/p>\n\n<p><strong>mode: one of {auto, min, max}. If save_best_only=True, the decision to overwrite the current save file is made based on either the maximization or the minimization of the monitored quantity. For val_acc, this should be max, for val_loss this should be min, etc. In auto mode, the direction is automatically inferred from the name of the monitored quantity.<\/strong> \nThe mode will be determined based on your monitoring metric, if it's a loss then the mode must be min, if it's something like accuracy, f1 score etc. then the mode must be max. (You want to save the weights which shows least loss, and best accuracy so far)<\/p>\n\n<p><strong>verbose: verbosity mode, 0 or 1.<\/strong> verbose determines how much information you want to get printed about your metrics (0 means nothing will be printed, 1 means some information will be printed)<\/p>\n\n<p>Other parameters should be very easy to understand. <\/p>\n"}]}
+{"QuestionId":58047314,"AnswerCount":1,"Tags":"<tensorflow><ubuntu><build><bazel>","CreationDate":"2019-09-22T08:14:14.237","AcceptedAnswerId":null,"OwnerUserId":6509922.0,"Title":"Building Tensorflow 2.0 from source","Body":"<p>I've built Tensorflow 2.0 from source using the below commands:<\/p>\n\n<pre><code>-march=native -Wno-sign-compare -mssse3 -mcx16 -msse4.1 -msse4.2 -mpopcnt -mno-avx\n\nbazel build --config=opt  \/\/tensorflow\/tools\/pip_package:build_pip_package --local_ram_resources=1024 --jobs=1 --define=tf_api_version=2\n<\/code><\/pre>\n\n<p>The build apparently completed successfully, that is without any errors, but when I go to install, using the below command:<\/p>\n\n<pre><code>python \/tensorflow\/tools\/pip_package\/setup.py install\n<\/code><\/pre>\n\n<p>It installs the egg in site-packages and checks for dependencies, but does nothing else. Nor does it return any error. The end of the output is below:<\/p>\n\n<pre><code>Searching for Markdown==3.1.1\nBest match: Markdown 3.1.1\nAdding Markdown 3.1.1 to easy-install.pth file\nInstalling markdown_py script to \/var\/...\n\nUsing \/var\/...\/python3.6\/site-packages\nSearching for h5py==2.9.0\nBest match: h5py 2.9.0\nAdding h5py 2.9.0 to easy-install.pth file\n\nUsing \/var\/...\/python3.6\/site-packages\nFinished processing dependencies for tensorflow==2.0.0rc0\n<\/code><\/pre>\n\n<p>and of course when I run python:<\/p>\n\n<pre><code>Python 3.6.7 (default, Oct 22 2018, 11:32:17)\n[GCC 8.2.0] on linux\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\n&gt;&gt;&gt; import tensorflow as tf\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\nModuleNotFoundError: No module named 'tensorflow'\n<\/code><\/pre>\n\n<p>What am I missing here?<\/p>\n","answers":[{"AnswerId":"58195366","CreationDate":"2019-10-02T04:49:44.670","ParentId":null,"OwnerUserId":"11414441","Title":null,"Body":"<p>I could not comment at the moment. Can you verify the installation is done successfully ? That is after running: <\/p>\n\n<pre><code>python \/tensorflow\/tools\/pip_package\/setup.py install\n<\/code><\/pre>\n\n<p>you can check the folder tensorflow package in Python lib folders.<\/p>\n"}]}
+{"QuestionId":58047324,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2019-09-22T08:16:05.970","AcceptedAnswerId":58838359.0,"OwnerUserId":11423459.0,"Title":"How do I generate data with Tensorflow, if I have a bunch of example?","Body":"<p>I Am an absolute beginner with Tensorflow. I have searched, but did not found how to do this:<\/p>\n\n<p>If I have a list of strings like this:<\/p>\n\n<pre><code>[\"sentace1\", \"...\", \"sentance5000\"]\n<\/code><\/pre>\n\n<blockquote>\n  <p>How do I train a neural network to create similar sentences? What is the logic of generating data, text, images?  Can someone explain to me using code, through this relatively basic example?<\/p>\n<\/blockquote>\n\n<p>Also, If I'd add more layers and different types of data, could it create for example pictures or music?<\/p>\n\n<p><strong>A thousand thanks :)<\/strong><\/p>\n","answers":[{"AnswerId":"58047415","CreationDate":"2019-09-22T08:29:10.970","ParentId":null,"OwnerUserId":"8516049","Title":null,"Body":"<p>Popular model for generating data in natural language processing is word2vec. It encodes words to vector space, then similar words to given can be generated. Basicly you can vectorise almost anything so variation you are looking for is sentence2vec which works similar but you give sentence as input and it encodes sentences to vectors.<\/p>\n\n<p>Here is tutorial for tensorflow: tensorflow.org\/tutorials\/representation\/word2vec<\/p>\n\n<p>You can also try gensim implementation of w2v:\n<a href=\"https:\/\/radimrehurek.com\/gensim\/models\/word2vec.html\" rel=\"nofollow noreferrer\">https:\/\/radimrehurek.com\/gensim\/models\/word2vec.html<\/a><\/p>\n\n<p>About:<\/p>\n\n<blockquote>\n  <p>Also, If I'd add more layers and different types of data, could it\n  create for example pictures or music?<\/p>\n<\/blockquote>\n\n<p>For generating images you should read about Auto Encoders (eg. VAE) and Generative Adversarial Networks (GANs). These architectures works different from NLP architectures.<\/p>\n"},{"AnswerId":"58838359","CreationDate":"2019-11-13T13:35:35.827","ParentId":null,"OwnerUserId":"7372766","Title":null,"Body":"<p>Generation of music and text differs form generation of images. Text and music generation can be done with sequence models (LSTM, RNN, GRU etc), while image generation can be done with <a href=\"https:\/\/arxiv.org\/abs\/1406.2661\" rel=\"nofollow noreferrer\">GAN - Generative Adversarial Network<\/a><br>\n<br><em>Text generation:<\/em><br>\nFor text generation, first step is to create <a href=\"https:\/\/blogs.oracle.com\/datascience\/introduction-to-embedding-in-natural-language-processing\" rel=\"nofollow noreferrer\">embeddings<\/a> form your sentence either from pre-trained embedding models (word2vec, GloVe etc) then apply this embedding to your sentences. There are several other embedding techniques that you can explore.\n<br>Next step is to fit your embedded features to a sequence model. Probably <a href=\"https:\/\/machinelearningmastery.com\/text-generation-lstm-recurrent-neural-networks-python-keras\/\" rel=\"nofollow noreferrer\">this<\/a> one you can refer as starting point.\n<br><br>Music generation:<br>\n<a href=\"https:\/\/arxiv.org\/abs\/1709.01620\" rel=\"nofollow noreferrer\">Music generation<\/a> can be done with sequence models, difference being instead of sequence of words you have sound waveform, spectrogram, note, chord etc. <br>\n<br>Image generation:<br>\nThis one is different then above two.<br><\/p>\n\n<blockquote>\n  <p>We propose a new framework for estimating generative models via an adversarial process, in which we simultaneously train two models: a generative model G that captures the data distribution, and a discriminative model D that estimates the probability that a sample came from the training data rather than G. The training procedure for G is to maximize the probability of D making a mistake. <br><a href=\"https:\/\/arxiv.org\/abs\/1406.2661\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1406.2661<\/a><\/p>\n<\/blockquote>\n\n<p>Coming to your questions:<br><\/p>\n\n<blockquote>\n  <p>If I'd add more layers and different types of data, could it create for example pictures or music?<\/p>\n<\/blockquote>\n\n<p>As stated, music and text generation can be done with similar type of network architecture (as both follows a sequence), while images needs to be treated differently.<\/p>\n"}]}
+{"QuestionId":58047398,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2019-09-22T08:27:12.020","AcceptedAnswerId":58067780.0,"OwnerUserId":7721538.0,"Title":"Why does \"tf.data.Dataset.from_tensor_slices\" print all paths of images in output?","Body":"<p>I'm writing a code to read images in Tensorflow. I use <a href=\"https:\/\/www.tensorflow.org\/tutorials\/load_data\/images\" rel=\"nofollow noreferrer\">this<\/a> tutorial to do that. The problem is when I'm using this command: <code>tf.data.Dataset.from_tensor_slices((all_image_paths, all_image_labels))<\/code>, the whole paths of images showed up in the output console.<\/p>\n\n<p>This is the code which I'm using:<\/p>\n\n<pre><code>def get_image_info(dir_path, file_url=None, file_name=None, untar=True):\n    if file_url != None:\n        dir_path = tf.keras.utils.get_file(fname=file_name, origin=file_url, untar=untar)\n\n    data_root = pathlib.Path(dir_path)\n    all_image_paths = list(data_root.glob('*\/*'))\n    label_names = sorted(item.name for item in data_root.glob('*\/') if item.is_dir())\n    label_dict = dict((name, index) for index, name in enumerate(label_names))\n    all_image_labels = [label_dict[pathlib.Path(path).parent.name] for path in all_image_paths]\n\n    return data_root, label_dict, all_image_paths, all_image_labels\n\ndef load_image_dataset(dir_path, file_url=None, file_name=None, untar=True):\n    def load_and_preprocess_from_path_label(path, label):\n        return load_and_preprocess_image(path), label\n\n    data_root, label_dict, all_image_paths, all_image_labels = get_image_info(dir_path, file_url, file_name, untar)\n    image_label_ds = tf.data.Dataset.from_tensor_slices((all_image_paths, all_image_labels))\n    # image_label_ds = ds.map(load_and_preprocess_from_path_label)\n\n    return image_label_ds, label_dict\n\nimage_label_ds, label_dict= load_image_dataset('', 'https:\/\/storage.googleapis.com\/download.tensorflow.org\/example_images\/flower_photos.tgz', 'flower_photos')\n<\/code><\/pre>\n\n<p>and this is a section of my output:<\/p>\n\n<pre><code>, WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/8838914676_8ef4db7f50_n.jpg'), WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/8838975946_f54194894e_m.jpg'), WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/8838983024_5c1a767878_n.jpg'), WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/8892851067_79242a7362_n.jpg'), WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/8904780994_8867d64155_n.jpg'), WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/8908062479_449200a1b4.jpg'), WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/8908097235_c3e746d36e_n.jpg'), WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/9019694597_2d3bbedb17.jpg'), WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/9030467406_05e93ff171_n.jpg'), WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/9048307967_40a164a459_m.jpg'), WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/924782410_94ed7913ca_m.jpg'), WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/9378657435_89fabf13c9_n.jpg'), WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/9444202147_405290415b_n.jpg'), WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/9446982168_06c4d71da3_n.jpg'), WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/9831362123_5aac525a99_n.jpg'), WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/9870557734_88eb3b9e3b_n.jpg'), WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/9947374414_fdf1d0861c_n.jpg'), WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/9947385346_3a8cacea02_n.jpg'), WindowsPath('C:\/Users\/hajba\/.keras\/datasets\/flower_photos\/tulips\/9976515506_d496c5e72c.jpg')]. Consider casting elements to a supported type.\n<\/code><\/pre>\n","answers":[{"AnswerId":"58067780","CreationDate":"2019-09-23T17:56:04.717","ParentId":null,"OwnerUserId":"7721538","Title":null,"Body":"<p>For those who face this problem: this is a kind of error which Tensorflow try to show the image paths tensor in output. I use windows os and to solve this error I convert paths type (WindowsPath) to string with this line of code:<\/p>\n\n<pre><code>all_image_paths_str = list(map(lambda x: str(x), all_image_paths))\n<\/code><\/pre>\n\n<p>and then use that for generating output dataset tensor:<\/p>\n\n<pre><code>image_label_ds = tf.data.Dataset.from_tensor_slices((all_image_paths_str, all_image_labels))\n<\/code><\/pre>\n"}]}
+{"QuestionId":58047454,"AnswerCount":3,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2019-09-22T08:34:26.437","AcceptedAnswerId":58639231.0,"OwnerUserId":12102026.0,"Title":"How to fix ' module 'keras.backend.tensorflow_backend' has no attribute '_is_tf_1''","Body":"<p>While training the yolov3 framework, there's always this module error<\/p>\n\n<p>I have tried reinstalling keras and tensorflow, and the version of keras is 2.3.0 and the version of tensorflow is 1.14.0.<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"train.py\", line 6, in &lt;module&gt;\n    import keras.backend as K\n  File \"F:\\Anacoda\\lib\\site-packages\\keras\\__init__.py\", line 3, in &lt;module&gt;\n    from . import utils\n  File \"F:\\Anacoda\\lib\\site-packages\\keras\\utils\\__init__.py\", line 27, in &lt;module&gt;\n    from .multi_gpu_utils import multi_gpu_model\n  File \"F:\\Anacoda\\lib\\site-packages\\keras\\utils\\multi_gpu_utils.py\", line 7, in &lt;module&gt;\n    from ..layers.merge import concatenate\n  File \"F:\\Anacoda\\lib\\site-packages\\keras\\layers\\__init__.py\", line 4, in &lt;module&gt;\n    from ..engine.base_layer import Layer\n  File \"F:\\Anacoda\\lib\\site-packages\\keras\\engine\\__init__.py\", line 8, in &lt;module&gt;\n    from .training import Model\n  File \"F:\\Anacoda\\lib\\site-packages\\keras\\engine\\training.py\", line 21, in &lt;module&gt;\n    from . import training_arrays\n  File \"F:\\Anacoda\\lib\\site-packages\\keras\\engine\\training_arrays.py\", line 14, in &lt;module&gt;\n    from .. import callbacks as cbks\n  File \"F:\\Anacoda\\lib\\site-packages\\keras\\callbacks\\__init__.py\", line 19, in &lt;module&gt;\n    if K.backend() == 'tensorflow' and not K.tensorflow_backend._is_tf_1():\nAttributeError: module 'keras.backend.tensorflow_backend' has no attribute '_is_tf_1'\n<\/code><\/pre>\n","answers":[{"AnswerId":"58095221","CreationDate":"2019-09-25T09:24:09.070","ParentId":null,"OwnerUserId":"10222903","Title":null,"Body":"<p>I fix this problem by replacing keras.XXX to tensorflow.keras.XXX<\/p>\n\n<p>try replace<\/p>\n\n<pre><code>import keras.backend as K\n<\/code><\/pre>\n\n<p>to <\/p>\n\n<pre><code>import tensorflow.keras.backend as K\n<\/code><\/pre>\n"},{"AnswerId":"58155583","CreationDate":"2019-09-29T13:18:35.303","ParentId":null,"OwnerUserId":"12134189","Title":null,"Body":"<p>make sure your keras version is right. if your backbend is tensorflow,you  can<\/p>\n\n<pre><code>import tensorflow as tf\n    print(tf.VERSION)\n    print(tf.keras.__version__)\n<\/code><\/pre>\n\n<p>get the right version of  keras,then install this version,I fix this problem by this way,i hope my answer can help you.<\/p>\n"},{"AnswerId":"58639231","CreationDate":"2019-10-31T08:17:24.550","ParentId":null,"OwnerUserId":"12301807","Title":null,"Body":"<blockquote>\n<pre><code>pip3 uninstall keras\npip3 install keras --upgrade\n<\/code><\/pre>\n<\/blockquote>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/13352\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/13352<\/a><\/p>\n"}]}
+{"QuestionId":58047736,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><tensor><tensorflow-estimator>","CreationDate":"2019-09-22T09:08:51.393","AcceptedAnswerId":58048268.0,"OwnerUserId":7862197.0,"Title":"What's the meaning of the number before the progress bar when tensorflow is training","Body":"<p>Could anyone tell me what's the meaning of  '10' and '49' in the following log of tensorflow?\nMuch Thanks<\/p>\n\n<p>INFO:tensorflow:Started compiling<\/p>\n\n<p>INFO:tensorflow:Finished compiling. Time elapsed: 5.899410247802734 secs<\/p>\n\n<p><strong>10\/10<\/strong> [==============================] - 23s 2s\/step - loss: 2.6726 - acc: 0.1459<\/p>\n\n<p><strong>49\/49<\/strong> [==============================] - 108s 2s\/step - loss: 2.3035 - acc: 0.2845 - val_loss: 2.6726 - val_acc: 0.1459<\/p>\n\n<p>Epoch 2\/100<\/p>\n\n<p><strong>10\/10<\/strong> [==============================] - 1s 133ms\/step - loss: 2.8799 - acc: 0.1693<\/p>\n\n<p><strong>49\/49<\/strong> [==============================] - 17s 337ms\/step - loss: 1.9664 - acc: 0.4042 - val_loss: 2.8799 - val_acc: 0.1693<\/p>\n","answers":[{"AnswerId":"58048268","CreationDate":"2019-09-22T10:28:11.963","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>10 and 49 corresponds to the number of batches which your dataset has been divided into in each epoch. <\/p>\n\n<p>For example, in your train dataset, there are totally 10000 images and your batch size is 64, then there will be totally <code>math.ceil(10000\/64) = 157<\/code> batches possible in each epoch.<\/p>\n"}]}
+{"QuestionId":58047997,"AnswerCount":2,"Tags":"<python><keras><deep-learning><lstm>","CreationDate":"2019-09-22T09:44:47.840","AcceptedAnswerId":null,"OwnerUserId":12102214.0,"Title":"LSTM time series forcasting - starting with low loss, and accuracy does not change","Body":"<p>I'm trying to predict network traffic based on past values. I built an LSTM network, and tried several parameters, however I always end up with the same very low accuracy (0.108).<\/p>\n\n<pre><code>scaler = MinMaxScaler(feature_range = (0, 1))\ndataset = scaler.fit_transform(dataset)\ntrain_size = int(len(dataset) * 0.67)\ntest_size = len(dataset) - train_size\ntrain, test = dataset[0:train_size,:], dataset[train_size:len(dataset),:]\nprint(len(train), len(test))\n\ndef 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\nwindow_size = 1\ntrain_X, train_Y = create_dataset(train, window_size)\ntest_X, test_Y = create_dataset(test, window_size)\nprint(\"Original training data shape:\")\nprint(train_X.shape)\n\n# Reshape the input data into appropriate form for Keras.\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(4, input_shape = (1, window_size)))\nmodel.add(Dense(1))\nopt = optimizers.SGD(lr=0.01, momentum=0.9)\nmodel.compile(loss = \"mean_squared_error\", optimizer = opt, metrics = ['accuracy'])\n<\/code><\/pre>\n\n<p>As you can see my loss starts from quite a low value, and my accuracy constant over time. What am I doing wrong?<\/p>\n\n<p>Thanks in advance. :)<\/p>\n\n<p>You can find the loss and accuracy graph here:\n<a href=\"https:\/\/i.stack.imgur.com\/Otcsc.png\" rel=\"nofollow noreferrer\">loss<\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/RsNEE.png\" rel=\"nofollow noreferrer\">accuracy<\/a><\/p>\n","answers":[{"AnswerId":"58050221","CreationDate":"2019-09-22T14:36:50.233","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>If <code>(13942, 1, 1)<\/code> is your entire dataset, it's far too small for deep learning; you're better off using 'shallow' methods, e.g. Support Vector Machines (SVM). Alternatively, consider my answer <a href=\"https:\/\/stackoverflow.com\/a\/58044089\/10133797\">here<\/a>.<\/p>\n\n<p><strong>EDIT<\/strong>: I just noticed that you use the <code>accuracy<\/code> metric; accuracy for regression is undefined - I'm surprised an error wasn't thrown. If it uses <code>prediction==true<\/code> to compute accuracy, you're lucky your accuracy isn't 0. Based on the loss, your model appears to be actually doing quite well; to double-check, plot <code>predictions<\/code> vs <code>true<\/code> and compare. (In general, <code>mse &lt; .3<\/code> is good, and <code>mse &lt; .1<\/code> is excellent)<\/p>\n"},{"AnswerId":"58051754","CreationDate":"2019-09-22T17:32:28.117","ParentId":null,"OwnerUserId":"4622046","Title":null,"Body":"<ol>\n<li>Try to randomly shuffle the data.<\/li>\n<li>LSTMs are best used when sequential data is used. Try to replace LSTM with the Dense layer. Or change your inputs. You need to pass a sequence of past values to the LSTM so that it can predict the next value. So, (1,1) is not a sequence, LSTM is not useful here.\nMetric accuracy is useless in this context, try mean absolute error or something.<\/li>\n<\/ol>\n"}]}
+{"QuestionId":58048414,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras><tpu>","CreationDate":"2019-09-22T10:46:37.410","AcceptedAnswerId":null,"OwnerUserId":9566491.0,"Title":"Tensorflow TPU Training. InvalidArgumentError: Undeclared output of TPU computation","Body":"<p>I created a ResNet with visual attention layer, trying to train it in TPU device.I compile it successfully, but when I call model.fit, returns <\/p>\n\n<pre><code>Traceback (most recent call last):\n\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/client\/session.py\", line 1370, in _do_call\n    raise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: Undeclared output of TPU computation. A common cause of this error is variable initializers that depend on the TPU computation. Edge: node global_max_pooling2d_12\/Max (defined at attention_resnet.py:268) :0 -&gt; node tf_op_layer_Reshape_1\/Reshape_1 (defined at attention_resnet.py:268) :0\n<\/code><\/pre>\n\n<p>The model's code is the following without intermediate ResBlocks <\/p>\n\n<pre><code>\ndef conv(Input, kernel, filt, stride, dilation, pad='same'):\n    x = layers.Conv2D(filters=filt, kernel_size=kernel, strides=stride, dilation_rate=dilation, padding=pad)(Input)\n    return x\n\ndef conv_group(Input, kernel, filt, stride, dilation, activation=True, pad='same'):\n    x = conv(Input, kernel, filt, stride, dilation, pad=pad)\n    x = layers.BatchNormalization(axis=-1, fused=True)(x)\n    if activation:\n        x = layers.LeakyReLU(alpha=0.1)(x)\n    return x\n\ndef conv_down(Input, filters):\n    x = conv(Input, 3, filters, 2 ,1, pad='valid')\n    x = layers.BatchNormalization(axis=-1, fused=True)(x)\n    x = layers.LeakyReLU(alpha=0.1)(x) \n    return x\n\ndef attention(Input):\n    x = channel_att(Input)\n    x = spatial_att(x)\n    return x\n\ndef channel_att(Input, ratio=8):\n    channel = Input.get_shape()[-1]\n    mlp_0 = layers.Dense(units=channel\/\/ratio, activation=layers.ReLU())\n    mlp_1 = layers.Dense(units=channel, activation=layers.ReLU())\n    ####\n    avg_pool = tf.keras.layers.GlobalAveragePooling2D()(Input)\n    avg_pool = tf.reshape(avg_pool, ([-1,1,1,channel]))\n    avg_ = mlp_1(mlp_0(avg_pool))\n    #####\n    max_pool = tf.keras.layers.GlobalMaxPooling2D()(Input)\n    max_pool = tf.reshape(max_pool, ([-1,1,1,channel]))\n    max_ = mlp_1(mlp_0(max_pool))\n    scale = keras.activations.sigmoid(avg_+max_)\n    return Input*scale\n\n\ndef spatial_att(Input, kernel=7):\n    avg_pool = tf.math.reduce_mean(Input, axis=[3], keepdims=True)\n    max_pool = tf.math.reduce_max(Input, axis=[3], keepdims=True)\n    concat = tf.concat([avg_pool, max_pool], axis=3)\n    concat = layers.Conv2D(filters=1, kernel_size=kernel, padding='same',use_bias=False)(concat)\n    concat = keras.activations.sigmoid(concat)\n    return Input*concat \n\ndef resblock(Input, filters, activ=True):\n    y = conv_group(Input, 3, filters, 1, 2)\n    x = attention(y)\n    x = layers.Add()([x, y])\n    x = conv_group(x, 3, filters, 1, 2, pad='valid', activation=activ)\n    return x\n\n\ndef create_model():\n    Input = layers.Input(shape=(540, 540, 3))\n    x = resblock(Input, 32)\n    ####################\n    x = conv_down(x, 128)\n    ####################\n    x = resblock(x, 128)\n    x = resblock(x, 128)\n    x = conv_down(x, 256)\n    x = resblock(x, 256)\n    x = resblock(x, 128)\n    ####################\n    x = conv_down(x, 64)\n    ####################\n    x = resblock(x, 45, activ=False)\n    model = tf.keras.Model(inputs=Input, outputs=x)\n    model.compile(optimizer=keras.optimizers.SGD(), loss=custom_loss)\n    return model\n\nwith tpu_strategy.scope():\n    model=create_model()\n\nmodel.fit(get_training_dataset(), validation_data=get_validation_dataset(),  initial_epoch=0, steps_per_epoch=steps_per_epoch ,validation_steps=val_steps, epochs=EPOCHS, verbose=1, callbacks=clbk)\n\n<\/code><\/pre>\n\n<p>I use Tensorflow 1.14, but the same error returns with TF-nightly too. I am able to train successfully with GPUs. <\/p>\n","answers":[]}
+{"QuestionId":58048486,"AnswerCount":0,"Tags":"<python><c++><tensorflow><gradient-descent><backpropagation>","CreationDate":"2019-09-22T10:54:21.500","AcceptedAnswerId":null,"OwnerUserId":12044023.0,"Title":"Output shape is (0, 0) when backpropagating","Body":"<p>I created a custom dense to sparse op for tensorflow in c++. I also created a reverse one for the gradients. I registered them in python. But when I train, the output of my reverse op is of shape (0, 0) instead of the input shape of the forward op and I don't understand why.<\/p>\n\n<p>What is strange, however is that my unit test works.<\/p>\n\n<p>I have my forward op make a sparse tensor with dense_shape (b, n, k) from a (b, n) shaped dense tensor  and my reverse op should obviously do the reverse (but doesn't).<\/p>\n\n<p>This is the code I use to register the reverse op in c++.<\/p>\n\n<pre class=\"lang-cpp prettyprint-override\"><code>REGISTER_OP(\"MyOpReverse\")\n        .Input(\"indices: Tint\")\n        .Input(\"values: T\")\n        .Input(\"dense_shape: Tint\")\n        .Output(\"decoded: T\")\n        .Attr(\"T: {bfloat16, float, double}\")\n        .Attr(\"Tint: {int32, int64}\")\n        .SetShapeFn([](InferenceContext *c) {\n            ShapeHandle encoded;\n            TF_RETURN_IF_ERROR(c-&gt;MakeShapeFromShapeTensor(2, &amp;encoded));\n            ShapeHandle out;\n            TF_RETURN_IF_ERROR(c-&gt;Subshape(encoded, 0, 2, &amp;out));\n            c-&gt;set_output(0, out);\n            return Status::OK();\n        });\n<\/code><\/pre>\n\n<p>I expect this to give a shape of (b, n) but apparently it gives (0, 0) when training, even though the forward input was something like (b, 10000), b standing for the batch size.\nIn python I register the gradient like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>@ops.RegisterGradient(\"MyOp\")\ndef _my_op_grad(op, *grad):\n    result = my_op_module.my_op_reverse(*grad)\n    return result\n<\/code><\/pre>\n\n<p>which outputs a Tensor of shape (0, 0) which leads to errors.<\/p>\n\n<p>Any idea how to fix this or where else the error could come from would be appreciated.<\/p>\n","answers":[]}
+{"QuestionId":58049090,"AnswerCount":2,"Tags":"<python-3.x><tensorflow><keras><deep-learning><face-recognition>","CreationDate":"2019-09-22T12:12:44.420","AcceptedAnswerId":null,"OwnerUserId":9401616.0,"Title":"How do i retrain the model without losing the earlier model data with new set of data","Body":"<p>for my current requirement, I'm having a dataset of 10k+ faces from 100 different people from which I have trained a model for recognizing the face(s). The model was trained by getting the 128 vectors from the facenet_keras.h5 model and feeding those vector value to the Dense layer for classifying the faces.\nBut the issue I'm facing currently is<\/p>\n\n<ol>\n<li>if want to train one person face, I have to retrain the whole model once again.<\/li>\n<\/ol>\n\n<p>How should I get on with this challenge? I have read about a concept called transfer learning but I have no clues about how to implement it. Please give your suggestion on this issue. What can be the possible solutions to it?<\/p>\n","answers":[{"AnswerId":"58049181","CreationDate":"2019-09-22T12:24:31.927","ParentId":null,"OwnerUserId":"12097394","Title":null,"Body":"<p>You can save your training results by saving your weights with:<\/p>\n\n<pre><code>model.save_weights('my_model_weights.h5')\n<\/code><\/pre>\n\n<p>And load them again later to resume your training after you added a new image to the dataset with:<\/p>\n\n<pre><code>model.load_weights('my_model_weights.h5')\n<\/code><\/pre>\n"},{"AnswerId":"58049249","CreationDate":"2019-09-22T12:33:10.170","ParentId":null,"OwnerUserId":"9149062","Title":null,"Body":"<p>With transfer learning you would copy an existing pre-trained model and use it for a different, but similar, dataset from the original one. In your case this would be what you need to do if you want to train the model to recognize your specific 100 people.<\/p>\n\n<p>If you already did this and you want to add another person to the database without having to retrain the complete model, then I would freeze all layers (set <code>layer.trainable = False<\/code> for all layers) except for the final fully-connected layer (or the final few layers). Then I would replace the last layer (which had 100 nodes) to a layer with 101 nodes. You could even copy the weights to the first 100 nodes and maybe freeze those too (I'm not sure if this is possible in Keras). In this case you would re-use all the trained convolutional layers etc. and teach the model to recognise this new face.<\/p>\n"}]}
+{"QuestionId":58049311,"AnswerCount":0,"Tags":"<python><numpy><pytorch><noise-generator>","CreationDate":"2019-09-22T12:40:16.847","AcceptedAnswerId":null,"OwnerUserId":4283749.0,"Title":"How do I add Poisson noise to a greyscale image in Python using numpy?","Body":"<p>I'm trying to implement adding Poisson noise to a greyscale image using <code>numpy<\/code> as a <code>Pytorch<\/code> transformer but so far my results have been very disappointing. Here are my various implementations with results ranging from seeing no visible changes to only seeing noise without an image. What is the correct way of doing this?<\/p>\n\n<pre><code>class Noise(object):\n  def __init__(self, noise_type):\n    self.noise_type = noise_type\n\n  def __call__(self, sample):\n    noise = np.random.poisson(1, sample.shape)\n    return sample + noise\n<\/code><\/pre>\n\n<p>but as I understand it, Poisson noise can't be simply added to an image which is why all I see here is noise (kind of like TV static) without the original greyscale image. <\/p>\n\n<p>The following version of the code yields no visible changes in the image which leads me to believe it's some sort of scaling issue since in Poisson noise, noise is added relative to a pixel's value. <\/p>\n\n<pre><code>vals = len(np.unique(sample))\nvals = 2 ** np.ceil(np.log2(vals))\nnoisy = np.random.poisson(sample * vals) \/ float(vals)\nreturn noisy\n<\/code><\/pre>\n\n<p>Any help or insight would be greatly appreciated.<\/p>\n","answers":[]}
+{"QuestionId":58049454,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-22T12:58:09.850","AcceptedAnswerId":58049659.0,"OwnerUserId":10295417.0,"Title":"is this model overfitting or underfitting?","Body":"<p>in the case of overfitting, to my knowledge the val_loss has to soar as opposed to the train_loss\nbut how about this?? is this model underfitting horribly? or is it some completely different case?\npreviously my models would overfit badly so i added the dropout of 0.3 (4 CuDNNGRU layers with 64 neurons and one Dense layer and batchsize of 64), so should i reduce the dropout? <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/1SoVK.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1SoVK.png\" alt=\"train_loss vs. validation_loss\"><\/a><\/p>\n","answers":[{"AnswerId":"58049659","CreationDate":"2019-09-22T13:27:04.033","ParentId":null,"OwnerUserId":"1729802","Title":null,"Body":"<p>This is neither overfitting nor underfitting. Some people refer to it as <em><a href=\"https:\/\/stats.stackexchange.com\/questions\/187335\/validation-error-less-than-training-error\">Unknown fit<\/a><\/em>. Validation &lt;&lt; training loss happens when you apply regularization (L1, L2, Dropout, ...) in keras because they are applied to training only and not on testing (validating). So it makes sense that your training loss is bigger (not all neurons are available for feed forward due to dropout for example).<\/p>\n\n<p>But what is clear is that your model is not being optimized for your validation set, (almost a flat line). This can be due to many things:<\/p>\n\n<ul>\n<li>Your validation set is badly representative of your dataset, has a very easy predictions or very small.<\/li>\n<li>decrease learning rate or add more regularization (recurrent_regularization since you are using CuDNNGRU)<\/li>\n<li>Your loss function is not appropriate for the problem you are trying to solve.<\/li>\n<\/ul>\n\n<p>Hope these tips help you out.<\/p>\n"}]}
+{"QuestionId":58049485,"AnswerCount":0,"Tags":"<python><tensorflow><graph>","CreationDate":"2019-09-22T13:02:36.947","AcceptedAnswerId":null,"OwnerUserId":6077080.0,"Title":"How does Tensorflow construct a graph from your code arguments?","Body":"<p>I am trying to understand how Tensorflow is able to accept an expression as an argument and turn it into a graph in Python<\/p>\n\n<p>I have tried looking at the Tensorflow code base (e.g. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.13.1\/tensorflow\/python\/ops\/nn_ops.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.13.1\/tensorflow\/python\/ops\/nn_ops.py<\/a>), and maybe I'm just being dense but I can't see how they are doing it from this.<\/p>\n\n<p>An example might be:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>b = tf.Variable(tf.zeros((100,)))\nW = tf.Variable(tf.random_uniform((784, 100), -1, 1))\nx = tf.placeholder(tf.float32,(100,784))\n\nh = tf.nn.relu(tf.matmul(x, W) + b)\n<\/code><\/pre>\n\n<p>It is the definition of 'h' here that interests me. How does tf\/Python facilitate turning matmul + b into 2 operands and an add operation in the graph? I haven't found any declaration for Python that allows <code>tf.matmul(x, W) + b<\/code> to be tokenized and registered as graph elements as I thought I might expect to see in the declaration of tf.nn.relu().<\/p>\n","answers":[]}
+{"QuestionId":58049625,"AnswerCount":1,"Tags":"<python><computer-vision><pytorch>","CreationDate":"2019-09-22T13:22:35.547","AcceptedAnswerId":null,"OwnerUserId":11929946.0,"Title":"Unet input and output","Body":"<p>I am totally new to computer vision, and am trying to train a model that takes an normal 3 channel image (RGB) and gives an output, which is just a simple binary mask. From my best knowledge, architecture called Unet does something like this. For example, TernausnetV2 worked quite well in a kaggle competition. In this kind of model architecture, is the input and output I'm thinking of correct ?<\/p>\n\n<p>I have 3 channel image (RGB) and the binary masked image (which is same size as the 3 channel image, and all the values are 1 or 0).\nThanks for the help in prior.<\/p>\n","answers":[{"AnswerId":"58050901","CreationDate":"2019-09-22T15:52:31.597","ParentId":null,"OwnerUserId":"12067496","Title":null,"Body":"<p>The computer vision task you are talking abour is called a semantic segmentation and the U-Net model is a basic autoencoder architecture used for problem. There are many models nowadays derived from the U-Net (Ternausnet one of them), some comparison can be found in the Papers with code SOTA comparison <a href=\"https:\/\/paperswithcode.com\/task\/semantic-segmentation\" rel=\"nofollow noreferrer\">link<\/a>.\nU-Net architecture usually uses 3 channel input size for RGB image processing and 1 channel output size - models usually use as a final layer a sigmoid activation convolution layer with core size 1x1 which basically goes through the entire image and on every pixel computes sigmoid function \n- wikipedia <a href=\"https:\/\/en.wikipedia.org\/wiki\/Sigmoid_function\" rel=\"nofollow noreferrer\">link<\/a> which assignes every pixel a values between 0 and 1 (depending on the data magnitude you can use also tanh - -1\/1). More information can be found in this awesome tutorial from Towards data science <a href=\"https:\/\/towardsdatascience.com\/understanding-semantic-segmentation-with-unet-6be4f42d4b47\" rel=\"nofollow noreferrer\">link<\/a><\/p>\n"}]}
+{"QuestionId":58049661,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-federated>","CreationDate":"2019-09-22T13:27:11.570","AcceptedAnswerId":null,"OwnerUserId":11283804.0,"Title":"federated_learning_for_image_classification.ipynb code error","Body":"<p>I use tensorflow-federated learning api.\nAnd recently, I update tensorflow-federated to 0.8.0 version.\nAnd then I run the federated_learning_for_image_classification.ipynb file.\nBut It did not work on \"<code>state = iterative_process.initialize()<\/code>\" code line.<\/p>\n\n<p>Some error is happened I can not do next.<\/p>\n\n<p>Why was it happened? Previous version of federated api 0.6.0 was working great.<\/p>\n\n<p>you can refer to the below code and error picture.<\/p>\n\n<pre><code>import nest_asyncio\nnest_asyncio.apply()\n\nfrom __future__ import absolute_import, division, print_function\n\nimport collections\n\nimport warnings\nfrom six.moves import range\nimport numpy as np\nimport six\nimport tensorflow as tf\nimport tensorflow_federated as tff\n\nwarnings.simplefilter('ignore');\n\ntf.compat.v1.enable_v2_behavior();\n\nnp.random.seed(0);\n\nif six.PY3:\ntff.framework.set_default_executor(tff.framework.create_local_executor());\n\ntff.federated_computation(lambda: 'Hello, World!')();\n\nemnist_train, emnist_test = tff.simulation.datasets.emnist.load_data();\n\nlen(emnist_train.client_ids);\n\nprint(emnist_train.client_ids, emnist_test.client_ids);\n\nemnist_train.output_types, emnist_train.output_shapes;\n\nexample_dataset = emnist_train.create_tf_dataset_for_client(\nemnist_train.client_ids[0]);\n\nexample_element = iter(example_dataset).next();\n\nexample_element['label'].numpy();\n\nfrom matplotlib import pyplot as plt;\nplt.imshow(example_element['pixels'].numpy(), cmap='gray', aspect='equal');\nplt.grid('off')\n_ = plt.show()\n\nNUM_CLIENTS = 5\nNUM_EPOCHS = 10\nBATCH_SIZE = 20\nSHUFFLE_BUFFER = 500\n\ndef preprocess(dataset):\n\n  def element_fn(element):\n    return collections.OrderedDict([\n        ('x', tf.reshape(element['pixels'], [-1])),\n        ('y', tf.reshape(element['label'], [1])),\n    ])\n\n  return dataset.repeat(NUM_EPOCHS).map(element_fn).shuffle(\n  SHUFFLE_BUFFER).batch(BATCH_SIZE)\n\npreprocessed_example_dataset = preprocess(example_dataset)\n\nsample_batch = tf.nest.map_structure(\n    lambda x: x.numpy(), iter(preprocessed_example_dataset).next())\n\nsample_batch\n\ndef make_federated_data(client_data, client_ids):\n  return [preprocess(client_data.create_tf_dataset_for_client(x))\n      for x in client_ids]\n\nsample_clients = emnist_train.client_ids[0:NUM_CLIENTS]\n\nprint(sample_clients)\n\nfederated_train_data = make_federated_data(emnist_train, sample_clients)\n\nlen(federated_train_data), federated_train_data[0]\n\ndef create_compiled_keras_model():\n  model = tf.keras.models.Sequential([\n      tf.keras.layers.Dense(\n          10, activation=tf.nn.softmax, kernel_initializer='zeros', \ninput_shape=(784,))])\n\n  model.compile(\n      loss=tf.keras.losses.SparseCategoricalCrossentropy(),\n      optimizer=tf.keras.optimizers.SGD(learning_rate=0.02),\n      metrics=[tf.keras.metrics.SparseCategoricalAccuracy()])\n  return model\n\ndef model_fn():\n  keras_model = create_compiled_keras_model()\n  return tff.learning.from_compiled_keras_model(keras_model, sample_batch)\n\niterative_process = tff.learning.build_federated_averaging_process(model_fn)\n\nstr(iterative_process.initialize.type_signature)\n\nstate = iterative_process.initialize()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/P3vGe.png\" rel=\"nofollow noreferrer\">error picture-1<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/DL5qo.png\" rel=\"nofollow noreferrer\">error picture-2<\/a><\/p>\n","answers":[{"AnswerId":"58087794","CreationDate":"2019-09-24T20:37:23.937","ParentId":null,"OwnerUserId":"7507932","Title":null,"Body":"<p>Please see the Github issue <a href=\"https:\/\/github.com\/tensorflow\/federated\/issues\/764\" rel=\"nofollow noreferrer\">here<\/a>; essentially, master is out of sync with the pip package. If you explicitly set the number of clients in the local executor, this should resolve itself.<\/p>\n\n<p>Thanks for your interest in TFF!<\/p>\n"}]}
+{"QuestionId":58049678,"AnswerCount":0,"Tags":"<tensorflow><artificial-intelligence><object-detection><image-recognition><faster-rcnn>","CreationDate":"2019-09-22T13:29:20.757","AcceptedAnswerId":null,"OwnerUserId":7340343.0,"Title":"Do i have to resize Image for SSD Object detector if i'm already using LabelImg annotation tool?","Body":"<p>Actually i'm having a hard time here and i would appreciate any help or explanation , first i tried using faster_Rcnn_inceptionV2 to train it on my custom annotated dataSet which contains images with different dimensions and it worked perfectly , but when i tried to train it again using SSD_MobileNetV1 i got really bad accuracy and when i checked the MobileNetV1.config file i saw that the model is resizing my images to \n300 x 300  <\/p>\n\n<pre><code>image_resizer {\n      fixed_shape_resizer {\n        height: 300\n        width: 300\n      }\n    }\n<\/code><\/pre>\n\n<p>While on the Faster Rcnn Model : <\/p>\n\n<pre><code>image_resizer {\n      keep_aspect_ratio_resizer {\n        min_dimension: 600\n        max_dimension: 1024\n      }\n    }\n<\/code><\/pre>\n\n<p>so what am i missing here ? do i have to resize all my images and annotate them again ?<\/p>\n\n<p>and if so that would be hard because i'm trying to detect small objects (BlackHeads \/ WhiteHeads ) <\/p>\n","answers":[]}
+{"QuestionId":58050020,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-09-22T14:12:52.890","AcceptedAnswerId":null,"OwnerUserId":10487245.0,"Title":"Is there a handy way to dump the running_stats for a pytorch model?","Body":"<p>I'm writing a C version of the pytorch model to run it on my special hardware.\nEverything looks ok so far, except the running_mean and running_var in every batchnorm layer.<\/p>\n\n<p>We have a python code to dump all named_parameters, but nothing to do for the running_stats, although we need to use it in the forwarding computation.<\/p>\n\n<p>So is there a way to dump it with sort of built-in features?\nI searched pytorch doc, no help on my task.\nOtherwise I might need to write a regexp code to recognize and dump them.<\/p>\n\n<p>Thanks a lot.\n\/Patrick<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>for name, param in model.named_parameters():\n    # here can dump weight and bias, but not running_stats\n    names.append(name)\n    shapes.append(list(param.data.numpy().shape))\n    values.append(param.data.numpy().flatten().tolist())\n<\/code><\/pre>\n","answers":[{"AnswerId":"58051263","CreationDate":"2019-09-22T16:34:19.857","ParentId":null,"OwnerUserId":"10886420","Title":null,"Body":"<p><code>running_mean<\/code> and others are <code>registered_buffers<\/code> in PyTorch. You can save (as you say dump) them with <code>torch.nn.Module<\/code>'s <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/saving_loading_models.html#what-is-a-state-dict\" rel=\"nofollow noreferrer\"><code>state_dict<\/code><\/a>:<\/p>\n\n<pre><code>torch.save(model.state_dict(), PATH) \n<\/code><\/pre>\n\n<p>You can iterate over named buffers and save each of them however you like similarly to parameters:<\/p>\n\n<pre><code>for name, buffer in model.named_buffers():\n    # do your thing with them\n<\/code><\/pre>\n"}]}
+{"QuestionId":58050105,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><deep-learning><pytorch>","CreationDate":"2019-09-22T14:22:56.617","AcceptedAnswerId":58050805.0,"OwnerUserId":11520760.0,"Title":"why does the Accuracy decrease when using a ReLu activation after Linear layers","Body":"<p>so I started using Pytorch and I'm building a very basic CNN on the FashionMNIST Dataset. I noticed some weird Behaviour while using the NN and I don't know why this is happening, in the Forward Function the Accuracy of the NN decrease when I use a Relu function after every Linear Layer. <\/p>\n\n<p>here is the Code for my custom NN:<\/p>\n\n<pre><code># custom class neural network \nclass FashionMnistClassifier(nn.Module):\n  def __init__(self, n_inputs, n_out):\n    super().__init__()\n    self.cnn1 = nn.Conv2d(n_inputs, out_channels=32, kernel_size=5).cuda(device)\n    self.cnn2 = nn.Conv2d(32, out_channels=64, kernel_size=5).cuda(device)\n    #self.cnn3 = nn.Conv2d(n_inputs, out_channels=32, kernel_size=5)\n    self.fc1 = nn.Linear(64*4*4, out_features=100).cuda(device)\n    self.fc2 = nn.Linear(100, out_features=n_out).cuda(device)\n    self.relu = nn.ReLU().cuda(device)\n    self.pool = nn.MaxPool2d(kernel_size=2).cuda(device)\n    self.soft_max = nn.Softmax().cuda(device)\n\n  def forward(self, x):\n    x.cuda(device)\n    out = self.relu(self.cnn1(x))\n    out = self.pool(out)\n    out = self.relu(self.cnn2(out))\n    out = self.pool(out)\n    #print(\"out shape in classifier forward func: \", out.shape)\n    out = self.fc1(out.view(out.size(0), -1))\n    #out = self.relu(out) # if I uncomment these then the Accuracy decrease from 90 to 50!!!\n    out = self.fc2(out)\n    #out = self.relu(out) # this too\n    return out\n\nn_batch = 100\nn_outputs = 10\nLR = 0.001\n\nmodel = FashionMnistClassifier(1, 10).cuda(device)\noptimizer = optim.Adam(model.parameters(), lr=LR)\ncriterion = nn.CrossEntropyLoss()\n<\/code><\/pre>\n\n<p>so if I use the ReLu only after the CNN layers I get an Accuracy of 90% but when I uncomment that section and use the Relu activation after the Linear Layers the Accuracy decrease to 50% I have no idea why is this happening since I thought it is always better to use activation after every linear layer to get a better accuracy for Classification. I thought always that we should always use the activation function if we have a Classification Problem and for Linear Regression we don't have to do this, but here in my case although it is a Classification Problem, I get a better performance if I don't use the Activation Function after the Linear Layers. Can someone maybe clarify this to me?<\/p>\n","answers":[{"AnswerId":"58050805","CreationDate":"2019-09-22T15:43:07.653","ParentId":null,"OwnerUserId":"10886420","Title":null,"Body":"<p><code>CrossEntropyLoss<\/code> requires you to pass in unnormalized logits (output from last <code>Linear<\/code> layer).<\/p>\n\n<p>If you use <code>ReLU<\/code> as the output from last layer you are only outputting values in the range <code>[0, inf)<\/code>, while neural network tends to go with small values for incorrect labels and high for the correct ones (we may say it's overconfident in it's predictions). Oh, and the one with highest logit value is chosen by <code>argmax<\/code> as correct label.<\/p>\n\n<p>So it won't definitely work with this line:<\/p>\n\n<pre><code># out = self.relu(out) # this too\n<\/code><\/pre>\n\n<p>Though it should with the <code>ReLU<\/code> before it. Just remember, more nonlinearity isn't always good for the network.<\/p>\n"}]}
+{"QuestionId":58050113,"AnswerCount":1,"Tags":"<python><tensorflow><keras><semantic-segmentation>","CreationDate":"2019-09-22T14:23:46.307","AcceptedAnswerId":58050474.0,"OwnerUserId":9376936.0,"Title":"ImageDataGenerator for semantic segmentation","Body":"<p>I am trying to do semantic segmentation with Keras and when trying to load the images i get this error using <code>flow_from_directory<\/code> method.<\/p>\n\n<pre><code>Found 0 images belonging to 0 classes.\nFound 0 images belonging to 0 classes.\n<\/code><\/pre>\n\n<p>This is my code.<\/p>\n\n<pre><code>from tensorflow.keras.applications.resnet50 import preprocess_input\nfrom tensorflow.keras.preprocessing.image import ImageDataGenerator\n\ndata_generator = ImageDataGenerator()\ntrain_generator = data_generator.flow_from_directory(\n                                        directory=\"..\/input\/Training_dataset\/Images\",\n                                        target_size=(IMG_SIZE, IMG_SIZE),\n                                        batch_size=16,\n                                        class_mode=None,\n                                        classes=None\n                                        )\n\nmask_generator = data_generator.flow_from_directory(\n    directory=\"..\/input\/Training_dataset\/Masks\/all\",\n    class_mode=None,\n    classes=None,\n    batch_size = 1,\n    )\n<\/code><\/pre>\n\n<p>I have read this question but solution  didnt work <a href=\"https:\/\/stackoverflow.com\/questions\/45510403\/keras-for-semantic-segmentation-flow-from-directory-error\">Keras for semantic segmentation, flow_from_directory() error<\/a><\/p>\n","answers":[{"AnswerId":"58050474","CreationDate":"2019-09-22T15:05:47.683","ParentId":null,"OwnerUserId":"7331631","Title":null,"Body":"<p>you need to keep your images inside one sub-folder like create a folder named \"img\" inside both image and mask directory.<\/p>\n\n<pre><code>-- image\n   -- img\n      -- 1.jpg\n      -- 2.jpg\n-- mask\n   -- img\n      -- 1.png\n      -- 2.png\n<\/code><\/pre>\n\n<p>Datagenerator should be like:- <\/p>\n\n<pre><code>seed = 909 # (IMPORTANT) to transform image and corresponding mask with same augmentation parameter.\nimage_datagen = ImageDataGenerator(width_shift_range=0.1,\n                 height_shift_range=0.1,\n                 preprocessing_function = image_preprocessing) # custom fuction for each image you can use resnet one too.\nmask_datagen = ImageDataGenerator(width_shift_range=0.1,\n                 height_shift_range=0.1,\n                 preprocessing_function = mask_preprocessing)  # to make mask as feedable formate (256,256,1)\n\nimage_generator =image_datagen.flow_from_directory(\"dataset\/image\/\",\n                                                    class_mode=None, seed=seed)\n\nmask_generator = mask_datagen.flow_from_directory(\"dataset\/mask\/\",\n                                                   class_mode=None, seed=seed)\n\ntrain_generator = zip(image_generator, mask_generator)\n<\/code><\/pre>\n\n<p>if you want to make your own custom data generator for semantic segmentation models to get better control over dataset, you can check my kaggle kernel where i have used camvid dataset to train UNET model.<\/p>\n\n<p><a href=\"https:\/\/www.kaggle.com\/mukulkr\/camvid-segmentation-using-unet\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/mukulkr\/camvid-segmentation-using-unet<\/a><\/p>\n\n<p>If you need better augmentation fuction you can check this awesome GitHub repo -\n<a href=\"https:\/\/github.com\/mdbloice\/Augmentor\" rel=\"nofollow noreferrer\">https:\/\/github.com\/mdbloice\/Augmentor<\/a><\/p>\n"}]}
+{"QuestionId":58050156,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-22T14:29:15.427","AcceptedAnswerId":58100429.0,"OwnerUserId":4969574.0,"Title":"Cannot load model in keras from Model.get_config() when the model has Attention layer","Body":"<p>I'm loading a model in keras from a config dict created by some function. I've tried this scheme with many other models without issues, but this is the first one were i'm using <code>tensorflow.keras.layers.Attention<\/code> and i'm getting an Unknown Layer exception when reading from the config. <\/p>\n\n<p>I know there is an API to serialize and load custom layers using JSON\/YAML, but this is a keras layer, am I doing something wrong?<\/p>\n\n<p>by the way this is using Tensorflow 1.14.0<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.keras import layers, models, utils\n\ndef my_model(max_len, vocab_size, embedding_dims):\n    sequence = layers.Input(shape=(max_len,), name='sequence')\n    feature = layers.Input(shape=(1,), name='another_feature')\n\n    x = layers.Embedding(input_dim=vocab_size,\n                         output_dim=embedding_dims,\n                         input_length=max_len)(sequence)\n    out, sh, sc = layers.LSTM(64, return_state=True)(x)\n    att = layers.Attention()([out, sh])\n    x = layers.concatenate([att, feature])\n    model = models.Model(inputs=[sequence, feature], outputs=[x])\n    model.summary()\n    return model.get_config()\n\ntitle_max_len = 50\nvocab_size = 35000\nembedding_dims = 30\n\nconfig = my_model(\n    title_max_len, \n    vocab_size, \n    embedding_dims\n)\nmodel = models.Model.from_config(config)  # Unknown layer: Attention\nutils.plot_model(\n    model, \n    show_shapes=True, \n    show_layer_names=True, \n    to_file='model.png'\n)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58100429","CreationDate":"2019-09-25T14:12:10.803","ParentId":null,"OwnerUserId":"4325492","Title":null,"Body":"<p>It looks like the problem exists for Tensorflow 1.14 but it works without any problems in 1.15. If you're able to upgrade then that may be the easiest solution. <\/p>\n"}]}
+{"QuestionId":58050284,"AnswerCount":0,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2019-09-22T14:45:32.063","AcceptedAnswerId":null,"OwnerUserId":10551411.0,"Title":"TensorBoard duplicate plugins","Body":"<p>I am using keras for my python project and I wanted to visualize how different networks work.<\/p>\n\n<pre><code>model.fit(x, y, batch_size=32, epochs=3, validation_split=0.3, callbacks=[tensorboard])\n<\/code><\/pre>\n\n<p>I wanted to use tensorboard so I installed tensorflow 1.14.0 but I kept getting this error <\/p>\n\n<blockquote>\n  <p>ImportError: You need the TensorFlow (v1) module installed to use TensorBoard.<\/p>\n<\/blockquote>\n\n<p>After that, I installed tensorflow 2.0.0 and the python project ran successfully. But when I tried to run tensorboad command in cmd I got this <\/p>\n\n<blockquote>\n  <p>ValueError: Duplicate plugins for name projector.<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":58050557,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning>","CreationDate":"2019-09-22T15:13:56.453","AcceptedAnswerId":null,"OwnerUserId":11247761.0,"Title":"Keras is unable to intialize","Body":"<p>I am trying to intialize withe the parameters but i get the error of \"<em>init<\/em>() got an unexpected keyword argument scale\"<\/p>\n\n<p>I am working on Neural collaborative filtering with keras and cannot initialize my parameters.<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nimport keras\nfrom keras import backend as K\nfrom keras import initializers\nfrom keras.regularizers import l1, l2, l1_l2\nfrom keras.models import Sequential, Model\nfrom keras.layers.core import Dense, Lambda, Activation\nfrom keras.layers import Embedding, Input, Dense, merge, Reshape, Merge, Flatten, Dropout\nfrom keras.optimizers import Adagrad, Adam, SGD, RMSprop\nfrom evaluate import evaluate_model\nfrom Dataset import Dataset\nfrom time import time\nimport sys\nimport GMF, MLP\nimport argparse\n\n#################### Arguments ####################\ndef parse_args():\n    parser = argparse.ArgumentParser(description=\"Run NeuMF.\")\n    parser.add_argument('--path', nargs='?', default='Data\/',\n                        help='Input data path.')\n    parser.add_argument('--dataset', nargs='?', default='ml-1m',\n                        help='Choose a dataset.')\n    parser.add_argument('--epochs', type=int, default=100,\n                        help='Number of epochs.')\n    parser.add_argument('--batch_size', type=int, default=256,\n                        help='Batch size.')\n    parser.add_argument('--num_factors', type=int, default=8,\n                        help='Embedding size of MF model.')\n    parser.add_argument('--layers', nargs='?', default='[64,32,16,8]',\n                        help=\"MLP layers. Note that the first layer is the concatenation of user and item embeddings. So layers[0]\/2 is the embedding size.\")\n    parser.add_argument('--reg_mf', type=float, default=0,\n                        help='Regularization for MF embeddings.')                    \n    parser.add_argument('--reg_layers', nargs='?', default='[0,0,0,0]',\n                        help=\"Regularization for each MLP layer. reg_layers[0] is the regularization for embeddings.\")\n    parser.add_argument('--num_neg', type=int, default=4,\n                        help='Number of negative instances to pair with a positive instance.')\n    parser.add_argument('--lr', type=float, default=0.001,\n                        help='Learning rate.')\n    parser.add_argument('--learner', nargs='?', default='adam',\n                        help='Specify an optimizer: adagrad, adam, rmsprop, sgd')\n    parser.add_argument('--verbose', type=int, default=1,\n                        help='Show performance per X iterations')\n    parser.add_argument('--out', type=int, default=1,\n                        help='Whether to save the trained model.')\n    parser.add_argument('--mf_pretrain', nargs='?', default='',\n                        help='Specify the pretrain model file for MF part. If empty, no pretrain will be used')\n    parser.add_argument('--mlp_pretrain', nargs='?', default='',\n                        help='Specify the pretrain model file for MLP part. If empty, no pretrain will be used')\n    return parser.parse_args()\n\ndef init_normal(shape, name=None):\n    return initializers.normal(shape, scale=0.01, name=name)\n<\/code><\/pre>\n\n<p>Error:<\/p>\n\n<blockquote>\n  <p>TypeError: <strong>init<\/strong>() got an unexpected keyword argument 'scale'<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"58052285","CreationDate":"2019-09-22T18:31:28.137","ParentId":null,"OwnerUserId":"11351803","Title":null,"Body":"<p><code>initializers.normal<\/code> doesn't contain the arguments <code>name<\/code> and <code>scale<\/code>. Only <code>shape<\/code>. <\/p>\n"}]}
+{"QuestionId":58051377,"AnswerCount":1,"Tags":"<r><machine-learning><keras><deep-learning>","CreationDate":"2019-09-22T16:47:58.113","AcceptedAnswerId":null,"OwnerUserId":5215498.0,"Title":"What can be the cause for difference in MAE outcome from deep-learing with R between these datasets?","Body":"<p>I\u2019m trying to replicate the deep learning example below with the same Boston housing dataset from another source.<\/p>\n\n<p><a href=\"https:\/\/jjallaire.github.io\/deep--with-r-notebooks\/notebooks\/3.6-predicting-house-prices.nb.html\" rel=\"nofollow noreferrer\">https:\/\/jjallaire.github.io\/deep--with-r-notebooks\/notebooks\/3.6-predicting-house-prices.nb.html<\/a><\/p>\n\n<p>Originally the data source is:<\/p>\n\n<pre><code>library(keras) dataset &lt;- dataset_boston_housing()\n<\/code><\/pre>\n\n<p>Alternatively I try to use:<\/p>\n\n<pre><code>library(mlbench)\ndata(BostonHousing)\n<\/code><\/pre>\n\n<p>The difference between the datasets are:<\/p>\n\n<ol>\n<li>the dataset from mlbench contains column names.<\/li>\n<li>the dataset from keras is already split between test and train.<\/li>\n<li>the set from keras is organised with lists containing matrices while the dataset from mlbench is a dataframe<\/li>\n<li>the fourth column contains a categorical variable \"chas\" which could not be preprocessed from the mlbench dataset while it can be preprocessed from the keras dataset. To compare apples with apples I have deleted this column from both datasets.<\/li>\n<\/ol>\n\n<p>In order to compare both datasets I have merged the train and testset from keras into 1 dataset. After this I have compared the merged dataset from keras with mlbench with summary() and these are identical for every feature (min, max, median, mean). <\/p>\n\n<p>Since the dataset from keras is already split between test and train (80-20), I can only use one training set for the deep learning proces. This training set gives a validation_mae of around 2.5. See this graph:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vHrnM.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/vHrnM.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>If I partition the data from mlbench at 0.8 to construct a training set of similar size, run the deep learing code and do this several times, I never reach a validation_mae of around 2.5. The range is between 4 and 6. An example of the output is this graph:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ZcKOo.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ZcKOo.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Does someone know what can be the cause for this difference?<\/p>\n\n<p>Code with dataset from keras:<\/p>\n\n<pre class=\"lang-r prettyprint-override\"><code>\nlibrary(keras)\ndataset &lt;- dataset_boston_housing()\n\nc(c(train_data, train_targets), c(test_data, test_targets)) %&lt;-% dataset\n\ntrain_data &lt;- train_data[,-4]\ntest_data &lt;- test_data[,-4]\n\nmean &lt;- apply(train_data, 2, mean)\nstd &lt;- apply(train_data, 2, sd)\ntrain_data &lt;- scale(train_data, center = mean, scale = std)\ntest_data &lt;- scale(test_data, center = mean, scale = std)\n\n# After this line the code is the same for both  code examples.\n# ========================================= \n\n# Because we will need to instantiate the same model multiple times,\n# we use a function to construct it.\nbuild_model &lt;- function() {\n  model &lt;- keras_model_sequential() %&gt;% \n    layer_dense(units = 64, activation = \"relu\", \n                input_shape = dim(train_data)[[2]]) %&gt;% \n    layer_dense(units = 64, activation = \"relu\") %&gt;% \n    layer_dense(units = 1) \n\n  model %&gt;% compile(\n    optimizer = \"rmsprop\", \n    loss = \"mse\", \n    metrics = c(\"mae\")\n  )\n}\n\n  k &lt;- 4\nindices &lt;- sample(1:nrow(train_data))\nfolds &lt;- cut(1:length(indices), breaks = k, labels = FALSE) \nnum_epochs &lt;- 100\nall_scores &lt;- c()\nfor (i in 1:k) {\n  cat(\"processing fold #\", i, \"\\n\")\n  # Prepare the validation data: data from partition # k\n  val_indices &lt;- which(folds == i, arr.ind = TRUE) \n  val_data &lt;- train_data[val_indices,]\n  val_targets &lt;- train_targets[val_indices]\n\n  # Prepare the training data: data from all other partitions\n  partial_train_data &lt;- train_data[-val_indices,]\n  partial_train_targets &lt;- train_targets[-val_indices]\n\n  # Build the Keras model (already compiled)\n  model &lt;- build_model()\n\n  # Train the model (in silent mode, verbose=0)\n  model %&gt;% fit(partial_train_data, partial_train_targets,\n                epochs = num_epochs, batch_size = 1, verbose = 0)\n\n  # Evaluate the model on the validation data\n  results &lt;- model %&gt;% evaluate(val_data, val_targets, verbose = 0)\n  all_scores &lt;- c(all_scores, results$mean_absolute_error)\n}  \nall_scores\nmean(all_scores)\n\n  # Some memory clean-up\n  k_clear_session()\nnum_epochs &lt;- 500\nall_mae_histories &lt;- NULL\nfor (i in 1:k) {\n  cat(\"processing fold #\", i, \"\\n\")\n\n  # Prepare the validation data: data from partition # k\n  val_indices &lt;- which(folds == i, arr.ind = TRUE)\n  val_data &lt;- train_data[val_indices,]\n  val_targets &lt;- train_targets[val_indices]\n\n  # Prepare the training data: data from all other partitions\n  partial_train_data &lt;- train_data[-val_indices,]\n  partial_train_targets &lt;- train_targets[-val_indices]\n\n  # Build the Keras model (already compiled)\n  model &lt;- build_model()\n\n  # Train the model (in silent mode, verbose=0)\n  history &lt;- model %&gt;% fit(\n    partial_train_data, partial_train_targets,\n    validation_data = list(val_data, val_targets),\n    epochs = num_epochs, batch_size = 1, verbose = 1\n  )\n  mae_history &lt;- history$metrics$val_mean_absolute_error\n  all_mae_histories &lt;- rbind(all_mae_histories, mae_history)\n}\n\n\n  average_mae_history &lt;- data.frame(\n    epoch = seq(1:ncol(all_mae_histories)),\n    validation_mae = apply(all_mae_histories, 2, mean)\n  )\n\n\n  library(ggplot2)\nggplot(average_mae_history, aes(x = epoch, y = validation_mae)) + geom_line()\n\n<\/code><\/pre>\n\n<p>Code with dataset from mlbench (after the line with \"=====\", the code is the same as in the code above:<\/p>\n\n<pre class=\"lang-r prettyprint-override\"><code>\nlibrary(dplyr)\nlibrary(mlbench)\nlibrary(groupdata2)\n\ndata(BostonHousing)\n\nparts &lt;- partition(BostonHousing, p = 0.2)\ntest_data &lt;- parts[[1]]\ntrain_data &lt;- parts[[2]]\n\n\ntrain_targets &lt;- train_data$medv\ntest_targets &lt;- test_data$medv\n\ntrain_data$medv &lt;- NULL\ntest_data$medv &lt;- NULL\n\n\ntrain_data$chas &lt;- NULL\ntest_data$chas &lt;- NULL\n\nmean &lt;- apply(train_data, 2, mean)\nstd &lt;- apply(train_data, 2, sd)\ntrain_data &lt;- scale(train_data, center = mean, scale = std)\ntest_data &lt;- scale(test_data, center = mean, scale = std)\n\nlibrary(keras)\n\n# After this line the code is the same for both  code examples.\n# ========================================= \n\nbuild_model &lt;- function() {\n  model &lt;- keras_model_sequential() %&gt;% \n    layer_dense(units = 64, activation = \"relu\", \n                input_shape = dim(train_data)[[2]]) %&gt;% \n    layer_dense(units = 64, activation = \"relu\") %&gt;% \n    layer_dense(units = 1) \n\n  model %&gt;% compile(\n    optimizer = \"rmsprop\", \n    loss = \"mse\", \n    metrics = c(\"mae\")\n  )\n}\n\n  k &lt;- 4\nindices &lt;- sample(1:nrow(train_data))\nfolds &lt;- cut(1:length(indices), breaks = k, labels = FALSE) \nnum_epochs &lt;- 100\nall_scores &lt;- c()\nfor (i in 1:k) {\n  cat(\"processing fold #\", i, \"\\n\")\n  # Prepare the validation data: data from partition # k\n  val_indices &lt;- which(folds == i, arr.ind = TRUE) \n  val_data &lt;- train_data[val_indices,]\n  val_targets &lt;- train_targets[val_indices]\n\n  # Prepare the training data: data from all other partitions\n  partial_train_data &lt;- train_data[-val_indices,]\n  partial_train_targets &lt;- train_targets[-val_indices]\n\n  # Build the Keras model (already compiled)\n  model &lt;- build_model()\n\n  # Train the model (in silent mode, verbose=0)\n  model %&gt;% fit(partial_train_data, partial_train_targets,\n                epochs = num_epochs, batch_size = 1, verbose = 0)\n\n  # Evaluate the model on the validation data\n  results &lt;- model %&gt;% evaluate(val_data, val_targets, verbose = 0)\n  all_scores &lt;- c(all_scores, results$mean_absolute_error)\n}  \nall_scores\nmean(all_scores)\n\n  # Some memory clean-up\n  k_clear_session()\nnum_epochs &lt;- 500\nall_mae_histories &lt;- NULL\nfor (i in 1:k) {\n  cat(\"processing fold #\", i, \"\\n\")\n\n  # Prepare the validation data: data from partition # k\n  val_indices &lt;- which(folds == i, arr.ind = TRUE)\n  val_data &lt;- train_data[val_indices,]\n  val_targets &lt;- train_targets[val_indices]\n\n  # Prepare the training data: data from all other partitions\n  partial_train_data &lt;- train_data[-val_indices,]\n  partial_train_targets &lt;- train_targets[-val_indices]\n\n  # Build the Keras model (already compiled)\n  model &lt;- build_model()\n\n  # Train the model (in silent mode, verbose=0)\n  history &lt;- model %&gt;% fit(\n    partial_train_data, partial_train_targets,\n    validation_data = list(val_data, val_targets),\n    epochs = num_epochs, batch_size = 1, verbose = 1\n  )\n  mae_history &lt;- history$metrics$val_mean_absolute_error\n  all_mae_histories &lt;- rbind(all_mae_histories, mae_history)\n}\n\n\n  average_mae_history &lt;- data.frame(\n    epoch = seq(1:ncol(all_mae_histories)),\n    validation_mae = apply(all_mae_histories, 2, mean)\n  )\n\n\n  library(ggplot2)\nggplot(average_mae_history, aes(x = epoch, y = validation_mae)) + geom_line()\n\n<\/code><\/pre>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"58059273","CreationDate":"2019-09-23T09:19:48.203","ParentId":null,"OwnerUserId":"10411744","Title":null,"Body":"<p>writing here because I can't comment... \nI checked the mlbench dataset <a href=\"https:\/\/rdrr.io\/cran\/mlbench\/man\/BostonHousing.html\" rel=\"nofollow noreferrer\">here<\/a> and it said, that it contains the 14 columns of the original boston data set and 5 additional columns. Not sure if you might have a faulty dataset because you state that there are no differences in the column counts of the datasets.<\/p>\n\n<p>Another guess might be, that the second example graph is from a model which is stuck in a local minima. To get more comparable models, you might want to work with the same seeds to make sure that the inizialisations of the weights etc. are the same to get the same results.<\/p>\n\n<p>Hope this helps and feel free to ask.<\/p>\n"}]}
+{"QuestionId":58051463,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-22T16:57:41.720","AcceptedAnswerId":null,"OwnerUserId":5677706.0,"Title":"How to use TPU with custom_objects","Body":"<p>I would like to run a model on a TPU, but this model use heavily custom layers.<\/p>\n\n<p>Here is how the model looks like:<\/p>\n\n<pre><code>def rdn_attention_learning(hidden_units, vocab_size, mixt_attention, mixt_output, seq_len,timesteps):\n    \"\"\"\n    This neural network don't take an input state and systematically return a sequence\n    it is used to learn.\n    \"\"\"\n    y = Input(shape=(timesteps, INPUT_DIM), name='Target_Sequence_RDN')\n    x = Input(shape=(timesteps, INPUT_DIM), name=\"Input_Sequence_RDN\")\n    c = Input(shape=(seq_len,vocab_size), name='Character_constant')\n\n    cells = [    \n        AttentionGRUCell(hidden_units,prevblock_dim=INPUT_DIM,vocab_size=vocab_size,k_mixture=mixt_attention),  \n        ResGRUCell(hidden_units, prevblock_dim=hidden_units, resblock_dim=INPUT_DIM, residual_att=vocab_size),\n        ResGRUCell(hidden_units,prevblock_dim=hidden_units, resblock_dim=hidden_units, residual_att=vocab_size)\n        ]\n\n    output = RNN(cells, input_shape=(timesteps, INPUT_DIM), return_sequences=True)(x,constants=c)\n\n    # Predicting\n    sample, loss = GMM(k_mixture=mixt_output, output_dim=2)([output,y]) \n    model = Model(inputs = [x,y,c], outputs = sample)\n    adam = Adam()\n    model.compile(optimizer = adam, loss=log_loss(loss))\n    return model\n\nrdna = rdn_attention_learning(hidden_units, vocab_size, mixt_attention, mixt_output, seq_len,timesteps)\n<\/code><\/pre>\n\n<p>As you notice I use 3 custom layers in my model: AttentionGRUCell, ResGRUCell, GMM.\nThe model work, I can export it in .h5 mode with keras API.<\/p>\n\n<p>Here is the code I use to load the model in the TPU architecture:<\/p>\n\n<pre><code>TPU_WORKER = 'grpc:\/\/' + os.environ['COLAB_TPU_ADDR']\ntf.logging.set_verbosity(tf.logging.INFO)\n\ntpu_model = tf.contrib.tpu.keras_to_tpu_model(\n    rdna,\n    strategy=tf.contrib.tpu.TPUDistributionStrategy(\n        tf.contrib.cluster_resolver.TPUClusterResolver(TPU_WORKER)))\n<\/code><\/pre>\n\n<p>However I get an error<\/p>\n\n<pre><code>ValueError: Unknown layer: AttentionGRUCell\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58051740,"AnswerCount":1,"Tags":"<keras><tensorboard><tensorflow2.0><graph-visualization><resnet>","CreationDate":"2019-09-22T17:30:07.287","AcceptedAnswerId":null,"OwnerUserId":12103455.0,"Title":"How to make a Graph Visualization in Tensorflow v2 using Tensorboard?","Body":"<p>Actually I have two related problems. First of all is it possible to use Graph Visualization in Tfv2 ? Because I couldn't write a code to show my architecure so I didn't know my design is correct.<\/p>\n\n<p>Second problem is my network has got training and validation accuracies are just %50. I used ResNet architecture. Here is my ResNet block and have used it in main network structure for 6 times. My GPU is not best so I have to use filter size 16 or 32:<\/p>\n\n<pre><code>class ResnetBlock(tf.keras.Model):\n\ndef __init__(self, residual_path=False):\n    super(ResnetBlock, self).__init__()\n\n    self.residual_path = residual_path\n\n    self.conv2 = tf.keras.layers.Conv2D(16, kernel_size = (3,3), strides=(1,1), padding='same', use_bias=False, kernel_initializer=tf.keras.initializers.he_normal(seed=13))\n    self.bn2 = tf.keras.layers.BatchNormalization(epsilon=1e-3, momentum=0.993)\n    self.relu2 = tf.keras.layers.Activation('relu')\n\n    self.conv3 = tf.keras.layers.Conv2D(32, kernel_size = (3,3), strides=(1,1), padding='same', use_bias=False, kernel_initializer=tf.keras.initializers.he_normal(seed=13))\n    self.bn3 = tf.keras.layers.BatchNormalization(epsilon=1e-3, momentum=0.993)\n    self.relu3 = tf.keras.layers.Activation('relu')\n\n\n\ndef call(self, inputs, training=None):\n\n    x0 = inputs\n\n    x = self.conv2(inputs)\n    x = self.bn2(x, training=training)\n    x = self.relu2(x)\n\n    x = self.conv3(x)\n    x = self.bn3(x, training=training)\n    x = self.relu3(x)        \n\n\n\n    x = x + x0    \n\n    return x\n<\/code><\/pre>\n\n<p>Also here is the main network structure:<\/p>\n\n<p>class ResNet(tf.keras.Model):<\/p>\n\n<pre><code>def __init__(self, num_blocks ,num_classes, **kwargs):\n    super(ResNet, self).__init__(**kwargs)\n\n    self.num_blocks = num_blocks\n\n    self.conv1 = tf.keras.layers.Conv2D(32, kernel_size = (7,7), strides=(2,2), padding='same', use_bias=False, kernel_initializer=tf.keras.initializers.he_normal(seed=13))\n    self.bn1 = tf.keras.layers.BatchNormalization(epsilon=1e-3, momentum=0.993)\n    self.relu1 = tf.keras.layers.Activation('relu')\n    self.mps1 = tf.keras.layers.MaxPooling2D(pool_size=(2, 2))\n\n\n    self.blocks = tf.keras.models.Sequential(name='dynamic-blocks')\n\n\n    for block_number in range(num_blocks):\n        block = ResnetBlock(residual_path=True)        \n        self.blocks.add(block)\n\n\n    self.final_bn = tf.keras.layers.BatchNormalization()\n    self.avg_pool = tf.keras.layers.GlobalAveragePooling2D()\n    self.fc = tf.keras.layers.Dense(num_classes)\n\ndef call(self, inputs, training=None):\n\n    x = self.conv1(inputs)\n    x = self.bn1(x, training=training)\n    x = self.relu1(x) \n    x = self.mps1(x)\n\n    out = self.blocks(x, training=training)\n\n\n\n    out = self.final_bn(out, training=training)\n    out = tf.nn.relu(out)\n\n    out = self.avg_pool(out)\n    out = self.fc(out)\n\n\n    return out\n<\/code><\/pre>\n","answers":[{"AnswerId":"58057144","CreationDate":"2019-09-23T06:59:08.843","ParentId":null,"OwnerUserId":"4373898","Title":null,"Body":"<p>For your graph visualization, easiest way is to call:<\/p>\n\n<pre><code>net = ResNet(num_blocks, num_classes)\nnet.compile(...)\nnet.summary(...)\n<\/code><\/pre>\n\n<p>You can look at the <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/Model#summary\" rel=\"nofollow noreferrer\">documentation<\/a> for the arguments.<\/p>\n\n<p>You can also test the <a href=\"https:\/\/keras.io\/visualization\/\" rel=\"nofollow noreferrer\"><code>keras.utils.plot_model<\/code> function<\/a> (I don't know if it works well with TF\u00a0networks.<\/p>\n\n<p>Finally, you fit your model for 1 epoch of 1 step, adding a <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/callbacks\/TensorBoard\" rel=\"nofollow noreferrer\">TensorBoard callback<\/a>. Then, using <a href=\"https:\/\/www.tensorflow.org\/tensorboard\/r2\/get_started\" rel=\"nofollow noreferrer\">TensorBoard<\/a> you can get a very detailled graphical representation of your model.<\/p>\n\n<p>Kind,\nAlexis.<\/p>\n\n<p>P.S: I do not really understand your second problem. Sorry.<\/p>\n"}]}
+{"QuestionId":58051767,"AnswerCount":1,"Tags":"<python><machine-learning><keras><neural-network>","CreationDate":"2019-09-22T17:33:33.523","AcceptedAnswerId":58052503.0,"OwnerUserId":2607447.0,"Title":"Keras - how to predict two values instead of one value?","Body":"<p>I'm learning machine learning and my dataset consists of 7 columns:<\/p>\n\n<pre><code>home_team, away_team, home_odds, away_odds, home_score, away_score, 1_if_home_wins_else_0\n<\/code><\/pre>\n\n<p>To be able to feed the Tensorflow with teams, I converted every team to integer so the first two columns are integers (like database ids)<\/p>\n\n<p>There is a 10k rows in the csv.<\/p>\n\n<p><strong>example<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/jMUjp.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/jMUjp.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I modified the code for <a href=\"https:\/\/machinelearningmastery.com\/tutorial-first-neural-network-python-keras\/\" rel=\"nofollow noreferrer\">pima indians diabetes<\/a> to predict <strong>winnings<\/strong> of home team.<\/p>\n\n<p>So now it \"predicts\" whether the home team win (1) otherwise 0.<\/p>\n\n<p>Now I would like to modify the algorithm to predict exact score <code>home_score<\/code>, <code>away_score<\/code>. I know the outputs will be wrong it's just learning.<\/p>\n\n<p><strong>code<\/strong><\/p>\n\n<pre><code># load the dataset\ndataset = loadtxt('football_data.csv', delimiter=',')\n# split into input (X) and output (y) variables\nX = dataset[:, 0:4]\ny = dataset[:, 6]\n# define the keras model\nmodel = Sequential()\nmodel.add(Dense(12, input_dim=4, activation='relu'))\nmodel.add(Dense(8, activation='relu'))\nmodel.add(Dense(1, activation='sigmoid'))\n# compile the keras model\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n# fit the keras model on the dataset\nmodel.fit(X, y, epochs=150, batch_size=10)\n# evaluate the keras model\n_, accuracy = model.evaluate(X, y)\nprint('Accuracy: %.2f' % (accuracy * 100))\n\n# make class predictions with the model\npredictions = model.predict_classes(X)\n# summarize the first 5 cases\nfor i in range(50):\n    print('%s =&gt; %d (expected %d)' % (X[i].tolist(), predictions[i], y[i]))\n<\/code><\/pre>\n\n<p>Do you know how to do that?<\/p>\n","answers":[{"AnswerId":"58052503","CreationDate":"2019-09-22T18:58:57.830","ParentId":null,"OwnerUserId":"4685471","Title":null,"Body":"<p>Since you now want to predict scores, i.e. a continuous quantity (although integer), this is no more a classification problem but a <em>regression<\/em> one.<\/p>\n\n<p>To do that, you need two changes in your existing model; the first is to modify your final layer to<\/p>\n\n<pre><code>model.add(Dense(2)) # final layer\n<\/code><\/pre>\n\n<p>i.e. 2 nodes (as many as the required outputs) without any activation (which means the default <code>linear<\/code>, i.e. the one we use for regression).<\/p>\n\n<p>The second modification should be in the loss:<\/p>\n\n<pre><code>model.compile(loss='mean_squared_error', optimizer='adam')\n<\/code><\/pre>\n\n<p>You need of course to modify accordingly your <code>y<\/code> data in order to contain 2 elements each, and use <code>model.predict<\/code> instead of <code>model.predict_classes<\/code> (since you don't have classes now).<\/p>\n\n<p>What you ask for is essentially a multi-output regression; see also this recent thread: <a href=\"https:\/\/stackoverflow.com\/questions\/58037682\/how-to-train-a-regression-model-for-single-input-and-multiple-output\">How to train a Regression model for single input and multiple output?<\/a> - it may be better indeed to use the functional Keras API, as shown there and already suggested in the comments.<\/p>\n"}]}
+{"QuestionId":58052057,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-22T18:07:44.460","AcceptedAnswerId":null,"OwnerUserId":1118236.0,"Title":"How to initialize a Tensorflow Variable with positive constraints?","Body":"<p>I am creating a Variable in Tensorflow as follows:<\/p>\n\n<pre><code>tf.get_variable('opt_res',\n                initializer=tf.truncated_normal(shape=[self.batch_size], mean=0.0, stddev=1e-1),\n                validate_shape=False,\n                trainable=trainable)\n<\/code><\/pre>\n\n<p>However, in my application, the variable (i.e., price) should be always positive. Negative values is meaningless. How should I implement such constraint in Tensorflow?<\/p>\n","answers":[]}
+{"QuestionId":58052160,"AnswerCount":0,"Tags":"<keras><deep-learning>","CreationDate":"2019-09-22T18:18:35.263","AcceptedAnswerId":null,"OwnerUserId":9126289.0,"Title":"how to update game state by using pixels as input?","Body":"<p>I'm currently writing a program using keras (python 3) to play a game similar to Atari games, only in this one there are objects moving in the screen in different angles and directions (in most of Atari games I've encountered the objects you need to shoot are static). The agent's aim is to shoot them. My Input is the pixels on the screen (and those are the states). after I get the screenshot I convert it into a pixels matrix, only I'm not sure how to efficiently get a new state and how to process the state's info. if anyone have a recommendation for an article or a book which could guide me how to efficiently use my information from the screenshot, I would be very grateful. <\/p>\n\n<p>Thank you<\/p>\n","answers":[]}
+{"QuestionId":58052588,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2019-09-22T19:08:44.847","AcceptedAnswerId":null,"OwnerUserId":12102214.0,"Title":"LSTM time series prediction shifts prediction","Body":"<p>I am using lstm to make a time series forecast. At time t I use t+1 to predict the next step.<\/p>\n\n<p>For example:<\/p>\n\n<pre><code>112  -&gt;   118\n118  -&gt;   132\n132  -&gt;   129\n129  -&gt;   121\n121  -&gt;   135\n<\/code><\/pre>\n\n<p>When I train my network and do a prediction I get the following result:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/wsUNx.png\" rel=\"nofollow noreferrer\">graph<\/a><\/p>\n\n<p>As you can see the test result is shifted by one because of the t+1 shift, but I do not understand why is the prediction is shifted and not aligned with the test line. <\/p>\n\n<p>Thanks in advance :)<\/p>\n","answers":[]}
+{"QuestionId":58052689,"AnswerCount":0,"Tags":"<tensorflow><installation><windows-10><anaconda>","CreationDate":"2019-09-22T19:23:12.443","AcceptedAnswerId":null,"OwnerUserId":10916087.0,"Title":"How to install tensorflow-gpu on Anaconda?","Body":"<p>I have installed tensorflow-gpu on my windows 10 (I have also installed python 3.6.6) . Now, I have downloaded and installed Anaconda 3, but I don't know how I can use tensorflow on spyder. I'm sorry but I'm really confused.\nCould Anyone give me some advices?<\/p>\n","answers":[]}
+{"QuestionId":58052770,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-22T19:35:36.400","AcceptedAnswerId":null,"OwnerUserId":11293481.0,"Title":"How are tensors added in TensorFlow?","Body":"<p>I am trying to understand some basics about tensorflow by reading this tutorial here: <a href=\"https:\/\/www.guru99.com\/tensor-tensorflow.html\" rel=\"nofollow noreferrer\">https:\/\/www.guru99.com\/tensor-tensorflow.html<\/a>\nWhat I cannot understand is why when running these commands:<\/p>\n\n<pre><code># Add\ntensor_a = tf.constant([[1,2]], dtype = tf.int32)\ntensor_b = tf.constant([[3, 4]], dtype = tf.int32)\ntensor_add = tf.add(tensor_a, tensor_b)\nprint(tensor_add)\n<\/code><\/pre>\n\n<p>I get this result:<\/p>\n\n<pre><code>Tensor(\"Add:0\", shape=(1, 2), dtype=int32)\n<\/code><\/pre>\n\n<p>I did the calculation on paper and when adding these 2 vectors, I get something complete different(4,6), why is that? What is a \"tensor\" anyway?<\/p>\n","answers":[{"AnswerId":"58052888","CreationDate":"2019-09-22T19:52:00.300","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>A \"tensor\" in TensorFlow is a computational object. What you get with <code>tf.add<\/code> is a NODE that adds its inputs, <code>tensor_a<\/code> and <code>tensor_b<\/code> - which is what you're seeing with <code>Tensor(\"Add:0\")<\/code> (the <code>:0<\/code> is its form of 'id'). This node, however, does nothing until <em>executed<\/em> - it's just \"there\" (see below). To execute, run<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with tf.Session() as sess:       # 'with' ensures computing resources are \n    print(sess.run(tensor_add))  # properly handled, but isn't required\n<\/code><\/pre>\n\n<p>I suggest you check out some starter tutorials, as TF isn't exactly intuitive - e.g. <a href=\"https:\/\/easy-tensorflow.com\/tf-tutorials\/basics\/graph-and-session\" rel=\"nofollow noreferrer\">here<\/a>. Good luck<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/juZ5K.png\" width=\"550\"><\/p>\n"}]}
+{"QuestionId":58052772,"AnswerCount":0,"Tags":"<python><tensorflow><keras><resnet><retinanet>","CreationDate":"2019-09-22T19:36:04.907","AcceptedAnswerId":null,"OwnerUserId":7837600.0,"Title":"ValueError: Cannot use 'filtered_detections\/map\/while\/strided_slice_1' as input to 'filtered_detections\/map\/while\/ones\/packed'","Body":"<p>Training on custom dataset by using Keras retinanet (<a href=\"https:\/\/github.com\/fizyr\/keras-retinanet\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fizyr\/keras-retinanet<\/a>). while running the retinanet-train command , i got the following error<\/p>\n\n<pre><code> File \"c:\\users\\samjith.cp\\anaconda3\\lib\\site-packages\\tensorflow\\python\\ops\\control_flow_util.py\", line 350, in CheckInputFromValidContext\nraise ValueError(error_msg + \" See info log for more details.\")\nValueError: Cannot use 'filtered_detections\/map\/while\/strided_slice_1' as \ninput to 'filtered_detections\/map\/while\/ones\/packed' \nbecause 'filtered_detections\/map\/while\/strided_slice_1' is in a while loop. \nSee info log for more details.\n<\/code><\/pre>\n\n<p>how to solve this error?<\/p>\n","answers":[]}
+{"QuestionId":58052841,"AnswerCount":1,"Tags":"<python><tensorflow><eager-execution>","CreationDate":"2019-09-22T19:45:08.550","AcceptedAnswerId":null,"OwnerUserId":1273987.0,"Title":"How to use complex variables in TensorFlow eager mode?","Body":"<p>In non-eager mode I can run this without issues:<\/p>\n\n<pre><code>s = tf.complex(tf.Variable(1.0), tf.Variable(1.0))\ntrain_op = tf.train.AdamOptimizer(0.01).minimize(tf.abs(s))\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    for i in range(5):\n        _, s_ = sess.run([train_op, s])\n        print(s_)\n\n&gt;(1+1j)\n(0.99+0.99j)\n(0.98+0.98j)\n(0.9700001+0.9700001j)\n(0.9600001+0.9600001j)\n<\/code><\/pre>\n\n<p>But I cannot seem to find the equivalent expression in eager mode. I've tried the following, but TF complains:<\/p>\n\n<pre><code>tfe = tf.contrib.eager\ns = tf.complex(tfe.Variable(1.0), tfe.Variable(1.0))\ndef obj(s):\n    return tf.abs(s)\nwith tf.GradientTape() as tape:\n    loss = obj(s)\n    grads = tape.gradient(loss, [s])\n    optimizer.apply_gradients(zip(grads, [s]))\n<\/code><\/pre>\n\n<blockquote>\n  <p>The dtype of the source tensor must be floating (e.g. <code>tf.float32<\/code>) when calling GradientTape.gradient, got <code>tf.complex64<\/code><\/p>\n<\/blockquote>\n\n<p>and <\/p>\n\n<blockquote>\n  <p>No gradients provided for any variable: <code>['tf.Tensor((1+1j), shape=(), dtype=complex64)']<\/code><\/p>\n<\/blockquote>\n\n<p>How does one train complex variables in eager mode?<\/p>\n","answers":[{"AnswerId":"59092718","CreationDate":"2019-11-28T16:05:46.007","ParentId":null,"OwnerUserId":"1316607","Title":null,"Body":"<p>Using eager mode in Tensorflow 2, you can take the real and imaginary parts as the real variables:<\/p>\n\n<pre><code>r, i = tf.Variable(1.0), tf.Variable(1.0)\ndef obj(s):\n    return tf.abs(s)\nwith tf.GradientTape() as tape:\n    s = tf.complex(r, i)\n    loss = obj(s)\n    grads = tape.gradient(loss, [r, i])\n    optimizer.apply_gradients(zip(grads, [r, i]))\n<\/code><\/pre>\n"}]}
+{"QuestionId":58052869,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-lite>","CreationDate":"2019-09-22T19:49:04.440","AcceptedAnswerId":58178610.0,"OwnerUserId":12103897.0,"Title":"TF Lite object detection only returning 10 detections","Body":"<p>I am using a custom object detection model with TensorFlow Lite on native Android. At this point I'm only detecting 2 custom objects. I am using the TensorFlow Object Detection API, and I have a pipeline in place that produces optimized .tflite files.<\/p>\n\n<p>However, at inference time, the model only returns up to 10 individual detections. According to <a href=\"https:\/\/www.tensorflow.org\/lite\/models\/object_detection\/overview\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/lite\/models\/object_detection\/overview<\/a>, this is expected. The problem is that my images have a relatively large object density. I need to be able to detect up to 30 individual objects per image\/inference call.<\/p>\n\n<p>If I change NUM_DETECTIONS in the sample Android app from the TF repo from 10 to, say, 20, I get a runtime exception due to shape mismatch. How can I produce .tflite files capable of yielding more than 10 object detection instances?<\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"58178610","CreationDate":"2019-10-01T05:35:29.557","ParentId":null,"OwnerUserId":"11031154","Title":null,"Body":"<p>Unfortunately, since TFLite prefers static-shaped Input\/Outputs, you would need to re-export a TFLite SSD graph with the required number of outputs. Instructions are <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/running_on_mobile_tensorflowlite.md\" rel=\"nofollow noreferrer\">here<\/a>. While invoking <code>object_detection\/export_tflite_ssd_graph.py<\/code>, you would need to pass in the parameter <code>--max_detections=20<\/code>. Then, your change of <code>NUM_DETECTIONS<\/code> should work as expected.<\/p>\n"}]}
+{"QuestionId":58052967,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-22T20:01:51.637","AcceptedAnswerId":null,"OwnerUserId":6263317.0,"Title":"How to select elements of a tensor along a specific axis in TensorFlow","Body":"<p>I have a tensor <code>t<\/code> of shape <code>(2, 3, 4)<\/code><\/p>\n\n<pre><code>t = tf.random.normal((2, 3, 4))\n\n&lt;tf.Tensor: id=55, shape=(2, 3, 4), dtype=float32, numpy=\narray([[[-0.86664855, -0.32786712, -0.9517335 ,  0.989722  ],\n        [-0.25011402, -0.35941386, -1.0808105 ,  0.60205466],\n        [ 0.07523973, -0.6512919 ,  1.3695312 , -1.5043781 ]],\n\n       [[ 0.33990988, -0.17364176,  0.72955394, -0.7119293 ],\n        [ 0.4013214 ,  0.5653289 ,  1.4327284 ,  1.2687784 ],\n        [-1.1986154 ,  1.3783301 ,  1.714094  ,  0.49866664]]],\n      dtype=float32)&gt;\n<\/code><\/pre>\n\n<p>and a set of indices <code>idx<\/code> of size <code>(2, 3)<\/code> with values that index along the last dimension  of <code>t<\/code><\/p>\n\n<pre><code>idx = tf.convert_to_tensor(np.random.randint(4, size=(2, 3)))\n\n&lt;tf.Tensor: id=56, shape=(2, 3), dtype=int64, numpy=\narray([[2, 2, 3],\n       [0, 3, 1]])&gt;\n<\/code><\/pre>\n\n<p>How can I extract the elements of <code>t<\/code> along it's last dimension at the indices specified by <code>idx<\/code>? The result should be the following tensor of shape <code>(2, 3)<\/code>.<\/p>\n\n<pre><code>&lt;tf.Tensor: id=57, shape=(2, 3), dtype=int64, numpy=\narray([[-0.9517335, -1.0808105, -1.5043781],\n       [0.33990988,  1.2687784,  1.3783301]])&gt;\n<\/code><\/pre>\n\n<p>I have been trying and failing with regular indexing<\/p>\n\n<pre><code>t[:, :, idx]  # error\nt[..., idx]   # error\n<\/code><\/pre>\n\n<p>and <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/gather\" rel=\"nofollow noreferrer\"><code>tf.gather<\/code><\/a> \/ <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/gather_nd\" rel=\"nofollow noreferrer\"><code>tf.gather_nd<\/code><\/a><\/p>\n\n<pre><code>tf.gather(t, idx, axis=2)  # has shape (2, 3, 2, 3)\ntf.gather_nd(t, idx)       # has shape (2, )\n<\/code><\/pre>\n\n<p>neither of which seem to accomplish this.<\/p>\n","answers":[{"AnswerId":"58054884","CreationDate":"2019-09-23T02:10:24.207","ParentId":null,"OwnerUserId":"1500585","Title":null,"Body":"<p>Think again about what you are trying to achieve. What are the indices of the elements you are trying to extract for the first and second axes? From your example, it seems that you are thinking about flattening the first two dimension so <code>t<\/code> is (6,4) and  extracting the elements whose first dimension indices are 0:6 and the second dimension indices are given by <code>idx<\/code>.<\/p>\n\n<p>To achieve this, you have to actually specify the indices for all dimensions. We can start by reshaping <code>t<\/code> to be 2D:<\/p>\n\n<pre><code>t_2d=tf.reshape(t,[-1,tf.shape(t)[-1]])\n\n&lt;tf.Tensor: id=55, shape=(6, 4), dtype=float32, numpy=\narray([[-0.86664855, -0.32786712, -0.9517335 ,  0.989722  ],\n       [-0.25011402, -0.35941386, -1.0808105 ,  0.60205466],\n       [ 0.07523973, -0.6512919 ,  1.3695312 , -1.5043781 ],\n       [ 0.33990988, -0.17364176,  0.72955394, -0.7119293 ],\n       [ 0.4013214 ,  0.5653289 ,  1.4327284 ,  1.2687784 ],\n       [-1.1986154 ,  1.3783301 ,  1.714094  ,  0.49866664]],\n      dtype=float32)&gt;\n<\/code><\/pre>\n\n<p>Now, we'll specify the indices of the first axis:<\/p>\n\n<pre><code>idx_0=tf.reshape(tf.range(t_2d.shape[0]),idx.shape)\n\n&lt;tf.Tensor: id=62, shape=(2, 3), dtype=int32, numpy=\narray([[0, 1, 2],\n       [3, 4, 5]], dtype=int32)&gt;\n<\/code><\/pre>\n\n<p>Join the indices of the first and second axes as expected by <code>tf.gather_nd<\/code>:<\/p>\n\n<pre><code>indices=tf.stack([idx_0,idx],axis=-1)\n\n&lt;tf.Tensor: id=64, shape=(2, 3, 2), dtype=int32, numpy=\narray([[[0, 2],\n        [1, 2],\n        [2, 3]],\n\n       [[3, 0],\n        [4, 3],\n        [5, 1]]], dtype=int32)&gt;\n<\/code><\/pre>\n\n<p>And finally:<\/p>\n\n<pre><code>tf.gather_nd(t_2d,indices)\n\n&lt;tf.Tensor: id=66, shape=(2, 3), dtype=float32, numpy=\narray([[-0.9517335 , -1.0808105 , -1.5043781 ],\n       [ 0.33990988,  1.2687784 ,  1.3783301 ]], dtype=float32)&gt;\n<\/code><\/pre>\n"}]}
+{"QuestionId":58053119,"AnswerCount":0,"Tags":"<python><tensorflow><object-detection>","CreationDate":"2019-09-22T20:25:35.850","AcceptedAnswerId":null,"OwnerUserId":8050649.0,"Title":"Error: object_detection.utils ModuleNotFoundError","Body":"<p>I'm following this tutorial (<a href=\"https:\/\/towardsdatascience.com\/creating-your-own-object-detector-ad69dda69c85\" rel=\"nofollow noreferrer\">https:\/\/towardsdatascience.com\/creating-your-own-object-detector-ad69dda69c85<\/a>) to create my own object detector.<\/p>\n\n<p>On the topic 'Generating TFRecords for training', whe i try run this code:<\/p>\n\n<pre><code>python generate_tfrecord.py --csv_input=images\\train_labels.csv --image_dir=images\\train --output_path=train.record\n<\/code><\/pre>\n\n<p>I got this error:<\/p>\n\n<pre><code>[...]\nTraceback (most recent call last):\n  File \"generate_tfrecord.py\", line 20, in &lt;module&gt;\n    from object_detection.utils import dataset_util\nModuleNotFoundError: No module named 'object_detection'\n<\/code><\/pre>\n\n<p>My folder structure:<\/p>\n\n<pre><code>-generate_tfrecord.py\n-images\/\n--test_labels.csv\n--train_labels.csv\n--test\/\n---*files*\n--train\/\n---*files*\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58053158,"AnswerCount":0,"Tags":"<python><tensorflow><keras><tensorboard>","CreationDate":"2019-09-22T20:31:01.950","AcceptedAnswerId":null,"OwnerUserId":8855704.0,"Title":"Keras tensorboard throwing Value Error for gradients with Batch Normalization layer","Body":"<p>I am training a convolutional neural network with Keras on Tensorflow, and want to see the histograms of the gradients on Tensorboard. <\/p>\n\n<p>My callback looks like this:<\/p>\n\n<pre><code>tensorboard = TensorBoard(log_dir=log_dir, histogram_freq=1, write_grads=True)\n<\/code><\/pre>\n\n<p>And I am using it when I fit the model:<\/p>\n\n<pre><code>history = model.fit_generator(\n    gen_train, steps_per_epoch=ceil(num_samples_train\/batch_size),\n    epochs=epochs, validation_data=gen_val, validation_steps=ceil(num_samples_val\/batch_size), verbose=1,\n    callbacks=[tensorboard, checkpoint])\n<\/code><\/pre>\n\n<p>However, if my model has a Batch normalization layer:<\/p>\n\n<pre><code>model.add(BatchNormalization(axis=channels))\n<\/code><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><code>  File \"\\.venv\\lib\\site-packages\\keras\\optimizers.py\", line 91, in get_gradients\nraise 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>If my model only has convolutions and activations, I don't get the error. But my understanding is that the gradients should be plotted even if the network has batch normalization. Any ideas why this would not work? And if this is, in fact, the expected behaviour, is there a workaround for getting the gradients on Tensorboard?<\/p>\n","answers":[]}
+{"QuestionId":58053330,"AnswerCount":1,"Tags":"<python><tensorflow><anaconda>","CreationDate":"2019-09-22T20:54:06.490","AcceptedAnswerId":null,"OwnerUserId":2410461.0,"Title":"Anaconda3 Pythonw.exe Ordinal 242 could not be located in the dynamic link library libiomp5md.dll","Body":"<p>I initially actually tried to install Tensorflow and Keras for R Studio. \nI followed the instructions online. I installed Python 64bit first and then installed Anaconda. \nFirst, it worked for a while. Then I was trying to install Tensorflow GPU version. Somehow, it got this weird problem. <\/p>\n\n<p>If I tried to run code associated with TensorFlow or even trying to open Spyder IDE. This error message pops up.<\/p>\n\n<p>I have tried uninstalled both Python and Anaconda many times. No success. What did I do wrong? Please help. Thank you so much, guys.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/r1OuN.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/r1OuN.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"58058288","CreationDate":"2019-09-23T08:18:58.217","ParentId":null,"OwnerUserId":"2410461","Title":null,"Body":"<p>I figured this out by myself. <\/p>\n\n<p>I copied libiomp5md.dll from Anaconda to Windows\\system32 directory and it worked. I don't know why but it worked.<\/p>\n"}]}
+{"QuestionId":58053509,"AnswerCount":1,"Tags":"<python><keras><tf.keras>","CreationDate":"2019-09-22T21:24:56.217","AcceptedAnswerId":58160270.0,"OwnerUserId":12104062.0,"Title":"Why does my Keras model train after I load it, even though I have not actually supplied any new training data?","Body":"<p>I am trying to train and make predictions with an LSTM model using tf.keras. I have written code in two different files, LSTMTraining.py which trains the Keras Model (and save it to a file), and Predict.py, which is supposed to load in the Keras model and use it to make predictions. For some reason, when I load the model in Predict.py, it starts training, even though I have not used the model.fit() command in that file. Why is this happening?<\/p>\n\n<p>I have saved the model into multiple different file formats. For example, I've tried saving the model's architecture into a JSON file (using model_to_json()), and saving the weights seperately, then loading both of these files in seperately and then combining them. I've also tried saving them together into one file (using model.save()), and loading that in.<\/p>\n\n<p>Creating and Training Model in LSTMTraining.py (Note: the log_similarity_loss was just a custom loss function I created for the model):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Machine learning\nimport tensorflow as tf\nfrom tensorflow.python.keras import layers\nimport numpy as np\n\n# Load\/save data\nimport pickle\nimport os\n\n# Shuffling\nfrom sklearn.utils import shuffle\n\n# Parameters\nepochs = 5\ndisplay_step = 1000\nn_input = 5\nwordvec_len = 5\nn_hidden = 512\nrecurrent_dropout = 0\ndropout = 0\n\n# Load data\nwith open(\"Vectorized_Word_By_Word.txt\", \"rb\") as data:\n    vectorized_txt = pickle.load(data)\n\n# Prepare data into format for training (x: [prev-words], y: [next-word])\nx_train, y_train = [], []\nfor n in range(0, len(vectorized_txt) - n_input - 1):\n    prev_words = vectorized_txt[n: n+5]\n    next_word = vectorized_txt[n+6]\n    x_train.append(prev_words)\n    y_train.append(next_word)\nx_train, y_train = np.array(x_train), np.array(y_train)\nx_train, y_train = shuffle(x_train, y_train, random_state=0)\n\n\ndef log_similarity_loss(y_actual, y_pred):\n    \"\"\"Log similarity loss calculation.\"\"\"\n    cos_similarity = tf.keras.losses.CosineSimilarity(axis=0)(y_actual, y_pred)\n    scaled_similarity = tf.add(tf.multiply(0.5, cos_similarity), 0.5)\n    return -0.5*tf.math.log(scaled_similarity)\n\n\nlog_similarity_loss(\n    [0.05, 0.01, 0.05, 1.2], [0.05, -0.01, 0.05, -1.2])\n\nmodel = tf.keras.Sequential([\n    layers.LSTM(n_hidden, input_shape=(n_input, wordvec_len),\n                dropout=dropout, recurrent_dropout=recurrent_dropout,\n                return_sequences=True),\n    layers.LSTM(n_hidden, dropout=dropout,\n                recurrent_dropout=recurrent_dropout),\n    layers.Dense(wordvec_len)\n])\n\nmodel.compile(loss=log_similarity_loss,\n              optimizer='adam', metrics=['cosine_proximity'])\n\nmodel.fit(x_train, y_train, epochs=epochs, batch_size=12)\n\nmodel.save(\"Keras_Model.h5\", include_optimizer=True, save_format='h5')\n\n# Save model weights and architecture\nmodel.save_weights('model_weights.h5')\nwith open(\"model_architecture.json\", \"w\") as json_file:\n    json_file.write(model.to_json())\n<\/code><\/pre>\n\n<p>Loading in the model in Predict.py (Note: All the functions imported from \"WordModel.py\" are just functions for text processing I've written that are unrelated to Keras):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from WordModel import word_by_word, word_to_vec, vec_to_word\nimport gensim\n\nimport tensorflow as tf\nfrom tensorflow.python.keras.models import load_model, model_from_json\n\nwith open('model_architecture.json', 'r') as json_file:\n    model_json = json_file.read()\n\nkeras_model = model_from_json(model_json)\nkeras_model.load_weights(\"model_weights.h5\")\n<\/code><\/pre>\n\n<p>I was expecting no output, just the model to be loaded. However, I got the verbose training output of the model like so (when running Predict.py):<\/p>\n\n<pre><code>  12\/1212 [..............................] - ETA: 3:32 - loss: 0.2656 - cosine_proximity: 0.0420\n  24\/1212 [..............................] - ETA: 1:55 - loss: 0.2712 - cosine_proximity: 0.2066\n  36\/1212 [..............................] - ETA: 1:24 - loss: 0.2703 - cosine_proximity: 0.2294\n  48\/1212 [&gt;.............................] - ETA: 1:08 - loss: 0.2394 - cosine_proximity: 0.2690\n  60\/1212 [&gt;.............................] - ETA: 58s - loss: 0.2286 - cosine_proximity: 0.2874 \n  72\/1212 [&gt;.............................] - ETA: 52s - loss: 0.2247 - cosine_proximity: 0.2750\n  84\/1212 [=&gt;............................] - ETA: 47s - loss: 0.2115 - cosine_proximity: 0.2924 \n<\/code><\/pre>\n\n<p>and so on.<\/p>\n\n<p>Note that I have not made any training command in my Predict.py file. I have rerun the code multiple times, and made sure that I was running the correct file. Still, nothing seems to be working.<\/p>\n\n<p>Thanks for the help!<\/p>\n","answers":[{"AnswerId":"58160270","CreationDate":"2019-09-30T00:35:56.163","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>The problem's likely with your VSCode IDE, which takes additional configuring to work both with Python and its packages -- when you run one script, you may be running <em>all<\/em> the scripts, thus the seen behavior. A solution I'd recommend is switching to <a href=\"https:\/\/www.spyder-ide.org\/\" rel=\"nofollow noreferrer\">Spyder<\/a> and installing your packages with <a href=\"https:\/\/anaconda.org\/\" rel=\"nofollow noreferrer\">Anaconda<\/a>. Once you've installed both, search \"anaconda command prompt\" or \"anaconda powershell\" on your PC, and in the terminal, type:<\/p>\n\n<pre><code>conda update conda\nconda update --all\nconda install numpy # optional (sort of)\nconda install matplotlib # optional (sort of)\n# SEE BELOW\nconda install -c conda-forge keras\nconda update --all # final 'cleanup' command - ensures package compatibility\n<\/code><\/pre>\n\n<p>If you plan on using a GPU (highly recommended), you'll need to first download CUDA - <a href=\"https:\/\/www.codingforentrepreneurs.com\/blog\/install-tensorflow-gpu-windows-cuda-cudnn\/\" rel=\"nofollow noreferrer\">instructions here<\/a> (get CUDA 10 instead of 9 in the article). Then run <code>conda install tensorflow-gpu<\/code> as in the article.<\/p>\n\n<p>Then, in Spyder: <code>Tools -&gt; Preferences -&gt; PYTHONPATH manager<\/code> -> add all folders of the modules\/data you plan to use, so you don't have to <code>%cd<\/code> each time or worry about relative pathing and can import directly. Lastly, make sure Anaconda &amp; Spyder use the right <a href=\"https:\/\/stackoverflow.com\/questions\/43592879\/how-to-change-python-version-in-anaconda-spyder\">Python interpreter<\/a>. <\/p>\n\n<p>Restart Spyder, run scripts - assuming no bugs, all should be well.<\/p>\n"}]}
+{"QuestionId":58053654,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-22T21:49:13.363","AcceptedAnswerId":58053677.0,"OwnerUserId":10112379.0,"Title":"Dot product consuming lot of memory","Body":"<p>I have two 2D tensors A and B of size [I,J] and I want to dot product each jth vector of A  with jth vector of B. It is consuming a lot of memory.\nI'm doing it like this:<\/p>\n\n<pre><code>for j in range (J):\n      K.dot(K.reshape(A[:,j], [1,-1]),K.reshape(B[:,j], [-1,1]))\n\n<\/code><\/pre>\n\n<p>Is there a better way to do that?<\/p>\n","answers":[{"AnswerId":"58053677","CreationDate":"2019-09-22T21:54:15.633","ParentId":null,"OwnerUserId":"3767239","Title":null,"Body":"<p>You can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/einsum\" rel=\"nofollow noreferrer\"><code>tf.einsum<\/code><\/a>:<\/p>\n\n<pre><code>tf.einsum('ij,ij-&gt;j', A, B)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58054035,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-22T23:12:55.643","AcceptedAnswerId":58054064.0,"OwnerUserId":4830432.0,"Title":"How does tf.losses.absolute_difference() work?","Body":"<p>I am working with some simple vectors that I want to do some math with. I want to compute the sum of differences between the vectors element-wise.<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>a = [1,2,3]\nb = [4,5,6]\n<\/code><\/pre>\n\n<p>evaluates to:<\/p>\n\n<pre><code>abs(a[0] - b[0]) + abs(a[1] - b[1]) + abs(a[2] - b[2]) = 3 + 3 + 3 = 9\n<\/code><\/pre>\n\n<p>When I went looking for a way to do this with Tensorflow if found the tf.losses.absolute_difference() function. However, when I started experimenting with it I get results that I don't quite understand.<\/p>\n\n<pre><code>d = [1,1]\ne = [2,3]\n\nwith tf.Session() as sess:\n\n  sess.run(init)\n\n  t = tf.losses.absolute_difference(e, d)\n\n  print(t.eval())\n<\/code><\/pre>\n\n<p>Here t evaluates to 1.5. I would have expected 3.<\/p>\n\n<pre><code>d = [1,1]\ne = [2,2]\n<\/code><\/pre>\n\n<p>Here t evaluates to 1.0. I would have expected 2.<\/p>\n\n<p>What function is tf.losses.absolute_difference() actually computing here?<\/p>\n","answers":[{"AnswerId":"58054064","CreationDate":"2019-09-22T23:21:13.800","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>From the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.14\/tensorflow\/python\/ops\/losses\/losses_impl.py#L206-L256\" rel=\"nofollow noreferrer\">source code<\/a>, <code>tf.losses.absolute_difference<\/code> computes the <em>weighted<\/em> absolute difference of its inputs - and by default, <code>weights=1.0 #broadcasted<\/code> (which computes the <em>average<\/em> difference):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>absolute_difference(a,b) = weights .* abs(a - b) \/ sum(weights) # .* = dot product\n                         = [1, 1] .* abs([2, 3] - [1, 1]) \/ (1 + 1) # lists for simplicity\n                         = [1, 1] .* abs([1, 2]) \/ 2\n                         = (1 + 2) \/ 2\n                         = 1.5\n<\/code><\/pre>\n"}]}
+{"QuestionId":58054115,"AnswerCount":1,"Tags":"<python><machine-learning><keras>","CreationDate":"2019-09-22T23:34:08.710","AcceptedAnswerId":58054467.0,"OwnerUserId":11896108.0,"Title":"Keras can't use callback to store checkpoint","Body":"<pre><code>train_data = ImageDataGenerator(rescale=1.\/255)\ntest_data = ImageDataGenerator(rescale=1.\/255)\ntrain_set = train_data.flow_from_directory('train_set',\n                                           target_size=(28,28),\n                                           batch_size=16,\n                                           class_mode='categorical')\ntest_set = train_data.flow_from_directory('test_set',\n                                           target_size=(28,28),\n                                           batch_size = 16,\n                                           class_mode='categorical')\ncheckpointer = ModelCheckpoint(filepath=\"best_weight.hdf5\",\n                               monitor='val_acc',\n                               verbose=1,\n                               save_best_only=True)\nhistory = classifier.fit_generator(train_set,\n                                   steps_per_epoch=210,\n                                   epochs=5,\n                                   callbacks=[checkpointer],\n                                   validation_data=test_set,\n                                   validation_steps=90)\n\nclassifier.load_weights('best_weights.hdf5')\n<\/code><\/pre>\n\n<p>When I tried to load the best weights, it shows no such file or directory. How should I make a change? Thank you guys so much!<\/p>\n","answers":[{"AnswerId":"58054467","CreationDate":"2019-09-23T00:43:50.933","ParentId":null,"OwnerUserId":"4685471","Title":null,"Body":"<p>You have a typo in the filename; you save at<\/p>\n\n<pre><code>filepath=\"best_weight.hdf5\"\n<\/code><\/pre>\n\n<p>while you try to load <code>'best_weights.hdf5'<\/code>.<\/p>\n\n<p>Remove the final <code>s<\/code> from the file name, i.e.:<\/p>\n\n<pre><code>classifier.load_weights('best_weight.hdf5')\n<\/code><\/pre>\n"}]}
+{"QuestionId":58054156,"AnswerCount":0,"Tags":"<opencv><keras><keras-layer><tf.keras>","CreationDate":"2019-09-22T23:41:26.963","AcceptedAnswerId":null,"OwnerUserId":282855.0,"Title":"How to add a Keras Layer to crop images?","Body":"<p>I need to add a layer to my <code>Sequential<\/code> model (actually will the first layer) in order to crop images to remove parts except than faces. For face recognition, I use <code>OpenCV<\/code>. Therefore, just like the <code>Transformations<\/code> in <code>PyTorch<\/code>, I need a layer that applies this transformation to the images and passes the output to the next layers.<\/p>\n\n<p>p.s. Using the latest stable releases of both of <code>Tensorflow<\/code> and <code>Keras<\/code> which are <code>1.14.0<\/code>, and <code>2.3.0<\/code>, respectively.<\/p>\n","answers":[]}
+{"QuestionId":58054247,"AnswerCount":1,"Tags":"<neural-network><deep-learning><pytorch><tensorboard>","CreationDate":"2019-09-23T00:00:19.187","AcceptedAnswerId":58059885.0,"OwnerUserId":5511906.0,"Title":"How to use tensorboard debugger with pytorch?","Body":"<p>I know pytorch starts to support tensorboard since the version 1.11. But I am wondering is it possible for us to use the debugger plugin for tensorboard with pytorch? I didn't find any information about this. If pytorch can also support tensorboard debugger, it would be extremely convenient and could save us a lot of time.<\/p>\n","answers":[{"AnswerId":"58059885","CreationDate":"2019-09-23T09:58:17.813","ParentId":null,"OwnerUserId":"5652313","Title":null,"Body":"<p>No, it is not and to doubt it ever will be possible. The plugin heavily relies on how TensorFlow internally represents graph nodes which is quite different in PyTorch.<\/p>\n"}]}
+{"QuestionId":58054251,"AnswerCount":0,"Tags":"<tensorflow><distributed-computing>","CreationDate":"2019-09-23T00:01:13.760","AcceptedAnswerId":null,"OwnerUserId":6450489.0,"Title":"Which way to proceed for distributed learning under tensorflow?","Body":"<p>I see a bunch of different methods in many web sites and Tensorflow guides to proceed for implementing a distributed learning example but can not clear my mind on them. Let me be simple and clear: I have model created under keras, have setup the TF_CONFIG to have a chief, a worker, a ps, and an evaluator server, and I assume that the below code should do the job:<\/p>\n\n<pre><code>runConfig = tf.estimator.RunConfig(\n        session_config=config,\n        model_dir=log_dir,\n        save_summary_steps=1,\n        save_checkpoints_steps=train_steps\n        )\nestimator = tf.keras.estimator.model_to_estimator(model, model_dir=log_dir, config=runConfig)\ntrain_spec = tf.estimator.TrainSpec(input_fn=lambda: read_datasets(...), max_steps=epochs*train_steps)\neval_spec = tf.estimator.EvalSpec(input_fn=lambda: read_datasets(...), steps=None)\ntf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)\n<\/code><\/pre>\n\n<p>And then I can see the performance graphs and values on Tensorboard. I assume that the code for all servers are exactly the same except the TF_CONFIG part, and just one copy of training data is needed to be placed in a shared folder, and one shared directory is set to save model parameters and checkpoints, and the chief is going to synchronize everything on splitting the data between workers and saving the model and summaries.<\/p>\n\n<p>There is something here that I am not clear about it: I want to evaluate the trained model on a test data, and I assume that I can just execute the below command to get the evaluated loss and accuracy in a dictionary object:<\/p>\n\n<pre><code>eval_dic = estimator.evaluate(input_fn=lambda: read_datasets(...), steps=test_steps)\n<\/code><\/pre>\n\n<p>But this code is run by all services (worker, chief, and evaluator), which I don't think should be. Also the result that I get is just null.<\/p>\n\n<p>In addition, I see pages about distribution strategy and its implications on keras and estimators, which make me so much confused. Do I really need all of these strategy stuff, or just above code is sufficient? What about the input_fn? I assumed that the chief is taking care of data split but in the tensorflow documentation (<a href=\"https:\/\/www.tensorflow.org\/guide\/distribute_strategy\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/distribute_strategy<\/a>) it says different things for keras and estimator implementation, and my application is a keras model encapsulated in an estimator! What should I do?<\/p>\n","answers":[]}
+{"QuestionId":58054328,"AnswerCount":1,"Tags":"<python><neural-network><conv-neural-network><pytorch><autoencoder>","CreationDate":"2019-09-23T00:14:14.127","AcceptedAnswerId":58055060.0,"OwnerUserId":2703038.0,"Title":"Extracting reduced dimension data from autoencoder in pytorch","Body":"<p>I have defined my autoencoder in pytorch as following:<\/p>\n\n<pre><code>self.encoder = nn.Sequential(\n    nn.Conv2d(input_shape[0], 32, kernel_size=1, stride=1),\n    nn.ReLU(),\n    nn.Conv2d(32, 64, kernel_size=1, stride=1),\n    nn.ReLU(),\n    nn.Conv2d(64, 64, kernel_size=1, stride=1),\n    nn.ReLU()\n)\n\nself.decoder = nn.Sequential(\n    nn.Conv2d(64, 64, kernel_size=1, stride=1),\n    nn.ReLU(),\n    nn.Conv2d(64, 32, kernel_size=1, stride=1),\n    nn.ReLU(),\n    nn.Conv2d(32, input_shape[0], kernel_size=1, stride=1),\n    nn.ReLU(),\n    nn.Sigmoid()\n)\n<\/code><\/pre>\n\n<p>I need to get a reduced dimension encoding which requires creating a new linear layer of the dimension N much lower than the image dimension so that I can extract the activations.<\/p>\n\n<p>If anybody can help me with fitting a linear layer in the decoder part I would appreciate (i know how to Flatten() the data, but I guess I need to \"unflatten\" it again to interface with the Conv2d layer again)<\/p>\n\n<p>Update:<\/p>\n\n<p>I have come up with the following based on the first answer (it gives me a 8-dimensional bottleneck at the output of the encoder which works fine torch.Size([1, 8, 1, 1]) ).<\/p>\n\n<pre><code>self.encoder = nn.Sequential(\n    nn.Conv2d(input_shape[0], 32, kernel_size=8, stride=4),\n    nn.ReLU(),\n    nn.Conv2d(32, 64, kernel_size=4, stride=2),\n    nn.ReLU(),\n    nn.Conv2d(64, 8, kernel_size=3, stride=1),\n    nn.ReLU(),\n    nn.MaxPool2d(7, stride=1)\n)\n\nself.decoder = nn.Sequential(\n    nn.ConvTranspose2d(8, 64, kernel_size=3, stride=1),\n    nn.ReLU(),\n    nn.Conv2d(64, 32, kernel_size=4, stride=2),\n    nn.ReLU(),\n    nn.Conv2d(32, input_shape[0], kernel_size=8, stride=4),\n    nn.ReLU(),\n    nn.Sigmoid()\n)\n<\/code><\/pre>\n\n<p>What I cannot do is train the autoencoder with <\/p>\n\n<pre><code>def forward(self, x):\n    x = self.encoder(x)\n    x = self.decoder(x)\n    return x\n<\/code><\/pre>\n\n<p>The decoder gives me an error:<\/p>\n\n<pre><code>Calculated padded input size per channel: (3 x 3). Kernel size: (4 x 4). Kernel size can't be greater than actual input size\n<\/code><\/pre>\n\n<p>I would like to thank the person who provided the first answer.<\/p>\n","answers":[{"AnswerId":"58055060","CreationDate":"2019-09-23T02:47:04.250","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>In the decoder part, you need to upsample to larger size, which can be done via <code>nn.ConvTranspose2d<\/code>. I notice that in your encoder part, seems like you didn't downsample your feature maps, because your stride is alwasy 1. Here is a toy example. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>self.encoder = nn.Sequential(\n            nn.Conv2d(32, 16, 3, stride=1, padding=1),  # b, 16, 32, 32\n            nn.ReLU(True),\n            nn.MaxPool2d(2, stride=2),  # b, 16, 16, 16\n            nn.Conv2d(16, 32, 3, stride=1, padding=1),  # b, 32, 16, 16\n            nn.ReLU(True),\n            nn.MaxPool2d(2, stride=2)  # b, 32, 8, 8\n        )\nself.decoder = nn.Sequential(\n            nn.ConvTranspose2d(32, 16, 3, stride=2,padding=1,output_padding=1),  # b, 16, 16, 16\n            nn.ReLU(True),\n            nn.ConvTranspose2d(16, 1, 3, stride=2, padding=1, output_padding=1),  # b, 1, 32, 32\n            nn.Sigmoid()\n        )\n<\/code><\/pre>\n"}]}
+{"QuestionId":58054933,"AnswerCount":0,"Tags":"<tensorflow><keras><tensorflow-estimator>","CreationDate":"2019-09-23T02:19:54.130","AcceptedAnswerId":null,"OwnerUserId":10867166.0,"Title":"How to get the tensor name from a estimator model of keras which can be recognized by LoggingTensorHook?","Body":"<p>I've build a keras model and converted it to an estimator model. In the training process, it prints only loss value. I also want to print metrics value like <code>accuracy<\/code>, so I want to add a <code>LoggingTensorHook<\/code> to train_spec. But the <code>LoggingTensorHook<\/code> takes tensor name as a parameter. I've tried several names and also checked the name from tensorboard. But it doesn't work, error is shown below. What shold I do to get the correct tensor name to make it work? <\/p>\n\n<p><strong>Error raised<\/strong><\/p>\n\n<pre><code>KeyError: \"The name 'dense_1_loss:0' refers to a Tensor which does not exist. The operation, 'dense_1_loss', does not exist in the graph.\"\n<\/code><\/pre>\n\n<p><strong>Here is the example code I've tried:<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from __future__ import absolute_import\nfrom __future__ import division\nfrom __future__ import print_function\n\nimport sys\n\nimport numpy as np\nimport tensorflow as tf\n\n\ndef input_fn():\n    x = np.random.random((1024, 10))\n    y = np.random.randint(2, size=(1024, 1))\n    x = tf.cast(x, tf.float32)\n    dataset = tf.data.Dataset.from_tensor_slices((x, y))\n    dataset = dataset.repeat(100)\n    dataset = dataset.batch(32)\n    return dataset\n\n\ndef main(args):\n    if len(args) &lt; 2:\n        print('You must specify model_dir for checkpoints such as'\n              ' \/tmp\/tfkeras_example\/.')\n        return\n\n    model_dir = args[1]\n    print('Using %s to store checkpoints.' % model_dir)\n\n    # Define a Keras Model.\n    model = tf.keras.Sequential()\n    model.add(tf.keras.layers.Dense(\n        16, activation='relu', input_shape=(10,)))\n    model.add(tf.keras.layers.Dense(1, activation='sigmoid'))\n\n    # Compile the model.\n    optimizer = tf.train.GradientDescentOptimizer(0.2)\n    model.compile(loss='binary_crossentropy',\n                  optimizer=optimizer, metrics=['accuracy'])\n\n    model.summary()\n    tf.keras.backend.set_learning_phase(True)\n\n    # Define DistributionStrategies and convert the Keras Model to an\n    # Estimator that utilizes these DistributionStrateges.\n    # Evaluator is a single worker, so using MirroredStrategy.\n    config = tf.estimator.RunConfig(\n        experimental_distribute=tf.contrib.distribute.DistributeConfig(\n            train_distribute=tf.contrib.distribute.CollectiveAllReduceStrategy(),\n            eval_distribute=tf.contrib.distribute.MirroredStrategy()))\n    keras_estimator = tf.keras.estimator.model_to_estimator(\n        keras_model=model, config=config, model_dir=model_dir)\n\n    # Train and evaluate the model. Evaluation will be skipped if there is not an\n    # \"evaluator\" job in the cluster.\n\n    logging_hook = tf.estimator.LoggingTensorHook(\n        ['dense_1_loss'], every_n_iter=10)\n\n    tf.estimator.train_and_evaluate(\n        keras_estimator,\n        train_spec=tf.estimator.TrainSpec(\n            input_fn=input_fn, hooks=[logging_hook]),\n        eval_spec=tf.estimator.EvalSpec(input_fn=input_fn))\n\n\nif __name__ == '__main__':\n    tf.logging.set_verbosity(tf.logging.INFO)\n    tf.app.run(argv=sys.argv)\n<\/code><\/pre>\n\n<p><strong>Here is a subgraph of the whole model<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/92ods.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/92ods.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><strong>Add metric to estimator<\/strong><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def my_auc(labels, predictions):\n    auc_metric = tf.keras.metrics.AUC(name=\"my_auc\")\n    auc_metric.update_state(\n        y_true=labels, y_pred=predictions['classes'])\n    return {'auc': auc_metric}\n# add metric to estimator\ntf.estimator.add_metrics(keras_estimator, my_auc)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58055095,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><nlp><recurrent-neural-network><text-classification>","CreationDate":"2019-09-23T02:53:15.263","AcceptedAnswerId":58070688.0,"OwnerUserId":5581893.0,"Title":"How to change this RNN text classification code to text generation?","Body":"<p>I have this code to do text classification with TensorFlow RNN, but how to change it to do text generation instead?<\/p>\n\n<p>The following text classification has 3D input, but 2D output. Should it be changed to 3D input and 3D output for text generation? and how?<\/p>\n\n<p>The example data are:<\/p>\n\n<pre><code>t0      t1      t2\nbritish gray    is =&gt; cat (y=0)\n0       1       2\nwhite   samoyed is =&gt; dog (y=1)\n3       4       2 \n<\/code><\/pre>\n\n<p>For classification feeding \"british gray is\" results in \"cat\". What I wish to get is feeding \"british\" should result in the next word \"gray\".<\/p>\n\n<pre><code>import tensorflow as tf;\ntf.reset_default_graph();\n\n#data\n'''\nt0      t1      t2\nbritish gray    is =&gt; cat (y=0)\n0       1       2\nwhite   samoyed is =&gt; dog (y=1)\n3       4       2 \n'''\nBsize = 2;\nTimes = 3;\nMax_X = 4;\nMax_Y = 1;\n\nX = [[[0],[1],[2]], [[3],[4],[2]]];\nY = [[0],           [1]          ];\n\n#normalise\nfor I in range(len(X)):\n  for J in range(len(X[I])):\n    X[I][J][0] \/= Max_X;\n\nfor I in range(len(Y)):\n  Y[I][0] \/= Max_Y;\n\n#model\nInputs   = tf.placeholder(tf.float32, [Bsize,Times,1]);\nExpected = tf.placeholder(tf.float32, [Bsize,      1]);\n\n#single LSTM layer\n#'''\nLayer1   = tf.keras.layers.LSTM(20);\nHidden1  = Layer1(Inputs);\n#'''\n\n#multi LSTM layers\n'''\nLayers = tf.keras.layers.RNN([\n  tf.keras.layers.LSTMCell(30), #hidden 1\n  tf.keras.layers.LSTMCell(20)  #hidden 2\n]);\nHidden2 = Layers(Inputs);\n'''\n\nWeight3  = tf.Variable(tf.random_uniform([20,1], -1,1));\nBias3    = tf.Variable(tf.random_uniform([   1], -1,1));\nOutput   = tf.sigmoid(tf.matmul(Hidden1,Weight3) + Bias3);\n\nLoss     = tf.reduce_sum(tf.square(Expected-Output));\nOptim    = tf.train.GradientDescentOptimizer(1e-1);\nTraining = Optim.minimize(Loss);\n\n#train\nSess = tf.Session();\nInit = tf.global_variables_initializer();\nSess.run(Init);\n\nFeed = {Inputs:X, Expected:Y};\nfor I in range(1000): #number of feeds, 1 feed = 1 batch\n  if I%100==0: \n    Lossvalue = Sess.run(Loss,Feed);\n    print(\"Loss:\",Lossvalue);\n  #end if\n\n  Sess.run(Training,Feed);\n#end for\n\nLastloss = Sess.run(Loss,Feed);\nprint(\"Loss:\",Lastloss,\"(Last)\");\n\n#eval\nResults = Sess.run(Output,Feed);\nprint(\"\\nEval:\");\nprint(Results);\n\nprint(\"\\nDone.\");\n#eof\n<\/code><\/pre>\n","answers":[{"AnswerId":"58070688","CreationDate":"2019-09-23T21:58:16.133","ParentId":null,"OwnerUserId":"5581893","Title":null,"Body":"<p>I found out how to switch it (the code) to do text generation task, use 3D input (X) and 3D labels (Y) as in the source code below:<\/p>\n\n<p>Source code:<\/p>\n\n<pre><code>import tensorflow as tf;\ntf.reset_default_graph();\n\n#data\n'''\nt0       t1       t2\nbritish  gray     is  cat\n0        1        2   (3)  &lt;=x\n1        2        3        &lt;=y\nwhite    samoyed  is  dog\n4        5        2   (6)  &lt;=x\n5        2        6        &lt;=y \n'''\nBsize = 2;\nTimes = 3;\nMax_X = 5;\nMax_Y = 6;\n\nX = [[[0],[1],[2]], [[4],[5],[2]]];\nY = [[[1],[2],[3]], [[5],[2],[6]]];\n\n#normalise\nfor I in range(len(X)):\n  for J in range(len(X[I])):\n    X[I][J][0] \/= Max_X;\n\nfor I in range(len(Y)):\n  for J in range(len(Y[I])):\n    Y[I][J][0] \/= Max_Y;\n\n#model\nInput    = tf.placeholder(tf.float32, [Bsize,Times,1]);\nExpected = tf.placeholder(tf.float32, [Bsize,Times,1]);\n\n#single LSTM layer\n'''\nLayer1   = tf.keras.layers.LSTM(20);\nHidden1  = Layer1(Input);\n'''\n\n#multi LSTM layers\n#'''\nLayers = tf.keras.layers.RNN([\n  tf.keras.layers.LSTMCell(30), #hidden 1\n  tf.keras.layers.LSTMCell(20)  #hidden 2\n],\nreturn_sequences=True);\nHidden2 = Layers(Input);\n#'''\n\nWeight3  = tf.Variable(tf.random_uniform([20,1], -1,1));\nBias3    = tf.Variable(tf.random_uniform([   1], -1,1));\nOutput   = tf.sigmoid(tf.matmul(Hidden2,Weight3) + Bias3); #sequence of 2d * 2d\n\nLoss     = tf.reduce_sum(tf.square(Expected-Output));\nOptim    = tf.train.GradientDescentOptimizer(1e-1);\nTraining = Optim.minimize(Loss);\n\n#train\nSess = tf.Session();\nInit = tf.global_variables_initializer();\nSess.run(Init);\n\nFeed   = {Input:X, Expected:Y};\nEpochs = 10000;\n\nfor I in range(Epochs): #number of feeds, 1 feed = 1 batch\n  if I%(Epochs\/10)==0: \n    Lossvalue = Sess.run(Loss,Feed);\n    print(\"Loss:\",Lossvalue);\n  #end if\n\n  Sess.run(Training,Feed);\n#end for\n\nLastloss = Sess.run(Loss,Feed);\nprint(\"Loss:\",Lastloss,\"(Last)\");\n\n#eval\nResults = Sess.run(Output,Feed).tolist();\nprint(\"\\nEval:\");\nfor I in range(len(Results)):\n  for J in range(len(Results[I])):\n    for K in range(len(Results[I][J])):\n      Results[I][J][K] = round(Results[I][J][K]*Max_Y);\n#end for i      \nprint(Results);\n\nprint(\"\\nDone.\");\n#eof\n<\/code><\/pre>\n"}]}
+{"QuestionId":58055105,"AnswerCount":1,"Tags":"<python><machine-learning><keras>","CreationDate":"2019-09-23T02:54:58.980","AcceptedAnswerId":58055980.0,"OwnerUserId":11506519.0,"Title":"Keras model producing same output","Body":"<p>I've seen a couple questions that have a similar problem, but none of them solved mine. I'm trying to fit a neural network in Keras to a dataset with 22 input features for binary classification. The problem is that I only have 195 training samples. I know it's a small dataset, but I don't know if it's possible to fit a model with reasonable accuracy (I'm aiming for >95% accuracy). The problem I'm having is that my model is only outputting 1 and getting 75% accuracy because my dataset is 75% positive cases. Here's the code I have:<\/p>\n\n<pre><code>data = pd.read_csv(\"\") #filename omitted, but it loads properly\nscaler = MinMaxScaler()\nX = scaler.fit_transform(X)\nY = data['status']\nX = data.drop(['status', 'name'], axis = 1)\nxTrain, xTest, yTrain, yTest = train_test_split(X, Y, train_size = 0.8)\nmodel = Sequential()\nmodel.add(Dense(48, input_shape=(22,), activation = 'relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(1, activation = 'softmax'))\noptim = keras.optimizers.adam(lr=0.0001)\nmodel.compile(optimizer = optim, loss = 'binary_crossentropy', metrics = ['accuracy'])\nmodel.fit(xTrain, yTrain, epochs = 20, batch_size = 5, validation_data = (xTest, yTest))\n<\/code><\/pre>\n\n<p>I've tried adding more hidden layers, increasing the number of training epochs, and increased and lowered the optimizer's learning rate, but the accuracy stays the same. Here's the link to the dataset: <a href=\"https:\/\/www.dropbox.com\/s\/c4td650b4z7aizc\/fixed.xlsx?dl=0\" rel=\"nofollow noreferrer\">https:\/\/www.dropbox.com\/s\/c4td650b4z7aizc\/fixed.xlsx?dl=0<\/a><\/p>\n","answers":[{"AnswerId":"58055980","CreationDate":"2019-09-23T05:10:24.763","ParentId":null,"OwnerUserId":"3187537","Title":null,"Body":"<p>Some things that you need to try out to get a better accuracy:<\/p>\n\n<ol>\n<li><p>Do not simply input the dataset as it is to the NN. Do some data prep like balancing the response class. Please take a look at various sampling techniques such as Undersampling, Oversampling, SMOTE etc. The accuracy can be improved well if your dataset has balanced class distribution. <\/p><\/li>\n<li><p>Instead of  <code>activation = 'softmax'<\/code>, you should use the <code>sigmoid<\/code> activation function.<\/p><\/li>\n<\/ol>\n\n<p>Apart from thses,you should try several other architectures, lr values, no. of epochs, batch_size,optimizers etc.<\/p>\n"}]}
+{"QuestionId":58055411,"AnswerCount":2,"Tags":"<python><tensorflow><image-processing><keras>","CreationDate":"2019-09-23T03:47:19.217","AcceptedAnswerId":58057267.0,"OwnerUserId":11817448.0,"Title":"NameError: name 'read_image' is not defined","Body":"<p>I want to test my image classification model with a image. \nBut I get some error: <\/p>\n\n<pre class=\"lang-none prettyprint-override\"><code> --&gt; path = 'E:\/My Work Elements\/Thesis Related\/Trash Classification\/Trash New Code\/test\/cardboard\/cardboard42.jpg'\n    test_single_image(path)\n        Traceback (most recent call last):\n\n          File \"&lt;ipython-input-15-4a6021aada0c&gt;\", line 2, in &lt;module&gt;\n            test_single_image(path)\n\n          File \"&lt;ipython-input-14-1654f8f7a46b&gt;\", line 3, in test_single_image\n            images = read_image(path)\n\n        NameError: name 'read_image' is not defined\n<\/code><\/pre>\n\n<p>I use this code:<\/p>\n\n<pre><code>def test_single_image(path):\n  Garbage = ['cardboard','Glass','Metal','paper','plastic','Trash']\n  images = read_image(path)\n  time.sleep(.5)\n  bt_prediction = vgg16.predict(images) \n  preds = model.predict_proba(bt_prediction)\n  for idx, Garbage, x in zip(range(0,6), Garbage , preds[0]):\n   print(\"ID: {}, Label: {} {}%\".format(idx, Garbage, round(x*100,2) ))\n  print('Final Decision:')\n  time.sleep(.5)\n  for x in range(3):\n   print('.'*(x+1))\n   time.sleep(.2)\n  class_predicted = model.predict_classes(bt_prediction)\n  class_dictionary = generator_top.class_indices\n  print(class_dictionary)\n  inv_map = {v: k for k, v in class_dictionary.items()} \n  print(\"Class: {}, prediction Result: {}\".format(class_predicted[0], inv_map[class_predicted[0]])) \n  return load_img(path, target_size=(224,224))\n\npath = 'E:\/My Work Elements\/Thesis Related\/Trash Classification\/Trash New Code\/test\/cardboard\/cardboard42.jpg'\ntest_single_image(path)\n<\/code><\/pre>\n\n<p>How do I fix the code?<\/p>\n","answers":[{"AnswerId":"58055542","CreationDate":"2019-09-23T04:07:58.823","ParentId":null,"OwnerUserId":"10656943","Title":null,"Body":"<p>if you are using opencv to read an image it is supposed to be cv2.imread(). as far as i used vgg16 there should be a read_image function defined before which takes in the image of any shape, resize's it into standard shape (224*224*3) and does any other type of preprocessing required.<\/p>\n\n<p>If you are using keras preprocessing it has a predefined load_img function which does the preprocessing for you .<\/p>\n\n<p>example :<\/p>\n\n<pre><code>from keras.preprocessing.image import load_img\nload_image(path,target_size= (224*224*3))\n<\/code><\/pre>\n\n<p>my sample load function<\/p>\n\n<pre><code>from keras.preprocessing import image \nfrom keras.applications.imagenet_utils import decode_predictions, preprocess_input\ndef load_image(path): \n  img = image.load_img(path,target_size=model.input_shape[1:3]) \n  x = image.img_to_array(img) \n  x = np.expand_dims(x,axis=0) \n  x = preprocess_input(x) \n  return img,x\n<\/code><\/pre>\n"},{"AnswerId":"58057267","CreationDate":"2019-09-23T07:07:55.863","ParentId":null,"OwnerUserId":"6227912","Title":null,"Body":"<p>you can use the keras VGG16 API to get the VGG16 model easily<\/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=False)\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\nfeatures = model.predict(x)\n<\/code><\/pre>\n\n<p>you <code>read_image<\/code> method change to the <code>load_image<\/code> method. The keras API will resize the image to met the vgg16 model request.<\/p>\n"}]}
+{"QuestionId":58055611,"AnswerCount":0,"Tags":"<python><tensorflow><object-detection-api>","CreationDate":"2019-09-23T04:18:31.573","AcceptedAnswerId":null,"OwnerUserId":11077232.0,"Title":"detection bounding box color based on the classes detected in detail or example","Body":"<p>how to code the video object detection with different bounding box color based on the classes detected<\/p>\n\n<p>I try to understand some others question answer but i still dun understand hope some one can show me some example or in more detail<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/56936088\/tensorflow-object-detection-api-setting-specific-color-to-bounding-boxes\">Tensorflow object detection API - Setting specific color to bounding boxes<\/a><\/p>\n","answers":[]}
+{"QuestionId":58055794,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-23T04:47:39.547","AcceptedAnswerId":null,"OwnerUserId":1042017.0,"Title":"Keras - Cannot constrain output by clipping","Body":"<p>I have been trying to constrain the output of the final layer by clipping the values of the output layer of a neural network to the [-1, 1] range. However, the clipping does not help in constraining the values. Is there a more effective way to constrain the output? <\/p>\n\n<pre><code>def encoder(self, x_next, x, y, keep_prob):\n    x_con = concatenate([x_next, x, y])\n    enc1 = Dense(self.encoder_size[0], activation='relu')(x_con)\n    drop = Dropout(keep_prob)(enc1)\n    enc2 = Dense(self.encoder_size[1], activation='relu')(drop)\n    drop = Dropout(keep_prob)(enc2)\n    enc3 = Dense(self.encoder_size[2], activation='relu')(drop)\n    mu = Dense(self.latent_dim, activation='linear', name='encoder_mean')(enc3)\n    mu_clipped = Lambda(self.clip_val)(mu)\n    log_var = Dense(self.latent_dim, activation='linear', name='encoder_log_var', kernel_initializer='zeros')(enc3)\n    return mu_clipped, log_var\n\ndef clip_val(self, args):\n    mu = args\n    mu_clip = K.clip(mu, min_value=-1.0, max_value=1.0)\n    return mu_clip\n<\/code><\/pre>\n\n<p>The output values I am getting range from [-7.1, 7.2] range. How can I constrain the output layer to produce values in the [-1, 1] range? Thanks. <\/p>\n","answers":[{"AnswerId":"58061905","CreationDate":"2019-09-23T11:54:31.960","ParentId":null,"OwnerUserId":"3336192","Title":null,"Body":"<p>Depending on your Problem you should consider to leave the Clipping of the network output to the Activation Function of your last layer: Using a tanh activation ensures y={-1...1}. <\/p>\n"}]}
+{"QuestionId":58055912,"AnswerCount":1,"Tags":"<python><pytorch>","CreationDate":"2019-09-23T05:01:49.040","AcceptedAnswerId":58056095.0,"OwnerUserId":2849157.0,"Title":"How to generate new tensor by given indexes and tensor in pytorch?","Body":"<p>I have a tensor like this:<\/p>\n\n<pre><code>x = torch.tensor([[3, 4, 2], [0, 1, 5]])\n<\/code><\/pre>\n\n<p>and I have a indexes like this:<\/p>\n\n<pre><code>ind = torch.tensor([[1, 1, 0], [0, 0, 1]])\n<\/code><\/pre>\n\n<p>then I want to generate a new tensor by <code>x<\/code> and <code>ind<\/code>:<\/p>\n\n<pre><code>z = torch.tensor([0, 1, 2], [3, 4, 5])\n<\/code><\/pre>\n\n<p>I impement it with python like this:<\/p>\n\n<pre><code># -*- coding: utf-8 -*-\n\nimport torch\n\nx = torch.tensor([[3, 4, 2], [0, 1, 5]])\nind = torch.tensor([[1, 1, 0], [0, 0, 1]])\n\nz = torch.zeros_like(x)\nfor i in range(x.shape[0]):\n    for j in range(x.shape[1]):\n        z[i, j] = x[ind[i][j]][j]\n\nprint(z)\n<\/code><\/pre>\n\n<p>I want to know how to solve this by <code>pytorch<\/code>?<\/p>\n","answers":[{"AnswerId":"58056095","CreationDate":"2019-09-23T05:24:36.217","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You are looking for <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.gather\" rel=\"nofollow noreferrer\"><code>torch.gather<\/code><\/a><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>In [1]: import torch\n\nIn [2]: x = torch.tensor([[3, 4, 2], [0, 1, 5]])\n\nIn [3]: ind = torch.tensor([[1, 1, 0], [0, 0, 1]])\n\nIn [4]: torch.gather(x, 0, ind)\nOut[4]:\ntensor([[0, 1, 2],\n        [3, 4, 5]])\n<\/code><\/pre>\n"}]}
+{"QuestionId":58056495,"AnswerCount":1,"Tags":"<keras><lstm><recurrent-neural-network>","CreationDate":"2019-09-23T06:07:09.890","AcceptedAnswerId":null,"OwnerUserId":3887432.0,"Title":"How to design a LSTM network , which accept multiple input","Body":"<p>Here is the scenario, I want to create a contextual chatbot, which means the bot will answer or reply based on context. As an example <\/p>\n\n<pre><code>Input :[\"text\": \"it was really nice\", \"topic\":movie]\nOutput:[\"text\": \"indeed,it was an awesome movie\",\"topic\":movie]\n<\/code><\/pre>\n\n<p>Whenever I have to consider the only one thing about the input, which is the sentence itself I can do it, all I need to do is to tokenize the sentences and feed into the input of LSTM. But how can I consider \"topic\"?<\/p>\n\n<p>I have already prepared a dataset, in such a format.<\/p>\n\n<p>I am using Keras to build such a bot.<\/p>\n","answers":[{"AnswerId":"58056643","CreationDate":"2019-09-23T06:20:55.870","ParentId":null,"OwnerUserId":"12097394","Title":null,"Body":"<p>I am not really sure what you want to build.\nThe first thing that comes to mind is a normal generativ lstm like this one\n<a href=\"https:\/\/keras.rstudio.com\/articles\/examples\/lstm_text_generation.html\" rel=\"nofollow noreferrer\">https:\/\/keras.rstudio.com\/articles\/examples\/lstm_text_generation.html<\/a><\/p>\n\n<p>wich generates text based on nietches works.\nTo use such a network you would need your training data in a question?, answer format.\nAnd you would need to set your question as the seed.\nYou do not need to load the topic seperatly, as the concept of a neural net is that it learns on its own to understand the data.<\/p>\n"}]}
+{"QuestionId":58056693,"AnswerCount":1,"Tags":"<pytorch><gradient><chainer>","CreationDate":"2019-09-23T06:26:33.523","AcceptedAnswerId":null,"OwnerUserId":12105327.0,"Title":"A gradient clipping in Chainer","Body":"<p>Can I get a Gradient Clipping function in Chainer?<\/p>\n\n<p>I found some codes in Pytorch Documentation : <a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/nn\/utils\/clip_grad.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/nn\/utils\/clip_grad.html<\/a><\/p>\n\n<p>Is there anything like alternative function in Chainer?\nI just found chainer.optimizer_hooks.GradientClipping, but this is hard to utilize.<\/p>\n\n<p>Thanks in advance. <\/p>\n","answers":[{"AnswerId":"58056842","CreationDate":"2019-09-23T06:37:10.787","ParentId":null,"OwnerUserId":"10651406","Title":null,"Body":"<p>how about trying this.\njust I just re-wrote pyTorch function in Chainer style.<\/p>\n\n<pre><code>import cupy\ndef clip_grad_norm(model, max_norm, norm_type=2):\n    params  = list( filter(lambda p : p.grad is not None ,  model.params()) )\n    max_norm = float(max_norm)\n    norm_type = float(norm_type)\n    total_norm = 0.0\n    for p in params:\n        g = p.grad\n        norm = cupy.linalg.norm(g)\n        total_norm += norm**(norm_type)\n    total_norm = total_norm **(1\/norm_type)\n    clip_coef = max_norm \/ (total_norm + 1e-6)\n    if clip_coef &lt; 1:\n        for p in params:\n            g = p.grad\n            p.grad = g * clip_coef\n<\/code><\/pre>\n"}]}
+{"QuestionId":58057590,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2019-09-23T07:31:20.553","AcceptedAnswerId":58064890.0,"OwnerUserId":12078446.0,"Title":"How to insert tensor array into tensor matrix after every second position","Body":"<p>I have a tensor array <em>a<\/em> and a tensor matrix <em>m<\/em>. Now I want to insert <em>a<\/em> into <em>m<\/em> after every second position started at index 0 ending with len(<em>m<\/em>)-2. Let's make an equivalent example using numpy and plain python:<\/p>\n\n<pre><code># define m\nm = np.array([[3,7,6],[4,3,1],[8,4,2],[2,8,7]])\nprint(m)\n#[[3 7 6]\n# [4 3 1]\n# [8 4 2]\n# [2 8 7]]\n\n# define a\na = np.array([1,2,3])\n#[1 2 3]\n\n# insert a into m\nresult = []\nfor i in range(len(m)):\n    result.append(a)\n    result.append(m[i])\nprint(np.array(result))\n#[[1 2 3]\n# [3 7 6]\n# [1 2 3]\n# [4 3 1]\n# [1 2 3]\n# [8 4 2]\n# [1 2 3]\n# [2 8 7]]\n<\/code><\/pre>\n\n<p>I am looking for a solution in tensorflow. I am convinced that there is a solution that doesn't need a loop but I am not able to find one. I hope someone can help me out with this!<\/p>\n","answers":[{"AnswerId":"58058179","CreationDate":"2019-09-23T08:11:33.810","ParentId":null,"OwnerUserId":"3146787","Title":null,"Body":"<p>This should work,<\/p>\n\n<pre><code>np.ravel(np.column_stack((m, np.tile(a, (4, 1))))).reshape(8, 3)\n<\/code><\/pre>\n\n<p>For idea, please refer to <a href=\"https:\/\/stackoverflow.com\/questions\/5347065\/interweaving-two-numpy-arrays\">Interweaving two numpy arrays<\/a>. Apply any solution described there, and reshape.<\/p>\n"},{"AnswerId":"58064890","CreationDate":"2019-09-23T14:42:10.993","ParentId":null,"OwnerUserId":"6537424","Title":null,"Body":"<p>You can concatenate your target vector at the beginning of each line of your matrix, and then reshape it.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\n\ninitial_array = tf.constant([\n    [3, 7, 6],\n    [4, 3, 1],\n    [8, 4, 2],\n    [2, 8, 7],\n])\n\nvector_to_add = [1, 2, 3]\nconcat = tf.concat([[vector_to_add] * initial_array.shape[0], initial_array], axis=1)  # Concatenate vector_to_add to each vector of initial_array\noutput = tf.reshape(concat, (2 * initial_array.shape[0], initial_array.shape[1]))  # Reshape\n<\/code><\/pre>\n"}]}
+{"QuestionId":58057708,"AnswerCount":2,"Tags":"<python><tensorflow><keras><python-3.6><tensorflow-serving>","CreationDate":"2019-09-23T07:40:53.813","AcceptedAnswerId":null,"OwnerUserId":1519955.0,"Title":"Issue with embedding layer when serving a Tensorflow\/Keras model with TF 2.0","Body":"<p>I followed the step in <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/keras\/feature_columns\" rel=\"nofollow noreferrer\">one of the TF beginner tutorial<\/a> to create a simple classification model. They are the following:<\/p>\n\n<pre><code>from __future__ import absolute_import, division, print_function, unicode_literals\nimport numpy as np\nimport pandas as pd\nimport tensorflow as tf\nfrom tensorflow import feature_column\nfrom tensorflow.keras import layers\nfrom sklearn.model_selection import train_test_split\n\nURL = 'https:\/\/storage.googleapis.com\/applied-dl\/heart.csv'\ndataframe = pd.read_csv(URL)\ndataframe.head()\n\ntrain, test = train_test_split(dataframe, test_size=0.2)\ntrain, val = train_test_split(train, test_size=0.2)\n\ndef df_to_dataset(dataframe, shuffle=True, batch_size=32):\n  dataframe = dataframe.copy()\n  labels = dataframe.pop('target')\n  ds = tf.data.Dataset.from_tensor_slices((dict(dataframe), labels))\n  if shuffle:\n    ds = ds.shuffle(buffer_size=len(dataframe))\n  ds = ds.batch(batch_size)\n  return ds\n\nbatch_size = 5 # A small batch sized is used for demonstration purposes\ntrain_ds = df_to_dataset(train, batch_size=batch_size)\nval_ds = df_to_dataset(val, shuffle=False, batch_size=batch_size)\ntest_ds = df_to_dataset(test, shuffle=False, batch_size=batch_size)\n\nfeature_columns = []\nfor header in ['age', 'trestbps', 'chol', 'thalach', 'oldpeak', 'slope', 'ca']:\n  feature_columns.append(feature_column.numeric_column(header))\nthal_embedding = feature_column.embedding_column(thal, dimension=8)\nfeature_columns.append(thal_embedding)\n\nfeature_layer = tf.keras.layers.DenseFeatures(feature_columns)\n\nbatch_size = 32\ntrain_ds = df_to_dataset(train, batch_size=batch_size)\nval_ds = df_to_dataset(val, shuffle=False, batch_size=batch_size)\ntest_ds = df_to_dataset(test, shuffle=False, batch_size=batch_size)\n\n\nmodel = tf.keras.Sequential([\n  feature_layer,\n  layers.Dense(128, activation='relu'),\n  layers.Dense(128, activation='relu'),\n  layers.Dense(1, activation='sigmoid')\n])\n\nmodel.compile(optimizer='adam',\n              loss='binary_crossentropy',\n              metrics=['accuracy'],\n              run_eagerly=True)\n\nmodel.fit(train_ds,\n          validation_data=val_ds,\n          epochs=5)\n<\/code><\/pre>\n\n<p>And I saved the model with:<\/p>\n\n<pre><code>model.save(\"model\/\", save_format='tf')\n<\/code><\/pre>\n\n<p>Then, I try to serve this model using this <a href=\"https:\/\/www.tensorflow.org\/tfx\/serving\/docker\" rel=\"nofollow noreferrer\">TF tutorial<\/a>. I do the following:<\/p>\n\n<pre><code>docker pull tensorflow\/serving\ndocker run -p 8501:8501 --mount type=bind,source=\/path\/to\/model\/,target=\/models\/model -e MODEL_NAME=mo\n<\/code><\/pre>\n\n<p>And I try to call the model this way:<\/p>\n\n<pre><code>curl -d '{\"inputs\": {\"age\": [0], \"trestbps\": [0], \"chol\": [0], \"thalach\": [0], \"oldpeak\": [0], \"slope\": [1], \"ca\": [0], \"exang\": [0], \"restecg\": [0], \"fbs\": [0], \"cp\": [0], \"sex\": [0], \"thal\": [\"normal\"], \"target\": [0] }}' -X POST http:\/\/localhost:8501\/v1\/models\/model:predict\n<\/code><\/pre>\n\n<p>I get the following error: <\/p>\n\n<blockquote>\n  <p>{ \"error\": \"indices = 1 is not in [0, 1)\\n\\t [[{{node StatefulPartitionedCall_51\/StatefulPartitionedCall\/sequential\/dense_features\/thal_embedding\/thal_embedding_weights\/GatherV2}}]]\" }<\/p>\n<\/blockquote>\n\n<p>It seems to be related to the embedding layer for the \"thal\" feature. But I have no idea what \"indices = 1 is not in [0, 1)\" means and why it happens.<\/p>\n\n<p>When the error occurs, here is what the TF docker server logs:<\/p>\n\n<blockquote>\n  <p>2019-09-23 12:50:43.921721: W external\/org_tensorflow\/tensorflow\/core\/framework\/op_kernel.cc:1502] OP_REQUIRES failed at lookup_table_op.cc:952 : Failed precondition: Table already initialized.<\/p>\n<\/blockquote>\n\n<p>Any idea where the error comes from and how I could fix it?<\/p>\n\n<p>Python version: 3.6<\/p>\n\n<p>tensorflow version: 2.0.0-rc0<\/p>\n\n<p>latest TensorFlow\/serving (as of 20\/09\/2019)<\/p>\n\n<p>Model signature:<\/p>\n\n<pre><code>signature_def['__saved_model_init_op']:\n  The given SavedModel SignatureDef contains the following input(s):\n  The given SavedModel SignatureDef contains the following output(s):\n    outputs['__saved_model_init_op'] tensor_info:\n        dtype: DT_INVALID\n        shape: unknown_rank\n        name: NoOp\n  Method name is: \n\nsignature_def['serving_default']:\n  The given SavedModel SignatureDef contains the following input(s):\n    inputs['age'] tensor_info:\n        dtype: DT_INT32\n        shape: (-1, 1)\n        name: serving_default_age:0\n    inputs['ca'] tensor_info:\n        dtype: DT_INT32\n        shape: (-1, 1)\n        name: serving_default_ca:0\n    inputs['chol'] tensor_info:\n        dtype: DT_INT32\n        shape: (-1, 1)\n        name: serving_default_chol:0\n    inputs['cp'] tensor_info:\n        dtype: DT_INT32\n        shape: (-1, 1)\n        name: serving_default_cp:0\n    inputs['exang'] tensor_info:\n        dtype: DT_INT32\n        shape: (-1, 1)\n        name: serving_default_exang:0\n    inputs['fbs'] tensor_info:\n        dtype: DT_INT32\n        shape: (-1, 1)\n        name: serving_default_fbs:0\n    inputs['oldpeak'] tensor_info:\n        dtype: DT_FLOAT\n        shape: (-1, 1)\n        name: serving_default_oldpeak:0\n    inputs['restecg'] tensor_info:\n        dtype: DT_INT32\n        shape: (-1, 1)\n        name: serving_default_restecg:0\n    inputs['sex'] tensor_info:\n        dtype: DT_INT32\n        shape: (-1, 1)\n        name: serving_default_sex:0\n    inputs['slope'] tensor_info:\n        dtype: DT_INT32\n        shape: (-1, 1)\n        name: serving_default_slope:0\n    inputs['thal'] tensor_info:\n        dtype: DT_STRING\n        shape: (-1, 1)\n        name: serving_default_thal:0\n    inputs['thalach'] tensor_info:\n        dtype: DT_INT32\n        shape: (-1, 1)\n        name: serving_default_thalach:0\n    inputs['trestbps'] tensor_info:\n        dtype: DT_INT32\n        shape: (-1, 1)\n        name: serving_default_trestbps:0\n  The given SavedModel SignatureDef contains the following output(s):\n    outputs['output_1'] tensor_info:\n        dtype: DT_FLOAT\n        shape: (-1, 1)\n        name: StatefulPartitionedCall:0\n  Method name is: tensorflow\/serving\/predict\n<\/code><\/pre>\n","answers":[{"AnswerId":"58337841","CreationDate":"2019-10-11T09:07:49.937","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>I am also trying to serve a model consisting an embedding layer, lstm layer etc but I am receing some other errors. I have even raised an <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/33136\" rel=\"nofollow noreferrer\">issue<\/a> on TF.<\/p>\n\n<p>Anyways, the problem I see in your code is with the type of saved model you are using for serving with Docker. If you read <a href=\"https:\/\/www.tensorflow.org\/tfx\/serving\/docker#running_a_serving_image\" rel=\"nofollow noreferrer\">here<\/a>, it says following point-<\/p>\n\n<p><code>A<\/code> <strong>SavedModel<\/strong> <code>to serve<\/code><\/p>\n\n<p>which is not the keras <code>model.save<\/code> but is another TF API, <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model#creating_a_savedmodel_from_keras\" rel=\"nofollow noreferrer\">here<\/a> is the way describing to create <strong>SavedModel<\/strong> from keras trained model. Give this a try and let us know the results.<\/p>\n"},{"AnswerId":"58254464","CreationDate":"2019-10-06T05:25:25.840","ParentId":null,"OwnerUserId":"7838040","Title":null,"Body":"<p>The problem seems with the format you're sending in. Can you post the signature of the model? \ncan't post it as a comment due to low reputation. <\/p>\n"}]}
+{"QuestionId":58058003,"AnswerCount":0,"Tags":"<tensorflow><keras><multipleoutputs>","CreationDate":"2019-09-23T08:01:17.440","AcceptedAnswerId":null,"OwnerUserId":8214268.0,"Title":"How to combine lots of outputs into a matrix with custom loss in Keras","Body":"<p>I am trying to train a model to read barcodes of fixed length. At the moment, my model looks like this (for 10-digits length):<\/p>\n\n<pre><code>    x = ly.Dense(units=4096)(x)\n    x = ly.Activation(\"relu\")(x)  \n    x = ly.Activation(\"relu\")(x)\n    d0 = ly.Dense(units= 10, activation=\"softmax\")(x)\n    d1 = ly.Dense(units= 10, activation=\"softmax\")(x)\n    d2 = ly.Dense(units= 10, activation=\"softmax\")(x)\n    d3 = ly.Dense(units= 10, activation=\"softmax\")(x)\n    d4 = ly.Dense(units= 10, activation=\"softmax\")(x)\n    d5 = ly.Dense(units= 10, activation=\"softmax\")(x)\n    d6 = ly.Dense(units= 10, activation=\"softmax\")(x)\n    d7 = ly.Dense(units= 10, activation=\"softmax\")(x)\n    d8 = ly.Dense(units= 10, activation=\"softmax\")(x)\n    d9 = ly.Dense(units= 10, activation=\"softmax\")(x)\n\n    model = tf.keras.Model(inputs=[inputs], outputs=[d0, d1, d2, d3, d4, d5, d6, d7, d8, d9]) \n<\/code><\/pre>\n\n<p>And my model is compiled with 'categorical_crossentropy' loss. It does work, however I have a lot of training, validation, testing loss and accuracy to monitor (about 30 in total). <\/p>\n\n<p>I would like to be able to scale this model to more digits (30, 100 ?). But it would be impractical to do it this way. <\/p>\n\n<p>I tried to implement this by stacking all the outputs into a single vector (size 100), as well as the labels, and compile the model with binary_crossentropy. However it never got to converge. I wonder how is it because it should be the same no ?  I verified that the labels were correctly stacked and yes they were. <\/p>\n\n<p>So I tried to reshape the output of Dense(4096) into a matrix of size (10, 10) with a Reshape layer, fed it with the labels reshaped as stacked columns (or rows I don't remember), and it would not work with either binary or categorical crossentropy, so I guess that I should implement a custom loss. <\/p>\n\n<p>What would be the best way to solve this problem  ?<\/p>\n\n<p>Thank you a lot.<\/p>\n","answers":[]}
+{"QuestionId":58058794,"AnswerCount":0,"Tags":"<machine-learning><keras><nlp><artificial-intelligence>","CreationDate":"2019-09-23T08:50:46.040","AcceptedAnswerId":null,"OwnerUserId":12105926.0,"Title":"Keras - fit_generator function not learning","Body":"<p>I am working on a NMT (Neural Machine Translation) model . I initially used the fit function of keras to train the model . It learned pretty well with an accuracy over 0.85 . But when i try to use the fit_generator function , the model does not seem to learn and the accuracy is stuck at 0.14 - 0.15 . Am I doing some mistake in the generator function or selecting batch_size or using steps_per_epoch ?<\/p>\n\n<p>BATCH SIZE:<\/p>\n\n<pre><code>batch = 32\n<\/code><\/pre>\n\n<p>DATA GENERATOR :<\/p>\n\n<pre><code>def data_generator():\n  while True:\n\n    idx = np.random.randint(len(eng),size=batch).tolist()\n\n    for i in idx:\n\n      X1 , X2 , Y = create_sequences([i] , batch)\n\n      X1 = pad_sequences(X1 , maxlen = Xlen , padding=\"post\")\n      X2 = pad_sequences(X2 , maxlen = Ylen , padding=\"post\")\n      Y = to_categorical(Y , Yvocab)\n\n      X1, X2, Y = sklearn.utils.shuffle(X1,X2, Y, random_state=0)\n\n      yield [[X1,X2] , Y]\n<\/code><\/pre>\n\n<p>CREATE SEQUENCES:<\/p>\n\n<pre><code>def create_sequences(inp ,batch):\n\n  X1 = []\n  X2 = []\n  Y = []\n\n  X1_tmp = []\n  X2_tmp = []\n  Y_tmp = []\n\n  for i in inp:\n\n    tmp = []\n\n    for j in ger[i]:\n      tmp.append(j)\n\n    for k in range(len(tmp) - 1):    \n      X1_tmp.append(eng[i])\n      X2_tmp.append(tmp[:k+1])\n      Y_tmp.append(tmp[k+1])\n\n  x = randint(0,len(X1_tmp)-1)\n  X1.append(X1_tmp[x])\n  X2.append(X2_tmp[x])\n  Y.append(Y_tmp[x])\n\n  return np.array(X1) , np.array(X2)  , np.array(Y)\n<\/code><\/pre>\n\n<p>MODEL:<\/p>\n\n<pre><code>inp1 = keras.layers.Input(shape=(Xlen , ))\nf1 = keras.layers.Embedding(Xvocab , 64 , input_length=Xlen)(inp1)\nf2 = keras.layers.Dropout(0.3)(f1)\nf3 = keras.layers.LSTM(256)(f2)\nf4 = keras.layers.Dropout(0.3)(f3)\nf5 = keras.layers.Dense(256 , activation='relu')(f4)\n\ninp2 = keras.layers.Input(shape=(Ylen, ))\ns1 = keras.layers.Embedding(Yvocab , 64 , input_length=Ylen)(inp2)\ns2 = keras.layers.Dropout(0.3)(s1)\ns3 = keras.layers.LSTM(256)(s2)\n\ndecoder1 = keras.layers.add([f5,s3])\ndecoder2 = keras.layers.Dense(256 , activation='relu')(decoder1)\nout = keras.layers.Dense(Yvocab , activation='softmax')(decoder2)\n\nmodel = keras.models.Model(inputs = [inp1,inp2] , outputs = out)\n\nmodel.compile(optimizer = 'adam' , loss = 'categorical_crossentropy' , metrics = ['accuracy'])\n<\/code><\/pre>\n\n<p>FIT_GENERATOR:<\/p>\n\n<pre><code>epochs = 10\n\nfor i in range(epochs):\n\n  generator = data_generator()\n\n  model.fit_generator(generator , epochs = 1 , steps_per_epoch=len(eng)\/\/batch)\n\n  model.save(\"drive\/My Drive\/Colab Notebooks\/nmtnew\" + str(i) + \".h5\")\n<\/code><\/pre>\n\n<p>RESULTS:<\/p>\n\n<pre><code>Epoch 1\/1\n3013\/3013 [==============================] - 1041s 346ms\/step - loss: 7.0601 - acc: 0.1437\nEpoch 1\/1\n3013\/3013 [==============================] - 1037s 344ms\/step - loss: 6.4778 - acc: 0.1543\nEpoch 1\/1\n3013\/3013 [==============================] - 1042s 346ms\/step - loss: 6.4813 - acc: 0.1580\nEpoch 1\/1\n3013\/3013 [==============================] - 1039s 345ms\/step - loss: 6.6110 - acc: 0.1577\nEpoch 1\/1\n3013\/3013 [==============================] - 1036s 344ms\/step - loss: 6.7250 - acc: 0.1500\nEpoch 1\/1\n3013\/3013 [==============================] - 1038s 345ms\/step - loss: 6.7857 - acc: 0.1507\nEpoch 1\/1\n3013\/3013 [==============================] - 1038s 345ms\/step - loss: 6.7567 - acc: 0.1487\nEpoch 1\/1\n3013\/3013 [==============================] - 1039s 345ms\/step - loss: 7.0087 - acc: 0.1424\nEpoch 1\/1\n3013\/3013 [==============================] - 1029s 342ms\/step - loss: 6.8201 - acc: 0.1484\nEpoch 1\/1\n3013\/3013 [==============================] - 1023s 340ms\/step - loss: 6.8712 - acc: 0.1540\n<\/code><\/pre>\n\n<p>I have also used different batch sizes like 256 , 512 , 10 etc.\nThe model seems to always predict 2 which is 'endseq'<\/p>\n","answers":[]}
+{"QuestionId":58059119,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><keras><conv-neural-network>","CreationDate":"2019-09-23T09:09:50.133","AcceptedAnswerId":58060079.0,"OwnerUserId":9813733.0,"Title":"Convert a tensor to numpy array in keras","Body":"<p>I want to convert the tensor of shape (?,224,224,3) to a numpy array in keras.<\/p>\n\n<pre><code>Tensor(\"add_1\/add_3:0\", shape=(?, 224, 224, 3), dtype=float32)\n<\/code><\/pre>\n\n<p>My basic CNN code is as below:<\/p>\n\n<pre><code>final_model = Model(input=[img_input], output=[act1,act2,act3,act4,act5])\naddops(final_model)    \n\ndef addops(model):\n        output = model.output\n        factor = 2\n        for i in range(len(output)):\n                output[i] = UpSampling2D(size = (factor**i,factor**i),interpolation='bilinear')(output[i])\n                output[i] = (ZeroPadding2D((1,1)))(output[i])\n                output[i] = (Conv2D(3,(3,3)))(output[i])\n        out = Add()([output[0],output[1],output[2],output[3],output[4]])\n        print(out)\n        sess = tf.InteractiveSession()\n        outarray = tf.Variable(out).eval()\n        print(outarray)\n<\/code><\/pre>\n\n<p>Basically, I am taking a pretrained model and capturing the feature maps of a few intermediate layers and doing those transformations on them to reshape them to the size of the input image. Later, i am adding those reshaped feature maps to a single tensor. I want to convert this tensor which is named as <strong><em>'out'<\/em><\/strong> in the above program to a numpy array. This is part of the project I am trying to implement but I am struck here. <\/p>\n\n<p>When I run this code, I get the below error <\/p>\n\n<pre><code>ValueError: initial_value must have a shape specified: Tensor(\"add_1\/add_3:0\", shape=(?, 224, 224, 3), dtype=float32)\n<\/code><\/pre>\n\n<p>The program gets compiled if I don't initialize the session and take just the tensor <strong><em>'out'<\/em><\/strong> as output. I tried several other ways and have referred to various questions on Stack Overflow, but none of them gave me the required results. What is the best way to convert this tensor <strong><em>'out'<\/em><\/strong> in the above code to a numpy array. <\/p>\n\n<p><strong><em>Note:<\/em><\/strong> I want to convert this numpy array to an image later. I am happy to know if there are any methods to directly convert this tensor to an image as well. <\/p>\n\n<p><strong><em>EDIT1<\/em><\/strong>: I passed an image to fill in the initial_value in the shape, but now I see another error which is basically an issue with converting this tensor to a numpy array.<\/p>\n\n<pre><code>    def addops(model,image):\n        image = load_img(image,target_size=(224,224))\n        image = img_to_array(image)\n        image = np.expand_dims(image,axis=0)\n        val =model.predict(image)\n#       print(val)\n        output = val\n        factor = 2\n        for i in range(len(output)):\n                output[i] = tf.convert_to_tensor(output[i],np.float32)\n                output[i] = UpSampling2D(size = (factor**i,factor**i),interpolation='bilinear')(output[i])\n                output[i] = (ZeroPadding2D((1,1)))(output[i])\n                output[i] = (Conv2D(3,(3,3)))(output[i])\n        out = Add()([output[0],output[1],output[2],output[3],output[4]])\n        print(out)\n        sess = tf.InteractiveSession()\n        sess.run(tf.global_variables_initializer())\n        sess.run(tf.local_variables_initializer())\n        outarray = tf.Variable(out).eval()\n        print(outarray)\n<\/code><\/pre>\n\n<p>The error message is: <\/p>\n\n<pre><code>    Using TensorFlow backend.\nWARNING:tensorflow:From \/home\/ssindhu\/deeplab_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/op_def_library.py:263: colocate_with (from tensorflow.python.framework.ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nColocations handled automatically by placer.\n2019-09-23 02:39:52.684145: I tensorflow\/core\/platform\/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\n2019-09-23 02:39:52.690019: I tensorflow\/core\/platform\/profile_utils\/cpu_utils.cc:94] CPU Frequency: 3493110000 Hz\n2019-09-23 02:39:52.692198: I tensorflow\/compiler\/xla\/service\/service.cc:150] XLA service 0x51c36f0 executing computations on platform Host. Devices:\n2019-09-23 02:39:52.692254: I tensorflow\/compiler\/xla\/service\/service.cc:158]   StreamExecutor device (0): &lt;undefined&gt;, &lt;undefined&gt;\nWARNING:tensorflow:From \/home\/ssindhu\/deeplab_env\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py:3445: calling dropout (from tensorflow.python.ops.nn_ops) with keep_prob is deprecated and will be removed in a future version.\nInstructions for updating:\nPlease use `rate` instead of `keep_prob`. Rate should be set to `rate = 1 - keep_prob`.\n\/home\/ssindhu\/hypercolumns.py:107: UserWarning: Update your `Model` call to the Keras 2 API: `Model(inputs=[&lt;tf.Tenso..., outputs=[&lt;tf.Tenso...)`\n  final_model = Model(input=[img_input], output=[act1,act2,act3,act4,act5])\nFrontend weights loaded.\nTensor(\"add_1\/add_3:0\", shape=(1, 224, 224, 3), dtype=float32)\nTraceback (most recent call last):\n  File \"\/home\/ssindhu\/deeplab_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1334, in _do_call\n    return fn(*args)\n  File \"\/home\/ssindhu\/deeplab_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1319, in _run_fn\n    options, feed_dict, fetch_list, target_list, run_metadata)\n  File \"\/home\/ssindhu\/deeplab_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1407, in _call_tf_sessionrun\n    run_metadata)\ntensorflow.python.framework.errors_impl.FailedPreconditionError: Attempting to use uninitialized value Variable\n         [[{{node _retval_Variable_0_0}}]]\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"training.py\", line 18, in &lt;module&gt;\n    gethypercols(model,'JPEGImages\/2007_000033.jpg')\n  File \"\/home\/ssindhu\/hypercolumns.py\", line 34, in gethypercols\n    outarray = tf.Variable(out).eval()\n  File \"\/home\/ssindhu\/deeplab_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/ops\/variables.py\", line 1695, in eval\n    return self._variable.eval(session=session)\n  File \"\/home\/ssindhu\/deeplab_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 695, in eval\n    return _eval_using_default_session(self, feed_dict, self.graph, session)\n  File \"\/home\/ssindhu\/deeplab_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 5181, in _eval_using_default_session\n    return session.run(tensors, feed_dict)\n  File \"\/home\/ssindhu\/deeplab_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 929, in run\n    run_metadata_ptr)\n  File \"\/home\/ssindhu\/deeplab_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1152, in _run\n    feed_dict_tensor, options, run_metadata)\n  File \"\/home\/ssindhu\/deeplab_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1328, in _do_run\n    run_metadata)\n  File \"\/home\/ssindhu\/deeplab_env\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1348, in _do_call\n    raise type(e)(node_def, op, message)\ntensorflow.python.framework.errors_impl.FailedPreconditionError: Attempting to use uninitialized value Variable\n         [[{{node _retval_Variable_0_0}}]]\n<\/code><\/pre>\n\n<p>It says that I am trying to use an uninitialized value variable. I have seen that we have to initialize the variables in the session for tensorflow and so I have included those lines as well.But still I see the same error. How could I resolve this.<\/p>\n","answers":[{"AnswerId":"58060079","CreationDate":"2019-09-23T10:09:27.010","ParentId":null,"OwnerUserId":"5652313","Title":null,"Body":"<p>If you want to compute the value of a tensor (here <code>out<\/code> is a tensor), you just run the tensor in a session, there is no need to create a variable out of it. You can just run <code>out<\/code> in a session.<\/p>\n\n<p>Variables are for storing trainable (and other) parameters of your model. What you do is that you create a new variable that should be initialized with <code>out<\/code> and run it without initialization.<\/p>\n"}]}
+{"QuestionId":58059130,"AnswerCount":0,"Tags":"<python><keras><dataset><mnist>","CreationDate":"2019-09-23T09:10:32.257","AcceptedAnswerId":null,"OwnerUserId":10345172.0,"Title":"How Can I Load My Own Images Instead Of Mnist Dataset Images?","Body":"<p>Hi I am very frustrated that all the ml examples I see only use the MNIST dataset and don't use custom images and I would like to load in my own image dataset of pokemon. Here is my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># -*- coding: utf-8 -*-\n\"\"\"autoencoder.ipynb\n\nAutomatically generated by Colaboratory.\n\nOriginal file is located at\n    https:\/\/colab.research.google.com\/drive\/1P5rdEhs3lzcNMK9SWsOXdNq9nl74E54D\n\"\"\"\n\nfrom keras.layers import Input, Dense\nfrom keras.models import Model\n\n# this is the size of our encoded representations\nencoding_dim = 32  # 32 floats -&gt; compression of factor 24.5, assuming the input is 784 floats\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\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))\n\nautoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')\nfrom keras.datasets import mnist\nimport numpy as np\n(x_train, _), (x_test, _) = mnist.load_data()\n\nx_train = x_train.astype('float32') \/ 255.\nx_test = x_test.astype('float32') \/ 255.\nx_train = x_train.reshape((len(x_train), np.prod(x_train.shape[1:])))\nx_test = x_test.reshape((len(x_test), np.prod(x_test.shape[1:])))\nprint(x_train.shape)\nprint(x_test.shape)\n\nautoencoder.fit(x_train, x_train,\n                epochs=10,\n                batch_size=256,\n                shuffle=True,\n                validation_data=(x_test, x_test))\n\n# encode and decode some digits\n# note that we take them from the *test* set\nencoded_imgs = encoder.predict(x_test)\ndecoded_imgs = decoder.predict(encoded_imgs)\n\n# use Matplotlib (don't ask)\nimport matplotlib.pyplot as plt\n\nn = 10  # how many digits we will display\nplt.figure(figsize=(20, 4))\nfor i in range(n):\n    # display original\n    ax = plt.subplot(2, n, i + 1)\n    plt.imshow(x_test[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, n, i + 1 + n)\n    plt.imshow(decoded_imgs[i].reshape(28, 28))\n    plt.gray()\n    ax.get_xaxis().set_visible(False)\n    ax.get_yaxis().set_visible(False)\nplt.show()\n<\/code><\/pre>\n\n<p>Here is the structure of my images<\/p>\n\n<pre><code>images\n  |_____abomasnow\n        |___image\n  |_____abra\n        |___image\n  |_____absol\n        |___image\n  |_____accelgor\n        |___image\n  ...\n  |_____zweilous\n        |___image\n  |_____zubat\n        |___image\n  |_____zorua\n        |___image\n<\/code><\/pre>\n\n<p>I've tried using <a href=\"https:\/\/github.com\/Arlen0615\/Convert-own-data-to-MNIST-format\" rel=\"nofollow noreferrer\">convert_to_mnist_format<\/a> but I got:\nValueError: could not broadcast input array from shape (120,120,3) into shape (120,120,4)<\/p>\n\n<p>So I would like some help to make <a href=\"https:\/\/drive.google.com\/drive\/folders\/1Tb7Rwek4cdl8WmIuVz_i6KZDn_s51BIC?usp=sharing\" rel=\"nofollow noreferrer\">this<\/a> dataset able to be read by the Autoencoder above<\/p>\n","answers":[]}
+{"QuestionId":58059221,"AnswerCount":1,"Tags":"<python-3.x><conv-neural-network><pytorch>","CreationDate":"2019-09-23T09:16:13.947","AcceptedAnswerId":58059613.0,"OwnerUserId":9297078.0,"Title":"Pytorch: ValueError: Expected input batch_size (32) to match target batch_size (64)","Body":"<p>Tried to run the CNN examples on MNIST dataset, batch size=64, channel =1, n_h=28, n_w=28, n_iters = 1000. The program runs for first 500 interation and then gives the above mentioned error.\nThere are same topics already being discussed on the forum such as : <a href=\"https:\/\/stackoverflow.com\/questions\/56719867\/pytorch-expected-input-batch-size-12-to-match-target-batch-size-64\">topic 1<\/a>\nand <a href=\"https:\/\/stackoverflow.com\/questions\/54928638\/pytorch-cnn-error-expected-input-batch-size-4-to-match-target-batch-size-64\">topic 2<\/a>, but none of them could help me identify the mistake in the following code:<\/p>\n\n<pre><code>class CNN_MNIST(nn.Module):\ndef __init__(self):\n    super(CNN_MNIST,self).__init__()\n\n    # convolution layer 1\n    self.cnn1 = nn.Conv2d(in_channels=1, out_channels= 32, kernel_size=5,\n                          stride=1,padding=2)\n\n    # ReLU activation \n    self.relu1 = nn.ReLU()\n\n    # maxpool 1\n    self.maxpool1 = nn.MaxPool2d(kernel_size=2,stride=2)\n\n    # convolution 2\n    self.cnn2 = nn.Conv2d(in_channels=32, out_channels=64, kernel_size=5,\n                          stride=1,padding=2)\n\n    # ReLU activation \n    self.relu2 = nn.ReLU()\n\n    # maxpool 2\n    self.maxpool2 = nn.MaxPool2d(kernel_size=2,stride=2)\n\n    # fully connected 1\n    self.fc1 = nn.Linear(7*7*64,1000)\n    # fully connected 2\n    self.fc2 = nn.Linear(1000,10)\n\ndef forward(self,x):\n\n    # convolution 1\n    out = self.cnn1(x)\n    # activation function\n    out = self.relu1(out)\n    # maxpool 1\n    out = self.maxpool1(out)\n\n    # convolution 2\n    out = self.cnn2(out)\n    # activation function\n    out = self.relu2(out)\n    # maxpool 2\n    out = self.maxpool2(out)\n\n    # flatten the output\n    out = out.view(out.size(0),-1)\n\n    # fully connected layers\n    out = self.fc1(out)\n    out = self.fc2(out)\n\n    return out\n# model trainning\ncount = 0\nloss_list = []\niteration_list = []\naccuracy_list = []\n\nfor epoch in range(int(n_epochs)):\n    for i, (image,labels) in enumerate(train_loader):\n\n    train = Variable(image)\n    labels = Variable(labels)\n\n    # clear gradient\n    optimizer.zero_grad()\n\n    # forward propagation\n    output = cnn_model(train)\n\n    # calculate softmax and cross entropy loss\n    loss = error(output,label)\n\n    # calculate gradients\n    loss.backward()\n\n    # update the optimizer\n    optimizer.step()\n\n    count += 1\n\n    if count % 50 ==0:\n        # calculate the accuracy\n        correct = 0\n        total = 0\n\n        # iterate through the test data\n        for image, labels in test_loader:\n\n            test = Variable(image)\n\n            # forward propagation\n            output = cnn_model(test)\n\n            # get prediction\n            predict = torch.max(output.data,1)[1]\n\n            # total number of labels\n            total += len(labels)\n\n            # correct prediction\n            correct += (predict==labels).sum()\n\n        # accuracy\n        accuracy = 100*correct\/float(total)\n\n        # store loss, number of iteration, and accuracy\n        loss_list.append(loss.data)\n        iteration_list.append(count)\n        accuracy_list.append(accuracy)\n\n        # print loss and accurcay as the algorithm progresses\n        if count % 500 ==0:\n            print('Iteration :{}    Loss :{}    Accuracy :\n\n{}'.format(count,loss.item(),accuracy))\n<\/code><\/pre>\n\n<p>The error is as follows:<\/p>\n\n<pre><code>    ---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-19-9e93a242961b&gt; in &lt;module&gt;\n     18 \n     19         # calculate softmax and cross entropy loss\n---&gt; 20         loss = error(output,label)\n     21 \n     22         # calculate gradients\n\n~\\Anaconda3\\lib\\site-packages\\torch\\nn\\modules\\module.py in __call__(self, *input, **kwargs)\n    545             result = self._slow_forward(*input, **kwargs)\n    546         else:\n--&gt; 547             result = self.forward(*input, **kwargs)\n    548         for hook in self._forward_hooks.values():\n    549             hook_result = hook(self, input, result)\n\n~\\Anaconda3\\lib\\site-packages\\torch\\nn\\modules\\loss.py in forward(self, input, target)\n    914     def forward(self, input, target):\n    915         return F.cross_entropy(input, target, weight=self.weight,\n--&gt; 916                                ignore_index=self.ignore_index, reduction=self.reduction)\n    917 \n    918 \n\n~\\Anaconda3\\lib\\site-packages\\torch\\nn\\functional.py in cross_entropy(input, target, weight, size_average, ignore_index, reduce, reduction)\n   1993     if size_average is not None or reduce is not None:\n   1994         reduction = _Reduction.legacy_get_string(size_average, reduce)\n-&gt; 1995     return nll_loss(log_softmax(input, 1), target, weight, None, ignore_index, None, reduction)\n   1996 \n   1997 \n\n~\\Anaconda3\\lib\\site-packages\\torch\\nn\\functional.py in nll_loss(input, target, weight, size_average, ignore_index, reduce, reduction)\n   1820     if input.size(0) != target.size(0):\n   1821         raise ValueError('Expected input batch_size ({}) to match target batch_size ({}).'\n-&gt; 1822                          .format(input.size(0), target.size(0)))\n   1823     if dim == 2:\n   1824         ret = torch._C._nn.nll_loss(input, target, weight, _Reduction.get_enum(reduction), ignore_index)\n\nValueError: Expected input batch_size (32) to match target batch_size (64).\n<\/code><\/pre>\n","answers":[{"AnswerId":"58059613","CreationDate":"2019-09-23T09:39:39.943","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You are providing the wrong target to your loss:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>loss = error(output, label)\n<\/code><\/pre>\n\n<p>While your loader gives you<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    for i, (image,labels) in enumerate(train_loader):\n\n    train = Variable(image)\n    labels = Variable(labels)\n<\/code><\/pre>\n\n<p>So you have a variable name <code>labels<\/code> (with <code>s<\/code>)  from the loader, yet you feed <code>label<\/code> (no <code>s<\/code>) to your loss. <\/p>\n\n<p>Batch size is the least of your worries.<\/p>\n"}]}
+{"QuestionId":58059268,"AnswerCount":2,"Tags":"<python-3.x><deep-learning><pytorch>","CreationDate":"2019-09-23T09:19:25.290","AcceptedAnswerId":58075433.0,"OwnerUserId":3293726.0,"Title":"Pytorch autograd.grad how to write the parameters for multiple outputs?","Body":"<p>In the <a href=\"https:\/\/pytorch.org\/docs\/master\/autograd.html?highlight=autograd%20grad#torch.autograd.grad\" rel=\"nofollow noreferrer\">documentation<\/a> of <code>torch.autograd.grad<\/code>, it is stated that, for parameters,<\/p>\n\n<blockquote>\n  <p>parameters: <\/p>\n  \n  <p>outputs (sequence of Tensor) \u2013 outputs of the differentiated function.<\/p>\n  \n  <p>inputs (sequence of Tensor) \u2013 Inputs w.r.t.\n  which the gradient will be returned (and not accumulated into .grad).<\/p>\n<\/blockquote>\n\n<p>I try the following:<\/p>\n\n<pre><code>a = torch.rand(2, requires_grad=True)\nb = torch.rand(2, requires_grad=True)\nc = a+b\nd = a-b\n\ntorch.autograd.grad([c, d], [a, b]) #ValueError: only one element tensors can be converted to Python scalars\ntorch.autograd.grad(torch.tensor([c, d]), torch.tensor([a, b])) #RuntimeError: grad can be implicitly created only for scalar outputs\n<\/code><\/pre>\n\n<p>I would like to get gradients of a list of tensors w.r.t another list of tensors. What is the correct way to feed the parameters?<\/p>\n","answers":[{"AnswerId":"58075433","CreationDate":"2019-09-24T07:46:30.407","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>As the <a href=\"https:\/\/pytorch.org\/docs\/stable\/autograd.html#torch.autograd.grad\" rel=\"nofollow noreferrer\">torch.autograd.grad<\/a> mentioned, <code>torch.autograd.grad<\/code> computes and returns the sum of gradients of outputs w.r.t. the inputs. Since your <code>c<\/code> and <code>d<\/code> are not scalar values, <code>grad_outputs<\/code> are required.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\n\na = torch.rand(2,requires_grad=True)\nb = torch.rand(2, requires_grad=True)\n\na\n# tensor([0.2308, 0.2388], requires_grad=True)\n\nb\n# tensor([0.6314, 0.7867], requires_grad=True)\n\nc = a*a + b*b\nd = 2*a+4*b\n\ntorch.autograd.grad([c,d], inputs=[a,b], grad_outputs=[torch.Tensor([1.,1.]), torch.Tensor([1.,1.])])\n# (tensor([2.4616, 2.4776]), tensor([5.2628, 5.5734]))\n<\/code><\/pre>\n\n<p>Explanation:\n<code>dc\/da = 2*a = [0.2308*2, 0.2388*2]<\/code>\n<code>dd\/da = [2.,2.]<\/code>\nSo the first output is <code>dc\/da*grad_outputs[0]+dd\/da*grad_outputs[1] = [2.4616, 2.4776]<\/code>. Same calculation for the second output. <\/p>\n\n<p>If you just want to get the gradient of <code>c<\/code> and <code>d<\/code> w.r.t. the inputs, probably you can do this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>a = torch.rand(2,requires_grad=True)\nb = torch.rand(2, requires_grad=True)\n\na\n# tensor([0.9566, 0.6066], requires_grad=True)\nb\n# tensor([0.5248, 0.4833], requires_grad=True)\n\nc = a*a + b*b\nd = 2*a+4*b\n\n[torch.autograd.grad(t, inputs=[a,b], grad_outputs=[torch.Tensor([1.,1.])]) for t in [c,d]]\n# [(tensor([1.9133, 1.2132]), tensor([1.0496, 0.9666])),\n# (tensor([2., 2.]), tensor([4., 4.]))]\n<\/code><\/pre>\n"},{"AnswerId":"58075067","CreationDate":"2019-09-24T07:22:47.640","ParentId":null,"OwnerUserId":"8119473","Title":null,"Body":"<p>Here you go\nIn the example you gave:<\/p>\n\n<pre><code>a = torch.rand(2, requires_grad=True)\nb = torch.rand(2, requires_grad=True)\nloss = a + b\n<\/code><\/pre>\n\n<p>As the loss is a vector with 2 elements, you can't perform the autograd operation at once.<\/p>\n\n<p>typically, <\/p>\n\n<pre><code>loss = torch.sum(a + b)\ntorch.autograd.grad([loss], [a, b])\n<\/code><\/pre>\n\n<p>This would return the correct value of gradient for the loss tensor which contains one element.\nYou can pass mutiple scalar tensors to outputs argument of the <code>torch.autograd.grad<\/code> method<\/p>\n"}]}
+{"QuestionId":58059685,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><tf.keras>","CreationDate":"2019-09-23T09:45:21.477","AcceptedAnswerId":null,"OwnerUserId":10150715.0,"Title":"Memory overflow during inference in tensorflow","Body":"<p>I have written these functions to carryout inference using the saved weights of a trained binary classifier. I have about 120k images to make inference. But the GPU freezes after getting to 82k images. Please is there anything I need to fix in my code to resolve this memory issue. Could the model be saving the checkpoints for the forwards pass nodes during each instance of inference? Please folks I need this fixed urgently, I have a huge number files to urgently sort.<\/p>\n\n<pre><code>def fully_frozen_MobileNetV2(inference = False, n_class=2):\n    image_size = 192\n    image_channels = 3\n    IMG_SHAPE = (image_size, image_size, image_channels)\n\n    # Create the base model from the pre-trained model MobileNet V2\n    base_model = tf.keras.applications.MobileNetV2(input_shape=IMG_SHAPE,\n                                               include_top=False,\n                                               weights='imagenet')\n\n    base_model.trainable = False\n\n    #create the top layers\n    global_average_layer = tf.keras.layers.GlobalAveragePooling2D()\n    prediction_layer = tf.keras.layers.Dense(n_class)\n\n    #add the top layers \n    model_fully_frozen = tf.keras.Sequential([\n      base_model,\n      global_average_layer,\n      prediction_layer\n    ])\n\n    if inference:\n        optimizer = tf.keras.optimizers.Adam(learning_rate = 0.001) \n   compute_loss=tf.keras.losses.CategoricalCrossentropy(from_logits=True) \n        compute_accuracy = tf.keras.metrics.CategoricalAccuracy()\n        model_fully_frozen.compile(optimizer, loss=compute_loss, metrics= \n    [compute_accuracy])\n\n    return model_fully_frozen\n\n\n\ndef read_and_preprocess_single_image_from_path(single_path, \n    image_size=192, image_channels=3):\n    #first read the image file\n    img_raw = tf.io.read_file(single_path)\n    image = tf.io.decode_jpeg(img_raw, channels=image_channels)\n    image = tf.image.resize(image, [image_size, image_size])\n    image \/= 255.0  # normalize to [0,1] range\n    return image\n\n\ndef get_path_list(path_to_image_folder):\n    data_root = pathlib.Path(path_to_image_folder)\n    #merge the folders, create a list of image paths and randomize\n    image_paths = list(data_root.glob('*\/*'))\n    image_paths = [str(path) for path in image_paths]\n    return image_paths\n\n\ndef classify_and_collect_images_with_bags(model, path_to_image_folder, \n    destination_folder, prnt_progr=True):\n    path_list = get_path_list(path_to_image_folder)\n    counter = 0\n    paths_of_images_with_bags = []\n    for path in path_list:\n        #create a single batch from the path\n        dataset = tf.data.Dataset.from_tensor_slices([path])\n        dataset = \ndataset.map(read_and_preprocess_single_image_from_path).batch(1)\n        image_class = np.argmax(model.predict(dataset))\n        print(image_class) ### test only\n        if int(image_class) &gt; 0:\n            copy_files(path, destination_folder)\n            paths_of_images_with_bags.append(path)\n        #print progress after each 1k steps if prnt_progr is true   \n        counter+=1\n        if prnt_progr and counter%1000==0:\n            print(counter)\n    return paths_of_images_with_bags\n\n\ndef copy_files(path, destination_folder):\n    #import shutil\n    shutil.copy(path, destination_folder)\n    return None\n<\/code><\/pre>\n\n<p>Make inference using images in the <code>in_folder<\/code> and copy files belonging to <code>class 1<\/code> into the <code>out_folder<\/code>.<\/p>\n\n<pre><code>classify_and_collect_images_with_bags(classifier, 'in_folder', 'out_folder')\n<\/code><\/pre>\n","answers":[{"AnswerId":"58090546","CreationDate":"2019-09-25T02:56:23.667","ParentId":null,"OwnerUserId":"10150715","Title":null,"Body":"<p>I finally found a work around for this issue. I had to break <code>classify_and_collect_image_with_bags<\/code> into these 3 simpler functions:<\/p>\n\n<pre><code>def create_inference_dataset(image_paths):\n    inference_dataset =tf.data.Dataset.from_tensor_slices(image_paths)\n    inference_dataset = \ninference_dataset.map(read_and_preprocess_single_image_from_path).batch(1)\n    return inference_dataset\n\ndef classify_single_image(model, single_batch):\n    image_class = int(np.argmax(model.predict(single_batch)))\n    return image_class\n\ndef classify_and_collect_images_with_bags(model, path_to_image_folder, \ndestination_folder, prnt_progr=True):\n    counter = 0\n    #collect image paths into a list and inference create dataset\n    path_list = get_path_list(path_to_image_folder)\n    inference_dataset = create_inference_dataset(path_list)\n    start_time = time.time()\n    for batch, path in zip(inference_dataset, path_list) :\n        image_class = classify_single_image(model, batch)\n        if image_class &gt; 0:\n            #print(path)\n            copy_files(path, destination_folder)\n        counter+=1\n        if prnt_progr and counter%1000==0:\n            duration = time.time() - start_time\n            print(counter, duration)\n<\/code><\/pre>\n\n<p>Rather than creating a new tf.dataset for each image, I created a single dataset with <code>bs = 1<\/code>and iterated through it. With this solution, inference was done one instance at a time for each batch. <\/p>\n\n<p>Aside from solving the memory issue, the inference time was also reduced from 75s\/1000 images to 20s\/1000 images.<\/p>\n\n<p>I trust there's a better solution to this issue the though, I do not believe creating multiple tf.dataset inside a loop would ground a GPU.<\/p>\n"}]}
+{"QuestionId":58060384,"AnswerCount":0,"Tags":"<tensorflow><deep-learning>","CreationDate":"2019-09-23T10:27:13.337","AcceptedAnswerId":null,"OwnerUserId":4117331.0,"Title":"Tensorflow error: Unsupported return value from function passed to Dataset.map()","Body":"<p>I'm using <code>tf.data.Dataset<\/code> for <code>input_fn<\/code> in a TensorFlow code. I need to read all channels separately as they are stored in different files. The code below show the loading and processing function.<\/p>\n\n<pre><code>def load_images_as_tensor(image_paths, dtype=np.uint8):\n    n_channels = 6\n    image_paths = image_paths\n    data = np.ndarray(shape=(512, 512, n_channels), dtype=dtype)\n    for ix, img_path in enumerate(image_paths):\n        data[:, :, ix] = load_image(img_path)\n    return(data)\n\ndef process(img, pixel_stats=GLOBAL_PIXEL_STATS, use_bfloat16 = True):\n    if pixel_stats is not None:\n        mean, std = pixel_stats\n        img = (tf.cast(img, tf.float32) - mean) \/ std\n    if use_bfloat16:\n        img = tf.image.convert_image_dtype(img, dtype=tf.bfloat16)\n\n    img = img.set_shape([512, 512, 6])\n    return(img)\n\ndef input_fn(params):\n  data = tf.data.Dataset.from_tensor_slices(tmp2)\n  data = data.map(lambda x: tf.py_func(load_images_as_tensor,[x], tf.uint8))\n  data = data.map(lambda x: process(x, GLOBAL_PIXEL_STATS, True))\n  return(data)\n<\/code><\/pre>\n\n<p>Due to the use of <code>tf.py_func<\/code>, TensorFlow is not able to get the shape of the input. After going through a TensorFlow issue on github, I thought that specifying the shape manually would fix the error. I'm still getting error <\/p>\n\n<blockquote>\n  <p>TypeError: Unsupported return value from function passed to\n  Dataset.map(): None.<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":58060620,"AnswerCount":1,"Tags":"<python><pytorch><torchvision>","CreationDate":"2019-09-23T10:41:01.007","AcceptedAnswerId":null,"OwnerUserId":8141850.0,"Title":"How to discover torch model name from the model itself?","Body":"<p>I am creating a built in torch model in this way:<br><\/p>\n\n<pre><code>m = torchvision.models.resnet50()\n<\/code><\/pre>\n\n<p>and I want to discover the model's name (resnet50 in this case) using the variable m afterwards.\nWhen calling <code>type(m)<\/code> I get <code>torchvision.models.resnet.ResNet<\/code> which is almost the answer but the depth is important to me as well. I could try to iterate <code>m.modules<\/code> (which is 151 modules long in this way) and infer the resblocks depth but that would work entirely different in other models (like inception, alexnet and so on).<\/p>\n","answers":[{"AnswerId":"58148947","CreationDate":"2019-09-28T17:55:31.057","ParentId":null,"OwnerUserId":"7872285","Title":null,"Body":"<p>ResNet50 is the name of a function if we look at the code <a href=\"https:\/\/github.com\/pytorch\/vision\/blob\/master\/torchvision\/models\/resnet.py\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>So the only way is the one suggested by Umang Gupta above : write a Wrapper model     <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = torchvision.models.resnet50()\nmodel.name = 'Resnet-50'\n<\/code><\/pre>\n"}]}
+{"QuestionId":58060885,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-09-23T10:56:19.990","AcceptedAnswerId":58061449.0,"OwnerUserId":8176285.0,"Title":"Pytorch GPU memory increase after load operation","Body":"<p>I have a pytorch model which is of the size 386MB, but when i load the model<\/p>\n\n<pre><code>state = torch.load(f, flair.device)\n<\/code><\/pre>\n\n<p>My GPU memory takes up to 900MB, why does this happen and is there a way to resolve this?<\/p>\n\n<p>This is how i save the model<\/p>\n\n<pre><code>model_state = self._get_state_dict()\n\n# additional fields for model checkpointing\nmodel_state[\"optimizer_state_dict\"] = optimizer_state\nmodel_state[\"scheduler_state_dict\"] = scheduler_state\nmodel_state[\"epoch\"] = epoch\nmodel_state[\"loss\"] = loss\n\ntorch.save(model_state, str(model_file), pickle_protocol=4)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58061449","CreationDate":"2019-09-23T11:26:56.487","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It's probably the <code>optimizer_state<\/code> that takes the extra space. Some optimizers (e.g., Adam) track statistics of each trainable parameter, such as first and second order moments. This information takes up space, as you can tell.<\/p>\n\n<p>You can load to CPU first:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>state = torch.load(f, map_location=torch.device('cpu'))\n<\/code><\/pre>\n"}]}
+{"QuestionId":58061031,"AnswerCount":1,"Tags":"<keras><deep-learning>","CreationDate":"2019-09-23T11:04:21.880","AcceptedAnswerId":58073900.0,"OwnerUserId":6144641.0,"Title":"Recommendations for implementing the architecture in Keras","Body":"<p>What is a simple way to implement the following architecture using Keras.<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/2sjZX.png\" alt=\"C2AE architecture\"><\/p>\n\n<p>There are two input encoders merged and the partial output is given to the decoder. The loss of the final layer propagated to the merged layer and the difference between the two encoders is back propagated to one encoder; the difference between the two network is added to the back-propagated output loss and passed to the other network.<\/p>\n","answers":[{"AnswerId":"58073900","CreationDate":"2019-09-24T05:59:39.370","ParentId":null,"OwnerUserId":"6144641","Title":null,"Body":"<p>I have started implementing it as follows: <\/p>\n\n<pre><code>Y=Input(shape=(100,),name='Fe_input')\ny=Dense(100, activation=LeakyReLU(alpha=0.0001))(Y)\ny=BatchNormalization()(y)\ny=Dense(60, activation=LeakyReLU(alpha=0.0001))(y)\ny=BatchNormalization()(y)\ny=Dense(20, activation=LeakyReLU(alpha=0.0001))(y)\ny=BatchNormalization()(y)\n\nX=Input(shape=(512,),name='Fx_input')\nx=Dense(100, activation=LeakyReLU(alpha=0.0001))(X)\nx=BatchNormalization()(x)\nx=Dense(60, activation=LeakyReLU(alpha=0.0001))(x)\nx=BatchNormalization()(x)\nx=Dense(20, activation=LeakyReLU(alpha=0.0001))(x)\nx=BatchNormalization()(x)\n\nmerge_layer=concatenate([x,y], name='merge_layer')\n# I need to push partial output to merge layer to the next layer. Not sure how to       \n#do that.\nx=Dense(60, activation=LeakyReLU(alpha=0.0001))(x)\nx=BatchNormalization()(x)\n\nFd=Dense(100, activation='sigmoid')(x)\nmodel=Model([X,Y],[merge_layer, Fd])\nmodel.compile(...)\n<\/code><\/pre>\n\n<p>How can I take partial output and send it across the decoder part? Moreover how can the gradient during backprop be summed and pushed to one of the branches?<\/p>\n"}]}
+{"QuestionId":58061104,"AnswerCount":2,"Tags":"<c++><tensorflow><tensorflow-lite><quantization>","CreationDate":"2019-09-23T11:08:22.250","AcceptedAnswerId":null,"OwnerUserId":6014338.0,"Title":"Dont understand mean values and std dev values when convertig .pb to ff-lite","Body":"<p>I am trying to quantizate a tensoflow graph stored in .pb. \nThe input of the network is a matrix, that each row is normalized with mean 0 and std 1.\nI want to create a tensorflow-lite model quantizate to inteference faster.\nI do not know how to pass the inputs to the line conversion. Is it a just one value? a vector with 64 values? how is it passed?<\/p>\n\n<p>The model is well converted without quantization.<\/p>\n\n<pre><code>tflite_convert \\\n    --output_file=model_simple_weight_q.tflite \\\n    --graph_def_file=model_simple.pb \\\n    --inference_type=QUANTIZED_UINT8 \\\n    --input_arrays=input \\\n    --output_arrays=LogSoftmax \\\n    --mean_values= # dont know  \\\n    --std_dev_values=# dont know\n<\/code><\/pre>\n\n<p>If I pass two single values, --mean_values=127 and --std_dev_values=128 for instance. Just to know what happens, I get the following error:<\/p>\n\n<pre><code>F tensorflow\/lite\/toco\/graph_transformations\/resolve_constant_gather.cc:108] Check failed: coords_array.data_type == ArrayDataType::kInt32 Only int32 indices are supported\nAborted (core dumped)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58107335","CreationDate":"2019-09-25T22:28:56.913","ParentId":null,"OwnerUserId":"834565","Title":null,"Body":"<p>The error you are seeing is not related to mean\/std_dev values. It is an error while converting your graph in general. It is saying that in one of your graph operations, an index is used in a <code>tf.gather<\/code> (or maybe a slicing) and it has a type that is not <code>int32<\/code>, but that's the only supported index type. <\/p>\n\n<p>Unfortunately the TFLite conversion errors will often not tell you the source code that caused the error, so you'll have to make informed guesses. <\/p>\n\n<p>Regards mean\/std_dev values, I suggest you take a look at this answer : <a href=\"https:\/\/stackoverflow.com\/a\/58096430\/834565\">https:\/\/stackoverflow.com\/a\/58096430\/834565<\/a> <\/p>\n"},{"AnswerId":"58162744","CreationDate":"2019-09-30T06:47:51.673","ParentId":null,"OwnerUserId":"1284889","Title":null,"Body":"<p>As each row of the input is normalized with mean 0 and std 1, you should give <code>--mean_values=0<\/code> and     <code>--std_dev_values=1<\/code> . Also, when you are doin a 'Dummy post traing quantization', you need to provide the minimum and maximum range for the activation functions. You need to provide --default_ranges_min= and --default_ranges_max= . <\/p>\n\n<p>As tensorflow doesn't provide very clear describtion about how to calculate the activation range, it is better to do 'post training full integer quantization' with a small calibration data- <a href=\"https:\/\/www.tensorflow.org\/lite\/performance\/post_training_integer_quant\" rel=\"nofollow noreferrer\">here<\/a> . In this methode the range of the activation function are calculated with the calibration images.<\/p>\n\n<p>Regards,<\/p>\n"}]}
+{"QuestionId":58061411,"AnswerCount":1,"Tags":"<tensorflow><keras><cudnn>","CreationDate":"2019-09-23T11:24:51.393","AcceptedAnswerId":null,"OwnerUserId":282855.0,"Title":"Tensorflow-GPU: ImportError: DLL load failed: The specified module could not be found","Body":"<p>I am facing the error below despite I have installed <code>CUDA<\/code>, <code>cuDNN<\/code>, and <code>tensorflow-gpu<\/code> package alongside adding them to the <code>PATH<\/code> when I try to run a simple <code>CNN<\/code> network. The detail of my software stack is added below:<\/p>\n\n<pre><code>TensorFlow: 1.14.0\nKeras: 2.3.0\nTensorFlow-GPU: 1.14.0\nCUDA: 10.1\ncuDNN: 7\n<\/code><\/pre>\n\n<p><strong>Full error stacktrace:<\/strong><\/p>\n\n<pre><code>Using TensorFlow backend.\nTraceback (most recent call last):\n  File \"D:\\.virtualenvs\\keras-fer-PqSTp19f\\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:\\.virtualenvs\\keras-fer-PqSTp19f\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"D:\\.virtualenvs\\keras-fer-PqSTp19f\\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 \"D:\\.virtualenvs\\keras-fer-PqSTp19f\\lib\\imp.py\", line 242, in load_module\n    return load_dynamic(name, filename, file)\n  File \"D:\\.virtualenvs\\keras-fer-PqSTp19f\\lib\\imp.py\", line 342, 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:\\Program Files\\JetBrains\\PyCharm 2018.1.2\\helpers\\pydev\\pydevd.py\", line 2066, in &lt;module&gt;\n    main()\n  File \"C:\\Program Files\\JetBrains\\PyCharm 2018.1.2\\helpers\\pydev\\pydevd.py\", line 2060, in main\n    globals = debugger.run(setup['file'], None, None, is_module)\n  File \"C:\\Program Files\\JetBrains\\PyCharm 2018.1.2\\helpers\\pydev\\pydevd.py\", line 1411, in run\n    return self._exec(is_module, entry_point_fn, module_name, file, globals, locals)\n  File \"C:\\Program Files\\JetBrains\\PyCharm 2018.1.2\\helpers\\pydev\\pydevd.py\", line 1418, in _exec\n    pydev_imports.execfile(file, globals, locals)  # execute the script\n  File \"C:\\Program Files\\JetBrains\\PyCharm 2018.1.2\\helpers\\pydev\\_pydev_imps\\_pydev_execfile.py\", line 18, in execfile\n    exec(compile(contents+\"\\n\", file, 'exec'), glob, loc)\n  File \"D:\/PyCharmProjects\/keras-fer\/fer_ck_1.py\", line 3, in &lt;module&gt;\n    from keras.models import Sequential\n  File \"D:\\.virtualenvs\\keras-fer-PqSTp19f\\lib\\site-packages\\keras\\__init__.py\", line 3, in &lt;module&gt;\n    from . import utils\n  File \"D:\\.virtualenvs\\keras-fer-PqSTp19f\\lib\\site-packages\\keras\\utils\\__init__.py\", line 6, in &lt;module&gt;\n    from . import conv_utils\n  File \"D:\\.virtualenvs\\keras-fer-PqSTp19f\\lib\\site-packages\\keras\\utils\\conv_utils.py\", line 9, in &lt;module&gt;\n    from .. import backend as K\n  File \"D:\\.virtualenvs\\keras-fer-PqSTp19f\\lib\\site-packages\\keras\\backend\\__init__.py\", line 1, in &lt;module&gt;\n    from .load_backend import epsilon\n  File \"D:\\.virtualenvs\\keras-fer-PqSTp19f\\lib\\site-packages\\keras\\backend\\load_backend.py\", line 90, in &lt;module&gt;\n    from .tensorflow_backend import *\n  File \"D:\\.virtualenvs\\keras-fer-PqSTp19f\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py\", line 5, in &lt;module&gt;\n    import tensorflow as tf\n  File \"D:\\.virtualenvs\\keras-fer-PqSTp19f\\lib\\site-packages\\tensorflow\\__init__.py\", line 28, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"D:\\.virtualenvs\\keras-fer-PqSTp19f\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"D:\\.virtualenvs\\keras-fer-PqSTp19f\\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:\\.virtualenvs\\keras-fer-PqSTp19f\\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:\\.virtualenvs\\keras-fer-PqSTp19f\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"D:\\.virtualenvs\\keras-fer-PqSTp19f\\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 \"D:\\.virtualenvs\\keras-fer-PqSTp19f\\lib\\imp.py\", line 242, in load_module\n    return load_dynamic(name, filename, file)\n  File \"D:\\.virtualenvs\\keras-fer-PqSTp19f\\lib\\imp.py\", line 342, 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\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":"58061889","CreationDate":"2019-09-23T11:53:28.813","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>You mentioned, Cuda version 10.1 which is not supported by TF-GPU yet. Please first completely remove Cuda 10.1 version and then install Cuda 10. Although, I am not sure if this will resolve your issue but at least it will not make you fall into Cuda 10.1 version usage issues. You can read the requirements <a href=\"https:\/\/www.tensorflow.org\/install\/gpu#software_requirements\" rel=\"nofollow noreferrer\">here<\/a>.\nLet us know if you are still facing the issue after installing Cuda 10.<\/p>\n"}]}
+{"QuestionId":58061500,"AnswerCount":1,"Tags":"<tensorflow><neural-network>","CreationDate":"2019-09-23T11:29:23.573","AcceptedAnswerId":null,"OwnerUserId":7879074.0,"Title":"Why my network needs so many epochs to give reasonable accuracy?","Body":"<p>I have a simple network with a few layers, I tried to reduce the layers (and the num_outputs) to speed the model, but the accuracy decreased as well\nmy question is how to speed the convergence and enhance the accuracy at the same time. Is there something wrong in the implementation that makes the network too slow <\/p>\n\n<pre><code>learning_rate=0.99\nn_x=50\nn_y=10\nlossList=[]\ng_tb = tf.Graph()\n#\nwith g_tb.as_default():\n    x = tf.placeholder(tf.float32, [None, n_x], name=\"x\")#Glove Vectors\n    y = tf.placeholder(tf.float32, [None, n_y], name=\"y\")#scores\n    with tf.name_scope('Neural_Nt'):\n        fully_connected1 = tf.contrib.layers.fully_connected(inputs=x, num_outputs=500, \n                                                         activation_fn=tf.nn.relu,scope=\"Fully_Conn1\")\n        fully_connected2 = tf.contrib.layers.fully_connected(inputs=fully_connected1, num_outputs=400, \n                                                             activation_fn=tf.nn.relu,scope=\"Fully_Conn2\")\n        fully_connected3 = tf.contrib.layers.fully_connected(inputs=fully_connected2, num_outputs=300, \n                                                             activation_fn=tf.nn.relu,scope=\"Fully_Conn3\")\n        fully_connected4 = tf.contrib.layers.fully_connected(inputs=fully_connected2, num_outputs=50, \n                                                             activation_fn=tf.nn.relu,scope=\"Fully_Conn4\")\n        fully_connected5 = tf.contrib.layers.fully_connected(inputs=fully_connected2, num_outputs=10, \n                                                             activation_fn=tf.nn.relu,scope=\"Fully_Conn5\")\n        prediction = tf.contrib.layers.fully_connected(inputs=fully_connected3, num_outputs=10, \n                                               activation_fn=tf.nn.softmax,scope=\"Out\")\n    with tf.name_scope('Cost'):\n        cost = tf.losses.softmax_cross_entropy(onehot_labels=y, logits=prediction,scope=\"Cost_Function\")\n    with tf.name_scope('Accuracy'):\n        correct_prediction = tf.equal(tf.argmax(prediction, 1, name=\"Argmax_Pred\"), tf.argmax(y, 1, name=\"Y_Pred\"), \n                                      name=\"Correct_Pred\")\n        accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32, name=\"Cast_Corr_Pred\"), name=\"Accuracy\")\n #       \n    with tf.name_scope('Optimization'):\n        optimizer = tf.train.AdagradOptimizer(learning_rate, name=\"Optimizer\").minimize(cost)\n\n    init = tf.global_variables_initializer()\n  #  \n\n\n# Start session\nwith tf.Session(graph=g_tb) as sess:\n#\n    lossList=[]\n\n    sess.run(init)\n    #Save the graph in the summary\n    #summary_writer = tf.summary.FileWriter(logs_path, graph=tf.get_default_graph())\n    training_epochs = 2000#4000\n    for epoch in range(training_epochs):\n        _, c ,prediction1,acc= sess.run([optimizer, cost,prediction,accuracy], feed_dict={x: np.array(gloveVectors),\n                                                     y: np.array(scores)})\n    # For every epoch save cost and accuracy\n        print(epoch, c)\n        lossList.append(c)\n    print(acc)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58061888","CreationDate":"2019-09-23T11:53:28.437","ParentId":null,"OwnerUserId":"10411744","Title":null,"Body":"<p>writing here my 5 cents because I can't comment.\nSo for starters you have a very high learning rate<code>learning_rate=0.99<\/code>.If you pick a high learning rate your model might learn faster but the trade off is a suboptimal set of weights or bad accuracy. You can read more about the learning rate <a href=\"https:\/\/machinelearningmastery.com\/learning-rate-for-deep-learning-neural-networks\/\" rel=\"nofollow noreferrer\">here.<\/a><\/p>\n\n<p>Another possibility to increase the learning speed is to increase the batch size. But be aware if you choose a very high batch size the numbers of weight updates decreases, so you might get also a bad accuracy score.<\/p>\n\n<p>Finally you can try <em>Dropout Layer<\/em> between your normal layers. This can decrease the learning speed, improve the accuracy and also prevent overfitting. You can read more about Dropout <a href=\"https:\/\/machinelearningmastery.com\/dropout-for-regularizing-deep-neural-networks\/\" rel=\"nofollow noreferrer\">here.<\/a><\/p>\n\n<p>As far as I can see there is nothing inherently wrong with your implementation, but to be sure you can share the data you use to check again.<\/p>\n\n<p>I hope this helps:)<\/p>\n"}]}
+{"QuestionId":58061538,"AnswerCount":1,"Tags":"<python><c++><tensorflow><mbed>","CreationDate":"2019-09-23T11:32:26.393","AcceptedAnswerId":58075104.0,"OwnerUserId":5308052.0,"Title":"uTensor model output does not equal to expected output","Body":"<p>I'm currently working on a project where <a href=\"https:\/\/github.com\/uTensor\/uTensor\" rel=\"nofollow noreferrer\">uTensor<\/a> is required. uTensor seems to work correctly, however I ran into an issue which I (apparently) cannot fix myself. <\/p>\n\n<h1>Problem definition<\/h1>\n\n<p>I've created a simple python script which generates and saves a model to a file. This saved model can later be converted into C++ code by using <a href=\"https:\/\/github.com\/mbartling\/uTensor-cli\" rel=\"nofollow noreferrer\">uTensor-cli<\/a>. The generated C++ code will be ran on a ARM dev board. <\/p>\n\n<p>Everything runs fine, no errors. However, when I create a model like: \"xW+b\" the output of the model on the devboard always equals some static value, which is not equal to the output of the model from the python script.<\/p>\n\n<p>The thing is, when a simple model like \"W+b\" is used (no input tensor used here) the output on the ARM devboard equals the output of the python script. And everything works as expected. <\/p>\n\n<h1>My findings<\/h1>\n\n<p>When using an input tensor (nothing large just a 1 dimensional array, like [1,0]) the ARM devboard always outputs some weird value compared to the output of the python script. When not using the input tensor, everything works as expected.<\/p>\n\n<h1>Other info<\/h1>\n\n<p>Because no wiki on uTensor exists yet, I've used a tutorial to learn about uTensor. The tutorial I've used can be found here: <a href=\"https:\/\/blog.hackster.io\/simple-neural-network-on-mcus-a7cbd3dc108c\" rel=\"nofollow noreferrer\">https:\/\/blog.hackster.io\/simple-neural-network-on-mcus-a7cbd3dc108c<\/a>\nThe code I've written is based on the tutorial and does not include any cost\/loss function and is not capable of 'learning' anything. The code is just for debugging.<\/p>\n\n<h1>The question<\/h1>\n\n<p>What is the reason an input tensor makes the application output unexpected values? How can I possibly fix this?<\/p>\n\n<h1>Code and output<\/h1>\n\n<p><strong>Python script<\/strong><\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nfrom tensorflow.python.framework import graph_util as gu\n\ndef weightVariable(shape, name):\n    initial = tf.truncated_normal(shape, stddev=0.1)\n    return tf.Variable(initial, name=name)\n\ndef createLayer(layerInput, inputSize, outputSize, layerNumber, dropout = -1):\n    layerNumber = str(layerNumber)\n\n    #Define weight and bias\n    W_fc = weightVariable([inputSize, outputSize], 'W_fc' + layerNumber)\n\n    #Formula Wx+b=y\n    a_fc = tf.matmul(layerInput, W_fc, name='y_pred' + layerNumber)\n    return a_fc\n\ndef saveGraph(saver, sess, y_pred):\n    outNodes = [y_pred.op.name]\n    subGraphDef = gu.remove_training_nodes(sess.graph_def)\n    subGraphDef = gu.convert_variables_to_constants(sess, subGraphDef, outNodes)\n\n    #Save the checkpoint\n    ckptPath = saver.save(sess, \".\/chkps\/model.ckpt\")\n\n    #Save the graph\n    graphPath = tf.train.write_graph(subGraphDef, \".\/graph\", \"mlp.pb\", as_text=False)\n\n    #Print some usefull messages\n    print(\"Saved checkpoint to: \" + ckptPath)\n    print(\"Saved graph to: \" + graphPath)\n    print(\"Output tensor: \" + y_pred.op.name)\n\n    def restoreGraph(saver, sess):\n    tf.reset_default_graph()\n    saver.restore(sess, \".\/chkps\/model.ckpt\")\n\ndef main():\n    data = [\n        [0,0],\n        [0,1],\n        [1,0],\n        [1,1]\n    ]\n    labels = [\n        [0],\n        [1],\n        [1],\n        [0]\n    ]\n    inputSize = 2\n    outputSize = 1\n\n    #Placeholders for the input and output\n    x_input = tf.placeholder(tf.float32, [None, inputSize], name='x_input')\n    y_output = tf.placeholder(tf.float32, [None, outputSize], name='y_output')\n\n    #Layers with relu activation\n    inputLayer = createLayer(x_input, inputSize, outputSize, 0)\n\n    #Start a session\n    sess = tf.Session()\n    saver = tf.train.Saver()\n\n    #Run the session\n    sess.run(tf.global_variables_initializer())\n    feed_dict = {x_input: data, y_output: labels}\n    sess.run(inputLayer, feed_dict=feed_dict)\n\n    #Save the graph\n    saveGraph(saver, sess, inputLayer)\n\n    #Test the algorithm\n    for i in range(0,4):\n        testInput = [data[i]]\n        output = sess.run(inputLayer, feed_dict={x_input: testInput})[0][0]\n        print(\"Test output \" + str(i) + \": \" + str(output))\n\n    #End the session\n    sess.close()\n\n#Execute the main function\nmain()\n<\/code><\/pre>\n\n<p><strong>Output<\/strong><\/p>\n\n<pre><code>Saved checkpoint to: .\/chkps\/model.ckpt\nSaved graph to: .\/graph\/mlp.pb\nOutput tensor: y_pred0\nTest output 0: 0.0\nTest output 1: 0.0034507334\nTest output 2: 0.07698402\nTest output 3: 0.080434754\n<\/code><\/pre>\n\n<p><strong>Converting to C++<\/strong><\/p>\n\n<pre><code>utensor-cli convert graph\/mlp.pb --output-nodes=y_pred0\n<\/code><\/pre>\n\n<p><strong>C++ code<\/strong><\/p>\n\n<pre><code>#include \"models\/mlp.hpp\"\n#include \"tensor.hpp\"\n#include \"mbed.h\"\n#include &lt;stdio.h&gt;\n\nconst int testData[4][2] = {{0,0},{0,1},{1,0},{1,1}};\n\nSerial uart(USBTX, USBRX, 115200);\n\nint main(void){\n    printf(\"Compiled at: \");\n    printf(__TIME__);\n    printf(\"\\n\");\n\n    for(int i = 0; i &lt; 4; i++){\n        \/\/Create the context class. \n        Context ctx;\n        Tensor *input_x = new WrappedRamTensor&lt;int&gt;({1, 2}, (int*) testData[i]);\n\n        get_mlp_ctx(ctx, input_x);          \/\/Pass the tensor to the context\n        S_TENSOR pred_tensor = ctx.get(\"y_pred0:0\");    \/\/Get the output tensor\n        ctx.eval();                 \/\/Trigger the inference\n\n        float prediction = *(pred_tensor-&gt;read&lt;float&gt;(0,0));    \/\/Get the result\n        printf(\"Test output %d: %f \\r\\n\", i,  prediction);  \/\/Print the result\n\n    }\n\n    printf(\"\\n\");\n    return 0;\n}\n<\/code><\/pre>\n\n<p><strong>Serial output<\/strong><\/p>\n\n<pre><code>Compiled at: (Compile time)\nTest output 0: 0.000000 \nTest output 1: 0.000000 \nTest output 2: 0.000000 \nTest output 3: 0.000000 \n<\/code><\/pre>\n","answers":[{"AnswerId":"58075104","CreationDate":"2019-09-24T07:24:53.983","ParentId":null,"OwnerUserId":"5308052","Title":null,"Body":"<p>Changing the datatype in the C++ code to float did the trick! I'm still not quite sure how I did not think about this.<\/p>\n"}]}
+{"QuestionId":58061775,"AnswerCount":2,"Tags":"<tensorflow><word><embedding>","CreationDate":"2019-09-23T11:46:49.643","AcceptedAnswerId":null,"OwnerUserId":11529280.0,"Title":"How to Extract Features from Text based on Fine-Tuned BERT Model","Body":"<p>I am trying to make a binary predictor on some data which has one columns with text and some additional columns with numerical values. My first solution was to use word2vec on the text to extract 30 features and use them with the other values in a Random Forest. It produces good result. I am interested in improving the TEXT to FEATURE model.<\/p>\n\n<p>I then wanted to improve the feature extraction algorithm by using BERT. I managed to implement a pre-trained BERT model for feature extraction with some improvement to the word2vec.<\/p>\n\n<p>Now I want to know, how can i fine-tune the BERT model on my data - to improve the feature extraction model - to get better text-to-features for my Random Forest algorithm. I know how to fine-tune BERT for a binary predictor (BertForSequenceClassification), but not how to fine-tune it for a making a better BERT text-to-feature extraction model. Can I use the layers in the BertForSequenceClassification somehow?? I spent 2 days trying to find a solution, but did not manage so far...<\/p>\n\n<p>Kind Regards, Peter<\/p>\n","answers":[{"AnswerId":"58141349","CreationDate":"2019-09-27T21:07:59.107","ParentId":null,"OwnerUserId":"5661543","Title":null,"Body":"<p>I am dealing with this problem too. As far I know, you must fine-tune the BERT language model; according to <a href=\"https:\/\/github.com\/google-research\/bert\/issues\/145\" rel=\"nofollow noreferrer\">this issue<\/a>, <a href=\"https:\/\/github.com\/google-research\/bert#pre-training-with-bert\" rel=\"nofollow noreferrer\">masked LM<\/a> is suggested. Then you can use <a href=\"https:\/\/bert-as-service.readthedocs.io\/en\/latest\/section\/faq.html#can-i-use-my-own-fine-tuned-bert-model\" rel=\"nofollow noreferrer\">Bert-as-service<\/a> to extract the features. Note that I haven't tested it yet, but I am going to. I thought it would be good to share it with you :)<\/p>\n"},{"AnswerId":"58187080","CreationDate":"2019-10-01T14:32:08.873","ParentId":null,"OwnerUserId":"8893595","Title":null,"Body":"<p>Yes, you can fine-tune BERT, and then extract the features. I have done it, but it really did not yield a good improvement. By fine-tuning and then extracting the text features, the text features are slightly adapted to your custom training data. It can still be done in 2 ways. Use fine-tuning and then extract the features, or you can continue pre-training from BERT's checkpoint (apply MLM over your custom data), and then followed by feature extraction. <\/p>\n"}]}
+{"QuestionId":58062040,"AnswerCount":0,"Tags":"<rest><tensorflow><tensorflow-serving>","CreationDate":"2019-09-23T12:03:53.820","AcceptedAnswerId":null,"OwnerUserId":9918667.0,"Title":"Sending batch of mixed data returns only one prediction","Body":"<p>I would like to send a batch of mixed data on a tensorflow model. I am using the API REST of <code>tensorflow serving<\/code> and it seems to do not work as expected.<\/p>\n\n<p>My POST request is:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>$ curl -XPOST -H 'Content-type:application.json' --data '{\"instances\": [\n    {\n        \"distances\": [0.015810276679841896,... ,0.011857707509881422],\n        \"settings\": [50,1,250]\n   },\n   {\n        \"distances\": [0.255810276679841896,... ,0.320857707509881422],\n        \"settings\": [50,1,250]\n   },\n   ]}' http:\/\/localhost:8501\/v1\/models\/mymodel:predict\n<\/code><\/pre>\n\n<p>and it returns:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>{\n    \"predictions\": [0.999931693, 6.81120073e-05, 2.74769889e-07]\n}\n<\/code><\/pre>\n\n<p>I am expecting something like (i.e as many responses as I send objects):<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>{\n    \"predictions\": [\n                    [0.999931693, 6.81120073e-05, 2.74769889e-07],\n                    [6.81120073e-05, 0.988888888, 2.74769889e-07]\n                 ]\n}\n<\/code><\/pre>\n\n<p>Here, I built a model which is waiting for mixed data, i.e an object containing two attributes: <code>distances<\/code> and <code>settings<\/code>. But I get only prediction (an array of 3 floats) where I send more than one object.<\/p>\n\n<hr>\n\n<p>Note: This kind of batch request works where the model waits for only one type of data (ex: an array of floats).<\/p>\n\n<p>Example:<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code>curl -XPOST -H 'Content-type:application.json' --data '{\"instances\":[\n    [0.015810276679841896,... ,0.011857707509881422],\n    [0.015810276679841896,... ,0.011857707509881422],\n    [0.015810276679841896,... ,0.011857707509881422]\n   ]}' http:\/\/localhost:8501\/v1\/models\/mymodel:predict\n\n{\n    \"predictions\": [\n                    [0.999931693, 6.81120073e-05, 2.74769889e-07],\n                    [6.81120073e-05, 0.988888888, 2.74769889e-07],\n                    [6.81120073e-05, 0.988888888, 2.74769889e-07]\n                 ]\n}\n<\/code><\/pre>\n\n<hr>\n\n<p>Here is my model architecture (with <code>keras<\/code> framework): <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    sess = tf.Session()\n    init_op = tf.initialize_all_variables()\n    sess.run(init_op)\n    K.set_session(sess)\n    K.set_learning_phase(0)\n\n    # defines two sets of inputs\n    input_distances = Input(shape=(100, ))\n    input_settings = Input(shape=(3, ))\n\n    # distances is the first branch which handles arrays of distances\n    distances = Dense(100,\n                      activation='relu',\n                      kernel_initializer='random_uniform')(input_distances)\n    distances = Dense(50,\n                      activation='relu',\n                      kernel_initializer='random_uniform')(distances)\n    distances = Model(inputs=input_distances, outputs=distances)\n\n    # settings is the second branch which handles arrays of distances\n    settings = Dense(3, activation='relu',\n                     kernel_initializer='random_uniform')(input_settings)\n    settings = Dense(6, activation='relu',\n                     kernel_initializer='random_uniform')(settings)\n    settings = Model(inputs=input_settings, outputs=settings)\n\n    # combine the output of the two branches\n    combined = concatenate([distances.output, settings.output])\n\n    result = Dense(2, activation='relu')(combined)\n    result = Dense(len(labels), activation='softmax')(result)\n\n    model = Model(inputs=[distances.input, settings.input], outputs=result)\n\n    opt = Adam(lr=1e-3, decay=1e-3 \/ 200)\n    model.compile(optimizer=opt,\n                  loss='sparse_categorical_crossentropy',\n                  metrics=['accuracy'])\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58062924,"AnswerCount":1,"Tags":"<keras><neural-network><conv-neural-network><autoencoder><deconvolution>","CreationDate":"2019-09-23T12:53:54.220","AcceptedAnswerId":58063176.0,"OwnerUserId":1252281.0,"Title":"How to train non-shared autoencoder networks parallelly using single loss function","Body":"<p>Train two non-shared neural networks together.<\/p>\n\n<p>I have two different auto-encoder networks, one for camera images, which takes an input image and try to reproduce the same image by using the transpose convolutional method. The second network takes a sketch image and does the same thing. Both networks have identical structures but learn different filters. <\/p>\n\n<pre><code>autoencoder_p = Model(input_img, decoder_p(encoder_p(input_img)))\nautoencoder_s = Model(input_img, decoder_s(encoder_s(input_img)))\n\nautoencoder_p.compile(optimizer='adam', loss='mse', metrics=['accuracy'])\nautoencoder_p.fit(X_train_real, X_train_real, epochs=10, batch_size=1, validation_split=0.1)\n\nautoencoder_s.compile(optimizer='adam', loss='mse', metrics=['accuracy'])\nautoencoder_s.fit(X_train_sim, X_train_sim, epochs=10, batch_size=1, validation_split=0.1)\n<\/code><\/pre>\n\n<p>The problem is I have to compile and train both networks one after another. Is there any possible way to compile and train both networks together. Because both networks are identical and use the same loss and training strategy.<\/p>\n","answers":[{"AnswerId":"58063176","CreationDate":"2019-09-23T13:06:21.240","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Make a single model that contains both.<\/p>\n\n<blockquote>\n  <p>Warning: your code seems to be using the same input image for both, which is not what you describe. Have two inputs <code>input_img_p<\/code> and <code>input_img_s<\/code>.   <\/p>\n<\/blockquote>\n\n<pre><code>twin_model = Model([input_img_p, input_img,_s],\n                   [\n                       decoder_p(encoder_p(input_img_p)),\n                       decoder_s(encoder_p(input_img_s)),                             \n                   ])\n\ntwin_model.compile(optimizer='adam', loss='mse', metrics=['accuracy'])\ntwin_model.fit([X_train_real, X_train_sim], [X_train_real, X_train_sim], ...)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58063749,"AnswerCount":1,"Tags":"<python><tensorflow><keras><layer><categorical-data>","CreationDate":"2019-09-23T13:37:26.903","AcceptedAnswerId":null,"OwnerUserId":11903541.0,"Title":"Functional API with mixed input data. Got this error: AttributeError: 'tuple' object has no attribute 'ndim'","Body":"<p>I have read and search the similar question but the answers did not work for me even If I try it in several ways. <\/p>\n\n<p>I have a data set that is this way \n4.000.000 different string representing a Cell (categorical data)\nand 4.000.000 longitude, latitude, representing where those cells are\nhere is a snippet of the dataset ( numerical data)<\/p>\n\n<pre><code>\ndevice_id                    seconds      latitude     longitude\n\njg4M\/taYRc2cBJPGa8c7vw==       752        53.392060    -2.069796\n\n<\/code><\/pre>\n\n<p>I create this model it works but i have problems when I try to fit it.\nThe model has a set of layers for categorical data and a separate set of layers for numerical data. And I merge them. \nAll work fine until the fit part\nHere is the code.<\/p>\n\n<pre><code>\ndevice_id = Input(name = 'data_tf', shape = [1,1])\nvisible = Input(name = 'X_train', shape=(3,1,1))\n\n#this is the categorical layers  \ntop_words = 10376\nembedding_vector_length = 22\nx = Embedding(top_words, embedding_vector_length)(device_id)\nx = Dense(2, activation='sigmoid')(x)\nmodelx = Model(inputs=device_id, outputs = x)\n\n\n#this are the numerical layers\nhidden1 = Dense(10, activation='relu')(visible)\nhidden2 = Dense(20, activation='relu')(hidden1)\nhidden3 = Dense(10, activation='relu')(hidden2)\noutput = Dense(2, activation='sigmoid')(hidden3)\n\nmodely = Model(inputs=visible, outputs=output)\n\n#merge the two set of layers and all this works perfectly \nmerge = concatenate([x, output], axis =1)\n\n#layers after concatenate \n\nhidden1 = Dense(10, activation='relu')(merge)\nhidden2 = Dense(10, activation='relu')(hidden1)\noutput = Dense(2, activation='sigmoid')(hidden2)\nmodel = Model(inputs=[device_id,visible], outputs=output)\n\n<\/code><\/pre>\n\n<p>here is a summary of the model:<\/p>\n\n<pre><code>\nLayer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\nX_train (InputLayer)            (None, 3, 1, 1)      0                                            \n__________________________________________________________________________________________________\ndense_2 (Dense)                 (None, 3, 1, 10)     20          X_train[0][0]                    \n__________________________________________________________________________________________________\ndata_tf (InputLayer)            (None, 1, 1)         0                                            \n__________________________________________________________________________________________________\ndense_3 (Dense)                 (None, 3, 1, 20)     220         dense_2[0][0]                    \n__________________________________________________________________________________________________\nembedding_1 (Embedding)         (None, 1, 1, 22)     228272      data_tf[0][0]                    \n__________________________________________________________________________________________________\ndense_4 (Dense)                 (None, 3, 1, 10)     210         dense_3[0][0]                    \n__________________________________________________________________________________________________\ndense_1 (Dense)                 (None, 1, 1, 2)      46          embedding_1[0][0]                \n__________________________________________________________________________________________________\ndense_5 (Dense)                 (None, 3, 1, 2)      22          dense_4[0][0]                    \n__________________________________________________________________________________________________\nconcatenate_1 (Concatenate)     (None, 4, 1, 2)      0           dense_1[0][0]                    \n                                                                 dense_5[0][0]                    \n__________________________________________________________________________________________________\ndense_6 (Dense)                 (None, 4, 1, 10)     30          concatenate_1[0][0]              \n__________________________________________________________________________________________________\ndense_7 (Dense)                 (None, 4, 1, 10)     110         dense_6[0][0]                    \n__________________________________________________________________________________________________\ndense_8 (Dense)                 (None, 4, 1, 2)      22          dense_7[0][0]                    \n==================================================================================================\nTotal params: 228,952\nTrainable params: 228,952\nNon-trainable params: 0\n__________________________________________________________________________________________________\nNone\n<\/code><\/pre>\n\n<p>then <\/p>\n\n<pre><code>model.compile(loss='binary_crossentropy',optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>but this is the part that does not work:<\/p>\n\n<pre><code>model.fit((visible,device_id),X_trainY, batch_size=10, epochs=5, verbose=1)\n\n\nAttributeError: 'tuple' object has no attribute 'ndim'\n\n<\/code><\/pre>\n\n<p>I have search for 4 days now similar question but none of them work for me.<\/p>\n\n<p>Can some one help?<\/p>\n\n<p>UPDATE \nI put the brackets as suggest below gently by Rishabh Sahrawat and have now this problem<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nValueError                                Traceback (most recent call last)\n&lt;ipython-input-53-9b35f717e3e7&gt; in &lt;module&gt;\n----&gt; 1 model.fit([visible,device_id],X_trainY, batch_size=10, epochs=5, verbose=1)\n\n\/\/anaconda3\/lib\/python3.7\/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\/lib\/python3.7\/site-packages\/keras\/engine\/training.py in _standardize_user_data(self, x, y, sample_weight, class_weight, check_array_lengths, batch_size)\n    749             feed_input_shapes,\n    750             check_batch_axis=False,  # Don't enforce the batch size.\n--&gt; 751             exception_prefix='input')\n    752 \n    753         if y is not None:\n\n\/\/anaconda3\/lib\/python3.7\/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\n\/\/anaconda3\/lib\/python3.7\/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\n\/\/anaconda3\/lib\/python3.7\/site-packages\/keras\/engine\/training_utils.py in standardize_single_array(x)\n     23                 'When feeding symbolic tensors to a model, we expect the'\n     24                 'tensors to have a static batch size. '\n---&gt; 25                 'Got tensor with shape: %s' % str(shape))\n     26         return x\n     27     elif x.ndim == 1:\n\nValueError: When feeding symbolic tensors to a model, we expect thetensors to have a static batch size. Got tensor with shape: (None, 3, 1, 1)\n\n<\/code><\/pre>\n","answers":[{"AnswerId":"58064077","CreationDate":"2019-09-23T13:55:13.870","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>Since you did not paste the full error log, from what you shared I think the problem is with the round brackets you used to input the data(categorical , numerical). \nTry with square brackets like <code>model.fit([visible,device_id], X_trainY, batch_size=10, epochs=5, verbose=1)<\/code><\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>In your definition of <code>device_id<\/code> the shape parameter should be with round brackets instead of square brackets. Or, maybe there is a reason you can explain.<\/p>\n\n<p>The reason behind the new error in your EDIT might be from using the layer name in the <code>model.fit<\/code> instead of feeding the actual data. Look <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">here<\/a> (skip to <strong>Shared layers<\/strong>) how and what to feed to shared layers model. Also, have a look at the similar <a href=\"https:\/\/stackoverflow.com\/questions\/51254382\/valueerror-when-feeding-symbolic-tensors-to-a-model-we-expect-the-tensors-to-h\">issue<\/a>. <\/p>\n\n<p>Another thing, I found in your code is here:\n<code>model = Model(inputs=[device_id,visible], outputs=output)<\/code>, here first, the model expects <code>device_id<\/code> data and then <code>visible<\/code> data but in your original code in <code>fit<\/code> you are feeding it the other way around. <code>model.fit((visible,device_id),X_trainY, batch_size=10, epochs=5, verbose=1)<\/code> and here <code>visible<\/code> and <code>device_id<\/code> must be actual data not layer names.<\/p>\n"}]}
+{"QuestionId":58063760,"AnswerCount":0,"Tags":"<python-3.x><keras><text-classification><embedding>","CreationDate":"2019-09-23T13:37:46.693","AcceptedAnswerId":null,"OwnerUserId":12107280.0,"Title":"keras embedding layer cannot be trained with vectors from googlenews word2vec","Body":"<p>I aim to do text classification, using keras in python, on plenty of positive and negative comments about some movies. I've done word2vec process on each comment, using Googlenews word2vec, and now I have some vectors with values between -1 to 1.I need to pass these vectors into a network which has an embedding layer as its first layer. What I realized is the embedding layer is unable to process the vectors as it givs a specific output in model.predict, for each input vector, which means it's not been trained at all as it does not matter what we have in the input.<\/p>\n\n<p>As I was completely confused with the results, I made some random input vectors of the size 300 (the same size of googlenews vectors), and built a simple network including an embedding layer, a flatten layer and a dense layer to examine the output of network after predicting. What I realized was if I change the domain of input vectors to, for example 0 to 2, the embedding layer will be traind well and the output of model.predict will be accurate enough. Here is a part of the code (p1 is the number of vectors and p2=300):<\/p>\n\n<pre><code>model = Sequential()\nembedding_layer = Embedding(input_dim=p1*p2, output_dim=5, input_length=p2, trainable=False, batch_input_shape=(None, p2))\nmodel.add(embedding_layer)\nmodel.add(Flatten())\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(optimizer='adam', loss='binary_crossentropy', metrics=['acc'])\n\nhistory = model.fit(x_train, y_train, epochs=100 , validation_data=(x_test,y_test), validation_split=0.3)\ntesting = model.predict(x_test)\n<\/code><\/pre>\n\n<p>I expect the output to has at least different values, while I get the same output for the all input vectors with valuee between 0 and 1(which are the result of word2vec).Does anyone have any idea about what's happening in embedding layer?<\/p>\n","answers":[]}
+{"QuestionId":58063826,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><pytorch>","CreationDate":"2019-09-23T13:41:36.280","AcceptedAnswerId":58065755.0,"OwnerUserId":8751092.0,"Title":"Compute cross entropy loss for classification in pytorch","Body":"<p>I am trying to build two neural network for classification. One for Binary and the second is for multi-class classification. I am trying to use the <code>torch.nn.CrossEntropyLoss()<\/code> as a loss function, but I try to train my first neural network I get the following error:<\/p>\n\n<pre><code>multi-target not supported at \/opt\/conda\/conda-bld\/pytorch_1565272271120\/work\/aten\/src\/THNN\/generic\/ClassNLLCriterion.c:22\n<\/code><\/pre>\n\n<p>From my analysis, I found that the my dataset has two problems that caused the error.<\/p>\n\n<ul>\n<li><strong>My data set is one hot encoded<\/strong>. I used one hot encoding to pre processes my dataset. The first target <code>Y_binary<\/code> variable has the shape of <code>torch.Size([125973, 1])<\/code> full of 0s and 1 indicating classes 'No' and 'Yes'.<\/li>\n<li><strong>My data has the wrong dimensions?<\/strong> I found that I can't use a simple vector with the cross entropy loss function. Some people used the following code to reshape their target vector before feeding to the loss function.<\/li>\n<\/ul>\n\n<p><code>out = out.permute(0, 2, 3, 1).contiguous().view(-1, class_number)<\/code> <\/p>\n\n<p>But I didn't really understand the reasoning behind this code. But it seems for my that I need to keep track of the following variables: <code>Class_Number<\/code>, <code>Batch_size<\/code>, <code>Dimension_Output<\/code>. For my code here are the dimensions <\/p>\n\n<pre><code>X_train.shape: (125973, 122)\nY_train2.shape: (125973, 1)\nbatch_size = 64\nK = len(set(Y_train2)) # Binary classification For multi class classification use K = len(set(Y_train5))\n<\/code><\/pre>\n\n<ul>\n<li>Should the target value be one hot encoded? If not, how I can feed a nominal feature to the loss function?<\/li>\n<li>If I use reshape the output, can you help me do this for my code ?<\/li>\n<\/ul>\n\n<p>I am trying to use this loss function for both my neural networks. <\/p>\n\n<p>Thank you in advance,<\/p>\n","answers":[{"AnswerId":"58065755","CreationDate":"2019-09-23T15:33:07.153","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p>The error is due to the usage of <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html?highlight=crossentropyloss#torch.nn.CrossEntropyLoss\" rel=\"nofollow noreferrer\">torch.nn.CrossEntropyLoss()<\/a> which can be used if you want to predict 1 class out of N classes. For multiclass classification, you should use <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.BCEWithLogitsLoss\" rel=\"nofollow noreferrer\">torch.nn.BCEWithLogitsLoss()<\/a> which combines a <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html?highlight=sigmoid#torch.nn.Sigmoid\" rel=\"nofollow noreferrer\">Sigmoid<\/a> layer and the <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html?highlight=bceloss#torch.nn.BCELoss\" rel=\"nofollow noreferrer\">BCELoss<\/a> in one single class. <\/p>\n\n<p>In case of multi-class, and if you use Sigmoid + BCELoss, then you need the target to be one-hot encoding, i.e. something like this per sample: <code>[0 1 0 0 0 1 0 0 1 0]<\/code>, where 1 will be at the locations of classes present.<\/p>\n"}]}
+{"QuestionId":58063921,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-23T13:46:37.010","AcceptedAnswerId":null,"OwnerUserId":11910970.0,"Title":"Keras Text Generator keeps repeating the same words","Body":"<p>I created a very simple Keras model for generating poems. I trained it over 300 epochs and it had a 0.99 accuracy but when I generate new text mostly it keeps repeating the same words over and over. What can I improve on the model at training or predicting?<\/p>\n\n<p>For example how long would be the optimal seq_len if my poems look like this:<\/p>\n\n<blockquote>\n  <p>\u201cif snow fell, it\u2019d fall black. please, don\u2019t call \\n\n   us dead, call us alive someplace better. \\n\n   we say our own names when we pray. \\n\n   we go out for sweets &amp; come back.\u201d<\/p>\n<\/blockquote>\n\n<p><strong>CREATING THE MODEL<\/strong><\/p>\n\n<pre><code>def create_text_sequences(tokens):\n    train_len = 25 + 1\n    text_sequences = []\n    for i in range(train_len, len(tokens)):\n        seq = tokens[i-train_len:i]\n        text_sequences.append(seq)\n    return text_sequences\n\ndef create_model(voc_size,seq_len):\n    SCHEMA = [\n        Embedding(voc_size, seq_len,input_length=seq_len),\n        LSTM(units=seq_len*2,return_sequences=True),\n        LSTM(units=seq_len),\n        Dense(50, activation='relu'),\n        Dense(voc_size,activation=\"softmax\")\n    ]\n    model = Sequential(SCHEMA)\n    model.compile(loss=\"categorical_crossentropy\", optimizer=\"adam\", metrics=[\"accuracy\"])\n    model.summary()\n    return model\n<\/code><\/pre>\n\n<p><strong>PREDICTING<\/strong><\/p>\n\n<pre><code>def generate_text(seq_len=15,num_words=10):\n    d = read_file(f\"poems.txt\")\n    tokens = separate_punc(d)\n    text_sequences = create_text_sequences(tokens)\n    model = load_model(f\"model_300.h5\") \n    tokenizer = load(open(f\"tokenizer_300\",\"rb\")) \n    output_text = []\n\n    random_pick = np.random.randint(0,len(text_sequences))\n    random_seed_text = text_sequences[random_pick]\n    seed_text = \" \".join(random_seed_text)\n    input_text = seed_text\n\n    for i in range(num_words):\n        encoded_text = tokenizer.texts_to_sequences([input_text])[0]\n        pad_encoding = pad_sequences([encoded_text], seq_len,truncating='pre')\n        pre_word_ind = model.predict_classes(pad_encoding,verbose=0)[0]\n        pre_word = tokenizer.index_word[pre_word_ind]\n        input_text = ' ' + pre_word\n        output_text.append(input_text)\n    return \" \".join(output_text)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58063936,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-23T13:47:29.017","AcceptedAnswerId":null,"OwnerUserId":2467772.0,"Title":"`Checkpoint` was expecting a trackable object (an object derived from `TrackableBase`), got Tensor","Body":"<p>I have a 10-layer model as follow.<\/p>\n\n<pre><code>def detectionmodel(self):\n    params = self.params\n    #layer 1\n    x = Input((3,int(params[\"input_size\"][0]),int(params[\"input_size\"][1])),int(params[\"batch_size\"]),dtype=params[\"dtype\"])\n    x = Conv2D(48,(5,5),strides=(2,2),padding='same',data_format='channels_first')(x)\n    x = BatchNormalization(trainable=True)(x)\n    x = Activation('relu')(x)\n    x = MaxPooling2D(pool_size=(2,2),strides=(2,2),padding='same',data_format='channels_first')(x)\n    x = Dropout(0.2)(x)\n    #layer 2\n    x = Conv2D(64,(5,5),strides=(1,1),padding='same',data_format='channels_first')(x)\n    x = BatchNormalization(trainable=True)(x)\n    x = Activation('relu')(x)\n    x = MaxPooling2D(pool_size=(2,2),strides=(1,1),padding='same',data_format='channels_first')(x)\n    x = Dropout(0.2)(x)\n    #layer 3\n    x = Conv2D(128,(5,5),strides=(1,1),padding='same',data_format='channels_first')(x)\n    x = BatchNormalization(trainable=True)(x)\n    x = Activation('relu')(x)\n    x = MaxPooling2D(pool_size=(2,2),strides=(2,2),padding='same',data_format='channels_first')(x)\n    x = Dropout(0.2)(x)\n    #layer 4\n    x = Conv2D(160,(5,5),strides=(1,1),padding='same',data_format='channels_first')(x)\n    x = BatchNormalization(trainable=True)(x)\n    x = Activation('relu')(x)\n    x = MaxPooling2D(pool_size=(2,2),strides=(1,1),padding='same',data_format='channels_first')(x)\n    x = Dropout(0.2)(x)\n    #layer 5\n    x = Conv2D(192,(5,5),strides=(1,1),padding='same',data_format='channels_first')(x)\n    x = BatchNormalization(trainable=True)(x)\n    x = Activation('relu')(x)\n    x = MaxPooling2D(pool_size=(2,2),strides=(2,2),padding='same',data_format='channels_first')(x)\n    x = Dropout(0.2)(x)\n    #layer 6\n    x = Conv2D(192,(5,5),strides=(1,1),padding='same',data_format='channels_first')(x)\n    x = BatchNormalization(trainable=True)(x)\n    x = Activation('relu')(x)\n    x = MaxPooling2D(pool_size=(2,2),strides=(1,1),padding='same',data_format='channels_first')(x)\n    x = Dropout(0.2)(x)\n    #layer 7\n    x = Conv2D(192,(5,5),strides=(1,1),padding='same',data_format='channels_first')(x)\n    x = BatchNormalization(trainable=True)(x)\n    x = Activation('relu')(x)\n    x = MaxPooling2D(pool_size=(2,2),strides=(2,2),padding='same',data_format='channels_first')(x)\n    x = Dropout(0.2)(x)\n    #layer 8\n    x = Conv2D(192,(5,5),strides=(1,1),padding='same',data_format='channels_first')(x)\n    x = BatchNormalization(trainable=True)(x)\n    x = Activation('relu')(x)\n    x = MaxPooling2D(pool_size=(2,2),strides=(1,1),padding='same',data_format='channels_first')(x)\n    x = Dropout(0.2)(x)\n    #layer 9\n    x = Conv2D(192,(5,5),strides=(1,1),padding='same',data_format='channels_first')(x)\n    x = BatchNormalization(trainable=True)(x)\n    x = Activation('relu')(x)\n    x = MaxPooling2D(pool_size=(2,2),strides=(2,2),padding='same',data_format='channels_first')(x)\n    x = Dropout(0.2)(x)\n     #layer 10\n    x = Dense(100,activation='linear',name='dense_1')(x)\n    x = Dense(100,activation='linear',name='dense_2')(x)\n    x = Dense(self.numClasses,activation='linear',name='dense_3')(x)\n    return x\n\nwith distribution_utils.get_strategy_scope(self.distribution_strategy):\n      model = self.detectionmodel()\n      opt = self._create_optimizer()\n      current_step = 0\n      checkpoint = tf.train.Checkpoint(model=model, optimizer=opt)\n<\/code><\/pre>\n\n<p>When I run this line<\/p>\n\n<pre><code>checkpoint = tf.train.Checkpoint(model=model, optimizer=opt)\n<\/code><\/pre>\n\n<p>got error as<\/p>\n\n<pre><code>ValueError: `Checkpoint` was expecting a trackable object (an object derived from `TrackableBase`), got Tensor(\"dense_3\/Identity:0\", shape=(4096, 192, 4, 12), dtype=float32). If you believe this object should be trackable (i.e. it is part of the TensorFlow Python API and manages state), please open an issue.\n&gt; \/home\/nyangpu\/PlateRecognition\/PlateDetection.py(154)train()\n-&gt; checkpoint = tf.train.Checkpoint(model=model, optimizer=opt)\n<\/code><\/pre>\n\n<p>What could be wrong?<\/p>\n","answers":[]}
+{"QuestionId":58064149,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2019-09-23T13:58:29.340","AcceptedAnswerId":null,"OwnerUserId":12107637.0,"Title":"In Keras can target be independent of batch size?","Body":"<p>The input to my model is sentences fed as (N,len), where N is the batch size of sentences. I have two loss functions in my model where for the first one I feed target (N,2), but for the second one as I want to get a document representation of shape (d,) from the input sentences embeddings, I want to feed target (d,). However, when I do this it shows me error that I'm expected to feed target (N,d). Is it possible to feed target independent of batch dimension in Keras? <\/p>\n","answers":[]}
+{"QuestionId":58064264,"AnswerCount":2,"Tags":"<keras>","CreationDate":"2019-09-23T14:04:59.780","AcceptedAnswerId":58064676.0,"OwnerUserId":6761328.0,"Title":"How does Keras load_data() know what part of the data is the train and test set?","Body":"<p>I'm quite new to Keras and I wanted to begin with a <a href=\"https:\/\/elitedatascience.com\/keras-tutorial-deep-learning-in-python\" rel=\"nofollow noreferrer\">tutorial<\/a>. There, let's say almost at the beginning, the code lines<\/p>\n\n<blockquote>\n  <p>Load pre-shuffled MNIST data into train and test sets <\/p>\n  \n  <p>(X_train, y_train), (X_test, y_test) = mnist.load_data()<\/p>\n<\/blockquote>\n\n<p>emerge. I wonder how Keras knows what of the data is part of the training and what is part of the testing? Though it's quite a basic question I'm not able to see the certain definition in the Keras documentation (the searching even does not provide any result there).\nTherefore, I appreciate any help as I often cannot find any command definitions in <a href=\"http:\/\/faroit.com\/keras-docs\/1.2.0\/search.html?q=load_data\" rel=\"nofollow noreferrer\">Keras<\/a>. For other languages, like C++, R, Python and so on it is quite easy to find some definitions. But for Keras, even google doesn't provide me useful searching results (at least not in the first 2 pages).<\/p>\n\n<p>TL;DR: How does load_data() know what is train and test of the data set?<\/p>\n","answers":[{"AnswerId":"58064676","CreationDate":"2019-09-23T14:30:30.687","ParentId":null,"OwnerUserId":"4727666","Title":null,"Body":"<p>The best way to find out is looking at Kera's <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/datasets\/mnist.py#L11\" rel=\"nofollow noreferrer\">code<\/a>:<\/p>\n\n<pre><code>def load_data(path='mnist.npz'):\n    path = get_file(path, origin='https:\/\/s3.amazonaws.com\/img-datasets\/mnist.npz', file_hash='8a61469f7ea1b51cbae51d4f78837e45')\n    with np.load(path, allow_pickle=True) as f:\n        x_train, y_train = f['x_train'], f['y_train']\n        x_test, y_test = f['x_test'], f['y_test']\n    return (x_train, y_train), (x_test, y_test)\n<\/code><\/pre>\n\n<p>You can see basically is downloading a file which contains the dataset, which is already separated in train and test data. \nThe only parameter (<code>path<\/code>) is basically where to store the downloaded dataset.<\/p>\n"},{"AnswerId":"58064617","CreationDate":"2019-09-23T14:25:46.777","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>For Keras source-stuff, I recommend searching the Github repository - e.g., Google \"keras mnist github\". From the <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/datasets\/mnist.py\" rel=\"nofollow noreferrer\">source code<\/a>, <code>mnist.load_data()<\/code> <em>unpacks<\/em> a dataset that was specifically <em>pickled<\/em> into a format that allows extracting the data as shown in the source code (also pre-sorted into train vs test, pre-shuffled, etc).<\/p>\n\n<p>Keras then returns the unpacked data in the form you used above.<\/p>\n"}]}
+{"QuestionId":58064701,"AnswerCount":1,"Tags":"<python><keras><neural-network>","CreationDate":"2019-09-23T14:31:53.627","AcceptedAnswerId":58066753.0,"OwnerUserId":9644683.0,"Title":"How to optimize accuracy of ANN","Body":"<p>I have 50 target classes of 300 datasets. <\/p>\n\n<p>This is my sample dataset, with 98 features:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Z0DaR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Z0DaR.png\" alt=\"Sample dataset\"><\/a><\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nimport seaborn as sns\nimport matplotlib.pyplot as plt\n%matplotlib inline\ndataset = pd.read_csv(root_path + 'pima-indians-diabetes.data.csv', header=None)\n\nX= dataset.iloc[:,0:8]\ny= dataset.iloc[:,8]\n\nfrom sklearn.preprocessing import StandardScaler\nsc = StandardScaler()\nX = sc.fit_transform(X)\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)\n\nfrom keras import Sequential\nfrom keras.layers import Dense\n\nclassifier = Sequential()\n#First Hidden Layer\nclassifier.add(Dense(units = 10, activation='relu',kernel_initializer='random_normal', input_dim=8))\n#Second  Hidden Layer\nclassifier.add(Dense(units = 10, activation='relu',kernel_initializer='random_normal'))\n#Output Layer\nclassifier.add(Dense(units = 1, activation='sigmoid',kernel_initializer='random_normal'))\n\n#Compiling the neural network\nclassifier.compile(optimizer ='adam',loss='binary_crossentropy', metrics =['accuracy'])\n\n#Fitting the data to the training dataset\nclassifier.fit(X_train,y_train, batch_size=2, epochs=10)\n<\/code><\/pre>\n\n<p>I get 19% accuracy here, and I don't know how to optimize my prediction result.<\/p>\n","answers":[{"AnswerId":"58066753","CreationDate":"2019-09-23T16:39:21.337","ParentId":null,"OwnerUserId":"5136550","Title":null,"Body":"<p>I am considering that you have performed the Dimentionality Reduction technique on your original data having 98 features and therefore you are using an 8-dimensional input feature in your model.<\/p>\n\n<p>I have a few observations on your implementation:<\/p>\n\n<h3>[As a Classification Problem]<\/h3>\n\n<p>As you have mentioned that your samples belong to 50 diffecent classes, the problem is certainly a multiclass classification problem. So, you need to encode your label first like:<\/p>\n\n<pre><code>from keras.utils import to_categorical\ny = to_categorical(y, num_classes=50, dtype='float32')\n<\/code><\/pre>\n\n<p>In this case, you need to change the number of output node (representing class) and activation function in the final layer as follows:<\/p>\n\n<pre><code>classifier.add(Dense(units = 50, activation='softmax'))\n<\/code><\/pre>\n\n<p>Furthermore, you have to ue <code>categorical_crossentropy<\/code> as a loss function while compiling your model.<\/p>\n\n<pre><code>classifier.compile(optimizer ='adam',loss='categorical_crossentropy', metrics =['accuracy'])\n<\/code><\/pre>\n\n<hr>\n\n<h3>[As a Regression Problem]<\/h3>\n\n<p>You can also consider this problem as a multiple regression problem as the output is within the range of 0 to 50 (continuous) and can keep a single output node in the final layer as you did. But in that case, you should use a <code>linear<\/code> activation function instead of <code>sigmoid<\/code>.<\/p>\n\n<p>So, the final layer should be like:<\/p>\n\n<pre><code>classifier.add(Dense(units = 1)) # default activation is linear\n<\/code><\/pre>\n\n<p>Additionally, In case of regression problem, <code>mean_squared_error<\/code> is the most relevant cost function to use (assuming not many outliers in your dataset) and <code>accuracy<\/code> as a performance metric is irrelevant (rather you may use <code>mean_absolute_error<\/code> which is analogous to loss). Hence, the second modification is:<\/p>\n\n<pre><code>classifier.compile(optimizer ='adam',loss='mean_squared_error')\n<\/code><\/pre>\n"}]}
+{"QuestionId":58064984,"AnswerCount":1,"Tags":"<keras><conv-neural-network><categorical-data><multiclass-classification>","CreationDate":"2019-09-23T14:47:48.993","AcceptedAnswerId":null,"OwnerUserId":7342371.0,"Title":"How to get probability of each class instead of one hot encoded array with one value 1 and others 0?","Body":"<p>My Sequential CNN model is trained on 39 classes as a multi-class classifier. As for predictions, it returns a one-hot encoded array like <code>[0,0,...1,0,...]<\/code> whereas I want something like <code>[0.012,0.022,0.067,...,0.997,0.0004,...]<\/code><\/p>\n\n<p>Is there a way to get this? if not what exactly should I make to get these? <\/p>\n\n<p>The reason I want it this way is to verify how close are other classes, so if one says 0.98 and others say 0.96 then I am doing something wrong, data isn't enough, etc..<\/p>\n\n<p>Thank you :)<\/p>\n\n<p>My model is basically a keras.model resnet50 with following configs : <\/p>\n\n<pre><code>model = keras.applications.resnet.ResNet50(include_top=False, weights=None, input_tensor=None, input_shape=(64,64,1), pooling='avg', classes=39)\n\nx = model.output\nx = Dropout(0.7)(x)\nnum_classes = 39\npredictions = Dense(num_classes, activation= 'softmax')(x)\nmodel = Model(inputs = model.input, outputs = predictions)\n\noptimizer = keras.optimizers.Adam(learning_rate=0.001, beta_1=0.9, beta_2=0.999, amsgrad=False)\nmodel.compile(optimizer, loss='categorical_crossentropy', metrics=['categorical_accuracy'], loss_weights=None, sample_weight_mode=None, weighted_metrics=None, target_tensors=None)\n\n<\/code><\/pre>\n\n<p>Sample input :<\/p>\n\n<pre><code>import cv2\nimg = cv2.imread(IMAGE_PATH, 0)\nimg = cv2.resize(img, (64,64))\nimg = np.reshape(img, (1,64,64,1))\npredicted_class_indices = np.argmax(model.predict(img, verbose = 1))\n<\/code><\/pre>\n\n<p>Sample output:<\/p>\n\n<pre><code>array([[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., 1., 0., 0., 0., 0., 0.,\n        0., 0., 0., 0., 0., 0., 0.]], dtype=float32)\n<\/code><\/pre>\n\n<p>Desired output (numbers are hypothetical):<\/p>\n\n<pre><code>array([[0.022, 0.353, 0.0535, 0.52, 0212., 0.822, 0.532, 0., 0., 0., 0., 0., 0., 0., 0., 0.,\n        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0.,\n        0., 0., 0., 0., 0., 0., 0.]], dtype=float32)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58065298","CreationDate":"2019-09-23T15:06:42.770","ParentId":null,"OwnerUserId":"6537424","Title":null,"Body":"<p>One way to do so is to remove the last activation layer (<a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/9370\" rel=\"nofollow noreferrer\">related issue<\/a>).<\/p>\n\n<p>You can do so by using <code>model.layers[-1].activation=None<\/code>.<\/p>\n\n<p>However, the softmax shouldn't output a one-hot vector but the prob distribution, you might want to check how your training is doing.<\/p>\n"}]}
+{"QuestionId":58065426,"AnswerCount":0,"Tags":"<python><tensorflow><pip><pycharm><pipenv>","CreationDate":"2019-09-23T15:13:35.767","AcceptedAnswerId":null,"OwnerUserId":282855.0,"Title":"pipenv - No module named tensorflow despite it is installed","Body":"<p>I am not able to import <code>tensorflow<\/code> (gives error when I do) despite it is already installed. Here is the content of my <code>Pipfile<\/code>:<\/p>\n\n<pre><code>[[source]]\nname = \"pypi\"\nurl = \"https:\/\/pypi.org\/simple\"\nverify_ssl = true\n\n[dev-packages]\n\n[packages]\ntensorflow = \"*\"\nnumpy = \"*\"\nscipy = \"*\"\nscikit-learn = \"*\"\npillow = \"*\"\nh5py = \"*\"\nkeras = \"*\"\nopencv-python = \"*\"\n\n[requires]\npython_version = \"3.7\"\n<\/code><\/pre>\n\n<p><strong>Things I have already tried to fix this issue:<\/strong><\/p>\n\n<ul>\n<li>Executed <code>pipenv update<\/code><\/li>\n<li>Executed <code>pipenv install tensorflow<\/code><\/li>\n<li>Click to <code>install package tensorflow<\/code> prompt in <code>PyCharm<\/code><\/li>\n<\/ul>\n\n<p>Python version: <code>Python 3.7.4 (tags\/v3.7.4:e09359112e, Jul  8 2019, 20:34:20) [MSC v.1916 64 bit (AMD64)] on win32<\/code><\/p>\n\n<p><strong>The output of <code>pip debug --verbose<\/code>:<\/strong><\/p>\n\n<pre><code>pip version: pip 19.2.3 from d:\\.virtualenvs\\keras-fer-pqstp19f\\lib\\site-packages\\pip (python 3.7)\nsys.version: 3.7.4 (tags\/v3.7.4:e09359112e, Jul  8 2019, 20:34:20) [MSC v.1916 64 bit (AMD64)]\nsys.executable: d:\\.virtualenvs\\keras-fer-pqstp19f\\scripts\\python.exe\nsys.getdefaultencoding: utf-8\nsys.getfilesystemencoding: utf-8\nlocale.getpreferredencoding: cp1252\nsys.platform: win32\nsys.implementation:\n  name: cpython\nConfig variable 'Py_DEBUG' is unset, Python ABI tag may be incorrect\nConfig variable 'WITH_PYMALLOC' is unset, Python ABI tag may be incorrect\nCompatible tags: 14\n  cp37-cp37m-win_amd64\n  cp37-none-win_amd64\n  py3-none-win_amd64\n  cp37-none-any\n  cp3-none-any\n  py37-none-any\n  py3-none-any\n  py36-none-any\n  py35-none-any\n  py34-none-any\n  py33-none-any\n  py32-none-any\n  py31-none-any\n  py30-none-any\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58065951,"AnswerCount":0,"Tags":"<python><artificial-intelligence><caffe2>","CreationDate":"2019-09-23T15:45:55.343","AcceptedAnswerId":null,"OwnerUserId":12108333.0,"Title":"assert 'SpatialNarrowAs' in workspace.RegisteredOperators() [Detectron test AssertionError]","Body":"<p>i got Caffe2 and Detectron installed on my Windows 10 machine today, but while running the test <code>\\Detectron\\detectron\\tests\\test_spatial_narrow_as_op.py<\/code> i got this error:<\/p>\n\n<pre><code>C:\\Users\\ilink\\source\\Detectron\\detectron\\tests&gt;python test_spatial_narrow_as_op.py\n[E ..\\caffe2\\core\\init_intrinsics_check.cc:43] CPU feature avx is present on your machine, but the Caffe2 binary is not compiled with it. It means you may not get the full speed of your CPU.\n[E ..\\caffe2\\core\\init_intrinsics_check.cc:43] CPU feature avx2 is present on your machine, but the Caffe2 binary is not compiled with it. It means you may not get the full speed of your CPU.\n[E ..\\caffe2\\core\\init_intrinsics_check.cc:43] CPU feature fma is present on your machine, but the Caffe2 binary is not compiled with it. It means you may not get the full speed of your CPU.\nIn Windows, Detectron ops are built-in. No need to load dynamically. Ignore the following warning.\nIgnoring  as it is not a valid file.\nTraceback (most recent call last):\n  File \"test_spatial_narrow_as_op.py\", line 89, in &lt;module&gt;\n    assert 'SpatialNarrowAs' in workspace.RegisteredOperators()\nAssertionError\n<\/code><\/pre>\n\n<p>If it helps, i used <code>http:\/\/gianni.rosagallina.com\/en\/posts\/2018\/10\/09\/caffe2-gpu-windows-2.html<\/code> to help install Detectron. I did <code>pip install pytorch<\/code> to install Caffe2.<\/p>\n\n<p>I'm pretty new to AI related programming, as well as stackoverflow in general, so if i'm missing something, please tell me. \nThanks in advance :)<\/p>\n","answers":[]}
+{"QuestionId":58065976,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-datasets><tfrecord>","CreationDate":"2019-09-23T15:47:53.223","AcceptedAnswerId":58076193.0,"OwnerUserId":5852258.0,"Title":"Unfurling TFRecords growing slower and slower","Body":"<p>I'm trying to convert a TFRecord dataset back to images and I'm using the following code to do so:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def get_im_and_label_helper(parsed_features, im_format, label_format):\n    im = tf.image.decode_png(parsed_features['image\/encoded'])\n    label = tf.image.decode_png(parsed_features['image\/segmentation\/class\/encoded'])\n    im, label = im.eval(), label.eval()\n    return im, label\n\nfor tfr_file_path_name in tfr_files_list:\n    tfr_file_path = os.path.join(sub_dataset_dir, tfr_file_path_name)\n    record_iterator = tf.python_io.tf_record_iterator(tfr_file_path)\n\n    for string_record in record_iterator:\n        parsed_features = tf.parse_single_example(string_record, READ_FEATURES)\n\n        filename = parsed_features['image\/filename'].eval().decode(\"utf-8\")\n        im, label = get_im_and_label_helper(parsed_features, im_format, label_format)\n        imageio.imwrite(os.path.join(target_dir, \"images\", filename + \".png\"), im)\n        imageio.imwrite(os.path.join(target_dir, \"labels\", filename + \".png\"), label)\n<\/code><\/pre>\n\n<p>It works fine and does what I expect - extracts the images and labels and saves them in the proper place. It starts fast and it gets slower and slower as it goes on. I'm inexperienced with tensorflow, so I assume I'm causing some computation graph to grow bigger and bigger, but I don't really know.<\/p>\n\n<p>Any ideas? <\/p>\n","answers":[{"AnswerId":"58076193","CreationDate":"2019-09-24T08:34:39.740","ParentId":null,"OwnerUserId":"5852258","Title":null,"Body":"<p>Using <code>tf.enable_eager_execution()<\/code> followed by <code>tf.executing_eagerly()<\/code>, and replacing all <code>.eval()<\/code>with <code>.numpy()<\/code>solved the problem.<\/p>\n"}]}
+{"QuestionId":58066194,"AnswerCount":1,"Tags":"<neural-network><pytorch><lstm><recurrent-neural-network>","CreationDate":"2019-09-23T16:01:25.593","AcceptedAnswerId":null,"OwnerUserId":7134235.0,"Title":"what does nn.Linear() do in pytorch's last, and why is it necessary?","Body":"<p>I am working with some code that trains an lstm to generate sequences. After training the model, the lstm() method is called:<\/p>\n\n<pre><code>x = some_input\nlstm_output, (h_n, c_n) = lstm(x, hc) \nfuncc = nn.Linear(in_features=lstm_num_hidden,\n                  output_features=vocab_size,\n                  bias=True)\nfunc_output = func(lstm_output)\n<\/code><\/pre>\n\n<p>I've looked at the documentation for <code>nn.Linear()<\/code> but I still don't understand what this transformation is doing and why it is necessary. If the lstm has already been trained, then the output it gives should already have a pre-established dimensionality. This output (lstm_output) would be the generated sequence, or in my case an array of vectors. Am I missing something here?<\/p>\n","answers":[{"AnswerId":"58066291","CreationDate":"2019-09-23T16:08:39.267","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p>Here, the Linear layer is transforming the hidden state representations (<code>lstm_output<\/code>) produced by the LSTM into a vector of size <code>vocab_size<\/code>. Your understanding is perhaps wrong. The <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html?highlight=linear#torch.nn.Linear\" rel=\"nofollow noreferrer\">Linear<\/a> layer should be trained along with the LSTM. <\/p>\n\n<p>And I guess you are trying to generate a sequence of tokens (words), so the Linear layer should be followed by a <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.functional.html?highlight=softmax#torch.nn.functional.softmax\" rel=\"nofollow noreferrer\">Softmax<\/a> operation to predict a probability distribution over the vocabulary.<\/p>\n"}]}
+{"QuestionId":58066673,"AnswerCount":1,"Tags":"<python><numpy><keras><categorical-data>","CreationDate":"2019-09-23T16:33:58.207","AcceptedAnswerId":58085265.0,"OwnerUserId":7215135.0,"Title":"How can I store one-hot-encodings as an object?","Body":"<p>First, some background on my model's architecture. <\/p>\n\n<p>The inputs to my keras the model are fairly simple:<\/p>\n\n<ul>\n<li>Categorical variable A<\/li>\n<li>Categorical variable B<\/li>\n<li>Numerical input C, on the range [0,1]. <\/li>\n<\/ul>\n\n<p>The model features a single output: <\/p>\n\n<ul>\n<li>a number on the [0,1] <\/li>\n<\/ul>\n\n<p>When training the model, my input data is a dataframe sourced from an SQL database using <code>pd.read_sql()<\/code>. I one-hot-encode the categorical variables A  and B (which are in <code>col1<\/code> adn <code>col2<\/code> of the dataframe <code>original_data<\/code>, respectively) with the following function:<\/p>\n\n<pre><code>from keras import utils as np_utils\n\ndef preprocess_categorical_features(self):\n        col1 = np_utils.to_categorical(np.copy(self.original_data.CURRENT_RTIF.values))\n        col2 = np_utils.to_categorical(np.copy(self.original_data.NEXT_RTIF.values))\n        cat_input_data = np.append(col1,col2,axis=1)\n        return cat_input_data\n<\/code><\/pre>\n\n<p>Later on, when I need to make predictions from this model, input data comes from a live feed from RabbitMQ in the form of a dictionary. This RabbitMQ data has to be processed by it's own (different) <code>reprocess_categorical_features()<\/code> function.<\/p>\n\n<p>This brings me to my question: How can I ensure that the one-hot-encodings are exactly the same, whether I'm preprocessing data from the database, or a rabbitMQ feed?<\/p>\n\n<p>The One-Hot Encoding for A applied to the database data:<\/p>\n\n<pre><code>|---------------------|------------------|\n|          A          | One-Hot-Encoding |\n|---------------------|------------------|\n|       \"coconut\"     |      &lt;0,1,0,0&gt;   |\n|---------------------|------------------|\n|       \"apple\"       |      &lt;1,0,0,0&gt;   |\n|---------------------|------------------|\n|       \"quince\"      |      &lt;0,0,0,1&gt;   |\n|---------------------|------------------|\n|       \"plum\"        |      &lt;0,1,0,0&gt;   |\n|---------------------|------------------|\n<\/code><\/pre>\n\n<p>The One-Hot Encoding for A applied to the RabbitMQ data (they have to be the same):<\/p>\n\n<pre><code>|---------------------|------------------|\n|          A          | One-Hot-Encoding |\n|---------------------|------------------|\n|       \"coconut\"     |      &lt;0,1,0,0&gt;   |\n|---------------------|------------------|\n|       \"apple\"       |      &lt;1,0,0,0&gt;   |\n|---------------------|------------------|\n|       \"quince\"      |      &lt;0,0,0,1&gt;   |\n|---------------------|------------------|\n|       \"plum\"        |      &lt;0,1,0,0&gt;   |\n|---------------------|------------------|\n<\/code><\/pre>\n\n<p>Is there a way for me to save the encodings as a dataframe, numpy ndarray, or dictionary, so that I could pass the encodings from the function that pre-processes my training data, to the function that pre-processes my input data? I'm open to using other libraries than Keras for OHE, but I'm interested to know if there's a way to accomplish this using keras' <a href=\"https:\/\/keras.io\/utils\/#to_categorical\" rel=\"nofollow noreferrer\">to_categorical<\/a> function I'm currently using.<\/p>\n","answers":[{"AnswerId":"58085265","CreationDate":"2019-09-24T17:14:59.580","ParentId":null,"OwnerUserId":"7215135","Title":null,"Body":"<p>Instead of relying on keras' <a href=\"https:\/\/keras.io\/utils\/\" rel=\"nofollow noreferrer\"><code>utils.to_categorical<\/code><\/a> method, I decided to use <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.OneHotEncoder.html\" rel=\"nofollow noreferrer\"><code>sklearn.preprocessing.OneHotEncoder<\/code><\/a>. This allowed me to declare a one-hot-encoder object, <code>self.encoder<\/code> when dealing with the training data:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class TrainingData:\n    def preprocess_categorical_features(self):\n        # declare OneHotEncoder object to save for later\n        self.encoder = OneHotEncoder(sparse=False)\n\n        # fit encoder to data\n        self.encoder.fit(self.original_data.CURRENT_RTIF.values.reshape(-1,1))\n\n        # perform one-hot-encoding on columns 1 and 2 of the training data\n        col1 = self.encoder.transform(self.original_data.CURRENT_RTIF.values.reshape(-1,1))\n        col2 = self.encoder.transform(self.original_data.NEXT_RTIF.values.reshape(-1,1))\n\n        # return on-hot-encoded data as a numpy ndarray\n        cat_input_data = np.append(col1,col2,axis=1)\n        return cat_input_data\n<\/code><\/pre>\n\n<p>Later on, I could re-use that encoder (by passing it as an argument, <code>training_data_ohe_encoder<\/code>) to the method that processes the input data needed to finally make my predictions. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class LiveData:\n    def preprocess_categorical_features(self, training_data_ohe_encoder):\n        # notice the training_data_ohe_encoder parameter; this is the \n        # encoder attribute from the Training Data Class.\n\n        # one-hot-encode the live data using the training_data_ohe_encoder encoder\n        col1 = training_data_ohe_encoder.transform(np.copy(self.preprocessed_data.CURRENT_RTIF.values).reshape(-1, 1))\n        col2 = training_data_ohe_encoder.transform(np.copy(self.preprocessed_data.NEXT_RTIF.values).reshape(-1, 1))\n\n        # return on-hot-encoded data as a numpy ndarray\n        cat_input_data = np.append(col1,col2,axis=1)\n        return cat_input_data\n<\/code><\/pre>\n"}]}
+{"QuestionId":58066803,"AnswerCount":1,"Tags":"<tensorflow><variable-assignment>","CreationDate":"2019-09-23T16:42:51.450","AcceptedAnswerId":58067855.0,"OwnerUserId":6714045.0,"Title":"How to assign an element of an tensor in tensorflow","Body":"<p>I want to understand the error message of the following code<\/p>\n\n<pre><code>M = K.eye(2)\nK.assign(M[0,1],1.0)\n<\/code><\/pre>\n\n<p>The message I got is \"Tried to convert 'input' to a tensor and failed. Error: None values not supported.\"<\/p>\n","answers":[{"AnswerId":"58067855","CreationDate":"2019-09-23T18:01:25.367","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>You can assign an element of a variable in tensorflow. Here is an example. (I didn't find K.assign this operation in my installed tensorflow version, btw)<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport keras.backend as K\n\nM = tf.Variable(K.eye(2), tf.float32)\nassign_op = tf.assign(M[0,1], 1.0)\n\nwith tf.Session() as sess:\n  sess.run(tf.global_variables_initializer())\n  print(sess.run(M))\n  sess.run(assign_op)\n  print(sess.run(M))\n\n#[[1. 0.]\n# [0. 1.]]\n\n#[[1. 1.]\n# [0. 1.]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":58066984,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-23T16:55:23.567","AcceptedAnswerId":null,"OwnerUserId":3103578.0,"Title":"How can i switch to tf.train.MonitoredTrainingSession?","Body":"<p>I'm new to python and tensorflow .. i tried to train a model but got <\/p>\n\n<pre><code>WARNING:tensorflow:From train.py:102: Supervisor.__init__ (from \ntensorflow.python.training.supervisor) is deprecated and will be removed in a future version.\nInstructions for updating:\nPlease switch to tf.train.MonitoredTrainingSession\nWARNING:tensorflow:Issue encountered when serializing global_step.\nType is unsupported, or the types of the items don't match field type in CollectionDef. Note \nthis is a warning and probably safe to ignore.\n'Tensor' object has no attribute 'to_proto \n<\/code><\/pre>\n\n<p>but couldn't find the way for switching .. Is that necessary to switch or it is just a warning ?<\/p>\n","answers":[]}
+{"QuestionId":58066994,"AnswerCount":0,"Tags":"<theano><gradient-descent><loss-function>","CreationDate":"2019-09-23T16:55:58.737","AcceptedAnswerId":null,"OwnerUserId":12103346.0,"Title":"Theano gradient descent and cost function issue","Body":"<p>I'm getting 'ValueError: Input dimension mis-match.' for my theano gradient descent implmentation.<\/p>\n\n<p>I tried transposing the input by using np.transpose, but the dimension remains unchanged.<\/p>\n\n<pre><code>import theano\nfrom theano import tensor as T\nimport numpy as np\nimport math \nimport cmath \nglobal W\nl=1\n\ndef inputvector(theta,phi,zeta):\n    global xreal,ximag\n    xreal=np.asarray([math.cos(2*theta-phi)*math.cos(phi), math.cos(2*theta-phi)*math.sin(phi)*math.cos(zeta), math.cos(2*theta-phi)*math.sin(phi)*math.sin(zeta)])\n    ximag=np.asarray([-math.sin(2*theta-phi)*math.sin(phi), math.sin(2*theta-phi)*math.cos(phi)*math.cos(zeta), math.sin(2*theta-phi)*math.cos(phi)*math.sin(zeta)])   \n    return xreal,ximag\n\n\ndef weights(w11r,w12r,w13r,w21r,w22r,w23r,w31r,w32r,w33r,w11i,w12i,w13i,w21i,w22i,w23i,w31i,w32i,w33i):\n    global W\n    W = theano.shared(np.asarray([w11r, w12r, w13r,w21r,w22r,w23r,w31r,w32r,w33r,w11i,w12i,w13i,w21i,w22i,w23i,w31i,w32i,w33i]), 'W')\n    return W\n\nxreal=T.dvector('xreal')\nximag=T.dvector('ximag')\nweights(np.random.rand(1),np.random.rand(1),np.random.rand(1),np.random.rand(1),np.random.rand(1),np.random.rand(1),np.random.rand(1),np.random.rand(1),np.random.rand(1),np.random.rand(1),np.random.rand(1),np.random.rand(1),np.random.rand(1),np.random.rand(1),np.random.rand(1),np.random.rand(1),np.random.rand(1),np.random.rand(1))\n\ncost=T.mean(T.sqr(np.dot(W[0:3],xreal)-np.dot(W[9:12],ximag)))\n\ngradients = theano.tensor.grad(cost, [W])\nW_updated = W - (0.1 * gradients[0])\nupdates = [(W, W_updated)]\ntrain=theano.function(inputs=[xreal,ximag],outputs=cost,updates=updates,allow_input_downcast=True)\nfor i in range(100):\n    inputvector(np.random.rand(1),np.random.rand(1),np.random.rand(1))\n    train(xreal,ximag)\n    print (W.get_value())\n<\/code><\/pre>\n\n<p>I don't see any reason that the input sizes should disagree. And I don't think I even have 4 inputs to begin with.<\/p>\n","answers":[]}
+{"QuestionId":58067146,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-23T17:07:17.900","AcceptedAnswerId":null,"OwnerUserId":3029270.0,"Title":"processing the output of model (prediction) before loss & optimization step in tensorflow","Body":"<p>I have model auto-encoder I want to fetch y_prediction value process it then return it back before loss &amp; optimizer step <\/p>\n\n<pre><code>with tf.device('\/gpu:'+GPU0):\n    ################################ ae loss\n    Y_ = tf.reshape(self.Y, shape=[-1, 64 ** 3])\n    Y_pred_modi_ = tf.reshape(self.D_1_Y, shape=[-1, 64 ** 3])\n    w = 0.85\n    self.aeu_loss = tf.reduce_mean(\n        -tf.reduce_mean(w * Y_ * tf.log(Y_pred_modi_ + 1e-8), reduction_indices=[1]) -\n        tf.reduce_mean((1 - w) * (1 - Y_) * tf.log(1 - Y_pred_modi_ + 1e-8), \n                       reduction_indices=[1]))\n    self.total_loss=self.aeu_loss\n\n\n    #################################  optimizer\n\n\nwith tf.device('\/gpu:'+GPU0):\n    aeu_var = [var for var in tf.trainable_variables() if var.name.startswith('aeu_2')]\n    print(aeu_var)\n    #dis_var = [var for var in tf.trainable_variables() if var.name.startswith('dis')]\n\n    self.aeu_g_optim = tf.train.AdamOptimizer(learning_rate=0.0001, beta1=0.9, beta2=0.999,\n                       epsilon=1e-8).minimize(self.total_loss, var_list=aeu_var)\n<\/code><\/pre>\n\n<p>before loss I want to take <code>Y_pred_modi_<\/code> value process it then return it back <\/p>\n","answers":[]}
+{"QuestionId":58067212,"AnswerCount":0,"Tags":"<python><tensorflow><crash><sparse-matrix>","CreationDate":"2019-09-23T17:13:05.187","AcceptedAnswerId":null,"OwnerUserId":12044023.0,"Title":"Custom OP gradient crashes tensorflow","Body":"<p>I created a custom op for tensorflow 2-beta1 with gradients but when I run the training algorithm, tensorflow crashes with exit code -1073740791 (0xC0000409).<\/p>\n\n<p>Before the crash I get an error saying it expected int32 but got int64 (<code>F tensorflow\/core\/framework\/tensor.cc:631] Check failed: dtype() == expected_dtype (9 vs. 3) int32 expected, got int64<\/code>). So I tried casting the only int64 values of my op to int32, but it still persists I have no idea why except there might be some unrelated error in my gradient. The unit test of my custom op is passing.<\/p>\n\n<p>The error appears to occur before returning from but after executing the train function which had worked before (although this might not be accurate, as it is cycled multiple times before returning\/crashing). The only thing I changed since it worked was switching to sparse tensors for a part of my neural network due to memory constraints. <\/p>\n\n<p>This is my gradient calculation:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>@ops.RegisterGradient(\"BatchSparseDenseMatmul\")\ndef _batch_sparse_dense_matmul_grad(op, grad):\n    a_indices, a_values, a_shape = op.inputs[:3]\n    b = op.inputs[3]\n    b_grad = batch_matmul_module.batch_sparse_dense_matmul(a_indices, a_values, a_shape, grad, transpose_a=True)\n    rows = a_indices[:, 1]\n    cols = a_indices[:, 2]\n    parts_a = array_ops.gather(grad, rows)\n    parts_b = array_ops.gather(b, cols)\n    a_values_grad = math_ops.reduce_sum(parts_a * parts_b, axis=1)\n    return tf.cast(a_indices, tf.int32), a_values_grad, tf.cast(a_shape, tf.int32), b_grad\n<\/code><\/pre>\n\n<p>As I copied a lot of the code for the gradients from the <code>spare_dense_matmul<\/code> gradient calculation it might be that I did something wrong when adjusting the code, as I do not fully understand what the <code>gather<\/code> and <code>reduced_sum<\/code> are good for. Any help is appreciated!<\/p>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>I cast the inputs to my feed forwad like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>batch_matmul_module.batch_sparse_dense_matmul(tf.cast(a.indices, tf.int32), a.values, tf.cast(a.dense_shape, tf.int32), b)\n<\/code><\/pre>\n\n<p>the error is gone but I get <code>Cannot reshape a tensor with 0 elements to shape [1,1,11520] (11520 elements) for 'Reshape_18' (op: 'Reshape') with input shapes: [0,0], [3] and with input tensors computed as partial shapes: input[1] = [1,1,11520].<\/code> but I think this is unrelated.<\/p>\n\n<p><strong>Another Edit:<\/strong><\/p>\n\n<p>Once I fixed the Error shown in the last edit I get the same error again, but I am sure the error is not coming from the op shown above, as for testing with small data I could use <code>to_dense<\/code> and <code>matmul<\/code>. Is there a way to debug this?<\/p>\n","answers":[]}
+{"QuestionId":58067427,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><one-hot-encoding>","CreationDate":"2019-09-23T17:28:54.297","AcceptedAnswerId":null,"OwnerUserId":10579777.0,"Title":"Tensorflow: OneHot-encoding with variable sized length","Body":"<p>I need to onehot-encode some positions with TensorFlow.<\/p>\n\n<p>However, the length of the input sequences (and therefore the <code>depth-argument<\/code> in <code>tf.one_hot<\/code>) is None as I work with variable sized inputs.<\/p>\n\n<p>This throws the following error: <\/p>\n\n<blockquote>\n  <p>\"ValueError: Tried to convert 'depth' to a tensor and failed. Error: None values not supported.\".<\/p>\n<\/blockquote>\n\n<p>Is there a workaround for this?<\/p>\n\n<p>I have already tried to set the depth to the correct sequence length before each individual call (through a variable that has some arbitrary initialization value) for a given sequence but as the computational graph is already built, the changes do not come into effect and the depth is stuck at the initialization value.<\/p>\n","answers":[{"AnswerId":"58080853","CreationDate":"2019-09-24T12:57:24.637","ParentId":null,"OwnerUserId":"10579777","Title":null,"Body":"<p>This is how I solved the issue: The problem was that I was using someone else's code and the <code>depth<\/code> argument in <code>tf.one_hot<\/code> was derived from <code>someTensor.get_shape().as_list()[1]<\/code>. The problem here is that if the shape of <code>someTensor<\/code> is unknown, the argument is a Python-<code>None<\/code> which is not a valid argument for <code>tf.one_hot<\/code>. However, using <code>tf.shape(someTensor)[1]<\/code> solved that problem as it returns a Dimension with unknown shape instead of a Python-<code>None<\/code>. A Dimension with unknown shape is a valid <code>depth<\/code>-argument for <code>tf.one_hot<\/code>.<\/p>\n"}]}
+{"QuestionId":58067508,"AnswerCount":0,"Tags":"<keras><conv-neural-network><loss-function><blending>","CreationDate":"2019-09-23T17:34:34.130","AcceptedAnswerId":null,"OwnerUserId":12050636.0,"Title":"Combinning Multiple Neural Networks","Body":"<p>can you combine multiple neural networks and have a totally new loss function for the combined network. <\/p>\n\n<p>I would like to assign a neural network for each part of the data (each part of an image for example) and optimize the solution with a totally new loss function.<\/p>\n","answers":[]}
+{"QuestionId":58067594,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-23T17:41:45.557","AcceptedAnswerId":58070302.0,"OwnerUserId":12108968.0,"Title":"Getting the image name using autoencoder on tensorflow","Body":"<p>I'm using this tensorflow image search script:\n<a href=\"https:\/\/www.kaggle.com\/jonmarty\/using-autoencoder-to-search-images\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/jonmarty\/using-autoencoder-to-search-images<\/a><\/p>\n\n<pre><code>def search(image):\n    hidden_states = [sess.run(hidden_state(X, mask, W, b),\n                     feed_dict={X: im.reshape(1, pixels), mask:  \n                     np.random.binomial(1, 1-corruption_level, (1, pixels))})\n                     for im in image_set]\n\n    query = sess.run(hidden_state(X, mask, W, b),\n                     feed_dict={X: image.reshape(1,pixels), mask: np.random.binomial(1, 1-corruption_level, (1, pixels))})\n\n    starting_state = int(np.random.random()*len(hidden_states)) #choose random starting state\n    best_states = [imported_images[starting_state]]\n    distance = euclidean_distance(query[0], hidden_states[starting_state][0]) #Calculate similarity between hidden states\n    for i in range(len(hidden_states)):\n        dist = euclidean_distance(query[0], hidden_states[i][0])\n        if dist &lt;= distance:\n            distance = dist #as the method progresses, it gets better at identifying similiar images\n            best_states.append(imported_images[i])\n    if len(best_states)&gt;0:\n        return best_states\n    else:\n        return best_states[len(best_states)-101:]\n<\/code><\/pre>\n\n<p>I'm wondering about if it's possible to know the image name (e.g: homer.jpg). I'm lost and I don't know what should I add in the code to know that.. That's the part of the script where I print the results:<\/p>\n\n<pre><code> print(len(results))\nslots = 0\nplt.figure(figsize = (125,125))\nfor im in results[::-1]: #reads through results backwards (more similiar images first)\n    plt.subplot(10, 10, slots+1) \n    plt.imshow(cv2.cvtColor(im, cv2.COLOR_BGR2RGB)); plt.axis('off')\n    slots += 1\n<\/code><\/pre>\n\n<p>Thank you a lot! :)<\/p>\n","answers":[{"AnswerId":"58070302","CreationDate":"2019-09-23T21:17:01.867","ParentId":null,"OwnerUserId":"10553098","Title":null,"Body":"<p>You need to modify the search function.<\/p>\n\n<p>Specifically, look at the line:<\/p>\n\n<pre><code>best_states.append(imported_images[i])\n<\/code><\/pre>\n\n<p>If you want to map between the images returned and the filenames, you need to record and return that index, <code>i<\/code>.  Consider adding a <code>best_states_index<\/code> variable and returning both or simply replacing <code>imported_images[i]<\/code> with <code>i<\/code> and using that to access both the filename and the image data.<\/p>\n\n<p>More explicitly:<\/p>\n\n<pre><code>def search(image):\n    hidden_states = [sess.run(hidden_state(X, mask, W, b),\n                     feed_dict={X: im.reshape(1, pixels), mask:  \n                     np.random.binomial(1, 1-corruption_level, (1, pixels))})\n                     for im in image_set]\n\n    query = sess.run(hidden_state(X, mask, W, b),\n                     feed_dict={X: image.reshape(1,pixels), mask: np.random.binomial(1, 1-corruption_level, (1, pixels))})\n\n    starting_state = int(np.random.random()*len(hidden_states)) #choose random starting state\n    best_states = [imported_images[starting_state]]\n    best_states_index = [starting_state]\n    distance = euclidean_distance(query[0], hidden_states[starting_state][0]) #Calculate similarity between hidden states\n    for i in range(len(hidden_states)):\n        dist = euclidean_distance(query[0], hidden_states[i][0])\n        if dist &lt;= distance:\n            distance = dist #as the method progresses, it gets better at identifying similiar images\n            best_states.append(imported_images[i])\n            best_states_index.append(i)\n    if len(best_states)&gt;0:\n        return best_states, best_states_index\n    else:\n        return best_states[len(best_states)-101:], best_states_index[len(best_states)-101:]\n<\/code><\/pre>\n"}]}
+{"QuestionId":58067839,"AnswerCount":0,"Tags":"<python><keras><missing-data>","CreationDate":"2019-09-23T18:00:50.690","AcceptedAnswerId":null,"OwnerUserId":10852215.0,"Title":"Mask value for only certain columns?","Body":"<p>I am trying to use a masking layer in Keras to deal with missing data however i am running into a problem because for some of my columns, a value of 0 does mean something. Given the nature of my features, there really isn't any value that would work for all columns. Is there a way to only apply the masking layer to certain columns? Or better yet is there a way to make your masking value NaN, or does it need a real number?<\/p>\n","answers":[]}
+{"QuestionId":58067995,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><machine-learning><neural-network><regression>","CreationDate":"2019-09-23T18:12:27.820","AcceptedAnswerId":null,"OwnerUserId":9625777.0,"Title":"Bare bone neural network unable to replicate the neural network implemented using Tensoflow","Body":"<p>I have sought some help and trained a regression model that takes in a single dependent variable <strong>Y<\/strong> and gives the three independent variable <strong>R, B and G<\/strong> as output. This has been done in attempt to approximate a function : <strong>Y = R+B+G<\/strong>.<\/p>\n\n<p>The neural network responsible for this is:<\/p>\n\n<pre><code>    input1 = Input(shape=(1,))\n    l1 = Dense(3, activation='relu')(input1)\n    out = Dense(3)(l1)\n\n    model = Model(inputs=input1, outputs=[out])\n    model.compile(\n                  optimizer='adam',\n                  loss=['mean_squared_error'],\n                  )\n<\/code><\/pre>\n\n<p>But, this is done using Keras and tensorflow. As this model ran successfully, I now needed to implement the same model (function approximation using regression). <\/p>\n\n<p>This is what I did to replicate the network entirely:<\/p>\n\n<pre><code>#ReLu Definition\ndef relu(x, deriv=False):\n  if(deriv==True):\n    x[x&lt;0] = 0.01\n    x[x&gt;=0] = 1\n    return x\n  return np.maximum(x,0.01*x)\n\n#Initializing Weights\nsyn0 = np.random.randn(1,3)*np.sqrt(1\/4)\n\n#Model Training\nfor iter in range(1000):\n    # forward propagation\n    l0 = X\n\n    l1 = relu(np.dot(l0,syn0))\n\n    # how much did we miss?\n    l1_error = y - l1\n    print(l1_error)\n\n    # multiply how much we missed by the \n    # slope of the relu at the values in l1\n    l1_delta = l1_error * relu(l1,True)\n\n    # update weights\n    syn0 += np.dot(l0.T,l1_delta)\n\n<\/code><\/pre>\n\n<p>But, the tensorflow's amazing results are nowhere to be replicated in the barebone model developed by me. I have not used <em>adam<\/em> optimizer, yet all other things are similar. I am unable to figure out the correct way of doing it. Any help?<\/p>\n","answers":[]}
+{"QuestionId":58068314,"AnswerCount":1,"Tags":"<python><keras><jupyter-notebook><shapes>","CreationDate":"2019-09-23T18:36:59.247","AcceptedAnswerId":58068586.0,"OwnerUserId":9363390.0,"Title":"Output layer's shape is always (1,)","Body":"<p>I am trying to make a simple Keras model. But regardless of what output shape I am specifying the output layer is always of the shape <code>(1,)<\/code> so I cannot train my model because of the output layer and target data shapes mismatch.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import keras\nfrom keras.models import Sequential\nfrom keras.layers import InputLayer, LSTM, Dense\n\n# 63 is the number of unique characters\n# 128 is the length of a sequence of characters\n\nX = ... # X is an one-hot ndarray; X.shape == (96092, 128, 63)\nY = ... # Y is an one-hot ndarray; Y.shape == (96092, 63)\n\nmodel = Sequential()\nmodel.add(InputLayer([128, 63]))\nmodel.add(LSTM(96))\nmodel.add(Dense(63))\n\nmodel.compile(\n  optimizer=keras.optimizers.RMSprop(1e-3, decay=1e-5),\n  loss=keras.losses.sparse_categorical_crossentropy,\n)\n\nmodel.fit(X, Y) # ValueError: Error when checking target: expected dense_4 to have shape (1,) but got array with shape (63,)\n<\/code><\/pre>\n\n<p>As you can see, the shape of the output dense layer is <code>(1,)<\/code> but it has to be of shape <code>(63,)<\/code>. What am I doing wrong?<\/p>\n\n<p>I am using Google Colab with preinstalled Keras.<\/p>\n","answers":[{"AnswerId":"58068586","CreationDate":"2019-09-23T18:58:06.667","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>What that error is saying is that the shape of the output layer <em>is<\/em> 63. However, for some reason, it <em>expects<\/em> 1.<\/p>\n\n<p>In this case, the reason it expects 1 is because you are using <code>sparse_categorical_crossentropy<\/code>, which expects an integer representing the index of the output category. Instead, use <code>categorical_crossentropy<\/code>, which expects a one-hot encoding of the output category.<\/p>\n"}]}
+{"QuestionId":58068400,"AnswerCount":0,"Tags":"<tensorflow><gradient><sparse-matrix>","CreationDate":"2019-09-23T18:43:30.623","AcceptedAnswerId":null,"OwnerUserId":12044023.0,"Title":"How does Tensorflow compute gradients for sparse dense matmul?","Body":"<p>I want to implement a batch version of \"sparse_dense_matmul\" and did so successfully. Now I want to implement the gradients for it but am failing, as I do not even fully understand how they are calculated for non batch matmul (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/df3a3375941b9e920667acfe72fb4c33a8f45503\/tensorflow\/python\/ops\/sparse_grad.py#L149\" rel=\"nofollow noreferrer\">Source Code<\/a>).<\/p>\n\n<p>I'd like to know especially why there are <code>None<\/code> values in the return instead of an indices and a shape tensor and what the following lines of code are doing:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>parts_a = array_ops.gather(grad, rows)\nparts_b = array_ops.gather(b, cols)\na_values_grad = math_ops.reduce_sum(parts_a * parts_b, axis=1)\n<\/code><\/pre>\n\n<p>I simplified them a little as my input will never be transposed.<\/p>\n","answers":[]}
+{"QuestionId":58068564,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-23T18:56:14.687","AcceptedAnswerId":null,"OwnerUserId":6244166.0,"Title":"How does tf.Graph().control_dependencies works?","Body":"<p>I am trying to understand how does dependencies function works on graph. By going through the documentation <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Graph#control_dependencies\" rel=\"nofollow noreferrer\">here<\/a> I am trying to replicate something similar to this\n<a href=\"https:\/\/i.stack.imgur.com\/SP6W7.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/SP6W7.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I tried below code but tf raised an error:<\/p>\n\n<pre><code>import tensorflow as tf\na = tf.Variable(10)\nc = tf.Variable(20)\n\nmy_graph = tf.Graph()\n\nwith my_graph.control_dependencies([a, c]):\n  b = tf.Variable(2) + a + c\n  sess.run(tf.global_variables_initializer())\n  sess.run(b)\n<\/code><\/pre>\n\n<p>But this is giving me an error <strong>Tensor Tensor(\"Variable_0:0\", shape=(), dtype=int32_ref) is not an element of this graph.<\/strong><\/p>\n\n<p>As per my understanding the error is pointing out that both variable a and b should be on same graph to perform the execution.\nSo I tweaked my code as per my understanding and tried running <\/p>\n\n<pre><code>import tensorflow as tf\n\nmy_graph = tf.Graph()\n\nwith my_graph.control_dependencies([a, c]):\n  a = tf.Variable(10)\n  c = tf.Variable(20)\n  b = tf.Variable(2) + a + c\n  sess.run(tf.global_variables_initializer())\n  sess.run(b)\n<\/code><\/pre>\n\n<p>But this again gave me the same error. \n<strong>Tensor Tensor(\"Variable_0:0\", shape=(), dtype=int32_ref) is not an element of this graph.<\/strong><\/p>\n\n<p>Can anyone please tell me the correct way of creating graph dependencies ? \nCan we create a dependencies from one graph to dependencies into other graph ?\nNote : By going through <a href=\"https:\/\/stackoverflow.com\/questions\/55094952\/understanding-tensorflow-control-dependencies\">this<\/a> I have understood how tf.control_dependencies work.<\/p>\n","answers":[]}
+{"QuestionId":58069052,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2019-09-23T19:33:04.007","AcceptedAnswerId":null,"OwnerUserId":10453757.0,"Title":"Large test error in simple dense neural net while training with SGD but small test error when trained with GD","Body":"<p>I'm trying to write a simple neural net with one hidden layer to learn a simple classification problem. The neural net is as follows: the input is real vector of dimension d. A fully-connected first layer of size 2*k followed by LeakyRelu activation with alpha = 0.1. The output layer is of size one and is connected to all the hidden nodes. The weights in the last output layer are not trained and are fixed as follows: the first k weights (connected to top k hidden nodes) are fixed to 1\/sqrt(2k) and the last k weights (connected to the bottom k hidden nodes) are set to -1\/sqrt(2k). The output loss is mean hinge loss: 1\/n \\sum max(0, 1 - y_i * NN_output), where NN_output is the output of the neural network. <\/p>\n\n<p>I choose a W of size dX1 such that each coordinate is a uniform random number between (-0.5,0.5). Then, I set the coordinates in W* above 15 to be 0. Thus, in the W vector only the first 15 coordinates have non-zero values and rest are zero. Then, I generate ith training data as follows: for each vector x_i of dimension dx1 I randomly choose each coordinate to lie in (-0.5,0.5). Y_i for the training data is set to be sign(W*x). So basically, W is the true hyper-plane that I want my neural net to learn. <\/p>\n\n<p>I first train with GD (by writing explicit formula) and then with SGD. In both cases the training error can be driven to zero. However, I observe that when I train with GD, the test (hinge) loss (on randomly chosen dx1 test vectors in (-0.5,0.5)) goes to zero. However, with SGD the test (hinge) loss is always \n large, say around 1. Can anyone help me figure out why GD works and SGD doesn't work? When I look at \n the weight vector of each hidden node (this is dx1 vector) when trained by GD, it seems to be a scaled version of W. However, for SGD trained network, it is just random. So it is clear that SGD test error will be high since it didn't really learn that the weight vector should be focused only on first 15 coordinates. My question is why SGD is not able to learn as opposed to GD? Changing epochs and using other optimizers such as Adam etc. doesn't change the picture above. Thanks so much for reading my question and helping me understand what I'm doing wrong.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import keras\nimport numpy as np\n# Importing the required Keras modules containing model and layers\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.layers import LeakyReLU\n\nd = 1000                # ambient dimension\nd_main = 15             \nk = 200                 # number of node in hidden layer = 2k\nn_train = 500          # size of training data\nn_test = 100            # size of test data\nm = 5\neta = 0.01              # learning rate\nv = 1\/np.sqrt(2*k)      # mult fact for second layer\n\nwstar = np.random.uniform(-0.5,0.5,(d,1))\nwstar[d_main:wstar.shape[0]] = 0\nwstar = wstar\/np.linalg.norm(wstar,ord=2)\nv1 = v*np.ones((k,1))\nv2 = -v*np.ones((k,1))\nvf = np.concatenate((v1,v2))\n\nX_train = np.zeros((d,n_train))\n\nfor i in range(n_train):\n    tempx = np.random.uniform(-0.5,0.5,(d,1))\n    while np.abs(np.dot(np.transpose(wstar),tempx)) &lt; 1:\n        tempx = np.random.uniform(-0.5,0.5,(d,1))    \n    X_train[:,i] = np.reshape(tempx,d)*m*np.maximum(1.0\/m,np.random.uniform(0,1))\n\n\nY_train = np.transpose(np.sign(np.matmul(np.transpose(wstar),X_train)))\n\nx_train = X_train.reshape(X_train.shape[1],d)\ny_train = Y_train\n\nopt = keras.optimizers.SGD(lr=eta)\ndef create_model():\n    model = Sequential()    \n    model.add(Dense(2*k, activation=None,use_bias=False,input_shape=(d,)))\n    model.add(LeakyReLU(alpha=0.1))\n    model.add(Dense(1, activation=None,use_bias=False))\n    model.compile(optimizer=opt, \n                  loss='hinge', \n                  metrics=None)\n    return model\n\nmodel = create_model()\nmodel.summary()\n\n# set the second layer to be non-trainable so that it is not updated during training\nlayer = model.layers[2]\n# set the weights of the layer to be v*(1,1,...1,-1,-1,..,-1)\nlayer.set_weights([vf])\nlayer.trainable = False\nmodel.compile(optimizer=opt, \n                  loss='hinge', \n                  metrics=None)\n\nmodel.fit(x=x_train,y=y_train, epochs=20)\n\n# standard test error\nX_test = np.zeros((d,n_test))\n\nfor i in range(n_test):\n    tempx = np.random.uniform(-0.5,0.5,(d,1))\n    while np.abs(np.dot(np.transpose(wstar),tempx)) &lt; 1:\n        tempx = np.random.uniform(-0.5,0.5,(d,1))     \n    X_test[:,i] = np.reshape(tempx,d)*m*np.maximum(1.0\/m,np.random.uniform(0,1))\n\nY_test = np.transpose(np.sign(np.matmul(np.transpose(wstar),X_test)))\n\nx_test = X_test.reshape(X_test.shape[1],d)\ny_test = Y_test\n\n\nmodel.evaluate(x_test,y_test)\n<\/code><\/pre>\n\n<p>The GD program is below: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\nimport numpy as np\n\nd = 1000     # ambient dimension\nd_main = 15  \nk = 200      # number of node in hidden layer = 2k\nn = 500      # size of training data\nm = 5\nepoch = 20\neta = 0.01\nv = 1\/np.sqrt(2*k)\nalpha = 0.1  # leaky ReLU with param alpha\n\nwstar = np.random.uniform(-0.5,0.5,(d,1))\nwstar[d_main-1:wstar.shape[0]-1] = 0\nwstar = wstar\/np.linalg.norm(wstar,ord=2)\n\nX = np.zeros((d,n))\n\nfor i in range(n):\n    tempx = np.random.uniform(-0.5,0.5,(d,1))\n    while np.abs(np.dot(np.transpose(wstar),tempx)) &lt; 1:\n       tempx = np.random.uniform(-0.5,0.5,(d,1))    \n    X[:,i] = np.reshape(tempx,d)*m*np.maximum(1.0\/m,np.random.uniform(0,1))\n\nY = np.transpose(np.sign(np.matmul(np.transpose(wstar),X)))\n\nW = np.random.uniform(-0.5,0.5,(d,k))\nU = np.random.uniform(-0.5,0.5,(d,k))\n\ndef sigmaprim(y,alpha):\n    z = np.zeros(len(y))\n    for i in range(len(y)):\n        if y[i] &gt;= 0:\n            z[i] = 1\n        else:\n            z[i] = alpha\n    return z\n\nfor e in range(epoch):\n    print('epoch:=',e)         \n    W_old = W         \n    grad_W = np.zeros((d,k))\n    grad_U = np.zeros((d,k))         \n    for i in range(n):\n        x = X[:,i]\n        y = Y[i]         \n        if  y*((np.sum(np.dot(np.transpose(W),x) - np.dot(np.transpose(U),x)))*v) &lt; 1:\n            xmat = np.tile(np.reshape(x,(d,1)),(1,k))\n            b = sigmaprim(np.dot(np.transpose(W),x),alpha)\n            grad_W = grad_W - y*np.matmul(xmat,np.diag(b))\n            b = sigmaprim(np.dot(np.transpose(U),x),alpha)\n            grad_U = grad_U + y*np.matmul(xmat,np.diag(b))                                                                                                 \n        W = W - eta*grad_W\n        U = U - eta*grad_U                                                               \n        convergence_error = np.linalg.norm(W\/np.linalg.norm(W,ord=2)-W_old\/np.linalg.norm(W_old,ord=2),ord=2)\n\nerror = 0\nfor i in range(100):\n    x = np.random.uniform(-0.5,0.5,(d,1))\n    while np.abs(np.dot(np.transpose(wstar),x)) &lt; 1:\n        x = np.random.uniform(-0.5,0.5,(d,1))\n\n    y = np.dot(np.transpose(wstar),x)\n\n    if y*((np.sum(np.dot(np.transpose(W),x) - np.dot(np.transpose(U),x)))*v) &lt; 1:\n        error = error + 1;\n\nprint('test error:=', error\/100)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58069259,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><pip>","CreationDate":"2019-09-23T19:48:02.137","AcceptedAnswerId":null,"OwnerUserId":282855.0,"Title":"TensorFlow - ModuleNotFoundError: No module named 'numpy.core._multiarray_umath'","Body":"<p>Getting the <code>ModuleNotFoundError: No module named 'numpy.core._multiarray_umath'<\/code> as soon as the first epoch starts. All modules (i.e. <code>tensorflow<\/code>, <code>numpy<\/code>, etc.) are up-to-date.<\/p>\n\n<p>p.s. None of the answers posted <a href=\"https:\/\/stackoverflow.com\/questions\/54721703\/modulenotfounderror-no-module-named-numpy-core-multiarray-umath-while-insta\">here for another question<\/a> did not work for me. So I've created a new question.<\/p>\n\n<p><strong>Software Stack:<\/strong><\/p>\n\n<pre><code>OS: Windows 10 1903\nPython: 3.7.4\nTensorFlow: 1.14.0\nnumpy: 1.17.2\n<\/code><\/pre>\n\n<p><strong>Full error stacktrace:<\/strong><\/p>\n\n<pre><code>ModuleNotFoundError: No module named 'numpy.core._multiarray_umath'\nImportError: numpy.core.multiarray failed to import\nImportError: numpy.core._multiarray_umath failed to import\nImportError: numpy.core.umath failed to import\n2019-09-23 22:41:24.328882: F tensorflow\/python\/lib\/core\/bfloat16.cc:675] Check failed: PyBfloat16_Type.tp_base != nullptr \nUsing TensorFlow backend.\nModuleNotFoundError: No module named 'numpy.core._multiarray_umath'\nImportError: numpy.core.multiarray failed to import\nImportError: numpy.core._multiarray_umath failed to import\nImportError: numpy.core.umath failed to import\nException in thread Thread-72:\nTraceback (most recent call last):\n  File \"C:\\Users\\talha\\AppData\\Local\\Programs\\Python\\Python37\\lib\\threading.py\", line 926, in _bootstrap_inner\n    self.run()\n  File \"C:\\Users\\talha\\AppData\\Local\\Programs\\Python\\Python37\\lib\\threading.py\", line 870, in run\n    self._target(*self._args, **self._kwargs)\n  File \"D:\\PyCharmProjects\\fer\\venv\\lib\\site-packages\\keras\\utils\\data_utils.py\", line 572, in _run\n    with closing(self.executor_fn(_SHARED_SEQUENCES)) as executor:\n  File \"D:\\PyCharmProjects\\fer\\venv\\lib\\site-packages\\keras\\utils\\data_utils.py\", line 554, in &lt;lambda&gt;\n    initargs=(seqs,))\n  File \"C:\\Users\\talha\\AppData\\Local\\Programs\\Python\\Python37\\lib\\multiprocessing\\context.py\", line 119, in Pool\n    context=self.get_context())\n  File \"C:\\Users\\talha\\AppData\\Local\\Programs\\Python\\Python37\\lib\\multiprocessing\\pool.py\", line 176, in __init__\n    self._repopulate_pool()\n  File \"C:\\Users\\talha\\AppData\\Local\\Programs\\Python\\Python37\\lib\\multiprocessing\\pool.py\", line 241, in _repopulate_pool\n    w.start()\n  File \"C:\\Users\\talha\\AppData\\Local\\Programs\\Python\\Python37\\lib\\multiprocessing\\process.py\", line 112, in start\n    self._popen = self._Popen(self)\n  File \"C:\\Users\\talha\\AppData\\Local\\Programs\\Python\\Python37\\lib\\multiprocessing\\context.py\", line 322, in _Popen\n    return Popen(process_obj)\n  File \"C:\\Users\\talha\\AppData\\Local\\Programs\\Python\\Python37\\lib\\multiprocessing\\popen_spawn_win32.py\", line 89, in __init__\n    reduction.dump(process_obj, to_child)\n  File \"C:\\Users\\talha\\AppData\\Local\\Programs\\Python\\Python37\\lib\\multiprocessing\\reduction.py\", line 60, in dump\n    ForkingPickler(file, protocol).dump(obj)\nTypeError: can't pickle _thread.lock objects\n\nException in thread Thread-74:\nTraceback (most recent call last):\n  File \"C:\\Users\\talha\\AppData\\Local\\Programs\\Python\\Python37\\lib\\threading.py\", line 926, in _bootstrap_inner\n    self.run()\n  File \"C:\\Users\\talha\\AppData\\Local\\Programs\\Python\\Python37\\lib\\threading.py\", line 870, in run\n    self._target(*self._args, **self._kwargs)\n  File \"D:\\PyCharmProjects\\fer\\venv\\lib\\site-packages\\keras\\utils\\data_utils.py\", line 572, in _run\n    with closing(self.executor_fn(_SHARED_SEQUENCES)) as executor:\n  File \"D:\\PyCharmProjects\\fer\\venv\\lib\\site-packages\\keras\\utils\\data_utils.py\", line 554, in &lt;lambda&gt;\n    initargs=(seqs,))\n  File \"C:\\Users\\talha\\AppData\\Local\\Programs\\Python\\Python37\\lib\\multiprocessing\\context.py\", line 119, in Pool\n    context=self.get_context())\n  File \"C:\\Users\\talha\\AppData\\Local\\Programs\\Python\\Python37\\lib\\multiprocessing\\pool.py\", line 176, in __init__\n    self._repopulate_pool()\n  File \"C:\\Users\\talha\\AppData\\Local\\Programs\\Python\\Python37\\lib\\multiprocessing\\pool.py\", line 241, in _repopulate_pool\n    w.start()\n  File \"C:\\Users\\talha\\AppData\\Local\\Programs\\Python\\Python37\\lib\\multiprocessing\\process.py\", line 112, in start\n    self._popen = self._Popen(self)\n  File \"C:\\Users\\talha\\AppData\\Local\\Programs\\Python\\Python37\\lib\\multiprocessing\\context.py\", line 322, in _Popen\n    return Popen(process_obj)\n  File \"C:\\Users\\talha\\AppData\\Local\\Programs\\Python\\Python37\\lib\\multiprocessing\\popen_spawn_win32.py\", line 89, in __init__\n    reduction.dump(process_obj, to_child)\n  File \"C:\\Users\\talha\\AppData\\Local\\Programs\\Python\\Python37\\lib\\multiprocessing\\reduction.py\", line 60, in dump\n    ForkingPickler(file, protocol).dump(obj)\nTypeError: can't pickle _thread.lock objects\n\n2019-09-23 22:41:25.141774: F tensorflow\/python\/lib\/core\/bfloat16.cc:675] Check failed: PyBfloat16_Type.tp_base != nullptr \n<\/code><\/pre>\n","answers":[{"AnswerId":"58070949","CreationDate":"2019-09-23T22:33:01.237","ParentId":null,"OwnerUserId":"4870005","Title":null,"Body":"<p>I had the same problem. followed this link <a href=\"https:\/\/github.com\/numpy\/numpy\/issues\/12977\" rel=\"nofollow noreferrer\">https:\/\/github.com\/numpy\/numpy\/issues\/12977<\/a><\/p>\n\n<pre><code>pip uninstall numpy\nconda install numpy\n<\/code><\/pre>\n"}]}
+{"QuestionId":58069452,"AnswerCount":2,"Tags":"<python><neural-network><conv-neural-network><pytorch>","CreationDate":"2019-09-23T20:02:32.917","AcceptedAnswerId":58069815.0,"OwnerUserId":2703038.0,"Title":"Upsampling an autoencoder in pytorch","Body":"<p>I have defined my autoencoder in pytorch as following (it gives me a 8-dimensional bottleneck at the output of the encoder which works fine torch.Size([1, 8, 1, 1])):<\/p>\n\n<pre><code>self.encoder = nn.Sequential(\n    nn.Conv2d(input_shape[0], 32, kernel_size=8, stride=4),\n    nn.ReLU(),\n    nn.Conv2d(32, 64, kernel_size=4, stride=2),\n    nn.ReLU(),\n    nn.Conv2d(64, 8, kernel_size=3, stride=1),\n    nn.ReLU(),\n    nn.MaxPool2d(7, stride=1)\n)\n\nself.decoder = nn.Sequential(\n    nn.ConvTranspose2d(8, 64, kernel_size=3, stride=1),\n    nn.ReLU(),\n    nn.Conv2d(64, 32, kernel_size=4, stride=2),\n    nn.ReLU(),\n    nn.Conv2d(32, input_shape[0], kernel_size=8, stride=4),\n    nn.ReLU(),\n    nn.Sigmoid()\n)\n<\/code><\/pre>\n\n<p>What I cannot do is train the autoencoder with <\/p>\n\n<pre><code>def forward(self, x):\n    x = self.encoder(x)\n    x = self.decoder(x)\n    return x\n<\/code><\/pre>\n\n<p>The decoder gives me an error that the decoder cannot upsample the tensor:<\/p>\n\n<pre><code>Calculated padded input size per channel: (3 x 3). Kernel size: (4 x 4). Kernel size can't be greater than actual input size\n<\/code><\/pre>\n","answers":[{"AnswerId":"58069815","CreationDate":"2019-09-23T20:31:05.897","ParentId":null,"OwnerUserId":"10886420","Title":null,"Body":"<p>You are not upsampling enough via <code>ConvTranspose2d<\/code>, shape of your encoder is only <code>1<\/code> pixel (<code>width x height<\/code>), see this example:<\/p>\n\n<pre><code>import torch\n\nlayer = torch.nn.ConvTranspose2d(8, 64, kernel_size=3, stride=1)\nprint(layer(torch.randn(64, 8, 1, 1)).shape)\n<\/code><\/pre>\n\n<p>This prints your exact <code>(3,3)<\/code> shape after upsampling.<\/p>\n\n<p>You can:<\/p>\n\n<ul>\n<li>Make the kernel smaller - instead of <code>4<\/code> in first <code>Conv2d<\/code> in decoder use <code>3<\/code> or <code>2<\/code> or even <code>1<\/code><\/li>\n<li>Upsample more, for example: <code>torch.nn.ConvTranspose2d(8, 64, kernel_size=7, stride=2)<\/code> would give you <code>7x7<\/code><\/li>\n<li>What I would do personally: downsample less in encoder, so output shape after it is at least <code>4x4<\/code> or maybe <code>5x5<\/code>. If you squash your image so much there is no way to encode enough information into one pixel, and even if the code passes the network won't learn any useful representation.<\/li>\n<\/ul>\n"},{"AnswerId":"58071002","CreationDate":"2019-09-23T22:41:14.093","ParentId":null,"OwnerUserId":"2703038","Title":null,"Body":"<p>I have managed to implement an autoencoder that provides an unsupervised clustering (in my case 8 classes)<\/p>\n\n<p>This is not an expert solution. I owe thanks to @Szymon Maszke for the suggestions.<\/p>\n\n<pre><code>self.encoder = nn.Sequential(\n    nn.Conv2d(1, 32, kernel_size=8, stride=4),\n    nn.ReLU(),\n    nn.Conv2d(32, 64, kernel_size=4, stride=2),\n    nn.ReLU(),\n    nn.Conv2d(64, 2, kernel_size=3, stride=1),\n    nn.ReLU(),\n    nn.MaxPool2d(6, stride=1)\n)\n\nself.decoder = nn.Sequential(\n    nn.ConvTranspose2d(2, 64, kernel_size=3, stride=1),\n    nn.ReLU(),\n    nn.ConvTranspose2d(64, 32, kernel_size=8, stride=4),\n    nn.ReLU(),\n    nn.ConvTranspose2d(32, 1, kernel_size=8, stride=4)\n)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58069572,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-23T20:11:46.993","AcceptedAnswerId":58069981.0,"OwnerUserId":12092080.0,"Title":"Workaround for lack of broadcast in TFLite","Body":"<p>I would like to run a TFLite model that requires me to produce a 3d output (the sample code is a minimum example generating the error). Is there a tensorflow equivalent to gather_nd that does not reduce the dimension by one?<\/p>\n\n<p>I've tried looking through the documentation for related functions that I can think of and haven't found a good option.<\/p>\n\n<pre><code>import tensorflow.compat.v1 as tf\nimport numpy as np\n\ntf.disable_v2_behavior()\ninitial_input = tf.placeholder(dtype=tf.float32, shape=(None,5,1024))\ncap_i = tf.gather_nd(initial_input, [[0,1]]) #[0,2],[0,3],[0,4],[0,5]\ncap_i_broadcast = tf.broadcast_to(cap_i, [1,5,1024])\ncap_iT = tf.transpose(cap_i_broadcast, perm=[0,2,1])\n\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\ntf.io.write_graph(sess.graph_def, '', 'train.pbtxt')\nconverter = tf.lite.TFLiteConverter.from_session(sess, [initial_input], [cap_iT])\nconverter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS, tf.lite.OpsSet.SELECT_TF_OPS]\ntflite_model = converter.convert()\nopen('converted_model.tflite', \"wb\").write(tflite_model)\nsess.close()\n<\/code><\/pre>\n\n<p>Some of the operators in the model are not supported by the standard TensorFlow Lite runtime and are not recognized by TensorFlow. If you have a custom implementation for them you can disable this error with --allow_custom_ops, or by setting allow_custom_ops=True when calling tf.lite.TFLiteConverter(). Here is a list of builtin operators you are using: GATHER_ND, TRANSPOSE. Here is a list of operators for which you will need custom implementations: BroadcastTo.<\/p>\n","answers":[{"AnswerId":"58069981","CreationDate":"2019-09-23T20:46:15.770","ParentId":null,"OwnerUserId":"12092080","Title":null,"Body":"<p>The following code has a solution using strided slice with dimensionality reduction and then reshape to get back the correct dimension.<\/p>\n\n<pre><code>import tensorflow.compat.v1 as tf\nimport numpy as np\n\ntf.disable_v2_behavior()\ninitial_input = tf.placeholder(dtype=tf.float32, shape=(None,5,1024))\ncap_i = tf.strided_slice(initial_input, [0,0,0], [0,5,1024], [1,1,1], \nshrink_axis_mask=1)\ncap_i_reshaped =tf.reshape(cap_i,[1,5,1024])\ncap_iT = tf.transpose(cap_i_reshaped, perm=[0,2,1])\n\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\ntf.io.write_graph(sess.graph_def, '', 'train.pbtxt')\nconverter = tf.lite.TFLiteConverter.from_session(sess, [initial_input], \n[cap_iT])\nconverter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS, \ntf.lite.OpsSet.SELECT_TF_OPS]\ntflite_model = converter.convert()\nopen('converted_model.tflite', \"wb\").write(tflite_model)\nsess.close()\n<\/code><\/pre>\n\n<p>Previously thought slice was supported in TFLite but only strided_slice is.<\/p>\n"}]}
+{"QuestionId":58069629,"AnswerCount":1,"Tags":"<tensorflow><keras><conv-neural-network><lstm>","CreationDate":"2019-09-23T20:16:52.610","AcceptedAnswerId":58069816.0,"OwnerUserId":6283174.0,"Title":"How does the LSTM know number of time steps and features in an Conv1D-LSTM network?","Body":"<p>I have a time series signal (n samples, each sample has 81 time steps and 3 features = n x 81 x 3).\nI am using an conv1D-LSTM network.  n_timesteps = 81, n_features = 3.\nNormal LSTM specifies both n_timesteps and n_features, however when combined with conv1d, these are not specified.  <\/p>\n\n<ol>\n<li><p>How does the LSTM know how many time steps and features there are in the input to it?<\/p><\/li>\n<li><p>How does the LSTM know the end of the sequence for each sample?<\/p><\/li>\n<li><p>Are the time steps \"stored up\" and them fed into the LSTM or are the processed one time step at a time and fed into the LSTM one time step at a time?<\/p><\/li>\n<li><p>If I include the \"flatten\" (below) it fails.  Why?<\/p><\/li>\n<li><p>Do the number of filters in the conv1d have to correspond to the number of filters in the LSTM?<\/p><\/li>\n<\/ol>\n\n<pre><code>model = Sequential()\n\nmodel.add(Conv1D(filters=32, kernel_size=3, activation='relu', input_shape=(n_timesteps,n_features)))\nmodel.add(Conv1D(filters=32, kernel_size=3, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(MaxPooling1D(pool_size=2))\n#model.add(Flatten())\n\n#model.add(LSTM(units=128, input_shape=(n_timesteps, n_features), return_sequences=True))\nmodel.add(LSTM(units=128, return_sequences=True))\nmodel.add(Dropout(0.3))\nmodel.add(LSTM(units=64, dropout=0.5, recurrent_dropout=0.5, return_sequences=True))\nmodel.add(LSTM(units=32, dropout=0.5, recurrent_dropout=0.5))\nmodel.add(Dense(16, activation='relu'))\nmodel.add(Dropout(0.3))\nmodel.add(Dense(1, activation='sigmoid'))\n<\/code><\/pre>\n","answers":[{"AnswerId":"58069816","CreationDate":"2019-09-23T20:31:15.787","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<h1>1 and 2<\/h1>\n\n<p>Everything is based on tensors (sort of like matrices, but with any number of dimensions).   <\/p>\n\n<p>The tensors have shapes and everything is based on the shapes. Your data tensors are three-dimensional: <code>(samples, time_steps, features)<\/code>. <\/p>\n\n<p>It happens that 1D convolutions also use the same 3D tensors: <code>(samples, length, channels)<\/code>. So:<\/p>\n\n<ul>\n<li>samples = examples = sequences    <\/li>\n<li>time_steps = length    <\/li>\n<li>features = channels<\/li>\n<\/ul>\n\n<p>There is no secret. The data is structured and the layers will use this structure. Look at your <code>model.summary()<\/code> and see the number of steps and features for every layer's output.<\/p>\n\n<h1>3<\/h1>\n\n<p>There is no interleaving between layers. <\/p>\n\n<ul>\n<li>The conv layer will process its entire input tensor and generate an output tensor.     <\/li>\n<li>The next conv layer will take this entire output and produce another entire output    <\/li>\n<li>The LSTM layer will do the same, take an entire input and output an entire tensor.    <\/li>\n<\/ul>\n\n<h1>4<\/h1>\n\n<p>If you flatten the data, your 3D tensors <code>(samples, steps, feats)<\/code> will become 2D tensors <code>(samples, something)<\/code>. No more data structure that can be understood by the layers.    <\/p>\n\n<h1>5<\/h1>\n\n<p>There is absolutely no requirement for number of filters or units. The only thing is that the final output of your model needs to have the same shape of your <code>y_train<\/code> data. <\/p>\n\n<p>Here is my model summary.  It appears that the number of features has changed from the original 3 (of the input) to 32 (for the conv1d).  Is it correct that the LSTM will now process then entire time steps (~81) on the 32 features of the conv1d instead of the 3 features of the input?<\/p>\n\n<h1>Example of summary:<\/h1>\n\n<p>The first LSTM will take an input shape of <code>(None, 38,32)<\/code>. This means this LSTM will process:<\/p>\n\n<ul>\n<li>38 steps   <\/li>\n<li>32 features   <\/li>\n<\/ul>\n\n<p>The convolutions are discarding border steps and the maxpooling is halving the steps. <\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nconv1d (Conv1D)              (None, 79, 32)            320       \n_________________________________________________________________\nconv1d_1 (Conv1D)            (None, 77, 32)            3104      \n_________________________________________________________________\ndropout (Dropout)            (None, 77, 32)            0         \n_________________________________________________________________\nmax_pooling1d (MaxPooling1D) (None, 38, 32)            0         \n_________________________________________________________________\nlstm (LSTM)                  (None, 38, 128)           82432     \n_________________________________________________________________\ndropout_1 (Dropout)          (None, 38, 128)           0         \n_________________________________________________________________\nlstm_1 (LSTM)                (None, 38, 64)            49408     \n_________________________________________________________________\nlstm_2 (LSTM)                (None, 32)                12416     \n_________________________________________________________________\ndense (Dense)                (None, 16)                528       \n_________________________________________________________________\ndropout_2 (Dropout)          (None, 16)                0         \n_________________________________________________________________\ndense_1 (Dense)              (None, 1)                 17        \n=================================================================\nTotal params: 148,225\nTrainable params: 148,225\nNon-trainable params: 0\n_________________________________________________________________```\n\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":58069871,"AnswerCount":0,"Tags":"<python><arrays><tensorflow><base64><tensorflow-serving>","CreationDate":"2019-09-23T20:36:23.960","AcceptedAnswerId":null,"OwnerUserId":6476080.0,"Title":"Tensorflow Serving: How to base64 encode\/decode arrays that are NOT images?","Body":"<p>For each sample, I have a 2D array that is NOT an image that I would like to do inference on via tensorflow serving. In the past, I have been able to deploy tensorflow serving successfully thanks to the answer to <a href=\"https:\/\/stackoverflow.com\/questions\/51432589\/how-do-i-get-a-tensorflow-keras-model-that-takes-images-as-input-to-serve-predic\/51432590#51432590\">this post<\/a> which uses the following <code>serving_input_receiver_fn<\/code>:<\/p>\n\n<pre><code>HEIGHT = 199\nWIDTH = 199\nCHANNELS = 1\n\ndef serving_input_receiver_fn():\n\n  def decode_and_resize(image_str_tensor):\n     \"\"\"Decodes jpeg string, resizes it and returns a uint8 tensor.\"\"\"\n     image = tf.image.decode_jpeg(image_str_tensor, channels=CHANNELS)\n     image = tf.expand_dims(image, 0)\n     image = tf.image.resize_bilinear(\n         image, [HEIGHT, WIDTH], align_corners=False)\n     image = tf.squeeze(image, squeeze_dims=[0])\n     image = tf.cast(image, dtype=tf.uint8)\n     return image\n\n # Optional; currently necessary for batch prediction.\n key_input = tf.placeholder(tf.string, shape=[None]) \n key_output = tf.identity(key_input)\n\n input_ph = tf.placeholder(tf.string, shape=[None], name='image_binary')\n images_tensor = tf.map_fn(\n      decode_and_resize, input_ph, back_prop=False, dtype=tf.uint8)\n images_tensor = tf.image.convert_image_dtype(images_tensor, dtype=tf.float32) \n\n return tf.estimator.export.ServingInputReceiver(\n     {'images': images_tensor},\n     {'bytes': input_ph})\n<\/code><\/pre>\n\n<p>However, for nonimage arrays, the following becomes unclear:<\/p>\n\n<ol>\n<li>How to decode the encoded string tensor. I took a look at <code>tf.io.decode_image<\/code>, but it doesn't seem to preserve the 2D array dimension.<\/li>\n<li>How to encode the array. In the case of images, I encoded the image data itself by <code>base64.b64encode(img_data)<\/code>. For general 2D arrays, how should I encode them?<\/li>\n<\/ol>\n\n<p>In short what is the way to generalize the linked post's answer to the non image array case?<\/p>\n","answers":[]}
+{"QuestionId":58070557,"AnswerCount":0,"Tags":"<algorithm><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2019-09-23T21:42:39.573","AcceptedAnswerId":null,"OwnerUserId":5590255.0,"Title":"Why does Keras (tf) Binary Classification predict() always gives extreme probabilities, either zero or one?","Body":"<p>I'm trying implement a NN model (wide and deep learning) with tf.keras v2.0 to do some binary classification. I have ~50k training samples and ~10k testing samples. The class distribution is fairly even (see image below). \nMy training and testing accuracy varies between 0.6-0.7 depending on my hyper parameters. Precision and Recall are both around 0.7 as well for testing. <\/p>\n\n<p>My NN's last layer is a 2-unit Softmax and I'm using <code>categorical_crossentropy<\/code> as my loss function. <\/p>\n\n<p>After training, I try to get the probabilities for each prediction using <code>model.predict()<\/code>, but the results for each probability is almost always <code>[0., 1.]<\/code> or <code>[1., 0.]<\/code>. I'm wondering why does this happen? <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/1VETR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1VETR.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/WrFQC.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/WrFQC.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/jyszt.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/jyszt.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I've tried using a 1-unit Sigmoid + <code>binary_classification<\/code> or even train for 1 epoch, the results have been the same.<\/p>\n\n<p>Here's the simplified structure of the NN:<\/p>\n\n<p>Deep component is being fed with Embeddings of continuous features, then passes through 3 feed-forward layers with Dropouts in between. <\/p>\n\n<p>Wide component is a one layer logistic regression that's being fed with cross product features of categorical data. <\/p>\n\n<pre><code>out_layer = tf.keras.layers.concatenate([deep_component_outlayer, logistic_input])\noutput = tf.keras.layers.Dense(2, activation='softmax')(out_layer)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58070576,"AnswerCount":0,"Tags":"<tensorflow><distributed-computing>","CreationDate":"2019-09-23T21:44:27.317","AcceptedAnswerId":null,"OwnerUserId":6450489.0,"Title":"Is distributed learning feasible for my application?","Body":"<p>I am running a LSTM network to classify a huge database of remote sensing sequences to their respective landcovers. Code is developed on Tensorflow\/Keras and I am trying models with 3 to 7 layers, each having 16 to 64 LSTM cells. Due to huge data size (around 13.5M sample sequences), I am interested to do my model training in distributed mode by data parallelization over multiple GPUs or multiple PCs.  However, my tries by now have not resulted in any improvement over single system version and I am thinking if it is feasible at all or not.<\/p>\n\n<p>My understanding from data parallelism under Tensorflow\/Keras is that there are different workers, a chief, and a parameter server (ps). Model parameters are stored in ps and communicated with workers and chief, and each worker works on a partition of the whole data. Each worker should get the model parameters from ps at the start of each data batch processing, do all the forward and backward propagation and calculate the weight updates and send them back to ps, where the updated model parameters are calculated by chief and stored. Therefore, each worker should read the model parameters before each batch processing.<\/p>\n\n<p>My simulation batch size is 1024, therefore there are about 13K batches in one epoch of data. Training of a model with 87,000 parameters took 412 seconds in my one-gpu system, which means about 30 ms per batch. A more complex model with 1M parameters took 1900 seconds to process one epoch, which means about 146 ms per batch.\nNow that I look at the numbers, I am in doubt if data parallelism is helpful. If two GPUs or CPUs want to communicate 100K or 1M parameters, can it be so fast to be done in a few milliseconds to be negligible? My experiments on a two-GPU system (which even doesn't require data transmission over network) resulted in batch processing time of twice single-GPU. Can it be attributed to the small batch processing time compared to data communication overhead on top of that? If it is, does it mean that my model is not as complex as necessary for distributed learning?<\/p>\n","answers":[]}
+{"QuestionId":58070590,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2019-09-23T21:46:09.437","AcceptedAnswerId":null,"OwnerUserId":2739945.0,"Title":"Tensorflow-serving: how to pass text or do word tokenization into the model","Body":"<p>I'm begining with tensorflow-serving, my goal is to predict text classification, but I'm not having a way to pass the text into the signature as this, without previously tokenize the words\nit's possible to do it into the predictor?<\/p>\n\n<p>The goal to use tf serving is to expose the service to be used from other services without need intermediate transformations, because if need to use flask, them is more profitable build some prediction models with scikit<\/p>\n","answers":[]}
+{"QuestionId":58070762,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-23T22:07:49.117","AcceptedAnswerId":58138285.0,"OwnerUserId":2241766.0,"Title":"Reduce 1D tensor based on similar values","Body":"<p>I'm trying to implement the following <code>numpy<\/code> code in <code>tensorflow<\/code>.<\/p>\n\n<pre><code>x = np.array([1,1,1,1,1,2,2,2,3,3,3])\nx_u = np.unique(x)\nr_indices = []\nfor v in x_u:\n    v_indices = np.argwhere(x==v).flatten()\n    selected_indices = np.random.choice(v_indices, size=int(0.5 * v_indices.shape[0]), replace=False)\n    r_indices.append(selected_indices)\nr_indices = np.concatenate(r_indices)\nprint(r_indices)\n<\/code><\/pre>\n\n<p>The code outputs an array of 4 elements. Basically, it randomly selects <code>50%<\/code> of each unique sets (sets contain only <code>1<\/code>, <code>2<\/code> or <code>3<\/code>, and returns the corresponding indices.\nI tried with <code>tf.map_fn<\/code> but it doesn't work because this function outputs consistent tensors. I also tried with <code>tf.while_loop<\/code> but it was not successful.<\/p>\n\n<pre><code>indicators = tf.constant([1,1,1,1,1,2,2,2,3,3,3])\nu_indicators = tf.unique(indicators)\nta = tf.Variable([], dtype=tf.int32)\nnum_elements = tf.shape(u_indicators)[0]\n\ndef body_func(i, ta):\n    v = tf.gather(u_indicators, i)\n    indices = tf.where(tf.equal(v, indicators))\n    indices = tf.squeeze(indices) #1D tensor\n    idxs = tf.range(tf.shape(indices)[0])\n    num_selected = tf.cast(tf.cast(tf.shape(indices)[0], tf.float32) * 0.5, tf.int32)\n    ridxs = tf.random_shuffle(idxs)[:num_selected]\n    ta = tf.concat([ta, ridxs], axis=0)\n    return (i+1, ta)\n\ni = tf.constant(0)\ninit_state = (i, ta)\ncondition = lambda i, _: i &lt; num_elements\nn, ta_final = tf.while_loop(condition, body_func, init_state, shape_invariants=[i.get_shape(), tf.TensorShape([None])])\n# get the final result\nta_final_result = ta_final.stack()\n\n# run the graph\nwith tf.Session() as sess:\n    # print the output of ta_final_result\n    print(sess.run(ta_final_result))\n<\/code><\/pre>\n\n<p>The code above throws the error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"test.py\", line 23, in &lt;module&gt;\n    n, ta_final = tf.while_loop(condition, body_func, init_state, shape_invariants=[i.get_shape(), tf.TensorShape([None])])\n  File \"\/Users\/tiendh\/deeplearning\/lib\/python3.7\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line 3484, in while_loop\n    loop_vars, shape_invariants, expand_composites=False)\n  File \"\/Users\/tiendh\/deeplearning\/lib\/python3.7\/site-packages\/tensorflow\/python\/util\/nest.py\", line 304, in assert_same_structure\n    % (str(e), str1, str2))\nTypeError: The two structures don't have the same nested structure.\n\nFirst structure: type=tuple str=(&lt;tf.Tensor 'Const_1:0' shape=() dtype=int32&gt;, &lt;tf.Variable 'Variable:0' shape=(0,) dtype=int32_ref&gt;)\n\nSecond structure: type=list str=[TensorShape([]), TensorShape([Dimension(None)])]\n<\/code><\/pre>\n\n<p>Any suggestion?<\/p>\n","answers":[{"AnswerId":"58138285","CreationDate":"2019-09-27T16:23:27.117","ParentId":null,"OwnerUserId":"2241766","Title":null,"Body":"<p>Finally, I got the code fixed. Here is the solution:<\/p>\n\n<pre><code>import tensorflow as tf\n\nindicators = tf.constant([1,1,1,1,1,2,2,2,3,3,3])\nu_indicators,_ = tf.unique(indicators)\nta = tf.Variable([], dtype=tf.int64)\nnum_elements = tf.shape(u_indicators)[0]\n\ndef body_func(i, ta):\n    v = tf.gather(u_indicators, i)\n    indices = tf.where(tf.equal(v, indicators))\n    indices = tf.squeeze(indices) #1D tensor\n    num_selected = tf.cast(tf.cast(tf.shape(indices)[0], tf.float32) * 0.5, tf.int32)\n    ridxs = tf.random_shuffle(indices)[:num_selected] # 1D tensor\n    ta = tf.concat([ta, ridxs], axis=0) # 1D tensor\n    return (i+1, ta)\n\ni = tf.constant(0)\ninit_state = [i, ta]\ncondition = lambda i, _: i &lt; num_elements\nn, ta_final = tf.while_loop(condition, body_func, init_state, shape_invariants=[i.get_shape(), tf.TensorShape([None])])\n\n# run the graph\nwith tf.Session() as sess:\n    # print the output of ta_final_result\n    print(sess.run(ta_final))\n<\/code><\/pre>\n"}]}
+{"QuestionId":58070782,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-23T22:11:02.553","AcceptedAnswerId":null,"OwnerUserId":7803582.0,"Title":"Error importing Tensorflow using Keras on Jupyter Notebook","Body":"<p>I'm trying to use keras for a school project, but during the import I get an error in one of the Tensor Flow dependencies (I'm using the code from this guide <a href=\"https:\/\/github.com\/ml4a\/ml4a-guides\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ml4a\/ml4a-guides<\/a>).<\/p>\n\n<p>I already tried reinstalling TF, but it still fails. As i'm new to machine learning in general I'm not really sure what's wrong and hot to fix it.<\/p>\n\n<p>This is the output that jupiter notebook is showing:<\/p>\n\n<pre><code>Using TensorFlow backend.\n---------------------------------------------------------------------------\nImportError                               Traceback (most recent call last)\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/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\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/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\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/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\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/imp.py in load_module(name, file, filename, details)\n    241         else:\n--&gt; 242             return load_dynamic(name, filename, file)\n    243     elif type_ == PKG_DIRECTORY:\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/imp.py in load_dynamic(name, path, file)\n    341             name=name, loader=loader, origin=path)\n--&gt; 342         return _load(spec)\n    343 \n\nImportError: dlopen(\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorflow\/python\/_pywrap_tensorflow_internal.so, 6): Symbol not found: _SecKeyCopyExternalRepresentation\n  Referenced from: \/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorflow\/python\/..\/libtensorflow_framework.1.dylib\n  Expected in: \/System\/Library\/Frameworks\/Security.framework\/Versions\/A\/Security\n in \/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorflow\/python\/..\/libtensorflow_framework.1.dylib\n\nDuring handling of the above exception, another exception occurred:\n\nImportError                               Traceback (most recent call last)\n&lt;ipython-input-4-cb02c0ec08f1&gt; in &lt;module&gt;\n      1 import os\n----&gt; 2 import keras\n      3 from keras.preprocessing import image\n      4 from keras.applications.imagenet_utils import decode_predictions, preprocess_input\n      5 from keras.models import Model\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/keras\/__init__.py in &lt;module&gt;\n      1 from __future__ import absolute_import\n      2 \n----&gt; 3 from . import utils\n      4 from . import activations\n      5 from . import applications\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/keras\/utils\/__init__.py in &lt;module&gt;\n      4 from . import data_utils\n      5 from . import io_utils\n----&gt; 6 from . import conv_utils\n      7 from . import losses_utils\n      8 from . import metrics_utils\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/keras\/utils\/conv_utils.py in &lt;module&gt;\n      7 from six.moves import range\n      8 import numpy as np\n----&gt; 9 from .. import backend as K\n     10 \n     11 \n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/keras\/backend\/__init__.py in &lt;module&gt;\n----&gt; 1 from .load_backend import epsilon\n      2 from .load_backend import set_epsilon\n      3 from .load_backend import floatx\n      4 from .load_backend import set_floatx\n      5 from .load_backend import cast_to_floatx\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/keras\/backend\/load_backend.py in &lt;module&gt;\n     88 elif _BACKEND == 'tensorflow':\n     89     sys.stderr.write('Using TensorFlow backend.\\n')\n---&gt; 90     from .tensorflow_backend import *\n     91 else:\n     92     # Try and load external backend.\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/keras\/backend\/tensorflow_backend.py in &lt;module&gt;\n      3 from __future__ import print_function\n      4 \n----&gt; 5 import tensorflow as tf\n      6 from tensorflow.python.eager import context\n      7 from tensorflow.python.framework import device as tfdev\n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorflow\/__init__.py in &lt;module&gt;\n     26 \n     27 # pylint: disable=g-bad-import-order\n---&gt; 28 from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n     29 from tensorflow.python.tools import module_util as _module_util\n     30 \n\n\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/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\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/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 \"\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/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\/3.7\/lib\/python3.7\/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\/3.7\/lib\/python3.7\/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\/3.7\/lib\/python3.7\/imp.py\", line 242, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/imp.py\", line 342, in load_dynamic\n    return _load(spec)\nImportError: dlopen(\/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorflow\/python\/_pywrap_tensorflow_internal.so, 6): Symbol not found: _SecKeyCopyExternalRepresentation\n  Referenced from: \/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorflow\/python\/..\/libtensorflow_framework.1.dylib\n  Expected in: \/System\/Library\/Frameworks\/Security.framework\/Versions\/A\/Security\n in \/Library\/Frameworks\/Python.framework\/Versions\/3.7\/lib\/python3.7\/site-packages\/tensorflow\/python\/..\/libtensorflow_framework.1.dylib\n\n\nFailed to load the native TensorFlow runtime.\n\nSee https:\/\/www.tensorflow.org\/install\/errors\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":58070886,"AnswerCount":0,"Tags":"<tensorflow><nvidia><tensorflow-estimator>","CreationDate":"2019-09-23T22:25:21.060","AcceptedAnswerId":null,"OwnerUserId":10323938.0,"Title":"Manual mixed precision training using tf estimators multi gpu","Body":"<p>I want to use mixed precision training on multi gpu. In documentation, nvidia mentioned automatic mixed precision cannot be used for multi gpu as of now.So, I went through nvidia documentation on manual mixed precision training and it was written in low level tensorflow. They didn't give clarity and didn't have a good example on how to use manual mixed precision training using tf estimators for multi gpu's. Can we use manual mixed precision with multi gpu's? If yes, can we use it with tf estimators?<\/p>\n","answers":[]}
+{"QuestionId":58070936,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-23T22:31:35.733","AcceptedAnswerId":58070987.0,"OwnerUserId":947960.0,"Title":"Training from remote resources","Body":"<p>All,<\/p>\n\n<p>I've researched this some and haven't found a clear answer anywhere.<\/p>\n\n<p>Using Keras with TF backend, how can you train a model using assets (like images for example) that are not local, but remote assets. <\/p>\n\n<p>For example, if you have 1M images on s3 that are labeled but not organized by folder, is there a practical way to stream in data in a way Keras can use to train a model?<\/p>\n\n<p>My thinking is that I would supply a file that was of the format:<\/p>\n\n<pre><code>{ label: \"Apple\", img: http:\/\/someurl\/img.jpg }\n{ label: \"Banana\", img: http:\/\/someurl\/img.jpg }\n{ label: \"Orange\", img: http:\/\/someurl\/img.jpg }\n<\/code><\/pre>\n\n<p>You could use <code>preprocessing.load_img<\/code> or <code>pillow<\/code> to grab and resize the url. <\/p>\n\n<p>This question is more about the correct process for this and the feasibility?<\/p>\n","answers":[{"AnswerId":"58070987","CreationDate":"2019-09-23T22:38:52.650","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>This would be possible by mirroring Keras' generator API. You can make a standard python generator that has an index of image URLs, and yields batches of images loaded from those URLs.<\/p>\n\n<p>However, I would not recommend this approach. Loading images from the web introduces extra latency, which has the potential to significantly slow down your model's training. The only case where it might be a good idea is if you literally do not have the space on your SSD to store the whole dataset, and\/or you find that the time it takes to load a batch of images images is short compared to the time it takes to train on that batch.<\/p>\n"}]}
+{"QuestionId":58071102,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-23T22:54:24.503","AcceptedAnswerId":58071422.0,"OwnerUserId":2886575.0,"Title":"What is the pattern of tensorflow fill_triangular?","Body":"<p>I'm sure that there is some logic here, but I can't figure out what the general pattern is.  If I apply <code>tf.contrib.distributions.fill_triangular<\/code> to a length <code>(n+1)n\/2<\/code> list, then in what pattern do the entries fill the upper\/lower triangle of the resulting matrix?<\/p>\n\n<p>To illustrate, here is a 5x5 matrix filled by <code>fill_triangular<\/code>:<\/p>\n\n<pre><code>import tensorflow as tf\n\nfoo = tf.contrib.distributions.fill_triangular(\n    [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15], upper=True\n)\nwith tf.Session('') as sesh:\n    print(sesh.run(foo))\n<\/code><\/pre>\n\n<p>gives<\/p>\n\n<pre><code>[[ 1  2  3  4  5]\n [ 0  7  8  9 10]\n [ 0  0 13 14 15]\n [ 0  0  0 12 11]\n [ 0  0  0  0  6]]\n<\/code><\/pre>\n\n<p>OTOH, one might expect a similar behavior to <code>np.triu_indices(5)<\/code>, e.g.:<\/p>\n\n<pre><code>import numpy as np\n\nbar = np.zeros((5,5))\nbar[np.triu_indices(5)] = [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15]\nprint(bar)\n<\/code><\/pre>\n\n<p>which gives<\/p>\n\n<pre><code>array([[ 1.,  2.,  3.,  4.,  5.],\n       [ 0.,  6.,  7.,  8.,  9.],\n       [ 0.,  0., 10., 11., 12.],\n       [ 0.,  0.,  0., 13., 14.],\n       [ 0.,  0.,  0.,  0., 15.]])\n<\/code><\/pre>\n\n<p>What is the general fill pattern for <code>tf.contrib.distributions.fill_triangular(list(range(int((n+1)*n\/2))))<\/code>?<\/p>\n","answers":[{"AnswerId":"58071422","CreationDate":"2019-09-23T23:51:13.560","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>As the tf official doc mentioned, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/distributions\/fill_triangular\" rel=\"nofollow noreferrer\">fill_triangular<\/a> fills the elements in a clockwise spiral. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\nfoo = tf.contrib.distributions.fill_triangular(\n    np.arange(1,22), upper=True\n)\nwith tf.Session('') as sesh:\n    print(sesh.run(foo))\n\n#[[ 1-&gt;2-&gt;3-&gt;4-&gt; 5 -&gt; 6]   &gt;----&gt;-----&gt;-------&gt;--------|\n# [ 0  8-&gt;9-&gt;10-&gt;11-&gt; 12]   &gt;-------&gt;------&gt;-------|   |   \n# [ 0  0 15-&gt;16-&gt;17-&gt; 18]  |   &gt;---&gt;---&gt;-----|     |   |  \n# [ 0  0  0  21&lt;-20&lt;- 19]  |  |  &lt;---&lt;---&lt;---|     |   |\n# [ 0  0  0  0   14&lt;- 13]  |  &lt;-----&lt;------&lt;-------|   |\n# [ 0  0  0  0  0  7]]     &lt;----&lt;-----&lt;-------&lt;--------|\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":58071132,"AnswerCount":1,"Tags":"<python><keras><deep-learning><epoch><batchsize>","CreationDate":"2019-09-23T22:59:24.530","AcceptedAnswerId":58071519.0,"OwnerUserId":11780764.0,"Title":"How to modify model.fit settings?","Body":"<p>Hi My code used to run fine untill I changed my data set. Now, I am getting an error at :<\/p>\n\n<blockquote>\n<pre><code>model.fit(train,train_df.iloc[:,-1],epochs=30, batch_size=20, verbose=1)\n<\/code><\/pre>\n<\/blockquote>\n\n<p>The error is:<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking input: expected dense_31_input to have\n  shape (1125,) but got array with shape (103,)<\/p>\n<\/blockquote>\n\n<p>The variables are :<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Jxyei.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<pre><code>scaler = StandardScaler()\n\ntrain=scaler.fit_transform(train_df.iloc[:,:-1])\ntest=scaler.fit_transform(test_df.iloc[:,:-1])\n\n# Creating Deep Model\n\n\n\nmodel = Sequential()\n\n# Add an input layer\nmodel.add(Dense(562, activation='relu', input_shape=(1125,)))\n\n# Add one hidden layer\nmodel.add(Dense(562, activation='relu'))\nmodel.add(BatchNormalization())\n\n# Add an output layer\nmodel.add(Dense(1, activation='sigmoid'))\n\n#add improvements \n\nmodel.add(Dropout(0.3))\n#Train the model\n\nmodel.compile(loss='binary_crossentropy',optimizer='adam',metrics=['accuracy'])\n\nmodel.fit(train,train_df.iloc[:,-1],epochs=30, batch_size=20, verbose=1)\n\n#TEst the model\n\ny_pred = model.predict(test_df.iloc[:,:-1])\n<\/code><\/pre>\n\n<p>I assume to fix it. I need to change the batch_size and epochs? but what numbers should be used?<\/p>\n","answers":[{"AnswerId":"58071519","CreationDate":"2019-09-24T00:09:34.713","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>In general, models assume that the first dimension of the input data is the batch size. Models don't care what the batch size is though, so you never set it when creating the model. Instead, you should set <code>input_shape<\/code> to the shape of <em>each sample<\/em> of your input data. In your case, each sample appears to be a vector of length 103, so set <code>input_shape<\/code> to <code>(103,)<\/code>.<\/p>\n"}]}
+{"QuestionId":58071255,"AnswerCount":0,"Tags":"<tensorflow><dataset><feed>","CreationDate":"2019-09-23T23:22:01.823","AcceptedAnswerId":null,"OwnerUserId":8510077.0,"Title":"how to fit tensorflow dataset","Body":"<p>I would like to feed model using tensorflow dataset iterator methods. However, I am not sure how to proceed that. Any suggestion will be highly appreciated. Thanks.<\/p>\n\n<pre><code>batch_size=10\ntf_X_train=tf.placeholder(tf.float32, shape=[None, 410,1,10])\ntf_Y_train=tf.placeholder(tf.float32, shape=[None])\ntrain_dataset = tf.data.Dataset.from_tensor_slices((tf_X_train, tf_Y_train))\ntrain_dataset = train_dataset.batch(batch_size)\n\niterator = tf.data.Iterator.from_structure(train_dataset.output_types, train_dataset.output_shapes)\n\ndata_X, data_y = iterator.get_next()\n\ntrain_iterator = iterator.make_initializer(train_dataset)\n\nwith tf.Session() as sess:\n    tf.global_variables_initializer()\n    learning_rate=0.0001\n    EPOCHS = 200\n    optimizer = tf.train.AdamOptimizer(learning_rate, 0.99)\n    model = cnn_model_fn(learning_rate)\n    model.compile(loss='mean_squared_error',\n                    optimizer=optimizer,\n                    metrics=['mean_absolute_error', 'mean_squared_error'])\nmodel.fit_generator(train_iterator,epochs=EPOCHS,steps_per_epoch=32,callbacks=[PrintDot()])\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58071570,"AnswerCount":0,"Tags":"<keras><terminal><sublimetext3>","CreationDate":"2019-09-24T00:18:15.590","AcceptedAnswerId":null,"OwnerUserId":9115361.0,"Title":"I can import keras from terminal, but I cannot from sublime text even though I am running the same python script?","Body":"<p>I have a simple python script that imports keras. I am using an anaconda env. I have the sublime text set up on the python I am using (3.6). When I run the script from the terminal, it works. When I run it from sublime, I get this:<\/p>\n\n<p>Traceback (most recent call last):\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 58, in \n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in \n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/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\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/imp.py\", line 242, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/imp.py\", line 342, in load_dynamic\n    return _load(spec)\nImportError: libcublas.so.10.0: 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):\n  File \"\/home\/squall\/Desktop\/delete.py\", line 2, in \n    import keras\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/site-packages\/keras\/<strong>init<\/strong>.py\", line 3, in \n    from . import utils\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/site-packages\/keras\/utils\/<strong>init<\/strong>.py\", line 6, in \n    from . import conv_utils\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/site-packages\/keras\/utils\/conv_utils.py\", line 9, in \n    from .. import backend as K\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/site-packages\/keras\/backend\/<strong>init<\/strong>.py\", line 1, in \n    from .load_backend import epsilon\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/site-packages\/keras\/backend\/load_backend.py\", line 90, in \n    from .tensorflow_backend import *\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 5, in \n    import tensorflow as tf\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/site-packages\/tensorflow\/<strong>init<\/strong>.py\", line 28, in \n    from tensorflow.python import pywrap_tensorflow  # pylint: disable=unused-import\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/site-packages\/tensorflow\/python\/<strong>init<\/strong>.py\", line 49, in \n    from tensorflow.python import pywrap_tensorflow\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 74, in \n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 58, in \n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/site-packages\/tensorflow\/python\/pywrap_tensorflow_internal.py\", line 28, in \n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/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\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/imp.py\", line 242, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/imp.py\", line 342, in load_dynamic\n    return _load(spec)\nImportError: libcublas.so.10.0: 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 <a href=\"https:\/\/www.tensorflow.org\/install\/errors\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/errors<\/a><\/p>\n\n<p>for some common reasons and solutions.  Include the entire stack trace\nabove this error message when asking for help.\n[Finished in 0.3s with exit code 1]\n[cmd: ['\/home\/squall\/anaconda3\/envs\/thunder\/bin\/python', '-u', '\/home\/squall\/Desktop\/delete.py']]<\/p>\n\n<p>Even weirder, sublime runs the script when I import numpy. This bug seems to pop up only when I import keras. Any ideas?<\/p>\n","answers":[]}
+{"QuestionId":58071808,"AnswerCount":0,"Tags":"<python><flask><keras>","CreationDate":"2019-09-24T01:06:24.590","AcceptedAnswerId":null,"OwnerUserId":10947056.0,"Title":"Keep a model initialized in flask","Body":"<p>I have a problem I feel will be easy to fix but I have no idea, I am fairly new to flask. <\/p>\n\n<p>I want to have a flask app that allows for an image to be uploaded by the user and tested on a specially trained keras model for detecting cat breeds. <\/p>\n\n<p>The code works already to preform the prediction without using flask, and I can get it to run if the model is initialized right before the predication is run. <\/p>\n\n<p>My goal is this : How do I get the model to stay initialized so that it can be run any number of times without having to re-initialize it each time?<\/p>\n\n<p>Here is the following code:<\/p>\n\n<pre><code>import os\n#import magic\nimport urllib.request\nfrom app import app\nfrom flask import Flask, flash, request, redirect, render_template\nfrom werkzeug.utils import secure_filename\nimport tensorflow as tf\nimport numpy as np\nimport requests\nfrom tensorflow.keras.preprocessing import image\nfrom tensorflow.keras.preprocessing.image import load_img\nfrom tensorflow.keras.models import load_model\nfrom tensorflow.keras.applications import xception\nfrom PIL import Image\n\nALLOWED_EXTENSIONS = set(['txt', 'pdf', 'png', 'jpg', 'jpeg', 'gif'])\n\nclass CatClass:\n    def __init__(self):\n        from tensorflow.python.lib.io import file_io\n        model_file = file_io.FileIO('gs:\/\/modelstorageforcats\/94Kitty.h5', mode='rb')\n\n        temp_model_location = '.\/94Kitty.h5'\n        temp_model_file = open(temp_model_location, 'wb')\n        try:\n            temp_model_file.write(model_file.read())\n            temp_model_file.close()\n            model_file.close()\n        except:\n            raise(\"Issues with getting the model\")\n        # get model\n        self.catModel = load_model(temp_model_location)\n\n    def catEstimator(self, catImage):\n        script_dir = os.path.dirname(__file__) #&lt;-- absolute dir the script is in\n        rel_path = \"uploads\/\" + catImage\n        abs_file_pathk = os.path.join(script_dir, rel_path)\n\n        #get picture the proper size for xception\n        try:\n            kittyPic = image.load_img(abs_file_pathk, target_size=(299,299))\n            x = xception.preprocess_input(np.expand_dims(kittyPic.copy(), axis=0))\n        except:\n            raise(\"Error with the images\")\n\n        #cat names the way the model learned it\n        catNames = [\"Bengal\",\"Abyssinian\",\"BritishShorthair\",\"Birman\",\"Sphynx\",\"Bombay\",\"EgyptianMau\",\"Persian\",\"Ragdoll\",\"MaineCoon\",\"Siamese\",\"RussianBlue\",\"AmericanBobtail\",\"DevonRex\",\"AmericanCurl\",\"DonSphynx\",\"Manx\",\"Balinese\",\"Burmilla\",\"Burmese\",\"KhaoManee\",\"Chausie\",\"AmericanShortHair\",\"Chartreux\",\"Pixiebob\",\"JapaneseBobtail\",\"BritishLonghair\",\"CornishRex\",\"Tabby\",\"Somali\",\"ExoticShortHair\",\"Tonkinese\",\"OrientalShortHair\",\"Minskin\",\"Korat\",\"Savannah\",\"Havana\",\"Singapura\",\"Nebelung\",\"OrientalLonghair\",\"TurkishAngora\",\"ScottishFold\",\"KurilianBobtail\",\"Lykoi\",\"ScottishFoldLonghair\",\"Ocicat\",\"Munchkin\",\"SelkirkRex\",\"AustralianMist\",\"AmericanWireHair\",\"TurkishVan\",\"SnowShoe\",\"Peterbald\",\"Siberian\",\"Toybob\",\"Himalayan\",\"LePerm\",\"NorwegianForestCat\"]\n\n        prediction = (self.catModel.predict(x))\n\n        label = int(np.argmax(prediction, axis=-1))\n\n        return(catNames[label])\n\ncatter = CatClass()\n\n\ndef allowed_file(filename):\n    return '.' in filename and filename.rsplit('.', 1)[1].lower() in ALLOWED_EXTENSIONS\n\n@app.route('\/')\ndef upload_form():\n    return render_template('upload.html')\n\n@app.route('\/', methods=['POST'])\ndef upload_file():\n    if request.method == 'POST':\n        # check if the post request has the file part\n        if 'file' not in request.files:\n            flash('No file part')\n            return redirect(request.url)\n        file = request.files['file']\n        if file.filename == '':\n            flash('No file selected for uploading')\n            return redirect(request.url)\n        if file and allowed_file(file.filename):\n            filename = secure_filename(file.filename)\n            file.save(os.path.join(app.config['UPLOAD_FOLDER'], filename))\n            flash('File successfully uploaded')\n            flash(catter.catEstimator(filename))\n            return redirect('\/')\n        else:\n            flash('Allowed file types are txt, pdf, png, jpg, jpeg, gif')\n            return redirect(request.url)\n\nif __name__ == \"__main__\":\n    app.run()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58071836,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><regression>","CreationDate":"2019-09-24T01:12:15.953","AcceptedAnswerId":null,"OwnerUserId":12101119.0,"Title":"Neural Network \u2013 Predicting Values of Multiple Variables","Body":"<p>I have data with columns A, B, C as inputs and columns D, E, F, G as outputs. The table has a shape (1000,7). I would like to train the model, validate and test it.<\/p>\n\n<p><strong><em>My data:<\/em><\/strong><\/p>\n\n<pre><code>A = [100, 120, 140, 160, 180, 200, 220, 240, 260, 280];\nB = [300, 320, 340, 360, 380, 400, 420, 440, 460, 480];\nC = [500, 520, 540, 560, 580, 600, 620, 640, 660, 680]; \n<\/code><\/pre>\n\n<p><strong><em>My desired outcome:<\/em><\/strong><\/p>\n\n<p>For each combination of A, B, C --> I get D, E, F, G as outputs <em>(for example)<\/em>:<\/p>\n\n<pre><code>D = 2.846485609 \nE = 5.06656901\nF = 3.255358183\nG = 5.464482379)\n<\/code><\/pre>\n\n<p>Also, for each different combination of A, B, C; I have a different set of outputs (D, E, F, G). <\/p>\n\n<p><strong><em>My Question:<\/em><\/strong>\nIs it possible to train a neural network, using this experienced network to predict new values of D, E, F, G; for new combination of A, B, C? <\/p>\n","answers":[{"AnswerId":"58074223","CreationDate":"2019-09-24T06:25:21.517","ParentId":null,"OwnerUserId":"5136550","Title":null,"Body":"<p>The problem falls into Multivariate Regression category since the outputs are continuous value. Therefore, you can train a neural network (NN) having 4 output nodes and input feature vector of size 4. A sample NN model having one hidden layer using tensorfow is as follows:<\/p>\n\n<pre><code>import itertools\nimport numpy as np\nfrom sklearn.preprocessing import StandardScaler\nfrom tensorflow.python.keras.layers import Input, Dense\nfrom tensorflow.python.keras.models import Model\n\nA = [100, 120, 140, 160, 180, 200, 220, 240, 260, 280]\nB = [300, 320, 340, 360, 380, 400, 420, 440, 460, 480]\nC = [500, 520, 540, 560, 580, 600, 620, 640, 660, 680]\n\nX_train = np.array(list(itertools.product(A, B, C)))\n# X_train = np.random.random(size=(1000,3))\nscaler = StandardScaler()\nX = scaler.fit_transform(X_train)\n\nY_train = np.random.randint(0, 100, size=(1000, 4)).astype(float)  # Should load original label\n\nX_test = np.random.random(size=(100, 3))\nY_test = np.random.randint(0, 100, size=(100, 4)).astype(float)\n\ninput = Input(shape=(3,))\nhidden_layer_1 = Dense(25, activation='relu')(input)\noutput = Dense(4)(hidden_layer_1)\n\nmodel = Model(inputs=input, outputs=output)\nmodel.compile(\n    optimizer='adam',\n    loss=['mean_squared_error']\n)\n\nhistory = model.fit(X_train, Y_train, epochs=1000, batch_size=8)\n\nresult = model.predict(X_test)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58071893,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-24T01:23:31.413","AcceptedAnswerId":null,"OwnerUserId":11674803.0,"Title":"version `GLIBC_2.23' not found and Failed to load the native TensorFlow runtime","Body":"<p>I want to install tensorflow1.13.1 in Anaconda virtual environment with python3.7 in Ubuntu.<\/p>\n\n<p>I have tried <code>pip install tensorflow==1.13.1<\/code> or <code>pip3 install tensorflow==1.13.1<\/code> or <code>conda install=1.13.1<\/code>\uff0cbut the terminal throw the error \"version `GLIBC_2.23' not found\" and \"Failed to load the native TensorFlow runtime\".<\/p>\n\n<p>Anyone can help me to resolve this problem? <\/p>\n","answers":[]}
+{"QuestionId":58071917,"AnswerCount":1,"Tags":"<tensorflow><keras><tensorflow2.0>","CreationDate":"2019-09-24T01:27:09.913","AcceptedAnswerId":58072987.0,"OwnerUserId":3594531.0,"Title":"Tensorflow 2.0 Keras won't save best model even with validation data, still gives me: Can save best model only with val_acc available, skipping","Body":"<p>I provide both training and validation data as Dataset iterators during training by <code>fit<\/code> function. I want to save best model by validation dataset accuracy. However, during training at each epoch, I get warning like <code>Can save best model only with val_acc, skipping.<\/code><\/p>\n\n<p>I compiled the model with <code>metrics = ['accuracy']<\/code>, declared <code>monitor='val_acc'<\/code> in callback, and provided validation data as Dataset iterator like <code>validation_data=my_val_dataset<\/code> in <code>fit<\/code>. What else am I missing here? <\/p>\n\n<p>I would really appreciate any help. <\/p>\n\n<p>My Jupyter notebook is in <a href=\"https:\/\/github.com\/shinchan75034\/testrep\/blob\/master\/tf20_dist_train.ipynb\" rel=\"nofollow noreferrer\">my GitHub repo<\/a>. <\/p>\n","answers":[{"AnswerId":"58072987","CreationDate":"2019-09-24T04:09:49.457","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Make sure <code>val_acc<\/code> exists in the logs.   <\/p>\n\n<p>Get the history from a short training:    <\/p>\n\n<pre><code>histCallback = model.fit(smallBatch_X, smallBatch_Y, epochs = 1)\n<\/code><\/pre>\n\n<p>Check what is in it:<\/p>\n\n<pre><code>for key in histCallback.history:\n    print(key)\n<\/code><\/pre>\n\n<p>Maybe you will find <code>val_accuracy<\/code> instead. Or maybe you have multiple outputs and will find <code>val_output_2_acc<\/code> or something in a similar style.<\/p>\n"}]}
+{"QuestionId":58072054,"AnswerCount":0,"Tags":"<tensorflow><distributed><tensorflow-estimator><parameter-server>","CreationDate":"2019-09-24T01:52:34.053","AcceptedAnswerId":null,"OwnerUserId":1944413.0,"Title":"TensorFlow Estimator: how to do prediction when using parameter server?","Body":"<p>TensorFlow Estimator is easy to use for distributed training with parameter server strategy. But I cannot do prediction with the parameter server strategy. I cannot find any resource to introduce the part.<\/p>\n\n<p>prediction sample code:<\/p>\n\n<pre><code>    run_config = tf.estimator.RunConfig()\n    model = tf.estimator.Estimator(\n        model_fn=self.model_fn,\n        model_dir=self._config.model_path,\n        config=run_config,\n        params=self.params())\n    results = model.predict(\n        input_fn=lambda: test_data.build(\n            batch_size=self._config.eval_batch_size,\n            num_epochs=1))\n<\/code><\/pre>\n\n<p>TF_CONFIG:<\/p>\n\n<pre><code>{'task': {'index': '0', 'type': 'ps'}, 'cluster': {'chief': ['127.0.0.1:2320'], 'ps': ['127.0.0.1:2220', '127.0.0.1:2221']}}\n{'task': {'index': '1', 'type': 'ps'}, 'cluster': {'chief': ['127.0.0.1:2320'], 'ps': ['127.0.0.1:2220', '127.0.0.1:2221']}}\n{'task': {'index': '0', 'type': 'chief'}, 'cluster': {'chief': ['127.0.0.1:2320'], 'ps': ['127.0.0.1:2220', '127.0.0.1:2221']}}\n<\/code><\/pre>\n\n<p>Result: Both PS and Woker did prediction.<\/p>\n\n<p>Any suggestion? Thanks a lot.<\/p>\n","answers":[]}
+{"QuestionId":58072114,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorflow-serving><resnet>","CreationDate":"2019-09-24T02:02:04.867","AcceptedAnswerId":58082057.0,"OwnerUserId":6476080.0,"Title":"Exporting VGGFace ResNet to Tensorflow Serving: ValueError: The last dimension of the inputs to `Dense` should be defined. Found `None`","Body":"<p>I have a code to load VGGFace pretrained ResNet model and export it to the tensorflow serving format. The snippet has worked on similar models that are not VGGFace ResNet. <\/p>\n\n<pre><code>import tensorflow as tf\nimport keras\nfrom keras_vggface.vggface import VGGFace\n\ntf.keras.backend.set_learning_phase(0) \nCHANNELS = 3\n\nmodel = VGGFace(model='resnet50')\nmodel.compile(\n    loss=keras.losses.categorical_crossentropy,\n    optimizer='adam',\n    metrics=['accuracy'])\nmodel.save('models\/VGGFaceResnet50_celebrity_classifier.h5')\n\ndef serving_input_receiver_fn():\n\n    def decode_and_resize(image_str_tensor):\n        \"\"\"Decodes jpeg string, resizes it and raeturns a uint8 tensor.\"\"\"\n        image = tf.image.decode_jpeg(image_str_tensor, channels=CHANNELS)\n        image = tf.cast(image, dtype=tf.uint8)\n        return image\n\n    # Optional; currently necessary for batch prediction.\n    key_input = tf.placeholder(tf.string, shape=[None]) \n    key_output = tf.identity(key_input)\n\n    input_ph = tf.placeholder(tf.string, shape=[None], name='image_binary')\n    images_tensor = tf.map_fn(\n            decode_and_resize, input_ph, back_prop=False, dtype=tf.uint8)\n    images_tensor = tf.image.convert_image_dtype(images_tensor, dtype=tf.float32) \n\n    return tf.estimator.export.ServingInputReceiver(\n         {'input_1': images_tensor},\n         {'bytes': input_ph})\n\n\nKERAS_MODEL_PATH='models\/VGGFaceResnet50_celebrity_classifier.h5'\nEXPORT_PATH='serving_model'\n\n# If you are invoking this from your training code, use `keras_model=model` instead.\nestimator = tf.keras.estimator.model_to_estimator(\n    keras_model_path=KERAS_MODEL_PATH)\nestimator.export_savedmodel(\n    EXPORT_PATH,\n    serving_input_receiver_fn=serving_input_receiver_fn\n) \n<\/code><\/pre>\n\n<p>However, I get the following error:<\/p>\n\n<pre><code>ValueError                  Traceback (most recent call last)\n&lt;ipython-input-1-edfbc77c298a&gt; in &lt;module&gt;\n     44 estimator.export_savedmodel(\n     45     EXPORT_PATH,\n---&gt; 46     serving_input_receiver_fn=serving_input_receiver_fn\n     47 ) \n...\nValueError: The last dimension of the inputs to `Dense` should be defined. Found `None`.\n\n<\/code><\/pre>\n\n<p>I checked the model summary and I don't see any <code>None<\/code> in the last dimension. Why is there an error?<\/p>\n","answers":[{"AnswerId":"58082057","CreationDate":"2019-09-24T14:02:17.780","ParentId":null,"OwnerUserId":"1849975","Title":null,"Body":"<p>Constructor of the <code>VGGFace<\/code> is defined as: <code>def VGGFace(include_top=True, model='vgg16', weights='vggface', input_tensor=None, input_shape=None, pooling=None, classes=None)<\/code>. The size of the last dimension is computed based on the <code>input_shape<\/code>, i.e. if it is <code>None<\/code> last dimension size will be <code>None<\/code> as well.<\/p>\n"}]}
+{"QuestionId":58072185,"AnswerCount":1,"Tags":"<deep-learning><pytorch><image-segmentation><unity3d-unet><semantic-segmentation>","CreationDate":"2019-09-24T02:13:02.990","AcceptedAnswerId":null,"OwnerUserId":6642287.0,"Title":"Pytorch:Apply cross entropy loss with custom weight map","Body":"<p>I am solving multi-class segmentation problem using u-net architecture in pytorch.\nAs specified in <a href=\"https:\/\/arxiv.org\/abs\/1505.04597\" rel=\"nofollow noreferrer\">U-NET<\/a> paper, I am trying to implement custom weight maps to counter class imbalances.<\/p>\n\n<p>Below is the opertion which I want to apply - \n<a href=\"https:\/\/i.stack.imgur.com\/Erp6K.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Erp6K.png\" alt=\"image\"><\/a><\/p>\n\n<p>Also, I reduced the <code>batch_size=1<\/code> so that I can remove that dimension while passing it to <code>precompute_to_masks<\/code> function.\nI tried the below approach-<\/p>\n\n<pre><code>def precompute_for_image(masks):\n    masks = masks.cpu()\n    cls = masks.unique()\n    res = torch.stack([torch.where(masks==cls_val, torch.tensor(1), torch.tensor(0)) for cls_val in cls])\n    return res\n\ndef train(n_epochs, loaders, model, optimizer, criterion, use_cuda, save_path):\n\n        ###################\n        # train the model #\n        ###################\n        model.train()\n        for batch_idx, (data, target) in enumerate(final_train_loader):\n            # move to GPU\n            if use_cuda:\n                data, target = data.cuda(), target.cuda()\n            optimizer.zero_grad()\n            output = model(data)\n            temp_target = precompute_for_image(target)\n            w = weight_map(temp_target)\n            loss = criterion(output,target)\n            loss = w*loss\n            loss.backward()\n            optimizer.step()\n\n            train_loss = train_loss + ((1 \/ (batch_idx + 1)) * (loss.data - train_loss))\n\n    return model\n<\/code><\/pre>\n\n<p>where weight_map is the function to calculate weight mask which I got from <a href=\"https:\/\/jaidevd.github.io\/posts\/weighted-loss-functions-for-instance-segmentation\/\" rel=\"nofollow noreferrer\">here<\/a>\nThe issue, I am facing is I am getting <code>memory error<\/code> when I apply the following method. <br>I am using 61gb RAM and Tesla V100 GPU.\nI really think I am applying it in incorrect way.\nHow to do it?\n<br>I am omitting the non-essential details from the training loop.\nBelow is my <code>weight_map<\/code> function:<\/p>\n\n<pre><code>from skimage.segmentation import find_boundaries\n\nw0 = 10\nsigma = 5\n\ndef make_weight_map(masks):\n    \"\"\"\n    Generate the weight maps as specified in the UNet paper\n    for a set of binary masks.\n\n    Parameters\n    ----------\n    masks: array-like\n        A 3D array of shape (n_masks, image_height, image_width),\n        where each slice of the matrix along the 0th axis represents one binary mask.\n\n    Returns\n    -------\n    array-like\n        A 2D array of shape (image_height, image_width)\n\n    \"\"\"\n    nrows, ncols = masks.shape[1:]\n    masks = (masks &gt; 0).astype(int)\n    distMap = np.zeros((nrows * ncols, masks.shape[0]))\n    X1, Y1 = np.meshgrid(np.arange(nrows), np.arange(ncols))\n    X1, Y1 = np.c_[X1.ravel(), Y1.ravel()].T\n    for i, mask in enumerate(masks):\n        # find the boundary of each mask,\n        # compute the distance of each pixel from this boundary\n        bounds = find_boundaries(mask, mode='inner')\n        X2, Y2 = np.nonzero(bounds)\n        xSum = (X2.reshape(-1, 1) - X1.reshape(1, -1)) ** 2\n        ySum = (Y2.reshape(-1, 1) - Y1.reshape(1, -1)) ** 2\n        distMap[:, i] = np.sqrt(xSum + ySum).min(axis=0)\n    ix = np.arange(distMap.shape[0])\n    if distMap.shape[1] == 1:\n        d1 = distMap.ravel()\n        border_loss_map = w0 * np.exp((-1 * (d1) ** 2) \/ (2 * (sigma ** 2)))\n    else:\n        if distMap.shape[1] == 2:\n            d1_ix, d2_ix = np.argpartition(distMap, 1, axis=1)[:, :2].T\n        else:\n            d1_ix, d2_ix = np.argpartition(distMap, 2, axis=1)[:, :2].T\n        d1 = distMap[ix, d1_ix]\n        d2 = distMap[ix, d2_ix]\n        border_loss_map = w0 * np.exp((-1 * (d1 + d2) ** 2) \/ (2 * (sigma ** 2)))\n    xBLoss = np.zeros((nrows, ncols))\n    xBLoss[X1, Y1] = border_loss_map\n    # class weight map\n    loss = np.zeros((nrows, ncols))\n    w_1 = 1 - masks.sum() \/ loss.size\n    w_0 = 1 - w_1\n    loss[masks.sum(0) == 1] = w_1\n    loss[masks.sum(0) == 0] = w_0\n    ZZ = xBLoss + loss\n    return ZZ\n<\/code><\/pre>\n\n<p>Traceback of the error-<\/p>\n\n<pre><code>MemoryError                               Traceback (most recent call last)\n&lt;ipython-input-30-f0a595b8de7e&gt; in &lt;module&gt;\n      1 # train the model\n      2 model_scratch = train(20, final_train_loader, unet, optimizer, \n----&gt; 3                       criterion, train_on_gpu, 'model_scratch.pt')\n\n&lt;ipython-input-29-b481b4f3120e&gt; in train(n_epochs, loaders, model, optimizer, criterion, use_cuda, save_path)\n     24             loss = criterion(output,target)\n     25             target.requires_grad = False\n---&gt; 26             w = make_weight_map(target)\n     27             loss = W*loss\n     28             loss.backward()\n\n&lt;ipython-input-5-e75a6281476f&gt; in make_weight_map(masks)\n     33         X2, Y2 = np.nonzero(bounds)\n     34         xSum = (X2.reshape(-1, 1) - X1.reshape(1, -1)) ** 2\n---&gt; 35         ySum = (Y2.reshape(-1, 1) - Y1.reshape(1, -1)) ** 2\n     36         distMap[:, i] = np.sqrt(xSum + ySum).min(axis=0)\n     37     ix = np.arange(distMap.shape[0])\n\nMemoryError:\n<\/code><\/pre>\n","answers":[{"AnswerId":"58080145","CreationDate":"2019-09-24T12:16:54.187","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Your <code>final_train_loader<\/code> provides you with an input image <code>data<\/code> and the expected pixel-wise labeling <code>target<\/code>. I assume (following pytorch's conventions) that <code>data<\/code> is of shape B-3-H-W and of <code>dtype=torch.float<\/code>.<br>\nMore importantly, <code>target<\/code> is of shape B-H-W and of <code>dtype=torch.long<\/code>.<\/p>\n\n<p>On the other hand <code>make_weight_map<\/code> expects its input to be C-H-W (with <code>C<\/code> = number of classes, NOT batch size), of type numpy array. <\/p>\n\n<p>Try providing <code>make_weight_map<\/code> the input mask <em>as it expects it<\/em> and see if you get similar errors.<br>\nI also recommend that you <em>visualize<\/em> the resulting weight map - to make sure your function does what you expect it to do.<\/p>\n"}]}
+{"QuestionId":58072263,"AnswerCount":1,"Tags":"<python><tensorflow><keras><loss-function>","CreationDate":"2019-09-24T02:26:40.930","AcceptedAnswerId":58072931.0,"OwnerUserId":7155684.0,"Title":"Create keras tensor with shape as same as model output for custom loss function","Body":"<p>I have a keras model with output shape of the last layer is <code>(None,574,6)<\/code>, which <code>None<\/code> is my batch size feed into the model.<\/p>\n\n<p>I also have a 2D numpy array called <code>anchors<\/code> with shape <code>(574,6)<\/code>.<\/p>\n\n<p>What I want is my output of each data minus that numpy array element wise.<\/p>\n\n<pre><code>import keras.backend as K\n\nanchor_tensor = K.cast(anchors, tf.float32)\nprint(K.int_shape(anchor_tensor))\n#(576, 4)\nprint(K.int_shape(y_pred))\n#(None, 574, 6)\ny_pred - anchor_tensor\n<\/code><\/pre>\n\n<p>The above code occured the following error due to the batch_size is unknown.<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError: Dimensions must be equal, but are 574 and 576\n  for 'sub_6' (op: 'Sub') with input shapes: [?,574,6], [576,4].<\/p>\n  \n  <p>During handling of the above exception, another exception occurred:<\/p>\n<\/blockquote>\n\n<p>How can I repeat <code>anchor_tensor<\/code> <code>None<\/code> times to make its shape as the same as <code>y_pred<\/code>?<\/p>\n","answers":[{"AnswerId":"58072931","CreationDate":"2019-09-24T04:03:00.080","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Tensorflow will easily do what it calls \"broadcasting\", which is automatically repeating the missing elements if possible. But for this to happen, it must confirm that the shapes allow that first. <\/p>\n\n<p>The safest way to assure the shapes are compatible, is to make them have the same length, and have value 1 in the dimension you want it to repeat.<\/p>\n\n<p>So, it's as simple as:<\/p>\n\n<pre><code>anchor_tensor = K.expand_dims(anchor_tensor, axis=0) #new shape is (1, 576, 4)   \nresult = y_pred - anchor tensor\n<\/code><\/pre>\n\n<p>Now Tensorflow can match the shapes and will repeat the tensor for the entire batch size.    <\/p>\n"}]}
+{"QuestionId":58072347,"AnswerCount":0,"Tags":"<c++><tensorflow>","CreationDate":"2019-09-24T02:37:53.770","AcceptedAnswerId":null,"OwnerUserId":9873377.0,"Title":"Calling custom op in tensorflow using c++","Body":"<p>I built a very simple custom op <code>zero_out<\/code> and try to run it using c++, the question is how to call this op, there is no doc about it.<\/p>\n\n<p>code of custom op <code>zero_out<\/code><\/p>\n\n<pre><code>#include \"tensorflow\/core\/framework\/op.h\"\n#include \"tensorflow\/core\/framework\/shape_inference.h\"\n#include \"tensorflow\/core\/framework\/op_kernel.h\"\nusing namespace tensorflow;\n\nREGISTER_OP(\"ZeroOut\")\n.Input(\"to_zero: float\")\n.Output(\"zeroed: float\")\n.SetShapeFn([](::tensorflow::shape_inference::InferenceContext* c) {\nc-&gt;set_output(0, c-&gt;input(0));\nreturn Status::OK();\n});\n\nclass ZeroOutOp : public OpKernel {\npublic:\n    explicit ZeroOutOp(OpKernelConstruction* context) : OpKernel(context) {}\n\n    void Compute(OpKernelContext* context) override {\n        const Tensor&amp; input_tensor = context-&gt;input(0);\n        auto input = input_tensor.flat&lt;float&gt;();\n        Tensor* output_tensor = NULL;\n        OP_REQUIRES_OK(context, context-&gt;allocate_output(0, input_tensor.shape(),                                                     &amp;output_tensor));\n        auto output_flat = output_tensor-&gt;flat&lt;float&gt;();\n        const int N = input.size();\n        for (int i = 1; i &lt; N; i++)\n            output_flat(i) = 0;\n        if (N &gt; 0) output_flat(0) = input(0);\n    }\n};\n\nREGISTER_KERNEL_BUILDER(Name(\"ZeroOut\").Device(DEVICE_CPU), ZeroOutOp);\n<\/code><\/pre>\n\n<p>After building, I got <code>zero_out.so<\/code>, then I load the lib using c++<\/p>\n\n<pre><code>    TF_Status* status_load = TF_NewStatus();\n    TF_Library* lib_handle = TF_LoadLibrary(\"libzero_out.so\", status_load);\n    TF_Code code = TF_GetCode(status_load);\n    cout &lt;&lt; \"code: \" &lt;&lt; code &lt;&lt; endl;  \/\/ output: 0\n\n    TF_Buffer op_list_buf = TF_GetOpList(lib_handle);\n    tensorflow::OpList op_list;\n    op_list.ParseFromArray(op_list_buf.data, op_list_buf.length);\n    cout &lt;&lt; \"oplist size = \" &lt;&lt; op_list.op_size() &lt;&lt; endl;    \/\/ output: 1\n    cout &lt;&lt; \"oplist name = \" &lt;&lt; op_list.op(0).name() &lt;&lt; endl;    \/\/ output: ZeroOut\n<\/code><\/pre>\n\n<p>I think the op should have been registered, then how to invoke the op function, Could you give some examples or code snippet?<\/p>\n","answers":[]}
+{"QuestionId":58072362,"AnswerCount":1,"Tags":"<keras><shapes><batchsize>","CreationDate":"2019-09-24T02:39:11.913","AcceptedAnswerId":null,"OwnerUserId":12110525.0,"Title":"keras compute_output_shape not working for Custom layer","Body":"<p>I customized a layer, merged the batch_size and the first dimension, the other dimensions remained unchanged, but compute_output_shape seemed to have no effect, resulting in the subsequent layer could not get accurate shape information, resulting in an error. How do I make compute_output_shape work?<\/p>\n\n<pre><code>import keras\nfrom keras import backend as K\n\nclass BatchMergeReshape(keras.layers.Layer):\n    def __init__(self, **kwargs):\n        super(BatchMergeReshape, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        super(BatchMergeReshape, self).build(input_shape)  \n\n    def call(self, x):\n        input_shape = K.shape(x)\n        batch_size, seq_len = input_shape[0], input_shape[1]\n        r = K.reshape(x, (batch_size*seq_len,)+input_shape[2:])\n        print(\"call_shape:\",r.shape)\n        return r\n\n    def compute_output_shape(self, input_shape):\n        if input_shape[0] is None:\n            r = (None,)+input_shape[2:]\n            print(\"compute_output_shape:\",r)\n            return r\n        else:\n            r = (input_shape[0]*input_shape[1],)+input_shape[2:]\n            return r\n\na = keras.layers.Input(shape=(3,4,5))\nb = BatchMergeReshape()(a)\nprint(b.shape)\n\n# call_shape: (?, ?)\n# compute_output_shape: (None, 4, 5)\n# (?, ?)\n<\/code><\/pre>\n\n<p>I need to get (None,4,5) but get (None,None),why compute_output_shape didn't work. My keras version is 2.2.4<\/p>\n","answers":[{"AnswerId":"58072900","CreationDate":"2019-09-24T03:58:52.340","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>The problem is probably that <code>K.shape<\/code> returns a <strong>tensor<\/strong> not a tuple. You can't do <code>(batch_size*seq_len,) + input_shape[2:]<\/code>. This is mixing a lot of things, tensors and tuples, the result will certainly be wrong.    <\/p>\n\n<p>Now the good thing is that, if you know the other dimensions and just not the batch size, you just need this layer:<\/p>\n\n<pre><code>Lambda(lambda x: K.reshape(x, (-1,) + other_dimensions_tuple))\n<\/code><\/pre>\n\n<p>If you don't then:<\/p>\n\n<pre><code>input_shape = K.shape(x)\nnew_batch_size = input_shape[0:1] * input_shape[1:2] #needs to keep a shape of an array   \n                 #new_batch_size.shape = (1,)\nnew_shape = K.concatenate([new_batch_size, input_shape[2:]]) #this is a tensor   \n                                                             #result of concatenating 2 tensors   \n\nr = K.reshape(x, new_shape)\n<\/code><\/pre>\n\n<p>Notice that this works in Tensorflow, but may not work in Theano. <\/p>\n\n<p>Notice also that Keras will demand that the batch size of the model's output is equal to the batch size of the model's inputs. This means that you will need to restore the original batch size before the end of the model.   <\/p>\n"}]}
+{"QuestionId":58072607,"AnswerCount":2,"Tags":"<keras><neural-network><lstm><recurrent-neural-network><lstm-stateful>","CreationDate":"2019-09-24T03:18:13.603","AcceptedAnswerId":null,"OwnerUserId":11617135.0,"Title":"Dense expects 2d but has got 3d in LSTM","Body":"<p>In my model <\/p>\n\n<pre><code>Xtrain shape : (62, 30, 100)\nYtrain shape : (62, 1, 100)\nXtest shape : (16, 30, 100)\nYtest shape : (16, 1, 100)\n<\/code><\/pre>\n\n<p>When I build my model,<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(units=100, return_sequences= True, input_shape=(x_train.shape[1],X_train.shape[2])))\nmodel.add(LSTM(units=100, return_sequences=True))\nmodel.add(Dense(units=100))\n\nmodel.fit(x_train,y_train,epochs=5,batch_size=13)\n<\/code><\/pre>\n\n<p>when I try to fit it throws a error,<\/p>\n\n<pre><code>ValueError: Error when checking target: expected dense_1 to have 2 dimensions, but got array with shape (62, 1, 100)\n<\/code><\/pre>\n\n<p>I need to predict for the next 1 time stamp for all 100 features.\nWhat are the changes needed to be done?<\/p>\n","answers":[{"AnswerId":"58072833","CreationDate":"2019-09-24T03:49:37.523","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>The code posted doesn't seem to be the same that generated the error. <\/p>\n\n<p>Print your <code>model.summary()<\/code>. You will see:<\/p>\n\n<ul>\n<li>LSTM 1: (None, 30, 100)   <\/li>\n<li>LSTM 2: (None, 30, 100)    <\/li>\n<li>Dense: (None, 30, 100)<\/li>\n<\/ul>\n\n<p>You didn't use anything to reduce the number of time steps to 1. Your error message, according to this model should definitely be complaining about trying <code>(None, 30, 100)<\/code> vs <code>(62, 1, 100)<\/code>.   <\/p>\n\n<p>To eliminate the timesteps, you need <code>return_sequences=False<\/code> in the last LSTM, so your model becomes:   <\/p>\n\n<ul>\n<li>(None, 30, 100)    <\/li>\n<li>(None, 100)    <\/li>\n<li>(None, 100) <\/li>\n<\/ul>\n\n<p>This way, you can have <code>Ytrain.shape == (62,100)<\/code>  <\/p>\n\n<p>If you really really need that middle dimension == 1, just use <code>Lambda(lambda x: K.expand_dims(x, 1))<\/code> after the dense. <\/p>\n"},{"AnswerId":"58124693","CreationDate":"2019-09-26T20:57:47.923","ParentId":null,"OwnerUserId":"2182857","Title":null,"Body":"<p>If you want to have the result of each LSTM step to be processed on its own (the most common use of placing a Dense layer after an LSTM or other RNNs), you need to wrap it, like so:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(units=100, return_sequences= True, input_shape=(x_train.shape[1],X_train.shape[2])))\nmodel.add(LSTM(units=100, return_sequences=True))\nmodel.add(TimeDistributed(Dense(units=100)))\n<\/code><\/pre>\n\n<p>Each output will be served on its own to the Dense layer (of course, it would be the same layer - all weights will be shared between each 'instance' of it).<\/p>\n"}]}
+{"QuestionId":58072977,"AnswerCount":1,"Tags":"<keras><coreml>","CreationDate":"2019-09-24T04:08:04.547","AcceptedAnswerId":58097482.0,"OwnerUserId":12110827.0,"Title":"Can't convert Keras model to Coreml","Body":"<p>I have tried various ways to convert my Keras model to core ml using core ml tools, but it gives me this error.<\/p>\n\n<blockquote>\n  <p>Keras layer '' not supported.<\/p>\n<\/blockquote>\n\n<p>I am trying to convert .h5 model to core ml so that I can use it in my app but it gives me some errors which I am not able to solve. Also, I have tried converting .h5 model to PB(frozen graph) but got errors over there. <\/p>\n\n<p>This is how my model looks.<\/p>\n\n<pre><code>img_input = layers.Input(shape=(224, 224, 3))\n\n\nseed = 230\n\nnumpy.random.seed(seed)\n\n\nx = layers.Conv2D(16, 3, activation='relu')(img_input)\n\nx = layers.MaxPooling2D(2)(x)\n\nx = layers.Conv2D(32, 3, activation='relu')(x)\n\nx = layers.MaxPooling2D(2)(x)\n\nx = layers.Flatten()(x)\n\nx = layers.Dense(128, activation='relu')(x)\n\nx = layers.Dropout(0.4)(x)\n\noutput = layers.Dense(3, activation='softmax')(x)\n\nmodel = Model(img_input, output)\n<\/code><\/pre>\n\n<p>model.compile(loss='sparse_categorical_crossentropy', optimizer='adam', metrics=['accuracy'])<\/p>\n\n<p>this is the code i found over the web to convert Keras model to core ml tools.<\/p>\n\n<p>import keras \nimport coremltools <\/p>\n\n<p>fcn_mlmodel = coremltools.converters.keras.convert( model, input_names = 'image', image_input_names = 'image', output_names = 'class_label' ) <\/p>\n\n<p>fcn_mlmodel.input_description['image']=\"Image size (224,224,3)\" <\/p>\n\n<p>fcn_mlmodel.output_description['class_label']=\" Class label\" <\/p>\n\n<p>fcn_mlmodel.save(\"Test_my.mlmodel\") <\/p>\n\n<p>Err: Keras layer '' not supported. \u2013<\/p>\n","answers":[{"AnswerId":"58097482","CreationDate":"2019-09-25T11:36:37.980","ParentId":null,"OwnerUserId":"10171211","Title":null,"Body":"<p>Can't recreate your problem, copied everything. Maybe it's a problem with your versions:<\/p>\n\n<p><code>pip install -U coremltools==3.0b6 tensorflow==1.13.1 keras==2.2.4<\/code> works good together. <\/p>\n"}]}
+{"QuestionId":58072978,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm><dimension>","CreationDate":"2019-09-24T04:08:07.287","AcceptedAnswerId":58078240.0,"OwnerUserId":10761367.0,"Title":"Model fit and dimension size error for kera LSTM network in python","Body":"<p>Hi I have been working on an LSTM network in python using Keras.  I created a 1D array for my training and test set.  When I try to fit the model I get the following error: <\/p>\n\n<p>ValueError: Error when checking input: expected lstm_31_input to have 3 dimensions, but got array with shape (599, 1)<\/p>\n\n<p>I have tried resizing the dimension and the add(Flatten) layer.  None of this work. My code is below: <\/p>\n\n<pre><code>#Setup\nimport pandas as pd\nimport numpy as np\nfrom numpy import array, zeros, newaxis\nfrom numpy import argmax\nfrom keras.layers.core import Dense, Activation, Dropout\nfrom keras.preprocessing import sequence\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Embedding, Flatten\nfrom keras.layers import LSTM\n\n#Used to ignore warning about some of the tensor command being depracated\n#Code from: \n#https:\/\/stackoverflow.com\/questions\/43819820\/how-to-disable-keras-warnings\n#import warnings\n#with warnings.catch_warnings():\n#    warnings.simplefilter(\"ignore\")\n\n\n\"\"\"\n#Allow use of modules from the Common_Functions Folder\nimport sys\nsys.path.append('..\/_Common_Functions')\nimport Hello_World as helloWorld\n\"\"\"\n\n#Creates dataset of random numbers\n#import numpy as np\nfrom random import random\ndef generateDatset(n):\n    val = np.array([])\n    typ = np.array([])\n    for i in range (1, n):\n        val = np.append(val, round(random()*10, 2))\n\n        if val[i-1] &lt; 3 or val[i-1] &gt; 7:\n            typ = np.append(typ, 'm')\n        else:\n            typ = np.append(typ, 'f')\n    return val, typ\n\n\n# Encode the output labels\ndef lable_encoding(gender_series):\n    labels = np.empty((0, 2))\n    for i in gender_series:\n        if i == 'm':\n            labels = np.append(labels, [[1,0]], axis=0)\n        else:\n            labels = np.append(labels, [[0,1]], axis=0)\n    return labels\n\n#Gets dataset in proper format for this program\nval, typ = generateDatset(1000)\ndf = pd.DataFrame( {\"first_name\": val[:], \"gender\": typ[:]} )\n\n# Split dataset in 60% train, 20% test and 20% validation\ntrain, validate, test = np.split(df.sample(frac=1), [int(.6*len(df)), int(.8*len(df))])\n\n# Convert both the input names as well as the output lables into the discussed machine readable vector format\ntrain_x = np.asarray(train.first_name)\n#train_x = np.reshape(train_x, train_x.shape + (1,))\n#train_x = np.reshape(train_x, (train_x.shape[0], 1, train_x.shape[1]))\n\ntrain_y = lable_encoding(train.gender)\n#train_y = np.reshape(train_y, train_y.shape + (1,))\n#train_y = np.reshape(train_y, (train_y.shape[0], 1, train_y.shape[1]))\n\nvalidate_x =  np.asarray(validate.first_name)\n#validate_x = np.reshape(validate_x, validate_x.shape + (1,))\nvalidate_y = lable_encoding(validate.gender)\n#validate_y = np.reshape(validate_y, validate_y.shape + (1,))\n\ntest_x =  np.asarray(test.first_name)\n#test_x = np.reshape(test_x, test_x.shape + (1,))\ntest_y = lable_encoding(test.gender)\n#test_x = np.reshape(test_x, test_x.shape + (1,))\n\n\"\"\"\nThe number of hidden nodes can be determined by the following equation: \nNh = (Ns\/ (alpha * Ni + No ) )\nWhere  Ni --&gt; number of input neurons\n       No --&gt; number of output neurons \n       Ns --&gt; number of samples\n       alph --&gt; scaling factor\n\nAlternatively the following equation can be used: \n    Nh = (2\/3)*(Ni + No)\nAs a not this equation is simpler but may not provide the best performance\n\n\"\"\"\n#Set a value for the scaling factor. \n#This typically ranges between 2 and 10\nalpha = 8\nhidden_nodes = int(np.size(train_x) \/ (alpha * ((len(df.columns)-1)+ 4)))\n\ninput_length = train_x.shape # Length of the character vector\noutput_labels = 2 # Number of output labels\n\nfrom keras import optimizers\n# Build the model\nprint('Building model...')\nmodel = Sequential()\n\n#print(train_x.shape)\n#\ndf = np.expand_dims(df, axis=2)\n\n\nmodel.add(LSTM(hidden_nodes, return_sequences=True, input_shape=(599, 1)))\n\nmodel.add(Dropout(0.2))\n\nmodel.add(Flatten())\nmodel.add(Dense(units=output_labels))\n\nmodel.add(Activation('softmax'))\nsgd = optimizers.SGD(lr=0.5, clipnorm=10.)\nmodel.compile(loss='categorical_crossentropy', optimizer= sgd, metrics=['acc'])\n#\nbatch_size=1000\n\n#x = train_x[..., newaxis, newaxis]\n#x.shape\n#y = train_y[..., newaxis, newaxis]\n#y.shape\nmodel.fit(train_x, train_y, batch_size=batch_size, epochs=10)\n\n#http:\/\/45.76.113.195\/?questions\/46616674\/expected-ndim-3-found-ndim-2-how-to-feed-sparse-matrix-to-lstm-layer-in-keras\n<\/code><\/pre>\n","answers":[{"AnswerId":"58078240","CreationDate":"2019-09-24T10:29:14.223","ParentId":null,"OwnerUserId":"6070667","Title":null,"Body":"<p><code>input_shape=(599, 1)<\/code> specifies the shape of one sample.<\/p>\n\n<p>In here your train batch size is 599 and shape of one sample is 1. Since the first layer is a LSTM layer it needs 3 dimensional input(batch_size,number_of_time_stamps_of_a_sample,diamentionality_of_one_time_stamp).<strong>But we don't mention the batch size in input_shape<\/strong>.So input shape of a LSTM layer should be <\/p>\n\n<p><strong>input_shape=(number_of_time_stamps_of_a_sample,diamentionality_of_one_time_stamp)<\/strong><\/p>\n\n<p>So u should <\/p>\n\n<p>1)Replace <code>input_shape=(599, 1)<\/code> with <code>input_shape=(1, 1)<\/code><\/p>\n\n<p>2)Add the following line before training <code>train_x=train_x.reshape(599,1,1)<\/code><\/p>\n\n<p>To be more clear please refer to this <a href=\"https:\/\/www.youtube.com\/watch?v=LZ-GS7LOyWs&amp;list=PLKaJlRMgdZumQc42kQy61v3woqs-3jpmt\" rel=\"nofollow noreferrer\">video<\/a> uploaded by me ;-)<\/p>\n"}]}
+{"QuestionId":58073032,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-09-24T04:17:07.660","AcceptedAnswerId":null,"OwnerUserId":7867162.0,"Title":"Same input results in different prediction in two times","Body":"<p>I read an image from file and call predict method of Keras Inception v3 model. And I found two different results from one input.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.applications.inception_v3 import InceptionV3, decode_predictions\nfrom keras.preprocessing import image\nimport numpy as np\n\ndef model():\n    model = InceptionV3(weights='imagenet')\n    def predict(x):\n        x *=  2\n        x -= 1\n        return model.predict(np.array([x]))[0]\n    return predict\n\nimg = image.load_img(\"2.jpg\", target_size=(299, 299))\nimg = image.img_to_array(img)\nimg \/= 255.\n\np = model()\n\nprint('Predicted:', decode_predictions(np.array([p(img)]), top=3)[0])\nprint('Predicted:', decode_predictions(np.array([p(img)]), top=3)[0])\n<\/code><\/pre>\n\n<p>The output is<\/p>\n\n<pre><code>Predicted: [('n01443537', 'goldfish', 0.98162466), ('n02701002', 'ambulance', 0.0010537759), ('n01440764', 'tench', 0.00027527584)]\nPredicted: [('n02606052', 'rock_beauty', 0.69015616), ('n01990800', 'isopod', 0.039278224), ('n01443537', 'goldfish', 0.03365362)]\n<\/code><\/pre>\n\n<p>where the first result is correct.<\/p>\n","answers":[{"AnswerId":"58074635","CreationDate":"2019-09-24T06:55:21.040","ParentId":null,"OwnerUserId":"7217806","Title":null,"Body":"<p>You are modifying your input (<code>img<\/code>) in the <code>predict<\/code> function not just locally, as you might expect. That modified input is used in the next predict, where it again is modified. So you are effectively appying the modifications once for in your first call to <code>predict<\/code>, but twice in the second call.\nYou can find more details about that behavior in this <a href=\"https:\/\/stackoverflow.com\/questions\/13299427\/python-functions-call-by-reference\">question<\/a>.<\/p>\n"}]}
+{"QuestionId":58074285,"AnswerCount":1,"Tags":"<tensorflow><deconvolution>","CreationDate":"2019-09-24T06:30:22.690","AcceptedAnswerId":null,"OwnerUserId":8342418.0,"Title":"Is there any way to implement the mathematical deconvolution(which exactly reverse the convolution) using tensorflow? Please let me know if there is","Body":"<p>I'm trying to make a software in which I need to reverse the convolution process. I haven't found anything useful.<\/p>\n","answers":[{"AnswerId":"58074489","CreationDate":"2019-09-24T06:44:32.153","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>Yes, it is called Transposed Convolution in Tensorflow and also in PyTorch. Here is the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/Conv2DTranspose\" rel=\"nofollow noreferrer\">link<\/a> for TF1.14.\n<a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/layers\/Conv2DTranspose\" rel=\"nofollow noreferrer\">Here<\/a> is for TF2.0.<\/p>\n"}]}
+{"QuestionId":58074474,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><recurrent-neural-network>","CreationDate":"2019-09-24T06:43:08.157","AcceptedAnswerId":null,"OwnerUserId":6997665.0,"Title":"RNN outputs NaN after weight update","Body":"<p>I'm implementing an auto encoder setup where both the encoder and decoders are Recurrent Neural Networks (RNNs). The encoder and decoder model is initialized like so:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def enc(message, weights, biases):\n    message = tf.unstack(message, 4, 1)\n    fw_cell = rnn.LSTMCell(num_hidden_enc)\n    with tf.variable_scope(\"encoder\"):\n        outputs, _ = rnn.static_rnn(fw_cell, message, dtype=tf.float32)\n    return tf.matmul(outputs[-1], weights) + biases\n\ndef dec(codeword, weights, biases, time_steps):\n    codeword = tf.expand_dims(codeword, axis=2)\n    codeword = tf.unstack(codeword, 7, 1)\n    fw_cell = rnn.LSTMCell(num_hidden_dec)\n    with tf.variable_scope(\"decoder\"):\n        outputs, _ = rnn.static_rnn(fw_cell, codeword, dtype=tf.float32)\n    a = tf.matmul(outputs[-1], weights) + biases\n    weight_fc = np.random.normal(loc=0.0, scale=0.01, size=[4, 4])\n    init = tf.constant_initializer(weight_fc)\n    return tf.layers.dense(a, units=4, activation=tf.nn.sigmoid, kernel_initializer=init)\n<\/code><\/pre>\n\n<p>And I am using a <code>mean square error<\/code> loss function along with <code>Adam<\/code> optimizer.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># message_hat is the output of the decoder neural network, input_bits is the input to the encoder network\nloss = tf.reduce_sum(0.5 * (tf.squeeze(input_bits) - message_hat) ** 2) \/ float(batch_size)\nopt = tf.train.AdamOptimizer().minimize(loss)\n<\/code><\/pre>\n\n<hr>\n\n<p>When I print the output of the decoder network at the end of epoch 0, I get some floating point values, but as soon as I run one more epoch, all the values default to <code>NaN<\/code>. This is true for both encoder and decoder. So I tried gradient clipping using the following code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>opt = tf.train.AdamOptimizer()\ngvs = opt.compute_gradients(loss)\ncapped_gvs = [(tf.clip_by_value(grad, -1., 1.), var) for grad, var in gvs]\ntrain_op = opt.apply_gradients(capped_gvs)\n<\/code><\/pre>\n\n<p>Still having the same problem. More precisely, this is the output from epoch 1 onwards:<\/p>\n\n<pre><code>Epoch: 1\nDecoding loss: nan\nDecoder output: [[nan nan nan nan nan nan nan]\n [nan nan nan nan nan nan nan]\n [nan nan nan nan nan nan nan]\n [nan nan nan nan nan nan nan]\n [nan nan nan nan nan nan nan]\n [nan nan nan nan nan nan nan]\n [nan nan nan nan nan nan nan]\n [nan nan nan nan nan nan nan]\n [nan nan nan nan nan nan nan]\n [nan nan nan nan nan nan nan]]\n<\/code><\/pre>\n\n<p>Is there something inherently wrong with the code? Suggestions are welcome. Thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":58074490,"AnswerCount":0,"Tags":"<tensorflow><tensorboard>","CreationDate":"2019-09-24T06:44:36.443","AcceptedAnswerId":null,"OwnerUserId":6244166.0,"Title":"Lazy loading not producing graph on TensorBoard","Body":"<p>I am trying to do an operations in tensorflow in two ways in order to understand\ndifference between normal loading and lazy loading.\nWhen I am running normal loading things are working as expected. Below is the code\nfor normal loading<\/p>\n\n<h1>Normal loading<\/h1>\n\n<pre><code>import tensorflow as tf\nwith tf.Graph().as_default() as my_graph:\n  a = tf.constant(10, name='constant_a')\n  b = tf.constant(30, name='constant_b')\n  c = tf.add(a, b, name='add_a_b_gen_c')\n\n  with tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    writer = tf.summary.FileWriter(\"graph_dir\", sess.graph)\n    for _ in range(10):\n      print(sess.run(c))\n    writer.close()\n  print(tf.get_default_graph().as_graph_def())\n<\/code><\/pre>\n\n<p>Below is the tensorboard screenshot of the operation\n<a href=\"https:\/\/i.stack.imgur.com\/zHtLi.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/zHtLi.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>But when I am channging my code where instead of storing value of tf.add(a, b)\nin c, I am calling this inside run session (lazy loading), then this is not \nproducing expected graph on graphboard.\nSee the changed code below<\/p>\n\n<h1>Lazy loading<\/h1>\n\n<pre><code>import tensorflow as tf\nwith tf.Graph().as_default() as my_graph:\n  a = tf.constant(10, name='constant_a')\n  b = tf.constant(30, name='constant_b')\n\n  with tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    writer = tf.summary.FileWriter(\"graph_dir\", sess.graph)\n    for _ in range(10):\n      print(sess.run(tf.add(a, b, name='add_a_b')))\n    writer.close()\n  print(tf.get_default_graph().as_graph_def())\n<\/code><\/pre>\n\n<p>The Tensorboard output is not as I expected. It is only showing the value init and not displaying a and b values. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/btqam.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/btqam.png\" alt=\"Screenshot1\"><\/a><\/p>\n\n<p>From <a href=\"https:\/\/web.stanford.edu\/class\/cs20si\/2017\/lectures\/notes_02.pdf\" rel=\"nofollow noreferrer\">this<\/a> reference the graph should look like below one\n<a href=\"https:\/\/i.stack.imgur.com\/jafgB.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/jafgB.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Also just for further clarification below function is producing the protobuf for the graph in lazy loading that contains has 10 copies of the node.<\/p>\n\n<pre><code>tf.get_default_graph().as_graph_def()\n<\/code><\/pre>\n\n<p>Can anyone please help me to understanding why the graph is not showing on tensorboard ?<\/p>\n","answers":[]}
+{"QuestionId":58074572,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-24T06:50:50.710","AcceptedAnswerId":null,"OwnerUserId":590335.0,"Title":"How to define a custom feature column in tensorflow?","Body":"<p>As much as I understand feature columns are intended to convert \"raw\" features to tensor encoding.<\/p>\n\n<p>The feature column provides several standard convertors, but it's not clear what is needed to implement a custom convertor? (specifically - string tokenisation \\ parsing)<\/p>\n","answers":[]}
+{"QuestionId":58074785,"AnswerCount":0,"Tags":"<tensorflow><bazel>","CreationDate":"2019-09-24T07:05:32.023","AcceptedAnswerId":null,"OwnerUserId":3951933.0,"Title":"TensorFlow Build Error: \"failed (Exit 1): crosstool_wrapper_driver_is_not_gcc\"","Body":"<p>I am currently trying to build TensorFlow from source using Bazel. I am trying with both TF 1.14 (using Bazel 0.24.1) and TF 1.13 (using Bazel 0.19.2) however I keep getting the same annoying errors. I see that people had these issues and mentioned about them online but TensorFlow guys kept saying \"we resolved this at the head\". But it still seems it doesn't works well anyway. Here is the errors and logs:<\/p>\n\n<p>When I try to build 1.13 with the below command;<\/p>\n\n<pre><code>bazel build --config=opt --config=nonccl --verbose_failures \/\/tensorflow\/tools\/pip_package:build_pip_package \/\/tensorflow:libtensorflow_cc.so \/\/tensorflow:libtensorflow_framework.so --cxxopt=-D_GLIBCXX_USE_CXX11_ABI=0\n<\/code><\/pre>\n\n<p>I get the following error after maybe 3-4 hours:<\/p>\n\n<pre><code>ERROR: \/root\/tensorflow\/tensorflow\/contrib\/tensorrt\/BUILD:268:1: C++ compilation of rule '\/\/tensorflow\/contrib\/tensorrt:trt_conversion' failed (Exit 1): crosstool_wrapper_driver_is_not_gcc failed: error executing command \n  (cd \/root\/.cache\/bazel\/_bazel_root\/efb88f6336d9c4a18216fb94287b8d97\/execroot\/org_tensorflow &amp;&amp; \\\n  exec env - \\\n    CUDA_TOOLKIT_PATH=\/usr\/local\/cuda-10.0 \\\n    CUDNN_INSTALL_PATH=\/usr\/lib\/aarch64-linux-gnu \\\n    GCC_HOST_COMPILER_PATH=\/usr\/bin\/gcc-5 \\\n    LD_LIBRARY_PATH=\/usr\/local\/cuda-10.0\/targets\/aarch64-linux\/lib: \\\n    PATH=\/root\/bazel\/output:\/usr\/local\/cuda-10.0\/bin:\/usr\/local\/cuda\/bin:\/usr\/local\/sbin:\/usr\/local\/bin:\/usr\/sbin:\/usr\/bin:\/sbin:\/bin \\\n    PWD=\/proc\/self\/cwd \\\n    PYTHON_BIN_PATH=\/usr\/bin\/python3 \\\n    PYTHON_LIB_PATH=\/usr\/lib\/python3.6\/dist-packages \\\n    TENSORRT_INSTALL_PATH=\/usr\/lib\/aarch64-linux-gnu \\\n    TF_CUDA_CLANG=0 \\\n    TF_CUDA_COMPUTE_CAPABILITIES=7.2 \\\n    TF_CUDA_VERSION=10.0 \\\n    TF_CUDNN_VERSION=7 \\\n    TF_NCCL_VERSION='' \\\n    TF_NEED_CUDA=1 \\\n    TF_NEED_OPENCL_SYCL=0 \\\n    TF_NEED_ROCM=0 \\\n    TF_TENSORRT_VERSION=5.1.6 \\\n  external\/local_config_cuda\/crosstool\/clang\/bin\/crosstool_wrapper_driver_is_not_gcc -MD -MF bazel-out\/aarch64-opt\/bin\/tensorflow\/contrib\/tensorrt\/_objs\/trt_conversion\/convert_nodes.pic.d '-frandom-seed=bazel-out\/aarch64-opt\/bin\/tensorflow\/contrib\/tensorrt\/_objs\/trt_conversion\/convert_nodes.pic.o' -D__CLANG_SUPPORT_DYN_ANNOTATION__ -DEIGEN_MPL2_ONLY '-DEIGEN_MAX_ALIGN_BYTES=64' '-DEIGEN_HAS_TYPE_TRAITS=0' -DTF_USE_SNAPPY -iquote . -iquote bazel-out\/aarch64-opt\/genfiles -iquote bazel-out\/aarch64-opt\/bin -iquote external\/com_google_absl -iquote bazel-out\/aarch64-opt\/genfiles\/external\/com_google_absl -iquote bazel-out\/aarch64-opt\/bin\/external\/com_google_absl -iquote external\/bazel_tools -iquote bazel-out\/aarch64-opt\/genfiles\/external\/bazel_tools -iquote bazel-out\/aarch64-opt\/bin\/external\/bazel_tools -iquote external\/eigen_archive -iquote bazel-out\/aarch64-opt\/genfiles\/external\/eigen_archive -iquote bazel-out\/aarch64-opt\/bin\/external\/eigen_archive -iquote external\/local_config_sycl -iquote bazel-out\/aarch64-opt\/genfiles\/external\/local_config_sycl -iquote bazel-out\/aarch64-opt\/bin\/external\/local_config_sycl -iquote external\/nsync -iquote bazel-out\/aarch64-opt\/genfiles\/external\/nsync -iquote bazel-out\/aarch64-opt\/bin\/external\/nsync -iquote external\/gif_archive -iquote bazel-out\/aarch64-opt\/genfiles\/external\/gif_archive -iquote bazel-out\/aarch64-opt\/bin\/external\/gif_archive -iquote external\/jpeg -iquote bazel-out\/aarch64-opt\/genfiles\/external\/jpeg -iquote bazel-out\/aarch64-opt\/bin\/external\/jpeg -iquote external\/protobuf_archive -iquote bazel-out\/aarch64-opt\/genfiles\/external\/protobuf_archive -iquote bazel-out\/aarch64-opt\/bin\/external\/protobuf_archive -iquote external\/com_googlesource_code_re2 -iquote bazel-out\/aarch64-opt\/genfiles\/external\/com_googlesource_code_re2 -iquote bazel-out\/aarch64-opt\/bin\/external\/com_googlesource_code_re2 -iquote external\/farmhash_archive -iquote bazel-out\/aarch64-opt\/genfiles\/external\/farmhash_archive -iquote bazel-out\/aarch64-opt\/bin\/external\/farmhash_archive -iquote external\/fft2d -iquote bazel-out\/aarch64-opt\/genfiles\/external\/fft2d -iquote bazel-out\/aarch64-opt\/bin\/external\/fft2d -iquote external\/highwayhash -iquote bazel-out\/aarch64-opt\/genfiles\/external\/highwayhash -iquote bazel-out\/aarch64-opt\/bin\/external\/highwayhash -iquote external\/zlib_archive -iquote bazel-out\/aarch64-opt\/genfiles\/external\/zlib_archive -iquote bazel-out\/aarch64-opt\/bin\/external\/zlib_archive -iquote external\/local_config_cuda -iquote bazel-out\/aarch64-opt\/genfiles\/external\/local_config_cuda -iquote bazel-out\/aarch64-opt\/bin\/external\/local_config_cuda -iquote external\/double_conversion -iquote bazel-out\/aarch64-opt\/genfiles\/external\/double_conversion -iquote bazel-out\/aarch64-opt\/bin\/external\/double_conversion -iquote external\/local_config_tensorrt -iquote bazel-out\/aarch64-opt\/genfiles\/external\/local_config_tensorrt -iquote bazel-out\/aarch64-opt\/bin\/external\/local_config_tensorrt -isystem external\/eigen_archive -isystem bazel-out\/aarch64-opt\/genfiles\/external\/eigen_archive -isystem bazel-out\/aarch64-opt\/bin\/external\/eigen_archive -isystem external\/nsync\/public -isystem bazel-out\/aarch64-opt\/genfiles\/external\/nsync\/public -isystem bazel-out\/aarch64-opt\/bin\/external\/nsync\/public -isystem external\/gif_archive\/lib -isystem bazel-out\/aarch64-opt\/genfiles\/external\/gif_archive\/lib -isystem bazel-out\/aarch64-opt\/bin\/external\/gif_archive\/lib -isystem external\/protobuf_archive\/src -isystem bazel-out\/aarch64-opt\/genfiles\/external\/protobuf_archive\/src -isystem bazel-out\/aarch64-opt\/bin\/external\/protobuf_archive\/src -isystem external\/farmhash_archive\/src -isystem bazel-out\/aarch64-opt\/genfiles\/external\/farmhash_archive\/src -isystem bazel-out\/aarch64-opt\/bin\/external\/farmhash_archive\/src -isystem external\/zlib_archive -isystem bazel-out\/aarch64-opt\/genfiles\/external\/zlib_archive -isystem bazel-out\/aarch64-opt\/bin\/external\/zlib_archive -isystem external\/local_config_cuda\/cuda -isystem bazel-out\/aarch64-opt\/genfiles\/external\/local_config_cuda\/cuda -isystem bazel-out\/aarch64-opt\/bin\/external\/local_config_cuda\/cuda -isystem external\/local_config_cuda\/cuda\/cuda\/include -isystem bazel-out\/aarch64-opt\/genfiles\/external\/local_config_cuda\/cuda\/cuda\/include -isystem bazel-out\/aarch64-opt\/bin\/external\/local_config_cuda\/cuda\/cuda\/include -isystem external\/local_config_cuda\/cuda\/cuda\/include\/crt -isystem bazel-out\/aarch64-opt\/genfiles\/external\/local_config_cuda\/cuda\/cuda\/include\/crt -isystem bazel-out\/aarch64-opt\/bin\/external\/local_config_cuda\/cuda\/cuda\/include\/crt -isystem external\/double_conversion -isystem bazel-out\/aarch64-opt\/genfiles\/external\/double_conversion -isystem bazel-out\/aarch64-opt\/bin\/external\/double_conversion -isystem external\/local_config_tensorrt\/include -isystem bazel-out\/aarch64-opt\/genfiles\/external\/local_config_tensorrt\/include -isystem bazel-out\/aarch64-opt\/bin\/external\/local_config_tensorrt\/include '-std=c++11' -Wno-builtin-macro-redefined '-D__DATE__=\"redacted\"' '-D__TIMESTAMP__=\"redacted\"' '-D__TIME__=\"redacted\"' -fPIC -U_FORTIFY_SOURCE '-D_FORTIFY_SOURCE=1' -fstack-protector -Wall -fno-omit-frame-pointer -no-canonical-prefixes -fno-canonical-system-headers -DNDEBUG -g0 -O2 -ffunction-sections -fdata-sections '-march=native' -Wno-sign-compare '-D_GLIBCXX_USE_CXX11_ABI=0' -DEIGEN_AVOID_STL_ARRAY -Iexternal\/gemmlowp -Wno-sign-compare -fno-exceptions '-ftemplate-depth=900' '-DGOOGLE_CUDA=1' '-DGOOGLE_TENSORRT=1' -pthread '-DGOOGLE_CUDA=1' '-DGOOGLE_TENSORRT=1' -c tensorflow\/contrib\/tensorrt\/convert\/convert_nodes.cc -o bazel-out\/aarch64-opt\/bin\/tensorflow\/contrib\/tensorrt\/_objs\/trt_conversion\/convert_nodes.pic.o)\ntensorflow\/contrib\/tensorrt\/convert\/convert_nodes.cc: In constructor 'tensorflow::tensorrt::convert::TRT_TensorOrWeights::TRT_TensorOrWeights(nvinfer1::DataType, const nvinfer1::Dims&amp;, int)':\ntensorflow\/contrib\/tensorrt\/convert\/convert_nodes.cc:516:60: error: invalid new-expression of abstract class type 'tensorflow::tensorrt::convert::TRT_TensorOrWeights::SimpleITensor'\n     : simple_itensor_(new SimpleITensor(trt_dtype, trt_dims)),\n                                                            ^\ntensorflow\/contrib\/tensorrt\/convert\/convert_nodes.cc:459:28: note:   because the following virtual functions are pure within 'tensorflow::tensorrt::convert::TRT_TensorOrWeights::SimpleITensor':\n class TRT_TensorOrWeights::SimpleITensor : public nvinfer1::ITensor {\n                            ^\nIn file included from .\/tensorflow\/contrib\/tensorrt\/log\/trt_logger.h:23:0,\n                 from .\/tensorflow\/contrib\/tensorrt\/convert\/convert_nodes.h:26,\n                 from tensorflow\/contrib\/tensorrt\/convert\/convert_nodes.cc:16:\nbazel-out\/aarch64-opt\/genfiles\/external\/local_config_tensorrt\/tensorrt\/include\/NvInfer.h:774:18: note:  virtual bool nvinfer1::ITensor::dynamicRangeIsSet() const\n     virtual bool dynamicRangeIsSet() const = 0;\n                  ^\nbazel-out\/aarch64-opt\/genfiles\/external\/local_config_tensorrt\/tensorrt\/include\/NvInfer.h:779:18: note:  virtual void nvinfer1::ITensor::resetDynamicRange()\n     virtual void resetDynamicRange() = 0;\n                  ^\nbazel-out\/aarch64-opt\/genfiles\/external\/local_config_tensorrt\/tensorrt\/include\/NvInfer.h:786:19: note:  virtual float nvinfer1::ITensor::getDynamicRangeMin() const\n     virtual float getDynamicRangeMin() const = 0;\n                   ^\nbazel-out\/aarch64-opt\/genfiles\/external\/local_config_tensorrt\/tensorrt\/include\/NvInfer.h:793:19: note:  virtual float nvinfer1::ITensor::getDynamicRangeMax() const\n     virtual float getDynamicRangeMax() const = 0;\n<\/code><\/pre>\n\n<p>And if I try to build 1.14 with the command below;<\/p>\n\n<pre><code>bazel build --config=opt --config=nonccl \/\/tensorflow\/tools\/pip_package:build_pip_package --verbose_failures --cxxopt=\"-D_GLIBCXX_USE_CXX11_ABI=0\"\n<\/code><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><code>ERROR: \/opt\/tf13\/tensorflow\/tensorflow\/lite\/kernels\/BUILD:286:1: C++ compilation of rule '\/\/tensorflow\/lite\/kernels:builtin_op_kernels' failed (Exit 1): crosstool_wrapper_driver_is_not_gcc failed: error executing command \n  (cd \/root\/.cache\/bazel\/_bazel_root\/8c266c5a221eef177229796f3ca6ace6\/execroot\/org_tensorflow &amp;&amp; \\\n  exec env - \\\n    LD_LIBRARY_PATH=\/usr\/local\/cuda-10.0\/targets\/aarch64-linux\/lib: \\\n    PATH=\/root\/bazel\/output:\/usr\/local\/cuda-10.0\/bin:\/usr\/local\/cuda\/bin:\/usr\/local\/sbin:\/usr\/local\/bin:\/usr\/sbin:\/usr\/bin:\/sbin:\/bin \\\n    PWD=\/proc\/self\/cwd \\\n  external\/local_config_cuda\/crosstool\/clang\/bin\/crosstool_wrapper_driver_is_not_gcc -MD -MF bazel-out\/host\/bin\/tensorflow\/lite\/kernels\/_objs\/builtin_op_kernels\/depthwise_conv.pic.d '-frandom-seed=bazel-out\/host\/bin\/tensorflow\/lite\/kernels\/_objs\/builtin_op_kernels\/depthwise_conv.pic.o' -DEIGEN_MPL2_ONLY '-DEIGEN_MAX_ALIGN_BYTES=64' '-DEIGEN_HAS_TYPE_TRAITS=0' -iquote . -iquote bazel-out\/host\/genfiles -iquote bazel-out\/host\/bin -iquote external\/gemmlowp -iquote bazel-out\/host\/genfiles\/external\/gemmlowp -iquote bazel-out\/host\/bin\/external\/gemmlowp -iquote external\/com_google_absl -iquote bazel-out\/host\/genfiles\/external\/com_google_absl -iquote bazel-out\/host\/bin\/external\/com_google_absl -iquote external\/arm_neon_2_x86_sse -iquote bazel-out\/host\/genfiles\/external\/arm_neon_2_x86_sse -iquote bazel-out\/host\/bin\/external\/arm_neon_2_x86_sse -iquote external\/eigen_archive -iquote bazel-out\/host\/genfiles\/external\/eigen_archive -iquote bazel-out\/host\/bin\/external\/eigen_archive -iquote external\/local_config_sycl -iquote bazel-out\/host\/genfiles\/external\/local_config_sycl -iquote bazel-out\/host\/bin\/external\/local_config_sycl -iquote external\/flatbuffers -iquote bazel-out\/host\/genfiles\/external\/flatbuffers -iquote bazel-out\/host\/bin\/external\/flatbuffers -iquote external\/fft2d -iquote bazel-out\/host\/genfiles\/external\/fft2d -iquote bazel-out\/host\/bin\/external\/fft2d -iquote external\/farmhash_archive -iquote bazel-out\/host\/genfiles\/external\/farmhash_archive -iquote bazel-out\/host\/bin\/external\/farmhash_archive -isystem external\/eigen_archive -isystem bazel-out\/host\/genfiles\/external\/eigen_archive -isystem bazel-out\/host\/bin\/external\/eigen_archive -isystem tensorflow\/lite\/schema -isystem bazel-out\/host\/genfiles\/tensorflow\/lite\/schema -isystem bazel-out\/host\/bin\/tensorflow\/lite\/schema -isystem external\/flatbuffers\/include -isystem bazel-out\/host\/genfiles\/external\/flatbuffers\/include -isystem bazel-out\/host\/bin\/external\/flatbuffers\/include -isystem external\/farmhash_archive\/src -isystem bazel-out\/host\/genfiles\/external\/farmhash_archive\/src -isystem bazel-out\/host\/bin\/external\/farmhash_archive\/src '-std=c++11' -Wno-builtin-macro-redefined '-D__DATE__=\"redacted\"' '-D__TIMESTAMP__=\"redacted\"' '-D__TIME__=\"redacted\"' -fPIC -U_FORTIFY_SOURCE '-D_FORTIFY_SOURCE=1' -fstack-protector -Wall -fno-omit-frame-pointer -no-canonical-prefixes -fno-canonical-system-headers -DNDEBUG -g0 -O2 -ffunction-sections -fdata-sections -g0 '-march=native' -g0 -DFARMHASH_NO_CXX_STRING -Wno-sign-compare '-Wno-error=reorder' -c tensorflow\/lite\/kernels\/depthwise_conv.cc -o bazel-out\/host\/bin\/tensorflow\/lite\/kernels\/_objs\/builtin_op_kernels\/depthwise_conv.pic.o)\nExecution platform: @bazel_tools\/\/platforms:host_platform\nIn file included from .\/tensorflow\/lite\/kernels\/internal\/optimized\/depthwiseconv_uint8.h:23:0,\n                 from .\/tensorflow\/lite\/kernels\/internal\/optimized\/depthwiseconv_multithread.h:21,\n                 from tensorflow\/lite\/kernels\/depthwise_conv.cc:28:\n.\/tensorflow\/lite\/kernels\/internal\/optimized\/depthwiseconv_uint8_3x3_filter.h: In static member function 'static void tflite::optimized_ops::depthwise_conv::PackMacroBlock&lt;(tflite::DepthwiseConvImplementation)3, (tflite::DepthwiseConvDepthMultiplication)0, 0&gt;::PackMacroBlockNeon(const uint8*, int8*, const tflite::optimized_ops::depthwise_conv::DepthwiseConvDotProdParams*)':\n.\/tensorflow\/lite\/kernels\/internal\/optimized\/depthwiseconv_uint8_3x3_filter.h:5981:24: note: use -flax-vector-conversions to permit conversions between vectors with differing element types or numbers of subparts\n           input_data_a = vld1q_u8(input_data_0);\n                        ^\n.\/tensorflow\/lite\/kernels\/internal\/optimized\/depthwiseconv_uint8_3x3_filter.h:5981:24: error: cannot convert 'uint8x16_t {aka __vector(16) unsigned char}' to 'int8x16_t {aka __vector(16) signed char}' in assignment\n.\/tensorflow\/lite\/kernels\/internal\/optimized\/depthwiseconv_uint8_3x3_filter.h:5982:24: error: cannot convert 'uint8x16_t {aka __vector(16) unsigned char}' to 'int8x16_t {aka __vector(16) signed char}' in assignment\n           input_data_b = vld1q_u8(input_data_0 + 1 * input_depth);\n                        ^\n.\/tensorflow\/lite\/kernels\/internal\/optimized\/depthwiseconv_uint8_3x3_filter.h:5983:24: error: cannot convert 'uint8x16_t {aka __vector(16) unsigned char}' to 'int8x16_t {aka __vector(16) signed char}' in assignment\n           input_data_c = vld1q_u8(input_data_0 + 2 * input_depth);\n                        ^\n.\/tensorflow\/lite\/kernels\/internal\/optimized\/depthwiseconv_uint8_3x3_filter.h:5984:24: error: cannot convert 'uint8x16_t {aka __vector(16) unsigned char}' to 'int8x16_t {aka __vector(16) signed char}' in assignment\n           input_data_d = vld1q_u8(input_data_0 + 3 * input_depth);\n                        ^\n.\/tensorflow\/lite\/kernels\/internal\/optimized\/depthwiseconv_uint8_3x3_filter.h:5993:55: error: cannot convert 'const uint8x16_t {aka const __vector(16) unsigned char}' to 'int8x16_t {aka __vector(16) signed char}' for argument '2' to 'int8x16_t veorq_s8(int8x16_t, int8x16_t)'\n             work_reg_a = veorq_s8(work_reg_a, sign_bit);\n                                                       ^\n.\/tensorflow\/lite\/kernels\/internal\/optimized\/depthwiseconv_uint8_3x3_filter.h:5994:55: error: cannot convert 'const uint8x16_t {aka const __vector(16) unsigned char}' to 'int8x16_t {aka __vector(16) signed char}' for argument '2' to 'int8x16_t veorq_s8(int8x16_t, int8x16_t)'\n             work_reg_b = veorq_s8(work_reg_b, sign_bit);\n                                                       ^\n.\/tensorflow\/lite\/kernels\/internal\/optimized\/depthwiseconv_uint8_3x3_filter.h:6000:26: error: cannot convert 'uint8x16_t {aka __vector(16) unsigned char}' to 'int8x16_t {aka __vector(16) signed char}' in assignment\n             input_data_a = vld1q_u8(input_data_0);\n<\/code><\/pre>\n\n<p>What is the problem here? It is so frustrating to wait 4 hours and see it explodes at the exact same point. I am also using -flax-vector-conversions -fomit-frame-pointer flags within the tensorflow\/lite\/tools\/make\/targets\/aarch64_makefile.inc file but this doesn't resolve the issue. What am I missing?<\/p>\n","answers":[]}
+{"QuestionId":58075020,"AnswerCount":0,"Tags":"<tensorflow><tensorboard>","CreationDate":"2019-09-24T07:19:45.000","AcceptedAnswerId":null,"OwnerUserId":6244166.0,"Title":"Default graph is producing results different from tf.Graph.as_default context manager (Incorrect result in default graph)","Body":"<p>I am getting very different results while using Default graph and tf.Graph.as_default context manager. The results that I am getting using tf.Graph.as_default context manager is align with what was expected. I am not able to understand why the result are different in case of Default graph(i.e. without using context manager tf.Graph.as_default see second code)<\/p>\n\n<p><strong>Normal loading with tf.Graph.as_default context manager<\/strong> (i.e. without default graph)<\/p>\n\n<pre><code>import tensorflow as tf\nwith tf.Graph().as_default() as my_graph:\n  a = tf.constant(10, name='constant_a')\n  b = tf.constant(30, name='constant_b')\n  c = tf.add(a, b, name='add_a_b_gen_c')\n\n  with tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    writer = tf.summary.FileWriter(\"graph_dir\", sess.graph)\n    for _ in range(10):\n      print(sess.run(c))\n    writer.close()\n  print(tf.get_default_graph().as_graph_def())\n<\/code><\/pre>\n\n<p>The output is as per expectation.\n<a href=\"https:\/\/i.stack.imgur.com\/P9TJ5.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/P9TJ5.png\" alt=\"Screenshot1\"><\/a><\/p>\n\n<p>Whereas in case I use <strong>Default graph<\/strong> (See below code)<\/p>\n\n<pre><code>import tensorflow as tf\na = tf.constant(10, name='a')\nb = tf.constant(30, name='b')\nc = tf.add(a, b, name=\"add_a_b\")\nwith tf.Session() as sess:\n  sess.run(tf.global_variables_initializer())\n  #sess.run([a, b])\n  writer = tf.summary.FileWriter(\"graph_dir\", sess.graph)\n  for _ in range(10):\n    print(sess.run(c))\n  writer.close()\nprint(tf.get_default_graph().as_graph_def())\n<\/code><\/pre>\n\n<p>Then my output is very different and not understandable to me.\nWhy is this happening and what does the output from second code represent? <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Dbuwx.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Dbuwx.png\" alt=\"Screenshot 2\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":58075061,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2019-09-24T07:22:29.063","AcceptedAnswerId":null,"OwnerUserId":11192771.0,"Title":"TypeError: Error converting shape to a TensorShape: Only size 1 arrays are accepted","Body":"<p>I'm currently stuck on an error regarding my neural network. The dataset consists of a 2418 arrays with each containing 400 numbers. The shape is (2418,400). After trial and error I reshaped this to (2418,400,1) because I thought that would be better for the model. The error I receive when running is:<\/p>\n\n<pre><code>Traceback (most recent call last):\n\n  File \"&lt;ipython-input-15-370efbb6619c&gt;\", line 9, in &lt;module&gt;\n    model.add(LSTM(128, input_shape=(Xreshaped), activation='relu', return_sequences=True))\n\n  File \"C:\\Users\\Tijev\\Anaconda3\\envs\\tfp3.6\\lib\\site-packages\\tensorflow\\python\\training\\checkpointable\\base.py\", line 442, in _method_wrapper\n    method(self, *args, **kwargs)\n\n  File \"C:\\Users\\Tijev\\Anaconda3\\envs\\tfp3.6\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\sequential.py\", line 160, in add\n    name=layer.name + '_input')\n\n  File \"C:\\Users\\Tijev\\Anaconda3\\envs\\tfp3.6\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\input_layer.py\", line 231, in Input\n    input_tensor=tensor)\n\n  File \"C:\\Users\\Tijev\\Anaconda3\\envs\\tfp3.6\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\input_layer.py\", line 107, in __init__\n    name=self.name)\n\n  File \"C:\\Users\\Tijev\\Anaconda3\\envs\\tfp3.6\\lib\\site-packages\\tensorflow\\python\\keras\\backend.py\", line 876, in placeholder\n    x = array_ops.placeholder(dtype, shape=shape, name=name)\n\n  File \"C:\\Users\\Tijev\\Anaconda3\\envs\\tfp3.6\\lib\\site-packages\\tensorflow\\python\\ops\\array_ops.py\", line 2077, in placeholder\n    return gen_array_ops.placeholder(dtype=dtype, shape=shape, name=name)\n\n  File \"C:\\Users\\Tijev\\Anaconda3\\envs\\tfp3.6\\lib\\site-packages\\tensorflow\\python\\ops\\gen_array_ops.py\", line 5788, in placeholder\n    shape = _execute.make_shape(shape, \"shape\")\n\n  File \"C:\\Users\\Tijev\\Anaconda3\\envs\\tfp3.6\\lib\\site-packages\\tensorflow\\python\\eager\\execute.py\", line 147, in make_shape\n    raise TypeError(\"Error converting %s to a TensorShape: %s.\" % (arg_name, e))\n\nTypeError: Error converting shape to a TensorShape: only size-1 arrays can be converted to Python scalars.\n<\/code><\/pre>\n\n<p>I've looked at a similar post on Stackoverflow but the answer didn't work out for this case. My code looks like this:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(128, input_shape=(Xreshaped), activation='relu', return_sequences=True))\nmodel.add(Dropout(0,2))\nmodel.add(LSTM(128, activation='relu'))\nmodel.add(Dropout(0,2))\nmodel.add(Dense(32, activation='relu'))\nmodel.add(Dropout(0,2))\nmodel.add(Dense(2, activation='sigmoid'))\n\nopt = tf.keras.optimizers.Adam(lr=1e-3, decay=1e-5)\n\n# mean_squared_error = mse\nmodel.compile(loss='sparse_categorical_crossentropy',\n              optimizer=opt,\n              metrics=['accuracy'])\n\nmodel.fit(X, Y, epochs=3, validation_data=(X, Y))\n<\/code><\/pre>\n","answers":[{"AnswerId":"58081836","CreationDate":"2019-09-24T13:49:47.283","ParentId":null,"OwnerUserId":"11192771","Title":null,"Body":"<p>Seems the problem is fixed with using:<\/p>\n\n<pre><code>Xreshaped = X.reshape((2418,400,1))\n\ninput_shape = (400,1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58075325,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><neural-network><text-classification>","CreationDate":"2019-09-24T07:39:24.380","AcceptedAnswerId":null,"OwnerUserId":12111421.0,"Title":"One huge or multiple small models for text classification","Body":"<p>I am working on classification of large text database (millions of textes) labled to some thousands of categories \/ subcategories (very similar to the amazon product dataset). <\/p>\n\n<p>My question is, if i could get a higher (in general) accuracy by the creation of multiple models e.g. <\/p>\n\n<ul>\n<li>One model to get the main category (one main-model)<\/li>\n<li><p>Separate models for each main category to get the final subcategory. (many category-models)<\/p>\n\n<p>compared to <\/p><\/li>\n<li><p>One huge model over all textes and all subcategories. <\/p><\/li>\n<\/ul>\n\n<p>I am wondering if the vocabulary that is used in the textes of a main category and their subcategories is easier (and of smaller size) and more precise to learn for a model? But there will be an additional problem if the main-model predicts a wrong main-category - then the search for the right subcategory will have no chance to predict the right result.<\/p>\n\n<p>Is there any research on this? \nAny rules of thumb?\nMany thanks!<\/p>\n","answers":[{"AnswerId":"58075565","CreationDate":"2019-09-24T07:54:33.827","ParentId":null,"OwnerUserId":"9317034","Title":null,"Body":"<p>don't know about research, but can suggest that first use precision and recall matrix for the main model (with multiple modelling techniques) and if any model for all categories gives precision and recall values above the threshold lets say 99.5% then you gotta go with your idea. If you fail to achieve this don't worry just try another approach.It's about trying and observing. let me know for anything else. Thank you.<\/p>\n"}]}
+{"QuestionId":58075431,"AnswerCount":1,"Tags":"<tensorflow><keras><normalization>","CreationDate":"2019-09-24T07:46:09.143","AcceptedAnswerId":58076395.0,"OwnerUserId":3306721.0,"Title":"Tensorflow \/ Keras: Normalize train \/ test \/ realtime Data or how to handle reality?","Body":"<p>I started developing some LSTM-models and now have some questions about normalization. <\/p>\n\n<p>Lets pretend I have some time series data that is roughly ranging between +500 and -500. Would it be more realistic to scale the Data from -1 to 1, or is 0 to 1 a better way, I tested it and 0 to 1 seemed to be faster. Is there a wrong way to do it? Or would it just be slower to learn?<\/p>\n\n<p>Second question: When do I normalize the data? I split the data into training and testdata, do I have to scale \/ normalize this data seperately? maybe the trainingdata is only ranging between +300 to -200 and the testdata ranges from +600 to -100. Thats not very good I guess.<\/p>\n\n<p>But on the other hand... If I scale \/ normalize the entire dataframe and split it after that, the data is fine for training and test, but how do I handle real new incomming data? The model is trained to scaled data, so I have to scale the new data as well, right? But what if the new Data is 1000? the normalization would turn this into something more then 1, because its a bigger number then everything else before.<\/p>\n\n<p>To make a long story short, when do I normalize data and what happens to completely new data? <\/p>\n\n<p>I hope I could make it clear what my problem is :D<\/p>\n\n<p>Thank you very much!<\/p>\n","answers":[{"AnswerId":"58076395","CreationDate":"2019-09-24T08:46:15.020","ParentId":null,"OwnerUserId":"10886420","Title":null,"Body":"<p>Would like to know how to handle reality as well tbh...<\/p>\n\n<p>On a serious note though:<\/p>\n\n<h2>1. How to normalize data<\/h2>\n\n<p>Usually, neural networks benefit from data coming from Gaussian Standard distribution (mean <code>0<\/code> and variance <code>1<\/code>).<\/p>\n\n<p>Techniques like Batch Normalization (simplifying), help neural net to have this trait throughout the whole network, so it's usually beneficial.<\/p>\n\n<p>There are other approaches that you mentioned, to tell reliably what helps for which problem and specified architecture you just have to check and measure.<\/p>\n\n<h2>2. What about test data?<\/h2>\n\n<p>Mean to subtract and variance to divide each instance by (or any other statistic you gather by any normalization scheme mentioned previously) should be gathered from your training dataset. If you take them from test, you perform data leakage (info about test distribution is incorporated into training) and you may get false impression your algorithm performs better than in reality.<\/p>\n\n<p>So just compute statistics over training dataset and use them on incoming\/validation\/test data as well. <\/p>\n"}]}
+{"QuestionId":58075445,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-24T07:47:10.357","AcceptedAnswerId":null,"OwnerUserId":4890400.0,"Title":"How to install Tensorflow with python 2.7 in Linux?","Body":"<p>when I install tensorflow(CPU) with python 2.7 in linux, I got this error:<\/p>\n\n<pre><code>Complete output from command python setup.py egg_info:\nTraceback (most recent call last):\n  File \"&lt;string&gt;\", line 1, in &lt;module&gt;\n  File \"\/tmp\/pip-build-QbJljI\/numpy\/setup.py\", line 31, in &lt;module&gt;\n    raise RuntimeError(\"Python version &gt;= 3.5 required.\")\nRuntimeError: Python version &gt;= 3.5 required.\n----------------------------------------\n<\/code><\/pre>\n\n<p>Command \"python setup.py egg_info\" failed with error code 1 in \/tmp\/pip-build-QbJljI\/numpy\/<\/p>\n\n<p>I used this command to install:\npip install --user --upgrade tensorflow<\/p>\n\n<p>my system is:\nLinux VM_73_170_centos 3.10.107-1-tlinux2_kvm_guest-0049<\/p>\n\n<p>And I do not want to exchange the version of python,\nso what I can do with this error? Thx<\/p>\n","answers":[]}
+{"QuestionId":58075740,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-24T08:06:01.473","AcceptedAnswerId":null,"OwnerUserId":12107342.0,"Title":"How can I obtain reproducible results using Keras during development?","Body":"<p>I followed this <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-reproducible-results-using-keras-during-development\" rel=\"nofollow noreferrer\">link<\/a> but I can't set <code>PYTHONHASHSEED=0<\/code> ,or <code>PYTHONHASHSEED=0<\/code>, it won't work. My code is as follows\uff1a  <\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport random as rn\nsd = 1\nnp.random.seed(sd)\nrn.seed(sd)\nconfig = tf.ConfigProto(intra_op_parallelism_threads=1,inter_op_parallelism_threads=1)\nfrom keras import backend as K\ntf.set_random_seed(sd)\nsess = tf.Session(graph=tf.get_default_graph(), config=config)\nK.set_session(sess)\nimport os\nos.environ[\"CUDA_DEVICE_ORDER\"] = \"PCI_BUS_ID\"\nos.environ[\"CUDA_VISIBLE_DEVICES\"] = \"\"\nos.environ['PYTHONHASHSEED'] = '0'\n<\/code><\/pre>\n\n<p>Every time I ran the code, it gave different results. My environment consists of<\/p>\n\n<ul>\n<li>Keras 2.2.4<\/li>\n<li>with tensorflow-CPU 1.14.0 as backend.<\/li>\n<li>python 3.7<\/li>\n<li>windows 10<\/li>\n<\/ul>\n\n<p>Can anyone help me get a reproducible result?<\/p>\n","answers":[{"AnswerId":"58076204","CreationDate":"2019-09-24T08:35:32.820","ParentId":null,"OwnerUserId":"10886420","Title":null,"Body":"<p>You can't reliably to be honest, apparently Keras devs don't seem concerned about that (see <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/2280\" rel=\"nofollow noreferrer\">this issue<\/a>).<\/p>\n\n<p>You can find some hacks to <strong>maybe<\/strong> do it over there, been trying most of them on collab with no luck though, sorry.<\/p>\n"}]}
+{"QuestionId":58075896,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><tensorflow.js>","CreationDate":"2019-09-24T08:15:26.700","AcceptedAnswerId":58079410.0,"OwnerUserId":2876990.0,"Title":"TensorflowJs conv2d - Tensor Shapes","Body":"<p>I want to create a machine learning model for audio files. I converted the audio files into a (spectrogram) tensor. My feature tensor (the audio files) has the following shape <code>[119, 241, 125]<\/code> (119 files, 241 samples\/file, 125 frequencies\/sample). By sample, I define the samples I took in a timespan e.g. 16ms. My output shape will be <code>[119, numOptions]<\/code>.<\/p>\n\n<p>I followed this <a href=\"https:\/\/youtu.be\/-1QGEQWhmSI?t=655\" rel=\"nofollow noreferrer\">tutorial from Tensorflow.js<\/a> on audio recognition. They build this model:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/kLsC4.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/kLsC4.png\" alt=\"Model\"><\/a><\/p>\n\n<p>I reshape my features tensor to be 4D: \n<code>this.features = this.features.reshape([this.features.shape[0],this.features.shape[1],this.features.shape[2],1])<\/code>for the 2Dconv.<\/p>\n\n<pre><code>  buildModel() {\n        const inputShape1 = [this.features.shape[1], this.features.shape[2],this.features.shape[3]];\n        this.model = tfNode.sequential();\n        \/\/ filter to the image =&gt; feature extractor, edge detector, sharpener (depends on the models understanding)\n        this.model.add(tfNode.layers.conv2d(\n            {filters: 8, kernelSize: [4, 2], activation: 'relu', inputShape: inputShape1}\n        ));\n\n        \/\/ see the image at a higher level, generalize it more, prevent overfit\n        this.model.add(tfNode.layers.maxPooling2d(\n            {poolSize: [2, 2], strides: [2, 2]}\n        ));\n\n        \/\/ filter to the image =&gt; feature extractor, edge detector, sharpener (depends on the models understanding)\n        const inputShape2 = [119,62,8];\n        this.model.add(tfNode.layers.conv2d(\n            {filters: 32, kernelSize: [4, 2], activation: 'relu', inputShape: inputShape2}\n        ));\n\n        \/\/ see the image at a higher level, generalize it more, prevent overfit\n        this.model.add(tfNode.layers.maxPooling2d(\n            {poolSize: [2, 2], strides: [2, 2]}\n        ));\n\n        \/\/ filter to the image =&gt; feature extractor, edge detector, sharpener (depends on the models understanding)\n        const inputShape3 = [58,30,32];\n        this.model.add(tfNode.layers.conv2d(\n            {filters: 32, kernelSize: [4, 2], activation: 'relu', inputShape: inputShape3}\n        ));\n\n        \/\/ see the image at a higher level, generalize it more, prevent overfit\n        this.model.add(tfNode.layers.maxPooling2d(\n            {poolSize: [2, 2], strides: [2, 2]}\n        ));\n\n        \/\/ 1D output, =&gt; final output score of labels\n        this.model.add(tfNode.layers.flatten({}));\n\n        \/\/ prevents overfitting, randomly set 0\n        this.model.add(tfNode.layers.dropout({rate: 0.25}));\n\n        \/\/ learn anything linear, non linear comb. from conv. and soft pool\n        this.model.add(tfNode.layers.dense({units: 2000, activation: 'relu'}));\n\n        this.model.add(tfNode.layers.dropout({rate: 0.25}));\n\n        \/\/ give probability for each label\n        this.model.add(tfNode.layers.dense({units: this.labels.shape[1], activation: 'softmax'}));\n\n        this.model.summary();\n\n        \/\/ compile the model\n        this.model.compile({loss: 'meanSquaredError', optimizer: 'adam'});\n        this.model.summary()\n    };\n<\/code><\/pre>\n\n<h2>Model summary:<\/h2>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output shape              Param #   \n=================================================================\nconv2d_Conv2D1 (Conv2D)      [null,238,124,8]          72        \n_________________________________________________________________\nmax_pooling2d_MaxPooling2D1  [null,119,62,8]           0         \n_________________________________________________________________\nconv2d_Conv2D2 (Conv2D)      [null,116,61,32]          2080      \n_________________________________________________________________\nmax_pooling2d_MaxPooling2D2  [null,58,30,32]           0         \n_________________________________________________________________\nconv2d_Conv2D3 (Conv2D)      [null,55,29,32]           8224      \n_________________________________________________________________\nmax_pooling2d_MaxPooling2D3  [null,27,14,32]           0         \n_________________________________________________________________\nflatten_Flatten1 (Flatten)   [null,12096]              0         \n_________________________________________________________________\ndropout_Dropout1 (Dropout)   [null,12096]              0         \n_________________________________________________________________\ndense_Dense1 (Dense)         [null,2000]               24194000  \n_________________________________________________________________\ndropout_Dropout2 (Dropout)   [null,2000]               0         \n_________________________________________________________________\ndense_Dense2 (Dense)         [null,2]                  4002      \n=================================================================\nTotal params: 24208378\nTrainable params: 24208378\nNon-trainable params: 0\n_________________________________________________________________\n    Epoch 1 \/ 10\neta=0.0 ======================================&gt;----------------------------------------------------------------------------- loss=0.515 0.51476\neta=0.8 ============================================================================&gt;--------------------------------------- loss=0.442 0.44186\neta=0.0 ===================================================================================================================&gt; \n3449ms 32236us\/step - loss=0.485 val_loss=0.958 \nEpoch 2 \/ 10\neta=0.0 ======================================&gt;----------------------------------------------------------------------------- loss=0.422 0.42188\neta=0.9 ============================================================================&gt;--------------------------------------- loss=0.395 0.39535\neta=0.0 ===================================================================================================================&gt; \n3643ms 34043us\/step - loss=0.411 val_loss=0.958 \nEpoch 3 \/ 10\n<\/code><\/pre>\n\n<p>1) The first input size is my features tensor shape. The other two inputShapes <code>(inputShape2, inputShape3)<\/code> where defined by the error message I got. How to determine the following two input sizes in advance?<\/p>\n","answers":[{"AnswerId":"58079410","CreationDate":"2019-09-24T11:35:08.993","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p><strong>How the inputShape is calculated ?<\/strong><\/p>\n\n<p>It is not the inputShape that is calculated. It is the dataset that is passed to the model that has to match the <code>inputShape<\/code>. While defining the model, the inputShape is of 3D. But looking at the model summary, there is a fourth dimension with value <code>null<\/code> that is the batchshape. As a result, the training data should be of 4D. The first dimension or batchshape can be whatever - what matters is for the features and the labels to have the same batchshape. There is a more detailed answer <a href=\"https:\/\/stackoverflow.com\/questions\/55737522\/tensorflow-js-inputshape-doesnt-match-model-input\">here<\/a><\/p>\n\n<p><strong>How the layers shape is calculated ?<\/strong><\/p>\n\n<p>It depends of the layers used. Layers such as <code>dropout<\/code>, <code>activation<\/code> don't change the input shape.<\/p>\n\n<ul>\n<li><p>Depending on the stride kernel, the convolution layer will change the input shape. This <a href=\"https:\/\/stackoverflow.com\/questions\/52729443\/new-shape-and-old-shape-must-have-the-same-number-of-elements\">answer<\/a> details how it is calculated.<\/p><\/li>\n<li><p>A flatten layer will simply reshape the inputShape to be of one dimension. In the model summary, there is the input shape <code>[null,27,14,32]<\/code> and the flatten layer has the shape <code>[null, 12096] (12096 = 27 * 14 *32)<\/code><\/p><\/li>\n<li><p>The dense layer will also change the input shape. The shape of the dense layer depends of the number of units of that layer.<\/p><\/li>\n<\/ul>\n"}]}
+{"QuestionId":58076216,"AnswerCount":1,"Tags":"<keras><neural-network><loss-function><multilabel-classification>","CreationDate":"2019-09-24T08:36:03.167","AcceptedAnswerId":58084318.0,"OwnerUserId":11276415.0,"Title":"Keras custom loss function that depends on the input features","Body":"<p>I have a multilabel classification problem with <code>K<\/code> labels and also I have a function, let's call it <code>f<\/code> that for each example in the dataset takes in two matrices, let's call them <code>H<\/code> and <code>P<\/code>. Both matrices are part of the input data.<\/p>\n\n<p>For each vector of labels <code>y<\/code> (for one example), i.e. <code>y<\/code> is a vector with dimension <code>(K \\times 1)<\/code>, I compute a scalar value <code>f_out = f(H, P, y)<\/code>. <\/p>\n\n<p>I want to define a loss function that minimizes the mean absolute percentage error between the two vectors formed by the values <code>f_out_true = f(H, P, y_true)<\/code> and <code>f_out_pred = f(H, P, y_pred)<\/code> for all examples.<\/p>\n\n<p>Seeing the documentation of Keras, I know that customized loss function comes in the form <code>custmLoss(y_pred, y_true)<\/code>, however, the loss function I want to define depends on the input data and these values <code>f_out_true<\/code> and <code>f_out_pred<\/code> need to be computed example by example to form the two vectors that I want to minimize the mean absolute percentage error.<\/p>\n","answers":[{"AnswerId":"58084318","CreationDate":"2019-09-24T16:08:13.547","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>As far as I am aware, there is no way to make a loss function that takes anything other than the model output and the corresponding ground truth. So, the only way to do what you want is to make the input part of your model's output. To do this, simply build your model with the functional API, and then add the input tensor to the list of outputs:<\/p>\n\n<pre><code>input = Input(input_shape)\n\n# build the rest of your model with the standard functional API here\n# this example model was taken from the Keras docs\nx = Dense(100, activation='relu')(input)\nx = Dense(100, activation='relu')(x)\nx = Dense(100, activation='relu')(x)\noutput = Dense(10, activation='softmax')(x)\n\nmodel = Model(inputs=[input], outputs=[output, input])\n<\/code><\/pre>\n\n<p>Then, make <code>y_true<\/code> a combination of your input data and the original ground truth.<\/p>\n\n<p>I don't have a whole lot of experience with the functional API so it's hard to be more specific, but hopefully this points you in the right direction.<\/p>\n"}]}
+{"QuestionId":58076296,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><tensorflow-hub>","CreationDate":"2019-09-24T08:40:19.170","AcceptedAnswerId":null,"OwnerUserId":8181833.0,"Title":"How to export tensorflow_hub module to use it in an installer","Body":"<p>I'm working on a app that use a tensorflow module, currently i load tensorflow like this<\/p>\n\n<pre><code>module_path = 'https:\/\/tfhub.dev\/deepmind\/biggan-deep-128\/1'\ntf.reset_default_graph()\nmodule = hub.Module(module_path)\n\ninputs = {k: tf.placeholder(v.dtype, v.get_shape().as_list(), k)\n          for k, v in module.get_input_info_dict().items()}\noutput = module(inputs)\n\ninput_z = inputs['z']\ninput_y = inputs['y']\ninput_trunc = inputs['truncation']\ndim_z = input_z.shape.as_list()[1]\nvocab_size = input_y.shape.as_list()[1]\n\ninitializer = tf.global_variables_initializer()\nsess = tf.Session()\nsess.run(initializer)\n<\/code><\/pre>\n\n<p>However i want to make my app usable for everyone, i plan to export my code to an exe (probably with pyinstaller), so i need to 'export' this model and make load it from the disk<\/p>\n\n<p>I didn't find any information on how to it online and i'm fairly new to tensorflow so i don't really get how to do it<\/p>\n","answers":[{"AnswerId":"58080697","CreationDate":"2019-09-24T12:48:54.083","ParentId":null,"OwnerUserId":"51685","Title":null,"Body":"<p>It's not exactly documented anywhere, but it <a href=\"https:\/\/github.com\/tensorflow\/hub\/blob\/25a740b3564db8ad70ea512b1ce090aafa6a9cce\/tensorflow_hub\/compressed_module_resolver.py#L41\" rel=\"nofollow noreferrer\">looks like<\/a> appending <code>?tf-hub-format=compressed<\/code> to a <code>https:\/\/tfhub.dev\/deepmind\/biggan-deep-128\/1<\/code> style URL causes a redirect to the actual download:<\/p>\n\n<pre><code>$ wget 'https:\/\/tfhub.dev\/deepmind\/biggan-deep-128\/1?tf-hub-format=compressed'\n...\nLength: 384022812 (366M) [application\/x-tar]\nSaving to: \u20181?tf-hub-format=compressed\u2019\n<\/code><\/pre>\n\n<p>Once you have that file (renamed to e.g. <code>biggan-deep-128.tar<\/code>), you can probably, if I'm reading the code right, pass it to <code>hub.Module()<\/code>, e.g.<\/p>\n\n<pre><code>module = hub.Module('.\/biggan-deep-128.tar')\n<\/code><\/pre>\n"}]}
+{"QuestionId":58076574,"AnswerCount":3,"Tags":"<python><tensorflow><coreml>","CreationDate":"2019-09-24T08:56:29.093","AcceptedAnswerId":58084415.0,"OwnerUserId":12111998.0,"Title":"Can I convert a \".py\" or \".ipynb\" model\/file created on Google Colab's Tensorflow to a coreML model\/ file?","Body":"<p>I created a CNN model on Google Colab's Tensorflow and downloaded the file as a \".py\" file. How can I convert this file into a \".coreml\" file?<\/p>\n\n<p>Any advice is much appreciated!<\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"58229685","CreationDate":"2019-10-04T03:52:57.933","ParentId":null,"OwnerUserId":"12111998","Title":null,"Body":"<p>@ Sebastian J. Vogt. This is my conversion code. Is it correct? Thank you!<\/p>\n\n<pre><code>import coremltools\n\ncategory_output_labels = ['dress', 'shirt', 'trousers']\ncolor_output_labels = ['black', 'blue', 'green', 'pink', 'red', \n'white']\n\ncoreml_model = coremltools.converters.keras.convert('.\/fashion_multi_output_v02.h5',\n<\/code><\/pre>\n\n<p>input_names='image', image_input_names='image', class_labels=['category_output_labels', 'color_output_labels'])<\/p>\n"},{"AnswerId":"58084415","CreationDate":"2019-09-24T16:14:14.993","ParentId":null,"OwnerUserId":"3312122","Title":null,"Body":"<p>Documentation on how to convert into Core ML's <code>.mlmodel<\/code>s: <a href=\"https:\/\/developer.apple.com\/documentation\/coreml\/converting_trained_models_to_core_ml\" rel=\"nofollow noreferrer\">https:\/\/developer.apple.com\/documentation\/coreml\/converting_trained_models_to_core_ml<\/a><\/p>\n\n<p>For TensorFlow models, you need to convert it to <code>frozen.pb<\/code> model as an input for the convert function. Here are some examples: <a href=\"https:\/\/github.com\/tf-coreml\/tf-coreml\/tree\/master\/examples\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tf-coreml\/tf-coreml\/tree\/master\/examples<\/a><\/p>\n"},{"AnswerId":"58097364","CreationDate":"2019-09-25T11:30:03.743","ParentId":null,"OwnerUserId":"10171211","Title":null,"Body":"<p>If you have downloaded a <code>.py<\/code> file you haven't downloaded the model, but the whole Python script. <\/p>\n\n<p>When you have your model compiled, you can use the python package <code>coremltools<\/code> to convert the TensorFlow\/Keras model into a <code>.mlmodel<\/code> file. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import coremltools\n\ncore_mlmodel = coremltools.converters.keras.convert(your_keras_model, respect_trainable=True)\ncore_mlmodel.save(\"YourModel.mlmodel\")\n<\/code><\/pre>\n\n<p>Then you will find your model on the left in <code>Files<\/code>. <\/p>\n"}]}
+{"QuestionId":58076586,"AnswerCount":0,"Tags":"<python><tensorflow><object-detection-api>","CreationDate":"2019-09-24T08:56:47.253","AcceptedAnswerId":null,"OwnerUserId":2467772.0,"Title":"Keep on evaluating in Training with Tensorflow Object Detection","Body":"<p>Using <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/object_detection\" rel=\"nofollow noreferrer\">Tensorflow object detection training<\/a>,training on FRCNN inception net v2.\nThe configuration for training is as follows.<\/p>\n\n<pre><code>optimizer {\n    momentum_optimizer: {\n      learning_rate: {\n        manual_step_learning_rate {\n          initial_learning_rate: 0.0002\n          schedule {\n            step: 50000\n            learning_rate: .00002\n          }\n          schedule {\n            step: 60000\n            learning_rate: .000002\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: \"\/home\/itc\/Data\/Cheers_Store\/TrainedModels\/FRcnnInceptionNet\/pretrained\/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 COCO 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: 70000\n<\/code><\/pre>\n\n<p>Trained for 70000 iterations with 30000 images.<\/p>\n\n<p>Evaluation with 3000 images and setting is as follow.<\/p>\n\n<pre><code>eval_config: {\n  num_examples: 3000\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\n<p>During training, it trained upto 4200 iterations. Then keep on evaluating and no more training. What could be the issue?<\/p>\n","answers":[]}
+{"QuestionId":58077635,"AnswerCount":1,"Tags":"<python><keras><computer-vision><classification><data-science>","CreationDate":"2019-09-24T09:53:59.503","AcceptedAnswerId":null,"OwnerUserId":11175275.0,"Title":"Why is my CNN pre trained image classifier overfitting?","Body":"<p>I have just started with Computer Vision and in the current task i am classifying images in 4 categories.<\/p>\n\n<p>Total number of image files=1043<\/p>\n\n<p>I am using pretrained InceptionV3 and fine tuning it on my dataset.<\/p>\n\n<p>This is what i have after the epoch:\nEpoch 1\/5\n320\/320 [==============================] - 1925s 6s\/step - loss: 0.4318 - acc: 0.8526 - val_loss: 1.1202 - val_acc: 0.5557<\/p>\n\n<p>Epoch 2\/5\n320\/320 [==============================] - 1650s 5s\/step - loss: 0.1807 - acc: 0.9446 - val_loss: 1.2694 - val_acc: 0.5436<\/p>\n\n<p>Epoch 3\/5\n320\/320 [==============================] - 1603s 5s\/step - loss: 0.1236 - acc: 0.9572 - val_loss: 1.2597 - val_acc: 0.5546<\/p>\n\n<p>Epoch 4\/5\n320\/320 [==============================] - 1582s 5s\/step - loss: 0.1057 - acc: 0.9671 - val_loss: 1.3845 - val_acc: 0.5457<\/p>\n\n<p>Epoch 5\/5\n320\/320 [==============================] - 1580s 5s\/step - loss: 0.0982 - acc: 0.9700 - val_loss: 1.2771 - val_acc: 0.5572\nThat is a huge difference. Kindly help me to figure out why is my model not able to generalize as it is fitting quite well on the train data.<\/p>\n\n<p>my code for reference:-<\/p>\n\n<pre><code>from keras.utils import to_categorical\nfrom keras.models import Model\nfrom keras.layers import Dense, GlobalAveragePooling2D, Dropout\nfrom keras.applications.inception_v3 import InceptionV3, preprocess_input\n\nCLASSES = 4\n\n# setup model\nbase_model = InceptionV3(weights='imagenet', include_top=False)\nfrom sklearn.preprocessing import OneHotEncoder\nx = base_model.output\nx = GlobalAveragePooling2D(name='avg_pool')(x)\nx = Dropout(0.4)(x)\npredictions = Dense(CLASSES, activation='softmax')(x)\nmodel = Model(inputs=base_model.input, outputs=predictions)\n\nfor layer in base_model.layers:\n    layer.trainable = False\n\nmodel.compile(optimizer='rmsprop',\n              loss='categorical_crossentropy',\n              metrics=['accuracy'])        \n\nfrom sklearn.preprocessing import LabelEncoder\nencoder = LabelEncoder()\ndf['Category']= encoder.fit_transform(df['Category'])\n\n\n\nfrom keras.preprocessing.image import ImageDataGenerator\n\nWIDTH = 299\nHEIGHT = 299\nBATCH_SIZE = 32\n\n\ntrain_datagen = ImageDataGenerator(rescale=1.\/255,preprocessing_function=preprocess_input)\n\n\n\nvalidation_datagen = ImageDataGenerator(rescale=1.\/255)\n\n\ndf['Category'] =df['Category'].astype(str)\n#dfval['Category'] = dfval['Category'].astype(str)\n\n\n\n\nfrom sklearn.utils import shuffle\ndf = shuffle(df)\n\nfrom sklearn.model_selection import train_test_split\n\ndftrain,dftest = train_test_split(df, test_size = 0.2, random_state = 0)\n\ntrain_generator = train_datagen.flow_from_dataframe(dftrain,target_size=(HEIGHT, WIDTH),batch_size=BATCH_SIZE,class_mode='categorical', x_col='Path', y_col='Category')\n\nvalidation_generator = validation_datagen.flow_from_dataframe(dftest,target_size=(HEIGHT, WIDTH),batch_size=BATCH_SIZE,class_mode='categorical', x_col='Path', y_col='Category')\n\n\n\nEPOCHS = 5\nBATCH_SIZE = 32\nSTEPS_PER_EPOCH = 320\nVALIDATION_STEPS = 64\n\nMODEL_FILE = 'filename.model'\n\nhistory = model.fit_generator(\n    train_generator,\n    epochs=EPOCHS,\n    steps_per_epoch=STEPS_PER_EPOCH,\n    validation_data=validation_generator,\n    validation_steps=VALIDATION_STEPS)\n<\/code><\/pre>\n\n<p>Any help would be appreciated :)<\/p>\n","answers":[{"AnswerId":"58081345","CreationDate":"2019-09-24T13:23:54.660","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>If you don't use <code>preprocess_input<\/code> in \"all\" your data, you will get terrible results. <\/p>\n\n<p>Look at these:<\/p>\n\n<pre><code>train_datagen = ImageDataGenerator(\n    preprocessing_function=preprocess_input, \n    ...)\n\nvalidation_datagen = ImageDataGenerator()\n<\/code><\/pre>\n\n<p>Now, I notice you are using <code>rescale<\/code>. Since you imported the correct <code>preprocess_input<\/code> function from the inception code, I really think you should not be using this rescale. The <code>preprocess_input<\/code> function is supposed to do all the necessary preprocessing. (Not all models were trained with normalized inputs)    <\/p>\n\n<p><strong>But would rescale be a problem if you're applying it to both batasets?<\/strong><\/p>\n\n<p>Well... if the <code>trainable=False<\/code> applied correctly to the <code>BatchNormalization<\/code> layers, this means that these layers have stored values for mean and variation which will only work well if the data is within the expected range.<\/p>\n"}]}
+{"QuestionId":58078858,"AnswerCount":1,"Tags":"<image><tensorflow><coreml>","CreationDate":"2019-09-24T11:02:30.343","AcceptedAnswerId":null,"OwnerUserId":11856751.0,"Title":"Is it possible to change the type of the input tensor in the coreml model from multiarray to image?","Body":"<p>I have a coreml model with those parameters and I want to change the type of the input tensor from Multiarray with shapes (112,112,3) to RGB image(112,112)<\/p>\n\n<pre><code>Core ML input(s):\n[name: \"input_to_float__0\"\ntype {\nmultiArrayType {\nshape: 3\nshape: 112\nshape: 112\ndataType: DOUBLE\n}\n}\n]\nCore ML output(s):\n[name: \"logits__BiasAdd__0\"\ntype {\nmultiArrayType {\nshape: 80\ndataType: DOUBLE\n} \n}\n]\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/viH8q.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/viH8q.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Is it possible to do that?<\/p>\n\n<p>How can I do that?<\/p>\n","answers":[{"AnswerId":"58084344","CreationDate":"2019-09-24T16:09:58.103","ParentId":null,"OwnerUserId":"7501629","Title":null,"Body":"<p>It's easiest to specify this when you convert the model to Core ML (using <code>image_input_names<\/code>). <\/p>\n\n<p>But you can also fix this using a bit of Python:<\/p>\n\n<pre><code>import coremltools\nimport coremltools.proto.FeatureTypes_pb2 as ft\nspec = coremltools.utils.load_spec(\"YourModel.mlmodel\")\ninput = spec.description.input[0]\ninput.type.imageType.colorSpace = ft.ImageFeatureType.RGB\ninput.type.imageType.height = 112\ninput.type.imageType.width = 112\ncoremltools.utils.save_spec(spec, \"YourNewModel.mlmodel\")\n<\/code><\/pre>\n"}]}
+{"QuestionId":58080527,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras><deep-learning>","CreationDate":"2019-09-24T12:38:33.257","AcceptedAnswerId":null,"OwnerUserId":11591599.0,"Title":"How to fix 'IndexError: tuple index out of range' in model.fit with tensorflow?","Body":"<p>I am trying to create a CNN with Tensorflow and keras with Sequential method.\nThe inputs have a (size, 50, 50, 1) shape and the labels have (size,). Size is the number of data in the dataset.\nThe problem is, after compilation, when I call the fit method with my model, I get an index error. See the code bellow :<\/p>\n\n<pre><code>model = tf.keras.models.Sequential()\nmodel.add(tf.keras.layers.Conv2D(32, 4, input_shape=(50, 50, 1), activation='relu', name=\"conv1\"))\nmodel.add(tf.keras.layers.Conv2D(64, 3, activation='relu', name=\"conv2\"))\nmodel.add(tf.keras.layers.Conv2D(128, 3, activation='relu', name=\"conv3\"))\nmodel.add(tf.keras.layers.Flatten(name='Flatten'))\nmodel.add(tf.keras.layers.Dense(128, activation='relu', name=\"d1\"))\nmodel.add(tf.keras.layers.Dense(4, activation='softmax', name=\"output\"))\n\n# Compile the model\nmodel.compile(\n    loss=\"sparse_categorical_crossentropy\",\n    optimizer=\"Adam\",\n    metrics=[\"accuracy\"]\n)\n\nmodel.fit(x_trains, y_labels, epochs=5, verbose=2, validation_data=0.33, shuffle=True)\n<\/code><\/pre>\n\n<p>And the error :<\/p>\n\n<pre><code>2019-09-24 13:59:40.902561: I tensorflow\/core\/platform\/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2\nTraceback (most recent call last):\n  File \"C:\/Users\/eloim\/Documents\/Programmation\/Python\/CNN_tf\/face_train_seq.py\", line 49, in &lt;module&gt;\n    model.fit(x_trains, y_labels, epochs=5, verbose=2, validation_data=0.33, shuffle=True)\n  File \"C:\\Users\\{}\\Anaconda3\\envs\\CNN_tf\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training.py\", line 728, in fit\n    use_multiprocessing=use_multiprocessing)\n  File \"C:\\Users\\{}\\Anaconda3\\envs\\CNN_tf\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2.py\", line 224, in fit\n    distribution_strategy=strategy)\n  File \"C:\\Users\\{}\\Anaconda3\\envs\\CNN_tf\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2.py\", line 564, in _process_training_inputs\n    distribution_strategy=distribution_strategy)\n  File \"C:\\Users\\{}\\Anaconda3\\envs\\CNN_tf\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\training_v2.py\", line 606, in _process_inputs\n    use_multiprocessing=use_multiprocessing)\n  File \"C:\\Users\\{}\\Anaconda3\\envs\\CNN_tf\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\data_adapter.py\", line 479, in __init__\n    batch_size=batch_size, shuffle=shuffle, **kwargs)\n  File \"C:\\Users\\{}\\Anaconda3\\envs\\CNN_tf\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\data_adapter.py\", line 238, in __init__\n    num_samples = set(int(i.shape[0]) for i in nest.flatten(inputs))\n  File \"C:\\Users\\{}\\Anaconda3\\envs\\CNN_tf\\lib\\site-packages\\tensorflow_core\\python\\keras\\engine\\data_adapter.py\", line 238, in &lt;genexpr&gt;\n    num_samples = set(int(i.shape[0]) for i in nest.flatten(inputs))\nIndexError: tuple index out of range\n<\/code><\/pre>\n\n<p>At the line :<\/p>\n\n<pre><code>model.fit(x_trains, y_labels, epochs=5, verbose=2, validation_data=0.33, shuffle=True)\n<\/code><\/pre>\n\n<p>I do not understand the nature of this error. How can I be rid of it ?\nThanks for your help.<\/p>\n","answers":[]}
+{"QuestionId":58081042,"AnswerCount":1,"Tags":"<python><tensorflow><cmd>","CreationDate":"2019-09-24T13:07:35.460","AcceptedAnswerId":58140591.0,"OwnerUserId":11053554.0,"Title":"How to test a tflite model with an image in the cmd line?","Body":"<p>I recently converted my 'output_graph.pb' to 'Skystone.tflite', but when I put it into my application it showed no objects detected.  How can I test that I trained the model properly?  Is there a way use the cmd line to give it an image and get a confidence level?<\/p>\n","answers":[{"AnswerId":"58140591","CreationDate":"2019-09-27T19:47:52.400","ParentId":null,"OwnerUserId":"11053554","Title":null,"Body":"<p>Alright, so after a long time of research I found a script called Label_Image, from the Tensorflow for poets 2.  This script's inputs were 'graph' and 'image', which worked out perfectly.<\/p>\n\n<p><a href=\"https:\/\/codelabs.developers.google.com\/codelabs\/tensorflow-for-poets-2\" rel=\"nofollow noreferrer\">tensorflow for poets 2<\/a><\/p>\n"}]}
+{"QuestionId":58081599,"AnswerCount":0,"Tags":"<tensorflow><ubuntu><ubuntu-18.04>","CreationDate":"2019-09-24T13:36:58.837","AcceptedAnswerId":null,"OwnerUserId":12113433.0,"Title":"I do not know if my EC2 instance (Ubuntu) 18.04 has GPU support (Tensorflow)","Body":"<p>When I run <code>lspci<\/code> on the Ubuntu prompt, it returns<\/p>\n\n<pre><code>00:00.0 Host bridge: Intel Corporation 440FX - 82441FX PMC [Natoma] (rev 02)\n00:01.0 ISA bridge: Intel Corporation 82371SB PIIX3 ISA [Natoma\/Triton II]\n00:01.1 IDE interface: Intel Corporation 82371SB PIIX3 IDE [Natoma\/Triton II]\n00:01.3 Bridge: Intel Corporation 82371AB\/EB\/MB PIIX4 ACPI (rev 01)\n00:02.0 VGA compatible controller: Cirrus Logic GD 5446\n00:03.0 Unassigned class [ff80]: XenSource, Inc. Xen Platform Device (rev 01)\n<\/code><\/pre>\n\n<p>One tutorial says: Make sure nothing will return when you run (not sure what this means) <code>lspci<\/code><\/p>\n\n<p>Another one says after I run <\/p>\n\n<blockquote>\n  <p>lspci<\/p>\n<\/blockquote>\n\n<p>: If you do not see any settings, update the PCI hardware database that Linux maintains by entering update-pciids (generally found in \/sbin) at the command line and rerun the previous lspci command.<\/p>\n\n<p>So just wondering if I set things up correctly<\/p>\n","answers":[]}
+{"QuestionId":58081650,"AnswerCount":1,"Tags":"<python-2.7><machine-learning><keras><entropy><softmax>","CreationDate":"2019-09-24T13:39:29.100","AcceptedAnswerId":58081723.0,"OwnerUserId":10623255.0,"Title":"Extracting output before the softmax layer, then manually calculating softmax gives a different result","Body":"<p>I have a model trained to classify rgb values into 1000 categories. <\/p>\n\n<pre><code>#Model architecture\nmodel = Sequential()\nmodel.add(Dense(512,input_shape=(3,),activation=\"relu\"))\nmodel.add(BatchNormalization())\nmodel.add(Dense(512,activation=\"relu\"))\nmodel.add(BatchNormalization())\nmodel.add(Dense(1000,activation=\"relu\"))\nmodel.add(Dense(1000,activation=\"softmax\"))\n<\/code><\/pre>\n\n<p>I want to be able to extract the output before the softmax layer so I can conduct analyses on different samples of categories within the model. I want execute softmax for each sample, and conduct analyses using a function named getinfo(). <\/p>\n\n<ol>\n<li><p>Model\nInitially, I enter X_train data into model.predict, to get a vector of 1000 probabilities for each input. I execute getinfo() on this array to get the desired result.<\/p><\/li>\n<li><p>Pop1\nI then use model.pop() to remove the softmax layer. I get new predictions for the popped model, and execute scipy.special.softmax. However, getinfo() produces an entirely different result on this array.<\/p><\/li>\n<li><p>Pop2\nI write my own softmax function to validate the 2nd result, and I receive an almost identical answer to Pop1.<\/p><\/li>\n<li><p>Pop3\nHowever, when I simply calculate getinfo() on the output of model.pop() with no softmax function, I get the same result as the initial Model.<\/p><\/li>\n<\/ol>\n\n<pre><code>data = np.loadtxt(\"allData.csv\",delimiter=\",\")\nmodel = load_model(\"model.h5\")\n\ndef getinfo(data):\n    objects = scipy.stats.entropy(np.mean(data, axis=0), base=2)\n    print(('objects_mean',objects))\n    colours_entropy = []\n    for i in data:\n        e = scipy.stats.entropy(i, base=2)\n        colours_entropy.append(e)\n    colours = np.mean(np.array(colours_entropy))\n    print(('colours_mean',colours))\n    info = objects - colours\n    print(('objects-colours',info))\n    return info\n\ndef softmax_max(data):\n    # calculate softmax whilst subtracting the max values (axis=1)\n    sm = []\n    count = 0\n    for row in data:\n        max = np.argmax(row)\n        e = np.exp(row-data[count,max])\n        s = np.sum(e)\n        sm.append(e\/s)\n    sm = np.asarray(sm)\n    return sm\n\n#model\npreds = model.predict(X_train)\ngetinfo(preds)\n\n#pop1\nmodel.pop()\npreds1 = model.predict(X_train)\nsm1 = scipy.special.softmax(preds1,axis=1)\ngetinfo(sm1)\n\n#pop2\nsm2 = softmax_max(preds1)\ngetinfo(sm2)\n\n#pop3\ngetinfo(preds1)\n<\/code><\/pre>\n\n<p>I expect to get the same output from Model, Pop1 and Pop2, but a different answer to Pop3, as I did not compute softmax here. I wonder if the issue is with computing softmax after model.predict? And whether I am getting the same result in Model and Pop3 because softmax is constraining the values between 0-1, so for the purpose of the getinfo() function, the result is mathematically equivalent? <\/p>\n\n<p>If this is the case, then how do I execute softmax before model.predict? <\/p>\n\n<p>I've gone around in circles with this, so any help or insight would be much appreciated. Please let me know if anything is unclear. Thank you!<\/p>\n","answers":[{"AnswerId":"58081723","CreationDate":"2019-09-24T13:43:23.997","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p><code>model.pop()<\/code> does not immediately have an effect. You need to run <code>model.compile()<\/code> again to recompile the new model that doesn't include the last layer.<\/p>\n\n<p>Without the recompile, you're essentially running <code>model.predict()<\/code> twice in a row on the exact same model, which explains why Model and Pop3 give the same result. Pop1 and Pop2 give weird results because they are calculating the softmax of a softmax.<\/p>\n\n<p>In addition, your model does not have the softmax as a separate layer, so <code>pop<\/code> takes off the entire last <code>Dense<\/code> layer. To fix this, add the softmax as a separate layer like so:<\/p>\n\n<pre><code>model.add(Dense(1000))           # softmax removed from this layer...\nmodel.add(Activation('softmax')) # ...and added to its own layer\n<\/code><\/pre>\n"}]}
+{"QuestionId":58081798,"AnswerCount":1,"Tags":"<tensorflow><keras><tensorflow2.0>","CreationDate":"2019-09-24T13:47:33.597","AcceptedAnswerId":58081853.0,"OwnerUserId":7657215.0,"Title":"Keras at TF2 metrics not added","Body":"<p>I'm using Tensorflow 2.0 nightly build, on google colab.<\/p>\n\n<p>I made simple CNN model, and than compiled it, and fit it.<\/p>\n\n<p>Here's code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = tf.keras.models.Sequential([\n    tf.keras.layers.Reshape((28, 28, 1)),\n    tf.keras.layers.Conv2D(filters=32, kernel_size=(3, 3), padding='SAME',\n                           activation=tf.nn.relu),\n    tf.keras.layers.MaxPooling2D(pool_size=(2, 2)),\n    tf.keras.layers.Conv2D(filters=64, kernel_size=(3, 3), padding='SAME',\n                           activation=tf.nn.relu),\n    tf.keras.layers.MaxPooling2D(pool_size=(2, 2)),\n    tf.keras.layers.Flatten(),\n    tf.keras.layers.Dense(256, activation=tf.nn.relu),\n    tf.keras.layers.Dropout(0.3),\n    tf.keras.layers.Dense(10),\n    tf.keras.layers.Softmax()\n])\n\n\noptimizer = tf.keras.optimizers.Adam(0.001)\nmodel.compile(optimizer=optimizer,\n              loss=tf.keras.losses.CategoricalCrossentropy(),\n              matrics=['accuracy'])\nlog_dir='.\/logs'\ntensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=log_dir,\n                                                      histogram_freq=2,\n                                                      write_images=True,\n                                                      update_freq='batch',\n                                                      profile_batch=0)\nmodel.fit(x=x_train, y=y_train, batch_size=100, epochs=15,\n          callbacks=[tensorboard_callback], validation_data=(x_test, y_test))\n<\/code><\/pre>\n\n<p>And it don't give me accuracy information.<\/p>\n\n<p>I evaluated model, and it supposed to give me accuracy information, but it only gives me loss information.<\/p>\n\n<p>I printed <code>model.metrics<\/code>, and it was just <code>[]<\/code>.<\/p>\n\n<p>Is it bug? Or I missed something? <\/p>\n","answers":[{"AnswerId":"58081853","CreationDate":"2019-09-24T13:50:36.440","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>You misspelled <code>metrics<\/code> as <code>matrics<\/code>. Change <code>matrics=['accuracy']<\/code> to <code>metrics=['accuracy']<\/code>.<\/p>\n"}]}
+{"QuestionId":58082406,"AnswerCount":0,"Tags":"<tensorflow><neural-network><graph-algorithm>","CreationDate":"2019-09-24T14:19:56.137","AcceptedAnswerId":null,"OwnerUserId":6333797.0,"Title":"how do data pass forward and backward in computational graph of TensorFlow if there is a branch?","Body":"<p>I do understand how sequential model work, data just pass through each layer in order one by one, but if there is a branch, for example, \nnode1->node2->node3->node4, node5->node3.\nI don't know how to compute the output of node4 properly. Using a recursive function is not smart because when the graph is very large, it wastes much time, I think. <\/p>\n\n<p>I want to write a simple program in python to learn this algorithm. Can anyone give me some advice or tell me what kind of theory I should learn?<\/p>\n","answers":[]}
+{"QuestionId":58082452,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-24T14:21:51.993","AcceptedAnswerId":null,"OwnerUserId":7853071.0,"Title":"ValueError: Only call `softmax_cross_entropy_with_logits","Body":"<p>I'm new to python and tensorflow\ni got this error when training a model <\/p>\n\n<pre><code> ValueError: Only call `softmax_cross_entropy_with_logits` with named arguments (labels=..., \n logits=..., ...)\n<\/code><\/pre>\n\n<p>i tried to change this line <\/p>\n\n<pre><code>losses = tf.nn.softmax_cross_entropy_with_logits(scores, self.y_input)\n<\/code><\/pre>\n\n<p>to <\/p>\n\n<pre><code> losses = tf.nn.softmax_cross_entropy_with_logits(logits=scores, labels=self.y_input)\n<\/code><\/pre>\n\n<p>without any effect .. <\/p>\n\n<pre><code>pip list | grep tensorflow\ntensorflow            1.14.0             \ntensorflow-estimator  1.14.0             \ntensorflow-gpu        1.14.0        \n<\/code><\/pre>\n\n<p>i tried the solutions i read it here but no effect .. any help ..     <\/p>\n","answers":[{"AnswerId":"58085754","CreationDate":"2019-09-24T17:50:48.227","ParentId":null,"OwnerUserId":"7853071","Title":null,"Body":"<p>i tried this line and worked <\/p>\n\n<pre><code>losses = tf.nn.softmax_cross_entropy_with_logits(_sentinel=None, labels=scores, logits = self.y_input, dim=-1, name=None )\n<\/code><\/pre>\n\n<p>means write all parameters even it's not will be use .. <\/p>\n"}]}
+{"QuestionId":58082522,"AnswerCount":0,"Tags":"<tensorflow><keras><tensorflow2.0>","CreationDate":"2019-09-24T14:25:18.993","AcceptedAnswerId":null,"OwnerUserId":287238.0,"Title":"In tensorflow keras, are there any good patterns to train the regularization parameters over a validation data but not on train data?","Body":"<p>Am thinking of something where you have a choice of either<\/p>\n\n<p>a) leaving out the l1, l1 etc params on the trainable_variables list in tape.gradient(loss, trainable_variables)<\/p>\n\n<p>or <\/p>\n\n<p>b) using stop-gradient with some train\/val trigger.<\/p>\n\n<p>There is a bit of book-keeping and am wondering if there are good patterns for this or if it known to be a bad idea for some other reason. <\/p>\n","answers":[]}
+{"QuestionId":58082534,"AnswerCount":1,"Tags":"<python-3.x><pytorch>","CreationDate":"2019-09-24T14:25:56.100","AcceptedAnswerId":58089590.0,"OwnerUserId":12113684.0,"Title":"Concept of mini batch in deep generative model using pyro","Body":"<p>I am new to probabilistic programming and ML. I am following a code on deep Markov model given on pyro's website. The link to the github page to that code is:<\/p>\n\n<p><a href=\"https:\/\/github.com\/pyro-ppl\/pyro\/blob\/dev\/examples\/dmm\/dmm.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pyro-ppl\/pyro\/blob\/dev\/examples\/dmm\/dmm.py<\/a><\/p>\n\n<p>I understand most part of the code. The part I don't understand is mini batch idea they are using from line 175. <\/p>\n\n<p><strong>Question 1:<\/strong>\n<strong>Could someone explain what are they doing there when they are using mini-batch?<\/strong><\/p>\n\n<p>In pyro documentation they say <\/p>\n\n<blockquote>\n  <p>mini_batch is a three dimensional tensor, with the first dimension being the batch dimension, the second dimension being the temporal dimension, and the final dimension being the features (88-dimensional in our case)'<\/p>\n<\/blockquote>\n\n<p><strong>Question 2:<\/strong> \n<strong>What does temporal dimension means here?<\/strong> <\/p>\n\n<p>Because I want to use this code on my dataset which is a sequential data. I have done one hot encoding of my data such that it's dimension is <strong>(10000,500,20)<\/strong> where 10000 is the number of examples\/Sequences, 500 is the length of each of these sequences and 20 is the number of features. <\/p>\n\n<p><strong>Question 3:<\/strong>\n<strong>How can I use my one hot encoded data as mini batch here?<\/strong> <\/p>\n\n<p>I'm sorry if it is a really basic question but, insights will be appreciated. <\/p>\n\n<p>Link to that documentation is:<\/p>\n\n<p><a href=\"https:\/\/pyro.ai\/examples\/dmm.html\" rel=\"nofollow noreferrer\">https:\/\/pyro.ai\/examples\/dmm.html<\/a><\/p>\n","answers":[{"AnswerId":"58089590","CreationDate":"2019-09-25T00:16:58.503","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<blockquote>\n  <p>Question 1: Could someone explain what are they doing there when they are using mini-batch?<\/p>\n<\/blockquote>\n\n<p>To optimize most of the deep learning models, we use mini-batch gradient descent. Here, A <code>mini_batch<\/code> refers to a small number of examples. Let's say, we have 10,000 training examples and we want to create mini-batches of 50 examples. So, in total there will be 200 mini-batches and we will perform 200 parameter updates during one iteration over the entire dataset.<\/p>\n\n<blockquote>\n  <p>Question 2: What does the temporal dimension mean here?<\/p>\n<\/blockquote>\n\n<p>In your data: <code>(10000, 500, 20)<\/code>, the second dimension refers to the temporal dimension. You can consider you have examples with 500 timesteps <code>(t1, t2, ..., t500)<\/code>.<\/p>\n\n<blockquote>\n  <p>Question 3: How can I use my one-hot encoded data as mini-batch here?<\/p>\n<\/blockquote>\n\n<p>In your scenario, you can split your data <code>(10000, 500, 20)<\/code> into 200 small batches of size <code>(50, 500, 20)<\/code> where 50 is the number of examples\/Sequences in the mini-batch, 500 is the length of each of these sequences and 20 is the number of features.<\/p>\n\n<p>How do we decide the mini-batch size? Basically, we can tune the batch size just like any other hyperparameters of our model.<\/p>\n"}]}
+{"QuestionId":58082856,"AnswerCount":0,"Tags":"<java><python><tensorflow>","CreationDate":"2019-09-24T14:42:42.013","AcceptedAnswerId":null,"OwnerUserId":5587080.0,"Title":"No Operation named [*] in the Graph","Body":"<p>I have training a model with python, I want to use is with java.<\/p>\n\n<p>python 3.7.4, java 8<\/p>\n\n<p>tf in python 1.14.0 | tf in java 1.14.0<\/p>\n\n<p>I have write this code to make sure states is included in graph<\/p>\n\n<pre><code>states = tf.identity(states, name=\"states\")\n<\/code><\/pre>\n\n<p>Here states is the result from <code>dynamic_rnn<\/code><\/p>\n\n<p>Before I save this model, I check all name of node<\/p>\n\n<pre><code>print(\"\\n\".join([n.name for n in sess.graph.as_graph_def().node]))\n<\/code><\/pre>\n\n<p>And yes, <code>states<\/code> is in.<\/p>\n\n<p>And I load pb file in java.<\/p>\n\n<p>An error be raised. <code>No Operation named [*] in the Graph<\/code><\/p>\n\n<p>So, For make sure it again, I use this code to check.<\/p>\n\n<pre><code>graph.operations().forEachRemaining(c-&gt;out.println(c.name()));\n<\/code><\/pre>\n\n<p>And No, not only the <code>states<\/code> not in this graph, but also a lot of other operators are not in.<\/p>\n\n<p>For example<\/p>\n\n<pre><code>gradients\/rnn\/while\/Enter_3_grad\/Exit\ndense\/kernel\/Initializer\/random_uniform\/min\nReshape\/shape\nLSTMCellZeroState\/ExpandDims_1\n<\/code><\/pre>\n\n<p>And so on that I have never explicit declare it.<\/p>\n\n<p>So I want to ask, if there are some node be ignored when load pb file with java.<\/p>\n","answers":[]}
+{"QuestionId":58083288,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-24T15:06:22.000","AcceptedAnswerId":null,"OwnerUserId":9306691.0,"Title":"Trained network predicts NaN output","Body":"<p>I am working on an autoencoder to encode\/decode german titles of articles (~2k samples). I am using byte pair encoding as a compromise between character encoding and word encoding. This is the model i am using for this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def create_model(characters, enc_size, latent_dim): #my parameter values: 32, 512, 32\n    inputs = tf.keras.layers.Input(shape=(characters, enc_size))\n    encoded = tf.keras.layers.CuDNNLSTM(128, return_sequences=True)(inputs)\n    encoded = tf.keras.layers.CuDNNLSTM(latent_dim)(encoded)\n\n    vector = tf.keras.layers.Input(shape=(latent_dim,))\n    decoded = tf.keras.layers.RepeatVector(characters)(vector)\n    decoded = tf.keras.layers.CuDNNLSTM(128, return_sequences=True)(decoded)\n    decoded = tf.keras.layers.CuDNNLSTM(enc_size, return_sequences=True)(decoded)\n\n    encoder = tf.keras.models.Model(inputs, encoded)\n    decoder = tf.keras.models.Model(vector, decoded)\n\n    auto = encoder(inputs)\n    auto = decoder(auto)\n    auto = tf.keras.models.Model(inputs, auto)\n\n    opt = tf.keras.optimizers.Adam(lr=1e-3)\n    auto.compile(optimizer=opt,\n                  loss='categorical_crossentropy',\n                  metrics=[\"accuracy\"])\n    auto.summary()\n\n    return auto, encoder, decoder\n<\/code><\/pre>\n\n<p>I created its structure, by modifying the Sequence-to-Sequence example in this keras <a href=\"https:\/\/blog.keras.io\/building-autoencoders-in-keras.html\" rel=\"nofollow noreferrer\">article<\/a>. Training this model is no problem. The loss decreases, while the accuracy increases. During this whole process both never become NaN. Now, when i try to use it in prediction (model.predict) the output is an array filled with NaNs. The strange thing is, that if i use \"binary_crossentropy\" (although it is not really suited for this task, because byte pairs are one hot encoded) instead of \"categorical_crossentropy\" this does not happen and the prediction after training is not NaN. But it seems, that the actual output does not really improve, even after 200 epochs. I cant really tell by the \"binary_accuracy\" because of the fact, that it is unsuited.<\/p>\n\n<p>One of the best predicted examples:<\/p>\n\n<blockquote>\n  <p>Original: \"luegde einbruch in wohnhaus\" | Prediction: \"moren einbruch in einbruch in diebstagesuchteinbruchbrandeses\"<\/p>\n<\/blockquote>\n\n<p>I am also certain, that the dataset is not contaminated with NaNs or Infs.\nDoes anyone knwo a solution\/reason for this strange behaviour, or knows a working example of an autoencoder for text with encoder and decoder model.\nThanks in advance!<\/p>\n","answers":[]}
+{"QuestionId":58083399,"AnswerCount":0,"Tags":"<python><tensorflow><keras><keras-layer><tf.keras>","CreationDate":"2019-09-24T15:14:44.230","AcceptedAnswerId":null,"OwnerUserId":2572750.0,"Title":"How to get Pairwise Euclidean Distance of Two tensors (shape=[?,256,512]) inside of Lambda layer in tensorflow keras?","Body":"<p>I have two tensors of shape (?,256,512). My aim is to calculate the Pairwise Euclidean Distance along axis=2 (third axis). I have tried <a href=\"https:\/\/stackoverflow.com\/questions\/43839120\/compute-the-pairwise-distance-between-each-pair-of-the-two-collections-of-inputs?fbclid=IwAR0YnmM3JAisf29kdmfknIJYw-VEH8tZ1QG5ekerR13zdpIYWFbdUB9t--M\">this<\/a> solution. In this solution they used 2-D tensors of known shape (e.g: (2,3)). However, in my case, I am trying to build a custom layer using input shapes of both (?,256,512) where ?=batch number, 256=feature number and 512=feature depth. The expected output shape is (256,256).<\/p>\n\n<p>The problem I am facing is due to the '?' in the mentioned shapes while building the model. I am looking for a solution that can be implemented specifically using tensorflow.<\/p>\n","answers":[]}
+{"QuestionId":58083502,"AnswerCount":1,"Tags":"<python><keras><deep-learning>","CreationDate":"2019-09-24T15:19:29.570","AcceptedAnswerId":null,"OwnerUserId":11192771.0,"Title":"Binary Classification Neural Network: both Nan loss and NaN predictions","Body":"<p>This model tries to predict two states based on an array with 400 numbers. During the first training round the model starts with loss on the first +- 200 samples, and then goes into Nan loss. The accuracy stays around 50% and when I print the predictions for the test set, it only predicts NaN. My <code>X_train<\/code> has a shape of <code>(1934, 400, 1)<\/code> and my <code>y_train<\/code> a shape of <code>(1934,)<\/code>. I already tried checking for NaNs in the dataset, but there were none.<\/p>\n\n<p>My model looks like this:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(128, input_shape=(400,1), activation='relu', return_sequences=True))\nmodel.add(Dropout(0,2))\nmodel.add(LSTM(128, activation='relu'))\nmodel.add(Dropout(0,2))\nmodel.add(Dense(32, activation='relu'))\nmodel.add(Dropout(0,2))\nmodel.add(Dense(1, activation='sigmoid'))\n\nopt = tf.keras.optimizers.Adam(lr=0.01)\n\n# mean_squared_error = mse\nmodel.compile(loss='binary_crossentropy',\n              optimizer=opt,\n              metrics=['accuracy'])\n\nhistory = model.fit(X_train, y_train, epochs=3, validation_split = 0.1, shuffle=True, batch_size = 64)\n<\/code><\/pre>\n\n<p>edit: Solved by changing the activation function to tanh. Sigmoid stays sigmoid!<\/p>\n","answers":[{"AnswerId":"58098196","CreationDate":"2019-09-25T12:17:19.347","ParentId":null,"OwnerUserId":"11192771","Title":null,"Body":"<p>The problem was solved with changing the activation functions to \"tanh\". Seems that dropout should be 0.2 instead of 0,2 also, but this wasn't the cause of the problem.<\/p>\n"}]}
+{"QuestionId":58083713,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><predict>","CreationDate":"2019-09-24T15:32:37.567","AcceptedAnswerId":58108780.0,"OwnerUserId":6724265.0,"Title":"When i use estimator with same checkpoints to predict same file for several times, the predicted result varies","Body":"<p>I'm using textcnn model in estimator to classify some text. After i train the model, the trained model was stored in the form of  checkpoints. But when i try to predict the same test file with same checkpoints,the predicted result(porbalility and logits) varies slightly.<\/p>\n\n<ol>\n<li>I have set the dropout_keep_prob=1 in dropout layer<\/li>\n<li>checkpoints and test file remain the same one.<\/li>\n<li>I have used the LoggingTensorHook to check the tensor values during the predict, two values begin to vary at the max_pool step(at least the conv values are same but i am not sure)<\/li>\n<\/ol>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\n\ndef line_parser(line, vocab):\n    def parse_content(record):\n        items = record.decode().strip().split()\n        cat = int(items[-1])\n        tokens = items[:-1]\n        token_length = len(tokens)\n        if token_length &gt; FLAGS.max_sequence_length:\n            tokens = tokens[:FLAGS.max_sequence_length]\n        if token_length &lt; FLAGS.max_sequence_length:\n            tokens += [FLAGS.pad_word]*(FLAGS.max_sequence_length-token_length)\n        return [tokens, cat]\n\n    result = tf.py_func(parse_content, [line], [tf.string, tf.int64])\n    ids = vocab.lookup(result[0])\n    ids = tf.cast(ids, tf.int64)\n    ids = tf.reshape(ids, [FLAGS.max_sequence_length])\n    label = tf.one_hot(result[1], FLAGS.num_classes, dtype=tf.int32)\n    return [ids, label]\n\n\ndef predict_line_parser(line, vocab):\n    def parse_content(record):\n        feature = record.decode().strip()\n        tokens = feature.split()\n        token_length = len(tokens)\n        if token_length &gt; FLAGS.max_sequence_length:\n            tokens = tokens[:FLAGS.max_sequence_length]\n        if token_length &lt; FLAGS.max_sequence_length:\n            tokens += [FLAGS.pad_word]*(FLAGS.max_sequence_length-token_length)\n        return [tokens]\n\n    result = tf.py_func(parse_content, [line], [tf.string])\n    ids = vocab.lookup(result[0])\n    ids = tf.cast(ids, tf.int64)\n    ids = tf.reshape(ids, [FLAGS.max_sequence_length])\n    return ids\n\n\ndef train_input_fn(file_paths, batch_size):\n    vocab = tf.contrib.lookup.index_table_from_file(FLAGS.vocab_path)\n    dataset = tf.data.TextLineDataset(file_paths)\n    dataset = dataset.map(lambda line: line_parser(line, vocab))\n    dataset = dataset.shuffle(1000)\n    dataset = dataset.batch(batch_size).repeat()\n    return dataset\n\n\ndef eval_input_fn(file_paths, batch_size):\n    vocab = tf.contrib.lookup.index_table_from_file(FLAGS.vocab_path)\n    dataset = tf.data.TextLineDataset(file_paths)\n    dataset = dataset.map(lambda line: line_parser(line, vocab))\n    dataset = dataset.batch(batch_size=batch_size)\n    return dataset\n\n\ndef predict_input_fn(file_paths, batch_size):\n    vocab = tf.contrib.lookup.index_table_from_file(FLAGS.vocab_path)\n    dataset = tf.data.TextLineDataset(file_paths)\n    dataset = dataset.map(lambda line:predict_line_parser(line, vocab))\n    dataset = dataset.batch(batch_size=batch_size)\n    return dataset\n\n\ndef create_model(features, params):\n    # projection from sentence with id to embedding\n    embedding_inputs = tf.nn.embedding_lookup(params[\"embedding\"], features)\n    embedding_inputs = tf.expand_dims(embedding_inputs, axis=-1)\n    l2_loss = tf.constant(0.0, name=\"l2_loss\", dtype=\"float64\")\n\n    # convolutional layer and pooling layer\n    pooled_outputs = list()\n    for i, filter_size in enumerate(params[\"filter_sizes\"]):\n        with tf.name_scope(\"conv_{}\".format(filter_size)):\n            filter_shape = [filter_size, params[\"embedding_size\"], 1, params[\"num_filters\"]]\n            W = tf.Variable(tf.truncated_normal(filter_shape, stddev=0.1, dtype=\"float64\"), name=\"W\")\n            b = tf.Variable(tf.constant(0.1, shape=[params[\"num_filters\"]], dtype=\"float64\"), name=\"b\")\n            conv = tf.nn.conv2d(embedding_inputs, W, strides=[1, 1, 1, 1], padding=\"VALID\", use_cudnn_on_gpu=True,\n                                name=\"conv\".format(filter_size))\n            h = tf.nn.relu(tf.nn.bias_add(conv, b), name=\"relu\".format(filter_size))\n            pooled = tf.nn.max_pool(\n                h,\n                ksize=[1, params[\"sequence_length\"] - filter_size + 1, 1, 1],\n                strides=[1, 1, 1, 1],\n                padding=\"VALID\",\n                name=\"pool\".format(filter_size)\n            )\n            pooled_outputs.append(pooled)\n\n    # concatenate all feature vector\n    number_filters_total = params[\"num_filters\"] * len(params[\"filter_sizes\"])\n    h_pool = tf.concat(pooled_outputs, 3)\n    h_pool_flat = tf.reshape(h_pool, [-1, number_filters_total])\n\n    # dropout\n    with tf.name_scope(\"dropout\"):\n        # h_drop = tf.nn.dropout(h_pool_flat, params[\"dropout_keep_prob\"])\n        h_drop = tf.nn.dropout(h_pool_flat, 1)\n\n    # fully connected layer\n    with tf.name_scope(\"output\"):\n        W = tf.Variable(\n            tf.truncated_normal(shape=[number_filters_total, params[\"num_classes\"]], stddev=0.1, dtype=\"float64\"),\n            name=\"W\")\n        b = tf.Variable(tf.constant(0.1, shape=[params[\"num_classes\"]], dtype=\"float64\"), name=\"b\")\n        l2_loss += tf.nn.l2_loss(W)\n        l2_loss += tf.nn.l2_loss(b)\n        logits = tf.nn.xw_plus_b(h_drop, W, b, name=\"scores\")\n    return logits, l2_loss\n\n\ndef model_fn_builder():\n    def text_cnn_model_fn(features, labels, mode, params):\n        logits, l2_loss = create_model(features, params)\n        # train mode branch\n        if mode == tf.estimator.ModeKeys.TRAIN:\n            # loss\n            with tf.name_scope(\"loss\"):\n                losses = tf.nn.softmax_cross_entropy_with_logits_v2(labels=labels, logits=logits)\n                loss = tf.reduce_mean(losses) + params[\"l2_reg_lambda\"] * l2_loss\n            # optimizer function\n            with tf.name_scope(\"optimizer\"):\n                optimizer = tf.train.AdamOptimizer(params[\"learning_rate\"])\n                grads_and_vars = optimizer.compute_gradients(loss)\n                train_op = optimizer.apply_gradients(grads_and_vars, global_step=tf.train.get_global_step())\n            return tf.estimator.EstimatorSpec(mode=mode, loss=loss, train_op=train_op)\n\n        # eval mode branch\n        if mode == tf.estimator.ModeKeys.EVAL:\n            # loss\n            with tf.name_scope(\"loss\"):\n                losses = tf.nn.softmax_cross_entropy_with_logits_v2(labels=labels, logits=logits)\n                loss = tf.reduce_mean(losses) + params[\"l2_reg_lambda\"] * l2_loss\n\n            # predictions\n            with tf.name_scope(\"prediction\"):\n                probability = tf.nn.softmax(logits, axis=1, name=\"probability\")\n                pred = tf.argmax(probability, axis=1, name=\"predictions\")\n\n            # metrics\n            with tf.name_scope(\"metrics\"):\n                accuracy = tf.metrics.accuracy(labels=tf.argmax(labels, axis=1), predictions=pred)\n                precision = tf.metrics.precision(labels=tf.argmax(labels, axis=1), predictions=pred)\n                recall = tf.metrics.recall(labels=tf.argmax(labels, axis=1), predictions=pred)\n                tf.summary.scalar(\"accuracy\", accuracy[1])\n                tf.summary.scalar(\"precision\", precision[1])\n                tf.summary.scalar(\"recall\", recall[1])\n                tf.summary.scalar(\"loss\", loss)\n                metrics = {\"accuracy\": accuracy, \"precision\": precision, \"recall\": recall}\n                metric_hook = tf.train.LoggingTensorHook(\n                    {\"f1-score\": 2 * precision[1] * recall[1] \/ (precision[1] + recall[1]), \"precision\": precision[1],\n                     \"recall\": recall[1]}, every_n_iter=100)\n\n            return tf.estimator.EstimatorSpec(mode=mode, loss=loss, eval_metric_ops=metrics,\n                                              evaluation_hooks=[metric_hook])\n\n        # predict mode branch\n        if mode == tf.estimator.ModeKeys.PREDICT:\n            # predictions\n            with tf.name_scope(\"prediction\"):\n                probability = tf.nn.softmax(logits, axis=1, name=\"probability\")\n                pred = tf.argmax(probability, axis=1, name=\"predictions\")\n                predictions = {\n                    \"class\": pred,\n                    \"probability\": probability,\n                }\n            return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions)\n\n    return text_cnn_model_fn\n<\/code><\/pre>\n\n<p>I expect the same output of twice predictions,but it varies like the following:<br>\nfirst time<br>\n0\\0.02336916147480053<br>\n0\\0.29461604884471243<br>\n0\\0.04555523004833724<br>\n1\\0.5450933830551228<br>\n0\\0.042727966035733034<br>\n0\\0.032764190484837884<br>\n0\\0.11542703615898613<br>\n0\\0.12662708812885717<br>\n0\\0.01605587344580832<br>\n0\\0.006454832043875243<br>\nsecond time<br>\n0\\0.03389085341620636<br>\n0\\0.31563690653966603<br>\n0\\0.06185060165562852<br>\n1\\0.5891016184323346<br>\n0\\0.07184752629327144<br>\n0\\0.04355442431024522<br>\n0\\0.16290306166502935<br>\n0\\0.17214872864042816<br>\n0\\0.02437323886282706<br>\n0\\0.0109889405648392<\/p>\n","answers":[{"AnswerId":"58108780","CreationDate":"2019-09-26T02:29:20.603","ParentId":null,"OwnerUserId":"6724265","Title":null,"Body":"<p>Autually,i figured out this problem. The variety is resulted by  word embedding vectors which are generated randomly every time.<\/p>\n"}]}
+{"QuestionId":58083752,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-24T15:34:20.197","AcceptedAnswerId":null,"OwnerUserId":5772365.0,"Title":"apply tf.GradientTape to gradient descent with low gpu utility","Body":"<p>Below code is using gradient descent to calculate the minimum <code>y=x^2<\/code>. However it runs with <code>167Mib<\/code> gpu memory and <code>0%<\/code> gpu utility.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport tensorflow as tf\n\nsess = tf.InteractiveSession()\nx = tf.constant(3.0)\nfor _ in range(10000):\n    with tf.GradientTape() as g:\n        g.watch(x)\n        y = x * x\n    dy_dx = g.gradient(y, x)\n    x -= dy_dx * 0.001\nprint(x.eval())\n<\/code><\/pre>\n\n<p>I expect a non zero gpu utility. So is there any wrong with this code?<\/p>\n\n<p><strong>Platform Info<\/strong>:<\/p>\n\n<ul>\n<li>gpu: GeForce GTX 1080<\/li>\n<li>cuda 9.0<\/li>\n<li>python 3.5.2<\/li>\n<li>tensorflow 1.12.0<\/li>\n<\/ul>\n","answers":[]}
+{"QuestionId":58083832,"AnswerCount":0,"Tags":"<machine-learning><pytorch><spacy>","CreationDate":"2019-09-24T15:39:15.163","AcceptedAnswerId":null,"OwnerUserId":1455800.0,"Title":"spacy-pytorch-transformers: loss = 0 constantly during training","Body":"<p>I'm using <code>spacy 2.1.8<\/code> and <code>spacy-pytorch-transformers 0.4.0<\/code> to train a text classifier. My code is strongly inspired from their <a href=\"https:\/\/github.com\/explosion\/spacy-pytorch-transformers\/blob\/master\/examples\/train_textcat.py\" rel=\"nofollow noreferrer\">example<\/a> but the model is not learning anything, which seems to be caused by the loss being 0 all the time. A minimum (not-)working example of my code is as follows:<\/p>\n\n<pre><code>nlp = spacy.load(\"en_pytt_xlnetbasecased_lg\")\ntextcategorizer = nlp.create_pipe(\"pytt_textcat\", config={\"exclusive_classes\": True, \"architecture\": \"softmax_last_hidden\"})\n\nfor label in labels:\n    textcategorizer.add_label(label)\nnlp.add_pipe(textcategorizer, last=True)\n\noptimizer = nlp.resume_training()\n\nfor epoch in range(num_of_epochs):\n    np.random.shuffle(train)\n    losses = Counter()\n\n    for step, batch in enumerate(minibatch(train, size=batch_size)):\n        optimizer.pytt_lr = 0.005\n        texts, cats = zip(*batch)\n        _, cats = preprocessed_labels_to_categories_for_training_and_eval(cats)\n\n        nlp.update(texts, cats, sgd=optimizer, losses=losses, drop=0.1)\n<\/code><\/pre>\n\n<p>I already double and triple checked that relevant variables, such as <code>cats<\/code> and <code>texts<\/code>, contain valid and correct values. <\/p>\n\n<p>Is there anything I'm missing?<\/p>\n","answers":[]}
+{"QuestionId":58084056,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><pip>","CreationDate":"2019-09-24T15:51:32.607","AcceptedAnswerId":null,"OwnerUserId":10031044.0,"Title":"Can't install tensorflow on Windows","Body":"<p>This is the Error I am getting. Please help me out with this. I tried all methods of installing Tensorflow. I am on windows 10, 64bit.\nAlso, I am getting similar error on ubuntu as well, but I want the solution for windows for now. Please help me out with this.<\/p>\n\n<p>Commands: <\/p>\n\n<pre><code>C:\\Users\\chatbot&gt;py -m pip install tensorflow\nCollecting tensorflow\n  Downloading https:\/\/files.pythonhosted.org\/packages\/f7\/08\/25e47a53692c2e0dcd2211a493ddfe9007a5cd92e175d6dffa6169a0b392\/tensorflow-1.14.0-cp37-cp37m-win_amd64.whl (68.3MB)\n     |\u2588\u2588                              | 4.7MB 17kB\/s eta 1:00:29\n<\/code><\/pre>\n\n<p>ERROR:<\/p>\n\n<pre><code>Collecting tensorflow\n  Downloading https:\/\/files.pythonhosted.org\/packages\/bf\/4a\/5c86ed8b245aa48f9f819b13a0a9039e9126ba19fdd0c7e0b8026c12315a\/tensorflow-1.14.0-cp36-cp36m-win_amd64.whl (68.3MB)\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 | 66.1MB 32kB\/s eta 0:01:08ERROR: Exception:\nTraceback (most recent call last):\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_vendor\\urllib3\\response.py\", line 397, in _error_catcher\n    yield\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_vendor\\urllib3\\response.py\", line 479, in read\n    data = self._fp.read(amt)\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_vendor\\cachecontrol\\filewrapper.py\", line 62, in read\n    data = self.__fp.read(amt)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\http\\client.py\", line 449, in read\n    n = self.readinto(b)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\http\\client.py\", line 493, in readinto\n    n = self.fp.readinto(b)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\socket.py\", line 586, in readinto\n    return self._sock.recv_into(b)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\ssl.py\", line 1009, in recv_into\n    return self.read(nbytes, buffer)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\ssl.py\", line 871, in read\n    return self._sslobj.read(len, buffer)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\ssl.py\", line 631, in read\n    v = self._sslobj.read(len, buffer)\nsocket.timeout: The read operation timed out\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_internal\\cli\\base_command.py\", line 188, in main\n    status = self.run(options, args)\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_internal\\commands\\install.py\", line 345, in run\n    resolver.resolve(requirement_set)\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_internal\\legacy_resolve.py\", line 196, in resolve\n    self._resolve_one(requirement_set, req)\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_internal\\legacy_resolve.py\", line 359, in _resolve_one\n    abstract_dist = self._get_abstract_dist_for(req_to_install)\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_internal\\legacy_resolve.py\", line 307, in _get_abstract_dist_for\n    self.require_hashes\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_internal\\operations\\prepare.py\", line 199, in prepare_linked_requirement\n    progress_bar=self.progress_bar\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_internal\\download.py\", line 1064, in unpack_url\n    progress_bar=progress_bar\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_internal\\download.py\", line 924, in unpack_http_url\n    progress_bar)\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_internal\\download.py\", line 1152, in _download_http_url\n    _download_url(resp, link, content_file, hashes, progress_bar)\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_internal\\download.py\", line 861, in _download_url\n    hashes.check_against_chunks(downloaded_chunks)\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_internal\\utils\\hashes.py\", line 75, in check_against_chunks\n    for chunk in chunks:\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_internal\\download.py\", line 829, in written_chunks\n    for chunk in chunks:\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_internal\\utils\\ui.py\", line 156, in iter\n    for x in it:\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_internal\\download.py\", line 818, in resp_read\n    decode_content=False):\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_vendor\\urllib3\\response.py\", line 531, in stream\n    data = self.read(amt=amt, decode_content=decode_content)\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_vendor\\urllib3\\response.py\", line 496, in read\n    raise IncompleteRead(self._fp_bytes_read, self.length_remaining)\n  File \"C:\\ProgramData\\Anaconda3\\lib\\contextlib.py\", line 99, in __exit__\n    self.gen.throw(type, value, traceback)\n  File \"C:\\Users\\samra\\AppData\\Roaming\\Python\\Python36\\site-packages\\pip\\_vendor\\urllib3\\response.py\", line 402, in _error_catcher\n    raise ReadTimeoutError(self._pool, None, 'Read timed out.')\npip._vendor.urllib3.exceptions.ReadTimeoutError: HTTPSConnectionPool(host='files.pythonhosted.org', port=443): Read timed out.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58084102,"AnswerCount":0,"Tags":"<tensorflow><generative-adversarial-network><tpu><gan>","CreationDate":"2019-09-24T15:54:04.553","AcceptedAnswerId":null,"OwnerUserId":4279650.0,"Title":"Preparing data for GAN training on TPU","Body":"<p>I want to use TPUGANEstimator for training a modified version of pix2pix GAN on TPU with additional conditional images (you can find the original publication (non-TPU) here: <a href=\"https:\/\/phillipi.github.io\/pix2pix\/\" rel=\"nofollow noreferrer\">https:\/\/phillipi.github.io\/pix2pix\/<\/a>). To train the generator I need to feed it input_image and condition_image, and for the discriminator I feed the target_image and condition_image. My problem is how the Estimator finds the proper input to the model for the dictionary that I create in input_fn. Here is the psudo-code for my input function code:<\/p>\n\n<pre><code>def input_fn(mode, params):\n    is_train = mode == tf.estimator.ModeKeys.TRAIN\n\n    # Yields a triplet of input_image, condition_image, target_image\n    data_gen = training_generator if is_train else test_generator \n\n    dataset = tf.data.Dataset.from_generator(data_gen,\n        ({'input': tf.float32, 'condition': tf.float32, 'target': tf.float32}),\n        output_shapes=({'input': tf.TensorShape(shape), 'condition': tf.TensorShape(shape),\n                        'target': tf.TensorShape(shape)}))\n    if is_train:\n        dataset = dataset.shuffle(buffer_size=int(config[\"shuffle_buffer_ratio\"] * config[\"batch_size\"]))\n\n    dataset = dataset.prefetch(config[\"max_buffer\"])\n    dataset = dataset.batch(config[\"batch_size\"])\n    # return dataset\n\n    iterator = dataset.make_one_shot_iterator()\n    next_item = iterator.get_next()\n    return next_item\n<\/code><\/pre>\n\n<p>The first issue is to define the generator to use the two inputs from the dictionary:<\/p>\n\n<pre><code>def generator_fn(input_dict, mode='TRAIN', scope='Generator'):\n    input = input_dict[\u2018input\u2019]\n    condition = input_dict[\u2018condition\u2019]\n    with tf.compat.v1.variable_scope(scope, reuse=tf.compat.v1.AUTO_REUSE):\n        x = generator_network(x)\n    return x\n<\/code><\/pre>\n\n<p>And the second problem is how to switch the sequence between fake and valid images for the discriminator (using the joint_train parameter probably):<\/p>\n\n<pre><code>def discriminator_fn(image, input_dict, scope='Discriminator'):\n    x = discriminator_network(x)\nreturn x\n<\/code><\/pre>\n\n<p>I defined the TPUGANEstimator according to the documentations, but I didn\u2019t find a good tutorial\/example on how to generate the correct sequence. <\/p>\n\n<pre><code>gan_estimator = tfgan.estimator.TPUGANEstimator(\n    generator_fn=generator_fn,\n    discriminator_fn=discriminator_fn,\n    generator_loss_fn=tfgan.losses.minimax_generator_loss,\n    discriminator_loss_fn=tfgan.losses.minimax_discriminator_loss,\n    generator_optimizer=tf.compat.v1.train.AdamOptimizer(0.1, 0.5),\n    discriminator_optimizer=tf.compat.v1.train.AdamOptimizer(0.1, 0.5),\n    joint_train=False,\n    gan_train_steps=tfgan_tuples.GANTrainSteps(1, 1),\n    model_dir=config['model_dir'],\n    params=params,\n    use_tpu=False,\n    train_batch_size=2,\n    eval_batch_size=2,\n    config=t_config)\n\nwhile cur_step &lt; number_of_steps:\n    print(\"Running gan estimator: {}\".format(cur_step))\n    gan_estimator.train(train_input_fn, steps=cur_step)\n<\/code><\/pre>\n\n<p>I highly appreciate your kind help in advance.<\/p>\n","answers":[]}
+{"QuestionId":58084325,"AnswerCount":1,"Tags":"<python><deep-learning><pytorch><autograd><autodiff>","CreationDate":"2019-09-24T16:08:36.917","AcceptedAnswerId":null,"OwnerUserId":5293739.0,"Title":"Computing Output Pixel-wise Gradient Norm in PyTorch","Body":"<p>I am looking for an efficient way of computing \\hat{x} of dimensions (b x c x h x w) defined per sample as:<\/p>\n\n<p><img src=\"https:\/\/discuss.pytorch.org\/uploads\/default\/original\/3X\/0\/3\/0326eba6ed3046f60f0765238def1e75c9ac33d9.png\" alt=\"eq\"><\/p>\n\n<p>where x is the output of the same dimensions generated by a model with parameters \\theta, and<\/p>\n\n<p>i,j: index the height and width of the 2D output feature map<\/p>\n\n<p>k: indexes the channel dimension<\/p>\n\n<p>l: indexes the parameters.<\/p>\n\n<p>How do I accomplish this with x.backward()? Is there an efficient way of doing this?<\/p>\n\n<p>If I did x.backward(torch.ones_like(x)), I would compute the sum of the gradients in the above equation in the place of sum of their absolute values.<\/p>\n\n<p>Brute-force way to do this would be to loop over each pixel in the output map and compute .backward(). But, that would be incredibly expensive. Is there an efficient way which avoids the loop?<\/p>\n","answers":[{"AnswerId":"58089717","CreationDate":"2019-09-25T00:40:07.213","ParentId":null,"OwnerUserId":"300908","Title":null,"Body":"<p>After you call <code>.backward()<\/code> on your loss, PyTorch calculates all the gradients for you each of which has a <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.norm\" rel=\"nofollow noreferrer\"><code>.norm()<\/code><\/a> function. Perhaps this is what you're looking for?<\/p>\n\n<p>For example:<\/p>\n\n<pre><code>x = torch.rand((4,3,3), requires_grad=True)   # Some output volume\ny = torch.rand((4,3,3), requires_grad=True)   # Some output volume\nz = x + y\nloss = z.sum()\nloss.backward() \n\nprint(x.grad)          # Print gradients for x\nprint(x.grad.norm(1))  # Print the L1 norm for these gradients\n<\/code><\/pre>\n"}]}
+{"QuestionId":58084466,"AnswerCount":0,"Tags":"<python><performance><tensorflow><keras>","CreationDate":"2019-09-24T16:18:29.257","AcceptedAnswerId":null,"OwnerUserId":12114255.0,"Title":"Performance issues when calling model.detect withMask_RCNN","Body":"<p>I was trying to run a sample python code in Mask_RCNN project.\n<a href=\"https:\/\/github.com\/matterport\/Mask_RCNN\/blob\/master\/samples\/demo.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/matterport\/Mask_RCNN\/blob\/master\/samples\/demo.ipynb<\/a><\/p>\n\n<p>My problem is that the execution of the part of the code below take more than 20 seconds to process an image ( all the cpu cores was running 100% during image detection)<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Run detection\nresults = model.detect([image], verbose=1)\n<\/code><\/pre>\n\n<p>I am running the sample on a MacBook pro CPU i7, 4 cores (no GPU). I checked the versions of tensorflow , opencv ..., and I had the latest versions.<\/p>\n","answers":[]}
+{"QuestionId":58084535,"AnswerCount":1,"Tags":"<python><python-3.x><anaconda><pytorch><conda>","CreationDate":"2019-09-24T16:23:03.017","AcceptedAnswerId":58085296.0,"OwnerUserId":2653567.0,"Title":"Not Able to Import PyTorch in Conda Env","Body":"<p>Last week I had a working conda env I was using for a project. I have not touched the project in a week. I just went to run a python file (<code>python file.py<\/code>) that had been running with no errors. <\/p>\n\n<p>Now I get the following error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Traceback (most recent call last):\n  File \"file.py\", line 2, in &lt;module&gt;\n    from torch.utils.data import Dataset, DataLoader\nModuleNotFoundError: No module named 'torch'\n<\/code><\/pre>\n\n<p>In an attempt to troubleshoot, I opened a python console and ran the following code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>&gt;&gt;&gt; import torch\n<\/code><\/pre>\n\n<p>The result was the following error message:<\/p>\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;\nModuleNotFoundError: No module named 'torch'\n<\/code><\/pre>\n\n<p>If I check all the installed packages using <code>conda list -n &lt;env_name&gt;<\/code>, I can see that PyTorch is in fact installed, just as it was last week. <\/p>\n\n<pre><code>...\npytorch                   1.2.0           py3.7_cuda9.2.148_cudnn7.6.2_0    pytorch\n...\ntorchvision               0.4.0                 py37_cu92    pytorch\n...\n<\/code><\/pre>\n\n<p>Here is what I see when I start python console using <code>python<\/code>:<\/p>\n\n<pre><code>Python 3.7.4 (default, Aug 13 2019, 20:35:49) \n[GCC 7.3.0] :: Anaconda, Inc. on linux\n<\/code><\/pre>\n\n<p>The output of <code>python -c 'import sys; print(sys.path)<\/code> in the base env is:<\/p>\n\n<p><code>['', '\/home\/&lt;name&gt;\/anaconda3\/lib\/python37.zip', '\/home\/&lt;name&gt;\/anaconda3\/lib\/python3.7', '\/home\/&lt;name&gt;\/anaconda3\/lib\/python3.7\/lib-dynload', '\/home\/&lt;name&gt;\/anaconda3\/lib\/python3.7\/site-packages']<\/code> I have not personally made any chages to <code>PYTHONPATH<\/code>.<\/p>\n\n<p>If I run <code>python -c 'import sys; print(sys.path)'<\/code> with my conda env (non-base) active, I get:<\/p>\n\n<p><code>['', '\/home\/&lt;name&gt;\/anaconda3\/envs\/&lt;env_name&gt;\/lib\/python37.zip', '\/home\/&lt;name&gt;\/anaconda3\/envs\/&lt;env_name&gt;\/lib\/python3.7', '\/home\/&lt;name&gt;\/anaconda3\/envs\/&lt;env_name&gt;\/lib\/python3.7\/lib-dynload', '\/home\/&lt;name&gt;\/anaconda3\/envs\/&lt;env_name&gt;\/lib\/python3.7\/site-packages']<\/code><\/p>\n\n<p>This is totally bizarre, can't figure out what is going on, and what could have happened, over the course of the last week, without me touching the code or making any changes to Anaconda.<\/p>\n","answers":[{"AnswerId":"58085296","CreationDate":"2019-09-24T17:16:38.967","ParentId":null,"OwnerUserId":"11383441","Title":null,"Body":"<ol>\n<li><p>open <strong>anaconda-prompt<\/strong> then run this<\/p>\n\n<p><code>conda install PyTorch -c PyTorch<\/code><\/p><\/li>\n<li><p>If you didn't upgrade your pip.use this command to update<\/p>\n\n<p><code>python -m pip install \u2013upgrade pip<\/code><\/p><\/li>\n<li><p>After first step run this<\/p>\n\n<p><code>pip3 install torchvision<\/code><\/p><\/li>\n<\/ol>\n\n<p>hope it will work.<\/p>\n"}]}
+{"QuestionId":58084579,"AnswerCount":0,"Tags":"<tensorflow><keras><object-detection><object-detection-api><resnet>","CreationDate":"2019-09-24T16:26:34.903","AcceptedAnswerId":null,"OwnerUserId":6499890.0,"Title":"Import feature extractor into Tensorflow Object Detection API","Body":"<p>I'm trying to import a <strong>ResNet<\/strong> based architecture as a feature extractor for <strong>SSD<\/strong>.\nAs my application requires real-time, I was trying with different ResNet depths (resnet-10, resnet-16, resnet-24, etc).<\/p>\n\n<p>I've read the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/v1.12.0\/research\/object_detection\/g3doc\/defining_your_own_model.md\" rel=\"nofollow noreferrer\">Defining a new Faster R-CNN or SSD Feature Extractor<\/a> tutorial, and tried to guide myself via <code>object_detection\/models\/ssd_resnet_v1_ppn_feature_extractor<\/code> (and fpn too). <\/p>\n\n<p>The thing is, I'm getting the resnet-10 (for example) architecture from <a href=\"https:\/\/github.com\/osmr\/imgclsmob\/tree\/master\/tensorflow_\" rel=\"nofollow noreferrer\">tensorflowcv<\/a> python package, which is implemented via Keras, and I'm defining the function <code>resnet10_base<\/code> as a redefinition of <a href=\"https:\/\/github.com\/osmr\/imgclsmob\/blob\/master\/tensorflow_\/tensorflowcv\/models\/resnet.py\" rel=\"nofollow noreferrer\">get_resnet<\/a> as seen below so that it ignores the last layers that make it a classifier.<\/p>\n\n<p><strong>My redefinition of <code>get_resnet<\/code>:<\/strong><\/p>\n\n<pre><code> 01  def resnet10_base(model_name='resnet10', root=os.path.join(\"~\", \".tensorflow\", \"    models\")):\n 02   layers = [1,1,1,1]\n 03   assert (sum(layers) * 2 + 2 == blocks)\n 04 \n 05   init_block_channels = 64\n 06   channels_per_layers = [64, 128, 256, 512]\n 07 \n 08   channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)]\n 09 \n 10   if pretrained:\n 11     if (model_name is None) or (not model_name):\n 12       raise ValueError(\"Parameter `model_name` should be properly initialized fo    r loading pretrained model.\")\n 13     from .model_store import download_state_dict\n 14     net.state_dict, net.file_path = download_state_dict(\n 15       model_name=model_name,\n 16       local_model_store_dir_path=root)\n 17 \n 18   return net\n<\/code><\/pre>\n\n<p>And I can't understand how to create the <code>extract_features()<\/code> function that <code>SSDKerasFeatureExtractor<\/code> asks for. For example, do I have to create the TFSession and run the net in there? Or do I have to return a tensor? How do I achieve that?<\/p>\n\n<p>Thanks in advance, sorry for the inconvenience.<\/p>\n","answers":[]}
+{"QuestionId":58084661,"AnswerCount":0,"Tags":"<python><nlp><pytorch><spacy><spacy-pytorch-transformers>","CreationDate":"2019-09-24T16:32:50.127","AcceptedAnswerId":null,"OwnerUserId":3607203.0,"Title":"How are token vectors calculated in spacy-pytorch-transformers","Body":"<p>I am currently working with the <a href=\"https:\/\/github.com\/explosion\/spacy-pytorch-transformers\" rel=\"nofollow noreferrer\"><code>spacy-pytorch-transformer<\/code><\/a> package to experiment with the respective embeddings.<br\/>\nWhen reading the introductionary article (essentially the GitHub README), my understanding was that the token-level embeddings are the mean over the embeddings of all corresponding word pieces, i.e. <code>embed(complex)<\/code> would be the same as <code>1\/2 * embed(comp#) * embed(#lex)<\/code>.<\/p>\n\n<p>According to the BERT paper, this should simply utilize the <code>last_hidden_state<\/code> property of the network, but my MCVE below shows that this is not the same for Spacy 2.1.8 and spacy-pytorch-transformers 0.4.0, for at least BERT and RoBERTa (have not verified it for more models):<\/p>\n\n<pre><code>import spacy\nimport numpy as np\nnlp = spacy.load(\"en_pytt_robertabase_lg\")  # either this or the BERT model\ntest = \"This is a test\"  # Note that all tokens are directly aligned, so no mean has to be calculated.\ndoc = nlp(test)\n# doc[0].vector and doc.tensor[0] are equal, so the results are equivalent.\nprint(np.allclose(doc[0].vector, doc._.pytt_last_hidden_state[1, :]))\n# returns False\n<\/code><\/pre>\n\n<p>The offset of 1 for the hidden states is due to the <code>&lt;CLS&gt;<\/code> token as the first input, which corresponds to the sentence classification task; I even checked with any available other token for my sentence (which has no token alignment problems according to <code>doc._.pytt_alignment<\/code>), so there is no way I missed something here.<\/p>\n\n<p><a href=\"https:\/\/github.com\/explosion\/spacy-pytorch-transformers\/blob\/master\/spacy_pytorch_transformers\/pipeline\/tok2vec.py\" rel=\"nofollow noreferrer\">According to the source code<\/a>, the corresponding hook overwrites simply to return the corresponding row in the tensor, so I do not see any transformation here. Is there something obvious that I am missing here, or is this deviating from the expected behavior?<\/p>\n","answers":[]}
+{"QuestionId":58085131,"AnswerCount":1,"Tags":"<python><nlp><pytorch><reasoning>","CreationDate":"2019-09-24T17:05:53.330","AcceptedAnswerId":null,"OwnerUserId":12114507.0,"Title":"how to deal with uncertain amount of sentences in different instances within a batch?","Body":"<p>I encountered a coding problem. In my dataset, an instance includes several sentences (different amounts in different instances). They can not be concatenated to serve as a single one. How can I effectively process this kind of data with PyTorch? Or I have to process instance one by one?<\/p>\n","answers":[{"AnswerId":"58090524","CreationDate":"2019-09-25T02:52:07.813","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p>This is a very broad question. However, I can think of two less complex solutions.<\/p>\n\n<ol>\n<li>Use a dummy sentence to pad the instances and mask the dummy sentences while learning representations for the instance.<\/li>\n<li>You can group the instances based on the number of sentences to create mini-batches to avoid padding. However, if this is not a case, at least try to group instances which are similar in the number of sentences to minimize the amount of padding.<\/li>\n<\/ol>\n\n<p>You can study existing implementations of learning document representations, for example, <a href=\"https:\/\/github.com\/cedias\/Hierarchical-Sentiment\" rel=\"nofollow noreferrer\">Hierarchical Attention Networks for Document Classification paper<\/a>.<\/p>\n"}]}
+{"QuestionId":58085460,"AnswerCount":0,"Tags":"<tensorflow><distributed-computing>","CreationDate":"2019-09-24T17:28:44.410","AcceptedAnswerId":null,"OwnerUserId":6450489.0,"Title":"Distributed training on different computer configurations in Tensorflow","Body":"<p>If I have one PC with 2 RTX 2080 Ti GPU and one PC with 1 GTX 1080 Ti GPU, what is the best approach to use them with maximum utilization for doing a distributed learning of a keras model in tensorflow? I am inclined to encapsulate keras model in an estimator class but then I don't know how to define the above GPU configuration in TF_CONFIG dictionary.<\/p>\n","answers":[]}
+{"QuestionId":58085722,"AnswerCount":2,"Tags":"<machine-learning><keras><deep-learning><computer-vision><rgb>","CreationDate":"2019-09-24T17:48:21.767","AcceptedAnswerId":null,"OwnerUserId":11599414.0,"Title":"Are RGB-images in Keras represented as [R,G,B], [R,G,B], ... or as [RR.., GG.., BB..]?","Body":"<p>Say we have a 3x3 RGB image.<\/p>\n\n<p>Which of the following representations is correct? <\/p>\n\n<pre><code>[ R, G, B],               [ R, R, R],\n[ R, G, B],     OR        [ G, G, G], \n[ R, G, B]                [ B, B, B]\n<\/code><\/pre>\n\n<p>I purposely left out the outer parenthesis as how they are set is part of the anwer to my question<\/p>\n","answers":[{"AnswerId":"58085967","CreationDate":"2019-09-24T18:07:04.120","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>In Keras (and most other places), \"R\", \"G\", and \"B\" are referred to as \"channels\". By default, the channels are the last dimension of the tensor. This means that each image is an array of rows, and each row is an array of pixels, and each pixel is an array of 3 colors\/channels. To answer your exact question, that is most closely represented by your first example rather than your second.<\/p>\n\n<p>However, Keras can also be configured to use <code>channels_first<\/code>, which means that it's the opposite. In this case, an image is an array of 3 channels, and each channel is a monochrome 2D array of pixels.<\/p>\n"},{"AnswerId":"58086077","CreationDate":"2019-09-24T18:15:45.543","ParentId":null,"OwnerUserId":"12067496","Title":null,"Body":"<p>From my point of view you are asking unclear question:<\/p>\n\n<ul>\n<li><strong>3x3 RGB image is usually represented as a 3x3x3 array<\/strong><\/li>\n<\/ul>\n\n<p>that means if your question is about a 3x3 RGB image it should contain three 3x3 array each containing values of one color which can be visualized as:<\/p>\n\n<pre><code>{[(R, R, R),   [(G, G, G),   [(B, B, B),\n  (R, R, R),    (G, G, G),    (B, B, B),\n  (R, R, R)],   (G, G, G)],   (B, B, B)]}\n<\/code><\/pre>\n\n<ul>\n<li><strong>Your question actually visualises 3x1 RGB image<\/strong><\/li>\n<\/ul>\n\n<p>in which case (3x1 RGB image) this version in your question should be correct (maybe after transposition):<\/p>\n\n<pre><code>{[ R, R, R],\n [ G, G, G], \n [ B, B, B]}\n<\/code><\/pre>\n\n<p>If I am wrong, please feel free to edit or correct my answer!<\/p>\n"}]}
+{"QuestionId":58085880,"AnswerCount":0,"Tags":"<tensorflow><tf.keras>","CreationDate":"2019-09-24T18:01:07.810","AcceptedAnswerId":null,"OwnerUserId":7765850.0,"Title":"Loading file from path contained in tf.Tensor","Body":"<p>I build a simple model using <code>tf.keras<\/code> and a <code>tf.data.Dataset<\/code> for efficient loading as the dataset is a couple of GBs big.<\/p>\n\n<p>The images are in <code>tiff<\/code> format and therefore need to be loaded directly as <code>numpy.array<\/code>.<\/p>\n\n<p>I do have a dataset of labels and file paths and want to map a function for loading on that dataset. Therefore, I somehow have to get the python string representation out of the tensor.  <\/p>\n\n<p>I tried using the usual <code>tf.Tensor.eval()<\/code> and then joining the chars to a full string but getting the err: <code>ValueError: Cannot evaluate tensor using<\/code>eval()<code>: No default session is registered.<\/code> which does make sense as there is no session before the keras model is being executed <\/p>\n\n<p>Then I tried putting <code>tf.enable_eager_execution()<\/code> right below my tensorflow import in the dataset file (and changing to from <code>.eval()<\/code> to <code>.numpy()<\/code>) but am getting err: <code>AttributeError: 'Tensor' object has not attribute 'numpy'<\/code> hinting <code>tf.enable_eager_execution()<\/code> did not work<\/p>\n\n<p>Basically I'm trying to read a string contained in a tensor a below:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>path = tf.decode_raw(path, tf.uint8)\npath = ''.join(map(chr, path.eval()))  # with session\npath = ''.join(map(chr, path.numpy()))  # with eager execution\n\nimage = PIL.Image.open(path)\nimage = numpy.array(image)\n<\/code><\/pre>\n\n<p>Both work fine when 'prototyping' in a single file without the rest but do not work when having e.g. my model in a <code>model.py<\/code> and dataset in <code>dataset.py<\/code> but having <code>tf.enable_eager_execution()<\/code> in both<\/p>\n","answers":[]}
+{"QuestionId":58086325,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-24T18:34:44.687","AcceptedAnswerId":58086557.0,"OwnerUserId":8651466.0,"Title":"Tensor flow reshape usage","Body":"<p>I saw in an example the following code :<\/p>\n\n<pre><code>X = tf.placeholder(tf.float32, [None, 784], name='InputData') \n# image shape 28*28=784\nXX = tf.reshape(X, [-1, 784]) \n# reshape input\n\n<\/code><\/pre>\n\n<p>What is the usage of tf.reshape here as we already have input data set coming from mnist data set and each image has a size of 784 (28*28).  So what are we gaining by reshaping the image again ?<\/p>\n","answers":[{"AnswerId":"58086557","CreationDate":"2019-09-24T18:52:59.233","ParentId":null,"OwnerUserId":"6866309","Title":null,"Body":"<blockquote>\n  <p>If one component of shape is the special value -1, the size of that\n  dimension is computed so that the total size remains constant<\/p>\n<\/blockquote>\n\n<p>In your example, 784 in both places may confuse you.<\/p>\n\n<p>But if you want a 28 * 28 dimensions, you many need to reshape as shown below. Then the <strong>-1<\/strong> is inferred to <strong>28<\/strong>.<\/p>\n\n<pre><code>X = tf.placeholder(tf.float32, [None, 784], name='InputData') \nXX = tf.reshape(X, [-1, 28]) \n<\/code><\/pre>\n"}]}
+{"QuestionId":58086601,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-24T18:56:06.593","AcceptedAnswerId":null,"OwnerUserId":9734248.0,"Title":"Xavier initialization in Tensorflow 2.0","Body":"<p>I have built a simple ANN model:<\/p>\n\n<pre><code>#ANN:\nmodel = tf.keras.models.Sequential([\n  tf.keras.layers.Dense(units = 128,activation = 'relu',input_shape = (784,)),#input layer\n  tf.keras.layers.BatchNormalization(),#batch normalization\n  tf.keras.layers.Dropout(0.2), #dropout technique\n\n  tf.keras.layers.Dense(units = 64,activation = 'relu'), #second fully connected layer\n  tf.keras.layers.BatchNormalization(),#batch normalization\n  tf.keras.layers.Dropout(0.2),#dropout technique\n\n  tf.keras.layers.Dense(units = 10,activation = 'softmax') #output layer\n    ])\n\nmodel.compile(optimizer = tf.train.AdamOptimizer(0.001),\n             loss='sparse_categorical_crossentropy',\n              metrics=['sparse_categorical_accuracy']) #compiling the model\n\n\nmodel.fit(X_train,Y_train,batch_size = 64,epochs = 100)\n<\/code><\/pre>\n\n<p>But I want to use Xavier initialization for weights but I didn't find how to do it in tensorflow 2.0<\/p>\n","answers":[{"AnswerId":"58087880","CreationDate":"2019-09-24T20:44:53.470","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>In tf2.0, the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/75a91fac16493b6f68c0740b214ff18f5d88e5a8\/tensorflow\/python\/ops\/init_ops_v2.py#L557\" rel=\"nofollow noreferrer\">GlorotUniform<\/a> \/ <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/75a91fac16493b6f68c0740b214ff18f5d88e5a8\/tensorflow\/python\/ops\/init_ops_v2.py#L586\" rel=\"nofollow noreferrer\">GlorotNormal<\/a> initializer , are also called Xavier uniform\/normal initializer. So you can define a layer like this,<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.add(Dense(64,kernel_initializer='glorot_uniform',bias_initializer='zeros'))\n<\/code><\/pre>\n"}]}
+{"QuestionId":58087084,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><edward>","CreationDate":"2019-09-24T19:33:32.120","AcceptedAnswerId":58095543.0,"OwnerUserId":10614851.0,"Title":"Edward2 implementation of Edward's HMC method","Body":"<p>Edward1 contained a method HMC with with the following syntax:<\/p>\n\n<pre><code>inference = ed.HMC({z: qz})\ninference.run()\n<\/code><\/pre>\n\n<p>I'm having difficulty tracking down the migration for this. Neither Edward2 nor TPF contain this package. When looking up the migration instruction on the Edward project's page, the documentation 404s out. The closest implementation in TFP that I have found is:<\/p>\n\n<pre><code>ed2.mcmc.HamiltonianMonteCarlo()\n<\/code><\/pre>\n\n<p>However this take in params that are significantly different from the Edward1 package. The code example I am working off is <a href=\"https:\/\/github.com\/blei-lab\/edward\/blob\/master\/examples\/normal.py\" rel=\"nofollow noreferrer\">linked<\/a>. There is no updated version of this available in the TFP example repos. <\/p>\n\n<p>Can anyone explain what the implementation of this method is in Edward2? Has it been entirely replaced\/no longer exist?<\/p>\n","answers":[{"AnswerId":"58095543","CreationDate":"2019-09-25T09:42:20.537","ParentId":null,"OwnerUserId":"10675934","Title":null,"Body":"<p>According to the docs here - \n<a href=\"https:\/\/github.com\/google\/edward2\/blob\/master\/Upgrading_From_Edward_To_Edward2.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/google\/edward2\/blob\/master\/Upgrading_From_Edward_To_Edward2.md<\/a><\/p>\n\n<p>the HMC available through tensorflow-probability replaces the edward implementation. Details here:\n<a href=\"https:\/\/www.tensorflow.org\/probability\/api_docs\/python\/tfp\/mcmc\/HamiltonianMonteCarlo\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/probability\/api_docs\/python\/tfp\/mcmc\/HamiltonianMonteCarlo<\/a><\/p>\n"}]}
+{"QuestionId":58087732,"AnswerCount":2,"Tags":"<python><tensorflow><anaconda>","CreationDate":"2019-09-24T20:32:28.580","AcceptedAnswerId":null,"OwnerUserId":9914673.0,"Title":"python2.7 pip install tensorflow when installed Anaconda with python3.5","Body":"<p>I had a large python 2.7 project which require tensorflow.  But I can not pip install it when I have  installed Anaconda with python3.5.<\/p>\n\n<pre><code>pip install tensorflow\n<\/code><\/pre>\n\n<p>Given error:<\/p>\n\n<p>Requirement already satisfied: tensorflow in c:\\users\\victo\\anaconda3\\lib\\site-packages (1.14.0)\ns\\victo\\anaconda3\\lib\\site-packages (from tensorboard&lt;1.15.0,>=1.14.0->tensorflow) (3.1.1)\nCollecting setuptools>=41.0.0 (from tensorboard&lt;1.15.0,>=1.14.0->tensorflow)\n..\n  Using cached <a href=\"https:\/\/files.pythonhosted.org\/packages\/b2\/86\/095d2f7829badc207c893dd4ac767e871f6cd547145df797ea26baea4e2e\/setuptools-41.2.0-py2.py3-none-any.whl\" rel=\"nofollow noreferrer\">https:\/\/files.pythonhosted.org\/packages\/b2\/86\/095d2f7829badc207c893dd4ac767e871f6cd547145df797ea26baea4e2e\/setuptools-41.2.0-py2.py3-none-any.whl<\/a><br>\n  Found existing installation: setuptools 27.2.0\nERROR: Cannot remove entries from nonexistent file c:\\users\\victo\\anaconda3\\lib\\site-packages\\easy-install.pth<\/p>\n","answers":[{"AnswerId":"58088066","CreationDate":"2019-09-24T20:59:58.543","ParentId":null,"OwnerUserId":"12115549","Title":null,"Body":"<p>I used to have both installations. Just install it in another folder, and run the anaconda version you need according to the project. <\/p>\n"},{"AnswerId":"58096723","CreationDate":"2019-09-25T10:50:13.567","ParentId":null,"OwnerUserId":"5101148","Title":null,"Body":"<p>You should create different environments for different projects, which is a very basic function included in <code>conda<\/code>. <code>conda<\/code> is both a package manager and a environment manager.<\/p>\n\n<p><a href=\"https:\/\/packaging.python.org\/guides\/installing-using-pip-and-virtual-environments\/#creating-a-virtual-environment\" rel=\"nofollow noreferrer\">What's a virtual environment<\/a><\/p>\n\n<blockquote>\n  <p>venv (for Python 3) and virtualenv (for Python 2) allow you to <strong>manage separate package installations for different projects<\/strong>. They essentially allow you to create a <strong>\u201cvirtual\u201d isolated Python installation<\/strong> and install packages into that virtual installation. When you switch projects, you can simply create a new virtual environment and not have to worry about breaking the packages installed in the other environments. It is always recommended to use a virtual environment while developing Python applications.<\/p>\n<\/blockquote>\n\n<p>Examples<\/p>\n\n<pre class=\"lang-sh prettyprint-override\"><code># create an virtual environment named py2 with Python 2.7 installed\nconda create -n py2 python=2.7\n\n# another env named py3 with Python 3.7\nconda create -n py3 python=3.7\n\n# activate an env and install package into this very env\nconda activate py3\n\n# install packages with pip, or conda\npip install tensorflow\n# or\nconda install tensorflow\n\n# exit the env\nconda deactivate\n<\/code><\/pre>\n\n<p><strong>Tips<\/strong>: After an env is activated, you can use <code>pip<\/code> or <code>conda<\/code> to install packages into it. But never use both of them within a virtual environment.<\/p>\n\n<p>Extended Reading<\/p>\n\n<ul>\n<li><a href=\"https:\/\/conda.io\/projects\/conda\/en\/latest\/user-guide\/getting-started.html\" rel=\"nofollow noreferrer\">get started with <code>conda<\/code><\/a><\/li>\n<li><a href=\"https:\/\/www.anaconda.com\/using-pip-in-a-conda-environment\/\" rel=\"nofollow noreferrer\">Using Pip in a Conda Environment<\/a><\/li>\n<\/ul>\n"}]}
+{"QuestionId":58087923,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-24T20:48:22.303","AcceptedAnswerId":null,"OwnerUserId":11836594.0,"Title":"Keras Learning Won't Start Training","Body":"<p>I am trying to train a model to recognize different cloud types and select the area where those clouds are located. To do this, I am sending an image file(350*550, grayscale) to a model, which outputs which pixels are to be selected(350*550). I have attached a link to my code below. This is part of a competition with Kaggle (<a href=\"https:\/\/www.kaggle.com\/c\/understanding_cloud_organization\/data\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/c\/understanding_cloud_organization\/data<\/a>), and I am not looking for someone to rewrite my entire code. <\/p>\n\n<p>When trying to train my data, the data is loaded and then sent to the model to be fit. Then I get a bunch of lines from tensorflows bfc allocation, which quits after a few minutes with no obvious error(for me, at least). I'm really at a loss on what to do.<\/p>\n\n<p>Code: <a href=\"https:\/\/github.com\/abritton99999999\/KaggleCode\" rel=\"nofollow noreferrer\">https:\/\/github.com\/abritton99999999\/KaggleCode<\/a>\nConsole output is posted in the github as well now!<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/q8Kj4.png\" rel=\"nofollow noreferrer\">Some error code<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/MRQjC.png\" rel=\"nofollow noreferrer\">More errors<\/a><\/p>\n","answers":[{"AnswerId":"58098614","CreationDate":"2019-09-25T12:39:22.103","ParentId":null,"OwnerUserId":"10675934","Title":null,"Body":"<p>You're trying to create a very large tensor (50000x192500 floats) and are exceeding memory limits. Some suggestions for workarounds are here\n<a href=\"https:\/\/stackoverflow.com\/questions\/46941747\/tensorflow-error-cannot-parse-tensor-from-proto\">Tensorflow Error: &quot;Cannot parse tensor from proto&quot;<\/a><\/p>\n\n<p>Essentially you're not going to be able to create a network like this<\/p>\n\n<pre><code>model.add(Dense(100000, input_shape=(192500,), activation=\"sigmoid\"))\nmodel.add(Dense(50000, activation=\"sigmoid\"))\nmodel.add(Dense(192500, activation=\"softmax\"))\n<\/code><\/pre>\n\n<p>Instead you'll need to use convolutional layers<\/p>\n"}]}
+{"QuestionId":58088208,"AnswerCount":2,"Tags":"<docker><opencv><tensorflow><jupyter-notebook>","CreationDate":"2019-09-24T21:12:41.803","AcceptedAnswerId":null,"OwnerUserId":10402016.0,"Title":"Dockerfile for tensorflow GPU + opencv3 + Jupyter Notebook?","Body":"<p>I am a bit confused. Now I am not using a dockerfile but the command:<\/p>\n\n<pre><code>docker run -it --rm -v $(realpath ~\/path\/of\/directory):\/tf\/notebooks -p 8888:8888 tensorflow\/tensorflow:latest-py3-jupyter\n<\/code><\/pre>\n\n<p>But I don't have access to cv2 module. Maybe add something to the run command or write my own dockerfile? but I don't know how. Do I have to pip install and RUN tensorflow GPU and jupyter notebook in the dockerfile ?<\/p>\n","answers":[{"AnswerId":"58516638","CreationDate":"2019-10-23T06:23:51.283","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>Tensorflow image by default does not have OpenCV in it. <\/p>\n\n<p>There are two ways you can set it up : <\/p>\n\n<p>A - Setup tensorflow docker container and then install using commands<\/p>\n\n<p>OR<\/p>\n\n<p>B - Create a docker file which takes tensorflow image as base image and install opencv using commands while building the image. <\/p>\n\n<p>Please refer to the below link to do the same. <\/p>\n\n<p><a href=\"https:\/\/github.com\/fbcotter\/docker-tensorflow-opencv\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fbcotter\/docker-tensorflow-opencv<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/fbcotter\/docker-tensorflow-opencv\/blob\/master\/Dockerfile\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fbcotter\/docker-tensorflow-opencv\/blob\/master\/Dockerfile<\/a><\/p>\n"},{"AnswerId":"58105677","CreationDate":"2019-09-25T19:52:28.867","ParentId":null,"OwnerUserId":"4948432","Title":null,"Body":"<p>Check if this image <em>tensorflow\/tensorflow:latest-py3-jupyter<\/em> has installed opencv and python module cv2.\nIf it doesn't you should find another image or install it by yourself.<\/p>\n"}]}
+{"QuestionId":58088290,"AnswerCount":0,"Tags":"<keras><neural-network>","CreationDate":"2019-09-24T21:19:55.700","AcceptedAnswerId":null,"OwnerUserId":11892298.0,"Title":"Error creating h5 file using Keras \"model.save()\"","Body":"<p>I'm currently training a network with keras, which is working completely fine. The issue is once I try to save the model using <code>model.save('model.h5')<\/code>, I get the following error:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Olg0t.png\" rel=\"nofollow noreferrer\">https:\/\/i.stack.imgur.com\/Olg0t.png<\/a><\/p>\n\n<p>This is not happening with identical code on my original computer. It also seems that my user account has full access to the directory. I have tried using absolute path but that also did not work.<\/p>\n","answers":[]}
+{"QuestionId":58088301,"AnswerCount":1,"Tags":"<python><keras><neural-network><lstm>","CreationDate":"2019-09-24T21:21:16.920","AcceptedAnswerId":null,"OwnerUserId":1493120.0,"Title":"Reshape input data for LSTM autoencoder","Body":"<p>I used to use <code>model.add(Reshape())<\/code> from Keras but for some reason the same code that worked before does not work now. I am trying to reshape the input for LSTM without reshaping the data prior; without using <code>np.reshape<\/code> command.\nMy code is:<\/p>\n\n<pre><code>def model_lstm_ae ():\n    model=Sequential()\n    model.add(Reshape((1,X_train_.shape[1])))\n    model.add(LSTM(128, activation='relu', \n                   input_shape=(1,X_train_.shape[1]), return_sequences=True))\n\n    model.add(LSTM(64, activation='relu', \n                   return_sequences=False))\n\n    model.add(RepeatVector(timesteps))\n\n    model.add(LSTM(64, activation='relu', \n                   return_sequences=True))\n\n    model.add(LSTM(128, activation='relu', \n                   return_sequences=False))\n\n    model.add(TimeDistributed(Dense(1)))\n\n    model.compile(loss='mean_squared_error', \n                  optimizer=k.optimizers.Adam(lr=1e-4, amsgrad=True))\n\n    return model\n\nmodel_lstm_ae = KerasRegressor(build_fn=model_lstm_ae)\nhistory=model_lstm_ae.fit(X_train_, X_train_, \n                          epochs=100, \n                          batch_size=256, \n                          validation_data = (X_test_, X_test_),\n                          verbose = 1,\n                          callbacks = [early_stopping_monitor, lr_reduce])\n<\/code><\/pre>\n\n<p>However, I get the following error:<\/p>\n\n<pre><code>ValueError: The first layer in a Sequential model must get an `input_shape` or `batch_input_shape` argument.\n<\/code><\/pre>\n\n<p>Any suggestions on how to make it work?<\/p>\n","answers":[{"AnswerId":"58107148","CreationDate":"2019-09-25T22:06:36.730","ParentId":null,"OwnerUserId":"12121788","Title":null,"Body":"<p>I run your model (inclduing the KerasRegressor and Reshape) with TF2 RC2 on a linux machine. To be sure, please provide some data for your variables. I suppose you just need to adapt your imports.<\/p>\n"}]}
+{"QuestionId":58088560,"AnswerCount":0,"Tags":"<keras><conv-neural-network><convolution><autoencoder>","CreationDate":"2019-09-24T21:48:21.217","AcceptedAnswerId":null,"OwnerUserId":5425165.0,"Title":"High loss from convolutional autoencoder keras","Body":"<p>I am training a convolutional autoencoder and I am having trouble getting the loss to decrease and was hoping someone could point out some possible improvements.<\/p>\n\n<p>I have greyscale images of 1024x1024 (I've tried the same thing with 512x512 too) and I'm hoping to compress them for unsupervised clustering. My full model is below but it follows a pretty basic pattern of several Conv2D layers w\/ max pooling, followed by a dense layer, then reshape and Conv2D layers back up to the original image size. <\/p>\n\n<p>Some things I've tried so far:<\/p>\n\n<p>1) I found that mse worked better as a loss function than binary cross entropy, because the pixel luminosity values are so far from evenly distributed (binary cross entropy gets stuck assigning everything to a value of 1, which has low error but isn't useful).<\/p>\n\n<p>2) I can easily achieve very low error and near perfect (to my eye at least) image reconstruction if I just get rid of the dense layer in the middle and compress the image slightly less. This is fairly obvious, but I guess it shows that I'm not making some kind of error resulting in nonsense outputs.<\/p>\n\n<p>3) My loss doesn't really get below 0.02-0.03. Still, at 0.025 or so the image is reconstructed enough that it's obvious that the output is coming from the input, not some kind of random noise (like making every pixel the same intensity or something). I think getting it below 0.01 would be enough that I can cluster on. My lowest (albeit on a slightly easier subset of my data) was 0.018 and I could see clear clustering in samples when I plotted my encoded values in a heatmap.<\/p>\n\n<p>4) I get a lot of dying ReLUs when my middle dense layer uses a ReLU activation, which makes it less useful for the eventual clustering. I use tanh instead. I also found that \"he_normal\" works better as an initialization for the dense layer.<\/p>\n\n<p>5) Adding more dense layers in the middle doesn't seem to help at all.<\/p>\n\n<p>6) Reversing the shape of the encoder (so that it goes from fewer kernels per layer to more kernels) doesn't help either, even though I know that's traditionally how a convolutional autoencoder looks.<\/p>\n\n<p>Here is the full model (the output from model.summary()<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_4 (InputLayer)         (None, 1024, 1024, 1)     0       \n_________________________________________________________________\nconv2d_40 (Conv2D)           (None, 1024, 1024, 128)   1280      \n_________________________________________________________________\nmax_pooling2d_19 (MaxPooling (None, 512, 512, 128)     0         \n_________________________________________________________________\nbatch_normalization_4 (Batch (None, 512, 512, 128)     512       \n_________________________________________________________________\nconv2d_41 (Conv2D)           (None, 512, 512, 64)      73792\n_________________________________________________________________\nmax_pooling2d_20 (MaxPooling (None, 256, 256, 64)      0         \n_________________________________________________________________\nconv2d_42 (Conv2D)           (None, 256, 256, 32)      18464\n_________________________________________________________________\nmax_pooling2d_21 (MaxPooling (None, 128, 128, 32)      0         \n_________________________________________________________________\nconv2d_43 (Conv2D)           (None, 128, 128, 16)      4624      \n_________________________________________________________________\nmax_pooling2d_22 (MaxPooling (None, 64, 64, 16)        0         \n_________________________________________________________________\nconv2d_44 (Conv2D)           (None, 64, 64, 8)         1160      \n_________________________________________________________________\nmax_pooling2d_23 (MaxPooling (None, 32, 32, 8)         0         \n_________________________________________________________________\nflatten_4 (Flatten)          (None, 8192)              0         \n_________________________________________________________________\ndense_5 (Dense)              (None, 512)               4194816   \n_________________________________________________________________\nreshape_4 (Reshape)          (None, 8, 8, 8)           0         \n_________________________________________________________________\nup_sampling2d_22 (UpSampling (None, 16, 16, 8)         0         \n_________________________________________________________________\nconv2d_45 (Conv2D)           (None, 16, 16, 16)        1168      \n_________________________________________________________________\nup_sampling2d_23 (UpSampling (None, 32, 32, 16)        0         \n_________________________________________________________________\nconv2d_46 (Conv2D)           (None, 32, 32, 16)        2320      \n_________________________________________________________________\nup_sampling2d_24 (UpSampling (None, 64, 64, 16)        0         \n_________________________________________________________________\nconv2d_47 (Conv2D)           (None, 64, 64, 32)        4640      \n_________________________________________________________________\nup_sampling2d_25 (UpSampling (None, 128, 128, 32)      0         \n_________________________________________________________________\nconv2d_48 (Conv2D)           (None, 128, 128, 64)      18496\n_________________________________________________________________\nup_sampling2d_26 (UpSampling (None, 256, 256, 64)      0         \n_________________________________________________________________\nconv2d_49 (Conv2D)           (None, 256, 256, 128)     73856     \n_________________________________________________________________\nup_sampling2d_27 (UpSampling (None, 512, 512, 128)     0         \n_________________________________________________________________\nconv2d_50 (Conv2D)           (None, 512, 512, 128)     147584\n_________________________________________________________________\nup_sampling2d_28 (UpSampling (None, 1024, 1024, 128)   0\n_________________________________________________________________\nconv2d_51 (Conv2D)           (None, 1024, 1024, 1)     1153\n=================================================================\nTotal params: 4,543,865\nTrainable params: 4,543,609\nNon-trainable params: 256\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58088837,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><keras><neural-network>","CreationDate":"2019-09-24T22:20:58.460","AcceptedAnswerId":58089508.0,"OwnerUserId":5601592.0,"Title":"Keras with activity_regularizer that is updated every iteration","Body":"<p>I am building a simple neural network using Keras. It has activity regularization so that the output of the only hidden layer is forced to have small values. Here is the code:<\/p>\n\n<pre><code>import numpy as np\nimport math\nimport keras\nfrom keras.models import Model, Sequential\nfrom keras.layers import Input, Dense, Activation\nfrom keras import regularizers\nfrom keras import backend as K\n\na=1\n\ndef my_regularizer(inputs):\n     means=K.mean((inputs),axis=1)\n     return a*K.sum(means)**2\n\nx_train=np.random.uniform(low=-1,high=1,size=(200,2))\n\nmodel=Sequential([\n     Dense(20,input_shape=(2,),activity_regularizer=my_regularizer),\n     Activation('tanh'),\n     Dense(2,),\n     Activation('linear')\n])\n\nmodel.compile(optimizer='adam',loss='mean_squared_error')\nmodel.fit(x_train,x_train,epochs=20,validation_split=0.1)\n<\/code><\/pre>\n\n<p><strong>Questions:<\/strong><\/p>\n\n<p>1) Currently, parameter <em>a<\/em> is set at the beginning and it does not change. How can I change the code such that the parameter <em>a<\/em> is updated after each iteration such that<\/p>\n\n<p><em>a_new=f(a_old,input)<\/em><\/p>\n\n<p>where input is the values at the hidden layer and <em>f(.)<\/em> is an arbitrary function.<\/p>\n\n<p>2) I want my activity regularizer to be applied after the first activation function tanh is applied. Have I written my code correctly? The term \"activity_regularizer=my_regularizer\" in<\/p>\n\n<pre><code>Dense(20,input_sahpe=(2,),activity_regularizer=my_regularizer)\n<\/code><\/pre>\n\n<p>makes me feel that the regularizer is being applied to values before the activation function tanh.<\/p>\n","answers":[{"AnswerId":"58089508","CreationDate":"2019-09-25T00:03:27.287","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>You can - but first, you need a valid Keras <code>Regularizer<\/code> object (your function won't work):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class MyActivityRegularizer(Regularizer):\n    def __init__(self, a=1):\n        self.a = K.variable(a, name='a')\n\n    # gets called at each train iteration\n    def __call__(self, x): # your custom function here\n        means = K.mean(x, axis=1)\n        return self.a * K.sum(means)**2\n\n    def get_config(self): # required class method\n        return {\"a\": float(K.get_value(self.a))}\n<\/code><\/pre>\n\n<p>Next, to work with <code>.fit<\/code>, you need a custom Keras <code>Callback<\/code> object (see alternative at bottom):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class ActivityRegularizerScheduler(Callback):\n    \"\"\" 'on_batch_end' gets automatically called by .fit when finishing\n    iterating over a batch. The model, and its attributes, are inherited by \n    'Callback' (except at __init__) and can be accessed via, e.g., self.model \"\"\"\n\n    def __init__(self, model, update_fn):\n        self.update_fn=update_fn\n        self.activity_regularizers=_get_activity_regularizers(model)\n\n    def on_batch_end(self, batch, logs=None):\n        iteration = K.get_value(self.model.optimizer.iterations)\n        new_activity_reg = self.update_fn(iteration) \n\n        # 'activity_regularizer' references model layer's activity_regularizer (in this \n        # case 'MyActivityRegularizer'), so its attributes ('a') can be set directly\n        for activity_regularizer in self.activity_regularizers:\n            K.set_value(activity_regularizer.a, new_activity_reg)\n\ndef _get_activity_regularizers(model):\n    activity_regularizers = []\n    for layer in model.layers:\n        a_reg = getattr(layer,'activity_regularizer',None)\n        if a_reg is not None:\n            activity_regularizers.append(a_reg)\n    return activity_regularizers\n<\/code><\/pre>\n\n<p>Lastly, you'll need to create your model within the Keras <code>CustomObjectScope<\/code> - see in full ex. below.\n<hr>\n<strong>Example usage<\/strong>:<\/p>\n\n<pre><code>from keras.layers import Dense\nfrom keras.models import Sequential\nfrom keras.regularizers import Regularizer\nfrom keras.callbacks import Callback\nfrom keras.utils import CustomObjectScope\nfrom keras.optimizers import Adam\nimport keras.backend as K\nimport numpy as np\n\ndef make_model(my_reg):\n    return Sequential([\n        Dense(20, activation='tanh', input_shape=(2,), activity_regularizer=my_reg),\n        Dense(2,  activation='linear'),\n        ])\n<\/code><\/pre>\n\n<pre><code>my_reg = MyActivityRegularizer(a=1)\n\nwith CustomObjectScope({'MyActivityRegularizer':my_reg}): # required for Keras to recognize\n    model = make_model(my_reg)\nopt = Adam(lr=1e-4)\nmodel.compile(optimizer=opt, loss='mse')\n<\/code><\/pre>\n\n<pre><code>x = np.random.randn(320,2) # dummy data\ny = np.random.randn(320,2) # dummy labels\n\nupdate_fn = lambda x: .5 + .4*np.cos(x) #x = number of train updates (optimizer.iterations)\nactivity_regularizer_scheduler = ActivityRegularizerScheduler(model, update_fn)\n\nmodel.fit(x,y,batch_size=32,callbacks=[activity_regularizer_scheduler],\n          epochs=4,verbose=1)\n<\/code><\/pre>\n\n<p>To TRACK your <code>a<\/code> and make sure it's changing, you can get its value at, e.g., each epoch end via:<\/p>\n\n<pre><code>for epoch in range(4):\n    model.fit(x,y,batch_size=32,callbacks=[activity_regularizer_scheduler],epochs=1)\n    print(\"Epoch {} activity_regularizer 'a': {}\".format(epoch,\n            K.get_value(_get_activity_regularizers(model)[0].a)))\n# My output:\n# Epoch 0 activity_regularizer 'a': 0.7190816402435303\n# Epoch 1 activity_regularizer 'a': 0.4982417821884155\n# Epoch 2 activity_regularizer 'a': 0.2838689386844635\n# Epoch 3 activity_regularizer 'a': 0.8644570708274841\n<\/code><\/pre>\n\n<p><hr>\nRegarding <strong>(2)<\/strong>, I'm afraid you're right - the <code>'tanh'<\/code> outputs won't be used; you'll need to pass <code>activation='tanh'<\/code> instead.\n<hr>\nLastly, you <em>can<\/em> do it without a callback, via <code>train_on_batch<\/code> - but a drawback is, you'll need to feed data to the model yourself (and shuffle it, etc):<\/p>\n\n<pre><code>activity_regularizers = _get_activity_regularizers(model)\n\nfor iteration in range(100):\n   x, y = get_data()\n   model.train_on_batch(x,y)\n   iteration = K.get_value(model.optimizer.iterations)\n\n   for activity_regularizer in activity_regularizers:\n       K.set_value(activity_regularizer, update_fn(iteration))\n<\/code><\/pre>\n"}]}
+{"QuestionId":58088867,"AnswerCount":0,"Tags":"<tensorflow><loss-function><discriminator><cross-entropy><gan>","CreationDate":"2019-09-24T22:25:01.527","AcceptedAnswerId":null,"OwnerUserId":10330077.0,"Title":"Calculating loss in dcgan follow up","Body":"<p>This is a follow up to the post:\n<a href=\"https:\/\/stackoverflow.com\/questions\/56550407\/calculating-loss-in-dcgan\">Calculating loss in dcgan<\/a> which got an answer selected, but it doesn't seem right to me.<\/p>\n\n<p>To summarize: In the <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/generative\/dcgan\" rel=\"nofollow noreferrer\">tensor flow tutorial about dcgan<\/a> they defined a loss function like this:<\/p>\n\n<pre><code>def generator_loss(generated_output):\n    return tf.losses.sigmoid_cross_entropy(tf.ones_like(generated_output), generated_output)\n<\/code><\/pre>\n\n<p>The question is why they use \"tf.ones_like(generated_output)\". The answer given by David Parks to the mentioned previous post, was that it's since<\/p>\n\n<blockquote>\n  <p>The discriminators output is in the range [0, 1]<\/p>\n<\/blockquote>\n\n<p>I thought this was the case, but then I noticed the discriminator does, in fact, not return a value between 0 and 1.<\/p>\n\n<p>The definition of the discriminator ends with:<\/p>\n\n<pre><code>model.add(tf.keras.layers.Dense(1))\n<\/code><\/pre>\n\n<p>and not with:<\/p>\n\n<pre><code>model.add(tf.keras.layers.Dense(1))\nmodel.add(tf.keras.layers.Activation(\"sigmoid\"))\n<\/code><\/pre>\n\n<p>as I would have expected.<\/p>\n\n<p>The output of the discriminator for 4 generated and 4 real pictures (after a few epochs) actually looks like looks like:<\/p>\n\n<pre><code>GENERATED:  tf.Tensor(\n[[-1.5261561 ]\n [-2.6044064 ]\n [-0.55789095]\n [-1.0553275 ]], shape=(4, 1), dtype=float32)\nREAL:  tf.Tensor(\n[[ 2.048546 ]\n [ 2.900073 ]\n [-2.3295398]\n [-1.098418 ]], shape=(4, 1), dtype=float32)\n<\/code><\/pre>\n\n<p>I thought this design of the discriminator might be a mistake, and the fact that it still works perhaps a fluke. So I tried adding the sigmoid activation layer to the discriminator, but that actually seriously worsened the performance. It's output looks something like this after a few epochs:<\/p>\n\n<pre><code>GENERATED:  tf.Tensor(\n[[1.5318394e-05]\n [5.3527951e-04]\n [3.1036139e-04]\n [6.9439411e-06]], shape=(4, 1), dtype=float32)\nREAL:  tf.Tensor(\n[[1.1146069e-05]\n [1.4841557e-05]\n [5.9306622e-06]\n [5.2452087e-06]], shape=(4, 1), dtype=float32)\n<\/code><\/pre>\n\n<p>So now I'm totally lost. Is there another explanation? I'm trying to remake the script to generate other (color) images, but it won't learn (it only generates crazy noise), and I suspect this may be part of the reason.<\/p>\n","answers":[]}
+{"QuestionId":58089002,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving><tensorflow-transform>","CreationDate":"2019-09-24T22:44:38.497","AcceptedAnswerId":58102234.0,"OwnerUserId":3864587.0,"Title":"tf.estimator serving function failing","Body":"<p>I am using the tf.estimator to train and serve my tensorflow model. the training completed as expected, but fails in serving. I read my data in as a TFRecordDataset. My parsing function applies a transformation to feature \"x2\". \"x2\" is a string that is split. the tranformed feature is \"x3\".<\/p>\n\n<pre><code>def parse_function(example_proto):\n    features={\"x1\":tf.FixedLenFeature((), tf.string), \"x2\":tf.FixedLenFeature((), \n tf.string),\n         \"label\":tf.FixedLenFeature((), tf.int64)}\n    parsed_features = tf.parse_example(example_proto, features)\n    x3=tf.string_split(parsed_features[\"string\"],',')\n    parsed_features[\"x3\"]=x3\n    return parsed_features, parsed_features[\"label\"]\n<\/code><\/pre>\n\n<p>My serving fucnction is <\/p>\n\n<pre><code>def serving_input_fn():\nreceiver_tensor = {}\nfor feature_name in record_columns:\n    if feature_name in {\"x1\", \"x2\",\"x3\"}:\n        dtype = tf.string \n    else:\n        dtype=tf.int32\n    receiver_tensor[feature_name] = tf.placeholder(dtype, shape=[None])\nfeatures = {\n    key: tf.expand_dims(tensor, -1)\n    for key, tensor in receiver_tensor.items()\n}\nreturn tf.estimator.export.ServingInputReceiver(features, receiver_tensor)\n<\/code><\/pre>\n\n<p>It always worked in the past when I didn't have any transformations in my parsing function, but it fails now with the error.<\/p>\n\n<pre><code>cloud.ml.prediction.prediction_utils.PredictionError: Failed to run the provided model: Exception during running the graph: Cannot feed value of shape (2, 1) for Tensor u'Placeholder_2:0', which has shape '(?,)' (Error code: 2)\n<\/code><\/pre>\n\n<p>I think I have to apply the transformation to \"x2\" in my serving function, but I don't know how. Any help would be greatly appreciated <\/p>\n","answers":[{"AnswerId":"58102234","CreationDate":"2019-09-25T15:49:35.887","ParentId":null,"OwnerUserId":"3864587","Title":null,"Body":"<p>Following this <a href=\"https:\/\/github.com\/GoogleCloudPlatform\/cloudml-samples\/blob\/master\/cloudml-template\/examples\/classification\/census\/trainer\/inputs.py\" rel=\"nofollow noreferrer\">link<\/a><\/p>\n\n<p>I processed feature \"x3\" after creating the receiver_tensor. Splitting the string in the serving fucntion required squeezing the tensor before splitting<\/p>\n\n<pre><code>def serving_input_fn():\n    receiver_tensor = {}\n    receiver_tensor[\"x1\"] = tf.placeholder(tf.string, shape=[None], name=\"x1\")\n    receiver_tensor[\"label\"] = tf.placeholder(tf.int32, shape=[None], name=\"x2\")\n    receiver_tensor[\"x2\"] = tf.placeholder(tf.string, shape=[None], \n      name=\"string\")\n\n    features = {\n        key: tf.expand_dims(tensor, -1)\n        for key, tensor in receiver_tensor.items()\n    }\n    features[\"x3\"]=tf.string_split(tf.squeeze(features[\"x2\"]),',')\n    return tf.estimator.export.ServingInputReceiver(features, receiver_tensor)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58089109,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-24T23:00:11.403","AcceptedAnswerId":null,"OwnerUserId":4107770.0,"Title":"Tensorflow use neural network many times within the graph","Body":"<p>I have the following issue. I have a neural network in tensorflow with a placeholder as the input, and outputs a scalar:<\/p>\n\n<pre><code>N = 10 \n\nn = tf.placeholder(tf.float32, [None, N]) # input data\n\n### Layer 1: ###\nW1 = tf.Variable( tf.random_normal([N, 2], mean=0.0, stddev=0.01, dtype=tf.float32) )\nb1 = tf.Variable( tf.zeros([2]) )\nz1 = tf.matmul(n, W1) + b1\na1 = tf.nn.relu( z1 )\n\n### Network output: ###\naL = tf.cast(a1, tf.complex64)\n\n### Wave function: ###\npsi = tf.abs(tf.exp(aL[:, 0] + 1j*aL[:, 1]))\n<\/code><\/pre>\n\n<p>I want to build a cost function that is constructed as follows:<\/p>\n\n<p>Given some distribution, sample vectors with the shape of the NN's input. Then, compute the output for each sampled vector and perform some calculation, which gives a number I want to minimize. In my case, the computation is as follows (run from a tf.Session):<\/p>\n\n<pre><code>average = 0\nP = 1000\nV = 1\nU = -3\nfor i in range(P + 1):\n    config = np.array(sample[i])\n    psi_i = sess.run(psi, feed_dict={n: config.reshape(1, N)})[0]\n    interaction = np.sum(V * config + 0.5 * U * config * (config - 1)) * psi_i\n    average += interaction\naverage \/= P\nprint(average)\n<\/code><\/pre>\n\n<p>where <code>sample<\/code> contains the sampled vectors. What I am doing here is to call many times the result of the NN for each sampled vector. However, I would like my model to minimize the <code>average<\/code>, for which I need the average to be computed as a tensorflow quantity within the graph.  For this I would have to add the second part of my code within the graph of the first part of my code, but I don't know how to call psi many times in order to feed <code>n<\/code> with the values from the <code>sample<\/code>. Here <code>sample<\/code> is a list of vectors that change for every epoch. My question is: how do I include the for loop in the graph in order to feed n with the vectors from the sample?<\/p>\n","answers":[]}
+{"QuestionId":58089184,"AnswerCount":0,"Tags":"<python><tensorflow><windows-10><openpose>","CreationDate":"2019-09-24T23:12:03.480","AcceptedAnswerId":null,"OwnerUserId":12115851.0,"Title":"TensorRT for python on windows","Body":"<p>I'm having problems using TensorRT for python on windows.<\/p>\n\n<p>I want to use openpose on windows however it requires TensorRT for python. when I'm trying to execute file run_webcam.py (openpose code), I have the following error:<\/p>\n\n<pre><code>ModuleNotFoundError: No module named \"tensorflow.contrib.tensorrt\"\n<\/code><\/pre>\n\n<p>I installed tensorflow package (last version) but it doesn't contain tensorrt class\nLooking for a solution, I found the following information on nvidia website <a href=\"https:\/\/devtalk.nvidia.com\/default\/topic\/1044414\/tensorrt\/tensorrt-python-and-windows\/\" rel=\"nofollow noreferrer\">https:\/\/devtalk.nvidia.com\/default\/topic\/1044414\/tensorrt\/tensorrt-python-and-windows\/<\/a> <\/p>\n\n<p>How does OpenPose work on windows if windows doesn't support TensorRT for python?????<\/p>\n\n<p>Any tips or help is greatly appreciated<\/p>\n","answers":[]}
+{"QuestionId":58089254,"AnswerCount":2,"Tags":"<tensorflow><keras><cudnn>","CreationDate":"2019-09-24T23:22:50.833","AcceptedAnswerId":null,"OwnerUserId":9115361.0,"Title":"Keras says my cuda drivers are incompatibile, but they are in fact compatibile versions","Body":"<p>I ran a simple keras script that trains a conv net on the MNIST database. This script works on my laptop yet not on my PC with the GeForce RTX 2070 graphics card. <\/p>\n\n<p>The error is this:<\/p>\n\n<pre><code>  File \"\/home\/squall\/spencer\/kaggle\/understanding_cloud_organization\/mnist_model.py\", line 67, in &lt;module&gt;\n    validation_data=(x_test, y_test))\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/site-packages\/keras\/engine\/training.py\", line 1239, in fit\n    validation_freq=validation_freq)\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/site-packages\/keras\/engine\/training_arrays.py\", line 196, in fit_loop\n    outs = fit_function(ins_batch)\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/site-packages\/tensorflow\/python\/keras\/backend.py\", line 3292, in __call__\n    run_metadata=self.run_metadata)\n  File \"\/home\/squall\/anaconda3\/envs\/thunder\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1458, in __call__\n    run_metadata_ptr)\ntensorflow.python.framework.errors_impl.UnknownError: 2 root error(s) found.\n  (0) Unknown: 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 conv2d_1\/convolution}}]]\n     [[metrics\/accuracy\/Identity\/_91]]\n  (1) Unknown: 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 conv2d_1\/convolution}}]]\n0 successful operations.\n0 derived errors ignored.\n<\/code><\/pre>\n\n<p>Cuda is 10.1. Driver is 418.56. CuDNN is 7.4.2. Tensorflow is 1.14. According to the official Nvidia chart, these are all compatible versions. <\/p>\n\n<p>Any ideas?<\/p>\n","answers":[{"AnswerId":"58092794","CreationDate":"2019-09-25T07:02:55.843","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>For tensorflow to work, you need Cuda 10 version. Uninstall the Cuda 10.1 completely and install supported Cuda 10. You can read about requirements for TF <a href=\"https:\/\/www.tensorflow.org\/install\/gpu#software_requirements\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"58091069","CreationDate":"2019-09-25T04:15:17.403","ParentId":null,"OwnerUserId":"4887423","Title":null,"Body":"<p>Try this<br>\n<a href=\"https:\/\/i.stack.imgur.com\/0Gq0V.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0Gq0V.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/3Ymp3.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3Ymp3.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>PS: CUDA is 10.0 and cuDNN is 7.6.3 for CUDA10.0<\/p>\n"}]}
+{"QuestionId":58089423,"AnswerCount":0,"Tags":"<python><tensorflow><keras><time-series><lstm>","CreationDate":"2019-09-24T23:50:39.600","AcceptedAnswerId":null,"OwnerUserId":4458718.0,"Title":"Keras - how can LSTM for time series be so accurate?","Body":"<p>SO I'm starting to test LSTM for time series prediction, and I've found a few different notebooks to use with my own data (here's one <a href=\"https:\/\/towardsdatascience.com\/time-series-analysis-visualization-forecasting-with-lstm-77a905180eba\" rel=\"nofollow noreferrer\">example<\/a>)<\/p>\n\n<p>What they all have in common is that they predict one timestep into the future, and do a really good job at matching the test data. I tried forcing an outlier in there, and the prediction almost perfectly matched it:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/5YnQT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5YnQT.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>What's going on here? There's no way the model can learn this from the pattern of the data since it's a made up point, but supposedly by looking at the previous time steps this model will \"know\" an outlier is coming next? I must be missing something, because it predicts the data with an outlier just as well as the data without an outlier... <\/p>\n","answers":[]}
+{"QuestionId":58089499,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-25T00:02:39.267","AcceptedAnswerId":58091351.0,"OwnerUserId":990639.0,"Title":"How to combine many numpy arrays efficiently?","Body":"<p>I am having difficulty trying to load 18k of training data for training with tensorflow. The files are npy files named as such: 0.npy, 1.npy...18000.npy.<\/p>\n\n<p>I was looking around the web and came up with a simple code to first read the files in the correct sequence and trying to concatenate the training data together but it takes forever.. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport glob\nimport re\nimport tensorflow as tf\n\nprint(\"TensorFlow version: {}\".format(tf.__version__))\n\nfiles = glob.glob('D:\/project\/train\/*.npy')\nfiles.sort(key=lambda var:[int(x) if x.isdigit() else x for x in \n           re.findall(r'[^0-9]|[0-9]+', var)])\n# print(files)\n\nfinal_dataset = []\ni = 0\nfor file in files:    \n    dataset = np.load(file, mmap_mode='r')\n    print(i)\n    #print(\"Size of dataset: {} \".format(dataset.shape))\n    if (i==0):\n      final_dataset = dataset\n    else: \n      final_dataset = np.concatenate((final_dataset, dataset), axis = 0)\n    i = i + 1\n\nprint(\"Size of final_dataset: {} \".format(final_dataset.shape))\nnp.save('combined_train.npy', final_dataset)\n\n<\/code><\/pre>\n","answers":[{"AnswerId":"58091351","CreationDate":"2019-09-25T04:50:07.487","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>'Combining' arrays in any way involves (1), creating an array with the two arrays' total size; (2), copying their contents into the array. If you do this each time you load in an array, it repeats 18000 times - with time per iteration <em>growing<\/em> for each iteration (due to larger <code>final_dataset<\/code>). <\/p>\n\n<p>A simple workaround is to <em>append<\/em> the arrays to a list - and then combine them all <em>once<\/em> at the end:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>dataset = []\nfor file in files:\n    data = np.load(file, mmap_mode='r')\n    dataset.append(data)\n\nfinal_dataset = np.concatenate(dataset, axis=0)\n<\/code><\/pre>\n\n<p>But <em>beware<\/em>: be sure <code>final_dataset<\/code> can actually fit in your RAM, else the program will crash. You can find out via <code>ram_required = size_per_file * number_of_files<\/code>. Relevant <a href=\"https:\/\/stackoverflow.com\/questions\/7133885\/fastest-way-to-grow-a-numpy-numeric-array\">SO<\/a>. (To speed things up even further, you can look into <a href=\"https:\/\/stackoverflow.com\/questions\/14677287\/multiprocessing-with-large-data\">multiprocessing<\/a> - but not simple to get working)<\/p>\n"}]}
+{"QuestionId":58089541,"AnswerCount":0,"Tags":"<machine-learning><keras><recurrent-neural-network>","CreationDate":"2019-09-25T00:08:15.000","AcceptedAnswerId":null,"OwnerUserId":866082.0,"Title":"What is the effect of using TimeDistributed layer wrapper?","Body":"<p>Consider the following two models:<\/p>\n\n<pre><code>from tensorflow.python.keras.layers import Input, GRU, Dense, TimeDistributed\nfrom tensorflow.python.keras.models import Model\n\ninputs = Input(batch_shape=(None, None, 100)) \ngru_out = GRU(32, return_sequences=True)(inputs)\ndense = Dense(200, activation='softmax')\ndecoder_pred = TimeDistributed(dense)(gru_out)\nmodel = Model(inputs=inputs, outputs=decoder_pred)\nmodel.summary()\n<\/code><\/pre>\n\n<p>with the output:<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_1 (InputLayer)         (None, None, 100)         0         \n_________________________________________________________________\ngru (GRU)                    (None, None, 32)          12768     \n_________________________________________________________________\ntime_distributed (TimeDistri (None, None, 200)         6600      \n=================================================================\nTotal params: 19,368\nTrainable params: 19,368\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>And the second model:<\/p>\n\n<pre><code>from tensorflow.python.keras.layers import Input, GRU, Dense\nfrom tensorflow.python.keras.models import Model\n\ninputs = Input(batch_shape=(None, None, 100)) \ngru_out = GRU(32, return_sequences=True)(inputs)\ndecoder_pred = Dense(200, activation='softmax')(gru_out)\nmodel = Model(inputs=inputs, outputs=decoder_pred)\nmodel.summary()\n<\/code><\/pre>\n\n<p>with the output:<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_2 (InputLayer)         (None, None, 100)         0         \n_________________________________________________________________\ngru_1 (GRU)                  (None, None, 32)          12768     \n_________________________________________________________________\ndense_1 (Dense)              (None, None, 200)         6600      \n=================================================================\nTotal params: 19,368\nTrainable params: 19,368\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n\n<p>My question is, is the <code>TimeDistributed<\/code> layer wrapper doing anything to the first model? Are these two different in any aspect (considering that their total number of params are identical)?<\/p>\n","answers":[]}
+{"QuestionId":58089579,"AnswerCount":2,"Tags":"<tensorflow><keras><deep-learning><gamma-distribution><tensorflow-probability>","CreationDate":"2019-09-25T00:15:29.797","AcceptedAnswerId":58272738.0,"OwnerUserId":4099168.0,"Title":"How to fit a Keras model to a Gamma Distribution?","Body":"<p>I am trying to fit a keras model in which my output variable is always positive. I want to use a gamma distribution to model this problem. The problem is that the loss always ouptputs NAN.<\/p>\n\n<p>I built the following keras model:<\/p>\n\n<pre><code>model_max = tf.keras.Sequential([\n            tf.keras.layers.Dense(20,input_dim=10, activation=\"relu\"),    \n            tf.keras.layers.Dense(15,activation=\"relu\"),\n            tf.keras.layers.Dense(10,activation=\"relu\"),\n            tf.keras.layers.Dense(5,activation=\"relu\"),\n            tf.keras.layers.Dense(2),\n            tfp.layers.DistributionLambda(lambda t:\n            tfd.Gamma(concentration = tf.math.softplus(0.005*t[...,:1])+0.001,\n             rate = tf.math.softplus(0.005*t[...,1:])+0.001)\n            ),\n])            \n<\/code><\/pre>\n\n<p>Notice that I used softplus because both arguments of the distribution must be positive. Also I added 0.001 to make sure the arguments are always greater than zero.<\/p>\n\n<p>My loss function is as follows:<\/p>\n\n<pre><code>def gamma_loss(y_true, my_dist):\n\n    dist_mean = my_dist.mean()\n    dist_stddev = my_dist.stddev()\n    alpha = (dist_mean \/ dist_stddev)**2\n    beta = dist_mean \/ dist_stddev**2\n    gamma_distr = tfd.Gamma(concentration=alpha, rate=beta)\n    return -tf.reduce_mean(gamma_distr.log_prob(y_true))\n\n<\/code><\/pre>\n\n<p>This function seems to work fine. For example, if I run the following code it runs fine:<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorflow_probability as tfp\ntfd = tfp.distributions\n\ndef gamma_loss(y_true, my_dist):\n\n    dist_mean = my_dist.mean()\n    dist_stddev = my_dist.stddev()\n    alpha = (dist_mean \/ dist_stddev)**2\n    beta = dist_mean \/ dist_stddev**2\n    #print(alpha)\n    gamma_distr = tfd.Gamma(concentration=alpha, rate=beta)\n    return -tf.reduce_mean(gamma_distr.log_prob(y_true)).numpy()\n\ndist = tfd.Gamma(1,1)\n\ngamma_loss(100, dist)\n<\/code><\/pre>\n\n<p>However, if I compile it with the following line:<\/p>\n\n<pre><code>model_max.compile(optimizer=tf.optimizers.Adam(learning_rate = 0.001),loss=gamma_loss)\n<\/code><\/pre>\n\n<p>The loss always outputs nan<\/p>\n\n<p>What am I doing wrong? I have tried different froms of the loss funcion but nothing seems to work. I think it is realted to the <em>concentration<\/em> argument since I already have a similar model to this working with a normal distribution. In that model, I did not use softplus for the mean (loc) because that distribution accepts any positive or negative value. I used the exact structure for the standard deviation as it must also be possitive in the Normal distribution. It works just fine. Why it doesn't work for the Gamma Distribution?<\/p>\n\n<p>Thank you in advice to anyone who can help me understand what I'm doing wrong.<\/p>\n","answers":[{"AnswerId":"58272738","CreationDate":"2019-10-07T15:31:38.493","ParentId":null,"OwnerUserId":"4099168","Title":null,"Body":"<p>I want to share with you all what I did to get my code to work: <\/p>\n\n<p>1) I made sure each layer had a <code>kernel_initializer='random_uniform'<\/code> statement and,\n2) I turned my whole gamma_loss function into: <code>lambda y, p_y: -p_y.log_prob(y)<\/code>v<\/p>\n\n<p>I'm not sure if the gamma_loss was the problem, but I found examples of people doing the same thing I was doing and the much simpler lambda <code>y, p_y: -p_y.log_prob(y)<\/code> function was working fine, so I went with that. I think my main problem was that the weights were not being randomly initialized. <\/p>\n\n<p>Also, I would like to echo some advice I found online during my search for answers: try fitting one single example and make sure that works well before using the real training data. In my case, I implemented this by taking one single training example and replicating that row thousands of times (creating a dataset in which all the rows are equal) and then training my model with only that. When my model was not being able to fit to that, it was easier to go layer by layer analyzing what the outcome of each layer should be.<\/p>\n\n<p>The answer given by Brian Patton was really helpful as it did point me in the right direction which is, try to understand what each layer is outputting and testing your assumptions with a simple example.<\/p>\n\n<p>For future reference, here is what my code now looks like:<\/p>\n\n<pre><code>model_max = tf.keras.Sequential([\n            tf.keras.layers.Dense(20,input_dim=10, activation=\"relu\", kernel_initializer='random_uniform' ),   \n            tf.keras.layers.Dense(15,activation=\"relu\",kernel_initializer='random_uniform' ),\n            tf.keras.layers.Dense(10,activation=\"relu\",kernel_initializer='random_uniform' ),\n            tf.keras.layers.Dense(5,activation=\"relu\",kernel_initializer='random_uniform' ),\n            tf.keras.layers.Dense(2, kernel_initializer='random_uniform' ),\n            tfp.layers.DistributionLambda(lambda t:\n            tfd.Gamma(concentration = tf.math.softplus(t[:,0])+0.000000001,\n            rate = tf.math.softplus(t[:,1])+0.000000001),\n            ),\n])     \n\n\nnegloglik = lambda y, p_y: -p_y.log_prob(y)\n\nmodel_max.compile(optimizer=tf.optimizers.Adamax(learning_rate = 0.0001),loss=negloglik)\n\n<\/code><\/pre>\n"},{"AnswerId":"58118905","CreationDate":"2019-09-26T14:21:09.273","ParentId":null,"OwnerUserId":"9946929","Title":null,"Body":"<p>Definitely drop the <code>.numpy()<\/code> from the end of <code>gamma_loss<\/code>, as that will break gradient backprop.<\/p>\n\n<p>You might want slightly more generous minimal values for gamma parameters, as they can make the distribution quite sharp. In particular a concentration parameter as low as 0.5 makes the distribution extremely concentrated at 0. (This is the one called 'shape\/alpha\/k' on wikipedia <a href=\"https:\/\/en.wikipedia.org\/wiki\/Gamma_distribution\" rel=\"nofollow noreferrer\">https:\/\/en.wikipedia.org\/wiki\/Gamma_distribution<\/a>).<\/p>\n\n<p>This could easily lead to +\/-inf somewhere, which then yields a nan somewhere else.<\/p>\n"}]}
+{"QuestionId":58090424,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-25T02:36:29.103","AcceptedAnswerId":null,"OwnerUserId":11388756.0,"Title":"Trying to use tensorflow.io.gfile.GFile and tensorflow.dataset.map, getting \"Expected binary or unicode string, got Tensor dtype=string","Body":"<p>I'm quite new to using Tensorflow, and imagine someone will quickly tell me I'm doing something stupid so here goes. <\/p>\n\n<p>I'm working with the MSTAR dataset and trying to get it read in. The files have a very strange format, but suffice it to say that if eager execution is on the following code reads and displays an image from the dataset. <\/p>\n\n<pre><code>import tensorflow as tf\nimport matplotlib.pyplot as plt\ntf.enable_eager_execution()\n\nimg1Path='HB15000.018'\nimg2Path='HB15001.018'\n\ndef pathToImgTF(path):\n    with tf.io.gfile.GFile(path,'rb') as filePath:\n        step1=filePath.readlines()\n        step2=[x.strip(b'\\n') for x in step1]\n        for x in step2:\n            if b'PhoenixHeaderLength' in x:\n                line=x.strip().split(b'=')\n                PhoenixHeaderLength=int(line[1])\n            elif b'native_header_length' in x:\n                line=x.strip().split(b'=')\n                native_header_length=int(line[1])\n            elif b'NumberOfColumns' in x:\n                line=x.strip().split(b'=')\n                NumberOfColumns=int(line[1])\n            elif b'NumberOfRows' in x:\n                line=x.strip().split(b'=')\n                NumberOfRows=int(line[1])\n        filePath.seek(PhoenixHeaderLength+native_header_length)\n        step3=tf.decode_raw(filePath.read(),out_type=tf.float32,little_endian=False)\n        depth_major=tf.reshape(step3,[2,NumberOfRows,NumberOfColumns])\n        image=tf.transpose(depth_major,[1,2,0])[:,:,0] #Cut off phase for now\n    return image\n\nimg=pathToImgTF(imgPath)\nplt.imshow(img,cmap='gray')\n<\/code><\/pre>\n\n<p>I would like to use tf.dataset.from_tensor_slices, but it appears that isn't an option because the following code:<\/p>\n\n<pre><code>ds=tf.data.Dataset.from_tensor_slices([img1Path,img2Path])\nds=ds.map(pathToImgTF)\n<\/code><\/pre>\n\n<p>Gives the error \"TypeError: Expected binary or unicode string, got tf.Tensor 'args_0:0' shape=() dtype=string\"<\/p>\n\n<p>The traceback looks to me like it's breaking specifically on 'filePath.readlines()', any help would be greatly appreciated. <\/p>\n\n<p>Full error output:<\/p>\n\n<blockquote>\n  <p>--------------------------------------------------------------------------- TypeError                                 Traceback (most recent call\n  last) ipython-input-6-e12909fb73cd in module\n        1 ds=tf.data.Dataset.from_tensor_slices([img1Path,img2Path])\n  ----> 2 ds=ds.map(pathToImgTF)<\/p>\n  \n  <p>~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\data\\ops\\dataset_ops.py\n  in map(self, map_func, num_parallel_calls)    1770     if\n  num_parallel_calls is None:    1771       return DatasetV1Adapter(\n  -> 1772           MapDataset(self, map_func, preserve_cardinality=False))    1773     else:    1774       return\n  DatasetV1Adapter(<\/p>\n  \n  <p>~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\data\\ops\\dataset_ops.py\n  in <strong>init<\/strong>(self, input_dataset, map_func, use_inter_op_parallelism,\n  preserve_cardinality, use_legacy_function)    3188<br>\n  self._transformation_name(),    3189         dataset=input_dataset,\n  -> 3190         use_legacy_function=use_legacy_function)    3191     variant_tensor = gen_dataset_ops.map_dataset(    3192<br>\n  input_dataset._variant_tensor,  # pylint: disable=protected-access<\/p>\n  \n  <p>~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\data\\ops\\dataset_ops.py\n  in <strong>init<\/strong>(self, func, transformation_name, dataset, input_classes,\n  input_shapes, input_types, input_structure, add_to_graph,\n  use_legacy_function, defun_kwargs)    2553       resource_tracker =\n  tracking.ResourceTracker()    2554       with\n  tracking.resource_tracker_scope(resource_tracker):\n  -> 2555         self._function = wrapper_fn._get_concrete_function_internal()    2556         if\n  add_to_graph:    2557<br>\n  self._function.add_to_graph(ops.get_default_graph())<\/p>\n  \n  <p>~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\eager\\function.py\n  in _get_concrete_function_internal(self, *args, **kwargs)    1353<br>\n  \"\"\"Bypasses error checking when getting a graph function.\"\"\"    1354<br>\n  graph_function =\n  self._get_concrete_function_internal_garbage_collected(\n  -> 1355         *args, **kwargs)    1356     # We're returning this concrete function to someone, and they may keep a    1357     #\n  reference to the FuncGraph without keeping a reference to the<\/p>\n  \n  <p>~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\eager\\function.py\n  in _get_concrete_function_internal_garbage_collected(self, *args,\n  **kwargs)    1347     if self.input_signature:    1348       args, kwargs = None, None\n  -> 1349     graph_function, _, _ = self._maybe_define_function(args, kwargs)    1350     return graph_function    1351 <\/p>\n  \n  <p>~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\eager\\function.py\n  in _maybe_define_function(self, args, kwargs)    1650<br>\n  graph_function = self._function_cache.primary.get(cache_key, None)<br>\n  1651       if graph_function is None:\n  -> 1652         graph_function = self._create_graph_function(args, kwargs)    1653         self._function_cache.primary[cache_key] =\n  graph_function    1654       return graph_function, args, kwargs<\/p>\n  \n  <p>~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\eager\\function.py\n  in _create_graph_function(self, args, kwargs,\n  override_flat_arg_shapes)    1543             arg_names=arg_names,<br>\n  1544             override_flat_arg_shapes=override_flat_arg_shapes,\n  -> 1545             capture_by_value=self._capture_by_value),    1546         self._function_attributes)    1547 <\/p>\n  \n  <p>~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\func_graph.py\n  in func_graph_from_py_func(name, python_func, args, kwargs, signature,\n  func_graph, autograph, autograph_options, add_control_dependencies,\n  arg_names, op_return_value, collections, capture_by_value,\n  override_flat_arg_shapes)\n      713                                           converted_func)\n      714 \n  --> 715       func_outputs = python_func(*func_args, **func_kwargs)\n      716 \n      717       # invariant: <code>func_outputs<\/code> contains only Tensors, CompositeTensors,<\/p>\n  \n  <p>~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\data\\ops\\dataset_ops.py\n  in wrapper_fn(*args)    2547           attributes=defun_kwargs)<br>\n  2548       def wrapper_fn(*args):  # pylint: disable=missing-docstring\n  -> 2549         ret = _wrapper_helper(*args)    2550         ret = self._output_structure._to_tensor_list(ret)    2551         return\n  [ops.convert_to_tensor(t) for t in ret]<\/p>\n  \n  <p>~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\data\\ops\\dataset_ops.py\n  in _wrapper_helper(*args)    2487         nested_args = (nested_args,)\n  2488 \n  -> 2489       ret = func(*nested_args)    2490       # If <code>func<\/code> returns a list of tensors, <code>nest.flatten()<\/code> and    2491       #\n  <code>ops.convert_to_tensor()<\/code> would conspire to attempt to stack<\/p>\n  \n  <p> in pathToImgTF(path)\n        9 def pathToImgTF(path):\n       10     with tf.io.gfile.GFile(path,'rb') as filePath:\n  ---> 11         step1=filePath.readlines()\n       12         step2=[x.strip(b'\\n') for x in step1]\n       13         for x in step2:<\/p>\n  \n  <p>~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\lib\\io\\file_io.py\n  in readlines(self)\n      181   def readlines(self):\n      182     \"\"\"Returns all lines from the file in a list.\"\"\"\n  --> 183     self._preread_check()\n      184     lines = []\n      185     while True:<\/p>\n  \n  <p>~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\lib\\io\\file_io.py\n  in _preread_check(self)\n       82                                            \"File isn't open for reading\")\n       83       self._read_buf = pywrap_tensorflow.CreateBufferedInputStream(\n  ---> 84           compat.as_bytes(self.__name), 1024 * 512)\n       85 \n       86   def _prewrite_check(self):<\/p>\n  \n  <p>~\\AppData\\Local\\Continuum\\anaconda3\\lib\\site-packages\\tensorflow\\python\\util\\compat.py\n  in as_bytes(bytes_or_text, encoding)\n       63   else:\n       64     raise TypeError('Expected binary or unicode string, got %r' %\n  ---> 65                     (bytes_or_text,))\n       66 \n       67 <\/p>\n  \n  <p>TypeError: Expected binary or unicode string, got tf.Tensor\n  'args_0:0' shape=() dtype=string<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"58092743","CreationDate":"2019-09-25T06:58:59.737","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>I am not sure what kind of data do <code>imgPath1<\/code> and <code>imgPath2<\/code> have. Anyways, they must be  of tensor type with predefined data type. These two consists the actual data.<\/p>\n\n<p>If that is the case, then change the square brackets to circular brackets as follows-\n<code>ds=tf.data.Dataset.from_tensor_slices((imgPath1,imgPath2))<\/code><\/p>\n\n<p>If this still throwing the same error, then please give information where exactly it is throwing the error.<\/p>\n"}]}
+{"QuestionId":58090432,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-lite>","CreationDate":"2019-09-25T02:37:54.680","AcceptedAnswerId":null,"OwnerUserId":9034499.0,"Title":"why tensorflow TFLiteConverter.from_session require the same size for input and output","Body":"<p>I am trying to use TFLiteConverter to convert my network. So I tried the sample code first. It works. But after some modification, it sends back error. Seems the input_array and output_array must be the same size. I just don't understand why. Can anybody help me?<\/p>\n\n<p>I modified the size of img from and the size of var from [1,64,64,3 to [1,64,3,1]\nthe complete code is pasted bellow<code>enter code here<\/code><\/p>\n\n<pre><code>import tensorflow as tf\n\nimg = tf.placeholder(name=\"img\", dtype=tf.float32, shape=(1, 64, 64, 1))\nvar = tf.get_variable(\"weights\", dtype=tf.float32, shape=(1, 64, 3, 1))\nval = tf.matmul(img,var)\nout = tf.identity(val, name=\"out\")\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print(val.shape)\n    converter = tf.lite.TFLiteConverter.from_session(sess, [img], [out])\n    tflite_model = converter.convert()\n    open(\"converted_model.tflite\", \"wb\").write(tflite_model)\n<\/code><\/pre>\n\n<p>The ERROR message:\nValueError: Dimensions must be equal, but are 1 and 3 for 'MatMul' (op: 'BatchMatMulV2') with input shapes: [1,64,64,1], [1,64,3,1].<\/p>\n","answers":[{"AnswerId":"58107293","CreationDate":"2019-09-25T22:23:09.843","ParentId":null,"OwnerUserId":"834565","Title":null,"Body":"<p>The problem is not with the TFLite conversion, but with build the graph in the first place.<\/p>\n\n<p><code>tf.matmul<\/code> operates on the inner-most 2D matrices in your tensors. So in your case, you are trying to matrix multiply a matrix of shape <code>64x1<\/code> by a matrix of size <code>3x1<\/code>, which is not valid. Matrix multiplication requires that columns of the first operand is equal to the rows in the second operand, but here <code>1 != 3<\/code> so it doesn't work.<\/p>\n\n<p>For example, replace the <code>3<\/code> by a <code>1<\/code> then it will work :<\/p>\n\n<pre><code>import tensorflow as tf\n\nimg = tf.placeholder(name=\"img\", dtype=tf.float32, shape=(1, 64, 64, 1))\nvar = tf.get_variable(\"weights\", dtype=tf.float32, shape=(1, 64, 3, 1))\nval = tf.matmul(img,var)\nout = tf.identity(val, name=\"out\")\n<\/code><\/pre>\n"}]}
+{"QuestionId":58090549,"AnswerCount":0,"Tags":"<python><tensorflow><keras><base64><tensorflow-serving>","CreationDate":"2019-09-25T02:56:42.593","AcceptedAnswerId":null,"OwnerUserId":6476080.0,"Title":"tensorflow serving with estimator: serving_input_receiver_fn equivalent for output?","Body":"<p>Is there an equivalent of <code>serving_input_receiver_fn<\/code> for processing outputs? It doesn't seem to be the case. What is the best workaround? The use case I have is the following:<\/p>\n\n<ul>\n<li>Given a base64 encoded image as input, do some manipulation (e.g. adversarial attack) on that image using keras pretrained models. Then base64 encode that manipulated image and return.<\/li>\n<\/ul>\n\n<p>Keras model can only compile <strong>without<\/strong> the base64 encoding because standard loss functions don't allow string outputs. So, in the following code, I'm not sure where I can insert the output base64 encoding.<\/p>\n\n<pre><code>estimator = keras.estimator.model_to_estimator(\n    keras_model_path=KERAS_MODEL_PATH,\n    model_dir=MODEL_DIR)\nestimator.export_savedmodel(\n    EXPORT_PATH,\n    serving_input_receiver_fn=serving_input_receiver_fn) \n<\/code><\/pre>\n\n<p>I am also fine rewriting my Keras model in vanilla tensorflow, if that let's me base64 encode the output image. I would like to know the general approach for doing so.<\/p>\n","answers":[]}
+{"QuestionId":58090737,"AnswerCount":1,"Tags":"<javascript><tensorflow><web-worker><tensorflow.js>","CreationDate":"2019-09-25T03:24:35.947","AcceptedAnswerId":58104489.0,"OwnerUserId":10061137.0,"Title":"TensorflowJS universal sentence encoder in WebWorker","Body":"<p>Trying to run USE in webworker to avoid performance issues, but it lacks embed() method.<\/p>\n\n<h3>WebWorker<\/h3>\n\n<pre><code>importScripts('https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/tfjs@1.2.9\/dist\/tf.min.js');\nimportScripts('https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow-models\/universal-sentence-encoder@1.1.1\/dist\/universal-sentence-encoder.min.js');\ntf.setBackend('cpu');\n\nonmessage = message =&gt; {\n  use.load().then(model =&gt; {\n    const sentences = [\n      'I like my phone.',\n      'Your cellphone looks great.'\n    ];\n    \/**\n      model = { model, tokenizer }\n    *\/\n    model.embed(sentences).then(async embeddings =&gt; {\n      const vec = await embeddings.array();\n      const cosine = tf.losses.cosineDistance(vec[0], vec[1], 0);\n      const result = await cosine.data();\n      console.log(`${Math.round((1 - result) * 100)}%`);\n    });\n  });\n}\n<\/code><\/pre>\n\n<p>There is no embed() inside model which is the case for main thread<br>\nIs it even possible to run USE in webworkers?<\/p>\n","answers":[{"AnswerId":"58104489","CreationDate":"2019-09-25T18:25:14.183","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>Universal Sentence Encoder (USE) can be used in a webworker. See below this working snippet with the result sent from the worker to the main thread<\/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;head&gt;\r\n &lt;script src=\"https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/tfjs@0.14.2\/dist\/tf.min.js\"&gt;&lt;\/script&gt;\r\n &lt;script&gt;\r\n        const worker_function = () =&gt;\r\n{\r\n importScripts('https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/tfjs@1.2.9\/dist\/tf.min.js');\r\n importScripts('https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow-models\/universal-sentence-encoder@1.1.1\/dist\/universal-sentence-encoder.min.js');\r\n tf.setBackend('cpu');\r\n onmessage = message =&gt;\r\n {\r\n        console.log(\"from worker\")\r\n  use.load().then(model =&gt;\r\n  {\r\n   const sentences = [\r\n    'I like my phone.',\r\n    'Your cellphone looks great.'\r\n   ];\r\n   model.embed(sentences).then(async embeddings =&gt;\r\n   { \r\n                console.log(await embeddings)\r\n    const vec = await embeddings.array();\r\n    const cosine = tf.losses.cosineDistance(vec[0], vec[1], 0);\r\n    const result = await cosine.data();\r\n                console.log(`${Math.round((1 - result) * 100)}%`);\r\n                \r\n                postMessage({res: cosine.dataSync(), shape: cosine.shape})\r\n            });\r\n            \r\n  });\r\n };\r\n}\r\nif (window != self)\r\n worker_function();\r\n    &lt;\/script&gt;\r\n &lt;script&gt;\r\n        const worker = new Worker(URL.createObjectURL(new Blob([\"(\" + worker_function.toString() + \")()\"], { type: 'text\/javascript' })));\r\n        worker.postMessage({});\r\n        worker.onmessage = (message) =&gt; {\r\n            console.log('from main thread')\r\n            const {data} = message\r\n         tf.tensor(message.data.res, message.data.shape).print()\r\n        }\r\n    &lt;\/script&gt;\r\n&lt;\/head&gt;<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n"}]}
+{"QuestionId":58090895,"AnswerCount":0,"Tags":"<python><tensorflow><face>","CreationDate":"2019-09-25T03:48:50.717","AcceptedAnswerId":null,"OwnerUserId":11674803.0,"Title":"IndexError: index 3 is out of bounds for axis 0 with size 1","Body":"<p>I'm trying run the Github project 'cleverhans', the script is in the folder <code>cleverhans\/example\/face_adversarial_faces\/facenet_fgsm.py<\/code>, all dependencies has been installed but when I run <code>python facenet_fgsm.py<\/code> in virtual enviroment, the terminal throws this error 'IndexError: index 3 is out of bounds for axis 0 with size 1', how to get rid of this issue?<\/p>\n\n<p>The script of <code>facenet_fgsm.py<\/code> is as following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow.python.util.deprecation as deprecation\ndeprecation._PRINT_DEPRECATION_WARNINGS = False\n\nimport cv2\nfrom skimage.transform import resize\nimport matplotlib.pyplot as plt\nimport tensorflow as tf\nimport numpy as np\nfrom cleverhans.cleverhans.model import Model\nfrom cleverhans.cleverhans.attacks.fast_gradient_method import FastGradientMethod\nfrom cleverhans.examples.facenet_adversarial_faces.set_loader import load_testset\nfrom facenet.src.facenet import load_model\n\nclass InceptionResnetV1Model(Model):\n    model_path = \"~\/Documents\/fnatk\/facenet\/src\/models\/20190702-103304\"\n\n    def __init__(self):\n        super(InceptionResnetV1Model, self).__init__(scope='model')\n\n        # Load Facenet CNN\n        load_model(self.model_path)\n        # Save input and output tensors references\n        graph = tf.get_default_graph()\n        self.face_input = graph.get_tensor_by_name(\"input:0\")\n        self.embedding_output = graph.get_tensor_by_name(\"embeddings:0\")\n\n    def convert_to_classifier(self):\n        # Create victim_embedding placeholder\n        self.victim_embedding_input = tf.placeholder(\n            tf.float32,\n            shape=(None, 128))\n\n        # Squared Euclidean Distance between embeddings\n        distance = tf.reduce_sum(\n            tf.square(self.embedding_output - self.victim_embedding_input),\n            axis=1)\n\n        # Convert distance to a softmax vector\n        # 0.99 out of 4 is the distance threshold for the Facenet CNN\n        threshold = 0.99\n        score = tf.where(\n            distance &gt; threshold,\n            0.5 + ((distance - threshold) * 0.5) \/ (4.0 - threshold),\n            0.5 * distance \/ threshold)\n        reverse_score = 1.0 - score\n        self.softmax_output = tf.transpose(tf.stack([reverse_score, score]))\n\n        # Save softmax layer\n        self.layer_names = []\n        self.layers = []\n        self.layers.append(self.softmax_output)\n        self.layer_names.append('logits')\n\n        # define logits\n        def calc(self, image):\n            model_def = \"models.inception_resnet_v1\"\n            network = importlib.import_module(model_def)\n            prelogits, _ = network.inference(images_placeholder, 1.0,\n                                             phase_train=False, weight_decay=0.0, reuse=False)\n            batch_norm_params = {\n                # Decay for the moving averages\n                'decay': 0.995,\n                # epsilon to prevent 0s in variance\n                'epsilon': 0.001,\n                # force in-place updates of mean and variance estimates\n                'updates_collections': None,\n                # Moving averages ends up in the trainable variables collection\n                'variables_collections': [tf.GraphKeys.TRAINABLE_VARIABLES],\n                # Only update statistics during training mode\n                'is_training': False\n            }\n            bottleneck = slim.fully_connected(prelogits, 128, activation_fn=None,\n                                              weights_initializer=tf.truncated_normal_initializer(stddev=0.1),\n                                              weights_regularizer=slim.l2_regularizer(0.0),\n                                              normalizer_fn=slim.batch_norm,\n                                              normalizer_params=batch_norm_params,\n                                              scope='Bottleneck', reuse=False)\n            logits = slim.fully_connected(bottleneck, 2, activation_fn=None,\n                                          weights_initializer=tf.truncated_normal_initializer(stddev=0.1),\n                                          weights_regularizer=slim.l2_regularizer(0.),\n                                          scope='Logits', reuse=False)\n            return logits\n\n    def fprop(self, x, set_ref=False):\n        return dict(zip(self.layer_names, self.layers))\n\n\nwith tf.Graph().as_default():\n  with tf.Session() as sess:\n    # Load model\n    model = InceptionResnetV1Model()\n    # Convert to classifier\n    model.convert_to_classifier()\n\n    # Load pairs of faces and their labels in one-hot encoding\n    faces1, faces2, labels = load_testset(10)\n\n    # Create victims' embeddings using Facenet itself\n    graph = tf.get_default_graph()\n    phase_train_placeholder = graph.get_tensor_by_name(\"phase_train:0\")\n    feed_dict = {model.face_input: faces2,\n                 phase_train_placeholder: False}\n    victims_embeddings = sess.run(\n        model.embedding_output, feed_dict=feed_dict)\n\n    def fgsm_attack(eps, steps):\n        # Define FGSM for the model\n        alpha = eps \/ steps\n        fgsm = FastGradientMethod(model)\n        fgsm_params = {'eps': alpha,\n                       'clip_min': 0.,\n                       'clip_max': 1.}\n        adv_x = fgsm.generate(model.face_input, **fgsm_params)\n\n        # Run FGSM\n        adv = faces1\n        for i in range(steps):\n          print(\"FGSM step \" + str(i + 1))\n          feed_dict = {model.face_input: adv,\n                       model.victim_embedding_input: victims_embeddings,\n                       phase_train_placeholder: False}\n          adv = sess.run(adv_x, feed_dict=feed_dict)\n\n        # Test accuracy of the model\n        batch_size = graph.get_tensor_by_name(\"batch_size:0\")\n\n        feed_dict = {model.face_input: faces1,\n                     model.victim_embedding_input: victims_embeddings,\n                     phase_train_placeholder: False,\n                     batch_size: 64}\n        real_labels = sess.run(model.softmax_output, feed_dict=feed_dict)\n\n        accuracy = np.mean(\n            (np.argmax(labels, axis=-1)) == (np.argmax(real_labels, axis=-1))\n        )\n        print('Accuracy: ' + str(accuracy * 100) + '%')\n\n        # Test accuracy against adversarial examples\n        feed_dict = {model.face_input: adv,\n                     model.victim_embedding_input: victims_embeddings,\n                     phase_train_placeholder: False,\n                     batch_size: 64}\n        adversarial_labels = sess.run(\n            model.softmax_output, feed_dict=feed_dict)\n\n        same_faces_index = np.where((np.argmax(labels, axis=-1) == 0))\n        different_faces_index = np.where((np.argmax(labels, axis=-1) == 1))\n\n        dodge_acc = np.mean(\n            (np.argmax(labels[same_faces_index], axis=-1)) ==\n            (np.argmax(adversarial_labels[same_faces_index], axis=-1))\n        )\n        print('Accuracy against adversarial examples for '\n              + 'same person faces (dodging): '\n              + str(dodge_acc * 100)\n              + '%')\n\n        impersonate_acc = np.mean(\n            (np.argmax(labels[different_faces_index], axis=-1)) == (\n                np.argmax(adversarial_labels[different_faces_index], axis=-1))\n        )\n        print('Accuracy against adversarial examples for '\n              + 'different people faces (impersonation): '\n              + str(impersonate_acc * 100)\n              + '%')\n\n        return dodge_acc, impersonate_acc\n\n    # Doesn't work atm\n    def plot_graph(eps, dodge_accs, impersonate_accs):\n        plt.plot(eps, dodge_accs, label = \"dodge\")\n        plt.plot(eps, impersonate_accs, label = \"impersonation\")\n\n        plt.xlabel(\"Epsilon\")\n        plt.ylabel(\"Accuracy (%)\")\n        plt.title(\"Accuracy against Epsilon graph\")\n        plt.legend()\n        plt.show()\n\n\n    #all_eps = [0.01, 0.05, 0.1, 0.2, 0.3, 0.5]\n    all_eps = [0.01]\n    all_steps = [1]\n    dodge_accs = []\n    impersonate_accs = []\n    #steps = 10\n    # eps = 0.3\n    for eps in all_eps:\n        for steps in all_steps:\n            print(\"eps: \" + str(eps) + \"    steps: \" + str(steps))\n            dodge_acc, impersonate_acc = fgsm_attack(eps, steps)\n            dodge_accs.append(dodge_acc)\n            impersonate_accs.append(impersonate_acc)\n    #plot_graph(np.array(eps), np.array(dodge_accs), np.array(impersonate_accs))\n\n<\/code><\/pre>\n\n<p>And the terminal throw out the following error:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Traceback (most recent call last):\n  File \"facenet_fgsm.py\", line 191, in &lt;module&gt;\n    dodge_acc, impersonate_acc = fgsm_attack(eps, steps)\n  File \"facenet_fgsm.py\", line 151, in fgsm_attack\n    (np.argmax(adversarial_labels[same_faces_index], axis=-1))\nIndexError: index 3 is out of bounds for axis 0 with size 1\n<\/code><\/pre>\n\n<p>I expect to run this script successfully.<\/p>\n","answers":[]}
+{"QuestionId":58091138,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><deep-learning><conv-neural-network><layer>","CreationDate":"2019-09-25T04:25:49.257","AcceptedAnswerId":null,"OwnerUserId":11175276.0,"Title":"filter size in convolution layers","Body":"<p>what happen if i add another convolution layer after convolution layer with same filter size<\/p>\n\n<p>for example inn a network of around 20 layers i have to chose different filter size among convolution layers. so what will be the impact if  I do something like this<\/p>\n\n<p>\"\"\"<\/p>\n\n<p>inner = MaxPooling2D(pool_size=(1,2),strides=2,padding='valid', name ='max3')(inner)<\/p>\n\n<p>inner = Conv2D(64, (3,3),strides=1,padding= 'same', name='conv6', kernel_initializer='he_normal')(inner)<\/p>\n\n<p>inner = Conv2D(64, (3,3),strides=1,padding= 'same', name='conv7', kernel_initializer='he_normal')(inner)<\/p>\n\n<p>inner = Conv2D(64, (3,3),strides=1,padding= 'same', name='conv8', kernel_initializer='he_normal')(inner)<br>\ninner = BatchNormalization()(inner)<\/p>\n\n<p>\"\"\"<\/p>\n","answers":[]}
+{"QuestionId":58091166,"AnswerCount":0,"Tags":"<python><tensorflow><conv-neural-network>","CreationDate":"2019-09-25T04:29:32.477","AcceptedAnswerId":null,"OwnerUserId":12116660.0,"Title":"What are the parameters for slim.repeat() and max_pool3d() in TensorFlow?","Body":"<p>This is a code for a 3D CNN for video processing. I'm having difficulties understanding the input parameters of the following functions: <\/p>\n\n<pre><code>slim.repeat(inputs, 1, slim.conv3d, 16, [1, 3, 3], scope='conv1')\n\nnet = PReLU(net, 'conv1_activation')\n\ntf.nn.max_pool3d(net, strides=[1, 1, 2, 2, 1], ksize=[1, 1, 3, 3, 1],padding='VALID', name='pool1')\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58091168,"AnswerCount":1,"Tags":"<python><tensorflow><conv-neural-network>","CreationDate":"2019-09-25T04:29:39.980","AcceptedAnswerId":null,"OwnerUserId":11118799.0,"Title":"What are the parameters in slim.conv3d?","Body":"<p>This is code for a layer in 3D CNN where my input is video frames. I am having difficulty in understanding what the parameters mean.<\/p>\n\n<pre><code>net = slim.conv3d(net, 64, [1, 3, 3], scope='conv32')\n\nnet = PReLU(net, 'conv32_activation')\n\nnet = tf.nn.max_pool3d(net, strides=[1, 1, 2, 2, 1], ksize=[1, 1, 3, 3, 1], padding='VALID', name='pool3')\n<\/code><\/pre>\n\n<p>Does [1,3,3] refer to my filter size?\nand what are the 5 numbers in ksize?    <\/p>\n","answers":[{"AnswerId":"58091276","CreationDate":"2019-09-25T04:42:51.827","ParentId":null,"OwnerUserId":"7454706","Title":null,"Body":"<p>Welcome to SO.<\/p>\n\n<p>Here you are using Conv3D, \nThere are a lot of parameters, you can see them <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\/conv3d\" rel=\"nofollow noreferrer\">here conv3d<\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/max_pool3d\" rel=\"nofollow noreferrer\">maxpool3d<\/a>.<\/p>\n\n<p>Yes, [1,3,3] refers to your filter size.<br>\nksize is the size of the window, essentially the dimensions that you want to consider while using maxpool.<\/p>\n\n<p><a href=\"http:\/\/cs231n.github.io\/convolutional-networks\/\" rel=\"nofollow noreferrer\">Here<\/a> is an awesome guide to CNN with all things explained.<\/p>\n"}]}
+{"QuestionId":58092004,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><pytorch><text-classification><transformer>","CreationDate":"2019-09-25T06:02:53.337","AcceptedAnswerId":null,"OwnerUserId":6303639.0,"Title":"How to do sequence classification with pytorch nn.Transformer?","Body":"<p>I am doing a sequence classification task using <code>nn.TransformerEncoder()<\/code>. Whose pipeline is similar to <code>nn.LSTM()<\/code>.<\/p>\n\n<p>I have tried several temporal features fusion methods:<\/p>\n\n<ol>\n<li><p>Selecting the final outputs as the representation of the whole sequence.<\/p><\/li>\n<li><p>Using an affine transformation to fuse these features.<\/p><\/li>\n<li><p>Classifying the sequence frame by frame, and then select the max values to be the category of the whole sequence.<\/p><\/li>\n<\/ol>\n\n<p>But, all these 3 methods got a terrible accuracy, only <strong>25%<\/strong> for 4 categories classification. While using nn.LSTM with the last hidden state, I can achieve <strong>83%<\/strong> accuracy easily. I tried plenty of hyperparameters of <code>nn.TransformerEncoder()<\/code>, but without any improvement for the accuracy.<\/p>\n\n<p>I have no idea about how to adjust this model now. Could you give me some practical advice? Thanks.<\/p>\n\n<p>For <code>LSTM<\/code>: the <code>forward()<\/code> is:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    def forward(self, x_in, x_lengths, apply_softmax=False):\n\n        # Embed\n        x_in = self.embeddings(x_in)\n\n        # Feed into RNN\n        out, h_n = self.LSTM(x_in) #shape of out: T*N*D\n\n        # Gather the last relevant hidden state\n        out = out[-1,:,:] # N*D\n\n        # FC layers\n        z = self.dropout(out)\n        z = self.fc1(z)\n        z = self.dropout(z)\n        y_pred = self.fc2(z)\n\n        if apply_softmax:\n            y_pred = F.softmax(y_pred, dim=1)\n        return y_pred\n<\/code><\/pre>\n\n<p>For <code>transformer<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    def forward(self, x_in, x_lengths, apply_softmax=False):\n\n        # Embed\n        x_in = self.embeddings(x_in)\n\n        # Feed into RNN\n        out = self.transformer(x_in)#shape of out T*N*D\n\n        # Gather the last relevant hidden state\n        out = out[-1,:,:] # N*D\n\n        # FC layers\n        z = self.dropout(out)\n        z = self.fc1(z)\n        z = self.dropout(z)\n        y_pred = self.fc2(z)\n\n        if apply_softmax:\n            y_pred = F.softmax(y_pred, dim=1)\n        return y_pred\n<\/code><\/pre>\n","answers":[{"AnswerId":"58103966","CreationDate":"2019-09-25T17:47:05.910","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p>The accuracy you mentioned indicates that something is wrong. Since you are comparing LSTM with TransformerEncoder, I want to point to some crucial differences. <\/p>\n\n<ol>\n<li><p><strong>Positional embeddings<\/strong>: This is very important since the Transformer does not have recurrence concept and so it doesn't capture sequence information. So, make sure you add positional information along with the input embeddings.<\/p><\/li>\n<li><p><strong>Model architecture<\/strong>: <code>d_model<\/code>, <code>n_head<\/code>, <code>num_encoder_layers<\/code> are important. Go with the default size as used in Vaswani et al., 2017. (<code>d_model=512<\/code>, <code>n_head=8<\/code>, <code>num_encoder_layers=6<\/code>)<\/p><\/li>\n<li><p><strong>Optimization<\/strong>: In many scenarios, it has been found that the Transformer needs to be trained with smaller learning rate, large batch size, WarmUpScheduling.<\/p><\/li>\n<\/ol>\n\n<p>Last but not least, for a sanity check, just make sure the parameters of the model is updating. You can also check the training accuracy to make sure the accuracy keeps increasing as the training proceeds.<\/p>\n\n<p>Although it is difficult to say what is exactly wrong in your code but I hope that the above points will help!<\/p>\n"}]}
+{"QuestionId":58092087,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tf.keras>","CreationDate":"2019-09-25T06:09:58.127","AcceptedAnswerId":58143583.0,"OwnerUserId":5189929.0,"Title":"tf and tf.keras Dense layer shows completely different behavior in my setup","Body":"<p>While using tensorflow 1.14, I noticed some very strange behavior when using tf.layers.Dense vs tf.keras.layers.Dense. People on Stackoverflow say that these two layers are exactly the same, and I basically would agree, but having a look at the discounted reward while training an AC agent results in the following graph:<\/p>\n\n<p><img src=\"https:\/\/code.julien.li\/so-tf-problem.png\" alt=\"tf vs tf.keras\"><\/p>\n\n<p>The arguments are exactly the same. Repeated runs lead to the same result (see differently colored data in image). As far as I understand the code, one of the Dense layers inherits from the other: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.14.0\/tensorflow\/python\/keras\/layers\/core.py#L913\" rel=\"nofollow noreferrer\">tf.keras.layers.core<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.14.0\/tensorflow\/python\/layers\/core.py#L33\" rel=\"nofollow noreferrer\">tf.layers.core<\/a>.<\/p>\n\n<p>Is anyone able to explain this behavior?<\/p>\n","answers":[{"AnswerId":"58143583","CreationDate":"2019-09-28T05:19:43.637","ParentId":null,"OwnerUserId":"5189929","Title":null,"Body":"<p>According to a response to a <a href=\"https:\/\/github.com\/hill-a\/stable-baselines\/issues\/220#issuecomment-470083299\" rel=\"nofollow noreferrer\">similar issue on the stable_baseline repository<\/a>, it seems that keras does not support shared weights between multiple agents. Therefore, when training an actor-critic network with multiple instances, every environment has its own network which leads to completely different results. The fix is to only use tensorflow layers directly which support reuse of the same weights.<\/p>\n"}]}
+{"QuestionId":58092176,"AnswerCount":2,"Tags":"<python><keras><functional-programming><neural-network><sequential>","CreationDate":"2019-09-25T06:17:45.167","AcceptedAnswerId":null,"OwnerUserId":3119718.0,"Title":"Keras' Sequential vs Functional API for Multi-Task Learning Neural Network","Body":"<p>I would like to design a neural network for a multi-task deep learning task. Within the Keras API we can either use the \"Sequential\" or \"Functional\" approach to build such a neural network. Underneath I provide the code I used to build a network using both approaches to build a network with two outputs:<\/p>\n\n<p><strong>Sequential<\/strong>  <\/p>\n\n<pre><code>seq_model = Sequential()\nseq_model.add(LSTM(32, input_shape=(10,2)))\nseq_model.add(Dense(8))\nseq_model.add(Dense(2))\nseq_model.summary()\n<\/code><\/pre>\n\n<p><strong>Functional<\/strong><\/p>\n\n<pre><code>input1 = Input(shape=(10,2))\nlay1 = LSTM(32, input_shape=(10,2))(input1)\nlay2 = Dense(8)(lay1)\nout1 = Dense(1)(lay2)\nout2 = Dense(1)(lay2)\nfunc_model = Model(inputs=input1, outputs=[out1, out2])\nfunc_model.summary()\n<\/code><\/pre>\n\n<p>When I look at both the summary outputs for the models, each of them contains identical number of trainable params: <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/gh3SJ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/gh3SJ.png\" alt=\"Sequential and Functional .summary()\"><\/a><\/p>\n\n<p>Up until now, this looks fine - however I start doubting myself when I plot both models (using <code>keras.utils.plot_model<\/code>) which results in the followings graphs: \n<a href=\"https:\/\/i.stack.imgur.com\/WCess.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/WCess.png\" alt=\"Sequential and Functional plot_model()\"><\/a><\/p>\n\n<p>Personally I do not know how to interpret these. When using a multi-task learning approach, I want all neurons (in my case 8) of the layer before the output-layer to connect to <strong>both<\/strong> output neurons. For me this clearly shows in the Functional API (where I have two <code>Dense(1)<\/code> instances), but this is not very clear from the Sequential API. Nevertheless, the amount of trainable params is identical; suggesting that also the Sequential API the last layer is fully connected to both neurons in the Dense output layer. <\/p>\n\n<p>Could anybody explain to me the differences between those two examples, or are those fully identical and result in the same neural network architecture? Also, which one would be preferred in this case? <\/p>\n\n<p>Thank you a lot in advance. <\/p>\n","answers":[{"AnswerId":"58092554","CreationDate":"2019-09-25T06:46:02.710","ParentId":null,"OwnerUserId":"5563864","Title":null,"Body":"<p>The difference between Sequential and functional keras API:<\/p>\n\n<blockquote>\n  <p>The sequential API allows you to create models layer-by-layer for most\n  problems. It is limited in that it does not allow you to create models\n  that share layers or have multiple inputs or outputs.<\/p>\n  \n  <p>the functional API allows you to create models that have a lot more\n  flexibility as you can easily define models where layers connect to\n  more than just the previous and next layers. In fact, you can connect\n  layers to (literally) any other layer. As a result, creating complex\n  networks such as siamese networks and residual networks become\n  possible.<\/p>\n<\/blockquote>\n\n<p>To answer your question:<\/p>\n\n<p>No these APIs are not the same and the number of layers is normal that are the same number.\nWhich one to use? It depends on the use you want to make of this network. What are you doing the training for? What do you want the output to be?<\/p>\n\n<p>I recommend this link to make the most of the concept.<\/p>\n\n<p><a href=\"https:\/\/jovianlin.io\/keras-models-sequential-vs-functional\/\" rel=\"nofollow noreferrer\">Sequential Models &amp; Functional Models <\/a><\/p>\n\n<p>I hope I helped you understand better.<\/p>\n"},{"AnswerId":"58098916","CreationDate":"2019-09-25T12:53:52.743","ParentId":null,"OwnerUserId":"9742432","Title":null,"Body":"<p>Both models are (in theory) equivalent, as the two output nodes do not have any interaction between them.<\/p>\n\n<p>It is just that the required outputs have a different shape <\/p>\n\n<pre><code>[(batch_size,2)]\n<\/code><\/pre>\n\n<p>vs<\/p>\n\n<pre><code>[(batch_size,),(batch_size,)]\n<\/code><\/pre>\n\n<p>and thus, the loss will be different.<\/p>\n\n<p>The total loss is averaged for the sequential model in this example, whereas it is summed up for the functional model with two outputs (at least with a default loss such as MSE).<\/p>\n\n<p>Of course, you can also adapt the functional model to be exactly equivalent to the sequential model:<\/p>\n\n<pre><code>out1 = Dense(2)(lay2)\n#out2 = Dense(1)(lay2)\nfunc_model = Model(inputs=input1, outputs=out1)\n<\/code><\/pre>\n\n<p>Maybe you will also need some activations after the Dense layers.<\/p>\n"}]}
+{"QuestionId":58092553,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-25T06:46:01.933","AcceptedAnswerId":58110335.0,"OwnerUserId":7098148.0,"Title":"verify TensorFlow is maximizing use of NVIDIA GPU","Body":"<p>I recently setup an Ubuntu 19.04 machine with a single NVIDIA GeForce GTX 1060 6GB card.  This is my first experience using all the deep learning stuff in Linux.  I followed several different instructional blogs to get what I believe is a functional CUDA\/Python environment.  My next step was to run some sample \"large jobs\" and verify I am getting the expected performance.<\/p>\n\n<p>I came across this link - <a href=\"https:\/\/medium.com\/@andriylazorenko\/tensorflow-performance-test-cpu-vs-gpu-79fcd39170c\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@andriylazorenko\/tensorflow-performance-test-cpu-vs-gpu-79fcd39170c<\/a>.  It seems to have example\/tutorial code which will exercise my system and also has some benchmark performance stats.<\/p>\n\n<p>When I run python cifar10_train.py as instructed from tutorials\/image\/cifar10, I see the following:<\/p>\n\n<pre><code>Every 1.0s: nvidia-smi                                                                                                                                                  prospector: Wed Sep 25 06:33:56 2019\n\nWed Sep 25 06:33:56 2019\n+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 418.87.00    Driver Version: 418.87.00    CUDA Version: 10.1     |\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 106...  On   | 00000000:41:00.0 Off |                  N\/A |\n| 33%   33C    P2    31W \/ 120W |   6034MiB \/  6075MiB |      6%      Default |\n+-------------------------------+----------------------+----------------------+\n\n+-----------------------------------------------------------------------------+\n| Processes:                                                       GPU Memory |\n|  GPU       PID   Type   Process name                             Usage      |\n|=============================================================================|\n|    0    101339      C   python                                      6023MiB |\n+-----------------------------------------------------------------------------+\n<\/code><\/pre>\n\n<p>So it appears the job is running on the GPU and using all available memory.  However, this test also pins all 24 cores of my Threadripper 1920X.  And the GPU temp only increases by a few degrees with the fan never really going past 33%.  Some relevant output from the python script:<\/p>\n\n<pre><code>2019-09-25 06:33:30.860034: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1326] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 5702 MB memory) -&gt; physical GPU (device: 0, name: GeForce GTX 1060 6GB, pci bus id: 0000:41:00.0, compute capability: 6.1)\n2019-09-25 06:33:30.862049: I tensorflow\/compiler\/xla\/service\/service.cc:168] XLA service 0x55f63af4ce80 executing computations on platform CUDA. Devices:\n2019-09-25 06:33:30.862071: I tensorflow\/compiler\/xla\/service\/service.cc:175]   StreamExecutor device (0): GeForce GTX 1060 6GB, Compute Capability 6.1\n2019-09-25 06:33:30.951617: W tensorflow\/compiler\/jit\/mark_for_compilation_pass.cc:1412] (One-time warning): Not using XLA:CPU for cluster because envvar TF_XLA_FLAGS=--tf_xla_cpu_global_jit was not set.  If you want XLA:CPU, either set that envvar, or use experimental_jit_scope to enable XLA:CPU.  To confirm that XLA is active, pass --vmodule=xla_compilation_cache=1 (as a proper command-line flag, not via TF_XLA_FLAGS) or set the envvar XLA_FLAGS=--xla_hlo_profile.\nI0925 06:33:31.438272 140501771114304 session_manager.py:500] Running local_init_op.\nI0925 06:33:31.456041 140501771114304 session_manager.py:502] Done running local_init_op.\nI0925 06:33:31.804769 140501771114304 basic_session_run_hooks.py:606] Saving checkpoints for 0 into \/tmp\/cifar10_train\/model.ckpt.\n2019-09-25 06:33:32.217337: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcublas.so.10\n2019-09-25 06:33:32.521519: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudnn.so.7\n2019-09-25 06:33:33.481558: step 0, loss = 4.68 (453.2 examples\/sec; 0.282 sec\/batch)\n2019-09-25 06:33:36.138869: step 10, loss = 4.59 (481.7 examples\/sec; 0.266 sec\/batch)\n2019-09-25 06:33:38.594918: step 20, loss = 4.39 (521.2 examples\/sec; 0.246 sec\/batch)\n2019-09-25 06:33:41.108821: step 30, loss = 4.49 (509.2 examples\/sec; 0.251 sec\/batch)\n2019-09-25 06:33:43.613224: step 40, loss = 4.40 (511.1 examples\/sec; 0.250 sec\/batch)\n2019-09-25 06:33:46.074761: step 50, loss = 4.36 (520.0 examples\/sec; 0.246 sec\/batch)\n2019-09-25 06:33:48.577182: step 60, loss = 4.07 (511.5 examples\/sec; 0.250 sec\/batch)\n2019-09-25 06:33:51.072065: step 70, loss = 4.19 (513.1 examples\/sec; 0.249 sec\/batch)\n2019-09-25 06:33:53.582682: step 80, loss = 4.14 (509.8 examples\/sec; 0.251 sec\/batch)\n2019-09-25 06:33:55.993107: step 90, loss = 4.14 (531.0 examples\/sec; 0.241 sec\/batch)\n<\/code><\/pre>\n\n<p>I had some install issues the first time I tried this and had to reinstall everything to get it to see the GPU.  However, from the above output and the nvidia-smi output, the test is clearly using the GPU.  However, when I look at the examples\/sec the author of the post gets when using a single 1070 card, he claims he is getting over 6,000 which is 12X what I see!<\/p>\n\n<p>My questions are:<\/p>\n\n<p>1) Am I getting the performance from this test I should expect based on my hardware (32MB RAM, 6GB VRAM)?<\/p>\n\n<p>2) Should I expect this test to also max out all 24 of my Threadripper cores and bring the CPU to almost max temp at 68C?<\/p>\n\n<p>3) Is there a better validation test I can run?<\/p>\n\n<p>I am hoping to make sure my system is setup correctly and if not what I should do to track down any bottlenecks or misconfigurations.<\/p>\n\n<p>PS - I also noticed something called Tensor RT.  How does that fit in to the performance picture?<\/p>\n\n<p>Thanks RaJa for responding.  I ran the first test from the link you sent and, just like the output from the TensorFlow load test I ran above, it shows that it recognizes and is using the GPU:<\/p>\n\n<pre><code>2019-09-25 07:55:24.868796: 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\n2019-09-25 07:55:24.989039: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1405] Found device 0 with properties:\nname: GeForce GTX 1060 6GB major: 6 minor: 1 memoryClockRate(GHz): 1.7085\npciBusID: 0000:41:00.0\ntotalMemory: 5.93GiB freeMemory: 5.86GiB\n2019-09-25 07:55:24.989066: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1484] Adding visible gpu devices: 0\n2019-09-25 07:55:25.302006: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:965] Device interconnect StreamExecutor with strength 1 edge matrix:\n2019-09-25 07:55:25.302044: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:971]      0\n2019-09-25 07:55:25.302050: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:984] 0:   N\n2019-09-25 07:55:25.302169: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1097] Created TensorFlow device (\/job:localhost\/replica:0\/task:0\/device:GPU:0 with 5640 MB memory) -&gt; physical GPU (device: 0, name: GeForce GTX 1060 6GB, pci bus id: 0000:41:00.0, compute capability: 6.1)\nDevice mapping:\n\/job:localhost\/replica:0\/task:0\/device:GPU:0 -&gt; device: 0, name: GeForce GTX 1060 6GB, pci bus id: 0000:41:00.0, compute capability: 6.1\n2019-09-25 07:55:25.360593: I tensorflow\/core\/common_runtime\/direct_session.cc:288] Device mapping:\n\/job:localhost\/replica:0\/task:0\/device:GPU:0 -&gt; device: 0, name: GeForce GTX 1060 6GB, pci bus id: 0000:41:00.0, compute capability: 6.1\n\nMatMul: (MatMul): \/job:localhost\/replica:0\/task:0\/device:GPU:0\n2019-09-25 07:55:25.361267: I tensorflow\/core\/common_runtime\/placer.cc:935] MatMul: (MatMul)\/job:localhost\/replica:0\/task:0\/device:GPU:0\na: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0\n2019-09-25 07:55:25.361289: I tensorflow\/core\/common_runtime\/placer.cc:935] a: (Const)\/job:localhost\/replica:0\/task:0\/device:GPU:0\nb: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0\n2019-09-25 07:55:25.361302: I tensorflow\/core\/common_runtime\/placer.cc:935] b: (Const)\/job:localhost\/replica:0\/task:0\/device:GPU:0\n[[22. 28.]\n [49. 64.]]\n<\/code><\/pre>\n\n<p>I'm less concerned about whether it recognizes the GPU and is using it and am more concerned about whether it is <em>maximizing<\/em> its use of the GPU.  See my 3 questions above.  I'm most concerned about the fact that I'm only seeing 1\/12th the performance that the author of the TensorFlow load test is seeing with what I believe should be a similar card (1060 vs 1070).  Am I missing something?  Should I only expect this level of performance?  What do other people see for performance using the same load test and a similar card?  Would TensorRT help?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"58110335","CreationDate":"2019-09-26T05:50:42.980","ParentId":null,"OwnerUserId":"4141279","Title":null,"Body":"<p>So I have found the reason or a better check. The official cifar10-test is somewhat bugged as there is already an issue in the tensorflow-repository.<\/p>\n\n<p>I have found a better, working example here: <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/6_MultiGPU\/multigpu_cnn.py\" rel=\"nofollow noreferrer\">multi-gpu example<\/a><\/p>\n\n<p>Just run the file with <code>python3 multigpu_cnn.py<\/code>and watch the number of samples per sec. Then you can change the number of gpus in the file and check again. <\/p>\n\n<p>My stats (4 Titan Z):<\/p>\n\n<ul>\n<li>2 GPUs -> 8800 samples\/sec<\/li>\n<li>4 GPUs -> 15500 samples\/sec<\/li>\n<\/ul>\n\n<p>Checking with <code>nvidia-smi<\/code>shows 100% load on all GPUs for the last example.<\/p>\n"}]}
+{"QuestionId":58092938,"AnswerCount":0,"Tags":"<tensorflow><keras><tf.keras>","CreationDate":"2019-09-25T07:11:38.620","AcceptedAnswerId":null,"OwnerUserId":11939660.0,"Title":"tf keras, custom loss function that require multiple network outputs as inputs","Body":"<p>I've recently decided to try tf2.0, especially the high level keras API. I am not too familiar with keras either, have mainly been using tensorflow before. However, I cannot find a good way of using custom loss functions. Take a simple example, let's suppose we have a multi-input, multi-output model defined as follow:<\/p>\n\n<pre><code># define model\n... your network definition\n...\n\nmodel = tf.keras.Model(inputs=[input_1, input_2], outputs=[output_1, output_2, output_3])\n<\/code><\/pre>\n\n<p>Let's suppose I have two custom loss functions: <code>my_loss_1<\/code> requires <code>output_1, output_2<\/code> and <code>targets_1<\/code> in order to compute the loss. Similarly <code>my_loss_2<\/code> requires <code>output_2, output_3<\/code> and <code>targets_2<\/code>. My questions are:<\/p>\n\n<ol>\n<li>Is there a way to do this by sticking to the high level API?<\/li>\n<li>How can I pass training data to the model? If I have constructed a dataset as<\/li>\n<\/ol>\n\n<pre><code>train_ds = tf.data.Dataset.zip((input_ds_1, input_ds_2), (target_ds_1, target_ds_2))\n<\/code><\/pre>\n\n<p>and pass it to <code>model.fit<\/code>, then the model will assume to compute <code>loss_1<\/code> using <code>output_1<\/code> and <code>target_ds_1<\/code>, which is not what I intended.<\/p>\n\n<ol start=\"3\">\n<li><p>I'm not even clear how the data correspondence rule is when using tf.data with keras, its not explained anywhere. My guess is that a training dataset should be a tuple of <code>(inputs, targets)<\/code>. If your network takes multiple inputs, then the inputs is another tuple of <code>(input_1, input_2, ...)<\/code>, same with targets. Is this correct?<\/p><\/li>\n<li><p>Is there a way to reuse the data passed to the model? Suppose now both <code>my_loss_1<\/code> and <code>my_loss_2<\/code> require <code>target_1<\/code>, do I have to change my <code>train_ds<\/code> to contain duplicates? <\/p><\/li>\n<\/ol>\n\n<pre><code>train_ds = tf.data.Dataset.zip((input_ds_1, input_ds_2), (target_ds_1, (target_ds_1, target_ds_2)))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58093506,"AnswerCount":2,"Tags":"<python><numpy><machine-learning><keras>","CreationDate":"2019-09-25T07:42:13.817","AcceptedAnswerId":58104522.0,"OwnerUserId":10766295.0,"Title":"Generator not being recognized when passing validation data to .fit in Keras Sequential","Body":"<p>The exact error: <\/p>\n\n<blockquote>\n  <p>ValueError: When passing validation_data, it must contain 2 (x_val,\n  y_val) or 3 (x_val, y_val, val_sample_weights) items, however it\n  contains 39 items<\/p>\n<\/blockquote>\n\n<p>I literally cannot find that error anywhere except the source code.<\/p>\n\n<pre><code>model.fit(  train_x\n            , train_y\n            , epochs=1\n            , validation_data=validation_data_flow\n            , callbacks=[checkpointer]\n        )\n<\/code><\/pre>\n\n<p>validation_data is a DirectoryIterator, by flow_from_directory<\/p>\n\n<pre><code>validation_data_flow = ImageDataGenerator().flow_from_directory(\n        validation_data_dir,\n        target_size = (img_width, img_height),\n        batch_size = batch_size,\n        class_mode = 'categorical')\n<\/code><\/pre>\n","answers":[{"AnswerId":"58105591","CreationDate":"2019-09-25T19:47:03.273","ParentId":null,"OwnerUserId":"3472735","Title":null,"Body":"<p>It's not exactly true that you have to train and validate on the same kind data, such as array or generator, but you can't do that with the same function.<\/p>\n\n<p>You could train on your array, and validate on your generator but this would take 2 different function calls, and means you won't get validation metrics after every epoch if you're using multiple epoch. You can work around this, with something like:<\/p>\n\n<pre><code>for i in range(epochs):\n    model.fit(train_x, train_y, epochs=1, callbacks=[checkpointer])\n    loss = model.evaluate_generator(validation_data_flow)\n\n    print(\"Validation loss for epoch %s was %s\" % (i, loss))\n<\/code><\/pre>\n"},{"AnswerId":"58104522","CreationDate":"2019-09-25T18:26:55.660","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>Validation and training data need to be the same type, either both generators or both ndarrays. To fix this, you need to convert one to the other type. Look at <a href=\"https:\/\/stackoverflow.com\/q\/367565\/2846923\">this<\/a> answer for how to convert a generator to an ndarray. To convert an ndarray to a generator, use <a href=\"https:\/\/keras.io\/preprocessing\/image\/#flow\" rel=\"nofollow noreferrer\"><code>ImageDataGenerator.flow()<\/code><\/a>.<\/p>\n"}]}
+{"QuestionId":58093761,"AnswerCount":0,"Tags":"<keras><deep-learning><batch-normalization>","CreationDate":"2019-09-25T07:56:05.960","AcceptedAnswerId":null,"OwnerUserId":10452691.0,"Title":"BatchNormalization makes poor result","Body":"<p>In dcgan with keras, I have source which is working well.<\/p>\n\n<p>But I found that when I add the batch normalization to generator and discriminator,\nit doesn't work.<\/p>\n\n<p>If I remove them, it works. again.<\/p>\n\n<p>can someone explain why?<\/p>\n","answers":[]}
+{"QuestionId":58093766,"AnswerCount":0,"Tags":"<python><memory><torch>","CreationDate":"2019-09-25T07:56:25.357","AcceptedAnswerId":null,"OwnerUserId":12117431.0,"Title":"Use memory_profiler to Increment a lot, the Mem usage is unchanged","Body":"<p>Please note 128, 129, 130 lines   <\/p>\n\n<p>I am using memory_profiler but I have some doubts about the memory situation I observed. The increase is very large, but the occupancy has not changed. This problem is caused by the memory_profiler problem or the internal memory mechanism of python.<\/p>\n\n<pre><code>line  Mem usage    Increment    Line Contents\n================================================\n 58  675.719 MiB  675.719 MiB   @profile\n 59                             def train(dataset, embedding, tokenizer, entity_linker, min_count, max_word_length, max_entity_length, batch_size,\n 60                                       patience, optimizer, learning_rate, weight_decay, warmup_epochs, dropout_prob, use_gpu, use_word):\n 61  675.719 MiB    0.000 MiB       batch_size = int(batch_size)\n 62  675.719 MiB    0.000 MiB       learning_rate = float(learning_rate)\n 63  675.719 MiB    0.000 MiB       weight_decay = float(weight_decay)\n 64  675.719 MiB    0.000 MiB       if use_gpu:\n 65                                     device = torch.device('cuda')\n 66                                 else:\n 67  675.719 MiB    0.000 MiB           device = torch.device('cpu')\n 68\n 69  947.828 MiB  272.109 MiB       data = generate_features(dataset, tokenizer, entity_linker, min_count, max_word_length, max_entity_length)\n 70  947.828 MiB    0.000 MiB       word_vocab = data['word_vocab']\n 71  947.828 MiB    0.000 MiB       entity_vocab = data['entity_vocab']\n 72\n 73  947.828 MiB    0.000 MiB       train_data_loader = DataLoader(data['train'], shuffle=True, batch_size=batch_size)\n 74  947.828 MiB    0.000 MiB       dev_data_loader = DataLoader(data['dev'], shuffle=False, batch_size=batch_size)\n 75\n 76  947.828 MiB    0.000 MiB       dim_size = embedding.syn0.shape[1]\n 77  976.418 MiB   28.590 MiB       word_embedding = np.random.uniform(low=-0.05, high=0.05, size=(len(word_vocab), dim_size))\n 78  976.418 MiB    0.000 MiB       word_embedding[0] = np.zeros(dim_size)\n 79 1043.434 MiB    0.000 MiB       for word, index in word_vocab.items():\n 80 1043.434 MiB    0.000 MiB           try:\n 81 1043.434 MiB    0.066 MiB               word_embedding[index] = embedding.get_word_vector(word)\n 82 1043.434 MiB    0.000 MiB           except KeyError:\n 83 1043.434 MiB    0.000 MiB               continue\n 84 1099.840 MiB   56.406 MiB       entity_embedding = np.random.uniform(low=-0.05, high=0.05, size=(len(entity_vocab), dim_size))\n 85 1099.840 MiB    0.000 MiB       entity_embedding[0] = np.zeros(dim_size)\n 86 1213.281 MiB    0.000 MiB       for entity, index in entity_vocab.items():\n 87 1213.281 MiB    0.000 MiB           try:\n 88 1213.281 MiB    0.027 MiB               entity_embedding[index] = embedding.get_entity_vector(entity)\n 89 1213.281 MiB    0.000 MiB           except KeyError:\n 90 1213.281 MiB    0.000 MiB               continue\n 91\n 92 1270.773 MiB   57.492 MiB       model = NABoE(word_embedding, entity_embedding, len(dataset.label_names), dropout_prob, use_word)\n 93 1270.773 MiB    0.000 MiB       if optimizer == 'adamW':\n 94                                     optimizer = AdamW(model.parameters(), lr=learning_rate, weight_decay=weight_decay,\n 95                                                       warmup=warmup_epochs * len(train_data_loader))\n 96 1270.773 MiB    0.000 MiB       elif optimizer == 'adam':\n 97 1270.773 MiB    0.000 MiB           optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)\n 98                                 elif optimizer == 'adagrad':\n 99                                     optimizer = torch.optim.Adagrad(model.parameters(), lr=learning_rate)\n100                                 elif optimizer == 'sgd':\n101                                     optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)\n102                                 elif optimizer == 'radam':\n103                                     optimizer = RAdam(model.parameters(), lr=learning_rate, weight_decay=weight_decay)\n104                                 elif optimizer == 'ranger':\n105                                     optimizer = Ranger(model.parameters(), lr=learning_rate, weight_decay=weight_decay)\n106 1270.809 MiB    0.035 MiB       model.to(device)\n107\n108 1270.809 MiB    0.000 MiB       epoch = 0\n109 1270.809 MiB    0.000 MiB       best_val_acc = 0.0\n110 1270.809 MiB    0.000 MiB       best_weights = None\n111 1270.809 MiB    0.000 MiB       num_epochs_without_improvement = 0\n112\n113 1270.809 MiB    0.000 MiB       total_memory = 0\n114\n115 1270.809 MiB    0.000 MiB       while True:\n116 1658.176 MiB    0.004 MiB           with tqdm(train_data_loader) as pbar:\n117 1658.176 MiB    0.000 MiB               model.train()\n118 1658.176 MiB    0.285 MiB               for batch in pbar:\n119\n120 1658.176 MiB    0.000 MiB                   memory = cpu_stats()\n121                                             # disk = psutil.disk_usage('\/').percent\n122                                             # cpu = psutil.cpu_percent(0)\n123                                             # cpu_kernel = psutil.cpu_percent(percpu=True)\n124\n125 1658.176 MiB    0.000 MiB                   args = {k: v.to(device) for k, v in batch.items() if k != 'label'}\n126 1892.559 MiB  263.688 MiB                   logits = model(**args)\n127 1892.559 MiB    0.102 MiB                   loss = F.cross_entropy(logits, batch['label'].to(device))\n128 1658.176 MiB    0.000 MiB                   loss.backward()\n129 1658.176 MiB  127.637 MiB                   optimizer.step()\n130 1658.176 MiB    0.008 MiB                   model.zero_grad()\n131 1658.176 MiB    0.000 MiB                   total_memory += memory\n132 1658.176 MiB    0.000 MiB                   pbar.set_description(f'epoch: {epoch} loss: {loss.item():.8f} The Memory Usage: {memory}'\n133                                                                  f' Total Memory Usage: {total_memory}')\n134\n135 1658.176 MiB    0.000 MiB           epoch += 1\n136 1658.176 MiB    0.312 MiB           val_acc = evaluate(model, dev_data_loader, device, 'dev')[0]\n137 1658.176 MiB    0.000 MiB           if val_acc &gt; best_val_acc:\n138 1658.176 MiB    0.000 MiB               best_val_acc = val_acc\n139 1658.176 MiB    0.004 MiB               best_weights = {k: v.to('cpu').clone() for k, v in model.state_dict().items()}\n140 1658.176 MiB    0.000 MiB               num_epochs_without_improvement = 0\n141                                     else:\n142 1653.926 MiB    0.000 MiB               num_epochs_without_improvement += 1\n143\n144 1658.176 MiB    0.000 MiB           if num_epochs_without_improvement &gt;= patience:\n145 1653.859 MiB    0.000 MiB               model.load_state_dict(best_weights)\n146 1653.859 MiB    0.000 MiB               break\n147\n148 1653.859 MiB    0.000 MiB       test_data_loader = DataLoader(data['test'], shuffle=False, batch_size=batch_size)   149 1653.863 MiB    0.004 MiB       return evaluate(model, test_data_loader, device, 'test')\n<\/code><\/pre>\n\n<p>129 lines increase 127M occupancy, but the memory footprint has not changed, even Python internal recycling mechanism is not exactly the same<\/p>\n","answers":[]}
+{"QuestionId":58094212,"AnswerCount":1,"Tags":"<python><keras><neural-network><deep-learning><sequential>","CreationDate":"2019-09-25T08:25:56.463","AcceptedAnswerId":null,"OwnerUserId":11945002.0,"Title":"Drop inactive features in Keras","Body":"<p>I'm building a Sequential NN model in Keras for binary classification. The training data has about 600,000 rows and 2,000 features, so every epoch and every layer is very time consuming. I believe many of the features are not relevant to the model, and can be dropped altogether, to make the model thinner, so it it would be faster to work with.\nI run a simple model with one hidden-layer of 200 neurons. How can I tell which of the features (which are actually the nodes in the input layer) are meaningless, so I could drop them from the data set and re run the model without them?<\/p>\n","answers":[{"AnswerId":"58096024","CreationDate":"2019-09-25T10:12:13.713","ParentId":null,"OwnerUserId":"5270291","Title":null,"Body":"<p>There is a very big topic in machine learning called <code>feature selection<\/code>. Though, neural networks are considered to automatically choose the best features for the problem, to an extent, by using their weights, to either consider more or less some of them. Neural networks also need a lot of experience to be tuned correctly. I would definitely suggest you to increase the layers of the network, because you have a lot of data and features and use <code>l1 regularisation<\/code>, in order to get <strong>sparse<\/strong> weights and exclude most of the features. Also, these information are indicative, since I do not know anything about your dataset and your network architecture. At last, I would suggest you to study more about the basics of machine learning and then continue learning about neural networks, before practicing with real data.<\/p>\n"}]}
+{"QuestionId":58095627,"AnswerCount":1,"Tags":"<python-3.x><gpu><pytorch><lstm>","CreationDate":"2019-09-25T09:47:15.553","AcceptedAnswerId":58104513.0,"OwnerUserId":12080262.0,"Title":"How to fix 'Input and hidden tensors are not at the same device' in pytorch","Body":"<p>When I want to put the model on the GPU, there is an error!\nIt said the Inputs is on GPU, but the hidden state is on CPU.\nHowever, all of them had been put on the GPU.\nI use <\/p>\n\n<pre><code>for m in model.parameters():\n    print(m.device) #return cuda:0\n<\/code><\/pre>\n\n<p>to see all of the state on the model is on the GPU device.\nThe error is \"RuntimeError: Input and hidden tensors are not at the same device, found input tensor at cuda:0 and hidden tensor at cpu\"<\/p>\n\n<p>Windows 10 server\nPytorch 1.2.0 + cuda 9.2\ncuda 9.2\ncudnn 7.6.3 for cuda 9.2<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>if torch.cuda.is_available():\n\n    model = model.cuda()\n\nif torch.cuda.is_available():\n    `test= test.cuda() #test is the Input`\n<\/code><\/pre>\n","answers":[{"AnswerId":"58104513","CreationDate":"2019-09-25T18:25:58.590","ParentId":null,"OwnerUserId":"7872285","Title":null,"Body":"<p>You need to move the model , the inputs and the targets to Cuda <\/p>\n\n<pre><code>if(torch.cuda.is_available()):\n\n   rnn.cuda()# rnn is your model \n   inputs = inputs.cuda() \n   target = target.cuda()\n<\/code><\/pre>\n"}]}
+{"QuestionId":58095727,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-lite><onnx>","CreationDate":"2019-09-25T09:53:26.650","AcceptedAnswerId":null,"OwnerUserId":10538466.0,"Title":"Convertion of tflite model to ONNX model","Body":"<p>I have a tflite model and i want to convert it into tensorflow or keras or ONNX format. Is there a way?<\/p>\n\n<p>I can import it using tf interpreter and run it on python. However i want to convert it into one of the above formats.<\/p>\n\n<pre><code>import tensorflow as tf\ninterpreter = tf.lite.Interpreter(model_path=\"conv_actions_frozen.tflite\")\ntensors = interpreter.get_tensor_details()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58096095,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-25T10:15:21.383","AcceptedAnswerId":58097046.0,"OwnerUserId":12118244.0,"Title":"How does tf.audio.decode_wav get its contents?","Body":"<p>I'm trying to pull some audio files into Tensorflow by using <code>tf.audio.decode_wav<\/code>.<\/p>\n\n<p>I can see someone is looking into providing more info in the docs, but does anyone have any examples of how this should work?<\/p>\n\n<pre><code>tf.audio.decode_wav(\n contents,\n desired_channels=-1,\n desired_samples=-1,\n name=None\n)\n<\/code><\/pre>\n\n<p><strong>Args:<\/strong><\/p>\n\n<ul>\n<li>contents: A Tensor of type string. The WAV-encoded audio, usually from a file.<\/li>\n<li>desired_channels: An optional int. Defaults to -1. Number of sample channels wanted.<\/li>\n<li>desired_samples: An optional int. Defaults to -1. Length of audio requested.<\/li>\n<li>name: A name for the operation (optional).<\/li>\n<\/ul>\n\n<p>I'm guessing the contents is a tensor which has already been pulled from a file rather than a path?<\/p>\n","answers":[{"AnswerId":"58097046","CreationDate":"2019-09-25T11:11:36.727","ParentId":null,"OwnerUserId":"5786339","Title":null,"Body":"<p>You're right, <code>tf.audio.decode_wav()<\/code> requires a tensor. You can provide one with \n<code>tf.io.read_file()<\/code> which reads wav file into tensor of type string.<\/p>\n\n<pre><code>raw_audio = tf.io.read_file(filename)\nwaveform = tf.audio.decode_wav(raw_audio)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58096102,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-09-25T10:15:55.423","AcceptedAnswerId":null,"OwnerUserId":12118247.0,"Title":"How to load images and mask for unet ,with keras imagedatagenerator","Body":"<p>I need to load images and mask for unet model,i dont wanna use cv2 or any libaraires to load images.I wanna use easy way to load up these images.\nMy dataset strucuture<\/p>\n\n<pre><code>data\/train\n    1.png\n    2.png\n\ndata\/mask\n    1_mask.png\n<\/code><\/pre>\n\n<p>while googling i found keras docs for images and mask code.but it need images and mask stored in array.Or any other method to load up images and mask.<\/p>\n\n<p>and how to work if below is case.<\/p>\n\n<pre><code>\/train:\n      \/dir_name\/\n                \/images\n                \/masks\n      \/dir_name2\/\n                \/images\n                \/masks\n      \/dir_name3\/\n                \/images\n                \/masks\n      \/dir_name4\/\n                \/images\n                \/masks\n<\/code><\/pre>\n","answers":[{"AnswerId":"58126772","CreationDate":"2019-09-27T01:57:22.570","ParentId":null,"OwnerUserId":"12118247","Title":null,"Body":"<p>check out:<\/p>\n\n<p>for first type of dataset<\/p>\n\n<p>Click <a href=\"https:\/\/stackoverflow.com\/questions\/58050113\/imagedatagenerator-for-semantic-segmentation\/58050474#58050474\">here<\/a>!\nusing keras imagedatagenrator code and dataset structure.<\/p>\n\n<p>for second type:\nclick <a href=\"https:\/\/stackoverflow.com\/questions\/47645797\/keras-model-unet-image-segmentation\">here<\/a><\/p>\n"}]}
+{"QuestionId":58096219,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-25T10:21:43.983","AcceptedAnswerId":58096299.0,"OwnerUserId":5048010.0,"Title":"Save a tensorflow model after a fixed training time","Body":"<p>I'm training a model on a server that allows me only one hour of computation: At the end of that time, it will simply kill my job. I would like tensorflow to save the results of its training after, say, 58 minutes of training, no matter what is the current state. I'm OK with it saving the status at the last completed epoch, I just want to have an idea what's going on. How can I do that?<\/p>\n","answers":[{"AnswerId":"58096299","CreationDate":"2019-09-25T10:26:09.420","ParentId":null,"OwnerUserId":"9361998","Title":null,"Body":"<p>Of course, you can define a callback function delegated to stop the training phase.<\/p>\n\n<p>You can have a look here for further information:<br>\n<a href=\"https:\/\/towardsdatascience.com\/neural-network-with-tensorflow-how-to-stop-training-using-callback-5c8d575c18a9\" rel=\"nofollow noreferrer\">https:\/\/towardsdatascience.com\/neural-network-with-tensorflow-how-to-stop-training-using-callback-5c8d575c18a9<\/a><\/p>\n\n<p>In this example, is created a callback function in order to stop the training phase when the 'ACCURACY' exceeds the threshold. You can modify the function in order to make a time computation in order to verify the elapsed time. <\/p>\n\n<hr>\n\n<p>This is a working piece of code:<\/p>\n\n<pre><code>class TimeOut(Callback):\n    def __init__(self, t0, timeout):\n        super().__init__()\n        self.t0 = t0\n        self.timeout = timeout  # time in minutes\n\n    def on_train_batch_end(self, batch, logs=None):\n        if time.time() - self.t0 &gt; self.timeout * 60:  # 58 minutes\n            print(f\"\\nReached {(time.time() - self.t0) \/ 60:.3f} minutes of training, stopping\")\n            self.model.stop_training = True\n\ncallbacks = [TimeOut(t0=time.time(), timeout=58)]\n<\/code><\/pre>\n"}]}
+{"QuestionId":58096806,"AnswerCount":0,"Tags":"<android-studio><tensorflow><arcore>","CreationDate":"2019-09-25T10:55:10.843","AcceptedAnswerId":null,"OwnerUserId":12053540.0,"Title":"ARCore compatibility with tensorflow in android studio?","Body":"<p>i'm trying to make a project in android studio that detect object and then, from an object detect, display a 3D image.\nI used tensorflow to detect the object in the world and ARCore to display 3D image, but i'm finding some difficulties, so i would know if it possible realize this idea of project with this two technologies.<\/p>\n","answers":[]}
+{"QuestionId":58097348,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-25T11:29:09.890","AcceptedAnswerId":null,"OwnerUserId":1179925.0,"Title":"What's the differences between tf.GraphKeys.GLOBAL_VARIABLES and tf.GraphKeys.TRAINABLE_VARIABLES in tensorflow?","Body":"<p>From <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/GraphKeys\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/GraphKeys<\/a><\/p>\n\n<blockquote>\n  <p>GLOBAL_VARIABLES: the default collection of Variable objects, shared\n  across distributed environment (model variables are subset of these).\n  See tf.compat.v1.global_variables for more details. Commonly, all\n  TRAINABLE_VARIABLES variables will be in MODEL_VARIABLES, and all\n  MODEL_VARIABLES variables will be in GLOBAL_VARIABLES<\/p>\n  \n  <p>TRAINABLE_VARIABLES: the subset of Variable objects that will be\n  trained by an optimizer. See tf.compat.v1.trainable_variables for more\n  details.<\/p>\n<\/blockquote>\n\n<p>So as I understand <code>TRAINABLE_VARIABLES<\/code> is subset of <code>GLOBAL_VARIABLES<\/code>, so what else <code>GLOBAL_VARIABLES<\/code> contain?<\/p>\n\n<p>Also for this simple example statement <code>Commonly, all TRAINABLE_VARIABLES variables will be in MODEL_VARIABLES, and all MODEL_VARIABLES variables will be in GLOBAL_VARIABLES<\/code> don't hold:<\/p>\n\n<pre><code>IMAGE_HEIGHT = 5\nIMAGE_WIDTH = 5\nwith tf.Graph().as_default():\n    with tf.variable_scope('my_scope', reuse=tf.AUTO_REUSE):\n        x_ph = tf.placeholder(\n                dtype=tf.float32,\n                shape=[None, IMAGE_HEIGHT, IMAGE_WIDTH, 3],\n                name='input'\n            )\n\n        x_tf = tf.layers.conv2d(x_ph, 32, 1, 1, padding='valid')\n\n    with tf.Session() as sess:\n        sess.run(tf.global_variables_initializer())\n\n        x_np = np.random.rand(1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)\n\n        out_np = sess.run(x_tf, {x_ph:x_np})\n\n        print('out_np.shape', out_np.shape)\n\n        print('-'*60)\n        global_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)\n        print('len(global_vars)', len(global_vars))\n        print('global_vars params:', sum([np.prod(var.shape) for var in global_vars]))\n        print(global_vars)\n\n        print('-'*60)\n        model_vars = tf.get_collection(tf.GraphKeys.MODEL_VARIABLES)\n        print('len(model_vars)', len(model_vars))\n        print('model_vars params:', sum([np.prod(var.shape) for var in model_vars]))\n        print(model_vars)\n\n        print('-'*60)\n        trainable_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)\n        print('len(trainable_vars)', len(trainable_vars))\n        print('trainable_vars params:', sum([np.prod(var.shape) for var in trainable_vars]))\n        print(trainable_vars)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>out_np.shape (1, 5, 5, 32)\n------------------------------------------------------------\nlen(global_vars) 2\nglobal_vars params: 128\n[&lt;tf.Variable 'my_scope\/conv2d\/kernel:0' shape=(1, 1, 3, 32) dtype=float32_ref&gt;, &lt;tf.Variable 'my_scope\/conv2d\/bias:0' shape=(32,) dtype=float32_ref&gt;]\n------------------------------------------------------------\nlen(model_vars) 0\nmodel_vars params: 0\n[]\n------------------------------------------------------------\nlen(trainable_vars) 2\ntrainable_vars params: 128\n[&lt;tf.Variable 'my_scope\/conv2d\/kernel:0' shape=(1, 1, 3, 32) dtype=float32_ref&gt;, &lt;tf.Variable 'my_scope\/conv2d\/bias:0' shape=(32,) dtype=float32_ref&gt;]\n<\/code><\/pre>\n\n<p>So questions are:<\/p>\n\n<ol>\n<li><p>Why <code>Commonly, all TRAINABLE_VARIABLES variables will be in MODEL_VARIABLES, and all MODEL_VARIABLES variables will be in GLOBAL_VARIABLES<\/code> don't hold for this example.<\/p><\/li>\n<li><p>What additional variables <code>GLOBAL_VARIABLES<\/code> contain besides <code>TRAINABLE_VARIABLES<\/code>? Is it true that <code>TRAINABLE_VARIABLES<\/code> will be always subset of <code>GLOBAL_VARIABLES<\/code> or they can just intersect partially?<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"58097882","CreationDate":"2019-09-25T12:01:16.797","ParentId":null,"OwnerUserId":"3214872","Title":null,"Body":"<p><strong>Note:<\/strong> All of this applies to TF version <strong>1<\/strong> only, as all variable collections have been deprecated and (IIRC) won't be in TF v2.<\/p>\n\n<h3>Starting from question 2:<\/h3>\n\n<blockquote>\n  <p>What additional variables GLOBAL_VARIABLES contain besides TRAINABLE_VARIABLES?<\/p>\n<\/blockquote>\n\n<p><code>global_step<\/code>, for example, is a global variable which is not trainable.\nIt's a variable, because you update it at every step, it's not trainable because it's not part of the optimization process (as in, it's not a weight\/bias that is altered in order to minimize the loss).<\/p>\n\n<blockquote>\n  <p>Is it true that TRAINABLE_VARIABLES will be always subset of GLOBAL_VARIABLES or they can just intersect partially?<\/p>\n<\/blockquote>\n\n<p>In principle, the two groups can just intersect partially, although this would be very odd. An example I can think of would be something like a custom distributed training environment where each machine has its own optimizer and where some of the trainable variables are defined as local variables (i.e., each machine has its own copy and these copies are not kept in sync). Why would you do that? No clue. But in principle, it's possible.<\/p>\n\n<h3>Then about question 1:<\/h3>\n\n<p>I believe the statement you quote is missing an important clarification: <strong>you<\/strong> need to put the variables in the <code>MODEL_VARIABLES<\/code> collection, by default variables are only added to the <code>GLOBAL_VARIABLES<\/code> collection and, if <code>trainable=True<\/code> also to <code>TRAINABLE_VARIABLS<\/code>. TF can't know on its own which variables are necessary for inference and which are only for training (e.g., networks with auxiliary heads used for training only), so it's left to the network architect. As a side note, I've never seen that collection used anywhere and I believe it's currently unused.<\/p>\n"}]}
+{"QuestionId":58097924,"AnswerCount":1,"Tags":"<python><neural-network><pytorch><torch>","CreationDate":"2019-09-25T12:03:17.917","AcceptedAnswerId":58098215.0,"OwnerUserId":7648650.0,"Title":"How to create variable names in loop for layers in pytorch neural network","Body":"<p>I am implementing a straightforward feedforward neural newtork in PyTorch. However I am wondern if theres a nicer way to add a flexible amount of layer to the network? Maybe by naming them during a loop, but i heard thats impossible?<\/p>\n\n<p>Currently I am doing it like this<\/p>\n\n<pre><code>import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nclass Net(nn.Module):\n\n    def __init__(self, input_dim, output_dim, hidden_dim):\n        super(Net, self).__init__()\n        self.input_dim = input_dim\n        self.output_dim = output_dim\n        self.hidden_dim = hidden_dim\n        self.layer_dim = len(hidden_dim)\n        self.fc1 = nn.Linear(self.input_dim, self.hidden_dim[0])\n        i = 1\n        if self.layer_dim &gt; i:\n            self.fc2 = nn.Linear(self.hidden_dim[i-1], self.hidden_dim[i])\n            i += 1\n        if self.layer_dim &gt; i:\n            self.fc3 = nn.Linear(self.hidden_dim[i-1], self.hidden_dim[i])\n            i += 1\n        if self.layer_dim &gt; i:\n            self.fc4 = nn.Linear(self.hidden_dim[i-1], self.hidden_dim[i])\n            i += 1\n        if self.layer_dim &gt; i:\n            self.fc5 = nn.Linear(self.hidden_dim[i-1], self.hidden_dim[i])\n            i += 1\n        if self.layer_dim &gt; i:\n            self.fc6 = nn.Linear(self.hidden_dim[i-1], self.hidden_dim[i])\n            i += 1\n        if self.layer_dim &gt; i:\n            self.fc7 = nn.Linear(self.hidden_dim[i-1], self.hidden_dim[i])\n            i += 1\n        if self.layer_dim &gt; i:\n            self.fc8 = nn.Linear(self.hidden_dim[i-1], self.hidden_dim[i])\n            i += 1\n        self.fcn = nn.Linear(self.hidden_dim[-1], self.output_dim)\n\n    def forward(self, x):\n        # Max pooling over a (2, 2) window\n        x = F.relu(self.fc1(x))\n        i = 1\n        if self.layer_dim &gt; i:\n            x = F.relu(self.fc2(x))\n            i += 1\n        if self.layer_dim &gt; i:\n            x = F.relu(self.fc3(x))\n            i += 1\n        if self.layer_dim &gt; i:\n            x = F.relu(self.fc4(x))\n            i += 1\n        if self.layer_dim &gt; i:\n            x = F.relu(self.fc5(x))\n            i += 1\n        if self.layer_dim &gt; i:\n            x = F.relu(self.fc6(x))\n            i += 1\n        if self.layer_dim &gt; i:\n            x = F.relu(self.fc7(x))\n            i += 1\n        if self.layer_dim &gt; i:\n            x = F.relu(self.fc8(x))\n            i += 1\n        x = F.softmax(self.fcn(x))\n        return x\n<\/code><\/pre>\n","answers":[{"AnswerId":"58098215","CreationDate":"2019-09-25T12:18:13.740","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can put your layers in a <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.ModuleList\" rel=\"nofollow noreferrer\"><code>ModuleList<\/code><\/a> container:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nclass Net(nn.Module):\n\n    def __init__(self, input_dim, output_dim, hidden_dim):\n        super(Net, self).__init__()\n        self.input_dim = input_dim\n        self.output_dim = output_dim\n        self.hidden_dim = hidden_dim\n        current_dim = input_dim\n        self.layers = nn.ModuleList()\n        for hdim in hidden_dim:\n            self.layers.append(nn.Linear(current_dim, hdim))\n            current_dim = hdim\n        self.layers.append(nn.Linear(current_dim, output_dim))\n\n    def forward(self, x):\n        for layer in self.layers[:-1]:\n            x = F.relu(layer(x))\n        out = F.softmax(self.layers[-1](x))\n        return out    \n<\/code><\/pre>\n\n<p>It is very important to use <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#containers\" rel=\"nofollow noreferrer\">pytorch Containers<\/a> for the layers, and not just a simple python lists. Please see <a href=\"https:\/\/stackoverflow.com\/a\/54684415\/1714410\">this answer<\/a> to know why.<\/p>\n"}]}
+{"QuestionId":58099106,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><gpu><cudnn>","CreationDate":"2019-09-25T13:03:48.100","AcceptedAnswerId":null,"OwnerUserId":6420063.0,"Title":"InvalidArgumentError: No OpKernel was registered to support Op 'CudnnRNN'","Body":"<p>I followed the steps provided here : <a href=\"https:\/\/www.tensorflow.org\/install\/gpu\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/gpu<\/a>\nto install tensorflow that works with my gpu, and use the cuDNN SDK\nbut for some reasons I still can not use it, and am still getting this error:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: No OpKernel was registered to support Op 'CudnnRNN' used by node sequential\/cu_dnnlstm\/CudnnRNN (defined at cudrnn.py:27) with these attrs: [input_mode=\"linear_input\", T=DT_FLOAT, direction=\"unidirectional\", rnn_mode=\"lstm\", seed2=0, is_training=true, seed=0, dropout=0]\nRegistered devices: [CPU, XLA_CPU, XLA_GPU]\nRegistered kernels:\n  device='GPU'; T in [DT_DOUBLE]\n  device='GPU'; T in [DT_FLOAT]\n  device='GPU'; T in [DT_HALF]\n<\/code><\/pre>\n\n<p>even though the GPU is registered but tensorflow claims that they can not find an OpKernel.<\/p>\n\n<p>I am using:<\/p>\n\n<p>tensoflow-gpu 1.14.0<\/p>\n\n<p>Cuda 10.0<\/p>\n\n<p>Ubuntu 18.04<\/p>\n\n<p>P.S: I can run other code using tensorflow and by checking nvidia-smi I can see that the GPUs are getting used actually.<\/p>\n\n<p>can anyone please help me knowing why I am getting this problem ?<\/p>\n\n<p>this is the code I am trying to run:<\/p>\n\n<pre><code>model = tf.keras.Sequential([\n    tf.keras.layers.CuDNNLSTM(\n        3,\n        return_sequences=False),\n    tf.keras.layers.Dense(1)\n])\n\nmodel.compile(\n    optimizer=tf.keras.optimizers.Adam(0.1, decay=0.003), \n    loss=tf.keras.losses.mse)\n\nmodel.fit(train_x, train_y, epochs=100, validation_split=0, shuffle=False)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58538483","CreationDate":"2019-10-24T09:37:59.800","ParentId":null,"OwnerUserId":"12198627","Title":null,"Body":"<p>Your Registered devices showing <code>XLA_GPU<\/code> which is optimized compiler but not same as <code>GPU<\/code> which the tensorflow library you are using expects,<\/p>\n\n<p>You can check couple of ways to resolve,<\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.tensorflow.org\/xla#enable_xla_for_tensorflow_models\" rel=\"nofollow noreferrer\">Enable XLA for TF models<\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/install\/source#tested_build_configurations\" rel=\"nofollow noreferrer\">Check for build configurations of tensorflow for your device<\/a><\/li>\n<\/ul>\n"}]}
+{"QuestionId":58099153,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-25T13:06:10.900","AcceptedAnswerId":null,"OwnerUserId":1179925.0,"Title":"Tensorflow: Is it possible to replace part of network with placeholder?","Body":"<p>At training stage I have network A(Embedder) and network B(Generator) which are chained and model A->B is trained end-to-end.<\/p>\n\n<p>Is it possible at test time to replace A network (i.e. input to B network) with placeholder or with constant?<\/p>\n","answers":[]}
+{"QuestionId":58099264,"AnswerCount":0,"Tags":"<tensorflow><keras><onnx>","CreationDate":"2019-09-25T13:12:26.350","AcceptedAnswerId":null,"OwnerUserId":8838476.0,"Title":"How to convert Tensorflow v2 models to Onnx","Body":"<p>I am trying to convert my TF 2.0 RC1 models to onnx (using keras2onnx or tf2onnx). When I execute this code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import keras2onnx\n<\/code><\/pre>\n\n<p>or this code<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tf2onnx\n<\/code><\/pre>\n\n<p>I get the error:<\/p>\n\n<pre><code>ModuleNotFoundError: No module named 'tensorflow.tools.graph_transforms'\n<\/code><\/pre>\n\n<p>How can I resolve this error or otherwise convert my Tensorflow v2.0 models to onnx?<\/p>\n","answers":[]}
+{"QuestionId":58099533,"AnswerCount":1,"Tags":"<python><keras><gradient>","CreationDate":"2019-09-25T13:27:06.713","AcceptedAnswerId":null,"OwnerUserId":4903201.0,"Title":"How to get gradients of weights w.r.t. to a target neuron?","Body":"<p>I have found code online to get the derivative of the total loss with respect to the deep learning weights. I am trying to find the derivative of the weights with respect to the loss of a single class instead of all classes. <\/p>\n\n<p>I used the following code to get the gradient of an input image with respect to the total loss. If I visualize it, it shows the importance of the pixels for all predictions. But, I would like to compute the derivative of the input image with respect to a particular class (e.g. \"lady_bug\"). This should show the importance of the pixels for the prediction of lady_bug. Do you have an idea how I can do that?<\/p>\n\n<pre><code>from keras.applications.vgg19 import VGG19\nimport numpy as np\nimport cv2\nfrom keras import backend as K\nimport matplotlib.pyplot as plt\n\nfrom keras.applications.inception_v3 import decode_predictions\n\n\ndef get_model():\n    model = VGG19(include_top=True, weights='imagenet')\n    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n    return model\n\n\ndef predict(model, images):\n    numeric_prediction = model.predict(images)\n    categorical_prediction = decode_predictions(numeric_prediction, top=1)\n    return [(x[0][1], x[0][2]) for x in categorical_prediction]\n\n\ndef get_test_image():\n    # Image\n    image_path = \"lady_bug.jpg\"\n    image = cv2.imread(image_path)\n    my_image = cv2.resize(image, (224,224))\n    my_image = np.expand_dims(my_image, axis=0)\n    return my_image\n\n\ndef visualize_sample(sample, file_path):\n    plt.figure()\n    plt.imshow(sample)\n    plt.savefig(file_path, bbox_inches='tight')\n\n\ndef test_input_gradient():\n    images = get_test_image()\n    model = get_model()\n\n    prediction = predict(model, images)\n    print(prediction)\n\n    gradients = K.gradients(model.output, model.input)              #Gradient of output wrt the input of the model (Tensor)\n    print(gradients)\n\n    sess = K.get_session()\n    evaluated_gradients = sess.run(gradients[0], feed_dict={model.input:\n    images})\n\n    visualize_sample((evaluated_gradients[0]*(10**9.5)).clip(0,255), \"test.png\")\n\n\nif __name__ == \"__main__\":\n    test_input_gradient()\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>[('ladybug', 0.53532666)]\n[&lt;tf.Tensor 'gradients\/block1_conv1\/convolution_grad\/Conv2DBackpropInput:0' shape=(?, 224, 224, 3) dtype=float32&gt;]\n<\/code><\/pre>\n","answers":[{"AnswerId":"58100675","CreationDate":"2019-09-25T14:25:02.360","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>It seems the code is taking the gradients of the outputs wrt the inputs.<br>\nSo, this is just taking a single slice from the outputs. <\/p>\n\n<blockquote>\n  <p>Warning: This considers a regular model output. I have no idea of what you're doing in decode predictions and the following list. <\/p>\n<\/blockquote>\n\n<pre><code>gradients = K.gradients(model.output[:, lady_bug_class], model.input)   \n<\/code><\/pre>\n"}]}
+{"QuestionId":58099575,"AnswerCount":0,"Tags":"<python><tensorflow><keras><classification>","CreationDate":"2019-09-25T13:29:43.210","AcceptedAnswerId":null,"OwnerUserId":2277994.0,"Title":"Get output confidence in binary classification with Keras","Body":"<p>I am trying to build a really shallow binary classification network in Keras in order to discriminate between two domains.<\/p>\n\n<p>I want to get the probability (or some sort of confidence score) for each prediction.<\/p>\n\n<p>I read the data from folders like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>train_it = datagen.flow_from_directory('sim2city\/train\/', class_mode='binary', batch_size=32)\nval_it = datagen.flow_from_directory('sim2city\/val\/', class_mode='binary', batch_size=32)\ntest_it = datagen.flow_from_directory('sim2city\/test\/', class_mode='binary', batch_size=1)\n<\/code><\/pre>\n\n<p>I haven't done any other processing.\nI tried to train something like this for 5 epochs:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = keras.Sequential([\n      keras.layers.Flatten(input_shape=(256, 256, 3)),\n      keras.layers.Dense(32, activation=tf.nn.relu),\n    #keras.layers.Dense(16, activation=tf.nn.relu),\n      keras.layers.Dense(2, activation=tf.nn.softmax)\n   ])\n<\/code><\/pre>\n\n<p>then<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.compile(optimizer='adam',\n              loss='sparse_categorical_crossentropy',\n              metrics=['accuracy'])\npredict = model.predict_generator(test_it, steps=NumImgs)\n<\/code><\/pre>\n\n<p>but the values of the results are 0\/1 or 1\/0, not something in between to be used as a confidence score. Also, for me it is pretty weird because the total accuracy is not 100%, so if it has some wrong classifications, I guess, at least, it should not be so sure of it.<\/p>\n","answers":[]}
+{"QuestionId":58099666,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-25T13:34:18.387","AcceptedAnswerId":null,"OwnerUserId":10215420.0,"Title":"My Keras yolov3-tiny.h5 model has an input tensor with dimensions ?,?,?,3 while I would expect ?,416,416,3","Body":"<p>When inspecting the Keras model yolov3-tiny.h5 using netron<\/p>\n\n<p>I see that the input node is called input_1  and has type float32[?,?,?,3]. I would expect float32[?,416,416,3]<\/p>\n\n<p>How can I force it to be float32[?,416,416,3]?\nThis is needed for downstream processing. The Keras model has to be converted to a frozen_model.pb in Tensorflow and then be further processed for deployement.\nThe deployement tools cannot handle an input with unknow w,h size.<\/p>\n\n<p>Here is how I generated the Keras model. I downloaded the yolov3-tiny.cfg (<a href=\"https:\/\/github.com\/pjreddie\/darknet\/blob\/master\/cfg\/yolov3-tiny.cfg\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pjreddie\/darknet\/blob\/master\/cfg\/yolov3-tiny.cfg<\/a>) and yolov3-tiny.weigths (<a href=\"https:\/\/pjreddie.com\/media\/files\/yolov3-tiny.weights\" rel=\"nofollow noreferrer\">https:\/\/pjreddie.com\/media\/files\/yolov3-tiny.weights<\/a>)<\/p>\n\n<p>And then converted the model to a keras model using the following command : <\/p>\n\n<p>python convert.py -p yolov3-tiny.cfg yolov3-tiny.weights model_data\/yolov3-tiny.h5<\/p>\n\n<p>(this code is obtained by cloning <a href=\"https:\/\/github.com\/qqwweee\/keras-yolo3\" rel=\"nofollow noreferrer\">https:\/\/github.com\/qqwweee\/keras-yolo3<\/a>)<\/p>\n\n<p>Making a prediction using the saved Keras model works fine :\npython yolo_video.py --image --model model_data\/yolov3-tiny.h5<\/p>\n\n<p>However when inspecting the Keras model yolov3-tiny.h5 using netron<\/p>\n\n<p>I see that the input node is called input_1  and has type float32[?,?,?,3]<\/p>\n\n<p>I would expect float32[?,416,416,3]<\/p>\n\n<p>How can I force it to be float32[?,416,416,3]?<\/p>\n","answers":[]}
+{"QuestionId":58099864,"AnswerCount":1,"Tags":"<python><machine-learning><image-processing><keras><vgg-net>","CreationDate":"2019-09-25T13:45:13.540","AcceptedAnswerId":58099974.0,"OwnerUserId":12096782.0,"Title":"How to fix (\"ValueError: You are trying to load a weight file containing 16 layers into a model with 0 layers\")","Body":"<p>I'm using vgg16 like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = VGG16()\ndata, labels = ReadImages(TRAIN_DIR)\n\nvgg16 = VGG16()\n\nmodel = Sequential()\n\n#Converting VGG16 into Sequential model\nfor layer in vgg16.layers[:-1]:\n    model.add(layer)\n\n#Freezing all layers except last layer for transfer learning\nfor layer in model.layers:\n    layer.trainable = False\n\n#Adding custom softmax layer\nmodel.add(Dense(1,activation='sigmoid'))\n\n#Compiling our model\nmodel.compile(optimizer='adam',loss='binary_crossentropy',metrics=['accuracy'])\n\nmodel.fit(np.array(data), np.array(labels), batch_size=32, epochs=3)\n\nmodel.save('model.h5')\n<\/code><\/pre>\n\n<p>and when I tried to load this model in another py file..:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = load_model('model.h5')\n<\/code><\/pre>\n\n<p><strong>I'd already tried load_weights and throws an error too<\/strong><\/p>\n\n<p>... returns this error:<\/p>\n\n<pre><code>ValueError: You are trying to load a weight file containing 16 layers into a model with 0 layers\n<\/code><\/pre>\n\n<p>What should I do to load this model to make my predictions?<\/p>\n\n<p>versions: keras 2.2.4 tensorflow 1.14.0<\/p>\n","answers":[{"AnswerId":"58099974","CreationDate":"2019-09-25T13:49:45.360","ParentId":null,"OwnerUserId":"5597718","Title":null,"Body":"<p>Known issue: <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/10417\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/10417<\/a><\/p>\n\n<p>There are three options: 1. Recreate the model architecture and use '<code>load_weights<\/code>'. This is good if you only want to do predictions. 2. Downgrade to Keras version <code>2.1.6.<\/code> 3. A workaround available at this link <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/10417#issuecomment-435620108\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras\/issues\/10417#issuecomment-435620108<\/a>. I adapted this for VGG16. This updates the h5 file. <\/p>\n\n<pre><code>def fix_layer0(filename, batch_input_shape, dtype):\n    with h5py.File(filename, 'r+') as f:\n        model_config = json.loads(f.attrs['model_config'].decode('utf-8'))\n        layer0 = model_config['config']['layers'][0]['config']\n        layer0['batch_input_shape'] = batch_input_shape\n        layer0['dtype'] = dtype\n        f.attrs['model_config'] = json.dumps(model_config).encode('utf-8')\n\nfix_layer0('model.h5', [None, 224, 224, 3], 'float32')\n\nloaded_model = load_model('model.h5')\n<\/code><\/pre>\n"}]}
+{"QuestionId":58100071,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-25T13:54:34.953","AcceptedAnswerId":null,"OwnerUserId":2206908.0,"Title":"What is the correct way to implement a 'useless loss' with Keras?","Body":"<p>I have a Keras model that has two outputs:<\/p>\n\n<ul>\n<li><p><code>output<\/code> is the true output of the network on which the loss is going to be computed<\/p><\/li>\n<li><p><code>additional<\/code> is used to make an external task during inference (no loss should be computed with this output)<\/p><\/li>\n<\/ul>\n\n<p>When I build the model, I write something like that:<\/p>\n\n<pre><code>model = Model(inputs=inp, outputs=[output, additional])\n<\/code><\/pre>\n\n<p>Since my <code>Model<\/code> has two outputs, I need to provide two losses when compiling the model so I created a useless loss like this:<\/p>\n\n<pre><code>class NoopLoss(object):\n\n    def __call__(self, y_true, y_pred, **kwargs):\n        return self.compute_loss(y_true, y_pred)\n\n    def compute_loss(self, y_true, y_pred):\n        return tf.math.square(0.0)\n\n<\/code><\/pre>\n\n<p>Which I integrate in the compile step like this:<\/p>\n\n<pre><code>loss = UsefulLoss()  # the real loss I'm using\nnoop_loss = NoopLoss()\n\nmodel.compile(loss=[loss, noop_loss], optimizer=optimizer, metrics=['binary_accuracy'])\n<\/code><\/pre>\n\n<p>It works, but I feel it is a bit hackish, is there a correct way to implement this behavior? I didn't find any official useless loss in the Keras documentation.<\/p>\n","answers":[{"AnswerId":"58100391","CreationDate":"2019-09-25T14:10:27.997","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>In my opinion, Keras was not thought to consider things like this.<br>\nI often use these hacks myself too.   <\/p>\n\n<p>But, not sure it's a better solution, actually it might not be, you can create a training model and an inference model, both sharing the trainable part:<\/p>\n\n<pre><code>inputs = Input(...)\ntrainable_out = SomeLayer(...)(inputs)\n....\ntrainable_out = ....\n\nextra_output = SomeLayer(...)(something)    \n\ntraining_model = Model(inputs, trainable_out)\ninference_model = Model(inputs, [trainable_out, extra_output])   \n<\/code><\/pre>\n\n<p>You can train <code>training_model<\/code> and automatically the other model will be trained as well. <\/p>\n"}]}
+{"QuestionId":58100108,"AnswerCount":0,"Tags":"<tensorflow><lstm><backpropagation>","CreationDate":"2019-09-25T13:56:49.363","AcceptedAnswerId":null,"OwnerUserId":9041648.0,"Title":"I am using LSTM , I want to know how to control BPTT in TensorFlow?","Body":"<p>I want to put <code>n<\/code> input into a LSTM, then get every output of this LSTM, then sum these output, then get a cost then BPTT.<\/p>\n\n<p>But for LSTM I know when I input, I must need to do BPPT, I don't know how to get sum of output to do BPTT.<\/p>\n","answers":[]}
+{"QuestionId":58100305,"AnswerCount":1,"Tags":"<json><tensorflow><tensorflow.js>","CreationDate":"2019-09-25T14:06:18.543","AcceptedAnswerId":null,"OwnerUserId":11335442.0,"Title":"TensorflowJS: Failed to parse model.json","Body":"<p>I have retrained and converted a Mobilenet model into a model.json-file. This file is loaded into my image classification platform (got a 200 OK status code), but is not recognized as proper json. <\/p>\n\n<p>I have used these sites to validate my json file<\/p>\n\n<p><a href=\"https:\/\/jsonlint.com\/\" rel=\"nofollow noreferrer\">https:\/\/jsonlint.com\/<\/a><\/p>\n\n<p><a href=\"https:\/\/jsonformatter.curiousconcept.com\/\" rel=\"nofollow noreferrer\">https:\/\/jsonformatter.curiousconcept.com\/<\/a><\/p>\n\n<p>And both return my file as being validated json. <\/p>\n\n<p>Two other JSON-files are loaded into the page in the same way, and they return no errors<\/p>\n\n<p>Yet I still keep getting this error:<\/p>\n\n<p><code>Uncaught (in promise) Error: Failed to parse model JSON of response from https:\/\/localhost\/pathto\/model.json. Please make sure the server is serving valid JSON for this request.<\/code><\/p>\n\n<p>Does anyone know what might be wrong with the file?<\/p>\n\n<p>EDIT:\nI haven't touched the model.json in any way, I simply loaded it into my index.js file and got the error. <\/p>\n\n<p>Also this my index.js (the full version):<\/p>\n\n<pre><code>  const MODEL_URL =\n      \"https:\/\/localhost\/tfjs_models\/model.json\";\n    const WEIGHTS_URL =\n      \"https:\/\/localhost\/tfjs_models\/weights_manifest.json\";\n    let model;\n    let IMAGENET_CLASSES = [];\n    let offset = tf.scalar(128);\n    async function loadModelAndClasses() {\n  $.getJSON(\n    \"https:\/\/localhost\/tfjs_models\/labels.json\",\n    function(data) {\n      $.each(data, function(key, val) {\n        IMAGENET_CLASSES.push(val);\n      });\n    }\n  );\n  model = await tf.loadGraphModel(MODEL_URL, WEIGHTS_URL);\n  \/\/console.log(\"After model is loaded: \" + tf.memory().numTensors);\n  $(\".loadingDiv\").hide();\n  $(\"#inputImage\").attr(\"disabled\", false);\n}\nloadModelAndClasses();\nfunction readURL(input) {\n  if (input.files &amp;&amp; input.files[0]) {\n    var reader = new FileReader();\n\n    reader.onload = function(e) {\n      $(\"#imageSrc\")\n        .attr(\"src\", e.target.result)\n        .width(224)\n        .height(224);\n    };\n\n    reader.readAsDataURL(input.files[0]);\n    \/\/console.log(\"After image is loaded: \" + tf.memory().numTensors);\n\n    reader.onloadend = async function() {\n      console.log(\"Before predictions: \" + tf.memory().numTensors);\n\n      let imageData = document.getElementById(\"imageSrc\");\n\n      \/\/console.log(\"After offset: \" + tf.memory().numTensors);\n      let pixels1 = tf.fromPixels(imageData);\n      let pixel2 = pixels1.resizeNearestNeighbor([224, 224]);\n      let pixel3 = pixel2.toFloat();\n      console.log(\"After pixels are formed: \" + tf.memory().numTensors);\n\n      let pixels = pixel3.sub(offset);\n      let pixels4 = pixels.div(offset);\n      let pixels5 = pixels4.expandDims();\n      console.log(\"After pre-processing: \" + tf.memory().numTensors);\n\n      const output = await model.predict(pixels5);\n      console.log(\"After output: \" + tf.memory().numTensors);\n      const predictions = Array.from(output.dataSync())\n        .map(function(p, i) {\n          return {\n            probabilty: p,\n            classname: IMAGENET_CLASSES[i]\n          };\n        })\n        .sort((a, b) =&gt; b.probabilty - a.probabilty)\n        .slice(0, 10);\n\n      \/\/console.log(predictions);\n      var html = \"\";\n      for (let i = 0; i &lt; 10; i++) {\n        html += \"&lt;li&gt;\" + predictions[i].classname + \"&lt;\/li&gt;\";\n      }\n      $(\".predictionList\").html(html);\n      console.log(\"After predictions: \" + tf.memory().numTensors);\n\n      pixels.dispose();\n      pixels1.dispose();\n      pixel2.dispose();\n      pixel3.dispose();\n      pixels4.dispose();\n      pixels5.dispose();\n      output.dispose();\n      console.log(\"After dispose: \" + tf.memory().numTensors);\n    };\n  }\n}\n<\/code><\/pre>\n\n<p>This is the link to the model as it was generated by TensorflowJS Converter. <\/p>\n\n<p><a href=\"https:\/\/github.com\/dvbeelen\/tfjs_model\" rel=\"nofollow noreferrer\">https:\/\/github.com\/dvbeelen\/tfjs_model<\/a><\/p>\n","answers":[{"AnswerId":"58582006","CreationDate":"2019-10-27T18:08:07.707","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>The model is wrongly loaded. Here is the <a href=\"https:\/\/js.tensorflow.org\/api\/latest\/#loadGraphModel\" rel=\"nofollow noreferrer\">doc<\/a><\/p>\n\n<p>This should be enough to load a model<\/p>\n\n<pre><code>model = await tf.loadGraphModel(MODEL_URL)\n<\/code><\/pre>\n\n<p>The <code>weights_manifest.json<\/code> needs not to be supplied<\/p>\n"}]}
+{"QuestionId":58100322,"AnswerCount":1,"Tags":"<keras><deep-learning><lstm>","CreationDate":"2019-09-25T14:07:06.020","AcceptedAnswerId":58102380.0,"OwnerUserId":9084663.0,"Title":"Top k categorical accuracy for Time Distributed LSTM results","Body":"<p>I'm trying to evaluate the results of an LSTM using <code>top_k_categorical_accuracy<\/code>.<\/p>\n\n<p>For each One-Hot encoded token, I try to predict the next token. In order to do this I take the output for each instance in the sequence by using the <code>TimeDistributed<\/code> layer wrapper, and pass it to a Dense layer to re-encode the results into the same One-Hot encoding.<\/p>\n\n<p>While using the built in accuracy metric <code>metrics=['accuracy']<\/code> works without a hitch, using <code>top_k_categorical_accuracy<\/code> fails, giving me the error message:<\/p>\n\n<pre><code>ValueError: Shape must be rank 2 but is rank 3 for 'metrics\/my_acc\/in_top_k\/InTopKV2' (op: 'InTopKV2') with input shapes: [?,?,501], [?,?], [].\n<\/code><\/pre>\n\n<p>What do I need to change in order to make this metric work?<\/p>\n\n<p>My code is a follows:<\/p>\n\n<pre><code>import numpy as np\nimport glob\n\nimport keras\nfrom keras.models import Sequential\nfrom keras.layers import LSTM, Dense, TimeDistributed,Lambda, Dropout, Activation\nfrom keras.metrics import top_k_categorical_accuracy\n\n\n\n\ntrain_val_split=0.2 # portion to be placed in validation\n\n\ntrain_control_number=0\nval_control_number=0\n\n\ndef my_acc(y_true, y_pred):\n    return top_k_categorical_accuracy(y_true, y_pred, k=5)\n\n\ndef basic_LSTM(features_num):\n    model = Sequential()\n    model.add(LSTM(40, return_sequences=True, input_shape=(None, features_num)))\n    model.add(LSTM(40, return_sequences=True))\n    model.add(LSTM(40, return_sequences=True))\n\n    model.add(TimeDistributed(Dense(features_num)))\n    model.add(Activation('linear')) \n\n    print(model.summary())\n    model.compile(loss='categorical_crossentropy', optimizer='adam',metrics=[my_acc])   \n    return (model)\n\n\ndef main ():\n    input_files=glob.glob('*npy')\n    data_list,dim=loader(input_files)\n    train_list,val_list=data_spliter(data_list)\n    model=basic_LSTM(dim)\n    model.fit_generator(train_generator(train_list), steps_per_epoch=len(train_list), epochs=10, verbose=1,validation_data=val_generator(val_list),validation_steps=len(val_list))\n\n\n\n\ndef train_generator(data_list):\n    while True:\n        global train_control_number\n        train_control_number=cycle_throught(len(data_list),train_control_number)    \n        this=data_list[train_control_number]\n        x_train = this [:,:-1,:] # all but the last 1\n        y_train = this [:,1:,:] # all but the first 1\n\n        yield (x_train, y_train)\n\n\n\n\ndef val_generator(data_list):\n    while True:\n        global val_control_number\n        val_control_number=cycle_throught(len(data_list),val_control_number)    \n        this=data_list[val_control_number]\n        x_train = this [:,:-1,:] # all but the last 1\n        y_train = this [:,1:,:] # all but the first 1\n\n        yield (x_train, y_train)\n\n\n\ndef cycle_throught (total,current):\n    current+=1\n    if (current==total):\n        current=0\n    return (current)\n\n\ndef loader(input_files):\n\n    data_list=[]\n\n    for input_file in input_files:\n        a=np.load (input_file)\n        incoming_shape=list(a.shape)\n        requested_shape=[1]+incoming_shape\n        a=a.reshape(requested_shape)\n        data_list.append(a)\n\n\n    return (data_list,incoming_shape[-1])\n\n\ndef data_spliter(input_list):\n    val_num=int(len(input_list)*train_val_split)\n    validation=input_list[:val_num]\n    train=input_list[val_num:]\n\n    return (train,validation)\n\n\n\nmain()\n<\/code><\/pre>\n\n<p>Many thanks.<\/p>\n","answers":[{"AnswerId":"58102380","CreationDate":"2019-09-25T15:58:54.637","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>You can transform your data in a 2D tensor within a custom metric so it suits to the required shape, keeping the last axis untouched:<\/p>\n\n<pre><code>import keras.backend as K #or tf.keras.backend as K    \n\ndef 3D_top_k(true, pred):\n    true = K.reshape(true, (-1, features_num))   \n    pred = K.reshape(pred, (-1, features_num))\n    return top_k_categorical_accuracy(true, pred, k=5)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58100370,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><scikit-image>","CreationDate":"2019-09-25T14:09:30.260","AcceptedAnswerId":null,"OwnerUserId":9582953.0,"Title":"how to access image from a dataset using skimage? Is there any other alternate?","Body":"<pre><code>import tensorflow as tf\nimport os\nimport skimage\ntrain_data_directory = \"F:\/Projects\/Tensorflow\/Code\/Training\"\ndirectories = [d for d in os.listdir(train_data_directory)\nif os.path.isdir(os.path.join(train_data_directory, d))]\nlabels =[] \nimages = []\nfor d in directories:\n    label_directory = os.path.join(train_data_directory, d)\n    file_name = [os.path.join(train_data_directory, f)\n                for f in os.listdir(label_directory)\n                if f.endswith(\".ppm\")]\n    labels.append(int(d))\n\nfor i in file_name:\n    skimage.io.imread(i)\n<\/code><\/pre>\n\n<p>this is my code, skimage.io.imread() this line is giving multiple error, i don't know why?<\/p>\n\n<p>The error I'm getting is :<\/p>\n\n<p><strong>FileNotFoundError: No such file: 'F:\\Projects\\Tensorflow\\Code\\Training\\00032_00000.ppm'<\/strong><\/p>\n","answers":[]}
+{"QuestionId":58101091,"AnswerCount":1,"Tags":"<pytorch><lstm><word2vec>","CreationDate":"2019-09-25T14:45:38.813","AcceptedAnswerId":58104247.0,"OwnerUserId":7134235.0,"Title":"Choosing loss function for lstm trained on word2vec vectors when target is a vector of same dimensions","Body":"<p>I have an lstm I'm using as a sequence-generator trained on word2vec vectors. The previous implementation produced a probability distribution for all the different labels. There was one label for every word in the vocabulary. This implementation used Pytorch's CrossEntropyLoss. I now want to change this so the lstm outputs a vector that has the same dimensions as the vectors used for training. This way I could use the euclydian distance measure to match wit to nearby vectors in the vocabulary. The problem is that in order to do this, I have to use a different loss function, because CrossEntropyLoss is appropriate for classifiers and not for regression problems. <\/p>\n\n<p>I tried changing the format of the target vector, but torch's CrossEntropyLoss function requires integer input, and I have a word vector. Having looked at a few options it seems Cosine Embedding Loss might be a good idea but I don't understand how it works and what kind of input it takes. <\/p>\n\n<p>I have already changed my fully connected layer to output vectors of the same dimensions as the Word Embeddings used for training:   <\/p>\n\n<pre><code>nn.Linear(in_features=self.cfg.lstm.lstm_num_hidden,out_features=self.cfg.lstm.embedding_dim,bias=True)\n<\/code><\/pre>\n\n<p>Any advice and examples would be much appreciated.<\/p>\n","answers":[{"AnswerId":"58104247","CreationDate":"2019-09-25T18:07:27.603","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p>As the documentation of <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html?highlight=cosine%20embedding%20loss#torch.nn.CosineEmbeddingLoss\" rel=\"nofollow noreferrer\">CosineEmbeddingLoss<\/a> says:<\/p>\n\n<blockquote>\n  <p>Creates a criterion that measures the loss given two input tensors and a Tensor label with values 1 or -1.<\/p>\n<\/blockquote>\n\n<p>In your scenario, you should always provide 1 as the Tensor label.<\/p>\n\n<pre><code>batch_size, seq_len, w2v_dim = 32, 100, 200\nx1 = torch.randn(batch_size, seq_len, w2v_dim)\nx2 = torch.randn(batch_size, seq_len, w2v_dim)\ny = torch.ones(batch_size, seq_len)\nloss_fn = torch.nn.CosineEmbeddingLoss(reduction='none')\nloss = loss_fn(x1.view(-1, w2v_dim), \n               x2.view(-1, w2v_dim),  \n               y.view(-1))\nloss = loss.view(batch_size, seq_len)\n<\/code><\/pre>\n\n<p>Here, I assume <code>x1<\/code> is the word embeddings, <code>x2<\/code> is the output of the LSTM followed by some transformation.<\/p>\n\n<blockquote>\n  <p>Why should I always provide 1 as the Tensor label?<\/p>\n<\/blockquote>\n\n<p>First, you should see the loss function.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/W63PK.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/W63PK.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>In your scenario, the higher the cosine similarity is, the lower the loss should be. In other words, you want to maximize the cosine similarity. So, you need to provide 1 as the label.<\/p>\n\n<p>On the other hand, if you want to minimize the cosine similarity, you need to provide -1 as the label.<\/p>\n"}]}
+{"QuestionId":58101262,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><object-detection><autoencoder>","CreationDate":"2019-09-25T14:54:55.483","AcceptedAnswerId":null,"OwnerUserId":7456788.0,"Title":"How to train a neural network without lots of labelled data","Body":"<p>I'm new to machine learning and I'm trying to build a object detection neural network.<\/p>\n\n<p>I have 3000 images but only 150 are labelled.\nMy images are composed of web ui pages where I label buttons and logos.<\/p>\n\n<p>For now, I have tried to train it using the faster_rcnn_inception_resnet_v2_coco model but the results are not great.\nThe features are found but not perfectly where they need to be.\n<a href=\"https:\/\/i.stack.imgur.com\/U74V1.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/U74V1.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Is there a way to improve results? maybe with another model?<\/p>\n\n<p>If I use an auto-encoder for unsupervised training with all the 3000 images and then use transfer learning into a deep learning model with the labelled data will I get better results? If so, how can I implement it using python with tensorflow?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"58110846","CreationDate":"2019-09-26T06:35:21.530","ParentId":null,"OwnerUserId":"4622046","Title":null,"Body":"<p>First, look into COCO and VOC datasets.<\/p>\n\n<p><a href=\"http:\/\/cocodataset.org\/#home\" rel=\"nofollow noreferrer\">http:\/\/cocodataset.org\/#home<\/a><\/p>\n\n<p><a href=\"http:\/\/host.robots.ox.ac.uk\/pascal\/VOC\/\" rel=\"nofollow noreferrer\">http:\/\/host.robots.ox.ac.uk\/pascal\/VOC\/<\/a><\/p>\n\n<p>If their classes are very similar to yours, then you can take a pre-trained model (deeplab, yolo, maskrnn, etc.) and train on your dataset, the performance should be great. Here are some repositories to check in github:<\/p>\n\n<p><a href=\"https:\/\/github.com\/search?l=Python&amp;q=object+detection&amp;type=Repositories\" rel=\"nofollow noreferrer\">https:\/\/github.com\/search?l=Python&amp;q=object+detection&amp;type=Repositories<\/a><\/p>\n\n<p>If your dataset is very different from these, you may not expect good performance. In that case, you have to redesign the network by changing the number of classes. You can use data augmentation (auto-encoder may not be great choice here for data generation as it's complicated to generate good quality images). Here is a good library for data augmentation which augments samples with masks.<\/p>\n\n<p><a href=\"https:\/\/github.com\/mdbloice\/Augmentor\" rel=\"nofollow noreferrer\">https:\/\/github.com\/mdbloice\/Augmentor<\/a><\/p>\n"}]}
+{"QuestionId":58101445,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-25T15:04:07.723","AcceptedAnswerId":null,"OwnerUserId":4969574.0,"Title":"Keras functional API and TensorFlow Hub","Body":"<p>I'm trying to use a <a href=\"https:\/\/tfhub.dev\/google\/universal-sentence-encoder-multilingual-large\/1\" rel=\"nofollow noreferrer\">Universal Sentence Encoder<\/a> from TF Hub as a keras layer in a functional way. I would like to use <code>hub.KerasLayer<\/code> with Keras Functional API, but i'm not sure how to achieve that, so far I've only seen exmaples of hub.KerasLayer with the Sequential API<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow_hub as hub\nimport tensorflow as tf\nfrom tensorflow.keras import layers\nimport tf_sentencepiece\n\n\nuse_url = 'https:\/\/tfhub.dev\/google\/universal-sentence-encoder-multilingual-large\/1'\n\nenglish_sentences = [\"dog\", \"Puppies are nice.\", \"I enjoy taking long walks along the beach with my dog.\"]\nenglish_sentences = np.array(english_sentences, dtype=object)[:, np.newaxis]\n\n\nseq = layers.Input(shape=(None, ), name='sentence', dtype=tf.string)\nmodule = hub.KerasLayer(hub.Module(use_url))(seq)\nmodel = tf.keras.models.Model(inputs=[seq], outputs=[module])\nmodel.summary()\n\nx = model.predict(english_sentences)\nprint(x)\n\n<\/code><\/pre>\n\n<p>the code above runs into this error when passing the input layer to the embedding: <code>TypeError: Can't convert 'inputs': Shape TensorShape([Dimension(None), Dimension(None)]) is incompatible with TensorShape([Dimension(None)])<\/code><\/p>\n\n<p>Is it possible to use hub.KerasLayer with keras functional API in TensorFlow 1.x? if it can be done, how?<\/p>\n","answers":[{"AnswerId":"58240550","CreationDate":"2019-10-04T17:03:52.803","ParentId":null,"OwnerUserId":"7760431","Title":null,"Body":"<p>First shape is what you are passing into the model, <code>Shape TensorShape([Dimension(None), Dimension(None)])<\/code>. Second shape is what you are expecting, <code>TensorShape([Dimension(None)])<\/code>. So in this error, its telling you it expecting a shape of ()...<\/p>\n\n<p>Or<\/p>\n\n<p>If you are expecting to do batches of text, perhaps do <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/TimeDistributed#class_timedistributed\" rel=\"nofollow noreferrer\">TimeDistributed<\/a> layer, like so... <\/p>\n\n<pre><code>module = tf.keras.layers.TimeDistributed(hub.KerasLayer(hub.Module(use_url)))(seq)\n<\/code><\/pre>\n\n<p>However you maybe forced to do specific size for text length...<\/p>\n"}]}
+{"QuestionId":58101620,"AnswerCount":2,"Tags":"<python><tensorflow><keras><conv-neural-network>","CreationDate":"2019-09-25T15:13:16.137","AcceptedAnswerId":null,"OwnerUserId":12119935.0,"Title":"Can input to 3D convolution be a list of 4D arrays, where rows of list are samples?","Body":"<p>I have 3d models obtained from .mat file of dimensions 182 x 199x3 and 22 frames. I have a total sample of 100 such data. \nAfter loading to the python I obtained a list of 100 arrays [Numpy array, numpy array...] whcih i converted to arrays and dimesions are [100,22,182,199,3]. Technically this looks like 5D input which conv3d requires but, i was wondering cant i feed list of arrays to input keeping the bachsize whatever I wish like 20\/32 or anything.<\/p>\n","answers":[{"AnswerId":"58101953","CreationDate":"2019-09-25T15:32:55.057","ParentId":null,"OwnerUserId":"9297078","Title":null,"Body":"<p>As per your specification of the data.<\/p>\n\n<pre><code>==&gt;&gt; Basic size of the data is 182 X 199 X 3\n==&gt;&gt; 22 is the total frames\n==&gt;&gt; so one batch of the data is [22(n_channel)X 182 X 199 X 3]\n==&gt;&gt; if you want to know the total shape of the data it would be 100 X 22 X 182 X 199 X 3\n==&gt;&gt; and if you wish for a batch_size of say 30 the shape would be [30 X 22 X 182 X 199 X 3]\n<\/code><\/pre>\n\n<p>So yes you can feed the data as per your requirement of batch_size<\/p>\n"},{"AnswerId":"58102570","CreationDate":"2019-09-25T16:10:22.163","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>You can input any 5D array.   <\/p>\n\n<p>You cannot input \"lists\" of arrays, but \"an\" array with the required batch size:<\/p>\n\n<pre><code>batch = full_array[:batch_size]   \nbatch2 = full_array[batch_size : 2* batch_size]\n<\/code><\/pre>\n\n<p>But Keras automatically iterates the batches from a single array in:   <\/p>\n\n<pre><code>model.fit(full_inputs_array, full_outputs_array, batch_size=32, ...)   \n<\/code><\/pre>\n"}]}
+{"QuestionId":58101761,"AnswerCount":1,"Tags":"<python><dataframe><keras><generator>","CreationDate":"2019-09-25T15:21:25.853","AcceptedAnswerId":58113197.0,"OwnerUserId":3169677.0,"Title":"How to enforce mapping of class_indices in flow_from_dataframe","Body":"<p>I am setting up a keras data generator using flow_from_dataframe. \nThe data are images, and there are 2000 different classes. I have a dataframe which maps the image files to the 2000 categories (integers from 0 to 1999). Both columns (links to images and labels) are formatted as strings.<\/p>\n\n<p>I have a net that was trained by someone else with a defined mapping from class_indices to labels. Here's the problem: When I set up my data generator with data that I want to test the trained model with, it enforces an alphanumeric ordering on the labels that is different from the one the trained net works with. I now get that \"0\" is 0, \"1\" is 1, but \"100\" is 2, \"1000\" is 3 etc, but I want that \"2\" is 2, \"3\" is 3 etc. <\/p>\n\n<p>As a consequence, the trained model performs at 0% accuracy.<\/p>\n\n<p>How can I circumvent the alphanumeric ordering in the data generator (such that \"2\" is 2, \"3\" is 3)?<\/p>\n\n<p>I am using keras 2.2.4. I have tried the following (unsuccesfully):<\/p>\n\n<ul>\n<li><p>transform the column in the pandas dataframe with the labels to %04d-format such that, if alphanumerically sorted, it should result in the desired order -- however, this results in a different order ('0000': 0, '0001': 1, '0007': 2, '0008': 3, '0011' ...)<\/p><\/li>\n<li><p>provide the \"classes\" argument to the construction of the data generator with a list in the desired order -- however, this order gets overwritten<\/p><\/li>\n<li><p>provide the \"classes\" argument to the construction of the data generator with a dictionary of the desired mapping -- however, again, this gets overwritten and results in the initial problem\n(it seems to work for \"flow_from_directory\": <a href=\"https:\/\/stackoverflow.com\/questions\/53892271\/is-it-possible-to-change-class-indices-of-keras-flow-from-directory\">Is it possible to change class indices of Keras flow from directory<\/a>, but it doesn't work for me with flow_from_dataframe)<\/p><\/li>\n<\/ul>\n\n<p>What I am trying basically looks like this:<\/p>\n\n<pre><code>    my_generator = my_datagen.flow_from_dataframe(\n        dataframe=my_df,\n        target_size = (224,224),\n        directory=None,\n        x_col='filename',\n        y_col='yID',\n        class_mode='categorical',\n        classes=classLabels,\n        validate_filenames=False,\n        batch_size=128)\n<\/code><\/pre>\n\n<p>where classLabels can be<\/p>\n\n<pre><code>classLabels = list(map(str,range(2000)))\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>classLabels = list(map(str,range(2000)))\nclassLabels = dict(zip(classLabels,list(range(2000))))\n<\/code><\/pre>\n\n<p>Expected result is that in <code>my_generator.class_indices<\/code>, I get <code>{'0': 0, '1': 1, '2': 2, '3': 3, ...}<\/code>.\n Actual results are e.g. <\/p>\n\n<ul>\n<li><code>{'0000': 0, '0001': 1, '0007': 2, '0008': 3, '0011' ...}<\/code> when transforming the dataframe column to have 4 digits<\/li>\n<li><code>{'0': 0, '1': 1, '100': 2, '1000': 3, '1002': 4, ...}<\/code> when I do or do not provide the \"classes\" argument [whether as dict or list doesn't matter] when creating the generator<\/li>\n<\/ul>\n\n<p>It could very well be that I am overseeing the actual problem (perhaps this is very well possible but I am just missing a detail that I might not even be reporting here -- please excuse me if this is the case, and do not hesitate to think of solutions to my problem that are simpler than I thought).\nMany thanks!<\/p>\n","answers":[{"AnswerId":"58113197","CreationDate":"2019-09-26T09:02:40.547","ParentId":null,"OwnerUserId":"3169677","Title":null,"Body":"<p>A kind and clever colleague came to help. <\/p>\n\n<p>That it gives the weird order when transforming the label column to 4 digit strings was because the specific (testing) dataframe didn't have all 2000 classes, but just a subset.<\/p>\n\n<p>So in principle, the 4 digit transformation was the right way to go (in case someone needs this: <code>my_df['yID'] = my_df['yID'].astype(str).str.zfill(4)<\/code>). <\/p>\n\n<p>The only thing that was missing then was additionally specifying a <code>classLabels<\/code> dictionary where the strings are in 4 digit format as well:<\/p>\n\n<pre><code>classLabels = list(map(lambda x: \"{:04d}\".format(x),range(2000)))\nclassLabels = dict(zip(classLabels,list(range(2000))))\n<\/code><\/pre>\n\n<p>So basically just combining the things I was trying. <\/p>\n\n<p>Would still be cool if this was possible directly within keras, i.e., if flow_from_dataframe had an option to turn off the alphanumeric sorting.<\/p>\n"}]}
+{"QuestionId":58102016,"AnswerCount":1,"Tags":"<keras><tensorboard>","CreationDate":"2019-09-25T15:36:23.917","AcceptedAnswerId":58110102.0,"OwnerUserId":11554051.0,"Title":"How visualize in Tensorboard a metric callback?","Body":"<p>I have a model in keras in which I use my custom metric as:<\/p>\n\n<pre><code>class MyMetrics(keras.callbacks.Callback):\n    def __init__(self):\n        initial_value = 0\n    def on_train_begin(self, logs={}):\n        ...\n    def on_epoch_end(self, batch, logs={}):\n        here I calculate my important values\n\n<\/code><\/pre>\n\n<p>Now, there is a way to visualize them in Tensorboard?\nFor example if my metric was something like:<\/p>\n\n<pre><code>def mymetric(y_true,y_pred):\n    return myImportantValues\n\n<\/code><\/pre>\n\n<p>I could visualize them in Tensorboard through\n<code>mymodel.compile(..., metrics = mymetric)<\/code><\/p>\n\n<p>Is there something similar with a metric callback? \nI tried to create a function inside the class MyMetric and pass it to the <code>mymodel.compile<\/code> but it does not update the values.<\/p>\n","answers":[{"AnswerId":"58110102","CreationDate":"2019-09-26T05:27:13.373","ParentId":null,"OwnerUserId":"5597718","Title":null,"Body":"<p>You can create an event file with the custom metrics and visualize it in tensorboard directly.<\/p>\n\n<p>This works for Tensorflow 2.0. In this example, the accuracy\/metrics are logged from training history. In your case, you can do it from the <code>on_epoch_end<\/code> callback. <\/p>\n\n<pre><code>import datetime\ncurrent_time = datetime.datetime.now().strftime(\"%Y%m%d-%H%M%S\")\ntrain_log_dir = 'logs\/train\/' + current_time\ntrain_summary_writer = tf.summary.create_file_writer(train_log_dir)\n\nhistory = model.fit(x=X, y=y, epochs=100, verbose=1)\nfor epoch in range(len(history.history['accuracy'])):\n    with train_summary_writer.as_default():\n        tf.summary.scalar('loss', history.history['loss'][epoch], step=epoch)\n        tf.summary.scalar('accuracy', history.history['accuracy'][epoch], step=epoch)\n<\/code><\/pre>\n\n<p>After script execution,<\/p>\n\n<blockquote>\n  <p>tensorboard --logdir logs\/train<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/tensorboard\/r2\/get_started#using_tensorboard_with_other_methods\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tensorboard\/r2\/get_started#using_tensorboard_with_other_methods<\/a><\/p>\n"}]}
+{"QuestionId":58102755,"AnswerCount":0,"Tags":"<python-3.x><tensorflow>","CreationDate":"2019-09-25T16:24:02.950","AcceptedAnswerId":null,"OwnerUserId":11073614.0,"Title":"Cudnn problem tensorflow object detection api with resnet module","Body":"<p>I wanna to use the object detection api of tensorflow running the \"ssd_resnet50_v1_fpn_shared_box_predictor_640x640_coco14_sync_2018_07_03\" module, but i get errors.<\/p>\n\n<pre><code>tensorflow-gpu 1.14\nCudnn: 7.6.3.30\nCuda: 10.0\nUbuntu 18.04\nRTX 2070 Super \nDriver Version: 430.26 \n\n\n2019-09-25 13:18:37.848463: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcublas.so.10.0\n2019-09-25 13:18:38.989178: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudnn.so.7\n2019-09-25 13:18:39.288055: E tensorflow\/stream_executor\/cuda\/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR\n2019-09-25 13:18:39.313052: E tensorflow\/stream_executor\/cuda\/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR\nTraceback (most recent call last):\n  File \"\/home\/luiz\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1356, in _do_call\n    return fn(*args)\n  File \"\/home\/luiz\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1341, in _run_fn\n    options, feed_dict, fetch_list, target_list, run_metadata)\n  File \"\/home\/luiz\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1429, in _call_tf_sessionrun\n    run_metadata)\ntensorflow.python.framework.errors_impl.UnknownError: 2 root error(s) found.\n  (0) Unknown: 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 FeatureExtractor\/resnet_v1_50\/resnet_v1_50\/conv1\/Conv2D}}]]\n     [[control_dependency_1\/_7983]]\n  (1) Unknown: 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 FeatureExtractor\/resnet_v1_50\/resnet_v1_50\/conv1\/Conv2D}}]]\n0 successful operations.\n0 derived errors ignored.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58102771,"AnswerCount":0,"Tags":"<tensorflow><keras><tensorflow2.0>","CreationDate":"2019-09-25T16:24:42.770","AcceptedAnswerId":null,"OwnerUserId":866082.0,"Title":"How to create a zero-filled variable without mentioning the batch size","Body":"<p>If I have a tensor (let's call it <code>src<\/code>) of shape <code>(?, 10)<\/code> I can create a zero-filled variable of the same shape like this:<\/p>\n\n<pre><code>from tensorflow.python.keras import backend as K\nzeros = K.zeros_like(src)\n<\/code><\/pre>\n\n<p>But if I don't have a <code>src<\/code>, how can I can create a similar variable? I've tried this and failed:<\/p>\n\n<pre><code>from tensorflow.python.keras import backend as K\nzeros = K.zeros(shape=(None, 10))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58102782,"AnswerCount":1,"Tags":"<tensorflow><keras><complex-numbers>","CreationDate":"2019-09-25T16:25:52.880","AcceptedAnswerId":null,"OwnerUserId":1920178.0,"Title":"Simple custom layer in keras, tensorflow confusion","Body":"<p>First before I explain, here's the relevant code snippet:<\/p>\n\n<pre><code>input = Input(shape=(784, ))\nhidden1 = Dense(784, activation='relu')(input)\nhidden2 = Dense(784, activation='relu')(hidden1)\nhidden3 = Dense(1568, activation='relu')(hidden2)\nhidden4 = Lambda(lambda x: makeComplex(x))(hidden3)\nhidden5 = Reshape((1, 28, 28))(hidden4)\nhidden6 = Lambda(lambda x: ifft2(x))(hidden5)\nhidden7 = Flatten()(hidden6)\noutput = Dense(train_targets.shape[1], activation='linear')(hidden7)\nmodel = Model(inputs=input, outputs=output)\nprint(model.summary())\n<\/code><\/pre>\n\n<p>where ifft2(x) is <\/p>\n\n<pre><code>def ifft2(x):\n    import tensorflow as tf\n    return tf.cast(tf.spectral.ifft2d(tf.cast(x,dtype=tf.complex64)),tf.float32)\n<\/code><\/pre>\n\n<p>My goal now is to implement the makeComplex method.<\/p>\n\n<p>Basically, it gets a vector of size 1568, and I want it to return a vector of size 784 in the following very simple fashion:<\/p>\n\n<p><code>new[k] = old[k] + old[k + 1] * i<\/code> where <code>i<\/code> is the imaginary unit<\/p>\n\n<p>Here's my attempt:<\/p>\n\n<pre><code>def makeComplex(x):\n    y = np.zeros((1, 784))\n    for i in range(784):\n        y[i] = np.complex(x[i], x[i + 1])\n    return y\n<\/code><\/pre>\n\n<p>ofcourse this doesnt work because x is not actually a vector but rather a tensorflow tensor. Something I know nothing about. How can I make this work?<\/p>\n","answers":[{"AnswerId":"58104165","CreationDate":"2019-09-25T18:01:37.413","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>An example of tensor [1.,2.,3.,4.], what you want is [1.+2.j,3.+4.j]. I think you can use <code>tf.gather<\/code> to get two tensors <code>[1.,2.]<\/code> and <code>[3.,4.]<\/code>, then use <code>tf.complex<\/code> to get the answer. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\na = tf.constant([1.,2.,3.,4.])\nreal = tf.gather(a,np.arange(0,a.get_shape().as_list()[0],2)) \nimag = tf.gather(a,np.arange(1,a.get_shape().as_list()[0],2))\n\nres = tf.complex(real, imag)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58102824,"AnswerCount":1,"Tags":"<javascript><tensorflow><user-interface><p5.js>","CreationDate":"2019-09-25T16:28:46.580","AcceptedAnswerId":58103496.0,"OwnerUserId":7238787.0,"Title":"How can I access the real-time sketch images on a p5.js canvas?","Body":"<p>Recently I am attempting to modify the source codes of <a href=\"https:\/\/magenta.tensorflow.org\/assets\/sketch_rnn_demo\/index.html\" rel=\"nofollow noreferrer\">this page<\/a>. The underlying technique of this interactive programe is called sketch-rnn, a deep learning algorithm that can generate sketches. I need to access the real time images on the canvas so that I can use convolutional neural network (CNN), and feed the image as a 2d array to the neural network so that I can further improve the programe. Is there any p5.js function that can help me achieve that?<\/p>\n","answers":[{"AnswerId":"58103496","CreationDate":"2019-09-25T17:12:57.743","ParentId":null,"OwnerUserId":"89766","Title":null,"Body":"<p>It depends in what format the CNN accepts input.<\/p>\n\n<p>The simplest thing I can think of is using plain JavaScript (outside of p5.js) to access the <code>&lt;canvas \/&gt;<\/code> element.<\/p>\n\n<p>For example this is something you can try in your browser console on the sketch_rnn_demo page:<\/p>\n\n<pre><code>\/\/ access the default p5.js Canvas\ncanvasElement = document.querySelector('#defaultCanvas0')\n\/\/ export the data as needed, for example encoded as a Base64 string:\ncanvasElement.toDataURL()\n<\/code><\/pre>\n\n<p>If you want to access pixels, you can via the <a href=\"https:\/\/developer.mozilla.org\/en-US\/docs\/Web\/API\/CanvasRenderingContext2D\" rel=\"nofollow noreferrer\">Canvas context<\/a> and <a href=\"https:\/\/developer.mozilla.org\/en-US\/docs\/Web\/API\/CanvasRenderingContext2D\/getImageData\" rel=\"nofollow noreferrer\">getImageData()<\/a>:<\/p>\n\n<pre><code>\/\/access &lt;canvas\/&gt; context\nvar context = canvasElement.getContext('2d');\n\/\/access pixels:\ncontext.getImageData(0,0,canvasElement.width,canvasElement.height);\n<\/code><\/pre>\n\n<p>This will return a 1D array of unsigned 8-bit integers (e.g. values from 0-255) in R,G,B,A order\n(e.g. pixel0R,pixel0G,pixel0B,pixel0A,pixel1R,pixel1G,pixel1B,pixel1A...etc.)<\/p>\n\n<p>If you want to use p5.js instead, call <a href=\"https:\/\/p5js.org\/reference\/#\/p5\/loadPixels\" rel=\"nofollow noreferrer\">loadPixels()<\/a> first, then access the <a href=\"https:\/\/p5js.org\/reference\/#\/p5\/pixels\" rel=\"nofollow noreferrer\">pixels[]<\/a> array which is the same format as above.<\/p>\n\n<p>You can also use <a href=\"https:\/\/p5js.org\/reference\/#\/p5\/get\" rel=\"nofollow noreferrer\">get(x,y)<\/a> in p5.js which allows a 2D way to access to pixel data, however this is much slower.<\/p>\n\n<p>If you CNN takes in a 2D array you still need to create this 2D array yourself and populate it pixel values (using <code>pixels[]<\/code> or <code>get()<\/code> for example). Be sure to double check the CNN input: <\/p>\n\n<ul>\n<li>it is a 2D array of 32-bit integers (e.g. R,G,B,A or A,R,G,B as a single int (0xAARRGGBB or 0xRRGGBBAA), just RGB, etc.)<\/li>\n<li>what resolution should the 2d array be ? (your sketch-rnn canvas may be a different size and you might need to resize it to match what the CNN expects as an input)<\/li>\n<\/ul>\n"}]}
+{"QuestionId":58103022,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><weighted>","CreationDate":"2019-09-25T16:42:20.077","AcceptedAnswerId":null,"OwnerUserId":11243873.0,"Title":"Incompatible shapes error when using tf.feature_column.weighted_categorical_column","Body":"<p>I use tensorflow 1.13 to build a linear classifier with a categorical action value weighted by an action power value (both are included in my CSV data file). I tried to build a weighted categorical column feature as following :<\/p>\n\n<pre><code>    import tensorflow as tf\n\n      Player_1_Action = tf.feature_column.categorical_column_with_vocabulary_list(key=\"Player_1_Action\", vocabulary_list=Action_vocab2)\n\n      feature_columns_complex_standardized = [\n          (...)\n          tf.feature_column.weighted_categorical_column(Player_1_Action,\"Player_5_Power\",dtype=tf.dtypes.float32),\n          ]\n<\/code><\/pre>\n\n<p>When i start training python stops and returns : <\/p>\n\n<p>InvalidArgumentError (see above for traceback): Incompatible shapes: [32] vs. [2\n8]<\/p>\n\n<p>Is there a solution for that ?<\/p>\n","answers":[]}
+{"QuestionId":58103035,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-09-25T16:43:12.297","AcceptedAnswerId":58103272.0,"OwnerUserId":4186233.0,"Title":"Keras Early Stopping with \"train_on_batch\"","Body":"<p>This question seems to be very basic, yet I didn't find anything useful on google.<\/p>\n\n<p>I manually run the epochs in the loop, as well as further nested mini batches in the loop.\nAt each mini batch I need to call train_on_batch, to enable the training of a customized model.<\/p>\n\n<p>Is there a manual way to restore the functionality of early stopping, i.e. to break the loop?<\/p>\n","answers":[{"AnswerId":"58103272","CreationDate":"2019-09-25T16:57:57.633","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>In practice, 'early stopping' is largely done via: (1) train for X epochs, (2) <em>save<\/em> the model each time it achieves a new best performance, (3) select the best model. \"Best performance\" defined as achieving the highest (e.g. accuracy) or lowest (e.g. loss) validation metric - example script below:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>best_val_loss = 999 # arbitrary init - should be high if 'best' is low, and vice versa\nnum_epochs = 5\nepoch = 0\n\nwhile epoch &lt; num_epochs:\n    model.train_on_batch(x_train, y_train)  # get x, y somewhere in the loop\n    val_loss = model.evaluate(x_val, y_val)\n\n    if val_loss &lt; best_val_loss:\n        model.save(best_model_path) # OR model.save_weights()\n        print(\"Best model w\/ val loss {} saved to {}\".format(val_loss, best_model_path))\n    # ...\n    epoch += 1\n<\/code><\/pre>\n\n<p>See <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-save-a-keras-model\" rel=\"nofollow noreferrer\">saving Keras models<\/a>. If you rather early-stop directly, then define some metric - i.e. condition - that'll end the train loop. For example,<\/p>\n\n<pre><code>while True:\n    loss = model.train_on_batch(...)\n    if loss &lt; .02:\n        break\n<\/code><\/pre>\n"}]}
+{"QuestionId":58103128,"AnswerCount":1,"Tags":"<tensorflow><pytorch><tensor>","CreationDate":"2019-09-25T16:49:13.593","AcceptedAnswerId":null,"OwnerUserId":8919443.0,"Title":"Finding subtraction of shifted tensor","Body":"<p>I'm trying to figure out how to do shifting on a tensor that has <code>b (batch size), d (depth), h (hight) and w (width)<\/code> represented as following:<\/p>\n\n<pre><code>b, d, h, w = tensor.size()\n<\/code><\/pre>\n\n<p>So, I need to find the subtract between the shifted tensor and the tensor itself.<\/p>\n\n<p>I'm thinking of using <code>torch.narrow<\/code> or <code>torch.concat<\/code> to do it for each side (shift the right, left, up then down side) and at each time I subtract from the same tensor side (tensor itself side), then at the end I will add\/sum the differences\/subtractions of each side (so I will have the final subtraction between the shifted and the tensor itself.<\/p>\n\n<p>I'm new to PyTorch, it's easy to understand but struggling to implemented and maybe there is a simpler way (directly do the subtraction rather than working on each side and so on .....)<\/p>\n\n<p>Any help on that please?<\/p>\n","answers":[{"AnswerId":"58107910","CreationDate":"2019-09-26T00:00:45.550","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>Basically, you can split the tensor first, and then cat them in reverse order. I write a function to implement your thoughts. The <code>shift<\/code> should be a non-negative number and less than or equal to the size of <code>dim<\/code>.  <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def tensor_shift(t, dim, shift):\n    \"\"\"\n    t (tensor): tensor to be shifted. \n    dim (int): the dimension apply shift.\n    shift (int): shift distance.\n    \"\"\"\n    assert 0 &lt;= shift &lt;= t.size(dim), \"shift distance should be smaller than or equal to the dim length.\"\n\n    overflow = t.index_select(dim, torch.arange(t.size(dim)-shift, t.size(dim)))\n    remain = t.index_select(dim, torch.arange(t.size(dim)-shift))\n\n    return torch.cat((overflow, remain),dim=dim)\n<\/code><\/pre>\n\n<p>Here are some test results. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>a = torch.arange(1,13).view(-1,3)\na\n#tensor([[ 1,  2,  3],\n#        [ 4,  5,  6],\n#        [ 7,  8,  9],\n#        [10, 11, 12]])\n\nshift(a, 0, 1) # shift 1 unit along dim=0\n#tensor([[10, 11, 12],\n#        [ 1,  2,  3],\n#        [ 4,  5,  6],\n#        [ 7,  8,  9]])\n\nb = torch.arange(1,13).view(-1,2,3)\nb\n#tensor([[[ 1,  2,  3],\n#         [ 4,  5,  6]],\n#\n#        [[ 7,  8,  9],\n#         [10, 11, 12]]])\n\nshift(b, 1, 1) # shift 1 unit along dim=1\n#tensor([[[ 4,  5,  6],\n#         [ 1,  2,  3]],\n#\n#        [[10, 11, 12],\n#         [ 7,  8,  9]]])\n<\/code><\/pre>\n"}]}
+{"QuestionId":58103200,"AnswerCount":1,"Tags":"<keras><deep-learning><lstm><lstm-stateful>","CreationDate":"2019-09-25T16:53:04.363","AcceptedAnswerId":58103321.0,"OwnerUserId":1449734.0,"Title":"Training many-to-many stateful LSTM with and without final dense layer","Body":"<p>I am trying to train a recurrent model in Keras containing an LSTM for regression purposes.\nI would like to use the model online and, as far as I understood, I need to train a stateful LSTM. \nSince the model has to output a sequence of values, I hope it computes the loss on each of the expected output vector.\nHowever, I fear my code is not working this way and I would be grateful if anyone would help me to understand if I am doing right or if there is some better approach.<\/p>\n\n<p>The input to the model is a sequence of 128-dimensional vectors. Each sequence in the training set has a different lenght.\nAt each time, the model should output a vector of 3 elements. <\/p>\n\n<p>I am trying to train and compare two models:\nA) a simple LSTM with 128 inputs and 3 outputs;\nB) a simple LSTM with 128 inputs and 100 outputs + a dense layer with 3 outputs;<\/p>\n\n<p>For model A) I wrote the following code:<\/p>\n\n<pre><code># Model\nmodel = Sequential()\nmodel.add(LSTM(3, batch_input_shape=(1, None, 128),  return_sequences=True, activation = \"linear\", stateful = True))`\nmodel.compile(loss='mean_squared_error', optimizer=Adam())\n\n# Training\nfor i in range(n_epoch):\n    for j in np.random.permutation(n_sequences):\n        X = data[j] # j-th sequences\n        X = X[np.newaxis, ...] # X has size 1 x NTimes x 128\n\n        Y = dataY[j] # Y has size NTimes x 3\n\n        history = model.fit(X, Y, epochs=1, batch_size=1, verbose=0, shuffle=False)\n        model.reset_states()\n<\/code><\/pre>\n\n<p>With this code, model A) seems to train fine because the output sequence approaches the ground-truth sequence on the training set. \nHowever, I wonder if the loss is really computed by considering all NTimes output vectors.<\/p>\n\n<p>For model B), I could not find any way to get the entire output sequence due to the dense layer. Hence, I wrote:<\/p>\n\n<pre><code># Model\nmodel = Sequential()\nmodel.add(LSTM(100, batch_input_shape=(1, None, 128), , stateful = True))\nmodel.add(Dense(3,   activation=\"linear\"))\nmodel.compile(loss='mean_squared_error', optimizer=Adam())\n\n# Training\nfor i in range(n_epoch):\n    for j in np.random.permutation(n_sequences):\n        X = data[j]  #j-th sequence\n        X = X[np.newaxis, ...] # X has size 1 x NTimes x 128\n\n        Y = dataY[j] # Y has size NTimes x 3\n\n        for h in range(X.shape[1]):\n            x = X[0,h,:]\n            x = x[np.newaxis, np.newaxis, ...] # h-th vector in j-th sequence\n            y = Y[h,:]\n            y = y[np.newaxis, ...]\n            loss += model.train_on_batch(x,y)\n        model.reset_states() #After the end of the sequence\n<\/code><\/pre>\n\n<p>With this code, model B) does not train fine. It seems to me the training does not converge and loss values increase and decrease cyclically\nI have also tried to use as Y only the last vector and them calling the fit function on the Whole training sequence X, but no improvements.<\/p>\n\n<p>Any idea? Thank you!<\/p>\n","answers":[{"AnswerId":"58103321","CreationDate":"2019-09-25T17:00:42.377","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<p>If you want to still have three outputs per step of your sequence, you need to TimeDistribute your Dense layer like so:<\/p>\n\n<pre><code>model.add(TimeDistributed(Dense(3, activation=\"linear\")))\n<\/code><\/pre>\n\n<p>This applies the dense layer to each timestep independently.<\/p>\n\n<p>See <a href=\"https:\/\/keras.io\/layers\/wrappers\/#timedistributed\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/wrappers\/#timedistributed<\/a><\/p>\n"}]}
+{"QuestionId":58104147,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-25T18:00:48.483","AcceptedAnswerId":null,"OwnerUserId":11666502.0,"Title":"Tensorflow error: ValueError: Shapes must be equal rank, but are 1 and 2","Body":"<p>This is a snippet of a tensorflow model I am trying to run but keep getting errors. I am using tensorflow 1.14.0 and I believe the author of the code used 1.08, if that makes a difference.<\/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>def select(parameters, length):\r\n    \"\"\"Select the last valid time step output as the sentence embedding\r\n    :params parameters: [batch, seq_len, hidden_dims]\r\n    :params length: [batch]\r\n    :Returns : [batch, hidden_dims]\r\n    \"\"\"\r\n    shape = tf.shape(parameters)\r\n    idx = tf.range(shape[0])\r\n    idx = tf.stack([idx, length - 1], axis = 1)\r\n    return tf.gather_nd(parameters, idx)\r\n\r\nhidden_size_lstm_1 = 200\r\nhidden_size_lstm_2 = 200\r\ntags = 39\r\nword_dim = 300\r\nproj1 = 200\r\nproj2 = 100\r\nwords = 20001\r\nbatchSize = 2\r\nlog_dir = \"train\"\r\nmodel_dir = \"DAModel\"\r\nmodel_name = \"ckpt\"\r\n\r\nclass DAModel():\r\n    def __init__(self):\r\n        with tf.variable_scope(\"placeholder\"):\r\n\r\n            self.dialogue_lengths = tf.placeholder(tf.int32, shape = [None], name = \"dialogue_lengths\")\r\n            self.word_ids = tf.placeholder(tf.int32, shape = [None,None,None], name = \"word_ids\")\r\n            self.utterance_lengths = tf.placeholder(tf.int32, shape = [None, None], name = \"utterance_lengths\")\r\n            # self.utterance_lengths = tf.placeholder(tf.float32, shape = [None, None], name = \"utterance_lengths\")\r\n            self.labels = tf.placeholder(tf.int32, shape = [None, None], name = \"labels\")\r\n            self.clip = tf.placeholder(tf.float32, shape = [], name = 'clip')\r\n        \r\n        with tf.variable_scope(\"embeddings\"):\r\n            _word_embeddings = tf.get_variable(\r\n                name = \"_word_embeddings\",\r\n                dtype = tf.float32,\r\n                shape = [words, word_dim],\r\n                initializer = tf.random_uniform_initializer()\r\n                )\r\n            word_embeddings = tf.nn.embedding_lookup(_word_embeddings,self.word_ids, name=\"word_embeddings\")\r\n            self.word_embeddings = tf.nn.dropout(word_embeddings, 0.8)\r\n                    \r\n        with tf.variable_scope(\"utterance_encoder\"):\r\n            s = tf.shape(self.word_embeddings)\r\n            batch_size = s[0] * s[1]\r\n            \r\n            time_step = s[-2]\r\n            word_embeddings = tf.reshape(self.word_embeddings, [batch_size, time_step, word_dim])\r\n            length = tf.reshape(self.utterance_lengths, [batch_size])\r\n\r\n            fw = tf.nn.rnn_cell.LSTMCell(hidden_size_lstm_1, forget_bias=0.8, state_is_tuple= True)\r\n            bw = tf.nn.rnn_cell.LSTMCell(hidden_size_lstm_1, forget_bias=0.8, state_is_tuple= True)\r\n            \r\n            output, _ = tf.nn.bidirectional_dynamic_rnn(fw, bw, word_embeddings,sequence_length=length, dtype = tf.float32)\r\n            output = tf.concat(output, axis = -1) # [batch_size, time_step, dim]\r\n\r\n            # Select the last valid time step output as the utterance embedding, \r\n            # this method is more concise than TensorArray with while_loop\r\n            output = select(output, self.utterance_lengths) # [batch_size, dim]\r\n            output = tf.reshape(output, s[0], s[1], 2 * hidden_size_lstm_1)\r\n            output = tf.nn.dropout(output, 0.8)<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>I keep getting this error:<\/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>Traceback (most recent call last):\r\n  File \"HBLSTM-CRF.py\", line 308, in &lt;module&gt;\r\n    main()\r\n  File \"HBLSTM-CRF.py\", line 251, in main\r\n    model = DAModel()\r\n  File \"HBLSTM-CRF.py\", line 134, in __init__\r\n    output = select(output, self.utterance_lengths) # [batch_size, dim]\r\n  File \"HBLSTM-CRF.py\", line 82, in select\r\n    idx = tf.stack([idx, length - 1], axis = 1)\r\n  File \"C:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\util\\dispatch.py\", line 180, in wrapper\r\n    return target(*args, **kwargs)\r\n  File \"C:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\ops\\array_ops.py\", line 1046, in stack\r\n    return gen_array_ops.pack(values, axis=axis, name=name)\r\n  File \"C:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\ops\\gen_array_ops.py\", line 6971, in pack\r\n    \"Pack\", values=values, axis=axis, name=name)\r\n  File \"C:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\op_def_library.py\", line 788, in _apply_op_helper\r\n    op_def=op_def)\r\n  File \"C:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\util\\deprecation.py\", line 507, in new_func\r\n    return func(*args, **kwargs)\r\n  File \"C:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 3616, in create_op\r\n    op_def=op_def)\r\n  File \"C:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 2027, in __init__\r\n    control_input_ops)\r\n  File \"C:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\ops.py\", line 1867, in _create_c_op\r\n    raise ValueError(str(e))\r\nValueError: Shapes must be equal rank, but are 1 and 2\r\n        From merging shape 0 with other shapes. for 'utterance_encoder\/stack' (op: 'Pack') with input shapes: [?], [?,?].<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>I have tried to fix this by changing 'utterance_lengths' to float32, thinking that it was erroring when trying to stack a float32 and int32, but this has not worked. Any recommendations would be appreciated. If it would aid in diagnosing the problem, the entire code can be found in this link below:<\/p>\n\n<p><a href=\"https:\/\/github.com\/YanWenqiang\/HBLSTM-CRF\/blob\/master\/HBLSTM-CRF.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/YanWenqiang\/HBLSTM-CRF\/blob\/master\/HBLSTM-CRF.py<\/a><\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
+{"QuestionId":58104491,"AnswerCount":0,"Tags":"<python><tensorflow><autoencoder><valueerror>","CreationDate":"2019-09-25T18:25:19.783","AcceptedAnswerId":null,"OwnerUserId":11547121.0,"Title":"Value Error in python tensorflow variable scope trying to share variable","Body":"<p>I am trying to train a variational autoencoder, and I want to train it each time with different number of latent space dimensions using a for loop. Inside that loop I have another loop to iterate over different coefficients for KL parameter, so I had to set all reuse flags as True.\nIt's Ok when I run the code with only one latent space dimension but when I write the for loop over it, I get this error:<\/p>\n\n<blockquote>\n  <p>ValueError: Trying to share variable gaussian_MLP_encoder\/wo, but specified shape (500, 8) and found shape (500, 4).<\/p>\n<\/blockquote>\n\n<p>Also it might be useful to know that n_hidden which is the number of hidden units in my encoder and decoder layers is equal to 500, and the for loop which causes the error is:<\/p>\n\n<pre><code>    dim_z = [2, 4, 6, 8, 10, 12, 14, 16, 18, 20]\n    for s in range(len(dim_z)):\n<\/code><\/pre>\n\n<p>Here is my code where get this error for:<\/p>\n\n<pre><code>wo = tf.get_variable('wo', [h1.get_shape()[1], n_output * 2], initializer=w_init)\n<\/code><\/pre>\n\n<p>in :<\/p>\n\n<pre><code>def gaussian_MLP_encoder(x, n_hidden, n_output, keep_prob):\n    with tf.variable_scope(\"gaussian_MLP_encoder\", reuse=tf.AUTO_REUSE):\n        # initializers\n        w_init = tf.contrib.layers.variance_scaling_initializer()\n        b_init = tf.constant_initializer(0.)\n\n        # 1st hidden layer\n        w0 = tf.get_variable('w0', [x.get_shape()[1], n_hidden], initializer=w_init)\n        b0 = tf.get_variable('b0', [n_hidden], initializer=b_init)\n        h0 = tf.matmul(x, w0) + b0\n        h0 = tf.nn.elu(h0)\n        h0 = tf.nn.dropout(h0, keep_prob)\n\n        # 2nd hidden layer\n        w1 = tf.get_variable('w1', [h0.get_shape()[1], n_hidden], initializer=w_init)\n        b1 = tf.get_variable('b1', [n_hidden], initializer=b_init)\n        h1 = tf.matmul(h0, w1) + b1\n        h1 = tf.nn.tanh(h1)\n        h1 = tf.nn.dropout(h1, keep_prob)\n\n        # output layer\n        # borrowed from https: \/\/ github.com \/ altosaar \/ vae \/ blob \/ master \/ vae.py\n        wo = tf.get_variable('wo', [h1.get_shape()[1], n_output * 2], initializer=w_init)\n        bo = tf.get_variable('bo', [n_output * 2], initializer=b_init)\n        gaussian_params = tf.matmul(h1, wo) + bo\n\n        # The mean parameter is unconstrained\n        mean = gaussian_params[:, :n_output]\n        # The standard deviation must be positive. Parametrize with a softplus and\n        # add a small epsilon for numerical stability\n        stddev = 1e-6 + tf.nn.softplus(gaussian_params[:, n_output:])\n\n    return mean, stddev\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58104512,"AnswerCount":0,"Tags":"<python><keras><neural-network><hyperparameters><talos>","CreationDate":"2019-09-25T18:25:57.927","AcceptedAnswerId":null,"OwnerUserId":10619823.0,"Title":"It's possible load a dynamic best model in Keras: EarlyStoping + Talos?","Body":"<p>I have a simple neural net and I'm optimicing the learning rate. The problem is that i'm using talos for change the learning rate parameter and i use earlystopping saving the weights of the best model. Then, Talos continue his loop over the best model, improving the metrics but i dont know if it's posible to save this model and load again.<\/p>\n\n<p>I tested a model.evaluate with the weights of the best model saved in talos but the results are poor. Are there some method for load the configuration? It's a dinamic configuration always over the best model<\/p>\n","answers":[]}
+{"QuestionId":58104698,"AnswerCount":1,"Tags":"<python><tensorflow><keras><neural-network><deep-learning>","CreationDate":"2019-09-25T18:40:09.217","AcceptedAnswerId":58105247.0,"OwnerUserId":11918892.0,"Title":"How to add custom metric to keras? (Percent Mean Absolute Error)","Body":"<p>I am trying to add <strong>percent mean absolute error<\/strong> (pmae) as a custom metric in keras. This is defined as (MAE divided by the mean absolute y-value multiplied by 100). I have tried:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def pmae(y_true,y_pred):\n    return K.mean(K.abs(y_pred - y_true)) \/ K.mean(K.abs(y_true)) * 100\n...\nmodel.compile(loss='mse', optimizer=Adam(),metrics=[pmae])\n<\/code><\/pre>\n\n<p>which runs, but the value is many orders of magnitude off (when I look at <code>model.history.history.pmae<\/code>)<\/p>\n\n<p>The working numpy version (on a test sample) is:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>y_pred = model.predict(X_test)\npmae = abs(y_pred - y_test).mean() \/ abs(y_true).mean() * 100\n<\/code><\/pre>\n\n<p>I've also tried adding <code>, axis=-1<\/code> to the <code>K.mean()<\/code> calls, with no improvement (as suggested in other stackoverflow answers). Does anyone know what's wrong?<\/p>\n\n<p><strong>Resources<\/strong><\/p>\n\n<ol>\n<li>The <a href=\"https:\/\/keras.io\/metrics\/#custom-metrics\" rel=\"nofollow noreferrer\">keras website<\/a> gives mean value of y as an example:<\/li>\n<\/ol>\n\n<pre class=\"lang-py prettyprint-override\"><code>import keras.backend as K\n\ndef mean_pred(y_true, y_pred):\n    return K.mean(y_pred)\n\nmodel.compile(optimizer='rmsprop',\n              loss='binary_crossentropy',\n              metrics=['accuracy', mean_pred])\n<\/code><\/pre>\n\n<ol start=\"2\">\n<li>Others have answered on stackoverflow about other custom metrics (eg <a href=\"https:\/\/stackoverflow.com\/q\/47582982\/11918892\">Keras custom RMSLE metric<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/37657260\/how-to-implement-custom-metric-in-keras\">how to implement custom metric in keras?<\/a>), but the responses there haven't helped me with pmae calculation.<\/li>\n<\/ol>\n","answers":[{"AnswerId":"58105247","CreationDate":"2019-09-25T19:21:22.470","ParentId":null,"OwnerUserId":"7469434","Title":null,"Body":"<p>Let's compare your implementation with the <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/losses.py#L616\" rel=\"nofollow noreferrer\"><code>mean_absolute_percentage_error<\/code><\/a> in Keras:<\/p>\n\n<pre><code>def mean_absolute_percentage_error(y_true, y_pred):\n    if not K.is_tensor(y_pred):\n        y_pred = K.constant(y_pred)\n    y_true = K.cast(y_true, y_pred.dtype)\n    diff = K.abs((y_true - y_pred) \/ K.clip(K.abs(y_true),\n                                            K.epsilon(),\n                                            None))\n    return 100. * K.mean(diff, axis=-1)\n<\/code><\/pre>\n\n<p>Based on this, the following should work for your case:<\/p>\n\n<pre><code>def percent_mean_absolute_error(y_true, y_pred):\n    if not K.is_tensor(y_pred):\n        y_pred = K.constant(y_pred)\n    y_true = K.cast(y_true, y_pred.dtype)\n    diff = K.mean(K.abs((y_pred - y_true)) \/ K.mean(K.clip(K.abs(y_true),\n                                                           K.epsilon(),\n                                                           None)))\n    return 100. * K.mean(diff)\n<\/code><\/pre>\n\n<p>The main difference to your attempt is that here both <code>y_true<\/code> and <code>y_pred<\/code> are cast to the same datatype and the denominator is at least <code>K.epsilon()<\/code> (which is <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/backend\/common.py#L9\" rel=\"nofollow noreferrer\">set to <code>1e-7<\/code> by default<\/a>), so the error does not go to infinity if <code>y_true<\/code> approaches <code>0<\/code>.<\/p>\n"}]}
+{"QuestionId":58105073,"AnswerCount":4,"Tags":"<python><pytorch>","CreationDate":"2019-09-25T19:07:07.867","AcceptedAnswerId":null,"OwnerUserId":2769240.0,"Title":"Splitting a directory with images into sub folders using Pytorch or Python","Body":"<p>I have a directory with two sub directories in it. Each sub-dir has bunch of images. These sub-dir also specify the two classes of images. <\/p>\n\n<p>I want to have 3 directories (train, validation, test) and within each of these 3 sub dir, I want to 2 sub directory of each class respectively with images. <\/p>\n\n<p>I want to split the no.of images into train, val, test directories by random sampling. SO some 60% images go to train, 20, val and 20 in test. <\/p>\n\n<p>Initial Structure: <\/p>\n\n<pre><code>Main_folder\n - Good\n - Bad \n<\/code><\/pre>\n\n<p>What I want: <\/p>\n\n<pre><code>Main_folder: \n - Train \n  - Good\n  - Bad\n - Val\n  - Good \n  - Bad\n - Test \n  - Good \n  - Bad\n<\/code><\/pre>\n\n<p>I want to split each sub directories into two directories within it with random assignment of images (<\/p>\n","answers":[{"AnswerId":"58105213","CreationDate":"2019-09-25T19:18:56.897","ParentId":null,"OwnerUserId":"12121150","Title":null,"Body":"<p>There are two parts of this problem:<\/p>\n\n<ol>\n<li>Given a list of files, find an algorithm that randomly generates subsets of given proportion.<\/li>\n<li>Shuffle the actual files around based on the subsets generated.<\/li>\n<\/ol>\n\n<p>Two is a trivial use case of <code>os<\/code> package and answered elsewhere. For 1, the algorithm is (pseudocode):<\/p>\n\n<pre><code>for d in [Good, Bad]:\n    n['all'] = total images in d\n    for c in [Train, Val, Test]:\n        if not last_item:\n            n[c] = round(n['all'] * proportion[c])\n        else:\n            n[c] = n['all'] - all other n's\n\n    shuffle list of all images in d\n    for c in [Train, Val, Test]:\n        take first n[c] images and put them in c\n        remove them from d\n<\/code><\/pre>\n\n<p>In this case, the Good\/Bad ratio will be the same in every subset as in the original input, which is not a bad idea. The split can be done in other ways, but that is a broad discussion I won't get into here. However, you might need at least 100 images or more for this to work - splitting just 40 images into 6 subsets doesn't really make much sense, they'd be too small.<\/p>\n"},{"AnswerId":"58113503","CreationDate":"2019-09-26T09:19:44.620","ParentId":null,"OwnerUserId":"10886420","Title":null,"Body":"<p>There is a <code>pytorch<\/code> way to do that (and I would advise using single library for such easy task).<\/p>\n\n<h2>Creating dataset<\/h2>\n\n<p>There is a pre-made one inside <code>torchvision<\/code>, namely <a href=\"https:\/\/pytorch.org\/docs\/stable\/torchvision\/datasets.html#imagefolder\" rel=\"nofollow noreferrer\"><code>ImageFolder<\/code><\/a>.<\/p>\n\n<p>Simply use it like this:<\/p>\n\n<pre><code>import torchvision\n\ndataset = torchvision.datasets.ImageFolder(\".\/my_data\")\n<\/code><\/pre>\n\n<p>This will create a dataset, where <code>Good<\/code> folder has <code>0<\/code> label and <code>Bad<\/code> has label of <code>1<\/code> for each image in those folder respectively.<\/p>\n\n<h2>Dividing into train, validation, test<\/h2>\n\n<p>No need for <code>sklearn<\/code> for such an easy job + <code>torch<\/code> has most of the <code>numpy<\/code> functionality as well and I would rather stick with one library instead of 3 (though <code>train_test_split<\/code> could be used twice consecutively, similar to what <a href=\"https:\/\/stackoverflow.com\/questions\/58105073\/splitting-a-directory-with-images-into-sub-folders-using-pytorch-or-python\/58105788#58105788\">@ESZ<\/a> proposed).<\/p>\n\n<p>IMO simpler would be:<\/p>\n\n<pre><code>def get_subset(indices, start, end):\n    return indices[start : start + end]\n\n\nTRAIN_PCT, VALIDATION_PCT = 0.6, 0.2  # rest will go for test\ntrain_count = int(len(dataset) * TRAIN_PCT)\nvalidation_count = int(len(dataset) * TRAIN_PCT)\n\nindices = torch.randperm(len(dataset))\n\ntrain_indices = get_subset(indices, 0, train_count)\nvalidation_indices = get_subset(indices, train_count, validation_count)\ntest_indices = get_subset(indices, train_count + validation_count, len(dataset))\n<\/code><\/pre>\n\n<p>This will create indices for <code>SubsetRandomSampler<\/code> and <code>torch.utils.data.DataLoader<\/code>.\nSo, similarly to <a href=\"https:\/\/stackoverflow.com\/questions\/58105073\/splitting-a-directory-with-images-into-sub-folders-using-pytorch-or-python\/58105788#58105788\">@ESZ<\/a> once again:<\/p>\n\n<pre><code>dataloaders = {\n    \"train\": torch.utils.data.DataLoader(\n        dataset, sampler=SubsetRandomSampler(train_indices)\n    ),\n    \"validation\": torch.utils.data.DataLoader(\n        dataset, sampler=SubsetRandomSampler(validation_indices)\n    ),\n    \"test\": torch.utils.data.DataLoader(\n        dataset, sampler=SubsetRandomSampler(test_indices)\n    ),\n}\n<\/code><\/pre>\n\n<p>You can specify <code>batch_size<\/code> and other argument to <code>DataLoader<\/code>, see <a href=\"https:\/\/pytorch.org\/docs\/stable\/data.html#torch.utils.data.DataLoader\" rel=\"nofollow noreferrer\">documentation<\/a> if you need more info.<\/p>\n"},{"AnswerId":"58105788","CreationDate":"2019-09-25T20:00:00.990","ParentId":null,"OwnerUserId":"7872285","Title":null,"Body":"<p>You can use the SKLearn's train_test_split + SubSetRandomSampler from torch as follow :    <\/p>\n\n<pre><code>from torch.utils.data.sampler import SubsetRandomSampler\nfrom sklearn.model_selection import train_test_split\nfrom torch.utils.data.sampler import SubsetRandomSampler\nvalid_size_and_test = 0.4 # then I will keep 0.1 for testing \n\n## split the dataset into train and rest data\ntargets = dataset.targets\n\ntrain_idx, rest_idx = train_test_split(np.arange(len(targets)), test_size= \n                    valid_size_and_test, random_state=42, shuffle=True, stratify=targets)\ntest_size = int(len(rest_idx) * valid_size_and_test)\nvalid_idx, test_idx = rest_idx[:test_size], rest_idx[test_size:]\n\ndataloaders = {'trainLoader' : torch.utils.data.DataLoader(dataset,batch_size=32,sampler=SubsetRandomSampler(train_idx)),\n               'validLoader' : torch.utils.data.DataLoader(dataset,batch_size=32,sampler=SubsetRandomSampler(valid_idx)),\n               'testLoader' :  torch.utils.data.DataLoader(dataset,batch_size=32,sampler=SubsetRandomSampler(test_idx)),\n              }\n<\/code><\/pre>\n"},{"AnswerId":"58105163","CreationDate":"2019-09-25T19:14:40.220","ParentId":null,"OwnerUserId":"12118774","Title":null,"Body":"<p>Well I'm not sure about the Good-Bad stuff but I'm pretty sure that you can use scipy train-test split function.<\/p>\n\n<p>You can find detailed information in <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.model_selection.train_test_split.html\" rel=\"nofollow noreferrer\">sklearn.model_selection.train_test_split<\/a>\nI guess it only creates two samples of image sets like train-test. However, you can divide the test part again to achieve your goal. Also, as I remember you can set the train-test percentage in the function as well.<\/p>\n"}]}
+{"QuestionId":58105126,"AnswerCount":0,"Tags":"<python><machine-learning><keras><classification><lstm-stateful>","CreationDate":"2019-09-25T19:12:06.517","AcceptedAnswerId":null,"OwnerUserId":10952488.0,"Title":"Video classification using stateful LSTM , validation accuracy doesn't improve","Body":"<p>I am trying to classify some video into 3 different classes. Each video has different length of frame. The training data has the shape of (104, None, 528) where: <\/p>\n\n<ul>\n<li>104 = Number of videos<\/li>\n<li>None = number of frames for each video which are different<\/li>\n<li>528 = Number of features for each frame<\/li>\n<\/ul>\n\n<p>As the sequence of frames for each video is long I am using \"stateful LSTM\" to manage the length of sequences. I have defined my model same as below:<\/p>\n\n<pre><code>def LSTM_Model():\n\nmodel = Sequential()\nmodel.add(LSTM(units = 256, input_shape=(None, 528),\\\n                       return_sequences=False, stateful=True, batch_size = 1))\nmodel.add(Dropout(0.4))\nmodel.add(Dense(3, activation='softmax'))\nopt = keras.optimizers.SGD(lr=0.00005, decay = 1e-6, momentum=0.9, nesterov=True)\nmodel.compile(loss = 'categorical_crossentropy', optimizer = opt, metrics = ['accuracy'])\nmodel.summary()\n\nreturn model\n<\/code><\/pre>\n\n<p>Then I trained and tested the model:<\/p>\n\n<pre><code>def train_model(X, y, X_test, y_test, model):\n    np.random.seed(200)\n    epochs = 100\n    maxlen = 500\n\n\n    for epoch in range(epochs):\n\n        mean_tr_loss, mean_tr_acc =[],[]\n        print('Epoch: ', epoch + 1 )\n\n        for sbj in range(X.shape[0]):\n\n            video = X[sbj]\n            y_sbj = y[sbj,:]\n            y_new = y_sbj\n            nb_frame = video.shape[0]\n\n            for count in range(nb_frame \/\/ maxlen +1):\n\n                if count == nb_frame \/\/ maxlen :\n                    seq = video[count*maxlen + count:, :]\n\n                else:\n                    seq = video[count*maxlen+count : (count+1)*maxlen+count, :]\n                    seq = np.expand_dims(seq, axis=0)\n\n      #   ''' Using train_on_batch '''\n\n                    tr_loss, tr_acc = model.train_on_batch(seq, np.array([y_new])) \n                    mean_tr_loss.append(tr_loss)\n                    mean_tr_acc.append(tr_acc)\n\n\n            print('Training on subject', sbj+1, 'done' )\n            model.reset_states() \n\n\n        print('accuracy training = {}'.format(np.mean(mean_tr_acc)))\n        print('loss training = {}'.format(np.mean(mean_tr_loss)))\n        print('___________________________________')\n\n\n\n        print('Testing....')\n\n\n        mean_te_loss, mean_te_acc =[],[]\n\n        for sbj_test in range(X_test.shape[0]):\n\n            video_test = X_test[sbj_test]\n            y_new_test = y_test[sbj_test]\n            nb_frame_test = video_test.shape[0]\n\n            for i in range(nb_frame_test \/\/ maxlen + 1):\n\n                if i == nb_frame_test \/\/ maxlen :\n                    seq_test = video_test[i*maxlen + i:, :]\n                else:\n                    seq_test = video_test[i*maxlen+i : (i+1)*maxlen+i, :]\n                    seq_test = np.expand_dims(seq_test, axis=0)\n                    te_loss, te_acc = model.test_on_batch(seq_test, np.array([y_new_test])) \n                    mean_te_loss.append(te_loss)\n                    mean_te_acc.append(te_acc)                \n            print('Testing on subject', sbj_test+1, 'done' )\n            model.reset_states()   \n\n\n\n        print('accuracy testing = {}'.format(np.mean(mean_te_acc)))\n        print('loss testing = {}'.format(np.mean(mean_te_loss)))\n<\/code><\/pre>\n\n<p>In the above code I considered each video separately and then each video was divided to different frame sequences with length 500 (except last sequence frame for each video because the length of frames are not divisible by 500). The training accuracy and test accuracy are same as below.<\/p>\n\n<pre><code>Epoch1 : accuracy training = 0.3694     accuracy testing = 0.3927\n         loss training = 1.146          loss testing = 1.109\nEpoch2 : accuracy training = 0.4423     accuracy testing = 0.4048\n         loss training = 1.053          loss testing = 1.109\nEpoch3 : accuracy training = 0.5017     accuracy testing = 0.4236\n         loss training = 0.994          loss testing = 1.115\nEpoch4 : accuracy training = 0.5491     accuracy testing = 0.4099\n         loss training = 0.94           loss testing = 1.124\nEpoch5: accuracy training = 0.5612      accuracy testing = 0.4013\n        loss training = 0.924           loss testing = 1.128\nEpoch6 : accuracy training = 0.6142     accuracy testing = 0.4113\n         loss training = 0.859          loss testing = 1.137\nEpoch7 : accuracy training = 0.6263     accuracy testing = 0.4116\n         loss training = 0.824          loss testing = 1.142\nEpoch8 : accuracy training = 0.6659     accuracy testing = 0.415\n         loss training = 0.775          loss testing = 1.152\n<\/code><\/pre>\n\n<p>After 100 epochs training accuracy increases while testing accuracy doesn't improve. If the case is \"overfitting\" adding dropout layer should help which didn't. So, I am confused about the cause. <\/p>\n\n<p>Any idea or suggestion would be appreciated.  <\/p>\n","answers":[]}
+{"QuestionId":58105563,"AnswerCount":0,"Tags":"<keras><python-imaging-library><generator><conv-neural-network><16-bit>","CreationDate":"2019-09-25T19:45:25.823","AcceptedAnswerId":null,"OwnerUserId":12120781.0,"Title":"How to make keras ImageGenerator load 16-bit images properly?","Body":"<p>I've searched for similar questions but found no solution to what i am trying to do.\nI have 16-bit grayscale images, and i am trying to fit them in keras ImageDataGenerator. When using functions like: flow_from_dataframe, it yields images all with the same pixel value (incorrect). <\/p>\n\n<p>I tried using the keras preprocess_input, rescale to [0,1], to [-1,1] with a custom preprocessing function, but none of this worked. I also set the color_mode='grayscale' in ImageDataGenerator. <\/p>\n\n<p>I further tested converting to 8-bit, and it worked. I triplicated the number of channels, and it was not the issue, since it still worked with 8-bit. I've read that keras in this case uses PIL library to read images, and since it doesn't treat 16-bit correctly, it returns that error. I saw that we could set a different library for loading images, but i don't know how to do that.<\/p>\n\n<p>Doas anyone know an alternative for using 16-bit images? In the last case, i would try a custom generator, but i would really like to profit from already built and tested functions for this purpose.<\/p>\n\n<p>I aim to use those images to fine-tune a pre-trained network, so i would like to standardize the type of input i use.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[]}
+{"QuestionId":58105702,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-25T19:54:21.843","AcceptedAnswerId":null,"OwnerUserId":6450489.0,"Title":"Setting GPU in Tensorflow with python multiprocessing module","Body":"<p>I have a system with two GPUs and want to use each GPU independently through python multiprocessing capabilities to train two different models in parallel (I don't want to use other ways of parallel learning in this case). To this end, I wrote below code:<\/p>\n\n<pre><code>def doWork(gpuID,...)\n   config = tf.ConfigProto()\n   config.gpu_options.visible_device_list = gpuID\n   with tf.Session(config=config) as sess:\n      ...\n      do tensorflow model setup and training\n      ...\n\nif _name__ == '__main__':\n   pool = Pool(processes = 2)\n   a1 = ('0',...)\n   a2 = ('1',...)\n   pool.map(doWork, (a1,a2))\n<\/code><\/pre>\n\n<p>However, Tensorflow returns me below error when running my code:<\/p>\n\n<pre><code>tensorflow.python.framework.errros_impl.InvalidArgumentError: Invalid device ordinal value (1). Valid range is [0,0].\nwhile setting up XLA_GPU_JIT device number 1\n<\/code><\/pre>\n\n<p>How can I change my code to run properly?<\/p>\n","answers":[]}
+{"QuestionId":58106068,"AnswerCount":3,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-25T20:25:05.043","AcceptedAnswerId":58119057.0,"OwnerUserId":2034650.0,"Title":"Keras ValueError: Error when checking target: expected dense_1 to have 3 dimensions","Body":"<p>I'm using keras with tensorflow backend &amp; have faced problem figuring out the right shapes for layers for my model.<\/p>\n\n<p>I have already read <a href=\"https:\/\/stackoverflow.com\/a\/44748370\">this <\/a> useful explanation on the difference of various keras layers' attributes.<\/p>\n\n<p>This is the architecture of my model:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/WLVwg.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/WLVwg.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I'm trying to do a binary classification (logistic regression) using categorical labels &amp; hence the last layer is a Dense layer with 1 unit which I assumed that would evaluate to 1 for positive class and 0 for negative class.<\/p>\n\n<p>And this is my model's summary:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/q03fN.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/q03fN.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>My input in one side of the net is 10158 &amp; for the other-side is 20316. I have total 1370 samples. My train_data's shape is (1370, 1, 10158) &amp; the label's shape is (1, 1370) &amp; the batch size is 100.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>input_layer = Input(shape=(1,no_terms), name='docs')\ns = Lambda(lambda x: x+1)(input_layer)\nlog_layer = Lambda(log, name='tf_output')(input_layer)\n\ntpr_fpr = np.zeros((2, no_terms))\ntpr_fpr[0,:] = np.sum(train_docs[np.where(train_label&gt;0), :]&gt;0, axis=1\n                      )\/np.sum(train_label&gt;0) * (1000)\ntpr_fpr[1,:] = np.sum(train_docs[np.where(train_label&gt;0), :]&gt;0, axis=1\n                     )\/np.sum(train_label &lt;= 0) * (1000)\n\nk_constants = backend.constant(np.reshape(tpr_fpr.T, (1,2*no_terms)))\nfixed_input = Input(tensor=k_constants, shape=(1, 2*no_terms), name='tpr_fpr')\nh = Dense(int(300), activation='relu', name='hidden', input_shape=(1, 2*no_terms), \n          trainable=True)(fixed_input)\nh = Dropout(0.2, name=\"D\")(h)\ncd = Dense(units=no_terms, activation='relu', name='cd', trainable=True)(h)\n\n\nprod = Multiply()([log_layer, cd])\no = Lambda(lambda x:(x\/backend.sqrt(backend.sum(x * x,axis=1,keepdims=True))))(prod)\no = ReLU(max_value=None, negative_slope=0.0, threshold=0.0)(o)\no = Dense(1, activation='sigmoid', input_shape=(no_terms,))(o)\n\n\nmodel_const = Model(fixed_input,cd)\nmodel = Model([input_layer, fixed_input], o)\n\nop = optimizers.RMSprop(learning_rate=.1, rho=0.9)\nmodel.compile(optimizer=op, loss=mean_squared_error, metrics=['accuracy'])\nplot_model(model, to_file='model.png')\nmodel.summary()\nbatchSize = 100\n\ncheckpoint = ModelCheckpoint(filepath=\"a.hdf5\",monitor='val_acc', mode='max', \n                             save_best_only=True)\nearlystop=EarlyStopping(monitor='val_loss', patience=20)\n\ntrain_docs.shape = (train_docs.shape[0], 1, train_docs.shape[1])\ntrain_label = to_categorical(train_label, num_classes=2, dtype='float32')\nmodel.fit(train_docs, train_label, epochs=10, batch_size=batchSize, \n          validation_data=(test_docs, test_label), \n          callbacks=[earlystop, checkpoint], verbose=1)\n<\/code><\/pre>\n\n<p>And here's the error I got:\n\"ValueError: Error when checking target: expected dense_1 to have 3 dimensions, but got array with shape (1430, 2)\"<\/p>\n\n<p>I have no Idea what is the what the (1430, 2) shape refers to &amp; why I got this error.<\/p>\n","answers":[{"AnswerId":"58112321","CreationDate":"2019-09-26T08:11:33.403","ParentId":null,"OwnerUserId":"9569642","Title":null,"Body":"<p>checking target: expected dense_1 to have 3 dimensions, but got array with shape (1430, 2)\"<\/p>\n\n<p>That means  dense_1  have 3 dimensions but your input  have only 2 dimensions if your design this model in Image processing in that case you have declare the image shape yhat is ((48,48),1) &amp; ((48,48),3) here 1 is for gray scale &amp; 3 is for rgb images <\/p>\n"},{"AnswerId":"58112293","CreationDate":"2019-09-26T08:10:02.940","ParentId":null,"OwnerUserId":"2034650","Title":null,"Body":"<p>Well I found the solution to the error but still can't understand why that previous shape didn't work out &amp; quite frankly this fix was a success among numerous trial &amp; error.<\/p>\n\n<p>I changed the following layers input from the shape format (1, x) to the format of (x,):<\/p>\n\n<pre><code>    input_layer = Input(shape=(no_terms,), name='docs')\n\n    k_constants = backend.constant(np.reshape(tpr_fpr.T, (1,2*no_terms)))\n    fixed_input = Input(tensor=k_constants, shape=(2*no_terms,), name='tpr_fpr')\n    h = Dense(int(300), activation='relu', name='hidden', input_shape=(2*no_terms,), trainable=True)(fixed_input)\n\n    o = ReLU(max_value=None, negative_slope=0.0, threshold=0.0)(o)\n    o = Dense(1, activation='sigmoid', input_shape=(no_terms,))(o)\n<\/code><\/pre>\n\n<p>and also removed the following lines from the code:<\/p>\n\n<pre><code>    train_docs.shape = (train_docs.shape[0], 1, train_docs.shape[1])\n    train_label = to_categorical(train_label, num_classes=2, dtype='float32')\n<\/code><\/pre>\n\n<p>Now I'm just using a label of shape (#no_of_samples, 1) which is binary &amp; not a categorical label.<\/p>\n\n<p>So the new structure is:\n<a href=\"https:\/\/i.stack.imgur.com\/G4yUn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/G4yUn.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I hope somebody can explain what was wrong with the previous model, so I would avoid making the same mistake again.<\/p>\n\n<p>Thanks.<\/p>\n"},{"AnswerId":"58119057","CreationDate":"2019-09-26T14:29:15.120","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>You did indeed directly solve the problem - here's how:<\/p>\n\n<ul>\n<li>Keras binary classification expects labels ('targets') shaped as <code>(batch_size, 1)<\/code>. Reason: the goal of the final layer is to output <em>predictions<\/em>, which will be compared against <em>labels<\/em> to compute metrics (loss, accuracy, etc) - and labels are shaped <code>(batch_size, 1)<\/code><\/li>\n<li>Above is also why <code>to_categorical<\/code> was a problem - see snippet from <a href=\"https:\/\/keras.io\/utils\/#to_categorical\" rel=\"nofollow noreferrer\">docs<\/a> below; for binary classification, one-hot encoding is redundant, as <code>binary_crossentropy<\/code> directly compares labels against predictions as supplied<\/li>\n<li>Keras <code>Dense<\/code> expects inputs to be 2D: <code>(batch_size, input_dim)<\/code>. Your reshaping was making the input 3D: <code>(batch_size, 1, input_dim)<\/code><\/li>\n<li>Above is also why <code>shape=(1, no_terms)<\/code> --> <code>shape=(no_terms,)<\/code> helped; both are, in fact, correct for the data shapes you were feeding at the time. The full batch shape simply includes the batch dim: <code>(batch_size, no_terms)<\/code> (<code>no_terms == input_dim<\/code>)<\/li>\n<li>Lastly, for binary classification, use <code>loss='binary_crossentropy'<\/code> - and never mean square error for classification problems (unless for very specific reasons)<\/li>\n<\/ul>\n\n<pre class=\"lang-py prettyprint-override\"><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"}]}
+{"QuestionId":58106507,"AnswerCount":1,"Tags":"<python><keras><object-detection>","CreationDate":"2019-09-25T21:02:57.453","AcceptedAnswerId":null,"OwnerUserId":6499558.0,"Title":"Keras, Face Detection with only one image","Body":"<p>I am new to object detection. And I can't figure out how to train a model to recognize a face in a picture. There are many mnist tutorials and pretrained model tutorials but not a case like this.<\/p>\n\n<p>I will assign numbers to my questions. <\/p>\n\n<pre><code>folderpath = \"G:\/Datasets\/FACE\/\"\nface = \"face.JPG\"\npeople = \"people.JPG\"\n\nmaskdict = {}\nnfaces = 0\n\nimage= Image.open(folderpath+face)\nimage = image.convert('L')\nimage= np.array(image)\nprint(image.shape) # (239, 162)\n<\/code><\/pre>\n\n<p>I turned the Face into grayscale so it would be easier to work.<\/p>\n\n<pre><code>def create_mask(image, plotimg=False):\n    height, width = image.shape\n    mask = np.empty((height, width))\n    topcroph = int(height*0.9)\n    botcroph = height-topcroph\n    rightcropw = int(width*0.9)\n    leftcropw = width - rightcropw\n\n    mask[botcroph:topcroph, leftcropw:rightcropw] = 1\n    img = image.copy()\n    if plotimg:\n        img[botcroph:topcroph, leftcropw:rightcropw] = 1\n        plt.imshow(img, cmap='gray')\n    return mask.astype(np.uint8)\n<\/code><\/pre>\n\n<p><strong>Creating mask<\/strong><\/p>\n\n<p>Assuming the input always will be a closeup face, create_mask function creates a mask very roughly where the face is, so I can use that mask as 'y' (label)<\/p>\n\n<p><em>1- Is this the correct way of creating a 'y'(label) for detection?<\/em><\/p>\n\n<pre><code>def resizer(img, msk, size):\n    img = np.array(Image.fromarray(img).resize((size,size)))\n    msk = np.array(Image.fromarray(msk).resize((size,size)))\n    return img, msk\n<\/code><\/pre>\n\n<p>Resize both image and it's mask.<\/p>\n\n<pre><code>mask = create_mask(image, plotimg=True)\nimg, msk = resizer(image, mask, 128) # img &amp; mask.shape would be (128,128)\n\nimg = np.expand_dims(img, axis=2) # add one more dimension\nmsk = np.expand_dims(msk, axis=2) # (128, 128, 1)\n<\/code><\/pre>\n\n<p><strong>Augmentation<\/strong><\/p>\n\n<p>In the below code, I wanted to multiply the image, so I could have a dataset. <\/p>\n\n<p><em>2- Technically, you can't train with one image, right?<\/em><\/p>\n\n<pre><code>datasize=120\ndata = np.empty((datasize, 128, 128, 1))\nmaskage = np.empty((datasize, 128, 128, 1))\nfor i in range(0, 40):\n    np.random.seed(i)\n    img_r = keras.preprocessing.image.random_rotation(img, 30)\n    msk_r = keras.preprocessing.image.random_rotation(msk, 30)\n    maskage[i,:,:] = msk_r\n    data[i,:,:] = img_r\nfor i in range(40,80):\n    np.random.seed(i)\n    img_b = keras.preprocessing.image.random_brightness(img, (.5, 1.5))\n    maskage[i,:,:] = msk\n    data[i,:,:] = img_b\nfor i in range(80,120):\n    np.random.seed(i)\n    img_z = keras.preprocessing.image.random_zoom(img, (.7, 1))\n    msk_r = keras.preprocessing.image.random_zoom(msk, (.7, 1))\n    data[i,:,:] = img_z\n    maskage[i,:,:] = msk_r\n<\/code><\/pre>\n\n<p><strong>Model<\/strong><\/p>\n\n<p>Keras model throws ValueError. I can't figure out how to configure layers, how does this work?<\/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(16, (3,3), padding=\"same\", input_shape=(128,128,1)))\nmodel.add(layers.Activation('relu'))\n\nmodel.add(layers.Conv2D(16, (3,3), padding=\"same\"))\nmodel.add(layers.Activation(\"sigmoid\"))\n\nmodel.compile(optimizer='adam', loss='binary_crossentropy', metrics=['acc'])\nmodel.summary()\n\n    _________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nconv2d_1 (Conv2D)            (None, 128, 128, 16)      160       \n_________________________________________________________________\nactivation_1 (Activation)    (None, 128, 128, 16)      0         \n_________________________________________________________________\nconv2d_2 (Conv2D)            (None, 128, 128, 16)      2320      \n_________________________________________________________________\nactivation_2 (Activation)    (None, 128, 128, 16)      0         \n=================================================================\nTotal params: 2,480\nTrainable params: 2,480\nNon-trainable params: 0\n\n# Input shapes are:\ndata.shape  # (120, 128, 128, 1) \nmaskage.shape  # (120, 128, 128, 1)\n\n#Run the model\nmodel.fit(data, maskage)\n<\/code><\/pre>\n\n<blockquote>\n  <p>ValueError: Error when checking target: expected activation_6 to have\n  shape (128, 128, 16) but got array with shape (128, 128, 1)<\/p>\n<\/blockquote>\n\n<p><em>3. What should I do at this point?<\/em><\/p>\n\n<p><em>4. Even if the model manages to train, will model.predict on \"people.JPG\" work?\n How can I draw a rectangle on face prediction output? Will this method work?<\/em><\/p>\n\n<p>Edit: Model started to train after I changed the second conv2d layer variable from 16 to 1. But It can't find the face.<\/p>\n","answers":[{"AnswerId":"58107427","CreationDate":"2019-09-25T22:41:56.500","ParentId":null,"OwnerUserId":"12118774","Title":null,"Body":"<p>Before starting your questions let me some important thing to be more clear. As far as I understand your goal is to achieve both <code>classification<\/code> and <code>localization<\/code>. These both needs different models to be trained on and it is more advanced to start from scratch. Assuming you only need classification let me try to answer your questions.<\/p>\n\n<p>1) y labels should be the binary for your case <code>face<\/code> or <code>background<\/code> and these should be 1D column array.<\/p>\n\n<p>2) Yes you are right, flipping or rotating image is a good strategy to enrich your dataset.<\/p>\n\n<p>3) This error happens because of the second argument of <code>model.fit(data, maskage)<\/code>, i.e <strong>maskage<\/strong>. Since your label needs to be a column, not a 128x128x1 array, this error occurs.<\/p>\n\n<p>4) This method won't work as I explained before the questions. You are looking for both <code>localization<\/code> and <code>classification<\/code><\/p>\n\n<p>As far as I understand, you have checked the MNIST tutorials to modify to your problem, but first and most important thing is you should change the y labels. Further, since I have not sufficient knowledge to do how to properly localize objects, I won't be able to give much information about how you succeed the localization part of your project.<\/p>\n\n<p>EDIT : For the ones who are not familiar with the <code>localization<\/code>, there is a helpful guide with Keras in the <a href=\"https:\/\/towardsdatascience.com\/getting-started-with-bounding-box-regression-in-tensorflow-743e22d0ccb3\" rel=\"nofollow noreferrer\">link<\/a><\/p>\n"}]}
+{"QuestionId":58106650,"AnswerCount":1,"Tags":"<python><tensorflow><gpu><conda>","CreationDate":"2019-09-25T21:18:53.937","AcceptedAnswerId":null,"OwnerUserId":4706745.0,"Title":"How to judge from the output of a log whether Tensorflow is working with GPU or not?","Body":"<p>If I try to run a command that tells you whether Tensorflow using GPU or not. <\/p>\n\n<pre><code>&gt;&gt;&gt; sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n<\/code><\/pre>\n\n<p>Command returns this, now I am not able to understand whether tensor flow using GPU or not. <\/p>\n\n<pre><code>2019-09-25 17:08:47.509729: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1640] Found device 0 with properties: \nname: Quadro P4000 major: 6 minor: 1 memoryClockRate(GHz): 1.48\npciBusID: 0000:03:00.0\n2019-09-25 17:08:47.509929: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:53] Could not dlopen library 'libcudart.so.10.0'; dlerror: libcudart.so.10.0: cannot open shared object file: No such file or directory\n2019-09-25 17:08:47.510040: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:53] Could not dlopen library 'libcublas.so.10.0'; dlerror: libcublas.so.10.0: cannot open shared object file: No such file or directory\n2019-09-25 17:08:47.510139: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:53] Could not dlopen library 'libcufft.so.10.0'; dlerror: libcufft.so.10.0: cannot open shared object file: No such file or directory\n2019-09-25 17:08:47.510234: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:53] Could not dlopen library 'libcurand.so.10.0'; dlerror: libcurand.so.10.0: cannot open shared object file: No such file or directory\n2019-09-25 17:08:47.510328: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:53] Could not dlopen library 'libcusolver.so.10.0'; dlerror: libcusolver.so.10.0: cannot open shared object file: No such file or directory\n2019-09-25 17:08:47.510440: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:53] Could not dlopen library 'libcusparse.so.10.0'; dlerror: libcusparse.so.10.0: cannot open shared object file: No such file or directory\n2019-09-25 17:08:47.510483: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudnn.so.7\n2019-09-25 17:08:47.510498: W tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1663] Cannot dlopen some GPU libraries. Skipping registering GPU devices...\n2019-09-25 17:08:47.510524: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1181] Device interconnect StreamExecutor with strength 1 edge matrix:\n2019-09-25 17:08:47.510536: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1187]      0 \n2019-09-25 17:08:47.510556: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1200] 0:   N \nDevice mapping:\n\/job:localhost\/replica:0\/task:0\/device:XLA_CPU:0 -&gt; device: XLA_CPU device\n\/job:localhost\/replica:0\/task:0\/device:XLA_GPU:0 -&gt; device: XLA_GPU device\n2019-09-25 17:08:47.510713: I tensorflow\/core\/common_runtime\/direct_session.cc:296] Device mapping:\n\/job:localhost\/replica:0\/task:0\/device:XLA_CPU:0 -&gt; device: XLA_CPU device\n\/job:localhost\/replica:0\/task:0\/device:XLA_GPU:0 -&gt; device: XLA_GPU device\n<\/code><\/pre>\n\n<p>nvidia-smi command return this..<\/p>\n\n<pre><code>+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 387.26                 Driver Version: 387.26                    |\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  Quadro P4000        Off  | 00000000:03:00.0  On |                  N\/A |\n| 46%   40C    P8    11W \/ 105W |   1240MiB \/  8111MiB |      0%      Default |\n+-------------------------------+----------------------+----------------------+\n\n+-----------------------------------------------------------------------------+\n| Processes:                                                       GPU Memory |\n|  GPU       PID   Type   Process name                             Usage      |\n|=============================================================================|\n|    0      1943      G   \/usr\/libexec\/Xorg                            317MiB |\n|    0      2047      G   \/usr\/bin\/gnome-shell                           5MiB |\n|    0      5505      G   \/usr\/libexec\/Xorg                            109MiB |\n|    0     10091      G   \/usr\/libexec\/Xorg                            165MiB |\n|    0     10952      G   ...uest-channel-token=14061294616847102337    87MiB |\n+-----------------------------------------------------------------------------+\n<\/code><\/pre>\n","answers":[{"AnswerId":"58106907","CreationDate":"2019-09-25T21:42:14.470","ParentId":null,"OwnerUserId":"11667119","Title":null,"Body":"<h2>tensorflow has multiple stages before running a learning kernel(function) on GPU<\/h2>\n\n<ol>\n<li>check available devices physically<\/li>\n<li>check library files relavent to them (CuDNN files and etc)<\/li>\n<li>allocate needed memory<\/li>\n<li>start the process<\/li>\n<\/ol>\n\n<p>yours stuck in stage 2 <\/p>\n\n<h2>running tensorflow with GPU needs couple of things<\/h2>\n\n<ol>\n<li>Nvidia driver<\/li>\n<li>Cuda (compiler)<\/li>\n<li>CuDNN library files<\/li>\n<\/ol>\n\n<h2>how to test them<\/h2>\n\n<ol>\n<li>run <code>Nvidia-smi<\/code> to test driver availability <\/li>\n<li>run <code>nvcc --version<\/code> to check cuda compiler availability<\/li>\n<li>run <code>import tensorflow as tf<\/code><\/li>\n<li>run a session in tensorflow<\/li>\n<\/ol>\n\n<p>if your getting error in stage 3 its because that tensorflow didn't find the CuDNN files <\/p>\n\n<p>if your getting error in stage 4 its because you have version incompatability, you can check <a href=\"https:\/\/stackoverflow.com\/questions\/50622525\/which-tensorflow-and-cuda-version-combinations-are-compatible\">this solution<\/a>  inside stackoverflow, or search the problem<\/p>\n\n<h2>how to know when tensorflow is not using GPU<\/h2>\n\n<p>there are several ways to know:<\/p>\n\n<ul>\n<li><p>the log like the one inside yours: <code>Cannot dlopen some GPU libraries. Skipping registering GPU devices...<\/code><\/p><\/li>\n<li><p>it can be detected from <code>nvidia-smi<\/code> command:\nif your using python, it will show a python process inside processes with very large memory consumption (it will consume almost all of your available GPU memory)<\/p><\/li>\n<\/ul>\n"}]}
+{"QuestionId":58106929,"AnswerCount":0,"Tags":"<python><function><csv><tensorflow>","CreationDate":"2019-09-25T21:43:58.157","AcceptedAnswerId":null,"OwnerUserId":12121629.0,"Title":"How does the tensorflow.python.data.ops.dataset_ops.DatasetV1Adapter work?","Body":"<p>I am trying to wrap my head around ML and AI using TensorFlow. There is an example problem on the website which discusses the processing of .CSV data. The .CVS data is said to have been taken from the titanic and essentially contains categorical and numerical features that will be used to label a passenger as dead or alive.<\/p>\n\n<p>First of all, if anyone know or has any resources or references that discusses that example in more detail than is done on the <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/load_data\/csv\" rel=\"nofollow noreferrer\">TensorFlow website<\/a>, please could you kindly refer me to that?<\/p>\n\n<p>Secondly, to my more important question. There is a method which is used in the example to pack the individual numeric data into a single numeric key while maintaining the categorical column as is.<\/p>\n\n<p>The function used to do this is as bellow:<\/p>\n\n<pre><code>class PackNumericFeatures(object):\n    def __init__(self, names):\n        self.names = names\n\n    def __call__(self, features, labels):\n        numeric_freatures = [features.pop(name) for name in self.names]\n        numeric_features = [tf.cast(feat, tf.float32) for feat in numeric_freatures]\n        numeric_features = tf.stack(numeric_features, axis=-1)\n        features['numeric'] = numeric_features\n\n        return features, labels\n<\/code><\/pre>\n\n<p>The above function is called like this:<\/p>\n\n<pre><code>\nNUMERIC_FEATURES = ['age','n_siblings_spouses','parch', 'fare']\n\npacked_train_data = raw_train_data.map(PackNumericFeatures(NUMERIC_FEATURES))\n\n<\/code><\/pre>\n\n<p>The output <code>packed_train_data<\/code> looks something like this:<\/p>\n\n<pre><code>sex                 : [b'male' b'female' b'female' b'female' b'male']\nclass               : [b'Third' b'First' b'Third' b'First' b'Third']\ndeck                : [b'unknown' b'C' b'unknown' b'C' b'unknown']\nembark_town         : [b'Southampton' b'Cherbourg' b'Southampton' b'Southampton' b'Queenstown']\nalone               : [b'n' b'n' b'y' b'n' b'y']\nnumeric             : [[22.      1.      0.      7.25  ]\n [38.      1.      0.     71.2833]\n [26.      0.      0.      7.925 ]\n [35.      1.      0.     53.1   ]\n [28.      0.      0.      8.4583]]\n\n<\/code><\/pre>\n\n<p>The above output is produced by passing a single batch of data (i.e. <code>packed_train_data<\/code>) to the function:<\/p>\n\n<pre><code>\ndef show_batch(dataset):\n    for batch, head in dataset.take(1):\n        for labels, value in batch.items():\n            print(\"{:20s}: {}\".format(labels, value.numpy()))\n<\/code><\/pre>\n\n<p>like this:<\/p>\n\n<pre><code>show_batch(packed_train_data)\n\n<\/code><\/pre>\n\n<p>What I do not understand is how the <code>map<\/code> function is working to generate that output. More generally I don't get how the call containing the <code>map<\/code> function is interacting with the <code>PackNumericFeatures(object)<\/code> class. <\/p>\n\n<p>I know this problem is specific but any help would be appreciated. Cheers.<\/p>\n","answers":[]}
+{"QuestionId":58107053,"AnswerCount":0,"Tags":"<pytorch><labelimg>","CreationDate":"2019-09-25T21:56:39.693","AcceptedAnswerId":null,"OwnerUserId":10493650.0,"Title":"How to use annotated images (made using labelImg) on a pretrained pytorch model?","Body":"<p>So I have a set of images which I am annotating using labelImg. Its in the PascalVoc format. I have an image folder and annotation folder. How do I load such dataset of labelled images onto a pytorch model like inception v3. Help with Tensorflow or any other framework will also work. I just wanna know how to use such a data. I can't understand any of the google results.<\/p>\n\n<p>I am kind of new to this. So a detailed help would really be appreciated.<\/p>\n","answers":[]}
+{"QuestionId":58107693,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-25T23:19:03.557","AcceptedAnswerId":58108093.0,"OwnerUserId":9798029.0,"Title":"TFrecord pre-process data before feeding to a model","Body":"<p>I have a TFrecord file that consist of such features<\/p>\n\n<pre><code>features = {\n  'features': tf.FixedLenFeature([17920], tf.float32),\n  'video_id': tf.FixedLenFeature([], tf.string),\n  'split_id': tf.FixedLenFeature([], tf.int64),\n  'audio': tf.FixedLenFeature([44100], tf.float32)\n}\n<\/code><\/pre>\n\n<p>The way I currently get the iterator is as such:<\/p>\n\n<pre><code> def getiterator(batch_size, decode_parallel_calls=1,repeat=False,\n    shuffle=False,\n    shuffle_buffer_size=None,\n    prefetch_size=None,\n    prefetch_gpu_num=None):\ndataset = tf.data.TFRecordDataset(['..\/main.tfrecords'])\n\n  # Shuffle all filepaths every epoch\n  if shuffle:\n    dataset = dataset.shuffle(buffer_size=sum(1 for _ in tf.python_io.tf_record_iterator('..\/main.tfrecords')))\n\n  # Repeat\n  if repeat:\n    dataset = dataset.repeat()\n\n  # Shuffle examples\n  if shuffle:\n    dataset = dataset.shuffle(buffer_size=shuffle_buffer_size)\n\n  def extract_fn(data_record):\n    features = {\n      'features': tf.FixedLenFeature([17920], tf.float32),\n      'video_id': tf.FixedLenFeature([], tf.string),\n      'split_id': tf.FixedLenFeature([], tf.int64),\n      'audio': tf.FixedLenFeature([44100], tf.float32)\n    }\n    return tf.parse_single_example(data_record, features)\n\n  dataset = dataset.map(extract_fn, num_parallel_calls=decode_parallel_calls)\n\n  # Make batches\n  dataset = dataset.batch(batch_size, drop_remainder=True)\n\n  # Get tensors\n  iterator = dataset.make_one_shot_iterator()\n\n  return iterator.get_next()\n<\/code><\/pre>\n\n<p>One thing that I am interested in doing is preprocessing the audio data by downsampling to 16000Hz. However, I am not sure how to do it so as I currently only get Tensors. What would be a good approach?<\/p>\n","answers":[{"AnswerId":"58108093","CreationDate":"2019-09-26T00:32:52.620","ParentId":null,"OwnerUserId":"1667256","Title":null,"Body":"<p>You could simply place your downsampling logic inside <code>extract_fn<\/code>. <\/p>\n\n<pre><code>def extract_fn(data_record):\n    features = {\n      'features': tf.FixedLenFeature([17920], tf.float32),\n      'video_id': tf.FixedLenFeature([], tf.string),\n      'split_id': tf.FixedLenFeature([], tf.int64),\n      'audio': tf.FixedLenFeature([44100], tf.float32)\n    }\n    features_in = tf.parse_single_example(data_record, features)\n\n    # This is your audio tensor before downsampling.\n    audio_tensor = features_in['audio'] \n    # Doing something to downsample audio_tensor\n    downsampled_audio_tensor = do_something(audio_tensor)\n\n    # Now I assume you need every other tensors as is, but replace the audio with\n    # the one we downsampled above.\n    features_out = {\n      'features': features_in['features'],\n      'video_id': features_in['video_id'],\n      'split_id': features_in['split_id'],\n      'audio': downsampled_audio_tensor,\n    }\n    return features_out\n\n<\/code><\/pre>\n"}]}
+{"QuestionId":58107942,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-26T00:07:20.683","AcceptedAnswerId":58108030.0,"OwnerUserId":3195597.0,"Title":"How to add a constant tensor to Input tensor in Keras\/tensorflow","Body":"<p>I have a simple CNN with inputs of shape (5,5,3). As a first step I want to add a constant tensor to the input. \nWith the code below, I get\nAttributeError: 'NoneType' object has no attribute '_inbound_nodes' <\/p>\n\n<p>I have tried a few things like <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>const_change = Input(tensor=tf.constant([ ...\n<\/code><\/pre>\n\n<p>or <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>const_change = Input(tensor=K.variable([ ...\n<\/code><\/pre>\n\n<p>but nothing seems to work. Any help is highly appreciated.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from __future__ import print_function\n\nimport tensorflow as tf\nimport numpy as np\nimport keras\nfrom keras import backend as K\nfrom keras.models import Model\nfrom keras.layers import Input\nfrom keras.layers import Conv2D, MaxPooling2D\nfrom keras.layers import Activation, Dropout, Flatten, Dense\n\n# Python 2.7.10\n# keras version 2.2.0\n# tf.VERSION '1.8.0'\n\nraw_input = Input(shape=(5, 5, 3))\n\nconst_change = tf.constant([\n        [[5.0,0.0,0.0],[0.0,0.0,-3.0],[-10.0,0.0,0.0],[0.0,0.0,4.0],[-20.0,0.0,0.0]],\n        [[-15.0,0.0,12.0],[0.0,4.0,0.0],[-3.0,0.0,10.0],[-18.0,0.0,0.0],[20.0,0.0,-6.0]],\n        [[0.0,0.0,6.0],[0.0,-2.0,-6.0],[0.0,0.0,2.0],[0.0,0.0,-9.0],[7.0,-6.0,0.0]],\n        [[-3.0,4.0,0.0],[11.0,-12.0,0.0],[0.0,0.0,0.0],[0.0,0.0,7.0],[0.0,0.0,2.0]],\n        [[0.0,0.0,0.0],[0.0,1.0,-2.0],[4.0,0.0,3.0],[0.0,0.0,0.0],[0.0,0.0,0.0]]])\n\ncnn_layer1 = Conv2D(32, (4, 4), activation='relu')\ncnn_layer2 = MaxPooling2D(pool_size=(2, 2))\n\ncnn_layer3 = Dense(128, activation='relu')\ncnn_layer4 = Dropout(0.1)\ncnn_output = Dense(4, activation='softmax')\n\nproc_input = keras.layers.Add()([raw_input, const_change])\n# proc_input = keras.layers.add([raw_input, const_change])  -&gt; leads to the same error (see below)\nlay1 = cnn_layer1(proc_input)\nlay2 = cnn_layer2(lay1)\nlay3 = Flatten()(lay2)\nlay4 = cnn_layer3(lay3)\nlay5 = cnn_layer4(lay4)\nlay_out = cnn_output(lay5)\n\nmodel = Model(inputs=raw_input, outputs=lay_out)\n# -&gt; AttributeError: 'NoneType' object has no attribute '_inbound_nodes'\n<\/code><\/pre>\n","answers":[{"AnswerId":"58108030","CreationDate":"2019-09-26T00:21:37.860","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>The <code>const_change<\/code> should be also <code>Input<\/code> just like <code>raw_input<\/code>. You can create another input layer named <code>const_input<\/code>, and feed <code>raw_input<\/code> and <code>const_input<\/code> together into model. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>...\nconst_input = Input(tensor=const_change)\n...\nproc_input = keras.layers.Add()[raw_input, const_input]\n...\nmodel = Model(inputs=[raw_input, const_input], outputs=lay_out)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58108109,"AnswerCount":0,"Tags":"<numpy><tensorflow><unique><indices>","CreationDate":"2019-09-26T00:35:52.957","AcceptedAnswerId":null,"OwnerUserId":2402023.0,"Title":"How to get a map for a lookup table in tensorflow","Body":"<p>I have two different datasets with categorical data and two corresponding keys for each dataset. <\/p>\n\n<p>So imagine dataset1 with values <code>[0,1,1,0,2]<\/code>\nwith keys that represent the categoryKey:<code>['cat','dog','mouse']<\/code><\/p>\n\n<p>and dataset2 with values <code>[0,1,2,2,1,2]<\/code>\nwith categoryKey:<code>['dog','mouse','cat']<\/code><\/p>\n\n<p>I would want to transform dataset2 to match dataset1, by finding the proper key mapping. In this case, the map for dataset 2 to match the keys to dataset 1 is <code>[1,2,0]<\/code>. Applied as a lookup for dataset2, it would transform it thus: <code>[0,1,2,2,1,2]-&gt;[1,2,0,0,2,0]<\/code><\/p>\n\n<p>and then dataset1 and 2 would have the correct categories w.r.t. each other.<\/p>\n\n<p>This is pretty straightforward with a list comprehension in normal python, but instead, I am working with tensorflow, so imagine dataset1 and dataset2 are tensors of shape <code>[100,100]<\/code>. What tensorflow operation can I do to produce a map <code>[1,2,0]<\/code>?<\/p>\n\n<p>In truth, datasets 1 and 2 are batches of data, as well as <code>categorieKeys<\/code> are also batches with shapes of >(3,3). So we cannot convert from tensors to numpy then to lists and do list comprehension.<\/p>\n\n<p>What is the correct tensorflow operation to find the map, and then to transform the data?<\/p>\n","answers":[]}
+{"QuestionId":58108351,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><keras>","CreationDate":"2019-09-26T01:21:14.690","AcceptedAnswerId":null,"OwnerUserId":11939235.0,"Title":"Tensorflow ValueError: could not convert string to float: 'MXHbacc3'","Body":"<p>I tried to build artificial neural network,\nbut it's not working.\nHow do I do it? <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential()\nmodel.add(Dense(output_dim=12, input_dim=4, init='normal', activation='sigmoid'))\n\n__main__:1: UserWarning: Update your `Dense` call to the Keras 2 API: `Dense(input_dim=4, activation=\"sigmoid\", kernel_initializer=\"normal\", units=12)`\n\nmodel.add(Dense(output_dim=1, init='normal', activation='linear'))\nmodel.compile(loss='mean_squared_error', optimizer='adam')\nmodel.fit(X_train, Y_train, nb_epoch=1000, batch_size=100, verbose=0)\n<\/code><\/pre>\n\n<pre><code>2019-09-25 22:11:17.983801: I tensorflow\/core\/platform\/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.2\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/Users\/kimsohyeon\/anaconda3\/envs\/tensorflow\/lib\/python3.5\/site-packages\/keras\/engine\/training.py\", line 1039, in fit\n    validation_steps=validation_steps)\n  File \"\/Users\/kimsohyeon\/anaconda3\/envs\/tensorflow\/lib\/python3.5\/site-packages\/keras\/engine\/training_arrays.py\", line 199, in fit_loop\n    outs = f(ins_batch)\n  File \"\/Users\/kimsohyeon\/anaconda3\/envs\/tensorflow\/lib\/python3.5\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 2721, in __call__\n    return self._legacy_call(inputs)\n  File \"\/Users\/kimsohyeon\/anaconda3\/envs\/tensorflow\/lib\/python3.5\/site-packages\/keras\/backend\/tensorflow_backend.py\", line 2693, in _legacy_call\n    **self.session_kwargs)\n  File \"\/Users\/kimsohyeon\/anaconda3\/envs\/tensorflow\/lib\/python3.5\/site-packages\/tensorflow\/python\/client\/session.py\", line 895, in run\n    run_metadata_ptr)\n  File \"\/Users\/kimsohyeon\/anaconda3\/envs\/tensorflow\/lib\/python3.5\/site-packages\/tensorflow\/python\/client\/session.py\", line 1097, in _run\n    np_val = np.asarray(subfeed_val, dtype=subfeed_dtype)\n  File \"\/Users\/kimsohyeon\/anaconda3\/envs\/tensorflow\/lib\/python3.5\/site-packages\/numpy\/core\/numeric.py\", line 492, in asarray\n    return array(a, dtype, copy=False, order=order)\n\nValueError: could not convert string to float: 'MXHbacc3'\n<\/code><\/pre>\n\n<p>Then, Is it a good idea to delete the conda virtual environment and reinstall new one?<\/p>\n\n<pre><code>&gt;&gt;&gt; tf.__version__\n'1.5.0\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58108479,"AnswerCount":0,"Tags":"<opencv><tensorflow><google-colaboratory>","CreationDate":"2019-09-26T01:43:00.280","AcceptedAnswerId":null,"OwnerUserId":12122305.0,"Title":"Using socketed video in Google Colab","Body":"<p>I'm currently working on an object detection project. I need to use the socketed video from raspberry pi in my model for the live feed. Can anyone help me how to use that socketed video in Google Colab?<\/p>\n","answers":[]}
+{"QuestionId":58108543,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning>","CreationDate":"2019-09-26T01:54:34.527","AcceptedAnswerId":null,"OwnerUserId":8158060.0,"Title":"Training accuracy while fitting the model not reflected in confusion matrix","Body":"<p>I am training an image classification model to classify certain images containing mountain, waterfall and people classes. Currently, I am using Vgg-net (transfer learning) to train this data. While training, I am getting almost 93% accuracy on training data and 90% accuracy on validation data. However, when I want to check the classes being wrongly classified during training by using confusion matrix on the training data, the accuracy seems to be much less. Only about 30% of the data is classified correctly.<\/p>\n\n<p>I have tried checking the confusion matrix on other similar image classification problems but there it seems to be showing the accuracy that we see while training. <\/p>\n\n<p>Code to create ImageDataGenerator objects for training and validation<\/p>\n\n<pre><code>from keras.applications.inception_v3 import preprocess_input, decode_predictions\n#Train DataSet Generator with Augmentation\nprint(\"\\nTraining Data Set\")\ntrain_generator = ImageDataGenerator(preprocessing_function=preprocess_input)\ntrain_flow = train_generator.flow(\n    train_images, train_labels,\n    batch_size = BATCH_SIZE\n)\n\n#Validation DataSet Generator with Augmentation\nprint(\"\\nValidation Data Set\")\nval_generator = ImageDataGenerator(preprocessing_function=preprocess_input)\nval_flow = val_generator.flow(\n    validation_images,validation_labels,\n    batch_size = BATCH_SIZE\n)\n<\/code><\/pre>\n\n<p>Code to build the model and compile it<\/p>\n\n<pre><code># Initialize InceptionV3 with transfer learning\nbase_model = applications.vgg16.VGG16(weights='imagenet', \n                                include_top=False, \n                                input_shape=(WIDTH, HEIGHT,3))\n\n# add a global spatial average pooling layer\nx = base_model.output\n\nx = GlobalAveragePooling2D()(x)\n# and a dense layer\nx = Dense(1024, activation='relu')(x)\npredictions = Dense(len(categories), activation='softmax')(x)\n\n# first: train only the top layers (which were randomly initialized)\n# i.e. freeze all convolutional InceptionV3 layers\nfor layer in base_model.layers:\n    layer.trainable = False\n\n\n# this is the model we will train\nmodel = Model(inputs=base_model.input, outputs=predictions)\n\n# compile the model (should be done *after* setting layers to non-trainable)\nmodel.compile(optimizer=optimizers.RMSprop(lr=1e-4), metrics=['accuracy'], loss='categorical_crossentropy')\nmodel.summary()\n<\/code><\/pre>\n\n<p>Code to fit the model to the training dataset and validate using validation dataset<\/p>\n\n<pre><code>import math\ntop_layers_file_path=r\"\/content\/drive\/My Drive\/intel-image-classification\/intel-image-classification\/top_layers.iv3.hdf5\"\n\ncheckpoint = ModelCheckpoint(top_layers_file_path, monitor='loss', verbose=1, save_best_only=True, mode='min')\ntb = TensorBoard(log_dir=r'\/content\/drive\/My Drive\/intel-image-classification\/intel-image-classification\/logs', batch_size=BATCH_SIZE, write_graph=True, update_freq='batch')\nearly = EarlyStopping(monitor=\"loss\", mode=\"min\", patience=5)\ncsv_logger = CSVLogger(r'\/content\/drive\/My Drive\/intel-image-classification\/intel-image-classification\/logs\/iv3-log.csv', append=True)\n\nhistory = model.fit_generator(train_flow, \n                              epochs=30, \n                              verbose=1,\n                              validation_data=val_flow, \n                              validation_steps=math.ceil(validation_images.shape[0]\/BATCH_SIZE),                              \n                              steps_per_epoch=math.ceil(train_images.shape[0]\/BATCH_SIZE),\n                              callbacks=[checkpoint, early, tb, csv_logger])\n<\/code><\/pre>\n\n<hr>\n\n<p>training steps show the following accuracy:<\/p>\n\n<pre><code>Epoch 1\/30\n\n91\/91 [==============================] - 44s 488ms\/step - loss: 0.6757 - acc: 0.7709 - val_loss: 0.4982 - val_acc: 0.8513\n\nEpoch 2\/30\n\n91\/91 [==============================] - 32s 349ms\/step - loss: 0.4454 - acc: 0.8395 - val_loss: 0.3980 - val_acc: 0.8557\n\n.\n.\n\nEpoch 20\/30\n\n91\/91 [==============================] - 32s 349ms\/step - loss: 0.2026 - acc: 0.9238 - val_loss: 0.2684 - val_acc: 0.8940\n\n.\n.\n\nEpoch 30\/30\n\n91\/91 [==============================] - 32s 349ms\/step - loss: 0.1739 - acc: 0.9364 - val_loss: 0.2616 - val_acc: 0.8984\n<\/code><\/pre>\n\n<p>Ran predictions on the training dataset itself:<\/p>\n\n<pre><code>import math\nimport numpy as np\npredictions = model.predict_generator(\n    train_flow,\n    verbose=1,\n    steps=math.ceil(train_images.shape[0]\/BATCH_SIZE))\npredicted_classes = [x[0] for x in enc.inverse_transform(predictions)]\ntrue_classes = [x[0] for x in enc.inverse_transform(train_labels)]\n<\/code><\/pre>\n\n<p>enc - OneHotEncoder<\/p>\n\n<p>However, the confusion matrix looks like the following:<\/p>\n\n<pre><code>import sklearn\n\nsklearn.metrics.confusion_matrix(\n    true_classes, \n    predicted_classes,\n    labels=['mountain','people','waterfall'])\n<\/code><\/pre>\n\n<p>Confusion Matrix (Could not upload a better looking picture)<\/p>\n\n<pre><code>  ([[315, 314, 283],\n\n   [334, 309, 273],\n\n   [337, 280, 263]])\n<\/code><\/pre>\n\n<p>The complete code has been uploaded on <a href=\"https:\/\/nbviewer.jupyter.org\/github\/paridhichoudhary\/scene-image-classification\/blob\/master\/Classification_v7.ipynb\" rel=\"nofollow noreferrer\">https:\/\/nbviewer.jupyter.org\/github\/paridhichoudhary\/scene-image-classification\/blob\/master\/Classification_v7.ipynb<\/a><\/p>\n","answers":[{"AnswerId":"58122434","CreationDate":"2019-09-26T18:06:14.320","ParentId":null,"OwnerUserId":"9477338","Title":null,"Body":"<p>I believe it's because of the <code>train_generator.flow<\/code> has <code>shuffle=True<\/code> by <a href=\"https:\/\/github.com\/keras-team\/keras-preprocessing\/blob\/master\/keras_preprocessing\/image\/image_data_generator.py#L371\" rel=\"nofollow noreferrer\">default<\/a>. This result in <code>predicted_classes<\/code> doesn't match the <code>train_labels<\/code>. <\/p>\n\n<p>Maybe set <code>shuffle=False<\/code> on <code>train_generator.flow<\/code> should help this, or instead use something like this might easier to understand. <\/p>\n\n<pre><code>predicted_classes = []\ntrue_classes = []\nfor i in range(len(train_flow)):  # you can just use `len(train_flow)` instead of `math.ceil(train_images.shape[0]\/BATCH_SIZE))`\n    x, y = train_flow[i]  # you can use `train_flow[i]` like this\n    z = model.predict(x)\n    for j in range(x.shape[0]):\n        predicted_classes.append(z[j])\n        true_classes.append(y[j])\n<\/code><\/pre>\n\n<p>I didn't try this yet but it should work though.<\/p>\n"}]}
+{"QuestionId":58108817,"AnswerCount":0,"Tags":"<python><django><tensorflow><keras>","CreationDate":"2019-09-26T02:33:52.933","AcceptedAnswerId":null,"OwnerUserId":12122369.0,"Title":"Python crashes while closing and open new Keras(tensorflow) session","Body":"<p>I'm making a model to service with Keras &amp; Django.\nThe purpose is initializing models with various hyperparams or load models, which are same architecture but, weights(parameters) and hyperparameters are different. <strong>Because weights are valid in each session, 'model_init' or 'load_model' method close and open tf.Session() each time it is called.<\/strong><\/p>\n\n<p><strong>However, After initializing new model or load other models, Python crashes and Django service shuts down. It seems to be related with session &amp; memory.<\/strong><\/p>\n\n<p>Below is my code:<\/p>\n\n<pre><code>class keras_model:\n    session = None\n    graph = None\n    model = None\n\n    def model_init(~):\n        # hyper parameter setting\n        tf.reset_default_graph()\n        self.graph = None\n        if self.session is not None:\n            self.session.close()\n            # del self.session\n            self.session = None\n\n        self.graph = tf.Graph()\n            with self.graph.as_default():\n                self.session = tf.Session(graph=self.graph)\n                    with self.session.as_default():\n                        # model definition\n                        self.model = models.Model(~)\n                        self.model.compile(~)\n\n    def train():\n        with tf.device(self.device):\n            with self.graph.as_default():\n                with self.session.as_default():\n                    training_log = self.model.fit(~)\n        return training_log\n\n    def predict():\n        ~\n\n    def load_model():\n        tf.reset_default_graph()\n        self.graph = None\n        if self.session is not None:\n            self.session.close()\n            self.session = None\n\n        self.graph = tf.Graph()\n        with self.graph.as_default():\n            self.session = tf.Session(graph=self.graph)\n            with self.session.as_default():\n                # load model architecture\n                self.model = model_from_json(~)\n                # load model weight\n                self.model.load_weights(~)\n\n    def save_model():\n        ~\n<\/code><\/pre>\n\n<p>Any idea or suggestion would be appreciated.<\/p>\n","answers":[]}
+{"QuestionId":58108943,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-26T02:53:38.023","AcceptedAnswerId":null,"OwnerUserId":1516854.0,"Title":"Finding differences in images","Body":"<p>I got images of an item labeled as \"good\". And I got images labeled as \"bad\".<\/p>\n\n<p>A good image means all perfect. A bad image means it has a defect. The defect could be a scratch (any size, any shape), a color distortion, a shape distortion (contour), ...<\/p>\n\n<p>I remember Lawrence saying in ancient programming we used data and rules to get answers, in AI we use data and answers to get rules. I wonder if it is possible to do that here too.<\/p>\n\n<p>I think to get a feature vector of the good (object recognition) and a feature vector of the bad images and can then overlay a frame(s) on the bad picture to mark the difference(s). Further more to extract that difference into a mask and classify this now as \"scratch\".<\/p>\n\n<p>After that I could find on new images with R-CNN and the masks of defects which item is good or bad.<\/p>\n\n<p>This is a beginning of the project. I ask if this way is the right approach or is there a better way?<\/p>\n","answers":[]}
+{"QuestionId":58108953,"AnswerCount":0,"Tags":"<docker><tensorflow><tensorflow-serving><docker-for-windows>","CreationDate":"2019-09-26T02:55:00.863","AcceptedAnswerId":null,"OwnerUserId":11053801.0,"Title":"How to fix invalid mount config for type \"bind\": source path must be a directory in Windows docker Container?","Body":"<p>I have a keras model (Windows Spyder IDE) and saved for tensorflow serving, and followed the steps in <a href=\"https:\/\/www.youtube.com\/watch?v=CUpUPY5g8NE\" rel=\"nofollow noreferrer\">https:\/\/www.youtube.com\/watch?v=CUpUPY5g8NE<\/a> for deploy.<\/p>\n\n<p>I am using docker-for-windows and running Windows Container (so no shared drive settings in docker)<\/p>\n\n<p>I pulled TF serving and ran docker with parameters<\/p>\n\n<pre><code>&gt; docker pull tensorflow\/serving\n\n&gt; docker run -t --rm -d -p 8500:8500 -p 8501:8501 --name tf_model --mount type=bind,source=$LOCALMODELDIR\/tf_model,target=\/models\/mymodel --mount type=bind,source=$LOCALMODELDIR\/models.config,target=\/models\/models.config -e MODEL_NAME=mymodel tensorflow\/serving --model_config_file=\/models\/models.config\n<\/code><\/pre>\n\n<p>I ran to the Error in docker as<\/p>\n\n<pre><code>C:\\Program Files\\Docker\\Docker\\Resources\\bin\\docker.exe: Error response from daemon: invalid mount config for type \"bind\": source path must be a directory.\nSee 'C:\\Program Files\\Docker\\Docker\\Resources\\bin\\docker.exe run --help'.\n<\/code><\/pre>\n\n<p>May I know any fixes.<\/p>\n","answers":[]}
+{"QuestionId":58109257,"AnswerCount":1,"Tags":"<tensorflow><keras><conv-neural-network><convolution>","CreationDate":"2019-09-26T03:42:32.837","AcceptedAnswerId":58109542.0,"OwnerUserId":4426009.0,"Title":"How convolutional layer filter size affects input length and how to keep labels the same size","Body":"<p>Assume we have a one dimensional convolutional layer in tensorflow.keras with input size of 100 (ignoring the batch and channel dimensions for simplicity), the kernel size is 51 (in reality it could happen by applying smaller filters in multiple layers) and we want to keep the length of output the same of input (padding=same). I would like to know what indexes will be filled with zero after applying the filter, would it be 50 most right ones or 50 most left hand ones or 25 on each sides. Also if I have a label per each element what would be the best way to reduce the number of labels to 50 to still best represent the input.  <\/p>\n","answers":[{"AnswerId":"58109542","CreationDate":"2019-09-26T04:25:36.393","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>The padding is both sides, before applying the convolution, you can easily see this with this code of convolving ones with ones. (The expected result is the kernel size for the center)<\/p>\n\n<pre><code>import numpy as np\nfrom keras.layers import Input, Conv1D\nfrom keras.models import Model\n\nlength = 100\nfilt = 51\n\nins = Input((length,1))\nouts = Conv1D(1, filt, padding='same', kernel_initializer='ones', use_bias=False)(ins)\nmodel = Model(ins, outs)\n\npreds = model.predict(np.ones((1,length,1)))[0, :, 0]\nprint('result length = ', len(preds))\nprint(preds)\n<\/code><\/pre>\n\n<p>Outputs:<\/p>\n\n<pre><code>result length =  100\n[26. 27. 28. 29. 30. 31. 32. 33. 34. 35. 36. 37. 38. 39. 40. 41. 42. 43.\n 44. 45. 46. 47. 48. 49. 50. 51. 51. 51. 51. 51. 51. 51. 51. 51. 51. 51.\n 51. 51. 51. 51. 51. 51. 51. 51. 51. 51. 51. 51. 51. 51. 51. 51. 51. 51.\n 51. 51. 51. 51. 51. 51. 51. 51. 51. 51. 51. 51. 51. 51. 51. 51. 51. 51.\n 51. 51. 51. 50. 49. 48. 47. 46. 45. 44. 43. 42. 41. 40. 39. 38. 37. 36.\n 35. 34. 33. 32. 31. 30. 29. 28. 27. 26.]\n<\/code><\/pre>\n\n<p>Notice how it gradually decreases to the borders. These are the increasing number of zeros as the filters slides.<\/p>\n\n<hr>\n\n<p>There isn't a clear answer to the last question. It depends on what you need to do. \nWhat does \"best represent\" mean? You can use a <code>Dense(50)<\/code>, would it work? It depends on what you want. You could reshape 100 to 50x2 and take a mean of the new axis.... there are really a lot of solutions, each one with its own properties. <\/p>\n"}]}
+{"QuestionId":58109557,"AnswerCount":0,"Tags":"<python><tensorflow><histogram><tensorboard>","CreationDate":"2019-09-26T04:27:59.773","AcceptedAnswerId":null,"OwnerUserId":5229251.0,"Title":"Understanding vertical axis of Tensorboard histogram","Body":"<p>I was trying to use the tensorboard histogram facility to plot the weights of my trained network. My CNN is based on AlexNet, and is being used as an estimator (motivation for this can found <a href=\"https:\/\/arxiv.org\/pdf\/1808.02369.pdf\" rel=\"nofollow noreferrer\">here<\/a>). However I am having trouble understanding the histogram generated; specifically what the vertical axis means. I have viewed similar posts by others(<a href=\"https:\/\/stackoverflow.com\/questions\/42315202\/understanding-tensorboard-weight-histograms\">this one<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/35567132\/meaning-of-histogram-on-tensorboard\">this one<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/43671461\/tensorboard-v1-0-histogram-tab-interpretation\">this one<\/a>), however those either:<\/p>\n\n<ul>\n<li>talk about changes in the histogram across layers<\/li>\n<li>refer to the distribution tab on tensorboard (used to be called histogram on previous tensorboard...I think)<\/li>\n<\/ul>\n\n<p>The weights histogram at the first convolutional layer is shown below.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/BBRHL.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/BBRHL.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>What does the 61.8 at 0.0490 actually mean? I know what histograms are in general, but how can you have 61.8 of something? There was a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/5381\" rel=\"nofollow noreferrer\">github<\/a> post that mentioned that it was related to density, but doesn't really make clear what they mean by density (other than vaguely mentioning normalised density...whatever that means)? If it was probability density, then shouldn't the area under the graph equal 1? Any clarification on what the vertical axis means would be much appreciated, especially how it's supposed to be interpreted. Also, if I have missed another post that my question appears to duplicate, please let me know.  <\/p>\n","answers":[]}
+{"QuestionId":58109618,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-26T04:33:53.773","AcceptedAnswerId":null,"OwnerUserId":12122782.0,"Title":"How to connect a graph after load model with keras?","Body":"<p>I'm setting up a custom training environment in tensorflow-keras, and I want to know if is possible to reconnect shared weights of models that were saved onto different files.<\/p>\n\n<p>I have an attention encoder-decoder model, and as is known, the training model of attention is a little bit different from the predict model, but those models are sharing the same weights. First i save a non-trained model into 3 files with tf.keras.models.save_model:<\/p>\n\n<ul>\n<li>Full Model ( Training Model )<\/li>\n<li>Encoder Model ( Prediction Model )<\/li>\n<li>Decoder Model ( Prediction Model )<\/li>\n<\/ul>\n\n<p>and then, I try to use tf.keras.models.load_model to load the three above models and train only the full model as usual.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>full_model = tf.keras.models.load_model(\n    'full_model.h5'), \n    custom_objects={'AttentionLayer': AttentionLayer}\n)\nencoder_model = tf.keras.models.load_model(\n    os.path.join('encoder.h5'), \n    custom_objects={'AttentionLayer': AttentionLayer}\n)\ndecoder_model = tf.keras.models.load_model(\n    os.path.join('decoder.h5'), \n    custom_objects={'AttentionLayer': AttentionLayer}\n)\n\nfull_model.train()...\n<\/code><\/pre>\n\n<p>So, the full_model weights are updating as expected... but, the encoder and decoder weights are still freezing. Is there some way to reconnect the graph of those models?<\/p>\n","answers":[]}
+{"QuestionId":58109632,"AnswerCount":1,"Tags":"<javascript><node.js><tensorflow><tensorflow.js>","CreationDate":"2019-09-26T04:36:21.057","AcceptedAnswerId":null,"OwnerUserId":12122800.0,"Title":"How to convert wav file to spectrogram for tensorflowjs with columnTruncateLength: 232 and numFramesPerSpectrogram: 43?","Body":"<p>i'm trying to use tensorflowjs speech recognition in offline mode. online mode using microphone is working fine. but for offline mode i'm not able to find any reliable library for converting wav\/mp3 file to spectrogram according to the required specs of array as \nffttsize:1024 , \ncolumnTruncateLength: 232, \nnumFramesPerSpectrogram: 43.<\/p>\n\n<p>All libraries like spectrogram.js that i tried dont have those conversin options. while tensorlfowjs speech clearly mentions to have following specs for spectrograph tensor<\/p>\n\n<pre><code>const mic = await tf.data.microphone({\n  fftSize: 1024,\n  columnTruncateLength: 232,\n  numFramesPerSpectrogram: 43,\n  sampleRateHz:44100,\n  includeSpectrogram: true,\n  includeWaveform: true\n});\n<\/code><\/pre>\n\n<p>Getting error as Error: tensor4d() requires shape to be provided when <code>values<\/code> are a flat array in following<\/p>\n\n<pre><code>await recognizer.ensureModelLoaded();\n    var audiocaptcha = await response.buffer();\n    fs.writeFile(\".\/afterverify.mp3\", audiocaptcha, function (err) {\n        if (err) {}\n    });\n    var bufferNewSamples =  new Float32Array(audiocaptcha);\n\n    const buffersliced = bufferNewSamples.slice(0,bufferNewSamples .length-(bufferNewSamples .length%9976));\n    const xtensor = tf.tensor(bufferNewSamples).reshape([-1, \n...recognizer.modelInputShape().slice(1)]);\n<\/code><\/pre>\n\n<p>got this error after slicing and correcting to tensor<\/p>\n\n<pre><code>output.scores\n[ Float32Array [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0 ],\n  Float32Array [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0 ],\n  Float32Array [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0 ],\n  Float32Array [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0 ],\n  Float32Array [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0 ] ]\nscore for word '_background_noise_' = 0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0\nscore for word '_unknown_' = 0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0\nscore for word 'down' = 0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0\nscore for word 'eight' = 0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0\nscore for word 'five' = 0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0\nscore for word 'four' = undefined\nscore for word 'go' = undefined\nscore for word 'left' = undefined\nscore for word 'nine' = undefined\nscore for word 'no' = undefined\nscore for word 'one' = undefined\nscore for word 'right' = undefined\nscore for word 'seven' = undefined\nscore for word 'six' = undefined\nscore for word 'stop' = undefined\nscore for word 'three' = undefined\nscore for word 'two' = undefined\nscore for word 'up' = undefined\nscore for word 'yes' = undefined\nscore for word 'zero' = undefined\n<\/code><\/pre>\n","answers":[{"AnswerId":"58112756","CreationDate":"2019-09-26T08:37:06.470","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>The only requirement when working with offline recognition is to have an input tensor of shape <code>[null, 43, 232, 1]<\/code>.<\/p>\n\n<p>1 - Read the wav file and get the array of data<\/p>\n\n<pre><code>var spectrogram = require('spectrogram');\n\nvar spectro = Spectrogram(document.getElementById('canvas'), {\n  audio: {\n    enable: false\n  }\n});\n\nvar audioContext = new AudioContext();\n\nreadWavFile() {\nreturn new Promise(resove =&gt; {\nvar request = new XMLHttpRequest();\nrequest.open('GET', 'audio.mp3', true);\nrequest.responseType = 'arraybuffer';\n\nrequest.onload = function() {\n  audioContext.decodeAudioData(request.response, function(buffer) {\n    resolve(buffer)\n  });\n};\nrequest.send()\n})\n\n}\n\nconst buffer = await readWavFile()\n<\/code><\/pre>\n\n<p>The same thing can be done without using the third party library. 2 options are possible.<\/p>\n\n<ul>\n<li><p>Read the file using <code>&lt;input type=\"file\"&gt;<\/code>. In that case, this <a href=\"https:\/\/stackoverflow.com\/questions\/29317866\/read-samples-from-wav-file\">answer<\/a> shows how to get the typedarray. <\/p><\/li>\n<li><p>Serve and read the wav file using a http request<\/p><\/li>\n<\/ul>\n\n<pre class=\"lang-js prettyprint-override\"><code>var req = new XMLHttpRequest();\nreq.open(\"GET\", \"file.wav\", true);\nreq.responseType = \"arraybuffer\";\n\nreq.onload = function () {\n  var arrayBuffer = req.response;\n  if (arrayBuffer) {\n    var byteArray = new Float32Array(arrayBuffer);\n  }\n};\n\nreq.send(null);\n<\/code><\/pre>\n\n<p>2- convert the buffer to typedarray<\/p>\n\n<pre><code>const data = Float32Array(buffer)\n<\/code><\/pre>\n\n<p>3- convert the array to a tensor using the shape of the speech recognition model<\/p>\n\n<pre><code>const x = tf.tensor(\n   data).reshape([-1, ...recognizer.modelInputShape().slice(1));\n<\/code><\/pre>\n\n<p>If the latter commands fails, it means that the data does not have the shape needed for the model. The tensor needs to be sliced to have the appropriate shape or the recording made should respect the <code>fft<\/code> and other parameters.<\/p>\n"}]}
+{"QuestionId":58109638,"AnswerCount":2,"Tags":"<docker><tensorflow><tensorflow-serving><docker-for-windows><docker-image>","CreationDate":"2019-09-26T04:37:27.940","AcceptedAnswerId":58109706.0,"OwnerUserId":11053801.0,"Title":"Docker container ls not listing my pulled container?","Body":"<p>I pull the tensorflow\/serving in docker-for-windows Linux containers<\/p>\n\n<pre><code>PS C:\\WINDOWS\\system32&gt; docker pull tensorflow\/serving\nUsing default tag: latest\nlatest: Pulling from tensorflow\/serving\nDigest: sha256:f7e59a29cbc17a6b507751cddde37bccad4407c05ebf2c13b8e6ccb7d2e9affb\nStatus: Image is up to date for tensorflow\/serving:latest\ndocker.io\/tensorflow\/serving:latest\n<\/code><\/pre>\n\n<p>After that for any following commands the container is not listing up<\/p>\n\n<pre><code>PS C:\\WINDOWS\\system32&gt; docker ps\nCONTAINER ID        IMAGE               COMMAND             CREATED             STATUS              PORTS               NAMES\nPS C:\\WINDOWS\\system32&gt; docker container ls -a\nCONTAINER ID        IMAGE               COMMAND             CREATED             STATUS              PORTS               NAMES\nPS C:\\WINDOWS\\system32&gt; docker container ls\nCONTAINER ID        IMAGE               COMMAND             CREATED             STATUS              PORTS               NAMES\n<\/code><\/pre>\n\n<p>I tried restarting docker as well, may I know how to fix it?<\/p>\n","answers":[{"AnswerId":"58109763","CreationDate":"2019-09-26T04:52:29.827","ParentId":null,"OwnerUserId":"6790948","Title":null,"Body":"<p>When you run <code>docker pull tensorflow\/serving<\/code> the Docker image will get pulled, which can be listed using <code>docker images<\/code> command.<\/p>\n\n<p>While <code>docker ps<\/code>, <code>docker container ls -a<\/code>, <code>docker container ls<\/code> will list running docker container. You need to run your Docker image using <code>docker run image-name<\/code> then the container will get listed using the mentioned commands.<\/p>\n\n<p>For more info on Docker, please refer <a href=\"https:\/\/docs.docker.com\/get-started\/\" rel=\"nofollow noreferrer\">this<\/a> official guide.<\/p>\n"},{"AnswerId":"58109706","CreationDate":"2019-09-26T04:45:27.040","ParentId":null,"OwnerUserId":"9202256","Title":null,"Body":"<p><code>docker pull<\/code> is pulling the image you selected\nYou don't have a container yet.<\/p>\n\n<p><code>docker ps<\/code> and the other commands you used are referring to containers.<\/p>\n\n<p>So to run the container use:<\/p>\n\n<p><code>docker run {options} image<\/code><\/p>\n\n<p>After that you be able to see the container using <code>docker ps<\/code><\/p>\n"}]}
+{"QuestionId":58110118,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-26T05:29:16.220","AcceptedAnswerId":null,"OwnerUserId":12095031.0,"Title":"TensorFlow Dataset Method Calling Sequence In A Input Functoin","Body":"<p>There are a lot of methods in tf.data.Dataset, like batch(), shard(), shuffle(), prefetch(), map()..., etc. Usually while we implement an input_fn we will call them depends on our wish.<\/p>\n\n<p>I wonder if there is any affect on program when we call these methods in different sequence? For instance, are they same in the following two calling sequence?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>dataset = dataset.shuffle().batch()\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code>dataset = dataset.batch().shuffle()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58110308,"AnswerCount":1,"Tags":"<tensorflow><gradient><derivative><activation><relu>","CreationDate":"2019-09-26T05:48:37.497","AcceptedAnswerId":58110923.0,"OwnerUserId":5581893.0,"Title":"How to simulate ReLU gradient with tf.GradientTape","Body":"<p>TensorFlow has a feature called <strong>GradientTape<\/strong>, kinda getting gradients using Monte Carlo method(?).<\/p>\n\n<p>I'm trying to simulate the gradient of ReLU but this doesn't work on the negative half of X.<\/p>\n\n<pre><code>#colab or ipython reset\n%reset -f\n\n#libs\nimport tensorflow as tf;\n\n#init\ntf.enable_eager_execution();\n\n#code\nx = tf.convert_to_tensor([-3,-2,-1,0,1,2,3],dtype=tf.float32);\n\nwith tf.GradientTape() as t:\n  t.watch(x);\n  y = fx = x; #THIS IS JUST THE POSITIVE HALF OF X\n\ndy_dx = t.gradient(y,x);\nprint(dy_dx); \n<\/code><\/pre>\n\n<p>Guess I have to change something at the line <code>y = fx = x<\/code>, like adding a <code>if x&lt;=0<\/code> but can't figure out how.<\/p>\n\n<p>The above code prints out:<\/p>\n\n<pre><code>tf.Tensor([1. 1. 1. 1. 1. 1. 1.], shape=(7,), dtype=float32)\n<\/code><\/pre>\n\n<p>But it is wanted to be:<\/p>\n\n<pre><code>tf.Tensor([0. 0. 0. 0. 1. 1. 1.], shape=(7,), dtype=float32)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58110923","CreationDate":"2019-09-26T06:42:28.877","ParentId":null,"OwnerUserId":"5581893","Title":null,"Body":"<p>The following <code>grad<\/code> function simulate the conditional X of ReLU function but I don't know whether it's the proposed, suggested way to do it:<\/p>\n\n<pre><code>#ipython\n%reset -f\n\n#libs\nimport tensorflow as tf;\nimport numpy      as np;\n\n#init\ntf.enable_eager_execution();\n\n#code\nX = tf.convert_to_tensor([-3,-2,-1,0,1,2,3], dtype=tf.float32);\n\nwith tf.GradientTape() as T:\n  T.watch(X);\n  Y = Fx = X;\n#end with\n\nDy_Dx = T.gradient(Y,X);\n#print(Dy_Dx);\n\n#get gradient of function Fx with conditional X\ndef grad(Y,At):\n  if (At&lt;=0): return 0;\n\n  for I in range(len(X)):\n    if X[I].numpy()==At:\n      return Dy_Dx[I].numpy();\n#end def\n\nprint(grad(Y,-3));\nprint(grad(Y,-2));\nprint(grad(Y,-1));\nprint(grad(Y,-0));\nprint(grad(Y,1));\nprint(grad(Y,2));\nprint(grad(Y,3));\n\nprint(\"\\nDone.\");\n#eof\n<\/code><\/pre>\n"}]}
+{"QuestionId":58110454,"AnswerCount":1,"Tags":"<python><docker><tensorflow><deep-learning><tensorflow-serving>","CreationDate":"2019-09-26T06:02:19.160","AcceptedAnswerId":null,"OwnerUserId":4663377.0,"Title":"Tensorflow serving object detection client","Body":"<p>I have a tensorflow serving docker container running a object detector model. To generate this model I have used the following code <\/p>\n\n<pre><code>  input_img = sess.graph.get_tensor_by_name('Placeholder:0')\n  output_cls_prob = sess.graph.get_tensor_by_name('Reshape_2:0')\n  output_box_pred = sess.graph.get_tensor_by_name('rpn_bbox_pred\/Reshape_1:0')\n\n  builder = tf.saved_model.builder.SavedModelBuilder('.\/export\/1')\n\n  imageplaceholder_info = tf.saved_model.utils.build_tensor_info(input_img)\n  cls_prob_info = tf.saved_model.utils.build_tensor_info(output_cls_prob)\n  box_pred_info = tf.saved_model.utils.build_tensor_info(output_box_pred)\n  prediction_signature = (\n    tf.saved_model.signature_def_utils.build_signature_def(\n      inputs={\n        'image': imageplaceholder_info\n      },\n      outputs={\n        'output_cls_prob': cls_prob_info,\n        'output_box_pred': box_pred_info\n      },\n      method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME\n    )\n  )\n  init_op = tf.group(tf.tables_initializer(), name='legacy_init_op')\n  builder.add_meta_graph_and_variables(sess, [tf.saved_model.tag_constants.SERVING],\n       signature_def_map={'ctpn_recs_predict': prediction_signature}, legacy_init_op=init_op)\n  builder.save()\n<\/code><\/pre>\n\n<p>The docker is running this tensorflow serving model on localhost: port 9000. How do I send an image on this port number and get back a appropriate response (in my case output_cls_prob, and output_box_pred) ?<\/p>\n\n<p>So far I have this grpc client code to read an image:<\/p>\n\n<pre><code>def run(host, port, image, model, signature_name):\n\n    channel = grpc.insecure_channel('{host}:{port}'.format(host=host, port=port))\n    stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)\n\n\n    # Read an image\n    data = imread(image)\n    data = data.astype(np.float32)\n    print(data)\n\n\n\n\n    if __name__ == '__main__':\n        parser = argparse.ArgumentParser()\n        parser.add_argument('--host', help='Tensorflow server host name', default='localhost', type=str)\n        parser.add_argument('--port', help='Tensorflow server port number', default=9000, type=int)\n        parser.add_argument('--image', help='input image', type=str, default='1.jpg')\n        parser.add_argument('--model', help='model name', type=str, default='serve\/test')\n        parser.add_argument('--signature_name', help='Signature name of saved TF model',\n                            default='prediction_signature', type=str)\n\n        args = parser.parse_args()\n        run(args.host, args.port, args.image, args.model, args.signature_name)\n<\/code><\/pre>\n\n<p>what should I do so that the image read is sent to tensorflow serving docker container for prediction ?<\/p>\n","answers":[{"AnswerId":"58517958","CreationDate":"2019-10-23T07:46:27.747","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>You can use a python script that queries the TF Serving using REST and gets the response from the model.<\/p>\n\n<p>Please refer to the below blog from tensorflow, where they have mentioned complete details from hosting a TF Serving model in Docker and sending requests for prediction. <\/p>\n\n<p><a href=\"https:\/\/medium.com\/tensorflow\/serving-ml-quickly-with-tensorflow-serving-and-docker-7df7094aa008\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/tensorflow\/serving-ml-quickly-with-tensorflow-serving-and-docker-7df7094aa008<\/a><\/p>\n"}]}
+{"QuestionId":58110466,"AnswerCount":0,"Tags":"<python><keras><reinforcement-learning>","CreationDate":"2019-09-26T06:03:32.383","AcceptedAnswerId":null,"OwnerUserId":10515022.0,"Title":"Understanding monte carlo tree search","Body":"<p>So, I'm trying to create an implementation of AlphaZero using keras. However, I am not too sure about MCTS. My understanding and coding of Monte Carlo Tree Search is as follows:<\/p>\n\n<pre><code>\nclass MCTS(object):\n    def __init__(self, action_size, movesets, nsims, ndepth):\n        self.nsims = nsims\n        self.ndepth = ndepth\n        self.movesets = movesets\n        self.action_size = action_size\n\n    def evaluate_and_act(self, agent, stateizer, critic, state):\n        sims = []\n\n        print(\"Beginning monte carlo tree search\")\n\n        true_state = state\n        for i in range(self.nsims):\n            random_walk = []\n            for j in range(self.ndepth):\n                random_actions = []\n                print(\"Searching depth\", j, \"of simulation\", i)\n\n                for k in range(self.movesets):\n                    rand_move = np.random.choice(self.action_size)\n                    rand_move_matrix = cp.add(cp.zeros((1, self.action_size)), .0001)\n                    rand_move_matrix[0][rand_move] = critic.predict(state, batch_size=64)[0][0]\n                    random_actions.append(cp.asnumpy(rand_move_matrix))\n                random_action_concat = np.concatenate(random_actions, -1)\n                state = stateizer.predict(cp.asnumpy(random_action_concat), batch_size=64)\n                random_walk.append(random_actions)\n            sims.append(random_walk)\n            state = true_state\n\n        best_reward = -1000000.0\n        for walk in sims:\n\n            sum_reward = np.sum(walk)\n            if sum_reward &gt;= best_reward:\n                best_walk = walk\n                best_reward = sum_reward\n\n        return best_walk[0]\n<\/code><\/pre>\n\n<p>It seems like I don't need the policy network at all in this implementation, just the critic. Can someone please help me with understanding whether or not my implementation is correct, and why it's incorrect in terms of AlphaZero? Thanks.  <\/p>\n","answers":[]}
+{"QuestionId":58110537,"AnswerCount":1,"Tags":"<python><tensorflow><syntax><gradient><with-statement>","CreationDate":"2019-09-26T06:11:29.400","AcceptedAnswerId":58110699.0,"OwnerUserId":5581893.0,"Title":"tf.GradientTape can't watch outside 'with' block","Body":"<p>TensorFlow tutorials on tensorflow.org show the way to use <code>tf.GradientTape<\/code> as:<\/p>\n\n<pre><code>x = tf.convert_to_tensor([1,2,3]);\nwith tf.GradientTape() as t:\n  t.watch(x); \n<\/code><\/pre>\n\n<p>I wonder why I can't move the <code>t.watch(x)<\/code> outside of <code>with<\/code> block like this:<\/p>\n\n<pre><code>x = tf.convert_to_tensor([1,2,3]);\nt = tf.GradientTape();\nt.watch(x); #ERROR\n<\/code><\/pre>\n\n<p>The error is:<\/p>\n\n<pre><code>tape.py (59):\npywrap_tensorflow.TFE_Py_TapeWatch(tape._tape, tensor)\n\nAttributeError: 'NoneType' object has no attribute '_tape'\n<\/code><\/pre>\n","answers":[{"AnswerId":"58110699","CreationDate":"2019-09-26T06:25:01.390","ParentId":null,"OwnerUserId":"5581893","Title":null,"Body":"<p>I found out how. The <code>tf.GradientTape<\/code> class is designed to work within <code>with<\/code> block, ie. it has <strong>enter<\/strong> and <strong>exit<\/strong> methods. <\/p>\n\n<p>So to have it working outside 'with' block, method <code>__enter__<\/code> must be explicitly called, however, avoid calling <code>__enter__<\/code> directly:<\/p>\n\n<pre><code>x = tf.convert_to_tensor([1,2,3]);\nt = tf.GradientTape();\nt.__enter__();\nt.watch(x); \n<\/code><\/pre>\n\n<p>Reference:\n<a href=\"https:\/\/stackoverflow.com\/questions\/1984325\/explaining-pythons-enter-and-exit\">Explaining Python&#39;s &#39;__enter__&#39; and &#39;__exit__&#39;<\/a><\/p>\n"}]}
+{"QuestionId":58111350,"AnswerCount":1,"Tags":"<tensorflow><keras><conv-neural-network><loss-function>","CreationDate":"2019-09-26T07:11:16.790","AcceptedAnswerId":null,"OwnerUserId":1031417.0,"Title":"Why does the output layer is simply zero at the end of the network?","Body":"<p>I am trying to train a model that takes a 15x15 image and classify each pixel into two classes (1\/0).<\/p>\n\n<p>This is my loss function:<\/p>\n\n<pre><code>smooth = 1\ndef tversky(y_true, y_pred):\n    y_true_pos = K.flatten(y_true)\n    y_pred_pos = K.flatten(y_pred)\n    true_pos = K.sum(y_true_pos * y_pred_pos)\n    false_neg = K.sum(y_true_pos * (1-y_pred_pos))\n    false_pos = K.sum((1-y_true_pos)*y_pred_pos)\n    alpha = 0.5\n    return (true_pos + smooth)\/(true_pos + alpha*false_neg + (1-alpha)*false_pos + smooth)\n\ndef tversky_loss2(y_true, y_pred):\n    return 1 - tversky(y_true,y_pred)\n<\/code><\/pre>\n\n<p>This is the model:<\/p>\n\n<pre><code>input_image = layers.Input(shape=(size, size, 1))\n\nb2 = layers.Conv2D(128, (3,3), padding='same',  activation='relu')(input_image)\nb2 = layers.Conv2D(128, (3,3), padding='same',  activation='relu')(b2)\nb2 = layers.Conv2D(128, (3,3), padding='same',  activation='relu')(b2)\n\noutput = layers.Conv2D(1, (1,1), activation='sigmoid', padding='same')(b2)\n\nmodel = models.Model(input_image, output)\nmodel.compile(optimizer='adam', loss=tversky_loss2, metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>The model left is the input and the label is the middle column and the prediction is always zero on the right column:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Qnm3M.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Qnm3M.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The training performs really poorly:<\/p>\n\n<pre><code>Epoch 1\/10\n100\/100 [==============================] - 4s 38ms\/step - loss: 0.9269 - acc: 0.1825\nEpoch 2\/10\n100\/100 [==============================] - 3s 29ms\/step - loss: 0.9277 - acc: 0.0238\nEpoch 3\/10\n100\/100 [==============================] - 3s 29ms\/step - loss: 0.9276 - acc: 0.0239\nEpoch 4\/10\n100\/100 [==============================] - 3s 29ms\/step - loss: 0.9270 - acc: 0.0241\nEpoch 5\/10\n100\/100 [==============================] - 3s 30ms\/step - loss: 0.9274 - acc: 0.0240\nEpoch 6\/10\n100\/100 [==============================] - 3s 29ms\/step - loss: 0.9269 - acc: 0.0242\nEpoch 7\/10\n100\/100 [==============================] - 3s 29ms\/step - loss: 0.9270 - acc: 0.0241\nEpoch 8\/10\n100\/100 [==============================] - 3s 29ms\/step - loss: 0.9271 - acc: 0.0241\nEpoch 9\/10\n100\/100 [==============================] - 3s 29ms\/step - loss: 0.9276 - acc: 0.0239\nEpoch 10\/10\n100\/100 [==============================] - 3s 29ms\/step - loss: 0.9266 - acc: 0.0242\n<\/code><\/pre>\n","answers":[{"AnswerId":"58117420","CreationDate":"2019-09-26T13:03:23.093","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>This sounds like a very imbalanced dataset with very tiny true regions. This might be hard to train indeed.   <\/p>\n\n<p>You may want to increase <code>alpha<\/code> to penalize more false negatives than false positives. Anyway, unless alpha is big enough, it's very normal that in the beginning your model first goes to all neg because it's definitely a great way to decrease the loss.<\/p>\n\n<p>Now, there is a conceptual mistake regarding how Keras works in that loss. You need to keep the \"samples\" separate. Otherwise you are calculating a loss as if all images were one image. (Thus, it's probable that images with many positives have a reasoable result, while images with few positives don't, and this will be a good solution)<\/p>\n\n<p>Fix the loss as:<\/p>\n\n<pre><code>\ndef tversky(y_true, y_pred):\n    y_true_pos = K.batch_flatten(y_true) #keep the batch dimension\n    y_pred_pos = K.batch_flatten(y_pred)\n\n    true_pos = K.sum(y_true_pos * y_pred_pos, axis=-1) #don't sum over the batch dimension   \n    false_neg = K.sum(y_true_pos * (1-y_pred_pos), axis=-1)\n    false_pos = K.sum((1-y_true_pos)*y_pred_pos, axis=-1)\n    alpha = 0.5\n    return (true_pos + smooth)\/(true_pos + alpha*false_neg + (1-alpha)*false_pos + smooth)\n\n<\/code><\/pre>\n\n<p>This way you have an individual loss value for each image, so the exitence of images with many positives don't affect the results of images with few positives. <\/p>\n"}]}
+{"QuestionId":58111456,"AnswerCount":2,"Tags":"<python><python-3.x><list><tensorflow>","CreationDate":"2019-09-26T07:17:44.090","AcceptedAnswerId":58113535.0,"OwnerUserId":12123347.0,"Title":"How to print out the correct predicted category?","Body":"<p>Currently able to get the correct prediction, however the printed category string is wrong. <\/p>\n\n<p>If there is only 2 categories then this code will run perfectly, however now im using 3 categories.<\/p>\n\n<pre><code>CATEGORIES = [\"RGB images score 1\", \"RGB images score 2\", \"RGB images score 3\"]\n\nprediction = model.predict([prepare('score3.png')])\n\nprint(prediction[0])  # will be a list in a list.\n\nprint(CATEGORIES[int(prediction[0][1])])\n<\/code><\/pre>\n\n<p>OUTPUT<\/p>\n\n<pre><code>[0. 0. 1.]\n\nRGB images score 1\n<\/code><\/pre>\n\n<p>Actual output should be \"RGB images score 3\". However I got \"RGB images score 1\" instead. Only score 3 images have this problem.<\/p>\n","answers":[{"AnswerId":"58113533","CreationDate":"2019-09-26T09:21:19.880","ParentId":null,"OwnerUserId":"12123347","Title":null,"Body":"<p>Managed to solve it using an if else <\/p>\n\n<pre><code>CATEGORIES = [\"RGB images score 1\", \"RGB images score 2\", \"RGB images score 3\"]\n\nprediction = model.predict([prepare('score3.png')])\n\nif prediction[0][2] == 1.0:\n    print(\"RGB images score 3\")\n\nelse:\n    print(CATEGORIES[int(prediction[0][1])])\n<\/code><\/pre>\n"},{"AnswerId":"58113535","CreationDate":"2019-09-26T09:21:23.610","ParentId":null,"OwnerUserId":"8244247","Title":null,"Body":"<p>You are looking at <code>prediction[0][1]<\/code> which is litterally : did the network predict <code>2<\/code>.<\/p>\n\n<p>It indeed works for 2 categories, but not for more ! You need to find the index for which prediction[0] is equal to 1.<\/p>\n\n<p>You could use for instance <code>print(CATEGORIES[int(np.argmax(prediction[0]))])<\/code><\/p>\n"}]}
+{"QuestionId":58112355,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-26T08:13:33.637","AcceptedAnswerId":null,"OwnerUserId":6691564.0,"Title":"What, exactly, is eager execution from a programming point of view?","Body":"<p>I am trying to understand eager execution. Pages returned by Google describe what it does for you, and I'm ok with that. I am trying to understand it from the point of view of program code. Here is an example from <a href=\"https:\/\/towardsdatascience.com\/eager-execution-tensorflow-8042128ca7be\" rel=\"nofollow noreferrer\">this article<\/a>.<\/p>\n\n<pre><code>a = tf.constant([[1,2],[3,4]])\n<\/code><\/pre>\n\n<p>The article says this statement does something different depending on whether you are in eager mode or not. Without eager mode, print(a) gives:<\/p>\n\n<pre><code>Tensor(\"Const:0\", shape=(2, 2), dtype=int32)\n<\/code><\/pre>\n\n<p>With eager mode, print(a) gives:<\/p>\n\n<pre><code>tf.Tensor(\n[[1 2]\n [3 4]], shape=(2, 2), dtype=int32)\n<\/code><\/pre>\n\n<p>Please could someone explain what these two return values are. If they are two different object types, a Tensor and a tf.Tensor, what is the difference between these objects?<\/p>\n\n<p>I have searched the TensorFlow documentation and can't see anything that addresses this distinction. Any pointers gratefully received.<\/p>\n\n<p>Thanks,<\/p>\n\n<p>Julian<\/p>\n","answers":[{"AnswerId":"58112468","CreationDate":"2019-09-26T08:19:43.560","ParentId":null,"OwnerUserId":"240443","Title":null,"Body":"<p>Lazy execution delays evaluation as much as possible. In context of TensorFlow, it will create a plan of execution (a graph) before it does anything, then when everything is ready, it will feed the input into the graph and calculate everything to return the output. <code>tf.constant<\/code> creates an execution node in the graph that will receive a constant value when the execution starts. You can compare lazy evaluation to a Rube Goldberg machine: you build the whole thing, then you drop a marble into it and watch the magic unfold.<\/p>\n\n<p>Eager execution evaluates immediately. In context of TensorFlow, it does not create a graph. <code>tf.constant<\/code> creates a constant, with a certain value, right then and there. If compared to the Rube Goldberg machine, you could imagine building the first part of the machine, drop the marble in, then stop it when the marble exits, before building the next part of the machine, and reintroducing the marble with the same position and velocity. :P<\/p>\n\n<p>TensorFlow is much more efficient in the lazy evaluation mode, but the eager mode is easier to develop and debug on.<\/p>\n"}]}
+{"QuestionId":58112426,"AnswerCount":0,"Tags":"<python><deep-learning><theano><reinforcement-learning><lasagne>","CreationDate":"2019-09-26T08:17:43.550","AcceptedAnswerId":null,"OwnerUserId":7012361.0,"Title":"Lasagne + Theano Dimshuffle ValueError drop a non-broadcastable dimension","Body":"<p>I'm new to deep reinforcement learning and have an issue using Theano an Lasagne to reproduce the model described here : <a href=\"https:\/\/openreview.net\/pdf?id=SklPFVXUsN\" rel=\"nofollow noreferrer\">https:\/\/openreview.net\/pdf?id=SklPFVXUsN<\/a> page 5 as it is close to my own context.<\/p>\n\n<p>Here is the code I use to create the model layers:<\/p>\n\n<pre><code>    def build_crafted_network(self, l_carte, h_carte, l_minimap, h_minimap, l_tinymap, h_tinymap, nb_properties,\n                          output_dim, batch_size):\n\n    self.l_in_carte = lasagne.layers.InputLayer(\n        shape=(batch_size, l_carte, h_carte)\n    )\n    self.l_in_minimap = lasagne.layers.InputLayer(\n        shape=(batch_size, l_minimap, h_minimap)\n    )\n    self.l_in_tinymap = lasagne.layers.InputLayer(\n        shape=(batch_size, l_tinymap, h_tinymap)\n    )\n    self.l_in_properties = lasagne.layers.InputLayer(\n        shape=(batch_size, 1, nb_properties)\n    )\n    l_embedding_carte = lasagne.layers.EmbeddingLayer(\n        self.l_in_carte,\n        input_size=l_carte*h_carte,\n        output_size=32\n    )\n    l_embedding_minimap = lasagne.layers.EmbeddingLayer(\n        self.l_in_minimap,\n        input_size=l_minimap*h_minimap,\n        output_size=32\n    )\n    l_embedding_tinymap = lasagne.layers.EmbeddingLayer(\n        self.l_in_tinymap,\n        input_size=l_tinymap*h_tinymap,\n        output_size=32\n    )\n    l_embedding_properties = lasagne.layers.EmbeddingLayer(\n        self.l_in_properties,\n        input_size=nb_properties,\n        output_size=64\n    )\n    l_dimshuffle_carte = lasagne.layers.DimshuffleLayer(\n        l_embedding_carte,\n        (0,3,1,2)\n    )\n    l_dimshuffle_minimap= lasagne.layers.DimshuffleLayer(\n        l_embedding_minimap,\n        (0,3,1,2)\n    )\n    l_dimshuffle_tinymap = lasagne.layers.DimshuffleLayer(\n        l_embedding_tinymap,\n        (0,3,1,2)\n    )\n    l_conv_carte = lasagne.layers.Conv2DLayer(\n        l_dimshuffle_carte,\n        num_filters=32,\n        filter_size=(3, 3),\n        nonlinearity=lasagne.nonlinearities.rectify\n    )\n    l_conv_minimap = lasagne.layers.Conv2DLayer(\n        l_dimshuffle_minimap,\n        num_filters=32,\n        filter_size=(3, 3),\n        nonlinearity=lasagne.nonlinearities.rectify\n    )\n    l_conv_tinymap = lasagne.layers.Conv2DLayer(\n        l_dimshuffle_tinymap,\n        num_filters=32,\n        filter_size=(3, 3),\n        nonlinearity=lasagne.nonlinearities.rectify,\n        pad=1\n    )\n    l_conv2_carte = lasagne.layers.Conv2DLayer(\n        l_conv_carte,\n        num_filters=64,\n        filter_size=(3, 3),\n        nonlinearity=lasagne.nonlinearities.rectify\n    )\n    l_conv2_minimap = lasagne.layers.Conv2DLayer(\n        l_conv_minimap,\n        num_filters=64,\n        filter_size=(3, 3),\n        nonlinearity=lasagne.nonlinearities.rectify\n    )\n    l_conv2_tinymap = lasagne.layers.Conv2DLayer(\n        l_conv_tinymap,\n        num_filters=64,\n        filter_size=(3, 3),\n        nonlinearity=lasagne.nonlinearities.rectify,\n        pad=1\n    )\n    l_fc_properties = lasagne.layers.DenseLayer(\n        l_embedding_properties,\n        num_units=256,\n        nonlinearity=None,\n        W=lasagne.init.Constant(0.0),\n        b=None\n    )\n    l_flatten_carte = lasagne.layers.FlattenLayer(\n        l_conv2_carte, 2\n    )\n    l_flatten_minimap = lasagne.layers.FlattenLayer(\n        l_conv2_minimap, 2\n    )\n    l_flatten_tinymap = lasagne.layers.FlattenLayer(\n        l_conv2_tinymap, 2\n    )\n    l_merge_layer = lasagne.layers.ConcatLayer(\n        (l_flatten_carte, l_flatten_minimap, l_flatten_tinymap, l_fc_properties),\n        axis=1\n    )\n    l_hidden_layer = lasagne.layers.DenseLayer(\n        l_merge_layer,\n        num_units=256,\n        nonlinearity=None,\n        W=lasagne.init.Constant(0.0),\n        b=lasagne.init.Constant(.1)\n    )\n\n    l_out = lasagne.layers.DenseLayer(\n        l_hidden_layer,\n        num_units=output_dim,\n        nonlinearity=T.nnet.softmax,\n        W=lasagne.init.Normal(.01),\n        b=lasagne.init.Constant(.1)\n    )\n    return l_out\n<\/code><\/pre>\n\n<p>The build_network completes without errors but then it crashes at<\/p>\n\n<pre><code>states_carte = T.tensor4('states_carte', dtype='int8')      \nstates_minimap = T.tensor4('states_minimap', dtype='int8')\nstates_tinymap = T.tensor4('states_tinymap', dtype='int8')\nstates_properties = T.tensor4('states_properties', dtype='int16')\nstates_dict = {self.l_in_carte:states_carte, self.l_in_minimap:states_minimap, self.l_in_tinymap:states_tinymap, self.l_in_properties:states_properties}\nq_vals = lasagne.layers.get_output(self.l_out, inputs = states_dict)\n<\/code><\/pre>\n\n<p>with the following error :<\/p>\n\n<pre><code>    Traceback (most recent call last):\n  File \"stester.py\", line 99, in &lt;module&gt;\n    np.random.RandomState(123456))\n  File \"\/home\/barre\/Private\/DeepDofus\/deep_q_learn.py\", line 121, in __init__\n    q_vals = lasagne.layers.get_output(self.l_out, inputs = states_dict)\n  File \"\/home\/barre\/anaconda3\/envs\/deepdofus\/lib\/python3.7\/site-packages\/lasagne\/layers\/helper.py\", line 197, in get_output\n    all_outputs[layer] = layer.get_output_for(layer_inputs, **kwargs)\n  File \"\/home\/barre\/anaconda3\/envs\/deepdofus\/lib\/python3.7\/site-packages\/lasagne\/layers\/shape.py\", line 284, in get_output_for\n    return input.dimshuffle(self.pattern)\n  File \"\/home\/barre\/anaconda3\/envs\/deepdofus\/lib\/python3.7\/site-packages\/theano\/tensor\/var.py\", line 358, in dimshuffle\n    pattern)\n  File \"\/home\/barre\/anaconda3\/envs\/deepdofus\/lib\/python3.7\/site-packages\/theano\/tensor\/elemwise.py\", line 197, in __init__\n    (input_broadcastable, new_order))\nValueError: ('You cannot drop a non-broadcastable dimension.', ([False, False, False, False, False], (0, 3, 1, 2)))\n<\/code><\/pre>\n\n<p>It seems that the dimshuffle layer is not properly working, but I don't understand why as the 2D convolutional layer requires a input like (batch_size, num_input_channels, input_rows, input_columns)<\/p>\n\n<p>I'm using those versions:\ntheano                    1.0.4+21.g8f510a1f0          pypi_0    pypi\nlasagne                   0.2.dev1                 pypi_0    pypi\npython                    3.7.4                h265db76_1  <\/p>\n\n<p>The full code is available at <a href=\"https:\/\/pastebin.com\/7z1uwvs8\" rel=\"nofollow noreferrer\">https:\/\/pastebin.com\/7z1uwvs8<\/a><\/p>\n","answers":[]}
+{"QuestionId":58112658,"AnswerCount":1,"Tags":"<python><pytorch><training-data><dataloader>","CreationDate":"2019-09-26T08:31:40.243","AcceptedAnswerId":58112772.0,"OwnerUserId":7648650.0,"Title":"How to use DataLoader for PyTorch on iPython Console of Spyder","Body":"<p>I check <a href=\"https:\/\/www.youtube.com\/watch?v=zN49HdDxHi8\" rel=\"nofollow noreferrer\">this Tutorial<\/a> and can't figure out a way to actually use my DataLoader to train a ANN.\nWhen iterating over my DataLoader a cmd prompt pops up and immediately closes itself, afterwards nothing happens. My original data are both np.arrays.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ED8CS.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ED8CS.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><code>import torch\nfrom torch.utils import data\nimport numpy as np\n\nclass Dataset(data.Dataset):\n  'Characterizes a dataset for PyTorch'\n  def __init__(self, datax, labels):\n        'Initialization'\n        self.labels = torch.tensor(labels)\n        self.datax = torch.tensor(datax)\n        self.len = len(datax)\n\n  def __len__(self):\n        'Denotes the total number of samples'\n        return self.len\n\n  def __getitem__(self, index):\n        'Generates one sample of data'\n        # Load data and get label\n        X = self.datax[index]\n        y = self.labels[index]\n        return X, y\n\nparams = {'batch_size': 64,\n          'shuffle': True,\n          'num_workers': 1}\ntraining_set = Dataset(datax=X, labels=labels)\ntraining_generator = data.DataLoader(training_set, **params)\n\nfor x in training_generator:\n    print(1)\n<\/code><\/pre>\n\n<p>I tried many times and had a glimpse at the commandprompt which says something like<\/p>\n\n<pre><code>OMP: Info #212: KMP_AFFINITY: decoding x2APIC ids.\nOMP: Info #210: KMP_AFFINITY: Affinity capable, using global cpuid leaf 11 info\nOMP: Info #154: KMP_AFFINITY: Initial OS proc set respected: 0\nOMP: Info #156: KMP_AFFINITY: 4 available OS procs\nOMP: Info #157: KMP_AFFINITY: Uniform topology\nOMP: Info #179: KMP_AFFINITY: 1 packages x 2 cores\/pkg x 2 threads\/core (2 total cores)\nOMP: Info #214: KMP_AFFINITY: OS proc to physical thread map:\nOMP: Info #171: KMP_AFFINITY: OS proc 0 maps to package 0 core 0 thread 0 \nOMP: Info #171: KMP_AFFINITY: OS proc 1 maps to package 0 core 0 thread 1 \nOMP: Info #171: KMP_AFFINITY: OS proc 2 maps to package 0 core 1 thread 0 \nOMP: Info #171: KMP_AFFINITY: OS proc 3 maps to package 0 core 1 thread 1 \nOMP: Info #250: KMP_AFFINITY: pid 10264 tid 2388 thread 0 bound to OS proc set 0\nOMP: Info #250: KMP_AFFINITY: pid 10264 tid 3288 thread 1 bound to OS proc set 2\n<\/code><\/pre>\n","answers":[{"AnswerId":"58112772","CreationDate":"2019-09-26T08:38:08.843","ParentId":null,"OwnerUserId":"5324310","Title":null,"Body":"<p>Here is how I do that:<\/p>\n\n<pre><code>class myDataset(Dataset):\n    '''\n    a dataset for PyTorch \n    '''\n    def __init__(self, X, y):\n        self.X = X\n        self.y = y\n    def __getitem__(self, index):\n        return self.X[index], self.y[index]\n    def __len__(self):\n        return len(self.X)\n<\/code><\/pre>\n\n<p>then you can simple add to the loader: <\/p>\n\n<pre><code>full_dataset = myDataset(X,y)\ntrain_loader = DataLoader(full_dataset, batch_size=batch_size)\n<\/code><\/pre>\n\n<p>Also, X, y are just <strong>numpy arrays<\/strong>. <\/p>\n\n<p>And for the training you can access your data with a for loop: <\/p>\n\n<pre><code>for data, target in train_loader:\n        if train_on_gpu:\n            data, target = data.double().cuda(), target.double().cuda()\n<\/code><\/pre>\n"}]}
+{"QuestionId":58112840,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><text-classification><tflearn>","CreationDate":"2019-09-26T08:41:50.873","AcceptedAnswerId":null,"OwnerUserId":8470765.0,"Title":"How to update existing model with new training data in TFlearn","Body":"<p>I'm new for Tensorflow , I'm using This tutorial- <a href=\"https:\/\/sourcedexter.com\/tensorflow-text-classification-python\/\" rel=\"nofollow noreferrer\">https:\/\/sourcedexter.com\/tensorflow-text-classification-python\/<\/a> for text classification build model.<\/p>\n\n<p>i'm saving the model using below code.<\/p>\n\n<pre><code>net = 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)\nmodel = tflearn.DNN(net, tensorboard_dir='tflearn_logs')\nmodel.fit(train_x, train_y, n_epoch=1000, batch_size=8, show_metric=True)\nmodel.save('.\/model.tflearn',)\n<\/code><\/pre>\n\n<p>This gives below files.<\/p>\n\n<pre><code>model.tflearn.data-00000-of-00001\nmodel.tflearn.index\nmodel.tflearn.meta\ntflearn_logs folder\n<\/code><\/pre>\n\n<p>But when i save Existing model(with new Training dataset). i'm getting prediction value only for new trained datas not for existing datas. How to append new train dataset in existing save model.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[]}
+{"QuestionId":58113043,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><tensorboard><graph-visualization>","CreationDate":"2019-09-26T08:54:03.863","AcceptedAnswerId":null,"OwnerUserId":8270077.0,"Title":"Visualizing a Keras Model with Tensorboard in Tensorflow 2.0","Body":"<p>I am learning to visualize a Keras model using Tensorboard with Tensorflow 2.0<\/p>\n\n<p>For a reproducible example:<\/p>\n\n<pre><code>inputs = Input(shape = (train_data.shape[1], ))\nx1 = Dense(100, kernel_initializer = 'he_normal', activation = 'elu')(inputs)\nx1a = Dropout(0.5)(x1)\nx2 = Dense(100, kernel_initializer = 'he_normal', activation = 'elu')(x1a)\nx2a = Dropout(0.5)(x2)\nx3 = Dense(100, kernel_initializer = 'he_normal', activation = 'elu')(x2a)\nx3a = Dropout(0.5)(x3)\nx4 = Dense(100, kernel_initializer = 'he_normal', activation = 'elu')(x3a)\nx4a = Dropout(0.5)(x4)\nx5 = Dense(100, kernel_initializer = 'he_normal', activation = 'elu')(x4a)\npredictions = Dense(1)(x5)\nmodel = Model(inputs = inputs, outputs = predictions)\n\n@tf.function\ndef traceme(x):\n    return model(x)\n\n\nlogdir = \"log\"\nwriter = tf.summary.create_file_writer(logdir)\ntf.summary.trace_on(graph=True, profiler=True)\n# Forward pass\ntraceme(tf.zeros((1, train_data.shape[1])))\nwith writer.as_default():\n    tf.summary.trace_export(name=\"model_trace\", step=0, profiler_outdir=logdir)\n<\/code><\/pre>\n\n<p>How can I proceed from this point to get the visualization?<\/p>\n\n<hr>\n\n<p>When I run the command from  my Jupyter Notebook I get an error message:<\/p>\n\n<pre><code>%load_ext tensorboard.notebook\n%tensorboard --logdir logs\n\nERROR: Failed to launch TensorBoard (exited with 1).\nContents of stderr:\nc:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\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\nTraceback (most recent call last):\n  File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\runpy.py\", line 193, in _run_module_as_main\n    \"__main__\", mod_spec)\n  File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\runpy.py\", line 85, in _run_code\n    exec(code, run_globals)\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Anaconda3_64\\Scripts\\tensorboard.exe\\__main__.py\", line 5, in &lt;module&gt;\n  File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\site-packages\\tensorboard\\main.py\", line 44, in &lt;module&gt;\n    from tensorboard import default\n  File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\site-packages\\tensorboard\\default.py\", line 36, in &lt;module&gt;\n    from tensorboard.plugins.beholder import beholder_plugin\n  File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\site-packages\\tensorboard\\plugins\\beholder\\__init__.py\", line 15, in &lt;module&gt;\n    from tensorboard.plugins.beholder.beholder import Beholder\n  File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\site-packages\\tensorboard\\plugins\\beholder\\beholder.py\", line 199, in &lt;module&gt;\n    class BeholderHook(tf.estimator.SessionRunHook):\nAttributeError: module 'tensorflow' has no attribute 'estimator'\n<\/code><\/pre>\n\n<p>When I run it from the terminal window again I get an exception:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ny8SB.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ny8SB.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><code>(base) C:\\WINDOWS\\system32&gt;tensorboard --logdir log\nc:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\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\nTraceback (most recent call last):\n  File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\runpy.py\", line 193, in _run_module_as_main\n    \"__main__\", mod_spec)\n  File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\runpy.py\", line 85, in _run_code\n    exec(code, run_globals)\n  File \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Anaconda3_64\\Scripts\\tensorboard.exe\\__main__.py\", line 5, in &lt;module&gt;\n  File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\site-packages\\tensorboard\\main.py\", line 44, in &lt;module&gt;\n    from tensorboard import default\n  File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\site-packages\\tensorboard\\default.py\", line 36, in &lt;module&gt;\n    from tensorboard.plugins.beholder import beholder_plugin\n  File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\site-packages\\tensorboard\\plugins\\beholder\\__init__.py\", line 15, in &lt;module&gt;\n    from tensorboard.plugins.beholder.beholder import Beholder\n  File \"c:\\program files (x86)\\microsoft visual studio\\shared\\anaconda3_64\\lib\\site-packages\\tensorboard\\plugins\\beholder\\beholder.py\", line 199, in &lt;module&gt;\n    class BeholderHook(tf.estimator.SessionRunHook):\nAttributeError: module 'tensorflow' has no attribute 'estimator'\n<\/code><\/pre>\n","answers":[{"AnswerId":"58113279","CreationDate":"2019-09-26T09:06:45.640","ParentId":null,"OwnerUserId":"5597718","Title":null,"Body":"<blockquote>\n  <p>Could you advise me how I proceed from this point to get the\n  visualization?<\/p>\n<\/blockquote>\n\n<p>In jupyter notebook, <\/p>\n\n<pre><code># Load the TensorBoard notebook extension.\n%load_ext tensorboard\n\n%tensorboard --logdir log\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/tensorboard\/r2\/graphs#graphs_of_tffunctions\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tensorboard\/r2\/graphs#graphs_of_tffunctions<\/a><\/p>\n\n<p>You can also do this from a terminal by running this command,<\/p>\n\n<pre><code>tensorboard --logdir log\n<\/code><\/pre>\n\n<p>and opening <code>'localhost:6006'<\/code> from a web browser.<\/p>\n\n<p>If the Anaconda environment is correct, check this link for a way to resolve the exception: <a href=\"https:\/\/github.com\/tensorflow\/tensorboard\/issues\/2067#issuecomment-478140779\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorboard\/issues\/2067#issuecomment-478140779<\/a><\/p>\n"}]}
+{"QuestionId":58113098,"AnswerCount":1,"Tags":"<python><python-3.x><pytorch>","CreationDate":"2019-09-26T08:57:07.310","AcceptedAnswerId":null,"OwnerUserId":2769240.0,"Title":"Increasing the size of images displayed in Pytorch","Body":"<p>I want to display few images and their respective labels using Pytorch dataloader. <\/p>\n\n<p>However the image displayed is very tiny grid. <\/p>\n\n<p>How do I increase the width of each image so it's bigger. <\/p>\n\n<p>Here's the code I used: <\/p>\n\n<pre><code>mean_nums = [0.485, 0.456, 0.406]\nstd_nums = [0.229, 0.224, 0.225]\n\ndef imshow(inp, title=None):\n    \"\"\"Imshow for Tensor.\"\"\"\n    inp = inp.numpy().transpose((1, 2, 0))\n    mean = np.array(mean_nums)\n    std = np.array(std_nums)\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['trainLoader']))\n\n# Make a grid from batch\nout = torchvision.utils.make_grid(inputs,nrow=2)\n\nimshow(out, title=[image_datasets['train'].classes[x] for x in classes])\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/XcxSt.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/XcxSt.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"58115647","CreationDate":"2019-09-26T11:17:11.523","ParentId":null,"OwnerUserId":"12054165","Title":null,"Body":"<p>Try to insert <code>plt.figure(figsize=[width, height])<\/code> before <code>plt.imshow<\/code> And choose the <code>width, height<\/code>  that will satisfy you.<\/p>\n\n<p>So, for example, the <code>imshow<\/code> function may be:<\/p>\n\n<pre><code>def imshow(inp, title=None):\n    \"\"\"Imshow for Tensor.\"\"\"\n    inp = inp.numpy().transpose((1, 2, 0))\n    mean = np.array(mean_nums)\n    std = np.array(std_nums)\n    inp = std * inp + mean\n    inp = np.clip(inp, 0, 1)\n    plt.figure(figsize=[20, 20])\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<\/code><\/pre>\n"}]}
+{"QuestionId":58113387,"AnswerCount":1,"Tags":"<python><tensorflow><keras><version><optimizer-statistics>","CreationDate":"2019-09-26T09:12:29.713","AcceptedAnswerId":58114679.0,"OwnerUserId":5581893.0,"Title":"What can be used to replace tf.train.GradientDescentOptimizer in TensorFlow 2","Body":"<p>I'm trying to convert my TensorFlow 1.4 code to TensorFlow 2 but in there's no more <code>tf.train.GradientDescentOptimizer<\/code> in TF 2.<\/p>\n\n<p>What should be the replacement for tf.train.GradientDescentOptimizer?<\/p>\n\n<p>I found the same thing in <code>tf.compat.v1.train.<\/code> but I shouldn't use it as those stuff in <code>tf.compat<\/code> will be removed sooner or later.<\/p>\n","answers":[{"AnswerId":"58114679","CreationDate":"2019-09-26T10:21:32.300","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>You can replace it with, <code>tf.keras.optimizers.SGD()<\/code> defined <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/keras\/optimizers\/SGD\" rel=\"nofollow noreferrer\">here<\/a>.\n<a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/migration_guide#keras_optimizers\" rel=\"nofollow noreferrer\">Here<\/a> (skip to third point) is the official message, where TF team mentioned to use this keras optimizer. <\/p>\n"}]}
+{"QuestionId":58113394,"AnswerCount":0,"Tags":"<keras><deep-learning><batch-normalization><dcgan><gan>","CreationDate":"2019-09-26T09:13:08.503","AcceptedAnswerId":null,"OwnerUserId":10452691.0,"Title":"Deep Learning, Strange batch normalization","Body":"<p>The below is my Generator structure.\nAs you see, if I remove the batch normalization, GAN extremely works well.\nBut if I add Batch Normalization at the commented place, it shows only noise.<\/p>\n\n<p>I don't know why. I have tried add BN to generator only or discriminator only or both too.\nEven when I add only one BN layer into anywhere, it never works.<\/p>\n\n<p>I am trying to fix the source at git <a href=\"https:\/\/github.com\/rickiepark\/deep-learning-with-python-notebooks\/blob\/master\/8.5-introduction-to-gans.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/rickiepark\/deep-learning-with-python-notebooks\/blob\/master\/8.5-introduction-to-gans.ipynb<\/a><\/p>\n\n<p>I know it is explained with KOREAN that you guys are not easy to understand but it's very simple source for keras text book.<\/p>\n\n<pre><code>Dense(128 * 16 * 16)\nLeakyReLU\nReshape((16, 16, 128))\nConv2D(256, 5, padding='same')\nLeakyReLU\nConv2DTranspose(256, 4, strides=2, padding='same')\n# BatchNormalization(momentum=0.8)\nLeakyReLU\nConv2D(256, 5, padding='same')\n# BatchNormalization(momentum=0.8)\nLeakyReLU\nConv2D(256, 5, padding='same')\nlayers.LeakyReLU\nlayers.Conv2D(3, 7, activation='tanh', padding='same')\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58114048,"AnswerCount":0,"Tags":"<javascript><node.js><tensorflow><tensorflow.js>","CreationDate":"2019-09-26T09:47:00.653","AcceptedAnswerId":null,"OwnerUserId":9018649.0,"Title":"How to import tensorflow in javascript? Import in file, served by local http-server","Body":"<p>I'm following this tutorial:\n<a href=\"https:\/\/codelabs.developers.google.com\/codelabs\/tfjs-training-classfication\/index.html#2\" rel=\"nofollow noreferrer\">https:\/\/codelabs.developers.google.com\/codelabs\/tfjs-training-classfication\/index.html#2<\/a><\/p>\n\n<p>I have set up a local HTTP-server, and that work and the app is running.\nHowever, when I try to execute step 3 (Load the data), I get the following error when loading the data:<\/p>\n\n<pre><code>Uncaught TypeError: Failed to resolve module specifier \"@tensorflow\/tfjs\". Relative references must start with either \"\/\", \".\/\", or \"..\/\".\n<\/code><\/pre>\n\n<p>If i comment out the import statment:<\/p>\n\n<pre><code>import * as tf from '@tensorflow\/tfjs';\n<\/code><\/pre>\n\n<p>the page actually loads and shows the sidebar with 16 handwritten digits.<\/p>\n\n<ol>\n<li>Does this mean that TensorFlow is not loaded?<\/li>\n<li><p>Is TensorFlow loaded, and I do not need this import statement?<\/p><\/li>\n<li><p>Or maybe most important, why does the import not work??<\/p><\/li>\n<\/ol>\n","answers":[]}
+{"QuestionId":58114217,"AnswerCount":1,"Tags":"<machine-learning><keras><neural-network>","CreationDate":"2019-09-26T09:56:15.517","AcceptedAnswerId":null,"OwnerUserId":1113684.0,"Title":"How to check the predicted output during fitting of the model in Keras?","Body":"<p>I am new in Keras and I learned fitting and evaluating the model. \nAfter evaluating  the model one can see the actual predictions made by model.<\/p>\n\n<p>I am wondering Is it also possible to see the predictions during fitting in Keras? Till now I cant find any code doing this.<\/p>\n","answers":[{"AnswerId":"58117092","CreationDate":"2019-09-26T12:43:33.867","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Since this question doesn't specify \"epochs\", and since using callbacks may represent extra computation, I don't think it's exactly a duplication.    <\/p>\n\n<p>With tensorflow, you can use a custom training loop with eager execution turned on. A simple tutorial for creating a custom training loop: <a href=\"https:\/\/www.tensorflow.org\/tutorials\/eager\/custom_training_walkthrough\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/eager\/custom_training_walkthrough<\/a><\/p>\n\n<p>Basically you will:<\/p>\n\n<pre><code>#transform your data in to a Dataset:\ndataset = tf.data.Dataset.from_tensor_slices(\n    (x_train, y_train)).shuffle(some_buffer).batch(batchSize) \n     #the above is buggy in some versions regarding shuffling, you may need to shuffle\n     #again between each epoch    \n\n\n#create an optimizer    \noptimizer = tf.keras.optimizers.Adam()\n\n#create an epoch loop:\nfor e in range(epochs):\n\n    #create a batch loop\n    for i, (x, y_true) in enumerate(dataset):\n\n        #create a tape to record actions\n        with  tf.GradientTape() as tape:\n\n            #take the model's predictions\n            y_pred = model(x)\n\n            #calculate loss\n            loss = tf.keras.losses.binary_crossentropy(y_true, y_pred)\n\n        #calculate gradients\n        gradients = tape.gradient(loss, model.trainable_weights)\n\n        #apply gradients\n        optimizer.apply_gradients(zip(gradients, model.trainable_weights)\n<\/code><\/pre>\n\n<p>You can use the <code>y_pred<\/code> var for doing anything, including getting its <code>numpy_pred = y_pred.numpy()<\/code> value. <\/p>\n\n<p>The tutorial gives some more details about metrics and validation loop. <\/p>\n"}]}
+{"QuestionId":58114381,"AnswerCount":1,"Tags":"<python><keras><conv-neural-network>","CreationDate":"2019-09-26T10:05:32.213","AcceptedAnswerId":null,"OwnerUserId":11701704.0,"Title":"How to train model to add new classes?","Body":"<p>My trained model has 10 classes ( ie. output layer has 10 classes). I want to add 3 more classes to it without training the whole model again.<br>\nI want to use the old trained model and add new classes to it. <\/p>\n\n<p>This is the code I had already tried but it shows an error. <\/p>\n\n<pre><code>from keras.models import load_model\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\nbase_model = load_model('hand_gest.h5')\n\nnew_model = Sequential()\n\nfor layer in base_model.layers[:-2]:\n    new_model.add(layer)\n\nfor layer in new_model.layers:\n    layer.trainable = False\n\nweights_training = base_model.layers[-2].get_weights()\nnew_model.layers[-2].set_weights(weights_training) \n\n\nnew_model.add(Dense(units = 3, activation = 'softmax'))    \n<\/code><\/pre>\n\n<p>But when I train this model it shows the following error. <\/p>\n\n<pre><code>ValueError: You called `set_weights(weights)` on layer \"max_pooling2d_2\" with a  weight list of length 2, but the layer was expecting 0 weights. Provided weights: [array([[-0.01650696,  0.01082378,  0.0149541 , .....\n<\/code><\/pre>\n","answers":[{"AnswerId":"58117470","CreationDate":"2019-09-26T13:05:39.580","ParentId":null,"OwnerUserId":"5597718","Title":null,"Body":"<p>As the number of classes is changed from 10 to 13, the last layer of the previous network needs to be changed. <\/p>\n\n<pre><code>base_model = load_model('hand_gest.h5')\nbase_model.pop() #remove the last layer - 'Dense' layer with 10 units\nfor layer in base_model.layers:\n    layer.trainable = False\nbase_model.add(Dense(units = 13, activation = 'softmax'))\nbase_model.summary() #Check architecture before starting the fine-tuning\n<\/code><\/pre>\n"}]}
+{"QuestionId":58114417,"AnswerCount":1,"Tags":"<tensorflow><keras><neural-network><deep-learning><nlp>","CreationDate":"2019-09-26T10:07:25.567","AcceptedAnswerId":null,"OwnerUserId":9080239.0,"Title":"My model's loss value decreases slowly .how to reduce my loss faster while training?","Body":"<p>when I train the model the loss decrease from 0.9 to 0.5 in 2500 epochs. Is it normal? <\/p>\n\n<p>my model:<\/p>\n\n<pre><code>    model = Sequential()\n\n    model.add(Embedding(vocab_size , emd_dim, weights=[emd_matrix], input_length=maxLen,trainable=False))\n\n    model.add(LSTM(256,return_sequences=True,activation=\"relu\",kernel_regularizer=regularizers.l2(0.01),kernel_initializer=keras.initializers.glorot_normal(seed=None)))\n    model.add(LSTM(256,return_sequences=True,activation=\"relu\",kernel_regularizer=regularizers.l2(0.01),kernel_initializer=keras.initializers.glorot_normal(seed=None)))\n\n    model.add(LSTM(256,return_sequences=False,activation=\"relu\",kernel_regularizer=regularizers.l2(0.01),kernel_initializer=keras.initializers.glorot_normal(seed=None)))\n    model.add(Dense(l_h2i,activation='softmax'))\n    model.compile(loss='categorical_crossentropy', optimizer=\"adam\", metrics=['accuracy'])\n    filepath = \"F:\/checkpoints\/\"+modelname+\"\/lstm-{epoch:02d}-{loss:0.3f}-{acc:0.3f}-{val_loss:0.3f}-{val_acc:0.3f}.hdf5\"\n    checkpoint = ModelCheckpoint(filepath, monitor=\"loss\", verbose=1, save_best_only=True, mode='min')\n    reduce_lr = ReduceLROnPlateau(monitor='loss', factor=0.5, patience=2, min_lr=0.000001)\n    print(model.summary())\n    history=model.fit(X_train_indices, Y_train_oh, batch_size=batch_size ,\n                      epochs=epochs , validation_split=0.1, shuffle=True,\n                      callbacks=[checkpoint, reduce_lr])\n<\/code><\/pre>\n\n<p>some part of the results are as shown here : <\/p>\n\n<pre><code>loss improved from 0.54275 to 0.54272\nloss: 0.5427 - acc: 0.8524 - val_loss: 1.1198 - val_acc: 0.7610\n\nloss improved from 0.54272 to 0.54268\nloss: 0.5427 - acc: 0.8525 - val_loss: 1.1195 - val_acc: 0.7311\n\nloss improved from 0.54268 to 0.54251\nloss: 0.5425 - acc: 0.8519 - val_loss: 1.1218 - val_acc: 0.7420\n\nloss improved from 0.54251 to 0.54249\nloss: 0.5425 - acc: 0.8517 - val_loss: 1.1210 - val_acc: 0.7518\n<\/code><\/pre>\n","answers":[{"AnswerId":"58116910","CreationDate":"2019-09-26T12:33:11.807","ParentId":null,"OwnerUserId":"9241362","Title":null,"Body":"<p>consider updating ReduceLROnPlateau parameters as in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/callbacks\/ReduceLROnPlateau\" rel=\"nofollow noreferrer\">TensorFlow documentation<\/a>. Factor should be a larger and Patience should be smaller numbers<\/p>\n\n<pre><code>reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.2,\n                              patience=5, min_lr=0.001)\nmodel.fit(X_train, Y_train, callbacks=[reduce_lr])\n<\/code><\/pre>\n\n<p><strong>Arguments:<\/strong><\/p>\n\n<blockquote>\n  <ul>\n  <li>monitor: quantity to be monitored. <\/li>\n  <li><strong>factor<\/strong>: factor by which the learning rate will be reduced. new_lr = lr * factor <\/li>\n  <li><strong>patience<\/strong>: number of epochs with no improvement after which learning rate will be    reduced. <\/li>\n  <li>verbose: int. 0: quiet, 1: update messages. <\/li>\n  <li>mode: one of {auto, min, max}. In min mode, lr will be reduced when the quantity    monitored has stopped decreasing; in max mode it will\n  be reduced when the quantity     monitored has stopped increasing; in\n  auto mode, the direction is automatically     inferred from the name\n  of the monitored quantity. <\/li>\n  <li>min_delta: threshold for measuring the new optimum, to only    focus on significant changes. <\/li>\n  <li>cooldown: number of epochs to wait before resuming normal operation after lr has been     reduced. <\/li>\n  <li>min_lr: lower bound on the learning rate.<\/li>\n  <\/ul>\n<\/blockquote>\n"}]}
+{"QuestionId":58114504,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras>","CreationDate":"2019-09-26T10:11:25.967","AcceptedAnswerId":58116874.0,"OwnerUserId":8857724.0,"Title":"Keras Nested Models save and load weights separately or view Summary of all nested models","Body":"<p>I am trying to train a Keras Model which include two nested models, and I want to save the weights of both inner models separately. Right now I am able to save weights of the whole model, but I am unable to load the weights of nested models within big model.<\/p>\n\n<p>Output of <code>Big_model.summary<\/code> looks like this<\/p>\n\n<pre><code>Model: \"model_3\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_4 (InputLayer)         [(None, 128, 128, 1)]     0         \n_________________________________________________________________\nmodel (Model)                (None, 16, 16, 512)       170369024 \n_________________________________________________________________\nmodel_1 (Model)              (None, 128, 128, 1)       15342209  \n=================================================================\nTotal params: 185,711,233\nTrainable params: 185,711,233\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>How can I even see the summary of both inner models, e.g <code>Big_model.inner_Model1.summary()<\/code> something like that, Or save the weights of both inner models separately after training using <code>Big_model.inner_Model1.save_weights()<\/code> and <code>Big_model.inner_Model2.save_weights()<\/code>or <code>callbacks<\/code> during <code>model.fit<\/code>.<\/p>\n\n<p>What I am getting is <code>Big_model<\/code> has no module as <code>inner_Model1<\/code>, Any help Please ??<\/p>\n\n<p>PS: There is no problem with training or anything, I can run the training, also I am using Tensorflow version <code>tf.keras.models.Model<\/code> for models. <\/p>\n\n<p>This is how I am creating models<\/p>\n\n<pre><code>inner_Model1 = tf.keras.models.Model()\ninner_Model2 = tf.keras.models.Model()\n\nx = tf.keras.layers.Input(shape=IMAGE_SHAPE)\nBig_model = tf.keras.models.Model(x, inner_model2(inner_model1(x)))\nBig_model.compile(optimizer=optimizer, loss='mean_absolute_error')\n<\/code><\/pre>\n","answers":[{"AnswerId":"58116874","CreationDate":"2019-09-26T12:31:35.700","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>In that summary you posted, <code>model<\/code> is layer 1, and <code>model_1<\/code> is layer 2:<\/p>\n\n<pre><code>Big_model.layers[1].summary()   #this is inner_Model1.summary()\nBig_model.layers[2].summary()   #this is inner_Model2.summary()\n<\/code><\/pre>\n\n<p>Do whatever you want with them. <\/p>\n\n<hr>\n\n<p>If you created the model like you did, there is nothing wrong with simply doing:   <\/p>\n\n<pre><code>inner_Model1.save_weights(...)\ninner_Model2.save_weights(...)\n<\/code><\/pre>\n\n<p>It will also work OK if you load the weights outside the big model, it will see the changes.<\/p>\n\n<pre><code>inner_Model1.load_weights(...)\ninner_Model2.load_weights(...)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58115212,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><tensorboard>","CreationDate":"2019-09-26T10:52:23.743","AcceptedAnswerId":null,"OwnerUserId":8270077.0,"Title":"Using Tensorboard to monitor training real time and visualize the model architecture","Body":"<p>I am learning to use Tensorboard -- Tensorflow 2.0.<\/p>\n\n<p>In particular, I would like to monitor the learning curves realtime and also to visually inspect and communicate the architecture of my model.<\/p>\n\n<p>Below I will provide code for a reproducible example.<\/p>\n\n<p>I have three problems:<\/p>\n\n<ol>\n<li><p>Although I get the learning curves once the training is over I don't know what I should do to monitor them in real time<\/p><\/li>\n<li><p>The learning curve I get from Tensorboard does not agree with the plot of history.history.  In fact is bizarre and difficult to interpret its reversals.<\/p><\/li>\n<li><p>I can not make sense of the graph.  I have trained a sequential model with 5 dense layers and dropout layers in between.  What Tensorboard shows me is something which much more elements in it.<\/p><\/li>\n<\/ol>\n\n<p>My code is the following:<\/p>\n\n<pre><code>from keras.datasets import boston_housing\n\n(train_data, train_targets), (test_data, test_targets) = boston_housing.load_data()\n\ninputs = Input(shape = (train_data.shape[1], ))\nx1 = Dense(100, kernel_initializer = 'he_normal', activation = 'elu')(inputs)\nx1a = Dropout(0.5)(x1)\nx2 = Dense(100, kernel_initializer = 'he_normal', activation = 'elu')(x1a)\nx2a = Dropout(0.5)(x2)\nx3 = Dense(100, kernel_initializer = 'he_normal', activation = 'elu')(x2a)\nx3a = Dropout(0.5)(x3)\nx4 = Dense(100, kernel_initializer = 'he_normal', activation = 'elu')(x3a)\nx4a = Dropout(0.5)(x4)\nx5 = Dense(100, kernel_initializer = 'he_normal', activation = 'elu')(x4a)\npredictions = Dense(1)(x5)\nmodel = Model(inputs = inputs, outputs = predictions)\n\nmodel.compile(optimizer = 'Adam', loss = 'mse')\n\nlogdir=\"logs\\\\fit\\\\\" + datetime.now().strftime(\"%Y%m%d-%H%M%S\")\ntensorboard_callback = keras.callbacks.TensorBoard(log_dir=logdir)\n\nhistory = model.fit(train_data, train_targets,\n          batch_size= 32,\n          epochs= 20,\n          validation_data=(test_data, test_targets),\n          shuffle=True,\n          callbacks=[tensorboard_callback ])\n\nplt.plot(history.history['loss'])\nplt.plot(history.history['val_loss'])\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/SxyYc.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/SxyYc.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/QsiH0.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/QsiH0.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><code>plt.plot(history.history['val_loss'])\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/CQH64.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/CQH64.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/BeLkz.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/BeLkz.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/oq00z.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/oq00z.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":58115407,"AnswerCount":1,"Tags":"<python><tensorflow><gpu>","CreationDate":"2019-09-26T11:03:01.833","AcceptedAnswerId":null,"OwnerUserId":12124369.0,"Title":"How can I solve it?","Body":"<p>I don't figure out what problem is here happening. After installing TensorFlow-GPU 2.0, I just check that TensorFlow work properly. And get this sort of problem.<\/p>\n\n<pre><code>import tensorflow as tf\nif tf.test.is_gpu_available():\n    with tf.device(\"\/gpu:0\"):\n        tf_string_array = tf.constant([\"TensorFlow\", \"Deep Learning\", \"AI\"])\n    tf_string_array.device\n<\/code><\/pre>\n\n<p>After execution my code. I get this error:<\/p>\n\n<p>RuntimeError: Can't copy Tensor with type string to device \/job:localhost\/replica:0\/task:0\/device:GPU:0.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/yFsRK.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p>And I was installing tensorflow to follow this blog: \n<a href=\"https:\/\/towardsdatascience.com\/installing-tensorflow-with-cuda-cudnn-and-gpu-support-on-windows-10-60693e46e781\" rel=\"nofollow noreferrer\">enter link description here<\/a> <\/p>\n","answers":[{"AnswerId":"58538037","CreationDate":"2019-10-24T09:13:30.547","ParentId":null,"OwnerUserId":"12198627","Title":null,"Body":"<p>Verify proper installation of tensorflow-GPU by using following steps,<\/p>\n\n<ul>\n<li><p>Can you check if tensorflow is using GPU,<\/p>\n\n<p><code>import tensorflow as tf\nprint(tf.test.is_gpu_available())<\/code><\/p><\/li>\n<li><p>To find out which device is used, you can enable log device placement\nlike this:<\/p>\n\n<p><code>sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))<\/code><\/p><\/li>\n<li><p>you can try to assign a computation to the gpu and see whether you have an error<\/p>\n\n<p><code>import tensorflow as tf\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')\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)\nwith tf.Session() as sess:\n    print (sess.run(c))<\/code><\/p><\/li>\n<\/ul>\n"}]}
+{"QuestionId":58115412,"AnswerCount":0,"Tags":"<python><tensorflow><plot><keras>","CreationDate":"2019-09-26T11:03:20.023","AcceptedAnswerId":null,"OwnerUserId":8270077.0,"Title":"\"ImportError: Failed to import pydot. You must install pydot and graphviz for `pydotprint` to work.\" However both are installed","Body":"<p>I have installed pydotplus and graphviz (I run Windows 10 and Python 3.6).<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/uDH8z.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/uDH8z.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>However, when I try to plot my Keras model I am getting an exception and I am asked to install pytdot and graphviz.<\/p>\n\n<pre><code>keras.utils.plot_model(model, 'my_first_model.png')\n\n---------------------------------------------------------------------------\nFileNotFoundError                         Traceback (most recent call last)\nC:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Anaconda3_64\\lib\\site-packages\\pydot.py in create(self, prog, format, encoding)\n   1914                 arguments=arguments,\n-&gt; 1915                 working_dir=tmp_dir,\n   1916             )\n\nC:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Anaconda3_64\\lib\\site-packages\\pydot.py in call_graphviz(program, arguments, working_dir, **kwargs)\n    135         stdout=subprocess.PIPE,\n--&gt; 136         **kwargs\n    137     )\n\nC:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Anaconda3_64\\lib\\subprocess.py in __init__(self, args, bufsize, executable, stdin, stdout, stderr, preexec_fn, close_fds, shell, cwd, env, universal_newlines, startupinfo, creationflags, restore_signals, start_new_session, pass_fds, encoding, errors)\n    708                                 errread, errwrite,\n--&gt; 709                                 restore_signals, start_new_session)\n    710         except:\n\nC:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Anaconda3_64\\lib\\subprocess.py in _execute_child(self, args, executable, preexec_fn, close_fds, pass_fds, cwd, env, startupinfo, creationflags, shell, p2cread, p2cwrite, c2pread, c2pwrite, errread, errwrite, unused_restore_signals, unused_start_new_session)\n    996                                          os.fspath(cwd) if cwd is not None else None,\n--&gt; 997                                          startupinfo)\n    998             finally:\n\nFileNotFoundError: [WinError 2] The system cannot find the file specified\n\nDuring handling of the above exception, another exception occurred:\n\nFileNotFoundError                         Traceback (most recent call last)\nC:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Anaconda3_64\\lib\\site-packages\\tensorflow\\python\\keras\\utils\\vis_utils.py in _check_pydot()\n     44     # to check the pydot\/graphviz installation.\n---&gt; 45     pydot.Dot.create(pydot.Dot())\n     46   except Exception:\n\nC:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Anaconda3_64\\lib\\site-packages\\pydot.py in create(self, prog, format, encoding)\n   1921                     prog=prog)\n-&gt; 1922                 raise OSError(*args)\n   1923             else:\n\nFileNotFoundError: [WinError 2] \"dot\" not found in path.\n\nDuring handling of the above exception, another exception occurred:\n\nImportError                               Traceback (most recent call last)\n&lt;ipython-input-58-f0e169c88003&gt; in &lt;module&gt;()\n----&gt; 1 keras.utils.plot_model(model, 'my_first_model.png')\n\nC:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Anaconda3_64\\lib\\site-packages\\tensorflow\\python\\keras\\utils\\vis_utils.py in plot_model(model, to_file, show_shapes, show_layer_names, rankdir)\n    151       This enables in-line display of the model plots in notebooks.\n    152   \"\"\"\n--&gt; 153   dot = model_to_dot(model, show_shapes, show_layer_names, rankdir)\n    154   _, extension = os.path.splitext(to_file)\n    155   if not extension:\n\nC:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Anaconda3_64\\lib\\site-packages\\tensorflow\\python\\keras\\utils\\vis_utils.py in model_to_dot(model, show_shapes, show_layer_names, rankdir)\n     70   from tensorflow.python.util import nest\n     71 \n---&gt; 72   _check_pydot()\n     73   dot = pydot.Dot()\n     74   dot.set('rankdir', rankdir)\n\nC:\\Program Files (x86)\\Microsoft Visual Studio\\Shared\\Anaconda3_64\\lib\\site-packages\\tensorflow\\python\\keras\\utils\\vis_utils.py in _check_pydot()\n     47     # pydot raises a generic Exception here,\n     48     # so no specific class can be caught.\n---&gt; 49     raise ImportError('Failed to import pydot. You must install pydot'\n     50                       ' and graphviz for `pydotprint` to work.')\n     51 \n\nImportError: Failed to import pydot. You must install pydot and graphviz for `pydotprint` to work.\n<\/code><\/pre>\n\n<p>It is not clear why I am getting this error message. Since I have already installed the packages, what else is left to do?<\/p>\n\n<hr>\n\n<p><strong>graphviz has been downloaded<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/kZAjS.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/kZAjS.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/623Jf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/623Jf.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":58115439,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-26T11:05:14.747","AcceptedAnswerId":null,"OwnerUserId":8211216.0,"Title":"How to reuse the BERT model using tensorflow.contrib","Body":"<p>I have tried below codes for reusing the saved BERT model.<\/p>\n\n<pre><code>def serving_input_receiver_fn():\n  feature_spec = {\n    \"input_ids\" : tf.FixedLenFeature([MAX_SEQ_LENGTH], tf.int64),\n    \"input_mask\" : tf.FixedLenFeature([MAX_SEQ_LENGTH], tf.int64),\n    \"segment_ids\" : tf.FixedLenFeature([MAX_SEQ_LENGTH], tf.int64),\n    \"label_ids\" :  tf.FixedLenFeature([], tf.int64)\n  }\n  serialized_tf_example = tf.placeholder(dtype=tf.string,                                         \n                                     shape=[None],\n                                     name='input_example_tensor')\n  print(serialized_tf_example, \"serialized_tf_example\")\n  print(serialized_tf_example.shape, \"Shape\")\n  receiver_tensors = {'example': serialized_tf_example}\n  print(receiver_tensors, \"receiver_tensors\")\n  features = tf.parse_example(serialized_tf_example, feature_spec)\n  return tf.estimator.export.ServingInputReceiver(features, receiver_tensors)\n\n  export_path = '.\/BERTmodel\/Data\/'\n<\/code><\/pre>\n\n<p><strong>But I am receiving the error below: ' Cannot feed value of shape () for Tensor 'input_example_tensor:0', which has shape '(?,)'<\/strong><\/p>\n\n<p>I tried below codes for prediction.<\/p>\n\n<p>Can someone advise me on this?<\/p>\n\n<pre><code>pred_sentences = [\"The site is great\", \"I think it's not good\"]\ndef getPrediction(in_sentences):\n labels = [\"Negative\", \"Positive\", \"Neutral\"]\n input_examples = [run_classifier.InputExample(guid=\"\", text_a = x, text_b = None, label = 0)    for x in in_sentences]\n input_features = run_classifier.convert_examples_to_features(input_examples, label_list, MAX_SEQ_LENGTH, tokenizer)\n predict_input_fn = run_classifier.input_fn_builder(features=input_features,   seq_length=MAX_SEQ_LENGTH, is_training=False, drop_remainder=False)\n return predict_input_fn\n\n from tensorflow.contrib import predictor \n with tf.Session() as sess:\n  predict_fn = predictor.from_saved_model('model_path')\n  predictions = predict_fn({\"example\": getPrediction(pred_sentences)})\n  print(predictions)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58115477,"AnswerCount":1,"Tags":"<python><tensorflow><caffe><yolo>","CreationDate":"2019-09-26T11:08:09.367","AcceptedAnswerId":null,"OwnerUserId":1632479.0,"Title":"Pretrained CNN(tensorflow\/darknet\/caffe) weights for human\/vehicle detection only","Body":"<p>I am making use of tensorflow's pretrained weights from <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/object_detection\/g3doc\/detection_model_zoo.md#coco-trained-models-coco-models\" rel=\"nofollow noreferrer\">Tensorflow detection model zoo<\/a> , which is primarily trained on <a href=\"http:\/\/cocodataset.org\/\" rel=\"nofollow noreferrer\">COCO<\/a> datasets , which covers about 80 different classes including humans, so making use of these models certainly results in higher computational trade-off, so is there any publicly open pre-trained weights which only focuses on one class in this case either human or vehicle(car).<\/p>\n\n<p>If there is no such models available then how it is possible to fine-tune or customize these existing models like,<\/p>\n\n<p>\"<strong>ssd_inception_v2_coco_2018_01_28<\/strong>\" , which performs pretty well with <em>mAP of  32 and computationally efficient as well<\/em> , so how such models can be utilized to detect only humans but not any other objects.<\/p>\n","answers":[{"AnswerId":"58115847","CreationDate":"2019-09-26T11:29:09.993","ParentId":null,"OwnerUserId":"1969069","Title":null,"Body":"<p>Have you tried the model <a href=\"https:\/\/github.com\/opencv\/open_model_zoo\/blob\/develop\/models\/intel\/person-vehicle-bike-detection-crossroad-1016\/description\/person-vehicle-bike-detection-crossroad-1016.md\" rel=\"nofollow noreferrer\">person-vehicle-bike-detection-crossroad-1016<\/a> ?<\/p>\n"}]}
+{"QuestionId":58115660,"AnswerCount":0,"Tags":"<python><numpy><torch>","CreationDate":"2019-09-26T11:17:40.400","AcceptedAnswerId":null,"OwnerUserId":4629668.0,"Title":"Transform same values with softmax","Body":"<p>Hy guys, I have a big problem.\nI have an N-dimensional array of numpy with 0 and 1.\nThe numbers that is one are contigous. For the region of the ones I have to calcolate the softmax function(present only in torch of pytorch)and then the sum of all NDArray is 1.\nCan you help me?<\/p>\n","answers":[]}
+{"QuestionId":58115801,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-09-26T11:26:24.310","AcceptedAnswerId":58855239.0,"OwnerUserId":3513738.0,"Title":"How to use the lbfgs optimizer with pytorch-lightning?","Body":"<p>I have a problem in using the LBFGS optimizer from pytorch with lightning.\nI use the template from <a href=\"https:\/\/github.com\/williamFalcon\/pytorch-lightning-conference-seed\" rel=\"nofollow noreferrer\">here<\/a> to start a new project and here is the code that I tried (only the training portion):<\/p>\n\n<pre><code>def training_step(self, batch, batch_nb):\n    x, y = batch\n    x = x.float()\n    y = y.float()\n    y_hat = self.forward(x)\n    return {'loss': F.mse_loss(y_hat, y)}\n\ndef configure_optimizers(self):\n    optimizer = torch.optim.LBFGS(self.parameters())\n    return optimizer\n\ndef optimizer_step(self, epoch_nb, batch_nb, optimizer, optimizer_i):\n    def closure():\n        optimizer.zero_grad()\n        l = self.training_step(batch, batch_nb)\n        loss = l['loss']\n        loss.backward()\n        return loss\n\n    optimizer.step(closure)\n<\/code><\/pre>\n\n<p>The LBFGS optimizer from pytorch requires a closure function (see <a href=\"https:\/\/pytorch.org\/docs\/stable\/_modules\/torch\/optim\/lbfgs.html\" rel=\"nofollow noreferrer\">here<\/a> and <a href=\"https:\/\/pytorch.org\/docs\/stable\/optim.html#torch.optim.LBFGS.step\" rel=\"nofollow noreferrer\">here<\/a>), but I don't know how to define it inside the template, specially I don't know how the batch data is passed to the optimizer. I tried to define a custom optimizer_step function but I have some problems to passing the batch inside the closure function.<\/p>\n\n<p>I will be very thankful of any advise that helps me to solve this problem or points me to the right direction.<\/p>\n\nEnvironment:\n\n<ul>\n<li>PyTorch version:  1.2.0+cpu<\/li>\n<li>Lightning version: 0.4.9   <\/li>\n<li>Test-tube version: 0.7.1<\/li>\n<\/ul>\n","answers":[{"AnswerId":"58855239","CreationDate":"2019-11-14T11:01:03.677","ParentId":null,"OwnerUserId":"3513738","Title":null,"Body":"<p>The support for lbfgs optimizer was added on <a href=\"https:\/\/github.com\/williamFalcon\/pytorch-lightning\/pull\/310\" rel=\"nofollow noreferrer\">#310<\/a>, now it is not required to define a closure function.<\/p>\n"}]}
+{"QuestionId":58116359,"AnswerCount":0,"Tags":"<python><tensorflow><machine-learning><keras><computer-vision>","CreationDate":"2019-09-26T11:58:10.050","AcceptedAnswerId":null,"OwnerUserId":12124818.0,"Title":"Is there a simple way to use only half of the images in underlying directories using Keras flow from directory?","Body":"<p>I am trying to work out how to use a subset of the images within directories in order to test if a larger dataset will actually help me classify some images (and for context, they are obscure images with seemingly random blobs (but they are not actually random) rather than shapes a human could distinguish by eye). I am using ImageDataGenerator with flow from directory in Keras, and so would like to test this on various different amounts of data.<\/p>\n\n<p>This is the simple code I have so far too pull the data from the directories using generators.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def load_the_data(batch_size):\n\n    train_datagen = ImageDataGenerator(\n        rescale=1.\/255,\n        horizontal_flip=True,\n        rotation_range=50,\n        validation_split=0.6,\n        )\n\n    train_generator = train_datagen.flow_from_directory(\n        train_data_dir,\n        target_size=(21, 21),\n        batch_size=batch_size,\n        class_mode='categorical',\n        subset='training'\n        )\n\n    validation_generator = train_datagen.flow_from_directory(\n        train_data_dir, # same directory as training data\n        target_size=(21, 21),\n        batch_size=batch_size,\n        class_mode='categorical',\n        subset='validation'\n        )\n\n    return train_datagen, train_generator, validation_generator\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58116824,"AnswerCount":1,"Tags":"<python><machine-learning><keras><neural-network><evaluation>","CreationDate":"2019-09-26T12:28:42.103","AcceptedAnswerId":58117124.0,"OwnerUserId":9080239.0,"Title":"How to get train and dev error in Keras neural net?","Body":"<p>How should I get these values for the purpose of understanding whether my model is overfitting or underfitting? \nFor instance, when <code>train_error= 0.1<\/code> and <code>dev_error=0.25<\/code>, it is understandable that the model is overfitted. But I don't know which of the provided results of my model indicates the mentioned errors.\nIs the loss value indicating this measure or the (1-accuracy) one?<\/p>\n\n<p>This is a part of my model's training result: <\/p>\n\n<pre><code>Epoch 152\/250\n100253\/100253 [==============================] - 12s 118us\/step - loss: \n0.5306 - acc: 0.8568 - val_loss: 1.1438 - val_acc: 0.7550\n<\/code><\/pre>\n","answers":[{"AnswerId":"58117124","CreationDate":"2019-09-26T12:45:07.747","ParentId":null,"OwnerUserId":"4685471","Title":null,"Body":"<ol>\n<li><p>the terms <em>error<\/em> and <em>loss<\/em> are synonyms<\/p><\/li>\n<li><p>Keras indicates the training loss (respectively accuracy) as <code>loss<\/code> (<code>acc<\/code>); validation loss and accuracy are indicated as <code>val_loss<\/code> and <code>val_acc<\/code> respectively.<\/p><\/li>\n<li><p><code>train_error= 0.1<\/code> and <code>dev_error=0.25<\/code>, or <code>loss: 0.5306<\/code> and  <code>val_loss: 1.1438<\/code> as here, does <strong>not<\/strong> indicate overfitting; this is just the <strong>generalization gap<\/strong>, i.e. the expected gap in the performance between the training and validation sets. Quoting from a recent <a href=\"https:\/\/ai.googleblog.com\/2019\/07\/predicting-generalization-gap-in-deep.html\" rel=\"nofollow noreferrer\">blog post by Google AI<\/a>:<\/p>\n\n<blockquote>\n  <p>An important concept for understanding generalization is the <em>generalization gap<\/em>, i.e., the difference between a model\u2019s performance on training data and its performance on unseen data drawn from the same distribution.<\/p>\n<\/blockquote><\/li>\n<\/ol>\n\n<p>The telltale <strong>signature<\/strong> of overfitting is when your validation loss starts increasing, while your training loss continues decreasing, i.e.:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/KLM4h.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/KLM4h.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Notice the <strong>gap<\/strong> (pun intended) between the curves in the above plot (adapted from the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Overfitting\" rel=\"nofollow noreferrer\">Wikipedia entry on overfitting<\/a>).<\/p>\n\n<p>Depending on the context, one may prefer to monitor the <em>metric<\/em> instead of the loss as correctly pointed out in the comment below, but this is another discussion - hopefully you get the idea here...<\/p>\n"}]}
+{"QuestionId":58117591,"AnswerCount":0,"Tags":"<python><tensorflow><object-detection>","CreationDate":"2019-09-26T13:11:29.173","AcceptedAnswerId":null,"OwnerUserId":7136516.0,"Title":"Object detection with background class","Body":"<p>I need to get the output of object detection with the background class \nI'm using frozen model (.pb) base on faster RCNN 101<\/p>\n\n<p>I know which layers you need to inference on .pb model file to get the output without the background\nex:<\/p>\n\n<pre><code>    scores = det_graph.get_tensor_by_name('detection_scores:0')\n    classes = det_graph.get_tensor_by_name('detection_classes:0')\n    num_detections = det_graph.get_tensor_by_name('num_detections:0')\n<\/code><\/pre>\n\n<p>is it possible to get the output with the background class?<\/p>\n","answers":[]}
+{"QuestionId":58117763,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><ner>","CreationDate":"2019-09-26T13:20:32.327","AcceptedAnswerId":null,"OwnerUserId":10183880.0,"Title":"How to train your model in deeppavlov (NER) Python 3","Body":"<p>First of all, sorry for any newbie mistakes that I've made. But I couldn't figure out and couldn't find a source specifically for <a href=\"http:\/\/docs.deeppavlov.ai\/en\/master\/features\/models\/ner.html\" rel=\"nofollow noreferrer\">deeppavlov (NER)<\/a> library. I'm trying to train ner_ontonotes_bert_mult as described <a href=\"http:\/\/docs.deeppavlov.ai\/en\/master\/features\/models\/ner.html#train-and-use-the-model\" rel=\"nofollow noreferrer\">here<\/a>. I guess it can be trained from its checkpoint to make it recognize some specific patterns like;<\/p>\n\n<pre><code>\"Round 23\/22; 24,9 x 12,2 x 12,3\"\n<\/code><\/pre>\n\n<p>as<\/p>\n\n<pre><code>[[['Round', '23\/22', ';', '24,9 x 12,2 x 12,3']], [['B-PRODUCT', 'I-PRODUCT', 'B-QUANTITY']]]\n<\/code><\/pre>\n\n<p>My questions are (before I dig into details):<\/p>\n\n<ol>\n<li><s> Is it possible?<\/s>  And I realized I can't use samples like \" Round 23\/22; 24,9 x 12,2 x 12,3 \". I need them to be in full sentences. <\/li>\n<li>Where can I find more info about it specifically related to deeppavlov's model(s)?<\/li>\n<li>How can I train pre-trained deeppavlov model to recognize my custom patterns?<\/li>\n<\/ol>\n\n<p>I don't even understand if it is possible but I've decided to give it go and prepared 3 <code>.txt<\/code> files as <code>\"train.txt\"<\/code>, <code>\"test.txt\"<\/code> and <code>\"validation.txt\"<\/code> as <a href=\"http:\/\/docs.deeppavlov.ai\/en\/master\/features\/models\/ner.html#training-data\" rel=\"nofollow noreferrer\">described in deeppovlov web page<\/a>. And I put them under the folder <code>'~\/.deeppavlov\/downloads\/ontonotes\/ner_ontonotes_bert_mult'<\/code>. My dataset looks like this:<\/p>\n\n<pre><code>Round B-PRODUCT\n23\/22 I-PRODUCT\n24,9 x 12,2 x 12,3 B-QUANTITY\nRing B-PRODUCT\nHDFAA I-PRODUCT\n12,7 x 10 B-QUANTITY\n<\/code><\/pre>\n\n<p>and so on... This is the code I am trying to train it:<\/p>\n\n<pre><code>import os\n# Force tensorflow to use CPU instead of GPU.\nos.environ['CUDA_VISIBLE_DEVICES'] = '-1'\nfrom deeppavlov import configs, train_model\nfrom deeppavlov.core.commands.utils import parse_config\n\nconfig_dict = parse_config(configs.ner.ner_ontonotes_bert_mult)\n\nprint(config_dict['dataset_reader']['data_path'])\n\nfrom deeppavlov import configs, train_model\n\nner_model = train_model(configs.ner.ner_ontonotes_bert_mult)\n<\/code><\/pre>\n\n<p>But I am getting this error:<\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: Assign requires shapes of both tensors to match. lhs shape= [3] rhs shape= [37]\n     [[{{node save\/Assign_280}}]]\n<\/code><\/pre>\n\n<p>Full traceback:<\/p>\n\n<pre><code>2019-09-26 15:50:27.63 ERROR in 'deeppavlov.core.common.params'['params'] at line 110: Exception in &lt;class 'deeppavlov.models.bert.bert_ner.BertNerModel'&gt;\nTraceback (most recent call last):\n  File \"\/home\/custom_user\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1356, in _do_call\n    return fn(*args)\n  File \"\/home\/custom_user\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1341, in _run_fn\n    options, feed_dict, fetch_list, target_list, run_metadata)\n  File \"\/home\/custom_user\/.local\/lib\/python3.6\/site-packages\/tensorflow\/python\/client\/session.py\", line 1429, in _call_tf_sessionrun\n    run_metadata)\ntensorflow.python.framework.errors_impl.InvalidArgumentError: Assign requires shapes of both tensors to match. lhs shape= [3] rhs shape= [37]\n     [[{{node save\/Assign_280}}]]\n<\/code><\/pre>\n\n<h1><strong>UPDATE 2:<\/strong><\/h1>\n\n<p>And I realized I can't use samples like \" Round 23\/22; 24,9 x 12,2 x 12,3 \". I need them to be in full sentences. <\/p>\n\n<h1><strong>UPDATE:<\/strong><\/h1>\n\n<p>It seems like this is happening due to my dataset. My custom dataset only has 3 tags (<code>B-PRODUCT<\/code>, <code>I-PRODUCT<\/code> and <code>B-QUANTITY<\/code>) but the pre-trained model has 37 of them. All available tags can be found <a href=\"http:\/\/docs.deeppavlov.ai\/en\/master\/features\/models\/ner.html#multilingual-bert-zero-shot-transfer\" rel=\"nofollow noreferrer\">here<\/a> under the sentence of <code>\"The list of available tags and their descriptions are presented below.\"<\/code>. 18 main tags(with <code>B<\/code> and <code>I<\/code> 36 tags), and <code>O<\/code> tag (\u201cO\u201d means the absence of entity.)). <strong>Total of all of the 37 tags needs to be present in the dataset.<\/strong> I was able to pass that error by adding dummy sentences by tagging them all with the missing tags. This is a terrible workaround since I'm willingly disrupting my own data-set. I'm still looking for a 'logical' way to train...<\/p>\n\n<p>PS: Now I am getting this error. <\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/home\/custom_user\/.PyCharm2019.2\/config\/scratches\/scratch_9.py\", line 13, in &lt;module&gt;\n    ner_model = train_model(configs.ner.ner_ontonotes_bert_mult)\n  File \"\/home\/custom_user\/.local\/lib\/python3.6\/site-packages\/deeppavlov\/__init__.py\", line 31, in train_model\n    train_evaluate_model_from_config(config, download=download, recursive=recursive)\n  File \"\/home\/custom_user\/.local\/lib\/python3.6\/site-packages\/deeppavlov\/core\/commands\/train.py\", line 121, in train_evaluate_model_from_config\n    trainer.train(iterator)\n  File \"\/home\/custom_user\/.local\/lib\/python3.6\/site-packages\/deeppavlov\/core\/trainers\/nn_trainer.py\", line 294, in train\n    self.train_on_batches(iterator)\n  File \"\/home\/custom_user\/.local\/lib\/python3.6\/site-packages\/deeppavlov\/core\/trainers\/nn_trainer.py\", line 234, in train_on_batches\n    self._validate(iterator)\n  File \"\/home\/custom_user\/.local\/lib\/python3.6\/site-packages\/deeppavlov\/core\/trainers\/nn_trainer.py\", line 150, in _validate\n    metrics = list(report['metrics'].items())\nAttributeError: 'NoneType' object has no attribute 'items'\n<\/code><\/pre>\n","answers":[{"AnswerId":"58857980","CreationDate":"2019-11-14T13:31:54.937","ParentId":null,"OwnerUserId":"12371705","Title":null,"Body":"<p>There are at least two problems here:<br>\n1. instead of <code>validation.txt<\/code> there should be a <code>valid.txt<\/code> file;<br>\n2. you are trying to retrain a model that was pretrained on a different dataset with a different set of tags, it's not necessary.  <\/p>\n\n<p>To train your model from scratch you can do something like:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import json\nfrom deeppavlov import configs, build_model, train_model\n\nwith configs.ner.ner_ontonotes_bert_mult.open(encoding='utf8') as f:\n    ner_config = json.load(f)\n\nner_config['dataset_reader']['data_path'] = '~\/my_data_dir\/'  # directory with train.txt, valid.txt and test.txt files\nner_config['metadata']['variables']['NER_PATH'] = '~\/where_to_save_the_model\/'\nner_config['metadata']['download'] = [ner_config['metadata']['download'][-1]]  # do not download the pretrained ontonotes model\n\nner_model = train_model(ner_config, download=True)\n<\/code><\/pre>\n\n<p><br\/><br\/>\nThe other thing that could go wrong is tokenization: <code>\"Round 23\/22; 24,9 x 12,2 x 12,3\"<\/code> will be split by the model to <code>['Round', '23', '\/', '22', ';', '24', ',', '9', 'x', '12', ',', '2', 'x', '12', ',', '3']<\/code> and not <code>['Round', '23\/22', ';', '24,9 x 12,2 x 12,3']<\/code>.  <\/p>\n\n<p>But you can tokenize your texts beforehand:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>ner_model([['Round', '23\/22', ';', '24,9 x 12,2 x 12,3']])\n<\/code><\/pre>\n"}]}
+{"QuestionId":58117928,"AnswerCount":2,"Tags":"<pytorch>","CreationDate":"2019-09-26T13:28:11.050","AcceptedAnswerId":null,"OwnerUserId":7295409.0,"Title":"RuntimeError: CUDA error: invalid argument","Body":"<p>It can run epoch 1 and eval successfully, but it fails when run epoch 2.<\/p>\n\n<pre><code>Train Epoch:1[655200\/655800(100%)] loss:26.4959 lr:0.2050\nTest Epoch:1 acc:0.973 val:0.895\n\nTrain Epoch:2[0\/655800(0%)] loss:26.8068 lr:0.2051\nFile \"train_11w.py\", line 244, in main\n    train(train_loader, model, optimizer, epoch, lr_decay_type, logger, args.log_interval, args)\n  File \"train_11w.py\", line 305, in train\n    prediction, ex, exnorm = model(img, mode=6, y=label)\n  File \"\/home\/luban\/anaconda2\/lib\/python2.7\/site-packages\/torch\/nn\/modules\/module.py\", line 489, in __call__\n    result = self.forward(*input, **kwargs)\n  File \"\/home\/luban\/anaconda2\/lib\/python2.7\/site-packages\/torch\/nn\/parallel\/data_parallel.py\", line 143, in forward\n    outputs = self.parallel_apply(replicas, inputs, kwargs)\n  File \"\/home\/luban\/anaconda2\/lib\/python2.7\/site-packages\/torch\/nn\/parallel\/data_parallel.py\", line 153, in parallel_apply\n    return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)])\n  File \"\/home\/luban\/anaconda2\/lib\/python2.7\/site-packages\/torch\/nn\/parallel\/parallel_apply.py\", line 83, in parallel_apply\n    raise output\nRuntimeError: CUDA error: invalid argument\n<\/code><\/pre>\n\n<pre><code>Driver Version: 418.67\nCUDA Version 10.0.130\npython 2.7.3\ntorch 1.0.0\n<\/code><\/pre>\n","answers":[{"AnswerId":"58126368","CreationDate":"2019-09-27T00:43:01.293","ParentId":null,"OwnerUserId":"7295409","Title":null,"Body":"<p>I used CUDA_LAUNCH_BLOCKING=1, also failed.\nTraceback (most recent call last):\n  File \"train_11w.py\", line 691, in \n    main(args)\n  File \"train_11w.py\", line 244, in main\n    train(train_loader, model, optimizer, epoch, lr_decay_type, logger, args.log_interval, args)\n  File \"train_11w.py\", line 307, in train\n    prediction, ex, exnorm = model(img, mode=6, y=label)\n  File \"\/home\/luban\/anaconda2\/lib\/python2.7\/site-packages\/torch\/nn\/modules\/module.py\", line 489, in <strong>call<\/strong>\n    result = self.forward(*input, **kwargs)\n  File \"\/home\/luban\/anaconda2\/lib\/python2.7\/site-packages\/torch\/nn\/parallel\/data_parallel.py\", line 142, in forward\n    replicas = self.replicate(self.module, self.device_ids[:len(inputs)])\n  File \"\/home\/luban\/anaconda2\/lib\/python2.7\/site-packages\/torch\/nn\/parallel\/data_parallel.py\", line 147, in replicate\n    return replicate(module, device_ids)\n  File \"\/home\/luban\/anaconda2\/lib\/python2.7\/site-packages\/torch\/nn\/parallel\/replicate.py\", line 13, in replicate\n    param_copies = Broadcast.apply(devices, *params)\n  File \"\/home\/luban\/anaconda2\/lib\/python2.7\/site-packages\/torch\/nn\/parallel\/_functions.py\", line 21, in forward\n    outputs = comm.broadcast_coalesced(inputs, ctx.target_gpus)\n  File \"\/home\/luban\/anaconda2\/lib\/python2.7\/site-packages\/torch\/cuda\/comm.py\", line 40, in broadcast_coalesced\n    return torch._C._broadcast_coalesced(tensors, devices, buffer_size)\nRuntimeError: NCCL Error 3: internal error<\/p>\n"},{"AnswerId":"58124368","CreationDate":"2019-09-26T20:32:01.820","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p>Although it is difficult to understand what's going wrong, I would suggest you do the following.<\/p>\n\n<ol>\n<li>Can you try to run your code with <code>CUDA_LAUNCH_BLOCKING=1 python script_name args<\/code>? The <code>CUDA_LAUNCH_BLOCKING=1<\/code> env variable makes sure to call all CUDA operations synchronously so that an error message should point to the right line of code in the stack trace.<\/li>\n<li>Try setting <code>torch.backends.cudnn.benchmark<\/code> to <code>True\/False<\/code> to check if it works.<\/li>\n<li>Train the model without using DataParallel.<\/li>\n<li>Check if you set <code>drop_last=True<\/code> while creating DataLoader, does training work?<\/li>\n<\/ol>\n"}]}
+{"QuestionId":58118334,"AnswerCount":1,"Tags":"<tensorflow><keras><tensorflow2.0><tf.keras>","CreationDate":"2019-09-26T13:50:36.600","AcceptedAnswerId":null,"OwnerUserId":4388164.0,"Title":"When should we inherits keras.Model instead of keras.layers.Layer even if we don't use model.fit?","Body":"<p>In some Tensorflow tutorials with tf2 (e.g. <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/text\/nmt_with_attention\" rel=\"nofollow noreferrer\">Neural Machine Translation with Attention<\/a> and <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/eager\" rel=\"nofollow noreferrer\">Eager essentials<\/a>), they define custom <code>tf.keras.Model<\/code>s instead of <code>tf.keras.layers.Layer<\/code>s (e.g. <code>BahdanauAttention(tf.keras.Model):<\/code>)<\/p>\n\n<p>Also, <a href=\"https:\/\/www.tensorflow.org\/tutorials\/eager\/custom_layers#models_composing_layers\" rel=\"nofollow noreferrer\">Models: composing layers<\/a> doc uses <code>tf.keras.Model<\/code> explicitly. The section says:<\/p>\n\n<blockquote>\n  <p>The main class used when creating a layer-like thing which contains other layers is tf.keras.Model. Implementing one is done by inheriting from tf.keras.Model.<\/p>\n<\/blockquote>\n\n<p>It sounds we need to inherit <code>tf.keras.Model<\/code> to define a layer which compose child layers.<\/p>\n\n<p>However, as far as I checked, this code works even if I define <code>ResnetIdentityBlock<\/code> as a child class of <code>tf.keras.layers.Layer<\/code>. Other two tutorials work with <code>Layer<\/code> too.\nIn addition to that, <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/keras\/custom_layers_and_models\" rel=\"nofollow noreferrer\">another tutorial<\/a> says<\/p>\n\n<blockquote>\n  <p>Model is just like a Layer, but with added training and serialization utilities.<\/p>\n<\/blockquote>\n\n<p>Thus, I have no idea what is the real difference between <code>tf.keras.Model<\/code> and <code>tf.keras.layers.Layer<\/code> and why those three tutorial with Eager execution uses <code>tf.keras.Model<\/code> though they don't use <em>training and serialization utilities<\/em> of <code>tf.keras.Model<\/code>.<\/p>\n\n<p>Why do we need to inherit <code>tf.keras.Model<\/code> in those tutorials?<\/p>\n\n<p><strong>Additional comment<\/strong><\/p>\n\n<p>utilities of <code>Model<\/code> work only with special subsets of <code>Layer<\/code> (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/c8ef33dd913463ced8cc347c03945a88b34da7f8\/tensorflow\/python\/keras\/engine\/training.py#L1461\" rel=\"nofollow noreferrer\">Layers whose <code>call<\/code> receive only one input<\/a>). Thus, I think the idea like <em>\"Always extend Model because Model has more features\"<\/em> is not correct. Also, it violates a basic programming program like <a href=\"https:\/\/en.wikipedia.org\/wiki\/Single_responsibility_principle\" rel=\"nofollow noreferrer\">SRP<\/a>.<\/p>\n","answers":[{"AnswerId":"58126729","CreationDate":"2019-09-27T01:50:03.063","ParentId":null,"OwnerUserId":"9828197","Title":null,"Body":"<p><strong><em>Update<\/em><\/strong><\/p>\n\n<p>So the comment was: <code>Yes, I know training and serialization utilities exist in Model as I wrote in the question. My question is why TF tutorials need to use Model though they don't use these methods.<\/code><\/p>\n\n<p>The best answer can be provided by the authors in this case because your question asks why they chose one method over the other where both of them have them can do the job equally well. Why can do the job equally well? Well, because <code>Model is just like a Layer, but with added training and serialization utilities.<\/code><\/p>\n\n<p>We can argue that using model when just layer can do the job is an overkill, but then it may be a matter of taste.<\/p>\n\n<p>Hope it helps<\/p>\n\n<p>PS.<\/p>\n\n<p>In the <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/eager\" rel=\"nofollow noreferrer\">eager example<\/a> and <a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/keras\/custom_layers_and_models\" rel=\"nofollow noreferrer\">custom layer writing<\/a> tutorials that you provided we cannot replace model with layer, so these tutorials do not apply to your question<\/p>\n\n<hr>\n\n<p>With model you can train but with layer only you cannot. See list of their methods below (excluding inner and inherited ones):<\/p>\n\n<p><code>tf.keras.layers.Layer<\/code><\/p>\n\n<pre><code>activity_regularizer\nactivity_regularizer\nadd_loss\nadd_metric\nadd_update\nadd_variable\nadd_weight\napply\nbuild\ncall\ncompute_mask\ncompute_output_shape\ncount_params\ndtype\ndynamic\nfrom_config\nget_config\nget_input_at\nget_input_mask_at\nget_input_shape_at\nget_losses_for\nget_output_at\nget_output_mask_at\nget_output_shape_at\nget_updates_for\nget_weights\ninbound_nodes\ninput\ninput_mask\ninput_shape\nlosses\nmetrics\nname\nnon_trainable_variables\nnon_trainable_weights\noutbound_nodes\noutput\noutput_mask\noutput_shape\nset_weights\ntrainable\ntrainable\ntrainable_variables\ntrainable_weights\nupdates\nvariables\nweights\n<\/code><\/pre>\n\n<p>see? no fit or evaluate method there.\n<code>tf.keras.Model<\/code><\/p>\n\n<pre><code>\ncompile\nevaluate\nevaluate_generator\nfit\nfit_generator\nget_weights\nload_weights\nmetrics\nmetrics_names\npredict\npredict_generator\npredict_on_batch\nreset_metrics\nrun_eagerly\nrun_eagerly\nsample_weights\ntest_on_batch\ntrain_on_batch\n<\/code><\/pre>\n"}]}
+{"QuestionId":58118438,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-26T13:56:01.777","AcceptedAnswerId":null,"OwnerUserId":886724.0,"Title":"Tensorflow : bucketize a tensor, then produce categorical tensor","Body":"<p>I want to transform a tensor of a continuous variable, into buckets, then into a categorical tensor (one-hot encoding based on boundaries)<\/p>\n\n<p>Say I have this<\/p>\n\n<pre><code>input = [1,2,3,4,5,6,7,8,9]\nboundaries = [3.5, 5.5]\n<\/code><\/pre>\n\n<p>I want to have the following output:<\/p>\n\n<pre><code>buckets = [0, 0, 0, 1, 1, 2, 2, 2, 2]\ncategorical = [[1, 0, 0],\n               [1, 0, 0],\n               [1, 0, 0],\n               [0, 1, 0],\n               [0, 1, 0],\n               [0, 0, 1],\n               [0, 0, 1],\n               [0, 0, 1],\n               [0, 0, 1]]\n<\/code><\/pre>\n\n<p>I am aware of <code>tf.feature_column.bucketized_column<\/code> and <code>tf.feature_column.categorical_column_with_identity<\/code>, but I want to use functions on pure tensors, without transforming data into <code>numeric_column<\/code>.<\/p>\n\n<p>I am also aware of <code>tf.one_hot<\/code> that allows to obtain <code>categorical<\/code> from <code>buckets<\/code>, but no idea on how to obtain <code>buckets<\/code> from the input tensor.<\/p>\n\n<p>Any help please?<\/p>\n","answers":[{"AnswerId":"58120570","CreationDate":"2019-09-26T15:53:40.233","ParentId":null,"OwnerUserId":"886724","Title":null,"Body":"<p>As said in the question, use <code>tf.one_hot<\/code> to get from buckets to categories, and use <code>tf.python.ops.math_ops._bucketize<\/code> to get the buckets<\/p>\n"}]}
+{"QuestionId":58118462,"AnswerCount":2,"Tags":"<tensorflow><keras><bayesian-networks><tensorflow-probability>","CreationDate":"2019-09-26T13:57:17.910","AcceptedAnswerId":null,"OwnerUserId":7542624.0,"Title":"Bayesian Model does not learn with tensorflow probability and keras","Body":"<p>I want to estimate epistemic uncertainty of my model. So I converted all layers into tensorflow probability layers. The model gives no errors back, but it also not learning anything. The model has two outputs and the losses of both outputs do not change at all. On the other hand, the overall loss of the model is shrinking, but seems not be related to the other losses at all, which I cant explain. <\/p>\n\n<pre><code>import numpy as np\nfrom tensorflow import keras\nimport tensorflow_probability as tfp\nimport tensorflow as tf\nfrom plot.plot_utils import plot_model_metrics\nfrom Custom_Keras_layers.ProbSqueezeExcite import squeeze_excite_block\n\ninp = keras.layers.Input(shape=[self.timesteps, self.features])\n    # left side\n    # 1 Conv1D block\n    l = tfp.layers.Convolution1DFlipout(filters=2*self.features, kernel_size=2, padding='same', activation=tf.nn.relu)(inp)\n    l = keras.layers.BatchNormalization()(l)\n    if squeeze_excite == 1:\n        l = squeeze_excite_block(l)\n    l = keras.layers.Dropout(dropout_rate)(l, training=True)\n\n    # 1 Conv1D block\n    l = tfp.layers.Convolution1DFlipout(filters=4 * self.features, kernel_size=4, padding='same', activation=tf.nn.relu)(l)\n    l = keras.layers.BatchNormalization()(l)\n    if squeeze_excite == 1:\n        l = squeeze_excite_block(l)\n    l = keras.layers.Dropout(dropout_rate)(l, training=True)\n\n    # 1 lstm bock\n    l = keras.layers.LSTM(32, recurrent_dropout=dropout_rate, dropout=dropout_rate)(l, training=True)\n\n    # letf output layer\n    l = tfp.layers.DenseFlipout(self.classes, activation=tf.nn.softmax, name='left')(l)\n\n    # right side\n    # 1 Conv1D block\n    r = tfp.layers.Convolution1DFlipout(filters=2 * self.features, kernel_size=2, padding='same', activation=tf.nn.relu)(inp)\n    r = keras.layers.BatchNormalization()(r)\n    if squeeze_excite == 1:\n        r = squeeze_excite_block(r)\n    r = keras.layers.Dropout(dropout_rate)(r, training=True)\n\n    # 1 Conv1D block\n    r = tfp.layers.Convolution1DFlipout(filters=4 * self.features, kernel_size=4, padding='same', activation=tf.nn.relu)(r)\n    r = keras.layers.BatchNormalization()(r)\n    if squeeze_excite == 1:\n        r = squeeze_excite_block(r)\n    r = keras.layers.Dropout(dropout_rate)(r, training=True)\n\n    # 1 lstm bock\n    r = keras.layers.LSTM(32, recurrent_dropout=dropout_rate, dropout=dropout_rate)(r, training=True)\n\n    # letf output layer\n    r = tfp.layers.DenseFlipout(self.classes, activation=tf.nn.softmax, name='right')(r)\n\n    model = keras.models.Model(inputs=inp, outputs=[l, r])\n\n    optimizer = tf.train.AdamOptimizer(learning_rate=lr)\n    losses = {\n        \"left\": self._neg_log_likelihood_bayesian,\n        \"right\": self._neg_log_likelihood_bayesian}\n    model.compile(optimizer=optimizer, loss=losses, metrics=['accuracy'])\n    self.model = model\n<\/code><\/pre>\n\n<p>and the loss function is defined as follows:<\/p>\n\n<pre><code>    def _neg_log_likelihood_bayesian(self, y_true, y_pred):\n    labels_distribution = tfp.distributions.Categorical(logits=y_pred)\n    neg_log_likelihood = -tf.reduce_mean(labels_distribution.log_prob(tf.argmax(y_true, axis=-1)))\n    kl = sum(self.model.losses) \/ self.trainNUM\n    loss = neg_log_likelihood + kl\n    return loss\n<\/code><\/pre>\n\n<p>Any help would be appreciated. The overall loss starts at 45000, whereas the two outputs losses are at around 1,3. It is very strange to me.<\/p>\n","answers":[{"AnswerId":"58135874","CreationDate":"2019-09-27T13:48:00.527","ParentId":null,"OwnerUserId":"7542624","Title":null,"Body":"<p>Thanks to this post on the issues forum of github tensorflow, I found out how to solve it <a href=\"https:\/\/github.com\/tensorflow\/probability\/issues\/282\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/probability\/issues\/282<\/a>\nYou have to scale the KL sum within each tfp layer: <\/p>\n\n<pre><code> kernel_divergence_fn=lambda q, p, _: tfp.distributions.kl_divergence(q, p) \/ tf.to_float(train.num_examples))\n<\/code><\/pre>\n\n<p>Furthermore I changed the loss function to :<\/p>\n\n<blockquote>\n  <p>neg_log_likelihood =\n  tf.nn.softmax_cross_entropy_with_logits_v2(labels=y_true,\n  logits=y_pred)<\/p>\n<\/blockquote>\n\n<p>That did it for me, now my model is properly training.<\/p>\n"},{"AnswerId":"58133109","CreationDate":"2019-09-27T10:54:56.187","ParentId":null,"OwnerUserId":"9946929","Title":null,"Body":"<p>The overall loss might already include the prior loss kl(sampled weights||prior), so could be doubke counted? (I'm not sure how Keras handles this.) Another thought is to try using reduce_sum instead of reduce_mean.<\/p>\n"}]}
+{"QuestionId":58119099,"AnswerCount":1,"Tags":"<python><machine-learning><keras><image-segmentation>","CreationDate":"2019-09-26T14:30:55.463","AcceptedAnswerId":null,"OwnerUserId":11486611.0,"Title":"Multi-class segmentation UNet gives unexpected output","Body":"<p>I'm trying to segment labeled objects (nuclei) using the UNet architecture in Keras and am using the Kaggle DSB18 dataset before I change to my own. To improve on the separation between two close objects, I generate a weighted map (jaidevd's implementation). I therefore end up with a input label looking like this:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/W3ym2.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/W3ym2.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Background is 0, 'real positives' are 1, spaces between objects are 2-11, depending on distance. I am trying to use a U-Net architecture as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Inputs\nX_train # list of all images, shaped (256, 256, 3)\nY_train # list of weighted mask, as shown above\nY_train = to_categorical(Y_train, num_classes=None)\n\n# Dice coefficient\ndef dice_coef(y_true, y_pred):\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 + 1.) \/ (K.sum(y_true_f) + K.sum(y_pred_f) + 1.)\n\n# Define model\ninputs = Input((256, 256, 3))\ns = Lambda(lambda x: x \/ 255) (inputs)\n# ... convolutional etc. layers\noutputs = Conv2D(12, (1, 1), activation='softmax') (c9)\nmodel = Model(inputs=[inputs], outputs=[outputs])\nmodel.compile(optimizer='adam', loss='categorical_crossentropy', metrics=[dice_coef])\n\n# Train model\nModelCheckpoint('model.hf', verbose=1, save_best_only=True)\nresults = model.fit(X_train, Y_train, validation_split=0.2, batch_size=10, epochs=15, callbacks=[checkpointer])\n\n# Test\ntest_image = # test image of shape (256,256,3)\ntest_image = np.expand_dims(test_image, axis=0)\nprediction = model.predict(test_image)\nprint(prediction.shape)\n&gt;&gt;&gt; (1, 256, 256, 12)\nimshow(np.sum(prediction[0,:,:,1:], axis=-1)) # or argmax or a single layer \n<\/code><\/pre>\n\n<p>All of my predicted outputs are either blank or have the borders highlighted as is shown here.\n<a href=\"https:\/\/i.stack.imgur.com\/yFoPz.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/yFoPz.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Does anyone see something wrong with my code or architecture or has another idea to separate close objects? As my first project I'm having a hard time figuring out what things to look out for.<\/p>\n\n<p>Gigantic thank you in advance. BBQuercus<\/p>\n","answers":[{"AnswerId":"58426828","CreationDate":"2019-10-17T07:11:31.203","ParentId":null,"OwnerUserId":"12231013","Title":null,"Body":"<p>im not sure, but \"outputs = Conv2D(12, (1, 1), activation='softmax') (c9)\" is not the correct way. Softmax is scaling\/sum your whole output to 1. \nConclusion: Class 1 = 0.1, Class 2 = 0.001, Class 3 = 0.3 --> Sum: 1 and your results make sence. Softmax get you the probabilities for each class. <\/p>\n\n<p>I think you can change it to \"sigmoid\" and hop it helps.<\/p>\n\n<p>Best regards<\/p>\n\n<p><em>Update: Now i run into the same issues.<\/em><\/p>\n"}]}
+{"QuestionId":58119155,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><tensorflow2.0>","CreationDate":"2019-09-26T14:33:08.980","AcceptedAnswerId":null,"OwnerUserId":2052105.0,"Title":"Freezing graph to pb in Tensorflow2","Body":"<p>We deploy lot of our models from TF1 by saving them through graph freezing:<\/p>\n\n<pre><code>tf.train.write_graph(self.session.graph_def, some_path)\n\n# get graph definitions with weights\noutput_graph_def = tf.graph_util.convert_variables_to_constants(\n        self.session,  # The session is used to retrieve the weights\n        self.session.graph.as_graph_def(),  # The graph_def is used to retrieve the nodes\n        output_nodes,  # The output node names are used to select the usefull nodes\n)\n\n# optimize graph\nif optimize:\n    output_graph_def = optimize_for_inference_lib.optimize_for_inference(\n            output_graph_def, input_nodes, output_nodes, tf.float32.as_datatype_enum\n    )\n\nwith open(path, \"wb\") as f:\n    f.write(output_graph_def.SerializeToString())\n<\/code><\/pre>\n\n<p>and then loading them through:<\/p>\n\n<pre><code>with tf.Graph().as_default() as graph:\n    with graph.device(\"\/\" + args[name].processing_unit):\n        tf.import_graph_def(graph_def, name=\"\")\n            for key, value in inputs.items():\n                self.input[key] = graph.get_tensor_by_name(value + \":0\")\n<\/code><\/pre>\n\n<p>We would like to save TF2 models in similar way. One protobuf file which will include graph and weights. How can I achieve this?<\/p>\n\n<p>I know that there are some methods for saving:<\/p>\n\n<ul>\n<li><p><code>keras.experimental.export_saved_model(model, 'path_to_saved_model')<\/code><\/p>\n\n<p>Which is experimental and creates multiple files :(.<\/p><\/li>\n<li><p><code>model.save('path_to_my_model.h5')<\/code><\/p>\n\n<p>Which saves h5 format :(.<\/p><\/li>\n<li><p><code>tf.saved_model.save(self.model, \"test_x_model\")<\/code><\/p>\n\n<p>Which agains save multiple files :(.<\/p><\/li>\n<\/ul>\n","answers":[{"AnswerId":"58119259","CreationDate":"2019-09-26T14:38:55.117","ParentId":null,"OwnerUserId":"3214872","Title":null,"Body":"<p>The way I do it at the moment is TF2 -> SavedModel (via <code>keras.experimental.export_saved_model<\/code>) -> frozen_graph.pb (via the <code>freeze_graph<\/code> tools, which can take a <code>SavedModel<\/code> as input). I don't know if this is the \"recommended\" way to do this though.<\/p>\n\n<p>Also, I still don't know how to load back the frozen model and run inference \"the TF2 way\" (aka no graphs, sessions, etc).<\/p>\n\n<p>You may also take a look at <code>keras.save_model('path', save_format='tf')<\/code> which seems to produce checkpoint files (you still need to freeze them, though, so I personally think the saved model path is better)<\/p>\n"}]}
+{"QuestionId":58119320,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm>","CreationDate":"2019-09-26T14:42:22.937","AcceptedAnswerId":58119434.0,"OwnerUserId":12125447.0,"Title":"ValueError : Input 0 of layer lstm is incompatible with the layer: expected ndim=3, found ndim=2. Full shape received: [None, 18]","Body":"<p>I'm new with Keras and I'm trying to build a model for personal use\/future learning. I've just started with python and I came up with this code (with help of videos and tutorials). I have a data of 16324 instances, each instance consists of 18 features and 1 dependent variable.<\/p>\n\n<pre><code>import pandas as pd\nimport os\nimport time\nimport tensorflow as tf\nfrom tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import Dense, Dropout, LSTM, BatchNormalization\nfrom tensorflow.keras.callbacks import TensorBoard, ModelCheckpoint\n\nEPOCHS = 10\nBATCH_SIZE = 64\nNAME = f\"-TEST-{int(time.time())}\"\n\ndf = pd.read_csv(\"EntryData.csv\", names=['1SH5', '1SHA', '1SA5', '1SAA', '1WH5', '1WHA', '2SA5', '2SAA', '2SH5', '2SHA', '2WA5', '2WAA', '3R1', '3R2', '3R3', '3R4', '3R5', '3R6', 'Target'])\n\ndf_val = 14554 \n\nvalidation_df = df[df.index &gt; df_val]\ndf = df[df.index &lt;= df_val]\n\ntrain_x = df.drop(columns=['Target'])\ntrain_y = df[['Target']]\nvalidation_x = validation_df.drop(columns=['Target'])\nvalidation_y = validation_df[['Target']]\n\nmodel = Sequential()\nmodel.add(LSTM(128, input_shape=(train_x.shape[1:]), return_sequences=True))\nmodel.add(Dropout(0.2))\nmodel.add(BatchNormalization())\n\nmodel.add(LSTM(128, return_sequences=True))\nmodel.add(Dropout(0.1))\nmodel.add(BatchNormalization())\n\nmodel.add(LSTM(128))\nmodel.add(Dropout(0.2))\nmodel.add(BatchNormalization())\n\nmodel.add(Dense(32, activation='relu'))\nmodel.add(Dropout(0.2))\n\nmodel.add(Dense(2, activation='softmax'))\n\nopt = tf.keras.optimizers.Adam(lr=0.001, decay=1e-6)\n\nmodel.compile(loss='sparse_categorical_crossentropy',\n              optimizer=opt,\n              metrics=['accuracy'])\n\ntensorboard = TensorBoard(log_dir=f'logs\/{NAME}')\n\nfilepath = \"RNN_Final-{epoch:02d}-{val_acc:.3f}\"  \ncheckpoint = ModelCheckpoint(\"models\/{}.model\".format(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')) # saves only the best ones\n\nhistory = model.fit(\n    train_x, train_y,\n    batch_size=BATCH_SIZE,\n    epochs=EPOCHS,\n    validation_data=(validation_x, validation_y),\n    callbacks=[tensorboard, checkpoint],)\n\nscore = model.evaluate(validation_x, validation_y, verbose=0)\nprint('Test loss:', score[0])\nprint('Test accuracy:', score[1])\n\nmodel.save(\"models\/{}\".format(NAME))\n<\/code><\/pre>\n\n<p>In line <\/p>\n\n<p><code>model.add(LSTM(128, input_shape=(train_x.shape[1:]), return_sequences=True))<\/code><\/p>\n\n<p>is throwing an error: <\/p>\n\n<blockquote>\n  <p>ValueError: Input 0 of layer lstm is incompatible with the layer:\n  expected ndim=3, found ndim=2. Full shape received: [None, 18]<\/p>\n<\/blockquote>\n\n<p>I was searching for solution on this site and on google for few hours now and I was not able to find proper answer for this or I was not able to implement the solution for similar problem.<\/p>\n\n<p>Thank you for any tips.<\/p>\n","answers":[{"AnswerId":"58119434","CreationDate":"2019-09-26T14:49:04.967","ParentId":null,"OwnerUserId":"10908375","Title":null,"Body":"<p>An LSTM network expects a three dimensional input. Here is the code that worked for me:<\/p>\n\n<p>Before feeding it to the neural network: <\/p>\n\n<p><code>X_train = X_train.reshape(X_train.shape[0], 1, X_train.shape[1])<\/code><\/p>\n\n<p>In the neural network: <\/p>\n\n<p><code>LSTM(64, input_shape=(X_train.shape[1:])),<\/code><\/p>\n"}]}
+{"QuestionId":58119390,"AnswerCount":0,"Tags":"<tensorflow><keras><tf.keras>","CreationDate":"2019-09-26T14:46:39.623","AcceptedAnswerId":null,"OwnerUserId":11312376.0,"Title":"Trouble Converting a LSTM model written in Tensorflow API to Keras API","Body":"<p>I am trying to convert some open source code that was written in the TensorFlow API into Keras API. I am getting the correct input and output shapes for each layer, however, it complains when I try to do a test on the model itself. Please excuse the code, I simplified a lot for brevity.<\/p>\n\n<p>Here is the TensorFlow API:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>        hidden_units_g = 100\n        num_generated_features = 1\n        seq_length = 30\n        batch_size = 28\n        latent_dim = 5\n\n        W_out_G = tf.get_variable(name='W_out_G', shape=[hidden_units_g, num_generated_features], initializer=W_out_G_initializer)\n        b_out_G = tf.get_variable(name='b_out_G', shape=num_generated_features, initializer=b_out_G_initializer)\n\n        # shape: (28, 30, 5)\n        inputs = tf.random_normal(shape=(batch_size, seq_length, latent_dim))\n\n        cell = LSTMCell(num_units=hidden_units_g,\n                           state_is_tuple=True,\n                           initializer=lstm_initializer,\n                           bias_start=bias_start,\n                           reuse=reuse)\n\n        # rnn_output shape: (28, 30, 100)\n        rnn_outputs, rnn_states = tf.nn.dynamic_rnn(\n            cell=cell,\n            dtype=tf.float32,\n            sequence_length=[seq_length]*batch_size,\n            inputs=inputs)\n\n        # reshape shape: (840, 100)\n        rnn_outputs_2d = tf.reshape(rnn_outputs, [-1, hidden_units_g])\n\n        # shape: ( 840, 1)\n        logits_2d = tf.matmul(rnn_outputs_2d, W_out_G) + b_out_G\n\n        # shape: (840, 1)\n        output_2d = tf.nn.tanh(logits_2d)\n\n        # shape: (28, 30, 1)\n        output_3d = tf.reshape(output_2d, [-1, seq_length, num_generated_features])\n\n<\/code><\/pre>\n\n<p>Here is my attempt to recreate this in Keras:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>gen = tf.keras.models.Sequential([\n    tf.keras.layers.LSTM(100, input_shape=(30, 5), return_sequences=False),\n    #tf.keras.layers.Dense(1, activation='tanh'),\n    tf.keras.layers.Dense(1, activation='tanh'),\n    tf.keras.layers.Reshape(target_shape=(30, 1))\n])\nModel: \"sequential\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nlstm (LSTM)                  (None, 100)               42400     \n_________________________________________________________________\ndense (Dense)                (None, 1)                 101       \n_________________________________________________________________\nreshape (Reshape)            (None, 30, 1)             0         \n=================================================================\nTotal params: 42,501\nTrainable params: 42,501\nNon-trainable params: 0\n<\/code><\/pre>\n\n<p>Also applying <code>tf.keras.utils.plot_model()<\/code> confirms that my input\/output shapes are exactly the same. However, when I try to do a mock test using: <code>np.random.randn(28, 30, 5)<\/code> I get an error saying that it fails to reshape the output of dense, it expects 840 values, but only saw 28. <\/p>\n\n<p>Is there a way to print out the actual batch shapes as it moves through the model in Keras?<\/p>\n\n<p>Thank you<\/p>\n\n<p>Edit: Here is the full input\/output when using <code>keras.utils.plot_model()<\/code><\/p>\n\n<pre><code>LSTM: input: (?, 30, 5)\n      output:(?, 100)\n\nDense: Input(?, 100)\n     : Output: (?, 1)\n\nReshape: Input: (?, 1)\n         Output: (?, 30, 1)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58119492,"AnswerCount":2,"Tags":"<python-3.x><tensorflow><nvidia><cudnn>","CreationDate":"2019-09-26T14:52:18.837","AcceptedAnswerId":null,"OwnerUserId":7099139.0,"Title":"Is tensorflow-gpu (=<1.14) with CUDA & cuDDN supported in Red Hat Enterprise Linux 7\/8 or cent os 7?","Body":"<p>I have cent os 7 with Nvidia 2080 super graphic &amp; i installed the following:<\/p>\n\n<ol>\n<li>Graphics Driver vision 430.4<\/li>\n<li>CUDA 10.1<\/li>\n<li>CuDNN 7.6<\/li>\n<li>conda env python 3.7<\/li>\n<li>tensorflow-gpu 1.14 (pip install tensorflow-gpu)<\/li>\n<\/ol>\n\n<p>then pip instill the required libraries for the code.<\/p>\n\n<p>but when i tried to run this code <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/image_ocr.py\" rel=\"nofollow noreferrer\">link<\/a> i get cuDNN initialization error..<\/p>\n\n<p>any guide or suggestions are much appreciated :-)  <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/larwP.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/larwP.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"58119861","CreationDate":"2019-09-26T15:12:54.687","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>I can not say about CentOS, since I have never installed on it but after looking at the information you provided. The error in your picture says, Cudnn can not be initialised this is most probably bc of compatibility issues. You need to follow <a href=\"https:\/\/www.tensorflow.org\/install\/gpu#software_requirements\" rel=\"nofollow noreferrer\">these<\/a> requirements.\nSo, remove Cuda 10.1 completely and install Cuda 10. See if the problem still exists.<\/p>\n"},{"AnswerId":"58125654","CreationDate":"2019-09-26T22:51:25.613","ParentId":null,"OwnerUserId":"10543528","Title":null,"Body":"<p>If you're using a conda environment, you should install packages with conda. Try installing tensorflow-gpu with conda; it should take care of all of your dependencies aside from the NVidia driver, which you already have. Doing things in this way allows conda to take care of additional package dependencies, like the CUDA toolkit and cudnn, among other things. <\/p>\n\n<p>Source: Installing just the 10.1 version of the CUDA toolkit (with the driver) to a system location and tensorflow-gpu in an anaconda environment, I have tensorflow-gpu 1.14 running with Cuda 10.1 on a CentOS 7 install as well as on a RHEL7 install. <\/p>\n"}]}
+{"QuestionId":58119629,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><keras><deep-learning>","CreationDate":"2019-09-26T14:59:59.103","AcceptedAnswerId":null,"OwnerUserId":11551046.0,"Title":"Validation Accuracy of CNN not increasing","Body":"<p>I have around 8200 images for face detection task. 4800 of them contain human faces. The other 3400 images contain images of 3D human face masks(which are made of rubber\/latex), human cartoon faces, faces of monkeys. I want to detect whether the given image contains a real human face or not.<\/p>\n\n<p>I have trained numerous networks, changing hyper parameters but every time my training accuracy shots up to over 98% and validation accuracy stays at around 60-70%. I have tried out networks containing 3-5 Conv layers and one FC layers. I used L2 regularization, batch norm, data augmentation and dropout to remove overfitting. I then tried out reducing the learning rate of Adam optimizer as the training progressed. I trained the network for more than 100 epochs and sometimes upto 200 epochs. However, the best validation accuracy(20% of dataset) I could achieve was 71%. Is there anyway out to improve the validation accuracy above 85%?\nI used the following architecture with input image size of 256*256*3 and trained them with a batch size of 16.<\/p>\n\n<pre><code>regularizer = tf.keras.regularizers.l2(l=0.005)\nmodel = tf.keras.models.Sequential([\n\n    tf.keras.layers.Conv2D(64, (5, 5),strides=(2, 2), activation='relu', input_shape=(256, 256, 3), kernel_regularizer=regularizer),\n    tf.keras.layers.MaxPooling2D(2, 2),\n\n    tf.keras.layers.Conv2D(96, (5, 5), padding='same', activation='relu', kernel_regularizer=None),\n    tf.keras.layers.MaxPooling2D(2, 2),\n\n    tf.keras.layers.Conv2D(128, (3, 3), padding='same', activation='relu', kernel_regularizer=None),\n    tf.keras.layers.MaxPooling2D(2, 2),\n\n    tf.keras.layers.Conv2D(256, (3, 3), padding='same', activation='relu', kernel_regularizer=None),\n    tf.keras.layers.MaxPooling2D(2, 2),\n\n    tf.keras.layers.Flatten(),\n\n    #tf.keras.layers.Dense(2048, activation='relu', kernel_regularizer=regularizer),\n    tf.keras.layers.Dense(4096, activation='relu', kernel_regularizer=None),\n    tf.keras.layers.Dropout(0.4),\n    tf.keras.layers.Dense(1, activation='sigmoid', kernel_regularizer=regularizer)\n])\n<\/code><\/pre>\n","answers":[{"AnswerId":"58120077","CreationDate":"2019-09-26T15:24:49.293","ParentId":null,"OwnerUserId":"2846923","Title":null,"Body":"<ul>\n<li>Make sure you are using all available forms of data augmentation (scaling, rotation, translation, flips, etc.).<\/li>\n<li>Use kernel regularizers on all layers.<\/li>\n<li>Add <code>SpatialDropout2D<\/code> after all Conv layers.<\/li>\n<li>Add <code>BatchNormalization<\/code> after all Conv and Dense layers (except for the last Dense\/sigmoid one, obviously).<\/li>\n<li>Reduce the size of your network (fewer layers and\/or fewer filters\/units per layer); you want the smallest possible network that can still learn the training data.<\/li>\n<\/ul>\n\n<p>If all of those combined are not enough to get good validation accuracy, then you probably just don't have enough data.<\/p>\n\n<hr>\n\n<p>A few tips that probably won't reduce overfitting, but tend to be helpful in general:<\/p>\n\n<ul>\n<li>Prefer sequences of 3x3 kernel conv layers rather than single conv layers with 5x5 or larger kernels.<\/li>\n<li>Replace the Flatten layer with a GlobalAveragePooling layer, and probably remove all Dense layers except the last one.<\/li>\n<li>Use either stride=2 <em>or<\/em> MaxPooling, not both.<\/li>\n<\/ul>\n"}]}
+{"QuestionId":58119638,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2019-09-26T15:00:33.173","AcceptedAnswerId":null,"OwnerUserId":774133.0,"Title":"Restore original text from Keras\u2019s imdb dataset - AGAIN","Body":"<p>I have an issue with the keras IMDB database.<\/p>\n\n<p>The code I am starting from is in the accepted answer here:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/42821330\/restore-original-text-from-keras-s-imdb-dataset\">Restore original text from Keras\u2019s imdb dataset<\/a><\/p>\n\n<pre><code>import keras\nNUM_WORDS=1000 # only use top 1000 words\nINDEX_FROM=3   # word index offset\n\ntrain,test = keras.datasets.imdb.load_data(num_words=NUM_WORDS, index_from=INDEX_FROM)\ntrain_x,train_y = train\ntest_x,test_y = test\n\nword_to_id = keras.datasets.imdb.get_word_index()\nword_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\nid_to_word = {value:key for key,value in word_to_id.items()}\nprint(' '.join(id_to_word[id] for id in train_x[0] ))\n<\/code><\/pre>\n\n<p>However, I agree with the comment by Nate Raw saying:<\/p>\n\n<blockquote>\n  <p>This code is actually incorrect. One line should be changed to\n  word_to_id={k:(v+INDEX_FROM-1) for k,v in word_to_id.items()}. The\n  indexes in the downloaded word_to_id dictionary are actually starting\n  at 1. So, when you add INDEX_FROM to the indexes, it causes there to\n  be a gap between id_to_word[2] and id_to_word[4]. There is no value\n  for id_to_word[3]<\/p>\n<\/blockquote>\n\n<p>If I follow this comment and use <code>INDEX_FROM -  1<\/code>, the reconstructed review text does not make any sense.<\/p>\n\n<p>What about the id_to_word[3]?<\/p>\n\n<p>Is there anyone who tried to solve this issue?<\/p>\n","answers":[]}
+{"QuestionId":58119930,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-26T15:16:32.073","AcceptedAnswerId":58122687.0,"OwnerUserId":3195597.0,"Title":"Keras fit_generator with images from directory and a constant tensor","Body":"<p>I have a simple CNN with input images of shape (5,5,3). As a first step I want to add a constant tensor to the input.\nAccording to the answer in my previous <a href=\"https:\/\/stackoverflow.com\/q\/58107942\/3195597\">SO question<\/a>, I have to define the constant tensor as an input layer (const_input), so that I can Add() it to the image data (raw_input). The model is compiled without errors:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from __future__ import print_function\nimport tensorflow as tf\nimport numpy as np\nimport keras\nfrom keras import backend as K\nfrom keras.models import Model\nfrom keras.layers import Input, Add\nfrom keras.layers import Conv2D, MaxPooling2D\nfrom keras.layers import Dropout, Flatten, Dense\nfrom keras.preprocessing.image import ImageDataGenerator\n# Python 2.7.10 - keras version 2.2.0 - tf.VERSION '1.8.0'\n\ncnn_layer1 = Conv2D(32, (4, 4), activation='relu')\ncnn_layer2 = MaxPooling2D(pool_size=(2, 2))\ncnn_layer3 = Dense(64, activation='relu')\ncnn_layer4 = Dropout(0.1)\ncnn_output = Dense(2, activation='softmax')\n\nraw_input   = Input(shape=(5, 5, 3))\nconst_input = Input(shape=(5, 5, 3))\n\npre_proc = Add()([raw_input, const_input])\nlay1 = cnn_layer1(pre_proc)\nlay2 = cnn_layer2(lay1)\nlay3 = Flatten()(lay2)\nlay4 = cnn_layer3(lay3)\nlay5 = cnn_layer4(lay4)\nlay_out = cnn_output(lay5)\n\nmodel = Model(inputs=[raw_input, const_input], outputs=lay_out)\n\nmodel.compile(optimizer='adam',\n              loss='categorical_crossentropy',\n              metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>Now I try to provide the constant tensor as an input along with the images that are read from directory:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>batch_size = 10\ntrain_datagen      = ImageDataGenerator(rescale=1.\/255)\nvalidation_datagen = ImageDataGenerator(rescale=1.\/255)\n\ntrain_generator = train_datagen.flow_from_directory(\n        'cd_data\/train', \n        target_size=(5, 5), \n        classes=['cat', 'dog'],\n        batch_size=batch_size) \n\nvalidation_generator = validation_datagen.flow_from_directory(\n        'cd_data\/validation',\n        target_size=(5, 5),\n        classes=['cat', 'dog'],\n        batch_size=batch_size)\n\nconst_array = np.array(\n    [[[5.0,0.0,0.0],[0.0,0.0,-3.0],[-10.0,0.0,0.0],[0.0,0.0,4.0],[-20.0,0.0,0.0]],\n    [[-15.0,0.0,12.0],[0.0,4.0,0.0],[-3.0,0.0,10.0],[-18.0,0.0,0.0],[20.0,0.0,-6.0]],\n    [[0.0,0.0,6.0],[0.0,-2.0,-6.0],[0.0,0.0,2.0],[0.0,0.0,-9.0],[7.0,-6.0,0.0]],\n    [[-3.0,4.0,0.0],[11.0,-12.0,0.0],[0.0,0.0,0.0],[0.0,0.0,7.0],[0.0,0.0,2.0]],\n    [[0.0,0.0,0.0],[0.0,1.0,-2.0],[4.0,0.0,3.0],[0.0,0.0,0.0],[0.0,0.0,0.0]]])\n\ndef merge_generator():\n        while True:\n                next_image = train_generator.next() \n                yield [next_image[0], const_array], next_image[1]\n\ntrain_gen_with_const = merge_generator()\n<\/code><\/pre>\n\n<p>Executing the fit_generator leads to error below<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.fit_generator(\n        train_gen_with_const,\n        steps_per_epoch=2,\n        epochs=1,\n        verbose=2, # one line per epoch\n        validation_data=validation_generator,\n        validation_steps=2)\n<\/code><\/pre>\n\n<blockquote>\n  <p>ValueError: Error when checking input: expected input_2 to have 4 dimensions, but got array with shape (5, 5, 3)<\/p>\n<\/blockquote>\n\n<p>I tried to provide the missing dimension like this<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>const_batch = np.broadcast_to(const_array, (batch_size, 5, 5, 3))\n\ndef merge_generator():\n        while True:\n                next_image = train_generator.next() \n                yield [next_image[0], const_batch], next_image[1]\n<\/code><\/pre>\n\n<p>but this leads to <\/p>\n\n<blockquote>\n  <p>ValueError: All input arrays (x) should have the same number of samples. Got array shapes: [(2, 5, 5, 3), (10, 5, 5, 3)]<\/p>\n<\/blockquote>\n\n<p>What is the right way to provide this constant tensor input?\nAny help is highly appreciated!<\/p>\n","answers":[{"AnswerId":"58122687","CreationDate":"2019-09-26T18:24:07.527","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>The problem lies with your <code>validation_data=<\/code> argument; your model expects <em>two<\/em> input arrays, whereas <code>validation_generator<\/code> supplies only <em>one<\/em>. You fixed this with <code>train_gen_with_const<\/code> - just extend it to val:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def merge_generator(): # const_batch inside the function to apply to each image\n    while True:\n        next_image = train_generator.next()                      \n        const_batch = np.broadcast_to(const_array, (len(next_image[0]), 5, 5, 3))\n        yield [next_image[0], const_batch], next_image[1]\n\ndef val_merge_generator():  \n    while True:\n        next_image = validation_generator.next()\n        const_batch = np.broadcast_to(const_array, (len(next_image[0]), 5, 5, 3))\n        yield [next_image[0], const_batch], next_image[1]\n<\/code><\/pre>\n\n<p>Remember, internally, <code>fit_generator<\/code> calls <code>train_on_batch(x, y)<\/code> and <code>evaluate(x, y)<\/code> - so each must receive the same dimensionality for <code>x<\/code> and <code>y<\/code> from both generators.<\/p>\n"}]}
+{"QuestionId":58120324,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2019-09-26T15:38:22.277","AcceptedAnswerId":null,"OwnerUserId":12126046.0,"Title":"Tensorflow Summary Scalar is not being displayed","Body":"<p>I am using the following code to produce scalar graphs for my accuracy and cost, but the scalar summaries are not being displayed at the tensorboard. It gives me an error saying <code>No scalar data was found<\/code>. Can someone have a look please? code for the model:<\/p>\n\n<pre><code>def train_neural_network(x):\n    prediction = convolutional_neural_network(x)\n    merged_summary_op = tf.summary.merge_all()\n\n    with tf.name_scope(\"cost\"):\n        cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=prediction, labels=y))\n        optimizer = tf.train.AdamOptimizer(learning_rate=1e-3).minimize(cost)\n        tf.summary.scalar(\"cost\", cost)\n\n    hm_epochs = 10\n    with tf.Session() as sess:\n\n        writer = tf.summary.FileWriter('C:\/Thesis\/Conv3d\/69', sess.graph)\n        sess.run(tf.initialize_all_variables())\n\n        successful_runs = 0\n        total_runs = 0\n\n        for epoch in range(hm_epochs):\n            epoch_loss = 0\n            for data in train_data:\n                total_runs += 1\n                try:\n                    X = data[0]\n\n\n                    Y = data[1]\n                    _, c = sess.run([optimizer, cost], feed_dict={x: X, y: Y})\n                    writer.add_summary(summary, global_step=epoch)\n                    epoch_loss += c\n\n                    successful_runs += 1\n\n                except Exception as e:\n\n                    pass\n\n            print('Epoch', epoch + 1, 'completed out of', hm_epochs, 'loss:', epoch_loss)\n            with tf.name_scope(\"accuracy\"):\n                correct = tf.equal(tf.argmax(prediction, 1), tf.argmax(y, 1))\n                accuracy = tf.reduce_mean(tf.cast(correct, 'float'))\n                tf.summary.scalar(\"accuracy\", accuracy)\n\n\n\n            print('Accuracy:', accuracy.eval({x: [i[0] for i in validation_data], y: [i[1] for i in validation_data]}))\n        print('Done. Finishing accuracy:')\n        print('Accuracy:', accuracy.eval({x: [i[0] for i in validation_data], y: [i[1] for i in validation_data]}))\n\n\n        print('fitment percent:', successful_runs \/ total_runs)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58122132","CreationDate":"2019-09-26T17:44:05.683","ParentId":null,"OwnerUserId":"9393102","Title":null,"Body":"<p>You need to call <code>merge_all<\/code> <em>after<\/em> defining the summary ops. What happens right now is that the op simply doesn't have anything to summarize (since it's called first) and unfortunately isn't \"smart\" enough to add summary ops that are defined later.<\/p>\n\n<p>Note that I also \"fixed\" the code in that you generally shouldn't run TF ops in a loop; I moved all the accuracy stuff before the training loop.<\/p>\n\n<pre><code>def train_neural_network(x):\n    prediction = convolutional_neural_network(x)\n\n    with tf.name_scope(\"cost\"):\n        cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=prediction, labels=y))\n        optimizer = tf.train.AdamOptimizer(learning_rate=1e-3).minimize(cost)\n        tf.summary.scalar(\"cost\", cost)\n\n    with tf.name_scope(\"accuracy\"):\n        correct = tf.equal(tf.argmax(prediction, 1), tf.argmax(y, 1))\n        accuracy = tf.reduce_mean(tf.cast(correct, 'float'))\n        tf.summary.scalar(\"accuracy\", accuracy)\n\n    merged_summary_op = tf.summary.merge_all()\n    hm_epochs = 10\n    with tf.Session() as sess:\n\n        writer = tf.summary.FileWriter('C:\/Thesis\/Conv3d\/69', sess.graph)\n        sess.run(tf.initialize_all_variables())\n\n        successful_runs = 0\n        total_runs = 0\n\n        for epoch in range(hm_epochs):\n            epoch_loss = 0\n            for data in train_data:\n                total_runs += 1\n                try:\n                    X = data[0]\n\n\n                    Y = data[1]\n                    _, c = sess.run([optimizer, cost], feed_dict={x: X, y: Y})\n                    writer.add_summary(summary, global_step=epoch)\n                    epoch_loss += c\n\n                    successful_runs += 1\n\n                except Exception as e:\n\n                    pass\n\n            print('Epoch', epoch + 1, 'completed out of', hm_epochs, 'loss:', epoch_loss)\n\n            print('Accuracy:', accuracy.eval({x: [i[0] for i in validation_data], y: [i[1] for i in validation_data]}))\n        print('Done. Finishing accuracy:')\n        print('Accuracy:', accuracy.eval({x: [i[0] for i in validation_data], y: [i[1] for i in validation_data]}))\n\n\n        print('fitment percent:', successful_runs \/ total_runs)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58120369,"AnswerCount":1,"Tags":"<python><keras><neural-network><autoencoder>","CreationDate":"2019-09-26T15:40:32.987","AcceptedAnswerId":58120897.0,"OwnerUserId":1692072.0,"Title":"Which parameters prohibit the autoencoder from learning the right distribution?","Body":"<p>I have a very minimal example of an autoencoder:<\/p>\n\n<pre><code>from tensorflow.keras.models import Model\nfrom tensorflow.keras.layers import Input, Dense\nimport pandas as pd\nimport numpy as np\nimport seaborn as sns\nimport matplotlib.pyplot as plt\n%matplotlib inline\n<\/code><\/pre>\n\n<p>First I create a data set with the highly correlated variables A and B (that are already normalized)<\/p>\n\n<pre><code>X = pd.DataFrame( (np.random.randn(1000,2)), columns=[\"A\", \"B\"] )\nX[\"B\"] = X[\"A\"] + X[\"B\"]\/4\n<\/code><\/pre>\n\n<p>Then I setup the autoencoder and train it<\/p>\n\n<pre><code>aeInput = Input(2)\nencode = Dense(2, activation='relu')(aeInput)\naeOutput = Dense(2, activation='relu')(encode)\n\nAE = Model(aeInput, aeOutput, name=\"autoencoder\")\nAE.compile(optimizer='adam', loss=\"mean_squared_error\", )\nTrainAE = AE.fit( x=X, y=X, epochs=100, batch_size=2**5,)\n<\/code><\/pre>\n\n<p>training looks good and converges smoothly, but when I look at the result the output is mainly zeros. <\/p>\n\n<pre><code>f, ax = plt.subplots(figsize=(8, 8))\nsns.kdeplot( X, shade=False, axis=ax)\nsns.kdeplot( AE.predict(X), shade=False, axis=ax)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/jUpKU.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/jUpKU.png\" alt=\"Blue is the original distribution, orange the autoencoded one\"><\/a><\/p>\n\n<p>This seems very odd to me, because the encoding layer is as large as the input, so a trivial and loss-free solution would simple be to wire the first A neuron straight through, with an activation of 1 and same for the second neuron encoding for B. Why is this not happening? Is there any parameter I use falsely?<\/p>\n","answers":[{"AnswerId":"58120897","CreationDate":"2019-09-26T16:14:05.573","ParentId":null,"OwnerUserId":"4531017","Title":null,"Body":"<p>One issue is that your final layer has the <code>relu<\/code> activation, which has a minimum of 0.  If you would like to predict numbers less than 0 on the final layer, you can change the activation to \"linear,\" like this<\/p>\n\n<p><code>aeOutput = Dense(2, activation='linear')(encode)<\/code><\/p>\n"}]}
+{"QuestionId":58120448,"AnswerCount":2,"Tags":"<python><machine-learning><pytorch>","CreationDate":"2019-09-26T15:45:54.953","AcceptedAnswerId":null,"OwnerUserId":12126090.0,"Title":"Pytorch :Why my dataset variance do not get the correct result?","Body":"<p>Here is the function I wrote:<\/p>\n\n<pre><code>def channel_var(image_dataset):\n    res = image_dataset[0]\n    for image in image_dataset[1:]:\n        res += image\n    return tuple(map(lambda x: x\/len(image_dataset),\n     (torch.var(res[0]),\n      torch.var(res[1]),\n      torch.var(res[2]))))\n<\/code><\/pre>\n\n<p>then I tested it with a Normal distribution : <\/p>\n\n<pre><code>m = normal.Normal(0, 3)\nm.sample((1, 3, 32, 32))\n<\/code><\/pre>\n\n<p>And I get this wrong result : <\/p>\n\n<pre><code>channel_var(list_test)\n\n&gt;&gt;(tensor(0.0338), tensor(0.0352), tensor(0.0365))\n<\/code><\/pre>\n\n<p>Thank you<\/p>\n","answers":[{"AnswerId":"58121458","CreationDate":"2019-09-26T16:51:57.523","ParentId":null,"OwnerUserId":"3236925","Title":null,"Body":"<p>Your function is wrong. And that's because you are computing the average image and then computing the channel variance in the average image. I don't think you want that. You can just find the variance in each channel by using<\/p>\n\n<p><code>torch.var(img, dim=[0,2,3])<\/code><\/p>\n\n<p>assuming <code>dim=1<\/code> is the channel dimension and img is a torch tensor. If img is not a torch tensor, you can concatenate list of imgs to make a tensor. <\/p>\n\n<p>You can do this as <code>torch.var(torch.cat(img, dim=0), dim=[0,2,3])<\/code> <code>cat<\/code> operation concatenates list to a tensor. <\/p>\n"},{"AnswerId":"58125605","CreationDate":"2019-09-26T22:44:43.593","ParentId":null,"OwnerUserId":"12126090","Title":null,"Body":"<p>I want to calculate the variance of all my dataset for each channel . Each element is a torch tensor. You're method seems good for 1 image. Can you show me an implementation? <\/p>\n\n<p>Thanks<\/p>\n"}]}
+{"QuestionId":58120738,"AnswerCount":0,"Tags":"<python><keras><lstm>","CreationDate":"2019-09-26T16:04:09.190","AcceptedAnswerId":null,"OwnerUserId":3552144.0,"Title":"Having trouble tuning LSTM with an increasing val_loss issue","Body":"<p>Problem:<\/p>\n\n<ul>\n<li>Im looking to forecast the number of transcriptions for a particular user. I am using an LSTM to do so but as I fit the model, my training error continues to decrease but my validation error increases. I tried to handle for overfitting (see below) but that doesn't seem to be working. <\/li>\n<\/ul>\n\n<p><strong>Summary:<\/strong><\/p>\n\n<p>Training Samples: 409<\/p>\n\n<p>Value Counts of response variable (transactions in next quarter):<\/p>\n\n<ul>\n<li>0: 212<\/li>\n<li>1: 125<\/li>\n<li>2: 56<\/li>\n<li>3: 13<\/li>\n<li>4: 3<\/li>\n<\/ul>\n\n<p>Training Set:<\/p>\n\n<ul>\n<li>Look back on the number of transcriptions over the last 4 weeks from time, t.<\/li>\n<\/ul>\n\n<p>Here is the code:<\/p>\n\n<pre><code># create and fit the LSTM network\nmodel = Sequential()\nmodel.add(GRU(5,  input_shape=(1, look_back)))\nmodel.add(Dropout(0.6))\nmodel.add(Dense(1))\nmodel.compile(loss='mean_squared_error', optimizer='adam')\nmodel.fit(trainX, trainY, epochs=100, batch_size=1,shuffle = False, verbose=2,validation_split = 0.10)\n<\/code><\/pre>\n\n<p>Any help would be much appreciated and please let me know if any other details are required. <\/p>\n","answers":[]}
+{"QuestionId":58120931,"AnswerCount":1,"Tags":"<keras><deep-learning><lstm>","CreationDate":"2019-09-26T16:16:17.980","AcceptedAnswerId":58216763.0,"OwnerUserId":9084663.0,"Title":"The loss and accuracy of this LSTM both drop to nearly 0 at the same epoch","Body":"<p>I'm trying to train an LSTM to predict the the <code>Nth<\/code> token using the <code>N-1<\/code> tokens preceding it<\/p>\n\n<p>For each One-Hot encoded token, I try to predict the next token. After three layers of LSTM, the results are fed to a Dense layer (wrapped in he <code>TimeDistributed<\/code> layer wrapper) to re-encode the results into the same One-Hot encoding.<\/p>\n\n<p>Oddly enough, after a few epochs the accuracy (in both train and validation) drops to nearly 0 (bad news), while the loss also drops to nearly 0 (good news?).<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/TNZrF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/TNZrF.png\" alt=\"Representative results\"><\/a><\/p>\n\n<p>Why does this happen?\nI know that I cannot expect the loss and accuracy to always go in the opposite directions (seeing as loss uses categorical cross-entropy over all categories, while accuracy uses merely the best or k best categories), but still - this behavior is highly unexpected and unexplainable.<\/p>\n\n<p>What causes this?\nAm I'm doing something wrong?\nHow should I change my code to allow my network to progress towards more and more accurate predictions? <\/p>\n\n<p>My code is as follows:<\/p>\n\n<pre><code>import numpy as np\nimport glob\n\nimport keras\nfrom keras.models import Sequential\nfrom keras.layers import LSTM, Dense, TimeDistributed,Lambda, Dropout, Activation\nfrom keras.metrics import top_k_categorical_accuracy\nfrom keras.callbacks import ModelCheckpoint\n\n###\nimport matplotlib\nmatplotlib.use('Agg') # prevents it from failing when there is no display\nimport matplotlib.pyplot as plt\nimport keras.backend as K\n###\n\nname='Try_6'\nmodel_designation=str(name)+'_'\n\ntrain_val_split=0.2 # portion to be placed in validation\n\n\ntrain_control_number=0\nval_control_number=0\nbatch_size = 16\n\ndef my_3D_top_5(true, pred):\n    features_num=int(list(pred.shape)[-1])\n\n    true = K.reshape(true, (-1, features_num))   \n    pred = K.reshape(pred, (-1, features_num))\n    return top_k_categorical_accuracy(true, pred, k=5)\n\ndef my_3D_top_10(true, pred):\n    features_num=int(list(pred.shape)[-1])\n\n    true = K.reshape(true, (-1, features_num))   \n    pred = K.reshape(pred, (-1, features_num))\n    return top_k_categorical_accuracy(true, pred, k=10)\n\n\n\ndef basic_LSTM(features_num):\n    model = Sequential()\n    model.add(LSTM(40, return_sequences=True, input_shape=(None, features_num)))\n    model.add(LSTM(40, return_sequences=True))\n    model.add(LSTM(40, return_sequences=True))\n\n    model.add(TimeDistributed(Dense(features_num)))\n    model.add(Activation('linear')) \n\n    print(model.summary())\n    model.compile(loss='categorical_crossentropy', optimizer='adam',metrics=['accuracy',my_3D_top_5,my_3D_top_10])  \n    return (model)\n\n\ndef main ():\n    input_files=glob.glob('*npy')\n    data_list,dim=loader(input_files)\n    train_list,val_list=data_spliter(data_list)\n\n    train_list=group_data(train_list,batch_size)\n    val_list=group_data(val_list,batch_size)\n\n    filepath = \"saved-model-\"+model_designation+\"-{epoch:02d}.hdf5\"\n    checkpoint = ModelCheckpoint(filepath, save_best_only=False)\n    callbacks_list=[checkpoint] \n\n\n    model=basic_LSTM(dim)\n    history=model.fit_generator(train_generator(train_list), steps_per_epoch=len(train_list), epochs=30, verbose=1,validation_data=val_generator(val_list),validation_steps=len(val_list),callbacks=callbacks_list)\n    report(history)\n\n\ndef group_data(data_list,size):  # groups data and elongate it to match\n    output=[]\n    list_of_sizes=[]\n    for data in data_list:\n        list_of_sizes.append(list(data.shape)[1]) \n\n    data_list = [x for _, x in sorted(zip(list_of_sizes,data_list), key=lambda pair: pair[0])]\n\n    while len(data_list)&gt;size:\n        this=data_list[:size]\n        data_list=data_list[size:]\n        combined=(elongate_and_combine(this))\n        output.append(combined)\n\n\n    combined=(elongate_and_combine(data_list))\n    output.append(combined)\n\n\n    return (output)\n\ndef elongate_and_combine(data_list):\n\n    max_length= (list(data_list[-1].shape)[1]) \n    last_element=list.pop(data_list)\n    output=last_element\n\n\n\n\n    stop_codon=last_element[0,(max_length-1),:]\n    stop_codon=stop_codon.reshape(1,1,stop_codon.size)\n\n    for data in data_list:\n        size_of_data=list(data.shape)[1]\n        while size_of_data&lt;max_length:\n            data=np.append(data, stop_codon, axis=1)\n            size_of_data=list(data.shape)[1]\n        output=np.append(output, data, axis=0)\n\n\n    return (output)\n\n\ndef train_generator(data_list):\n    while True:\n        global train_control_number\n        train_control_number=cycle_throught(len(data_list),train_control_number)\n        #print (train_control_number)       \n        this=data_list[train_control_number]\n\n\n        x_train = this [:,:-1,:] # all but the last 1\n        y_train = this [:,1:,:] # all but the first 1\n\n        yield (x_train, y_train)\n\n\n\n\ndef val_generator(data_list):\n    while True:\n        global val_control_number\n        val_control_number=cycle_throught(len(data_list),val_control_number)\n        #print (val_control_number)     \n        this=data_list[val_control_number]\n        x_train = this [:,:-1,:] # all but the last 1\n        y_train = this [:,1:,:] # all but the first 1\n\n        yield (x_train, y_train)\n\n\n\ndef cycle_throught (total,current):\n    current+=1\n    if (current==total):\n        current=0\n    return (current)\n\n\ndef loader(input_files):\n\n    data_list=[]\n\n    for input_file in input_files:\n        a=np.load (input_file)\n        incoming_shape=list(a.shape)\n        requested_shape=[1]+incoming_shape\n        a=a.reshape(requested_shape)\n        #print (a.shape)\n        data_list.append(a)\n\n\n    return (data_list,incoming_shape[-1])\n\n\ndef data_spliter(input_list):\n    val_num=int(len(input_list)*train_val_split)\n    validation=input_list[:val_num]\n    train=input_list[val_num:]\n\n    return (train,validation)\n\ndef report(history) :\n\n\n    print(history.history.keys())\n\n\n    acc = history.history['acc']\n    val_acc = history.history['val_acc']\n\n    loss = history.history['loss']\n    val_loss = history.history['val_loss']\n\n    acc_5=history.history['my_3D_top_5']\n    val_acc_5=history.history['val_my_3D_top_5']\n\n    acc_10=history.history['my_3D_top_10']\n    val_acc_10=history.history['val_my_3D_top_10']\n\n\n\n    epochs = range(1, len(acc) + 1)\n\n    fig, axes = plt.subplots(nrows=2, ncols=2, figsize=(10, 6))\n\n\n\n    axes[0][0].plot(epochs, acc, 'bo', label='Training acc')\n    axes[0][0].plot(epochs, val_acc, 'b', label='Validation acc')\n    axes[0][0].set_title('Training and validation accuracy')\n    axes[0][0].legend()\n\n\n\n    axes[0][1].plot(epochs, loss, 'ro', label='Training loss')\n    axes[0][1].plot(epochs, val_loss, 'r', label='Validation loss')\n    axes[0][1].set_title('Training and validation loss')\n    axes[0][1].legend()\n\n    axes[1][0].plot(epochs, acc_5, 'go', label='Training acc over top 5')\n    axes[1][0].plot(epochs, val_acc_5, 'g', label='Validation acc over top 5')\n    axes[1][0].set_title('Training and validation accuracy over top 5')\n    axes[1][0].legend()\n\n    axes[1][1].plot(epochs, acc_10, 'mo', label='Training acc over top 10')\n    axes[1][1].plot(epochs, val_acc_10, 'm', label='Validation acc over top 10')\n    axes[1][1].set_title('Training and validation accuracy over top 10')\n    axes[1][1].legend()\n\n    fig.tight_layout()\n    fig.savefig('fig_'+name+'.png')   # save the figure to file\n\n\n\nmain()\n<\/code><\/pre>\n","answers":[{"AnswerId":"58216763","CreationDate":"2019-10-03T10:03:26.680","ParentId":null,"OwnerUserId":"9084663","Title":null,"Body":"<p>The reason both accuracy and loss drop to 0 in the graph is that at that point the output becomes <code>nan<\/code>. \nThis causes the loss to become <code>nan<\/code> as well, which apparently <code>matplotlib<\/code> regards as 0. Accuracy, in this case, is of course zero.<\/p>\n\n<p>My mistake was using linear rather than softmax activation for the decoding layer. <\/p>\n\n<p>After replacing <\/p>\n\n<pre><code>model.add(Activation('linear'))\n<\/code><\/pre>\n\n<p>with <\/p>\n\n<pre><code>model.add(Activation('softmax')) \n<\/code><\/pre>\n\n<p>the model does not collapses into <code>nan<\/code> but rather improves in accuracy. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vED1J.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/vED1J.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
+{"QuestionId":58121025,"AnswerCount":1,"Tags":"<python><tensorflow><bounding-box><object-detection-api>","CreationDate":"2019-09-26T16:22:22.017","AcceptedAnswerId":58183868.0,"OwnerUserId":5445079.0,"Title":"How to make tensorflow bounding boxes to show up every \"n\" seconds?","Body":"<p>Currently I'm playing with Tensorflow Object Detection Api with my own dataset. I want to \"hide\" detection bounding boxes for every 5 frames. Thus, those bounding boxes will be displayed as \"blinking\" and they will get much more attention while using detection framework.<\/p>\n\n<p>I've already messed up with <em>visualization_utils.py<\/em> and tried to have another method for visualizing bounding boxes for this purpose and used it with a while loop:<\/p>\n\n<pre><code>def draw_bounding_box_every5(image,\n                         ymin,\n                         xmin,\n                         ymax,\n                         xmax,\n                         count,                            \n                         use_normalized_coordinates=True):\n\n\n#while True:\n\n  count+=1\n\n  draw = ImageDraw.Draw(image)\n  im_width, im_height = image.size\n\n  ## with the \"if\" below, I'm aiming to have only 1 bounding box to display from every 5 frames.##\n\n  if(count %5 == 0):\n      if use_normalized_coordinates:\n          (left, right, top, bottom) = (xmin * im_width, xmax * im_width,\n                                        ymin * im_height, ymax * im_height)       \n      else:\n          (left, right, top, bottom) = (xmin, xmax, ymin, ymax)\n      draw.line([(left, top), (left, bottom), (right, bottom),\n                (right, top), (left, top)])\n\n\n\n      print(\"line drawed\")\n<\/code><\/pre>\n\n<p>If I don't use <strong>while True<\/strong> loop at the beginning, Tensorflow will continue to display detected objects as expected. But when I use that, it crashes. I guess creating an infinite loop before displaying bounding boxes disabling some callback functions when drawing.\nIf anyone knows how to make blinking bounding boxes, i'm all ears.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"58183868","CreationDate":"2019-10-01T11:27:23.240","ParentId":null,"OwnerUserId":"5445079","Title":null,"Body":"<p>I gave up on this. Instead of getting rid of bounding boxes, I've just added another color into them. Bounding boxes are now changing color within a fixed time.  <\/p>\n"}]}
+{"QuestionId":58121054,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><keras><svm><tensorflow-serving>","CreationDate":"2019-09-26T16:24:31.510","AcceptedAnswerId":null,"OwnerUserId":11765356.0,"Title":"TensorFlow Keras SavedModel empty variables folder","Body":"<p>I have a TensorFlow keras model that looks like the following:<\/p>\n\n<pre><code>model = tf.keras.Sequential([\n    feature_layer,\n    tf.keras.layers.Dense(64, activation='relu'),\n    tf.keras.layers.Dense(\n        6, kernel_regularizer=l2(0.01), activation=tf.nn.softmax\n    )\n])\nmodel.compile(loss='squared_hinge',\n              optimizer='adadelta',\n              metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>I save it to a format servable with TensorFlow Serving, with the following code and inputs:<\/p>\n\n<pre><code>tf.saved_model.simple_save(\n    keras.backend.get_session(),\n    export_path,\n    inputs=model.inputs,\n    outputs={t.name:t for t in model.outputs})\n\n#model.inputs:\n#{\n# 'int1': &lt;tf.Tensor 'int1_5:0' shape=(?,) dtype=int32&gt;, \n# 'int2': &lt;tf.Tensor 'int2_5:0' shape=(?,) dtype=int32&gt;,\n# 'int3': &lt;tf.Tensor 'int3_5:0' shape=(?,) dtype=int32&gt;,\n# 'string1': &lt;tf.Tensor 'string1_5:0' shape=(?,) dtype=string&gt;\n#}\n\n#model.outputs:\n# [&lt;tf.Tensor 'sequential\/Identity:0' shape=(?, 6) dtype=float32&gt;]\n<\/code><\/pre>\n\n<p>I expect 6 outputs, each cooresponding to a class and their likelyhood. Using <code>model.predict()<\/code>, this seems to work fine: the model returns arrays like this <code>[0.15914398 0.152271   0.18949589 0.14985411 0.17449048 0.1747446 ]<\/code>.<\/p>\n\n<p>However, the SavedModel that I want to use with TensorFlow Serving that is produced by the code has an empty 'variables' folder upon generation, leaving only the saved_model.pb file. When I request the model with the saved_model_cli I get the following:<\/p>\n\n<pre><code>RuntimeError: The Session graph is empty.  Add operations to the graph before calling run().\n<\/code><\/pre>\n\n<p>Why is this the case and how can I get the behaviour I want? \nThanks!<\/p>\n","answers":[]}
+{"QuestionId":58121185,"AnswerCount":1,"Tags":"<tensorflow><cuda><nvidia>","CreationDate":"2019-09-26T16:32:31.747","AcceptedAnswerId":null,"OwnerUserId":4706745.0,"Title":"How can I resolve this error for conda tensorflow installation error?","Body":"<p>How can I resolve this error? <\/p>\n\n<pre><code>&gt;&gt;&gt; import tensorflow as tf\n&gt;&gt;&gt; tf.Session(config=tf.ConfigProto(log_device_placement=True))\n2019-09-26 12:28:37.749941: I tensorflow\/core\/platform\/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA\n2019-09-26 12:28:37.756897: I tensorflow\/core\/platform\/profile_utils\/cpu_utils.cc:94] CPU Frequency: 2993275000 Hz\n2019-09-26 12:28:37.758229: I tensorflow\/compiler\/xla\/service\/service.cc:168] XLA service 0x5636b6113300 executing computations on platform Host. Devices:\n2019-09-26 12:28:37.758293: I tensorflow\/compiler\/xla\/service\/service.cc:175]   StreamExecutor device (0): &lt;undefined&gt;, &lt;undefined&gt;\n2019-09-26 12:28:37.759374: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcuda.so.1\n2019-09-26 12:28:37.782412: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1640] Found device 0 with properties: \nname: Quadro P4000 major: 6 minor: 1 memoryClockRate(GHz): 1.48\npciBusID: 0000:03:00.0\n2019-09-26 12:28:37.782612: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudart.so.10.1\n2019-09-26 12:28:37.784210: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcublas.so.10\n2019-09-26 12:28:37.785800: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcufft.so.10\n2019-09-26 12:28:37.786057: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcurand.so.10\n2019-09-26 12:28:37.787715: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcusolver.so.10\n2019-09-26 12:28:37.788633: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcusparse.so.10\n2019-09-26 12:28:37.792187: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudnn.so.7\n2019-09-26 12:28:37.793278: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1763] Adding visible gpu devices: 0\n2019-09-26 12:28:37.793314: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudart.so.10.1\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/home\/joshij\/yes\/envs\/TF\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 1570, in __init__\n    super(Session, self).__init__(target, graph, config=config)\n  File \"\/home\/joshij\/yes\/envs\/TF\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py\", line 693, in __init__\n    self._session = tf_session.TF_NewSessionRef(self._graph._c_graph, opts)\ntensorflow.python.framework.errors_impl.InternalError: cudaGetDevice() failed. Status: CUDA driver version is insufficient for CUDA runtime version\n&gt;&gt;&gt; \n<\/code><\/pre>\n\n<p>Output for nvidia-smi command<\/p>\n\n<pre><code>Thu Sep 26 12:53:39 2019       \n+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 387.26                 Driver Version: 387.26                    |\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  Quadro P4000        Off  | 00000000:03:00.0  On |                  N\/A |\n| 46%   43C    P0    29W \/ 105W |   1023MiB \/  8080MiB |      0%      Default |\n+-------------------------------+----------------------+----------------------+\n\n+-----------------------------------------------------------------------------+\n| Processes:                                                       GPU Memory |\n|  GPU       PID   Type   Process name                             Usage      |\n|=============================================================================|\n|    0       677      G   ...quest-channel-token=5147126485013694581    54MiB |\n|    0      1900      G   \/usr\/libexec\/Xorg                            214MiB |\n|    0      2182      G   \/usr\/libexec\/Xorg                            331MiB |\n|    0      2416      G   \/usr\/bin\/gnome-shell                         172MiB |\n|    0      3154      G   ...uest-channel-token=13678957064081588319   207MiB |\n+-----------------------------------------------------------------------------+\n<\/code><\/pre>\n","answers":[{"AnswerId":"58121460","CreationDate":"2019-09-26T16:52:02.463","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>Use <code>nvidia-smi<\/code> command to check your driver version. Looks like your nvidia-driver version is kind of out of date. It should be larger than 410 because you are using cuda 10. Try to upgrade your driver version. <\/p>\n"}]}
+{"QuestionId":58121188,"AnswerCount":0,"Tags":"<keras><tensorflow2.0>","CreationDate":"2019-09-26T16:32:50.133","AcceptedAnswerId":null,"OwnerUserId":287238.0,"Title":"In tensorflow 2.0 keras, how can we get gradients for the regularisation parameters like l1 and l2?","Body":"<p>For example if we build a network using<\/p>\n\n<pre><code>tf.keras.layers.Dense(16, kernel_regularizer=tf.keras.regularizers.L1L1(l1=0.1, l2=0.2))\n<\/code><\/pre>\n\n<p>You get entries appearing in the <code>model.losses<\/code> however they do not appear to get recalculated when changing l1, and l2. Furthermore, it seems like the loss function is not differentiable w.r.t l1 and l2 even after replace l1, l2 with<\/p>\n\n<pre><code>l1 = tf.keras.backend.variable(name='l1', value=0.1)\nl2 = tf.keras.backend.variable(name='l2', value=0.1)\n<\/code><\/pre>\n\n<p>The variables <em>are<\/em> trainable via some function that explicitly uses them.<\/p>\n\n<p>UPDATE: <\/p>\n\n<p>I have a feeling it is the <code>_gather_children...<\/code> function that has a non-differentiable boolean in it unfortunately.<\/p>\n\n<pre><code>OperatorNotAllowedInGraphError: in converted code:\n\n&lt;ipython-input-272-b7a1c0374804&gt;:7 train_one  *\n    d = loss(x, y)\n&lt;ipython-input-222-e2bf7d12dcd8&gt;:10 loss  *\n    layer_losses = tf.reduce_mean(agent.predictor.losses)\nasdf\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/keras\/engine\/base_layer.py:1015 losses\n    return collected_losses + self._gather_children_attribute('losses')\nasdf\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/keras\/engine\/base_layer.py:2336 _gather_children_attribute\n    getattr(layer, attribute) for layer in nested_layers))\nasdf\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/keras\/engine\/base_layer.py:2336 &lt;genexpr&gt;\n    getattr(layer, attribute) for layer in nested_layers))\nasdf\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/keras\/engine\/base_layer.py:1012 losses\n    loss_tensor = regularizer()\nasdf\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/keras\/engine\/base_layer.py:1087 _tag_unconditional\n    loss = loss()\nasdf\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/keras\/engine\/base_layer.py:1916 _loss_for_variable\n    regularization = regularizer(v)\nasdf\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/keras\/regularizers.py:57 __call__\n    if not self.l1 and not self.l2:\nasdf\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/ops\/resource_variable_ops.py:450 __bool__\n    return bool(self.read_value())\nasdf\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/framework\/ops.py:765 __bool__\n    self._disallow_bool_casting()\nasdf\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/framework\/ops.py:531 _disallow_bool_casting\n    \"using a `tf.Tensor` as a Python `bool`\")\nasdf\/lib\/python3.7\/site-packages\/tensorflow_core\/python\/framework\/ops.py:518 _disallow_when_autograph_enabled\n    \" decorating it directly with @tf.function.\".format(task))\n\nOperatorNotAllowedInGraphError: using a `tf.Tensor` as a Python `bool` is not allowed: AutoGraph did not convert this function. Try decorating it directly with @tf.function.\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58121654,"AnswerCount":1,"Tags":"<python><arrays><numpy><pytorch>","CreationDate":"2019-09-26T17:06:51.637","AcceptedAnswerId":58121969.0,"OwnerUserId":6642287.0,"Title":"Convert image of dimension height,width,number of channels to n_masks, image_height, image_width","Body":"<p>I have an RGB image, it contains masks of different colors, each color represents a particular class.\nI want to convert it into the format - <code>n_masks, image_height, image_width<\/code> where n_masks is number of masks present in the image. and each slice of the matrix along the 0th axis represents one binary mask.<br>\nSo far, I have been able to convert it into the format of <code>image_height, image_width<\/code> where each array value represents which class it belongs to but I am kind of struck after it.\nBelow is my code to convert it into <code>image_height,image_width<\/code> format-<\/p>\n\n<pre><code>def mask_to_class(mask):\n    target = torch.from_numpy(mask)\n    h,w = target.shape[0],target.shape[1]\n    masks = torch.empty(h, w, dtype=torch.long)\n    colors = torch.unique(target.view(-1,target.size(2)),dim=0).numpy()\n    target = target.permute(2, 0, 1).contiguous()\n    mapping = {tuple(c): t for c, t in zip(colors.tolist(), range(len(colors)))}\n    for k in mapping:\n        idx = (target==torch.tensor(k, dtype=torch.uint8).unsqueeze(1).unsqueeze(2))\n        validx = (idx.sum(0) == 3) \n        masks[validx] = torch.tensor(mapping[k], dtype=torch.long)\n    return masks\n<\/code><\/pre>\n\n<p>It converts an image of let's say format (512,512,3) to (512,512) where each pixel values represents the class it belongs to, but I don't have any idea how to proceed further.<br>\nP.S- I am coding it in <code>pytorch<\/code> but any approach involving <code>numpy<\/code> is also welcomed.<br>\nAttaching the image for representational purpose<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Bqz2O.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Bqz2O.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"58121969","CreationDate":"2019-09-26T17:32:14.817","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>Suppose you already have (512,512) mask. You can first use <code>mask.unique()<\/code> to get all classes pixel values. Then for each class value, <code>torch.where(mask==cls_val, torch.tensor(1), torch.tensor(0))<\/code> will return the mask of one certain class. In the end, you stack all outputs. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>mask = torch.tensor([[1,2,3],\n                     [2,4,5],\n                     [1,2,3]])\ncls = mask.unique()\nres = torch.stack([torch.where(mask==cls_val, torch.tensor(1), torch.tensor(0)) for cls_val in cls])\n\n#tensor([[[1, 0, 0],\n#         [0, 0, 0],\n#         [1, 0, 0]],\n#\n#        [[0, 1, 0],\n#         [1, 0, 0],\n#         [0, 1, 0]],\n#\n#        [[0, 0, 1],\n#         [0, 0, 0],\n#         [0, 0, 1]],\n#\n#        [[0, 0, 0],\n#         [0, 1, 0],\n#         [0, 0, 0]],\n#\n#        [[0, 0, 0],\n#         [0, 0, 1],\n#         [0, 0, 0]]])\n<\/code><\/pre>\n"}]}
+{"QuestionId":58121729,"AnswerCount":0,"Tags":"<python><c++><tensorflow><keras><snpe>","CreationDate":"2019-09-26T17:13:24.833","AcceptedAnswerId":null,"OwnerUserId":4910495.0,"Title":"SNPE - How can I predict on a model with two (or more) inputs?","Body":"<p>I've built a Keras model with two inputs that I'd like to predict on with SNPE on my phone. I've already converted it successfully, it's just the C++ code that I'm having trouble with right now. I'm able to predict on a model with one input with any shape 1D array, but I now have a model that takes two 1D arrays of size 1.<\/p>\n\n<p>So in Keras, predicting looks like this: <code>model.predict([np.array([.4]), np.array([.6])])<\/code><\/p>\n\n<p>And the SNPE code I have to predict:<\/p>\n\n<pre><code>void init_model(){\n  zdl::DlSystem::Runtime_t runt=checkRuntime();\n  initializeSNPE(runt);\n}\n\nfloat run_model(float a, float b){\n  std::vector&lt;float&gt; inputVec;\n  std::vector&lt;float&gt; inputVec2;\n  inputVec.push_back(a);\n  inputVec2.push_back(b);\n  std::unique_ptr&lt;zdl::DlSystem::ITensor&gt; inputTensor = loadInputTensor(snpe, inputVec);\n  std::unique_ptr&lt;zdl::DlSystem::ITensor&gt; inputTensor2 = loadInputTensor(snpe, inputVec2);  \/\/ what do I do with this?\n  zdl::DlSystem::ITensor* oTensor = executeNetwork(snpe, inputTensor);\n  return returnOutput(oTensor);\n}\n<\/code><\/pre>\n\n<p>The functions I'm using are modified from SNPE's website. It works well with my prior uses of predicting on a single array:<\/p>\n\n<pre><code>zdl::DlSystem::Runtime_t checkRuntime()\n{\n    static zdl::DlSystem::Version_t Version = zdl::SNPE::SNPEFactory::getLibraryVersion();\n    static zdl::DlSystem::Runtime_t Runtime;\n    std::cout &lt;&lt; \"SNPE Version: \" &lt;&lt; Version.asString().c_str() &lt;&lt; std::endl; \/\/Print Version number\n    std::cout &lt;&lt; \"\\ntest\";\n    if (zdl::SNPE::SNPEFactory::isRuntimeAvailable(zdl::DlSystem::Runtime_t::GPU)) {\n        Runtime = zdl::DlSystem::Runtime_t::GPU;\n    } else {\n        Runtime = zdl::DlSystem::Runtime_t::CPU;\n    }\n\n    return Runtime;\n}\n\nvoid initializeSNPE(zdl::DlSystem::Runtime_t runtime) {\n  std::unique_ptr&lt;zdl::DlContainer::IDlContainer&gt; container;\n  container = zdl::DlContainer::IDlContainer::open(\"\/path\/to\/model.dlc\");\n  \/\/printf(\"loaded model\\n\");\n  int counter = 0;\n  zdl::SNPE::SNPEBuilder snpeBuilder(container.get());\n  snpe = snpeBuilder.setOutputLayers({})\n                      .setRuntimeProcessor(runtime)\n                      .setUseUserSuppliedBuffers(false)\n                      .setPerformanceProfile(zdl::DlSystem::PerformanceProfile_t::HIGH_PERFORMANCE)\n                      .build();\n}\n\nstd::unique_ptr&lt;zdl::DlSystem::ITensor&gt; loadInputTensor(std::unique_ptr&lt;zdl::SNPE::SNPE&gt; &amp;snpe, std::vector&lt;float&gt; inputVec) {\n  std::unique_ptr&lt;zdl::DlSystem::ITensor&gt; input;\n  const auto &amp;strList_opt = snpe-&gt;getInputTensorNames();\n  if (!strList_opt) throw std::runtime_error(\"Error obtaining Input tensor names\");\n  const auto &amp;strList = *strList_opt;\n\n  const auto &amp;inputDims_opt = snpe-&gt;getInputDimensions(strList.at(0));\n  const auto &amp;inputShape = *inputDims_opt;\n\n  input = zdl::SNPE::SNPEFactory::getTensorFactory().createTensor(inputShape);\n  std::copy(inputVec.begin(), inputVec.end(), input-&gt;begin());\n\n  return input;\n}\n\nfloat returnOutput(const zdl::DlSystem::ITensor* tensor) {\n  float op = *tensor-&gt;cbegin();\n  return op;\n}\n\nzdl::DlSystem::ITensor* executeNetwork(std::unique_ptr&lt;zdl::SNPE::SNPE&gt;&amp; snpe,\n                    std::unique_ptr&lt;zdl::DlSystem::ITensor&gt;&amp; input) {\n  static zdl::DlSystem::TensorMap outputTensorMap;\n  snpe-&gt;execute(input.get(), outputTensorMap);\n  zdl::DlSystem::StringList tensorNames = outputTensorMap.getTensorNames();\n\n  const char* name = tensorNames.at(0);  \/\/ only should the first\n  auto tensorPtr = outputTensorMap.getTensor(name);\n  return tensorPtr;\n}\n<\/code><\/pre>\n\n<p>But I have no clue how to combine the two input tensors I've gotten to use with the <code>executeNetwork<\/code> function. Any help would be appreciated.<\/p>\n","answers":[]}
+{"QuestionId":58121828,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-datasets>","CreationDate":"2019-09-26T17:21:47.760","AcceptedAnswerId":58152174.0,"OwnerUserId":3263841.0,"Title":"How to get a windowed dataset in tensorflow 2 from an array of numpy arrays?","Body":"<p>Imagine I have some data: <\/p>\n\n<pre><code>some_data = np.array([[1,2,3,4], [5, 6, 7,8]])\n<\/code><\/pre>\n\n<p>It looks like this: <\/p>\n\n<pre><code>array([[1, 2, 3, 4],\n       [5, 6, 7, 8]])\n<\/code><\/pre>\n\n<p>Each row represents a different observation, so they should not be combined. I want to create a windowed dataset, each window of size 3, shifted by 1. When I pass a single observation, I get what I want, like this: <\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices(some_data[0])\ndataset = dataset.window(size=3, shift=1, drop_remainder=True)\ndataset = dataset.flat_map(lambda window: window.batch(3))\n<\/code><\/pre>\n\n<p>The result: <\/p>\n\n<pre><code>for x in dataset:\n    print(x.numpy())\n\n[1 2 3]\n[2 3 4]\n<\/code><\/pre>\n\n<p>But when I pass the whole numpy array of arrays, I don't get back anything. <\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices(some_data)\ndataset = dataset.window(size=3, shift=1, drop_remainder=True)\ndataset = dataset.flat_map(lambda window: window.batch(3))\n<\/code><\/pre>\n\n<p>This is what I would expect:<\/p>\n\n<pre><code>for x in dataset:\n    print(x.numpy())\n\n[1 2 3]\n[2 3 4]\n[5 6 7]\n[6 7 8]\n<\/code><\/pre>\n\n<p>I guess I could loop over <code>some_data<\/code> and pass one array at a time, and then concatenate the datasets, but this seems like a bad solution. What's the right way to do it? <\/p>\n\n<p>I'm using Tensorflow 2.0. Thanks! <\/p>\n","answers":[{"AnswerId":"58152174","CreationDate":"2019-09-29T04:11:47.363","ParentId":null,"OwnerUserId":"7389608","Title":null,"Body":"<p>Each row of dataset has only one element when you use <code>dataset = tf.data.Dataset.from_tensor_slices(some_data[0])<\/code>.<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices(some_data[0])\nfor x in dataset:\n    print(x.numpy())\n1\n2\n3\n4\n<\/code><\/pre>\n\n<p>But each row of dataset has four elements when you use <code>dataset = tf.data.Dataset.from_tensor_slices(some_data)<\/code>.<\/p>\n\n<pre><code>dataset = tf.data.Dataset.from_tensor_slices(some_data)\nfor x in dataset:\n    print(x.numpy())\n[1 2 3 4]\n[5 6 7 8]\n<\/code><\/pre>\n\n<p>So what you need to do is convert each row and merge it.<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nsome_data = np.array([[1,2,3,4], [5, 6, 7,8]])\ndataset = tf.data.Dataset.from_tensor_slices(some_data)\n\ndef parse_samples(x):\n    return tf.data.Dataset.from_tensor_slices(x)\\\n        .window(size=3, shift=1, drop_remainder=True)\\\n        .flat_map(lambda window: window.batch(3))\n\ndataset = dataset.flat_map(parse_samples)\n\nfor x in dataset:\n    print(x.numpy())\n\n[1 2 3]\n[2 3 4]\n[5 6 7]\n[6 7 8]\n<\/code><\/pre>\n"}]}
+{"QuestionId":58121936,"AnswerCount":0,"Tags":"<python><tensorflow><distributed-computing>","CreationDate":"2019-09-26T17:29:57.357","AcceptedAnswerId":null,"OwnerUserId":919794.0,"Title":"How to set up \"worker\" machine in tensorflow","Body":"<p>I've been trying to follow the tensorflow tutorial at <a href=\"https:\/\/www.tensorflow.org\/guide\/distribute_strategy\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/distribute_strategy<\/a>   to distribute training a model between two machines. But I'm not sure which code runs on which machine. (In this case, I'm trying to use the MultiWorkerMirroredStrategy.)<\/p>\n\n<p>The code in the tutorial sets up the model within mirrored_strategy.scope(), then runs something like model.fit(). Let's call the machine running that code the \"host\" machine. What code do I run on the OTHER machine?<\/p>\n\n<p>Is there some simple python code or command-line tool to say \"please listen to port 12345 and service any tensorflow requests that come in\"?<\/p>\n","answers":[]}
+{"QuestionId":58122094,"AnswerCount":0,"Tags":"<tensorflow><seq2seq><encoder-decoder><beam-search>","CreationDate":"2019-09-26T17:41:19.497","AcceptedAnswerId":null,"OwnerUserId":10516607.0,"Title":"Can beam search in a seq2seq rnn network be used in non NLP model?","Body":"<p>It seems that <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.14\/tensorflow\/contrib\/seq2seq\/python\/ops\/beam_search_decoder.py#L589-L750\" rel=\"nofollow noreferrer\">tf.contrib.seq2seq.BeamSearchDecoder<\/a> cannot be used in fields other than modeling words, as it needs to input the starting and end tokens, and search the best K (beam width) words in the total dictionary. What if I doo not have a dictionary, but only a input sequence with scalars to put into the rnn encoder and output a sequence with the encoded state? Is tf.contrib.seq2seq.BeamSearchDecoder still useful, or I misunderstand it? Any answers appreciated.<\/p>\n","answers":[]}
+{"QuestionId":58122191,"AnswerCount":0,"Tags":"<python><machine-learning><neural-network><pytorch><lstm>","CreationDate":"2019-09-26T17:48:48.747","AcceptedAnswerId":null,"OwnerUserId":12012482.0,"Title":"Neural net not training at all","Body":"<p>my neural net accuracy and test loss are stuck while building a model on <a href=\"https:\/\/archive.ics.uci.edu\/ml\/datasets\/Human+Activity+Recognition+Using+Smartphones\" rel=\"nofollow noreferrer\">HAR<\/a>. The model was built using only time series variables(expert features were not used)<\/p>\n\n<p>I have tried many learning rates, and optimizers, even then from epoch 1 to 200 they are stuck at 18% accuracy<\/p>\n\n<p>The architecture used was from <a href=\"https:\/\/www.youtube.com\/watch?v=PWMlR8eVgHc&amp;t=983s\" rel=\"nofollow noreferrer\">here<\/a>\nAnd I couldn't figure out if architecture is really bad or my code has a bug in it. <\/p>\n\n<pre><code>class Net(nn.Module):\n\n  def __init__(self):\n    super().__init__()\n    self.model = nn.ModuleDict({\n            'lstm': nn.LSTM(\n                input_size=64,    \n                hidden_size=64,batch_first=True\n            ),\n            'linear': nn.Linear(\n                in_features=128,\n                out_features=6),\n            'dropout': nn.Dropout(0.7),\n            'lstm2': nn.LSTM(\n                input_size=64,    \n                hidden_size=128,batch_first=True),\n            'linear2': nn.Linear(\n                in_features=9,\n                out_features=64)})\n\n  def forward(self,x):\n    x = x.permute(1,0,2)\n    x = x.contiguous().view(-1,9)\n    x = self.model['linear2'](x)\n\n    x = x.contiguous().view(-1,128,64)\n    x, _ = self.model['lstm'](x)\n    x = self.model['dropout'](x)\n    x, _ = self.model['lstm2'](x)\n    x = self.model['dropout'](x)\n    x = self.model['linear'](x[:,-1,:])\n    return x\n\n\nnet=Net().to(device)\n\ndef train( model, device, train_loader, optimizer, epoch,):\n    model.train()\n    train_loss=0\n    for batch_idx, (data, target) in enumerate(train_loader):\n        data, target = data.to(device), target.view(-1).to(device)\n        optimizer.zero_grad()\n\n        output = model(data.float())\n\n        train_loss+=F.cross_entropy(output, target).item()\n        loss = F.cross_entropy(output, target)\n        loss.backward()\n        optimizer.step()\n\n    train_loss \/= np.ceil(len(train_loader.dataset)\/bs)\n    print('Train Epoch: {} \\tLoss per Batch: {:.6f}'.format(epoch,train_loss))\n\ndef test( model, device, test_loader):\n    model.eval()\n    test_loss = 0\n    correct = 0\n    with torch.no_grad():\n        for batch_idx,(data, target) in enumerate(test_loader):\n            data, target = data.to(device), target.to(device)\n            output = model(data.float())\n            test_loss += F.cross_entropy(output, target).item() # sum up batch loss\n            pred = output.argmax(dim=1, keepdim=True) # get the index of the max log-probability\n            correct += pred.eq(target.view_as(pred)).sum().item()\n    test_loss \/= np.ceil(len(test_loader.dataset)\/bs)\n\n\n    print('\\nTest set: Loss per Batch: {:.4f}, Accuracy: {}\/{} ({:.0f}%)\\n'.format(\n        test_loss, correct, len(test_loader.dataset),\n        100. * correct \/ len(test_loader.dataset)))\n\n\nimport torch.optim as optim\noptimizer = optim.Adam(net.parameters(), lr=.01)\n<\/code><\/pre>\n\n<hr>\n\n<p>By using Adam with lr=.01, I got the following results<\/p>\n\n<p>loss per batch and accuracy of Test set have remained constant during the whole training period of 70 epochs<\/p>\n\n<pre><code>Train Epoch: 0 Loss per Batch: 1.788564\n\nTest set: Loss per Batch: 1.7989, Accuracy: 533\/2947 (18%)\n\nTrain Epoch: 1  Loss per Batch: 1.786124\n\nTest set: Loss per Batch: 1.7926, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 2  Loss per Batch: 1.785584\n\nTest set: Loss per Batch: 1.7947, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 3  Loss per Batch: 1.785016\n\nTest set: Loss per Batch: 1.8117, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 4  Loss per Batch: 1.786215\n\nTest set: Loss per Batch: 1.7921, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 5  Loss per Batch: 1.785138\n\nTest set: Loss per Batch: 1.7939, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 6  Loss per Batch: 1.784805\n\nTest set: Loss per Batch: 1.7932, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 7  Loss per Batch: 1.784633\n\nTest set: Loss per Batch: 1.7932, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 8  Loss per Batch: 1.784855\n\nTest set: Loss per Batch: 1.7938, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 9  Loss per Batch: 1.785028\n\nTest set: Loss per Batch: 1.7936, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 10     Loss per Batch: 1.784683\n\nTest set: Loss per Batch: 1.7930, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 11     Loss per Batch: 1.785211\n\nTest set: Loss per Batch: 1.7930, Accuracy: 527\/2947 (18%)\n\nTrain Epoch: 12     Loss per Batch: 1.784648\n\nTest set: Loss per Batch: 1.7912, Accuracy: 514\/2947 (17%)\n\nTrain Epoch: 13     Loss per Batch: 1.784625\n\nTest set: Loss per Batch: 1.7909, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 14     Loss per Batch: 1.785082\n\nTest set: Loss per Batch: 1.7921, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 15     Loss per Batch: 1.785162\n\nTest set: Loss per Batch: 1.7931, Accuracy: 509\/2947 (17%)\n\nTrain Epoch: 16     Loss per Batch: 1.784732\n\nTest set: Loss per Batch: 1.7932, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 17     Loss per Batch: 1.784819\n\nTest set: Loss per Batch: 1.7949, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 18     Loss per Batch: 1.785135\n\nTest set: Loss per Batch: 1.7927, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 19     Loss per Batch: 1.784840\n\nTest set: Loss per Batch: 1.7933, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 20     Loss per Batch: 1.784750\n\nTest set: Loss per Batch: 1.7939, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 21     Loss per Batch: 1.784810\n\nTest set: Loss per Batch: 1.7940, Accuracy: 517\/2947 (18%)\n\nTrain Epoch: 22     Loss per Batch: 1.784939\n\nTest set: Loss per Batch: 1.7872, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 23     Loss per Batch: 1.785099\n\nTest set: Loss per Batch: 1.7926, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 24     Loss per Batch: 1.784571\n\nTest set: Loss per Batch: 1.7918, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 25     Loss per Batch: 1.784607\n\nTest set: Loss per Batch: 1.7922, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 26     Loss per Batch: 1.784505\n\nTest set: Loss per Batch: 1.7902, Accuracy: 514\/2947 (17%)\n\nTrain Epoch: 27     Loss per Batch: 1.785079\n\nTest set: Loss per Batch: 1.7932, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 28     Loss per Batch: 1.784903\n\nTest set: Loss per Batch: 1.7933, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 29     Loss per Batch: 1.784797\n\nTest set: Loss per Batch: 1.7938, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 30     Loss per Batch: 1.784668\n\nTest set: Loss per Batch: 1.7940, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 31     Loss per Batch: 1.784762\n\nTest set: Loss per Batch: 1.7934, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 32     Loss per Batch: 1.784556\n\nTest set: Loss per Batch: 1.7947, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 33     Loss per Batch: 1.784983\n\nTest set: Loss per Batch: 1.7935, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 34     Loss per Batch: 1.785202\n\nTest set: Loss per Batch: 1.7945, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 35     Loss per Batch: 1.785080\n\nTest set: Loss per Batch: 1.7921, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 36     Loss per Batch: 1.784834\n\nTest set: Loss per Batch: 1.7925, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 37     Loss per Batch: 1.784711\n\nTest set: Loss per Batch: 1.7938, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 38     Loss per Batch: 1.784832\n\nTest set: Loss per Batch: 1.7922, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 39     Loss per Batch: 1.784580\n\nTest set: Loss per Batch: 1.7931, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 40     Loss per Batch: 1.784810\n\nTest set: Loss per Batch: 1.7928, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 41     Loss per Batch: 1.784549\n\nTest set: Loss per Batch: 1.7934, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 42     Loss per Batch: 1.784458\n\nTest set: Loss per Batch: 1.7920, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 43     Loss per Batch: 1.785136\n\nTest set: Loss per Batch: 1.7974, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 44     Loss per Batch: 1.784797\n\nTest set: Loss per Batch: 1.7762, Accuracy: 548\/2947 (19%)\n\nTrain Epoch: 45     Loss per Batch: 1.785898\n\nTest set: Loss per Batch: 1.7944, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 46     Loss per Batch: 1.784899\n\nTest set: Loss per Batch: 1.7951, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 47     Loss per Batch: 1.784710\n\nTest set: Loss per Batch: 1.7934, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 48     Loss per Batch: 1.785024\n\nTest set: Loss per Batch: 1.7946, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 49     Loss per Batch: 1.784930\n\nTest set: Loss per Batch: 1.7926, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 50     Loss per Batch: 1.784564\n\nTest set: Loss per Batch: 1.7913, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 51     Loss per Batch: 1.784940\n\nTest set: Loss per Batch: 1.7936, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 52     Loss per Batch: 1.785000\n\nTest set: Loss per Batch: 1.7921, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 53     Loss per Batch: 1.784505\n\nTest set: Loss per Batch: 1.7947, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 54     Loss per Batch: 1.784886\n\nTest set: Loss per Batch: 1.7914, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 55     Loss per Batch: 1.784894\n\nTest set: Loss per Batch: 1.7920, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 56     Loss per Batch: 1.784917\n\nTest set: Loss per Batch: 1.7920, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 57     Loss per Batch: 1.784813\n\nTest set: Loss per Batch: 1.7941, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 58     Loss per Batch: 1.784905\n\nTest set: Loss per Batch: 1.7933, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 59     Loss per Batch: 1.784620\n\nTest set: Loss per Batch: 1.7914, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 60     Loss per Batch: 1.784722\n\nTest set: Loss per Batch: 1.7934, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 61     Loss per Batch: 1.784789\n\nTest set: Loss per Batch: 1.7929, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 62     Loss per Batch: 1.784626\n\nTest set: Loss per Batch: 1.7933, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 63     Loss per Batch: 1.784967\n\nTest set: Loss per Batch: 1.7932, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 64     Loss per Batch: 1.784695\n\nTest set: Loss per Batch: 1.7918, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 65     Loss per Batch: 1.784872\n\nTest set: Loss per Batch: 1.7939, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 66     Loss per Batch: 1.784757\n\nTest set: Loss per Batch: 1.7925, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 67     Loss per Batch: 1.784619\n\nTest set: Loss per Batch: 1.7912, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 68     Loss per Batch: 1.784919\n\nTest set: Loss per Batch: 1.7927, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 69     Loss per Batch: 1.784839\n\nTest set: Loss per Batch: 1.7927, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 70     Loss per Batch: 1.784806\n\nTest set: Loss per Batch: 1.7937, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 71     Loss per Batch: 1.784729\n\nTest set: Loss per Batch: 1.7910, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 72     Loss per Batch: 1.784705\n\nTest set: Loss per Batch: 1.7937, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 73     Loss per Batch: 1.785041\n\nTest set: Loss per Batch: 1.7936, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 74     Loss per Batch: 1.784568\n\nTest set: Loss per Batch: 1.7919, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 75     Loss per Batch: 1.784705\n\nTest set: Loss per Batch: 1.7946, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 76     Loss per Batch: 1.784397\n\nTest set: Loss per Batch: 1.7896, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 77     Loss per Batch: 1.785445\n\nTest set: Loss per Batch: 1.7932, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 78     Loss per Batch: 1.785322\n\nTest set: Loss per Batch: 1.7935, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 79     Loss per Batch: 1.784972\n\nTest set: Loss per Batch: 1.7929, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 80     Loss per Batch: 1.785531\n\nTest set: Loss per Batch: 1.7935, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 81     Loss per Batch: 1.784897\n\nTest set: Loss per Batch: 1.7930, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 82     Loss per Batch: 1.784814\n\nTest set: Loss per Batch: 1.7935, Accuracy: 536\/2947 (18%)\n\nTrain Epoch: 83     Loss per Batch: 1.784614\n\nTest set: Loss per Batch: 1.7960, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 84     Loss per Batch: 1.785118\n\nTest set: Loss per Batch: 1.7930, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 85     Loss per Batch: 1.785127\n\nTest set: Loss per Batch: 1.7919, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 86     Loss per Batch: 1.784900\n\nTest set: Loss per Batch: 1.7934, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 87     Loss per Batch: 1.784775\n\nTest set: Loss per Batch: 1.7954, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 88     Loss per Batch: 1.784831\n\nTest set: Loss per Batch: 1.7931, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 89     Loss per Batch: 1.784919\n\nTest set: Loss per Batch: 1.7937, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 90     Loss per Batch: 1.784544\n\nTest set: Loss per Batch: 1.7935, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 91     Loss per Batch: 1.784984\n\nTest set: Loss per Batch: 1.7938, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 92     Loss per Batch: 1.784681\n\nTest set: Loss per Batch: 1.7919, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 93     Loss per Batch: 1.784792\n\nTest set: Loss per Batch: 1.7928, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 94     Loss per Batch: 1.785022\n\nTest set: Loss per Batch: 1.7918, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 95     Loss per Batch: 1.785049\n\nTest set: Loss per Batch: 1.7926, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 96     Loss per Batch: 1.784717\n\nTest set: Loss per Batch: 1.7946, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 97     Loss per Batch: 1.784922\n\nTest set: Loss per Batch: 1.7912, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 98     Loss per Batch: 1.784669\n\nTest set: Loss per Batch: 1.7917, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 99     Loss per Batch: 1.784664\n\nTest set: Loss per Batch: 1.7927, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 100    Loss per Batch: 1.784869\n\nTest set: Loss per Batch: 1.7937, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 101    Loss per Batch: 1.784915\n\nTest set: Loss per Batch: 1.7954, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 102    Loss per Batch: 1.784592\n\nTest set: Loss per Batch: 1.7913, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 103    Loss per Batch: 1.784817\n\nTest set: Loss per Batch: 1.7920, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 104    Loss per Batch: 1.784869\n\nTest set: Loss per Batch: 1.7941, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 105    Loss per Batch: 1.787049\n\nTest set: Loss per Batch: 1.7915, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 106    Loss per Batch: 1.785012\n\nTest set: Loss per Batch: 1.7919, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 107    Loss per Batch: 1.784757\n\nTest set: Loss per Batch: 1.7927, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 108    Loss per Batch: 1.784881\n\nTest set: Loss per Batch: 1.7939, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 109    Loss per Batch: 1.785723\n\nTest set: Loss per Batch: 1.7926, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 110    Loss per Batch: 1.784988\n\nTest set: Loss per Batch: 1.7914, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 111    Loss per Batch: 1.785197\n\nTest set: Loss per Batch: 1.7928, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 112    Loss per Batch: 1.784856\n\nTest set: Loss per Batch: 1.7937, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 113    Loss per Batch: 1.785120\n\nTest set: Loss per Batch: 1.7932, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 114    Loss per Batch: 1.784842\n\nTest set: Loss per Batch: 1.7906, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 115    Loss per Batch: 1.785119\n\nTest set: Loss per Batch: 1.7918, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 116    Loss per Batch: 1.784935\n\nTest set: Loss per Batch: 1.7917, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 117    Loss per Batch: 1.784734\n\nTest set: Loss per Batch: 1.7931, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 118    Loss per Batch: 1.784930\n\nTest set: Loss per Batch: 1.7916, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 119    Loss per Batch: 1.785058\n\nTest set: Loss per Batch: 1.7931, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 120    Loss per Batch: 1.784399\n\nTest set: Loss per Batch: 1.7955, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 121    Loss per Batch: 1.784580\n\nTest set: Loss per Batch: 1.7921, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 122    Loss per Batch: 1.784730\n\nTest set: Loss per Batch: 1.7926, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 123    Loss per Batch: 1.784921\n\nTest set: Loss per Batch: 1.7946, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 124    Loss per Batch: 1.785217\n\nTest set: Loss per Batch: 1.7918, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 125    Loss per Batch: 1.784670\n\nTest set: Loss per Batch: 1.7926, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 126    Loss per Batch: 1.784825\n\nTest set: Loss per Batch: 1.7935, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 127    Loss per Batch: 1.785658\n\nTest set: Loss per Batch: 1.7926, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 128    Loss per Batch: 1.785025\n\nTest set: Loss per Batch: 1.7924, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 129    Loss per Batch: 1.784739\n\nTest set: Loss per Batch: 1.7949, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 130    Loss per Batch: 1.785050\n\nTest set: Loss per Batch: 1.7932, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 131    Loss per Batch: 1.784816\n\nTest set: Loss per Batch: 1.7909, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 132    Loss per Batch: 1.784678\n\nTest set: Loss per Batch: 1.7943, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 133    Loss per Batch: 1.784638\n\nTest set: Loss per Batch: 1.7934, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 134    Loss per Batch: 1.784962\n\nTest set: Loss per Batch: 1.7941, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 135    Loss per Batch: 1.784798\n\nTest set: Loss per Batch: 1.7941, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 136    Loss per Batch: 1.784776\n\nTest set: Loss per Batch: 1.7932, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 137    Loss per Batch: 1.785209\n\nTest set: Loss per Batch: 1.7908, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 138    Loss per Batch: 1.784884\n\nTest set: Loss per Batch: 1.7926, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 139    Loss per Batch: 1.784842\n\nTest set: Loss per Batch: 1.7931, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 140    Loss per Batch: 1.784622\n\nTest set: Loss per Batch: 1.7919, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 141    Loss per Batch: 1.785023\n\nTest set: Loss per Batch: 1.7925, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 142    Loss per Batch: 1.784689\n\nTest set: Loss per Batch: 1.7944, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 143    Loss per Batch: 1.784810\n\nTest set: Loss per Batch: 1.7931, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 144    Loss per Batch: 1.784905\n\nTest set: Loss per Batch: 1.7920, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 145    Loss per Batch: 1.784574\n\nTest set: Loss per Batch: 1.7946, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 146    Loss per Batch: 1.784829\n\nTest set: Loss per Batch: 1.7930, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 147    Loss per Batch: 1.784832\n\nTest set: Loss per Batch: 1.7933, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 148    Loss per Batch: 1.784582\n\nTest set: Loss per Batch: 1.7941, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 149    Loss per Batch: 1.784732\n\nTest set: Loss per Batch: 1.7926, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 150    Loss per Batch: 1.784652\n\nTest set: Loss per Batch: 1.7947, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 151    Loss per Batch: 1.784722\n\nTest set: Loss per Batch: 1.7926, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 152    Loss per Batch: 1.784994\n\nTest set: Loss per Batch: 1.7951, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 153    Loss per Batch: 1.784773\n\nTest set: Loss per Batch: 1.7938, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 154    Loss per Batch: 1.784962\n\nTest set: Loss per Batch: 1.7934, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 155    Loss per Batch: 1.784522\n\nTest set: Loss per Batch: 1.7934, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 156    Loss per Batch: 1.784805\n\nTest set: Loss per Batch: 1.7921, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 157    Loss per Batch: 1.785455\n\nTest set: Loss per Batch: 1.7944, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 158    Loss per Batch: 1.784509\n\nTest set: Loss per Batch: 1.7943, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 159    Loss per Batch: 1.784982\n\nTest set: Loss per Batch: 1.7926, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 160    Loss per Batch: 1.784593\n\nTest set: Loss per Batch: 1.7938, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 161    Loss per Batch: 1.785101\n\nTest set: Loss per Batch: 1.7930, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 162    Loss per Batch: 1.785277\n\nTest set: Loss per Batch: 1.7939, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 163    Loss per Batch: 1.784970\n\nTest set: Loss per Batch: 1.7935, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 164    Loss per Batch: 1.784984\n\nTest set: Loss per Batch: 1.7943, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 165    Loss per Batch: 1.784928\n\nTest set: Loss per Batch: 1.7931, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 166    Loss per Batch: 1.784774\n\nTest set: Loss per Batch: 1.7923, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 167    Loss per Batch: 1.785107\n\nTest set: Loss per Batch: 1.7935, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 168    Loss per Batch: 1.784620\n\nTest set: Loss per Batch: 1.7941, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 169    Loss per Batch: 1.784884\n\nTest set: Loss per Batch: 1.7931, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 170    Loss per Batch: 1.784452\n\nTest set: Loss per Batch: 1.7923, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 171    Loss per Batch: 1.784965\n\nTest set: Loss per Batch: 1.7935, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 172    Loss per Batch: 1.784781\n\nTest set: Loss per Batch: 1.7924, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 173    Loss per Batch: 1.784724\n\nTest set: Loss per Batch: 1.7941, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 174    Loss per Batch: 1.784639\n\nTest set: Loss per Batch: 1.7912, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 175    Loss per Batch: 1.784592\n\nTest set: Loss per Batch: 1.7936, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 176    Loss per Batch: 1.784642\n\nTest set: Loss per Batch: 1.7922, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 177    Loss per Batch: 1.785098\n\nTest set: Loss per Batch: 1.7931, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 178    Loss per Batch: 1.784758\n\nTest set: Loss per Batch: 1.7938, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 179    Loss per Batch: 1.784484\n\nTest set: Loss per Batch: 1.7910, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 180    Loss per Batch: 1.784845\n\nTest set: Loss per Batch: 1.7949, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 181    Loss per Batch: 1.784970\n\nTest set: Loss per Batch: 1.7926, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 182    Loss per Batch: 1.785243\n\nTest set: Loss per Batch: 1.7917, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 183    Loss per Batch: 1.784563\n\nTest set: Loss per Batch: 1.7932, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 184    Loss per Batch: 1.784880\n\nTest set: Loss per Batch: 1.7936, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 185    Loss per Batch: 1.784850\n\nTest set: Loss per Batch: 1.7961, Accuracy: 537\/2947 (18%)\n\nTrain Epoch: 186    Loss per Batch: 1.784834\n\nTest set: Loss per Batch: 1.7921, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 187    Loss per Batch: 1.784880\n\nTest set: Loss per Batch: 1.7917, Accuracy: 532\/2947 (18%)\n\nTrain Epoch: 188    Loss per Batch: 1.784855\n\nTest set: Loss per Batch: 1.7919, Accuracy: 532\/2947 (18%)\n<\/code><\/pre>\n\n<p>edit1:\nNow i have tried to check if params are updating at all as suggested by Ahmad.<\/p>\n\n<pre><code>       x1=net.model.linear2.weight.data # weights at 150 epoch\n       x3=net.model.linear2.weight.data # weights at 152 epoch\n       x1-x3\ntensor([[0., 0., 0., 0., 0., 0., 0., 0., 0.],\n        [0., 0., 0., 0., 0., 0., 0., 0., 0.],\n        [0., 0., 0., 0., 0., 0., 0., 0., 0.],\n        [0., 0., 0., 0., 0., 0., 0., 0., 0.],\n        [0., 0., 0., 0., 0., 0., 0., 0., 0.],\n        [0., 0., 0., 0., 0., 0., 0., 0., 0.],\n        [0., 0., 0., 0., 0., 0., 0., 0., 0.],\n        [0., 0., 0., 0., 0., 0., 0., 0., 0.],\n        [0., 0., 0., 0., 0., 0., 0., 0., 0.],\n        [0., 0., 0., 0., 0., 0., 0., 0., 0.],\n        [0., 0., 0., 0., 0., 0., 0., 0., 0.],\n        [0., 0., 0., 0., 0., 0., 0., 0., 0.]], device='cuda:0')```\n\nSo,now i tried to check grad at 152 epoch\nnet.model.linear2.weight.grad\n\ntensor([[-5.2324e-14, -1.8329e-14, -4.5901e-15,  3.9225e-14, -4.8923e-14,\n         -3.7663e-14, -6.6461e-14, -1.3109e-12, -1.4465e-12],\n        [ 3.0959e-14,  2.7681e-14, -1.2809e-14, -2.1927e-14,  4.5212e-14,\n          3.6410e-14, -1.3647e-13, -1.2680e-12, -1.4908e-12],\n        [-7.7439e-15, -1.8330e-14,  1.2772e-14,  4.4520e-15, -2.0769e-14,\n         -1.6550e-14,  1.3497e-13,  1.6057e-12,  1.8513e-12],\n        [-7.0256e-15, -3.3731e-14,  2.6417e-14,  1.2869e-14, -4.3155e-14,\n         -3.2898e-14,  3.1167e-13,  3.5165e-12,  4.0472e-12],\n        [ 4.1122e-14,  1.7677e-14,  4.3402e-16, -2.3387e-14,  3.5274e-14,\n          2.8663e-14,  3.6638e-14,  6.8818e-13,  7.4608e-13],\n        [-7.3279e-14, -3.9648e-14,  6.5658e-15,  4.4829e-14, -7.3947e-14,\n         -5.9679e-14,  2.8081e-14, -2.1435e-13, -1.6659e-13],\n        [ 3.9702e-15,  1.8115e-14, -1.4382e-14, -1.4078e-15,  1.8547e-14,\n          1.5202e-14, -1.5490e-13, -1.8519e-12, -2.1300e-12],\n        [ 1.9415e-14,  1.8774e-14, -8.1929e-15, -1.9029e-14,  3.2933e-14,\n          2.4555e-14, -9.4528e-14, -8.9978e-13, -1.0555e-12],\n        [-9.5713e-15, -3.9856e-14,  3.1216e-14,  1.8188e-14, -5.4189e-14,\n         -4.1472e-14,  3.7540e-13,  4.1245e-12,  4.7497e-12],\n        [ 6.3520e-15, -6.6954e-15,  7.5276e-15, -1.1499e-16, -5.6589e-15,\n         -3.1609e-15,  9.2859e-14,  1.1613e-12,  1.3267e-12],\n        [ 2.0287e-14,  1.1300e-14, -2.4096e-15, -1.6815e-14,  2.5231e-14,\n          1.9853e-14, -2.9268e-14, -4.7256e-14, -7.8026e-14],\n        [-7.6004e-14, -4.6867e-14,  1.1355e-14,  5.1983e-14, -8.6401e-14,\n         -6.8230e-14,  9.0062e-14,  4.3427e-13,  5.8239e-13],\n        [ 3.0111e-14,  3.5032e-14, -1.8398e-14, -2.6007e-14,  5.4304e-14,\n          4.1803e-14, -2.0308e-13, -2.1058e-12, -2.4519e12]],device='cuda:0')\n\nNow i know that this is the case of vanishing gradient.\nHow should i proceed to remedy vanishing gradients for this architecture??\n\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58122198,"AnswerCount":1,"Tags":"<keras><neural-network>","CreationDate":"2019-09-26T17:49:13.810","AcceptedAnswerId":58122656.0,"OwnerUserId":11861082.0,"Title":"Why MSE from Keras is not the same as what I calculate?","Body":"<p>I use the following code:<\/p>\n\n<pre><code>import numpy as np\nimport math\nimport keras\nfrom keras.models import Model, Sequential\nfrom keras.layers import Input, Dense, Activation\nfrom keras import regularizers\nfrom keras import backend as K\n\na=1\n\ndef my_regularizer(inputs):\n    means=K.mean((inputs),axis=1)\n    return a*K.sum(means)**2\n\nx_train=np.random.uniform(low=-1,high=1,size=(200,2))\nx_test=np.random.uniform(low=-1,high=1,size=(20,2))\nmodel=Sequential([\n     Dense(20,input_shape=(2,),activity_regularizer=my_regularizer),\n     Activation('tanh'),\n     Dense(2,),\n     Activation('linear')\n])\n\nmodel.compile(optimizer='adam',loss='mean_squared_error')\nhist=model.fit(x_train,x_train,epochs=10,verbose=1,validation_data=(x_test,x_test))\n\nprint('MSE from Keras: ',hist.history['val_loss'][-1])\ny_pred=model.predict(x_test)\nprint('Calculated MSE: ', np.mean((y_pred-x_test)**2))\n<\/code><\/pre>\n\n<p>The output is:<\/p>\n\n<pre><code>MSE from Keras:  0.1555381715297699\nCalculated MSE:  0.12031101597786406\n<\/code><\/pre>\n\n<p>If I remove <code>activity_regularizer=my_regularizer<\/code>, then they will be closer, but still different:<\/p>\n\n<pre><code>MSE from Keras:  0.09773887693881989\nCalculated MSE:  0.09773887699599623\n<\/code><\/pre>\n","answers":[{"AnswerId":"58122656","CreationDate":"2019-09-26T18:21:34.237","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Well, the answer is clear. You have a regularizer. \nThe role of the regularizer is to add a term to the loss function, so it's the expected behavior to have a greater loss. <\/p>\n\n<p>For the other little difference, it's just precision. Maybe using float 32 vs float 64, or doing calculations on the GPU x CPU with different algorythms. I would not worry about that difference. <\/p>\n"}]}
+{"QuestionId":58122223,"AnswerCount":2,"Tags":"<numpy><machine-learning><pytorch><feature-extraction><torch>","CreationDate":"2019-09-26T17:50:53.357","AcceptedAnswerId":null,"OwnerUserId":6894995.0,"Title":"rebuild torch tensor from its batchified version","Body":"<p>This is very nice example of how to build a 3D tensor:<\/p>\n\n<pre><code>import torch\ny = torch.rand(100, 1)\nbatch_size = 10\nbatched_data = y.contiguous().view(batch_size, -1, y.size(-1)).transpose(0,1)\nbatched_data.shape\n<\/code><\/pre>\n\n<p>the output is:<\/p>\n\n<pre><code>torch.Size([10, 10, 1])\n<\/code><\/pre>\n\n<p>ok, now what I want to do is, starting from batched_data I want to build y.\nThe other way around.\nAny good suggestion with a powerful pytorch streamlined code?<\/p>\n\n<p>==== Additional input =====<\/p>\n\n<p>I am using this for RNN and now I have some doubts, becaus eif you consider the following code:<\/p>\n\n<pre><code>import torch\ny = torch.arange(100).view(100,1)\nbatch_size = 10\nbatched_data = y.contiguous().view(batch_size, -1, y.size(-1)).transpose(0,1)\nbatched_data.shape\n<\/code><\/pre>\n\n<p>The output is:<\/p>\n\n<pre><code>tensor([[[ 0],\n         [10],\n         [20],\n         [30],\n         [40],\n         [50],\n         [60],\n         [70],\n         [80],\n         [90]],\n\n        [[ 1],\n         [11],\n         [21],\n         [31],\n         [41],\n         [51],\n         [61],\n         [71],\n         [81],\n         [91]],\n<\/code><\/pre>\n\n<p>Which I would not expect. I would expect something like: \n<code>[[1,2,3,4,5,6,7,8,9,10],[11,12,13,14,15,16,17,18,19,20],....<\/code><\/p>\n","answers":[{"AnswerId":"58124171","CreationDate":"2019-09-26T20:16:58.700","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p>If you want to prepare inputs for RNN, you need to know that <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html?highlight=rnn#torch.nn.RNN\" rel=\"nofollow noreferrer\">RNN<\/a> takes 3d tensors of shape, <code>seq_len, batch, input_size<\/code>. Here, <code>input_size<\/code> refers to the number of features and in your scenario, it is 1. So, the input tensor of shape <code>10, 10, 1<\/code> can still be a valid input for an RNN.<\/p>\n\n<p><strong>Example<\/strong><\/p>\n\n<pre><code>rnn = nn.RNN(input_size=1, hidden_size=20, num_layers=1)\ninput = torch.randn(10, 10, 1)\noutput, hn = rnn(input)\nprint(output.size()) # 10, 10, 20\n<\/code><\/pre>\n\n<p>RNN's output is of shape, <code>seq_len, batch, num_directions * hidden_size<\/code>.<\/p>\n"},{"AnswerId":"58122892","CreationDate":"2019-09-26T18:37:49.107","ParentId":null,"OwnerUserId":"12054165","Title":null,"Body":"<p>Suppose you want to do something like this to rebuild y:<\/p>\n\n<pre><code>rebuilded_y = batched_data.transpose(0,1).view(*y.shape)\n<\/code><\/pre>\n\n<p>To make the input look like you expected you need to remove transpose and additional dimension in batched_data:<\/p>\n\n<pre><code>batched_data = y.contiguous().view(batch_size, -1)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58122505,"AnswerCount":2,"Tags":"<python><neural-network><pytorch><softmax><cross-entropy>","CreationDate":"2019-09-26T18:11:18.613","AcceptedAnswerId":58123713.0,"OwnerUserId":7648650.0,"Title":"Suppress use of Softmax in CrossEntropyLoss for PyTorch Neural Net","Body":"<p>I know theres no need to use a <code>nn.Softmax()<\/code> Function in the output layer for a neural net when using <code>nn.CrossEntropyLoss<\/code> as a loss function.<\/p>\n\n<p>However I need to do so, is there a way to suppress the implemented use of softmax in <code>nn.CrossEntropyLoss<\/code> and instead use <code>nn.Softmax()<\/code> on my output layer of the neural network itself?<\/p>\n\n<p><strong>Motivation<\/strong>: I am using <code>shap<\/code> package to analyze the features influences afterwards, where I can only feed my trained model as an input. The outputs however don't make any sense then because I am looking at unbound values instead of probabilites. <\/p>\n\n<p><strong>Example<\/strong>: Instead of -69.36 as an output value for one class of my model, I want something between 0 and 1, summing up to 1 for all classes. As I can't alter it afterwards, the outputs need to be like this already during training.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/7OSyU.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/7OSyU.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"58123088","CreationDate":"2019-09-26T18:54:17.900","ParentId":null,"OwnerUserId":"6418218","Title":null,"Body":"<p>You can use <code>nn.NLLLoss()<\/code>. \n<code>nn.CrossEntropyLoss<\/code> computes log softmax of the input scores and computes the negative log-likelihood loss. If you already have log probabilities, you can just use <code>nn.NLLLoss()<\/code>. <\/p>\n\n<p><a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html\" rel=\"nofollow noreferrer\">Here<\/a> is an example from PyTorchs documentation <\/p>\n\n<pre><code>m = nn.LogSoftmax(dim=1)\nloss = nn.NLLLoss()\ninput = torch.randn(3, 5, requires_grad=True)\ntarget = torch.tensor([1, 0, 4])\noutput = loss(m(input), target)\n<\/code><\/pre>\n"},{"AnswerId":"58123713","CreationDate":"2019-09-26T19:42:30.437","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p>The documentation of <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html?highlight=crossentropyloss#torch.nn.CrossEntropyLoss\" rel=\"nofollow noreferrer\">nn.CrossEntropyLoss<\/a> says, <\/p>\n\n<blockquote>\n  <p>This criterion combines <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html?highlight=logsoftmax#torch.nn.LogSoftmax\" rel=\"nofollow noreferrer\">nn.LogSoftmax()<\/a> and <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html?highlight=nllloss#torch.nn.NLLLoss\" rel=\"nofollow noreferrer\">nn.NLLLoss()<\/a> in one single class.<\/p>\n<\/blockquote>\n\n<p>I suggest you stick to the use of <code>CrossEntropyLoss<\/code> as the loss criterion. However, you can convert the output of your model into probability values by using the <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.functional.html?highlight=softmax#torch.nn.functional.softmax\" rel=\"nofollow noreferrer\">softmax<\/a> function. <\/p>\n\n<p><strong>Please note<\/strong>, you can always play with the output values of your model, you do not need to change the loss criterion for that.<\/p>\n\n<p>But if you still want to use <code>Softmax()<\/code> in your network, then you can use the <code>NLLLoss()<\/code> as the loss criterion, only apply <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html?highlight=log#torch.log\" rel=\"nofollow noreferrer\">log()<\/a> before feeding model's output to the criterion function. Similarly, if you use <code>LogSoftmax<\/code> instead in your network, you can apply <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html?highlight=exp#torch.exp\" rel=\"nofollow noreferrer\">exp()<\/a> to get the probability values.<\/p>\n\n<p><strong>Update<\/strong>:<\/p>\n\n<p>To use <code>log()<\/code> on the <code>Softmax<\/code> output, please do:<\/p>\n\n<pre><code>torch.log(prob_scores + 1e-20)\n<\/code><\/pre>\n\n<p>By adding a very small number (1e-20) to the <code>prob_scores<\/code>, we can avoid the <code>log(0)<\/code> issue.<\/p>\n"}]}
+{"QuestionId":58122557,"AnswerCount":2,"Tags":"<python><keras><neural-network>","CreationDate":"2019-09-26T18:14:38.633","AcceptedAnswerId":null,"OwnerUserId":5601592.0,"Title":"Customized derivative for Keras when back propagation?","Body":"<p>This is related to <a href=\"https:\/\/stackoverflow.com\/questions\/43915482\/how-do-you-create-a-custom-activation-function-with-keras\">How do you create a custom activation function with Keras?<\/a><\/p>\n\n<p>I have implemented my own cost function<\/p>\n\n<pre><code>import numpy as np\nimport math\nimport keras\nfrom keras.models import Model, Sequential\nfrom keras.layers import Input, Dense, Activation\nfrom keras import regularizers\nfrom keras import backend as K\n\ndef custom_activation(x):\n    return (K.sigmoid(x) *2-1 ) \n\nx_train=np.random.uniform(low=-1,high=1,size=(200,2))\n\nmodel=Sequential([\n     Dense(20,input_shape=(2,)),\n     Activation(custom_activation),\n     Dense(2,),\n     Activation('linear')\n])\n\nmodel.compile(optimizer='adam',loss='mean_squared_error')\nmodel.fit(x_train,x_train,epochs=20,validation_split=0.1)\n<\/code><\/pre>\n\n<p>Instead of letting Keras automatically take the derivative of my approximation function, can I give it the derivative? <\/p>\n\n<p>Please note that this is only an example. My real custom_activation is way more complicated.<\/p>\n","answers":[{"AnswerId":"58124560","CreationDate":"2019-09-26T20:47:53.003","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Use the <code>@tf.custom_gradient<\/code> decorator on your function and define a <code>grad(dy)<\/code> function inside it to be returned:<\/p>\n\n<pre><code>#works only with tensorflow\nfrom keras.backend import tf\n\n@tf.custom_gradient\ndef custom_activation(x):\n    #... do things ...\n\n    def grad(dy):\n        #... do things ...\n        return dy * the_derivative(x)\n\n    #... do things ...\n\n    return result, grad #return the result and the gradient function\n<\/code><\/pre>\n\n<p>Adapted from: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/custom_gradient\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/custom_gradient<\/a><\/p>\n\n<hr>\n\n<p>I never used this in Keras, but if it doesn't work instantly, you can try a hack of putting this function inside a standard Keras function:<\/p>\n\n<pre><code>from keras.layers import Lambda\n\nlayer = Lambda(lambda x: custom_activation(x))\n<\/code><\/pre>\n"},{"AnswerId":"58124677","CreationDate":"2019-09-26T20:56:30.053","ParentId":null,"OwnerUserId":"12118774","Title":null,"Body":"<p>I guess your problem is pretty similar with this problem <a href=\"https:\/\/stackoverflow.com\/questions\/51754639\/how-to-define-the-derivative-of-a-custom-activation-function-in-keras\">how-to-define-the-derivative-of-a-custom-activation-function-in-keras<\/a>. Accepted answer in that link is self-explanatory and useful information to utilize.<\/p>\n\n<p>To be short, you can check <a href=\"https:\/\/www.tensorflow.org\/guide\/extend\/op\" rel=\"nofollow noreferrer\">TF Add New Op.<\/a>.<\/p>\n"}]}
+{"QuestionId":58122591,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><binary-reproducibility>","CreationDate":"2019-09-26T18:17:08.003","AcceptedAnswerId":null,"OwnerUserId":6826845.0,"Title":"Why can't this code get me reproducibility from my Keras LSTM?","Body":"<p>I have tried using the various resources here on StackExchange and those provided by Keras to obtain \nreproducible results from my LSTM. I am currently doing the following (and failing) to ensure reproducibility, in this order:<\/p>\n\n<ul>\n<li><p>I set the PYTHONHASHSEED (via a magic, as you will see I am working in a Jupyter notebook).<\/p><\/li>\n<li><p>I ensure that I am only using CPU and not GPU.<\/p><\/li>\n<li><p>I set random seeds for the python libraries in which any randomization might be introduced- (numpy, python random module, tensorflow).<\/p><\/li>\n<li><p>I force TensorFlow to use a single thread.<\/p><\/li>\n<li><p>While you may notice RandomizedSearchCV and TimeSeriesSplit are imported in the below code, I am not using them in the generation of my results, as it occurred to me that this might introduce other randomization pitfalls. Instead, I am running one training and one test set for my data, broken up by hardcoded indices, and training for only one epoch to make isolating the source of this unwanted variability easier.<\/p><\/li>\n<\/ul>\n\n<p>Instituting the above controls has really constricted the variability for the better, but despite this, my val_loss vacillates between 2 values, and is thus not truly reproducible. I have no idea why one run will give me the one value while the next one will give me the same value or the other one with no rhyme or reason. Can someone tell me what I'm doing wrong please?<\/p>\n\n<pre><code>%env PYTHONHASHSEED=0\n\nimport os\nos.environ[\"CUDA_DEVICE_ORDER\"] = \"PCI_BUS_ID\"\nos.environ[\"CUDA_VISIBLE_DEVICES\"] = \"\"\n\nimport numpy as np\nimport tensorflow as tf\nimport random as rn\n\nseed = 2\n\n# The below is necessary for starting Numpy generated random numbers\n# in a well-defined initial state.\n\nnp.random.seed(seed)\n\n# The below is necessary for starting core Python generated random numbers\n# in a well-defined state.\n\nrn.seed(seed)\n\n# The below tf.set_random_seed() will make random number generation\n# in the TensorFlow backend have a well-defined initial state.\n# For further details, see:\n# https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/set_random_seed\n\ntf.set_random_seed(seed)\n\n# Force TensorFlow to use single thread.\n# Multiple threads are a potential source of non-reproducible results.\n# For further details, see: https:\/\/stackoverflow.com\/questions\/42022950\/\n\nsession_conf = tf.ConfigProto(intra_op_parallelism_threads=1,\n                              inter_op_parallelism_threads=1)\n\nfrom keras import backend as K\n\nsess = tf.Session(graph=tf.get_default_graph(), config=session_conf)\nK.set_session(sess)\n\n# Rest of code follows ..\n\n\n\n\nimport pandas as pd, matplotlib.pyplot as plt, scipy.stats\nfrom keras.models import Sequential\nfrom keras.layers import Dense, BatchNormalization, Embedding, LSTM, Dropout\nfrom keras.callbacks import EarlyStopping\nfrom keras.wrappers.scikit_learn import KerasClassifier\nfrom sklearn.model_selection import RandomizedSearchCV, TimeSeriesSplit\nfrom keras.optimizers import Adam\nfrom keras.losses import binary_crossentropy\n\nmodel = Sequential()\nmodel.add(LSTM(30, input_shape = (sequences.shape[1], sequences.shape[2]), activation = 'relu'))\nmodel.add(BatchNormalization())\nmodel.add(Dense(30, activation = 'relu'))\nmodel.add(BatchNormalization())\nmodel.add(Dense(25, activation = 'relu'))\nmodel.add(BatchNormalization())\nmodel.add(Dense(1, activation = 'sigmoid'))\n\n#compile model\nmodel.compile(optimizer = Adam(lr=.001), loss = binary_crossentropy, metrics=['accuracy'])\n\n<\/code><\/pre>\n","answers":[{"AnswerId":"58125912","CreationDate":"2019-09-26T23:26:44.403","ParentId":null,"OwnerUserId":"815724","Title":null,"Body":"<p>I suggest checking whether your model contains nondeterministic operations. Notably, <code>reduce_sum<\/code> is one such operation. These operations are nondeterministic because floating-point addition and multiplication are nonassociative (the order in which floating-point numbers are added or multiplied affects the result) and because such operations don't guarantee their inputs are added or multiplied in the same order every time. I'm not aware of a complete list of nondeterministic TensorFlow operations.  To investigate which operation may be causing this issue, try reducing the number of layers in the model or changing which layers or operations are used in that model, and see whether doing so gives you consistent results.  See also <a href=\"https:\/\/stackoverflow.com\/questions\/50744565\/how-to-handle-non-determinism-when-training-on-a-gpu\">this question<\/a>.<\/p>\n"}]}
+{"QuestionId":58122611,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-26T18:18:45.763","AcceptedAnswerId":58123458.0,"OwnerUserId":7197896.0,"Title":"tensorflow learnable weight coeffecient with keras API","Body":"<p>I would like to make a loss function of the following form:<\/p>\n\n<p>L = l1(x) + (1\/w1)*l2(x) + (1\/w2)*l3(x) + ln(1+(w1*w1 + w2*w2)<\/p>\n\n<p>where w1 and w2 are learn-able.  Usually, these weight coefs are fixed.  I can create a tf model where these variables are trainable, but I am unsure how to do this with the keras API.  <\/p>\n\n<p>How does one specify this using the tf.keras api and how can I examine their values while training?<\/p>\n","answers":[{"AnswerId":"58123458","CreationDate":"2019-09-26T19:22:59.883","ParentId":null,"OwnerUserId":"5597718","Title":null,"Body":"<p>You can define a custom layer containing the weights w1 &amp; w2.<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nfrom tensorflow.keras import layers\nfrom tensorflow.keras.models import Sequential\nfrom keras import backend as K\n\nclass CustomLayer(layers.Layer):\n\n  def __init__(self):\n    super(CustomLayer, self).__init__()\n\n  def build(self, input_shape):\n    self.w1 = self.add_weight(shape=(1,),\n                             initializer='ones',\n                             trainable=True)\n\n    self.w2 = self.add_weight(shape=(1,),\n                             initializer='ones',\n                             trainable=True)\n\n  def call(self, inputs):\n    y_true = inputs[:,:1]\n    y_pred = inputs[:,1:]\n    loss = K.sum((y_pred - y_true) ** 2. + self.w1 + self.w2, -1) #loss calculation\n    self.add_loss(loss, inputs=inputs) # add the loss\n    return  K.square(inputs) # not used\n\n\nmodel = Sequential()\nmodel.add(layers.Input(shape=(2,)))\nmodel.add(CustomLayer())\nmodel.compile(optimizer='adam', loss=None) #no loss updation here\nX = np.random.randn(10, 2)\nmodel.fit(X, epochs=2)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>10\/10 [==============================] - 0s 6ms\/sample - loss: 3.6391\nEpoch 2\/2\n10\/10 [==============================] - 0s 103us\/sample - loss: 3.6371\n<\/code><\/pre>\n\n<p>Complete example available here: <a href=\"https:\/\/github.com\/yaringal\/multi-task-learning-example\" rel=\"nofollow noreferrer\">https:\/\/github.com\/yaringal\/multi-task-learning-example<\/a><\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/keras\/custom_layers_and_models#layers_encapsulate_a_state_weights_and_some_computation\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/guide\/keras\/custom_layers_and_models#layers_encapsulate_a_state_weights_and_some_computation<\/a><\/p>\n"}]}
+{"QuestionId":58122653,"AnswerCount":0,"Tags":"<python><keras><hdf5>","CreationDate":"2019-09-26T18:21:28.233","AcceptedAnswerId":null,"OwnerUserId":11736428.0,"Title":"Keras saves model up until a certain point, then permission denied","Body":"<p>So I'm currently training a model with Keras, and saving the best to-epoch model every epoch. It works fine most of the time, but now, after a random number of epochs, it just suddenly gives me a permission denied error.<\/p>\n\n<p>I've tried running in administrator, restart the IDE and computer, re-running.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Epoch 329\/500\n90\/99 [==========================&gt;...] - ETA: 0s - loss: 509.8072 - \nmean_absolute_percentage_error: 39.9497 - mean_squared_error: 509.8072\n\nEpoch 00329: mean_absolute_percentage_error improved from 39.62766 to 38.46789, \nsaving model to .mdl_wts.hdf5\n<\/code><\/pre>\n\n<pre><code>OSError: Unable to create file (unable to open file: name = '.mdl_wts.hdf5', errno = 13, \nerror message = 'Permission denied', flags = 13, o_flags = 302)\n<\/code><\/pre>\n\n<p>I'm trying to figure out how to prevent this permission denied error from popping up, to no avail.<\/p>\n","answers":[]}
+{"QuestionId":58122871,"AnswerCount":1,"Tags":"<javascript><tensorflow><machine-learning><deep-learning><tensorflow.js>","CreationDate":"2019-09-26T18:35:41.563","AcceptedAnswerId":null,"OwnerUserId":12043568.0,"Title":"How to use tensorflow model to train and predict mouse movement?","Body":"<pre><code>&lt;html&gt;\n&lt;body id='body'&gt;\n    &lt;button onclick=\"StartData(event)\"&gt; Start&lt;\/button&gt;\n    &lt;button onclick=\"getStopCordinates(event)\"&gt;Stop&lt;\/button&gt;\n    &lt;script&gt;\n        let inputs = [];\n        let labels = [];\n        function Mouse(event) {\n            inputs.push({ x: event.clientX, y: event.clientY })\n        }\n        function StartData() {\n            document.getElementById('body').addEventListener(\"mouseover\", Mouse())\n        }\n\n        function getStopCordinates(event) {\n            labels.push({ x: event.clientX, y: event.clientY })\n            document.getElementById('body').removeEventListener(\"mouseover\", Mouse())\n        }\n\n    &lt;\/script&gt;\n&lt;\/body&gt;\n&lt;\/html&gt;\n<\/code><\/pre>\n\n<p>I am using above code to capture all the x y coordinates of the mouse in the body . When the user moves the pointer towards the stop button i am capturing all the x,y coordinates for this . and when user clicks stop i am capturing stop coordinates also . now i want to train the tensorflow js model from the captured points so that when user moves the mouse with same trajectory i can predict  that he will click the stop button .<\/p>\n\n<p>tensorflow code : <\/p>\n\n<pre><code> const model = tf.sequential(); \n\n  \/\/ Add a single hidden layer\n  model.add(tf.layers.dense({inputShape: [2], units: 1, useBias: true}));\n\n  \/\/ Add an output layer\n  model.add(tf.layers.dense({units: 1, useBias: true}));\n\n  const inputTensor = tf.tensor2d(inputs, [inputs.length, 1]);\n    const labelTensor = tf.tensor2d(labels, [labels.length, 1]);\n\ntrainModel(model,inputs,labels)\nasync function trainModel(model, inputs, labels) {\n  \/\/ Prepare the model for training.  \n  model.compile({\n    optimizer: tf.train.adam(),\n    loss: tf.losses.meanSquaredError,\n    metrics: ['mse'],\n  });\n\n  const batchSize = 32;\n  const epochs = 50;\n\n  return await model.fit(inputs, labels, {\n    batchSize,\n    epochs,\n    shuffle: true,\n    callbacks: tfvis.show.fitCallbacks(\n      { name: 'Training Performance' },\n      ['loss', 'mse'], \n      { height: 200, callbacks: ['onEpochEnd'] }\n    )\n  });\n}\n<\/code><\/pre>\n\n<p>but this code gives error as the inputs and labels are not the same so how to correct this code for the above result ?<\/p>\n","answers":[{"AnswerId":"58130514","CreationDate":"2019-09-27T08:16:52.010","ParentId":null,"OwnerUserId":"5069957","Title":null,"Body":"<p>The inputs and labels should be an array of array and not an array of object. The inputs should rather be <\/p>\n\n<pre><code> [[1, 2], [4, 6], ...]\n<\/code><\/pre>\n\n<p>The same thing holds for the labels.<\/p>\n\n<p>Since your are predicting 2 values, the last layer should have as number of units 2<\/p>\n\n<pre><code> const model = tf.sequential(); \n\n \/\/ Add a single hidden layer\n model.add(tf.layers.dense({inputShape: [2], units: 1, useBias: true}));\n\n \/\/ Add an output layer\n model.add(tf.layers.dense({units: 1, useBias: true}));\n<\/code><\/pre>\n\n<p>The last thing - surely not the least - is to add activation in order to add non linearity to the model<\/p>\n"}]}
+{"QuestionId":58123276,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><data-augmentation>","CreationDate":"2019-09-26T19:08:20.557","AcceptedAnswerId":null,"OwnerUserId":1889418.0,"Title":"Keras - CNN - adjust dataset - remove biased class and data augmentation attempt","Body":"<p>I have a a predicament with the model I am developing based on a popular dataset of skin-cancer images. \nI have to points I'd like some guidance on - <\/p>\n\n<p><strong>A.<\/strong><\/p>\n\n<p>The original dataset is over +10K images with which almost 7000 images belong to one of the seven classes. I've created a subset of 4948 random images with which I ran a function to convert the images into a list of lists - first list contains the image and the latter the class as well as dismiss any images that are of class (5 - the class with the +6800K images). Thought process was to normalise the distribution across the classes.<\/p>\n\n<p>Re-running the original model with an output (Dense layer of 6 neurons instead of 7) - retrieves an error.<\/p>\n\n<p>Am I missing a step to 'indicate' to the model that there are only six possible classes? The model runs only when the output layer has seven neurons.<\/p>\n\n<p>error:<\/p>\n\n<pre><code>Train on 1245 samples, validate on 312 samples\nEpoch 1\/30\n---------------------------------------------------------------------------\nInvalidArgumentError                      Traceback (most recent call last)\n&lt;ipython-input-138-8a3b40a69e37&gt; in &lt;module&gt;\n     25              metrics=[\"accuracy\"])\n     26 \n---&gt; 27 model.fit(X_train, y_train, batch_size=32, epochs=30, validation_split=0.2)\n\n\/anaconda3\/lib\/python3.7\/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, validation_freq, max_queue_size, workers, use_multiprocessing, **kwargs)\n    778           validation_steps=validation_steps,\n    779           validation_freq=validation_freq,\n--&gt; 780           steps_name='steps_per_epoch')\n    781 \n    782   def evaluate(self,\n\n\/anaconda3\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/engine\/training_arrays.py in model_iteration(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, validation_freq, mode, validation_in_fit, prepared_feed_values_from_dataset, steps_name, **kwargs)\n    361 \n    362         # Get outputs.\n--&gt; 363         batch_outs = f(ins_batch)\n    364         if not isinstance(batch_outs, list):\n    365           batch_outs = [batch_outs]\n\n\/anaconda3\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/backend.py in __call__(self, inputs)\n   3290 \n   3291     fetched = self._callable_fn(*array_vals,\n-&gt; 3292                                 run_metadata=self.run_metadata)\n   3293     self._call_fetch_callbacks(fetched[-len(self._fetches):])\n   3294     output_structure = nest.pack_sequence_as(\n\n\/anaconda3\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py in __call__(self, *args, **kwargs)\n   1456         ret = tf_session.TF_SessionRunCallable(self._session._session,\n   1457                                                self._handle, args,\n-&gt; 1458                                                run_metadata_ptr)\n   1459         if run_metadata:\n   1460           proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\nInvalidArgumentError: Received a label value of 6 which is outside the valid range of [0, 6).  Label values: 1 1 2 4 2 1 2 1 2 1 2 2 4 2 2 1 3 1 4 6 0 2 4 2 0 4 2 4 4 0 2 4\n     [[{{node loss_15\/activation_63_loss\/SparseSoftmaxCrossEntropyWithLogits\/SparseSoftmaxCrossEntropyWithLogits}}]]\n<\/code><\/pre>\n\n<p><strong>B.<\/strong><\/p>\n\n<p>I'm attempting to add Data Augmentation as the dataset is relatively small taking into account the number of classes and the sparsity of the images across the classes. Once I try to run the generator I receive the error message below which suggests that there is something wrong with one of the variables in the <code>validation_data<\/code> tuple. I cannot understand what the issue is.<\/p>\n\n<p>Example values of the test set look like:<\/p>\n\n<pre><code>[[[[0.41568627]\n   [0.4       ]\n   [0.43137255]\n   ...\n   [0.54509804]\n   [0.54901961]\n   [0.54509804]]\n\n  [[0.42352941]\n   [0.43137255]\n   [0.43921569]\n   ...\n   [0.56078431]\n   [0.54117647]\n   [0.55294118]]\n\n  [[0.41960784]\n   [0.41960784]\n   [0.45490196]\n   ...\n   [0.51764706]\n   [0.57254902]\n   [0.50588235]]\n\n  ...\n\n  [[0.30980392]\n   [0.36470588]\n   [0.36470588]\n   ...\n   [0.47058824]\n   [0.44705882]\n   [0.41960784]]\n\n  [[0.29803922]\n   [0.31764706]\n   [0.34509804]\n   ...\n   [0.45098039]\n   [0.43921569]\n   [0.4       ]]\n\n  [[0.25882353]\n   [0.30196078]\n   [0.31764706]\n   ...\n   [0.45490196]\n   [0.42745098]\n   [0.36078431]]]\n\n\n [[[0.60784314]\n   [0.59215686]\n   [0.56862745]\n   ...\n   [0.59607843]\n   [0.63921569]\n   [0.63529412]]\n\n  [[0.6627451 ]\n   [0.63137255]\n   [0.62352941]\n   ...\n   [0.67843137]\n   [0.60784314]\n   [0.63529412]]\n\n  [[0.62745098]\n   [0.65098039]\n   [0.6       ]\n   ...\n   [0.61568627]\n   [0.63921569]\n   [0.67058824]]\n\n  ...\n\n  [[0.62352941]\n   [0.6       ]\n   [0.59607843]\n   ...\n   [0.6627451 ]\n   [0.71372549]\n   [0.6745098 ]]\n\n  [[0.61568627]\n   [0.58431373]\n   [0.61568627]\n   ...\n   [0.67058824]\n   [0.65882353]\n   [0.68235294]]\n\n  [[0.61176471]\n   [0.60392157]\n   [0.61960784]\n   ...\n   [0.65490196]\n   [0.6627451 ]\n   [0.66666667]]]]\n\n[2, 1, 4, 4, 2]\n<\/code><\/pre>\n\n<p><strong>Error:<\/strong><\/p>\n\n<pre><code>Epoch 1\/10\n  1\/155 [..............................] - ETA: 11s - loss: 1.7916 - acc: 0.3000\n---------------------------------------------------------------------------\nInvalidArgumentError                      Traceback (most recent call last)\n&lt;ipython-input-139-8f19a958861f&gt; in &lt;module&gt;\n     12 history = model.fit_generator(trainAug.flow(X_train, y_train, batch_size=batch_size)                           \n     13                              ,epochs = 10, validation_data = (X_test, y_test),\n---&gt; 14                               steps_per_epoch= X_train.shape[0]\/\/ batch_size\n     15                              )\n     16 #epochs = epochs, validation_data = (X_test, y_test),\n\n\/anaconda3\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/engine\/training.py in fit_generator(self, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, validation_freq, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch)\n   1431         shuffle=shuffle,\n   1432         initial_epoch=initial_epoch,\n-&gt; 1433         steps_name='steps_per_epoch')\n   1434 \n   1435   def evaluate_generator(self,\n\n\/anaconda3\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/engine\/training_generator.py in model_iteration(model, data, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, validation_freq, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch, mode, batch_size, steps_name, **kwargs)\n    262 \n    263       is_deferred = not model._is_compiled\n--&gt; 264       batch_outs = batch_function(*batch_data)\n    265       if not isinstance(batch_outs, list):\n    266         batch_outs = [batch_outs]\n\n\/anaconda3\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/engine\/training.py in train_on_batch(self, x, y, sample_weight, class_weight, reset_metrics)\n   1173       self._update_sample_weight_modes(sample_weights=sample_weights)\n   1174       self._make_train_function()\n-&gt; 1175       outputs = self.train_function(ins)  # pylint: disable=not-callable\n   1176 \n   1177     if reset_metrics:\n\n\/anaconda3\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/backend.py in __call__(self, inputs)\n   3290 \n   3291     fetched = self._callable_fn(*array_vals,\n-&gt; 3292                                 run_metadata=self.run_metadata)\n   3293     self._call_fetch_callbacks(fetched[-len(self._fetches):])\n   3294     output_structure = nest.pack_sequence_as(\n\n\/anaconda3\/lib\/python3.7\/site-packages\/tensorflow\/python\/client\/session.py in __call__(self, *args, **kwargs)\n   1456         ret = tf_session.TF_SessionRunCallable(self._session._session,\n   1457                                                self._handle, args,\n-&gt; 1458                                                run_metadata_ptr)\n   1459         if run_metadata:\n   1460           proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)\n\nInvalidArgumentError: Received a label value of 6 which is outside the valid range of [0, 6).  Label values: 0 1 6 4 2 4 2 0 1 2\n     [[{{node loss_15\/activation_63_loss\/SparseSoftmaxCrossEntropyWithLogits\/SparseSoftmaxCrossEntropyWithLogits}}]]\n<\/code><\/pre>\n\n<p><strong>Code:<\/strong><\/p>\n\n<pre><code>import tensorflow as tf\n        from tensorflow.keras.models import Sequential\n        from tensorflow.keras.layers import Dense, Dropout, Activation, Flatten, Conv2D, MaxPooling2D\n\n        import numpy as np\n        import pandas as pd\n\n        import matplotlib.pyplot as plt\n\n        import sys\n        import os\n        import cv2\n\n        DATA_DIR = \"\/Users\/namefolder\/PycharmProjects\/skin-cancer\/HAM10000_images_part_1\"\n\n        metadata = pd.read_csv(os.path.join(DATA_DIR, 'HAM10000_metadata.csv'))\n\n        lesion_type_dict = {'nv': 'Melanocytic nevi',\n            'mel': 'Melanoma',\n            'bkl': 'Benign keratosis-like lesions ',\n            'bcc': 'Basal cell carcinoma',\n            'akiec': 'Actinic keratoses',\n            'vasc': 'Vascular lesions',\n            'df': 'Dermatofibroma'}\n\n        metadata['cell_type'] = metadata['dx'].map(lesion_type_dict.get)\n        metadata['dx_code'] = pd.Categorical(metadata['dx']).codes\n\n        # save array of image-id and diagnosis-type (categorical)\n        metadata = metadata[['image_id', 'dx', 'dx_type', 'dx_code']]\n\n        training_data = []\n\n        IMG_SIZE=50\n\n        # preparing training data\n\n        def creating_training_data(path):\n            for img in os.listdir(path):\n                try:\n                    img_array = cv2.imread(os.path.join(path, img), cv2.IMREAD_GRAYSCALE)\n                    new_array = cv2.resize(img_array, (IMG_SIZE, IMG_SIZE))\n                    for index, row in metadata.iterrows():\n                        if (img == row['image_id']+'.jpg') &amp; (row['dx_code'] != 5):\n                            try:\n                                training_data.append([new_array, row['dx_code']])\n                            except Exception as ee:\n                                pass\n                except Exception as e:\n                    pass\n\n            return training_data\n\n        training_data = creating_training_data(DATA_DIR)\n\n        import random\n\n        random.shuffle(training_data)\n\n        # Splitting data into X features and Y label\n        X_train = []\n        y_train = []\n        for features, label in training_data:\n            X_train.append(features)\n            y_train.append(label)\n\n        # Reshaping of the data - required by Tensorflow and Keras (*necessary step of deep-learning using these repos)\n        X_train = np.array(X_train).reshape(-1, IMG_SIZE, IMG_SIZE, 1)\n\n        # Normalize data - to reduce processing requirements\n        X_train = X_train\/255.0\n\n        # model configuration\n        model = Sequential()\n        model.add(Conv2D(64, (3,3), input_shape = X_train.shape[1:]))\n        model.add(Activation(\"relu\"))\n        model.add(MaxPooling2D(pool_size=(2, 2)))\n\n        model.add(Conv2D(64, (3,3)))\n        model.add(Activation(\"relu\"))\n        model.add(MaxPooling2D(pool_size=(2, 2)))\n\n        model.add(Flatten())\n        model.add(Dense(64))\n\n        model.add(Dense(6))\n        model.add(Activation(\"softmax\"))\n\n        model.compile(loss=\"mean_squared_error\",\n                     optimizer=\"adam\",\n                     metrics=[\"accuracy\"])\n\n    # Data Augmentation - Repo enabler\n    from keras.preprocessing.image import ImageDataGenerator\n    from keras.callbacks import ReduceLROnPlateau\n\n    # initialize the training training data augmentation object\n    trainAug = ImageDataGenerator(\n        rescale=1 \/ 255.0,\n        rotation_range=20,\n        zoom_range=0.05,\n        width_shift_range=0.05,\n        height_shift_range=0.05,\n        shear_range=0.05,\n        horizontal_flip=True,\n        fill_mode=\"nearest\")\n\n    # initialize the validation (and testing) data augmentation object\n    valAug = ImageDataGenerator(rescale=1 \/ 255.0)\n\n    #set a leraning rate annealer\n    learning_rate_reduction = ReduceLROnPlateau(monitor='val_acc',\n                                               patience=3,\n                                               verbose=1,\n                                               factor=0.5,\n                                               min_lr=0.00001)\n\n    #Augmented Images model development\n   )\n\n    trainAug.fit(X_train)\n\n    #Fit the model\n    epochs = 10\n    batch_size= 10\n\n    history = model.fit_generator(trainAug.flow(X_train, y_train, batch_size=batch_size),epochs = 10, validation_data = (X_test, y_test), steps_per_epoch= X_train.shape[0]\/\/ batch_size)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58124717","CreationDate":"2019-09-26T21:00:17.643","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Initially you had 7 labels: your code was then expecting labels 0, 1, 2, 3, 4, 5, 6   <\/p>\n\n<p>You removed label 5 from the dataset, ok. Now you have 6 labels in total.<br>\nYour code is expecting: 0, 1, 2, 3, 4, 5   <\/p>\n\n<p>But what you have in your data is: 0, 1, 2, 3, 4, 6<\/p>\n\n<p>After removing the label 5, you need to transform the labels 6 into 5. <\/p>\n\n<hr>\n\n<p>Something in the lines of:<\/p>\n\n<pre><code>if (img == row['image_id']+'.jpg') &amp; (row['dx_code'] &gt; 5):\n    try:\n        training_data.append([new_array, row['dx_code'] - 1])\n    except Exception as ee:\n        pass\nelif (img == row['image_id']+'.jpg') &amp; (row['dx_code'] &lt; 5):\n    try:\n        training_data.append([new_array, row['dx_code']])\n    except Exception as ee:\n        pass\n<\/code><\/pre>\n"}]}
+{"QuestionId":58123324,"AnswerCount":0,"Tags":"<keras><nlp><tensorflow2.0><transfer-learning><elmo>","CreationDate":"2019-09-26T19:12:06.420","AcceptedAnswerId":null,"OwnerUserId":10286806.0,"Title":"Problems with variable_scope in ELMo Embedding layer with Keras","Body":"<p>I'm trying to use a custom ElmoEmbeddingLayer for my model in Keras with tensorflow-gpu==2.0.0-rc0.<\/p>\n\n<p><strong>This is the class:<\/strong><\/p>\n\n<pre><code>class ElmoEmbeddingLayer(Layer):\n    def __init__(self, trainable=True, output_mode=\"default\", **kwargs):\n        assert output_mode in [\"default\", \"word_emb\", \"lstm_outputs1\", \"lstm_outputs2\", \"elmo\"]\n        assert trainable in [True, False]\n        self.output_mode = output_mode\n        self.trainable = trainable\n        super(ElmoEmbeddingLayer, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        self.elmo = hub.Module('https:\/\/tfhub.dev\/google\/elmo\/2', trainable=self.trainable,\n                               name=\"{}_module\".format(self.name))\n\n        self.trainable_weights += K.tf.trainable_variables(scope=\"^{}_module\/.*\".format(self.name))\n        super(ElmoEmbeddingLayer, self).build(input_shape)\n\n    def call(self, x, mask=None):\n        result = self.elmo(\n            K.squeeze(K.cast(x, tf.string), axis=1),\n            as_dict=True,\n            signature='default',                      \n            )['elmo']\n        return result\n\n    def compute_mask(self, inputs, mask=None):\n        return K.not_equal(inputs, '--PAD--')\n\n    def compute_output_shape(self, input_shape):\n        if self.output_mode == \"default\":\n            return (input_shape[0], 1024)\n        if self.output_mode == \"word_emb\":\n            return (input_shape[0], self.max_length, 512)\n        if self.output_mode == \"lstm_outputs1\":\n            return (input_shape[0], self.max_length, 1024)\n        if self.output_mode == \"lstm_outputs2\":\n            return (input_shape[0], self.max_length, 1024)\n        if self.output_mode == \"elmo\":\n            return (input_shape[0], self.max_length, 1024)\n\n    def get_config(self):\n        config = {\n            'output_mode': self.output_mode \n        }\n        return list(config.items()) \n<\/code><\/pre>\n\n<p><strong>And this is my model:<\/strong><\/p>\n\n<pre><code>def train_model(lstm_layers = 300, dropout_rate = 0.1, learning_rate = 0.01,\n                batch_s = 25, verbose = 0, save_model = bool(0), epchs = 10):  \n\n\n  model = tf.keras.Sequential([\n    tf.keras.layers.Input(shape=(max_lenght,), dtype=\"string\"),\n    ElmoEmbeddingLayer(output_mode=\"default\", trainable=True),                           \n    #tf.keras.layers.Embedding(max_lenght+1 , embedding_dim, input_length=max_lenght), \n    #tf.keras.layers.Reshape((max_lenght, embedding_dim)),            \n    #tf.keras.layers.Bidirectional(LSTM(int(lstm_layers))),\n    tf.keras.layers.Dropout(dropout_rate),   \n    tf.keras.layers.Dense(1, activation='sigmoid')\n  ])\n\n  model.summary()\n\n  adam=Adam(lr=learning_rate)\n  model.compile(loss='binary_crossentropy', optimizer=adam, metrics=['accuracy'])\n  history = model.fit(np.array(data_train), np.array(labels_train), epochs=epchs,\n                    batch_size = int(batch_s), validation_data=(np.array(data_test), np.array(labels_test)), verbose = verbose)\n<\/code><\/pre>\n\n<p><strong>However, when i try to fit my model i'm getting this error, how could i fix it?<\/strong><\/p>\n\n<p><strong>RuntimeError: variable_scope elmo_embedding_layer_14_module\/ was unused but the corresponding name_scope was already taken.<\/strong><\/p>\n","answers":[]}
+{"QuestionId":58123393,"AnswerCount":1,"Tags":"<tensorflow><nlp><transformer><bert-language-model>","CreationDate":"2019-09-26T19:18:15.977","AcceptedAnswerId":null,"OwnerUserId":2056969.0,"Title":"How to use Transformers for text classification?","Body":"<p>I have two questions about how to use Tensorflow implementation of the Transformers for text classifications. <\/p>\n\n<ul>\n<li><strong>First<\/strong>, it seems people mostly used only the encoder layer to do the text classification task. However, encoder layer generates one prediction for each input word. Based on my understanding of transformers, the input to the encoder each time is one word from the input sentence. Then, the attention weights and the output is calculated using the current input word. And we can repeat this process for all of the words in the input sentence. As a result we'll end up with pairs of (attention weights, outputs) for each word in the input sentence. Is that correct? Then how would you use this pairs to perform a text classification?  <\/li>\n<li><strong>Second<\/strong>, based on the Tensorflow implementation of transformer <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/text\/transformer\" rel=\"nofollow noreferrer\">here<\/a>, they embed the whole input sentence to one vector and feed a batch of these vectors to the Transformer. However, I expected the input to be a batch of words instead of sentences based on what I've learned from <a href=\"http:\/\/jalammar.github.io\/illustrated-transformer\/\" rel=\"nofollow noreferrer\">The Illustrated Transformer<\/a><\/li>\n<\/ul>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"58131999","CreationDate":"2019-09-27T09:44:54.113","ParentId":null,"OwnerUserId":"5652313","Title":null,"Body":"<p>There are two approaches, you can take:<\/p>\n\n<ol>\n<li>Just average the states you get from the encoder;<\/li>\n<li>Prepend a special token <code>[CLS]<\/code> (or whatever you like to call it) and use the hidden state for the special token as input to your classifier.<\/li>\n<\/ol>\n\n<p>The second approach is used by <a href=\"https:\/\/www.aclweb.org\/anthology\/N19-1423\" rel=\"nofollow noreferrer\">BERT<\/a>. When pre-training, the hidden state corresponding to this special token is used for predicting whether two sentences are consecutive. In the downstream tasks, it is also used for sentence classification. However, my experience is that sometimes, averaging the hidden states give a better result.<\/p>\n\n<p>Instead of training a Transformer model from scratch, it is probably more convenient to use (and eventually finetune) a pre-trained model (BERT, XLNet, DistilBERT, ...) from the <a href=\"https:\/\/github.com\/huggingface\/transformers\" rel=\"nofollow noreferrer\">transformers package<\/a>. It has pre-trained models ready to use in PyTorch and TensorFlow 2.0.<\/p>\n"}]}
+{"QuestionId":58123573,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorflow-lite><tpu>","CreationDate":"2019-09-26T19:31:31.437","AcceptedAnswerId":null,"OwnerUserId":5282060.0,"Title":"Convert Keras model to quantized Tensorflow Lite model that can be used on Edge TPU","Body":"<p>I have a Keras model that I want to run on the Coral Edge TPU device. To do this, it needs to be a Tensorflow Lite model with full integer quantization. I was able to convert the model to a TFLite model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model.save('keras_model.h5')\n\nconverter = tf.lite.TFLiteConverter.from_keras_model_file(\"keras_model.h5\")\ntflite_model = converter.convert()\nopen(\"converted_model.tflite\", \"wb\").write(tflite_model)\n<\/code><\/pre>\n\n<p>But when I run <code>edgetpu_compiler converted_model.tflite<\/code>, I get this error:<\/p>\n\n<pre><code>Edge TPU Compiler version 2.0.267685300\nInvalid model: converted_model.tflite\nModel not quantized\n<\/code><\/pre>\n\n<p>This is because I need to quantize the model, but I'm not sure how to do that. I found <a href=\"https:\/\/www.tensorflow.org\/lite\/performance\/post_training_quantization#full_integer_quantization_of_weights_and_activations\" rel=\"nofollow noreferrer\">this page<\/a> which tells me how to do this, but it wants me to make an input data generator. This is the example it provides:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def representative_dataset_gen():\n  for _ in range(num_calibration_steps):\n    # Get sample input data as a numpy array in a method of your choosing.\n    yield [input]\n<\/code><\/pre>\n\n<p>How can I adapt this code to work with my input data? Where does <code>num_calibration_steps<\/code> come from? Is there a better way to do this? (I saw references to <code>tf.contrib.tpu.keras_to_tpu_model<\/code> but it has been deprecated)<\/p>\n","answers":[{"AnswerId":"58666545","CreationDate":"2019-11-01T22:24:10.463","ParentId":null,"OwnerUserId":"5011438","Title":null,"Body":"<p>I believe <code>num_calibration_steps<\/code> is just the number of times the converter uses your rep set to determine the quantization levels. Just a guess, but maybe it subsamples from your rep set multiple times (bootstrapping or jackknifing). I'm still investigating the whole process myself, but it seems to work for me if I just pass it a single image for each yield, and use <code>num_calibration_steps<\/code> at about 100 (e.g., 100 representative images). You can see my demo script <a href=\"https:\/\/github.com\/mattroos\/EdgeTpuTesting\/blob\/master\/test_edgetpu.ipynb\" rel=\"nofollow noreferrer\">on github<\/a>.<\/p>\n\n<p>The key part is:<\/p>\n\n<pre><code>image_shape = (56, 56, 32)\n\ndef representative_dataset_gen():\n    num_calibration_images = 10\n    for i in range(num_calibration_images):\n        image = tf.random.normal([1] + list(image_shape))\n        yield [image]\n<\/code><\/pre>\n\n<p>Also see my similar response to this question: \n<a href=\"https:\/\/stackoverflow.com\/questions\/58623192\/post-training-full-integer-quantization-of-keras-model\/58666509#58666509\">Post-training full integer quantization of Keras model<\/a><\/p>\n"}]}
+{"QuestionId":58123825,"AnswerCount":1,"Tags":"<python-3.x><numpy><tensorflow><keras>","CreationDate":"2019-09-26T19:51:03.923","AcceptedAnswerId":58124372.0,"OwnerUserId":9841023.0,"Title":"Input to reshape is a tensor with 'batch_size' values, but the requested shape requires a multiple of 'n_features'","Body":"<p>I'm trying to make own attention model and I found example code in here: \n<a href=\"https:\/\/www.kaggle.com\/takuok\/bidirectional-lstm-and-attention-lb-0-043\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/takuok\/bidirectional-lstm-and-attention-lb-0-043<\/a><\/p>\n\n<p>and it works just fine when I run it without modification.<\/p>\n\n<p>But my own data contain only numeric values, I had to change example code. <\/p>\n\n<p>so I erase embedding part in example code and plus, this is what I fixed.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>xtr = np.reshape(xtr, (xtr.shape[0], 1, xtr.shape[1])) \n# xtr.shape() = (n_sample_train, 1, 150), y.shape() = (n_sample_train, 6)\nxte = np.reshape(xte, (xte.shape[0], 1, xte.shape[1]))\n# xtr.shape() = (n_sample_test, 1, 150)\n\nmodel = BidLstm(maxlen, max_features)\nmodel.compile(loss='binary_crossentropy', optimizer='adam',\n              metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>and my BidLstm func looks like,<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>\ndef BidLstm(maxlen, max_features):\n    inp = Input(shape=(1,150))\n    #x = Embedding(max_features, embed_size, weights=[embedding_matrix], trainable=False)(inp) -&gt; I don't need embedding since my own data is numeric.\n    x = Bidirectional(LSTM(300, return_sequences=True, dropout=0.25,\n                           recurrent_dropout=0.25))(inp)\n    x = Attention(maxlen)(x)\n    x = Dense(256, activation=\"relu\")(x)\n    x = Dropout(0.25)(x)\n    x = Dense(6, activation=\"sigmoid\")(x)\n    model = Model(inputs=inp, outputs=x)\n\n    return model\n<\/code><\/pre>\n\n<p>and it said,<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>InvalidArgumentErrorTraceback (most recent call last)\n&lt;ipython-input-62-929955370368&gt; in &lt;module&gt;\n     29 \n     30     early = EarlyStopping(monitor=\"val_loss\", mode=\"min\", patience=1)\n---&gt; 31     model.fit(xtr, y, batch_size=128, epochs=15, validation_split=0.1, callbacks=[early])\n     32     #model.fit(xtr, y, batch_size=256, epochs=1, validation_split=0.1)\n     33 \n\n\/usr\/local\/lib\/python3.5\/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\/python3.5\/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\/python3.5\/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\/python3.5\/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\/python3.5\/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\/python3.5\/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: Input to reshape is a tensor with 128 values, but the requested shape requires a multiple of 150\n     [[{{node attention_16\/Reshape_2}}]]\n     [[{{node loss_5\/mul}}]]\n<\/code><\/pre>\n\n<p>I think something wrong in loss function saids in here:\n<a href=\"https:\/\/stackoverflow.com\/questions\/42115585\/input-to-reshape-is-a-tensor-with-2-batch-size-values-but-the-requested-sha\">Input to reshape is a tensor with 2 * &quot;batch_size&quot; values, but the requested shape has &quot;batch_size&quot;<\/a><\/p>\n\n<p>but I don't know which part to fix it.<\/p>\n\n<p>my keras and tensorflow versions are 2.2.4 and 1.13.0-rc0<\/p>\n\n<p>please help. thanks.<\/p>\n\n<p><strong>Edit 1<\/strong><\/p>\n\n<p>I've change my batch size, like keras saids, multiple of 150(batch_size = 150). than it reports<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Train on 143613 samples, validate on 15958 samples\nEpoch 1\/15\n143400\/143613 [============================&gt;.] - ETA: 0s - loss: 0.1505 - acc: 0.9619\n\n\nInvalidArgumentError: Input to reshape is a tensor with 63 values, but the requested shape requires a multiple of 150\n     [[{{node attention_18\/Reshape_2}}]]\n     [[{{node metrics_6\/acc\/Mean_1}}]]\n<\/code><\/pre>\n\n<p>and details is same as before. what should I do?<\/p>\n","answers":[{"AnswerId":"58124372","CreationDate":"2019-09-26T20:32:17.190","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Your input shape must be <code>(150,1)<\/code>.  <\/p>\n\n<p>LSTM shapes are <code>(batch, steps, features)<\/code>. It's pointless to use LSTMs with 1 step only. (Unless you are using custom training loops with <code>stateful=True<\/code>, which is not your case).    <\/p>\n"}]}
+{"QuestionId":58123842,"AnswerCount":1,"Tags":"<python><tensorflow><keras><raspberry-pi><raspberry-pi-zero>","CreationDate":"2019-09-26T19:51:59.887","AcceptedAnswerId":null,"OwnerUserId":5129982.0,"Title":"Is there a way to use a compiled keras model on the RPI Zero?","Body":"<p>I am working on a Letter Recognition Application for a robot. I used my home PC for training the model and wanted the recognition to be on the RPI Zero W with the already trained model.<\/p>\n\n<p>I got an HDF model. When I try to install Tensorflow on the RPI zero, it's throwing a hash error, as far as I found it this is due to TF beeing for 64bit machines. When I try to install Tensorflow Lite, the installation stocks and crashes.<\/p>\n\n<p>For saving the model I use:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>classifier.save('test2.h5')\n<\/code><\/pre>\n\n<p>That are the Prediction lines:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>test_image = ks.preprocessing.image.load_img('image.jpg')\ntest_image = ks.preprocessing.image.img_to_array(test_image)\n\nresult = classifier.predict(test_image)\n<\/code><\/pre>\n\n<p>I also tried to compile the python script via Nuitka, but as the RPI is ARM and nuitka is not offering cross-compile, this possibility felt out.<\/p>\n","answers":[{"AnswerId":"58391838","CreationDate":"2019-10-15T09:49:30.947","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>You can use already available TFLite to solve your issue. <\/p>\n\n<p>If that does not help, you can also build TFLite from source. <\/p>\n\n<p>Please refer to below links:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/lite\/guide\/build_rpi\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/lite\/guide\/build_rpi<\/a><\/p>\n\n<p><a href=\"https:\/\/medium.com\/@haraldfernengel\/compiling-tensorflow-lite-for-a-raspberry-pi-786b1b98e646\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@haraldfernengel\/compiling-tensorflow-lite-for-a-raspberry-pi-786b1b98e646<\/a> <\/p>\n"}]}
+{"QuestionId":58125212,"AnswerCount":0,"Tags":"<python><machine-learning><deep-learning><pytorch>","CreationDate":"2019-09-26T21:52:31.373","AcceptedAnswerId":null,"OwnerUserId":6271800.0,"Title":"Shared weights with model parallelism in PyTorch","Body":"<p>Our setup involves initial part of the network (input interface) which run on separate GPU cards. Each GPU gets its own portion of data (model parallelism) and process it separately.<\/p>\n\n<p>Each input interface, in turn, it itself a complex nn.Module. Every input interface can occupy one or several cards (say, interface_1 runs on GPU 0 and 1, interface_2 - on GPU 2 and 3 and so on).<\/p>\n\n<p>We need to keep the weights of these input interface the same all over the training. We also need them to run in parallel to save training time which is already weeks for our scenario.<\/p>\n\n<p>The best idea we can think of was initializing the interfaces with the same weights and then average the gradients for them. As the interfaces are identical, updating same weights with the same gradients should keep them the same all over the training process thus achieving desired \u201cshared weights\u201d mode.<\/p>\n\n<p>However, I cannot find any good solution for changing values of these weights and their gradients represented as Parameter in PyTorch. Apparently, PyTorch does not allow to do so.<\/p>\n\n<p>Our current state is: if we copy.deepcopy the \u2018parameter.data\u2019 of the \u201cmaster\u201d interface and assign it to \u2018parameter.data\u2019 of the \"slave\" interface, the values are indeed changed but .to(device_id) does not work and keeps them at the \u201cmaster\u201d device. However, we need them to move to a \u201cslave\u201d device.<\/p>\n\n<p>Could someone please tell me if it is possible at all or, if not, if there is a better way to implement shared weights along with the parallel execution for our scenario?<\/p>\n","answers":[]}
+{"QuestionId":58125498,"AnswerCount":1,"Tags":"<python><keras><neural-network><anaconda><lstm>","CreationDate":"2019-09-26T22:27:02.420","AcceptedAnswerId":null,"OwnerUserId":477228.0,"Title":"Predict the next outcome MLP neural network Python","Body":"<p>I am trying to predict the next set of numbers in my dataset sequence however using the predict function classifys the entire dataset, how can I change my code so that it predicts the next outcome in the sequence?<\/p>\n\n<p>I was following this tutorial and his model outputs 80 based on 50,60,70 as his dataset. Yet mine just predicts the entire dataset? <a href=\"https:\/\/machinelearningmastery.com\/how-to-get-started-with-deep-learning-for-time-series-forecasting-7-day-mini-course\/\" rel=\"nofollow noreferrer\">How to Get Started with Deep Learning for Time Series Forecasting (7-Day Mini-Course)<\/a><\/p>\n\n<p>This is my dataset <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ztuvb.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ztuvb.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>and this is the code:<\/p>\n\n<pre><code>import pandas as pd\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.preprocessing import StandardScaler\nfrom sklearn.neural_network import MLPClassifier\nfrom sklearn.metrics import classification_report,confusion_matrix\n\nfrom numpy import loadtxt\nfrom keras.models import Sequential\nfrom keras.layers import Dense\n# load the dataset\ncol_names = ['N1', 'N2', 'N3', 'N4', 'N5', 'L1', 'L2','label']\n# load dataset \npima = pd.read_csv(\"dataset.csv\", header=None, names=col_names)\npima.head()\n\nfeature_cols = ['N1', 'N2', 'N3', 'N4', 'N5', 'L1', 'L2']\nX = pima[feature_cols] # Features\ny = pima.label \n\nmodel = Sequential()\nmodel.add(Dense(122, input_dim=7, activation='relu'))\nmodel.add(Dense(20, activation='relu'))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\nmodel.fit(X, y, epochs=2000, batch_size=10)\n# =======================\n_, accuracy = model.evaluate(X, y)\nprint('Accuracy: %.2f' % (accuracy*100))\n\nyhat = model.predict(X, verbose=0)\nprint(yhat) &lt;- this outputs the predictions for the entire dataset and not the next prediction\n<\/code><\/pre>\n\n<p>Edit:<\/p>\n\n<p>The output I am getting is this, for the entire dataset.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9U3Yx.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9U3Yx.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The dataset runs from 1 to 1251 rows and I want to predict row 1252 with an output of N1,N2,N3,N4,N5,L1,L2.<\/p>\n","answers":[{"AnswerId":"58125624","CreationDate":"2019-09-26T22:47:26.633","ParentId":null,"OwnerUserId":"7114624","Title":null,"Body":"<p>You have sigmoid activation as your final layer, which gives outputs in range [-1, 1], which is not what you want for your final layer as you are predicting the next number.<\/p>\n\n<pre><code>model.add(Dense(1, activation='sigmoid'))\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>try changing this to <\/p>\n\n<pre><code>model.add(Dense(1))\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>Also, you better use MSE loss function for this type of tasks, BCE is used more for classification tasks, you can check that here <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/36515202\/why-is-the-cross-entropy-method-preferred-over-mean-squared-error-in-what-cases\">Why is the Cross Entropy method preferred over Mean Squared Error? In what cases does this doesn&#39;t hold up?<\/a><\/p>\n\n<pre><code>model.compile(loss='mse', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>hope this solves your problem<\/p>\n"}]}
+{"QuestionId":58125674,"AnswerCount":1,"Tags":"<python><list><dictionary><pytorch><tensor>","CreationDate":"2019-09-26T22:53:31.977","AcceptedAnswerId":58125780.0,"OwnerUserId":5552971.0,"Title":"Remove certain elements of all the tensors in a list of dictionary of tensors in pytorch","Body":"<p>I am new to pytorch, and trying to learn by playing with simple codes. What I have is a list of length zero, populated by a dictionary, where values in this dictionary are tensors. This is what this list looks like: <\/p>\n\n<pre><code>A = [{'boxes': tensor([[ 142.1232,  142.9373, 1106.0452,  971.3792],\n    [ 259.1277,  618.4834, 1100.1293, 1028.8989],\n    [ 232.1346,  692.5888,  763.3408, 1028.6766],\n    [ 206.8070,  312.2080, 1137.1434, 1013.4373],\n    [ 495.9471,  675.7287,  978.5932, 1012.7568]], grad_fn=&lt;StackBackward&gt;), 'labels': tensor([16,  1,  1,  1,  1]), 'scores': tensor([0.9988, 0.9489, 0.5228, 0.3500, 0.0639], grad_fn=&lt;IndexBackward&gt;), 'masks': tensor([[[[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\n\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\n\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\n\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\n\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.]]]], grad_fn=&lt;UnsqueezeBackward0&gt;)}]\n<\/code><\/pre>\n\n<p>This list is the output of Mask-RCNN and I want to remove certain element of all the tensors in the nested dictionary. In this case, I only want to keep information related with class label \"1\". Information related to each class ('boxes', 'labels', 'scores', and 'masks') are all in the same location(s) (index) in each of the tensors. So, I find the index of all the \"1\"s in the tensor with key \"labels\":<\/p>\n\n<pre><code>idxOfClass = [i for i, x in enumerate(list(pred[0]['labels'])) if x == 1]\n<\/code><\/pre>\n\n<p>which gives me: <em>[1, 2, 3, 4]<\/em>. \nThen, I want to keep all the values located at indices in idxOfClass in all of the tensors in the nested dictionary. If I do something like this:<\/p>\n\n<pre><code>Anew = [{pred[0]['boxes'][idxOfClass],pred[0]['labels'][idxOfClass],pred[0]['masks'][idxOfClass],pred[0]['scores'][idxOfClass]}]\n<\/code><\/pre>\n\n<p>I get:<\/p>\n\n<pre><code>[{tensor([[ 259.1277,  618.4834, 1100.1293, 1028.8989],\n    [ 232.1346,  692.5888,  763.3408, 1028.6766],\n    [ 206.8070,  312.2080, 1137.1434, 1013.4373],\n    [ 495.9471,  675.7287,  978.5932, 1012.7568]], grad_fn= \n    &lt;IndexBackward&gt;), tensor([0.9489, 0.5228, 0.3500, 0.0639], \n    grad_fn= \n    &lt;IndexBackward&gt;), tensor([[[[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\n\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\n\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\n\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.]]]], grad_fn=&lt;IndexBackward&gt;), \n    tensor([1, 1, 1, 1])}]\n<\/code><\/pre>\n\n<p>But, this is not a list of dictionary with tensors as the values of the dictionary. This is a list of tensors, with no key-value structure of the nested dictionary. My question is, \"is there any way to keep the original structure of the list when I remove certain elements of all the tensors using indices of specific elements?\".<\/p>\n","answers":[{"AnswerId":"58125780","CreationDate":"2019-09-26T23:06:28.583","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>You can just add key names when you construct the new pred res. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>Anew = [{'boxes': pred[0]['boxes'][idxOfClass],'labels': pred[0]['labels'][idxOfClass],'masks': pred[0]['masks'][idxOfClass],'scores': pred[0]['scores'][idxOfClass]}]\n<\/code><\/pre>\n"}]}
+{"QuestionId":58125788,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><keras><gpu><conv-neural-network>","CreationDate":"2019-09-26T23:06:56.307","AcceptedAnswerId":58126614.0,"OwnerUserId":7421340.0,"Title":"Keras CNN how can i reduce gpu memory usage with large image sizes?","Body":"<p>I am trying to train a cnn-lstm model, my sample image sizes are 640x640.<\/p>\n\n<p>I have a GTX 1080 ti 11GB.<\/p>\n\n<p>I am using Keras with tensorflow backend.<\/p>\n\n<p>Here is the model.<\/p>\n\n<pre><code>img_input_1 = Input(shape=(1, n_width, n_height, n_channels))\n\nconv_1 = TimeDistributed(Conv2D(96, (11,11), activation='relu', padding='same'))(img_input_1)\n\npool_1 = TimeDistributed(MaxPooling2D((3,3)))(conv_1)\n\nconv_2 = TimeDistributed(Conv2D(128, (11,11), activation='relu', padding='same'))(pool_1)\n\nflat_1 = TimeDistributed(Flatten())(conv_2)\n\ndense_1 = TimeDistributed(Dense(4096, activation='relu'))(flat_1)\n\ndrop_1 = TimeDistributed(Dropout(0.5))(dense_1)\n\nlstm_1 = LSTM(17, activation='linear')(drop_1)\n\ndense_2 = Dense(4096, activation='relu')(lstm_1)\n\ndense_output_2 = Dense(1, activation='sigmoid')(dense_2)\n\nmodel = Model(inputs=img_input_1, outputs=dense_output_2)\n\nop = optimizers.Adam(lr=0.00001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.001)\n\nmodel.compile(loss='mean_absolute_error', optimizer=op, metrics=['accuracy'])\n\nmodel.fit(X, Y, epochs=3, batch_size=1)\n<\/code><\/pre>\n\n<p>Right now using this model i can only use the training data when the images are resized to 60x60, any larger and i run out of GPU memory.<\/p>\n\n<p>I want to use the largest possible size as i want to retain as much discriminatory information as possible. (The y labels will be mouse screen coordinates between 0 - 640)<\/p>\n\n<p>Among many others, i found this answer:\n<a href=\"https:\/\/ai.stackexchange.com\/questions\/3938\/how-to-handle-images-of-large-sizes-in-cnn\">https:\/\/ai.stackexchange.com\/questions\/3938\/how-to-handle-images-of-large-sizes-in-cnn<\/a><\/p>\n\n<p>Though i am not sure how i can \"restrict your CNN\" or \"stream your data in each epoch\" or if these would help.<\/p>\n\n<p>How can i reduce the amount of memory used so i can increase the image sizes?<\/p>\n\n<p>Is it possible to sacrifice training time\/computation speed in favor of higher resolution data whilst retaining model effectiveness?<\/p>\n\n<p>Note: The above model is not final, just a basic outlay.<\/p>\n","answers":[{"AnswerId":"58126614","CreationDate":"2019-09-27T01:31:42.740","ParentId":null,"OwnerUserId":"1699075","Title":null,"Body":"<p>Your <code>Dense<\/code> layer is probably blowing up the training. To give some context, let's assume you are using the <code>640x640x3<\/code> image size. And let's forget about the <code>LSTM<\/code> layer as well and pretend this is a non time-series task (and of course, being a time-series problem complexity becomes worse).<\/p>\n\n<p>Here are the output sizes.<\/p>\n\n<ul>\n<li><code>Conv1<\/code> -> <code>640x640x96<\/code><\/li>\n<li><code>Maxpool1<\/code> -> <code>210x210x96<\/code> (appx)<\/li>\n<li><code>Conv2<\/code> -> <code>210x210x128<\/code><\/li>\n<\/ul>\n\n<p>Now comes the bottleneck. You are then <code>flatten()<\/code>-ing the output and sending it to a <code>Dense<\/code> layer. This dense layer has <code>210x210x128x4096<\/code> parameters (which is <code>23,121,100,800<\/code>). Assuming <code>32-bit<\/code> precision, your dense layer will take around ~86GB (I hope my calculations are correct, but I guarantee you that is not a small number).<\/p>\n\n<p>So you got few options.<\/p>\n\n<ul>\n<li>First and most obvious, reduce your <code>Dense<\/code> layer size.<\/li>\n<li>Reduce the size of the minibatches<\/li>\n<li>Reduce the channel depth of the <code>Conv<\/code> layers.<\/li>\n<li>And you might want to think if you really want the input to be <code>640x640x3<\/code>. Depending on what you're trying to achieve, you might be able to do that with a smaller image.<\/li>\n<\/ul>\n"}]}
+{"QuestionId":58125963,"AnswerCount":0,"Tags":"<python><tensorflow><computer-vision>","CreationDate":"2019-09-26T23:34:17.663","AcceptedAnswerId":null,"OwnerUserId":12125847.0,"Title":"NotFoundError: Failed to create a directory:","Body":"<p>when i try to train the model with ModelTraining class: \nps: i use windows 10\n<a href=\"https:\/\/i.stack.imgur.com\/q7TGx.png\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p>It's show the following error:<\/p>\n\n<blockquote>\n  <p>Using Enhanced Data Generation Found 14 images belonging to 2 classes.\n  Found 8 images belonging to 2 classes. JSON Mapping for the model\n  classes saved to  C:\/Users\/Ihab\/Desktop\/Object\n  train.predict\/dataset\\json\\model_class.json Number of experiments\n  (Epochs) :  100 Epoch 1\/100 13\/14 [==========================>...] -\n  ETA: 6s - loss: 2.0208 - acc: 0.6154 Epoch 1\/100  8\/14\n  [================>.............] - ETA: 4s - loss: 0.7097 - acc:\n  0.5000 Epoch 00001: val_acc improved from -inf to 0.50000, saving model to C:\/Users\/Ihab\/Desktop\/Object\n  train.predict\/dataset\\models\\model_ex-001_acc-0.500000.h5<\/p>\n  \n  <hr>\n  \n  <p>NotFoundError: Failed to create a directory:\n  C:\/Users\/Ihab\/Desktop\/Object\n  train.predict\/dataset\\logs\\resnet_lr-0.001_2019-09-25-01-24-54\\train;\n  No such file or directory [Op:CreateSummaryFileWriter]<\/p>\n<\/blockquote>\n","answers":[]}
+{"QuestionId":58126125,"AnswerCount":1,"Tags":"<python><keras><convolution><max-pooling>","CreationDate":"2019-09-27T00:01:06.617","AcceptedAnswerId":null,"OwnerUserId":7092731.0,"Title":"Is it possible to extend \"im2col\" and \"col2im\" to N-D images?","Body":"<p>\"Im2col\" has already been implemented, <a href=\"https:\/\/stackoverflow.com\/questions\/30109068\/implement-matlabs-im2col-sliding-in-python\" title=\"Python\">Implement MATLAB's im2col 'sliding' in Python<\/a>, efficiently for 2-D images in Python. I was wondering whether it is possible to extend this to arbitrary N-D images? Many applications involve high-dimensional data (e.g. convolutions, filtering, max pooling, etc.).<\/p>\n","answers":[{"AnswerId":"58126294","CreationDate":"2019-09-27T00:31:25.743","ParentId":null,"OwnerUserId":"7092731","Title":null,"Body":"<p>So the purpose of this question was really just to post my solution to this problem publicly. I could not seem to find such a solution on Google, so I decided to take a stab at it myself. Turns out the implementation is actually quite simple to extend from \"Approach #2\" in the post referenced in my question!<\/p>\n\n<p><strong>Efficient Implementation of N-D \"im2col\"<\/strong><\/p>\n\n<pre><code>def im2col(im, win, strides = 1):\n    # Dimensions\n    ext_shp = tuple(np.subtract(im.shape, win) + 1)\n    shp = tuple(win) + ext_shp\n    strd = im.strides*2\n    win_len = np.prod(win)\n    try:\n        len(strides)\n    except:\n        strides = [strides]*im.ndim\n    strides = [min(i, s) for i, s in zip(im.shape, strides)]\n\n    # Stack all possible patches as an N-D array using a strided view followed by reshaping\n    col = np.lib.stride_tricks.as_strided(im, shape = shp, strides = strd).reshape(win_len, -1).reshape(-1, *ext_shp)\n\n    # Extract patches with stride and reshape into columns\n    slcs = tuple([slice(None, None, None)] + [slice(None, None, s) for s in strides])\n    col = col[slcs].reshape(win_len, -1)\n\n    return col\n<\/code><\/pre>\n\n<p><strong>Efficient Implementation of N-D \"col2im\"<\/strong><\/p>\n\n<pre><code>def col2im(col, im_shp, win, strides = 1):\n    # Dimensions\n    try:\n        len(strides)\n    except:\n        strides = [strides]*len(im_shp)\n    strides = [min(i, s) for i, s in zip(im_shp, strides)]\n\n    # Reshape columns into image\n    if col.ndim &gt; 1:\n        im = col.reshape((-1, ) + tuple(np.subtract(im_shp, win)\/\/np.array(strides) + 1))[0]\n    else:\n        im = col.reshape(tuple(np.subtract(im_shp, win)\/\/np.array(strides) + 1))\n\n    return im\n<\/code><\/pre>\n\n<p><strong>Verification That It Works<\/strong><\/p>\n\n<p>Let's define an arbitrary 3-D input:<\/p>\n\n<pre><code>x = np.arange(216).reshape(6, 6, 6)\nprint(x)\n\n[[[  0   1   2   3   4   5]\n  [  6   7   8   9  10  11]\n  [ 12  13  14  15  16  17]\n  [ 18  19  20  21  22  23]\n  [ 24  25  26  27  28  29]\n  [ 30  31  32  33  34  35]]\n\n [[ 36  37  38  39  40  41]\n  [ 42  43  44  45  46  47]\n  [ 48  49  50  51  52  53]\n  [ 54  55  56  57  58  59]\n  [ 60  61  62  63  64  65]\n  [ 66  67  68  69  70  71]]\n\n [[ 72  73  74  75  76  77]\n  [ 78  79  80  81  82  83]\n  [ 84  85  86  87  88  89]\n  [ 90  91  92  93  94  95]\n  [ 96  97  98  99 100 101]\n  [102 103 104 105 106 107]]\n\n [[108 109 110 111 112 113]\n  [114 115 116 117 118 119]\n  [120 121 122 123 124 125]\n  [126 127 128 129 130 131]\n  [132 133 134 135 136 137]\n  [138 139 140 141 142 143]]\n\n [[144 145 146 147 148 149]\n  [150 151 152 153 154 155]\n  [156 157 158 159 160 161]\n  [162 163 164 165 166 167]\n  [168 169 170 171 172 173]\n  [174 175 176 177 178 179]]\n\n [[180 181 182 183 184 185]\n  [186 187 188 189 190 191]\n  [192 193 194 195 196 197]\n  [198 199 200 201 202 203]\n  [204 205 206 207 208 209]\n  [210 211 212 213 214 215]]]\n<\/code><\/pre>\n\n<p>Let's extract all the patches with a non-uniform window and equal stride:<\/p>\n\n<pre><code>y = im2col(x, [1, 3, 2], strides = [1, 3, 2])\nprint(y.T) # transposed for ease of visualization\n\n[[  0   1   6   7  12  13]\n [  2   3   8   9  14  15]\n [  4   5  10  11  16  17]\n [ 18  19  24  25  30  31]\n [ 20  21  26  27  32  33]\n [ 22  23  28  29  34  35]\n [ 36  37  42  43  48  49]\n [ 38  39  44  45  50  51]\n [ 40  41  46  47  52  53]\n [ 54  55  60  61  66  67]\n [ 56  57  62  63  68  69]\n [ 58  59  64  65  70  71]\n [ 72  73  78  79  84  85]\n [ 74  75  80  81  86  87]\n [ 76  77  82  83  88  89]\n [ 90  91  96  97 102 103]\n [ 92  93  98  99 104 105]\n [ 94  95 100 101 106 107]\n [108 109 114 115 120 121]\n [110 111 116 117 122 123]\n [112 113 118 119 124 125]\n [126 127 132 133 138 139]\n [128 129 134 135 140 141]\n [130 131 136 137 142 143]\n [144 145 150 151 156 157]\n [146 147 152 153 158 159]\n [148 149 154 155 160 161]\n [162 163 168 169 174 175]\n [164 165 170 171 176 177]\n [166 167 172 173 178 179]\n [180 181 186 187 192 193]\n [182 183 188 189 194 195]\n [184 185 190 191 196 197]\n [198 199 204 205 210 211]\n [200 201 206 207 212 213]\n [202 203 208 209 214 215]]\n<\/code><\/pre>\n\n<p>Let's convert this back to a (downsampled) image:<\/p>\n\n<pre><code>z = col2im(y, x.shape, [1, 3, 2], strides = [1, 3, 2])\nprint(z)\n\n[[[  0   2   4]\n  [ 18  20  22]]\n\n [[ 36  38  40]\n  [ 54  56  58]]\n\n [[ 72  74  76]\n  [ 90  92  94]]\n\n [[108 110 112]\n  [126 128 130]]\n\n [[144 146 148]\n  [162 164 166]]\n\n [[180 182 184]\n  [198 200 202]]]\n<\/code><\/pre>\n\n<p>As you can see, the final output is indeed the downsampled image that we expect (you can easily check this by going value by value). The dimensionality and strides I chose were purely illustrative. There's no reason why the window size has to be the same as your stride or that you can't go higher than 3 dimensions.<\/p>\n\n<p><strong>Applications<\/strong><\/p>\n\n<p>If you want to use this practically, all you have to do is intercept the output of im2col before turning it back into an image. For example, if you want to do pooling, you could take the mean or the maximum across the 0th axis. If you want to do a convolution, you just need to multiply this by your flattened convolutional filter.<\/p>\n\n<p>There may be more efficient alternatives to this already implemented under the hood of Tensorflow, etc. that are faster than \"im2col.\" This is not meant to be the MOST efficient implementation. And of course, you could possibly optimize my code further by eliminating the intermediate reshaping step in \"im2col,\" but it wasn't immediately obvious to me so I just left it at that. If you have a better solution, let me know. Anyways, hope this helps someone else looking for the same answer!<\/p>\n"}]}
+{"QuestionId":58126267,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-estimator>","CreationDate":"2019-09-27T00:26:02.050","AcceptedAnswerId":null,"OwnerUserId":2930156.0,"Title":"Tensorflow metrics drop precipitously after restarting from a saved checkpoint","Body":"<p>This is a question that has been nagging colleagues past and present for quite some time. If I kill and tensorflow training instance and restart, the documentation states that it will resume from the latest checkpoint saved under model_dir. Usually however we see a sizable dip in the metrics such as AUC, logloss etc. I never really understood the reason behind such dip. All the relevant information is saved in the checkpoint file for tensorflow to reconstruct the state before it gets killed, except possibly the iterator of the input dataset (under estimator framework). Historically this has discouraged me and others from restarting tensorflow training by loading from existing checkpoints, because metrics are almost guaranteed to worsen.<\/p>\n\n<p>Today I did something slightly more aggressive by bringing up a training instance with 1 chief and 7 workers on a single machine, without parameter server. This is supposed to run under the STANDALONE distribution strategy with the chief spawning multiple worker threads. When I killed training and restarted from the latest checkpoint with the same distributed setup, however, the metrics dropped by almost half of its entire gain since inception! Note that the multiple jumps in the tensorboard plots below are due to multiple restart I initiated from the same saved checkpoint.<\/p>\n\n<p>I then started looking at the gradient histogram\/distribution plots, and noticed that the gradient of all the bias weights, especially the last one bias_3, increased significantly in variance after restart. Clearly something is wrong, but comparing the variables in the checkpoints before and after restart reveals that the values are still evolving continuously, albeit with a bigger than normal jump. To prevent inadvertent overfitting, I shuffled my input files (about 100 of those) with a different random seed so that the restart does not go back to the beginning of the original training epoch. In any case, the training has lasted for more than a few epochs so that it's highly unlikely to be a case of overfitting on seen training data.<\/p>\n\n<p>Besides that, I can't think of any reason why the gradients would experience such enormous jumps, causing the metrics to tank. <\/p>\n\n<p>Update:<\/p>\n\n<p>Curiously, I found that the Adagrad of the bias vector of the last layer of my DNN stays constant at 0.1, even though the gradients are clearly nonzero, albeit extremely small. I don't think this explains the jump after restart because the serialized Adagrad value should be close to the one stored in memory.<\/p>\n\n<pre><code>python -m inspect_checkpoint --file_name model.ckpt-720730 --tensor_name bias_3\/Adagrad\ntensor_name:  bias_3\/Adagrad\n[0.1]\n\npython -m inspect_checkpoint --file_name model.ckpt-591678 --tensor_name bias_3\/Adagrad\ntensor_name:  bias_3\/Adagrad\n[0.1]\n<\/code><\/pre>\n\n<p>For comparison the Adagrad of bias2 appears to be change dramatically over 200k steps (batch size 1024). <\/p>\n\n<pre><code>python -m inspect_checkpoint --file_name model.ckpt-720730 --tensor_name bias_2\/Adagrad\ntensor_name:  bias_2\/Adagrad\n[80.608215    1.0615484   1.1355964  89.86574    90.919426   80.83253\n  1.7569603   1.2538469  69.11257    77.651375   89.71191     0.7492978\n...\n\npython -m inspect_checkpoint --file_name model.ckpt-591678 --tensor_name bias_2\/Adagrad\ntensor_name:  bias_2\/Adagrad\n[47.73241     1.0530229   1.1187999  50.532574   51.123966   40.66552\n  1.7377361   1.2532469  36.016315   45.788876   50.346928    0.7473773\n...\n<\/code><\/pre>\n\n<p>For completeness, I am using dropout throughout the DNN, but no regularization of any kind. No fancy gradient multiplier, gradient decay function, etc. Just plain old Adagrad. Any help or suggestion is highly appreciated.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/HpRnd.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/HpRnd.png\" alt=\"enter image description here\"><\/a> <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/GfTLV.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/GfTLV.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/N4Kwm.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/N4Kwm.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/xSCCc.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/xSCCc.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/RSZXl.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/RSZXl.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/AtZcZ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/AtZcZ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/KYZ9y.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/KYZ9y.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":58126269,"AnswerCount":0,"Tags":"<machine-learning><keras><neural-network><multilabel-classification>","CreationDate":"2019-09-27T00:26:12.440","AcceptedAnswerId":null,"OwnerUserId":9376040.0,"Title":"Issue with multilabel classification","Body":"<p>I followed this tutorial: <a href=\"https:\/\/medium.com\/@vijayabhaskar96\/multi-label-image-classification-tutorial-with-keras-imagedatagenerator-cd541f8eaf24\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@vijayabhaskar96\/multi-label-image-classification-tutorial-with-keras-imagedatagenerator-cd541f8eaf24<\/a>\nand wrote some of my code for multilabel classification. I had it working with one-hot encoding on a small scale but I had to move to option 2 mentioned in the article because I have 6000 classes and therefore one hot was not viable. I managed to train the network and it said 99% accuracy and 83% f1 score. However, when I'm trying to test the network, for every image it's outputting some combination of only 3 labels when there are 6000 possible labels. I wondered if maybe the code to test the model was incorrect. I tried using the code mentioned in the post and it doesn't work:<\/p>\n\n<pre><code>test_generator.reset()\npred = model.predict_generator(test_generator, steps=STEP_SIZE_TEST, verbose=1);\npred_bool = (pred &gt; 0.5)\n<\/code><\/pre>\n\n<blockquote>\n  <p>unorderable <code>types: list() &gt; float()<\/code><\/p>\n<\/blockquote>\n\n<p>I've tried hard to fix this and not figured it out and I can't find any examples online of anyone doing something similar. Does anyone have an idea of how to get this prediction part working using this code block (I had it with another 2 options and was getting that issue printing one or several labels) or why the model might be failing in training with this behavior?<\/p>\n\n<p>EDIT: for more context on the training issue, here is all the training code:<\/p>\n\n<pre><code>import json\ninput_file = open ('class_names_6000.json')\njson_array = json.load(input_file)\n#print(str(json_array))\n\nargs = parser.parse_args()\ngpu_options = tf.GPUOptions(allow_growth=True)\nsess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))\nprint('Loading Data...')\ndf = pd.read_csv('dataset_train.csv')\ndf[\"labels\"]=df[\"labels\"].apply(lambda x:x.split(\",\"))\ndatagen=ImageDataGenerator(rescale=1.\/255.)\ntest_datagen=ImageDataGenerator(rescale=1.\/255.)\ntrain_generator=datagen.flow_from_dataframe(\ndataframe=df,\ndirectory=\"\",\nx_col=\"Filepaths\",\ny_col=\"labels\",\nbatch_size=128,\nseed=42,\nshuffle=True,\nclass_mode=\"categorical\",\nclasses=json_array,\ntarget_size=(100,100))\n\ndf = pd.read_csv('dataset_test.csv')\ndf[\"labels\"]=df[\"labels\"].apply(lambda x:x.split(\",\"))\ntest_generator=test_datagen.flow_from_dataframe(\ndataframe=df,\ndirectory=\"\",\nx_col=\"Filepaths\",\ny_col=\"labels\",\nbatch_size=128,\nseed=42,\nshuffle=True,\nclass_mode=\"categorical\",\nclasses=json_array,\ntarget_size=(100,100))\n\ndf = pd.read_csv('dataset_validation.csv')\ndf[\"labels\"]=df[\"labels\"].apply(lambda x:x.split(\",\"))\nvalid_generator=test_datagen.flow_from_dataframe(\ndataframe=df,\ndirectory=\"\",\nx_col=\"Filepaths\",\ny_col=\"labels\",\nbatch_size=128,\nseed=42,\nshuffle=True,\nclass_mode=\"categorical\",\nclasses=json_array,\ntarget_size=(100,100))\nprint('Data Loaded.')\n\nf1_score_callback = ComputeF1()\nmodel = build_model('train', numclasses=len(json_array), model_name = args.model)\n\nImageFile.LOAD_TRUNCATED_IMAGES = True\n<\/code><\/pre>\n\n<p><strong>Also, an important detail, when training, it says the accuracy is 99% and the f1 score is 84% with an validation f1 score at 84% as well.<\/strong><\/p>\n","answers":[]}
+{"QuestionId":58126270,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-27T00:26:31.583","AcceptedAnswerId":null,"OwnerUserId":6052531.0,"Title":"Optimizing a custom conv2d layer in tf.keras","Body":"<p>I am trying to efficiently implement the following kind of conv2d layer. The current implementation I believe works but is very inefficient.<\/p>\n\n<p>Input tensor of size<\/p>\n\n<p>(batch_size x W x H x C_in) <\/p>\n\n<p>Output tensor<\/p>\n\n<p>(batch_size x W x H x C_out) <\/p>\n\n<p>The layer takes two parameter, number of units (C_u), and a list of K conv kernels (known ahead of time). Each conv kernel is size (W,H,1,N) where N is the number of out channels (in channels being 1). Note, different kernels in the same like have different Ns!<\/p>\n\n<p>First we apply a densely connected layer (trainable) transforming  input shape to <\/p>\n\n<p>(batch_size x W x H x C_u) <\/p>\n\n<p>Then, I want to apply each of the convolutional kernels to each of the channels.<\/p>\n\n<p>This results in C_u x K x (batch_size x W x H x N)<\/p>\n\n<p>I then want to take a max along N (so I get (batch_size x W x H x 1)) and concatenate everything to get<\/p>\n\n<p>(batch_size x W x H x (C_u x K))<\/p>\n\n<p>(so C_out = C_u x K)<\/p>\n\n<p>Here is one way to implement this but training time is super slow and this does not play well with being put on the GPU:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.keras import layers\n\nclass fixedConvLayer(layers.Dense):\n\n    def __init__(self, units, conv_kernels, **params):\n        params['units']=units\n        self.conv_kernels_numpy = conv_kernels\n        super().__init__(**params)\n        return\n\n    def build(self, input_shape):\n\n        super().build(input_shape)\n        self.conv_kernels = [tf.convert_to_tensor(np.reshape(kernels,[3,3,1,-1]))                                                           \n                                  for kernels in self.conv_kernels_numpy]\n\n        return \n\n    def comp_filters(self,channel):\n\n        return tf.concat([\n                    tf.math.reduce_max(tf.nn.conv2d(channel,\n                                      filter=kernel,\n                                      strides=1,\n                                      padding='SAME'),axis=3,keepdims=True)\n\n                        for kernel in self.conv_kernels],axis=3)\n\n\n    def call(self,inputs):\n\n        #take from Dense definition and slightly modify\n        inputs = tf.convert_to_tensor(inputs)\n        rank = tf.rank(inputs)\n        if rank != 4:\n            assert 'Rank expected to be 4'\n\n\n        # Broadcasting is required for the inputs.\n        outputs = tf.tensordot(inputs, self.kernel, [[3], [0]])\n\n        # Reshape the output back to the original ndim of the input.\n        shape = inputs.shape.as_list()\n        output_shape = shape[:-1] + [self.units]\n        outputs.set_shape(output_shape)\n\n        if self.use_bias:\n            outputs = tf.nn.bias_add(outputs, self.bias)\n        if self.activation is not None:\n            outputs = self.activation(outputs)  \n\n        #apply the conv filters\n        channel_list = tf.split(outputs,num_or_size_splits= self.units,axis = -1)\n        max_layers = tf.concat([self.comp_filters(channel) for channel in channel_list],axis=3)\n\n        return max_layers\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58126494,"AnswerCount":1,"Tags":"<python><pandas><csv><tensorflow>","CreationDate":"2019-09-27T01:06:27.243","AcceptedAnswerId":58130094.0,"OwnerUserId":9873122.0,"Title":"How to Translate CSV Data into TFRecord Files","Body":"<p>Currently I am working on a system that can take data from a CSV file and import it into a TFRecord file, However I have a few questions.<\/p>\n\n<p>For starters, I need to know what type a TFRecord file can take, when using CSV types are removed.<\/p>\n\n<p>Secondly, How can I convert data type:object into a type that a TFRecord can take?<\/p>\n\n<p>I have two columns (will post example below) of two objects types that are strings, How can I convert that data to the correct type for TFRecords?<\/p>\n\n<p>When importing Im hoping to append data from each row at a time into the TFRecord file, any advice or documentation would be great, I have been looking for some time at this problem and it seems there can only be ints,floats inputted into a TFRecord but what about a list\/array of Integers?<\/p>\n\n<p>Thankyou for reading!<\/p>\n\n<p>Quick Note, I am using PANDAS to create a dataframe of the CSV file<\/p>\n\n<p>Some Example Code Im using <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import pandas as pd\nfrom ast import literal_eval\nimport numpy as np\nimport tensorflow as tf\n\n\ntf.compat.v1.enable_eager_execution()\n\n\ndef Start():\n    db = pd.read_csv(\"I:\\Github\\ClubKeno\\Keno Project\\Database\\..\\LotteryDatabase.csv\")\n\n    pd.DataFrame = db\n    print(db['Winning_Numbers'])\n    print(db.dtypes)\n\n    training_dataset = (\n        tf.data.Dataset.from_tensor_slices(\n            (\n                tf.cast(db['Draw_Number'].values, tf.int64),\n                tf.cast(db['Winning_Numbers'].values, tf.int64),\n                tf.cast(db['Extra_Numbers'].values, tf.int64),\n                tf.cast(db['Kicker'].values, tf.int64)\n            )\n        )\n    )\n\n    for features_tensor, target_tensor in training_dataset:\n        print(f'features:{features_tensor} target:{target_tensor}')\n<\/code><\/pre>\n\n<p>Error Message:<\/p>\n\n<p><img src=\"https:\/\/cdn.discordapp.com\/attachments\/279786369902051328\/626967249395122213\/Capture.PNG\" alt=\"Error Message\"><\/p>\n\n<p><a href=\"https:\/\/cdn.discordapp.com\/attachments\/502661247809093673\/626946732239880194\/LotteryDatabase.csv\" rel=\"nofollow noreferrer\">CSV Data<\/a><\/p>\n\n<p>Update:\nGot Two Columns of dating working using the following function...<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>dataset = tf.data.experimental.make_csv_dataset(\n        file_pattern=databasefile,\n        column_names=['Draw_Number', 'Kicker'],\n        column_defaults=[tf.int64, tf.int64],\n    )\n<\/code><\/pre>\n\n<p>However when trying to include my two other column object types\n(What data looks like in both those columns)\n<code>\"3,9,11,16,25,26,28,29,36,40,41,46,63,66,67,69,72,73,78,80\"<\/code><\/p>\n\n<p>I get an error, here is the function I tried for that<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    dataset = tf.data.experimental.make_csv_dataset(\n        file_pattern=databasefile,\n        column_names=['Draw_Number', 'Winning_Numbers', 'Extra_Numbers', 'Kicker'],\n        column_defaults=[tf.int64, tf.compat.as_bytes, tf.compat.as_bytes, tf.int64],\n        header=True,\n        batch_size=100,\n        field_delim=',',\n        na_value='NA'\n    )\n<\/code><\/pre>\n\n<p>This Error Appears:<\/p>\n\n<pre><code>TypeError: Failed to convert object of type &lt;class 'function'&gt; to Tensor. Contents: &lt;function as_bytes at 0x000000EA530908C8&gt;. Consider casting elements to a supported type.\n<\/code><\/pre>\n\n<p>Should I try to Cast those two types outside the function and try combining it later into the TFRecord file alongside the tf.data from the <code>make_csv_dataset<\/code> function? <\/p>\n","answers":[{"AnswerId":"58130094","CreationDate":"2019-09-27T07:47:40.987","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<blockquote>\n  <p>For starters, I need to know what type a TFRecord file can take, when using CSV types are removed.<\/p>\n<\/blockquote>\n\n<p>TFRecord accepts following datatypes-\nstring, byte, float32, float 64, bool, enum, int32, int64, uint32, uint64\nTalked <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/load_data\/tf_records#data_types_for_tfexample\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<blockquote>\n  <p>Secondly, How can I convert data type:object into a type that a TFRecord can take?<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/load_data\/tf_records#creating_a_tfexample_message\" rel=\"nofollow noreferrer\">Here<\/a> is an example from TF, it is a bit complicated to digest it at once but if you read it carefully it is easy.<\/p>\n\n<blockquote>\n  <p>have two columns (will post example below) of two objects types that are strings, How can I convert that data to the correct type for TFRecords?<\/p>\n<\/blockquote>\n\n<p>For string type data, you require <code>tf.train.BytesList<\/code> which returns a bytes_list from a string.<\/p>\n\n<blockquote>\n  <p>When importing Im hoping to append data from each row at a time into the TFRecord file, any advice or documentation would be great, I have been looking for some time at this problem and it seems there can only be ints,floats inputted into a TFRecord but what about a list\/array of Integers?<\/p>\n  \n  <p>Quick Note, I am using PANDAS to create a dataframe of the CSV file<\/p>\n<\/blockquote>\n\n<p>Instead of reading csv file using Pandas, I would recommend you to use <code>tf.data.experimental.make_csv_dataset<\/code> defined <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/data\/experimental\/make_csv_dataset\" rel=\"nofollow noreferrer\">here<\/a>. This will make this conversion process very faster than Pandas and will give you less compatibility issues to work with TF classes. If you use this function, then you will not need to read the csv file row by row but all at once using <code>map()<\/code> which uses <code>eager execution<\/code>. <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/load_data\/tf_records#tfrecord_files_using_tfdata\" rel=\"nofollow noreferrer\">This<\/a> is a good tutorial to get started. <\/p>\n\n<p>Accidentally edited wrong section of the post<\/p>\n"}]}
+{"QuestionId":58126617,"AnswerCount":1,"Tags":"<tensorflow><object><tracking><bounding-box>","CreationDate":"2019-09-27T01:32:12.893","AcceptedAnswerId":null,"OwnerUserId":10598627.0,"Title":"Bounding box coordinates via video object detection","Body":"<p>Is it possible to get the coordinates of the detected object through video object detection?  I used OpenCV to change video to images and LabelImg to put the bounding box to them. For the output, I want to be able to read (for a video file) the coordinates of the bounding box so I could get the centre of the box.<\/p>\n","answers":[{"AnswerId":"58388708","CreationDate":"2019-10-15T06:45:53.883","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>Object Detection works on per image or per frame basis. <\/p>\n\n<p>The basic object detection model takes image as input and gives back the detected bounding boxes. <\/p>\n\n<p>Now, what you need to do is, once your model is trained, read a video, frame by frame, or by skippng 'n' frames, pass the frame or image to the object detector and get the coordinates from it and show it into the output video frame. <\/p>\n\n<p>That is how object detection works for a video. <\/p>\n\n<p>Please refer to below links for references : \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>\n<a href=\"https:\/\/www.edureka.co\/blog\/tensorflow-object-detection-tutorial\/\" rel=\"nofollow noreferrer\">https:\/\/www.edureka.co\/blog\/tensorflow-object-detection-tutorial\/<\/a><\/p>\n"}]}
+{"QuestionId":58126678,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-datasets><tensorflow-estimator>","CreationDate":"2019-09-27T01:41:24.173","AcceptedAnswerId":null,"OwnerUserId":2930156.0,"Title":"Tensorflow steps\/sec gradually drops","Body":"<p>Hi Tensorflow experts:<\/p>\n\n<p>I have noticed that steps\/sec of my training job with multiple workers and parameter servers tend to stabilize to a lower value than initially. Is there any explanation for that? I don't think it has to do with evaluators since I configured them to start after 1000 seconds. Also I am using tensorflow 1.12 with estimator and dataset API. Thanks!<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ehFm9.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ehFm9.png\" alt=\"global steps \/ sec plot\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":58126978,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2019-09-27T02:30:39.603","AcceptedAnswerId":null,"OwnerUserId":9187036.0,"Title":"Errors running Tensorflow on Windows with Python 3.7.4","Body":"<p><strong>PROBLEM REGARDING TFLEARN AND PYTHON 3.7.4<\/strong>\ni believe that my code is fine but i am facing tflearn OR TENSORFLOW  issue in my python 3.7.4 how can i solve it?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/1APL6.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1APL6.jpg\" alt=\"1st image is  here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/G4ISZ.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/G4ISZ.jpg\" alt=\"2nd image is here\"><\/a><\/p>\n","answers":[{"AnswerId":"58388643","CreationDate":"2019-10-15T06:41:54.273","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>Based on the screenshots you have put in here, looks like tensorflow does not have any issue, it is just showing you warnings. <\/p>\n\n<p>The error you have got is because nltk requires a library punkt. Please download it using<\/p>\n\n<pre><code>nltk.download('punkt')\n<\/code><\/pre>\n\n<p>inside the python console and it should solve your issue.<\/p>\n"}]}
+{"QuestionId":58127059,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><transformer><attention-model>","CreationDate":"2019-09-27T02:40:48.810","AcceptedAnswerId":null,"OwnerUserId":12128213.0,"Title":"How to understand masked multi-head attention in transformer","Body":"<p>I'm currently studying code of transformer, but I can not understand the masked multi-head of decoder. The paper said that it is to prevent you from seeing the generating word, but I can not unserstand if the words after generating word have not been generated, how can them be seen? <\/p>\n\n<p>I try to read the code of transformer (link:<a href=\"https:\/\/github.com\/Kyubyong\/transformer\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Kyubyong\/transformer<\/a>). The code achieved mask is shown below. It uses the lower triangular matrix to mask, I can not understand why.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>padding_num = -2 ** 32 + 1\ndiag_vals = tf.ones_like(inputs[0, :, :])  # (T_q, T_k)\ntril = tf.linalg.LinearOperatorLowerTriangular(diag_vals).to_dense()  # (T_q, T_k)\nmasks = tf.tile(tf.expand_dims(tril, 0), [tf.shape(inputs)[0], 1, 1])  # (N, T_q, T_k)\npaddings = tf.ones_like(masks) * padding_num\noutputs = tf.where(tf.equal(masks, 0), paddings, inputs)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58127277,"AnswerCount":1,"Tags":"<python-2.7><tensorflow><keras>","CreationDate":"2019-09-27T03:08:18.730","AcceptedAnswerId":null,"OwnerUserId":5825021.0,"Title":"Matrix size-incompatible between Flatten and Dense layer","Body":"<p>I'm trying to implement multi-agent A2C algorithm with Keras.<br>\nThere is an issue when I use Keras backend function (K.fuction).  <\/p>\n\n<p>I gave input whose shape is (None, 3, 4, 5).<br>\nThis input passes through <code>Flatten()<\/code>, and then it is connected to <code>Dense()<\/code>.  <\/p>\n\n<p>It works fine when I use <code>model.predict(input)<\/code>, but a problem happened when I use customized optimizer using K.function.<\/p>\n\n<p>I gave input whose shape is (1000, 3, 4, 5) into <code>model.input<\/code> in K.function, but Matrix size-incompatible error appear:  <\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: Matrix size-incompatible: In[0]: [3,20], In[1]: [60,64]\n    [[{{node dense_1\/Relu}}]]\n<\/code><\/pre>\n\n<p>It looks <code>Flatten()<\/code> doesn't work properly.  <\/p>\n\n<p>Network looks like this:  <\/p>\n\n<pre><code>def build_network(self):\n    inp = Input(shape=(3,4,5))\n    x = Flatten()(inp)\n    x = Dense(64, activation='relu')(x)\n    x = Dense(128, activation='relu')(x)\n    x = Dense(128, activation='relu')(x)\n    out = Dense(9, activation='softmax')(x)\n\n    return Model(inp, out)\n\nself.model = self.build_network()\n<\/code><\/pre>\n\n<p><em>model.summary()<\/em> :<br>\n<a href=\"https:\/\/i.stack.imgur.com\/iI5UH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/iI5UH.png\" alt=\"\"><\/a>  <\/p>\n\n<p>The customized optimizer is here:  <\/p>\n\n<pre><code>import keras.backend as K\n\ndef optimizer(self):\n    weighted_actions = K.sum(self.action_pl * self.model.ouput, axis=1)\n    eligibility = K.log(weighted_actions + 1e-10) * K.stop_gradient(self.advantages_pl)\n    entropy = K.sum(self.model.output * K.log(self.model.output + 1e-10), axis=1)\n    loss = 0.001 * entropy - K.sum(eligibility)\n\n    _out = K.zeros(shape=(1))\n    updates = self.rms_optimizer.get_updates(self.model.trainable_weights, [], loss)\n\n    return K.fucntion([self.model.input, self.action_pl, self.advantages_pl], _out, updates=updates)\n\nself.opt = self.optimizer()\n\nself.opt([input, actions, advantages]) # Error occured here. input.shape = (1000,3,4,5)\n<\/code><\/pre>\n\n<p>I don't know why it happens and how to solve this issue.<\/p>\n\n<p>Any advice will be appreciated.<\/p>\n\n<p>Environment:\nUbuntu16.04\nPython2.7\nKeras: 2.3.0\ntensorflow 1.14.0<\/p>\n","answers":[{"AnswerId":"58391240","CreationDate":"2019-10-15T09:16:36.123","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>Issue seems to be that the input is taking of shape [3,20], but it should have been [1000, 60] after flatten. <\/p>\n\n<p>Please check the shape of your input. Also you can use <\/p>\n\n<pre><code>tf.expand_dims(\n    input,\n    axis,\n    name=None\n)\n<\/code><\/pre>\n\n<p>to add dimension for your input and try. It should resolve your issue.<\/p>\n"}]}
+{"QuestionId":58127401,"AnswerCount":1,"Tags":"<cuda><pytorch><torchvision>","CreationDate":"2019-09-27T03:22:08.623","AcceptedAnswerId":58127855.0,"OwnerUserId":8410477.0,"Title":"libcudart.so.9.0: cannot open shared object file: No such file or directory","Body":"<p>I'm using Pytorch  under Ubuntu 18.04 and trying to <code>import torchvision<\/code>, but I get an error <code>libcudart.so.9.0: cannot open shared object file: No such file or directory<\/code>.<\/p>\n\n<p>Someone could help to fix it? Thanks.<\/p>\n\n<p>The infos below are detailed error logs:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-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-2-6dd351122000&gt;\", line 1, in &lt;module&gt;\n    import torchvision\n  File \"\/home\/x\/pycharm-2019.2\/helpers\/pydev\/_pydev_bundle\/pydev_import_hook.py\", line 21, in do_import\n    module = self._system_import(name, *args, **kwargs)\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/torchvision\/__init__.py\", line 1, in &lt;module&gt;\n    from torchvision import models\n  File \"\/home\/x\/pycharm-2019.2\/helpers\/pydev\/_pydev_bundle\/pydev_import_hook.py\", line 21, in do_import\n    module = self._system_import(name, *args, **kwargs)\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/torchvision\/models\/__init__.py\", line 11, in &lt;module&gt;\n    from . import detection\n  File \"\/home\/x\/pycharm-2019.2\/helpers\/pydev\/_pydev_bundle\/pydev_import_hook.py\", line 21, in do_import\n    module = self._system_import(name, *args, **kwargs)\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/torchvision\/models\/detection\/__init__.py\", line 1, in &lt;module&gt;\n    from .faster_rcnn import *\n  File \"\/home\/x\/pycharm-2019.2\/helpers\/pydev\/_pydev_bundle\/pydev_import_hook.py\", line 21, in do_import\n    module = self._system_import(name, *args, **kwargs)\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/torchvision\/models\/detection\/faster_rcnn.py\", line 7, in &lt;module&gt;\n    from torchvision.ops import misc as misc_nn_ops\n  File \"\/home\/x\/pycharm-2019.2\/helpers\/pydev\/_pydev_bundle\/pydev_import_hook.py\", line 21, in do_import\n    module = self._system_import(name, *args, **kwargs)\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/torchvision\/ops\/__init__.py\", line 1, in &lt;module&gt;\n    from .boxes import nms, box_iou\n  File \"\/home\/x\/pycharm-2019.2\/helpers\/pydev\/_pydev_bundle\/pydev_import_hook.py\", line 21, in do_import\n    module = self._system_import(name, *args, **kwargs)\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/torchvision\/ops\/boxes.py\", line 2, in &lt;module&gt;\n    from torchvision import _C\n  File \"\/home\/x\/pycharm-2019.2\/helpers\/pydev\/_pydev_bundle\/pydev_import_hook.py\", line 21, in do_import\n    module = self._system_import(name, *args, **kwargs)\nImportError: libcudart.so.9.0: cannot open shared object file: No such file or directory\nimport torch\nimport torch.nn as nn\nimport torchvision.transforms as transforms\nTraceback (most recent call last):\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-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-4-677acbcfae34&gt;\", line 1, in &lt;module&gt;\n    import torchvision.transforms as transforms\n  File \"\/home\/x\/pycharm-2019.2\/helpers\/pydev\/_pydev_bundle\/pydev_import_hook.py\", line 21, in do_import\n    module = self._system_import(name, *args, **kwargs)\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/torchvision\/__init__.py\", line 1, in &lt;module&gt;\n    from torchvision import models\n  File \"\/home\/x\/pycharm-2019.2\/helpers\/pydev\/_pydev_bundle\/pydev_import_hook.py\", line 21, in do_import\n    module = self._system_import(name, *args, **kwargs)\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/torchvision\/models\/__init__.py\", line 11, in &lt;module&gt;\n    from . import detection\n  File \"\/home\/x\/pycharm-2019.2\/helpers\/pydev\/_pydev_bundle\/pydev_import_hook.py\", line 21, in do_import\n    module = self._system_import(name, *args, **kwargs)\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/torchvision\/models\/detection\/__init__.py\", line 1, in &lt;module&gt;\n    from .faster_rcnn import *\n  File \"\/home\/x\/pycharm-2019.2\/helpers\/pydev\/_pydev_bundle\/pydev_import_hook.py\", line 21, in do_import\n    module = self._system_import(name, *args, **kwargs)\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/torchvision\/models\/detection\/faster_rcnn.py\", line 7, in &lt;module&gt;\n    from torchvision.ops import misc as misc_nn_ops\n  File \"\/home\/x\/pycharm-2019.2\/helpers\/pydev\/_pydev_bundle\/pydev_import_hook.py\", line 21, in do_import\n    module = self._system_import(name, *args, **kwargs)\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/torchvision\/ops\/__init__.py\", line 1, in &lt;module&gt;\n    from .boxes import nms, box_iou\n  File \"\/home\/x\/pycharm-2019.2\/helpers\/pydev\/_pydev_bundle\/pydev_import_hook.py\", line 21, in do_import\n    module = self._system_import(name, *args, **kwargs)\n  File \"\/home\/x\/.local\/lib\/python3.6\/site-packages\/torchvision\/ops\/boxes.py\", line 2, in &lt;module&gt;\n    from torchvision import _C\n  File \"\/home\/x\/pycharm-2019.2\/helpers\/pydev\/_pydev_bundle\/pydev_import_hook.py\", line 21, in do_import\n    module = self._system_import(name, *args, **kwargs)\nImportError: libcudart.so.9.0: cannot open shared object file: No such file or directory\n<\/code><\/pre>\n","answers":[{"AnswerId":"58127855","CreationDate":"2019-09-27T04:23:12.047","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p>If you are using <code>anaconda<\/code>, the following may fix your problem.<\/p>\n\n<pre><code>conda install -c anaconda cudatoolkit==9.0\n<\/code><\/pre>\n\n<hr>\n\n<p>You can also try the followings.<\/p>\n\n<p>Make sure the CUDA version is 9.0. And add the following 2 lines to <code>~\/.bashrc<\/code>.<\/p>\n\n<pre><code>export PATH=\/usr\/local\/cuda\/bin:$PATH\nexport LD_LIBRARY_PATH=\/usr\/local\/cuda\/lib64:$LD_LIBRARY_PATH\n<\/code><\/pre>\n\n<p>Then, run:<\/p>\n\n<pre><code>source  ~\/.bashrc\n<\/code><\/pre>\n\n<p>Add the following lines to <code>\/etc\/ld.so.conf.d\/cuda.conf<\/code><\/p>\n\n<pre><code>\/usr\/local\/cuda\/lib64\n<\/code><\/pre>\n\n<p>And run:<\/p>\n\n<pre><code>sudo ldconfig\n<\/code><\/pre>\n"}]}
+{"QuestionId":58127635,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-27T03:53:59.583","AcceptedAnswerId":58127971.0,"OwnerUserId":11181287.0,"Title":"Keras Tensorflow warning","Body":"<p>Can someone help me fix the following warning?<\/p>\n\n<p>I have looked to the backend.py but cant seem to work out what the problem is, and the solution?<\/p>\n\n<blockquote>\n  <p>WARNING:tensorflow:From \/Users\/temp\/Desktop\/python_stuff\/env\/lib\/python3.7\/site-packages\/tensorflow\/python\/keras\/backend.py:4149: add_dispatch_support..wrapper (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.\n  Instructions for updating:\n  Use tf.where in 2.0, which has the same broadcast rule as np.where<\/p>\n<\/blockquote>\n\n<p>I also have the following warning, but saw another question on stack, didn't understand it fully so I'll add to this question if anyone knows how to fix it...<\/p>\n\n<blockquote>\n  <p>tensorflow\/core\/platform\/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA<\/p>\n<\/blockquote>\n\n<p>I have used pip to install everything and virtual env for my development\nThank you.<\/p>\n","answers":[{"AnswerId":"58127971","CreationDate":"2019-09-27T04:38:55.863","ParentId":null,"OwnerUserId":"9828197","Title":null,"Body":"<p>I will quote my answer from this <a href=\"https:\/\/stackoverflow.com\/questions\/57951499\/depreciation-warnings-in-tensorflow-keras-official-tutorial\/57952272#57952272\">thread<\/a>:<\/p>\n\n<blockquote>\n  <p>tensorflow is notorious for not being backward compatible and worse yet there are warnings like this now for a long time.<\/p>\n  \n  <p>See here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/25996\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/25996<\/a><\/p>\n  \n  <p>So your best bet is just ignore those warnings and, if you wish to be proactive, then contribute with a fix. <\/p>\n<\/blockquote>\n\n<p>Your other queston has also been answered here and you may safely ignore such warning too, especially if you use GPU:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/47068709\/your-cpu-supports-instructions-that-this-tensorflow-binary-was-not-compiled-to-u\">Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX AVX2<\/a><\/p>\n"}]}
+{"QuestionId":58127685,"AnswerCount":1,"Tags":"<tensorflow><debugging><keras>","CreationDate":"2019-09-27T03:59:42.960","AcceptedAnswerId":58127823.0,"OwnerUserId":9276708.0,"Title":"How to avoid reduce_mean being an empty tensor in Keras or Tensorflow","Body":"<p>I use <code>reduce_mean<\/code> to get the mean value of <code>tensor_a<\/code>.\nIf <code>tensor_a<\/code> is an empty tensor, I get <code>nan<\/code> value.<\/p>\n\n<pre><code>tensor_a = K.variable([])\nprint(K.get_value(tf.reduce_mean(tensor_a)))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>nan\n<\/code><\/pre>\n\n<p>However, if <code>tensor_a<\/code> is a empty tensor, I'd like to get zero instead of <code>nan<\/code> value. \nHow to solve it?<\/p>\n","answers":[{"AnswerId":"58127823","CreationDate":"2019-09-27T04:17:50.287","ParentId":null,"OwnerUserId":"1699075","Title":null,"Body":"<p>Is this what you're looking for? Essentially as long as <code>tf.size(tensor_a)<\/code> is greater than 0, you'll get the <code>reduce_mean()<\/code> called. Otherwise, you get 0 out.<\/p>\n\n<pre><code>tensor_a = K.variable([])\n\nprint(\n  K.get_value(\n    tf.cond(\n      tf.equal(tf.size(tensor_a), 0), \n      lambda : tf.constant(0.0), lambda: tf.reduce_mean(tensor_a)\n    )\n  )\n)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58127902,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-27T04:28:35.843","AcceptedAnswerId":null,"OwnerUserId":9828197.0,"Title":"Why do I get Operation 'cond\/IsVariableInitialized' has been marked as not fetchable","Body":"<p>So when I run the below code:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom keras import backend as K\n\ntensor_a = K.variable([])\ndef return_zero():\n    return 0\ndef get_mean():\n    return K.get_value(K.tf.reduce_mean(tensor_a))\n\nis_equal = tf.equal(tf.size(tensor_a), 0)\nr = tf.cond(is_equal, return_zero, get_mean)\n\nprint(r)\n<\/code><\/pre>\n\n<p>I get the error:<\/p>\n\n<pre><code>ValueError: Operation 'cond\/IsVariableInitialized' has been marked as not fetchable.\n<\/code><\/pre>\n\n<p>However, if this code is run, there is no error. What gives?<\/p>\n\n<pre><code>tensor_a = K.variable([])\n\nprint(\n  K.get_value(\n    tf.cond(\n      tf.equal(tf.size(tensor_a), 0), \n      lambda : tf.constant(0.0), lambda: tf.reduce_mean(tensor_a)\n    )\n  )\n)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58391582","CreationDate":"2019-10-15T09:35:09.737","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>tf.cond takes operations as lambda functions. <\/p>\n\n<p>Please read <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4094\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4094<\/a> for more details on it<\/p>\n"}]}
+{"QuestionId":58127997,"AnswerCount":0,"Tags":"<python><tensorflow><classification><conv-neural-network>","CreationDate":"2019-09-27T04:42:33.133","AcceptedAnswerId":null,"OwnerUserId":2908528.0,"Title":"Classification Module from Capsule Network Code","Body":"<p>Refer to the <a href=\"https:\/\/github.com\/ageron\/handson-ml\/blob\/master\/extra_capsnets.ipynb\" rel=\"nofollow noreferrer\">Capsule Network Code<\/a>, I am using just the classification module from the mentioned code, So following is the complete classification code that I extracted from the link.<\/p>\n\n<pre><code>from __future__ import division, print_function, unicode_literals\nimport matplotlib\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport tensorflow as tf\n\ntf.reset_default_graph()\n\nnp.random.seed(42)\ntf.set_random_seed(42)\n\nfrom tensorflow.examples.tutorials.mnist import input_data\nmnist = input_data.read_data_sets(\"\/tmp\/data\/\")\n\nX = tf.placeholder(shape=[None, 28, 28, 1], dtype=tf.float32, name=\"X\")\n\ncaps1_n_maps = 32\ncaps1_n_caps = caps1_n_maps * 6 * 6  # 1152 primary capsules\ncaps1_n_dims = 8\n\nconv1_params = {\n    \"filters\": 256,\n    \"kernel_size\": 9,\n    \"strides\": 1,\n    \"padding\": \"valid\",\n    \"activation\": tf.nn.relu,\n}\n\nconv2_params = {\n    \"filters\": caps1_n_maps * caps1_n_dims, # 256 convolutional filters\n    \"kernel_size\": 9,\n    \"strides\": 2,\n    \"padding\": \"valid\",\n    \"activation\": tf.nn.relu\n}\n\nconv1 = tf.layers.conv2d(X, name=\"conv1\", **conv1_params)\nconv2 = tf.layers.conv2d(conv1, name=\"conv2\", **conv2_params)\n\ncaps1_raw = tf.reshape(conv2, [-1, caps1_n_caps, caps1_n_dims],name=\"caps1_raw\")\n\n\ndef squash(s, axis=-1, epsilon=1e-7, name=None):\n    with tf.name_scope(name, default_name=\"squash\"):\n        squared_norm = tf.reduce_sum(tf.square(s), axis=axis,keep_dims=True)\n        safe_norm = tf.sqrt(squared_norm + epsilon)\n        squash_factor = squared_norm \/ (1. + squared_norm)\n        unit_vector = s \/ safe_norm\n        return squash_factor * unit_vector\n\ncaps1_output = squash(caps1_raw, name=\"caps1_output\")\n\ncaps2_n_caps = 10\ncaps2_n_dims = 16\n\ninit_sigma = 0.1\n\nW_init = tf.random_normal(\n    shape=(1, caps1_n_caps, caps2_n_caps, caps2_n_dims, caps1_n_dims),\n    stddev=init_sigma, dtype=tf.float32, name=\"W_init\")\nW = tf.Variable(W_init, name=\"W\")\n\nbatch_size = tf.shape(X)[0]\nW_tiled = tf.tile(W, [batch_size, 1, 1, 1, 1], name=\"W_tiled\")\n\ncaps1_output_expanded = tf.expand_dims(caps1_output, -1,\n                                       name=\"caps1_output_expanded\")\ncaps1_output_tile = tf.expand_dims(caps1_output_expanded, 2,\n                                   name=\"caps1_output_tile\")\ncaps1_output_tiled = tf.tile(caps1_output_tile, [1, 1, caps2_n_caps, 1, 1],\n                             name=\"caps1_output_tiled\")\n\ncaps2_predicted = tf.matmul(W_tiled, caps1_output_tiled,\n                            name=\"caps2_predicted\")\n\nraw_weights = tf.zeros([batch_size, caps1_n_caps, caps2_n_caps, 1, 1],\n                       dtype=np.float32, name=\"raw_weights\")\n\n#ROUND 1\n\nrouting_weights = tf.nn.softmax(raw_weights, dim=2, name=\"routing_weights\")\n\nweighted_predictions = tf.multiply(routing_weights, caps2_predicted,\n                                   name=\"weighted_predictions\")\nweighted_sum = tf.reduce_sum(weighted_predictions, axis=1, keep_dims=True,\n                             name=\"weighted_sum\")\n\ncaps2_output_round_1 = squash(weighted_sum, axis=-2,\n                              name=\"caps2_output_round_1\")\n\n\n#ROUND 2\ncaps2_output_round_1_tiled = tf.tile(\n    caps2_output_round_1, [1, caps1_n_caps, 1, 1, 1],\n    name=\"caps2_output_round_1_tiled\")\n\nagreement = tf.matmul(caps2_predicted, caps2_output_round_1_tiled,\n                      transpose_a=True, name=\"agreement\")\n\nraw_weights_round_2 = tf.add(raw_weights, agreement,\n                             name=\"raw_weights_round_2\")\n\nrouting_weights_round_2 = tf.nn.softmax(raw_weights_round_2,\n                                        dim=2,\n                                        name=\"routing_weights_round_2\")\nweighted_predictions_round_2 = tf.multiply(routing_weights_round_2,\n                                           caps2_predicted,\n                                           name=\"weighted_predictions_round_2\")\nweighted_sum_round_2 = tf.reduce_sum(weighted_predictions_round_2,\n                                     axis=1, keep_dims=True,\n                                     name=\"weighted_sum_round_2\")\ncaps2_output_round_2 = squash(weighted_sum_round_2,\n                              axis=-2,\n                              name=\"caps2_output_round_2\")\n\ncaps2_output = caps2_output_round_2\n#ESTIMATE CLASS PROBABILITIES\n\ndef safe_norm(s, axis=-1, epsilon=1e-7, keep_dims=False, name=None):\n    with tf.name_scope(name, default_name=\"safe_norm\"):\n        squared_norm = tf.reduce_sum(tf.square(s), axis=axis,\n                                     keep_dims=keep_dims)\n        return tf.sqrt(squared_norm + epsilon)\n\ny_proba = safe_norm(caps2_output, axis=-2, name=\"y_proba\")\ny_proba_argmax = tf.argmax(y_proba, axis=2, name=\"y_proba\")\n\ny_pred = tf.squeeze(y_proba_argmax, axis=[1,2], name=\"y_pred\")\n\n\ny = tf.placeholder(shape=[None], dtype=tf.int64, name=\"y\")\n\nm_plus = 0.9\nm_minus = 0.1\nlambda_ = 0.5\n\nT = tf.one_hot(y, depth=caps2_n_caps, name=\"T\")\n\ncaps2_output_norm = safe_norm(caps2_output, axis=-2, keep_dims=True,\n                              name=\"caps2_output_norm\")\n\npresent_error_raw = tf.square(tf.maximum(0., m_plus - caps2_output_norm),\n                              name=\"present_error_raw\")\npresent_error = tf.reshape(present_error_raw, shape=(-1, 10),\n                           name=\"present_error\")\n\nabsent_error_raw = tf.square(tf.maximum(0., caps2_output_norm - m_minus),\n                             name=\"absent_error_raw\")\nabsent_error = tf.reshape(absent_error_raw, shape=(-1, 10),\n                          name=\"absent_error\")\n\nL = tf.add(T * present_error, lambda_ * (1.0 - T) * absent_error,\n           name=\"L\")\nmargin_loss = tf.reduce_mean(tf.reduce_sum(L, axis=1), name=\"margin_loss\")\n\n\nloss = tf.add(margin_loss, 0, name=\"loss\")\n\ncorrect = tf.equal(y, y_pred, name=\"correct\")\naccuracy = tf.reduce_mean(tf.cast(correct, tf.float32), name=\"accuracy\")\n\noptimizer = tf.train.AdamOptimizer()\ntraining_op = optimizer.minimize(loss, name=\"training_op\")\n\ninit = tf.global_variables_initializer()\nsaver = tf.train.Saver()\n\nn_epochs = 10\nbatch_size = 50\nrestore_checkpoint = True\n\nn_iterations_per_epoch = mnist.train.num_examples \/\/ batch_size\nn_iterations_validation = mnist.validation.num_examples \/\/ batch_size\nbest_loss_val = np.infty\ncheckpoint_path = \".\/my_capsule_network\"\n\nwith tf.Session() as sess:\n    if restore_checkpoint and tf.train.checkpoint_exists(checkpoint_path):\n        saver.restore(sess, checkpoint_path)\n    else:\n        init.run()\n\n    for epoch in range(n_epochs):\n        for iteration in range(1, n_iterations_per_epoch + 1):\n            X_batch, y_batch = mnist.train.next_batch(batch_size)\n            # Run the training operation and measure the loss:\n            _, loss_train = sess.run(\n                [training_op, loss],\n                feed_dict={X: X_batch.reshape([-1, 28, 28, 1]),\n                           y: y_batch})\n            print(\"\\rIteration: {}\/{} ({:.1f}%)  Loss: {:.5f}\".format(\n                      iteration, n_iterations_per_epoch,\n                      iteration * 100 \/ n_iterations_per_epoch,\n                      loss_train),\n                  end=\"\")\n\n        # At the end of each epoch,\n        # measure the validation loss and accuracy:\n        loss_vals = []\n        acc_vals = []\n        for iteration in range(1, n_iterations_validation + 1):\n            X_batch, y_batch = mnist.validation.next_batch(batch_size)\n            loss_val, acc_val = sess.run(\n                    [loss, accuracy],\n                    feed_dict={X: X_batch.reshape([-1, 28, 28, 1]),\n                               y: y_batch})\n            loss_vals.append(loss_val)\n            acc_vals.append(acc_val)\n            print(\"\\rEvaluating the model: {}\/{} ({:.1f}%)\".format(\n                      iteration, n_iterations_validation,\n                      iteration * 100 \/ n_iterations_validation),\n                  end=\" \" * 10)\n        loss_val = np.mean(loss_vals)\n        acc_val = np.mean(acc_vals)\n        print(\"\\rEpoch: {}  Val accuracy: {:.4f}%  Loss: {:.6f}{}\".format(\n            epoch + 1, acc_val * 100, loss_val,\n            \" (improved)\" if loss_val &lt; best_loss_val else \"\"))\n\n        # And save the model if it improved:\n        if loss_val &lt; best_loss_val:\n            save_path = saver.save(sess, checkpoint_path)\n            best_loss_val = loss_val\n\n\n\nn_iterations_test = mnist.test.num_examples \/\/ batch_size\n\nwith tf.Session() as sess:\n    saver.restore(sess, checkpoint_path)\n\n    loss_tests = []\n    acc_tests = []\n    for iteration in range(1, n_iterations_test + 1):\n        X_batch, y_batch = mnist.test.next_batch(batch_size)\n        loss_test, acc_test = sess.run(\n                [loss, accuracy],\n                feed_dict={X: X_batch.reshape([-1, 28, 28, 1]),\n                           y: y_batch})\n        loss_tests.append(loss_test)\n        acc_tests.append(acc_test)\n        print(\"\\rEvaluating the model: {}\/{} ({:.1f}%)\".format(\n                  iteration, n_iterations_test,\n                  iteration * 100 \/ n_iterations_test),\n              end=\" \" * 10)\n    loss_test = np.mean(loss_tests)\n    acc_test = np.mean(acc_tests)\n    print(\"\\rFinal test accuracy: {:.4f}%  Loss: {:.6f}\".format(\n        acc_test * 100, loss_test))\n<\/code><\/pre>\n\n<p>Since reconstruction module is not required so in the final loss function I have set loss = 0 for the reconstruction part so that optimizer only work on the classification loss.<\/p>\n\n<pre><code>loss = tf.add(margin_loss, 0, name=\"loss\")\n<\/code><\/pre>\n\n<p>Now the classification module is working but the accuracy is pathetically low even for MNIST dataset, I got around 10% validation accuracy, Can somebody explain to me what is the problem here because on MNIST the Capsule Networks perform well in classification as discussed in <a href=\"https:\/\/pdfs.semanticscholar.org\/de7f\/27677ae04bf1f09c223f68cea71af90be3d4.pdf\" rel=\"nofollow noreferrer\">Paper<\/a>?<\/p>\n\n<p>Regards<\/p>\n","answers":[]}
+{"QuestionId":58128078,"AnswerCount":1,"Tags":"<python><tensorflow><keras><deep-learning><conv-neural-network>","CreationDate":"2019-09-27T04:54:10.603","AcceptedAnswerId":null,"OwnerUserId":9813733.0,"Title":"Keras VGG16 modified model giving the same prediction every time","Body":"<p>I have taken the inbuilt keras.applications.vgg16.VGG16(weights='imagenet', include_top=True,input_shape=(224,224,3)) model and did transfer learning for PASCAL VOC 2012 dataset with 20 classes by including Global Average Pooling Layer as given below:<\/p>\n\n<pre><code>def VGG16_modified():\n    base_model = vgg16.VGG16(include_top=True,weights='imagenet',input_shape=(224,224,3))\n    print(base_model.summary())\n    x = base_model.get_layer('block5_pool').output\n    x = (GlobalAveragePooling2D())(x)\n    predictions = Dense(20,activation='sigmoid')(x)\n\n    final_model = Model(input = base_model.input, output = predictions)\n    print(final_model.get_weights())\n    return final_model\n<\/code><\/pre>\n\n<p>Now, I want to take the Class Activation Map based on <a href=\"http:\/\/cnnlocalization.csail.mit.edu\/Zhou_Learning_Deep_Features_CVPR_2016_paper.pdf\" rel=\"nofollow noreferrer\">this<\/a> paper. For this, my code is as given below:<\/p>\n\n<pre><code> def get_CAM(model,img):\n        model = load_model(model)\n        im = image.load_img(img,target_size=(224,224))\n        im = image.img_to_array(im)\n        im = np.expand_dims(im,axis=0)\n        class_weights = model.layers[-1].get_weights()[0]\n        final_conv_layer = model.get_layer('block5_pool')\n        cam_model = Model(inputs = model.input,outputs=(final_conv_layer.output,model.layers[-1].output))\n        conv_outputs, predictions = cam_model.predict(im)\n        conv_outputs = np.squeeze(conv_outputs)\n        prediction = np.argmax(predictions)\n        print(predictions)\n        print(prediction)\n        print(conv_outputs)\n        print(conv_outputs.shape)\n        class_weights = class_weights[:,prediction]\n        mat_for_mult = scipy.ndimage.zoom(conv_outputs,(32,32,1),order=1)\n        final_output = np.dot(mat_for_mult.reshape((224*224, 512)),class_weights).reshape((224,224))\n        print(final_output)\n        return final_output\n<\/code><\/pre>\n\n<p>But <strong><em>cam_model.predict(im)<\/em><\/strong> is always giving the same class for all images. I am not sure where have I wrong with this. As the pascal voc 2012 contains multi label images, I have used 'sigmoid' in the last layer of the modified_vgg16 rather than 'softmax'. Can you let me know where have I gone wrong. <\/p>\n","answers":[{"AnswerId":"58390920","CreationDate":"2019-10-15T09:01:04.077","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>PASCAL VOC has imbalanced distribution of images between classses. <\/p>\n\n<p>Most of the images in dataset are for person which leads to imbalance data distribution and creates bias towards person class. <\/p>\n\n<p>In order to avoid it, you can use the method of weighting the classes which helps to remove imbalance between classes<\/p>\n\n<p>Please refer to <a href=\"https:\/\/datascience.stackexchange.com\/questions\/13490\/how-to-set-class-weights-for-imbalanced-classes-in-keras\">this<\/a> for more information on how to set class weights for imbalanced class.<\/p>\n"}]}
+{"QuestionId":58128970,"AnswerCount":3,"Tags":"<pytorch><conda>","CreationDate":"2019-09-27T06:28:53.923","AcceptedAnswerId":null,"OwnerUserId":8417524.0,"Title":"Conda install of pytorch fails","Body":"<p>I created an environment with conda and I want to install pytorch in it, but it doesn't work. After I get inside my environment with <code>source activate env_name<\/code> I tried this: <code>conda install pytorch torchvision -c pytorch<\/code> (I also tried it like this: <code>conda install -c pytorch pytorch torchvision<\/code>) but I am getting this error:<\/p>\n\n<pre><code>Using Anaconda Cloud api site https:\/\/api.anaconda.org\nFetching package metadata: ......\nSolving package specifications: ......\nError: Could not find some dependencies for pytorch: mkl &gt;=2018, cudatoolkit &gt;=9.0,&lt;9.1, blas * mkl, cudatoolkit &gt;=10.0,&lt;10.1, cudatoolkit &gt;=9.2,&lt;9.3, blas * openblas, cudnn 7.0.*, cudatoolkit 9.*\n\nDid you mean one of these?\n\n    pytorch, pytorch-gpu, pytorch-cpu\n\nDid you mean one of these?\n\n    cudatoolkit\n\nYou can search for this package on anaconda.org with\n\n    anaconda search -t conda cudatoolkit 9.*\n\n (and similarly for the other packages)\n<\/code><\/pre>\n\n<p>Here are my installed packages:<\/p>\n\n<pre><code>backports                 1.0                      py34_0  \nbackports.shutil-get-terminal-size 1.0.0                     &lt;pip&gt;\ndecorator                 4.0.11                   py34_0  \nget_terminal_size         1.0.0                    py34_0  \nipython                   4.2.0                    py34_0  \nipython-genutils          0.1.0                     &lt;pip&gt;\nipython_genutils          0.1.0                    py34_0  \nlibgfortran               1.0                           0  \nnumpy                     1.9.2                    py34_0  \nopenssl                   1.0.2l                        0  \npath.py                   10.0                     py34_0  \npexpect                   4.2.1                    py34_0  \npickleshare               0.7.4                    py34_0  \npip                       9.0.1                    py34_1  \nptyprocess                0.5.1                    py34_0  \npython                    3.4.5                         0  \nreadline                  6.2                           2  \nscipy                     0.16.0               np19py34_0  \nsetuptools                27.2.0                   py34_0  \nsimplegeneric             0.8.1                    py34_1  \nsix                       1.10.0                   py34_0  \nsqlite                    3.13.0                        0  \ntk                        8.5.18                        0  \ntraitlets                 4.3.1                    py34_0  \nwheel                     0.29.0                   py34_0  \nxz                        5.2.3                         0  \nzlib                      1.2.11                        0  \n<\/code><\/pre>\n\n<p>What should I do? Thank you!<\/p>\n","answers":[{"AnswerId":"58133670","CreationDate":"2019-09-27T11:29:00.443","ParentId":null,"OwnerUserId":"9427019","Title":null,"Body":"<p>Please try the following steps.It worked fine for me.<\/p>\n\n<p><code>source activate env_name\nconda install -c pytorch pytorch<\/code><\/p>\n\n<p>open python shell\n<code>import torch<\/code><\/p>\n"},{"AnswerId":"58129992","CreationDate":"2019-09-27T07:40:39.527","ParentId":null,"OwnerUserId":"5101148","Title":null,"Body":"<p>I can't give you a definite answer cause you didn't provided the info about the Python version, platform you're using.<\/p>\n\n<p>Go to the <a href=\"https:\/\/pytorch.org\/\" rel=\"nofollow noreferrer\">official website for Pytorch<\/a>, choose a installation method according to your platform, Python version and whether you need CUDA.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/cmtBx.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/cmtBx.png\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"58149814","CreationDate":"2019-09-28T19:53:12.200","ParentId":null,"OwnerUserId":"570918","Title":null,"Body":"<p>Pytorch's <code>vision<\/code> package (aka <code>torchvision<\/code>) was developed post-Python 3.4, and so only has versions supporting Python 2.7, 3.5-7. Please create a new environment with a later Python version. Note it is always better to include the packages you care about in the creation of the environment, e.g., <\/p>\n\n<pre><code>conda create -n env_name -c pytorch torchvision\n<\/code><\/pre>\n\n<p>and Conda will figure the rest out. If you need to have a specific version of Python, you can include that as well (e.g., <code>python=3.6<\/code>).<\/p>\n"}]}
+{"QuestionId":58129371,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-27T06:58:44.410","AcceptedAnswerId":null,"OwnerUserId":12128882.0,"Title":"Tenseflow and Keras have different results for the same CNN source","Body":"<p>I used CNN method with tensorflow and got an appropriate result.(like 95% acc)\nNext, I used same model structure with keras. I got a different result. It show horrible accuray and it didn't got better while epochs.<\/p>\n\n<p>I changed lr, pooling_size, filters, etc. but not going better.\nI used keras model for mnist. this model works well at mnist dataset.<\/p>\n\n<p>This is CNN model with tensorflow:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import what i needed #(pseudo)\n\ndata_list = glob.glob('images dir') #(pseudo)\ntf.set_random_seed(777)\nrandom.shuffule(data_list)\n\n# hyper parameters\nlearning_rate = 0.001\ntraining_epochs = 10\nbatch_size = 64\nbatch_size_test = 1000\n\n# dropout\nkeep_prob = tf.placeholder(tf.float32)\n\ntrain_x, train_y, test_x, test_y = data_list #(pseudo)\n#image data shape is [none, 256, 256, 3]\n#result shape is [4, ]\n#train set: 500, test set: 230 ... (i know the tiny resource, but getting more data is not available.)\n\n#model\nclass Model:\n    def __init__(self, sess, name):\n        self.sess = sess\n        self.name = name\n        self._build_net()\n\n    def _build_net(self):\n        with tf.variable_scope(self.name):\n            # dropout (keep_prob) rate  0.7~0.5 on training, but should be 1\n            # for testing\n            self.training = tf.placeholder(tf.bool)\n\n            # input place holders\n            self.X = tf.placeholder(tf.float32, [None, 256, 256, 3])\n\n            # img 256x256x3 , Input Layer\n            X_img = tf.reshape(self.X, [-1, 256, 256, 3])\n            self.Y = tf.placeholder(tf.float32, [None, 4])\n\n            # Convolutional Layer #1\n            conv1 = tf.layers.conv2d(inputs=X_img, filters=32, kernel_size=[3, 3],\n                                     padding=\"SAME\", activation=tf.nn.relu)\n            # Pooling Layer #1\n            pool1 = tf.layers.max_pooling2d(inputs=conv1, pool_size=[2, 2],\n                                            padding=\"SAME\", strides=2)\n            dropout1 = tf.layers.dropout(inputs=pool1,\n                                         rate=0.3, training=self.training)\n            #None*256*256*3 -&gt; 256*256*32 -&gt; 128*128*32\n\n            # Convolutional Layer #2 and Pooling Layer #2\n            conv2 = tf.layers.conv2d(inputs=dropout1, filters=64, kernel_size=[3, 3],\n                                     padding=\"SAME\", activation=tf.nn.relu)\n            pool2 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2],\n                                            padding=\"SAME\", strides=2)\n            dropout2 = tf.layers.dropout(inputs=pool2,\n                                         rate=0.3, training=self.training)\n            #128*128*32 -&gt; 128*128*64 -&gt; 64*64*64\n\n            # Convolutional Layer #2 and Pooling Layer #2\n            conv3 = tf.layers.conv2d(inputs=dropout2, filters=128, kernel_size=[3, 3],\n                                     padding=\"same\", activation=tf.nn.relu)\n            pool3 = tf.layers.max_pooling2d(inputs=conv3, pool_size=[2, 2],\n                                            padding=\"same\", strides=2)\n            dropout3 = tf.layers.dropout(inputs=pool3,\n                                         rate=0.3, training=self.training)\n            #64*64*64 -&gt; 64*64*128 -&gt; 32*32*128\n\n            # Dense Layer with Relu\n            flat = tf.reshape(dropout3, [-1, 128 * 32 * 32])\n            dense4 = tf.layers.dense(inputs=flat,\n                                     units=625, activation=tf.nn.relu)\n            dropout4 = tf.layers.dropout(inputs=dense4,\n                                         rate=0.5, training=self.training)\n\n            # Logits (no activation) Layer: L5 Final FC 625 inputs -&gt; 4 outputs\n            self.logits = tf.layers.dense(inputs=dropout4, units=4)\n\n        # define cost\/loss &amp; optimizer\n        self.cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(\n            logits=self.logits, labels=self.Y))\n        self.optimizer = tf.train.AdamOptimizer(\n            learning_rate=learning_rate).minimize(self.cost)\n\n        correct_prediction = tf.equal(\n            tf.argmax(self.logits, 1), tf.argmax(self.Y, 1))\n        self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n\n    def predict(self, x_test, training=False):\n        return self.sess.run(self.logits,\n                             feed_dict={self.X: x_test, self.training: training})\n\n    def get_accuracy(self, x_test, y_test, training=False):\n        return self.sess.run(self.accuracy,\n                             feed_dict={self.X: x_test,\n                                        self.Y: y_test, self.training: training})\n\n    def train(self, x_data, y_data, training=True):\n        return self.sess.run([self.cost, self.optimizer], feed_dict={\n            self.X: x_data, self.Y: y_data, self.training: training})\n\n#-----------------session-----------------------------------------\nsess = tf.Session()\nm1 = Model(sess, \"m1\")\n\nsess.run(tf.global_variables_initializer())\n#-----------------------------------------------------------------\nprint('Learning Started!')\n\n# train my model\nfor epoch in range(training_epochs):\n    cost_result = 0\n    total_batch = int(len(data_list)\/batch_size)\n\n    pbar = tqdm(range(0, len(data_list), batch_size))\n    for i in pbar:\n        train_x, train_y = get_next_data_set(i)\n        c, _ = m1.train(train_x, train_y)\n        cost_result += c \/total_batch\n        pbar.set_description(\"cost : %f\" % cost_result)\n    print('Accuracy:', m1.get_accuracy(train_x, train_y))\n\nprint('Learning Finished!')\n\n<\/code><\/pre>\n\n<p>and result is:\nLearning Started!\ncost : 4465.586227: 100%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 8\/8 [02:55&lt;00:00, 20.01s\/it]<\/p>\n\n<p>Accuracy: 0.1923077<\/p>\n\n<p>cost : 19.377403: 100%|\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| 8\/8 [02:28&lt;00:00, 17.93s\/it]<\/p>\n\n<p>Accuracy: 0.78846157<\/p>\n\n<p>cost : 0.977960: 100%|\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| 8\/8 [02:29&lt;00:00, 18.21s\/it]<\/p>\n\n<p>Accuracy: 0.3653846<\/p>\n\n<p>cost : 1.106399: 100%|\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| 8\/8 [02:30&lt;00:00, 18.09s\/it]<\/p>\n\n<p>Accuracy: 0.63461536<\/p>\n\n<p>cost : 0.896358: 100%|\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| 8\/8 [02:28&lt;00:00, 18.02s\/it]<\/p>\n\n<p>Accuracy: 0.84615386<\/p>\n\n<p>cost : 0.622414: 100%|\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| 8\/8 [02:33&lt;00:00, 18.56s\/it]<\/p>\n\n<p>Accuracy: 0.88461536<\/p>\n\n<p>cost : 0.521675: 100%|\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| 8\/8 [02:34&lt;00:00, 18.57s\/it]<\/p>\n\n<p>Accuracy: 0.8269231<\/p>\n\n<p>cost : 0.421264: 100%|\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| 8\/8 [02:22&lt;00:00, 17.09s\/it]<\/p>\n\n<p>Accuracy: 0.9423077<\/p>\n\n<p>cost : 0.314349: 100%|\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| 8\/8 [02:20&lt;00:00, 16.99s\/it]<\/p>\n\n<p>Accuracy: 0.9423077<\/p>\n\n<p>cost : 0.240451: 100%|\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| 8\/8 [02:25&lt;00:00, 17.43s\/it]<\/p>\n\n<p>Accuracy: 0.9423077<\/p>\n\n<p>Learning Finished!<\/p>\n\n<p>This is Keras: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import what i needed #(pseudo)\n\n#model\nmodel = keras.models.Sequential()\nmodel.add(keras.layers.Conv2D(filters = 32, \n                              kernel_size=(3,3), \n                              padding='same', \n                              activation='relu', \n                              input_shape=(256, 256, 3)))\nmodel.add(keras.layers.MaxPooling2D(pool_size=(3, 3),\n                                   strides=3,\n                                   padding='same'))\nmodel.add(keras.layers.Dropout(0.5))\nmodel.add(keras.layers.Conv2D(filters=64,\n                              kernel_size=(3,3), \n                              padding='same', \n                              activation='relu'))\nmodel.add(keras.layers.MaxPooling2D(pool_size=(3, 3),\n                                   strides=3,\n                                   padding='same'))\nmodel.add(keras.layers.Dropout(0.5))\nmodel.add(keras.layers.Conv2D(filters=128, \n                              kernel_size=(3,3), \n                              padding='same', \n                              activation='relu'))\nmodel.add(keras.layers.MaxPooling2D((3, 3),\n                                   strides=3,\n                                   padding='same'))\nmodel.add(keras.layers.Dropout(0.5))\nmodel.add(keras.layers.Flatten())\nmodel.add(keras.layers.Dense(625, activation='relu'))\nmodel.add(keras.layers.Dropout(0.8))\nmodel.add(keras.layers.Dense(4, activation='softmax'))\n\n#data set\ndata_list = glob.glob('images dir') #(pseudo)\ntf.set_random_seed(777)\nrandom.shuffule(data_list)\n\ntrain_x, train_y, test_x, test_y = data_list #(pseudo)\n#image data shape is [none, 256, 256, 3]\n#result shape is [4, ]\n#train set: 500, test set: 230 ... (i know the tiny resource, but getting more data is not available.)\n\nmodel.compile(optimizer=keras.optimizers.Adam(lr=0.001),\n             loss='categorical_crossentropy',\n             metrics=['accuracy'])\n\nhistory = model.fit(train_x,\n                   train_y,\n                   batch_size=64,\n                   epochs=20)\n\n<\/code><\/pre>\n\n<p>result is:\nEpoch 1\/20<\/p>\n\n<p>500\/500 [==============================] - 64s 128ms\/step - loss: 8.2847 - acc: 0.4860<\/p>\n\n<p>Epoch 2\/20<\/p>\n\n<p>500\/500 [==============================] - 60s 120ms\/step - loss: 8.1880 - acc: 0.4920<\/p>\n\n<p>Epoch 3\/20<\/p>\n\n<p>128\/500 [======>.......................] - ETA: 45s - loss: 8.4368 - acc: 0.4766<\/p>\n\n<p>.. it never gonna better than this loss and acc...<\/p>\n","answers":[]}
+{"QuestionId":58129484,"AnswerCount":0,"Tags":"<python><opencv><machine-learning><keras><deep-learning>","CreationDate":"2019-09-27T07:06:16.893","AcceptedAnswerId":null,"OwnerUserId":11499971.0,"Title":"Generate a unique hash value of unique cropped person face image using OpenCV","Body":"<p>I want to generate a unique hash value for a person's cropped face and store it in DB. So next time the person appears the hashing algorithm will generate an exact hash value which was stored in the database.<\/p>\n\n<p>I am using OpenCV and Python and Keras.  <\/p>\n\n<p>SO can someone help me on this?   <\/p>\n","answers":[]}
+{"QuestionId":58129748,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-27T07:24:12.353","AcceptedAnswerId":58130383.0,"OwnerUserId":12039381.0,"Title":"(keras) input a tensor and a index, get the tensor i want","Body":"<p>My Input is like a (3,3,2) array and a (3,3) array:<\/p>\n\n<pre><code>img = np.array([[[1,1],[2,2],[3,3]],\n                [[4,4],[5,5],[6,6]],\n                [[7,7],[8,8],[9,9]]])\n\nidx = np.array([[1,0,0],\n                [0,0,1],\n                [1,1,0]])\n<\/code><\/pre>\n\n<p>My ideal output should be:<\/p>\n\n<pre><code>[[1 1]\n [6 6]\n [7 7]\n [8 8]]\n<\/code><\/pre>\n\n<p>I want to do this by a custom layer:<\/p>\n\n<ol>\n<li>make a layer:<\/li>\n<\/ol>\n\n<pre><code>def extract_layer(data, idx):\n\n    idx = tf.where(idx)\n    data = tf.gather_nd(data,idx)\n    data = tf.reshape(data,[-1,2])\n\n    return data\n<\/code><\/pre>\n\n<ol start=\"2\">\n<li>make into model:<\/li>\n<\/ol>\n\n<pre><code>input_data = kl.Input(shape=(3,3,2))\ninput_idxs = kl.Input(shape=(3,3))\nextraction = kl.Lambda(lambda x:extract_layer(*x),name='extraction')([input_data,input_idxs])\n<\/code><\/pre>\n\n<p>I can build the model , and i can see the keras summary of the model,\nthe output is <\/p>\n\n<pre><code>model = Model(inputs=([input_data,input_idxs]), outputs=extraction)\nmodel.summary()\n\n...\ninput_1 (InputLayer)            (None, 3, 3, 2) \ninput_2 (InputLayer)            (None, 3, 3) \nextraction (Lambda)             (None, 2)\nTotal params: 0\n...\n<\/code><\/pre>\n\n<p>but when i start to predict like :<\/p>\n\n<pre><code>'i have already made the two inputs into (1,3,3,2) and (1,3,3) shape'\nresult = model.predict(x=([img,idx]))\n<\/code><\/pre>\n\n<p>it gets error:<\/p>\n\n<pre><code>'ValueError: could not broadcast input array from shape (4,2) into shape (1,2)'\n<\/code><\/pre>\n\n<p>i think the tensor of shape(4,2) is the value i want\nbut i don't know why keras broadcast it to (1,2)<\/p>\n\n<p>is there anyone who can help me ??<\/p>\n\n<p>thanks very much !<\/p>\n","answers":[{"AnswerId":"58130383","CreationDate":"2019-09-27T08:07:34.353","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>In your <code>extract_layer()<\/code> function, the <code>data<\/code> is two dims tensor. But <code>model.predict<\/code> is supposed to return results with an extra batch dim. Just expand the dim when return <code>data<\/code> in <code>extract_layer()<\/code> can fix this error.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def extract_layer(data, idx):\n\n    idx = tf.where(idx)\n    data = tf.gather_nd(data,idx)\n    data = tf.reshape(data,[-1,2])\n\n    return tf.expand_dims(data, axis=0)\n<\/code><\/pre>\n\n<p><strong>Note<\/strong>: Since the results returned by <code>tf.gather_nd<\/code> might have different length, I think the batch size would only be 1. Please correct me if I'm wrong. <\/p>\n"}]}
+{"QuestionId":58130047,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2019-09-27T07:44:39.147","AcceptedAnswerId":null,"OwnerUserId":9714034.0,"Title":"How to save a model in Tensorflow?","Body":"<p>Hi I am using pertained weights for an image classification task. I am using Tensorflow version 1.14.0, I am using mobilenetv1_050_224 for this task from the following <a href=\"https:\/\/tfhub.dev\/google\/imagenet\/mobilenet_v1_050_224\/classification\/3\" rel=\"nofollow noreferrer\">source<\/a>. <\/p>\n\n<pre><code>IMAGE_SHAPE = (400, 400)\nn_classes = 10\nclassifier_url = 'https:\/\/tfhub.dev\/google\/imagenet\/mobilenet_v1_050_224\/classification\/3'\nbase_model = hub.Module(classifier_url, tags=['train'])\nbase_model.trainable = False\nclassifier = tf.keras.Sequential([\n    hub.KerasLayer(base_model, input_shape=IMAGE_SHAPE+(3,)),\n    keras.layers.Dense(n_classes, activation='softmax')\n])\n#print (base_model.summary())\nprint (classifier.summary())\n<\/code><\/pre>\n\n<p>I train this model and I am able to get a good train\/validation accuracy on my dataset using transfer learning. The following is the code for learning part. <\/p>\n\n<pre><code>train_datagen = keras.preprocessing.image.ImageDataGenerator(\n                rescale=1.\/255)\n\nvalidation_datagen = keras.preprocessing.image.ImageDataGenerator(rescale=1.\/255)\n\ntrain_generator = train_datagen.flow(\n    x = train_dataset,\n    y = train_labels,\n    batch_size=batch_size,\n    seed=1)\n\nvalidation_generator = validation_datagen.flow(\n    x = validation_dataset, # Source directory for the validation images\n    y = valid_labels,\n    batch_size=batch_size,\n    seed=1)\n\nclassifier.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.01, beta_1=0.9, beta_2=0.999),\n              loss='categorical_crossentropy',\n              metrics=['accuracy'])\nepochs = 2\nsteps_per_epoch = train_generator.n \/\/ batch_size\nvalidation_steps = validation_generator.n \/\/ batch_size\n\nmodel = classifier.fit_generator(train_generator,\n                                 steps_per_epoch = steps_per_epoch,\n                                 epochs=epochs,\n                                 workers=4,\n                                 validation_data=validation_generator,\n                                 validation_steps=validation_steps)\n<\/code><\/pre>\n\n<p>However when I try to save the model: <\/p>\n\n<pre><code>export_path = '\/tmp\/simple_keras_model.h5'\nclassifier.save(export_path, save_format='h5')\n<\/code><\/pre>\n\n<p>I'm getting the following error:<\/p>\n\n<blockquote>\n  <p>NotImplementedError: Can only generate a valid config for\n  <code>hub.KerasLayer(handle, ...)<\/code>that uses a string <code>handle<\/code>.<\/p>\n  \n  <p>Got <code>type(handle)<\/code>: <\/p>\n<\/blockquote>\n\n<p>I am stuck on it and unable to get my head around it. Any leads in this regard would be helpful. Thank you. <\/p>\n","answers":[{"AnswerId":"58388496","CreationDate":"2019-10-15T06:31:49.103","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>As it is mentioned in the <a href=\"https:\/\/github.com\/tensorflow\/hub\/issues\/343\" rel=\"nofollow noreferrer\">link<\/a>, only TF2-compatible Hub modules work with hub.Keras.<\/p>\n\n<p>But can use SavedModel to save it. Please refer to this <a href=\"https:\/\/github.com\/tensorflow\/docs\/blob\/master\/site\/en\/guide\/saved_model.ipynb\" rel=\"nofollow noreferrer\">link<\/a> on how to use SavedModel API <\/p>\n"},{"AnswerId":"58130467","CreationDate":"2019-09-27T08:13:08.297","ParentId":null,"OwnerUserId":"6013016","Title":null,"Body":"<p>How about this?<\/p>\n\n<pre><code># Export the model to a SavedModel\nkeras.experimental.export_saved_model(classifier, '\/tmp\/simple_keras_model.h5')\n\n# Recreate the exact same model\nnew_model = keras.experimental.load_from_saved_model('\/tmp\/simple_keras_model.h5')\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/beta\/guide\/keras\/saving_and_serializing#export_to_savedmodel\" rel=\"nofollow noreferrer\">Source<\/a><\/p>\n"},{"AnswerId":"58130152","CreationDate":"2019-09-27T07:51:50.630","ParentId":null,"OwnerUserId":"8366805","Title":null,"Body":"<p>You can save the <code>classifier<\/code> by<\/p>\n\n<pre><code># save the underlying  tensorflow graph\nmodel_file = classifier.to_json()\nwith open(\"model.json\", \"w\") as source:\n    source.write(model_file)\n# save model parameter \nclassifier.save_weights(\"model_weights.h5\")\n<\/code><\/pre>\n\n<p>Then the saved model can be loaded by <\/p>\n\n<pre><code>from keras.models import model_from_json\n\nwith open(\"model.json\", \"r\") as f:\n    classifier = model_from_json(f.read())\nclassifier.load_weights(\"model_weights.h5\")\n<\/code><\/pre>\n"}]}
+{"QuestionId":58130203,"AnswerCount":2,"Tags":"<tensorflow><artificial-intelligence><transfer-learning><google-coral>","CreationDate":"2019-09-27T07:55:44.327","AcceptedAnswerId":null,"OwnerUserId":515958.0,"Title":"Distributed retraining (TF & Google Coral)","Body":"<p>Assuming I have tens of Google Coral devices doing object detection (using same trained model), every once in awhile we will retraining a device for new object (transfer learning), lets say this device is Coral1, now I wonder how would I transfer learning, what Coral1 learned to all devices (without the need to retrain those devices)?<\/p>\n\n<p>for sure devices can be Google Coral or any other device<\/p>\n","answers":[{"AnswerId":"58388442","CreationDate":"2019-10-15T06:27:48.197","ParentId":null,"OwnerUserId":"12192957","Title":null,"Body":"<p>Since what we are assuming is that, all the devices will inference on the same model at start, <\/p>\n\n<p>whenever there is a new learning done by any device, the updated trained model should be pushed to all other devices, helping them to start recognizing the new objects. <\/p>\n\n<p>Every device does not need to train individually, if we can maintain synchronization among the devices.<\/p>\n\n<p>For further information, go through this <a href=\"https:\/\/coral.withgoogle.com\/docs\/edgetpu\/retrain-detection\/\" rel=\"nofollow noreferrer\">link<\/a><\/p>\n"},{"AnswerId":"58385281","CreationDate":"2019-10-14T23:32:59.390","ParentId":null,"OwnerUserId":"6262757","Title":null,"Body":"<p>You shouldn't be \"retraining a device\", but rather retraining a model. Take a look at this guide on how to re-train a model:\n<a href=\"https:\/\/coral.withgoogle.com\/docs\/edgetpu\/retrain-detection\/\" rel=\"nofollow noreferrer\">https:\/\/coral.withgoogle.com\/docs\/edgetpu\/retrain-detection\/<\/a><\/p>\n\n<p>Once you've finished retraining a model, you can scp it into the other board and reload.<\/p>\n"}]}
+{"QuestionId":58130237,"AnswerCount":1,"Tags":"<pytorch>","CreationDate":"2019-09-27T07:58:01.883","AcceptedAnswerId":58130331.0,"OwnerUserId":8410477.0,"Title":"RuntimeError: Given groups=1, weight of size 16 1 5 5, expected input[100, 3, 256, 256] to have 1 channels, but got 3 channels instead","Body":"<p>I try to run the following programe for images classification problem in Pytorch:<\/p>\n\n<pre><code>import torch\nimport torch.nn as nn\nimport torchvision\nimport torchvision.transforms as transforms\nimport torch.utils.data as data\n\n# Device configuration\ndevice = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')\n\n# Hyper parameters\nnum_epochs = 5\nnum_classes = 10\nbatch_size = 100\nlearning_rate = 0.001\n\nTRAIN_DATA_PATH = \"train\/\"\nTEST_DATA_PATH = \"test\/\"\nTRANSFORM_IMG = transforms.Compose([\n    transforms.Resize(256),\n    transforms.CenterCrop(256),\n    transforms.ToTensor(),\n    transforms.Normalize(mean=[0.485, 0.456, 0.406],\n                         std=[0.229, 0.224, 0.225] )\n    ])\n\ntrain_dataset = torchvision.datasets.ImageFolder(root=TRAIN_DATA_PATH, transform=TRANSFORM_IMG)\ntrain_loader = data.DataLoader(train_dataset, batch_size=batch_size, shuffle=True,  num_workers=4)\ntest_dataset = torchvision.datasets.ImageFolder(root=TEST_DATA_PATH, transform=TRANSFORM_IMG)\ntest_loader  = data.DataLoader(test_dataset, batch_size=batch_size, shuffle=True, num_workers=4)\n\n# Convolutional neural network (two convolutional layers)\nclass ConvNet(nn.Module):\n    def __init__(self, num_classes=10):\n        super(ConvNet, self).__init__()\n        self.layer1 = nn.Sequential(\n            nn.Conv2d(1, 16, kernel_size=5, stride=1, padding=2),\n            nn.BatchNorm2d(16),\n            nn.ReLU(),\n            nn.MaxPool2d(kernel_size=2, stride=2))\n        self.layer2 = nn.Sequential(\n            nn.Conv2d(16, 32, kernel_size=5, stride=1, padding=2),\n            nn.BatchNorm2d(32),\n            nn.ReLU(),\n            nn.MaxPool2d(kernel_size=2, stride=2))\n        self.fc = nn.Linear(7 * 7 * 32, num_classes)\n\n    def forward(self, x):\n        out = self.layer1(x)\n        out = self.layer2(out)\n        out = out.reshape(out.size(0), -1)\n        out = self.fc(out)\n        return out\n\n\nmodel = ConvNet(num_classes).to(device)\n\n# Loss and optimizer\ncriterion = nn.CrossEntropyLoss()\noptimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)\n\n# Train the model\ntotal_step = len(train_loader)\nfor epoch in range(num_epochs):\n    for i, (images, labels) in enumerate(train_loader):\n        images = images.to(device)\n        labels = labels.to(device)\n\n        # Forward pass\n        outputs = model(images)\n        loss = criterion(outputs, labels)\n\n        # Backward and optimize\n        optimizer.zero_grad()\n        loss.backward()\n        optimizer.step()\n\n        if (i + 1) % 100 == 0:\n            print('Epoch [{}\/{}], Step [{}\/{}], Loss: {:.4f}'\n                  .format(epoch + 1, num_epochs, i + 1, total_step, loss.item()))\n\n# Test the model\nmodel.eval()  # eval mode (batchnorm uses moving mean\/variance instead of mini-batch mean\/variance)\nwith torch.no_grad():\n    correct = 0\n    total = 0\n    for images, labels in test_loader:\n        images = images.to(device)\n        labels = labels.to(device)\n        outputs = model(images)\n        _, predicted = torch.max(outputs.data, 1)\n        total += labels.size(0)\n        correct += (predicted == labels).sum().item()\n\n    print('Test Accuracy of the model on the 10000 test images: {} %'.format(100 * correct \/ total))\n\n# Save the model checkpoint\ntorch.save(model.state_dict(), 'model\/model.ckpt')\n<\/code><\/pre>\n\n<p>But I get a <code>RuntimeError<\/code>:<\/p>\n\n<pre><code>Traceback (most recent call last):\n\nRuntimeError: Given groups=1, weight of size 16 1 5 5, expected input[100, 3, 256, 256] to have 1 channels, but got 3 channels instead\n<\/code><\/pre>\n\n<p>Someone could help to fix the bug? Thanks a lot.<\/p>\n\n<p><strong>Reference related:<\/strong><\/p>\n\n<p><a href=\"https:\/\/discuss.pytorch.org\/t\/given-groups-1-weight-16-1-5-5-so-expected-input-100-3-64-64-to-have-1-channels-but-got-3-channels-instead\/28831\/17\" rel=\"nofollow noreferrer\">https:\/\/discuss.pytorch.org\/t\/given-groups-1-weight-16-1-5-5-so-expected-input-100-3-64-64-to-have-1-channels-but-got-3-channels-instead\/28831\/17<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/53416833\/runtimeerror-given-groups-1-weight-of-size-64-3-7-7-expected-input3-1\">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<\/a><\/p>\n","answers":[{"AnswerId":"58130331","CreationDate":"2019-09-27T08:03:47.387","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Your input layer <code>self.layer1<\/code> starts with a 2d convolution <code>nn.Conv2d(1, 16, kernel_size=5, stride=1, padding=2)<\/code>. This conv layer expects an input with two spatial dimensions and one channel, and outputs a tesnor with the same spatial dimensions and 16 channels.<br>\nHowever, your input has three channels and not one (RGB image instead of gray level image).<\/p>\n\n<p>Make sure your net and data are in synch.<\/p>\n"}]}
+{"QuestionId":58130241,"AnswerCount":0,"Tags":"<echo><pytorch><data-fitting>","CreationDate":"2019-09-27T07:58:18.160","AcceptedAnswerId":null,"OwnerUserId":11270627.0,"Title":"training fit in echostatenet","Body":"<p>I am using an Echo-state-net code implemented on PyTorch, and even though the prediction is accurate enough, the training fit looks shifted, can someone help to understand why this can happen? The shift function is deactivated in the code.\n<a href=\"https:\/\/i.stack.imgur.com\/gkQI3.png\" rel=\"nofollow noreferrer\">Example image<\/a><\/p>\n","answers":[]}
+{"QuestionId":58130277,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><lstm>","CreationDate":"2019-09-27T08:00:34.633","AcceptedAnswerId":null,"OwnerUserId":7850499.0,"Title":"Implement a customized LSTM in Tensorflow","Body":"<p>I'm trying to implement LSTM with layer normalization and masking in Tensorflow. Here's my code, roughly following <a href=\"https:\/\/github.com\/openai\/baselines\/blob\/8c2aea2addc9f3ba36d4a0c937e6a2d09830afc7\/baselines\/a2c\/utils.py#L81\" rel=\"nofollow noreferrer\">OpenAI's implementation<\/a> with some modifications.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def lstm_norm(xs, units, masks):\n    \"\"\"lstm with masks and layer normalization\n\n    Arguments:\n        xs      3d Tensor   --  input data of shape [n_batch, n_seq, dim]\n        units   int         --  size of hidden\/cell state\n        masks   2d Tensor   --  masks, must match the first 2 dimensions of xs\n\n    Returns:\n        ys      3d Tensor   -- output date of shape [n_batch, n_seq, units]\n        initial_state, final_state\n    \"\"\"\n    def ln(x, gamma, beta, eps=1e-5, axes=[1]):\n        mean, var = tf.nn.moments(x, axes=axes, keep_dims=True)\n        x = (x - mean) \/ tf.sqrt(var + eps)\n        x = gamma * x + beta\n        return x\n\n    n_batch, n_steps, x_dim = xs.shape.as_list()\n\n    xw_shape = [x_dim, 4*units]\n    xb_shape = [4*units]\n    hw_shape = [units, 4*units]\n    hb_shape = [4*units]\n\n    with tf.variable_scope('lstm_norm'):\n        x_w = tf.get_variable('x_w', shape=xw_shape)\n        x_g = tf.get_variable('x_g', [4*units], initializer=tf.constant_initializer(1.0))\n        x_b = tf.get_variable('x_b', shape=xb_shape, \n                              initializer=tf.zeros_initializer())\n\n        h_w = tf.get_variable('h_w', shape=hw_shape)\n        h_g = tf.get_variable('h_g', [4*units], initializer=tf.constant_initializer(1.0))\n        h_b = tf.get_variable('h_b', shape=hb_shape, \n                              initializer=tf.zeros_initializer())\n\n        b = tf.get_variable('b', [4*units], initializer=tf.zeros_initializer())\n\n        c_g = tf.get_variable('c_g', [units], initializer=tf.constant_initializer(1.0))\n        c_b = tf.get_variable('c_b', [units], initializer=tf.zeros_initializer())\n\n        initial_state = tf.zeros([n_batch, 2*units], name='initial_state')\n        h, c = tf.split(value=initial_state, num_or_size_splits=2, axis=1)\n\n        # unstack xs and masks so that we can iterate through next\n        xs = tf.unstack(xs, n_steps, axis=1)\n        if masks is not None:\n            masks = tf.unstack(masks, n_steps, axis=1)\n\n        ys = []\n        for x, m in zip(xs, masks):\n            m = m[..., None]\n            c *= m\n            h *= m\n            # implementation of lstm with layer norm starts here\n            z = ln(tf.matmul(x, x_w), x_g, x_b) + ln(tf.matmul(h, h_w), h_g, h_b)\n            f, i, o, u = tf.split(value=z, num_or_size_splits=4, axis=1)\n            f = tf.nn.sigmoid(f)\n            i = tf.nn.sigmoid(i)\n            o = tf.nn.sigmoid(o)\n            u = tf.tanh(u)\n            c = f * c + i * u\n            h = o * tf.tanh(ln(c, c_g, c_b))\n            ys.append(h)\n\n        final_state = (h, c)\n        ys = tf.stack(ys, 1)\n\n    return ys, (initial_state, final_state)\n\ntf.reset_default_graph()\nxs = tf.placeholder(tf.float32, [64, 1000, 256])\nmasks = tf.placeholder(tf.float32, [64, 1000])\nstart = time.time()\n# takes 36s to construct the graph\nys, _ =  lstm_norm(xs, 126, masks)\nprint(time.time() - start)\n<\/code><\/pre>\n\n<p>This code takes significantly longer time to construct the graph than the official Tensorflow implementation in the following(36s vs 180ms on my laptop with no CUDA GPU). <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># official implementation of LSTM, taking 180ms to construct the graph\ntf.reset_default_graph()\nxs = tf.placeholder(tf.float32, [64, 1000, 256])\nlstm_keras = tf.keras.layers.LSTM(126, return_sequences=True, return_state=True)\nys_keras, _, _ = lstm_keras(xs)\n<\/code><\/pre>\n\n<p>I also test their run time with the following code<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># use my implementation of LSTM, taking 24s\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    start = time.time()\n    y = sess.run(ys, feed_dict={xs: np.random.rand(64, 1000, 256), masks: np.random.randint(2, size=(64, 1000))})\n    print(time.time() - start)\n    print(y.shape)\n<\/code><\/pre>\n\n<pre class=\"lang-py prettyprint-override\"><code># use the official implementation, using 0.8s\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    start = time.time()\n    y = sess.run(ys_keras, feed_dict={xs: np.random.rand(64, 1000, 256)})\n    print(time.time() - start)\n    print(y.shape)\n<\/code><\/pre>\n\n<p>Furthermore, my implementation has a significant drawback that it should know the length of sequence before hand. This limits its use in cases where we may pass sequences of different length --- it may be bypassed by reusing variables, which, however, will complicate the code. Is there any other way to construct an LSTM that meet my needs? More specifically, I want to implement a LSTM with the following properties<\/p>\n\n<ol>\n<li>Layer normalization can be applied<\/li>\n<li>Masking can be applied<\/li>\n<li>It can be used with varied length of sequence<\/li>\n<\/ol>\n\n<p>Any advice will be appreciated. Thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":58130281,"AnswerCount":1,"Tags":"<python><tensorflow><horovod>","CreationDate":"2019-09-27T08:00:56.223","AcceptedAnswerId":null,"OwnerUserId":3182473.0,"Title":"tensorflow: tf.set_random_seed() same code, but got different results","Body":"<p><strong>In short, in tensorflow, except for tf.set_random_seed(), is there any other config I should set to reproduce the same result? no numpy operation in my code.<\/strong><\/p>\n\n<p>Long version:\nI am training a model with horovod, and I want to reproduce the results for debug. I set tf.set_random_seed(1) before build the graph. All operations are in tensorflow, no numpy.<\/p>\n\n<p>And I just rerun the training script for several times, code has not been modified.<\/p>\n\n<p>As I understand, they should produce the same results, since the graph is build by the same order and the random_seed is set the same. The session.run order will not cause the different, not to mention I didn't change the code.<\/p>\n\n<p>However, I still got different results for the same code on multiple runs. <\/p>\n\n<p>Specifically, I can see the input image is read by the same order, and the step 'tf.image.random_flip_left_right' produce random flip with the same seed.<\/p>\n\n<p>But I test the tf.image.random_flip_left_right by the following code, which gives the same results. So there is no bug in this function. It did accept the global tf random_seed and produce same results.<\/p>\n\n<pre><code>import tensorflow as tf                                                                                                              \ntf.set_random_seed(0)\nimg_raw =tf.read_file('test.jpg')\nimage = tf.image.decode_jpeg(img_raw,channels=3,dct_method='INTEGER_ACCURATE')\nimage = tf.image.random_flip_left_right(image)\n\nsess=tf.Session()\nprint(sess.run(image))\n<\/code><\/pre>\n\n<p>How can it be random flip with the same seed and graph? Any suggestion would be appreciated.<\/p>\n","answers":[{"AnswerId":"58130407","CreationDate":"2019-09-27T08:09:27.783","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p>Can you set the seed in <code>tf.image.random_flip_left_right(image, seed = 0)<\/code> and test?<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>okay, you said you are not using numpy in the code but TF uses it for some internal tasks. So, can you again try and this time fixing the numpy seed too.<\/p>\n\n<pre><code>import numpy as np\nfrom numpy.random import seed\nseed(0)\n<\/code><\/pre>\n\n<p>If this doesn't also work then maybe the problem is with <code>Horovod<\/code> as talked about it <a href=\"https:\/\/github.com\/horovod\/horovod\/issues\/173#issuecomment-369247966\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":58130484,"AnswerCount":1,"Tags":"<image><tensorflow><nlp>","CreationDate":"2019-09-27T08:14:20.417","AcceptedAnswerId":null,"OwnerUserId":10833336.0,"Title":"I want to build a code to recognize some information from a form picture with tensorflow","Body":"<p>Can someone introduce some guide about this question? I have looking tutorial from official site of tensorflow. <\/p>\n","answers":[{"AnswerId":"58130712","CreationDate":"2019-09-27T08:30:27.280","ParentId":null,"OwnerUserId":"8842694","Title":null,"Body":"<p><a href=\"https:\/\/towardsdatascience.com\/image-classification-in-10-minutes-with-mnist-dataset-54c35b77a38d\" rel=\"nofollow noreferrer\">Here<\/a> is a good example for recognising digits from a picture. It is in keras which is also supported by TensorFlow.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/keras\/basic_classification\" rel=\"nofollow noreferrer\">Here<\/a> is another tutorial for recognizing clothes from pictures.<\/p>\n"}]}
+{"QuestionId":58130966,"AnswerCount":0,"Tags":"<python-3.x><keras><model><jupyter-notebook>","CreationDate":"2019-09-27T08:46:15.470","AcceptedAnswerId":null,"OwnerUserId":8544660.0,"Title":"model = keras.models.load_model(model_file) kills kernel in jupyter notebook","Body":"<p>Every time I try to load my model file the kernel dies and has to restart. Absolutely no idea what to do here.<\/p>\n\n<pre><code>model_file = cur_dir+model_names[i]\n\nprint(model_file)\n<\/code><\/pre>\n\n<p>Output: \/home\/jack\/Desktop\/output\/models\/2019-09-27 00-05-20\/02-val_acc-0.69-val_loss-0.66.hdf5<\/p>\n\n<pre><code>model = keras.models.load_model(model_file)\n<\/code><\/pre>\n\n<p>Then the kernel dies.\nPlease help.<\/p>\n","answers":[]}
+{"QuestionId":58131278,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><gradient>","CreationDate":"2019-09-27T09:04:30.970","AcceptedAnswerId":null,"OwnerUserId":6384539.0,"Title":"Why tensorflow does not use numerical differentiation?","Body":"<p>I know TensorFlow uses automatic differentiation to calculate gradients using the computational graph. Is there any specific reason for not using the numerical differential method. Won't it save all the trouble of applying chain rules for computing gradients in the graph?<\/p>\n","answers":[]}
+{"QuestionId":58131442,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-27T09:14:01.940","AcceptedAnswerId":58131719.0,"OwnerUserId":7296485.0,"Title":"how can i sum first n items in tensorflow?","Body":"<p>example<\/p>\n\n<pre><code>a = [[1, 0, 1, 0], [0, 1, 1 ,0]]\nresult = [[a[0,0], a[0,0]+a[0,1], a[0,0]+a[0,1]+a[0,2], a[0,0]+a[0,1]+a[0,2]+a[0,3]], [a[1,0], a[1,0]+a[1,1], a[1,0]+a[1,1]+a[1,2], a[1,0]+a[1,1]+a[1,2]+a[1,3]]\n       = [[1, 1, 2, 2], [0, 1, 2 ,2]]\n<\/code><\/pre>\n\n<p>I don't known how to do this, please help me. thx<\/p>\n","answers":[{"AnswerId":"58131719","CreationDate":"2019-09-27T09:30:29.603","ParentId":null,"OwnerUserId":"9698684","Title":null,"Body":"<p>You're lookink for the cumulative sum, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/math\/cumsum\" rel=\"nofollow noreferrer\"><code>tf.cumsum<\/code><\/a>, along the second axis:<\/p>\n\n<pre><code>a = tf.Variable([[1, 0, 1, 0], [0, 1, 1 ,0]])\ninit_op = tf.initialize_all_variables()\n\n# run the graph and calculate the cumsum\nwith tf.Session() as sess:\n    sess.run(init_op) #execute init_op\n    print (sess.run(tf.cumsum(a, axis=1)))):\n\n[[1 1 2 2]\n [0 1 2 2]]\n<\/code><\/pre>\n"}]}
+{"QuestionId":58131581,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning><seq2seq>","CreationDate":"2019-09-27T09:21:43.713","AcceptedAnswerId":null,"OwnerUserId":11368194.0,"Title":"What to do when GPU memory is fully used","Body":"<p>Recently, I have started working with <code>1 Tesla T4 GPU with 12 vCpu and 60 GB RAM<\/code>. I am training an <code>Seq2Seq bidirectional LSTM with attention layer<\/code> and have <code>38,863,916 training parameters<\/code>. While training my Seq2Seq model I am getting following errors `GPU sync failed. I searched for the error and got to know it means my gpu memory is full. Following is my code<\/p>\n\n<pre><code>encoder_inputs = Input(shape=(max_x_len,))\n\nemb1 = Embedding(len(x_voc), 100, weights=[x_voc], trainable = False)(encoder_inputs)\n\nencoder = Bidirectional(LSTM(latent_dim, return_state=True, return_sequences =True))\nencoder_outputs0, _, _, _, _ = encoder(emb1)\n\nencoder = Bidirectional(LSTM(latent_dim, return_state=True, return_sequences =True))\nencoder_outputs2, forward_h, forward_c, backward_h, backward_c = encoder(encoder_outputs0)\n\nencoder_states = [forward_h, forward_c, backward_h, backward_c]\n\n\n### Decoder\ndecoder_inputs = Input(shape=(None,))\n\nemb2 = Embedding(len(y_voc), 100, weights=[y_voc], trainable = False)(decoder_inputs)\n\ndecoder_lstm = Bidirectional(LSTM(latent_dim, return_sequences=True, return_state=True))\ndecoder_outputs2, _, _, _, _ = decoder_lstm(emb2, initial_state=encoder_states)\n\n\nattn_layer = AttentionLayer(name='attention_layer')\nattn_out, attn_states = attn_layer([encoder_outputs2, decoder_outputs2]) \ndecoder_concat_input = Concatenate(axis=-1, name='concat_layer')([decoder_outputs2, attn_out])\n\ndecoder_dense = Dense(len(y_voc), activation='softmax')\ndecoder_outputs = decoder_dense(decoder_concat_input)\n\n\nmodel = Model([encoder_inputs, decoder_inputs], decoder_outputs)\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\nmodel.summary()\n\nmodel.fit([x_inc,x_dec],y_dec, batch_size = 32, epochs=500)\n<\/code><\/pre>\n\n<pre><code>x_inc.shape =&gt; (1356, 433)\nx_dec.shape =&gt;(1356, 131)\ny_dec.shape =&gt;(1356, 131, 10633)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58131591,"AnswerCount":0,"Tags":"<python><tensorflow><object-detection><object-detection-api>","CreationDate":"2019-09-27T09:22:08.683","AcceptedAnswerId":null,"OwnerUserId":2277994.0,"Title":"Build custom function to feed training data to Object Detection API","Body":"<p>I am using the Object Detection API with Faster RCNN with a COCO-style dataset. I run the training from the model_main.py file.<\/p>\n\n<p>However, I do not want to use the random training provided, but I want to customise it.<\/p>\n\n<p>Basically, I have a file, scores.txt, which contains the name of the image file with a score. I need to sort the files based on that score. Then, for the first 30k iterations I need to train only on the first half of the images, then on all images.<\/p>\n\n<p>From what I understand, currently, the training data is fed like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>train_and_eval_dict = model_lib.create_estimator_and_inputs(\n      run_config=config,\n      hparams=model_hparams.create_hparams(FLAGS.hparams_overrides),\n      pipeline_config_path=FLAGS.pipeline_config_path,\n      train_steps=FLAGS.num_train_steps,\n      sample_1_of_n_eval_examples=FLAGS.sample_1_of_n_eval_examples,\n      sample_1_of_n_eval_on_train_examples=(\n          FLAGS.sample_1_of_n_eval_on_train_examples))\n...\ntrain_input_fn = train_and_eval_dict['train_input_fn']\n...\ntrain_spec, eval_specs = model_lib.create_train_and_eval_specs(\n        train_input_fn,\n        eval_input_fns,\n        eval_on_train_input_fn,\n        predict_input_fn,\n        train_steps,\n        eval_on_train_data=False)\n\ntf.estimator.train_and_evaluate(estimator, train_spec, eval_specs[0])\n<\/code><\/pre>\n\n<p>where train_input_fn actually call the train_input function defined as<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def train_input(train_config, train_input_config,\n                model_config, model=None, params=None):\n  \"\"\"Returns `features` and `labels` tensor dictionaries for training.\n\n  Args:\n    train_config: A train_pb2.TrainConfig.\n    train_input_config: An input_reader_pb2.InputReader.\n    model_config: A model_pb2.DetectionModel.\n    model: A pre-constructed Detection Model.\n      If None, one will be created from the config.\n    params: Parameter dictionary passed from the estimator.\n\n  Returns:\n    A tf.data.Dataset that holds (features, labels) tuple.\n\n    features: Dictionary of feature tensors.\n      features[fields.InputDataFields.image] is a [batch_size, H, W, C]\n        float32 tensor with preprocessed images.\n      features[HASH_KEY] is a [batch_size] int32 tensor representing unique\n        identifiers for the images.\n      features[fields.InputDataFields.true_image_shape] is a [batch_size, 3]\n        int32 tensor representing the true image shapes, as preprocessed\n        images could be padded.\n      features[fields.InputDataFields.original_image] (optional) is a\n        [batch_size, H, W, C] float32 tensor with original images.\n    labels: Dictionary of groundtruth tensors.\n      labels[fields.InputDataFields.num_groundtruth_boxes] is a [batch_size]\n        int32 tensor indicating the number of groundtruth boxes.\n      labels[fields.InputDataFields.groundtruth_boxes] is a\n        [batch_size, num_boxes, 4] float32 tensor containing the corners of\n        the groundtruth boxes.\n      labels[fields.InputDataFields.groundtruth_classes] is a\n        [batch_size, num_boxes, num_classes] float32 one-hot tensor of\n        classes.\n      labels[fields.InputDataFields.groundtruth_weights] is a\n        [batch_size, num_boxes] float32 tensor containing groundtruth weights\n        for the boxes.\n      -- Optional --\n      labels[fields.InputDataFields.groundtruth_instance_masks] is a\n        [batch_size, num_boxes, H, W] float32 tensor containing only binary\n        values, which represent instance masks for objects.\n      labels[fields.InputDataFields.groundtruth_keypoints] is a\n        [batch_size, num_boxes, num_keypoints, 2] float32 tensor containing\n        keypoints for each box.\n\n  Raises:\n    TypeError: if the `train_config`, `train_input_config` or `model_config`\n      are not of the correct type.\n  \"\"\"\n  if not isinstance(train_config, train_pb2.TrainConfig):\n    raise TypeError('For training mode, the `train_config` must be a '\n                    'train_pb2.TrainConfig.')\n  if not isinstance(train_input_config, input_reader_pb2.InputReader):\n    raise TypeError('The `train_input_config` must be a '\n                    'input_reader_pb2.InputReader.')\n  if not isinstance(model_config, model_pb2.DetectionModel):\n    raise TypeError('The `model_config` must be a '\n                    'model_pb2.DetectionModel.')\n\n  if model is None:\n    model_preprocess_fn = INPUT_BUILDER_UTIL_MAP['model_build'](\n        model_config, is_training=True).preprocess\n  else:\n    model_preprocess_fn = model.preprocess\n\n  def transform_and_pad_input_data_fn(tensor_dict):\n    \"\"\"Combines transform and pad operation.\"\"\"\n    data_augmentation_options = [\n        preprocessor_builder.build(step)\n        for step in train_config.data_augmentation_options\n    ]\n    data_augmentation_fn = functools.partial(\n        augment_input_data,\n        data_augmentation_options=data_augmentation_options)\n\n    image_resizer_config = config_util.get_image_resizer_config(model_config)\n    image_resizer_fn = image_resizer_builder.build(image_resizer_config)\n    transform_data_fn = functools.partial(\n        transform_input_data, model_preprocess_fn=model_preprocess_fn,\n        image_resizer_fn=image_resizer_fn,\n        num_classes=config_util.get_number_of_classes(model_config),\n        data_augmentation_fn=data_augmentation_fn,\n        merge_multiple_boxes=train_config.merge_multiple_label_boxes,\n        retain_original_image=train_config.retain_original_images,\n        use_multiclass_scores=train_config.use_multiclass_scores,\n        use_bfloat16=train_config.use_bfloat16)\n\n    tensor_dict = pad_input_data_to_static_shapes(\n        tensor_dict=transform_data_fn(tensor_dict),\n        max_num_boxes=train_input_config.max_number_of_boxes,\n        num_classes=config_util.get_number_of_classes(model_config),\n        spatial_image_shape=config_util.get_spatial_image_size(\n            image_resizer_config))\n    return (_get_features_dict(tensor_dict), _get_labels_dict(tensor_dict))\n\n  dataset = INPUT_BUILDER_UTIL_MAP['dataset_build'](\n      train_input_config,\n      transform_input_data_fn=transform_and_pad_input_data_fn,\n      batch_size=params['batch_size'] if params else train_config.batch_size)\n  for label in dataset.take(10):\n    print(label)\n  import pdb\n  pdb.set_trace()\n  return dataset\n\n\ndef create_eval_input_fn(eval_config, eval_input_config, model_config):\n  \"\"\"Creates an eval `input` function for `Estimator`.\n\n  Args:\n    eval_config: An eval_pb2.EvalConfig.\n    eval_input_config: An input_reader_pb2.InputReader.\n    model_config: A model_pb2.DetectionModel.\n\n  Returns:\n    `input_fn` for `Estimator` in EVAL mode.\n  \"\"\"\n\n  def _eval_input_fn(params=None):\n    return eval_input(eval_config, eval_input_config, model_config,\n                      params=params)\n\n  return _eval_input_fn\n\n\ndef eval_input(eval_config, eval_input_config, model_config,\n               model=None, params=None):\n  \"\"\"Returns `features` and `labels` tensor dictionaries for evaluation.\n\n  Args:\n    eval_config: An eval_pb2.EvalConfig.\n    eval_input_config: An input_reader_pb2.InputReader.\n    model_config: A model_pb2.DetectionModel.\n    model: A pre-constructed Detection Model.\n      If None, one will be created from the config.\n    params: Parameter dictionary passed from the estimator.\n\n  Returns:\n    A tf.data.Dataset that holds (features, labels) tuple.\n\n    features: Dictionary of feature tensors.\n      features[fields.InputDataFields.image] is a [1, H, W, C] float32 tensor\n        with preprocessed images.\n      features[HASH_KEY] is a [1] int32 tensor representing unique\n        identifiers for the images.\n      features[fields.InputDataFields.true_image_shape] is a [1, 3]\n        int32 tensor representing the true image shapes, as preprocessed\n        images could be padded.\n      features[fields.InputDataFields.original_image] is a [1, H', W', C]\n        float32 tensor with the original image.\n    labels: Dictionary of groundtruth tensors.\n      labels[fields.InputDataFields.groundtruth_boxes] is a [1, num_boxes, 4]\n        float32 tensor containing the corners of the groundtruth boxes.\n      labels[fields.InputDataFields.groundtruth_classes] is a\n        [num_boxes, num_classes] float32 one-hot tensor of classes.\n      labels[fields.InputDataFields.groundtruth_area] is a [1, num_boxes]\n        float32 tensor containing object areas.\n      labels[fields.InputDataFields.groundtruth_is_crowd] is a [1, num_boxes]\n        bool tensor indicating if the boxes enclose a crowd.\n      labels[fields.InputDataFields.groundtruth_difficult] is a [1, num_boxes]\n        int32 tensor indicating if the boxes represent difficult instances.\n      -- Optional --\n      labels[fields.InputDataFields.groundtruth_instance_masks] is a\n        [1, num_boxes, H, W] float32 tensor containing only binary values,\n        which represent instance masks for objects.\n\n  Raises:\n    TypeError: if the `eval_config`, `eval_input_config` or `model_config`\n      are not of the correct type.\n  \"\"\"\n  params = params or {}\n  if not isinstance(eval_config, eval_pb2.EvalConfig):\n    raise TypeError('For eval mode, the `eval_config` must be a '\n                    'train_pb2.EvalConfig.')\n  if not isinstance(eval_input_config, input_reader_pb2.InputReader):\n    raise TypeError('The `eval_input_config` must be a '\n                    'input_reader_pb2.InputReader.')\n  if not isinstance(model_config, model_pb2.DetectionModel):\n    raise TypeError('The `model_config` must be a '\n                    'model_pb2.DetectionModel.')\n\n  if model is None:\n    model_preprocess_fn = INPUT_BUILDER_UTIL_MAP['model_build'](\n        model_config, is_training=False).preprocess\n  else:\n    model_preprocess_fn = model.preprocess\n\n  def transform_and_pad_input_data_fn(tensor_dict):\n    \"\"\"Combines transform and pad operation.\"\"\"\n    num_classes = config_util.get_number_of_classes(model_config)\n\n    image_resizer_config = config_util.get_image_resizer_config(model_config)\n    image_resizer_fn = image_resizer_builder.build(image_resizer_config)\n\n    transform_data_fn = functools.partial(\n        transform_input_data, model_preprocess_fn=model_preprocess_fn,\n        image_resizer_fn=image_resizer_fn,\n        num_classes=num_classes,\n        data_augmentation_fn=None,\n        retain_original_image=eval_config.retain_original_images)\n    tensor_dict = pad_input_data_to_static_shapes(\n        tensor_dict=transform_data_fn(tensor_dict),\n        max_num_boxes=eval_input_config.max_number_of_boxes,\n        num_classes=config_util.get_number_of_classes(model_config),\n        spatial_image_shape=config_util.get_spatial_image_size(\n            image_resizer_config))\n    return (_get_features_dict(tensor_dict), _get_labels_dict(tensor_dict))\n  dataset = INPUT_BUILDER_UTIL_MAP['dataset_build'](\n      eval_input_config,\n      batch_size=params['batch_size'] if params else eval_config.batch_size,\n      transform_input_data_fn=transform_and_pad_input_data_fn)\n  return dataset\n<\/code><\/pre>\n\n<p>I do not know if I can override the train_input_fn from model_main, how to create it to read everything needed from the config file, how to access the actual dataset to sort it as mentioned and how to pass the current iteration.<\/p>\n\n<p>I will be glad to answer to other questions.<\/p>\n\n<p>Thank you,\nPetru<\/p>\n","answers":[]}
+{"QuestionId":58131629,"AnswerCount":1,"Tags":"<python><machine-learning><keras>","CreationDate":"2019-09-27T09:24:50.400","AcceptedAnswerId":null,"OwnerUserId":11609346.0,"Title":"Custom RMSE not the same as taking the root of built-in Keras MSE for same prediction","Body":"<p>I have defined a custom RMSE function:<\/p>\n\n<pre><code>def rmse(y_pred, y_true):\n    return K.sqrt(K.mean(K.square(y_pred - y_true)))\n<\/code><\/pre>\n\n<p>I was evaluating it against the mean squared error provided by Keras:<\/p>\n\n<p><code>keras.losses.mean_squared_error(y_true, y_pred)<\/code><\/p>\n\n<p>The values I get for MSE and RMSE metrics respectively for some (the same) prediction are:<\/p>\n\n<p><code>mse: 115.7218 - rmse: 8.0966<\/code><\/p>\n\n<p>Now, when I take the root of the MSE, I get <code>10.7574<\/code>, which is obviously higher than the RMSE the custom RMSE function outputs. I haven't been able to figure out why this is so, nor have I found any related posts on this particular topic. Is there maybe a mistake in the RMSE function that I'm simply not seeing? Or is it somehow related to how Keras defines <code>axis=-1<\/code> in the MSE function (purpose of which I haven't fully understood yet)?<\/p>\n\n<p>Here is where I invoke the RMSE and MSE:<\/p>\n\n<p><code>model.compile(loss=\"mae\", optimizer=\"adam\", metrics=[\"mse\", rmse])<\/code><\/p>\n\n<p>So I would expect the root of MSE to be the same as the RMSE.<\/p>\n\n<p>I originally asked this question on Cross Validated but it was put on hold as off-topic.<\/p>\n","answers":[{"AnswerId":"58133546","CreationDate":"2019-09-27T11:21:36.977","ParentId":null,"OwnerUserId":"5597718","Title":null,"Body":"<blockquote>\n  <p>Is there maybe a mistake in the RMSE loss function that I'm simply not\n  seeing? Or is it somehow related to how Keras defines axis=-1 in the\n  MSE loss function (purpose of which I haven't fully understood yet)?<\/p>\n<\/blockquote>\n\n<p>When Keras does the loss calculation, the batch dimension is retained which is the reason for <code>axis=-1<\/code>. The returned value is a tensor. This is because the loss for each sample <strong>may<\/strong> have to be weighted before taking the mean depending on whether certain arguments are passed in the <code>fit()<\/code> method like <code>sample_weight<\/code>.<\/p>\n\n<p>I get the same results with both the approaches.<\/p>\n\n<pre><code>from tensorflow import keras\nimport numpy as np\nfrom keras import backend as K\n\ndef rmse(y_pred, y_true):\n    return K.sqrt(K.mean(K.square(y_pred - y_true)))\n\nl1 = keras.layers.Input(shape=(32))\nl2 = keras.layers.Dense(10)(l1)\nmodel = keras.Model(inputs=l1, outputs=l2)\n\ntrain_examples = np.random.randn(5,32)\ntrain_labels=np.random.randn(5,10)\n<\/code><\/pre>\n\n<p><strong>MSE approach<\/strong><\/p>\n\n<pre><code>model.compile(loss='mse', optimizer='adam')\nmodel.evaluate(train_examples, train_labels)\n<\/code><\/pre>\n\n<p><strong>RMSE approach<\/strong><\/p>\n\n<pre><code>model.compile(loss=rmse, optimizer='adam')\nmodel.evaluate(train_examples, train_labels)\n<\/code><\/pre>\n\n<p><strong>Output<\/strong><\/p>\n\n<pre><code>5\/5 [==============================] - 0s 8ms\/sample - loss: 1.9011\n5\/5 [==============================] - 0s 2ms\/sample - loss: 1.3788\n<\/code><\/pre>\n\n<p>sqrt(1.9011) = 1.3788<\/p>\n"}]}
+{"QuestionId":58131814,"AnswerCount":3,"Tags":"<machine-learning><keras><scikit-learn><neural-network>","CreationDate":"2019-09-27T09:36:01.113","AcceptedAnswerId":null,"OwnerUserId":12078387.0,"Title":"I was training an Ann machine learning model using GridSearchCV and got stuck with an IndexError in gridSearchCV","Body":"<p>My model starts to train and while executing for sometime it gives an error :-\nIndexError: index 37 is out of bounds for axis 0 with size 37<\/p>\n\n<p>It executes properly for my model without using gridsearchCV with fixed parameters<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/7dgxI.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/7dgxI.jpg\" alt=\"\"><\/a><\/p>\n\n<p>Here is my code <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>    from keras.wrappers.scikit_learn import KerasClassifier\n    from sklearn.model_selection import GridSearchCV\n    from keras.models import Sequential\n    from keras.layers import Dense\n    def build_classifier(optimizer, nb_layers,unit):\n        classifier = Sequential()\n        classifier.add(Dense(units = unit, kernel_initializer = 'uniform', activation = 'relu', input_dim = 14))\n        i = 1\n        while i &lt;= nb_layers:\n            classifier.add(Dense(activation=\"relu\", units=unit, kernel_initializer=\"uniform\"))\n            i += 1\n        classifier.add(Dense(units = 38, kernel_initializer = 'uniform', activation = 'softmax'))\n        classifier.compile(optimizer = optimizer, loss = 'sparse_categorical_crossentropy', metrics = ['accuracy'])\n        return classifier\n    classifier = KerasClassifier(build_fn = build_classifier)\n    parameters = {'batch_size': [10,25],\n                  'epochs': [100,200],\n                  'optimizer': ['adam'],\n                  'nb_layers': [5,6,7],\n                  'unit':[48,57,76]\n                 }\n    grid_search = GridSearchCV(estimator = classifier,\n                               param_grid = parameters,\n                               scoring = 'accuracy',\n                              cv=5,n_jobs=-1)\n    grid_search = grid_search.fit(X_train, y_train)\n    best_parameters = grid_search.best_params_\n    best_accuracy = grid_search.best_score_\n<\/code><\/pre>\n","answers":[{"AnswerId":"58133175","CreationDate":"2019-09-27T10:58:13.317","ParentId":null,"OwnerUserId":"3660313","Title":null,"Body":"<p>This is a multi-class classifier with a huge Number of Classes (38 classes). It seems like <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.model_selection.GridSearchCV.html\" rel=\"nofollow noreferrer\">GridSearchCV<\/a> isn't spliting your dataset by stratified sampling, may be because you haven't enough data and\/or your dataset isn't class-balanced.<\/p>\n\n<p>According to the <a href=\"https:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.model_selection.GridSearchCV.html\" rel=\"nofollow noreferrer\">documentation<\/a>: <\/p>\n\n<blockquote>\n  <p>For integer\/None inputs, if the estimator is a classifier and y is\n  either binary or multiclass, StratifiedKFold is used. In all other\n  cases, KFold is used.<\/p>\n<\/blockquote>\n\n<p>By using <strong>categorical_crossentropy<\/strong>, <a href=\"https:\/\/keras.io\/scikit-learn-api\/\" rel=\"nofollow noreferrer\">KerasClassifier<\/a> will convert targets (a class vector (integers)) to binary class matrix using <a href=\"https:\/\/keras.io\/utils\/\" rel=\"nofollow noreferrer\">keras.utils.to_categorical<\/a>. Since there are 38 classes, each target will be converted to a binary vector of dimension 38 (index from 0 to 37).<\/p>\n\n<p>I guess that in some splits, the validation set doesn't have samples from all the 38 classes, so targets are converted to vectors of dimension &lt; 38, but since GridSearchCV is fitted with samples from all the 38 classes, it expects vectors of dimension = 38, which causes this error.<\/p>\n"},{"AnswerId":"58137004","CreationDate":"2019-09-27T14:57:44.950","ParentId":null,"OwnerUserId":"11395634","Title":null,"Body":"<p>Take a look at the shape of your y_train. It need to be a some sort of one hot with shape (,37) <\/p>\n"},{"AnswerId":"58132737","CreationDate":"2019-09-27T10:29:09.237","ParentId":null,"OwnerUserId":"9875707","Title":null,"Body":"<p>The error <code>IndexError: index 37 is out of bounds for axis 0 with size 37<\/code> means that there is no element with index 37 in your object.<\/p>\n\n<p>In python, if you have an object like array or list, which has elements indexed numerically, if it has <code>n<\/code> elements, indexes will go from <code>0<\/code> to <code>n-1<\/code> (this is the general case, with the exception of reindexing in dataframes).<\/p>\n\n<p>So, if you ahve 37 elements you can only retrieve elements from 0-36.<\/p>\n"}]}
+{"QuestionId":58131815,"AnswerCount":1,"Tags":"<opencv><tensorflow><image-processing>","CreationDate":"2019-09-27T09:36:05.617","AcceptedAnswerId":58169044.0,"OwnerUserId":8244247.0,"Title":"How to perform a linear homography of an image in tensorflow","Body":"<p>I would like to be able to replicate the behaviour of the opencv function <code>warpPerspective<\/code> which takes as input an image and an homography matrix, and projects the image according to the homography matrix (more details here : <a href=\"https:\/\/docs.opencv.org\/2.4\/modules\/imgproc\/doc\/geometric_transformations.html\" rel=\"nofollow noreferrer\">https:\/\/docs.opencv.org\/2.4\/modules\/imgproc\/doc\/geometric_transformations.html<\/a>). <\/p>\n\n<p>It seems like <code>tf.contrib.image.sparse_image_warp<\/code> should do the job, but I am unable to replicate the behaviour of <code>warpPerspective<\/code>. The output I get is distorted in a non-linear fashion despite the use of the parameter <code>interpolation_order=1<\/code>. <\/p>\n\n<p>With some further research, I suspect this is due to the fact that <code>tf.contrib.image.interpolate_spline<\/code> does not perform linear interpolation even when its order is 1 but rather uses some RBF kernels.<\/p>\n\n<p>I can't see a way around this except encoding it with a <code>dense_image_warp<\/code>, but it seems a bit overkill and maybe costly. Does anyone has another solution ?<\/p>\n","answers":[{"AnswerId":"58169044","CreationDate":"2019-09-30T13:35:15.310","ParentId":null,"OwnerUserId":"8244247","Title":null,"Body":"<p>After some research, here is a solution. it uses the <code>tf.contrib.image.dense_image_warp<\/code> function and is not really pretty, but still, it works :<\/p>\n\n<p>This first function computes the optical flow needed to perform the homography :<\/p>\n\n<pre><code>    def homography_matrix_to_flow(tf_homography_matrix, im_shape1, im_shape2):\n        Y, X = np.meshgrid(range(im_shape1), range(im_shape2))\n        Z = np.ones_like(X)\n        XYZ = np.stack((X, Y, Z), axis=-1)\n        tf_XYZ = tf.constant(XYZ.astype(\"float64\"))\n        tf_XYZ = tf_XYZ[tf.newaxis,:,:, :, tf.newaxis]\n\n        tf_homography_matrix = tf.tile(tf_homography_matrix[tf.newaxis, tf.newaxis], (1, im_shape2, im_shape1, 1, 1))\n        tf_unnormalized_transformed_XYZ = tf.matmul(tf_homography_matrix, tf_XYZ, transpose_b=False)\n        tf_transformed_XYZ = tf_unnormalized_transformed_XYZ \/ tf_unnormalized_transformed_XYZ[:,:,:, -1][:,:,:, tf.newaxis]\n        flow = -tf.squeeze(tf_transformed_XYZ-tf_XYZ)[..., :2]\n\n        return flow\n<\/code><\/pre>\n\n<p>Then, it used to warp the original image to the distorted image.<\/p>\n\n<p>There is one trick : due to how the <code>tf.contrib.image.dense_image_warp<\/code> function works, you need to pass the inverse of the homography matrix to find the correct optical flow to use.  <\/p>\n\n<pre><code>        homography_matrix = np.array([[-4.86219067e-01, -2.20871298e+00,  4.08214879e+02],\n        [-1.02940133e-01, -5.60378659e+00,  3.87573763e+02],\n        [-1.35051362e-04, -6.59600583e-03,  2.91244998e-01]])\n\n        inv_homography_matrix = np.linalg.inv(homography_matrix)\n\n        tf_inv_homography_matrix = tf.constant(inv_homography_matrix)[tf.newaxis]\n        flow = homography_matrix_to_flow(tf_inv_homography_matrix, img.shape[1], img.shape[2])[tf.newaxis]\n        flow =tf.tile(flow, (self.bs, 1,1,1))\n        image_warped = tf.contrib.image.dense_image_warp(tf.transpose(img, (0,2,1,3)), flow)\n        image_warped = tf.transpose(image_warped, (0,2,1,3))\n\n<\/code><\/pre>\n\n<p>I still hope to find a better answer (one which does not have to compute a whole tensor of flow), therefore, I leave the question unanswered for now.<\/p>\n"}]}
+{"QuestionId":58132183,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><object-detection><tensorboard><faster-rcnn>","CreationDate":"2019-09-27T09:54:37.053","AcceptedAnswerId":null,"OwnerUserId":1083662.0,"Title":"How to get validation loss alongside total loss and loss in tensorboard","Body":"<p>I am using a pretrained model to train custom data-set using TensorFlow Object detection API. For this purpose I chose Faster_RCNN_inception_v2_pets model and did the training using model_main.py in order to have training and evaluation simultaneously. <\/p>\n\n<pre><code> python model_main.py --pipeline_config_path=training\/faster_rcnn_inception_v2_pets.config --model_dir=training\/ --sample_1_of_n_eval_examples=1 --alsologtostderr\n<\/code><\/pre>\n\n<p>I am using PAScal_VOC metrics for evaluation. \nHowever, I am interested in having both training and validation losses and be able to compare them, but all tensorborad gives me is training losses, the bounding box IOU AP, validation image comparison to the predicted ones and the total loss and then in the end loss. \nI read some resources on how to include the validation loss like in <a href=\"https:\/\/stackoverflow.com\/questions\/34471563\/logging-training-and-validation-loss-in-tensorboard\">here<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/49905773\/how-to-log-validation-loss-and-accuracy-using-tfslim\">here<\/a>, but I have no clue to which file I should attach these lines of code to, or even which file to edit in my folder. Any suggestions? <\/p>\n","answers":[]}
+{"QuestionId":58132955,"AnswerCount":1,"Tags":"<machine-learning><keras><neural-network>","CreationDate":"2019-09-27T10:44:37.623","AcceptedAnswerId":58135171.0,"OwnerUserId":19808.0,"Title":"How to evaluate Keras model precision for only some of the classes","Body":"<p>I have the following Keras model, whose output has 3 classes (0, 1, 2):<\/p>\n\n<pre><code>    model = Sequential()\n    model.add(LSTM(100, input_shape=(n_time_steps,n_features)))\n    model.add(Dropout(0.5))\n    model.add(Dense(100, activation='relu'))\n    model.add(Dense(n_classes, activation='softmax'))\n    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>I'm interested in evaluating the model for precision of classes 1 and 2, not 0, i.e. the model is good if it has the smallest number of false positives in classes 1 and 2, while I don't really care about class 0.<\/p>\n\n<p>How can I write such metrics in Keras?<\/p>\n","answers":[{"AnswerId":"58135171","CreationDate":"2019-09-27T13:04:25.717","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Create any function taking the ground truth and the predicted results and calculate your metrics:<\/p>\n\n<pre><code>def false_positives(y_true, y_pred):\n    negatives = 1 - y_true   \n    y_pred = K.cast(K.greater(y_pred, 0.5), K.floatx()) #round to 0 or 1\n    #if you don't round y_pred, it might even serve as a loss function\n\n    falsePositives = y_pred * negatives\n    falsePositives1 = falsePositives[:,1]\n    falsePositives2 = falsePositives[:,2]\n\n    return something\n<\/code><\/pre>\n\n<p>Use <code>metrics = [false_positives]<\/code>, or <code>metrics=['accuracy', false_positives_1, false_positives_2]<\/code>, etc. <\/p>\n"}]}
+{"QuestionId":58133056,"AnswerCount":2,"Tags":"<tensorflow><image-processing><keras><conv-neural-network>","CreationDate":"2019-09-27T10:51:19.800","AcceptedAnswerId":null,"OwnerUserId":8577917.0,"Title":"Increasing epochs vs increasing training data","Body":"<p>For image analysis:<\/p>\n\n<p>Is it better to increase the epochs, let's say from 2 to 4, for 40000 images. This takes double the time.<\/p>\n\n<p>OR<\/p>\n\n<p>Is it better to increase the size of the training data, from 40000 to 80000 but with 2 epochs only. This will also take double the time.<\/p>\n\n<p>Since increasing both the number of epochs and the training data will take a lot of time, I can only do one.<\/p>\n\n<p>What should be the choice? <\/p>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"58133116","CreationDate":"2019-09-27T10:55:20.653","ParentId":null,"OwnerUserId":"3214872","Title":null,"Body":"<p>Caveats aside (bad\/confusing samples, ...), increasing data is always preferred.\nThe reason for this is generalization: you can show the same image N times to the network, or N different images. In the first case, it will overfit to the training dataset and fail to generalize to new images.<\/p>\n\n<p>That's also the reason why data augmentation techniques exist: if you don't have any \"new\" data to train on, you can attemp to generate \"new\" samples applying transformations to the ones you have.<\/p>\n\n<p>Of course, more data means bigger datasets to gather, clean, annotate, store, distribute, which eventually puts a limit to the size of real-world datasets.\nBut if in your case you have available data to train on, use it.<\/p>\n"},{"AnswerId":"58407663","CreationDate":"2019-10-16T07:11:00.590","ParentId":null,"OwnerUserId":"12192748","Title":null,"Body":"<p>Having more data is always a good approach but also having more epoch will lead to overfitting while less epochs will lead to underfitting. You can choose to have <strong>EarlyStopping<\/strong> in Keras which will stop training model at certain epoch once model performance stops improving.<\/p>\n\n<p>Also if the data is limited you can augment data which will boost the number of different images you will use to train your model.<\/p>\n\n<p>Please refer to openCV and scikit-image for different image transformations techniques like:<\/p>\n\n<ul>\n<li><strong>Affine transformations<\/strong><\/li>\n<li><strong>Perspective transformations<\/strong><\/li>\n<\/ul>\n"}]}
+{"QuestionId":58133430,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tf.keras>","CreationDate":"2019-09-27T11:13:43.760","AcceptedAnswerId":58136256.0,"OwnerUserId":3884340.0,"Title":"How to substitute `keras.layers.merge._Merge` in `tensorflow.keras`","Body":"<p>I want to create a custom Merge layer using the <code>tf.keras<\/code> API. However, the new API hides the <code>keras.layers.merge._Merge<\/code> class that I want to inherit from.<\/p>\n\n<p>The purpose of this is to create a Layer that can perform a weighted sum\/merge of the outputs of two different layers. Before, and in <code>keras<\/code> python API (not the one included in <code>tensorflow.keras<\/code>) I could inherit from <code>keras.layers.merge._Merge<\/code> class, which is now not accessible from <code>tensorflow.keras<\/code>.<\/p>\n\n<p>Where before I could do this<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class RandomWeightedAverage(keras.layers.merge._Merge):\n    def __init__(self, batch_size):\n        super().__init__()\n        self.batch_size = batch_size\n    def _merge_function(self, inputs):\n        alpha = K.random_uniform((self.batch_size, 1, 1, 1))\n        return (alpha * inputs[0]) + ((1 - alpha) * inputs[1])\n<\/code><\/pre>\n\n<p>Now I cannot use the same logic if using <code>tensorflow.keras<\/code><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class RandomWeightedAverage(tf.keras.layers.merge._Merge):\n    def __init__(self, batch_size):\n        super().__init__()\n        self.batch_size = batch_size\n    def _merge_function(self, inputs):\n        alpha = K.random_uniform((self.batch_size, 1, 1, 1))\n        return (alpha * inputs[0]) + ((1 - alpha) * inputs[1])\n<\/code><\/pre>\n\n<p>Produces<\/p>\n\n<pre><code>AttributeError: module 'tensorflow.python.keras.api._v1.keras.layers' has no attribute 'merge'\n<\/code><\/pre>\n\n<p>I have also tried inheriting from <code>Layer<\/code> class instead<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class RandomWeightedAverage(tensorflow.keras.layers.Layer):\n    def __init__(self, batch_size):\n        super().__init__()\n        self.batch_size = batch_size\n    def call(self, inputs):\n        alpha = K.random_uniform((self.batch_size, 1, 1, 1))\n        return (alpha * inputs[0]) + ((1 - alpha) * inputs[1])\n<\/code><\/pre>\n\n<p>which gives me a layer with output shape equals to <code>multiple<\/code>, whereas I want the output shape to be well defined. I further attempted<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class RandomWeightedAverage(tensorflow.keras.layers.Layer):\n    def __init__(self, batch_size):\n        super().__init__()\n        self.batch_size = batch_size\n    def call(self, inputs):\n        alpha = K.random_uniform((self.batch_size, 1, 1, 1))\n        return (alpha * inputs[0]) + ((1 - alpha) * inputs[1])\n\n    def compute_output_shape(self, input_shape):\n        return input_shape[0]\n<\/code><\/pre>\n\n<p>But this did not solve the <code>multiple<\/code> ambiguity as output shape.<\/p>\n","answers":[{"AnswerId":"58136256","CreationDate":"2019-09-27T14:11:31.240","ParentId":null,"OwnerUserId":"6537424","Title":null,"Body":"<p>I have slightly modified your code to use <code>tf.random_uniform<\/code> instead of <code>K.random_uniform<\/code> and it's working fine on 1.13.1 and 1.14.0 (full snippet and resulting <code>model.summary()<\/code> below). <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nprint(tf.__version__)\n\n\nclass RandomWeightedAverage(tf.keras.layers.Layer):\n    def __init__(self, batch_size):\n        super().__init__()\n        self.batch_size = batch_size\n\n    def call(self, inputs, **kwargs):\n        alpha = tf.random_uniform((self.batch_size, 1, 1, 1))\n        return (alpha * inputs[0]) + ((1 - alpha) * inputs[1])\n\n    def compute_output_shape(self, input_shape):\n        return input_shape[0]\n\n\nx1 = tf.keras.layers.Input((32, 32, 1))\nx2 = tf.keras.layers.Input((32, 32, 1))\n\ny = RandomWeightedAverage(4)(inputs=[x1, x2])\n\nmodel = tf.keras.Model(inputs=[x1, x2], outputs=[y])\n\nprint(model.summary())\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/lcmrt.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lcmrt.png\" alt=\"model summary\"><\/a><\/p>\n"}]}
+{"QuestionId":58133681,"AnswerCount":3,"Tags":"<tensorflow><tensorflow-lite>","CreationDate":"2019-09-27T11:29:41.800","AcceptedAnswerId":null,"OwnerUserId":9337714.0,"Title":"Tensorflow frozen_graph to tflite for android app","Body":"<p>I am a newbie on the TensorFlow object detection library. I have a specific data set what I have to produce myself and labeled it with thousands of jpg. I have run the file to detect the object from these images.. any way.The end of the process i have gotten frozen_graph and from it I exported model.ckpl file to inference graph folder everything goes fine, and I have tested model.ckpl model on the object_detection.ipynb file it works fine. Until this step, there is no problem. \nHowever,Am not able to understand how could convert that model.ckpl file to model.tflite file to use on android studio app. <\/p>\n\n<p>I have see many things <a href=\"https:\/\/stackoverflow.com\/questions\/50632152\/tensorflow-convert-pb-file-to-tflite-using-python\">like<\/a> but I am no idea what is the input_tensors = [...]\noutput_tensors = [...] \nI may already know but what it was actually...<\/p>\n\n<p>Could you show me how could I convert it? <\/p>\n","answers":[{"AnswerId":"58134588","CreationDate":"2019-09-27T12:29:50.410","ParentId":null,"OwnerUserId":"8105979","Title":null,"Body":"<p>Use tensorboard to find out your input and output layer. For reference follow these links - <\/p>\n\n<p><a href=\"https:\/\/heartbeat.fritz.ai\/intro-to-machine-learning-on-android-how-to-convert-a-custom-model-to-tensorflow-lite-e07d2d9d50e3\" rel=\"nofollow noreferrer\">https:\/\/heartbeat.fritz.ai\/intro-to-machine-learning-on-android-how-to-convert-a-custom-model-to-tensorflow-lite-e07d2d9d50e3<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/50632152\/tensorflow-convert-pb-file-to-tflite-using-python\">Tensorflow Convert pb file to TFLITE using python<\/a><\/p>\n"},{"AnswerId":"58407319","CreationDate":"2019-10-16T06:50:37.913","ParentId":null,"OwnerUserId":"12192748","Title":null,"Body":"<p>If you have trained your model from scratch you must be having the .meta file. Also you need to specify the output node names using which you can create a .pb file. Please refer to below link on steps to create this file:<\/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\">Tensorflow: How to convert .meta, .data and .index model files into one graph.pb file<\/a><\/p>\n\n<p>Once this is created you can further convert your .pb to tflite as below:<\/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)\n\ntflite_model = converter.convert()\n\nopen(\"converted_model.tflite\", \"wb\").write(tflite_model)\n<\/code><\/pre>\n"},{"AnswerId":"58210930","CreationDate":"2019-10-03T00:31:28.917","ParentId":null,"OwnerUserId":"10952155","Title":null,"Body":"<p>If you don't know your inputs and outputs, use summarize_graph tool and feed it your frozen model.\nSee command here\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/tools\/graph_transforms#inspecting-graphs\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/tools\/graph_transforms#inspecting-graphs<\/a><\/p>\n"}]}
+{"QuestionId":58133909,"AnswerCount":0,"Tags":"<deep-learning><pytorch><resnet><densenet><hinge-loss>","CreationDate":"2019-09-27T11:44:38.913","AcceptedAnswerId":null,"OwnerUserId":570593.0,"Title":"Densenet with Hinge loss on CIFAR dataset","Body":"<p>I am trying to use Hinge loss with densenet on the CIFAR 100 dataset. The learning converges to some point and after that there is no learning. The accuracy is much less than Densenet with CrossEntropy loss function. I tried with different learning rates and weight decays. <\/p>\n\n<p>Any ideas on why I am unable to train properly Densenet with Hinge loss? I am able to use Hinge loss with Resnet without any problem.<\/p>\n","answers":[]}
+{"QuestionId":58134005,"AnswerCount":1,"Tags":"<python><keras><deep-learning><image-segmentation>","CreationDate":"2019-09-27T11:51:00.727","AcceptedAnswerId":58159647.0,"OwnerUserId":11486611.0,"Title":"Keras U-Net weighted loss implementation","Body":"<p>I'm trying to separate close objects as was shown in the U-Net paper (<a href=\"https:\/\/arxiv.org\/pdf\/1505.04597.pdf\" rel=\"nofollow noreferrer\">here<\/a>). For this, one generates weight maps which can be used for pixel-wise losses. The following code describes the network I use from <a href=\"https:\/\/jaidevd.github.io\/posts\/weighted-loss-functions-for-instance-segmentation\/\" rel=\"nofollow noreferrer\">this<\/a> blog post.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x_train_val = # list of images (imgs, 256, 256, 3)\ny_train_val = # list of masks (imgs, 256, 256, 1)\ny_weights = # list of weight maps (imgs, 256, 256, 1) according to the blog post \n# visual inspection confirms the correct calculation of these maps\n\n# Blog posts' loss function\ndef my_loss(target, output):\n    return - tf.reduce_sum(target * output,\n                           len(output.get_shape()) - 1)\n\n# Standard Unet model from blog post\n_epsilon = tf.convert_to_tensor(K.epsilon(), np.float32)\n\ndef make_weighted_loss_unet(input_shape, n_classes):\n    ip = L.Input(shape=input_shape)\n    weight_ip = L.Input(shape=input_shape[:2] + (n_classes,))\n\n    conv1 = L.Conv2D(64, 3, activation='relu', padding='same', kernel_initializer='he_normal')(ip)\n    conv1 = L.Conv2D(64, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv1)\n    conv1 = L.Dropout(0.1)(conv1)\n    mpool1 = L.MaxPool2D()(conv1)\n\n    conv2 = L.Conv2D(128, 3, activation='relu', padding='same', kernel_initializer='he_normal')(mpool1)\n    conv2 = L.Conv2D(128, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv2)\n    conv2 = L.Dropout(0.2)(conv2)\n    mpool2 = L.MaxPool2D()(conv2)\n\n    conv3 = L.Conv2D(256, 3, activation='relu', padding='same', kernel_initializer='he_normal')(mpool2)\n    conv3 = L.Conv2D(256, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv3)\n    conv3 = L.Dropout(0.3)(conv3)\n    mpool3 = L.MaxPool2D()(conv3)\n\n    conv4 = L.Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(mpool3)\n    conv4 = L.Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv4)\n    conv4 = L.Dropout(0.4)(conv4)\n    mpool4 = L.MaxPool2D()(conv4)\n\n    conv5 = L.Conv2D(1024, 3, activation='relu', padding='same', kernel_initializer='he_normal')(mpool4)\n    conv5 = L.Conv2D(1024, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv5)\n    conv5 = L.Dropout(0.5)(conv5)\n\n    up6 = L.Conv2DTranspose(512, 2, strides=2, kernel_initializer='he_normal', padding='same')(conv5)\n    conv6 = L.Concatenate()([up6, conv4])\n    conv6 = L.Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv6)\n    conv6 = L.Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv6)\n    conv6 = L.Dropout(0.4)(conv6)\n\n    up7 = L.Conv2DTranspose(256, 2, strides=2, kernel_initializer='he_normal', padding='same')(conv6)\n    conv7 = L.Concatenate()([up7, conv3])\n    conv7 = L.Conv2D(256, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv7)\n    conv7 = L.Conv2D(256, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv7)\n    conv7 = L.Dropout(0.3)(conv7)\n\n    up8 = L.Conv2DTranspose(128, 2, strides=2, kernel_initializer='he_normal', padding='same')(conv7)\n    conv8 = L.Concatenate()([up8, conv2])\n    conv8 = L.Conv2D(128, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv8)\n    conv8 = L.Conv2D(128, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv8)\n    conv8 = L.Dropout(0.2)(conv8)\n\n    up9 = L.Conv2DTranspose(64, 2, strides=2, kernel_initializer='he_normal', padding='same')(conv8)\n    conv9 = L.Concatenate()([up9, conv1])\n    conv9 = L.Conv2D(64, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv9)\n    conv9 = L.Conv2D(64, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv9)\n    conv9 = L.Dropout(0.1)(conv9)\n\n    c10 = L.Conv2D(n_classes, 1, activation='softmax', kernel_initializer='he_normal')(conv9)\n\n    # Mimic crossentropy loss\n    c11 = L.Lambda(lambda x: x \/ tf.reduce_sum(x, len(x.get_shape()) - 1, True))(c10)\n    c11 = L.Lambda(lambda x: tf.clip_by_value(x, _epsilon, 1. - _epsilon))(c11)\n    c11 = L.Lambda(lambda x: K.log(x))(c11)\n    weighted_sm = L.multiply([c11, weight_ip])\n\n    model = Model(inputs=[ip, weight_ip], outputs=[weighted_sm])\n    return model\n<\/code><\/pre>\n\n<p>I then compile and fit the model as is shown below:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = make_weighted_loss_unet((256, 256, 3), 1) # shape of input, number of classes\nmodel.compile(optimizer='adam', loss=my_loss, metrics=['acc'])\nmodel.fit([x_train_val, y_weights], y_train_val, validation_split=0.1, epochs=1)\n<\/code><\/pre>\n\n<p>The model can then train as usual. However, the loss doesn't seem to improve much. Furthermore, when I try to predict on new images, I obviously don't have the weight maps (because they are calculated on the labeled masks). I tried to use empty \/ zero arrays shaped like the weight map but that only yields in blank \/ zero predictions. I also tried different metrics and more standards losses without any success.<\/p>\n\n<p>Did anyone face the same issue or have an alternative in implementing this weighted loss? Thanks in advance. BBQuercus<\/p>\n","answers":[{"AnswerId":"58159647","CreationDate":"2019-09-29T22:20:32.253","ParentId":null,"OwnerUserId":"834565","Title":null,"Body":"<h1>A simpler way to write custom loss with pixel weights<\/h1>\n\n<p>In your code, the loss is scattered around, between <code>my_loss<\/code> and <code>make_weighted_loss_unet<\/code> functions. You can add targets as an input and use <code>model.add_loss<\/code> to structure the code better :<\/p>\n\n<pre><code>def make_weighted_loss_unet(input_shape, n_classes):\n    ip = L.Input(shape=input_shape)\n    weight_ip = L.Input(shape=input_shape[:2] + (n_classes,))\n    targets   = L.input(shape=input_shape[:2] + (n_classes,))\n    # .... rest of your model definition code ...\n\n    c10 = L.Conv2D(n_classes, 1, activation='softmax', kernel_initializer='he_normal')(conv9)\n    model.add_loss(pixel_weighted_cross_entropy(weights_ip, targets, c10))\n    # .... return Model .... NO NEED to specify loss in model.compile\n\ndef pixel_weighted_cross_entropy(weights, targets, predictions)\n    loss_val = keras.losses.categorical_crossentropy(targets, predictions)\n    weighted_loss_val = weights * loss_val\n    return K.mean(weighted_loss_val)\n<\/code><\/pre>\n\n<p>If you don't refactor your code to the above approach, next section shows how to still run inference without issues<\/p>\n\n<h1>How to run your model in inference<\/h1>\n\n<h2>Option 1 : Use another <code>Model<\/code> object for inference<\/h2>\n\n<p>You can create a <code>Model<\/code> used for training and another used for inference. Both are largely the same except that the inference <code>Model<\/code> does not take <code>weights_ip<\/code>, and gives an early output <code>c10<\/code>.<\/p>\n\n<p>Here's an example code that adds an argument <code>is_training=True<\/code> to decide which <code>Model<\/code> to return :<\/p>\n\n<pre><code>def make_weighted_loss_unet(input_shape, n_classes, is_training=True):\n    ip = L.Input(shape=input_shape)\n\n    conv1 = L.Conv2D(64, 3, activation='relu', padding='same', kernel_initializer='he_normal')(ip)\n    # .... rest of your model definition code ...\n    c10 = L.Conv2D(n_classes, 1, activation='softmax', kernel_initializer='he_normal')(conv9)\n\n    if is_training:\n        # Mimic crossentropy loss\n        c11 = L.Lambda(lambda x: x \/ tf.reduce_sum(x, len(x.get_shape()) - 1, True))(c10)\n        c11 = L.Lambda(lambda x: tf.clip_by_value(x, _epsilon, 1. - _epsilon))(c11)\n        c11 = L.Lambda(lambda x: K.log(x))(c11)\n        weight_ip = L.Input(shape=input_shape[:2] + (n_classes,))\n        weighted_sm = L.multiply([c11, weight_ip])\n        return Model(inputs=[ip, weight_ip], outputs=[weighted_sm])\n    else:\n        return Model(inputs=[ip], outputs=[c10]) \n    return model\n<\/code><\/pre>\n\n<h2>Option 2 : Use <code>K.function<\/code><\/h2>\n\n<p>If you don't want to mess with your Model definition method (<code>make_weighted_loss_unet<\/code>) and want to achieve the same result outside, you can use a function that extracts the subgraph relevant for inference.<\/p>\n\n<p>In your inference function:<\/p>\n\n<pre><code>from keras import backend as K\n\nmodel = make_weighted_loss_unet(input_shape, n_classes)\ninference_function = K.function([model.get_layer(\"input_layer\").input], \n                                [model.get_layer(\"output_softmax_layer\").output])\npredicted_heatmap = inference_function(new_image)\n<\/code><\/pre>\n\n<p>Note that you'll have to give <code>name=<\/code> to your <code>ip<\/code> layer and <code>c10<\/code> layer to be able to retrieve them via <code>model.get_layer(name)<\/code> :<\/p>\n\n<pre><code>ip = L.Input(shape=input_shape, name=\"input_layer\")\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code>c10 = L.Conv2D(n_classes, 1, activation='softmax', kernel_initializer='he_normal', name=\"output_softmax_layer\")(conv9)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58134030,"AnswerCount":1,"Tags":"<python><keras><deep-learning>","CreationDate":"2019-09-27T11:52:45.687","AcceptedAnswerId":null,"OwnerUserId":3630804.0,"Title":"Diverging loss in Keras with custom loss","Body":"<p>I have a fully connected feed-forward network implemented with Keras. Initially, I used binary cross-entropy as the loss and the metric, and Adam optimizer as follows<\/p>\n\n<pre><code>adam = keras.optimizers.Adam(lr=0.01, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False)\nmodel.compile(optimizer=adam, loss='binary_crossentropy', metrics=['binary_crossentropy']) \n<\/code><\/pre>\n\n<p>This model trains well and gives good results. In order to get better results I want to use a different loss function and metric as below,<\/p>\n\n<pre><code>import keras.backend as K\n\ndef soft_bit_error_loss(yTrue, yPred):\n    loss = K.pow(1 - yPred, yTrue) * K.pow(yPred, 1-yTrue)\n    return K.mean(loss)\n\ndef ber(yTrue, yPred):\n    x_hat_train = K.cast(K.greater(yPred, 0.5), 'uint8')\n    train_errors = K.cast(K.not_equal(K.cast(yTrue, 'uint8'), x_hat_train), 'float32')\n    train_ber = K.mean(train_errors)\n    return train_ber\n<\/code><\/pre>\n\n<p>I use it to compile my model as below<\/p>\n\n<pre><code>model.compile(optimizer=adam, loss=soft_bit_error_loss, metrics=[ber])\n<\/code><\/pre>\n\n<p>However, when I do that, the loss and the metric diverge after some training, everytime as in the following pictures.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/XoPod.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/XoPod.png\" alt=\"custom loss\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/VdXC3.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/VdXC3.png\" alt=\"ber\"><\/a><\/p>\n\n<p>What can be the cause of this?<\/p>\n","answers":[{"AnswerId":"58139766","CreationDate":"2019-09-27T18:26:30.937","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Your loss function is very unstable, look at it:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/nWdO9.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/nWdO9.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Where I replaced <code>y_pred<\/code> (variable) with <code>x<\/code> and <code>y_true<\/code> (constant) with <code>c<\/code> for simplicity.<\/p>\n\n<p>As your predictions approach zero, at least one operation will tend to 1\/0, which is infinite. Although by the limits theory you can know the result is ok, Keras doesn't know the \"whole\" function as one, it calculates derivatives based on the basic operations used. <\/p>\n\n<p>So, one easy solution is the one pointed by @today:<\/p>\n\n<pre><code>loss = K.switch(yTrue == 1, 1 - yPred, yPred)\n<\/code><\/pre>\n\n<p>It's exactly the same function (difference only when <code>c<\/code> is not zero or 1).   <\/p>\n\n<p>Also, even easier, for <code>c=0<\/code> or <code>c=1<\/code>, it's just a plain <code>loss='mae'<\/code>.   <\/p>\n"}]}
+{"QuestionId":58134103,"AnswerCount":1,"Tags":"<python><docker><tensorflow><tensorflow-serving>","CreationDate":"2019-09-27T11:57:12.493","AcceptedAnswerId":58369500.0,"OwnerUserId":5502377.0,"Title":"TF - how do I setup the model signature correctly for serving with Docker?","Body":"<p>I am trying to understand how to setup a TF model for serving with docker. I manged to install docker, I know how to export a trained model as a .pb. What I dont understand is how to correctly define the model signature for serving. I want to simply call a trained model in the terminal with docker. Could you explain me what I have to change in the following example?<\/p>\n\n<p>I am doing the following steps:<\/p>\n\n<p>1) create directory <code>\/tmp\/serving_minimal<\/code>, cd to it in the terminal <code>$cd \/tmp\/serving_minimal<\/code><\/p>\n\n<p>2) save in <code>\/tmp\/serving_minimal<\/code> the file <code>generate_model.py<\/code> for the following code<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport tensorflow as tf\nimport os, shutil\n\n#%% Data\n\n# Input (2D)\nx = np.array([[x1,x2] for x1 in np.linspace(10,20,4) for x2 in np.linspace(-7,-3,3)])\n\n# Output (3D)\nf = np.array([[np.sin(np.sum(xx)),np.cos(np.sum(xx)),np.cos(np.sum(xx))**2] for xx in x])\n\n#%% Model\n\nprint('**********************************************')\nprint('TF - save')\n\n# Dimension of input x and output f\nd_x = x.shape[-1]\nd_f = f.shape[-1]\n\n# Placeholders\nx_p = tf.placeholder(tf.float64,[None,d_x],'my_x_p')\nf_p = tf.placeholder(tf.float64,[None,d_f],'my_f_p')\n\n# Model\nmodel = x_p\nmodel = tf.layers.dense(model,7,tf.tanh)\nmodel = tf.layers.dense(model,5,tf.tanh)\nmodel = tf.layers.dense(model,d_f,None)\nmodel = tf.identity(model,'my_model')\n\n# Session\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\n\n# Evaluate for later check of serving\nf_model = sess.run(model,{x_p:x})\nfolder = 'data'\nif not os.path.exists(folder):\n    os.mkdir(folder)\nnp.savetxt('data\/x.dat',f_model)\nnp.savetxt('data\/f_model.dat',f_model)\n\n# Save model\nfolder = 'saved\/model\/001'\nif os.path.exists(folder):\n    shutil.rmtree(folder)\n    print('Old model deleted')\nsaver = tf.saved_model.builder.SavedModelBuilder(folder)\n############################################\n# HOW DO I SET UP THE SIGNATURE CORRECTLY?\n############################################\ninfo_input = tf.saved_model.utils.build_tensor_info(x_p)\ninfo_output = tf.saved_model.utils.build_tensor_info(model)\nsignature = tf.saved_model.signature_def_utils.build_signature_def(\n        inputs={'x':info_input}\n        ,outputs={'f':info_output}\n        ,method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME\n        )\nsaver.add_meta_graph_and_variables(\n        sess\n        ,[tf.saved_model.tag_constants.SERVING]\n        ,signature_def_map={'predict':signature}\n        ####################################################################\n        ### WHAT DO I NEED TO PUT HERE IN ORDER TO CALL THE MODEL LATER ON \n        ### WHILE SERVING WITH DOCKER AND HOW DO I CALL IT IN DOCKER??\n        ####################################################################\n        )\nsaver.save()\n\n# Close and clean up\nsess.close()\ntf.reset_default_graph()\n\n#%% Load in Python and check\n\nprint('**********************************************')\nprint('TF - load in Python')\n\n# Session\nsess = tf.Session()\n\n# Load\ntf.saved_model.loader.load(\n        sess\n        ,[tf.saved_model.tag_constants.SERVING]\n        ,folder\n        )\n\n# Extract operations from graph\ngraph = tf.get_default_graph()\nx_p = graph.get_tensor_by_name('my_x_p:0')\nf_p = graph.get_tensor_by_name('my_f_p:0')\nmodel = graph.get_tensor_by_name('my_model:0')\n\n# Evaluate model\nf_model2 = sess.run(model,{x_p:x})\nprint(f_model - f_model2)\n\n# Close and clean up\nsess.close()\ntf.reset_default_graph()\n<\/code><\/pre>\n\n<p>4) run script in the terminal <code>$python generate_model.py<\/code> (exports the model and loads it in Python for checking)<\/p>\n\n<p>5) start docker in the terminal <code>$sudo docker ps<\/code><\/p>\n\n<p>6) run model in docker<\/p>\n\n<pre><code>$ sudo docker run \\\n    -p 8501:8501 \\\n    --name my_container \\\n    --mount type=bind,source=\/tmp\/serving_minimal\/saved\/model,target=\/models\/model1 \\\n    -e MODEL_NAME=model1 \\\n    -t tensorflow\/serving &amp;\n<\/code><\/pre>\n\n<p>7) Check that model is active (it is)<\/p>\n\n<pre><code>$ sudo docker ps\n<\/code><\/pre>\n\n<p>8) [Error] Try to evaluate active model<\/p>\n\n<pre><code>$ curl -d '{\"x\": [[1.0,2.0],[10.0,20.0]]}' -X POST http:\/\/localhost:8501\/v1\/models\/model1:predict\n<\/code><\/pre>\n\n<p>At 8) I get the error <\/p>\n\n<pre><code>{ \"error\": \"Serving signature name: \\\"serving_default\\\" not found in signature def\" }\n<\/code><\/pre>\n\n<p>but I do not understand the TF signature definition commands well enough in order to know what to do. Can you show me what needs to be corrected? Thank you!<\/p>\n","answers":[{"AnswerId":"58369500","CreationDate":"2019-10-14T02:47:11.827","ParentId":null,"OwnerUserId":"10340061","Title":null,"Body":"<p>See this line in your training code<\/p>\n\n<pre><code>signature_def_map={'predict':signature}\n<\/code><\/pre>\n\n<p>So you should change your post request body like this<\/p>\n\n<pre><code>{\n  'signature_name': 'predict',\n  'x': [[1.0,2.0],[10.0,20.0]]\n}\n<\/code><\/pre>\n\n<p>Or change 'predict' to 'serving_default' in your training code, and you could not specify signature_name in post request<\/p>\n"}]}
+{"QuestionId":58134293,"AnswerCount":1,"Tags":"<python><linux><tensorflow><shared-libraries>","CreationDate":"2019-09-27T12:10:27.253","AcceptedAnswerId":58155520.0,"OwnerUserId":1179925.0,"Title":"How to list .so files that are loaded on python script run?","Body":"<p>How to list all .so files (full path) that are loaded on python script run?<\/p>\n\n<p>Is it possible to get full path of (for example) <code>libcudart.so.10.1<\/code>?\nIs it possible to get list of .so (full path) that are loaded?<\/p>\n\n<p>For example I'm running:<\/p>\n\n<pre><code>python -c \"import tensorflow as tf; tf.test.is_gpu_available()\"\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>2019-09-27 15:02:27.186029: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcuda.so.1\n2019-09-27 15:02:27.657901: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1640] Found device 0 with properties:\nname: TITAN RTX major: 7 minor: 5 memoryClockRate(GHz): 1.77\npciBusID: 0000:17:00.0\n2019-09-27 15:02:27.658580: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1640] Found device 1 with properties:\nname: TITAN RTX major: 7 minor: 5 memoryClockRate(GHz): 1.77\npciBusID: 0000:65:00.0\n2019-09-27 15:02:27.658766: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudart.so.10.1\n2019-09-27 15:02:27.659868: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcublas.so.10\n2019-09-27 15:02:27.661073: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcufft.so.10\n2019-09-27 15:02:27.661305: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcurand.so.10\n2019-09-27 15:02:27.662477: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcusolver.so.10\n2019-09-27 15:02:27.663054: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcusparse.so.10\n2019-09-27 15:02:27.665455: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudnn.so.7\n2019-09-27 15:02:27.667986: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1763] Adding visible gpu devices: 0, 1\n2019-09-27 15:02:27.668031: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:42] Successfully opened dynamic library libcudart.so.10.1\n2019-09-27 15:02:27.669523: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1181] Device interconnect StreamExecutor with strength 1 edge matrix:\n2019-09-27 15:02:27.669535: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1187]      0 1\n2019-09-27 15:02:27.669542: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1200] 0:   N N\n2019-09-27 15:02:27.669546: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1200] 1:   N N\n2019-09-27 15:02:27.674204: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1326] Created TensorFlow device (\/device:GPU:0 with 22845 MB memory) -&gt; physical GPU (device: 0, name: TITAN RTX, pci bus id: 0000:17:00.0, compute capability: 7.5)\n2019-09-27 15:02:27.675838: I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1326] Created TensorFlow device (\/device:GPU:1 with 22823 MB memory) -&gt; physical GPU (device: 1, name: TITAN RTX, pci bus id: 0000:65:00.0, compute capability: 7.5)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58155520","CreationDate":"2019-09-29T13:09:57.470","ParentId":null,"OwnerUserId":"12098405","Title":null,"Body":"<p>When running under Linux, depending on you environment, you can get the information using one of those alternatives:<\/p>\n\n<ol>\n<li>ld.so debug flags<\/li>\n<li>strace<\/li>\n<\/ol>\n\n<p>With ld.so debug flags, you can activate debug mode by setting the LD_DEBUG environment variable to 'libs' (or even 'all') before your run the python code (depending on your shell: export LD_DEBUG=libs, or setenv LD_DEBUG libs). Dynamically loaded libraries search path will be displayed to stderr.<\/p>\n\n<p>Look at ld.so man page for additional options.<\/p>\n\n<p>Sample output - from LD_DEBUG=libs (using python 'import json')<\/p>\n\n<pre><code>owner@vm1:~\/Project\/stackoverflow\/58152351$ python\n     28796: find library=libc.so.6 [0]; searching\n     28796:  search cache=\/etc\/ld.so.cache\n     28796:   trying file=\/lib\/x86_64-linux-gnu\/libc.so.6\n     28796: \n     28796: find library=libpthread.so.0 [0]; searching\n     28796:  search cache=\/etc\/ld.so.cache\n     28796:   trying file=\/lib\/x86_64-linux-gnu\/libpthread.so.0\n     28796: \n     28796: find library=libdl.so.2 [0]; searching\n     28796:  search cache=\/etc\/ld.so.cache\n     28796:   trying file=\/lib\/x86_64-linux-gnu\/libdl.so.2\n...\n&gt;&gt;&gt; import json\n     28796: \n     28796: calling init: \/usr\/lib\/python2.7\/lib-dynload\/_json.x86_64-linux-gnu.so\n     28796: \n<\/code><\/pre>\n\n<p>As an alternative, strace can be used, but parsing it's output requires significantly more effort. The 'stat' lines may provide information about the search path for each library.<\/p>\n"}]}
+{"QuestionId":58134311,"AnswerCount":1,"Tags":"<pytorch><lstm><operation><sequence-modeling>","CreationDate":"2019-09-27T12:11:34.057","AcceptedAnswerId":58139582.0,"OwnerUserId":12129529.0,"Title":"How the function nn.LSTM behaves within the batches\/ seq_len?","Body":"<p>I\u2019m currently learning to use <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#lstm\" rel=\"nofollow noreferrer\">nn.LSTM<\/a> with pytorch and had to ask how the function is working.<\/p>\n\n<p>Basically I\u2019m trying to feed my dataset matrix (M x N).\nSince the dataset is a matrix, I wanted to feed the dataset recursively(as timesteps) into the LSTM network with <strong>Dataloader<\/strong>(utils.data.Dataset).<\/p>\n\n<p>The point where i got confused was the size of <strong>input<\/strong>(seq_len, batch, input_size)<\/p>\n\n<p>Let\u2019s say I\u2019m getting my data_loader with batch_size=10.\nIn order to generate the train_loader with the right form, I had to make the previous size of (M x N) into the size including the sequence_length which could simply be transformed to (M\/seq_len, seq_len, N).<\/p>\n\n<p>Then the input size of my nn.LSTM would be like:\n(M\/seq_len\/batch_size, seq_len, N)\n<br\/><br\/><\/p>\n\n<h2>So, my main question comes:<\/h2>\n\n<ol>\n<li><p>If i feed this data size into the LSTM model nn.LSTM(N, hidden_size),\nis the LSTM model already doing the recursive feed-forward within the whole batch?<\/p><\/li>\n<li><p>I'm also confused with the seq_len, while seq_len>1, the output will get the dimension of seq_len.\nWould that mean the output contains the recursive operations of sequences?<\/p><\/li>\n<\/ol>\n\n<p>I\u2019m not sure i made the questions clear, but my understanding is getting quite messed up..lol\nHope somebody could help me organizing the right understanding.<\/p>\n","answers":[{"AnswerId":"58139582","CreationDate":"2019-09-27T18:11:37.710","ParentId":null,"OwnerUserId":"10886420","Title":null,"Body":"<ol>\n<li><p>Yes, provided each sample's sequence length is the same (which seems to be the case here). If not, you have to pad with <a href=\"https:\/\/pytorch.org\/docs\/stable\/nn.html#torch.nn.utils.rnn.pad_sequence\" rel=\"nofollow noreferrer\"><code>torch.nn.utils.rnn.pad_sequence<\/code><\/a> for example.<\/p><\/li>\n<li><p>Yes, LSTM is expanded to each timestep and there is output for each timestep already. Hence you don't have to apply it for each element separately. <\/p><\/li>\n<\/ol>\n"}]}
+{"QuestionId":58134455,"AnswerCount":0,"Tags":"<python-3.x><keras>","CreationDate":"2019-09-27T12:21:16.383","AcceptedAnswerId":null,"OwnerUserId":9311692.0,"Title":"How can I suppress the output image generation in detectObjectsFromImage()?","Body":"<p>I am using ImageAI's detectObjectsFromImage function to run a pre-trained keras model on certain images. It works fine but it generates an output image with enclosed boxes against objects. Ideally, I dont want the output image to be generated as I just need to detect in code and perform action on it. I tried not to give output argument but it still created a jpg file.<\/p>\n\n<p>Any ideas how to suppress the output image?<\/p>\n\n<p>code: <\/p>\n\n<pre><code>detections = detector.detectObjectsFromImage(input_image=os.path.join(\"pics\",filename))\nfor eachObject in detections:\n    if(eachObject[\"name\"]==\"person\"):\n        print(\"yes\")\n    else:\n        print(\"no\")\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58134524,"AnswerCount":1,"Tags":"<keras><conv-neural-network>","CreationDate":"2019-09-27T12:25:23.237","AcceptedAnswerId":58138561.0,"OwnerUserId":11990380.0,"Title":"CNN alternates between good performance and chance","Body":"<p>I have a binary classification problem I am trying to solve with a CNN written in Keras. The input are very sparse <code>200X125X2<\/code> tensors (can be though of as two images stacked together), and its nonzero elements are only ones (representing neuron spike trains). The input is generated using a data generator that I have built, so the model is trained using the <code>fit_generator<\/code> function. <\/p>\n\n<p>I have tried various architectures, and some show a decent performance (~88%), but the thing is that sometimes when I train new models, they don't seem to work at all, <strong>giving a chance (50%) result every epoch<\/strong>. The weird thing is that <strong>it happens sometimes to the same architectures that worked well before<\/strong>. I am running the code on Google Colab (GPU) with TensorFlow 2.0. I have check multiple times that I haven't changed anything in the code. I know that random initialization of the weights and biases may cause slight changes in the performance, but it looks like something else.<\/p>\n\n<p>Any ideas will be very helpful. Thanks!                 <\/p>\n\n<p>Here is the relevant code for one of the models that had this problem (I am using unusual kernels, I know):<\/p>\n\n<pre><code># General settings\n\nx_max = 10 \nx_size, t_size, N_features = parameters(x_max)\nbatch_size = 64\nN_epochs = 10\nN_final = 10*N_features                                    \nN_final = int(N_final - N_final%(batch_size))\nN_val = 100*batch_size\nN_test = N_final\/5\n\n# Setting up the architecture of the network and compiling\n\nmodel = Sequential()\nmodel.add(SeparableConv2D(50, (50,30), data_format='channels_first', input_shape=(2,x_size, t_size)))\nmodel.add(MaxPooling2D(pool_size=2, data_format='channels_first'))\nmodel.add(SeparableConv2D(100, (10,6), data_format='channels_first', input_shape=(2,x_size, t_size)))\nmodel.add(MaxPooling2D(pool_size=2, data_format='channels_first'))\nmodel.add(Flatten())\nmodel.add(Dense(100, activation='relu'))\nmodel.add(Dense(1, activation='sigmoid'))\n\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n\n# Fitiing the model on generated data\n\nfilepath=\"......hdf5\"\ncheckpoint = ModelCheckpoint(filepath, monitor='val_accuracy', verbose=1, save_best_only=True, mode='max')\n\nstart = time.time()\nfit_history = model_delta.fit_generator(generator = data_generator(batch_size,x_max,'delta','_',100),\n                    steps_per_epoch = N_final\/\/batch_size,\n                    validation_data = data_generator(batch_size,x_max,'delta','_',100),\n                    validation_steps = N_val\/\/batch_size,\n                    callbacks = [checkpoint],\n                    epochs = N_epochs)\nend = time.time()\n<\/code><\/pre>\n","answers":[{"AnswerId":"58138561","CreationDate":"2019-09-27T16:44:15.917","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>The most suspicious thing I see is a 'relu' near the end of the model. Depending on the initialization and on the learning rate, ReLUs can be unlucky and fall into an all-zeros case. When this happens, they completely stop gradients and don't train anymore.   <\/p>\n\n<p>By the looks of your problem (sometimes it works, sometimes it doesn't), it seems very plausible that it's the relu.<\/p>\n\n<p>So, the first suggesion (this always solves it) is to add a batch normalization before the activation:<\/p>\n\n<pre><code>model.add(Dense(100))\nmodel.add(BatchNormalization())\nmodel.add(Activation('relu'))\nmodel.add(Dense(1, activation='sigmoid')\n<\/code><\/pre>\n\n<p>Hint, if you are going to use it with the 4D tensors before the flatten, remember to use the channels dimension: <code>BatchNormalization(1)<\/code>. <\/p>\n"}]}
+{"QuestionId":58134946,"AnswerCount":1,"Tags":"<python-3.x><numpy><tensorflow><scipy><image-resizing>","CreationDate":"2019-09-27T12:51:31.680","AcceptedAnswerId":58394008.0,"OwnerUserId":6831406.0,"Title":"Resize float32 array with K-nearest neighbour in the same way as scipy.misc.imresize or tf.image.resize","Body":"<p>I am to create a network using much of the same characteristics as pix2pix: <a href=\"https:\/\/github.com\/affinelayer\/pix2pix-tensorflow\" rel=\"nofollow noreferrer\">https:\/\/github.com\/affinelayer\/pix2pix-tensorflow<\/a>.<\/p>\n\n<p>My adjustment is that I will not be using images, but matrices with <strong>float32<\/strong> values. This introduces a lot of problems and there is a lot to rewrite. Most of the code can easily be rewritten, but I've encountered a problem. <\/p>\n\n<p>The network has a separable convolutional layer where the image is resized using <code>tf.image.resize<\/code>. This function uses different resize methods, such as K-Nearest Neighbors, and I don't want to loose that feature. Both <code>scipy.misc.imresize<\/code> and <code>tf.image.resize<\/code> are limited to int values and does not support any higher than uint16. If I were to transform the data to said formats, I will loose precision.<\/p>\n\n<p><strong><em>Is there a way to create this efficiently in numpy (or any equivalent) supporting float32?<\/em><\/strong><\/p>\n\n<p>Sorry for not introducing any code, but the problem more or less explains itself without (I hope).<\/p>\n","answers":[{"AnswerId":"58394008","CreationDate":"2019-10-15T11:48:42.323","ParentId":null,"OwnerUserId":"12192748","Title":null,"Body":"<p>Try using  <strong>scipy.ndimage.interpolation.zoom<\/strong>. This works for float number images.\nUse it as below:<\/p>\n\n<p><strong>image = scipy.ndimage.interpolation.zoom(image, 0.5)<\/strong><\/p>\n"}]}
+{"QuestionId":58135790,"AnswerCount":2,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2019-09-27T13:43:05.117","AcceptedAnswerId":null,"OwnerUserId":10133689.0,"Title":"While using custom callback in Earlystopping callback not works","Body":"<p>I am working on a project where I am using custom callback with earlystopping callback, in this my model training not stops even <code>val_loss<\/code> not improving much.<\/p>\n\n<p>Here is my implmentation:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class CustomCallback(keras.callbacks.Callback):\n    def __init__(self, x, y):\n        self.x = x\n        self.y = y\n\n    def on_epoch_end(self, epoch, logs={}):\n        y_pred = self.model.predict(self.x)\n        error_rate =  np.sum(self.y == y_pred)\n        print(f'Error number:: {error_rate}')\n        logs['error_rate'] = error_rate\n\n\nearly_stop = EarlyStopping(monitor='val_loss', mode='min', verbose=1, patience=2)\ncustom_callback = CustomCallback(X_data, y_data)\n\n\nmodel.fit(train_data, y_train, epochs=100, batch_size=32, validation_data=(cv_data, y_cv), callbacks=[early_stop, custom_callback])\n<\/code><\/pre>\n\n<p>What is wrong in my implementation?<\/p>\n","answers":[{"AnswerId":"58393794","CreationDate":"2019-10-15T11:37:10.747","ParentId":null,"OwnerUserId":"12192748","Title":null,"Body":"<p>Try passing the <strong>min_delta<\/strong> argument in <strong>EarlyStopping<\/strong> with some value so that an absolute change less than min_delta will count as no improvement and it will stop training.<\/p>\n"},{"AnswerId":"58136570","CreationDate":"2019-09-27T14:29:27.423","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Why not use a custom metric instead of a callback?<\/p>\n\n<pre><code>def error_rate(y_true, y_pred):\n    rate = K.cast(K.equal(y_true, y_pred), K.floatx())\n    return keras.backend.sum(rate)\n<\/code><\/pre>\n\n<p>Are you passing label numbers or one hot tensors as y?? Usually it should be rounding first (there will be nothing equal if you don't)   <\/p>\n\n<pre><code>def error_rate(y_true, y_pred):\n    y_pred = K.cast(K.greater(y_pred, 0.5), K.floatx())\n    ate = K.cast(K.equal(y_true, y_pred), K.floatx())\n    return keras.backend.sum(rate)\n<\/code><\/pre>\n\n<p>Use it as a metric:<\/p>\n\n<pre><code>model.compile(......, metrics=[error_rate, ...])   \n<\/code><\/pre>\n"}]}
+{"QuestionId":58135906,"AnswerCount":1,"Tags":"<python><keras><classification><loss-function><imbalanced-data>","CreationDate":"2019-09-27T13:50:04.550","AcceptedAnswerId":null,"OwnerUserId":11388837.0,"Title":"Maximize a function in a binary classification problem in keras","Body":"<p>This is my first time here. I don't know if my tittle was clear. <\/p>\n\n<p>I am trying to apply deep learning (FCN) for a binary classification problem with high class imbalance <code>(0: 48887, 1:8862)<\/code>. <\/p>\n\n<p>I want to maximize the following function:  <code>(#true_positive*165 - #false_positive*203) \/ #cases to predict<\/code>. <\/p>\n\n<p>For class imbalance I tried <code>class_weight<\/code> and I think the way to attack this problem will be changing the loss function but I'm not sure about it. I will really appreciate any pointers on how I can solve this task. Thank you!<\/p>\n","answers":[{"AnswerId":"58139243","CreationDate":"2019-09-27T17:42:56.973","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Use this as loss:<\/p>\n\n<pre><code>def customLoss(true, pred):\n    true_positive = true * pred\n    false_positive = (1-true) * pred\n\n    #returning negative because Keras will minimize this instead of maximize\n    return - (true_positive - false_positive)\n<\/code><\/pre>\n\n<p>The mean value of the samples (dividing by cases to predict) is automatic later. <\/p>\n\n<hr>\n\n<p>Suggestion against imbalance: divide each term by it's expected total:<\/p>\n\n<pre><code>def balancedLoss(true, pred):\n    negatives = 1-true\n    true_positive = (true * pred) + K.epsilon()\n    false_positive = (negatives * pred) + K.epsilon()\n\n    total_positive = K.sum(true) + K.epsilon()\n    total_negative = K.sum(negatives) + K.epsilon()\n\n    return (false_positive \/ total_negative) - (true_positive \/ total_positive)   \n<\/code><\/pre>\n\n<p>This suggestion requires big batch sizes. <\/p>\n"}]}
+{"QuestionId":58136311,"AnswerCount":0,"Tags":"<keras><recurrent-neural-network>","CreationDate":"2019-09-27T14:13:48.480","AcceptedAnswerId":null,"OwnerUserId":1857860.0,"Title":"Binary Classification Model Works Decently But Confidence Scores Are Low","Body":"<p>Below I have the model\/network I've built to solve a binary classification problem<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Embedding(10_000, 50, input_length=80))\nmodel.add(GRU(32))\nmodel.add(layers.Dense(32, activation='relu'))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(optimizer='adam', loss='binary_crossentropy', metrics=['mae'])\nmodel.summary()\nhistory = model.fit(x_train, y_train, epochs=4, batch_size=16, validation_split=0.2)\n<\/code><\/pre>\n\n<p>The input data is one-hot encoded text with each sample being labeled as 1 for 'in' and 0 for 'out'.<\/p>\n\n<p>It works pretty well. When I give it an input I expect to be 'in' category it tends to give me answers greater than 0.35 and for inputs not in the 'in' category it gives me answers much lower, less than 0.1. So, it seems clear to me that my network has a pretty good idea of what's in and what's out<\/p>\n\n<p>This is my problem: I'm curious as to why this threshold point is so low? Especially given that the expected results are exactly 1. It makes me think there's some kind of layer or constraint that should be in my model that I'm forgetting that would make the answers given to be much more 'confident', ie closer to 1.<\/p>\n\n<p>Edit: It's worth noting that the highest confidence I get, which is when I give the model a sample it trained out, is ~0.55.<\/p>\n","answers":[]}
+{"QuestionId":58136592,"AnswerCount":4,"Tags":"<python><multidimensional-array><pytorch><tensor><tensor-indexing>","CreationDate":"2019-09-27T14:31:36.210","AcceptedAnswerId":58138266.0,"OwnerUserId":1150683.0,"Title":"For a given condition, get indices of values in 2D tensor A, use those to index a 3D tensor B","Body":"<p>For a given 2D tensor I want to retrieve all indices where the value is <code>1<\/code>. I expected to be able to simply use <code>torch.nonzero(a == 1).squeeze()<\/code>, which would return <code>tensor([1, 3, 2])<\/code>. However, instead, <code>torch.nonzero(a == 1)<\/code> returns a 2D tensor (that's okay), with two values per row (that's not what I expected). The returned indices should then be used to index the second dimension (index 1) of a 3D tensor, again returning a 2D tensor.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\n\na = torch.Tensor([[12, 1, 0, 0],\n                  [4, 9, 21, 1],\n                  [10, 2, 1, 0]])\n\nb = torch.rand(3, 4, 8)\n\nprint('a_size', a.size())\n# a_size torch.Size([3, 4])\nprint('b_size', b.size())\n# b_size torch.Size([3, 4, 8])\n\nidxs = torch.nonzero(a == 1)\nprint('idxs_size', idxs.size())\n# idxs_size torch.Size([3, 2])\n\nprint(b.gather(1, idxs))\n<\/code><\/pre>\n\n<p>Evidently, this does not work, leading to aRunTimeError:<\/p>\n\n<blockquote>\n  <p>RuntimeError: invalid argument 4: Index tensor must have same\n  dimensions as input tensor at\n  C:\\w\\1\\s\\windows\\pytorch\\aten\\src\\TH\/generic\/THTensorEvenMoreMath.cpp:453<\/p>\n<\/blockquote>\n\n<p>It seems that <code>idxs<\/code> is not what I expect it to be, nor can I use it the way I thought. <code>idxs<\/code> is<\/p>\n\n<pre><code>tensor([[0, 1],\n        [1, 3],\n        [2, 2]])\n<\/code><\/pre>\n\n<p>but reading through the <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.nonzero\" rel=\"nofollow noreferrer\">documentation<\/a> I don't understand why I also get back the row indices in the resulting tensor. Now, I know I can get the correct idxs by slicing <code>idxs[:, 1]<\/code> but then still, I cannot use those values as indices for the 3D tensor because the same error as before is raised. Is it possible to use the 1D tensor of indices to select items across a given dimension?<\/p>\n","answers":[{"AnswerId":"58139844","CreationDate":"2019-09-27T18:33:39.210","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>As a supplementary of @kmario23's solution, you can still achieve the same results like<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>b[torch.nonzero(a==1,as_tuple=True)]\n<\/code><\/pre>\n"},{"AnswerId":"58136829","CreationDate":"2019-09-27T14:46:39.317","ParentId":null,"OwnerUserId":"2066337","Title":null,"Body":"<pre><code>import torch\n\na = torch.Tensor([[12, 1, 0, 0],\n                  [4, 9, 21, 1],\n                  [10, 2, 1, 0]])\n\nb = torch.rand(3, 4, 8)\n\nprint('a_size', a.size())\n# a_size torch.Size([3, 4])\nprint('b_size', b.size())\n# b_size torch.Size([3, 4, 8])\n\n#idxs = torch.nonzero(a == 1, as_tuple=True)\nidxs = torch.nonzero(a == 1)\n#print('idxs_size', idxs.size())\n\nprint(torch.index_select(b,1,idxs[:,1]))\n<\/code><\/pre>\n"},{"AnswerId":"58137269","CreationDate":"2019-09-27T15:15:27.700","ParentId":null,"OwnerUserId":"1150683","Title":null,"Body":"<p>Assuming that <code>b<\/code>'s three dimensions are <code>batch_size x sequence_length x features<\/code> (b x s x feats), the expected results can be achieved as follows.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import torch\n\na = torch.Tensor([[12, 1, 0, 0],\n                  [4, 9, 21, 1],\n                  [10, 2, 1, 0]])\n\nb = torch.rand(3, 4, 8)\nprint(b.size())\n# b x s x feats\nidxs = torch.nonzero(a == 1)[:, 1]\nprint(idxs.size())\n# b\nc = b[torch.arange(b.size(0)), idxs]\nprint(c.size())\n# b x feats\n<\/code><\/pre>\n"},{"AnswerId":"58138266","CreationDate":"2019-09-27T16:21:54.740","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<p>You could simply slice them and pass it as the indices as in:<\/p>\n\n<pre><code>In [193]: idxs = torch.nonzero(a == 1)     \nIn [194]: c = b[idxs[:, 0], idxs[:, 1]]  \n\nIn [195]: c   \nOut[195]: \ntensor([[0.3411, 0.3944, 0.8108, 0.3986, 0.3917, 0.1176, 0.6252, 0.4885],\n        [0.5698, 0.3140, 0.6525, 0.7724, 0.3751, 0.3376, 0.5425, 0.1062],\n        [0.7780, 0.4572, 0.5645, 0.5759, 0.5957, 0.2750, 0.6429, 0.1029]])\n<\/code><\/pre>\n\n<hr>\n\n<p>Alternatively, an even simpler &amp; my preferred approach would be to just use <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.where\" rel=\"nofollow noreferrer\"><code>torch.where()<\/code><\/a> and then directly index into the tensor <code>b<\/code> as in:<\/p>\n\n<pre><code>In [196]: b[torch.where(a == 1)]  \nOut[196]: \ntensor([[0.3411, 0.3944, 0.8108, 0.3986, 0.3917, 0.1176, 0.6252, 0.4885],\n        [0.5698, 0.3140, 0.6525, 0.7724, 0.3751, 0.3376, 0.5425, 0.1062],\n        [0.7780, 0.4572, 0.5645, 0.5759, 0.5957, 0.2750, 0.6429, 0.1029]])\n<\/code><\/pre>\n\n<p>A bit more explanation about the above approach of using <code>torch.where()<\/code>: It works based on the concept of <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.indexing.html#advanced-indexing\" rel=\"nofollow noreferrer\">advanced indexing<\/a>. That is, when we index into the tensor using a tuple of sequence objects such as tuple of tensors, tuple of lists, tuple of tuples etc.<\/p>\n\n<pre><code># some input tensor\nIn [207]: a  \nOut[207]: \ntensor([[12.,  1.,  0.,  0.],\n        [ 4.,  9., 21.,  1.],\n        [10.,  2.,  1.,  0.]])\n<\/code><\/pre>\n\n<p>For basic slicing, we would need a tuple of integer indices:<\/p>\n\n<pre><code>   In [212]: a[(1, 2)] \n   Out[212]: tensor(21.)\n<\/code><\/pre>\n\n<p>To achieve the same using advanced indexing, we would need a tuple of sequence objects:<\/p>\n\n<pre><code># adv. indexing using a tuple of lists\nIn [213]: a[([1,], [2,])] \nOut[213]: tensor([21.])\n\n# adv. indexing using a tuple of tuples\nIn [215]: a[((1,), (2,))]  \nOut[215]: tensor([21.])\n\n# adv. indexing using a tuple of tensors\nIn [214]: a[(torch.tensor([1,]), torch.tensor([2,]))] \nOut[214]: tensor([21.])\n<\/code><\/pre>\n\n<p>And the dimension of the returned tensor would always be one dimension less than the dimension of the input tensor.<\/p>\n"}]}
+{"QuestionId":58136770,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-27T14:42:40.610","AcceptedAnswerId":58136904.0,"OwnerUserId":11409756.0,"Title":"Setting weights to model from another one with additional layers","Body":"<p>I was trying to update weights in Keras model(let it be model A). <\/p>\n\n<p>Weights are given from another model (model B- it's an extension of model A, two extra layers at the end) which is in training loop process(train_on_batch), im updating weights using method:<\/p>\n\n<pre><code>modelA.set_weights(modelB.get_weights())\n<\/code><\/pre>\n\n<p>And surprisingly it worked (the process run), even if network's architectures are different. How's that possible? Is set_weights() automatically cutting of additional part of weights? Or something is wrong and updated weight are mixed up?<\/p>\n","answers":[{"AnswerId":"58136904","CreationDate":"2019-09-27T14:51:09.803","ParentId":null,"OwnerUserId":"6537424","Title":null,"Body":"<p>It seems the <code>set_weights<\/code> method loads a list of float regarding of the layers' name. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/QfkAy.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/QfkAy.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>What you might be looking for is the <code>load_weights<\/code> method with argument <code>by_name=True<\/code> (<a href=\"https:\/\/keras.io\/getting-started\/faq\/#savingloading-only-a-models-weights\" rel=\"nofollow noreferrer\">documentation<\/a>)<\/p>\n"}]}
+{"QuestionId":58137140,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2019-09-27T15:06:48.413","AcceptedAnswerId":null,"OwnerUserId":12131277.0,"Title":"RandomizedSearchCV with Keras LSTM (Regression)","Body":"<p>I'm trying to tune hyperparameters for LSTM using RandomizedSearchCV. My code:<\/p>\n\n<pre><code>train_X, train_y = train[:, :-1], train[:, -1]\ntrain_X = train_X.reshape((train_X.shape[0], timesteps, features))\n\n\ndef create_model(neurons =50, hidden_layers=1):\n    model = Sequential()\n    model.add(LSTM(neurons, \n                   return_sequences = True, \n                   input_shape=(train_X.shape[1], train_X.shape[2])))\n    model.add(Dropout(0.2))\n    for i in range(hidden_layers):\n        model.add(LSTM(neurons))\n        model.add(Dropout(0.2))\n\n    model.add(Dense(1))\n    model.compile(loss='mean_squared_error', optimizer='adam', metrics=['acc'])\n    return model\n\nmy_regressor = KerasRegressor(build_fn=create_model, verbose=0)\n\n# Create hyperparameter space\nepochs = [100, 150]\nbatches = [5, 10]\nhidden_layers = [1, 2]\nneurons = [50, 100]\n\nhyperparameters = dict(neurons=neurons,\n                       epochs=epochs, \n                       batch_size=batches, \n                       hidden_layers = hidden_layers)\n\ngrid = RandomizedSearchCV(estimator=my_regressor, \n                          param_distributions=hyperparameters)\n\n\n# Fit grid search\n\n\ngrid_result = grid.fit(train_X, train_y)\n<\/code><\/pre>\n\n<p>This throws the error:<\/p>\n\n<blockquote>\n  <p>ValueError: Input 0 is incompatible with layer lstm_93: expected ndim=3, found ndim=2<\/p>\n<\/blockquote>\n\n<p>Any idea what I am doing wrong here? <\/p>\n","answers":[]}
+{"QuestionId":58137414,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-27T15:24:21.533","AcceptedAnswerId":null,"OwnerUserId":12131427.0,"Title":"TensorFLow: ConnectionResetError: [WinError 10054] An existing connection was forcibly closed by the remote host","Body":"<p>I am the beginner in studying tensorflow.\nI am making the code listed bellow, but facing the problem when running it, written that:<\/p>\n\n<pre><code>\"ConnectionResetError: [WinError 10054] An existing connection was forcibly closed by the remote host\"\n<\/code><\/pre>\n\n<p>Could anyone help me how to deal with that with clear explanation? because I am the beginner so I don't know much yet about tensorflow. Tnank you so much in advance. Below is my code:<\/p>\n\n<pre><code>import tensorflow as tf\nprint(tf.__version__)\nmnist = tf.keras.datasets.fashion_mnist\n(training_images, training_labels), (test_images, test_labels)= mnist.load_data()\nimport matplotlib.pyplot as plt\nplt.imshow(training_images[0])\nprint(training_labels[0])\nprint(training_images[0])\ntraining_images = training_images \/ 255.0\ntest_images = test_images \/ 255.0\nmodel = tf.keras.models.Sequential([tf.keras.layers.Flatten(),\n                                    tf.keras.layers.Dense(128, activation=tf.nn.relu),\n                                    tf.keras.layers.Dense(10, activation=tf.nn.softmax)])\nmodel.compile(optimizer=tf.train.AdamOPtimizer(),\n              loss= 'sparse_categorical_crossentropy',\n              metrics= ['accuracy'])\nmodel.fit(training_images, training_labels, epochs=5)\nmodel.evaluate(test_images, test_labels)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58137422,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-27T15:24:44.983","AcceptedAnswerId":null,"OwnerUserId":1036017.0,"Title":"Conv1D over an embedded text sentence","Body":"<p>I am looking for some clarification on the shape of input taken by the <code>Conv1D<\/code> layer. Let's say I have a model that partially looks like this.<\/p>\n\n<pre><code>tf.keras.layers.Embedding(vocab_size, embedding_dim, input_length=words_per_sentence),\n#128 filters. Window size 5 words at a time.\ntf.keras.layers.Conv1D(128, 5, activation='relu'),\n<\/code><\/pre>\n\n<p>After embedding each sentence will be a 2D matrix of shape (words_per_sentence, embedding_dim). I would have expected Conv1D to be able to work on an 1D vector only. Does it also work on a 2D matrix as is the case here? If each filter is a 1D vector then I can't even imagine how it will convolve over a 2D matrix. The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/Conv1D\" rel=\"nofollow noreferrer\">documentation<\/a> tries to touch on the subject. But it's less than clear to me.<\/p>\n","answers":[]}
+{"QuestionId":58137627,"AnswerCount":0,"Tags":"<python><installation><conda><theano>","CreationDate":"2019-09-27T15:38:46.253","AcceptedAnswerId":null,"OwnerUserId":9916514.0,"Title":"after installing packages, conda is suddenly not working","Body":"<p>I installed theano and pygpu using conda:<\/p>\n\n<pre><code>conda install theano pygpu\n<\/code><\/pre>\n\n<p>The installation seemed to finish ok, but after the installation, when I want to use conda I get the following error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/home\/arsalan\/anaconda3\/bin\/conda\", line 7, in &lt;module&gt;\n    from conda.cli import main\nModuleNotFoundError: No module named 'conda'\n<\/code><\/pre>\n\n<p>Also, I'm not able to import the packages that I had previously installed with conda. For example when I try to import pandas in python, I get the following:<\/p>\n\n<pre><code>    import pandas\n\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\nModuleNotFoundError: No module named 'pandas'\n<\/code><\/pre>\n\n<p>How can I solve this problem?<\/p>\n","answers":[]}
+{"QuestionId":58137677,"AnswerCount":1,"Tags":"<python><tensorflow><memory><keras><checkpoint>","CreationDate":"2019-09-27T15:42:28.343","AcceptedAnswerId":null,"OwnerUserId":12125447.0,"Title":"Keras model training memory leak","Body":"<p>I'm new with Keras, Tensorflow, Python and I'm trying to build a model for personal use\/future learning. I've just started with python and I came up with this code (with help of videos and tutorials). My problem is that my memory usage of Python is slowly creeping up with each epoch and even after constructing new model. Once the memory is at 100% the training just stop with no error\/warning. I don\u00b4t know too much but the issue should be somewhere within the loop (If I\u00b4m not mistaken). I know about<\/p>\n\n<blockquote>\n  <p>k.clear.session()<\/p>\n<\/blockquote>\n\n<p>but either the issue was not removed or I don\u00b4t know how to integrate it in my code. \nI have:\nPython v 3.6.4,\nTensorflow 2.0.0rc1 (cpu version),\nKeras 2.3.0<\/p>\n\n<p>This is my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import pandas as pd\nimport os\nimport time\nimport tensorflow as tf\nimport numpy as np\nimport keras\nfrom tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import Dense, Dropout, LSTM, BatchNormalization\nfrom tensorflow.keras.callbacks import TensorBoard, ModelCheckpoint\n\nEPOCHS = 25\nBATCH_SIZE = 32           \n\ndf = pd.read_csv(\"EntryData.csv\", names=['1SH5', '1SHA', '1SA5', '1SAA', '1WH5', '1WHA',\n                                         '2SA5', '2SAA', '2SH5', '2SHA', '2WA5', '2WAA',\n                                         '3R1', '3R2', '3R3', '3R4', '3R5', '3R6',\n                                         'Target'])\n\ndf_val = 14554 \n\nvalidation_df = df[df.index &gt; df_val]\ndf = df[df.index &lt;= df_val]\n\ntrain_x = df.drop(columns=['Target'])\ntrain_y = df[['Target']]\nvalidation_x = validation_df.drop(columns=['Target'])\nvalidation_y = validation_df[['Target']]\n\ntrain_x = np.asarray(train_x)\ntrain_y = np.asarray(train_y)\nvalidation_x = np.asarray(validation_x)\nvalidation_y = np.asarray(validation_y)\n\ntrain_x = train_x.reshape(train_x.shape[0], 1, train_x.shape[1])\nvalidation_x = validation_x.reshape(validation_x.shape[0], 1, validation_x.shape[1])\n\ndense_layers = [0, 1, 2]\nlayer_sizes = [32, 64, 128]\nconv_layers = [1, 2, 3]\n\nfor dense_layer in dense_layers:\n    for layer_size in layer_sizes:\n        for conv_layer in conv_layers:\n            NAME = \"{}-conv-{}-nodes-{}-dense-{}\".format(conv_layer, layer_size, \n                    dense_layer, int(time.time()))\n            tensorboard = TensorBoard(log_dir=\"logs\\{}\".format(NAME))\n            print(NAME)\n\n            model = Sequential()\n            model.add(LSTM(layer_size, input_shape=(train_x.shape[1:]), \n                                       return_sequences=True))\n            model.add(Dropout(0.2))\n            model.add(BatchNormalization())\n\n            for l in range(conv_layer-1):\n                model.add(LSTM(layer_size, return_sequences=True))\n                model.add(Dropout(0.1))\n                model.add(BatchNormalization())\n\n            for l in range(dense_layer):\n                model.add(Dense(layer_size, activation='relu'))\n                model.add(Dropout(0.2))\n\n            model.add(Dense(2, activation='softmax'))\n\n            opt = tf.keras.optimizers.Adam(lr=0.001, decay=1e-6)\n\n            # Compile model\n            model.compile(loss='sparse_categorical_crossentropy',\n                          optimizer=opt,\n                          metrics=['accuracy'])\n\n            # unique file name that will include the epoch \n            # and the validation acc for that epoch\n            filepath = \"RNN_Final.{epoch:02d}-{val_accuracy:.3f}\"  \n            checkpoint = ModelCheckpoint(\"models\\{}.model\".format(filepath, \n                         monitor='val_acc', verbose=0, save_best_only=True, \n                         mode='max')) # saves only the best ones\n\n            # Train model\n            history = model.fit(\n                train_x, train_y,\n                batch_size=BATCH_SIZE,\n                epochs=EPOCHS,\n                validation_data=(validation_x, validation_y),\n                callbacks=[tensorboard, checkpoint])\n\n# Score model\nscore = model.evaluate(validation_x, validation_y, verbose=2)\nprint('Test loss:', score[0])\nprint('Test accuracy:', score[1])\n# Save model\nmodel.save(\"models\\{}\".format(NAME))\n<\/code><\/pre>\n\n<p>Also I don\u00b4t know If it\u00b4s possible to ask 2 problems within 1 question (I don\u00b4t want to spam it here with my problems which anyone with any python experience can resolve within a minute), but I also have problem with checkpoint saving. I want to save only the best performing model (1 model per 1 NN specification - number of nodes\/layers) but currently it is saved after every epoch. If this is inappropriate to ask I can create another question for this.<\/p>\n\n<p>Thank you very much for any help.<\/p>\n","answers":[{"AnswerId":"58138230","CreationDate":"2019-09-27T16:18:15.343","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>One source of the problem is, a new loop of <code>model = Sequential()<\/code> does <em>not<\/em> remove the previous model; it remains built within its TensorFlow graph scope, and every new <code>model = Sequential()<\/code> adds another lingering construction which eventually overflows memory. To ensure a model is properly destroyed in full, run below once you're done with a model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import gc\ndel model\ngc.collect()\nK.clear_session()\ntf.compat.v1.reset_default_graph() # TF graph isn't same as Keras graph\n<\/code><\/pre>\n\n<p><code>gc<\/code> is Python's garbage collection module, which clears remnant traces of <code>model<\/code> after <code>del<\/code>. <code>K.clear_session()<\/code> is the main call, and clears the TensorFlow graph. <\/p>\n\n<p>Also, while your idea for model checkpointing, logging, and hyperparameter search is quite sound, it's quite faultily executed; you will actually be testing only <em>one<\/em> hyperparameter combination for the entire nested loop you've set up there. But this should be asked in a separate question.<\/p>\n\n<hr>\n\n<p><strong>UPDATE<\/strong>: just encountered the same problem, on a fully properly setup environment; the likeliest conclusion is, it's a bug - and a definite culprit is <em>Eager execution<\/em>. To work around, use<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tf.compat.v1.disable_eager_execution() # right after `import tensorflow as tf`\n<\/code><\/pre>\n\n<p>to switch to Graph mode, which can also run <a href=\"https:\/\/stackoverflow.com\/questions\/58441514\/why-is-tensorflow-2-much-slower-than-tensorflow-1\/58653632#58653632\">significantly faster<\/a>. Also see updated clear code above.<\/p>\n"}]}
+{"QuestionId":58137816,"AnswerCount":1,"Tags":"<docker><tensorflow>","CreationDate":"2019-09-27T15:51:04.383","AcceptedAnswerId":58167401.0,"OwnerUserId":12131586.0,"Title":"using tensorflow inside docker with gpu","Body":"<p>I'm trying to use tensorflow in docker, following the instructions at<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/install\/docker\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/docker<\/a><\/p>\n\n<p>I have docker 19.03 installed on a centos 7.7.1908 machine, and I can run nvidia-smi<\/p>\n\n<pre><code>docker run --gpus all nvidia\/cuda:10.1-base nvidia-smi\n<\/code><\/pre>\n\n<p>my understanding is that the use of nvidia-docker is deprecated. then I download tensorflow<\/p>\n\n<pre><code>docker pull tensorflow\/tensorflow:latest-devel-py3\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>docker pull tensorflow\/tensorflow:devel-gpu\n<\/code><\/pre>\n\n<p>but when I run one of them<\/p>\n\n<pre><code>docker run -it --rm tensorflow\/tensorflow:latest-devel-py3 python -c \"import tensorflow as tf;\"\n<\/code><\/pre>\n\n<p>I get<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File \"\", line 1, in\n   ModuleNotFoundError: No module named 'tensorflow' failed to\n  resize tty, using default size<\/p>\n<\/blockquote>\n\n<p>what am I missing?<\/p>\n\n<p>well, I've since found out that<\/p>\n\n<pre><code>latest-gpu-py3\n<\/code><\/pre>\n\n<p>allows me to import tensorflow, whereas<\/p>\n\n<pre><code>latest-devel-py3\n<\/code><\/pre>\n\n<p>doesn't. I'd've thought that all these images would work...<\/p>\n","answers":[{"AnswerId":"58167401","CreationDate":"2019-09-30T11:53:33.457","ParentId":null,"OwnerUserId":"9218240","Title":null,"Body":"<p>\"devel\" image contains tensorflow sources, see: <a href=\"https:\/\/hub.docker.com\/r\/tensorflow\/tensorflow\/\" rel=\"nofollow noreferrer\">https:\/\/hub.docker.com\/r\/tensorflow\/tensorflow\/<\/a><\/p>\n\n<p>Quote: <\/p>\n\n<blockquote>\n  <p>devel images come with Bazel and are ideal for developing changes to\n  TensorFlow at master. \/tensorflow_src includes the TensorFlow source\n  tree at the latest nightly commit where the Pip package built\n  successfully in the container. We no longer provide images for\n  developing on top of older versions of TF (1.12.0 was the last release\n  where this was the case).<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":58137990,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning>","CreationDate":"2019-09-27T16:01:51.840","AcceptedAnswerId":58138160.0,"OwnerUserId":9734248.0,"Title":"Best practices in Tensorflow 2.0(Training step)","Body":"<p>In tensorflow 2.0 you don't have to worry about training phase(batch size, number of epochs etc), because everything can be defined in <code>compile<\/code> method: <code>model.fit(X_train,Y_train,batch_size = 64,epochs = 100)<\/code>.<\/p>\n\n<p>But I have seen the following code style:<\/p>\n\n<pre><code>optimizer = tf.keras.optimizers.Adam(0.001)\nloss_fn = tf.keras.losses.SparseCategoricalCrossentropy()\n\n@tf.function\ndef train_step(inputs, labels):\n  with tf.GradientTape() as tape:\n    predictions = model(inputs, training=True)\n    regularization_loss = tf.math.add_n(model.losses)\n    pred_loss = loss_fn(labels, predictions)\n    total_loss = pred_loss + regularization_loss\n\n  gradients = tape.gradient(total_loss, model.trainable_variables)\n  optimizer.apply_gradients(zip(gradients, model.trainable_variables))\n\nfor epoch in range(NUM_EPOCHS):\n  for inputs, labels in train_data:\n    train_step(inputs, labels)\n  print(\"Finished epoch\", epoch)\n<\/code><\/pre>\n\n<p>So here you can observe \"more detailed\" code, where you manually define by for loops you training procedure.<\/p>\n\n<p>I have following question: what is the best practice in Tensorflow 2.0? I haven't found a any complete tutorial.<\/p>\n","answers":[{"AnswerId":"58138160","CreationDate":"2019-09-27T16:13:39.727","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Use what is best for your needs.<\/p>\n\n<p>Both methods are documented in Tensorflow tutorials.    <\/p>\n\n<p>If you don't need anything special, no extra losses, strange metrics or intricate gradient computation, just use a <code>model.fit()<\/code> or a <code>model.fit_generator()<\/code>. This is totally ok and makes your life easier. <\/p>\n\n<p>A custom training loop might come in handy when you have complicated models with non-trivial loss\/gradients calculation. <\/p>\n\n<p>Up to now, two applications I tried were easier with this:<\/p>\n\n<ul>\n<li>Training a GAN's generator and discriminator simultaneously without having to do the generation step twice. (It's complicated because you have a loss function that applies to different <code>y_true<\/code> values, and each case should update only a part of the model) - The other option would require to have a few separate models, each model with its own <code>trainable=True\/False<\/code> configuration, and train then in separate phases.   <\/li>\n<li>Training inputs (good for style transfer models) -- Alternatively, create a custom layer that takes dummy inputs and that outputs its own trainable weights. But it gets complicated to compile several loss functions for each of the outputs of the base and style networks. <\/li>\n<\/ul>\n"}]}
+{"QuestionId":58138071,"AnswerCount":2,"Tags":"<python><tensorflow><keras><version>","CreationDate":"2019-09-27T16:07:20.803","AcceptedAnswerId":null,"OwnerUserId":12131632.0,"Title":"'tuple' object has no attribute 'layer'","Body":"<p>I am having many troubles trying to start training my model (a DCGAN). It is sending me the error : \"'tuple' object has no attribute 'layer'\". I read this could be due to having both the tensorflow version 1.14.0 and the keras version 2.2 or superior. \nI tried to fix this by downgrading the keras version to the 2.1.5, but is is sending me the same problem.\nPlease help!\nThanks<\/p>\n\n<p>I tried to fix this by downgrading the keras version to the 2.1.5, but is is sending me the same problem.<\/p>\n\n<pre><code>from google.colab import drive \ndrive.mount('\/mntDrive')\n\nimport os,sys      #os es para gestionar directorios (trabajar con archivos); sys es para trabajar con variables del sistema\n\n# numpy\nimport numpy as np\n\n# data processing, CSV file I\/O (e.g. pd.read_csv)\nimport pandas as pd\nfrom sklearn.model_selection import train_test_split\n\n# Charts\nimport matplotlib.pyplot as plt\nfrom matplotlib.pyplot import imread\n#process\nfrom tqdm import tqdm\n\n# Image IO\nfrom PIL import Image\nimport skimage.io\nimport skimage.transform\nfrom skimage.transform import resize\n\n# Deep learning\nimport tensorflow as tf\n\nfrom tensorflow.keras import layers\nfrom keras import optimizers\nfrom keras.models import Sequential, Model\nfrom keras.layers.advanced_activations import LeakyReLU\nfrom keras.layers import Dense, Conv2D, Flatten, MaxPool2D, Dropout, Input\nfrom keras.callbacks import ModelCheckpoint\nfrom keras.preprocessing.image import ImageDataGenerator\nfrom keras.optimizers import Adam\nfrom keras import initializers\n\nfrom __future__ import absolute_import, division, print_function, unicode_literals\n\n\n# To make sure that we can reproduce the experiment and get the same results\nnp.random.seed(21)\n\n\n# The dimension of our random noise vector.\nrandom_dim = 100\nfrom skimage.color import rgb2gray\nimport scipy.ndimage\nimport scipy.misc\nimport re\nimages = []\nfor root, dirnames, filenames in os.walk(\"\/mntDrive\/My Drive\/Colab Notebooks\/cubism\"):\n    for filename in filenames:\n        if re.search(\"\\.(jpg|jpeg|png)$\", filename):\n            filepath = os.path.join(root, filename)\n            image = plt.imread(filepath, )\n            image = (image.astype(np.float32) - 127.5)\/127.5\n            image_resized = resize(image, (112, 112))\n            images.append(image_resized)\nimages = np.array(images)\n\n\nprint('Original image shape: {}'.format(images.shape))\nim_gray = rgb2gray(images)\nprint('New image shape: {}'.format(im_gray.shape))\nimages_resized = im_gray.reshape(320,12544)\n\ndef get_optimizer():\n  optimizer=tf.keras.optimizers.Adam(0.001)\n  return optimizer\n\n\ndef make_generator_model(optimizer):\n    generator = tf.keras.Sequential()\n    generator.add(layers.Dense(7*7*256, use_bias=False, input_shape=(random_dim,)))\n    generator.add(layers.BatchNormalization())\n    generator.add(layers.LeakyReLU())\n\n    generator.add(layers.Reshape((7, 7, 256)))\n    assert generator.output_shape == (None, 7, 7, 256) # Note: None is the batch size\n\n    generator.add(layers.Conv2DTranspose(128, (5, 5), strides=(2, 2), padding='same', use_bias=False))\n    assert generator.output_shape == (None, 14, 14, 128)\n    generator.add(layers.BatchNormalization())\n    generator.add(layers.LeakyReLU())\n\n    generator.add(layers.Conv2DTranspose(64, (5, 5), strides=(2, 2), padding='same', use_bias=False))\n    assert generator.output_shape == (None, 28, 28, 64)\n    generator.add(layers.BatchNormalization())\n    generator.add(layers.LeakyReLU())\n\n    generator.add(layers.Conv2DTranspose(32, (5, 5), strides=(2, 2), padding='same', use_bias=False))\n    assert generator.output_shape == (None, 56, 56, 32)\n    generator.add(layers.BatchNormalization())\n    generator.add(layers.LeakyReLU())\n\n    generator.add(layers.Conv2DTranspose(1, (5, 5), strides=(2, 2), padding='same', use_bias=False, activation='tanh'))\n    assert generator.output_shape == (None, 112, 112, 1)\n    generator.compile(loss='binary_crossentropy', optimizer=tf.keras.optimizers.Adam(0.001))\n\n    return generator    \n\ndef make_discriminator_model(optimizer):\n    discriminator = tf.keras.Sequential()\n    discriminator.add(layers.Conv2D(64, (5, 5), strides=(2, 2), padding='same', input_shape=[112, 112, 1]))\n    discriminator.add(layers.LeakyReLU())\n    discriminator.add(layers.Dropout(0.3))\n\n    discriminator.add(layers.Conv2D(128, (5, 5), strides=(2, 2), padding='same'))\n    discriminator.add(layers.LeakyReLU())\n    discriminator.add(layers.Dropout(0.3))\n\n    discriminator.add(layers.Conv2D(256, (5, 5), strides=(2, 2), padding='same'))   \n    discriminator.add(layers.LeakyReLU())\n    discriminator.add(layers.Dropout(0.3))\n\n    discriminator.add(layers.Flatten())\n    discriminator.add(layers.Dense(1, activation='sigmoid' ))\n    discriminator.compile(loss='binary_crossentropy', optimizer=tf.keras.optimizers.Adam(0.001))    \n\n    return discriminator\n\ndef get_gan_network(discriminator, random_dim, generator, optimizer):\n    # We initially set trainable to False since we only want to train either the\n    # generator or discriminator at a time\n    discriminator.trainable = False\n    # gan input (noise) will be 100-dimensional vectors\n    gan_input = Input(shape=(random_dim,))\n    # the output of the generator (an image)\n    x = generator(gan_input)\n    # get the output of the discriminator (probability if the image is real or not)\n    gan_output = discriminator(x)\n    gan = Model(inputs=gan_input, outputs=gan_output)\n    gan.compile(loss='binary_crossentropy', optimizer=tf.keras.optimizers.Adam(0.001))\n    return gan\n\ndef plot_generated_images(epoch, generator, examples=64, dim=(10, 10), figsize=(100, 100)):\n    noise = np.random.normal(0, 1, size=[examples, random_dim]) #mean, std deviation, size\n    generated_images = generator.predict(noise)\n    generated_images = generated_images.reshape(examples, 112, 112)\n\n    plt.figure(figsize=figsize)\n    for i in range(generated_images.shape[0]):\n        plt.subplot(dim[0], dim[1], i+1)\n        plt.imshow(generated_images[i], interpolation='nearest', cmap='gray_r')\n        plt.axis('off')\n    plt.tight_layout()\n    plt.savefig('dcgan_generated_Originals_2_epoch_%d.png' % epoch)\n\n\n#you can create a function which will save your generated images every 20 epochs\n    # Create a wall of generated MNIST images\ndef train(epochs=15000, batch_size=80):\n    # Get the training and testing data\n    images_resized\n    # Split the training data into batches of size 80\n    batch_count = images_resized.shape[0] \/\/ batch_size \n\n    # Build our GAN network\n    optimizer=get_optimizer()\n    generator = make_generator_model(optimizer)\n    discriminator = make_discriminator_model(optimizer)\n    gan = get_gan_network(discriminator, random_dim, generator, optimizer)\n\n    for e in range(1, epochs+1):\n        print ('-'*15, 'Epoch %d' % e, '-'*15)\n        for _ in tqdm(range(batch_count)):\n            # Get a random set of input noise and images\n            noise = np.random.normal(0, 1, size=[batch_size, random_dim])\n            image_batch = images_resized[np.random.randint(0, images_resized.shape[0], size=batch_size)]\n\n            # Generate fake MNIST images\n            generated_images = generator.predict(noise)\n            X = np.concatenate([image_batch, generated_images])  #128x4096 --- 'a' x 4096\n\n            # Labels for generated and real data\n            y_dis = np.zeros(2*batch_size)\n            y_dis[:batch_size] = 0.9    # One-sided label smoothing. Se refiere a que la mitad de y_dis seran 0.9 (reales) y la otra mitad seran 0 (falsos)\n\n\n            # Train discriminator\n            discriminator.trainable = True\n            discriminator.train_on_batch(X, y_dis)      #Se entrena el Discriminador con dos vectores: X(tiene un batch(128) imagenes reales y un batch de imagenes generadas); y_dis(tiene un batch de 128 0.9(serian las reales) y otro batch de 128 ceros(serian las generadas). De esta manera, el algoritmo generador, aunque las genere bien, va a seguir optimizandose el numero de epochs que haga falta(seguimos en la fase de entrenamiento), ya que sus imagenes son continuamente rechazadas, comparandose con las reales y sacando continuamente diferencias. ) \n                                                        #Entra X(imagenes reales, imagenes generadas)---&gt;y_dis(0.9 , 0)\n                                                        #train_on_batch(x,y)---&gt; realiza una actualizacion del gradiente en un batch----&gt; x=array de training data (si el modelo tiene varias entradas pueden ser varios); y= array de target data (si el modelo tiene varias salidas pueden ser varios)\n\n            # Train generator\n            noise = np.random.normal(0, 1, size=[batch_size, random_dim])\n            y_gen = np.ones(batch_size)\n            discriminator.trainable = False\n            gan.train_on_batch(noise, y_gen)            #Entra noise(pixeles desordenados)----&gt;y_gen(todo son 1)\n\n        if e == 1 or e % 50 == 0:\n            plot_generated_images(e, generator)\n\n\nif __name__ == '__main__':\n   train(30000, 80)\n<\/code><\/pre>\n\n<p>This is the output:<\/p>\n\n<pre><code>Drive already mounted at \/mntDrive; to attempt to forcibly remount, call drive.mount(\"\/mntDrive\", force_remount=True).\nOriginal image shape: (320, 112, 112, 3)\nNew image shape: (320, 112, 112)\nWARNING:tensorflow:From \/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/init_ops.py:1251: calling VarianceScaling.__init__ (from tensorflow.python.ops.init_ops) with dtype is deprecated and will be removed in a future version.\nInstructions for updating:\nCall initializer instance with the dtype argument instead of passing it to the constructor\nWARNING:tensorflow:From \/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/ops\/nn_impl.py:180: add_dispatch_support.&lt;locals&gt;.wrapper (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse tf.where in 2.0, which has the same broadcast rule as np.where\nWARNING:tensorflow:From \/usr\/local\/lib\/python3.6\/dist-packages\/keras\/backend\/tensorflow_backend.py:68: The name tf.get_default_graph is deprecated. Please use tf.compat.v1.get_default_graph instead.\n\nWARNING:tensorflow:From \/usr\/local\/lib\/python3.6\/dist-packages\/keras\/backend\/tensorflow_backend.py:507: The name tf.placeholder is deprecated. Please use tf.compat.v1.placeholder instead.\n\n    ---------------------------------------------------------------------------\n    AttributeError                            Traceback (most recent call last)\n    &lt;ipython-input-18-6e3c9ece87ff&gt; in &lt;module&gt;()\n        196 \n        197 if __name__ == '__main__':\n    --&gt; 198    train(30000, 80)\n\n    11 frames\n    &lt;ipython-input-18-6e3c9ece87ff&gt; in train(epochs, batch_size)\n        161     generator = make_generator_model(optimizer)\n        162     discriminator = make_discriminator_model(optimizer)\n    --&gt; 163     gan = get_gan_network(discriminator, random_dim, generator, optimizer)\n        164 \n        165     for e in range(1, epochs+1):\n\n    &lt;ipython-input-18-6e3c9ece87ff&gt; in get_gan_network(discriminator, random_dim, generator, optimizer)\n        128     gan_input = Input(shape=(random_dim,))\n        129     # the output of the generator (an image)\n    --&gt; 130     x = generator(gan_input)\n        131     # get the output of the discriminator (probability if the image is real or not)\n        132     gan_output = discriminator(x)\n\n    \/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in __call__(self, inputs, *args, **kwargs)\n        632                     outputs = base_layer_utils.mark_as_return(outputs, acd)\n        633                 else:\n    --&gt; 634                   outputs = call_fn(inputs, *args, **kwargs)\n        635 \n        636             except TypeError as e:\n\n    \/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/sequential.py in call(self, inputs, training, mask)\n        245       if not self.built:\n        246         self._init_graph_network(self.inputs, self.outputs, name=self.name)\n    --&gt; 247       return super(Sequential, self).call(inputs, training=training, mask=mask)\n        248 \n        249     outputs = inputs  # handle the corner case where self.layers is empty\n\n    \/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/network.py in call(self, inputs, training, mask)\n        749                                 ' implement a `call` method.')\n        750 \n    --&gt; 751     return self._run_internal_graph(inputs, training=training, mask=mask)\n        752 \n        753   def compute_output_shape(self, input_shape):\n\n    \/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/network.py in _run_internal_graph(self, inputs, training, mask)\n        891 \n        892           # Compute outputs.\n    --&gt; 893           output_tensors = layer(computed_tensors, **kwargs)\n        894 \n        895           # Update tensor_dict.\n\n    \/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in __call__(self, inputs, *args, **kwargs)\n        661               kwargs.pop('training')\n        662             inputs, outputs = self._set_connectivity_metadata_(\n    --&gt; 663                 inputs, outputs, args, kwargs)\n        664           self._handle_activity_regularization(inputs, outputs)\n        665           self._set_mask_metadata(inputs, outputs, previous_mask)\n\n    \/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in _set_connectivity_metadata_(self, inputs, outputs, args, kwargs)\n       1706     kwargs.pop('mask', None)  # `mask` should not be serialized.\n       1707     self._add_inbound_node(\n    -&gt; 1708         input_tensors=inputs, output_tensors=outputs, arguments=kwargs)\n       1709     return inputs, outputs\n       1710 \n\n    \/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in _add_inbound_node(self, input_tensors, output_tensors, arguments)\n       1793     \"\"\"\n       1794     inbound_layers = nest.map_structure(lambda t: t._keras_history.layer,\n    -&gt; 1795                                         input_tensors)\n       1796     node_indices = nest.map_structure(lambda t: t._keras_history.node_index,\n       1797                                       input_tensors)\n\n    \/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/util\/nest.py in map_structure(func, *structure, **kwargs)\n        513 \n        514   return pack_sequence_as(\n    --&gt; 515       structure[0], [func(*x) for x in entries],\n        516       expand_composites=expand_composites)\n        517 \n\n    \/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/util\/nest.py in &lt;listcomp&gt;(.0)\n        513 \n        514   return pack_sequence_as(\n    --&gt; 515       structure[0], [func(*x) for x in entries],\n        516       expand_composites=expand_composites)\n        517 \n\n    \/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in &lt;lambda&gt;(t)\n       1792             `call` method of the layer at the call that created the node.\n       1793     \"\"\"\n    -&gt; 1794     inbound_layers = nest.map_structure(lambda t: t._keras_history.layer,\n       1795                                         input_tensors)\n       1796     node_indices = nest.map_structure(lambda t: t._keras_history.node_index,\n    AttributeError: 'tuple' object has no attribute 'layer'\n    AttributeError: 'tuple' object has no attribute 'layer'\n<\/code><\/pre>\n","answers":[{"AnswerId":"58392672","CreationDate":"2019-10-15T10:35:12.913","ParentId":null,"OwnerUserId":"12192748","Title":null,"Body":"<p>You have imported layers from <strong>tensorflow.keras<\/strong> while other functions you have imported from <strong>keras<\/strong>. You can either import your layers from keras or try importing other functions from tensorflow.keras which might work.<\/p>\n"},{"AnswerId":"58157329","CreationDate":"2019-09-29T16:55:03.943","ParentId":null,"OwnerUserId":"10938918","Title":null,"Body":"<p>I am only new in Python and deep learning. But I wanted to share this, as it helped me with a similar error. \nMaybe, in the importing section, try:<\/p>\n\n<pre><code>import tensorflow.keras\n<\/code><\/pre>\n\n<p>For all the rest you have to import from keras, keep the <strong>tensorflow.<\/strong> in front. \nAnd type it as a whole when using <strong>anything<\/strong> from keras later. Eg:<\/p>\n\n<pre><code>model.add(tensorflow.keras.layers.Conv2D())\n<\/code><\/pre>\n\n<p>Then you will not need: import tensorflow as tf<\/p>\n\n<p>The <em>tuple<\/em> probably refers to your <em>input_shape<\/em>. What I mention above solved my problem.<\/p>\n\n<p>I cannot answer why exactly this happens, though.<\/p>\n"}]}
+{"QuestionId":58138099,"AnswerCount":1,"Tags":"<python><tensorflow><keras><softmax>","CreationDate":"2019-09-27T16:09:44.183","AcceptedAnswerId":null,"OwnerUserId":11417079.0,"Title":"softmax score by a cnn model","Body":"<p>I have build a CNN model for image classification. I want to pick five images that are correctly classified by the network and have the maximum softmax scores(for each class).<\/p>\n\n<p>I have tried to check the model.evaluate(X_test,y_test) but it gives the overall softmax score of model.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential()\nmodel.add(Conv2D(96, (7, 7),  kernel_initializer='he_uniform', padding='same',strides=2, input_shape=(32, 32, 3)))\nmodel.add(Activation(\"relu\"))\nmodel.add(Conv2D(64, (5, 5), kernel_initializer='he_uniform', padding='same',strides=2))\nmodel.add(Activation(\"relu\"))\n\nmodel.add(Conv2D(128, (3, 3), kernel_initializer='he_uniform', padding='same',strides=2))\nmodel.add(Activation(\"relu\"))\n\nmodel.add(MaxPooling2D(pool_size=(3, 3),padding='same',strides=3))\n\nmodel.add(Flatten())\n\nmodel.add(Dense(128))\nmodel.add(Activation(\"relu\"))\n\nmodel.add(Dense(10, activation='softmax'))\n<\/code><\/pre>\n\n<h1>compile model<\/h1>\n\n<pre class=\"lang-py prettyprint-override\"><code>opt = SGD(lr=0.0001, momentum=0.9)\nmodel.compile(optimizer=opt, loss='categorical_crossentropy', metrics['accuracy'])\n<\/code><\/pre>\n\n<p>I want to see the softmax scores of correctly classified images.<\/p>\n","answers":[{"AnswerId":"58142667","CreationDate":"2019-09-28T01:17:56.313","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>For this you have to use <code>model.predict<\/code>, not <code>model.evaluate<\/code>:<\/p>\n\n<pre><code>model.predict(X_test)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58138108,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2019-09-27T16:10:20.607","AcceptedAnswerId":58139305.0,"OwnerUserId":12131650.0,"Title":"Multiply the outputs custom loss function","Body":"<p>So my question is, if I have something like:<\/p>\n\n<pre><code>model = Model(inputs = input, outputs = [y1,y2])\n\nmodel.compile(loss = my_loss ...) \n<\/code><\/pre>\n\n<p>I have only seen <code>my_loss<\/code> as a dictionary of independent losses and, then, the final loss is defined as the sum of those. But, can I define in a multitask model a loss function that take all the predicted\/true values and then I can multiply them (for instance)?<\/p>\n\n<p>This is the loss I am trying to define:<\/p>\n\n<pre><code>def my_loss(y_true1, y_true2, y_pred1, y_pred2):\n   final_loss = binary_crossentropy(y_true1, y_pred1) + y_true1 * categorical_crossentropy(y_true2, y_pred2)\n   return final_loss\n<\/code><\/pre>\n\n<p>Usually, your paramaters are <code>y_true, y_pred<\/code> in the loss function, where <code>y_pred<\/code> is either <code>y1<\/code> or <code>y2<\/code>. But now I need both to compute the loss, so how can I define this loss function and pass all the parameters to the function: <code>y_true1, y_true2, y_pred1, y_pred2<\/code>.<\/p>\n\n<p>My current model that I want to change its loss:<\/p>\n\n<pre><code>x = Input(shape=(n, ))\nshared = Dense(32)(x)\nsub1 = Dense(16)(shared)\nsub2 = Dense(16)(shared)\ny1 = Dense(1)(sub1, activation='sigmoid')\ny2 = Dense(4)(sub2, activation='softmax')\n\nmodel = Model(inputs = input, outputs = [y1,y2])\n\nmodel.compile(loss = ['binary_crossentropy', 'categorical_crossentropy'] ...) #THIS LINE I WANT TO CHANGE IT\n<\/code><\/pre>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"58139305","CreationDate":"2019-09-27T17:48:21.420","ParentId":null,"OwnerUserId":"4531017","Title":null,"Body":"<p>I'm not sure if I'm understanding correctly, but I'll try.<\/p>\n\n<p>The loss function must contain both the predicted and the actual data -- it's a way to measure the error between what your model is predicting and the true data.  However, the predicted and actual data do not need to be one-dimensional. You can make <code>y_pred<\/code> a tensor that contains both <code>y_pred1<\/code> and <code>y_pred2<\/code>.  Likewise, <code>y_true<\/code> can be a tensor that contains both <code>y_true1<\/code> and <code>y_true2<\/code>. <\/p>\n\n<p>As far as I know, loss functions should return a single number. That's why loss functions often have a mean or a sum to add up all of the losses for individual data points.<\/p>\n\n<p>Here's an example of mean square error that will work for more than 1D:<\/p>\n\n<pre><code>import keras.backend as K\n\ndef my_loss(y_true, y_pred):\n    # this example is mean squared error\n    # works if if y_pred and y_true are greater than 1D\n    return K.mean(K.square(y_pred - y_true))\n<\/code><\/pre>\n\n<p>Here's another example of a loss function that I think is closer to your question (although I cannot comment on whether or not it's a good loss function):<\/p>\n\n<pre><code>def my_loss(y_true, y_pred):\n    # calculate mean(abs(y_pred1*y_pred2 - y_true1*ytrue2)) \n    # this will work for 2D inputs of y_pred and y_true\n    return K.mean(K.abs(K.prod(y_pred, axis = 1) - K.prod(y_true, axis = 1)))\n<\/code><\/pre>\n\n<p><strong>Update:<\/strong> <\/p>\n\n<p>You can concatenate two outputs into a single tensor with <code>keras.layers.Concatenate<\/code>.  That way you can still have a loss function with only two arguments. <\/p>\n\n<p>In the model you wrote above, the <code>y1<\/code> output shape is <code>(None, 1)<\/code> and the <code>y2<\/code> output shape is <code>(None, 4)<\/code>.  Here's an example of how you could write your model so that the output is a single tensor that concatenates <code>y1<\/code> and <code>y1<\/code> into a shape of <code>(None, 5)<\/code>: <\/p>\n\n<pre><code>from keras import Model\nfrom keras.layers import Input, Dense\nfrom keras.layers import Concatenate\n\ninput_layer = Input(shape=(n, ))\nshared = Dense(32)(input_layer)\nsub1 = Dense(16)(shared)\nsub2 = Dense(16)(shared)\ny1 = Dense(1, activation='sigmoid')(sub1)\ny2 = Dense(4, activation='softmax')(sub2)\nmergedOutput = Concatenate()([y1, y2])\n<\/code><\/pre>\n\n<p>Below, I show an example for how you could rewrite your loss function. I wasn't sure which of the 5 columns of the output to call <code>y_true1<\/code> vs. <code>y_true2<\/code>, so I guessed that <code>y_true1<\/code> was column 1 and <code>y_true2<\/code> was the remaining 4 columns. The same column structure would apply to <code>y_pred1<\/code> and <code>y_pred2<\/code>. <\/p>\n\n<pre><code>from keras import losses\n\ndef my_loss(y_true, y_pred):\n    final_loss = (losses.binary_crossentropy(y_true[:, 0], y_pred[:, 0]) + \n                    y_true[:, 0] * \n                  losses.categorical_crossentropy(y_true[:, 1:], y_pred[:,1:]))\n    return final_loss\n<\/code><\/pre>\n\n<p>Finally, you can compile the model without any major changes from normal: <\/p>\n\n<pre><code>model.compile(optimizer='adam', loss=my_loss)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58138274,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-27T16:22:30.143","AcceptedAnswerId":null,"OwnerUserId":6264041.0,"Title":"TensorFlow prints log information twice","Body":"<p>I'm running TensorFlow official model from <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/official\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/official<\/a><\/p>\n\n<p>When I run the official resnet model, each tensorflow INFO message is printed twice as displayed below:<\/p>\n\n<pre><code>INFO:tensorflow:cross_entropy = 2.053623, learning_rate = 0.025, train_accuracy = 0.19732143 (0.102 sec)\nI0927 11:58:25.770764 140597122463552 basic_session_run_hooks.py:247] cross_entropy = 2.053623, learning_rate = 0.025, train_accuracy = 0.19732143 (0.102 sec)\nINFO:tensorflow:cross_entropy = 1.7440678, learning_rate = 0.025, train_accuracy = 0.19938381 (0.107 sec)\nI0927 11:58:25.878196 140597122463552 basic_session_run_hooks.py:247] cross_entropy = 1.7440678, learning_rate = 0.025, train_accuracy = 0.19938381 (0.107 sec)                                                                                   \nINFO:tensorflow:cross_entropy = 2.116714, learning_rate = 0.025, train_accuracy = 0.1983507 (0.102 sec)\nI0927 11:58:25.980251 140597122463552 basic_session_run_hooks.py:247] cross_entropy = 2.116714, learning_rate = 0.025, train_accuracy = 0.1983507 (0.102 sec)\n<\/code><\/pre>\n\n<p>My current log level is<\/p>\n\n<p><code>export TF_CPP_MIN_LOG_LEVEL=2<\/code><\/p>\n\n<p>Can someone help with the issue? Thanks!<\/p>\n","answers":[]}
+{"QuestionId":58138286,"AnswerCount":0,"Tags":"<tensorflow><keras><conv-neural-network><recurrent-neural-network>","CreationDate":"2019-09-27T16:23:27.563","AcceptedAnswerId":null,"OwnerUserId":9047382.0,"Title":"How to build a pretrained CNN-LSTM network with Keras","Body":"<p>I'm trying to use a CNN-LSTM network with Keras in order to analyze videos. I read about it and run into <code>TimeDistributed<\/code> function and some examples.<\/p>\n\n<p>Actually, I tried the network described below, which is in fact composed by a convolutional and pooling layers followed by recurrent and dense layers.<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(TimeDistributed(Conv2D(2, (2,2), activation= 'relu' ), input_shape=(None, IMG_SIZE, IMG_SIZE, 3)))\nmodel.add(TimeDistributed(MaxPooling2D(pool_size=(2, 2))))\nmodel.add(TimeDistributed(Flatten()))\nmodel.add(LSTM(50))\nmodel.add(Dense(50, activation = 'softmax'))\nmodel.compile(loss = 'categorical_crossentropy' , optimizer = 'adam' , metrics = ['acc'])\n<\/code><\/pre>\n\n<p>I haven't tested properly the model, since my dataset is too small. However, during training process the network reaches <strong>accuracy 0.98<\/strong> in 4-5 epochs (perhaps it is overfitting, but it isn't a problem yet because I hope to get a suitable dataset later).<\/p>\n\n<p>Then, I read about how to use a pretrained convolutional network (MobileNet, ResNet or Inception) as a feature extractor for LSTM network, such that I use the following code:<\/p>\n\n<pre><code>inputs = Input(shape = (frames, IMG_SIZE, IMG_SIZE, 3))\ncnn_base = InceptionV3(include_top = False, weights='imagenet', input_shape = (IMG_SIZE, IMG_SIZE, 3))\n\ncnn_out = GlobalAveragePooling2D()(cnn_base.output)\ncnn = Model(inputs=cnn_base.input, outputs=cnn_out)\nencoded_frames = TimeDistributed(cnn)(inputs)\nencoded_sequence = LSTM(256)(encoded_frames)\n\nhidden_layer = Dense(1024, activation=\"relu\")(encoded_sequence)\noutputs = Dense(50, activation=\"softmax\")(hidden_layer)\nmodel = Model([inputs], outputs)\n<\/code><\/pre>\n\n<p>In this case, when training the model it always shows <strong>accuracy ~0.02<\/strong> (it is the baseline 1\/50).<\/p>\n\n<p>Since the first model at least learned anything, I am wondering if there is any error with the way the network is build in the second case.<\/p>\n\n<p>Has anybody faced this situation? Any advice?<\/p>\n\n<p>Thank you.<\/p>\n","answers":[]}
+{"QuestionId":58138609,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-27T16:47:43.567","AcceptedAnswerId":null,"OwnerUserId":11938775.0,"Title":"Tensorflow\/Keras: Unknown: Failed to get convolution algorithm","Body":"<p>I have the following hardware\/Software: <\/p>\n\n<ul>\n<li>NVIDIA RTX2060 6GB<\/li>\n<li>Cuda 10<\/li>\n<\/ul>\n\n<p>I would like run the <a href=\"https:\/\/keras.io\/examples\/mnist_cnn\/\" rel=\"nofollow noreferrer\">Keras mnist example<\/a>. On my other Computer there works with the same software. I obtain this Error: <\/p>\n\n<pre><code>tensorflow.python.framework.errors_impl.UnknownError: 2 root error(s) found.\n  (0) Unknown: 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 conv2d_1\/convolution}}]]\n         [[conv2d_19\/Sigmoid\/_1405]]\n  (1) Unknown: 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 conv2d_1\/convolution}}]]\n<\/code><\/pre>\n\n<p>Where is my problem?<\/p>\n","answers":[{"AnswerId":"58392318","CreationDate":"2019-10-15T10:15:45.057","ParentId":null,"OwnerUserId":"12192748","Title":null,"Body":"<p>This error usually occurs due to vesion mismatch of CUDA\/cudnn. Please follow instructions on link provided: <a href=\"https:\/\/www.tensorflow.org\/install\/gpu\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/gpu<\/a>. Also if your default python is coming from Anaconda you should install evrything using conda.<\/p>\n"}]}
+{"QuestionId":58138663,"AnswerCount":0,"Tags":"<python><tensorflow><keras><autoencoder>","CreationDate":"2019-09-27T16:51:53.370","AcceptedAnswerId":null,"OwnerUserId":8057661.0,"Title":"s2s, auto encoders\/RNN\/LSTM for simulating mouse movements","Body":"<p>Does anyone have any advice\/blobs as to whether or not it is possible to build a s2s neural network model to simulate real human mouse behaviour?<\/p>\n\n<p>My current thinking is that an auto-encoder could take a final destination <code>(x,y)<\/code> point and be given a sequence of <code>(x,y)<\/code> coordinates that lead up to that destination.<\/p>\n\n<p>Do you guys think it would be possible to have a 1 to many model that given a destination to get to it provides a sequence of mouse <code>(x,y)<\/code> coordinates to get there?<\/p>\n\n<p>Training data looks like:<\/p>\n\n<pre><code>destination(x,y)\nsequence[(x1,y1), .... (x10,y10)]...\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58138758,"AnswerCount":1,"Tags":"<python><loops><logging><pytorch>","CreationDate":"2019-09-27T16:59:27.837","AcceptedAnswerId":null,"OwnerUserId":12114770.0,"Title":"Python (Pytorch) - Looping and logging","Body":"<p>I am very new to Python and would need some help on looping ANN with different number of epoch.<\/p>\n\n<p>The code starts with <\/p>\n\n<pre><code>import sys\nfrom tqdm import tqdm\nimport numpy as np\n\nimport torch\nimport torch.nn.functional as F\n\ntry:\n<\/code><\/pre>\n\n<p>and the epoch part<\/p>\n\n<pre><code>for t in range(2000):\nprediction = net(x)\n<\/code><\/pre>\n\n<p>May I ask what should I add (before and after the current code) in order to loop the programme for 3 times, then add 1000 epoch and run again till epoch reaches 10000? <\/p>\n\n<p>Thanks very much. <\/p>\n\n<p>Also, I would like to know after looping, how to keep the log of the results?<\/p>\n\n<pre><code>loss_func = torch.nn.MSELoss() \n<\/code><\/pre>\n","answers":[{"AnswerId":"58140041","CreationDate":"2019-09-27T18:50:43.433","ParentId":null,"OwnerUserId":"9374613","Title":null,"Body":"<p>Update your for loop as:<\/p>\n\n<pre><code>import sys\n\n# To log your outputs in a separate file \"myErrorLog.txt\"\nsys.stdout = open(\"myErrorLog.txt\", \"w\")\nfor t in range(2000, 10000, 1000):\n    print (\"iteration {}\".format(t))\n    # your work here\n\n# end of your code\nsys.stdout.close()\n<\/code><\/pre>\n"}]}
+{"QuestionId":58138898,"AnswerCount":2,"Tags":"<python><machine-learning><keras><conv-neural-network><vgg-net>","CreationDate":"2019-09-27T17:13:52.270","AcceptedAnswerId":58139845.0,"OwnerUserId":12096782.0,"Title":"What do I do to improve my Keras CNN VGG16 model","Body":"<p>I'm working in a project that has 700 images for 2 classes (1400 total). I'm using VGG16 but i'm new with this model and I don't know what could I do to improve this model..<\/p>\n\n<p>This is my model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>vgg16_model = VGG16(weights=\"imagenet\", include_top=True)\n\n# (1) visualize layers\nprint(\"VGG16 model layers\")\nfor i, layer in enumerate(vgg16_model.layers):\n    print(i, layer.name, layer.output_shape)\n\n# (2) remove the top layer\nbase_model = Model(input=vgg16_model.input, \n                   output=vgg16_model.get_layer(\"block5_pool\").output)\n\n# (3) attach a new top layer\nbase_out = base_model.output\nbase_out = Reshape((25088,))(base_out)\ntop_fc1 = Dense(256, activation=\"relu\")(base_out)\ntop_fc1 = Dropout(0.5)(top_fc1)\n# output layer: (None, 5)\ntop_preds = Dense(1, activation=\"sigmoid\")(top_fc1)\n\n# (4) freeze weights until the last but one convolution layer (block4_pool)\nfor layer in base_model.layers[0:14]:\n    layer.trainable = False\n\n# (5) create new hybrid model\nmodel = Model(input=base_model.input, output=top_preds)\n\n# (6) compile and train the model\nsgd = SGD(lr=1e-4, momentum=0.9)\nmodel.compile(optimizer=sgd, loss=\"binary_crossentropy\", metrics=[\"accuracy\"])\n\nhistory = model.fit([data], [labels], nb_epoch=NUM_EPOCHS, \n                    batch_size=BATCH_SIZE, validation_split=0.1)\n\n# evaluate final model\nvlabels = model.predict(np.array(valid))\n\nmodel.save('model.h5')\n<\/code><\/pre>\n\n<p>... that gives me the follow return:<\/p>\n\n<pre><code>Train on 1260 samples, validate on 140 samples\nEpoch 1\/5\n1260\/1260 [==============================] - 437s 347ms\/step - loss: 0.2200 - acc: 0.9746 - val_loss: 2.4432e-05 - val_acc: 1.0000\nEpoch 2\/5\n1260\/1260 [==============================] - 456s 362ms\/step - loss: 0.0090 - acc: 0.9984 - val_loss: 1.5452e-04 - val_acc: 1.0000\nEpoch 3\/5\n1260\/1260 [==============================] - 438s 347ms\/step - loss: 1.3702e-07 - acc: 1.0000 - val_loss: 8.4489e-05 - val_acc: 1.0000\nEpoch 4\/5\n1260\/1260 [==============================] - 446s 354ms\/step - loss: 4.2592e-06 - acc: 1.0000 - val_loss: 7.6768e-05 - val_acc: 1.0000\nEpoch 5\/5\n1260\/1260 [==============================] - 457s 363ms\/step - loss: 0.0017 - acc: 0.9992 - val_loss: 1.1921e-07 - val_acc: 1.0000\n<\/code><\/pre>\n\n<p>It seems to be a bit overfitting..<\/p>\n\n<p>My predict.py:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def fix_layer0(filename, batch_input_shape, dtype):\n    with h5py.File(filename, 'r+') as f:\n        model_config = json.loads(f.attrs['model_config'].decode('utf-8'))\n        layer0 = model_config['config']['layers'][0]['config']\n        layer0['batch_input_shape'] = batch_input_shape\n        layer0['dtype'] = dtype\n        f.attrs['model_config'] = json.dumps(model_config).encode('utf-8')\n\nfix_layer0('model.h5', [None, 224, 224, 3], 'float32')\n\nmodel = load_model('model.h5')\n\nfor filename in os.listdir(r'v\/'):\n    if filename.endswith(\".jpg\") or filename.endswith(\".ppm\") or filename.endswith(\".jpeg\") or filename.endswith(\".png\"):\n        ImageCV = cv2.resize(cv2.imread(os.path.join(TEST_DIR) + filename), (224,224))\n        ImageCV = cv2.addWeighted(ImageCV,4, cv2.GaussianBlur(ImageCV,(0,0), 224\/25), -4, 120) #The same process made when I get data in train\n        ImageCV = ImageCV.reshape(-1,224,224,3)\n        print(model.predict(ImageCV))\n<\/code><\/pre>\n\n<p>And the results are strange because only the 2 first images are of 'class 0'.. the others are 'class 1':<\/p>\n\n<pre><code>[[0.99905235]]\n[[0.]]\n[[1.]]\n[[0.012198]]\n[[0.]]\n[[1.]]\n[[1.6363418e-07]]\n[[0.99997246]]\n[[0.00433112]]\n[[0.9996668]]\n[[1.]]\n[[6.183685e-08]]\n<\/code><\/pre>\n\n<p>What can I do to improve it? I'm a little confused..<\/p>\n","answers":[{"AnswerId":"58139845","CreationDate":"2019-09-27T18:33:40.847","ParentId":null,"OwnerUserId":"12081962","Title":null,"Body":"\n\n<p>First of all, Keras <code>predict<\/code> will return the scores of the regression (probabilities for each class) and predict_classes will return the most likely class of your prediction. For example, if you classify between cats and dogs, <code>predict<\/code> could output 0.2 for cat and 0.8 for dog. \nSo, if you use <code>predict<\/code>, there should be two values per picture, one for each class.<\/p>\n\n<p>The reason why you only have one value is that your network only has one output neuron. It should have two, as there are two classes. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>top_preds = Dense(2, activation=\"sigmoid\")(top_fc1)\n<\/code><\/pre>\n\n<p>If you now want to see most likely class, not the probabilities, you should use <code>predict_classes<\/code>.<\/p>\n"},{"AnswerId":"58139870","CreationDate":"2019-09-27T18:36:01.400","ParentId":null,"OwnerUserId":"5597718","Title":null,"Body":"<blockquote>\n<pre><code>ImageCV = cv2.addWeighted(ImageCV,4, cv2.GaussianBlur(ImageCV,(0,0),\n                          224\/25), -4, 120)\n<\/code><\/pre>\n<\/blockquote>\n\n<p>Not sure why you do this for the test data. For the validation\/test data, usually only normalization is done. During training as well you need to apply the same normalization as a final step before feeding the data to the network.<\/p>\n\n<p>Refer to this example for fine-tuning VGG16 for a two class problem(dogs vs cats)\n<a href=\"https:\/\/gist.github.com\/fchollet\/7eb39b44eb9e16e59632d25fb3119975\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/fchollet\/7eb39b44eb9e16e59632d25fb3119975<\/a><\/p>\n\n<p><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<p>To reduce overfitting, you can do data augmentation for the training data i.e. feed the original data &amp; augmented data(apply operations like flip, zoom etc.). Keras <code>ImageDataGenerator<\/code>s make it easy to do the augmentation. Explored in the above tutorial as well.<\/p>\n\n<p><a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/preprocessing\/image\/<\/a><\/p>\n"}]}
+{"QuestionId":58138947,"AnswerCount":0,"Tags":"<python><keras><image-recognition>","CreationDate":"2019-09-27T17:17:44.227","AcceptedAnswerId":null,"OwnerUserId":10301261.0,"Title":"KeyError: 'val_acc'","Body":"<p>Below are some parameters for the model.fit_generator() function. Theses objects are saved in a list labeled callbacks.     <\/p>\n\n<pre><code>checkpoint = ModelCheckpoint(\n        model_file, \n        monitor= 'val_acc', \n        save_best_only=True)\n\nearly_stopping = EarlyStopping(\n    monitor='val_loss',\n    patience=5,\n    verbose=1,\n    restore_best_weights=True)\n\n\ntensorboard = TensorBoard(\n    log_dir=log_dir,\n    batch_size=batch_size,\n    update_freq = 'batch')\n\n\nreduce_lr = ReduceLROnPlateau(\n    monitor='val_loss',\n    patience=5,\n    cooldown=2,\n    min_lr=0.0000000001,\n    verbose=1)\n\n\n#-----------------------------------------------------------------------------------------------------------------#\ncallbacks = [checkpoint, reduce_lr, early_stopping, tensorboard]\n<\/code><\/pre>\n\n<p>After creating the callback objects and parameters for the objects, I implement the layers and compile(which is not shown because it is irrelevant to the problem I am having). Then I run the model.fit_generator function (which uses the callback arguments above): <\/p>\n\n<pre><code>history = model.fit_generator(\n    train_generator,\n    steps_per_epoch = steps_per_epoch,\n    epochs=epochs,\n    verbose=2,\n    callbacks=callbacks,\n    validation_data=validation_generator,\n    validation_steps=validation_steps,\n    class_weight=class_weight)\n<\/code><\/pre>\n\n<p>The error that I am getting is: <\/p>\n\n<pre><code>KeyError: 'val_acc'\n<\/code><\/pre>\n\n<p>From my understanding this means that val_acc is not in the list. But it is.. so need help to understanding why I am getting this error. <\/p>\n\n<p>Edit: <\/p>\n\n<p>Picture of the result before the error shows..[<a href=\"https:\/\/i.stack.imgur.com\/5lheg.png]\" rel=\"nofollow noreferrer\">https:\/\/i.stack.imgur.com\/5lheg.png]<\/a><\/p>\n","answers":[]}
+{"QuestionId":58139334,"AnswerCount":0,"Tags":"<keras><plaidml>","CreationDate":"2019-09-27T17:50:38.260","AcceptedAnswerId":null,"OwnerUserId":348501.0,"Title":"PlaidML Keras backend: Does it compile the entire model to a single tile function or to multiple tile functions?","Body":"<p>I am currently looking at the <a href=\"https:\/\/en.wikipedia.org\/wiki\/PlaidML\" rel=\"nofollow noreferrer\">PlaidML<\/a> Keras backend.<\/p>\n\n<p>I understand that with PlaidML tensor operations are internally programmed in tile code.<\/p>\n\n<p>Can somebody tell me, how the keras model is internally mapped to the tile language?<\/p>\n\n<p>Say, I have 100 layers in my Neural network, which in theory could be written entirely in tile language. (Matrix multiplications &amp; sigmoid functions) Does the Keras backend \"compile\" the entire network into a single monolithic tile function or does each layer and operation get compiled to its own tile function?<\/p>\n\n<p>Looking at the code, I suspect, the latter is the case. If that is so, is the context switching between python on the CPU and the operations on the GPU not a bottleneck?<\/p>\n","answers":[]}
+{"QuestionId":58139652,"AnswerCount":1,"Tags":"<java><android><tensorflow><keras><dl4j>","CreationDate":"2019-09-27T18:16:32.403","AcceptedAnswerId":null,"OwnerUserId":10276794.0,"Title":"Converting tensorflow keras LSTM model into .tflite or any working format","Body":"<p>I have trouble transforming my model into a one that can be deployed on android app via android-studio. My model is construed as follows:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(70, activation='relu', input_shape=(n_steps, n_features)))\nmodel.add(Dense(1))\nmodel.compile(optimizer='adam', loss='mse')\n<\/code><\/pre>\n\n<p>Then saved by .save function:<\/p>\n\n<pre><code>model.save(\"my.h5\")\n<\/code><\/pre>\n\n<p>How can I open this model on android app?  I have seen that tflite has problems with LSTM and my attempts to load it with dl4j failed.<\/p>\n","answers":[{"AnswerId":"58392070","CreationDate":"2019-10-15T10:02:03.313","ParentId":null,"OwnerUserId":"12192748","Title":null,"Body":"<p>Use the Tfite converter from <strong>tensorflow contrib(lite.TFLiteConverter.from_keras_model_file( 'model.h5'))<\/strong> to convert HDF5 to Tflite model.<\/p>\n"}]}
+{"QuestionId":58139679,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-27T18:18:43.317","AcceptedAnswerId":null,"OwnerUserId":231238.0,"Title":"Does the way I create and store layers in subclassed Keras `Model` have any effect?","Body":"<p>AKA Keras <code>Model<\/code> subclassing magic.<\/p>\n\n<p>While playing with Keras, I noticed, that <code>ResNetBlock.layers<\/code> gets populated as I put new instances of layers into collections I previously put into my custom model.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class ResNetBlock(Model):\n    PART_COUNT = 3\n\n    def __init__(self, kernel_size, filters):\n        super().__init__()\n        self.convs = []\n        self.batchNorms = []\n\n        for part in range(ResNetBlock.PART_COUNT):\n            if part == 1:\n                conv = Conv2D(filters[part], kernel_size=kernel_size, padding=\"same\")\n            else:\n                conv = Conv2D(filters[part], kernel_size=(1,1))\n            self.convs.append(conv)\n            self.batchNorms.append(BatchNormalization())\n\nresnet = ResNetBlock(1, [1, 2, 3])\nprint(resnet.layers) # actually prints non-empty list\n                     # filled with Conv2Ds and BNs from above\n<\/code><\/pre>\n\n<p>Adopted from official tutorial: <a href=\"https:\/\/www.tensorflow.org\/beta\/tutorials\/eager\/custom_layers\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/beta\/tutorials\/eager\/custom_layers<\/a><\/p>\n\n<p>A bit of digging into TensorFlow source showed, that some kind of tracking is used via <code>__setattr__<\/code> in <code>Network<\/code> class.<\/p>\n\n<p>Now the code is not trivial, documentation lacking, and it seems unclear if the order of creating new layers\/adding them to respective collections matters at all? E.g. if I first fill in <code>convs<\/code> collection, and only then <code>batchNorms<\/code> collection, would it still be the same model?<\/p>\n\n<p>In most tutorials each layer is actually put into its own attribute.<\/p>\n\n<p>Bonus question is: why is it done so implicitly? This kind of magic kinda breaks the motto to prefer explicit over implicit. What if for some reason I'd need to use a custom collection type not derived from <code>list<\/code>? How would I ensure these magic operations are done properly?<\/p>\n","answers":[{"AnswerId":"58141897","CreationDate":"2019-09-27T22:26:17.743","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>The order won't matter. What really changes your model is the <code>call<\/code> method. This stores the order of the operations (even if the order of the weights were variable, they would be applied in the same graph with the same functions)    <\/p>\n\n<p>Now, if you suspect that not using a \"property\", but using another kind of storage for the layers, would not register the layer for some reason, you can double check with:<\/p>\n\n<pre><code>print(len(resnet.trainable_weights))\n<\/code><\/pre>\n\n<p>The count should be <code>6 * PART_COUNT<\/code>:    <\/p>\n\n<ul>\n<li>2 tensors for the conv layers (kernel and bias)   <\/li>\n<li>4 tensors for the BatchNormalization layers (mean, variance, scale and offset)    <\/li>\n<\/ul>\n"}]}
+{"QuestionId":58139803,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><machine-learning><ludwig>","CreationDate":"2019-09-27T18:29:56.370","AcceptedAnswerId":58390486.0,"OwnerUserId":11064696.0,"Title":"TypeError: 'Tensor' object is not callable Ludwig Uber","Body":"<p>I'm new to Uber Ludwig, an I'm getting the error in the Title,\nThe StackTrace is:<\/p>\n\n<pre><code>Epoch   1\nTraining: 100%|\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\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\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\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\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\u2588\u2588| 100\/100 [00:56&lt;00:00,  1.76it\/s]\nTraceback (most recent call last):\n  File \"\/usr\/local\/bin\/ludwig\", line 10, in &lt;module&gt;\n    sys.exit(main())\n  File \"\/usr\/local\/lib\/python3.7\/dist-packages\/ludwig\/cli.py\", line 108, in main\n    CLI()\n  File \"\/usr\/local\/lib\/python3.7\/dist-packages\/ludwig\/cli.py\", line 64, in __init__\n    getattr(self, args.command)()\n  File \"\/usr\/local\/lib\/python3.7\/dist-packages\/ludwig\/cli.py\", line 69, in experiment\n    experiment.cli(sys.argv[2:])\n  File \"\/usr\/local\/lib\/python3.7\/dist-packages\/ludwig\/experiment.py\", line 472, in cli\n    experiment(**vars(args))\n  File \"\/usr\/local\/lib\/python3.7\/dist-packages\/ludwig\/experiment.py\", line 203, in experiment\n    **kwargs\n  File \"\/usr\/local\/lib\/python3.7\/dist-packages\/ludwig\/train.py\", line 311, in full_train\n    debug=debug\n  File \"\/usr\/local\/lib\/python3.7\/dist-packages\/ludwig\/train.py\", line 475, in train\n    **model_definition['training']\n  File \"\/usr\/local\/lib\/python3.7\/dist-packages\/ludwig\/models\/model.py\", line 576, in train\n    bucketing_field\nTypeError: 'Tensor' object is not callable\n\n<\/code><\/pre>\n\n<p>So I have a CSV file with data in the following format:<\/p>\n\n<pre><code>elo\\\\,evaluation\n1063\\\\,\"12\\\\, 30\\\\, 16\\\\, 2\\\\, 17\\\\, 28\\\\, 1\\\\, 5\\\\, 7\\\\, 0\\\\, 39\\\\, 0\"\n1034\\\\,\"18\\\\, 13\\\\, 13\\\\, 1\\\\, 8\\\\, 20\\\\, 35\\\\, 0\\\\, 20\\\\, 11\\\\, 8\\\\, 9\"\n1713\\\\,\"23\\\\, 13\\\\, 26\\\\, 35\\\\, 36\\\\, 33\\\\, 14\\\\, 20\\\\, 7\\\\, 11\\\\, 1\\\\, 3\\\\, 5\\\\, 14\\\\, 16\\\\, 2\\\\, 6\\\\, 19\\\\, 60\\\\, 193\\\\, 58\\\\, 100\\\\, 82\\\\, 24\\\\, 2985\\\\, 1\\\\, 230\\\\, 0\\\\, 0\\\\, 0\\\\, 0\\\\, 0\\\\, 0\\\\, 0\"\n1812\\\\,\"1\\\\, 2\\\\, 53\\\\, 3\\\\, 66\\\\, 5\\\\, 14\\\\, 8\\\\, 56\\\\, 29\\\\, 7\\\\, 30\\\\, 14\\\\, 15\\\\, 10\\\\, 14\\\\, 4\\\\, 34\\\\, 1\\\\, 22\\\\, 15\\\\, 26\\\\, 133\\\\, 6\\\\, 59\\\\, 0\\\\, 257\\\\, 553\\\\, 0\\\\, 0\\\\, 0\\\\, 0\\\\, 0\\\\, 0\"\n1718\\\\,\"21\\\\, 83\\\\, 23\\\\, 6\\\\, 93\\\\, 8\\\\, 26\\\\, 25\\\\, 19\\\\, 12\\\\, 21\\\\, 23\\\\, 10\\\\, 5\\\\, 6\\\\, 12\\\\, 6\\\\, 11\\\\, 7\\\\, 38\\\\, 0\\\\, 3\\\\, 33\\\\, 18\\\\, 11\\\\, 6\\\\, 22\\\\, 14\\\\, 21\\\\, 6\\\\, 19\\\\, 13\\\\, 8\\\\, 0\\\\, 4158\\\\, 0\\\\, 5\\\\, 1\\\\, 0\\\\, 1\\\\, 0\\\\, 0\\\\, 0\\\\, 0\\\\, 0\\\\, 0\\\\, 0\\\\, 0\\\\, 0\\\\, 0\\\\, 0\\\\, 0\\\\, 0\"\n1743\\\\,\"4\\\\, 17\\\\, 22\\\\, 25\\\\, 36\\\\, 7\\\\, 63\\\\, 5\\\\, 4\\\\, 11\\\\, 26\\\\, 12\\\\, 0\\\\, 17\\\\, 3\\\\, 2\\\\, 10\\\\, 16\\\\, 3\\\\, 19\\\\, 17\\\\, 13\\\\, 2\\\\, 5\\\\, 49\\\\, 40\\\\, 7\\\\, 1\\\\, 42\\\\, 10\\\\, 2\\\\, 36\\\\, 48\\\\, 8\\\\, 1175\\\\, 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>Now my Model definition is:<\/p>\n\n<pre><code>input_features:\n  -\n    name: evaluation\n    type: sequence\noutput_features:\n  -\n    name: elo\n    type: numerical\n<\/code><\/pre>\n\n<p>and the command I run on cli:<\/p>\n\n<pre><code>ludwig experiment --data_csv evaluated.csv --model_definition_file modelDefinition.yml --output_directory results\n<\/code><\/pre>\n\n<p>Now I have no idea whats why the error occurs can someone help me?<\/p>\n","answers":[{"AnswerId":"58390486","CreationDate":"2019-10-15T08:36:14.670","ParentId":null,"OwnerUserId":"12192748","Title":null,"Body":"<p>As evaluation is protected keyword in ludwig try changing the name of input feature to something other than evaluation.<\/p>\n"}]}
+{"QuestionId":58140866,"AnswerCount":2,"Tags":"<python><arrays><vectorization><pytorch><tensor>","CreationDate":"2019-09-27T20:14:13.370","AcceptedAnswerId":null,"OwnerUserId":1667423.0,"Title":"Vectorizing assignment of a tensor to a slice in PyTorch","Body":"<p>I'm trying to vectorize a slice assignment of the form<\/p>\n\n<pre><code>for i in range(a.shape[1]):\n    for j in range(a.shape[2]):\n        a[:,i,j,:,i:i+b.shape[2],j:j+b.shape[3]] = b\n<\/code><\/pre>\n\n<p>where <code>b<\/code> itself is an array. This is because the nested Python loop is too inefficient and is taking up most of the runtime. Is there a way to do this?<\/p>\n\n<p>For a simpler case, consider the following:<\/p>\n\n<pre><code>for i in range(a.shape[1]):\n    a[:,i,:,i:i+b.shape[2]] = b\n<\/code><\/pre>\n\n<p>This is what <code>b<\/code> and <code>a<\/code> might look like:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/owWNh.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/owWNh.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>You can see the diagonal, \"sliding\" structure of the resulting matrix.<\/p>\n","answers":[{"AnswerId":"58141376","CreationDate":"2019-09-27T21:11:21.417","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>Assuming <code>b<\/code> has a shape <code>(2,3,x1)<\/code>, and <code>a<\/code> has a shape <code>(2,x2-x1+1,3,x2)<\/code>. In your screenshot, we can infer that <code>x1=4<\/code>, <code>x2=6<\/code>.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\n\nb_shape = (2,3,4)\na_shape = (2,3,3,6)\n\nb = np.arange(1,25).reshape(b_shape)\n#array([[[ 1,  2,  3,  4],\n#        [ 5,  6,  7,  8],\n#        [ 9, 10, 11, 12]],\n#\n#       [[13, 14, 15, 16],\n#        [17, 18, 19, 20],\n#        [21, 22, 23, 24]]])\n\nc = np.pad(b, (*[(0,0) for _ in range(len(b_shape[:-1]))], (0,a_shape[-1]-b_shape[-1])), 'constant')\n\n#array([[[ 1,  2,  3,  4,  0,  0],\n#        [ 5,  6,  7,  8,  0,  0],\n#        [ 9, 10, 11, 12,  0,  0]],\n#\n#       [[13, 14, 15, 16,  0,  0],\n#        [17, 18, 19, 20,  0,  0],\n#        [21, 22, 23, 24,  0,  0]]])\n\na = np.stack([np.roll(c, shift=i) for i in range(a_shape[-1]-b_shape[-1]+1)], axis=1)\n\n# array([[[[ 1,  2,  3,  4,  0,  0],\n#          [ 5,  6,  7,  8,  0,  0],\n#          [ 9, 10, 11, 12,  0,  0]],\n\n#         [[ 0,  1,  2,  3,  4,  0],\n#          [ 0,  5,  6,  7,  8,  0],\n#          [ 0,  9, 10, 11, 12,  0]],\n\n#         [[ 0,  0,  1,  2,  3,  4],\n#          [ 0,  0,  5,  6,  7,  8],\n#          [ 0,  0,  9, 10, 11, 12]]],\n\n\n#        [[[13, 14, 15, 16,  0,  0],\n#          [17, 18, 19, 20,  0,  0],\n#          [21, 22, 23, 24,  0,  0]],\n\n#         [[ 0, 13, 14, 15, 16,  0],\n#          [ 0, 17, 18, 19, 20,  0],\n#          [ 0, 21, 22, 23, 24,  0]],\n\n#         [[ 0,  0, 13, 14, 15, 16],\n#          [ 0,  0, 17, 18, 19, 20],\n#          [ 0,  0, 21, 22, 23, 24]]]])\n<\/code><\/pre>\n"},{"AnswerId":"58141846","CreationDate":"2019-09-27T22:18:54.220","ParentId":null,"OwnerUserId":"3293881","Title":null,"Body":"<p>We can leverage <a href=\"http:\/\/www.scipy-lectures.org\/advanced\/advanced_numpy\/#indexing-scheme-strides\" rel=\"nofollow noreferrer\"><code>np.lib.stride_tricks.as_strided<\/code><\/a> based <a href=\"http:\/\/scikit-image.org\/docs\/dev\/api\/skimage.util.html#skimage.util.view_as_windows\" rel=\"nofollow noreferrer\"><code>scikit-image's view_as_windows<\/code><\/a> to get sliding windowed views into a <code>0s<\/code> padded version of the input and being a view would be efficient on memory and performance.  <a href=\"https:\/\/stackoverflow.com\/a\/51890064\/\">More info on use of <code>as_strided<\/code> based <code>view_as_windows<\/code><\/a>.<\/p>\n\n<p>Hence, for the simpler case, it would be -<\/p>\n\n<pre><code>from skimage.util.shape import view_as_windows\n\ndef sliding_2D_windows(b, outshp_axis1):\n    # outshp_axis1 is desired output's shape along axis=1\n    n = outshp_axis1-1\n    b1 = np.pad(b,((0,0),(0,0),(n,n)),'constant')\n    w_shp = (1,b1.shape[1],b.shape[2]+n)\n    return view_as_windows(b1,w_shp)[...,0,::-1,0,:,:]\n<\/code><\/pre>\n\n<p>Sample run -<\/p>\n\n<pre><code>In [192]: b\nOut[192]: \narray([[[54, 57, 74, 77],\n        [77, 19, 93, 31],\n        [46, 97, 80, 98]],\n\n       [[98, 22, 68, 75],\n        [49, 97, 56, 98],\n        [91, 47, 35, 87]]])\n\nIn [193]: sliding_2D_windows(b, outshp_axis1=3)\nOut[193]: \narray([[[[54, 57, 74, 77,  0,  0],\n         [77, 19, 93, 31,  0,  0],\n         [46, 97, 80, 98,  0,  0]],\n\n        [[ 0, 54, 57, 74, 77,  0],\n         [ 0, 77, 19, 93, 31,  0],\n         [ 0, 46, 97, 80, 98,  0]],\n\n        [[ 0,  0, 54, 57, 74, 77],\n         [ 0,  0, 77, 19, 93, 31],\n         [ 0,  0, 46, 97, 80, 98]]],\n\n\n       [[[98, 22, 68, 75,  0,  0],\n         [49, 97, 56, 98,  0,  0],\n         [91, 47, 35, 87,  0,  0]],\n\n        ....\n\n        [[ 0,  0, 98, 22, 68, 75],\n         [ 0,  0, 49, 97, 56, 98],\n         [ 0,  0, 91, 47, 35, 87]]]])\n<\/code><\/pre>\n"}]}
+{"QuestionId":58141405,"AnswerCount":0,"Tags":"<pytorch>","CreationDate":"2019-09-27T21:14:36.960","AcceptedAnswerId":null,"OwnerUserId":9105277.0,"Title":"Turn list of tuples into binary tensors?","Body":"<p>I have a list of tuples in the form below. A given tuple represents a given pair of movies that a given user liked. All tuples, together, capture every combination of movie likes found in my data. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>[(movie_a,movie_b),...(movie_a,movie_b)] \n<\/code><\/pre>\n\n<p>My task is to create movie embeddings, similar to word embeddings. The idea is to train a single hidden layer NN to predict the most likely movie which any user might like, given a movie supplied. Much like word embeddings, the task is inconsequential; it's the weight matrix I'm after, which maps movies to vectors. <\/p>\n\n<p>Reference: <a href=\"https:\/\/arxiv.org\/vc\/arxiv\/papers\/1603\/1603.04259v2.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/vc\/arxiv\/papers\/1603\/1603.04259v2.pdf<\/a><\/p>\n\n<p>In total, there are 19,000,000 tuples (training examples.) Likewise, there are 9,000 unique movie IDs in my data. My initial goal was to create an input variable, X where each row represented a unique movie_id, and each column represented a unique observation. In any given column, only one cell would be set to 1, with all other values set to 0.<\/p>\n\n<p>As an intermediate step, I tried creating a matrix of zeros of the right dimensions<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>X = np.zeros([9000,19000000])\n<\/code><\/pre>\n\n<p>Understandably, my computer crashed, simply trying to allocate sufficient memory to X. <\/p>\n\n<p>Is there a memory efficient way to pass my list of values into PyTorch, such that a binary vector is created for every training example?<\/p>\n\n<p>Likewise, I tried randomly sampling 500,000 observations. But similarly, passing 9000,500000 to np.zeroes() resulted in another crash.<\/p>\n\n<p>My university has a GPU server available, and that's my next stop. But I'd like to know if there's a memory efficient way that I should be doing this, especially since I'll be using shared resources.  <\/p>\n","answers":[]}
+{"QuestionId":58141530,"AnswerCount":1,"Tags":"<tensorflow><ml.net>","CreationDate":"2019-09-27T21:30:38.350","AcceptedAnswerId":59074571.0,"OwnerUserId":2416394.0,"Title":"Inception settings changed: Computed output size would be negative: -4 [input_size: 2, effective_filter_size: 7, stride: 1]","Body":"<p>I was working along the image classifier ML.net sample code over at <a href=\"https:\/\/docs.microsoft.com\/en-US\/dotnet\/machine-learning\/tutorials\/image-classification\" rel=\"nofollow noreferrer\">https:\/\/docs.microsoft.com\/en-US\/dotnet\/machine-learning\/tutorials\/image-classification<\/a><\/p>\n\n<p>The classification there uses the following inception settings<\/p>\n\n<pre><code>private struct InceptionSettings\n{\n    public const int ImageHeight = 224;\n    public const int ImageWidth = 224;\n    public const float Mean = 117;\n    public const float Scale = 1;\n    public const bool ChannelsLast = true;\n}\n<\/code><\/pre>\n\n<p>while using the tensorflow inception5h model.\nIt appears to be working. Unclear for me is however, what breaks when I change <code>Height<\/code> and <code>Width<\/code> from <strong>224<\/strong> to say <strong>64<\/strong> to reduce the load and precision of the prediction the <code>reuse and tune inception model<\/code> part nearly instantly crashes with<\/p>\n\n<pre><code>System.InvalidOperationException: Splitter\/consolidator worker encountered exception while consuming source data ---&gt; Microsoft.ML.Transforms.TensorFlow.TFException: Computed output size would be negative: -4 [input_size: 2, effective_filter_size: 7, stride: 1]\n     [[{{node avgpool0}}]]\n   at Microsoft.ML.Transforms.TensorFlow.TFStatus.CheckMaybeRaise(TFStatus incomingStatus, Boolean last)\n   at Microsoft.ML.Transforms.TensorFlow.TFSession.Run(TFOutput[] inputs, TFTensor[] inputValues, TFOutput[] outputs, TFOperation[] targetOpers, TFBuffer runMetadata, TFBuffer runOptions, TFStatus status)\n   at Microsoft.ML.Transforms.TensorFlow.TFSession.Runner.Run(TFStatus status)\n   at Microsoft.ML.Transforms.TensorFlowTransformer.Mapper.UpdateCacheIfNeeded(Int64 position, ITensorValueGetter[] srcTensorGetters, String[] activeOutputColNames, OutputCache outputCache)\n   at Microsoft.ML.Transforms.TensorFlowTransformer.Mapper.&lt;&gt;c__DisplayClass8_0`1.&lt;MakeGetter&gt;b__3(VBuffer`1&amp; dst)\n   at Microsoft.ML.Data.DataViewUtils.Splitter.InPipe.Impl`1.Fill()\n   at Microsoft.ML.Data.DataViewUtils.Splitter.&lt;&gt;c__DisplayClass5_1.&lt;ConsolidateCore&gt;b__2()\n   --- End of inner exception stack trace ---\n   at Microsoft.ML.Data.DataViewUtils.Splitter.Batch.SetAll(OutPipe[] pipes)\n   at Microsoft.ML.Data.DataViewUtils.Splitter.Cursor.MoveNextCore()\n   at Microsoft.ML.Data.RootCursorBase.MoveNext()\n   at Microsoft.ML.Trainers.TrainingCursorBase.MoveNext()\n   at Microsoft.ML.Trainers.LbfgsTrainerBase`3.TrainCore(IChannel ch, RoleMappedData data)\n   at Microsoft.ML.Trainers.LbfgsTrainerBase`3.TrainModelCore(TrainContext context)\n   at Microsoft.ML.Trainers.TrainerEstimatorBase`2.TrainTransformer(IDataView trainSet, IDataView validationSet, IPredictor initPredictor)\n   at Microsoft.ML.Data.EstimatorChain`1.Fit(IDataView input)\n   at D:\\My\\MLTrainer.Program.ReuseAndTuneInceptionModel(MLContext mlContext, TrainerData trainerData, String dataLocation, String inputModelLocation, String outputModelLocation) in MLTrainer\\Program.cs:line 66\n   at MLTrainer.Program.Main(String[] args) in D:\\My\\MLTrainer\\Program.cs:line 29\n<\/code><\/pre>\n\n<p>Now I don't get what I can do and where in the details the issue is buried. Is the pre-trained model already fixed to this somewhat strange resolution? The resolution itself seems not to be used somewhere else, nor do I get why the splitter don't like it.<\/p>\n\n<p>Do I hit some sort of min size condition I am not aware of? If so, what are the boundaries? I tried 1024x1024 for instance, which failed with another error.<\/p>\n\n<p>Any hints on that are appreciated :)<\/p>\n","answers":[{"AnswerId":"59074571","CreationDate":"2019-11-27T16:26:51.273","ParentId":null,"OwnerUserId":"12446553","Title":null,"Body":"<p>You could install Netron to look at your model.<\/p>\n\n<p>You'll see that on the first layer will stand something like Nx224x224x3.<\/p>\n\n<p>The model is fixed to this input resolution, because it is trained with a large dataset in exactly this resolution. You could change the input layer with e.g. keras. \nBut I never tried to change the input layer in ML.NET.<\/p>\n"}]}
+{"QuestionId":58141812,"AnswerCount":0,"Tags":"<xml><python-3.x><tensorflow><labelimg>","CreationDate":"2019-09-27T22:13:27.640","AcceptedAnswerId":null,"OwnerUserId":12132881.0,"Title":"Export labeled object after object detection as XML (Pascal VOC) to reimport it to LabelImage","Body":"<p>i recently trained a network for object detection using tensorflow with faster_rcnn_inception_v2_coco  following the guide: \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>Now I would love to use my network to iterate through a folder of images, creating a XML file for each Image in the PascalVOC Format, which can be opened with LabelImage. Can somebody help me with this Problem? I think it would be a great solution to increase the number of labeled images fast. <\/p>\n\n<p>This is the code that i use to iterate through a folder, labeling each image and displaying it:<\/p>\n\n<pre><code>for filename in glob.iglob('Folder_Path\/*', recursive=True):\n    if os.path.isfile(filename): # filter dirs\n        image = cv2.imread(filename)\n        image_expanded = np.expand_dims(image, 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: image_expanded})\n\n# Draw the results of the detection\n\n        vis_util.visualize_boxes_and_labels_on_image_array(\n            image,\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=0.80)\n\n# All the results have been drawn on image. Now display the image.\n        cv2.imshow('Object detector', image)\n\n# Press any key to close the image\n        cv2.waitKey(0)\n\n# Clean up\n        cv2.destroyAllWindows()\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58141906,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-27T22:27:55.890","AcceptedAnswerId":null,"OwnerUserId":10677155.0,"Title":"How can I grab the index of the value from a 1D Tensor","Body":"<p>I am trying to calculate the MRR and for that, I need to find the rank of a given value (y_actual) in a tensor (y_pred). \ny_actual is a string ('42' for example) where y_pred is a list containing multiple string ids. The goal is to find the position of that value in the y_pred (1D TENSOR)<\/p>\n","answers":[]}
+{"QuestionId":58142459,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><tensorflow-datasets><mnist>","CreationDate":"2019-09-28T00:21:55.870","AcceptedAnswerId":null,"OwnerUserId":752843.0,"Title":"Unexpected dimensions with TensorFlow dataset","Body":"<p>I am trying to do transfer learning with InceptionV3 on the MNIST dataset.<\/p>\n\n<p>The plan is to read in the MNIST dataset, resize the images, and then use these to train, like so:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport os\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom PIL import Image\n\nimport tensorflow.compat.v2 as tf\nimport tensorflow.compat.v1 as tfv1\nfrom tensorflow.python.keras.applications import InceptionV3\n\ntfv1.enable_v2_behavior()\n\nprint(tf.version.VERSION)\n\nimg_size = 299\n\ndef preprocess_tf_image(image, label):\n  image = tf.image.grayscale_to_rgb(image)\n  image = tf.image.resize(image, [img_size, img_size])\n  return image, label\n\n#Acquire MNIST data\nmnist = tf.keras.datasets.mnist\n(x_train, y_train), (x_test, y_test) = mnist.load_data()\n#Convert data to [0,1] range\nx_train, x_test = x_train \/ 255.0, x_test \/ 255.0\n\n#Add extra dimension to images so that they can be converted to RGB\nx_train = x_train.reshape(x_train.shape[0], 28, 28, 1)\nx_test  = x_test.reshape (x_test.shape[0],  28, 28, 1)\n\nx_train = tf.data.Dataset.from_tensor_slices((x_train, y_train))\nx_test  = tf.data.Dataset.from_tensor_slices((x_test, y_test))\n\n#Convert images to RGB space and resize\nx_train = x_train.map(preprocess_tf_image)\nx_test  = x_test.map(preprocess_tf_image)\n\nimg_shape = (img_size, img_size, 3)\n\n#Get trained model, but leave off the head\nbase_model = InceptionV3(input_shape = img_shape, weights='imagenet', include_top=False)\nbase_model.trainable = False\n\n#Make a model with a new head\nmodel = tf.keras.Sequential([\n  base_model,\n  tf.keras.layers.GlobalAveragePooling2D(),\n  tf.keras.layers.Dropout(0.2),\n  tf.keras.layers.Dense(10, activation='softmax')\n])\n\n#Compile model\nmodel.compile(\n    optimizer='adam', #tf.keras.optimizers.RMSprop(lr=BASE_LEARNING_RATE),\n    loss='binary_crossentropy',\n    metrics=['accuracy']\n)\n\nmodel.fit(x_train, epochs=5)\n\nmodel.evaluate(x_test)\n<\/code><\/pre>\n\n<p>But, when I run this, things stop at <code>model.fit()<\/code> with the error:<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking input: expected inception_v3_input to have 4 dimensions, but got array with shape (299, 299, 3)<\/p>\n<\/blockquote>\n\n<p>What's going on?<\/p>\n","answers":[{"AnswerId":"58169330","CreationDate":"2019-09-30T13:51:27.677","ParentId":null,"OwnerUserId":"483157","Title":null,"Body":"<p>After you apply <code>map<\/code> to a dataset the response has no information about the batch size, you have to invoke the <code>batch<\/code> function to add it:<\/p>\n\n<pre><code>x_train = x_train.batch(batch_size = BATCH_SIZE) # adds batch size dimension to train dataset\n\nx_test = x_test.batch(batch_size = BATCH_SIZE) # idem for test.\n<\/code><\/pre>\n\n<p>After that I could fully train and evaluate the model using Google's Colab as you can check <a href=\"https:\/\/colab.research.google.com\/drive\/1ol06nDxJeUU044DSIyw_8G8rA8xzyiUS\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
+{"QuestionId":58142567,"AnswerCount":0,"Tags":"<python><numpy><tensorflow>","CreationDate":"2019-09-28T00:50:24.670","AcceptedAnswerId":null,"OwnerUserId":11876924.0,"Title":"Streaming NumPy data as input to Tensorflow","Body":"<p>I was reading <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/datasets<\/a> to look for a solution to stream NumPy arrays stored in npz files, which may be too large to fit in memory. This snippet is provided in the documentation:<\/p>\n\n<pre><code># Load the training data into two NumPy arrays, for example using `np.load()`.\nwith np.load(\"\/var\/data\/training_data.npy\") as data:\n  features = data[\"features\"]\n  labels = data[\"labels\"]\n\n# Assume that each row of `features` corresponds to the same row as `labels`.\nassert features.shape[0] == labels.shape[0]\n\nfeatures_placeholder = tf.placeholder(features.dtype, features.shape)\nlabels_placeholder = tf.placeholder(labels.dtype, labels.shape)\n\ndataset = tf.data.Dataset.from_tensor_slices((features_placeholder, labels_placeholder))\n# [Other transformations on `dataset`...]\ndataset = ...\niterator = dataset.make_initializable_iterator()\n\nsess.run(iterator.initializer, feed_dict={features_placeholder: features,\n                                          labels_placeholder: labels})\n<\/code><\/pre>\n\n<p>Does this method really allow you to stream NumPy data? Doesn't <code>features = data[\"features\"]<\/code> load the data entirely into memory?<\/p>\n","answers":[]}
+{"QuestionId":58142890,"AnswerCount":2,"Tags":"<python><keras>","CreationDate":"2019-09-28T02:24:26.197","AcceptedAnswerId":null,"OwnerUserId":8684109.0,"Title":"In Keras, I want to save a trained model on my disk (Windows), but got error message \"Invalid argument\"?","Body":"<p>This is my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>sava_path = 'D:\/5-PyCharm\/unet-membrane\/membrane\/data\/models\/' + model_name\ncheckpoint = ModelCheckpoint(filepath=sava_path, monitor='binary_accuracy', verbose=1,\n                             save_best_only=True, mode='max')\nhistory = unet_model.fit(x=x_train, y=y_train, batch_size=batch_size, callbacks=[checkpoint],\n                         validation_split=0.2, epochs=epochs, shuffle=False)\n<\/code><\/pre>\n\n<p>but got an error:<\/p>\n\n<pre><code>OSError: Unable to create file (unable to open file: name = 'D:\/5-PyCharm\/unet-membrane\/membrane\/data\/models\/Unet_membrane2019-09-28_09:50:25.h5', errno = 22, error message = 'Invalid argument', flags = 13, o_flags = 302)\n<\/code><\/pre>\n\n<p>If the <code>save_path<\/code> is <code>.\/membrane\/data\/models\/unet_membrane.h5<\/code>, everything is ok.\nThe most important difference between the two is whether the path is preceded by a Windows disk character.\nSo, I want to ask whether there is a <code>D:<\/code> in <code>save_path<\/code> to cause this error? <\/p>\n\n<p>Click this link to see my captured images:\n<a href=\"https:\/\/i.stack.imgur.com\/chiAu.png\" rel=\"nofollow noreferrer\">save successful<\/a>, <a href=\"https:\/\/i.stack.imgur.com\/GBGUQ.png\" rel=\"nofollow noreferrer\">got an error \"invalid argument\"<\/a><\/p>\n","answers":[{"AnswerId":"58143858","CreationDate":"2019-09-28T06:11:49.457","ParentId":null,"OwnerUserId":"6117017","Title":null,"Body":"<p>Choose another model name, or <strong>try replacing '\/' with '\\\\'<\/strong>. This also occurred to me while creating the 'logs' folder from the Tensorboard callback.<\/p>\n\n<p>The method that I mentioned helped me solve me issue.<\/p>\n"},{"AnswerId":"58160486","CreationDate":"2019-09-30T01:20:08.317","ParentId":null,"OwnerUserId":"8684109","Title":null,"Body":"<p><code>Invalid argument<\/code> is caused by illegal character. The file name of Windows file can't include <code>\\\/:*?\"&lt;&gt;|<\/code>. Your file name has the character <code>:<\/code>, please remove it.<\/p>\n"}]}
+{"QuestionId":58143008,"AnswerCount":2,"Tags":"<tensorflow><gradient><data-science><gradient-descent><backpropagation>","CreationDate":"2019-09-28T02:55:31.527","AcceptedAnswerId":58182837.0,"OwnerUserId":5581893.0,"Title":"Where is backpropagation performed in this example","Body":"<p>I have an example of DNN learning XOR (right click to open in new tab): <a href=\"https:\/\/colab.research.google.com\/drive\/1M5xFp4gaXPCbnejM8-5_yLp1B6UvwdL8\" rel=\"nofollow noreferrer\">https:\/\/colab.research.google.com\/drive\/1M5xFp4gaXPCbnejM8-5_yLp1B6UvwdL8<\/a><\/p>\n\n<p>I'm confused in these 2 lines (related to backpropagation):<\/p>\n\n<pre><code>Grads = T.gradient(Loss,[W1,B1,W2,B2]);\nOptim.apply_gradients(zip(Grads,[W1,B1,W2,B2]));\n<\/code><\/pre>\n\n<p>I'm guessing the backward loop is at <code>T.gradient<\/code> because those are gradient values related to loss, but I'm still not clear. The questions are:<\/p>\n\n<ul>\n<li>Question1. Is there backpropagation (the backward loop) in those 2 lines?<\/li>\n<li>Question2. If there is backpropagation, it's at <code>T.gradient<\/code> or <code>Optim.apply_gradients<\/code>?<\/li>\n<li>Question3. Because backpropagation is done backward, is the order of <code>[W1,B1,W2,B2]<\/code> important? I believe, eg. this shuffled <code>[B1,W2,B2,W1]<\/code> can't be the same, because backpropagation needs layer order from output back to input.<\/li>\n<\/ul>\n\n<p>From my trying, when shuffling the order of weights and biases in the variable array, the optimisation process is still working. <strong>But backpropagation needs layer order from output back to input, I don't get this<\/strong>.<\/p>\n\n<p>Source code:<\/p>\n\n<pre><code>#!pip install tensorflow==2.0.0rc2\n%tensorflow_version 2.x\n%reset -f\n\n#libs\nimport tensorflow as tf;\n\n#data\nX = [[0,0],[0,1],[1,0],[1,1]];\nY = [[0],  [1],  [1],  [0]  ];\nX = tf.convert_to_tensor(X,tf.float32);\nY = tf.convert_to_tensor(Y,tf.float32);\n\n#model\nW1 = tf.Variable(tf.random.uniform([2,20],-1,1));\nB1 = tf.Variable(tf.random.uniform([  20],-1,1));\n\nW2 = tf.Variable(tf.random.uniform([20,1],-1,1));\nB2 = tf.Variable(tf.random.uniform([   1],-1,1));\n\n@tf.function\ndef feedforward(X):\n  H1  = tf.nn.leaky_relu(tf.matmul(X,W1) + B1);\n  Out = tf.sigmoid(tf.matmul(H1,W2) + B2);\n  return Out;\n#end def\n\n#train\nOptim = tf.keras.optimizers.SGD(1e-1);\nSteps = 1000;\n\nfor I in range(Steps):\n  if I%(Steps\/10)==0:\n    Out  = feedforward(X);\n    Loss = tf.reduce_sum(tf.square(Y-Out));\n    print(\"Loss:\",Loss.numpy());\n  #end if\n\n  with tf.GradientTape() as T:\n    Out  = feedforward(X);\n    Loss = tf.reduce_sum(tf.square(Y-Out));\n  #end with\n\n  #BACKPROPAGATION HERE?\n  Grads = T.gradient(Loss,[W1,B1,W2,B2]);\n  Optim.apply_gradients(zip(Grads,[W1,B1,W2,B2]));\n#end for\n\nOut  = feedforward(X);\nLoss = tf.reduce_sum(tf.square(Y-Out));\nprint(\"Loss:\",Loss.numpy(),\"(Last)\");\n\nprint(\"\\nDone.\");\n#eof\n<\/code><\/pre>\n","answers":[{"AnswerId":"58143069","CreationDate":"2019-09-28T03:15:35.843","ParentId":null,"OwnerUserId":"5533928","Title":null,"Body":"<p>Let's take this one step at a time.<\/p>\n\n<p>Step 1: Calculation of Gradients:<\/p>\n\n<pre><code>Grads = T.gradient(Loss,[W1,B1,W2,B2])\n<\/code><\/pre>\n\n<p>Here, we calculate the gradients of the loss with respect to the variables in the provided list. The list of gradients is indexed based on the indices of the variables. This means that <code>Grads[0]<\/code> will be the gradients with respect to <code>W1<\/code>, and so on.<\/p>\n\n<p>Step 2: Next, we perform the update. This is done in:<\/p>\n\n<pre><code>Optim.apply_gradients(zip(Grads,[W1,B1,W2,B2]))\n<\/code><\/pre>\n\n<p>Here, <code>Grads[0]<\/code> are used to update <code>W1<\/code>, <code>Grads[1]<\/code> to update <code>B1<\/code> and so on.<\/p>\n\n<p>Note that gradient calculation and the update steps are performed separately. So as long as the variables appear in the same order in both lists, there shouldn't be any problems.<\/p>\n\n<p>Also, <code>GradientTape<\/code> has to be used with Eager Execution.<\/p>\n"},{"AnswerId":"58182837","CreationDate":"2019-10-01T10:25:23.593","ParentId":null,"OwnerUserId":"5581893","Title":null,"Body":"<p>With TensorFlow 2 in default eager mode, and even without the @tf.function decorator to make graph. TensorFlow is still tracking the relation between tensors while calculation: <a href=\"https:\/\/stats.stackexchange.com\/a\/272000\/142160\">https:\/\/stats.stackexchange.com\/a\/272000\/142160<\/a><\/p>\n\n<p>TensorFlow tracks every variables here:<\/p>\n\n<pre><code>with tf.GradientTape() as T:\n    Out  = feedforward(X);\n    Loss = tf.reduce_sum(tf.square(Y-Out));\n<\/code><\/pre>\n\n<p>It is <strong>automatic differentiation<\/strong> (kinda Monte Carlo method) instead of mathematical differentiation, and thus, all gradients obtained by the following function is already at their proper depths in backpropagation (just like the backward loop to calculate errors at all layers):<\/p>\n\n<pre><code>Grads = T.gradient(Loss,[W1,B1,W2,B2]);\n<\/code><\/pre>\n\n<p>After that, optimiser will apply gradients to change weights and biases:<\/p>\n\n<pre><code>Optim.apply_gradients(zip(Grads,[W1,B1,W2,B2]));\n<\/code><\/pre>\n"}]}
+{"QuestionId":58143041,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><multilabel-classification><multiclass-classification>","CreationDate":"2019-09-28T03:05:23.690","AcceptedAnswerId":58143318.0,"OwnerUserId":7310340.0,"Title":"\"NaN\" result when running multi class classification","Body":"<p>when i run these lines of code for binary classification it is running well without any problem and get a good result, but when i try to make it for many classes e.g 3 classes it give \"NaN\" in predict result    <\/p>\n\n<pre><code># Importing the Keras libraries and packages\n\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\n\n# Initialising the CNN\nclassifier = Sequential()\n\n# Step 1 - Convolution\nclassifier.add(Conv2D(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(Conv2D(32, (3, 3), activation = 'relu'))\nclassifier.add(MaxPooling2D(pool_size = (2, 2)))\n\nclassifier.add(Conv2enter code hereD(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\n# Step 3 - Flattening\nclassifier.add(Flatten())\n\n# Step 4 - Full connection\nclassifier.add(Dense(units = 128, activation = 'relu'))\nclassifier.add(Dense(units = 3, activation = 'sigmoid'))\n\n# Compiling the CNN\nclassifier.compile(optimizer = 'adam', loss = 'categorical_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('data\/train',\n                                                 target_size = (64, 64),\n                                                 batch_size = 32,\n                                                 class_mode = 'categorical')\n\ntest_set = test_datagen.flow_from_directory('data\/test',\n                                            target_size = (64, 64),\n                                            batch_size = 32,\n                                            class_mode = 'categorical')\n\nclassifier.fit_generator(training_set,\n                         steps_per_epoch = 240 ,\n                         epochs = 25,\n                         validation_data = test_set,\n                         validation_steps = 30)\n\n\n\nimport numpy as np\nfrom keras.preprocessing import image\ntest_image = image.load_img('2.jpeg', 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\n<\/code><\/pre>\n\n<p>i tried these lines of code with loss function \"binary\" with 2 classes it worked well without any problems and get a good result that helped me with my work and the accuracy approximately  '93%' .<br>\nbut my project based on multi class classification, so i tried to change the loss function to <code>'categorical_crossentropy'<\/code> and the class mod in <code>fit_generator<\/code> to <code>'categorical'<\/code> to make it multi class, the accuracy start with 60% and grows up to 99 and suddenly drop down to 33%.<\/p>\n\n<p>the expected result the labels of the classes \nthe actual result is \"NaN\".<\/p>\n\n<p>thanks in advance.<\/p>\n","answers":[{"AnswerId":"58143318","CreationDate":"2019-09-28T04:19:37.140","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>For multi-class classification, usually <code>softmax<\/code> is applied on the last dense layer instead of <code>sigmoid<\/code>. Change it to <code>softmax<\/code> to see whether the issue is still there. <\/p>\n"}]}
+{"QuestionId":58143119,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-28T03:29:37.637","AcceptedAnswerId":null,"OwnerUserId":4927466.0,"Title":"keras installed but can't be loaded","Body":"<p>I am new to using keras or tensorflow.<\/p>\n\n<p>I first used <code>conda install keras<\/code> in Anaconda Prompt to install keras. Then I go to Jupyter Notebook to call the keras using <code>import keras<\/code> or <code>from keras.datasets import mnist<\/code>. The error message showed as: No module named 'tensorflow'.<\/p>\n\n<p>I didn't know whether keras must have tensorflow installed, then it can be called. Then I used <code>conda install tensorflow<\/code> in Anaconda Prompt again. The result was like this:\n<a href=\"https:\/\/i.stack.imgur.com\/ZYfkp.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ZYfkp.jpg\" alt=\"enter image description here\"><\/a>  <\/p>\n\n<p>Then I rerun those statement to call keras again in Jupyter Notebook. Now the error message was: ImportError: DLL load failed: The specified module could not be found.Failed to load the native TensorFlow runtime.<\/p>\n\n<p>I didn't know what to do next to solve this problem to use keras.<\/p>\n\n<p>Thanks for your help.<\/p>\n","answers":[{"AnswerId":"58389830","CreationDate":"2019-10-15T07:57:30.817","ParentId":null,"OwnerUserId":"12192748","Title":null,"Body":"<p>As Keras requires backend framework like Tensorflow you need to install Tensorflow first and then install Keras which will solve the issue. <\/p>\n\n<p>After this you can either import keras from <strong>import tensorflow.keras<\/strong> or directly <strong>import Keras<\/strong><\/p>\n"}]}
+{"QuestionId":58143135,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2019-09-28T03:33:49.107","AcceptedAnswerId":58148886.0,"OwnerUserId":8462092.0,"Title":"Which Tensorflow Package for Keras 2.3","Body":"<p>I am trying to use Keras on my machine but I keep getting the error message \n\"AttributeError: module 'tensorflow.python.keras.backend' has no attribute 'get_graph'\". <\/p>\n\n<p>Stack Trace is:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\n AttributeError                            Traceback (most recent call last)\n&lt;ipython-input-1-dd13cebc6f04&gt; in &lt;module&gt;\n----&gt; 1 import acgan\n\n~\/Documents\/GitHub\/Keras-GAN\/acgan\/acgan.py in &lt;module&gt;\n  1 from __future__ import print_function, division\n  2 \n ----&gt; 3 from keras.datasets import mnist\n  4 from keras.layers import Input, Dense, Reshape, Flatten, Dropout, multiply\n  5 from keras.layers import BatchNormalization, Activation, Embedding, ZeroPadding2D\n\n\/anaconda3\/lib\/python3.6\/site-packages\/keras\/__init__.py in &lt;module&gt;\n  1 from __future__ import absolute_import\n  2 \n ----&gt; 3 from . import utils\n  4 from . import activations\n  5 from . import applications\n\n\/anaconda3\/lib\/python3.6\/site-packages\/keras\/utils\/__init__.py in &lt;module&gt;\n  4 from . import data_utils\n  5 from . import io_utils\n ----&gt; 6 from . import conv_utils\n  7 from . import losses_utils\n  8 from . import metrics_utils\n\n \/anaconda3\/lib\/python3.6\/site-packages\/keras\/utils\/conv_utils.py in &lt;module&gt;\n  7 from six.moves import range\n  8 import numpy as np\n  ----&gt; 9 from .. import backend as K\n 10 \n 11 \n\n \/anaconda3\/lib\/python3.6\/site-packages\/keras\/backend\/__init__.py in &lt;module&gt;\n ----&gt; 1 from .load_backend import epsilon\n  2 from .load_backend import set_epsilon\n  3 from .load_backend import floatx\n  4 from .load_backend import set_floatx\n  5 from .load_backend import cast_to_floatx\n\n \/anaconda3\/lib\/python3.6\/site-packages\/keras\/backend\/load_backend.py in &lt;module&gt;\n 88 elif _BACKEND == 'tensorflow':\n 89     sys.stderr.write('Using TensorFlow backend.\\n')\n ---&gt; 90     from .tensorflow_backend import *\n 91 else:\n 92     # Try and load external backend.\n\n \/anaconda3\/lib\/python3.6\/site-packages\/keras\/backend\/tensorflow_backend.py in &lt;module&gt;\n 52 \n 53 # Private TF Keras utils\n ---&gt; 54 get_graph = tf_keras_backend.get_graph\n 55 # learning_phase_scope = tf_keras_backend.learning_phase_scope  # \nTODO\n 56 name_scope = tf.name_scope\n\nAttributeError: module 'tensorflow.python.keras.backend' has no attribute 'get_graph'\n<\/code><\/pre>\n\n<p>I googled the issue and found this <\/p>\n\n<p><a href=\"https:\/\/forums.xilinx.com\/t5\/Deephi-DNNDK\/Module-tensorflow-python-keras-backend-has-no-attribute-get\/td-p\/1022555\" rel=\"nofollow noreferrer\">https:\/\/forums.xilinx.com\/t5\/Deephi-DNNDK\/Module-tensorflow-python-keras-backend-has-no-attribute-get\/td-p\/1022555<\/a> <\/p>\n\n<p>which tells me that it is tensorflow version related, but not which version to use. I am using Python3 on MacOS.<\/p>\n\n<p>Thank you for your help.<\/p>\n","answers":[{"AnswerId":"58148886","CreationDate":"2019-09-28T17:49:09.853","ParentId":null,"OwnerUserId":"5597718","Title":null,"Body":"<p>I couldn't reproduce the error you observe with Keras 2.3.0 &amp; tensorflow 2.0.0rc1. Although, I was able to start the training of acgan, I got a different error after one iteration: <code>Resource localhost\/_AnonymousVar84\/N10tensorflow3VarE does not exist<\/code>.<\/p>\n\n<p>I could successfully run the example with the following versions:<\/p>\n\n<ol>\n<li>Keras 2.2.4\/tensorflow 1.14.0<\/li>\n<li>tensorflow 2.0.0rc1 and replace <code>keras<\/code> with <code>tensorflow.keras<\/code> in the imports.<\/li>\n<\/ol>\n"}]}
+{"QuestionId":58143170,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2019-09-28T03:41:09.900","AcceptedAnswerId":null,"OwnerUserId":10520937.0,"Title":"Interfacing with tensorflow model results in an error","Body":"<p>I'm currently using DNN regressor with 11 input neurons multiple hidden layers, and 11 output neurons.\nI've have finished training the model. However, when I try to interface with the model with this code:<\/p>\n\n<pre><code> a = np.array(0,0,0,0,0,0,0,0,0,0,0) \n    y_pred = regressor.predict(a, as_iterable=False) \n    print(y_pred)\n<\/code><\/pre>\n\n<p>However this generates an error:<\/p>\n\n<pre><code>InvalidArgumentError (see above for traceback): Input to reshape is a tensor with 11 values, but the requested shape has 121\n         [[node dnn\/input_from_feature_columns\/input_layer\/X\/Reshape (defined at regressor_full.py:177) ]]\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"regressor_full.py\", line 206, in &lt;module&gt;\n    a = np.array(0,0,0,0,0,0,0,0,0,0,0)\nValueError: only 2 non-keyword arguments accepted.\n<\/code><\/pre>\n\n<p>I'm just trying to generate a prediction for 11 input variables and predict 11 output variables(as I trained the model), however, it generates an error.\nI've tried changing the as_iterable to True, however this doesn't change anything. What is currently causing the error, and how do I fix it? Thank you for your time.<\/p>\n\n<p>As requested, I included the code that defines the model. <\/p>\n\n<pre><code>import tensorflow.contrib.learn as skflow\n\nregressor = skflow.DNNRegressor(feature_columns=feature_columns,\n                label_dimension=11,\n                hidden_units=hidden_layers,\n                model_dir=MODEL_PATH,\n                dropout=dropout,\n                config=test_config,\n                activation_fn = tf.nn.relu,\n                optimizer  = tf.compat.v1.train.AdamOptimizer(learning_rate = learning_rate)\n)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58389740","CreationDate":"2019-10-15T07:51:46.857","ParentId":null,"OwnerUserId":"12192748","Title":null,"Body":"<p>Try changing the input array to <strong>a = np.array([0,0,0,0,0,0,0,0,0,0,0],dtype=\" \")<\/strong> and also define the dtype which might solve the issue. <\/p>\n"}]}
+{"QuestionId":58143637,"AnswerCount":2,"Tags":"<tensorflow><nvidia><tensorflow2.0><cudnn>","CreationDate":"2019-09-28T05:29:08.740","AcceptedAnswerId":null,"OwnerUserId":11129589.0,"Title":"Tensorflow 2.0 can't use GPU, something wrong in cuDNN? :Failed to get convolution algorithm. This is probably because cuDNN failed to initialize","Body":"<p><strong>I tried to predict with my tf2.0\/tf.keras CNN model on GPU, but get those error messages.\ncould someone help me to fix it?<\/strong><\/p>\n\n<p>**enviroments:<\/p>\n\n<p>python 3.6.8<\/p>\n\n<p>tensorflow-gpu 2.0.0-rc0<\/p>\n\n<p>nvidia 418.x<\/p>\n\n<p>CUDA 10.0<\/p>\n\n<p>cuDNN 7.6+**<\/p>\n\n<p>===========================================================<\/p>\n\n<p><strong>LOGs:<\/strong><\/p>\n\n<p>2019-09-28 13:10:59.833892: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:44] Successfully opened dynamic library libcublas.so.10.0\n2019-09-28 13:11:00.228025: I tensorflow\/stream_executor\/platform\/default\/dso_loader.cc:44] Successfully opened dynamic library libcudnn.so.7\n2019-09-28 13:11:00.957534: E tensorflow\/stream_executor\/cuda\/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR\n2019-09-28 13:11:00.963310: E tensorflow\/stream_executor\/cuda\/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR\n2019-09-28 13:11:00.963416: W tensorflow\/core\/common_runtime\/base_collective_executor.cc:216] BaseCollectiveExecutor::StartAbort Unknown: 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 mobilenetv2_1.00_192\/Conv1\/Conv2D}}]]\nmobilenetv2_1.00_192\/block_15_expand_BN\/cond\/then\/_630\/Const: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0=====>GPU Available:  True\n=====> 4 Physical GPUs, 1 Logical GPUs<\/p>\n\n<p>mobilenetv2_1.00_192\/block_15_expand_BN\/cond\/then\/_630\/Const_1: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nmobilenetv2_1.00_192\/block_15_depthwise_BN\/cond\/then\/_644\/Const: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nmobilenetv2_1.00_192\/block_15_depthwise_BN\/cond\/then\/_644\/Const_1: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nmobilenetv2_1.00_192\/block_15_project_BN\/cond\/then\/_658\/Const: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nmobilenetv2_1.00_192\/block_15_project_BN\/cond\/then\/_658\/Const_1: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nmobilenetv2_1.00_192\/block_16_expand_BN\/cond\/then\/_672\/Const: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nmobilenetv2_1.00_192\/block_16_expand_BN\/cond\/then\/_672\/Const_1: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nmobilenetv2_1.00_192\/block_16_depthwise_BN\/cond\/then\/_686\/Const: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nmobilenetv2_1.00_192\/block_16_depthwise_BN\/cond\/then\/_686\/Const_1: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nmobilenetv2_1.00_192\/block_16_project_BN\/cond\/then\/_700\/Const: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nmobilenetv2_1.00_192\/block_16_project_BN\/cond\/then\/_700\/Const_1: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nmobilenetv2_1.00_192\/Conv_1_bn\/cond\/then\/_714\/Const: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nmobilenetv2_1.00_192\/Conv_1_bn\/cond\/then\/_714\/Const_1: (Const): \/job:localhost\/replica:0\/task:0\/device:GPU:0\nTraceback (most recent call last):\n  File \"NSFW_Server.py\", line 162, in \n    model.predict(initial_tensor)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow_core\/python\/keras\/engine\/training.py\", line 915, in predict\n    use_multiprocessing=use_multiprocessing)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow_core\/python\/keras\/engine\/training_arrays.py\", line 722, in predict\n    callbacks=callbacks)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow_core\/python\/keras\/engine\/training_arrays.py\", line 393, in model_iteration\n    batch_outs = f(ins_batch)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow_core\/python\/keras\/backend.py\", line 3625, in <strong>call<\/strong>\n    outputs = self._graph_fn(*converted_inputs)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow_core\/python\/eager\/function.py\", line 1081, in <strong>call<\/strong>\n    return self._call_impl(args, kwargs)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow_core\/python\/eager\/function.py\", line 1121, in _call_impl\n    return self._call_flat(args, self.captured_inputs, cancellation_manager)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow_core\/python\/eager\/function.py\", line 1224, in _call_flat\n    ctx, args, cancellation_manager=cancellation_manager)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow_core\/python\/eager\/function.py\", line 511, in call\n    ctx=ctx)\n  File \"\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow_core\/python\/eager\/execute.py\", line 67, in quick_execute\n    six.raise_from(core._status_to_exception(e.code, message), None)\n  File \"\", line 3, in raise_from\ntensorflow.python.framework.errors_impl.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 mobilenetv2_1.00_192\/Conv1\/Conv2D (defined at \/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow_core\/python\/framework\/ops.py:1751) ]] [Op:__inference_keras_scratch_graph_10727]<\/p>\n\n<p>Function call stack:\nkeras_scratch_graph<\/p>\n\n<p>===========================================================<\/p>\n\n<p><strong>RELATED CODE:<\/strong><\/p>\n\n<p>if <strong>name<\/strong> == \"<strong>main<\/strong>\":<\/p>\n\n<pre><code>print(\"=====&gt;GPU Available: \", tf.test.is_gpu_available())\ntf.debugging.set_log_device_placement(True)\n\ngpus = tf.config.experimental.list_physical_devices('GPU')\nif gpus:\n    try:\n        # Currently, memory growth needs to be the same across GPUs\n\n        tf.config.experimental.set_visible_devices(gpus[0], 'GPU')\n        tf.config.experimental.set_memory_growth(gpus[0], True)\n        logical_gpus = tf.config.experimental.list_logical_devices('GPU')\n        print(\"=====&gt;\", len(gpus), \"Physical GPUs,\", len(logical_gpus), \"Logical GPUs\")\n    except RuntimeError as e:\n        # Memory growth must be set before GPUs have been initialized\n        print(e)\n\nparas_path = \".\/paras\/{}\".format(int(2011))\nmodel = tf.keras.experimental.load_from_saved_model(paras_path)\ninitial_tensor = np.zeros((1, INPUT_SHAPE, INPUT_SHAPE, 3))\nmodel.predict(initial_tensor)\n<\/code><\/pre>\n\n<p>===========================================================<\/p>\n\n<p><strong>pls help, it bothers me couple days...<\/strong><\/p>\n","answers":[{"AnswerId":"58456367","CreationDate":"2019-10-18T18:27:57.457","ParentId":null,"OwnerUserId":"11002096","Title":null,"Body":"<p>Check the instructions on this TensorFlow GPU instruction <a href=\"https:\/\/www.tensorflow.org\/install\/gpu#ubuntu_1604_cuda_90_for_tensorflow_1130\" rel=\"nofollow noreferrer\">page<\/a> for your OS. It resolved issue for me on Ubuntu 16.04.6 LTS and Tensorflow 2.0<\/p>\n"},{"AnswerId":"58143787","CreationDate":"2019-09-28T05:58:36.413","ParentId":null,"OwnerUserId":"6117017","Title":null,"Body":"<p>You have to check that you have the right version of CUDA + CUDNN + TensorFlow(also ensure that you have all installed.<\/p>\n\n<p>Two examples of running configurations are presented below<\/p>\n\n<p>Eg. Cuda 10.0 + CuDNN 7.6.3  + \/ TensorFlow 1.13\/1.14 \/ TensorFlow 2.0.<\/p>\n\n<p>Eg2 Cuda 9 + CuDNN 7.0.5 + TensorFlow 1.10 works<\/p>\n\n<p>Usually this error appears when you have an incompatible version of TensorFlow\/CuDNN installed. In my case, this appeared when I tried using an older TensorFlow with a newer version of CuDNN.<\/p>\n"}]}
+{"QuestionId":58143742,"AnswerCount":2,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2019-09-28T05:47:30.740","AcceptedAnswerId":null,"OwnerUserId":9126289.0,"Title":"problems importing tensorflow","Body":"<p>I'm working on a project using keras (python 3), and I've encountered a problem - I've installed using pip tensorflow, and imported it into my prject, but whenether I try to run it, I get an error saying:<\/p>\n\n<pre><code>ModuleNotFoundError: No module named 'tensorflow'\n<\/code><\/pre>\n\n<p>it seems my installation completed successfully, and I think I have the right PATH since I installed few other things such as numpy and their installation worked well. does anyone have a clue what did I do wrong? <\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"58408273","CreationDate":"2019-10-16T07:47:18.887","ParentId":null,"OwnerUserId":"12199329","Title":null,"Body":"<p>Faced similar issue when installing Tensorflow, following might the reasons:<\/p>\n\n<ul>\n<li>Installing IDLE before Anaconda.<\/li>\n<li><p>Incorrect path variable while installing.<\/p>\n\n<p><strong>Solution for this is<\/strong>,<\/p><\/li>\n<li>Uninstall Anaconda and IDLE(if installed)<\/li>\n<li>Setup Anaconda and add path variable while installing<\/li>\n<li>Create environment in conda prompt <code>conda create -n env_name python=3.7<\/code><\/li>\n<li>Activate environment <code>conda activate env_name<\/code><\/li>\n<li>Install Pandas, Tensorflow and Keras. (Other required libraries come with env)<\/li>\n<\/ul>\n"},{"AnswerId":"58143778","CreationDate":"2019-09-28T05:57:08.840","ParentId":null,"OwnerUserId":"12102176","Title":null,"Body":"<p>secondary create backup for itenary such as vm <\/p>\n"}]}
+{"QuestionId":58144009,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><keras>","CreationDate":"2019-09-28T06:41:21.983","AcceptedAnswerId":null,"OwnerUserId":11493507.0,"Title":"How to get the gradients in Keras?","Body":"<p>I want to get the gradients computed in every layer in Keras, and I have implemented the solution described <a href=\"https:\/\/stackoverflow.com\/questions\/51140950\/how-to-obtain-the-gradients-in-keras\/51144278\">here<\/a>, using the TensorFlow 2.0 version of keras, which as of now is <code>2.2.4-tf<\/code>.<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nfrom tensorflow.keras.models import Sequential\nfrom tensorflow.keras import backend as K\n\nmodel = keras.Sequential()\nmodel.add(keras.layers.Dense(20, input_shape = (10, )))\nmodel.add(keras.layers.Dense(5))\nmodel.compile('adam', 'mse')\n\ndummy_in = np.ones((4, 10))\ndummy_out = np.ones((4, 5))\ndummy_loss = model.train_on_batch(dummy_in, dummy_out)\n\ndef get_weight_grad(model, inputs, outputs):\n    \"\"\" Gets gradient of model for given inputs and outputs for all weights\"\"\"\n    grads = model.optimizer.get_gradients(model.total_loss, model.trainable_weights)\n    symb_inputs = (model._feed_inputs + model._feed_targets + model._feed_sample_weights)\n    f = K.function(symb_inputs, grads)\n    x, y, sample_weight = model._standardize_user_data(inputs, outputs)\n    output_grad = f(x + y + sample_weight)\n    return output_grad\n\n\ndef get_layer_output_grad(model, inputs, outputs, layer=-1):\n    \"\"\" Gets gradient a layer output for given inputs and outputs\"\"\"\n    grads = model.optimizer.get_gradients(model.total_loss, model.layers[layer].output)\n    symb_inputs = (model._feed_inputs + model._feed_targets + model._feed_sample_weights)\n    f = K.function(symb_inputs, grads)\n    x, y, sample_weight = model._standardize_user_data(inputs, outputs)\n    output_grad = f(x + y + sample_weight)\n    return output_grad\n\n\nweight_grads = get_weight_grad(model, dummy_in, dummy_out)\noutput_grad = get_layer_output_grad(model, dummy_in, dummy_out)\n<\/code><\/pre>\n\n<p>The problem is that the <code>model._standardize_user_data(dummy_in, dummy_out)<\/code> function returns a <code>NoneType<\/code> for <code>sample_weight<\/code>, thus, I cannot concateneta it with <code>x<\/code> and <code>y<\/code> in <code>f(x + y + sample_weight)<\/code>. I get a <code>TypeError: can only concatenate list (not \"NoneType\") to list<\/code>.<\/p>\n","answers":[]}
+{"QuestionId":58144274,"AnswerCount":0,"Tags":"<python><tensorflow><mnist>","CreationDate":"2019-09-28T07:28:43.657","AcceptedAnswerId":null,"OwnerUserId":12131427.0,"Title":"TensorFlow EOFError: Compressed file ended before the end-of-stream marker was reached","Body":"<p>I am new to Tensorflow. I am making the code listed bellow, but facing the following problem when running it:<\/p>\n\n<blockquote>\n  <p>EOFError: Compressed file ended before the end-of-stream marker was\n  reached<\/p>\n<\/blockquote>\n\n<p>Could anyone help me understand why this occurs ?<\/p>\n\n<p>The code is below:<\/p>\n\n<pre><code>import tensorflow as tf\n\nprint(tf.__version__)\nmnist = tf.keras.datasets.fashion_mnist\n(training_images, training_labels), (test_images, test_labels)= mnist.load_data()\nimport matplotlib.pyplot as plt\nplt.imshow(training_images[0])\nprint(training_labels[0])\nprint(training_images[0])\ntraining_images = training_images \/ 255.0\ntest_images = test_images \/ 255.0\nmodel = tf.keras.models.Sequential([tf.keras.layers.Flatten(),\n                                    tf.keras.layers.Dense(128, activation=tf.nn.relu),\n                                    tf.keras.layers.Dense(10, activation=tf.nn.softmax)])\nmodel.compile(optimizer=tf.train.AdamOPtimizer(),\n              loss= 'sparse_categorical_crossentropy',\n              metrics= ['accuracy'])\nmodel.fit(training_images, training_labels, epochs=5)\nmodel.evaluate(test_images, test_labels)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58144336,"AnswerCount":0,"Tags":"<keras><lstm><masking>","CreationDate":"2019-09-28T07:40:12.940","AcceptedAnswerId":null,"OwnerUserId":8278303.0,"Title":"How to mask the inputs in an LSTM autoencoder having a RepeatVector() layer?","Body":"<p>I have been trying to obtaining a vector representation of a sequence of vectors using an LSTM autoencoder so that I can classify the sequence using a SVM or other such supervised algorithms. The amount of data is preventing me from using a fully connected dense layer for classification.<\/p>\n\n<p>The shortest size of my input is 7 timesteps and the longest sequence is 356 timesteps. Accordingly, I have padded the shorter sequences with zeros to obtain a final x_train of shape (1326, 356, 8) where 1326 is the number of training samples and 8 is the dimension of one timestep. I am trying to encode these sequences into a single vector using the given LSTM autoencoder.<\/p>\n\n<pre><code>model.add(Masking(mask_value=0.0, input_shape=(max_len, 8)))\nmodel.add(LSTM(100, activation='relu'))\nmodel.add(RepeatVector(max_len))\nmodel.add(LSTM(8, activation='relu', return_sequences=True))\nmodel.compile(optimizer='adam', loss='mse')\nmodel.fit(x_train, x_train, batch_size=32, callbacks=[chk], epochs=1000, validation_split=0.05, shuffle=True)\n<\/code><\/pre>\n\n<p>I am trying to mask the zero padded results but the RepeatVector() layer may be hindering the process. Hence, after sometime the mean square error loss is becoming <code>nan<\/code>. Can anyone help me out as to how I can only include the relevant timestep in calculating the loss function and ignore the other timesteps?<\/p>\n","answers":[]}
+{"QuestionId":58145301,"AnswerCount":0,"Tags":"<tensorflow><keras><prediction><tensorflow-datasets><loss-function>","CreationDate":"2019-09-28T10:13:45.540","AcceptedAnswerId":null,"OwnerUserId":12134411.0,"Title":"Keras prediction: is it possible to access an element of the prediction vector and return it as scalar?","Body":"<p>I am using Keras for a DNN for a resource allocation problem, where the dimension of the output of the DNN is equal to the number of resources (let's say there are R resources). Each of the R output nodes contains the user that a resource is assigned to (i.e., having 10 users, each output node takes an integer value between 1 and 10). In the loss function I want to access these integers but I am not able to do that.\nMy loss function is the following, where y_pred is the output of the DNN:<\/p>\n\n<pre><code># Define custom loss\ndef custom_loss(num_outputs):\n\n    # Create a loss function \n    def loss(y_true,y_pred):\n        for k in range(num_outputs\/2):            \n            user_k = y_pred[0][k]\n            user_k_integer = int(user_k)\n\n#... return loss\n<\/code><\/pre>\n\n<p>When I try to compile the model, I get the following error message:\nTypeError: int() argument must be a string or a number, not 'Tensor'<\/p>\n\n<p>What I want, is the value stored in y_pred[0][k] as a scalar and use it then as an integer for indexing.<\/p>\n\n<p>Anybody has an idea how to solve my problem?<\/p>\n\n<p>Many thanks in advance!<\/p>\n","answers":[]}
+{"QuestionId":58145727,"AnswerCount":1,"Tags":"<python><pytorch><freeze><pruning>","CreationDate":"2019-09-28T11:10:27.807","AcceptedAnswerId":58154243.0,"OwnerUserId":8333506.0,"Title":"Freezing Individual Weights in Pytorch","Body":"<p>The following question is <strong>not a duplicate<\/strong> of <a href=\"https:\/\/stackoverflow.com\/questions\/51801648\/how-to-apply-layer-wise-learning-rate-in-pytorch\">How to apply layer-wise learning rate in Pytorch?<\/a> because this question aims at freezing a subset of a tensor from training rather than the entire layer.<\/p>\n\n<p>I am trying out a PyTorch implementation of Lottery Ticket Hypothesis. <\/p>\n\n<p>For that, I want to freeze the weights in a model that are zero. Is the following a correct way to implement it?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>for name, p in model.named_parameters():\n            if 'weight' in name:\n                tensor = p.data.cpu().numpy()\n                grad_tensor = p.grad.data.cpu().numpy()\n                grad_tensor = np.where(tensor == 0, 0, grad_tensor)\n                p.grad.data = torch.from_numpy(grad_tensor).to(device)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58154243","CreationDate":"2019-09-29T10:12:00.267","ParentId":null,"OwnerUserId":"2790047","Title":null,"Body":"<p>What you have seems like it would work provided you did it after <code>loss.backward()<\/code> and before <code>optimizer.step()<\/code> (referring to the common usage for these variable names). That said, it seems a bit convoluted. Also, if your weights are floating point values then <a href=\"https:\/\/stackoverflow.com\/questions\/588004\/is-floating-point-math-broken\">comparing them to exactly zero is probably a bad idea<\/a>, we could introduce an epsilon to account for this. IMO the following is a little cleaner than the solution you proposed:<\/p>\n\n<pre><code># locate zero-value weights before training loop\nEPS = 1e-6\nlocked_masks = {n: torch.abs(w) &lt; EPS for n, w in model.named_parameters() if n.endswith('weight')}\n\n...\n\nfor ... #training loop\n\n    ...\n\n    optimizer.zero_grad()\n    loss.backward()\n    # zero the gradients of interest\n    for n, w in model.named_parameters():                                                                                                                                                                           \n        if w.grad is not None and n in locked_masks:                                                                                                                                                                                   \n            w.grad[locked_masks[n]] = 0 \n    optimizer.step()\n<\/code><\/pre>\n"}]}
+{"QuestionId":58145774,"AnswerCount":1,"Tags":"<python><tensorflow><object-detection-api><tensorflow-lite><toco>","CreationDate":"2019-09-28T11:16:59.383","AcceptedAnswerId":null,"OwnerUserId":9786502.0,"Title":"Can't convert Frozen Inference Graph to .tflite","Body":"<p>I am new to the object detection API and TensorFlow in general. I followed <a href=\"https:\/\/www.youtube.com\/watch?v=HjiBbChYRDw&amp;t=1248s\" rel=\"nofollow noreferrer\">this<\/a> tutorial and in the end I produced a <code>frozen_inference_graph.pb<\/code>. I want to run this object detection model on my phone, which in my understanding requires me to convert it to .tflite (please lmk if this doesn't make any sense).<\/p>\n\n<p>When I tried to convert it using this standard code here:<\/p>\n\n<pre><code>import tensorflow as tf\ngraph = 'pathtomygraph'\ninput_arrays = ['image_tensor']\noutput_arrays = ['all_class_predictions_with_background']\n\n\nconverter = tf.lite.TFLiteConverter.from_frozen_graph(graph, input_arrays, output_arrays)\ntflite_model = converter.convert()\nopen(\"converted_model.tflite\", \"wb\").write(tflite_model)\n\n<\/code><\/pre>\n\n<p>It throws an error, saying: <\/p>\n\n<blockquote>\n  <p>ValueError: None is only supported in the 1st dimension. Tensor\n  'image_tensor' has invalid shape '[None, None, None, 3]'<\/p>\n<\/blockquote>\n\n<p>This is a common error I found on the internet, and after searching through many threads, I tried to give an extra parameter to the code: <\/p>\n\n<pre><code>converter = tf.lite.TFLiteConverter.from_frozen_graph(\n        graph, input_arrays, output_arrays,input_shapes={\"image_tensor\":[1,600,600,3]})\n<\/code><\/pre>\n\n<p>Now it looks like this:<\/p>\n\n<pre><code>import tensorflow as tf\ngraph = 'pathtomygraph'\ninput_arrays = ['image_tensor']\noutput_arrays = ['all_class_predictions_with_background']\n\n\nconverter = tf.lite.TFLiteConverter.from_frozen_graph(\n        graph, input_arrays, output_arrays,input_shapes={\"image_tensor\":[1,600,600,3]})\ntflite_model = converter.convert()\nopen(\"converted_model.tflite\", \"wb\").write(tflite_model)\n\n<\/code><\/pre>\n\n<p>This <strong>works<\/strong> at first, but throws another error at the end, saying: <\/p>\n\n<blockquote>\n  <p>Check failed: array.data_type == array.final_data_type Array\n  \"image_tensor\" has mis-matching actual and final data types\n  (data_type=uint8, final_data_type=float). Fatal Error: Aborted<\/p>\n<\/blockquote>\n\n<p>I understand that my input tensor has the data type of uint8 and this causes a mismatch, I guess. My question would be, is this the correct way to approach things? (I want to run my model on my phone). If it is, how do I then fix the error? :\/<\/p>\n\n<p>Thank you very much.<\/p>\n","answers":[{"AnswerId":"58159666","CreationDate":"2019-09-29T22:23:21.317","ParentId":null,"OwnerUserId":"834565","Title":null,"Body":"<p>Change your model input (the <code>image_tensor<\/code> placeholder) to have data type <code>tf.float32<\/code>.<\/p>\n"}]}
+{"QuestionId":58146063,"AnswerCount":0,"Tags":"<python><tensorflow><object-detection><object-detection-api>","CreationDate":"2019-09-28T11:54:37.293","AcceptedAnswerId":null,"OwnerUserId":9109572.0,"Title":"SSD-mobilenet class score and bbox coordinates computing","Body":"<p>I\u2019m implementing an object detection task using <strong>SSD-mobilenet<\/strong>,\nIm ok with the results, but i want to understand more about the theoretical principle.<\/p>\n\n<p>How does <strong>SSD-monilenet<\/strong> computes the  the confidence score and the bounding box coordinates, referring to <strong>SSD paper<\/strong>, it said that this process is done with the aid of <strong>small convolutional Kernels.<\/strong> With no more info about the way. \nCan someone explain how this process is done with more clarification?<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[]}
+{"QuestionId":58146169,"AnswerCount":0,"Tags":"<python><tensorflow><lstm><tensorflow-serving><tensorflow-lite>","CreationDate":"2019-09-28T12:07:58.350","AcceptedAnswerId":null,"OwnerUserId":8087807.0,"Title":"How to deploy a LSTM tensorflow model which uses native tensorflow functions","Body":"<p>I came across a tensorflow model for OMR (Optical Music Recognition). This model performs very well on this task and I would like to make an Android app that would use this model. My question is: Is it possible to run this model on the android device or do I need to make an API for this?<\/p>\n\n<p>I've tried converting the model to tensorflow-lite. But the model is quite complicated and there doesn't seem to be equivalent in tensorflow-lite to native tensorflow functions used here.<\/p>\n\n<p>The model is now stored as meta graph. So the loading goes like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Restore weights\nsaver = tf.train.import_meta_graph(args.model)\nsaver.restore(sess,args.model[:-5])\n\ngraph = tf.get_default_graph()\n<\/code><\/pre>\n\n<p>This is the code for inference:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>input = graph.get_tensor_by_name(\"model_input:0\")\nseq_len = graph.get_tensor_by_name(\"seq_lengths:0\")\nrnn_keep_prob = graph.get_tensor_by_name(\"keep_prob:0\")\nheight_tensor = graph.get_tensor_by_name(\"input_height:0\")\nwidth_reduction_tensor = graph.get_tensor_by_name(\"width_reduction:0\")\nlogits = tf.get_collection(\"logits\")[0]\n\n# Constants that are saved inside the model itself\nWIDTH_REDUCTION, HEIGHT = sess.run([width_reduction_tensor, height_tensor])\n\ndecoded, _ = tf.nn.ctc_greedy_decoder(logits, seq_len)\n\nimage = cv2.imread(args.image,False)\nimage = ctc_utils.resize(image, HEIGHT)\nimage = ctc_utils.normalize(image)\nimage = np.asarray(image).reshape(1,image.shape[0],image.shape[1],1)\n\nseq_lengths = [ image.shape[2] \/ WIDTH_REDUCTION ]\n\nprediction = sess.run(decoded,\n                      feed_dict={\n                          input: image,\n                          seq_len: seq_lengths,\n                          rnn_keep_prob: 1.0,\n                      })\n<\/code><\/pre>\n\n<p>this is the model definition:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>input = tf.placeholder(shape=(None,\n                                   params['img_height'],\n                                   params['img_width'],\n                                   params['img_channels']),  # [batch, height, width, channels]\n                            dtype=tf.float32,\n                            name='model_input')\n\n    input_shape = tf.shape(input)\n\n    width_reduction = 1\n    height_reduction = 1\n\n\n    # Convolutional blocks\n    x = input\n    for i in range(params['conv_blocks']):\n\n        x = tf.layers.conv2d(\n            inputs=x,\n            filters=params['conv_filter_n'][i],\n            kernel_size=params['conv_filter_size'][i],\n            padding=\"same\",\n            activation=None)\n\n        x = tf.layers.batch_normalization(x)\n        x = leaky_relu(x)\n\n        x = tf.layers.max_pooling2d(inputs=x,\n                                    pool_size=params['conv_pooling_size'][i],\n                                    strides=params['conv_pooling_size'][i])\n\n        width_reduction = width_reduction * params['conv_pooling_size'][i][1]\n        height_reduction = height_reduction * params['conv_pooling_size'][i][0]\n\n\n    # Prepare output of conv block for recurrent blocks\n    features = tf.transpose(x, perm=[2, 0, 3, 1])  # -&gt; [width, batch, height, channels] (time_major=True)\n    feature_dim = params['conv_filter_n'][-1] * (params['img_height'] \/ height_reduction)\n    feature_width = input_shape[2] \/ width_reduction\n    features = tf.reshape(features, tf.stack([tf.cast(feature_width,'int32'), input_shape[0], tf.cast(feature_dim,'int32')]))  # -&gt; [width, batch, features]\n\n    tf.constant(params['img_height'],name='input_height')\n    tf.constant(width_reduction,name='width_reduction')\n\n    # Recurrent block\n    rnn_keep_prob = tf.placeholder(dtype=tf.float32, name=\"keep_prob\")\n    rnn_hidden_units = params['rnn_units']\n    rnn_hidden_layers = params['rnn_layers']\n\n    rnn_outputs, _ = tf.nn.bidirectional_dynamic_rnn(\n        tf.contrib.rnn.MultiRNNCell(\n            [tf.nn.rnn_cell.DropoutWrapper(tf.contrib.rnn.BasicLSTMCell(rnn_hidden_units), input_keep_prob=rnn_keep_prob)\n             for _ in range(rnn_hidden_layers)]),\n        tf.contrib.rnn.MultiRNNCell(\n            [tf.nn.rnn_cell.DropoutWrapper(tf.contrib.rnn.BasicLSTMCell(rnn_hidden_units), input_keep_prob=rnn_keep_prob)\n             for _ in range(rnn_hidden_layers)]),\n        features,\n        dtype=tf.float32,\n        time_major=True,\n    )\n\n    rnn_outputs = tf.concat(rnn_outputs, 2)\n\n    logits = tf.contrib.layers.fully_connected(\n        rnn_outputs,\n        params['vocabulary_size'] + 1,  # BLANK\n        activation_fn=None,\n    )\n\n<\/code><\/pre>\n\n<p>The whole source code can be found at: <a href=\"https:\/\/github.com\/calvozaragoza\/tf-deep-omr\" rel=\"nofollow noreferrer\">https:\/\/github.com\/calvozaragoza\/tf-deep-omr<\/a><\/p>\n","answers":[]}
+{"QuestionId":58146801,"AnswerCount":0,"Tags":"<python><tensorflow><keras><deep-learning>","CreationDate":"2019-09-28T13:33:40.780","AcceptedAnswerId":null,"OwnerUserId":1745048.0,"Title":"Tensorflow model, 'tuple' object has no attribute 'layer'","Body":"<p>I am trying to build a siamese network model. There whole model takes in two inputs. The two inputs will be passed into two twin network (in other words, a shared network). And their outputs will be compared in a distance layer, which then connects to an output layer.<\/p>\n\n<p>The error seems not happy about how I pass an input layer to the shared network. Here is the code for building the model<\/p>\n\n<pre><code>def get_siamese_model(input_shape):\n  \"\"\"\n        Model architecture based on the one provided in: http:\/\/www.cs.utoronto.ca\/~gkoch\/files\/msc-thesis.pdf\n  \"\"\"\n  #2 inputs\n  left_input = Input(input_shape)\n  right_input = Input(input_shape)\n\n  #currently not convolutional NN\n  model = tf.keras.models.Sequential([\n    tf.keras.layers.Flatten(input_shape=input_shape),\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\n  #encoding layer\n  encoded_l = model(left_input)\n  encoded_r = model(right_input)\n\n  #distance layer\n  L1_layer = Lambda(lambda tensors: K.abs(tensors[0]- tensors[1]))\n  L1_distance = L1_layer([encoded_l,encoded_r])\n\n  #prediction layer\n  prediction = Dense(1, activation = 'sigmoid', bias_initializer = initialize_bias)(L1_distance)\n\n  #connect the inputs with the outputs\n  siamese_net = Model(input=[left_input,right_input], outputs = prediction)\n\n  return siamese_net\n\nmodel = get_siamese_model((28, 28,1))\nmodel.summary()\n<\/code><\/pre>\n\n<p>And here is the full stack of errors:<\/p>\n\n<pre><code>---------------------------------------------------------------------------\nAttributeError                            Traceback (most recent call last)\n&lt;ipython-input-43-655eb0c236b9&gt; in &lt;module&gt;()\n----&gt; 1 model = get_siamese_model((28, 28,1))\n      2 model.summary()\n\n10 frames\n&lt;ipython-input-42-af8702cd4619&gt; in get_siamese_model(input_shape)\n     33   #                 kernel_initializer=initialize_weights,bias_initializer=initialize_bias))\n     34   #encoding layer\n---&gt; 35   encoded_l = model(left_input)\n     36   encoded_r = model(right_input)\n     37 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in __call__(self, inputs, *args, **kwargs)\n    632                     outputs = base_layer_utils.mark_as_return(outputs, acd)\n    633                 else:\n--&gt; 634                   outputs = call_fn(inputs, *args, **kwargs)\n    635 \n    636             except TypeError as e:\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/sequential.py in call(self, inputs, training, mask)\n    245       if not self.built:\n    246         self._init_graph_network(self.inputs, self.outputs, name=self.name)\n--&gt; 247       return super(Sequential, self).call(inputs, training=training, mask=mask)\n    248 \n    249     outputs = inputs  # handle the corner case where self.layers is empty\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/network.py in call(self, inputs, training, mask)\n    749                                 ' implement a `call` method.')\n    750 \n--&gt; 751     return self._run_internal_graph(inputs, training=training, mask=mask)\n    752 \n    753   def compute_output_shape(self, input_shape):\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/network.py in _run_internal_graph(self, inputs, training, mask)\n    891 \n    892           # Compute outputs.\n--&gt; 893           output_tensors = layer(computed_tensors, **kwargs)\n    894 \n    895           # Update tensor_dict.\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in __call__(self, inputs, *args, **kwargs)\n    661               kwargs.pop('training')\n    662             inputs, outputs = self._set_connectivity_metadata_(\n--&gt; 663                 inputs, outputs, args, kwargs)\n    664           self._handle_activity_regularization(inputs, outputs)\n    665           self._set_mask_metadata(inputs, outputs, previous_mask)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in _set_connectivity_metadata_(self, inputs, outputs, args, kwargs)\n   1706     kwargs.pop('mask', None)  # `mask` should not be serialized.\n   1707     self._add_inbound_node(\n-&gt; 1708         input_tensors=inputs, output_tensors=outputs, arguments=kwargs)\n   1709     return inputs, outputs\n   1710 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in _add_inbound_node(self, input_tensors, output_tensors, arguments)\n   1793     \"\"\"\n   1794     inbound_layers = nest.map_structure(lambda t: t._keras_history.layer,\n-&gt; 1795                                         input_tensors)\n   1796     node_indices = nest.map_structure(lambda t: t._keras_history.node_index,\n   1797                                       input_tensors)\n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/util\/nest.py in map_structure(func, *structure, **kwargs)\n    513 \n    514   return pack_sequence_as(\n--&gt; 515       structure[0], [func(*x) for x in entries],\n    516       expand_composites=expand_composites)\n    517 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/util\/nest.py in &lt;listcomp&gt;(.0)\n    513 \n    514   return pack_sequence_as(\n--&gt; 515       structure[0], [func(*x) for x in entries],\n    516       expand_composites=expand_composites)\n    517 \n\n\/usr\/local\/lib\/python3.6\/dist-packages\/tensorflow\/python\/keras\/engine\/base_layer.py in &lt;lambda&gt;(t)\n   1792             `call` method of the layer at the call that created the node.\n   1793     \"\"\"\n-&gt; 1794     inbound_layers = nest.map_structure(lambda t: t._keras_history.layer,\n   1795                                         input_tensors)\n   1796     node_indices = nest.map_structure(lambda t: t._keras_history.node_index,\n\nAttributeError: 'tuple' object has no attribute 'layer'\n<\/code><\/pre>\n\n<p>An post which might seem relevant talks about I may need to downgrade my tensorflow. I haven't tried that because I want to use the latest version. I am wondering if I am doing something wrong while building the model.<\/p>\n","answers":[]}
+{"QuestionId":58146901,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-28T13:45:12.473","AcceptedAnswerId":58149166.0,"OwnerUserId":10192485.0,"Title":"How to use only max\/average pooling layer in Tensorflow for 1D array?","Body":"<p>I have a .wav file which I read into an array in python. <\/p>\n\n<pre><code>import wave as wav\npath = \"Casio-Celesta-C5.wav\"\nf = wav.open(path)\ndata_sound = f.readframes(-1)\ndata_sound = np.frombuffer(data_sound,\"Int16\")\n<\/code><\/pre>\n\n<p>I want to perform average\/max pooling on it with Tensorflow or Keras. But I'm not familiar with this framework so can anyone show me how to implement it? Or do anyone know another way to do it without Tensorflow?<\/p>\n","answers":[{"AnswerId":"58149166","CreationDate":"2019-09-28T18:24:56.137","ParentId":null,"OwnerUserId":"5597718","Title":null,"Body":"<h3>Install Tensorflow<\/h3>\n\n<p>pip install tensorflow==2.0.0-rc1<\/p>\n\n<p>Maxpooling - <code>tf.nn.max_pool1d<\/code><\/p>\n\n<p>Avgpooling - <code>tf.nn.avg_pool1d<\/code><\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/max_pool1d\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/max_pool1d<\/a>\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/MaxPool1D\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/MaxPool1D<\/a><\/p>\n\n<h3>Input data<\/h3>\n\n<pre><code>import numpy as np\na = np.random.randn(4).astype('float32')\nprint(a)\nbac = a[None][:,:,None] # convert to batch\/features\/channels format\n# here batch size = 1, no. of channels = 1\n<\/code><\/pre>\n\n<p>[0.02703167 0.382881   0.57891446 0.58068216]<\/p>\n\n<h3>Tensorflow<\/h3>\n\n<pre><code>import tensorflow as tf\noutput = tf.nn.max_pool1d(bac, 2, 2, padding='VALID')\nprint(np.squeeze(output)) #strip off the additional dimensions\n<\/code><\/pre>\n\n<p>[0.382881   0.58068216]<\/p>\n\n<h3>Keras<\/h3>\n\n<pre><code>from tensorflow.keras import layers\noutput = layers.MaxPool1D(pool_size=2, strides=2)(bac)\nprint(np.squeeze(output)) #strip off the additional dimensions\n<\/code><\/pre>\n\n<p>[0.382881   0.58068216]<\/p>\n"}]}
+{"QuestionId":58146902,"AnswerCount":0,"Tags":"<python><tensorflow><keras><loss>","CreationDate":"2019-09-28T13:45:21.817","AcceptedAnswerId":null,"OwnerUserId":11409756.0,"Title":"Creating custom loss which does not include network output (y_pred)","Body":"<p>Im trying to create custom loss function which is not considering my real neural network output:<\/p>\n\n<pre><code>def loss(y_true, y_pred):\n    y_pred = np.ones([20])#batch size is 20\n    y_pred = K.variable(y_pred)\n    return K.mean(K.square(y_pred - y_true))\n<\/code><\/pre>\n\n<p>Yes, I'm aware what idea is behind gradient descent and back propagation... don't need comments and opinion about this loss function.<\/p>\n\n<p>y_true and y_pred are arrays shape[20,] with values between &lt;0,1><\/p>\n\n<p>when trying to start training using fit() function got this error:<\/p>\n\n<p>ValueError: An operation has <code>None<\/code> 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. <\/p>\n\n<p>How that possible that mean squared error is not differentiable?\nI'm guessing that something is wrong with my loss function but not sure what (the model compiled successfully). Please help me correct my loss function.<\/p>\n","answers":[]}
+{"QuestionId":58146919,"AnswerCount":2,"Tags":"<python><tensorflow><model><tensorboard>","CreationDate":"2019-09-28T13:47:50.107","AcceptedAnswerId":null,"OwnerUserId":12032861.0,"Title":"I cant convert .pb file to .tflite","Body":"<p>I need .tflite file to use the model that i trained in android. <\/p>\n\n<p>I will use this code but i don\u2019t know the input array and the output array values:<\/p>\n\n<pre><code>  tflite_convert \\ \n--output_file=object_detection\\outputtflite  \\ \n--graph_def_file=object_detection\\out\\frozen_inference_graph.pb\\ \n--input_arrays=input\\ \n--output_arrays=output\n<\/code><\/pre>\n\n<p>I heard that i can learn input_arrays and\noutput_arrays values from tensorboard but they don\u2019t appear in there. I use this code:<\/p>\n\n<pre><code>tensorboard --logdir=object_detection\\out\\\n<\/code><\/pre>\n\n<p>There are these errors in my tensorboard screen:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/rOPxW.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rOPxW.png\" alt=\"GRAPHS\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/9oDVZ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9oDVZ.png\" alt=\"\u0130MAGES\"><\/a><\/p>\n\n<p>My out folder looks like this:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/yJR7p.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/yJR7p.png\" alt=\"outFolder\"><\/a><\/p>\n\n<p>What can i do about this?<\/p>\n","answers":[{"AnswerId":"58793643","CreationDate":"2019-11-10T22:51:41.510","ParentId":null,"OwnerUserId":"7697591","Title":null,"Body":"<p>to check for input_arrays and output_arrays. make a python file inside where frozen_inference_graph.pb folder is located and then paste this code in the py<\/p>\n\n<pre><code>import tensorflow as tf\ngf = tf.GraphDef()\nm_file = open('tflite_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>Once you have run the file it will show the input_arrays and output_arrays.<\/p>\n\n<p>furthe info check at : <a href=\"https:\/\/stackoverflow.com\/questions\/52918051\/how-to-convert-pb-to-tflite-format\">How to convert .pb to TFLite format?<\/a><\/p>\n"},{"AnswerId":"58210873","CreationDate":"2019-10-03T00:22:46.783","ParentId":null,"OwnerUserId":"10952155","Title":null,"Body":"<p>You can use the summarize_graph to inspect the graph.\nThis can help you solve the problem\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/tools\/graph_transforms#inspecting-graphs\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/tools\/graph_transforms#inspecting-graphs<\/a><\/p>\n"}]}
+{"QuestionId":58146950,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2019-09-28T13:50:05.930","AcceptedAnswerId":null,"OwnerUserId":7886651.0,"Title":"Error when using tf.image.per_image_standardization with keras api in tensorflow 2.0","Body":"<p>I am using keras api to build a model in tensorflow 2.0. However, I would like to normalize my inputs. Therefore, I want to use <code>tf.image.per_image_standardization<\/code>. Here is my code:<\/p>\n\n<pre><code>inputs = k.Input(shape=(130, 267, 3))\ninputs_normalized = tf.image.per_image_standardization(inputs)\n\nlayer = inputs_normalized\n\nlayer = k.layers.Conv2D(32, (3, 3), activation='relu')(layer)\nlayer = k.layers.MaxPooling2D(2, 2)(layer)\nlayer = k.layers.Conv2D(64, (3, 3), activation='relu')(layer)\nlayer = k.layers.MaxPooling2D(2, 2)(layer)\nlayer = k.layers.Conv2D(64, (3, 3), activation='relu')(layer)\nlayer = k.layers.MaxPooling2D(2, 2)(layer)\nlayer = k.layers.Conv2D(64, (3, 3), activation='relu')(layer)\n\nlayer = k.layers.Flatten()(layer)\nlayer = k.layers.Dense(64, activation='relu')\noutputs = k.layers.Dense(5, activation='softmax')\n\nmodel = k.Model(inputs=inputs, outputs=outputs, name=\"custom model\")\n<\/code><\/pre>\n\n<p>However, I got the following error:<\/p>\n\n<pre><code>OperatorNotAllowedInGraphError: using a `tf.Tensor` as a Python `bool` is not allowed in Graph execution. Use Eager execution or decorate this function with @tf.function.\n<\/code><\/pre>\n\n<p>So I would like to know why I cannot use <code>tf.image.per_image_standardization(inputs)<\/code>. <\/p>\n\n<p>Any help is much appreciated!!<\/p>\n","answers":[]}
+{"QuestionId":58146961,"AnswerCount":1,"Tags":"<python><image><keras>","CreationDate":"2019-09-28T13:51:01.760","AcceptedAnswerId":58154379.0,"OwnerUserId":11486611.0,"Title":"Keras fit_generator doesn't predict the same as fit","Body":"<p>I'm trying to train a model in Keras for image segmentation (U-Net) and started off by generating two lists containing my train (and validation) images and masks. I then trained the model as is shown below.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x_train_val = # list of images (nr_images, 256, 256, 3)\ny_train_val = # list of masks (nr_images, 256, 256, 1)\n\n# Define model\ndef standard_unet():\n    inputs = Input((img_size, img_size, 3))\n\n    s = Lambda(lambda x: x \/ 255) (inputs)\n\n    c1 = Conv2D(8, (3, 3), activation='relu', padding='same') (inputs)\n    c1 = Conv2D(8, (3, 3), activation='relu', padding='same') (c1)\n    p1 = MaxPooling2D((2, 2)) (c1)\n\n    c2 = Conv2D(16, (3, 3), activation='relu', padding='same') (p1)\n    c2 = Conv2D(16, (3, 3), activation='relu', padding='same') (c2)\n    p2 = MaxPooling2D((2, 2)) (c2)\n\n    c3 = Conv2D(32, (3, 3), activation='relu', padding='same') (p2)\n    c3 = Conv2D(32, (3, 3), activation='relu', padding='same') (c3)\n    p3 = MaxPooling2D((2, 2)) (c3)\n\n    c4 = Conv2D(64, (3, 3), activation='relu', padding='same') (p3)\n    c4 = Conv2D(64, (3, 3), activation='relu', padding='same') (c4)\n    p4 = MaxPooling2D(pool_size=(2, 2)) (c4)\n\n    c5 = Conv2D(128, (3, 3), activation='relu', padding='same') (p4)\n    c5 = Conv2D(128, (3, 3), activation='relu', padding='same') (c5)\n\n    u6 = Conv2DTranspose(64, (2, 2), strides=(2, 2), padding='same') (c5)\n    u6 = concatenate([u6, c4])\n    c6 = Conv2D(64, (3, 3), activation='relu', padding='same') (u6)\n    c6 = Conv2D(64, (3, 3), activation='relu', padding='same') (c6)\n\n    u7 = Conv2DTranspose(32, (2, 2), strides=(2, 2), padding='same') (c6)\n    u7 = concatenate([u7, c3])\n    c7 = Conv2D(32, (3, 3), activation='relu', padding='same') (u7)\n    c7 = Conv2D(32, (3, 3), activation='relu', padding='same') (c7)\n\n    u8 = Conv2DTranspose(16, (2, 2), strides=(2, 2), padding='same') (c7)\n    u8 = concatenate([u8, c2])\n    c8 = Conv2D(16, (3, 3), activation='relu', padding='same') (u8)\n    c8 = Conv2D(16, (3, 3), activation='relu', padding='same') (c8)\n\n    u9 = Conv2DTranspose(8, (2, 2), strides=(2, 2), padding='same') (c8)\n    u9 = concatenate([u9, c1], axis=3)\n    c9 = Conv2D(8, (3, 3), activation='relu', padding='same') (u9)\n    c9 = Conv2D(8, (3, 3), activation='relu', padding='same') (c9)\n\n    outputs = Conv2D(1, (1, 1), activation='sigmoid') (c9)\n\n    model = Model(inputs=[inputs], outputs=[outputs])\n\n    return model\n\n# IoU metric\ndef mean_iou(y_true, y_pred):\n    prec = []\n    for t in np.arange(0.5, 1.0, 0.05):\n        y_pred_ = tf.to_int32(y_pred &gt; t)\n        score, up_opt = tf.metrics.mean_iou(y_true, y_pred_, 2)\n        K.get_session().run(tf.local_variables_initializer())\n        with tf.control_dependencies([up_opt]):\n            score = tf.identity(score)\n        prec.append(score)\n    return K.mean(K.stack(prec), axis=0)\n\n# Dice coef loss\ndef dice_coef(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    return (2. * intersection + smooth) \/ (K.sum(y_true_f) + K.sum(y_pred_f) + smooth)\n\ndef bce_dice_loss(y_true, y_pred):\n    return 0.5 * binary_crossentropy(y_true, y_pred) - dice_coef(y_true, y_pred)\n\n# Model compiling\nK.clear_session()\nmodel = standard_unet()\nmodel.compile(optimizer='adam', loss=bce_dice_loss, metrics=[mean_iou])\n\n# Fitting\nmodel.fit(x_train_val, y_train_val, validation_split=0.1, epochs=20)\n<\/code><\/pre>\n\n<p>This works exactly as intended and I get decent results when I try to predict on test images. As I want to increase the number of training images, I tried using <code>ImageDataGenerator<\/code> and <code>train_generator<\/code> using the following function.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Runtime data augmentation\ndef get_train_test_augmented(x_data=x_train_val, y_data=y_train_val, validation_split=0.1, batch_size=32):\n    x_train, x_valid, y_train, y_valid = train_test_split(x_data, y_data,\n                                                          train_size=1-validation_split,\n                                                          test_size=validation_split)\n\n    data_gen_args = dict(rotation_range=45.,\n                         width_shift_range=0.1,\n                         height_shift_range=0.1,\n                         horizontal_flip=True,\n                         vertical_flip=True,\n                         fill_mode='reflect')  #use 'constant'??\n\n    x_datagen = ImageDataGenerator(**data_gen_args)\n    y_datagen = ImageDataGenerator(**data_gen_args)\n    x_datagen.fit(x_train, augment=True)\n    y_datagen.fit(y_train, augment=True)\n    x_train_augmented = x_datagen.flow(x_train, batch_size=batch_size, shuffle=True)\n    y_train_augmented = y_datagen.flow(y_train, batch_size=batch_size, shuffle=True)\n\n    # combine generators into one which yields image and masks\n    train_generator = zip(x_train_augmented, y_train_augmented)\n\n    return train_generator\n<\/code><\/pre>\n\n<p>Visual inspection of these images shows, that they contain what I expect (augmented images and masks). However, when I now fit my model, my predictions are always blank.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>train_generator = get_train_test_augmented()\nmodel.fit_generator(train_generator, epochs=20)\n<\/code><\/pre>\n\n<p>Has anyone experienced the same issue with blank image prediction or know how to solve it?\nThanks, BBQuercus.<\/p>\n","answers":[{"AnswerId":"58154379","CreationDate":"2019-09-29T10:30:33.267","ParentId":null,"OwnerUserId":"4332585","Title":null,"Body":"<p>You are doing the image generation separately with the image and mask generators, what will happen is that the random transform will not be the same for the input image and the label (mask). \nEven more than that, you are shuffling both generators, so they do not even correspond to one another (the image and the mask in fit generator).<\/p>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/3059#issuecomment-364787723\" rel=\"nofollow noreferrer\">This github issue comment<\/a> also talks about that, and suggest creating an extra generator to merge both.<\/p>\n\n<p>Try seeding both generators with the same seed and see if it changes anything. <\/p>\n\n<hr>\n\n<p>EDIT<\/p>\n\n<p>As I was working with image denoising, I noticed that the solution using <code>zip<\/code> is suboptimal because you can't make use of <code>use_multiprocessing=True<\/code> when fitting. A solution is to implement a custom generator merger:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class MergedGenerators(Sequence):\n    def __init__(self, *generators):\n        self.generators = generators\n        # TODO add a check to verify that all generators have the same length\n\n    def __len__(self):\n        return len(self.generators[0])\n\n    def __getitem__(self, index):\n        return [generator[index] for generator in self.generators]\n\ntrain_generator = MergedGenerators(image_generator, mask_generator)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58147133,"AnswerCount":0,"Tags":"<android><tensorflow><encryption><decompiler>","CreationDate":"2019-09-28T14:10:46.730","AcceptedAnswerId":null,"OwnerUserId":9298828.0,"Title":"How to encrypt and decrypt tensorflow model in android successfully","Body":"<p>As files in android assets can be easily hacked by converting .apk to .zip. I have some propitiatory model which I need to use in android application. I am working on how to encrypt the assets of the application. ( I will use firebase for retreiving  the key for decryption to make it more secure ) \nI am pretty successful in encrypting and decrypting images (png and jpg) using cipher library in android. <\/p>\n\n<p>Following is the code which I used for the purpose.<\/p>\n\n<pre><code>    String fake = \"sadhkja6\";\n    byte[] encodedKey     = Base64.decode(fake, Base64.DEFAULT);\n    yourKey= new SecretKeySpec(encodedKey, 0, encodedKey.length, \"AES\");\n    Bitmap bitmap = BitmapFactory.decodeResource(getResources(), id);\n    bitmap = Bitmap.createScaledBitmap(bitmap, 500,500, true);\n\n   byte[] output = getBytes(bitmap);\n\n   byte[] output = getBytesFile();\n   saveFile(output);\n\n\npublic static byte[] encodeFile(SecretKey yourKey, byte[] fileData)\n        throws Exception {\n    byte[] encrypted = null;\n    byte[] data = yourKey.getEncoded();\n    SecretKeySpec skeySpec = new SecretKeySpec(data, 0, data.length,\n            algorithm);\n    Cipher cipher = Cipher.getInstance(algorithm);\n    cipher.init(Cipher.ENCRYPT_MODE, skeySpec, new IvParameterSpec(\n            new byte[cipher.getBlockSize()]));\n    encrypted = cipher.doFinal(fileData);\n    return encrypted;\n}\n\n\npublic byte[] getBytes(Bitmap bitmap){\n\n    ByteArrayOutputStream stream = new ByteArrayOutputStream();\n    bitmap.compress(    Bitmap.CompressFormat.PNG, 100, stream);\n    byte[] byteArray = stream.toByteArray();\n    bitmap.recycle();\n\n    return byteArray;\n}\n\npublic static byte[] encodeFile(SecretKey yourKey, byte[] fileData)\n        throws Exception {\n    byte[] encrypted = null;\n    byte[] data = yourKey.getEncoded();\n    SecretKeySpec skeySpec = new SecretKeySpec(data, 0, data.length,\n            algorithm);\n    Cipher cipher = Cipher.getInstance(algorithm);\n    cipher.init(Cipher.ENCRYPT_MODE, skeySpec, new IvParameterSpec(\n            new byte[cipher.getBlockSize()]));\n    encrypted = cipher.doFinal(fileData);\n    return encrypted;\n}\n\nvoid saveFile(byte[] array) {\n    String decrypted_file = \"encypted.png\";\n\n    try {\n        File file = new File(Environment.getExternalStorageDirectory()\n                + File.separator, decrypted_file);\n        BufferedOutputStream bos = new BufferedOutputStream(\n                new FileOutputStream(file));\n        yourKey = generateKey();\n        byte[] filesBytes = encodeFile(yourKey, array);\n        bos.write(filesBytes);\n        bos.flush();\n        bos.close();\n    } catch (FileNotFoundException e) {\n        e.printStackTrace();\n    } catch (IOException e) {\n        e.printStackTrace();\n    } catch (Exception e) {\n        e.printStackTrace();\n    }\n}\n<\/code><\/pre>\n\n<p>But when I use the same code for encyrpting tensorflow models, the model is not successfully decompiled resulting in app crash. Most probably because we are loading the model in bytes which may cause problems.<\/p>\n\n<p>Please tell me how can I use the cyphering technique for securing my tensorflow model. Thank you very much.<\/p>\n","answers":[]}
+{"QuestionId":58147364,"AnswerCount":0,"Tags":"<python><tensorflow><shapes>","CreationDate":"2019-09-28T14:38:44.760","AcceptedAnswerId":null,"OwnerUserId":735008.0,"Title":"What is the difference between x.shape and tf.shape() in tensorflow 2.0?","Body":"<p>I wonder when I need to use tf.shape() and x.shape(). I'm currently using tensorflow 2.0 rc0<\/p>\n\n<p>The following is an example code.<\/p>\n\n<pre><code>#!\/usr\/bin\/python3\n\nimport tensorflow as tf\n\na = tf.zeros((4, 3, 1)) \nprint (tf.shape(a).numpy())\nprint (a.shape)\n<\/code><\/pre>\n\n<p>The result of the above code is as follows:<\/p>\n\n<pre><code>[4 3 1]\n(4, 3, 1)\n<\/code><\/pre>\n\n<p><code>tf.shape(a).numpy()<\/code> returns the numpy array whereas <code>a.shape<\/code> returns a tuple, but I cannot easily find which one is better and which one should be preferred.<\/p>\n\n<p>Could anyone please give some advice on this?<\/p>\n","answers":[]}
+{"QuestionId":58147366,"AnswerCount":1,"Tags":"<tensorflow><keras><neural-network><importerror>","CreationDate":"2019-09-28T14:38:57.927","AcceptedAnswerId":null,"OwnerUserId":9126289.0,"Title":"Problem with importing tensorflow and testing NN","Body":"<p>I'm currently working on a program to play a game similar to atari-games. I'm using keras (python 3). I finished writing the code and I want to test it, and I have few questions about the process:\nfirst of all, I have trouble importing tesnorflow for some reason. I've installed it using pip. I've made sure to created new env. before the installation (which finished successfully), but when I try to run my program it says:<\/p>\n\n<pre><code>ModuleNotFoundError: No module named 'tensorflow'\n<\/code><\/pre>\n\n<p>I also, tried to install the package from within pycharm, but then I get this error:<\/p>\n\n<pre><code>Could not find a version that satisfies the requirement tensorflow (from versions: )\nNo matching distribution found for tensorflow\n<\/code><\/pre>\n\n<p>I've checked program requirements (such as pip, python, virtualenv and setuptools versions) and everything seems up to date. perhaps someone could point out what else might be the problem?<\/p>\n\n<p>Is there any other way I can test the performance of my program?<\/p>\n\n<p>Thank you very much for your time and attention.<\/p>\n","answers":[{"AnswerId":"58167308","CreationDate":"2019-09-30T11:47:58.293","ParentId":null,"OwnerUserId":"12118774","Title":null,"Body":"<p><strong>Anaconda<\/strong> is a complete time-saver. I suggest create an enviornment using Anaconda and install the tensorflow by <code>conda install tensorflow<\/code> If you would like to use the gpu version, conda automatically installs the CUDA and cudnn for you too.<\/p>\n"}]}
+{"QuestionId":58147888,"AnswerCount":1,"Tags":"<python><tensorflow><keras><multiprocessing>","CreationDate":"2019-09-28T15:42:22.607","AcceptedAnswerId":58238373.0,"OwnerUserId":6394941.0,"Title":"Keras + multiprocessing - correctly generating sessions, but only one processor","Body":"<p>I have a problem with Keras and multiprocessing. I have already searched a lot and I found a lot of questions with the same subjects:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/50262215\/importing-keras-breaks-multiprocessing\">Importing Keras breaks multiprocessing<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/42504669\/keras-tensorflow-and-multiprocessing-in-python\">Keras + Tensorflow and Multiprocessing in Python<\/a><\/li>\n<\/ul>\n\n<p>(and lot more)<\/p>\n\n<p>I tried these solutions, so basically importing Keras after the multiprocessing has been instantiated. In actual fact, I see this message:<\/p>\n\n<pre><code>Using TensorFlow backend.\nUsing TensorFlow backend.\nUsing TensorFlow backend.\nUsing TensorFlow backend.\nUsing TensorFlow backend.\nUsing TensorFlow backend.\nUsing TensorFlow backend.\nUsing TensorFlow backend.\nUsing TensorFlow backend.\nUsing TensorFlow backend. \n<\/code><\/pre>\n\n<p>Before this message was only printed one time, so I assume that the backend is different, however, my processes are running on the same core. If run again the main process, it creates more processes that run in the same processor as well. It seems that something blocks the execution on different processors.<\/p>\n\n<p>Any idea on how to fix it?<\/p>\n\n<p>PS: I am using the second solution I have linked, in particular the following :<\/p>\n\n<pre><code>DO NOT LOAD KERAS TO YOUR MAIN ENVIRONMENT\n<\/code><\/pre>\n","answers":[{"AnswerId":"58238373","CreationDate":"2019-10-04T14:30:33.957","ParentId":null,"OwnerUserId":"6394941","Title":null,"Body":"<p>The problem was in the installation of <code>tensorflow<\/code> and <code>keras<\/code>. The methods for achieving parallelization are correct.<\/p>\n\n<p>The <code>tensorflow<\/code> documentation clearly states that is highly suggested to install the package using <code>pip<\/code> as the <code>conda<\/code> package is maintained only by the community (<a href=\"https:\/\/www.tensorflow.org\/install\/pip\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/pip<\/a>). <\/p>\n\n<p>I fixed the problem uninstalling <code>keras<\/code> and <code>tensorflow<\/code> and reinstalling them with:<\/p>\n\n<pre><code>pip install tensorflow\npip install keras\n<\/code><\/pre>\n"}]}
+{"QuestionId":58148079,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><raspberry-pi3><raspbian-stretch>","CreationDate":"2019-09-28T16:06:06.103","AcceptedAnswerId":null,"OwnerUserId":9009851.0,"Title":"ImportError: \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/_pywrap_tensorflow.so: cannot open shared object file: No such file or directory","Body":"<p>Working on a model that uses webcam to detect and recognize face on Raspbian Stretch Desktop(Run on virtualBox). Have installed tensorflow, keras, opencv, mtcnn etc. While importing tensorflow, the following error is obtained.<\/p>\n\n<p>Works fine on Ubuntu, but Raspbian only is throwing this ImportError.\nTried running the same with sudo, but in vain. <\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"webcam.py\", line 3, in &lt;module&gt;\n    from mtcnn.mtcnn import MTCNN\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/mtcnn\/mtcnn.py\", line 38, in &lt;module&gt;\n    import tensorflow as tf\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/__init__.py\", line 23, in &lt;module&gt;\n    from tensorflow.python import *\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow = swig_import_helper()\n  File \"\/usr\/local\/lib\/python3.5\/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\/python3.5\/imp.py\", line 242, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/usr\/lib\/python3.5\/imp.py\", line 342, in load_dynamic\n    return _load(spec)\nImportError: \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/_pywrap_tensorflow.so: cannot open shared object file: No such file or directory\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58148396,"AnswerCount":2,"Tags":"<tensorflow><object-detection-api>","CreationDate":"2019-09-28T16:45:12.247","AcceptedAnswerId":null,"OwnerUserId":12009032.0,"Title":"What should I add so that when the system detects some specific object it will play an alert sound?","Body":"<p>I have done the object detection api using webcam, the system was run successful detected object, now i want to add, when the system detected some specific object it will play diff alert sound <\/p>\n\n<pre><code>while True:\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 = video.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=0.60)\n    #if xxxxx:\n    #    alert.play()\n    #else:\n    #    pass\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    # Press 'q' to quit\n    if cv2.waitKey(1) == ord('q'):\n        break\n<\/code><\/pre>\n\n<p>my system is auto detect dangerous weapon, when my system detected \"Gun\", or \"Knife\" it will alert security by using sound alarm.<\/p>\n","answers":[{"AnswerId":"58407597","CreationDate":"2019-10-16T07:07:22.460","ParentId":null,"OwnerUserId":"12199329","Title":null,"Body":"<p>Insert loop<\/p>\n\n<pre><code>import winsound\nif category_index=='index of knife':\n    windsound.beep(duration(ms),frequency)\nelse:\n    pass\n<\/code><\/pre>\n"},{"AnswerId":"58148810","CreationDate":"2019-09-28T17:38:45.147","ParentId":null,"OwnerUserId":"11156408","Title":null,"Body":"<p><strong>For Windows<\/strong><\/p>\n\n<pre><code>import winsound\n\nwinsound.PlaySound(\"sound_file.wav\", FLAG)\n<\/code><\/pre>\n\n<p>or just a beep<\/p>\n\n<pre><code>import winsound\n\ndur = 500 # as millisecond\nfreq = 2000 # sound frequency\nwinsound.Beep(freq, dur)\n<\/code><\/pre>\n\n<p>you can check the document \n<a href=\"https:\/\/docs.python.org\/3.7\/library\/winsound.html\" rel=\"nofollow noreferrer\">https:\/\/docs.python.org\/3.7\/library\/winsound.html<\/a><\/p>\n\n<p><strong>For other OS (Linux, Mac etc.)<\/strong><\/p>\n\n<pre><code>import os\n\nos.system(\"sound_file.wav&amp;\")\n<\/code><\/pre>\n"}]}
+{"QuestionId":58148597,"AnswerCount":0,"Tags":"<matplotlib><neural-network><pytorch>","CreationDate":"2019-09-28T17:11:26.060","AcceptedAnswerId":null,"OwnerUserId":10240085.0,"Title":"How to draw decision boundary for 2 layer Neural Network","Body":"<p>I have a two layer Neural Network. Given the weights and biases predicted by Neural Network, how to draw the decision boundary on this dataset? <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/wgfge.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/wgfge.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>My Neural network is defined as follows:<\/p>\n\n<pre><code>class Classifier(nn.Module):\n    def __init__(self):\n        super().__init__()\n        self.fc1 = nn.Linear(2,2)\n        self.fc2 = nn.Linear(2,1)\n    def forward(self,x):\n        x = F.relu(self.fc1(x))\n        x = self.fc2(x)\n        return x\n<\/code><\/pre>\n\n<p><code>list(model.parameters())<\/code> returns me <\/p>\n\n<pre><code>[Parameter containing:\n tensor([[-0.5128, -0.3776],\n         [ 0.4611,  0.4625]], requires_grad=True), Parameter containing:\n tensor([ 0.6443, -3.6282], requires_grad=True), Parameter containing:\n tensor([[-0.1985,  2.8158]], requires_grad=True), Parameter containing:\n tensor([-5.4537], requires_grad=True)]\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/colab.research.google.com\/drive\/1iMsdz1N338boPoBlOH9D0-P8rTmjuKsD\" rel=\"nofollow noreferrer\">The Google Colab notebook is attached<\/a>. You will find the complete code under <strong>Toy Dataset 3 predicting logits<\/strong>. Besides, I have drawn 1 layer neural network decision boundary as an example. Find it under <strong>Toy Dataset 2 predicting logits<\/strong>. For your ease, 1 layer neural network dataset and its decision boundary is attached here as well.\n<a href=\"https:\/\/i.stack.imgur.com\/TKJHX.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/TKJHX.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/2VnYV.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/2VnYV.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
+{"QuestionId":58149052,"AnswerCount":0,"Tags":"<tensorflow><keras><tf.keras>","CreationDate":"2019-09-28T18:08:44.537","AcceptedAnswerId":null,"OwnerUserId":10895879.0,"Title":"Training in TF2.0.0b1 on multi_gpu_model, does not work","Body":"<p>I want to train my model with the multi_gpu_model() function. However, this does not work as \nI get the following error <\/p>\n\n<pre><code>ValueError: To call `multi_gpu_model` with `gpus=2`, we expect the following devices to be available: ['\/cpu:0', '\/gpu:0', '\/gpu:1']. However this machine only has: ['\/cpu:0', '\/xla_gpu:0', '\/xla_gpu:1', '\/xla_gpu:2', '\/xla_gpu:3', '\/xla_cpu:0']. Try reducing `gpus`.\n<\/code><\/pre>\n\n<p>Is there a way to fix this?<\/p>\n\n<p>For PC specs: I am using the AWS EC2 AMI <code>Deep Learning AMI (Ubuntu) Version 24.2 (ami-0ba6d589ad99d7604)<\/code><\/p>\n\n<p>Using MirroredStrategy() yields a different error, something with sample_weights not being abled to be unpacked.<\/p>\n\n<pre><code>Train on 10 steps, validate on 10 steps\nTraceback (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\/train\/code\/sc-ai-model-training\/src\/main\/scripts\/run_training.py\", line 94, in &lt;module&gt;\n    main()\n  File \"\/home\/train\/code\/sc-ai-model-training\/src\/main\/scripts\/run_training.py\", line 79, in main\n    verbose=1)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/training.py\", line 643, in fit\n    use_multiprocessing=use_multiprocessing)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/training_distributed.py\", line 681, in fit\n    steps_name='steps_per_epoch')\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/engine\/training_arrays.py\", line 294, in model_iteration\n    batch_outs = f(actual_inputs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras\/distribute\/distributed_training_utils.py\", line 814, in execution_function\n    return [out.numpy() for out in distributed_function(input_fn)]\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/eager\/def_function.py\", line 416, in __call__\n    self._initialize(args, kwds, add_initializers_to=initializer_map)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/eager\/def_function.py\", line 359, in _initialize\n    *args, **kwds))\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/eager\/function.py\", line 1360, in _get_concrete_function_internal_garbage_collected\n    graph_function, _, _ = self._maybe_define_function(args, kwargs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/eager\/function.py\", line 1648, in _maybe_define_function\n    graph_function = self._create_graph_function(args, kwargs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/eager\/function.py\", line 1541, in _create_graph_function\n    capture_by_value=self._capture_by_value),\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/framework\/func_graph.py\", line 716, in func_graph_from_py_func\n    func_outputs = python_func(*func_args, **func_kwargs)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/eager\/def_function.py\", line 309, in wrapped_fn\n    return weak_wrapped_fn().__wrapped__(*args, **kwds)\n  File \"\/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/framework\/func_graph.py\", line 706, in wrapper\n    raise e.ag_error_metadata.to_exception(type(e))\nValueError: in converted code:\n    relative to \/usr\/local\/lib\/python3.5\/dist-packages\/tensorflow\/python\/keras:\n\n    distribute\/distributed_training_utils.py:799 distributed_function  *\n        x, y, sample_weights = input_fn()\n    engine\/training_arrays.py:506 get_distributed_inputs\n        model, inputs, targets, sample_weights, mode)\n    distribute\/distributed_training_utils.py:580 _prepare_feed_values\n        inputs, targets, sample_weights = _get_input_from_iterator(inputs, model)\n    distribute\/distributed_training_utils.py:558 _get_input_from_iterator\n        x, y, sample_weights = next_element\n\n    ValueError: not enough values to unpack (expected 3, got 2)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58149306,"AnswerCount":0,"Tags":"<python><tensorflow><python-3.7.4>","CreationDate":"2019-09-28T18:46:29.777","AcceptedAnswerId":null,"OwnerUserId":7376038.0,"Title":"Tensorflow support for python 3.7.4","Body":"<p>I'm trying to pip install tensor flow on python 3.7.4 and I get the following error:<\/p>\n\n<blockquote>\n  <p>ERROR: Could not find a version that satisfies the requirement\n  tensorflow (from versions: none) ERROR: No matching distribution found\n  for tensorflow<\/p>\n<\/blockquote>\n\n<p>From tensorflow installation site:<\/p>\n\n<blockquote>\n  <p>Windows Python 3.5\n  CPU-only  <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.14.0-cp35-cp35m-win_amd64.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.14.0-cp35-cp35m-win_amd64.whl<\/a><\/p>\n  \n  <p>Python 3.5 GPU\n  support   <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.14.0-cp35-cp35m-win_amd64.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.14.0-cp35-cp35m-win_amd64.whl<\/a><\/p>\n  \n  <p>Python 3.6\n  CPU-only  <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.14.0-cp36-cp36m-win_amd64.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.14.0-cp36-cp36m-win_amd64.whl<\/a><\/p>\n  \n  <p>Python 3.6 GPU\n  support   <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.14.0-cp36-cp36m-win_amd64.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.14.0-cp36-cp36m-win_amd64.whl<\/a><\/p>\n<\/blockquote>\n\n<p>It seems that there is not version for Python 3.7 in Windows. Is there any info about the release date of the new version?<\/p>\n","answers":[]}
+{"QuestionId":58149401,"AnswerCount":1,"Tags":"<python><memory><gpu><pytorch>","CreationDate":"2019-09-28T18:59:17.103","AcceptedAnswerId":58150831.0,"OwnerUserId":8040538.0,"Title":"Why pytorch needs much more memory than it should?","Body":"<p>I'm just playing around with pytorch and I'm wondering why it consumes so much memory of my GPU?<\/p>\n\n<p>I'm using Cuda 10.0 with pythorch 1.2.0 and torchvision 0.4.0.<\/p>\n\n<pre><code>import torch\ngpu = torch.device(\"cuda\")\nx = torch.ones(int(4e8), device=gpu)\ny = torch.ones(int(1e5), device=gpu)\n<\/code><\/pre>\n\n<p>Running the above code I get the error:\nRuntimeError: CUDA out of memory. Tried to allocate 2.00 MiB (GPU 0; 2.00 GiB total capacity; 1.49 GiB already allocated; 0 bytes free; 0 bytes cached)<\/p>\n\n<p>So, does pytorch needs ~500MB of the gpu memory as overhead? Or what is the problem here?<\/p>\n","answers":[{"AnswerId":"58150831","CreationDate":"2019-09-28T22:40:20.667","ParentId":null,"OwnerUserId":"12118774","Title":null,"Body":"<p>More information and testing done by <code>xymeng<\/code> in github could be seen in the given <a href=\"https:\/\/github.com\/pytorch\/pytorch\/issues\/12873\" rel=\"nofollow noreferrer\">link<\/a><\/p>\n\n<p>Referencing <code>xymeng<\/code>'s words : <\/p>\n\n<blockquote>\n  <p>PyTorch has its own cuda kernels. From my measurement the cuda runtime allocates ~1GB memory for them. If you compile pytorch with cudnn enabled the total memory usage is 1GB + 750M + others = 2GB+\n  Note that this is just my speculation as there is no official documentation about this. What puzzles me is that the cuda runtime allocates much more memory than the actual code size (they are approx. linearly correlated. If I remove half of pytorch's kernels the memory usage is also reduced by half). I suspect either the kernel binaries have been compressed or they have to be post-processed by the runtime.<\/p>\n<\/blockquote>\n\n<p>Seems it suits your situation.<\/p>\n"}]}
+{"QuestionId":58149560,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-28T19:19:36.243","AcceptedAnswerId":null,"OwnerUserId":4830432.0,"Title":"Dynamically creating Tensorflow operations","Body":"<p>I am writing an algorithm that generates a series of same-size vectors that I would like to multiply together with Tensorflow repeatedly.<\/p>\n\n<p>The pseudocode of the algorithm goes like this:<\/p>\n\n<pre><code>1. Generate N randomly initialized vectors of the same size\n2. Multiply all the vectors together with a seed to get a result\n3. Store the result\n3. Numerically alter the vectors\n4. Repeat steps 2 to 3 N times (N is very large)\n<\/code><\/pre>\n\n<p>The best way I can think of is to create a tensor with these vectors and then use it like so:<\/p>\n\n<pre><code>def generateTFOperation(inputPlaceholder, vectorOfVectors) :\n\n    operation = tf.multiply(inputPlaceholder, vectorOfVectors[0])\n\n    for index in range(1, len(vectorOfVectors)) :\n        operation = tf.multiply(operation, vectorOfVectors[index])\n\n    return operation\n<\/code><\/pre>\n\n<p>Using the operation:<\/p>\n\n<pre><code>def mutliplyVectorsWithSeed(seed, vectorOfVectors) :\n    inputPlaceholder = tf.compat.v1.placeholder(tf.float32)\n    operation = generateTFOperation(inputPlaceholder, vectorOfVectors)\n\n    with tf.compat.v1.Session() as sess:\n\n        resultant = sess.run(operation, feed_dict = {inputPlaceholder: seed}\n\n    return resultant\n<\/code><\/pre>\n\n<p>And the calling function:<\/p>\n\n<pre><code>def callingFunction() :\n\n    results = []\n    vectorOfVectors = generateRandomVectors(vectorSize) #This is implemented elsewhere\n    for i in range(aLargeNumber) :\n\n        result = multiplyVectorsWithSeed(someSeed, vectorOfVectors)\n        results.append(result)\n\n        vectorOfVectors = numericallyAlterVectors(vectorOfVectors)\n\n    return results\n<\/code><\/pre>\n\n<p>I've implemented this and it is taking a horrendous amount of time to run. Is there a more efficient \/ idiomatic way to do this sort of thing?<\/p>\n\n<p>I am concerned that dynamically creating a tensor operation like this has a huge overhead. I've thought of trying to do this with placeholders and just create the operation once but I have no idea how to do that. Im not sure how I could feed the set of vectors to a set of placeholders.<\/p>\n\n<p>Another thought I had to eliminate the overhead of this approach would be to figure out how to load the vectors into a rank 2 tensor in Tensorflow and somehow alter them in place to avoid having to repeatedly create the series of multiplications by loading the vectors into them.<\/p>\n\n<p>I am rather new to Tensorflow and am not even sure if it is a suitable library to use for what I am trying to do. Any help\/direction would be appreciated!<\/p>\n","answers":[]}
+{"QuestionId":58149572,"AnswerCount":2,"Tags":"<tensorflow><keras><cpu><compiler-optimization><python-3.7>","CreationDate":"2019-09-28T19:21:27.627","AcceptedAnswerId":null,"OwnerUserId":11808133.0,"Title":"How to run TensorFlow for SSE4.1 SSE4.2 AVX AVX2 FMA on Python with Spyder on MacOs","Body":"<p>I am trying to run the code: <\/p>\n\n<pre><code>from keras.datasets import imdb as im\nfrom keras.preprocessing import sequence as seq\nfrom keras.models import Sequential\nfrom keras.layers import Embedding\nfrom keras.layers import LSTM\nfrom keras.layers import Dense\n\ntrain_set, test_set = im.load_data(num_words = 10000)\nX_train, y_train = train_set\nX_test, y_test = test_set\n\nX_train_padded = seq.pad_sequences(X_train, maxlen = 100)\nX_test_padded = seq.pad_sequences(X_test, maxlen = 100)\n\nmodel = Sequential()\nmodel.add(Embedding(input_dim=10000, output_dim=128))\nmodel.add(LSTM(units=128))\nmodel.add(Dense(units=1, activation='sigmoid'))\nmodel.compile(loss='binary_crossentropy',\n                  optimizer='sgd',\n                  metrics=['accuracy'])\nscores = model.fit(X_train_padded,y_train)\n\n<\/code><\/pre>\n\n<p>When I run the code, it gives me a message:<\/p>\n\n<blockquote>\n  <p>I tensorflow\/core\/platform\/cpu_feature_guard.cc:145] This TensorFlow binary is optimized with Intel(R) MKL-DNN to use the following CPU instructions in performance critical operations:  SSE4.1 SSE4.2 AVX AVX2 FMA<\/p>\n  \n  <p>To enable them in non-MKL-DNN operations, rebuild TensorFlow with the appropriate compiler flags.<\/p>\n  \n  <p>I tensorflow\/core\/common_runtime\/process_util.cc:115] Creating new thread pool with default inter op setting: 4. Tune using inter_op_parallelism_threads for best performance.<\/p>\n<\/blockquote>\n\n<p>I don't understand what the issue is and what I am supposed to do next. I installed the \"tenserflow\" package (1.14.0) but that doesn't solve the issue.<\/p>\n\n<p>I have looked at this reference but I don't know what I am looking for:<\/p>\n\n<pre>\nhttps:\/\/stackoverflow.com\/questions\/41293077\/how-to-compile-tensorflow-with-sse4-2-and-avx-instructions\n<\/pre>\n\n<p>Can someone please help me. Thanks.<\/p>\n\n<p>my config: osx-64, MacOS Mojave v.10.14.6, Python 3.7 with Spyder with Anaconda, conda version : 4.7.12<\/p>\n","answers":[{"AnswerId":"58973670","CreationDate":"2019-11-21T11:15:39.120","ParentId":null,"OwnerUserId":"5445215","Title":null,"Body":"<p>This might not be answering the exact question you have put, but I had a very similar error message when running a similar task.<\/p>\n\n<p>In addition to the error message above, I also had the following error message:<\/p>\n\n<blockquote>\n  <p>OMP: Error #15: Initializing libiomp5.dylib, but found libiomp5.dylib already initialized.\n  OMP: Hint This means that multiple copies of the OpenMP runtime have been linked into the program. That is dangerous, since it can degrade performance or cause incorrect results. The best thing to do is to ensure that only a single OpenMP runtime is linked into the process, e.g. by avoiding static linking of the OpenMP runtime in any library. As an unsafe, unsupported, undocumented workaround you can set the environment variable KMP_DUPLICATE_LIB_OK=TRUE to allow the program to continue to execute, but that may cause crashes or silently produce incorrect results. For more information, please see <a href=\"http:\/\/www.intel.com\/software\/products\/support\/\" rel=\"nofollow noreferrer\">http:\/\/www.intel.com\/software\/products\/support\/<\/a>.<\/p>\n<\/blockquote>\n\n<p>Error was solved with:<\/p>\n\n<pre><code>conda install nomkl\n<\/code><\/pre>\n\n<p>This is as per this <a href=\"https:\/\/stackoverflow.com\/questions\/53648730\/omp-error-15-initializing-libiomp5-dylib-but-found-libiomp5-dylib-already-in\">stackoverflow<\/a> post<\/p>\n"},{"AnswerId":"58988211","CreationDate":"2019-11-22T05:46:52.273","ParentId":null,"OwnerUserId":"5091738","Title":null,"Body":"<p>You can ignore the message and everything will work fine.<\/p>\n\n<p>As far as I can gather from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/24782\/commits\/7faefa4bb665e115cc744d7895a407338624993f\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/pull\/24782\/commits\/7faefa4bb665e115cc744d7895a407338624993f<\/a>, when TensorFlow is compiled with MKL-DNN support (which it is, according to your message), MKL-DNN will take care of using all available CPU performance features. So it doesn't matter that TensorFlow wasn't compiled to use them.<\/p>\n"}]}
+{"QuestionId":58149839,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2019-09-28T19:55:58.983","AcceptedAnswerId":58151051.0,"OwnerUserId":11847592.0,"Title":"Learning rate of custom training loop for tensorflow 2.0","Body":"<p>Are there any functions or methods which can show the learning rate when I use the tensorflow 2.0 custom training loop? <\/p>\n\n<p>Here is an example of tensorflow guide: <\/p>\n\n<pre><code>def train_step(images, labels):\n  with tf.GradientTape() as tape:\n    predictions = model(images)\n    loss = loss_object(labels, predictions)\n  gradients = tape.gradient(loss, model.trainable_variables)\n  optimizer.apply_gradients(zip(gradients, model.trainable_variables))\n\n  train_loss(loss)\n  train_accuracy(labels, predictions)\n<\/code><\/pre>\n\n<p>How can I retrieve the current learning rate from the optimizer when the model is training?<\/p>\n\n<p>I will be grateful for any help you can provide. :)<\/p>\n","answers":[{"AnswerId":"58151051","CreationDate":"2019-09-28T23:30:00.880","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>In custom training loop setting, you can <code>print(optimizer.lr.numpy())<\/code> to get the learning rate. <\/p>\n\n<p>If you are using keras api, you can define your own callback that records the current learning rate. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow.keras.callbacks import Callback\n\nclass LRRecorder(Callback):\n    \"\"\"Record current learning rate. \"\"\"\n    def on_epoch_begin(self, epoch, logs=None):\n        lr = self.model.optimizer.lr\n        print(\"The current learning rate is {}\".format(lr.numpy()))\n\n# your other callbacks \ncallbacks.append(LRRecorder())\n<\/code><\/pre>\n\n<h3>Update<\/h3>\n\n<p><code>w := w - (base_lr*m\/sqrt(v))*grad = w - act_lr*grad<\/code>\nThe learning rate we get above is the <code>base_lr<\/code>. However, <code>act_lr<\/code> is adaptive changed during training. Take Adam optimizer as an example, <code>act_lr<\/code> is determined by <code>base_lr<\/code>, <code>m<\/code> and <code>v<\/code>. <code>m<\/code> and <code>v<\/code> are the first and second momentums of parameters. Different parameters have different <code>m<\/code> and <code>v<\/code> values. So if you would like to know the <code>act_lr<\/code>, you need to know the variable's name. For example, you want to know the <code>act_lr<\/code> of the variable <code>Adam\/dense\/kernel<\/code>, you can access the <code>m<\/code> and <code>v<\/code> like this,<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>for var in optimizer.variables():\n  if 'Adam\/dense\/kernel\/m' in var.name:\n    print(var.name, var.numpy())\n\n  if 'Adam\/dense\/kernel\/v' in var.name:\n    print(var.name, var.numpy())\n<\/code><\/pre>\n\n<p>Then you can easily calculate the <code>act_lr<\/code> using above formula. <\/p>\n"}]}
+{"QuestionId":58149843,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning><matrix-multiplication><keras-layer>","CreationDate":"2019-09-28T19:56:11.217","AcceptedAnswerId":null,"OwnerUserId":4515527.0,"Title":"How to create a trainable linear layer for input with unknown batch size? (Keras\/Tensorflow)","Body":"<p>I have an input x of the following shape [?, m, n] and I would like to create a new layer of form x<em>w + b that retain the original shape of x.\nso basically x<\/em>w should be an element-wise multiplication.<\/p>\n\n<p>but because the batch size is unknown I'm not quite sure how to keep w as the same dimension as x. <\/p>\n\n<p>I can't force w to be [?, m, n] and if w is [m, n] then I get error for trying multiplication between 3D and 2D<\/p>\n\n<p><strong>Solved:<\/strong> <\/p>\n\n<pre><code>class LinearLayer(Layer):\n    def __init__(self, **kwargs):\n        super(LinearLayer, self).__init__(**kwargs)\n\n    def build(self, input_shape):\n        self.w = self.add_weight(name='w',\n                                      shape=(input_shape[1], input_shape[2]),\n                                      initializer=RandomNormal(),\n                                      trainable=True)\n        self.b = self.add_weight(name='b',\n                                      shape=(input_shape[1], input_shape[2]),\n                                      initializer=RandomNormal(),\n                                      trainable=True)\n        super(LinearLayer, self).build(input_shape)\n\n    def call(self, x, **kwargs):\n        batch_size = tf.shape(x)[0]\n        w_expand = tf.expand_dims(self.w, axis=0)\n        w_tile = tf.tile(w_expand, multiples=[batch_size, 1, 1])\n        return tf.multiply(x, w_tile) + self.b\n\n    def compute_output_shape(self, input_shape):\n        return input_shape\n<\/code><\/pre>\n","answers":[{"AnswerId":"58151921","CreationDate":"2019-09-29T03:08:41.927","ParentId":null,"OwnerUserId":"1500585","Title":null,"Body":"<p>Make <code>w<\/code> to be [1,m,n], and then the multiplication with a [?,m,n] tensor should be straightforward. <\/p>\n\n<p>Even better, if you can use Tensorflow operations instead of pure Keras, <code>x*self.W<\/code> (or more explicitly <code>tf.math.multiply(x,self.W)<\/code>) should work just fine (Tensorflow's broadcasting is less lame than Keras').<\/p>\n"}]}
+{"QuestionId":58150186,"AnswerCount":1,"Tags":"<python><parallel-processing><multiprocessing><pytorch><fast-ai>","CreationDate":"2019-09-28T20:48:21.147","AcceptedAnswerId":null,"OwnerUserId":7298325.0,"Title":"Run inference on CPU using pytorch and multiprocessing","Body":"<p>I have trained a CNN model on GPU using FastAI (PyTorch backend). I am now trying to use that model for inference on the same machine, but using CPU instead of GPU. Along with that, I am also trying to make use of multiple CPU cores using the multiprocessing module. Now here is the issue,<\/p>\n\n<blockquote>\n  <p><strong>Running the code on single CPU (without multiprocessing) takes only 40 seconds to process nearly 50 \n  images<\/strong><\/p>\n  \n  <p><strong>Running the code on multiple CPUs using torch multiprocessing takes more than 6 minutes to process the same 50 images<\/strong><\/p>\n<\/blockquote>\n\n<pre><code>from torch.multiprocessing import Pool, set_start_method\nos.environ['CUDA_VISIBLE_DEVICES']=\"\"\nfrom fastai.vision import *\nfrom fastai.text import *\ndefaults.device = torch.device('cpu')\n\ndef process_image_batch(batch):\n\n    learn_cnn  = load_learner(scripts_folder, 'cnn_model.pkl')\n    learn_cnn.model.training = False    \n    learn_cnn.model = learn_cnn.model.eval()\n    # for image in batch: \n    #     prediction = ... # predicting the image here\n    #     return prediction\n\nif __name__ == '__main__':\n    #\n    # image_batches = ..... # retrieving the image batches (It is a list of 5 lists)\n    # n_processes = 5\n    set_start_method('spawn', force=True)\n    try:\n        pool = Pool(n_processes)\n        pool.map(process_image_batch, image_batches)\n    except Exception as e:\n        print('Main Pool Error: ', e)\n    except KeyboardInterrupt:\n        exit()\n    finally:\n        pool.terminate()\n        pool.join()\n<\/code><\/pre>\n\n<p>I am not sure what's causing this slowdown in multiprocessing mode. I've read a lot of posts discussing similar issue but couldn't find a proper solution anywhere.<\/p>\n","answers":[{"AnswerId":"58164748","CreationDate":"2019-09-30T09:11:41.033","ParentId":null,"OwnerUserId":"8119473","Title":null,"Body":"<p>I think you have done a very naive mistake here, you are reading the model object in the function which you are parallelizing.<\/p>\n\n<p>Meaning for every single image, you are reloading the model from the disk. \nDepending on your model object size, IO is gonna be more time consuming then running a forward step.<\/p>\n\n<p>Please consider reading the model once in the main thread and then make the object available for inference in the parallel function.     <\/p>\n"}]}
+{"QuestionId":58150403,"AnswerCount":1,"Tags":"<keras><python-3.6><autoencoder><tensorflow2.0>","CreationDate":"2019-09-28T21:21:52.177","AcceptedAnswerId":58150665.0,"OwnerUserId":7663393.0,"Title":"Obtaining hidden layer outputs in a denoising autoencoder using Keras","Body":"<p>I have built a <code>Sequential<\/code> Keras model with three layers: A <code>Gaussian Noise<\/code> layer, a hidden layer, and the output layer with the same dimension as the input layer.  For this, I'm using the <strong>Keras<\/strong> package that comes with <strong>Tensorflow 2.0.0-beta1<\/strong>.  Thus, I'd like to get the output of the hidden layer, such that I circumvent the <code>Gaussian Noise<\/code> layer since it's only necessary in the training phase.<\/p>\n\n<p>To achieve my goal, I followed the instructions in <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-the-output-of-an-intermediate-layer\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-the-output-of-an-intermediate-layer<\/a>, which are pretty much described in <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> too.<\/p>\n\n<p>I have tried the following example from the official Keras documentation:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from tensorflow import keras\nfrom tensorflow.keras import backend as K\n\ndae = keras.Sequential([\n    keras.layers.GaussianNoise( 0.001, input_shape=(10,) ),\n    keras.layers.Dense( 80, name=\"hidden\", activation=\"relu\" ),\n    keras.layers.Dense( 10 )\n])\n\noptimizer = keras.optimizers.Adam()\ndae.compile( loss=\"mse\", optimizer=optimizer, metrics=[\"mae\"] )\n\n# Here the fitting process...\n# dae.fit( \u00b7 )\n\n# Attempting to retrieve a decoder functor.\nencoder = K.function([dae.input, K.learning_phase()], \n                               [dae.get_layer(\"hidden\").output])\n<\/code><\/pre>\n\n<p>However, when <code>K.learning_phase()<\/code> is used to create the Keras backend functor, I get the error:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/anaconda3\/lib\/python3.6\/contextlib.py\", line 99, in __exit__\n    self.gen.throw(type, value, traceback)\n  File \"\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/keras\/backend.py\", line 534, in _scratch_graph\n    yield graph\n  File \"\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/keras\/backend.py\", line 3670, in __init__\n    base_graph=source_graph)\n  File \"\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/eager\/lift_to_graph.py\", line 249, in lift_to_graph\n    visited_ops = set([x.op for x in sources])\n  File \"\/anaconda3\/lib\/python3.6\/site-packages\/tensorflow_core\/python\/eager\/lift_to_graph.py\", line 249, in &lt;listcomp&gt;\n    visited_ops = set([x.op for x in sources])\nAttributeError: 'int' object has no attribute 'op'\n<\/code><\/pre>\n\n<p>The code works great if I don't include <code>K.learning_phase()<\/code>, but I need to make sure that the output from my hidden layer is evaluated over an input that is not polluted with noise (i.e. in \"test\" mode -- not \"training\" mode).<\/p>\n\n<p>I know my other option is to create a model from the original denoising autoencoder, but can anyone point me into why my approach from the officially documented functor creation fails?<\/p>\n","answers":[{"AnswerId":"58150665","CreationDate":"2019-09-28T22:05:49.067","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>Firstly, ensure your packages are up-to-date, as your script works fine for me. Second, <code>encoder<\/code> won't get the outputs - continuing from your snippet after <code># Here is the fitting process...<\/code>, <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = np.random.randn(32, 10) # toy data\ny = np.random.randn(32, 10) # toy labels\ndae.fit(x, y) # run one iteration\n\nencoder = K.function([dae.input, K.learning_phase()], [dae.get_layer(\"hidden\").output])\noutputs = [encoder([x, int(False)])][0][0] # [0][0] to index into nested list of len 1\nprint(outputs.shape)\n# (32, 80)\n<\/code><\/pre>\n\n<p>However, as of <strong>Tensorflow 2.0.0-rc2<\/strong>, this will <em>not<\/em> work with eager execution enabled - disable via:<\/p>\n\n<pre><code>tf.compat.v1.disable_eager_execution()\n<\/code><\/pre>\n"}]}
+{"QuestionId":58150480,"AnswerCount":0,"Tags":"<python><machine-learning><keras><neural-network><forecast>","CreationDate":"2019-09-28T21:33:53.693","AcceptedAnswerId":null,"OwnerUserId":4508829.0,"Title":"Nueral Net Forecast for specific time periods","Body":"<p>I would like to know conceptually how to set up an analysis to run a neural net, that would forecast a probability or sales for a specified time period. Specifically, how can I forecast if a given sold product will be brought back for a warranty claim in the next 6 months or n months.<\/p>\n\n<p>I have all original order data, as well as warranty claim data to work with. I can forecast overall probability or a claim, but don't understand how to adjust it to specific time periods.<\/p>\n\n<p>This is the neural net I'm using, but I think the solution will be in the pre-processing <\/p>\n\n<pre><code>classifier = Sequential()\n\nclassifier.add(Dense(units = 12000, kernel_initializer = 'uniform', activation = 'relu', input_dim = X_train.shape[1]))\nclassifier.add(Dropout(0.2))\nclassifier.add(BatchNormalization())\n\nclassifier.add(Dense(units = 500, kernel_initializer = 'uniform', activation = 'relu'))\nclassifier.add(Dropout(0.2))\nclassifier.add(BatchNormalization())\n\nclassifier.add(Dense(units = 250, kernel_initializer = 'uniform', activation = 'relu'))\nclassifier.add(Dropout(0.2))\nclassifier.add(BatchNormalization())\nclassifier.add(Dense(units = 1, kernel_initializer = 'uniform', activation = 'sigmoid'))\nclassifier.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])\n<\/code><\/pre>\n\n<p>I would expect the results of the NN to be interpreted as the probability of something occurring in a user defined time period. ie. there's an X percent chance of Y occurring in the next N months.<\/p>\n","answers":[]}
+{"QuestionId":58150519,"AnswerCount":2,"Tags":"<tensorflow><keras>","CreationDate":"2019-09-28T21:40:59.603","AcceptedAnswerId":58150624.0,"OwnerUserId":4597273.0,"Title":"ResourceExhaustedError: OOM when allocating tensor in Keras","Body":"<p>I am training a model with the following summary using tf.keras:<\/p>\n\n<pre><code>Model: \"sequential\"\n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nconv2d (Conv2D)              (None, 28, 28, 20)        520       \n_________________________________________________________________\nflatten (Flatten)            (None, 15680)             0         \n_________________________________________________________________\ndense (Dense)                (None, 5)                 78405     \n=================================================================\nTotal params: 78,925\nTrainable params: 78,925\nNon-trainable params: 0\n_________________________________________________________________ \n<\/code><\/pre>\n\n<p>I am calling the fit method using <\/p>\n\n<p><code>model.fit(X_train, y_train, batch_size=32, steps_per_epoch=125, epochs=5, use_multiprocessing=True)<\/code><\/p>\n\n<p>where X_train is a tensorflow Variable with shape [900000,32,32,1].<\/p>\n\n<p>I am running into the following error:<\/p>\n\n<pre><code>Resource exhausted: OOM when allocating tensor with shape[900000,28,28,20] and type float on \/job:localhost\/replica:0\/task:0\/device:GPU:0 by allocator GPU_0_bfc\n     [[{{node conv2d_1\/Conv2D}}]]\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.\n\n     [[metrics_2\/acc\/Identity\/_53]]\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.\n\n  (1) Resource exhausted: OOM when allocating tensor with shape[900000,28,28,20] and type float on \/job:localhost\/replica:0\/task:0\/device:GPU:0 by allocator GPU_0_bfc\n     [[{{node conv2d_1\/Conv2D}}]]\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.\n\n0 successful operations.\n0 derived errors ignored.\n<\/code><\/pre>\n\n<p>I am unable to understand why it's allocating a tensor with shape [900000,28,28,20] when the batch size is 32. I was expecting [32,28,28,20]. <\/p>\n\n<hr>\n\n<p>Full code:<\/p>\n\n<pre><code>IMAGE_SIZE = 32\nBATCH_SIZE = 32\n\nconv1_layer = keras.layers.Conv2D(input_shape=(IMAGE_SIZE, IMAGE_SIZE, 1), filters=20, kernel_size=[5, 5], activation='relu')\nf = keras.layers.Flatten()\noutput_layer = keras.layers.Dense(units=5, activation='softmax')\n\nmodel = keras.models.Sequential(layers=[conv1_layer, f, output_layer])\nmodel.compile(optimizer='sgd', loss='categorical_crossentropy', metrics=['accuracy'])\nmodel.fit(X_train, y_train, batch_size=BATCH_SIZE, steps_per_epoch=int(X_train.shape[0])\/\/BATCH_SIZE, epochs=5, use_multiprocessing=True)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58150624","CreationDate":"2019-09-28T21:57:57.717","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>The problem's most likely in <code>use_multiprocessing=True<\/code>, which expects inputs as a generator object (see <a href=\"https:\/\/keras.io\/models\/sequential\/#fit\" rel=\"nofollow noreferrer\">docs<\/a>) - so if feeding the <code>X_train<\/code> array doesn't immediately yield an error, it may be treating it as a generator by iterating over axis 0 and feeding all 900,000 samples at once - thus the error. Try <code>use_multiprocessing=False<\/code>, or feed your data with a <a href=\"https:\/\/stanford.edu\/~shervine\/blog\/keras-how-to-generate-data-on-the-fly\" rel=\"nofollow noreferrer\">generator<\/a>.<\/p>\n\n<p>Also, <code>steps_per_epoch<\/code> may be an additional source - omit them from <code>fit<\/code> - then, <em>evaluate<\/em> your <code>tf.Variable<\/code> inputs before passing them in, as <code>fit<\/code> doesn't handle that automatically.<\/p>\n"},{"AnswerId":"58150938","CreationDate":"2019-09-28T23:05:36.310","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>I will suggest to write your own training data generator or <code>keras.utils.Sequence<\/code>. Directly feed the entire training data is not a good idea especially when your dataset is very large. In additionally, <strong><code>use_multiprocessing<\/code> argument is used for generator or  keras.utils.Sequence input only<\/strong> as explained in <a href=\"https:\/\/keras.io\/models\/sequential\/#fit\" rel=\"nofollow noreferrer\">.fit() doc<\/a><\/p>\n\n<p>The following code is adapted from <a href=\"https:\/\/stanford.edu\/~shervine\/blog\/keras-how-to-generate-data-on-the-fly\" rel=\"nofollow noreferrer\">A detailed example of how to use data generators with Keras<\/a>.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport keras\n\nclass DataGenerator(keras.utils.Sequence):\n    'Generates data for Keras'\n    def __init__(self, list_IDs, labels, batch_size=32, dim=(32,32), n_channels=1, 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 __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, 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,] = np.load('data\/' + ID + '.npy')\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\nparams = {'dim': (32,32),\n          'batch_size': 32,\n          'n_classes': 5,\n          'n_channels': 1,\n          'shuffle': True}\n\ntrain_data_path = # (list of string) ['0302','0325','3103',...] 'data\/0302.npy' is the path where stores the '0302' image. \ntrain_labels = # (list of int) [1,0,4,...] corresponding labels of 900,000 images.\n\n# Generators\ntrain_gen = DataGenerator(train_data_path, train_labels, **params)\n\nmodel.fit_generator(generator=train_gen,\n                    epochs=5,\n                    use_multiprocessing=True,\n                    workers=2)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58150670,"AnswerCount":1,"Tags":"<python><machine-learning><keras><neural-network><regression>","CreationDate":"2019-09-28T22:06:53.793","AcceptedAnswerId":58150816.0,"OwnerUserId":10997265.0,"Title":"Evaluation of an ANN Linear Regression Model","Body":"<p>I am kind of newbie in Machine Learning business and recently I made my first ANN with Keras in Python. It's a Linear Regression Model with 5 features and 1 output. I made a plot with \"MSE\" and \"Loss function\" and these are the results. Can we say that it is a good model? In addition R^2 = 0.91 .\nIs this the right way?<\/p>\n\n<pre><code>classifier = Sequential()\n\nclassifier.add(Dense(5, input_dim=5,kernel_initializer='normal',activation='relu'))\n\nclassifier.add(Dense(5, activation='relu'))\n\nclassifier.add(Dense(1,activation='linear'))\n\n\nclassifier.compile(loss='mse', optimizer='adam', metrics=['mse','mae'])\n\nhistory = classifier.fit(X_train, y_train, batch_size=10, validation_data=(X_test, y_test), epochs=200, verbose=0)\n\ny_pred=classifier.predict(X_test)\n\ntrain_mse=classifier.evaluate(X_train, y_train, verbose=0)\n\nplt.title('Loss \/ Mean Squared Error')\nplt.plot(history.history['loss'], label='train')\nplt.plot(history.history['val_loss'], label='test')\nplt.legend()\nplt.show()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/SVYbE.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/SVYbE.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"58150816","CreationDate":"2019-09-28T22:37:11.730","ParentId":null,"OwnerUserId":"4685471","Title":null,"Body":"<p>Apart from some terminology details (NN regression is not <em>linear<\/em> regression, and usually we don't call such model a <code>classifier<\/code>), your model looks good indeed, with both errors (training &amp; test) reducing smoothly and no signs of <a href=\"https:\/\/stackoverflow.com\/questions\/56780450\/is-this-overfitting\/56805078#56805078\">overfitting<\/a>.<\/p>\n\n<p>Although  an R^2 value of 0.91 sounds pretty good, the use of the metric in <em>predictive<\/em> settings, like here, is quite problematic; quoting from my own answer in <a href=\"https:\/\/stackoverflow.com\/questions\/54614157\/scikit-learn-statsmodels-which-r-squared-is-correct\/54618898#54618898\">another SO thread<\/a>:<\/p>\n\n<blockquote>\n  <p>the whole R-squared concept comes in fact directly from the world of statistics, where the emphasis is on <em>interpretative<\/em> models, and it has little use in machine learning contexts, where the emphasis is clearly on <em>predictive<\/em> models; at least AFAIK, and beyond some very introductory courses, I have never (I mean <em>never<\/em>...) seen a predictive modeling problem where the R-squared is used for any kind of performance assessment; neither it's an accident that popular <em>machine learning<\/em> introductions, such as Andrew Ng's <a href=\"https:\/\/www.coursera.org\/learn\/machine-learning\" rel=\"nofollow noreferrer\">Machine Learning<\/a> at Coursera, do not even bother to mention it. And, as noted in the <a href=\"https:\/\/github.com\/scikit-learn\/scikit-learn\/issues\/5570\" rel=\"nofollow noreferrer\">Github thread<\/a> above (emphasis added):<\/p>\n  \n  <blockquote>\n    <p>In particular when using a <strong>test<\/strong> set, it's a bit unclear to me what the R^2 means.<\/p>\n  <\/blockquote>\n  \n  <p>with which I certainly concur.<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":58150760,"AnswerCount":0,"Tags":"<keras><neural-network><save><lstm><sequence-to-sequence>","CreationDate":"2019-09-28T22:26:04.030","AcceptedAnswerId":null,"OwnerUserId":7486814.0,"Title":"Keras, sequence-to-sequence, save, load, continue training","Body":"<p>I followed <a href=\"https:\/\/blog.keras.io\/a-ten-minute-introduction-to-sequence-to-sequence-learning-in-keras.html\" rel=\"nofollow noreferrer\">this tutorial from F. Chollet<\/a> and implemented and trained a sequence-to-sequence model in Keras. \nIt also works fine and I was able to save and load the model and make predictions. <\/p>\n\n<p>However, I wonder how I would proceed if I wanted to load the model and add some more rounds of training to it. <\/p>\n\n<p>I am fine with calling the fit-method after loading the model called <strong><em>model<\/em><\/strong> in the tutorial. But since there are two related models, namely <strong><em>encoder_model<\/em><\/strong> and <strong><em>decoder_model<\/em><\/strong> that are sharing states and inputs with <em>model<\/em> and that I need to perform the predictions, I doubt that this procedure actually works...\nRather, I guess that after serializing\/deserializing training on <em>model<\/em> does not affect <em>encoder_model<\/em> and <em>decoder_model<\/em>... <\/p>\n\n<p>But how would I proceed? How extract the necessary outputs and inputs from <em>model<\/em> to reconstruct or update <em>encoder_model<\/em> and <em>decoder_model<\/em>?\nThanks for any hints.<\/p>\n","answers":[]}
+{"QuestionId":58150861,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><keras-vggface>","CreationDate":"2019-09-28T22:49:07.600","AcceptedAnswerId":58151004.0,"OwnerUserId":10544891.0,"Title":"How to correctly concatenate a Flatten layer and a feature vector in Keras","Body":"<p><strong>I just need to concatenate a flatten layer and a feature vector in Keras. This is the code:<\/strong><\/p>\n\n<pre><code>#custom parameters\nn_features = 38\n\nvgg_model = VGGFace(include_top=False, input_shape=(224, 224, 3))\nlast_layer = vgg_model.get_layer('pool5').output\n\nx = Flatten(name='flatten')(last_layer)\n\n# feature vector\nfeature_vector = Input(shape = (n_features,))\n\nconc = concatenate(([x, feature_vector]), axis=1)\n\n\nlayer_intermediate = Dense(128, activation='relu', name='fc6')(conc)\nlayer_intermediate1 = Dense(32, activation='relu', name='fc7')(layer_intermediate)\nout = Dense(5, activation='softmax', name='fc8')(layer_intermediate1)\n\ncustom_vgg_model = Model(vgg_model.input, out)\n<\/code><\/pre>\n\n<p><strong>But I'm getting this error:<\/strong><\/p>\n\n<p>---> 20 custom_vgg_model = Model(vgg_model.input, out)<\/p>\n\n<p><em>ValueError: Graph disconnected: cannot obtain value for tensor Tensor(\"input_88:0\", shape=(?, 38), dtype=float32) at layer \"input_88\". The following previous layers were accessed without issue: ['input_87', 'conv1_1', 'conv1_2', 'pool1', 'conv2_1', 'conv2_2', 'pool2', 'conv3_1', 'conv3_2', 'conv3_3', 'pool3', 'conv4_1', 'conv4_2', 'conv4_3', 'pool4', 'conv5_1', 'conv5_2', 'conv5_3', 'pool5', 'flatten']<\/em><\/p>\n\n<p><strong>Btw the shape of the flatten layer is (None, 25088)<\/strong><\/p>\n","answers":[{"AnswerId":"58151004","CreationDate":"2019-09-28T23:20:46.230","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>Since your <code>feature_vector<\/code> is also <code>Input<\/code>. Try to add <code>feature_vector<\/code> into inputs when you define the Model. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>custom_vgg_model = Model([vgg_model.input,feature_vector], out)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58151285,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-29T00:33:18.680","AcceptedAnswerId":null,"OwnerUserId":10515435.0,"Title":"\"Tensor must be from the same graph as Tensor\" when calling keras model train_on_batch","Body":"<pre><code>gen_examples = self.generator.predict([xmal_batch, noise])\n<\/code><\/pre>\n\n<p>Generator is a keras model which adds noise to original data and generate new examples<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with tf.Session() as sess:         \n    sess.run(tf.global_variables_initializer())\n    sess.run(tf.local_variables_initializer())\n\n    saver.restore(sess, PATH)\n    print (\"load model from:\", PATH)\n\n    ymal_batch = sess.run(model.y_pred, feed_dict = {model.x_input: gen_examples})\n\n    self.substitute_detector.train_on_batch(gen_examples, ymal_batch)\n<\/code><\/pre>\n\n<blockquote>\n  <p>ValueError: Tensor(\"training\/Adam\/Const:0\", shape=(), dtype=int64) must be from \n  the same graph as Tensor(\"Adam\/iterations:0\", shape=(), dtype=resource).<\/p>\n<\/blockquote>\n\n<p>I'm new to Tensorflow and I don't know why there's this error, since <code>gen_examples<\/code> and <code>ymal_batch<\/code> are basically numpy arrays.<\/p>\n\n<p>It seems like there's new graph created, but I'm so confused.<\/p>\n","answers":[]}
+{"QuestionId":58151307,"AnswerCount":0,"Tags":"<c++><tensorflow>","CreationDate":"2019-09-29T00:37:41.510","AcceptedAnswerId":null,"OwnerUserId":4830432.0,"Title":"Create C++ Tensorflow Tensors from float arrays","Body":"<p>I am beginning to work with the C++ Tensorflow API. I can't seem to find any direct documentation on building Tensor objects from primitives.<\/p>\n\n<p>For instance, I would like to convert primitive float arrays to Tensors. How would I go about doing that?<\/p>\n","answers":[]}
+{"QuestionId":58151405,"AnswerCount":1,"Tags":"<python><machine-learning><keras><deep-learning><computer-vision>","CreationDate":"2019-09-29T01:05:03.867","AcceptedAnswerId":null,"OwnerUserId":12136797.0,"Title":"My CNN is over-fitting even in simple architecture and lots of data","Body":"<p>I have created a CNN for multi-labelled images, (of different shoes) and their size is 100x100x3. I have 50000 images,with 5 classes (which I think are enough), but the issue is, while training:<\/p>\n\n<ul>\n<li><p>Both test accuracy and validation accuracy rise together in the start but Val acc becomes almost constant at 58% approx, and the train acc keeps on increasing up to 85-99% approx.<\/p><\/li>\n<li><p>seems like over fitting, I have tried using drop outs, l1 regularization etc. <\/p><\/li>\n<\/ul>\n\n<p>Here is my architecture. What could be wrong with this?<\/p>\n\n<pre><code>**Conv Layer , 16 ,(3,3), input =100,100,3 ,padding = same**\n**LeakyRelu and MaxPooling()**\n**Dropouts **\n**Conv layer , 32, (9,9), pd = same** \n**LeakyRelu and maxpooling()**\n**Dropouts**\n**Flatten()**\n**Dense 5, leakyRelu, l1(0.001), dropout**\n**Dense 5, leakyRelu,l1(0.001)**\n**Sigmoid**\n<\/code><\/pre>\n","answers":[{"AnswerId":"58152690","CreationDate":"2019-09-29T05:59:23.953","ParentId":null,"OwnerUserId":"9149062","Title":null,"Body":"<p>A few things you can try:<\/p>\n\n<ul>\n<li>a filter of size <code>9x9<\/code> is not that common, certainly not for such small images. I'd use <code>3x3<\/code> in that case<\/li>\n<li>have you tried just using ReLU instead of its leaky version? I think it's better to start with the simple version of a nonlinearity and later on try different versions to see which performs best (and if there <em>is<\/em> any difference)<\/li>\n<li>I would drop the L1 regularizations and only add them later on to see if they help. First try to get reasonable results before you try to get state of art results<\/li>\n<li>currently you have a dropout after each layer, which seems to be a lot to me. Maybe you can try removing the first one? Also: what are your dropout rates?<\/li>\n<li>you have a pooling layer after each convolutional layer, have you tried removing the first one? As with the dropouts, you seem to be over-using the pooling layers.<\/li>\n<li>your first fully-connected (dense) layer has very few nodes. Note that the output of the <code>Flatten()<\/code> layer is 20000-dimensional (twice pooling -> <code>100\/2\/2 x 100\/2\/2 x 32 = 20000<\/code>) so it would make more sense to use a larger number of nodes in the layer between the flatten and the last fully-connected. I would suggest to try 1024 or even 256 would be better I guess.<\/li>\n<\/ul>\n\n<p>Also you have very little convolutional layers. Even for the CIFAR-10 dataset (60000 images of size <code>32x32x3<\/code> for 10 classes) I would use at least 2 blocks of 2 convolutional layers each, something like this:<\/p>\n\n<pre><code>input(32, 32, 3)\n\nconv_2D(32, (3,3), activation='relu', padding='same')\nconv_2D(32, (3,3), activation='relu', padding='same')\ndropout(0.25)\nmaxpooling\n\nconv_2D(64, (3,3), activation='relu', padding='same')\nconv_2D(64, (3,3), activation='relu', padding='same')\ndropout(0.25)\nmaxpooling\n\nflatten\ndense(1024, activation='relu')\ndense(10, activation='sigmoid')\n<\/code><\/pre>\n"}]}
+{"QuestionId":58151501,"AnswerCount":0,"Tags":"<machine-learning><keras><keras-layer>","CreationDate":"2019-09-29T01:30:06.737","AcceptedAnswerId":null,"OwnerUserId":12136837.0,"Title":"Keras: Identifying relationship between shared layers","Body":"<p>I have a network which is like this..<\/p>\n\n<pre><code>InputA      InputB\n  \\           \/\n   Shared Layer\n    \/        \\\nOutput   Output for\nfor A        B\n<\/code><\/pre>\n\n<p>Putting in a different form, the above is equivalent of<\/p>\n\n<pre><code>    InputA          InputB\n      |               |\n   Shared Layer   Duplicate of Shared Layer\n       |              |\n    Output         Output for\n    for A              B\n<\/code><\/pre>\n\n<p>With the original network, I think there must be some way to identify how Shared Layer is connected to output A and Input A, then input B and output B<\/p>\n\n<p>But I just couldnt figure out from the documentation<\/p>\n\n<p>May I ask for your help?<\/p>\n","answers":[]}
+{"QuestionId":58151507,"AnswerCount":1,"Tags":"<python><pytorch><normalize>","CreationDate":"2019-09-29T01:31:04.473","AcceptedAnswerId":58151903.0,"OwnerUserId":9066532.0,"Title":"Why Pytorch officially use mean=[0.485, 0.456, 0.406] and std=[0.229, 0.224, 0.225] to normalize images?","Body":"<p>In this page (<a href=\"https:\/\/pytorch.org\/docs\/stable\/torchvision\/models.html\" rel=\"nofollow noreferrer\">https:\/\/pytorch.org\/docs\/stable\/torchvision\/models.html<\/a>), it says that \"All pre-trained models expect input images normalized in the same way, i.e. mini-batches of 3-channel RGB images of shape (3 x H x W), where H and W are expected to be at least 224. The images have to be loaded in to a range of [0, 1] and then normalized using <code>mean = [0.485, 0.456, 0.406]<\/code> and <code>std = [0.229, 0.224, 0.225]<\/code>\". <\/p>\n\n<p>Shouldn't the usual <code>mean<\/code> and <code>std<\/code> of normalization be <code>[0.5, 0.5, 0.5]<\/code> and <code>[0.5, 0.5, 0.5]<\/code>? Why is it setting such strange values?<\/p>\n","answers":[{"AnswerId":"58151903","CreationDate":"2019-09-29T03:03:04.713","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>Using the mean and std of Imagenet is a common practice. They are calculated based on millions of images. If you want to train from scratch on your own dataset, you can calculate the new mean and std. Otherwise, using the Imagenet pretrianed model with its own mean and std is recommended. <\/p>\n"}]}
+{"QuestionId":58151839,"AnswerCount":0,"Tags":"<python><keras><deep-learning><resnet><transfer-learning>","CreationDate":"2019-09-29T02:47:50.710","AcceptedAnswerId":null,"OwnerUserId":11726179.0,"Title":"Validation Accuracy Static and Validation Loss Increasing Keras Resnet50 Transfer Learning","Body":"<p>I'm currently trying to use transfer learning to train a cnn on the cifar10 dataset using a frozen renset50 model with weights from imagenet.<\/p>\n\n<p>Using this code:<\/p>\n\n<pre><code>import keras\nfrom keras import optimizers\nimport numpy as np\nfrom keras.optimizers import SGD\nfrom keras.datasets import cifar10\nfrom keras.applications.resnet50 import ResNet50\nfrom keras.layers import GlobalAveragePooling2D, Dense\nfrom keras.preprocessing.image import ImageDataGenerator\nfrom keras.models import Model\nfrom skimage.transform import resize\nfrom IPython import embed\n\nNUM_CLASSES = 10\nBATCH_SIZE = 256\nNUM_EPOCHS = 15\n\n(x_train, y_train), (x_test, y_test) = cifar10.load_data()\n\n# Convert 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\n# Normalize the data\nx_train = x_train.astype('float32')\nx_test = x_test.astype('float32')\nx_train \/= 255\nx_test \/= 255\n\nbase_model = ResNet50(include_top=False, weights='imagenet')\n\nfor layer in base_model.layers:\n    layer.trainable=False\n\n# add a global spatial average pooling layer\nx = base_model.output\nx = GlobalAveragePooling2D()(x)\n# let's add a fully-connected layer\nx = Dense(64, activation='relu')(x)\nx = Dropout(0.5)(x)\n# and a logistic layer -- 10 classes for CIFAR10\nx = Dense(64, activation='relu')(x)\npredictions = Dense(NUM_CLASSES, activation='softmax')(x)\n\n# this is the model we will train\nmodel = Model(inputs=base_model.input, outputs=predictions)\n\n# Let's train the model using RMSprop\nmodel.compile(loss='binary_crossentropy', optimizer='sgd', metrics=['accuracy'])\n\nmodel.fit(x_train, y_train, batch_size=BATCH_SIZE, epochs=NUM_EPOCHS, validation_data=(x_test, y_test), shuffle=False)\n<\/code><\/pre>\n\n<p>Here is the training output:<\/p>\n\n<pre><code>Train on 50000 samples, validate on 10000 samples\nEpoch 1\/15\n50000\/50000 [==============================] - 162s 3ms\/step - loss: 0.3966 - acc: 0.8868 - val_loss: 0.3543 - val_acc: 0.9000\nEpoch 2\/15\n50000\/50000 [==============================] - 151s 3ms\/step - loss: 0.3520 - acc: 0.8959 - val_loss: 0.3429 - val_acc: 0.9000\nEpoch 3\/15\n50000\/50000 [==============================] - 152s 3ms\/step - loss: 0.3393 - acc: 0.8979 - val_loss: 0.3385 - val_acc: 0.9000\nEpoch 4\/15\n50000\/50000 [==============================] - 151s 3ms\/step - loss: 0.3334 - acc: 0.8987 - val_loss: 0.3365 - val_acc: 0.9000\nEpoch 5\/15\n50000\/50000 [==============================] - 152s 3ms\/step - loss: 0.3286 - acc: 0.8991 - val_loss: 0.3356 - val_acc: 0.9000\nEpoch 6\/15\n50000\/50000 [==============================] - 151s 3ms\/step - loss: 0.3251 - acc: 0.8994 - val_loss: 0.3349 - val_acc: 0.9000\nEpoch 7\/15\n50000\/50000 [==============================] - 150s 3ms\/step - loss: 0.3223 - acc: 0.8994 - val_loss: 0.3345 - val_acc: 0.9000\nEpoch 8\/15\n50000\/50000 [==============================] - 152s 3ms\/step - loss: 0.3194 - acc: 0.8995 - val_loss: 0.3341 - val_acc: 0.9000\nEpoch 9\/15\n50000\/50000 [==============================] - 151s 3ms\/step - loss: 0.3175 - acc: 0.8996 - val_loss: 0.3338 - val_acc: 0.9000\nEpoch 10\/15\n50000\/50000 [==============================] - 153s 3ms\/step - loss: 0.3151 - acc: 0.8997 - val_loss: 0.3337 - val_acc: 0.9000\n50000\/50000 [==============================] - 153s 3ms\/step - loss: 0.3132 - acc: 0.8997 - val_loss: 0.3337 - val_acc: 0.9000\nEpoch 12\/15\n50000\/50000 [==============================] - 153s 3ms\/step - loss: 0.3107 - acc: 0.8998 - val_loss: 0.3339 - val_acc: 0.9000\nEpoch 13\/15\n50000\/50000 [==============================] - 157s 3ms\/step - loss: 0.3089 - acc: 0.8998 - val_loss: 0.3341 - val_acc: 0.9000\nEpoch 14\/15\n50000\/50000 [==============================] - 155s 3ms\/step - loss: 0.3070 - acc: 0.8998 - val_loss: 0.3342 - val_acc: 0.9000\nEpoch 15\/15\n50000\/50000 [==============================] - 154s 3ms\/step - loss: 0.3056 - acc: 0.8998 - val_loss: 0.3346 - val_acc: 0.9000\n<\/code><\/pre>\n\n<p>The loss is decreasing, and accuracy increasing. However, the validation loss is decreasing ONLY up to the 13th epoch, and the validation accuracy is not moving at all!<\/p>\n\n<p>What could be causing this reversal of validation loss in the 13th epoch and why is the validation accuracy static? <\/p>\n\n<p>Any help would be amazing!<\/p>\n","answers":[]}
+{"QuestionId":58151875,"AnswerCount":0,"Tags":"<tensorflow><cublas>","CreationDate":"2019-09-29T02:54:16.753","AcceptedAnswerId":null,"OwnerUserId":9873377.0,"Title":"How to get cublas handle in tensorflow?","Body":"<p>I write a custom op using cublas function <code>cublasCgetrfBatched<\/code> and <code>cublasCgetriBatched<\/code>, the functions use cublas handle as a input param, however the <code>cublasCreate(&amp;handle);<\/code> cost nearly 100ms. \nI think the TF has already integrate CUBLAS module, <code>cublasCreate(&amp;handle)<\/code> must have been invoked in the init process, then how to get the handle? An example may be the best!<\/p>\n","answers":[]}
+{"QuestionId":58152118,"AnswerCount":0,"Tags":"<python><pandas><numpy><keras>","CreationDate":"2019-09-29T03:55:58.070","AcceptedAnswerId":null,"OwnerUserId":12128685.0,"Title":"ValueError: Error when checking target: expected lstm_2 to have 3 dimensions, but got array with shape (1, 184800)","Body":"<p>My data consists of 26 columns and 231000 rows. I have tried to fix the error for a few days now but I am still unsure on what needs to be done. Please do give your insight on this problem. Below is my full code <\/p>\n\n<pre><code>import pandas as pd\nimport numpy as np\nfrom sklearn.model_selection import train_test_split\nfrom sklearn import datasets, linear_model\nfrom matplotlib import pyplot as plt\nfrom keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.layers import LSTM\nfrom keras.layers import RepeatVector\nfrom keras.layers import TimeDistributed\n\n# load dataset\ndataframe = pd.read_csv(\"AHPS_26C_Norm.csv\", header=None)\ndataset = dataframe.values\nX = dataset[1:, 1:26].astype(float)\nY = dataset[1:, 0]\n\nX = np.array(X)\nY = np.array(Y)\nX_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size=0.20, random_state=42)\n\nn_steps = len(X_train)\nn_features = 25\nbatch_size = 1\nX_train = X_train.reshape(1, n_steps, n_features)\nY_train = Y_train.reshape(1, n_steps)\n\n# create model\nmodel = Sequential()\nmodel.add(LSTM(100, activation='relu', batch_input_shape=(1, n_steps, n_features)))\nmodel.add(RepeatVector(n_steps))\nmodel.add(LSTM(100, activation='softmax', return_sequences=True))\nmodel.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])\n\n# fit model\nmodel.fit(X_train, Y_train, epochs=300, validation_data = (X_test, Y_test), batch_size= batch_size, verbose=0)\nprint(model.summary())\n\n\n# Final evaluation of the model\nscores = model.evaluate(X_train, Y_train, verbose=0)\nprint(\"Accuracy: %.2f%%\" % (scores[1] * 100))\n<\/code><\/pre>\n\n<p>Below is the error I received:<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"C:\/Users\/Iliya Juhari\/.PyCharmCE2018.3\/config\/scratches\/scratch_3.py\", line 40, in &lt;module&gt;\n    model.fit(X_train, Y_train, epochs=300, validation_data = (X_test, Y_test), batch_size= batch_size, verbose=0)\n  File \"C:\\Users\\Iliya Juhari\\PycharmProjects\\untitled\\venv\\lib\\site-packages\\keras\\engine\\training.py\", line 1154, in fit\n    batch_size=batch_size)\n  File \"C:\\Users\\Iliya Juhari\\PycharmProjects\\untitled\\venv\\lib\\site-packages\\keras\\engine\\training.py\", line 621, in _standardize_user_data\n    exception_prefix='target')\n  File \"C:\\Users\\Iliya Juhari\\PycharmProjects\\untitled\\venv\\lib\\site-packages\\keras\\engine\\training_utils.py\", line 135, in standardize_input_data\n    'with shape ' + str(data_shape))\nValueError: Error when checking target: expected lstm_2 to have 3 dimensions, but got array with shape (1, 184800)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58152158,"AnswerCount":0,"Tags":"<python><tensorflow><jupyter-notebook><artificial-intelligence><google-colaboratory>","CreationDate":"2019-09-29T04:06:48.357","AcceptedAnswerId":null,"OwnerUserId":12126174.0,"Title":"ipykernel_launcher.py error on google colab","Body":"<p>I'm trying to recreate <a href=\"https:\/\/github.com\/gentaiscool\/lstm-attention\" rel=\"nofollow noreferrer\">this project<\/a> on Google Colab (because it has Tensorflow\/Keras built-in).\nWhen running the main file I get this error:<\/p>\n\n<pre><code>Usage: ipykernel_launcher.py [options]\n\nipykernel_launcher.py: error: no such option: -f\nAn exception has occurred, use %tb to see the full traceback.\n\nSystemExit: 2\n<\/code><\/pre>\n\n<p>I'm not sure what it's referring to as I don't have any files like that in the folder. I'm new to AI\/TF\/Keras and Google Colab so any help\/suggestions would be great. <\/p>\n","answers":[]}
+{"QuestionId":58152167,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2019-09-29T04:09:19.427","AcceptedAnswerId":null,"OwnerUserId":12009032.0,"Title":"How to get the detected result to do somethings in object detection API?","Body":"<p>I have done the object detection using webcam. Can I know, how to do when the system detected some specific object(Gun\/knife), the system will alert the user by using alert sound<\/p>\n\n<p>have a nice day <\/p>\n\n<p>I have imported pygame be my sound system, but when I combine the result and pygame it's doesn't work.<\/p>\n\n<pre><code>image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')\ndetection_boxes = detection_graph.get_tensor_by_name('detection_boxes:0')\ndetection_scores =detection_graph.get_tensor_by_name('detection_scores:0')\ndetection_class=detection_graph.get_tensor_by_name('detection_classes:0')\nnum_detections = detection_graph.get_tensor_by_name('num_detections:0')\n\nvideo = cv2.VideoCapture(0)\nret = video.set(3,1280)\nret = video.set(4,720)\nwhile(True):\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\nret, frame = video.read()\nframe_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')\nvis_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=0.60)\n\n# All the results have been drawn on the frame, so it's time to display it.\ncv2.imshow('Object detector', frame)\n\n# Press 'q' to quit\nif cv2.waitKey(1) == ord('q'):\n    break\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58152402,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2019-09-29T05:08:43.977","AcceptedAnswerId":null,"OwnerUserId":4160995.0,"Title":"How to create hybrid sparse tensor?","Body":"<p>I'm porting a model from pytorch to tensorflow. Pytorch has type of tensor called hybrid sparse which mean only the first few dimensions are sparse, the last dimensions can be dense. Is there any thing like that in tensorflow? I've checked tf.sparse.SparseTensor api and it doesn't support hybrid tensor. How can i get around this? I'm doing relation embedding and node embedding on knowledge graph so i need to work with hybrid sparse tensors.<\/p>\n\n<p>I'm thinking about separate the hybrid sparse tensor into multiple 'normal' sparse tensors, do some computation and concat the result back together. I'm wondering if this is a good work around?<\/p>\n","answers":[]}
+{"QuestionId":58152427,"AnswerCount":0,"Tags":"<python><arrays><tensorflow><keras>","CreationDate":"2019-09-29T05:12:28.930","AcceptedAnswerId":null,"OwnerUserId":11724935.0,"Title":"How to enter array into a Dense neural network?","Body":"<p>I try to build a deep reinforcement learning whit neural network but I have trouble to enter the state of the game into the neural network I always get this <strong>error:<\/strong>\n<code>ValueError: setting an array element with a sequence.<\/code>\nThis is how the state look like:<\/p>\n\n<pre><code>[[array([[3775.74129204, 1946.76168772],\n       [3121.9310053 ,  699.72903429],\n       [4366.58744074,  735.83476952]])\n  array([[-2318.83417005,  1468.79312881],\n       [-2212.77983243,   366.27498107]])\n  array([[-3124.,   800.],\n       [ -645.,   800.]]) -24.0]]\n<\/code><\/pre>\n\n<p>Full code(this isn't the full class this just the important thing that do the trouble):<\/p>\n\n<pre><code>class DQNAgent:\n    def __init__(self, state_size, action_size):\n        self.state_size = state_size\n        self.action_size = action_size\n        self.memory = deque(maxlen=2000)\n        self.gamma = 0.95    # discount rate\n        self.epsilon = 1.0  # exploration rate\n        self.epsilon_min = 0.01\n        self.epsilon_decay = 0.995\n        self.learning_rate = 0.001\n        self.model = self._build_model()\n\n    def _build_model(self):\n        # Neural Net for Deep-Q learning Model\n        model = Sequential()\n        model.add(Dense(24, input_shape=(self.state_size,) ,activation='relu'))\n        model.add(Dense(24, activation='relu'))\n        model.add(Dense(self.action_size, activation='linear'))\n        model.compile(loss='mse',\n                      optimizer=Adam(lr=self.learning_rate))\n        return model\n\n    def remember(self, state, action, reward, next_state, done):\n        self.memory.append((state, action, reward, next_state, done))\n    def replay(self, batch_size):\n        minibatch = random.sample(self.memory, batch_size)\n        for state, action, reward, next_state, done in minibatch:\n            target = reward\n            if not done:\n                print(next_state)\n                target = (reward + self.gamma *\n                          np.amax(self.model.predict(next_state)[0]))\n            target_f = self.model.predict(state)\n            target_f[0][action] = target\n            self.model.fit(state, target_f, epochs=1, verbose=0)\n        if self.epsilon &gt; self.epsilon_min:\n            self.epsilon *= self.epsilon_decay\nif __name__ == \"__main__\":\n    Init()\n    done=False\n    state_size=4\n    action_size=4\n    agent = DQNAgent(state_size, action_size)\n    batch_size = 32\n    EPISODES=50\n    for e in range(EPISODES):\n        for stpf in range(5):\n            r_locs, i_locs, c_locs, ang, score = Game_step(1)\n            state=np.array([r_locs, i_locs, c_locs,ang])\n            state = np.reshape(state, [1, state_size])\n        for stp in range(100):\n            action =  agent.act(state)\n            r_locs, i_locs, c_locs, ang, score = Game_step(action)\n            next_state=np.array([r_locs, i_locs, c_locs,ang])\n            reward=score\n            next_state = np.reshape(next_state, [1, state_size])\n            agent.remember(state, action, reward, next_state, done)\n            state = next_state\n            if len(agent.memory) &gt; batch_size:\n                agent.replay(batch_size)\n        print(\"episode: {}\/{}, score: {}\".format(e, EPISODES))\n        done=True\n<\/code><\/pre>\n\n<p>How I fix it?\nThanks!<\/p>\n","answers":[]}
+{"QuestionId":58152849,"AnswerCount":1,"Tags":"<r><tensorflow><keras>","CreationDate":"2019-09-29T06:27:42.647","AcceptedAnswerId":58170770.0,"OwnerUserId":10466306.0,"Title":"Get activation functions (and hyperparamters) of trained model keras\/tensorflow in R Studio","Body":"<p>I have trained a model with keras\/tensorflow in R Studio and stored it.\nI now reload the model with <\/p>\n\n<pre><code> my_model &lt;- load_model_hdf5(\"my_model.h5\")\n<\/code><\/pre>\n\n<p>with<\/p>\n\n<pre><code>summary(my_model)\n<\/code><\/pre>\n\n<p>I get a summary of the model, including number and size of layers.\nYet, I don't see what the activation functions are.<\/p>\n\n<p>Is there a way to the access the activation functions?<\/p>\n\n<p>Secondly, is there also a way to access the hyperparameters as epoch number, batch size,.. that were used in training this model?<\/p>\n","answers":[{"AnswerId":"58170770","CreationDate":"2019-09-30T15:16:25.427","ParentId":null,"OwnerUserId":"12118774","Title":null,"Body":"<p>You can check <a href=\"https:\/\/lutzroeder.github.io\/netron\/\" rel=\"nofollow noreferrer\">Netron<\/a> with loading your <code>.h5<\/code> file. Netron is highly useful. <\/p>\n"}]}
+{"QuestionId":58153085,"AnswerCount":1,"Tags":"<python><python-3.x><keras><deep-learning><conv-neural-network>","CreationDate":"2019-09-29T07:08:36.473","AcceptedAnswerId":58154199.0,"OwnerUserId":7390581.0,"Title":"ImageDataGenerator with additional text input","Body":"<p>The architecture I'm trying to implement is here:\n<a href=\"https:\/\/i.stack.imgur.com\/ky5Cq.png\" rel=\"nofollow noreferrer\">Patient-data adapted model architecture: ResNet-50<\/a>. My images are split into folders by labels as such:<\/p>\n\n<pre><code>root\/\n    \u251c\u2500\u2500 train\/\n    \u2502   \u251c\u2500\u2500 class1\/\n    \u2502   \u251c\u2500\u2500 class2\/\n    \u2502   ...\n    \u2502\n    \u2514\u2500\u2500 validation\/\n       \u251c\u2500\u2500 class1\/\n       \u251c\u2500\u2500 class2\/\n       ...\n<\/code><\/pre>\n\n<p>I also have a CSV file that contains the image name, image labels (an image can have multiple class labels) and additional information:<\/p>\n\n<pre><code>+--------+---------------+-------+------+\n| File  |    Labels     | Info1 | Info2 |\n+-------+---------------+-------+-------+\n| 1.png | class1        | 0.512 |     1 |\n| 2.png | class2        |   0.4 |     0 |\n| 3.png | class1|class2 |  0.64 |     1 |\n+-------+---------------+-------+-------+\n<\/code><\/pre>\n\n<p>My network model has two inputs, one which will be used to process the image and another one which will concatenate to the final layer before the dense layer:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>input_shape = (img_height, img_width, 1)\n\nimg_input= Input(input_shape)\nvec_input = Input((2,))\n\nres = ZeroPadding2D((3, 3))(img_input)\n\n# Processing ...\n\nres = Flatten()(res)\nres = Concatenate()([res, vec_input])\nres = Dense(classes, activation='softmax', name='fc' + str(classes))(res)\n<\/code><\/pre>\n\n<p>To get the images I am using ImageDataGenerator with flow_from_directory, which works fine for getting only the image data:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>validation_datagen = ImageDataGenerator(rescale=1. \/ 255)\nvalidation_generator = validation_datagen.flow_from_directory(\n        validation_dir,\n        target_size=(target_size, target_size),\n        batch_size=batch_size,\n        class_mode=class_mode,\n        color_mode=color_mode)\n\n# Similarly for the train data generator ...\n\n# Train the model using above defined data generators\nhistory = model.fit_generator(\n    train_generator,\n    epochs=epochs,\n    validation_data=validation_generator)\n\n<\/code><\/pre>\n\n<p>I now need to use the additional information for each image as the vec_input in my model. I have looked at using flow_from_dataframe and creating custom generators, but am unsure how to go about this. I can restructure the images by putting them in the same folder if needed, although then I suppose I cannot use flow_from_directory. Any ideas on how I can achieve this?<\/p>\n\n<p>EDIT:<\/p>\n\n<p>In case somebody needs a solution, here's what I've been able to come up with:<\/p>\n\n<pre><code>class CustomSequenceGenerator(Sequence):\n\n    def __init__(self, image_dir, csv_file_path, label_path, dim=448, batch_size=8,\n                 n_classes=15, n_channels=1, vec_size=3, shuffle=True):\n        # Keras generator\n        self.image_dir = image_dir\n        self.image_file_list = os.listdir(image_dir)\n        self.batch_size = batch_size\n        self.csv_file = pd.read_csv(csv_file_path)\n        self.n_classes = n_classes\n        self.dim = dim\n        self.n_channels = n_channels\n        self.shuffle = shuffle\n        self.vec_size = vec_size\n        self.labels = get_class_labels(label_path)\n        self.labels_dict = dict(zip(self.labels, range(0, len(self.labels))))\n\n        self.csv_file.set_index('File', inplace=True, drop=True)\n\n    def __len__(self):\n        \"\"\"It is mandatory to implement it on Keras Sequence\"\"\"\n        return int(np.ceil(len(self.image_file_list) \/ float(self.batch_size)))\n\n    def __getitem__(self, index):\n\n        # Generate indexes of the batch\n        samples = self.image_file_list[index * self.batch_size:(index + 1) * self.batch_size]\n\n        x, y = self.__data_generation(samples, index)\n        return x, y\n\n    def __data_generation(self, samples, start_index):\n\n        x_batch_image = np.empty((self.batch_size, self.dim, self.dim, self.n_channels))\n        x_batch_vector = np.empty((self.batch_size, self.vec_size))\n        y_batch = np.empty((self.batch_size, self.n_classes))\n        self.csv_file.reindex()\n        for i, sample in enumerate(samples):\n            image_file_path = self.image_dir + \"\/\" + sample\n            image = self.preprocess_image(Image.open(image_file_path), 448)\n\n            features, labels = self.preprocess_csv(self.csv_file, sample, self.labels_dict, self.n_classes)\n            x_batch_image[i] = image\n            x_batch_vector[i] = features\n            y_batch[i] = labels\n\n        return [x_batch_image, x_batch_vector], y_batch\n<\/code><\/pre>\n","answers":[{"AnswerId":"58154199","CreationDate":"2019-09-29T10:05:08.497","ParentId":null,"OwnerUserId":"4332585","Title":null,"Body":"<p>I think the best way to achieve this would be to implement a custom <a href=\"https:\/\/keras.io\/utils\/#sequence\" rel=\"nofollow noreferrer\"><code>Sequence<\/code> object<\/a>, potentially inheriting the methods of <code>ImageDataGenerator<\/code>. Maybe what you need doesn't require all the complexity of <code>ImageDataGenerator<\/code> (i.e. random transforms, image saving, interpolation) and in this case you wouldn't need to inherit it.<\/p>\n"}]}
+{"QuestionId":58153146,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><multiprocessing>","CreationDate":"2019-09-29T07:18:34.067","AcceptedAnswerId":null,"OwnerUserId":8958254.0,"Title":"Already fit model taking space in GPU memory even without loading","Body":"<pre><code>def get_model():\n  model = Sequential()\n  model.add(Dense(20, input_shape=(1,), activation='relu'))\n  model.add(Dense(1))\n  print(model.summary())\n  model.compile(loss='mse', optimizer='adam')\n  return model\n\ndef train_model():\n  \/\/ X and y are input and output labels generated using some logic\n  epochs = 20\n  batch_size = 1000\n  verbose = 1\n  queue = multiprocessing.Queue()\n  validation_split = 0.2\n  model = get_model()\n  p1 = Process(target = fit_model, args=(model, X, y, validation_split, \n        epochs, batch_size, verbose, queue))\n  p1.start()\n  p1.join()\n  time.sleep(7)\n  queue.get()\n  \/\/check nvidia-smi after this\n  #model_name = str(time.time()).split('.')[0]\n  #pickle.dump(model, open('nn_sq_'+model_name, 'wb'))\n\ndef fit_model(model, X, y, validation_split, epochs, batch_size, verbose, \n                queue):\n  h = model.fit(X, y, validation_split=0.2,\n               epochs=epochs,\n               batch_size=batch_size,\n               verbose=1)\n  queue.put(model)\n\nif __name__ == '__main__' :\n    train_model()\n<\/code><\/pre>\n\n<p>Consider the code above. In the above code:<\/p>\n\n<p>1) calling train_model in main method<\/p>\n\n<p>2) train_model calls another process for model.fit for fit_model method<\/p>\n\n<p>3) Once that process completes, I pass the fit model to the queue to pass back to the main process so that the child process is completed and GPU memory can be freed.<\/p>\n\n<p>But one problem which I am facing is that if in the parent process (train_model method) if I do queue.get(model), the GPU memory becomes filled up again(watching nvidia-smi) and doesn't get freed up.<\/p>\n\n<p>A similar thing happens if I just try to dump the model using pickle (commented code). Even if I don't call multiprocessing method and directly try to pickle the model, I get memory filled up again.<\/p>\n\n<p>If I pickle the model in the child process(fit_model) itself, then there is no problem and I don't face any issues with memory fill.<\/p>\n\n<p>So, in short, my question is: when the child process gets completed, then memory gets freed, then why does going back to main process and just getting the model from the queue or pickling the model object fills up the GPU memory?<\/p>\n","answers":[]}
+{"QuestionId":58153177,"AnswerCount":0,"Tags":"<keras><loss-function><objective-function>","CreationDate":"2019-09-29T07:26:15.483","AcceptedAnswerId":null,"OwnerUserId":11544257.0,"Title":"Do I need one alpha and beta for Tversky loss for all classes, or I need a pair of alpha and beta for each class?","Body":"<p>I want to use Tversky loos for my multi class segmentation problem (Need to segment many classes in the image. Pixels are not multi label).\nIn the paper - <a href=\"https:\/\/arxiv.org\/pdf\/1706.05721.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1706.05721.pdf<\/a> said they used two hyper parameters alpha and beta.<\/p>\n\n<p>My question is do I need to use one alpha and beta for all classes or I need to pass pair of alpha and beta for each class ?<\/p>\n","answers":[]}
+{"QuestionId":58153205,"AnswerCount":1,"Tags":"<python><opencv><tensorflow><object-detection><tflearn>","CreationDate":"2019-09-29T07:32:22.693","AcceptedAnswerId":null,"OwnerUserId":11631791.0,"Title":"How to use tflearn model to perform blob image processing in python","Body":"<p>I want to find blob in a frame (or image), and apply the trained DNN to classify the regions.\nHowever, I am using tflearn, and I don't know how to use the model trained using tflearn.<\/p>\n\n<pre><code>.\n.\n.\nmodel.load('model.tflearn')\nimage_test = cv2.imread(\"test_fish.jpg\", 0)\nblob = cv2.dnn.blobFromImage(image_test, 1.0, (300, 300), [128, 128, 128], False, False)\n# I want to do something like model.setInput(blob)\n.\n.\n.\n<\/code><\/pre>\n\n<p>Also, if anyone have time, I would like to seek for some advice on how to do self-learning in coding? <\/p>\n\n<p>There are a lot of resources online but I just cannot develop my own program or solve difficult questions without tutorials. Am I too relying on existing codes? How can I improve and become a solid programmer?<\/p>\n","answers":[{"AnswerId":"58517031","CreationDate":"2019-10-23T06:53:15.520","ParentId":null,"OwnerUserId":"12192748","Title":null,"Body":"<p>I would suggest you need not write the whole code from scratch but you can start with some use-case and getting piece of codes as required and stitching them accordingly to fit the solution. This way you will get to know how and where to use specific functions\/piece of code. Also you can get ready code and try changing different things as per the result you require and you will end up having your own different version of code.\nPlease refer to code below for your problem:<\/p>\n\n<pre><code>import numpy as np\nimport cv2\n\nrows = open(\"synset_words.txt\").read().strip().split(\"\\n\")\nclasses = [r[r.find(\" \") + 1:].split(\",\")[0] for r in rows]\n\nnet = cv2.dnn.readNetFromCaffe(\"bvlc_googlenet.prototxt\",\"bvlc_googlenet.caffemodel\")\n#Here if you want to import a tensorflow model you can do it as below\n#cv2.dnn.readNetFromTensorflow('frozen_inference_graph.pb', 'graph.pbtxt')\n\nimage = cv2.imread('image.jpg')\nresized = cv2.resize(image, (224, 224))\nblob = cv2.dnn.blobFromImage(resized, 1, (224, 224), (104, 117, 123))\nprint(\"First Blob: {}\".format(blob.shape))\n\nnet.setInput(blob)\npreds = net.forward()\n\nidx = np.argsort(preds[0])[::-1][0]\ntext = \"Label: {}, {:.2f}%\".format(classes[idx],preds[0][idx] * 100)\ncv2.putText(image, text, (5, 25),  cv2.FONT_HERSHEY_SIMPLEX,0.7, (0, 0, 255), 2)\n\n# show the output image\ncv2.imshow(\"Image\", image)\ncv2.waitKey(0)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58153216,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><pip>","CreationDate":"2019-09-29T07:33:42.330","AcceptedAnswerId":null,"OwnerUserId":8589884.0,"Title":"Why can't I use pip3 after installing pip?","Body":"<p>I want to install tensorflow-cpu using pip on my Ubuntu 18.04, 64-bit system and it required pip>=19.0.\nI was having pip 9.0.x and was using pip3 to install packages but to fulfill the requirement I ran<\/p>\n\n<pre><code>pip3 install pip\n<\/code><\/pre>\n\n<p>which installed pip 19.2.3 but after doing it I can no longer use pip3. Using pip3 now gives this error<\/p>\n\n<pre><code>Traceback (most recent call last):\n  File \"\/usr\/bin\/pip3\", line 9, in &lt;module&gt;\n    from pip import main\nImportError: cannot import name 'main'\n<\/code><\/pre>\n\n<p>Earlier using pip would give this error<\/p>\n\n<pre><code>Command 'pip' not found, but can be installed with:\n\nsudo apt install python-pip\n<\/code><\/pre>\n\n<p>but now using it to install packages works totally fine.<\/p>\n\n<p>I read a lot of articles after which I realised that we use pip for python 2.x and for python 3.x we use pip3. But with the change in behaviour after upgrading pip I'm totally confused and can't find any help. Is using pip to install a package same as installing it using pip3. What are the differences between them?<\/p>\n","answers":[]}
+{"QuestionId":58153229,"AnswerCount":0,"Tags":"<tensorflow><keras><lstm><cudnn>","CreationDate":"2019-09-29T07:35:06.323","AcceptedAnswerId":null,"OwnerUserId":9012535.0,"Title":"The difference between the number of params of LSTM and CuDNNLSTM in Kreras","Body":"<p>If an input dimension is 26 and the cell size is 100, the number of parameters in a bi-directional CTC network is 101,600.\nWhen using Keras, I confirmed the number of parameters of a bi-directional LSTM layer was exactly the same as what I expected.\nHowever, CuDNNLSTM contains 102,400 parameters. \nI would appreciate it if someone could provide a clue about the 800 parameters for a quick understanding before reviewing the source code.<\/p>\n","answers":[]}
+{"QuestionId":58153231,"AnswerCount":1,"Tags":"<javascript><tensorflow><neural-network><tensorflow-serving><gcp-ai-platform-training>","CreationDate":"2019-09-29T07:35:14.570","AcceptedAnswerId":null,"OwnerUserId":1726407.0,"Title":"how to predict online tensorflow model using JS","Body":"<p>We build a NN model in tensorflow similar to \n<a href=\"https:\/\/github.com\/GoogleCloudPlatform\/cloudml-samples\/tree\/master\/census\/tf-keras\" rel=\"nofollow noreferrer\">https:\/\/github.com\/GoogleCloudPlatform\/cloudml-samples\/tree\/master\/census\/tf-keras<\/a><\/p>\n\n<p>Then we exported it to GCP using the following <code>gcloud ai-platform models create<\/code> commands\nand predict using <code>gcloud ai-platform predict<\/code> works fine<\/p>\n\n<p>Now we want to do online prediction using javascript, seems to have several options:<\/p>\n\n<ol>\n<li><p>using <a href=\"https:\/\/www.npmjs.com\/package\/@google-cloud\/automl\" rel=\"nofollow noreferrer\">https:\/\/www.npmjs.com\/package\/@google-cloud\/automl<\/a>, but \nChecking <a href=\"https:\/\/googleapis.dev\/nodejs\/automl\/latest\/v1beta1.PredictionServiceClient.html#predict\" rel=\"nofollow noreferrer\">https:\/\/googleapis.dev\/nodejs\/automl\/latest\/v1beta1.PredictionServiceClient.html#predict<\/a> \nit seems that it can not be used for neural networks(?)\nWhen we run the example code \n<a href=\"https:\/\/www.npmjs.com\/package\/@google-cloud\/automl\" rel=\"nofollow noreferrer\">https:\/\/www.npmjs.com\/package\/@google-cloud\/automl<\/a>\nwith our parameters we get 'Invalid resource ID' error<\/p><\/li>\n<li><p>Using REST API, curl works fine but how do we set permanent authorization in the server ?<\/p><\/li>\n<li><p>Any other suggestion ?<\/p><\/li>\n<\/ol>\n\n<p>Amir<\/p>\n","answers":[{"AnswerId":"58406454","CreationDate":"2019-10-16T05:44:23.570","ParentId":null,"OwnerUserId":"1726407","Title":null,"Body":"<p>I solved the problem by building Python container that run on gcloud\n(<a href=\"https:\/\/cloud.google.com\/run\/docs\/quickstarts\/build-and-deploy\" rel=\"nofollow noreferrer\">https:\/\/cloud.google.com\/run\/docs\/quickstarts\/build-and-deploy<\/a>)<\/p>\n\n<p>The JS call the container using REST API, the python container do the prediction and reply to results the JS<\/p>\n"}]}
+{"QuestionId":58153345,"AnswerCount":1,"Tags":"<tensorflow><keras><deep-learning>","CreationDate":"2019-09-29T07:54:56.700","AcceptedAnswerId":58159197.0,"OwnerUserId":8023232.0,"Title":"can i mark the dog with this model?","Body":"<p>I saw this code on the youtube <a href=\"https:\/\/youtu.be\/A4K6D_gx2Iw\" rel=\"nofollow noreferrer\">How to use your trained model - Deep Learning basics with Python, TensorFlow and Keras p.6<\/a><\/p>\n\n<p>If I bring a picture, he tells me if the picture is cat or dog.<\/p>\n\n<p>I can use this code and mark the dog\/cat in the picture?<\/p>\n\n<p>Or do I need to run a brand-new model?<\/p>\n\n<p>in the code I get all pic of dog and cat and put it in X  and y<\/p>\n\n<p>X = it's the features<\/p>\n\n<p>y = it the label<\/p>\n\n<p><strong>the code<\/strong><\/p>\n\n<pre><code>dense_layers = [0]\nlayer_sizes = [64]\nconv_layers = [3]\n\n\n\nmodel = Sequential()\n\nmodel.add(Conv2D(layer_size, (3, 3), input_shape=X.shape[1:]))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Conv2D(layer_size, (3, 3)))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\n\nmodel.add(Flatten())\n\nmodel.add(Dense(layer_size))\n model.add(Activation('relu'))\n\nmodel.add(Dense(1))\nmodel.add(Activation('sigmoid'))\n\nmodel.compile(loss='binary_crossentropy',\n               optimizer='adam',\n               metrics=['accuracy'],\n                )\n\nmodel.fit(X, y,\n          batch_size=32,\n          epochs=10,\n          validation_split=0.3,\n          )\n\nmodel.save('64x3-CNN.model')\n<\/code><\/pre>\n\n<p>after he save it he can use it on some picture and get the write answer<\/p>\n\n<p>I want to add to it the possibility of marking the dog or cat.<\/p>\n","answers":[{"AnswerId":"58159197","CreationDate":"2019-09-29T21:07:07.400","ParentId":null,"OwnerUserId":"12118774","Title":null,"Body":"<p>Questions are not clear but, Let me answer one by one.<\/p>\n\n<p><strong>1)<\/strong> <em>Q: can I use this code and mark the dog\/cat in the picture?<\/em> <\/p>\n\n<blockquote>\n  <p><em>A: Yes you can. However, you should have the saved model to use it. Otherwise you have to train the model and save it again.<\/em><\/p>\n<\/blockquote>\n\n<p><strong>2)<\/strong> <em>Q: I want to add to it the possibility of marking the dog or cat. Is this possible?<\/em><\/p>\n\n<blockquote>\n  <p>*A: Yes, it is possible. <code>model.predict<\/code> would yield you the probabilities.<\/p>\n<\/blockquote>\n"}]}
+{"QuestionId":58153888,"AnswerCount":1,"Tags":"<tensorflow><keras><subclass>","CreationDate":"2019-09-29T09:18:13.780","AcceptedAnswerId":null,"OwnerUserId":10867166.0,"Title":"How to set the input of a keras subclass model in tensorflow?","Body":"<p>I've created a keras subclass model using tensorflow. Snippets are shown below. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>class SubModel(Model):\n    def call(self, inputs):\n        print(inputs)\n\nmodel = SubModel()\nmodel.fit(data, labels, ...)\n<\/code><\/pre>\n\n<p>When <code>fit<\/code> the model, it will get the inputs and input_shape itself. What I want to do is pass inputs to the model myself.Just like the functional API does.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>inputs = tf.keras.input(shape=(100,))\nmodel = tf.keras.Model(inputs=inputs, outputs=outputs)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58154335","CreationDate":"2019-09-29T10:24:49.383","ParentId":null,"OwnerUserId":"6082275","Title":null,"Body":"<p>Something like that?<\/p>\n\n<pre><code>model_ = SubModel()\ninputs = tf.keras.input(shape=(100,))\noutputs = model_(inputs)\nmodel = tf.keras.Model(inputs=inputs, outputs=outputs)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58154104,"AnswerCount":1,"Tags":"<python><machine-learning><keras>","CreationDate":"2019-09-29T09:50:23.873","AcceptedAnswerId":null,"OwnerUserId":12078589.0,"Title":"NameError: name 'X_train' is not defined. this error comes in the remaaining code also. please do rectify this","Body":"<pre><code>from sklearn.model_selection import train_test_split  \n\ndata,Label = shuffle(M, label, random_state = 2)\nlabelled_data = [data, Label]\nX,Y = [labelled_data[0],labelled_data[1]]\n\nX_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size=0.4, random_state=4)\nx_test, x_validation, y_test, y_validation=train_test_split(X_test, Y_test, test_size=0.5,\nrandom_state=4)\n\nX_train = X_train.reshape((X_train.shape[0],256,256,3))\nx_validation = x_validation.reshape((x_validation.shape[0],256,256,3))\n\nx_test =x_test.reshape((x_test.shape[0],256,256,3))\nX_train = X_train.astype('float32')\n\nx_validation = x_validation.astype('float32')\nx_test = x_test.astype('float32')\n\nX_train = X_train\/255\nx_validation = x_validation\/255\nx_test =x_test\/255\n\nfrom keras.utils import np_utils\n\nY_validation = np_utils.to_categorical(Y_train,8)\ny_validation =np_utils.to_categorical(y_validation,8)\ny_test =np_utils.to_categorical(y_test,8)\n\nfrom keras.models import Sequential\nfrom keras.layers import Activation, Dense\n\nmodel = Sequential()\nmodel.add(Dense(units=256,input_shape=(1000,),activation='relu'))\nmodel.add(Dense(units=64,activation='tanh'))\nmodel.add(Dense(units=8,activation='softmax'))\nmodel.compile(loss='categorical_crossentropy', optimizer='sgd', \nmetrics=['categorical_accuracy']) \nmodel.fit(X_train,Y_train,epochs=5,batch_size=32)\nmodel.predict(x_test,batch_size=32)\nreturn model\n\n\nfrom keras.layers.convolutional import Convolution2D\nmodel = Sequential()\nmodel.add(convolution2D(filters=(6,3,3),input_shape=(256,256,1),activation='relu'))\n\nfrom keras.layers.convolutional import MaxPooling2D\nmodel = Sequential()\nmodel.add(MaxPooling2D(pool_size=(2,2),strides=2))\n\nfrom keras.layers import Dropout\nmodel = Sequential()\nmodel.add(Convolution2D(filters=6, nb_row=3, nb_col=3,subsample=(2,2),\n    input_shape=(256, 256, 1,), activation='relu', border_mode='same'))\nmodel.add(Dropout(0.1))\nmodel.add(Flatten())\nmodel.add(Dense(8, activation='softmax'))\n\n\nmodel= fish_model()\nprint(model.summary())\nhistory = model.fit(X_train,Y_train,validation_data=(x_validation, y_validation),epochs=5,batch_size=32)\n\nModel: \"sequential_42\"\n<\/code><\/pre>\n\n<hr>\n\n<h1>      Layer (type)                 Output Shape              Param #<\/h1>\n\n<pre><code> dense_67 (Dense)             (None, 256)               256256    \n<\/code><\/pre>\n\n<hr>\n\n<pre><code> dense_68 (Dense)             (None, 64)                16448     \n<\/code><\/pre>\n\n<hr>\n\n<h1>     dense_69 (Dense)             (None, 8)                 520<\/h1>\n\n<pre><code> Total params: 273,224\n Trainable params: 273,224\n  Non-trainable params: 0\n<\/code><\/pre>\n\n<hr>\n\n<h2>     None<\/h2>\n\n<pre><code>NameError                                 Traceback (most recent call last)\n &lt;ipython-input-131-ab439073340b&gt; in &lt;module&gt;\n       1 model= fish_model()\n       2 print(model.summary())\n ----&gt; 3 history = model.fit(X_train,Y_train,validation_data=(x_validation, y_validation),epochs=5,batch_size=32)\n\n NameError: name 'X_train' is not defined\n<\/code><\/pre>\n\n<p>I have already defined X_train but it shows an error like ir is not defined.\nWhen i tried to test the accuracy also i got the same error like x_test is not defined.<\/p>\n","answers":[{"AnswerId":"58154354","CreationDate":"2019-09-29T10:27:26.997","ParentId":null,"OwnerUserId":"7231848","Title":null,"Body":"<p>the line number shown in the error <\/p>\n\n<hr>\n\n<p>'3' <\/p>\n\n<hr>\n\n<p>does not match the line number in the code you posted.<\/p>\n\n<hr>\n\n<p>Start running your code in chunks to get the 'real' line number of the error. For example, run the first 10 lines of code, then the next 10 lines of code, or run the first module then run the second module. When you get to the error by chunking, the line number should be correct or you may find the error when running the chunks before you get to the end of your code.  <\/p>\n"}]}
+{"QuestionId":58154211,"AnswerCount":2,"Tags":"<python><keras><neural-network><deep-learning><linear-regression>","CreationDate":"2019-09-29T10:07:17.020","AcceptedAnswerId":null,"OwnerUserId":12137797.0,"Title":"Training loss goes up and down during an epoch in a regression model. Also, model performance on a test set is relatively poor (~40% average error)","Body":"<p>I'm trying to build a regression model for apartment prices prediction using Keras. During fitting training loss goes up and down a lot with little to no improvement. What is the problem here? I've tried dropout, regularization, making the network deeper, shallower, wider, narrower, increasing and decreasing learning rate, increasing and decreasing validation set, increasing and decreasing batch size.<\/p>\n\n<pre><code>import pandas as pd\nfrom keras import Sequential\nfrom keras import regularizers\nfrom keras.callbacks import EarlyStopping\nfrom keras.wrappers.scikit_learn import KerasRegressor\nfrom keras.layers import Dense, Dropout, BatchNormalization\nfrom keras.constraints import NonNeg\nfrom keras.optimizers import Adam, RMSprop, SGD\nimport numpy as np\nfrom sklearn.preprocessing import StandardScaler, MinMaxScaler\n\ndf = pd.read_excel('data.xlsx', sheet_name='\u041b\u0438\u0441\u04421')\ndf = df.sample(frac=1, random_state=1).reset_index(drop=True)\n\ntrain_data = df.loc[0:145008, 'id_poligon_level_3':'kitchen_area']\ny_label = df.loc[0:145008, 'price_usd':'price_usd']\n\npred_test = df.loc[145008:, 'id_poligon_level_3':'kitchen_area']\npred_y_label = df.loc[145008:, 'price_usd':'price_usd']\n\n#maxPrice = np.max(df.loc[0:146008, 'price_1sq_meter_usd':'price_1sq_meter_usd'].values)\n\nscale_train_X = StandardScaler()\ntrain_data = scale_train_X.fit_transform(train_data.values)\nscale_train_Y = StandardScaler()\ny_label = scale_train_Y.fit_transform(y_label.values)\n#y_label \/= maxPrice\npred_test = scale_train_X.fit_transform(pred_test.values)\n#pred_y_label \/= maxPrice\npred_y_label = scale_train_Y.transform(pred_y_label.values)\n\n'''\ndata = xl.load_workbook(filename='data.xlsx')\nsheet = data['\u041b\u0438\u0441\u04421']'''\n\n#maxPrice = np.max([val[0].value for val in sheet['H2:H146009']])\n\n'''\ndef get_training_data():\n\n    train_data = []\n    for j in range(2, 146010):\n        for val in sheet['B'+str(j)+':H'+str(j)]:\n            train_set = []\n            train_set.append([cell.value for cell in val])\n            train_data.append(train_set)\n    try:\n        assert train_data is not None\n        print('Getting training data process finished successfully')\n        return train_data\n    except AssertionError:\n        print(\"Error\")\n\n\ndef get_y_label():\n\n    train_set = [val[0].value for val in sheet['L2:L146010']]\n    train_set \/= maxPrice\n\n    try:\n        assert train_set is not None\n        print('Getting y label data process finished successfully')\n        return train_set\n    except AssertionError:\n        print(AssertionError)\n\n\ntrain_data = get_training_data()\n\ntrain_data = np.asarray(train_data)\ntrain_data = np.reshape(train_data, [1, train_data.shape[0], 7])\ntrain_data = train_data[0]\n\nfor i in range(0, train_data.shape[0]):\n    summ = 0\n    for j in str(int(train_data[i][0])):\n        summ += int(j)\n    train_data[i][0] = summ\n\ny_label = train_data.T[6]\ny_label = np.reshape(y_label, [1, y_label.shape[0]])\n\ntest_data = train_data[87000:]\ntrain_data = train_data[:87000]\n\ntest_data = test_data.T[:6]\ntrain_data = train_data.T[:6]\n\ny_label_test = y_label[0][87000:]\ny_label = y_label[0][:87000]\n\n#y_label_test \/= maxPrice\n#y_label \/= maxPrice\ny_label = np.reshape(y_label, [1, y_label.shape[0]])\ny_label_test = np.reshape(y_label_test, [1, y_label_test.shape[0]])\n\n'''\n'''\ndef get_test_data():\n\n    test_data = []\n    for j in range(100001, 146010):\n        for val in sheet['B'+str(j)+':I'+str(j)]:\n            train_set = []\n            train_set.append([cell.value for cell in val])\n            test_data.append(train_set)\n\n    try:\n        assert test_data is not None\n        print('Getting test data process finished successfully')\n        return test_data\n    except AssertionError:\n        print(\"Error\")\n\n\ndef get_test_label():\n\n    test_set = [val[0].value for val in sheet['L100001:L146009']]\n    test_set \/= maxPrice\n\n    try:\n        assert test_set is not None\n        print('Getting y label test data process finished successfully')\n        return test_set\n    except AssertionError:\n        print(AssertionError)\n\n\ntest_data = get_test_data()\ny_label_test = get_test_label()\n\ntest_data = np.asarray(test_data)\ntest_data = np.reshape(test_data, [1, test_data.shape[0], 8])\ntest_data = test_data[0]\n\ny_label_test = np.asarray(y_label_test)\ny_label_test = np.reshape(y_label_test, [1, y_label_test.shape[0]])\n\nassert test_data.shape[0] == y_label_test.shape[1]\n\n\nsc_train_X = StandardScaler().fit(train_data)\nsc_train_Y = StandardScaler().fit(y_label)\nsc_test_X = StandardScaler().fit(test_data)\nsc_test_Y = StandardScaler().fit(y_label_test)\n\ntrain_data = sc_train_X.transform(train_data)\ntest_data = sc_test_X.transform(test_data)\ny_label_test = sc_test_Y.transform(y_label_test)\ny_label = sc_train_Y.transform(y_label)\n\n'''\ndef build_regression():\n\n    model = Sequential([\n        Dense(64, activation='relu', input_shape=(6,), kernel_initializer='random_normal', bias_initializer='Zeros'),\n        #BatchNormalization(),\n        #Dropout(0.2),\n        Dense(64, activation='relu', kernel_initializer='random_normal', bias_initializer='Zeros'),\n        #BatchNormalization(),\n        #Dropout(0.1),\n        Dense(1, activation='linear', kernel_initializer='random_normal', bias_initializer='Zeros'),\n        ])\n\n    opt = Adam(lr=1e-4, decay=1e-3 \/ 200) #Adam(lr=0.001, beta_1=0.9, beta_2=0.98)\n    model.compile(optimizer=opt, loss='mse')\n    return model\n\nestimator = build_regression()#KerasRegressor(build_regression, batch_size=16, epochs=100, validation_split=0.3)\n\nestimator.fit(x=train_data, y=y_label, batch_size=16, epochs=100, validation_split=0.3, shuffle=False, callbacks=[EarlyStopping(monitor='loss', patience=5)])\n\n#estimator.fit(x=train_data, y=y_label, batch_size=96, epochs=250, validation_split=0.2, shuffle=True)\n\npredict = estimator.predict(pred_test)\npredict = scale_train_Y.inverse_transform(predict)\npred_y_label = scale_train_Y.inverse_transform(pred_y_label)\n\ndiff = predict - pred_y_label\npercDiff = (diff \/ pred_y_label) * 100\nabsPercDiff = np.abs(percDiff)\nmean_perc = np.mean(absPercDiff)\nstd_perc = np.std(absPercDiff)\n\nprint('Mean error percentage: {} \\nStandard error percentage: {}'.format(mean_perc, std_perc))\n\n'''\nmodel_json = estimator.to_json()\nwith open(\"model.json\", \"w\") as json_file:\n    json_file.write(model_json)\n# serialize weights to HDF5\nregressor.save_weights(\"model.h5\")\nprint(\"Saved model to disk\")\n'''\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>Train on 101506 samples, validate on 43503 samples\nEpoch 1\/100\n2019-09-29 12:41:56.065057: I tensorflow\/core\/platform\/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2\n\n    16\/101506 [..............................] - ETA: 36:14 - loss: 0.1636\n   896\/101506 [..............................] - ETA: 44s - loss: 0.6881  \n  1520\/101506 [..............................] - ETA: 29s - loss: 0.5158\n  2192\/101506 [..............................] - ETA: 22s - loss: 0.6153\n  3088\/101506 [..............................] - ETA: 17s - loss: 0.8603\n  4016\/101506 [&gt;.............................] - ETA: 14s - loss: 0.7997\n  4944\/101506 [&gt;.............................] - ETA: 12s - loss: 0.9174\n  5840\/101506 [&gt;.............................] - ETA: 11s - loss: 1.0359\n  6688\/101506 [&gt;.............................] - ETA: 10s - loss: 0.9711\n  7472\/101506 [=&gt;............................] - ETA: 10s - loss: 0.9186\n  8336\/101506 [=&gt;............................] - ETA: 9s - loss: 0.8943 \n  9232\/101506 [=&gt;............................] - ETA: 8s - loss: 0.8843\n 10128\/101506 [=&gt;............................] - ETA: 8s - loss: 0.8377\n 11056\/101506 [==&gt;...........................] - ETA: 8s - loss: 0.7994\n 11984\/101506 [==&gt;...........................] - ETA: 7s - loss: 0.7640\n 12912\/101506 [==&gt;...........................] - ETA: 7s - loss: 0.7327\n 13808\/101506 [===&gt;..........................] - ETA: 7s - loss: 0.7647\n 14736\/101506 [===&gt;..........................] - ETA: 7s - loss: 0.7474\n 15664\/101506 [===&gt;..........................] - ETA: 6s - loss: 0.7393\n 16576\/101506 [===&gt;..........................] - ETA: 6s - loss: 0.7204\n 17504\/101506 [====&gt;.........................] - ETA: 6s - loss: 0.6941\n 18448\/101506 [====&gt;.........................] - ETA: 6s - loss: 0.6858\n 19392\/101506 [====&gt;.........................] - ETA: 6s - loss: 0.6624\n 20320\/101506 [=====&gt;........................] - ETA: 5s - loss: 0.6431\n 21248\/101506 [=====&gt;........................] - ETA: 5s - loss: 0.6349\n 22192\/101506 [=====&gt;........................] - ETA: 5s - loss: 0.6385\n 23120\/101506 [=====&gt;........................] - ETA: 5s - loss: 0.6409\n 24064\/101506 [======&gt;.......................] - ETA: 5s - loss: 0.6310\n 25008\/101506 [======&gt;.......................] - ETA: 5s - loss: 0.6638\n 25904\/101506 [======&gt;.......................] - ETA: 5s - loss: 0.6495\n 26752\/101506 [======&gt;.......................] - ETA: 5s - loss: 0.6451\n 27616\/101506 [=======&gt;......................] - ETA: 5s - loss: 0.6402\n 28448\/101506 [=======&gt;......................] - ETA: 5s - loss: 0.6285\n 29360\/101506 [=======&gt;......................] - ETA: 4s - loss: 0.6177\n 30208\/101506 [=======&gt;......................] - ETA: 4s - loss: 0.6194\n 31056\/101506 [========&gt;.....................] - ETA: 4s - loss: 0.6105\n 31936\/101506 [========&gt;.....................] - ETA: 4s - loss: 0.6181\n 32800\/101506 [========&gt;.....................] - ETA: 4s - loss: 0.6073\n 33632\/101506 [========&gt;.....................] - ETA: 4s - loss: 0.5983\n 34496\/101506 [=========&gt;....................] - ETA: 4s - loss: 0.5917\n 35360\/101506 [=========&gt;....................] - ETA: 4s - loss: 0.5834\n 36208\/101506 [=========&gt;....................] - ETA: 4s - loss: 0.5833\n 37104\/101506 [=========&gt;....................] - ETA: 4s - loss: 0.5940\n 37952\/101506 [==========&gt;...................] - ETA: 4s - loss: 0.5937\n 38848\/101506 [==========&gt;...................] - ETA: 4s - loss: 0.5950\n 39744\/101506 [==========&gt;...................] - ETA: 4s - loss: 0.5859\n 40656\/101506 [===========&gt;..................] - ETA: 3s - loss: 0.5809\n 41568\/101506 [===========&gt;..................] - ETA: 3s - loss: 0.5721\n 42464\/101506 [===========&gt;..................] - ETA: 3s - loss: 0.5673\n 43376\/101506 [===========&gt;..................] - ETA: 3s - loss: 0.5745\n 44272\/101506 [============&gt;.................] - ETA: 3s - loss: 0.5830\n 45168\/101506 [============&gt;.................] - ETA: 3s - loss: 0.5803\n 46016\/101506 [============&gt;.................] - ETA: 3s - loss: 0.5893\n 46992\/101506 [============&gt;.................] - ETA: 3s - loss: 0.5844\n 47920\/101506 [=============&gt;................] - ETA: 3s - loss: 0.5794\n 48816\/101506 [=============&gt;................] - ETA: 3s - loss: 0.5725\n 49712\/101506 [=============&gt;................] - ETA: 3s - loss: 0.5688\n 50656\/101506 [=============&gt;................] - ETA: 3s - loss: 0.5686\n 51568\/101506 [==============&gt;...............] - ETA: 3s - loss: 0.5622\n 52496\/101506 [==============&gt;...............] - ETA: 3s - loss: 0.5594\n 53392\/101506 [==============&gt;...............] - ETA: 3s - loss: 0.5529\n 54320\/101506 [===============&gt;..............] - ETA: 2s - loss: 0.5494\n 55248\/101506 [===============&gt;..............] - ETA: 2s - loss: 0.5438\n 56176\/101506 [===============&gt;..............] - ETA: 2s - loss: 0.5374\n 57104\/101506 [===============&gt;..............] - ETA: 2s - loss: 0.5347\n 58032\/101506 [================&gt;.............] - ETA: 2s - loss: 0.5325\n 58960\/101506 [================&gt;.............] - ETA: 2s - loss: 0.5272\n 59904\/101506 [================&gt;.............] - ETA: 2s - loss: 0.5218\n 60848\/101506 [================&gt;.............] - ETA: 2s - loss: 0.5171\n 61776\/101506 [=================&gt;............] - ETA: 2s - loss: 0.5180\n 62704\/101506 [=================&gt;............] - ETA: 2s - loss: 0.5206\n 63648\/101506 [=================&gt;............] - ETA: 2s - loss: 0.5192\n 64528\/101506 [==================&gt;...........] - ETA: 2s - loss: 0.5205\n 65392\/101506 [==================&gt;...........] - ETA: 2s - loss: 0.5179\n 66304\/101506 [==================&gt;...........] - ETA: 2s - loss: 0.5147\n 67216\/101506 [==================&gt;...........] - ETA: 2s - loss: 0.5106\n 68096\/101506 [===================&gt;..........] - ETA: 2s - loss: 0.5122\n 68960\/101506 [===================&gt;..........] - ETA: 1s - loss: 0.5087\n 69824\/101506 [===================&gt;..........] - ETA: 1s - loss: 0.5048\n 70704\/101506 [===================&gt;..........] - ETA: 1s - loss: 0.5066\n 71584\/101506 [====================&gt;.........] - ETA: 1s - loss: 0.5114\n 72480\/101506 [====================&gt;.........] - ETA: 1s - loss: 0.5068\n 73344\/101506 [====================&gt;.........] - ETA: 1s - loss: 0.5029\n 74224\/101506 [====================&gt;.........] - ETA: 1s - loss: 0.5022\n 75104\/101506 [=====================&gt;........] - ETA: 1s - loss: 0.5036\n 75984\/101506 [=====================&gt;........] - ETA: 1s - loss: 0.5017\n 76816\/101506 [=====================&gt;........] - ETA: 1s - loss: 0.4990\n 77600\/101506 [=====================&gt;........] - ETA: 1s - loss: 0.4959\n 78416\/101506 [======================&gt;.......] - ETA: 1s - loss: 0.5045\n 79248\/101506 [======================&gt;.......] - ETA: 1s - loss: 0.5044\n 80080\/101506 [======================&gt;.......] - ETA: 1s - loss: 0.5015\n 80912\/101506 [======================&gt;.......] - ETA: 1s - loss: 0.5000\n 81728\/101506 [=======================&gt;......] - ETA: 1s - loss: 0.4969\n 82576\/101506 [=======================&gt;......] - ETA: 1s - loss: 0.4954\n 83424\/101506 [=======================&gt;......] - ETA: 1s - loss: 0.4917\n 84288\/101506 [=======================&gt;......] - ETA: 1s - loss: 0.4908\n 85200\/101506 [========================&gt;.....] - ETA: 0s - loss: 0.4895\n 86128\/101506 [========================&gt;.....] - ETA: 0s - loss: 0.4898\n 87040\/101506 [========================&gt;.....] - ETA: 0s - loss: 0.4914\n 87936\/101506 [========================&gt;.....] - ETA: 0s - loss: 0.5128\n 88848\/101506 [=========================&gt;....] - ETA: 0s - loss: 0.5126\n 89744\/101506 [=========================&gt;....] - ETA: 0s - loss: 0.5132\n 90624\/101506 [=========================&gt;....] - ETA: 0s - loss: 0.5102\n 91536\/101506 [==========================&gt;...] - ETA: 0s - loss: 0.5073\n 92448\/101506 [==========================&gt;...] - ETA: 0s - loss: 0.5036\n 93344\/101506 [==========================&gt;...] - ETA: 0s - loss: 0.5030\n 94240\/101506 [==========================&gt;...] - ETA: 0s - loss: 0.5010\n 95104\/101506 [===========================&gt;..] - ETA: 0s - loss: 0.4980\n 95984\/101506 [===========================&gt;..] - ETA: 0s - loss: 0.4956\n 96896\/101506 [===========================&gt;..] - ETA: 0s - loss: 0.4931\n 97776\/101506 [===========================&gt;..] - ETA: 0s - loss: 0.4928\n 98656\/101506 [============================&gt;.] - ETA: 0s - loss: 0.4901\n 99552\/101506 [============================&gt;.] - ETA: 0s - loss: 0.4872\n100448\/101506 [============================&gt;.] - ETA: 0s - loss: 0.4875\n101328\/101506 [============================&gt;.] - ETA: 0s - loss: 0.4850\n101506\/101506 [==============================] - 7s 69us\/step - loss: 0.4844 - val_loss: 0.3809\nEpoch 2\/100\n\n    16\/101506 [..............................] - ETA: 18s - loss: 0.0778\n   928\/101506 [..............................] - ETA: 5s - loss: 0.1726 \n  1840\/101506 [..............................] - ETA: 5s - loss: 0.1622\n  2688\/101506 [..............................] - ETA: 5s - loss: 0.2834\n  3568\/101506 [&gt;.............................] - ETA: 5s - loss: 0.2775\n  4448\/101506 [&gt;.............................] - ETA: 5s - loss: 0.2740\n  5360\/101506 [&gt;.............................] - ETA: 5s - loss: 0.3466\n  6288\/101506 [&gt;.............................] - ETA: 5s - loss: 0.3877\n  7232\/101506 [=&gt;............................] - ETA: 5s - loss: 0.3647\n  8144\/101506 [=&gt;............................] - ETA: 5s - loss: 0.3463\n  9072\/101506 [=&gt;............................] - ETA: 5s - loss: 0.3689\n  9984\/101506 [=&gt;............................] - ETA: 5s - loss: 0.3552\n 10928\/101506 [==&gt;...........................] - ETA: 5s - loss: 0.3515\n 11856\/101506 [==&gt;...........................] - ETA: 4s - loss: 0.3333\n 12784\/101506 [==&gt;...........................] - ETA: 4s - loss: 0.3246\n 13696\/101506 [===&gt;..........................] - ETA: 4s - loss: 0.3461\n 14640\/101506 [===&gt;..........................] - ETA: 4s - loss: 0.3441\n 15568\/101506 [===&gt;..........................] - ETA: 4s - loss: 0.3416\n 16464\/101506 [===&gt;..........................] - ETA: 4s - loss: 0.3337\n 17392\/101506 [====&gt;.........................] - ETA: 4s - loss: 0.3213\n 18288\/101506 [====&gt;.........................] - ETA: 4s - loss: 0.3251\n 19200\/101506 [====&gt;.........................] - ETA: 4s - loss: 0.3165\n 20096\/101506 [====&gt;.........................] - ETA: 4s - loss: 0.3106\n 21008\/101506 [=====&gt;........................] - ETA: 4s - loss: 0.3036\n 21920\/101506 [=====&gt;........................] - ETA: 4s - loss: 0.3175\n 22816\/101506 [=====&gt;........................] - ETA: 4s - loss: 0.3159\n 23712\/101506 [======&gt;.......................] - ETA: 4s - loss: 0.3225\n 24608\/101506 [======&gt;.......................] - ETA: 4s - loss: 0.3269\n 25520\/101506 [======&gt;.......................] - ETA: 4s - loss: 0.3477\n 26400\/101506 [======&gt;.......................] - ETA: 4s - loss: 0.3490\n 27328\/101506 [=======&gt;......................] - ETA: 4s - loss: 0.3496\n 28272\/101506 [=======&gt;......................] - ETA: 4s - loss: 0.3469\n 29200\/101506 [=======&gt;......................] - ETA: 4s - loss: 0.3442\n 30112\/101506 [=======&gt;......................] - ETA: 3s - loss: 0.3502\n 31008\/101506 [========&gt;.....................] - ETA: 3s - loss: 0.3480\n 31904\/101506 [========&gt;.....................] - ETA: 3s - loss: 0.3549\n 32832\/101506 [========&gt;.....................] - ETA: 3s - loss: 0.3537\n 33792\/101506 [========&gt;.....................] - ETA: 3s - loss: 0.3507\n 34704\/101506 [=========&gt;....................] - ETA: 3s - loss: 0.3510\n 35616\/101506 [=========&gt;....................] - ETA: 3s - loss: 0.3505\n 36528\/101506 [=========&gt;....................] - ETA: 3s - loss: 0.3498\n 37488\/101506 [==========&gt;...................] - ETA: 3s - loss: 0.3664\n 38416\/101506 [==========&gt;...................] - ETA: 3s - loss: 0.3635\n 39344\/101506 [==========&gt;...................] - ETA: 3s - loss: 0.3681\n 40272\/101506 [==========&gt;...................] - ETA: 3s - loss: 0.3648\n 41232\/101506 [===========&gt;..................] - ETA: 3s - loss: 0.3614\n 42192\/101506 [===========&gt;..................] - ETA: 3s - loss: 0.3578\n 43152\/101506 [===========&gt;..................] - ETA: 3s - loss: 0.3639\n 44080\/101506 [============&gt;.................] - ETA: 3s - loss: 0.3728\n 45008\/101506 [============&gt;.................] - ETA: 3s - loss: 0.3785\n 45968\/101506 [============&gt;.................] - ETA: 3s - loss: 0.3899\n 46864\/101506 [============&gt;.................] - ETA: 3s - loss: 0.3891\n 47776\/101506 [=============&gt;................] - ETA: 2s - loss: 0.3859\n 48672\/101506 [=============&gt;................] - ETA: 2s - loss: 0.3831\n 49568\/101506 [=============&gt;................] - ETA: 2s - loss: 0.3818\n 50448\/101506 [=============&gt;................] - ETA: 2s - loss: 0.3825\n 51376\/101506 [==============&gt;...............] - ETA: 2s - loss: 0.3814\n 52288\/101506 [==============&gt;...............] - ETA: 2s - loss: 0.3807\n 53200\/101506 [==============&gt;...............] - ETA: 2s - loss: 0.3764\n 54144\/101506 [===============&gt;..............] - ETA: 2s - loss: 0.3758\n 55072\/101506 [===============&gt;..............] - ETA: 2s - loss: 0.3733\n 55968\/101506 [===============&gt;..............] - ETA: 2s - loss: 0.3696\n 56864\/101506 [===============&gt;..............] - ETA: 2s - loss: 0.3698\n 57824\/101506 [================&gt;.............] - ETA: 2s - loss: 0.3687\n 58768\/101506 [================&gt;.............] - ETA: 2s - loss: 0.3659\n 59696\/101506 [================&gt;.............] - ETA: 2s - loss: 0.3630\n 60624\/101506 [================&gt;.............] - ETA: 2s - loss: 0.3606\n 61552\/101506 [=================&gt;............] - ETA: 2s - loss: 0.3633\n 62464\/101506 [=================&gt;............] - ETA: 2s - loss: 0.3683\n 63376\/101506 [=================&gt;............] - ETA: 2s - loss: 0.3688\n 64320\/101506 [==================&gt;...........] - ETA: 2s - loss: 0.3670\n 65248\/101506 [==================&gt;...........] - ETA: 1s - loss: 0.3695\n 66176\/101506 [==================&gt;...........] - ETA: 1s - loss: 0.3686\n 67088\/101506 [==================&gt;...........] - ETA: 1s - loss: 0.3656\n 67936\/101506 [===================&gt;..........] - ETA: 1s - loss: 0.3676\n 68816\/101506 [===================&gt;..........] - ETA: 1s - loss: 0.3651\n 69712\/101506 [===================&gt;..........] - ETA: 1s - loss: 0.3639\n 70624\/101506 [===================&gt;..........] - ETA: 1s - loss: 0.3671\n 71536\/101506 [====================&gt;.........] - ETA: 1s - loss: 0.3728\n 72448\/101506 [====================&gt;.........] - ETA: 1s - loss: 0.3699\n 73344\/101506 [====================&gt;.........] - ETA: 1s - loss: 0.3676\n 74256\/101506 [====================&gt;.........] - ETA: 1s - loss: 0.3686\n 75184\/101506 [=====================&gt;........] - ETA: 1s - loss: 0.3707\n 76112\/101506 [=====================&gt;........] - ETA: 1s - loss: 0.3698\n 77040\/101506 [=====================&gt;........] - ETA: 1s - loss: 0.3684\n 77968\/101506 [======================&gt;.......] - ETA: 1s - loss: 0.3680\n 78960\/101506 [======================&gt;.......] - ETA: 1s - loss: 0.3762\n 79872\/101506 [======================&gt;.......] - ETA: 1s - loss: 0.3766\n 80816\/101506 [======================&gt;.......] - ETA: 1s - loss: 0.3755\n 81728\/101506 [=======================&gt;......] - ETA: 1s - loss: 0.3736\n 82672\/101506 [=======================&gt;......] - ETA: 1s - loss: 0.3733\n 83552\/101506 [=======================&gt;......] - ETA: 0s - loss: 0.3708\n 84528\/101506 [=======================&gt;......] - ETA: 0s - loss: 0.3706\n 85456\/101506 [========================&gt;.....] - ETA: 0s - loss: 0.3699\n 86336\/101506 [========================&gt;.....] - ETA: 0s - loss: 0.3712\n 87264\/101506 [========================&gt;.....] - ETA: 0s - loss: 0.3740\n 88192\/101506 [=========================&gt;....] - ETA: 0s - loss: 0.3960\n 89072\/101506 [=========================&gt;....] - ETA: 0s - loss: 0.3997\n 89984\/101506 [=========================&gt;....] - ETA: 0s - loss: 0.3976\n 90928\/101506 [=========================&gt;....] - ETA: 0s - loss: 0.3951\n 91840\/101506 [==========================&gt;...] - ETA: 0s - loss: 0.3932\n 92768\/101506 [==========================&gt;...] - ETA: 0s - loss: 0.3909\n 93664\/101506 [==========================&gt;...] - ETA: 0s - loss: 0.3916\n 94576\/101506 [==========================&gt;...] - ETA: 0s - loss: 0.3896\n 95488\/101506 [===========================&gt;..] - ETA: 0s - loss: 0.3882\n 96416\/101506 [===========================&gt;..] - ETA: 0s - loss: 0.3860\n 97328\/101506 [===========================&gt;..] - ETA: 0s - loss: 0.3867\n 98240\/101506 [============================&gt;.] - ETA: 0s - loss: 0.3851\n 99152\/101506 [============================&gt;.] - ETA: 0s - loss: 0.3825\n100048\/101506 [============================&gt;.] - ETA: 0s - loss: 0.3818\n100960\/101506 [============================&gt;.] - ETA: 0s - loss: 0.3817\n101506\/101506 [==============================] - 6s 64us\/step - loss: 0.3808 - val_loss: 0.3671\n<\/code><\/pre>\n\n<p>After 100 epochs the loss stays around 0.36<\/p>\n","answers":[{"AnswerId":"58159034","CreationDate":"2019-09-29T20:41:39.510","ParentId":null,"OwnerUserId":"12118774","Title":null,"Body":"<p>Seems like you just hit the local minima that is not intended. The things that you can also try is:<\/p>\n\n<ul>\n<li>Change the optimizer.<\/li>\n<li>Increase the learning rate in a way that you just overshoot the local minimums with a <strong>well-designed decay<\/strong> (You might need high decay) to keep track on better loss value.<\/li>\n<li>Change the activation functions.<\/li>\n<\/ul>\n\n<p>OR<\/p>\n\n<p>A possible thing that you have to suspect is the <strong>data itself<\/strong>. It might be so noisy to be fitted in. You have to check whether you have a noisy data or not by simply check with <code>matplotlib<\/code> in this case. However do not forget that, as I've understand from your code, it may be computationally expensive since you have like 200k points.<\/p>\n\n<p>If it is so hard to check with the plotting, you can eliminate the noisy data with some filters that may differ with the situation itself.<\/p>\n"},{"AnswerId":"58159275","CreationDate":"2019-09-29T21:18:04.377","ParentId":null,"OwnerUserId":"10133797","Title":null,"Body":"<p>Your results appear fine, if not great for the data you have; <code>val_loss<\/code> is very close to <code>loss<\/code>, which is basically ideal. In absolute terms, <code>.36<\/code> validation loss is pretty good - <code>&lt;.25<\/code> would be excellent - but the main indicator is that it's close to your train loss, so <em>data<\/em> is the most likely limiting factor.<\/p>\n\n<p>The confusion may be from the title: \"`~40% average error\": <strong>loss != error<\/strong>. For regression, \"accuracy\" is not defined (you can only measure the 'closeness' of predictions vs. labels; if you measured accuracy as with classification, it'd always be close to zero as you'd have to match labels, which are float32 real-numbers, <em>exactly<\/em>). If you'd like to get an even better idea of how your model is doing, <em>plot<\/em> predictions vs. labels.<\/p>\n"}]}
+{"QuestionId":58154472,"AnswerCount":1,"Tags":"<python><list><tensorflow><deep-learning><tflearn>","CreationDate":"2019-09-29T10:47:05.097","AcceptedAnswerId":null,"OwnerUserId":12138185.0,"Title":"Python Deep Learning: Shape of irregular multidimensional data sets","Body":"<p>I am using TFLearn and python 3.7 to create a my first neural network, which goal is to play snake. \nThe game consists of a 10x10 board, with empty spaces shown as a 0, the goal the snake is trying to eat shown as -1, the head of the snake shown as a 1, and each consecutive part of the snake shown as the previous part of the snake + 1. For example, a game state could be:<\/p>\n\n<pre><code>0 0 0 0 0 0 0 0 0 0\n\n0 0 0 0 0 0 0 0 0 0\n\n-1 0 0 0 0 0 0 0 5 0\n\n0 0 0 0 0 0 2 3 4 0\n\n0 0 0 0 0 0 1 0 0 0\n\n0 0 0 0 0 0 0 0 0 0\n\n0 0 0 0 0 0 0 0 0 0\n\n0 0 0 0 0 0 0 0 0 0\n\n0 0 0 0 0 0 0 0 0 0\n\n0 0 0 0 0 0 0 0 0 0\n<\/code><\/pre>\n\n<p>The state is stored as a single array, of size 100. <\/p>\n\n<p>I have created the game and the training data, which is a list of lists containing a state and the action taken (0 - 3 inclusive, representing the 4 directions). An element of the input data could be:<\/p>\n\n<pre><code>[[0, 0, ... 0], 2]\n<\/code><\/pre>\n\n<p>In the above element of input data, the state was [0, 0, ... 0] and the action taken was 2 (move down). <\/p>\n\n<p>I am having difficulty determining the shape of this data, as the action is an int not a list of size 100.  I assume the shape of the data would be:<\/p>\n\n<pre><code>[None, numberOfInputElements, 2, 100],\n<\/code><\/pre>\n\n<p>though this seems odd as the action taken is an integer and not a list of size 100 as is defined in the data shape. <\/p>\n\n<p>Would this be correct, and if not, what would the correct shape be (if possible, please provide an explanation as to why).<\/p>\n\n<p>Thanks for any help you can give!<\/p>\n","answers":[{"AnswerId":"58155774","CreationDate":"2019-09-29T13:38:59.797","ParentId":null,"OwnerUserId":"1500585","Title":null,"Body":"<p>It would be better to have one tensor representing the input image with the shape <code>[None, 100]<\/code> (or better, <code>[None, 10, 10]<\/code> so you can try a convolutional approach) and another tensor representing the action with the shape <code>[None]<\/code>, or perhaps <code>[None,4]<\/code>, where the action is represented in one-hot fashion (this is useful for defining cost functions such as the cross-entropy).<\/p>\n\n<p>An unrelated idea - assuming the game arena is cyclic, always center the matrix around the snake's head (so the snake's head is always at the same location).<\/p>\n"}]}
+{"QuestionId":58154611,"AnswerCount":2,"Tags":"<python><keras><deep-learning><conv-neural-network><autoencoder>","CreationDate":"2019-09-29T11:08:18.657","AcceptedAnswerId":null,"OwnerUserId":8066148.0,"Title":"Keras CONV1D: Error when checking target: expected dense_output to have 2 dimensions, but got array with shape (35206, 50, 1)","Body":"<p>I have this trouble:\nError when checking target: expected dense_output to have 2 dimensions, but got array with shape (35206, 50, 1)\nWith this code, an autoencoder with CONV1D and two output, but the trouble is the reconstruction output (dense_output):<\/p>\n\n<pre><code>X_train, X_test, y_train, y_test = train_test_split(X, other_output, test_size=0.3, random_state=42)\n\nTAM_VECTOR = X_train.shape[1]\n\ninput_tweet = Input(shape=(TAM_VECTOR,X_train.shape[2]))\n\nencoded = Conv1D(64, kernel_size=1, activation='relu')(input_tweet)\nencoded = Conv1D(32, kernel_size=1, activation='relu')(encoded)\n\ndecoded = Conv1D(32, kernel_size=1, activation='relu')(encoded)\ndecoded = Conv1D(64, kernel_size=1, activation='relu')(decoded)\ndecoded = Flatten()(decoded)\ndecoded = Dense(TAM_VECTOR, activation='relu', name='dense_output')(decoded)\n\nencoded = Flatten()(encoded)\nsecond_output = Dense(1, activation='linear', name='second_output')(encoded)\n\nautoencoder = Model(inputs=input_tweet, outputs=[decoded, second_output])\n\nautoencoder.compile(optimizer=\"adam\",\n                    loss={'dense_output': 'mse', 'second_output': 'mse'},\n                    loss_weights={'dense_output': 0.001, 'second_output': 0.999},\n                    metrics=[\"mae\"])\n\nautoencoder.fit([X_train], [X_train, y_train], epochs=10, batch_size=32)\n<\/code><\/pre>\n\n<p>The input (X) has shape (50000,50) and I reshape to:<\/p>\n\n<pre><code>X = np.reshape(X, (X.shape[0], X.shape[1], -1))\n<\/code><\/pre>\n\n<p>(50000,50,1)<\/p>\n\n<p>And other_output is<\/p>\n\n<pre><code>other_output.shape\n<\/code><\/pre>\n\n<p>(50000,1)<\/p>\n\n<p>And here the model summary<\/p>\n\n<pre><code>Layer (type)                    Output Shape         Param #     Connected to                     \n==================================================================================================\ninput_27 (InputLayer)           (None, 50, 1)        0                                            \n__________________________________________________________________________________________________\nconv1d_105 (Conv1D)             (None, 50, 64)       128         input_27[0][0]                   \n__________________________________________________________________________________________________\nconv1d_106 (Conv1D)             (None, 50, 32)       2080        conv1d_105[0][0]                 \n__________________________________________________________________________________________________\nconv1d_107 (Conv1D)             (None, 50, 32)       1056        conv1d_106[0][0]                 \n__________________________________________________________________________________________________\nconv1d_108 (Conv1D)             (None, 50, 64)       2112        conv1d_107[0][0]                 \n__________________________________________________________________________________________________\nflatten_42 (Flatten)            (None, 3200)         0           conv1d_108[0][0]                 \n__________________________________________________________________________________________________\nflatten_43 (Flatten)            (None, 1600)         0           conv1d_106[0][0]                 \n__________________________________________________________________________________________________\ndense_output (Dense)            (None, 50)           160050      flatten_42[0][0]                 \n__________________________________________________________________________________________________\nsecond_output (Dense)           (None, 1)            1601        flatten_43[0][0]                 \n==================================================================================================\nTotal params: 167,027\nTrainable params: 167,027\nNon-trainable params: 0\n<\/code><\/pre>\n","answers":[{"AnswerId":"58155388","CreationDate":"2019-09-29T12:54:06.330","ParentId":null,"OwnerUserId":"8066148","Title":null,"Body":"<p>Now, changing the last Dense by a Conv the code looks like:<\/p>\n\n<pre><code>input_tweet = Input(shape=(TAM_VECTOR,X_train.shape[2]))\n\nencoded = Conv1D(64, kernel_size=1, activation='relu')(input_tweet)\nencoded = Conv1D(32, kernel_size=1, activation='relu')(encoded)\n\ndecoded = Conv1D(32, kernel_size=1, activation='relu')(encoded)\ndecoded = Conv1D(64, kernel_size=1, activation='relu')(decoded)\ndecoded = Conv1D(TAM_VECTOR, kernel_size=1, activation='relu', name='decoded_output')(decoded)\n\nencoded = Flatten()(encoded)\nsecond_output = Dense(1, activation='linear', name='second_output')(encoded)\n<\/code><\/pre>\n\n<p>But with this error:<\/p>\n\n<pre><code>ValueError: Error when checking target: expected tweet_output to have shape (50, 50) but got array with shape (50, 1)\n<\/code><\/pre>\n\n<p>I also had this error in other testings when I put a Dense layer without a Flatten one previously<\/p>\n"},{"AnswerId":"58155147","CreationDate":"2019-09-29T12:19:41.633","ParentId":null,"OwnerUserId":"8366805","Title":null,"Body":"<p>You are training a convolutional autoencoder, so the output layer for decoding cannot be a dense layer (it does not have a channel). Instead, you need a <code>Conv1D<\/code> layer at the end.<\/p>\n\n<p>Here is the code that works:<\/p>\n\n<pre><code>tf.reset_default_graph()\n\nX = np.random.rand(50000, 50, 1); y = np.linspace(1, 50000, 50000)\nX_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=42)\n\nTAM_VECTOR = X_train.shape[1]\n\ninput_tweet = Input(shape=(TAM_VECTOR, X_train.shape[2]))\nencoded = Conv1D(64, kernel_size=1, activation='relu')(input_tweet)\nencoded = Conv1D(32, kernel_size=1, activation='relu')(encoded)\n\ndecoded = Conv1D(32, kernel_size=1, activation='relu')(encoded)\ndecoded = Conv1D(64, kernel_size=1, activation='relu')(decoded)\ndecoded = Conv1D(1, kernel_size=1, activation='sigmoid', name='dense_output')(decoded)\n\nencoded = Flatten()(encoded)\nsecond_output = Dense(1, activation='linear', name='second_output')(encoded)\n\nautoencoder = Model(inputs=input_tweet, outputs=[decoded, second_output])\nautoencoder.compile(optimizer='adam', loss='mse')\n\nplot_model(autoencoder, show_shapes=True)\nautoencoder.fit(X_train, {'dense_output': X_train, 'second_output': y_train}, epochs=10, batch_size=32)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58154802,"AnswerCount":0,"Tags":"<python-3.x><keras><categorical-data>","CreationDate":"2019-09-29T11:37:37.753","AcceptedAnswerId":null,"OwnerUserId":6278135.0,"Title":"Keras to_categoral and predictions","Body":"<p>about using the to_categorical function of the keras.utils package, the documentation provides the following example:<\/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>In this example, the categories are the 0, 1, and 2 values in labels.<\/p>\n\n<p>From the result given in the example, I believe that to_categorical returns the values of categories in ascending order. <strong>Is this understanding correct?<\/strong> Another question is, when making predictions, in something like<\/p>\n\n<pre><code>predictions = model.predict(X_teste)\ndf_keras = pd.DataFrame(predictions, columns=[0, 1, 2, 3, 4, 5, 6, 7, 8])\ndf_keras['result'] = df_keras.filter(regex='\\d').agg('idxmax', axis=1)\n<\/code><\/pre>\n\n<p>Is it guaranteed that the resulting vector will have the values of the <strong>categories in ascending order<\/strong>?<\/p>\n","answers":[]}
+{"QuestionId":58154906,"AnswerCount":0,"Tags":"<tensorflow><keras><vgg-net><faster-rcnn>","CreationDate":"2019-09-29T11:50:33.237","AcceptedAnswerId":null,"OwnerUserId":11118799.0,"Title":"Error in bottleneck_model.predict_generator","Body":"<p>I am trying to run the lip reading code from <a href=\"https:\/\/github.com\/zrobert7\/cs231N_lip_reading\/blob\/master\/main.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/zrobert7\/cs231N_lip_reading\/blob\/master\/main.py<\/a><\/p>\n\n<p>Here the author has made a bottleneck_feature.npy file using <\/p>\n\n<pre><code>bottleneck_features_train = bottleneck_model.predict_generator(self.training_generator(),steps=np.shape(self.X_train)[0] \/ self.config.batch_size)\n<\/code><\/pre>\n\n<p>where training_generator() is<\/p>\n\n<pre><code>def training_generator(self):\nwhile True:\nfor i in range(int(np.shape(self.X_train)[0]\/ self.config.batch_size)):\n  x = self.X_train[i * self.config.batch_size: (i + 1) * self.config.batch_size]\n  y = self.y_train[i * self.config.batch_size: (i + 1) * self.config.batch_size]\n  one_hot_labels_train = keras.utils.to_categorical(y, num_classes=self.config.num_classes)\n  yield (x, one_hot_labels_train)\n\n<\/code><\/pre>\n\n<p>It stops abruptly without an Error. Please Help<\/p>\n","answers":[]}
+{"QuestionId":58155073,"AnswerCount":1,"Tags":"<python><http><flask><keras>","CreationDate":"2019-09-29T12:10:34.547","AcceptedAnswerId":58155204.0,"OwnerUserId":11613489.0,"Title":"Python Flask File Upload fails with 404 error","Body":"<p>While following the Flask documentation \/ tutorials on how to upload a file I used this code:<\/p>\n\n<p><strong>main.py:<\/strong>    <\/p>\n\n<pre><code>from flask import render_template, jsonify, Flask, redirect, url_for, request\nfrom app import app\nimport random\nimport os\nfrom keras.applications.resnet50 import ResNet50\nfrom keras.preprocessing import image\nfrom keras.applications.resnet50 import preprocess_input, decode_predictions\nimport numpy as np\n\n\n@app.route('\/')\n\n\n\n@app.route('\/upload')\ndef upload_file2():\n   return render_template('index.html')\n\n@app.route('\/uploaded', methods = ['GET', 'POST'])\ndef upload_file():\n   if request.method == 'POST':\n      f = request.files['file']\n      path = os.path.join(app.config['UPLOAD_FOLDER'], f.filename)\n      model= ResNet50(weights='imagenet')\n      img = image.load_img(path, target_size=(224,224))\n      x = image.img_to_array(img)\n      x = np.expand_dims(x, axis=0)\n      x = preprocess_input(x)\n      preds = model.predict(x)\n      preds_decoded = decode_predictions(preds, top=3)[0] \n      print(decode_predictions(preds, top=3)[0])\n      f.save(path)\n      return render_template('uploaded.html', title='Success', predictions=preds_decoded, user_image=f.filename)\n\n\n@app.route('\/index')\ndef index():\n    return render_template('index.html', title='Home')\n\n@app.route('\/map')\ndef map():\n    return render_template('map.html', title='Map')\n\n\n@app.route('\/map\/refresh', methods=['POST'])\ndef map_refresh():\n    points = [(random.uniform(48.8434100, 48.8634100),\n               random.uniform(2.3388000, 2.3588000))\n              for _ in range(random.randint(2, 9))]\n    return jsonify({'points': points})\n\n\n@app.route('\/contact')\ndef contact():\n    return render_template('contact.html', title='Contact')\n<\/code><\/pre>\n\n<p>Everything seems to be fine in the code. I have checked several times.<\/p>\n\n<p>Well when running it on localhost on my Windows machine. <\/p>\n\n<p>I get an 404 http-status-code (page not found).\n<a href=\"https:\/\/i.stack.imgur.com\/39nRG.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/39nRG.png\" alt=\"e 404 POST\"><\/a><\/p>\n\n<p>I tried: on <strong><a href=\"http:\/\/localhost:5000\/\" rel=\"nofollow noreferrer\">http:\/\/localhost:5000\/<\/a><\/strong> and <strong><a href=\"http:\/\/localhost:5500\/\" rel=\"nofollow noreferrer\">http:\/\/localhost:5500\/<\/a><\/strong><\/p>\n\n<p>Also I cleaned my browser...<\/p>\n","answers":[{"AnswerId":"58155204","CreationDate":"2019-09-29T12:26:53.180","ParentId":null,"OwnerUserId":"11059704","Title":null,"Body":"<p>Change\n@app.route('\/uploaded', methods = ['GET', 'POST'])\nTo\n@app.route('\/upload.php', methods = ['GET', 'POST'])<\/p>\n\n<p>Or, change your front end form to post to uploaded.<\/p>\n"}]}
+{"QuestionId":58155252,"AnswerCount":2,"Tags":"<python><pytorch>","CreationDate":"2019-09-29T12:35:03.953","AcceptedAnswerId":null,"OwnerUserId":11287895.0,"Title":"How to increase numerical accuracy of Pytorch model?","Body":"<p>With the following line:     <\/p>\n\n<pre><code>print(self.netG(self.real_A)-self.netG(self.real_A))\n<\/code><\/pre>\n\n<p>I think I can get a all zero tensor, but no.<\/p>\n\n<pre><code>tensor([[ [[-0.0032,  0.0089, -0.0085,  ..., -0.0027,  0.0004, -0.0022],\n          [-0.0019, -0.0022,  0.0775,  ...,  0.0236, -0.0277, -0.0125],\n          [ 0.0049,  0.0159,  0.0203,  ..., -0.0212,  0.0010, -0.0069],\n          ...,\n          [ 0.0042,  0.0081, -0.0127,  ..., -0.0097,  0.0136, -0.0002],\n          [-0.0010,  0.0020, -0.0066,  ...,  0.0260,  0.0433,  0.0088],\n          [-0.0023,  0.0095,  0.0125,  ...,  0.0005,  0.0090,  0.0029]]]],\n       device='cuda:0', grad_fn=&lt;SubBackward0&gt;)\n<\/code><\/pre>\n","answers":[{"AnswerId":"58174628","CreationDate":"2019-09-30T20:06:10.957","ParentId":null,"OwnerUserId":"4280242","Title":null,"Body":"<p>When working with floating point values, <a href=\"https:\/\/en.wikipedia.org\/wiki\/Associative_property\" rel=\"nofollow noreferrer\">associativity<\/a> of some real-valued operations is not preserved. This means for instance, that there is no guarantee that <code>(a + b) + c == a + (b + c)<\/code>.<\/p>\n\n<p>In most code you deal with daily the order of operations is fixed, so you only ever get to observe <code>(a + b) + c<\/code> or <code>a + (b + c)<\/code> (depending on the programmer wrote it), but in PyTorch, on CUDA backend, there are series of such operations which are not deterministically ordered (due to parallelism). This means that on one run of your <code>self.netG(self.real_A)<\/code> you can observe <code>(a + b) + c<\/code> and on another <code>a + (b + c)<\/code>. Without seeing your code one can't tell, but it is a likely the source of your surprise. A bit more is given in <a href=\"https:\/\/pytorch.org\/docs\/stable\/notes\/randomness.html#pytorch\" rel=\"nofollow noreferrer\">PyTorch docs<\/a>.<\/p>\n\n<p>EDIT: obviously, you can also switch your computations to 64-bit floating point numbers, which will improve the numerical accuracy (as it is commonly defined) of your calculations but is unlikely to help with nondeterminism (which is what you're actually complaining about).<\/p>\n"},{"AnswerId":"58201931","CreationDate":"2019-10-02T12:53:30.020","ParentId":null,"OwnerUserId":"11287895","Title":null,"Body":"<p>model.eval()\n This has any effect only on certain modules. See documentations of\n        particular modules for details of their behaviors in training\/evaluation\n        mode, if they are affected, e.g. :class:<code>Dropout<\/code>, :class:<code>BatchNorm<\/code>,\n        etc.<\/p>\n"}]}
+{"QuestionId":58155383,"AnswerCount":0,"Tags":"<pytorch>","CreationDate":"2019-09-29T12:53:12.897","AcceptedAnswerId":null,"OwnerUserId":10855529.0,"Title":"How do I load omniglot dataset in few shot setting?","Body":"<p>That is in 1 shot 5 way, or 5 shot 5 way, omniglot dataset is included in PyTorch 1.2.0, but how to load it in few shot setting?<\/p>\n","answers":[]}
+{"QuestionId":58155390,"AnswerCount":2,"Tags":"<python><tensorflow><nlp><finetunning>","CreationDate":"2019-09-29T12:54:17.467","AcceptedAnswerId":58200119.0,"OwnerUserId":3894673.0,"Title":"Improving BERT by training on additional data","Body":"<p>I have a BERT multilanguage model from Google. And I have a lot of text data in my language (Korean). I want BERT to make better vectors for texts in this language. So I want to additionally train BERT on that text corpus I have. Like if I would have w2v model trained on some data and would want to continue training it. Is it possible with BERT?<\/p>\n\n<p>There are a lot of examples of \"fine-tuning\" BERT on some specific tasks like even the <a href=\"https:\/\/github.com\/google-research\/bert#fine-tuning-with-bert\" rel=\"nofollow noreferrer\">original one from Google<\/a> where you can train BERT further on your data. But as far as I understand it (I might be wrong) we do it within our task-specified model (for classification task for example). So... we do it at the same time as training our classifier (??)<\/p>\n\n<p>What I want is to train BERT further separately and then get fixed vectors for my data. Not to build it  into some task-specified model. But just get vector representation for my data (using get_features function) like they do in <a href=\"https:\/\/towardsdatascience.com\/nlp-extract-contextualized-word-embeddings-from-bert-keras-tf-67ef29f60a7b\" rel=\"nofollow noreferrer\">here<\/a>. I just need to train the BERT model additionally on more data of the specific language.<\/p>\n\n<p>Would be endlessly grateful for any suggestions\/links on how to train BURT model further (preferably Tensorflow). Thank you.<\/p>\n","answers":[{"AnswerId":"58169024","CreationDate":"2019-09-30T13:34:01.243","ParentId":null,"OwnerUserId":"5652313","Title":null,"Body":"<p>Package <a href=\"https:\/\/github.com\/huggingface\/transformers\" rel=\"nofollow noreferrer\"><code>transformers<\/code><\/a> provides code for using and fine-tuning of most currently popular pre-trained Transformers including BERT, XLNet, GPT-2, ... You can easily load the model and continue training.<\/p>\n\n<p>You can get the multilingual BERT model:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tokenizer = BertTokenizer.from_pretrained('bert-base-multiligual-cased')\nmodel = TFBertForSequenceClassification.from_pretrained('bert-base-multiligual-cased')\n<\/code><\/pre>\n\n<p>The tokenizer is used both for tokenizing the input and for converting the sub-words into embedding ids. Calling the model on the subword indices will give you hidden states of the model.<\/p>\n\n<p>Unfortunately, the package does not implement the training procedure, i.e., the masked language model and the next sentence prediction. You will need to write it yourself, but the training procedure well described in the paper and the implementation will be straightforward.<\/p>\n"},{"AnswerId":"58200119","CreationDate":"2019-10-02T10:59:18.260","ParentId":null,"OwnerUserId":"3894673","Title":null,"Body":"<p>Everything appeared to be way easier than I thought.. I just followed <a href=\"https:\/\/github.com\/google-research\/bert#pre-training-with-bert\" rel=\"nofollow noreferrer\">these instructions<\/a>.. originally I thought they are only for training from scratch but actually you can continue training with this.\nIf you are getting OOM error just decrease the batch size.<\/p>\n"}]}
+{"QuestionId":58155502,"AnswerCount":0,"Tags":"<python><numpy><machine-learning><keras><neural-network>","CreationDate":"2019-09-29T13:07:45.427","AcceptedAnswerId":null,"OwnerUserId":12120467.0,"Title":"How to resolve, ValueError: setting an array element with a sequence?","Body":"<p>This is my code, getting error even when I did <code>one-hot encoding<\/code>, but that doesn't worked for me.Than I converted sequence to number by self But while compiling  error occurs \"ValueError: setting an array element with a sequence\"<\/p>\n\n<p>Please guide me how to fit data into Neural Netwrok, I have list of protein sequences in one column of csv file, and and in 2nd column their structures named <code>a, ab, b<\/code><\/p>\n\n<pre><code>df= pd.read_csv('\/home\/alpha\/mk fyp\/whole\/DATASET2.csv',names=('X1','Y'),delimiter=',')\nX1=df.iloc[:,0].values\nY=df.iloc[:,1:2].values\ndef convert(sequence):\nfor item in sequence:\n    try:\n        yield float(item)\n    except ValueError as e:\n        yield item\n\na = [[item for item in convert(sublist)] for sublist in X1]\nX = np.array(a)\nX_train, X_test, y_train, y_test = train_test_split(X,Y, test_size=0.20)\n# define a simple CNN model\ndef baseline_model():\n# create model\nmodel = Sequential()\nmodel.add(Dense(input_dim = 1, units = 10, activation='relu', kernel_initializer='uniform'))\n\nmodel.add(Dense(1, activation='relu'))\n# Compile model\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\nreturn model \n# build the model\nmodel = baseline_model()\n# Fit the model\nmodel.fit(X, Y, validation_data=(X1, Y), epochs=10, batch_size=200)\n# Final evaluation of the model\nscores = model.evaluate(X_test, y_test, verbose=0)\nprint(\"CNN Error: %.2f%%\" % (100-scores[1]*100))\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58155557,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><keras><deep-learning><loss-function>","CreationDate":"2019-09-29T13:15:48.893","AcceptedAnswerId":null,"OwnerUserId":7065438.0,"Title":"Why the training loss is fractuation when the validation loss is like a flat curve?","Body":"<p>I have Replaced the last two fully connected layers of VGG16 model with global average pooling, and use this CNN for binary classification (Class 0: benign, Class 1: Malignant) using mammography image. The learning rate was set as 0.0000001 and epoch= 500. The training loss curve is fluctuating within a large range (like a large zigzag), and the validation loss curve is like a flat straight line. I attach the figure of training loss vs validation loss curve as a figure for better understanding.<a href=\"https:\/\/i.stack.imgur.com\/L2LFc.jpg\" rel=\"nofollow noreferrer\">training loss vs validation loss curve<\/a><\/p>\n\n<p>On the other hand, The training accuracy curve is also fluctuating within a large range (like a large zigzag), and the validation accuracy curve is like a straight line.I also attach the figure of training accuracy vs validation accuracy curve as a figure for better understanding.<a href=\"https:\/\/i.stack.imgur.com\/rRf9s.jpg\" rel=\"nofollow noreferrer\">training accuracy vs validation accuracy curve<\/a><\/p>\n\n<p>For training 280(benign: 136, malignant: 144) images were used data augmentation applied. For validation, 27 (benign: 14, malignant: 13)images were used. And for testing, 21  (benign: 5, malignant: 16) images were used. The Confusion matrix looks OK : <a href=\"https:\/\/i.stack.imgur.com\/si3xU.jpg\" rel=\"nofollow noreferrer\">Confusion matrix<\/a><\/p>\n\n<p>My question is, why the training loss curve is fluctuating, when the validation loss curve is always flat? And also, why my training accuracy curve is fluctuating, when the validation accuracy curve is always flat? <\/p>\n\n<pre><code># Show the image paths\ntrain_path = 'Digital_Mamo\/OPTIMAM'      # Relative Path\nvalid_path = 'Digital_Mamo\/InBreast'\ntest_path = 'Digital_Mamo\/BCDR'\n\n\nvalid_batches = ImageDataGenerator().flow_from_directory(valid_path, target_size=(224, 224), classes=['Benign', 'Malignant'], batch_size=9) #, color_mode=\"grayscale\"\ntest_batches = ImageDataGenerator().flow_from_directory(test_path, target_size=(224, 224), classes=['Benign', 'Malignant'], batch_size=7)\n\ndatagen = ImageDataGenerator(rotation_range=10, width_shift_range=0.1,\n   height_shift_range=0.1, shear_range=0.15, zoom_range=0.1,\n   channel_shift_range=10., horizontal_flip=True)\n\ntrain_batches = datagen.flow_from_directory(\n    train_path,\n    target_size=(224, 224),\n    batch_size=10,\n    classes=['Benign','Malignant'])\n\n\nvgg16_model= load_model('Fetched_VGG.h5')\n\n# transform the model to Sequential\nfor layer in vgg16_model.layers[:-4]:\n    model.add(layer)\n\nmodel.add(GlobalAveragePooling2D())\n\n# Freezing the layers (Oppose weights to be updated)\nfor layer in model.layers:\nlayer.trainable = False\n\nmodel.add(Dense(2, activation='softmax', name='predictions'))\n\n ### Compile the model\n model.compile(Adam(lr=.0000001), loss='categorical_crossentropy',        metrics=['accuracy'])   # lr=.0001\n\n# train the model\nhist = model.fit_generator(train_batches, steps_per_epoch=28,   validation_data=valid_batches, validation_steps=3, epochs=500, verbose=2)  # epochs=5\n\nscoreSeg = model.evaluate_generator(test_batches, steps=3)\nprint('Test Loss:', scoreSeg[0])\nprint('Test accuracy:', scoreSeg[1])\n\n#test\npredictions = model.predict_generator(test_batches, steps=3, verbose=0)\n#print(predictions)\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58155561,"AnswerCount":0,"Tags":"<python><tensorflow><jupyter-notebook>","CreationDate":"2019-09-29T13:16:00.120","AcceptedAnswerId":null,"OwnerUserId":11927380.0,"Title":"How to a Trained model from Jupyter notebook to android studio using tensorflow?","Body":"<p>I've trained a model using AlexNet on Tensorflow. Now that I've trained it successfully, I'm not sure how am I going to save it and deploy it for my android app. Can anyone help me with this? There are many help available online but I'm doubting where to start and how to do it really. <\/p>\n\n<p>This is the code for training it:<\/p>\n\n<pre><code>%%time\n#Changing settings for GPU running.\nconfig = tf.ConfigProto(allow_soft_placement=True)\nconfig.gpu_options.allow_growth = True\nconfig.gpu_options.allocator_type = 'BFC'\n#Training and saving the result\nwith tf.Session(config=config) as sess:\nsess.run(init)\nsummary_writer = tf.summary.FileWriter(TRAIN_DIR, graph=tf.get_default_graph())\nfor i in range(epochs):\nfor j in range(0,steps,step_size):       \n _ , c , summary,d = sess.run([train,cross_entropy,merged_summary_op,acc],feed_dict= \n{x:X[j:j+step_size] , y_true:Y[j:j+step_size] \n,hold_prob1:0.5,hold_prob2:0.5,hold_prob3:0.5,hold_prob4:0.5})\n        summary_writer.add_summary(summary, i * total_batch + j)\n        acc_train.append(d)\n        mean_of_cross_entropy = sess.run(cross_entropy,feed_dict={x:cv_x,y_true:cv_y \n,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0})\n        mean_of_acc = sess.run(acc,feed_dict={x:cv_x ,y_true:cv_y \n,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0})\n        cross_entropy_list.append(mean_of_cross_entropy)\n        acc_list.append(mean_of_acc)\nprint(i,mean_of_cross_entropy,mean_of_acc)\nsaver.save(sess, \"C:\\\\Users\\\\blessie\\\\Desktop\\\\LEAF RECOGNITION v5\\\\Models\\\\CNN_MC.ckpt\")\nprint(\"test accuracy = \",np.mean([sess.run(acc,feed_dict={x:test_x[:230],y_true:test_y[:230] \n<\/code><\/pre>\n\n<p>,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0}),sess.run(acc,feed_dict= \n   {x:test_x[230:460],y_true:test_y[230:460] \n   ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0}),sess.run(acc,feed_dict= \n   {x:test_x[460:],y_true:test_y[460:] ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0})]))<br>\n    print(\"cross_entropy loss = \",np.mean([sess.run(cross_entropy,feed_dict= \n   {x:test_x[:230],y_true:test_y[:230] \n   ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0}),sess.run(cross_entropy,feed_dict= \n   {x:test_x[230:460],y_true:test_y[230:460] \n   ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0}),sess.run(cross_entropy,feed_dict= \n   {x:test_x[460:],y_true:test_y[460:] ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0})]))<\/p>\n","answers":[]}
+{"QuestionId":58155844,"AnswerCount":0,"Tags":"<python><tensorflow><keras><tf.keras>","CreationDate":"2019-09-29T13:47:53.700","AcceptedAnswerId":null,"OwnerUserId":3235916.0,"Title":"Custom RNN with in tensorflow 2 (python)","Body":"<p>I am trying to implement a very simple RNN using the keras model class in tensorflow 2 (Python 3.7) that predicts an output given a sequence of inputs. My goal here is mainly didactic as its obviously fairly trivial to implement these models using the keras layers.<\/p>\n\n<p>Using the code I have written I can call an instance of the model with input and receive an output. (Data here is obviously just noise)<\/p>\n\n<p>i.e. <\/p>\n\n<pre><code>x_train = np.random.normal(size=[20,1000,1])\ny_train = np.random.normal(size=1000)\nbatch_size = 5\nrnn = simpleRNN(20,batch_size)\nrnn.compile(optimizer='adam', loss='mae')\nrnn(x_train[:,:5])\n<\/code><\/pre>\n\n<p>works just fine<\/p>\n\n<p>however, when I'm trying to fit the model parameters by running something like<\/p>\n\n<pre><code>epochs= 1\nN = 50\nrnn.fit(tf.cast(x_train[:,:N],dtype=tf.float64),\n        tf.cast(y_train[:N].reshape(1,-1),dtype=tf.float64),\n        batch_size=batch_size, epochs=epochs, verbose=1)\n<\/code><\/pre>\n\n<p>I get all sorts of errors. I have tried a bunch of different things and always seems to get different errors which leads me to think there is something I haven't understood.<\/p>\n\n<p>The current error I'm getting is<\/p>\n\n<pre><code> ValueError: Dimensions must be equal, but are 5 and 50 for 'add_2' (op: 'AddV2') with input shapes: [5,13], [50,13].\n<\/code><\/pre>\n\n<p>Model class and imports below. <\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nimport tensorflow.keras as keras\n\nclass simpleRNN(keras.Model):\n\n    def __init__(self, units,batch_size):\n        super(simpleRNN, self).__init__()\n        self.batch_size = batch_size\n        self.init_state = tf.random.normal(shape=(batch_size,units),dtype=tf.float64)\n\n        self.inW = tf.Variable(np.random.rand(1,units))\n        self.inb = tf.Variable(np.random.rand(1,units))\n\n        self.W = tf.Variable(np.random.rand(units, units), dtype=tf.float64)\n        self.b = tf.Variable(np.zeros((1,units)), dtype=tf.float64)\n\n        self.W2 = tf.Variable(np.random.rand(units,1),dtype=tf.float64)\n        self.b2 = tf.Variable(np.zeros((1)), dtype=tf.float64)\n\n\n    def call(self, inputs):\n        inputs = tf.cast(inputs,tf.float64)\n        current_state = self.init_state\n        #print(cur)\n\n        all_pred = []\n        out = tf.zeros([self.batch_size,1],tf.float64)\n        for x in inputs:\n            print(x.shape)\n            intoRec = tf.matmul(x,self.inW) + self.inb\n\n            current_state = tf.tanh(tf.matmul(current_state, self.W)+self.b +intoRec) \n            print(current_state.shape,intoRec.shape)\n            logits = tf.matmul(current_state,self.W2) + self.b2\n            out = tf.nn.tanh(logits)\n\n        return out\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58155920,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><deep-learning><conv-neural-network><pytorch>","CreationDate":"2019-09-29T13:56:57.570","AcceptedAnswerId":58165847.0,"OwnerUserId":7512010.0,"Title":"Should I transpose a Tensor when feeding it into a CNN","Body":"<p>I am using a custom dataset with images of different sizes in the Lab format (Lightness, a, b) which are feed into a CNN. The input layer has 3 in-channels and so my idea was to split all 3 channels (L, a, b) and feed those into the network. Next I was wondering if each tensor needs to be transposed? My doubt is that it would lose its dimensions which are variable from image to image and I would not be able to reconstruct the image in the end. Any thoughts or ideas how I should normalize the image?<\/p>\n","answers":[{"AnswerId":"58165847","CreationDate":"2019-09-30T10:19:27.253","ParentId":null,"OwnerUserId":"7872285","Title":null,"Body":"<p>You can normalise without the need of transposing the image or splitting it based on its channels <\/p>\n\n<p><code>torchvision.transforms.Normalize(mean=[l_channel_mean, a_channel_mean , b_channel_mean], std= [l_channel_mean, a_channel_mean , b_channel_mean])\n<\/code>\nThe only required transform is the one that converts the images to tensors : \n<code>\ntorchvision.transforms.ToTensor()\n<\/code><\/p>\n"}]}
+{"QuestionId":58155999,"AnswerCount":2,"Tags":"<python><arrays><pytorch><slice><tensor>","CreationDate":"2019-09-29T14:07:47.620","AcceptedAnswerId":null,"OwnerUserId":1667423.0,"Title":"Assign a tensor to multiple slices","Body":"<p>Let<\/p>\n\n<pre><code>a = tensor([[0, 0, 0, 0],\n            [0, 0, 0, 0],\n            [0, 0, 0, 0]])\nb = torch.tensor([1, 2])\nc = tensor([[1, 2, 0, 0],\n            [0, 1, 2, 0],\n            [0, 0, 1, 2]])\n<\/code><\/pre>\n\n<p>Is there a way to obtain <code>c<\/code> by assigning <code>b<\/code> to slices of <code>a<\/code> without any loops? That is, <code>a[indices] = b<\/code> for some <code>indices<\/code> or something similar?<\/p>\n","answers":[{"AnswerId":"58157213","CreationDate":"2019-09-29T16:40:38.697","ParentId":null,"OwnerUserId":"7212365","Title":null,"Body":"<p>You can use <code>scatter<\/code> method in pytorch. <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>a = torch.tensor([[0, 0, 0, 0],\n                 [0, 0, 0, 0],\n                 [0, 0, 0, 0]])\n\nb = torch.tensor([1, 2])\n\nindex = torch.tensor([[0,1],[1,2],[2,3]])\n\na.scatter_(1, index, b.view(-1,2).repeat(3,1))\n# tensor([[1, 2, 0, 0],\n#         [0, 1, 2, 0],\n#         [0, 0, 1, 2]])\n<\/code><\/pre>\n"},{"AnswerId":"58165827","CreationDate":"2019-09-30T10:18:19.697","ParentId":null,"OwnerUserId":"5218354","Title":null,"Body":"<p>The logic behind this operation is a bit iffy in the sense that it is not clear what the parameters of the operation are.\nHowever, one way of obtaining the desired output from the input with vectorized operations only is:<\/p>\n\n<ul>\n<li>determine how many rows are needed (<code>3<\/code> for your example)<\/li>\n<li>create a <code>a<\/code> with a number of columns such that <code>b<\/code> is followed by as many zeros as the number of rows (<code>2 + 3<\/code>), and the choosen number of rows (<code>3<\/code>)<\/li>\n<li>assign <code>b<\/code> to the beginning of <code>a<\/code> for each<\/li>\n<li>flatten the array, cut <code>num_rows<\/code> zeros, and reshape to the target shape.<\/li>\n<\/ul>\n\n<p>In NumPy, this can be implemented as:<\/p>\n\n<pre><code>import numpy as np\n\n\nb = np.array([1, 2])\nc = np.array([[1, 2, 0, 0],\n              [0, 1, 2, 0],\n              [0, 0, 1, 2]])\n\nnum_rows = 3\na = np.zeros((num_rows, len(b) + num_rows), dtype=b.dtype)\na[:, :len(b)] = b\na = a.ravel()[:-num_rows].reshape((num_rows, len(b) + num_rows - 1))\n\nprint(a)\n# [[1 2 0 0]\n#  [0 1 2 0]\n#  [0 0 1 2]]\n\nprint(np.all(a == c))\n# True\n<\/code><\/pre>\n\n<hr>\n\n<h1>EDIT<\/h1>\n\n<p>The same approach implemented in Torch:<\/p>\n\n<pre><code>import torch as to\n\n\nb = to.tensor([1, 2])\nc = to.tensor([[1, 2, 0, 0],\n               [0, 1, 2, 0],\n               [0, 0, 1, 2]])\n\nnum_rows = 3\na = to.zeros((num_rows, len(b) + num_rows), dtype=b.dtype)\na[:, :len(b)] = b\na = a.flatten()[:-num_rows].reshape((num_rows, len(b) + num_rows - 1))\n\nprint(a)\n# tensor([[1, 2, 0, 0],\n#         [0, 1, 2, 0],\n#         [0, 0, 1, 2]])\n\nprint(to.all(a == c))\n# tensor(1, dtype=torch.uint8)\n<\/code><\/pre>\n"}]}
+{"QuestionId":58156129,"AnswerCount":0,"Tags":"<python><tensorflow><eager-execution>","CreationDate":"2019-09-29T14:23:39.247","AcceptedAnswerId":null,"OwnerUserId":12138827.0,"Title":"Training in Tensorflow 2.0 with eager execution without keras","Body":"<p>I am new to machine learning and trying to understand how tensorflow 2.0 works, but can't find any comprehensive resouces in the internet which use new API. So, I has started to explore it by myself and that's the big question I have found - how can I train model with eager execution? That's my code that I've written implementing SVM linear model (using example from Tensorflow cookbook), and now I just don't understand what to do next in order to train this model. Thank you for your help.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport random\n\n\nalpha_ = 0.1\n\n\nnumber_of_predictors = 20\nnumber_of_samples = 100\nnumber_of_predictions = 3\n\n\n#\n# set up dataset\n#\n\n\npredictors = [np.random.normal(size = number_of_predictors).astype('f') for i in range(number_of_samples)]\nlabels = [[random.choice([-1., 1.]) for i in range(number_of_predictions)] for i in range(number_of_samples)]\n\n\n#\n# set up model\n#\n\n\nslope = tf.Variable(tf.random.normal(shape=[number_of_predictors, number_of_predictions]))\nintercept = tf.Variable(tf.random.normal(shape=[1, number_of_predictions]))\n\n\n# model prediction\npredictions = tf.subtract(tf.matmul(predictors, slope), intercept)\n\n\n# regularization\nl2_norm = tf.reduce_sum(tf.square(slope))\nalpha = tf.constant([alpha_])\n\n\n# assemble model prediction and regularization\nclassification_term = tf.reduce_mean(tf.maximum(0., tf.subtract(1., tf.multiply(predictions, labels))))\nloss = tf.add(classification_term, tf.multiply(l2_norm, alpha))\n\n\nprediction = tf.sign(predictions)\nprint(f\"Predictions: {predictions}\")\naccuracy = tf.reduce_mean(tf.cast(tf.equal(predictions, labels), dtype = tf.float32))\nprint(f\"Accuracy: {accuracy}\")\n<\/code><\/pre>\n","answers":[]}
+{"QuestionId":58156202,"AnswerCount":1,"Tags":"<tensorflow><tensorflow2.0>","CreationDate":"2019-09-29T14:32:54.703","AcceptedAnswerId":null,"OwnerUserId":8917781.0,"Title":"a bizarre mistake-----InvalidArgumentError: slice index -1 of dimension 0 out of bounds.\uff08\uff09","Body":"<p>My computer is installed with win10, tf-nightly-2.0-preview==2.0.0dev20190926\uff0cpython3.7&amp;python3.6.8<\/p>\n\n<p>Tensorflow standard template: <a href=\"https:\/\/tensorflow.google.cn\/tutorials\/load_data\/images#setup\" rel=\"nofollow noreferrer\">https:\/\/tensorflow.google.cn\/tutorials\/load_data\/images#setup<\/a><\/p>\n\n<p>When I run the tensorflow standard template, I get an error in<\/p>\n\n<pre><code>for image, label in labeled_ds.take(1):\n    print(\"Image shape: \", image.numpy().shape)\n    print(\"Label: \", label.numpy())\n<\/code><\/pre>\n\n<p>the error message:<\/p>\n\n<pre><code>InvalidArgumentError: slice index -1 of dimension 0 out of bounds. [[{{node strided_slice}}]]\nEncountered when executing an operation using EagerExecutor. \n<\/code><\/pre>\n\n<p>This error cancels all future operations and poisons their output tensors.<\/p>\n\n<p>However, in google colab I installed tf-nightly-2.0-preview==2.0.0dev20190926, python==3.6,\nRunning the code but it runs fine without errors<\/p>\n\n<p>If I want to run this standard template code on my computer, how should I modify it?<\/p>\n","answers":[{"AnswerId":"58738093","CreationDate":"2019-11-06T20:34:54.633","ParentId":null,"OwnerUserId":"1184789","Title":null,"Body":"<p>I found that this was a problem with <code>tf.strings.split<\/code> within the provided <code>get_label<\/code> method. On Colab splitting on a single forward slash worked, but on my local Windows machine I needed to split on a backslash.<\/p>\n\n<p>So if your Colab code looks like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def get_label(file_path):\n    parts = tf.strings.split(file_path, '\/')\n    return parts[-2] == settings.CLASS_NAMES\n<\/code><\/pre>\n\n<p>Change it to look like this when running locally:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def get_label(file_path):\n    parts = tf.strings.split(file_path, '\\\\')\n    return parts[-2] == settings.CLASS_NAMES\n<\/code><\/pre>\n"}]}
+{"QuestionId":58156378,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2019-09-29T14:57:47.750","AcceptedAnswerId":58156655.0,"OwnerUserId":11298280.0,"Title":"What is Coded Feature in Auto Encoder ? Is there any formula used to calculate the Coded Feature?","Body":"<p>I am working on AutoCoder. I have drawn graphs for Input values and Reconstructed values using Matplotlib. How to draw graph for Coded Features and what is Coded Features?<\/p>\n","answers":[{"AnswerId":"58156655","CreationDate":"2019-09-29T15:27:43.700","ParentId":null,"OwnerUserId":"9149062","Title":null,"Body":"<p>An auto-encoder is a type of neural network where you combine an encoder and a decoder. Typically the output of the encoder has less parameters than the input and the decoder takes this reduced number of parameters to recreate the original input as closely as possible. You can think of the output of the encoder (also named 'latent vector') as a lossy compression of the input. There are however multiple other use-cases and types of auto-encoders. You can read more about them on <a href=\"https:\/\/en.wikipedia.org\/wiki\/Autoencoder\" rel=\"nofollow noreferrer\">Wikipedia<\/a> and <a href=\"https:\/\/towardsdatascience.com\/auto-encoder-what-is-it-and-what-is-it-used-for-part-1-3e5c6f017726\" rel=\"nofollow noreferrer\">this blogpost on towardsdatascience.com<\/a> also gives a very good explanation with examples.<\/p>\n\n<p>If you want an explanation on how to draw graphs, you will need to give us more information on your input dataset and maybe show us your graphs of the input and output, because I'm not sure what you mean.<\/p>\n"}]}
+{"QuestionId":58156384,"AnswerCount":1,"Tags":"<python><tensorflow><keras><tensorboard><keras-rl>","CreationDate":"2019-09-29T14:58:58.920","AcceptedAnswerId":58277912.0,"OwnerUserId":9704382.0,"Title":"TypeError: 'module' object is not callable Tensorboard in Keras","Body":"<p>I am implementing a RL agent with policy gradient method. I define a dense network for actor and another dense network for critic. For example, my critic network is:<\/p>\n\n<pre><code>state_input = Input(shape=(self.num_states,))\nx = Dense(self.hidden_size, activation='tanh')(state_input)\nfor _ in range(self.num_layers - 1):\n    x = Dense(self.hidden_size, activation='tanh')(x)\n\nout_value = Dense(1)(x)\n\nmodel = Model(inputs=[state_input], outputs=[out_value])\nmodel.compile(optimizer=SGD(lr=self.learning_rate), loss='mse')\n<\/code><\/pre>\n\n<p>In training phase I'm calling tensorboard:<\/p>\n\n<pre><code>from keras.callbacks import TensorBoard\n\ntensorboard = TensorBoard(log_dir=\"\/logs\/{}\".format(time()), \n                          histogram_freq=1, batch_size=32, \n                          write_graph=True, write_grads=True,\n                          write_images=True, embeddings_freq=0, \n                          embeddings_layer_names=None, \n                          embeddings_metadata=None, \n                          embeddings_data=None, update_freq='epoch')\n\ncritic_loss = self.critic.fit([obs, advantage, old_prediction], [action], \n                                 batch_size=self.batch_size,\n                                 shuffle=True, epochs=self.epochs, verbose=False,\n                                 callbacks=[tensorboard_actor])\n<\/code><\/pre>\n\n<p>But I'm getting this error:<\/p>\n\n<pre><code>TypeError: 'module' object is not callable\n<\/code><\/pre>\n","answers":[{"AnswerId":"58277912","CreationDate":"2019-10-07T22:29:39.790","ParentId":null,"OwnerUserId":"9037648","Title":null,"Body":"<blockquote>\n  <p>TypeError: 'module' object is not callable\n  in your case is caused by \n  time\n  module<\/p>\n<\/blockquote>\n\n<p>I am assuming that you imported time module as <\/p>\n\n<pre><code>import time\n<\/code><\/pre>\n\n<p>and called the function time()<\/p>\n\n<pre><code>tensorboard = TensorBoard(log_dir=\"\/logs\/{}\".format(time())\n<\/code><\/pre>\n\n<p>It can be solved easily by importing:<\/p>\n\n<pre><code>from time import time\n<\/code><\/pre>\n"}]}